1457c8996SThomas Gleixner // SPDX-License-Identifier: GPL-2.0-only 21da177e4SLinus Torvalds /* 31da177e4SLinus Torvalds * Generic hugetlb support. 46d49e352SNadia Yvette Chambers * (C) Nadia Yvette Chambers, April 2004 51da177e4SLinus Torvalds */ 61da177e4SLinus Torvalds #include <linux/list.h> 71da177e4SLinus Torvalds #include <linux/init.h> 81da177e4SLinus Torvalds #include <linux/mm.h> 9e1759c21SAlexey Dobriyan #include <linux/seq_file.h> 101da177e4SLinus Torvalds #include <linux/sysctl.h> 111da177e4SLinus Torvalds #include <linux/highmem.h> 12cddb8a5cSAndrea Arcangeli #include <linux/mmu_notifier.h> 131da177e4SLinus Torvalds #include <linux/nodemask.h> 1463551ae0SDavid Gibson #include <linux/pagemap.h> 155da7ca86SChristoph Lameter #include <linux/mempolicy.h> 163b32123dSGideon Israel Dsouza #include <linux/compiler.h> 17aea47ff3SChristoph Lameter #include <linux/cpuset.h> 183935baa9SDavid Gibson #include <linux/mutex.h> 1997ad1087SMike Rapoport #include <linux/memblock.h> 20a3437870SNishanth Aravamudan #include <linux/sysfs.h> 215a0e3ad6STejun Heo #include <linux/slab.h> 22bbe88753SJoonsoo Kim #include <linux/sched/mm.h> 2363489f8eSMike Kravetz #include <linux/mmdebug.h> 24174cd4b1SIngo Molnar #include <linux/sched/signal.h> 250fe6e20bSNaoya Horiguchi #include <linux/rmap.h> 26c6247f72SMatthew Wilcox #include <linux/string_helpers.h> 27fd6a03edSNaoya Horiguchi #include <linux/swap.h> 28fd6a03edSNaoya Horiguchi #include <linux/swapops.h> 298382d914SDavidlohr Bueso #include <linux/jhash.h> 3098fa15f3SAnshuman Khandual #include <linux/numa.h> 31c77c0a8aSWaiman Long #include <linux/llist.h> 32cf11e85fSRoman Gushchin #include <linux/cma.h> 338cc5fcbbSMina Almasry #include <linux/migrate.h> 34f9317f77SMike Kravetz #include <linux/nospec.h> 35662ce1dcSYang Yang #include <linux/delayacct.h> 36b958d4d0SMuchun Song #include <linux/memory.h> 37d6606683SLinus Torvalds 3863551ae0SDavid Gibson #include <asm/page.h> 39ca15ca40SMike Rapoport #include <asm/pgalloc.h> 4024669e58SAneesh Kumar K.V #include <asm/tlb.h> 4163551ae0SDavid Gibson 4224669e58SAneesh Kumar K.V #include <linux/io.h> 4363551ae0SDavid Gibson #include <linux/hugetlb.h> 449dd540e2SAneesh Kumar K.V #include <linux/hugetlb_cgroup.h> 459a305230SLee Schermerhorn #include <linux/node.h> 46ab5ac90aSMichal Hocko #include <linux/page_owner.h> 477835e98bSNick Piggin #include "internal.h" 48f41f2ed4SMuchun Song #include "hugetlb_vmemmap.h" 491da177e4SLinus Torvalds 50c3f38a38SAneesh Kumar K.V int hugetlb_max_hstate __read_mostly; 51e5ff2159SAndi Kleen unsigned int default_hstate_idx; 52e5ff2159SAndi Kleen struct hstate hstates[HUGE_MAX_HSTATE]; 53cf11e85fSRoman Gushchin 54dbda8feaSBarry Song #ifdef CONFIG_CMA 55cf11e85fSRoman Gushchin static struct cma *hugetlb_cma[MAX_NUMNODES]; 5638e719abSBaolin Wang static unsigned long hugetlb_cma_size_in_node[MAX_NUMNODES] __initdata; 572f6c57d6SSidhartha Kumar static bool hugetlb_cma_folio(struct folio *folio, unsigned int order) 58a01f4390SMike Kravetz { 592f6c57d6SSidhartha Kumar return cma_pages_valid(hugetlb_cma[folio_nid(folio)], &folio->page, 60a01f4390SMike Kravetz 1 << order); 61a01f4390SMike Kravetz } 62a01f4390SMike Kravetz #else 632f6c57d6SSidhartha Kumar static bool hugetlb_cma_folio(struct folio *folio, unsigned int order) 64a01f4390SMike Kravetz { 65a01f4390SMike Kravetz return false; 66a01f4390SMike Kravetz } 67dbda8feaSBarry Song #endif 68dbda8feaSBarry Song static unsigned long hugetlb_cma_size __initdata; 69cf11e85fSRoman Gushchin 7053ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 7153ba51d2SJon Tollefson 72e5ff2159SAndi Kleen /* for command line parsing */ 73e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 74e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 759fee021dSVaishali Thakkar static bool __initdata parsed_valid_hugepagesz = true; 76282f4214SMike Kravetz static bool __initdata parsed_default_hugepagesz; 77b5389086SZhenguo Yao static unsigned int default_hugepages_in_node[MAX_NUMNODES] __initdata; 78e5ff2159SAndi Kleen 793935baa9SDavid Gibson /* 8031caf665SNaoya Horiguchi * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, 8131caf665SNaoya Horiguchi * free_huge_pages, and surplus_huge_pages. 823935baa9SDavid Gibson */ 83c3f38a38SAneesh Kumar K.V DEFINE_SPINLOCK(hugetlb_lock); 840bd0f9fbSEric Paris 858382d914SDavidlohr Bueso /* 868382d914SDavidlohr Bueso * Serializes faults on the same logical page. This is used to 878382d914SDavidlohr Bueso * prevent spurious OOMs when the hugepage pool is fully utilized. 888382d914SDavidlohr Bueso */ 898382d914SDavidlohr Bueso static int num_fault_mutexes; 90c672c7f2SMike Kravetz struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; 918382d914SDavidlohr Bueso 927ca02d0aSMike Kravetz /* Forward declaration */ 937ca02d0aSMike Kravetz static int hugetlb_acct_memory(struct hstate *h, long delta); 948d9bfb26SMike Kravetz static void hugetlb_vma_lock_free(struct vm_area_struct *vma); 958d9bfb26SMike Kravetz static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma); 96ecfbd733SMike Kravetz static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma); 97b30c14cdSJames Houghton static void hugetlb_unshare_pmds(struct vm_area_struct *vma, 98b30c14cdSJames Houghton unsigned long start, unsigned long end); 997ca02d0aSMike Kravetz 1001d88433bSMiaohe Lin static inline bool subpool_is_free(struct hugepage_subpool *spool) 1011d88433bSMiaohe Lin { 1021d88433bSMiaohe Lin if (spool->count) 1031d88433bSMiaohe Lin return false; 1041d88433bSMiaohe Lin if (spool->max_hpages != -1) 1051d88433bSMiaohe Lin return spool->used_hpages == 0; 1061d88433bSMiaohe Lin if (spool->min_hpages != -1) 1071d88433bSMiaohe Lin return spool->rsv_hpages == spool->min_hpages; 1081d88433bSMiaohe Lin 1091d88433bSMiaohe Lin return true; 1101d88433bSMiaohe Lin } 1111d88433bSMiaohe Lin 112db71ef79SMike Kravetz static inline void unlock_or_release_subpool(struct hugepage_subpool *spool, 113db71ef79SMike Kravetz unsigned long irq_flags) 11490481622SDavid Gibson { 115db71ef79SMike Kravetz spin_unlock_irqrestore(&spool->lock, irq_flags); 11690481622SDavid Gibson 11790481622SDavid Gibson /* If no pages are used, and no other handles to the subpool 1187c8de358SEthon Paul * remain, give up any reservations based on minimum size and 1197ca02d0aSMike Kravetz * free the subpool */ 1201d88433bSMiaohe Lin if (subpool_is_free(spool)) { 1217ca02d0aSMike Kravetz if (spool->min_hpages != -1) 1227ca02d0aSMike Kravetz hugetlb_acct_memory(spool->hstate, 1237ca02d0aSMike Kravetz -spool->min_hpages); 12490481622SDavid Gibson kfree(spool); 12590481622SDavid Gibson } 1267ca02d0aSMike Kravetz } 12790481622SDavid Gibson 1287ca02d0aSMike Kravetz struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, 1297ca02d0aSMike Kravetz long min_hpages) 13090481622SDavid Gibson { 13190481622SDavid Gibson struct hugepage_subpool *spool; 13290481622SDavid Gibson 133c6a91820SMike Kravetz spool = kzalloc(sizeof(*spool), GFP_KERNEL); 13490481622SDavid Gibson if (!spool) 13590481622SDavid Gibson return NULL; 13690481622SDavid Gibson 13790481622SDavid Gibson spin_lock_init(&spool->lock); 13890481622SDavid Gibson spool->count = 1; 1397ca02d0aSMike Kravetz spool->max_hpages = max_hpages; 1407ca02d0aSMike Kravetz spool->hstate = h; 1417ca02d0aSMike Kravetz spool->min_hpages = min_hpages; 1427ca02d0aSMike Kravetz 1437ca02d0aSMike Kravetz if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { 1447ca02d0aSMike Kravetz kfree(spool); 1457ca02d0aSMike Kravetz return NULL; 1467ca02d0aSMike Kravetz } 1477ca02d0aSMike Kravetz spool->rsv_hpages = min_hpages; 14890481622SDavid Gibson 14990481622SDavid Gibson return spool; 15090481622SDavid Gibson } 15190481622SDavid Gibson 15290481622SDavid Gibson void hugepage_put_subpool(struct hugepage_subpool *spool) 15390481622SDavid Gibson { 154db71ef79SMike Kravetz unsigned long flags; 155db71ef79SMike Kravetz 156db71ef79SMike Kravetz spin_lock_irqsave(&spool->lock, flags); 15790481622SDavid Gibson BUG_ON(!spool->count); 15890481622SDavid Gibson spool->count--; 159db71ef79SMike Kravetz unlock_or_release_subpool(spool, flags); 16090481622SDavid Gibson } 16190481622SDavid Gibson 1621c5ecae3SMike Kravetz /* 1631c5ecae3SMike Kravetz * Subpool accounting for allocating and reserving pages. 1641c5ecae3SMike Kravetz * Return -ENOMEM if there are not enough resources to satisfy the 1659e7ee400SRandy Dunlap * request. Otherwise, return the number of pages by which the 1661c5ecae3SMike Kravetz * global pools must be adjusted (upward). The returned value may 1671c5ecae3SMike Kravetz * only be different than the passed value (delta) in the case where 1687c8de358SEthon Paul * a subpool minimum size must be maintained. 1691c5ecae3SMike Kravetz */ 1701c5ecae3SMike Kravetz static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, 17190481622SDavid Gibson long delta) 17290481622SDavid Gibson { 1731c5ecae3SMike Kravetz long ret = delta; 17490481622SDavid Gibson 17590481622SDavid Gibson if (!spool) 1761c5ecae3SMike Kravetz return ret; 17790481622SDavid Gibson 178db71ef79SMike Kravetz spin_lock_irq(&spool->lock); 17990481622SDavid Gibson 1801c5ecae3SMike Kravetz if (spool->max_hpages != -1) { /* maximum size accounting */ 1811c5ecae3SMike Kravetz if ((spool->used_hpages + delta) <= spool->max_hpages) 1821c5ecae3SMike Kravetz spool->used_hpages += delta; 1831c5ecae3SMike Kravetz else { 1841c5ecae3SMike Kravetz ret = -ENOMEM; 1851c5ecae3SMike Kravetz goto unlock_ret; 1861c5ecae3SMike Kravetz } 1871c5ecae3SMike Kravetz } 1881c5ecae3SMike Kravetz 18909a95e29SMike Kravetz /* minimum size accounting */ 19009a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->rsv_hpages) { 1911c5ecae3SMike Kravetz if (delta > spool->rsv_hpages) { 1921c5ecae3SMike Kravetz /* 1931c5ecae3SMike Kravetz * Asking for more reserves than those already taken on 1941c5ecae3SMike Kravetz * behalf of subpool. Return difference. 1951c5ecae3SMike Kravetz */ 1961c5ecae3SMike Kravetz ret = delta - spool->rsv_hpages; 1971c5ecae3SMike Kravetz spool->rsv_hpages = 0; 1981c5ecae3SMike Kravetz } else { 1991c5ecae3SMike Kravetz ret = 0; /* reserves already accounted for */ 2001c5ecae3SMike Kravetz spool->rsv_hpages -= delta; 2011c5ecae3SMike Kravetz } 2021c5ecae3SMike Kravetz } 2031c5ecae3SMike Kravetz 2041c5ecae3SMike Kravetz unlock_ret: 205db71ef79SMike Kravetz spin_unlock_irq(&spool->lock); 20690481622SDavid Gibson return ret; 20790481622SDavid Gibson } 20890481622SDavid Gibson 2091c5ecae3SMike Kravetz /* 2101c5ecae3SMike Kravetz * Subpool accounting for freeing and unreserving pages. 2111c5ecae3SMike Kravetz * Return the number of global page reservations that must be dropped. 2121c5ecae3SMike Kravetz * The return value may only be different than the passed value (delta) 2131c5ecae3SMike Kravetz * in the case where a subpool minimum size must be maintained. 2141c5ecae3SMike Kravetz */ 2151c5ecae3SMike Kravetz static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, 21690481622SDavid Gibson long delta) 21790481622SDavid Gibson { 2181c5ecae3SMike Kravetz long ret = delta; 219db71ef79SMike Kravetz unsigned long flags; 2201c5ecae3SMike Kravetz 22190481622SDavid Gibson if (!spool) 2221c5ecae3SMike Kravetz return delta; 22390481622SDavid Gibson 224db71ef79SMike Kravetz spin_lock_irqsave(&spool->lock, flags); 2251c5ecae3SMike Kravetz 2261c5ecae3SMike Kravetz if (spool->max_hpages != -1) /* maximum size accounting */ 22790481622SDavid Gibson spool->used_hpages -= delta; 2281c5ecae3SMike Kravetz 22909a95e29SMike Kravetz /* minimum size accounting */ 23009a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { 2311c5ecae3SMike Kravetz if (spool->rsv_hpages + delta <= spool->min_hpages) 2321c5ecae3SMike Kravetz ret = 0; 2331c5ecae3SMike Kravetz else 2341c5ecae3SMike Kravetz ret = spool->rsv_hpages + delta - spool->min_hpages; 2351c5ecae3SMike Kravetz 2361c5ecae3SMike Kravetz spool->rsv_hpages += delta; 2371c5ecae3SMike Kravetz if (spool->rsv_hpages > spool->min_hpages) 2381c5ecae3SMike Kravetz spool->rsv_hpages = spool->min_hpages; 2391c5ecae3SMike Kravetz } 2401c5ecae3SMike Kravetz 2411c5ecae3SMike Kravetz /* 2421c5ecae3SMike Kravetz * If hugetlbfs_put_super couldn't free spool due to an outstanding 2431c5ecae3SMike Kravetz * quota reference, free it now. 2441c5ecae3SMike Kravetz */ 245db71ef79SMike Kravetz unlock_or_release_subpool(spool, flags); 2461c5ecae3SMike Kravetz 2471c5ecae3SMike Kravetz return ret; 24890481622SDavid Gibson } 24990481622SDavid Gibson 25090481622SDavid Gibson static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 25190481622SDavid Gibson { 25290481622SDavid Gibson return HUGETLBFS_SB(inode->i_sb)->spool; 25390481622SDavid Gibson } 25490481622SDavid Gibson 25590481622SDavid Gibson static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 25690481622SDavid Gibson { 257496ad9aaSAl Viro return subpool_inode(file_inode(vma->vm_file)); 25890481622SDavid Gibson } 25990481622SDavid Gibson 260e700898fSMike Kravetz /* 261e700898fSMike Kravetz * hugetlb vma_lock helper routines 262e700898fSMike Kravetz */ 263e700898fSMike Kravetz void hugetlb_vma_lock_read(struct vm_area_struct *vma) 264e700898fSMike Kravetz { 265e700898fSMike Kravetz if (__vma_shareable_lock(vma)) { 266e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 267e700898fSMike Kravetz 268e700898fSMike Kravetz down_read(&vma_lock->rw_sema); 269e700898fSMike Kravetz } 270e700898fSMike Kravetz } 271e700898fSMike Kravetz 272e700898fSMike Kravetz void hugetlb_vma_unlock_read(struct vm_area_struct *vma) 273e700898fSMike Kravetz { 274e700898fSMike Kravetz if (__vma_shareable_lock(vma)) { 275e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 276e700898fSMike Kravetz 277e700898fSMike Kravetz up_read(&vma_lock->rw_sema); 278e700898fSMike Kravetz } 279e700898fSMike Kravetz } 280e700898fSMike Kravetz 281e700898fSMike Kravetz void hugetlb_vma_lock_write(struct vm_area_struct *vma) 282e700898fSMike Kravetz { 283e700898fSMike Kravetz if (__vma_shareable_lock(vma)) { 284e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 285e700898fSMike Kravetz 286e700898fSMike Kravetz down_write(&vma_lock->rw_sema); 287e700898fSMike Kravetz } 288e700898fSMike Kravetz } 289e700898fSMike Kravetz 290e700898fSMike Kravetz void hugetlb_vma_unlock_write(struct vm_area_struct *vma) 291e700898fSMike Kravetz { 292e700898fSMike Kravetz if (__vma_shareable_lock(vma)) { 293e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 294e700898fSMike Kravetz 295e700898fSMike Kravetz up_write(&vma_lock->rw_sema); 296e700898fSMike Kravetz } 297e700898fSMike Kravetz } 298e700898fSMike Kravetz 299e700898fSMike Kravetz int hugetlb_vma_trylock_write(struct vm_area_struct *vma) 300e700898fSMike Kravetz { 301e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 302e700898fSMike Kravetz 303e700898fSMike Kravetz if (!__vma_shareable_lock(vma)) 304e700898fSMike Kravetz return 1; 305e700898fSMike Kravetz 306e700898fSMike Kravetz return down_write_trylock(&vma_lock->rw_sema); 307e700898fSMike Kravetz } 308e700898fSMike Kravetz 309e700898fSMike Kravetz void hugetlb_vma_assert_locked(struct vm_area_struct *vma) 310e700898fSMike Kravetz { 311e700898fSMike Kravetz if (__vma_shareable_lock(vma)) { 312e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 313e700898fSMike Kravetz 314e700898fSMike Kravetz lockdep_assert_held(&vma_lock->rw_sema); 315e700898fSMike Kravetz } 316e700898fSMike Kravetz } 317e700898fSMike Kravetz 318e700898fSMike Kravetz void hugetlb_vma_lock_release(struct kref *kref) 319e700898fSMike Kravetz { 320e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = container_of(kref, 321e700898fSMike Kravetz struct hugetlb_vma_lock, refs); 322e700898fSMike Kravetz 323e700898fSMike Kravetz kfree(vma_lock); 324e700898fSMike Kravetz } 325e700898fSMike Kravetz 326e700898fSMike Kravetz static void __hugetlb_vma_unlock_write_put(struct hugetlb_vma_lock *vma_lock) 327e700898fSMike Kravetz { 328e700898fSMike Kravetz struct vm_area_struct *vma = vma_lock->vma; 329e700898fSMike Kravetz 330e700898fSMike Kravetz /* 331e700898fSMike Kravetz * vma_lock structure may or not be released as a result of put, 332e700898fSMike Kravetz * it certainly will no longer be attached to vma so clear pointer. 333e700898fSMike Kravetz * Semaphore synchronizes access to vma_lock->vma field. 334e700898fSMike Kravetz */ 335e700898fSMike Kravetz vma_lock->vma = NULL; 336e700898fSMike Kravetz vma->vm_private_data = NULL; 337e700898fSMike Kravetz up_write(&vma_lock->rw_sema); 338e700898fSMike Kravetz kref_put(&vma_lock->refs, hugetlb_vma_lock_release); 339e700898fSMike Kravetz } 340e700898fSMike Kravetz 341e700898fSMike Kravetz static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma) 342e700898fSMike Kravetz { 343e700898fSMike Kravetz if (__vma_shareable_lock(vma)) { 344e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 345e700898fSMike Kravetz 346e700898fSMike Kravetz __hugetlb_vma_unlock_write_put(vma_lock); 347e700898fSMike Kravetz } 348e700898fSMike Kravetz } 349e700898fSMike Kravetz 350e700898fSMike Kravetz static void hugetlb_vma_lock_free(struct vm_area_struct *vma) 351e700898fSMike Kravetz { 352e700898fSMike Kravetz /* 353e700898fSMike Kravetz * Only present in sharable vmas. 354e700898fSMike Kravetz */ 355e700898fSMike Kravetz if (!vma || !__vma_shareable_lock(vma)) 356e700898fSMike Kravetz return; 357e700898fSMike Kravetz 358e700898fSMike Kravetz if (vma->vm_private_data) { 359e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 360e700898fSMike Kravetz 361e700898fSMike Kravetz down_write(&vma_lock->rw_sema); 362e700898fSMike Kravetz __hugetlb_vma_unlock_write_put(vma_lock); 363e700898fSMike Kravetz } 364e700898fSMike Kravetz } 365e700898fSMike Kravetz 366e700898fSMike Kravetz static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma) 367e700898fSMike Kravetz { 368e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock; 369e700898fSMike Kravetz 370e700898fSMike Kravetz /* Only establish in (flags) sharable vmas */ 371e700898fSMike Kravetz if (!vma || !(vma->vm_flags & VM_MAYSHARE)) 372e700898fSMike Kravetz return; 373e700898fSMike Kravetz 374e700898fSMike Kravetz /* Should never get here with non-NULL vm_private_data */ 375e700898fSMike Kravetz if (vma->vm_private_data) 376e700898fSMike Kravetz return; 377e700898fSMike Kravetz 378e700898fSMike Kravetz vma_lock = kmalloc(sizeof(*vma_lock), GFP_KERNEL); 379e700898fSMike Kravetz if (!vma_lock) { 380e700898fSMike Kravetz /* 381e700898fSMike Kravetz * If we can not allocate structure, then vma can not 382e700898fSMike Kravetz * participate in pmd sharing. This is only a possible 383e700898fSMike Kravetz * performance enhancement and memory saving issue. 384e700898fSMike Kravetz * However, the lock is also used to synchronize page 385e700898fSMike Kravetz * faults with truncation. If the lock is not present, 386e700898fSMike Kravetz * unlikely races could leave pages in a file past i_size 387e700898fSMike Kravetz * until the file is removed. Warn in the unlikely case of 388e700898fSMike Kravetz * allocation failure. 389e700898fSMike Kravetz */ 390e700898fSMike Kravetz pr_warn_once("HugeTLB: unable to allocate vma specific lock\n"); 391e700898fSMike Kravetz return; 392e700898fSMike Kravetz } 393e700898fSMike Kravetz 394e700898fSMike Kravetz kref_init(&vma_lock->refs); 395e700898fSMike Kravetz init_rwsem(&vma_lock->rw_sema); 396e700898fSMike Kravetz vma_lock->vma = vma; 397e700898fSMike Kravetz vma->vm_private_data = vma_lock; 398e700898fSMike Kravetz } 399e700898fSMike Kravetz 4000db9d74eSMina Almasry /* Helper that removes a struct file_region from the resv_map cache and returns 4010db9d74eSMina Almasry * it for use. 4020db9d74eSMina Almasry */ 4030db9d74eSMina Almasry static struct file_region * 4040db9d74eSMina Almasry get_file_region_entry_from_cache(struct resv_map *resv, long from, long to) 4050db9d74eSMina Almasry { 4063259914fSXU pengfei struct file_region *nrg; 4070db9d74eSMina Almasry 4080db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count <= 0); 4090db9d74eSMina Almasry 4100db9d74eSMina Almasry resv->region_cache_count--; 4110db9d74eSMina Almasry nrg = list_first_entry(&resv->region_cache, struct file_region, link); 4120db9d74eSMina Almasry list_del(&nrg->link); 4130db9d74eSMina Almasry 4140db9d74eSMina Almasry nrg->from = from; 4150db9d74eSMina Almasry nrg->to = to; 4160db9d74eSMina Almasry 4170db9d74eSMina Almasry return nrg; 4180db9d74eSMina Almasry } 4190db9d74eSMina Almasry 420075a61d0SMina Almasry static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, 421075a61d0SMina Almasry struct file_region *rg) 422075a61d0SMina Almasry { 423075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 424075a61d0SMina Almasry nrg->reservation_counter = rg->reservation_counter; 425075a61d0SMina Almasry nrg->css = rg->css; 426075a61d0SMina Almasry if (rg->css) 427075a61d0SMina Almasry css_get(rg->css); 428075a61d0SMina Almasry #endif 429075a61d0SMina Almasry } 430075a61d0SMina Almasry 431075a61d0SMina Almasry /* Helper that records hugetlb_cgroup uncharge info. */ 432075a61d0SMina Almasry static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, 433075a61d0SMina Almasry struct hstate *h, 434075a61d0SMina Almasry struct resv_map *resv, 435075a61d0SMina Almasry struct file_region *nrg) 436075a61d0SMina Almasry { 437075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 438075a61d0SMina Almasry if (h_cg) { 439075a61d0SMina Almasry nrg->reservation_counter = 440075a61d0SMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 441075a61d0SMina Almasry nrg->css = &h_cg->css; 442d85aecf2SMiaohe Lin /* 443d85aecf2SMiaohe Lin * The caller will hold exactly one h_cg->css reference for the 444d85aecf2SMiaohe Lin * whole contiguous reservation region. But this area might be 445d85aecf2SMiaohe Lin * scattered when there are already some file_regions reside in 446d85aecf2SMiaohe Lin * it. As a result, many file_regions may share only one css 447d85aecf2SMiaohe Lin * reference. In order to ensure that one file_region must hold 448d85aecf2SMiaohe Lin * exactly one h_cg->css reference, we should do css_get for 449d85aecf2SMiaohe Lin * each file_region and leave the reference held by caller 450d85aecf2SMiaohe Lin * untouched. 451d85aecf2SMiaohe Lin */ 452d85aecf2SMiaohe Lin css_get(&h_cg->css); 453075a61d0SMina Almasry if (!resv->pages_per_hpage) 454075a61d0SMina Almasry resv->pages_per_hpage = pages_per_huge_page(h); 455075a61d0SMina Almasry /* pages_per_hpage should be the same for all entries in 456075a61d0SMina Almasry * a resv_map. 457075a61d0SMina Almasry */ 458075a61d0SMina Almasry VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); 459075a61d0SMina Almasry } else { 460075a61d0SMina Almasry nrg->reservation_counter = NULL; 461075a61d0SMina Almasry nrg->css = NULL; 462075a61d0SMina Almasry } 463075a61d0SMina Almasry #endif 464075a61d0SMina Almasry } 465075a61d0SMina Almasry 466d85aecf2SMiaohe Lin static void put_uncharge_info(struct file_region *rg) 467d85aecf2SMiaohe Lin { 468d85aecf2SMiaohe Lin #ifdef CONFIG_CGROUP_HUGETLB 469d85aecf2SMiaohe Lin if (rg->css) 470d85aecf2SMiaohe Lin css_put(rg->css); 471d85aecf2SMiaohe Lin #endif 472d85aecf2SMiaohe Lin } 473d85aecf2SMiaohe Lin 474a9b3f867SMina Almasry static bool has_same_uncharge_info(struct file_region *rg, 475a9b3f867SMina Almasry struct file_region *org) 476a9b3f867SMina Almasry { 477a9b3f867SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 4780739eb43SBaolin Wang return rg->reservation_counter == org->reservation_counter && 479a9b3f867SMina Almasry rg->css == org->css; 480a9b3f867SMina Almasry 481a9b3f867SMina Almasry #else 482a9b3f867SMina Almasry return true; 483a9b3f867SMina Almasry #endif 484a9b3f867SMina Almasry } 485a9b3f867SMina Almasry 486a9b3f867SMina Almasry static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) 487a9b3f867SMina Almasry { 4883259914fSXU pengfei struct file_region *nrg, *prg; 489a9b3f867SMina Almasry 490a9b3f867SMina Almasry prg = list_prev_entry(rg, link); 491a9b3f867SMina Almasry if (&prg->link != &resv->regions && prg->to == rg->from && 492a9b3f867SMina Almasry has_same_uncharge_info(prg, rg)) { 493a9b3f867SMina Almasry prg->to = rg->to; 494a9b3f867SMina Almasry 495a9b3f867SMina Almasry list_del(&rg->link); 496d85aecf2SMiaohe Lin put_uncharge_info(rg); 497a9b3f867SMina Almasry kfree(rg); 498a9b3f867SMina Almasry 4997db5e7b6SWei Yang rg = prg; 500a9b3f867SMina Almasry } 501a9b3f867SMina Almasry 502a9b3f867SMina Almasry nrg = list_next_entry(rg, link); 503a9b3f867SMina Almasry if (&nrg->link != &resv->regions && nrg->from == rg->to && 504a9b3f867SMina Almasry has_same_uncharge_info(nrg, rg)) { 505a9b3f867SMina Almasry nrg->from = rg->from; 506a9b3f867SMina Almasry 507a9b3f867SMina Almasry list_del(&rg->link); 508d85aecf2SMiaohe Lin put_uncharge_info(rg); 509a9b3f867SMina Almasry kfree(rg); 510a9b3f867SMina Almasry } 511a9b3f867SMina Almasry } 512a9b3f867SMina Almasry 5132103cf9cSPeter Xu static inline long 51484448c8eSJakob Koschel hugetlb_resv_map_add(struct resv_map *map, struct list_head *rg, long from, 5152103cf9cSPeter Xu long to, struct hstate *h, struct hugetlb_cgroup *cg, 5162103cf9cSPeter Xu long *regions_needed) 5172103cf9cSPeter Xu { 5182103cf9cSPeter Xu struct file_region *nrg; 5192103cf9cSPeter Xu 5202103cf9cSPeter Xu if (!regions_needed) { 5212103cf9cSPeter Xu nrg = get_file_region_entry_from_cache(map, from, to); 5222103cf9cSPeter Xu record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg); 52384448c8eSJakob Koschel list_add(&nrg->link, rg); 5242103cf9cSPeter Xu coalesce_file_region(map, nrg); 5252103cf9cSPeter Xu } else 5262103cf9cSPeter Xu *regions_needed += 1; 5272103cf9cSPeter Xu 5282103cf9cSPeter Xu return to - from; 5292103cf9cSPeter Xu } 5302103cf9cSPeter Xu 531972a3da3SWei Yang /* 532972a3da3SWei Yang * Must be called with resv->lock held. 533972a3da3SWei Yang * 534972a3da3SWei Yang * Calling this with regions_needed != NULL will count the number of pages 535972a3da3SWei Yang * to be added but will not modify the linked list. And regions_needed will 536972a3da3SWei Yang * indicate the number of file_regions needed in the cache to carry out to add 537972a3da3SWei Yang * the regions for this range. 538d75c6af9SMina Almasry */ 539d75c6af9SMina Almasry static long add_reservation_in_range(struct resv_map *resv, long f, long t, 540075a61d0SMina Almasry struct hugetlb_cgroup *h_cg, 541972a3da3SWei Yang struct hstate *h, long *regions_needed) 542d75c6af9SMina Almasry { 5430db9d74eSMina Almasry long add = 0; 544d75c6af9SMina Almasry struct list_head *head = &resv->regions; 5450db9d74eSMina Almasry long last_accounted_offset = f; 54684448c8eSJakob Koschel struct file_region *iter, *trg = NULL; 54784448c8eSJakob Koschel struct list_head *rg = NULL; 548d75c6af9SMina Almasry 5490db9d74eSMina Almasry if (regions_needed) 5500db9d74eSMina Almasry *regions_needed = 0; 551d75c6af9SMina Almasry 5520db9d74eSMina Almasry /* In this loop, we essentially handle an entry for the range 55384448c8eSJakob Koschel * [last_accounted_offset, iter->from), at every iteration, with some 5540db9d74eSMina Almasry * bounds checking. 5550db9d74eSMina Almasry */ 55684448c8eSJakob Koschel list_for_each_entry_safe(iter, trg, head, link) { 5570db9d74eSMina Almasry /* Skip irrelevant regions that start before our range. */ 55884448c8eSJakob Koschel if (iter->from < f) { 5590db9d74eSMina Almasry /* If this region ends after the last accounted offset, 5600db9d74eSMina Almasry * then we need to update last_accounted_offset. 5610db9d74eSMina Almasry */ 56284448c8eSJakob Koschel if (iter->to > last_accounted_offset) 56384448c8eSJakob Koschel last_accounted_offset = iter->to; 5640db9d74eSMina Almasry continue; 5650db9d74eSMina Almasry } 566d75c6af9SMina Almasry 5670db9d74eSMina Almasry /* When we find a region that starts beyond our range, we've 5680db9d74eSMina Almasry * finished. 5690db9d74eSMina Almasry */ 57084448c8eSJakob Koschel if (iter->from >= t) { 57184448c8eSJakob Koschel rg = iter->link.prev; 572d75c6af9SMina Almasry break; 57384448c8eSJakob Koschel } 574d75c6af9SMina Almasry 57584448c8eSJakob Koschel /* Add an entry for last_accounted_offset -> iter->from, and 5760db9d74eSMina Almasry * update last_accounted_offset. 577d75c6af9SMina Almasry */ 57884448c8eSJakob Koschel if (iter->from > last_accounted_offset) 57984448c8eSJakob Koschel add += hugetlb_resv_map_add(resv, iter->link.prev, 5802103cf9cSPeter Xu last_accounted_offset, 58184448c8eSJakob Koschel iter->from, h, h_cg, 5822103cf9cSPeter Xu regions_needed); 583d75c6af9SMina Almasry 58484448c8eSJakob Koschel last_accounted_offset = iter->to; 5850db9d74eSMina Almasry } 5860db9d74eSMina Almasry 5870db9d74eSMina Almasry /* Handle the case where our range extends beyond 5880db9d74eSMina Almasry * last_accounted_offset. 5890db9d74eSMina Almasry */ 59084448c8eSJakob Koschel if (!rg) 59184448c8eSJakob Koschel rg = head->prev; 5922103cf9cSPeter Xu if (last_accounted_offset < t) 5932103cf9cSPeter Xu add += hugetlb_resv_map_add(resv, rg, last_accounted_offset, 5942103cf9cSPeter Xu t, h, h_cg, regions_needed); 5950db9d74eSMina Almasry 5960db9d74eSMina Almasry return add; 5970db9d74eSMina Almasry } 5980db9d74eSMina Almasry 5990db9d74eSMina Almasry /* Must be called with resv->lock acquired. Will drop lock to allocate entries. 6000db9d74eSMina Almasry */ 6010db9d74eSMina Almasry static int allocate_file_region_entries(struct resv_map *resv, 6020db9d74eSMina Almasry int regions_needed) 6030db9d74eSMina Almasry __must_hold(&resv->lock) 6040db9d74eSMina Almasry { 60534665341SMiaohe Lin LIST_HEAD(allocated_regions); 6060db9d74eSMina Almasry int to_allocate = 0, i = 0; 6070db9d74eSMina Almasry struct file_region *trg = NULL, *rg = NULL; 6080db9d74eSMina Almasry 6090db9d74eSMina Almasry VM_BUG_ON(regions_needed < 0); 6100db9d74eSMina Almasry 6110db9d74eSMina Almasry /* 6120db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 6130db9d74eSMina Almasry * the number of in progress add operations plus regions_needed. 6140db9d74eSMina Almasry * 6150db9d74eSMina Almasry * This is a while loop because when we drop the lock, some other call 6160db9d74eSMina Almasry * to region_add or region_del may have consumed some region_entries, 6170db9d74eSMina Almasry * so we keep looping here until we finally have enough entries for 6180db9d74eSMina Almasry * (adds_in_progress + regions_needed). 6190db9d74eSMina Almasry */ 6200db9d74eSMina Almasry while (resv->region_cache_count < 6210db9d74eSMina Almasry (resv->adds_in_progress + regions_needed)) { 6220db9d74eSMina Almasry to_allocate = resv->adds_in_progress + regions_needed - 6230db9d74eSMina Almasry resv->region_cache_count; 6240db9d74eSMina Almasry 6250db9d74eSMina Almasry /* At this point, we should have enough entries in the cache 626f0953a1bSIngo Molnar * for all the existing adds_in_progress. We should only be 6270db9d74eSMina Almasry * needing to allocate for regions_needed. 6280db9d74eSMina Almasry */ 6290db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); 6300db9d74eSMina Almasry 6310db9d74eSMina Almasry spin_unlock(&resv->lock); 6320db9d74eSMina Almasry for (i = 0; i < to_allocate; i++) { 6330db9d74eSMina Almasry trg = kmalloc(sizeof(*trg), GFP_KERNEL); 6340db9d74eSMina Almasry if (!trg) 6350db9d74eSMina Almasry goto out_of_memory; 6360db9d74eSMina Almasry list_add(&trg->link, &allocated_regions); 6370db9d74eSMina Almasry } 6380db9d74eSMina Almasry 6390db9d74eSMina Almasry spin_lock(&resv->lock); 6400db9d74eSMina Almasry 641d3ec7b6eSWei Yang list_splice(&allocated_regions, &resv->region_cache); 642d3ec7b6eSWei Yang resv->region_cache_count += to_allocate; 6430db9d74eSMina Almasry } 6440db9d74eSMina Almasry 6450db9d74eSMina Almasry return 0; 6460db9d74eSMina Almasry 6470db9d74eSMina Almasry out_of_memory: 6480db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, &allocated_regions, link) { 649d75c6af9SMina Almasry list_del(&rg->link); 650d75c6af9SMina Almasry kfree(rg); 651d75c6af9SMina Almasry } 6520db9d74eSMina Almasry return -ENOMEM; 653d75c6af9SMina Almasry } 654d75c6af9SMina Almasry 6551dd308a7SMike Kravetz /* 6561dd308a7SMike Kravetz * Add the huge page range represented by [f, t) to the reserve 6570db9d74eSMina Almasry * map. Regions will be taken from the cache to fill in this range. 6580db9d74eSMina Almasry * Sufficient regions should exist in the cache due to the previous 6590db9d74eSMina Almasry * call to region_chg with the same range, but in some cases the cache will not 6600db9d74eSMina Almasry * have sufficient entries due to races with other code doing region_add or 6610db9d74eSMina Almasry * region_del. The extra needed entries will be allocated. 662cf3ad20bSMike Kravetz * 6630db9d74eSMina Almasry * regions_needed is the out value provided by a previous call to region_chg. 6640db9d74eSMina Almasry * 6650db9d74eSMina Almasry * Return the number of new huge pages added to the map. This number is greater 6660db9d74eSMina Almasry * than or equal to zero. If file_region entries needed to be allocated for 6677c8de358SEthon Paul * this operation and we were not able to allocate, it returns -ENOMEM. 6680db9d74eSMina Almasry * region_add of regions of length 1 never allocate file_regions and cannot 6690db9d74eSMina Almasry * fail; region_chg will always allocate at least 1 entry and a region_add for 6700db9d74eSMina Almasry * 1 page will only require at most 1 entry. 6711dd308a7SMike Kravetz */ 6720db9d74eSMina Almasry static long region_add(struct resv_map *resv, long f, long t, 673075a61d0SMina Almasry long in_regions_needed, struct hstate *h, 674075a61d0SMina Almasry struct hugetlb_cgroup *h_cg) 67596822904SAndy Whitcroft { 6760db9d74eSMina Almasry long add = 0, actual_regions_needed = 0; 67796822904SAndy Whitcroft 6787b24d861SDavidlohr Bueso spin_lock(&resv->lock); 6790db9d74eSMina Almasry retry: 6800db9d74eSMina Almasry 6810db9d74eSMina Almasry /* Count how many regions are actually needed to execute this add. */ 682972a3da3SWei Yang add_reservation_in_range(resv, f, t, NULL, NULL, 683972a3da3SWei Yang &actual_regions_needed); 68496822904SAndy Whitcroft 6855e911373SMike Kravetz /* 6860db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 6870db9d74eSMina Almasry * this add operation. Note that actual_regions_needed may be greater 6880db9d74eSMina Almasry * than in_regions_needed, as the resv_map may have been modified since 6890db9d74eSMina Almasry * the region_chg call. In this case, we need to make sure that we 6900db9d74eSMina Almasry * allocate extra entries, such that we have enough for all the 6910db9d74eSMina Almasry * existing adds_in_progress, plus the excess needed for this 6920db9d74eSMina Almasry * operation. 6935e911373SMike Kravetz */ 6940db9d74eSMina Almasry if (actual_regions_needed > in_regions_needed && 6950db9d74eSMina Almasry resv->region_cache_count < 6960db9d74eSMina Almasry resv->adds_in_progress + 6970db9d74eSMina Almasry (actual_regions_needed - in_regions_needed)) { 6980db9d74eSMina Almasry /* region_add operation of range 1 should never need to 6990db9d74eSMina Almasry * allocate file_region entries. 7000db9d74eSMina Almasry */ 7010db9d74eSMina Almasry VM_BUG_ON(t - f <= 1); 7025e911373SMike Kravetz 7030db9d74eSMina Almasry if (allocate_file_region_entries( 7040db9d74eSMina Almasry resv, actual_regions_needed - in_regions_needed)) { 7050db9d74eSMina Almasry return -ENOMEM; 7065e911373SMike Kravetz } 7075e911373SMike Kravetz 7080db9d74eSMina Almasry goto retry; 7090db9d74eSMina Almasry } 710cf3ad20bSMike Kravetz 711972a3da3SWei Yang add = add_reservation_in_range(resv, f, t, h_cg, h, NULL); 7120db9d74eSMina Almasry 7130db9d74eSMina Almasry resv->adds_in_progress -= in_regions_needed; 7140db9d74eSMina Almasry 7157b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 716cf3ad20bSMike Kravetz return add; 71796822904SAndy Whitcroft } 71896822904SAndy Whitcroft 7191dd308a7SMike Kravetz /* 7201dd308a7SMike Kravetz * Examine the existing reserve map and determine how many 7211dd308a7SMike Kravetz * huge pages in the specified range [f, t) are NOT currently 7221dd308a7SMike Kravetz * represented. This routine is called before a subsequent 7231dd308a7SMike Kravetz * call to region_add that will actually modify the reserve 7241dd308a7SMike Kravetz * map to add the specified range [f, t). region_chg does 7251dd308a7SMike Kravetz * not change the number of huge pages represented by the 7260db9d74eSMina Almasry * map. A number of new file_region structures is added to the cache as a 7270db9d74eSMina Almasry * placeholder, for the subsequent region_add call to use. At least 1 7280db9d74eSMina Almasry * file_region structure is added. 7290db9d74eSMina Almasry * 7300db9d74eSMina Almasry * out_regions_needed is the number of regions added to the 7310db9d74eSMina Almasry * resv->adds_in_progress. This value needs to be provided to a follow up call 7320db9d74eSMina Almasry * to region_add or region_abort for proper accounting. 7335e911373SMike Kravetz * 7345e911373SMike Kravetz * Returns the number of huge pages that need to be added to the existing 7355e911373SMike Kravetz * reservation map for the range [f, t). This number is greater or equal to 7365e911373SMike Kravetz * zero. -ENOMEM is returned if a new file_region structure or cache entry 7375e911373SMike Kravetz * is needed and can not be allocated. 7381dd308a7SMike Kravetz */ 7390db9d74eSMina Almasry static long region_chg(struct resv_map *resv, long f, long t, 7400db9d74eSMina Almasry long *out_regions_needed) 74196822904SAndy Whitcroft { 74296822904SAndy Whitcroft long chg = 0; 74396822904SAndy Whitcroft 7447b24d861SDavidlohr Bueso spin_lock(&resv->lock); 7455e911373SMike Kravetz 746972a3da3SWei Yang /* Count how many hugepages in this range are NOT represented. */ 747075a61d0SMina Almasry chg = add_reservation_in_range(resv, f, t, NULL, NULL, 748972a3da3SWei Yang out_regions_needed); 7495e911373SMike Kravetz 7500db9d74eSMina Almasry if (*out_regions_needed == 0) 7510db9d74eSMina Almasry *out_regions_needed = 1; 7525e911373SMike Kravetz 7530db9d74eSMina Almasry if (allocate_file_region_entries(resv, *out_regions_needed)) 7545e911373SMike Kravetz return -ENOMEM; 7555e911373SMike Kravetz 7560db9d74eSMina Almasry resv->adds_in_progress += *out_regions_needed; 75796822904SAndy Whitcroft 7587b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 75996822904SAndy Whitcroft return chg; 76096822904SAndy Whitcroft } 76196822904SAndy Whitcroft 7621dd308a7SMike Kravetz /* 7635e911373SMike Kravetz * Abort the in progress add operation. The adds_in_progress field 7645e911373SMike Kravetz * of the resv_map keeps track of the operations in progress between 7655e911373SMike Kravetz * calls to region_chg and region_add. Operations are sometimes 7665e911373SMike Kravetz * aborted after the call to region_chg. In such cases, region_abort 7670db9d74eSMina Almasry * is called to decrement the adds_in_progress counter. regions_needed 7680db9d74eSMina Almasry * is the value returned by the region_chg call, it is used to decrement 7690db9d74eSMina Almasry * the adds_in_progress counter. 7705e911373SMike Kravetz * 7715e911373SMike Kravetz * NOTE: The range arguments [f, t) are not needed or used in this 7725e911373SMike Kravetz * routine. They are kept to make reading the calling code easier as 7735e911373SMike Kravetz * arguments will match the associated region_chg call. 7745e911373SMike Kravetz */ 7750db9d74eSMina Almasry static void region_abort(struct resv_map *resv, long f, long t, 7760db9d74eSMina Almasry long regions_needed) 7775e911373SMike Kravetz { 7785e911373SMike Kravetz spin_lock(&resv->lock); 7795e911373SMike Kravetz VM_BUG_ON(!resv->region_cache_count); 7800db9d74eSMina Almasry resv->adds_in_progress -= regions_needed; 7815e911373SMike Kravetz spin_unlock(&resv->lock); 7825e911373SMike Kravetz } 7835e911373SMike Kravetz 7845e911373SMike Kravetz /* 785feba16e2SMike Kravetz * Delete the specified range [f, t) from the reserve map. If the 786feba16e2SMike Kravetz * t parameter is LONG_MAX, this indicates that ALL regions after f 787feba16e2SMike Kravetz * should be deleted. Locate the regions which intersect [f, t) 788feba16e2SMike Kravetz * and either trim, delete or split the existing regions. 789feba16e2SMike Kravetz * 790feba16e2SMike Kravetz * Returns the number of huge pages deleted from the reserve map. 791feba16e2SMike Kravetz * In the normal case, the return value is zero or more. In the 792feba16e2SMike Kravetz * case where a region must be split, a new region descriptor must 793feba16e2SMike Kravetz * be allocated. If the allocation fails, -ENOMEM will be returned. 794feba16e2SMike Kravetz * NOTE: If the parameter t == LONG_MAX, then we will never split 795feba16e2SMike Kravetz * a region and possibly return -ENOMEM. Callers specifying 796feba16e2SMike Kravetz * t == LONG_MAX do not need to check for -ENOMEM error. 7971dd308a7SMike Kravetz */ 798feba16e2SMike Kravetz static long region_del(struct resv_map *resv, long f, long t) 79996822904SAndy Whitcroft { 8001406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 80196822904SAndy Whitcroft struct file_region *rg, *trg; 802feba16e2SMike Kravetz struct file_region *nrg = NULL; 803feba16e2SMike Kravetz long del = 0; 80496822904SAndy Whitcroft 805feba16e2SMike Kravetz retry: 8067b24d861SDavidlohr Bueso spin_lock(&resv->lock); 807feba16e2SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 808dbe409e4SMike Kravetz /* 809dbe409e4SMike Kravetz * Skip regions before the range to be deleted. file_region 810dbe409e4SMike Kravetz * ranges are normally of the form [from, to). However, there 811dbe409e4SMike Kravetz * may be a "placeholder" entry in the map which is of the form 812dbe409e4SMike Kravetz * (from, to) with from == to. Check for placeholder entries 813dbe409e4SMike Kravetz * at the beginning of the range to be deleted. 814dbe409e4SMike Kravetz */ 815dbe409e4SMike Kravetz if (rg->to <= f && (rg->to != rg->from || rg->to != f)) 816feba16e2SMike Kravetz continue; 817dbe409e4SMike Kravetz 818feba16e2SMike Kravetz if (rg->from >= t) 81996822904SAndy Whitcroft break; 82096822904SAndy Whitcroft 821feba16e2SMike Kravetz if (f > rg->from && t < rg->to) { /* Must split region */ 822feba16e2SMike Kravetz /* 823feba16e2SMike Kravetz * Check for an entry in the cache before dropping 824feba16e2SMike Kravetz * lock and attempting allocation. 825feba16e2SMike Kravetz */ 826feba16e2SMike Kravetz if (!nrg && 827feba16e2SMike Kravetz resv->region_cache_count > resv->adds_in_progress) { 828feba16e2SMike Kravetz nrg = list_first_entry(&resv->region_cache, 829feba16e2SMike Kravetz struct file_region, 830feba16e2SMike Kravetz link); 831feba16e2SMike Kravetz list_del(&nrg->link); 832feba16e2SMike Kravetz resv->region_cache_count--; 83396822904SAndy Whitcroft } 83496822904SAndy Whitcroft 835feba16e2SMike Kravetz if (!nrg) { 836feba16e2SMike Kravetz spin_unlock(&resv->lock); 837feba16e2SMike Kravetz nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 838feba16e2SMike Kravetz if (!nrg) 839feba16e2SMike Kravetz return -ENOMEM; 840feba16e2SMike Kravetz goto retry; 841feba16e2SMike Kravetz } 842feba16e2SMike Kravetz 843feba16e2SMike Kravetz del += t - f; 84479aa925bSMike Kravetz hugetlb_cgroup_uncharge_file_region( 845d85aecf2SMiaohe Lin resv, rg, t - f, false); 846feba16e2SMike Kravetz 847feba16e2SMike Kravetz /* New entry for end of split region */ 848feba16e2SMike Kravetz nrg->from = t; 849feba16e2SMike Kravetz nrg->to = rg->to; 850075a61d0SMina Almasry 851075a61d0SMina Almasry copy_hugetlb_cgroup_uncharge_info(nrg, rg); 852075a61d0SMina Almasry 853feba16e2SMike Kravetz INIT_LIST_HEAD(&nrg->link); 854feba16e2SMike Kravetz 855feba16e2SMike Kravetz /* Original entry is trimmed */ 856feba16e2SMike Kravetz rg->to = f; 857feba16e2SMike Kravetz 858feba16e2SMike Kravetz list_add(&nrg->link, &rg->link); 859feba16e2SMike Kravetz nrg = NULL; 86096822904SAndy Whitcroft break; 861feba16e2SMike Kravetz } 862feba16e2SMike Kravetz 863feba16e2SMike Kravetz if (f <= rg->from && t >= rg->to) { /* Remove entire region */ 864feba16e2SMike Kravetz del += rg->to - rg->from; 865075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 866d85aecf2SMiaohe Lin rg->to - rg->from, true); 86796822904SAndy Whitcroft list_del(&rg->link); 86896822904SAndy Whitcroft kfree(rg); 869feba16e2SMike Kravetz continue; 87096822904SAndy Whitcroft } 8717b24d861SDavidlohr Bueso 872feba16e2SMike Kravetz if (f <= rg->from) { /* Trim beginning of region */ 873075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 874d85aecf2SMiaohe Lin t - rg->from, false); 875075a61d0SMina Almasry 87679aa925bSMike Kravetz del += t - rg->from; 87779aa925bSMike Kravetz rg->from = t; 87879aa925bSMike Kravetz } else { /* Trim end of region */ 879075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 880d85aecf2SMiaohe Lin rg->to - f, false); 88179aa925bSMike Kravetz 88279aa925bSMike Kravetz del += rg->to - f; 88379aa925bSMike Kravetz rg->to = f; 884feba16e2SMike Kravetz } 885feba16e2SMike Kravetz } 886feba16e2SMike Kravetz 8877b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 888feba16e2SMike Kravetz kfree(nrg); 889feba16e2SMike Kravetz return del; 89096822904SAndy Whitcroft } 89196822904SAndy Whitcroft 8921dd308a7SMike Kravetz /* 893b5cec28dSMike Kravetz * A rare out of memory error was encountered which prevented removal of 894b5cec28dSMike Kravetz * the reserve map region for a page. The huge page itself was free'ed 895b5cec28dSMike Kravetz * and removed from the page cache. This routine will adjust the subpool 896b5cec28dSMike Kravetz * usage count, and the global reserve count if needed. By incrementing 897b5cec28dSMike Kravetz * these counts, the reserve map entry which could not be deleted will 898b5cec28dSMike Kravetz * appear as a "reserved" entry instead of simply dangling with incorrect 899b5cec28dSMike Kravetz * counts. 900b5cec28dSMike Kravetz */ 90172e2936cSzhong jiang void hugetlb_fix_reserve_counts(struct inode *inode) 902b5cec28dSMike Kravetz { 903b5cec28dSMike Kravetz struct hugepage_subpool *spool = subpool_inode(inode); 904b5cec28dSMike Kravetz long rsv_adjust; 905da56388cSMiaohe Lin bool reserved = false; 906b5cec28dSMike Kravetz 907b5cec28dSMike Kravetz rsv_adjust = hugepage_subpool_get_pages(spool, 1); 908da56388cSMiaohe Lin if (rsv_adjust > 0) { 909b5cec28dSMike Kravetz struct hstate *h = hstate_inode(inode); 910b5cec28dSMike Kravetz 911da56388cSMiaohe Lin if (!hugetlb_acct_memory(h, 1)) 912da56388cSMiaohe Lin reserved = true; 913da56388cSMiaohe Lin } else if (!rsv_adjust) { 914da56388cSMiaohe Lin reserved = true; 915b5cec28dSMike Kravetz } 916da56388cSMiaohe Lin 917da56388cSMiaohe Lin if (!reserved) 918da56388cSMiaohe Lin pr_warn("hugetlb: Huge Page Reserved count may go negative.\n"); 919b5cec28dSMike Kravetz } 920b5cec28dSMike Kravetz 921b5cec28dSMike Kravetz /* 9221dd308a7SMike Kravetz * Count and return the number of huge pages in the reserve map 9231dd308a7SMike Kravetz * that intersect with the range [f, t). 9241dd308a7SMike Kravetz */ 9251406ec9bSJoonsoo Kim static long region_count(struct resv_map *resv, long f, long t) 92684afd99bSAndy Whitcroft { 9271406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 92884afd99bSAndy Whitcroft struct file_region *rg; 92984afd99bSAndy Whitcroft long chg = 0; 93084afd99bSAndy Whitcroft 9317b24d861SDavidlohr Bueso spin_lock(&resv->lock); 93284afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 93384afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 934f2135a4aSWang Sheng-Hui long seg_from; 935f2135a4aSWang Sheng-Hui long seg_to; 93684afd99bSAndy Whitcroft 93784afd99bSAndy Whitcroft if (rg->to <= f) 93884afd99bSAndy Whitcroft continue; 93984afd99bSAndy Whitcroft if (rg->from >= t) 94084afd99bSAndy Whitcroft break; 94184afd99bSAndy Whitcroft 94284afd99bSAndy Whitcroft seg_from = max(rg->from, f); 94384afd99bSAndy Whitcroft seg_to = min(rg->to, t); 94484afd99bSAndy Whitcroft 94584afd99bSAndy Whitcroft chg += seg_to - seg_from; 94684afd99bSAndy Whitcroft } 9477b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 94884afd99bSAndy Whitcroft 94984afd99bSAndy Whitcroft return chg; 95084afd99bSAndy Whitcroft } 95184afd99bSAndy Whitcroft 95296822904SAndy Whitcroft /* 953e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 954e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 955e7c4b0bfSAndy Whitcroft */ 956a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 957a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 958e7c4b0bfSAndy Whitcroft { 959a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 960a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 961e7c4b0bfSAndy Whitcroft } 962e7c4b0bfSAndy Whitcroft 9630fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 9640fe6e20bSNaoya Horiguchi unsigned long address) 9650fe6e20bSNaoya Horiguchi { 9660fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 9670fe6e20bSNaoya Horiguchi } 968dee41079SDan Williams EXPORT_SYMBOL_GPL(linear_hugepage_index); 9690fe6e20bSNaoya Horiguchi 97084afd99bSAndy Whitcroft /* 97108fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 97208fba699SMel Gorman * cases this will be same size as used by the page table entries. 97308fba699SMel Gorman */ 97408fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 97508fba699SMel Gorman { 97605ea8860SDan Williams if (vma->vm_ops && vma->vm_ops->pagesize) 97705ea8860SDan Williams return vma->vm_ops->pagesize(vma); 97808fba699SMel Gorman return PAGE_SIZE; 97908fba699SMel Gorman } 980f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 98108fba699SMel Gorman 98208fba699SMel Gorman /* 9833340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 9843340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 98509135cc5SDan Williams * architectures where it differs, an architecture-specific 'strong' 98609135cc5SDan Williams * version of this symbol is required. 9873340289dSMel Gorman */ 98809135cc5SDan Williams __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 9893340289dSMel Gorman { 9903340289dSMel Gorman return vma_kernel_pagesize(vma); 9913340289dSMel Gorman } 9923340289dSMel Gorman 9933340289dSMel Gorman /* 99484afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 99584afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 99684afd99bSAndy Whitcroft * alignment. 99784afd99bSAndy Whitcroft */ 99884afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 99984afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 100004f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 100184afd99bSAndy Whitcroft 1002a1e78772SMel Gorman /* 1003a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 1004a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 1005a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 1006a1e78772SMel Gorman * 10078d9bfb26SMike Kravetz * With the exception of hugetlb_dup_vma_private() which is called at fork(), 1008a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 1009a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 1010a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 101184afd99bSAndy Whitcroft * 101284afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 101384afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 101484afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 101584afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 101684afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 101784afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 101884afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 101984afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 1020a1e78772SMel Gorman */ 1021e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 1022e7c4b0bfSAndy Whitcroft { 1023e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 1024e7c4b0bfSAndy Whitcroft } 1025e7c4b0bfSAndy Whitcroft 1026e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 1027e7c4b0bfSAndy Whitcroft unsigned long value) 1028e7c4b0bfSAndy Whitcroft { 1029e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 1030e7c4b0bfSAndy Whitcroft } 1031e7c4b0bfSAndy Whitcroft 1032e9fe92aeSMina Almasry static void 1033e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, 1034e9fe92aeSMina Almasry struct hugetlb_cgroup *h_cg, 1035e9fe92aeSMina Almasry struct hstate *h) 1036e9fe92aeSMina Almasry { 1037e9fe92aeSMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 1038e9fe92aeSMina Almasry if (!h_cg || !h) { 1039e9fe92aeSMina Almasry resv_map->reservation_counter = NULL; 1040e9fe92aeSMina Almasry resv_map->pages_per_hpage = 0; 1041e9fe92aeSMina Almasry resv_map->css = NULL; 1042e9fe92aeSMina Almasry } else { 1043e9fe92aeSMina Almasry resv_map->reservation_counter = 1044e9fe92aeSMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 1045e9fe92aeSMina Almasry resv_map->pages_per_hpage = pages_per_huge_page(h); 1046e9fe92aeSMina Almasry resv_map->css = &h_cg->css; 1047e9fe92aeSMina Almasry } 1048e9fe92aeSMina Almasry #endif 1049e9fe92aeSMina Almasry } 1050e9fe92aeSMina Almasry 10519119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 105284afd99bSAndy Whitcroft { 105384afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 10545e911373SMike Kravetz struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 10555e911373SMike Kravetz 10565e911373SMike Kravetz if (!resv_map || !rg) { 10575e911373SMike Kravetz kfree(resv_map); 10585e911373SMike Kravetz kfree(rg); 105984afd99bSAndy Whitcroft return NULL; 10605e911373SMike Kravetz } 106184afd99bSAndy Whitcroft 106284afd99bSAndy Whitcroft kref_init(&resv_map->refs); 10637b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 106484afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 106584afd99bSAndy Whitcroft 10665e911373SMike Kravetz resv_map->adds_in_progress = 0; 1067e9fe92aeSMina Almasry /* 1068e9fe92aeSMina Almasry * Initialize these to 0. On shared mappings, 0's here indicate these 1069e9fe92aeSMina Almasry * fields don't do cgroup accounting. On private mappings, these will be 1070e9fe92aeSMina Almasry * re-initialized to the proper values, to indicate that hugetlb cgroup 1071e9fe92aeSMina Almasry * reservations are to be un-charged from here. 1072e9fe92aeSMina Almasry */ 1073e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); 10745e911373SMike Kravetz 10755e911373SMike Kravetz INIT_LIST_HEAD(&resv_map->region_cache); 10765e911373SMike Kravetz list_add(&rg->link, &resv_map->region_cache); 10775e911373SMike Kravetz resv_map->region_cache_count = 1; 10785e911373SMike Kravetz 107984afd99bSAndy Whitcroft return resv_map; 108084afd99bSAndy Whitcroft } 108184afd99bSAndy Whitcroft 10829119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 108384afd99bSAndy Whitcroft { 108484afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 10855e911373SMike Kravetz struct list_head *head = &resv_map->region_cache; 10865e911373SMike Kravetz struct file_region *rg, *trg; 108784afd99bSAndy Whitcroft 108884afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 1089feba16e2SMike Kravetz region_del(resv_map, 0, LONG_MAX); 10905e911373SMike Kravetz 10915e911373SMike Kravetz /* ... and any entries left in the cache */ 10925e911373SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 10935e911373SMike Kravetz list_del(&rg->link); 10945e911373SMike Kravetz kfree(rg); 10955e911373SMike Kravetz } 10965e911373SMike Kravetz 10975e911373SMike Kravetz VM_BUG_ON(resv_map->adds_in_progress); 10985e911373SMike Kravetz 109984afd99bSAndy Whitcroft kfree(resv_map); 110084afd99bSAndy Whitcroft } 110184afd99bSAndy Whitcroft 11024e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 11034e35f483SJoonsoo Kim { 1104f27a5136SMike Kravetz /* 1105f27a5136SMike Kravetz * At inode evict time, i_mapping may not point to the original 1106f27a5136SMike Kravetz * address space within the inode. This original address space 1107f27a5136SMike Kravetz * contains the pointer to the resv_map. So, always use the 1108f27a5136SMike Kravetz * address space embedded within the inode. 1109f27a5136SMike Kravetz * The VERY common case is inode->mapping == &inode->i_data but, 1110f27a5136SMike Kravetz * this may not be true for device special inodes. 1111f27a5136SMike Kravetz */ 1112f27a5136SMike Kravetz return (struct resv_map *)(&inode->i_data)->private_data; 11134e35f483SJoonsoo Kim } 11144e35f483SJoonsoo Kim 111584afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 1116a1e78772SMel Gorman { 111781d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 11184e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 11194e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 11204e35f483SJoonsoo Kim struct inode *inode = mapping->host; 11214e35f483SJoonsoo Kim 11224e35f483SJoonsoo Kim return inode_resv_map(inode); 11234e35f483SJoonsoo Kim 11244e35f483SJoonsoo Kim } else { 112584afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 112684afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 11274e35f483SJoonsoo Kim } 1128a1e78772SMel Gorman } 1129a1e78772SMel Gorman 113084afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 1131a1e78772SMel Gorman { 113281d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 113381d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 1134a1e78772SMel Gorman 113584afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 113684afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 113704f2cbe3SMel Gorman } 113804f2cbe3SMel Gorman 113904f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 114004f2cbe3SMel Gorman { 114181d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 114281d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 1143e7c4b0bfSAndy Whitcroft 1144e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 114504f2cbe3SMel Gorman } 114604f2cbe3SMel Gorman 114704f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 114804f2cbe3SMel Gorman { 114981d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 1150e7c4b0bfSAndy Whitcroft 1151e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 1152a1e78772SMel Gorman } 1153a1e78772SMel Gorman 11548d9bfb26SMike Kravetz void hugetlb_dup_vma_private(struct vm_area_struct *vma) 1155a1e78772SMel Gorman { 115681d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 11578d9bfb26SMike Kravetz /* 11588d9bfb26SMike Kravetz * Clear vm_private_data 1159612b8a31SMike Kravetz * - For shared mappings this is a per-vma semaphore that may be 1160612b8a31SMike Kravetz * allocated in a subsequent call to hugetlb_vm_op_open. 1161612b8a31SMike Kravetz * Before clearing, make sure pointer is not associated with vma 1162612b8a31SMike Kravetz * as this will leak the structure. This is the case when called 1163612b8a31SMike Kravetz * via clear_vma_resv_huge_pages() and hugetlb_vm_op_open has already 1164612b8a31SMike Kravetz * been called to allocate a new structure. 11658d9bfb26SMike Kravetz * - For MAP_PRIVATE mappings, this is the reserve map which does 11668d9bfb26SMike Kravetz * not apply to children. Faults generated by the children are 11678d9bfb26SMike Kravetz * not guaranteed to succeed, even if read-only. 11688d9bfb26SMike Kravetz */ 1169612b8a31SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 1170612b8a31SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 1171612b8a31SMike Kravetz 1172612b8a31SMike Kravetz if (vma_lock && vma_lock->vma != vma) 1173612b8a31SMike Kravetz vma->vm_private_data = NULL; 1174612b8a31SMike Kravetz } else 1175612b8a31SMike Kravetz vma->vm_private_data = NULL; 1176a1e78772SMel Gorman } 1177a1e78772SMel Gorman 1178550a7d60SMina Almasry /* 1179550a7d60SMina Almasry * Reset and decrement one ref on hugepage private reservation. 11808651a137SLorenzo Stoakes * Called with mm->mmap_lock writer semaphore held. 1181550a7d60SMina Almasry * This function should be only used by move_vma() and operate on 1182550a7d60SMina Almasry * same sized vma. It should never come here with last ref on the 1183550a7d60SMina Almasry * reservation. 1184550a7d60SMina Almasry */ 1185550a7d60SMina Almasry void clear_vma_resv_huge_pages(struct vm_area_struct *vma) 1186550a7d60SMina Almasry { 1187550a7d60SMina Almasry /* 1188550a7d60SMina Almasry * Clear the old hugetlb private page reservation. 1189550a7d60SMina Almasry * It has already been transferred to new_vma. 1190550a7d60SMina Almasry * 1191550a7d60SMina Almasry * During a mremap() operation of a hugetlb vma we call move_vma() 1192550a7d60SMina Almasry * which copies vma into new_vma and unmaps vma. After the copy 1193550a7d60SMina Almasry * operation both new_vma and vma share a reference to the resv_map 1194550a7d60SMina Almasry * struct, and at that point vma is about to be unmapped. We don't 1195550a7d60SMina Almasry * want to return the reservation to the pool at unmap of vma because 1196550a7d60SMina Almasry * the reservation still lives on in new_vma, so simply decrement the 1197550a7d60SMina Almasry * ref here and remove the resv_map reference from this vma. 1198550a7d60SMina Almasry */ 1199550a7d60SMina Almasry struct resv_map *reservations = vma_resv_map(vma); 1200550a7d60SMina Almasry 1201afe041c2SBui Quang Minh if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 1202afe041c2SBui Quang Minh resv_map_put_hugetlb_cgroup_uncharge_info(reservations); 1203550a7d60SMina Almasry kref_put(&reservations->refs, resv_map_release); 1204afe041c2SBui Quang Minh } 1205550a7d60SMina Almasry 12068d9bfb26SMike Kravetz hugetlb_dup_vma_private(vma); 1207550a7d60SMina Almasry } 1208550a7d60SMina Almasry 1209a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 1210559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 1211a1e78772SMel Gorman { 1212af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 1213af0ed73eSJoonsoo Kim /* 1214af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 1215af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 1216af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 1217af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 1218af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 1219af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 1220af0ed73eSJoonsoo Kim * properly, so add work-around here. 1221af0ed73eSJoonsoo Kim */ 1222af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 1223559ec2f8SNicholas Krause return true; 1224af0ed73eSJoonsoo Kim else 1225559ec2f8SNicholas Krause return false; 1226af0ed73eSJoonsoo Kim } 1227a63884e9SJoonsoo Kim 1228a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 12291fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 12301fb1b0e9SMike Kravetz /* 12311fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 12321fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 12331fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 12347c8de358SEthon Paul * fallocate. In this case, there really are no reserves to 12351fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 12361fb1b0e9SMike Kravetz */ 12371fb1b0e9SMike Kravetz if (chg) 12381fb1b0e9SMike Kravetz return false; 12391fb1b0e9SMike Kravetz else 1240559ec2f8SNicholas Krause return true; 12411fb1b0e9SMike Kravetz } 1242a63884e9SJoonsoo Kim 1243a63884e9SJoonsoo Kim /* 1244a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 1245a63884e9SJoonsoo Kim * private mappings. 1246a63884e9SJoonsoo Kim */ 124767961f9dSMike Kravetz if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 124867961f9dSMike Kravetz /* 124967961f9dSMike Kravetz * Like the shared case above, a hole punch or truncate 125067961f9dSMike Kravetz * could have been performed on the private mapping. 125167961f9dSMike Kravetz * Examine the value of chg to determine if reserves 125267961f9dSMike Kravetz * actually exist or were previously consumed. 125367961f9dSMike Kravetz * Very Subtle - The value of chg comes from a previous 125467961f9dSMike Kravetz * call to vma_needs_reserves(). The reserve map for 125567961f9dSMike Kravetz * private mappings has different (opposite) semantics 125667961f9dSMike Kravetz * than that of shared mappings. vma_needs_reserves() 125767961f9dSMike Kravetz * has already taken this difference in semantics into 125867961f9dSMike Kravetz * account. Therefore, the meaning of chg is the same 125967961f9dSMike Kravetz * as in the shared case above. Code could easily be 126067961f9dSMike Kravetz * combined, but keeping it separate draws attention to 126167961f9dSMike Kravetz * subtle differences. 126267961f9dSMike Kravetz */ 126367961f9dSMike Kravetz if (chg) 126467961f9dSMike Kravetz return false; 126567961f9dSMike Kravetz else 1266559ec2f8SNicholas Krause return true; 126767961f9dSMike Kravetz } 1268a63884e9SJoonsoo Kim 1269559ec2f8SNicholas Krause return false; 1270a1e78772SMel Gorman } 1271a1e78772SMel Gorman 1272240d67a8SSidhartha Kumar static void enqueue_hugetlb_folio(struct hstate *h, struct folio *folio) 12731da177e4SLinus Torvalds { 1274240d67a8SSidhartha Kumar int nid = folio_nid(folio); 12759487ca60SMike Kravetz 12769487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1277240d67a8SSidhartha Kumar VM_BUG_ON_FOLIO(folio_ref_count(folio), folio); 1278b65a4edaSMike Kravetz 1279240d67a8SSidhartha Kumar list_move(&folio->lru, &h->hugepage_freelists[nid]); 1280a5516438SAndi Kleen h->free_huge_pages++; 1281a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 1282240d67a8SSidhartha Kumar folio_set_hugetlb_freed(folio); 12831da177e4SLinus Torvalds } 12841da177e4SLinus Torvalds 1285a36f1e90SSidhartha Kumar static struct folio *dequeue_hugetlb_folio_node_exact(struct hstate *h, 1286a36f1e90SSidhartha Kumar int nid) 1287bf50bab2SNaoya Horiguchi { 1288a36f1e90SSidhartha Kumar struct folio *folio; 12891a08ae36SPavel Tatashin bool pin = !!(current->flags & PF_MEMALLOC_PIN); 1290bf50bab2SNaoya Horiguchi 12919487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1292a36f1e90SSidhartha Kumar list_for_each_entry(folio, &h->hugepage_freelists[nid], lru) { 1293a36f1e90SSidhartha Kumar if (pin && !folio_is_longterm_pinnable(folio)) 1294bbe88753SJoonsoo Kim continue; 1295bbe88753SJoonsoo Kim 1296a36f1e90SSidhartha Kumar if (folio_test_hwpoison(folio)) 12976664bfc8SWei Yang continue; 1298bbe88753SJoonsoo Kim 1299a36f1e90SSidhartha Kumar list_move(&folio->lru, &h->hugepage_activelist); 1300a36f1e90SSidhartha Kumar folio_ref_unfreeze(folio, 1); 1301a36f1e90SSidhartha Kumar folio_clear_hugetlb_freed(folio); 1302bf50bab2SNaoya Horiguchi h->free_huge_pages--; 1303bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 1304a36f1e90SSidhartha Kumar return folio; 1305bf50bab2SNaoya Horiguchi } 1306bf50bab2SNaoya Horiguchi 13076664bfc8SWei Yang return NULL; 13086664bfc8SWei Yang } 13096664bfc8SWei Yang 1310a36f1e90SSidhartha Kumar static struct folio *dequeue_hugetlb_folio_nodemask(struct hstate *h, gfp_t gfp_mask, 1311a36f1e90SSidhartha Kumar int nid, nodemask_t *nmask) 131294310cbcSAnshuman Khandual { 13133e59fcb0SMichal Hocko unsigned int cpuset_mems_cookie; 13143e59fcb0SMichal Hocko struct zonelist *zonelist; 13153e59fcb0SMichal Hocko struct zone *zone; 13163e59fcb0SMichal Hocko struct zoneref *z; 131798fa15f3SAnshuman Khandual int node = NUMA_NO_NODE; 13183e59fcb0SMichal Hocko 13193e59fcb0SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 13203e59fcb0SMichal Hocko 13213e59fcb0SMichal Hocko retry_cpuset: 13223e59fcb0SMichal Hocko cpuset_mems_cookie = read_mems_allowed_begin(); 13233e59fcb0SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 1324a36f1e90SSidhartha Kumar struct folio *folio; 132594310cbcSAnshuman Khandual 13263e59fcb0SMichal Hocko if (!cpuset_zone_allowed(zone, gfp_mask)) 13273e59fcb0SMichal Hocko continue; 13283e59fcb0SMichal Hocko /* 13293e59fcb0SMichal Hocko * no need to ask again on the same node. Pool is node rather than 13303e59fcb0SMichal Hocko * zone aware 13313e59fcb0SMichal Hocko */ 13323e59fcb0SMichal Hocko if (zone_to_nid(zone) == node) 13333e59fcb0SMichal Hocko continue; 13343e59fcb0SMichal Hocko node = zone_to_nid(zone); 133594310cbcSAnshuman Khandual 1336a36f1e90SSidhartha Kumar folio = dequeue_hugetlb_folio_node_exact(h, node); 1337a36f1e90SSidhartha Kumar if (folio) 1338a36f1e90SSidhartha Kumar return folio; 133994310cbcSAnshuman Khandual } 13403e59fcb0SMichal Hocko if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 13413e59fcb0SMichal Hocko goto retry_cpuset; 13423e59fcb0SMichal Hocko 134394310cbcSAnshuman Khandual return NULL; 134494310cbcSAnshuman Khandual } 134594310cbcSAnshuman Khandual 13468346d69dSXin Hao static unsigned long available_huge_pages(struct hstate *h) 13478346d69dSXin Hao { 13488346d69dSXin Hao return h->free_huge_pages - h->resv_huge_pages; 13498346d69dSXin Hao } 13508346d69dSXin Hao 1351ff7d853bSSidhartha Kumar static struct folio *dequeue_hugetlb_folio_vma(struct hstate *h, 1352a5516438SAndi Kleen struct vm_area_struct *vma, 1353af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 1354af0ed73eSJoonsoo Kim long chg) 13551da177e4SLinus Torvalds { 1356a36f1e90SSidhartha Kumar struct folio *folio = NULL; 1357480eccf9SLee Schermerhorn struct mempolicy *mpol; 135804ec6264SVlastimil Babka gfp_t gfp_mask; 13593e59fcb0SMichal Hocko nodemask_t *nodemask; 136004ec6264SVlastimil Babka int nid; 13611da177e4SLinus Torvalds 1362a1e78772SMel Gorman /* 1363a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 1364a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 1365a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 1366a1e78772SMel Gorman */ 13678346d69dSXin Hao if (!vma_has_reserves(vma, chg) && !available_huge_pages(h)) 1368c0ff7453SMiao Xie goto err; 1369a1e78772SMel Gorman 137004f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 13718346d69dSXin Hao if (avoid_reserve && !available_huge_pages(h)) 13726eab04a8SJustin P. Mattock goto err; 137304f2cbe3SMel Gorman 137404ec6264SVlastimil Babka gfp_mask = htlb_alloc_mask(h); 137504ec6264SVlastimil Babka nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 1376cfcaa66fSBen Widawsky 1377cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 1378a36f1e90SSidhartha Kumar folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask, 1379a36f1e90SSidhartha Kumar nid, nodemask); 1380cfcaa66fSBen Widawsky 1381cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 1382cfcaa66fSBen Widawsky nodemask = NULL; 1383cfcaa66fSBen Widawsky } 1384cfcaa66fSBen Widawsky 1385a36f1e90SSidhartha Kumar if (!folio) 1386a36f1e90SSidhartha Kumar folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask, 1387a36f1e90SSidhartha Kumar nid, nodemask); 1388cfcaa66fSBen Widawsky 1389a36f1e90SSidhartha Kumar if (folio && !avoid_reserve && vma_has_reserves(vma, chg)) { 1390a36f1e90SSidhartha Kumar folio_set_hugetlb_restore_reserve(folio); 1391a63884e9SJoonsoo Kim h->resv_huge_pages--; 1392bf50bab2SNaoya Horiguchi } 1393cc9a6c87SMel Gorman 1394cc9a6c87SMel Gorman mpol_cond_put(mpol); 1395ff7d853bSSidhartha Kumar return folio; 1396cc9a6c87SMel Gorman 1397c0ff7453SMiao Xie err: 1398cc9a6c87SMel Gorman return NULL; 13991da177e4SLinus Torvalds } 14001da177e4SLinus Torvalds 14011cac6f2cSLuiz Capitulino /* 14021cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 14031cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 14041cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 14051cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 14061cac6f2cSLuiz Capitulino * node for alloc or free. 14071cac6f2cSLuiz Capitulino */ 14081cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 14091cac6f2cSLuiz Capitulino { 14100edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 14111cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 14121cac6f2cSLuiz Capitulino 14131cac6f2cSLuiz Capitulino return nid; 14141cac6f2cSLuiz Capitulino } 14151cac6f2cSLuiz Capitulino 14161cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 14171cac6f2cSLuiz Capitulino { 14181cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 14191cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 14201cac6f2cSLuiz Capitulino return nid; 14211cac6f2cSLuiz Capitulino } 14221cac6f2cSLuiz Capitulino 14231cac6f2cSLuiz Capitulino /* 14241cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 14251cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 14261cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 14271cac6f2cSLuiz Capitulino * mask. 14281cac6f2cSLuiz Capitulino */ 14291cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 14301cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 14311cac6f2cSLuiz Capitulino { 14321cac6f2cSLuiz Capitulino int nid; 14331cac6f2cSLuiz Capitulino 14341cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 14351cac6f2cSLuiz Capitulino 14361cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 14371cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 14381cac6f2cSLuiz Capitulino 14391cac6f2cSLuiz Capitulino return nid; 14401cac6f2cSLuiz Capitulino } 14411cac6f2cSLuiz Capitulino 14421cac6f2cSLuiz Capitulino /* 144310c6ec49SMike Kravetz * helper for remove_pool_huge_page() - return the previously saved 14441cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 14451cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 14461cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 14471cac6f2cSLuiz Capitulino */ 14481cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 14491cac6f2cSLuiz Capitulino { 14501cac6f2cSLuiz Capitulino int nid; 14511cac6f2cSLuiz Capitulino 14521cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 14531cac6f2cSLuiz Capitulino 14541cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 14551cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 14561cac6f2cSLuiz Capitulino 14571cac6f2cSLuiz Capitulino return nid; 14581cac6f2cSLuiz Capitulino } 14591cac6f2cSLuiz Capitulino 14601cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 14611cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 14621cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 14631cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 14641cac6f2cSLuiz Capitulino nr_nodes--) 14651cac6f2cSLuiz Capitulino 14661cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 14671cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 14681cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 14691cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 14701cac6f2cSLuiz Capitulino nr_nodes--) 14711cac6f2cSLuiz Capitulino 14728531fc6fSMike Kravetz /* used to demote non-gigantic_huge pages as well */ 1473911565b8SSidhartha Kumar static void __destroy_compound_gigantic_folio(struct folio *folio, 147434d9e35bSMike Kravetz unsigned int order, bool demote) 1475944d9fecSLuiz Capitulino { 1476944d9fecSLuiz Capitulino int i; 1477944d9fecSLuiz Capitulino int nr_pages = 1 << order; 147814455eabSCheng Li struct page *p; 1479944d9fecSLuiz Capitulino 148046f27228SMatthew Wilcox (Oracle) atomic_set(&folio->_entire_mapcount, 0); 1481eec20426SMatthew Wilcox (Oracle) atomic_set(&folio->_nr_pages_mapped, 0); 148294688e8eSMatthew Wilcox (Oracle) atomic_set(&folio->_pincount, 0); 148347e29d32SJohn Hubbard 148414455eabSCheng Li for (i = 1; i < nr_pages; i++) { 1485911565b8SSidhartha Kumar p = folio_page(folio, i); 1486a01f4390SMike Kravetz p->mapping = NULL; 14871d798ca3SKirill A. Shutemov clear_compound_head(p); 148834d9e35bSMike Kravetz if (!demote) 1489944d9fecSLuiz Capitulino set_page_refcounted(p); 1490944d9fecSLuiz Capitulino } 1491944d9fecSLuiz Capitulino 149204a42e72SSidhartha Kumar folio_set_order(folio, 0); 1493911565b8SSidhartha Kumar __folio_clear_head(folio); 1494944d9fecSLuiz Capitulino } 1495944d9fecSLuiz Capitulino 1496911565b8SSidhartha Kumar static void destroy_compound_hugetlb_folio_for_demote(struct folio *folio, 14978531fc6fSMike Kravetz unsigned int order) 14988531fc6fSMike Kravetz { 1499911565b8SSidhartha Kumar __destroy_compound_gigantic_folio(folio, order, true); 15008531fc6fSMike Kravetz } 15018531fc6fSMike Kravetz 15028531fc6fSMike Kravetz #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 1503911565b8SSidhartha Kumar static void destroy_compound_gigantic_folio(struct folio *folio, 150434d9e35bSMike Kravetz unsigned int order) 150534d9e35bSMike Kravetz { 1506911565b8SSidhartha Kumar __destroy_compound_gigantic_folio(folio, order, false); 150734d9e35bSMike Kravetz } 150834d9e35bSMike Kravetz 15097f325a8dSSidhartha Kumar static void free_gigantic_folio(struct folio *folio, unsigned int order) 1510944d9fecSLuiz Capitulino { 1511cf11e85fSRoman Gushchin /* 1512cf11e85fSRoman Gushchin * If the page isn't allocated using the cma allocator, 1513cf11e85fSRoman Gushchin * cma_release() returns false. 1514cf11e85fSRoman Gushchin */ 1515dbda8feaSBarry Song #ifdef CONFIG_CMA 15167f325a8dSSidhartha Kumar int nid = folio_nid(folio); 15177f325a8dSSidhartha Kumar 15187f325a8dSSidhartha Kumar if (cma_release(hugetlb_cma[nid], &folio->page, 1 << order)) 1519cf11e85fSRoman Gushchin return; 1520dbda8feaSBarry Song #endif 1521cf11e85fSRoman Gushchin 15227f325a8dSSidhartha Kumar free_contig_range(folio_pfn(folio), 1 << order); 1523944d9fecSLuiz Capitulino } 1524944d9fecSLuiz Capitulino 15254eb0716eSAlexandre Ghiti #ifdef CONFIG_CONTIG_ALLOC 152619fc1a7eSSidhartha Kumar static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask, 1527d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) 1528944d9fecSLuiz Capitulino { 152919fc1a7eSSidhartha Kumar struct page *page; 153004adbc3fSMiaohe Lin unsigned long nr_pages = pages_per_huge_page(h); 1531953f064aSLi Xinhai if (nid == NUMA_NO_NODE) 1532953f064aSLi Xinhai nid = numa_mem_id(); 1533944d9fecSLuiz Capitulino 1534dbda8feaSBarry Song #ifdef CONFIG_CMA 1535dbda8feaSBarry Song { 1536cf11e85fSRoman Gushchin int node; 1537cf11e85fSRoman Gushchin 1538953f064aSLi Xinhai if (hugetlb_cma[nid]) { 1539953f064aSLi Xinhai page = cma_alloc(hugetlb_cma[nid], nr_pages, 1540953f064aSLi Xinhai huge_page_order(h), true); 1541953f064aSLi Xinhai if (page) 154219fc1a7eSSidhartha Kumar return page_folio(page); 1543953f064aSLi Xinhai } 1544953f064aSLi Xinhai 1545953f064aSLi Xinhai if (!(gfp_mask & __GFP_THISNODE)) { 1546cf11e85fSRoman Gushchin for_each_node_mask(node, *nodemask) { 1547953f064aSLi Xinhai if (node == nid || !hugetlb_cma[node]) 1548cf11e85fSRoman Gushchin continue; 1549cf11e85fSRoman Gushchin 1550cf11e85fSRoman Gushchin page = cma_alloc(hugetlb_cma[node], nr_pages, 1551cf11e85fSRoman Gushchin huge_page_order(h), true); 1552cf11e85fSRoman Gushchin if (page) 155319fc1a7eSSidhartha Kumar return page_folio(page); 1554cf11e85fSRoman Gushchin } 1555cf11e85fSRoman Gushchin } 1556953f064aSLi Xinhai } 1557dbda8feaSBarry Song #endif 1558cf11e85fSRoman Gushchin 155919fc1a7eSSidhartha Kumar page = alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); 156019fc1a7eSSidhartha Kumar return page ? page_folio(page) : NULL; 1561944d9fecSLuiz Capitulino } 1562944d9fecSLuiz Capitulino 15634eb0716eSAlexandre Ghiti #else /* !CONFIG_CONTIG_ALLOC */ 156419fc1a7eSSidhartha Kumar static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask, 15654eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 15664eb0716eSAlexandre Ghiti { 15674eb0716eSAlexandre Ghiti return NULL; 15684eb0716eSAlexandre Ghiti } 15694eb0716eSAlexandre Ghiti #endif /* CONFIG_CONTIG_ALLOC */ 1570944d9fecSLuiz Capitulino 1571e1073d1eSAneesh Kumar K.V #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 157219fc1a7eSSidhartha Kumar static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask, 15734eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 15744eb0716eSAlexandre Ghiti { 15754eb0716eSAlexandre Ghiti return NULL; 15764eb0716eSAlexandre Ghiti } 15777f325a8dSSidhartha Kumar static inline void free_gigantic_folio(struct folio *folio, 15787f325a8dSSidhartha Kumar unsigned int order) { } 1579911565b8SSidhartha Kumar static inline void destroy_compound_gigantic_folio(struct folio *folio, 1580d00181b9SKirill A. Shutemov unsigned int order) { } 1581944d9fecSLuiz Capitulino #endif 1582944d9fecSLuiz Capitulino 15836eb4e88aSMike Kravetz /* 1584cfd5082bSSidhartha Kumar * Remove hugetlb folio from lists, and update dtor so that the folio appears 158534d9e35bSMike Kravetz * as just a compound page. 158634d9e35bSMike Kravetz * 1587cfd5082bSSidhartha Kumar * A reference is held on the folio, except in the case of demote. 15886eb4e88aSMike Kravetz * 15896eb4e88aSMike Kravetz * Must be called with hugetlb lock held. 15906eb4e88aSMike Kravetz */ 1591cfd5082bSSidhartha Kumar static void __remove_hugetlb_folio(struct hstate *h, struct folio *folio, 159234d9e35bSMike Kravetz bool adjust_surplus, 159334d9e35bSMike Kravetz bool demote) 15946eb4e88aSMike Kravetz { 1595cfd5082bSSidhartha Kumar int nid = folio_nid(folio); 15966eb4e88aSMike Kravetz 1597f074732dSSidhartha Kumar VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio(folio), folio); 1598f074732dSSidhartha Kumar VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio_rsvd(folio), folio); 15996eb4e88aSMike Kravetz 16009487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 16016eb4e88aSMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 16026eb4e88aSMike Kravetz return; 16036eb4e88aSMike Kravetz 1604cfd5082bSSidhartha Kumar list_del(&folio->lru); 16056eb4e88aSMike Kravetz 1606cfd5082bSSidhartha Kumar if (folio_test_hugetlb_freed(folio)) { 16076eb4e88aSMike Kravetz h->free_huge_pages--; 16086eb4e88aSMike Kravetz h->free_huge_pages_node[nid]--; 16096eb4e88aSMike Kravetz } 16106eb4e88aSMike Kravetz if (adjust_surplus) { 16116eb4e88aSMike Kravetz h->surplus_huge_pages--; 16126eb4e88aSMike Kravetz h->surplus_huge_pages_node[nid]--; 16136eb4e88aSMike Kravetz } 16146eb4e88aSMike Kravetz 1615e32d20c0SMike Kravetz /* 1616e32d20c0SMike Kravetz * Very subtle 1617e32d20c0SMike Kravetz * 1618e32d20c0SMike Kravetz * For non-gigantic pages set the destructor to the normal compound 1619e32d20c0SMike Kravetz * page dtor. This is needed in case someone takes an additional 1620e32d20c0SMike Kravetz * temporary ref to the page, and freeing is delayed until they drop 1621e32d20c0SMike Kravetz * their reference. 1622e32d20c0SMike Kravetz * 1623e32d20c0SMike Kravetz * For gigantic pages set the destructor to the null dtor. This 1624e32d20c0SMike Kravetz * destructor will never be called. Before freeing the gigantic 1625911565b8SSidhartha Kumar * page destroy_compound_gigantic_folio will turn the folio into a 1626911565b8SSidhartha Kumar * simple group of pages. After this the destructor does not 1627e32d20c0SMike Kravetz * apply. 1628e32d20c0SMike Kravetz * 1629e32d20c0SMike Kravetz * This handles the case where more than one ref is held when and 1630d6ef19e2SSidhartha Kumar * after update_and_free_hugetlb_folio is called. 163134d9e35bSMike Kravetz * 163234d9e35bSMike Kravetz * In the case of demote we do not ref count the page as it will soon 163334d9e35bSMike Kravetz * be turned into a page of smaller size. 1634e32d20c0SMike Kravetz */ 163534d9e35bSMike Kravetz if (!demote) 1636cfd5082bSSidhartha Kumar folio_ref_unfreeze(folio, 1); 1637e32d20c0SMike Kravetz if (hstate_is_gigantic(h)) 1638cfd5082bSSidhartha Kumar folio_set_compound_dtor(folio, NULL_COMPOUND_DTOR); 1639e32d20c0SMike Kravetz else 1640cfd5082bSSidhartha Kumar folio_set_compound_dtor(folio, COMPOUND_PAGE_DTOR); 16416eb4e88aSMike Kravetz 16426eb4e88aSMike Kravetz h->nr_huge_pages--; 16436eb4e88aSMike Kravetz h->nr_huge_pages_node[nid]--; 16446eb4e88aSMike Kravetz } 16456eb4e88aSMike Kravetz 1646cfd5082bSSidhartha Kumar static void remove_hugetlb_folio(struct hstate *h, struct folio *folio, 164734d9e35bSMike Kravetz bool adjust_surplus) 164834d9e35bSMike Kravetz { 1649cfd5082bSSidhartha Kumar __remove_hugetlb_folio(h, folio, adjust_surplus, false); 165034d9e35bSMike Kravetz } 165134d9e35bSMike Kravetz 1652cfd5082bSSidhartha Kumar static void remove_hugetlb_folio_for_demote(struct hstate *h, struct folio *folio, 16538531fc6fSMike Kravetz bool adjust_surplus) 16548531fc6fSMike Kravetz { 1655cfd5082bSSidhartha Kumar __remove_hugetlb_folio(h, folio, adjust_surplus, true); 16568531fc6fSMike Kravetz } 16578531fc6fSMike Kravetz 16582f6c57d6SSidhartha Kumar static void add_hugetlb_folio(struct hstate *h, struct folio *folio, 1659ad2fa371SMuchun Song bool adjust_surplus) 1660ad2fa371SMuchun Song { 1661ad2fa371SMuchun Song int zeroed; 16622f6c57d6SSidhartha Kumar int nid = folio_nid(folio); 1663ad2fa371SMuchun Song 16642f6c57d6SSidhartha Kumar VM_BUG_ON_FOLIO(!folio_test_hugetlb_vmemmap_optimized(folio), folio); 1665ad2fa371SMuchun Song 1666ad2fa371SMuchun Song lockdep_assert_held(&hugetlb_lock); 1667ad2fa371SMuchun Song 16682f6c57d6SSidhartha Kumar INIT_LIST_HEAD(&folio->lru); 1669ad2fa371SMuchun Song h->nr_huge_pages++; 1670ad2fa371SMuchun Song h->nr_huge_pages_node[nid]++; 1671ad2fa371SMuchun Song 1672ad2fa371SMuchun Song if (adjust_surplus) { 1673ad2fa371SMuchun Song h->surplus_huge_pages++; 1674ad2fa371SMuchun Song h->surplus_huge_pages_node[nid]++; 1675ad2fa371SMuchun Song } 1676ad2fa371SMuchun Song 16772f6c57d6SSidhartha Kumar folio_set_compound_dtor(folio, HUGETLB_PAGE_DTOR); 16782f6c57d6SSidhartha Kumar folio_change_private(folio, NULL); 1679a9e1eab2SMiaohe Lin /* 16802f6c57d6SSidhartha Kumar * We have to set hugetlb_vmemmap_optimized again as above 16812f6c57d6SSidhartha Kumar * folio_change_private(folio, NULL) cleared it. 1682a9e1eab2SMiaohe Lin */ 16832f6c57d6SSidhartha Kumar folio_set_hugetlb_vmemmap_optimized(folio); 1684ad2fa371SMuchun Song 1685ad2fa371SMuchun Song /* 16862f6c57d6SSidhartha Kumar * This folio is about to be managed by the hugetlb allocator and 1687b65a4edaSMike Kravetz * should have no users. Drop our reference, and check for others 1688b65a4edaSMike Kravetz * just in case. 1689ad2fa371SMuchun Song */ 16902f6c57d6SSidhartha Kumar zeroed = folio_put_testzero(folio); 16912f6c57d6SSidhartha Kumar if (unlikely(!zeroed)) 1692b65a4edaSMike Kravetz /* 1693b65a4edaSMike Kravetz * It is VERY unlikely soneone else has taken a ref on 1694b65a4edaSMike Kravetz * the page. In this case, we simply return as the 1695b65a4edaSMike Kravetz * hugetlb destructor (free_huge_page) will be called 1696b65a4edaSMike Kravetz * when this other ref is dropped. 1697b65a4edaSMike Kravetz */ 1698b65a4edaSMike Kravetz return; 1699b65a4edaSMike Kravetz 17002f6c57d6SSidhartha Kumar arch_clear_hugepage_flags(&folio->page); 1701240d67a8SSidhartha Kumar enqueue_hugetlb_folio(h, folio); 1702ad2fa371SMuchun Song } 1703ad2fa371SMuchun Song 17046f6956cfSSidhartha Kumar static void __update_and_free_hugetlb_folio(struct hstate *h, 17056f6956cfSSidhartha Kumar struct folio *folio) 17066af2acb6SAdam Litke { 17076af2acb6SAdam Litke int i; 170814455eabSCheng Li struct page *subpage; 1709a5516438SAndi Kleen 17104eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1711944d9fecSLuiz Capitulino return; 171218229df5SAndy Whitcroft 1713161df60eSNaoya Horiguchi /* 1714161df60eSNaoya Horiguchi * If we don't know which subpages are hwpoisoned, we can't free 1715161df60eSNaoya Horiguchi * the hugepage, so it's leaked intentionally. 1716161df60eSNaoya Horiguchi */ 17177f325a8dSSidhartha Kumar if (folio_test_hugetlb_raw_hwp_unreliable(folio)) 1718161df60eSNaoya Horiguchi return; 1719161df60eSNaoya Horiguchi 17206f6956cfSSidhartha Kumar if (hugetlb_vmemmap_restore(h, &folio->page)) { 1721ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 1722ad2fa371SMuchun Song /* 1723ad2fa371SMuchun Song * If we cannot allocate vmemmap pages, just refuse to free the 1724ad2fa371SMuchun Song * page and put the page back on the hugetlb free list and treat 1725ad2fa371SMuchun Song * as a surplus page. 1726ad2fa371SMuchun Song */ 17277f325a8dSSidhartha Kumar add_hugetlb_folio(h, folio, true); 1728ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 1729ad2fa371SMuchun Song return; 1730ad2fa371SMuchun Song } 1731ad2fa371SMuchun Song 1732161df60eSNaoya Horiguchi /* 1733161df60eSNaoya Horiguchi * Move PageHWPoison flag from head page to the raw error pages, 1734161df60eSNaoya Horiguchi * which makes any healthy subpages reusable. 1735161df60eSNaoya Horiguchi */ 1736911565b8SSidhartha Kumar if (unlikely(folio_test_hwpoison(folio))) 17372ff6ceceSSidhartha Kumar folio_clear_hugetlb_hwpoison(folio); 1738161df60eSNaoya Horiguchi 173914455eabSCheng Li for (i = 0; i < pages_per_huge_page(h); i++) { 17407f325a8dSSidhartha Kumar subpage = folio_page(folio, i); 1741dbfee5aeSMike Kravetz subpage->flags &= ~(1 << PG_locked | 1 << PG_error | 174232f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1743a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1744a7407a27SLuiz Capitulino 1 << PG_writeback); 17456af2acb6SAdam Litke } 1746a01f4390SMike Kravetz 1747a01f4390SMike Kravetz /* 1748a01f4390SMike Kravetz * Non-gigantic pages demoted from CMA allocated gigantic pages 17497f325a8dSSidhartha Kumar * need to be given back to CMA in free_gigantic_folio. 1750a01f4390SMike Kravetz */ 1751a01f4390SMike Kravetz if (hstate_is_gigantic(h) || 17522f6c57d6SSidhartha Kumar hugetlb_cma_folio(folio, huge_page_order(h))) { 1753911565b8SSidhartha Kumar destroy_compound_gigantic_folio(folio, huge_page_order(h)); 17547f325a8dSSidhartha Kumar free_gigantic_folio(folio, huge_page_order(h)); 1755944d9fecSLuiz Capitulino } else { 17566f6956cfSSidhartha Kumar __free_pages(&folio->page, huge_page_order(h)); 17576af2acb6SAdam Litke } 1758944d9fecSLuiz Capitulino } 17596af2acb6SAdam Litke 1760b65d4adbSMuchun Song /* 1761d6ef19e2SSidhartha Kumar * As update_and_free_hugetlb_folio() can be called under any context, so we cannot 1762b65d4adbSMuchun Song * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the 1763b65d4adbSMuchun Song * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate 1764b65d4adbSMuchun Song * the vmemmap pages. 1765b65d4adbSMuchun Song * 1766b65d4adbSMuchun Song * free_hpage_workfn() locklessly retrieves the linked list of pages to be 1767b65d4adbSMuchun Song * freed and frees them one-by-one. As the page->mapping pointer is going 1768b65d4adbSMuchun Song * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node 1769b65d4adbSMuchun Song * structure of a lockless linked list of huge pages to be freed. 1770b65d4adbSMuchun Song */ 1771b65d4adbSMuchun Song static LLIST_HEAD(hpage_freelist); 1772b65d4adbSMuchun Song 1773b65d4adbSMuchun Song static void free_hpage_workfn(struct work_struct *work) 1774b65d4adbSMuchun Song { 1775b65d4adbSMuchun Song struct llist_node *node; 1776b65d4adbSMuchun Song 1777b65d4adbSMuchun Song node = llist_del_all(&hpage_freelist); 1778b65d4adbSMuchun Song 1779b65d4adbSMuchun Song while (node) { 1780b65d4adbSMuchun Song struct page *page; 1781b65d4adbSMuchun Song struct hstate *h; 1782b65d4adbSMuchun Song 1783b65d4adbSMuchun Song page = container_of((struct address_space **)node, 1784b65d4adbSMuchun Song struct page, mapping); 1785b65d4adbSMuchun Song node = node->next; 1786b65d4adbSMuchun Song page->mapping = NULL; 1787b65d4adbSMuchun Song /* 1788b65d4adbSMuchun Song * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate() 1789b65d4adbSMuchun Song * is going to trigger because a previous call to 1790cfd5082bSSidhartha Kumar * remove_hugetlb_folio() will call folio_set_compound_dtor 1791cfd5082bSSidhartha Kumar * (folio, NULL_COMPOUND_DTOR), so do not use page_hstate() 1792cfd5082bSSidhartha Kumar * directly. 1793b65d4adbSMuchun Song */ 1794b65d4adbSMuchun Song h = size_to_hstate(page_size(page)); 1795b65d4adbSMuchun Song 17966f6956cfSSidhartha Kumar __update_and_free_hugetlb_folio(h, page_folio(page)); 1797b65d4adbSMuchun Song 1798b65d4adbSMuchun Song cond_resched(); 1799b65d4adbSMuchun Song } 1800b65d4adbSMuchun Song } 1801b65d4adbSMuchun Song static DECLARE_WORK(free_hpage_work, free_hpage_workfn); 1802b65d4adbSMuchun Song 1803b65d4adbSMuchun Song static inline void flush_free_hpage_work(struct hstate *h) 1804b65d4adbSMuchun Song { 18056213834cSMuchun Song if (hugetlb_vmemmap_optimizable(h)) 1806b65d4adbSMuchun Song flush_work(&free_hpage_work); 1807b65d4adbSMuchun Song } 1808b65d4adbSMuchun Song 1809d6ef19e2SSidhartha Kumar static void update_and_free_hugetlb_folio(struct hstate *h, struct folio *folio, 1810b65d4adbSMuchun Song bool atomic) 1811b65d4adbSMuchun Song { 1812d6ef19e2SSidhartha Kumar if (!folio_test_hugetlb_vmemmap_optimized(folio) || !atomic) { 18136f6956cfSSidhartha Kumar __update_and_free_hugetlb_folio(h, folio); 1814b65d4adbSMuchun Song return; 1815b65d4adbSMuchun Song } 1816b65d4adbSMuchun Song 1817b65d4adbSMuchun Song /* 1818b65d4adbSMuchun Song * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages. 1819b65d4adbSMuchun Song * 1820b65d4adbSMuchun Song * Only call schedule_work() if hpage_freelist is previously 1821b65d4adbSMuchun Song * empty. Otherwise, schedule_work() had been called but the workfn 1822b65d4adbSMuchun Song * hasn't retrieved the list yet. 1823b65d4adbSMuchun Song */ 1824d6ef19e2SSidhartha Kumar if (llist_add((struct llist_node *)&folio->mapping, &hpage_freelist)) 1825b65d4adbSMuchun Song schedule_work(&free_hpage_work); 1826b65d4adbSMuchun Song } 1827b65d4adbSMuchun Song 182810c6ec49SMike Kravetz static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list) 182910c6ec49SMike Kravetz { 183010c6ec49SMike Kravetz struct page *page, *t_page; 1831d6ef19e2SSidhartha Kumar struct folio *folio; 183210c6ec49SMike Kravetz 183310c6ec49SMike Kravetz list_for_each_entry_safe(page, t_page, list, lru) { 1834d6ef19e2SSidhartha Kumar folio = page_folio(page); 1835d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, folio, false); 183610c6ec49SMike Kravetz cond_resched(); 183710c6ec49SMike Kravetz } 183810c6ec49SMike Kravetz } 183910c6ec49SMike Kravetz 1840e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1841e5ff2159SAndi Kleen { 1842e5ff2159SAndi Kleen struct hstate *h; 1843e5ff2159SAndi Kleen 1844e5ff2159SAndi Kleen for_each_hstate(h) { 1845e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1846e5ff2159SAndi Kleen return h; 1847e5ff2159SAndi Kleen } 1848e5ff2159SAndi Kleen return NULL; 1849e5ff2159SAndi Kleen } 1850e5ff2159SAndi Kleen 1851db71ef79SMike Kravetz void free_huge_page(struct page *page) 185227a85ef1SDavid Gibson { 1853a5516438SAndi Kleen /* 1854a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1855a5516438SAndi Kleen * compound page destructor. 1856a5516438SAndi Kleen */ 18570356c4b9SSidhartha Kumar struct folio *folio = page_folio(page); 18580356c4b9SSidhartha Kumar struct hstate *h = folio_hstate(folio); 18590356c4b9SSidhartha Kumar int nid = folio_nid(folio); 18600356c4b9SSidhartha Kumar struct hugepage_subpool *spool = hugetlb_folio_subpool(folio); 186107443a85SJoonsoo Kim bool restore_reserve; 1862db71ef79SMike Kravetz unsigned long flags; 186327a85ef1SDavid Gibson 18640356c4b9SSidhartha Kumar VM_BUG_ON_FOLIO(folio_ref_count(folio), folio); 18650356c4b9SSidhartha Kumar VM_BUG_ON_FOLIO(folio_mapcount(folio), folio); 18668ace22bcSYongkai Wu 18670356c4b9SSidhartha Kumar hugetlb_set_folio_subpool(folio, NULL); 18680356c4b9SSidhartha Kumar if (folio_test_anon(folio)) 18690356c4b9SSidhartha Kumar __ClearPageAnonExclusive(&folio->page); 18700356c4b9SSidhartha Kumar folio->mapping = NULL; 18710356c4b9SSidhartha Kumar restore_reserve = folio_test_hugetlb_restore_reserve(folio); 18720356c4b9SSidhartha Kumar folio_clear_hugetlb_restore_reserve(folio); 187327a85ef1SDavid Gibson 18741c5ecae3SMike Kravetz /* 1875d6995da3SMike Kravetz * If HPageRestoreReserve was set on page, page allocation consumed a 18760919e1b6SMike Kravetz * reservation. If the page was associated with a subpool, there 18770919e1b6SMike Kravetz * would have been a page reserved in the subpool before allocation 18780919e1b6SMike Kravetz * via hugepage_subpool_get_pages(). Since we are 'restoring' the 18796c26d310SMiaohe Lin * reservation, do not call hugepage_subpool_put_pages() as this will 18800919e1b6SMike Kravetz * remove the reserved page from the subpool. 18810919e1b6SMike Kravetz */ 18820919e1b6SMike Kravetz if (!restore_reserve) { 18830919e1b6SMike Kravetz /* 18840919e1b6SMike Kravetz * A return code of zero implies that the subpool will be 18850919e1b6SMike Kravetz * under its minimum size if the reservation is not restored 18860919e1b6SMike Kravetz * after page is free. Therefore, force restore_reserve 18870919e1b6SMike Kravetz * operation. 18881c5ecae3SMike Kravetz */ 18891c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 18901c5ecae3SMike Kravetz restore_reserve = true; 18910919e1b6SMike Kravetz } 18921c5ecae3SMike Kravetz 1893db71ef79SMike Kravetz spin_lock_irqsave(&hugetlb_lock, flags); 18940356c4b9SSidhartha Kumar folio_clear_hugetlb_migratable(folio); 1895d4ab0316SSidhartha Kumar hugetlb_cgroup_uncharge_folio(hstate_index(h), 1896d4ab0316SSidhartha Kumar pages_per_huge_page(h), folio); 1897d4ab0316SSidhartha Kumar hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h), 1898d4ab0316SSidhartha Kumar pages_per_huge_page(h), folio); 189907443a85SJoonsoo Kim if (restore_reserve) 190007443a85SJoonsoo Kim h->resv_huge_pages++; 190107443a85SJoonsoo Kim 19020356c4b9SSidhartha Kumar if (folio_test_hugetlb_temporary(folio)) { 1903cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, false); 1904db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1905d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, folio, true); 1906ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 19070edaecfaSAneesh Kumar K.V /* remove the page from active list */ 1908cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, true); 1909db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1910d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, folio, true); 19117893d1d5SAdam Litke } else { 19125d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1913240d67a8SSidhartha Kumar enqueue_hugetlb_folio(h, folio); 1914db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 191527a85ef1SDavid Gibson } 19161121828aSMike Kravetz } 191727a85ef1SDavid Gibson 1918d3d99fccSOscar Salvador /* 1919d3d99fccSOscar Salvador * Must be called with the hugetlb lock held 1920d3d99fccSOscar Salvador */ 1921d3d99fccSOscar Salvador static void __prep_account_new_huge_page(struct hstate *h, int nid) 1922d3d99fccSOscar Salvador { 1923d3d99fccSOscar Salvador lockdep_assert_held(&hugetlb_lock); 1924d3d99fccSOscar Salvador h->nr_huge_pages++; 1925d3d99fccSOscar Salvador h->nr_huge_pages_node[nid]++; 1926d3d99fccSOscar Salvador } 1927d3d99fccSOscar Salvador 1928de656ed3SSidhartha Kumar static void __prep_new_hugetlb_folio(struct hstate *h, struct folio *folio) 1929b7ba30c6SAndi Kleen { 1930de656ed3SSidhartha Kumar hugetlb_vmemmap_optimize(h, &folio->page); 1931de656ed3SSidhartha Kumar INIT_LIST_HEAD(&folio->lru); 19329fd33058SSidhartha Kumar folio_set_compound_dtor(folio, HUGETLB_PAGE_DTOR); 1933de656ed3SSidhartha Kumar hugetlb_set_folio_subpool(folio, NULL); 1934de656ed3SSidhartha Kumar set_hugetlb_cgroup(folio, NULL); 1935de656ed3SSidhartha Kumar set_hugetlb_cgroup_rsvd(folio, NULL); 1936d3d99fccSOscar Salvador } 1937d3d99fccSOscar Salvador 1938d1c60955SSidhartha Kumar static void prep_new_hugetlb_folio(struct hstate *h, struct folio *folio, int nid) 1939d3d99fccSOscar Salvador { 1940de656ed3SSidhartha Kumar __prep_new_hugetlb_folio(h, folio); 1941db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 1942d3d99fccSOscar Salvador __prep_account_new_huge_page(h, nid); 1943db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1944b7ba30c6SAndi Kleen } 1945b7ba30c6SAndi Kleen 1946d1c60955SSidhartha Kumar static bool __prep_compound_gigantic_folio(struct folio *folio, 1947d1c60955SSidhartha Kumar unsigned int order, bool demote) 194820a0307cSWu Fengguang { 19497118fc29SMike Kravetz int i, j; 195020a0307cSWu Fengguang int nr_pages = 1 << order; 195114455eabSCheng Li struct page *p; 195220a0307cSWu Fengguang 1953d1c60955SSidhartha Kumar __folio_clear_reserved(folio); 1954d1c60955SSidhartha Kumar __folio_set_head(folio); 1955c45bc55aSSidhartha Kumar /* we rely on prep_new_hugetlb_folio to set the destructor */ 195604a42e72SSidhartha Kumar folio_set_order(folio, order); 19572b21624fSMike Kravetz for (i = 0; i < nr_pages; i++) { 1958d1c60955SSidhartha Kumar p = folio_page(folio, i); 195914455eabSCheng Li 1960ef5a22beSAndrea Arcangeli /* 1961ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1962ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1963ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 19647c8de358SEthon Paul * too. Otherwise drivers using get_user_pages() to access tail 1965ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1966ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1967ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1968ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1969ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1970ef5a22beSAndrea Arcangeli * after get_user_pages(). 1971ef5a22beSAndrea Arcangeli */ 19727fb0728aSMike Kravetz if (i != 0) /* head page cleared above */ 1973ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 19747118fc29SMike Kravetz /* 19757118fc29SMike Kravetz * Subtle and very unlikely 19767118fc29SMike Kravetz * 19777118fc29SMike Kravetz * Gigantic 'page allocators' such as memblock or cma will 19787118fc29SMike Kravetz * return a set of pages with each page ref counted. We need 19797118fc29SMike Kravetz * to turn this set of pages into a compound page with tail 19807118fc29SMike Kravetz * page ref counts set to zero. Code such as speculative page 19817118fc29SMike Kravetz * cache adding could take a ref on a 'to be' tail page. 19827118fc29SMike Kravetz * We need to respect any increased ref count, and only set 19837118fc29SMike Kravetz * the ref count to zero if count is currently 1. If count 1984416d85edSMike Kravetz * is not 1, we return an error. An error return indicates 1985416d85edSMike Kravetz * the set of pages can not be converted to a gigantic page. 1986416d85edSMike Kravetz * The caller who allocated the pages should then discard the 1987416d85edSMike Kravetz * pages using the appropriate free interface. 198834d9e35bSMike Kravetz * 198934d9e35bSMike Kravetz * In the case of demote, the ref count will be zero. 19907118fc29SMike Kravetz */ 199134d9e35bSMike Kravetz if (!demote) { 19927118fc29SMike Kravetz if (!page_ref_freeze(p, 1)) { 1993416d85edSMike Kravetz pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n"); 19947118fc29SMike Kravetz goto out_error; 19957118fc29SMike Kravetz } 199634d9e35bSMike Kravetz } else { 199734d9e35bSMike Kravetz VM_BUG_ON_PAGE(page_count(p), p); 199834d9e35bSMike Kravetz } 19992b21624fSMike Kravetz if (i != 0) 2000d1c60955SSidhartha Kumar set_compound_head(p, &folio->page); 200120a0307cSWu Fengguang } 200246f27228SMatthew Wilcox (Oracle) atomic_set(&folio->_entire_mapcount, -1); 2003eec20426SMatthew Wilcox (Oracle) atomic_set(&folio->_nr_pages_mapped, 0); 200494688e8eSMatthew Wilcox (Oracle) atomic_set(&folio->_pincount, 0); 20057118fc29SMike Kravetz return true; 20067118fc29SMike Kravetz 20077118fc29SMike Kravetz out_error: 20082b21624fSMike Kravetz /* undo page modifications made above */ 20092b21624fSMike Kravetz for (j = 0; j < i; j++) { 2010d1c60955SSidhartha Kumar p = folio_page(folio, j); 20112b21624fSMike Kravetz if (j != 0) 20127118fc29SMike Kravetz clear_compound_head(p); 20137118fc29SMike Kravetz set_page_refcounted(p); 20147118fc29SMike Kravetz } 20157118fc29SMike Kravetz /* need to clear PG_reserved on remaining tail pages */ 201614455eabSCheng Li for (; j < nr_pages; j++) { 2017d1c60955SSidhartha Kumar p = folio_page(folio, j); 20187118fc29SMike Kravetz __ClearPageReserved(p); 201914455eabSCheng Li } 202004a42e72SSidhartha Kumar folio_set_order(folio, 0); 2021d1c60955SSidhartha Kumar __folio_clear_head(folio); 20227118fc29SMike Kravetz return false; 202320a0307cSWu Fengguang } 202420a0307cSWu Fengguang 2025d1c60955SSidhartha Kumar static bool prep_compound_gigantic_folio(struct folio *folio, 20268531fc6fSMike Kravetz unsigned int order) 20278531fc6fSMike Kravetz { 2028d1c60955SSidhartha Kumar return __prep_compound_gigantic_folio(folio, order, false); 2029d1c60955SSidhartha Kumar } 2030d1c60955SSidhartha Kumar 2031d1c60955SSidhartha Kumar static bool prep_compound_gigantic_folio_for_demote(struct folio *folio, 2032d1c60955SSidhartha Kumar unsigned int order) 2033d1c60955SSidhartha Kumar { 2034d1c60955SSidhartha Kumar return __prep_compound_gigantic_folio(folio, order, true); 20358531fc6fSMike Kravetz } 20368531fc6fSMike Kravetz 20377795912cSAndrew Morton /* 20387795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 20397795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 20407795912cSAndrew Morton * details. 20417795912cSAndrew Morton */ 204220a0307cSWu Fengguang int PageHuge(struct page *page) 204320a0307cSWu Fengguang { 20442d678c64SMatthew Wilcox (Oracle) struct folio *folio; 20452d678c64SMatthew Wilcox (Oracle) 204620a0307cSWu Fengguang if (!PageCompound(page)) 204720a0307cSWu Fengguang return 0; 20482d678c64SMatthew Wilcox (Oracle) folio = page_folio(page); 20492d678c64SMatthew Wilcox (Oracle) return folio->_folio_dtor == HUGETLB_PAGE_DTOR; 205020a0307cSWu Fengguang } 205143131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 205243131e14SNaoya Horiguchi 2053957ebbdfSMatthew Wilcox (Oracle) /** 2054957ebbdfSMatthew Wilcox (Oracle) * folio_test_hugetlb - Determine if the folio belongs to hugetlbfs 2055957ebbdfSMatthew Wilcox (Oracle) * @folio: The folio to test. 2056957ebbdfSMatthew Wilcox (Oracle) * 2057957ebbdfSMatthew Wilcox (Oracle) * Context: Any context. Caller should have a reference on the folio to 2058957ebbdfSMatthew Wilcox (Oracle) * prevent it from being turned into a tail page. 2059957ebbdfSMatthew Wilcox (Oracle) * Return: True for hugetlbfs folios, false for anon folios or folios 2060957ebbdfSMatthew Wilcox (Oracle) * belonging to other filesystems. 206127c73ae7SAndrea Arcangeli */ 2062957ebbdfSMatthew Wilcox (Oracle) bool folio_test_hugetlb(struct folio *folio) 206327c73ae7SAndrea Arcangeli { 20642d678c64SMatthew Wilcox (Oracle) if (!folio_test_large(folio)) 2065957ebbdfSMatthew Wilcox (Oracle) return false; 206627c73ae7SAndrea Arcangeli 20672d678c64SMatthew Wilcox (Oracle) return folio->_folio_dtor == HUGETLB_PAGE_DTOR; 206827c73ae7SAndrea Arcangeli } 2069957ebbdfSMatthew Wilcox (Oracle) EXPORT_SYMBOL_GPL(folio_test_hugetlb); 207027c73ae7SAndrea Arcangeli 2071c0d0381aSMike Kravetz /* 2072c0d0381aSMike Kravetz * Find and lock address space (mapping) in write mode. 2073c0d0381aSMike Kravetz * 2074336bf30eSMike Kravetz * Upon entry, the page is locked which means that page_mapping() is 2075336bf30eSMike Kravetz * stable. Due to locking order, we can only trylock_write. If we can 2076336bf30eSMike Kravetz * not get the lock, simply return NULL to caller. 2077c0d0381aSMike Kravetz */ 2078c0d0381aSMike Kravetz struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) 2079c0d0381aSMike Kravetz { 2080336bf30eSMike Kravetz struct address_space *mapping = page_mapping(hpage); 2081c0d0381aSMike Kravetz 2082c0d0381aSMike Kravetz if (!mapping) 2083c0d0381aSMike Kravetz return mapping; 2084c0d0381aSMike Kravetz 2085c0d0381aSMike Kravetz if (i_mmap_trylock_write(mapping)) 2086c0d0381aSMike Kravetz return mapping; 2087c0d0381aSMike Kravetz 2088c0d0381aSMike Kravetz return NULL; 2089c0d0381aSMike Kravetz } 2090c0d0381aSMike Kravetz 2091fe19bd3dSHugh Dickins pgoff_t hugetlb_basepage_index(struct page *page) 209213d60f4bSZhang Yi { 209313d60f4bSZhang Yi struct page *page_head = compound_head(page); 209413d60f4bSZhang Yi pgoff_t index = page_index(page_head); 209513d60f4bSZhang Yi unsigned long compound_idx; 209613d60f4bSZhang Yi 209723baf831SKirill A. Shutemov if (compound_order(page_head) > MAX_ORDER) 209813d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 209913d60f4bSZhang Yi else 210013d60f4bSZhang Yi compound_idx = page - page_head; 210113d60f4bSZhang Yi 210213d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 210313d60f4bSZhang Yi } 210413d60f4bSZhang Yi 210519fc1a7eSSidhartha Kumar static struct folio *alloc_buddy_hugetlb_folio(struct hstate *h, 2106f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 2107f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 21081da177e4SLinus Torvalds { 2109af0fb9dfSMichal Hocko int order = huge_page_order(h); 21101da177e4SLinus Torvalds struct page *page; 2111f60858f9SMike Kravetz bool alloc_try_hard = true; 21122b21624fSMike Kravetz bool retry = true; 2113f96efd58SJoe Jin 2114f60858f9SMike Kravetz /* 2115f60858f9SMike Kravetz * By default we always try hard to allocate the page with 2116f60858f9SMike Kravetz * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in 2117f60858f9SMike Kravetz * a loop (to adjust global huge page counts) and previous allocation 2118f60858f9SMike Kravetz * failed, do not continue to try hard on the same node. Use the 2119f60858f9SMike Kravetz * node_alloc_noretry bitmap to manage this state information. 2120f60858f9SMike Kravetz */ 2121f60858f9SMike Kravetz if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) 2122f60858f9SMike Kravetz alloc_try_hard = false; 2123f60858f9SMike Kravetz gfp_mask |= __GFP_COMP|__GFP_NOWARN; 2124f60858f9SMike Kravetz if (alloc_try_hard) 2125f60858f9SMike Kravetz gfp_mask |= __GFP_RETRY_MAYFAIL; 2126af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 2127af0fb9dfSMichal Hocko nid = numa_mem_id(); 21282b21624fSMike Kravetz retry: 212984172f4bSMatthew Wilcox (Oracle) page = __alloc_pages(gfp_mask, order, nid, nmask); 21302b21624fSMike Kravetz 21312b21624fSMike Kravetz /* Freeze head page */ 21322b21624fSMike Kravetz if (page && !page_ref_freeze(page, 1)) { 21332b21624fSMike Kravetz __free_pages(page, order); 21342b21624fSMike Kravetz if (retry) { /* retry once */ 21352b21624fSMike Kravetz retry = false; 21362b21624fSMike Kravetz goto retry; 21372b21624fSMike Kravetz } 21382b21624fSMike Kravetz /* WOW! twice in a row. */ 21392b21624fSMike Kravetz pr_warn("HugeTLB head page unexpected inflated ref count\n"); 21402b21624fSMike Kravetz page = NULL; 21412b21624fSMike Kravetz } 21422b21624fSMike Kravetz 2143f60858f9SMike Kravetz /* 2144f60858f9SMike Kravetz * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this 2145f60858f9SMike Kravetz * indicates an overall state change. Clear bit so that we resume 2146f60858f9SMike Kravetz * normal 'try hard' allocations. 2147f60858f9SMike Kravetz */ 2148f60858f9SMike Kravetz if (node_alloc_noretry && page && !alloc_try_hard) 2149f60858f9SMike Kravetz node_clear(nid, *node_alloc_noretry); 2150f60858f9SMike Kravetz 2151f60858f9SMike Kravetz /* 2152f60858f9SMike Kravetz * If we tried hard to get a page but failed, set bit so that 2153f60858f9SMike Kravetz * subsequent attempts will not try as hard until there is an 2154f60858f9SMike Kravetz * overall state change. 2155f60858f9SMike Kravetz */ 2156f60858f9SMike Kravetz if (node_alloc_noretry && !page && alloc_try_hard) 2157f60858f9SMike Kravetz node_set(nid, *node_alloc_noretry); 2158f60858f9SMike Kravetz 215919fc1a7eSSidhartha Kumar if (!page) { 216019fc1a7eSSidhartha Kumar __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 216119fc1a7eSSidhartha Kumar return NULL; 216219fc1a7eSSidhartha Kumar } 216319fc1a7eSSidhartha Kumar 216419fc1a7eSSidhartha Kumar __count_vm_event(HTLB_BUDDY_PGALLOC); 216519fc1a7eSSidhartha Kumar return page_folio(page); 216663b4613cSNishanth Aravamudan } 216763b4613cSNishanth Aravamudan 2168af0fb9dfSMichal Hocko /* 21690c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 21700c397daeSMichal Hocko * should use this function to get new hugetlb pages 21712b21624fSMike Kravetz * 21722b21624fSMike Kravetz * Note that returned page is 'frozen': ref count of head page and all tail 21732b21624fSMike Kravetz * pages is zero. 21740c397daeSMichal Hocko */ 217519fc1a7eSSidhartha Kumar static struct folio *alloc_fresh_hugetlb_folio(struct hstate *h, 2176f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 2177f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 21780c397daeSMichal Hocko { 21797f325a8dSSidhartha Kumar struct folio *folio; 21807118fc29SMike Kravetz bool retry = false; 21810c397daeSMichal Hocko 21827118fc29SMike Kravetz retry: 21830c397daeSMichal Hocko if (hstate_is_gigantic(h)) 218419fc1a7eSSidhartha Kumar folio = alloc_gigantic_folio(h, gfp_mask, nid, nmask); 21850c397daeSMichal Hocko else 218619fc1a7eSSidhartha Kumar folio = alloc_buddy_hugetlb_folio(h, gfp_mask, 2187f60858f9SMike Kravetz nid, nmask, node_alloc_noretry); 218819fc1a7eSSidhartha Kumar if (!folio) 21890c397daeSMichal Hocko return NULL; 21907118fc29SMike Kravetz if (hstate_is_gigantic(h)) { 2191d1c60955SSidhartha Kumar if (!prep_compound_gigantic_folio(folio, huge_page_order(h))) { 21927118fc29SMike Kravetz /* 21937118fc29SMike Kravetz * Rare failure to convert pages to compound page. 21947118fc29SMike Kravetz * Free pages and try again - ONCE! 21957118fc29SMike Kravetz */ 21967f325a8dSSidhartha Kumar free_gigantic_folio(folio, huge_page_order(h)); 21977118fc29SMike Kravetz if (!retry) { 21987118fc29SMike Kravetz retry = true; 21997118fc29SMike Kravetz goto retry; 22007118fc29SMike Kravetz } 22017118fc29SMike Kravetz return NULL; 22027118fc29SMike Kravetz } 22037118fc29SMike Kravetz } 2204d1c60955SSidhartha Kumar prep_new_hugetlb_folio(h, folio, folio_nid(folio)); 22050c397daeSMichal Hocko 220619fc1a7eSSidhartha Kumar return folio; 22070c397daeSMichal Hocko } 22080c397daeSMichal Hocko 22090c397daeSMichal Hocko /* 2210af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 2211af0fb9dfSMichal Hocko * manner. 2212af0fb9dfSMichal Hocko */ 2213f60858f9SMike Kravetz static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 2214f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 2215b2261026SJoonsoo Kim { 221619fc1a7eSSidhartha Kumar struct folio *folio; 2217b2261026SJoonsoo Kim int nr_nodes, node; 2218af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2219b2261026SJoonsoo Kim 2220b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 222119fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, node, 222219fc1a7eSSidhartha Kumar nodes_allowed, node_alloc_noretry); 222319fc1a7eSSidhartha Kumar if (folio) { 222419fc1a7eSSidhartha Kumar free_huge_page(&folio->page); /* free it into the hugepage allocator */ 222519fc1a7eSSidhartha Kumar return 1; 222619fc1a7eSSidhartha Kumar } 2227b2261026SJoonsoo Kim } 2228b2261026SJoonsoo Kim 2229af0fb9dfSMichal Hocko return 0; 2230b2261026SJoonsoo Kim } 2231b2261026SJoonsoo Kim 2232e8c5c824SLee Schermerhorn /* 223310c6ec49SMike Kravetz * Remove huge page from pool from next node to free. Attempt to keep 223410c6ec49SMike Kravetz * persistent huge pages more or less balanced over allowed nodes. 223510c6ec49SMike Kravetz * This routine only 'removes' the hugetlb page. The caller must make 223610c6ec49SMike Kravetz * an additional call to free the page to low level allocators. 2237e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 2238e8c5c824SLee Schermerhorn */ 223910c6ec49SMike Kravetz static struct page *remove_pool_huge_page(struct hstate *h, 224010c6ec49SMike Kravetz nodemask_t *nodes_allowed, 22416ae11b27SLee Schermerhorn bool acct_surplus) 2242e8c5c824SLee Schermerhorn { 2243b2261026SJoonsoo Kim int nr_nodes, node; 224410c6ec49SMike Kravetz struct page *page = NULL; 2245cfd5082bSSidhartha Kumar struct folio *folio; 2246e8c5c824SLee Schermerhorn 22479487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2248b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2249685f3457SLee Schermerhorn /* 2250685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 2251685f3457SLee Schermerhorn * nodes with surplus pages. 2252685f3457SLee Schermerhorn */ 2253b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 2254b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 225510c6ec49SMike Kravetz page = list_entry(h->hugepage_freelists[node].next, 2256e8c5c824SLee Schermerhorn struct page, lru); 2257cfd5082bSSidhartha Kumar folio = page_folio(page); 2258cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, acct_surplus); 22599a76db09SLee Schermerhorn break; 2260e8c5c824SLee Schermerhorn } 2261b2261026SJoonsoo Kim } 2262e8c5c824SLee Schermerhorn 226310c6ec49SMike Kravetz return page; 2264e8c5c824SLee Schermerhorn } 2265e8c5c824SLee Schermerhorn 2266c8721bbbSNaoya Horiguchi /* 2267c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 2268faf53defSNaoya Horiguchi * nothing for in-use hugepages and non-hugepages. 2269faf53defSNaoya Horiguchi * This function returns values like below: 2270faf53defSNaoya Horiguchi * 2271ad2fa371SMuchun Song * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages 2272ad2fa371SMuchun Song * when the system is under memory pressure and the feature of 2273ad2fa371SMuchun Song * freeing unused vmemmap pages associated with each hugetlb page 2274ad2fa371SMuchun Song * is enabled. 2275faf53defSNaoya Horiguchi * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 2276faf53defSNaoya Horiguchi * (allocated or reserved.) 2277faf53defSNaoya Horiguchi * 0: successfully dissolved free hugepages or the page is not a 2278faf53defSNaoya Horiguchi * hugepage (considered as already dissolved) 2279c8721bbbSNaoya Horiguchi */ 2280c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 2281c8721bbbSNaoya Horiguchi { 22826bc9b564SNaoya Horiguchi int rc = -EBUSY; 22831a7cdab5SSidhartha Kumar struct folio *folio = page_folio(page); 2284082d5b6bSGerald Schaefer 22857ffddd49SMuchun Song retry: 2286faf53defSNaoya Horiguchi /* Not to disrupt normal path by vainly holding hugetlb_lock */ 22871a7cdab5SSidhartha Kumar if (!folio_test_hugetlb(folio)) 2288faf53defSNaoya Horiguchi return 0; 2289faf53defSNaoya Horiguchi 2290db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 22911a7cdab5SSidhartha Kumar if (!folio_test_hugetlb(folio)) { 2292faf53defSNaoya Horiguchi rc = 0; 2293faf53defSNaoya Horiguchi goto out; 2294faf53defSNaoya Horiguchi } 2295faf53defSNaoya Horiguchi 22961a7cdab5SSidhartha Kumar if (!folio_ref_count(folio)) { 22971a7cdab5SSidhartha Kumar struct hstate *h = folio_hstate(folio); 22988346d69dSXin Hao if (!available_huge_pages(h)) 2299082d5b6bSGerald Schaefer goto out; 23007ffddd49SMuchun Song 23017ffddd49SMuchun Song /* 23027ffddd49SMuchun Song * We should make sure that the page is already on the free list 23037ffddd49SMuchun Song * when it is dissolved. 23047ffddd49SMuchun Song */ 23051a7cdab5SSidhartha Kumar if (unlikely(!folio_test_hugetlb_freed(folio))) { 2306db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 23077ffddd49SMuchun Song cond_resched(); 23087ffddd49SMuchun Song 23097ffddd49SMuchun Song /* 23107ffddd49SMuchun Song * Theoretically, we should return -EBUSY when we 23117ffddd49SMuchun Song * encounter this race. In fact, we have a chance 23127ffddd49SMuchun Song * to successfully dissolve the page if we do a 23137ffddd49SMuchun Song * retry. Because the race window is quite small. 23147ffddd49SMuchun Song * If we seize this opportunity, it is an optimization 23157ffddd49SMuchun Song * for increasing the success rate of dissolving page. 23167ffddd49SMuchun Song */ 23177ffddd49SMuchun Song goto retry; 23187ffddd49SMuchun Song } 23197ffddd49SMuchun Song 2320cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, false); 2321ad2fa371SMuchun Song h->max_huge_pages--; 2322ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2323ad2fa371SMuchun Song 2324c3114a84SAnshuman Khandual /* 2325d6ef19e2SSidhartha Kumar * Normally update_and_free_hugtlb_folio will allocate required vmemmmap 2326d6ef19e2SSidhartha Kumar * before freeing the page. update_and_free_hugtlb_folio will fail to 2327ad2fa371SMuchun Song * free the page if it can not allocate required vmemmap. We 2328ad2fa371SMuchun Song * need to adjust max_huge_pages if the page is not freed. 2329ad2fa371SMuchun Song * Attempt to allocate vmemmmap here so that we can take 2330ad2fa371SMuchun Song * appropriate action on failure. 2331ad2fa371SMuchun Song */ 23321a7cdab5SSidhartha Kumar rc = hugetlb_vmemmap_restore(h, &folio->page); 2333ad2fa371SMuchun Song if (!rc) { 2334d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, folio, false); 2335ad2fa371SMuchun Song } else { 2336ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 23372f6c57d6SSidhartha Kumar add_hugetlb_folio(h, folio, false); 2338ad2fa371SMuchun Song h->max_huge_pages++; 2339ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2340ad2fa371SMuchun Song } 2341ad2fa371SMuchun Song 2342ad2fa371SMuchun Song return rc; 2343c8721bbbSNaoya Horiguchi } 2344082d5b6bSGerald Schaefer out: 2345db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2346082d5b6bSGerald Schaefer return rc; 2347c8721bbbSNaoya Horiguchi } 2348c8721bbbSNaoya Horiguchi 2349c8721bbbSNaoya Horiguchi /* 2350c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 2351c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 23522247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 23532247bb33SGerald Schaefer * part of it lies within the given range. 2354082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 2355082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 2356c8721bbbSNaoya Horiguchi */ 2357082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 2358c8721bbbSNaoya Horiguchi { 2359c8721bbbSNaoya Horiguchi unsigned long pfn; 2360eb03aa00SGerald Schaefer struct page *page; 2361082d5b6bSGerald Schaefer int rc = 0; 2362dc2628f3SMuchun Song unsigned int order; 2363dc2628f3SMuchun Song struct hstate *h; 2364c8721bbbSNaoya Horiguchi 2365d0177639SLi Zhong if (!hugepages_supported()) 2366082d5b6bSGerald Schaefer return rc; 2367d0177639SLi Zhong 2368dc2628f3SMuchun Song order = huge_page_order(&default_hstate); 2369dc2628f3SMuchun Song for_each_hstate(h) 2370dc2628f3SMuchun Song order = min(order, huge_page_order(h)); 2371dc2628f3SMuchun Song 2372dc2628f3SMuchun Song for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) { 2373eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 2374eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 2375eb03aa00SGerald Schaefer if (rc) 2376082d5b6bSGerald Schaefer break; 2377eb03aa00SGerald Schaefer } 2378082d5b6bSGerald Schaefer 2379082d5b6bSGerald Schaefer return rc; 2380c8721bbbSNaoya Horiguchi } 2381c8721bbbSNaoya Horiguchi 2382ab5ac90aSMichal Hocko /* 2383ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 2384ab5ac90aSMichal Hocko */ 23853a740e8bSSidhartha Kumar static struct folio *alloc_surplus_hugetlb_folio(struct hstate *h, 23863a740e8bSSidhartha Kumar gfp_t gfp_mask, int nid, nodemask_t *nmask) 23877893d1d5SAdam Litke { 238819fc1a7eSSidhartha Kumar struct folio *folio = NULL; 23897893d1d5SAdam Litke 2390bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2391aa888a74SAndi Kleen return NULL; 2392aa888a74SAndi Kleen 2393db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 23949980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 23959980d744SMichal Hocko goto out_unlock; 2396db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2397d1c3fb1fSNishanth Aravamudan 239819fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL); 239919fc1a7eSSidhartha Kumar if (!folio) 24000c397daeSMichal Hocko return NULL; 2401d1c3fb1fSNishanth Aravamudan 2402db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 24039980d744SMichal Hocko /* 24049980d744SMichal Hocko * We could have raced with the pool size change. 24059980d744SMichal Hocko * Double check that and simply deallocate the new page 24069980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 24079980d744SMichal Hocko * temporary page to workaround the nasty free_huge_page 24089980d744SMichal Hocko * codeflow 24099980d744SMichal Hocko */ 24109980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 241119fc1a7eSSidhartha Kumar folio_set_hugetlb_temporary(folio); 2412db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 241319fc1a7eSSidhartha Kumar free_huge_page(&folio->page); 24142bf753e6SKai Shen return NULL; 2415b65a4edaSMike Kravetz } 2416b65a4edaSMike Kravetz 24179980d744SMichal Hocko h->surplus_huge_pages++; 241819fc1a7eSSidhartha Kumar h->surplus_huge_pages_node[folio_nid(folio)]++; 24199980d744SMichal Hocko 24209980d744SMichal Hocko out_unlock: 2421db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 24227893d1d5SAdam Litke 24233a740e8bSSidhartha Kumar return folio; 24247893d1d5SAdam Litke } 24257893d1d5SAdam Litke 2426e37d3e83SSidhartha Kumar static struct folio *alloc_migrate_hugetlb_folio(struct hstate *h, gfp_t gfp_mask, 2427ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 2428ab5ac90aSMichal Hocko { 242919fc1a7eSSidhartha Kumar struct folio *folio; 2430ab5ac90aSMichal Hocko 2431ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 2432ab5ac90aSMichal Hocko return NULL; 2433ab5ac90aSMichal Hocko 243419fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL); 243519fc1a7eSSidhartha Kumar if (!folio) 2436ab5ac90aSMichal Hocko return NULL; 2437ab5ac90aSMichal Hocko 24382b21624fSMike Kravetz /* fresh huge pages are frozen */ 243919fc1a7eSSidhartha Kumar folio_ref_unfreeze(folio, 1); 2440ab5ac90aSMichal Hocko /* 2441ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 2442ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 2443ab5ac90aSMichal Hocko */ 244419fc1a7eSSidhartha Kumar folio_set_hugetlb_temporary(folio); 2445ab5ac90aSMichal Hocko 2446e37d3e83SSidhartha Kumar return folio; 2447ab5ac90aSMichal Hocko } 2448ab5ac90aSMichal Hocko 2449e4e574b7SAdam Litke /* 2450099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 2451099730d6SDave Hansen */ 2452e0ec90eeSDave Hansen static 2453ff7d853bSSidhartha Kumar struct folio *alloc_buddy_hugetlb_folio_with_mpol(struct hstate *h, 2454099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 2455099730d6SDave Hansen { 24563a740e8bSSidhartha Kumar struct folio *folio = NULL; 2457aaf14e40SMichal Hocko struct mempolicy *mpol; 2458aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 2459aaf14e40SMichal Hocko int nid; 2460aaf14e40SMichal Hocko nodemask_t *nodemask; 2461aaf14e40SMichal Hocko 2462aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 2463cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 2464cfcaa66fSBen Widawsky gfp_t gfp = gfp_mask | __GFP_NOWARN; 2465cfcaa66fSBen Widawsky 2466cfcaa66fSBen Widawsky gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL); 24673a740e8bSSidhartha Kumar folio = alloc_surplus_hugetlb_folio(h, gfp, nid, nodemask); 2468cfcaa66fSBen Widawsky 2469cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 2470cfcaa66fSBen Widawsky nodemask = NULL; 2471cfcaa66fSBen Widawsky } 2472cfcaa66fSBen Widawsky 24733a740e8bSSidhartha Kumar if (!folio) 24743a740e8bSSidhartha Kumar folio = alloc_surplus_hugetlb_folio(h, gfp_mask, nid, nodemask); 2475aaf14e40SMichal Hocko mpol_cond_put(mpol); 2476ff7d853bSSidhartha Kumar return folio; 2477099730d6SDave Hansen } 2478099730d6SDave Hansen 2479e37d3e83SSidhartha Kumar /* folio migration callback function */ 2480e37d3e83SSidhartha Kumar struct folio *alloc_hugetlb_folio_nodemask(struct hstate *h, int preferred_nid, 2481d92bbc27SJoonsoo Kim nodemask_t *nmask, gfp_t gfp_mask) 24824db9b2efSMichal Hocko { 2483db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 24848346d69dSXin Hao if (available_huge_pages(h)) { 2485a36f1e90SSidhartha Kumar struct folio *folio; 24863e59fcb0SMichal Hocko 2487a36f1e90SSidhartha Kumar folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask, 2488a36f1e90SSidhartha Kumar preferred_nid, nmask); 2489a36f1e90SSidhartha Kumar if (folio) { 2490db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2491e37d3e83SSidhartha Kumar return folio; 24924db9b2efSMichal Hocko } 24934db9b2efSMichal Hocko } 2494db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 24954db9b2efSMichal Hocko 2496e37d3e83SSidhartha Kumar return alloc_migrate_hugetlb_folio(h, gfp_mask, preferred_nid, nmask); 24974db9b2efSMichal Hocko } 24984db9b2efSMichal Hocko 2499ebd63723SMichal Hocko /* mempolicy aware migration callback */ 2500d0ce0e47SSidhartha Kumar struct folio *alloc_hugetlb_folio_vma(struct hstate *h, struct vm_area_struct *vma, 2501389c8178SMichal Hocko unsigned long address) 2502ebd63723SMichal Hocko { 2503ebd63723SMichal Hocko struct mempolicy *mpol; 2504ebd63723SMichal Hocko nodemask_t *nodemask; 2505e37d3e83SSidhartha Kumar struct folio *folio; 2506ebd63723SMichal Hocko gfp_t gfp_mask; 2507ebd63723SMichal Hocko int node; 2508ebd63723SMichal Hocko 2509ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 2510ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 2511e37d3e83SSidhartha Kumar folio = alloc_hugetlb_folio_nodemask(h, node, nodemask, gfp_mask); 2512ebd63723SMichal Hocko mpol_cond_put(mpol); 2513ebd63723SMichal Hocko 2514d0ce0e47SSidhartha Kumar return folio; 2515ebd63723SMichal Hocko } 2516ebd63723SMichal Hocko 2517bf50bab2SNaoya Horiguchi /* 251825985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 2519e4e574b7SAdam Litke * of size 'delta'. 2520e4e574b7SAdam Litke */ 25210a4f3d1bSLiu Xiang static int gather_surplus_pages(struct hstate *h, long delta) 25221b2a1e7bSJules Irenge __must_hold(&hugetlb_lock) 2523e4e574b7SAdam Litke { 252434665341SMiaohe Lin LIST_HEAD(surplus_list); 25253a740e8bSSidhartha Kumar struct folio *folio; 2526e4e574b7SAdam Litke struct page *page, *tmp; 25270a4f3d1bSLiu Xiang int ret; 25280a4f3d1bSLiu Xiang long i; 25290a4f3d1bSLiu Xiang long needed, allocated; 253028073b02SHillf Danton bool alloc_ok = true; 2531e4e574b7SAdam Litke 25329487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2533a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 2534ac09b3a1SAdam Litke if (needed <= 0) { 2535a5516438SAndi Kleen h->resv_huge_pages += delta; 2536e4e574b7SAdam Litke return 0; 2537ac09b3a1SAdam Litke } 2538e4e574b7SAdam Litke 2539e4e574b7SAdam Litke allocated = 0; 2540e4e574b7SAdam Litke 2541e4e574b7SAdam Litke ret = -ENOMEM; 2542e4e574b7SAdam Litke retry: 2543db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2544e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 25453a740e8bSSidhartha Kumar folio = alloc_surplus_hugetlb_folio(h, htlb_alloc_mask(h), 25462b21624fSMike Kravetz NUMA_NO_NODE, NULL); 25473a740e8bSSidhartha Kumar if (!folio) { 254828073b02SHillf Danton alloc_ok = false; 254928073b02SHillf Danton break; 255028073b02SHillf Danton } 25513a740e8bSSidhartha Kumar list_add(&folio->lru, &surplus_list); 255269ed779aSDavid Rientjes cond_resched(); 2553e4e574b7SAdam Litke } 255428073b02SHillf Danton allocated += i; 2555e4e574b7SAdam Litke 2556e4e574b7SAdam Litke /* 2557e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 2558e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 2559e4e574b7SAdam Litke */ 2560db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2561a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 2562a5516438SAndi Kleen (h->free_huge_pages + allocated); 256328073b02SHillf Danton if (needed > 0) { 256428073b02SHillf Danton if (alloc_ok) 2565e4e574b7SAdam Litke goto retry; 256628073b02SHillf Danton /* 256728073b02SHillf Danton * We were not able to allocate enough pages to 256828073b02SHillf Danton * satisfy the entire reservation so we free what 256928073b02SHillf Danton * we've allocated so far. 257028073b02SHillf Danton */ 257128073b02SHillf Danton goto free; 257228073b02SHillf Danton } 2573e4e574b7SAdam Litke /* 2574e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 257525985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 2576e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 2577ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 2578ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 2579ac09b3a1SAdam Litke * before they are reserved. 2580e4e574b7SAdam Litke */ 2581e4e574b7SAdam Litke needed += allocated; 2582a5516438SAndi Kleen h->resv_huge_pages += delta; 2583e4e574b7SAdam Litke ret = 0; 2584a9869b83SNaoya Horiguchi 258519fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 258619fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 258719fc3f0aSAdam Litke if ((--needed) < 0) 258819fc3f0aSAdam Litke break; 2589b65a4edaSMike Kravetz /* Add the page to the hugetlb allocator */ 2590240d67a8SSidhartha Kumar enqueue_hugetlb_folio(h, page_folio(page)); 259119fc3f0aSAdam Litke } 259228073b02SHillf Danton free: 2593db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 259419fc3f0aSAdam Litke 2595b65a4edaSMike Kravetz /* 2596b65a4edaSMike Kravetz * Free unnecessary surplus pages to the buddy allocator. 2597b65a4edaSMike Kravetz * Pages have no ref count, call free_huge_page directly. 2598b65a4edaSMike Kravetz */ 2599c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 2600b65a4edaSMike Kravetz free_huge_page(page); 2601db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2602e4e574b7SAdam Litke 2603e4e574b7SAdam Litke return ret; 2604e4e574b7SAdam Litke } 2605e4e574b7SAdam Litke 2606e4e574b7SAdam Litke /* 2607e5bbc8a6SMike Kravetz * This routine has two main purposes: 2608e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 2609e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 2610e5bbc8a6SMike Kravetz * to the associated reservation map. 2611e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 2612e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 2613e4e574b7SAdam Litke */ 2614a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 2615a5516438SAndi Kleen unsigned long unused_resv_pages) 2616e4e574b7SAdam Litke { 2617e4e574b7SAdam Litke unsigned long nr_pages; 261810c6ec49SMike Kravetz struct page *page; 261910c6ec49SMike Kravetz LIST_HEAD(page_list); 262010c6ec49SMike Kravetz 26219487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 262210c6ec49SMike Kravetz /* Uncommit the reservation */ 262310c6ec49SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 2624e4e574b7SAdam Litke 2625c0531714SNaoya Horiguchi if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 2626e5bbc8a6SMike Kravetz goto out; 2627aa888a74SAndi Kleen 2628e5bbc8a6SMike Kravetz /* 2629e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 2630e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 2631e5bbc8a6SMike Kravetz */ 2632a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 2633e4e574b7SAdam Litke 2634685f3457SLee Schermerhorn /* 2635685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 26369b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 26379b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 26389b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 263910c6ec49SMike Kravetz * remove_pool_huge_page() will balance the freed pages across the 26409b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 2641685f3457SLee Schermerhorn */ 2642685f3457SLee Schermerhorn while (nr_pages--) { 264310c6ec49SMike Kravetz page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1); 264410c6ec49SMike Kravetz if (!page) 2645e5bbc8a6SMike Kravetz goto out; 264610c6ec49SMike Kravetz 264710c6ec49SMike Kravetz list_add(&page->lru, &page_list); 2648e4e574b7SAdam Litke } 2649e5bbc8a6SMike Kravetz 2650e5bbc8a6SMike Kravetz out: 2651db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 265210c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 2653db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2654e4e574b7SAdam Litke } 2655e4e574b7SAdam Litke 26565e911373SMike Kravetz 2657c37f9fb1SAndy Whitcroft /* 2658feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 26595e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 2660cf3ad20bSMike Kravetz * 2661cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 2662cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 2663cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 2664cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 2665cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 2666feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 2667feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 2668feba16e2SMike Kravetz * is called in such cases. 2669cf3ad20bSMike Kravetz * 2670cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 2671cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 2672cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 2673cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 2674cf3ad20bSMike Kravetz * take appropriate action. 267596b96a96SMike Kravetz * 267696b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 267796b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 267896b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 267996b96a96SMike Kravetz * reservation exists. 2680846be085SMike Kravetz * 2681846be085SMike Kravetz * vma_del_reservation is used in error paths where an entry in the reserve 2682846be085SMike Kravetz * map was created during huge page allocation and must be removed. It is to 2683846be085SMike Kravetz * be called after calling vma_needs_reservation to determine if a reservation 2684846be085SMike Kravetz * exists. 2685c37f9fb1SAndy Whitcroft */ 26865e911373SMike Kravetz enum vma_resv_mode { 26875e911373SMike Kravetz VMA_NEEDS_RESV, 26885e911373SMike Kravetz VMA_COMMIT_RESV, 2689feba16e2SMike Kravetz VMA_END_RESV, 269096b96a96SMike Kravetz VMA_ADD_RESV, 2691846be085SMike Kravetz VMA_DEL_RESV, 26925e911373SMike Kravetz }; 2693cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 2694cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 26955e911373SMike Kravetz enum vma_resv_mode mode) 2696c37f9fb1SAndy Whitcroft { 26974e35f483SJoonsoo Kim struct resv_map *resv; 26984e35f483SJoonsoo Kim pgoff_t idx; 2699cf3ad20bSMike Kravetz long ret; 27000db9d74eSMina Almasry long dummy_out_regions_needed; 2701c37f9fb1SAndy Whitcroft 27024e35f483SJoonsoo Kim resv = vma_resv_map(vma); 27034e35f483SJoonsoo Kim if (!resv) 2704c37f9fb1SAndy Whitcroft return 1; 2705c37f9fb1SAndy Whitcroft 27064e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 27075e911373SMike Kravetz switch (mode) { 27085e911373SMike Kravetz case VMA_NEEDS_RESV: 27090db9d74eSMina Almasry ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); 27100db9d74eSMina Almasry /* We assume that vma_reservation_* routines always operate on 27110db9d74eSMina Almasry * 1 page, and that adding to resv map a 1 page entry can only 27120db9d74eSMina Almasry * ever require 1 region. 27130db9d74eSMina Almasry */ 27140db9d74eSMina Almasry VM_BUG_ON(dummy_out_regions_needed != 1); 27155e911373SMike Kravetz break; 27165e911373SMike Kravetz case VMA_COMMIT_RESV: 2717075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 27180db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 27190db9d74eSMina Almasry VM_BUG_ON(ret < 0); 27205e911373SMike Kravetz break; 2721feba16e2SMike Kravetz case VMA_END_RESV: 27220db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 27235e911373SMike Kravetz ret = 0; 27245e911373SMike Kravetz break; 272596b96a96SMike Kravetz case VMA_ADD_RESV: 27260db9d74eSMina Almasry if (vma->vm_flags & VM_MAYSHARE) { 2727075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 27280db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 27290db9d74eSMina Almasry VM_BUG_ON(ret < 0); 27300db9d74eSMina Almasry } else { 27310db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 273296b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 273396b96a96SMike Kravetz } 273496b96a96SMike Kravetz break; 2735846be085SMike Kravetz case VMA_DEL_RESV: 2736846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 2737846be085SMike Kravetz region_abort(resv, idx, idx + 1, 1); 2738846be085SMike Kravetz ret = region_del(resv, idx, idx + 1); 2739846be085SMike Kravetz } else { 2740846be085SMike Kravetz ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 2741846be085SMike Kravetz /* region_add calls of range 1 should never fail. */ 2742846be085SMike Kravetz VM_BUG_ON(ret < 0); 2743846be085SMike Kravetz } 2744846be085SMike Kravetz break; 27455e911373SMike Kravetz default: 27465e911373SMike Kravetz BUG(); 27475e911373SMike Kravetz } 274884afd99bSAndy Whitcroft 2749846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV) 2750cf3ad20bSMike Kravetz return ret; 275167961f9dSMike Kravetz /* 2752bf3d12b9SMiaohe Lin * We know private mapping must have HPAGE_RESV_OWNER set. 2753bf3d12b9SMiaohe Lin * 275467961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 275567961f9dSMike Kravetz * However, a file associated with mapping could have been 275667961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 275767961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 275867961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 275967961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 276067961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 276167961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 276267961f9dSMike Kravetz * reservation has already been consumed. As a result, the 276367961f9dSMike Kravetz * return value of this routine is the opposite of the 276467961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 276567961f9dSMike Kravetz */ 2766bf3d12b9SMiaohe Lin if (ret > 0) 276767961f9dSMike Kravetz return 0; 2768bf3d12b9SMiaohe Lin if (ret == 0) 276967961f9dSMike Kravetz return 1; 2770bf3d12b9SMiaohe Lin return ret; 277184afd99bSAndy Whitcroft } 2772cf3ad20bSMike Kravetz 2773cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 2774a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 2775c37f9fb1SAndy Whitcroft { 27765e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 2777cf3ad20bSMike Kravetz } 2778c37f9fb1SAndy Whitcroft 2779cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 2780cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 2781cf3ad20bSMike Kravetz { 27825e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 27835e911373SMike Kravetz } 27845e911373SMike Kravetz 2785feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 27865e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 27875e911373SMike Kravetz { 2788feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 2789c37f9fb1SAndy Whitcroft } 2790c37f9fb1SAndy Whitcroft 279196b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 279296b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 279396b96a96SMike Kravetz { 279496b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 279596b96a96SMike Kravetz } 279696b96a96SMike Kravetz 2797846be085SMike Kravetz static long vma_del_reservation(struct hstate *h, 2798846be085SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 279996b96a96SMike Kravetz { 2800846be085SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV); 2801846be085SMike Kravetz } 2802846be085SMike Kravetz 2803846be085SMike Kravetz /* 2804846be085SMike Kravetz * This routine is called to restore reservation information on error paths. 2805d0ce0e47SSidhartha Kumar * It should ONLY be called for folios allocated via alloc_hugetlb_folio(), 2806d0ce0e47SSidhartha Kumar * and the hugetlb mutex should remain held when calling this routine. 2807846be085SMike Kravetz * 2808846be085SMike Kravetz * It handles two specific cases: 2809d2d7bb44SSidhartha Kumar * 1) A reservation was in place and the folio consumed the reservation. 2810d2d7bb44SSidhartha Kumar * hugetlb_restore_reserve is set in the folio. 2811d2d7bb44SSidhartha Kumar * 2) No reservation was in place for the page, so hugetlb_restore_reserve is 2812d0ce0e47SSidhartha Kumar * not set. However, alloc_hugetlb_folio always updates the reserve map. 2813846be085SMike Kravetz * 2814846be085SMike Kravetz * In case 1, free_huge_page later in the error path will increment the 2815846be085SMike Kravetz * global reserve count. But, free_huge_page does not have enough context 2816846be085SMike Kravetz * to adjust the reservation map. This case deals primarily with private 2817846be085SMike Kravetz * mappings. Adjust the reserve map here to be consistent with global 2818846be085SMike Kravetz * reserve count adjustments to be made by free_huge_page. Make sure the 2819846be085SMike Kravetz * reserve map indicates there is a reservation present. 2820846be085SMike Kravetz * 2821d0ce0e47SSidhartha Kumar * In case 2, simply undo reserve map modifications done by alloc_hugetlb_folio. 2822846be085SMike Kravetz */ 2823846be085SMike Kravetz void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma, 2824d2d7bb44SSidhartha Kumar unsigned long address, struct folio *folio) 2825846be085SMike Kravetz { 282696b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 282796b96a96SMike Kravetz 28280ffdc38eSSidhartha Kumar if (folio_test_hugetlb_restore_reserve(folio)) { 2829846be085SMike Kravetz if (unlikely(rc < 0)) 283096b96a96SMike Kravetz /* 283196b96a96SMike Kravetz * Rare out of memory condition in reserve map 28320ffdc38eSSidhartha Kumar * manipulation. Clear hugetlb_restore_reserve so 28330ffdc38eSSidhartha Kumar * that global reserve count will not be incremented 283496b96a96SMike Kravetz * by free_huge_page. This will make it appear 28350ffdc38eSSidhartha Kumar * as though the reservation for this folio was 283696b96a96SMike Kravetz * consumed. This may prevent the task from 28370ffdc38eSSidhartha Kumar * faulting in the folio at a later time. This 283896b96a96SMike Kravetz * is better than inconsistent global huge page 283996b96a96SMike Kravetz * accounting of reserve counts. 284096b96a96SMike Kravetz */ 28410ffdc38eSSidhartha Kumar folio_clear_hugetlb_restore_reserve(folio); 2842846be085SMike Kravetz else if (rc) 2843846be085SMike Kravetz (void)vma_add_reservation(h, vma, address); 2844846be085SMike Kravetz else 2845846be085SMike Kravetz vma_end_reservation(h, vma, address); 2846846be085SMike Kravetz } else { 2847846be085SMike Kravetz if (!rc) { 284896b96a96SMike Kravetz /* 2849846be085SMike Kravetz * This indicates there is an entry in the reserve map 2850d0ce0e47SSidhartha Kumar * not added by alloc_hugetlb_folio. We know it was added 2851d0ce0e47SSidhartha Kumar * before the alloc_hugetlb_folio call, otherwise 28520ffdc38eSSidhartha Kumar * hugetlb_restore_reserve would be set on the folio. 2853846be085SMike Kravetz * Remove the entry so that a subsequent allocation 2854846be085SMike Kravetz * does not consume a reservation. 285596b96a96SMike Kravetz */ 2856846be085SMike Kravetz rc = vma_del_reservation(h, vma, address); 2857846be085SMike Kravetz if (rc < 0) 2858846be085SMike Kravetz /* 2859846be085SMike Kravetz * VERY rare out of memory condition. Since 2860846be085SMike Kravetz * we can not delete the entry, set 28610ffdc38eSSidhartha Kumar * hugetlb_restore_reserve so that the reserve 28620ffdc38eSSidhartha Kumar * count will be incremented when the folio 2863846be085SMike Kravetz * is freed. This reserve will be consumed 2864846be085SMike Kravetz * on a subsequent allocation. 2865846be085SMike Kravetz */ 28660ffdc38eSSidhartha Kumar folio_set_hugetlb_restore_reserve(folio); 2867846be085SMike Kravetz } else if (rc < 0) { 2868846be085SMike Kravetz /* 2869846be085SMike Kravetz * Rare out of memory condition from 2870846be085SMike Kravetz * vma_needs_reservation call. Memory allocation is 2871846be085SMike Kravetz * only attempted if a new entry is needed. Therefore, 2872846be085SMike Kravetz * this implies there is not an entry in the 2873846be085SMike Kravetz * reserve map. 2874846be085SMike Kravetz * 2875846be085SMike Kravetz * For shared mappings, no entry in the map indicates 2876846be085SMike Kravetz * no reservation. We are done. 2877846be085SMike Kravetz */ 2878846be085SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 2879846be085SMike Kravetz /* 2880846be085SMike Kravetz * For private mappings, no entry indicates 2881846be085SMike Kravetz * a reservation is present. Since we can 28820ffdc38eSSidhartha Kumar * not add an entry, set hugetlb_restore_reserve 28830ffdc38eSSidhartha Kumar * on the folio so reserve count will be 2884846be085SMike Kravetz * incremented when freed. This reserve will 2885846be085SMike Kravetz * be consumed on a subsequent allocation. 2886846be085SMike Kravetz */ 28870ffdc38eSSidhartha Kumar folio_set_hugetlb_restore_reserve(folio); 288896b96a96SMike Kravetz } else 2889846be085SMike Kravetz /* 2890846be085SMike Kravetz * No reservation present, do nothing 2891846be085SMike Kravetz */ 289296b96a96SMike Kravetz vma_end_reservation(h, vma, address); 289396b96a96SMike Kravetz } 289496b96a96SMike Kravetz } 289596b96a96SMike Kravetz 2896369fa227SOscar Salvador /* 289719fc1a7eSSidhartha Kumar * alloc_and_dissolve_hugetlb_folio - Allocate a new folio and dissolve 289819fc1a7eSSidhartha Kumar * the old one 2899369fa227SOscar Salvador * @h: struct hstate old page belongs to 290019fc1a7eSSidhartha Kumar * @old_folio: Old folio to dissolve 2901ae37c7ffSOscar Salvador * @list: List to isolate the page in case we need to 2902369fa227SOscar Salvador * Returns 0 on success, otherwise negated error. 2903369fa227SOscar Salvador */ 290419fc1a7eSSidhartha Kumar static int alloc_and_dissolve_hugetlb_folio(struct hstate *h, 290519fc1a7eSSidhartha Kumar struct folio *old_folio, struct list_head *list) 2906369fa227SOscar Salvador { 2907369fa227SOscar Salvador gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2908de656ed3SSidhartha Kumar int nid = folio_nid(old_folio); 2909de656ed3SSidhartha Kumar struct folio *new_folio; 2910369fa227SOscar Salvador int ret = 0; 2911369fa227SOscar Salvador 2912369fa227SOscar Salvador /* 291319fc1a7eSSidhartha Kumar * Before dissolving the folio, we need to allocate a new one for the 291419fc1a7eSSidhartha Kumar * pool to remain stable. Here, we allocate the folio and 'prep' it 2915f41f2ed4SMuchun Song * by doing everything but actually updating counters and adding to 2916f41f2ed4SMuchun Song * the pool. This simplifies and let us do most of the processing 2917f41f2ed4SMuchun Song * under the lock. 2918369fa227SOscar Salvador */ 291919fc1a7eSSidhartha Kumar new_folio = alloc_buddy_hugetlb_folio(h, gfp_mask, nid, NULL, NULL); 292019fc1a7eSSidhartha Kumar if (!new_folio) 2921369fa227SOscar Salvador return -ENOMEM; 2922de656ed3SSidhartha Kumar __prep_new_hugetlb_folio(h, new_folio); 2923369fa227SOscar Salvador 2924369fa227SOscar Salvador retry: 2925369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2926de656ed3SSidhartha Kumar if (!folio_test_hugetlb(old_folio)) { 2927369fa227SOscar Salvador /* 292819fc1a7eSSidhartha Kumar * Freed from under us. Drop new_folio too. 2929369fa227SOscar Salvador */ 2930369fa227SOscar Salvador goto free_new; 2931de656ed3SSidhartha Kumar } else if (folio_ref_count(old_folio)) { 29329747b9e9SBaolin Wang bool isolated; 29339747b9e9SBaolin Wang 2934369fa227SOscar Salvador /* 293519fc1a7eSSidhartha Kumar * Someone has grabbed the folio, try to isolate it here. 2936ae37c7ffSOscar Salvador * Fail with -EBUSY if not possible. 2937369fa227SOscar Salvador */ 2938ae37c7ffSOscar Salvador spin_unlock_irq(&hugetlb_lock); 29399747b9e9SBaolin Wang isolated = isolate_hugetlb(old_folio, list); 29409747b9e9SBaolin Wang ret = isolated ? 0 : -EBUSY; 2941ae37c7ffSOscar Salvador spin_lock_irq(&hugetlb_lock); 2942369fa227SOscar Salvador goto free_new; 2943de656ed3SSidhartha Kumar } else if (!folio_test_hugetlb_freed(old_folio)) { 2944369fa227SOscar Salvador /* 294519fc1a7eSSidhartha Kumar * Folio's refcount is 0 but it has not been enqueued in the 2946369fa227SOscar Salvador * freelist yet. Race window is small, so we can succeed here if 2947369fa227SOscar Salvador * we retry. 2948369fa227SOscar Salvador */ 2949369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2950369fa227SOscar Salvador cond_resched(); 2951369fa227SOscar Salvador goto retry; 2952369fa227SOscar Salvador } else { 2953369fa227SOscar Salvador /* 295419fc1a7eSSidhartha Kumar * Ok, old_folio is still a genuine free hugepage. Remove it from 2955369fa227SOscar Salvador * the freelist and decrease the counters. These will be 2956369fa227SOscar Salvador * incremented again when calling __prep_account_new_huge_page() 2957240d67a8SSidhartha Kumar * and enqueue_hugetlb_folio() for new_folio. The counters will 2958240d67a8SSidhartha Kumar * remain stable since this happens under the lock. 2959369fa227SOscar Salvador */ 2960cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, old_folio, false); 2961369fa227SOscar Salvador 2962369fa227SOscar Salvador /* 296319fc1a7eSSidhartha Kumar * Ref count on new_folio is already zero as it was dropped 2964b65a4edaSMike Kravetz * earlier. It can be directly added to the pool free list. 2965369fa227SOscar Salvador */ 2966369fa227SOscar Salvador __prep_account_new_huge_page(h, nid); 2967240d67a8SSidhartha Kumar enqueue_hugetlb_folio(h, new_folio); 2968369fa227SOscar Salvador 2969369fa227SOscar Salvador /* 297019fc1a7eSSidhartha Kumar * Folio has been replaced, we can safely free the old one. 2971369fa227SOscar Salvador */ 2972369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2973d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, old_folio, false); 2974369fa227SOscar Salvador } 2975369fa227SOscar Salvador 2976369fa227SOscar Salvador return ret; 2977369fa227SOscar Salvador 2978369fa227SOscar Salvador free_new: 2979369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 298019fc1a7eSSidhartha Kumar /* Folio has a zero ref count, but needs a ref to be freed */ 2981de656ed3SSidhartha Kumar folio_ref_unfreeze(new_folio, 1); 2982d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, new_folio, false); 2983369fa227SOscar Salvador 2984369fa227SOscar Salvador return ret; 2985369fa227SOscar Salvador } 2986369fa227SOscar Salvador 2987ae37c7ffSOscar Salvador int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list) 2988369fa227SOscar Salvador { 2989369fa227SOscar Salvador struct hstate *h; 2990d5e33bd8SSidhartha Kumar struct folio *folio = page_folio(page); 2991ae37c7ffSOscar Salvador int ret = -EBUSY; 2992369fa227SOscar Salvador 2993369fa227SOscar Salvador /* 2994369fa227SOscar Salvador * The page might have been dissolved from under our feet, so make sure 2995369fa227SOscar Salvador * to carefully check the state under the lock. 2996369fa227SOscar Salvador * Return success when racing as if we dissolved the page ourselves. 2997369fa227SOscar Salvador */ 2998369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2999d5e33bd8SSidhartha Kumar if (folio_test_hugetlb(folio)) { 3000d5e33bd8SSidhartha Kumar h = folio_hstate(folio); 3001369fa227SOscar Salvador } else { 3002369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 3003369fa227SOscar Salvador return 0; 3004369fa227SOscar Salvador } 3005369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 3006369fa227SOscar Salvador 3007369fa227SOscar Salvador /* 3008369fa227SOscar Salvador * Fence off gigantic pages as there is a cyclic dependency between 3009369fa227SOscar Salvador * alloc_contig_range and them. Return -ENOMEM as this has the effect 3010369fa227SOscar Salvador * of bailing out right away without further retrying. 3011369fa227SOscar Salvador */ 3012369fa227SOscar Salvador if (hstate_is_gigantic(h)) 3013369fa227SOscar Salvador return -ENOMEM; 3014369fa227SOscar Salvador 30159747b9e9SBaolin Wang if (folio_ref_count(folio) && isolate_hugetlb(folio, list)) 3016ae37c7ffSOscar Salvador ret = 0; 3017d5e33bd8SSidhartha Kumar else if (!folio_ref_count(folio)) 301819fc1a7eSSidhartha Kumar ret = alloc_and_dissolve_hugetlb_folio(h, folio, list); 3019ae37c7ffSOscar Salvador 3020ae37c7ffSOscar Salvador return ret; 3021369fa227SOscar Salvador } 3022369fa227SOscar Salvador 3023d0ce0e47SSidhartha Kumar struct folio *alloc_hugetlb_folio(struct vm_area_struct *vma, 302404f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 3025348ea204SAdam Litke { 302690481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 3027a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3028d4ab0316SSidhartha Kumar struct folio *folio; 3029d85f69b0SMike Kravetz long map_chg, map_commit; 3030d85f69b0SMike Kravetz long gbl_chg; 30316d76dcf4SAneesh Kumar K.V int ret, idx; 3032d0ce0e47SSidhartha Kumar struct hugetlb_cgroup *h_cg = NULL; 303308cf9fafSMina Almasry bool deferred_reserve; 30342fc39cecSAdam Litke 30356d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 3036a1e78772SMel Gorman /* 3037d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 3038d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 3039d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 3040a1e78772SMel Gorman */ 3041d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 3042d85f69b0SMike Kravetz if (map_chg < 0) 304376dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 3044d85f69b0SMike Kravetz 3045d85f69b0SMike Kravetz /* 3046d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 3047d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 3048d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 3049d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 3050d85f69b0SMike Kravetz * checked against any subpool limit. 3051d85f69b0SMike Kravetz */ 3052d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 3053d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 3054d85f69b0SMike Kravetz if (gbl_chg < 0) { 3055feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 305676dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 30575e911373SMike Kravetz } 305890d8b7e6SAdam Litke 3059d85f69b0SMike Kravetz /* 3060d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 3061d85f69b0SMike Kravetz * map, there could be reservations associated with the 3062d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 3063d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 3064d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 3065d85f69b0SMike Kravetz * the subpool reservations. 3066d85f69b0SMike Kravetz */ 3067d85f69b0SMike Kravetz if (avoid_reserve) 3068d85f69b0SMike Kravetz gbl_chg = 1; 3069d85f69b0SMike Kravetz } 3070d85f69b0SMike Kravetz 307108cf9fafSMina Almasry /* If this allocation is not consuming a reservation, charge it now. 307208cf9fafSMina Almasry */ 30736501fe5fSMiaohe Lin deferred_reserve = map_chg || avoid_reserve; 307408cf9fafSMina Almasry if (deferred_reserve) { 307508cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 307608cf9fafSMina Almasry idx, pages_per_huge_page(h), &h_cg); 30778f34af6fSJianyu Zhan if (ret) 30788f34af6fSJianyu Zhan goto out_subpool_put; 307908cf9fafSMina Almasry } 308008cf9fafSMina Almasry 308108cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 308208cf9fafSMina Almasry if (ret) 308308cf9fafSMina Almasry goto out_uncharge_cgroup_reservation; 30848f34af6fSJianyu Zhan 3085db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3086d85f69b0SMike Kravetz /* 3087d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 3088d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 3089d85f69b0SMike Kravetz * a reservation exists for the allocation. 3090d85f69b0SMike Kravetz */ 3091ff7d853bSSidhartha Kumar folio = dequeue_hugetlb_folio_vma(h, vma, addr, avoid_reserve, gbl_chg); 3092ff7d853bSSidhartha Kumar if (!folio) { 3093db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3094ff7d853bSSidhartha Kumar folio = alloc_buddy_hugetlb_folio_with_mpol(h, vma, addr); 3095ff7d853bSSidhartha Kumar if (!folio) 30968f34af6fSJianyu Zhan goto out_uncharge_cgroup; 309712df140fSRik van Riel spin_lock_irq(&hugetlb_lock); 3098a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 3099ff7d853bSSidhartha Kumar folio_set_hugetlb_restore_reserve(folio); 3100a88c7695SNaoya Horiguchi h->resv_huge_pages--; 3101a88c7695SNaoya Horiguchi } 3102ff7d853bSSidhartha Kumar list_add(&folio->lru, &h->hugepage_activelist); 3103ff7d853bSSidhartha Kumar folio_ref_unfreeze(folio, 1); 310481a6fcaeSJoonsoo Kim /* Fall through */ 3105a1e78772SMel Gorman } 3106ff7d853bSSidhartha Kumar 3107ff7d853bSSidhartha Kumar hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, folio); 310808cf9fafSMina Almasry /* If allocation is not consuming a reservation, also store the 310908cf9fafSMina Almasry * hugetlb_cgroup pointer on the page. 311008cf9fafSMina Almasry */ 311108cf9fafSMina Almasry if (deferred_reserve) { 311208cf9fafSMina Almasry hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), 3113ff7d853bSSidhartha Kumar h_cg, folio); 311408cf9fafSMina Almasry } 311508cf9fafSMina Almasry 3116db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3117a1e78772SMel Gorman 3118ff7d853bSSidhartha Kumar hugetlb_set_folio_subpool(folio, spool); 3119a1e78772SMel Gorman 3120d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 3121d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 312233039678SMike Kravetz /* 312333039678SMike Kravetz * The page was added to the reservation map between 312433039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 312533039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 312633039678SMike Kravetz * Adjust for the subpool count incremented above AND 312733039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 312833039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 312933039678SMike Kravetz * no longer applies. 313033039678SMike Kravetz */ 313133039678SMike Kravetz long rsv_adjust; 313233039678SMike Kravetz 313333039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 313433039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 313579aa925bSMike Kravetz if (deferred_reserve) 3136d4ab0316SSidhartha Kumar hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h), 3137d4ab0316SSidhartha Kumar pages_per_huge_page(h), folio); 313833039678SMike Kravetz } 3139d0ce0e47SSidhartha Kumar return folio; 31408f34af6fSJianyu Zhan 31418f34af6fSJianyu Zhan out_uncharge_cgroup: 31428f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 314308cf9fafSMina Almasry out_uncharge_cgroup_reservation: 314408cf9fafSMina Almasry if (deferred_reserve) 314508cf9fafSMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), 314608cf9fafSMina Almasry h_cg); 31478f34af6fSJianyu Zhan out_subpool_put: 3148d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 31498f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 3150feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 31518f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 3152b45b5bd6SDavid Gibson } 3153b45b5bd6SDavid Gibson 3154b5389086SZhenguo Yao int alloc_bootmem_huge_page(struct hstate *h, int nid) 3155e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 3156b5389086SZhenguo Yao int __alloc_bootmem_huge_page(struct hstate *h, int nid) 3157aa888a74SAndi Kleen { 3158b5389086SZhenguo Yao struct huge_bootmem_page *m = NULL; /* initialize for clang */ 3159b2261026SJoonsoo Kim int nr_nodes, node; 3160aa888a74SAndi Kleen 3161b5389086SZhenguo Yao /* do node specific alloc */ 3162b5389086SZhenguo Yao if (nid != NUMA_NO_NODE) { 3163b5389086SZhenguo Yao m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h), 3164b5389086SZhenguo Yao 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid); 3165b5389086SZhenguo Yao if (!m) 3166b5389086SZhenguo Yao return 0; 3167b5389086SZhenguo Yao goto found; 3168b5389086SZhenguo Yao } 3169b5389086SZhenguo Yao /* allocate from next node when distributing huge pages */ 3170b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 3171b5389086SZhenguo Yao m = memblock_alloc_try_nid_raw( 31728b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 317397ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 3174aa888a74SAndi Kleen /* 3175aa888a74SAndi Kleen * Use the beginning of the huge page to store the 3176aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 3177aa888a74SAndi Kleen * puts them into the mem_map). 3178aa888a74SAndi Kleen */ 3179b5389086SZhenguo Yao if (!m) 3180b5389086SZhenguo Yao return 0; 3181aa888a74SAndi Kleen goto found; 3182aa888a74SAndi Kleen } 3183aa888a74SAndi Kleen 3184aa888a74SAndi Kleen found: 3185aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 3186330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 3187aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 3188aa888a74SAndi Kleen m->hstate = h; 3189aa888a74SAndi Kleen return 1; 3190aa888a74SAndi Kleen } 3191aa888a74SAndi Kleen 319248b8d744SMike Kravetz /* 319348b8d744SMike Kravetz * Put bootmem huge pages into the standard lists after mem_map is up. 319448b8d744SMike Kravetz * Note: This only applies to gigantic (order > MAX_ORDER) pages. 319548b8d744SMike Kravetz */ 3196aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 3197aa888a74SAndi Kleen { 3198aa888a74SAndi Kleen struct huge_bootmem_page *m; 3199aa888a74SAndi Kleen 3200aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 320140d18ebfSMike Kravetz struct page *page = virt_to_page(m); 32027f325a8dSSidhartha Kumar struct folio *folio = page_folio(page); 3203aa888a74SAndi Kleen struct hstate *h = m->hstate; 3204ee8f248dSBecky Bruce 320548b8d744SMike Kravetz VM_BUG_ON(!hstate_is_gigantic(h)); 3206d1c60955SSidhartha Kumar WARN_ON(folio_ref_count(folio) != 1); 3207d1c60955SSidhartha Kumar if (prep_compound_gigantic_folio(folio, huge_page_order(h))) { 3208d1c60955SSidhartha Kumar WARN_ON(folio_test_reserved(folio)); 3209d1c60955SSidhartha Kumar prep_new_hugetlb_folio(h, folio, folio_nid(folio)); 32102b21624fSMike Kravetz free_huge_page(page); /* add to the hugepage allocator */ 32117118fc29SMike Kravetz } else { 3212416d85edSMike Kravetz /* VERY unlikely inflated ref count on a tail page */ 32137f325a8dSSidhartha Kumar free_gigantic_folio(folio, huge_page_order(h)); 32147118fc29SMike Kravetz } 3215af0fb9dfSMichal Hocko 3216b0320c7bSRafael Aquini /* 321748b8d744SMike Kravetz * We need to restore the 'stolen' pages to totalram_pages 321848b8d744SMike Kravetz * in order to fix confusing memory reports from free(1) and 321948b8d744SMike Kravetz * other side-effects, like CommitLimit going negative. 3220b0320c7bSRafael Aquini */ 3221c78a7f36SMiaohe Lin adjust_managed_page_count(page, pages_per_huge_page(h)); 3222520495feSCannon Matthews cond_resched(); 3223aa888a74SAndi Kleen } 3224aa888a74SAndi Kleen } 3225b5389086SZhenguo Yao static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid) 3226b5389086SZhenguo Yao { 3227b5389086SZhenguo Yao unsigned long i; 3228b5389086SZhenguo Yao char buf[32]; 3229b5389086SZhenguo Yao 3230b5389086SZhenguo Yao for (i = 0; i < h->max_huge_pages_node[nid]; ++i) { 3231b5389086SZhenguo Yao if (hstate_is_gigantic(h)) { 3232b5389086SZhenguo Yao if (!alloc_bootmem_huge_page(h, nid)) 3233b5389086SZhenguo Yao break; 3234b5389086SZhenguo Yao } else { 323519fc1a7eSSidhartha Kumar struct folio *folio; 3236b5389086SZhenguo Yao gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 3237b5389086SZhenguo Yao 323819fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, 3239b5389086SZhenguo Yao &node_states[N_MEMORY], NULL); 324019fc1a7eSSidhartha Kumar if (!folio) 3241b5389086SZhenguo Yao break; 324219fc1a7eSSidhartha Kumar free_huge_page(&folio->page); /* free it into the hugepage allocator */ 3243b5389086SZhenguo Yao } 3244b5389086SZhenguo Yao cond_resched(); 3245b5389086SZhenguo Yao } 3246b5389086SZhenguo Yao if (i == h->max_huge_pages_node[nid]) 3247b5389086SZhenguo Yao return; 3248b5389086SZhenguo Yao 3249b5389086SZhenguo Yao string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3250b5389086SZhenguo Yao pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n", 3251b5389086SZhenguo Yao h->max_huge_pages_node[nid], buf, nid, i); 3252b5389086SZhenguo Yao h->max_huge_pages -= (h->max_huge_pages_node[nid] - i); 3253b5389086SZhenguo Yao h->max_huge_pages_node[nid] = i; 3254b5389086SZhenguo Yao } 3255aa888a74SAndi Kleen 32568faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 32571da177e4SLinus Torvalds { 32581da177e4SLinus Torvalds unsigned long i; 3259f60858f9SMike Kravetz nodemask_t *node_alloc_noretry; 3260b5389086SZhenguo Yao bool node_specific_alloc = false; 3261f60858f9SMike Kravetz 3262b5389086SZhenguo Yao /* skip gigantic hugepages allocation if hugetlb_cma enabled */ 3263b5389086SZhenguo Yao if (hstate_is_gigantic(h) && hugetlb_cma_size) { 3264b5389086SZhenguo Yao pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); 3265b5389086SZhenguo Yao return; 3266b5389086SZhenguo Yao } 3267b5389086SZhenguo Yao 3268b5389086SZhenguo Yao /* do node specific alloc */ 32690a7a0f6fSPeng Liu for_each_online_node(i) { 3270b5389086SZhenguo Yao if (h->max_huge_pages_node[i] > 0) { 3271b5389086SZhenguo Yao hugetlb_hstate_alloc_pages_onenode(h, i); 3272b5389086SZhenguo Yao node_specific_alloc = true; 3273b5389086SZhenguo Yao } 3274b5389086SZhenguo Yao } 3275b5389086SZhenguo Yao 3276b5389086SZhenguo Yao if (node_specific_alloc) 3277b5389086SZhenguo Yao return; 3278b5389086SZhenguo Yao 3279b5389086SZhenguo Yao /* below will do all node balanced alloc */ 3280f60858f9SMike Kravetz if (!hstate_is_gigantic(h)) { 3281f60858f9SMike Kravetz /* 3282f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3283f60858f9SMike Kravetz * Ignore errors as lower level routines can deal with 3284f60858f9SMike Kravetz * node_alloc_noretry == NULL. If this kmalloc fails at boot 3285f60858f9SMike Kravetz * time, we are likely in bigger trouble. 3286f60858f9SMike Kravetz */ 3287f60858f9SMike Kravetz node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), 3288f60858f9SMike Kravetz GFP_KERNEL); 3289f60858f9SMike Kravetz } else { 3290f60858f9SMike Kravetz /* allocations done at boot time */ 3291f60858f9SMike Kravetz node_alloc_noretry = NULL; 3292f60858f9SMike Kravetz } 3293f60858f9SMike Kravetz 3294f60858f9SMike Kravetz /* bit mask controlling how hard we retry per-node allocations */ 3295f60858f9SMike Kravetz if (node_alloc_noretry) 3296f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 32971da177e4SLinus Torvalds 3298e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 3299bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 3300b5389086SZhenguo Yao if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE)) 3301aa888a74SAndi Kleen break; 33020c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 3303f60858f9SMike Kravetz &node_states[N_MEMORY], 3304f60858f9SMike Kravetz node_alloc_noretry)) 33051da177e4SLinus Torvalds break; 330669ed779aSDavid Rientjes cond_resched(); 33071da177e4SLinus Torvalds } 3308d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 3309d715cf80SLiam R. Howlett char buf[32]; 3310d715cf80SLiam R. Howlett 3311c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3312d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 3313d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 33148faa8b07SAndi Kleen h->max_huge_pages = i; 3315e5ff2159SAndi Kleen } 3316f60858f9SMike Kravetz kfree(node_alloc_noretry); 3317d715cf80SLiam R. Howlett } 3318e5ff2159SAndi Kleen 3319e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 3320e5ff2159SAndi Kleen { 332179dfc695SMike Kravetz struct hstate *h, *h2; 3322e5ff2159SAndi Kleen 3323e5ff2159SAndi Kleen for_each_hstate(h) { 33248faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 3325bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 33268faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 332779dfc695SMike Kravetz 332879dfc695SMike Kravetz /* 332979dfc695SMike Kravetz * Set demote order for each hstate. Note that 333079dfc695SMike Kravetz * h->demote_order is initially 0. 333179dfc695SMike Kravetz * - We can not demote gigantic pages if runtime freeing 333279dfc695SMike Kravetz * is not supported, so skip this. 3333a01f4390SMike Kravetz * - If CMA allocation is possible, we can not demote 3334a01f4390SMike Kravetz * HUGETLB_PAGE_ORDER or smaller size pages. 333579dfc695SMike Kravetz */ 333679dfc695SMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 333779dfc695SMike Kravetz continue; 3338a01f4390SMike Kravetz if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER) 3339a01f4390SMike Kravetz continue; 334079dfc695SMike Kravetz for_each_hstate(h2) { 334179dfc695SMike Kravetz if (h2 == h) 334279dfc695SMike Kravetz continue; 334379dfc695SMike Kravetz if (h2->order < h->order && 334479dfc695SMike Kravetz h2->order > h->demote_order) 334579dfc695SMike Kravetz h->demote_order = h2->order; 334679dfc695SMike Kravetz } 3347e5ff2159SAndi Kleen } 3348e5ff2159SAndi Kleen } 3349e5ff2159SAndi Kleen 3350e5ff2159SAndi Kleen static void __init report_hugepages(void) 3351e5ff2159SAndi Kleen { 3352e5ff2159SAndi Kleen struct hstate *h; 3353e5ff2159SAndi Kleen 3354e5ff2159SAndi Kleen for_each_hstate(h) { 33554abd32dbSAndi Kleen char buf[32]; 3356c6247f72SMatthew Wilcox 3357c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 33586213834cSMuchun Song pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n", 3359c6247f72SMatthew Wilcox buf, h->free_huge_pages); 33606213834cSMuchun Song pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n", 33616213834cSMuchun Song hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf); 3362e5ff2159SAndi Kleen } 3363e5ff2159SAndi Kleen } 3364e5ff2159SAndi Kleen 33651da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 33666ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 33676ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 33681da177e4SLinus Torvalds { 33694415cc8dSChristoph Lameter int i; 33701121828aSMike Kravetz LIST_HEAD(page_list); 33714415cc8dSChristoph Lameter 33729487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 3373bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3374aa888a74SAndi Kleen return; 3375aa888a74SAndi Kleen 33761121828aSMike Kravetz /* 33771121828aSMike Kravetz * Collect pages to be freed on a list, and free after dropping lock 33781121828aSMike Kravetz */ 33796ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 338010c6ec49SMike Kravetz struct page *page, *next; 3381a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 3382a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 3383a5516438SAndi Kleen if (count >= h->nr_huge_pages) 33841121828aSMike Kravetz goto out; 33851da177e4SLinus Torvalds if (PageHighMem(page)) 33861da177e4SLinus Torvalds continue; 3387cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, page_folio(page), false); 33881121828aSMike Kravetz list_add(&page->lru, &page_list); 33891121828aSMike Kravetz } 33901121828aSMike Kravetz } 33911121828aSMike Kravetz 33921121828aSMike Kravetz out: 3393db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 339410c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3395db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 33961da177e4SLinus Torvalds } 33971da177e4SLinus Torvalds #else 33986ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 33996ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 34001da177e4SLinus Torvalds { 34011da177e4SLinus Torvalds } 34021da177e4SLinus Torvalds #endif 34031da177e4SLinus Torvalds 340420a0307cSWu Fengguang /* 340520a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 340620a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 340720a0307cSWu Fengguang * Returns 1 if an adjustment was made. 340820a0307cSWu Fengguang */ 34096ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 34106ae11b27SLee Schermerhorn int delta) 341120a0307cSWu Fengguang { 3412b2261026SJoonsoo Kim int nr_nodes, node; 341320a0307cSWu Fengguang 34149487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 341520a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 341620a0307cSWu Fengguang 3417e8c5c824SLee Schermerhorn if (delta < 0) { 3418b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 3419b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 3420b2261026SJoonsoo Kim goto found; 3421b2261026SJoonsoo Kim } 3422b2261026SJoonsoo Kim } else { 3423b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 3424b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 3425b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 3426b2261026SJoonsoo Kim goto found; 3427e8c5c824SLee Schermerhorn } 34289a76db09SLee Schermerhorn } 3429b2261026SJoonsoo Kim return 0; 343020a0307cSWu Fengguang 3431b2261026SJoonsoo Kim found: 343220a0307cSWu Fengguang h->surplus_huge_pages += delta; 3433b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 3434b2261026SJoonsoo Kim return 1; 343520a0307cSWu Fengguang } 343620a0307cSWu Fengguang 3437a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 3438fd875dcaSMike Kravetz static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 34396ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 34401da177e4SLinus Torvalds { 34417893d1d5SAdam Litke unsigned long min_count, ret; 344210c6ec49SMike Kravetz struct page *page; 344310c6ec49SMike Kravetz LIST_HEAD(page_list); 3444f60858f9SMike Kravetz NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); 3445f60858f9SMike Kravetz 3446f60858f9SMike Kravetz /* 3447f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3448f60858f9SMike Kravetz * If we can not allocate the bit mask, do not attempt to allocate 3449f60858f9SMike Kravetz * the requested huge pages. 3450f60858f9SMike Kravetz */ 3451f60858f9SMike Kravetz if (node_alloc_noretry) 3452f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 3453f60858f9SMike Kravetz else 3454f60858f9SMike Kravetz return -ENOMEM; 34551da177e4SLinus Torvalds 345629383967SMike Kravetz /* 345729383967SMike Kravetz * resize_lock mutex prevents concurrent adjustments to number of 345829383967SMike Kravetz * pages in hstate via the proc/sysfs interfaces. 345929383967SMike Kravetz */ 346029383967SMike Kravetz mutex_lock(&h->resize_lock); 3461b65d4adbSMuchun Song flush_free_hpage_work(h); 3462db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 34634eb0716eSAlexandre Ghiti 34644eb0716eSAlexandre Ghiti /* 3465fd875dcaSMike Kravetz * Check for a node specific request. 3466fd875dcaSMike Kravetz * Changing node specific huge page count may require a corresponding 3467fd875dcaSMike Kravetz * change to the global count. In any case, the passed node mask 3468fd875dcaSMike Kravetz * (nodes_allowed) will restrict alloc/free to the specified node. 3469fd875dcaSMike Kravetz */ 3470fd875dcaSMike Kravetz if (nid != NUMA_NO_NODE) { 3471fd875dcaSMike Kravetz unsigned long old_count = count; 3472fd875dcaSMike Kravetz 3473fd875dcaSMike Kravetz count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 3474fd875dcaSMike Kravetz /* 3475fd875dcaSMike Kravetz * User may have specified a large count value which caused the 3476fd875dcaSMike Kravetz * above calculation to overflow. In this case, they wanted 3477fd875dcaSMike Kravetz * to allocate as many huge pages as possible. Set count to 3478fd875dcaSMike Kravetz * largest possible value to align with their intention. 3479fd875dcaSMike Kravetz */ 3480fd875dcaSMike Kravetz if (count < old_count) 3481fd875dcaSMike Kravetz count = ULONG_MAX; 3482fd875dcaSMike Kravetz } 3483fd875dcaSMike Kravetz 3484fd875dcaSMike Kravetz /* 34854eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 34864eb0716eSAlexandre Ghiti * page range allocation. 34874eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 34884eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 34894eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 34904eb0716eSAlexandre Ghiti */ 34914eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 34924eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 3493db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 349429383967SMike Kravetz mutex_unlock(&h->resize_lock); 3495f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 34964eb0716eSAlexandre Ghiti return -EINVAL; 34974eb0716eSAlexandre Ghiti } 34984eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 34994eb0716eSAlexandre Ghiti } 3500aa888a74SAndi Kleen 35017893d1d5SAdam Litke /* 35027893d1d5SAdam Litke * Increase the pool size 35037893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 35047893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 3505d1c3fb1fSNishanth Aravamudan * 35063a740e8bSSidhartha Kumar * We might race with alloc_surplus_hugetlb_folio() here and be unable 3507d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 3508d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 3509d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 3510d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 35117893d1d5SAdam Litke */ 3512a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 35136ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 35147893d1d5SAdam Litke break; 35157893d1d5SAdam Litke } 35167893d1d5SAdam Litke 3517a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 35187893d1d5SAdam Litke /* 35197893d1d5SAdam Litke * If this allocation races such that we no longer need the 35207893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 35217893d1d5SAdam Litke * and reducing the surplus. 35227893d1d5SAdam Litke */ 3523db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3524649920c6SJia He 3525649920c6SJia He /* yield cpu to avoid soft lockup */ 3526649920c6SJia He cond_resched(); 3527649920c6SJia He 3528f60858f9SMike Kravetz ret = alloc_pool_huge_page(h, nodes_allowed, 3529f60858f9SMike Kravetz node_alloc_noretry); 3530db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 35317893d1d5SAdam Litke if (!ret) 35327893d1d5SAdam Litke goto out; 35337893d1d5SAdam Litke 3534536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 3535536240f2SMel Gorman if (signal_pending(current)) 3536536240f2SMel Gorman goto out; 35377893d1d5SAdam Litke } 35387893d1d5SAdam Litke 35397893d1d5SAdam Litke /* 35407893d1d5SAdam Litke * Decrease the pool size 35417893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 35427893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 35437893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 35447893d1d5SAdam Litke * to the desired size as pages become free. 3545d1c3fb1fSNishanth Aravamudan * 3546d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 3547d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 3548d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 35493a740e8bSSidhartha Kumar * alloc_surplus_hugetlb_folio() is checking the global counter, 3550d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 3551d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 3552d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 35537893d1d5SAdam Litke */ 3554a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 35556b0c880dSAdam Litke min_count = max(count, min_count); 35566ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 355710c6ec49SMike Kravetz 355810c6ec49SMike Kravetz /* 355910c6ec49SMike Kravetz * Collect pages to be removed on list without dropping lock 356010c6ec49SMike Kravetz */ 3561a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 356210c6ec49SMike Kravetz page = remove_pool_huge_page(h, nodes_allowed, 0); 356310c6ec49SMike Kravetz if (!page) 35641da177e4SLinus Torvalds break; 356510c6ec49SMike Kravetz 356610c6ec49SMike Kravetz list_add(&page->lru, &page_list); 35671da177e4SLinus Torvalds } 356810c6ec49SMike Kravetz /* free the pages after dropping lock */ 3569db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 357010c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3571b65d4adbSMuchun Song flush_free_hpage_work(h); 3572db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 357310c6ec49SMike Kravetz 3574a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 35756ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 35767893d1d5SAdam Litke break; 35777893d1d5SAdam Litke } 35787893d1d5SAdam Litke out: 35794eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 3580db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 358129383967SMike Kravetz mutex_unlock(&h->resize_lock); 35824eb0716eSAlexandre Ghiti 3583f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 3584f60858f9SMike Kravetz 35854eb0716eSAlexandre Ghiti return 0; 35861da177e4SLinus Torvalds } 35871da177e4SLinus Torvalds 3588bdd7be07SSidhartha Kumar static int demote_free_hugetlb_folio(struct hstate *h, struct folio *folio) 35898531fc6fSMike Kravetz { 3590bdd7be07SSidhartha Kumar int i, nid = folio_nid(folio); 35918531fc6fSMike Kravetz struct hstate *target_hstate; 359231731452SDoug Berger struct page *subpage; 3593bdd7be07SSidhartha Kumar struct folio *inner_folio; 35948531fc6fSMike Kravetz int rc = 0; 35958531fc6fSMike Kravetz 35968531fc6fSMike Kravetz target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order); 35978531fc6fSMike Kravetz 3598cfd5082bSSidhartha Kumar remove_hugetlb_folio_for_demote(h, folio, false); 35998531fc6fSMike Kravetz spin_unlock_irq(&hugetlb_lock); 36008531fc6fSMike Kravetz 3601bdd7be07SSidhartha Kumar rc = hugetlb_vmemmap_restore(h, &folio->page); 36028531fc6fSMike Kravetz if (rc) { 3603bdd7be07SSidhartha Kumar /* Allocation of vmemmmap failed, we can not demote folio */ 36048531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 3605bdd7be07SSidhartha Kumar folio_ref_unfreeze(folio, 1); 3606bdd7be07SSidhartha Kumar add_hugetlb_folio(h, folio, false); 36078531fc6fSMike Kravetz return rc; 36088531fc6fSMike Kravetz } 36098531fc6fSMike Kravetz 36108531fc6fSMike Kravetz /* 3611911565b8SSidhartha Kumar * Use destroy_compound_hugetlb_folio_for_demote for all huge page 3612bdd7be07SSidhartha Kumar * sizes as it will not ref count folios. 36138531fc6fSMike Kravetz */ 3614911565b8SSidhartha Kumar destroy_compound_hugetlb_folio_for_demote(folio, huge_page_order(h)); 36158531fc6fSMike Kravetz 36168531fc6fSMike Kravetz /* 36178531fc6fSMike Kravetz * Taking target hstate mutex synchronizes with set_max_huge_pages. 36188531fc6fSMike Kravetz * Without the mutex, pages added to target hstate could be marked 36198531fc6fSMike Kravetz * as surplus. 36208531fc6fSMike Kravetz * 36218531fc6fSMike Kravetz * Note that we already hold h->resize_lock. To prevent deadlock, 36228531fc6fSMike Kravetz * use the convention of always taking larger size hstate mutex first. 36238531fc6fSMike Kravetz */ 36248531fc6fSMike Kravetz mutex_lock(&target_hstate->resize_lock); 36258531fc6fSMike Kravetz for (i = 0; i < pages_per_huge_page(h); 36268531fc6fSMike Kravetz i += pages_per_huge_page(target_hstate)) { 3627bdd7be07SSidhartha Kumar subpage = folio_page(folio, i); 3628bdd7be07SSidhartha Kumar inner_folio = page_folio(subpage); 36298531fc6fSMike Kravetz if (hstate_is_gigantic(target_hstate)) 3630bdd7be07SSidhartha Kumar prep_compound_gigantic_folio_for_demote(inner_folio, 36318531fc6fSMike Kravetz target_hstate->order); 36328531fc6fSMike Kravetz else 363331731452SDoug Berger prep_compound_page(subpage, target_hstate->order); 3634bdd7be07SSidhartha Kumar folio_change_private(inner_folio, NULL); 3635bdd7be07SSidhartha Kumar prep_new_hugetlb_folio(target_hstate, inner_folio, nid); 36362b21624fSMike Kravetz free_huge_page(subpage); 36378531fc6fSMike Kravetz } 36388531fc6fSMike Kravetz mutex_unlock(&target_hstate->resize_lock); 36398531fc6fSMike Kravetz 36408531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 36418531fc6fSMike Kravetz 36428531fc6fSMike Kravetz /* 36438531fc6fSMike Kravetz * Not absolutely necessary, but for consistency update max_huge_pages 36448531fc6fSMike Kravetz * based on pool changes for the demoted page. 36458531fc6fSMike Kravetz */ 36468531fc6fSMike Kravetz h->max_huge_pages--; 3647a43a83c7SMiaohe Lin target_hstate->max_huge_pages += 3648a43a83c7SMiaohe Lin pages_per_huge_page(h) / pages_per_huge_page(target_hstate); 36498531fc6fSMike Kravetz 36508531fc6fSMike Kravetz return rc; 36518531fc6fSMike Kravetz } 36528531fc6fSMike Kravetz 365379dfc695SMike Kravetz static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 365479dfc695SMike Kravetz __must_hold(&hugetlb_lock) 365579dfc695SMike Kravetz { 36568531fc6fSMike Kravetz int nr_nodes, node; 3657bdd7be07SSidhartha Kumar struct folio *folio; 365879dfc695SMike Kravetz 365979dfc695SMike Kravetz lockdep_assert_held(&hugetlb_lock); 366079dfc695SMike Kravetz 366179dfc695SMike Kravetz /* We should never get here if no demote order */ 366279dfc695SMike Kravetz if (!h->demote_order) { 366379dfc695SMike Kravetz pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n"); 366479dfc695SMike Kravetz return -EINVAL; /* internal error */ 366579dfc695SMike Kravetz } 366679dfc695SMike Kravetz 36678531fc6fSMike Kravetz for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 3668bdd7be07SSidhartha Kumar list_for_each_entry(folio, &h->hugepage_freelists[node], lru) { 3669bdd7be07SSidhartha Kumar if (folio_test_hwpoison(folio)) 36705a317412SMike Kravetz continue; 3671bdd7be07SSidhartha Kumar return demote_free_hugetlb_folio(h, folio); 36728531fc6fSMike Kravetz } 36738531fc6fSMike Kravetz } 36748531fc6fSMike Kravetz 36755a317412SMike Kravetz /* 36765a317412SMike Kravetz * Only way to get here is if all pages on free lists are poisoned. 36775a317412SMike Kravetz * Return -EBUSY so that caller will not retry. 36785a317412SMike Kravetz */ 36795a317412SMike Kravetz return -EBUSY; 368079dfc695SMike Kravetz } 368179dfc695SMike Kravetz 3682a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 3683a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 3684a3437870SNishanth Aravamudan 368579dfc695SMike Kravetz #define HSTATE_ATTR_WO(_name) \ 368679dfc695SMike Kravetz static struct kobj_attribute _name##_attr = __ATTR_WO(_name) 368779dfc695SMike Kravetz 3688a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 368998bc26acSMiaohe Lin static struct kobj_attribute _name##_attr = __ATTR_RW(_name) 3690a3437870SNishanth Aravamudan 3691a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 3692a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 3693a3437870SNishanth Aravamudan 36949a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 36959a305230SLee Schermerhorn 36969a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 3697a3437870SNishanth Aravamudan { 3698a3437870SNishanth Aravamudan int i; 36999a305230SLee Schermerhorn 3700a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 37019a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 37029a305230SLee Schermerhorn if (nidp) 37039a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 3704a3437870SNishanth Aravamudan return &hstates[i]; 37059a305230SLee Schermerhorn } 37069a305230SLee Schermerhorn 37079a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 3708a3437870SNishanth Aravamudan } 3709a3437870SNishanth Aravamudan 371006808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 3711a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3712a3437870SNishanth Aravamudan { 37139a305230SLee Schermerhorn struct hstate *h; 37149a305230SLee Schermerhorn unsigned long nr_huge_pages; 37159a305230SLee Schermerhorn int nid; 37169a305230SLee Schermerhorn 37179a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 37189a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 37199a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 37209a305230SLee Schermerhorn else 37219a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 37229a305230SLee Schermerhorn 3723ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", nr_huge_pages); 3724a3437870SNishanth Aravamudan } 3725adbe8726SEric B Munson 3726238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 3727238d3c13SDavid Rientjes struct hstate *h, int nid, 3728238d3c13SDavid Rientjes unsigned long count, size_t len) 3729a3437870SNishanth Aravamudan { 3730a3437870SNishanth Aravamudan int err; 37312d0adf7eSOscar Salvador nodemask_t nodes_allowed, *n_mask; 3732a3437870SNishanth Aravamudan 37332d0adf7eSOscar Salvador if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 37342d0adf7eSOscar Salvador return -EINVAL; 3735adbe8726SEric B Munson 37369a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 37379a305230SLee Schermerhorn /* 37389a305230SLee Schermerhorn * global hstate attribute 37399a305230SLee Schermerhorn */ 37409a305230SLee Schermerhorn if (!(obey_mempolicy && 37412d0adf7eSOscar Salvador init_nodemask_of_mempolicy(&nodes_allowed))) 37422d0adf7eSOscar Salvador n_mask = &node_states[N_MEMORY]; 37432d0adf7eSOscar Salvador else 37442d0adf7eSOscar Salvador n_mask = &nodes_allowed; 37452d0adf7eSOscar Salvador } else { 37469a305230SLee Schermerhorn /* 3747fd875dcaSMike Kravetz * Node specific request. count adjustment happens in 3748fd875dcaSMike Kravetz * set_max_huge_pages() after acquiring hugetlb_lock. 37499a305230SLee Schermerhorn */ 37502d0adf7eSOscar Salvador init_nodemask_of_node(&nodes_allowed, nid); 37512d0adf7eSOscar Salvador n_mask = &nodes_allowed; 3752fd875dcaSMike Kravetz } 37539a305230SLee Schermerhorn 37542d0adf7eSOscar Salvador err = set_max_huge_pages(h, count, nid, n_mask); 375506808b08SLee Schermerhorn 37564eb0716eSAlexandre Ghiti return err ? err : len; 375706808b08SLee Schermerhorn } 375806808b08SLee Schermerhorn 3759238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 3760238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 3761238d3c13SDavid Rientjes size_t len) 3762238d3c13SDavid Rientjes { 3763238d3c13SDavid Rientjes struct hstate *h; 3764238d3c13SDavid Rientjes unsigned long count; 3765238d3c13SDavid Rientjes int nid; 3766238d3c13SDavid Rientjes int err; 3767238d3c13SDavid Rientjes 3768238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 3769238d3c13SDavid Rientjes if (err) 3770238d3c13SDavid Rientjes return err; 3771238d3c13SDavid Rientjes 3772238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 3773238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 3774238d3c13SDavid Rientjes } 3775238d3c13SDavid Rientjes 377606808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 377706808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 377806808b08SLee Schermerhorn { 377906808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 378006808b08SLee Schermerhorn } 378106808b08SLee Schermerhorn 378206808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 378306808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 378406808b08SLee Schermerhorn { 3785238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 3786a3437870SNishanth Aravamudan } 3787a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 3788a3437870SNishanth Aravamudan 378906808b08SLee Schermerhorn #ifdef CONFIG_NUMA 379006808b08SLee Schermerhorn 379106808b08SLee Schermerhorn /* 379206808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 379306808b08SLee Schermerhorn * huge page alloc/free. 379406808b08SLee Schermerhorn */ 379506808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 3796ae7a927dSJoe Perches struct kobj_attribute *attr, 3797ae7a927dSJoe Perches char *buf) 379806808b08SLee Schermerhorn { 379906808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 380006808b08SLee Schermerhorn } 380106808b08SLee Schermerhorn 380206808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 380306808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 380406808b08SLee Schermerhorn { 3805238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 380606808b08SLee Schermerhorn } 380706808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 380806808b08SLee Schermerhorn #endif 380906808b08SLee Schermerhorn 381006808b08SLee Schermerhorn 3811a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 3812a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3813a3437870SNishanth Aravamudan { 38149a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3815ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); 3816a3437870SNishanth Aravamudan } 3817adbe8726SEric B Munson 3818a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 3819a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 3820a3437870SNishanth Aravamudan { 3821a3437870SNishanth Aravamudan int err; 3822a3437870SNishanth Aravamudan unsigned long input; 38239a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3824a3437870SNishanth Aravamudan 3825bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3826adbe8726SEric B Munson return -EINVAL; 3827adbe8726SEric B Munson 38283dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 3829a3437870SNishanth Aravamudan if (err) 383073ae31e5SEric B Munson return err; 3831a3437870SNishanth Aravamudan 3832db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3833a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 3834db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3835a3437870SNishanth Aravamudan 3836a3437870SNishanth Aravamudan return count; 3837a3437870SNishanth Aravamudan } 3838a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 3839a3437870SNishanth Aravamudan 3840a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 3841a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3842a3437870SNishanth Aravamudan { 38439a305230SLee Schermerhorn struct hstate *h; 38449a305230SLee Schermerhorn unsigned long free_huge_pages; 38459a305230SLee Schermerhorn int nid; 38469a305230SLee Schermerhorn 38479a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 38489a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 38499a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 38509a305230SLee Schermerhorn else 38519a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 38529a305230SLee Schermerhorn 3853ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", free_huge_pages); 3854a3437870SNishanth Aravamudan } 3855a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 3856a3437870SNishanth Aravamudan 3857a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 3858a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3859a3437870SNishanth Aravamudan { 38609a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3861ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); 3862a3437870SNishanth Aravamudan } 3863a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 3864a3437870SNishanth Aravamudan 3865a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 3866a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3867a3437870SNishanth Aravamudan { 38689a305230SLee Schermerhorn struct hstate *h; 38699a305230SLee Schermerhorn unsigned long surplus_huge_pages; 38709a305230SLee Schermerhorn int nid; 38719a305230SLee Schermerhorn 38729a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 38739a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 38749a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 38759a305230SLee Schermerhorn else 38769a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 38779a305230SLee Schermerhorn 3878ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", surplus_huge_pages); 3879a3437870SNishanth Aravamudan } 3880a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 3881a3437870SNishanth Aravamudan 388279dfc695SMike Kravetz static ssize_t demote_store(struct kobject *kobj, 388379dfc695SMike Kravetz struct kobj_attribute *attr, const char *buf, size_t len) 388479dfc695SMike Kravetz { 388579dfc695SMike Kravetz unsigned long nr_demote; 388679dfc695SMike Kravetz unsigned long nr_available; 388779dfc695SMike Kravetz nodemask_t nodes_allowed, *n_mask; 388879dfc695SMike Kravetz struct hstate *h; 38898eeda55fSLi zeming int err; 389079dfc695SMike Kravetz int nid; 389179dfc695SMike Kravetz 389279dfc695SMike Kravetz err = kstrtoul(buf, 10, &nr_demote); 389379dfc695SMike Kravetz if (err) 389479dfc695SMike Kravetz return err; 389579dfc695SMike Kravetz h = kobj_to_hstate(kobj, &nid); 389679dfc695SMike Kravetz 389779dfc695SMike Kravetz if (nid != NUMA_NO_NODE) { 389879dfc695SMike Kravetz init_nodemask_of_node(&nodes_allowed, nid); 389979dfc695SMike Kravetz n_mask = &nodes_allowed; 390079dfc695SMike Kravetz } else { 390179dfc695SMike Kravetz n_mask = &node_states[N_MEMORY]; 390279dfc695SMike Kravetz } 390379dfc695SMike Kravetz 390479dfc695SMike Kravetz /* Synchronize with other sysfs operations modifying huge pages */ 390579dfc695SMike Kravetz mutex_lock(&h->resize_lock); 390679dfc695SMike Kravetz spin_lock_irq(&hugetlb_lock); 390779dfc695SMike Kravetz 390879dfc695SMike Kravetz while (nr_demote) { 390979dfc695SMike Kravetz /* 391079dfc695SMike Kravetz * Check for available pages to demote each time thorough the 391179dfc695SMike Kravetz * loop as demote_pool_huge_page will drop hugetlb_lock. 391279dfc695SMike Kravetz */ 391379dfc695SMike Kravetz if (nid != NUMA_NO_NODE) 391479dfc695SMike Kravetz nr_available = h->free_huge_pages_node[nid]; 391579dfc695SMike Kravetz else 391679dfc695SMike Kravetz nr_available = h->free_huge_pages; 391779dfc695SMike Kravetz nr_available -= h->resv_huge_pages; 391879dfc695SMike Kravetz if (!nr_available) 391979dfc695SMike Kravetz break; 392079dfc695SMike Kravetz 392179dfc695SMike Kravetz err = demote_pool_huge_page(h, n_mask); 392279dfc695SMike Kravetz if (err) 392379dfc695SMike Kravetz break; 392479dfc695SMike Kravetz 392579dfc695SMike Kravetz nr_demote--; 392679dfc695SMike Kravetz } 392779dfc695SMike Kravetz 392879dfc695SMike Kravetz spin_unlock_irq(&hugetlb_lock); 392979dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 393079dfc695SMike Kravetz 393179dfc695SMike Kravetz if (err) 393279dfc695SMike Kravetz return err; 393379dfc695SMike Kravetz return len; 393479dfc695SMike Kravetz } 393579dfc695SMike Kravetz HSTATE_ATTR_WO(demote); 393679dfc695SMike Kravetz 393779dfc695SMike Kravetz static ssize_t demote_size_show(struct kobject *kobj, 393879dfc695SMike Kravetz struct kobj_attribute *attr, char *buf) 393979dfc695SMike Kravetz { 394012658abfSMiaohe Lin struct hstate *h = kobj_to_hstate(kobj, NULL); 394179dfc695SMike Kravetz unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K; 394279dfc695SMike Kravetz 394379dfc695SMike Kravetz return sysfs_emit(buf, "%lukB\n", demote_size); 394479dfc695SMike Kravetz } 394579dfc695SMike Kravetz 394679dfc695SMike Kravetz static ssize_t demote_size_store(struct kobject *kobj, 394779dfc695SMike Kravetz struct kobj_attribute *attr, 394879dfc695SMike Kravetz const char *buf, size_t count) 394979dfc695SMike Kravetz { 395079dfc695SMike Kravetz struct hstate *h, *demote_hstate; 395179dfc695SMike Kravetz unsigned long demote_size; 395279dfc695SMike Kravetz unsigned int demote_order; 395379dfc695SMike Kravetz 395479dfc695SMike Kravetz demote_size = (unsigned long)memparse(buf, NULL); 395579dfc695SMike Kravetz 395679dfc695SMike Kravetz demote_hstate = size_to_hstate(demote_size); 395779dfc695SMike Kravetz if (!demote_hstate) 395879dfc695SMike Kravetz return -EINVAL; 395979dfc695SMike Kravetz demote_order = demote_hstate->order; 3960a01f4390SMike Kravetz if (demote_order < HUGETLB_PAGE_ORDER) 3961a01f4390SMike Kravetz return -EINVAL; 396279dfc695SMike Kravetz 396379dfc695SMike Kravetz /* demote order must be smaller than hstate order */ 396412658abfSMiaohe Lin h = kobj_to_hstate(kobj, NULL); 396579dfc695SMike Kravetz if (demote_order >= h->order) 396679dfc695SMike Kravetz return -EINVAL; 396779dfc695SMike Kravetz 396879dfc695SMike Kravetz /* resize_lock synchronizes access to demote size and writes */ 396979dfc695SMike Kravetz mutex_lock(&h->resize_lock); 397079dfc695SMike Kravetz h->demote_order = demote_order; 397179dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 397279dfc695SMike Kravetz 397379dfc695SMike Kravetz return count; 397479dfc695SMike Kravetz } 397579dfc695SMike Kravetz HSTATE_ATTR(demote_size); 397679dfc695SMike Kravetz 3977a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 3978a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 3979a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 3980a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 3981a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 3982a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 398306808b08SLee Schermerhorn #ifdef CONFIG_NUMA 398406808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 398506808b08SLee Schermerhorn #endif 3986a3437870SNishanth Aravamudan NULL, 3987a3437870SNishanth Aravamudan }; 3988a3437870SNishanth Aravamudan 398967e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 3990a3437870SNishanth Aravamudan .attrs = hstate_attrs, 3991a3437870SNishanth Aravamudan }; 3992a3437870SNishanth Aravamudan 399379dfc695SMike Kravetz static struct attribute *hstate_demote_attrs[] = { 399479dfc695SMike Kravetz &demote_size_attr.attr, 399579dfc695SMike Kravetz &demote_attr.attr, 399679dfc695SMike Kravetz NULL, 399779dfc695SMike Kravetz }; 399879dfc695SMike Kravetz 399979dfc695SMike Kravetz static const struct attribute_group hstate_demote_attr_group = { 400079dfc695SMike Kravetz .attrs = hstate_demote_attrs, 400179dfc695SMike Kravetz }; 400279dfc695SMike Kravetz 4003094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 40049a305230SLee Schermerhorn struct kobject **hstate_kobjs, 400567e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 4006a3437870SNishanth Aravamudan { 4007a3437870SNishanth Aravamudan int retval; 4008972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 4009a3437870SNishanth Aravamudan 40109a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 40119a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 4012a3437870SNishanth Aravamudan return -ENOMEM; 4013a3437870SNishanth Aravamudan 40149a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 4015cc2205a6SMiaohe Lin if (retval) { 40169a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 4017cc2205a6SMiaohe Lin hstate_kobjs[hi] = NULL; 40183a6bdda0SMiaohe Lin return retval; 4019cc2205a6SMiaohe Lin } 4020a3437870SNishanth Aravamudan 402179dfc695SMike Kravetz if (h->demote_order) { 402201088a60SMiaohe Lin retval = sysfs_create_group(hstate_kobjs[hi], 402301088a60SMiaohe Lin &hstate_demote_attr_group); 402401088a60SMiaohe Lin if (retval) { 402579dfc695SMike Kravetz pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name); 402601088a60SMiaohe Lin sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group); 402701088a60SMiaohe Lin kobject_put(hstate_kobjs[hi]); 402801088a60SMiaohe Lin hstate_kobjs[hi] = NULL; 402901088a60SMiaohe Lin return retval; 403001088a60SMiaohe Lin } 403179dfc695SMike Kravetz } 403279dfc695SMike Kravetz 403301088a60SMiaohe Lin return 0; 4034a3437870SNishanth Aravamudan } 4035a3437870SNishanth Aravamudan 40369a305230SLee Schermerhorn #ifdef CONFIG_NUMA 4037a4a00b45SMuchun Song static bool hugetlb_sysfs_initialized __ro_after_init; 40389a305230SLee Schermerhorn 40399a305230SLee Schermerhorn /* 40409a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 404110fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 404210fbcf4cSKay Sievers * index of a node device or _hstate == node id. 404310fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 40449a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 40459a305230SLee Schermerhorn */ 40469a305230SLee Schermerhorn struct node_hstate { 40479a305230SLee Schermerhorn struct kobject *hugepages_kobj; 40489a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 40499a305230SLee Schermerhorn }; 4050b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 40519a305230SLee Schermerhorn 40529a305230SLee Schermerhorn /* 405310fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 40549a305230SLee Schermerhorn */ 40559a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 40569a305230SLee Schermerhorn &nr_hugepages_attr.attr, 40579a305230SLee Schermerhorn &free_hugepages_attr.attr, 40589a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 40599a305230SLee Schermerhorn NULL, 40609a305230SLee Schermerhorn }; 40619a305230SLee Schermerhorn 406267e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 40639a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 40649a305230SLee Schermerhorn }; 40659a305230SLee Schermerhorn 40669a305230SLee Schermerhorn /* 406710fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 40689a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 40699a305230SLee Schermerhorn */ 40709a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 40719a305230SLee Schermerhorn { 40729a305230SLee Schermerhorn int nid; 40739a305230SLee Schermerhorn 40749a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 40759a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 40769a305230SLee Schermerhorn int i; 40779a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 40789a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 40799a305230SLee Schermerhorn if (nidp) 40809a305230SLee Schermerhorn *nidp = nid; 40819a305230SLee Schermerhorn return &hstates[i]; 40829a305230SLee Schermerhorn } 40839a305230SLee Schermerhorn } 40849a305230SLee Schermerhorn 40859a305230SLee Schermerhorn BUG(); 40869a305230SLee Schermerhorn return NULL; 40879a305230SLee Schermerhorn } 40889a305230SLee Schermerhorn 40899a305230SLee Schermerhorn /* 409010fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 40919a305230SLee Schermerhorn * No-op if no hstate attributes attached. 40929a305230SLee Schermerhorn */ 4093a4a00b45SMuchun Song void hugetlb_unregister_node(struct node *node) 40949a305230SLee Schermerhorn { 40959a305230SLee Schermerhorn struct hstate *h; 409610fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 40979a305230SLee Schermerhorn 40989a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 40999b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 41009a305230SLee Schermerhorn 4101972dc4deSAneesh Kumar K.V for_each_hstate(h) { 4102972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 410301088a60SMiaohe Lin struct kobject *hstate_kobj = nhs->hstate_kobjs[idx]; 410401088a60SMiaohe Lin 410501088a60SMiaohe Lin if (!hstate_kobj) 410601088a60SMiaohe Lin continue; 410701088a60SMiaohe Lin if (h->demote_order) 410801088a60SMiaohe Lin sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group); 410901088a60SMiaohe Lin sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group); 411001088a60SMiaohe Lin kobject_put(hstate_kobj); 4111972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 4112972dc4deSAneesh Kumar K.V } 41139a305230SLee Schermerhorn 41149a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 41159a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 41169a305230SLee Schermerhorn } 41179a305230SLee Schermerhorn 41189a305230SLee Schermerhorn 41199a305230SLee Schermerhorn /* 412010fbcf4cSKay Sievers * Register hstate attributes for a single node device. 41219a305230SLee Schermerhorn * No-op if attributes already registered. 41229a305230SLee Schermerhorn */ 4123a4a00b45SMuchun Song void hugetlb_register_node(struct node *node) 41249a305230SLee Schermerhorn { 41259a305230SLee Schermerhorn struct hstate *h; 412610fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 41279a305230SLee Schermerhorn int err; 41289a305230SLee Schermerhorn 4129a4a00b45SMuchun Song if (!hugetlb_sysfs_initialized) 4130a4a00b45SMuchun Song return; 4131a4a00b45SMuchun Song 41329a305230SLee Schermerhorn if (nhs->hugepages_kobj) 41339a305230SLee Schermerhorn return; /* already allocated */ 41349a305230SLee Schermerhorn 41359a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 413610fbcf4cSKay Sievers &node->dev.kobj); 41379a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 41389a305230SLee Schermerhorn return; 41399a305230SLee Schermerhorn 41409a305230SLee Schermerhorn for_each_hstate(h) { 41419a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 41429a305230SLee Schermerhorn nhs->hstate_kobjs, 41439a305230SLee Schermerhorn &per_node_hstate_attr_group); 41449a305230SLee Schermerhorn if (err) { 4145282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s for node %d\n", 414610fbcf4cSKay Sievers h->name, node->dev.id); 41479a305230SLee Schermerhorn hugetlb_unregister_node(node); 41489a305230SLee Schermerhorn break; 41499a305230SLee Schermerhorn } 41509a305230SLee Schermerhorn } 41519a305230SLee Schermerhorn } 41529a305230SLee Schermerhorn 41539a305230SLee Schermerhorn /* 41549b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 415510fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 415610fbcf4cSKay Sievers * registered their associated device by this time. 41579a305230SLee Schermerhorn */ 41587d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 41599a305230SLee Schermerhorn { 41609a305230SLee Schermerhorn int nid; 41619a305230SLee Schermerhorn 4162a4a00b45SMuchun Song for_each_online_node(nid) 4163b958d4d0SMuchun Song hugetlb_register_node(node_devices[nid]); 41649a305230SLee Schermerhorn } 41659a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 41669a305230SLee Schermerhorn 41679a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 41689a305230SLee Schermerhorn { 41699a305230SLee Schermerhorn BUG(); 41709a305230SLee Schermerhorn if (nidp) 41719a305230SLee Schermerhorn *nidp = -1; 41729a305230SLee Schermerhorn return NULL; 41739a305230SLee Schermerhorn } 41749a305230SLee Schermerhorn 41759a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 41769a305230SLee Schermerhorn 41779a305230SLee Schermerhorn #endif 41789a305230SLee Schermerhorn 4179263b8998SMiaohe Lin #ifdef CONFIG_CMA 4180263b8998SMiaohe Lin static void __init hugetlb_cma_check(void); 4181263b8998SMiaohe Lin #else 4182263b8998SMiaohe Lin static inline __init void hugetlb_cma_check(void) 4183263b8998SMiaohe Lin { 4184263b8998SMiaohe Lin } 4185263b8998SMiaohe Lin #endif 4186263b8998SMiaohe Lin 4187a4a00b45SMuchun Song static void __init hugetlb_sysfs_init(void) 4188a4a00b45SMuchun Song { 4189a4a00b45SMuchun Song struct hstate *h; 4190a4a00b45SMuchun Song int err; 4191a4a00b45SMuchun Song 4192a4a00b45SMuchun Song hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 4193a4a00b45SMuchun Song if (!hugepages_kobj) 4194a4a00b45SMuchun Song return; 4195a4a00b45SMuchun Song 4196a4a00b45SMuchun Song for_each_hstate(h) { 4197a4a00b45SMuchun Song err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 4198a4a00b45SMuchun Song hstate_kobjs, &hstate_attr_group); 4199a4a00b45SMuchun Song if (err) 4200a4a00b45SMuchun Song pr_err("HugeTLB: Unable to add hstate %s", h->name); 4201a4a00b45SMuchun Song } 4202a4a00b45SMuchun Song 4203a4a00b45SMuchun Song #ifdef CONFIG_NUMA 4204a4a00b45SMuchun Song hugetlb_sysfs_initialized = true; 4205a4a00b45SMuchun Song #endif 4206a4a00b45SMuchun Song hugetlb_register_all_nodes(); 4207a4a00b45SMuchun Song } 4208a4a00b45SMuchun Song 4209962de548SKefeng Wang #ifdef CONFIG_SYSCTL 4210962de548SKefeng Wang static void hugetlb_sysctl_init(void); 4211962de548SKefeng Wang #else 4212962de548SKefeng Wang static inline void hugetlb_sysctl_init(void) { } 4213962de548SKefeng Wang #endif 4214962de548SKefeng Wang 4215a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 4216a3437870SNishanth Aravamudan { 42178382d914SDavidlohr Bueso int i; 42188382d914SDavidlohr Bueso 4219d6995da3SMike Kravetz BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < 4220d6995da3SMike Kravetz __NR_HPAGEFLAGS); 4221d6995da3SMike Kravetz 4222c2833a5bSMike Kravetz if (!hugepages_supported()) { 4223c2833a5bSMike Kravetz if (hugetlb_max_hstate || default_hstate_max_huge_pages) 4224c2833a5bSMike Kravetz pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); 42250ef89d25SBenjamin Herrenschmidt return 0; 4226d715cf80SLiam R. Howlett } 4227d715cf80SLiam R. Howlett 4228282f4214SMike Kravetz /* 4229282f4214SMike Kravetz * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some 4230282f4214SMike Kravetz * architectures depend on setup being done here. 4231282f4214SMike Kravetz */ 4232a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 4233282f4214SMike Kravetz if (!parsed_default_hugepagesz) { 4234282f4214SMike Kravetz /* 4235282f4214SMike Kravetz * If we did not parse a default huge page size, set 4236282f4214SMike Kravetz * default_hstate_idx to HPAGE_SIZE hstate. And, if the 4237282f4214SMike Kravetz * number of huge pages for this default size was implicitly 4238282f4214SMike Kravetz * specified, set that here as well. 4239282f4214SMike Kravetz * Note that the implicit setting will overwrite an explicit 4240282f4214SMike Kravetz * setting. A warning will be printed in this case. 4241282f4214SMike Kravetz */ 4242282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); 4243f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 4244282f4214SMike Kravetz if (default_hstate.max_huge_pages) { 4245282f4214SMike Kravetz char buf[32]; 4246282f4214SMike Kravetz 4247282f4214SMike Kravetz string_get_size(huge_page_size(&default_hstate), 4248282f4214SMike Kravetz 1, STRING_UNITS_2, buf, 32); 4249282f4214SMike Kravetz pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", 4250282f4214SMike Kravetz default_hstate.max_huge_pages, buf); 4251282f4214SMike Kravetz pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", 4252282f4214SMike Kravetz default_hstate_max_huge_pages); 4253282f4214SMike Kravetz } 4254282f4214SMike Kravetz default_hstate.max_huge_pages = 4255282f4214SMike Kravetz default_hstate_max_huge_pages; 4256b5389086SZhenguo Yao 42570a7a0f6fSPeng Liu for_each_online_node(i) 4258b5389086SZhenguo Yao default_hstate.max_huge_pages_node[i] = 4259b5389086SZhenguo Yao default_hugepages_in_node[i]; 4260282f4214SMike Kravetz } 4261f8b74815SVaishali Thakkar } 4262a3437870SNishanth Aravamudan 4263cf11e85fSRoman Gushchin hugetlb_cma_check(); 4264a3437870SNishanth Aravamudan hugetlb_init_hstates(); 4265aa888a74SAndi Kleen gather_bootmem_prealloc(); 4266a3437870SNishanth Aravamudan report_hugepages(); 4267a3437870SNishanth Aravamudan 4268a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 42697179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 4270962de548SKefeng Wang hugetlb_sysctl_init(); 42719a305230SLee Schermerhorn 42728382d914SDavidlohr Bueso #ifdef CONFIG_SMP 42738382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 42748382d914SDavidlohr Bueso #else 42758382d914SDavidlohr Bueso num_fault_mutexes = 1; 42768382d914SDavidlohr Bueso #endif 4277c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 42786da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 42796da2ec56SKees Cook GFP_KERNEL); 4280c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 42818382d914SDavidlohr Bueso 42828382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 4283c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 4284a3437870SNishanth Aravamudan return 0; 4285a3437870SNishanth Aravamudan } 42863e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 4287a3437870SNishanth Aravamudan 4288ae94da89SMike Kravetz /* Overwritten by architectures with more huge page sizes */ 4289ae94da89SMike Kravetz bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) 42909fee021dSVaishali Thakkar { 4291ae94da89SMike Kravetz return size == HPAGE_SIZE; 42929fee021dSVaishali Thakkar } 42939fee021dSVaishali Thakkar 4294d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 4295a3437870SNishanth Aravamudan { 4296a3437870SNishanth Aravamudan struct hstate *h; 42978faa8b07SAndi Kleen unsigned long i; 42988faa8b07SAndi Kleen 4299a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 4300a3437870SNishanth Aravamudan return; 4301a3437870SNishanth Aravamudan } 430247d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 4303a3437870SNishanth Aravamudan BUG_ON(order == 0); 430447d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 430529383967SMike Kravetz mutex_init(&h->resize_lock); 4306a3437870SNishanth Aravamudan h->order = order; 4307aca78307SMiaohe Lin h->mask = ~(huge_page_size(h) - 1); 43088faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 43098faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 43100edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 431154f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 431254f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 4313a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 4314c2c3a60aSMiaohe Lin huge_page_size(h)/SZ_1K); 43158faa8b07SAndi Kleen 4316a3437870SNishanth Aravamudan parsed_hstate = h; 4317a3437870SNishanth Aravamudan } 4318a3437870SNishanth Aravamudan 4319b5389086SZhenguo Yao bool __init __weak hugetlb_node_alloc_supported(void) 4320b5389086SZhenguo Yao { 4321b5389086SZhenguo Yao return true; 4322b5389086SZhenguo Yao } 4323f87442f4SPeng Liu 4324f87442f4SPeng Liu static void __init hugepages_clear_pages_in_node(void) 4325f87442f4SPeng Liu { 4326f87442f4SPeng Liu if (!hugetlb_max_hstate) { 4327f87442f4SPeng Liu default_hstate_max_huge_pages = 0; 4328f87442f4SPeng Liu memset(default_hugepages_in_node, 0, 432910395680SMiaohe Lin sizeof(default_hugepages_in_node)); 4330f87442f4SPeng Liu } else { 4331f87442f4SPeng Liu parsed_hstate->max_huge_pages = 0; 4332f87442f4SPeng Liu memset(parsed_hstate->max_huge_pages_node, 0, 433310395680SMiaohe Lin sizeof(parsed_hstate->max_huge_pages_node)); 4334f87442f4SPeng Liu } 4335f87442f4SPeng Liu } 4336f87442f4SPeng Liu 4337282f4214SMike Kravetz /* 4338282f4214SMike Kravetz * hugepages command line processing 4339282f4214SMike Kravetz * hugepages normally follows a valid hugepagsz or default_hugepagsz 4340282f4214SMike Kravetz * specification. If not, ignore the hugepages value. hugepages can also 4341282f4214SMike Kravetz * be the first huge page command line option in which case it implicitly 4342282f4214SMike Kravetz * specifies the number of huge pages for the default size. 4343282f4214SMike Kravetz */ 4344282f4214SMike Kravetz static int __init hugepages_setup(char *s) 4345a3437870SNishanth Aravamudan { 4346a3437870SNishanth Aravamudan unsigned long *mhp; 43478faa8b07SAndi Kleen static unsigned long *last_mhp; 4348b5389086SZhenguo Yao int node = NUMA_NO_NODE; 4349b5389086SZhenguo Yao int count; 4350b5389086SZhenguo Yao unsigned long tmp; 4351b5389086SZhenguo Yao char *p = s; 4352a3437870SNishanth Aravamudan 43539fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 4354282f4214SMike Kravetz pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); 43559fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 4356f81f6e4bSPeng Liu return 1; 43579fee021dSVaishali Thakkar } 4358282f4214SMike Kravetz 4359a3437870SNishanth Aravamudan /* 4360282f4214SMike Kravetz * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter 4361282f4214SMike Kravetz * yet, so this hugepages= parameter goes to the "default hstate". 4362282f4214SMike Kravetz * Otherwise, it goes with the previously parsed hugepagesz or 4363282f4214SMike Kravetz * default_hugepagesz. 4364a3437870SNishanth Aravamudan */ 43659fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 4366a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 4367a3437870SNishanth Aravamudan else 4368a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 4369a3437870SNishanth Aravamudan 43708faa8b07SAndi Kleen if (mhp == last_mhp) { 4371282f4214SMike Kravetz pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); 4372f81f6e4bSPeng Liu return 1; 43738faa8b07SAndi Kleen } 43748faa8b07SAndi Kleen 4375b5389086SZhenguo Yao while (*p) { 4376b5389086SZhenguo Yao count = 0; 4377b5389086SZhenguo Yao if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4378b5389086SZhenguo Yao goto invalid; 4379b5389086SZhenguo Yao /* Parameter is node format */ 4380b5389086SZhenguo Yao if (p[count] == ':') { 4381b5389086SZhenguo Yao if (!hugetlb_node_alloc_supported()) { 4382b5389086SZhenguo Yao pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n"); 4383f81f6e4bSPeng Liu return 1; 4384b5389086SZhenguo Yao } 43850a7a0f6fSPeng Liu if (tmp >= MAX_NUMNODES || !node_online(tmp)) 4386e79ce983SLiu Yuntao goto invalid; 43870a7a0f6fSPeng Liu node = array_index_nospec(tmp, MAX_NUMNODES); 4388b5389086SZhenguo Yao p += count + 1; 4389b5389086SZhenguo Yao /* Parse hugepages */ 4390b5389086SZhenguo Yao if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4391b5389086SZhenguo Yao goto invalid; 4392b5389086SZhenguo Yao if (!hugetlb_max_hstate) 4393b5389086SZhenguo Yao default_hugepages_in_node[node] = tmp; 4394b5389086SZhenguo Yao else 4395b5389086SZhenguo Yao parsed_hstate->max_huge_pages_node[node] = tmp; 4396b5389086SZhenguo Yao *mhp += tmp; 4397b5389086SZhenguo Yao /* Go to parse next node*/ 4398b5389086SZhenguo Yao if (p[count] == ',') 4399b5389086SZhenguo Yao p += count + 1; 4400b5389086SZhenguo Yao else 4401b5389086SZhenguo Yao break; 4402b5389086SZhenguo Yao } else { 4403b5389086SZhenguo Yao if (p != s) 4404b5389086SZhenguo Yao goto invalid; 4405b5389086SZhenguo Yao *mhp = tmp; 4406b5389086SZhenguo Yao break; 4407b5389086SZhenguo Yao } 4408b5389086SZhenguo Yao } 4409a3437870SNishanth Aravamudan 44108faa8b07SAndi Kleen /* 44118faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 441204adbc3fSMiaohe Lin * But we need to allocate gigantic hstates here early to still 44138faa8b07SAndi Kleen * use the bootmem allocator. 44148faa8b07SAndi Kleen */ 441504adbc3fSMiaohe Lin if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) 44168faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 44178faa8b07SAndi Kleen 44188faa8b07SAndi Kleen last_mhp = mhp; 44198faa8b07SAndi Kleen 4420a3437870SNishanth Aravamudan return 1; 4421b5389086SZhenguo Yao 4422b5389086SZhenguo Yao invalid: 4423b5389086SZhenguo Yao pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p); 4424f87442f4SPeng Liu hugepages_clear_pages_in_node(); 4425f81f6e4bSPeng Liu return 1; 4426a3437870SNishanth Aravamudan } 4427282f4214SMike Kravetz __setup("hugepages=", hugepages_setup); 4428e11bfbfcSNick Piggin 4429282f4214SMike Kravetz /* 4430282f4214SMike Kravetz * hugepagesz command line processing 4431282f4214SMike Kravetz * A specific huge page size can only be specified once with hugepagesz. 4432282f4214SMike Kravetz * hugepagesz is followed by hugepages on the command line. The global 4433282f4214SMike Kravetz * variable 'parsed_valid_hugepagesz' is used to determine if prior 4434282f4214SMike Kravetz * hugepagesz argument was valid. 4435282f4214SMike Kravetz */ 4436359f2544SMike Kravetz static int __init hugepagesz_setup(char *s) 4437e11bfbfcSNick Piggin { 4438359f2544SMike Kravetz unsigned long size; 4439282f4214SMike Kravetz struct hstate *h; 4440282f4214SMike Kravetz 4441282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4442359f2544SMike Kravetz size = (unsigned long)memparse(s, NULL); 4443359f2544SMike Kravetz 4444359f2544SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4445282f4214SMike Kravetz pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); 4446f81f6e4bSPeng Liu return 1; 4447359f2544SMike Kravetz } 4448359f2544SMike Kravetz 4449282f4214SMike Kravetz h = size_to_hstate(size); 4450282f4214SMike Kravetz if (h) { 4451282f4214SMike Kravetz /* 4452282f4214SMike Kravetz * hstate for this size already exists. This is normally 4453282f4214SMike Kravetz * an error, but is allowed if the existing hstate is the 4454282f4214SMike Kravetz * default hstate. More specifically, it is only allowed if 4455282f4214SMike Kravetz * the number of huge pages for the default hstate was not 4456282f4214SMike Kravetz * previously specified. 4457282f4214SMike Kravetz */ 4458282f4214SMike Kravetz if (!parsed_default_hugepagesz || h != &default_hstate || 4459282f4214SMike Kravetz default_hstate.max_huge_pages) { 4460282f4214SMike Kravetz pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); 4461f81f6e4bSPeng Liu return 1; 446238237830SMike Kravetz } 446338237830SMike Kravetz 4464282f4214SMike Kravetz /* 4465282f4214SMike Kravetz * No need to call hugetlb_add_hstate() as hstate already 4466282f4214SMike Kravetz * exists. But, do set parsed_hstate so that a following 4467282f4214SMike Kravetz * hugepages= parameter will be applied to this hstate. 4468282f4214SMike Kravetz */ 4469282f4214SMike Kravetz parsed_hstate = h; 4470282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4471e11bfbfcSNick Piggin return 1; 4472e11bfbfcSNick Piggin } 4473282f4214SMike Kravetz 4474359f2544SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4475282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4476359f2544SMike Kravetz return 1; 4477359f2544SMike Kravetz } 4478359f2544SMike Kravetz __setup("hugepagesz=", hugepagesz_setup); 4479359f2544SMike Kravetz 4480282f4214SMike Kravetz /* 4481282f4214SMike Kravetz * default_hugepagesz command line input 4482282f4214SMike Kravetz * Only one instance of default_hugepagesz allowed on command line. 4483282f4214SMike Kravetz */ 4484ae94da89SMike Kravetz static int __init default_hugepagesz_setup(char *s) 4485e11bfbfcSNick Piggin { 4486ae94da89SMike Kravetz unsigned long size; 4487b5389086SZhenguo Yao int i; 4488ae94da89SMike Kravetz 4489282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4490282f4214SMike Kravetz if (parsed_default_hugepagesz) { 4491282f4214SMike Kravetz pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); 4492f81f6e4bSPeng Liu return 1; 4493282f4214SMike Kravetz } 4494282f4214SMike Kravetz 4495282f4214SMike Kravetz size = (unsigned long)memparse(s, NULL); 4496282f4214SMike Kravetz 4497282f4214SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4498282f4214SMike Kravetz pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); 4499f81f6e4bSPeng Liu return 1; 4500282f4214SMike Kravetz } 4501282f4214SMike Kravetz 4502282f4214SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4503282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4504282f4214SMike Kravetz parsed_default_hugepagesz = true; 4505282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(size)); 4506282f4214SMike Kravetz 4507282f4214SMike Kravetz /* 4508282f4214SMike Kravetz * The number of default huge pages (for this size) could have been 4509282f4214SMike Kravetz * specified as the first hugetlb parameter: hugepages=X. If so, 4510282f4214SMike Kravetz * then default_hstate_max_huge_pages is set. If the default huge 451123baf831SKirill A. Shutemov * page size is gigantic (> MAX_ORDER), then the pages must be 4512282f4214SMike Kravetz * allocated here from bootmem allocator. 4513282f4214SMike Kravetz */ 4514282f4214SMike Kravetz if (default_hstate_max_huge_pages) { 4515282f4214SMike Kravetz default_hstate.max_huge_pages = default_hstate_max_huge_pages; 45160a7a0f6fSPeng Liu for_each_online_node(i) 4517b5389086SZhenguo Yao default_hstate.max_huge_pages_node[i] = 4518b5389086SZhenguo Yao default_hugepages_in_node[i]; 4519282f4214SMike Kravetz if (hstate_is_gigantic(&default_hstate)) 4520282f4214SMike Kravetz hugetlb_hstate_alloc_pages(&default_hstate); 4521282f4214SMike Kravetz default_hstate_max_huge_pages = 0; 4522282f4214SMike Kravetz } 4523282f4214SMike Kravetz 4524e11bfbfcSNick Piggin return 1; 4525e11bfbfcSNick Piggin } 4526ae94da89SMike Kravetz __setup("default_hugepagesz=", default_hugepagesz_setup); 4527a3437870SNishanth Aravamudan 4528d2226ebdSFeng Tang static nodemask_t *policy_mbind_nodemask(gfp_t gfp) 4529d2226ebdSFeng Tang { 4530d2226ebdSFeng Tang #ifdef CONFIG_NUMA 4531d2226ebdSFeng Tang struct mempolicy *mpol = get_task_policy(current); 4532d2226ebdSFeng Tang 4533d2226ebdSFeng Tang /* 4534d2226ebdSFeng Tang * Only enforce MPOL_BIND policy which overlaps with cpuset policy 4535d2226ebdSFeng Tang * (from policy_nodemask) specifically for hugetlb case 4536d2226ebdSFeng Tang */ 4537d2226ebdSFeng Tang if (mpol->mode == MPOL_BIND && 4538d2226ebdSFeng Tang (apply_policy_zone(mpol, gfp_zone(gfp)) && 4539d2226ebdSFeng Tang cpuset_nodemask_valid_mems_allowed(&mpol->nodes))) 4540d2226ebdSFeng Tang return &mpol->nodes; 4541d2226ebdSFeng Tang #endif 4542d2226ebdSFeng Tang return NULL; 4543d2226ebdSFeng Tang } 4544d2226ebdSFeng Tang 45458ca39e68SMuchun Song static unsigned int allowed_mems_nr(struct hstate *h) 45468a213460SNishanth Aravamudan { 45478a213460SNishanth Aravamudan int node; 45488a213460SNishanth Aravamudan unsigned int nr = 0; 4549d2226ebdSFeng Tang nodemask_t *mbind_nodemask; 45508ca39e68SMuchun Song unsigned int *array = h->free_huge_pages_node; 45518ca39e68SMuchun Song gfp_t gfp_mask = htlb_alloc_mask(h); 45528a213460SNishanth Aravamudan 4553d2226ebdSFeng Tang mbind_nodemask = policy_mbind_nodemask(gfp_mask); 45548ca39e68SMuchun Song for_each_node_mask(node, cpuset_current_mems_allowed) { 4555d2226ebdSFeng Tang if (!mbind_nodemask || node_isset(node, *mbind_nodemask)) 45568a213460SNishanth Aravamudan nr += array[node]; 45578ca39e68SMuchun Song } 45588a213460SNishanth Aravamudan 45598a213460SNishanth Aravamudan return nr; 45608a213460SNishanth Aravamudan } 45618a213460SNishanth Aravamudan 45628a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 456317743798SMuchun Song static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, 456417743798SMuchun Song void *buffer, size_t *length, 456517743798SMuchun Song loff_t *ppos, unsigned long *out) 456617743798SMuchun Song { 456717743798SMuchun Song struct ctl_table dup_table; 456817743798SMuchun Song 456917743798SMuchun Song /* 457017743798SMuchun Song * In order to avoid races with __do_proc_doulongvec_minmax(), we 457117743798SMuchun Song * can duplicate the @table and alter the duplicate of it. 457217743798SMuchun Song */ 457317743798SMuchun Song dup_table = *table; 457417743798SMuchun Song dup_table.data = out; 457517743798SMuchun Song 457617743798SMuchun Song return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); 457717743798SMuchun Song } 457817743798SMuchun Song 457906808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 458006808b08SLee Schermerhorn struct ctl_table *table, int write, 458132927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 45821da177e4SLinus Torvalds { 4583e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 4584238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 458508d4a246SMichal Hocko int ret; 4586e5ff2159SAndi Kleen 4587457c1b27SNishanth Aravamudan if (!hugepages_supported()) 458886613628SJan Stancek return -EOPNOTSUPP; 4589457c1b27SNishanth Aravamudan 459017743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 459117743798SMuchun Song &tmp); 459208d4a246SMichal Hocko if (ret) 459308d4a246SMichal Hocko goto out; 4594e5ff2159SAndi Kleen 4595238d3c13SDavid Rientjes if (write) 4596238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 4597238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 459808d4a246SMichal Hocko out: 459908d4a246SMichal Hocko return ret; 46001da177e4SLinus Torvalds } 4601396faf03SMel Gorman 4602962de548SKefeng Wang static int hugetlb_sysctl_handler(struct ctl_table *table, int write, 460332927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 460406808b08SLee Schermerhorn { 460506808b08SLee Schermerhorn 460606808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 460706808b08SLee Schermerhorn buffer, length, ppos); 460806808b08SLee Schermerhorn } 460906808b08SLee Schermerhorn 461006808b08SLee Schermerhorn #ifdef CONFIG_NUMA 4611962de548SKefeng Wang static int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 461232927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 461306808b08SLee Schermerhorn { 461406808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 461506808b08SLee Schermerhorn buffer, length, ppos); 461606808b08SLee Schermerhorn } 461706808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 461806808b08SLee Schermerhorn 4619962de548SKefeng Wang static int hugetlb_overcommit_handler(struct ctl_table *table, int write, 462032927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 4621a3d0c6aaSNishanth Aravamudan { 4622a5516438SAndi Kleen struct hstate *h = &default_hstate; 4623e5ff2159SAndi Kleen unsigned long tmp; 462408d4a246SMichal Hocko int ret; 4625e5ff2159SAndi Kleen 4626457c1b27SNishanth Aravamudan if (!hugepages_supported()) 462786613628SJan Stancek return -EOPNOTSUPP; 4628457c1b27SNishanth Aravamudan 4629e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 4630e5ff2159SAndi Kleen 4631bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 4632adbe8726SEric B Munson return -EINVAL; 4633adbe8726SEric B Munson 463417743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 463517743798SMuchun Song &tmp); 463608d4a246SMichal Hocko if (ret) 463708d4a246SMichal Hocko goto out; 4638e5ff2159SAndi Kleen 4639e5ff2159SAndi Kleen if (write) { 4640db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4641e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 4642db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4643e5ff2159SAndi Kleen } 464408d4a246SMichal Hocko out: 464508d4a246SMichal Hocko return ret; 4646a3d0c6aaSNishanth Aravamudan } 4647a3d0c6aaSNishanth Aravamudan 4648962de548SKefeng Wang static struct ctl_table hugetlb_table[] = { 4649962de548SKefeng Wang { 4650962de548SKefeng Wang .procname = "nr_hugepages", 4651962de548SKefeng Wang .data = NULL, 4652962de548SKefeng Wang .maxlen = sizeof(unsigned long), 4653962de548SKefeng Wang .mode = 0644, 4654962de548SKefeng Wang .proc_handler = hugetlb_sysctl_handler, 4655962de548SKefeng Wang }, 4656962de548SKefeng Wang #ifdef CONFIG_NUMA 4657962de548SKefeng Wang { 4658962de548SKefeng Wang .procname = "nr_hugepages_mempolicy", 4659962de548SKefeng Wang .data = NULL, 4660962de548SKefeng Wang .maxlen = sizeof(unsigned long), 4661962de548SKefeng Wang .mode = 0644, 4662962de548SKefeng Wang .proc_handler = &hugetlb_mempolicy_sysctl_handler, 4663962de548SKefeng Wang }, 4664962de548SKefeng Wang #endif 4665962de548SKefeng Wang { 4666962de548SKefeng Wang .procname = "hugetlb_shm_group", 4667962de548SKefeng Wang .data = &sysctl_hugetlb_shm_group, 4668962de548SKefeng Wang .maxlen = sizeof(gid_t), 4669962de548SKefeng Wang .mode = 0644, 4670962de548SKefeng Wang .proc_handler = proc_dointvec, 4671962de548SKefeng Wang }, 4672962de548SKefeng Wang { 4673962de548SKefeng Wang .procname = "nr_overcommit_hugepages", 4674962de548SKefeng Wang .data = NULL, 4675962de548SKefeng Wang .maxlen = sizeof(unsigned long), 4676962de548SKefeng Wang .mode = 0644, 4677962de548SKefeng Wang .proc_handler = hugetlb_overcommit_handler, 4678962de548SKefeng Wang }, 4679962de548SKefeng Wang { } 4680962de548SKefeng Wang }; 4681962de548SKefeng Wang 4682962de548SKefeng Wang static void hugetlb_sysctl_init(void) 4683962de548SKefeng Wang { 4684962de548SKefeng Wang register_sysctl_init("vm", hugetlb_table); 4685962de548SKefeng Wang } 46861da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 46871da177e4SLinus Torvalds 4688e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 46891da177e4SLinus Torvalds { 4690fcb2b0c5SRoman Gushchin struct hstate *h; 4691fcb2b0c5SRoman Gushchin unsigned long total = 0; 4692fcb2b0c5SRoman Gushchin 4693457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4694457c1b27SNishanth Aravamudan return; 4695fcb2b0c5SRoman Gushchin 4696fcb2b0c5SRoman Gushchin for_each_hstate(h) { 4697fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 4698fcb2b0c5SRoman Gushchin 4699aca78307SMiaohe Lin total += huge_page_size(h) * count; 4700fcb2b0c5SRoman Gushchin 4701fcb2b0c5SRoman Gushchin if (h == &default_hstate) 4702e1759c21SAlexey Dobriyan seq_printf(m, 47031da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 47041da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 4705b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 47067893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 47074f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 4708fcb2b0c5SRoman Gushchin count, 4709a5516438SAndi Kleen h->free_huge_pages, 4710a5516438SAndi Kleen h->resv_huge_pages, 4711a5516438SAndi Kleen h->surplus_huge_pages, 4712aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4713fcb2b0c5SRoman Gushchin } 4714fcb2b0c5SRoman Gushchin 4715aca78307SMiaohe Lin seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); 47161da177e4SLinus Torvalds } 47171da177e4SLinus Torvalds 47187981593bSJoe Perches int hugetlb_report_node_meminfo(char *buf, int len, int nid) 47191da177e4SLinus Torvalds { 4720a5516438SAndi Kleen struct hstate *h = &default_hstate; 47217981593bSJoe Perches 4722457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4723457c1b27SNishanth Aravamudan return 0; 47247981593bSJoe Perches 47257981593bSJoe Perches return sysfs_emit_at(buf, len, 47261da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 4727a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 4728a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 4729a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 4730a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 4731a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 47321da177e4SLinus Torvalds } 47331da177e4SLinus Torvalds 4734dcadcf1cSGang Li void hugetlb_show_meminfo_node(int nid) 4735949f7ec5SDavid Rientjes { 4736949f7ec5SDavid Rientjes struct hstate *h; 4737949f7ec5SDavid Rientjes 4738457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4739457c1b27SNishanth Aravamudan return; 4740457c1b27SNishanth Aravamudan 4741949f7ec5SDavid Rientjes for_each_hstate(h) 4742dcadcf1cSGang Li printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 4743949f7ec5SDavid Rientjes nid, 4744949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 4745949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 4746949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 4747aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4748949f7ec5SDavid Rientjes } 4749949f7ec5SDavid Rientjes 47505d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 47515d317b2bSNaoya Horiguchi { 47525d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 47535d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 47545d317b2bSNaoya Horiguchi } 47555d317b2bSNaoya Horiguchi 47561da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 47571da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 47581da177e4SLinus Torvalds { 4759d0028588SWanpeng Li struct hstate *h; 4760d0028588SWanpeng Li unsigned long nr_total_pages = 0; 4761d0028588SWanpeng Li 4762d0028588SWanpeng Li for_each_hstate(h) 4763d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 4764d0028588SWanpeng Li return nr_total_pages; 47651da177e4SLinus Torvalds } 47661da177e4SLinus Torvalds 4767a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 4768fc1b8a73SMel Gorman { 4769fc1b8a73SMel Gorman int ret = -ENOMEM; 4770fc1b8a73SMel Gorman 47710aa7f354SMiaohe Lin if (!delta) 47720aa7f354SMiaohe Lin return 0; 47730aa7f354SMiaohe Lin 4774db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4775fc1b8a73SMel Gorman /* 4776fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 4777fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 4778fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 4779fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 4780fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 4781fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 4782fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 4783fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 4784fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 4785fc1b8a73SMel Gorman * 4786fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 4787fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 4788fc1b8a73SMel Gorman * we fall back to check against current free page availability as 4789fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 4790fc1b8a73SMel Gorman * semantics that cpuset has. 47918ca39e68SMuchun Song * 47928ca39e68SMuchun Song * Apart from cpuset, we also have memory policy mechanism that 47938ca39e68SMuchun Song * also determines from which node the kernel will allocate memory 47948ca39e68SMuchun Song * in a NUMA system. So similar to cpuset, we also should consider 47958ca39e68SMuchun Song * the memory policy of the current task. Similar to the description 47968ca39e68SMuchun Song * above. 4797fc1b8a73SMel Gorman */ 4798fc1b8a73SMel Gorman if (delta > 0) { 4799a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 4800fc1b8a73SMel Gorman goto out; 4801fc1b8a73SMel Gorman 48028ca39e68SMuchun Song if (delta > allowed_mems_nr(h)) { 4803a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 4804fc1b8a73SMel Gorman goto out; 4805fc1b8a73SMel Gorman } 4806fc1b8a73SMel Gorman } 4807fc1b8a73SMel Gorman 4808fc1b8a73SMel Gorman ret = 0; 4809fc1b8a73SMel Gorman if (delta < 0) 4810a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 4811fc1b8a73SMel Gorman 4812fc1b8a73SMel Gorman out: 4813db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4814fc1b8a73SMel Gorman return ret; 4815fc1b8a73SMel Gorman } 4816fc1b8a73SMel Gorman 481784afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 481884afd99bSAndy Whitcroft { 4819f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 482084afd99bSAndy Whitcroft 482184afd99bSAndy Whitcroft /* 4822612b8a31SMike Kravetz * HPAGE_RESV_OWNER indicates a private mapping. 482384afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 482484afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 482584afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 482625985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 482784afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 482884afd99bSAndy Whitcroft * new reference here without additional locking. 482984afd99bSAndy Whitcroft */ 483009a26e83SMike Kravetz if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 483109a26e83SMike Kravetz resv_map_dup_hugetlb_cgroup_uncharge_info(resv); 4832f522c3acSJoonsoo Kim kref_get(&resv->refs); 483384afd99bSAndy Whitcroft } 48348d9bfb26SMike Kravetz 4835131a79b4SMike Kravetz /* 4836131a79b4SMike Kravetz * vma_lock structure for sharable mappings is vma specific. 4837612b8a31SMike Kravetz * Clear old pointer (if copied via vm_area_dup) and allocate 4838612b8a31SMike Kravetz * new structure. Before clearing, make sure vma_lock is not 4839612b8a31SMike Kravetz * for this vma. 4840131a79b4SMike Kravetz */ 4841131a79b4SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 4842612b8a31SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 4843612b8a31SMike Kravetz 4844612b8a31SMike Kravetz if (vma_lock) { 4845612b8a31SMike Kravetz if (vma_lock->vma != vma) { 4846131a79b4SMike Kravetz vma->vm_private_data = NULL; 48478d9bfb26SMike Kravetz hugetlb_vma_lock_alloc(vma); 4848612b8a31SMike Kravetz } else 4849612b8a31SMike Kravetz pr_warn("HugeTLB: vma_lock already exists in %s.\n", __func__); 4850612b8a31SMike Kravetz } else 4851612b8a31SMike Kravetz hugetlb_vma_lock_alloc(vma); 485209a26e83SMike Kravetz } 4853131a79b4SMike Kravetz } 485484afd99bSAndy Whitcroft 4855a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 4856a1e78772SMel Gorman { 4857a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 48588d9bfb26SMike Kravetz struct resv_map *resv; 485990481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 48604e35f483SJoonsoo Kim unsigned long reserve, start, end; 48611c5ecae3SMike Kravetz long gbl_reserve; 486284afd99bSAndy Whitcroft 48638d9bfb26SMike Kravetz hugetlb_vma_lock_free(vma); 48648d9bfb26SMike Kravetz 48658d9bfb26SMike Kravetz resv = vma_resv_map(vma); 48664e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 48674e35f483SJoonsoo Kim return; 48684e35f483SJoonsoo Kim 4869a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 4870a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 487184afd99bSAndy Whitcroft 48724e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 4873e9fe92aeSMina Almasry hugetlb_cgroup_uncharge_counter(resv, start, end); 48747251ff78SAdam Litke if (reserve) { 48751c5ecae3SMike Kravetz /* 48761c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 48771c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 48781c5ecae3SMike Kravetz */ 48791c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 48801c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 48817251ff78SAdam Litke } 4882e9fe92aeSMina Almasry 4883e9fe92aeSMina Almasry kref_put(&resv->refs, resv_map_release); 4884a1e78772SMel Gorman } 4885a1e78772SMel Gorman 488631383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 488731383c68SDan Williams { 488831383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 488931383c68SDan Williams return -EINVAL; 4890b30c14cdSJames Houghton 4891b30c14cdSJames Houghton /* 4892b30c14cdSJames Houghton * PMD sharing is only possible for PUD_SIZE-aligned address ranges 4893b30c14cdSJames Houghton * in HugeTLB VMAs. If we will lose PUD_SIZE alignment due to this 4894b30c14cdSJames Houghton * split, unshare PMDs in the PUD_SIZE interval surrounding addr now. 4895b30c14cdSJames Houghton */ 4896b30c14cdSJames Houghton if (addr & ~PUD_MASK) { 4897b30c14cdSJames Houghton /* 4898b30c14cdSJames Houghton * hugetlb_vm_op_split is called right before we attempt to 4899b30c14cdSJames Houghton * split the VMA. We will need to unshare PMDs in the old and 4900b30c14cdSJames Houghton * new VMAs, so let's unshare before we split. 4901b30c14cdSJames Houghton */ 4902b30c14cdSJames Houghton unsigned long floor = addr & PUD_MASK; 4903b30c14cdSJames Houghton unsigned long ceil = floor + PUD_SIZE; 4904b30c14cdSJames Houghton 4905b30c14cdSJames Houghton if (floor >= vma->vm_start && ceil <= vma->vm_end) 4906b30c14cdSJames Houghton hugetlb_unshare_pmds(vma, floor, ceil); 4907b30c14cdSJames Houghton } 4908b30c14cdSJames Houghton 490931383c68SDan Williams return 0; 491031383c68SDan Williams } 491131383c68SDan Williams 491205ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 491305ea8860SDan Williams { 4914aca78307SMiaohe Lin return huge_page_size(hstate_vma(vma)); 491505ea8860SDan Williams } 491605ea8860SDan Williams 49171da177e4SLinus Torvalds /* 49181da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 49191da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 49206c26d310SMiaohe Lin * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get 49211da177e4SLinus Torvalds * this far. 49221da177e4SLinus Torvalds */ 4923b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 49241da177e4SLinus Torvalds { 49251da177e4SLinus Torvalds BUG(); 4926d0217ac0SNick Piggin return 0; 49271da177e4SLinus Torvalds } 49281da177e4SLinus Torvalds 4929eec3636aSJane Chu /* 4930eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 4931eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 4932eec3636aSJane Chu * This is because under System V memory model, mappings created via 4933eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 4934eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 4935eec3636aSJane Chu */ 4936f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 4937d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 493884afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 4939a1e78772SMel Gorman .close = hugetlb_vm_op_close, 4940dd3b614fSDmitry Safonov .may_split = hugetlb_vm_op_split, 494105ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 49421da177e4SLinus Torvalds }; 49431da177e4SLinus Torvalds 49441e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 49451e8f889bSDavid Gibson int writable) 494663551ae0SDavid Gibson { 494763551ae0SDavid Gibson pte_t entry; 494879c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 494963551ae0SDavid Gibson 49501e8f889bSDavid Gibson if (writable) { 4951106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 4952106c992aSGerald Schaefer vma->vm_page_prot))); 495363551ae0SDavid Gibson } else { 4954106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 4955106c992aSGerald Schaefer vma->vm_page_prot)); 495663551ae0SDavid Gibson } 495763551ae0SDavid Gibson entry = pte_mkyoung(entry); 495879c1c594SChristophe Leroy entry = arch_make_huge_pte(entry, shift, vma->vm_flags); 495963551ae0SDavid Gibson 496063551ae0SDavid Gibson return entry; 496163551ae0SDavid Gibson } 496263551ae0SDavid Gibson 49631e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 49641e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 49651e8f889bSDavid Gibson { 49661e8f889bSDavid Gibson pte_t entry; 49671e8f889bSDavid Gibson 4968106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 496932f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 49704b3073e1SRussell King update_mmu_cache(vma, address, ptep); 49711e8f889bSDavid Gibson } 49721e8f889bSDavid Gibson 4973d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 49744a705fefSNaoya Horiguchi { 49754a705fefSNaoya Horiguchi swp_entry_t swp; 49764a705fefSNaoya Horiguchi 49774a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 4978d5ed7444SAneesh Kumar K.V return false; 49794a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4980d79d176aSBaoquan He if (is_migration_entry(swp)) 4981d5ed7444SAneesh Kumar K.V return true; 49824a705fefSNaoya Horiguchi else 4983d5ed7444SAneesh Kumar K.V return false; 49844a705fefSNaoya Horiguchi } 49854a705fefSNaoya Horiguchi 49863e5c3600SBaoquan He static bool is_hugetlb_entry_hwpoisoned(pte_t pte) 49874a705fefSNaoya Horiguchi { 49884a705fefSNaoya Horiguchi swp_entry_t swp; 49894a705fefSNaoya Horiguchi 49904a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 49913e5c3600SBaoquan He return false; 49924a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4993d79d176aSBaoquan He if (is_hwpoison_entry(swp)) 49943e5c3600SBaoquan He return true; 49954a705fefSNaoya Horiguchi else 49963e5c3600SBaoquan He return false; 49974a705fefSNaoya Horiguchi } 49981e8f889bSDavid Gibson 49994eae4efaSPeter Xu static void 5000ea4c353dSSidhartha Kumar hugetlb_install_folio(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, 50015a2f8d22SPeter Xu struct folio *new_folio, pte_t old) 50024eae4efaSPeter Xu { 50035a2f8d22SPeter Xu pte_t newpte = make_huge_pte(vma, &new_folio->page, 1); 50045a2f8d22SPeter Xu 5005ea4c353dSSidhartha Kumar __folio_mark_uptodate(new_folio); 5006d0ce0e47SSidhartha Kumar hugepage_add_new_anon_rmap(new_folio, vma, addr); 50075a2f8d22SPeter Xu if (userfaultfd_wp(vma) && huge_pte_uffd_wp(old)) 50085a2f8d22SPeter Xu newpte = huge_pte_mkuffd_wp(newpte); 50095a2f8d22SPeter Xu set_huge_pte_at(vma->vm_mm, addr, ptep, newpte); 50104eae4efaSPeter Xu hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); 5011ea4c353dSSidhartha Kumar folio_set_hugetlb_migratable(new_folio); 50124eae4efaSPeter Xu } 50134eae4efaSPeter Xu 501463551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 5015bc70fbf2SPeter Xu struct vm_area_struct *dst_vma, 5016bc70fbf2SPeter Xu struct vm_area_struct *src_vma) 501763551ae0SDavid Gibson { 50183aa4ed80SMiaohe Lin pte_t *src_pte, *dst_pte, entry; 5019ad27ce20SZhangPeng struct folio *pte_folio; 50201c59827dSHugh Dickins unsigned long addr; 5021bc70fbf2SPeter Xu bool cow = is_cow_mapping(src_vma->vm_flags); 5022bc70fbf2SPeter Xu struct hstate *h = hstate_vma(src_vma); 5023a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 50244eae4efaSPeter Xu unsigned long npages = pages_per_huge_page(h); 5025ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 5026e95a9851SMike Kravetz unsigned long last_addr_mask; 5027e8569dd2SAndreas Sandberg int ret = 0; 50281e8f889bSDavid Gibson 5029ac46d4f3SJérôme Glisse if (cow) { 50307d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src, 5031bc70fbf2SPeter Xu src_vma->vm_start, 5032bc70fbf2SPeter Xu src_vma->vm_end); 5033ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 5034623a1ddfSDavid Hildenbrand mmap_assert_write_locked(src); 5035623a1ddfSDavid Hildenbrand raw_write_seqcount_begin(&src->write_protect_seq); 503640549ba8SMike Kravetz } else { 503740549ba8SMike Kravetz /* 503840549ba8SMike Kravetz * For shared mappings the vma lock must be held before 50399c67a207SPeter Xu * calling hugetlb_walk() in the src vma. Otherwise, the 504040549ba8SMike Kravetz * returned ptep could go away if part of a shared pmd and 504140549ba8SMike Kravetz * another thread calls huge_pmd_unshare. 504240549ba8SMike Kravetz */ 504340549ba8SMike Kravetz hugetlb_vma_lock_read(src_vma); 5044ac46d4f3SJérôme Glisse } 5045e8569dd2SAndreas Sandberg 5046e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 5047bc70fbf2SPeter Xu for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) { 5048cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 50499c67a207SPeter Xu src_pte = hugetlb_walk(src_vma, addr, sz); 5050e95a9851SMike Kravetz if (!src_pte) { 5051e95a9851SMike Kravetz addr |= last_addr_mask; 5052c74df32cSHugh Dickins continue; 5053e95a9851SMike Kravetz } 5054bc70fbf2SPeter Xu dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz); 5055e8569dd2SAndreas Sandberg if (!dst_pte) { 5056e8569dd2SAndreas Sandberg ret = -ENOMEM; 5057e8569dd2SAndreas Sandberg break; 5058e8569dd2SAndreas Sandberg } 5059c5c99429SLarry Woodman 50605e41540cSMike Kravetz /* 50615e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 50625e41540cSMike Kravetz * 50633aa4ed80SMiaohe Lin * dst_pte == src_pte is the common case of src/dest sharing. 50645e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 50653aa4ed80SMiaohe Lin * another vma. So page_count of ptep page is checked instead 50663aa4ed80SMiaohe Lin * to reliably determine whether pte is shared. 50675e41540cSMike Kravetz */ 50683aa4ed80SMiaohe Lin if (page_count(virt_to_page(dst_pte)) > 1) { 5069e95a9851SMike Kravetz addr |= last_addr_mask; 5070c5c99429SLarry Woodman continue; 5071e95a9851SMike Kravetz } 5072c5c99429SLarry Woodman 5073cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 5074cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 5075cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 50764a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 50774eae4efaSPeter Xu again: 50783aa4ed80SMiaohe Lin if (huge_pte_none(entry)) { 50795e41540cSMike Kravetz /* 50803aa4ed80SMiaohe Lin * Skip if src entry none. 50815e41540cSMike Kravetz */ 50824a705fefSNaoya Horiguchi ; 5083c2cb0dccSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) { 50845a2f8d22SPeter Xu if (!userfaultfd_wp(dst_vma)) 5085c2cb0dccSNaoya Horiguchi entry = huge_pte_clear_uffd_wp(entry); 5086c2cb0dccSNaoya Horiguchi set_huge_pte_at(dst, addr, dst_pte, entry); 5087c2cb0dccSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry))) { 50884a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 50895a2f8d22SPeter Xu bool uffd_wp = pte_swp_uffd_wp(entry); 50904a705fefSNaoya Horiguchi 50916c287605SDavid Hildenbrand if (!is_readable_migration_entry(swp_entry) && cow) { 50924a705fefSNaoya Horiguchi /* 50934a705fefSNaoya Horiguchi * COW mappings require pages in both 50944a705fefSNaoya Horiguchi * parent and child to be set to read. 50954a705fefSNaoya Horiguchi */ 50964dd845b5SAlistair Popple swp_entry = make_readable_migration_entry( 50974dd845b5SAlistair Popple swp_offset(swp_entry)); 50984a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 5099bc70fbf2SPeter Xu if (userfaultfd_wp(src_vma) && uffd_wp) 51005a2f8d22SPeter Xu entry = pte_swp_mkuffd_wp(entry); 510118f39629SQi Zheng set_huge_pte_at(src, addr, src_pte, entry); 51024a705fefSNaoya Horiguchi } 51035a2f8d22SPeter Xu if (!userfaultfd_wp(dst_vma)) 5104bc70fbf2SPeter Xu entry = huge_pte_clear_uffd_wp(entry); 510518f39629SQi Zheng set_huge_pte_at(dst, addr, dst_pte, entry); 5106bc70fbf2SPeter Xu } else if (unlikely(is_pte_marker(entry))) { 51077e3ce3f8SPeter Xu /* No swap on hugetlb */ 51087e3ce3f8SPeter Xu WARN_ON_ONCE( 51097e3ce3f8SPeter Xu is_swapin_error_entry(pte_to_swp_entry(entry))); 5110bc70fbf2SPeter Xu /* 5111bc70fbf2SPeter Xu * We copy the pte marker only if the dst vma has 5112bc70fbf2SPeter Xu * uffd-wp enabled. 5113bc70fbf2SPeter Xu */ 5114bc70fbf2SPeter Xu if (userfaultfd_wp(dst_vma)) 5115bc70fbf2SPeter Xu set_huge_pte_at(dst, addr, dst_pte, entry); 51164a705fefSNaoya Horiguchi } else { 51174eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 5118ad27ce20SZhangPeng pte_folio = page_folio(pte_page(entry)); 5119ad27ce20SZhangPeng folio_get(pte_folio); 51204eae4efaSPeter Xu 51214eae4efaSPeter Xu /* 5122fb3d824dSDavid Hildenbrand * Failing to duplicate the anon rmap is a rare case 5123fb3d824dSDavid Hildenbrand * where we see pinned hugetlb pages while they're 5124fb3d824dSDavid Hildenbrand * prone to COW. We need to do the COW earlier during 5125fb3d824dSDavid Hildenbrand * fork. 51264eae4efaSPeter Xu * 51274eae4efaSPeter Xu * When pre-allocating the page or copying data, we 51284eae4efaSPeter Xu * need to be without the pgtable locks since we could 51294eae4efaSPeter Xu * sleep during the process. 51304eae4efaSPeter Xu */ 5131ad27ce20SZhangPeng if (!folio_test_anon(pte_folio)) { 5132ad27ce20SZhangPeng page_dup_file_rmap(&pte_folio->page, true); 5133ad27ce20SZhangPeng } else if (page_try_dup_anon_rmap(&pte_folio->page, 5134ad27ce20SZhangPeng true, src_vma)) { 51354eae4efaSPeter Xu pte_t src_pte_old = entry; 5136d0ce0e47SSidhartha Kumar struct folio *new_folio; 51374eae4efaSPeter Xu 51384eae4efaSPeter Xu spin_unlock(src_ptl); 51394eae4efaSPeter Xu spin_unlock(dst_ptl); 51404eae4efaSPeter Xu /* Do not use reserve as it's private owned */ 5141d0ce0e47SSidhartha Kumar new_folio = alloc_hugetlb_folio(dst_vma, addr, 1); 5142d0ce0e47SSidhartha Kumar if (IS_ERR(new_folio)) { 5143ad27ce20SZhangPeng folio_put(pte_folio); 5144d0ce0e47SSidhartha Kumar ret = PTR_ERR(new_folio); 51454eae4efaSPeter Xu break; 51464eae4efaSPeter Xu } 51471cb9dc4bSLiu Shixin ret = copy_user_large_folio(new_folio, 5148ad27ce20SZhangPeng pte_folio, 5149c0e8150eSZhangPeng addr, dst_vma); 5150ad27ce20SZhangPeng folio_put(pte_folio); 51511cb9dc4bSLiu Shixin if (ret) { 51521cb9dc4bSLiu Shixin folio_put(new_folio); 51531cb9dc4bSLiu Shixin break; 51541cb9dc4bSLiu Shixin } 51554eae4efaSPeter Xu 5156d0ce0e47SSidhartha Kumar /* Install the new hugetlb folio if src pte stable */ 51574eae4efaSPeter Xu dst_ptl = huge_pte_lock(h, dst, dst_pte); 51584eae4efaSPeter Xu src_ptl = huge_pte_lockptr(h, src, src_pte); 51594eae4efaSPeter Xu spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 51604eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 51614eae4efaSPeter Xu if (!pte_same(src_pte_old, entry)) { 5162bc70fbf2SPeter Xu restore_reserve_on_error(h, dst_vma, addr, 5163d2d7bb44SSidhartha Kumar new_folio); 5164d0ce0e47SSidhartha Kumar folio_put(new_folio); 51653aa4ed80SMiaohe Lin /* huge_ptep of dst_pte won't change as in child */ 51664eae4efaSPeter Xu goto again; 51674eae4efaSPeter Xu } 51685a2f8d22SPeter Xu hugetlb_install_folio(dst_vma, dst_pte, addr, 51695a2f8d22SPeter Xu new_folio, src_pte_old); 51704eae4efaSPeter Xu spin_unlock(src_ptl); 51714eae4efaSPeter Xu spin_unlock(dst_ptl); 51724eae4efaSPeter Xu continue; 51734eae4efaSPeter Xu } 51744eae4efaSPeter Xu 517534ee645eSJoerg Roedel if (cow) { 51760f10851eSJérôme Glisse /* 51770f10851eSJérôme Glisse * No need to notify as we are downgrading page 51780f10851eSJérôme Glisse * table protection not changing it to point 51790f10851eSJérôme Glisse * to a new page. 51800f10851eSJérôme Glisse * 5181ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 51820f10851eSJérôme Glisse */ 51837f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 518484894e1cSPeter Xu entry = huge_pte_wrprotect(entry); 518534ee645eSJoerg Roedel } 51864eae4efaSPeter Xu 51875a2f8d22SPeter Xu if (!userfaultfd_wp(dst_vma)) 51885a2f8d22SPeter Xu entry = huge_pte_clear_uffd_wp(entry); 51895a2f8d22SPeter Xu 519063551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 51914eae4efaSPeter Xu hugetlb_count_add(npages, dst); 51921c59827dSHugh Dickins } 5193cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 5194cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 519563551ae0SDavid Gibson } 519663551ae0SDavid Gibson 5197623a1ddfSDavid Hildenbrand if (cow) { 5198623a1ddfSDavid Hildenbrand raw_write_seqcount_end(&src->write_protect_seq); 5199ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 520040549ba8SMike Kravetz } else { 520140549ba8SMike Kravetz hugetlb_vma_unlock_read(src_vma); 5202623a1ddfSDavid Hildenbrand } 5203e8569dd2SAndreas Sandberg 5204e8569dd2SAndreas Sandberg return ret; 520563551ae0SDavid Gibson } 520663551ae0SDavid Gibson 5207550a7d60SMina Almasry static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr, 5208db110a99SAneesh Kumar K.V unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte) 5209550a7d60SMina Almasry { 5210550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 5211550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 5212550a7d60SMina Almasry spinlock_t *src_ptl, *dst_ptl; 5213db110a99SAneesh Kumar K.V pte_t pte; 5214550a7d60SMina Almasry 5215550a7d60SMina Almasry dst_ptl = huge_pte_lock(h, mm, dst_pte); 5216550a7d60SMina Almasry src_ptl = huge_pte_lockptr(h, mm, src_pte); 5217550a7d60SMina Almasry 5218550a7d60SMina Almasry /* 5219550a7d60SMina Almasry * We don't have to worry about the ordering of src and dst ptlocks 52208651a137SLorenzo Stoakes * because exclusive mmap_lock (or the i_mmap_lock) prevents deadlock. 5221550a7d60SMina Almasry */ 5222550a7d60SMina Almasry if (src_ptl != dst_ptl) 5223550a7d60SMina Almasry spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 5224550a7d60SMina Almasry 5225550a7d60SMina Almasry pte = huge_ptep_get_and_clear(mm, old_addr, src_pte); 5226550a7d60SMina Almasry set_huge_pte_at(mm, new_addr, dst_pte, pte); 5227550a7d60SMina Almasry 5228550a7d60SMina Almasry if (src_ptl != dst_ptl) 5229550a7d60SMina Almasry spin_unlock(src_ptl); 5230550a7d60SMina Almasry spin_unlock(dst_ptl); 5231550a7d60SMina Almasry } 5232550a7d60SMina Almasry 5233550a7d60SMina Almasry int move_hugetlb_page_tables(struct vm_area_struct *vma, 5234550a7d60SMina Almasry struct vm_area_struct *new_vma, 5235550a7d60SMina Almasry unsigned long old_addr, unsigned long new_addr, 5236550a7d60SMina Almasry unsigned long len) 5237550a7d60SMina Almasry { 5238550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 5239550a7d60SMina Almasry struct address_space *mapping = vma->vm_file->f_mapping; 5240550a7d60SMina Almasry unsigned long sz = huge_page_size(h); 5241550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 5242550a7d60SMina Almasry unsigned long old_end = old_addr + len; 5243e95a9851SMike Kravetz unsigned long last_addr_mask; 5244550a7d60SMina Almasry pte_t *src_pte, *dst_pte; 5245550a7d60SMina Almasry struct mmu_notifier_range range; 52463d0b95cdSBaolin Wang bool shared_pmd = false; 5247550a7d60SMina Almasry 52487d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, old_addr, 5249550a7d60SMina Almasry old_end); 5250550a7d60SMina Almasry adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 52513d0b95cdSBaolin Wang /* 52523d0b95cdSBaolin Wang * In case of shared PMDs, we should cover the maximum possible 52533d0b95cdSBaolin Wang * range. 52543d0b95cdSBaolin Wang */ 52553d0b95cdSBaolin Wang flush_cache_range(vma, range.start, range.end); 52563d0b95cdSBaolin Wang 5257550a7d60SMina Almasry mmu_notifier_invalidate_range_start(&range); 5258e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 5259550a7d60SMina Almasry /* Prevent race with file truncation */ 526040549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 5261550a7d60SMina Almasry i_mmap_lock_write(mapping); 5262550a7d60SMina Almasry for (; old_addr < old_end; old_addr += sz, new_addr += sz) { 52639c67a207SPeter Xu src_pte = hugetlb_walk(vma, old_addr, sz); 5264e95a9851SMike Kravetz if (!src_pte) { 5265e95a9851SMike Kravetz old_addr |= last_addr_mask; 5266e95a9851SMike Kravetz new_addr |= last_addr_mask; 5267550a7d60SMina Almasry continue; 5268e95a9851SMike Kravetz } 5269550a7d60SMina Almasry if (huge_pte_none(huge_ptep_get(src_pte))) 5270550a7d60SMina Almasry continue; 5271550a7d60SMina Almasry 52724ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) { 52733d0b95cdSBaolin Wang shared_pmd = true; 52744ddb4d91SMike Kravetz old_addr |= last_addr_mask; 52754ddb4d91SMike Kravetz new_addr |= last_addr_mask; 5276550a7d60SMina Almasry continue; 52773d0b95cdSBaolin Wang } 5278550a7d60SMina Almasry 5279550a7d60SMina Almasry dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz); 5280550a7d60SMina Almasry if (!dst_pte) 5281550a7d60SMina Almasry break; 5282550a7d60SMina Almasry 5283db110a99SAneesh Kumar K.V move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte); 5284550a7d60SMina Almasry } 52853d0b95cdSBaolin Wang 52863d0b95cdSBaolin Wang if (shared_pmd) 52873d0b95cdSBaolin Wang flush_tlb_range(vma, range.start, range.end); 52883d0b95cdSBaolin Wang else 5289550a7d60SMina Almasry flush_tlb_range(vma, old_end - len, old_end); 5290550a7d60SMina Almasry mmu_notifier_invalidate_range_end(&range); 529113e4ad2cSNadav Amit i_mmap_unlock_write(mapping); 529240549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 5293550a7d60SMina Almasry 5294550a7d60SMina Almasry return len + old_addr - old_end; 5295550a7d60SMina Almasry } 5296550a7d60SMina Almasry 529773c54763SPeter Xu static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 529824669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 529905e90bd0SPeter Xu struct page *ref_page, zap_flags_t zap_flags) 530063551ae0SDavid Gibson { 530163551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 530263551ae0SDavid Gibson unsigned long address; 5303c7546f8fSDavid Gibson pte_t *ptep; 530463551ae0SDavid Gibson pte_t pte; 5305cb900f41SKirill A. Shutemov spinlock_t *ptl; 530663551ae0SDavid Gibson struct page *page; 5307a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 5308a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 5309e95a9851SMike Kravetz unsigned long last_addr_mask; 5310a4a118f2SNadav Amit bool force_flush = false; 5311a5516438SAndi Kleen 531263551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 5313a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 5314a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 531563551ae0SDavid Gibson 531607e32661SAneesh Kumar K.V /* 531707e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 531807e32661SAneesh Kumar K.V * to huge page. 531907e32661SAneesh Kumar K.V */ 5320ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 532124669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 5322dff11abeSMike Kravetz 5323e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 5324569f48b8SHillf Danton address = start; 5325569f48b8SHillf Danton for (; address < end; address += sz) { 53269c67a207SPeter Xu ptep = hugetlb_walk(vma, address, sz); 5327e95a9851SMike Kravetz if (!ptep) { 5328e95a9851SMike Kravetz address |= last_addr_mask; 5329c7546f8fSDavid Gibson continue; 5330e95a9851SMike Kravetz } 5331c7546f8fSDavid Gibson 5332cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 53334ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, address, ptep)) { 533431d49da5SAneesh Kumar K.V spin_unlock(ptl); 5335a4a118f2SNadav Amit tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE); 5336a4a118f2SNadav Amit force_flush = true; 53374ddb4d91SMike Kravetz address |= last_addr_mask; 533831d49da5SAneesh Kumar K.V continue; 533931d49da5SAneesh Kumar K.V } 534039dde65cSChen, Kenneth W 53416629326bSHillf Danton pte = huge_ptep_get(ptep); 534231d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 534331d49da5SAneesh Kumar K.V spin_unlock(ptl); 534431d49da5SAneesh Kumar K.V continue; 534531d49da5SAneesh Kumar K.V } 53466629326bSHillf Danton 53476629326bSHillf Danton /* 53489fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 53499fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 53506629326bSHillf Danton */ 53519fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 535205e90bd0SPeter Xu /* 535305e90bd0SPeter Xu * If the pte was wr-protected by uffd-wp in any of the 535405e90bd0SPeter Xu * swap forms, meanwhile the caller does not want to 535505e90bd0SPeter Xu * drop the uffd-wp bit in this zap, then replace the 535605e90bd0SPeter Xu * pte with a marker. 535705e90bd0SPeter Xu */ 535805e90bd0SPeter Xu if (pte_swp_uffd_wp_any(pte) && 535905e90bd0SPeter Xu !(zap_flags & ZAP_FLAG_DROP_MARKER)) 536005e90bd0SPeter Xu set_huge_pte_at(mm, address, ptep, 536105e90bd0SPeter Xu make_pte_marker(PTE_MARKER_UFFD_WP)); 536205e90bd0SPeter Xu else 53639386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 536431d49da5SAneesh Kumar K.V spin_unlock(ptl); 536531d49da5SAneesh Kumar K.V continue; 53668c4894c6SNaoya Horiguchi } 53676629326bSHillf Danton 53686629326bSHillf Danton page = pte_page(pte); 536904f2cbe3SMel Gorman /* 537004f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 537104f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 537204f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 537304f2cbe3SMel Gorman */ 537404f2cbe3SMel Gorman if (ref_page) { 537531d49da5SAneesh Kumar K.V if (page != ref_page) { 537631d49da5SAneesh Kumar K.V spin_unlock(ptl); 537731d49da5SAneesh Kumar K.V continue; 537831d49da5SAneesh Kumar K.V } 537904f2cbe3SMel Gorman /* 538004f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 538104f2cbe3SMel Gorman * future faults in this VMA will fail rather than 538204f2cbe3SMel Gorman * looking like data was lost 538304f2cbe3SMel Gorman */ 538404f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 538504f2cbe3SMel Gorman } 538604f2cbe3SMel Gorman 5387c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 5388b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 5389106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 53906649a386SKen Chen set_page_dirty(page); 539105e90bd0SPeter Xu /* Leave a uffd-wp pte marker if needed */ 539205e90bd0SPeter Xu if (huge_pte_uffd_wp(pte) && 539305e90bd0SPeter Xu !(zap_flags & ZAP_FLAG_DROP_MARKER)) 539405e90bd0SPeter Xu set_huge_pte_at(mm, address, ptep, 539505e90bd0SPeter Xu make_pte_marker(PTE_MARKER_UFFD_WP)); 53965d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 5397cea86fe2SHugh Dickins page_remove_rmap(page, vma, true); 539831d49da5SAneesh Kumar K.V 5399cb900f41SKirill A. Shutemov spin_unlock(ptl); 5400e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 540124669e58SAneesh Kumar K.V /* 540231d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 540324669e58SAneesh Kumar K.V */ 540431d49da5SAneesh Kumar K.V if (ref_page) 540531d49da5SAneesh Kumar K.V break; 5406fe1668aeSChen, Kenneth W } 540724669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 5408a4a118f2SNadav Amit 5409a4a118f2SNadav Amit /* 5410a4a118f2SNadav Amit * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We 5411a4a118f2SNadav Amit * could defer the flush until now, since by holding i_mmap_rwsem we 5412a4a118f2SNadav Amit * guaranteed that the last refernece would not be dropped. But we must 5413a4a118f2SNadav Amit * do the flushing before we return, as otherwise i_mmap_rwsem will be 5414a4a118f2SNadav Amit * dropped and the last reference to the shared PMDs page might be 5415a4a118f2SNadav Amit * dropped as well. 5416a4a118f2SNadav Amit * 5417a4a118f2SNadav Amit * In theory we could defer the freeing of the PMD pages as well, but 5418a4a118f2SNadav Amit * huge_pmd_unshare() relies on the exact page_count for the PMD page to 5419a4a118f2SNadav Amit * detect sharing, so we cannot defer the release of the page either. 5420a4a118f2SNadav Amit * Instead, do flush now. 5421a4a118f2SNadav Amit */ 5422a4a118f2SNadav Amit if (force_flush) 5423a4a118f2SNadav Amit tlb_flush_mmu_tlbonly(tlb); 54241da177e4SLinus Torvalds } 542563551ae0SDavid Gibson 5426d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 5427d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 542805e90bd0SPeter Xu unsigned long end, struct page *ref_page, 542905e90bd0SPeter Xu zap_flags_t zap_flags) 5430d833352aSMel Gorman { 5431131a79b4SMike Kravetz hugetlb_vma_lock_write(vma); 5432131a79b4SMike Kravetz i_mmap_lock_write(vma->vm_file->f_mapping); 5433131a79b4SMike Kravetz 5434369258ceSMike Kravetz /* mmu notification performed in caller */ 543505e90bd0SPeter Xu __unmap_hugepage_range(tlb, vma, start, end, ref_page, zap_flags); 5436d833352aSMel Gorman 543704ada095SMike Kravetz if (zap_flags & ZAP_FLAG_UNMAP) { /* final unmap */ 5438d833352aSMel Gorman /* 543904ada095SMike Kravetz * Unlock and free the vma lock before releasing i_mmap_rwsem. 544004ada095SMike Kravetz * When the vma_lock is freed, this makes the vma ineligible 544104ada095SMike Kravetz * for pmd sharing. And, i_mmap_rwsem is required to set up 544204ada095SMike Kravetz * pmd sharing. This is important as page tables for this 544304ada095SMike Kravetz * unmapped range will be asynchrously deleted. If the page 544404ada095SMike Kravetz * tables are shared, there will be issues when accessed by 544504ada095SMike Kravetz * someone else. 5446d833352aSMel Gorman */ 5447ecfbd733SMike Kravetz __hugetlb_vma_unlock_write_free(vma); 5448131a79b4SMike Kravetz i_mmap_unlock_write(vma->vm_file->f_mapping); 544904ada095SMike Kravetz } else { 545004ada095SMike Kravetz i_mmap_unlock_write(vma->vm_file->f_mapping); 545104ada095SMike Kravetz hugetlb_vma_unlock_write(vma); 545204ada095SMike Kravetz } 5453d833352aSMel Gorman } 5454d833352aSMel Gorman 5455502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 545605e90bd0SPeter Xu unsigned long end, struct page *ref_page, 545705e90bd0SPeter Xu zap_flags_t zap_flags) 5458502717f4SChen, Kenneth W { 5459369258ceSMike Kravetz struct mmu_notifier_range range; 546024669e58SAneesh Kumar K.V struct mmu_gather tlb; 5461dff11abeSMike Kravetz 54627d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm, 5463369258ceSMike Kravetz start, end); 5464369258ceSMike Kravetz adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 5465369258ceSMike Kravetz mmu_notifier_invalidate_range_start(&range); 5466a72afd87SWill Deacon tlb_gather_mmu(&tlb, vma->vm_mm); 5467369258ceSMike Kravetz 546805e90bd0SPeter Xu __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags); 5469369258ceSMike Kravetz 5470369258ceSMike Kravetz mmu_notifier_invalidate_range_end(&range); 5471ae8eba8bSWill Deacon tlb_finish_mmu(&tlb); 5472502717f4SChen, Kenneth W } 5473502717f4SChen, Kenneth W 547404f2cbe3SMel Gorman /* 547504f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 5476578b7725SZhiyuan Dai * mapping it owns the reserve page for. The intention is to unmap the page 547704f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 547804f2cbe3SMel Gorman * same region. 547904f2cbe3SMel Gorman */ 54802f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 54812a4b3dedSHarvey Harrison struct page *page, unsigned long address) 548204f2cbe3SMel Gorman { 54837526674dSAdam Litke struct hstate *h = hstate_vma(vma); 548404f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 548504f2cbe3SMel Gorman struct address_space *mapping; 548604f2cbe3SMel Gorman pgoff_t pgoff; 548704f2cbe3SMel Gorman 548804f2cbe3SMel Gorman /* 548904f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 549004f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 549104f2cbe3SMel Gorman */ 54927526674dSAdam Litke address = address & huge_page_mask(h); 549336e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 549436e4f20aSMichal Hocko vma->vm_pgoff; 549593c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 549604f2cbe3SMel Gorman 54974eb2b1dcSMel Gorman /* 54984eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 54994eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 55004eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 55014eb2b1dcSMel Gorman */ 550283cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 55036b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 550404f2cbe3SMel Gorman /* Do not unmap the current VMA */ 550504f2cbe3SMel Gorman if (iter_vma == vma) 550604f2cbe3SMel Gorman continue; 550704f2cbe3SMel Gorman 550804f2cbe3SMel Gorman /* 55092f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 55102f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 55112f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 55122f84a899SMel Gorman */ 55132f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 55142f84a899SMel Gorman continue; 55152f84a899SMel Gorman 55162f84a899SMel Gorman /* 551704f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 551804f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 551904f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 552004f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 552104f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 552204f2cbe3SMel Gorman */ 552304f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 552424669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 552505e90bd0SPeter Xu address + huge_page_size(h), page, 0); 552604f2cbe3SMel Gorman } 552783cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 552804f2cbe3SMel Gorman } 552904f2cbe3SMel Gorman 55300fe6e20bSNaoya Horiguchi /* 5531c89357e2SDavid Hildenbrand * hugetlb_wp() should be called with page lock of the original hugepage held. 5532aa6d2e8cSBaolin Wang * Called with hugetlb_fault_mutex_table held and pte_page locked so we 5533ef009b25SMichal Hocko * cannot race with other handlers or page migration. 5534ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 55350fe6e20bSNaoya Horiguchi */ 5536c89357e2SDavid Hildenbrand static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma, 5537c89357e2SDavid Hildenbrand unsigned long address, pte_t *ptep, unsigned int flags, 5538371607a3SSidhartha Kumar struct folio *pagecache_folio, spinlock_t *ptl) 55391e8f889bSDavid Gibson { 5540c89357e2SDavid Hildenbrand const bool unshare = flags & FAULT_FLAG_UNSHARE; 554160d5b473SPeter Xu pte_t pte = huge_ptep_get(ptep); 5542a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 5543*959a78b6SZhangPeng struct folio *old_folio; 5544d0ce0e47SSidhartha Kumar struct folio *new_folio; 55452b740303SSouptick Joarder int outside_reserve = 0; 55462b740303SSouptick Joarder vm_fault_t ret = 0; 5547974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 5548ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 55491e8f889bSDavid Gibson 55501d8d1464SDavid Hildenbrand /* 555160d5b473SPeter Xu * Never handle CoW for uffd-wp protected pages. It should be only 555260d5b473SPeter Xu * handled when the uffd-wp protection is removed. 555360d5b473SPeter Xu * 555460d5b473SPeter Xu * Note that only the CoW optimization path (in hugetlb_no_page()) 555560d5b473SPeter Xu * can trigger this, because hugetlb_fault() will always resolve 555660d5b473SPeter Xu * uffd-wp bit first. 555760d5b473SPeter Xu */ 555860d5b473SPeter Xu if (!unshare && huge_pte_uffd_wp(pte)) 555960d5b473SPeter Xu return 0; 556060d5b473SPeter Xu 556160d5b473SPeter Xu /* 55621d8d1464SDavid Hildenbrand * hugetlb does not support FOLL_FORCE-style write faults that keep the 55631d8d1464SDavid Hildenbrand * PTE mapped R/O such as maybe_mkwrite() would do. 55641d8d1464SDavid Hildenbrand */ 55651d8d1464SDavid Hildenbrand if (WARN_ON_ONCE(!unshare && !(vma->vm_flags & VM_WRITE))) 55661d8d1464SDavid Hildenbrand return VM_FAULT_SIGSEGV; 55671d8d1464SDavid Hildenbrand 55681d8d1464SDavid Hildenbrand /* Let's take out MAP_SHARED mappings first. */ 55691d8d1464SDavid Hildenbrand if (vma->vm_flags & VM_MAYSHARE) { 55701d8d1464SDavid Hildenbrand set_huge_ptep_writable(vma, haddr, ptep); 55711d8d1464SDavid Hildenbrand return 0; 55721d8d1464SDavid Hildenbrand } 55731d8d1464SDavid Hildenbrand 5574*959a78b6SZhangPeng old_folio = page_folio(pte_page(pte)); 55751e8f889bSDavid Gibson 5576662ce1dcSYang Yang delayacct_wpcopy_start(); 5577662ce1dcSYang Yang 557804f2cbe3SMel Gorman retry_avoidcopy: 5579c89357e2SDavid Hildenbrand /* 5580c89357e2SDavid Hildenbrand * If no-one else is actually using this page, we're the exclusive 5581c89357e2SDavid Hildenbrand * owner and can reuse this page. 5582c89357e2SDavid Hildenbrand */ 5583*959a78b6SZhangPeng if (folio_mapcount(old_folio) == 1 && folio_test_anon(old_folio)) { 5584*959a78b6SZhangPeng if (!PageAnonExclusive(&old_folio->page)) 5585*959a78b6SZhangPeng page_move_anon_rmap(&old_folio->page, vma); 5586c89357e2SDavid Hildenbrand if (likely(!unshare)) 55875b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 5588662ce1dcSYang Yang 5589662ce1dcSYang Yang delayacct_wpcopy_end(); 559083c54070SNick Piggin return 0; 55911e8f889bSDavid Gibson } 5592*959a78b6SZhangPeng VM_BUG_ON_PAGE(folio_test_anon(old_folio) && 5593*959a78b6SZhangPeng PageAnonExclusive(&old_folio->page), &old_folio->page); 55941e8f889bSDavid Gibson 559504f2cbe3SMel Gorman /* 559604f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 559704f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 559804f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 559904f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 560004f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 560104f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 560204f2cbe3SMel Gorman * of the full address range. 560304f2cbe3SMel Gorman */ 56045944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 5605*959a78b6SZhangPeng old_folio != pagecache_folio) 560604f2cbe3SMel Gorman outside_reserve = 1; 560704f2cbe3SMel Gorman 5608*959a78b6SZhangPeng folio_get(old_folio); 5609b76c8cfbSLarry Woodman 5610ad4404a2SDavidlohr Bueso /* 5611ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 5612ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 5613ad4404a2SDavidlohr Bueso */ 5614cb900f41SKirill A. Shutemov spin_unlock(ptl); 5615d0ce0e47SSidhartha Kumar new_folio = alloc_hugetlb_folio(vma, haddr, outside_reserve); 56161e8f889bSDavid Gibson 5617d0ce0e47SSidhartha Kumar if (IS_ERR(new_folio)) { 561804f2cbe3SMel Gorman /* 561904f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 562004f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 562104f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 562204f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 562304f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 562404f2cbe3SMel Gorman */ 562504f2cbe3SMel Gorman if (outside_reserve) { 562640549ba8SMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 562740549ba8SMike Kravetz pgoff_t idx; 562840549ba8SMike Kravetz u32 hash; 562940549ba8SMike Kravetz 5630*959a78b6SZhangPeng folio_put(old_folio); 563140549ba8SMike Kravetz /* 563240549ba8SMike Kravetz * Drop hugetlb_fault_mutex and vma_lock before 563340549ba8SMike Kravetz * unmapping. unmapping needs to hold vma_lock 563440549ba8SMike Kravetz * in write mode. Dropping vma_lock in read mode 563540549ba8SMike Kravetz * here is OK as COW mappings do not interact with 563640549ba8SMike Kravetz * PMD sharing. 563740549ba8SMike Kravetz * 563840549ba8SMike Kravetz * Reacquire both after unmap operation. 563940549ba8SMike Kravetz */ 564040549ba8SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 564140549ba8SMike Kravetz hash = hugetlb_fault_mutex_hash(mapping, idx); 564240549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 564340549ba8SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 564440549ba8SMike Kravetz 5645*959a78b6SZhangPeng unmap_ref_private(mm, vma, &old_folio->page, haddr); 564640549ba8SMike Kravetz 564740549ba8SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 564840549ba8SMike Kravetz hugetlb_vma_lock_read(vma); 5649cb900f41SKirill A. Shutemov spin_lock(ptl); 56509c67a207SPeter Xu ptep = hugetlb_walk(vma, haddr, huge_page_size(h)); 5651a9af0c5dSNaoya Horiguchi if (likely(ptep && 5652a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 565304f2cbe3SMel Gorman goto retry_avoidcopy; 5654a734bcc8SHillf Danton /* 5655cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 5656cb900f41SKirill A. Shutemov * lock, and our job is done. 5657a734bcc8SHillf Danton */ 5658662ce1dcSYang Yang delayacct_wpcopy_end(); 5659a734bcc8SHillf Danton return 0; 566004f2cbe3SMel Gorman } 566104f2cbe3SMel Gorman 5662d0ce0e47SSidhartha Kumar ret = vmf_error(PTR_ERR(new_folio)); 5663ad4404a2SDavidlohr Bueso goto out_release_old; 56641e8f889bSDavid Gibson } 56651e8f889bSDavid Gibson 56660fe6e20bSNaoya Horiguchi /* 56670fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 56680fe6e20bSNaoya Horiguchi * anon_vma prepared. 56690fe6e20bSNaoya Horiguchi */ 567044e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 5671ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 5672ad4404a2SDavidlohr Bueso goto out_release_all; 567344e2aa93SDean Nelson } 56740fe6e20bSNaoya Horiguchi 5675*959a78b6SZhangPeng if (copy_user_large_folio(new_folio, old_folio, address, vma)) { 56761cb9dc4bSLiu Shixin ret = VM_FAULT_HWPOISON_LARGE; 56771cb9dc4bSLiu Shixin goto out_release_all; 56781cb9dc4bSLiu Shixin } 5679d0ce0e47SSidhartha Kumar __folio_mark_uptodate(new_folio); 56801e8f889bSDavid Gibson 56817d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, haddr, 56826f4f13e8SJérôme Glisse haddr + huge_page_size(h)); 5683ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 5684ad4404a2SDavidlohr Bueso 5685b76c8cfbSLarry Woodman /* 5686cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 5687b76c8cfbSLarry Woodman * before the page tables are altered 5688b76c8cfbSLarry Woodman */ 5689cb900f41SKirill A. Shutemov spin_lock(ptl); 56909c67a207SPeter Xu ptep = hugetlb_walk(vma, haddr, huge_page_size(h)); 5691a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 56920f230bc2SPeter Xu pte_t newpte = make_huge_pte(vma, &new_folio->page, !unshare); 56930f230bc2SPeter Xu 5694c89357e2SDavid Hildenbrand /* Break COW or unshare */ 56955b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 5696ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, range.end); 5697*959a78b6SZhangPeng page_remove_rmap(&old_folio->page, vma, true); 5698d0ce0e47SSidhartha Kumar hugepage_add_new_anon_rmap(new_folio, vma, haddr); 56990f230bc2SPeter Xu if (huge_pte_uffd_wp(pte)) 57000f230bc2SPeter Xu newpte = huge_pte_mkuffd_wp(newpte); 57010f230bc2SPeter Xu set_huge_pte_at(mm, haddr, ptep, newpte); 5702d0ce0e47SSidhartha Kumar folio_set_hugetlb_migratable(new_folio); 57031e8f889bSDavid Gibson /* Make the old page be freed below */ 5704*959a78b6SZhangPeng new_folio = old_folio; 57051e8f889bSDavid Gibson } 5706cb900f41SKirill A. Shutemov spin_unlock(ptl); 5707ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 5708ad4404a2SDavidlohr Bueso out_release_all: 5709c89357e2SDavid Hildenbrand /* 5710c89357e2SDavid Hildenbrand * No restore in case of successful pagetable update (Break COW or 5711c89357e2SDavid Hildenbrand * unshare) 5712c89357e2SDavid Hildenbrand */ 5713*959a78b6SZhangPeng if (new_folio != old_folio) 5714d2d7bb44SSidhartha Kumar restore_reserve_on_error(h, vma, haddr, new_folio); 5715d0ce0e47SSidhartha Kumar folio_put(new_folio); 5716ad4404a2SDavidlohr Bueso out_release_old: 5717*959a78b6SZhangPeng folio_put(old_folio); 57188312034fSJoonsoo Kim 5719ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 5720662ce1dcSYang Yang 5721662ce1dcSYang Yang delayacct_wpcopy_end(); 5722ad4404a2SDavidlohr Bueso return ret; 57231e8f889bSDavid Gibson } 57241e8f889bSDavid Gibson 57253ae77f43SHugh Dickins /* 57263ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 57273ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 57283ae77f43SHugh Dickins */ 57293ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 57302a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 57312a15efc9SHugh Dickins { 573291a2fb95SSidhartha Kumar struct address_space *mapping = vma->vm_file->f_mapping; 573391a2fb95SSidhartha Kumar pgoff_t idx = vma_hugecache_offset(h, vma, address); 573491a2fb95SSidhartha Kumar bool present; 57352a15efc9SHugh Dickins 573691a2fb95SSidhartha Kumar rcu_read_lock(); 573791a2fb95SSidhartha Kumar present = page_cache_next_miss(mapping, idx, 1) != idx; 573891a2fb95SSidhartha Kumar rcu_read_unlock(); 57392a15efc9SHugh Dickins 574091a2fb95SSidhartha Kumar return present; 57412a15efc9SHugh Dickins } 57422a15efc9SHugh Dickins 57439b91c0e2SSidhartha Kumar int hugetlb_add_to_page_cache(struct folio *folio, struct address_space *mapping, 5744ab76ad54SMike Kravetz pgoff_t idx) 5745ab76ad54SMike Kravetz { 5746ab76ad54SMike Kravetz struct inode *inode = mapping->host; 5747ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 5748d9ef44deSMatthew Wilcox (Oracle) int err; 5749ab76ad54SMike Kravetz 5750d9ef44deSMatthew Wilcox (Oracle) __folio_set_locked(folio); 5751d9ef44deSMatthew Wilcox (Oracle) err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL); 5752d9ef44deSMatthew Wilcox (Oracle) 5753d9ef44deSMatthew Wilcox (Oracle) if (unlikely(err)) { 5754d9ef44deSMatthew Wilcox (Oracle) __folio_clear_locked(folio); 5755ab76ad54SMike Kravetz return err; 5756d9ef44deSMatthew Wilcox (Oracle) } 57579b91c0e2SSidhartha Kumar folio_clear_hugetlb_restore_reserve(folio); 5758ab76ad54SMike Kravetz 575922146c3cSMike Kravetz /* 5760d9ef44deSMatthew Wilcox (Oracle) * mark folio dirty so that it will not be removed from cache/file 576122146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 576222146c3cSMike Kravetz */ 5763d9ef44deSMatthew Wilcox (Oracle) folio_mark_dirty(folio); 576422146c3cSMike Kravetz 5765ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 5766ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 5767ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 5768ab76ad54SMike Kravetz return 0; 5769ab76ad54SMike Kravetz } 5770ab76ad54SMike Kravetz 57717677f7fdSAxel Rasmussen static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma, 57727677f7fdSAxel Rasmussen struct address_space *mapping, 57737677f7fdSAxel Rasmussen pgoff_t idx, 57747677f7fdSAxel Rasmussen unsigned int flags, 57757677f7fdSAxel Rasmussen unsigned long haddr, 5776824ddc60SNadav Amit unsigned long addr, 57777677f7fdSAxel Rasmussen unsigned long reason) 57787677f7fdSAxel Rasmussen { 57797677f7fdSAxel Rasmussen u32 hash; 57807677f7fdSAxel Rasmussen struct vm_fault vmf = { 57817677f7fdSAxel Rasmussen .vma = vma, 57827677f7fdSAxel Rasmussen .address = haddr, 5783824ddc60SNadav Amit .real_address = addr, 57847677f7fdSAxel Rasmussen .flags = flags, 57857677f7fdSAxel Rasmussen 57867677f7fdSAxel Rasmussen /* 57877677f7fdSAxel Rasmussen * Hard to debug if it ends up being 57887677f7fdSAxel Rasmussen * used by a callee that assumes 57897677f7fdSAxel Rasmussen * something about the other 57907677f7fdSAxel Rasmussen * uninitialized fields... same as in 57917677f7fdSAxel Rasmussen * memory.c 57927677f7fdSAxel Rasmussen */ 57937677f7fdSAxel Rasmussen }; 57947677f7fdSAxel Rasmussen 57957677f7fdSAxel Rasmussen /* 5796958f32ceSLiu Shixin * vma_lock and hugetlb_fault_mutex must be dropped before handling 5797958f32ceSLiu Shixin * userfault. Also mmap_lock could be dropped due to handling 5798958f32ceSLiu Shixin * userfault, any vma operation should be careful from here. 57997677f7fdSAxel Rasmussen */ 580040549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 58017677f7fdSAxel Rasmussen hash = hugetlb_fault_mutex_hash(mapping, idx); 58027677f7fdSAxel Rasmussen mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5803958f32ceSLiu Shixin return handle_userfault(&vmf, reason); 58047677f7fdSAxel Rasmussen } 58057677f7fdSAxel Rasmussen 58062ea7ff1eSPeter Xu /* 58072ea7ff1eSPeter Xu * Recheck pte with pgtable lock. Returns true if pte didn't change, or 58082ea7ff1eSPeter Xu * false if pte changed or is changing. 58092ea7ff1eSPeter Xu */ 58102ea7ff1eSPeter Xu static bool hugetlb_pte_stable(struct hstate *h, struct mm_struct *mm, 58112ea7ff1eSPeter Xu pte_t *ptep, pte_t old_pte) 58122ea7ff1eSPeter Xu { 58132ea7ff1eSPeter Xu spinlock_t *ptl; 58142ea7ff1eSPeter Xu bool same; 58152ea7ff1eSPeter Xu 58162ea7ff1eSPeter Xu ptl = huge_pte_lock(h, mm, ptep); 58172ea7ff1eSPeter Xu same = pte_same(huge_ptep_get(ptep), old_pte); 58182ea7ff1eSPeter Xu spin_unlock(ptl); 58192ea7ff1eSPeter Xu 58202ea7ff1eSPeter Xu return same; 58212ea7ff1eSPeter Xu } 58222ea7ff1eSPeter Xu 58232b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 58242b740303SSouptick Joarder struct vm_area_struct *vma, 58258382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 5826c64e912cSPeter Xu unsigned long address, pte_t *ptep, 5827c64e912cSPeter Xu pte_t old_pte, unsigned int flags) 5828ac9b9c66SHugh Dickins { 5829a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 58302b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 5831409eb8c2SHillf Danton int anon_rmap = 0; 58324c887265SAdam Litke unsigned long size; 5833d0ce0e47SSidhartha Kumar struct folio *folio; 58341e8f889bSDavid Gibson pte_t new_pte; 5835cb900f41SKirill A. Shutemov spinlock_t *ptl; 5836285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 5837d0ce0e47SSidhartha Kumar bool new_folio, new_pagecache_folio = false; 5838958f32ceSLiu Shixin u32 hash = hugetlb_fault_mutex_hash(mapping, idx); 58394c887265SAdam Litke 584004f2cbe3SMel Gorman /* 584104f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 584204f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 5843c89357e2SDavid Hildenbrand * COW/unsharing. Warn that such a situation has occurred as it may not 5844c89357e2SDavid Hildenbrand * be obvious. 584504f2cbe3SMel Gorman */ 584604f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 5847910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 584804f2cbe3SMel Gorman current->pid); 5849958f32ceSLiu Shixin goto out; 585004f2cbe3SMel Gorman } 585104f2cbe3SMel Gorman 58524c887265SAdam Litke /* 5853188a3972SMike Kravetz * Use page lock to guard against racing truncation 5854188a3972SMike Kravetz * before we get page_table_lock. 58554c887265SAdam Litke */ 5856d0ce0e47SSidhartha Kumar new_folio = false; 5857d0ce0e47SSidhartha Kumar folio = filemap_lock_folio(mapping, idx); 585866dabbb6SChristoph Hellwig if (IS_ERR(folio)) { 5859188a3972SMike Kravetz size = i_size_read(mapping->host) >> huge_page_shift(h); 5860188a3972SMike Kravetz if (idx >= size) 5861188a3972SMike Kravetz goto out; 58627677f7fdSAxel Rasmussen /* Check for page in userfault range */ 58632ea7ff1eSPeter Xu if (userfaultfd_missing(vma)) { 58642ea7ff1eSPeter Xu /* 58652ea7ff1eSPeter Xu * Since hugetlb_no_page() was examining pte 58662ea7ff1eSPeter Xu * without pgtable lock, we need to re-test under 58672ea7ff1eSPeter Xu * lock because the pte may not be stable and could 58682ea7ff1eSPeter Xu * have changed from under us. Try to detect 58692ea7ff1eSPeter Xu * either changed or during-changing ptes and retry 58702ea7ff1eSPeter Xu * properly when needed. 58712ea7ff1eSPeter Xu * 58722ea7ff1eSPeter Xu * Note that userfaultfd is actually fine with 58732ea7ff1eSPeter Xu * false positives (e.g. caused by pte changed), 58742ea7ff1eSPeter Xu * but not wrong logical events (e.g. caused by 58752ea7ff1eSPeter Xu * reading a pte during changing). The latter can 58762ea7ff1eSPeter Xu * confuse the userspace, so the strictness is very 58772ea7ff1eSPeter Xu * much preferred. E.g., MISSING event should 58782ea7ff1eSPeter Xu * never happen on the page after UFFDIO_COPY has 58792ea7ff1eSPeter Xu * correctly installed the page and returned. 58802ea7ff1eSPeter Xu */ 58812ea7ff1eSPeter Xu if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) { 58822ea7ff1eSPeter Xu ret = 0; 58832ea7ff1eSPeter Xu goto out; 58842ea7ff1eSPeter Xu } 58852ea7ff1eSPeter Xu 58862ea7ff1eSPeter Xu return hugetlb_handle_userfault(vma, mapping, idx, flags, 58872ea7ff1eSPeter Xu haddr, address, 58887677f7fdSAxel Rasmussen VM_UFFD_MISSING); 58892ea7ff1eSPeter Xu } 58901a1aad8aSMike Kravetz 5891d0ce0e47SSidhartha Kumar folio = alloc_hugetlb_folio(vma, haddr, 0); 5892d0ce0e47SSidhartha Kumar if (IS_ERR(folio)) { 58934643d67eSMike Kravetz /* 58944643d67eSMike Kravetz * Returning error will result in faulting task being 58954643d67eSMike Kravetz * sent SIGBUS. The hugetlb fault mutex prevents two 58964643d67eSMike Kravetz * tasks from racing to fault in the same page which 58974643d67eSMike Kravetz * could result in false unable to allocate errors. 58984643d67eSMike Kravetz * Page migration does not take the fault mutex, but 58994643d67eSMike Kravetz * does a clear then write of pte's under page table 59004643d67eSMike Kravetz * lock. Page fault code could race with migration, 59014643d67eSMike Kravetz * notice the clear pte and try to allocate a page 59024643d67eSMike Kravetz * here. Before returning error, get ptl and make 59034643d67eSMike Kravetz * sure there really is no pte entry. 59044643d67eSMike Kravetz */ 5905f9bf6c03SPeter Xu if (hugetlb_pte_stable(h, mm, ptep, old_pte)) 5906d0ce0e47SSidhartha Kumar ret = vmf_error(PTR_ERR(folio)); 5907f9bf6c03SPeter Xu else 5908f9bf6c03SPeter Xu ret = 0; 59096bda666aSChristoph Lameter goto out; 59106bda666aSChristoph Lameter } 5911d0ce0e47SSidhartha Kumar clear_huge_page(&folio->page, address, pages_per_huge_page(h)); 5912d0ce0e47SSidhartha Kumar __folio_mark_uptodate(folio); 5913d0ce0e47SSidhartha Kumar new_folio = true; 5914ac9b9c66SHugh Dickins 5915f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 59169b91c0e2SSidhartha Kumar int err = hugetlb_add_to_page_cache(folio, mapping, idx); 59176bda666aSChristoph Lameter if (err) { 59183a5497a2SMiaohe Lin /* 59193a5497a2SMiaohe Lin * err can't be -EEXIST which implies someone 59203a5497a2SMiaohe Lin * else consumed the reservation since hugetlb 59213a5497a2SMiaohe Lin * fault mutex is held when add a hugetlb page 59223a5497a2SMiaohe Lin * to the page cache. So it's safe to call 59233a5497a2SMiaohe Lin * restore_reserve_on_error() here. 59243a5497a2SMiaohe Lin */ 5925d2d7bb44SSidhartha Kumar restore_reserve_on_error(h, vma, haddr, folio); 5926d0ce0e47SSidhartha Kumar folio_put(folio); 59276bda666aSChristoph Lameter goto out; 59286bda666aSChristoph Lameter } 5929d0ce0e47SSidhartha Kumar new_pagecache_folio = true; 593023be7468SMel Gorman } else { 5931d0ce0e47SSidhartha Kumar folio_lock(folio); 59320fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 59330fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 59340fe6e20bSNaoya Horiguchi goto backout_unlocked; 593523be7468SMel Gorman } 5936409eb8c2SHillf Danton anon_rmap = 1; 59370fe6e20bSNaoya Horiguchi } 59380fe6e20bSNaoya Horiguchi } else { 593957303d80SAndy Whitcroft /* 5940998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 5941998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 5942998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 5943fd6a03edSNaoya Horiguchi */ 5944d0ce0e47SSidhartha Kumar if (unlikely(folio_test_hwpoison(folio))) { 59450eb98f15SMiaohe Lin ret = VM_FAULT_HWPOISON_LARGE | 5946972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 5947fd6a03edSNaoya Horiguchi goto backout_unlocked; 59486bda666aSChristoph Lameter } 59497677f7fdSAxel Rasmussen 59507677f7fdSAxel Rasmussen /* Check for page in userfault range. */ 59517677f7fdSAxel Rasmussen if (userfaultfd_minor(vma)) { 5952d0ce0e47SSidhartha Kumar folio_unlock(folio); 5953d0ce0e47SSidhartha Kumar folio_put(folio); 59542ea7ff1eSPeter Xu /* See comment in userfaultfd_missing() block above */ 59552ea7ff1eSPeter Xu if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) { 59562ea7ff1eSPeter Xu ret = 0; 59572ea7ff1eSPeter Xu goto out; 59582ea7ff1eSPeter Xu } 59592ea7ff1eSPeter Xu return hugetlb_handle_userfault(vma, mapping, idx, flags, 59602ea7ff1eSPeter Xu haddr, address, 59617677f7fdSAxel Rasmussen VM_UFFD_MINOR); 59627677f7fdSAxel Rasmussen } 5963998b4382SNaoya Horiguchi } 59641e8f889bSDavid Gibson 596557303d80SAndy Whitcroft /* 596657303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 596757303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 596857303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 596957303d80SAndy Whitcroft * the spinlock. 597057303d80SAndy Whitcroft */ 59715e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 5972285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 59732b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 59742b26736cSAndy Whitcroft goto backout_unlocked; 59752b26736cSAndy Whitcroft } 59765e911373SMike Kravetz /* Just decrements count, does not deallocate */ 5977285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 59785e911373SMike Kravetz } 597957303d80SAndy Whitcroft 59808bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 598183c54070SNick Piggin ret = 0; 5982c64e912cSPeter Xu /* If pte changed from under us, retry */ 5983c64e912cSPeter Xu if (!pte_same(huge_ptep_get(ptep), old_pte)) 59844c887265SAdam Litke goto backout; 59854c887265SAdam Litke 59864781593dSPeter Xu if (anon_rmap) 5987d0ce0e47SSidhartha Kumar hugepage_add_new_anon_rmap(folio, vma, haddr); 59884781593dSPeter Xu else 5989d0ce0e47SSidhartha Kumar page_dup_file_rmap(&folio->page, true); 5990d0ce0e47SSidhartha Kumar new_pte = make_huge_pte(vma, &folio->page, ((vma->vm_flags & VM_WRITE) 59911e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 5992c64e912cSPeter Xu /* 5993c64e912cSPeter Xu * If this pte was previously wr-protected, keep it wr-protected even 5994c64e912cSPeter Xu * if populated. 5995c64e912cSPeter Xu */ 5996c64e912cSPeter Xu if (unlikely(pte_marker_uffd_wp(old_pte))) 5997f1eb1bacSPeter Xu new_pte = huge_pte_mkuffd_wp(new_pte); 5998285b8dcaSHuang Ying set_huge_pte_at(mm, haddr, ptep, new_pte); 59991e8f889bSDavid Gibson 60005d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 6001788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 60021e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 6003371607a3SSidhartha Kumar ret = hugetlb_wp(mm, vma, address, ptep, flags, folio, ptl); 60041e8f889bSDavid Gibson } 60051e8f889bSDavid Gibson 6006cb900f41SKirill A. Shutemov spin_unlock(ptl); 6007cb6acd01SMike Kravetz 6008cb6acd01SMike Kravetz /* 6009d0ce0e47SSidhartha Kumar * Only set hugetlb_migratable in newly allocated pages. Existing pages 6010d0ce0e47SSidhartha Kumar * found in the pagecache may not have hugetlb_migratable if they have 60118f251a3dSMike Kravetz * been isolated for migration. 6012cb6acd01SMike Kravetz */ 6013d0ce0e47SSidhartha Kumar if (new_folio) 6014d0ce0e47SSidhartha Kumar folio_set_hugetlb_migratable(folio); 6015cb6acd01SMike Kravetz 6016d0ce0e47SSidhartha Kumar folio_unlock(folio); 60174c887265SAdam Litke out: 6018958f32ceSLiu Shixin hugetlb_vma_unlock_read(vma); 6019958f32ceSLiu Shixin mutex_unlock(&hugetlb_fault_mutex_table[hash]); 6020ac9b9c66SHugh Dickins return ret; 60214c887265SAdam Litke 60224c887265SAdam Litke backout: 6023cb900f41SKirill A. Shutemov spin_unlock(ptl); 60242b26736cSAndy Whitcroft backout_unlocked: 6025d0ce0e47SSidhartha Kumar if (new_folio && !new_pagecache_folio) 6026d2d7bb44SSidhartha Kumar restore_reserve_on_error(h, vma, haddr, folio); 6027fa27759aSMike Kravetz 6028d0ce0e47SSidhartha Kumar folio_unlock(folio); 6029d0ce0e47SSidhartha Kumar folio_put(folio); 60304c887265SAdam Litke goto out; 6031ac9b9c66SHugh Dickins } 6032ac9b9c66SHugh Dickins 60338382d914SDavidlohr Bueso #ifdef CONFIG_SMP 6034188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 60358382d914SDavidlohr Bueso { 60368382d914SDavidlohr Bueso unsigned long key[2]; 60378382d914SDavidlohr Bueso u32 hash; 60388382d914SDavidlohr Bueso 60398382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 60408382d914SDavidlohr Bueso key[1] = idx; 60418382d914SDavidlohr Bueso 604255254636SMike Kravetz hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); 60438382d914SDavidlohr Bueso 60448382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 60458382d914SDavidlohr Bueso } 60468382d914SDavidlohr Bueso #else 60478382d914SDavidlohr Bueso /* 60486c26d310SMiaohe Lin * For uniprocessor systems we always use a single mutex, so just 60498382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 60508382d914SDavidlohr Bueso */ 6051188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 60528382d914SDavidlohr Bueso { 60538382d914SDavidlohr Bueso return 0; 60548382d914SDavidlohr Bueso } 60558382d914SDavidlohr Bueso #endif 60568382d914SDavidlohr Bueso 60572b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 6058788c7df4SHugh Dickins unsigned long address, unsigned int flags) 605986e5216fSAdam Litke { 60608382d914SDavidlohr Bueso pte_t *ptep, entry; 6061cb900f41SKirill A. Shutemov spinlock_t *ptl; 60622b740303SSouptick Joarder vm_fault_t ret; 60638382d914SDavidlohr Bueso u32 hash; 60648382d914SDavidlohr Bueso pgoff_t idx; 60650fe6e20bSNaoya Horiguchi struct page *page = NULL; 6066371607a3SSidhartha Kumar struct folio *pagecache_folio = NULL; 6067a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 60688382d914SDavidlohr Bueso struct address_space *mapping; 60690f792cf9SNaoya Horiguchi int need_wait_lock = 0; 6070285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 607186e5216fSAdam Litke 60723935baa9SDavid Gibson /* 60733935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 60743935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 60753935baa9SDavid Gibson * the same page in the page cache. 60763935baa9SDavid Gibson */ 607740549ba8SMike Kravetz mapping = vma->vm_file->f_mapping; 607840549ba8SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 6079188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 6080c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 60818382d914SDavidlohr Bueso 608240549ba8SMike Kravetz /* 608340549ba8SMike Kravetz * Acquire vma lock before calling huge_pte_alloc and hold 608440549ba8SMike Kravetz * until finished with ptep. This prevents huge_pmd_unshare from 608540549ba8SMike Kravetz * being called elsewhere and making the ptep no longer valid. 608640549ba8SMike Kravetz */ 608740549ba8SMike Kravetz hugetlb_vma_lock_read(vma); 608840549ba8SMike Kravetz ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); 608940549ba8SMike Kravetz if (!ptep) { 609040549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 609140549ba8SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 609240549ba8SMike Kravetz return VM_FAULT_OOM; 609340549ba8SMike Kravetz } 609440549ba8SMike Kravetz 60957f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 6096c64e912cSPeter Xu /* PTE markers should be handled the same way as none pte */ 6097958f32ceSLiu Shixin if (huge_pte_none_mostly(entry)) 6098958f32ceSLiu Shixin /* 6099958f32ceSLiu Shixin * hugetlb_no_page will drop vma lock and hugetlb fault 6100958f32ceSLiu Shixin * mutex internally, which make us return immediately. 6101958f32ceSLiu Shixin */ 6102958f32ceSLiu Shixin return hugetlb_no_page(mm, vma, mapping, idx, address, ptep, 6103c64e912cSPeter Xu entry, flags); 610486e5216fSAdam Litke 610583c54070SNick Piggin ret = 0; 61061e8f889bSDavid Gibson 610757303d80SAndy Whitcroft /* 61080f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 61090f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 61107c8de358SEthon Paul * an active hugepage in pagecache. This goto expects the 2nd page 61117c8de358SEthon Paul * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will 61127c8de358SEthon Paul * properly handle it. 61130f792cf9SNaoya Horiguchi */ 6114fcd48540SPeter Xu if (!pte_present(entry)) { 6115fcd48540SPeter Xu if (unlikely(is_hugetlb_entry_migration(entry))) { 6116fcd48540SPeter Xu /* 6117fcd48540SPeter Xu * Release the hugetlb fault lock now, but retain 6118fcd48540SPeter Xu * the vma lock, because it is needed to guard the 6119fcd48540SPeter Xu * huge_pte_lockptr() later in 6120fcd48540SPeter Xu * migration_entry_wait_huge(). The vma lock will 6121fcd48540SPeter Xu * be released there. 6122fcd48540SPeter Xu */ 6123fcd48540SPeter Xu mutex_unlock(&hugetlb_fault_mutex_table[hash]); 6124fcd48540SPeter Xu migration_entry_wait_huge(vma, ptep); 6125fcd48540SPeter Xu return 0; 6126fcd48540SPeter Xu } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 6127fcd48540SPeter Xu ret = VM_FAULT_HWPOISON_LARGE | 6128fcd48540SPeter Xu VM_FAULT_SET_HINDEX(hstate_index(h)); 61290f792cf9SNaoya Horiguchi goto out_mutex; 6130fcd48540SPeter Xu } 61310f792cf9SNaoya Horiguchi 61320f792cf9SNaoya Horiguchi /* 6133c89357e2SDavid Hildenbrand * If we are going to COW/unshare the mapping later, we examine the 6134c89357e2SDavid Hildenbrand * pending reservations for this page now. This will ensure that any 613557303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 61361d8d1464SDavid Hildenbrand * spinlock. Also lookup the pagecache page now as it is used to 61371d8d1464SDavid Hildenbrand * determine if a reservation has been consumed. 613857303d80SAndy Whitcroft */ 6139c89357e2SDavid Hildenbrand if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) && 61401d8d1464SDavid Hildenbrand !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(entry)) { 6141285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 61422b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 6143b4d1d99fSDavid Gibson goto out_mutex; 61442b26736cSAndy Whitcroft } 61455e911373SMike Kravetz /* Just decrements count, does not deallocate */ 6146285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 614757303d80SAndy Whitcroft 6148371607a3SSidhartha Kumar pagecache_folio = filemap_lock_folio(mapping, idx); 614966dabbb6SChristoph Hellwig if (IS_ERR(pagecache_folio)) 615066dabbb6SChristoph Hellwig pagecache_folio = NULL; 615157303d80SAndy Whitcroft } 615257303d80SAndy Whitcroft 61530f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 61540fe6e20bSNaoya Horiguchi 6155c89357e2SDavid Hildenbrand /* Check for a racing update before calling hugetlb_wp() */ 6156b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 6157cb900f41SKirill A. Shutemov goto out_ptl; 6158b4d1d99fSDavid Gibson 6159166f3eccSPeter Xu /* Handle userfault-wp first, before trying to lock more pages */ 6160166f3eccSPeter Xu if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) && 6161166f3eccSPeter Xu (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 6162166f3eccSPeter Xu struct vm_fault vmf = { 6163166f3eccSPeter Xu .vma = vma, 6164166f3eccSPeter Xu .address = haddr, 6165166f3eccSPeter Xu .real_address = address, 6166166f3eccSPeter Xu .flags = flags, 6167166f3eccSPeter Xu }; 6168166f3eccSPeter Xu 6169166f3eccSPeter Xu spin_unlock(ptl); 6170371607a3SSidhartha Kumar if (pagecache_folio) { 6171371607a3SSidhartha Kumar folio_unlock(pagecache_folio); 6172371607a3SSidhartha Kumar folio_put(pagecache_folio); 6173166f3eccSPeter Xu } 617440549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 6175166f3eccSPeter Xu mutex_unlock(&hugetlb_fault_mutex_table[hash]); 6176166f3eccSPeter Xu return handle_userfault(&vmf, VM_UFFD_WP); 6177166f3eccSPeter Xu } 6178166f3eccSPeter Xu 61790f792cf9SNaoya Horiguchi /* 6180c89357e2SDavid Hildenbrand * hugetlb_wp() requires page locks of pte_page(entry) and 6181371607a3SSidhartha Kumar * pagecache_folio, so here we need take the former one 6182371607a3SSidhartha Kumar * when page != pagecache_folio or !pagecache_folio. 61830f792cf9SNaoya Horiguchi */ 61840f792cf9SNaoya Horiguchi page = pte_page(entry); 6185371607a3SSidhartha Kumar if (page_folio(page) != pagecache_folio) 61860f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 61870f792cf9SNaoya Horiguchi need_wait_lock = 1; 61880f792cf9SNaoya Horiguchi goto out_ptl; 61890f792cf9SNaoya Horiguchi } 61900f792cf9SNaoya Horiguchi 61910f792cf9SNaoya Horiguchi get_page(page); 6192b4d1d99fSDavid Gibson 6193c89357e2SDavid Hildenbrand if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) { 6194106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 6195c89357e2SDavid Hildenbrand ret = hugetlb_wp(mm, vma, address, ptep, flags, 6196371607a3SSidhartha Kumar pagecache_folio, ptl); 61970f792cf9SNaoya Horiguchi goto out_put_page; 6198c89357e2SDavid Hildenbrand } else if (likely(flags & FAULT_FLAG_WRITE)) { 6199106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 6200b4d1d99fSDavid Gibson } 6201c89357e2SDavid Hildenbrand } 6202b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 6203285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 6204788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 6205285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 62060f792cf9SNaoya Horiguchi out_put_page: 6207371607a3SSidhartha Kumar if (page_folio(page) != pagecache_folio) 62080f792cf9SNaoya Horiguchi unlock_page(page); 62090f792cf9SNaoya Horiguchi put_page(page); 6210cb900f41SKirill A. Shutemov out_ptl: 6211cb900f41SKirill A. Shutemov spin_unlock(ptl); 621257303d80SAndy Whitcroft 6213371607a3SSidhartha Kumar if (pagecache_folio) { 6214371607a3SSidhartha Kumar folio_unlock(pagecache_folio); 6215371607a3SSidhartha Kumar folio_put(pagecache_folio); 621657303d80SAndy Whitcroft } 6217b4d1d99fSDavid Gibson out_mutex: 621840549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 6219c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 62200f792cf9SNaoya Horiguchi /* 62210f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 62220f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 62230f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 62240f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 62250f792cf9SNaoya Horiguchi * here without taking refcount. 62260f792cf9SNaoya Horiguchi */ 62270f792cf9SNaoya Horiguchi if (need_wait_lock) 62280f792cf9SNaoya Horiguchi wait_on_page_locked(page); 62291e8f889bSDavid Gibson return ret; 623086e5216fSAdam Litke } 623186e5216fSAdam Litke 6232714c1891SAxel Rasmussen #ifdef CONFIG_USERFAULTFD 62338fb5debcSMike Kravetz /* 6234a734991cSAxel Rasmussen * Used by userfaultfd UFFDIO_* ioctls. Based on userfaultfd's mfill_atomic_pte 6235a734991cSAxel Rasmussen * with modifications for hugetlb pages. 62368fb5debcSMike Kravetz */ 623761c50040SAxel Rasmussen int hugetlb_mfill_atomic_pte(pte_t *dst_pte, 62388fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 62398fb5debcSMike Kravetz unsigned long dst_addr, 62408fb5debcSMike Kravetz unsigned long src_addr, 6241d9712937SAxel Rasmussen uffd_flags_t flags, 62420169fd51SZhangPeng struct folio **foliop) 62438fb5debcSMike Kravetz { 624461c50040SAxel Rasmussen struct mm_struct *dst_mm = dst_vma->vm_mm; 6245d9712937SAxel Rasmussen bool is_continue = uffd_flags_mode_is(flags, MFILL_ATOMIC_CONTINUE); 6246d9712937SAxel Rasmussen bool wp_enabled = (flags & MFILL_ATOMIC_WP); 62478cc5fcbbSMina Almasry struct hstate *h = hstate_vma(dst_vma); 62488cc5fcbbSMina Almasry struct address_space *mapping = dst_vma->vm_file->f_mapping; 62498cc5fcbbSMina Almasry pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr); 62501e392147SAndrea Arcangeli unsigned long size; 62511c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 62528fb5debcSMike Kravetz pte_t _dst_pte; 62538fb5debcSMike Kravetz spinlock_t *ptl; 62548cc5fcbbSMina Almasry int ret = -ENOMEM; 6255d0ce0e47SSidhartha Kumar struct folio *folio; 6256f6191471SAxel Rasmussen int writable; 6257d0ce0e47SSidhartha Kumar bool folio_in_pagecache = false; 62588fb5debcSMike Kravetz 6259f6191471SAxel Rasmussen if (is_continue) { 6260f6191471SAxel Rasmussen ret = -EFAULT; 6261d0ce0e47SSidhartha Kumar folio = filemap_lock_folio(mapping, idx); 626266dabbb6SChristoph Hellwig if (IS_ERR(folio)) 6263f6191471SAxel Rasmussen goto out; 6264d0ce0e47SSidhartha Kumar folio_in_pagecache = true; 62650169fd51SZhangPeng } else if (!*foliop) { 62660169fd51SZhangPeng /* If a folio already exists, then it's UFFDIO_COPY for 6267d84cf06eSMina Almasry * a non-missing case. Return -EEXIST. 6268d84cf06eSMina Almasry */ 6269d84cf06eSMina Almasry if (vm_shared && 6270d84cf06eSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 6271d84cf06eSMina Almasry ret = -EEXIST; 62728fb5debcSMike Kravetz goto out; 6273d84cf06eSMina Almasry } 6274d84cf06eSMina Almasry 6275d0ce0e47SSidhartha Kumar folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0); 6276d0ce0e47SSidhartha Kumar if (IS_ERR(folio)) { 6277d84cf06eSMina Almasry ret = -ENOMEM; 6278d84cf06eSMina Almasry goto out; 6279d84cf06eSMina Almasry } 62808fb5debcSMike Kravetz 6281e87340caSZhangPeng ret = copy_folio_from_user(folio, (const void __user *) src_addr, 6282e87340caSZhangPeng false); 62838fb5debcSMike Kravetz 6284c1e8d7c6SMichel Lespinasse /* fallback to copy_from_user outside mmap_lock */ 62858fb5debcSMike Kravetz if (unlikely(ret)) { 62869e368259SAndrea Arcangeli ret = -ENOENT; 6287d0ce0e47SSidhartha Kumar /* Free the allocated folio which may have 62888cc5fcbbSMina Almasry * consumed a reservation. 62898cc5fcbbSMina Almasry */ 6290d2d7bb44SSidhartha Kumar restore_reserve_on_error(h, dst_vma, dst_addr, folio); 6291d0ce0e47SSidhartha Kumar folio_put(folio); 62928cc5fcbbSMina Almasry 6293d0ce0e47SSidhartha Kumar /* Allocate a temporary folio to hold the copied 62948cc5fcbbSMina Almasry * contents. 62958cc5fcbbSMina Almasry */ 6296d0ce0e47SSidhartha Kumar folio = alloc_hugetlb_folio_vma(h, dst_vma, dst_addr); 6297d0ce0e47SSidhartha Kumar if (!folio) { 62988cc5fcbbSMina Almasry ret = -ENOMEM; 62998cc5fcbbSMina Almasry goto out; 63008cc5fcbbSMina Almasry } 63010169fd51SZhangPeng *foliop = folio; 63020169fd51SZhangPeng /* Set the outparam foliop and return to the caller to 63038cc5fcbbSMina Almasry * copy the contents outside the lock. Don't free the 63040169fd51SZhangPeng * folio. 63058cc5fcbbSMina Almasry */ 63068fb5debcSMike Kravetz goto out; 63078fb5debcSMike Kravetz } 63088fb5debcSMike Kravetz } else { 63098cc5fcbbSMina Almasry if (vm_shared && 63108cc5fcbbSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 63110169fd51SZhangPeng folio_put(*foliop); 63128cc5fcbbSMina Almasry ret = -EEXIST; 63130169fd51SZhangPeng *foliop = NULL; 63148cc5fcbbSMina Almasry goto out; 63158cc5fcbbSMina Almasry } 63168cc5fcbbSMina Almasry 6317d0ce0e47SSidhartha Kumar folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0); 6318d0ce0e47SSidhartha Kumar if (IS_ERR(folio)) { 63190169fd51SZhangPeng folio_put(*foliop); 63208cc5fcbbSMina Almasry ret = -ENOMEM; 63210169fd51SZhangPeng *foliop = NULL; 63228cc5fcbbSMina Almasry goto out; 63238cc5fcbbSMina Almasry } 63241cb9dc4bSLiu Shixin ret = copy_user_large_folio(folio, *foliop, dst_addr, dst_vma); 63250169fd51SZhangPeng folio_put(*foliop); 63260169fd51SZhangPeng *foliop = NULL; 63271cb9dc4bSLiu Shixin if (ret) { 63281cb9dc4bSLiu Shixin folio_put(folio); 63291cb9dc4bSLiu Shixin goto out; 63301cb9dc4bSLiu Shixin } 63318fb5debcSMike Kravetz } 63328fb5debcSMike Kravetz 63338fb5debcSMike Kravetz /* 6334d0ce0e47SSidhartha Kumar * The memory barrier inside __folio_mark_uptodate makes sure that 63358fb5debcSMike Kravetz * preceding stores to the page contents become visible before 63368fb5debcSMike Kravetz * the set_pte_at() write. 63378fb5debcSMike Kravetz */ 6338d0ce0e47SSidhartha Kumar __folio_mark_uptodate(folio); 63398fb5debcSMike Kravetz 6340f6191471SAxel Rasmussen /* Add shared, newly allocated pages to the page cache. */ 6341f6191471SAxel Rasmussen if (vm_shared && !is_continue) { 63421e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 63431e392147SAndrea Arcangeli ret = -EFAULT; 63441e392147SAndrea Arcangeli if (idx >= size) 63451e392147SAndrea Arcangeli goto out_release_nounlock; 63461c9e8defSMike Kravetz 63471e392147SAndrea Arcangeli /* 63481e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 63497e1813d4SMike Kravetz * hugetlb_add_to_page_cache() below happens through the 63501e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 63511e392147SAndrea Arcangeli * the caller. 63521e392147SAndrea Arcangeli */ 63539b91c0e2SSidhartha Kumar ret = hugetlb_add_to_page_cache(folio, mapping, idx); 63541c9e8defSMike Kravetz if (ret) 63551c9e8defSMike Kravetz goto out_release_nounlock; 6356d0ce0e47SSidhartha Kumar folio_in_pagecache = true; 63571c9e8defSMike Kravetz } 63581c9e8defSMike Kravetz 6359bcc66543SMiaohe Lin ptl = huge_pte_lock(h, dst_mm, dst_pte); 63608fb5debcSMike Kravetz 63618625147cSJames Houghton ret = -EIO; 6362d0ce0e47SSidhartha Kumar if (folio_test_hwpoison(folio)) 63638625147cSJames Houghton goto out_release_unlock; 63648625147cSJames Houghton 63651e392147SAndrea Arcangeli /* 63666041c691SPeter Xu * We allow to overwrite a pte marker: consider when both MISSING|WP 63676041c691SPeter Xu * registered, we firstly wr-protect a none pte which has no page cache 63686041c691SPeter Xu * page backing it, then access the page. 63696041c691SPeter Xu */ 6370fa27759aSMike Kravetz ret = -EEXIST; 63716041c691SPeter Xu if (!huge_pte_none_mostly(huge_ptep_get(dst_pte))) 63728fb5debcSMike Kravetz goto out_release_unlock; 63738fb5debcSMike Kravetz 6374d0ce0e47SSidhartha Kumar if (folio_in_pagecache) 6375d0ce0e47SSidhartha Kumar page_dup_file_rmap(&folio->page, true); 63764781593dSPeter Xu else 6377d0ce0e47SSidhartha Kumar hugepage_add_new_anon_rmap(folio, dst_vma, dst_addr); 63788fb5debcSMike Kravetz 63796041c691SPeter Xu /* 63806041c691SPeter Xu * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY 63816041c691SPeter Xu * with wp flag set, don't set pte write bit. 63826041c691SPeter Xu */ 6383d9712937SAxel Rasmussen if (wp_enabled || (is_continue && !vm_shared)) 6384f6191471SAxel Rasmussen writable = 0; 6385f6191471SAxel Rasmussen else 6386f6191471SAxel Rasmussen writable = dst_vma->vm_flags & VM_WRITE; 6387f6191471SAxel Rasmussen 6388d0ce0e47SSidhartha Kumar _dst_pte = make_huge_pte(dst_vma, &folio->page, writable); 63896041c691SPeter Xu /* 63906041c691SPeter Xu * Always mark UFFDIO_COPY page dirty; note that this may not be 63916041c691SPeter Xu * extremely important for hugetlbfs for now since swapping is not 63926041c691SPeter Xu * supported, but we should still be clear in that this page cannot be 63936041c691SPeter Xu * thrown away at will, even if write bit not set. 63946041c691SPeter Xu */ 63958fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 63968fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 63978fb5debcSMike Kravetz 6398d9712937SAxel Rasmussen if (wp_enabled) 63996041c691SPeter Xu _dst_pte = huge_pte_mkuffd_wp(_dst_pte); 64006041c691SPeter Xu 64018fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 64028fb5debcSMike Kravetz 64038fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 64048fb5debcSMike Kravetz 64058fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 64068fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 64078fb5debcSMike Kravetz 64088fb5debcSMike Kravetz spin_unlock(ptl); 6409f6191471SAxel Rasmussen if (!is_continue) 6410d0ce0e47SSidhartha Kumar folio_set_hugetlb_migratable(folio); 6411f6191471SAxel Rasmussen if (vm_shared || is_continue) 6412d0ce0e47SSidhartha Kumar folio_unlock(folio); 64138fb5debcSMike Kravetz ret = 0; 64148fb5debcSMike Kravetz out: 64158fb5debcSMike Kravetz return ret; 64168fb5debcSMike Kravetz out_release_unlock: 64178fb5debcSMike Kravetz spin_unlock(ptl); 6418f6191471SAxel Rasmussen if (vm_shared || is_continue) 6419d0ce0e47SSidhartha Kumar folio_unlock(folio); 64205af10dfdSAndrea Arcangeli out_release_nounlock: 6421d0ce0e47SSidhartha Kumar if (!folio_in_pagecache) 6422d2d7bb44SSidhartha Kumar restore_reserve_on_error(h, dst_vma, dst_addr, folio); 6423d0ce0e47SSidhartha Kumar folio_put(folio); 64248fb5debcSMike Kravetz goto out; 64258fb5debcSMike Kravetz } 6426714c1891SAxel Rasmussen #endif /* CONFIG_USERFAULTFD */ 64278fb5debcSMike Kravetz 6428b2cac248SLorenzo Stoakes static void record_subpages(struct page *page, struct vm_area_struct *vma, 6429b2cac248SLorenzo Stoakes int refs, struct page **pages) 643082e5d378SJoao Martins { 643182e5d378SJoao Martins int nr; 643282e5d378SJoao Martins 643382e5d378SJoao Martins for (nr = 0; nr < refs; nr++) { 643482e5d378SJoao Martins if (likely(pages)) 643514455eabSCheng Li pages[nr] = nth_page(page, nr); 643682e5d378SJoao Martins } 643782e5d378SJoao Martins } 643882e5d378SJoao Martins 643984209e87SDavid Hildenbrand static inline bool __follow_hugetlb_must_fault(struct vm_area_struct *vma, 644084209e87SDavid Hildenbrand unsigned int flags, pte_t *pte, 6441a7f22660SDavid Hildenbrand bool *unshare) 6442a7f22660SDavid Hildenbrand { 6443a7f22660SDavid Hildenbrand pte_t pteval = huge_ptep_get(pte); 6444a7f22660SDavid Hildenbrand 6445a7f22660SDavid Hildenbrand *unshare = false; 6446a7f22660SDavid Hildenbrand if (is_swap_pte(pteval)) 6447a7f22660SDavid Hildenbrand return true; 6448a7f22660SDavid Hildenbrand if (huge_pte_write(pteval)) 6449a7f22660SDavid Hildenbrand return false; 6450a7f22660SDavid Hildenbrand if (flags & FOLL_WRITE) 6451a7f22660SDavid Hildenbrand return true; 645284209e87SDavid Hildenbrand if (gup_must_unshare(vma, flags, pte_page(pteval))) { 6453a7f22660SDavid Hildenbrand *unshare = true; 6454a7f22660SDavid Hildenbrand return true; 6455a7f22660SDavid Hildenbrand } 6456a7f22660SDavid Hildenbrand return false; 6457a7f22660SDavid Hildenbrand } 6458a7f22660SDavid Hildenbrand 645957a196a5SMike Kravetz struct page *hugetlb_follow_page_mask(struct vm_area_struct *vma, 646057a196a5SMike Kravetz unsigned long address, unsigned int flags) 646157a196a5SMike Kravetz { 646257a196a5SMike Kravetz struct hstate *h = hstate_vma(vma); 646357a196a5SMike Kravetz struct mm_struct *mm = vma->vm_mm; 646457a196a5SMike Kravetz unsigned long haddr = address & huge_page_mask(h); 646557a196a5SMike Kravetz struct page *page = NULL; 646657a196a5SMike Kravetz spinlock_t *ptl; 646757a196a5SMike Kravetz pte_t *pte, entry; 646857a196a5SMike Kravetz 646957a196a5SMike Kravetz /* 647057a196a5SMike Kravetz * FOLL_PIN is not supported for follow_page(). Ordinary GUP goes via 647157a196a5SMike Kravetz * follow_hugetlb_page(). 647257a196a5SMike Kravetz */ 647357a196a5SMike Kravetz if (WARN_ON_ONCE(flags & FOLL_PIN)) 647457a196a5SMike Kravetz return NULL; 647557a196a5SMike Kravetz 64767d049f3aSPeter Xu hugetlb_vma_lock_read(vma); 64779c67a207SPeter Xu pte = hugetlb_walk(vma, haddr, huge_page_size(h)); 647857a196a5SMike Kravetz if (!pte) 64797d049f3aSPeter Xu goto out_unlock; 648057a196a5SMike Kravetz 648157a196a5SMike Kravetz ptl = huge_pte_lock(h, mm, pte); 648257a196a5SMike Kravetz entry = huge_ptep_get(pte); 648357a196a5SMike Kravetz if (pte_present(entry)) { 648457a196a5SMike Kravetz page = pte_page(entry) + 648557a196a5SMike Kravetz ((address & ~huge_page_mask(h)) >> PAGE_SHIFT); 648657a196a5SMike Kravetz /* 648757a196a5SMike Kravetz * Note that page may be a sub-page, and with vmemmap 648857a196a5SMike Kravetz * optimizations the page struct may be read only. 648957a196a5SMike Kravetz * try_grab_page() will increase the ref count on the 649057a196a5SMike Kravetz * head page, so this will be OK. 649157a196a5SMike Kravetz * 6492e2ca6ba6SLinus Torvalds * try_grab_page() should always be able to get the page here, 6493e2ca6ba6SLinus Torvalds * because we hold the ptl lock and have verified pte_present(). 649457a196a5SMike Kravetz */ 6495e2ca6ba6SLinus Torvalds if (try_grab_page(page, flags)) { 649657a196a5SMike Kravetz page = NULL; 649757a196a5SMike Kravetz goto out; 649857a196a5SMike Kravetz } 649957a196a5SMike Kravetz } 650057a196a5SMike Kravetz out: 650157a196a5SMike Kravetz spin_unlock(ptl); 65027d049f3aSPeter Xu out_unlock: 65037d049f3aSPeter Xu hugetlb_vma_unlock_read(vma); 650457a196a5SMike Kravetz return page; 650557a196a5SMike Kravetz } 650657a196a5SMike Kravetz 650728a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 6508b2cac248SLorenzo Stoakes struct page **pages, unsigned long *position, 6509b2cac248SLorenzo Stoakes unsigned long *nr_pages, long i, unsigned int flags, 6510b2cac248SLorenzo Stoakes int *locked) 651163551ae0SDavid Gibson { 6512d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 6513d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 651428a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 6515a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 65160fa5bc40SJoao Martins int err = -EFAULT, refs; 651763551ae0SDavid Gibson 651863551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 651963551ae0SDavid Gibson pte_t *pte; 6520cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 6521a7f22660SDavid Hildenbrand bool unshare = false; 65222a15efc9SHugh Dickins int absent; 652363551ae0SDavid Gibson struct page *page; 652463551ae0SDavid Gibson 65254c887265SAdam Litke /* 652602057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 652702057967SDavid Rientjes * potentially allocating memory. 652802057967SDavid Rientjes */ 6529fa45f116SDavidlohr Bueso if (fatal_signal_pending(current)) { 653002057967SDavid Rientjes remainder = 0; 653102057967SDavid Rientjes break; 653202057967SDavid Rientjes } 653302057967SDavid Rientjes 6534eefc7fa5SPeter Xu hugetlb_vma_lock_read(vma); 653502057967SDavid Rientjes /* 65364c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 65372a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 65384c887265SAdam Litke * first, for the page indexing below to work. 6539cb900f41SKirill A. Shutemov * 6540cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 65414c887265SAdam Litke */ 65429c67a207SPeter Xu pte = hugetlb_walk(vma, vaddr & huge_page_mask(h), 65437868a208SPunit Agrawal huge_page_size(h)); 6544cb900f41SKirill A. Shutemov if (pte) 6545cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 65462a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 654763551ae0SDavid Gibson 65482a15efc9SHugh Dickins /* 65492a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 65503ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 65513ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 65523ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 65533ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 65542a15efc9SHugh Dickins */ 65553ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 65563ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 6557cb900f41SKirill A. Shutemov if (pte) 6558cb900f41SKirill A. Shutemov spin_unlock(ptl); 6559eefc7fa5SPeter Xu hugetlb_vma_unlock_read(vma); 65602a15efc9SHugh Dickins remainder = 0; 65612a15efc9SHugh Dickins break; 65622a15efc9SHugh Dickins } 65632a15efc9SHugh Dickins 65649cc3a5bdSNaoya Horiguchi /* 65659cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 65669cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 65679cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 65689cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 65699cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 65709cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 65719cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 65729cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 65739cc3a5bdSNaoya Horiguchi */ 6574a7f22660SDavid Hildenbrand if (absent || 657584209e87SDavid Hildenbrand __follow_hugetlb_must_fault(vma, flags, pte, &unshare)) { 65762b740303SSouptick Joarder vm_fault_t ret; 657787ffc118SAndrea Arcangeli unsigned int fault_flags = 0; 65784c887265SAdam Litke 6579cb900f41SKirill A. Shutemov if (pte) 6580cb900f41SKirill A. Shutemov spin_unlock(ptl); 6581eefc7fa5SPeter Xu hugetlb_vma_unlock_read(vma); 6582eefc7fa5SPeter Xu 658387ffc118SAndrea Arcangeli if (flags & FOLL_WRITE) 658487ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_WRITE; 6585a7f22660SDavid Hildenbrand else if (unshare) 6586a7f22660SDavid Hildenbrand fault_flags |= FAULT_FLAG_UNSHARE; 658793c5c61dSPeter Xu if (locked) { 658871335f37SPeter Xu fault_flags |= FAULT_FLAG_ALLOW_RETRY | 658971335f37SPeter Xu FAULT_FLAG_KILLABLE; 659093c5c61dSPeter Xu if (flags & FOLL_INTERRUPTIBLE) 659193c5c61dSPeter Xu fault_flags |= FAULT_FLAG_INTERRUPTIBLE; 659293c5c61dSPeter Xu } 659387ffc118SAndrea Arcangeli if (flags & FOLL_NOWAIT) 659487ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY | 659587ffc118SAndrea Arcangeli FAULT_FLAG_RETRY_NOWAIT; 659687ffc118SAndrea Arcangeli if (flags & FOLL_TRIED) { 65974426e945SPeter Xu /* 65984426e945SPeter Xu * Note: FAULT_FLAG_ALLOW_RETRY and 65994426e945SPeter Xu * FAULT_FLAG_TRIED can co-exist 66004426e945SPeter Xu */ 660187ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_TRIED; 660287ffc118SAndrea Arcangeli } 660387ffc118SAndrea Arcangeli ret = hugetlb_fault(mm, vma, vaddr, fault_flags); 660487ffc118SAndrea Arcangeli if (ret & VM_FAULT_ERROR) { 66052be7cfedSDaniel Jordan err = vm_fault_to_errno(ret, flags); 66061c59827dSHugh Dickins remainder = 0; 66071c59827dSHugh Dickins break; 66081c59827dSHugh Dickins } 660987ffc118SAndrea Arcangeli if (ret & VM_FAULT_RETRY) { 66104f6da934SPeter Xu if (locked && 66111ac25013SAndrea Arcangeli !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) 66124f6da934SPeter Xu *locked = 0; 661387ffc118SAndrea Arcangeli *nr_pages = 0; 661487ffc118SAndrea Arcangeli /* 661587ffc118SAndrea Arcangeli * VM_FAULT_RETRY must not return an 661687ffc118SAndrea Arcangeli * error, it will return zero 661787ffc118SAndrea Arcangeli * instead. 661887ffc118SAndrea Arcangeli * 661987ffc118SAndrea Arcangeli * No need to update "position" as the 662087ffc118SAndrea Arcangeli * caller will not check it after 662187ffc118SAndrea Arcangeli * *nr_pages is set to 0. 662287ffc118SAndrea Arcangeli */ 662387ffc118SAndrea Arcangeli return i; 662487ffc118SAndrea Arcangeli } 662587ffc118SAndrea Arcangeli continue; 662687ffc118SAndrea Arcangeli } 662763551ae0SDavid Gibson 6628a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 66297f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 66308fde12caSLinus Torvalds 6631b6a2619cSDavid Hildenbrand VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) && 6632b6a2619cSDavid Hildenbrand !PageAnonExclusive(page), page); 6633b6a2619cSDavid Hildenbrand 66348fde12caSLinus Torvalds /* 6635acbfb087SZhigang Lu * If subpage information not requested, update counters 6636acbfb087SZhigang Lu * and skip the same_page loop below. 6637acbfb087SZhigang Lu */ 6638b2cac248SLorenzo Stoakes if (!pages && !pfn_offset && 6639acbfb087SZhigang Lu (vaddr + huge_page_size(h) < vma->vm_end) && 6640acbfb087SZhigang Lu (remainder >= pages_per_huge_page(h))) { 6641acbfb087SZhigang Lu vaddr += huge_page_size(h); 6642acbfb087SZhigang Lu remainder -= pages_per_huge_page(h); 6643acbfb087SZhigang Lu i += pages_per_huge_page(h); 6644acbfb087SZhigang Lu spin_unlock(ptl); 6645eefc7fa5SPeter Xu hugetlb_vma_unlock_read(vma); 6646acbfb087SZhigang Lu continue; 6647acbfb087SZhigang Lu } 6648acbfb087SZhigang Lu 6649d08af0a5SJoao Martins /* vaddr may not be aligned to PAGE_SIZE */ 6650d08af0a5SJoao Martins refs = min3(pages_per_huge_page(h) - pfn_offset, remainder, 6651d08af0a5SJoao Martins (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT); 66520fa5bc40SJoao Martins 6653b2cac248SLorenzo Stoakes if (pages) 6654b2cac248SLorenzo Stoakes record_subpages(nth_page(page, pfn_offset), 665582e5d378SJoao Martins vma, refs, 6656b2cac248SLorenzo Stoakes likely(pages) ? pages + i : NULL); 665763551ae0SDavid Gibson 665882e5d378SJoao Martins if (pages) { 66590fa5bc40SJoao Martins /* 6660822951d8SMatthew Wilcox (Oracle) * try_grab_folio() should always succeed here, 66610fa5bc40SJoao Martins * because: a) we hold the ptl lock, and b) we've just 66620fa5bc40SJoao Martins * checked that the huge page is present in the page 66630fa5bc40SJoao Martins * tables. If the huge page is present, then the tail 66640fa5bc40SJoao Martins * pages must also be present. The ptl prevents the 66650fa5bc40SJoao Martins * head page and tail pages from being rearranged in 66664003f107SLogan Gunthorpe * any way. As this is hugetlb, the pages will never 66674003f107SLogan Gunthorpe * be p2pdma or not longterm pinable. So this page 66684003f107SLogan Gunthorpe * must be available at this point, unless the page 66694003f107SLogan Gunthorpe * refcount overflowed: 66700fa5bc40SJoao Martins */ 6671822951d8SMatthew Wilcox (Oracle) if (WARN_ON_ONCE(!try_grab_folio(pages[i], refs, 66720fa5bc40SJoao Martins flags))) { 66730fa5bc40SJoao Martins spin_unlock(ptl); 6674eefc7fa5SPeter Xu hugetlb_vma_unlock_read(vma); 66750fa5bc40SJoao Martins remainder = 0; 66760fa5bc40SJoao Martins err = -ENOMEM; 66770fa5bc40SJoao Martins break; 66780fa5bc40SJoao Martins } 6679d5d4b0aaSChen, Kenneth W } 668082e5d378SJoao Martins 668182e5d378SJoao Martins vaddr += (refs << PAGE_SHIFT); 668282e5d378SJoao Martins remainder -= refs; 668382e5d378SJoao Martins i += refs; 668482e5d378SJoao Martins 6685cb900f41SKirill A. Shutemov spin_unlock(ptl); 6686eefc7fa5SPeter Xu hugetlb_vma_unlock_read(vma); 668763551ae0SDavid Gibson } 668828a35716SMichel Lespinasse *nr_pages = remainder; 668987ffc118SAndrea Arcangeli /* 669087ffc118SAndrea Arcangeli * setting position is actually required only if remainder is 669187ffc118SAndrea Arcangeli * not zero but it's faster not to add a "if (remainder)" 669287ffc118SAndrea Arcangeli * branch. 669387ffc118SAndrea Arcangeli */ 669463551ae0SDavid Gibson *position = vaddr; 669563551ae0SDavid Gibson 66962be7cfedSDaniel Jordan return i ? i : err; 669763551ae0SDavid Gibson } 66988f860591SZhang, Yanmin 6699a79390f5SPeter Xu long hugetlb_change_protection(struct vm_area_struct *vma, 67005a90d5a1SPeter Xu unsigned long address, unsigned long end, 67015a90d5a1SPeter Xu pgprot_t newprot, unsigned long cp_flags) 67028f860591SZhang, Yanmin { 67038f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 67048f860591SZhang, Yanmin unsigned long start = address; 67058f860591SZhang, Yanmin pte_t *ptep; 67068f860591SZhang, Yanmin pte_t pte; 6707a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 6708a79390f5SPeter Xu long pages = 0, psize = huge_page_size(h); 6709dff11abeSMike Kravetz bool shared_pmd = false; 6710ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 6711e95a9851SMike Kravetz unsigned long last_addr_mask; 67125a90d5a1SPeter Xu bool uffd_wp = cp_flags & MM_CP_UFFD_WP; 67135a90d5a1SPeter Xu bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE; 6714dff11abeSMike Kravetz 6715dff11abeSMike Kravetz /* 6716dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 6717ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 6718dff11abeSMike Kravetz * range if PMD sharing is possible. 6719dff11abeSMike Kravetz */ 67207269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 67217d4a8be0SAlistair Popple 0, mm, start, end); 6722ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 67238f860591SZhang, Yanmin 67248f860591SZhang, Yanmin BUG_ON(address >= end); 6725ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 67268f860591SZhang, Yanmin 6727ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 672840549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 672983cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 673040549ba8SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 673160dfaad6SPeter Xu for (; address < end; address += psize) { 6732cb900f41SKirill A. Shutemov spinlock_t *ptl; 67339c67a207SPeter Xu ptep = hugetlb_walk(vma, address, psize); 6734e95a9851SMike Kravetz if (!ptep) { 6735fed15f13SPeter Xu if (!uffd_wp) { 6736e95a9851SMike Kravetz address |= last_addr_mask; 67378f860591SZhang, Yanmin continue; 6738e95a9851SMike Kravetz } 6739fed15f13SPeter Xu /* 6740fed15f13SPeter Xu * Userfaultfd wr-protect requires pgtable 6741fed15f13SPeter Xu * pre-allocations to install pte markers. 6742fed15f13SPeter Xu */ 6743fed15f13SPeter Xu ptep = huge_pte_alloc(mm, vma, address, psize); 6744d1751118SPeter Xu if (!ptep) { 6745d1751118SPeter Xu pages = -ENOMEM; 6746fed15f13SPeter Xu break; 6747fed15f13SPeter Xu } 6748d1751118SPeter Xu } 6749cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 67504ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, address, ptep)) { 675160dfaad6SPeter Xu /* 675260dfaad6SPeter Xu * When uffd-wp is enabled on the vma, unshare 675360dfaad6SPeter Xu * shouldn't happen at all. Warn about it if it 675460dfaad6SPeter Xu * happened due to some reason. 675560dfaad6SPeter Xu */ 675660dfaad6SPeter Xu WARN_ON_ONCE(uffd_wp || uffd_wp_resolve); 67577da4d641SPeter Zijlstra pages++; 6758cb900f41SKirill A. Shutemov spin_unlock(ptl); 6759dff11abeSMike Kravetz shared_pmd = true; 67604ddb4d91SMike Kravetz address |= last_addr_mask; 676139dde65cSChen, Kenneth W continue; 67627da4d641SPeter Zijlstra } 6763a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 6764a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 67650e678153SDavid Hildenbrand /* Nothing to do. */ 67660e678153SDavid Hildenbrand } else if (unlikely(is_hugetlb_entry_migration(pte))) { 6767a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 67686c287605SDavid Hildenbrand struct page *page = pfn_swap_entry_to_page(entry); 676944f86392SDavid Hildenbrand pte_t newpte = pte; 6770a8bda28dSNaoya Horiguchi 677144f86392SDavid Hildenbrand if (is_writable_migration_entry(entry)) { 67726c287605SDavid Hildenbrand if (PageAnon(page)) 67736c287605SDavid Hildenbrand entry = make_readable_exclusive_migration_entry( 67746c287605SDavid Hildenbrand swp_offset(entry)); 67756c287605SDavid Hildenbrand else 67764dd845b5SAlistair Popple entry = make_readable_migration_entry( 67774dd845b5SAlistair Popple swp_offset(entry)); 6778a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 677944f86392SDavid Hildenbrand pages++; 678044f86392SDavid Hildenbrand } 678144f86392SDavid Hildenbrand 67825a90d5a1SPeter Xu if (uffd_wp) 67835a90d5a1SPeter Xu newpte = pte_swp_mkuffd_wp(newpte); 67845a90d5a1SPeter Xu else if (uffd_wp_resolve) 67855a90d5a1SPeter Xu newpte = pte_swp_clear_uffd_wp(newpte); 678644f86392SDavid Hildenbrand if (!pte_same(pte, newpte)) 678718f39629SQi Zheng set_huge_pte_at(mm, address, ptep, newpte); 67880e678153SDavid Hildenbrand } else if (unlikely(is_pte_marker(pte))) { 67890e678153SDavid Hildenbrand /* No other markers apply for now. */ 67900e678153SDavid Hildenbrand WARN_ON_ONCE(!pte_marker_uffd_wp(pte)); 679160dfaad6SPeter Xu if (uffd_wp_resolve) 67920e678153SDavid Hildenbrand /* Safe to modify directly (non-present->none). */ 679360dfaad6SPeter Xu huge_pte_clear(mm, address, ptep, psize); 67940e678153SDavid Hildenbrand } else if (!huge_pte_none(pte)) { 6795023bdd00SAneesh Kumar K.V pte_t old_pte; 679679c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 6797023bdd00SAneesh Kumar K.V 6798023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 679916785bd7SAnshuman Khandual pte = huge_pte_modify(old_pte, newprot); 680079c1c594SChristophe Leroy pte = arch_make_huge_pte(pte, shift, vma->vm_flags); 68015a90d5a1SPeter Xu if (uffd_wp) 6802f1eb1bacSPeter Xu pte = huge_pte_mkuffd_wp(pte); 68035a90d5a1SPeter Xu else if (uffd_wp_resolve) 68045a90d5a1SPeter Xu pte = huge_pte_clear_uffd_wp(pte); 6805023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 68067da4d641SPeter Zijlstra pages++; 680760dfaad6SPeter Xu } else { 680860dfaad6SPeter Xu /* None pte */ 680960dfaad6SPeter Xu if (unlikely(uffd_wp)) 681060dfaad6SPeter Xu /* Safe to modify directly (none->non-present). */ 681160dfaad6SPeter Xu set_huge_pte_at(mm, address, ptep, 681260dfaad6SPeter Xu make_pte_marker(PTE_MARKER_UFFD_WP)); 68138f860591SZhang, Yanmin } 6814cb900f41SKirill A. Shutemov spin_unlock(ptl); 68158f860591SZhang, Yanmin } 6816d833352aSMel Gorman /* 6817c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 6818d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 6819c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 6820dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 6821dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 6822d833352aSMel Gorman */ 6823dff11abeSMike Kravetz if (shared_pmd) 6824ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 6825dff11abeSMike Kravetz else 68265491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 68270f10851eSJérôme Glisse /* 68280f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 68290f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 68300f10851eSJérôme Glisse * 6831ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 68320f10851eSJérôme Glisse */ 683383cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 683440549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 6835ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 68367da4d641SPeter Zijlstra 6837d1751118SPeter Xu return pages > 0 ? (pages << h->order) : pages; 68388f860591SZhang, Yanmin } 68398f860591SZhang, Yanmin 684033b8f84aSMike Kravetz /* Return true if reservation was successful, false otherwise. */ 684133b8f84aSMike Kravetz bool hugetlb_reserve_pages(struct inode *inode, 6842a1e78772SMel Gorman long from, long to, 68435a6fe125SMel Gorman struct vm_area_struct *vma, 6844ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 6845e4e574b7SAdam Litke { 6846c5094ec7SMike Kravetz long chg = -1, add = -1; 6847a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 684890481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 68499119a41eSJoonsoo Kim struct resv_map *resv_map; 6850075a61d0SMina Almasry struct hugetlb_cgroup *h_cg = NULL; 68510db9d74eSMina Almasry long gbl_reserve, regions_needed = 0; 6852e4e574b7SAdam Litke 685363489f8eSMike Kravetz /* This should never happen */ 685463489f8eSMike Kravetz if (from > to) { 685563489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 685633b8f84aSMike Kravetz return false; 685763489f8eSMike Kravetz } 685863489f8eSMike Kravetz 6859a1e78772SMel Gorman /* 6860e700898fSMike Kravetz * vma specific semaphore used for pmd sharing and fault/truncation 6861e700898fSMike Kravetz * synchronization 68628d9bfb26SMike Kravetz */ 68638d9bfb26SMike Kravetz hugetlb_vma_lock_alloc(vma); 68648d9bfb26SMike Kravetz 68658d9bfb26SMike Kravetz /* 686617c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 686717c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 686890481622SDavid Gibson * without using reserves 686917c9d12eSMel Gorman */ 6870ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 687133b8f84aSMike Kravetz return true; 687217c9d12eSMel Gorman 687317c9d12eSMel Gorman /* 6874a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 6875a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 6876a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 6877a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 6878a1e78772SMel Gorman */ 68799119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 6880f27a5136SMike Kravetz /* 6881f27a5136SMike Kravetz * resv_map can not be NULL as hugetlb_reserve_pages is only 6882f27a5136SMike Kravetz * called for inodes for which resv_maps were created (see 6883f27a5136SMike Kravetz * hugetlbfs_get_inode). 6884f27a5136SMike Kravetz */ 68854e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 68869119a41eSJoonsoo Kim 68870db9d74eSMina Almasry chg = region_chg(resv_map, from, to, ®ions_needed); 68889119a41eSJoonsoo Kim } else { 6889e9fe92aeSMina Almasry /* Private mapping. */ 68909119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 68915a6fe125SMel Gorman if (!resv_map) 68928d9bfb26SMike Kravetz goto out_err; 68935a6fe125SMel Gorman 689417c9d12eSMel Gorman chg = to - from; 689517c9d12eSMel Gorman 68965a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 68975a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 68985a6fe125SMel Gorman } 68995a6fe125SMel Gorman 690033b8f84aSMike Kravetz if (chg < 0) 6901c50ac050SDave Hansen goto out_err; 690217c9d12eSMel Gorman 690333b8f84aSMike Kravetz if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), 690433b8f84aSMike Kravetz chg * pages_per_huge_page(h), &h_cg) < 0) 6905075a61d0SMina Almasry goto out_err; 6906075a61d0SMina Almasry 6907075a61d0SMina Almasry if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { 6908075a61d0SMina Almasry /* For private mappings, the hugetlb_cgroup uncharge info hangs 6909075a61d0SMina Almasry * of the resv_map. 6910075a61d0SMina Almasry */ 6911075a61d0SMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); 6912075a61d0SMina Almasry } 6913075a61d0SMina Almasry 69141c5ecae3SMike Kravetz /* 69151c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 69161c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 69171c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 69181c5ecae3SMike Kravetz */ 69191c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 692033b8f84aSMike Kravetz if (gbl_reserve < 0) 6921075a61d0SMina Almasry goto out_uncharge_cgroup; 692217c9d12eSMel Gorman 692317c9d12eSMel Gorman /* 692417c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 692590481622SDavid Gibson * Hand the pages back to the subpool if there are not 692617c9d12eSMel Gorman */ 692733b8f84aSMike Kravetz if (hugetlb_acct_memory(h, gbl_reserve) < 0) 6928075a61d0SMina Almasry goto out_put_pages; 692917c9d12eSMel Gorman 693017c9d12eSMel Gorman /* 693117c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 693217c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 693317c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 693417c9d12eSMel Gorman * the reservation was and the page cache tells how much of 693517c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 693617c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 693717c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 693817c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 693917c9d12eSMel Gorman * else has to be done for private mappings here 694017c9d12eSMel Gorman */ 694133039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 6942075a61d0SMina Almasry add = region_add(resv_map, from, to, regions_needed, h, h_cg); 694333039678SMike Kravetz 69440db9d74eSMina Almasry if (unlikely(add < 0)) { 69450db9d74eSMina Almasry hugetlb_acct_memory(h, -gbl_reserve); 6946075a61d0SMina Almasry goto out_put_pages; 69470db9d74eSMina Almasry } else if (unlikely(chg > add)) { 694833039678SMike Kravetz /* 694933039678SMike Kravetz * pages in this range were added to the reserve 695033039678SMike Kravetz * map between region_chg and region_add. This 6951d0ce0e47SSidhartha Kumar * indicates a race with alloc_hugetlb_folio. Adjust 695233039678SMike Kravetz * the subpool and reserve counts modified above 695333039678SMike Kravetz * based on the difference. 695433039678SMike Kravetz */ 695533039678SMike Kravetz long rsv_adjust; 695633039678SMike Kravetz 6957d85aecf2SMiaohe Lin /* 6958d85aecf2SMiaohe Lin * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the 6959d85aecf2SMiaohe Lin * reference to h_cg->css. See comment below for detail. 6960d85aecf2SMiaohe Lin */ 6961075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd( 6962075a61d0SMina Almasry hstate_index(h), 6963075a61d0SMina Almasry (chg - add) * pages_per_huge_page(h), h_cg); 6964075a61d0SMina Almasry 696533039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 696633039678SMike Kravetz chg - add); 696733039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 6968d85aecf2SMiaohe Lin } else if (h_cg) { 6969d85aecf2SMiaohe Lin /* 6970d85aecf2SMiaohe Lin * The file_regions will hold their own reference to 6971d85aecf2SMiaohe Lin * h_cg->css. So we should release the reference held 6972d85aecf2SMiaohe Lin * via hugetlb_cgroup_charge_cgroup_rsvd() when we are 6973d85aecf2SMiaohe Lin * done. 6974d85aecf2SMiaohe Lin */ 6975d85aecf2SMiaohe Lin hugetlb_cgroup_put_rsvd_cgroup(h_cg); 697633039678SMike Kravetz } 697733039678SMike Kravetz } 697833b8f84aSMike Kravetz return true; 697933b8f84aSMike Kravetz 6980075a61d0SMina Almasry out_put_pages: 6981075a61d0SMina Almasry /* put back original number of pages, chg */ 6982075a61d0SMina Almasry (void)hugepage_subpool_put_pages(spool, chg); 6983075a61d0SMina Almasry out_uncharge_cgroup: 6984075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), 6985075a61d0SMina Almasry chg * pages_per_huge_page(h), h_cg); 6986c50ac050SDave Hansen out_err: 69878d9bfb26SMike Kravetz hugetlb_vma_lock_free(vma); 69885e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 69890db9d74eSMina Almasry /* Only call region_abort if the region_chg succeeded but the 69900db9d74eSMina Almasry * region_add failed or didn't run. 69910db9d74eSMina Almasry */ 69920db9d74eSMina Almasry if (chg >= 0 && add < 0) 69930db9d74eSMina Almasry region_abort(resv_map, from, to, regions_needed); 6994f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 6995f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 699633b8f84aSMike Kravetz return false; 6997a43a8c39SChen, Kenneth W } 6998a43a8c39SChen, Kenneth W 6999b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 7000b5cec28dSMike Kravetz long freed) 7001a43a8c39SChen, Kenneth W { 7002a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 70034e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 70049119a41eSJoonsoo Kim long chg = 0; 700590481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 70061c5ecae3SMike Kravetz long gbl_reserve; 700745c682a6SKen Chen 7008f27a5136SMike Kravetz /* 7009f27a5136SMike Kravetz * Since this routine can be called in the evict inode path for all 7010f27a5136SMike Kravetz * hugetlbfs inodes, resv_map could be NULL. 7011f27a5136SMike Kravetz */ 7012b5cec28dSMike Kravetz if (resv_map) { 7013b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 7014b5cec28dSMike Kravetz /* 7015b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 7016b5cec28dSMike Kravetz * must be split and another region descriptor can not be 7017b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 7018b5cec28dSMike Kravetz */ 7019b5cec28dSMike Kravetz if (chg < 0) 7020b5cec28dSMike Kravetz return chg; 7021b5cec28dSMike Kravetz } 7022b5cec28dSMike Kravetz 702345c682a6SKen Chen spin_lock(&inode->i_lock); 7024e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 702545c682a6SKen Chen spin_unlock(&inode->i_lock); 702645c682a6SKen Chen 70271c5ecae3SMike Kravetz /* 70281c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 70291c5ecae3SMike Kravetz * reservations to be released may be adjusted. 7030dddf31a4SMiaohe Lin * 7031dddf31a4SMiaohe Lin * Note that !resv_map implies freed == 0. So (chg - freed) 7032dddf31a4SMiaohe Lin * won't go negative. 70331c5ecae3SMike Kravetz */ 70341c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 70351c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 7036b5cec28dSMike Kravetz 7037b5cec28dSMike Kravetz return 0; 7038a43a8c39SChen, Kenneth W } 703993f70f90SNaoya Horiguchi 70403212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 70413212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 70423212b535SSteve Capper struct vm_area_struct *vma, 70433212b535SSteve Capper unsigned long addr, pgoff_t idx) 70443212b535SSteve Capper { 70453212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 70463212b535SSteve Capper svma->vm_start; 70473212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 70483212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 70493212b535SSteve Capper 70503212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 7051e430a95aSSuren Baghdasaryan unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED_MASK; 7052e430a95aSSuren Baghdasaryan unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED_MASK; 70533212b535SSteve Capper 70543212b535SSteve Capper /* 70553212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 70563212b535SSteve Capper * page table page. 7057131a79b4SMike Kravetz * 7058131a79b4SMike Kravetz * Also, vma_lock (vm_private_data) is required for sharing. 70593212b535SSteve Capper */ 70603212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 70613212b535SSteve Capper vm_flags != svm_flags || 7062131a79b4SMike Kravetz !range_in_vma(svma, sbase, s_end) || 7063131a79b4SMike Kravetz !svma->vm_private_data) 70643212b535SSteve Capper return 0; 70653212b535SSteve Capper 70663212b535SSteve Capper return saddr; 70673212b535SSteve Capper } 70683212b535SSteve Capper 7069bbff39ccSMike Kravetz bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 70703212b535SSteve Capper { 7071bbff39ccSMike Kravetz unsigned long start = addr & PUD_MASK; 7072bbff39ccSMike Kravetz unsigned long end = start + PUD_SIZE; 7073bbff39ccSMike Kravetz 70748d9bfb26SMike Kravetz #ifdef CONFIG_USERFAULTFD 70758d9bfb26SMike Kravetz if (uffd_disable_huge_pmd_share(vma)) 70768d9bfb26SMike Kravetz return false; 70778d9bfb26SMike Kravetz #endif 70783212b535SSteve Capper /* 70793212b535SSteve Capper * check on proper vm_flags and page table alignment 70803212b535SSteve Capper */ 70818d9bfb26SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 708231aafb45SNicholas Krause return false; 7083bbff39ccSMike Kravetz if (!vma->vm_private_data) /* vma lock required for sharing */ 70848d9bfb26SMike Kravetz return false; 70858d9bfb26SMike Kravetz if (!range_in_vma(vma, start, end)) 70868d9bfb26SMike Kravetz return false; 70878d9bfb26SMike Kravetz return true; 70888d9bfb26SMike Kravetz } 70898d9bfb26SMike Kravetz 70903212b535SSteve Capper /* 7091017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 7092017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 7093017b1660SMike Kravetz * shared pmd mappings. 7094017b1660SMike Kravetz */ 7095017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 7096017b1660SMike Kravetz unsigned long *start, unsigned long *end) 7097017b1660SMike Kravetz { 7098a1ba9da8SLi Xinhai unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), 7099a1ba9da8SLi Xinhai v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 7100017b1660SMike Kravetz 7101a1ba9da8SLi Xinhai /* 7102f0953a1bSIngo Molnar * vma needs to span at least one aligned PUD size, and the range 7103f0953a1bSIngo Molnar * must be at least partially within in. 7104a1ba9da8SLi Xinhai */ 7105a1ba9da8SLi Xinhai if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || 7106a1ba9da8SLi Xinhai (*end <= v_start) || (*start >= v_end)) 7107017b1660SMike Kravetz return; 7108017b1660SMike Kravetz 710975802ca6SPeter Xu /* Extend the range to be PUD aligned for a worst case scenario */ 7110a1ba9da8SLi Xinhai if (*start > v_start) 7111a1ba9da8SLi Xinhai *start = ALIGN_DOWN(*start, PUD_SIZE); 7112017b1660SMike Kravetz 7113a1ba9da8SLi Xinhai if (*end < v_end) 7114a1ba9da8SLi Xinhai *end = ALIGN(*end, PUD_SIZE); 7115017b1660SMike Kravetz } 7116017b1660SMike Kravetz 7117017b1660SMike Kravetz /* 71183212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 71193212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 71203212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 71213a47c54fSMike Kravetz * code much cleaner. pmd allocation is essential for the shared case because 71223a47c54fSMike Kravetz * pud has to be populated inside the same i_mmap_rwsem section - otherwise 71233a47c54fSMike Kravetz * racing tasks could either miss the sharing (see huge_pte_offset) or select a 71243a47c54fSMike Kravetz * bad pmd for sharing. 71253212b535SSteve Capper */ 7126aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 7127aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 71283212b535SSteve Capper { 71293212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 71303212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 71313212b535SSteve Capper vma->vm_pgoff; 71323212b535SSteve Capper struct vm_area_struct *svma; 71333212b535SSteve Capper unsigned long saddr; 71343212b535SSteve Capper pte_t *spte = NULL; 71353212b535SSteve Capper pte_t *pte; 7136cb900f41SKirill A. Shutemov spinlock_t *ptl; 71373212b535SSteve Capper 71383a47c54fSMike Kravetz i_mmap_lock_read(mapping); 71393212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 71403212b535SSteve Capper if (svma == vma) 71413212b535SSteve Capper continue; 71423212b535SSteve Capper 71433212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 71443212b535SSteve Capper if (saddr) { 71459c67a207SPeter Xu spte = hugetlb_walk(svma, saddr, 71467868a208SPunit Agrawal vma_mmu_pagesize(svma)); 71473212b535SSteve Capper if (spte) { 71483212b535SSteve Capper get_page(virt_to_page(spte)); 71493212b535SSteve Capper break; 71503212b535SSteve Capper } 71513212b535SSteve Capper } 71523212b535SSteve Capper } 71533212b535SSteve Capper 71543212b535SSteve Capper if (!spte) 71553212b535SSteve Capper goto out; 71563212b535SSteve Capper 71578bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 7158dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 71593212b535SSteve Capper pud_populate(mm, pud, 71603212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 7161c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 7162dc6c9a35SKirill A. Shutemov } else { 71633212b535SSteve Capper put_page(virt_to_page(spte)); 7164dc6c9a35SKirill A. Shutemov } 7165cb900f41SKirill A. Shutemov spin_unlock(ptl); 71663212b535SSteve Capper out: 71673212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 71683a47c54fSMike Kravetz i_mmap_unlock_read(mapping); 71693212b535SSteve Capper return pte; 71703212b535SSteve Capper } 71713212b535SSteve Capper 71723212b535SSteve Capper /* 71733212b535SSteve Capper * unmap huge page backed by shared pte. 71743212b535SSteve Capper * 71753212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 71763212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 71773212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 71783212b535SSteve Capper * 71793a47c54fSMike Kravetz * Called with page table lock held. 71803212b535SSteve Capper * 71813212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 71823212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 71833212b535SSteve Capper */ 718434ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 71854ddb4d91SMike Kravetz unsigned long addr, pte_t *ptep) 71863212b535SSteve Capper { 71874ddb4d91SMike Kravetz pgd_t *pgd = pgd_offset(mm, addr); 71884ddb4d91SMike Kravetz p4d_t *p4d = p4d_offset(pgd, addr); 71894ddb4d91SMike Kravetz pud_t *pud = pud_offset(p4d, addr); 71903212b535SSteve Capper 719134ae204fSMike Kravetz i_mmap_assert_write_locked(vma->vm_file->f_mapping); 719240549ba8SMike Kravetz hugetlb_vma_assert_locked(vma); 71933212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 71943212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 71953212b535SSteve Capper return 0; 71963212b535SSteve Capper 71973212b535SSteve Capper pud_clear(pud); 71983212b535SSteve Capper put_page(virt_to_page(ptep)); 7199dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 72003212b535SSteve Capper return 1; 72013212b535SSteve Capper } 7202c1991e07SPeter Xu 72039e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 72048d9bfb26SMike Kravetz 7205aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 7206aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 72079e5fc74cSSteve Capper { 72089e5fc74cSSteve Capper return NULL; 72099e5fc74cSSteve Capper } 7210e81f2d22SZhang Zhen 721134ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 72124ddb4d91SMike Kravetz unsigned long addr, pte_t *ptep) 7213e81f2d22SZhang Zhen { 7214e81f2d22SZhang Zhen return 0; 7215e81f2d22SZhang Zhen } 7216017b1660SMike Kravetz 7217017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 7218017b1660SMike Kravetz unsigned long *start, unsigned long *end) 7219017b1660SMike Kravetz { 7220017b1660SMike Kravetz } 7221c1991e07SPeter Xu 7222c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 7223c1991e07SPeter Xu { 7224c1991e07SPeter Xu return false; 7225c1991e07SPeter Xu } 72263212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 72273212b535SSteve Capper 72289e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 7229aec44e0fSPeter Xu pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, 72309e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 72319e5fc74cSSteve Capper { 72329e5fc74cSSteve Capper pgd_t *pgd; 7233c2febafcSKirill A. Shutemov p4d_t *p4d; 72349e5fc74cSSteve Capper pud_t *pud; 72359e5fc74cSSteve Capper pte_t *pte = NULL; 72369e5fc74cSSteve Capper 72379e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 7238f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 7239f4f0a3d8SKirill A. Shutemov if (!p4d) 7240f4f0a3d8SKirill A. Shutemov return NULL; 7241c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 72429e5fc74cSSteve Capper if (pud) { 72439e5fc74cSSteve Capper if (sz == PUD_SIZE) { 72449e5fc74cSSteve Capper pte = (pte_t *)pud; 72459e5fc74cSSteve Capper } else { 72469e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 7247c1991e07SPeter Xu if (want_pmd_share(vma, addr) && pud_none(*pud)) 7248aec44e0fSPeter Xu pte = huge_pmd_share(mm, vma, addr, pud); 72499e5fc74cSSteve Capper else 72509e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 72519e5fc74cSSteve Capper } 72529e5fc74cSSteve Capper } 72534e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 72549e5fc74cSSteve Capper 72559e5fc74cSSteve Capper return pte; 72569e5fc74cSSteve Capper } 72579e5fc74cSSteve Capper 72589b19df29SPunit Agrawal /* 72599b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 72609b19df29SPunit Agrawal * entry at address @addr 72619b19df29SPunit Agrawal * 72628ac0b81aSLi Xinhai * Return: Pointer to page table entry (PUD or PMD) for 72638ac0b81aSLi Xinhai * address @addr, or NULL if a !p*d_present() entry is encountered and the 72649b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 72659b19df29SPunit Agrawal * table. 72669b19df29SPunit Agrawal */ 72677868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 72687868a208SPunit Agrawal unsigned long addr, unsigned long sz) 72699e5fc74cSSteve Capper { 72709e5fc74cSSteve Capper pgd_t *pgd; 7271c2febafcSKirill A. Shutemov p4d_t *p4d; 72728ac0b81aSLi Xinhai pud_t *pud; 72738ac0b81aSLi Xinhai pmd_t *pmd; 72749e5fc74cSSteve Capper 72759e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 7276c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 7277c2febafcSKirill A. Shutemov return NULL; 7278c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 7279c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 7280c2febafcSKirill A. Shutemov return NULL; 72819b19df29SPunit Agrawal 7282c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 72838ac0b81aSLi Xinhai if (sz == PUD_SIZE) 72848ac0b81aSLi Xinhai /* must be pud huge, non-present or none */ 72859e5fc74cSSteve Capper return (pte_t *)pud; 72868ac0b81aSLi Xinhai if (!pud_present(*pud)) 72878ac0b81aSLi Xinhai return NULL; 72888ac0b81aSLi Xinhai /* must have a valid entry and size to go further */ 72899b19df29SPunit Agrawal 72909e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 72918ac0b81aSLi Xinhai /* must be pmd huge, non-present or none */ 72929e5fc74cSSteve Capper return (pte_t *)pmd; 72939e5fc74cSSteve Capper } 72949e5fc74cSSteve Capper 7295e95a9851SMike Kravetz /* 7296e95a9851SMike Kravetz * Return a mask that can be used to update an address to the last huge 7297e95a9851SMike Kravetz * page in a page table page mapping size. Used to skip non-present 7298e95a9851SMike Kravetz * page table entries when linearly scanning address ranges. Architectures 7299e95a9851SMike Kravetz * with unique huge page to page table relationships can define their own 7300e95a9851SMike Kravetz * version of this routine. 7301e95a9851SMike Kravetz */ 7302e95a9851SMike Kravetz unsigned long hugetlb_mask_last_page(struct hstate *h) 7303e95a9851SMike Kravetz { 7304e95a9851SMike Kravetz unsigned long hp_size = huge_page_size(h); 7305e95a9851SMike Kravetz 7306e95a9851SMike Kravetz if (hp_size == PUD_SIZE) 7307e95a9851SMike Kravetz return P4D_SIZE - PUD_SIZE; 7308e95a9851SMike Kravetz else if (hp_size == PMD_SIZE) 7309e95a9851SMike Kravetz return PUD_SIZE - PMD_SIZE; 7310e95a9851SMike Kravetz else 7311e95a9851SMike Kravetz return 0UL; 7312e95a9851SMike Kravetz } 7313e95a9851SMike Kravetz 7314e95a9851SMike Kravetz #else 7315e95a9851SMike Kravetz 7316e95a9851SMike Kravetz /* See description above. Architectures can provide their own version. */ 7317e95a9851SMike Kravetz __weak unsigned long hugetlb_mask_last_page(struct hstate *h) 7318e95a9851SMike Kravetz { 73194ddb4d91SMike Kravetz #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 73204ddb4d91SMike Kravetz if (huge_page_size(h) == PMD_SIZE) 73214ddb4d91SMike Kravetz return PUD_SIZE - PMD_SIZE; 73224ddb4d91SMike Kravetz #endif 7323e95a9851SMike Kravetz return 0UL; 7324e95a9851SMike Kravetz } 7325e95a9851SMike Kravetz 732661f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 732761f77edaSNaoya Horiguchi 732861f77edaSNaoya Horiguchi /* 732961f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 733061f77edaSNaoya Horiguchi * behavior. 733161f77edaSNaoya Horiguchi */ 73329747b9e9SBaolin Wang bool isolate_hugetlb(struct folio *folio, struct list_head *list) 733331caf665SNaoya Horiguchi { 73349747b9e9SBaolin Wang bool ret = true; 7335bcc54222SNaoya Horiguchi 7336db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 73376aa3a920SSidhartha Kumar if (!folio_test_hugetlb(folio) || 73386aa3a920SSidhartha Kumar !folio_test_hugetlb_migratable(folio) || 73396aa3a920SSidhartha Kumar !folio_try_get(folio)) { 73409747b9e9SBaolin Wang ret = false; 7341bcc54222SNaoya Horiguchi goto unlock; 7342bcc54222SNaoya Horiguchi } 73436aa3a920SSidhartha Kumar folio_clear_hugetlb_migratable(folio); 73446aa3a920SSidhartha Kumar list_move_tail(&folio->lru, list); 7345bcc54222SNaoya Horiguchi unlock: 7346db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7347bcc54222SNaoya Horiguchi return ret; 734831caf665SNaoya Horiguchi } 734931caf665SNaoya Horiguchi 735004bac040SSidhartha Kumar int get_hwpoison_hugetlb_folio(struct folio *folio, bool *hugetlb, bool unpoison) 735125182f05SNaoya Horiguchi { 735225182f05SNaoya Horiguchi int ret = 0; 735325182f05SNaoya Horiguchi 735425182f05SNaoya Horiguchi *hugetlb = false; 735525182f05SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 735604bac040SSidhartha Kumar if (folio_test_hugetlb(folio)) { 735725182f05SNaoya Horiguchi *hugetlb = true; 735804bac040SSidhartha Kumar if (folio_test_hugetlb_freed(folio)) 7359b283d983SNaoya Horiguchi ret = 0; 736004bac040SSidhartha Kumar else if (folio_test_hugetlb_migratable(folio) || unpoison) 736104bac040SSidhartha Kumar ret = folio_try_get(folio); 73620ed950d1SNaoya Horiguchi else 73630ed950d1SNaoya Horiguchi ret = -EBUSY; 736425182f05SNaoya Horiguchi } 736525182f05SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 736625182f05SNaoya Horiguchi return ret; 736725182f05SNaoya Horiguchi } 736825182f05SNaoya Horiguchi 7369e591ef7dSNaoya Horiguchi int get_huge_page_for_hwpoison(unsigned long pfn, int flags, 7370e591ef7dSNaoya Horiguchi bool *migratable_cleared) 7371405ce051SNaoya Horiguchi { 7372405ce051SNaoya Horiguchi int ret; 7373405ce051SNaoya Horiguchi 7374405ce051SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 7375e591ef7dSNaoya Horiguchi ret = __get_huge_page_for_hwpoison(pfn, flags, migratable_cleared); 7376405ce051SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 7377405ce051SNaoya Horiguchi return ret; 7378405ce051SNaoya Horiguchi } 7379405ce051SNaoya Horiguchi 7380ea8e72f4SSidhartha Kumar void folio_putback_active_hugetlb(struct folio *folio) 738131caf665SNaoya Horiguchi { 7382db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 7383ea8e72f4SSidhartha Kumar folio_set_hugetlb_migratable(folio); 7384ea8e72f4SSidhartha Kumar list_move_tail(&folio->lru, &(folio_hstate(folio))->hugepage_activelist); 7385db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7386ea8e72f4SSidhartha Kumar folio_put(folio); 738731caf665SNaoya Horiguchi } 7388ab5ac90aSMichal Hocko 7389345c62d1SSidhartha Kumar void move_hugetlb_state(struct folio *old_folio, struct folio *new_folio, int reason) 7390ab5ac90aSMichal Hocko { 7391345c62d1SSidhartha Kumar struct hstate *h = folio_hstate(old_folio); 7392ab5ac90aSMichal Hocko 7393345c62d1SSidhartha Kumar hugetlb_cgroup_migrate(old_folio, new_folio); 7394345c62d1SSidhartha Kumar set_page_owner_migrate_reason(&new_folio->page, reason); 7395ab5ac90aSMichal Hocko 7396ab5ac90aSMichal Hocko /* 7397345c62d1SSidhartha Kumar * transfer temporary state of the new hugetlb folio. This is 7398ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 7399ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 7400ab5ac90aSMichal Hocko * the temporary status. 7401ab5ac90aSMichal Hocko * 7402ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 7403ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 7404ab5ac90aSMichal Hocko * the per-node's. 7405ab5ac90aSMichal Hocko */ 7406345c62d1SSidhartha Kumar if (folio_test_hugetlb_temporary(new_folio)) { 7407345c62d1SSidhartha Kumar int old_nid = folio_nid(old_folio); 7408345c62d1SSidhartha Kumar int new_nid = folio_nid(new_folio); 7409ab5ac90aSMichal Hocko 7410345c62d1SSidhartha Kumar folio_set_hugetlb_temporary(old_folio); 7411345c62d1SSidhartha Kumar folio_clear_hugetlb_temporary(new_folio); 7412345c62d1SSidhartha Kumar 7413ab5ac90aSMichal Hocko 74145af1ab1dSMiaohe Lin /* 74155af1ab1dSMiaohe Lin * There is no need to transfer the per-node surplus state 74165af1ab1dSMiaohe Lin * when we do not cross the node. 74175af1ab1dSMiaohe Lin */ 74185af1ab1dSMiaohe Lin if (new_nid == old_nid) 74195af1ab1dSMiaohe Lin return; 7420db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 7421ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 7422ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 7423ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 7424ab5ac90aSMichal Hocko } 7425db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7426ab5ac90aSMichal Hocko } 7427ab5ac90aSMichal Hocko } 7428cf11e85fSRoman Gushchin 7429b30c14cdSJames Houghton static void hugetlb_unshare_pmds(struct vm_area_struct *vma, 7430b30c14cdSJames Houghton unsigned long start, 7431b30c14cdSJames Houghton unsigned long end) 74326dfeaff9SPeter Xu { 74336dfeaff9SPeter Xu struct hstate *h = hstate_vma(vma); 74346dfeaff9SPeter Xu unsigned long sz = huge_page_size(h); 74356dfeaff9SPeter Xu struct mm_struct *mm = vma->vm_mm; 74366dfeaff9SPeter Xu struct mmu_notifier_range range; 7437b30c14cdSJames Houghton unsigned long address; 74386dfeaff9SPeter Xu spinlock_t *ptl; 74396dfeaff9SPeter Xu pte_t *ptep; 74406dfeaff9SPeter Xu 74416dfeaff9SPeter Xu if (!(vma->vm_flags & VM_MAYSHARE)) 74426dfeaff9SPeter Xu return; 74436dfeaff9SPeter Xu 74446dfeaff9SPeter Xu if (start >= end) 74456dfeaff9SPeter Xu return; 74466dfeaff9SPeter Xu 74479c8bbfacSBaolin Wang flush_cache_range(vma, start, end); 74486dfeaff9SPeter Xu /* 74496dfeaff9SPeter Xu * No need to call adjust_range_if_pmd_sharing_possible(), because 74506dfeaff9SPeter Xu * we have already done the PUD_SIZE alignment. 74516dfeaff9SPeter Xu */ 74527d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, 74536dfeaff9SPeter Xu start, end); 74546dfeaff9SPeter Xu mmu_notifier_invalidate_range_start(&range); 745540549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 74566dfeaff9SPeter Xu i_mmap_lock_write(vma->vm_file->f_mapping); 74576dfeaff9SPeter Xu for (address = start; address < end; address += PUD_SIZE) { 74589c67a207SPeter Xu ptep = hugetlb_walk(vma, address, sz); 74596dfeaff9SPeter Xu if (!ptep) 74606dfeaff9SPeter Xu continue; 74616dfeaff9SPeter Xu ptl = huge_pte_lock(h, mm, ptep); 74624ddb4d91SMike Kravetz huge_pmd_unshare(mm, vma, address, ptep); 74636dfeaff9SPeter Xu spin_unlock(ptl); 74646dfeaff9SPeter Xu } 74656dfeaff9SPeter Xu flush_hugetlb_tlb_range(vma, start, end); 74666dfeaff9SPeter Xu i_mmap_unlock_write(vma->vm_file->f_mapping); 746740549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 74686dfeaff9SPeter Xu /* 74696dfeaff9SPeter Xu * No need to call mmu_notifier_invalidate_range(), see 7470ee65728eSMike Rapoport * Documentation/mm/mmu_notifier.rst. 74716dfeaff9SPeter Xu */ 74726dfeaff9SPeter Xu mmu_notifier_invalidate_range_end(&range); 74736dfeaff9SPeter Xu } 74746dfeaff9SPeter Xu 7475b30c14cdSJames Houghton /* 7476b30c14cdSJames Houghton * This function will unconditionally remove all the shared pmd pgtable entries 7477b30c14cdSJames Houghton * within the specific vma for a hugetlbfs memory range. 7478b30c14cdSJames Houghton */ 7479b30c14cdSJames Houghton void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) 7480b30c14cdSJames Houghton { 7481b30c14cdSJames Houghton hugetlb_unshare_pmds(vma, ALIGN(vma->vm_start, PUD_SIZE), 7482b30c14cdSJames Houghton ALIGN_DOWN(vma->vm_end, PUD_SIZE)); 7483b30c14cdSJames Houghton } 7484b30c14cdSJames Houghton 7485cf11e85fSRoman Gushchin #ifdef CONFIG_CMA 7486cf11e85fSRoman Gushchin static bool cma_reserve_called __initdata; 7487cf11e85fSRoman Gushchin 7488cf11e85fSRoman Gushchin static int __init cmdline_parse_hugetlb_cma(char *p) 7489cf11e85fSRoman Gushchin { 749038e719abSBaolin Wang int nid, count = 0; 749138e719abSBaolin Wang unsigned long tmp; 749238e719abSBaolin Wang char *s = p; 749338e719abSBaolin Wang 749438e719abSBaolin Wang while (*s) { 749538e719abSBaolin Wang if (sscanf(s, "%lu%n", &tmp, &count) != 1) 749638e719abSBaolin Wang break; 749738e719abSBaolin Wang 749838e719abSBaolin Wang if (s[count] == ':') { 7499f9317f77SMike Kravetz if (tmp >= MAX_NUMNODES) 750038e719abSBaolin Wang break; 7501f9317f77SMike Kravetz nid = array_index_nospec(tmp, MAX_NUMNODES); 750238e719abSBaolin Wang 750338e719abSBaolin Wang s += count + 1; 750438e719abSBaolin Wang tmp = memparse(s, &s); 750538e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = tmp; 750638e719abSBaolin Wang hugetlb_cma_size += tmp; 750738e719abSBaolin Wang 750838e719abSBaolin Wang /* 750938e719abSBaolin Wang * Skip the separator if have one, otherwise 751038e719abSBaolin Wang * break the parsing. 751138e719abSBaolin Wang */ 751238e719abSBaolin Wang if (*s == ',') 751338e719abSBaolin Wang s++; 751438e719abSBaolin Wang else 751538e719abSBaolin Wang break; 751638e719abSBaolin Wang } else { 7517cf11e85fSRoman Gushchin hugetlb_cma_size = memparse(p, &p); 751838e719abSBaolin Wang break; 751938e719abSBaolin Wang } 752038e719abSBaolin Wang } 752138e719abSBaolin Wang 7522cf11e85fSRoman Gushchin return 0; 7523cf11e85fSRoman Gushchin } 7524cf11e85fSRoman Gushchin 7525cf11e85fSRoman Gushchin early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); 7526cf11e85fSRoman Gushchin 7527cf11e85fSRoman Gushchin void __init hugetlb_cma_reserve(int order) 7528cf11e85fSRoman Gushchin { 7529cf11e85fSRoman Gushchin unsigned long size, reserved, per_node; 753038e719abSBaolin Wang bool node_specific_cma_alloc = false; 7531cf11e85fSRoman Gushchin int nid; 7532cf11e85fSRoman Gushchin 7533cf11e85fSRoman Gushchin cma_reserve_called = true; 7534cf11e85fSRoman Gushchin 7535cf11e85fSRoman Gushchin if (!hugetlb_cma_size) 7536cf11e85fSRoman Gushchin return; 7537cf11e85fSRoman Gushchin 753838e719abSBaolin Wang for (nid = 0; nid < MAX_NUMNODES; nid++) { 753938e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 754038e719abSBaolin Wang continue; 754138e719abSBaolin Wang 754230a51400SPeng Liu if (!node_online(nid)) { 754338e719abSBaolin Wang pr_warn("hugetlb_cma: invalid node %d specified\n", nid); 754438e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 754538e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 754638e719abSBaolin Wang continue; 754738e719abSBaolin Wang } 754838e719abSBaolin Wang 754938e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) { 755038e719abSBaolin Wang pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n", 755138e719abSBaolin Wang nid, (PAGE_SIZE << order) / SZ_1M); 755238e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 755338e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 755438e719abSBaolin Wang } else { 755538e719abSBaolin Wang node_specific_cma_alloc = true; 755638e719abSBaolin Wang } 755738e719abSBaolin Wang } 755838e719abSBaolin Wang 755938e719abSBaolin Wang /* Validate the CMA size again in case some invalid nodes specified. */ 756038e719abSBaolin Wang if (!hugetlb_cma_size) 756138e719abSBaolin Wang return; 756238e719abSBaolin Wang 7563cf11e85fSRoman Gushchin if (hugetlb_cma_size < (PAGE_SIZE << order)) { 7564cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", 7565cf11e85fSRoman Gushchin (PAGE_SIZE << order) / SZ_1M); 7566a01f4390SMike Kravetz hugetlb_cma_size = 0; 7567cf11e85fSRoman Gushchin return; 7568cf11e85fSRoman Gushchin } 7569cf11e85fSRoman Gushchin 757038e719abSBaolin Wang if (!node_specific_cma_alloc) { 7571cf11e85fSRoman Gushchin /* 7572cf11e85fSRoman Gushchin * If 3 GB area is requested on a machine with 4 numa nodes, 7573cf11e85fSRoman Gushchin * let's allocate 1 GB on first three nodes and ignore the last one. 7574cf11e85fSRoman Gushchin */ 7575cf11e85fSRoman Gushchin per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); 7576cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", 7577cf11e85fSRoman Gushchin hugetlb_cma_size / SZ_1M, per_node / SZ_1M); 757838e719abSBaolin Wang } 7579cf11e85fSRoman Gushchin 7580cf11e85fSRoman Gushchin reserved = 0; 758130a51400SPeng Liu for_each_online_node(nid) { 7582cf11e85fSRoman Gushchin int res; 75832281f797SBarry Song char name[CMA_MAX_NAME]; 7584cf11e85fSRoman Gushchin 758538e719abSBaolin Wang if (node_specific_cma_alloc) { 758638e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 758738e719abSBaolin Wang continue; 758838e719abSBaolin Wang 758938e719abSBaolin Wang size = hugetlb_cma_size_in_node[nid]; 759038e719abSBaolin Wang } else { 7591cf11e85fSRoman Gushchin size = min(per_node, hugetlb_cma_size - reserved); 759238e719abSBaolin Wang } 759338e719abSBaolin Wang 7594cf11e85fSRoman Gushchin size = round_up(size, PAGE_SIZE << order); 7595cf11e85fSRoman Gushchin 75962281f797SBarry Song snprintf(name, sizeof(name), "hugetlb%d", nid); 7597a01f4390SMike Kravetz /* 7598a01f4390SMike Kravetz * Note that 'order per bit' is based on smallest size that 7599a01f4390SMike Kravetz * may be returned to CMA allocator in the case of 7600a01f4390SMike Kravetz * huge page demotion. 7601a01f4390SMike Kravetz */ 7602a01f4390SMike Kravetz res = cma_declare_contiguous_nid(0, size, 0, 7603a01f4390SMike Kravetz PAGE_SIZE << HUGETLB_PAGE_ORDER, 760429d0f41dSBarry Song 0, false, name, 7605cf11e85fSRoman Gushchin &hugetlb_cma[nid], nid); 7606cf11e85fSRoman Gushchin if (res) { 7607cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: reservation failed: err %d, node %d", 7608cf11e85fSRoman Gushchin res, nid); 7609cf11e85fSRoman Gushchin continue; 7610cf11e85fSRoman Gushchin } 7611cf11e85fSRoman Gushchin 7612cf11e85fSRoman Gushchin reserved += size; 7613cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", 7614cf11e85fSRoman Gushchin size / SZ_1M, nid); 7615cf11e85fSRoman Gushchin 7616cf11e85fSRoman Gushchin if (reserved >= hugetlb_cma_size) 7617cf11e85fSRoman Gushchin break; 7618cf11e85fSRoman Gushchin } 7619a01f4390SMike Kravetz 7620a01f4390SMike Kravetz if (!reserved) 7621a01f4390SMike Kravetz /* 7622a01f4390SMike Kravetz * hugetlb_cma_size is used to determine if allocations from 7623a01f4390SMike Kravetz * cma are possible. Set to zero if no cma regions are set up. 7624a01f4390SMike Kravetz */ 7625a01f4390SMike Kravetz hugetlb_cma_size = 0; 7626cf11e85fSRoman Gushchin } 7627cf11e85fSRoman Gushchin 7628263b8998SMiaohe Lin static void __init hugetlb_cma_check(void) 7629cf11e85fSRoman Gushchin { 7630cf11e85fSRoman Gushchin if (!hugetlb_cma_size || cma_reserve_called) 7631cf11e85fSRoman Gushchin return; 7632cf11e85fSRoman Gushchin 7633cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); 7634cf11e85fSRoman Gushchin } 7635cf11e85fSRoman Gushchin 7636cf11e85fSRoman Gushchin #endif /* CONFIG_CMA */ 7637