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 205327c73ae7SAndrea Arcangeli /* 205427c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 205527c73ae7SAndrea Arcangeli * normal or transparent huge pages. 205627c73ae7SAndrea Arcangeli */ 205727c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 205827c73ae7SAndrea Arcangeli { 20592d678c64SMatthew Wilcox (Oracle) struct folio *folio = (struct folio *)page_head; 20602d678c64SMatthew Wilcox (Oracle) if (!folio_test_large(folio)) 206127c73ae7SAndrea Arcangeli return 0; 206227c73ae7SAndrea Arcangeli 20632d678c64SMatthew Wilcox (Oracle) return folio->_folio_dtor == HUGETLB_PAGE_DTOR; 206427c73ae7SAndrea Arcangeli } 20654e936eccSDavid Howells EXPORT_SYMBOL_GPL(PageHeadHuge); 206627c73ae7SAndrea Arcangeli 2067c0d0381aSMike Kravetz /* 2068c0d0381aSMike Kravetz * Find and lock address space (mapping) in write mode. 2069c0d0381aSMike Kravetz * 2070336bf30eSMike Kravetz * Upon entry, the page is locked which means that page_mapping() is 2071336bf30eSMike Kravetz * stable. Due to locking order, we can only trylock_write. If we can 2072336bf30eSMike Kravetz * not get the lock, simply return NULL to caller. 2073c0d0381aSMike Kravetz */ 2074c0d0381aSMike Kravetz struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) 2075c0d0381aSMike Kravetz { 2076336bf30eSMike Kravetz struct address_space *mapping = page_mapping(hpage); 2077c0d0381aSMike Kravetz 2078c0d0381aSMike Kravetz if (!mapping) 2079c0d0381aSMike Kravetz return mapping; 2080c0d0381aSMike Kravetz 2081c0d0381aSMike Kravetz if (i_mmap_trylock_write(mapping)) 2082c0d0381aSMike Kravetz return mapping; 2083c0d0381aSMike Kravetz 2084c0d0381aSMike Kravetz return NULL; 2085c0d0381aSMike Kravetz } 2086c0d0381aSMike Kravetz 2087fe19bd3dSHugh Dickins pgoff_t hugetlb_basepage_index(struct page *page) 208813d60f4bSZhang Yi { 208913d60f4bSZhang Yi struct page *page_head = compound_head(page); 209013d60f4bSZhang Yi pgoff_t index = page_index(page_head); 209113d60f4bSZhang Yi unsigned long compound_idx; 209213d60f4bSZhang Yi 209313d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 209413d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 209513d60f4bSZhang Yi else 209613d60f4bSZhang Yi compound_idx = page - page_head; 209713d60f4bSZhang Yi 209813d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 209913d60f4bSZhang Yi } 210013d60f4bSZhang Yi 210119fc1a7eSSidhartha Kumar static struct folio *alloc_buddy_hugetlb_folio(struct hstate *h, 2102f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 2103f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 21041da177e4SLinus Torvalds { 2105af0fb9dfSMichal Hocko int order = huge_page_order(h); 21061da177e4SLinus Torvalds struct page *page; 2107f60858f9SMike Kravetz bool alloc_try_hard = true; 21082b21624fSMike Kravetz bool retry = true; 2109f96efd58SJoe Jin 2110f60858f9SMike Kravetz /* 2111f60858f9SMike Kravetz * By default we always try hard to allocate the page with 2112f60858f9SMike Kravetz * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in 2113f60858f9SMike Kravetz * a loop (to adjust global huge page counts) and previous allocation 2114f60858f9SMike Kravetz * failed, do not continue to try hard on the same node. Use the 2115f60858f9SMike Kravetz * node_alloc_noretry bitmap to manage this state information. 2116f60858f9SMike Kravetz */ 2117f60858f9SMike Kravetz if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) 2118f60858f9SMike Kravetz alloc_try_hard = false; 2119f60858f9SMike Kravetz gfp_mask |= __GFP_COMP|__GFP_NOWARN; 2120f60858f9SMike Kravetz if (alloc_try_hard) 2121f60858f9SMike Kravetz gfp_mask |= __GFP_RETRY_MAYFAIL; 2122af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 2123af0fb9dfSMichal Hocko nid = numa_mem_id(); 21242b21624fSMike Kravetz retry: 212584172f4bSMatthew Wilcox (Oracle) page = __alloc_pages(gfp_mask, order, nid, nmask); 21262b21624fSMike Kravetz 21272b21624fSMike Kravetz /* Freeze head page */ 21282b21624fSMike Kravetz if (page && !page_ref_freeze(page, 1)) { 21292b21624fSMike Kravetz __free_pages(page, order); 21302b21624fSMike Kravetz if (retry) { /* retry once */ 21312b21624fSMike Kravetz retry = false; 21322b21624fSMike Kravetz goto retry; 21332b21624fSMike Kravetz } 21342b21624fSMike Kravetz /* WOW! twice in a row. */ 21352b21624fSMike Kravetz pr_warn("HugeTLB head page unexpected inflated ref count\n"); 21362b21624fSMike Kravetz page = NULL; 21372b21624fSMike Kravetz } 21382b21624fSMike Kravetz 2139f60858f9SMike Kravetz /* 2140f60858f9SMike Kravetz * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this 2141f60858f9SMike Kravetz * indicates an overall state change. Clear bit so that we resume 2142f60858f9SMike Kravetz * normal 'try hard' allocations. 2143f60858f9SMike Kravetz */ 2144f60858f9SMike Kravetz if (node_alloc_noretry && page && !alloc_try_hard) 2145f60858f9SMike Kravetz node_clear(nid, *node_alloc_noretry); 2146f60858f9SMike Kravetz 2147f60858f9SMike Kravetz /* 2148f60858f9SMike Kravetz * If we tried hard to get a page but failed, set bit so that 2149f60858f9SMike Kravetz * subsequent attempts will not try as hard until there is an 2150f60858f9SMike Kravetz * overall state change. 2151f60858f9SMike Kravetz */ 2152f60858f9SMike Kravetz if (node_alloc_noretry && !page && alloc_try_hard) 2153f60858f9SMike Kravetz node_set(nid, *node_alloc_noretry); 2154f60858f9SMike Kravetz 215519fc1a7eSSidhartha Kumar if (!page) { 215619fc1a7eSSidhartha Kumar __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 215719fc1a7eSSidhartha Kumar return NULL; 215819fc1a7eSSidhartha Kumar } 215919fc1a7eSSidhartha Kumar 216019fc1a7eSSidhartha Kumar __count_vm_event(HTLB_BUDDY_PGALLOC); 216119fc1a7eSSidhartha Kumar return page_folio(page); 216263b4613cSNishanth Aravamudan } 216363b4613cSNishanth Aravamudan 2164af0fb9dfSMichal Hocko /* 21650c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 21660c397daeSMichal Hocko * should use this function to get new hugetlb pages 21672b21624fSMike Kravetz * 21682b21624fSMike Kravetz * Note that returned page is 'frozen': ref count of head page and all tail 21692b21624fSMike Kravetz * pages is zero. 21700c397daeSMichal Hocko */ 217119fc1a7eSSidhartha Kumar static struct folio *alloc_fresh_hugetlb_folio(struct hstate *h, 2172f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 2173f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 21740c397daeSMichal Hocko { 21757f325a8dSSidhartha Kumar struct folio *folio; 21767118fc29SMike Kravetz bool retry = false; 21770c397daeSMichal Hocko 21787118fc29SMike Kravetz retry: 21790c397daeSMichal Hocko if (hstate_is_gigantic(h)) 218019fc1a7eSSidhartha Kumar folio = alloc_gigantic_folio(h, gfp_mask, nid, nmask); 21810c397daeSMichal Hocko else 218219fc1a7eSSidhartha Kumar folio = alloc_buddy_hugetlb_folio(h, gfp_mask, 2183f60858f9SMike Kravetz nid, nmask, node_alloc_noretry); 218419fc1a7eSSidhartha Kumar if (!folio) 21850c397daeSMichal Hocko return NULL; 21867118fc29SMike Kravetz if (hstate_is_gigantic(h)) { 2187d1c60955SSidhartha Kumar if (!prep_compound_gigantic_folio(folio, huge_page_order(h))) { 21887118fc29SMike Kravetz /* 21897118fc29SMike Kravetz * Rare failure to convert pages to compound page. 21907118fc29SMike Kravetz * Free pages and try again - ONCE! 21917118fc29SMike Kravetz */ 21927f325a8dSSidhartha Kumar free_gigantic_folio(folio, huge_page_order(h)); 21937118fc29SMike Kravetz if (!retry) { 21947118fc29SMike Kravetz retry = true; 21957118fc29SMike Kravetz goto retry; 21967118fc29SMike Kravetz } 21977118fc29SMike Kravetz return NULL; 21987118fc29SMike Kravetz } 21997118fc29SMike Kravetz } 2200d1c60955SSidhartha Kumar prep_new_hugetlb_folio(h, folio, folio_nid(folio)); 22010c397daeSMichal Hocko 220219fc1a7eSSidhartha Kumar return folio; 22030c397daeSMichal Hocko } 22040c397daeSMichal Hocko 22050c397daeSMichal Hocko /* 2206af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 2207af0fb9dfSMichal Hocko * manner. 2208af0fb9dfSMichal Hocko */ 2209f60858f9SMike Kravetz static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 2210f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 2211b2261026SJoonsoo Kim { 221219fc1a7eSSidhartha Kumar struct folio *folio; 2213b2261026SJoonsoo Kim int nr_nodes, node; 2214af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2215b2261026SJoonsoo Kim 2216b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 221719fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, node, 221819fc1a7eSSidhartha Kumar nodes_allowed, node_alloc_noretry); 221919fc1a7eSSidhartha Kumar if (folio) { 222019fc1a7eSSidhartha Kumar free_huge_page(&folio->page); /* free it into the hugepage allocator */ 222119fc1a7eSSidhartha Kumar return 1; 222219fc1a7eSSidhartha Kumar } 2223b2261026SJoonsoo Kim } 2224b2261026SJoonsoo Kim 2225af0fb9dfSMichal Hocko return 0; 2226b2261026SJoonsoo Kim } 2227b2261026SJoonsoo Kim 2228e8c5c824SLee Schermerhorn /* 222910c6ec49SMike Kravetz * Remove huge page from pool from next node to free. Attempt to keep 223010c6ec49SMike Kravetz * persistent huge pages more or less balanced over allowed nodes. 223110c6ec49SMike Kravetz * This routine only 'removes' the hugetlb page. The caller must make 223210c6ec49SMike Kravetz * an additional call to free the page to low level allocators. 2233e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 2234e8c5c824SLee Schermerhorn */ 223510c6ec49SMike Kravetz static struct page *remove_pool_huge_page(struct hstate *h, 223610c6ec49SMike Kravetz nodemask_t *nodes_allowed, 22376ae11b27SLee Schermerhorn bool acct_surplus) 2238e8c5c824SLee Schermerhorn { 2239b2261026SJoonsoo Kim int nr_nodes, node; 224010c6ec49SMike Kravetz struct page *page = NULL; 2241cfd5082bSSidhartha Kumar struct folio *folio; 2242e8c5c824SLee Schermerhorn 22439487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2244b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2245685f3457SLee Schermerhorn /* 2246685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 2247685f3457SLee Schermerhorn * nodes with surplus pages. 2248685f3457SLee Schermerhorn */ 2249b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 2250b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 225110c6ec49SMike Kravetz page = list_entry(h->hugepage_freelists[node].next, 2252e8c5c824SLee Schermerhorn struct page, lru); 2253cfd5082bSSidhartha Kumar folio = page_folio(page); 2254cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, acct_surplus); 22559a76db09SLee Schermerhorn break; 2256e8c5c824SLee Schermerhorn } 2257b2261026SJoonsoo Kim } 2258e8c5c824SLee Schermerhorn 225910c6ec49SMike Kravetz return page; 2260e8c5c824SLee Schermerhorn } 2261e8c5c824SLee Schermerhorn 2262c8721bbbSNaoya Horiguchi /* 2263c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 2264faf53defSNaoya Horiguchi * nothing for in-use hugepages and non-hugepages. 2265faf53defSNaoya Horiguchi * This function returns values like below: 2266faf53defSNaoya Horiguchi * 2267ad2fa371SMuchun Song * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages 2268ad2fa371SMuchun Song * when the system is under memory pressure and the feature of 2269ad2fa371SMuchun Song * freeing unused vmemmap pages associated with each hugetlb page 2270ad2fa371SMuchun Song * is enabled. 2271faf53defSNaoya Horiguchi * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 2272faf53defSNaoya Horiguchi * (allocated or reserved.) 2273faf53defSNaoya Horiguchi * 0: successfully dissolved free hugepages or the page is not a 2274faf53defSNaoya Horiguchi * hugepage (considered as already dissolved) 2275c8721bbbSNaoya Horiguchi */ 2276c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 2277c8721bbbSNaoya Horiguchi { 22786bc9b564SNaoya Horiguchi int rc = -EBUSY; 22791a7cdab5SSidhartha Kumar struct folio *folio = page_folio(page); 2280082d5b6bSGerald Schaefer 22817ffddd49SMuchun Song retry: 2282faf53defSNaoya Horiguchi /* Not to disrupt normal path by vainly holding hugetlb_lock */ 22831a7cdab5SSidhartha Kumar if (!folio_test_hugetlb(folio)) 2284faf53defSNaoya Horiguchi return 0; 2285faf53defSNaoya Horiguchi 2286db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 22871a7cdab5SSidhartha Kumar if (!folio_test_hugetlb(folio)) { 2288faf53defSNaoya Horiguchi rc = 0; 2289faf53defSNaoya Horiguchi goto out; 2290faf53defSNaoya Horiguchi } 2291faf53defSNaoya Horiguchi 22921a7cdab5SSidhartha Kumar if (!folio_ref_count(folio)) { 22931a7cdab5SSidhartha Kumar struct hstate *h = folio_hstate(folio); 22948346d69dSXin Hao if (!available_huge_pages(h)) 2295082d5b6bSGerald Schaefer goto out; 22967ffddd49SMuchun Song 22977ffddd49SMuchun Song /* 22987ffddd49SMuchun Song * We should make sure that the page is already on the free list 22997ffddd49SMuchun Song * when it is dissolved. 23007ffddd49SMuchun Song */ 23011a7cdab5SSidhartha Kumar if (unlikely(!folio_test_hugetlb_freed(folio))) { 2302db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 23037ffddd49SMuchun Song cond_resched(); 23047ffddd49SMuchun Song 23057ffddd49SMuchun Song /* 23067ffddd49SMuchun Song * Theoretically, we should return -EBUSY when we 23077ffddd49SMuchun Song * encounter this race. In fact, we have a chance 23087ffddd49SMuchun Song * to successfully dissolve the page if we do a 23097ffddd49SMuchun Song * retry. Because the race window is quite small. 23107ffddd49SMuchun Song * If we seize this opportunity, it is an optimization 23117ffddd49SMuchun Song * for increasing the success rate of dissolving page. 23127ffddd49SMuchun Song */ 23137ffddd49SMuchun Song goto retry; 23147ffddd49SMuchun Song } 23157ffddd49SMuchun Song 2316cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, false); 2317ad2fa371SMuchun Song h->max_huge_pages--; 2318ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2319ad2fa371SMuchun Song 2320c3114a84SAnshuman Khandual /* 2321d6ef19e2SSidhartha Kumar * Normally update_and_free_hugtlb_folio will allocate required vmemmmap 2322d6ef19e2SSidhartha Kumar * before freeing the page. update_and_free_hugtlb_folio will fail to 2323ad2fa371SMuchun Song * free the page if it can not allocate required vmemmap. We 2324ad2fa371SMuchun Song * need to adjust max_huge_pages if the page is not freed. 2325ad2fa371SMuchun Song * Attempt to allocate vmemmmap here so that we can take 2326ad2fa371SMuchun Song * appropriate action on failure. 2327ad2fa371SMuchun Song */ 23281a7cdab5SSidhartha Kumar rc = hugetlb_vmemmap_restore(h, &folio->page); 2329ad2fa371SMuchun Song if (!rc) { 2330d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, folio, false); 2331ad2fa371SMuchun Song } else { 2332ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 23332f6c57d6SSidhartha Kumar add_hugetlb_folio(h, folio, false); 2334ad2fa371SMuchun Song h->max_huge_pages++; 2335ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2336ad2fa371SMuchun Song } 2337ad2fa371SMuchun Song 2338ad2fa371SMuchun Song return rc; 2339c8721bbbSNaoya Horiguchi } 2340082d5b6bSGerald Schaefer out: 2341db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2342082d5b6bSGerald Schaefer return rc; 2343c8721bbbSNaoya Horiguchi } 2344c8721bbbSNaoya Horiguchi 2345c8721bbbSNaoya Horiguchi /* 2346c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 2347c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 23482247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 23492247bb33SGerald Schaefer * part of it lies within the given range. 2350082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 2351082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 2352c8721bbbSNaoya Horiguchi */ 2353082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 2354c8721bbbSNaoya Horiguchi { 2355c8721bbbSNaoya Horiguchi unsigned long pfn; 2356eb03aa00SGerald Schaefer struct page *page; 2357082d5b6bSGerald Schaefer int rc = 0; 2358dc2628f3SMuchun Song unsigned int order; 2359dc2628f3SMuchun Song struct hstate *h; 2360c8721bbbSNaoya Horiguchi 2361d0177639SLi Zhong if (!hugepages_supported()) 2362082d5b6bSGerald Schaefer return rc; 2363d0177639SLi Zhong 2364dc2628f3SMuchun Song order = huge_page_order(&default_hstate); 2365dc2628f3SMuchun Song for_each_hstate(h) 2366dc2628f3SMuchun Song order = min(order, huge_page_order(h)); 2367dc2628f3SMuchun Song 2368dc2628f3SMuchun Song for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) { 2369eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 2370eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 2371eb03aa00SGerald Schaefer if (rc) 2372082d5b6bSGerald Schaefer break; 2373eb03aa00SGerald Schaefer } 2374082d5b6bSGerald Schaefer 2375082d5b6bSGerald Schaefer return rc; 2376c8721bbbSNaoya Horiguchi } 2377c8721bbbSNaoya Horiguchi 2378ab5ac90aSMichal Hocko /* 2379ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 2380ab5ac90aSMichal Hocko */ 23813a740e8bSSidhartha Kumar static struct folio *alloc_surplus_hugetlb_folio(struct hstate *h, 23823a740e8bSSidhartha Kumar gfp_t gfp_mask, int nid, nodemask_t *nmask) 23837893d1d5SAdam Litke { 238419fc1a7eSSidhartha Kumar struct folio *folio = NULL; 23857893d1d5SAdam Litke 2386bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2387aa888a74SAndi Kleen return NULL; 2388aa888a74SAndi Kleen 2389db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 23909980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 23919980d744SMichal Hocko goto out_unlock; 2392db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2393d1c3fb1fSNishanth Aravamudan 239419fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL); 239519fc1a7eSSidhartha Kumar if (!folio) 23960c397daeSMichal Hocko return NULL; 2397d1c3fb1fSNishanth Aravamudan 2398db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 23999980d744SMichal Hocko /* 24009980d744SMichal Hocko * We could have raced with the pool size change. 24019980d744SMichal Hocko * Double check that and simply deallocate the new page 24029980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 24039980d744SMichal Hocko * temporary page to workaround the nasty free_huge_page 24049980d744SMichal Hocko * codeflow 24059980d744SMichal Hocko */ 24069980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 240719fc1a7eSSidhartha Kumar folio_set_hugetlb_temporary(folio); 2408db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 240919fc1a7eSSidhartha Kumar free_huge_page(&folio->page); 24102bf753e6SKai Shen return NULL; 2411b65a4edaSMike Kravetz } 2412b65a4edaSMike Kravetz 24139980d744SMichal Hocko h->surplus_huge_pages++; 241419fc1a7eSSidhartha Kumar h->surplus_huge_pages_node[folio_nid(folio)]++; 24159980d744SMichal Hocko 24169980d744SMichal Hocko out_unlock: 2417db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 24187893d1d5SAdam Litke 24193a740e8bSSidhartha Kumar return folio; 24207893d1d5SAdam Litke } 24217893d1d5SAdam Litke 2422e37d3e83SSidhartha Kumar static struct folio *alloc_migrate_hugetlb_folio(struct hstate *h, gfp_t gfp_mask, 2423ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 2424ab5ac90aSMichal Hocko { 242519fc1a7eSSidhartha Kumar struct folio *folio; 2426ab5ac90aSMichal Hocko 2427ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 2428ab5ac90aSMichal Hocko return NULL; 2429ab5ac90aSMichal Hocko 243019fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL); 243119fc1a7eSSidhartha Kumar if (!folio) 2432ab5ac90aSMichal Hocko return NULL; 2433ab5ac90aSMichal Hocko 24342b21624fSMike Kravetz /* fresh huge pages are frozen */ 243519fc1a7eSSidhartha Kumar folio_ref_unfreeze(folio, 1); 2436ab5ac90aSMichal Hocko /* 2437ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 2438ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 2439ab5ac90aSMichal Hocko */ 244019fc1a7eSSidhartha Kumar folio_set_hugetlb_temporary(folio); 2441ab5ac90aSMichal Hocko 2442e37d3e83SSidhartha Kumar return folio; 2443ab5ac90aSMichal Hocko } 2444ab5ac90aSMichal Hocko 2445e4e574b7SAdam Litke /* 2446099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 2447099730d6SDave Hansen */ 2448e0ec90eeSDave Hansen static 2449ff7d853bSSidhartha Kumar struct folio *alloc_buddy_hugetlb_folio_with_mpol(struct hstate *h, 2450099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 2451099730d6SDave Hansen { 24523a740e8bSSidhartha Kumar struct folio *folio = NULL; 2453aaf14e40SMichal Hocko struct mempolicy *mpol; 2454aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 2455aaf14e40SMichal Hocko int nid; 2456aaf14e40SMichal Hocko nodemask_t *nodemask; 2457aaf14e40SMichal Hocko 2458aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 2459cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 2460cfcaa66fSBen Widawsky gfp_t gfp = gfp_mask | __GFP_NOWARN; 2461cfcaa66fSBen Widawsky 2462cfcaa66fSBen Widawsky gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL); 24633a740e8bSSidhartha Kumar folio = alloc_surplus_hugetlb_folio(h, gfp, nid, nodemask); 2464cfcaa66fSBen Widawsky 2465cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 2466cfcaa66fSBen Widawsky nodemask = NULL; 2467cfcaa66fSBen Widawsky } 2468cfcaa66fSBen Widawsky 24693a740e8bSSidhartha Kumar if (!folio) 24703a740e8bSSidhartha Kumar folio = alloc_surplus_hugetlb_folio(h, gfp_mask, nid, nodemask); 2471aaf14e40SMichal Hocko mpol_cond_put(mpol); 2472ff7d853bSSidhartha Kumar return folio; 2473099730d6SDave Hansen } 2474099730d6SDave Hansen 2475e37d3e83SSidhartha Kumar /* folio migration callback function */ 2476e37d3e83SSidhartha Kumar struct folio *alloc_hugetlb_folio_nodemask(struct hstate *h, int preferred_nid, 2477d92bbc27SJoonsoo Kim nodemask_t *nmask, gfp_t gfp_mask) 24784db9b2efSMichal Hocko { 2479db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 24808346d69dSXin Hao if (available_huge_pages(h)) { 2481a36f1e90SSidhartha Kumar struct folio *folio; 24823e59fcb0SMichal Hocko 2483a36f1e90SSidhartha Kumar folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask, 2484a36f1e90SSidhartha Kumar preferred_nid, nmask); 2485a36f1e90SSidhartha Kumar if (folio) { 2486db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2487e37d3e83SSidhartha Kumar return folio; 24884db9b2efSMichal Hocko } 24894db9b2efSMichal Hocko } 2490db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 24914db9b2efSMichal Hocko 2492e37d3e83SSidhartha Kumar return alloc_migrate_hugetlb_folio(h, gfp_mask, preferred_nid, nmask); 24934db9b2efSMichal Hocko } 24944db9b2efSMichal Hocko 2495ebd63723SMichal Hocko /* mempolicy aware migration callback */ 2496389c8178SMichal Hocko struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, 2497389c8178SMichal Hocko unsigned long address) 2498ebd63723SMichal Hocko { 2499ebd63723SMichal Hocko struct mempolicy *mpol; 2500ebd63723SMichal Hocko nodemask_t *nodemask; 2501e37d3e83SSidhartha Kumar struct folio *folio; 2502ebd63723SMichal Hocko gfp_t gfp_mask; 2503ebd63723SMichal Hocko int node; 2504ebd63723SMichal Hocko 2505ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 2506ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 2507e37d3e83SSidhartha Kumar folio = alloc_hugetlb_folio_nodemask(h, node, nodemask, gfp_mask); 2508ebd63723SMichal Hocko mpol_cond_put(mpol); 2509ebd63723SMichal Hocko 2510e37d3e83SSidhartha Kumar return &folio->page; 2511ebd63723SMichal Hocko } 2512ebd63723SMichal Hocko 2513bf50bab2SNaoya Horiguchi /* 251425985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 2515e4e574b7SAdam Litke * of size 'delta'. 2516e4e574b7SAdam Litke */ 25170a4f3d1bSLiu Xiang static int gather_surplus_pages(struct hstate *h, long delta) 25181b2a1e7bSJules Irenge __must_hold(&hugetlb_lock) 2519e4e574b7SAdam Litke { 252034665341SMiaohe Lin LIST_HEAD(surplus_list); 25213a740e8bSSidhartha Kumar struct folio *folio; 2522e4e574b7SAdam Litke struct page *page, *tmp; 25230a4f3d1bSLiu Xiang int ret; 25240a4f3d1bSLiu Xiang long i; 25250a4f3d1bSLiu Xiang long needed, allocated; 252628073b02SHillf Danton bool alloc_ok = true; 2527e4e574b7SAdam Litke 25289487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2529a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 2530ac09b3a1SAdam Litke if (needed <= 0) { 2531a5516438SAndi Kleen h->resv_huge_pages += delta; 2532e4e574b7SAdam Litke return 0; 2533ac09b3a1SAdam Litke } 2534e4e574b7SAdam Litke 2535e4e574b7SAdam Litke allocated = 0; 2536e4e574b7SAdam Litke 2537e4e574b7SAdam Litke ret = -ENOMEM; 2538e4e574b7SAdam Litke retry: 2539db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2540e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 25413a740e8bSSidhartha Kumar folio = alloc_surplus_hugetlb_folio(h, htlb_alloc_mask(h), 25422b21624fSMike Kravetz NUMA_NO_NODE, NULL); 25433a740e8bSSidhartha Kumar if (!folio) { 254428073b02SHillf Danton alloc_ok = false; 254528073b02SHillf Danton break; 254628073b02SHillf Danton } 25473a740e8bSSidhartha Kumar list_add(&folio->lru, &surplus_list); 254869ed779aSDavid Rientjes cond_resched(); 2549e4e574b7SAdam Litke } 255028073b02SHillf Danton allocated += i; 2551e4e574b7SAdam Litke 2552e4e574b7SAdam Litke /* 2553e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 2554e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 2555e4e574b7SAdam Litke */ 2556db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2557a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 2558a5516438SAndi Kleen (h->free_huge_pages + allocated); 255928073b02SHillf Danton if (needed > 0) { 256028073b02SHillf Danton if (alloc_ok) 2561e4e574b7SAdam Litke goto retry; 256228073b02SHillf Danton /* 256328073b02SHillf Danton * We were not able to allocate enough pages to 256428073b02SHillf Danton * satisfy the entire reservation so we free what 256528073b02SHillf Danton * we've allocated so far. 256628073b02SHillf Danton */ 256728073b02SHillf Danton goto free; 256828073b02SHillf Danton } 2569e4e574b7SAdam Litke /* 2570e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 257125985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 2572e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 2573ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 2574ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 2575ac09b3a1SAdam Litke * before they are reserved. 2576e4e574b7SAdam Litke */ 2577e4e574b7SAdam Litke needed += allocated; 2578a5516438SAndi Kleen h->resv_huge_pages += delta; 2579e4e574b7SAdam Litke ret = 0; 2580a9869b83SNaoya Horiguchi 258119fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 258219fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 258319fc3f0aSAdam Litke if ((--needed) < 0) 258419fc3f0aSAdam Litke break; 2585b65a4edaSMike Kravetz /* Add the page to the hugetlb allocator */ 2586240d67a8SSidhartha Kumar enqueue_hugetlb_folio(h, page_folio(page)); 258719fc3f0aSAdam Litke } 258828073b02SHillf Danton free: 2589db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 259019fc3f0aSAdam Litke 2591b65a4edaSMike Kravetz /* 2592b65a4edaSMike Kravetz * Free unnecessary surplus pages to the buddy allocator. 2593b65a4edaSMike Kravetz * Pages have no ref count, call free_huge_page directly. 2594b65a4edaSMike Kravetz */ 2595c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 2596b65a4edaSMike Kravetz free_huge_page(page); 2597db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2598e4e574b7SAdam Litke 2599e4e574b7SAdam Litke return ret; 2600e4e574b7SAdam Litke } 2601e4e574b7SAdam Litke 2602e4e574b7SAdam Litke /* 2603e5bbc8a6SMike Kravetz * This routine has two main purposes: 2604e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 2605e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 2606e5bbc8a6SMike Kravetz * to the associated reservation map. 2607e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 2608e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 2609e4e574b7SAdam Litke */ 2610a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 2611a5516438SAndi Kleen unsigned long unused_resv_pages) 2612e4e574b7SAdam Litke { 2613e4e574b7SAdam Litke unsigned long nr_pages; 261410c6ec49SMike Kravetz struct page *page; 261510c6ec49SMike Kravetz LIST_HEAD(page_list); 261610c6ec49SMike Kravetz 26179487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 261810c6ec49SMike Kravetz /* Uncommit the reservation */ 261910c6ec49SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 2620e4e574b7SAdam Litke 2621c0531714SNaoya Horiguchi if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 2622e5bbc8a6SMike Kravetz goto out; 2623aa888a74SAndi Kleen 2624e5bbc8a6SMike Kravetz /* 2625e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 2626e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 2627e5bbc8a6SMike Kravetz */ 2628a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 2629e4e574b7SAdam Litke 2630685f3457SLee Schermerhorn /* 2631685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 26329b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 26339b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 26349b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 263510c6ec49SMike Kravetz * remove_pool_huge_page() will balance the freed pages across the 26369b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 2637685f3457SLee Schermerhorn */ 2638685f3457SLee Schermerhorn while (nr_pages--) { 263910c6ec49SMike Kravetz page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1); 264010c6ec49SMike Kravetz if (!page) 2641e5bbc8a6SMike Kravetz goto out; 264210c6ec49SMike Kravetz 264310c6ec49SMike Kravetz list_add(&page->lru, &page_list); 2644e4e574b7SAdam Litke } 2645e5bbc8a6SMike Kravetz 2646e5bbc8a6SMike Kravetz out: 2647db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 264810c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 2649db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2650e4e574b7SAdam Litke } 2651e4e574b7SAdam Litke 26525e911373SMike Kravetz 2653c37f9fb1SAndy Whitcroft /* 2654feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 26555e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 2656cf3ad20bSMike Kravetz * 2657cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 2658cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 2659cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 2660cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 2661cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 2662feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 2663feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 2664feba16e2SMike Kravetz * is called in such cases. 2665cf3ad20bSMike Kravetz * 2666cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 2667cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 2668cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 2669cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 2670cf3ad20bSMike Kravetz * take appropriate action. 267196b96a96SMike Kravetz * 267296b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 267396b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 267496b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 267596b96a96SMike Kravetz * reservation exists. 2676846be085SMike Kravetz * 2677846be085SMike Kravetz * vma_del_reservation is used in error paths where an entry in the reserve 2678846be085SMike Kravetz * map was created during huge page allocation and must be removed. It is to 2679846be085SMike Kravetz * be called after calling vma_needs_reservation to determine if a reservation 2680846be085SMike Kravetz * exists. 2681c37f9fb1SAndy Whitcroft */ 26825e911373SMike Kravetz enum vma_resv_mode { 26835e911373SMike Kravetz VMA_NEEDS_RESV, 26845e911373SMike Kravetz VMA_COMMIT_RESV, 2685feba16e2SMike Kravetz VMA_END_RESV, 268696b96a96SMike Kravetz VMA_ADD_RESV, 2687846be085SMike Kravetz VMA_DEL_RESV, 26885e911373SMike Kravetz }; 2689cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 2690cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 26915e911373SMike Kravetz enum vma_resv_mode mode) 2692c37f9fb1SAndy Whitcroft { 26934e35f483SJoonsoo Kim struct resv_map *resv; 26944e35f483SJoonsoo Kim pgoff_t idx; 2695cf3ad20bSMike Kravetz long ret; 26960db9d74eSMina Almasry long dummy_out_regions_needed; 2697c37f9fb1SAndy Whitcroft 26984e35f483SJoonsoo Kim resv = vma_resv_map(vma); 26994e35f483SJoonsoo Kim if (!resv) 2700c37f9fb1SAndy Whitcroft return 1; 2701c37f9fb1SAndy Whitcroft 27024e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 27035e911373SMike Kravetz switch (mode) { 27045e911373SMike Kravetz case VMA_NEEDS_RESV: 27050db9d74eSMina Almasry ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); 27060db9d74eSMina Almasry /* We assume that vma_reservation_* routines always operate on 27070db9d74eSMina Almasry * 1 page, and that adding to resv map a 1 page entry can only 27080db9d74eSMina Almasry * ever require 1 region. 27090db9d74eSMina Almasry */ 27100db9d74eSMina Almasry VM_BUG_ON(dummy_out_regions_needed != 1); 27115e911373SMike Kravetz break; 27125e911373SMike Kravetz case VMA_COMMIT_RESV: 2713075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 27140db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 27150db9d74eSMina Almasry VM_BUG_ON(ret < 0); 27165e911373SMike Kravetz break; 2717feba16e2SMike Kravetz case VMA_END_RESV: 27180db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 27195e911373SMike Kravetz ret = 0; 27205e911373SMike Kravetz break; 272196b96a96SMike Kravetz case VMA_ADD_RESV: 27220db9d74eSMina Almasry if (vma->vm_flags & VM_MAYSHARE) { 2723075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 27240db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 27250db9d74eSMina Almasry VM_BUG_ON(ret < 0); 27260db9d74eSMina Almasry } else { 27270db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 272896b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 272996b96a96SMike Kravetz } 273096b96a96SMike Kravetz break; 2731846be085SMike Kravetz case VMA_DEL_RESV: 2732846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 2733846be085SMike Kravetz region_abort(resv, idx, idx + 1, 1); 2734846be085SMike Kravetz ret = region_del(resv, idx, idx + 1); 2735846be085SMike Kravetz } else { 2736846be085SMike Kravetz ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 2737846be085SMike Kravetz /* region_add calls of range 1 should never fail. */ 2738846be085SMike Kravetz VM_BUG_ON(ret < 0); 2739846be085SMike Kravetz } 2740846be085SMike Kravetz break; 27415e911373SMike Kravetz default: 27425e911373SMike Kravetz BUG(); 27435e911373SMike Kravetz } 274484afd99bSAndy Whitcroft 2745846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV) 2746cf3ad20bSMike Kravetz return ret; 274767961f9dSMike Kravetz /* 2748bf3d12b9SMiaohe Lin * We know private mapping must have HPAGE_RESV_OWNER set. 2749bf3d12b9SMiaohe Lin * 275067961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 275167961f9dSMike Kravetz * However, a file associated with mapping could have been 275267961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 275367961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 275467961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 275567961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 275667961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 275767961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 275867961f9dSMike Kravetz * reservation has already been consumed. As a result, the 275967961f9dSMike Kravetz * return value of this routine is the opposite of the 276067961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 276167961f9dSMike Kravetz */ 2762bf3d12b9SMiaohe Lin if (ret > 0) 276367961f9dSMike Kravetz return 0; 2764bf3d12b9SMiaohe Lin if (ret == 0) 276567961f9dSMike Kravetz return 1; 2766bf3d12b9SMiaohe Lin return ret; 276784afd99bSAndy Whitcroft } 2768cf3ad20bSMike Kravetz 2769cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 2770a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 2771c37f9fb1SAndy Whitcroft { 27725e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 2773cf3ad20bSMike Kravetz } 2774c37f9fb1SAndy Whitcroft 2775cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 2776cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 2777cf3ad20bSMike Kravetz { 27785e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 27795e911373SMike Kravetz } 27805e911373SMike Kravetz 2781feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 27825e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 27835e911373SMike Kravetz { 2784feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 2785c37f9fb1SAndy Whitcroft } 2786c37f9fb1SAndy Whitcroft 278796b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 278896b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 278996b96a96SMike Kravetz { 279096b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 279196b96a96SMike Kravetz } 279296b96a96SMike Kravetz 2793846be085SMike Kravetz static long vma_del_reservation(struct hstate *h, 2794846be085SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 279596b96a96SMike Kravetz { 2796846be085SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV); 2797846be085SMike Kravetz } 2798846be085SMike Kravetz 2799846be085SMike Kravetz /* 2800846be085SMike Kravetz * This routine is called to restore reservation information on error paths. 2801846be085SMike Kravetz * It should ONLY be called for pages allocated via alloc_huge_page(), and 2802846be085SMike Kravetz * the hugetlb mutex should remain held when calling this routine. 2803846be085SMike Kravetz * 2804846be085SMike Kravetz * It handles two specific cases: 2805846be085SMike Kravetz * 1) A reservation was in place and the page consumed the reservation. 2806846be085SMike Kravetz * HPageRestoreReserve is set in the page. 2807846be085SMike Kravetz * 2) No reservation was in place for the page, so HPageRestoreReserve is 2808846be085SMike Kravetz * not set. However, alloc_huge_page always updates the reserve map. 2809846be085SMike Kravetz * 2810846be085SMike Kravetz * In case 1, free_huge_page later in the error path will increment the 2811846be085SMike Kravetz * global reserve count. But, free_huge_page does not have enough context 2812846be085SMike Kravetz * to adjust the reservation map. This case deals primarily with private 2813846be085SMike Kravetz * mappings. Adjust the reserve map here to be consistent with global 2814846be085SMike Kravetz * reserve count adjustments to be made by free_huge_page. Make sure the 2815846be085SMike Kravetz * reserve map indicates there is a reservation present. 2816846be085SMike Kravetz * 2817846be085SMike Kravetz * In case 2, simply undo reserve map modifications done by alloc_huge_page. 2818846be085SMike Kravetz */ 2819846be085SMike Kravetz void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma, 2820846be085SMike Kravetz unsigned long address, struct page *page) 2821846be085SMike Kravetz { 28220ffdc38eSSidhartha Kumar struct folio *folio = page_folio(page); 282396b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 282496b96a96SMike Kravetz 28250ffdc38eSSidhartha Kumar if (folio_test_hugetlb_restore_reserve(folio)) { 2826846be085SMike Kravetz if (unlikely(rc < 0)) 282796b96a96SMike Kravetz /* 282896b96a96SMike Kravetz * Rare out of memory condition in reserve map 28290ffdc38eSSidhartha Kumar * manipulation. Clear hugetlb_restore_reserve so 28300ffdc38eSSidhartha Kumar * that global reserve count will not be incremented 283196b96a96SMike Kravetz * by free_huge_page. This will make it appear 28320ffdc38eSSidhartha Kumar * as though the reservation for this folio was 283396b96a96SMike Kravetz * consumed. This may prevent the task from 28340ffdc38eSSidhartha Kumar * faulting in the folio at a later time. This 283596b96a96SMike Kravetz * is better than inconsistent global huge page 283696b96a96SMike Kravetz * accounting of reserve counts. 283796b96a96SMike Kravetz */ 28380ffdc38eSSidhartha Kumar folio_clear_hugetlb_restore_reserve(folio); 2839846be085SMike Kravetz else if (rc) 2840846be085SMike Kravetz (void)vma_add_reservation(h, vma, address); 2841846be085SMike Kravetz else 2842846be085SMike Kravetz vma_end_reservation(h, vma, address); 2843846be085SMike Kravetz } else { 2844846be085SMike Kravetz if (!rc) { 284596b96a96SMike Kravetz /* 2846846be085SMike Kravetz * This indicates there is an entry in the reserve map 2847c7b1850dSMike Kravetz * not added by alloc_huge_page. We know it was added 2848846be085SMike Kravetz * before the alloc_huge_page call, otherwise 28490ffdc38eSSidhartha Kumar * hugetlb_restore_reserve would be set on the folio. 2850846be085SMike Kravetz * Remove the entry so that a subsequent allocation 2851846be085SMike Kravetz * does not consume a reservation. 285296b96a96SMike Kravetz */ 2853846be085SMike Kravetz rc = vma_del_reservation(h, vma, address); 2854846be085SMike Kravetz if (rc < 0) 2855846be085SMike Kravetz /* 2856846be085SMike Kravetz * VERY rare out of memory condition. Since 2857846be085SMike Kravetz * we can not delete the entry, set 28580ffdc38eSSidhartha Kumar * hugetlb_restore_reserve so that the reserve 28590ffdc38eSSidhartha Kumar * count will be incremented when the folio 2860846be085SMike Kravetz * is freed. This reserve will be consumed 2861846be085SMike Kravetz * on a subsequent allocation. 2862846be085SMike Kravetz */ 28630ffdc38eSSidhartha Kumar folio_set_hugetlb_restore_reserve(folio); 2864846be085SMike Kravetz } else if (rc < 0) { 2865846be085SMike Kravetz /* 2866846be085SMike Kravetz * Rare out of memory condition from 2867846be085SMike Kravetz * vma_needs_reservation call. Memory allocation is 2868846be085SMike Kravetz * only attempted if a new entry is needed. Therefore, 2869846be085SMike Kravetz * this implies there is not an entry in the 2870846be085SMike Kravetz * reserve map. 2871846be085SMike Kravetz * 2872846be085SMike Kravetz * For shared mappings, no entry in the map indicates 2873846be085SMike Kravetz * no reservation. We are done. 2874846be085SMike Kravetz */ 2875846be085SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 2876846be085SMike Kravetz /* 2877846be085SMike Kravetz * For private mappings, no entry indicates 2878846be085SMike Kravetz * a reservation is present. Since we can 28790ffdc38eSSidhartha Kumar * not add an entry, set hugetlb_restore_reserve 28800ffdc38eSSidhartha Kumar * on the folio so reserve count will be 2881846be085SMike Kravetz * incremented when freed. This reserve will 2882846be085SMike Kravetz * be consumed on a subsequent allocation. 2883846be085SMike Kravetz */ 28840ffdc38eSSidhartha Kumar folio_set_hugetlb_restore_reserve(folio); 288596b96a96SMike Kravetz } else 2886846be085SMike Kravetz /* 2887846be085SMike Kravetz * No reservation present, do nothing 2888846be085SMike Kravetz */ 288996b96a96SMike Kravetz vma_end_reservation(h, vma, address); 289096b96a96SMike Kravetz } 289196b96a96SMike Kravetz } 289296b96a96SMike Kravetz 2893369fa227SOscar Salvador /* 289419fc1a7eSSidhartha Kumar * alloc_and_dissolve_hugetlb_folio - Allocate a new folio and dissolve 289519fc1a7eSSidhartha Kumar * the old one 2896369fa227SOscar Salvador * @h: struct hstate old page belongs to 289719fc1a7eSSidhartha Kumar * @old_folio: Old folio to dissolve 2898ae37c7ffSOscar Salvador * @list: List to isolate the page in case we need to 2899369fa227SOscar Salvador * Returns 0 on success, otherwise negated error. 2900369fa227SOscar Salvador */ 290119fc1a7eSSidhartha Kumar static int alloc_and_dissolve_hugetlb_folio(struct hstate *h, 290219fc1a7eSSidhartha Kumar struct folio *old_folio, struct list_head *list) 2903369fa227SOscar Salvador { 2904369fa227SOscar Salvador gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2905de656ed3SSidhartha Kumar int nid = folio_nid(old_folio); 2906de656ed3SSidhartha Kumar struct folio *new_folio; 2907369fa227SOscar Salvador int ret = 0; 2908369fa227SOscar Salvador 2909369fa227SOscar Salvador /* 291019fc1a7eSSidhartha Kumar * Before dissolving the folio, we need to allocate a new one for the 291119fc1a7eSSidhartha Kumar * pool to remain stable. Here, we allocate the folio and 'prep' it 2912f41f2ed4SMuchun Song * by doing everything but actually updating counters and adding to 2913f41f2ed4SMuchun Song * the pool. This simplifies and let us do most of the processing 2914f41f2ed4SMuchun Song * under the lock. 2915369fa227SOscar Salvador */ 291619fc1a7eSSidhartha Kumar new_folio = alloc_buddy_hugetlb_folio(h, gfp_mask, nid, NULL, NULL); 291719fc1a7eSSidhartha Kumar if (!new_folio) 2918369fa227SOscar Salvador return -ENOMEM; 2919de656ed3SSidhartha Kumar __prep_new_hugetlb_folio(h, new_folio); 2920369fa227SOscar Salvador 2921369fa227SOscar Salvador retry: 2922369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2923de656ed3SSidhartha Kumar if (!folio_test_hugetlb(old_folio)) { 2924369fa227SOscar Salvador /* 292519fc1a7eSSidhartha Kumar * Freed from under us. Drop new_folio too. 2926369fa227SOscar Salvador */ 2927369fa227SOscar Salvador goto free_new; 2928de656ed3SSidhartha Kumar } else if (folio_ref_count(old_folio)) { 2929369fa227SOscar Salvador /* 293019fc1a7eSSidhartha Kumar * Someone has grabbed the folio, try to isolate it here. 2931ae37c7ffSOscar Salvador * Fail with -EBUSY if not possible. 2932369fa227SOscar Salvador */ 2933ae37c7ffSOscar Salvador spin_unlock_irq(&hugetlb_lock); 29346aa3a920SSidhartha Kumar ret = isolate_hugetlb(old_folio, list); 2935ae37c7ffSOscar Salvador spin_lock_irq(&hugetlb_lock); 2936369fa227SOscar Salvador goto free_new; 2937de656ed3SSidhartha Kumar } else if (!folio_test_hugetlb_freed(old_folio)) { 2938369fa227SOscar Salvador /* 293919fc1a7eSSidhartha Kumar * Folio's refcount is 0 but it has not been enqueued in the 2940369fa227SOscar Salvador * freelist yet. Race window is small, so we can succeed here if 2941369fa227SOscar Salvador * we retry. 2942369fa227SOscar Salvador */ 2943369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2944369fa227SOscar Salvador cond_resched(); 2945369fa227SOscar Salvador goto retry; 2946369fa227SOscar Salvador } else { 2947369fa227SOscar Salvador /* 294819fc1a7eSSidhartha Kumar * Ok, old_folio is still a genuine free hugepage. Remove it from 2949369fa227SOscar Salvador * the freelist and decrease the counters. These will be 2950369fa227SOscar Salvador * incremented again when calling __prep_account_new_huge_page() 2951240d67a8SSidhartha Kumar * and enqueue_hugetlb_folio() for new_folio. The counters will 2952240d67a8SSidhartha Kumar * remain stable since this happens under the lock. 2953369fa227SOscar Salvador */ 2954cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, old_folio, false); 2955369fa227SOscar Salvador 2956369fa227SOscar Salvador /* 295719fc1a7eSSidhartha Kumar * Ref count on new_folio is already zero as it was dropped 2958b65a4edaSMike Kravetz * earlier. It can be directly added to the pool free list. 2959369fa227SOscar Salvador */ 2960369fa227SOscar Salvador __prep_account_new_huge_page(h, nid); 2961240d67a8SSidhartha Kumar enqueue_hugetlb_folio(h, new_folio); 2962369fa227SOscar Salvador 2963369fa227SOscar Salvador /* 296419fc1a7eSSidhartha Kumar * Folio has been replaced, we can safely free the old one. 2965369fa227SOscar Salvador */ 2966369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2967d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, old_folio, false); 2968369fa227SOscar Salvador } 2969369fa227SOscar Salvador 2970369fa227SOscar Salvador return ret; 2971369fa227SOscar Salvador 2972369fa227SOscar Salvador free_new: 2973369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 297419fc1a7eSSidhartha Kumar /* Folio has a zero ref count, but needs a ref to be freed */ 2975de656ed3SSidhartha Kumar folio_ref_unfreeze(new_folio, 1); 2976d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, new_folio, false); 2977369fa227SOscar Salvador 2978369fa227SOscar Salvador return ret; 2979369fa227SOscar Salvador } 2980369fa227SOscar Salvador 2981ae37c7ffSOscar Salvador int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list) 2982369fa227SOscar Salvador { 2983369fa227SOscar Salvador struct hstate *h; 2984d5e33bd8SSidhartha Kumar struct folio *folio = page_folio(page); 2985ae37c7ffSOscar Salvador int ret = -EBUSY; 2986369fa227SOscar Salvador 2987369fa227SOscar Salvador /* 2988369fa227SOscar Salvador * The page might have been dissolved from under our feet, so make sure 2989369fa227SOscar Salvador * to carefully check the state under the lock. 2990369fa227SOscar Salvador * Return success when racing as if we dissolved the page ourselves. 2991369fa227SOscar Salvador */ 2992369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2993d5e33bd8SSidhartha Kumar if (folio_test_hugetlb(folio)) { 2994d5e33bd8SSidhartha Kumar h = folio_hstate(folio); 2995369fa227SOscar Salvador } else { 2996369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2997369fa227SOscar Salvador return 0; 2998369fa227SOscar Salvador } 2999369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 3000369fa227SOscar Salvador 3001369fa227SOscar Salvador /* 3002369fa227SOscar Salvador * Fence off gigantic pages as there is a cyclic dependency between 3003369fa227SOscar Salvador * alloc_contig_range and them. Return -ENOMEM as this has the effect 3004369fa227SOscar Salvador * of bailing out right away without further retrying. 3005369fa227SOscar Salvador */ 3006369fa227SOscar Salvador if (hstate_is_gigantic(h)) 3007369fa227SOscar Salvador return -ENOMEM; 3008369fa227SOscar Salvador 30096aa3a920SSidhartha Kumar if (folio_ref_count(folio) && !isolate_hugetlb(folio, list)) 3010ae37c7ffSOscar Salvador ret = 0; 3011d5e33bd8SSidhartha Kumar else if (!folio_ref_count(folio)) 301219fc1a7eSSidhartha Kumar ret = alloc_and_dissolve_hugetlb_folio(h, folio, list); 3013ae37c7ffSOscar Salvador 3014ae37c7ffSOscar Salvador return ret; 3015369fa227SOscar Salvador } 3016369fa227SOscar Salvador 301770c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 301804f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 3019348ea204SAdam Litke { 302090481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 3021a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3022d4ab0316SSidhartha Kumar struct folio *folio; 3023d85f69b0SMike Kravetz long map_chg, map_commit; 3024d85f69b0SMike Kravetz long gbl_chg; 30256d76dcf4SAneesh Kumar K.V int ret, idx; 30266d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 302708cf9fafSMina Almasry bool deferred_reserve; 30282fc39cecSAdam Litke 30296d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 3030a1e78772SMel Gorman /* 3031d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 3032d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 3033d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 3034a1e78772SMel Gorman */ 3035d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 3036d85f69b0SMike Kravetz if (map_chg < 0) 303776dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 3038d85f69b0SMike Kravetz 3039d85f69b0SMike Kravetz /* 3040d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 3041d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 3042d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 3043d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 3044d85f69b0SMike Kravetz * checked against any subpool limit. 3045d85f69b0SMike Kravetz */ 3046d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 3047d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 3048d85f69b0SMike Kravetz if (gbl_chg < 0) { 3049feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 305076dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 30515e911373SMike Kravetz } 305290d8b7e6SAdam Litke 3053d85f69b0SMike Kravetz /* 3054d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 3055d85f69b0SMike Kravetz * map, there could be reservations associated with the 3056d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 3057d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 3058d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 3059d85f69b0SMike Kravetz * the subpool reservations. 3060d85f69b0SMike Kravetz */ 3061d85f69b0SMike Kravetz if (avoid_reserve) 3062d85f69b0SMike Kravetz gbl_chg = 1; 3063d85f69b0SMike Kravetz } 3064d85f69b0SMike Kravetz 306508cf9fafSMina Almasry /* If this allocation is not consuming a reservation, charge it now. 306608cf9fafSMina Almasry */ 30676501fe5fSMiaohe Lin deferred_reserve = map_chg || avoid_reserve; 306808cf9fafSMina Almasry if (deferred_reserve) { 306908cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 307008cf9fafSMina Almasry idx, pages_per_huge_page(h), &h_cg); 30718f34af6fSJianyu Zhan if (ret) 30728f34af6fSJianyu Zhan goto out_subpool_put; 307308cf9fafSMina Almasry } 307408cf9fafSMina Almasry 307508cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 307608cf9fafSMina Almasry if (ret) 307708cf9fafSMina Almasry goto out_uncharge_cgroup_reservation; 30788f34af6fSJianyu Zhan 3079db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3080d85f69b0SMike Kravetz /* 3081d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 3082d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 3083d85f69b0SMike Kravetz * a reservation exists for the allocation. 3084d85f69b0SMike Kravetz */ 3085ff7d853bSSidhartha Kumar folio = dequeue_hugetlb_folio_vma(h, vma, addr, avoid_reserve, gbl_chg); 3086ff7d853bSSidhartha Kumar if (!folio) { 3087db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3088ff7d853bSSidhartha Kumar folio = alloc_buddy_hugetlb_folio_with_mpol(h, vma, addr); 3089ff7d853bSSidhartha Kumar if (!folio) 30908f34af6fSJianyu Zhan goto out_uncharge_cgroup; 309112df140fSRik van Riel spin_lock_irq(&hugetlb_lock); 3092a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 3093ff7d853bSSidhartha Kumar folio_set_hugetlb_restore_reserve(folio); 3094a88c7695SNaoya Horiguchi h->resv_huge_pages--; 3095a88c7695SNaoya Horiguchi } 3096ff7d853bSSidhartha Kumar list_add(&folio->lru, &h->hugepage_activelist); 3097ff7d853bSSidhartha Kumar folio_ref_unfreeze(folio, 1); 309881a6fcaeSJoonsoo Kim /* Fall through */ 3099a1e78772SMel Gorman } 3100ff7d853bSSidhartha Kumar 3101ff7d853bSSidhartha Kumar hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, folio); 310208cf9fafSMina Almasry /* If allocation is not consuming a reservation, also store the 310308cf9fafSMina Almasry * hugetlb_cgroup pointer on the page. 310408cf9fafSMina Almasry */ 310508cf9fafSMina Almasry if (deferred_reserve) { 310608cf9fafSMina Almasry hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), 3107ff7d853bSSidhartha Kumar h_cg, folio); 310808cf9fafSMina Almasry } 310908cf9fafSMina Almasry 3110db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3111a1e78772SMel Gorman 3112ff7d853bSSidhartha Kumar hugetlb_set_folio_subpool(folio, spool); 3113a1e78772SMel Gorman 3114d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 3115d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 311633039678SMike Kravetz /* 311733039678SMike Kravetz * The page was added to the reservation map between 311833039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 311933039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 312033039678SMike Kravetz * Adjust for the subpool count incremented above AND 312133039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 312233039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 312333039678SMike Kravetz * no longer applies. 312433039678SMike Kravetz */ 312533039678SMike Kravetz long rsv_adjust; 312633039678SMike Kravetz 312733039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 312833039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 312979aa925bSMike Kravetz if (deferred_reserve) 3130d4ab0316SSidhartha Kumar hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h), 3131d4ab0316SSidhartha Kumar pages_per_huge_page(h), folio); 313233039678SMike Kravetz } 3133ff7d853bSSidhartha Kumar return &folio->page; 31348f34af6fSJianyu Zhan 31358f34af6fSJianyu Zhan out_uncharge_cgroup: 31368f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 313708cf9fafSMina Almasry out_uncharge_cgroup_reservation: 313808cf9fafSMina Almasry if (deferred_reserve) 313908cf9fafSMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), 314008cf9fafSMina Almasry h_cg); 31418f34af6fSJianyu Zhan out_subpool_put: 3142d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 31438f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 3144feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 31458f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 3146b45b5bd6SDavid Gibson } 3147b45b5bd6SDavid Gibson 3148b5389086SZhenguo Yao int alloc_bootmem_huge_page(struct hstate *h, int nid) 3149e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 3150b5389086SZhenguo Yao int __alloc_bootmem_huge_page(struct hstate *h, int nid) 3151aa888a74SAndi Kleen { 3152b5389086SZhenguo Yao struct huge_bootmem_page *m = NULL; /* initialize for clang */ 3153b2261026SJoonsoo Kim int nr_nodes, node; 3154aa888a74SAndi Kleen 3155b5389086SZhenguo Yao /* do node specific alloc */ 3156b5389086SZhenguo Yao if (nid != NUMA_NO_NODE) { 3157b5389086SZhenguo Yao m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h), 3158b5389086SZhenguo Yao 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid); 3159b5389086SZhenguo Yao if (!m) 3160b5389086SZhenguo Yao return 0; 3161b5389086SZhenguo Yao goto found; 3162b5389086SZhenguo Yao } 3163b5389086SZhenguo Yao /* allocate from next node when distributing huge pages */ 3164b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 3165b5389086SZhenguo Yao m = memblock_alloc_try_nid_raw( 31668b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 316797ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 3168aa888a74SAndi Kleen /* 3169aa888a74SAndi Kleen * Use the beginning of the huge page to store the 3170aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 3171aa888a74SAndi Kleen * puts them into the mem_map). 3172aa888a74SAndi Kleen */ 3173b5389086SZhenguo Yao if (!m) 3174b5389086SZhenguo Yao return 0; 3175aa888a74SAndi Kleen goto found; 3176aa888a74SAndi Kleen } 3177aa888a74SAndi Kleen 3178aa888a74SAndi Kleen found: 3179aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 3180330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 3181aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 3182aa888a74SAndi Kleen m->hstate = h; 3183aa888a74SAndi Kleen return 1; 3184aa888a74SAndi Kleen } 3185aa888a74SAndi Kleen 318648b8d744SMike Kravetz /* 318748b8d744SMike Kravetz * Put bootmem huge pages into the standard lists after mem_map is up. 318848b8d744SMike Kravetz * Note: This only applies to gigantic (order > MAX_ORDER) pages. 318948b8d744SMike Kravetz */ 3190aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 3191aa888a74SAndi Kleen { 3192aa888a74SAndi Kleen struct huge_bootmem_page *m; 3193aa888a74SAndi Kleen 3194aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 319540d18ebfSMike Kravetz struct page *page = virt_to_page(m); 31967f325a8dSSidhartha Kumar struct folio *folio = page_folio(page); 3197aa888a74SAndi Kleen struct hstate *h = m->hstate; 3198ee8f248dSBecky Bruce 319948b8d744SMike Kravetz VM_BUG_ON(!hstate_is_gigantic(h)); 3200d1c60955SSidhartha Kumar WARN_ON(folio_ref_count(folio) != 1); 3201d1c60955SSidhartha Kumar if (prep_compound_gigantic_folio(folio, huge_page_order(h))) { 3202d1c60955SSidhartha Kumar WARN_ON(folio_test_reserved(folio)); 3203d1c60955SSidhartha Kumar prep_new_hugetlb_folio(h, folio, folio_nid(folio)); 32042b21624fSMike Kravetz free_huge_page(page); /* add to the hugepage allocator */ 32057118fc29SMike Kravetz } else { 3206416d85edSMike Kravetz /* VERY unlikely inflated ref count on a tail page */ 32077f325a8dSSidhartha Kumar free_gigantic_folio(folio, huge_page_order(h)); 32087118fc29SMike Kravetz } 3209af0fb9dfSMichal Hocko 3210b0320c7bSRafael Aquini /* 321148b8d744SMike Kravetz * We need to restore the 'stolen' pages to totalram_pages 321248b8d744SMike Kravetz * in order to fix confusing memory reports from free(1) and 321348b8d744SMike Kravetz * other side-effects, like CommitLimit going negative. 3214b0320c7bSRafael Aquini */ 3215c78a7f36SMiaohe Lin adjust_managed_page_count(page, pages_per_huge_page(h)); 3216520495feSCannon Matthews cond_resched(); 3217aa888a74SAndi Kleen } 3218aa888a74SAndi Kleen } 3219b5389086SZhenguo Yao static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid) 3220b5389086SZhenguo Yao { 3221b5389086SZhenguo Yao unsigned long i; 3222b5389086SZhenguo Yao char buf[32]; 3223b5389086SZhenguo Yao 3224b5389086SZhenguo Yao for (i = 0; i < h->max_huge_pages_node[nid]; ++i) { 3225b5389086SZhenguo Yao if (hstate_is_gigantic(h)) { 3226b5389086SZhenguo Yao if (!alloc_bootmem_huge_page(h, nid)) 3227b5389086SZhenguo Yao break; 3228b5389086SZhenguo Yao } else { 322919fc1a7eSSidhartha Kumar struct folio *folio; 3230b5389086SZhenguo Yao gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 3231b5389086SZhenguo Yao 323219fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, 3233b5389086SZhenguo Yao &node_states[N_MEMORY], NULL); 323419fc1a7eSSidhartha Kumar if (!folio) 3235b5389086SZhenguo Yao break; 323619fc1a7eSSidhartha Kumar free_huge_page(&folio->page); /* free it into the hugepage allocator */ 3237b5389086SZhenguo Yao } 3238b5389086SZhenguo Yao cond_resched(); 3239b5389086SZhenguo Yao } 3240b5389086SZhenguo Yao if (i == h->max_huge_pages_node[nid]) 3241b5389086SZhenguo Yao return; 3242b5389086SZhenguo Yao 3243b5389086SZhenguo Yao string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3244b5389086SZhenguo Yao pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n", 3245b5389086SZhenguo Yao h->max_huge_pages_node[nid], buf, nid, i); 3246b5389086SZhenguo Yao h->max_huge_pages -= (h->max_huge_pages_node[nid] - i); 3247b5389086SZhenguo Yao h->max_huge_pages_node[nid] = i; 3248b5389086SZhenguo Yao } 3249aa888a74SAndi Kleen 32508faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 32511da177e4SLinus Torvalds { 32521da177e4SLinus Torvalds unsigned long i; 3253f60858f9SMike Kravetz nodemask_t *node_alloc_noretry; 3254b5389086SZhenguo Yao bool node_specific_alloc = false; 3255f60858f9SMike Kravetz 3256b5389086SZhenguo Yao /* skip gigantic hugepages allocation if hugetlb_cma enabled */ 3257b5389086SZhenguo Yao if (hstate_is_gigantic(h) && hugetlb_cma_size) { 3258b5389086SZhenguo Yao pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); 3259b5389086SZhenguo Yao return; 3260b5389086SZhenguo Yao } 3261b5389086SZhenguo Yao 3262b5389086SZhenguo Yao /* do node specific alloc */ 32630a7a0f6fSPeng Liu for_each_online_node(i) { 3264b5389086SZhenguo Yao if (h->max_huge_pages_node[i] > 0) { 3265b5389086SZhenguo Yao hugetlb_hstate_alloc_pages_onenode(h, i); 3266b5389086SZhenguo Yao node_specific_alloc = true; 3267b5389086SZhenguo Yao } 3268b5389086SZhenguo Yao } 3269b5389086SZhenguo Yao 3270b5389086SZhenguo Yao if (node_specific_alloc) 3271b5389086SZhenguo Yao return; 3272b5389086SZhenguo Yao 3273b5389086SZhenguo Yao /* below will do all node balanced alloc */ 3274f60858f9SMike Kravetz if (!hstate_is_gigantic(h)) { 3275f60858f9SMike Kravetz /* 3276f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3277f60858f9SMike Kravetz * Ignore errors as lower level routines can deal with 3278f60858f9SMike Kravetz * node_alloc_noretry == NULL. If this kmalloc fails at boot 3279f60858f9SMike Kravetz * time, we are likely in bigger trouble. 3280f60858f9SMike Kravetz */ 3281f60858f9SMike Kravetz node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), 3282f60858f9SMike Kravetz GFP_KERNEL); 3283f60858f9SMike Kravetz } else { 3284f60858f9SMike Kravetz /* allocations done at boot time */ 3285f60858f9SMike Kravetz node_alloc_noretry = NULL; 3286f60858f9SMike Kravetz } 3287f60858f9SMike Kravetz 3288f60858f9SMike Kravetz /* bit mask controlling how hard we retry per-node allocations */ 3289f60858f9SMike Kravetz if (node_alloc_noretry) 3290f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 32911da177e4SLinus Torvalds 3292e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 3293bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 3294b5389086SZhenguo Yao if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE)) 3295aa888a74SAndi Kleen break; 32960c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 3297f60858f9SMike Kravetz &node_states[N_MEMORY], 3298f60858f9SMike Kravetz node_alloc_noretry)) 32991da177e4SLinus Torvalds break; 330069ed779aSDavid Rientjes cond_resched(); 33011da177e4SLinus Torvalds } 3302d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 3303d715cf80SLiam R. Howlett char buf[32]; 3304d715cf80SLiam R. Howlett 3305c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3306d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 3307d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 33088faa8b07SAndi Kleen h->max_huge_pages = i; 3309e5ff2159SAndi Kleen } 3310f60858f9SMike Kravetz kfree(node_alloc_noretry); 3311d715cf80SLiam R. Howlett } 3312e5ff2159SAndi Kleen 3313e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 3314e5ff2159SAndi Kleen { 331579dfc695SMike Kravetz struct hstate *h, *h2; 3316e5ff2159SAndi Kleen 3317e5ff2159SAndi Kleen for_each_hstate(h) { 33188faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 3319bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 33208faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 332179dfc695SMike Kravetz 332279dfc695SMike Kravetz /* 332379dfc695SMike Kravetz * Set demote order for each hstate. Note that 332479dfc695SMike Kravetz * h->demote_order is initially 0. 332579dfc695SMike Kravetz * - We can not demote gigantic pages if runtime freeing 332679dfc695SMike Kravetz * is not supported, so skip this. 3327a01f4390SMike Kravetz * - If CMA allocation is possible, we can not demote 3328a01f4390SMike Kravetz * HUGETLB_PAGE_ORDER or smaller size pages. 332979dfc695SMike Kravetz */ 333079dfc695SMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 333179dfc695SMike Kravetz continue; 3332a01f4390SMike Kravetz if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER) 3333a01f4390SMike Kravetz continue; 333479dfc695SMike Kravetz for_each_hstate(h2) { 333579dfc695SMike Kravetz if (h2 == h) 333679dfc695SMike Kravetz continue; 333779dfc695SMike Kravetz if (h2->order < h->order && 333879dfc695SMike Kravetz h2->order > h->demote_order) 333979dfc695SMike Kravetz h->demote_order = h2->order; 334079dfc695SMike Kravetz } 3341e5ff2159SAndi Kleen } 3342e5ff2159SAndi Kleen } 3343e5ff2159SAndi Kleen 3344e5ff2159SAndi Kleen static void __init report_hugepages(void) 3345e5ff2159SAndi Kleen { 3346e5ff2159SAndi Kleen struct hstate *h; 3347e5ff2159SAndi Kleen 3348e5ff2159SAndi Kleen for_each_hstate(h) { 33494abd32dbSAndi Kleen char buf[32]; 3350c6247f72SMatthew Wilcox 3351c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 33526213834cSMuchun Song pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n", 3353c6247f72SMatthew Wilcox buf, h->free_huge_pages); 33546213834cSMuchun Song pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n", 33556213834cSMuchun Song hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf); 3356e5ff2159SAndi Kleen } 3357e5ff2159SAndi Kleen } 3358e5ff2159SAndi Kleen 33591da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 33606ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 33616ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 33621da177e4SLinus Torvalds { 33634415cc8dSChristoph Lameter int i; 33641121828aSMike Kravetz LIST_HEAD(page_list); 33654415cc8dSChristoph Lameter 33669487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 3367bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3368aa888a74SAndi Kleen return; 3369aa888a74SAndi Kleen 33701121828aSMike Kravetz /* 33711121828aSMike Kravetz * Collect pages to be freed on a list, and free after dropping lock 33721121828aSMike Kravetz */ 33736ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 337410c6ec49SMike Kravetz struct page *page, *next; 3375a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 3376a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 3377a5516438SAndi Kleen if (count >= h->nr_huge_pages) 33781121828aSMike Kravetz goto out; 33791da177e4SLinus Torvalds if (PageHighMem(page)) 33801da177e4SLinus Torvalds continue; 3381cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, page_folio(page), false); 33821121828aSMike Kravetz list_add(&page->lru, &page_list); 33831121828aSMike Kravetz } 33841121828aSMike Kravetz } 33851121828aSMike Kravetz 33861121828aSMike Kravetz out: 3387db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 338810c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3389db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 33901da177e4SLinus Torvalds } 33911da177e4SLinus Torvalds #else 33926ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 33936ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 33941da177e4SLinus Torvalds { 33951da177e4SLinus Torvalds } 33961da177e4SLinus Torvalds #endif 33971da177e4SLinus Torvalds 339820a0307cSWu Fengguang /* 339920a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 340020a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 340120a0307cSWu Fengguang * Returns 1 if an adjustment was made. 340220a0307cSWu Fengguang */ 34036ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 34046ae11b27SLee Schermerhorn int delta) 340520a0307cSWu Fengguang { 3406b2261026SJoonsoo Kim int nr_nodes, node; 340720a0307cSWu Fengguang 34089487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 340920a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 341020a0307cSWu Fengguang 3411e8c5c824SLee Schermerhorn if (delta < 0) { 3412b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 3413b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 3414b2261026SJoonsoo Kim goto found; 3415b2261026SJoonsoo Kim } 3416b2261026SJoonsoo Kim } else { 3417b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 3418b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 3419b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 3420b2261026SJoonsoo Kim goto found; 3421e8c5c824SLee Schermerhorn } 34229a76db09SLee Schermerhorn } 3423b2261026SJoonsoo Kim return 0; 342420a0307cSWu Fengguang 3425b2261026SJoonsoo Kim found: 342620a0307cSWu Fengguang h->surplus_huge_pages += delta; 3427b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 3428b2261026SJoonsoo Kim return 1; 342920a0307cSWu Fengguang } 343020a0307cSWu Fengguang 3431a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 3432fd875dcaSMike Kravetz static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 34336ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 34341da177e4SLinus Torvalds { 34357893d1d5SAdam Litke unsigned long min_count, ret; 343610c6ec49SMike Kravetz struct page *page; 343710c6ec49SMike Kravetz LIST_HEAD(page_list); 3438f60858f9SMike Kravetz NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); 3439f60858f9SMike Kravetz 3440f60858f9SMike Kravetz /* 3441f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3442f60858f9SMike Kravetz * If we can not allocate the bit mask, do not attempt to allocate 3443f60858f9SMike Kravetz * the requested huge pages. 3444f60858f9SMike Kravetz */ 3445f60858f9SMike Kravetz if (node_alloc_noretry) 3446f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 3447f60858f9SMike Kravetz else 3448f60858f9SMike Kravetz return -ENOMEM; 34491da177e4SLinus Torvalds 345029383967SMike Kravetz /* 345129383967SMike Kravetz * resize_lock mutex prevents concurrent adjustments to number of 345229383967SMike Kravetz * pages in hstate via the proc/sysfs interfaces. 345329383967SMike Kravetz */ 345429383967SMike Kravetz mutex_lock(&h->resize_lock); 3455b65d4adbSMuchun Song flush_free_hpage_work(h); 3456db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 34574eb0716eSAlexandre Ghiti 34584eb0716eSAlexandre Ghiti /* 3459fd875dcaSMike Kravetz * Check for a node specific request. 3460fd875dcaSMike Kravetz * Changing node specific huge page count may require a corresponding 3461fd875dcaSMike Kravetz * change to the global count. In any case, the passed node mask 3462fd875dcaSMike Kravetz * (nodes_allowed) will restrict alloc/free to the specified node. 3463fd875dcaSMike Kravetz */ 3464fd875dcaSMike Kravetz if (nid != NUMA_NO_NODE) { 3465fd875dcaSMike Kravetz unsigned long old_count = count; 3466fd875dcaSMike Kravetz 3467fd875dcaSMike Kravetz count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 3468fd875dcaSMike Kravetz /* 3469fd875dcaSMike Kravetz * User may have specified a large count value which caused the 3470fd875dcaSMike Kravetz * above calculation to overflow. In this case, they wanted 3471fd875dcaSMike Kravetz * to allocate as many huge pages as possible. Set count to 3472fd875dcaSMike Kravetz * largest possible value to align with their intention. 3473fd875dcaSMike Kravetz */ 3474fd875dcaSMike Kravetz if (count < old_count) 3475fd875dcaSMike Kravetz count = ULONG_MAX; 3476fd875dcaSMike Kravetz } 3477fd875dcaSMike Kravetz 3478fd875dcaSMike Kravetz /* 34794eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 34804eb0716eSAlexandre Ghiti * page range allocation. 34814eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 34824eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 34834eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 34844eb0716eSAlexandre Ghiti */ 34854eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 34864eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 3487db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 348829383967SMike Kravetz mutex_unlock(&h->resize_lock); 3489f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 34904eb0716eSAlexandre Ghiti return -EINVAL; 34914eb0716eSAlexandre Ghiti } 34924eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 34934eb0716eSAlexandre Ghiti } 3494aa888a74SAndi Kleen 34957893d1d5SAdam Litke /* 34967893d1d5SAdam Litke * Increase the pool size 34977893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 34987893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 3499d1c3fb1fSNishanth Aravamudan * 35003a740e8bSSidhartha Kumar * We might race with alloc_surplus_hugetlb_folio() here and be unable 3501d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 3502d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 3503d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 3504d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 35057893d1d5SAdam Litke */ 3506a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 35076ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 35087893d1d5SAdam Litke break; 35097893d1d5SAdam Litke } 35107893d1d5SAdam Litke 3511a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 35127893d1d5SAdam Litke /* 35137893d1d5SAdam Litke * If this allocation races such that we no longer need the 35147893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 35157893d1d5SAdam Litke * and reducing the surplus. 35167893d1d5SAdam Litke */ 3517db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3518649920c6SJia He 3519649920c6SJia He /* yield cpu to avoid soft lockup */ 3520649920c6SJia He cond_resched(); 3521649920c6SJia He 3522f60858f9SMike Kravetz ret = alloc_pool_huge_page(h, nodes_allowed, 3523f60858f9SMike Kravetz node_alloc_noretry); 3524db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 35257893d1d5SAdam Litke if (!ret) 35267893d1d5SAdam Litke goto out; 35277893d1d5SAdam Litke 3528536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 3529536240f2SMel Gorman if (signal_pending(current)) 3530536240f2SMel Gorman goto out; 35317893d1d5SAdam Litke } 35327893d1d5SAdam Litke 35337893d1d5SAdam Litke /* 35347893d1d5SAdam Litke * Decrease the pool size 35357893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 35367893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 35377893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 35387893d1d5SAdam Litke * to the desired size as pages become free. 3539d1c3fb1fSNishanth Aravamudan * 3540d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 3541d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 3542d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 35433a740e8bSSidhartha Kumar * alloc_surplus_hugetlb_folio() is checking the global counter, 3544d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 3545d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 3546d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 35477893d1d5SAdam Litke */ 3548a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 35496b0c880dSAdam Litke min_count = max(count, min_count); 35506ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 355110c6ec49SMike Kravetz 355210c6ec49SMike Kravetz /* 355310c6ec49SMike Kravetz * Collect pages to be removed on list without dropping lock 355410c6ec49SMike Kravetz */ 3555a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 355610c6ec49SMike Kravetz page = remove_pool_huge_page(h, nodes_allowed, 0); 355710c6ec49SMike Kravetz if (!page) 35581da177e4SLinus Torvalds break; 355910c6ec49SMike Kravetz 356010c6ec49SMike Kravetz list_add(&page->lru, &page_list); 35611da177e4SLinus Torvalds } 356210c6ec49SMike Kravetz /* free the pages after dropping lock */ 3563db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 356410c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3565b65d4adbSMuchun Song flush_free_hpage_work(h); 3566db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 356710c6ec49SMike Kravetz 3568a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 35696ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 35707893d1d5SAdam Litke break; 35717893d1d5SAdam Litke } 35727893d1d5SAdam Litke out: 35734eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 3574db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 357529383967SMike Kravetz mutex_unlock(&h->resize_lock); 35764eb0716eSAlexandre Ghiti 3577f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 3578f60858f9SMike Kravetz 35794eb0716eSAlexandre Ghiti return 0; 35801da177e4SLinus Torvalds } 35811da177e4SLinus Torvalds 3582bdd7be07SSidhartha Kumar static int demote_free_hugetlb_folio(struct hstate *h, struct folio *folio) 35838531fc6fSMike Kravetz { 3584bdd7be07SSidhartha Kumar int i, nid = folio_nid(folio); 35858531fc6fSMike Kravetz struct hstate *target_hstate; 358631731452SDoug Berger struct page *subpage; 3587bdd7be07SSidhartha Kumar struct folio *inner_folio; 35888531fc6fSMike Kravetz int rc = 0; 35898531fc6fSMike Kravetz 35908531fc6fSMike Kravetz target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order); 35918531fc6fSMike Kravetz 3592cfd5082bSSidhartha Kumar remove_hugetlb_folio_for_demote(h, folio, false); 35938531fc6fSMike Kravetz spin_unlock_irq(&hugetlb_lock); 35948531fc6fSMike Kravetz 3595bdd7be07SSidhartha Kumar rc = hugetlb_vmemmap_restore(h, &folio->page); 35968531fc6fSMike Kravetz if (rc) { 3597bdd7be07SSidhartha Kumar /* Allocation of vmemmmap failed, we can not demote folio */ 35988531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 3599bdd7be07SSidhartha Kumar folio_ref_unfreeze(folio, 1); 3600bdd7be07SSidhartha Kumar add_hugetlb_folio(h, folio, false); 36018531fc6fSMike Kravetz return rc; 36028531fc6fSMike Kravetz } 36038531fc6fSMike Kravetz 36048531fc6fSMike Kravetz /* 3605911565b8SSidhartha Kumar * Use destroy_compound_hugetlb_folio_for_demote for all huge page 3606bdd7be07SSidhartha Kumar * sizes as it will not ref count folios. 36078531fc6fSMike Kravetz */ 3608911565b8SSidhartha Kumar destroy_compound_hugetlb_folio_for_demote(folio, huge_page_order(h)); 36098531fc6fSMike Kravetz 36108531fc6fSMike Kravetz /* 36118531fc6fSMike Kravetz * Taking target hstate mutex synchronizes with set_max_huge_pages. 36128531fc6fSMike Kravetz * Without the mutex, pages added to target hstate could be marked 36138531fc6fSMike Kravetz * as surplus. 36148531fc6fSMike Kravetz * 36158531fc6fSMike Kravetz * Note that we already hold h->resize_lock. To prevent deadlock, 36168531fc6fSMike Kravetz * use the convention of always taking larger size hstate mutex first. 36178531fc6fSMike Kravetz */ 36188531fc6fSMike Kravetz mutex_lock(&target_hstate->resize_lock); 36198531fc6fSMike Kravetz for (i = 0; i < pages_per_huge_page(h); 36208531fc6fSMike Kravetz i += pages_per_huge_page(target_hstate)) { 3621bdd7be07SSidhartha Kumar subpage = folio_page(folio, i); 3622bdd7be07SSidhartha Kumar inner_folio = page_folio(subpage); 36238531fc6fSMike Kravetz if (hstate_is_gigantic(target_hstate)) 3624bdd7be07SSidhartha Kumar prep_compound_gigantic_folio_for_demote(inner_folio, 36258531fc6fSMike Kravetz target_hstate->order); 36268531fc6fSMike Kravetz else 362731731452SDoug Berger prep_compound_page(subpage, target_hstate->order); 3628bdd7be07SSidhartha Kumar folio_change_private(inner_folio, NULL); 3629bdd7be07SSidhartha Kumar prep_new_hugetlb_folio(target_hstate, inner_folio, nid); 36302b21624fSMike Kravetz free_huge_page(subpage); 36318531fc6fSMike Kravetz } 36328531fc6fSMike Kravetz mutex_unlock(&target_hstate->resize_lock); 36338531fc6fSMike Kravetz 36348531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 36358531fc6fSMike Kravetz 36368531fc6fSMike Kravetz /* 36378531fc6fSMike Kravetz * Not absolutely necessary, but for consistency update max_huge_pages 36388531fc6fSMike Kravetz * based on pool changes for the demoted page. 36398531fc6fSMike Kravetz */ 36408531fc6fSMike Kravetz h->max_huge_pages--; 3641a43a83c7SMiaohe Lin target_hstate->max_huge_pages += 3642a43a83c7SMiaohe Lin pages_per_huge_page(h) / pages_per_huge_page(target_hstate); 36438531fc6fSMike Kravetz 36448531fc6fSMike Kravetz return rc; 36458531fc6fSMike Kravetz } 36468531fc6fSMike Kravetz 364779dfc695SMike Kravetz static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 364879dfc695SMike Kravetz __must_hold(&hugetlb_lock) 364979dfc695SMike Kravetz { 36508531fc6fSMike Kravetz int nr_nodes, node; 3651bdd7be07SSidhartha Kumar struct folio *folio; 365279dfc695SMike Kravetz 365379dfc695SMike Kravetz lockdep_assert_held(&hugetlb_lock); 365479dfc695SMike Kravetz 365579dfc695SMike Kravetz /* We should never get here if no demote order */ 365679dfc695SMike Kravetz if (!h->demote_order) { 365779dfc695SMike Kravetz pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n"); 365879dfc695SMike Kravetz return -EINVAL; /* internal error */ 365979dfc695SMike Kravetz } 366079dfc695SMike Kravetz 36618531fc6fSMike Kravetz for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 3662bdd7be07SSidhartha Kumar list_for_each_entry(folio, &h->hugepage_freelists[node], lru) { 3663bdd7be07SSidhartha Kumar if (folio_test_hwpoison(folio)) 36645a317412SMike Kravetz continue; 3665bdd7be07SSidhartha Kumar return demote_free_hugetlb_folio(h, folio); 36668531fc6fSMike Kravetz } 36678531fc6fSMike Kravetz } 36688531fc6fSMike Kravetz 36695a317412SMike Kravetz /* 36705a317412SMike Kravetz * Only way to get here is if all pages on free lists are poisoned. 36715a317412SMike Kravetz * Return -EBUSY so that caller will not retry. 36725a317412SMike Kravetz */ 36735a317412SMike Kravetz return -EBUSY; 367479dfc695SMike Kravetz } 367579dfc695SMike Kravetz 3676a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 3677a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 3678a3437870SNishanth Aravamudan 367979dfc695SMike Kravetz #define HSTATE_ATTR_WO(_name) \ 368079dfc695SMike Kravetz static struct kobj_attribute _name##_attr = __ATTR_WO(_name) 368179dfc695SMike Kravetz 3682a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 368398bc26acSMiaohe Lin static struct kobj_attribute _name##_attr = __ATTR_RW(_name) 3684a3437870SNishanth Aravamudan 3685a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 3686a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 3687a3437870SNishanth Aravamudan 36889a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 36899a305230SLee Schermerhorn 36909a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 3691a3437870SNishanth Aravamudan { 3692a3437870SNishanth Aravamudan int i; 36939a305230SLee Schermerhorn 3694a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 36959a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 36969a305230SLee Schermerhorn if (nidp) 36979a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 3698a3437870SNishanth Aravamudan return &hstates[i]; 36999a305230SLee Schermerhorn } 37009a305230SLee Schermerhorn 37019a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 3702a3437870SNishanth Aravamudan } 3703a3437870SNishanth Aravamudan 370406808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 3705a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3706a3437870SNishanth Aravamudan { 37079a305230SLee Schermerhorn struct hstate *h; 37089a305230SLee Schermerhorn unsigned long nr_huge_pages; 37099a305230SLee Schermerhorn int nid; 37109a305230SLee Schermerhorn 37119a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 37129a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 37139a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 37149a305230SLee Schermerhorn else 37159a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 37169a305230SLee Schermerhorn 3717ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", nr_huge_pages); 3718a3437870SNishanth Aravamudan } 3719adbe8726SEric B Munson 3720238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 3721238d3c13SDavid Rientjes struct hstate *h, int nid, 3722238d3c13SDavid Rientjes unsigned long count, size_t len) 3723a3437870SNishanth Aravamudan { 3724a3437870SNishanth Aravamudan int err; 37252d0adf7eSOscar Salvador nodemask_t nodes_allowed, *n_mask; 3726a3437870SNishanth Aravamudan 37272d0adf7eSOscar Salvador if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 37282d0adf7eSOscar Salvador return -EINVAL; 3729adbe8726SEric B Munson 37309a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 37319a305230SLee Schermerhorn /* 37329a305230SLee Schermerhorn * global hstate attribute 37339a305230SLee Schermerhorn */ 37349a305230SLee Schermerhorn if (!(obey_mempolicy && 37352d0adf7eSOscar Salvador init_nodemask_of_mempolicy(&nodes_allowed))) 37362d0adf7eSOscar Salvador n_mask = &node_states[N_MEMORY]; 37372d0adf7eSOscar Salvador else 37382d0adf7eSOscar Salvador n_mask = &nodes_allowed; 37392d0adf7eSOscar Salvador } else { 37409a305230SLee Schermerhorn /* 3741fd875dcaSMike Kravetz * Node specific request. count adjustment happens in 3742fd875dcaSMike Kravetz * set_max_huge_pages() after acquiring hugetlb_lock. 37439a305230SLee Schermerhorn */ 37442d0adf7eSOscar Salvador init_nodemask_of_node(&nodes_allowed, nid); 37452d0adf7eSOscar Salvador n_mask = &nodes_allowed; 3746fd875dcaSMike Kravetz } 37479a305230SLee Schermerhorn 37482d0adf7eSOscar Salvador err = set_max_huge_pages(h, count, nid, n_mask); 374906808b08SLee Schermerhorn 37504eb0716eSAlexandre Ghiti return err ? err : len; 375106808b08SLee Schermerhorn } 375206808b08SLee Schermerhorn 3753238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 3754238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 3755238d3c13SDavid Rientjes size_t len) 3756238d3c13SDavid Rientjes { 3757238d3c13SDavid Rientjes struct hstate *h; 3758238d3c13SDavid Rientjes unsigned long count; 3759238d3c13SDavid Rientjes int nid; 3760238d3c13SDavid Rientjes int err; 3761238d3c13SDavid Rientjes 3762238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 3763238d3c13SDavid Rientjes if (err) 3764238d3c13SDavid Rientjes return err; 3765238d3c13SDavid Rientjes 3766238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 3767238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 3768238d3c13SDavid Rientjes } 3769238d3c13SDavid Rientjes 377006808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 377106808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 377206808b08SLee Schermerhorn { 377306808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 377406808b08SLee Schermerhorn } 377506808b08SLee Schermerhorn 377606808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 377706808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 377806808b08SLee Schermerhorn { 3779238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 3780a3437870SNishanth Aravamudan } 3781a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 3782a3437870SNishanth Aravamudan 378306808b08SLee Schermerhorn #ifdef CONFIG_NUMA 378406808b08SLee Schermerhorn 378506808b08SLee Schermerhorn /* 378606808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 378706808b08SLee Schermerhorn * huge page alloc/free. 378806808b08SLee Schermerhorn */ 378906808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 3790ae7a927dSJoe Perches struct kobj_attribute *attr, 3791ae7a927dSJoe Perches char *buf) 379206808b08SLee Schermerhorn { 379306808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 379406808b08SLee Schermerhorn } 379506808b08SLee Schermerhorn 379606808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 379706808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 379806808b08SLee Schermerhorn { 3799238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 380006808b08SLee Schermerhorn } 380106808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 380206808b08SLee Schermerhorn #endif 380306808b08SLee Schermerhorn 380406808b08SLee Schermerhorn 3805a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 3806a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3807a3437870SNishanth Aravamudan { 38089a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3809ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); 3810a3437870SNishanth Aravamudan } 3811adbe8726SEric B Munson 3812a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 3813a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 3814a3437870SNishanth Aravamudan { 3815a3437870SNishanth Aravamudan int err; 3816a3437870SNishanth Aravamudan unsigned long input; 38179a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3818a3437870SNishanth Aravamudan 3819bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3820adbe8726SEric B Munson return -EINVAL; 3821adbe8726SEric B Munson 38223dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 3823a3437870SNishanth Aravamudan if (err) 382473ae31e5SEric B Munson return err; 3825a3437870SNishanth Aravamudan 3826db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3827a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 3828db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3829a3437870SNishanth Aravamudan 3830a3437870SNishanth Aravamudan return count; 3831a3437870SNishanth Aravamudan } 3832a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 3833a3437870SNishanth Aravamudan 3834a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 3835a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3836a3437870SNishanth Aravamudan { 38379a305230SLee Schermerhorn struct hstate *h; 38389a305230SLee Schermerhorn unsigned long free_huge_pages; 38399a305230SLee Schermerhorn int nid; 38409a305230SLee Schermerhorn 38419a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 38429a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 38439a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 38449a305230SLee Schermerhorn else 38459a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 38469a305230SLee Schermerhorn 3847ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", free_huge_pages); 3848a3437870SNishanth Aravamudan } 3849a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 3850a3437870SNishanth Aravamudan 3851a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 3852a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3853a3437870SNishanth Aravamudan { 38549a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3855ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); 3856a3437870SNishanth Aravamudan } 3857a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 3858a3437870SNishanth Aravamudan 3859a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 3860a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3861a3437870SNishanth Aravamudan { 38629a305230SLee Schermerhorn struct hstate *h; 38639a305230SLee Schermerhorn unsigned long surplus_huge_pages; 38649a305230SLee Schermerhorn int nid; 38659a305230SLee Schermerhorn 38669a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 38679a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 38689a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 38699a305230SLee Schermerhorn else 38709a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 38719a305230SLee Schermerhorn 3872ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", surplus_huge_pages); 3873a3437870SNishanth Aravamudan } 3874a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 3875a3437870SNishanth Aravamudan 387679dfc695SMike Kravetz static ssize_t demote_store(struct kobject *kobj, 387779dfc695SMike Kravetz struct kobj_attribute *attr, const char *buf, size_t len) 387879dfc695SMike Kravetz { 387979dfc695SMike Kravetz unsigned long nr_demote; 388079dfc695SMike Kravetz unsigned long nr_available; 388179dfc695SMike Kravetz nodemask_t nodes_allowed, *n_mask; 388279dfc695SMike Kravetz struct hstate *h; 38838eeda55fSLi zeming int err; 388479dfc695SMike Kravetz int nid; 388579dfc695SMike Kravetz 388679dfc695SMike Kravetz err = kstrtoul(buf, 10, &nr_demote); 388779dfc695SMike Kravetz if (err) 388879dfc695SMike Kravetz return err; 388979dfc695SMike Kravetz h = kobj_to_hstate(kobj, &nid); 389079dfc695SMike Kravetz 389179dfc695SMike Kravetz if (nid != NUMA_NO_NODE) { 389279dfc695SMike Kravetz init_nodemask_of_node(&nodes_allowed, nid); 389379dfc695SMike Kravetz n_mask = &nodes_allowed; 389479dfc695SMike Kravetz } else { 389579dfc695SMike Kravetz n_mask = &node_states[N_MEMORY]; 389679dfc695SMike Kravetz } 389779dfc695SMike Kravetz 389879dfc695SMike Kravetz /* Synchronize with other sysfs operations modifying huge pages */ 389979dfc695SMike Kravetz mutex_lock(&h->resize_lock); 390079dfc695SMike Kravetz spin_lock_irq(&hugetlb_lock); 390179dfc695SMike Kravetz 390279dfc695SMike Kravetz while (nr_demote) { 390379dfc695SMike Kravetz /* 390479dfc695SMike Kravetz * Check for available pages to demote each time thorough the 390579dfc695SMike Kravetz * loop as demote_pool_huge_page will drop hugetlb_lock. 390679dfc695SMike Kravetz */ 390779dfc695SMike Kravetz if (nid != NUMA_NO_NODE) 390879dfc695SMike Kravetz nr_available = h->free_huge_pages_node[nid]; 390979dfc695SMike Kravetz else 391079dfc695SMike Kravetz nr_available = h->free_huge_pages; 391179dfc695SMike Kravetz nr_available -= h->resv_huge_pages; 391279dfc695SMike Kravetz if (!nr_available) 391379dfc695SMike Kravetz break; 391479dfc695SMike Kravetz 391579dfc695SMike Kravetz err = demote_pool_huge_page(h, n_mask); 391679dfc695SMike Kravetz if (err) 391779dfc695SMike Kravetz break; 391879dfc695SMike Kravetz 391979dfc695SMike Kravetz nr_demote--; 392079dfc695SMike Kravetz } 392179dfc695SMike Kravetz 392279dfc695SMike Kravetz spin_unlock_irq(&hugetlb_lock); 392379dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 392479dfc695SMike Kravetz 392579dfc695SMike Kravetz if (err) 392679dfc695SMike Kravetz return err; 392779dfc695SMike Kravetz return len; 392879dfc695SMike Kravetz } 392979dfc695SMike Kravetz HSTATE_ATTR_WO(demote); 393079dfc695SMike Kravetz 393179dfc695SMike Kravetz static ssize_t demote_size_show(struct kobject *kobj, 393279dfc695SMike Kravetz struct kobj_attribute *attr, char *buf) 393379dfc695SMike Kravetz { 393412658abfSMiaohe Lin struct hstate *h = kobj_to_hstate(kobj, NULL); 393579dfc695SMike Kravetz unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K; 393679dfc695SMike Kravetz 393779dfc695SMike Kravetz return sysfs_emit(buf, "%lukB\n", demote_size); 393879dfc695SMike Kravetz } 393979dfc695SMike Kravetz 394079dfc695SMike Kravetz static ssize_t demote_size_store(struct kobject *kobj, 394179dfc695SMike Kravetz struct kobj_attribute *attr, 394279dfc695SMike Kravetz const char *buf, size_t count) 394379dfc695SMike Kravetz { 394479dfc695SMike Kravetz struct hstate *h, *demote_hstate; 394579dfc695SMike Kravetz unsigned long demote_size; 394679dfc695SMike Kravetz unsigned int demote_order; 394779dfc695SMike Kravetz 394879dfc695SMike Kravetz demote_size = (unsigned long)memparse(buf, NULL); 394979dfc695SMike Kravetz 395079dfc695SMike Kravetz demote_hstate = size_to_hstate(demote_size); 395179dfc695SMike Kravetz if (!demote_hstate) 395279dfc695SMike Kravetz return -EINVAL; 395379dfc695SMike Kravetz demote_order = demote_hstate->order; 3954a01f4390SMike Kravetz if (demote_order < HUGETLB_PAGE_ORDER) 3955a01f4390SMike Kravetz return -EINVAL; 395679dfc695SMike Kravetz 395779dfc695SMike Kravetz /* demote order must be smaller than hstate order */ 395812658abfSMiaohe Lin h = kobj_to_hstate(kobj, NULL); 395979dfc695SMike Kravetz if (demote_order >= h->order) 396079dfc695SMike Kravetz return -EINVAL; 396179dfc695SMike Kravetz 396279dfc695SMike Kravetz /* resize_lock synchronizes access to demote size and writes */ 396379dfc695SMike Kravetz mutex_lock(&h->resize_lock); 396479dfc695SMike Kravetz h->demote_order = demote_order; 396579dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 396679dfc695SMike Kravetz 396779dfc695SMike Kravetz return count; 396879dfc695SMike Kravetz } 396979dfc695SMike Kravetz HSTATE_ATTR(demote_size); 397079dfc695SMike Kravetz 3971a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 3972a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 3973a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 3974a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 3975a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 3976a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 397706808b08SLee Schermerhorn #ifdef CONFIG_NUMA 397806808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 397906808b08SLee Schermerhorn #endif 3980a3437870SNishanth Aravamudan NULL, 3981a3437870SNishanth Aravamudan }; 3982a3437870SNishanth Aravamudan 398367e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 3984a3437870SNishanth Aravamudan .attrs = hstate_attrs, 3985a3437870SNishanth Aravamudan }; 3986a3437870SNishanth Aravamudan 398779dfc695SMike Kravetz static struct attribute *hstate_demote_attrs[] = { 398879dfc695SMike Kravetz &demote_size_attr.attr, 398979dfc695SMike Kravetz &demote_attr.attr, 399079dfc695SMike Kravetz NULL, 399179dfc695SMike Kravetz }; 399279dfc695SMike Kravetz 399379dfc695SMike Kravetz static const struct attribute_group hstate_demote_attr_group = { 399479dfc695SMike Kravetz .attrs = hstate_demote_attrs, 399579dfc695SMike Kravetz }; 399679dfc695SMike Kravetz 3997094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 39989a305230SLee Schermerhorn struct kobject **hstate_kobjs, 399967e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 4000a3437870SNishanth Aravamudan { 4001a3437870SNishanth Aravamudan int retval; 4002972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 4003a3437870SNishanth Aravamudan 40049a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 40059a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 4006a3437870SNishanth Aravamudan return -ENOMEM; 4007a3437870SNishanth Aravamudan 40089a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 4009cc2205a6SMiaohe Lin if (retval) { 40109a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 4011cc2205a6SMiaohe Lin hstate_kobjs[hi] = NULL; 40123a6bdda0SMiaohe Lin return retval; 4013cc2205a6SMiaohe Lin } 4014a3437870SNishanth Aravamudan 401579dfc695SMike Kravetz if (h->demote_order) { 401601088a60SMiaohe Lin retval = sysfs_create_group(hstate_kobjs[hi], 401701088a60SMiaohe Lin &hstate_demote_attr_group); 401801088a60SMiaohe Lin if (retval) { 401979dfc695SMike Kravetz pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name); 402001088a60SMiaohe Lin sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group); 402101088a60SMiaohe Lin kobject_put(hstate_kobjs[hi]); 402201088a60SMiaohe Lin hstate_kobjs[hi] = NULL; 402301088a60SMiaohe Lin return retval; 402401088a60SMiaohe Lin } 402579dfc695SMike Kravetz } 402679dfc695SMike Kravetz 402701088a60SMiaohe Lin return 0; 4028a3437870SNishanth Aravamudan } 4029a3437870SNishanth Aravamudan 40309a305230SLee Schermerhorn #ifdef CONFIG_NUMA 4031a4a00b45SMuchun Song static bool hugetlb_sysfs_initialized __ro_after_init; 40329a305230SLee Schermerhorn 40339a305230SLee Schermerhorn /* 40349a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 403510fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 403610fbcf4cSKay Sievers * index of a node device or _hstate == node id. 403710fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 40389a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 40399a305230SLee Schermerhorn */ 40409a305230SLee Schermerhorn struct node_hstate { 40419a305230SLee Schermerhorn struct kobject *hugepages_kobj; 40429a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 40439a305230SLee Schermerhorn }; 4044b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 40459a305230SLee Schermerhorn 40469a305230SLee Schermerhorn /* 404710fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 40489a305230SLee Schermerhorn */ 40499a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 40509a305230SLee Schermerhorn &nr_hugepages_attr.attr, 40519a305230SLee Schermerhorn &free_hugepages_attr.attr, 40529a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 40539a305230SLee Schermerhorn NULL, 40549a305230SLee Schermerhorn }; 40559a305230SLee Schermerhorn 405667e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 40579a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 40589a305230SLee Schermerhorn }; 40599a305230SLee Schermerhorn 40609a305230SLee Schermerhorn /* 406110fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 40629a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 40639a305230SLee Schermerhorn */ 40649a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 40659a305230SLee Schermerhorn { 40669a305230SLee Schermerhorn int nid; 40679a305230SLee Schermerhorn 40689a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 40699a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 40709a305230SLee Schermerhorn int i; 40719a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 40729a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 40739a305230SLee Schermerhorn if (nidp) 40749a305230SLee Schermerhorn *nidp = nid; 40759a305230SLee Schermerhorn return &hstates[i]; 40769a305230SLee Schermerhorn } 40779a305230SLee Schermerhorn } 40789a305230SLee Schermerhorn 40799a305230SLee Schermerhorn BUG(); 40809a305230SLee Schermerhorn return NULL; 40819a305230SLee Schermerhorn } 40829a305230SLee Schermerhorn 40839a305230SLee Schermerhorn /* 408410fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 40859a305230SLee Schermerhorn * No-op if no hstate attributes attached. 40869a305230SLee Schermerhorn */ 4087a4a00b45SMuchun Song void hugetlb_unregister_node(struct node *node) 40889a305230SLee Schermerhorn { 40899a305230SLee Schermerhorn struct hstate *h; 409010fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 40919a305230SLee Schermerhorn 40929a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 40939b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 40949a305230SLee Schermerhorn 4095972dc4deSAneesh Kumar K.V for_each_hstate(h) { 4096972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 409701088a60SMiaohe Lin struct kobject *hstate_kobj = nhs->hstate_kobjs[idx]; 409801088a60SMiaohe Lin 409901088a60SMiaohe Lin if (!hstate_kobj) 410001088a60SMiaohe Lin continue; 410101088a60SMiaohe Lin if (h->demote_order) 410201088a60SMiaohe Lin sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group); 410301088a60SMiaohe Lin sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group); 410401088a60SMiaohe Lin kobject_put(hstate_kobj); 4105972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 4106972dc4deSAneesh Kumar K.V } 41079a305230SLee Schermerhorn 41089a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 41099a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 41109a305230SLee Schermerhorn } 41119a305230SLee Schermerhorn 41129a305230SLee Schermerhorn 41139a305230SLee Schermerhorn /* 411410fbcf4cSKay Sievers * Register hstate attributes for a single node device. 41159a305230SLee Schermerhorn * No-op if attributes already registered. 41169a305230SLee Schermerhorn */ 4117a4a00b45SMuchun Song void hugetlb_register_node(struct node *node) 41189a305230SLee Schermerhorn { 41199a305230SLee Schermerhorn struct hstate *h; 412010fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 41219a305230SLee Schermerhorn int err; 41229a305230SLee Schermerhorn 4123a4a00b45SMuchun Song if (!hugetlb_sysfs_initialized) 4124a4a00b45SMuchun Song return; 4125a4a00b45SMuchun Song 41269a305230SLee Schermerhorn if (nhs->hugepages_kobj) 41279a305230SLee Schermerhorn return; /* already allocated */ 41289a305230SLee Schermerhorn 41299a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 413010fbcf4cSKay Sievers &node->dev.kobj); 41319a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 41329a305230SLee Schermerhorn return; 41339a305230SLee Schermerhorn 41349a305230SLee Schermerhorn for_each_hstate(h) { 41359a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 41369a305230SLee Schermerhorn nhs->hstate_kobjs, 41379a305230SLee Schermerhorn &per_node_hstate_attr_group); 41389a305230SLee Schermerhorn if (err) { 4139282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s for node %d\n", 414010fbcf4cSKay Sievers h->name, node->dev.id); 41419a305230SLee Schermerhorn hugetlb_unregister_node(node); 41429a305230SLee Schermerhorn break; 41439a305230SLee Schermerhorn } 41449a305230SLee Schermerhorn } 41459a305230SLee Schermerhorn } 41469a305230SLee Schermerhorn 41479a305230SLee Schermerhorn /* 41489b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 414910fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 415010fbcf4cSKay Sievers * registered their associated device by this time. 41519a305230SLee Schermerhorn */ 41527d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 41539a305230SLee Schermerhorn { 41549a305230SLee Schermerhorn int nid; 41559a305230SLee Schermerhorn 4156a4a00b45SMuchun Song for_each_online_node(nid) 4157b958d4d0SMuchun Song hugetlb_register_node(node_devices[nid]); 41589a305230SLee Schermerhorn } 41599a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 41609a305230SLee Schermerhorn 41619a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 41629a305230SLee Schermerhorn { 41639a305230SLee Schermerhorn BUG(); 41649a305230SLee Schermerhorn if (nidp) 41659a305230SLee Schermerhorn *nidp = -1; 41669a305230SLee Schermerhorn return NULL; 41679a305230SLee Schermerhorn } 41689a305230SLee Schermerhorn 41699a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 41709a305230SLee Schermerhorn 41719a305230SLee Schermerhorn #endif 41729a305230SLee Schermerhorn 4173263b8998SMiaohe Lin #ifdef CONFIG_CMA 4174263b8998SMiaohe Lin static void __init hugetlb_cma_check(void); 4175263b8998SMiaohe Lin #else 4176263b8998SMiaohe Lin static inline __init void hugetlb_cma_check(void) 4177263b8998SMiaohe Lin { 4178263b8998SMiaohe Lin } 4179263b8998SMiaohe Lin #endif 4180263b8998SMiaohe Lin 4181a4a00b45SMuchun Song static void __init hugetlb_sysfs_init(void) 4182a4a00b45SMuchun Song { 4183a4a00b45SMuchun Song struct hstate *h; 4184a4a00b45SMuchun Song int err; 4185a4a00b45SMuchun Song 4186a4a00b45SMuchun Song hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 4187a4a00b45SMuchun Song if (!hugepages_kobj) 4188a4a00b45SMuchun Song return; 4189a4a00b45SMuchun Song 4190a4a00b45SMuchun Song for_each_hstate(h) { 4191a4a00b45SMuchun Song err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 4192a4a00b45SMuchun Song hstate_kobjs, &hstate_attr_group); 4193a4a00b45SMuchun Song if (err) 4194a4a00b45SMuchun Song pr_err("HugeTLB: Unable to add hstate %s", h->name); 4195a4a00b45SMuchun Song } 4196a4a00b45SMuchun Song 4197a4a00b45SMuchun Song #ifdef CONFIG_NUMA 4198a4a00b45SMuchun Song hugetlb_sysfs_initialized = true; 4199a4a00b45SMuchun Song #endif 4200a4a00b45SMuchun Song hugetlb_register_all_nodes(); 4201a4a00b45SMuchun Song } 4202a4a00b45SMuchun Song 4203a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 4204a3437870SNishanth Aravamudan { 42058382d914SDavidlohr Bueso int i; 42068382d914SDavidlohr Bueso 4207d6995da3SMike Kravetz BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < 4208d6995da3SMike Kravetz __NR_HPAGEFLAGS); 4209d6995da3SMike Kravetz 4210c2833a5bSMike Kravetz if (!hugepages_supported()) { 4211c2833a5bSMike Kravetz if (hugetlb_max_hstate || default_hstate_max_huge_pages) 4212c2833a5bSMike Kravetz pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); 42130ef89d25SBenjamin Herrenschmidt return 0; 4214d715cf80SLiam R. Howlett } 4215d715cf80SLiam R. Howlett 4216282f4214SMike Kravetz /* 4217282f4214SMike Kravetz * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some 4218282f4214SMike Kravetz * architectures depend on setup being done here. 4219282f4214SMike Kravetz */ 4220a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 4221282f4214SMike Kravetz if (!parsed_default_hugepagesz) { 4222282f4214SMike Kravetz /* 4223282f4214SMike Kravetz * If we did not parse a default huge page size, set 4224282f4214SMike Kravetz * default_hstate_idx to HPAGE_SIZE hstate. And, if the 4225282f4214SMike Kravetz * number of huge pages for this default size was implicitly 4226282f4214SMike Kravetz * specified, set that here as well. 4227282f4214SMike Kravetz * Note that the implicit setting will overwrite an explicit 4228282f4214SMike Kravetz * setting. A warning will be printed in this case. 4229282f4214SMike Kravetz */ 4230282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); 4231f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 4232282f4214SMike Kravetz if (default_hstate.max_huge_pages) { 4233282f4214SMike Kravetz char buf[32]; 4234282f4214SMike Kravetz 4235282f4214SMike Kravetz string_get_size(huge_page_size(&default_hstate), 4236282f4214SMike Kravetz 1, STRING_UNITS_2, buf, 32); 4237282f4214SMike Kravetz pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", 4238282f4214SMike Kravetz default_hstate.max_huge_pages, buf); 4239282f4214SMike Kravetz pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", 4240282f4214SMike Kravetz default_hstate_max_huge_pages); 4241282f4214SMike Kravetz } 4242282f4214SMike Kravetz default_hstate.max_huge_pages = 4243282f4214SMike Kravetz default_hstate_max_huge_pages; 4244b5389086SZhenguo Yao 42450a7a0f6fSPeng Liu for_each_online_node(i) 4246b5389086SZhenguo Yao default_hstate.max_huge_pages_node[i] = 4247b5389086SZhenguo Yao default_hugepages_in_node[i]; 4248282f4214SMike Kravetz } 4249f8b74815SVaishali Thakkar } 4250a3437870SNishanth Aravamudan 4251cf11e85fSRoman Gushchin hugetlb_cma_check(); 4252a3437870SNishanth Aravamudan hugetlb_init_hstates(); 4253aa888a74SAndi Kleen gather_bootmem_prealloc(); 4254a3437870SNishanth Aravamudan report_hugepages(); 4255a3437870SNishanth Aravamudan 4256a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 42577179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 42589a305230SLee Schermerhorn 42598382d914SDavidlohr Bueso #ifdef CONFIG_SMP 42608382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 42618382d914SDavidlohr Bueso #else 42628382d914SDavidlohr Bueso num_fault_mutexes = 1; 42638382d914SDavidlohr Bueso #endif 4264c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 42656da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 42666da2ec56SKees Cook GFP_KERNEL); 4267c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 42688382d914SDavidlohr Bueso 42698382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 4270c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 4271a3437870SNishanth Aravamudan return 0; 4272a3437870SNishanth Aravamudan } 42733e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 4274a3437870SNishanth Aravamudan 4275ae94da89SMike Kravetz /* Overwritten by architectures with more huge page sizes */ 4276ae94da89SMike Kravetz bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) 42779fee021dSVaishali Thakkar { 4278ae94da89SMike Kravetz return size == HPAGE_SIZE; 42799fee021dSVaishali Thakkar } 42809fee021dSVaishali Thakkar 4281d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 4282a3437870SNishanth Aravamudan { 4283a3437870SNishanth Aravamudan struct hstate *h; 42848faa8b07SAndi Kleen unsigned long i; 42858faa8b07SAndi Kleen 4286a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 4287a3437870SNishanth Aravamudan return; 4288a3437870SNishanth Aravamudan } 428947d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 4290a3437870SNishanth Aravamudan BUG_ON(order == 0); 429147d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 429229383967SMike Kravetz mutex_init(&h->resize_lock); 4293a3437870SNishanth Aravamudan h->order = order; 4294aca78307SMiaohe Lin h->mask = ~(huge_page_size(h) - 1); 42958faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 42968faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 42970edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 429854f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 429954f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 4300a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 4301c2c3a60aSMiaohe Lin huge_page_size(h)/SZ_1K); 43028faa8b07SAndi Kleen 4303a3437870SNishanth Aravamudan parsed_hstate = h; 4304a3437870SNishanth Aravamudan } 4305a3437870SNishanth Aravamudan 4306b5389086SZhenguo Yao bool __init __weak hugetlb_node_alloc_supported(void) 4307b5389086SZhenguo Yao { 4308b5389086SZhenguo Yao return true; 4309b5389086SZhenguo Yao } 4310f87442f4SPeng Liu 4311f87442f4SPeng Liu static void __init hugepages_clear_pages_in_node(void) 4312f87442f4SPeng Liu { 4313f87442f4SPeng Liu if (!hugetlb_max_hstate) { 4314f87442f4SPeng Liu default_hstate_max_huge_pages = 0; 4315f87442f4SPeng Liu memset(default_hugepages_in_node, 0, 431610395680SMiaohe Lin sizeof(default_hugepages_in_node)); 4317f87442f4SPeng Liu } else { 4318f87442f4SPeng Liu parsed_hstate->max_huge_pages = 0; 4319f87442f4SPeng Liu memset(parsed_hstate->max_huge_pages_node, 0, 432010395680SMiaohe Lin sizeof(parsed_hstate->max_huge_pages_node)); 4321f87442f4SPeng Liu } 4322f87442f4SPeng Liu } 4323f87442f4SPeng Liu 4324282f4214SMike Kravetz /* 4325282f4214SMike Kravetz * hugepages command line processing 4326282f4214SMike Kravetz * hugepages normally follows a valid hugepagsz or default_hugepagsz 4327282f4214SMike Kravetz * specification. If not, ignore the hugepages value. hugepages can also 4328282f4214SMike Kravetz * be the first huge page command line option in which case it implicitly 4329282f4214SMike Kravetz * specifies the number of huge pages for the default size. 4330282f4214SMike Kravetz */ 4331282f4214SMike Kravetz static int __init hugepages_setup(char *s) 4332a3437870SNishanth Aravamudan { 4333a3437870SNishanth Aravamudan unsigned long *mhp; 43348faa8b07SAndi Kleen static unsigned long *last_mhp; 4335b5389086SZhenguo Yao int node = NUMA_NO_NODE; 4336b5389086SZhenguo Yao int count; 4337b5389086SZhenguo Yao unsigned long tmp; 4338b5389086SZhenguo Yao char *p = s; 4339a3437870SNishanth Aravamudan 43409fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 4341282f4214SMike Kravetz pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); 43429fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 4343f81f6e4bSPeng Liu return 1; 43449fee021dSVaishali Thakkar } 4345282f4214SMike Kravetz 4346a3437870SNishanth Aravamudan /* 4347282f4214SMike Kravetz * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter 4348282f4214SMike Kravetz * yet, so this hugepages= parameter goes to the "default hstate". 4349282f4214SMike Kravetz * Otherwise, it goes with the previously parsed hugepagesz or 4350282f4214SMike Kravetz * default_hugepagesz. 4351a3437870SNishanth Aravamudan */ 43529fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 4353a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 4354a3437870SNishanth Aravamudan else 4355a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 4356a3437870SNishanth Aravamudan 43578faa8b07SAndi Kleen if (mhp == last_mhp) { 4358282f4214SMike Kravetz pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); 4359f81f6e4bSPeng Liu return 1; 43608faa8b07SAndi Kleen } 43618faa8b07SAndi Kleen 4362b5389086SZhenguo Yao while (*p) { 4363b5389086SZhenguo Yao count = 0; 4364b5389086SZhenguo Yao if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4365b5389086SZhenguo Yao goto invalid; 4366b5389086SZhenguo Yao /* Parameter is node format */ 4367b5389086SZhenguo Yao if (p[count] == ':') { 4368b5389086SZhenguo Yao if (!hugetlb_node_alloc_supported()) { 4369b5389086SZhenguo Yao pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n"); 4370f81f6e4bSPeng Liu return 1; 4371b5389086SZhenguo Yao } 43720a7a0f6fSPeng Liu if (tmp >= MAX_NUMNODES || !node_online(tmp)) 4373e79ce983SLiu Yuntao goto invalid; 43740a7a0f6fSPeng Liu node = array_index_nospec(tmp, MAX_NUMNODES); 4375b5389086SZhenguo Yao p += count + 1; 4376b5389086SZhenguo Yao /* Parse hugepages */ 4377b5389086SZhenguo Yao if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4378b5389086SZhenguo Yao goto invalid; 4379b5389086SZhenguo Yao if (!hugetlb_max_hstate) 4380b5389086SZhenguo Yao default_hugepages_in_node[node] = tmp; 4381b5389086SZhenguo Yao else 4382b5389086SZhenguo Yao parsed_hstate->max_huge_pages_node[node] = tmp; 4383b5389086SZhenguo Yao *mhp += tmp; 4384b5389086SZhenguo Yao /* Go to parse next node*/ 4385b5389086SZhenguo Yao if (p[count] == ',') 4386b5389086SZhenguo Yao p += count + 1; 4387b5389086SZhenguo Yao else 4388b5389086SZhenguo Yao break; 4389b5389086SZhenguo Yao } else { 4390b5389086SZhenguo Yao if (p != s) 4391b5389086SZhenguo Yao goto invalid; 4392b5389086SZhenguo Yao *mhp = tmp; 4393b5389086SZhenguo Yao break; 4394b5389086SZhenguo Yao } 4395b5389086SZhenguo Yao } 4396a3437870SNishanth Aravamudan 43978faa8b07SAndi Kleen /* 43988faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 439904adbc3fSMiaohe Lin * But we need to allocate gigantic hstates here early to still 44008faa8b07SAndi Kleen * use the bootmem allocator. 44018faa8b07SAndi Kleen */ 440204adbc3fSMiaohe Lin if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) 44038faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 44048faa8b07SAndi Kleen 44058faa8b07SAndi Kleen last_mhp = mhp; 44068faa8b07SAndi Kleen 4407a3437870SNishanth Aravamudan return 1; 4408b5389086SZhenguo Yao 4409b5389086SZhenguo Yao invalid: 4410b5389086SZhenguo Yao pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p); 4411f87442f4SPeng Liu hugepages_clear_pages_in_node(); 4412f81f6e4bSPeng Liu return 1; 4413a3437870SNishanth Aravamudan } 4414282f4214SMike Kravetz __setup("hugepages=", hugepages_setup); 4415e11bfbfcSNick Piggin 4416282f4214SMike Kravetz /* 4417282f4214SMike Kravetz * hugepagesz command line processing 4418282f4214SMike Kravetz * A specific huge page size can only be specified once with hugepagesz. 4419282f4214SMike Kravetz * hugepagesz is followed by hugepages on the command line. The global 4420282f4214SMike Kravetz * variable 'parsed_valid_hugepagesz' is used to determine if prior 4421282f4214SMike Kravetz * hugepagesz argument was valid. 4422282f4214SMike Kravetz */ 4423359f2544SMike Kravetz static int __init hugepagesz_setup(char *s) 4424e11bfbfcSNick Piggin { 4425359f2544SMike Kravetz unsigned long size; 4426282f4214SMike Kravetz struct hstate *h; 4427282f4214SMike Kravetz 4428282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4429359f2544SMike Kravetz size = (unsigned long)memparse(s, NULL); 4430359f2544SMike Kravetz 4431359f2544SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4432282f4214SMike Kravetz pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); 4433f81f6e4bSPeng Liu return 1; 4434359f2544SMike Kravetz } 4435359f2544SMike Kravetz 4436282f4214SMike Kravetz h = size_to_hstate(size); 4437282f4214SMike Kravetz if (h) { 4438282f4214SMike Kravetz /* 4439282f4214SMike Kravetz * hstate for this size already exists. This is normally 4440282f4214SMike Kravetz * an error, but is allowed if the existing hstate is the 4441282f4214SMike Kravetz * default hstate. More specifically, it is only allowed if 4442282f4214SMike Kravetz * the number of huge pages for the default hstate was not 4443282f4214SMike Kravetz * previously specified. 4444282f4214SMike Kravetz */ 4445282f4214SMike Kravetz if (!parsed_default_hugepagesz || h != &default_hstate || 4446282f4214SMike Kravetz default_hstate.max_huge_pages) { 4447282f4214SMike Kravetz pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); 4448f81f6e4bSPeng Liu return 1; 444938237830SMike Kravetz } 445038237830SMike Kravetz 4451282f4214SMike Kravetz /* 4452282f4214SMike Kravetz * No need to call hugetlb_add_hstate() as hstate already 4453282f4214SMike Kravetz * exists. But, do set parsed_hstate so that a following 4454282f4214SMike Kravetz * hugepages= parameter will be applied to this hstate. 4455282f4214SMike Kravetz */ 4456282f4214SMike Kravetz parsed_hstate = h; 4457282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4458e11bfbfcSNick Piggin return 1; 4459e11bfbfcSNick Piggin } 4460282f4214SMike Kravetz 4461359f2544SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4462282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4463359f2544SMike Kravetz return 1; 4464359f2544SMike Kravetz } 4465359f2544SMike Kravetz __setup("hugepagesz=", hugepagesz_setup); 4466359f2544SMike Kravetz 4467282f4214SMike Kravetz /* 4468282f4214SMike Kravetz * default_hugepagesz command line input 4469282f4214SMike Kravetz * Only one instance of default_hugepagesz allowed on command line. 4470282f4214SMike Kravetz */ 4471ae94da89SMike Kravetz static int __init default_hugepagesz_setup(char *s) 4472e11bfbfcSNick Piggin { 4473ae94da89SMike Kravetz unsigned long size; 4474b5389086SZhenguo Yao int i; 4475ae94da89SMike Kravetz 4476282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4477282f4214SMike Kravetz if (parsed_default_hugepagesz) { 4478282f4214SMike Kravetz pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); 4479f81f6e4bSPeng Liu return 1; 4480282f4214SMike Kravetz } 4481282f4214SMike Kravetz 4482282f4214SMike Kravetz size = (unsigned long)memparse(s, NULL); 4483282f4214SMike Kravetz 4484282f4214SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4485282f4214SMike Kravetz pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); 4486f81f6e4bSPeng Liu return 1; 4487282f4214SMike Kravetz } 4488282f4214SMike Kravetz 4489282f4214SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4490282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4491282f4214SMike Kravetz parsed_default_hugepagesz = true; 4492282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(size)); 4493282f4214SMike Kravetz 4494282f4214SMike Kravetz /* 4495282f4214SMike Kravetz * The number of default huge pages (for this size) could have been 4496282f4214SMike Kravetz * specified as the first hugetlb parameter: hugepages=X. If so, 4497282f4214SMike Kravetz * then default_hstate_max_huge_pages is set. If the default huge 4498282f4214SMike Kravetz * page size is gigantic (>= MAX_ORDER), then the pages must be 4499282f4214SMike Kravetz * allocated here from bootmem allocator. 4500282f4214SMike Kravetz */ 4501282f4214SMike Kravetz if (default_hstate_max_huge_pages) { 4502282f4214SMike Kravetz default_hstate.max_huge_pages = default_hstate_max_huge_pages; 45030a7a0f6fSPeng Liu for_each_online_node(i) 4504b5389086SZhenguo Yao default_hstate.max_huge_pages_node[i] = 4505b5389086SZhenguo Yao default_hugepages_in_node[i]; 4506282f4214SMike Kravetz if (hstate_is_gigantic(&default_hstate)) 4507282f4214SMike Kravetz hugetlb_hstate_alloc_pages(&default_hstate); 4508282f4214SMike Kravetz default_hstate_max_huge_pages = 0; 4509282f4214SMike Kravetz } 4510282f4214SMike Kravetz 4511e11bfbfcSNick Piggin return 1; 4512e11bfbfcSNick Piggin } 4513ae94da89SMike Kravetz __setup("default_hugepagesz=", default_hugepagesz_setup); 4514a3437870SNishanth Aravamudan 4515d2226ebdSFeng Tang static nodemask_t *policy_mbind_nodemask(gfp_t gfp) 4516d2226ebdSFeng Tang { 4517d2226ebdSFeng Tang #ifdef CONFIG_NUMA 4518d2226ebdSFeng Tang struct mempolicy *mpol = get_task_policy(current); 4519d2226ebdSFeng Tang 4520d2226ebdSFeng Tang /* 4521d2226ebdSFeng Tang * Only enforce MPOL_BIND policy which overlaps with cpuset policy 4522d2226ebdSFeng Tang * (from policy_nodemask) specifically for hugetlb case 4523d2226ebdSFeng Tang */ 4524d2226ebdSFeng Tang if (mpol->mode == MPOL_BIND && 4525d2226ebdSFeng Tang (apply_policy_zone(mpol, gfp_zone(gfp)) && 4526d2226ebdSFeng Tang cpuset_nodemask_valid_mems_allowed(&mpol->nodes))) 4527d2226ebdSFeng Tang return &mpol->nodes; 4528d2226ebdSFeng Tang #endif 4529d2226ebdSFeng Tang return NULL; 4530d2226ebdSFeng Tang } 4531d2226ebdSFeng Tang 45328ca39e68SMuchun Song static unsigned int allowed_mems_nr(struct hstate *h) 45338a213460SNishanth Aravamudan { 45348a213460SNishanth Aravamudan int node; 45358a213460SNishanth Aravamudan unsigned int nr = 0; 4536d2226ebdSFeng Tang nodemask_t *mbind_nodemask; 45378ca39e68SMuchun Song unsigned int *array = h->free_huge_pages_node; 45388ca39e68SMuchun Song gfp_t gfp_mask = htlb_alloc_mask(h); 45398a213460SNishanth Aravamudan 4540d2226ebdSFeng Tang mbind_nodemask = policy_mbind_nodemask(gfp_mask); 45418ca39e68SMuchun Song for_each_node_mask(node, cpuset_current_mems_allowed) { 4542d2226ebdSFeng Tang if (!mbind_nodemask || node_isset(node, *mbind_nodemask)) 45438a213460SNishanth Aravamudan nr += array[node]; 45448ca39e68SMuchun Song } 45458a213460SNishanth Aravamudan 45468a213460SNishanth Aravamudan return nr; 45478a213460SNishanth Aravamudan } 45488a213460SNishanth Aravamudan 45498a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 455017743798SMuchun Song static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, 455117743798SMuchun Song void *buffer, size_t *length, 455217743798SMuchun Song loff_t *ppos, unsigned long *out) 455317743798SMuchun Song { 455417743798SMuchun Song struct ctl_table dup_table; 455517743798SMuchun Song 455617743798SMuchun Song /* 455717743798SMuchun Song * In order to avoid races with __do_proc_doulongvec_minmax(), we 455817743798SMuchun Song * can duplicate the @table and alter the duplicate of it. 455917743798SMuchun Song */ 456017743798SMuchun Song dup_table = *table; 456117743798SMuchun Song dup_table.data = out; 456217743798SMuchun Song 456317743798SMuchun Song return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); 456417743798SMuchun Song } 456517743798SMuchun Song 456606808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 456706808b08SLee Schermerhorn struct ctl_table *table, int write, 456832927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 45691da177e4SLinus Torvalds { 4570e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 4571238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 457208d4a246SMichal Hocko int ret; 4573e5ff2159SAndi Kleen 4574457c1b27SNishanth Aravamudan if (!hugepages_supported()) 457586613628SJan Stancek return -EOPNOTSUPP; 4576457c1b27SNishanth Aravamudan 457717743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 457817743798SMuchun Song &tmp); 457908d4a246SMichal Hocko if (ret) 458008d4a246SMichal Hocko goto out; 4581e5ff2159SAndi Kleen 4582238d3c13SDavid Rientjes if (write) 4583238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 4584238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 458508d4a246SMichal Hocko out: 458608d4a246SMichal Hocko return ret; 45871da177e4SLinus Torvalds } 4588396faf03SMel Gorman 458906808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 459032927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 459106808b08SLee Schermerhorn { 459206808b08SLee Schermerhorn 459306808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 459406808b08SLee Schermerhorn buffer, length, ppos); 459506808b08SLee Schermerhorn } 459606808b08SLee Schermerhorn 459706808b08SLee Schermerhorn #ifdef CONFIG_NUMA 459806808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 459932927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 460006808b08SLee Schermerhorn { 460106808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 460206808b08SLee Schermerhorn buffer, length, ppos); 460306808b08SLee Schermerhorn } 460406808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 460506808b08SLee Schermerhorn 4606a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 460732927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 4608a3d0c6aaSNishanth Aravamudan { 4609a5516438SAndi Kleen struct hstate *h = &default_hstate; 4610e5ff2159SAndi Kleen unsigned long tmp; 461108d4a246SMichal Hocko int ret; 4612e5ff2159SAndi Kleen 4613457c1b27SNishanth Aravamudan if (!hugepages_supported()) 461486613628SJan Stancek return -EOPNOTSUPP; 4615457c1b27SNishanth Aravamudan 4616e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 4617e5ff2159SAndi Kleen 4618bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 4619adbe8726SEric B Munson return -EINVAL; 4620adbe8726SEric B Munson 462117743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 462217743798SMuchun Song &tmp); 462308d4a246SMichal Hocko if (ret) 462408d4a246SMichal Hocko goto out; 4625e5ff2159SAndi Kleen 4626e5ff2159SAndi Kleen if (write) { 4627db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4628e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 4629db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4630e5ff2159SAndi Kleen } 463108d4a246SMichal Hocko out: 463208d4a246SMichal Hocko return ret; 4633a3d0c6aaSNishanth Aravamudan } 4634a3d0c6aaSNishanth Aravamudan 46351da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 46361da177e4SLinus Torvalds 4637e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 46381da177e4SLinus Torvalds { 4639fcb2b0c5SRoman Gushchin struct hstate *h; 4640fcb2b0c5SRoman Gushchin unsigned long total = 0; 4641fcb2b0c5SRoman Gushchin 4642457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4643457c1b27SNishanth Aravamudan return; 4644fcb2b0c5SRoman Gushchin 4645fcb2b0c5SRoman Gushchin for_each_hstate(h) { 4646fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 4647fcb2b0c5SRoman Gushchin 4648aca78307SMiaohe Lin total += huge_page_size(h) * count; 4649fcb2b0c5SRoman Gushchin 4650fcb2b0c5SRoman Gushchin if (h == &default_hstate) 4651e1759c21SAlexey Dobriyan seq_printf(m, 46521da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 46531da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 4654b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 46557893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 46564f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 4657fcb2b0c5SRoman Gushchin count, 4658a5516438SAndi Kleen h->free_huge_pages, 4659a5516438SAndi Kleen h->resv_huge_pages, 4660a5516438SAndi Kleen h->surplus_huge_pages, 4661aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4662fcb2b0c5SRoman Gushchin } 4663fcb2b0c5SRoman Gushchin 4664aca78307SMiaohe Lin seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); 46651da177e4SLinus Torvalds } 46661da177e4SLinus Torvalds 46677981593bSJoe Perches int hugetlb_report_node_meminfo(char *buf, int len, int nid) 46681da177e4SLinus Torvalds { 4669a5516438SAndi Kleen struct hstate *h = &default_hstate; 46707981593bSJoe Perches 4671457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4672457c1b27SNishanth Aravamudan return 0; 46737981593bSJoe Perches 46747981593bSJoe Perches return sysfs_emit_at(buf, len, 46751da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 4676a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 4677a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 4678a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 4679a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 4680a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 46811da177e4SLinus Torvalds } 46821da177e4SLinus Torvalds 4683dcadcf1cSGang Li void hugetlb_show_meminfo_node(int nid) 4684949f7ec5SDavid Rientjes { 4685949f7ec5SDavid Rientjes struct hstate *h; 4686949f7ec5SDavid Rientjes 4687457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4688457c1b27SNishanth Aravamudan return; 4689457c1b27SNishanth Aravamudan 4690949f7ec5SDavid Rientjes for_each_hstate(h) 4691dcadcf1cSGang Li printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 4692949f7ec5SDavid Rientjes nid, 4693949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 4694949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 4695949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 4696aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4697949f7ec5SDavid Rientjes } 4698949f7ec5SDavid Rientjes 46995d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 47005d317b2bSNaoya Horiguchi { 47015d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 47025d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 47035d317b2bSNaoya Horiguchi } 47045d317b2bSNaoya Horiguchi 47051da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 47061da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 47071da177e4SLinus Torvalds { 4708d0028588SWanpeng Li struct hstate *h; 4709d0028588SWanpeng Li unsigned long nr_total_pages = 0; 4710d0028588SWanpeng Li 4711d0028588SWanpeng Li for_each_hstate(h) 4712d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 4713d0028588SWanpeng Li return nr_total_pages; 47141da177e4SLinus Torvalds } 47151da177e4SLinus Torvalds 4716a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 4717fc1b8a73SMel Gorman { 4718fc1b8a73SMel Gorman int ret = -ENOMEM; 4719fc1b8a73SMel Gorman 47200aa7f354SMiaohe Lin if (!delta) 47210aa7f354SMiaohe Lin return 0; 47220aa7f354SMiaohe Lin 4723db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4724fc1b8a73SMel Gorman /* 4725fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 4726fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 4727fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 4728fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 4729fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 4730fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 4731fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 4732fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 4733fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 4734fc1b8a73SMel Gorman * 4735fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 4736fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 4737fc1b8a73SMel Gorman * we fall back to check against current free page availability as 4738fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 4739fc1b8a73SMel Gorman * semantics that cpuset has. 47408ca39e68SMuchun Song * 47418ca39e68SMuchun Song * Apart from cpuset, we also have memory policy mechanism that 47428ca39e68SMuchun Song * also determines from which node the kernel will allocate memory 47438ca39e68SMuchun Song * in a NUMA system. So similar to cpuset, we also should consider 47448ca39e68SMuchun Song * the memory policy of the current task. Similar to the description 47458ca39e68SMuchun Song * above. 4746fc1b8a73SMel Gorman */ 4747fc1b8a73SMel Gorman if (delta > 0) { 4748a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 4749fc1b8a73SMel Gorman goto out; 4750fc1b8a73SMel Gorman 47518ca39e68SMuchun Song if (delta > allowed_mems_nr(h)) { 4752a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 4753fc1b8a73SMel Gorman goto out; 4754fc1b8a73SMel Gorman } 4755fc1b8a73SMel Gorman } 4756fc1b8a73SMel Gorman 4757fc1b8a73SMel Gorman ret = 0; 4758fc1b8a73SMel Gorman if (delta < 0) 4759a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 4760fc1b8a73SMel Gorman 4761fc1b8a73SMel Gorman out: 4762db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4763fc1b8a73SMel Gorman return ret; 4764fc1b8a73SMel Gorman } 4765fc1b8a73SMel Gorman 476684afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 476784afd99bSAndy Whitcroft { 4768f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 476984afd99bSAndy Whitcroft 477084afd99bSAndy Whitcroft /* 4771612b8a31SMike Kravetz * HPAGE_RESV_OWNER indicates a private mapping. 477284afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 477384afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 477484afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 477525985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 477684afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 477784afd99bSAndy Whitcroft * new reference here without additional locking. 477884afd99bSAndy Whitcroft */ 477909a26e83SMike Kravetz if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 478009a26e83SMike Kravetz resv_map_dup_hugetlb_cgroup_uncharge_info(resv); 4781f522c3acSJoonsoo Kim kref_get(&resv->refs); 478284afd99bSAndy Whitcroft } 47838d9bfb26SMike Kravetz 4784131a79b4SMike Kravetz /* 4785131a79b4SMike Kravetz * vma_lock structure for sharable mappings is vma specific. 4786612b8a31SMike Kravetz * Clear old pointer (if copied via vm_area_dup) and allocate 4787612b8a31SMike Kravetz * new structure. Before clearing, make sure vma_lock is not 4788612b8a31SMike Kravetz * for this vma. 4789131a79b4SMike Kravetz */ 4790131a79b4SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 4791612b8a31SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 4792612b8a31SMike Kravetz 4793612b8a31SMike Kravetz if (vma_lock) { 4794612b8a31SMike Kravetz if (vma_lock->vma != vma) { 4795131a79b4SMike Kravetz vma->vm_private_data = NULL; 47968d9bfb26SMike Kravetz hugetlb_vma_lock_alloc(vma); 4797612b8a31SMike Kravetz } else 4798612b8a31SMike Kravetz pr_warn("HugeTLB: vma_lock already exists in %s.\n", __func__); 4799612b8a31SMike Kravetz } else 4800612b8a31SMike Kravetz hugetlb_vma_lock_alloc(vma); 480109a26e83SMike Kravetz } 4802131a79b4SMike Kravetz } 480384afd99bSAndy Whitcroft 4804a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 4805a1e78772SMel Gorman { 4806a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 48078d9bfb26SMike Kravetz struct resv_map *resv; 480890481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 48094e35f483SJoonsoo Kim unsigned long reserve, start, end; 48101c5ecae3SMike Kravetz long gbl_reserve; 481184afd99bSAndy Whitcroft 48128d9bfb26SMike Kravetz hugetlb_vma_lock_free(vma); 48138d9bfb26SMike Kravetz 48148d9bfb26SMike Kravetz resv = vma_resv_map(vma); 48154e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 48164e35f483SJoonsoo Kim return; 48174e35f483SJoonsoo Kim 4818a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 4819a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 482084afd99bSAndy Whitcroft 48214e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 4822e9fe92aeSMina Almasry hugetlb_cgroup_uncharge_counter(resv, start, end); 48237251ff78SAdam Litke if (reserve) { 48241c5ecae3SMike Kravetz /* 48251c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 48261c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 48271c5ecae3SMike Kravetz */ 48281c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 48291c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 48307251ff78SAdam Litke } 4831e9fe92aeSMina Almasry 4832e9fe92aeSMina Almasry kref_put(&resv->refs, resv_map_release); 4833a1e78772SMel Gorman } 4834a1e78772SMel Gorman 483531383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 483631383c68SDan Williams { 483731383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 483831383c68SDan Williams return -EINVAL; 4839b30c14cdSJames Houghton 4840b30c14cdSJames Houghton /* 4841b30c14cdSJames Houghton * PMD sharing is only possible for PUD_SIZE-aligned address ranges 4842b30c14cdSJames Houghton * in HugeTLB VMAs. If we will lose PUD_SIZE alignment due to this 4843b30c14cdSJames Houghton * split, unshare PMDs in the PUD_SIZE interval surrounding addr now. 4844b30c14cdSJames Houghton */ 4845b30c14cdSJames Houghton if (addr & ~PUD_MASK) { 4846b30c14cdSJames Houghton /* 4847b30c14cdSJames Houghton * hugetlb_vm_op_split is called right before we attempt to 4848b30c14cdSJames Houghton * split the VMA. We will need to unshare PMDs in the old and 4849b30c14cdSJames Houghton * new VMAs, so let's unshare before we split. 4850b30c14cdSJames Houghton */ 4851b30c14cdSJames Houghton unsigned long floor = addr & PUD_MASK; 4852b30c14cdSJames Houghton unsigned long ceil = floor + PUD_SIZE; 4853b30c14cdSJames Houghton 4854b30c14cdSJames Houghton if (floor >= vma->vm_start && ceil <= vma->vm_end) 4855b30c14cdSJames Houghton hugetlb_unshare_pmds(vma, floor, ceil); 4856b30c14cdSJames Houghton } 4857b30c14cdSJames Houghton 485831383c68SDan Williams return 0; 485931383c68SDan Williams } 486031383c68SDan Williams 486105ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 486205ea8860SDan Williams { 4863aca78307SMiaohe Lin return huge_page_size(hstate_vma(vma)); 486405ea8860SDan Williams } 486505ea8860SDan Williams 48661da177e4SLinus Torvalds /* 48671da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 48681da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 48696c26d310SMiaohe Lin * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get 48701da177e4SLinus Torvalds * this far. 48711da177e4SLinus Torvalds */ 4872b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 48731da177e4SLinus Torvalds { 48741da177e4SLinus Torvalds BUG(); 4875d0217ac0SNick Piggin return 0; 48761da177e4SLinus Torvalds } 48771da177e4SLinus Torvalds 4878eec3636aSJane Chu /* 4879eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 4880eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 4881eec3636aSJane Chu * This is because under System V memory model, mappings created via 4882eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 4883eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 4884eec3636aSJane Chu */ 4885f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 4886d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 488784afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 4888a1e78772SMel Gorman .close = hugetlb_vm_op_close, 4889dd3b614fSDmitry Safonov .may_split = hugetlb_vm_op_split, 489005ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 48911da177e4SLinus Torvalds }; 48921da177e4SLinus Torvalds 48931e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 48941e8f889bSDavid Gibson int writable) 489563551ae0SDavid Gibson { 489663551ae0SDavid Gibson pte_t entry; 489779c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 489863551ae0SDavid Gibson 48991e8f889bSDavid Gibson if (writable) { 4900106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 4901106c992aSGerald Schaefer vma->vm_page_prot))); 490263551ae0SDavid Gibson } else { 4903106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 4904106c992aSGerald Schaefer vma->vm_page_prot)); 490563551ae0SDavid Gibson } 490663551ae0SDavid Gibson entry = pte_mkyoung(entry); 490779c1c594SChristophe Leroy entry = arch_make_huge_pte(entry, shift, vma->vm_flags); 490863551ae0SDavid Gibson 490963551ae0SDavid Gibson return entry; 491063551ae0SDavid Gibson } 491163551ae0SDavid Gibson 49121e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 49131e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 49141e8f889bSDavid Gibson { 49151e8f889bSDavid Gibson pte_t entry; 49161e8f889bSDavid Gibson 4917106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 491832f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 49194b3073e1SRussell King update_mmu_cache(vma, address, ptep); 49201e8f889bSDavid Gibson } 49211e8f889bSDavid Gibson 4922d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 49234a705fefSNaoya Horiguchi { 49244a705fefSNaoya Horiguchi swp_entry_t swp; 49254a705fefSNaoya Horiguchi 49264a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 4927d5ed7444SAneesh Kumar K.V return false; 49284a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4929d79d176aSBaoquan He if (is_migration_entry(swp)) 4930d5ed7444SAneesh Kumar K.V return true; 49314a705fefSNaoya Horiguchi else 4932d5ed7444SAneesh Kumar K.V return false; 49334a705fefSNaoya Horiguchi } 49344a705fefSNaoya Horiguchi 49353e5c3600SBaoquan He static bool is_hugetlb_entry_hwpoisoned(pte_t pte) 49364a705fefSNaoya Horiguchi { 49374a705fefSNaoya Horiguchi swp_entry_t swp; 49384a705fefSNaoya Horiguchi 49394a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 49403e5c3600SBaoquan He return false; 49414a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4942d79d176aSBaoquan He if (is_hwpoison_entry(swp)) 49433e5c3600SBaoquan He return true; 49444a705fefSNaoya Horiguchi else 49453e5c3600SBaoquan He return false; 49464a705fefSNaoya Horiguchi } 49471e8f889bSDavid Gibson 49484eae4efaSPeter Xu static void 4949*ea4c353dSSidhartha Kumar hugetlb_install_folio(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, 4950*ea4c353dSSidhartha Kumar struct folio *new_folio) 49514eae4efaSPeter Xu { 4952*ea4c353dSSidhartha Kumar __folio_mark_uptodate(new_folio); 4953*ea4c353dSSidhartha Kumar hugepage_add_new_anon_rmap(&new_folio->page, vma, addr); 4954*ea4c353dSSidhartha Kumar set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, &new_folio->page, 1)); 49554eae4efaSPeter Xu hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); 4956*ea4c353dSSidhartha Kumar folio_set_hugetlb_migratable(new_folio); 49574eae4efaSPeter Xu } 49584eae4efaSPeter Xu 495963551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 4960bc70fbf2SPeter Xu struct vm_area_struct *dst_vma, 4961bc70fbf2SPeter Xu struct vm_area_struct *src_vma) 496263551ae0SDavid Gibson { 49633aa4ed80SMiaohe Lin pte_t *src_pte, *dst_pte, entry; 496463551ae0SDavid Gibson struct page *ptepage; 49651c59827dSHugh Dickins unsigned long addr; 4966bc70fbf2SPeter Xu bool cow = is_cow_mapping(src_vma->vm_flags); 4967bc70fbf2SPeter Xu struct hstate *h = hstate_vma(src_vma); 4968a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 49694eae4efaSPeter Xu unsigned long npages = pages_per_huge_page(h); 4970ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 4971e95a9851SMike Kravetz unsigned long last_addr_mask; 4972e8569dd2SAndreas Sandberg int ret = 0; 49731e8f889bSDavid Gibson 4974ac46d4f3SJérôme Glisse if (cow) { 49757d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src, 4976bc70fbf2SPeter Xu src_vma->vm_start, 4977bc70fbf2SPeter Xu src_vma->vm_end); 4978ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4979623a1ddfSDavid Hildenbrand mmap_assert_write_locked(src); 4980623a1ddfSDavid Hildenbrand raw_write_seqcount_begin(&src->write_protect_seq); 498140549ba8SMike Kravetz } else { 498240549ba8SMike Kravetz /* 498340549ba8SMike Kravetz * For shared mappings the vma lock must be held before 49849c67a207SPeter Xu * calling hugetlb_walk() in the src vma. Otherwise, the 498540549ba8SMike Kravetz * returned ptep could go away if part of a shared pmd and 498640549ba8SMike Kravetz * another thread calls huge_pmd_unshare. 498740549ba8SMike Kravetz */ 498840549ba8SMike Kravetz hugetlb_vma_lock_read(src_vma); 4989ac46d4f3SJérôme Glisse } 4990e8569dd2SAndreas Sandberg 4991e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 4992bc70fbf2SPeter Xu for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) { 4993cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 49949c67a207SPeter Xu src_pte = hugetlb_walk(src_vma, addr, sz); 4995e95a9851SMike Kravetz if (!src_pte) { 4996e95a9851SMike Kravetz addr |= last_addr_mask; 4997c74df32cSHugh Dickins continue; 4998e95a9851SMike Kravetz } 4999bc70fbf2SPeter Xu dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz); 5000e8569dd2SAndreas Sandberg if (!dst_pte) { 5001e8569dd2SAndreas Sandberg ret = -ENOMEM; 5002e8569dd2SAndreas Sandberg break; 5003e8569dd2SAndreas Sandberg } 5004c5c99429SLarry Woodman 50055e41540cSMike Kravetz /* 50065e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 50075e41540cSMike Kravetz * 50083aa4ed80SMiaohe Lin * dst_pte == src_pte is the common case of src/dest sharing. 50095e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 50103aa4ed80SMiaohe Lin * another vma. So page_count of ptep page is checked instead 50113aa4ed80SMiaohe Lin * to reliably determine whether pte is shared. 50125e41540cSMike Kravetz */ 50133aa4ed80SMiaohe Lin if (page_count(virt_to_page(dst_pte)) > 1) { 5014e95a9851SMike Kravetz addr |= last_addr_mask; 5015c5c99429SLarry Woodman continue; 5016e95a9851SMike Kravetz } 5017c5c99429SLarry Woodman 5018cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 5019cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 5020cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 50214a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 50224eae4efaSPeter Xu again: 50233aa4ed80SMiaohe Lin if (huge_pte_none(entry)) { 50245e41540cSMike Kravetz /* 50253aa4ed80SMiaohe Lin * Skip if src entry none. 50265e41540cSMike Kravetz */ 50274a705fefSNaoya Horiguchi ; 5028c2cb0dccSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) { 5029c2cb0dccSNaoya Horiguchi bool uffd_wp = huge_pte_uffd_wp(entry); 5030c2cb0dccSNaoya Horiguchi 5031c2cb0dccSNaoya Horiguchi if (!userfaultfd_wp(dst_vma) && uffd_wp) 5032c2cb0dccSNaoya Horiguchi entry = huge_pte_clear_uffd_wp(entry); 5033c2cb0dccSNaoya Horiguchi set_huge_pte_at(dst, addr, dst_pte, entry); 5034c2cb0dccSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry))) { 50354a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 5036bc70fbf2SPeter Xu bool uffd_wp = huge_pte_uffd_wp(entry); 50374a705fefSNaoya Horiguchi 50386c287605SDavid Hildenbrand if (!is_readable_migration_entry(swp_entry) && cow) { 50394a705fefSNaoya Horiguchi /* 50404a705fefSNaoya Horiguchi * COW mappings require pages in both 50414a705fefSNaoya Horiguchi * parent and child to be set to read. 50424a705fefSNaoya Horiguchi */ 50434dd845b5SAlistair Popple swp_entry = make_readable_migration_entry( 50444dd845b5SAlistair Popple swp_offset(swp_entry)); 50454a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 5046bc70fbf2SPeter Xu if (userfaultfd_wp(src_vma) && uffd_wp) 5047bc70fbf2SPeter Xu entry = huge_pte_mkuffd_wp(entry); 504818f39629SQi Zheng set_huge_pte_at(src, addr, src_pte, entry); 50494a705fefSNaoya Horiguchi } 5050bc70fbf2SPeter Xu if (!userfaultfd_wp(dst_vma) && uffd_wp) 5051bc70fbf2SPeter Xu entry = huge_pte_clear_uffd_wp(entry); 505218f39629SQi Zheng set_huge_pte_at(dst, addr, dst_pte, entry); 5053bc70fbf2SPeter Xu } else if (unlikely(is_pte_marker(entry))) { 50547e3ce3f8SPeter Xu /* No swap on hugetlb */ 50557e3ce3f8SPeter Xu WARN_ON_ONCE( 50567e3ce3f8SPeter Xu is_swapin_error_entry(pte_to_swp_entry(entry))); 5057bc70fbf2SPeter Xu /* 5058bc70fbf2SPeter Xu * We copy the pte marker only if the dst vma has 5059bc70fbf2SPeter Xu * uffd-wp enabled. 5060bc70fbf2SPeter Xu */ 5061bc70fbf2SPeter Xu if (userfaultfd_wp(dst_vma)) 5062bc70fbf2SPeter Xu set_huge_pte_at(dst, addr, dst_pte, entry); 50634a705fefSNaoya Horiguchi } else { 50644eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 50654eae4efaSPeter Xu ptepage = pte_page(entry); 50664eae4efaSPeter Xu get_page(ptepage); 50674eae4efaSPeter Xu 50684eae4efaSPeter Xu /* 5069fb3d824dSDavid Hildenbrand * Failing to duplicate the anon rmap is a rare case 5070fb3d824dSDavid Hildenbrand * where we see pinned hugetlb pages while they're 5071fb3d824dSDavid Hildenbrand * prone to COW. We need to do the COW earlier during 5072fb3d824dSDavid Hildenbrand * fork. 50734eae4efaSPeter Xu * 50744eae4efaSPeter Xu * When pre-allocating the page or copying data, we 50754eae4efaSPeter Xu * need to be without the pgtable locks since we could 50764eae4efaSPeter Xu * sleep during the process. 50774eae4efaSPeter Xu */ 5078fb3d824dSDavid Hildenbrand if (!PageAnon(ptepage)) { 5079fb3d824dSDavid Hildenbrand page_dup_file_rmap(ptepage, true); 5080bc70fbf2SPeter Xu } else if (page_try_dup_anon_rmap(ptepage, true, 5081bc70fbf2SPeter Xu src_vma)) { 50824eae4efaSPeter Xu pte_t src_pte_old = entry; 50834eae4efaSPeter Xu struct page *new; 50844eae4efaSPeter Xu 50854eae4efaSPeter Xu spin_unlock(src_ptl); 50864eae4efaSPeter Xu spin_unlock(dst_ptl); 50874eae4efaSPeter Xu /* Do not use reserve as it's private owned */ 5088bc70fbf2SPeter Xu new = alloc_huge_page(dst_vma, addr, 1); 50894eae4efaSPeter Xu if (IS_ERR(new)) { 50904eae4efaSPeter Xu put_page(ptepage); 50914eae4efaSPeter Xu ret = PTR_ERR(new); 50924eae4efaSPeter Xu break; 50934eae4efaSPeter Xu } 5094bc70fbf2SPeter Xu copy_user_huge_page(new, ptepage, addr, dst_vma, 50954eae4efaSPeter Xu npages); 50964eae4efaSPeter Xu put_page(ptepage); 50974eae4efaSPeter Xu 50984eae4efaSPeter Xu /* Install the new huge page if src pte stable */ 50994eae4efaSPeter Xu dst_ptl = huge_pte_lock(h, dst, dst_pte); 51004eae4efaSPeter Xu src_ptl = huge_pte_lockptr(h, src, src_pte); 51014eae4efaSPeter Xu spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 51024eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 51034eae4efaSPeter Xu if (!pte_same(src_pte_old, entry)) { 5104bc70fbf2SPeter Xu restore_reserve_on_error(h, dst_vma, addr, 5105846be085SMike Kravetz new); 51064eae4efaSPeter Xu put_page(new); 51073aa4ed80SMiaohe Lin /* huge_ptep of dst_pte won't change as in child */ 51084eae4efaSPeter Xu goto again; 51094eae4efaSPeter Xu } 5110*ea4c353dSSidhartha Kumar hugetlb_install_folio(dst_vma, dst_pte, addr, page_folio(new)); 51114eae4efaSPeter Xu spin_unlock(src_ptl); 51124eae4efaSPeter Xu spin_unlock(dst_ptl); 51134eae4efaSPeter Xu continue; 51144eae4efaSPeter Xu } 51154eae4efaSPeter Xu 511634ee645eSJoerg Roedel if (cow) { 51170f10851eSJérôme Glisse /* 51180f10851eSJérôme Glisse * No need to notify as we are downgrading page 51190f10851eSJérôme Glisse * table protection not changing it to point 51200f10851eSJérôme Glisse * to a new page. 51210f10851eSJérôme Glisse * 5122ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 51230f10851eSJérôme Glisse */ 51247f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 512584894e1cSPeter Xu entry = huge_pte_wrprotect(entry); 512634ee645eSJoerg Roedel } 51274eae4efaSPeter Xu 512863551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 51294eae4efaSPeter Xu hugetlb_count_add(npages, dst); 51301c59827dSHugh Dickins } 5131cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 5132cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 513363551ae0SDavid Gibson } 513463551ae0SDavid Gibson 5135623a1ddfSDavid Hildenbrand if (cow) { 5136623a1ddfSDavid Hildenbrand raw_write_seqcount_end(&src->write_protect_seq); 5137ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 513840549ba8SMike Kravetz } else { 513940549ba8SMike Kravetz hugetlb_vma_unlock_read(src_vma); 5140623a1ddfSDavid Hildenbrand } 5141e8569dd2SAndreas Sandberg 5142e8569dd2SAndreas Sandberg return ret; 514363551ae0SDavid Gibson } 514463551ae0SDavid Gibson 5145550a7d60SMina Almasry static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr, 5146db110a99SAneesh Kumar K.V unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte) 5147550a7d60SMina Almasry { 5148550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 5149550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 5150550a7d60SMina Almasry spinlock_t *src_ptl, *dst_ptl; 5151db110a99SAneesh Kumar K.V pte_t pte; 5152550a7d60SMina Almasry 5153550a7d60SMina Almasry dst_ptl = huge_pte_lock(h, mm, dst_pte); 5154550a7d60SMina Almasry src_ptl = huge_pte_lockptr(h, mm, src_pte); 5155550a7d60SMina Almasry 5156550a7d60SMina Almasry /* 5157550a7d60SMina Almasry * We don't have to worry about the ordering of src and dst ptlocks 51588651a137SLorenzo Stoakes * because exclusive mmap_lock (or the i_mmap_lock) prevents deadlock. 5159550a7d60SMina Almasry */ 5160550a7d60SMina Almasry if (src_ptl != dst_ptl) 5161550a7d60SMina Almasry spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 5162550a7d60SMina Almasry 5163550a7d60SMina Almasry pte = huge_ptep_get_and_clear(mm, old_addr, src_pte); 5164550a7d60SMina Almasry set_huge_pte_at(mm, new_addr, dst_pte, pte); 5165550a7d60SMina Almasry 5166550a7d60SMina Almasry if (src_ptl != dst_ptl) 5167550a7d60SMina Almasry spin_unlock(src_ptl); 5168550a7d60SMina Almasry spin_unlock(dst_ptl); 5169550a7d60SMina Almasry } 5170550a7d60SMina Almasry 5171550a7d60SMina Almasry int move_hugetlb_page_tables(struct vm_area_struct *vma, 5172550a7d60SMina Almasry struct vm_area_struct *new_vma, 5173550a7d60SMina Almasry unsigned long old_addr, unsigned long new_addr, 5174550a7d60SMina Almasry unsigned long len) 5175550a7d60SMina Almasry { 5176550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 5177550a7d60SMina Almasry struct address_space *mapping = vma->vm_file->f_mapping; 5178550a7d60SMina Almasry unsigned long sz = huge_page_size(h); 5179550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 5180550a7d60SMina Almasry unsigned long old_end = old_addr + len; 5181e95a9851SMike Kravetz unsigned long last_addr_mask; 5182550a7d60SMina Almasry pte_t *src_pte, *dst_pte; 5183550a7d60SMina Almasry struct mmu_notifier_range range; 51843d0b95cdSBaolin Wang bool shared_pmd = false; 5185550a7d60SMina Almasry 51867d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, old_addr, 5187550a7d60SMina Almasry old_end); 5188550a7d60SMina Almasry adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 51893d0b95cdSBaolin Wang /* 51903d0b95cdSBaolin Wang * In case of shared PMDs, we should cover the maximum possible 51913d0b95cdSBaolin Wang * range. 51923d0b95cdSBaolin Wang */ 51933d0b95cdSBaolin Wang flush_cache_range(vma, range.start, range.end); 51943d0b95cdSBaolin Wang 5195550a7d60SMina Almasry mmu_notifier_invalidate_range_start(&range); 5196e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 5197550a7d60SMina Almasry /* Prevent race with file truncation */ 519840549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 5199550a7d60SMina Almasry i_mmap_lock_write(mapping); 5200550a7d60SMina Almasry for (; old_addr < old_end; old_addr += sz, new_addr += sz) { 52019c67a207SPeter Xu src_pte = hugetlb_walk(vma, old_addr, sz); 5202e95a9851SMike Kravetz if (!src_pte) { 5203e95a9851SMike Kravetz old_addr |= last_addr_mask; 5204e95a9851SMike Kravetz new_addr |= last_addr_mask; 5205550a7d60SMina Almasry continue; 5206e95a9851SMike Kravetz } 5207550a7d60SMina Almasry if (huge_pte_none(huge_ptep_get(src_pte))) 5208550a7d60SMina Almasry continue; 5209550a7d60SMina Almasry 52104ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) { 52113d0b95cdSBaolin Wang shared_pmd = true; 52124ddb4d91SMike Kravetz old_addr |= last_addr_mask; 52134ddb4d91SMike Kravetz new_addr |= last_addr_mask; 5214550a7d60SMina Almasry continue; 52153d0b95cdSBaolin Wang } 5216550a7d60SMina Almasry 5217550a7d60SMina Almasry dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz); 5218550a7d60SMina Almasry if (!dst_pte) 5219550a7d60SMina Almasry break; 5220550a7d60SMina Almasry 5221db110a99SAneesh Kumar K.V move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte); 5222550a7d60SMina Almasry } 52233d0b95cdSBaolin Wang 52243d0b95cdSBaolin Wang if (shared_pmd) 52253d0b95cdSBaolin Wang flush_tlb_range(vma, range.start, range.end); 52263d0b95cdSBaolin Wang else 5227550a7d60SMina Almasry flush_tlb_range(vma, old_end - len, old_end); 5228550a7d60SMina Almasry mmu_notifier_invalidate_range_end(&range); 522913e4ad2cSNadav Amit i_mmap_unlock_write(mapping); 523040549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 5231550a7d60SMina Almasry 5232550a7d60SMina Almasry return len + old_addr - old_end; 5233550a7d60SMina Almasry } 5234550a7d60SMina Almasry 523573c54763SPeter Xu static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 523624669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 523705e90bd0SPeter Xu struct page *ref_page, zap_flags_t zap_flags) 523863551ae0SDavid Gibson { 523963551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 524063551ae0SDavid Gibson unsigned long address; 5241c7546f8fSDavid Gibson pte_t *ptep; 524263551ae0SDavid Gibson pte_t pte; 5243cb900f41SKirill A. Shutemov spinlock_t *ptl; 524463551ae0SDavid Gibson struct page *page; 5245a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 5246a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 5247e95a9851SMike Kravetz unsigned long last_addr_mask; 5248a4a118f2SNadav Amit bool force_flush = false; 5249a5516438SAndi Kleen 525063551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 5251a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 5252a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 525363551ae0SDavid Gibson 525407e32661SAneesh Kumar K.V /* 525507e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 525607e32661SAneesh Kumar K.V * to huge page. 525707e32661SAneesh Kumar K.V */ 5258ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 525924669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 5260dff11abeSMike Kravetz 5261e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 5262569f48b8SHillf Danton address = start; 5263569f48b8SHillf Danton for (; address < end; address += sz) { 52649c67a207SPeter Xu ptep = hugetlb_walk(vma, address, sz); 5265e95a9851SMike Kravetz if (!ptep) { 5266e95a9851SMike Kravetz address |= last_addr_mask; 5267c7546f8fSDavid Gibson continue; 5268e95a9851SMike Kravetz } 5269c7546f8fSDavid Gibson 5270cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 52714ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, address, ptep)) { 527231d49da5SAneesh Kumar K.V spin_unlock(ptl); 5273a4a118f2SNadav Amit tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE); 5274a4a118f2SNadav Amit force_flush = true; 52754ddb4d91SMike Kravetz address |= last_addr_mask; 527631d49da5SAneesh Kumar K.V continue; 527731d49da5SAneesh Kumar K.V } 527839dde65cSChen, Kenneth W 52796629326bSHillf Danton pte = huge_ptep_get(ptep); 528031d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 528131d49da5SAneesh Kumar K.V spin_unlock(ptl); 528231d49da5SAneesh Kumar K.V continue; 528331d49da5SAneesh Kumar K.V } 52846629326bSHillf Danton 52856629326bSHillf Danton /* 52869fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 52879fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 52886629326bSHillf Danton */ 52899fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 529005e90bd0SPeter Xu /* 529105e90bd0SPeter Xu * If the pte was wr-protected by uffd-wp in any of the 529205e90bd0SPeter Xu * swap forms, meanwhile the caller does not want to 529305e90bd0SPeter Xu * drop the uffd-wp bit in this zap, then replace the 529405e90bd0SPeter Xu * pte with a marker. 529505e90bd0SPeter Xu */ 529605e90bd0SPeter Xu if (pte_swp_uffd_wp_any(pte) && 529705e90bd0SPeter Xu !(zap_flags & ZAP_FLAG_DROP_MARKER)) 529805e90bd0SPeter Xu set_huge_pte_at(mm, address, ptep, 529905e90bd0SPeter Xu make_pte_marker(PTE_MARKER_UFFD_WP)); 530005e90bd0SPeter Xu else 53019386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 530231d49da5SAneesh Kumar K.V spin_unlock(ptl); 530331d49da5SAneesh Kumar K.V continue; 53048c4894c6SNaoya Horiguchi } 53056629326bSHillf Danton 53066629326bSHillf Danton page = pte_page(pte); 530704f2cbe3SMel Gorman /* 530804f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 530904f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 531004f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 531104f2cbe3SMel Gorman */ 531204f2cbe3SMel Gorman if (ref_page) { 531331d49da5SAneesh Kumar K.V if (page != ref_page) { 531431d49da5SAneesh Kumar K.V spin_unlock(ptl); 531531d49da5SAneesh Kumar K.V continue; 531631d49da5SAneesh Kumar K.V } 531704f2cbe3SMel Gorman /* 531804f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 531904f2cbe3SMel Gorman * future faults in this VMA will fail rather than 532004f2cbe3SMel Gorman * looking like data was lost 532104f2cbe3SMel Gorman */ 532204f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 532304f2cbe3SMel Gorman } 532404f2cbe3SMel Gorman 5325c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 5326b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 5327106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 53286649a386SKen Chen set_page_dirty(page); 532905e90bd0SPeter Xu /* Leave a uffd-wp pte marker if needed */ 533005e90bd0SPeter Xu if (huge_pte_uffd_wp(pte) && 533105e90bd0SPeter Xu !(zap_flags & ZAP_FLAG_DROP_MARKER)) 533205e90bd0SPeter Xu set_huge_pte_at(mm, address, ptep, 533305e90bd0SPeter Xu make_pte_marker(PTE_MARKER_UFFD_WP)); 53345d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 5335cea86fe2SHugh Dickins page_remove_rmap(page, vma, true); 533631d49da5SAneesh Kumar K.V 5337cb900f41SKirill A. Shutemov spin_unlock(ptl); 5338e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 533924669e58SAneesh Kumar K.V /* 534031d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 534124669e58SAneesh Kumar K.V */ 534231d49da5SAneesh Kumar K.V if (ref_page) 534331d49da5SAneesh Kumar K.V break; 5344fe1668aeSChen, Kenneth W } 534524669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 5346a4a118f2SNadav Amit 5347a4a118f2SNadav Amit /* 5348a4a118f2SNadav Amit * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We 5349a4a118f2SNadav Amit * could defer the flush until now, since by holding i_mmap_rwsem we 5350a4a118f2SNadav Amit * guaranteed that the last refernece would not be dropped. But we must 5351a4a118f2SNadav Amit * do the flushing before we return, as otherwise i_mmap_rwsem will be 5352a4a118f2SNadav Amit * dropped and the last reference to the shared PMDs page might be 5353a4a118f2SNadav Amit * dropped as well. 5354a4a118f2SNadav Amit * 5355a4a118f2SNadav Amit * In theory we could defer the freeing of the PMD pages as well, but 5356a4a118f2SNadav Amit * huge_pmd_unshare() relies on the exact page_count for the PMD page to 5357a4a118f2SNadav Amit * detect sharing, so we cannot defer the release of the page either. 5358a4a118f2SNadav Amit * Instead, do flush now. 5359a4a118f2SNadav Amit */ 5360a4a118f2SNadav Amit if (force_flush) 5361a4a118f2SNadav Amit tlb_flush_mmu_tlbonly(tlb); 53621da177e4SLinus Torvalds } 536363551ae0SDavid Gibson 5364d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 5365d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 536605e90bd0SPeter Xu unsigned long end, struct page *ref_page, 536705e90bd0SPeter Xu zap_flags_t zap_flags) 5368d833352aSMel Gorman { 5369131a79b4SMike Kravetz hugetlb_vma_lock_write(vma); 5370131a79b4SMike Kravetz i_mmap_lock_write(vma->vm_file->f_mapping); 5371131a79b4SMike Kravetz 5372369258ceSMike Kravetz /* mmu notification performed in caller */ 537305e90bd0SPeter Xu __unmap_hugepage_range(tlb, vma, start, end, ref_page, zap_flags); 5374d833352aSMel Gorman 537504ada095SMike Kravetz if (zap_flags & ZAP_FLAG_UNMAP) { /* final unmap */ 5376d833352aSMel Gorman /* 537704ada095SMike Kravetz * Unlock and free the vma lock before releasing i_mmap_rwsem. 537804ada095SMike Kravetz * When the vma_lock is freed, this makes the vma ineligible 537904ada095SMike Kravetz * for pmd sharing. And, i_mmap_rwsem is required to set up 538004ada095SMike Kravetz * pmd sharing. This is important as page tables for this 538104ada095SMike Kravetz * unmapped range will be asynchrously deleted. If the page 538204ada095SMike Kravetz * tables are shared, there will be issues when accessed by 538304ada095SMike Kravetz * someone else. 5384d833352aSMel Gorman */ 5385ecfbd733SMike Kravetz __hugetlb_vma_unlock_write_free(vma); 5386131a79b4SMike Kravetz i_mmap_unlock_write(vma->vm_file->f_mapping); 538704ada095SMike Kravetz } else { 538804ada095SMike Kravetz i_mmap_unlock_write(vma->vm_file->f_mapping); 538904ada095SMike Kravetz hugetlb_vma_unlock_write(vma); 539004ada095SMike Kravetz } 5391d833352aSMel Gorman } 5392d833352aSMel Gorman 5393502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 539405e90bd0SPeter Xu unsigned long end, struct page *ref_page, 539505e90bd0SPeter Xu zap_flags_t zap_flags) 5396502717f4SChen, Kenneth W { 5397369258ceSMike Kravetz struct mmu_notifier_range range; 539824669e58SAneesh Kumar K.V struct mmu_gather tlb; 5399dff11abeSMike Kravetz 54007d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm, 5401369258ceSMike Kravetz start, end); 5402369258ceSMike Kravetz adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 5403369258ceSMike Kravetz mmu_notifier_invalidate_range_start(&range); 5404a72afd87SWill Deacon tlb_gather_mmu(&tlb, vma->vm_mm); 5405369258ceSMike Kravetz 540605e90bd0SPeter Xu __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags); 5407369258ceSMike Kravetz 5408369258ceSMike Kravetz mmu_notifier_invalidate_range_end(&range); 5409ae8eba8bSWill Deacon tlb_finish_mmu(&tlb); 5410502717f4SChen, Kenneth W } 5411502717f4SChen, Kenneth W 541204f2cbe3SMel Gorman /* 541304f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 5414578b7725SZhiyuan Dai * mapping it owns the reserve page for. The intention is to unmap the page 541504f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 541604f2cbe3SMel Gorman * same region. 541704f2cbe3SMel Gorman */ 54182f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 54192a4b3dedSHarvey Harrison struct page *page, unsigned long address) 542004f2cbe3SMel Gorman { 54217526674dSAdam Litke struct hstate *h = hstate_vma(vma); 542204f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 542304f2cbe3SMel Gorman struct address_space *mapping; 542404f2cbe3SMel Gorman pgoff_t pgoff; 542504f2cbe3SMel Gorman 542604f2cbe3SMel Gorman /* 542704f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 542804f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 542904f2cbe3SMel Gorman */ 54307526674dSAdam Litke address = address & huge_page_mask(h); 543136e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 543236e4f20aSMichal Hocko vma->vm_pgoff; 543393c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 543404f2cbe3SMel Gorman 54354eb2b1dcSMel Gorman /* 54364eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 54374eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 54384eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 54394eb2b1dcSMel Gorman */ 544083cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 54416b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 544204f2cbe3SMel Gorman /* Do not unmap the current VMA */ 544304f2cbe3SMel Gorman if (iter_vma == vma) 544404f2cbe3SMel Gorman continue; 544504f2cbe3SMel Gorman 544604f2cbe3SMel Gorman /* 54472f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 54482f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 54492f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 54502f84a899SMel Gorman */ 54512f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 54522f84a899SMel Gorman continue; 54532f84a899SMel Gorman 54542f84a899SMel Gorman /* 545504f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 545604f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 545704f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 545804f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 545904f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 546004f2cbe3SMel Gorman */ 546104f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 546224669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 546305e90bd0SPeter Xu address + huge_page_size(h), page, 0); 546404f2cbe3SMel Gorman } 546583cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 546604f2cbe3SMel Gorman } 546704f2cbe3SMel Gorman 54680fe6e20bSNaoya Horiguchi /* 5469c89357e2SDavid Hildenbrand * hugetlb_wp() should be called with page lock of the original hugepage held. 5470aa6d2e8cSBaolin Wang * Called with hugetlb_fault_mutex_table held and pte_page locked so we 5471ef009b25SMichal Hocko * cannot race with other handlers or page migration. 5472ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 54730fe6e20bSNaoya Horiguchi */ 5474c89357e2SDavid Hildenbrand static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma, 5475c89357e2SDavid Hildenbrand unsigned long address, pte_t *ptep, unsigned int flags, 5476cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 54771e8f889bSDavid Gibson { 5478c89357e2SDavid Hildenbrand const bool unshare = flags & FAULT_FLAG_UNSHARE; 54793999f52eSAneesh Kumar K.V pte_t pte; 5480a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 54811e8f889bSDavid Gibson struct page *old_page, *new_page; 54822b740303SSouptick Joarder int outside_reserve = 0; 54832b740303SSouptick Joarder vm_fault_t ret = 0; 5484974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 5485ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 54861e8f889bSDavid Gibson 54871d8d1464SDavid Hildenbrand /* 54881d8d1464SDavid Hildenbrand * hugetlb does not support FOLL_FORCE-style write faults that keep the 54891d8d1464SDavid Hildenbrand * PTE mapped R/O such as maybe_mkwrite() would do. 54901d8d1464SDavid Hildenbrand */ 54911d8d1464SDavid Hildenbrand if (WARN_ON_ONCE(!unshare && !(vma->vm_flags & VM_WRITE))) 54921d8d1464SDavid Hildenbrand return VM_FAULT_SIGSEGV; 54931d8d1464SDavid Hildenbrand 54941d8d1464SDavid Hildenbrand /* Let's take out MAP_SHARED mappings first. */ 54951d8d1464SDavid Hildenbrand if (vma->vm_flags & VM_MAYSHARE) { 54961d8d1464SDavid Hildenbrand set_huge_ptep_writable(vma, haddr, ptep); 54971d8d1464SDavid Hildenbrand return 0; 54981d8d1464SDavid Hildenbrand } 54991d8d1464SDavid Hildenbrand 55003999f52eSAneesh Kumar K.V pte = huge_ptep_get(ptep); 55011e8f889bSDavid Gibson old_page = pte_page(pte); 55021e8f889bSDavid Gibson 5503662ce1dcSYang Yang delayacct_wpcopy_start(); 5504662ce1dcSYang Yang 550504f2cbe3SMel Gorman retry_avoidcopy: 5506c89357e2SDavid Hildenbrand /* 5507c89357e2SDavid Hildenbrand * If no-one else is actually using this page, we're the exclusive 5508c89357e2SDavid Hildenbrand * owner and can reuse this page. 5509c89357e2SDavid Hildenbrand */ 551037a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 5511c89357e2SDavid Hildenbrand if (!PageAnonExclusive(old_page)) 55125a49973dSHugh Dickins page_move_anon_rmap(old_page, vma); 5513c89357e2SDavid Hildenbrand if (likely(!unshare)) 55145b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 5515662ce1dcSYang Yang 5516662ce1dcSYang Yang delayacct_wpcopy_end(); 551783c54070SNick Piggin return 0; 55181e8f889bSDavid Gibson } 55196c287605SDavid Hildenbrand VM_BUG_ON_PAGE(PageAnon(old_page) && PageAnonExclusive(old_page), 55206c287605SDavid Hildenbrand old_page); 55211e8f889bSDavid Gibson 552204f2cbe3SMel Gorman /* 552304f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 552404f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 552504f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 552604f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 552704f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 552804f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 552904f2cbe3SMel Gorman * of the full address range. 553004f2cbe3SMel Gorman */ 55315944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 553204f2cbe3SMel Gorman old_page != pagecache_page) 553304f2cbe3SMel Gorman outside_reserve = 1; 553404f2cbe3SMel Gorman 553509cbfeafSKirill A. Shutemov get_page(old_page); 5536b76c8cfbSLarry Woodman 5537ad4404a2SDavidlohr Bueso /* 5538ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 5539ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 5540ad4404a2SDavidlohr Bueso */ 5541cb900f41SKirill A. Shutemov spin_unlock(ptl); 55425b7a1d40SHuang Ying new_page = alloc_huge_page(vma, haddr, outside_reserve); 55431e8f889bSDavid Gibson 55442fc39cecSAdam Litke if (IS_ERR(new_page)) { 554504f2cbe3SMel Gorman /* 554604f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 554704f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 554804f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 554904f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 555004f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 555104f2cbe3SMel Gorman */ 555204f2cbe3SMel Gorman if (outside_reserve) { 555340549ba8SMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 555440549ba8SMike Kravetz pgoff_t idx; 555540549ba8SMike Kravetz u32 hash; 555640549ba8SMike Kravetz 555709cbfeafSKirill A. Shutemov put_page(old_page); 555840549ba8SMike Kravetz /* 555940549ba8SMike Kravetz * Drop hugetlb_fault_mutex and vma_lock before 556040549ba8SMike Kravetz * unmapping. unmapping needs to hold vma_lock 556140549ba8SMike Kravetz * in write mode. Dropping vma_lock in read mode 556240549ba8SMike Kravetz * here is OK as COW mappings do not interact with 556340549ba8SMike Kravetz * PMD sharing. 556440549ba8SMike Kravetz * 556540549ba8SMike Kravetz * Reacquire both after unmap operation. 556640549ba8SMike Kravetz */ 556740549ba8SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 556840549ba8SMike Kravetz hash = hugetlb_fault_mutex_hash(mapping, idx); 556940549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 557040549ba8SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 557140549ba8SMike Kravetz 55725b7a1d40SHuang Ying unmap_ref_private(mm, vma, old_page, haddr); 557340549ba8SMike Kravetz 557440549ba8SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 557540549ba8SMike Kravetz hugetlb_vma_lock_read(vma); 5576cb900f41SKirill A. Shutemov spin_lock(ptl); 55779c67a207SPeter Xu ptep = hugetlb_walk(vma, haddr, huge_page_size(h)); 5578a9af0c5dSNaoya Horiguchi if (likely(ptep && 5579a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 558004f2cbe3SMel Gorman goto retry_avoidcopy; 5581a734bcc8SHillf Danton /* 5582cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 5583cb900f41SKirill A. Shutemov * lock, and our job is done. 5584a734bcc8SHillf Danton */ 5585662ce1dcSYang Yang delayacct_wpcopy_end(); 5586a734bcc8SHillf Danton return 0; 558704f2cbe3SMel Gorman } 558804f2cbe3SMel Gorman 55892b740303SSouptick Joarder ret = vmf_error(PTR_ERR(new_page)); 5590ad4404a2SDavidlohr Bueso goto out_release_old; 55911e8f889bSDavid Gibson } 55921e8f889bSDavid Gibson 55930fe6e20bSNaoya Horiguchi /* 55940fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 55950fe6e20bSNaoya Horiguchi * anon_vma prepared. 55960fe6e20bSNaoya Horiguchi */ 559744e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 5598ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 5599ad4404a2SDavidlohr Bueso goto out_release_all; 560044e2aa93SDean Nelson } 56010fe6e20bSNaoya Horiguchi 5602974e6d66SHuang Ying copy_user_huge_page(new_page, old_page, address, vma, 560347ad8475SAndrea Arcangeli pages_per_huge_page(h)); 56040ed361deSNick Piggin __SetPageUptodate(new_page); 56051e8f889bSDavid Gibson 56067d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, haddr, 56076f4f13e8SJérôme Glisse haddr + huge_page_size(h)); 5608ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 5609ad4404a2SDavidlohr Bueso 5610b76c8cfbSLarry Woodman /* 5611cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 5612b76c8cfbSLarry Woodman * before the page tables are altered 5613b76c8cfbSLarry Woodman */ 5614cb900f41SKirill A. Shutemov spin_lock(ptl); 56159c67a207SPeter Xu ptep = hugetlb_walk(vma, haddr, huge_page_size(h)); 5616a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 5617c89357e2SDavid Hildenbrand /* Break COW or unshare */ 56185b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 5619ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, range.end); 5620cea86fe2SHugh Dickins page_remove_rmap(old_page, vma, true); 56215b7a1d40SHuang Ying hugepage_add_new_anon_rmap(new_page, vma, haddr); 56221eba86c0SPasha Tatashin set_huge_pte_at(mm, haddr, ptep, 5623c89357e2SDavid Hildenbrand make_huge_pte(vma, new_page, !unshare)); 56248f251a3dSMike Kravetz SetHPageMigratable(new_page); 56251e8f889bSDavid Gibson /* Make the old page be freed below */ 56261e8f889bSDavid Gibson new_page = old_page; 56271e8f889bSDavid Gibson } 5628cb900f41SKirill A. Shutemov spin_unlock(ptl); 5629ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 5630ad4404a2SDavidlohr Bueso out_release_all: 5631c89357e2SDavid Hildenbrand /* 5632c89357e2SDavid Hildenbrand * No restore in case of successful pagetable update (Break COW or 5633c89357e2SDavid Hildenbrand * unshare) 5634c89357e2SDavid Hildenbrand */ 5635c7b1850dSMike Kravetz if (new_page != old_page) 56365b7a1d40SHuang Ying restore_reserve_on_error(h, vma, haddr, new_page); 563709cbfeafSKirill A. Shutemov put_page(new_page); 5638ad4404a2SDavidlohr Bueso out_release_old: 563909cbfeafSKirill A. Shutemov put_page(old_page); 56408312034fSJoonsoo Kim 5641ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 5642662ce1dcSYang Yang 5643662ce1dcSYang Yang delayacct_wpcopy_end(); 5644ad4404a2SDavidlohr Bueso return ret; 56451e8f889bSDavid Gibson } 56461e8f889bSDavid Gibson 56473ae77f43SHugh Dickins /* 56483ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 56493ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 56503ae77f43SHugh Dickins */ 56513ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 56522a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 56532a15efc9SHugh Dickins { 56542a15efc9SHugh Dickins struct address_space *mapping; 56552a15efc9SHugh Dickins pgoff_t idx; 56562a15efc9SHugh Dickins struct page *page; 56572a15efc9SHugh Dickins 56582a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 56592a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 56602a15efc9SHugh Dickins 56612a15efc9SHugh Dickins page = find_get_page(mapping, idx); 56622a15efc9SHugh Dickins if (page) 56632a15efc9SHugh Dickins put_page(page); 56642a15efc9SHugh Dickins return page != NULL; 56652a15efc9SHugh Dickins } 56662a15efc9SHugh Dickins 56677e1813d4SMike Kravetz int hugetlb_add_to_page_cache(struct page *page, struct address_space *mapping, 5668ab76ad54SMike Kravetz pgoff_t idx) 5669ab76ad54SMike Kravetz { 5670d9ef44deSMatthew Wilcox (Oracle) struct folio *folio = page_folio(page); 5671ab76ad54SMike Kravetz struct inode *inode = mapping->host; 5672ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 5673d9ef44deSMatthew Wilcox (Oracle) int err; 5674ab76ad54SMike Kravetz 5675d9ef44deSMatthew Wilcox (Oracle) __folio_set_locked(folio); 5676d9ef44deSMatthew Wilcox (Oracle) err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL); 5677d9ef44deSMatthew Wilcox (Oracle) 5678d9ef44deSMatthew Wilcox (Oracle) if (unlikely(err)) { 5679d9ef44deSMatthew Wilcox (Oracle) __folio_clear_locked(folio); 5680ab76ad54SMike Kravetz return err; 5681d9ef44deSMatthew Wilcox (Oracle) } 5682d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 5683ab76ad54SMike Kravetz 568422146c3cSMike Kravetz /* 5685d9ef44deSMatthew Wilcox (Oracle) * mark folio dirty so that it will not be removed from cache/file 568622146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 568722146c3cSMike Kravetz */ 5688d9ef44deSMatthew Wilcox (Oracle) folio_mark_dirty(folio); 568922146c3cSMike Kravetz 5690ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 5691ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 5692ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 5693ab76ad54SMike Kravetz return 0; 5694ab76ad54SMike Kravetz } 5695ab76ad54SMike Kravetz 56967677f7fdSAxel Rasmussen static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma, 56977677f7fdSAxel Rasmussen struct address_space *mapping, 56987677f7fdSAxel Rasmussen pgoff_t idx, 56997677f7fdSAxel Rasmussen unsigned int flags, 57007677f7fdSAxel Rasmussen unsigned long haddr, 5701824ddc60SNadav Amit unsigned long addr, 57027677f7fdSAxel Rasmussen unsigned long reason) 57037677f7fdSAxel Rasmussen { 57047677f7fdSAxel Rasmussen u32 hash; 57057677f7fdSAxel Rasmussen struct vm_fault vmf = { 57067677f7fdSAxel Rasmussen .vma = vma, 57077677f7fdSAxel Rasmussen .address = haddr, 5708824ddc60SNadav Amit .real_address = addr, 57097677f7fdSAxel Rasmussen .flags = flags, 57107677f7fdSAxel Rasmussen 57117677f7fdSAxel Rasmussen /* 57127677f7fdSAxel Rasmussen * Hard to debug if it ends up being 57137677f7fdSAxel Rasmussen * used by a callee that assumes 57147677f7fdSAxel Rasmussen * something about the other 57157677f7fdSAxel Rasmussen * uninitialized fields... same as in 57167677f7fdSAxel Rasmussen * memory.c 57177677f7fdSAxel Rasmussen */ 57187677f7fdSAxel Rasmussen }; 57197677f7fdSAxel Rasmussen 57207677f7fdSAxel Rasmussen /* 5721958f32ceSLiu Shixin * vma_lock and hugetlb_fault_mutex must be dropped before handling 5722958f32ceSLiu Shixin * userfault. Also mmap_lock could be dropped due to handling 5723958f32ceSLiu Shixin * userfault, any vma operation should be careful from here. 57247677f7fdSAxel Rasmussen */ 572540549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 57267677f7fdSAxel Rasmussen hash = hugetlb_fault_mutex_hash(mapping, idx); 57277677f7fdSAxel Rasmussen mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5728958f32ceSLiu Shixin return handle_userfault(&vmf, reason); 57297677f7fdSAxel Rasmussen } 57307677f7fdSAxel Rasmussen 57312ea7ff1eSPeter Xu /* 57322ea7ff1eSPeter Xu * Recheck pte with pgtable lock. Returns true if pte didn't change, or 57332ea7ff1eSPeter Xu * false if pte changed or is changing. 57342ea7ff1eSPeter Xu */ 57352ea7ff1eSPeter Xu static bool hugetlb_pte_stable(struct hstate *h, struct mm_struct *mm, 57362ea7ff1eSPeter Xu pte_t *ptep, pte_t old_pte) 57372ea7ff1eSPeter Xu { 57382ea7ff1eSPeter Xu spinlock_t *ptl; 57392ea7ff1eSPeter Xu bool same; 57402ea7ff1eSPeter Xu 57412ea7ff1eSPeter Xu ptl = huge_pte_lock(h, mm, ptep); 57422ea7ff1eSPeter Xu same = pte_same(huge_ptep_get(ptep), old_pte); 57432ea7ff1eSPeter Xu spin_unlock(ptl); 57442ea7ff1eSPeter Xu 57452ea7ff1eSPeter Xu return same; 57462ea7ff1eSPeter Xu } 57472ea7ff1eSPeter Xu 57482b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 57492b740303SSouptick Joarder struct vm_area_struct *vma, 57508382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 5751c64e912cSPeter Xu unsigned long address, pte_t *ptep, 5752c64e912cSPeter Xu pte_t old_pte, unsigned int flags) 5753ac9b9c66SHugh Dickins { 5754a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 57552b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 5756409eb8c2SHillf Danton int anon_rmap = 0; 57574c887265SAdam Litke unsigned long size; 57584c887265SAdam Litke struct page *page; 57591e8f889bSDavid Gibson pte_t new_pte; 5760cb900f41SKirill A. Shutemov spinlock_t *ptl; 5761285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 5762c7b1850dSMike Kravetz bool new_page, new_pagecache_page = false; 5763958f32ceSLiu Shixin u32 hash = hugetlb_fault_mutex_hash(mapping, idx); 57644c887265SAdam Litke 576504f2cbe3SMel Gorman /* 576604f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 576704f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 5768c89357e2SDavid Hildenbrand * COW/unsharing. Warn that such a situation has occurred as it may not 5769c89357e2SDavid Hildenbrand * be obvious. 577004f2cbe3SMel Gorman */ 577104f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 5772910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 577304f2cbe3SMel Gorman current->pid); 5774958f32ceSLiu Shixin goto out; 577504f2cbe3SMel Gorman } 577604f2cbe3SMel Gorman 57774c887265SAdam Litke /* 5778188a3972SMike Kravetz * Use page lock to guard against racing truncation 5779188a3972SMike Kravetz * before we get page_table_lock. 57804c887265SAdam Litke */ 5781c7b1850dSMike Kravetz new_page = false; 578287bf91d3SMike Kravetz page = find_lock_page(mapping, idx); 578387bf91d3SMike Kravetz if (!page) { 5784188a3972SMike Kravetz size = i_size_read(mapping->host) >> huge_page_shift(h); 5785188a3972SMike Kravetz if (idx >= size) 5786188a3972SMike Kravetz goto out; 57877677f7fdSAxel Rasmussen /* Check for page in userfault range */ 57882ea7ff1eSPeter Xu if (userfaultfd_missing(vma)) { 57892ea7ff1eSPeter Xu /* 57902ea7ff1eSPeter Xu * Since hugetlb_no_page() was examining pte 57912ea7ff1eSPeter Xu * without pgtable lock, we need to re-test under 57922ea7ff1eSPeter Xu * lock because the pte may not be stable and could 57932ea7ff1eSPeter Xu * have changed from under us. Try to detect 57942ea7ff1eSPeter Xu * either changed or during-changing ptes and retry 57952ea7ff1eSPeter Xu * properly when needed. 57962ea7ff1eSPeter Xu * 57972ea7ff1eSPeter Xu * Note that userfaultfd is actually fine with 57982ea7ff1eSPeter Xu * false positives (e.g. caused by pte changed), 57992ea7ff1eSPeter Xu * but not wrong logical events (e.g. caused by 58002ea7ff1eSPeter Xu * reading a pte during changing). The latter can 58012ea7ff1eSPeter Xu * confuse the userspace, so the strictness is very 58022ea7ff1eSPeter Xu * much preferred. E.g., MISSING event should 58032ea7ff1eSPeter Xu * never happen on the page after UFFDIO_COPY has 58042ea7ff1eSPeter Xu * correctly installed the page and returned. 58052ea7ff1eSPeter Xu */ 58062ea7ff1eSPeter Xu if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) { 58072ea7ff1eSPeter Xu ret = 0; 58082ea7ff1eSPeter Xu goto out; 58092ea7ff1eSPeter Xu } 58102ea7ff1eSPeter Xu 58112ea7ff1eSPeter Xu return hugetlb_handle_userfault(vma, mapping, idx, flags, 58122ea7ff1eSPeter Xu haddr, address, 58137677f7fdSAxel Rasmussen VM_UFFD_MISSING); 58142ea7ff1eSPeter Xu } 58151a1aad8aSMike Kravetz 5816285b8dcaSHuang Ying page = alloc_huge_page(vma, haddr, 0); 58172fc39cecSAdam Litke if (IS_ERR(page)) { 58184643d67eSMike Kravetz /* 58194643d67eSMike Kravetz * Returning error will result in faulting task being 58204643d67eSMike Kravetz * sent SIGBUS. The hugetlb fault mutex prevents two 58214643d67eSMike Kravetz * tasks from racing to fault in the same page which 58224643d67eSMike Kravetz * could result in false unable to allocate errors. 58234643d67eSMike Kravetz * Page migration does not take the fault mutex, but 58244643d67eSMike Kravetz * does a clear then write of pte's under page table 58254643d67eSMike Kravetz * lock. Page fault code could race with migration, 58264643d67eSMike Kravetz * notice the clear pte and try to allocate a page 58274643d67eSMike Kravetz * here. Before returning error, get ptl and make 58284643d67eSMike Kravetz * sure there really is no pte entry. 58294643d67eSMike Kravetz */ 5830f9bf6c03SPeter Xu if (hugetlb_pte_stable(h, mm, ptep, old_pte)) 58312b740303SSouptick Joarder ret = vmf_error(PTR_ERR(page)); 5832f9bf6c03SPeter Xu else 5833f9bf6c03SPeter Xu ret = 0; 58346bda666aSChristoph Lameter goto out; 58356bda666aSChristoph Lameter } 583647ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 58370ed361deSNick Piggin __SetPageUptodate(page); 5838cb6acd01SMike Kravetz new_page = true; 5839ac9b9c66SHugh Dickins 5840f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 58417e1813d4SMike Kravetz int err = hugetlb_add_to_page_cache(page, mapping, idx); 58426bda666aSChristoph Lameter if (err) { 58433a5497a2SMiaohe Lin /* 58443a5497a2SMiaohe Lin * err can't be -EEXIST which implies someone 58453a5497a2SMiaohe Lin * else consumed the reservation since hugetlb 58463a5497a2SMiaohe Lin * fault mutex is held when add a hugetlb page 58473a5497a2SMiaohe Lin * to the page cache. So it's safe to call 58483a5497a2SMiaohe Lin * restore_reserve_on_error() here. 58493a5497a2SMiaohe Lin */ 58503a5497a2SMiaohe Lin restore_reserve_on_error(h, vma, haddr, page); 58516bda666aSChristoph Lameter put_page(page); 58526bda666aSChristoph Lameter goto out; 58536bda666aSChristoph Lameter } 5854c7b1850dSMike Kravetz new_pagecache_page = true; 585523be7468SMel Gorman } else { 58566bda666aSChristoph Lameter lock_page(page); 58570fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 58580fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 58590fe6e20bSNaoya Horiguchi goto backout_unlocked; 586023be7468SMel Gorman } 5861409eb8c2SHillf Danton anon_rmap = 1; 58620fe6e20bSNaoya Horiguchi } 58630fe6e20bSNaoya Horiguchi } else { 586457303d80SAndy Whitcroft /* 5865998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 5866998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 5867998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 5868fd6a03edSNaoya Horiguchi */ 5869fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 58700eb98f15SMiaohe Lin ret = VM_FAULT_HWPOISON_LARGE | 5871972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 5872fd6a03edSNaoya Horiguchi goto backout_unlocked; 58736bda666aSChristoph Lameter } 58747677f7fdSAxel Rasmussen 58757677f7fdSAxel Rasmussen /* Check for page in userfault range. */ 58767677f7fdSAxel Rasmussen if (userfaultfd_minor(vma)) { 58777677f7fdSAxel Rasmussen unlock_page(page); 58787677f7fdSAxel Rasmussen put_page(page); 58792ea7ff1eSPeter Xu /* See comment in userfaultfd_missing() block above */ 58802ea7ff1eSPeter Xu if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) { 58812ea7ff1eSPeter Xu ret = 0; 58822ea7ff1eSPeter Xu goto out; 58832ea7ff1eSPeter Xu } 58842ea7ff1eSPeter Xu return hugetlb_handle_userfault(vma, mapping, idx, flags, 58852ea7ff1eSPeter Xu haddr, address, 58867677f7fdSAxel Rasmussen VM_UFFD_MINOR); 58877677f7fdSAxel Rasmussen } 5888998b4382SNaoya Horiguchi } 58891e8f889bSDavid Gibson 589057303d80SAndy Whitcroft /* 589157303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 589257303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 589357303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 589457303d80SAndy Whitcroft * the spinlock. 589557303d80SAndy Whitcroft */ 58965e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 5897285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 58982b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 58992b26736cSAndy Whitcroft goto backout_unlocked; 59002b26736cSAndy Whitcroft } 59015e911373SMike Kravetz /* Just decrements count, does not deallocate */ 5902285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 59035e911373SMike Kravetz } 590457303d80SAndy Whitcroft 59058bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 590683c54070SNick Piggin ret = 0; 5907c64e912cSPeter Xu /* If pte changed from under us, retry */ 5908c64e912cSPeter Xu if (!pte_same(huge_ptep_get(ptep), old_pte)) 59094c887265SAdam Litke goto backout; 59104c887265SAdam Litke 59114781593dSPeter Xu if (anon_rmap) 5912285b8dcaSHuang Ying hugepage_add_new_anon_rmap(page, vma, haddr); 59134781593dSPeter Xu else 5914fb3d824dSDavid Hildenbrand page_dup_file_rmap(page, true); 59151e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 59161e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 5917c64e912cSPeter Xu /* 5918c64e912cSPeter Xu * If this pte was previously wr-protected, keep it wr-protected even 5919c64e912cSPeter Xu * if populated. 5920c64e912cSPeter Xu */ 5921c64e912cSPeter Xu if (unlikely(pte_marker_uffd_wp(old_pte))) 5922f1eb1bacSPeter Xu new_pte = huge_pte_mkuffd_wp(new_pte); 5923285b8dcaSHuang Ying set_huge_pte_at(mm, haddr, ptep, new_pte); 59241e8f889bSDavid Gibson 59255d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 5926788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 59271e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 5928c89357e2SDavid Hildenbrand ret = hugetlb_wp(mm, vma, address, ptep, flags, page, ptl); 59291e8f889bSDavid Gibson } 59301e8f889bSDavid Gibson 5931cb900f41SKirill A. Shutemov spin_unlock(ptl); 5932cb6acd01SMike Kravetz 5933cb6acd01SMike Kravetz /* 59348f251a3dSMike Kravetz * Only set HPageMigratable in newly allocated pages. Existing pages 59358f251a3dSMike Kravetz * found in the pagecache may not have HPageMigratableset if they have 59368f251a3dSMike Kravetz * been isolated for migration. 5937cb6acd01SMike Kravetz */ 5938cb6acd01SMike Kravetz if (new_page) 59398f251a3dSMike Kravetz SetHPageMigratable(page); 5940cb6acd01SMike Kravetz 59414c887265SAdam Litke unlock_page(page); 59424c887265SAdam Litke out: 5943958f32ceSLiu Shixin hugetlb_vma_unlock_read(vma); 5944958f32ceSLiu Shixin mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5945ac9b9c66SHugh Dickins return ret; 59464c887265SAdam Litke 59474c887265SAdam Litke backout: 5948cb900f41SKirill A. Shutemov spin_unlock(ptl); 59492b26736cSAndy Whitcroft backout_unlocked: 5950c7b1850dSMike Kravetz if (new_page && !new_pagecache_page) 5951285b8dcaSHuang Ying restore_reserve_on_error(h, vma, haddr, page); 5952fa27759aSMike Kravetz 5953fa27759aSMike Kravetz unlock_page(page); 59544c887265SAdam Litke put_page(page); 59554c887265SAdam Litke goto out; 5956ac9b9c66SHugh Dickins } 5957ac9b9c66SHugh Dickins 59588382d914SDavidlohr Bueso #ifdef CONFIG_SMP 5959188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 59608382d914SDavidlohr Bueso { 59618382d914SDavidlohr Bueso unsigned long key[2]; 59628382d914SDavidlohr Bueso u32 hash; 59638382d914SDavidlohr Bueso 59648382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 59658382d914SDavidlohr Bueso key[1] = idx; 59668382d914SDavidlohr Bueso 596755254636SMike Kravetz hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); 59688382d914SDavidlohr Bueso 59698382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 59708382d914SDavidlohr Bueso } 59718382d914SDavidlohr Bueso #else 59728382d914SDavidlohr Bueso /* 59736c26d310SMiaohe Lin * For uniprocessor systems we always use a single mutex, so just 59748382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 59758382d914SDavidlohr Bueso */ 5976188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 59778382d914SDavidlohr Bueso { 59788382d914SDavidlohr Bueso return 0; 59798382d914SDavidlohr Bueso } 59808382d914SDavidlohr Bueso #endif 59818382d914SDavidlohr Bueso 59822b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 5983788c7df4SHugh Dickins unsigned long address, unsigned int flags) 598486e5216fSAdam Litke { 59858382d914SDavidlohr Bueso pte_t *ptep, entry; 5986cb900f41SKirill A. Shutemov spinlock_t *ptl; 59872b740303SSouptick Joarder vm_fault_t ret; 59888382d914SDavidlohr Bueso u32 hash; 59898382d914SDavidlohr Bueso pgoff_t idx; 59900fe6e20bSNaoya Horiguchi struct page *page = NULL; 599157303d80SAndy Whitcroft struct page *pagecache_page = NULL; 5992a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 59938382d914SDavidlohr Bueso struct address_space *mapping; 59940f792cf9SNaoya Horiguchi int need_wait_lock = 0; 5995285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 599686e5216fSAdam Litke 59973935baa9SDavid Gibson /* 59983935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 59993935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 60003935baa9SDavid Gibson * the same page in the page cache. 60013935baa9SDavid Gibson */ 600240549ba8SMike Kravetz mapping = vma->vm_file->f_mapping; 600340549ba8SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 6004188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 6005c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 60068382d914SDavidlohr Bueso 600740549ba8SMike Kravetz /* 600840549ba8SMike Kravetz * Acquire vma lock before calling huge_pte_alloc and hold 600940549ba8SMike Kravetz * until finished with ptep. This prevents huge_pmd_unshare from 601040549ba8SMike Kravetz * being called elsewhere and making the ptep no longer valid. 601140549ba8SMike Kravetz */ 601240549ba8SMike Kravetz hugetlb_vma_lock_read(vma); 601340549ba8SMike Kravetz ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); 601440549ba8SMike Kravetz if (!ptep) { 601540549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 601640549ba8SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 601740549ba8SMike Kravetz return VM_FAULT_OOM; 601840549ba8SMike Kravetz } 601940549ba8SMike Kravetz 60207f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 6021c64e912cSPeter Xu /* PTE markers should be handled the same way as none pte */ 6022958f32ceSLiu Shixin if (huge_pte_none_mostly(entry)) 6023958f32ceSLiu Shixin /* 6024958f32ceSLiu Shixin * hugetlb_no_page will drop vma lock and hugetlb fault 6025958f32ceSLiu Shixin * mutex internally, which make us return immediately. 6026958f32ceSLiu Shixin */ 6027958f32ceSLiu Shixin return hugetlb_no_page(mm, vma, mapping, idx, address, ptep, 6028c64e912cSPeter Xu entry, flags); 602986e5216fSAdam Litke 603083c54070SNick Piggin ret = 0; 60311e8f889bSDavid Gibson 603257303d80SAndy Whitcroft /* 60330f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 60340f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 60357c8de358SEthon Paul * an active hugepage in pagecache. This goto expects the 2nd page 60367c8de358SEthon Paul * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will 60377c8de358SEthon Paul * properly handle it. 60380f792cf9SNaoya Horiguchi */ 6039fcd48540SPeter Xu if (!pte_present(entry)) { 6040fcd48540SPeter Xu if (unlikely(is_hugetlb_entry_migration(entry))) { 6041fcd48540SPeter Xu /* 6042fcd48540SPeter Xu * Release the hugetlb fault lock now, but retain 6043fcd48540SPeter Xu * the vma lock, because it is needed to guard the 6044fcd48540SPeter Xu * huge_pte_lockptr() later in 6045fcd48540SPeter Xu * migration_entry_wait_huge(). The vma lock will 6046fcd48540SPeter Xu * be released there. 6047fcd48540SPeter Xu */ 6048fcd48540SPeter Xu mutex_unlock(&hugetlb_fault_mutex_table[hash]); 6049fcd48540SPeter Xu migration_entry_wait_huge(vma, ptep); 6050fcd48540SPeter Xu return 0; 6051fcd48540SPeter Xu } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 6052fcd48540SPeter Xu ret = VM_FAULT_HWPOISON_LARGE | 6053fcd48540SPeter Xu VM_FAULT_SET_HINDEX(hstate_index(h)); 60540f792cf9SNaoya Horiguchi goto out_mutex; 6055fcd48540SPeter Xu } 60560f792cf9SNaoya Horiguchi 60570f792cf9SNaoya Horiguchi /* 6058c89357e2SDavid Hildenbrand * If we are going to COW/unshare the mapping later, we examine the 6059c89357e2SDavid Hildenbrand * pending reservations for this page now. This will ensure that any 606057303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 60611d8d1464SDavid Hildenbrand * spinlock. Also lookup the pagecache page now as it is used to 60621d8d1464SDavid Hildenbrand * determine if a reservation has been consumed. 606357303d80SAndy Whitcroft */ 6064c89357e2SDavid Hildenbrand if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) && 60651d8d1464SDavid Hildenbrand !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(entry)) { 6066285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 60672b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 6068b4d1d99fSDavid Gibson goto out_mutex; 60692b26736cSAndy Whitcroft } 60705e911373SMike Kravetz /* Just decrements count, does not deallocate */ 6071285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 607257303d80SAndy Whitcroft 607329be8426SMiaohe Lin pagecache_page = find_lock_page(mapping, idx); 607457303d80SAndy Whitcroft } 607557303d80SAndy Whitcroft 60760f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 60770fe6e20bSNaoya Horiguchi 6078c89357e2SDavid Hildenbrand /* Check for a racing update before calling hugetlb_wp() */ 6079b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 6080cb900f41SKirill A. Shutemov goto out_ptl; 6081b4d1d99fSDavid Gibson 6082166f3eccSPeter Xu /* Handle userfault-wp first, before trying to lock more pages */ 6083166f3eccSPeter Xu if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) && 6084166f3eccSPeter Xu (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 6085166f3eccSPeter Xu struct vm_fault vmf = { 6086166f3eccSPeter Xu .vma = vma, 6087166f3eccSPeter Xu .address = haddr, 6088166f3eccSPeter Xu .real_address = address, 6089166f3eccSPeter Xu .flags = flags, 6090166f3eccSPeter Xu }; 6091166f3eccSPeter Xu 6092166f3eccSPeter Xu spin_unlock(ptl); 6093166f3eccSPeter Xu if (pagecache_page) { 6094166f3eccSPeter Xu unlock_page(pagecache_page); 6095166f3eccSPeter Xu put_page(pagecache_page); 6096166f3eccSPeter Xu } 609740549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 6098166f3eccSPeter Xu mutex_unlock(&hugetlb_fault_mutex_table[hash]); 6099166f3eccSPeter Xu return handle_userfault(&vmf, VM_UFFD_WP); 6100166f3eccSPeter Xu } 6101166f3eccSPeter Xu 61020f792cf9SNaoya Horiguchi /* 6103c89357e2SDavid Hildenbrand * hugetlb_wp() requires page locks of pte_page(entry) and 61040f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 61050f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 61060f792cf9SNaoya Horiguchi */ 61070f792cf9SNaoya Horiguchi page = pte_page(entry); 61080f792cf9SNaoya Horiguchi if (page != pagecache_page) 61090f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 61100f792cf9SNaoya Horiguchi need_wait_lock = 1; 61110f792cf9SNaoya Horiguchi goto out_ptl; 61120f792cf9SNaoya Horiguchi } 61130f792cf9SNaoya Horiguchi 61140f792cf9SNaoya Horiguchi get_page(page); 6115b4d1d99fSDavid Gibson 6116c89357e2SDavid Hildenbrand if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) { 6117106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 6118c89357e2SDavid Hildenbrand ret = hugetlb_wp(mm, vma, address, ptep, flags, 6119cb900f41SKirill A. Shutemov pagecache_page, ptl); 61200f792cf9SNaoya Horiguchi goto out_put_page; 6121c89357e2SDavid Hildenbrand } else if (likely(flags & FAULT_FLAG_WRITE)) { 6122106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 6123b4d1d99fSDavid Gibson } 6124c89357e2SDavid Hildenbrand } 6125b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 6126285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 6127788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 6128285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 61290f792cf9SNaoya Horiguchi out_put_page: 61300f792cf9SNaoya Horiguchi if (page != pagecache_page) 61310f792cf9SNaoya Horiguchi unlock_page(page); 61320f792cf9SNaoya Horiguchi put_page(page); 6133cb900f41SKirill A. Shutemov out_ptl: 6134cb900f41SKirill A. Shutemov spin_unlock(ptl); 613557303d80SAndy Whitcroft 613657303d80SAndy Whitcroft if (pagecache_page) { 613757303d80SAndy Whitcroft unlock_page(pagecache_page); 613857303d80SAndy Whitcroft put_page(pagecache_page); 613957303d80SAndy Whitcroft } 6140b4d1d99fSDavid Gibson out_mutex: 614140549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 6142c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 61430f792cf9SNaoya Horiguchi /* 61440f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 61450f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 61460f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 61470f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 61480f792cf9SNaoya Horiguchi * here without taking refcount. 61490f792cf9SNaoya Horiguchi */ 61500f792cf9SNaoya Horiguchi if (need_wait_lock) 61510f792cf9SNaoya Horiguchi wait_on_page_locked(page); 61521e8f889bSDavid Gibson return ret; 615386e5216fSAdam Litke } 615486e5216fSAdam Litke 6155714c1891SAxel Rasmussen #ifdef CONFIG_USERFAULTFD 61568fb5debcSMike Kravetz /* 61578fb5debcSMike Kravetz * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with 61588fb5debcSMike Kravetz * modifications for huge pages. 61598fb5debcSMike Kravetz */ 61608fb5debcSMike Kravetz int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, 61618fb5debcSMike Kravetz pte_t *dst_pte, 61628fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 61638fb5debcSMike Kravetz unsigned long dst_addr, 61648fb5debcSMike Kravetz unsigned long src_addr, 6165f6191471SAxel Rasmussen enum mcopy_atomic_mode mode, 61666041c691SPeter Xu struct page **pagep, 61676041c691SPeter Xu bool wp_copy) 61688fb5debcSMike Kravetz { 6169f6191471SAxel Rasmussen bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE); 61708cc5fcbbSMina Almasry struct hstate *h = hstate_vma(dst_vma); 61718cc5fcbbSMina Almasry struct address_space *mapping = dst_vma->vm_file->f_mapping; 61728cc5fcbbSMina Almasry pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr); 61731e392147SAndrea Arcangeli unsigned long size; 61741c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 61758fb5debcSMike Kravetz pte_t _dst_pte; 61768fb5debcSMike Kravetz spinlock_t *ptl; 61778cc5fcbbSMina Almasry int ret = -ENOMEM; 61788fb5debcSMike Kravetz struct page *page; 6179f6191471SAxel Rasmussen int writable; 6180cc30042dSMina Almasry bool page_in_pagecache = false; 61818fb5debcSMike Kravetz 6182f6191471SAxel Rasmussen if (is_continue) { 6183f6191471SAxel Rasmussen ret = -EFAULT; 6184f6191471SAxel Rasmussen page = find_lock_page(mapping, idx); 6185f6191471SAxel Rasmussen if (!page) 6186f6191471SAxel Rasmussen goto out; 6187cc30042dSMina Almasry page_in_pagecache = true; 6188f6191471SAxel Rasmussen } else if (!*pagep) { 6189d84cf06eSMina Almasry /* If a page already exists, then it's UFFDIO_COPY for 6190d84cf06eSMina Almasry * a non-missing case. Return -EEXIST. 6191d84cf06eSMina Almasry */ 6192d84cf06eSMina Almasry if (vm_shared && 6193d84cf06eSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 6194d84cf06eSMina Almasry ret = -EEXIST; 61958fb5debcSMike Kravetz goto out; 6196d84cf06eSMina Almasry } 6197d84cf06eSMina Almasry 6198d84cf06eSMina Almasry page = alloc_huge_page(dst_vma, dst_addr, 0); 6199d84cf06eSMina Almasry if (IS_ERR(page)) { 6200d84cf06eSMina Almasry ret = -ENOMEM; 6201d84cf06eSMina Almasry goto out; 6202d84cf06eSMina Almasry } 62038fb5debcSMike Kravetz 62048fb5debcSMike Kravetz ret = copy_huge_page_from_user(page, 62058fb5debcSMike Kravetz (const void __user *) src_addr, 6206810a56b9SMike Kravetz pages_per_huge_page(h), false); 62078fb5debcSMike Kravetz 6208c1e8d7c6SMichel Lespinasse /* fallback to copy_from_user outside mmap_lock */ 62098fb5debcSMike Kravetz if (unlikely(ret)) { 62109e368259SAndrea Arcangeli ret = -ENOENT; 62118cc5fcbbSMina Almasry /* Free the allocated page which may have 62128cc5fcbbSMina Almasry * consumed a reservation. 62138cc5fcbbSMina Almasry */ 62148cc5fcbbSMina Almasry restore_reserve_on_error(h, dst_vma, dst_addr, page); 62158cc5fcbbSMina Almasry put_page(page); 62168cc5fcbbSMina Almasry 62178cc5fcbbSMina Almasry /* Allocate a temporary page to hold the copied 62188cc5fcbbSMina Almasry * contents. 62198cc5fcbbSMina Almasry */ 62208cc5fcbbSMina Almasry page = alloc_huge_page_vma(h, dst_vma, dst_addr); 62218cc5fcbbSMina Almasry if (!page) { 62228cc5fcbbSMina Almasry ret = -ENOMEM; 62238cc5fcbbSMina Almasry goto out; 62248cc5fcbbSMina Almasry } 62258fb5debcSMike Kravetz *pagep = page; 62268cc5fcbbSMina Almasry /* Set the outparam pagep and return to the caller to 62278cc5fcbbSMina Almasry * copy the contents outside the lock. Don't free the 62288cc5fcbbSMina Almasry * page. 62298cc5fcbbSMina Almasry */ 62308fb5debcSMike Kravetz goto out; 62318fb5debcSMike Kravetz } 62328fb5debcSMike Kravetz } else { 62338cc5fcbbSMina Almasry if (vm_shared && 62348cc5fcbbSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 62358cc5fcbbSMina Almasry put_page(*pagep); 62368cc5fcbbSMina Almasry ret = -EEXIST; 62378cc5fcbbSMina Almasry *pagep = NULL; 62388cc5fcbbSMina Almasry goto out; 62398cc5fcbbSMina Almasry } 62408cc5fcbbSMina Almasry 62418cc5fcbbSMina Almasry page = alloc_huge_page(dst_vma, dst_addr, 0); 62428cc5fcbbSMina Almasry if (IS_ERR(page)) { 6243da9a298fSMiaohe Lin put_page(*pagep); 62448cc5fcbbSMina Almasry ret = -ENOMEM; 62458cc5fcbbSMina Almasry *pagep = NULL; 62468cc5fcbbSMina Almasry goto out; 62478cc5fcbbSMina Almasry } 624834892366SMuchun Song copy_user_huge_page(page, *pagep, dst_addr, dst_vma, 624934892366SMuchun Song pages_per_huge_page(h)); 62508cc5fcbbSMina Almasry put_page(*pagep); 62518fb5debcSMike Kravetz *pagep = NULL; 62528fb5debcSMike Kravetz } 62538fb5debcSMike Kravetz 62548fb5debcSMike Kravetz /* 62558fb5debcSMike Kravetz * The memory barrier inside __SetPageUptodate makes sure that 62568fb5debcSMike Kravetz * preceding stores to the page contents become visible before 62578fb5debcSMike Kravetz * the set_pte_at() write. 62588fb5debcSMike Kravetz */ 62598fb5debcSMike Kravetz __SetPageUptodate(page); 62608fb5debcSMike Kravetz 6261f6191471SAxel Rasmussen /* Add shared, newly allocated pages to the page cache. */ 6262f6191471SAxel Rasmussen if (vm_shared && !is_continue) { 62631e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 62641e392147SAndrea Arcangeli ret = -EFAULT; 62651e392147SAndrea Arcangeli if (idx >= size) 62661e392147SAndrea Arcangeli goto out_release_nounlock; 62671c9e8defSMike Kravetz 62681e392147SAndrea Arcangeli /* 62691e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 62707e1813d4SMike Kravetz * hugetlb_add_to_page_cache() below happens through the 62711e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 62721e392147SAndrea Arcangeli * the caller. 62731e392147SAndrea Arcangeli */ 62747e1813d4SMike Kravetz ret = hugetlb_add_to_page_cache(page, mapping, idx); 62751c9e8defSMike Kravetz if (ret) 62761c9e8defSMike Kravetz goto out_release_nounlock; 6277cc30042dSMina Almasry page_in_pagecache = true; 62781c9e8defSMike Kravetz } 62791c9e8defSMike Kravetz 6280bcc66543SMiaohe Lin ptl = huge_pte_lock(h, dst_mm, dst_pte); 62818fb5debcSMike Kravetz 62828625147cSJames Houghton ret = -EIO; 62838625147cSJames Houghton if (PageHWPoison(page)) 62848625147cSJames Houghton goto out_release_unlock; 62858625147cSJames Houghton 62861e392147SAndrea Arcangeli /* 62876041c691SPeter Xu * We allow to overwrite a pte marker: consider when both MISSING|WP 62886041c691SPeter Xu * registered, we firstly wr-protect a none pte which has no page cache 62896041c691SPeter Xu * page backing it, then access the page. 62906041c691SPeter Xu */ 6291fa27759aSMike Kravetz ret = -EEXIST; 62926041c691SPeter Xu if (!huge_pte_none_mostly(huge_ptep_get(dst_pte))) 62938fb5debcSMike Kravetz goto out_release_unlock; 62948fb5debcSMike Kravetz 62954781593dSPeter Xu if (page_in_pagecache) 6296fb3d824dSDavid Hildenbrand page_dup_file_rmap(page, true); 62974781593dSPeter Xu else 62988fb5debcSMike Kravetz hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); 62998fb5debcSMike Kravetz 63006041c691SPeter Xu /* 63016041c691SPeter Xu * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY 63026041c691SPeter Xu * with wp flag set, don't set pte write bit. 63036041c691SPeter Xu */ 63046041c691SPeter Xu if (wp_copy || (is_continue && !vm_shared)) 6305f6191471SAxel Rasmussen writable = 0; 6306f6191471SAxel Rasmussen else 6307f6191471SAxel Rasmussen writable = dst_vma->vm_flags & VM_WRITE; 6308f6191471SAxel Rasmussen 6309f6191471SAxel Rasmussen _dst_pte = make_huge_pte(dst_vma, page, writable); 63106041c691SPeter Xu /* 63116041c691SPeter Xu * Always mark UFFDIO_COPY page dirty; note that this may not be 63126041c691SPeter Xu * extremely important for hugetlbfs for now since swapping is not 63136041c691SPeter Xu * supported, but we should still be clear in that this page cannot be 63146041c691SPeter Xu * thrown away at will, even if write bit not set. 63156041c691SPeter Xu */ 63168fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 63178fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 63188fb5debcSMike Kravetz 63196041c691SPeter Xu if (wp_copy) 63206041c691SPeter Xu _dst_pte = huge_pte_mkuffd_wp(_dst_pte); 63216041c691SPeter Xu 63228fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 63238fb5debcSMike Kravetz 63248fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 63258fb5debcSMike Kravetz 63268fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 63278fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 63288fb5debcSMike Kravetz 63298fb5debcSMike Kravetz spin_unlock(ptl); 6330f6191471SAxel Rasmussen if (!is_continue) 63318f251a3dSMike Kravetz SetHPageMigratable(page); 6332f6191471SAxel Rasmussen if (vm_shared || is_continue) 63331c9e8defSMike Kravetz unlock_page(page); 63348fb5debcSMike Kravetz ret = 0; 63358fb5debcSMike Kravetz out: 63368fb5debcSMike Kravetz return ret; 63378fb5debcSMike Kravetz out_release_unlock: 63388fb5debcSMike Kravetz spin_unlock(ptl); 6339f6191471SAxel Rasmussen if (vm_shared || is_continue) 63401c9e8defSMike Kravetz unlock_page(page); 63415af10dfdSAndrea Arcangeli out_release_nounlock: 6342cc30042dSMina Almasry if (!page_in_pagecache) 6343846be085SMike Kravetz restore_reserve_on_error(h, dst_vma, dst_addr, page); 63448fb5debcSMike Kravetz put_page(page); 63458fb5debcSMike Kravetz goto out; 63468fb5debcSMike Kravetz } 6347714c1891SAxel Rasmussen #endif /* CONFIG_USERFAULTFD */ 63488fb5debcSMike Kravetz 634982e5d378SJoao Martins static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma, 635082e5d378SJoao Martins int refs, struct page **pages, 635182e5d378SJoao Martins struct vm_area_struct **vmas) 635282e5d378SJoao Martins { 635382e5d378SJoao Martins int nr; 635482e5d378SJoao Martins 635582e5d378SJoao Martins for (nr = 0; nr < refs; nr++) { 635682e5d378SJoao Martins if (likely(pages)) 635714455eabSCheng Li pages[nr] = nth_page(page, nr); 635882e5d378SJoao Martins if (vmas) 635982e5d378SJoao Martins vmas[nr] = vma; 636082e5d378SJoao Martins } 636182e5d378SJoao Martins } 636282e5d378SJoao Martins 636384209e87SDavid Hildenbrand static inline bool __follow_hugetlb_must_fault(struct vm_area_struct *vma, 636484209e87SDavid Hildenbrand unsigned int flags, pte_t *pte, 6365a7f22660SDavid Hildenbrand bool *unshare) 6366a7f22660SDavid Hildenbrand { 6367a7f22660SDavid Hildenbrand pte_t pteval = huge_ptep_get(pte); 6368a7f22660SDavid Hildenbrand 6369a7f22660SDavid Hildenbrand *unshare = false; 6370a7f22660SDavid Hildenbrand if (is_swap_pte(pteval)) 6371a7f22660SDavid Hildenbrand return true; 6372a7f22660SDavid Hildenbrand if (huge_pte_write(pteval)) 6373a7f22660SDavid Hildenbrand return false; 6374a7f22660SDavid Hildenbrand if (flags & FOLL_WRITE) 6375a7f22660SDavid Hildenbrand return true; 637684209e87SDavid Hildenbrand if (gup_must_unshare(vma, flags, pte_page(pteval))) { 6377a7f22660SDavid Hildenbrand *unshare = true; 6378a7f22660SDavid Hildenbrand return true; 6379a7f22660SDavid Hildenbrand } 6380a7f22660SDavid Hildenbrand return false; 6381a7f22660SDavid Hildenbrand } 6382a7f22660SDavid Hildenbrand 638357a196a5SMike Kravetz struct page *hugetlb_follow_page_mask(struct vm_area_struct *vma, 638457a196a5SMike Kravetz unsigned long address, unsigned int flags) 638557a196a5SMike Kravetz { 638657a196a5SMike Kravetz struct hstate *h = hstate_vma(vma); 638757a196a5SMike Kravetz struct mm_struct *mm = vma->vm_mm; 638857a196a5SMike Kravetz unsigned long haddr = address & huge_page_mask(h); 638957a196a5SMike Kravetz struct page *page = NULL; 639057a196a5SMike Kravetz spinlock_t *ptl; 639157a196a5SMike Kravetz pte_t *pte, entry; 639257a196a5SMike Kravetz 639357a196a5SMike Kravetz /* 639457a196a5SMike Kravetz * FOLL_PIN is not supported for follow_page(). Ordinary GUP goes via 639557a196a5SMike Kravetz * follow_hugetlb_page(). 639657a196a5SMike Kravetz */ 639757a196a5SMike Kravetz if (WARN_ON_ONCE(flags & FOLL_PIN)) 639857a196a5SMike Kravetz return NULL; 639957a196a5SMike Kravetz 64007d049f3aSPeter Xu hugetlb_vma_lock_read(vma); 64019c67a207SPeter Xu pte = hugetlb_walk(vma, haddr, huge_page_size(h)); 640257a196a5SMike Kravetz if (!pte) 64037d049f3aSPeter Xu goto out_unlock; 640457a196a5SMike Kravetz 640557a196a5SMike Kravetz ptl = huge_pte_lock(h, mm, pte); 640657a196a5SMike Kravetz entry = huge_ptep_get(pte); 640757a196a5SMike Kravetz if (pte_present(entry)) { 640857a196a5SMike Kravetz page = pte_page(entry) + 640957a196a5SMike Kravetz ((address & ~huge_page_mask(h)) >> PAGE_SHIFT); 641057a196a5SMike Kravetz /* 641157a196a5SMike Kravetz * Note that page may be a sub-page, and with vmemmap 641257a196a5SMike Kravetz * optimizations the page struct may be read only. 641357a196a5SMike Kravetz * try_grab_page() will increase the ref count on the 641457a196a5SMike Kravetz * head page, so this will be OK. 641557a196a5SMike Kravetz * 6416e2ca6ba6SLinus Torvalds * try_grab_page() should always be able to get the page here, 6417e2ca6ba6SLinus Torvalds * because we hold the ptl lock and have verified pte_present(). 641857a196a5SMike Kravetz */ 6419e2ca6ba6SLinus Torvalds if (try_grab_page(page, flags)) { 642057a196a5SMike Kravetz page = NULL; 642157a196a5SMike Kravetz goto out; 642257a196a5SMike Kravetz } 642357a196a5SMike Kravetz } 642457a196a5SMike Kravetz out: 642557a196a5SMike Kravetz spin_unlock(ptl); 64267d049f3aSPeter Xu out_unlock: 64277d049f3aSPeter Xu hugetlb_vma_unlock_read(vma); 642857a196a5SMike Kravetz return page; 642957a196a5SMike Kravetz } 643057a196a5SMike Kravetz 643128a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 643263551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 643328a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 64344f6da934SPeter Xu long i, unsigned int flags, int *locked) 643563551ae0SDavid Gibson { 6436d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 6437d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 643828a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 6439a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 64400fa5bc40SJoao Martins int err = -EFAULT, refs; 644163551ae0SDavid Gibson 644263551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 644363551ae0SDavid Gibson pte_t *pte; 6444cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 6445a7f22660SDavid Hildenbrand bool unshare = false; 64462a15efc9SHugh Dickins int absent; 644763551ae0SDavid Gibson struct page *page; 644863551ae0SDavid Gibson 64494c887265SAdam Litke /* 645002057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 645102057967SDavid Rientjes * potentially allocating memory. 645202057967SDavid Rientjes */ 6453fa45f116SDavidlohr Bueso if (fatal_signal_pending(current)) { 645402057967SDavid Rientjes remainder = 0; 645502057967SDavid Rientjes break; 645602057967SDavid Rientjes } 645702057967SDavid Rientjes 6458eefc7fa5SPeter Xu hugetlb_vma_lock_read(vma); 645902057967SDavid Rientjes /* 64604c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 64612a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 64624c887265SAdam Litke * first, for the page indexing below to work. 6463cb900f41SKirill A. Shutemov * 6464cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 64654c887265SAdam Litke */ 64669c67a207SPeter Xu pte = hugetlb_walk(vma, vaddr & huge_page_mask(h), 64677868a208SPunit Agrawal huge_page_size(h)); 6468cb900f41SKirill A. Shutemov if (pte) 6469cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 64702a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 647163551ae0SDavid Gibson 64722a15efc9SHugh Dickins /* 64732a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 64743ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 64753ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 64763ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 64773ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 64782a15efc9SHugh Dickins */ 64793ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 64803ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 6481cb900f41SKirill A. Shutemov if (pte) 6482cb900f41SKirill A. Shutemov spin_unlock(ptl); 6483eefc7fa5SPeter Xu hugetlb_vma_unlock_read(vma); 64842a15efc9SHugh Dickins remainder = 0; 64852a15efc9SHugh Dickins break; 64862a15efc9SHugh Dickins } 64872a15efc9SHugh Dickins 64889cc3a5bdSNaoya Horiguchi /* 64899cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 64909cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 64919cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 64929cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 64939cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 64949cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 64959cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 64969cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 64979cc3a5bdSNaoya Horiguchi */ 6498a7f22660SDavid Hildenbrand if (absent || 649984209e87SDavid Hildenbrand __follow_hugetlb_must_fault(vma, flags, pte, &unshare)) { 65002b740303SSouptick Joarder vm_fault_t ret; 650187ffc118SAndrea Arcangeli unsigned int fault_flags = 0; 65024c887265SAdam Litke 6503cb900f41SKirill A. Shutemov if (pte) 6504cb900f41SKirill A. Shutemov spin_unlock(ptl); 6505eefc7fa5SPeter Xu hugetlb_vma_unlock_read(vma); 6506eefc7fa5SPeter Xu 650787ffc118SAndrea Arcangeli if (flags & FOLL_WRITE) 650887ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_WRITE; 6509a7f22660SDavid Hildenbrand else if (unshare) 6510a7f22660SDavid Hildenbrand fault_flags |= FAULT_FLAG_UNSHARE; 651193c5c61dSPeter Xu if (locked) { 651271335f37SPeter Xu fault_flags |= FAULT_FLAG_ALLOW_RETRY | 651371335f37SPeter Xu FAULT_FLAG_KILLABLE; 651493c5c61dSPeter Xu if (flags & FOLL_INTERRUPTIBLE) 651593c5c61dSPeter Xu fault_flags |= FAULT_FLAG_INTERRUPTIBLE; 651693c5c61dSPeter Xu } 651787ffc118SAndrea Arcangeli if (flags & FOLL_NOWAIT) 651887ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY | 651987ffc118SAndrea Arcangeli FAULT_FLAG_RETRY_NOWAIT; 652087ffc118SAndrea Arcangeli if (flags & FOLL_TRIED) { 65214426e945SPeter Xu /* 65224426e945SPeter Xu * Note: FAULT_FLAG_ALLOW_RETRY and 65234426e945SPeter Xu * FAULT_FLAG_TRIED can co-exist 65244426e945SPeter Xu */ 652587ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_TRIED; 652687ffc118SAndrea Arcangeli } 652787ffc118SAndrea Arcangeli ret = hugetlb_fault(mm, vma, vaddr, fault_flags); 652887ffc118SAndrea Arcangeli if (ret & VM_FAULT_ERROR) { 65292be7cfedSDaniel Jordan err = vm_fault_to_errno(ret, flags); 65301c59827dSHugh Dickins remainder = 0; 65311c59827dSHugh Dickins break; 65321c59827dSHugh Dickins } 653387ffc118SAndrea Arcangeli if (ret & VM_FAULT_RETRY) { 65344f6da934SPeter Xu if (locked && 65351ac25013SAndrea Arcangeli !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) 65364f6da934SPeter Xu *locked = 0; 653787ffc118SAndrea Arcangeli *nr_pages = 0; 653887ffc118SAndrea Arcangeli /* 653987ffc118SAndrea Arcangeli * VM_FAULT_RETRY must not return an 654087ffc118SAndrea Arcangeli * error, it will return zero 654187ffc118SAndrea Arcangeli * instead. 654287ffc118SAndrea Arcangeli * 654387ffc118SAndrea Arcangeli * No need to update "position" as the 654487ffc118SAndrea Arcangeli * caller will not check it after 654587ffc118SAndrea Arcangeli * *nr_pages is set to 0. 654687ffc118SAndrea Arcangeli */ 654787ffc118SAndrea Arcangeli return i; 654887ffc118SAndrea Arcangeli } 654987ffc118SAndrea Arcangeli continue; 655087ffc118SAndrea Arcangeli } 655163551ae0SDavid Gibson 6552a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 65537f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 65548fde12caSLinus Torvalds 6555b6a2619cSDavid Hildenbrand VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) && 6556b6a2619cSDavid Hildenbrand !PageAnonExclusive(page), page); 6557b6a2619cSDavid Hildenbrand 65588fde12caSLinus Torvalds /* 6559acbfb087SZhigang Lu * If subpage information not requested, update counters 6560acbfb087SZhigang Lu * and skip the same_page loop below. 6561acbfb087SZhigang Lu */ 6562acbfb087SZhigang Lu if (!pages && !vmas && !pfn_offset && 6563acbfb087SZhigang Lu (vaddr + huge_page_size(h) < vma->vm_end) && 6564acbfb087SZhigang Lu (remainder >= pages_per_huge_page(h))) { 6565acbfb087SZhigang Lu vaddr += huge_page_size(h); 6566acbfb087SZhigang Lu remainder -= pages_per_huge_page(h); 6567acbfb087SZhigang Lu i += pages_per_huge_page(h); 6568acbfb087SZhigang Lu spin_unlock(ptl); 6569eefc7fa5SPeter Xu hugetlb_vma_unlock_read(vma); 6570acbfb087SZhigang Lu continue; 6571acbfb087SZhigang Lu } 6572acbfb087SZhigang Lu 6573d08af0a5SJoao Martins /* vaddr may not be aligned to PAGE_SIZE */ 6574d08af0a5SJoao Martins refs = min3(pages_per_huge_page(h) - pfn_offset, remainder, 6575d08af0a5SJoao Martins (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT); 65760fa5bc40SJoao Martins 657782e5d378SJoao Martins if (pages || vmas) 657814455eabSCheng Li record_subpages_vmas(nth_page(page, pfn_offset), 657982e5d378SJoao Martins vma, refs, 658082e5d378SJoao Martins likely(pages) ? pages + i : NULL, 658182e5d378SJoao Martins vmas ? vmas + i : NULL); 658263551ae0SDavid Gibson 658382e5d378SJoao Martins if (pages) { 65840fa5bc40SJoao Martins /* 6585822951d8SMatthew Wilcox (Oracle) * try_grab_folio() should always succeed here, 65860fa5bc40SJoao Martins * because: a) we hold the ptl lock, and b) we've just 65870fa5bc40SJoao Martins * checked that the huge page is present in the page 65880fa5bc40SJoao Martins * tables. If the huge page is present, then the tail 65890fa5bc40SJoao Martins * pages must also be present. The ptl prevents the 65900fa5bc40SJoao Martins * head page and tail pages from being rearranged in 65914003f107SLogan Gunthorpe * any way. As this is hugetlb, the pages will never 65924003f107SLogan Gunthorpe * be p2pdma or not longterm pinable. So this page 65934003f107SLogan Gunthorpe * must be available at this point, unless the page 65944003f107SLogan Gunthorpe * refcount overflowed: 65950fa5bc40SJoao Martins */ 6596822951d8SMatthew Wilcox (Oracle) if (WARN_ON_ONCE(!try_grab_folio(pages[i], refs, 65970fa5bc40SJoao Martins flags))) { 65980fa5bc40SJoao Martins spin_unlock(ptl); 6599eefc7fa5SPeter Xu hugetlb_vma_unlock_read(vma); 66000fa5bc40SJoao Martins remainder = 0; 66010fa5bc40SJoao Martins err = -ENOMEM; 66020fa5bc40SJoao Martins break; 66030fa5bc40SJoao Martins } 6604d5d4b0aaSChen, Kenneth W } 660582e5d378SJoao Martins 660682e5d378SJoao Martins vaddr += (refs << PAGE_SHIFT); 660782e5d378SJoao Martins remainder -= refs; 660882e5d378SJoao Martins i += refs; 660982e5d378SJoao Martins 6610cb900f41SKirill A. Shutemov spin_unlock(ptl); 6611eefc7fa5SPeter Xu hugetlb_vma_unlock_read(vma); 661263551ae0SDavid Gibson } 661328a35716SMichel Lespinasse *nr_pages = remainder; 661487ffc118SAndrea Arcangeli /* 661587ffc118SAndrea Arcangeli * setting position is actually required only if remainder is 661687ffc118SAndrea Arcangeli * not zero but it's faster not to add a "if (remainder)" 661787ffc118SAndrea Arcangeli * branch. 661887ffc118SAndrea Arcangeli */ 661963551ae0SDavid Gibson *position = vaddr; 662063551ae0SDavid Gibson 66212be7cfedSDaniel Jordan return i ? i : err; 662263551ae0SDavid Gibson } 66238f860591SZhang, Yanmin 6624a79390f5SPeter Xu long hugetlb_change_protection(struct vm_area_struct *vma, 66255a90d5a1SPeter Xu unsigned long address, unsigned long end, 66265a90d5a1SPeter Xu pgprot_t newprot, unsigned long cp_flags) 66278f860591SZhang, Yanmin { 66288f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 66298f860591SZhang, Yanmin unsigned long start = address; 66308f860591SZhang, Yanmin pte_t *ptep; 66318f860591SZhang, Yanmin pte_t pte; 6632a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 6633a79390f5SPeter Xu long pages = 0, psize = huge_page_size(h); 6634dff11abeSMike Kravetz bool shared_pmd = false; 6635ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 6636e95a9851SMike Kravetz unsigned long last_addr_mask; 66375a90d5a1SPeter Xu bool uffd_wp = cp_flags & MM_CP_UFFD_WP; 66385a90d5a1SPeter Xu bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE; 6639dff11abeSMike Kravetz 6640dff11abeSMike Kravetz /* 6641dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 6642ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 6643dff11abeSMike Kravetz * range if PMD sharing is possible. 6644dff11abeSMike Kravetz */ 66457269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 66467d4a8be0SAlistair Popple 0, mm, start, end); 6647ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 66488f860591SZhang, Yanmin 66498f860591SZhang, Yanmin BUG_ON(address >= end); 6650ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 66518f860591SZhang, Yanmin 6652ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 665340549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 665483cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 665540549ba8SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 665660dfaad6SPeter Xu for (; address < end; address += psize) { 6657cb900f41SKirill A. Shutemov spinlock_t *ptl; 66589c67a207SPeter Xu ptep = hugetlb_walk(vma, address, psize); 6659e95a9851SMike Kravetz if (!ptep) { 6660fed15f13SPeter Xu if (!uffd_wp) { 6661e95a9851SMike Kravetz address |= last_addr_mask; 66628f860591SZhang, Yanmin continue; 6663e95a9851SMike Kravetz } 6664fed15f13SPeter Xu /* 6665fed15f13SPeter Xu * Userfaultfd wr-protect requires pgtable 6666fed15f13SPeter Xu * pre-allocations to install pte markers. 6667fed15f13SPeter Xu */ 6668fed15f13SPeter Xu ptep = huge_pte_alloc(mm, vma, address, psize); 6669d1751118SPeter Xu if (!ptep) { 6670d1751118SPeter Xu pages = -ENOMEM; 6671fed15f13SPeter Xu break; 6672fed15f13SPeter Xu } 6673d1751118SPeter Xu } 6674cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 66754ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, address, ptep)) { 667660dfaad6SPeter Xu /* 667760dfaad6SPeter Xu * When uffd-wp is enabled on the vma, unshare 667860dfaad6SPeter Xu * shouldn't happen at all. Warn about it if it 667960dfaad6SPeter Xu * happened due to some reason. 668060dfaad6SPeter Xu */ 668160dfaad6SPeter Xu WARN_ON_ONCE(uffd_wp || uffd_wp_resolve); 66827da4d641SPeter Zijlstra pages++; 6683cb900f41SKirill A. Shutemov spin_unlock(ptl); 6684dff11abeSMike Kravetz shared_pmd = true; 66854ddb4d91SMike Kravetz address |= last_addr_mask; 668639dde65cSChen, Kenneth W continue; 66877da4d641SPeter Zijlstra } 6688a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 6689a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 66900e678153SDavid Hildenbrand /* Nothing to do. */ 66910e678153SDavid Hildenbrand } else if (unlikely(is_hugetlb_entry_migration(pte))) { 6692a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 66936c287605SDavid Hildenbrand struct page *page = pfn_swap_entry_to_page(entry); 669444f86392SDavid Hildenbrand pte_t newpte = pte; 6695a8bda28dSNaoya Horiguchi 669644f86392SDavid Hildenbrand if (is_writable_migration_entry(entry)) { 66976c287605SDavid Hildenbrand if (PageAnon(page)) 66986c287605SDavid Hildenbrand entry = make_readable_exclusive_migration_entry( 66996c287605SDavid Hildenbrand swp_offset(entry)); 67006c287605SDavid Hildenbrand else 67014dd845b5SAlistair Popple entry = make_readable_migration_entry( 67024dd845b5SAlistair Popple swp_offset(entry)); 6703a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 670444f86392SDavid Hildenbrand pages++; 670544f86392SDavid Hildenbrand } 670644f86392SDavid Hildenbrand 67075a90d5a1SPeter Xu if (uffd_wp) 67085a90d5a1SPeter Xu newpte = pte_swp_mkuffd_wp(newpte); 67095a90d5a1SPeter Xu else if (uffd_wp_resolve) 67105a90d5a1SPeter Xu newpte = pte_swp_clear_uffd_wp(newpte); 671144f86392SDavid Hildenbrand if (!pte_same(pte, newpte)) 671218f39629SQi Zheng set_huge_pte_at(mm, address, ptep, newpte); 67130e678153SDavid Hildenbrand } else if (unlikely(is_pte_marker(pte))) { 67140e678153SDavid Hildenbrand /* No other markers apply for now. */ 67150e678153SDavid Hildenbrand WARN_ON_ONCE(!pte_marker_uffd_wp(pte)); 671660dfaad6SPeter Xu if (uffd_wp_resolve) 67170e678153SDavid Hildenbrand /* Safe to modify directly (non-present->none). */ 671860dfaad6SPeter Xu huge_pte_clear(mm, address, ptep, psize); 67190e678153SDavid Hildenbrand } else if (!huge_pte_none(pte)) { 6720023bdd00SAneesh Kumar K.V pte_t old_pte; 672179c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 6722023bdd00SAneesh Kumar K.V 6723023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 672416785bd7SAnshuman Khandual pte = huge_pte_modify(old_pte, newprot); 672579c1c594SChristophe Leroy pte = arch_make_huge_pte(pte, shift, vma->vm_flags); 67265a90d5a1SPeter Xu if (uffd_wp) 6727f1eb1bacSPeter Xu pte = huge_pte_mkuffd_wp(pte); 67285a90d5a1SPeter Xu else if (uffd_wp_resolve) 67295a90d5a1SPeter Xu pte = huge_pte_clear_uffd_wp(pte); 6730023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 67317da4d641SPeter Zijlstra pages++; 673260dfaad6SPeter Xu } else { 673360dfaad6SPeter Xu /* None pte */ 673460dfaad6SPeter Xu if (unlikely(uffd_wp)) 673560dfaad6SPeter Xu /* Safe to modify directly (none->non-present). */ 673660dfaad6SPeter Xu set_huge_pte_at(mm, address, ptep, 673760dfaad6SPeter Xu make_pte_marker(PTE_MARKER_UFFD_WP)); 67388f860591SZhang, Yanmin } 6739cb900f41SKirill A. Shutemov spin_unlock(ptl); 67408f860591SZhang, Yanmin } 6741d833352aSMel Gorman /* 6742c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 6743d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 6744c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 6745dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 6746dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 6747d833352aSMel Gorman */ 6748dff11abeSMike Kravetz if (shared_pmd) 6749ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 6750dff11abeSMike Kravetz else 67515491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 67520f10851eSJérôme Glisse /* 67530f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 67540f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 67550f10851eSJérôme Glisse * 6756ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 67570f10851eSJérôme Glisse */ 675883cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 675940549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 6760ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 67617da4d641SPeter Zijlstra 6762d1751118SPeter Xu return pages > 0 ? (pages << h->order) : pages; 67638f860591SZhang, Yanmin } 67648f860591SZhang, Yanmin 676533b8f84aSMike Kravetz /* Return true if reservation was successful, false otherwise. */ 676633b8f84aSMike Kravetz bool hugetlb_reserve_pages(struct inode *inode, 6767a1e78772SMel Gorman long from, long to, 67685a6fe125SMel Gorman struct vm_area_struct *vma, 6769ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 6770e4e574b7SAdam Litke { 6771c5094ec7SMike Kravetz long chg = -1, add = -1; 6772a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 677390481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 67749119a41eSJoonsoo Kim struct resv_map *resv_map; 6775075a61d0SMina Almasry struct hugetlb_cgroup *h_cg = NULL; 67760db9d74eSMina Almasry long gbl_reserve, regions_needed = 0; 6777e4e574b7SAdam Litke 677863489f8eSMike Kravetz /* This should never happen */ 677963489f8eSMike Kravetz if (from > to) { 678063489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 678133b8f84aSMike Kravetz return false; 678263489f8eSMike Kravetz } 678363489f8eSMike Kravetz 6784a1e78772SMel Gorman /* 6785e700898fSMike Kravetz * vma specific semaphore used for pmd sharing and fault/truncation 6786e700898fSMike Kravetz * synchronization 67878d9bfb26SMike Kravetz */ 67888d9bfb26SMike Kravetz hugetlb_vma_lock_alloc(vma); 67898d9bfb26SMike Kravetz 67908d9bfb26SMike Kravetz /* 679117c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 679217c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 679390481622SDavid Gibson * without using reserves 679417c9d12eSMel Gorman */ 6795ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 679633b8f84aSMike Kravetz return true; 679717c9d12eSMel Gorman 679817c9d12eSMel Gorman /* 6799a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 6800a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 6801a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 6802a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 6803a1e78772SMel Gorman */ 68049119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 6805f27a5136SMike Kravetz /* 6806f27a5136SMike Kravetz * resv_map can not be NULL as hugetlb_reserve_pages is only 6807f27a5136SMike Kravetz * called for inodes for which resv_maps were created (see 6808f27a5136SMike Kravetz * hugetlbfs_get_inode). 6809f27a5136SMike Kravetz */ 68104e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 68119119a41eSJoonsoo Kim 68120db9d74eSMina Almasry chg = region_chg(resv_map, from, to, ®ions_needed); 68139119a41eSJoonsoo Kim } else { 6814e9fe92aeSMina Almasry /* Private mapping. */ 68159119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 68165a6fe125SMel Gorman if (!resv_map) 68178d9bfb26SMike Kravetz goto out_err; 68185a6fe125SMel Gorman 681917c9d12eSMel Gorman chg = to - from; 682017c9d12eSMel Gorman 68215a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 68225a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 68235a6fe125SMel Gorman } 68245a6fe125SMel Gorman 682533b8f84aSMike Kravetz if (chg < 0) 6826c50ac050SDave Hansen goto out_err; 682717c9d12eSMel Gorman 682833b8f84aSMike Kravetz if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), 682933b8f84aSMike Kravetz chg * pages_per_huge_page(h), &h_cg) < 0) 6830075a61d0SMina Almasry goto out_err; 6831075a61d0SMina Almasry 6832075a61d0SMina Almasry if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { 6833075a61d0SMina Almasry /* For private mappings, the hugetlb_cgroup uncharge info hangs 6834075a61d0SMina Almasry * of the resv_map. 6835075a61d0SMina Almasry */ 6836075a61d0SMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); 6837075a61d0SMina Almasry } 6838075a61d0SMina Almasry 68391c5ecae3SMike Kravetz /* 68401c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 68411c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 68421c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 68431c5ecae3SMike Kravetz */ 68441c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 684533b8f84aSMike Kravetz if (gbl_reserve < 0) 6846075a61d0SMina Almasry goto out_uncharge_cgroup; 684717c9d12eSMel Gorman 684817c9d12eSMel Gorman /* 684917c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 685090481622SDavid Gibson * Hand the pages back to the subpool if there are not 685117c9d12eSMel Gorman */ 685233b8f84aSMike Kravetz if (hugetlb_acct_memory(h, gbl_reserve) < 0) 6853075a61d0SMina Almasry goto out_put_pages; 685417c9d12eSMel Gorman 685517c9d12eSMel Gorman /* 685617c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 685717c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 685817c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 685917c9d12eSMel Gorman * the reservation was and the page cache tells how much of 686017c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 686117c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 686217c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 686317c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 686417c9d12eSMel Gorman * else has to be done for private mappings here 686517c9d12eSMel Gorman */ 686633039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 6867075a61d0SMina Almasry add = region_add(resv_map, from, to, regions_needed, h, h_cg); 686833039678SMike Kravetz 68690db9d74eSMina Almasry if (unlikely(add < 0)) { 68700db9d74eSMina Almasry hugetlb_acct_memory(h, -gbl_reserve); 6871075a61d0SMina Almasry goto out_put_pages; 68720db9d74eSMina Almasry } else if (unlikely(chg > add)) { 687333039678SMike Kravetz /* 687433039678SMike Kravetz * pages in this range were added to the reserve 687533039678SMike Kravetz * map between region_chg and region_add. This 687633039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 687733039678SMike Kravetz * the subpool and reserve counts modified above 687833039678SMike Kravetz * based on the difference. 687933039678SMike Kravetz */ 688033039678SMike Kravetz long rsv_adjust; 688133039678SMike Kravetz 6882d85aecf2SMiaohe Lin /* 6883d85aecf2SMiaohe Lin * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the 6884d85aecf2SMiaohe Lin * reference to h_cg->css. See comment below for detail. 6885d85aecf2SMiaohe Lin */ 6886075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd( 6887075a61d0SMina Almasry hstate_index(h), 6888075a61d0SMina Almasry (chg - add) * pages_per_huge_page(h), h_cg); 6889075a61d0SMina Almasry 689033039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 689133039678SMike Kravetz chg - add); 689233039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 6893d85aecf2SMiaohe Lin } else if (h_cg) { 6894d85aecf2SMiaohe Lin /* 6895d85aecf2SMiaohe Lin * The file_regions will hold their own reference to 6896d85aecf2SMiaohe Lin * h_cg->css. So we should release the reference held 6897d85aecf2SMiaohe Lin * via hugetlb_cgroup_charge_cgroup_rsvd() when we are 6898d85aecf2SMiaohe Lin * done. 6899d85aecf2SMiaohe Lin */ 6900d85aecf2SMiaohe Lin hugetlb_cgroup_put_rsvd_cgroup(h_cg); 690133039678SMike Kravetz } 690233039678SMike Kravetz } 690333b8f84aSMike Kravetz return true; 690433b8f84aSMike Kravetz 6905075a61d0SMina Almasry out_put_pages: 6906075a61d0SMina Almasry /* put back original number of pages, chg */ 6907075a61d0SMina Almasry (void)hugepage_subpool_put_pages(spool, chg); 6908075a61d0SMina Almasry out_uncharge_cgroup: 6909075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), 6910075a61d0SMina Almasry chg * pages_per_huge_page(h), h_cg); 6911c50ac050SDave Hansen out_err: 69128d9bfb26SMike Kravetz hugetlb_vma_lock_free(vma); 69135e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 69140db9d74eSMina Almasry /* Only call region_abort if the region_chg succeeded but the 69150db9d74eSMina Almasry * region_add failed or didn't run. 69160db9d74eSMina Almasry */ 69170db9d74eSMina Almasry if (chg >= 0 && add < 0) 69180db9d74eSMina Almasry region_abort(resv_map, from, to, regions_needed); 6919f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 6920f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 692133b8f84aSMike Kravetz return false; 6922a43a8c39SChen, Kenneth W } 6923a43a8c39SChen, Kenneth W 6924b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 6925b5cec28dSMike Kravetz long freed) 6926a43a8c39SChen, Kenneth W { 6927a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 69284e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 69299119a41eSJoonsoo Kim long chg = 0; 693090481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 69311c5ecae3SMike Kravetz long gbl_reserve; 693245c682a6SKen Chen 6933f27a5136SMike Kravetz /* 6934f27a5136SMike Kravetz * Since this routine can be called in the evict inode path for all 6935f27a5136SMike Kravetz * hugetlbfs inodes, resv_map could be NULL. 6936f27a5136SMike Kravetz */ 6937b5cec28dSMike Kravetz if (resv_map) { 6938b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 6939b5cec28dSMike Kravetz /* 6940b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 6941b5cec28dSMike Kravetz * must be split and another region descriptor can not be 6942b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 6943b5cec28dSMike Kravetz */ 6944b5cec28dSMike Kravetz if (chg < 0) 6945b5cec28dSMike Kravetz return chg; 6946b5cec28dSMike Kravetz } 6947b5cec28dSMike Kravetz 694845c682a6SKen Chen spin_lock(&inode->i_lock); 6949e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 695045c682a6SKen Chen spin_unlock(&inode->i_lock); 695145c682a6SKen Chen 69521c5ecae3SMike Kravetz /* 69531c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 69541c5ecae3SMike Kravetz * reservations to be released may be adjusted. 6955dddf31a4SMiaohe Lin * 6956dddf31a4SMiaohe Lin * Note that !resv_map implies freed == 0. So (chg - freed) 6957dddf31a4SMiaohe Lin * won't go negative. 69581c5ecae3SMike Kravetz */ 69591c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 69601c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 6961b5cec28dSMike Kravetz 6962b5cec28dSMike Kravetz return 0; 6963a43a8c39SChen, Kenneth W } 696493f70f90SNaoya Horiguchi 69653212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 69663212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 69673212b535SSteve Capper struct vm_area_struct *vma, 69683212b535SSteve Capper unsigned long addr, pgoff_t idx) 69693212b535SSteve Capper { 69703212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 69713212b535SSteve Capper svma->vm_start; 69723212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 69733212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 69743212b535SSteve Capper 69753212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 6976e430a95aSSuren Baghdasaryan unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED_MASK; 6977e430a95aSSuren Baghdasaryan unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED_MASK; 69783212b535SSteve Capper 69793212b535SSteve Capper /* 69803212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 69813212b535SSteve Capper * page table page. 6982131a79b4SMike Kravetz * 6983131a79b4SMike Kravetz * Also, vma_lock (vm_private_data) is required for sharing. 69843212b535SSteve Capper */ 69853212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 69863212b535SSteve Capper vm_flags != svm_flags || 6987131a79b4SMike Kravetz !range_in_vma(svma, sbase, s_end) || 6988131a79b4SMike Kravetz !svma->vm_private_data) 69893212b535SSteve Capper return 0; 69903212b535SSteve Capper 69913212b535SSteve Capper return saddr; 69923212b535SSteve Capper } 69933212b535SSteve Capper 6994bbff39ccSMike Kravetz bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 69953212b535SSteve Capper { 6996bbff39ccSMike Kravetz unsigned long start = addr & PUD_MASK; 6997bbff39ccSMike Kravetz unsigned long end = start + PUD_SIZE; 6998bbff39ccSMike Kravetz 69998d9bfb26SMike Kravetz #ifdef CONFIG_USERFAULTFD 70008d9bfb26SMike Kravetz if (uffd_disable_huge_pmd_share(vma)) 70018d9bfb26SMike Kravetz return false; 70028d9bfb26SMike Kravetz #endif 70033212b535SSteve Capper /* 70043212b535SSteve Capper * check on proper vm_flags and page table alignment 70053212b535SSteve Capper */ 70068d9bfb26SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 700731aafb45SNicholas Krause return false; 7008bbff39ccSMike Kravetz if (!vma->vm_private_data) /* vma lock required for sharing */ 70098d9bfb26SMike Kravetz return false; 70108d9bfb26SMike Kravetz if (!range_in_vma(vma, start, end)) 70118d9bfb26SMike Kravetz return false; 70128d9bfb26SMike Kravetz return true; 70138d9bfb26SMike Kravetz } 70148d9bfb26SMike Kravetz 70153212b535SSteve Capper /* 7016017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 7017017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 7018017b1660SMike Kravetz * shared pmd mappings. 7019017b1660SMike Kravetz */ 7020017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 7021017b1660SMike Kravetz unsigned long *start, unsigned long *end) 7022017b1660SMike Kravetz { 7023a1ba9da8SLi Xinhai unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), 7024a1ba9da8SLi Xinhai v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 7025017b1660SMike Kravetz 7026a1ba9da8SLi Xinhai /* 7027f0953a1bSIngo Molnar * vma needs to span at least one aligned PUD size, and the range 7028f0953a1bSIngo Molnar * must be at least partially within in. 7029a1ba9da8SLi Xinhai */ 7030a1ba9da8SLi Xinhai if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || 7031a1ba9da8SLi Xinhai (*end <= v_start) || (*start >= v_end)) 7032017b1660SMike Kravetz return; 7033017b1660SMike Kravetz 703475802ca6SPeter Xu /* Extend the range to be PUD aligned for a worst case scenario */ 7035a1ba9da8SLi Xinhai if (*start > v_start) 7036a1ba9da8SLi Xinhai *start = ALIGN_DOWN(*start, PUD_SIZE); 7037017b1660SMike Kravetz 7038a1ba9da8SLi Xinhai if (*end < v_end) 7039a1ba9da8SLi Xinhai *end = ALIGN(*end, PUD_SIZE); 7040017b1660SMike Kravetz } 7041017b1660SMike Kravetz 7042017b1660SMike Kravetz /* 70433212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 70443212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 70453212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 70463a47c54fSMike Kravetz * code much cleaner. pmd allocation is essential for the shared case because 70473a47c54fSMike Kravetz * pud has to be populated inside the same i_mmap_rwsem section - otherwise 70483a47c54fSMike Kravetz * racing tasks could either miss the sharing (see huge_pte_offset) or select a 70493a47c54fSMike Kravetz * bad pmd for sharing. 70503212b535SSteve Capper */ 7051aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 7052aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 70533212b535SSteve Capper { 70543212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 70553212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 70563212b535SSteve Capper vma->vm_pgoff; 70573212b535SSteve Capper struct vm_area_struct *svma; 70583212b535SSteve Capper unsigned long saddr; 70593212b535SSteve Capper pte_t *spte = NULL; 70603212b535SSteve Capper pte_t *pte; 7061cb900f41SKirill A. Shutemov spinlock_t *ptl; 70623212b535SSteve Capper 70633a47c54fSMike Kravetz i_mmap_lock_read(mapping); 70643212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 70653212b535SSteve Capper if (svma == vma) 70663212b535SSteve Capper continue; 70673212b535SSteve Capper 70683212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 70693212b535SSteve Capper if (saddr) { 70709c67a207SPeter Xu spte = hugetlb_walk(svma, saddr, 70717868a208SPunit Agrawal vma_mmu_pagesize(svma)); 70723212b535SSteve Capper if (spte) { 70733212b535SSteve Capper get_page(virt_to_page(spte)); 70743212b535SSteve Capper break; 70753212b535SSteve Capper } 70763212b535SSteve Capper } 70773212b535SSteve Capper } 70783212b535SSteve Capper 70793212b535SSteve Capper if (!spte) 70803212b535SSteve Capper goto out; 70813212b535SSteve Capper 70828bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 7083dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 70843212b535SSteve Capper pud_populate(mm, pud, 70853212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 7086c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 7087dc6c9a35SKirill A. Shutemov } else { 70883212b535SSteve Capper put_page(virt_to_page(spte)); 7089dc6c9a35SKirill A. Shutemov } 7090cb900f41SKirill A. Shutemov spin_unlock(ptl); 70913212b535SSteve Capper out: 70923212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 70933a47c54fSMike Kravetz i_mmap_unlock_read(mapping); 70943212b535SSteve Capper return pte; 70953212b535SSteve Capper } 70963212b535SSteve Capper 70973212b535SSteve Capper /* 70983212b535SSteve Capper * unmap huge page backed by shared pte. 70993212b535SSteve Capper * 71003212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 71013212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 71023212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 71033212b535SSteve Capper * 71043a47c54fSMike Kravetz * Called with page table lock held. 71053212b535SSteve Capper * 71063212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 71073212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 71083212b535SSteve Capper */ 710934ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 71104ddb4d91SMike Kravetz unsigned long addr, pte_t *ptep) 71113212b535SSteve Capper { 71124ddb4d91SMike Kravetz pgd_t *pgd = pgd_offset(mm, addr); 71134ddb4d91SMike Kravetz p4d_t *p4d = p4d_offset(pgd, addr); 71144ddb4d91SMike Kravetz pud_t *pud = pud_offset(p4d, addr); 71153212b535SSteve Capper 711634ae204fSMike Kravetz i_mmap_assert_write_locked(vma->vm_file->f_mapping); 711740549ba8SMike Kravetz hugetlb_vma_assert_locked(vma); 71183212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 71193212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 71203212b535SSteve Capper return 0; 71213212b535SSteve Capper 71223212b535SSteve Capper pud_clear(pud); 71233212b535SSteve Capper put_page(virt_to_page(ptep)); 7124dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 71253212b535SSteve Capper return 1; 71263212b535SSteve Capper } 7127c1991e07SPeter Xu 71289e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 71298d9bfb26SMike Kravetz 7130aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 7131aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 71329e5fc74cSSteve Capper { 71339e5fc74cSSteve Capper return NULL; 71349e5fc74cSSteve Capper } 7135e81f2d22SZhang Zhen 713634ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 71374ddb4d91SMike Kravetz unsigned long addr, pte_t *ptep) 7138e81f2d22SZhang Zhen { 7139e81f2d22SZhang Zhen return 0; 7140e81f2d22SZhang Zhen } 7141017b1660SMike Kravetz 7142017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 7143017b1660SMike Kravetz unsigned long *start, unsigned long *end) 7144017b1660SMike Kravetz { 7145017b1660SMike Kravetz } 7146c1991e07SPeter Xu 7147c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 7148c1991e07SPeter Xu { 7149c1991e07SPeter Xu return false; 7150c1991e07SPeter Xu } 71513212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 71523212b535SSteve Capper 71539e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 7154aec44e0fSPeter Xu pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, 71559e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 71569e5fc74cSSteve Capper { 71579e5fc74cSSteve Capper pgd_t *pgd; 7158c2febafcSKirill A. Shutemov p4d_t *p4d; 71599e5fc74cSSteve Capper pud_t *pud; 71609e5fc74cSSteve Capper pte_t *pte = NULL; 71619e5fc74cSSteve Capper 71629e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 7163f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 7164f4f0a3d8SKirill A. Shutemov if (!p4d) 7165f4f0a3d8SKirill A. Shutemov return NULL; 7166c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 71679e5fc74cSSteve Capper if (pud) { 71689e5fc74cSSteve Capper if (sz == PUD_SIZE) { 71699e5fc74cSSteve Capper pte = (pte_t *)pud; 71709e5fc74cSSteve Capper } else { 71719e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 7172c1991e07SPeter Xu if (want_pmd_share(vma, addr) && pud_none(*pud)) 7173aec44e0fSPeter Xu pte = huge_pmd_share(mm, vma, addr, pud); 71749e5fc74cSSteve Capper else 71759e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 71769e5fc74cSSteve Capper } 71779e5fc74cSSteve Capper } 71784e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 71799e5fc74cSSteve Capper 71809e5fc74cSSteve Capper return pte; 71819e5fc74cSSteve Capper } 71829e5fc74cSSteve Capper 71839b19df29SPunit Agrawal /* 71849b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 71859b19df29SPunit Agrawal * entry at address @addr 71869b19df29SPunit Agrawal * 71878ac0b81aSLi Xinhai * Return: Pointer to page table entry (PUD or PMD) for 71888ac0b81aSLi Xinhai * address @addr, or NULL if a !p*d_present() entry is encountered and the 71899b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 71909b19df29SPunit Agrawal * table. 71919b19df29SPunit Agrawal */ 71927868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 71937868a208SPunit Agrawal unsigned long addr, unsigned long sz) 71949e5fc74cSSteve Capper { 71959e5fc74cSSteve Capper pgd_t *pgd; 7196c2febafcSKirill A. Shutemov p4d_t *p4d; 71978ac0b81aSLi Xinhai pud_t *pud; 71988ac0b81aSLi Xinhai pmd_t *pmd; 71999e5fc74cSSteve Capper 72009e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 7201c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 7202c2febafcSKirill A. Shutemov return NULL; 7203c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 7204c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 7205c2febafcSKirill A. Shutemov return NULL; 72069b19df29SPunit Agrawal 7207c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 72088ac0b81aSLi Xinhai if (sz == PUD_SIZE) 72098ac0b81aSLi Xinhai /* must be pud huge, non-present or none */ 72109e5fc74cSSteve Capper return (pte_t *)pud; 72118ac0b81aSLi Xinhai if (!pud_present(*pud)) 72128ac0b81aSLi Xinhai return NULL; 72138ac0b81aSLi Xinhai /* must have a valid entry and size to go further */ 72149b19df29SPunit Agrawal 72159e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 72168ac0b81aSLi Xinhai /* must be pmd huge, non-present or none */ 72179e5fc74cSSteve Capper return (pte_t *)pmd; 72189e5fc74cSSteve Capper } 72199e5fc74cSSteve Capper 7220e95a9851SMike Kravetz /* 7221e95a9851SMike Kravetz * Return a mask that can be used to update an address to the last huge 7222e95a9851SMike Kravetz * page in a page table page mapping size. Used to skip non-present 7223e95a9851SMike Kravetz * page table entries when linearly scanning address ranges. Architectures 7224e95a9851SMike Kravetz * with unique huge page to page table relationships can define their own 7225e95a9851SMike Kravetz * version of this routine. 7226e95a9851SMike Kravetz */ 7227e95a9851SMike Kravetz unsigned long hugetlb_mask_last_page(struct hstate *h) 7228e95a9851SMike Kravetz { 7229e95a9851SMike Kravetz unsigned long hp_size = huge_page_size(h); 7230e95a9851SMike Kravetz 7231e95a9851SMike Kravetz if (hp_size == PUD_SIZE) 7232e95a9851SMike Kravetz return P4D_SIZE - PUD_SIZE; 7233e95a9851SMike Kravetz else if (hp_size == PMD_SIZE) 7234e95a9851SMike Kravetz return PUD_SIZE - PMD_SIZE; 7235e95a9851SMike Kravetz else 7236e95a9851SMike Kravetz return 0UL; 7237e95a9851SMike Kravetz } 7238e95a9851SMike Kravetz 7239e95a9851SMike Kravetz #else 7240e95a9851SMike Kravetz 7241e95a9851SMike Kravetz /* See description above. Architectures can provide their own version. */ 7242e95a9851SMike Kravetz __weak unsigned long hugetlb_mask_last_page(struct hstate *h) 7243e95a9851SMike Kravetz { 72444ddb4d91SMike Kravetz #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 72454ddb4d91SMike Kravetz if (huge_page_size(h) == PMD_SIZE) 72464ddb4d91SMike Kravetz return PUD_SIZE - PMD_SIZE; 72474ddb4d91SMike Kravetz #endif 7248e95a9851SMike Kravetz return 0UL; 7249e95a9851SMike Kravetz } 7250e95a9851SMike Kravetz 725161f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 725261f77edaSNaoya Horiguchi 725361f77edaSNaoya Horiguchi /* 725461f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 725561f77edaSNaoya Horiguchi * behavior. 725661f77edaSNaoya Horiguchi */ 72576aa3a920SSidhartha Kumar int isolate_hugetlb(struct folio *folio, struct list_head *list) 725831caf665SNaoya Horiguchi { 72597ce82f4cSMiaohe Lin int ret = 0; 7260bcc54222SNaoya Horiguchi 7261db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 72626aa3a920SSidhartha Kumar if (!folio_test_hugetlb(folio) || 72636aa3a920SSidhartha Kumar !folio_test_hugetlb_migratable(folio) || 72646aa3a920SSidhartha Kumar !folio_try_get(folio)) { 72657ce82f4cSMiaohe Lin ret = -EBUSY; 7266bcc54222SNaoya Horiguchi goto unlock; 7267bcc54222SNaoya Horiguchi } 72686aa3a920SSidhartha Kumar folio_clear_hugetlb_migratable(folio); 72696aa3a920SSidhartha Kumar list_move_tail(&folio->lru, list); 7270bcc54222SNaoya Horiguchi unlock: 7271db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7272bcc54222SNaoya Horiguchi return ret; 727331caf665SNaoya Horiguchi } 727431caf665SNaoya Horiguchi 727504bac040SSidhartha Kumar int get_hwpoison_hugetlb_folio(struct folio *folio, bool *hugetlb, bool unpoison) 727625182f05SNaoya Horiguchi { 727725182f05SNaoya Horiguchi int ret = 0; 727825182f05SNaoya Horiguchi 727925182f05SNaoya Horiguchi *hugetlb = false; 728025182f05SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 728104bac040SSidhartha Kumar if (folio_test_hugetlb(folio)) { 728225182f05SNaoya Horiguchi *hugetlb = true; 728304bac040SSidhartha Kumar if (folio_test_hugetlb_freed(folio)) 7284b283d983SNaoya Horiguchi ret = 0; 728504bac040SSidhartha Kumar else if (folio_test_hugetlb_migratable(folio) || unpoison) 728604bac040SSidhartha Kumar ret = folio_try_get(folio); 72870ed950d1SNaoya Horiguchi else 72880ed950d1SNaoya Horiguchi ret = -EBUSY; 728925182f05SNaoya Horiguchi } 729025182f05SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 729125182f05SNaoya Horiguchi return ret; 729225182f05SNaoya Horiguchi } 729325182f05SNaoya Horiguchi 7294e591ef7dSNaoya Horiguchi int get_huge_page_for_hwpoison(unsigned long pfn, int flags, 7295e591ef7dSNaoya Horiguchi bool *migratable_cleared) 7296405ce051SNaoya Horiguchi { 7297405ce051SNaoya Horiguchi int ret; 7298405ce051SNaoya Horiguchi 7299405ce051SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 7300e591ef7dSNaoya Horiguchi ret = __get_huge_page_for_hwpoison(pfn, flags, migratable_cleared); 7301405ce051SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 7302405ce051SNaoya Horiguchi return ret; 7303405ce051SNaoya Horiguchi } 7304405ce051SNaoya Horiguchi 730531caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 730631caf665SNaoya Horiguchi { 7307db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 73088f251a3dSMike Kravetz SetHPageMigratable(page); 730931caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 7310db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 731131caf665SNaoya Horiguchi put_page(page); 731231caf665SNaoya Horiguchi } 7313ab5ac90aSMichal Hocko 7314345c62d1SSidhartha Kumar void move_hugetlb_state(struct folio *old_folio, struct folio *new_folio, int reason) 7315ab5ac90aSMichal Hocko { 7316345c62d1SSidhartha Kumar struct hstate *h = folio_hstate(old_folio); 7317ab5ac90aSMichal Hocko 7318345c62d1SSidhartha Kumar hugetlb_cgroup_migrate(old_folio, new_folio); 7319345c62d1SSidhartha Kumar set_page_owner_migrate_reason(&new_folio->page, reason); 7320ab5ac90aSMichal Hocko 7321ab5ac90aSMichal Hocko /* 7322345c62d1SSidhartha Kumar * transfer temporary state of the new hugetlb folio. This is 7323ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 7324ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 7325ab5ac90aSMichal Hocko * the temporary status. 7326ab5ac90aSMichal Hocko * 7327ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 7328ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 7329ab5ac90aSMichal Hocko * the per-node's. 7330ab5ac90aSMichal Hocko */ 7331345c62d1SSidhartha Kumar if (folio_test_hugetlb_temporary(new_folio)) { 7332345c62d1SSidhartha Kumar int old_nid = folio_nid(old_folio); 7333345c62d1SSidhartha Kumar int new_nid = folio_nid(new_folio); 7334ab5ac90aSMichal Hocko 7335345c62d1SSidhartha Kumar folio_set_hugetlb_temporary(old_folio); 7336345c62d1SSidhartha Kumar folio_clear_hugetlb_temporary(new_folio); 7337345c62d1SSidhartha Kumar 7338ab5ac90aSMichal Hocko 73395af1ab1dSMiaohe Lin /* 73405af1ab1dSMiaohe Lin * There is no need to transfer the per-node surplus state 73415af1ab1dSMiaohe Lin * when we do not cross the node. 73425af1ab1dSMiaohe Lin */ 73435af1ab1dSMiaohe Lin if (new_nid == old_nid) 73445af1ab1dSMiaohe Lin return; 7345db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 7346ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 7347ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 7348ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 7349ab5ac90aSMichal Hocko } 7350db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7351ab5ac90aSMichal Hocko } 7352ab5ac90aSMichal Hocko } 7353cf11e85fSRoman Gushchin 7354b30c14cdSJames Houghton static void hugetlb_unshare_pmds(struct vm_area_struct *vma, 7355b30c14cdSJames Houghton unsigned long start, 7356b30c14cdSJames Houghton unsigned long end) 73576dfeaff9SPeter Xu { 73586dfeaff9SPeter Xu struct hstate *h = hstate_vma(vma); 73596dfeaff9SPeter Xu unsigned long sz = huge_page_size(h); 73606dfeaff9SPeter Xu struct mm_struct *mm = vma->vm_mm; 73616dfeaff9SPeter Xu struct mmu_notifier_range range; 7362b30c14cdSJames Houghton unsigned long address; 73636dfeaff9SPeter Xu spinlock_t *ptl; 73646dfeaff9SPeter Xu pte_t *ptep; 73656dfeaff9SPeter Xu 73666dfeaff9SPeter Xu if (!(vma->vm_flags & VM_MAYSHARE)) 73676dfeaff9SPeter Xu return; 73686dfeaff9SPeter Xu 73696dfeaff9SPeter Xu if (start >= end) 73706dfeaff9SPeter Xu return; 73716dfeaff9SPeter Xu 73729c8bbfacSBaolin Wang flush_cache_range(vma, start, end); 73736dfeaff9SPeter Xu /* 73746dfeaff9SPeter Xu * No need to call adjust_range_if_pmd_sharing_possible(), because 73756dfeaff9SPeter Xu * we have already done the PUD_SIZE alignment. 73766dfeaff9SPeter Xu */ 73777d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, 73786dfeaff9SPeter Xu start, end); 73796dfeaff9SPeter Xu mmu_notifier_invalidate_range_start(&range); 738040549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 73816dfeaff9SPeter Xu i_mmap_lock_write(vma->vm_file->f_mapping); 73826dfeaff9SPeter Xu for (address = start; address < end; address += PUD_SIZE) { 73839c67a207SPeter Xu ptep = hugetlb_walk(vma, address, sz); 73846dfeaff9SPeter Xu if (!ptep) 73856dfeaff9SPeter Xu continue; 73866dfeaff9SPeter Xu ptl = huge_pte_lock(h, mm, ptep); 73874ddb4d91SMike Kravetz huge_pmd_unshare(mm, vma, address, ptep); 73886dfeaff9SPeter Xu spin_unlock(ptl); 73896dfeaff9SPeter Xu } 73906dfeaff9SPeter Xu flush_hugetlb_tlb_range(vma, start, end); 73916dfeaff9SPeter Xu i_mmap_unlock_write(vma->vm_file->f_mapping); 739240549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 73936dfeaff9SPeter Xu /* 73946dfeaff9SPeter Xu * No need to call mmu_notifier_invalidate_range(), see 7395ee65728eSMike Rapoport * Documentation/mm/mmu_notifier.rst. 73966dfeaff9SPeter Xu */ 73976dfeaff9SPeter Xu mmu_notifier_invalidate_range_end(&range); 73986dfeaff9SPeter Xu } 73996dfeaff9SPeter Xu 7400b30c14cdSJames Houghton /* 7401b30c14cdSJames Houghton * This function will unconditionally remove all the shared pmd pgtable entries 7402b30c14cdSJames Houghton * within the specific vma for a hugetlbfs memory range. 7403b30c14cdSJames Houghton */ 7404b30c14cdSJames Houghton void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) 7405b30c14cdSJames Houghton { 7406b30c14cdSJames Houghton hugetlb_unshare_pmds(vma, ALIGN(vma->vm_start, PUD_SIZE), 7407b30c14cdSJames Houghton ALIGN_DOWN(vma->vm_end, PUD_SIZE)); 7408b30c14cdSJames Houghton } 7409b30c14cdSJames Houghton 7410cf11e85fSRoman Gushchin #ifdef CONFIG_CMA 7411cf11e85fSRoman Gushchin static bool cma_reserve_called __initdata; 7412cf11e85fSRoman Gushchin 7413cf11e85fSRoman Gushchin static int __init cmdline_parse_hugetlb_cma(char *p) 7414cf11e85fSRoman Gushchin { 741538e719abSBaolin Wang int nid, count = 0; 741638e719abSBaolin Wang unsigned long tmp; 741738e719abSBaolin Wang char *s = p; 741838e719abSBaolin Wang 741938e719abSBaolin Wang while (*s) { 742038e719abSBaolin Wang if (sscanf(s, "%lu%n", &tmp, &count) != 1) 742138e719abSBaolin Wang break; 742238e719abSBaolin Wang 742338e719abSBaolin Wang if (s[count] == ':') { 7424f9317f77SMike Kravetz if (tmp >= MAX_NUMNODES) 742538e719abSBaolin Wang break; 7426f9317f77SMike Kravetz nid = array_index_nospec(tmp, MAX_NUMNODES); 742738e719abSBaolin Wang 742838e719abSBaolin Wang s += count + 1; 742938e719abSBaolin Wang tmp = memparse(s, &s); 743038e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = tmp; 743138e719abSBaolin Wang hugetlb_cma_size += tmp; 743238e719abSBaolin Wang 743338e719abSBaolin Wang /* 743438e719abSBaolin Wang * Skip the separator if have one, otherwise 743538e719abSBaolin Wang * break the parsing. 743638e719abSBaolin Wang */ 743738e719abSBaolin Wang if (*s == ',') 743838e719abSBaolin Wang s++; 743938e719abSBaolin Wang else 744038e719abSBaolin Wang break; 744138e719abSBaolin Wang } else { 7442cf11e85fSRoman Gushchin hugetlb_cma_size = memparse(p, &p); 744338e719abSBaolin Wang break; 744438e719abSBaolin Wang } 744538e719abSBaolin Wang } 744638e719abSBaolin Wang 7447cf11e85fSRoman Gushchin return 0; 7448cf11e85fSRoman Gushchin } 7449cf11e85fSRoman Gushchin 7450cf11e85fSRoman Gushchin early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); 7451cf11e85fSRoman Gushchin 7452cf11e85fSRoman Gushchin void __init hugetlb_cma_reserve(int order) 7453cf11e85fSRoman Gushchin { 7454cf11e85fSRoman Gushchin unsigned long size, reserved, per_node; 745538e719abSBaolin Wang bool node_specific_cma_alloc = false; 7456cf11e85fSRoman Gushchin int nid; 7457cf11e85fSRoman Gushchin 7458cf11e85fSRoman Gushchin cma_reserve_called = true; 7459cf11e85fSRoman Gushchin 7460cf11e85fSRoman Gushchin if (!hugetlb_cma_size) 7461cf11e85fSRoman Gushchin return; 7462cf11e85fSRoman Gushchin 746338e719abSBaolin Wang for (nid = 0; nid < MAX_NUMNODES; nid++) { 746438e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 746538e719abSBaolin Wang continue; 746638e719abSBaolin Wang 746730a51400SPeng Liu if (!node_online(nid)) { 746838e719abSBaolin Wang pr_warn("hugetlb_cma: invalid node %d specified\n", nid); 746938e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 747038e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 747138e719abSBaolin Wang continue; 747238e719abSBaolin Wang } 747338e719abSBaolin Wang 747438e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) { 747538e719abSBaolin Wang pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n", 747638e719abSBaolin Wang nid, (PAGE_SIZE << order) / SZ_1M); 747738e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 747838e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 747938e719abSBaolin Wang } else { 748038e719abSBaolin Wang node_specific_cma_alloc = true; 748138e719abSBaolin Wang } 748238e719abSBaolin Wang } 748338e719abSBaolin Wang 748438e719abSBaolin Wang /* Validate the CMA size again in case some invalid nodes specified. */ 748538e719abSBaolin Wang if (!hugetlb_cma_size) 748638e719abSBaolin Wang return; 748738e719abSBaolin Wang 7488cf11e85fSRoman Gushchin if (hugetlb_cma_size < (PAGE_SIZE << order)) { 7489cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", 7490cf11e85fSRoman Gushchin (PAGE_SIZE << order) / SZ_1M); 7491a01f4390SMike Kravetz hugetlb_cma_size = 0; 7492cf11e85fSRoman Gushchin return; 7493cf11e85fSRoman Gushchin } 7494cf11e85fSRoman Gushchin 749538e719abSBaolin Wang if (!node_specific_cma_alloc) { 7496cf11e85fSRoman Gushchin /* 7497cf11e85fSRoman Gushchin * If 3 GB area is requested on a machine with 4 numa nodes, 7498cf11e85fSRoman Gushchin * let's allocate 1 GB on first three nodes and ignore the last one. 7499cf11e85fSRoman Gushchin */ 7500cf11e85fSRoman Gushchin per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); 7501cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", 7502cf11e85fSRoman Gushchin hugetlb_cma_size / SZ_1M, per_node / SZ_1M); 750338e719abSBaolin Wang } 7504cf11e85fSRoman Gushchin 7505cf11e85fSRoman Gushchin reserved = 0; 750630a51400SPeng Liu for_each_online_node(nid) { 7507cf11e85fSRoman Gushchin int res; 75082281f797SBarry Song char name[CMA_MAX_NAME]; 7509cf11e85fSRoman Gushchin 751038e719abSBaolin Wang if (node_specific_cma_alloc) { 751138e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 751238e719abSBaolin Wang continue; 751338e719abSBaolin Wang 751438e719abSBaolin Wang size = hugetlb_cma_size_in_node[nid]; 751538e719abSBaolin Wang } else { 7516cf11e85fSRoman Gushchin size = min(per_node, hugetlb_cma_size - reserved); 751738e719abSBaolin Wang } 751838e719abSBaolin Wang 7519cf11e85fSRoman Gushchin size = round_up(size, PAGE_SIZE << order); 7520cf11e85fSRoman Gushchin 75212281f797SBarry Song snprintf(name, sizeof(name), "hugetlb%d", nid); 7522a01f4390SMike Kravetz /* 7523a01f4390SMike Kravetz * Note that 'order per bit' is based on smallest size that 7524a01f4390SMike Kravetz * may be returned to CMA allocator in the case of 7525a01f4390SMike Kravetz * huge page demotion. 7526a01f4390SMike Kravetz */ 7527a01f4390SMike Kravetz res = cma_declare_contiguous_nid(0, size, 0, 7528a01f4390SMike Kravetz PAGE_SIZE << HUGETLB_PAGE_ORDER, 752929d0f41dSBarry Song 0, false, name, 7530cf11e85fSRoman Gushchin &hugetlb_cma[nid], nid); 7531cf11e85fSRoman Gushchin if (res) { 7532cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: reservation failed: err %d, node %d", 7533cf11e85fSRoman Gushchin res, nid); 7534cf11e85fSRoman Gushchin continue; 7535cf11e85fSRoman Gushchin } 7536cf11e85fSRoman Gushchin 7537cf11e85fSRoman Gushchin reserved += size; 7538cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", 7539cf11e85fSRoman Gushchin size / SZ_1M, nid); 7540cf11e85fSRoman Gushchin 7541cf11e85fSRoman Gushchin if (reserved >= hugetlb_cma_size) 7542cf11e85fSRoman Gushchin break; 7543cf11e85fSRoman Gushchin } 7544a01f4390SMike Kravetz 7545a01f4390SMike Kravetz if (!reserved) 7546a01f4390SMike Kravetz /* 7547a01f4390SMike Kravetz * hugetlb_cma_size is used to determine if allocations from 7548a01f4390SMike Kravetz * cma are possible. Set to zero if no cma regions are set up. 7549a01f4390SMike Kravetz */ 7550a01f4390SMike Kravetz hugetlb_cma_size = 0; 7551cf11e85fSRoman Gushchin } 7552cf11e85fSRoman Gushchin 7553263b8998SMiaohe Lin static void __init hugetlb_cma_check(void) 7554cf11e85fSRoman Gushchin { 7555cf11e85fSRoman Gushchin if (!hugetlb_cma_size || cma_reserve_called) 7556cf11e85fSRoman Gushchin return; 7557cf11e85fSRoman Gushchin 7558cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); 7559cf11e85fSRoman Gushchin } 7560cf11e85fSRoman Gushchin 7561cf11e85fSRoman Gushchin #endif /* CONFIG_CMA */ 7562