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 128594310cbcSAnshuman Khandual static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) 1286bf50bab2SNaoya Horiguchi { 1287bf50bab2SNaoya Horiguchi struct page *page; 12881a08ae36SPavel Tatashin bool pin = !!(current->flags & PF_MEMALLOC_PIN); 1289bf50bab2SNaoya Horiguchi 12909487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1291bbe88753SJoonsoo Kim list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { 12926077c943SAlex Sierra if (pin && !is_longterm_pinnable_page(page)) 1293bbe88753SJoonsoo Kim continue; 1294bbe88753SJoonsoo Kim 12956664bfc8SWei Yang if (PageHWPoison(page)) 12966664bfc8SWei Yang continue; 1297bbe88753SJoonsoo Kim 12980edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 1299a9869b83SNaoya Horiguchi set_page_refcounted(page); 13006c037149SMike Kravetz ClearHPageFreed(page); 1301bf50bab2SNaoya Horiguchi h->free_huge_pages--; 1302bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 1303bf50bab2SNaoya Horiguchi return page; 1304bf50bab2SNaoya Horiguchi } 1305bf50bab2SNaoya Horiguchi 13066664bfc8SWei Yang return NULL; 13076664bfc8SWei Yang } 13086664bfc8SWei Yang 13093e59fcb0SMichal Hocko static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, 13103e59fcb0SMichal Hocko nodemask_t *nmask) 131194310cbcSAnshuman Khandual { 13123e59fcb0SMichal Hocko unsigned int cpuset_mems_cookie; 13133e59fcb0SMichal Hocko struct zonelist *zonelist; 13143e59fcb0SMichal Hocko struct zone *zone; 13153e59fcb0SMichal Hocko struct zoneref *z; 131698fa15f3SAnshuman Khandual int node = NUMA_NO_NODE; 13173e59fcb0SMichal Hocko 13183e59fcb0SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 13193e59fcb0SMichal Hocko 13203e59fcb0SMichal Hocko retry_cpuset: 13213e59fcb0SMichal Hocko cpuset_mems_cookie = read_mems_allowed_begin(); 13223e59fcb0SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 132394310cbcSAnshuman Khandual struct page *page; 132494310cbcSAnshuman Khandual 13253e59fcb0SMichal Hocko if (!cpuset_zone_allowed(zone, gfp_mask)) 13263e59fcb0SMichal Hocko continue; 13273e59fcb0SMichal Hocko /* 13283e59fcb0SMichal Hocko * no need to ask again on the same node. Pool is node rather than 13293e59fcb0SMichal Hocko * zone aware 13303e59fcb0SMichal Hocko */ 13313e59fcb0SMichal Hocko if (zone_to_nid(zone) == node) 13323e59fcb0SMichal Hocko continue; 13333e59fcb0SMichal Hocko node = zone_to_nid(zone); 133494310cbcSAnshuman Khandual 133594310cbcSAnshuman Khandual page = dequeue_huge_page_node_exact(h, node); 133694310cbcSAnshuman Khandual if (page) 133794310cbcSAnshuman Khandual return page; 133894310cbcSAnshuman Khandual } 13393e59fcb0SMichal Hocko if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 13403e59fcb0SMichal Hocko goto retry_cpuset; 13413e59fcb0SMichal Hocko 134294310cbcSAnshuman Khandual return NULL; 134394310cbcSAnshuman Khandual } 134494310cbcSAnshuman Khandual 13458346d69dSXin Hao static unsigned long available_huge_pages(struct hstate *h) 13468346d69dSXin Hao { 13478346d69dSXin Hao return h->free_huge_pages - h->resv_huge_pages; 13488346d69dSXin Hao } 13498346d69dSXin Hao 1350a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 1351a5516438SAndi Kleen struct vm_area_struct *vma, 1352af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 1353af0ed73eSJoonsoo Kim long chg) 13541da177e4SLinus Torvalds { 1355cfcaa66fSBen Widawsky struct page *page = NULL; 1356480eccf9SLee Schermerhorn struct mempolicy *mpol; 135704ec6264SVlastimil Babka gfp_t gfp_mask; 13583e59fcb0SMichal Hocko nodemask_t *nodemask; 135904ec6264SVlastimil Babka int nid; 13601da177e4SLinus Torvalds 1361a1e78772SMel Gorman /* 1362a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 1363a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 1364a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 1365a1e78772SMel Gorman */ 13668346d69dSXin Hao if (!vma_has_reserves(vma, chg) && !available_huge_pages(h)) 1367c0ff7453SMiao Xie goto err; 1368a1e78772SMel Gorman 136904f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 13708346d69dSXin Hao if (avoid_reserve && !available_huge_pages(h)) 13716eab04a8SJustin P. Mattock goto err; 137204f2cbe3SMel Gorman 137304ec6264SVlastimil Babka gfp_mask = htlb_alloc_mask(h); 137404ec6264SVlastimil Babka nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 1375cfcaa66fSBen Widawsky 1376cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 13773e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 1378cfcaa66fSBen Widawsky 1379cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 1380cfcaa66fSBen Widawsky nodemask = NULL; 1381cfcaa66fSBen Widawsky } 1382cfcaa66fSBen Widawsky 1383cfcaa66fSBen Widawsky if (!page) 1384cfcaa66fSBen Widawsky page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 1385cfcaa66fSBen Widawsky 13863e59fcb0SMichal Hocko if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { 1387d6995da3SMike Kravetz SetHPageRestoreReserve(page); 1388a63884e9SJoonsoo Kim h->resv_huge_pages--; 1389bf50bab2SNaoya Horiguchi } 1390cc9a6c87SMel Gorman 1391cc9a6c87SMel Gorman mpol_cond_put(mpol); 1392cc9a6c87SMel Gorman return page; 1393cc9a6c87SMel Gorman 1394c0ff7453SMiao Xie err: 1395cc9a6c87SMel Gorman return NULL; 13961da177e4SLinus Torvalds } 13971da177e4SLinus Torvalds 13981cac6f2cSLuiz Capitulino /* 13991cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 14001cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 14011cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 14021cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 14031cac6f2cSLuiz Capitulino * node for alloc or free. 14041cac6f2cSLuiz Capitulino */ 14051cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 14061cac6f2cSLuiz Capitulino { 14070edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 14081cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 14091cac6f2cSLuiz Capitulino 14101cac6f2cSLuiz Capitulino return nid; 14111cac6f2cSLuiz Capitulino } 14121cac6f2cSLuiz Capitulino 14131cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 14141cac6f2cSLuiz Capitulino { 14151cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 14161cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 14171cac6f2cSLuiz Capitulino return nid; 14181cac6f2cSLuiz Capitulino } 14191cac6f2cSLuiz Capitulino 14201cac6f2cSLuiz Capitulino /* 14211cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 14221cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 14231cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 14241cac6f2cSLuiz Capitulino * mask. 14251cac6f2cSLuiz Capitulino */ 14261cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 14271cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 14281cac6f2cSLuiz Capitulino { 14291cac6f2cSLuiz Capitulino int nid; 14301cac6f2cSLuiz Capitulino 14311cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 14321cac6f2cSLuiz Capitulino 14331cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 14341cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 14351cac6f2cSLuiz Capitulino 14361cac6f2cSLuiz Capitulino return nid; 14371cac6f2cSLuiz Capitulino } 14381cac6f2cSLuiz Capitulino 14391cac6f2cSLuiz Capitulino /* 144010c6ec49SMike Kravetz * helper for remove_pool_huge_page() - return the previously saved 14411cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 14421cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 14431cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 14441cac6f2cSLuiz Capitulino */ 14451cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 14461cac6f2cSLuiz Capitulino { 14471cac6f2cSLuiz Capitulino int nid; 14481cac6f2cSLuiz Capitulino 14491cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 14501cac6f2cSLuiz Capitulino 14511cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 14521cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 14531cac6f2cSLuiz Capitulino 14541cac6f2cSLuiz Capitulino return nid; 14551cac6f2cSLuiz Capitulino } 14561cac6f2cSLuiz Capitulino 14571cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 14581cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 14591cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 14601cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 14611cac6f2cSLuiz Capitulino nr_nodes--) 14621cac6f2cSLuiz Capitulino 14631cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 14641cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 14651cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 14661cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 14671cac6f2cSLuiz Capitulino nr_nodes--) 14681cac6f2cSLuiz Capitulino 14698531fc6fSMike Kravetz /* used to demote non-gigantic_huge pages as well */ 1470911565b8SSidhartha Kumar static void __destroy_compound_gigantic_folio(struct folio *folio, 147134d9e35bSMike Kravetz unsigned int order, bool demote) 1472944d9fecSLuiz Capitulino { 1473944d9fecSLuiz Capitulino int i; 1474944d9fecSLuiz Capitulino int nr_pages = 1 << order; 147514455eabSCheng Li struct page *p; 1476944d9fecSLuiz Capitulino 147746f27228SMatthew Wilcox (Oracle) atomic_set(&folio->_entire_mapcount, 0); 1478eec20426SMatthew Wilcox (Oracle) atomic_set(&folio->_nr_pages_mapped, 0); 147994688e8eSMatthew Wilcox (Oracle) atomic_set(&folio->_pincount, 0); 148047e29d32SJohn Hubbard 148114455eabSCheng Li for (i = 1; i < nr_pages; i++) { 1482911565b8SSidhartha Kumar p = folio_page(folio, i); 1483a01f4390SMike Kravetz p->mapping = NULL; 14841d798ca3SKirill A. Shutemov clear_compound_head(p); 148534d9e35bSMike Kravetz if (!demote) 1486944d9fecSLuiz Capitulino set_page_refcounted(p); 1487944d9fecSLuiz Capitulino } 1488944d9fecSLuiz Capitulino 148904a42e72SSidhartha Kumar folio_set_order(folio, 0); 1490911565b8SSidhartha Kumar __folio_clear_head(folio); 1491944d9fecSLuiz Capitulino } 1492944d9fecSLuiz Capitulino 1493911565b8SSidhartha Kumar static void destroy_compound_hugetlb_folio_for_demote(struct folio *folio, 14948531fc6fSMike Kravetz unsigned int order) 14958531fc6fSMike Kravetz { 1496911565b8SSidhartha Kumar __destroy_compound_gigantic_folio(folio, order, true); 14978531fc6fSMike Kravetz } 14988531fc6fSMike Kravetz 14998531fc6fSMike Kravetz #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 1500911565b8SSidhartha Kumar static void destroy_compound_gigantic_folio(struct folio *folio, 150134d9e35bSMike Kravetz unsigned int order) 150234d9e35bSMike Kravetz { 1503911565b8SSidhartha Kumar __destroy_compound_gigantic_folio(folio, order, false); 150434d9e35bSMike Kravetz } 150534d9e35bSMike Kravetz 15067f325a8dSSidhartha Kumar static void free_gigantic_folio(struct folio *folio, unsigned int order) 1507944d9fecSLuiz Capitulino { 1508cf11e85fSRoman Gushchin /* 1509cf11e85fSRoman Gushchin * If the page isn't allocated using the cma allocator, 1510cf11e85fSRoman Gushchin * cma_release() returns false. 1511cf11e85fSRoman Gushchin */ 1512dbda8feaSBarry Song #ifdef CONFIG_CMA 15137f325a8dSSidhartha Kumar int nid = folio_nid(folio); 15147f325a8dSSidhartha Kumar 15157f325a8dSSidhartha Kumar if (cma_release(hugetlb_cma[nid], &folio->page, 1 << order)) 1516cf11e85fSRoman Gushchin return; 1517dbda8feaSBarry Song #endif 1518cf11e85fSRoman Gushchin 15197f325a8dSSidhartha Kumar free_contig_range(folio_pfn(folio), 1 << order); 1520944d9fecSLuiz Capitulino } 1521944d9fecSLuiz Capitulino 15224eb0716eSAlexandre Ghiti #ifdef CONFIG_CONTIG_ALLOC 152319fc1a7eSSidhartha Kumar static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask, 1524d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) 1525944d9fecSLuiz Capitulino { 152619fc1a7eSSidhartha Kumar struct page *page; 152704adbc3fSMiaohe Lin unsigned long nr_pages = pages_per_huge_page(h); 1528953f064aSLi Xinhai if (nid == NUMA_NO_NODE) 1529953f064aSLi Xinhai nid = numa_mem_id(); 1530944d9fecSLuiz Capitulino 1531dbda8feaSBarry Song #ifdef CONFIG_CMA 1532dbda8feaSBarry Song { 1533cf11e85fSRoman Gushchin int node; 1534cf11e85fSRoman Gushchin 1535953f064aSLi Xinhai if (hugetlb_cma[nid]) { 1536953f064aSLi Xinhai page = cma_alloc(hugetlb_cma[nid], nr_pages, 1537953f064aSLi Xinhai huge_page_order(h), true); 1538953f064aSLi Xinhai if (page) 153919fc1a7eSSidhartha Kumar return page_folio(page); 1540953f064aSLi Xinhai } 1541953f064aSLi Xinhai 1542953f064aSLi Xinhai if (!(gfp_mask & __GFP_THISNODE)) { 1543cf11e85fSRoman Gushchin for_each_node_mask(node, *nodemask) { 1544953f064aSLi Xinhai if (node == nid || !hugetlb_cma[node]) 1545cf11e85fSRoman Gushchin continue; 1546cf11e85fSRoman Gushchin 1547cf11e85fSRoman Gushchin page = cma_alloc(hugetlb_cma[node], nr_pages, 1548cf11e85fSRoman Gushchin huge_page_order(h), true); 1549cf11e85fSRoman Gushchin if (page) 155019fc1a7eSSidhartha Kumar return page_folio(page); 1551cf11e85fSRoman Gushchin } 1552cf11e85fSRoman Gushchin } 1553953f064aSLi Xinhai } 1554dbda8feaSBarry Song #endif 1555cf11e85fSRoman Gushchin 155619fc1a7eSSidhartha Kumar page = alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); 155719fc1a7eSSidhartha Kumar return page ? page_folio(page) : NULL; 1558944d9fecSLuiz Capitulino } 1559944d9fecSLuiz Capitulino 15604eb0716eSAlexandre Ghiti #else /* !CONFIG_CONTIG_ALLOC */ 156119fc1a7eSSidhartha Kumar static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask, 15624eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 15634eb0716eSAlexandre Ghiti { 15644eb0716eSAlexandre Ghiti return NULL; 15654eb0716eSAlexandre Ghiti } 15664eb0716eSAlexandre Ghiti #endif /* CONFIG_CONTIG_ALLOC */ 1567944d9fecSLuiz Capitulino 1568e1073d1eSAneesh Kumar K.V #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 156919fc1a7eSSidhartha Kumar static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask, 15704eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 15714eb0716eSAlexandre Ghiti { 15724eb0716eSAlexandre Ghiti return NULL; 15734eb0716eSAlexandre Ghiti } 15747f325a8dSSidhartha Kumar static inline void free_gigantic_folio(struct folio *folio, 15757f325a8dSSidhartha Kumar unsigned int order) { } 1576911565b8SSidhartha Kumar static inline void destroy_compound_gigantic_folio(struct folio *folio, 1577d00181b9SKirill A. Shutemov unsigned int order) { } 1578944d9fecSLuiz Capitulino #endif 1579944d9fecSLuiz Capitulino 15806eb4e88aSMike Kravetz /* 1581cfd5082bSSidhartha Kumar * Remove hugetlb folio from lists, and update dtor so that the folio appears 158234d9e35bSMike Kravetz * as just a compound page. 158334d9e35bSMike Kravetz * 1584cfd5082bSSidhartha Kumar * A reference is held on the folio, except in the case of demote. 15856eb4e88aSMike Kravetz * 15866eb4e88aSMike Kravetz * Must be called with hugetlb lock held. 15876eb4e88aSMike Kravetz */ 1588cfd5082bSSidhartha Kumar static void __remove_hugetlb_folio(struct hstate *h, struct folio *folio, 158934d9e35bSMike Kravetz bool adjust_surplus, 159034d9e35bSMike Kravetz bool demote) 15916eb4e88aSMike Kravetz { 1592cfd5082bSSidhartha Kumar int nid = folio_nid(folio); 15936eb4e88aSMike Kravetz 1594f074732dSSidhartha Kumar VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio(folio), folio); 1595f074732dSSidhartha Kumar VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio_rsvd(folio), folio); 15966eb4e88aSMike Kravetz 15979487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 15986eb4e88aSMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 15996eb4e88aSMike Kravetz return; 16006eb4e88aSMike Kravetz 1601cfd5082bSSidhartha Kumar list_del(&folio->lru); 16026eb4e88aSMike Kravetz 1603cfd5082bSSidhartha Kumar if (folio_test_hugetlb_freed(folio)) { 16046eb4e88aSMike Kravetz h->free_huge_pages--; 16056eb4e88aSMike Kravetz h->free_huge_pages_node[nid]--; 16066eb4e88aSMike Kravetz } 16076eb4e88aSMike Kravetz if (adjust_surplus) { 16086eb4e88aSMike Kravetz h->surplus_huge_pages--; 16096eb4e88aSMike Kravetz h->surplus_huge_pages_node[nid]--; 16106eb4e88aSMike Kravetz } 16116eb4e88aSMike Kravetz 1612e32d20c0SMike Kravetz /* 1613e32d20c0SMike Kravetz * Very subtle 1614e32d20c0SMike Kravetz * 1615e32d20c0SMike Kravetz * For non-gigantic pages set the destructor to the normal compound 1616e32d20c0SMike Kravetz * page dtor. This is needed in case someone takes an additional 1617e32d20c0SMike Kravetz * temporary ref to the page, and freeing is delayed until they drop 1618e32d20c0SMike Kravetz * their reference. 1619e32d20c0SMike Kravetz * 1620e32d20c0SMike Kravetz * For gigantic pages set the destructor to the null dtor. This 1621e32d20c0SMike Kravetz * destructor will never be called. Before freeing the gigantic 1622911565b8SSidhartha Kumar * page destroy_compound_gigantic_folio will turn the folio into a 1623911565b8SSidhartha Kumar * simple group of pages. After this the destructor does not 1624e32d20c0SMike Kravetz * apply. 1625e32d20c0SMike Kravetz * 1626e32d20c0SMike Kravetz * This handles the case where more than one ref is held when and 1627d6ef19e2SSidhartha Kumar * after update_and_free_hugetlb_folio is called. 162834d9e35bSMike Kravetz * 162934d9e35bSMike Kravetz * In the case of demote we do not ref count the page as it will soon 163034d9e35bSMike Kravetz * be turned into a page of smaller size. 1631e32d20c0SMike Kravetz */ 163234d9e35bSMike Kravetz if (!demote) 1633cfd5082bSSidhartha Kumar folio_ref_unfreeze(folio, 1); 1634e32d20c0SMike Kravetz if (hstate_is_gigantic(h)) 1635cfd5082bSSidhartha Kumar folio_set_compound_dtor(folio, NULL_COMPOUND_DTOR); 1636e32d20c0SMike Kravetz else 1637cfd5082bSSidhartha Kumar folio_set_compound_dtor(folio, COMPOUND_PAGE_DTOR); 16386eb4e88aSMike Kravetz 16396eb4e88aSMike Kravetz h->nr_huge_pages--; 16406eb4e88aSMike Kravetz h->nr_huge_pages_node[nid]--; 16416eb4e88aSMike Kravetz } 16426eb4e88aSMike Kravetz 1643cfd5082bSSidhartha Kumar static void remove_hugetlb_folio(struct hstate *h, struct folio *folio, 164434d9e35bSMike Kravetz bool adjust_surplus) 164534d9e35bSMike Kravetz { 1646cfd5082bSSidhartha Kumar __remove_hugetlb_folio(h, folio, adjust_surplus, false); 164734d9e35bSMike Kravetz } 164834d9e35bSMike Kravetz 1649cfd5082bSSidhartha Kumar static void remove_hugetlb_folio_for_demote(struct hstate *h, struct folio *folio, 16508531fc6fSMike Kravetz bool adjust_surplus) 16518531fc6fSMike Kravetz { 1652cfd5082bSSidhartha Kumar __remove_hugetlb_folio(h, folio, adjust_surplus, true); 16538531fc6fSMike Kravetz } 16548531fc6fSMike Kravetz 16552f6c57d6SSidhartha Kumar static void add_hugetlb_folio(struct hstate *h, struct folio *folio, 1656ad2fa371SMuchun Song bool adjust_surplus) 1657ad2fa371SMuchun Song { 1658ad2fa371SMuchun Song int zeroed; 16592f6c57d6SSidhartha Kumar int nid = folio_nid(folio); 1660ad2fa371SMuchun Song 16612f6c57d6SSidhartha Kumar VM_BUG_ON_FOLIO(!folio_test_hugetlb_vmemmap_optimized(folio), folio); 1662ad2fa371SMuchun Song 1663ad2fa371SMuchun Song lockdep_assert_held(&hugetlb_lock); 1664ad2fa371SMuchun Song 16652f6c57d6SSidhartha Kumar INIT_LIST_HEAD(&folio->lru); 1666ad2fa371SMuchun Song h->nr_huge_pages++; 1667ad2fa371SMuchun Song h->nr_huge_pages_node[nid]++; 1668ad2fa371SMuchun Song 1669ad2fa371SMuchun Song if (adjust_surplus) { 1670ad2fa371SMuchun Song h->surplus_huge_pages++; 1671ad2fa371SMuchun Song h->surplus_huge_pages_node[nid]++; 1672ad2fa371SMuchun Song } 1673ad2fa371SMuchun Song 16742f6c57d6SSidhartha Kumar folio_set_compound_dtor(folio, HUGETLB_PAGE_DTOR); 16752f6c57d6SSidhartha Kumar folio_change_private(folio, NULL); 1676a9e1eab2SMiaohe Lin /* 16772f6c57d6SSidhartha Kumar * We have to set hugetlb_vmemmap_optimized again as above 16782f6c57d6SSidhartha Kumar * folio_change_private(folio, NULL) cleared it. 1679a9e1eab2SMiaohe Lin */ 16802f6c57d6SSidhartha Kumar folio_set_hugetlb_vmemmap_optimized(folio); 1681ad2fa371SMuchun Song 1682ad2fa371SMuchun Song /* 16832f6c57d6SSidhartha Kumar * This folio is about to be managed by the hugetlb allocator and 1684b65a4edaSMike Kravetz * should have no users. Drop our reference, and check for others 1685b65a4edaSMike Kravetz * just in case. 1686ad2fa371SMuchun Song */ 16872f6c57d6SSidhartha Kumar zeroed = folio_put_testzero(folio); 16882f6c57d6SSidhartha Kumar if (unlikely(!zeroed)) 1689b65a4edaSMike Kravetz /* 1690b65a4edaSMike Kravetz * It is VERY unlikely soneone else has taken a ref on 1691b65a4edaSMike Kravetz * the page. In this case, we simply return as the 1692b65a4edaSMike Kravetz * hugetlb destructor (free_huge_page) will be called 1693b65a4edaSMike Kravetz * when this other ref is dropped. 1694b65a4edaSMike Kravetz */ 1695b65a4edaSMike Kravetz return; 1696b65a4edaSMike Kravetz 16972f6c57d6SSidhartha Kumar arch_clear_hugepage_flags(&folio->page); 1698240d67a8SSidhartha Kumar enqueue_hugetlb_folio(h, folio); 1699ad2fa371SMuchun Song } 1700ad2fa371SMuchun Song 1701b65d4adbSMuchun Song static void __update_and_free_page(struct hstate *h, struct page *page) 17026af2acb6SAdam Litke { 17036af2acb6SAdam Litke int i; 1704911565b8SSidhartha Kumar struct folio *folio = page_folio(page); 170514455eabSCheng Li struct page *subpage; 1706a5516438SAndi Kleen 17074eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1708944d9fecSLuiz Capitulino return; 170918229df5SAndy Whitcroft 1710161df60eSNaoya Horiguchi /* 1711161df60eSNaoya Horiguchi * If we don't know which subpages are hwpoisoned, we can't free 1712161df60eSNaoya Horiguchi * the hugepage, so it's leaked intentionally. 1713161df60eSNaoya Horiguchi */ 17147f325a8dSSidhartha Kumar if (folio_test_hugetlb_raw_hwp_unreliable(folio)) 1715161df60eSNaoya Horiguchi return; 1716161df60eSNaoya Horiguchi 17176213834cSMuchun Song if (hugetlb_vmemmap_restore(h, page)) { 1718ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 1719ad2fa371SMuchun Song /* 1720ad2fa371SMuchun Song * If we cannot allocate vmemmap pages, just refuse to free the 1721ad2fa371SMuchun Song * page and put the page back on the hugetlb free list and treat 1722ad2fa371SMuchun Song * as a surplus page. 1723ad2fa371SMuchun Song */ 17247f325a8dSSidhartha Kumar add_hugetlb_folio(h, folio, true); 1725ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 1726ad2fa371SMuchun Song return; 1727ad2fa371SMuchun Song } 1728ad2fa371SMuchun Song 1729161df60eSNaoya Horiguchi /* 1730161df60eSNaoya Horiguchi * Move PageHWPoison flag from head page to the raw error pages, 1731161df60eSNaoya Horiguchi * which makes any healthy subpages reusable. 1732161df60eSNaoya Horiguchi */ 1733911565b8SSidhartha Kumar if (unlikely(folio_test_hwpoison(folio))) 1734*2ff6ceceSSidhartha Kumar folio_clear_hugetlb_hwpoison(folio); 1735161df60eSNaoya Horiguchi 173614455eabSCheng Li for (i = 0; i < pages_per_huge_page(h); i++) { 17377f325a8dSSidhartha Kumar subpage = folio_page(folio, i); 1738dbfee5aeSMike Kravetz subpage->flags &= ~(1 << PG_locked | 1 << PG_error | 173932f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1740a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1741a7407a27SLuiz Capitulino 1 << PG_writeback); 17426af2acb6SAdam Litke } 1743a01f4390SMike Kravetz 1744a01f4390SMike Kravetz /* 1745a01f4390SMike Kravetz * Non-gigantic pages demoted from CMA allocated gigantic pages 17467f325a8dSSidhartha Kumar * need to be given back to CMA in free_gigantic_folio. 1747a01f4390SMike Kravetz */ 1748a01f4390SMike Kravetz if (hstate_is_gigantic(h) || 17492f6c57d6SSidhartha Kumar hugetlb_cma_folio(folio, huge_page_order(h))) { 1750911565b8SSidhartha Kumar destroy_compound_gigantic_folio(folio, huge_page_order(h)); 17517f325a8dSSidhartha Kumar free_gigantic_folio(folio, huge_page_order(h)); 1752944d9fecSLuiz Capitulino } else { 1753a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 17546af2acb6SAdam Litke } 1755944d9fecSLuiz Capitulino } 17566af2acb6SAdam Litke 1757b65d4adbSMuchun Song /* 1758d6ef19e2SSidhartha Kumar * As update_and_free_hugetlb_folio() can be called under any context, so we cannot 1759b65d4adbSMuchun Song * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the 1760b65d4adbSMuchun Song * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate 1761b65d4adbSMuchun Song * the vmemmap pages. 1762b65d4adbSMuchun Song * 1763b65d4adbSMuchun Song * free_hpage_workfn() locklessly retrieves the linked list of pages to be 1764b65d4adbSMuchun Song * freed and frees them one-by-one. As the page->mapping pointer is going 1765b65d4adbSMuchun Song * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node 1766b65d4adbSMuchun Song * structure of a lockless linked list of huge pages to be freed. 1767b65d4adbSMuchun Song */ 1768b65d4adbSMuchun Song static LLIST_HEAD(hpage_freelist); 1769b65d4adbSMuchun Song 1770b65d4adbSMuchun Song static void free_hpage_workfn(struct work_struct *work) 1771b65d4adbSMuchun Song { 1772b65d4adbSMuchun Song struct llist_node *node; 1773b65d4adbSMuchun Song 1774b65d4adbSMuchun Song node = llist_del_all(&hpage_freelist); 1775b65d4adbSMuchun Song 1776b65d4adbSMuchun Song while (node) { 1777b65d4adbSMuchun Song struct page *page; 1778b65d4adbSMuchun Song struct hstate *h; 1779b65d4adbSMuchun Song 1780b65d4adbSMuchun Song page = container_of((struct address_space **)node, 1781b65d4adbSMuchun Song struct page, mapping); 1782b65d4adbSMuchun Song node = node->next; 1783b65d4adbSMuchun Song page->mapping = NULL; 1784b65d4adbSMuchun Song /* 1785b65d4adbSMuchun Song * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate() 1786b65d4adbSMuchun Song * is going to trigger because a previous call to 1787cfd5082bSSidhartha Kumar * remove_hugetlb_folio() will call folio_set_compound_dtor 1788cfd5082bSSidhartha Kumar * (folio, NULL_COMPOUND_DTOR), so do not use page_hstate() 1789cfd5082bSSidhartha Kumar * directly. 1790b65d4adbSMuchun Song */ 1791b65d4adbSMuchun Song h = size_to_hstate(page_size(page)); 1792b65d4adbSMuchun Song 1793b65d4adbSMuchun Song __update_and_free_page(h, page); 1794b65d4adbSMuchun Song 1795b65d4adbSMuchun Song cond_resched(); 1796b65d4adbSMuchun Song } 1797b65d4adbSMuchun Song } 1798b65d4adbSMuchun Song static DECLARE_WORK(free_hpage_work, free_hpage_workfn); 1799b65d4adbSMuchun Song 1800b65d4adbSMuchun Song static inline void flush_free_hpage_work(struct hstate *h) 1801b65d4adbSMuchun Song { 18026213834cSMuchun Song if (hugetlb_vmemmap_optimizable(h)) 1803b65d4adbSMuchun Song flush_work(&free_hpage_work); 1804b65d4adbSMuchun Song } 1805b65d4adbSMuchun Song 1806d6ef19e2SSidhartha Kumar static void update_and_free_hugetlb_folio(struct hstate *h, struct folio *folio, 1807b65d4adbSMuchun Song bool atomic) 1808b65d4adbSMuchun Song { 1809d6ef19e2SSidhartha Kumar if (!folio_test_hugetlb_vmemmap_optimized(folio) || !atomic) { 1810d6ef19e2SSidhartha Kumar __update_and_free_page(h, &folio->page); 1811b65d4adbSMuchun Song return; 1812b65d4adbSMuchun Song } 1813b65d4adbSMuchun Song 1814b65d4adbSMuchun Song /* 1815b65d4adbSMuchun Song * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages. 1816b65d4adbSMuchun Song * 1817b65d4adbSMuchun Song * Only call schedule_work() if hpage_freelist is previously 1818b65d4adbSMuchun Song * empty. Otherwise, schedule_work() had been called but the workfn 1819b65d4adbSMuchun Song * hasn't retrieved the list yet. 1820b65d4adbSMuchun Song */ 1821d6ef19e2SSidhartha Kumar if (llist_add((struct llist_node *)&folio->mapping, &hpage_freelist)) 1822b65d4adbSMuchun Song schedule_work(&free_hpage_work); 1823b65d4adbSMuchun Song } 1824b65d4adbSMuchun Song 182510c6ec49SMike Kravetz static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list) 182610c6ec49SMike Kravetz { 182710c6ec49SMike Kravetz struct page *page, *t_page; 1828d6ef19e2SSidhartha Kumar struct folio *folio; 182910c6ec49SMike Kravetz 183010c6ec49SMike Kravetz list_for_each_entry_safe(page, t_page, list, lru) { 1831d6ef19e2SSidhartha Kumar folio = page_folio(page); 1832d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, folio, false); 183310c6ec49SMike Kravetz cond_resched(); 183410c6ec49SMike Kravetz } 183510c6ec49SMike Kravetz } 183610c6ec49SMike Kravetz 1837e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1838e5ff2159SAndi Kleen { 1839e5ff2159SAndi Kleen struct hstate *h; 1840e5ff2159SAndi Kleen 1841e5ff2159SAndi Kleen for_each_hstate(h) { 1842e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1843e5ff2159SAndi Kleen return h; 1844e5ff2159SAndi Kleen } 1845e5ff2159SAndi Kleen return NULL; 1846e5ff2159SAndi Kleen } 1847e5ff2159SAndi Kleen 1848db71ef79SMike Kravetz void free_huge_page(struct page *page) 184927a85ef1SDavid Gibson { 1850a5516438SAndi Kleen /* 1851a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1852a5516438SAndi Kleen * compound page destructor. 1853a5516438SAndi Kleen */ 18540356c4b9SSidhartha Kumar struct folio *folio = page_folio(page); 18550356c4b9SSidhartha Kumar struct hstate *h = folio_hstate(folio); 18560356c4b9SSidhartha Kumar int nid = folio_nid(folio); 18570356c4b9SSidhartha Kumar struct hugepage_subpool *spool = hugetlb_folio_subpool(folio); 185807443a85SJoonsoo Kim bool restore_reserve; 1859db71ef79SMike Kravetz unsigned long flags; 186027a85ef1SDavid Gibson 18610356c4b9SSidhartha Kumar VM_BUG_ON_FOLIO(folio_ref_count(folio), folio); 18620356c4b9SSidhartha Kumar VM_BUG_ON_FOLIO(folio_mapcount(folio), folio); 18638ace22bcSYongkai Wu 18640356c4b9SSidhartha Kumar hugetlb_set_folio_subpool(folio, NULL); 18650356c4b9SSidhartha Kumar if (folio_test_anon(folio)) 18660356c4b9SSidhartha Kumar __ClearPageAnonExclusive(&folio->page); 18670356c4b9SSidhartha Kumar folio->mapping = NULL; 18680356c4b9SSidhartha Kumar restore_reserve = folio_test_hugetlb_restore_reserve(folio); 18690356c4b9SSidhartha Kumar folio_clear_hugetlb_restore_reserve(folio); 187027a85ef1SDavid Gibson 18711c5ecae3SMike Kravetz /* 1872d6995da3SMike Kravetz * If HPageRestoreReserve was set on page, page allocation consumed a 18730919e1b6SMike Kravetz * reservation. If the page was associated with a subpool, there 18740919e1b6SMike Kravetz * would have been a page reserved in the subpool before allocation 18750919e1b6SMike Kravetz * via hugepage_subpool_get_pages(). Since we are 'restoring' the 18766c26d310SMiaohe Lin * reservation, do not call hugepage_subpool_put_pages() as this will 18770919e1b6SMike Kravetz * remove the reserved page from the subpool. 18780919e1b6SMike Kravetz */ 18790919e1b6SMike Kravetz if (!restore_reserve) { 18800919e1b6SMike Kravetz /* 18810919e1b6SMike Kravetz * A return code of zero implies that the subpool will be 18820919e1b6SMike Kravetz * under its minimum size if the reservation is not restored 18830919e1b6SMike Kravetz * after page is free. Therefore, force restore_reserve 18840919e1b6SMike Kravetz * operation. 18851c5ecae3SMike Kravetz */ 18861c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 18871c5ecae3SMike Kravetz restore_reserve = true; 18880919e1b6SMike Kravetz } 18891c5ecae3SMike Kravetz 1890db71ef79SMike Kravetz spin_lock_irqsave(&hugetlb_lock, flags); 18910356c4b9SSidhartha Kumar folio_clear_hugetlb_migratable(folio); 1892d4ab0316SSidhartha Kumar hugetlb_cgroup_uncharge_folio(hstate_index(h), 1893d4ab0316SSidhartha Kumar pages_per_huge_page(h), folio); 1894d4ab0316SSidhartha Kumar hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h), 1895d4ab0316SSidhartha Kumar pages_per_huge_page(h), folio); 189607443a85SJoonsoo Kim if (restore_reserve) 189707443a85SJoonsoo Kim h->resv_huge_pages++; 189807443a85SJoonsoo Kim 18990356c4b9SSidhartha Kumar if (folio_test_hugetlb_temporary(folio)) { 1900cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, false); 1901db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1902d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, folio, true); 1903ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 19040edaecfaSAneesh Kumar K.V /* remove the page from active list */ 1905cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, true); 1906db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1907d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, folio, true); 19087893d1d5SAdam Litke } else { 19095d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1910240d67a8SSidhartha Kumar enqueue_hugetlb_folio(h, folio); 1911db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 191227a85ef1SDavid Gibson } 19131121828aSMike Kravetz } 191427a85ef1SDavid Gibson 1915d3d99fccSOscar Salvador /* 1916d3d99fccSOscar Salvador * Must be called with the hugetlb lock held 1917d3d99fccSOscar Salvador */ 1918d3d99fccSOscar Salvador static void __prep_account_new_huge_page(struct hstate *h, int nid) 1919d3d99fccSOscar Salvador { 1920d3d99fccSOscar Salvador lockdep_assert_held(&hugetlb_lock); 1921d3d99fccSOscar Salvador h->nr_huge_pages++; 1922d3d99fccSOscar Salvador h->nr_huge_pages_node[nid]++; 1923d3d99fccSOscar Salvador } 1924d3d99fccSOscar Salvador 1925de656ed3SSidhartha Kumar static void __prep_new_hugetlb_folio(struct hstate *h, struct folio *folio) 1926b7ba30c6SAndi Kleen { 1927de656ed3SSidhartha Kumar hugetlb_vmemmap_optimize(h, &folio->page); 1928de656ed3SSidhartha Kumar INIT_LIST_HEAD(&folio->lru); 19299fd33058SSidhartha Kumar folio_set_compound_dtor(folio, HUGETLB_PAGE_DTOR); 1930de656ed3SSidhartha Kumar hugetlb_set_folio_subpool(folio, NULL); 1931de656ed3SSidhartha Kumar set_hugetlb_cgroup(folio, NULL); 1932de656ed3SSidhartha Kumar set_hugetlb_cgroup_rsvd(folio, NULL); 1933d3d99fccSOscar Salvador } 1934d3d99fccSOscar Salvador 1935d1c60955SSidhartha Kumar static void prep_new_hugetlb_folio(struct hstate *h, struct folio *folio, int nid) 1936d3d99fccSOscar Salvador { 1937de656ed3SSidhartha Kumar __prep_new_hugetlb_folio(h, folio); 1938db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 1939d3d99fccSOscar Salvador __prep_account_new_huge_page(h, nid); 1940db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1941b7ba30c6SAndi Kleen } 1942b7ba30c6SAndi Kleen 1943d1c60955SSidhartha Kumar static bool __prep_compound_gigantic_folio(struct folio *folio, 1944d1c60955SSidhartha Kumar unsigned int order, bool demote) 194520a0307cSWu Fengguang { 19467118fc29SMike Kravetz int i, j; 194720a0307cSWu Fengguang int nr_pages = 1 << order; 194814455eabSCheng Li struct page *p; 194920a0307cSWu Fengguang 1950d1c60955SSidhartha Kumar __folio_clear_reserved(folio); 1951d1c60955SSidhartha Kumar __folio_set_head(folio); 1952c45bc55aSSidhartha Kumar /* we rely on prep_new_hugetlb_folio to set the destructor */ 195304a42e72SSidhartha Kumar folio_set_order(folio, order); 19542b21624fSMike Kravetz for (i = 0; i < nr_pages; i++) { 1955d1c60955SSidhartha Kumar p = folio_page(folio, i); 195614455eabSCheng Li 1957ef5a22beSAndrea Arcangeli /* 1958ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1959ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1960ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 19617c8de358SEthon Paul * too. Otherwise drivers using get_user_pages() to access tail 1962ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1963ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1964ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1965ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1966ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1967ef5a22beSAndrea Arcangeli * after get_user_pages(). 1968ef5a22beSAndrea Arcangeli */ 19697fb0728aSMike Kravetz if (i != 0) /* head page cleared above */ 1970ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 19717118fc29SMike Kravetz /* 19727118fc29SMike Kravetz * Subtle and very unlikely 19737118fc29SMike Kravetz * 19747118fc29SMike Kravetz * Gigantic 'page allocators' such as memblock or cma will 19757118fc29SMike Kravetz * return a set of pages with each page ref counted. We need 19767118fc29SMike Kravetz * to turn this set of pages into a compound page with tail 19777118fc29SMike Kravetz * page ref counts set to zero. Code such as speculative page 19787118fc29SMike Kravetz * cache adding could take a ref on a 'to be' tail page. 19797118fc29SMike Kravetz * We need to respect any increased ref count, and only set 19807118fc29SMike Kravetz * the ref count to zero if count is currently 1. If count 1981416d85edSMike Kravetz * is not 1, we return an error. An error return indicates 1982416d85edSMike Kravetz * the set of pages can not be converted to a gigantic page. 1983416d85edSMike Kravetz * The caller who allocated the pages should then discard the 1984416d85edSMike Kravetz * pages using the appropriate free interface. 198534d9e35bSMike Kravetz * 198634d9e35bSMike Kravetz * In the case of demote, the ref count will be zero. 19877118fc29SMike Kravetz */ 198834d9e35bSMike Kravetz if (!demote) { 19897118fc29SMike Kravetz if (!page_ref_freeze(p, 1)) { 1990416d85edSMike Kravetz pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n"); 19917118fc29SMike Kravetz goto out_error; 19927118fc29SMike Kravetz } 199334d9e35bSMike Kravetz } else { 199434d9e35bSMike Kravetz VM_BUG_ON_PAGE(page_count(p), p); 199534d9e35bSMike Kravetz } 19962b21624fSMike Kravetz if (i != 0) 1997d1c60955SSidhartha Kumar set_compound_head(p, &folio->page); 199820a0307cSWu Fengguang } 199946f27228SMatthew Wilcox (Oracle) atomic_set(&folio->_entire_mapcount, -1); 2000eec20426SMatthew Wilcox (Oracle) atomic_set(&folio->_nr_pages_mapped, 0); 200194688e8eSMatthew Wilcox (Oracle) atomic_set(&folio->_pincount, 0); 20027118fc29SMike Kravetz return true; 20037118fc29SMike Kravetz 20047118fc29SMike Kravetz out_error: 20052b21624fSMike Kravetz /* undo page modifications made above */ 20062b21624fSMike Kravetz for (j = 0; j < i; j++) { 2007d1c60955SSidhartha Kumar p = folio_page(folio, j); 20082b21624fSMike Kravetz if (j != 0) 20097118fc29SMike Kravetz clear_compound_head(p); 20107118fc29SMike Kravetz set_page_refcounted(p); 20117118fc29SMike Kravetz } 20127118fc29SMike Kravetz /* need to clear PG_reserved on remaining tail pages */ 201314455eabSCheng Li for (; j < nr_pages; j++) { 2014d1c60955SSidhartha Kumar p = folio_page(folio, j); 20157118fc29SMike Kravetz __ClearPageReserved(p); 201614455eabSCheng Li } 201704a42e72SSidhartha Kumar folio_set_order(folio, 0); 2018d1c60955SSidhartha Kumar __folio_clear_head(folio); 20197118fc29SMike Kravetz return false; 202020a0307cSWu Fengguang } 202120a0307cSWu Fengguang 2022d1c60955SSidhartha Kumar static bool prep_compound_gigantic_folio(struct folio *folio, 20238531fc6fSMike Kravetz unsigned int order) 20248531fc6fSMike Kravetz { 2025d1c60955SSidhartha Kumar return __prep_compound_gigantic_folio(folio, order, false); 2026d1c60955SSidhartha Kumar } 2027d1c60955SSidhartha Kumar 2028d1c60955SSidhartha Kumar static bool prep_compound_gigantic_folio_for_demote(struct folio *folio, 2029d1c60955SSidhartha Kumar unsigned int order) 2030d1c60955SSidhartha Kumar { 2031d1c60955SSidhartha Kumar return __prep_compound_gigantic_folio(folio, order, true); 20328531fc6fSMike Kravetz } 20338531fc6fSMike Kravetz 20347795912cSAndrew Morton /* 20357795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 20367795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 20377795912cSAndrew Morton * details. 20387795912cSAndrew Morton */ 203920a0307cSWu Fengguang int PageHuge(struct page *page) 204020a0307cSWu Fengguang { 20412d678c64SMatthew Wilcox (Oracle) struct folio *folio; 20422d678c64SMatthew Wilcox (Oracle) 204320a0307cSWu Fengguang if (!PageCompound(page)) 204420a0307cSWu Fengguang return 0; 20452d678c64SMatthew Wilcox (Oracle) folio = page_folio(page); 20462d678c64SMatthew Wilcox (Oracle) return folio->_folio_dtor == HUGETLB_PAGE_DTOR; 204720a0307cSWu Fengguang } 204843131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 204943131e14SNaoya Horiguchi 205027c73ae7SAndrea Arcangeli /* 205127c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 205227c73ae7SAndrea Arcangeli * normal or transparent huge pages. 205327c73ae7SAndrea Arcangeli */ 205427c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 205527c73ae7SAndrea Arcangeli { 20562d678c64SMatthew Wilcox (Oracle) struct folio *folio = (struct folio *)page_head; 20572d678c64SMatthew Wilcox (Oracle) if (!folio_test_large(folio)) 205827c73ae7SAndrea Arcangeli return 0; 205927c73ae7SAndrea Arcangeli 20602d678c64SMatthew Wilcox (Oracle) return folio->_folio_dtor == HUGETLB_PAGE_DTOR; 206127c73ae7SAndrea Arcangeli } 20624e936eccSDavid Howells EXPORT_SYMBOL_GPL(PageHeadHuge); 206327c73ae7SAndrea Arcangeli 2064c0d0381aSMike Kravetz /* 2065c0d0381aSMike Kravetz * Find and lock address space (mapping) in write mode. 2066c0d0381aSMike Kravetz * 2067336bf30eSMike Kravetz * Upon entry, the page is locked which means that page_mapping() is 2068336bf30eSMike Kravetz * stable. Due to locking order, we can only trylock_write. If we can 2069336bf30eSMike Kravetz * not get the lock, simply return NULL to caller. 2070c0d0381aSMike Kravetz */ 2071c0d0381aSMike Kravetz struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) 2072c0d0381aSMike Kravetz { 2073336bf30eSMike Kravetz struct address_space *mapping = page_mapping(hpage); 2074c0d0381aSMike Kravetz 2075c0d0381aSMike Kravetz if (!mapping) 2076c0d0381aSMike Kravetz return mapping; 2077c0d0381aSMike Kravetz 2078c0d0381aSMike Kravetz if (i_mmap_trylock_write(mapping)) 2079c0d0381aSMike Kravetz return mapping; 2080c0d0381aSMike Kravetz 2081c0d0381aSMike Kravetz return NULL; 2082c0d0381aSMike Kravetz } 2083c0d0381aSMike Kravetz 2084fe19bd3dSHugh Dickins pgoff_t hugetlb_basepage_index(struct page *page) 208513d60f4bSZhang Yi { 208613d60f4bSZhang Yi struct page *page_head = compound_head(page); 208713d60f4bSZhang Yi pgoff_t index = page_index(page_head); 208813d60f4bSZhang Yi unsigned long compound_idx; 208913d60f4bSZhang Yi 209013d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 209113d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 209213d60f4bSZhang Yi else 209313d60f4bSZhang Yi compound_idx = page - page_head; 209413d60f4bSZhang Yi 209513d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 209613d60f4bSZhang Yi } 209713d60f4bSZhang Yi 209819fc1a7eSSidhartha Kumar static struct folio *alloc_buddy_hugetlb_folio(struct hstate *h, 2099f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 2100f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 21011da177e4SLinus Torvalds { 2102af0fb9dfSMichal Hocko int order = huge_page_order(h); 21031da177e4SLinus Torvalds struct page *page; 2104f60858f9SMike Kravetz bool alloc_try_hard = true; 21052b21624fSMike Kravetz bool retry = true; 2106f96efd58SJoe Jin 2107f60858f9SMike Kravetz /* 2108f60858f9SMike Kravetz * By default we always try hard to allocate the page with 2109f60858f9SMike Kravetz * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in 2110f60858f9SMike Kravetz * a loop (to adjust global huge page counts) and previous allocation 2111f60858f9SMike Kravetz * failed, do not continue to try hard on the same node. Use the 2112f60858f9SMike Kravetz * node_alloc_noretry bitmap to manage this state information. 2113f60858f9SMike Kravetz */ 2114f60858f9SMike Kravetz if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) 2115f60858f9SMike Kravetz alloc_try_hard = false; 2116f60858f9SMike Kravetz gfp_mask |= __GFP_COMP|__GFP_NOWARN; 2117f60858f9SMike Kravetz if (alloc_try_hard) 2118f60858f9SMike Kravetz gfp_mask |= __GFP_RETRY_MAYFAIL; 2119af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 2120af0fb9dfSMichal Hocko nid = numa_mem_id(); 21212b21624fSMike Kravetz retry: 212284172f4bSMatthew Wilcox (Oracle) page = __alloc_pages(gfp_mask, order, nid, nmask); 21232b21624fSMike Kravetz 21242b21624fSMike Kravetz /* Freeze head page */ 21252b21624fSMike Kravetz if (page && !page_ref_freeze(page, 1)) { 21262b21624fSMike Kravetz __free_pages(page, order); 21272b21624fSMike Kravetz if (retry) { /* retry once */ 21282b21624fSMike Kravetz retry = false; 21292b21624fSMike Kravetz goto retry; 21302b21624fSMike Kravetz } 21312b21624fSMike Kravetz /* WOW! twice in a row. */ 21322b21624fSMike Kravetz pr_warn("HugeTLB head page unexpected inflated ref count\n"); 21332b21624fSMike Kravetz page = NULL; 21342b21624fSMike Kravetz } 21352b21624fSMike Kravetz 2136f60858f9SMike Kravetz /* 2137f60858f9SMike Kravetz * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this 2138f60858f9SMike Kravetz * indicates an overall state change. Clear bit so that we resume 2139f60858f9SMike Kravetz * normal 'try hard' allocations. 2140f60858f9SMike Kravetz */ 2141f60858f9SMike Kravetz if (node_alloc_noretry && page && !alloc_try_hard) 2142f60858f9SMike Kravetz node_clear(nid, *node_alloc_noretry); 2143f60858f9SMike Kravetz 2144f60858f9SMike Kravetz /* 2145f60858f9SMike Kravetz * If we tried hard to get a page but failed, set bit so that 2146f60858f9SMike Kravetz * subsequent attempts will not try as hard until there is an 2147f60858f9SMike Kravetz * overall state change. 2148f60858f9SMike Kravetz */ 2149f60858f9SMike Kravetz if (node_alloc_noretry && !page && alloc_try_hard) 2150f60858f9SMike Kravetz node_set(nid, *node_alloc_noretry); 2151f60858f9SMike Kravetz 215219fc1a7eSSidhartha Kumar if (!page) { 215319fc1a7eSSidhartha Kumar __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 215419fc1a7eSSidhartha Kumar return NULL; 215519fc1a7eSSidhartha Kumar } 215619fc1a7eSSidhartha Kumar 215719fc1a7eSSidhartha Kumar __count_vm_event(HTLB_BUDDY_PGALLOC); 215819fc1a7eSSidhartha Kumar return page_folio(page); 215963b4613cSNishanth Aravamudan } 216063b4613cSNishanth Aravamudan 2161af0fb9dfSMichal Hocko /* 21620c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 21630c397daeSMichal Hocko * should use this function to get new hugetlb pages 21642b21624fSMike Kravetz * 21652b21624fSMike Kravetz * Note that returned page is 'frozen': ref count of head page and all tail 21662b21624fSMike Kravetz * pages is zero. 21670c397daeSMichal Hocko */ 216819fc1a7eSSidhartha Kumar static struct folio *alloc_fresh_hugetlb_folio(struct hstate *h, 2169f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 2170f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 21710c397daeSMichal Hocko { 21727f325a8dSSidhartha Kumar struct folio *folio; 21737118fc29SMike Kravetz bool retry = false; 21740c397daeSMichal Hocko 21757118fc29SMike Kravetz retry: 21760c397daeSMichal Hocko if (hstate_is_gigantic(h)) 217719fc1a7eSSidhartha Kumar folio = alloc_gigantic_folio(h, gfp_mask, nid, nmask); 21780c397daeSMichal Hocko else 217919fc1a7eSSidhartha Kumar folio = alloc_buddy_hugetlb_folio(h, gfp_mask, 2180f60858f9SMike Kravetz nid, nmask, node_alloc_noretry); 218119fc1a7eSSidhartha Kumar if (!folio) 21820c397daeSMichal Hocko return NULL; 21837118fc29SMike Kravetz if (hstate_is_gigantic(h)) { 2184d1c60955SSidhartha Kumar if (!prep_compound_gigantic_folio(folio, huge_page_order(h))) { 21857118fc29SMike Kravetz /* 21867118fc29SMike Kravetz * Rare failure to convert pages to compound page. 21877118fc29SMike Kravetz * Free pages and try again - ONCE! 21887118fc29SMike Kravetz */ 21897f325a8dSSidhartha Kumar free_gigantic_folio(folio, huge_page_order(h)); 21907118fc29SMike Kravetz if (!retry) { 21917118fc29SMike Kravetz retry = true; 21927118fc29SMike Kravetz goto retry; 21937118fc29SMike Kravetz } 21947118fc29SMike Kravetz return NULL; 21957118fc29SMike Kravetz } 21967118fc29SMike Kravetz } 2197d1c60955SSidhartha Kumar prep_new_hugetlb_folio(h, folio, folio_nid(folio)); 21980c397daeSMichal Hocko 219919fc1a7eSSidhartha Kumar return folio; 22000c397daeSMichal Hocko } 22010c397daeSMichal Hocko 22020c397daeSMichal Hocko /* 2203af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 2204af0fb9dfSMichal Hocko * manner. 2205af0fb9dfSMichal Hocko */ 2206f60858f9SMike Kravetz static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 2207f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 2208b2261026SJoonsoo Kim { 220919fc1a7eSSidhartha Kumar struct folio *folio; 2210b2261026SJoonsoo Kim int nr_nodes, node; 2211af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2212b2261026SJoonsoo Kim 2213b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 221419fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, node, 221519fc1a7eSSidhartha Kumar nodes_allowed, node_alloc_noretry); 221619fc1a7eSSidhartha Kumar if (folio) { 221719fc1a7eSSidhartha Kumar free_huge_page(&folio->page); /* free it into the hugepage allocator */ 221819fc1a7eSSidhartha Kumar return 1; 221919fc1a7eSSidhartha Kumar } 2220b2261026SJoonsoo Kim } 2221b2261026SJoonsoo Kim 2222af0fb9dfSMichal Hocko return 0; 2223b2261026SJoonsoo Kim } 2224b2261026SJoonsoo Kim 2225e8c5c824SLee Schermerhorn /* 222610c6ec49SMike Kravetz * Remove huge page from pool from next node to free. Attempt to keep 222710c6ec49SMike Kravetz * persistent huge pages more or less balanced over allowed nodes. 222810c6ec49SMike Kravetz * This routine only 'removes' the hugetlb page. The caller must make 222910c6ec49SMike Kravetz * an additional call to free the page to low level allocators. 2230e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 2231e8c5c824SLee Schermerhorn */ 223210c6ec49SMike Kravetz static struct page *remove_pool_huge_page(struct hstate *h, 223310c6ec49SMike Kravetz nodemask_t *nodes_allowed, 22346ae11b27SLee Schermerhorn bool acct_surplus) 2235e8c5c824SLee Schermerhorn { 2236b2261026SJoonsoo Kim int nr_nodes, node; 223710c6ec49SMike Kravetz struct page *page = NULL; 2238cfd5082bSSidhartha Kumar struct folio *folio; 2239e8c5c824SLee Schermerhorn 22409487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2241b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2242685f3457SLee Schermerhorn /* 2243685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 2244685f3457SLee Schermerhorn * nodes with surplus pages. 2245685f3457SLee Schermerhorn */ 2246b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 2247b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 224810c6ec49SMike Kravetz page = list_entry(h->hugepage_freelists[node].next, 2249e8c5c824SLee Schermerhorn struct page, lru); 2250cfd5082bSSidhartha Kumar folio = page_folio(page); 2251cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, acct_surplus); 22529a76db09SLee Schermerhorn break; 2253e8c5c824SLee Schermerhorn } 2254b2261026SJoonsoo Kim } 2255e8c5c824SLee Schermerhorn 225610c6ec49SMike Kravetz return page; 2257e8c5c824SLee Schermerhorn } 2258e8c5c824SLee Schermerhorn 2259c8721bbbSNaoya Horiguchi /* 2260c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 2261faf53defSNaoya Horiguchi * nothing for in-use hugepages and non-hugepages. 2262faf53defSNaoya Horiguchi * This function returns values like below: 2263faf53defSNaoya Horiguchi * 2264ad2fa371SMuchun Song * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages 2265ad2fa371SMuchun Song * when the system is under memory pressure and the feature of 2266ad2fa371SMuchun Song * freeing unused vmemmap pages associated with each hugetlb page 2267ad2fa371SMuchun Song * is enabled. 2268faf53defSNaoya Horiguchi * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 2269faf53defSNaoya Horiguchi * (allocated or reserved.) 2270faf53defSNaoya Horiguchi * 0: successfully dissolved free hugepages or the page is not a 2271faf53defSNaoya Horiguchi * hugepage (considered as already dissolved) 2272c8721bbbSNaoya Horiguchi */ 2273c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 2274c8721bbbSNaoya Horiguchi { 22756bc9b564SNaoya Horiguchi int rc = -EBUSY; 22761a7cdab5SSidhartha Kumar struct folio *folio = page_folio(page); 2277082d5b6bSGerald Schaefer 22787ffddd49SMuchun Song retry: 2279faf53defSNaoya Horiguchi /* Not to disrupt normal path by vainly holding hugetlb_lock */ 22801a7cdab5SSidhartha Kumar if (!folio_test_hugetlb(folio)) 2281faf53defSNaoya Horiguchi return 0; 2282faf53defSNaoya Horiguchi 2283db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 22841a7cdab5SSidhartha Kumar if (!folio_test_hugetlb(folio)) { 2285faf53defSNaoya Horiguchi rc = 0; 2286faf53defSNaoya Horiguchi goto out; 2287faf53defSNaoya Horiguchi } 2288faf53defSNaoya Horiguchi 22891a7cdab5SSidhartha Kumar if (!folio_ref_count(folio)) { 22901a7cdab5SSidhartha Kumar struct hstate *h = folio_hstate(folio); 22918346d69dSXin Hao if (!available_huge_pages(h)) 2292082d5b6bSGerald Schaefer goto out; 22937ffddd49SMuchun Song 22947ffddd49SMuchun Song /* 22957ffddd49SMuchun Song * We should make sure that the page is already on the free list 22967ffddd49SMuchun Song * when it is dissolved. 22977ffddd49SMuchun Song */ 22981a7cdab5SSidhartha Kumar if (unlikely(!folio_test_hugetlb_freed(folio))) { 2299db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 23007ffddd49SMuchun Song cond_resched(); 23017ffddd49SMuchun Song 23027ffddd49SMuchun Song /* 23037ffddd49SMuchun Song * Theoretically, we should return -EBUSY when we 23047ffddd49SMuchun Song * encounter this race. In fact, we have a chance 23057ffddd49SMuchun Song * to successfully dissolve the page if we do a 23067ffddd49SMuchun Song * retry. Because the race window is quite small. 23077ffddd49SMuchun Song * If we seize this opportunity, it is an optimization 23087ffddd49SMuchun Song * for increasing the success rate of dissolving page. 23097ffddd49SMuchun Song */ 23107ffddd49SMuchun Song goto retry; 23117ffddd49SMuchun Song } 23127ffddd49SMuchun Song 2313cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, false); 2314ad2fa371SMuchun Song h->max_huge_pages--; 2315ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2316ad2fa371SMuchun Song 2317c3114a84SAnshuman Khandual /* 2318d6ef19e2SSidhartha Kumar * Normally update_and_free_hugtlb_folio will allocate required vmemmmap 2319d6ef19e2SSidhartha Kumar * before freeing the page. update_and_free_hugtlb_folio will fail to 2320ad2fa371SMuchun Song * free the page if it can not allocate required vmemmap. We 2321ad2fa371SMuchun Song * need to adjust max_huge_pages if the page is not freed. 2322ad2fa371SMuchun Song * Attempt to allocate vmemmmap here so that we can take 2323ad2fa371SMuchun Song * appropriate action on failure. 2324ad2fa371SMuchun Song */ 23251a7cdab5SSidhartha Kumar rc = hugetlb_vmemmap_restore(h, &folio->page); 2326ad2fa371SMuchun Song if (!rc) { 2327d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, folio, false); 2328ad2fa371SMuchun Song } else { 2329ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 23302f6c57d6SSidhartha Kumar add_hugetlb_folio(h, folio, false); 2331ad2fa371SMuchun Song h->max_huge_pages++; 2332ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2333ad2fa371SMuchun Song } 2334ad2fa371SMuchun Song 2335ad2fa371SMuchun Song return rc; 2336c8721bbbSNaoya Horiguchi } 2337082d5b6bSGerald Schaefer out: 2338db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2339082d5b6bSGerald Schaefer return rc; 2340c8721bbbSNaoya Horiguchi } 2341c8721bbbSNaoya Horiguchi 2342c8721bbbSNaoya Horiguchi /* 2343c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 2344c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 23452247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 23462247bb33SGerald Schaefer * part of it lies within the given range. 2347082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 2348082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 2349c8721bbbSNaoya Horiguchi */ 2350082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 2351c8721bbbSNaoya Horiguchi { 2352c8721bbbSNaoya Horiguchi unsigned long pfn; 2353eb03aa00SGerald Schaefer struct page *page; 2354082d5b6bSGerald Schaefer int rc = 0; 2355dc2628f3SMuchun Song unsigned int order; 2356dc2628f3SMuchun Song struct hstate *h; 2357c8721bbbSNaoya Horiguchi 2358d0177639SLi Zhong if (!hugepages_supported()) 2359082d5b6bSGerald Schaefer return rc; 2360d0177639SLi Zhong 2361dc2628f3SMuchun Song order = huge_page_order(&default_hstate); 2362dc2628f3SMuchun Song for_each_hstate(h) 2363dc2628f3SMuchun Song order = min(order, huge_page_order(h)); 2364dc2628f3SMuchun Song 2365dc2628f3SMuchun Song for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) { 2366eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 2367eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 2368eb03aa00SGerald Schaefer if (rc) 2369082d5b6bSGerald Schaefer break; 2370eb03aa00SGerald Schaefer } 2371082d5b6bSGerald Schaefer 2372082d5b6bSGerald Schaefer return rc; 2373c8721bbbSNaoya Horiguchi } 2374c8721bbbSNaoya Horiguchi 2375ab5ac90aSMichal Hocko /* 2376ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 2377ab5ac90aSMichal Hocko */ 23780c397daeSMichal Hocko static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, 23792b21624fSMike Kravetz int nid, nodemask_t *nmask) 23807893d1d5SAdam Litke { 238119fc1a7eSSidhartha Kumar struct folio *folio = NULL; 23827893d1d5SAdam Litke 2383bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2384aa888a74SAndi Kleen return NULL; 2385aa888a74SAndi Kleen 2386db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 23879980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 23889980d744SMichal Hocko goto out_unlock; 2389db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2390d1c3fb1fSNishanth Aravamudan 239119fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL); 239219fc1a7eSSidhartha Kumar if (!folio) 23930c397daeSMichal Hocko return NULL; 2394d1c3fb1fSNishanth Aravamudan 2395db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 23969980d744SMichal Hocko /* 23979980d744SMichal Hocko * We could have raced with the pool size change. 23989980d744SMichal Hocko * Double check that and simply deallocate the new page 23999980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 24009980d744SMichal Hocko * temporary page to workaround the nasty free_huge_page 24019980d744SMichal Hocko * codeflow 24029980d744SMichal Hocko */ 24039980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 240419fc1a7eSSidhartha Kumar folio_set_hugetlb_temporary(folio); 2405db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 240619fc1a7eSSidhartha Kumar free_huge_page(&folio->page); 24072bf753e6SKai Shen return NULL; 2408b65a4edaSMike Kravetz } 2409b65a4edaSMike Kravetz 24109980d744SMichal Hocko h->surplus_huge_pages++; 241119fc1a7eSSidhartha Kumar h->surplus_huge_pages_node[folio_nid(folio)]++; 24129980d744SMichal Hocko 24139980d744SMichal Hocko out_unlock: 2414db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 24157893d1d5SAdam Litke 241619fc1a7eSSidhartha Kumar return &folio->page; 24177893d1d5SAdam Litke } 24187893d1d5SAdam Litke 2419bbe88753SJoonsoo Kim static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, 2420ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 2421ab5ac90aSMichal Hocko { 242219fc1a7eSSidhartha Kumar struct folio *folio; 2423ab5ac90aSMichal Hocko 2424ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 2425ab5ac90aSMichal Hocko return NULL; 2426ab5ac90aSMichal Hocko 242719fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL); 242819fc1a7eSSidhartha Kumar if (!folio) 2429ab5ac90aSMichal Hocko return NULL; 2430ab5ac90aSMichal Hocko 24312b21624fSMike Kravetz /* fresh huge pages are frozen */ 243219fc1a7eSSidhartha Kumar folio_ref_unfreeze(folio, 1); 2433ab5ac90aSMichal Hocko /* 2434ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 2435ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 2436ab5ac90aSMichal Hocko */ 243719fc1a7eSSidhartha Kumar folio_set_hugetlb_temporary(folio); 2438ab5ac90aSMichal Hocko 243919fc1a7eSSidhartha Kumar return &folio->page; 2440ab5ac90aSMichal Hocko } 2441ab5ac90aSMichal Hocko 2442e4e574b7SAdam Litke /* 2443099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 2444099730d6SDave Hansen */ 2445e0ec90eeSDave Hansen static 24460c397daeSMichal Hocko struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, 2447099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 2448099730d6SDave Hansen { 2449cfcaa66fSBen Widawsky struct page *page = NULL; 2450aaf14e40SMichal Hocko struct mempolicy *mpol; 2451aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 2452aaf14e40SMichal Hocko int nid; 2453aaf14e40SMichal Hocko nodemask_t *nodemask; 2454aaf14e40SMichal Hocko 2455aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 2456cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 2457cfcaa66fSBen Widawsky gfp_t gfp = gfp_mask | __GFP_NOWARN; 2458cfcaa66fSBen Widawsky 2459cfcaa66fSBen Widawsky gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL); 24602b21624fSMike Kravetz page = alloc_surplus_huge_page(h, gfp, nid, nodemask); 2461cfcaa66fSBen Widawsky 2462cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 2463cfcaa66fSBen Widawsky nodemask = NULL; 2464cfcaa66fSBen Widawsky } 2465cfcaa66fSBen Widawsky 2466cfcaa66fSBen Widawsky if (!page) 24672b21624fSMike Kravetz page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); 2468aaf14e40SMichal Hocko mpol_cond_put(mpol); 2469aaf14e40SMichal Hocko return page; 2470099730d6SDave Hansen } 2471099730d6SDave Hansen 2472ab5ac90aSMichal Hocko /* page migration callback function */ 24733e59fcb0SMichal Hocko struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, 2474d92bbc27SJoonsoo Kim nodemask_t *nmask, gfp_t gfp_mask) 24754db9b2efSMichal Hocko { 2476db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 24778346d69dSXin Hao if (available_huge_pages(h)) { 24783e59fcb0SMichal Hocko struct page *page; 24793e59fcb0SMichal Hocko 24803e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); 24813e59fcb0SMichal Hocko if (page) { 2482db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 24833e59fcb0SMichal Hocko return page; 24844db9b2efSMichal Hocko } 24854db9b2efSMichal Hocko } 2486db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 24874db9b2efSMichal Hocko 24880c397daeSMichal Hocko return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); 24894db9b2efSMichal Hocko } 24904db9b2efSMichal Hocko 2491ebd63723SMichal Hocko /* mempolicy aware migration callback */ 2492389c8178SMichal Hocko struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, 2493389c8178SMichal Hocko unsigned long address) 2494ebd63723SMichal Hocko { 2495ebd63723SMichal Hocko struct mempolicy *mpol; 2496ebd63723SMichal Hocko nodemask_t *nodemask; 2497ebd63723SMichal Hocko struct page *page; 2498ebd63723SMichal Hocko gfp_t gfp_mask; 2499ebd63723SMichal Hocko int node; 2500ebd63723SMichal Hocko 2501ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 2502ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 2503d92bbc27SJoonsoo Kim page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); 2504ebd63723SMichal Hocko mpol_cond_put(mpol); 2505ebd63723SMichal Hocko 2506ebd63723SMichal Hocko return page; 2507ebd63723SMichal Hocko } 2508ebd63723SMichal Hocko 2509bf50bab2SNaoya Horiguchi /* 251025985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 2511e4e574b7SAdam Litke * of size 'delta'. 2512e4e574b7SAdam Litke */ 25130a4f3d1bSLiu Xiang static int gather_surplus_pages(struct hstate *h, long delta) 25141b2a1e7bSJules Irenge __must_hold(&hugetlb_lock) 2515e4e574b7SAdam Litke { 251634665341SMiaohe Lin LIST_HEAD(surplus_list); 2517e4e574b7SAdam Litke struct page *page, *tmp; 25180a4f3d1bSLiu Xiang int ret; 25190a4f3d1bSLiu Xiang long i; 25200a4f3d1bSLiu Xiang long needed, allocated; 252128073b02SHillf Danton bool alloc_ok = true; 2522e4e574b7SAdam Litke 25239487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2524a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 2525ac09b3a1SAdam Litke if (needed <= 0) { 2526a5516438SAndi Kleen h->resv_huge_pages += delta; 2527e4e574b7SAdam Litke return 0; 2528ac09b3a1SAdam Litke } 2529e4e574b7SAdam Litke 2530e4e574b7SAdam Litke allocated = 0; 2531e4e574b7SAdam Litke 2532e4e574b7SAdam Litke ret = -ENOMEM; 2533e4e574b7SAdam Litke retry: 2534db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2535e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 25360c397daeSMichal Hocko page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), 25372b21624fSMike Kravetz NUMA_NO_NODE, NULL); 253828073b02SHillf Danton if (!page) { 253928073b02SHillf Danton alloc_ok = false; 254028073b02SHillf Danton break; 254128073b02SHillf Danton } 2542e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 254369ed779aSDavid Rientjes cond_resched(); 2544e4e574b7SAdam Litke } 254528073b02SHillf Danton allocated += i; 2546e4e574b7SAdam Litke 2547e4e574b7SAdam Litke /* 2548e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 2549e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 2550e4e574b7SAdam Litke */ 2551db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2552a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 2553a5516438SAndi Kleen (h->free_huge_pages + allocated); 255428073b02SHillf Danton if (needed > 0) { 255528073b02SHillf Danton if (alloc_ok) 2556e4e574b7SAdam Litke goto retry; 255728073b02SHillf Danton /* 255828073b02SHillf Danton * We were not able to allocate enough pages to 255928073b02SHillf Danton * satisfy the entire reservation so we free what 256028073b02SHillf Danton * we've allocated so far. 256128073b02SHillf Danton */ 256228073b02SHillf Danton goto free; 256328073b02SHillf Danton } 2564e4e574b7SAdam Litke /* 2565e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 256625985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 2567e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 2568ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 2569ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 2570ac09b3a1SAdam Litke * before they are reserved. 2571e4e574b7SAdam Litke */ 2572e4e574b7SAdam Litke needed += allocated; 2573a5516438SAndi Kleen h->resv_huge_pages += delta; 2574e4e574b7SAdam Litke ret = 0; 2575a9869b83SNaoya Horiguchi 257619fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 257719fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 257819fc3f0aSAdam Litke if ((--needed) < 0) 257919fc3f0aSAdam Litke break; 2580b65a4edaSMike Kravetz /* Add the page to the hugetlb allocator */ 2581240d67a8SSidhartha Kumar enqueue_hugetlb_folio(h, page_folio(page)); 258219fc3f0aSAdam Litke } 258328073b02SHillf Danton free: 2584db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 258519fc3f0aSAdam Litke 2586b65a4edaSMike Kravetz /* 2587b65a4edaSMike Kravetz * Free unnecessary surplus pages to the buddy allocator. 2588b65a4edaSMike Kravetz * Pages have no ref count, call free_huge_page directly. 2589b65a4edaSMike Kravetz */ 2590c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 2591b65a4edaSMike Kravetz free_huge_page(page); 2592db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2593e4e574b7SAdam Litke 2594e4e574b7SAdam Litke return ret; 2595e4e574b7SAdam Litke } 2596e4e574b7SAdam Litke 2597e4e574b7SAdam Litke /* 2598e5bbc8a6SMike Kravetz * This routine has two main purposes: 2599e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 2600e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 2601e5bbc8a6SMike Kravetz * to the associated reservation map. 2602e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 2603e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 2604e4e574b7SAdam Litke */ 2605a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 2606a5516438SAndi Kleen unsigned long unused_resv_pages) 2607e4e574b7SAdam Litke { 2608e4e574b7SAdam Litke unsigned long nr_pages; 260910c6ec49SMike Kravetz struct page *page; 261010c6ec49SMike Kravetz LIST_HEAD(page_list); 261110c6ec49SMike Kravetz 26129487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 261310c6ec49SMike Kravetz /* Uncommit the reservation */ 261410c6ec49SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 2615e4e574b7SAdam Litke 2616c0531714SNaoya Horiguchi if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 2617e5bbc8a6SMike Kravetz goto out; 2618aa888a74SAndi Kleen 2619e5bbc8a6SMike Kravetz /* 2620e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 2621e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 2622e5bbc8a6SMike Kravetz */ 2623a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 2624e4e574b7SAdam Litke 2625685f3457SLee Schermerhorn /* 2626685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 26279b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 26289b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 26299b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 263010c6ec49SMike Kravetz * remove_pool_huge_page() will balance the freed pages across the 26319b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 2632685f3457SLee Schermerhorn */ 2633685f3457SLee Schermerhorn while (nr_pages--) { 263410c6ec49SMike Kravetz page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1); 263510c6ec49SMike Kravetz if (!page) 2636e5bbc8a6SMike Kravetz goto out; 263710c6ec49SMike Kravetz 263810c6ec49SMike Kravetz list_add(&page->lru, &page_list); 2639e4e574b7SAdam Litke } 2640e5bbc8a6SMike Kravetz 2641e5bbc8a6SMike Kravetz out: 2642db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 264310c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 2644db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2645e4e574b7SAdam Litke } 2646e4e574b7SAdam Litke 26475e911373SMike Kravetz 2648c37f9fb1SAndy Whitcroft /* 2649feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 26505e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 2651cf3ad20bSMike Kravetz * 2652cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 2653cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 2654cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 2655cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 2656cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 2657feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 2658feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 2659feba16e2SMike Kravetz * is called in such cases. 2660cf3ad20bSMike Kravetz * 2661cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 2662cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 2663cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 2664cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 2665cf3ad20bSMike Kravetz * take appropriate action. 266696b96a96SMike Kravetz * 266796b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 266896b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 266996b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 267096b96a96SMike Kravetz * reservation exists. 2671846be085SMike Kravetz * 2672846be085SMike Kravetz * vma_del_reservation is used in error paths where an entry in the reserve 2673846be085SMike Kravetz * map was created during huge page allocation and must be removed. It is to 2674846be085SMike Kravetz * be called after calling vma_needs_reservation to determine if a reservation 2675846be085SMike Kravetz * exists. 2676c37f9fb1SAndy Whitcroft */ 26775e911373SMike Kravetz enum vma_resv_mode { 26785e911373SMike Kravetz VMA_NEEDS_RESV, 26795e911373SMike Kravetz VMA_COMMIT_RESV, 2680feba16e2SMike Kravetz VMA_END_RESV, 268196b96a96SMike Kravetz VMA_ADD_RESV, 2682846be085SMike Kravetz VMA_DEL_RESV, 26835e911373SMike Kravetz }; 2684cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 2685cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 26865e911373SMike Kravetz enum vma_resv_mode mode) 2687c37f9fb1SAndy Whitcroft { 26884e35f483SJoonsoo Kim struct resv_map *resv; 26894e35f483SJoonsoo Kim pgoff_t idx; 2690cf3ad20bSMike Kravetz long ret; 26910db9d74eSMina Almasry long dummy_out_regions_needed; 2692c37f9fb1SAndy Whitcroft 26934e35f483SJoonsoo Kim resv = vma_resv_map(vma); 26944e35f483SJoonsoo Kim if (!resv) 2695c37f9fb1SAndy Whitcroft return 1; 2696c37f9fb1SAndy Whitcroft 26974e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 26985e911373SMike Kravetz switch (mode) { 26995e911373SMike Kravetz case VMA_NEEDS_RESV: 27000db9d74eSMina Almasry ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); 27010db9d74eSMina Almasry /* We assume that vma_reservation_* routines always operate on 27020db9d74eSMina Almasry * 1 page, and that adding to resv map a 1 page entry can only 27030db9d74eSMina Almasry * ever require 1 region. 27040db9d74eSMina Almasry */ 27050db9d74eSMina Almasry VM_BUG_ON(dummy_out_regions_needed != 1); 27065e911373SMike Kravetz break; 27075e911373SMike Kravetz case VMA_COMMIT_RESV: 2708075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 27090db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 27100db9d74eSMina Almasry VM_BUG_ON(ret < 0); 27115e911373SMike Kravetz break; 2712feba16e2SMike Kravetz case VMA_END_RESV: 27130db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 27145e911373SMike Kravetz ret = 0; 27155e911373SMike Kravetz break; 271696b96a96SMike Kravetz case VMA_ADD_RESV: 27170db9d74eSMina Almasry if (vma->vm_flags & VM_MAYSHARE) { 2718075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 27190db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 27200db9d74eSMina Almasry VM_BUG_ON(ret < 0); 27210db9d74eSMina Almasry } else { 27220db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 272396b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 272496b96a96SMike Kravetz } 272596b96a96SMike Kravetz break; 2726846be085SMike Kravetz case VMA_DEL_RESV: 2727846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 2728846be085SMike Kravetz region_abort(resv, idx, idx + 1, 1); 2729846be085SMike Kravetz ret = region_del(resv, idx, idx + 1); 2730846be085SMike Kravetz } else { 2731846be085SMike Kravetz ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 2732846be085SMike Kravetz /* region_add calls of range 1 should never fail. */ 2733846be085SMike Kravetz VM_BUG_ON(ret < 0); 2734846be085SMike Kravetz } 2735846be085SMike Kravetz break; 27365e911373SMike Kravetz default: 27375e911373SMike Kravetz BUG(); 27385e911373SMike Kravetz } 273984afd99bSAndy Whitcroft 2740846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV) 2741cf3ad20bSMike Kravetz return ret; 274267961f9dSMike Kravetz /* 2743bf3d12b9SMiaohe Lin * We know private mapping must have HPAGE_RESV_OWNER set. 2744bf3d12b9SMiaohe Lin * 274567961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 274667961f9dSMike Kravetz * However, a file associated with mapping could have been 274767961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 274867961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 274967961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 275067961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 275167961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 275267961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 275367961f9dSMike Kravetz * reservation has already been consumed. As a result, the 275467961f9dSMike Kravetz * return value of this routine is the opposite of the 275567961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 275667961f9dSMike Kravetz */ 2757bf3d12b9SMiaohe Lin if (ret > 0) 275867961f9dSMike Kravetz return 0; 2759bf3d12b9SMiaohe Lin if (ret == 0) 276067961f9dSMike Kravetz return 1; 2761bf3d12b9SMiaohe Lin return ret; 276284afd99bSAndy Whitcroft } 2763cf3ad20bSMike Kravetz 2764cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 2765a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 2766c37f9fb1SAndy Whitcroft { 27675e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 2768cf3ad20bSMike Kravetz } 2769c37f9fb1SAndy Whitcroft 2770cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 2771cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 2772cf3ad20bSMike Kravetz { 27735e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 27745e911373SMike Kravetz } 27755e911373SMike Kravetz 2776feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 27775e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 27785e911373SMike Kravetz { 2779feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 2780c37f9fb1SAndy Whitcroft } 2781c37f9fb1SAndy Whitcroft 278296b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 278396b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 278496b96a96SMike Kravetz { 278596b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 278696b96a96SMike Kravetz } 278796b96a96SMike Kravetz 2788846be085SMike Kravetz static long vma_del_reservation(struct hstate *h, 2789846be085SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 279096b96a96SMike Kravetz { 2791846be085SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV); 2792846be085SMike Kravetz } 2793846be085SMike Kravetz 2794846be085SMike Kravetz /* 2795846be085SMike Kravetz * This routine is called to restore reservation information on error paths. 2796846be085SMike Kravetz * It should ONLY be called for pages allocated via alloc_huge_page(), and 2797846be085SMike Kravetz * the hugetlb mutex should remain held when calling this routine. 2798846be085SMike Kravetz * 2799846be085SMike Kravetz * It handles two specific cases: 2800846be085SMike Kravetz * 1) A reservation was in place and the page consumed the reservation. 2801846be085SMike Kravetz * HPageRestoreReserve is set in the page. 2802846be085SMike Kravetz * 2) No reservation was in place for the page, so HPageRestoreReserve is 2803846be085SMike Kravetz * not set. However, alloc_huge_page always updates the reserve map. 2804846be085SMike Kravetz * 2805846be085SMike Kravetz * In case 1, free_huge_page later in the error path will increment the 2806846be085SMike Kravetz * global reserve count. But, free_huge_page does not have enough context 2807846be085SMike Kravetz * to adjust the reservation map. This case deals primarily with private 2808846be085SMike Kravetz * mappings. Adjust the reserve map here to be consistent with global 2809846be085SMike Kravetz * reserve count adjustments to be made by free_huge_page. Make sure the 2810846be085SMike Kravetz * reserve map indicates there is a reservation present. 2811846be085SMike Kravetz * 2812846be085SMike Kravetz * In case 2, simply undo reserve map modifications done by alloc_huge_page. 2813846be085SMike Kravetz */ 2814846be085SMike Kravetz void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma, 2815846be085SMike Kravetz unsigned long address, struct page *page) 2816846be085SMike Kravetz { 281796b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 281896b96a96SMike Kravetz 2819846be085SMike Kravetz if (HPageRestoreReserve(page)) { 2820846be085SMike Kravetz if (unlikely(rc < 0)) 282196b96a96SMike Kravetz /* 282296b96a96SMike Kravetz * Rare out of memory condition in reserve map 2823d6995da3SMike Kravetz * manipulation. Clear HPageRestoreReserve so that 282496b96a96SMike Kravetz * global reserve count will not be incremented 282596b96a96SMike Kravetz * by free_huge_page. This will make it appear 282696b96a96SMike Kravetz * as though the reservation for this page was 282796b96a96SMike Kravetz * consumed. This may prevent the task from 282896b96a96SMike Kravetz * faulting in the page at a later time. This 282996b96a96SMike Kravetz * is better than inconsistent global huge page 283096b96a96SMike Kravetz * accounting of reserve counts. 283196b96a96SMike Kravetz */ 2832d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 2833846be085SMike Kravetz else if (rc) 2834846be085SMike Kravetz (void)vma_add_reservation(h, vma, address); 2835846be085SMike Kravetz else 2836846be085SMike Kravetz vma_end_reservation(h, vma, address); 2837846be085SMike Kravetz } else { 2838846be085SMike Kravetz if (!rc) { 283996b96a96SMike Kravetz /* 2840846be085SMike Kravetz * This indicates there is an entry in the reserve map 2841c7b1850dSMike Kravetz * not added by alloc_huge_page. We know it was added 2842846be085SMike Kravetz * before the alloc_huge_page call, otherwise 2843846be085SMike Kravetz * HPageRestoreReserve would be set on the page. 2844846be085SMike Kravetz * Remove the entry so that a subsequent allocation 2845846be085SMike Kravetz * does not consume a reservation. 284696b96a96SMike Kravetz */ 2847846be085SMike Kravetz rc = vma_del_reservation(h, vma, address); 2848846be085SMike Kravetz if (rc < 0) 2849846be085SMike Kravetz /* 2850846be085SMike Kravetz * VERY rare out of memory condition. Since 2851846be085SMike Kravetz * we can not delete the entry, set 2852846be085SMike Kravetz * HPageRestoreReserve so that the reserve 2853846be085SMike Kravetz * count will be incremented when the page 2854846be085SMike Kravetz * is freed. This reserve will be consumed 2855846be085SMike Kravetz * on a subsequent allocation. 2856846be085SMike Kravetz */ 2857846be085SMike Kravetz SetHPageRestoreReserve(page); 2858846be085SMike Kravetz } else if (rc < 0) { 2859846be085SMike Kravetz /* 2860846be085SMike Kravetz * Rare out of memory condition from 2861846be085SMike Kravetz * vma_needs_reservation call. Memory allocation is 2862846be085SMike Kravetz * only attempted if a new entry is needed. Therefore, 2863846be085SMike Kravetz * this implies there is not an entry in the 2864846be085SMike Kravetz * reserve map. 2865846be085SMike Kravetz * 2866846be085SMike Kravetz * For shared mappings, no entry in the map indicates 2867846be085SMike Kravetz * no reservation. We are done. 2868846be085SMike Kravetz */ 2869846be085SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 2870846be085SMike Kravetz /* 2871846be085SMike Kravetz * For private mappings, no entry indicates 2872846be085SMike Kravetz * a reservation is present. Since we can 2873846be085SMike Kravetz * not add an entry, set SetHPageRestoreReserve 2874846be085SMike Kravetz * on the page so reserve count will be 2875846be085SMike Kravetz * incremented when freed. This reserve will 2876846be085SMike Kravetz * be consumed on a subsequent allocation. 2877846be085SMike Kravetz */ 2878846be085SMike Kravetz SetHPageRestoreReserve(page); 287996b96a96SMike Kravetz } else 2880846be085SMike Kravetz /* 2881846be085SMike Kravetz * No reservation present, do nothing 2882846be085SMike Kravetz */ 288396b96a96SMike Kravetz vma_end_reservation(h, vma, address); 288496b96a96SMike Kravetz } 288596b96a96SMike Kravetz } 288696b96a96SMike Kravetz 2887369fa227SOscar Salvador /* 288819fc1a7eSSidhartha Kumar * alloc_and_dissolve_hugetlb_folio - Allocate a new folio and dissolve 288919fc1a7eSSidhartha Kumar * the old one 2890369fa227SOscar Salvador * @h: struct hstate old page belongs to 289119fc1a7eSSidhartha Kumar * @old_folio: Old folio to dissolve 2892ae37c7ffSOscar Salvador * @list: List to isolate the page in case we need to 2893369fa227SOscar Salvador * Returns 0 on success, otherwise negated error. 2894369fa227SOscar Salvador */ 289519fc1a7eSSidhartha Kumar static int alloc_and_dissolve_hugetlb_folio(struct hstate *h, 289619fc1a7eSSidhartha Kumar struct folio *old_folio, struct list_head *list) 2897369fa227SOscar Salvador { 2898369fa227SOscar Salvador gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2899de656ed3SSidhartha Kumar int nid = folio_nid(old_folio); 2900de656ed3SSidhartha Kumar struct folio *new_folio; 2901369fa227SOscar Salvador int ret = 0; 2902369fa227SOscar Salvador 2903369fa227SOscar Salvador /* 290419fc1a7eSSidhartha Kumar * Before dissolving the folio, we need to allocate a new one for the 290519fc1a7eSSidhartha Kumar * pool to remain stable. Here, we allocate the folio and 'prep' it 2906f41f2ed4SMuchun Song * by doing everything but actually updating counters and adding to 2907f41f2ed4SMuchun Song * the pool. This simplifies and let us do most of the processing 2908f41f2ed4SMuchun Song * under the lock. 2909369fa227SOscar Salvador */ 291019fc1a7eSSidhartha Kumar new_folio = alloc_buddy_hugetlb_folio(h, gfp_mask, nid, NULL, NULL); 291119fc1a7eSSidhartha Kumar if (!new_folio) 2912369fa227SOscar Salvador return -ENOMEM; 2913de656ed3SSidhartha Kumar __prep_new_hugetlb_folio(h, new_folio); 2914369fa227SOscar Salvador 2915369fa227SOscar Salvador retry: 2916369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2917de656ed3SSidhartha Kumar if (!folio_test_hugetlb(old_folio)) { 2918369fa227SOscar Salvador /* 291919fc1a7eSSidhartha Kumar * Freed from under us. Drop new_folio too. 2920369fa227SOscar Salvador */ 2921369fa227SOscar Salvador goto free_new; 2922de656ed3SSidhartha Kumar } else if (folio_ref_count(old_folio)) { 2923369fa227SOscar Salvador /* 292419fc1a7eSSidhartha Kumar * Someone has grabbed the folio, try to isolate it here. 2925ae37c7ffSOscar Salvador * Fail with -EBUSY if not possible. 2926369fa227SOscar Salvador */ 2927ae37c7ffSOscar Salvador spin_unlock_irq(&hugetlb_lock); 292819fc1a7eSSidhartha Kumar ret = isolate_hugetlb(&old_folio->page, list); 2929ae37c7ffSOscar Salvador spin_lock_irq(&hugetlb_lock); 2930369fa227SOscar Salvador goto free_new; 2931de656ed3SSidhartha Kumar } else if (!folio_test_hugetlb_freed(old_folio)) { 2932369fa227SOscar Salvador /* 293319fc1a7eSSidhartha Kumar * Folio's refcount is 0 but it has not been enqueued in the 2934369fa227SOscar Salvador * freelist yet. Race window is small, so we can succeed here if 2935369fa227SOscar Salvador * we retry. 2936369fa227SOscar Salvador */ 2937369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2938369fa227SOscar Salvador cond_resched(); 2939369fa227SOscar Salvador goto retry; 2940369fa227SOscar Salvador } else { 2941369fa227SOscar Salvador /* 294219fc1a7eSSidhartha Kumar * Ok, old_folio is still a genuine free hugepage. Remove it from 2943369fa227SOscar Salvador * the freelist and decrease the counters. These will be 2944369fa227SOscar Salvador * incremented again when calling __prep_account_new_huge_page() 2945240d67a8SSidhartha Kumar * and enqueue_hugetlb_folio() for new_folio. The counters will 2946240d67a8SSidhartha Kumar * remain stable since this happens under the lock. 2947369fa227SOscar Salvador */ 2948cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, old_folio, false); 2949369fa227SOscar Salvador 2950369fa227SOscar Salvador /* 295119fc1a7eSSidhartha Kumar * Ref count on new_folio is already zero as it was dropped 2952b65a4edaSMike Kravetz * earlier. It can be directly added to the pool free list. 2953369fa227SOscar Salvador */ 2954369fa227SOscar Salvador __prep_account_new_huge_page(h, nid); 2955240d67a8SSidhartha Kumar enqueue_hugetlb_folio(h, new_folio); 2956369fa227SOscar Salvador 2957369fa227SOscar Salvador /* 295819fc1a7eSSidhartha Kumar * Folio has been replaced, we can safely free the old one. 2959369fa227SOscar Salvador */ 2960369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2961d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, old_folio, false); 2962369fa227SOscar Salvador } 2963369fa227SOscar Salvador 2964369fa227SOscar Salvador return ret; 2965369fa227SOscar Salvador 2966369fa227SOscar Salvador free_new: 2967369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 296819fc1a7eSSidhartha Kumar /* Folio has a zero ref count, but needs a ref to be freed */ 2969de656ed3SSidhartha Kumar folio_ref_unfreeze(new_folio, 1); 2970d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, new_folio, false); 2971369fa227SOscar Salvador 2972369fa227SOscar Salvador return ret; 2973369fa227SOscar Salvador } 2974369fa227SOscar Salvador 2975ae37c7ffSOscar Salvador int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list) 2976369fa227SOscar Salvador { 2977369fa227SOscar Salvador struct hstate *h; 2978d5e33bd8SSidhartha Kumar struct folio *folio = page_folio(page); 2979ae37c7ffSOscar Salvador int ret = -EBUSY; 2980369fa227SOscar Salvador 2981369fa227SOscar Salvador /* 2982369fa227SOscar Salvador * The page might have been dissolved from under our feet, so make sure 2983369fa227SOscar Salvador * to carefully check the state under the lock. 2984369fa227SOscar Salvador * Return success when racing as if we dissolved the page ourselves. 2985369fa227SOscar Salvador */ 2986369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2987d5e33bd8SSidhartha Kumar if (folio_test_hugetlb(folio)) { 2988d5e33bd8SSidhartha Kumar h = folio_hstate(folio); 2989369fa227SOscar Salvador } else { 2990369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2991369fa227SOscar Salvador return 0; 2992369fa227SOscar Salvador } 2993369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2994369fa227SOscar Salvador 2995369fa227SOscar Salvador /* 2996369fa227SOscar Salvador * Fence off gigantic pages as there is a cyclic dependency between 2997369fa227SOscar Salvador * alloc_contig_range and them. Return -ENOMEM as this has the effect 2998369fa227SOscar Salvador * of bailing out right away without further retrying. 2999369fa227SOscar Salvador */ 3000369fa227SOscar Salvador if (hstate_is_gigantic(h)) 3001369fa227SOscar Salvador return -ENOMEM; 3002369fa227SOscar Salvador 3003d5e33bd8SSidhartha Kumar if (folio_ref_count(folio) && !isolate_hugetlb(&folio->page, list)) 3004ae37c7ffSOscar Salvador ret = 0; 3005d5e33bd8SSidhartha Kumar else if (!folio_ref_count(folio)) 300619fc1a7eSSidhartha Kumar ret = alloc_and_dissolve_hugetlb_folio(h, folio, list); 3007ae37c7ffSOscar Salvador 3008ae37c7ffSOscar Salvador return ret; 3009369fa227SOscar Salvador } 3010369fa227SOscar Salvador 301170c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 301204f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 3013348ea204SAdam Litke { 301490481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 3015a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3016348ea204SAdam Litke struct page *page; 3017d4ab0316SSidhartha Kumar struct folio *folio; 3018d85f69b0SMike Kravetz long map_chg, map_commit; 3019d85f69b0SMike Kravetz long gbl_chg; 30206d76dcf4SAneesh Kumar K.V int ret, idx; 30216d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 302208cf9fafSMina Almasry bool deferred_reserve; 30232fc39cecSAdam Litke 30246d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 3025a1e78772SMel Gorman /* 3026d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 3027d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 3028d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 3029a1e78772SMel Gorman */ 3030d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 3031d85f69b0SMike Kravetz if (map_chg < 0) 303276dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 3033d85f69b0SMike Kravetz 3034d85f69b0SMike Kravetz /* 3035d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 3036d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 3037d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 3038d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 3039d85f69b0SMike Kravetz * checked against any subpool limit. 3040d85f69b0SMike Kravetz */ 3041d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 3042d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 3043d85f69b0SMike Kravetz if (gbl_chg < 0) { 3044feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 304576dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 30465e911373SMike Kravetz } 304790d8b7e6SAdam Litke 3048d85f69b0SMike Kravetz /* 3049d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 3050d85f69b0SMike Kravetz * map, there could be reservations associated with the 3051d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 3052d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 3053d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 3054d85f69b0SMike Kravetz * the subpool reservations. 3055d85f69b0SMike Kravetz */ 3056d85f69b0SMike Kravetz if (avoid_reserve) 3057d85f69b0SMike Kravetz gbl_chg = 1; 3058d85f69b0SMike Kravetz } 3059d85f69b0SMike Kravetz 306008cf9fafSMina Almasry /* If this allocation is not consuming a reservation, charge it now. 306108cf9fafSMina Almasry */ 30626501fe5fSMiaohe Lin deferred_reserve = map_chg || avoid_reserve; 306308cf9fafSMina Almasry if (deferred_reserve) { 306408cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 306508cf9fafSMina Almasry idx, pages_per_huge_page(h), &h_cg); 30668f34af6fSJianyu Zhan if (ret) 30678f34af6fSJianyu Zhan goto out_subpool_put; 306808cf9fafSMina Almasry } 306908cf9fafSMina Almasry 307008cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 307108cf9fafSMina Almasry if (ret) 307208cf9fafSMina Almasry goto out_uncharge_cgroup_reservation; 30738f34af6fSJianyu Zhan 3074db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3075d85f69b0SMike Kravetz /* 3076d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 3077d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 3078d85f69b0SMike Kravetz * a reservation exists for the allocation. 3079d85f69b0SMike Kravetz */ 3080d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 308181a6fcaeSJoonsoo Kim if (!page) { 3082db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 30830c397daeSMichal Hocko page = alloc_buddy_huge_page_with_mpol(h, vma, addr); 30848f34af6fSJianyu Zhan if (!page) 30858f34af6fSJianyu Zhan goto out_uncharge_cgroup; 308612df140fSRik van Riel spin_lock_irq(&hugetlb_lock); 3087a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 3088d6995da3SMike Kravetz SetHPageRestoreReserve(page); 3089a88c7695SNaoya Horiguchi h->resv_huge_pages--; 3090a88c7695SNaoya Horiguchi } 309115a8d68eSWei Yang list_add(&page->lru, &h->hugepage_activelist); 30922b21624fSMike Kravetz set_page_refcounted(page); 309381a6fcaeSJoonsoo Kim /* Fall through */ 3094a1e78772SMel Gorman } 3095d4ab0316SSidhartha Kumar folio = page_folio(page); 309681a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 309708cf9fafSMina Almasry /* If allocation is not consuming a reservation, also store the 309808cf9fafSMina Almasry * hugetlb_cgroup pointer on the page. 309908cf9fafSMina Almasry */ 310008cf9fafSMina Almasry if (deferred_reserve) { 310108cf9fafSMina Almasry hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), 310208cf9fafSMina Almasry h_cg, page); 310308cf9fafSMina Almasry } 310408cf9fafSMina Almasry 3105db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3106a1e78772SMel Gorman 3107d6995da3SMike Kravetz hugetlb_set_page_subpool(page, spool); 3108a1e78772SMel Gorman 3109d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 3110d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 311133039678SMike Kravetz /* 311233039678SMike Kravetz * The page was added to the reservation map between 311333039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 311433039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 311533039678SMike Kravetz * Adjust for the subpool count incremented above AND 311633039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 311733039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 311833039678SMike Kravetz * no longer applies. 311933039678SMike Kravetz */ 312033039678SMike Kravetz long rsv_adjust; 312133039678SMike Kravetz 312233039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 312333039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 312479aa925bSMike Kravetz if (deferred_reserve) 3125d4ab0316SSidhartha Kumar hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h), 3126d4ab0316SSidhartha Kumar pages_per_huge_page(h), folio); 312733039678SMike Kravetz } 31287893d1d5SAdam Litke return page; 31298f34af6fSJianyu Zhan 31308f34af6fSJianyu Zhan out_uncharge_cgroup: 31318f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 313208cf9fafSMina Almasry out_uncharge_cgroup_reservation: 313308cf9fafSMina Almasry if (deferred_reserve) 313408cf9fafSMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), 313508cf9fafSMina Almasry h_cg); 31368f34af6fSJianyu Zhan out_subpool_put: 3137d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 31388f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 3139feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 31408f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 3141b45b5bd6SDavid Gibson } 3142b45b5bd6SDavid Gibson 3143b5389086SZhenguo Yao int alloc_bootmem_huge_page(struct hstate *h, int nid) 3144e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 3145b5389086SZhenguo Yao int __alloc_bootmem_huge_page(struct hstate *h, int nid) 3146aa888a74SAndi Kleen { 3147b5389086SZhenguo Yao struct huge_bootmem_page *m = NULL; /* initialize for clang */ 3148b2261026SJoonsoo Kim int nr_nodes, node; 3149aa888a74SAndi Kleen 3150b5389086SZhenguo Yao /* do node specific alloc */ 3151b5389086SZhenguo Yao if (nid != NUMA_NO_NODE) { 3152b5389086SZhenguo Yao m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h), 3153b5389086SZhenguo Yao 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid); 3154b5389086SZhenguo Yao if (!m) 3155b5389086SZhenguo Yao return 0; 3156b5389086SZhenguo Yao goto found; 3157b5389086SZhenguo Yao } 3158b5389086SZhenguo Yao /* allocate from next node when distributing huge pages */ 3159b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 3160b5389086SZhenguo Yao m = memblock_alloc_try_nid_raw( 31618b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 316297ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 3163aa888a74SAndi Kleen /* 3164aa888a74SAndi Kleen * Use the beginning of the huge page to store the 3165aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 3166aa888a74SAndi Kleen * puts them into the mem_map). 3167aa888a74SAndi Kleen */ 3168b5389086SZhenguo Yao if (!m) 3169b5389086SZhenguo Yao return 0; 3170aa888a74SAndi Kleen goto found; 3171aa888a74SAndi Kleen } 3172aa888a74SAndi Kleen 3173aa888a74SAndi Kleen found: 3174aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 3175330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 3176aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 3177aa888a74SAndi Kleen m->hstate = h; 3178aa888a74SAndi Kleen return 1; 3179aa888a74SAndi Kleen } 3180aa888a74SAndi Kleen 318148b8d744SMike Kravetz /* 318248b8d744SMike Kravetz * Put bootmem huge pages into the standard lists after mem_map is up. 318348b8d744SMike Kravetz * Note: This only applies to gigantic (order > MAX_ORDER) pages. 318448b8d744SMike Kravetz */ 3185aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 3186aa888a74SAndi Kleen { 3187aa888a74SAndi Kleen struct huge_bootmem_page *m; 3188aa888a74SAndi Kleen 3189aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 319040d18ebfSMike Kravetz struct page *page = virt_to_page(m); 31917f325a8dSSidhartha Kumar struct folio *folio = page_folio(page); 3192aa888a74SAndi Kleen struct hstate *h = m->hstate; 3193ee8f248dSBecky Bruce 319448b8d744SMike Kravetz VM_BUG_ON(!hstate_is_gigantic(h)); 3195d1c60955SSidhartha Kumar WARN_ON(folio_ref_count(folio) != 1); 3196d1c60955SSidhartha Kumar if (prep_compound_gigantic_folio(folio, huge_page_order(h))) { 3197d1c60955SSidhartha Kumar WARN_ON(folio_test_reserved(folio)); 3198d1c60955SSidhartha Kumar prep_new_hugetlb_folio(h, folio, folio_nid(folio)); 31992b21624fSMike Kravetz free_huge_page(page); /* add to the hugepage allocator */ 32007118fc29SMike Kravetz } else { 3201416d85edSMike Kravetz /* VERY unlikely inflated ref count on a tail page */ 32027f325a8dSSidhartha Kumar free_gigantic_folio(folio, huge_page_order(h)); 32037118fc29SMike Kravetz } 3204af0fb9dfSMichal Hocko 3205b0320c7bSRafael Aquini /* 320648b8d744SMike Kravetz * We need to restore the 'stolen' pages to totalram_pages 320748b8d744SMike Kravetz * in order to fix confusing memory reports from free(1) and 320848b8d744SMike Kravetz * other side-effects, like CommitLimit going negative. 3209b0320c7bSRafael Aquini */ 3210c78a7f36SMiaohe Lin adjust_managed_page_count(page, pages_per_huge_page(h)); 3211520495feSCannon Matthews cond_resched(); 3212aa888a74SAndi Kleen } 3213aa888a74SAndi Kleen } 3214b5389086SZhenguo Yao static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid) 3215b5389086SZhenguo Yao { 3216b5389086SZhenguo Yao unsigned long i; 3217b5389086SZhenguo Yao char buf[32]; 3218b5389086SZhenguo Yao 3219b5389086SZhenguo Yao for (i = 0; i < h->max_huge_pages_node[nid]; ++i) { 3220b5389086SZhenguo Yao if (hstate_is_gigantic(h)) { 3221b5389086SZhenguo Yao if (!alloc_bootmem_huge_page(h, nid)) 3222b5389086SZhenguo Yao break; 3223b5389086SZhenguo Yao } else { 322419fc1a7eSSidhartha Kumar struct folio *folio; 3225b5389086SZhenguo Yao gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 3226b5389086SZhenguo Yao 322719fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, 3228b5389086SZhenguo Yao &node_states[N_MEMORY], NULL); 322919fc1a7eSSidhartha Kumar if (!folio) 3230b5389086SZhenguo Yao break; 323119fc1a7eSSidhartha Kumar free_huge_page(&folio->page); /* free it into the hugepage allocator */ 3232b5389086SZhenguo Yao } 3233b5389086SZhenguo Yao cond_resched(); 3234b5389086SZhenguo Yao } 3235b5389086SZhenguo Yao if (i == h->max_huge_pages_node[nid]) 3236b5389086SZhenguo Yao return; 3237b5389086SZhenguo Yao 3238b5389086SZhenguo Yao string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3239b5389086SZhenguo Yao pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n", 3240b5389086SZhenguo Yao h->max_huge_pages_node[nid], buf, nid, i); 3241b5389086SZhenguo Yao h->max_huge_pages -= (h->max_huge_pages_node[nid] - i); 3242b5389086SZhenguo Yao h->max_huge_pages_node[nid] = i; 3243b5389086SZhenguo Yao } 3244aa888a74SAndi Kleen 32458faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 32461da177e4SLinus Torvalds { 32471da177e4SLinus Torvalds unsigned long i; 3248f60858f9SMike Kravetz nodemask_t *node_alloc_noretry; 3249b5389086SZhenguo Yao bool node_specific_alloc = false; 3250f60858f9SMike Kravetz 3251b5389086SZhenguo Yao /* skip gigantic hugepages allocation if hugetlb_cma enabled */ 3252b5389086SZhenguo Yao if (hstate_is_gigantic(h) && hugetlb_cma_size) { 3253b5389086SZhenguo Yao pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); 3254b5389086SZhenguo Yao return; 3255b5389086SZhenguo Yao } 3256b5389086SZhenguo Yao 3257b5389086SZhenguo Yao /* do node specific alloc */ 32580a7a0f6fSPeng Liu for_each_online_node(i) { 3259b5389086SZhenguo Yao if (h->max_huge_pages_node[i] > 0) { 3260b5389086SZhenguo Yao hugetlb_hstate_alloc_pages_onenode(h, i); 3261b5389086SZhenguo Yao node_specific_alloc = true; 3262b5389086SZhenguo Yao } 3263b5389086SZhenguo Yao } 3264b5389086SZhenguo Yao 3265b5389086SZhenguo Yao if (node_specific_alloc) 3266b5389086SZhenguo Yao return; 3267b5389086SZhenguo Yao 3268b5389086SZhenguo Yao /* below will do all node balanced alloc */ 3269f60858f9SMike Kravetz if (!hstate_is_gigantic(h)) { 3270f60858f9SMike Kravetz /* 3271f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3272f60858f9SMike Kravetz * Ignore errors as lower level routines can deal with 3273f60858f9SMike Kravetz * node_alloc_noretry == NULL. If this kmalloc fails at boot 3274f60858f9SMike Kravetz * time, we are likely in bigger trouble. 3275f60858f9SMike Kravetz */ 3276f60858f9SMike Kravetz node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), 3277f60858f9SMike Kravetz GFP_KERNEL); 3278f60858f9SMike Kravetz } else { 3279f60858f9SMike Kravetz /* allocations done at boot time */ 3280f60858f9SMike Kravetz node_alloc_noretry = NULL; 3281f60858f9SMike Kravetz } 3282f60858f9SMike Kravetz 3283f60858f9SMike Kravetz /* bit mask controlling how hard we retry per-node allocations */ 3284f60858f9SMike Kravetz if (node_alloc_noretry) 3285f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 32861da177e4SLinus Torvalds 3287e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 3288bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 3289b5389086SZhenguo Yao if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE)) 3290aa888a74SAndi Kleen break; 32910c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 3292f60858f9SMike Kravetz &node_states[N_MEMORY], 3293f60858f9SMike Kravetz node_alloc_noretry)) 32941da177e4SLinus Torvalds break; 329569ed779aSDavid Rientjes cond_resched(); 32961da177e4SLinus Torvalds } 3297d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 3298d715cf80SLiam R. Howlett char buf[32]; 3299d715cf80SLiam R. Howlett 3300c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3301d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 3302d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 33038faa8b07SAndi Kleen h->max_huge_pages = i; 3304e5ff2159SAndi Kleen } 3305f60858f9SMike Kravetz kfree(node_alloc_noretry); 3306d715cf80SLiam R. Howlett } 3307e5ff2159SAndi Kleen 3308e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 3309e5ff2159SAndi Kleen { 331079dfc695SMike Kravetz struct hstate *h, *h2; 3311e5ff2159SAndi Kleen 3312e5ff2159SAndi Kleen for_each_hstate(h) { 33138faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 3314bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 33158faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 331679dfc695SMike Kravetz 331779dfc695SMike Kravetz /* 331879dfc695SMike Kravetz * Set demote order for each hstate. Note that 331979dfc695SMike Kravetz * h->demote_order is initially 0. 332079dfc695SMike Kravetz * - We can not demote gigantic pages if runtime freeing 332179dfc695SMike Kravetz * is not supported, so skip this. 3322a01f4390SMike Kravetz * - If CMA allocation is possible, we can not demote 3323a01f4390SMike Kravetz * HUGETLB_PAGE_ORDER or smaller size pages. 332479dfc695SMike Kravetz */ 332579dfc695SMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 332679dfc695SMike Kravetz continue; 3327a01f4390SMike Kravetz if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER) 3328a01f4390SMike Kravetz continue; 332979dfc695SMike Kravetz for_each_hstate(h2) { 333079dfc695SMike Kravetz if (h2 == h) 333179dfc695SMike Kravetz continue; 333279dfc695SMike Kravetz if (h2->order < h->order && 333379dfc695SMike Kravetz h2->order > h->demote_order) 333479dfc695SMike Kravetz h->demote_order = h2->order; 333579dfc695SMike Kravetz } 3336e5ff2159SAndi Kleen } 3337e5ff2159SAndi Kleen } 3338e5ff2159SAndi Kleen 3339e5ff2159SAndi Kleen static void __init report_hugepages(void) 3340e5ff2159SAndi Kleen { 3341e5ff2159SAndi Kleen struct hstate *h; 3342e5ff2159SAndi Kleen 3343e5ff2159SAndi Kleen for_each_hstate(h) { 33444abd32dbSAndi Kleen char buf[32]; 3345c6247f72SMatthew Wilcox 3346c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 33476213834cSMuchun Song pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n", 3348c6247f72SMatthew Wilcox buf, h->free_huge_pages); 33496213834cSMuchun Song pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n", 33506213834cSMuchun Song hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf); 3351e5ff2159SAndi Kleen } 3352e5ff2159SAndi Kleen } 3353e5ff2159SAndi Kleen 33541da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 33556ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 33566ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 33571da177e4SLinus Torvalds { 33584415cc8dSChristoph Lameter int i; 33591121828aSMike Kravetz LIST_HEAD(page_list); 33604415cc8dSChristoph Lameter 33619487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 3362bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3363aa888a74SAndi Kleen return; 3364aa888a74SAndi Kleen 33651121828aSMike Kravetz /* 33661121828aSMike Kravetz * Collect pages to be freed on a list, and free after dropping lock 33671121828aSMike Kravetz */ 33686ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 336910c6ec49SMike Kravetz struct page *page, *next; 3370a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 3371a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 3372a5516438SAndi Kleen if (count >= h->nr_huge_pages) 33731121828aSMike Kravetz goto out; 33741da177e4SLinus Torvalds if (PageHighMem(page)) 33751da177e4SLinus Torvalds continue; 3376cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, page_folio(page), false); 33771121828aSMike Kravetz list_add(&page->lru, &page_list); 33781121828aSMike Kravetz } 33791121828aSMike Kravetz } 33801121828aSMike Kravetz 33811121828aSMike Kravetz out: 3382db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 338310c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3384db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 33851da177e4SLinus Torvalds } 33861da177e4SLinus Torvalds #else 33876ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 33886ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 33891da177e4SLinus Torvalds { 33901da177e4SLinus Torvalds } 33911da177e4SLinus Torvalds #endif 33921da177e4SLinus Torvalds 339320a0307cSWu Fengguang /* 339420a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 339520a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 339620a0307cSWu Fengguang * Returns 1 if an adjustment was made. 339720a0307cSWu Fengguang */ 33986ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 33996ae11b27SLee Schermerhorn int delta) 340020a0307cSWu Fengguang { 3401b2261026SJoonsoo Kim int nr_nodes, node; 340220a0307cSWu Fengguang 34039487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 340420a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 340520a0307cSWu Fengguang 3406e8c5c824SLee Schermerhorn if (delta < 0) { 3407b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 3408b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 3409b2261026SJoonsoo Kim goto found; 3410b2261026SJoonsoo Kim } 3411b2261026SJoonsoo Kim } else { 3412b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 3413b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 3414b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 3415b2261026SJoonsoo Kim goto found; 3416e8c5c824SLee Schermerhorn } 34179a76db09SLee Schermerhorn } 3418b2261026SJoonsoo Kim return 0; 341920a0307cSWu Fengguang 3420b2261026SJoonsoo Kim found: 342120a0307cSWu Fengguang h->surplus_huge_pages += delta; 3422b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 3423b2261026SJoonsoo Kim return 1; 342420a0307cSWu Fengguang } 342520a0307cSWu Fengguang 3426a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 3427fd875dcaSMike Kravetz static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 34286ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 34291da177e4SLinus Torvalds { 34307893d1d5SAdam Litke unsigned long min_count, ret; 343110c6ec49SMike Kravetz struct page *page; 343210c6ec49SMike Kravetz LIST_HEAD(page_list); 3433f60858f9SMike Kravetz NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); 3434f60858f9SMike Kravetz 3435f60858f9SMike Kravetz /* 3436f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3437f60858f9SMike Kravetz * If we can not allocate the bit mask, do not attempt to allocate 3438f60858f9SMike Kravetz * the requested huge pages. 3439f60858f9SMike Kravetz */ 3440f60858f9SMike Kravetz if (node_alloc_noretry) 3441f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 3442f60858f9SMike Kravetz else 3443f60858f9SMike Kravetz return -ENOMEM; 34441da177e4SLinus Torvalds 344529383967SMike Kravetz /* 344629383967SMike Kravetz * resize_lock mutex prevents concurrent adjustments to number of 344729383967SMike Kravetz * pages in hstate via the proc/sysfs interfaces. 344829383967SMike Kravetz */ 344929383967SMike Kravetz mutex_lock(&h->resize_lock); 3450b65d4adbSMuchun Song flush_free_hpage_work(h); 3451db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 34524eb0716eSAlexandre Ghiti 34534eb0716eSAlexandre Ghiti /* 3454fd875dcaSMike Kravetz * Check for a node specific request. 3455fd875dcaSMike Kravetz * Changing node specific huge page count may require a corresponding 3456fd875dcaSMike Kravetz * change to the global count. In any case, the passed node mask 3457fd875dcaSMike Kravetz * (nodes_allowed) will restrict alloc/free to the specified node. 3458fd875dcaSMike Kravetz */ 3459fd875dcaSMike Kravetz if (nid != NUMA_NO_NODE) { 3460fd875dcaSMike Kravetz unsigned long old_count = count; 3461fd875dcaSMike Kravetz 3462fd875dcaSMike Kravetz count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 3463fd875dcaSMike Kravetz /* 3464fd875dcaSMike Kravetz * User may have specified a large count value which caused the 3465fd875dcaSMike Kravetz * above calculation to overflow. In this case, they wanted 3466fd875dcaSMike Kravetz * to allocate as many huge pages as possible. Set count to 3467fd875dcaSMike Kravetz * largest possible value to align with their intention. 3468fd875dcaSMike Kravetz */ 3469fd875dcaSMike Kravetz if (count < old_count) 3470fd875dcaSMike Kravetz count = ULONG_MAX; 3471fd875dcaSMike Kravetz } 3472fd875dcaSMike Kravetz 3473fd875dcaSMike Kravetz /* 34744eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 34754eb0716eSAlexandre Ghiti * page range allocation. 34764eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 34774eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 34784eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 34794eb0716eSAlexandre Ghiti */ 34804eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 34814eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 3482db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 348329383967SMike Kravetz mutex_unlock(&h->resize_lock); 3484f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 34854eb0716eSAlexandre Ghiti return -EINVAL; 34864eb0716eSAlexandre Ghiti } 34874eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 34884eb0716eSAlexandre Ghiti } 3489aa888a74SAndi Kleen 34907893d1d5SAdam Litke /* 34917893d1d5SAdam Litke * Increase the pool size 34927893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 34937893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 3494d1c3fb1fSNishanth Aravamudan * 34950c397daeSMichal Hocko * We might race with alloc_surplus_huge_page() here and be unable 3496d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 3497d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 3498d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 3499d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 35007893d1d5SAdam Litke */ 3501a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 35026ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 35037893d1d5SAdam Litke break; 35047893d1d5SAdam Litke } 35057893d1d5SAdam Litke 3506a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 35077893d1d5SAdam Litke /* 35087893d1d5SAdam Litke * If this allocation races such that we no longer need the 35097893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 35107893d1d5SAdam Litke * and reducing the surplus. 35117893d1d5SAdam Litke */ 3512db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3513649920c6SJia He 3514649920c6SJia He /* yield cpu to avoid soft lockup */ 3515649920c6SJia He cond_resched(); 3516649920c6SJia He 3517f60858f9SMike Kravetz ret = alloc_pool_huge_page(h, nodes_allowed, 3518f60858f9SMike Kravetz node_alloc_noretry); 3519db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 35207893d1d5SAdam Litke if (!ret) 35217893d1d5SAdam Litke goto out; 35227893d1d5SAdam Litke 3523536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 3524536240f2SMel Gorman if (signal_pending(current)) 3525536240f2SMel Gorman goto out; 35267893d1d5SAdam Litke } 35277893d1d5SAdam Litke 35287893d1d5SAdam Litke /* 35297893d1d5SAdam Litke * Decrease the pool size 35307893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 35317893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 35327893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 35337893d1d5SAdam Litke * to the desired size as pages become free. 3534d1c3fb1fSNishanth Aravamudan * 3535d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 3536d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 3537d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 35380c397daeSMichal Hocko * alloc_surplus_huge_page() is checking the global counter, 3539d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 3540d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 3541d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 35427893d1d5SAdam Litke */ 3543a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 35446b0c880dSAdam Litke min_count = max(count, min_count); 35456ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 354610c6ec49SMike Kravetz 354710c6ec49SMike Kravetz /* 354810c6ec49SMike Kravetz * Collect pages to be removed on list without dropping lock 354910c6ec49SMike Kravetz */ 3550a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 355110c6ec49SMike Kravetz page = remove_pool_huge_page(h, nodes_allowed, 0); 355210c6ec49SMike Kravetz if (!page) 35531da177e4SLinus Torvalds break; 355410c6ec49SMike Kravetz 355510c6ec49SMike Kravetz list_add(&page->lru, &page_list); 35561da177e4SLinus Torvalds } 355710c6ec49SMike Kravetz /* free the pages after dropping lock */ 3558db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 355910c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3560b65d4adbSMuchun Song flush_free_hpage_work(h); 3561db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 356210c6ec49SMike Kravetz 3563a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 35646ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 35657893d1d5SAdam Litke break; 35667893d1d5SAdam Litke } 35677893d1d5SAdam Litke out: 35684eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 3569db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 357029383967SMike Kravetz mutex_unlock(&h->resize_lock); 35714eb0716eSAlexandre Ghiti 3572f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 3573f60858f9SMike Kravetz 35744eb0716eSAlexandre Ghiti return 0; 35751da177e4SLinus Torvalds } 35761da177e4SLinus Torvalds 35778531fc6fSMike Kravetz static int demote_free_huge_page(struct hstate *h, struct page *page) 35788531fc6fSMike Kravetz { 35798531fc6fSMike Kravetz int i, nid = page_to_nid(page); 35808531fc6fSMike Kravetz struct hstate *target_hstate; 3581911565b8SSidhartha Kumar struct folio *folio = page_folio(page); 358231731452SDoug Berger struct page *subpage; 35838531fc6fSMike Kravetz int rc = 0; 35848531fc6fSMike Kravetz 35858531fc6fSMike Kravetz target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order); 35868531fc6fSMike Kravetz 3587cfd5082bSSidhartha Kumar remove_hugetlb_folio_for_demote(h, folio, false); 35888531fc6fSMike Kravetz spin_unlock_irq(&hugetlb_lock); 35898531fc6fSMike Kravetz 35906213834cSMuchun Song rc = hugetlb_vmemmap_restore(h, page); 35918531fc6fSMike Kravetz if (rc) { 35928531fc6fSMike Kravetz /* Allocation of vmemmmap failed, we can not demote page */ 35938531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 35948531fc6fSMike Kravetz set_page_refcounted(page); 35952f6c57d6SSidhartha Kumar add_hugetlb_folio(h, page_folio(page), false); 35968531fc6fSMike Kravetz return rc; 35978531fc6fSMike Kravetz } 35988531fc6fSMike Kravetz 35998531fc6fSMike Kravetz /* 3600911565b8SSidhartha Kumar * Use destroy_compound_hugetlb_folio_for_demote for all huge page 36018531fc6fSMike Kravetz * sizes as it will not ref count pages. 36028531fc6fSMike Kravetz */ 3603911565b8SSidhartha Kumar destroy_compound_hugetlb_folio_for_demote(folio, huge_page_order(h)); 36048531fc6fSMike Kravetz 36058531fc6fSMike Kravetz /* 36068531fc6fSMike Kravetz * Taking target hstate mutex synchronizes with set_max_huge_pages. 36078531fc6fSMike Kravetz * Without the mutex, pages added to target hstate could be marked 36088531fc6fSMike Kravetz * as surplus. 36098531fc6fSMike Kravetz * 36108531fc6fSMike Kravetz * Note that we already hold h->resize_lock. To prevent deadlock, 36118531fc6fSMike Kravetz * use the convention of always taking larger size hstate mutex first. 36128531fc6fSMike Kravetz */ 36138531fc6fSMike Kravetz mutex_lock(&target_hstate->resize_lock); 36148531fc6fSMike Kravetz for (i = 0; i < pages_per_huge_page(h); 36158531fc6fSMike Kravetz i += pages_per_huge_page(target_hstate)) { 361631731452SDoug Berger subpage = nth_page(page, i); 3617d1c60955SSidhartha Kumar folio = page_folio(subpage); 36188531fc6fSMike Kravetz if (hstate_is_gigantic(target_hstate)) 3619d1c60955SSidhartha Kumar prep_compound_gigantic_folio_for_demote(folio, 36208531fc6fSMike Kravetz target_hstate->order); 36218531fc6fSMike Kravetz else 362231731452SDoug Berger prep_compound_page(subpage, target_hstate->order); 362331731452SDoug Berger set_page_private(subpage, 0); 3624d1c60955SSidhartha Kumar prep_new_hugetlb_folio(target_hstate, folio, nid); 36252b21624fSMike Kravetz free_huge_page(subpage); 36268531fc6fSMike Kravetz } 36278531fc6fSMike Kravetz mutex_unlock(&target_hstate->resize_lock); 36288531fc6fSMike Kravetz 36298531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 36308531fc6fSMike Kravetz 36318531fc6fSMike Kravetz /* 36328531fc6fSMike Kravetz * Not absolutely necessary, but for consistency update max_huge_pages 36338531fc6fSMike Kravetz * based on pool changes for the demoted page. 36348531fc6fSMike Kravetz */ 36358531fc6fSMike Kravetz h->max_huge_pages--; 3636a43a83c7SMiaohe Lin target_hstate->max_huge_pages += 3637a43a83c7SMiaohe Lin pages_per_huge_page(h) / pages_per_huge_page(target_hstate); 36388531fc6fSMike Kravetz 36398531fc6fSMike Kravetz return rc; 36408531fc6fSMike Kravetz } 36418531fc6fSMike Kravetz 364279dfc695SMike Kravetz static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 364379dfc695SMike Kravetz __must_hold(&hugetlb_lock) 364479dfc695SMike Kravetz { 36458531fc6fSMike Kravetz int nr_nodes, node; 36468531fc6fSMike Kravetz struct page *page; 364779dfc695SMike Kravetz 364879dfc695SMike Kravetz lockdep_assert_held(&hugetlb_lock); 364979dfc695SMike Kravetz 365079dfc695SMike Kravetz /* We should never get here if no demote order */ 365179dfc695SMike Kravetz if (!h->demote_order) { 365279dfc695SMike Kravetz pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n"); 365379dfc695SMike Kravetz return -EINVAL; /* internal error */ 365479dfc695SMike Kravetz } 365579dfc695SMike Kravetz 36568531fc6fSMike Kravetz for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 36575a317412SMike Kravetz list_for_each_entry(page, &h->hugepage_freelists[node], lru) { 36585a317412SMike Kravetz if (PageHWPoison(page)) 36595a317412SMike Kravetz continue; 36605a317412SMike Kravetz 36615a317412SMike Kravetz return demote_free_huge_page(h, page); 36628531fc6fSMike Kravetz } 36638531fc6fSMike Kravetz } 36648531fc6fSMike Kravetz 36655a317412SMike Kravetz /* 36665a317412SMike Kravetz * Only way to get here is if all pages on free lists are poisoned. 36675a317412SMike Kravetz * Return -EBUSY so that caller will not retry. 36685a317412SMike Kravetz */ 36695a317412SMike Kravetz return -EBUSY; 367079dfc695SMike Kravetz } 367179dfc695SMike Kravetz 3672a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 3673a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 3674a3437870SNishanth Aravamudan 367579dfc695SMike Kravetz #define HSTATE_ATTR_WO(_name) \ 367679dfc695SMike Kravetz static struct kobj_attribute _name##_attr = __ATTR_WO(_name) 367779dfc695SMike Kravetz 3678a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 367998bc26acSMiaohe Lin static struct kobj_attribute _name##_attr = __ATTR_RW(_name) 3680a3437870SNishanth Aravamudan 3681a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 3682a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 3683a3437870SNishanth Aravamudan 36849a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 36859a305230SLee Schermerhorn 36869a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 3687a3437870SNishanth Aravamudan { 3688a3437870SNishanth Aravamudan int i; 36899a305230SLee Schermerhorn 3690a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 36919a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 36929a305230SLee Schermerhorn if (nidp) 36939a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 3694a3437870SNishanth Aravamudan return &hstates[i]; 36959a305230SLee Schermerhorn } 36969a305230SLee Schermerhorn 36979a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 3698a3437870SNishanth Aravamudan } 3699a3437870SNishanth Aravamudan 370006808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 3701a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3702a3437870SNishanth Aravamudan { 37039a305230SLee Schermerhorn struct hstate *h; 37049a305230SLee Schermerhorn unsigned long nr_huge_pages; 37059a305230SLee Schermerhorn int nid; 37069a305230SLee Schermerhorn 37079a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 37089a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 37099a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 37109a305230SLee Schermerhorn else 37119a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 37129a305230SLee Schermerhorn 3713ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", nr_huge_pages); 3714a3437870SNishanth Aravamudan } 3715adbe8726SEric B Munson 3716238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 3717238d3c13SDavid Rientjes struct hstate *h, int nid, 3718238d3c13SDavid Rientjes unsigned long count, size_t len) 3719a3437870SNishanth Aravamudan { 3720a3437870SNishanth Aravamudan int err; 37212d0adf7eSOscar Salvador nodemask_t nodes_allowed, *n_mask; 3722a3437870SNishanth Aravamudan 37232d0adf7eSOscar Salvador if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 37242d0adf7eSOscar Salvador return -EINVAL; 3725adbe8726SEric B Munson 37269a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 37279a305230SLee Schermerhorn /* 37289a305230SLee Schermerhorn * global hstate attribute 37299a305230SLee Schermerhorn */ 37309a305230SLee Schermerhorn if (!(obey_mempolicy && 37312d0adf7eSOscar Salvador init_nodemask_of_mempolicy(&nodes_allowed))) 37322d0adf7eSOscar Salvador n_mask = &node_states[N_MEMORY]; 37332d0adf7eSOscar Salvador else 37342d0adf7eSOscar Salvador n_mask = &nodes_allowed; 37352d0adf7eSOscar Salvador } else { 37369a305230SLee Schermerhorn /* 3737fd875dcaSMike Kravetz * Node specific request. count adjustment happens in 3738fd875dcaSMike Kravetz * set_max_huge_pages() after acquiring hugetlb_lock. 37399a305230SLee Schermerhorn */ 37402d0adf7eSOscar Salvador init_nodemask_of_node(&nodes_allowed, nid); 37412d0adf7eSOscar Salvador n_mask = &nodes_allowed; 3742fd875dcaSMike Kravetz } 37439a305230SLee Schermerhorn 37442d0adf7eSOscar Salvador err = set_max_huge_pages(h, count, nid, n_mask); 374506808b08SLee Schermerhorn 37464eb0716eSAlexandre Ghiti return err ? err : len; 374706808b08SLee Schermerhorn } 374806808b08SLee Schermerhorn 3749238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 3750238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 3751238d3c13SDavid Rientjes size_t len) 3752238d3c13SDavid Rientjes { 3753238d3c13SDavid Rientjes struct hstate *h; 3754238d3c13SDavid Rientjes unsigned long count; 3755238d3c13SDavid Rientjes int nid; 3756238d3c13SDavid Rientjes int err; 3757238d3c13SDavid Rientjes 3758238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 3759238d3c13SDavid Rientjes if (err) 3760238d3c13SDavid Rientjes return err; 3761238d3c13SDavid Rientjes 3762238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 3763238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 3764238d3c13SDavid Rientjes } 3765238d3c13SDavid Rientjes 376606808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 376706808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 376806808b08SLee Schermerhorn { 376906808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 377006808b08SLee Schermerhorn } 377106808b08SLee Schermerhorn 377206808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 377306808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 377406808b08SLee Schermerhorn { 3775238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 3776a3437870SNishanth Aravamudan } 3777a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 3778a3437870SNishanth Aravamudan 377906808b08SLee Schermerhorn #ifdef CONFIG_NUMA 378006808b08SLee Schermerhorn 378106808b08SLee Schermerhorn /* 378206808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 378306808b08SLee Schermerhorn * huge page alloc/free. 378406808b08SLee Schermerhorn */ 378506808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 3786ae7a927dSJoe Perches struct kobj_attribute *attr, 3787ae7a927dSJoe Perches char *buf) 378806808b08SLee Schermerhorn { 378906808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 379006808b08SLee Schermerhorn } 379106808b08SLee Schermerhorn 379206808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 379306808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 379406808b08SLee Schermerhorn { 3795238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 379606808b08SLee Schermerhorn } 379706808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 379806808b08SLee Schermerhorn #endif 379906808b08SLee Schermerhorn 380006808b08SLee Schermerhorn 3801a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 3802a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3803a3437870SNishanth Aravamudan { 38049a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3805ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); 3806a3437870SNishanth Aravamudan } 3807adbe8726SEric B Munson 3808a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 3809a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 3810a3437870SNishanth Aravamudan { 3811a3437870SNishanth Aravamudan int err; 3812a3437870SNishanth Aravamudan unsigned long input; 38139a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3814a3437870SNishanth Aravamudan 3815bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3816adbe8726SEric B Munson return -EINVAL; 3817adbe8726SEric B Munson 38183dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 3819a3437870SNishanth Aravamudan if (err) 382073ae31e5SEric B Munson return err; 3821a3437870SNishanth Aravamudan 3822db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3823a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 3824db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3825a3437870SNishanth Aravamudan 3826a3437870SNishanth Aravamudan return count; 3827a3437870SNishanth Aravamudan } 3828a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 3829a3437870SNishanth Aravamudan 3830a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 3831a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3832a3437870SNishanth Aravamudan { 38339a305230SLee Schermerhorn struct hstate *h; 38349a305230SLee Schermerhorn unsigned long free_huge_pages; 38359a305230SLee Schermerhorn int nid; 38369a305230SLee Schermerhorn 38379a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 38389a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 38399a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 38409a305230SLee Schermerhorn else 38419a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 38429a305230SLee Schermerhorn 3843ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", free_huge_pages); 3844a3437870SNishanth Aravamudan } 3845a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 3846a3437870SNishanth Aravamudan 3847a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 3848a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3849a3437870SNishanth Aravamudan { 38509a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3851ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); 3852a3437870SNishanth Aravamudan } 3853a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 3854a3437870SNishanth Aravamudan 3855a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 3856a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3857a3437870SNishanth Aravamudan { 38589a305230SLee Schermerhorn struct hstate *h; 38599a305230SLee Schermerhorn unsigned long surplus_huge_pages; 38609a305230SLee Schermerhorn int nid; 38619a305230SLee Schermerhorn 38629a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 38639a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 38649a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 38659a305230SLee Schermerhorn else 38669a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 38679a305230SLee Schermerhorn 3868ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", surplus_huge_pages); 3869a3437870SNishanth Aravamudan } 3870a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 3871a3437870SNishanth Aravamudan 387279dfc695SMike Kravetz static ssize_t demote_store(struct kobject *kobj, 387379dfc695SMike Kravetz struct kobj_attribute *attr, const char *buf, size_t len) 387479dfc695SMike Kravetz { 387579dfc695SMike Kravetz unsigned long nr_demote; 387679dfc695SMike Kravetz unsigned long nr_available; 387779dfc695SMike Kravetz nodemask_t nodes_allowed, *n_mask; 387879dfc695SMike Kravetz struct hstate *h; 38798eeda55fSLi zeming int err; 388079dfc695SMike Kravetz int nid; 388179dfc695SMike Kravetz 388279dfc695SMike Kravetz err = kstrtoul(buf, 10, &nr_demote); 388379dfc695SMike Kravetz if (err) 388479dfc695SMike Kravetz return err; 388579dfc695SMike Kravetz h = kobj_to_hstate(kobj, &nid); 388679dfc695SMike Kravetz 388779dfc695SMike Kravetz if (nid != NUMA_NO_NODE) { 388879dfc695SMike Kravetz init_nodemask_of_node(&nodes_allowed, nid); 388979dfc695SMike Kravetz n_mask = &nodes_allowed; 389079dfc695SMike Kravetz } else { 389179dfc695SMike Kravetz n_mask = &node_states[N_MEMORY]; 389279dfc695SMike Kravetz } 389379dfc695SMike Kravetz 389479dfc695SMike Kravetz /* Synchronize with other sysfs operations modifying huge pages */ 389579dfc695SMike Kravetz mutex_lock(&h->resize_lock); 389679dfc695SMike Kravetz spin_lock_irq(&hugetlb_lock); 389779dfc695SMike Kravetz 389879dfc695SMike Kravetz while (nr_demote) { 389979dfc695SMike Kravetz /* 390079dfc695SMike Kravetz * Check for available pages to demote each time thorough the 390179dfc695SMike Kravetz * loop as demote_pool_huge_page will drop hugetlb_lock. 390279dfc695SMike Kravetz */ 390379dfc695SMike Kravetz if (nid != NUMA_NO_NODE) 390479dfc695SMike Kravetz nr_available = h->free_huge_pages_node[nid]; 390579dfc695SMike Kravetz else 390679dfc695SMike Kravetz nr_available = h->free_huge_pages; 390779dfc695SMike Kravetz nr_available -= h->resv_huge_pages; 390879dfc695SMike Kravetz if (!nr_available) 390979dfc695SMike Kravetz break; 391079dfc695SMike Kravetz 391179dfc695SMike Kravetz err = demote_pool_huge_page(h, n_mask); 391279dfc695SMike Kravetz if (err) 391379dfc695SMike Kravetz break; 391479dfc695SMike Kravetz 391579dfc695SMike Kravetz nr_demote--; 391679dfc695SMike Kravetz } 391779dfc695SMike Kravetz 391879dfc695SMike Kravetz spin_unlock_irq(&hugetlb_lock); 391979dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 392079dfc695SMike Kravetz 392179dfc695SMike Kravetz if (err) 392279dfc695SMike Kravetz return err; 392379dfc695SMike Kravetz return len; 392479dfc695SMike Kravetz } 392579dfc695SMike Kravetz HSTATE_ATTR_WO(demote); 392679dfc695SMike Kravetz 392779dfc695SMike Kravetz static ssize_t demote_size_show(struct kobject *kobj, 392879dfc695SMike Kravetz struct kobj_attribute *attr, char *buf) 392979dfc695SMike Kravetz { 393012658abfSMiaohe Lin struct hstate *h = kobj_to_hstate(kobj, NULL); 393179dfc695SMike Kravetz unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K; 393279dfc695SMike Kravetz 393379dfc695SMike Kravetz return sysfs_emit(buf, "%lukB\n", demote_size); 393479dfc695SMike Kravetz } 393579dfc695SMike Kravetz 393679dfc695SMike Kravetz static ssize_t demote_size_store(struct kobject *kobj, 393779dfc695SMike Kravetz struct kobj_attribute *attr, 393879dfc695SMike Kravetz const char *buf, size_t count) 393979dfc695SMike Kravetz { 394079dfc695SMike Kravetz struct hstate *h, *demote_hstate; 394179dfc695SMike Kravetz unsigned long demote_size; 394279dfc695SMike Kravetz unsigned int demote_order; 394379dfc695SMike Kravetz 394479dfc695SMike Kravetz demote_size = (unsigned long)memparse(buf, NULL); 394579dfc695SMike Kravetz 394679dfc695SMike Kravetz demote_hstate = size_to_hstate(demote_size); 394779dfc695SMike Kravetz if (!demote_hstate) 394879dfc695SMike Kravetz return -EINVAL; 394979dfc695SMike Kravetz demote_order = demote_hstate->order; 3950a01f4390SMike Kravetz if (demote_order < HUGETLB_PAGE_ORDER) 3951a01f4390SMike Kravetz return -EINVAL; 395279dfc695SMike Kravetz 395379dfc695SMike Kravetz /* demote order must be smaller than hstate order */ 395412658abfSMiaohe Lin h = kobj_to_hstate(kobj, NULL); 395579dfc695SMike Kravetz if (demote_order >= h->order) 395679dfc695SMike Kravetz return -EINVAL; 395779dfc695SMike Kravetz 395879dfc695SMike Kravetz /* resize_lock synchronizes access to demote size and writes */ 395979dfc695SMike Kravetz mutex_lock(&h->resize_lock); 396079dfc695SMike Kravetz h->demote_order = demote_order; 396179dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 396279dfc695SMike Kravetz 396379dfc695SMike Kravetz return count; 396479dfc695SMike Kravetz } 396579dfc695SMike Kravetz HSTATE_ATTR(demote_size); 396679dfc695SMike Kravetz 3967a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 3968a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 3969a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 3970a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 3971a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 3972a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 397306808b08SLee Schermerhorn #ifdef CONFIG_NUMA 397406808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 397506808b08SLee Schermerhorn #endif 3976a3437870SNishanth Aravamudan NULL, 3977a3437870SNishanth Aravamudan }; 3978a3437870SNishanth Aravamudan 397967e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 3980a3437870SNishanth Aravamudan .attrs = hstate_attrs, 3981a3437870SNishanth Aravamudan }; 3982a3437870SNishanth Aravamudan 398379dfc695SMike Kravetz static struct attribute *hstate_demote_attrs[] = { 398479dfc695SMike Kravetz &demote_size_attr.attr, 398579dfc695SMike Kravetz &demote_attr.attr, 398679dfc695SMike Kravetz NULL, 398779dfc695SMike Kravetz }; 398879dfc695SMike Kravetz 398979dfc695SMike Kravetz static const struct attribute_group hstate_demote_attr_group = { 399079dfc695SMike Kravetz .attrs = hstate_demote_attrs, 399179dfc695SMike Kravetz }; 399279dfc695SMike Kravetz 3993094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 39949a305230SLee Schermerhorn struct kobject **hstate_kobjs, 399567e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 3996a3437870SNishanth Aravamudan { 3997a3437870SNishanth Aravamudan int retval; 3998972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 3999a3437870SNishanth Aravamudan 40009a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 40019a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 4002a3437870SNishanth Aravamudan return -ENOMEM; 4003a3437870SNishanth Aravamudan 40049a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 4005cc2205a6SMiaohe Lin if (retval) { 40069a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 4007cc2205a6SMiaohe Lin hstate_kobjs[hi] = NULL; 40083a6bdda0SMiaohe Lin return retval; 4009cc2205a6SMiaohe Lin } 4010a3437870SNishanth Aravamudan 401179dfc695SMike Kravetz if (h->demote_order) { 401201088a60SMiaohe Lin retval = sysfs_create_group(hstate_kobjs[hi], 401301088a60SMiaohe Lin &hstate_demote_attr_group); 401401088a60SMiaohe Lin if (retval) { 401579dfc695SMike Kravetz pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name); 401601088a60SMiaohe Lin sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group); 401701088a60SMiaohe Lin kobject_put(hstate_kobjs[hi]); 401801088a60SMiaohe Lin hstate_kobjs[hi] = NULL; 401901088a60SMiaohe Lin return retval; 402001088a60SMiaohe Lin } 402179dfc695SMike Kravetz } 402279dfc695SMike Kravetz 402301088a60SMiaohe Lin return 0; 4024a3437870SNishanth Aravamudan } 4025a3437870SNishanth Aravamudan 40269a305230SLee Schermerhorn #ifdef CONFIG_NUMA 4027a4a00b45SMuchun Song static bool hugetlb_sysfs_initialized __ro_after_init; 40289a305230SLee Schermerhorn 40299a305230SLee Schermerhorn /* 40309a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 403110fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 403210fbcf4cSKay Sievers * index of a node device or _hstate == node id. 403310fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 40349a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 40359a305230SLee Schermerhorn */ 40369a305230SLee Schermerhorn struct node_hstate { 40379a305230SLee Schermerhorn struct kobject *hugepages_kobj; 40389a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 40399a305230SLee Schermerhorn }; 4040b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 40419a305230SLee Schermerhorn 40429a305230SLee Schermerhorn /* 404310fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 40449a305230SLee Schermerhorn */ 40459a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 40469a305230SLee Schermerhorn &nr_hugepages_attr.attr, 40479a305230SLee Schermerhorn &free_hugepages_attr.attr, 40489a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 40499a305230SLee Schermerhorn NULL, 40509a305230SLee Schermerhorn }; 40519a305230SLee Schermerhorn 405267e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 40539a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 40549a305230SLee Schermerhorn }; 40559a305230SLee Schermerhorn 40569a305230SLee Schermerhorn /* 405710fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 40589a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 40599a305230SLee Schermerhorn */ 40609a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 40619a305230SLee Schermerhorn { 40629a305230SLee Schermerhorn int nid; 40639a305230SLee Schermerhorn 40649a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 40659a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 40669a305230SLee Schermerhorn int i; 40679a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 40689a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 40699a305230SLee Schermerhorn if (nidp) 40709a305230SLee Schermerhorn *nidp = nid; 40719a305230SLee Schermerhorn return &hstates[i]; 40729a305230SLee Schermerhorn } 40739a305230SLee Schermerhorn } 40749a305230SLee Schermerhorn 40759a305230SLee Schermerhorn BUG(); 40769a305230SLee Schermerhorn return NULL; 40779a305230SLee Schermerhorn } 40789a305230SLee Schermerhorn 40799a305230SLee Schermerhorn /* 408010fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 40819a305230SLee Schermerhorn * No-op if no hstate attributes attached. 40829a305230SLee Schermerhorn */ 4083a4a00b45SMuchun Song void hugetlb_unregister_node(struct node *node) 40849a305230SLee Schermerhorn { 40859a305230SLee Schermerhorn struct hstate *h; 408610fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 40879a305230SLee Schermerhorn 40889a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 40899b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 40909a305230SLee Schermerhorn 4091972dc4deSAneesh Kumar K.V for_each_hstate(h) { 4092972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 409301088a60SMiaohe Lin struct kobject *hstate_kobj = nhs->hstate_kobjs[idx]; 409401088a60SMiaohe Lin 409501088a60SMiaohe Lin if (!hstate_kobj) 409601088a60SMiaohe Lin continue; 409701088a60SMiaohe Lin if (h->demote_order) 409801088a60SMiaohe Lin sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group); 409901088a60SMiaohe Lin sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group); 410001088a60SMiaohe Lin kobject_put(hstate_kobj); 4101972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 4102972dc4deSAneesh Kumar K.V } 41039a305230SLee Schermerhorn 41049a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 41059a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 41069a305230SLee Schermerhorn } 41079a305230SLee Schermerhorn 41089a305230SLee Schermerhorn 41099a305230SLee Schermerhorn /* 411010fbcf4cSKay Sievers * Register hstate attributes for a single node device. 41119a305230SLee Schermerhorn * No-op if attributes already registered. 41129a305230SLee Schermerhorn */ 4113a4a00b45SMuchun Song void hugetlb_register_node(struct node *node) 41149a305230SLee Schermerhorn { 41159a305230SLee Schermerhorn struct hstate *h; 411610fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 41179a305230SLee Schermerhorn int err; 41189a305230SLee Schermerhorn 4119a4a00b45SMuchun Song if (!hugetlb_sysfs_initialized) 4120a4a00b45SMuchun Song return; 4121a4a00b45SMuchun Song 41229a305230SLee Schermerhorn if (nhs->hugepages_kobj) 41239a305230SLee Schermerhorn return; /* already allocated */ 41249a305230SLee Schermerhorn 41259a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 412610fbcf4cSKay Sievers &node->dev.kobj); 41279a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 41289a305230SLee Schermerhorn return; 41299a305230SLee Schermerhorn 41309a305230SLee Schermerhorn for_each_hstate(h) { 41319a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 41329a305230SLee Schermerhorn nhs->hstate_kobjs, 41339a305230SLee Schermerhorn &per_node_hstate_attr_group); 41349a305230SLee Schermerhorn if (err) { 4135282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s for node %d\n", 413610fbcf4cSKay Sievers h->name, node->dev.id); 41379a305230SLee Schermerhorn hugetlb_unregister_node(node); 41389a305230SLee Schermerhorn break; 41399a305230SLee Schermerhorn } 41409a305230SLee Schermerhorn } 41419a305230SLee Schermerhorn } 41429a305230SLee Schermerhorn 41439a305230SLee Schermerhorn /* 41449b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 414510fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 414610fbcf4cSKay Sievers * registered their associated device by this time. 41479a305230SLee Schermerhorn */ 41487d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 41499a305230SLee Schermerhorn { 41509a305230SLee Schermerhorn int nid; 41519a305230SLee Schermerhorn 4152a4a00b45SMuchun Song for_each_online_node(nid) 4153b958d4d0SMuchun Song hugetlb_register_node(node_devices[nid]); 41549a305230SLee Schermerhorn } 41559a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 41569a305230SLee Schermerhorn 41579a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 41589a305230SLee Schermerhorn { 41599a305230SLee Schermerhorn BUG(); 41609a305230SLee Schermerhorn if (nidp) 41619a305230SLee Schermerhorn *nidp = -1; 41629a305230SLee Schermerhorn return NULL; 41639a305230SLee Schermerhorn } 41649a305230SLee Schermerhorn 41659a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 41669a305230SLee Schermerhorn 41679a305230SLee Schermerhorn #endif 41689a305230SLee Schermerhorn 4169263b8998SMiaohe Lin #ifdef CONFIG_CMA 4170263b8998SMiaohe Lin static void __init hugetlb_cma_check(void); 4171263b8998SMiaohe Lin #else 4172263b8998SMiaohe Lin static inline __init void hugetlb_cma_check(void) 4173263b8998SMiaohe Lin { 4174263b8998SMiaohe Lin } 4175263b8998SMiaohe Lin #endif 4176263b8998SMiaohe Lin 4177a4a00b45SMuchun Song static void __init hugetlb_sysfs_init(void) 4178a4a00b45SMuchun Song { 4179a4a00b45SMuchun Song struct hstate *h; 4180a4a00b45SMuchun Song int err; 4181a4a00b45SMuchun Song 4182a4a00b45SMuchun Song hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 4183a4a00b45SMuchun Song if (!hugepages_kobj) 4184a4a00b45SMuchun Song return; 4185a4a00b45SMuchun Song 4186a4a00b45SMuchun Song for_each_hstate(h) { 4187a4a00b45SMuchun Song err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 4188a4a00b45SMuchun Song hstate_kobjs, &hstate_attr_group); 4189a4a00b45SMuchun Song if (err) 4190a4a00b45SMuchun Song pr_err("HugeTLB: Unable to add hstate %s", h->name); 4191a4a00b45SMuchun Song } 4192a4a00b45SMuchun Song 4193a4a00b45SMuchun Song #ifdef CONFIG_NUMA 4194a4a00b45SMuchun Song hugetlb_sysfs_initialized = true; 4195a4a00b45SMuchun Song #endif 4196a4a00b45SMuchun Song hugetlb_register_all_nodes(); 4197a4a00b45SMuchun Song } 4198a4a00b45SMuchun Song 4199a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 4200a3437870SNishanth Aravamudan { 42018382d914SDavidlohr Bueso int i; 42028382d914SDavidlohr Bueso 4203d6995da3SMike Kravetz BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < 4204d6995da3SMike Kravetz __NR_HPAGEFLAGS); 4205d6995da3SMike Kravetz 4206c2833a5bSMike Kravetz if (!hugepages_supported()) { 4207c2833a5bSMike Kravetz if (hugetlb_max_hstate || default_hstate_max_huge_pages) 4208c2833a5bSMike Kravetz pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); 42090ef89d25SBenjamin Herrenschmidt return 0; 4210d715cf80SLiam R. Howlett } 4211d715cf80SLiam R. Howlett 4212282f4214SMike Kravetz /* 4213282f4214SMike Kravetz * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some 4214282f4214SMike Kravetz * architectures depend on setup being done here. 4215282f4214SMike Kravetz */ 4216a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 4217282f4214SMike Kravetz if (!parsed_default_hugepagesz) { 4218282f4214SMike Kravetz /* 4219282f4214SMike Kravetz * If we did not parse a default huge page size, set 4220282f4214SMike Kravetz * default_hstate_idx to HPAGE_SIZE hstate. And, if the 4221282f4214SMike Kravetz * number of huge pages for this default size was implicitly 4222282f4214SMike Kravetz * specified, set that here as well. 4223282f4214SMike Kravetz * Note that the implicit setting will overwrite an explicit 4224282f4214SMike Kravetz * setting. A warning will be printed in this case. 4225282f4214SMike Kravetz */ 4226282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); 4227f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 4228282f4214SMike Kravetz if (default_hstate.max_huge_pages) { 4229282f4214SMike Kravetz char buf[32]; 4230282f4214SMike Kravetz 4231282f4214SMike Kravetz string_get_size(huge_page_size(&default_hstate), 4232282f4214SMike Kravetz 1, STRING_UNITS_2, buf, 32); 4233282f4214SMike Kravetz pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", 4234282f4214SMike Kravetz default_hstate.max_huge_pages, buf); 4235282f4214SMike Kravetz pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", 4236282f4214SMike Kravetz default_hstate_max_huge_pages); 4237282f4214SMike Kravetz } 4238282f4214SMike Kravetz default_hstate.max_huge_pages = 4239282f4214SMike Kravetz default_hstate_max_huge_pages; 4240b5389086SZhenguo Yao 42410a7a0f6fSPeng Liu for_each_online_node(i) 4242b5389086SZhenguo Yao default_hstate.max_huge_pages_node[i] = 4243b5389086SZhenguo Yao default_hugepages_in_node[i]; 4244282f4214SMike Kravetz } 4245f8b74815SVaishali Thakkar } 4246a3437870SNishanth Aravamudan 4247cf11e85fSRoman Gushchin hugetlb_cma_check(); 4248a3437870SNishanth Aravamudan hugetlb_init_hstates(); 4249aa888a74SAndi Kleen gather_bootmem_prealloc(); 4250a3437870SNishanth Aravamudan report_hugepages(); 4251a3437870SNishanth Aravamudan 4252a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 42537179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 42549a305230SLee Schermerhorn 42558382d914SDavidlohr Bueso #ifdef CONFIG_SMP 42568382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 42578382d914SDavidlohr Bueso #else 42588382d914SDavidlohr Bueso num_fault_mutexes = 1; 42598382d914SDavidlohr Bueso #endif 4260c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 42616da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 42626da2ec56SKees Cook GFP_KERNEL); 4263c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 42648382d914SDavidlohr Bueso 42658382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 4266c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 4267a3437870SNishanth Aravamudan return 0; 4268a3437870SNishanth Aravamudan } 42693e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 4270a3437870SNishanth Aravamudan 4271ae94da89SMike Kravetz /* Overwritten by architectures with more huge page sizes */ 4272ae94da89SMike Kravetz bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) 42739fee021dSVaishali Thakkar { 4274ae94da89SMike Kravetz return size == HPAGE_SIZE; 42759fee021dSVaishali Thakkar } 42769fee021dSVaishali Thakkar 4277d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 4278a3437870SNishanth Aravamudan { 4279a3437870SNishanth Aravamudan struct hstate *h; 42808faa8b07SAndi Kleen unsigned long i; 42818faa8b07SAndi Kleen 4282a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 4283a3437870SNishanth Aravamudan return; 4284a3437870SNishanth Aravamudan } 428547d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 4286a3437870SNishanth Aravamudan BUG_ON(order == 0); 428747d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 428829383967SMike Kravetz mutex_init(&h->resize_lock); 4289a3437870SNishanth Aravamudan h->order = order; 4290aca78307SMiaohe Lin h->mask = ~(huge_page_size(h) - 1); 42918faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 42928faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 42930edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 429454f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 429554f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 4296a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 4297c2c3a60aSMiaohe Lin huge_page_size(h)/SZ_1K); 42988faa8b07SAndi Kleen 4299a3437870SNishanth Aravamudan parsed_hstate = h; 4300a3437870SNishanth Aravamudan } 4301a3437870SNishanth Aravamudan 4302b5389086SZhenguo Yao bool __init __weak hugetlb_node_alloc_supported(void) 4303b5389086SZhenguo Yao { 4304b5389086SZhenguo Yao return true; 4305b5389086SZhenguo Yao } 4306f87442f4SPeng Liu 4307f87442f4SPeng Liu static void __init hugepages_clear_pages_in_node(void) 4308f87442f4SPeng Liu { 4309f87442f4SPeng Liu if (!hugetlb_max_hstate) { 4310f87442f4SPeng Liu default_hstate_max_huge_pages = 0; 4311f87442f4SPeng Liu memset(default_hugepages_in_node, 0, 431210395680SMiaohe Lin sizeof(default_hugepages_in_node)); 4313f87442f4SPeng Liu } else { 4314f87442f4SPeng Liu parsed_hstate->max_huge_pages = 0; 4315f87442f4SPeng Liu memset(parsed_hstate->max_huge_pages_node, 0, 431610395680SMiaohe Lin sizeof(parsed_hstate->max_huge_pages_node)); 4317f87442f4SPeng Liu } 4318f87442f4SPeng Liu } 4319f87442f4SPeng Liu 4320282f4214SMike Kravetz /* 4321282f4214SMike Kravetz * hugepages command line processing 4322282f4214SMike Kravetz * hugepages normally follows a valid hugepagsz or default_hugepagsz 4323282f4214SMike Kravetz * specification. If not, ignore the hugepages value. hugepages can also 4324282f4214SMike Kravetz * be the first huge page command line option in which case it implicitly 4325282f4214SMike Kravetz * specifies the number of huge pages for the default size. 4326282f4214SMike Kravetz */ 4327282f4214SMike Kravetz static int __init hugepages_setup(char *s) 4328a3437870SNishanth Aravamudan { 4329a3437870SNishanth Aravamudan unsigned long *mhp; 43308faa8b07SAndi Kleen static unsigned long *last_mhp; 4331b5389086SZhenguo Yao int node = NUMA_NO_NODE; 4332b5389086SZhenguo Yao int count; 4333b5389086SZhenguo Yao unsigned long tmp; 4334b5389086SZhenguo Yao char *p = s; 4335a3437870SNishanth Aravamudan 43369fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 4337282f4214SMike Kravetz pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); 43389fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 4339f81f6e4bSPeng Liu return 1; 43409fee021dSVaishali Thakkar } 4341282f4214SMike Kravetz 4342a3437870SNishanth Aravamudan /* 4343282f4214SMike Kravetz * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter 4344282f4214SMike Kravetz * yet, so this hugepages= parameter goes to the "default hstate". 4345282f4214SMike Kravetz * Otherwise, it goes with the previously parsed hugepagesz or 4346282f4214SMike Kravetz * default_hugepagesz. 4347a3437870SNishanth Aravamudan */ 43489fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 4349a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 4350a3437870SNishanth Aravamudan else 4351a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 4352a3437870SNishanth Aravamudan 43538faa8b07SAndi Kleen if (mhp == last_mhp) { 4354282f4214SMike Kravetz pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); 4355f81f6e4bSPeng Liu return 1; 43568faa8b07SAndi Kleen } 43578faa8b07SAndi Kleen 4358b5389086SZhenguo Yao while (*p) { 4359b5389086SZhenguo Yao count = 0; 4360b5389086SZhenguo Yao if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4361b5389086SZhenguo Yao goto invalid; 4362b5389086SZhenguo Yao /* Parameter is node format */ 4363b5389086SZhenguo Yao if (p[count] == ':') { 4364b5389086SZhenguo Yao if (!hugetlb_node_alloc_supported()) { 4365b5389086SZhenguo Yao pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n"); 4366f81f6e4bSPeng Liu return 1; 4367b5389086SZhenguo Yao } 43680a7a0f6fSPeng Liu if (tmp >= MAX_NUMNODES || !node_online(tmp)) 4369e79ce983SLiu Yuntao goto invalid; 43700a7a0f6fSPeng Liu node = array_index_nospec(tmp, MAX_NUMNODES); 4371b5389086SZhenguo Yao p += count + 1; 4372b5389086SZhenguo Yao /* Parse hugepages */ 4373b5389086SZhenguo Yao if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4374b5389086SZhenguo Yao goto invalid; 4375b5389086SZhenguo Yao if (!hugetlb_max_hstate) 4376b5389086SZhenguo Yao default_hugepages_in_node[node] = tmp; 4377b5389086SZhenguo Yao else 4378b5389086SZhenguo Yao parsed_hstate->max_huge_pages_node[node] = tmp; 4379b5389086SZhenguo Yao *mhp += tmp; 4380b5389086SZhenguo Yao /* Go to parse next node*/ 4381b5389086SZhenguo Yao if (p[count] == ',') 4382b5389086SZhenguo Yao p += count + 1; 4383b5389086SZhenguo Yao else 4384b5389086SZhenguo Yao break; 4385b5389086SZhenguo Yao } else { 4386b5389086SZhenguo Yao if (p != s) 4387b5389086SZhenguo Yao goto invalid; 4388b5389086SZhenguo Yao *mhp = tmp; 4389b5389086SZhenguo Yao break; 4390b5389086SZhenguo Yao } 4391b5389086SZhenguo Yao } 4392a3437870SNishanth Aravamudan 43938faa8b07SAndi Kleen /* 43948faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 439504adbc3fSMiaohe Lin * But we need to allocate gigantic hstates here early to still 43968faa8b07SAndi Kleen * use the bootmem allocator. 43978faa8b07SAndi Kleen */ 439804adbc3fSMiaohe Lin if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) 43998faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 44008faa8b07SAndi Kleen 44018faa8b07SAndi Kleen last_mhp = mhp; 44028faa8b07SAndi Kleen 4403a3437870SNishanth Aravamudan return 1; 4404b5389086SZhenguo Yao 4405b5389086SZhenguo Yao invalid: 4406b5389086SZhenguo Yao pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p); 4407f87442f4SPeng Liu hugepages_clear_pages_in_node(); 4408f81f6e4bSPeng Liu return 1; 4409a3437870SNishanth Aravamudan } 4410282f4214SMike Kravetz __setup("hugepages=", hugepages_setup); 4411e11bfbfcSNick Piggin 4412282f4214SMike Kravetz /* 4413282f4214SMike Kravetz * hugepagesz command line processing 4414282f4214SMike Kravetz * A specific huge page size can only be specified once with hugepagesz. 4415282f4214SMike Kravetz * hugepagesz is followed by hugepages on the command line. The global 4416282f4214SMike Kravetz * variable 'parsed_valid_hugepagesz' is used to determine if prior 4417282f4214SMike Kravetz * hugepagesz argument was valid. 4418282f4214SMike Kravetz */ 4419359f2544SMike Kravetz static int __init hugepagesz_setup(char *s) 4420e11bfbfcSNick Piggin { 4421359f2544SMike Kravetz unsigned long size; 4422282f4214SMike Kravetz struct hstate *h; 4423282f4214SMike Kravetz 4424282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4425359f2544SMike Kravetz size = (unsigned long)memparse(s, NULL); 4426359f2544SMike Kravetz 4427359f2544SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4428282f4214SMike Kravetz pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); 4429f81f6e4bSPeng Liu return 1; 4430359f2544SMike Kravetz } 4431359f2544SMike Kravetz 4432282f4214SMike Kravetz h = size_to_hstate(size); 4433282f4214SMike Kravetz if (h) { 4434282f4214SMike Kravetz /* 4435282f4214SMike Kravetz * hstate for this size already exists. This is normally 4436282f4214SMike Kravetz * an error, but is allowed if the existing hstate is the 4437282f4214SMike Kravetz * default hstate. More specifically, it is only allowed if 4438282f4214SMike Kravetz * the number of huge pages for the default hstate was not 4439282f4214SMike Kravetz * previously specified. 4440282f4214SMike Kravetz */ 4441282f4214SMike Kravetz if (!parsed_default_hugepagesz || h != &default_hstate || 4442282f4214SMike Kravetz default_hstate.max_huge_pages) { 4443282f4214SMike Kravetz pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); 4444f81f6e4bSPeng Liu return 1; 444538237830SMike Kravetz } 444638237830SMike Kravetz 4447282f4214SMike Kravetz /* 4448282f4214SMike Kravetz * No need to call hugetlb_add_hstate() as hstate already 4449282f4214SMike Kravetz * exists. But, do set parsed_hstate so that a following 4450282f4214SMike Kravetz * hugepages= parameter will be applied to this hstate. 4451282f4214SMike Kravetz */ 4452282f4214SMike Kravetz parsed_hstate = h; 4453282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4454e11bfbfcSNick Piggin return 1; 4455e11bfbfcSNick Piggin } 4456282f4214SMike Kravetz 4457359f2544SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4458282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4459359f2544SMike Kravetz return 1; 4460359f2544SMike Kravetz } 4461359f2544SMike Kravetz __setup("hugepagesz=", hugepagesz_setup); 4462359f2544SMike Kravetz 4463282f4214SMike Kravetz /* 4464282f4214SMike Kravetz * default_hugepagesz command line input 4465282f4214SMike Kravetz * Only one instance of default_hugepagesz allowed on command line. 4466282f4214SMike Kravetz */ 4467ae94da89SMike Kravetz static int __init default_hugepagesz_setup(char *s) 4468e11bfbfcSNick Piggin { 4469ae94da89SMike Kravetz unsigned long size; 4470b5389086SZhenguo Yao int i; 4471ae94da89SMike Kravetz 4472282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4473282f4214SMike Kravetz if (parsed_default_hugepagesz) { 4474282f4214SMike Kravetz pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); 4475f81f6e4bSPeng Liu return 1; 4476282f4214SMike Kravetz } 4477282f4214SMike Kravetz 4478282f4214SMike Kravetz size = (unsigned long)memparse(s, NULL); 4479282f4214SMike Kravetz 4480282f4214SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4481282f4214SMike Kravetz pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); 4482f81f6e4bSPeng Liu return 1; 4483282f4214SMike Kravetz } 4484282f4214SMike Kravetz 4485282f4214SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4486282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4487282f4214SMike Kravetz parsed_default_hugepagesz = true; 4488282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(size)); 4489282f4214SMike Kravetz 4490282f4214SMike Kravetz /* 4491282f4214SMike Kravetz * The number of default huge pages (for this size) could have been 4492282f4214SMike Kravetz * specified as the first hugetlb parameter: hugepages=X. If so, 4493282f4214SMike Kravetz * then default_hstate_max_huge_pages is set. If the default huge 4494282f4214SMike Kravetz * page size is gigantic (>= MAX_ORDER), then the pages must be 4495282f4214SMike Kravetz * allocated here from bootmem allocator. 4496282f4214SMike Kravetz */ 4497282f4214SMike Kravetz if (default_hstate_max_huge_pages) { 4498282f4214SMike Kravetz default_hstate.max_huge_pages = default_hstate_max_huge_pages; 44990a7a0f6fSPeng Liu for_each_online_node(i) 4500b5389086SZhenguo Yao default_hstate.max_huge_pages_node[i] = 4501b5389086SZhenguo Yao default_hugepages_in_node[i]; 4502282f4214SMike Kravetz if (hstate_is_gigantic(&default_hstate)) 4503282f4214SMike Kravetz hugetlb_hstate_alloc_pages(&default_hstate); 4504282f4214SMike Kravetz default_hstate_max_huge_pages = 0; 4505282f4214SMike Kravetz } 4506282f4214SMike Kravetz 4507e11bfbfcSNick Piggin return 1; 4508e11bfbfcSNick Piggin } 4509ae94da89SMike Kravetz __setup("default_hugepagesz=", default_hugepagesz_setup); 4510a3437870SNishanth Aravamudan 4511d2226ebdSFeng Tang static nodemask_t *policy_mbind_nodemask(gfp_t gfp) 4512d2226ebdSFeng Tang { 4513d2226ebdSFeng Tang #ifdef CONFIG_NUMA 4514d2226ebdSFeng Tang struct mempolicy *mpol = get_task_policy(current); 4515d2226ebdSFeng Tang 4516d2226ebdSFeng Tang /* 4517d2226ebdSFeng Tang * Only enforce MPOL_BIND policy which overlaps with cpuset policy 4518d2226ebdSFeng Tang * (from policy_nodemask) specifically for hugetlb case 4519d2226ebdSFeng Tang */ 4520d2226ebdSFeng Tang if (mpol->mode == MPOL_BIND && 4521d2226ebdSFeng Tang (apply_policy_zone(mpol, gfp_zone(gfp)) && 4522d2226ebdSFeng Tang cpuset_nodemask_valid_mems_allowed(&mpol->nodes))) 4523d2226ebdSFeng Tang return &mpol->nodes; 4524d2226ebdSFeng Tang #endif 4525d2226ebdSFeng Tang return NULL; 4526d2226ebdSFeng Tang } 4527d2226ebdSFeng Tang 45288ca39e68SMuchun Song static unsigned int allowed_mems_nr(struct hstate *h) 45298a213460SNishanth Aravamudan { 45308a213460SNishanth Aravamudan int node; 45318a213460SNishanth Aravamudan unsigned int nr = 0; 4532d2226ebdSFeng Tang nodemask_t *mbind_nodemask; 45338ca39e68SMuchun Song unsigned int *array = h->free_huge_pages_node; 45348ca39e68SMuchun Song gfp_t gfp_mask = htlb_alloc_mask(h); 45358a213460SNishanth Aravamudan 4536d2226ebdSFeng Tang mbind_nodemask = policy_mbind_nodemask(gfp_mask); 45378ca39e68SMuchun Song for_each_node_mask(node, cpuset_current_mems_allowed) { 4538d2226ebdSFeng Tang if (!mbind_nodemask || node_isset(node, *mbind_nodemask)) 45398a213460SNishanth Aravamudan nr += array[node]; 45408ca39e68SMuchun Song } 45418a213460SNishanth Aravamudan 45428a213460SNishanth Aravamudan return nr; 45438a213460SNishanth Aravamudan } 45448a213460SNishanth Aravamudan 45458a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 454617743798SMuchun Song static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, 454717743798SMuchun Song void *buffer, size_t *length, 454817743798SMuchun Song loff_t *ppos, unsigned long *out) 454917743798SMuchun Song { 455017743798SMuchun Song struct ctl_table dup_table; 455117743798SMuchun Song 455217743798SMuchun Song /* 455317743798SMuchun Song * In order to avoid races with __do_proc_doulongvec_minmax(), we 455417743798SMuchun Song * can duplicate the @table and alter the duplicate of it. 455517743798SMuchun Song */ 455617743798SMuchun Song dup_table = *table; 455717743798SMuchun Song dup_table.data = out; 455817743798SMuchun Song 455917743798SMuchun Song return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); 456017743798SMuchun Song } 456117743798SMuchun Song 456206808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 456306808b08SLee Schermerhorn struct ctl_table *table, int write, 456432927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 45651da177e4SLinus Torvalds { 4566e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 4567238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 456808d4a246SMichal Hocko int ret; 4569e5ff2159SAndi Kleen 4570457c1b27SNishanth Aravamudan if (!hugepages_supported()) 457186613628SJan Stancek return -EOPNOTSUPP; 4572457c1b27SNishanth Aravamudan 457317743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 457417743798SMuchun Song &tmp); 457508d4a246SMichal Hocko if (ret) 457608d4a246SMichal Hocko goto out; 4577e5ff2159SAndi Kleen 4578238d3c13SDavid Rientjes if (write) 4579238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 4580238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 458108d4a246SMichal Hocko out: 458208d4a246SMichal Hocko return ret; 45831da177e4SLinus Torvalds } 4584396faf03SMel Gorman 458506808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 458632927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 458706808b08SLee Schermerhorn { 458806808b08SLee Schermerhorn 458906808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 459006808b08SLee Schermerhorn buffer, length, ppos); 459106808b08SLee Schermerhorn } 459206808b08SLee Schermerhorn 459306808b08SLee Schermerhorn #ifdef CONFIG_NUMA 459406808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 459532927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 459606808b08SLee Schermerhorn { 459706808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 459806808b08SLee Schermerhorn buffer, length, ppos); 459906808b08SLee Schermerhorn } 460006808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 460106808b08SLee Schermerhorn 4602a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 460332927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 4604a3d0c6aaSNishanth Aravamudan { 4605a5516438SAndi Kleen struct hstate *h = &default_hstate; 4606e5ff2159SAndi Kleen unsigned long tmp; 460708d4a246SMichal Hocko int ret; 4608e5ff2159SAndi Kleen 4609457c1b27SNishanth Aravamudan if (!hugepages_supported()) 461086613628SJan Stancek return -EOPNOTSUPP; 4611457c1b27SNishanth Aravamudan 4612e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 4613e5ff2159SAndi Kleen 4614bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 4615adbe8726SEric B Munson return -EINVAL; 4616adbe8726SEric B Munson 461717743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 461817743798SMuchun Song &tmp); 461908d4a246SMichal Hocko if (ret) 462008d4a246SMichal Hocko goto out; 4621e5ff2159SAndi Kleen 4622e5ff2159SAndi Kleen if (write) { 4623db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4624e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 4625db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4626e5ff2159SAndi Kleen } 462708d4a246SMichal Hocko out: 462808d4a246SMichal Hocko return ret; 4629a3d0c6aaSNishanth Aravamudan } 4630a3d0c6aaSNishanth Aravamudan 46311da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 46321da177e4SLinus Torvalds 4633e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 46341da177e4SLinus Torvalds { 4635fcb2b0c5SRoman Gushchin struct hstate *h; 4636fcb2b0c5SRoman Gushchin unsigned long total = 0; 4637fcb2b0c5SRoman Gushchin 4638457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4639457c1b27SNishanth Aravamudan return; 4640fcb2b0c5SRoman Gushchin 4641fcb2b0c5SRoman Gushchin for_each_hstate(h) { 4642fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 4643fcb2b0c5SRoman Gushchin 4644aca78307SMiaohe Lin total += huge_page_size(h) * count; 4645fcb2b0c5SRoman Gushchin 4646fcb2b0c5SRoman Gushchin if (h == &default_hstate) 4647e1759c21SAlexey Dobriyan seq_printf(m, 46481da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 46491da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 4650b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 46517893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 46524f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 4653fcb2b0c5SRoman Gushchin count, 4654a5516438SAndi Kleen h->free_huge_pages, 4655a5516438SAndi Kleen h->resv_huge_pages, 4656a5516438SAndi Kleen h->surplus_huge_pages, 4657aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4658fcb2b0c5SRoman Gushchin } 4659fcb2b0c5SRoman Gushchin 4660aca78307SMiaohe Lin seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); 46611da177e4SLinus Torvalds } 46621da177e4SLinus Torvalds 46637981593bSJoe Perches int hugetlb_report_node_meminfo(char *buf, int len, int nid) 46641da177e4SLinus Torvalds { 4665a5516438SAndi Kleen struct hstate *h = &default_hstate; 46667981593bSJoe Perches 4667457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4668457c1b27SNishanth Aravamudan return 0; 46697981593bSJoe Perches 46707981593bSJoe Perches return sysfs_emit_at(buf, len, 46711da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 4672a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 4673a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 4674a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 4675a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 4676a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 46771da177e4SLinus Torvalds } 46781da177e4SLinus Torvalds 4679dcadcf1cSGang Li void hugetlb_show_meminfo_node(int nid) 4680949f7ec5SDavid Rientjes { 4681949f7ec5SDavid Rientjes struct hstate *h; 4682949f7ec5SDavid Rientjes 4683457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4684457c1b27SNishanth Aravamudan return; 4685457c1b27SNishanth Aravamudan 4686949f7ec5SDavid Rientjes for_each_hstate(h) 4687dcadcf1cSGang Li printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 4688949f7ec5SDavid Rientjes nid, 4689949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 4690949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 4691949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 4692aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4693949f7ec5SDavid Rientjes } 4694949f7ec5SDavid Rientjes 46955d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 46965d317b2bSNaoya Horiguchi { 46975d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 46985d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 46995d317b2bSNaoya Horiguchi } 47005d317b2bSNaoya Horiguchi 47011da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 47021da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 47031da177e4SLinus Torvalds { 4704d0028588SWanpeng Li struct hstate *h; 4705d0028588SWanpeng Li unsigned long nr_total_pages = 0; 4706d0028588SWanpeng Li 4707d0028588SWanpeng Li for_each_hstate(h) 4708d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 4709d0028588SWanpeng Li return nr_total_pages; 47101da177e4SLinus Torvalds } 47111da177e4SLinus Torvalds 4712a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 4713fc1b8a73SMel Gorman { 4714fc1b8a73SMel Gorman int ret = -ENOMEM; 4715fc1b8a73SMel Gorman 47160aa7f354SMiaohe Lin if (!delta) 47170aa7f354SMiaohe Lin return 0; 47180aa7f354SMiaohe Lin 4719db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4720fc1b8a73SMel Gorman /* 4721fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 4722fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 4723fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 4724fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 4725fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 4726fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 4727fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 4728fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 4729fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 4730fc1b8a73SMel Gorman * 4731fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 4732fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 4733fc1b8a73SMel Gorman * we fall back to check against current free page availability as 4734fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 4735fc1b8a73SMel Gorman * semantics that cpuset has. 47368ca39e68SMuchun Song * 47378ca39e68SMuchun Song * Apart from cpuset, we also have memory policy mechanism that 47388ca39e68SMuchun Song * also determines from which node the kernel will allocate memory 47398ca39e68SMuchun Song * in a NUMA system. So similar to cpuset, we also should consider 47408ca39e68SMuchun Song * the memory policy of the current task. Similar to the description 47418ca39e68SMuchun Song * above. 4742fc1b8a73SMel Gorman */ 4743fc1b8a73SMel Gorman if (delta > 0) { 4744a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 4745fc1b8a73SMel Gorman goto out; 4746fc1b8a73SMel Gorman 47478ca39e68SMuchun Song if (delta > allowed_mems_nr(h)) { 4748a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 4749fc1b8a73SMel Gorman goto out; 4750fc1b8a73SMel Gorman } 4751fc1b8a73SMel Gorman } 4752fc1b8a73SMel Gorman 4753fc1b8a73SMel Gorman ret = 0; 4754fc1b8a73SMel Gorman if (delta < 0) 4755a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 4756fc1b8a73SMel Gorman 4757fc1b8a73SMel Gorman out: 4758db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4759fc1b8a73SMel Gorman return ret; 4760fc1b8a73SMel Gorman } 4761fc1b8a73SMel Gorman 476284afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 476384afd99bSAndy Whitcroft { 4764f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 476584afd99bSAndy Whitcroft 476684afd99bSAndy Whitcroft /* 4767612b8a31SMike Kravetz * HPAGE_RESV_OWNER indicates a private mapping. 476884afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 476984afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 477084afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 477125985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 477284afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 477384afd99bSAndy Whitcroft * new reference here without additional locking. 477484afd99bSAndy Whitcroft */ 477509a26e83SMike Kravetz if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 477609a26e83SMike Kravetz resv_map_dup_hugetlb_cgroup_uncharge_info(resv); 4777f522c3acSJoonsoo Kim kref_get(&resv->refs); 477884afd99bSAndy Whitcroft } 47798d9bfb26SMike Kravetz 4780131a79b4SMike Kravetz /* 4781131a79b4SMike Kravetz * vma_lock structure for sharable mappings is vma specific. 4782612b8a31SMike Kravetz * Clear old pointer (if copied via vm_area_dup) and allocate 4783612b8a31SMike Kravetz * new structure. Before clearing, make sure vma_lock is not 4784612b8a31SMike Kravetz * for this vma. 4785131a79b4SMike Kravetz */ 4786131a79b4SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 4787612b8a31SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 4788612b8a31SMike Kravetz 4789612b8a31SMike Kravetz if (vma_lock) { 4790612b8a31SMike Kravetz if (vma_lock->vma != vma) { 4791131a79b4SMike Kravetz vma->vm_private_data = NULL; 47928d9bfb26SMike Kravetz hugetlb_vma_lock_alloc(vma); 4793612b8a31SMike Kravetz } else 4794612b8a31SMike Kravetz pr_warn("HugeTLB: vma_lock already exists in %s.\n", __func__); 4795612b8a31SMike Kravetz } else 4796612b8a31SMike Kravetz hugetlb_vma_lock_alloc(vma); 479709a26e83SMike Kravetz } 4798131a79b4SMike Kravetz } 479984afd99bSAndy Whitcroft 4800a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 4801a1e78772SMel Gorman { 4802a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 48038d9bfb26SMike Kravetz struct resv_map *resv; 480490481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 48054e35f483SJoonsoo Kim unsigned long reserve, start, end; 48061c5ecae3SMike Kravetz long gbl_reserve; 480784afd99bSAndy Whitcroft 48088d9bfb26SMike Kravetz hugetlb_vma_lock_free(vma); 48098d9bfb26SMike Kravetz 48108d9bfb26SMike Kravetz resv = vma_resv_map(vma); 48114e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 48124e35f483SJoonsoo Kim return; 48134e35f483SJoonsoo Kim 4814a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 4815a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 481684afd99bSAndy Whitcroft 48174e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 4818e9fe92aeSMina Almasry hugetlb_cgroup_uncharge_counter(resv, start, end); 48197251ff78SAdam Litke if (reserve) { 48201c5ecae3SMike Kravetz /* 48211c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 48221c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 48231c5ecae3SMike Kravetz */ 48241c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 48251c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 48267251ff78SAdam Litke } 4827e9fe92aeSMina Almasry 4828e9fe92aeSMina Almasry kref_put(&resv->refs, resv_map_release); 4829a1e78772SMel Gorman } 4830a1e78772SMel Gorman 483131383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 483231383c68SDan Williams { 483331383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 483431383c68SDan Williams return -EINVAL; 4835b30c14cdSJames Houghton 4836b30c14cdSJames Houghton /* 4837b30c14cdSJames Houghton * PMD sharing is only possible for PUD_SIZE-aligned address ranges 4838b30c14cdSJames Houghton * in HugeTLB VMAs. If we will lose PUD_SIZE alignment due to this 4839b30c14cdSJames Houghton * split, unshare PMDs in the PUD_SIZE interval surrounding addr now. 4840b30c14cdSJames Houghton */ 4841b30c14cdSJames Houghton if (addr & ~PUD_MASK) { 4842b30c14cdSJames Houghton /* 4843b30c14cdSJames Houghton * hugetlb_vm_op_split is called right before we attempt to 4844b30c14cdSJames Houghton * split the VMA. We will need to unshare PMDs in the old and 4845b30c14cdSJames Houghton * new VMAs, so let's unshare before we split. 4846b30c14cdSJames Houghton */ 4847b30c14cdSJames Houghton unsigned long floor = addr & PUD_MASK; 4848b30c14cdSJames Houghton unsigned long ceil = floor + PUD_SIZE; 4849b30c14cdSJames Houghton 4850b30c14cdSJames Houghton if (floor >= vma->vm_start && ceil <= vma->vm_end) 4851b30c14cdSJames Houghton hugetlb_unshare_pmds(vma, floor, ceil); 4852b30c14cdSJames Houghton } 4853b30c14cdSJames Houghton 485431383c68SDan Williams return 0; 485531383c68SDan Williams } 485631383c68SDan Williams 485705ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 485805ea8860SDan Williams { 4859aca78307SMiaohe Lin return huge_page_size(hstate_vma(vma)); 486005ea8860SDan Williams } 486105ea8860SDan Williams 48621da177e4SLinus Torvalds /* 48631da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 48641da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 48656c26d310SMiaohe Lin * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get 48661da177e4SLinus Torvalds * this far. 48671da177e4SLinus Torvalds */ 4868b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 48691da177e4SLinus Torvalds { 48701da177e4SLinus Torvalds BUG(); 4871d0217ac0SNick Piggin return 0; 48721da177e4SLinus Torvalds } 48731da177e4SLinus Torvalds 4874eec3636aSJane Chu /* 4875eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 4876eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 4877eec3636aSJane Chu * This is because under System V memory model, mappings created via 4878eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 4879eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 4880eec3636aSJane Chu */ 4881f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 4882d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 488384afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 4884a1e78772SMel Gorman .close = hugetlb_vm_op_close, 4885dd3b614fSDmitry Safonov .may_split = hugetlb_vm_op_split, 488605ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 48871da177e4SLinus Torvalds }; 48881da177e4SLinus Torvalds 48891e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 48901e8f889bSDavid Gibson int writable) 489163551ae0SDavid Gibson { 489263551ae0SDavid Gibson pte_t entry; 489379c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 489463551ae0SDavid Gibson 48951e8f889bSDavid Gibson if (writable) { 4896106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 4897106c992aSGerald Schaefer vma->vm_page_prot))); 489863551ae0SDavid Gibson } else { 4899106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 4900106c992aSGerald Schaefer vma->vm_page_prot)); 490163551ae0SDavid Gibson } 490263551ae0SDavid Gibson entry = pte_mkyoung(entry); 490379c1c594SChristophe Leroy entry = arch_make_huge_pte(entry, shift, vma->vm_flags); 490463551ae0SDavid Gibson 490563551ae0SDavid Gibson return entry; 490663551ae0SDavid Gibson } 490763551ae0SDavid Gibson 49081e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 49091e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 49101e8f889bSDavid Gibson { 49111e8f889bSDavid Gibson pte_t entry; 49121e8f889bSDavid Gibson 4913106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 491432f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 49154b3073e1SRussell King update_mmu_cache(vma, address, ptep); 49161e8f889bSDavid Gibson } 49171e8f889bSDavid Gibson 4918d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 49194a705fefSNaoya Horiguchi { 49204a705fefSNaoya Horiguchi swp_entry_t swp; 49214a705fefSNaoya Horiguchi 49224a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 4923d5ed7444SAneesh Kumar K.V return false; 49244a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4925d79d176aSBaoquan He if (is_migration_entry(swp)) 4926d5ed7444SAneesh Kumar K.V return true; 49274a705fefSNaoya Horiguchi else 4928d5ed7444SAneesh Kumar K.V return false; 49294a705fefSNaoya Horiguchi } 49304a705fefSNaoya Horiguchi 49313e5c3600SBaoquan He static bool is_hugetlb_entry_hwpoisoned(pte_t pte) 49324a705fefSNaoya Horiguchi { 49334a705fefSNaoya Horiguchi swp_entry_t swp; 49344a705fefSNaoya Horiguchi 49354a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 49363e5c3600SBaoquan He return false; 49374a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4938d79d176aSBaoquan He if (is_hwpoison_entry(swp)) 49393e5c3600SBaoquan He return true; 49404a705fefSNaoya Horiguchi else 49413e5c3600SBaoquan He return false; 49424a705fefSNaoya Horiguchi } 49431e8f889bSDavid Gibson 49444eae4efaSPeter Xu static void 49454eae4efaSPeter Xu hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, 49464eae4efaSPeter Xu struct page *new_page) 49474eae4efaSPeter Xu { 49484eae4efaSPeter Xu __SetPageUptodate(new_page); 49494eae4efaSPeter Xu hugepage_add_new_anon_rmap(new_page, vma, addr); 49501eba86c0SPasha Tatashin set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1)); 49514eae4efaSPeter Xu hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); 49524eae4efaSPeter Xu SetHPageMigratable(new_page); 49534eae4efaSPeter Xu } 49544eae4efaSPeter Xu 495563551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 4956bc70fbf2SPeter Xu struct vm_area_struct *dst_vma, 4957bc70fbf2SPeter Xu struct vm_area_struct *src_vma) 495863551ae0SDavid Gibson { 49593aa4ed80SMiaohe Lin pte_t *src_pte, *dst_pte, entry; 496063551ae0SDavid Gibson struct page *ptepage; 49611c59827dSHugh Dickins unsigned long addr; 4962bc70fbf2SPeter Xu bool cow = is_cow_mapping(src_vma->vm_flags); 4963bc70fbf2SPeter Xu struct hstate *h = hstate_vma(src_vma); 4964a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 49654eae4efaSPeter Xu unsigned long npages = pages_per_huge_page(h); 4966ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 4967e95a9851SMike Kravetz unsigned long last_addr_mask; 4968e8569dd2SAndreas Sandberg int ret = 0; 49691e8f889bSDavid Gibson 4970ac46d4f3SJérôme Glisse if (cow) { 49717d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src, 4972bc70fbf2SPeter Xu src_vma->vm_start, 4973bc70fbf2SPeter Xu src_vma->vm_end); 4974ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4975623a1ddfSDavid Hildenbrand mmap_assert_write_locked(src); 4976623a1ddfSDavid Hildenbrand raw_write_seqcount_begin(&src->write_protect_seq); 497740549ba8SMike Kravetz } else { 497840549ba8SMike Kravetz /* 497940549ba8SMike Kravetz * For shared mappings the vma lock must be held before 49809c67a207SPeter Xu * calling hugetlb_walk() in the src vma. Otherwise, the 498140549ba8SMike Kravetz * returned ptep could go away if part of a shared pmd and 498240549ba8SMike Kravetz * another thread calls huge_pmd_unshare. 498340549ba8SMike Kravetz */ 498440549ba8SMike Kravetz hugetlb_vma_lock_read(src_vma); 4985ac46d4f3SJérôme Glisse } 4986e8569dd2SAndreas Sandberg 4987e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 4988bc70fbf2SPeter Xu for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) { 4989cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 49909c67a207SPeter Xu src_pte = hugetlb_walk(src_vma, addr, sz); 4991e95a9851SMike Kravetz if (!src_pte) { 4992e95a9851SMike Kravetz addr |= last_addr_mask; 4993c74df32cSHugh Dickins continue; 4994e95a9851SMike Kravetz } 4995bc70fbf2SPeter Xu dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz); 4996e8569dd2SAndreas Sandberg if (!dst_pte) { 4997e8569dd2SAndreas Sandberg ret = -ENOMEM; 4998e8569dd2SAndreas Sandberg break; 4999e8569dd2SAndreas Sandberg } 5000c5c99429SLarry Woodman 50015e41540cSMike Kravetz /* 50025e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 50035e41540cSMike Kravetz * 50043aa4ed80SMiaohe Lin * dst_pte == src_pte is the common case of src/dest sharing. 50055e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 50063aa4ed80SMiaohe Lin * another vma. So page_count of ptep page is checked instead 50073aa4ed80SMiaohe Lin * to reliably determine whether pte is shared. 50085e41540cSMike Kravetz */ 50093aa4ed80SMiaohe Lin if (page_count(virt_to_page(dst_pte)) > 1) { 5010e95a9851SMike Kravetz addr |= last_addr_mask; 5011c5c99429SLarry Woodman continue; 5012e95a9851SMike Kravetz } 5013c5c99429SLarry Woodman 5014cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 5015cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 5016cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 50174a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 50184eae4efaSPeter Xu again: 50193aa4ed80SMiaohe Lin if (huge_pte_none(entry)) { 50205e41540cSMike Kravetz /* 50213aa4ed80SMiaohe Lin * Skip if src entry none. 50225e41540cSMike Kravetz */ 50234a705fefSNaoya Horiguchi ; 5024c2cb0dccSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) { 5025c2cb0dccSNaoya Horiguchi bool uffd_wp = huge_pte_uffd_wp(entry); 5026c2cb0dccSNaoya Horiguchi 5027c2cb0dccSNaoya Horiguchi if (!userfaultfd_wp(dst_vma) && uffd_wp) 5028c2cb0dccSNaoya Horiguchi entry = huge_pte_clear_uffd_wp(entry); 5029c2cb0dccSNaoya Horiguchi set_huge_pte_at(dst, addr, dst_pte, entry); 5030c2cb0dccSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry))) { 50314a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 5032bc70fbf2SPeter Xu bool uffd_wp = huge_pte_uffd_wp(entry); 50334a705fefSNaoya Horiguchi 50346c287605SDavid Hildenbrand if (!is_readable_migration_entry(swp_entry) && cow) { 50354a705fefSNaoya Horiguchi /* 50364a705fefSNaoya Horiguchi * COW mappings require pages in both 50374a705fefSNaoya Horiguchi * parent and child to be set to read. 50384a705fefSNaoya Horiguchi */ 50394dd845b5SAlistair Popple swp_entry = make_readable_migration_entry( 50404dd845b5SAlistair Popple swp_offset(swp_entry)); 50414a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 5042bc70fbf2SPeter Xu if (userfaultfd_wp(src_vma) && uffd_wp) 5043bc70fbf2SPeter Xu entry = huge_pte_mkuffd_wp(entry); 504418f39629SQi Zheng set_huge_pte_at(src, addr, src_pte, entry); 50454a705fefSNaoya Horiguchi } 5046bc70fbf2SPeter Xu if (!userfaultfd_wp(dst_vma) && uffd_wp) 5047bc70fbf2SPeter Xu entry = huge_pte_clear_uffd_wp(entry); 504818f39629SQi Zheng set_huge_pte_at(dst, addr, dst_pte, entry); 5049bc70fbf2SPeter Xu } else if (unlikely(is_pte_marker(entry))) { 50507e3ce3f8SPeter Xu /* No swap on hugetlb */ 50517e3ce3f8SPeter Xu WARN_ON_ONCE( 50527e3ce3f8SPeter Xu is_swapin_error_entry(pte_to_swp_entry(entry))); 5053bc70fbf2SPeter Xu /* 5054bc70fbf2SPeter Xu * We copy the pte marker only if the dst vma has 5055bc70fbf2SPeter Xu * uffd-wp enabled. 5056bc70fbf2SPeter Xu */ 5057bc70fbf2SPeter Xu if (userfaultfd_wp(dst_vma)) 5058bc70fbf2SPeter Xu set_huge_pte_at(dst, addr, dst_pte, entry); 50594a705fefSNaoya Horiguchi } else { 50604eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 50614eae4efaSPeter Xu ptepage = pte_page(entry); 50624eae4efaSPeter Xu get_page(ptepage); 50634eae4efaSPeter Xu 50644eae4efaSPeter Xu /* 5065fb3d824dSDavid Hildenbrand * Failing to duplicate the anon rmap is a rare case 5066fb3d824dSDavid Hildenbrand * where we see pinned hugetlb pages while they're 5067fb3d824dSDavid Hildenbrand * prone to COW. We need to do the COW earlier during 5068fb3d824dSDavid Hildenbrand * fork. 50694eae4efaSPeter Xu * 50704eae4efaSPeter Xu * When pre-allocating the page or copying data, we 50714eae4efaSPeter Xu * need to be without the pgtable locks since we could 50724eae4efaSPeter Xu * sleep during the process. 50734eae4efaSPeter Xu */ 5074fb3d824dSDavid Hildenbrand if (!PageAnon(ptepage)) { 5075fb3d824dSDavid Hildenbrand page_dup_file_rmap(ptepage, true); 5076bc70fbf2SPeter Xu } else if (page_try_dup_anon_rmap(ptepage, true, 5077bc70fbf2SPeter Xu src_vma)) { 50784eae4efaSPeter Xu pte_t src_pte_old = entry; 50794eae4efaSPeter Xu struct page *new; 50804eae4efaSPeter Xu 50814eae4efaSPeter Xu spin_unlock(src_ptl); 50824eae4efaSPeter Xu spin_unlock(dst_ptl); 50834eae4efaSPeter Xu /* Do not use reserve as it's private owned */ 5084bc70fbf2SPeter Xu new = alloc_huge_page(dst_vma, addr, 1); 50854eae4efaSPeter Xu if (IS_ERR(new)) { 50864eae4efaSPeter Xu put_page(ptepage); 50874eae4efaSPeter Xu ret = PTR_ERR(new); 50884eae4efaSPeter Xu break; 50894eae4efaSPeter Xu } 5090bc70fbf2SPeter Xu copy_user_huge_page(new, ptepage, addr, dst_vma, 50914eae4efaSPeter Xu npages); 50924eae4efaSPeter Xu put_page(ptepage); 50934eae4efaSPeter Xu 50944eae4efaSPeter Xu /* Install the new huge page if src pte stable */ 50954eae4efaSPeter Xu dst_ptl = huge_pte_lock(h, dst, dst_pte); 50964eae4efaSPeter Xu src_ptl = huge_pte_lockptr(h, src, src_pte); 50974eae4efaSPeter Xu spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 50984eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 50994eae4efaSPeter Xu if (!pte_same(src_pte_old, entry)) { 5100bc70fbf2SPeter Xu restore_reserve_on_error(h, dst_vma, addr, 5101846be085SMike Kravetz new); 51024eae4efaSPeter Xu put_page(new); 51033aa4ed80SMiaohe Lin /* huge_ptep of dst_pte won't change as in child */ 51044eae4efaSPeter Xu goto again; 51054eae4efaSPeter Xu } 5106bc70fbf2SPeter Xu hugetlb_install_page(dst_vma, dst_pte, addr, new); 51074eae4efaSPeter Xu spin_unlock(src_ptl); 51084eae4efaSPeter Xu spin_unlock(dst_ptl); 51094eae4efaSPeter Xu continue; 51104eae4efaSPeter Xu } 51114eae4efaSPeter Xu 511234ee645eSJoerg Roedel if (cow) { 51130f10851eSJérôme Glisse /* 51140f10851eSJérôme Glisse * No need to notify as we are downgrading page 51150f10851eSJérôme Glisse * table protection not changing it to point 51160f10851eSJérôme Glisse * to a new page. 51170f10851eSJérôme Glisse * 5118ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 51190f10851eSJérôme Glisse */ 51207f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 512184894e1cSPeter Xu entry = huge_pte_wrprotect(entry); 512234ee645eSJoerg Roedel } 51234eae4efaSPeter Xu 512463551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 51254eae4efaSPeter Xu hugetlb_count_add(npages, dst); 51261c59827dSHugh Dickins } 5127cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 5128cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 512963551ae0SDavid Gibson } 513063551ae0SDavid Gibson 5131623a1ddfSDavid Hildenbrand if (cow) { 5132623a1ddfSDavid Hildenbrand raw_write_seqcount_end(&src->write_protect_seq); 5133ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 513440549ba8SMike Kravetz } else { 513540549ba8SMike Kravetz hugetlb_vma_unlock_read(src_vma); 5136623a1ddfSDavid Hildenbrand } 5137e8569dd2SAndreas Sandberg 5138e8569dd2SAndreas Sandberg return ret; 513963551ae0SDavid Gibson } 514063551ae0SDavid Gibson 5141550a7d60SMina Almasry static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr, 5142db110a99SAneesh Kumar K.V unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte) 5143550a7d60SMina Almasry { 5144550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 5145550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 5146550a7d60SMina Almasry spinlock_t *src_ptl, *dst_ptl; 5147db110a99SAneesh Kumar K.V pte_t pte; 5148550a7d60SMina Almasry 5149550a7d60SMina Almasry dst_ptl = huge_pte_lock(h, mm, dst_pte); 5150550a7d60SMina Almasry src_ptl = huge_pte_lockptr(h, mm, src_pte); 5151550a7d60SMina Almasry 5152550a7d60SMina Almasry /* 5153550a7d60SMina Almasry * We don't have to worry about the ordering of src and dst ptlocks 51548651a137SLorenzo Stoakes * because exclusive mmap_lock (or the i_mmap_lock) prevents deadlock. 5155550a7d60SMina Almasry */ 5156550a7d60SMina Almasry if (src_ptl != dst_ptl) 5157550a7d60SMina Almasry spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 5158550a7d60SMina Almasry 5159550a7d60SMina Almasry pte = huge_ptep_get_and_clear(mm, old_addr, src_pte); 5160550a7d60SMina Almasry set_huge_pte_at(mm, new_addr, dst_pte, pte); 5161550a7d60SMina Almasry 5162550a7d60SMina Almasry if (src_ptl != dst_ptl) 5163550a7d60SMina Almasry spin_unlock(src_ptl); 5164550a7d60SMina Almasry spin_unlock(dst_ptl); 5165550a7d60SMina Almasry } 5166550a7d60SMina Almasry 5167550a7d60SMina Almasry int move_hugetlb_page_tables(struct vm_area_struct *vma, 5168550a7d60SMina Almasry struct vm_area_struct *new_vma, 5169550a7d60SMina Almasry unsigned long old_addr, unsigned long new_addr, 5170550a7d60SMina Almasry unsigned long len) 5171550a7d60SMina Almasry { 5172550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 5173550a7d60SMina Almasry struct address_space *mapping = vma->vm_file->f_mapping; 5174550a7d60SMina Almasry unsigned long sz = huge_page_size(h); 5175550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 5176550a7d60SMina Almasry unsigned long old_end = old_addr + len; 5177e95a9851SMike Kravetz unsigned long last_addr_mask; 5178550a7d60SMina Almasry pte_t *src_pte, *dst_pte; 5179550a7d60SMina Almasry struct mmu_notifier_range range; 51803d0b95cdSBaolin Wang bool shared_pmd = false; 5181550a7d60SMina Almasry 51827d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, old_addr, 5183550a7d60SMina Almasry old_end); 5184550a7d60SMina Almasry adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 51853d0b95cdSBaolin Wang /* 51863d0b95cdSBaolin Wang * In case of shared PMDs, we should cover the maximum possible 51873d0b95cdSBaolin Wang * range. 51883d0b95cdSBaolin Wang */ 51893d0b95cdSBaolin Wang flush_cache_range(vma, range.start, range.end); 51903d0b95cdSBaolin Wang 5191550a7d60SMina Almasry mmu_notifier_invalidate_range_start(&range); 5192e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 5193550a7d60SMina Almasry /* Prevent race with file truncation */ 519440549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 5195550a7d60SMina Almasry i_mmap_lock_write(mapping); 5196550a7d60SMina Almasry for (; old_addr < old_end; old_addr += sz, new_addr += sz) { 51979c67a207SPeter Xu src_pte = hugetlb_walk(vma, old_addr, sz); 5198e95a9851SMike Kravetz if (!src_pte) { 5199e95a9851SMike Kravetz old_addr |= last_addr_mask; 5200e95a9851SMike Kravetz new_addr |= last_addr_mask; 5201550a7d60SMina Almasry continue; 5202e95a9851SMike Kravetz } 5203550a7d60SMina Almasry if (huge_pte_none(huge_ptep_get(src_pte))) 5204550a7d60SMina Almasry continue; 5205550a7d60SMina Almasry 52064ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) { 52073d0b95cdSBaolin Wang shared_pmd = true; 52084ddb4d91SMike Kravetz old_addr |= last_addr_mask; 52094ddb4d91SMike Kravetz new_addr |= last_addr_mask; 5210550a7d60SMina Almasry continue; 52113d0b95cdSBaolin Wang } 5212550a7d60SMina Almasry 5213550a7d60SMina Almasry dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz); 5214550a7d60SMina Almasry if (!dst_pte) 5215550a7d60SMina Almasry break; 5216550a7d60SMina Almasry 5217db110a99SAneesh Kumar K.V move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte); 5218550a7d60SMina Almasry } 52193d0b95cdSBaolin Wang 52203d0b95cdSBaolin Wang if (shared_pmd) 52213d0b95cdSBaolin Wang flush_tlb_range(vma, range.start, range.end); 52223d0b95cdSBaolin Wang else 5223550a7d60SMina Almasry flush_tlb_range(vma, old_end - len, old_end); 5224550a7d60SMina Almasry mmu_notifier_invalidate_range_end(&range); 522513e4ad2cSNadav Amit i_mmap_unlock_write(mapping); 522640549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 5227550a7d60SMina Almasry 5228550a7d60SMina Almasry return len + old_addr - old_end; 5229550a7d60SMina Almasry } 5230550a7d60SMina Almasry 523173c54763SPeter Xu static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 523224669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 523305e90bd0SPeter Xu struct page *ref_page, zap_flags_t zap_flags) 523463551ae0SDavid Gibson { 523563551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 523663551ae0SDavid Gibson unsigned long address; 5237c7546f8fSDavid Gibson pte_t *ptep; 523863551ae0SDavid Gibson pte_t pte; 5239cb900f41SKirill A. Shutemov spinlock_t *ptl; 524063551ae0SDavid Gibson struct page *page; 5241a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 5242a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 5243e95a9851SMike Kravetz unsigned long last_addr_mask; 5244a4a118f2SNadav Amit bool force_flush = false; 5245a5516438SAndi Kleen 524663551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 5247a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 5248a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 524963551ae0SDavid Gibson 525007e32661SAneesh Kumar K.V /* 525107e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 525207e32661SAneesh Kumar K.V * to huge page. 525307e32661SAneesh Kumar K.V */ 5254ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 525524669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 5256dff11abeSMike Kravetz 5257e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 5258569f48b8SHillf Danton address = start; 5259569f48b8SHillf Danton for (; address < end; address += sz) { 52609c67a207SPeter Xu ptep = hugetlb_walk(vma, address, sz); 5261e95a9851SMike Kravetz if (!ptep) { 5262e95a9851SMike Kravetz address |= last_addr_mask; 5263c7546f8fSDavid Gibson continue; 5264e95a9851SMike Kravetz } 5265c7546f8fSDavid Gibson 5266cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 52674ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, address, ptep)) { 526831d49da5SAneesh Kumar K.V spin_unlock(ptl); 5269a4a118f2SNadav Amit tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE); 5270a4a118f2SNadav Amit force_flush = true; 52714ddb4d91SMike Kravetz address |= last_addr_mask; 527231d49da5SAneesh Kumar K.V continue; 527331d49da5SAneesh Kumar K.V } 527439dde65cSChen, Kenneth W 52756629326bSHillf Danton pte = huge_ptep_get(ptep); 527631d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 527731d49da5SAneesh Kumar K.V spin_unlock(ptl); 527831d49da5SAneesh Kumar K.V continue; 527931d49da5SAneesh Kumar K.V } 52806629326bSHillf Danton 52816629326bSHillf Danton /* 52829fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 52839fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 52846629326bSHillf Danton */ 52859fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 528605e90bd0SPeter Xu /* 528705e90bd0SPeter Xu * If the pte was wr-protected by uffd-wp in any of the 528805e90bd0SPeter Xu * swap forms, meanwhile the caller does not want to 528905e90bd0SPeter Xu * drop the uffd-wp bit in this zap, then replace the 529005e90bd0SPeter Xu * pte with a marker. 529105e90bd0SPeter Xu */ 529205e90bd0SPeter Xu if (pte_swp_uffd_wp_any(pte) && 529305e90bd0SPeter Xu !(zap_flags & ZAP_FLAG_DROP_MARKER)) 529405e90bd0SPeter Xu set_huge_pte_at(mm, address, ptep, 529505e90bd0SPeter Xu make_pte_marker(PTE_MARKER_UFFD_WP)); 529605e90bd0SPeter Xu else 52979386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 529831d49da5SAneesh Kumar K.V spin_unlock(ptl); 529931d49da5SAneesh Kumar K.V continue; 53008c4894c6SNaoya Horiguchi } 53016629326bSHillf Danton 53026629326bSHillf Danton page = pte_page(pte); 530304f2cbe3SMel Gorman /* 530404f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 530504f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 530604f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 530704f2cbe3SMel Gorman */ 530804f2cbe3SMel Gorman if (ref_page) { 530931d49da5SAneesh Kumar K.V if (page != ref_page) { 531031d49da5SAneesh Kumar K.V spin_unlock(ptl); 531131d49da5SAneesh Kumar K.V continue; 531231d49da5SAneesh Kumar K.V } 531304f2cbe3SMel Gorman /* 531404f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 531504f2cbe3SMel Gorman * future faults in this VMA will fail rather than 531604f2cbe3SMel Gorman * looking like data was lost 531704f2cbe3SMel Gorman */ 531804f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 531904f2cbe3SMel Gorman } 532004f2cbe3SMel Gorman 5321c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 5322b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 5323106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 53246649a386SKen Chen set_page_dirty(page); 532505e90bd0SPeter Xu /* Leave a uffd-wp pte marker if needed */ 532605e90bd0SPeter Xu if (huge_pte_uffd_wp(pte) && 532705e90bd0SPeter Xu !(zap_flags & ZAP_FLAG_DROP_MARKER)) 532805e90bd0SPeter Xu set_huge_pte_at(mm, address, ptep, 532905e90bd0SPeter Xu make_pte_marker(PTE_MARKER_UFFD_WP)); 53305d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 5331cea86fe2SHugh Dickins page_remove_rmap(page, vma, true); 533231d49da5SAneesh Kumar K.V 5333cb900f41SKirill A. Shutemov spin_unlock(ptl); 5334e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 533524669e58SAneesh Kumar K.V /* 533631d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 533724669e58SAneesh Kumar K.V */ 533831d49da5SAneesh Kumar K.V if (ref_page) 533931d49da5SAneesh Kumar K.V break; 5340fe1668aeSChen, Kenneth W } 534124669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 5342a4a118f2SNadav Amit 5343a4a118f2SNadav Amit /* 5344a4a118f2SNadav Amit * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We 5345a4a118f2SNadav Amit * could defer the flush until now, since by holding i_mmap_rwsem we 5346a4a118f2SNadav Amit * guaranteed that the last refernece would not be dropped. But we must 5347a4a118f2SNadav Amit * do the flushing before we return, as otherwise i_mmap_rwsem will be 5348a4a118f2SNadav Amit * dropped and the last reference to the shared PMDs page might be 5349a4a118f2SNadav Amit * dropped as well. 5350a4a118f2SNadav Amit * 5351a4a118f2SNadav Amit * In theory we could defer the freeing of the PMD pages as well, but 5352a4a118f2SNadav Amit * huge_pmd_unshare() relies on the exact page_count for the PMD page to 5353a4a118f2SNadav Amit * detect sharing, so we cannot defer the release of the page either. 5354a4a118f2SNadav Amit * Instead, do flush now. 5355a4a118f2SNadav Amit */ 5356a4a118f2SNadav Amit if (force_flush) 5357a4a118f2SNadav Amit tlb_flush_mmu_tlbonly(tlb); 53581da177e4SLinus Torvalds } 535963551ae0SDavid Gibson 5360d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 5361d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 536205e90bd0SPeter Xu unsigned long end, struct page *ref_page, 536305e90bd0SPeter Xu zap_flags_t zap_flags) 5364d833352aSMel Gorman { 5365131a79b4SMike Kravetz hugetlb_vma_lock_write(vma); 5366131a79b4SMike Kravetz i_mmap_lock_write(vma->vm_file->f_mapping); 5367131a79b4SMike Kravetz 5368369258ceSMike Kravetz /* mmu notification performed in caller */ 536905e90bd0SPeter Xu __unmap_hugepage_range(tlb, vma, start, end, ref_page, zap_flags); 5370d833352aSMel Gorman 537104ada095SMike Kravetz if (zap_flags & ZAP_FLAG_UNMAP) { /* final unmap */ 5372d833352aSMel Gorman /* 537304ada095SMike Kravetz * Unlock and free the vma lock before releasing i_mmap_rwsem. 537404ada095SMike Kravetz * When the vma_lock is freed, this makes the vma ineligible 537504ada095SMike Kravetz * for pmd sharing. And, i_mmap_rwsem is required to set up 537604ada095SMike Kravetz * pmd sharing. This is important as page tables for this 537704ada095SMike Kravetz * unmapped range will be asynchrously deleted. If the page 537804ada095SMike Kravetz * tables are shared, there will be issues when accessed by 537904ada095SMike Kravetz * someone else. 5380d833352aSMel Gorman */ 5381ecfbd733SMike Kravetz __hugetlb_vma_unlock_write_free(vma); 5382131a79b4SMike Kravetz i_mmap_unlock_write(vma->vm_file->f_mapping); 538304ada095SMike Kravetz } else { 538404ada095SMike Kravetz i_mmap_unlock_write(vma->vm_file->f_mapping); 538504ada095SMike Kravetz hugetlb_vma_unlock_write(vma); 538604ada095SMike Kravetz } 5387d833352aSMel Gorman } 5388d833352aSMel Gorman 5389502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 539005e90bd0SPeter Xu unsigned long end, struct page *ref_page, 539105e90bd0SPeter Xu zap_flags_t zap_flags) 5392502717f4SChen, Kenneth W { 5393369258ceSMike Kravetz struct mmu_notifier_range range; 539424669e58SAneesh Kumar K.V struct mmu_gather tlb; 5395dff11abeSMike Kravetz 53967d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm, 5397369258ceSMike Kravetz start, end); 5398369258ceSMike Kravetz adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 5399369258ceSMike Kravetz mmu_notifier_invalidate_range_start(&range); 5400a72afd87SWill Deacon tlb_gather_mmu(&tlb, vma->vm_mm); 5401369258ceSMike Kravetz 540205e90bd0SPeter Xu __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags); 5403369258ceSMike Kravetz 5404369258ceSMike Kravetz mmu_notifier_invalidate_range_end(&range); 5405ae8eba8bSWill Deacon tlb_finish_mmu(&tlb); 5406502717f4SChen, Kenneth W } 5407502717f4SChen, Kenneth W 540804f2cbe3SMel Gorman /* 540904f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 5410578b7725SZhiyuan Dai * mapping it owns the reserve page for. The intention is to unmap the page 541104f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 541204f2cbe3SMel Gorman * same region. 541304f2cbe3SMel Gorman */ 54142f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 54152a4b3dedSHarvey Harrison struct page *page, unsigned long address) 541604f2cbe3SMel Gorman { 54177526674dSAdam Litke struct hstate *h = hstate_vma(vma); 541804f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 541904f2cbe3SMel Gorman struct address_space *mapping; 542004f2cbe3SMel Gorman pgoff_t pgoff; 542104f2cbe3SMel Gorman 542204f2cbe3SMel Gorman /* 542304f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 542404f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 542504f2cbe3SMel Gorman */ 54267526674dSAdam Litke address = address & huge_page_mask(h); 542736e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 542836e4f20aSMichal Hocko vma->vm_pgoff; 542993c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 543004f2cbe3SMel Gorman 54314eb2b1dcSMel Gorman /* 54324eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 54334eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 54344eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 54354eb2b1dcSMel Gorman */ 543683cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 54376b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 543804f2cbe3SMel Gorman /* Do not unmap the current VMA */ 543904f2cbe3SMel Gorman if (iter_vma == vma) 544004f2cbe3SMel Gorman continue; 544104f2cbe3SMel Gorman 544204f2cbe3SMel Gorman /* 54432f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 54442f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 54452f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 54462f84a899SMel Gorman */ 54472f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 54482f84a899SMel Gorman continue; 54492f84a899SMel Gorman 54502f84a899SMel Gorman /* 545104f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 545204f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 545304f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 545404f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 545504f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 545604f2cbe3SMel Gorman */ 545704f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 545824669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 545905e90bd0SPeter Xu address + huge_page_size(h), page, 0); 546004f2cbe3SMel Gorman } 546183cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 546204f2cbe3SMel Gorman } 546304f2cbe3SMel Gorman 54640fe6e20bSNaoya Horiguchi /* 5465c89357e2SDavid Hildenbrand * hugetlb_wp() should be called with page lock of the original hugepage held. 5466aa6d2e8cSBaolin Wang * Called with hugetlb_fault_mutex_table held and pte_page locked so we 5467ef009b25SMichal Hocko * cannot race with other handlers or page migration. 5468ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 54690fe6e20bSNaoya Horiguchi */ 5470c89357e2SDavid Hildenbrand static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma, 5471c89357e2SDavid Hildenbrand unsigned long address, pte_t *ptep, unsigned int flags, 5472cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 54731e8f889bSDavid Gibson { 5474c89357e2SDavid Hildenbrand const bool unshare = flags & FAULT_FLAG_UNSHARE; 54753999f52eSAneesh Kumar K.V pte_t pte; 5476a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 54771e8f889bSDavid Gibson struct page *old_page, *new_page; 54782b740303SSouptick Joarder int outside_reserve = 0; 54792b740303SSouptick Joarder vm_fault_t ret = 0; 5480974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 5481ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 54821e8f889bSDavid Gibson 54831d8d1464SDavid Hildenbrand /* 54841d8d1464SDavid Hildenbrand * hugetlb does not support FOLL_FORCE-style write faults that keep the 54851d8d1464SDavid Hildenbrand * PTE mapped R/O such as maybe_mkwrite() would do. 54861d8d1464SDavid Hildenbrand */ 54871d8d1464SDavid Hildenbrand if (WARN_ON_ONCE(!unshare && !(vma->vm_flags & VM_WRITE))) 54881d8d1464SDavid Hildenbrand return VM_FAULT_SIGSEGV; 54891d8d1464SDavid Hildenbrand 54901d8d1464SDavid Hildenbrand /* Let's take out MAP_SHARED mappings first. */ 54911d8d1464SDavid Hildenbrand if (vma->vm_flags & VM_MAYSHARE) { 54921d8d1464SDavid Hildenbrand set_huge_ptep_writable(vma, haddr, ptep); 54931d8d1464SDavid Hildenbrand return 0; 54941d8d1464SDavid Hildenbrand } 54951d8d1464SDavid Hildenbrand 54963999f52eSAneesh Kumar K.V pte = huge_ptep_get(ptep); 54971e8f889bSDavid Gibson old_page = pte_page(pte); 54981e8f889bSDavid Gibson 5499662ce1dcSYang Yang delayacct_wpcopy_start(); 5500662ce1dcSYang Yang 550104f2cbe3SMel Gorman retry_avoidcopy: 5502c89357e2SDavid Hildenbrand /* 5503c89357e2SDavid Hildenbrand * If no-one else is actually using this page, we're the exclusive 5504c89357e2SDavid Hildenbrand * owner and can reuse this page. 5505c89357e2SDavid Hildenbrand */ 550637a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 5507c89357e2SDavid Hildenbrand if (!PageAnonExclusive(old_page)) 55085a49973dSHugh Dickins page_move_anon_rmap(old_page, vma); 5509c89357e2SDavid Hildenbrand if (likely(!unshare)) 55105b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 5511662ce1dcSYang Yang 5512662ce1dcSYang Yang delayacct_wpcopy_end(); 551383c54070SNick Piggin return 0; 55141e8f889bSDavid Gibson } 55156c287605SDavid Hildenbrand VM_BUG_ON_PAGE(PageAnon(old_page) && PageAnonExclusive(old_page), 55166c287605SDavid Hildenbrand old_page); 55171e8f889bSDavid Gibson 551804f2cbe3SMel Gorman /* 551904f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 552004f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 552104f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 552204f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 552304f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 552404f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 552504f2cbe3SMel Gorman * of the full address range. 552604f2cbe3SMel Gorman */ 55275944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 552804f2cbe3SMel Gorman old_page != pagecache_page) 552904f2cbe3SMel Gorman outside_reserve = 1; 553004f2cbe3SMel Gorman 553109cbfeafSKirill A. Shutemov get_page(old_page); 5532b76c8cfbSLarry Woodman 5533ad4404a2SDavidlohr Bueso /* 5534ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 5535ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 5536ad4404a2SDavidlohr Bueso */ 5537cb900f41SKirill A. Shutemov spin_unlock(ptl); 55385b7a1d40SHuang Ying new_page = alloc_huge_page(vma, haddr, outside_reserve); 55391e8f889bSDavid Gibson 55402fc39cecSAdam Litke if (IS_ERR(new_page)) { 554104f2cbe3SMel Gorman /* 554204f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 554304f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 554404f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 554504f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 554604f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 554704f2cbe3SMel Gorman */ 554804f2cbe3SMel Gorman if (outside_reserve) { 554940549ba8SMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 555040549ba8SMike Kravetz pgoff_t idx; 555140549ba8SMike Kravetz u32 hash; 555240549ba8SMike Kravetz 555309cbfeafSKirill A. Shutemov put_page(old_page); 555440549ba8SMike Kravetz /* 555540549ba8SMike Kravetz * Drop hugetlb_fault_mutex and vma_lock before 555640549ba8SMike Kravetz * unmapping. unmapping needs to hold vma_lock 555740549ba8SMike Kravetz * in write mode. Dropping vma_lock in read mode 555840549ba8SMike Kravetz * here is OK as COW mappings do not interact with 555940549ba8SMike Kravetz * PMD sharing. 556040549ba8SMike Kravetz * 556140549ba8SMike Kravetz * Reacquire both after unmap operation. 556240549ba8SMike Kravetz */ 556340549ba8SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 556440549ba8SMike Kravetz hash = hugetlb_fault_mutex_hash(mapping, idx); 556540549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 556640549ba8SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 556740549ba8SMike Kravetz 55685b7a1d40SHuang Ying unmap_ref_private(mm, vma, old_page, haddr); 556940549ba8SMike Kravetz 557040549ba8SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 557140549ba8SMike Kravetz hugetlb_vma_lock_read(vma); 5572cb900f41SKirill A. Shutemov spin_lock(ptl); 55739c67a207SPeter Xu ptep = hugetlb_walk(vma, haddr, huge_page_size(h)); 5574a9af0c5dSNaoya Horiguchi if (likely(ptep && 5575a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 557604f2cbe3SMel Gorman goto retry_avoidcopy; 5577a734bcc8SHillf Danton /* 5578cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 5579cb900f41SKirill A. Shutemov * lock, and our job is done. 5580a734bcc8SHillf Danton */ 5581662ce1dcSYang Yang delayacct_wpcopy_end(); 5582a734bcc8SHillf Danton return 0; 558304f2cbe3SMel Gorman } 558404f2cbe3SMel Gorman 55852b740303SSouptick Joarder ret = vmf_error(PTR_ERR(new_page)); 5586ad4404a2SDavidlohr Bueso goto out_release_old; 55871e8f889bSDavid Gibson } 55881e8f889bSDavid Gibson 55890fe6e20bSNaoya Horiguchi /* 55900fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 55910fe6e20bSNaoya Horiguchi * anon_vma prepared. 55920fe6e20bSNaoya Horiguchi */ 559344e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 5594ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 5595ad4404a2SDavidlohr Bueso goto out_release_all; 559644e2aa93SDean Nelson } 55970fe6e20bSNaoya Horiguchi 5598974e6d66SHuang Ying copy_user_huge_page(new_page, old_page, address, vma, 559947ad8475SAndrea Arcangeli pages_per_huge_page(h)); 56000ed361deSNick Piggin __SetPageUptodate(new_page); 56011e8f889bSDavid Gibson 56027d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, haddr, 56036f4f13e8SJérôme Glisse haddr + huge_page_size(h)); 5604ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 5605ad4404a2SDavidlohr Bueso 5606b76c8cfbSLarry Woodman /* 5607cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 5608b76c8cfbSLarry Woodman * before the page tables are altered 5609b76c8cfbSLarry Woodman */ 5610cb900f41SKirill A. Shutemov spin_lock(ptl); 56119c67a207SPeter Xu ptep = hugetlb_walk(vma, haddr, huge_page_size(h)); 5612a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 5613c89357e2SDavid Hildenbrand /* Break COW or unshare */ 56145b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 5615ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, range.end); 5616cea86fe2SHugh Dickins page_remove_rmap(old_page, vma, true); 56175b7a1d40SHuang Ying hugepage_add_new_anon_rmap(new_page, vma, haddr); 56181eba86c0SPasha Tatashin set_huge_pte_at(mm, haddr, ptep, 5619c89357e2SDavid Hildenbrand make_huge_pte(vma, new_page, !unshare)); 56208f251a3dSMike Kravetz SetHPageMigratable(new_page); 56211e8f889bSDavid Gibson /* Make the old page be freed below */ 56221e8f889bSDavid Gibson new_page = old_page; 56231e8f889bSDavid Gibson } 5624cb900f41SKirill A. Shutemov spin_unlock(ptl); 5625ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 5626ad4404a2SDavidlohr Bueso out_release_all: 5627c89357e2SDavid Hildenbrand /* 5628c89357e2SDavid Hildenbrand * No restore in case of successful pagetable update (Break COW or 5629c89357e2SDavid Hildenbrand * unshare) 5630c89357e2SDavid Hildenbrand */ 5631c7b1850dSMike Kravetz if (new_page != old_page) 56325b7a1d40SHuang Ying restore_reserve_on_error(h, vma, haddr, new_page); 563309cbfeafSKirill A. Shutemov put_page(new_page); 5634ad4404a2SDavidlohr Bueso out_release_old: 563509cbfeafSKirill A. Shutemov put_page(old_page); 56368312034fSJoonsoo Kim 5637ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 5638662ce1dcSYang Yang 5639662ce1dcSYang Yang delayacct_wpcopy_end(); 5640ad4404a2SDavidlohr Bueso return ret; 56411e8f889bSDavid Gibson } 56421e8f889bSDavid Gibson 56433ae77f43SHugh Dickins /* 56443ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 56453ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 56463ae77f43SHugh Dickins */ 56473ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 56482a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 56492a15efc9SHugh Dickins { 56502a15efc9SHugh Dickins struct address_space *mapping; 56512a15efc9SHugh Dickins pgoff_t idx; 56522a15efc9SHugh Dickins struct page *page; 56532a15efc9SHugh Dickins 56542a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 56552a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 56562a15efc9SHugh Dickins 56572a15efc9SHugh Dickins page = find_get_page(mapping, idx); 56582a15efc9SHugh Dickins if (page) 56592a15efc9SHugh Dickins put_page(page); 56602a15efc9SHugh Dickins return page != NULL; 56612a15efc9SHugh Dickins } 56622a15efc9SHugh Dickins 56637e1813d4SMike Kravetz int hugetlb_add_to_page_cache(struct page *page, struct address_space *mapping, 5664ab76ad54SMike Kravetz pgoff_t idx) 5665ab76ad54SMike Kravetz { 5666d9ef44deSMatthew Wilcox (Oracle) struct folio *folio = page_folio(page); 5667ab76ad54SMike Kravetz struct inode *inode = mapping->host; 5668ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 5669d9ef44deSMatthew Wilcox (Oracle) int err; 5670ab76ad54SMike Kravetz 5671d9ef44deSMatthew Wilcox (Oracle) __folio_set_locked(folio); 5672d9ef44deSMatthew Wilcox (Oracle) err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL); 5673d9ef44deSMatthew Wilcox (Oracle) 5674d9ef44deSMatthew Wilcox (Oracle) if (unlikely(err)) { 5675d9ef44deSMatthew Wilcox (Oracle) __folio_clear_locked(folio); 5676ab76ad54SMike Kravetz return err; 5677d9ef44deSMatthew Wilcox (Oracle) } 5678d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 5679ab76ad54SMike Kravetz 568022146c3cSMike Kravetz /* 5681d9ef44deSMatthew Wilcox (Oracle) * mark folio dirty so that it will not be removed from cache/file 568222146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 568322146c3cSMike Kravetz */ 5684d9ef44deSMatthew Wilcox (Oracle) folio_mark_dirty(folio); 568522146c3cSMike Kravetz 5686ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 5687ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 5688ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 5689ab76ad54SMike Kravetz return 0; 5690ab76ad54SMike Kravetz } 5691ab76ad54SMike Kravetz 56927677f7fdSAxel Rasmussen static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma, 56937677f7fdSAxel Rasmussen struct address_space *mapping, 56947677f7fdSAxel Rasmussen pgoff_t idx, 56957677f7fdSAxel Rasmussen unsigned int flags, 56967677f7fdSAxel Rasmussen unsigned long haddr, 5697824ddc60SNadav Amit unsigned long addr, 56987677f7fdSAxel Rasmussen unsigned long reason) 56997677f7fdSAxel Rasmussen { 57007677f7fdSAxel Rasmussen u32 hash; 57017677f7fdSAxel Rasmussen struct vm_fault vmf = { 57027677f7fdSAxel Rasmussen .vma = vma, 57037677f7fdSAxel Rasmussen .address = haddr, 5704824ddc60SNadav Amit .real_address = addr, 57057677f7fdSAxel Rasmussen .flags = flags, 57067677f7fdSAxel Rasmussen 57077677f7fdSAxel Rasmussen /* 57087677f7fdSAxel Rasmussen * Hard to debug if it ends up being 57097677f7fdSAxel Rasmussen * used by a callee that assumes 57107677f7fdSAxel Rasmussen * something about the other 57117677f7fdSAxel Rasmussen * uninitialized fields... same as in 57127677f7fdSAxel Rasmussen * memory.c 57137677f7fdSAxel Rasmussen */ 57147677f7fdSAxel Rasmussen }; 57157677f7fdSAxel Rasmussen 57167677f7fdSAxel Rasmussen /* 5717958f32ceSLiu Shixin * vma_lock and hugetlb_fault_mutex must be dropped before handling 5718958f32ceSLiu Shixin * userfault. Also mmap_lock could be dropped due to handling 5719958f32ceSLiu Shixin * userfault, any vma operation should be careful from here. 57207677f7fdSAxel Rasmussen */ 572140549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 57227677f7fdSAxel Rasmussen hash = hugetlb_fault_mutex_hash(mapping, idx); 57237677f7fdSAxel Rasmussen mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5724958f32ceSLiu Shixin return handle_userfault(&vmf, reason); 57257677f7fdSAxel Rasmussen } 57267677f7fdSAxel Rasmussen 57272ea7ff1eSPeter Xu /* 57282ea7ff1eSPeter Xu * Recheck pte with pgtable lock. Returns true if pte didn't change, or 57292ea7ff1eSPeter Xu * false if pte changed or is changing. 57302ea7ff1eSPeter Xu */ 57312ea7ff1eSPeter Xu static bool hugetlb_pte_stable(struct hstate *h, struct mm_struct *mm, 57322ea7ff1eSPeter Xu pte_t *ptep, pte_t old_pte) 57332ea7ff1eSPeter Xu { 57342ea7ff1eSPeter Xu spinlock_t *ptl; 57352ea7ff1eSPeter Xu bool same; 57362ea7ff1eSPeter Xu 57372ea7ff1eSPeter Xu ptl = huge_pte_lock(h, mm, ptep); 57382ea7ff1eSPeter Xu same = pte_same(huge_ptep_get(ptep), old_pte); 57392ea7ff1eSPeter Xu spin_unlock(ptl); 57402ea7ff1eSPeter Xu 57412ea7ff1eSPeter Xu return same; 57422ea7ff1eSPeter Xu } 57432ea7ff1eSPeter Xu 57442b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 57452b740303SSouptick Joarder struct vm_area_struct *vma, 57468382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 5747c64e912cSPeter Xu unsigned long address, pte_t *ptep, 5748c64e912cSPeter Xu pte_t old_pte, unsigned int flags) 5749ac9b9c66SHugh Dickins { 5750a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 57512b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 5752409eb8c2SHillf Danton int anon_rmap = 0; 57534c887265SAdam Litke unsigned long size; 57544c887265SAdam Litke struct page *page; 57551e8f889bSDavid Gibson pte_t new_pte; 5756cb900f41SKirill A. Shutemov spinlock_t *ptl; 5757285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 5758c7b1850dSMike Kravetz bool new_page, new_pagecache_page = false; 5759958f32ceSLiu Shixin u32 hash = hugetlb_fault_mutex_hash(mapping, idx); 57604c887265SAdam Litke 576104f2cbe3SMel Gorman /* 576204f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 576304f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 5764c89357e2SDavid Hildenbrand * COW/unsharing. Warn that such a situation has occurred as it may not 5765c89357e2SDavid Hildenbrand * be obvious. 576604f2cbe3SMel Gorman */ 576704f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 5768910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 576904f2cbe3SMel Gorman current->pid); 5770958f32ceSLiu Shixin goto out; 577104f2cbe3SMel Gorman } 577204f2cbe3SMel Gorman 57734c887265SAdam Litke /* 5774188a3972SMike Kravetz * Use page lock to guard against racing truncation 5775188a3972SMike Kravetz * before we get page_table_lock. 57764c887265SAdam Litke */ 5777c7b1850dSMike Kravetz new_page = false; 577887bf91d3SMike Kravetz page = find_lock_page(mapping, idx); 577987bf91d3SMike Kravetz if (!page) { 5780188a3972SMike Kravetz size = i_size_read(mapping->host) >> huge_page_shift(h); 5781188a3972SMike Kravetz if (idx >= size) 5782188a3972SMike Kravetz goto out; 57837677f7fdSAxel Rasmussen /* Check for page in userfault range */ 57842ea7ff1eSPeter Xu if (userfaultfd_missing(vma)) { 57852ea7ff1eSPeter Xu /* 57862ea7ff1eSPeter Xu * Since hugetlb_no_page() was examining pte 57872ea7ff1eSPeter Xu * without pgtable lock, we need to re-test under 57882ea7ff1eSPeter Xu * lock because the pte may not be stable and could 57892ea7ff1eSPeter Xu * have changed from under us. Try to detect 57902ea7ff1eSPeter Xu * either changed or during-changing ptes and retry 57912ea7ff1eSPeter Xu * properly when needed. 57922ea7ff1eSPeter Xu * 57932ea7ff1eSPeter Xu * Note that userfaultfd is actually fine with 57942ea7ff1eSPeter Xu * false positives (e.g. caused by pte changed), 57952ea7ff1eSPeter Xu * but not wrong logical events (e.g. caused by 57962ea7ff1eSPeter Xu * reading a pte during changing). The latter can 57972ea7ff1eSPeter Xu * confuse the userspace, so the strictness is very 57982ea7ff1eSPeter Xu * much preferred. E.g., MISSING event should 57992ea7ff1eSPeter Xu * never happen on the page after UFFDIO_COPY has 58002ea7ff1eSPeter Xu * correctly installed the page and returned. 58012ea7ff1eSPeter Xu */ 58022ea7ff1eSPeter Xu if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) { 58032ea7ff1eSPeter Xu ret = 0; 58042ea7ff1eSPeter Xu goto out; 58052ea7ff1eSPeter Xu } 58062ea7ff1eSPeter Xu 58072ea7ff1eSPeter Xu return hugetlb_handle_userfault(vma, mapping, idx, flags, 58082ea7ff1eSPeter Xu haddr, address, 58097677f7fdSAxel Rasmussen VM_UFFD_MISSING); 58102ea7ff1eSPeter Xu } 58111a1aad8aSMike Kravetz 5812285b8dcaSHuang Ying page = alloc_huge_page(vma, haddr, 0); 58132fc39cecSAdam Litke if (IS_ERR(page)) { 58144643d67eSMike Kravetz /* 58154643d67eSMike Kravetz * Returning error will result in faulting task being 58164643d67eSMike Kravetz * sent SIGBUS. The hugetlb fault mutex prevents two 58174643d67eSMike Kravetz * tasks from racing to fault in the same page which 58184643d67eSMike Kravetz * could result in false unable to allocate errors. 58194643d67eSMike Kravetz * Page migration does not take the fault mutex, but 58204643d67eSMike Kravetz * does a clear then write of pte's under page table 58214643d67eSMike Kravetz * lock. Page fault code could race with migration, 58224643d67eSMike Kravetz * notice the clear pte and try to allocate a page 58234643d67eSMike Kravetz * here. Before returning error, get ptl and make 58244643d67eSMike Kravetz * sure there really is no pte entry. 58254643d67eSMike Kravetz */ 5826f9bf6c03SPeter Xu if (hugetlb_pte_stable(h, mm, ptep, old_pte)) 58272b740303SSouptick Joarder ret = vmf_error(PTR_ERR(page)); 5828f9bf6c03SPeter Xu else 5829f9bf6c03SPeter Xu ret = 0; 58306bda666aSChristoph Lameter goto out; 58316bda666aSChristoph Lameter } 583247ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 58330ed361deSNick Piggin __SetPageUptodate(page); 5834cb6acd01SMike Kravetz new_page = true; 5835ac9b9c66SHugh Dickins 5836f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 58377e1813d4SMike Kravetz int err = hugetlb_add_to_page_cache(page, mapping, idx); 58386bda666aSChristoph Lameter if (err) { 58393a5497a2SMiaohe Lin /* 58403a5497a2SMiaohe Lin * err can't be -EEXIST which implies someone 58413a5497a2SMiaohe Lin * else consumed the reservation since hugetlb 58423a5497a2SMiaohe Lin * fault mutex is held when add a hugetlb page 58433a5497a2SMiaohe Lin * to the page cache. So it's safe to call 58443a5497a2SMiaohe Lin * restore_reserve_on_error() here. 58453a5497a2SMiaohe Lin */ 58463a5497a2SMiaohe Lin restore_reserve_on_error(h, vma, haddr, page); 58476bda666aSChristoph Lameter put_page(page); 58486bda666aSChristoph Lameter goto out; 58496bda666aSChristoph Lameter } 5850c7b1850dSMike Kravetz new_pagecache_page = true; 585123be7468SMel Gorman } else { 58526bda666aSChristoph Lameter lock_page(page); 58530fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 58540fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 58550fe6e20bSNaoya Horiguchi goto backout_unlocked; 585623be7468SMel Gorman } 5857409eb8c2SHillf Danton anon_rmap = 1; 58580fe6e20bSNaoya Horiguchi } 58590fe6e20bSNaoya Horiguchi } else { 586057303d80SAndy Whitcroft /* 5861998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 5862998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 5863998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 5864fd6a03edSNaoya Horiguchi */ 5865fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 58660eb98f15SMiaohe Lin ret = VM_FAULT_HWPOISON_LARGE | 5867972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 5868fd6a03edSNaoya Horiguchi goto backout_unlocked; 58696bda666aSChristoph Lameter } 58707677f7fdSAxel Rasmussen 58717677f7fdSAxel Rasmussen /* Check for page in userfault range. */ 58727677f7fdSAxel Rasmussen if (userfaultfd_minor(vma)) { 58737677f7fdSAxel Rasmussen unlock_page(page); 58747677f7fdSAxel Rasmussen put_page(page); 58752ea7ff1eSPeter Xu /* See comment in userfaultfd_missing() block above */ 58762ea7ff1eSPeter Xu if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) { 58772ea7ff1eSPeter Xu ret = 0; 58782ea7ff1eSPeter Xu goto out; 58792ea7ff1eSPeter Xu } 58802ea7ff1eSPeter Xu return hugetlb_handle_userfault(vma, mapping, idx, flags, 58812ea7ff1eSPeter Xu haddr, address, 58827677f7fdSAxel Rasmussen VM_UFFD_MINOR); 58837677f7fdSAxel Rasmussen } 5884998b4382SNaoya Horiguchi } 58851e8f889bSDavid Gibson 588657303d80SAndy Whitcroft /* 588757303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 588857303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 588957303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 589057303d80SAndy Whitcroft * the spinlock. 589157303d80SAndy Whitcroft */ 58925e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 5893285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 58942b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 58952b26736cSAndy Whitcroft goto backout_unlocked; 58962b26736cSAndy Whitcroft } 58975e911373SMike Kravetz /* Just decrements count, does not deallocate */ 5898285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 58995e911373SMike Kravetz } 590057303d80SAndy Whitcroft 59018bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 590283c54070SNick Piggin ret = 0; 5903c64e912cSPeter Xu /* If pte changed from under us, retry */ 5904c64e912cSPeter Xu if (!pte_same(huge_ptep_get(ptep), old_pte)) 59054c887265SAdam Litke goto backout; 59064c887265SAdam Litke 59074781593dSPeter Xu if (anon_rmap) 5908285b8dcaSHuang Ying hugepage_add_new_anon_rmap(page, vma, haddr); 59094781593dSPeter Xu else 5910fb3d824dSDavid Hildenbrand page_dup_file_rmap(page, true); 59111e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 59121e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 5913c64e912cSPeter Xu /* 5914c64e912cSPeter Xu * If this pte was previously wr-protected, keep it wr-protected even 5915c64e912cSPeter Xu * if populated. 5916c64e912cSPeter Xu */ 5917c64e912cSPeter Xu if (unlikely(pte_marker_uffd_wp(old_pte))) 5918f1eb1bacSPeter Xu new_pte = huge_pte_mkuffd_wp(new_pte); 5919285b8dcaSHuang Ying set_huge_pte_at(mm, haddr, ptep, new_pte); 59201e8f889bSDavid Gibson 59215d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 5922788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 59231e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 5924c89357e2SDavid Hildenbrand ret = hugetlb_wp(mm, vma, address, ptep, flags, page, ptl); 59251e8f889bSDavid Gibson } 59261e8f889bSDavid Gibson 5927cb900f41SKirill A. Shutemov spin_unlock(ptl); 5928cb6acd01SMike Kravetz 5929cb6acd01SMike Kravetz /* 59308f251a3dSMike Kravetz * Only set HPageMigratable in newly allocated pages. Existing pages 59318f251a3dSMike Kravetz * found in the pagecache may not have HPageMigratableset if they have 59328f251a3dSMike Kravetz * been isolated for migration. 5933cb6acd01SMike Kravetz */ 5934cb6acd01SMike Kravetz if (new_page) 59358f251a3dSMike Kravetz SetHPageMigratable(page); 5936cb6acd01SMike Kravetz 59374c887265SAdam Litke unlock_page(page); 59384c887265SAdam Litke out: 5939958f32ceSLiu Shixin hugetlb_vma_unlock_read(vma); 5940958f32ceSLiu Shixin mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5941ac9b9c66SHugh Dickins return ret; 59424c887265SAdam Litke 59434c887265SAdam Litke backout: 5944cb900f41SKirill A. Shutemov spin_unlock(ptl); 59452b26736cSAndy Whitcroft backout_unlocked: 5946c7b1850dSMike Kravetz if (new_page && !new_pagecache_page) 5947285b8dcaSHuang Ying restore_reserve_on_error(h, vma, haddr, page); 5948fa27759aSMike Kravetz 5949fa27759aSMike Kravetz unlock_page(page); 59504c887265SAdam Litke put_page(page); 59514c887265SAdam Litke goto out; 5952ac9b9c66SHugh Dickins } 5953ac9b9c66SHugh Dickins 59548382d914SDavidlohr Bueso #ifdef CONFIG_SMP 5955188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 59568382d914SDavidlohr Bueso { 59578382d914SDavidlohr Bueso unsigned long key[2]; 59588382d914SDavidlohr Bueso u32 hash; 59598382d914SDavidlohr Bueso 59608382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 59618382d914SDavidlohr Bueso key[1] = idx; 59628382d914SDavidlohr Bueso 596355254636SMike Kravetz hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); 59648382d914SDavidlohr Bueso 59658382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 59668382d914SDavidlohr Bueso } 59678382d914SDavidlohr Bueso #else 59688382d914SDavidlohr Bueso /* 59696c26d310SMiaohe Lin * For uniprocessor systems we always use a single mutex, so just 59708382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 59718382d914SDavidlohr Bueso */ 5972188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 59738382d914SDavidlohr Bueso { 59748382d914SDavidlohr Bueso return 0; 59758382d914SDavidlohr Bueso } 59768382d914SDavidlohr Bueso #endif 59778382d914SDavidlohr Bueso 59782b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 5979788c7df4SHugh Dickins unsigned long address, unsigned int flags) 598086e5216fSAdam Litke { 59818382d914SDavidlohr Bueso pte_t *ptep, entry; 5982cb900f41SKirill A. Shutemov spinlock_t *ptl; 59832b740303SSouptick Joarder vm_fault_t ret; 59848382d914SDavidlohr Bueso u32 hash; 59858382d914SDavidlohr Bueso pgoff_t idx; 59860fe6e20bSNaoya Horiguchi struct page *page = NULL; 598757303d80SAndy Whitcroft struct page *pagecache_page = NULL; 5988a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 59898382d914SDavidlohr Bueso struct address_space *mapping; 59900f792cf9SNaoya Horiguchi int need_wait_lock = 0; 5991285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 599286e5216fSAdam Litke 59933935baa9SDavid Gibson /* 59943935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 59953935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 59963935baa9SDavid Gibson * the same page in the page cache. 59973935baa9SDavid Gibson */ 599840549ba8SMike Kravetz mapping = vma->vm_file->f_mapping; 599940549ba8SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 6000188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 6001c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 60028382d914SDavidlohr Bueso 600340549ba8SMike Kravetz /* 600440549ba8SMike Kravetz * Acquire vma lock before calling huge_pte_alloc and hold 600540549ba8SMike Kravetz * until finished with ptep. This prevents huge_pmd_unshare from 600640549ba8SMike Kravetz * being called elsewhere and making the ptep no longer valid. 600740549ba8SMike Kravetz */ 600840549ba8SMike Kravetz hugetlb_vma_lock_read(vma); 600940549ba8SMike Kravetz ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); 601040549ba8SMike Kravetz if (!ptep) { 601140549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 601240549ba8SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 601340549ba8SMike Kravetz return VM_FAULT_OOM; 601440549ba8SMike Kravetz } 601540549ba8SMike Kravetz 60167f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 6017c64e912cSPeter Xu /* PTE markers should be handled the same way as none pte */ 6018958f32ceSLiu Shixin if (huge_pte_none_mostly(entry)) 6019958f32ceSLiu Shixin /* 6020958f32ceSLiu Shixin * hugetlb_no_page will drop vma lock and hugetlb fault 6021958f32ceSLiu Shixin * mutex internally, which make us return immediately. 6022958f32ceSLiu Shixin */ 6023958f32ceSLiu Shixin return hugetlb_no_page(mm, vma, mapping, idx, address, ptep, 6024c64e912cSPeter Xu entry, flags); 602586e5216fSAdam Litke 602683c54070SNick Piggin ret = 0; 60271e8f889bSDavid Gibson 602857303d80SAndy Whitcroft /* 60290f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 60300f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 60317c8de358SEthon Paul * an active hugepage in pagecache. This goto expects the 2nd page 60327c8de358SEthon Paul * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will 60337c8de358SEthon Paul * properly handle it. 60340f792cf9SNaoya Horiguchi */ 6035fcd48540SPeter Xu if (!pte_present(entry)) { 6036fcd48540SPeter Xu if (unlikely(is_hugetlb_entry_migration(entry))) { 6037fcd48540SPeter Xu /* 6038fcd48540SPeter Xu * Release the hugetlb fault lock now, but retain 6039fcd48540SPeter Xu * the vma lock, because it is needed to guard the 6040fcd48540SPeter Xu * huge_pte_lockptr() later in 6041fcd48540SPeter Xu * migration_entry_wait_huge(). The vma lock will 6042fcd48540SPeter Xu * be released there. 6043fcd48540SPeter Xu */ 6044fcd48540SPeter Xu mutex_unlock(&hugetlb_fault_mutex_table[hash]); 6045fcd48540SPeter Xu migration_entry_wait_huge(vma, ptep); 6046fcd48540SPeter Xu return 0; 6047fcd48540SPeter Xu } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 6048fcd48540SPeter Xu ret = VM_FAULT_HWPOISON_LARGE | 6049fcd48540SPeter Xu VM_FAULT_SET_HINDEX(hstate_index(h)); 60500f792cf9SNaoya Horiguchi goto out_mutex; 6051fcd48540SPeter Xu } 60520f792cf9SNaoya Horiguchi 60530f792cf9SNaoya Horiguchi /* 6054c89357e2SDavid Hildenbrand * If we are going to COW/unshare the mapping later, we examine the 6055c89357e2SDavid Hildenbrand * pending reservations for this page now. This will ensure that any 605657303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 60571d8d1464SDavid Hildenbrand * spinlock. Also lookup the pagecache page now as it is used to 60581d8d1464SDavid Hildenbrand * determine if a reservation has been consumed. 605957303d80SAndy Whitcroft */ 6060c89357e2SDavid Hildenbrand if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) && 60611d8d1464SDavid Hildenbrand !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(entry)) { 6062285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 60632b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 6064b4d1d99fSDavid Gibson goto out_mutex; 60652b26736cSAndy Whitcroft } 60665e911373SMike Kravetz /* Just decrements count, does not deallocate */ 6067285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 606857303d80SAndy Whitcroft 606929be8426SMiaohe Lin pagecache_page = find_lock_page(mapping, idx); 607057303d80SAndy Whitcroft } 607157303d80SAndy Whitcroft 60720f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 60730fe6e20bSNaoya Horiguchi 6074c89357e2SDavid Hildenbrand /* Check for a racing update before calling hugetlb_wp() */ 6075b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 6076cb900f41SKirill A. Shutemov goto out_ptl; 6077b4d1d99fSDavid Gibson 6078166f3eccSPeter Xu /* Handle userfault-wp first, before trying to lock more pages */ 6079166f3eccSPeter Xu if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) && 6080166f3eccSPeter Xu (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 6081166f3eccSPeter Xu struct vm_fault vmf = { 6082166f3eccSPeter Xu .vma = vma, 6083166f3eccSPeter Xu .address = haddr, 6084166f3eccSPeter Xu .real_address = address, 6085166f3eccSPeter Xu .flags = flags, 6086166f3eccSPeter Xu }; 6087166f3eccSPeter Xu 6088166f3eccSPeter Xu spin_unlock(ptl); 6089166f3eccSPeter Xu if (pagecache_page) { 6090166f3eccSPeter Xu unlock_page(pagecache_page); 6091166f3eccSPeter Xu put_page(pagecache_page); 6092166f3eccSPeter Xu } 609340549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 6094166f3eccSPeter Xu mutex_unlock(&hugetlb_fault_mutex_table[hash]); 6095166f3eccSPeter Xu return handle_userfault(&vmf, VM_UFFD_WP); 6096166f3eccSPeter Xu } 6097166f3eccSPeter Xu 60980f792cf9SNaoya Horiguchi /* 6099c89357e2SDavid Hildenbrand * hugetlb_wp() requires page locks of pte_page(entry) and 61000f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 61010f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 61020f792cf9SNaoya Horiguchi */ 61030f792cf9SNaoya Horiguchi page = pte_page(entry); 61040f792cf9SNaoya Horiguchi if (page != pagecache_page) 61050f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 61060f792cf9SNaoya Horiguchi need_wait_lock = 1; 61070f792cf9SNaoya Horiguchi goto out_ptl; 61080f792cf9SNaoya Horiguchi } 61090f792cf9SNaoya Horiguchi 61100f792cf9SNaoya Horiguchi get_page(page); 6111b4d1d99fSDavid Gibson 6112c89357e2SDavid Hildenbrand if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) { 6113106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 6114c89357e2SDavid Hildenbrand ret = hugetlb_wp(mm, vma, address, ptep, flags, 6115cb900f41SKirill A. Shutemov pagecache_page, ptl); 61160f792cf9SNaoya Horiguchi goto out_put_page; 6117c89357e2SDavid Hildenbrand } else if (likely(flags & FAULT_FLAG_WRITE)) { 6118106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 6119b4d1d99fSDavid Gibson } 6120c89357e2SDavid Hildenbrand } 6121b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 6122285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 6123788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 6124285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 61250f792cf9SNaoya Horiguchi out_put_page: 61260f792cf9SNaoya Horiguchi if (page != pagecache_page) 61270f792cf9SNaoya Horiguchi unlock_page(page); 61280f792cf9SNaoya Horiguchi put_page(page); 6129cb900f41SKirill A. Shutemov out_ptl: 6130cb900f41SKirill A. Shutemov spin_unlock(ptl); 613157303d80SAndy Whitcroft 613257303d80SAndy Whitcroft if (pagecache_page) { 613357303d80SAndy Whitcroft unlock_page(pagecache_page); 613457303d80SAndy Whitcroft put_page(pagecache_page); 613557303d80SAndy Whitcroft } 6136b4d1d99fSDavid Gibson out_mutex: 613740549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 6138c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 61390f792cf9SNaoya Horiguchi /* 61400f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 61410f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 61420f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 61430f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 61440f792cf9SNaoya Horiguchi * here without taking refcount. 61450f792cf9SNaoya Horiguchi */ 61460f792cf9SNaoya Horiguchi if (need_wait_lock) 61470f792cf9SNaoya Horiguchi wait_on_page_locked(page); 61481e8f889bSDavid Gibson return ret; 614986e5216fSAdam Litke } 615086e5216fSAdam Litke 6151714c1891SAxel Rasmussen #ifdef CONFIG_USERFAULTFD 61528fb5debcSMike Kravetz /* 61538fb5debcSMike Kravetz * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with 61548fb5debcSMike Kravetz * modifications for huge pages. 61558fb5debcSMike Kravetz */ 61568fb5debcSMike Kravetz int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, 61578fb5debcSMike Kravetz pte_t *dst_pte, 61588fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 61598fb5debcSMike Kravetz unsigned long dst_addr, 61608fb5debcSMike Kravetz unsigned long src_addr, 6161f6191471SAxel Rasmussen enum mcopy_atomic_mode mode, 61626041c691SPeter Xu struct page **pagep, 61636041c691SPeter Xu bool wp_copy) 61648fb5debcSMike Kravetz { 6165f6191471SAxel Rasmussen bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE); 61668cc5fcbbSMina Almasry struct hstate *h = hstate_vma(dst_vma); 61678cc5fcbbSMina Almasry struct address_space *mapping = dst_vma->vm_file->f_mapping; 61688cc5fcbbSMina Almasry pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr); 61691e392147SAndrea Arcangeli unsigned long size; 61701c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 61718fb5debcSMike Kravetz pte_t _dst_pte; 61728fb5debcSMike Kravetz spinlock_t *ptl; 61738cc5fcbbSMina Almasry int ret = -ENOMEM; 61748fb5debcSMike Kravetz struct page *page; 6175f6191471SAxel Rasmussen int writable; 6176cc30042dSMina Almasry bool page_in_pagecache = false; 61778fb5debcSMike Kravetz 6178f6191471SAxel Rasmussen if (is_continue) { 6179f6191471SAxel Rasmussen ret = -EFAULT; 6180f6191471SAxel Rasmussen page = find_lock_page(mapping, idx); 6181f6191471SAxel Rasmussen if (!page) 6182f6191471SAxel Rasmussen goto out; 6183cc30042dSMina Almasry page_in_pagecache = true; 6184f6191471SAxel Rasmussen } else if (!*pagep) { 6185d84cf06eSMina Almasry /* If a page already exists, then it's UFFDIO_COPY for 6186d84cf06eSMina Almasry * a non-missing case. Return -EEXIST. 6187d84cf06eSMina Almasry */ 6188d84cf06eSMina Almasry if (vm_shared && 6189d84cf06eSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 6190d84cf06eSMina Almasry ret = -EEXIST; 61918fb5debcSMike Kravetz goto out; 6192d84cf06eSMina Almasry } 6193d84cf06eSMina Almasry 6194d84cf06eSMina Almasry page = alloc_huge_page(dst_vma, dst_addr, 0); 6195d84cf06eSMina Almasry if (IS_ERR(page)) { 6196d84cf06eSMina Almasry ret = -ENOMEM; 6197d84cf06eSMina Almasry goto out; 6198d84cf06eSMina Almasry } 61998fb5debcSMike Kravetz 62008fb5debcSMike Kravetz ret = copy_huge_page_from_user(page, 62018fb5debcSMike Kravetz (const void __user *) src_addr, 6202810a56b9SMike Kravetz pages_per_huge_page(h), false); 62038fb5debcSMike Kravetz 6204c1e8d7c6SMichel Lespinasse /* fallback to copy_from_user outside mmap_lock */ 62058fb5debcSMike Kravetz if (unlikely(ret)) { 62069e368259SAndrea Arcangeli ret = -ENOENT; 62078cc5fcbbSMina Almasry /* Free the allocated page which may have 62088cc5fcbbSMina Almasry * consumed a reservation. 62098cc5fcbbSMina Almasry */ 62108cc5fcbbSMina Almasry restore_reserve_on_error(h, dst_vma, dst_addr, page); 62118cc5fcbbSMina Almasry put_page(page); 62128cc5fcbbSMina Almasry 62138cc5fcbbSMina Almasry /* Allocate a temporary page to hold the copied 62148cc5fcbbSMina Almasry * contents. 62158cc5fcbbSMina Almasry */ 62168cc5fcbbSMina Almasry page = alloc_huge_page_vma(h, dst_vma, dst_addr); 62178cc5fcbbSMina Almasry if (!page) { 62188cc5fcbbSMina Almasry ret = -ENOMEM; 62198cc5fcbbSMina Almasry goto out; 62208cc5fcbbSMina Almasry } 62218fb5debcSMike Kravetz *pagep = page; 62228cc5fcbbSMina Almasry /* Set the outparam pagep and return to the caller to 62238cc5fcbbSMina Almasry * copy the contents outside the lock. Don't free the 62248cc5fcbbSMina Almasry * page. 62258cc5fcbbSMina Almasry */ 62268fb5debcSMike Kravetz goto out; 62278fb5debcSMike Kravetz } 62288fb5debcSMike Kravetz } else { 62298cc5fcbbSMina Almasry if (vm_shared && 62308cc5fcbbSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 62318cc5fcbbSMina Almasry put_page(*pagep); 62328cc5fcbbSMina Almasry ret = -EEXIST; 62338cc5fcbbSMina Almasry *pagep = NULL; 62348cc5fcbbSMina Almasry goto out; 62358cc5fcbbSMina Almasry } 62368cc5fcbbSMina Almasry 62378cc5fcbbSMina Almasry page = alloc_huge_page(dst_vma, dst_addr, 0); 62388cc5fcbbSMina Almasry if (IS_ERR(page)) { 6239da9a298fSMiaohe Lin put_page(*pagep); 62408cc5fcbbSMina Almasry ret = -ENOMEM; 62418cc5fcbbSMina Almasry *pagep = NULL; 62428cc5fcbbSMina Almasry goto out; 62438cc5fcbbSMina Almasry } 624434892366SMuchun Song copy_user_huge_page(page, *pagep, dst_addr, dst_vma, 624534892366SMuchun Song pages_per_huge_page(h)); 62468cc5fcbbSMina Almasry put_page(*pagep); 62478fb5debcSMike Kravetz *pagep = NULL; 62488fb5debcSMike Kravetz } 62498fb5debcSMike Kravetz 62508fb5debcSMike Kravetz /* 62518fb5debcSMike Kravetz * The memory barrier inside __SetPageUptodate makes sure that 62528fb5debcSMike Kravetz * preceding stores to the page contents become visible before 62538fb5debcSMike Kravetz * the set_pte_at() write. 62548fb5debcSMike Kravetz */ 62558fb5debcSMike Kravetz __SetPageUptodate(page); 62568fb5debcSMike Kravetz 6257f6191471SAxel Rasmussen /* Add shared, newly allocated pages to the page cache. */ 6258f6191471SAxel Rasmussen if (vm_shared && !is_continue) { 62591e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 62601e392147SAndrea Arcangeli ret = -EFAULT; 62611e392147SAndrea Arcangeli if (idx >= size) 62621e392147SAndrea Arcangeli goto out_release_nounlock; 62631c9e8defSMike Kravetz 62641e392147SAndrea Arcangeli /* 62651e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 62667e1813d4SMike Kravetz * hugetlb_add_to_page_cache() below happens through the 62671e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 62681e392147SAndrea Arcangeli * the caller. 62691e392147SAndrea Arcangeli */ 62707e1813d4SMike Kravetz ret = hugetlb_add_to_page_cache(page, mapping, idx); 62711c9e8defSMike Kravetz if (ret) 62721c9e8defSMike Kravetz goto out_release_nounlock; 6273cc30042dSMina Almasry page_in_pagecache = true; 62741c9e8defSMike Kravetz } 62751c9e8defSMike Kravetz 6276bcc66543SMiaohe Lin ptl = huge_pte_lock(h, dst_mm, dst_pte); 62778fb5debcSMike Kravetz 62788625147cSJames Houghton ret = -EIO; 62798625147cSJames Houghton if (PageHWPoison(page)) 62808625147cSJames Houghton goto out_release_unlock; 62818625147cSJames Houghton 62821e392147SAndrea Arcangeli /* 62836041c691SPeter Xu * We allow to overwrite a pte marker: consider when both MISSING|WP 62846041c691SPeter Xu * registered, we firstly wr-protect a none pte which has no page cache 62856041c691SPeter Xu * page backing it, then access the page. 62866041c691SPeter Xu */ 6287fa27759aSMike Kravetz ret = -EEXIST; 62886041c691SPeter Xu if (!huge_pte_none_mostly(huge_ptep_get(dst_pte))) 62898fb5debcSMike Kravetz goto out_release_unlock; 62908fb5debcSMike Kravetz 62914781593dSPeter Xu if (page_in_pagecache) 6292fb3d824dSDavid Hildenbrand page_dup_file_rmap(page, true); 62934781593dSPeter Xu else 62948fb5debcSMike Kravetz hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); 62958fb5debcSMike Kravetz 62966041c691SPeter Xu /* 62976041c691SPeter Xu * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY 62986041c691SPeter Xu * with wp flag set, don't set pte write bit. 62996041c691SPeter Xu */ 63006041c691SPeter Xu if (wp_copy || (is_continue && !vm_shared)) 6301f6191471SAxel Rasmussen writable = 0; 6302f6191471SAxel Rasmussen else 6303f6191471SAxel Rasmussen writable = dst_vma->vm_flags & VM_WRITE; 6304f6191471SAxel Rasmussen 6305f6191471SAxel Rasmussen _dst_pte = make_huge_pte(dst_vma, page, writable); 63066041c691SPeter Xu /* 63076041c691SPeter Xu * Always mark UFFDIO_COPY page dirty; note that this may not be 63086041c691SPeter Xu * extremely important for hugetlbfs for now since swapping is not 63096041c691SPeter Xu * supported, but we should still be clear in that this page cannot be 63106041c691SPeter Xu * thrown away at will, even if write bit not set. 63116041c691SPeter Xu */ 63128fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 63138fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 63148fb5debcSMike Kravetz 63156041c691SPeter Xu if (wp_copy) 63166041c691SPeter Xu _dst_pte = huge_pte_mkuffd_wp(_dst_pte); 63176041c691SPeter Xu 63188fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 63198fb5debcSMike Kravetz 63208fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 63218fb5debcSMike Kravetz 63228fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 63238fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 63248fb5debcSMike Kravetz 63258fb5debcSMike Kravetz spin_unlock(ptl); 6326f6191471SAxel Rasmussen if (!is_continue) 63278f251a3dSMike Kravetz SetHPageMigratable(page); 6328f6191471SAxel Rasmussen if (vm_shared || is_continue) 63291c9e8defSMike Kravetz unlock_page(page); 63308fb5debcSMike Kravetz ret = 0; 63318fb5debcSMike Kravetz out: 63328fb5debcSMike Kravetz return ret; 63338fb5debcSMike Kravetz out_release_unlock: 63348fb5debcSMike Kravetz spin_unlock(ptl); 6335f6191471SAxel Rasmussen if (vm_shared || is_continue) 63361c9e8defSMike Kravetz unlock_page(page); 63375af10dfdSAndrea Arcangeli out_release_nounlock: 6338cc30042dSMina Almasry if (!page_in_pagecache) 6339846be085SMike Kravetz restore_reserve_on_error(h, dst_vma, dst_addr, page); 63408fb5debcSMike Kravetz put_page(page); 63418fb5debcSMike Kravetz goto out; 63428fb5debcSMike Kravetz } 6343714c1891SAxel Rasmussen #endif /* CONFIG_USERFAULTFD */ 63448fb5debcSMike Kravetz 634582e5d378SJoao Martins static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma, 634682e5d378SJoao Martins int refs, struct page **pages, 634782e5d378SJoao Martins struct vm_area_struct **vmas) 634882e5d378SJoao Martins { 634982e5d378SJoao Martins int nr; 635082e5d378SJoao Martins 635182e5d378SJoao Martins for (nr = 0; nr < refs; nr++) { 635282e5d378SJoao Martins if (likely(pages)) 635314455eabSCheng Li pages[nr] = nth_page(page, nr); 635482e5d378SJoao Martins if (vmas) 635582e5d378SJoao Martins vmas[nr] = vma; 635682e5d378SJoao Martins } 635782e5d378SJoao Martins } 635882e5d378SJoao Martins 635984209e87SDavid Hildenbrand static inline bool __follow_hugetlb_must_fault(struct vm_area_struct *vma, 636084209e87SDavid Hildenbrand unsigned int flags, pte_t *pte, 6361a7f22660SDavid Hildenbrand bool *unshare) 6362a7f22660SDavid Hildenbrand { 6363a7f22660SDavid Hildenbrand pte_t pteval = huge_ptep_get(pte); 6364a7f22660SDavid Hildenbrand 6365a7f22660SDavid Hildenbrand *unshare = false; 6366a7f22660SDavid Hildenbrand if (is_swap_pte(pteval)) 6367a7f22660SDavid Hildenbrand return true; 6368a7f22660SDavid Hildenbrand if (huge_pte_write(pteval)) 6369a7f22660SDavid Hildenbrand return false; 6370a7f22660SDavid Hildenbrand if (flags & FOLL_WRITE) 6371a7f22660SDavid Hildenbrand return true; 637284209e87SDavid Hildenbrand if (gup_must_unshare(vma, flags, pte_page(pteval))) { 6373a7f22660SDavid Hildenbrand *unshare = true; 6374a7f22660SDavid Hildenbrand return true; 6375a7f22660SDavid Hildenbrand } 6376a7f22660SDavid Hildenbrand return false; 6377a7f22660SDavid Hildenbrand } 6378a7f22660SDavid Hildenbrand 637957a196a5SMike Kravetz struct page *hugetlb_follow_page_mask(struct vm_area_struct *vma, 638057a196a5SMike Kravetz unsigned long address, unsigned int flags) 638157a196a5SMike Kravetz { 638257a196a5SMike Kravetz struct hstate *h = hstate_vma(vma); 638357a196a5SMike Kravetz struct mm_struct *mm = vma->vm_mm; 638457a196a5SMike Kravetz unsigned long haddr = address & huge_page_mask(h); 638557a196a5SMike Kravetz struct page *page = NULL; 638657a196a5SMike Kravetz spinlock_t *ptl; 638757a196a5SMike Kravetz pte_t *pte, entry; 638857a196a5SMike Kravetz 638957a196a5SMike Kravetz /* 639057a196a5SMike Kravetz * FOLL_PIN is not supported for follow_page(). Ordinary GUP goes via 639157a196a5SMike Kravetz * follow_hugetlb_page(). 639257a196a5SMike Kravetz */ 639357a196a5SMike Kravetz if (WARN_ON_ONCE(flags & FOLL_PIN)) 639457a196a5SMike Kravetz return NULL; 639557a196a5SMike Kravetz 63967d049f3aSPeter Xu hugetlb_vma_lock_read(vma); 63979c67a207SPeter Xu pte = hugetlb_walk(vma, haddr, huge_page_size(h)); 639857a196a5SMike Kravetz if (!pte) 63997d049f3aSPeter Xu goto out_unlock; 640057a196a5SMike Kravetz 640157a196a5SMike Kravetz ptl = huge_pte_lock(h, mm, pte); 640257a196a5SMike Kravetz entry = huge_ptep_get(pte); 640357a196a5SMike Kravetz if (pte_present(entry)) { 640457a196a5SMike Kravetz page = pte_page(entry) + 640557a196a5SMike Kravetz ((address & ~huge_page_mask(h)) >> PAGE_SHIFT); 640657a196a5SMike Kravetz /* 640757a196a5SMike Kravetz * Note that page may be a sub-page, and with vmemmap 640857a196a5SMike Kravetz * optimizations the page struct may be read only. 640957a196a5SMike Kravetz * try_grab_page() will increase the ref count on the 641057a196a5SMike Kravetz * head page, so this will be OK. 641157a196a5SMike Kravetz * 6412e2ca6ba6SLinus Torvalds * try_grab_page() should always be able to get the page here, 6413e2ca6ba6SLinus Torvalds * because we hold the ptl lock and have verified pte_present(). 641457a196a5SMike Kravetz */ 6415e2ca6ba6SLinus Torvalds if (try_grab_page(page, flags)) { 641657a196a5SMike Kravetz page = NULL; 641757a196a5SMike Kravetz goto out; 641857a196a5SMike Kravetz } 641957a196a5SMike Kravetz } 642057a196a5SMike Kravetz out: 642157a196a5SMike Kravetz spin_unlock(ptl); 64227d049f3aSPeter Xu out_unlock: 64237d049f3aSPeter Xu hugetlb_vma_unlock_read(vma); 642457a196a5SMike Kravetz return page; 642557a196a5SMike Kravetz } 642657a196a5SMike Kravetz 642728a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 642863551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 642928a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 64304f6da934SPeter Xu long i, unsigned int flags, int *locked) 643163551ae0SDavid Gibson { 6432d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 6433d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 643428a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 6435a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 64360fa5bc40SJoao Martins int err = -EFAULT, refs; 643763551ae0SDavid Gibson 643863551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 643963551ae0SDavid Gibson pte_t *pte; 6440cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 6441a7f22660SDavid Hildenbrand bool unshare = false; 64422a15efc9SHugh Dickins int absent; 644363551ae0SDavid Gibson struct page *page; 644463551ae0SDavid Gibson 64454c887265SAdam Litke /* 644602057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 644702057967SDavid Rientjes * potentially allocating memory. 644802057967SDavid Rientjes */ 6449fa45f116SDavidlohr Bueso if (fatal_signal_pending(current)) { 645002057967SDavid Rientjes remainder = 0; 645102057967SDavid Rientjes break; 645202057967SDavid Rientjes } 645302057967SDavid Rientjes 6454eefc7fa5SPeter Xu hugetlb_vma_lock_read(vma); 645502057967SDavid Rientjes /* 64564c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 64572a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 64584c887265SAdam Litke * first, for the page indexing below to work. 6459cb900f41SKirill A. Shutemov * 6460cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 64614c887265SAdam Litke */ 64629c67a207SPeter Xu pte = hugetlb_walk(vma, vaddr & huge_page_mask(h), 64637868a208SPunit Agrawal huge_page_size(h)); 6464cb900f41SKirill A. Shutemov if (pte) 6465cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 64662a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 646763551ae0SDavid Gibson 64682a15efc9SHugh Dickins /* 64692a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 64703ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 64713ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 64723ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 64733ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 64742a15efc9SHugh Dickins */ 64753ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 64763ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 6477cb900f41SKirill A. Shutemov if (pte) 6478cb900f41SKirill A. Shutemov spin_unlock(ptl); 6479eefc7fa5SPeter Xu hugetlb_vma_unlock_read(vma); 64802a15efc9SHugh Dickins remainder = 0; 64812a15efc9SHugh Dickins break; 64822a15efc9SHugh Dickins } 64832a15efc9SHugh Dickins 64849cc3a5bdSNaoya Horiguchi /* 64859cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 64869cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 64879cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 64889cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 64899cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 64909cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 64919cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 64929cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 64939cc3a5bdSNaoya Horiguchi */ 6494a7f22660SDavid Hildenbrand if (absent || 649584209e87SDavid Hildenbrand __follow_hugetlb_must_fault(vma, flags, pte, &unshare)) { 64962b740303SSouptick Joarder vm_fault_t ret; 649787ffc118SAndrea Arcangeli unsigned int fault_flags = 0; 64984c887265SAdam Litke 6499cb900f41SKirill A. Shutemov if (pte) 6500cb900f41SKirill A. Shutemov spin_unlock(ptl); 6501eefc7fa5SPeter Xu hugetlb_vma_unlock_read(vma); 6502eefc7fa5SPeter Xu 650387ffc118SAndrea Arcangeli if (flags & FOLL_WRITE) 650487ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_WRITE; 6505a7f22660SDavid Hildenbrand else if (unshare) 6506a7f22660SDavid Hildenbrand fault_flags |= FAULT_FLAG_UNSHARE; 650793c5c61dSPeter Xu if (locked) { 650871335f37SPeter Xu fault_flags |= FAULT_FLAG_ALLOW_RETRY | 650971335f37SPeter Xu FAULT_FLAG_KILLABLE; 651093c5c61dSPeter Xu if (flags & FOLL_INTERRUPTIBLE) 651193c5c61dSPeter Xu fault_flags |= FAULT_FLAG_INTERRUPTIBLE; 651293c5c61dSPeter Xu } 651387ffc118SAndrea Arcangeli if (flags & FOLL_NOWAIT) 651487ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY | 651587ffc118SAndrea Arcangeli FAULT_FLAG_RETRY_NOWAIT; 651687ffc118SAndrea Arcangeli if (flags & FOLL_TRIED) { 65174426e945SPeter Xu /* 65184426e945SPeter Xu * Note: FAULT_FLAG_ALLOW_RETRY and 65194426e945SPeter Xu * FAULT_FLAG_TRIED can co-exist 65204426e945SPeter Xu */ 652187ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_TRIED; 652287ffc118SAndrea Arcangeli } 652387ffc118SAndrea Arcangeli ret = hugetlb_fault(mm, vma, vaddr, fault_flags); 652487ffc118SAndrea Arcangeli if (ret & VM_FAULT_ERROR) { 65252be7cfedSDaniel Jordan err = vm_fault_to_errno(ret, flags); 65261c59827dSHugh Dickins remainder = 0; 65271c59827dSHugh Dickins break; 65281c59827dSHugh Dickins } 652987ffc118SAndrea Arcangeli if (ret & VM_FAULT_RETRY) { 65304f6da934SPeter Xu if (locked && 65311ac25013SAndrea Arcangeli !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) 65324f6da934SPeter Xu *locked = 0; 653387ffc118SAndrea Arcangeli *nr_pages = 0; 653487ffc118SAndrea Arcangeli /* 653587ffc118SAndrea Arcangeli * VM_FAULT_RETRY must not return an 653687ffc118SAndrea Arcangeli * error, it will return zero 653787ffc118SAndrea Arcangeli * instead. 653887ffc118SAndrea Arcangeli * 653987ffc118SAndrea Arcangeli * No need to update "position" as the 654087ffc118SAndrea Arcangeli * caller will not check it after 654187ffc118SAndrea Arcangeli * *nr_pages is set to 0. 654287ffc118SAndrea Arcangeli */ 654387ffc118SAndrea Arcangeli return i; 654487ffc118SAndrea Arcangeli } 654587ffc118SAndrea Arcangeli continue; 654687ffc118SAndrea Arcangeli } 654763551ae0SDavid Gibson 6548a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 65497f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 65508fde12caSLinus Torvalds 6551b6a2619cSDavid Hildenbrand VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) && 6552b6a2619cSDavid Hildenbrand !PageAnonExclusive(page), page); 6553b6a2619cSDavid Hildenbrand 65548fde12caSLinus Torvalds /* 6555acbfb087SZhigang Lu * If subpage information not requested, update counters 6556acbfb087SZhigang Lu * and skip the same_page loop below. 6557acbfb087SZhigang Lu */ 6558acbfb087SZhigang Lu if (!pages && !vmas && !pfn_offset && 6559acbfb087SZhigang Lu (vaddr + huge_page_size(h) < vma->vm_end) && 6560acbfb087SZhigang Lu (remainder >= pages_per_huge_page(h))) { 6561acbfb087SZhigang Lu vaddr += huge_page_size(h); 6562acbfb087SZhigang Lu remainder -= pages_per_huge_page(h); 6563acbfb087SZhigang Lu i += pages_per_huge_page(h); 6564acbfb087SZhigang Lu spin_unlock(ptl); 6565eefc7fa5SPeter Xu hugetlb_vma_unlock_read(vma); 6566acbfb087SZhigang Lu continue; 6567acbfb087SZhigang Lu } 6568acbfb087SZhigang Lu 6569d08af0a5SJoao Martins /* vaddr may not be aligned to PAGE_SIZE */ 6570d08af0a5SJoao Martins refs = min3(pages_per_huge_page(h) - pfn_offset, remainder, 6571d08af0a5SJoao Martins (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT); 65720fa5bc40SJoao Martins 657382e5d378SJoao Martins if (pages || vmas) 657414455eabSCheng Li record_subpages_vmas(nth_page(page, pfn_offset), 657582e5d378SJoao Martins vma, refs, 657682e5d378SJoao Martins likely(pages) ? pages + i : NULL, 657782e5d378SJoao Martins vmas ? vmas + i : NULL); 657863551ae0SDavid Gibson 657982e5d378SJoao Martins if (pages) { 65800fa5bc40SJoao Martins /* 6581822951d8SMatthew Wilcox (Oracle) * try_grab_folio() should always succeed here, 65820fa5bc40SJoao Martins * because: a) we hold the ptl lock, and b) we've just 65830fa5bc40SJoao Martins * checked that the huge page is present in the page 65840fa5bc40SJoao Martins * tables. If the huge page is present, then the tail 65850fa5bc40SJoao Martins * pages must also be present. The ptl prevents the 65860fa5bc40SJoao Martins * head page and tail pages from being rearranged in 65874003f107SLogan Gunthorpe * any way. As this is hugetlb, the pages will never 65884003f107SLogan Gunthorpe * be p2pdma or not longterm pinable. So this page 65894003f107SLogan Gunthorpe * must be available at this point, unless the page 65904003f107SLogan Gunthorpe * refcount overflowed: 65910fa5bc40SJoao Martins */ 6592822951d8SMatthew Wilcox (Oracle) if (WARN_ON_ONCE(!try_grab_folio(pages[i], refs, 65930fa5bc40SJoao Martins flags))) { 65940fa5bc40SJoao Martins spin_unlock(ptl); 6595eefc7fa5SPeter Xu hugetlb_vma_unlock_read(vma); 65960fa5bc40SJoao Martins remainder = 0; 65970fa5bc40SJoao Martins err = -ENOMEM; 65980fa5bc40SJoao Martins break; 65990fa5bc40SJoao Martins } 6600d5d4b0aaSChen, Kenneth W } 660182e5d378SJoao Martins 660282e5d378SJoao Martins vaddr += (refs << PAGE_SHIFT); 660382e5d378SJoao Martins remainder -= refs; 660482e5d378SJoao Martins i += refs; 660582e5d378SJoao Martins 6606cb900f41SKirill A. Shutemov spin_unlock(ptl); 6607eefc7fa5SPeter Xu hugetlb_vma_unlock_read(vma); 660863551ae0SDavid Gibson } 660928a35716SMichel Lespinasse *nr_pages = remainder; 661087ffc118SAndrea Arcangeli /* 661187ffc118SAndrea Arcangeli * setting position is actually required only if remainder is 661287ffc118SAndrea Arcangeli * not zero but it's faster not to add a "if (remainder)" 661387ffc118SAndrea Arcangeli * branch. 661487ffc118SAndrea Arcangeli */ 661563551ae0SDavid Gibson *position = vaddr; 661663551ae0SDavid Gibson 66172be7cfedSDaniel Jordan return i ? i : err; 661863551ae0SDavid Gibson } 66198f860591SZhang, Yanmin 6620a79390f5SPeter Xu long hugetlb_change_protection(struct vm_area_struct *vma, 66215a90d5a1SPeter Xu unsigned long address, unsigned long end, 66225a90d5a1SPeter Xu pgprot_t newprot, unsigned long cp_flags) 66238f860591SZhang, Yanmin { 66248f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 66258f860591SZhang, Yanmin unsigned long start = address; 66268f860591SZhang, Yanmin pte_t *ptep; 66278f860591SZhang, Yanmin pte_t pte; 6628a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 6629a79390f5SPeter Xu long pages = 0, psize = huge_page_size(h); 6630dff11abeSMike Kravetz bool shared_pmd = false; 6631ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 6632e95a9851SMike Kravetz unsigned long last_addr_mask; 66335a90d5a1SPeter Xu bool uffd_wp = cp_flags & MM_CP_UFFD_WP; 66345a90d5a1SPeter Xu bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE; 6635dff11abeSMike Kravetz 6636dff11abeSMike Kravetz /* 6637dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 6638ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 6639dff11abeSMike Kravetz * range if PMD sharing is possible. 6640dff11abeSMike Kravetz */ 66417269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 66427d4a8be0SAlistair Popple 0, mm, start, end); 6643ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 66448f860591SZhang, Yanmin 66458f860591SZhang, Yanmin BUG_ON(address >= end); 6646ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 66478f860591SZhang, Yanmin 6648ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 664940549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 665083cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 665140549ba8SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 665260dfaad6SPeter Xu for (; address < end; address += psize) { 6653cb900f41SKirill A. Shutemov spinlock_t *ptl; 66549c67a207SPeter Xu ptep = hugetlb_walk(vma, address, psize); 6655e95a9851SMike Kravetz if (!ptep) { 6656fed15f13SPeter Xu if (!uffd_wp) { 6657e95a9851SMike Kravetz address |= last_addr_mask; 66588f860591SZhang, Yanmin continue; 6659e95a9851SMike Kravetz } 6660fed15f13SPeter Xu /* 6661fed15f13SPeter Xu * Userfaultfd wr-protect requires pgtable 6662fed15f13SPeter Xu * pre-allocations to install pte markers. 6663fed15f13SPeter Xu */ 6664fed15f13SPeter Xu ptep = huge_pte_alloc(mm, vma, address, psize); 6665d1751118SPeter Xu if (!ptep) { 6666d1751118SPeter Xu pages = -ENOMEM; 6667fed15f13SPeter Xu break; 6668fed15f13SPeter Xu } 6669d1751118SPeter Xu } 6670cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 66714ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, address, ptep)) { 667260dfaad6SPeter Xu /* 667360dfaad6SPeter Xu * When uffd-wp is enabled on the vma, unshare 667460dfaad6SPeter Xu * shouldn't happen at all. Warn about it if it 667560dfaad6SPeter Xu * happened due to some reason. 667660dfaad6SPeter Xu */ 667760dfaad6SPeter Xu WARN_ON_ONCE(uffd_wp || uffd_wp_resolve); 66787da4d641SPeter Zijlstra pages++; 6679cb900f41SKirill A. Shutemov spin_unlock(ptl); 6680dff11abeSMike Kravetz shared_pmd = true; 66814ddb4d91SMike Kravetz address |= last_addr_mask; 668239dde65cSChen, Kenneth W continue; 66837da4d641SPeter Zijlstra } 6684a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 6685a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 66860e678153SDavid Hildenbrand /* Nothing to do. */ 66870e678153SDavid Hildenbrand } else if (unlikely(is_hugetlb_entry_migration(pte))) { 6688a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 66896c287605SDavid Hildenbrand struct page *page = pfn_swap_entry_to_page(entry); 669044f86392SDavid Hildenbrand pte_t newpte = pte; 6691a8bda28dSNaoya Horiguchi 669244f86392SDavid Hildenbrand if (is_writable_migration_entry(entry)) { 66936c287605SDavid Hildenbrand if (PageAnon(page)) 66946c287605SDavid Hildenbrand entry = make_readable_exclusive_migration_entry( 66956c287605SDavid Hildenbrand swp_offset(entry)); 66966c287605SDavid Hildenbrand else 66974dd845b5SAlistair Popple entry = make_readable_migration_entry( 66984dd845b5SAlistair Popple swp_offset(entry)); 6699a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 670044f86392SDavid Hildenbrand pages++; 670144f86392SDavid Hildenbrand } 670244f86392SDavid Hildenbrand 67035a90d5a1SPeter Xu if (uffd_wp) 67045a90d5a1SPeter Xu newpte = pte_swp_mkuffd_wp(newpte); 67055a90d5a1SPeter Xu else if (uffd_wp_resolve) 67065a90d5a1SPeter Xu newpte = pte_swp_clear_uffd_wp(newpte); 670744f86392SDavid Hildenbrand if (!pte_same(pte, newpte)) 670818f39629SQi Zheng set_huge_pte_at(mm, address, ptep, newpte); 67090e678153SDavid Hildenbrand } else if (unlikely(is_pte_marker(pte))) { 67100e678153SDavid Hildenbrand /* No other markers apply for now. */ 67110e678153SDavid Hildenbrand WARN_ON_ONCE(!pte_marker_uffd_wp(pte)); 671260dfaad6SPeter Xu if (uffd_wp_resolve) 67130e678153SDavid Hildenbrand /* Safe to modify directly (non-present->none). */ 671460dfaad6SPeter Xu huge_pte_clear(mm, address, ptep, psize); 67150e678153SDavid Hildenbrand } else if (!huge_pte_none(pte)) { 6716023bdd00SAneesh Kumar K.V pte_t old_pte; 671779c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 6718023bdd00SAneesh Kumar K.V 6719023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 672016785bd7SAnshuman Khandual pte = huge_pte_modify(old_pte, newprot); 672179c1c594SChristophe Leroy pte = arch_make_huge_pte(pte, shift, vma->vm_flags); 67225a90d5a1SPeter Xu if (uffd_wp) 6723f1eb1bacSPeter Xu pte = huge_pte_mkuffd_wp(pte); 67245a90d5a1SPeter Xu else if (uffd_wp_resolve) 67255a90d5a1SPeter Xu pte = huge_pte_clear_uffd_wp(pte); 6726023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 67277da4d641SPeter Zijlstra pages++; 672860dfaad6SPeter Xu } else { 672960dfaad6SPeter Xu /* None pte */ 673060dfaad6SPeter Xu if (unlikely(uffd_wp)) 673160dfaad6SPeter Xu /* Safe to modify directly (none->non-present). */ 673260dfaad6SPeter Xu set_huge_pte_at(mm, address, ptep, 673360dfaad6SPeter Xu make_pte_marker(PTE_MARKER_UFFD_WP)); 67348f860591SZhang, Yanmin } 6735cb900f41SKirill A. Shutemov spin_unlock(ptl); 67368f860591SZhang, Yanmin } 6737d833352aSMel Gorman /* 6738c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 6739d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 6740c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 6741dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 6742dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 6743d833352aSMel Gorman */ 6744dff11abeSMike Kravetz if (shared_pmd) 6745ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 6746dff11abeSMike Kravetz else 67475491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 67480f10851eSJérôme Glisse /* 67490f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 67500f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 67510f10851eSJérôme Glisse * 6752ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 67530f10851eSJérôme Glisse */ 675483cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 675540549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 6756ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 67577da4d641SPeter Zijlstra 6758d1751118SPeter Xu return pages > 0 ? (pages << h->order) : pages; 67598f860591SZhang, Yanmin } 67608f860591SZhang, Yanmin 676133b8f84aSMike Kravetz /* Return true if reservation was successful, false otherwise. */ 676233b8f84aSMike Kravetz bool hugetlb_reserve_pages(struct inode *inode, 6763a1e78772SMel Gorman long from, long to, 67645a6fe125SMel Gorman struct vm_area_struct *vma, 6765ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 6766e4e574b7SAdam Litke { 6767c5094ec7SMike Kravetz long chg = -1, add = -1; 6768a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 676990481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 67709119a41eSJoonsoo Kim struct resv_map *resv_map; 6771075a61d0SMina Almasry struct hugetlb_cgroup *h_cg = NULL; 67720db9d74eSMina Almasry long gbl_reserve, regions_needed = 0; 6773e4e574b7SAdam Litke 677463489f8eSMike Kravetz /* This should never happen */ 677563489f8eSMike Kravetz if (from > to) { 677663489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 677733b8f84aSMike Kravetz return false; 677863489f8eSMike Kravetz } 677963489f8eSMike Kravetz 6780a1e78772SMel Gorman /* 6781e700898fSMike Kravetz * vma specific semaphore used for pmd sharing and fault/truncation 6782e700898fSMike Kravetz * synchronization 67838d9bfb26SMike Kravetz */ 67848d9bfb26SMike Kravetz hugetlb_vma_lock_alloc(vma); 67858d9bfb26SMike Kravetz 67868d9bfb26SMike Kravetz /* 678717c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 678817c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 678990481622SDavid Gibson * without using reserves 679017c9d12eSMel Gorman */ 6791ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 679233b8f84aSMike Kravetz return true; 679317c9d12eSMel Gorman 679417c9d12eSMel Gorman /* 6795a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 6796a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 6797a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 6798a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 6799a1e78772SMel Gorman */ 68009119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 6801f27a5136SMike Kravetz /* 6802f27a5136SMike Kravetz * resv_map can not be NULL as hugetlb_reserve_pages is only 6803f27a5136SMike Kravetz * called for inodes for which resv_maps were created (see 6804f27a5136SMike Kravetz * hugetlbfs_get_inode). 6805f27a5136SMike Kravetz */ 68064e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 68079119a41eSJoonsoo Kim 68080db9d74eSMina Almasry chg = region_chg(resv_map, from, to, ®ions_needed); 68099119a41eSJoonsoo Kim } else { 6810e9fe92aeSMina Almasry /* Private mapping. */ 68119119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 68125a6fe125SMel Gorman if (!resv_map) 68138d9bfb26SMike Kravetz goto out_err; 68145a6fe125SMel Gorman 681517c9d12eSMel Gorman chg = to - from; 681617c9d12eSMel Gorman 68175a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 68185a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 68195a6fe125SMel Gorman } 68205a6fe125SMel Gorman 682133b8f84aSMike Kravetz if (chg < 0) 6822c50ac050SDave Hansen goto out_err; 682317c9d12eSMel Gorman 682433b8f84aSMike Kravetz if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), 682533b8f84aSMike Kravetz chg * pages_per_huge_page(h), &h_cg) < 0) 6826075a61d0SMina Almasry goto out_err; 6827075a61d0SMina Almasry 6828075a61d0SMina Almasry if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { 6829075a61d0SMina Almasry /* For private mappings, the hugetlb_cgroup uncharge info hangs 6830075a61d0SMina Almasry * of the resv_map. 6831075a61d0SMina Almasry */ 6832075a61d0SMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); 6833075a61d0SMina Almasry } 6834075a61d0SMina Almasry 68351c5ecae3SMike Kravetz /* 68361c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 68371c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 68381c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 68391c5ecae3SMike Kravetz */ 68401c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 684133b8f84aSMike Kravetz if (gbl_reserve < 0) 6842075a61d0SMina Almasry goto out_uncharge_cgroup; 684317c9d12eSMel Gorman 684417c9d12eSMel Gorman /* 684517c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 684690481622SDavid Gibson * Hand the pages back to the subpool if there are not 684717c9d12eSMel Gorman */ 684833b8f84aSMike Kravetz if (hugetlb_acct_memory(h, gbl_reserve) < 0) 6849075a61d0SMina Almasry goto out_put_pages; 685017c9d12eSMel Gorman 685117c9d12eSMel Gorman /* 685217c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 685317c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 685417c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 685517c9d12eSMel Gorman * the reservation was and the page cache tells how much of 685617c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 685717c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 685817c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 685917c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 686017c9d12eSMel Gorman * else has to be done for private mappings here 686117c9d12eSMel Gorman */ 686233039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 6863075a61d0SMina Almasry add = region_add(resv_map, from, to, regions_needed, h, h_cg); 686433039678SMike Kravetz 68650db9d74eSMina Almasry if (unlikely(add < 0)) { 68660db9d74eSMina Almasry hugetlb_acct_memory(h, -gbl_reserve); 6867075a61d0SMina Almasry goto out_put_pages; 68680db9d74eSMina Almasry } else if (unlikely(chg > add)) { 686933039678SMike Kravetz /* 687033039678SMike Kravetz * pages in this range were added to the reserve 687133039678SMike Kravetz * map between region_chg and region_add. This 687233039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 687333039678SMike Kravetz * the subpool and reserve counts modified above 687433039678SMike Kravetz * based on the difference. 687533039678SMike Kravetz */ 687633039678SMike Kravetz long rsv_adjust; 687733039678SMike Kravetz 6878d85aecf2SMiaohe Lin /* 6879d85aecf2SMiaohe Lin * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the 6880d85aecf2SMiaohe Lin * reference to h_cg->css. See comment below for detail. 6881d85aecf2SMiaohe Lin */ 6882075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd( 6883075a61d0SMina Almasry hstate_index(h), 6884075a61d0SMina Almasry (chg - add) * pages_per_huge_page(h), h_cg); 6885075a61d0SMina Almasry 688633039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 688733039678SMike Kravetz chg - add); 688833039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 6889d85aecf2SMiaohe Lin } else if (h_cg) { 6890d85aecf2SMiaohe Lin /* 6891d85aecf2SMiaohe Lin * The file_regions will hold their own reference to 6892d85aecf2SMiaohe Lin * h_cg->css. So we should release the reference held 6893d85aecf2SMiaohe Lin * via hugetlb_cgroup_charge_cgroup_rsvd() when we are 6894d85aecf2SMiaohe Lin * done. 6895d85aecf2SMiaohe Lin */ 6896d85aecf2SMiaohe Lin hugetlb_cgroup_put_rsvd_cgroup(h_cg); 689733039678SMike Kravetz } 689833039678SMike Kravetz } 689933b8f84aSMike Kravetz return true; 690033b8f84aSMike Kravetz 6901075a61d0SMina Almasry out_put_pages: 6902075a61d0SMina Almasry /* put back original number of pages, chg */ 6903075a61d0SMina Almasry (void)hugepage_subpool_put_pages(spool, chg); 6904075a61d0SMina Almasry out_uncharge_cgroup: 6905075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), 6906075a61d0SMina Almasry chg * pages_per_huge_page(h), h_cg); 6907c50ac050SDave Hansen out_err: 69088d9bfb26SMike Kravetz hugetlb_vma_lock_free(vma); 69095e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 69100db9d74eSMina Almasry /* Only call region_abort if the region_chg succeeded but the 69110db9d74eSMina Almasry * region_add failed or didn't run. 69120db9d74eSMina Almasry */ 69130db9d74eSMina Almasry if (chg >= 0 && add < 0) 69140db9d74eSMina Almasry region_abort(resv_map, from, to, regions_needed); 6915f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 6916f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 691733b8f84aSMike Kravetz return false; 6918a43a8c39SChen, Kenneth W } 6919a43a8c39SChen, Kenneth W 6920b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 6921b5cec28dSMike Kravetz long freed) 6922a43a8c39SChen, Kenneth W { 6923a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 69244e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 69259119a41eSJoonsoo Kim long chg = 0; 692690481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 69271c5ecae3SMike Kravetz long gbl_reserve; 692845c682a6SKen Chen 6929f27a5136SMike Kravetz /* 6930f27a5136SMike Kravetz * Since this routine can be called in the evict inode path for all 6931f27a5136SMike Kravetz * hugetlbfs inodes, resv_map could be NULL. 6932f27a5136SMike Kravetz */ 6933b5cec28dSMike Kravetz if (resv_map) { 6934b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 6935b5cec28dSMike Kravetz /* 6936b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 6937b5cec28dSMike Kravetz * must be split and another region descriptor can not be 6938b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 6939b5cec28dSMike Kravetz */ 6940b5cec28dSMike Kravetz if (chg < 0) 6941b5cec28dSMike Kravetz return chg; 6942b5cec28dSMike Kravetz } 6943b5cec28dSMike Kravetz 694445c682a6SKen Chen spin_lock(&inode->i_lock); 6945e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 694645c682a6SKen Chen spin_unlock(&inode->i_lock); 694745c682a6SKen Chen 69481c5ecae3SMike Kravetz /* 69491c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 69501c5ecae3SMike Kravetz * reservations to be released may be adjusted. 6951dddf31a4SMiaohe Lin * 6952dddf31a4SMiaohe Lin * Note that !resv_map implies freed == 0. So (chg - freed) 6953dddf31a4SMiaohe Lin * won't go negative. 69541c5ecae3SMike Kravetz */ 69551c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 69561c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 6957b5cec28dSMike Kravetz 6958b5cec28dSMike Kravetz return 0; 6959a43a8c39SChen, Kenneth W } 696093f70f90SNaoya Horiguchi 69613212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 69623212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 69633212b535SSteve Capper struct vm_area_struct *vma, 69643212b535SSteve Capper unsigned long addr, pgoff_t idx) 69653212b535SSteve Capper { 69663212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 69673212b535SSteve Capper svma->vm_start; 69683212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 69693212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 69703212b535SSteve Capper 69713212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 6972de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 6973de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 69743212b535SSteve Capper 69753212b535SSteve Capper /* 69763212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 69773212b535SSteve Capper * page table page. 6978131a79b4SMike Kravetz * 6979131a79b4SMike Kravetz * Also, vma_lock (vm_private_data) is required for sharing. 69803212b535SSteve Capper */ 69813212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 69823212b535SSteve Capper vm_flags != svm_flags || 6983131a79b4SMike Kravetz !range_in_vma(svma, sbase, s_end) || 6984131a79b4SMike Kravetz !svma->vm_private_data) 69853212b535SSteve Capper return 0; 69863212b535SSteve Capper 69873212b535SSteve Capper return saddr; 69883212b535SSteve Capper } 69893212b535SSteve Capper 6990bbff39ccSMike Kravetz bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 69913212b535SSteve Capper { 6992bbff39ccSMike Kravetz unsigned long start = addr & PUD_MASK; 6993bbff39ccSMike Kravetz unsigned long end = start + PUD_SIZE; 6994bbff39ccSMike Kravetz 69958d9bfb26SMike Kravetz #ifdef CONFIG_USERFAULTFD 69968d9bfb26SMike Kravetz if (uffd_disable_huge_pmd_share(vma)) 69978d9bfb26SMike Kravetz return false; 69988d9bfb26SMike Kravetz #endif 69993212b535SSteve Capper /* 70003212b535SSteve Capper * check on proper vm_flags and page table alignment 70013212b535SSteve Capper */ 70028d9bfb26SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 700331aafb45SNicholas Krause return false; 7004bbff39ccSMike Kravetz if (!vma->vm_private_data) /* vma lock required for sharing */ 70058d9bfb26SMike Kravetz return false; 70068d9bfb26SMike Kravetz if (!range_in_vma(vma, start, end)) 70078d9bfb26SMike Kravetz return false; 70088d9bfb26SMike Kravetz return true; 70098d9bfb26SMike Kravetz } 70108d9bfb26SMike Kravetz 70113212b535SSteve Capper /* 7012017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 7013017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 7014017b1660SMike Kravetz * shared pmd mappings. 7015017b1660SMike Kravetz */ 7016017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 7017017b1660SMike Kravetz unsigned long *start, unsigned long *end) 7018017b1660SMike Kravetz { 7019a1ba9da8SLi Xinhai unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), 7020a1ba9da8SLi Xinhai v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 7021017b1660SMike Kravetz 7022a1ba9da8SLi Xinhai /* 7023f0953a1bSIngo Molnar * vma needs to span at least one aligned PUD size, and the range 7024f0953a1bSIngo Molnar * must be at least partially within in. 7025a1ba9da8SLi Xinhai */ 7026a1ba9da8SLi Xinhai if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || 7027a1ba9da8SLi Xinhai (*end <= v_start) || (*start >= v_end)) 7028017b1660SMike Kravetz return; 7029017b1660SMike Kravetz 703075802ca6SPeter Xu /* Extend the range to be PUD aligned for a worst case scenario */ 7031a1ba9da8SLi Xinhai if (*start > v_start) 7032a1ba9da8SLi Xinhai *start = ALIGN_DOWN(*start, PUD_SIZE); 7033017b1660SMike Kravetz 7034a1ba9da8SLi Xinhai if (*end < v_end) 7035a1ba9da8SLi Xinhai *end = ALIGN(*end, PUD_SIZE); 7036017b1660SMike Kravetz } 7037017b1660SMike Kravetz 7038017b1660SMike Kravetz /* 70393212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 70403212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 70413212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 70423a47c54fSMike Kravetz * code much cleaner. pmd allocation is essential for the shared case because 70433a47c54fSMike Kravetz * pud has to be populated inside the same i_mmap_rwsem section - otherwise 70443a47c54fSMike Kravetz * racing tasks could either miss the sharing (see huge_pte_offset) or select a 70453a47c54fSMike Kravetz * bad pmd for sharing. 70463212b535SSteve Capper */ 7047aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 7048aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 70493212b535SSteve Capper { 70503212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 70513212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 70523212b535SSteve Capper vma->vm_pgoff; 70533212b535SSteve Capper struct vm_area_struct *svma; 70543212b535SSteve Capper unsigned long saddr; 70553212b535SSteve Capper pte_t *spte = NULL; 70563212b535SSteve Capper pte_t *pte; 7057cb900f41SKirill A. Shutemov spinlock_t *ptl; 70583212b535SSteve Capper 70593a47c54fSMike Kravetz i_mmap_lock_read(mapping); 70603212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 70613212b535SSteve Capper if (svma == vma) 70623212b535SSteve Capper continue; 70633212b535SSteve Capper 70643212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 70653212b535SSteve Capper if (saddr) { 70669c67a207SPeter Xu spte = hugetlb_walk(svma, saddr, 70677868a208SPunit Agrawal vma_mmu_pagesize(svma)); 70683212b535SSteve Capper if (spte) { 70693212b535SSteve Capper get_page(virt_to_page(spte)); 70703212b535SSteve Capper break; 70713212b535SSteve Capper } 70723212b535SSteve Capper } 70733212b535SSteve Capper } 70743212b535SSteve Capper 70753212b535SSteve Capper if (!spte) 70763212b535SSteve Capper goto out; 70773212b535SSteve Capper 70788bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 7079dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 70803212b535SSteve Capper pud_populate(mm, pud, 70813212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 7082c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 7083dc6c9a35SKirill A. Shutemov } else { 70843212b535SSteve Capper put_page(virt_to_page(spte)); 7085dc6c9a35SKirill A. Shutemov } 7086cb900f41SKirill A. Shutemov spin_unlock(ptl); 70873212b535SSteve Capper out: 70883212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 70893a47c54fSMike Kravetz i_mmap_unlock_read(mapping); 70903212b535SSteve Capper return pte; 70913212b535SSteve Capper } 70923212b535SSteve Capper 70933212b535SSteve Capper /* 70943212b535SSteve Capper * unmap huge page backed by shared pte. 70953212b535SSteve Capper * 70963212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 70973212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 70983212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 70993212b535SSteve Capper * 71003a47c54fSMike Kravetz * Called with page table lock held. 71013212b535SSteve Capper * 71023212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 71033212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 71043212b535SSteve Capper */ 710534ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 71064ddb4d91SMike Kravetz unsigned long addr, pte_t *ptep) 71073212b535SSteve Capper { 71084ddb4d91SMike Kravetz pgd_t *pgd = pgd_offset(mm, addr); 71094ddb4d91SMike Kravetz p4d_t *p4d = p4d_offset(pgd, addr); 71104ddb4d91SMike Kravetz pud_t *pud = pud_offset(p4d, addr); 71113212b535SSteve Capper 711234ae204fSMike Kravetz i_mmap_assert_write_locked(vma->vm_file->f_mapping); 711340549ba8SMike Kravetz hugetlb_vma_assert_locked(vma); 71143212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 71153212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 71163212b535SSteve Capper return 0; 71173212b535SSteve Capper 71183212b535SSteve Capper pud_clear(pud); 71193212b535SSteve Capper put_page(virt_to_page(ptep)); 7120dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 71213212b535SSteve Capper return 1; 71223212b535SSteve Capper } 7123c1991e07SPeter Xu 71249e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 71258d9bfb26SMike Kravetz 7126aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 7127aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 71289e5fc74cSSteve Capper { 71299e5fc74cSSteve Capper return NULL; 71309e5fc74cSSteve Capper } 7131e81f2d22SZhang Zhen 713234ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 71334ddb4d91SMike Kravetz unsigned long addr, pte_t *ptep) 7134e81f2d22SZhang Zhen { 7135e81f2d22SZhang Zhen return 0; 7136e81f2d22SZhang Zhen } 7137017b1660SMike Kravetz 7138017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 7139017b1660SMike Kravetz unsigned long *start, unsigned long *end) 7140017b1660SMike Kravetz { 7141017b1660SMike Kravetz } 7142c1991e07SPeter Xu 7143c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 7144c1991e07SPeter Xu { 7145c1991e07SPeter Xu return false; 7146c1991e07SPeter Xu } 71473212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 71483212b535SSteve Capper 71499e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 7150aec44e0fSPeter Xu pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, 71519e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 71529e5fc74cSSteve Capper { 71539e5fc74cSSteve Capper pgd_t *pgd; 7154c2febafcSKirill A. Shutemov p4d_t *p4d; 71559e5fc74cSSteve Capper pud_t *pud; 71569e5fc74cSSteve Capper pte_t *pte = NULL; 71579e5fc74cSSteve Capper 71589e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 7159f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 7160f4f0a3d8SKirill A. Shutemov if (!p4d) 7161f4f0a3d8SKirill A. Shutemov return NULL; 7162c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 71639e5fc74cSSteve Capper if (pud) { 71649e5fc74cSSteve Capper if (sz == PUD_SIZE) { 71659e5fc74cSSteve Capper pte = (pte_t *)pud; 71669e5fc74cSSteve Capper } else { 71679e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 7168c1991e07SPeter Xu if (want_pmd_share(vma, addr) && pud_none(*pud)) 7169aec44e0fSPeter Xu pte = huge_pmd_share(mm, vma, addr, pud); 71709e5fc74cSSteve Capper else 71719e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 71729e5fc74cSSteve Capper } 71739e5fc74cSSteve Capper } 71744e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 71759e5fc74cSSteve Capper 71769e5fc74cSSteve Capper return pte; 71779e5fc74cSSteve Capper } 71789e5fc74cSSteve Capper 71799b19df29SPunit Agrawal /* 71809b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 71819b19df29SPunit Agrawal * entry at address @addr 71829b19df29SPunit Agrawal * 71838ac0b81aSLi Xinhai * Return: Pointer to page table entry (PUD or PMD) for 71848ac0b81aSLi Xinhai * address @addr, or NULL if a !p*d_present() entry is encountered and the 71859b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 71869b19df29SPunit Agrawal * table. 71879b19df29SPunit Agrawal */ 71887868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 71897868a208SPunit Agrawal unsigned long addr, unsigned long sz) 71909e5fc74cSSteve Capper { 71919e5fc74cSSteve Capper pgd_t *pgd; 7192c2febafcSKirill A. Shutemov p4d_t *p4d; 71938ac0b81aSLi Xinhai pud_t *pud; 71948ac0b81aSLi Xinhai pmd_t *pmd; 71959e5fc74cSSteve Capper 71969e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 7197c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 7198c2febafcSKirill A. Shutemov return NULL; 7199c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 7200c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 7201c2febafcSKirill A. Shutemov return NULL; 72029b19df29SPunit Agrawal 7203c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 72048ac0b81aSLi Xinhai if (sz == PUD_SIZE) 72058ac0b81aSLi Xinhai /* must be pud huge, non-present or none */ 72069e5fc74cSSteve Capper return (pte_t *)pud; 72078ac0b81aSLi Xinhai if (!pud_present(*pud)) 72088ac0b81aSLi Xinhai return NULL; 72098ac0b81aSLi Xinhai /* must have a valid entry and size to go further */ 72109b19df29SPunit Agrawal 72119e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 72128ac0b81aSLi Xinhai /* must be pmd huge, non-present or none */ 72139e5fc74cSSteve Capper return (pte_t *)pmd; 72149e5fc74cSSteve Capper } 72159e5fc74cSSteve Capper 7216e95a9851SMike Kravetz /* 7217e95a9851SMike Kravetz * Return a mask that can be used to update an address to the last huge 7218e95a9851SMike Kravetz * page in a page table page mapping size. Used to skip non-present 7219e95a9851SMike Kravetz * page table entries when linearly scanning address ranges. Architectures 7220e95a9851SMike Kravetz * with unique huge page to page table relationships can define their own 7221e95a9851SMike Kravetz * version of this routine. 7222e95a9851SMike Kravetz */ 7223e95a9851SMike Kravetz unsigned long hugetlb_mask_last_page(struct hstate *h) 7224e95a9851SMike Kravetz { 7225e95a9851SMike Kravetz unsigned long hp_size = huge_page_size(h); 7226e95a9851SMike Kravetz 7227e95a9851SMike Kravetz if (hp_size == PUD_SIZE) 7228e95a9851SMike Kravetz return P4D_SIZE - PUD_SIZE; 7229e95a9851SMike Kravetz else if (hp_size == PMD_SIZE) 7230e95a9851SMike Kravetz return PUD_SIZE - PMD_SIZE; 7231e95a9851SMike Kravetz else 7232e95a9851SMike Kravetz return 0UL; 7233e95a9851SMike Kravetz } 7234e95a9851SMike Kravetz 7235e95a9851SMike Kravetz #else 7236e95a9851SMike Kravetz 7237e95a9851SMike Kravetz /* See description above. Architectures can provide their own version. */ 7238e95a9851SMike Kravetz __weak unsigned long hugetlb_mask_last_page(struct hstate *h) 7239e95a9851SMike Kravetz { 72404ddb4d91SMike Kravetz #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 72414ddb4d91SMike Kravetz if (huge_page_size(h) == PMD_SIZE) 72424ddb4d91SMike Kravetz return PUD_SIZE - PMD_SIZE; 72434ddb4d91SMike Kravetz #endif 7244e95a9851SMike Kravetz return 0UL; 7245e95a9851SMike Kravetz } 7246e95a9851SMike Kravetz 724761f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 724861f77edaSNaoya Horiguchi 724961f77edaSNaoya Horiguchi /* 725061f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 725161f77edaSNaoya Horiguchi * behavior. 725261f77edaSNaoya Horiguchi */ 72537ce82f4cSMiaohe Lin int isolate_hugetlb(struct page *page, struct list_head *list) 725431caf665SNaoya Horiguchi { 72557ce82f4cSMiaohe Lin int ret = 0; 7256bcc54222SNaoya Horiguchi 7257db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 72588f251a3dSMike Kravetz if (!PageHeadHuge(page) || 72598f251a3dSMike Kravetz !HPageMigratable(page) || 72600eb2df2bSMuchun Song !get_page_unless_zero(page)) { 72617ce82f4cSMiaohe Lin ret = -EBUSY; 7262bcc54222SNaoya Horiguchi goto unlock; 7263bcc54222SNaoya Horiguchi } 72648f251a3dSMike Kravetz ClearHPageMigratable(page); 726531caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 7266bcc54222SNaoya Horiguchi unlock: 7267db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7268bcc54222SNaoya Horiguchi return ret; 726931caf665SNaoya Horiguchi } 727031caf665SNaoya Horiguchi 7271e591ef7dSNaoya Horiguchi int get_hwpoison_huge_page(struct page *page, bool *hugetlb, bool unpoison) 727225182f05SNaoya Horiguchi { 727325182f05SNaoya Horiguchi int ret = 0; 727425182f05SNaoya Horiguchi 727525182f05SNaoya Horiguchi *hugetlb = false; 727625182f05SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 727725182f05SNaoya Horiguchi if (PageHeadHuge(page)) { 727825182f05SNaoya Horiguchi *hugetlb = true; 7279b283d983SNaoya Horiguchi if (HPageFreed(page)) 7280b283d983SNaoya Horiguchi ret = 0; 7281e591ef7dSNaoya Horiguchi else if (HPageMigratable(page) || unpoison) 728225182f05SNaoya Horiguchi ret = get_page_unless_zero(page); 72830ed950d1SNaoya Horiguchi else 72840ed950d1SNaoya Horiguchi ret = -EBUSY; 728525182f05SNaoya Horiguchi } 728625182f05SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 728725182f05SNaoya Horiguchi return ret; 728825182f05SNaoya Horiguchi } 728925182f05SNaoya Horiguchi 7290e591ef7dSNaoya Horiguchi int get_huge_page_for_hwpoison(unsigned long pfn, int flags, 7291e591ef7dSNaoya Horiguchi bool *migratable_cleared) 7292405ce051SNaoya Horiguchi { 7293405ce051SNaoya Horiguchi int ret; 7294405ce051SNaoya Horiguchi 7295405ce051SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 7296e591ef7dSNaoya Horiguchi ret = __get_huge_page_for_hwpoison(pfn, flags, migratable_cleared); 7297405ce051SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 7298405ce051SNaoya Horiguchi return ret; 7299405ce051SNaoya Horiguchi } 7300405ce051SNaoya Horiguchi 730131caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 730231caf665SNaoya Horiguchi { 7303db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 73048f251a3dSMike Kravetz SetHPageMigratable(page); 730531caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 7306db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 730731caf665SNaoya Horiguchi put_page(page); 730831caf665SNaoya Horiguchi } 7309ab5ac90aSMichal Hocko 7310345c62d1SSidhartha Kumar void move_hugetlb_state(struct folio *old_folio, struct folio *new_folio, int reason) 7311ab5ac90aSMichal Hocko { 7312345c62d1SSidhartha Kumar struct hstate *h = folio_hstate(old_folio); 7313ab5ac90aSMichal Hocko 7314345c62d1SSidhartha Kumar hugetlb_cgroup_migrate(old_folio, new_folio); 7315345c62d1SSidhartha Kumar set_page_owner_migrate_reason(&new_folio->page, reason); 7316ab5ac90aSMichal Hocko 7317ab5ac90aSMichal Hocko /* 7318345c62d1SSidhartha Kumar * transfer temporary state of the new hugetlb folio. This is 7319ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 7320ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 7321ab5ac90aSMichal Hocko * the temporary status. 7322ab5ac90aSMichal Hocko * 7323ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 7324ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 7325ab5ac90aSMichal Hocko * the per-node's. 7326ab5ac90aSMichal Hocko */ 7327345c62d1SSidhartha Kumar if (folio_test_hugetlb_temporary(new_folio)) { 7328345c62d1SSidhartha Kumar int old_nid = folio_nid(old_folio); 7329345c62d1SSidhartha Kumar int new_nid = folio_nid(new_folio); 7330ab5ac90aSMichal Hocko 7331345c62d1SSidhartha Kumar folio_set_hugetlb_temporary(old_folio); 7332345c62d1SSidhartha Kumar folio_clear_hugetlb_temporary(new_folio); 7333345c62d1SSidhartha Kumar 7334ab5ac90aSMichal Hocko 73355af1ab1dSMiaohe Lin /* 73365af1ab1dSMiaohe Lin * There is no need to transfer the per-node surplus state 73375af1ab1dSMiaohe Lin * when we do not cross the node. 73385af1ab1dSMiaohe Lin */ 73395af1ab1dSMiaohe Lin if (new_nid == old_nid) 73405af1ab1dSMiaohe Lin return; 7341db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 7342ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 7343ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 7344ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 7345ab5ac90aSMichal Hocko } 7346db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7347ab5ac90aSMichal Hocko } 7348ab5ac90aSMichal Hocko } 7349cf11e85fSRoman Gushchin 7350b30c14cdSJames Houghton static void hugetlb_unshare_pmds(struct vm_area_struct *vma, 7351b30c14cdSJames Houghton unsigned long start, 7352b30c14cdSJames Houghton unsigned long end) 73536dfeaff9SPeter Xu { 73546dfeaff9SPeter Xu struct hstate *h = hstate_vma(vma); 73556dfeaff9SPeter Xu unsigned long sz = huge_page_size(h); 73566dfeaff9SPeter Xu struct mm_struct *mm = vma->vm_mm; 73576dfeaff9SPeter Xu struct mmu_notifier_range range; 7358b30c14cdSJames Houghton unsigned long address; 73596dfeaff9SPeter Xu spinlock_t *ptl; 73606dfeaff9SPeter Xu pte_t *ptep; 73616dfeaff9SPeter Xu 73626dfeaff9SPeter Xu if (!(vma->vm_flags & VM_MAYSHARE)) 73636dfeaff9SPeter Xu return; 73646dfeaff9SPeter Xu 73656dfeaff9SPeter Xu if (start >= end) 73666dfeaff9SPeter Xu return; 73676dfeaff9SPeter Xu 73689c8bbfacSBaolin Wang flush_cache_range(vma, start, end); 73696dfeaff9SPeter Xu /* 73706dfeaff9SPeter Xu * No need to call adjust_range_if_pmd_sharing_possible(), because 73716dfeaff9SPeter Xu * we have already done the PUD_SIZE alignment. 73726dfeaff9SPeter Xu */ 73737d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, 73746dfeaff9SPeter Xu start, end); 73756dfeaff9SPeter Xu mmu_notifier_invalidate_range_start(&range); 737640549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 73776dfeaff9SPeter Xu i_mmap_lock_write(vma->vm_file->f_mapping); 73786dfeaff9SPeter Xu for (address = start; address < end; address += PUD_SIZE) { 73799c67a207SPeter Xu ptep = hugetlb_walk(vma, address, sz); 73806dfeaff9SPeter Xu if (!ptep) 73816dfeaff9SPeter Xu continue; 73826dfeaff9SPeter Xu ptl = huge_pte_lock(h, mm, ptep); 73834ddb4d91SMike Kravetz huge_pmd_unshare(mm, vma, address, ptep); 73846dfeaff9SPeter Xu spin_unlock(ptl); 73856dfeaff9SPeter Xu } 73866dfeaff9SPeter Xu flush_hugetlb_tlb_range(vma, start, end); 73876dfeaff9SPeter Xu i_mmap_unlock_write(vma->vm_file->f_mapping); 738840549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 73896dfeaff9SPeter Xu /* 73906dfeaff9SPeter Xu * No need to call mmu_notifier_invalidate_range(), see 7391ee65728eSMike Rapoport * Documentation/mm/mmu_notifier.rst. 73926dfeaff9SPeter Xu */ 73936dfeaff9SPeter Xu mmu_notifier_invalidate_range_end(&range); 73946dfeaff9SPeter Xu } 73956dfeaff9SPeter Xu 7396b30c14cdSJames Houghton /* 7397b30c14cdSJames Houghton * This function will unconditionally remove all the shared pmd pgtable entries 7398b30c14cdSJames Houghton * within the specific vma for a hugetlbfs memory range. 7399b30c14cdSJames Houghton */ 7400b30c14cdSJames Houghton void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) 7401b30c14cdSJames Houghton { 7402b30c14cdSJames Houghton hugetlb_unshare_pmds(vma, ALIGN(vma->vm_start, PUD_SIZE), 7403b30c14cdSJames Houghton ALIGN_DOWN(vma->vm_end, PUD_SIZE)); 7404b30c14cdSJames Houghton } 7405b30c14cdSJames Houghton 7406cf11e85fSRoman Gushchin #ifdef CONFIG_CMA 7407cf11e85fSRoman Gushchin static bool cma_reserve_called __initdata; 7408cf11e85fSRoman Gushchin 7409cf11e85fSRoman Gushchin static int __init cmdline_parse_hugetlb_cma(char *p) 7410cf11e85fSRoman Gushchin { 741138e719abSBaolin Wang int nid, count = 0; 741238e719abSBaolin Wang unsigned long tmp; 741338e719abSBaolin Wang char *s = p; 741438e719abSBaolin Wang 741538e719abSBaolin Wang while (*s) { 741638e719abSBaolin Wang if (sscanf(s, "%lu%n", &tmp, &count) != 1) 741738e719abSBaolin Wang break; 741838e719abSBaolin Wang 741938e719abSBaolin Wang if (s[count] == ':') { 7420f9317f77SMike Kravetz if (tmp >= MAX_NUMNODES) 742138e719abSBaolin Wang break; 7422f9317f77SMike Kravetz nid = array_index_nospec(tmp, MAX_NUMNODES); 742338e719abSBaolin Wang 742438e719abSBaolin Wang s += count + 1; 742538e719abSBaolin Wang tmp = memparse(s, &s); 742638e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = tmp; 742738e719abSBaolin Wang hugetlb_cma_size += tmp; 742838e719abSBaolin Wang 742938e719abSBaolin Wang /* 743038e719abSBaolin Wang * Skip the separator if have one, otherwise 743138e719abSBaolin Wang * break the parsing. 743238e719abSBaolin Wang */ 743338e719abSBaolin Wang if (*s == ',') 743438e719abSBaolin Wang s++; 743538e719abSBaolin Wang else 743638e719abSBaolin Wang break; 743738e719abSBaolin Wang } else { 7438cf11e85fSRoman Gushchin hugetlb_cma_size = memparse(p, &p); 743938e719abSBaolin Wang break; 744038e719abSBaolin Wang } 744138e719abSBaolin Wang } 744238e719abSBaolin Wang 7443cf11e85fSRoman Gushchin return 0; 7444cf11e85fSRoman Gushchin } 7445cf11e85fSRoman Gushchin 7446cf11e85fSRoman Gushchin early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); 7447cf11e85fSRoman Gushchin 7448cf11e85fSRoman Gushchin void __init hugetlb_cma_reserve(int order) 7449cf11e85fSRoman Gushchin { 7450cf11e85fSRoman Gushchin unsigned long size, reserved, per_node; 745138e719abSBaolin Wang bool node_specific_cma_alloc = false; 7452cf11e85fSRoman Gushchin int nid; 7453cf11e85fSRoman Gushchin 7454cf11e85fSRoman Gushchin cma_reserve_called = true; 7455cf11e85fSRoman Gushchin 7456cf11e85fSRoman Gushchin if (!hugetlb_cma_size) 7457cf11e85fSRoman Gushchin return; 7458cf11e85fSRoman Gushchin 745938e719abSBaolin Wang for (nid = 0; nid < MAX_NUMNODES; nid++) { 746038e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 746138e719abSBaolin Wang continue; 746238e719abSBaolin Wang 746330a51400SPeng Liu if (!node_online(nid)) { 746438e719abSBaolin Wang pr_warn("hugetlb_cma: invalid node %d specified\n", nid); 746538e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 746638e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 746738e719abSBaolin Wang continue; 746838e719abSBaolin Wang } 746938e719abSBaolin Wang 747038e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) { 747138e719abSBaolin Wang pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n", 747238e719abSBaolin Wang nid, (PAGE_SIZE << order) / SZ_1M); 747338e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 747438e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 747538e719abSBaolin Wang } else { 747638e719abSBaolin Wang node_specific_cma_alloc = true; 747738e719abSBaolin Wang } 747838e719abSBaolin Wang } 747938e719abSBaolin Wang 748038e719abSBaolin Wang /* Validate the CMA size again in case some invalid nodes specified. */ 748138e719abSBaolin Wang if (!hugetlb_cma_size) 748238e719abSBaolin Wang return; 748338e719abSBaolin Wang 7484cf11e85fSRoman Gushchin if (hugetlb_cma_size < (PAGE_SIZE << order)) { 7485cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", 7486cf11e85fSRoman Gushchin (PAGE_SIZE << order) / SZ_1M); 7487a01f4390SMike Kravetz hugetlb_cma_size = 0; 7488cf11e85fSRoman Gushchin return; 7489cf11e85fSRoman Gushchin } 7490cf11e85fSRoman Gushchin 749138e719abSBaolin Wang if (!node_specific_cma_alloc) { 7492cf11e85fSRoman Gushchin /* 7493cf11e85fSRoman Gushchin * If 3 GB area is requested on a machine with 4 numa nodes, 7494cf11e85fSRoman Gushchin * let's allocate 1 GB on first three nodes and ignore the last one. 7495cf11e85fSRoman Gushchin */ 7496cf11e85fSRoman Gushchin per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); 7497cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", 7498cf11e85fSRoman Gushchin hugetlb_cma_size / SZ_1M, per_node / SZ_1M); 749938e719abSBaolin Wang } 7500cf11e85fSRoman Gushchin 7501cf11e85fSRoman Gushchin reserved = 0; 750230a51400SPeng Liu for_each_online_node(nid) { 7503cf11e85fSRoman Gushchin int res; 75042281f797SBarry Song char name[CMA_MAX_NAME]; 7505cf11e85fSRoman Gushchin 750638e719abSBaolin Wang if (node_specific_cma_alloc) { 750738e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 750838e719abSBaolin Wang continue; 750938e719abSBaolin Wang 751038e719abSBaolin Wang size = hugetlb_cma_size_in_node[nid]; 751138e719abSBaolin Wang } else { 7512cf11e85fSRoman Gushchin size = min(per_node, hugetlb_cma_size - reserved); 751338e719abSBaolin Wang } 751438e719abSBaolin Wang 7515cf11e85fSRoman Gushchin size = round_up(size, PAGE_SIZE << order); 7516cf11e85fSRoman Gushchin 75172281f797SBarry Song snprintf(name, sizeof(name), "hugetlb%d", nid); 7518a01f4390SMike Kravetz /* 7519a01f4390SMike Kravetz * Note that 'order per bit' is based on smallest size that 7520a01f4390SMike Kravetz * may be returned to CMA allocator in the case of 7521a01f4390SMike Kravetz * huge page demotion. 7522a01f4390SMike Kravetz */ 7523a01f4390SMike Kravetz res = cma_declare_contiguous_nid(0, size, 0, 7524a01f4390SMike Kravetz PAGE_SIZE << HUGETLB_PAGE_ORDER, 752529d0f41dSBarry Song 0, false, name, 7526cf11e85fSRoman Gushchin &hugetlb_cma[nid], nid); 7527cf11e85fSRoman Gushchin if (res) { 7528cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: reservation failed: err %d, node %d", 7529cf11e85fSRoman Gushchin res, nid); 7530cf11e85fSRoman Gushchin continue; 7531cf11e85fSRoman Gushchin } 7532cf11e85fSRoman Gushchin 7533cf11e85fSRoman Gushchin reserved += size; 7534cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", 7535cf11e85fSRoman Gushchin size / SZ_1M, nid); 7536cf11e85fSRoman Gushchin 7537cf11e85fSRoman Gushchin if (reserved >= hugetlb_cma_size) 7538cf11e85fSRoman Gushchin break; 7539cf11e85fSRoman Gushchin } 7540a01f4390SMike Kravetz 7541a01f4390SMike Kravetz if (!reserved) 7542a01f4390SMike Kravetz /* 7543a01f4390SMike Kravetz * hugetlb_cma_size is used to determine if allocations from 7544a01f4390SMike Kravetz * cma are possible. Set to zero if no cma regions are set up. 7545a01f4390SMike Kravetz */ 7546a01f4390SMike Kravetz hugetlb_cma_size = 0; 7547cf11e85fSRoman Gushchin } 7548cf11e85fSRoman Gushchin 7549263b8998SMiaohe Lin static void __init hugetlb_cma_check(void) 7550cf11e85fSRoman Gushchin { 7551cf11e85fSRoman Gushchin if (!hugetlb_cma_size || cma_reserve_called) 7552cf11e85fSRoman Gushchin return; 7553cf11e85fSRoman Gushchin 7554cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); 7555cf11e85fSRoman Gushchin } 7556cf11e85fSRoman Gushchin 7557cf11e85fSRoman Gushchin #endif /* CONFIG_CMA */ 7558