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> 37af19487fSAxel Rasmussen #include <linux/mm_inline.h> 38d6606683SLinus Torvalds 3963551ae0SDavid Gibson #include <asm/page.h> 40ca15ca40SMike Rapoport #include <asm/pgalloc.h> 4124669e58SAneesh Kumar K.V #include <asm/tlb.h> 4263551ae0SDavid Gibson 4324669e58SAneesh Kumar K.V #include <linux/io.h> 4463551ae0SDavid Gibson #include <linux/hugetlb.h> 459dd540e2SAneesh Kumar K.V #include <linux/hugetlb_cgroup.h> 469a305230SLee Schermerhorn #include <linux/node.h> 47ab5ac90aSMichal Hocko #include <linux/page_owner.h> 487835e98bSNick Piggin #include "internal.h" 49f41f2ed4SMuchun Song #include "hugetlb_vmemmap.h" 501da177e4SLinus Torvalds 51c3f38a38SAneesh Kumar K.V int hugetlb_max_hstate __read_mostly; 52e5ff2159SAndi Kleen unsigned int default_hstate_idx; 53e5ff2159SAndi Kleen struct hstate hstates[HUGE_MAX_HSTATE]; 54cf11e85fSRoman Gushchin 55dbda8feaSBarry Song #ifdef CONFIG_CMA 56cf11e85fSRoman Gushchin static struct cma *hugetlb_cma[MAX_NUMNODES]; 5738e719abSBaolin Wang static unsigned long hugetlb_cma_size_in_node[MAX_NUMNODES] __initdata; 582f6c57d6SSidhartha Kumar static bool hugetlb_cma_folio(struct folio *folio, unsigned int order) 59a01f4390SMike Kravetz { 602f6c57d6SSidhartha Kumar return cma_pages_valid(hugetlb_cma[folio_nid(folio)], &folio->page, 61a01f4390SMike Kravetz 1 << order); 62a01f4390SMike Kravetz } 63a01f4390SMike Kravetz #else 642f6c57d6SSidhartha Kumar static bool hugetlb_cma_folio(struct folio *folio, unsigned int order) 65a01f4390SMike Kravetz { 66a01f4390SMike Kravetz return false; 67a01f4390SMike Kravetz } 68dbda8feaSBarry Song #endif 69dbda8feaSBarry Song static unsigned long hugetlb_cma_size __initdata; 70cf11e85fSRoman Gushchin 7153ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 7253ba51d2SJon Tollefson 73e5ff2159SAndi Kleen /* for command line parsing */ 74e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 75e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 769fee021dSVaishali Thakkar static bool __initdata parsed_valid_hugepagesz = true; 77282f4214SMike Kravetz static bool __initdata parsed_default_hugepagesz; 78b5389086SZhenguo Yao static unsigned int default_hugepages_in_node[MAX_NUMNODES] __initdata; 79e5ff2159SAndi Kleen 803935baa9SDavid Gibson /* 8131caf665SNaoya Horiguchi * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, 8231caf665SNaoya Horiguchi * free_huge_pages, and surplus_huge_pages. 833935baa9SDavid Gibson */ 84c3f38a38SAneesh Kumar K.V DEFINE_SPINLOCK(hugetlb_lock); 850bd0f9fbSEric Paris 868382d914SDavidlohr Bueso /* 878382d914SDavidlohr Bueso * Serializes faults on the same logical page. This is used to 888382d914SDavidlohr Bueso * prevent spurious OOMs when the hugepage pool is fully utilized. 898382d914SDavidlohr Bueso */ 908382d914SDavidlohr Bueso static int num_fault_mutexes; 91c672c7f2SMike Kravetz struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; 928382d914SDavidlohr Bueso 937ca02d0aSMike Kravetz /* Forward declaration */ 947ca02d0aSMike Kravetz static int hugetlb_acct_memory(struct hstate *h, long delta); 958d9bfb26SMike Kravetz static void hugetlb_vma_lock_free(struct vm_area_struct *vma); 968d9bfb26SMike Kravetz static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma); 97ecfbd733SMike Kravetz static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma); 98b30c14cdSJames Houghton static void hugetlb_unshare_pmds(struct vm_area_struct *vma, 99b30c14cdSJames Houghton unsigned long start, unsigned long end); 100bf491692SRik van Riel static struct resv_map *vma_resv_map(struct vm_area_struct *vma); 1017ca02d0aSMike Kravetz 1021d88433bSMiaohe Lin static inline bool subpool_is_free(struct hugepage_subpool *spool) 1031d88433bSMiaohe Lin { 1041d88433bSMiaohe Lin if (spool->count) 1051d88433bSMiaohe Lin return false; 1061d88433bSMiaohe Lin if (spool->max_hpages != -1) 1071d88433bSMiaohe Lin return spool->used_hpages == 0; 1081d88433bSMiaohe Lin if (spool->min_hpages != -1) 1091d88433bSMiaohe Lin return spool->rsv_hpages == spool->min_hpages; 1101d88433bSMiaohe Lin 1111d88433bSMiaohe Lin return true; 1121d88433bSMiaohe Lin } 1131d88433bSMiaohe Lin 114db71ef79SMike Kravetz static inline void unlock_or_release_subpool(struct hugepage_subpool *spool, 115db71ef79SMike Kravetz unsigned long irq_flags) 11690481622SDavid Gibson { 117db71ef79SMike Kravetz spin_unlock_irqrestore(&spool->lock, irq_flags); 11890481622SDavid Gibson 11990481622SDavid Gibson /* If no pages are used, and no other handles to the subpool 1207c8de358SEthon Paul * remain, give up any reservations based on minimum size and 1217ca02d0aSMike Kravetz * free the subpool */ 1221d88433bSMiaohe Lin if (subpool_is_free(spool)) { 1237ca02d0aSMike Kravetz if (spool->min_hpages != -1) 1247ca02d0aSMike Kravetz hugetlb_acct_memory(spool->hstate, 1257ca02d0aSMike Kravetz -spool->min_hpages); 12690481622SDavid Gibson kfree(spool); 12790481622SDavid Gibson } 1287ca02d0aSMike Kravetz } 12990481622SDavid Gibson 1307ca02d0aSMike Kravetz struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, 1317ca02d0aSMike Kravetz long min_hpages) 13290481622SDavid Gibson { 13390481622SDavid Gibson struct hugepage_subpool *spool; 13490481622SDavid Gibson 135c6a91820SMike Kravetz spool = kzalloc(sizeof(*spool), GFP_KERNEL); 13690481622SDavid Gibson if (!spool) 13790481622SDavid Gibson return NULL; 13890481622SDavid Gibson 13990481622SDavid Gibson spin_lock_init(&spool->lock); 14090481622SDavid Gibson spool->count = 1; 1417ca02d0aSMike Kravetz spool->max_hpages = max_hpages; 1427ca02d0aSMike Kravetz spool->hstate = h; 1437ca02d0aSMike Kravetz spool->min_hpages = min_hpages; 1447ca02d0aSMike Kravetz 1457ca02d0aSMike Kravetz if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { 1467ca02d0aSMike Kravetz kfree(spool); 1477ca02d0aSMike Kravetz return NULL; 1487ca02d0aSMike Kravetz } 1497ca02d0aSMike Kravetz spool->rsv_hpages = min_hpages; 15090481622SDavid Gibson 15190481622SDavid Gibson return spool; 15290481622SDavid Gibson } 15390481622SDavid Gibson 15490481622SDavid Gibson void hugepage_put_subpool(struct hugepage_subpool *spool) 15590481622SDavid Gibson { 156db71ef79SMike Kravetz unsigned long flags; 157db71ef79SMike Kravetz 158db71ef79SMike Kravetz spin_lock_irqsave(&spool->lock, flags); 15990481622SDavid Gibson BUG_ON(!spool->count); 16090481622SDavid Gibson spool->count--; 161db71ef79SMike Kravetz unlock_or_release_subpool(spool, flags); 16290481622SDavid Gibson } 16390481622SDavid Gibson 1641c5ecae3SMike Kravetz /* 1651c5ecae3SMike Kravetz * Subpool accounting for allocating and reserving pages. 1661c5ecae3SMike Kravetz * Return -ENOMEM if there are not enough resources to satisfy the 1679e7ee400SRandy Dunlap * request. Otherwise, return the number of pages by which the 1681c5ecae3SMike Kravetz * global pools must be adjusted (upward). The returned value may 1691c5ecae3SMike Kravetz * only be different than the passed value (delta) in the case where 1707c8de358SEthon Paul * a subpool minimum size must be maintained. 1711c5ecae3SMike Kravetz */ 1721c5ecae3SMike Kravetz static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, 17390481622SDavid Gibson long delta) 17490481622SDavid Gibson { 1751c5ecae3SMike Kravetz long ret = delta; 17690481622SDavid Gibson 17790481622SDavid Gibson if (!spool) 1781c5ecae3SMike Kravetz return ret; 17990481622SDavid Gibson 180db71ef79SMike Kravetz spin_lock_irq(&spool->lock); 18190481622SDavid Gibson 1821c5ecae3SMike Kravetz if (spool->max_hpages != -1) { /* maximum size accounting */ 1831c5ecae3SMike Kravetz if ((spool->used_hpages + delta) <= spool->max_hpages) 1841c5ecae3SMike Kravetz spool->used_hpages += delta; 1851c5ecae3SMike Kravetz else { 1861c5ecae3SMike Kravetz ret = -ENOMEM; 1871c5ecae3SMike Kravetz goto unlock_ret; 1881c5ecae3SMike Kravetz } 1891c5ecae3SMike Kravetz } 1901c5ecae3SMike Kravetz 19109a95e29SMike Kravetz /* minimum size accounting */ 19209a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->rsv_hpages) { 1931c5ecae3SMike Kravetz if (delta > spool->rsv_hpages) { 1941c5ecae3SMike Kravetz /* 1951c5ecae3SMike Kravetz * Asking for more reserves than those already taken on 1961c5ecae3SMike Kravetz * behalf of subpool. Return difference. 1971c5ecae3SMike Kravetz */ 1981c5ecae3SMike Kravetz ret = delta - spool->rsv_hpages; 1991c5ecae3SMike Kravetz spool->rsv_hpages = 0; 2001c5ecae3SMike Kravetz } else { 2011c5ecae3SMike Kravetz ret = 0; /* reserves already accounted for */ 2021c5ecae3SMike Kravetz spool->rsv_hpages -= delta; 2031c5ecae3SMike Kravetz } 2041c5ecae3SMike Kravetz } 2051c5ecae3SMike Kravetz 2061c5ecae3SMike Kravetz unlock_ret: 207db71ef79SMike Kravetz spin_unlock_irq(&spool->lock); 20890481622SDavid Gibson return ret; 20990481622SDavid Gibson } 21090481622SDavid Gibson 2111c5ecae3SMike Kravetz /* 2121c5ecae3SMike Kravetz * Subpool accounting for freeing and unreserving pages. 2131c5ecae3SMike Kravetz * Return the number of global page reservations that must be dropped. 2141c5ecae3SMike Kravetz * The return value may only be different than the passed value (delta) 2151c5ecae3SMike Kravetz * in the case where a subpool minimum size must be maintained. 2161c5ecae3SMike Kravetz */ 2171c5ecae3SMike Kravetz static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, 21890481622SDavid Gibson long delta) 21990481622SDavid Gibson { 2201c5ecae3SMike Kravetz long ret = delta; 221db71ef79SMike Kravetz unsigned long flags; 2221c5ecae3SMike Kravetz 22390481622SDavid Gibson if (!spool) 2241c5ecae3SMike Kravetz return delta; 22590481622SDavid Gibson 226db71ef79SMike Kravetz spin_lock_irqsave(&spool->lock, flags); 2271c5ecae3SMike Kravetz 2281c5ecae3SMike Kravetz if (spool->max_hpages != -1) /* maximum size accounting */ 22990481622SDavid Gibson spool->used_hpages -= delta; 2301c5ecae3SMike Kravetz 23109a95e29SMike Kravetz /* minimum size accounting */ 23209a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { 2331c5ecae3SMike Kravetz if (spool->rsv_hpages + delta <= spool->min_hpages) 2341c5ecae3SMike Kravetz ret = 0; 2351c5ecae3SMike Kravetz else 2361c5ecae3SMike Kravetz ret = spool->rsv_hpages + delta - spool->min_hpages; 2371c5ecae3SMike Kravetz 2381c5ecae3SMike Kravetz spool->rsv_hpages += delta; 2391c5ecae3SMike Kravetz if (spool->rsv_hpages > spool->min_hpages) 2401c5ecae3SMike Kravetz spool->rsv_hpages = spool->min_hpages; 2411c5ecae3SMike Kravetz } 2421c5ecae3SMike Kravetz 2431c5ecae3SMike Kravetz /* 2441c5ecae3SMike Kravetz * If hugetlbfs_put_super couldn't free spool due to an outstanding 2451c5ecae3SMike Kravetz * quota reference, free it now. 2461c5ecae3SMike Kravetz */ 247db71ef79SMike Kravetz unlock_or_release_subpool(spool, flags); 2481c5ecae3SMike Kravetz 2491c5ecae3SMike Kravetz return ret; 25090481622SDavid Gibson } 25190481622SDavid Gibson 25290481622SDavid Gibson static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 25390481622SDavid Gibson { 25490481622SDavid Gibson return HUGETLBFS_SB(inode->i_sb)->spool; 25590481622SDavid Gibson } 25690481622SDavid Gibson 25790481622SDavid Gibson static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 25890481622SDavid Gibson { 259496ad9aaSAl Viro return subpool_inode(file_inode(vma->vm_file)); 26090481622SDavid Gibson } 26190481622SDavid Gibson 262e700898fSMike Kravetz /* 263e700898fSMike Kravetz * hugetlb vma_lock helper routines 264e700898fSMike Kravetz */ 265e700898fSMike Kravetz void hugetlb_vma_lock_read(struct vm_area_struct *vma) 266e700898fSMike Kravetz { 267e700898fSMike Kravetz if (__vma_shareable_lock(vma)) { 268e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 269e700898fSMike Kravetz 270e700898fSMike Kravetz down_read(&vma_lock->rw_sema); 271bf491692SRik van Riel } else if (__vma_private_lock(vma)) { 272bf491692SRik van Riel struct resv_map *resv_map = vma_resv_map(vma); 273bf491692SRik van Riel 274bf491692SRik van Riel down_read(&resv_map->rw_sema); 275e700898fSMike Kravetz } 276e700898fSMike Kravetz } 277e700898fSMike Kravetz 278e700898fSMike Kravetz void hugetlb_vma_unlock_read(struct vm_area_struct *vma) 279e700898fSMike Kravetz { 280e700898fSMike Kravetz if (__vma_shareable_lock(vma)) { 281e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 282e700898fSMike Kravetz 283e700898fSMike Kravetz up_read(&vma_lock->rw_sema); 284bf491692SRik van Riel } else if (__vma_private_lock(vma)) { 285bf491692SRik van Riel struct resv_map *resv_map = vma_resv_map(vma); 286bf491692SRik van Riel 287bf491692SRik van Riel up_read(&resv_map->rw_sema); 288e700898fSMike Kravetz } 289e700898fSMike Kravetz } 290e700898fSMike Kravetz 291e700898fSMike Kravetz void hugetlb_vma_lock_write(struct vm_area_struct *vma) 292e700898fSMike Kravetz { 293e700898fSMike Kravetz if (__vma_shareable_lock(vma)) { 294e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 295e700898fSMike Kravetz 296e700898fSMike Kravetz down_write(&vma_lock->rw_sema); 297bf491692SRik van Riel } else if (__vma_private_lock(vma)) { 298bf491692SRik van Riel struct resv_map *resv_map = vma_resv_map(vma); 299bf491692SRik van Riel 300bf491692SRik van Riel down_write(&resv_map->rw_sema); 301e700898fSMike Kravetz } 302e700898fSMike Kravetz } 303e700898fSMike Kravetz 304e700898fSMike Kravetz void hugetlb_vma_unlock_write(struct vm_area_struct *vma) 305e700898fSMike Kravetz { 306e700898fSMike Kravetz if (__vma_shareable_lock(vma)) { 307e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 308e700898fSMike Kravetz 309e700898fSMike Kravetz up_write(&vma_lock->rw_sema); 310bf491692SRik van Riel } else if (__vma_private_lock(vma)) { 311bf491692SRik van Riel struct resv_map *resv_map = vma_resv_map(vma); 312bf491692SRik van Riel 313bf491692SRik van Riel up_write(&resv_map->rw_sema); 314e700898fSMike Kravetz } 315e700898fSMike Kravetz } 316e700898fSMike Kravetz 317e700898fSMike Kravetz int hugetlb_vma_trylock_write(struct vm_area_struct *vma) 318e700898fSMike Kravetz { 319bf491692SRik van Riel 320bf491692SRik van Riel if (__vma_shareable_lock(vma)) { 321e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 322e700898fSMike Kravetz 323e700898fSMike Kravetz return down_write_trylock(&vma_lock->rw_sema); 324bf491692SRik van Riel } else if (__vma_private_lock(vma)) { 325bf491692SRik van Riel struct resv_map *resv_map = vma_resv_map(vma); 326bf491692SRik van Riel 327bf491692SRik van Riel return down_write_trylock(&resv_map->rw_sema); 328bf491692SRik van Riel } 329bf491692SRik van Riel 330bf491692SRik van Riel return 1; 331e700898fSMike Kravetz } 332e700898fSMike Kravetz 333e700898fSMike Kravetz void hugetlb_vma_assert_locked(struct vm_area_struct *vma) 334e700898fSMike Kravetz { 335e700898fSMike Kravetz if (__vma_shareable_lock(vma)) { 336e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 337e700898fSMike Kravetz 338e700898fSMike Kravetz lockdep_assert_held(&vma_lock->rw_sema); 339bf491692SRik van Riel } else if (__vma_private_lock(vma)) { 340bf491692SRik van Riel struct resv_map *resv_map = vma_resv_map(vma); 341bf491692SRik van Riel 342bf491692SRik van Riel lockdep_assert_held(&resv_map->rw_sema); 343e700898fSMike Kravetz } 344e700898fSMike Kravetz } 345e700898fSMike Kravetz 346e700898fSMike Kravetz void hugetlb_vma_lock_release(struct kref *kref) 347e700898fSMike Kravetz { 348e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = container_of(kref, 349e700898fSMike Kravetz struct hugetlb_vma_lock, refs); 350e700898fSMike Kravetz 351e700898fSMike Kravetz kfree(vma_lock); 352e700898fSMike Kravetz } 353e700898fSMike Kravetz 354e700898fSMike Kravetz static void __hugetlb_vma_unlock_write_put(struct hugetlb_vma_lock *vma_lock) 355e700898fSMike Kravetz { 356e700898fSMike Kravetz struct vm_area_struct *vma = vma_lock->vma; 357e700898fSMike Kravetz 358e700898fSMike Kravetz /* 359e700898fSMike Kravetz * vma_lock structure may or not be released as a result of put, 360e700898fSMike Kravetz * it certainly will no longer be attached to vma so clear pointer. 361e700898fSMike Kravetz * Semaphore synchronizes access to vma_lock->vma field. 362e700898fSMike Kravetz */ 363e700898fSMike Kravetz vma_lock->vma = NULL; 364e700898fSMike Kravetz vma->vm_private_data = NULL; 365e700898fSMike Kravetz up_write(&vma_lock->rw_sema); 366e700898fSMike Kravetz kref_put(&vma_lock->refs, hugetlb_vma_lock_release); 367e700898fSMike Kravetz } 368e700898fSMike Kravetz 369e700898fSMike Kravetz static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma) 370e700898fSMike Kravetz { 371e700898fSMike Kravetz if (__vma_shareable_lock(vma)) { 372e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 373e700898fSMike Kravetz 374e700898fSMike Kravetz __hugetlb_vma_unlock_write_put(vma_lock); 375bf491692SRik van Riel } else if (__vma_private_lock(vma)) { 376bf491692SRik van Riel struct resv_map *resv_map = vma_resv_map(vma); 377bf491692SRik van Riel 378bf491692SRik van Riel /* no free for anon vmas, but still need to unlock */ 379bf491692SRik van Riel up_write(&resv_map->rw_sema); 380e700898fSMike Kravetz } 381e700898fSMike Kravetz } 382e700898fSMike Kravetz 383e700898fSMike Kravetz static void hugetlb_vma_lock_free(struct vm_area_struct *vma) 384e700898fSMike Kravetz { 385e700898fSMike Kravetz /* 386e700898fSMike Kravetz * Only present in sharable vmas. 387e700898fSMike Kravetz */ 388e700898fSMike Kravetz if (!vma || !__vma_shareable_lock(vma)) 389e700898fSMike Kravetz return; 390e700898fSMike Kravetz 391e700898fSMike Kravetz if (vma->vm_private_data) { 392e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 393e700898fSMike Kravetz 394e700898fSMike Kravetz down_write(&vma_lock->rw_sema); 395e700898fSMike Kravetz __hugetlb_vma_unlock_write_put(vma_lock); 396e700898fSMike Kravetz } 397e700898fSMike Kravetz } 398e700898fSMike Kravetz 399e700898fSMike Kravetz static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma) 400e700898fSMike Kravetz { 401e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock; 402e700898fSMike Kravetz 403e700898fSMike Kravetz /* Only establish in (flags) sharable vmas */ 404e700898fSMike Kravetz if (!vma || !(vma->vm_flags & VM_MAYSHARE)) 405e700898fSMike Kravetz return; 406e700898fSMike Kravetz 407e700898fSMike Kravetz /* Should never get here with non-NULL vm_private_data */ 408e700898fSMike Kravetz if (vma->vm_private_data) 409e700898fSMike Kravetz return; 410e700898fSMike Kravetz 411e700898fSMike Kravetz vma_lock = kmalloc(sizeof(*vma_lock), GFP_KERNEL); 412e700898fSMike Kravetz if (!vma_lock) { 413e700898fSMike Kravetz /* 414e700898fSMike Kravetz * If we can not allocate structure, then vma can not 415e700898fSMike Kravetz * participate in pmd sharing. This is only a possible 416e700898fSMike Kravetz * performance enhancement and memory saving issue. 417e700898fSMike Kravetz * However, the lock is also used to synchronize page 418e700898fSMike Kravetz * faults with truncation. If the lock is not present, 419e700898fSMike Kravetz * unlikely races could leave pages in a file past i_size 420e700898fSMike Kravetz * until the file is removed. Warn in the unlikely case of 421e700898fSMike Kravetz * allocation failure. 422e700898fSMike Kravetz */ 423e700898fSMike Kravetz pr_warn_once("HugeTLB: unable to allocate vma specific lock\n"); 424e700898fSMike Kravetz return; 425e700898fSMike Kravetz } 426e700898fSMike Kravetz 427e700898fSMike Kravetz kref_init(&vma_lock->refs); 428e700898fSMike Kravetz init_rwsem(&vma_lock->rw_sema); 429e700898fSMike Kravetz vma_lock->vma = vma; 430e700898fSMike Kravetz vma->vm_private_data = vma_lock; 431e700898fSMike Kravetz } 432e700898fSMike Kravetz 4330db9d74eSMina Almasry /* Helper that removes a struct file_region from the resv_map cache and returns 4340db9d74eSMina Almasry * it for use. 4350db9d74eSMina Almasry */ 4360db9d74eSMina Almasry static struct file_region * 4370db9d74eSMina Almasry get_file_region_entry_from_cache(struct resv_map *resv, long from, long to) 4380db9d74eSMina Almasry { 4393259914fSXU pengfei struct file_region *nrg; 4400db9d74eSMina Almasry 4410db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count <= 0); 4420db9d74eSMina Almasry 4430db9d74eSMina Almasry resv->region_cache_count--; 4440db9d74eSMina Almasry nrg = list_first_entry(&resv->region_cache, struct file_region, link); 4450db9d74eSMina Almasry list_del(&nrg->link); 4460db9d74eSMina Almasry 4470db9d74eSMina Almasry nrg->from = from; 4480db9d74eSMina Almasry nrg->to = to; 4490db9d74eSMina Almasry 4500db9d74eSMina Almasry return nrg; 4510db9d74eSMina Almasry } 4520db9d74eSMina Almasry 453075a61d0SMina Almasry static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, 454075a61d0SMina Almasry struct file_region *rg) 455075a61d0SMina Almasry { 456075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 457075a61d0SMina Almasry nrg->reservation_counter = rg->reservation_counter; 458075a61d0SMina Almasry nrg->css = rg->css; 459075a61d0SMina Almasry if (rg->css) 460075a61d0SMina Almasry css_get(rg->css); 461075a61d0SMina Almasry #endif 462075a61d0SMina Almasry } 463075a61d0SMina Almasry 464075a61d0SMina Almasry /* Helper that records hugetlb_cgroup uncharge info. */ 465075a61d0SMina Almasry static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, 466075a61d0SMina Almasry struct hstate *h, 467075a61d0SMina Almasry struct resv_map *resv, 468075a61d0SMina Almasry struct file_region *nrg) 469075a61d0SMina Almasry { 470075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 471075a61d0SMina Almasry if (h_cg) { 472075a61d0SMina Almasry nrg->reservation_counter = 473075a61d0SMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 474075a61d0SMina Almasry nrg->css = &h_cg->css; 475d85aecf2SMiaohe Lin /* 476d85aecf2SMiaohe Lin * The caller will hold exactly one h_cg->css reference for the 477d85aecf2SMiaohe Lin * whole contiguous reservation region. But this area might be 478d85aecf2SMiaohe Lin * scattered when there are already some file_regions reside in 479d85aecf2SMiaohe Lin * it. As a result, many file_regions may share only one css 480d85aecf2SMiaohe Lin * reference. In order to ensure that one file_region must hold 481d85aecf2SMiaohe Lin * exactly one h_cg->css reference, we should do css_get for 482d85aecf2SMiaohe Lin * each file_region and leave the reference held by caller 483d85aecf2SMiaohe Lin * untouched. 484d85aecf2SMiaohe Lin */ 485d85aecf2SMiaohe Lin css_get(&h_cg->css); 486075a61d0SMina Almasry if (!resv->pages_per_hpage) 487075a61d0SMina Almasry resv->pages_per_hpage = pages_per_huge_page(h); 488075a61d0SMina Almasry /* pages_per_hpage should be the same for all entries in 489075a61d0SMina Almasry * a resv_map. 490075a61d0SMina Almasry */ 491075a61d0SMina Almasry VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); 492075a61d0SMina Almasry } else { 493075a61d0SMina Almasry nrg->reservation_counter = NULL; 494075a61d0SMina Almasry nrg->css = NULL; 495075a61d0SMina Almasry } 496075a61d0SMina Almasry #endif 497075a61d0SMina Almasry } 498075a61d0SMina Almasry 499d85aecf2SMiaohe Lin static void put_uncharge_info(struct file_region *rg) 500d85aecf2SMiaohe Lin { 501d85aecf2SMiaohe Lin #ifdef CONFIG_CGROUP_HUGETLB 502d85aecf2SMiaohe Lin if (rg->css) 503d85aecf2SMiaohe Lin css_put(rg->css); 504d85aecf2SMiaohe Lin #endif 505d85aecf2SMiaohe Lin } 506d85aecf2SMiaohe Lin 507a9b3f867SMina Almasry static bool has_same_uncharge_info(struct file_region *rg, 508a9b3f867SMina Almasry struct file_region *org) 509a9b3f867SMina Almasry { 510a9b3f867SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 5110739eb43SBaolin Wang return rg->reservation_counter == org->reservation_counter && 512a9b3f867SMina Almasry rg->css == org->css; 513a9b3f867SMina Almasry 514a9b3f867SMina Almasry #else 515a9b3f867SMina Almasry return true; 516a9b3f867SMina Almasry #endif 517a9b3f867SMina Almasry } 518a9b3f867SMina Almasry 519a9b3f867SMina Almasry static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) 520a9b3f867SMina Almasry { 5213259914fSXU pengfei struct file_region *nrg, *prg; 522a9b3f867SMina Almasry 523a9b3f867SMina Almasry prg = list_prev_entry(rg, link); 524a9b3f867SMina Almasry if (&prg->link != &resv->regions && prg->to == rg->from && 525a9b3f867SMina Almasry has_same_uncharge_info(prg, rg)) { 526a9b3f867SMina Almasry prg->to = rg->to; 527a9b3f867SMina Almasry 528a9b3f867SMina Almasry list_del(&rg->link); 529d85aecf2SMiaohe Lin put_uncharge_info(rg); 530a9b3f867SMina Almasry kfree(rg); 531a9b3f867SMina Almasry 5327db5e7b6SWei Yang rg = prg; 533a9b3f867SMina Almasry } 534a9b3f867SMina Almasry 535a9b3f867SMina Almasry nrg = list_next_entry(rg, link); 536a9b3f867SMina Almasry if (&nrg->link != &resv->regions && nrg->from == rg->to && 537a9b3f867SMina Almasry has_same_uncharge_info(nrg, rg)) { 538a9b3f867SMina Almasry nrg->from = rg->from; 539a9b3f867SMina Almasry 540a9b3f867SMina Almasry list_del(&rg->link); 541d85aecf2SMiaohe Lin put_uncharge_info(rg); 542a9b3f867SMina Almasry kfree(rg); 543a9b3f867SMina Almasry } 544a9b3f867SMina Almasry } 545a9b3f867SMina Almasry 5462103cf9cSPeter Xu static inline long 54784448c8eSJakob Koschel hugetlb_resv_map_add(struct resv_map *map, struct list_head *rg, long from, 5482103cf9cSPeter Xu long to, struct hstate *h, struct hugetlb_cgroup *cg, 5492103cf9cSPeter Xu long *regions_needed) 5502103cf9cSPeter Xu { 5512103cf9cSPeter Xu struct file_region *nrg; 5522103cf9cSPeter Xu 5532103cf9cSPeter Xu if (!regions_needed) { 5542103cf9cSPeter Xu nrg = get_file_region_entry_from_cache(map, from, to); 5552103cf9cSPeter Xu record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg); 55684448c8eSJakob Koschel list_add(&nrg->link, rg); 5572103cf9cSPeter Xu coalesce_file_region(map, nrg); 5582103cf9cSPeter Xu } else 5592103cf9cSPeter Xu *regions_needed += 1; 5602103cf9cSPeter Xu 5612103cf9cSPeter Xu return to - from; 5622103cf9cSPeter Xu } 5632103cf9cSPeter Xu 564972a3da3SWei Yang /* 565972a3da3SWei Yang * Must be called with resv->lock held. 566972a3da3SWei Yang * 567972a3da3SWei Yang * Calling this with regions_needed != NULL will count the number of pages 568972a3da3SWei Yang * to be added but will not modify the linked list. And regions_needed will 569972a3da3SWei Yang * indicate the number of file_regions needed in the cache to carry out to add 570972a3da3SWei Yang * the regions for this range. 571d75c6af9SMina Almasry */ 572d75c6af9SMina Almasry static long add_reservation_in_range(struct resv_map *resv, long f, long t, 573075a61d0SMina Almasry struct hugetlb_cgroup *h_cg, 574972a3da3SWei Yang struct hstate *h, long *regions_needed) 575d75c6af9SMina Almasry { 5760db9d74eSMina Almasry long add = 0; 577d75c6af9SMina Almasry struct list_head *head = &resv->regions; 5780db9d74eSMina Almasry long last_accounted_offset = f; 57984448c8eSJakob Koschel struct file_region *iter, *trg = NULL; 58084448c8eSJakob Koschel struct list_head *rg = NULL; 581d75c6af9SMina Almasry 5820db9d74eSMina Almasry if (regions_needed) 5830db9d74eSMina Almasry *regions_needed = 0; 584d75c6af9SMina Almasry 5850db9d74eSMina Almasry /* In this loop, we essentially handle an entry for the range 58684448c8eSJakob Koschel * [last_accounted_offset, iter->from), at every iteration, with some 5870db9d74eSMina Almasry * bounds checking. 5880db9d74eSMina Almasry */ 58984448c8eSJakob Koschel list_for_each_entry_safe(iter, trg, head, link) { 5900db9d74eSMina Almasry /* Skip irrelevant regions that start before our range. */ 59184448c8eSJakob Koschel if (iter->from < f) { 5920db9d74eSMina Almasry /* If this region ends after the last accounted offset, 5930db9d74eSMina Almasry * then we need to update last_accounted_offset. 5940db9d74eSMina Almasry */ 59584448c8eSJakob Koschel if (iter->to > last_accounted_offset) 59684448c8eSJakob Koschel last_accounted_offset = iter->to; 5970db9d74eSMina Almasry continue; 5980db9d74eSMina Almasry } 599d75c6af9SMina Almasry 6000db9d74eSMina Almasry /* When we find a region that starts beyond our range, we've 6010db9d74eSMina Almasry * finished. 6020db9d74eSMina Almasry */ 60384448c8eSJakob Koschel if (iter->from >= t) { 60484448c8eSJakob Koschel rg = iter->link.prev; 605d75c6af9SMina Almasry break; 60684448c8eSJakob Koschel } 607d75c6af9SMina Almasry 60884448c8eSJakob Koschel /* Add an entry for last_accounted_offset -> iter->from, and 6090db9d74eSMina Almasry * update last_accounted_offset. 610d75c6af9SMina Almasry */ 61184448c8eSJakob Koschel if (iter->from > last_accounted_offset) 61284448c8eSJakob Koschel add += hugetlb_resv_map_add(resv, iter->link.prev, 6132103cf9cSPeter Xu last_accounted_offset, 61484448c8eSJakob Koschel iter->from, h, h_cg, 6152103cf9cSPeter Xu regions_needed); 616d75c6af9SMina Almasry 61784448c8eSJakob Koschel last_accounted_offset = iter->to; 6180db9d74eSMina Almasry } 6190db9d74eSMina Almasry 6200db9d74eSMina Almasry /* Handle the case where our range extends beyond 6210db9d74eSMina Almasry * last_accounted_offset. 6220db9d74eSMina Almasry */ 62384448c8eSJakob Koschel if (!rg) 62484448c8eSJakob Koschel rg = head->prev; 6252103cf9cSPeter Xu if (last_accounted_offset < t) 6262103cf9cSPeter Xu add += hugetlb_resv_map_add(resv, rg, last_accounted_offset, 6272103cf9cSPeter Xu t, h, h_cg, regions_needed); 6280db9d74eSMina Almasry 6290db9d74eSMina Almasry return add; 6300db9d74eSMina Almasry } 6310db9d74eSMina Almasry 6320db9d74eSMina Almasry /* Must be called with resv->lock acquired. Will drop lock to allocate entries. 6330db9d74eSMina Almasry */ 6340db9d74eSMina Almasry static int allocate_file_region_entries(struct resv_map *resv, 6350db9d74eSMina Almasry int regions_needed) 6360db9d74eSMina Almasry __must_hold(&resv->lock) 6370db9d74eSMina Almasry { 63834665341SMiaohe Lin LIST_HEAD(allocated_regions); 6390db9d74eSMina Almasry int to_allocate = 0, i = 0; 6400db9d74eSMina Almasry struct file_region *trg = NULL, *rg = NULL; 6410db9d74eSMina Almasry 6420db9d74eSMina Almasry VM_BUG_ON(regions_needed < 0); 6430db9d74eSMina Almasry 6440db9d74eSMina Almasry /* 6450db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 6460db9d74eSMina Almasry * the number of in progress add operations plus regions_needed. 6470db9d74eSMina Almasry * 6480db9d74eSMina Almasry * This is a while loop because when we drop the lock, some other call 6490db9d74eSMina Almasry * to region_add or region_del may have consumed some region_entries, 6500db9d74eSMina Almasry * so we keep looping here until we finally have enough entries for 6510db9d74eSMina Almasry * (adds_in_progress + regions_needed). 6520db9d74eSMina Almasry */ 6530db9d74eSMina Almasry while (resv->region_cache_count < 6540db9d74eSMina Almasry (resv->adds_in_progress + regions_needed)) { 6550db9d74eSMina Almasry to_allocate = resv->adds_in_progress + regions_needed - 6560db9d74eSMina Almasry resv->region_cache_count; 6570db9d74eSMina Almasry 6580db9d74eSMina Almasry /* At this point, we should have enough entries in the cache 659f0953a1bSIngo Molnar * for all the existing adds_in_progress. We should only be 6600db9d74eSMina Almasry * needing to allocate for regions_needed. 6610db9d74eSMina Almasry */ 6620db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); 6630db9d74eSMina Almasry 6640db9d74eSMina Almasry spin_unlock(&resv->lock); 6650db9d74eSMina Almasry for (i = 0; i < to_allocate; i++) { 6660db9d74eSMina Almasry trg = kmalloc(sizeof(*trg), GFP_KERNEL); 6670db9d74eSMina Almasry if (!trg) 6680db9d74eSMina Almasry goto out_of_memory; 6690db9d74eSMina Almasry list_add(&trg->link, &allocated_regions); 6700db9d74eSMina Almasry } 6710db9d74eSMina Almasry 6720db9d74eSMina Almasry spin_lock(&resv->lock); 6730db9d74eSMina Almasry 674d3ec7b6eSWei Yang list_splice(&allocated_regions, &resv->region_cache); 675d3ec7b6eSWei Yang resv->region_cache_count += to_allocate; 6760db9d74eSMina Almasry } 6770db9d74eSMina Almasry 6780db9d74eSMina Almasry return 0; 6790db9d74eSMina Almasry 6800db9d74eSMina Almasry out_of_memory: 6810db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, &allocated_regions, link) { 682d75c6af9SMina Almasry list_del(&rg->link); 683d75c6af9SMina Almasry kfree(rg); 684d75c6af9SMina Almasry } 6850db9d74eSMina Almasry return -ENOMEM; 686d75c6af9SMina Almasry } 687d75c6af9SMina Almasry 6881dd308a7SMike Kravetz /* 6891dd308a7SMike Kravetz * Add the huge page range represented by [f, t) to the reserve 6900db9d74eSMina Almasry * map. Regions will be taken from the cache to fill in this range. 6910db9d74eSMina Almasry * Sufficient regions should exist in the cache due to the previous 6920db9d74eSMina Almasry * call to region_chg with the same range, but in some cases the cache will not 6930db9d74eSMina Almasry * have sufficient entries due to races with other code doing region_add or 6940db9d74eSMina Almasry * region_del. The extra needed entries will be allocated. 695cf3ad20bSMike Kravetz * 6960db9d74eSMina Almasry * regions_needed is the out value provided by a previous call to region_chg. 6970db9d74eSMina Almasry * 6980db9d74eSMina Almasry * Return the number of new huge pages added to the map. This number is greater 6990db9d74eSMina Almasry * than or equal to zero. If file_region entries needed to be allocated for 7007c8de358SEthon Paul * this operation and we were not able to allocate, it returns -ENOMEM. 7010db9d74eSMina Almasry * region_add of regions of length 1 never allocate file_regions and cannot 7020db9d74eSMina Almasry * fail; region_chg will always allocate at least 1 entry and a region_add for 7030db9d74eSMina Almasry * 1 page will only require at most 1 entry. 7041dd308a7SMike Kravetz */ 7050db9d74eSMina Almasry static long region_add(struct resv_map *resv, long f, long t, 706075a61d0SMina Almasry long in_regions_needed, struct hstate *h, 707075a61d0SMina Almasry struct hugetlb_cgroup *h_cg) 70896822904SAndy Whitcroft { 7090db9d74eSMina Almasry long add = 0, actual_regions_needed = 0; 71096822904SAndy Whitcroft 7117b24d861SDavidlohr Bueso spin_lock(&resv->lock); 7120db9d74eSMina Almasry retry: 7130db9d74eSMina Almasry 7140db9d74eSMina Almasry /* Count how many regions are actually needed to execute this add. */ 715972a3da3SWei Yang add_reservation_in_range(resv, f, t, NULL, NULL, 716972a3da3SWei Yang &actual_regions_needed); 71796822904SAndy Whitcroft 7185e911373SMike Kravetz /* 7190db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 7200db9d74eSMina Almasry * this add operation. Note that actual_regions_needed may be greater 7210db9d74eSMina Almasry * than in_regions_needed, as the resv_map may have been modified since 7220db9d74eSMina Almasry * the region_chg call. In this case, we need to make sure that we 7230db9d74eSMina Almasry * allocate extra entries, such that we have enough for all the 7240db9d74eSMina Almasry * existing adds_in_progress, plus the excess needed for this 7250db9d74eSMina Almasry * operation. 7265e911373SMike Kravetz */ 7270db9d74eSMina Almasry if (actual_regions_needed > in_regions_needed && 7280db9d74eSMina Almasry resv->region_cache_count < 7290db9d74eSMina Almasry resv->adds_in_progress + 7300db9d74eSMina Almasry (actual_regions_needed - in_regions_needed)) { 7310db9d74eSMina Almasry /* region_add operation of range 1 should never need to 7320db9d74eSMina Almasry * allocate file_region entries. 7330db9d74eSMina Almasry */ 7340db9d74eSMina Almasry VM_BUG_ON(t - f <= 1); 7355e911373SMike Kravetz 7360db9d74eSMina Almasry if (allocate_file_region_entries( 7370db9d74eSMina Almasry resv, actual_regions_needed - in_regions_needed)) { 7380db9d74eSMina Almasry return -ENOMEM; 7395e911373SMike Kravetz } 7405e911373SMike Kravetz 7410db9d74eSMina Almasry goto retry; 7420db9d74eSMina Almasry } 743cf3ad20bSMike Kravetz 744972a3da3SWei Yang add = add_reservation_in_range(resv, f, t, h_cg, h, NULL); 7450db9d74eSMina Almasry 7460db9d74eSMina Almasry resv->adds_in_progress -= in_regions_needed; 7470db9d74eSMina Almasry 7487b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 749cf3ad20bSMike Kravetz return add; 75096822904SAndy Whitcroft } 75196822904SAndy Whitcroft 7521dd308a7SMike Kravetz /* 7531dd308a7SMike Kravetz * Examine the existing reserve map and determine how many 7541dd308a7SMike Kravetz * huge pages in the specified range [f, t) are NOT currently 7551dd308a7SMike Kravetz * represented. This routine is called before a subsequent 7561dd308a7SMike Kravetz * call to region_add that will actually modify the reserve 7571dd308a7SMike Kravetz * map to add the specified range [f, t). region_chg does 7581dd308a7SMike Kravetz * not change the number of huge pages represented by the 7590db9d74eSMina Almasry * map. A number of new file_region structures is added to the cache as a 7600db9d74eSMina Almasry * placeholder, for the subsequent region_add call to use. At least 1 7610db9d74eSMina Almasry * file_region structure is added. 7620db9d74eSMina Almasry * 7630db9d74eSMina Almasry * out_regions_needed is the number of regions added to the 7640db9d74eSMina Almasry * resv->adds_in_progress. This value needs to be provided to a follow up call 7650db9d74eSMina Almasry * to region_add or region_abort for proper accounting. 7665e911373SMike Kravetz * 7675e911373SMike Kravetz * Returns the number of huge pages that need to be added to the existing 7685e911373SMike Kravetz * reservation map for the range [f, t). This number is greater or equal to 7695e911373SMike Kravetz * zero. -ENOMEM is returned if a new file_region structure or cache entry 7705e911373SMike Kravetz * is needed and can not be allocated. 7711dd308a7SMike Kravetz */ 7720db9d74eSMina Almasry static long region_chg(struct resv_map *resv, long f, long t, 7730db9d74eSMina Almasry long *out_regions_needed) 77496822904SAndy Whitcroft { 77596822904SAndy Whitcroft long chg = 0; 77696822904SAndy Whitcroft 7777b24d861SDavidlohr Bueso spin_lock(&resv->lock); 7785e911373SMike Kravetz 779972a3da3SWei Yang /* Count how many hugepages in this range are NOT represented. */ 780075a61d0SMina Almasry chg = add_reservation_in_range(resv, f, t, NULL, NULL, 781972a3da3SWei Yang out_regions_needed); 7825e911373SMike Kravetz 7830db9d74eSMina Almasry if (*out_regions_needed == 0) 7840db9d74eSMina Almasry *out_regions_needed = 1; 7855e911373SMike Kravetz 7860db9d74eSMina Almasry if (allocate_file_region_entries(resv, *out_regions_needed)) 7875e911373SMike Kravetz return -ENOMEM; 7885e911373SMike Kravetz 7890db9d74eSMina Almasry resv->adds_in_progress += *out_regions_needed; 79096822904SAndy Whitcroft 7917b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 79296822904SAndy Whitcroft return chg; 79396822904SAndy Whitcroft } 79496822904SAndy Whitcroft 7951dd308a7SMike Kravetz /* 7965e911373SMike Kravetz * Abort the in progress add operation. The adds_in_progress field 7975e911373SMike Kravetz * of the resv_map keeps track of the operations in progress between 7985e911373SMike Kravetz * calls to region_chg and region_add. Operations are sometimes 7995e911373SMike Kravetz * aborted after the call to region_chg. In such cases, region_abort 8000db9d74eSMina Almasry * is called to decrement the adds_in_progress counter. regions_needed 8010db9d74eSMina Almasry * is the value returned by the region_chg call, it is used to decrement 8020db9d74eSMina Almasry * the adds_in_progress counter. 8035e911373SMike Kravetz * 8045e911373SMike Kravetz * NOTE: The range arguments [f, t) are not needed or used in this 8055e911373SMike Kravetz * routine. They are kept to make reading the calling code easier as 8065e911373SMike Kravetz * arguments will match the associated region_chg call. 8075e911373SMike Kravetz */ 8080db9d74eSMina Almasry static void region_abort(struct resv_map *resv, long f, long t, 8090db9d74eSMina Almasry long regions_needed) 8105e911373SMike Kravetz { 8115e911373SMike Kravetz spin_lock(&resv->lock); 8125e911373SMike Kravetz VM_BUG_ON(!resv->region_cache_count); 8130db9d74eSMina Almasry resv->adds_in_progress -= regions_needed; 8145e911373SMike Kravetz spin_unlock(&resv->lock); 8155e911373SMike Kravetz } 8165e911373SMike Kravetz 8175e911373SMike Kravetz /* 818feba16e2SMike Kravetz * Delete the specified range [f, t) from the reserve map. If the 819feba16e2SMike Kravetz * t parameter is LONG_MAX, this indicates that ALL regions after f 820feba16e2SMike Kravetz * should be deleted. Locate the regions which intersect [f, t) 821feba16e2SMike Kravetz * and either trim, delete or split the existing regions. 822feba16e2SMike Kravetz * 823feba16e2SMike Kravetz * Returns the number of huge pages deleted from the reserve map. 824feba16e2SMike Kravetz * In the normal case, the return value is zero or more. In the 825feba16e2SMike Kravetz * case where a region must be split, a new region descriptor must 826feba16e2SMike Kravetz * be allocated. If the allocation fails, -ENOMEM will be returned. 827feba16e2SMike Kravetz * NOTE: If the parameter t == LONG_MAX, then we will never split 828feba16e2SMike Kravetz * a region and possibly return -ENOMEM. Callers specifying 829feba16e2SMike Kravetz * t == LONG_MAX do not need to check for -ENOMEM error. 8301dd308a7SMike Kravetz */ 831feba16e2SMike Kravetz static long region_del(struct resv_map *resv, long f, long t) 83296822904SAndy Whitcroft { 8331406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 83496822904SAndy Whitcroft struct file_region *rg, *trg; 835feba16e2SMike Kravetz struct file_region *nrg = NULL; 836feba16e2SMike Kravetz long del = 0; 83796822904SAndy Whitcroft 838feba16e2SMike Kravetz retry: 8397b24d861SDavidlohr Bueso spin_lock(&resv->lock); 840feba16e2SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 841dbe409e4SMike Kravetz /* 842dbe409e4SMike Kravetz * Skip regions before the range to be deleted. file_region 843dbe409e4SMike Kravetz * ranges are normally of the form [from, to). However, there 844dbe409e4SMike Kravetz * may be a "placeholder" entry in the map which is of the form 845dbe409e4SMike Kravetz * (from, to) with from == to. Check for placeholder entries 846dbe409e4SMike Kravetz * at the beginning of the range to be deleted. 847dbe409e4SMike Kravetz */ 848dbe409e4SMike Kravetz if (rg->to <= f && (rg->to != rg->from || rg->to != f)) 849feba16e2SMike Kravetz continue; 850dbe409e4SMike Kravetz 851feba16e2SMike Kravetz if (rg->from >= t) 85296822904SAndy Whitcroft break; 85396822904SAndy Whitcroft 854feba16e2SMike Kravetz if (f > rg->from && t < rg->to) { /* Must split region */ 855feba16e2SMike Kravetz /* 856feba16e2SMike Kravetz * Check for an entry in the cache before dropping 857feba16e2SMike Kravetz * lock and attempting allocation. 858feba16e2SMike Kravetz */ 859feba16e2SMike Kravetz if (!nrg && 860feba16e2SMike Kravetz resv->region_cache_count > resv->adds_in_progress) { 861feba16e2SMike Kravetz nrg = list_first_entry(&resv->region_cache, 862feba16e2SMike Kravetz struct file_region, 863feba16e2SMike Kravetz link); 864feba16e2SMike Kravetz list_del(&nrg->link); 865feba16e2SMike Kravetz resv->region_cache_count--; 86696822904SAndy Whitcroft } 86796822904SAndy Whitcroft 868feba16e2SMike Kravetz if (!nrg) { 869feba16e2SMike Kravetz spin_unlock(&resv->lock); 870feba16e2SMike Kravetz nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 871feba16e2SMike Kravetz if (!nrg) 872feba16e2SMike Kravetz return -ENOMEM; 873feba16e2SMike Kravetz goto retry; 874feba16e2SMike Kravetz } 875feba16e2SMike Kravetz 876feba16e2SMike Kravetz del += t - f; 87779aa925bSMike Kravetz hugetlb_cgroup_uncharge_file_region( 878d85aecf2SMiaohe Lin resv, rg, t - f, false); 879feba16e2SMike Kravetz 880feba16e2SMike Kravetz /* New entry for end of split region */ 881feba16e2SMike Kravetz nrg->from = t; 882feba16e2SMike Kravetz nrg->to = rg->to; 883075a61d0SMina Almasry 884075a61d0SMina Almasry copy_hugetlb_cgroup_uncharge_info(nrg, rg); 885075a61d0SMina Almasry 886feba16e2SMike Kravetz INIT_LIST_HEAD(&nrg->link); 887feba16e2SMike Kravetz 888feba16e2SMike Kravetz /* Original entry is trimmed */ 889feba16e2SMike Kravetz rg->to = f; 890feba16e2SMike Kravetz 891feba16e2SMike Kravetz list_add(&nrg->link, &rg->link); 892feba16e2SMike Kravetz nrg = NULL; 89396822904SAndy Whitcroft break; 894feba16e2SMike Kravetz } 895feba16e2SMike Kravetz 896feba16e2SMike Kravetz if (f <= rg->from && t >= rg->to) { /* Remove entire region */ 897feba16e2SMike Kravetz del += rg->to - rg->from; 898075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 899d85aecf2SMiaohe Lin rg->to - rg->from, true); 90096822904SAndy Whitcroft list_del(&rg->link); 90196822904SAndy Whitcroft kfree(rg); 902feba16e2SMike Kravetz continue; 90396822904SAndy Whitcroft } 9047b24d861SDavidlohr Bueso 905feba16e2SMike Kravetz if (f <= rg->from) { /* Trim beginning of region */ 906075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 907d85aecf2SMiaohe Lin t - rg->from, false); 908075a61d0SMina Almasry 90979aa925bSMike Kravetz del += t - rg->from; 91079aa925bSMike Kravetz rg->from = t; 91179aa925bSMike Kravetz } else { /* Trim end of region */ 912075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 913d85aecf2SMiaohe Lin rg->to - f, false); 91479aa925bSMike Kravetz 91579aa925bSMike Kravetz del += rg->to - f; 91679aa925bSMike Kravetz rg->to = f; 917feba16e2SMike Kravetz } 918feba16e2SMike Kravetz } 919feba16e2SMike Kravetz 9207b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 921feba16e2SMike Kravetz kfree(nrg); 922feba16e2SMike Kravetz return del; 92396822904SAndy Whitcroft } 92496822904SAndy Whitcroft 9251dd308a7SMike Kravetz /* 926b5cec28dSMike Kravetz * A rare out of memory error was encountered which prevented removal of 927b5cec28dSMike Kravetz * the reserve map region for a page. The huge page itself was free'ed 928b5cec28dSMike Kravetz * and removed from the page cache. This routine will adjust the subpool 929b5cec28dSMike Kravetz * usage count, and the global reserve count if needed. By incrementing 930b5cec28dSMike Kravetz * these counts, the reserve map entry which could not be deleted will 931b5cec28dSMike Kravetz * appear as a "reserved" entry instead of simply dangling with incorrect 932b5cec28dSMike Kravetz * counts. 933b5cec28dSMike Kravetz */ 93472e2936cSzhong jiang void hugetlb_fix_reserve_counts(struct inode *inode) 935b5cec28dSMike Kravetz { 936b5cec28dSMike Kravetz struct hugepage_subpool *spool = subpool_inode(inode); 937b5cec28dSMike Kravetz long rsv_adjust; 938da56388cSMiaohe Lin bool reserved = false; 939b5cec28dSMike Kravetz 940b5cec28dSMike Kravetz rsv_adjust = hugepage_subpool_get_pages(spool, 1); 941da56388cSMiaohe Lin if (rsv_adjust > 0) { 942b5cec28dSMike Kravetz struct hstate *h = hstate_inode(inode); 943b5cec28dSMike Kravetz 944da56388cSMiaohe Lin if (!hugetlb_acct_memory(h, 1)) 945da56388cSMiaohe Lin reserved = true; 946da56388cSMiaohe Lin } else if (!rsv_adjust) { 947da56388cSMiaohe Lin reserved = true; 948b5cec28dSMike Kravetz } 949da56388cSMiaohe Lin 950da56388cSMiaohe Lin if (!reserved) 951da56388cSMiaohe Lin pr_warn("hugetlb: Huge Page Reserved count may go negative.\n"); 952b5cec28dSMike Kravetz } 953b5cec28dSMike Kravetz 954b5cec28dSMike Kravetz /* 9551dd308a7SMike Kravetz * Count and return the number of huge pages in the reserve map 9561dd308a7SMike Kravetz * that intersect with the range [f, t). 9571dd308a7SMike Kravetz */ 9581406ec9bSJoonsoo Kim static long region_count(struct resv_map *resv, long f, long t) 95984afd99bSAndy Whitcroft { 9601406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 96184afd99bSAndy Whitcroft struct file_region *rg; 96284afd99bSAndy Whitcroft long chg = 0; 96384afd99bSAndy Whitcroft 9647b24d861SDavidlohr Bueso spin_lock(&resv->lock); 96584afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 96684afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 967f2135a4aSWang Sheng-Hui long seg_from; 968f2135a4aSWang Sheng-Hui long seg_to; 96984afd99bSAndy Whitcroft 97084afd99bSAndy Whitcroft if (rg->to <= f) 97184afd99bSAndy Whitcroft continue; 97284afd99bSAndy Whitcroft if (rg->from >= t) 97384afd99bSAndy Whitcroft break; 97484afd99bSAndy Whitcroft 97584afd99bSAndy Whitcroft seg_from = max(rg->from, f); 97684afd99bSAndy Whitcroft seg_to = min(rg->to, t); 97784afd99bSAndy Whitcroft 97884afd99bSAndy Whitcroft chg += seg_to - seg_from; 97984afd99bSAndy Whitcroft } 9807b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 98184afd99bSAndy Whitcroft 98284afd99bSAndy Whitcroft return chg; 98384afd99bSAndy Whitcroft } 98484afd99bSAndy Whitcroft 98596822904SAndy Whitcroft /* 986e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 987e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 988e7c4b0bfSAndy Whitcroft */ 989a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 990a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 991e7c4b0bfSAndy Whitcroft { 992a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 993a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 994e7c4b0bfSAndy Whitcroft } 995e7c4b0bfSAndy Whitcroft 9960fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 9970fe6e20bSNaoya Horiguchi unsigned long address) 9980fe6e20bSNaoya Horiguchi { 9990fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 10000fe6e20bSNaoya Horiguchi } 1001dee41079SDan Williams EXPORT_SYMBOL_GPL(linear_hugepage_index); 10020fe6e20bSNaoya Horiguchi 10038cfd014eSMatthew Wilcox (Oracle) /** 10048cfd014eSMatthew Wilcox (Oracle) * vma_kernel_pagesize - Page size granularity for this VMA. 10058cfd014eSMatthew Wilcox (Oracle) * @vma: The user mapping. 10068cfd014eSMatthew Wilcox (Oracle) * 10078cfd014eSMatthew Wilcox (Oracle) * Folios in this VMA will be aligned to, and at least the size of the 10088cfd014eSMatthew Wilcox (Oracle) * number of bytes returned by this function. 10098cfd014eSMatthew Wilcox (Oracle) * 10108cfd014eSMatthew Wilcox (Oracle) * Return: The default size of the folios allocated when backing a VMA. 101108fba699SMel Gorman */ 101208fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 101308fba699SMel Gorman { 101405ea8860SDan Williams if (vma->vm_ops && vma->vm_ops->pagesize) 101505ea8860SDan Williams return vma->vm_ops->pagesize(vma); 101608fba699SMel Gorman return PAGE_SIZE; 101708fba699SMel Gorman } 1018f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 101908fba699SMel Gorman 102008fba699SMel Gorman /* 10213340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 10223340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 102309135cc5SDan Williams * architectures where it differs, an architecture-specific 'strong' 102409135cc5SDan Williams * version of this symbol is required. 10253340289dSMel Gorman */ 102609135cc5SDan Williams __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 10273340289dSMel Gorman { 10283340289dSMel Gorman return vma_kernel_pagesize(vma); 10293340289dSMel Gorman } 10303340289dSMel Gorman 10313340289dSMel Gorman /* 103284afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 103384afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 103484afd99bSAndy Whitcroft * alignment. 103584afd99bSAndy Whitcroft */ 103684afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 103784afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 103804f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 103984afd99bSAndy Whitcroft 1040a1e78772SMel Gorman /* 1041a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 1042a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 1043a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 1044a1e78772SMel Gorman * 10458d9bfb26SMike Kravetz * With the exception of hugetlb_dup_vma_private() which is called at fork(), 1046a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 1047a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 1048a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 104984afd99bSAndy Whitcroft * 105084afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 105184afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 105284afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 105384afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 105484afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 105584afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 105684afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 105784afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 1058a1e78772SMel Gorman */ 1059e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 1060e7c4b0bfSAndy Whitcroft { 1061e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 1062e7c4b0bfSAndy Whitcroft } 1063e7c4b0bfSAndy Whitcroft 1064e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 1065e7c4b0bfSAndy Whitcroft unsigned long value) 1066e7c4b0bfSAndy Whitcroft { 1067e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 1068e7c4b0bfSAndy Whitcroft } 1069e7c4b0bfSAndy Whitcroft 1070e9fe92aeSMina Almasry static void 1071e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, 1072e9fe92aeSMina Almasry struct hugetlb_cgroup *h_cg, 1073e9fe92aeSMina Almasry struct hstate *h) 1074e9fe92aeSMina Almasry { 1075e9fe92aeSMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 1076e9fe92aeSMina Almasry if (!h_cg || !h) { 1077e9fe92aeSMina Almasry resv_map->reservation_counter = NULL; 1078e9fe92aeSMina Almasry resv_map->pages_per_hpage = 0; 1079e9fe92aeSMina Almasry resv_map->css = NULL; 1080e9fe92aeSMina Almasry } else { 1081e9fe92aeSMina Almasry resv_map->reservation_counter = 1082e9fe92aeSMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 1083e9fe92aeSMina Almasry resv_map->pages_per_hpage = pages_per_huge_page(h); 1084e9fe92aeSMina Almasry resv_map->css = &h_cg->css; 1085e9fe92aeSMina Almasry } 1086e9fe92aeSMina Almasry #endif 1087e9fe92aeSMina Almasry } 1088e9fe92aeSMina Almasry 10899119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 109084afd99bSAndy Whitcroft { 109184afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 10925e911373SMike Kravetz struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 10935e911373SMike Kravetz 10945e911373SMike Kravetz if (!resv_map || !rg) { 10955e911373SMike Kravetz kfree(resv_map); 10965e911373SMike Kravetz kfree(rg); 109784afd99bSAndy Whitcroft return NULL; 10985e911373SMike Kravetz } 109984afd99bSAndy Whitcroft 110084afd99bSAndy Whitcroft kref_init(&resv_map->refs); 11017b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 110284afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 1103bf491692SRik van Riel init_rwsem(&resv_map->rw_sema); 110484afd99bSAndy Whitcroft 11055e911373SMike Kravetz resv_map->adds_in_progress = 0; 1106e9fe92aeSMina Almasry /* 1107e9fe92aeSMina Almasry * Initialize these to 0. On shared mappings, 0's here indicate these 1108e9fe92aeSMina Almasry * fields don't do cgroup accounting. On private mappings, these will be 1109e9fe92aeSMina Almasry * re-initialized to the proper values, to indicate that hugetlb cgroup 1110e9fe92aeSMina Almasry * reservations are to be un-charged from here. 1111e9fe92aeSMina Almasry */ 1112e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); 11135e911373SMike Kravetz 11145e911373SMike Kravetz INIT_LIST_HEAD(&resv_map->region_cache); 11155e911373SMike Kravetz list_add(&rg->link, &resv_map->region_cache); 11165e911373SMike Kravetz resv_map->region_cache_count = 1; 11175e911373SMike Kravetz 111884afd99bSAndy Whitcroft return resv_map; 111984afd99bSAndy Whitcroft } 112084afd99bSAndy Whitcroft 11219119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 112284afd99bSAndy Whitcroft { 112384afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 11245e911373SMike Kravetz struct list_head *head = &resv_map->region_cache; 11255e911373SMike Kravetz struct file_region *rg, *trg; 112684afd99bSAndy Whitcroft 112784afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 1128feba16e2SMike Kravetz region_del(resv_map, 0, LONG_MAX); 11295e911373SMike Kravetz 11305e911373SMike Kravetz /* ... and any entries left in the cache */ 11315e911373SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 11325e911373SMike Kravetz list_del(&rg->link); 11335e911373SMike Kravetz kfree(rg); 11345e911373SMike Kravetz } 11355e911373SMike Kravetz 11365e911373SMike Kravetz VM_BUG_ON(resv_map->adds_in_progress); 11375e911373SMike Kravetz 113884afd99bSAndy Whitcroft kfree(resv_map); 113984afd99bSAndy Whitcroft } 114084afd99bSAndy Whitcroft 11414e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 11424e35f483SJoonsoo Kim { 1143f27a5136SMike Kravetz /* 1144f27a5136SMike Kravetz * At inode evict time, i_mapping may not point to the original 1145f27a5136SMike Kravetz * address space within the inode. This original address space 1146f27a5136SMike Kravetz * contains the pointer to the resv_map. So, always use the 1147f27a5136SMike Kravetz * address space embedded within the inode. 1148f27a5136SMike Kravetz * The VERY common case is inode->mapping == &inode->i_data but, 1149f27a5136SMike Kravetz * this may not be true for device special inodes. 1150f27a5136SMike Kravetz */ 1151f27a5136SMike Kravetz return (struct resv_map *)(&inode->i_data)->private_data; 11524e35f483SJoonsoo Kim } 11534e35f483SJoonsoo Kim 115484afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 1155a1e78772SMel Gorman { 115681d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 11574e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 11584e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 11594e35f483SJoonsoo Kim struct inode *inode = mapping->host; 11604e35f483SJoonsoo Kim 11614e35f483SJoonsoo Kim return inode_resv_map(inode); 11624e35f483SJoonsoo Kim 11634e35f483SJoonsoo Kim } else { 116484afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 116584afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 11664e35f483SJoonsoo Kim } 1167a1e78772SMel Gorman } 1168a1e78772SMel Gorman 116984afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 1170a1e78772SMel Gorman { 117181d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 117281d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 1173a1e78772SMel Gorman 117492fe9dcbSRik van Riel set_vma_private_data(vma, (unsigned long)map); 117504f2cbe3SMel Gorman } 117604f2cbe3SMel Gorman 117704f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 117804f2cbe3SMel Gorman { 117981d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 118081d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 1181e7c4b0bfSAndy Whitcroft 1182e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 118304f2cbe3SMel Gorman } 118404f2cbe3SMel Gorman 118504f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 118604f2cbe3SMel Gorman { 118781d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 1188e7c4b0bfSAndy Whitcroft 1189e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 1190a1e78772SMel Gorman } 1191a1e78772SMel Gorman 1192512b420aSMike Kravetz bool __vma_private_lock(struct vm_area_struct *vma) 1193512b420aSMike Kravetz { 1194512b420aSMike Kravetz return !(vma->vm_flags & VM_MAYSHARE) && 1195512b420aSMike Kravetz get_vma_private_data(vma) & ~HPAGE_RESV_MASK && 1196512b420aSMike Kravetz is_vma_resv_set(vma, HPAGE_RESV_OWNER); 1197512b420aSMike Kravetz } 1198512b420aSMike Kravetz 11998d9bfb26SMike Kravetz void hugetlb_dup_vma_private(struct vm_area_struct *vma) 1200a1e78772SMel Gorman { 120181d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 12028d9bfb26SMike Kravetz /* 12038d9bfb26SMike Kravetz * Clear vm_private_data 1204612b8a31SMike Kravetz * - For shared mappings this is a per-vma semaphore that may be 1205612b8a31SMike Kravetz * allocated in a subsequent call to hugetlb_vm_op_open. 1206612b8a31SMike Kravetz * Before clearing, make sure pointer is not associated with vma 1207612b8a31SMike Kravetz * as this will leak the structure. This is the case when called 1208612b8a31SMike Kravetz * via clear_vma_resv_huge_pages() and hugetlb_vm_op_open has already 1209612b8a31SMike Kravetz * been called to allocate a new structure. 12108d9bfb26SMike Kravetz * - For MAP_PRIVATE mappings, this is the reserve map which does 12118d9bfb26SMike Kravetz * not apply to children. Faults generated by the children are 12128d9bfb26SMike Kravetz * not guaranteed to succeed, even if read-only. 12138d9bfb26SMike Kravetz */ 1214612b8a31SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 1215612b8a31SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 1216612b8a31SMike Kravetz 1217612b8a31SMike Kravetz if (vma_lock && vma_lock->vma != vma) 1218612b8a31SMike Kravetz vma->vm_private_data = NULL; 1219612b8a31SMike Kravetz } else 1220612b8a31SMike Kravetz vma->vm_private_data = NULL; 1221a1e78772SMel Gorman } 1222a1e78772SMel Gorman 1223550a7d60SMina Almasry /* 1224550a7d60SMina Almasry * Reset and decrement one ref on hugepage private reservation. 12258651a137SLorenzo Stoakes * Called with mm->mmap_lock writer semaphore held. 1226550a7d60SMina Almasry * This function should be only used by move_vma() and operate on 1227550a7d60SMina Almasry * same sized vma. It should never come here with last ref on the 1228550a7d60SMina Almasry * reservation. 1229550a7d60SMina Almasry */ 1230550a7d60SMina Almasry void clear_vma_resv_huge_pages(struct vm_area_struct *vma) 1231550a7d60SMina Almasry { 1232550a7d60SMina Almasry /* 1233550a7d60SMina Almasry * Clear the old hugetlb private page reservation. 1234550a7d60SMina Almasry * It has already been transferred to new_vma. 1235550a7d60SMina Almasry * 1236550a7d60SMina Almasry * During a mremap() operation of a hugetlb vma we call move_vma() 1237550a7d60SMina Almasry * which copies vma into new_vma and unmaps vma. After the copy 1238550a7d60SMina Almasry * operation both new_vma and vma share a reference to the resv_map 1239550a7d60SMina Almasry * struct, and at that point vma is about to be unmapped. We don't 1240550a7d60SMina Almasry * want to return the reservation to the pool at unmap of vma because 1241550a7d60SMina Almasry * the reservation still lives on in new_vma, so simply decrement the 1242550a7d60SMina Almasry * ref here and remove the resv_map reference from this vma. 1243550a7d60SMina Almasry */ 1244550a7d60SMina Almasry struct resv_map *reservations = vma_resv_map(vma); 1245550a7d60SMina Almasry 1246afe041c2SBui Quang Minh if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 1247afe041c2SBui Quang Minh resv_map_put_hugetlb_cgroup_uncharge_info(reservations); 1248550a7d60SMina Almasry kref_put(&reservations->refs, resv_map_release); 1249afe041c2SBui Quang Minh } 1250550a7d60SMina Almasry 12518d9bfb26SMike Kravetz hugetlb_dup_vma_private(vma); 1252550a7d60SMina Almasry } 1253550a7d60SMina Almasry 1254a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 1255559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 1256a1e78772SMel Gorman { 1257af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 1258af0ed73eSJoonsoo Kim /* 1259af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 1260af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 1261af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 1262af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 1263af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 1264af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 1265af0ed73eSJoonsoo Kim * properly, so add work-around here. 1266af0ed73eSJoonsoo Kim */ 1267af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 1268559ec2f8SNicholas Krause return true; 1269af0ed73eSJoonsoo Kim else 1270559ec2f8SNicholas Krause return false; 1271af0ed73eSJoonsoo Kim } 1272a63884e9SJoonsoo Kim 1273a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 12741fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 12751fb1b0e9SMike Kravetz /* 12761fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 12771fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 12781fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 12797c8de358SEthon Paul * fallocate. In this case, there really are no reserves to 12801fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 12811fb1b0e9SMike Kravetz */ 12821fb1b0e9SMike Kravetz if (chg) 12831fb1b0e9SMike Kravetz return false; 12841fb1b0e9SMike Kravetz else 1285559ec2f8SNicholas Krause return true; 12861fb1b0e9SMike Kravetz } 1287a63884e9SJoonsoo Kim 1288a63884e9SJoonsoo Kim /* 1289a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 1290a63884e9SJoonsoo Kim * private mappings. 1291a63884e9SJoonsoo Kim */ 129267961f9dSMike Kravetz if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 129367961f9dSMike Kravetz /* 129467961f9dSMike Kravetz * Like the shared case above, a hole punch or truncate 129567961f9dSMike Kravetz * could have been performed on the private mapping. 129667961f9dSMike Kravetz * Examine the value of chg to determine if reserves 129767961f9dSMike Kravetz * actually exist or were previously consumed. 129867961f9dSMike Kravetz * Very Subtle - The value of chg comes from a previous 129967961f9dSMike Kravetz * call to vma_needs_reserves(). The reserve map for 130067961f9dSMike Kravetz * private mappings has different (opposite) semantics 130167961f9dSMike Kravetz * than that of shared mappings. vma_needs_reserves() 130267961f9dSMike Kravetz * has already taken this difference in semantics into 130367961f9dSMike Kravetz * account. Therefore, the meaning of chg is the same 130467961f9dSMike Kravetz * as in the shared case above. Code could easily be 130567961f9dSMike Kravetz * combined, but keeping it separate draws attention to 130667961f9dSMike Kravetz * subtle differences. 130767961f9dSMike Kravetz */ 130867961f9dSMike Kravetz if (chg) 130967961f9dSMike Kravetz return false; 131067961f9dSMike Kravetz else 1311559ec2f8SNicholas Krause return true; 131267961f9dSMike Kravetz } 1313a63884e9SJoonsoo Kim 1314559ec2f8SNicholas Krause return false; 1315a1e78772SMel Gorman } 1316a1e78772SMel Gorman 1317240d67a8SSidhartha Kumar static void enqueue_hugetlb_folio(struct hstate *h, struct folio *folio) 13181da177e4SLinus Torvalds { 1319240d67a8SSidhartha Kumar int nid = folio_nid(folio); 13209487ca60SMike Kravetz 13219487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1322240d67a8SSidhartha Kumar VM_BUG_ON_FOLIO(folio_ref_count(folio), folio); 1323b65a4edaSMike Kravetz 1324240d67a8SSidhartha Kumar list_move(&folio->lru, &h->hugepage_freelists[nid]); 1325a5516438SAndi Kleen h->free_huge_pages++; 1326a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 1327240d67a8SSidhartha Kumar folio_set_hugetlb_freed(folio); 13281da177e4SLinus Torvalds } 13291da177e4SLinus Torvalds 1330a36f1e90SSidhartha Kumar static struct folio *dequeue_hugetlb_folio_node_exact(struct hstate *h, 1331a36f1e90SSidhartha Kumar int nid) 1332bf50bab2SNaoya Horiguchi { 1333a36f1e90SSidhartha Kumar struct folio *folio; 13341a08ae36SPavel Tatashin bool pin = !!(current->flags & PF_MEMALLOC_PIN); 1335bf50bab2SNaoya Horiguchi 13369487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1337a36f1e90SSidhartha Kumar list_for_each_entry(folio, &h->hugepage_freelists[nid], lru) { 1338a36f1e90SSidhartha Kumar if (pin && !folio_is_longterm_pinnable(folio)) 1339bbe88753SJoonsoo Kim continue; 1340bbe88753SJoonsoo Kim 1341a36f1e90SSidhartha Kumar if (folio_test_hwpoison(folio)) 13426664bfc8SWei Yang continue; 1343bbe88753SJoonsoo Kim 1344a36f1e90SSidhartha Kumar list_move(&folio->lru, &h->hugepage_activelist); 1345a36f1e90SSidhartha Kumar folio_ref_unfreeze(folio, 1); 1346a36f1e90SSidhartha Kumar folio_clear_hugetlb_freed(folio); 1347bf50bab2SNaoya Horiguchi h->free_huge_pages--; 1348bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 1349a36f1e90SSidhartha Kumar return folio; 1350bf50bab2SNaoya Horiguchi } 1351bf50bab2SNaoya Horiguchi 13526664bfc8SWei Yang return NULL; 13536664bfc8SWei Yang } 13546664bfc8SWei Yang 1355a36f1e90SSidhartha Kumar static struct folio *dequeue_hugetlb_folio_nodemask(struct hstate *h, gfp_t gfp_mask, 1356a36f1e90SSidhartha Kumar int nid, nodemask_t *nmask) 135794310cbcSAnshuman Khandual { 13583e59fcb0SMichal Hocko unsigned int cpuset_mems_cookie; 13593e59fcb0SMichal Hocko struct zonelist *zonelist; 13603e59fcb0SMichal Hocko struct zone *zone; 13613e59fcb0SMichal Hocko struct zoneref *z; 136298fa15f3SAnshuman Khandual int node = NUMA_NO_NODE; 13633e59fcb0SMichal Hocko 13643e59fcb0SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 13653e59fcb0SMichal Hocko 13663e59fcb0SMichal Hocko retry_cpuset: 13673e59fcb0SMichal Hocko cpuset_mems_cookie = read_mems_allowed_begin(); 13683e59fcb0SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 1369a36f1e90SSidhartha Kumar struct folio *folio; 137094310cbcSAnshuman Khandual 13713e59fcb0SMichal Hocko if (!cpuset_zone_allowed(zone, gfp_mask)) 13723e59fcb0SMichal Hocko continue; 13733e59fcb0SMichal Hocko /* 13743e59fcb0SMichal Hocko * no need to ask again on the same node. Pool is node rather than 13753e59fcb0SMichal Hocko * zone aware 13763e59fcb0SMichal Hocko */ 13773e59fcb0SMichal Hocko if (zone_to_nid(zone) == node) 13783e59fcb0SMichal Hocko continue; 13793e59fcb0SMichal Hocko node = zone_to_nid(zone); 138094310cbcSAnshuman Khandual 1381a36f1e90SSidhartha Kumar folio = dequeue_hugetlb_folio_node_exact(h, node); 1382a36f1e90SSidhartha Kumar if (folio) 1383a36f1e90SSidhartha Kumar return folio; 138494310cbcSAnshuman Khandual } 13853e59fcb0SMichal Hocko if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 13863e59fcb0SMichal Hocko goto retry_cpuset; 13873e59fcb0SMichal Hocko 138894310cbcSAnshuman Khandual return NULL; 138994310cbcSAnshuman Khandual } 139094310cbcSAnshuman Khandual 13918346d69dSXin Hao static unsigned long available_huge_pages(struct hstate *h) 13928346d69dSXin Hao { 13938346d69dSXin Hao return h->free_huge_pages - h->resv_huge_pages; 13948346d69dSXin Hao } 13958346d69dSXin Hao 1396ff7d853bSSidhartha Kumar static struct folio *dequeue_hugetlb_folio_vma(struct hstate *h, 1397a5516438SAndi Kleen struct vm_area_struct *vma, 1398af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 1399af0ed73eSJoonsoo Kim long chg) 14001da177e4SLinus Torvalds { 1401a36f1e90SSidhartha Kumar struct folio *folio = NULL; 1402480eccf9SLee Schermerhorn struct mempolicy *mpol; 140304ec6264SVlastimil Babka gfp_t gfp_mask; 14043e59fcb0SMichal Hocko nodemask_t *nodemask; 140504ec6264SVlastimil Babka int nid; 14061da177e4SLinus Torvalds 1407a1e78772SMel Gorman /* 1408a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 1409a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 1410a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 1411a1e78772SMel Gorman */ 14128346d69dSXin Hao if (!vma_has_reserves(vma, chg) && !available_huge_pages(h)) 1413c0ff7453SMiao Xie goto err; 1414a1e78772SMel Gorman 141504f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 14168346d69dSXin Hao if (avoid_reserve && !available_huge_pages(h)) 14176eab04a8SJustin P. Mattock goto err; 141804f2cbe3SMel Gorman 141904ec6264SVlastimil Babka gfp_mask = htlb_alloc_mask(h); 142004ec6264SVlastimil Babka nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 1421cfcaa66fSBen Widawsky 1422cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 1423a36f1e90SSidhartha Kumar folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask, 1424a36f1e90SSidhartha Kumar nid, nodemask); 1425cfcaa66fSBen Widawsky 1426cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 1427cfcaa66fSBen Widawsky nodemask = NULL; 1428cfcaa66fSBen Widawsky } 1429cfcaa66fSBen Widawsky 1430a36f1e90SSidhartha Kumar if (!folio) 1431a36f1e90SSidhartha Kumar folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask, 1432a36f1e90SSidhartha Kumar nid, nodemask); 1433cfcaa66fSBen Widawsky 1434a36f1e90SSidhartha Kumar if (folio && !avoid_reserve && vma_has_reserves(vma, chg)) { 1435a36f1e90SSidhartha Kumar folio_set_hugetlb_restore_reserve(folio); 1436a63884e9SJoonsoo Kim h->resv_huge_pages--; 1437bf50bab2SNaoya Horiguchi } 1438cc9a6c87SMel Gorman 1439cc9a6c87SMel Gorman mpol_cond_put(mpol); 1440ff7d853bSSidhartha Kumar return folio; 1441cc9a6c87SMel Gorman 1442c0ff7453SMiao Xie err: 1443cc9a6c87SMel Gorman return NULL; 14441da177e4SLinus Torvalds } 14451da177e4SLinus Torvalds 14461cac6f2cSLuiz Capitulino /* 14471cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 14481cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 14491cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 14501cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 14511cac6f2cSLuiz Capitulino * node for alloc or free. 14521cac6f2cSLuiz Capitulino */ 14531cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 14541cac6f2cSLuiz Capitulino { 14550edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 14561cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 14571cac6f2cSLuiz Capitulino 14581cac6f2cSLuiz Capitulino return nid; 14591cac6f2cSLuiz Capitulino } 14601cac6f2cSLuiz Capitulino 14611cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 14621cac6f2cSLuiz Capitulino { 14631cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 14641cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 14651cac6f2cSLuiz Capitulino return nid; 14661cac6f2cSLuiz Capitulino } 14671cac6f2cSLuiz Capitulino 14681cac6f2cSLuiz Capitulino /* 14691cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 14701cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 14711cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 14721cac6f2cSLuiz Capitulino * mask. 14731cac6f2cSLuiz Capitulino */ 14741cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 14751cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 14761cac6f2cSLuiz Capitulino { 14771cac6f2cSLuiz Capitulino int nid; 14781cac6f2cSLuiz Capitulino 14791cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 14801cac6f2cSLuiz Capitulino 14811cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 14821cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 14831cac6f2cSLuiz Capitulino 14841cac6f2cSLuiz Capitulino return nid; 14851cac6f2cSLuiz Capitulino } 14861cac6f2cSLuiz Capitulino 14871cac6f2cSLuiz Capitulino /* 148810c6ec49SMike Kravetz * helper for remove_pool_huge_page() - return the previously saved 14891cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 14901cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 14911cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 14921cac6f2cSLuiz Capitulino */ 14931cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 14941cac6f2cSLuiz Capitulino { 14951cac6f2cSLuiz Capitulino int nid; 14961cac6f2cSLuiz Capitulino 14971cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 14981cac6f2cSLuiz Capitulino 14991cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 15001cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 15011cac6f2cSLuiz Capitulino 15021cac6f2cSLuiz Capitulino return nid; 15031cac6f2cSLuiz Capitulino } 15041cac6f2cSLuiz Capitulino 15051cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 15061cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 15071cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 15081cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 15091cac6f2cSLuiz Capitulino nr_nodes--) 15101cac6f2cSLuiz Capitulino 15111cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 15121cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 15131cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 15141cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 15151cac6f2cSLuiz Capitulino nr_nodes--) 15161cac6f2cSLuiz Capitulino 15178531fc6fSMike Kravetz /* used to demote non-gigantic_huge pages as well */ 1518911565b8SSidhartha Kumar static void __destroy_compound_gigantic_folio(struct folio *folio, 151934d9e35bSMike Kravetz unsigned int order, bool demote) 1520944d9fecSLuiz Capitulino { 1521944d9fecSLuiz Capitulino int i; 1522944d9fecSLuiz Capitulino int nr_pages = 1 << order; 152314455eabSCheng Li struct page *p; 1524944d9fecSLuiz Capitulino 152546f27228SMatthew Wilcox (Oracle) atomic_set(&folio->_entire_mapcount, 0); 1526eec20426SMatthew Wilcox (Oracle) atomic_set(&folio->_nr_pages_mapped, 0); 152794688e8eSMatthew Wilcox (Oracle) atomic_set(&folio->_pincount, 0); 152847e29d32SJohn Hubbard 152914455eabSCheng Li for (i = 1; i < nr_pages; i++) { 1530911565b8SSidhartha Kumar p = folio_page(folio, i); 15316c141973SMike Kravetz p->flags &= ~PAGE_FLAGS_CHECK_AT_FREE; 1532a01f4390SMike Kravetz p->mapping = NULL; 15331d798ca3SKirill A. Shutemov clear_compound_head(p); 153434d9e35bSMike Kravetz if (!demote) 1535944d9fecSLuiz Capitulino set_page_refcounted(p); 1536944d9fecSLuiz Capitulino } 1537944d9fecSLuiz Capitulino 1538911565b8SSidhartha Kumar __folio_clear_head(folio); 1539944d9fecSLuiz Capitulino } 1540944d9fecSLuiz Capitulino 1541911565b8SSidhartha Kumar static void destroy_compound_hugetlb_folio_for_demote(struct folio *folio, 15428531fc6fSMike Kravetz unsigned int order) 15438531fc6fSMike Kravetz { 1544911565b8SSidhartha Kumar __destroy_compound_gigantic_folio(folio, order, true); 15458531fc6fSMike Kravetz } 15468531fc6fSMike Kravetz 15478531fc6fSMike Kravetz #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 1548911565b8SSidhartha Kumar static void destroy_compound_gigantic_folio(struct folio *folio, 154934d9e35bSMike Kravetz unsigned int order) 155034d9e35bSMike Kravetz { 1551911565b8SSidhartha Kumar __destroy_compound_gigantic_folio(folio, order, false); 155234d9e35bSMike Kravetz } 155334d9e35bSMike Kravetz 15547f325a8dSSidhartha Kumar static void free_gigantic_folio(struct folio *folio, unsigned int order) 1555944d9fecSLuiz Capitulino { 1556cf11e85fSRoman Gushchin /* 1557cf11e85fSRoman Gushchin * If the page isn't allocated using the cma allocator, 1558cf11e85fSRoman Gushchin * cma_release() returns false. 1559cf11e85fSRoman Gushchin */ 1560dbda8feaSBarry Song #ifdef CONFIG_CMA 15617f325a8dSSidhartha Kumar int nid = folio_nid(folio); 15627f325a8dSSidhartha Kumar 15637f325a8dSSidhartha Kumar if (cma_release(hugetlb_cma[nid], &folio->page, 1 << order)) 1564cf11e85fSRoman Gushchin return; 1565dbda8feaSBarry Song #endif 1566cf11e85fSRoman Gushchin 15677f325a8dSSidhartha Kumar free_contig_range(folio_pfn(folio), 1 << order); 1568944d9fecSLuiz Capitulino } 1569944d9fecSLuiz Capitulino 15704eb0716eSAlexandre Ghiti #ifdef CONFIG_CONTIG_ALLOC 157119fc1a7eSSidhartha Kumar static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask, 1572d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) 1573944d9fecSLuiz Capitulino { 157419fc1a7eSSidhartha Kumar struct page *page; 157504adbc3fSMiaohe Lin unsigned long nr_pages = pages_per_huge_page(h); 1576953f064aSLi Xinhai if (nid == NUMA_NO_NODE) 1577953f064aSLi Xinhai nid = numa_mem_id(); 1578944d9fecSLuiz Capitulino 1579dbda8feaSBarry Song #ifdef CONFIG_CMA 1580dbda8feaSBarry Song { 1581cf11e85fSRoman Gushchin int node; 1582cf11e85fSRoman Gushchin 1583953f064aSLi Xinhai if (hugetlb_cma[nid]) { 1584953f064aSLi Xinhai page = cma_alloc(hugetlb_cma[nid], nr_pages, 1585953f064aSLi Xinhai huge_page_order(h), true); 1586953f064aSLi Xinhai if (page) 158719fc1a7eSSidhartha Kumar return page_folio(page); 1588953f064aSLi Xinhai } 1589953f064aSLi Xinhai 1590953f064aSLi Xinhai if (!(gfp_mask & __GFP_THISNODE)) { 1591cf11e85fSRoman Gushchin for_each_node_mask(node, *nodemask) { 1592953f064aSLi Xinhai if (node == nid || !hugetlb_cma[node]) 1593cf11e85fSRoman Gushchin continue; 1594cf11e85fSRoman Gushchin 1595cf11e85fSRoman Gushchin page = cma_alloc(hugetlb_cma[node], nr_pages, 1596cf11e85fSRoman Gushchin huge_page_order(h), true); 1597cf11e85fSRoman Gushchin if (page) 159819fc1a7eSSidhartha Kumar return page_folio(page); 1599cf11e85fSRoman Gushchin } 1600cf11e85fSRoman Gushchin } 1601953f064aSLi Xinhai } 1602dbda8feaSBarry Song #endif 1603cf11e85fSRoman Gushchin 160419fc1a7eSSidhartha Kumar page = alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); 160519fc1a7eSSidhartha Kumar return page ? page_folio(page) : NULL; 1606944d9fecSLuiz Capitulino } 1607944d9fecSLuiz Capitulino 16084eb0716eSAlexandre Ghiti #else /* !CONFIG_CONTIG_ALLOC */ 160919fc1a7eSSidhartha Kumar static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask, 16104eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 16114eb0716eSAlexandre Ghiti { 16124eb0716eSAlexandre Ghiti return NULL; 16134eb0716eSAlexandre Ghiti } 16144eb0716eSAlexandre Ghiti #endif /* CONFIG_CONTIG_ALLOC */ 1615944d9fecSLuiz Capitulino 1616e1073d1eSAneesh Kumar K.V #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 161719fc1a7eSSidhartha Kumar static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask, 16184eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 16194eb0716eSAlexandre Ghiti { 16204eb0716eSAlexandre Ghiti return NULL; 16214eb0716eSAlexandre Ghiti } 16227f325a8dSSidhartha Kumar static inline void free_gigantic_folio(struct folio *folio, 16237f325a8dSSidhartha Kumar unsigned int order) { } 1624911565b8SSidhartha Kumar static inline void destroy_compound_gigantic_folio(struct folio *folio, 1625d00181b9SKirill A. Shutemov unsigned int order) { } 1626944d9fecSLuiz Capitulino #endif 1627944d9fecSLuiz Capitulino 162832c87719SMike Kravetz static inline void __clear_hugetlb_destructor(struct hstate *h, 162932c87719SMike Kravetz struct folio *folio) 163032c87719SMike Kravetz { 163132c87719SMike Kravetz lockdep_assert_held(&hugetlb_lock); 163232c87719SMike Kravetz 16332431b5f2SMatthew Wilcox (Oracle) __folio_clear_hugetlb(folio); 163432c87719SMike Kravetz } 163532c87719SMike Kravetz 163632c87719SMike Kravetz /* 163732c87719SMike Kravetz * Remove hugetlb folio from lists. 163832c87719SMike Kravetz * If vmemmap exists for the folio, update dtor so that the folio appears 163932c87719SMike Kravetz * as just a compound page. Otherwise, wait until after allocating vmemmap 164032c87719SMike Kravetz * to update dtor. 164134d9e35bSMike Kravetz * 1642cfd5082bSSidhartha Kumar * A reference is held on the folio, except in the case of demote. 16436eb4e88aSMike Kravetz * 16446eb4e88aSMike Kravetz * Must be called with hugetlb lock held. 16456eb4e88aSMike Kravetz */ 1646cfd5082bSSidhartha Kumar static void __remove_hugetlb_folio(struct hstate *h, struct folio *folio, 164734d9e35bSMike Kravetz bool adjust_surplus, 164834d9e35bSMike Kravetz bool demote) 16496eb4e88aSMike Kravetz { 1650cfd5082bSSidhartha Kumar int nid = folio_nid(folio); 16516eb4e88aSMike Kravetz 1652f074732dSSidhartha Kumar VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio(folio), folio); 1653f074732dSSidhartha Kumar VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio_rsvd(folio), folio); 16546eb4e88aSMike Kravetz 16559487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 16566eb4e88aSMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 16576eb4e88aSMike Kravetz return; 16586eb4e88aSMike Kravetz 1659cfd5082bSSidhartha Kumar list_del(&folio->lru); 16606eb4e88aSMike Kravetz 1661cfd5082bSSidhartha Kumar if (folio_test_hugetlb_freed(folio)) { 16626eb4e88aSMike Kravetz h->free_huge_pages--; 16636eb4e88aSMike Kravetz h->free_huge_pages_node[nid]--; 16646eb4e88aSMike Kravetz } 16656eb4e88aSMike Kravetz if (adjust_surplus) { 16666eb4e88aSMike Kravetz h->surplus_huge_pages--; 16676eb4e88aSMike Kravetz h->surplus_huge_pages_node[nid]--; 16686eb4e88aSMike Kravetz } 16696eb4e88aSMike Kravetz 1670e32d20c0SMike Kravetz /* 167132c87719SMike Kravetz * We can only clear the hugetlb destructor after allocating vmemmap 167232c87719SMike Kravetz * pages. Otherwise, someone (memory error handling) may try to write 167332c87719SMike Kravetz * to tail struct pages. 167432c87719SMike Kravetz */ 167532c87719SMike Kravetz if (!folio_test_hugetlb_vmemmap_optimized(folio)) 167632c87719SMike Kravetz __clear_hugetlb_destructor(h, folio); 167732c87719SMike Kravetz 167832c87719SMike Kravetz /* 167934d9e35bSMike Kravetz * In the case of demote we do not ref count the page as it will soon 168034d9e35bSMike Kravetz * be turned into a page of smaller size. 1681e32d20c0SMike Kravetz */ 168234d9e35bSMike Kravetz if (!demote) 1683cfd5082bSSidhartha Kumar folio_ref_unfreeze(folio, 1); 16846eb4e88aSMike Kravetz 16856eb4e88aSMike Kravetz h->nr_huge_pages--; 16866eb4e88aSMike Kravetz h->nr_huge_pages_node[nid]--; 16876eb4e88aSMike Kravetz } 16886eb4e88aSMike Kravetz 1689cfd5082bSSidhartha Kumar static void remove_hugetlb_folio(struct hstate *h, struct folio *folio, 169034d9e35bSMike Kravetz bool adjust_surplus) 169134d9e35bSMike Kravetz { 1692cfd5082bSSidhartha Kumar __remove_hugetlb_folio(h, folio, adjust_surplus, false); 169334d9e35bSMike Kravetz } 169434d9e35bSMike Kravetz 1695cfd5082bSSidhartha Kumar static void remove_hugetlb_folio_for_demote(struct hstate *h, struct folio *folio, 16968531fc6fSMike Kravetz bool adjust_surplus) 16978531fc6fSMike Kravetz { 1698cfd5082bSSidhartha Kumar __remove_hugetlb_folio(h, folio, adjust_surplus, true); 16998531fc6fSMike Kravetz } 17008531fc6fSMike Kravetz 17012f6c57d6SSidhartha Kumar static void add_hugetlb_folio(struct hstate *h, struct folio *folio, 1702ad2fa371SMuchun Song bool adjust_surplus) 1703ad2fa371SMuchun Song { 1704ad2fa371SMuchun Song int zeroed; 17052f6c57d6SSidhartha Kumar int nid = folio_nid(folio); 1706ad2fa371SMuchun Song 17072f6c57d6SSidhartha Kumar VM_BUG_ON_FOLIO(!folio_test_hugetlb_vmemmap_optimized(folio), folio); 1708ad2fa371SMuchun Song 1709ad2fa371SMuchun Song lockdep_assert_held(&hugetlb_lock); 1710ad2fa371SMuchun Song 17112f6c57d6SSidhartha Kumar INIT_LIST_HEAD(&folio->lru); 1712ad2fa371SMuchun Song h->nr_huge_pages++; 1713ad2fa371SMuchun Song h->nr_huge_pages_node[nid]++; 1714ad2fa371SMuchun Song 1715ad2fa371SMuchun Song if (adjust_surplus) { 1716ad2fa371SMuchun Song h->surplus_huge_pages++; 1717ad2fa371SMuchun Song h->surplus_huge_pages_node[nid]++; 1718ad2fa371SMuchun Song } 1719ad2fa371SMuchun Song 17202431b5f2SMatthew Wilcox (Oracle) __folio_set_hugetlb(folio); 17212f6c57d6SSidhartha Kumar folio_change_private(folio, NULL); 1722a9e1eab2SMiaohe Lin /* 17232f6c57d6SSidhartha Kumar * We have to set hugetlb_vmemmap_optimized again as above 17242f6c57d6SSidhartha Kumar * folio_change_private(folio, NULL) cleared it. 1725a9e1eab2SMiaohe Lin */ 17262f6c57d6SSidhartha Kumar folio_set_hugetlb_vmemmap_optimized(folio); 1727ad2fa371SMuchun Song 1728ad2fa371SMuchun Song /* 17292f6c57d6SSidhartha Kumar * This folio is about to be managed by the hugetlb allocator and 1730b65a4edaSMike Kravetz * should have no users. Drop our reference, and check for others 1731b65a4edaSMike Kravetz * just in case. 1732ad2fa371SMuchun Song */ 17332f6c57d6SSidhartha Kumar zeroed = folio_put_testzero(folio); 17342f6c57d6SSidhartha Kumar if (unlikely(!zeroed)) 1735b65a4edaSMike Kravetz /* 1736454a00c4SMatthew Wilcox (Oracle) * It is VERY unlikely soneone else has taken a ref 1737454a00c4SMatthew Wilcox (Oracle) * on the folio. In this case, we simply return as 1738454a00c4SMatthew Wilcox (Oracle) * free_huge_folio() will be called when this other ref 1739454a00c4SMatthew Wilcox (Oracle) * is dropped. 1740b65a4edaSMike Kravetz */ 1741b65a4edaSMike Kravetz return; 1742b65a4edaSMike Kravetz 17432f6c57d6SSidhartha Kumar arch_clear_hugepage_flags(&folio->page); 1744240d67a8SSidhartha Kumar enqueue_hugetlb_folio(h, folio); 1745ad2fa371SMuchun Song } 1746ad2fa371SMuchun Song 17476f6956cfSSidhartha Kumar static void __update_and_free_hugetlb_folio(struct hstate *h, 17486f6956cfSSidhartha Kumar struct folio *folio) 17496af2acb6SAdam Litke { 175032c87719SMike Kravetz bool clear_dtor = folio_test_hugetlb_vmemmap_optimized(folio); 1751a5516438SAndi Kleen 17524eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1753944d9fecSLuiz Capitulino return; 175418229df5SAndy Whitcroft 1755161df60eSNaoya Horiguchi /* 1756161df60eSNaoya Horiguchi * If we don't know which subpages are hwpoisoned, we can't free 1757161df60eSNaoya Horiguchi * the hugepage, so it's leaked intentionally. 1758161df60eSNaoya Horiguchi */ 17597f325a8dSSidhartha Kumar if (folio_test_hugetlb_raw_hwp_unreliable(folio)) 1760161df60eSNaoya Horiguchi return; 1761161df60eSNaoya Horiguchi 17626f6956cfSSidhartha Kumar if (hugetlb_vmemmap_restore(h, &folio->page)) { 1763ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 1764ad2fa371SMuchun Song /* 1765ad2fa371SMuchun Song * If we cannot allocate vmemmap pages, just refuse to free the 1766ad2fa371SMuchun Song * page and put the page back on the hugetlb free list and treat 1767ad2fa371SMuchun Song * as a surplus page. 1768ad2fa371SMuchun Song */ 17697f325a8dSSidhartha Kumar add_hugetlb_folio(h, folio, true); 1770ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 1771ad2fa371SMuchun Song return; 1772ad2fa371SMuchun Song } 1773ad2fa371SMuchun Song 1774161df60eSNaoya Horiguchi /* 1775161df60eSNaoya Horiguchi * Move PageHWPoison flag from head page to the raw error pages, 1776161df60eSNaoya Horiguchi * which makes any healthy subpages reusable. 1777161df60eSNaoya Horiguchi */ 1778911565b8SSidhartha Kumar if (unlikely(folio_test_hwpoison(folio))) 17792ff6ceceSSidhartha Kumar folio_clear_hugetlb_hwpoison(folio); 1780161df60eSNaoya Horiguchi 178132c87719SMike Kravetz /* 178232c87719SMike Kravetz * If vmemmap pages were allocated above, then we need to clear the 178332c87719SMike Kravetz * hugetlb destructor under the hugetlb lock. 178432c87719SMike Kravetz */ 178532c87719SMike Kravetz if (clear_dtor) { 178632c87719SMike Kravetz spin_lock_irq(&hugetlb_lock); 178732c87719SMike Kravetz __clear_hugetlb_destructor(h, folio); 178832c87719SMike Kravetz spin_unlock_irq(&hugetlb_lock); 178932c87719SMike Kravetz } 179032c87719SMike Kravetz 1791a01f4390SMike Kravetz /* 1792a01f4390SMike Kravetz * Non-gigantic pages demoted from CMA allocated gigantic pages 17937f325a8dSSidhartha Kumar * need to be given back to CMA in free_gigantic_folio. 1794a01f4390SMike Kravetz */ 1795a01f4390SMike Kravetz if (hstate_is_gigantic(h) || 17962f6c57d6SSidhartha Kumar hugetlb_cma_folio(folio, huge_page_order(h))) { 1797911565b8SSidhartha Kumar destroy_compound_gigantic_folio(folio, huge_page_order(h)); 17987f325a8dSSidhartha Kumar free_gigantic_folio(folio, huge_page_order(h)); 1799944d9fecSLuiz Capitulino } else { 18006f6956cfSSidhartha Kumar __free_pages(&folio->page, huge_page_order(h)); 18016af2acb6SAdam Litke } 1802944d9fecSLuiz Capitulino } 18036af2acb6SAdam Litke 1804b65d4adbSMuchun Song /* 1805d6ef19e2SSidhartha Kumar * As update_and_free_hugetlb_folio() can be called under any context, so we cannot 1806b65d4adbSMuchun Song * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the 1807b65d4adbSMuchun Song * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate 1808b65d4adbSMuchun Song * the vmemmap pages. 1809b65d4adbSMuchun Song * 1810b65d4adbSMuchun Song * free_hpage_workfn() locklessly retrieves the linked list of pages to be 1811b65d4adbSMuchun Song * freed and frees them one-by-one. As the page->mapping pointer is going 1812b65d4adbSMuchun Song * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node 1813b65d4adbSMuchun Song * structure of a lockless linked list of huge pages to be freed. 1814b65d4adbSMuchun Song */ 1815b65d4adbSMuchun Song static LLIST_HEAD(hpage_freelist); 1816b65d4adbSMuchun Song 1817b65d4adbSMuchun Song static void free_hpage_workfn(struct work_struct *work) 1818b65d4adbSMuchun Song { 1819b65d4adbSMuchun Song struct llist_node *node; 1820b65d4adbSMuchun Song 1821b65d4adbSMuchun Song node = llist_del_all(&hpage_freelist); 1822b65d4adbSMuchun Song 1823b65d4adbSMuchun Song while (node) { 1824b65d4adbSMuchun Song struct page *page; 1825b65d4adbSMuchun Song struct hstate *h; 1826b65d4adbSMuchun Song 1827b65d4adbSMuchun Song page = container_of((struct address_space **)node, 1828b65d4adbSMuchun Song struct page, mapping); 1829b65d4adbSMuchun Song node = node->next; 1830b65d4adbSMuchun Song page->mapping = NULL; 1831b65d4adbSMuchun Song /* 1832affd26b1SSidhartha Kumar * The VM_BUG_ON_FOLIO(!folio_test_hugetlb(folio), folio) in 1833affd26b1SSidhartha Kumar * folio_hstate() is going to trigger because a previous call to 18349c5ccf2dSMatthew Wilcox (Oracle) * remove_hugetlb_folio() will clear the hugetlb bit, so do 18359c5ccf2dSMatthew Wilcox (Oracle) * not use folio_hstate() directly. 1836b65d4adbSMuchun Song */ 1837b65d4adbSMuchun Song h = size_to_hstate(page_size(page)); 1838b65d4adbSMuchun Song 18396f6956cfSSidhartha Kumar __update_and_free_hugetlb_folio(h, page_folio(page)); 1840b65d4adbSMuchun Song 1841b65d4adbSMuchun Song cond_resched(); 1842b65d4adbSMuchun Song } 1843b65d4adbSMuchun Song } 1844b65d4adbSMuchun Song static DECLARE_WORK(free_hpage_work, free_hpage_workfn); 1845b65d4adbSMuchun Song 1846b65d4adbSMuchun Song static inline void flush_free_hpage_work(struct hstate *h) 1847b65d4adbSMuchun Song { 18486213834cSMuchun Song if (hugetlb_vmemmap_optimizable(h)) 1849b65d4adbSMuchun Song flush_work(&free_hpage_work); 1850b65d4adbSMuchun Song } 1851b65d4adbSMuchun Song 1852d6ef19e2SSidhartha Kumar static void update_and_free_hugetlb_folio(struct hstate *h, struct folio *folio, 1853b65d4adbSMuchun Song bool atomic) 1854b65d4adbSMuchun Song { 1855d6ef19e2SSidhartha Kumar if (!folio_test_hugetlb_vmemmap_optimized(folio) || !atomic) { 18566f6956cfSSidhartha Kumar __update_and_free_hugetlb_folio(h, folio); 1857b65d4adbSMuchun Song return; 1858b65d4adbSMuchun Song } 1859b65d4adbSMuchun Song 1860b65d4adbSMuchun Song /* 1861b65d4adbSMuchun Song * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages. 1862b65d4adbSMuchun Song * 1863b65d4adbSMuchun Song * Only call schedule_work() if hpage_freelist is previously 1864b65d4adbSMuchun Song * empty. Otherwise, schedule_work() had been called but the workfn 1865b65d4adbSMuchun Song * hasn't retrieved the list yet. 1866b65d4adbSMuchun Song */ 1867d6ef19e2SSidhartha Kumar if (llist_add((struct llist_node *)&folio->mapping, &hpage_freelist)) 1868b65d4adbSMuchun Song schedule_work(&free_hpage_work); 1869b65d4adbSMuchun Song } 1870b65d4adbSMuchun Song 187110c6ec49SMike Kravetz static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list) 187210c6ec49SMike Kravetz { 187310c6ec49SMike Kravetz struct page *page, *t_page; 1874d6ef19e2SSidhartha Kumar struct folio *folio; 187510c6ec49SMike Kravetz 187610c6ec49SMike Kravetz list_for_each_entry_safe(page, t_page, list, lru) { 1877d6ef19e2SSidhartha Kumar folio = page_folio(page); 1878d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, folio, false); 187910c6ec49SMike Kravetz cond_resched(); 188010c6ec49SMike Kravetz } 188110c6ec49SMike Kravetz } 188210c6ec49SMike Kravetz 1883e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1884e5ff2159SAndi Kleen { 1885e5ff2159SAndi Kleen struct hstate *h; 1886e5ff2159SAndi Kleen 1887e5ff2159SAndi Kleen for_each_hstate(h) { 1888e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1889e5ff2159SAndi Kleen return h; 1890e5ff2159SAndi Kleen } 1891e5ff2159SAndi Kleen return NULL; 1892e5ff2159SAndi Kleen } 1893e5ff2159SAndi Kleen 1894454a00c4SMatthew Wilcox (Oracle) void free_huge_folio(struct folio *folio) 189527a85ef1SDavid Gibson { 1896a5516438SAndi Kleen /* 1897a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1898a5516438SAndi Kleen * compound page destructor. 1899a5516438SAndi Kleen */ 19000356c4b9SSidhartha Kumar struct hstate *h = folio_hstate(folio); 19010356c4b9SSidhartha Kumar int nid = folio_nid(folio); 19020356c4b9SSidhartha Kumar struct hugepage_subpool *spool = hugetlb_folio_subpool(folio); 190307443a85SJoonsoo Kim bool restore_reserve; 1904db71ef79SMike Kravetz unsigned long flags; 190527a85ef1SDavid Gibson 19060356c4b9SSidhartha Kumar VM_BUG_ON_FOLIO(folio_ref_count(folio), folio); 19070356c4b9SSidhartha Kumar VM_BUG_ON_FOLIO(folio_mapcount(folio), folio); 19088ace22bcSYongkai Wu 19090356c4b9SSidhartha Kumar hugetlb_set_folio_subpool(folio, NULL); 19100356c4b9SSidhartha Kumar if (folio_test_anon(folio)) 19110356c4b9SSidhartha Kumar __ClearPageAnonExclusive(&folio->page); 19120356c4b9SSidhartha Kumar folio->mapping = NULL; 19130356c4b9SSidhartha Kumar restore_reserve = folio_test_hugetlb_restore_reserve(folio); 19140356c4b9SSidhartha Kumar folio_clear_hugetlb_restore_reserve(folio); 191527a85ef1SDavid Gibson 19161c5ecae3SMike Kravetz /* 1917d6995da3SMike Kravetz * If HPageRestoreReserve was set on page, page allocation consumed a 19180919e1b6SMike Kravetz * reservation. If the page was associated with a subpool, there 19190919e1b6SMike Kravetz * would have been a page reserved in the subpool before allocation 19200919e1b6SMike Kravetz * via hugepage_subpool_get_pages(). Since we are 'restoring' the 19216c26d310SMiaohe Lin * reservation, do not call hugepage_subpool_put_pages() as this will 19220919e1b6SMike Kravetz * remove the reserved page from the subpool. 19230919e1b6SMike Kravetz */ 19240919e1b6SMike Kravetz if (!restore_reserve) { 19250919e1b6SMike Kravetz /* 19260919e1b6SMike Kravetz * A return code of zero implies that the subpool will be 19270919e1b6SMike Kravetz * under its minimum size if the reservation is not restored 19280919e1b6SMike Kravetz * after page is free. Therefore, force restore_reserve 19290919e1b6SMike Kravetz * operation. 19301c5ecae3SMike Kravetz */ 19311c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 19321c5ecae3SMike Kravetz restore_reserve = true; 19330919e1b6SMike Kravetz } 19341c5ecae3SMike Kravetz 1935db71ef79SMike Kravetz spin_lock_irqsave(&hugetlb_lock, flags); 19360356c4b9SSidhartha Kumar folio_clear_hugetlb_migratable(folio); 1937d4ab0316SSidhartha Kumar hugetlb_cgroup_uncharge_folio(hstate_index(h), 1938d4ab0316SSidhartha Kumar pages_per_huge_page(h), folio); 1939d4ab0316SSidhartha Kumar hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h), 1940d4ab0316SSidhartha Kumar pages_per_huge_page(h), folio); 194107443a85SJoonsoo Kim if (restore_reserve) 194207443a85SJoonsoo Kim h->resv_huge_pages++; 194307443a85SJoonsoo Kim 19440356c4b9SSidhartha Kumar if (folio_test_hugetlb_temporary(folio)) { 1945cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, false); 1946db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1947d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, folio, true); 1948ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 19490edaecfaSAneesh Kumar K.V /* remove the page from active list */ 1950cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, true); 1951db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1952d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, folio, true); 19537893d1d5SAdam Litke } else { 1954454a00c4SMatthew Wilcox (Oracle) arch_clear_hugepage_flags(&folio->page); 1955240d67a8SSidhartha Kumar enqueue_hugetlb_folio(h, folio); 1956db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 195727a85ef1SDavid Gibson } 19581121828aSMike Kravetz } 195927a85ef1SDavid Gibson 1960d3d99fccSOscar Salvador /* 1961d3d99fccSOscar Salvador * Must be called with the hugetlb lock held 1962d3d99fccSOscar Salvador */ 1963d3d99fccSOscar Salvador static void __prep_account_new_huge_page(struct hstate *h, int nid) 1964d3d99fccSOscar Salvador { 1965d3d99fccSOscar Salvador lockdep_assert_held(&hugetlb_lock); 1966d3d99fccSOscar Salvador h->nr_huge_pages++; 1967d3d99fccSOscar Salvador h->nr_huge_pages_node[nid]++; 1968d3d99fccSOscar Salvador } 1969d3d99fccSOscar Salvador 1970de656ed3SSidhartha Kumar static void __prep_new_hugetlb_folio(struct hstate *h, struct folio *folio) 1971b7ba30c6SAndi Kleen { 1972de656ed3SSidhartha Kumar hugetlb_vmemmap_optimize(h, &folio->page); 1973de656ed3SSidhartha Kumar INIT_LIST_HEAD(&folio->lru); 19742431b5f2SMatthew Wilcox (Oracle) __folio_set_hugetlb(folio); 1975de656ed3SSidhartha Kumar hugetlb_set_folio_subpool(folio, NULL); 1976de656ed3SSidhartha Kumar set_hugetlb_cgroup(folio, NULL); 1977de656ed3SSidhartha Kumar set_hugetlb_cgroup_rsvd(folio, NULL); 1978d3d99fccSOscar Salvador } 1979d3d99fccSOscar Salvador 1980d1c60955SSidhartha Kumar static void prep_new_hugetlb_folio(struct hstate *h, struct folio *folio, int nid) 1981d3d99fccSOscar Salvador { 1982de656ed3SSidhartha Kumar __prep_new_hugetlb_folio(h, folio); 1983db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 1984d3d99fccSOscar Salvador __prep_account_new_huge_page(h, nid); 1985db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1986b7ba30c6SAndi Kleen } 1987b7ba30c6SAndi Kleen 1988d1c60955SSidhartha Kumar static bool __prep_compound_gigantic_folio(struct folio *folio, 1989d1c60955SSidhartha Kumar unsigned int order, bool demote) 199020a0307cSWu Fengguang { 19917118fc29SMike Kravetz int i, j; 199220a0307cSWu Fengguang int nr_pages = 1 << order; 199314455eabSCheng Li struct page *p; 199420a0307cSWu Fengguang 1995d1c60955SSidhartha Kumar __folio_clear_reserved(folio); 19962b21624fSMike Kravetz for (i = 0; i < nr_pages; i++) { 1997d1c60955SSidhartha Kumar p = folio_page(folio, i); 199814455eabSCheng Li 1999ef5a22beSAndrea Arcangeli /* 2000ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 2001ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 2002ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 20037c8de358SEthon Paul * too. Otherwise drivers using get_user_pages() to access tail 2004ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 2005ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 2006ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 2007ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 2008ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 2009ef5a22beSAndrea Arcangeli * after get_user_pages(). 2010ef5a22beSAndrea Arcangeli */ 20117fb0728aSMike Kravetz if (i != 0) /* head page cleared above */ 2012ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 20137118fc29SMike Kravetz /* 20147118fc29SMike Kravetz * Subtle and very unlikely 20157118fc29SMike Kravetz * 20167118fc29SMike Kravetz * Gigantic 'page allocators' such as memblock or cma will 20177118fc29SMike Kravetz * return a set of pages with each page ref counted. We need 20187118fc29SMike Kravetz * to turn this set of pages into a compound page with tail 20197118fc29SMike Kravetz * page ref counts set to zero. Code such as speculative page 20207118fc29SMike Kravetz * cache adding could take a ref on a 'to be' tail page. 20217118fc29SMike Kravetz * We need to respect any increased ref count, and only set 20227118fc29SMike Kravetz * the ref count to zero if count is currently 1. If count 2023416d85edSMike Kravetz * is not 1, we return an error. An error return indicates 2024416d85edSMike Kravetz * the set of pages can not be converted to a gigantic page. 2025416d85edSMike Kravetz * The caller who allocated the pages should then discard the 2026416d85edSMike Kravetz * pages using the appropriate free interface. 202734d9e35bSMike Kravetz * 202834d9e35bSMike Kravetz * In the case of demote, the ref count will be zero. 20297118fc29SMike Kravetz */ 203034d9e35bSMike Kravetz if (!demote) { 20317118fc29SMike Kravetz if (!page_ref_freeze(p, 1)) { 2032416d85edSMike Kravetz pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n"); 20337118fc29SMike Kravetz goto out_error; 20347118fc29SMike Kravetz } 203534d9e35bSMike Kravetz } else { 203634d9e35bSMike Kravetz VM_BUG_ON_PAGE(page_count(p), p); 203734d9e35bSMike Kravetz } 20382b21624fSMike Kravetz if (i != 0) 2039d1c60955SSidhartha Kumar set_compound_head(p, &folio->page); 204020a0307cSWu Fengguang } 2041e3b7bf97STarun Sahu __folio_set_head(folio); 2042e3b7bf97STarun Sahu /* we rely on prep_new_hugetlb_folio to set the destructor */ 2043e3b7bf97STarun Sahu folio_set_order(folio, order); 204446f27228SMatthew Wilcox (Oracle) atomic_set(&folio->_entire_mapcount, -1); 2045eec20426SMatthew Wilcox (Oracle) atomic_set(&folio->_nr_pages_mapped, 0); 204694688e8eSMatthew Wilcox (Oracle) atomic_set(&folio->_pincount, 0); 20477118fc29SMike Kravetz return true; 20487118fc29SMike Kravetz 20497118fc29SMike Kravetz out_error: 20502b21624fSMike Kravetz /* undo page modifications made above */ 20512b21624fSMike Kravetz for (j = 0; j < i; j++) { 2052d1c60955SSidhartha Kumar p = folio_page(folio, j); 20532b21624fSMike Kravetz if (j != 0) 20547118fc29SMike Kravetz clear_compound_head(p); 20557118fc29SMike Kravetz set_page_refcounted(p); 20567118fc29SMike Kravetz } 20577118fc29SMike Kravetz /* need to clear PG_reserved on remaining tail pages */ 205814455eabSCheng Li for (; j < nr_pages; j++) { 2059d1c60955SSidhartha Kumar p = folio_page(folio, j); 20607118fc29SMike Kravetz __ClearPageReserved(p); 206114455eabSCheng Li } 20627118fc29SMike Kravetz return false; 206320a0307cSWu Fengguang } 206420a0307cSWu Fengguang 2065d1c60955SSidhartha Kumar static bool prep_compound_gigantic_folio(struct folio *folio, 20668531fc6fSMike Kravetz unsigned int order) 20678531fc6fSMike Kravetz { 2068d1c60955SSidhartha Kumar return __prep_compound_gigantic_folio(folio, order, false); 2069d1c60955SSidhartha Kumar } 2070d1c60955SSidhartha Kumar 2071d1c60955SSidhartha Kumar static bool prep_compound_gigantic_folio_for_demote(struct folio *folio, 2072d1c60955SSidhartha Kumar unsigned int order) 2073d1c60955SSidhartha Kumar { 2074d1c60955SSidhartha Kumar return __prep_compound_gigantic_folio(folio, order, true); 20758531fc6fSMike Kravetz } 20768531fc6fSMike Kravetz 20777795912cSAndrew Morton /* 2078c0d0381aSMike Kravetz * Find and lock address space (mapping) in write mode. 2079c0d0381aSMike Kravetz * 2080336bf30eSMike Kravetz * Upon entry, the page is locked which means that page_mapping() is 2081336bf30eSMike Kravetz * stable. Due to locking order, we can only trylock_write. If we can 2082336bf30eSMike Kravetz * not get the lock, simply return NULL to caller. 2083c0d0381aSMike Kravetz */ 2084c0d0381aSMike Kravetz struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) 2085c0d0381aSMike Kravetz { 2086336bf30eSMike Kravetz struct address_space *mapping = page_mapping(hpage); 2087c0d0381aSMike Kravetz 2088c0d0381aSMike Kravetz if (!mapping) 2089c0d0381aSMike Kravetz return mapping; 2090c0d0381aSMike Kravetz 2091c0d0381aSMike Kravetz if (i_mmap_trylock_write(mapping)) 2092c0d0381aSMike Kravetz return mapping; 2093c0d0381aSMike Kravetz 2094c0d0381aSMike Kravetz return NULL; 2095c0d0381aSMike Kravetz } 2096c0d0381aSMike Kravetz 2097fe19bd3dSHugh Dickins pgoff_t hugetlb_basepage_index(struct page *page) 209813d60f4bSZhang Yi { 209913d60f4bSZhang Yi struct page *page_head = compound_head(page); 210013d60f4bSZhang Yi pgoff_t index = page_index(page_head); 210113d60f4bSZhang Yi unsigned long compound_idx; 210213d60f4bSZhang Yi 210323baf831SKirill A. Shutemov if (compound_order(page_head) > MAX_ORDER) 210413d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 210513d60f4bSZhang Yi else 210613d60f4bSZhang Yi compound_idx = page - page_head; 210713d60f4bSZhang Yi 210813d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 210913d60f4bSZhang Yi } 211013d60f4bSZhang Yi 211119fc1a7eSSidhartha Kumar static struct folio *alloc_buddy_hugetlb_folio(struct hstate *h, 2112f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 2113f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 21141da177e4SLinus Torvalds { 2115af0fb9dfSMichal Hocko int order = huge_page_order(h); 21161da177e4SLinus Torvalds struct page *page; 2117f60858f9SMike Kravetz bool alloc_try_hard = true; 21182b21624fSMike Kravetz bool retry = true; 2119f96efd58SJoe Jin 2120f60858f9SMike Kravetz /* 2121f60858f9SMike Kravetz * By default we always try hard to allocate the page with 2122f60858f9SMike Kravetz * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in 2123f60858f9SMike Kravetz * a loop (to adjust global huge page counts) and previous allocation 2124f60858f9SMike Kravetz * failed, do not continue to try hard on the same node. Use the 2125f60858f9SMike Kravetz * node_alloc_noretry bitmap to manage this state information. 2126f60858f9SMike Kravetz */ 2127f60858f9SMike Kravetz if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) 2128f60858f9SMike Kravetz alloc_try_hard = false; 2129f60858f9SMike Kravetz gfp_mask |= __GFP_COMP|__GFP_NOWARN; 2130f60858f9SMike Kravetz if (alloc_try_hard) 2131f60858f9SMike Kravetz gfp_mask |= __GFP_RETRY_MAYFAIL; 2132af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 2133af0fb9dfSMichal Hocko nid = numa_mem_id(); 21342b21624fSMike Kravetz retry: 213584172f4bSMatthew Wilcox (Oracle) page = __alloc_pages(gfp_mask, order, nid, nmask); 21362b21624fSMike Kravetz 21372b21624fSMike Kravetz /* Freeze head page */ 21382b21624fSMike Kravetz if (page && !page_ref_freeze(page, 1)) { 21392b21624fSMike Kravetz __free_pages(page, order); 21402b21624fSMike Kravetz if (retry) { /* retry once */ 21412b21624fSMike Kravetz retry = false; 21422b21624fSMike Kravetz goto retry; 21432b21624fSMike Kravetz } 21442b21624fSMike Kravetz /* WOW! twice in a row. */ 21452b21624fSMike Kravetz pr_warn("HugeTLB head page unexpected inflated ref count\n"); 21462b21624fSMike Kravetz page = NULL; 21472b21624fSMike Kravetz } 21482b21624fSMike Kravetz 2149f60858f9SMike Kravetz /* 2150f60858f9SMike Kravetz * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this 2151f60858f9SMike Kravetz * indicates an overall state change. Clear bit so that we resume 2152f60858f9SMike Kravetz * normal 'try hard' allocations. 2153f60858f9SMike Kravetz */ 2154f60858f9SMike Kravetz if (node_alloc_noretry && page && !alloc_try_hard) 2155f60858f9SMike Kravetz node_clear(nid, *node_alloc_noretry); 2156f60858f9SMike Kravetz 2157f60858f9SMike Kravetz /* 2158f60858f9SMike Kravetz * If we tried hard to get a page but failed, set bit so that 2159f60858f9SMike Kravetz * subsequent attempts will not try as hard until there is an 2160f60858f9SMike Kravetz * overall state change. 2161f60858f9SMike Kravetz */ 2162f60858f9SMike Kravetz if (node_alloc_noretry && !page && alloc_try_hard) 2163f60858f9SMike Kravetz node_set(nid, *node_alloc_noretry); 2164f60858f9SMike Kravetz 216519fc1a7eSSidhartha Kumar if (!page) { 216619fc1a7eSSidhartha Kumar __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 216719fc1a7eSSidhartha Kumar return NULL; 216819fc1a7eSSidhartha Kumar } 216919fc1a7eSSidhartha Kumar 217019fc1a7eSSidhartha Kumar __count_vm_event(HTLB_BUDDY_PGALLOC); 217119fc1a7eSSidhartha Kumar return page_folio(page); 217263b4613cSNishanth Aravamudan } 217363b4613cSNishanth Aravamudan 2174af0fb9dfSMichal Hocko /* 21750c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 21760c397daeSMichal Hocko * should use this function to get new hugetlb pages 21772b21624fSMike Kravetz * 21782b21624fSMike Kravetz * Note that returned page is 'frozen': ref count of head page and all tail 21792b21624fSMike Kravetz * pages is zero. 21800c397daeSMichal Hocko */ 218119fc1a7eSSidhartha Kumar static struct folio *alloc_fresh_hugetlb_folio(struct hstate *h, 2182f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 2183f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 21840c397daeSMichal Hocko { 21857f325a8dSSidhartha Kumar struct folio *folio; 21867118fc29SMike Kravetz bool retry = false; 21870c397daeSMichal Hocko 21887118fc29SMike Kravetz retry: 21890c397daeSMichal Hocko if (hstate_is_gigantic(h)) 219019fc1a7eSSidhartha Kumar folio = alloc_gigantic_folio(h, gfp_mask, nid, nmask); 21910c397daeSMichal Hocko else 219219fc1a7eSSidhartha Kumar folio = alloc_buddy_hugetlb_folio(h, gfp_mask, 2193f60858f9SMike Kravetz nid, nmask, node_alloc_noretry); 219419fc1a7eSSidhartha Kumar if (!folio) 21950c397daeSMichal Hocko return NULL; 21967118fc29SMike Kravetz if (hstate_is_gigantic(h)) { 2197d1c60955SSidhartha Kumar if (!prep_compound_gigantic_folio(folio, huge_page_order(h))) { 21987118fc29SMike Kravetz /* 21997118fc29SMike Kravetz * Rare failure to convert pages to compound page. 22007118fc29SMike Kravetz * Free pages and try again - ONCE! 22017118fc29SMike Kravetz */ 22027f325a8dSSidhartha Kumar free_gigantic_folio(folio, huge_page_order(h)); 22037118fc29SMike Kravetz if (!retry) { 22047118fc29SMike Kravetz retry = true; 22057118fc29SMike Kravetz goto retry; 22067118fc29SMike Kravetz } 22077118fc29SMike Kravetz return NULL; 22087118fc29SMike Kravetz } 22097118fc29SMike Kravetz } 2210d1c60955SSidhartha Kumar prep_new_hugetlb_folio(h, folio, folio_nid(folio)); 22110c397daeSMichal Hocko 221219fc1a7eSSidhartha Kumar return folio; 22130c397daeSMichal Hocko } 22140c397daeSMichal Hocko 22150c397daeSMichal Hocko /* 2216af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 2217af0fb9dfSMichal Hocko * manner. 2218af0fb9dfSMichal Hocko */ 2219f60858f9SMike Kravetz static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 2220f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 2221b2261026SJoonsoo Kim { 222219fc1a7eSSidhartha Kumar struct folio *folio; 2223b2261026SJoonsoo Kim int nr_nodes, node; 2224af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2225b2261026SJoonsoo Kim 2226b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 222719fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, node, 222819fc1a7eSSidhartha Kumar nodes_allowed, node_alloc_noretry); 222919fc1a7eSSidhartha Kumar if (folio) { 2230454a00c4SMatthew Wilcox (Oracle) free_huge_folio(folio); /* free it into the hugepage allocator */ 223119fc1a7eSSidhartha Kumar return 1; 223219fc1a7eSSidhartha Kumar } 2233b2261026SJoonsoo Kim } 2234b2261026SJoonsoo Kim 2235af0fb9dfSMichal Hocko return 0; 2236b2261026SJoonsoo Kim } 2237b2261026SJoonsoo Kim 2238e8c5c824SLee Schermerhorn /* 223910c6ec49SMike Kravetz * Remove huge page from pool from next node to free. Attempt to keep 224010c6ec49SMike Kravetz * persistent huge pages more or less balanced over allowed nodes. 224110c6ec49SMike Kravetz * This routine only 'removes' the hugetlb page. The caller must make 224210c6ec49SMike Kravetz * an additional call to free the page to low level allocators. 2243e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 2244e8c5c824SLee Schermerhorn */ 224510c6ec49SMike Kravetz static struct page *remove_pool_huge_page(struct hstate *h, 224610c6ec49SMike Kravetz nodemask_t *nodes_allowed, 22476ae11b27SLee Schermerhorn bool acct_surplus) 2248e8c5c824SLee Schermerhorn { 2249b2261026SJoonsoo Kim int nr_nodes, node; 225010c6ec49SMike Kravetz struct page *page = NULL; 2251cfd5082bSSidhartha Kumar struct folio *folio; 2252e8c5c824SLee Schermerhorn 22539487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2254b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2255685f3457SLee Schermerhorn /* 2256685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 2257685f3457SLee Schermerhorn * nodes with surplus pages. 2258685f3457SLee Schermerhorn */ 2259b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 2260b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 226110c6ec49SMike Kravetz page = list_entry(h->hugepage_freelists[node].next, 2262e8c5c824SLee Schermerhorn struct page, lru); 2263cfd5082bSSidhartha Kumar folio = page_folio(page); 2264cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, acct_surplus); 22659a76db09SLee Schermerhorn break; 2266e8c5c824SLee Schermerhorn } 2267b2261026SJoonsoo Kim } 2268e8c5c824SLee Schermerhorn 226910c6ec49SMike Kravetz return page; 2270e8c5c824SLee Schermerhorn } 2271e8c5c824SLee Schermerhorn 2272c8721bbbSNaoya Horiguchi /* 2273c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 2274faf53defSNaoya Horiguchi * nothing for in-use hugepages and non-hugepages. 2275faf53defSNaoya Horiguchi * This function returns values like below: 2276faf53defSNaoya Horiguchi * 2277ad2fa371SMuchun Song * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages 2278ad2fa371SMuchun Song * when the system is under memory pressure and the feature of 2279ad2fa371SMuchun Song * freeing unused vmemmap pages associated with each hugetlb page 2280ad2fa371SMuchun Song * is enabled. 2281faf53defSNaoya Horiguchi * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 2282faf53defSNaoya Horiguchi * (allocated or reserved.) 2283faf53defSNaoya Horiguchi * 0: successfully dissolved free hugepages or the page is not a 2284faf53defSNaoya Horiguchi * hugepage (considered as already dissolved) 2285c8721bbbSNaoya Horiguchi */ 2286c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 2287c8721bbbSNaoya Horiguchi { 22886bc9b564SNaoya Horiguchi int rc = -EBUSY; 22891a7cdab5SSidhartha Kumar struct folio *folio = page_folio(page); 2290082d5b6bSGerald Schaefer 22917ffddd49SMuchun Song retry: 2292faf53defSNaoya Horiguchi /* Not to disrupt normal path by vainly holding hugetlb_lock */ 22931a7cdab5SSidhartha Kumar if (!folio_test_hugetlb(folio)) 2294faf53defSNaoya Horiguchi return 0; 2295faf53defSNaoya Horiguchi 2296db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 22971a7cdab5SSidhartha Kumar if (!folio_test_hugetlb(folio)) { 2298faf53defSNaoya Horiguchi rc = 0; 2299faf53defSNaoya Horiguchi goto out; 2300faf53defSNaoya Horiguchi } 2301faf53defSNaoya Horiguchi 23021a7cdab5SSidhartha Kumar if (!folio_ref_count(folio)) { 23031a7cdab5SSidhartha Kumar struct hstate *h = folio_hstate(folio); 23048346d69dSXin Hao if (!available_huge_pages(h)) 2305082d5b6bSGerald Schaefer goto out; 23067ffddd49SMuchun Song 23077ffddd49SMuchun Song /* 23087ffddd49SMuchun Song * We should make sure that the page is already on the free list 23097ffddd49SMuchun Song * when it is dissolved. 23107ffddd49SMuchun Song */ 23111a7cdab5SSidhartha Kumar if (unlikely(!folio_test_hugetlb_freed(folio))) { 2312db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 23137ffddd49SMuchun Song cond_resched(); 23147ffddd49SMuchun Song 23157ffddd49SMuchun Song /* 23167ffddd49SMuchun Song * Theoretically, we should return -EBUSY when we 23177ffddd49SMuchun Song * encounter this race. In fact, we have a chance 23187ffddd49SMuchun Song * to successfully dissolve the page if we do a 23197ffddd49SMuchun Song * retry. Because the race window is quite small. 23207ffddd49SMuchun Song * If we seize this opportunity, it is an optimization 23217ffddd49SMuchun Song * for increasing the success rate of dissolving page. 23227ffddd49SMuchun Song */ 23237ffddd49SMuchun Song goto retry; 23247ffddd49SMuchun Song } 23257ffddd49SMuchun Song 2326cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, false); 2327ad2fa371SMuchun Song h->max_huge_pages--; 2328ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2329ad2fa371SMuchun Song 2330c3114a84SAnshuman Khandual /* 2331d6ef19e2SSidhartha Kumar * Normally update_and_free_hugtlb_folio will allocate required vmemmmap 2332d6ef19e2SSidhartha Kumar * before freeing the page. update_and_free_hugtlb_folio will fail to 2333ad2fa371SMuchun Song * free the page if it can not allocate required vmemmap. We 2334ad2fa371SMuchun Song * need to adjust max_huge_pages if the page is not freed. 2335ad2fa371SMuchun Song * Attempt to allocate vmemmmap here so that we can take 2336ad2fa371SMuchun Song * appropriate action on failure. 2337ad2fa371SMuchun Song */ 23381a7cdab5SSidhartha Kumar rc = hugetlb_vmemmap_restore(h, &folio->page); 2339ad2fa371SMuchun Song if (!rc) { 2340d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, folio, false); 2341ad2fa371SMuchun Song } else { 2342ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 23432f6c57d6SSidhartha Kumar add_hugetlb_folio(h, folio, false); 2344ad2fa371SMuchun Song h->max_huge_pages++; 2345ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2346ad2fa371SMuchun Song } 2347ad2fa371SMuchun Song 2348ad2fa371SMuchun Song return rc; 2349c8721bbbSNaoya Horiguchi } 2350082d5b6bSGerald Schaefer out: 2351db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2352082d5b6bSGerald Schaefer return rc; 2353c8721bbbSNaoya Horiguchi } 2354c8721bbbSNaoya Horiguchi 2355c8721bbbSNaoya Horiguchi /* 2356c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 2357c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 23582247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 23592247bb33SGerald Schaefer * part of it lies within the given range. 2360082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 2361082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 2362c8721bbbSNaoya Horiguchi */ 2363082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 2364c8721bbbSNaoya Horiguchi { 2365c8721bbbSNaoya Horiguchi unsigned long pfn; 2366eb03aa00SGerald Schaefer struct page *page; 2367082d5b6bSGerald Schaefer int rc = 0; 2368dc2628f3SMuchun Song unsigned int order; 2369dc2628f3SMuchun Song struct hstate *h; 2370c8721bbbSNaoya Horiguchi 2371d0177639SLi Zhong if (!hugepages_supported()) 2372082d5b6bSGerald Schaefer return rc; 2373d0177639SLi Zhong 2374dc2628f3SMuchun Song order = huge_page_order(&default_hstate); 2375dc2628f3SMuchun Song for_each_hstate(h) 2376dc2628f3SMuchun Song order = min(order, huge_page_order(h)); 2377dc2628f3SMuchun Song 2378dc2628f3SMuchun Song for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) { 2379eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 2380eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 2381eb03aa00SGerald Schaefer if (rc) 2382082d5b6bSGerald Schaefer break; 2383eb03aa00SGerald Schaefer } 2384082d5b6bSGerald Schaefer 2385082d5b6bSGerald Schaefer return rc; 2386c8721bbbSNaoya Horiguchi } 2387c8721bbbSNaoya Horiguchi 2388ab5ac90aSMichal Hocko /* 2389ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 2390ab5ac90aSMichal Hocko */ 23913a740e8bSSidhartha Kumar static struct folio *alloc_surplus_hugetlb_folio(struct hstate *h, 23923a740e8bSSidhartha Kumar gfp_t gfp_mask, int nid, nodemask_t *nmask) 23937893d1d5SAdam Litke { 239419fc1a7eSSidhartha Kumar struct folio *folio = NULL; 23957893d1d5SAdam Litke 2396bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2397aa888a74SAndi Kleen return NULL; 2398aa888a74SAndi Kleen 2399db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 24009980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 24019980d744SMichal Hocko goto out_unlock; 2402db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2403d1c3fb1fSNishanth Aravamudan 240419fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL); 240519fc1a7eSSidhartha Kumar if (!folio) 24060c397daeSMichal Hocko return NULL; 2407d1c3fb1fSNishanth Aravamudan 2408db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 24099980d744SMichal Hocko /* 24109980d744SMichal Hocko * We could have raced with the pool size change. 24119980d744SMichal Hocko * Double check that and simply deallocate the new page 24129980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 2413454a00c4SMatthew Wilcox (Oracle) * temporary page to workaround the nasty free_huge_folio 24149980d744SMichal Hocko * codeflow 24159980d744SMichal Hocko */ 24169980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 241719fc1a7eSSidhartha Kumar folio_set_hugetlb_temporary(folio); 2418db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2419454a00c4SMatthew Wilcox (Oracle) free_huge_folio(folio); 24202bf753e6SKai Shen return NULL; 2421b65a4edaSMike Kravetz } 2422b65a4edaSMike Kravetz 24239980d744SMichal Hocko h->surplus_huge_pages++; 242419fc1a7eSSidhartha Kumar h->surplus_huge_pages_node[folio_nid(folio)]++; 24259980d744SMichal Hocko 24269980d744SMichal Hocko out_unlock: 2427db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 24287893d1d5SAdam Litke 24293a740e8bSSidhartha Kumar return folio; 24307893d1d5SAdam Litke } 24317893d1d5SAdam Litke 2432e37d3e83SSidhartha Kumar static struct folio *alloc_migrate_hugetlb_folio(struct hstate *h, gfp_t gfp_mask, 2433ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 2434ab5ac90aSMichal Hocko { 243519fc1a7eSSidhartha Kumar struct folio *folio; 2436ab5ac90aSMichal Hocko 2437ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 2438ab5ac90aSMichal Hocko return NULL; 2439ab5ac90aSMichal Hocko 244019fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL); 244119fc1a7eSSidhartha Kumar if (!folio) 2442ab5ac90aSMichal Hocko return NULL; 2443ab5ac90aSMichal Hocko 24442b21624fSMike Kravetz /* fresh huge pages are frozen */ 244519fc1a7eSSidhartha Kumar folio_ref_unfreeze(folio, 1); 2446ab5ac90aSMichal Hocko /* 2447ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 2448ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 2449ab5ac90aSMichal Hocko */ 245019fc1a7eSSidhartha Kumar folio_set_hugetlb_temporary(folio); 2451ab5ac90aSMichal Hocko 2452e37d3e83SSidhartha Kumar return folio; 2453ab5ac90aSMichal Hocko } 2454ab5ac90aSMichal Hocko 2455e4e574b7SAdam Litke /* 2456099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 2457099730d6SDave Hansen */ 2458e0ec90eeSDave Hansen static 2459ff7d853bSSidhartha Kumar struct folio *alloc_buddy_hugetlb_folio_with_mpol(struct hstate *h, 2460099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 2461099730d6SDave Hansen { 24623a740e8bSSidhartha Kumar struct folio *folio = NULL; 2463aaf14e40SMichal Hocko struct mempolicy *mpol; 2464aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 2465aaf14e40SMichal Hocko int nid; 2466aaf14e40SMichal Hocko nodemask_t *nodemask; 2467aaf14e40SMichal Hocko 2468aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 2469cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 2470cfcaa66fSBen Widawsky gfp_t gfp = gfp_mask | __GFP_NOWARN; 2471cfcaa66fSBen Widawsky 2472cfcaa66fSBen Widawsky gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL); 24733a740e8bSSidhartha Kumar folio = alloc_surplus_hugetlb_folio(h, gfp, nid, nodemask); 2474cfcaa66fSBen Widawsky 2475cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 2476cfcaa66fSBen Widawsky nodemask = NULL; 2477cfcaa66fSBen Widawsky } 2478cfcaa66fSBen Widawsky 24793a740e8bSSidhartha Kumar if (!folio) 24803a740e8bSSidhartha Kumar folio = alloc_surplus_hugetlb_folio(h, gfp_mask, nid, nodemask); 2481aaf14e40SMichal Hocko mpol_cond_put(mpol); 2482ff7d853bSSidhartha Kumar return folio; 2483099730d6SDave Hansen } 2484099730d6SDave Hansen 2485e37d3e83SSidhartha Kumar /* folio migration callback function */ 2486e37d3e83SSidhartha Kumar struct folio *alloc_hugetlb_folio_nodemask(struct hstate *h, int preferred_nid, 2487d92bbc27SJoonsoo Kim nodemask_t *nmask, gfp_t gfp_mask) 24884db9b2efSMichal Hocko { 2489db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 24908346d69dSXin Hao if (available_huge_pages(h)) { 2491a36f1e90SSidhartha Kumar struct folio *folio; 24923e59fcb0SMichal Hocko 2493a36f1e90SSidhartha Kumar folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask, 2494a36f1e90SSidhartha Kumar preferred_nid, nmask); 2495a36f1e90SSidhartha Kumar if (folio) { 2496db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2497e37d3e83SSidhartha Kumar return folio; 24984db9b2efSMichal Hocko } 24994db9b2efSMichal Hocko } 2500db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 25014db9b2efSMichal Hocko 2502e37d3e83SSidhartha Kumar return alloc_migrate_hugetlb_folio(h, gfp_mask, preferred_nid, nmask); 25034db9b2efSMichal Hocko } 25044db9b2efSMichal Hocko 2505ebd63723SMichal Hocko /* mempolicy aware migration callback */ 2506d0ce0e47SSidhartha Kumar struct folio *alloc_hugetlb_folio_vma(struct hstate *h, struct vm_area_struct *vma, 2507389c8178SMichal Hocko unsigned long address) 2508ebd63723SMichal Hocko { 2509ebd63723SMichal Hocko struct mempolicy *mpol; 2510ebd63723SMichal Hocko nodemask_t *nodemask; 2511e37d3e83SSidhartha Kumar struct folio *folio; 2512ebd63723SMichal Hocko gfp_t gfp_mask; 2513ebd63723SMichal Hocko int node; 2514ebd63723SMichal Hocko 2515ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 2516ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 2517e37d3e83SSidhartha Kumar folio = alloc_hugetlb_folio_nodemask(h, node, nodemask, gfp_mask); 2518ebd63723SMichal Hocko mpol_cond_put(mpol); 2519ebd63723SMichal Hocko 2520d0ce0e47SSidhartha Kumar return folio; 2521ebd63723SMichal Hocko } 2522ebd63723SMichal Hocko 2523bf50bab2SNaoya Horiguchi /* 252425985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 2525e4e574b7SAdam Litke * of size 'delta'. 2526e4e574b7SAdam Litke */ 25270a4f3d1bSLiu Xiang static int gather_surplus_pages(struct hstate *h, long delta) 25281b2a1e7bSJules Irenge __must_hold(&hugetlb_lock) 2529e4e574b7SAdam Litke { 253034665341SMiaohe Lin LIST_HEAD(surplus_list); 2531454a00c4SMatthew Wilcox (Oracle) struct folio *folio, *tmp; 25320a4f3d1bSLiu Xiang int ret; 25330a4f3d1bSLiu Xiang long i; 25340a4f3d1bSLiu Xiang long needed, allocated; 253528073b02SHillf Danton bool alloc_ok = true; 2536e4e574b7SAdam Litke 25379487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2538a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 2539ac09b3a1SAdam Litke if (needed <= 0) { 2540a5516438SAndi Kleen h->resv_huge_pages += delta; 2541e4e574b7SAdam Litke return 0; 2542ac09b3a1SAdam Litke } 2543e4e574b7SAdam Litke 2544e4e574b7SAdam Litke allocated = 0; 2545e4e574b7SAdam Litke 2546e4e574b7SAdam Litke ret = -ENOMEM; 2547e4e574b7SAdam Litke retry: 2548db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2549e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 25503a740e8bSSidhartha Kumar folio = alloc_surplus_hugetlb_folio(h, htlb_alloc_mask(h), 25512b21624fSMike Kravetz NUMA_NO_NODE, NULL); 25523a740e8bSSidhartha Kumar if (!folio) { 255328073b02SHillf Danton alloc_ok = false; 255428073b02SHillf Danton break; 255528073b02SHillf Danton } 25563a740e8bSSidhartha Kumar list_add(&folio->lru, &surplus_list); 255769ed779aSDavid Rientjes cond_resched(); 2558e4e574b7SAdam Litke } 255928073b02SHillf Danton allocated += i; 2560e4e574b7SAdam Litke 2561e4e574b7SAdam Litke /* 2562e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 2563e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 2564e4e574b7SAdam Litke */ 2565db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2566a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 2567a5516438SAndi Kleen (h->free_huge_pages + allocated); 256828073b02SHillf Danton if (needed > 0) { 256928073b02SHillf Danton if (alloc_ok) 2570e4e574b7SAdam Litke goto retry; 257128073b02SHillf Danton /* 257228073b02SHillf Danton * We were not able to allocate enough pages to 257328073b02SHillf Danton * satisfy the entire reservation so we free what 257428073b02SHillf Danton * we've allocated so far. 257528073b02SHillf Danton */ 257628073b02SHillf Danton goto free; 257728073b02SHillf Danton } 2578e4e574b7SAdam Litke /* 2579e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 258025985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 2581e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 2582ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 2583ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 2584ac09b3a1SAdam Litke * before they are reserved. 2585e4e574b7SAdam Litke */ 2586e4e574b7SAdam Litke needed += allocated; 2587a5516438SAndi Kleen h->resv_huge_pages += delta; 2588e4e574b7SAdam Litke ret = 0; 2589a9869b83SNaoya Horiguchi 259019fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 2591454a00c4SMatthew Wilcox (Oracle) list_for_each_entry_safe(folio, tmp, &surplus_list, lru) { 259219fc3f0aSAdam Litke if ((--needed) < 0) 259319fc3f0aSAdam Litke break; 2594b65a4edaSMike Kravetz /* Add the page to the hugetlb allocator */ 2595454a00c4SMatthew Wilcox (Oracle) enqueue_hugetlb_folio(h, folio); 259619fc3f0aSAdam Litke } 259728073b02SHillf Danton free: 2598db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 259919fc3f0aSAdam Litke 2600b65a4edaSMike Kravetz /* 2601b65a4edaSMike Kravetz * Free unnecessary surplus pages to the buddy allocator. 2602454a00c4SMatthew Wilcox (Oracle) * Pages have no ref count, call free_huge_folio directly. 2603b65a4edaSMike Kravetz */ 2604454a00c4SMatthew Wilcox (Oracle) list_for_each_entry_safe(folio, tmp, &surplus_list, lru) 2605454a00c4SMatthew Wilcox (Oracle) free_huge_folio(folio); 2606db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2607e4e574b7SAdam Litke 2608e4e574b7SAdam Litke return ret; 2609e4e574b7SAdam Litke } 2610e4e574b7SAdam Litke 2611e4e574b7SAdam Litke /* 2612e5bbc8a6SMike Kravetz * This routine has two main purposes: 2613e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 2614e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 2615e5bbc8a6SMike Kravetz * to the associated reservation map. 2616e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 2617e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 2618e4e574b7SAdam Litke */ 2619a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 2620a5516438SAndi Kleen unsigned long unused_resv_pages) 2621e4e574b7SAdam Litke { 2622e4e574b7SAdam Litke unsigned long nr_pages; 262310c6ec49SMike Kravetz struct page *page; 262410c6ec49SMike Kravetz LIST_HEAD(page_list); 262510c6ec49SMike Kravetz 26269487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 262710c6ec49SMike Kravetz /* Uncommit the reservation */ 262810c6ec49SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 2629e4e574b7SAdam Litke 2630c0531714SNaoya Horiguchi if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 2631e5bbc8a6SMike Kravetz goto out; 2632aa888a74SAndi Kleen 2633e5bbc8a6SMike Kravetz /* 2634e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 2635e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 2636e5bbc8a6SMike Kravetz */ 2637a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 2638e4e574b7SAdam Litke 2639685f3457SLee Schermerhorn /* 2640685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 26419b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 26429b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 26439b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 264410c6ec49SMike Kravetz * remove_pool_huge_page() will balance the freed pages across the 26459b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 2646685f3457SLee Schermerhorn */ 2647685f3457SLee Schermerhorn while (nr_pages--) { 264810c6ec49SMike Kravetz page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1); 264910c6ec49SMike Kravetz if (!page) 2650e5bbc8a6SMike Kravetz goto out; 265110c6ec49SMike Kravetz 265210c6ec49SMike Kravetz list_add(&page->lru, &page_list); 2653e4e574b7SAdam Litke } 2654e5bbc8a6SMike Kravetz 2655e5bbc8a6SMike Kravetz out: 2656db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 265710c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 2658db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2659e4e574b7SAdam Litke } 2660e4e574b7SAdam Litke 26615e911373SMike Kravetz 2662c37f9fb1SAndy Whitcroft /* 2663feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 26645e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 2665cf3ad20bSMike Kravetz * 2666cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 2667cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 2668cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 2669cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 2670cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 2671feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 2672feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 2673feba16e2SMike Kravetz * is called in such cases. 2674cf3ad20bSMike Kravetz * 2675cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 2676cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 2677cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 2678cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 2679cf3ad20bSMike Kravetz * take appropriate action. 268096b96a96SMike Kravetz * 268196b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 268296b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 268396b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 268496b96a96SMike Kravetz * reservation exists. 2685846be085SMike Kravetz * 2686846be085SMike Kravetz * vma_del_reservation is used in error paths where an entry in the reserve 2687846be085SMike Kravetz * map was created during huge page allocation and must be removed. It is to 2688846be085SMike Kravetz * be called after calling vma_needs_reservation to determine if a reservation 2689846be085SMike Kravetz * exists. 2690c37f9fb1SAndy Whitcroft */ 26915e911373SMike Kravetz enum vma_resv_mode { 26925e911373SMike Kravetz VMA_NEEDS_RESV, 26935e911373SMike Kravetz VMA_COMMIT_RESV, 2694feba16e2SMike Kravetz VMA_END_RESV, 269596b96a96SMike Kravetz VMA_ADD_RESV, 2696846be085SMike Kravetz VMA_DEL_RESV, 26975e911373SMike Kravetz }; 2698cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 2699cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 27005e911373SMike Kravetz enum vma_resv_mode mode) 2701c37f9fb1SAndy Whitcroft { 27024e35f483SJoonsoo Kim struct resv_map *resv; 27034e35f483SJoonsoo Kim pgoff_t idx; 2704cf3ad20bSMike Kravetz long ret; 27050db9d74eSMina Almasry long dummy_out_regions_needed; 2706c37f9fb1SAndy Whitcroft 27074e35f483SJoonsoo Kim resv = vma_resv_map(vma); 27084e35f483SJoonsoo Kim if (!resv) 2709c37f9fb1SAndy Whitcroft return 1; 2710c37f9fb1SAndy Whitcroft 27114e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 27125e911373SMike Kravetz switch (mode) { 27135e911373SMike Kravetz case VMA_NEEDS_RESV: 27140db9d74eSMina Almasry ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); 27150db9d74eSMina Almasry /* We assume that vma_reservation_* routines always operate on 27160db9d74eSMina Almasry * 1 page, and that adding to resv map a 1 page entry can only 27170db9d74eSMina Almasry * ever require 1 region. 27180db9d74eSMina Almasry */ 27190db9d74eSMina Almasry VM_BUG_ON(dummy_out_regions_needed != 1); 27205e911373SMike Kravetz break; 27215e911373SMike Kravetz case VMA_COMMIT_RESV: 2722075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 27230db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 27240db9d74eSMina Almasry VM_BUG_ON(ret < 0); 27255e911373SMike Kravetz break; 2726feba16e2SMike Kravetz case VMA_END_RESV: 27270db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 27285e911373SMike Kravetz ret = 0; 27295e911373SMike Kravetz break; 273096b96a96SMike Kravetz case VMA_ADD_RESV: 27310db9d74eSMina Almasry if (vma->vm_flags & VM_MAYSHARE) { 2732075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 27330db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 27340db9d74eSMina Almasry VM_BUG_ON(ret < 0); 27350db9d74eSMina Almasry } else { 27360db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 273796b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 273896b96a96SMike Kravetz } 273996b96a96SMike Kravetz break; 2740846be085SMike Kravetz case VMA_DEL_RESV: 2741846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 2742846be085SMike Kravetz region_abort(resv, idx, idx + 1, 1); 2743846be085SMike Kravetz ret = region_del(resv, idx, idx + 1); 2744846be085SMike Kravetz } else { 2745846be085SMike Kravetz ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 2746846be085SMike Kravetz /* region_add calls of range 1 should never fail. */ 2747846be085SMike Kravetz VM_BUG_ON(ret < 0); 2748846be085SMike Kravetz } 2749846be085SMike Kravetz break; 27505e911373SMike Kravetz default: 27515e911373SMike Kravetz BUG(); 27525e911373SMike Kravetz } 275384afd99bSAndy Whitcroft 2754846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV) 2755cf3ad20bSMike Kravetz return ret; 275667961f9dSMike Kravetz /* 2757bf3d12b9SMiaohe Lin * We know private mapping must have HPAGE_RESV_OWNER set. 2758bf3d12b9SMiaohe Lin * 275967961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 276067961f9dSMike Kravetz * However, a file associated with mapping could have been 276167961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 276267961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 276367961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 276467961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 276567961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 276667961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 276767961f9dSMike Kravetz * reservation has already been consumed. As a result, the 276867961f9dSMike Kravetz * return value of this routine is the opposite of the 276967961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 277067961f9dSMike Kravetz */ 2771bf3d12b9SMiaohe Lin if (ret > 0) 277267961f9dSMike Kravetz return 0; 2773bf3d12b9SMiaohe Lin if (ret == 0) 277467961f9dSMike Kravetz return 1; 2775bf3d12b9SMiaohe Lin return ret; 277684afd99bSAndy Whitcroft } 2777cf3ad20bSMike Kravetz 2778cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 2779a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 2780c37f9fb1SAndy Whitcroft { 27815e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 2782cf3ad20bSMike Kravetz } 2783c37f9fb1SAndy Whitcroft 2784cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 2785cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 2786cf3ad20bSMike Kravetz { 27875e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 27885e911373SMike Kravetz } 27895e911373SMike Kravetz 2790feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 27915e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 27925e911373SMike Kravetz { 2793feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 2794c37f9fb1SAndy Whitcroft } 2795c37f9fb1SAndy Whitcroft 279696b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 279796b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 279896b96a96SMike Kravetz { 279996b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 280096b96a96SMike Kravetz } 280196b96a96SMike Kravetz 2802846be085SMike Kravetz static long vma_del_reservation(struct hstate *h, 2803846be085SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 280496b96a96SMike Kravetz { 2805846be085SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV); 2806846be085SMike Kravetz } 2807846be085SMike Kravetz 2808846be085SMike Kravetz /* 2809846be085SMike Kravetz * This routine is called to restore reservation information on error paths. 2810d0ce0e47SSidhartha Kumar * It should ONLY be called for folios allocated via alloc_hugetlb_folio(), 2811d0ce0e47SSidhartha Kumar * and the hugetlb mutex should remain held when calling this routine. 2812846be085SMike Kravetz * 2813846be085SMike Kravetz * It handles two specific cases: 2814d2d7bb44SSidhartha Kumar * 1) A reservation was in place and the folio consumed the reservation. 2815d2d7bb44SSidhartha Kumar * hugetlb_restore_reserve is set in the folio. 2816d2d7bb44SSidhartha Kumar * 2) No reservation was in place for the page, so hugetlb_restore_reserve is 2817d0ce0e47SSidhartha Kumar * not set. However, alloc_hugetlb_folio always updates the reserve map. 2818846be085SMike Kravetz * 2819454a00c4SMatthew Wilcox (Oracle) * In case 1, free_huge_folio later in the error path will increment the 2820454a00c4SMatthew Wilcox (Oracle) * global reserve count. But, free_huge_folio does not have enough context 2821846be085SMike Kravetz * to adjust the reservation map. This case deals primarily with private 2822846be085SMike Kravetz * mappings. Adjust the reserve map here to be consistent with global 2823454a00c4SMatthew Wilcox (Oracle) * reserve count adjustments to be made by free_huge_folio. Make sure the 2824846be085SMike Kravetz * reserve map indicates there is a reservation present. 2825846be085SMike Kravetz * 2826d0ce0e47SSidhartha Kumar * In case 2, simply undo reserve map modifications done by alloc_hugetlb_folio. 2827846be085SMike Kravetz */ 2828846be085SMike Kravetz void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma, 2829d2d7bb44SSidhartha Kumar unsigned long address, struct folio *folio) 2830846be085SMike Kravetz { 283196b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 283296b96a96SMike Kravetz 28330ffdc38eSSidhartha Kumar if (folio_test_hugetlb_restore_reserve(folio)) { 2834846be085SMike Kravetz if (unlikely(rc < 0)) 283596b96a96SMike Kravetz /* 283696b96a96SMike Kravetz * Rare out of memory condition in reserve map 28370ffdc38eSSidhartha Kumar * manipulation. Clear hugetlb_restore_reserve so 28380ffdc38eSSidhartha Kumar * that global reserve count will not be incremented 2839454a00c4SMatthew Wilcox (Oracle) * by free_huge_folio. This will make it appear 28400ffdc38eSSidhartha Kumar * as though the reservation for this folio was 284196b96a96SMike Kravetz * consumed. This may prevent the task from 28420ffdc38eSSidhartha Kumar * faulting in the folio at a later time. This 284396b96a96SMike Kravetz * is better than inconsistent global huge page 284496b96a96SMike Kravetz * accounting of reserve counts. 284596b96a96SMike Kravetz */ 28460ffdc38eSSidhartha Kumar folio_clear_hugetlb_restore_reserve(folio); 2847846be085SMike Kravetz else if (rc) 2848846be085SMike Kravetz (void)vma_add_reservation(h, vma, address); 2849846be085SMike Kravetz else 2850846be085SMike Kravetz vma_end_reservation(h, vma, address); 2851846be085SMike Kravetz } else { 2852846be085SMike Kravetz if (!rc) { 285396b96a96SMike Kravetz /* 2854846be085SMike Kravetz * This indicates there is an entry in the reserve map 2855d0ce0e47SSidhartha Kumar * not added by alloc_hugetlb_folio. We know it was added 2856d0ce0e47SSidhartha Kumar * before the alloc_hugetlb_folio call, otherwise 28570ffdc38eSSidhartha Kumar * hugetlb_restore_reserve would be set on the folio. 2858846be085SMike Kravetz * Remove the entry so that a subsequent allocation 2859846be085SMike Kravetz * does not consume a reservation. 286096b96a96SMike Kravetz */ 2861846be085SMike Kravetz rc = vma_del_reservation(h, vma, address); 2862846be085SMike Kravetz if (rc < 0) 2863846be085SMike Kravetz /* 2864846be085SMike Kravetz * VERY rare out of memory condition. Since 2865846be085SMike Kravetz * we can not delete the entry, set 28660ffdc38eSSidhartha Kumar * hugetlb_restore_reserve so that the reserve 28670ffdc38eSSidhartha Kumar * count will be incremented when the folio 2868846be085SMike Kravetz * is freed. This reserve will be consumed 2869846be085SMike Kravetz * on a subsequent allocation. 2870846be085SMike Kravetz */ 28710ffdc38eSSidhartha Kumar folio_set_hugetlb_restore_reserve(folio); 2872846be085SMike Kravetz } else if (rc < 0) { 2873846be085SMike Kravetz /* 2874846be085SMike Kravetz * Rare out of memory condition from 2875846be085SMike Kravetz * vma_needs_reservation call. Memory allocation is 2876846be085SMike Kravetz * only attempted if a new entry is needed. Therefore, 2877846be085SMike Kravetz * this implies there is not an entry in the 2878846be085SMike Kravetz * reserve map. 2879846be085SMike Kravetz * 2880846be085SMike Kravetz * For shared mappings, no entry in the map indicates 2881846be085SMike Kravetz * no reservation. We are done. 2882846be085SMike Kravetz */ 2883846be085SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 2884846be085SMike Kravetz /* 2885846be085SMike Kravetz * For private mappings, no entry indicates 2886846be085SMike Kravetz * a reservation is present. Since we can 28870ffdc38eSSidhartha Kumar * not add an entry, set hugetlb_restore_reserve 28880ffdc38eSSidhartha Kumar * on the folio so reserve count will be 2889846be085SMike Kravetz * incremented when freed. This reserve will 2890846be085SMike Kravetz * be consumed on a subsequent allocation. 2891846be085SMike Kravetz */ 28920ffdc38eSSidhartha Kumar folio_set_hugetlb_restore_reserve(folio); 289396b96a96SMike Kravetz } else 2894846be085SMike Kravetz /* 2895846be085SMike Kravetz * No reservation present, do nothing 2896846be085SMike Kravetz */ 289796b96a96SMike Kravetz vma_end_reservation(h, vma, address); 289896b96a96SMike Kravetz } 289996b96a96SMike Kravetz } 290096b96a96SMike Kravetz 2901369fa227SOscar Salvador /* 290219fc1a7eSSidhartha Kumar * alloc_and_dissolve_hugetlb_folio - Allocate a new folio and dissolve 290319fc1a7eSSidhartha Kumar * the old one 2904369fa227SOscar Salvador * @h: struct hstate old page belongs to 290519fc1a7eSSidhartha Kumar * @old_folio: Old folio to dissolve 2906ae37c7ffSOscar Salvador * @list: List to isolate the page in case we need to 2907369fa227SOscar Salvador * Returns 0 on success, otherwise negated error. 2908369fa227SOscar Salvador */ 290919fc1a7eSSidhartha Kumar static int alloc_and_dissolve_hugetlb_folio(struct hstate *h, 291019fc1a7eSSidhartha Kumar struct folio *old_folio, struct list_head *list) 2911369fa227SOscar Salvador { 2912369fa227SOscar Salvador gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2913de656ed3SSidhartha Kumar int nid = folio_nid(old_folio); 2914de656ed3SSidhartha Kumar struct folio *new_folio; 2915369fa227SOscar Salvador int ret = 0; 2916369fa227SOscar Salvador 2917369fa227SOscar Salvador /* 291819fc1a7eSSidhartha Kumar * Before dissolving the folio, we need to allocate a new one for the 291919fc1a7eSSidhartha Kumar * pool to remain stable. Here, we allocate the folio and 'prep' it 2920f41f2ed4SMuchun Song * by doing everything but actually updating counters and adding to 2921f41f2ed4SMuchun Song * the pool. This simplifies and let us do most of the processing 2922f41f2ed4SMuchun Song * under the lock. 2923369fa227SOscar Salvador */ 292419fc1a7eSSidhartha Kumar new_folio = alloc_buddy_hugetlb_folio(h, gfp_mask, nid, NULL, NULL); 292519fc1a7eSSidhartha Kumar if (!new_folio) 2926369fa227SOscar Salvador return -ENOMEM; 2927de656ed3SSidhartha Kumar __prep_new_hugetlb_folio(h, new_folio); 2928369fa227SOscar Salvador 2929369fa227SOscar Salvador retry: 2930369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2931de656ed3SSidhartha Kumar if (!folio_test_hugetlb(old_folio)) { 2932369fa227SOscar Salvador /* 293319fc1a7eSSidhartha Kumar * Freed from under us. Drop new_folio too. 2934369fa227SOscar Salvador */ 2935369fa227SOscar Salvador goto free_new; 2936de656ed3SSidhartha Kumar } else if (folio_ref_count(old_folio)) { 29379747b9e9SBaolin Wang bool isolated; 29389747b9e9SBaolin Wang 2939369fa227SOscar Salvador /* 294019fc1a7eSSidhartha Kumar * Someone has grabbed the folio, try to isolate it here. 2941ae37c7ffSOscar Salvador * Fail with -EBUSY if not possible. 2942369fa227SOscar Salvador */ 2943ae37c7ffSOscar Salvador spin_unlock_irq(&hugetlb_lock); 29449747b9e9SBaolin Wang isolated = isolate_hugetlb(old_folio, list); 29459747b9e9SBaolin Wang ret = isolated ? 0 : -EBUSY; 2946ae37c7ffSOscar Salvador spin_lock_irq(&hugetlb_lock); 2947369fa227SOscar Salvador goto free_new; 2948de656ed3SSidhartha Kumar } else if (!folio_test_hugetlb_freed(old_folio)) { 2949369fa227SOscar Salvador /* 295019fc1a7eSSidhartha Kumar * Folio's refcount is 0 but it has not been enqueued in the 2951369fa227SOscar Salvador * freelist yet. Race window is small, so we can succeed here if 2952369fa227SOscar Salvador * we retry. 2953369fa227SOscar Salvador */ 2954369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2955369fa227SOscar Salvador cond_resched(); 2956369fa227SOscar Salvador goto retry; 2957369fa227SOscar Salvador } else { 2958369fa227SOscar Salvador /* 295919fc1a7eSSidhartha Kumar * Ok, old_folio is still a genuine free hugepage. Remove it from 2960369fa227SOscar Salvador * the freelist and decrease the counters. These will be 2961369fa227SOscar Salvador * incremented again when calling __prep_account_new_huge_page() 2962240d67a8SSidhartha Kumar * and enqueue_hugetlb_folio() for new_folio. The counters will 2963240d67a8SSidhartha Kumar * remain stable since this happens under the lock. 2964369fa227SOscar Salvador */ 2965cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, old_folio, false); 2966369fa227SOscar Salvador 2967369fa227SOscar Salvador /* 296819fc1a7eSSidhartha Kumar * Ref count on new_folio is already zero as it was dropped 2969b65a4edaSMike Kravetz * earlier. It can be directly added to the pool free list. 2970369fa227SOscar Salvador */ 2971369fa227SOscar Salvador __prep_account_new_huge_page(h, nid); 2972240d67a8SSidhartha Kumar enqueue_hugetlb_folio(h, new_folio); 2973369fa227SOscar Salvador 2974369fa227SOscar Salvador /* 297519fc1a7eSSidhartha Kumar * Folio has been replaced, we can safely free the old one. 2976369fa227SOscar Salvador */ 2977369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2978d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, old_folio, false); 2979369fa227SOscar Salvador } 2980369fa227SOscar Salvador 2981369fa227SOscar Salvador return ret; 2982369fa227SOscar Salvador 2983369fa227SOscar Salvador free_new: 2984369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 298519fc1a7eSSidhartha Kumar /* Folio has a zero ref count, but needs a ref to be freed */ 2986de656ed3SSidhartha Kumar folio_ref_unfreeze(new_folio, 1); 2987d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, new_folio, false); 2988369fa227SOscar Salvador 2989369fa227SOscar Salvador return ret; 2990369fa227SOscar Salvador } 2991369fa227SOscar Salvador 2992ae37c7ffSOscar Salvador int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list) 2993369fa227SOscar Salvador { 2994369fa227SOscar Salvador struct hstate *h; 2995d5e33bd8SSidhartha Kumar struct folio *folio = page_folio(page); 2996ae37c7ffSOscar Salvador int ret = -EBUSY; 2997369fa227SOscar Salvador 2998369fa227SOscar Salvador /* 2999369fa227SOscar Salvador * The page might have been dissolved from under our feet, so make sure 3000369fa227SOscar Salvador * to carefully check the state under the lock. 3001369fa227SOscar Salvador * Return success when racing as if we dissolved the page ourselves. 3002369fa227SOscar Salvador */ 3003369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 3004d5e33bd8SSidhartha Kumar if (folio_test_hugetlb(folio)) { 3005d5e33bd8SSidhartha Kumar h = folio_hstate(folio); 3006369fa227SOscar Salvador } else { 3007369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 3008369fa227SOscar Salvador return 0; 3009369fa227SOscar Salvador } 3010369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 3011369fa227SOscar Salvador 3012369fa227SOscar Salvador /* 3013369fa227SOscar Salvador * Fence off gigantic pages as there is a cyclic dependency between 3014369fa227SOscar Salvador * alloc_contig_range and them. Return -ENOMEM as this has the effect 3015369fa227SOscar Salvador * of bailing out right away without further retrying. 3016369fa227SOscar Salvador */ 3017369fa227SOscar Salvador if (hstate_is_gigantic(h)) 3018369fa227SOscar Salvador return -ENOMEM; 3019369fa227SOscar Salvador 30209747b9e9SBaolin Wang if (folio_ref_count(folio) && isolate_hugetlb(folio, list)) 3021ae37c7ffSOscar Salvador ret = 0; 3022d5e33bd8SSidhartha Kumar else if (!folio_ref_count(folio)) 302319fc1a7eSSidhartha Kumar ret = alloc_and_dissolve_hugetlb_folio(h, folio, list); 3024ae37c7ffSOscar Salvador 3025ae37c7ffSOscar Salvador return ret; 3026369fa227SOscar Salvador } 3027369fa227SOscar Salvador 3028d0ce0e47SSidhartha Kumar struct folio *alloc_hugetlb_folio(struct vm_area_struct *vma, 302904f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 3030348ea204SAdam Litke { 303190481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 3032a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3033d4ab0316SSidhartha Kumar struct folio *folio; 3034d85f69b0SMike Kravetz long map_chg, map_commit; 3035d85f69b0SMike Kravetz long gbl_chg; 30366d76dcf4SAneesh Kumar K.V int ret, idx; 3037d0ce0e47SSidhartha Kumar struct hugetlb_cgroup *h_cg = NULL; 303808cf9fafSMina Almasry bool deferred_reserve; 30392fc39cecSAdam Litke 30406d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 3041a1e78772SMel Gorman /* 3042d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 3043d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 3044d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 3045a1e78772SMel Gorman */ 3046d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 3047d85f69b0SMike Kravetz if (map_chg < 0) 304876dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 3049d85f69b0SMike Kravetz 3050d85f69b0SMike Kravetz /* 3051d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 3052d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 3053d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 3054d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 3055d85f69b0SMike Kravetz * checked against any subpool limit. 3056d85f69b0SMike Kravetz */ 3057d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 3058d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 3059d85f69b0SMike Kravetz if (gbl_chg < 0) { 3060feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 306176dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 30625e911373SMike Kravetz } 306390d8b7e6SAdam Litke 3064d85f69b0SMike Kravetz /* 3065d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 3066d85f69b0SMike Kravetz * map, there could be reservations associated with the 3067d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 3068d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 3069d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 3070d85f69b0SMike Kravetz * the subpool reservations. 3071d85f69b0SMike Kravetz */ 3072d85f69b0SMike Kravetz if (avoid_reserve) 3073d85f69b0SMike Kravetz gbl_chg = 1; 3074d85f69b0SMike Kravetz } 3075d85f69b0SMike Kravetz 307608cf9fafSMina Almasry /* If this allocation is not consuming a reservation, charge it now. 307708cf9fafSMina Almasry */ 30786501fe5fSMiaohe Lin deferred_reserve = map_chg || avoid_reserve; 307908cf9fafSMina Almasry if (deferred_reserve) { 308008cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 308108cf9fafSMina Almasry idx, pages_per_huge_page(h), &h_cg); 30828f34af6fSJianyu Zhan if (ret) 30838f34af6fSJianyu Zhan goto out_subpool_put; 308408cf9fafSMina Almasry } 308508cf9fafSMina Almasry 308608cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 308708cf9fafSMina Almasry if (ret) 308808cf9fafSMina Almasry goto out_uncharge_cgroup_reservation; 30898f34af6fSJianyu Zhan 3090db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3091d85f69b0SMike Kravetz /* 3092d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 3093d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 3094d85f69b0SMike Kravetz * a reservation exists for the allocation. 3095d85f69b0SMike Kravetz */ 3096ff7d853bSSidhartha Kumar folio = dequeue_hugetlb_folio_vma(h, vma, addr, avoid_reserve, gbl_chg); 3097ff7d853bSSidhartha Kumar if (!folio) { 3098db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3099ff7d853bSSidhartha Kumar folio = alloc_buddy_hugetlb_folio_with_mpol(h, vma, addr); 3100ff7d853bSSidhartha Kumar if (!folio) 31018f34af6fSJianyu Zhan goto out_uncharge_cgroup; 310212df140fSRik van Riel spin_lock_irq(&hugetlb_lock); 3103a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 3104ff7d853bSSidhartha Kumar folio_set_hugetlb_restore_reserve(folio); 3105a88c7695SNaoya Horiguchi h->resv_huge_pages--; 3106a88c7695SNaoya Horiguchi } 3107ff7d853bSSidhartha Kumar list_add(&folio->lru, &h->hugepage_activelist); 3108ff7d853bSSidhartha Kumar folio_ref_unfreeze(folio, 1); 310981a6fcaeSJoonsoo Kim /* Fall through */ 3110a1e78772SMel Gorman } 3111ff7d853bSSidhartha Kumar 3112ff7d853bSSidhartha Kumar hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, folio); 311308cf9fafSMina Almasry /* If allocation is not consuming a reservation, also store the 311408cf9fafSMina Almasry * hugetlb_cgroup pointer on the page. 311508cf9fafSMina Almasry */ 311608cf9fafSMina Almasry if (deferred_reserve) { 311708cf9fafSMina Almasry hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), 3118ff7d853bSSidhartha Kumar h_cg, folio); 311908cf9fafSMina Almasry } 312008cf9fafSMina Almasry 3121db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3122a1e78772SMel Gorman 3123ff7d853bSSidhartha Kumar hugetlb_set_folio_subpool(folio, spool); 3124a1e78772SMel Gorman 3125d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 3126d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 312733039678SMike Kravetz /* 312833039678SMike Kravetz * The page was added to the reservation map between 312933039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 313033039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 313133039678SMike Kravetz * Adjust for the subpool count incremented above AND 313233039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 313333039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 313433039678SMike Kravetz * no longer applies. 313533039678SMike Kravetz */ 313633039678SMike Kravetz long rsv_adjust; 313733039678SMike Kravetz 313833039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 313933039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 3140f6c5d21dSPeter Xu if (deferred_reserve) { 3141f6c5d21dSPeter Xu spin_lock_irq(&hugetlb_lock); 3142d4ab0316SSidhartha Kumar hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h), 3143d4ab0316SSidhartha Kumar pages_per_huge_page(h), folio); 3144f6c5d21dSPeter Xu spin_unlock_irq(&hugetlb_lock); 3145f6c5d21dSPeter Xu } 314633039678SMike Kravetz } 3147d0ce0e47SSidhartha Kumar return folio; 31488f34af6fSJianyu Zhan 31498f34af6fSJianyu Zhan out_uncharge_cgroup: 31508f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 315108cf9fafSMina Almasry out_uncharge_cgroup_reservation: 315208cf9fafSMina Almasry if (deferred_reserve) 315308cf9fafSMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), 315408cf9fafSMina Almasry h_cg); 31558f34af6fSJianyu Zhan out_subpool_put: 3156d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 31578f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 3158feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 31598f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 3160b45b5bd6SDavid Gibson } 3161b45b5bd6SDavid Gibson 3162b5389086SZhenguo Yao int alloc_bootmem_huge_page(struct hstate *h, int nid) 3163e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 3164b5389086SZhenguo Yao int __alloc_bootmem_huge_page(struct hstate *h, int nid) 3165aa888a74SAndi Kleen { 3166b5389086SZhenguo Yao struct huge_bootmem_page *m = NULL; /* initialize for clang */ 3167b2261026SJoonsoo Kim int nr_nodes, node; 3168aa888a74SAndi Kleen 3169b5389086SZhenguo Yao /* do node specific alloc */ 3170b5389086SZhenguo Yao if (nid != NUMA_NO_NODE) { 3171b5389086SZhenguo Yao m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h), 3172b5389086SZhenguo Yao 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid); 3173b5389086SZhenguo Yao if (!m) 3174b5389086SZhenguo Yao return 0; 3175b5389086SZhenguo Yao goto found; 3176b5389086SZhenguo Yao } 3177b5389086SZhenguo Yao /* allocate from next node when distributing huge pages */ 3178b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 3179b5389086SZhenguo Yao m = memblock_alloc_try_nid_raw( 31808b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 318197ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 3182aa888a74SAndi Kleen /* 3183aa888a74SAndi Kleen * Use the beginning of the huge page to store the 3184aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 3185aa888a74SAndi Kleen * puts them into the mem_map). 3186aa888a74SAndi Kleen */ 3187b5389086SZhenguo Yao if (!m) 3188b5389086SZhenguo Yao return 0; 3189aa888a74SAndi Kleen goto found; 3190aa888a74SAndi Kleen } 3191aa888a74SAndi Kleen 3192aa888a74SAndi Kleen found: 3193aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 3194330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 3195aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 3196aa888a74SAndi Kleen m->hstate = h; 3197aa888a74SAndi Kleen return 1; 3198aa888a74SAndi Kleen } 3199aa888a74SAndi Kleen 320048b8d744SMike Kravetz /* 320148b8d744SMike Kravetz * Put bootmem huge pages into the standard lists after mem_map is up. 320248b8d744SMike Kravetz * Note: This only applies to gigantic (order > MAX_ORDER) pages. 320348b8d744SMike Kravetz */ 3204aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 3205aa888a74SAndi Kleen { 3206aa888a74SAndi Kleen struct huge_bootmem_page *m; 3207aa888a74SAndi Kleen 3208aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 320940d18ebfSMike Kravetz struct page *page = virt_to_page(m); 32107f325a8dSSidhartha Kumar struct folio *folio = page_folio(page); 3211aa888a74SAndi Kleen struct hstate *h = m->hstate; 3212ee8f248dSBecky Bruce 321348b8d744SMike Kravetz VM_BUG_ON(!hstate_is_gigantic(h)); 3214d1c60955SSidhartha Kumar WARN_ON(folio_ref_count(folio) != 1); 3215d1c60955SSidhartha Kumar if (prep_compound_gigantic_folio(folio, huge_page_order(h))) { 3216d1c60955SSidhartha Kumar WARN_ON(folio_test_reserved(folio)); 3217d1c60955SSidhartha Kumar prep_new_hugetlb_folio(h, folio, folio_nid(folio)); 3218454a00c4SMatthew Wilcox (Oracle) free_huge_folio(folio); /* add to the hugepage allocator */ 32197118fc29SMike Kravetz } else { 3220416d85edSMike Kravetz /* VERY unlikely inflated ref count on a tail page */ 32217f325a8dSSidhartha Kumar free_gigantic_folio(folio, huge_page_order(h)); 32227118fc29SMike Kravetz } 3223af0fb9dfSMichal Hocko 3224b0320c7bSRafael Aquini /* 322548b8d744SMike Kravetz * We need to restore the 'stolen' pages to totalram_pages 322648b8d744SMike Kravetz * in order to fix confusing memory reports from free(1) and 322748b8d744SMike Kravetz * other side-effects, like CommitLimit going negative. 3228b0320c7bSRafael Aquini */ 3229c78a7f36SMiaohe Lin adjust_managed_page_count(page, pages_per_huge_page(h)); 3230520495feSCannon Matthews cond_resched(); 3231aa888a74SAndi Kleen } 3232aa888a74SAndi Kleen } 3233b5389086SZhenguo Yao static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid) 3234b5389086SZhenguo Yao { 3235b5389086SZhenguo Yao unsigned long i; 3236b5389086SZhenguo Yao char buf[32]; 3237b5389086SZhenguo Yao 3238b5389086SZhenguo Yao for (i = 0; i < h->max_huge_pages_node[nid]; ++i) { 3239b5389086SZhenguo Yao if (hstate_is_gigantic(h)) { 3240b5389086SZhenguo Yao if (!alloc_bootmem_huge_page(h, nid)) 3241b5389086SZhenguo Yao break; 3242b5389086SZhenguo Yao } else { 324319fc1a7eSSidhartha Kumar struct folio *folio; 3244b5389086SZhenguo Yao gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 3245b5389086SZhenguo Yao 324619fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, 3247b5389086SZhenguo Yao &node_states[N_MEMORY], NULL); 324819fc1a7eSSidhartha Kumar if (!folio) 3249b5389086SZhenguo Yao break; 3250454a00c4SMatthew Wilcox (Oracle) free_huge_folio(folio); /* free it into the hugepage allocator */ 3251b5389086SZhenguo Yao } 3252b5389086SZhenguo Yao cond_resched(); 3253b5389086SZhenguo Yao } 3254b5389086SZhenguo Yao if (i == h->max_huge_pages_node[nid]) 3255b5389086SZhenguo Yao return; 3256b5389086SZhenguo Yao 3257b5389086SZhenguo Yao string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3258b5389086SZhenguo Yao pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n", 3259b5389086SZhenguo Yao h->max_huge_pages_node[nid], buf, nid, i); 3260b5389086SZhenguo Yao h->max_huge_pages -= (h->max_huge_pages_node[nid] - i); 3261b5389086SZhenguo Yao h->max_huge_pages_node[nid] = i; 3262b5389086SZhenguo Yao } 3263aa888a74SAndi Kleen 32648faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 32651da177e4SLinus Torvalds { 32661da177e4SLinus Torvalds unsigned long i; 3267f60858f9SMike Kravetz nodemask_t *node_alloc_noretry; 3268b5389086SZhenguo Yao bool node_specific_alloc = false; 3269f60858f9SMike Kravetz 3270b5389086SZhenguo Yao /* skip gigantic hugepages allocation if hugetlb_cma enabled */ 3271b5389086SZhenguo Yao if (hstate_is_gigantic(h) && hugetlb_cma_size) { 3272b5389086SZhenguo Yao pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); 3273b5389086SZhenguo Yao return; 3274b5389086SZhenguo Yao } 3275b5389086SZhenguo Yao 3276b5389086SZhenguo Yao /* do node specific alloc */ 32770a7a0f6fSPeng Liu for_each_online_node(i) { 3278b5389086SZhenguo Yao if (h->max_huge_pages_node[i] > 0) { 3279b5389086SZhenguo Yao hugetlb_hstate_alloc_pages_onenode(h, i); 3280b5389086SZhenguo Yao node_specific_alloc = true; 3281b5389086SZhenguo Yao } 3282b5389086SZhenguo Yao } 3283b5389086SZhenguo Yao 3284b5389086SZhenguo Yao if (node_specific_alloc) 3285b5389086SZhenguo Yao return; 3286b5389086SZhenguo Yao 3287b5389086SZhenguo Yao /* below will do all node balanced alloc */ 3288f60858f9SMike Kravetz if (!hstate_is_gigantic(h)) { 3289f60858f9SMike Kravetz /* 3290f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3291f60858f9SMike Kravetz * Ignore errors as lower level routines can deal with 3292f60858f9SMike Kravetz * node_alloc_noretry == NULL. If this kmalloc fails at boot 3293f60858f9SMike Kravetz * time, we are likely in bigger trouble. 3294f60858f9SMike Kravetz */ 3295f60858f9SMike Kravetz node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), 3296f60858f9SMike Kravetz GFP_KERNEL); 3297f60858f9SMike Kravetz } else { 3298f60858f9SMike Kravetz /* allocations done at boot time */ 3299f60858f9SMike Kravetz node_alloc_noretry = NULL; 3300f60858f9SMike Kravetz } 3301f60858f9SMike Kravetz 3302f60858f9SMike Kravetz /* bit mask controlling how hard we retry per-node allocations */ 3303f60858f9SMike Kravetz if (node_alloc_noretry) 3304f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 33051da177e4SLinus Torvalds 3306e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 3307bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 3308b5389086SZhenguo Yao if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE)) 3309aa888a74SAndi Kleen break; 33100c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 3311f60858f9SMike Kravetz &node_states[N_MEMORY], 3312f60858f9SMike Kravetz node_alloc_noretry)) 33131da177e4SLinus Torvalds break; 331469ed779aSDavid Rientjes cond_resched(); 33151da177e4SLinus Torvalds } 3316d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 3317d715cf80SLiam R. Howlett char buf[32]; 3318d715cf80SLiam R. Howlett 3319c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3320d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 3321d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 33228faa8b07SAndi Kleen h->max_huge_pages = i; 3323e5ff2159SAndi Kleen } 3324f60858f9SMike Kravetz kfree(node_alloc_noretry); 3325d715cf80SLiam R. Howlett } 3326e5ff2159SAndi Kleen 3327e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 3328e5ff2159SAndi Kleen { 332979dfc695SMike Kravetz struct hstate *h, *h2; 3330e5ff2159SAndi Kleen 3331e5ff2159SAndi Kleen for_each_hstate(h) { 33328faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 3333bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 33348faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 333579dfc695SMike Kravetz 333679dfc695SMike Kravetz /* 333779dfc695SMike Kravetz * Set demote order for each hstate. Note that 333879dfc695SMike Kravetz * h->demote_order is initially 0. 333979dfc695SMike Kravetz * - We can not demote gigantic pages if runtime freeing 334079dfc695SMike Kravetz * is not supported, so skip this. 3341a01f4390SMike Kravetz * - If CMA allocation is possible, we can not demote 3342a01f4390SMike Kravetz * HUGETLB_PAGE_ORDER or smaller size pages. 334379dfc695SMike Kravetz */ 334479dfc695SMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 334579dfc695SMike Kravetz continue; 3346a01f4390SMike Kravetz if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER) 3347a01f4390SMike Kravetz continue; 334879dfc695SMike Kravetz for_each_hstate(h2) { 334979dfc695SMike Kravetz if (h2 == h) 335079dfc695SMike Kravetz continue; 335179dfc695SMike Kravetz if (h2->order < h->order && 335279dfc695SMike Kravetz h2->order > h->demote_order) 335379dfc695SMike Kravetz h->demote_order = h2->order; 335479dfc695SMike Kravetz } 3355e5ff2159SAndi Kleen } 3356e5ff2159SAndi Kleen } 3357e5ff2159SAndi Kleen 3358e5ff2159SAndi Kleen static void __init report_hugepages(void) 3359e5ff2159SAndi Kleen { 3360e5ff2159SAndi Kleen struct hstate *h; 3361e5ff2159SAndi Kleen 3362e5ff2159SAndi Kleen for_each_hstate(h) { 33634abd32dbSAndi Kleen char buf[32]; 3364c6247f72SMatthew Wilcox 3365c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 33666213834cSMuchun Song pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n", 3367c6247f72SMatthew Wilcox buf, h->free_huge_pages); 33686213834cSMuchun Song pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n", 33696213834cSMuchun Song hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf); 3370e5ff2159SAndi Kleen } 3371e5ff2159SAndi Kleen } 3372e5ff2159SAndi Kleen 33731da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 33746ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 33756ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 33761da177e4SLinus Torvalds { 33774415cc8dSChristoph Lameter int i; 33781121828aSMike Kravetz LIST_HEAD(page_list); 33794415cc8dSChristoph Lameter 33809487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 3381bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3382aa888a74SAndi Kleen return; 3383aa888a74SAndi Kleen 33841121828aSMike Kravetz /* 33851121828aSMike Kravetz * Collect pages to be freed on a list, and free after dropping lock 33861121828aSMike Kravetz */ 33876ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 338810c6ec49SMike Kravetz struct page *page, *next; 3389a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 3390a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 3391a5516438SAndi Kleen if (count >= h->nr_huge_pages) 33921121828aSMike Kravetz goto out; 33931da177e4SLinus Torvalds if (PageHighMem(page)) 33941da177e4SLinus Torvalds continue; 3395cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, page_folio(page), false); 33961121828aSMike Kravetz list_add(&page->lru, &page_list); 33971121828aSMike Kravetz } 33981121828aSMike Kravetz } 33991121828aSMike Kravetz 34001121828aSMike Kravetz out: 3401db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 340210c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3403db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 34041da177e4SLinus Torvalds } 34051da177e4SLinus Torvalds #else 34066ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 34076ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 34081da177e4SLinus Torvalds { 34091da177e4SLinus Torvalds } 34101da177e4SLinus Torvalds #endif 34111da177e4SLinus Torvalds 341220a0307cSWu Fengguang /* 341320a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 341420a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 341520a0307cSWu Fengguang * Returns 1 if an adjustment was made. 341620a0307cSWu Fengguang */ 34176ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 34186ae11b27SLee Schermerhorn int delta) 341920a0307cSWu Fengguang { 3420b2261026SJoonsoo Kim int nr_nodes, node; 342120a0307cSWu Fengguang 34229487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 342320a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 342420a0307cSWu Fengguang 3425e8c5c824SLee Schermerhorn if (delta < 0) { 3426b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 3427b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 3428b2261026SJoonsoo Kim goto found; 3429b2261026SJoonsoo Kim } 3430b2261026SJoonsoo Kim } else { 3431b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 3432b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 3433b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 3434b2261026SJoonsoo Kim goto found; 3435e8c5c824SLee Schermerhorn } 34369a76db09SLee Schermerhorn } 3437b2261026SJoonsoo Kim return 0; 343820a0307cSWu Fengguang 3439b2261026SJoonsoo Kim found: 344020a0307cSWu Fengguang h->surplus_huge_pages += delta; 3441b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 3442b2261026SJoonsoo Kim return 1; 344320a0307cSWu Fengguang } 344420a0307cSWu Fengguang 3445a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 3446fd875dcaSMike Kravetz static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 34476ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 34481da177e4SLinus Torvalds { 34497893d1d5SAdam Litke unsigned long min_count, ret; 345010c6ec49SMike Kravetz struct page *page; 345110c6ec49SMike Kravetz LIST_HEAD(page_list); 3452f60858f9SMike Kravetz NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); 3453f60858f9SMike Kravetz 3454f60858f9SMike Kravetz /* 3455f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3456f60858f9SMike Kravetz * If we can not allocate the bit mask, do not attempt to allocate 3457f60858f9SMike Kravetz * the requested huge pages. 3458f60858f9SMike Kravetz */ 3459f60858f9SMike Kravetz if (node_alloc_noretry) 3460f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 3461f60858f9SMike Kravetz else 3462f60858f9SMike Kravetz return -ENOMEM; 34631da177e4SLinus Torvalds 346429383967SMike Kravetz /* 346529383967SMike Kravetz * resize_lock mutex prevents concurrent adjustments to number of 346629383967SMike Kravetz * pages in hstate via the proc/sysfs interfaces. 346729383967SMike Kravetz */ 346829383967SMike Kravetz mutex_lock(&h->resize_lock); 3469b65d4adbSMuchun Song flush_free_hpage_work(h); 3470db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 34714eb0716eSAlexandre Ghiti 34724eb0716eSAlexandre Ghiti /* 3473fd875dcaSMike Kravetz * Check for a node specific request. 3474fd875dcaSMike Kravetz * Changing node specific huge page count may require a corresponding 3475fd875dcaSMike Kravetz * change to the global count. In any case, the passed node mask 3476fd875dcaSMike Kravetz * (nodes_allowed) will restrict alloc/free to the specified node. 3477fd875dcaSMike Kravetz */ 3478fd875dcaSMike Kravetz if (nid != NUMA_NO_NODE) { 3479fd875dcaSMike Kravetz unsigned long old_count = count; 3480fd875dcaSMike Kravetz 3481fd875dcaSMike Kravetz count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 3482fd875dcaSMike Kravetz /* 3483fd875dcaSMike Kravetz * User may have specified a large count value which caused the 3484fd875dcaSMike Kravetz * above calculation to overflow. In this case, they wanted 3485fd875dcaSMike Kravetz * to allocate as many huge pages as possible. Set count to 3486fd875dcaSMike Kravetz * largest possible value to align with their intention. 3487fd875dcaSMike Kravetz */ 3488fd875dcaSMike Kravetz if (count < old_count) 3489fd875dcaSMike Kravetz count = ULONG_MAX; 3490fd875dcaSMike Kravetz } 3491fd875dcaSMike Kravetz 3492fd875dcaSMike Kravetz /* 34934eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 34944eb0716eSAlexandre Ghiti * page range allocation. 34954eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 34964eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 34974eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 34984eb0716eSAlexandre Ghiti */ 34994eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 35004eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 3501db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 350229383967SMike Kravetz mutex_unlock(&h->resize_lock); 3503f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 35044eb0716eSAlexandre Ghiti return -EINVAL; 35054eb0716eSAlexandre Ghiti } 35064eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 35074eb0716eSAlexandre Ghiti } 3508aa888a74SAndi Kleen 35097893d1d5SAdam Litke /* 35107893d1d5SAdam Litke * Increase the pool size 35117893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 35127893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 3513d1c3fb1fSNishanth Aravamudan * 35143a740e8bSSidhartha Kumar * We might race with alloc_surplus_hugetlb_folio() here and be unable 3515d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 3516d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 3517d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 3518d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 35197893d1d5SAdam Litke */ 3520a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 35216ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 35227893d1d5SAdam Litke break; 35237893d1d5SAdam Litke } 35247893d1d5SAdam Litke 3525a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 35267893d1d5SAdam Litke /* 35277893d1d5SAdam Litke * If this allocation races such that we no longer need the 3528454a00c4SMatthew Wilcox (Oracle) * page, free_huge_folio will handle it by freeing the page 35297893d1d5SAdam Litke * and reducing the surplus. 35307893d1d5SAdam Litke */ 3531db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3532649920c6SJia He 3533649920c6SJia He /* yield cpu to avoid soft lockup */ 3534649920c6SJia He cond_resched(); 3535649920c6SJia He 3536f60858f9SMike Kravetz ret = alloc_pool_huge_page(h, nodes_allowed, 3537f60858f9SMike Kravetz node_alloc_noretry); 3538db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 35397893d1d5SAdam Litke if (!ret) 35407893d1d5SAdam Litke goto out; 35417893d1d5SAdam Litke 3542536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 3543536240f2SMel Gorman if (signal_pending(current)) 3544536240f2SMel Gorman goto out; 35457893d1d5SAdam Litke } 35467893d1d5SAdam Litke 35477893d1d5SAdam Litke /* 35487893d1d5SAdam Litke * Decrease the pool size 35497893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 35507893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 35517893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 35527893d1d5SAdam Litke * to the desired size as pages become free. 3553d1c3fb1fSNishanth Aravamudan * 3554d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 3555d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 3556d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 35573a740e8bSSidhartha Kumar * alloc_surplus_hugetlb_folio() is checking the global counter, 3558d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 3559d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 3560d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 35617893d1d5SAdam Litke */ 3562a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 35636b0c880dSAdam Litke min_count = max(count, min_count); 35646ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 356510c6ec49SMike Kravetz 356610c6ec49SMike Kravetz /* 356710c6ec49SMike Kravetz * Collect pages to be removed on list without dropping lock 356810c6ec49SMike Kravetz */ 3569a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 357010c6ec49SMike Kravetz page = remove_pool_huge_page(h, nodes_allowed, 0); 357110c6ec49SMike Kravetz if (!page) 35721da177e4SLinus Torvalds break; 357310c6ec49SMike Kravetz 357410c6ec49SMike Kravetz list_add(&page->lru, &page_list); 35751da177e4SLinus Torvalds } 357610c6ec49SMike Kravetz /* free the pages after dropping lock */ 3577db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 357810c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3579b65d4adbSMuchun Song flush_free_hpage_work(h); 3580db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 358110c6ec49SMike Kravetz 3582a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 35836ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 35847893d1d5SAdam Litke break; 35857893d1d5SAdam Litke } 35867893d1d5SAdam Litke out: 35874eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 3588db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 358929383967SMike Kravetz mutex_unlock(&h->resize_lock); 35904eb0716eSAlexandre Ghiti 3591f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 3592f60858f9SMike Kravetz 35934eb0716eSAlexandre Ghiti return 0; 35941da177e4SLinus Torvalds } 35951da177e4SLinus Torvalds 3596bdd7be07SSidhartha Kumar static int demote_free_hugetlb_folio(struct hstate *h, struct folio *folio) 35978531fc6fSMike Kravetz { 3598bdd7be07SSidhartha Kumar int i, nid = folio_nid(folio); 35998531fc6fSMike Kravetz struct hstate *target_hstate; 360031731452SDoug Berger struct page *subpage; 3601bdd7be07SSidhartha Kumar struct folio *inner_folio; 36028531fc6fSMike Kravetz int rc = 0; 36038531fc6fSMike Kravetz 36048531fc6fSMike Kravetz target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order); 36058531fc6fSMike Kravetz 3606cfd5082bSSidhartha Kumar remove_hugetlb_folio_for_demote(h, folio, false); 36078531fc6fSMike Kravetz spin_unlock_irq(&hugetlb_lock); 36088531fc6fSMike Kravetz 3609bdd7be07SSidhartha Kumar rc = hugetlb_vmemmap_restore(h, &folio->page); 36108531fc6fSMike Kravetz if (rc) { 3611bdd7be07SSidhartha Kumar /* Allocation of vmemmmap failed, we can not demote folio */ 36128531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 3613bdd7be07SSidhartha Kumar folio_ref_unfreeze(folio, 1); 3614bdd7be07SSidhartha Kumar add_hugetlb_folio(h, folio, false); 36158531fc6fSMike Kravetz return rc; 36168531fc6fSMike Kravetz } 36178531fc6fSMike Kravetz 36188531fc6fSMike Kravetz /* 3619911565b8SSidhartha Kumar * Use destroy_compound_hugetlb_folio_for_demote for all huge page 3620bdd7be07SSidhartha Kumar * sizes as it will not ref count folios. 36218531fc6fSMike Kravetz */ 3622911565b8SSidhartha Kumar destroy_compound_hugetlb_folio_for_demote(folio, huge_page_order(h)); 36238531fc6fSMike Kravetz 36248531fc6fSMike Kravetz /* 36258531fc6fSMike Kravetz * Taking target hstate mutex synchronizes with set_max_huge_pages. 36268531fc6fSMike Kravetz * Without the mutex, pages added to target hstate could be marked 36278531fc6fSMike Kravetz * as surplus. 36288531fc6fSMike Kravetz * 36298531fc6fSMike Kravetz * Note that we already hold h->resize_lock. To prevent deadlock, 36308531fc6fSMike Kravetz * use the convention of always taking larger size hstate mutex first. 36318531fc6fSMike Kravetz */ 36328531fc6fSMike Kravetz mutex_lock(&target_hstate->resize_lock); 36338531fc6fSMike Kravetz for (i = 0; i < pages_per_huge_page(h); 36348531fc6fSMike Kravetz i += pages_per_huge_page(target_hstate)) { 3635bdd7be07SSidhartha Kumar subpage = folio_page(folio, i); 3636bdd7be07SSidhartha Kumar inner_folio = page_folio(subpage); 36378531fc6fSMike Kravetz if (hstate_is_gigantic(target_hstate)) 3638bdd7be07SSidhartha Kumar prep_compound_gigantic_folio_for_demote(inner_folio, 36398531fc6fSMike Kravetz target_hstate->order); 36408531fc6fSMike Kravetz else 364131731452SDoug Berger prep_compound_page(subpage, target_hstate->order); 3642bdd7be07SSidhartha Kumar folio_change_private(inner_folio, NULL); 3643bdd7be07SSidhartha Kumar prep_new_hugetlb_folio(target_hstate, inner_folio, nid); 3644454a00c4SMatthew Wilcox (Oracle) free_huge_folio(inner_folio); 36458531fc6fSMike Kravetz } 36468531fc6fSMike Kravetz mutex_unlock(&target_hstate->resize_lock); 36478531fc6fSMike Kravetz 36488531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 36498531fc6fSMike Kravetz 36508531fc6fSMike Kravetz /* 36518531fc6fSMike Kravetz * Not absolutely necessary, but for consistency update max_huge_pages 36528531fc6fSMike Kravetz * based on pool changes for the demoted page. 36538531fc6fSMike Kravetz */ 36548531fc6fSMike Kravetz h->max_huge_pages--; 3655a43a83c7SMiaohe Lin target_hstate->max_huge_pages += 3656a43a83c7SMiaohe Lin pages_per_huge_page(h) / pages_per_huge_page(target_hstate); 36578531fc6fSMike Kravetz 36588531fc6fSMike Kravetz return rc; 36598531fc6fSMike Kravetz } 36608531fc6fSMike Kravetz 366179dfc695SMike Kravetz static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 366279dfc695SMike Kravetz __must_hold(&hugetlb_lock) 366379dfc695SMike Kravetz { 36648531fc6fSMike Kravetz int nr_nodes, node; 3665bdd7be07SSidhartha Kumar struct folio *folio; 366679dfc695SMike Kravetz 366779dfc695SMike Kravetz lockdep_assert_held(&hugetlb_lock); 366879dfc695SMike Kravetz 366979dfc695SMike Kravetz /* We should never get here if no demote order */ 367079dfc695SMike Kravetz if (!h->demote_order) { 367179dfc695SMike Kravetz pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n"); 367279dfc695SMike Kravetz return -EINVAL; /* internal error */ 367379dfc695SMike Kravetz } 367479dfc695SMike Kravetz 36758531fc6fSMike Kravetz for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 3676bdd7be07SSidhartha Kumar list_for_each_entry(folio, &h->hugepage_freelists[node], lru) { 3677bdd7be07SSidhartha Kumar if (folio_test_hwpoison(folio)) 36785a317412SMike Kravetz continue; 3679bdd7be07SSidhartha Kumar return demote_free_hugetlb_folio(h, folio); 36808531fc6fSMike Kravetz } 36818531fc6fSMike Kravetz } 36828531fc6fSMike Kravetz 36835a317412SMike Kravetz /* 36845a317412SMike Kravetz * Only way to get here is if all pages on free lists are poisoned. 36855a317412SMike Kravetz * Return -EBUSY so that caller will not retry. 36865a317412SMike Kravetz */ 36875a317412SMike Kravetz return -EBUSY; 368879dfc695SMike Kravetz } 368979dfc695SMike Kravetz 3690a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 3691a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 3692a3437870SNishanth Aravamudan 369379dfc695SMike Kravetz #define HSTATE_ATTR_WO(_name) \ 369479dfc695SMike Kravetz static struct kobj_attribute _name##_attr = __ATTR_WO(_name) 369579dfc695SMike Kravetz 3696a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 369798bc26acSMiaohe Lin static struct kobj_attribute _name##_attr = __ATTR_RW(_name) 3698a3437870SNishanth Aravamudan 3699a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 3700a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 3701a3437870SNishanth Aravamudan 37029a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 37039a305230SLee Schermerhorn 37049a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 3705a3437870SNishanth Aravamudan { 3706a3437870SNishanth Aravamudan int i; 37079a305230SLee Schermerhorn 3708a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 37099a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 37109a305230SLee Schermerhorn if (nidp) 37119a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 3712a3437870SNishanth Aravamudan return &hstates[i]; 37139a305230SLee Schermerhorn } 37149a305230SLee Schermerhorn 37159a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 3716a3437870SNishanth Aravamudan } 3717a3437870SNishanth Aravamudan 371806808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 3719a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3720a3437870SNishanth Aravamudan { 37219a305230SLee Schermerhorn struct hstate *h; 37229a305230SLee Schermerhorn unsigned long nr_huge_pages; 37239a305230SLee Schermerhorn int nid; 37249a305230SLee Schermerhorn 37259a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 37269a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 37279a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 37289a305230SLee Schermerhorn else 37299a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 37309a305230SLee Schermerhorn 3731ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", nr_huge_pages); 3732a3437870SNishanth Aravamudan } 3733adbe8726SEric B Munson 3734238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 3735238d3c13SDavid Rientjes struct hstate *h, int nid, 3736238d3c13SDavid Rientjes unsigned long count, size_t len) 3737a3437870SNishanth Aravamudan { 3738a3437870SNishanth Aravamudan int err; 37392d0adf7eSOscar Salvador nodemask_t nodes_allowed, *n_mask; 3740a3437870SNishanth Aravamudan 37412d0adf7eSOscar Salvador if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 37422d0adf7eSOscar Salvador return -EINVAL; 3743adbe8726SEric B Munson 37449a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 37459a305230SLee Schermerhorn /* 37469a305230SLee Schermerhorn * global hstate attribute 37479a305230SLee Schermerhorn */ 37489a305230SLee Schermerhorn if (!(obey_mempolicy && 37492d0adf7eSOscar Salvador init_nodemask_of_mempolicy(&nodes_allowed))) 37502d0adf7eSOscar Salvador n_mask = &node_states[N_MEMORY]; 37512d0adf7eSOscar Salvador else 37522d0adf7eSOscar Salvador n_mask = &nodes_allowed; 37532d0adf7eSOscar Salvador } else { 37549a305230SLee Schermerhorn /* 3755fd875dcaSMike Kravetz * Node specific request. count adjustment happens in 3756fd875dcaSMike Kravetz * set_max_huge_pages() after acquiring hugetlb_lock. 37579a305230SLee Schermerhorn */ 37582d0adf7eSOscar Salvador init_nodemask_of_node(&nodes_allowed, nid); 37592d0adf7eSOscar Salvador n_mask = &nodes_allowed; 3760fd875dcaSMike Kravetz } 37619a305230SLee Schermerhorn 37622d0adf7eSOscar Salvador err = set_max_huge_pages(h, count, nid, n_mask); 376306808b08SLee Schermerhorn 37644eb0716eSAlexandre Ghiti return err ? err : len; 376506808b08SLee Schermerhorn } 376606808b08SLee Schermerhorn 3767238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 3768238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 3769238d3c13SDavid Rientjes size_t len) 3770238d3c13SDavid Rientjes { 3771238d3c13SDavid Rientjes struct hstate *h; 3772238d3c13SDavid Rientjes unsigned long count; 3773238d3c13SDavid Rientjes int nid; 3774238d3c13SDavid Rientjes int err; 3775238d3c13SDavid Rientjes 3776238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 3777238d3c13SDavid Rientjes if (err) 3778238d3c13SDavid Rientjes return err; 3779238d3c13SDavid Rientjes 3780238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 3781238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 3782238d3c13SDavid Rientjes } 3783238d3c13SDavid Rientjes 378406808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 378506808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 378606808b08SLee Schermerhorn { 378706808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 378806808b08SLee Schermerhorn } 378906808b08SLee Schermerhorn 379006808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 379106808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 379206808b08SLee Schermerhorn { 3793238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 3794a3437870SNishanth Aravamudan } 3795a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 3796a3437870SNishanth Aravamudan 379706808b08SLee Schermerhorn #ifdef CONFIG_NUMA 379806808b08SLee Schermerhorn 379906808b08SLee Schermerhorn /* 380006808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 380106808b08SLee Schermerhorn * huge page alloc/free. 380206808b08SLee Schermerhorn */ 380306808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 3804ae7a927dSJoe Perches struct kobj_attribute *attr, 3805ae7a927dSJoe Perches char *buf) 380606808b08SLee Schermerhorn { 380706808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 380806808b08SLee Schermerhorn } 380906808b08SLee Schermerhorn 381006808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 381106808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 381206808b08SLee Schermerhorn { 3813238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 381406808b08SLee Schermerhorn } 381506808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 381606808b08SLee Schermerhorn #endif 381706808b08SLee Schermerhorn 381806808b08SLee Schermerhorn 3819a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 3820a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3821a3437870SNishanth Aravamudan { 38229a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3823ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); 3824a3437870SNishanth Aravamudan } 3825adbe8726SEric B Munson 3826a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 3827a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 3828a3437870SNishanth Aravamudan { 3829a3437870SNishanth Aravamudan int err; 3830a3437870SNishanth Aravamudan unsigned long input; 38319a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3832a3437870SNishanth Aravamudan 3833bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3834adbe8726SEric B Munson return -EINVAL; 3835adbe8726SEric B Munson 38363dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 3837a3437870SNishanth Aravamudan if (err) 383873ae31e5SEric B Munson return err; 3839a3437870SNishanth Aravamudan 3840db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3841a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 3842db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3843a3437870SNishanth Aravamudan 3844a3437870SNishanth Aravamudan return count; 3845a3437870SNishanth Aravamudan } 3846a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 3847a3437870SNishanth Aravamudan 3848a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 3849a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3850a3437870SNishanth Aravamudan { 38519a305230SLee Schermerhorn struct hstate *h; 38529a305230SLee Schermerhorn unsigned long free_huge_pages; 38539a305230SLee Schermerhorn int nid; 38549a305230SLee Schermerhorn 38559a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 38569a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 38579a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 38589a305230SLee Schermerhorn else 38599a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 38609a305230SLee Schermerhorn 3861ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", free_huge_pages); 3862a3437870SNishanth Aravamudan } 3863a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 3864a3437870SNishanth Aravamudan 3865a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 3866a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3867a3437870SNishanth Aravamudan { 38689a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3869ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); 3870a3437870SNishanth Aravamudan } 3871a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 3872a3437870SNishanth Aravamudan 3873a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 3874a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3875a3437870SNishanth Aravamudan { 38769a305230SLee Schermerhorn struct hstate *h; 38779a305230SLee Schermerhorn unsigned long surplus_huge_pages; 38789a305230SLee Schermerhorn int nid; 38799a305230SLee Schermerhorn 38809a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 38819a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 38829a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 38839a305230SLee Schermerhorn else 38849a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 38859a305230SLee Schermerhorn 3886ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", surplus_huge_pages); 3887a3437870SNishanth Aravamudan } 3888a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 3889a3437870SNishanth Aravamudan 389079dfc695SMike Kravetz static ssize_t demote_store(struct kobject *kobj, 389179dfc695SMike Kravetz struct kobj_attribute *attr, const char *buf, size_t len) 389279dfc695SMike Kravetz { 389379dfc695SMike Kravetz unsigned long nr_demote; 389479dfc695SMike Kravetz unsigned long nr_available; 389579dfc695SMike Kravetz nodemask_t nodes_allowed, *n_mask; 389679dfc695SMike Kravetz struct hstate *h; 38978eeda55fSLi zeming int err; 389879dfc695SMike Kravetz int nid; 389979dfc695SMike Kravetz 390079dfc695SMike Kravetz err = kstrtoul(buf, 10, &nr_demote); 390179dfc695SMike Kravetz if (err) 390279dfc695SMike Kravetz return err; 390379dfc695SMike Kravetz h = kobj_to_hstate(kobj, &nid); 390479dfc695SMike Kravetz 390579dfc695SMike Kravetz if (nid != NUMA_NO_NODE) { 390679dfc695SMike Kravetz init_nodemask_of_node(&nodes_allowed, nid); 390779dfc695SMike Kravetz n_mask = &nodes_allowed; 390879dfc695SMike Kravetz } else { 390979dfc695SMike Kravetz n_mask = &node_states[N_MEMORY]; 391079dfc695SMike Kravetz } 391179dfc695SMike Kravetz 391279dfc695SMike Kravetz /* Synchronize with other sysfs operations modifying huge pages */ 391379dfc695SMike Kravetz mutex_lock(&h->resize_lock); 391479dfc695SMike Kravetz spin_lock_irq(&hugetlb_lock); 391579dfc695SMike Kravetz 391679dfc695SMike Kravetz while (nr_demote) { 391779dfc695SMike Kravetz /* 391879dfc695SMike Kravetz * Check for available pages to demote each time thorough the 391979dfc695SMike Kravetz * loop as demote_pool_huge_page will drop hugetlb_lock. 392079dfc695SMike Kravetz */ 392179dfc695SMike Kravetz if (nid != NUMA_NO_NODE) 392279dfc695SMike Kravetz nr_available = h->free_huge_pages_node[nid]; 392379dfc695SMike Kravetz else 392479dfc695SMike Kravetz nr_available = h->free_huge_pages; 392579dfc695SMike Kravetz nr_available -= h->resv_huge_pages; 392679dfc695SMike Kravetz if (!nr_available) 392779dfc695SMike Kravetz break; 392879dfc695SMike Kravetz 392979dfc695SMike Kravetz err = demote_pool_huge_page(h, n_mask); 393079dfc695SMike Kravetz if (err) 393179dfc695SMike Kravetz break; 393279dfc695SMike Kravetz 393379dfc695SMike Kravetz nr_demote--; 393479dfc695SMike Kravetz } 393579dfc695SMike Kravetz 393679dfc695SMike Kravetz spin_unlock_irq(&hugetlb_lock); 393779dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 393879dfc695SMike Kravetz 393979dfc695SMike Kravetz if (err) 394079dfc695SMike Kravetz return err; 394179dfc695SMike Kravetz return len; 394279dfc695SMike Kravetz } 394379dfc695SMike Kravetz HSTATE_ATTR_WO(demote); 394479dfc695SMike Kravetz 394579dfc695SMike Kravetz static ssize_t demote_size_show(struct kobject *kobj, 394679dfc695SMike Kravetz struct kobj_attribute *attr, char *buf) 394779dfc695SMike Kravetz { 394812658abfSMiaohe Lin struct hstate *h = kobj_to_hstate(kobj, NULL); 394979dfc695SMike Kravetz unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K; 395079dfc695SMike Kravetz 395179dfc695SMike Kravetz return sysfs_emit(buf, "%lukB\n", demote_size); 395279dfc695SMike Kravetz } 395379dfc695SMike Kravetz 395479dfc695SMike Kravetz static ssize_t demote_size_store(struct kobject *kobj, 395579dfc695SMike Kravetz struct kobj_attribute *attr, 395679dfc695SMike Kravetz const char *buf, size_t count) 395779dfc695SMike Kravetz { 395879dfc695SMike Kravetz struct hstate *h, *demote_hstate; 395979dfc695SMike Kravetz unsigned long demote_size; 396079dfc695SMike Kravetz unsigned int demote_order; 396179dfc695SMike Kravetz 396279dfc695SMike Kravetz demote_size = (unsigned long)memparse(buf, NULL); 396379dfc695SMike Kravetz 396479dfc695SMike Kravetz demote_hstate = size_to_hstate(demote_size); 396579dfc695SMike Kravetz if (!demote_hstate) 396679dfc695SMike Kravetz return -EINVAL; 396779dfc695SMike Kravetz demote_order = demote_hstate->order; 3968a01f4390SMike Kravetz if (demote_order < HUGETLB_PAGE_ORDER) 3969a01f4390SMike Kravetz return -EINVAL; 397079dfc695SMike Kravetz 397179dfc695SMike Kravetz /* demote order must be smaller than hstate order */ 397212658abfSMiaohe Lin h = kobj_to_hstate(kobj, NULL); 397379dfc695SMike Kravetz if (demote_order >= h->order) 397479dfc695SMike Kravetz return -EINVAL; 397579dfc695SMike Kravetz 397679dfc695SMike Kravetz /* resize_lock synchronizes access to demote size and writes */ 397779dfc695SMike Kravetz mutex_lock(&h->resize_lock); 397879dfc695SMike Kravetz h->demote_order = demote_order; 397979dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 398079dfc695SMike Kravetz 398179dfc695SMike Kravetz return count; 398279dfc695SMike Kravetz } 398379dfc695SMike Kravetz HSTATE_ATTR(demote_size); 398479dfc695SMike Kravetz 3985a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 3986a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 3987a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 3988a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 3989a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 3990a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 399106808b08SLee Schermerhorn #ifdef CONFIG_NUMA 399206808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 399306808b08SLee Schermerhorn #endif 3994a3437870SNishanth Aravamudan NULL, 3995a3437870SNishanth Aravamudan }; 3996a3437870SNishanth Aravamudan 399767e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 3998a3437870SNishanth Aravamudan .attrs = hstate_attrs, 3999a3437870SNishanth Aravamudan }; 4000a3437870SNishanth Aravamudan 400179dfc695SMike Kravetz static struct attribute *hstate_demote_attrs[] = { 400279dfc695SMike Kravetz &demote_size_attr.attr, 400379dfc695SMike Kravetz &demote_attr.attr, 400479dfc695SMike Kravetz NULL, 400579dfc695SMike Kravetz }; 400679dfc695SMike Kravetz 400779dfc695SMike Kravetz static const struct attribute_group hstate_demote_attr_group = { 400879dfc695SMike Kravetz .attrs = hstate_demote_attrs, 400979dfc695SMike Kravetz }; 401079dfc695SMike Kravetz 4011094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 40129a305230SLee Schermerhorn struct kobject **hstate_kobjs, 401367e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 4014a3437870SNishanth Aravamudan { 4015a3437870SNishanth Aravamudan int retval; 4016972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 4017a3437870SNishanth Aravamudan 40189a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 40199a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 4020a3437870SNishanth Aravamudan return -ENOMEM; 4021a3437870SNishanth Aravamudan 40229a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 4023cc2205a6SMiaohe Lin if (retval) { 40249a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 4025cc2205a6SMiaohe Lin hstate_kobjs[hi] = NULL; 40263a6bdda0SMiaohe Lin return retval; 4027cc2205a6SMiaohe Lin } 4028a3437870SNishanth Aravamudan 402979dfc695SMike Kravetz if (h->demote_order) { 403001088a60SMiaohe Lin retval = sysfs_create_group(hstate_kobjs[hi], 403101088a60SMiaohe Lin &hstate_demote_attr_group); 403201088a60SMiaohe Lin if (retval) { 403379dfc695SMike Kravetz pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name); 403401088a60SMiaohe Lin sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group); 403501088a60SMiaohe Lin kobject_put(hstate_kobjs[hi]); 403601088a60SMiaohe Lin hstate_kobjs[hi] = NULL; 403701088a60SMiaohe Lin return retval; 403801088a60SMiaohe Lin } 403979dfc695SMike Kravetz } 404079dfc695SMike Kravetz 404101088a60SMiaohe Lin return 0; 4042a3437870SNishanth Aravamudan } 4043a3437870SNishanth Aravamudan 40449a305230SLee Schermerhorn #ifdef CONFIG_NUMA 4045a4a00b45SMuchun Song static bool hugetlb_sysfs_initialized __ro_after_init; 40469a305230SLee Schermerhorn 40479a305230SLee Schermerhorn /* 40489a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 404910fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 405010fbcf4cSKay Sievers * index of a node device or _hstate == node id. 405110fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 40529a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 40539a305230SLee Schermerhorn */ 40549a305230SLee Schermerhorn struct node_hstate { 40559a305230SLee Schermerhorn struct kobject *hugepages_kobj; 40569a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 40579a305230SLee Schermerhorn }; 4058b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 40599a305230SLee Schermerhorn 40609a305230SLee Schermerhorn /* 406110fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 40629a305230SLee Schermerhorn */ 40639a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 40649a305230SLee Schermerhorn &nr_hugepages_attr.attr, 40659a305230SLee Schermerhorn &free_hugepages_attr.attr, 40669a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 40679a305230SLee Schermerhorn NULL, 40689a305230SLee Schermerhorn }; 40699a305230SLee Schermerhorn 407067e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 40719a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 40729a305230SLee Schermerhorn }; 40739a305230SLee Schermerhorn 40749a305230SLee Schermerhorn /* 407510fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 40769a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 40779a305230SLee Schermerhorn */ 40789a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 40799a305230SLee Schermerhorn { 40809a305230SLee Schermerhorn int nid; 40819a305230SLee Schermerhorn 40829a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 40839a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 40849a305230SLee Schermerhorn int i; 40859a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 40869a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 40879a305230SLee Schermerhorn if (nidp) 40889a305230SLee Schermerhorn *nidp = nid; 40899a305230SLee Schermerhorn return &hstates[i]; 40909a305230SLee Schermerhorn } 40919a305230SLee Schermerhorn } 40929a305230SLee Schermerhorn 40939a305230SLee Schermerhorn BUG(); 40949a305230SLee Schermerhorn return NULL; 40959a305230SLee Schermerhorn } 40969a305230SLee Schermerhorn 40979a305230SLee Schermerhorn /* 409810fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 40999a305230SLee Schermerhorn * No-op if no hstate attributes attached. 41009a305230SLee Schermerhorn */ 4101a4a00b45SMuchun Song void hugetlb_unregister_node(struct node *node) 41029a305230SLee Schermerhorn { 41039a305230SLee Schermerhorn struct hstate *h; 410410fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 41059a305230SLee Schermerhorn 41069a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 41079b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 41089a305230SLee Schermerhorn 4109972dc4deSAneesh Kumar K.V for_each_hstate(h) { 4110972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 411101088a60SMiaohe Lin struct kobject *hstate_kobj = nhs->hstate_kobjs[idx]; 411201088a60SMiaohe Lin 411301088a60SMiaohe Lin if (!hstate_kobj) 411401088a60SMiaohe Lin continue; 411501088a60SMiaohe Lin if (h->demote_order) 411601088a60SMiaohe Lin sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group); 411701088a60SMiaohe Lin sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group); 411801088a60SMiaohe Lin kobject_put(hstate_kobj); 4119972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 4120972dc4deSAneesh Kumar K.V } 41219a305230SLee Schermerhorn 41229a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 41239a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 41249a305230SLee Schermerhorn } 41259a305230SLee Schermerhorn 41269a305230SLee Schermerhorn 41279a305230SLee Schermerhorn /* 412810fbcf4cSKay Sievers * Register hstate attributes for a single node device. 41299a305230SLee Schermerhorn * No-op if attributes already registered. 41309a305230SLee Schermerhorn */ 4131a4a00b45SMuchun Song void hugetlb_register_node(struct node *node) 41329a305230SLee Schermerhorn { 41339a305230SLee Schermerhorn struct hstate *h; 413410fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 41359a305230SLee Schermerhorn int err; 41369a305230SLee Schermerhorn 4137a4a00b45SMuchun Song if (!hugetlb_sysfs_initialized) 4138a4a00b45SMuchun Song return; 4139a4a00b45SMuchun Song 41409a305230SLee Schermerhorn if (nhs->hugepages_kobj) 41419a305230SLee Schermerhorn return; /* already allocated */ 41429a305230SLee Schermerhorn 41439a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 414410fbcf4cSKay Sievers &node->dev.kobj); 41459a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 41469a305230SLee Schermerhorn return; 41479a305230SLee Schermerhorn 41489a305230SLee Schermerhorn for_each_hstate(h) { 41499a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 41509a305230SLee Schermerhorn nhs->hstate_kobjs, 41519a305230SLee Schermerhorn &per_node_hstate_attr_group); 41529a305230SLee Schermerhorn if (err) { 4153282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s for node %d\n", 415410fbcf4cSKay Sievers h->name, node->dev.id); 41559a305230SLee Schermerhorn hugetlb_unregister_node(node); 41569a305230SLee Schermerhorn break; 41579a305230SLee Schermerhorn } 41589a305230SLee Schermerhorn } 41599a305230SLee Schermerhorn } 41609a305230SLee Schermerhorn 41619a305230SLee Schermerhorn /* 41629b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 416310fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 416410fbcf4cSKay Sievers * registered their associated device by this time. 41659a305230SLee Schermerhorn */ 41667d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 41679a305230SLee Schermerhorn { 41689a305230SLee Schermerhorn int nid; 41699a305230SLee Schermerhorn 4170a4a00b45SMuchun Song for_each_online_node(nid) 4171b958d4d0SMuchun Song hugetlb_register_node(node_devices[nid]); 41729a305230SLee Schermerhorn } 41739a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 41749a305230SLee Schermerhorn 41759a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 41769a305230SLee Schermerhorn { 41779a305230SLee Schermerhorn BUG(); 41789a305230SLee Schermerhorn if (nidp) 41799a305230SLee Schermerhorn *nidp = -1; 41809a305230SLee Schermerhorn return NULL; 41819a305230SLee Schermerhorn } 41829a305230SLee Schermerhorn 41839a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 41849a305230SLee Schermerhorn 41859a305230SLee Schermerhorn #endif 41869a305230SLee Schermerhorn 4187263b8998SMiaohe Lin #ifdef CONFIG_CMA 4188263b8998SMiaohe Lin static void __init hugetlb_cma_check(void); 4189263b8998SMiaohe Lin #else 4190263b8998SMiaohe Lin static inline __init void hugetlb_cma_check(void) 4191263b8998SMiaohe Lin { 4192263b8998SMiaohe Lin } 4193263b8998SMiaohe Lin #endif 4194263b8998SMiaohe Lin 4195a4a00b45SMuchun Song static void __init hugetlb_sysfs_init(void) 4196a4a00b45SMuchun Song { 4197a4a00b45SMuchun Song struct hstate *h; 4198a4a00b45SMuchun Song int err; 4199a4a00b45SMuchun Song 4200a4a00b45SMuchun Song hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 4201a4a00b45SMuchun Song if (!hugepages_kobj) 4202a4a00b45SMuchun Song return; 4203a4a00b45SMuchun Song 4204a4a00b45SMuchun Song for_each_hstate(h) { 4205a4a00b45SMuchun Song err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 4206a4a00b45SMuchun Song hstate_kobjs, &hstate_attr_group); 4207a4a00b45SMuchun Song if (err) 4208a4a00b45SMuchun Song pr_err("HugeTLB: Unable to add hstate %s", h->name); 4209a4a00b45SMuchun Song } 4210a4a00b45SMuchun Song 4211a4a00b45SMuchun Song #ifdef CONFIG_NUMA 4212a4a00b45SMuchun Song hugetlb_sysfs_initialized = true; 4213a4a00b45SMuchun Song #endif 4214a4a00b45SMuchun Song hugetlb_register_all_nodes(); 4215a4a00b45SMuchun Song } 4216a4a00b45SMuchun Song 4217962de548SKefeng Wang #ifdef CONFIG_SYSCTL 4218962de548SKefeng Wang static void hugetlb_sysctl_init(void); 4219962de548SKefeng Wang #else 4220962de548SKefeng Wang static inline void hugetlb_sysctl_init(void) { } 4221962de548SKefeng Wang #endif 4222962de548SKefeng Wang 4223a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 4224a3437870SNishanth Aravamudan { 42258382d914SDavidlohr Bueso int i; 42268382d914SDavidlohr Bueso 4227d6995da3SMike Kravetz BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < 4228d6995da3SMike Kravetz __NR_HPAGEFLAGS); 4229d6995da3SMike Kravetz 4230c2833a5bSMike Kravetz if (!hugepages_supported()) { 4231c2833a5bSMike Kravetz if (hugetlb_max_hstate || default_hstate_max_huge_pages) 4232c2833a5bSMike Kravetz pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); 42330ef89d25SBenjamin Herrenschmidt return 0; 4234d715cf80SLiam R. Howlett } 4235d715cf80SLiam R. Howlett 4236282f4214SMike Kravetz /* 4237282f4214SMike Kravetz * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some 4238282f4214SMike Kravetz * architectures depend on setup being done here. 4239282f4214SMike Kravetz */ 4240a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 4241282f4214SMike Kravetz if (!parsed_default_hugepagesz) { 4242282f4214SMike Kravetz /* 4243282f4214SMike Kravetz * If we did not parse a default huge page size, set 4244282f4214SMike Kravetz * default_hstate_idx to HPAGE_SIZE hstate. And, if the 4245282f4214SMike Kravetz * number of huge pages for this default size was implicitly 4246282f4214SMike Kravetz * specified, set that here as well. 4247282f4214SMike Kravetz * Note that the implicit setting will overwrite an explicit 4248282f4214SMike Kravetz * setting. A warning will be printed in this case. 4249282f4214SMike Kravetz */ 4250282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); 4251f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 4252282f4214SMike Kravetz if (default_hstate.max_huge_pages) { 4253282f4214SMike Kravetz char buf[32]; 4254282f4214SMike Kravetz 4255282f4214SMike Kravetz string_get_size(huge_page_size(&default_hstate), 4256282f4214SMike Kravetz 1, STRING_UNITS_2, buf, 32); 4257282f4214SMike Kravetz pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", 4258282f4214SMike Kravetz default_hstate.max_huge_pages, buf); 4259282f4214SMike Kravetz pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", 4260282f4214SMike Kravetz default_hstate_max_huge_pages); 4261282f4214SMike Kravetz } 4262282f4214SMike Kravetz default_hstate.max_huge_pages = 4263282f4214SMike Kravetz default_hstate_max_huge_pages; 4264b5389086SZhenguo Yao 42650a7a0f6fSPeng Liu for_each_online_node(i) 4266b5389086SZhenguo Yao default_hstate.max_huge_pages_node[i] = 4267b5389086SZhenguo Yao default_hugepages_in_node[i]; 4268282f4214SMike Kravetz } 4269f8b74815SVaishali Thakkar } 4270a3437870SNishanth Aravamudan 4271cf11e85fSRoman Gushchin hugetlb_cma_check(); 4272a3437870SNishanth Aravamudan hugetlb_init_hstates(); 4273aa888a74SAndi Kleen gather_bootmem_prealloc(); 4274a3437870SNishanth Aravamudan report_hugepages(); 4275a3437870SNishanth Aravamudan 4276a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 42777179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 4278962de548SKefeng Wang hugetlb_sysctl_init(); 42799a305230SLee Schermerhorn 42808382d914SDavidlohr Bueso #ifdef CONFIG_SMP 42818382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 42828382d914SDavidlohr Bueso #else 42838382d914SDavidlohr Bueso num_fault_mutexes = 1; 42848382d914SDavidlohr Bueso #endif 4285c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 42866da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 42876da2ec56SKees Cook GFP_KERNEL); 4288c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 42898382d914SDavidlohr Bueso 42908382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 4291c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 4292a3437870SNishanth Aravamudan return 0; 4293a3437870SNishanth Aravamudan } 42943e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 4295a3437870SNishanth Aravamudan 4296ae94da89SMike Kravetz /* Overwritten by architectures with more huge page sizes */ 4297ae94da89SMike Kravetz bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) 42989fee021dSVaishali Thakkar { 4299ae94da89SMike Kravetz return size == HPAGE_SIZE; 43009fee021dSVaishali Thakkar } 43019fee021dSVaishali Thakkar 4302d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 4303a3437870SNishanth Aravamudan { 4304a3437870SNishanth Aravamudan struct hstate *h; 43058faa8b07SAndi Kleen unsigned long i; 43068faa8b07SAndi Kleen 4307a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 4308a3437870SNishanth Aravamudan return; 4309a3437870SNishanth Aravamudan } 431047d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 4311a3437870SNishanth Aravamudan BUG_ON(order == 0); 431247d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 431329383967SMike Kravetz mutex_init(&h->resize_lock); 4314a3437870SNishanth Aravamudan h->order = order; 4315aca78307SMiaohe Lin h->mask = ~(huge_page_size(h) - 1); 43168faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 43178faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 43180edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 431954f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 432054f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 4321a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 4322c2c3a60aSMiaohe Lin huge_page_size(h)/SZ_1K); 43238faa8b07SAndi Kleen 4324a3437870SNishanth Aravamudan parsed_hstate = h; 4325a3437870SNishanth Aravamudan } 4326a3437870SNishanth Aravamudan 4327b5389086SZhenguo Yao bool __init __weak hugetlb_node_alloc_supported(void) 4328b5389086SZhenguo Yao { 4329b5389086SZhenguo Yao return true; 4330b5389086SZhenguo Yao } 4331f87442f4SPeng Liu 4332f87442f4SPeng Liu static void __init hugepages_clear_pages_in_node(void) 4333f87442f4SPeng Liu { 4334f87442f4SPeng Liu if (!hugetlb_max_hstate) { 4335f87442f4SPeng Liu default_hstate_max_huge_pages = 0; 4336f87442f4SPeng Liu memset(default_hugepages_in_node, 0, 433710395680SMiaohe Lin sizeof(default_hugepages_in_node)); 4338f87442f4SPeng Liu } else { 4339f87442f4SPeng Liu parsed_hstate->max_huge_pages = 0; 4340f87442f4SPeng Liu memset(parsed_hstate->max_huge_pages_node, 0, 434110395680SMiaohe Lin sizeof(parsed_hstate->max_huge_pages_node)); 4342f87442f4SPeng Liu } 4343f87442f4SPeng Liu } 4344f87442f4SPeng Liu 4345282f4214SMike Kravetz /* 4346282f4214SMike Kravetz * hugepages command line processing 4347282f4214SMike Kravetz * hugepages normally follows a valid hugepagsz or default_hugepagsz 4348282f4214SMike Kravetz * specification. If not, ignore the hugepages value. hugepages can also 4349282f4214SMike Kravetz * be the first huge page command line option in which case it implicitly 4350282f4214SMike Kravetz * specifies the number of huge pages for the default size. 4351282f4214SMike Kravetz */ 4352282f4214SMike Kravetz static int __init hugepages_setup(char *s) 4353a3437870SNishanth Aravamudan { 4354a3437870SNishanth Aravamudan unsigned long *mhp; 43558faa8b07SAndi Kleen static unsigned long *last_mhp; 4356b5389086SZhenguo Yao int node = NUMA_NO_NODE; 4357b5389086SZhenguo Yao int count; 4358b5389086SZhenguo Yao unsigned long tmp; 4359b5389086SZhenguo Yao char *p = s; 4360a3437870SNishanth Aravamudan 43619fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 4362282f4214SMike Kravetz pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); 43639fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 4364f81f6e4bSPeng Liu return 1; 43659fee021dSVaishali Thakkar } 4366282f4214SMike Kravetz 4367a3437870SNishanth Aravamudan /* 4368282f4214SMike Kravetz * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter 4369282f4214SMike Kravetz * yet, so this hugepages= parameter goes to the "default hstate". 4370282f4214SMike Kravetz * Otherwise, it goes with the previously parsed hugepagesz or 4371282f4214SMike Kravetz * default_hugepagesz. 4372a3437870SNishanth Aravamudan */ 43739fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 4374a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 4375a3437870SNishanth Aravamudan else 4376a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 4377a3437870SNishanth Aravamudan 43788faa8b07SAndi Kleen if (mhp == last_mhp) { 4379282f4214SMike Kravetz pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); 4380f81f6e4bSPeng Liu return 1; 43818faa8b07SAndi Kleen } 43828faa8b07SAndi Kleen 4383b5389086SZhenguo Yao while (*p) { 4384b5389086SZhenguo Yao count = 0; 4385b5389086SZhenguo Yao if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4386b5389086SZhenguo Yao goto invalid; 4387b5389086SZhenguo Yao /* Parameter is node format */ 4388b5389086SZhenguo Yao if (p[count] == ':') { 4389b5389086SZhenguo Yao if (!hugetlb_node_alloc_supported()) { 4390b5389086SZhenguo Yao pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n"); 4391f81f6e4bSPeng Liu return 1; 4392b5389086SZhenguo Yao } 43930a7a0f6fSPeng Liu if (tmp >= MAX_NUMNODES || !node_online(tmp)) 4394e79ce983SLiu Yuntao goto invalid; 43950a7a0f6fSPeng Liu node = array_index_nospec(tmp, MAX_NUMNODES); 4396b5389086SZhenguo Yao p += count + 1; 4397b5389086SZhenguo Yao /* Parse hugepages */ 4398b5389086SZhenguo Yao if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4399b5389086SZhenguo Yao goto invalid; 4400b5389086SZhenguo Yao if (!hugetlb_max_hstate) 4401b5389086SZhenguo Yao default_hugepages_in_node[node] = tmp; 4402b5389086SZhenguo Yao else 4403b5389086SZhenguo Yao parsed_hstate->max_huge_pages_node[node] = tmp; 4404b5389086SZhenguo Yao *mhp += tmp; 4405b5389086SZhenguo Yao /* Go to parse next node*/ 4406b5389086SZhenguo Yao if (p[count] == ',') 4407b5389086SZhenguo Yao p += count + 1; 4408b5389086SZhenguo Yao else 4409b5389086SZhenguo Yao break; 4410b5389086SZhenguo Yao } else { 4411b5389086SZhenguo Yao if (p != s) 4412b5389086SZhenguo Yao goto invalid; 4413b5389086SZhenguo Yao *mhp = tmp; 4414b5389086SZhenguo Yao break; 4415b5389086SZhenguo Yao } 4416b5389086SZhenguo Yao } 4417a3437870SNishanth Aravamudan 44188faa8b07SAndi Kleen /* 44198faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 442004adbc3fSMiaohe Lin * But we need to allocate gigantic hstates here early to still 44218faa8b07SAndi Kleen * use the bootmem allocator. 44228faa8b07SAndi Kleen */ 442304adbc3fSMiaohe Lin if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) 44248faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 44258faa8b07SAndi Kleen 44268faa8b07SAndi Kleen last_mhp = mhp; 44278faa8b07SAndi Kleen 4428a3437870SNishanth Aravamudan return 1; 4429b5389086SZhenguo Yao 4430b5389086SZhenguo Yao invalid: 4431b5389086SZhenguo Yao pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p); 4432f87442f4SPeng Liu hugepages_clear_pages_in_node(); 4433f81f6e4bSPeng Liu return 1; 4434a3437870SNishanth Aravamudan } 4435282f4214SMike Kravetz __setup("hugepages=", hugepages_setup); 4436e11bfbfcSNick Piggin 4437282f4214SMike Kravetz /* 4438282f4214SMike Kravetz * hugepagesz command line processing 4439282f4214SMike Kravetz * A specific huge page size can only be specified once with hugepagesz. 4440282f4214SMike Kravetz * hugepagesz is followed by hugepages on the command line. The global 4441282f4214SMike Kravetz * variable 'parsed_valid_hugepagesz' is used to determine if prior 4442282f4214SMike Kravetz * hugepagesz argument was valid. 4443282f4214SMike Kravetz */ 4444359f2544SMike Kravetz static int __init hugepagesz_setup(char *s) 4445e11bfbfcSNick Piggin { 4446359f2544SMike Kravetz unsigned long size; 4447282f4214SMike Kravetz struct hstate *h; 4448282f4214SMike Kravetz 4449282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4450359f2544SMike Kravetz size = (unsigned long)memparse(s, NULL); 4451359f2544SMike Kravetz 4452359f2544SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4453282f4214SMike Kravetz pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); 4454f81f6e4bSPeng Liu return 1; 4455359f2544SMike Kravetz } 4456359f2544SMike Kravetz 4457282f4214SMike Kravetz h = size_to_hstate(size); 4458282f4214SMike Kravetz if (h) { 4459282f4214SMike Kravetz /* 4460282f4214SMike Kravetz * hstate for this size already exists. This is normally 4461282f4214SMike Kravetz * an error, but is allowed if the existing hstate is the 4462282f4214SMike Kravetz * default hstate. More specifically, it is only allowed if 4463282f4214SMike Kravetz * the number of huge pages for the default hstate was not 4464282f4214SMike Kravetz * previously specified. 4465282f4214SMike Kravetz */ 4466282f4214SMike Kravetz if (!parsed_default_hugepagesz || h != &default_hstate || 4467282f4214SMike Kravetz default_hstate.max_huge_pages) { 4468282f4214SMike Kravetz pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); 4469f81f6e4bSPeng Liu return 1; 447038237830SMike Kravetz } 447138237830SMike Kravetz 4472282f4214SMike Kravetz /* 4473282f4214SMike Kravetz * No need to call hugetlb_add_hstate() as hstate already 4474282f4214SMike Kravetz * exists. But, do set parsed_hstate so that a following 4475282f4214SMike Kravetz * hugepages= parameter will be applied to this hstate. 4476282f4214SMike Kravetz */ 4477282f4214SMike Kravetz parsed_hstate = h; 4478282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4479e11bfbfcSNick Piggin return 1; 4480e11bfbfcSNick Piggin } 4481282f4214SMike Kravetz 4482359f2544SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4483282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4484359f2544SMike Kravetz return 1; 4485359f2544SMike Kravetz } 4486359f2544SMike Kravetz __setup("hugepagesz=", hugepagesz_setup); 4487359f2544SMike Kravetz 4488282f4214SMike Kravetz /* 4489282f4214SMike Kravetz * default_hugepagesz command line input 4490282f4214SMike Kravetz * Only one instance of default_hugepagesz allowed on command line. 4491282f4214SMike Kravetz */ 4492ae94da89SMike Kravetz static int __init default_hugepagesz_setup(char *s) 4493e11bfbfcSNick Piggin { 4494ae94da89SMike Kravetz unsigned long size; 4495b5389086SZhenguo Yao int i; 4496ae94da89SMike Kravetz 4497282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4498282f4214SMike Kravetz if (parsed_default_hugepagesz) { 4499282f4214SMike Kravetz pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); 4500f81f6e4bSPeng Liu return 1; 4501282f4214SMike Kravetz } 4502282f4214SMike Kravetz 4503282f4214SMike Kravetz size = (unsigned long)memparse(s, NULL); 4504282f4214SMike Kravetz 4505282f4214SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4506282f4214SMike Kravetz pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); 4507f81f6e4bSPeng Liu return 1; 4508282f4214SMike Kravetz } 4509282f4214SMike Kravetz 4510282f4214SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4511282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4512282f4214SMike Kravetz parsed_default_hugepagesz = true; 4513282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(size)); 4514282f4214SMike Kravetz 4515282f4214SMike Kravetz /* 4516282f4214SMike Kravetz * The number of default huge pages (for this size) could have been 4517282f4214SMike Kravetz * specified as the first hugetlb parameter: hugepages=X. If so, 4518282f4214SMike Kravetz * then default_hstate_max_huge_pages is set. If the default huge 451923baf831SKirill A. Shutemov * page size is gigantic (> MAX_ORDER), then the pages must be 4520282f4214SMike Kravetz * allocated here from bootmem allocator. 4521282f4214SMike Kravetz */ 4522282f4214SMike Kravetz if (default_hstate_max_huge_pages) { 4523282f4214SMike Kravetz default_hstate.max_huge_pages = default_hstate_max_huge_pages; 45240a7a0f6fSPeng Liu for_each_online_node(i) 4525b5389086SZhenguo Yao default_hstate.max_huge_pages_node[i] = 4526b5389086SZhenguo Yao default_hugepages_in_node[i]; 4527282f4214SMike Kravetz if (hstate_is_gigantic(&default_hstate)) 4528282f4214SMike Kravetz hugetlb_hstate_alloc_pages(&default_hstate); 4529282f4214SMike Kravetz default_hstate_max_huge_pages = 0; 4530282f4214SMike Kravetz } 4531282f4214SMike Kravetz 4532e11bfbfcSNick Piggin return 1; 4533e11bfbfcSNick Piggin } 4534ae94da89SMike Kravetz __setup("default_hugepagesz=", default_hugepagesz_setup); 4535a3437870SNishanth Aravamudan 4536d2226ebdSFeng Tang static nodemask_t *policy_mbind_nodemask(gfp_t gfp) 4537d2226ebdSFeng Tang { 4538d2226ebdSFeng Tang #ifdef CONFIG_NUMA 4539d2226ebdSFeng Tang struct mempolicy *mpol = get_task_policy(current); 4540d2226ebdSFeng Tang 4541d2226ebdSFeng Tang /* 4542d2226ebdSFeng Tang * Only enforce MPOL_BIND policy which overlaps with cpuset policy 4543d2226ebdSFeng Tang * (from policy_nodemask) specifically for hugetlb case 4544d2226ebdSFeng Tang */ 4545d2226ebdSFeng Tang if (mpol->mode == MPOL_BIND && 4546d2226ebdSFeng Tang (apply_policy_zone(mpol, gfp_zone(gfp)) && 4547d2226ebdSFeng Tang cpuset_nodemask_valid_mems_allowed(&mpol->nodes))) 4548d2226ebdSFeng Tang return &mpol->nodes; 4549d2226ebdSFeng Tang #endif 4550d2226ebdSFeng Tang return NULL; 4551d2226ebdSFeng Tang } 4552d2226ebdSFeng Tang 45538ca39e68SMuchun Song static unsigned int allowed_mems_nr(struct hstate *h) 45548a213460SNishanth Aravamudan { 45558a213460SNishanth Aravamudan int node; 45568a213460SNishanth Aravamudan unsigned int nr = 0; 4557d2226ebdSFeng Tang nodemask_t *mbind_nodemask; 45588ca39e68SMuchun Song unsigned int *array = h->free_huge_pages_node; 45598ca39e68SMuchun Song gfp_t gfp_mask = htlb_alloc_mask(h); 45608a213460SNishanth Aravamudan 4561d2226ebdSFeng Tang mbind_nodemask = policy_mbind_nodemask(gfp_mask); 45628ca39e68SMuchun Song for_each_node_mask(node, cpuset_current_mems_allowed) { 4563d2226ebdSFeng Tang if (!mbind_nodemask || node_isset(node, *mbind_nodemask)) 45648a213460SNishanth Aravamudan nr += array[node]; 45658ca39e68SMuchun Song } 45668a213460SNishanth Aravamudan 45678a213460SNishanth Aravamudan return nr; 45688a213460SNishanth Aravamudan } 45698a213460SNishanth Aravamudan 45708a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 457117743798SMuchun Song static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, 457217743798SMuchun Song void *buffer, size_t *length, 457317743798SMuchun Song loff_t *ppos, unsigned long *out) 457417743798SMuchun Song { 457517743798SMuchun Song struct ctl_table dup_table; 457617743798SMuchun Song 457717743798SMuchun Song /* 457817743798SMuchun Song * In order to avoid races with __do_proc_doulongvec_minmax(), we 457917743798SMuchun Song * can duplicate the @table and alter the duplicate of it. 458017743798SMuchun Song */ 458117743798SMuchun Song dup_table = *table; 458217743798SMuchun Song dup_table.data = out; 458317743798SMuchun Song 458417743798SMuchun Song return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); 458517743798SMuchun Song } 458617743798SMuchun Song 458706808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 458806808b08SLee Schermerhorn struct ctl_table *table, int write, 458932927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 45901da177e4SLinus Torvalds { 4591e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 4592238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 459308d4a246SMichal Hocko int ret; 4594e5ff2159SAndi Kleen 4595457c1b27SNishanth Aravamudan if (!hugepages_supported()) 459686613628SJan Stancek return -EOPNOTSUPP; 4597457c1b27SNishanth Aravamudan 459817743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 459917743798SMuchun Song &tmp); 460008d4a246SMichal Hocko if (ret) 460108d4a246SMichal Hocko goto out; 4602e5ff2159SAndi Kleen 4603238d3c13SDavid Rientjes if (write) 4604238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 4605238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 460608d4a246SMichal Hocko out: 460708d4a246SMichal Hocko return ret; 46081da177e4SLinus Torvalds } 4609396faf03SMel Gorman 4610962de548SKefeng Wang static int hugetlb_sysctl_handler(struct ctl_table *table, int write, 461132927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 461206808b08SLee Schermerhorn { 461306808b08SLee Schermerhorn 461406808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 461506808b08SLee Schermerhorn buffer, length, ppos); 461606808b08SLee Schermerhorn } 461706808b08SLee Schermerhorn 461806808b08SLee Schermerhorn #ifdef CONFIG_NUMA 4619962de548SKefeng Wang static int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 462032927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 462106808b08SLee Schermerhorn { 462206808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 462306808b08SLee Schermerhorn buffer, length, ppos); 462406808b08SLee Schermerhorn } 462506808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 462606808b08SLee Schermerhorn 4627962de548SKefeng Wang static int hugetlb_overcommit_handler(struct ctl_table *table, int write, 462832927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 4629a3d0c6aaSNishanth Aravamudan { 4630a5516438SAndi Kleen struct hstate *h = &default_hstate; 4631e5ff2159SAndi Kleen unsigned long tmp; 463208d4a246SMichal Hocko int ret; 4633e5ff2159SAndi Kleen 4634457c1b27SNishanth Aravamudan if (!hugepages_supported()) 463586613628SJan Stancek return -EOPNOTSUPP; 4636457c1b27SNishanth Aravamudan 4637e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 4638e5ff2159SAndi Kleen 4639bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 4640adbe8726SEric B Munson return -EINVAL; 4641adbe8726SEric B Munson 464217743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 464317743798SMuchun Song &tmp); 464408d4a246SMichal Hocko if (ret) 464508d4a246SMichal Hocko goto out; 4646e5ff2159SAndi Kleen 4647e5ff2159SAndi Kleen if (write) { 4648db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4649e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 4650db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4651e5ff2159SAndi Kleen } 465208d4a246SMichal Hocko out: 465308d4a246SMichal Hocko return ret; 4654a3d0c6aaSNishanth Aravamudan } 4655a3d0c6aaSNishanth Aravamudan 4656962de548SKefeng Wang static struct ctl_table hugetlb_table[] = { 4657962de548SKefeng Wang { 4658962de548SKefeng Wang .procname = "nr_hugepages", 4659962de548SKefeng Wang .data = NULL, 4660962de548SKefeng Wang .maxlen = sizeof(unsigned long), 4661962de548SKefeng Wang .mode = 0644, 4662962de548SKefeng Wang .proc_handler = hugetlb_sysctl_handler, 4663962de548SKefeng Wang }, 4664962de548SKefeng Wang #ifdef CONFIG_NUMA 4665962de548SKefeng Wang { 4666962de548SKefeng Wang .procname = "nr_hugepages_mempolicy", 4667962de548SKefeng Wang .data = NULL, 4668962de548SKefeng Wang .maxlen = sizeof(unsigned long), 4669962de548SKefeng Wang .mode = 0644, 4670962de548SKefeng Wang .proc_handler = &hugetlb_mempolicy_sysctl_handler, 4671962de548SKefeng Wang }, 4672962de548SKefeng Wang #endif 4673962de548SKefeng Wang { 4674962de548SKefeng Wang .procname = "hugetlb_shm_group", 4675962de548SKefeng Wang .data = &sysctl_hugetlb_shm_group, 4676962de548SKefeng Wang .maxlen = sizeof(gid_t), 4677962de548SKefeng Wang .mode = 0644, 4678962de548SKefeng Wang .proc_handler = proc_dointvec, 4679962de548SKefeng Wang }, 4680962de548SKefeng Wang { 4681962de548SKefeng Wang .procname = "nr_overcommit_hugepages", 4682962de548SKefeng Wang .data = NULL, 4683962de548SKefeng Wang .maxlen = sizeof(unsigned long), 4684962de548SKefeng Wang .mode = 0644, 4685962de548SKefeng Wang .proc_handler = hugetlb_overcommit_handler, 4686962de548SKefeng Wang }, 4687962de548SKefeng Wang { } 4688962de548SKefeng Wang }; 4689962de548SKefeng Wang 4690962de548SKefeng Wang static void hugetlb_sysctl_init(void) 4691962de548SKefeng Wang { 4692962de548SKefeng Wang register_sysctl_init("vm", hugetlb_table); 4693962de548SKefeng Wang } 46941da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 46951da177e4SLinus Torvalds 4696e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 46971da177e4SLinus Torvalds { 4698fcb2b0c5SRoman Gushchin struct hstate *h; 4699fcb2b0c5SRoman Gushchin unsigned long total = 0; 4700fcb2b0c5SRoman Gushchin 4701457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4702457c1b27SNishanth Aravamudan return; 4703fcb2b0c5SRoman Gushchin 4704fcb2b0c5SRoman Gushchin for_each_hstate(h) { 4705fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 4706fcb2b0c5SRoman Gushchin 4707aca78307SMiaohe Lin total += huge_page_size(h) * count; 4708fcb2b0c5SRoman Gushchin 4709fcb2b0c5SRoman Gushchin if (h == &default_hstate) 4710e1759c21SAlexey Dobriyan seq_printf(m, 47111da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 47121da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 4713b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 47147893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 47154f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 4716fcb2b0c5SRoman Gushchin count, 4717a5516438SAndi Kleen h->free_huge_pages, 4718a5516438SAndi Kleen h->resv_huge_pages, 4719a5516438SAndi Kleen h->surplus_huge_pages, 4720aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4721fcb2b0c5SRoman Gushchin } 4722fcb2b0c5SRoman Gushchin 4723aca78307SMiaohe Lin seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); 47241da177e4SLinus Torvalds } 47251da177e4SLinus Torvalds 47267981593bSJoe Perches int hugetlb_report_node_meminfo(char *buf, int len, int nid) 47271da177e4SLinus Torvalds { 4728a5516438SAndi Kleen struct hstate *h = &default_hstate; 47297981593bSJoe Perches 4730457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4731457c1b27SNishanth Aravamudan return 0; 47327981593bSJoe Perches 47337981593bSJoe Perches return sysfs_emit_at(buf, len, 47341da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 4735a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 4736a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 4737a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 4738a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 4739a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 47401da177e4SLinus Torvalds } 47411da177e4SLinus Torvalds 4742dcadcf1cSGang Li void hugetlb_show_meminfo_node(int nid) 4743949f7ec5SDavid Rientjes { 4744949f7ec5SDavid Rientjes struct hstate *h; 4745949f7ec5SDavid Rientjes 4746457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4747457c1b27SNishanth Aravamudan return; 4748457c1b27SNishanth Aravamudan 4749949f7ec5SDavid Rientjes for_each_hstate(h) 4750dcadcf1cSGang Li printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 4751949f7ec5SDavid Rientjes nid, 4752949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 4753949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 4754949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 4755aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4756949f7ec5SDavid Rientjes } 4757949f7ec5SDavid Rientjes 47585d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 47595d317b2bSNaoya Horiguchi { 47605d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 47616c1aa2d3SZhangPeng K(atomic_long_read(&mm->hugetlb_usage))); 47625d317b2bSNaoya Horiguchi } 47635d317b2bSNaoya Horiguchi 47641da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 47651da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 47661da177e4SLinus Torvalds { 4767d0028588SWanpeng Li struct hstate *h; 4768d0028588SWanpeng Li unsigned long nr_total_pages = 0; 4769d0028588SWanpeng Li 4770d0028588SWanpeng Li for_each_hstate(h) 4771d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 4772d0028588SWanpeng Li return nr_total_pages; 47731da177e4SLinus Torvalds } 47741da177e4SLinus Torvalds 4775a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 4776fc1b8a73SMel Gorman { 4777fc1b8a73SMel Gorman int ret = -ENOMEM; 4778fc1b8a73SMel Gorman 47790aa7f354SMiaohe Lin if (!delta) 47800aa7f354SMiaohe Lin return 0; 47810aa7f354SMiaohe Lin 4782db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4783fc1b8a73SMel Gorman /* 4784fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 4785fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 4786fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 4787fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 4788fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 4789fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 4790fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 4791fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 4792fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 4793fc1b8a73SMel Gorman * 4794fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 4795fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 4796fc1b8a73SMel Gorman * we fall back to check against current free page availability as 4797fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 4798fc1b8a73SMel Gorman * semantics that cpuset has. 47998ca39e68SMuchun Song * 48008ca39e68SMuchun Song * Apart from cpuset, we also have memory policy mechanism that 48018ca39e68SMuchun Song * also determines from which node the kernel will allocate memory 48028ca39e68SMuchun Song * in a NUMA system. So similar to cpuset, we also should consider 48038ca39e68SMuchun Song * the memory policy of the current task. Similar to the description 48048ca39e68SMuchun Song * above. 4805fc1b8a73SMel Gorman */ 4806fc1b8a73SMel Gorman if (delta > 0) { 4807a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 4808fc1b8a73SMel Gorman goto out; 4809fc1b8a73SMel Gorman 48108ca39e68SMuchun Song if (delta > allowed_mems_nr(h)) { 4811a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 4812fc1b8a73SMel Gorman goto out; 4813fc1b8a73SMel Gorman } 4814fc1b8a73SMel Gorman } 4815fc1b8a73SMel Gorman 4816fc1b8a73SMel Gorman ret = 0; 4817fc1b8a73SMel Gorman if (delta < 0) 4818a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 4819fc1b8a73SMel Gorman 4820fc1b8a73SMel Gorman out: 4821db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4822fc1b8a73SMel Gorman return ret; 4823fc1b8a73SMel Gorman } 4824fc1b8a73SMel Gorman 482584afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 482684afd99bSAndy Whitcroft { 4827f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 482884afd99bSAndy Whitcroft 482984afd99bSAndy Whitcroft /* 4830612b8a31SMike Kravetz * HPAGE_RESV_OWNER indicates a private mapping. 483184afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 483284afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 483384afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 483425985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 483584afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 483684afd99bSAndy Whitcroft * new reference here without additional locking. 483784afd99bSAndy Whitcroft */ 483809a26e83SMike Kravetz if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 483909a26e83SMike Kravetz resv_map_dup_hugetlb_cgroup_uncharge_info(resv); 4840f522c3acSJoonsoo Kim kref_get(&resv->refs); 484184afd99bSAndy Whitcroft } 48428d9bfb26SMike Kravetz 4843131a79b4SMike Kravetz /* 4844131a79b4SMike Kravetz * vma_lock structure for sharable mappings is vma specific. 4845612b8a31SMike Kravetz * Clear old pointer (if copied via vm_area_dup) and allocate 4846612b8a31SMike Kravetz * new structure. Before clearing, make sure vma_lock is not 4847612b8a31SMike Kravetz * for this vma. 4848131a79b4SMike Kravetz */ 4849131a79b4SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 4850612b8a31SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 4851612b8a31SMike Kravetz 4852612b8a31SMike Kravetz if (vma_lock) { 4853612b8a31SMike Kravetz if (vma_lock->vma != vma) { 4854131a79b4SMike Kravetz vma->vm_private_data = NULL; 48558d9bfb26SMike Kravetz hugetlb_vma_lock_alloc(vma); 4856612b8a31SMike Kravetz } else 4857612b8a31SMike Kravetz pr_warn("HugeTLB: vma_lock already exists in %s.\n", __func__); 4858612b8a31SMike Kravetz } else 4859612b8a31SMike Kravetz hugetlb_vma_lock_alloc(vma); 486009a26e83SMike Kravetz } 4861131a79b4SMike Kravetz } 486284afd99bSAndy Whitcroft 4863a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 4864a1e78772SMel Gorman { 4865a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 48668d9bfb26SMike Kravetz struct resv_map *resv; 486790481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 48684e35f483SJoonsoo Kim unsigned long reserve, start, end; 48691c5ecae3SMike Kravetz long gbl_reserve; 487084afd99bSAndy Whitcroft 48718d9bfb26SMike Kravetz hugetlb_vma_lock_free(vma); 48728d9bfb26SMike Kravetz 48738d9bfb26SMike Kravetz resv = vma_resv_map(vma); 48744e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 48754e35f483SJoonsoo Kim return; 48764e35f483SJoonsoo Kim 4877a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 4878a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 487984afd99bSAndy Whitcroft 48804e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 4881e9fe92aeSMina Almasry hugetlb_cgroup_uncharge_counter(resv, start, end); 48827251ff78SAdam Litke if (reserve) { 48831c5ecae3SMike Kravetz /* 48841c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 48851c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 48861c5ecae3SMike Kravetz */ 48871c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 48881c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 48897251ff78SAdam Litke } 4890e9fe92aeSMina Almasry 4891e9fe92aeSMina Almasry kref_put(&resv->refs, resv_map_release); 4892a1e78772SMel Gorman } 4893a1e78772SMel Gorman 489431383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 489531383c68SDan Williams { 489631383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 489731383c68SDan Williams return -EINVAL; 4898b30c14cdSJames Houghton 4899b30c14cdSJames Houghton /* 4900b30c14cdSJames Houghton * PMD sharing is only possible for PUD_SIZE-aligned address ranges 4901b30c14cdSJames Houghton * in HugeTLB VMAs. If we will lose PUD_SIZE alignment due to this 4902b30c14cdSJames Houghton * split, unshare PMDs in the PUD_SIZE interval surrounding addr now. 4903b30c14cdSJames Houghton */ 4904b30c14cdSJames Houghton if (addr & ~PUD_MASK) { 4905b30c14cdSJames Houghton /* 4906b30c14cdSJames Houghton * hugetlb_vm_op_split is called right before we attempt to 4907b30c14cdSJames Houghton * split the VMA. We will need to unshare PMDs in the old and 4908b30c14cdSJames Houghton * new VMAs, so let's unshare before we split. 4909b30c14cdSJames Houghton */ 4910b30c14cdSJames Houghton unsigned long floor = addr & PUD_MASK; 4911b30c14cdSJames Houghton unsigned long ceil = floor + PUD_SIZE; 4912b30c14cdSJames Houghton 4913b30c14cdSJames Houghton if (floor >= vma->vm_start && ceil <= vma->vm_end) 4914b30c14cdSJames Houghton hugetlb_unshare_pmds(vma, floor, ceil); 4915b30c14cdSJames Houghton } 4916b30c14cdSJames Houghton 491731383c68SDan Williams return 0; 491831383c68SDan Williams } 491931383c68SDan Williams 492005ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 492105ea8860SDan Williams { 4922aca78307SMiaohe Lin return huge_page_size(hstate_vma(vma)); 492305ea8860SDan Williams } 492405ea8860SDan Williams 49251da177e4SLinus Torvalds /* 49261da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 49271da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 49286c26d310SMiaohe Lin * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get 49291da177e4SLinus Torvalds * this far. 49301da177e4SLinus Torvalds */ 4931b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 49321da177e4SLinus Torvalds { 49331da177e4SLinus Torvalds BUG(); 4934d0217ac0SNick Piggin return 0; 49351da177e4SLinus Torvalds } 49361da177e4SLinus Torvalds 4937eec3636aSJane Chu /* 4938eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 4939eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 4940eec3636aSJane Chu * This is because under System V memory model, mappings created via 4941eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 4942eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 4943eec3636aSJane Chu */ 4944f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 4945d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 494684afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 4947a1e78772SMel Gorman .close = hugetlb_vm_op_close, 4948dd3b614fSDmitry Safonov .may_split = hugetlb_vm_op_split, 494905ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 49501da177e4SLinus Torvalds }; 49511da177e4SLinus Torvalds 49521e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 49531e8f889bSDavid Gibson int writable) 495463551ae0SDavid Gibson { 495563551ae0SDavid Gibson pte_t entry; 495679c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 495763551ae0SDavid Gibson 49581e8f889bSDavid Gibson if (writable) { 4959106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 4960106c992aSGerald Schaefer vma->vm_page_prot))); 496163551ae0SDavid Gibson } else { 4962106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 4963106c992aSGerald Schaefer vma->vm_page_prot)); 496463551ae0SDavid Gibson } 496563551ae0SDavid Gibson entry = pte_mkyoung(entry); 496679c1c594SChristophe Leroy entry = arch_make_huge_pte(entry, shift, vma->vm_flags); 496763551ae0SDavid Gibson 496863551ae0SDavid Gibson return entry; 496963551ae0SDavid Gibson } 497063551ae0SDavid Gibson 49711e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 49721e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 49731e8f889bSDavid Gibson { 49741e8f889bSDavid Gibson pte_t entry; 49751e8f889bSDavid Gibson 4976106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 497732f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 49784b3073e1SRussell King update_mmu_cache(vma, address, ptep); 49791e8f889bSDavid Gibson } 49801e8f889bSDavid Gibson 4981d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 49824a705fefSNaoya Horiguchi { 49834a705fefSNaoya Horiguchi swp_entry_t swp; 49844a705fefSNaoya Horiguchi 49854a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 4986d5ed7444SAneesh Kumar K.V return false; 49874a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4988d79d176aSBaoquan He if (is_migration_entry(swp)) 4989d5ed7444SAneesh Kumar K.V return true; 49904a705fefSNaoya Horiguchi else 4991d5ed7444SAneesh Kumar K.V return false; 49924a705fefSNaoya Horiguchi } 49934a705fefSNaoya Horiguchi 49943e5c3600SBaoquan He static bool is_hugetlb_entry_hwpoisoned(pte_t pte) 49954a705fefSNaoya Horiguchi { 49964a705fefSNaoya Horiguchi swp_entry_t swp; 49974a705fefSNaoya Horiguchi 49984a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 49993e5c3600SBaoquan He return false; 50004a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 5001d79d176aSBaoquan He if (is_hwpoison_entry(swp)) 50023e5c3600SBaoquan He return true; 50034a705fefSNaoya Horiguchi else 50043e5c3600SBaoquan He return false; 50054a705fefSNaoya Horiguchi } 50061e8f889bSDavid Gibson 50074eae4efaSPeter Xu static void 5008ea4c353dSSidhartha Kumar hugetlb_install_folio(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, 5009935d4f0cSRyan Roberts struct folio *new_folio, pte_t old, unsigned long sz) 50104eae4efaSPeter Xu { 50115a2f8d22SPeter Xu pte_t newpte = make_huge_pte(vma, &new_folio->page, 1); 50125a2f8d22SPeter Xu 5013ea4c353dSSidhartha Kumar __folio_mark_uptodate(new_folio); 5014d0ce0e47SSidhartha Kumar hugepage_add_new_anon_rmap(new_folio, vma, addr); 50155a2f8d22SPeter Xu if (userfaultfd_wp(vma) && huge_pte_uffd_wp(old)) 50165a2f8d22SPeter Xu newpte = huge_pte_mkuffd_wp(newpte); 5017935d4f0cSRyan Roberts set_huge_pte_at(vma->vm_mm, addr, ptep, newpte, sz); 50184eae4efaSPeter Xu hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); 5019ea4c353dSSidhartha Kumar folio_set_hugetlb_migratable(new_folio); 50204eae4efaSPeter Xu } 50214eae4efaSPeter Xu 502263551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 5023bc70fbf2SPeter Xu struct vm_area_struct *dst_vma, 5024bc70fbf2SPeter Xu struct vm_area_struct *src_vma) 502563551ae0SDavid Gibson { 50263aa4ed80SMiaohe Lin pte_t *src_pte, *dst_pte, entry; 5027ad27ce20SZhangPeng struct folio *pte_folio; 50281c59827dSHugh Dickins unsigned long addr; 5029bc70fbf2SPeter Xu bool cow = is_cow_mapping(src_vma->vm_flags); 5030bc70fbf2SPeter Xu struct hstate *h = hstate_vma(src_vma); 5031a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 50324eae4efaSPeter Xu unsigned long npages = pages_per_huge_page(h); 5033ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 5034e95a9851SMike Kravetz unsigned long last_addr_mask; 5035e8569dd2SAndreas Sandberg int ret = 0; 50361e8f889bSDavid Gibson 5037ac46d4f3SJérôme Glisse if (cow) { 50387d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src, 5039bc70fbf2SPeter Xu src_vma->vm_start, 5040bc70fbf2SPeter Xu src_vma->vm_end); 5041ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 5042e727bfd5SSuren Baghdasaryan vma_assert_write_locked(src_vma); 5043623a1ddfSDavid Hildenbrand raw_write_seqcount_begin(&src->write_protect_seq); 504440549ba8SMike Kravetz } else { 504540549ba8SMike Kravetz /* 504640549ba8SMike Kravetz * For shared mappings the vma lock must be held before 50479c67a207SPeter Xu * calling hugetlb_walk() in the src vma. Otherwise, the 504840549ba8SMike Kravetz * returned ptep could go away if part of a shared pmd and 504940549ba8SMike Kravetz * another thread calls huge_pmd_unshare. 505040549ba8SMike Kravetz */ 505140549ba8SMike Kravetz hugetlb_vma_lock_read(src_vma); 5052ac46d4f3SJérôme Glisse } 5053e8569dd2SAndreas Sandberg 5054e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 5055bc70fbf2SPeter Xu for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) { 5056cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 50579c67a207SPeter Xu src_pte = hugetlb_walk(src_vma, addr, sz); 5058e95a9851SMike Kravetz if (!src_pte) { 5059e95a9851SMike Kravetz addr |= last_addr_mask; 5060c74df32cSHugh Dickins continue; 5061e95a9851SMike Kravetz } 5062bc70fbf2SPeter Xu dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz); 5063e8569dd2SAndreas Sandberg if (!dst_pte) { 5064e8569dd2SAndreas Sandberg ret = -ENOMEM; 5065e8569dd2SAndreas Sandberg break; 5066e8569dd2SAndreas Sandberg } 5067c5c99429SLarry Woodman 50685e41540cSMike Kravetz /* 50695e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 50705e41540cSMike Kravetz * 50713aa4ed80SMiaohe Lin * dst_pte == src_pte is the common case of src/dest sharing. 50725e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 50733aa4ed80SMiaohe Lin * another vma. So page_count of ptep page is checked instead 50743aa4ed80SMiaohe Lin * to reliably determine whether pte is shared. 50755e41540cSMike Kravetz */ 50763aa4ed80SMiaohe Lin if (page_count(virt_to_page(dst_pte)) > 1) { 5077e95a9851SMike Kravetz addr |= last_addr_mask; 5078c5c99429SLarry Woodman continue; 5079e95a9851SMike Kravetz } 5080c5c99429SLarry Woodman 5081cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 5082cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 5083cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 50844a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 50854eae4efaSPeter Xu again: 50863aa4ed80SMiaohe Lin if (huge_pte_none(entry)) { 50875e41540cSMike Kravetz /* 50883aa4ed80SMiaohe Lin * Skip if src entry none. 50895e41540cSMike Kravetz */ 50904a705fefSNaoya Horiguchi ; 5091c2cb0dccSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) { 50925a2f8d22SPeter Xu if (!userfaultfd_wp(dst_vma)) 5093c2cb0dccSNaoya Horiguchi entry = huge_pte_clear_uffd_wp(entry); 5094935d4f0cSRyan Roberts set_huge_pte_at(dst, addr, dst_pte, entry, sz); 5095c2cb0dccSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry))) { 50964a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 50975a2f8d22SPeter Xu bool uffd_wp = pte_swp_uffd_wp(entry); 50984a705fefSNaoya Horiguchi 50996c287605SDavid Hildenbrand if (!is_readable_migration_entry(swp_entry) && cow) { 51004a705fefSNaoya Horiguchi /* 51014a705fefSNaoya Horiguchi * COW mappings require pages in both 51024a705fefSNaoya Horiguchi * parent and child to be set to read. 51034a705fefSNaoya Horiguchi */ 51044dd845b5SAlistair Popple swp_entry = make_readable_migration_entry( 51054dd845b5SAlistair Popple swp_offset(swp_entry)); 51064a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 5107bc70fbf2SPeter Xu if (userfaultfd_wp(src_vma) && uffd_wp) 51085a2f8d22SPeter Xu entry = pte_swp_mkuffd_wp(entry); 5109935d4f0cSRyan Roberts set_huge_pte_at(src, addr, src_pte, entry, sz); 51104a705fefSNaoya Horiguchi } 51115a2f8d22SPeter Xu if (!userfaultfd_wp(dst_vma)) 5112bc70fbf2SPeter Xu entry = huge_pte_clear_uffd_wp(entry); 5113935d4f0cSRyan Roberts set_huge_pte_at(dst, addr, dst_pte, entry, sz); 5114bc70fbf2SPeter Xu } else if (unlikely(is_pte_marker(entry))) { 5115af19487fSAxel Rasmussen pte_marker marker = copy_pte_marker( 5116af19487fSAxel Rasmussen pte_to_swp_entry(entry), dst_vma); 5117af19487fSAxel Rasmussen 5118af19487fSAxel Rasmussen if (marker) 5119af19487fSAxel Rasmussen set_huge_pte_at(dst, addr, dst_pte, 5120935d4f0cSRyan Roberts make_pte_marker(marker), sz); 51214a705fefSNaoya Horiguchi } else { 51224eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 5123ad27ce20SZhangPeng pte_folio = page_folio(pte_page(entry)); 5124ad27ce20SZhangPeng folio_get(pte_folio); 51254eae4efaSPeter Xu 51264eae4efaSPeter Xu /* 5127fb3d824dSDavid Hildenbrand * Failing to duplicate the anon rmap is a rare case 5128fb3d824dSDavid Hildenbrand * where we see pinned hugetlb pages while they're 5129fb3d824dSDavid Hildenbrand * prone to COW. We need to do the COW earlier during 5130fb3d824dSDavid Hildenbrand * fork. 51314eae4efaSPeter Xu * 51324eae4efaSPeter Xu * When pre-allocating the page or copying data, we 51334eae4efaSPeter Xu * need to be without the pgtable locks since we could 51344eae4efaSPeter Xu * sleep during the process. 51354eae4efaSPeter Xu */ 5136ad27ce20SZhangPeng if (!folio_test_anon(pte_folio)) { 5137ad27ce20SZhangPeng page_dup_file_rmap(&pte_folio->page, true); 5138ad27ce20SZhangPeng } else if (page_try_dup_anon_rmap(&pte_folio->page, 5139ad27ce20SZhangPeng true, src_vma)) { 51404eae4efaSPeter Xu pte_t src_pte_old = entry; 5141d0ce0e47SSidhartha Kumar struct folio *new_folio; 51424eae4efaSPeter Xu 51434eae4efaSPeter Xu spin_unlock(src_ptl); 51444eae4efaSPeter Xu spin_unlock(dst_ptl); 51454eae4efaSPeter Xu /* Do not use reserve as it's private owned */ 5146d0ce0e47SSidhartha Kumar new_folio = alloc_hugetlb_folio(dst_vma, addr, 1); 5147d0ce0e47SSidhartha Kumar if (IS_ERR(new_folio)) { 5148ad27ce20SZhangPeng folio_put(pte_folio); 5149d0ce0e47SSidhartha Kumar ret = PTR_ERR(new_folio); 51504eae4efaSPeter Xu break; 51514eae4efaSPeter Xu } 51521cb9dc4bSLiu Shixin ret = copy_user_large_folio(new_folio, 5153ad27ce20SZhangPeng pte_folio, 5154c0e8150eSZhangPeng addr, dst_vma); 5155ad27ce20SZhangPeng folio_put(pte_folio); 51561cb9dc4bSLiu Shixin if (ret) { 51571cb9dc4bSLiu Shixin folio_put(new_folio); 51581cb9dc4bSLiu Shixin break; 51591cb9dc4bSLiu Shixin } 51604eae4efaSPeter Xu 5161d0ce0e47SSidhartha Kumar /* Install the new hugetlb folio if src pte stable */ 51624eae4efaSPeter Xu dst_ptl = huge_pte_lock(h, dst, dst_pte); 51634eae4efaSPeter Xu src_ptl = huge_pte_lockptr(h, src, src_pte); 51644eae4efaSPeter Xu spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 51654eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 51664eae4efaSPeter Xu if (!pte_same(src_pte_old, entry)) { 5167bc70fbf2SPeter Xu restore_reserve_on_error(h, dst_vma, addr, 5168d2d7bb44SSidhartha Kumar new_folio); 5169d0ce0e47SSidhartha Kumar folio_put(new_folio); 51703aa4ed80SMiaohe Lin /* huge_ptep of dst_pte won't change as in child */ 51714eae4efaSPeter Xu goto again; 51724eae4efaSPeter Xu } 51735a2f8d22SPeter Xu hugetlb_install_folio(dst_vma, dst_pte, addr, 5174935d4f0cSRyan Roberts new_folio, src_pte_old, sz); 51754eae4efaSPeter Xu spin_unlock(src_ptl); 51764eae4efaSPeter Xu spin_unlock(dst_ptl); 51774eae4efaSPeter Xu continue; 51784eae4efaSPeter Xu } 51794eae4efaSPeter Xu 518034ee645eSJoerg Roedel if (cow) { 51810f10851eSJérôme Glisse /* 51820f10851eSJérôme Glisse * No need to notify as we are downgrading page 51830f10851eSJérôme Glisse * table protection not changing it to point 51840f10851eSJérôme Glisse * to a new page. 51850f10851eSJérôme Glisse * 5186ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 51870f10851eSJérôme Glisse */ 51887f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 518984894e1cSPeter Xu entry = huge_pte_wrprotect(entry); 519034ee645eSJoerg Roedel } 51914eae4efaSPeter Xu 51925a2f8d22SPeter Xu if (!userfaultfd_wp(dst_vma)) 51935a2f8d22SPeter Xu entry = huge_pte_clear_uffd_wp(entry); 51945a2f8d22SPeter Xu 5195935d4f0cSRyan Roberts set_huge_pte_at(dst, addr, dst_pte, entry, sz); 51964eae4efaSPeter Xu hugetlb_count_add(npages, dst); 51971c59827dSHugh Dickins } 5198cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 5199cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 520063551ae0SDavid Gibson } 520163551ae0SDavid Gibson 5202623a1ddfSDavid Hildenbrand if (cow) { 5203623a1ddfSDavid Hildenbrand raw_write_seqcount_end(&src->write_protect_seq); 5204ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 520540549ba8SMike Kravetz } else { 520640549ba8SMike Kravetz hugetlb_vma_unlock_read(src_vma); 5207623a1ddfSDavid Hildenbrand } 5208e8569dd2SAndreas Sandberg 5209e8569dd2SAndreas Sandberg return ret; 521063551ae0SDavid Gibson } 521163551ae0SDavid Gibson 5212550a7d60SMina Almasry static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr, 5213935d4f0cSRyan Roberts unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte, 5214935d4f0cSRyan Roberts unsigned long sz) 5215550a7d60SMina Almasry { 5216550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 5217550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 5218550a7d60SMina Almasry spinlock_t *src_ptl, *dst_ptl; 5219db110a99SAneesh Kumar K.V pte_t pte; 5220550a7d60SMina Almasry 5221550a7d60SMina Almasry dst_ptl = huge_pte_lock(h, mm, dst_pte); 5222550a7d60SMina Almasry src_ptl = huge_pte_lockptr(h, mm, src_pte); 5223550a7d60SMina Almasry 5224550a7d60SMina Almasry /* 5225550a7d60SMina Almasry * We don't have to worry about the ordering of src and dst ptlocks 52268651a137SLorenzo Stoakes * because exclusive mmap_lock (or the i_mmap_lock) prevents deadlock. 5227550a7d60SMina Almasry */ 5228550a7d60SMina Almasry if (src_ptl != dst_ptl) 5229550a7d60SMina Almasry spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 5230550a7d60SMina Almasry 5231550a7d60SMina Almasry pte = huge_ptep_get_and_clear(mm, old_addr, src_pte); 5232935d4f0cSRyan Roberts set_huge_pte_at(mm, new_addr, dst_pte, pte, sz); 5233550a7d60SMina Almasry 5234550a7d60SMina Almasry if (src_ptl != dst_ptl) 5235550a7d60SMina Almasry spin_unlock(src_ptl); 5236550a7d60SMina Almasry spin_unlock(dst_ptl); 5237550a7d60SMina Almasry } 5238550a7d60SMina Almasry 5239550a7d60SMina Almasry int move_hugetlb_page_tables(struct vm_area_struct *vma, 5240550a7d60SMina Almasry struct vm_area_struct *new_vma, 5241550a7d60SMina Almasry unsigned long old_addr, unsigned long new_addr, 5242550a7d60SMina Almasry unsigned long len) 5243550a7d60SMina Almasry { 5244550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 5245550a7d60SMina Almasry struct address_space *mapping = vma->vm_file->f_mapping; 5246550a7d60SMina Almasry unsigned long sz = huge_page_size(h); 5247550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 5248550a7d60SMina Almasry unsigned long old_end = old_addr + len; 5249e95a9851SMike Kravetz unsigned long last_addr_mask; 5250550a7d60SMina Almasry pte_t *src_pte, *dst_pte; 5251550a7d60SMina Almasry struct mmu_notifier_range range; 52523d0b95cdSBaolin Wang bool shared_pmd = false; 5253550a7d60SMina Almasry 52547d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, old_addr, 5255550a7d60SMina Almasry old_end); 5256550a7d60SMina Almasry adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 52573d0b95cdSBaolin Wang /* 52583d0b95cdSBaolin Wang * In case of shared PMDs, we should cover the maximum possible 52593d0b95cdSBaolin Wang * range. 52603d0b95cdSBaolin Wang */ 52613d0b95cdSBaolin Wang flush_cache_range(vma, range.start, range.end); 52623d0b95cdSBaolin Wang 5263550a7d60SMina Almasry mmu_notifier_invalidate_range_start(&range); 5264e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 5265550a7d60SMina Almasry /* Prevent race with file truncation */ 526640549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 5267550a7d60SMina Almasry i_mmap_lock_write(mapping); 5268550a7d60SMina Almasry for (; old_addr < old_end; old_addr += sz, new_addr += sz) { 52699c67a207SPeter Xu src_pte = hugetlb_walk(vma, old_addr, sz); 5270e95a9851SMike Kravetz if (!src_pte) { 5271e95a9851SMike Kravetz old_addr |= last_addr_mask; 5272e95a9851SMike Kravetz new_addr |= last_addr_mask; 5273550a7d60SMina Almasry continue; 5274e95a9851SMike Kravetz } 5275550a7d60SMina Almasry if (huge_pte_none(huge_ptep_get(src_pte))) 5276550a7d60SMina Almasry continue; 5277550a7d60SMina Almasry 52784ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) { 52793d0b95cdSBaolin Wang shared_pmd = true; 52804ddb4d91SMike Kravetz old_addr |= last_addr_mask; 52814ddb4d91SMike Kravetz new_addr |= last_addr_mask; 5282550a7d60SMina Almasry continue; 52833d0b95cdSBaolin Wang } 5284550a7d60SMina Almasry 5285550a7d60SMina Almasry dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz); 5286550a7d60SMina Almasry if (!dst_pte) 5287550a7d60SMina Almasry break; 5288550a7d60SMina Almasry 5289935d4f0cSRyan Roberts move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte, sz); 5290550a7d60SMina Almasry } 52913d0b95cdSBaolin Wang 52923d0b95cdSBaolin Wang if (shared_pmd) 5293f720b471SKefeng Wang flush_hugetlb_tlb_range(vma, range.start, range.end); 52943d0b95cdSBaolin Wang else 5295f720b471SKefeng Wang flush_hugetlb_tlb_range(vma, old_end - len, old_end); 5296550a7d60SMina Almasry mmu_notifier_invalidate_range_end(&range); 529713e4ad2cSNadav Amit i_mmap_unlock_write(mapping); 529840549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 5299550a7d60SMina Almasry 5300550a7d60SMina Almasry return len + old_addr - old_end; 5301550a7d60SMina Almasry } 5302550a7d60SMina Almasry 53032820b0f0SRik van Riel void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 530424669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 530505e90bd0SPeter Xu struct page *ref_page, zap_flags_t zap_flags) 530663551ae0SDavid Gibson { 530763551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 530863551ae0SDavid Gibson unsigned long address; 5309c7546f8fSDavid Gibson pte_t *ptep; 531063551ae0SDavid Gibson pte_t pte; 5311cb900f41SKirill A. Shutemov spinlock_t *ptl; 531263551ae0SDavid Gibson struct page *page; 5313a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 5314a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 5315e95a9851SMike Kravetz unsigned long last_addr_mask; 5316a4a118f2SNadav Amit bool force_flush = false; 5317a5516438SAndi Kleen 531863551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 5319a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 5320a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 532163551ae0SDavid Gibson 532207e32661SAneesh Kumar K.V /* 532307e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 532407e32661SAneesh Kumar K.V * to huge page. 532507e32661SAneesh Kumar K.V */ 5326ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 532724669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 5328dff11abeSMike Kravetz 5329e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 5330569f48b8SHillf Danton address = start; 5331569f48b8SHillf Danton for (; address < end; address += sz) { 53329c67a207SPeter Xu ptep = hugetlb_walk(vma, address, sz); 5333e95a9851SMike Kravetz if (!ptep) { 5334e95a9851SMike Kravetz address |= last_addr_mask; 5335c7546f8fSDavid Gibson continue; 5336e95a9851SMike Kravetz } 5337c7546f8fSDavid Gibson 5338cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 53394ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, address, ptep)) { 534031d49da5SAneesh Kumar K.V spin_unlock(ptl); 5341a4a118f2SNadav Amit tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE); 5342a4a118f2SNadav Amit force_flush = true; 53434ddb4d91SMike Kravetz address |= last_addr_mask; 534431d49da5SAneesh Kumar K.V continue; 534531d49da5SAneesh Kumar K.V } 534639dde65cSChen, Kenneth W 53476629326bSHillf Danton pte = huge_ptep_get(ptep); 534831d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 534931d49da5SAneesh Kumar K.V spin_unlock(ptl); 535031d49da5SAneesh Kumar K.V continue; 535131d49da5SAneesh Kumar K.V } 53526629326bSHillf Danton 53536629326bSHillf Danton /* 53549fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 53559fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 53566629326bSHillf Danton */ 53579fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 535805e90bd0SPeter Xu /* 535905e90bd0SPeter Xu * If the pte was wr-protected by uffd-wp in any of the 536005e90bd0SPeter Xu * swap forms, meanwhile the caller does not want to 536105e90bd0SPeter Xu * drop the uffd-wp bit in this zap, then replace the 536205e90bd0SPeter Xu * pte with a marker. 536305e90bd0SPeter Xu */ 536405e90bd0SPeter Xu if (pte_swp_uffd_wp_any(pte) && 536505e90bd0SPeter Xu !(zap_flags & ZAP_FLAG_DROP_MARKER)) 536605e90bd0SPeter Xu set_huge_pte_at(mm, address, ptep, 5367935d4f0cSRyan Roberts make_pte_marker(PTE_MARKER_UFFD_WP), 5368935d4f0cSRyan Roberts sz); 536905e90bd0SPeter Xu else 53709386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 537131d49da5SAneesh Kumar K.V spin_unlock(ptl); 537231d49da5SAneesh Kumar K.V continue; 53738c4894c6SNaoya Horiguchi } 53746629326bSHillf Danton 53756629326bSHillf Danton page = pte_page(pte); 537604f2cbe3SMel Gorman /* 537704f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 537804f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 537904f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 538004f2cbe3SMel Gorman */ 538104f2cbe3SMel Gorman if (ref_page) { 538231d49da5SAneesh Kumar K.V if (page != ref_page) { 538331d49da5SAneesh Kumar K.V spin_unlock(ptl); 538431d49da5SAneesh Kumar K.V continue; 538531d49da5SAneesh Kumar K.V } 538604f2cbe3SMel Gorman /* 538704f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 538804f2cbe3SMel Gorman * future faults in this VMA will fail rather than 538904f2cbe3SMel Gorman * looking like data was lost 539004f2cbe3SMel Gorman */ 539104f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 539204f2cbe3SMel Gorman } 539304f2cbe3SMel Gorman 5394c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 5395b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 5396106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 53976649a386SKen Chen set_page_dirty(page); 539805e90bd0SPeter Xu /* Leave a uffd-wp pte marker if needed */ 539905e90bd0SPeter Xu if (huge_pte_uffd_wp(pte) && 540005e90bd0SPeter Xu !(zap_flags & ZAP_FLAG_DROP_MARKER)) 540105e90bd0SPeter Xu set_huge_pte_at(mm, address, ptep, 5402935d4f0cSRyan Roberts make_pte_marker(PTE_MARKER_UFFD_WP), 5403935d4f0cSRyan Roberts sz); 54045d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 5405cea86fe2SHugh Dickins page_remove_rmap(page, vma, true); 540631d49da5SAneesh Kumar K.V 5407cb900f41SKirill A. Shutemov spin_unlock(ptl); 5408e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 540924669e58SAneesh Kumar K.V /* 541031d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 541124669e58SAneesh Kumar K.V */ 541231d49da5SAneesh Kumar K.V if (ref_page) 541331d49da5SAneesh Kumar K.V break; 5414fe1668aeSChen, Kenneth W } 541524669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 5416a4a118f2SNadav Amit 5417a4a118f2SNadav Amit /* 5418a4a118f2SNadav Amit * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We 5419a4a118f2SNadav Amit * could defer the flush until now, since by holding i_mmap_rwsem we 5420a4a118f2SNadav Amit * guaranteed that the last refernece would not be dropped. But we must 5421a4a118f2SNadav Amit * do the flushing before we return, as otherwise i_mmap_rwsem will be 5422a4a118f2SNadav Amit * dropped and the last reference to the shared PMDs page might be 5423a4a118f2SNadav Amit * dropped as well. 5424a4a118f2SNadav Amit * 5425a4a118f2SNadav Amit * In theory we could defer the freeing of the PMD pages as well, but 5426a4a118f2SNadav Amit * huge_pmd_unshare() relies on the exact page_count for the PMD page to 5427a4a118f2SNadav Amit * detect sharing, so we cannot defer the release of the page either. 5428a4a118f2SNadav Amit * Instead, do flush now. 5429a4a118f2SNadav Amit */ 5430a4a118f2SNadav Amit if (force_flush) 5431a4a118f2SNadav Amit tlb_flush_mmu_tlbonly(tlb); 54321da177e4SLinus Torvalds } 543363551ae0SDavid Gibson 54342820b0f0SRik van Riel void __hugetlb_zap_begin(struct vm_area_struct *vma, 54352820b0f0SRik van Riel unsigned long *start, unsigned long *end) 5436d833352aSMel Gorman { 54372820b0f0SRik van Riel if (!vma->vm_file) /* hugetlbfs_file_mmap error */ 54382820b0f0SRik van Riel return; 5439131a79b4SMike Kravetz 54402820b0f0SRik van Riel adjust_range_if_pmd_sharing_possible(vma, start, end); 54412820b0f0SRik van Riel hugetlb_vma_lock_write(vma); 54422820b0f0SRik van Riel if (vma->vm_file) 54432820b0f0SRik van Riel i_mmap_lock_write(vma->vm_file->f_mapping); 54442820b0f0SRik van Riel } 54452820b0f0SRik van Riel 54462820b0f0SRik van Riel void __hugetlb_zap_end(struct vm_area_struct *vma, 54472820b0f0SRik van Riel struct zap_details *details) 54482820b0f0SRik van Riel { 54492820b0f0SRik van Riel zap_flags_t zap_flags = details ? details->zap_flags : 0; 54502820b0f0SRik van Riel 54512820b0f0SRik van Riel if (!vma->vm_file) /* hugetlbfs_file_mmap error */ 54522820b0f0SRik van Riel return; 5453d833352aSMel Gorman 545404ada095SMike Kravetz if (zap_flags & ZAP_FLAG_UNMAP) { /* final unmap */ 5455d833352aSMel Gorman /* 545604ada095SMike Kravetz * Unlock and free the vma lock before releasing i_mmap_rwsem. 545704ada095SMike Kravetz * When the vma_lock is freed, this makes the vma ineligible 545804ada095SMike Kravetz * for pmd sharing. And, i_mmap_rwsem is required to set up 545904ada095SMike Kravetz * pmd sharing. This is important as page tables for this 546004ada095SMike Kravetz * unmapped range will be asynchrously deleted. If the page 546104ada095SMike Kravetz * tables are shared, there will be issues when accessed by 546204ada095SMike Kravetz * someone else. 5463d833352aSMel Gorman */ 5464ecfbd733SMike Kravetz __hugetlb_vma_unlock_write_free(vma); 546504ada095SMike Kravetz } else { 546604ada095SMike Kravetz hugetlb_vma_unlock_write(vma); 546704ada095SMike Kravetz } 54682820b0f0SRik van Riel 54692820b0f0SRik van Riel if (vma->vm_file) 54702820b0f0SRik van Riel i_mmap_unlock_write(vma->vm_file->f_mapping); 5471d833352aSMel Gorman } 5472d833352aSMel Gorman 5473502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 547405e90bd0SPeter Xu unsigned long end, struct page *ref_page, 547505e90bd0SPeter Xu zap_flags_t zap_flags) 5476502717f4SChen, Kenneth W { 5477369258ceSMike Kravetz struct mmu_notifier_range range; 547824669e58SAneesh Kumar K.V struct mmu_gather tlb; 5479dff11abeSMike Kravetz 54807d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm, 5481369258ceSMike Kravetz start, end); 5482369258ceSMike Kravetz adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 5483369258ceSMike Kravetz mmu_notifier_invalidate_range_start(&range); 5484a72afd87SWill Deacon tlb_gather_mmu(&tlb, vma->vm_mm); 5485369258ceSMike Kravetz 548605e90bd0SPeter Xu __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags); 5487369258ceSMike Kravetz 5488369258ceSMike Kravetz mmu_notifier_invalidate_range_end(&range); 5489ae8eba8bSWill Deacon tlb_finish_mmu(&tlb); 5490502717f4SChen, Kenneth W } 5491502717f4SChen, Kenneth W 549204f2cbe3SMel Gorman /* 549304f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 5494578b7725SZhiyuan Dai * mapping it owns the reserve page for. The intention is to unmap the page 549504f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 549604f2cbe3SMel Gorman * same region. 549704f2cbe3SMel Gorman */ 54982f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 54992a4b3dedSHarvey Harrison struct page *page, unsigned long address) 550004f2cbe3SMel Gorman { 55017526674dSAdam Litke struct hstate *h = hstate_vma(vma); 550204f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 550304f2cbe3SMel Gorman struct address_space *mapping; 550404f2cbe3SMel Gorman pgoff_t pgoff; 550504f2cbe3SMel Gorman 550604f2cbe3SMel Gorman /* 550704f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 550804f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 550904f2cbe3SMel Gorman */ 55107526674dSAdam Litke address = address & huge_page_mask(h); 551136e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 551236e4f20aSMichal Hocko vma->vm_pgoff; 551393c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 551404f2cbe3SMel Gorman 55154eb2b1dcSMel Gorman /* 55164eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 55174eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 55184eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 55194eb2b1dcSMel Gorman */ 552083cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 55216b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 552204f2cbe3SMel Gorman /* Do not unmap the current VMA */ 552304f2cbe3SMel Gorman if (iter_vma == vma) 552404f2cbe3SMel Gorman continue; 552504f2cbe3SMel Gorman 552604f2cbe3SMel Gorman /* 55272f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 55282f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 55292f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 55302f84a899SMel Gorman */ 55312f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 55322f84a899SMel Gorman continue; 55332f84a899SMel Gorman 55342f84a899SMel Gorman /* 553504f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 553604f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 553704f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 553804f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 553904f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 554004f2cbe3SMel Gorman */ 554104f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 554224669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 554305e90bd0SPeter Xu address + huge_page_size(h), page, 0); 554404f2cbe3SMel Gorman } 554583cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 554604f2cbe3SMel Gorman } 554704f2cbe3SMel Gorman 55480fe6e20bSNaoya Horiguchi /* 5549c89357e2SDavid Hildenbrand * hugetlb_wp() should be called with page lock of the original hugepage held. 5550aa6d2e8cSBaolin Wang * Called with hugetlb_fault_mutex_table held and pte_page locked so we 5551ef009b25SMichal Hocko * cannot race with other handlers or page migration. 5552ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 55530fe6e20bSNaoya Horiguchi */ 5554c89357e2SDavid Hildenbrand static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma, 5555c89357e2SDavid Hildenbrand unsigned long address, pte_t *ptep, unsigned int flags, 5556371607a3SSidhartha Kumar struct folio *pagecache_folio, spinlock_t *ptl) 55571e8f889bSDavid Gibson { 5558c89357e2SDavid Hildenbrand const bool unshare = flags & FAULT_FLAG_UNSHARE; 555960d5b473SPeter Xu pte_t pte = huge_ptep_get(ptep); 5560a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 5561959a78b6SZhangPeng struct folio *old_folio; 5562d0ce0e47SSidhartha Kumar struct folio *new_folio; 55632b740303SSouptick Joarder int outside_reserve = 0; 55642b740303SSouptick Joarder vm_fault_t ret = 0; 5565974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 5566ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 55671e8f889bSDavid Gibson 55681d8d1464SDavid Hildenbrand /* 556960d5b473SPeter Xu * Never handle CoW for uffd-wp protected pages. It should be only 557060d5b473SPeter Xu * handled when the uffd-wp protection is removed. 557160d5b473SPeter Xu * 557260d5b473SPeter Xu * Note that only the CoW optimization path (in hugetlb_no_page()) 557360d5b473SPeter Xu * can trigger this, because hugetlb_fault() will always resolve 557460d5b473SPeter Xu * uffd-wp bit first. 557560d5b473SPeter Xu */ 557660d5b473SPeter Xu if (!unshare && huge_pte_uffd_wp(pte)) 557760d5b473SPeter Xu return 0; 557860d5b473SPeter Xu 557960d5b473SPeter Xu /* 55801d8d1464SDavid Hildenbrand * hugetlb does not support FOLL_FORCE-style write faults that keep the 55811d8d1464SDavid Hildenbrand * PTE mapped R/O such as maybe_mkwrite() would do. 55821d8d1464SDavid Hildenbrand */ 55831d8d1464SDavid Hildenbrand if (WARN_ON_ONCE(!unshare && !(vma->vm_flags & VM_WRITE))) 55841d8d1464SDavid Hildenbrand return VM_FAULT_SIGSEGV; 55851d8d1464SDavid Hildenbrand 55861d8d1464SDavid Hildenbrand /* Let's take out MAP_SHARED mappings first. */ 55871d8d1464SDavid Hildenbrand if (vma->vm_flags & VM_MAYSHARE) { 55881d8d1464SDavid Hildenbrand set_huge_ptep_writable(vma, haddr, ptep); 55891d8d1464SDavid Hildenbrand return 0; 55901d8d1464SDavid Hildenbrand } 55911d8d1464SDavid Hildenbrand 5592959a78b6SZhangPeng old_folio = page_folio(pte_page(pte)); 55931e8f889bSDavid Gibson 5594662ce1dcSYang Yang delayacct_wpcopy_start(); 5595662ce1dcSYang Yang 559604f2cbe3SMel Gorman retry_avoidcopy: 5597c89357e2SDavid Hildenbrand /* 5598c89357e2SDavid Hildenbrand * If no-one else is actually using this page, we're the exclusive 5599c89357e2SDavid Hildenbrand * owner and can reuse this page. 5600c89357e2SDavid Hildenbrand */ 5601959a78b6SZhangPeng if (folio_mapcount(old_folio) == 1 && folio_test_anon(old_folio)) { 5602959a78b6SZhangPeng if (!PageAnonExclusive(&old_folio->page)) 5603959a78b6SZhangPeng page_move_anon_rmap(&old_folio->page, vma); 5604c89357e2SDavid Hildenbrand if (likely(!unshare)) 56055b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 5606662ce1dcSYang Yang 5607662ce1dcSYang Yang delayacct_wpcopy_end(); 560883c54070SNick Piggin return 0; 56091e8f889bSDavid Gibson } 5610959a78b6SZhangPeng VM_BUG_ON_PAGE(folio_test_anon(old_folio) && 5611959a78b6SZhangPeng PageAnonExclusive(&old_folio->page), &old_folio->page); 56121e8f889bSDavid Gibson 561304f2cbe3SMel Gorman /* 561404f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 561504f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 561604f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 561704f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 561804f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 561904f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 562004f2cbe3SMel Gorman * of the full address range. 562104f2cbe3SMel Gorman */ 56225944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 5623959a78b6SZhangPeng old_folio != pagecache_folio) 562404f2cbe3SMel Gorman outside_reserve = 1; 562504f2cbe3SMel Gorman 5626959a78b6SZhangPeng folio_get(old_folio); 5627b76c8cfbSLarry Woodman 5628ad4404a2SDavidlohr Bueso /* 5629ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 5630ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 5631ad4404a2SDavidlohr Bueso */ 5632cb900f41SKirill A. Shutemov spin_unlock(ptl); 5633d0ce0e47SSidhartha Kumar new_folio = alloc_hugetlb_folio(vma, haddr, outside_reserve); 56341e8f889bSDavid Gibson 5635d0ce0e47SSidhartha Kumar if (IS_ERR(new_folio)) { 563604f2cbe3SMel Gorman /* 563704f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 563804f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 563904f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 564004f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 564104f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 564204f2cbe3SMel Gorman */ 564304f2cbe3SMel Gorman if (outside_reserve) { 564440549ba8SMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 564540549ba8SMike Kravetz pgoff_t idx; 564640549ba8SMike Kravetz u32 hash; 564740549ba8SMike Kravetz 5648959a78b6SZhangPeng folio_put(old_folio); 564940549ba8SMike Kravetz /* 565040549ba8SMike Kravetz * Drop hugetlb_fault_mutex and vma_lock before 565140549ba8SMike Kravetz * unmapping. unmapping needs to hold vma_lock 565240549ba8SMike Kravetz * in write mode. Dropping vma_lock in read mode 565340549ba8SMike Kravetz * here is OK as COW mappings do not interact with 565440549ba8SMike Kravetz * PMD sharing. 565540549ba8SMike Kravetz * 565640549ba8SMike Kravetz * Reacquire both after unmap operation. 565740549ba8SMike Kravetz */ 565840549ba8SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 565940549ba8SMike Kravetz hash = hugetlb_fault_mutex_hash(mapping, idx); 566040549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 566140549ba8SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 566240549ba8SMike Kravetz 5663959a78b6SZhangPeng unmap_ref_private(mm, vma, &old_folio->page, haddr); 566440549ba8SMike Kravetz 566540549ba8SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 566640549ba8SMike Kravetz hugetlb_vma_lock_read(vma); 5667cb900f41SKirill A. Shutemov spin_lock(ptl); 56689c67a207SPeter Xu ptep = hugetlb_walk(vma, haddr, huge_page_size(h)); 5669a9af0c5dSNaoya Horiguchi if (likely(ptep && 5670a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 567104f2cbe3SMel Gorman goto retry_avoidcopy; 5672a734bcc8SHillf Danton /* 5673cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 5674cb900f41SKirill A. Shutemov * lock, and our job is done. 5675a734bcc8SHillf Danton */ 5676662ce1dcSYang Yang delayacct_wpcopy_end(); 5677a734bcc8SHillf Danton return 0; 567804f2cbe3SMel Gorman } 567904f2cbe3SMel Gorman 5680d0ce0e47SSidhartha Kumar ret = vmf_error(PTR_ERR(new_folio)); 5681ad4404a2SDavidlohr Bueso goto out_release_old; 56821e8f889bSDavid Gibson } 56831e8f889bSDavid Gibson 56840fe6e20bSNaoya Horiguchi /* 56850fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 56860fe6e20bSNaoya Horiguchi * anon_vma prepared. 56870fe6e20bSNaoya Horiguchi */ 568844e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 5689ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 5690ad4404a2SDavidlohr Bueso goto out_release_all; 569144e2aa93SDean Nelson } 56920fe6e20bSNaoya Horiguchi 5693959a78b6SZhangPeng if (copy_user_large_folio(new_folio, old_folio, address, vma)) { 56941cb9dc4bSLiu Shixin ret = VM_FAULT_HWPOISON_LARGE; 56951cb9dc4bSLiu Shixin goto out_release_all; 56961cb9dc4bSLiu Shixin } 5697d0ce0e47SSidhartha Kumar __folio_mark_uptodate(new_folio); 56981e8f889bSDavid Gibson 56997d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, haddr, 57006f4f13e8SJérôme Glisse haddr + huge_page_size(h)); 5701ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 5702ad4404a2SDavidlohr Bueso 5703b76c8cfbSLarry Woodman /* 5704cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 5705b76c8cfbSLarry Woodman * before the page tables are altered 5706b76c8cfbSLarry Woodman */ 5707cb900f41SKirill A. Shutemov spin_lock(ptl); 57089c67a207SPeter Xu ptep = hugetlb_walk(vma, haddr, huge_page_size(h)); 5709a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 57100f230bc2SPeter Xu pte_t newpte = make_huge_pte(vma, &new_folio->page, !unshare); 57110f230bc2SPeter Xu 5712c89357e2SDavid Hildenbrand /* Break COW or unshare */ 57135b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 5714959a78b6SZhangPeng page_remove_rmap(&old_folio->page, vma, true); 5715d0ce0e47SSidhartha Kumar hugepage_add_new_anon_rmap(new_folio, vma, haddr); 57160f230bc2SPeter Xu if (huge_pte_uffd_wp(pte)) 57170f230bc2SPeter Xu newpte = huge_pte_mkuffd_wp(newpte); 5718935d4f0cSRyan Roberts set_huge_pte_at(mm, haddr, ptep, newpte, huge_page_size(h)); 5719d0ce0e47SSidhartha Kumar folio_set_hugetlb_migratable(new_folio); 57201e8f889bSDavid Gibson /* Make the old page be freed below */ 5721959a78b6SZhangPeng new_folio = old_folio; 57221e8f889bSDavid Gibson } 5723cb900f41SKirill A. Shutemov spin_unlock(ptl); 5724ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 5725ad4404a2SDavidlohr Bueso out_release_all: 5726c89357e2SDavid Hildenbrand /* 5727c89357e2SDavid Hildenbrand * No restore in case of successful pagetable update (Break COW or 5728c89357e2SDavid Hildenbrand * unshare) 5729c89357e2SDavid Hildenbrand */ 5730959a78b6SZhangPeng if (new_folio != old_folio) 5731d2d7bb44SSidhartha Kumar restore_reserve_on_error(h, vma, haddr, new_folio); 5732d0ce0e47SSidhartha Kumar folio_put(new_folio); 5733ad4404a2SDavidlohr Bueso out_release_old: 5734959a78b6SZhangPeng folio_put(old_folio); 57358312034fSJoonsoo Kim 5736ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 5737662ce1dcSYang Yang 5738662ce1dcSYang Yang delayacct_wpcopy_end(); 5739ad4404a2SDavidlohr Bueso return ret; 57401e8f889bSDavid Gibson } 57411e8f889bSDavid Gibson 57423ae77f43SHugh Dickins /* 57433ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 57443ae77f43SHugh Dickins */ 57453ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 57462a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 57472a15efc9SHugh Dickins { 574891a2fb95SSidhartha Kumar struct address_space *mapping = vma->vm_file->f_mapping; 574991a2fb95SSidhartha Kumar pgoff_t idx = vma_hugecache_offset(h, vma, address); 5750fd4aed8dSMike Kravetz struct folio *folio; 57512a15efc9SHugh Dickins 5752fd4aed8dSMike Kravetz folio = filemap_get_folio(mapping, idx); 5753fd4aed8dSMike Kravetz if (IS_ERR(folio)) 5754fd4aed8dSMike Kravetz return false; 5755fd4aed8dSMike Kravetz folio_put(folio); 5756fd4aed8dSMike Kravetz return true; 57572a15efc9SHugh Dickins } 57582a15efc9SHugh Dickins 57599b91c0e2SSidhartha Kumar int hugetlb_add_to_page_cache(struct folio *folio, struct address_space *mapping, 5760ab76ad54SMike Kravetz pgoff_t idx) 5761ab76ad54SMike Kravetz { 5762ab76ad54SMike Kravetz struct inode *inode = mapping->host; 5763ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 5764d9ef44deSMatthew Wilcox (Oracle) int err; 5765ab76ad54SMike Kravetz 5766d9ef44deSMatthew Wilcox (Oracle) __folio_set_locked(folio); 5767d9ef44deSMatthew Wilcox (Oracle) err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL); 5768d9ef44deSMatthew Wilcox (Oracle) 5769d9ef44deSMatthew Wilcox (Oracle) if (unlikely(err)) { 5770d9ef44deSMatthew Wilcox (Oracle) __folio_clear_locked(folio); 5771ab76ad54SMike Kravetz return err; 5772d9ef44deSMatthew Wilcox (Oracle) } 57739b91c0e2SSidhartha Kumar folio_clear_hugetlb_restore_reserve(folio); 5774ab76ad54SMike Kravetz 577522146c3cSMike Kravetz /* 5776d9ef44deSMatthew Wilcox (Oracle) * mark folio dirty so that it will not be removed from cache/file 577722146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 577822146c3cSMike Kravetz */ 5779d9ef44deSMatthew Wilcox (Oracle) folio_mark_dirty(folio); 578022146c3cSMike Kravetz 5781ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 5782ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 5783ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 5784ab76ad54SMike Kravetz return 0; 5785ab76ad54SMike Kravetz } 5786ab76ad54SMike Kravetz 57877677f7fdSAxel Rasmussen static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma, 57887677f7fdSAxel Rasmussen struct address_space *mapping, 57897677f7fdSAxel Rasmussen pgoff_t idx, 57907677f7fdSAxel Rasmussen unsigned int flags, 57917677f7fdSAxel Rasmussen unsigned long haddr, 5792824ddc60SNadav Amit unsigned long addr, 57937677f7fdSAxel Rasmussen unsigned long reason) 57947677f7fdSAxel Rasmussen { 57957677f7fdSAxel Rasmussen u32 hash; 57967677f7fdSAxel Rasmussen struct vm_fault vmf = { 57977677f7fdSAxel Rasmussen .vma = vma, 57987677f7fdSAxel Rasmussen .address = haddr, 5799824ddc60SNadav Amit .real_address = addr, 58007677f7fdSAxel Rasmussen .flags = flags, 58017677f7fdSAxel Rasmussen 58027677f7fdSAxel Rasmussen /* 58037677f7fdSAxel Rasmussen * Hard to debug if it ends up being 58047677f7fdSAxel Rasmussen * used by a callee that assumes 58057677f7fdSAxel Rasmussen * something about the other 58067677f7fdSAxel Rasmussen * uninitialized fields... same as in 58077677f7fdSAxel Rasmussen * memory.c 58087677f7fdSAxel Rasmussen */ 58097677f7fdSAxel Rasmussen }; 58107677f7fdSAxel Rasmussen 58117677f7fdSAxel Rasmussen /* 5812958f32ceSLiu Shixin * vma_lock and hugetlb_fault_mutex must be dropped before handling 5813958f32ceSLiu Shixin * userfault. Also mmap_lock could be dropped due to handling 5814958f32ceSLiu Shixin * userfault, any vma operation should be careful from here. 58157677f7fdSAxel Rasmussen */ 581640549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 58177677f7fdSAxel Rasmussen hash = hugetlb_fault_mutex_hash(mapping, idx); 58187677f7fdSAxel Rasmussen mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5819958f32ceSLiu Shixin return handle_userfault(&vmf, reason); 58207677f7fdSAxel Rasmussen } 58217677f7fdSAxel Rasmussen 58222ea7ff1eSPeter Xu /* 58232ea7ff1eSPeter Xu * Recheck pte with pgtable lock. Returns true if pte didn't change, or 58242ea7ff1eSPeter Xu * false if pte changed or is changing. 58252ea7ff1eSPeter Xu */ 58262ea7ff1eSPeter Xu static bool hugetlb_pte_stable(struct hstate *h, struct mm_struct *mm, 58272ea7ff1eSPeter Xu pte_t *ptep, pte_t old_pte) 58282ea7ff1eSPeter Xu { 58292ea7ff1eSPeter Xu spinlock_t *ptl; 58302ea7ff1eSPeter Xu bool same; 58312ea7ff1eSPeter Xu 58322ea7ff1eSPeter Xu ptl = huge_pte_lock(h, mm, ptep); 58332ea7ff1eSPeter Xu same = pte_same(huge_ptep_get(ptep), old_pte); 58342ea7ff1eSPeter Xu spin_unlock(ptl); 58352ea7ff1eSPeter Xu 58362ea7ff1eSPeter Xu return same; 58372ea7ff1eSPeter Xu } 58382ea7ff1eSPeter Xu 58392b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 58402b740303SSouptick Joarder struct vm_area_struct *vma, 58418382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 5842c64e912cSPeter Xu unsigned long address, pte_t *ptep, 5843c64e912cSPeter Xu pte_t old_pte, unsigned int flags) 5844ac9b9c66SHugh Dickins { 5845a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 58462b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 5847409eb8c2SHillf Danton int anon_rmap = 0; 58484c887265SAdam Litke unsigned long size; 5849d0ce0e47SSidhartha Kumar struct folio *folio; 58501e8f889bSDavid Gibson pte_t new_pte; 5851cb900f41SKirill A. Shutemov spinlock_t *ptl; 5852285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 5853d0ce0e47SSidhartha Kumar bool new_folio, new_pagecache_folio = false; 5854958f32ceSLiu Shixin u32 hash = hugetlb_fault_mutex_hash(mapping, idx); 58554c887265SAdam Litke 585604f2cbe3SMel Gorman /* 585704f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 585804f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 5859c89357e2SDavid Hildenbrand * COW/unsharing. Warn that such a situation has occurred as it may not 5860c89357e2SDavid Hildenbrand * be obvious. 586104f2cbe3SMel Gorman */ 586204f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 5863910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 586404f2cbe3SMel Gorman current->pid); 5865958f32ceSLiu Shixin goto out; 586604f2cbe3SMel Gorman } 586704f2cbe3SMel Gorman 58684c887265SAdam Litke /* 5869188a3972SMike Kravetz * Use page lock to guard against racing truncation 5870188a3972SMike Kravetz * before we get page_table_lock. 58714c887265SAdam Litke */ 5872d0ce0e47SSidhartha Kumar new_folio = false; 5873d0ce0e47SSidhartha Kumar folio = filemap_lock_folio(mapping, idx); 587466dabbb6SChristoph Hellwig if (IS_ERR(folio)) { 5875188a3972SMike Kravetz size = i_size_read(mapping->host) >> huge_page_shift(h); 5876188a3972SMike Kravetz if (idx >= size) 5877188a3972SMike Kravetz goto out; 58787677f7fdSAxel Rasmussen /* Check for page in userfault range */ 58792ea7ff1eSPeter Xu if (userfaultfd_missing(vma)) { 58802ea7ff1eSPeter Xu /* 58812ea7ff1eSPeter Xu * Since hugetlb_no_page() was examining pte 58822ea7ff1eSPeter Xu * without pgtable lock, we need to re-test under 58832ea7ff1eSPeter Xu * lock because the pte may not be stable and could 58842ea7ff1eSPeter Xu * have changed from under us. Try to detect 58852ea7ff1eSPeter Xu * either changed or during-changing ptes and retry 58862ea7ff1eSPeter Xu * properly when needed. 58872ea7ff1eSPeter Xu * 58882ea7ff1eSPeter Xu * Note that userfaultfd is actually fine with 58892ea7ff1eSPeter Xu * false positives (e.g. caused by pte changed), 58902ea7ff1eSPeter Xu * but not wrong logical events (e.g. caused by 58912ea7ff1eSPeter Xu * reading a pte during changing). The latter can 58922ea7ff1eSPeter Xu * confuse the userspace, so the strictness is very 58932ea7ff1eSPeter Xu * much preferred. E.g., MISSING event should 58942ea7ff1eSPeter Xu * never happen on the page after UFFDIO_COPY has 58952ea7ff1eSPeter Xu * correctly installed the page and returned. 58962ea7ff1eSPeter Xu */ 58972ea7ff1eSPeter Xu if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) { 58982ea7ff1eSPeter Xu ret = 0; 58992ea7ff1eSPeter Xu goto out; 59002ea7ff1eSPeter Xu } 59012ea7ff1eSPeter Xu 59022ea7ff1eSPeter Xu return hugetlb_handle_userfault(vma, mapping, idx, flags, 59032ea7ff1eSPeter Xu haddr, address, 59047677f7fdSAxel Rasmussen VM_UFFD_MISSING); 59052ea7ff1eSPeter Xu } 59061a1aad8aSMike Kravetz 5907d0ce0e47SSidhartha Kumar folio = alloc_hugetlb_folio(vma, haddr, 0); 5908d0ce0e47SSidhartha Kumar if (IS_ERR(folio)) { 59094643d67eSMike Kravetz /* 59104643d67eSMike Kravetz * Returning error will result in faulting task being 59114643d67eSMike Kravetz * sent SIGBUS. The hugetlb fault mutex prevents two 59124643d67eSMike Kravetz * tasks from racing to fault in the same page which 59134643d67eSMike Kravetz * could result in false unable to allocate errors. 59144643d67eSMike Kravetz * Page migration does not take the fault mutex, but 59154643d67eSMike Kravetz * does a clear then write of pte's under page table 59164643d67eSMike Kravetz * lock. Page fault code could race with migration, 59174643d67eSMike Kravetz * notice the clear pte and try to allocate a page 59184643d67eSMike Kravetz * here. Before returning error, get ptl and make 59194643d67eSMike Kravetz * sure there really is no pte entry. 59204643d67eSMike Kravetz */ 5921f9bf6c03SPeter Xu if (hugetlb_pte_stable(h, mm, ptep, old_pte)) 5922d0ce0e47SSidhartha Kumar ret = vmf_error(PTR_ERR(folio)); 5923f9bf6c03SPeter Xu else 5924f9bf6c03SPeter Xu ret = 0; 59256bda666aSChristoph Lameter goto out; 59266bda666aSChristoph Lameter } 5927d0ce0e47SSidhartha Kumar clear_huge_page(&folio->page, address, pages_per_huge_page(h)); 5928d0ce0e47SSidhartha Kumar __folio_mark_uptodate(folio); 5929d0ce0e47SSidhartha Kumar new_folio = true; 5930ac9b9c66SHugh Dickins 5931f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 59329b91c0e2SSidhartha Kumar int err = hugetlb_add_to_page_cache(folio, mapping, idx); 59336bda666aSChristoph Lameter if (err) { 59343a5497a2SMiaohe Lin /* 59353a5497a2SMiaohe Lin * err can't be -EEXIST which implies someone 59363a5497a2SMiaohe Lin * else consumed the reservation since hugetlb 59373a5497a2SMiaohe Lin * fault mutex is held when add a hugetlb page 59383a5497a2SMiaohe Lin * to the page cache. So it's safe to call 59393a5497a2SMiaohe Lin * restore_reserve_on_error() here. 59403a5497a2SMiaohe Lin */ 5941d2d7bb44SSidhartha Kumar restore_reserve_on_error(h, vma, haddr, folio); 5942d0ce0e47SSidhartha Kumar folio_put(folio); 59436bda666aSChristoph Lameter goto out; 59446bda666aSChristoph Lameter } 5945d0ce0e47SSidhartha Kumar new_pagecache_folio = true; 594623be7468SMel Gorman } else { 5947d0ce0e47SSidhartha Kumar folio_lock(folio); 59480fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 59490fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 59500fe6e20bSNaoya Horiguchi goto backout_unlocked; 595123be7468SMel Gorman } 5952409eb8c2SHillf Danton anon_rmap = 1; 59530fe6e20bSNaoya Horiguchi } 59540fe6e20bSNaoya Horiguchi } else { 595557303d80SAndy Whitcroft /* 5956998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 5957998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 5958998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 5959fd6a03edSNaoya Horiguchi */ 5960d0ce0e47SSidhartha Kumar if (unlikely(folio_test_hwpoison(folio))) { 59610eb98f15SMiaohe Lin ret = VM_FAULT_HWPOISON_LARGE | 5962972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 5963fd6a03edSNaoya Horiguchi goto backout_unlocked; 59646bda666aSChristoph Lameter } 59657677f7fdSAxel Rasmussen 59667677f7fdSAxel Rasmussen /* Check for page in userfault range. */ 59677677f7fdSAxel Rasmussen if (userfaultfd_minor(vma)) { 5968d0ce0e47SSidhartha Kumar folio_unlock(folio); 5969d0ce0e47SSidhartha Kumar folio_put(folio); 59702ea7ff1eSPeter Xu /* See comment in userfaultfd_missing() block above */ 59712ea7ff1eSPeter Xu if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) { 59722ea7ff1eSPeter Xu ret = 0; 59732ea7ff1eSPeter Xu goto out; 59742ea7ff1eSPeter Xu } 59752ea7ff1eSPeter Xu return hugetlb_handle_userfault(vma, mapping, idx, flags, 59762ea7ff1eSPeter Xu haddr, address, 59777677f7fdSAxel Rasmussen VM_UFFD_MINOR); 59787677f7fdSAxel Rasmussen } 5979998b4382SNaoya Horiguchi } 59801e8f889bSDavid Gibson 598157303d80SAndy Whitcroft /* 598257303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 598357303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 598457303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 598557303d80SAndy Whitcroft * the spinlock. 598657303d80SAndy Whitcroft */ 59875e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 5988285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 59892b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 59902b26736cSAndy Whitcroft goto backout_unlocked; 59912b26736cSAndy Whitcroft } 59925e911373SMike Kravetz /* Just decrements count, does not deallocate */ 5993285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 59945e911373SMike Kravetz } 599557303d80SAndy Whitcroft 59968bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 599783c54070SNick Piggin ret = 0; 5998c64e912cSPeter Xu /* If pte changed from under us, retry */ 5999c64e912cSPeter Xu if (!pte_same(huge_ptep_get(ptep), old_pte)) 60004c887265SAdam Litke goto backout; 60014c887265SAdam Litke 60024781593dSPeter Xu if (anon_rmap) 6003d0ce0e47SSidhartha Kumar hugepage_add_new_anon_rmap(folio, vma, haddr); 60044781593dSPeter Xu else 6005d0ce0e47SSidhartha Kumar page_dup_file_rmap(&folio->page, true); 6006d0ce0e47SSidhartha Kumar new_pte = make_huge_pte(vma, &folio->page, ((vma->vm_flags & VM_WRITE) 60071e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 6008c64e912cSPeter Xu /* 6009c64e912cSPeter Xu * If this pte was previously wr-protected, keep it wr-protected even 6010c64e912cSPeter Xu * if populated. 6011c64e912cSPeter Xu */ 6012c64e912cSPeter Xu if (unlikely(pte_marker_uffd_wp(old_pte))) 6013f1eb1bacSPeter Xu new_pte = huge_pte_mkuffd_wp(new_pte); 6014935d4f0cSRyan Roberts set_huge_pte_at(mm, haddr, ptep, new_pte, huge_page_size(h)); 60151e8f889bSDavid Gibson 60165d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 6017788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 60181e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 6019371607a3SSidhartha Kumar ret = hugetlb_wp(mm, vma, address, ptep, flags, folio, ptl); 60201e8f889bSDavid Gibson } 60211e8f889bSDavid Gibson 6022cb900f41SKirill A. Shutemov spin_unlock(ptl); 6023cb6acd01SMike Kravetz 6024cb6acd01SMike Kravetz /* 6025d0ce0e47SSidhartha Kumar * Only set hugetlb_migratable in newly allocated pages. Existing pages 6026d0ce0e47SSidhartha Kumar * found in the pagecache may not have hugetlb_migratable if they have 60278f251a3dSMike Kravetz * been isolated for migration. 6028cb6acd01SMike Kravetz */ 6029d0ce0e47SSidhartha Kumar if (new_folio) 6030d0ce0e47SSidhartha Kumar folio_set_hugetlb_migratable(folio); 6031cb6acd01SMike Kravetz 6032d0ce0e47SSidhartha Kumar folio_unlock(folio); 60334c887265SAdam Litke out: 6034958f32ceSLiu Shixin hugetlb_vma_unlock_read(vma); 6035958f32ceSLiu Shixin mutex_unlock(&hugetlb_fault_mutex_table[hash]); 6036ac9b9c66SHugh Dickins return ret; 60374c887265SAdam Litke 60384c887265SAdam Litke backout: 6039cb900f41SKirill A. Shutemov spin_unlock(ptl); 60402b26736cSAndy Whitcroft backout_unlocked: 6041d0ce0e47SSidhartha Kumar if (new_folio && !new_pagecache_folio) 6042d2d7bb44SSidhartha Kumar restore_reserve_on_error(h, vma, haddr, folio); 6043fa27759aSMike Kravetz 6044d0ce0e47SSidhartha Kumar folio_unlock(folio); 6045d0ce0e47SSidhartha Kumar folio_put(folio); 60464c887265SAdam Litke goto out; 6047ac9b9c66SHugh Dickins } 6048ac9b9c66SHugh Dickins 60498382d914SDavidlohr Bueso #ifdef CONFIG_SMP 6050188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 60518382d914SDavidlohr Bueso { 60528382d914SDavidlohr Bueso unsigned long key[2]; 60538382d914SDavidlohr Bueso u32 hash; 60548382d914SDavidlohr Bueso 60558382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 60568382d914SDavidlohr Bueso key[1] = idx; 60578382d914SDavidlohr Bueso 605855254636SMike Kravetz hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); 60598382d914SDavidlohr Bueso 60608382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 60618382d914SDavidlohr Bueso } 60628382d914SDavidlohr Bueso #else 60638382d914SDavidlohr Bueso /* 60646c26d310SMiaohe Lin * For uniprocessor systems we always use a single mutex, so just 60658382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 60668382d914SDavidlohr Bueso */ 6067188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 60688382d914SDavidlohr Bueso { 60698382d914SDavidlohr Bueso return 0; 60708382d914SDavidlohr Bueso } 60718382d914SDavidlohr Bueso #endif 60728382d914SDavidlohr Bueso 60732b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 6074788c7df4SHugh Dickins unsigned long address, unsigned int flags) 607586e5216fSAdam Litke { 60768382d914SDavidlohr Bueso pte_t *ptep, entry; 6077cb900f41SKirill A. Shutemov spinlock_t *ptl; 60782b740303SSouptick Joarder vm_fault_t ret; 60798382d914SDavidlohr Bueso u32 hash; 60808382d914SDavidlohr Bueso pgoff_t idx; 6081061e62e8SZhangPeng struct folio *folio = NULL; 6082371607a3SSidhartha Kumar struct folio *pagecache_folio = NULL; 6083a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 60848382d914SDavidlohr Bueso struct address_space *mapping; 60850f792cf9SNaoya Horiguchi int need_wait_lock = 0; 6086285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 608786e5216fSAdam Litke 60884ec31152SMatthew Wilcox (Oracle) /* TODO: Handle faults under the VMA lock */ 60894ec31152SMatthew Wilcox (Oracle) if (flags & FAULT_FLAG_VMA_LOCK) { 60904ec31152SMatthew Wilcox (Oracle) vma_end_read(vma); 60914ec31152SMatthew Wilcox (Oracle) return VM_FAULT_RETRY; 60924ec31152SMatthew Wilcox (Oracle) } 60934ec31152SMatthew Wilcox (Oracle) 60943935baa9SDavid Gibson /* 60953935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 60963935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 60973935baa9SDavid Gibson * the same page in the page cache. 60983935baa9SDavid Gibson */ 609940549ba8SMike Kravetz mapping = vma->vm_file->f_mapping; 610040549ba8SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 6101188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 6102c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 61038382d914SDavidlohr Bueso 610440549ba8SMike Kravetz /* 610540549ba8SMike Kravetz * Acquire vma lock before calling huge_pte_alloc and hold 610640549ba8SMike Kravetz * until finished with ptep. This prevents huge_pmd_unshare from 610740549ba8SMike Kravetz * being called elsewhere and making the ptep no longer valid. 610840549ba8SMike Kravetz */ 610940549ba8SMike Kravetz hugetlb_vma_lock_read(vma); 611040549ba8SMike Kravetz ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); 611140549ba8SMike Kravetz if (!ptep) { 611240549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 611340549ba8SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 611440549ba8SMike Kravetz return VM_FAULT_OOM; 611540549ba8SMike Kravetz } 611640549ba8SMike Kravetz 61177f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 6118af19487fSAxel Rasmussen if (huge_pte_none_mostly(entry)) { 6119af19487fSAxel Rasmussen if (is_pte_marker(entry)) { 6120af19487fSAxel Rasmussen pte_marker marker = 6121af19487fSAxel Rasmussen pte_marker_get(pte_to_swp_entry(entry)); 6122af19487fSAxel Rasmussen 6123af19487fSAxel Rasmussen if (marker & PTE_MARKER_POISONED) { 6124af19487fSAxel Rasmussen ret = VM_FAULT_HWPOISON_LARGE; 6125af19487fSAxel Rasmussen goto out_mutex; 6126af19487fSAxel Rasmussen } 6127af19487fSAxel Rasmussen } 6128af19487fSAxel Rasmussen 6129958f32ceSLiu Shixin /* 6130af19487fSAxel Rasmussen * Other PTE markers should be handled the same way as none PTE. 6131af19487fSAxel Rasmussen * 6132958f32ceSLiu Shixin * hugetlb_no_page will drop vma lock and hugetlb fault 6133958f32ceSLiu Shixin * mutex internally, which make us return immediately. 6134958f32ceSLiu Shixin */ 6135958f32ceSLiu Shixin return hugetlb_no_page(mm, vma, mapping, idx, address, ptep, 6136c64e912cSPeter Xu entry, flags); 6137af19487fSAxel Rasmussen } 613886e5216fSAdam Litke 613983c54070SNick Piggin ret = 0; 61401e8f889bSDavid Gibson 614157303d80SAndy Whitcroft /* 61420f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 61430f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 61447c8de358SEthon Paul * an active hugepage in pagecache. This goto expects the 2nd page 61457c8de358SEthon Paul * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will 61467c8de358SEthon Paul * properly handle it. 61470f792cf9SNaoya Horiguchi */ 6148fcd48540SPeter Xu if (!pte_present(entry)) { 6149fcd48540SPeter Xu if (unlikely(is_hugetlb_entry_migration(entry))) { 6150fcd48540SPeter Xu /* 6151fcd48540SPeter Xu * Release the hugetlb fault lock now, but retain 6152fcd48540SPeter Xu * the vma lock, because it is needed to guard the 6153fcd48540SPeter Xu * huge_pte_lockptr() later in 6154fcd48540SPeter Xu * migration_entry_wait_huge(). The vma lock will 6155fcd48540SPeter Xu * be released there. 6156fcd48540SPeter Xu */ 6157fcd48540SPeter Xu mutex_unlock(&hugetlb_fault_mutex_table[hash]); 6158fcd48540SPeter Xu migration_entry_wait_huge(vma, ptep); 6159fcd48540SPeter Xu return 0; 6160fcd48540SPeter Xu } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 6161fcd48540SPeter Xu ret = VM_FAULT_HWPOISON_LARGE | 6162fcd48540SPeter Xu VM_FAULT_SET_HINDEX(hstate_index(h)); 61630f792cf9SNaoya Horiguchi goto out_mutex; 6164fcd48540SPeter Xu } 61650f792cf9SNaoya Horiguchi 61660f792cf9SNaoya Horiguchi /* 6167c89357e2SDavid Hildenbrand * If we are going to COW/unshare the mapping later, we examine the 6168c89357e2SDavid Hildenbrand * pending reservations for this page now. This will ensure that any 616957303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 61701d8d1464SDavid Hildenbrand * spinlock. Also lookup the pagecache page now as it is used to 61711d8d1464SDavid Hildenbrand * determine if a reservation has been consumed. 617257303d80SAndy Whitcroft */ 6173c89357e2SDavid Hildenbrand if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) && 61741d8d1464SDavid Hildenbrand !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(entry)) { 6175285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 61762b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 6177b4d1d99fSDavid Gibson goto out_mutex; 61782b26736cSAndy Whitcroft } 61795e911373SMike Kravetz /* Just decrements count, does not deallocate */ 6180285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 618157303d80SAndy Whitcroft 6182371607a3SSidhartha Kumar pagecache_folio = filemap_lock_folio(mapping, idx); 618366dabbb6SChristoph Hellwig if (IS_ERR(pagecache_folio)) 618466dabbb6SChristoph Hellwig pagecache_folio = NULL; 618557303d80SAndy Whitcroft } 618657303d80SAndy Whitcroft 61870f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 61880fe6e20bSNaoya Horiguchi 6189c89357e2SDavid Hildenbrand /* Check for a racing update before calling hugetlb_wp() */ 6190b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 6191cb900f41SKirill A. Shutemov goto out_ptl; 6192b4d1d99fSDavid Gibson 6193166f3eccSPeter Xu /* Handle userfault-wp first, before trying to lock more pages */ 6194166f3eccSPeter Xu if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) && 6195166f3eccSPeter Xu (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 6196166f3eccSPeter Xu struct vm_fault vmf = { 6197166f3eccSPeter Xu .vma = vma, 6198166f3eccSPeter Xu .address = haddr, 6199166f3eccSPeter Xu .real_address = address, 6200166f3eccSPeter Xu .flags = flags, 6201166f3eccSPeter Xu }; 6202166f3eccSPeter Xu 6203166f3eccSPeter Xu spin_unlock(ptl); 6204371607a3SSidhartha Kumar if (pagecache_folio) { 6205371607a3SSidhartha Kumar folio_unlock(pagecache_folio); 6206371607a3SSidhartha Kumar folio_put(pagecache_folio); 6207166f3eccSPeter Xu } 620840549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 6209166f3eccSPeter Xu mutex_unlock(&hugetlb_fault_mutex_table[hash]); 6210166f3eccSPeter Xu return handle_userfault(&vmf, VM_UFFD_WP); 6211166f3eccSPeter Xu } 6212166f3eccSPeter Xu 62130f792cf9SNaoya Horiguchi /* 6214c89357e2SDavid Hildenbrand * hugetlb_wp() requires page locks of pte_page(entry) and 6215371607a3SSidhartha Kumar * pagecache_folio, so here we need take the former one 6216061e62e8SZhangPeng * when folio != pagecache_folio or !pagecache_folio. 62170f792cf9SNaoya Horiguchi */ 6218061e62e8SZhangPeng folio = page_folio(pte_page(entry)); 6219061e62e8SZhangPeng if (folio != pagecache_folio) 6220061e62e8SZhangPeng if (!folio_trylock(folio)) { 62210f792cf9SNaoya Horiguchi need_wait_lock = 1; 62220f792cf9SNaoya Horiguchi goto out_ptl; 62230f792cf9SNaoya Horiguchi } 62240f792cf9SNaoya Horiguchi 6225061e62e8SZhangPeng folio_get(folio); 6226b4d1d99fSDavid Gibson 6227c89357e2SDavid Hildenbrand if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) { 6228106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 6229c89357e2SDavid Hildenbrand ret = hugetlb_wp(mm, vma, address, ptep, flags, 6230371607a3SSidhartha Kumar pagecache_folio, ptl); 62310f792cf9SNaoya Horiguchi goto out_put_page; 6232c89357e2SDavid Hildenbrand } else if (likely(flags & FAULT_FLAG_WRITE)) { 6233106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 6234b4d1d99fSDavid Gibson } 6235c89357e2SDavid Hildenbrand } 6236b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 6237285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 6238788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 6239285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 62400f792cf9SNaoya Horiguchi out_put_page: 6241061e62e8SZhangPeng if (folio != pagecache_folio) 6242061e62e8SZhangPeng folio_unlock(folio); 6243061e62e8SZhangPeng folio_put(folio); 6244cb900f41SKirill A. Shutemov out_ptl: 6245cb900f41SKirill A. Shutemov spin_unlock(ptl); 624657303d80SAndy Whitcroft 6247371607a3SSidhartha Kumar if (pagecache_folio) { 6248371607a3SSidhartha Kumar folio_unlock(pagecache_folio); 6249371607a3SSidhartha Kumar folio_put(pagecache_folio); 625057303d80SAndy Whitcroft } 6251b4d1d99fSDavid Gibson out_mutex: 625240549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 6253c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 62540f792cf9SNaoya Horiguchi /* 62550f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 62560f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 62570f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 62580f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 62590f792cf9SNaoya Horiguchi * here without taking refcount. 62600f792cf9SNaoya Horiguchi */ 62610f792cf9SNaoya Horiguchi if (need_wait_lock) 6262061e62e8SZhangPeng folio_wait_locked(folio); 62631e8f889bSDavid Gibson return ret; 626486e5216fSAdam Litke } 626586e5216fSAdam Litke 6266714c1891SAxel Rasmussen #ifdef CONFIG_USERFAULTFD 62678fb5debcSMike Kravetz /* 6268a734991cSAxel Rasmussen * Used by userfaultfd UFFDIO_* ioctls. Based on userfaultfd's mfill_atomic_pte 6269a734991cSAxel Rasmussen * with modifications for hugetlb pages. 62708fb5debcSMike Kravetz */ 627161c50040SAxel Rasmussen int hugetlb_mfill_atomic_pte(pte_t *dst_pte, 62728fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 62738fb5debcSMike Kravetz unsigned long dst_addr, 62748fb5debcSMike Kravetz unsigned long src_addr, 6275d9712937SAxel Rasmussen uffd_flags_t flags, 62760169fd51SZhangPeng struct folio **foliop) 62778fb5debcSMike Kravetz { 627861c50040SAxel Rasmussen struct mm_struct *dst_mm = dst_vma->vm_mm; 6279d9712937SAxel Rasmussen bool is_continue = uffd_flags_mode_is(flags, MFILL_ATOMIC_CONTINUE); 6280d9712937SAxel Rasmussen bool wp_enabled = (flags & MFILL_ATOMIC_WP); 62818cc5fcbbSMina Almasry struct hstate *h = hstate_vma(dst_vma); 62828cc5fcbbSMina Almasry struct address_space *mapping = dst_vma->vm_file->f_mapping; 62838cc5fcbbSMina Almasry pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr); 62841e392147SAndrea Arcangeli unsigned long size; 62851c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 62868fb5debcSMike Kravetz pte_t _dst_pte; 62878fb5debcSMike Kravetz spinlock_t *ptl; 62888cc5fcbbSMina Almasry int ret = -ENOMEM; 6289d0ce0e47SSidhartha Kumar struct folio *folio; 6290f6191471SAxel Rasmussen int writable; 6291d0ce0e47SSidhartha Kumar bool folio_in_pagecache = false; 62928fb5debcSMike Kravetz 62938a13897fSAxel Rasmussen if (uffd_flags_mode_is(flags, MFILL_ATOMIC_POISON)) { 62948a13897fSAxel Rasmussen ptl = huge_pte_lock(h, dst_mm, dst_pte); 62958a13897fSAxel Rasmussen 62968a13897fSAxel Rasmussen /* Don't overwrite any existing PTEs (even markers) */ 62978a13897fSAxel Rasmussen if (!huge_pte_none(huge_ptep_get(dst_pte))) { 62988a13897fSAxel Rasmussen spin_unlock(ptl); 62998a13897fSAxel Rasmussen return -EEXIST; 63008a13897fSAxel Rasmussen } 63018a13897fSAxel Rasmussen 63028a13897fSAxel Rasmussen _dst_pte = make_pte_marker(PTE_MARKER_POISONED); 6303935d4f0cSRyan Roberts set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte, 6304935d4f0cSRyan Roberts huge_page_size(h)); 63058a13897fSAxel Rasmussen 63068a13897fSAxel Rasmussen /* No need to invalidate - it was non-present before */ 63078a13897fSAxel Rasmussen update_mmu_cache(dst_vma, dst_addr, dst_pte); 63088a13897fSAxel Rasmussen 63098a13897fSAxel Rasmussen spin_unlock(ptl); 63108a13897fSAxel Rasmussen return 0; 63118a13897fSAxel Rasmussen } 63128a13897fSAxel Rasmussen 6313f6191471SAxel Rasmussen if (is_continue) { 6314f6191471SAxel Rasmussen ret = -EFAULT; 6315d0ce0e47SSidhartha Kumar folio = filemap_lock_folio(mapping, idx); 631666dabbb6SChristoph Hellwig if (IS_ERR(folio)) 6317f6191471SAxel Rasmussen goto out; 6318d0ce0e47SSidhartha Kumar folio_in_pagecache = true; 63190169fd51SZhangPeng } else if (!*foliop) { 63200169fd51SZhangPeng /* If a folio already exists, then it's UFFDIO_COPY for 6321d84cf06eSMina Almasry * a non-missing case. Return -EEXIST. 6322d84cf06eSMina Almasry */ 6323d84cf06eSMina Almasry if (vm_shared && 6324d84cf06eSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 6325d84cf06eSMina Almasry ret = -EEXIST; 63268fb5debcSMike Kravetz goto out; 6327d84cf06eSMina Almasry } 6328d84cf06eSMina Almasry 6329d0ce0e47SSidhartha Kumar folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0); 6330d0ce0e47SSidhartha Kumar if (IS_ERR(folio)) { 6331d84cf06eSMina Almasry ret = -ENOMEM; 6332d84cf06eSMina Almasry goto out; 6333d84cf06eSMina Almasry } 63348fb5debcSMike Kravetz 6335e87340caSZhangPeng ret = copy_folio_from_user(folio, (const void __user *) src_addr, 6336e87340caSZhangPeng false); 63378fb5debcSMike Kravetz 6338c1e8d7c6SMichel Lespinasse /* fallback to copy_from_user outside mmap_lock */ 63398fb5debcSMike Kravetz if (unlikely(ret)) { 63409e368259SAndrea Arcangeli ret = -ENOENT; 6341d0ce0e47SSidhartha Kumar /* Free the allocated folio which may have 63428cc5fcbbSMina Almasry * consumed a reservation. 63438cc5fcbbSMina Almasry */ 6344d2d7bb44SSidhartha Kumar restore_reserve_on_error(h, dst_vma, dst_addr, folio); 6345d0ce0e47SSidhartha Kumar folio_put(folio); 63468cc5fcbbSMina Almasry 6347d0ce0e47SSidhartha Kumar /* Allocate a temporary folio to hold the copied 63488cc5fcbbSMina Almasry * contents. 63498cc5fcbbSMina Almasry */ 6350d0ce0e47SSidhartha Kumar folio = alloc_hugetlb_folio_vma(h, dst_vma, dst_addr); 6351d0ce0e47SSidhartha Kumar if (!folio) { 63528cc5fcbbSMina Almasry ret = -ENOMEM; 63538cc5fcbbSMina Almasry goto out; 63548cc5fcbbSMina Almasry } 63550169fd51SZhangPeng *foliop = folio; 63560169fd51SZhangPeng /* Set the outparam foliop and return to the caller to 63578cc5fcbbSMina Almasry * copy the contents outside the lock. Don't free the 63580169fd51SZhangPeng * folio. 63598cc5fcbbSMina Almasry */ 63608fb5debcSMike Kravetz goto out; 63618fb5debcSMike Kravetz } 63628fb5debcSMike Kravetz } else { 63638cc5fcbbSMina Almasry if (vm_shared && 63648cc5fcbbSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 63650169fd51SZhangPeng folio_put(*foliop); 63668cc5fcbbSMina Almasry ret = -EEXIST; 63670169fd51SZhangPeng *foliop = NULL; 63688cc5fcbbSMina Almasry goto out; 63698cc5fcbbSMina Almasry } 63708cc5fcbbSMina Almasry 6371d0ce0e47SSidhartha Kumar folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0); 6372d0ce0e47SSidhartha Kumar if (IS_ERR(folio)) { 63730169fd51SZhangPeng folio_put(*foliop); 63748cc5fcbbSMina Almasry ret = -ENOMEM; 63750169fd51SZhangPeng *foliop = NULL; 63768cc5fcbbSMina Almasry goto out; 63778cc5fcbbSMina Almasry } 63781cb9dc4bSLiu Shixin ret = copy_user_large_folio(folio, *foliop, dst_addr, dst_vma); 63790169fd51SZhangPeng folio_put(*foliop); 63800169fd51SZhangPeng *foliop = NULL; 63811cb9dc4bSLiu Shixin if (ret) { 63821cb9dc4bSLiu Shixin folio_put(folio); 63831cb9dc4bSLiu Shixin goto out; 63841cb9dc4bSLiu Shixin } 63858fb5debcSMike Kravetz } 63868fb5debcSMike Kravetz 63878fb5debcSMike Kravetz /* 6388d0ce0e47SSidhartha Kumar * The memory barrier inside __folio_mark_uptodate makes sure that 63898fb5debcSMike Kravetz * preceding stores to the page contents become visible before 63908fb5debcSMike Kravetz * the set_pte_at() write. 63918fb5debcSMike Kravetz */ 6392d0ce0e47SSidhartha Kumar __folio_mark_uptodate(folio); 63938fb5debcSMike Kravetz 6394f6191471SAxel Rasmussen /* Add shared, newly allocated pages to the page cache. */ 6395f6191471SAxel Rasmussen if (vm_shared && !is_continue) { 63961e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 63971e392147SAndrea Arcangeli ret = -EFAULT; 63981e392147SAndrea Arcangeli if (idx >= size) 63991e392147SAndrea Arcangeli goto out_release_nounlock; 64001c9e8defSMike Kravetz 64011e392147SAndrea Arcangeli /* 64021e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 64037e1813d4SMike Kravetz * hugetlb_add_to_page_cache() below happens through the 64041e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 64051e392147SAndrea Arcangeli * the caller. 64061e392147SAndrea Arcangeli */ 64079b91c0e2SSidhartha Kumar ret = hugetlb_add_to_page_cache(folio, mapping, idx); 64081c9e8defSMike Kravetz if (ret) 64091c9e8defSMike Kravetz goto out_release_nounlock; 6410d0ce0e47SSidhartha Kumar folio_in_pagecache = true; 64111c9e8defSMike Kravetz } 64121c9e8defSMike Kravetz 6413bcc66543SMiaohe Lin ptl = huge_pte_lock(h, dst_mm, dst_pte); 64148fb5debcSMike Kravetz 64158625147cSJames Houghton ret = -EIO; 6416d0ce0e47SSidhartha Kumar if (folio_test_hwpoison(folio)) 64178625147cSJames Houghton goto out_release_unlock; 64188625147cSJames Houghton 64191e392147SAndrea Arcangeli /* 64206041c691SPeter Xu * We allow to overwrite a pte marker: consider when both MISSING|WP 64216041c691SPeter Xu * registered, we firstly wr-protect a none pte which has no page cache 64226041c691SPeter Xu * page backing it, then access the page. 64236041c691SPeter Xu */ 6424fa27759aSMike Kravetz ret = -EEXIST; 64256041c691SPeter Xu if (!huge_pte_none_mostly(huge_ptep_get(dst_pte))) 64268fb5debcSMike Kravetz goto out_release_unlock; 64278fb5debcSMike Kravetz 6428d0ce0e47SSidhartha Kumar if (folio_in_pagecache) 6429d0ce0e47SSidhartha Kumar page_dup_file_rmap(&folio->page, true); 64304781593dSPeter Xu else 6431d0ce0e47SSidhartha Kumar hugepage_add_new_anon_rmap(folio, dst_vma, dst_addr); 64328fb5debcSMike Kravetz 64336041c691SPeter Xu /* 64346041c691SPeter Xu * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY 64356041c691SPeter Xu * with wp flag set, don't set pte write bit. 64366041c691SPeter Xu */ 6437d9712937SAxel Rasmussen if (wp_enabled || (is_continue && !vm_shared)) 6438f6191471SAxel Rasmussen writable = 0; 6439f6191471SAxel Rasmussen else 6440f6191471SAxel Rasmussen writable = dst_vma->vm_flags & VM_WRITE; 6441f6191471SAxel Rasmussen 6442d0ce0e47SSidhartha Kumar _dst_pte = make_huge_pte(dst_vma, &folio->page, writable); 64436041c691SPeter Xu /* 64446041c691SPeter Xu * Always mark UFFDIO_COPY page dirty; note that this may not be 64456041c691SPeter Xu * extremely important for hugetlbfs for now since swapping is not 64466041c691SPeter Xu * supported, but we should still be clear in that this page cannot be 64476041c691SPeter Xu * thrown away at will, even if write bit not set. 64486041c691SPeter Xu */ 64498fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 64508fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 64518fb5debcSMike Kravetz 6452d9712937SAxel Rasmussen if (wp_enabled) 64536041c691SPeter Xu _dst_pte = huge_pte_mkuffd_wp(_dst_pte); 64546041c691SPeter Xu 6455935d4f0cSRyan Roberts set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte, huge_page_size(h)); 64568fb5debcSMike Kravetz 64578fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 64588fb5debcSMike Kravetz 64598fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 64608fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 64618fb5debcSMike Kravetz 64628fb5debcSMike Kravetz spin_unlock(ptl); 6463f6191471SAxel Rasmussen if (!is_continue) 6464d0ce0e47SSidhartha Kumar folio_set_hugetlb_migratable(folio); 6465f6191471SAxel Rasmussen if (vm_shared || is_continue) 6466d0ce0e47SSidhartha Kumar folio_unlock(folio); 64678fb5debcSMike Kravetz ret = 0; 64688fb5debcSMike Kravetz out: 64698fb5debcSMike Kravetz return ret; 64708fb5debcSMike Kravetz out_release_unlock: 64718fb5debcSMike Kravetz spin_unlock(ptl); 6472f6191471SAxel Rasmussen if (vm_shared || is_continue) 6473d0ce0e47SSidhartha Kumar folio_unlock(folio); 64745af10dfdSAndrea Arcangeli out_release_nounlock: 6475d0ce0e47SSidhartha Kumar if (!folio_in_pagecache) 6476d2d7bb44SSidhartha Kumar restore_reserve_on_error(h, dst_vma, dst_addr, folio); 6477d0ce0e47SSidhartha Kumar folio_put(folio); 64788fb5debcSMike Kravetz goto out; 64798fb5debcSMike Kravetz } 6480714c1891SAxel Rasmussen #endif /* CONFIG_USERFAULTFD */ 64818fb5debcSMike Kravetz 648257a196a5SMike Kravetz struct page *hugetlb_follow_page_mask(struct vm_area_struct *vma, 64835502ea44SPeter Xu unsigned long address, unsigned int flags, 64845502ea44SPeter Xu unsigned int *page_mask) 648557a196a5SMike Kravetz { 648657a196a5SMike Kravetz struct hstate *h = hstate_vma(vma); 648757a196a5SMike Kravetz struct mm_struct *mm = vma->vm_mm; 648857a196a5SMike Kravetz unsigned long haddr = address & huge_page_mask(h); 648957a196a5SMike Kravetz struct page *page = NULL; 649057a196a5SMike Kravetz spinlock_t *ptl; 649157a196a5SMike Kravetz pte_t *pte, entry; 6492458568c9SPeter Xu int ret; 649357a196a5SMike Kravetz 64947d049f3aSPeter Xu hugetlb_vma_lock_read(vma); 64959c67a207SPeter Xu pte = hugetlb_walk(vma, haddr, huge_page_size(h)); 649657a196a5SMike Kravetz if (!pte) 64977d049f3aSPeter Xu goto out_unlock; 649857a196a5SMike Kravetz 649957a196a5SMike Kravetz ptl = huge_pte_lock(h, mm, pte); 650057a196a5SMike Kravetz entry = huge_ptep_get(pte); 650157a196a5SMike Kravetz if (pte_present(entry)) { 6502458568c9SPeter Xu page = pte_page(entry); 6503458568c9SPeter Xu 6504458568c9SPeter Xu if (!huge_pte_write(entry)) { 6505458568c9SPeter Xu if (flags & FOLL_WRITE) { 6506458568c9SPeter Xu page = NULL; 6507458568c9SPeter Xu goto out; 6508458568c9SPeter Xu } 6509458568c9SPeter Xu 6510458568c9SPeter Xu if (gup_must_unshare(vma, flags, page)) { 6511458568c9SPeter Xu /* Tell the caller to do unsharing */ 6512458568c9SPeter Xu page = ERR_PTR(-EMLINK); 6513458568c9SPeter Xu goto out; 6514458568c9SPeter Xu } 6515458568c9SPeter Xu } 6516458568c9SPeter Xu 651718576d12SZi Yan page = nth_page(page, ((address & ~huge_page_mask(h)) >> PAGE_SHIFT)); 6518458568c9SPeter Xu 651957a196a5SMike Kravetz /* 652057a196a5SMike Kravetz * Note that page may be a sub-page, and with vmemmap 652157a196a5SMike Kravetz * optimizations the page struct may be read only. 652257a196a5SMike Kravetz * try_grab_page() will increase the ref count on the 652357a196a5SMike Kravetz * head page, so this will be OK. 652457a196a5SMike Kravetz * 6525e2ca6ba6SLinus Torvalds * try_grab_page() should always be able to get the page here, 6526e2ca6ba6SLinus Torvalds * because we hold the ptl lock and have verified pte_present(). 652757a196a5SMike Kravetz */ 6528458568c9SPeter Xu ret = try_grab_page(page, flags); 6529458568c9SPeter Xu 6530458568c9SPeter Xu if (WARN_ON_ONCE(ret)) { 6531458568c9SPeter Xu page = ERR_PTR(ret); 653257a196a5SMike Kravetz goto out; 653357a196a5SMike Kravetz } 65345502ea44SPeter Xu 65355502ea44SPeter Xu *page_mask = (1U << huge_page_order(h)) - 1; 653657a196a5SMike Kravetz } 653757a196a5SMike Kravetz out: 653857a196a5SMike Kravetz spin_unlock(ptl); 65397d049f3aSPeter Xu out_unlock: 65407d049f3aSPeter Xu hugetlb_vma_unlock_read(vma); 6541dd767aaaSPeter Xu 6542dd767aaaSPeter Xu /* 6543dd767aaaSPeter Xu * Fixup retval for dump requests: if pagecache doesn't exist, 6544dd767aaaSPeter Xu * don't try to allocate a new page but just skip it. 6545dd767aaaSPeter Xu */ 6546dd767aaaSPeter Xu if (!page && (flags & FOLL_DUMP) && 6547dd767aaaSPeter Xu !hugetlbfs_pagecache_present(h, vma, address)) 6548dd767aaaSPeter Xu page = ERR_PTR(-EFAULT); 6549dd767aaaSPeter Xu 655057a196a5SMike Kravetz return page; 655157a196a5SMike Kravetz } 655257a196a5SMike Kravetz 6553a79390f5SPeter Xu long hugetlb_change_protection(struct vm_area_struct *vma, 65545a90d5a1SPeter Xu unsigned long address, unsigned long end, 65555a90d5a1SPeter Xu pgprot_t newprot, unsigned long cp_flags) 65568f860591SZhang, Yanmin { 65578f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 65588f860591SZhang, Yanmin unsigned long start = address; 65598f860591SZhang, Yanmin pte_t *ptep; 65608f860591SZhang, Yanmin pte_t pte; 6561a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 6562a79390f5SPeter Xu long pages = 0, psize = huge_page_size(h); 6563dff11abeSMike Kravetz bool shared_pmd = false; 6564ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 6565e95a9851SMike Kravetz unsigned long last_addr_mask; 65665a90d5a1SPeter Xu bool uffd_wp = cp_flags & MM_CP_UFFD_WP; 65675a90d5a1SPeter Xu bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE; 6568dff11abeSMike Kravetz 6569dff11abeSMike Kravetz /* 6570dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 6571ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 6572dff11abeSMike Kravetz * range if PMD sharing is possible. 6573dff11abeSMike Kravetz */ 65747269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 65757d4a8be0SAlistair Popple 0, mm, start, end); 6576ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 65778f860591SZhang, Yanmin 65788f860591SZhang, Yanmin BUG_ON(address >= end); 6579ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 65808f860591SZhang, Yanmin 6581ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 658240549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 658383cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 658440549ba8SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 658560dfaad6SPeter Xu for (; address < end; address += psize) { 6586cb900f41SKirill A. Shutemov spinlock_t *ptl; 65879c67a207SPeter Xu ptep = hugetlb_walk(vma, address, psize); 6588e95a9851SMike Kravetz if (!ptep) { 6589fed15f13SPeter Xu if (!uffd_wp) { 6590e95a9851SMike Kravetz address |= last_addr_mask; 65918f860591SZhang, Yanmin continue; 6592e95a9851SMike Kravetz } 6593fed15f13SPeter Xu /* 6594fed15f13SPeter Xu * Userfaultfd wr-protect requires pgtable 6595fed15f13SPeter Xu * pre-allocations to install pte markers. 6596fed15f13SPeter Xu */ 6597fed15f13SPeter Xu ptep = huge_pte_alloc(mm, vma, address, psize); 6598d1751118SPeter Xu if (!ptep) { 6599d1751118SPeter Xu pages = -ENOMEM; 6600fed15f13SPeter Xu break; 6601fed15f13SPeter Xu } 6602d1751118SPeter Xu } 6603cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 66044ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, address, ptep)) { 660560dfaad6SPeter Xu /* 660660dfaad6SPeter Xu * When uffd-wp is enabled on the vma, unshare 660760dfaad6SPeter Xu * shouldn't happen at all. Warn about it if it 660860dfaad6SPeter Xu * happened due to some reason. 660960dfaad6SPeter Xu */ 661060dfaad6SPeter Xu WARN_ON_ONCE(uffd_wp || uffd_wp_resolve); 66117da4d641SPeter Zijlstra pages++; 6612cb900f41SKirill A. Shutemov spin_unlock(ptl); 6613dff11abeSMike Kravetz shared_pmd = true; 66144ddb4d91SMike Kravetz address |= last_addr_mask; 661539dde65cSChen, Kenneth W continue; 66167da4d641SPeter Zijlstra } 6617a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 6618a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 66190e678153SDavid Hildenbrand /* Nothing to do. */ 66200e678153SDavid Hildenbrand } else if (unlikely(is_hugetlb_entry_migration(pte))) { 6621a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 66226c287605SDavid Hildenbrand struct page *page = pfn_swap_entry_to_page(entry); 662344f86392SDavid Hildenbrand pte_t newpte = pte; 6624a8bda28dSNaoya Horiguchi 662544f86392SDavid Hildenbrand if (is_writable_migration_entry(entry)) { 66266c287605SDavid Hildenbrand if (PageAnon(page)) 66276c287605SDavid Hildenbrand entry = make_readable_exclusive_migration_entry( 66286c287605SDavid Hildenbrand swp_offset(entry)); 66296c287605SDavid Hildenbrand else 66304dd845b5SAlistair Popple entry = make_readable_migration_entry( 66314dd845b5SAlistair Popple swp_offset(entry)); 6632a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 663344f86392SDavid Hildenbrand pages++; 663444f86392SDavid Hildenbrand } 663544f86392SDavid Hildenbrand 66365a90d5a1SPeter Xu if (uffd_wp) 66375a90d5a1SPeter Xu newpte = pte_swp_mkuffd_wp(newpte); 66385a90d5a1SPeter Xu else if (uffd_wp_resolve) 66395a90d5a1SPeter Xu newpte = pte_swp_clear_uffd_wp(newpte); 664044f86392SDavid Hildenbrand if (!pte_same(pte, newpte)) 6641935d4f0cSRyan Roberts set_huge_pte_at(mm, address, ptep, newpte, psize); 66420e678153SDavid Hildenbrand } else if (unlikely(is_pte_marker(pte))) { 6643db01bfbdSPeter Xu /* 6644db01bfbdSPeter Xu * Do nothing on a poison marker; page is 6645db01bfbdSPeter Xu * corrupted, permissons do not apply. Here 6646db01bfbdSPeter Xu * pte_marker_uffd_wp()==true implies !poison 6647db01bfbdSPeter Xu * because they're mutual exclusive. 6648db01bfbdSPeter Xu */ 6649db01bfbdSPeter Xu if (pte_marker_uffd_wp(pte) && uffd_wp_resolve) 66500e678153SDavid Hildenbrand /* Safe to modify directly (non-present->none). */ 665160dfaad6SPeter Xu huge_pte_clear(mm, address, ptep, psize); 66520e678153SDavid Hildenbrand } else if (!huge_pte_none(pte)) { 6653023bdd00SAneesh Kumar K.V pte_t old_pte; 665479c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 6655023bdd00SAneesh Kumar K.V 6656023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 665716785bd7SAnshuman Khandual pte = huge_pte_modify(old_pte, newprot); 665879c1c594SChristophe Leroy pte = arch_make_huge_pte(pte, shift, vma->vm_flags); 66595a90d5a1SPeter Xu if (uffd_wp) 6660f1eb1bacSPeter Xu pte = huge_pte_mkuffd_wp(pte); 66615a90d5a1SPeter Xu else if (uffd_wp_resolve) 66625a90d5a1SPeter Xu pte = huge_pte_clear_uffd_wp(pte); 6663023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 66647da4d641SPeter Zijlstra pages++; 666560dfaad6SPeter Xu } else { 666660dfaad6SPeter Xu /* None pte */ 666760dfaad6SPeter Xu if (unlikely(uffd_wp)) 666860dfaad6SPeter Xu /* Safe to modify directly (none->non-present). */ 666960dfaad6SPeter Xu set_huge_pte_at(mm, address, ptep, 6670935d4f0cSRyan Roberts make_pte_marker(PTE_MARKER_UFFD_WP), 6671935d4f0cSRyan Roberts psize); 66728f860591SZhang, Yanmin } 6673cb900f41SKirill A. Shutemov spin_unlock(ptl); 66748f860591SZhang, Yanmin } 6675d833352aSMel Gorman /* 6676c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 6677d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 6678c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 6679dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 6680dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 6681d833352aSMel Gorman */ 6682dff11abeSMike Kravetz if (shared_pmd) 6683ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 6684dff11abeSMike Kravetz else 66855491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 66860f10851eSJérôme Glisse /* 66871af5a810SAlistair Popple * No need to call mmu_notifier_arch_invalidate_secondary_tlbs() we are 66881af5a810SAlistair Popple * downgrading page table protection not changing it to point to a new 66891af5a810SAlistair Popple * page. 66900f10851eSJérôme Glisse * 6691ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 66920f10851eSJérôme Glisse */ 669383cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 669440549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 6695ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 66967da4d641SPeter Zijlstra 6697d1751118SPeter Xu return pages > 0 ? (pages << h->order) : pages; 66988f860591SZhang, Yanmin } 66998f860591SZhang, Yanmin 670033b8f84aSMike Kravetz /* Return true if reservation was successful, false otherwise. */ 670133b8f84aSMike Kravetz bool hugetlb_reserve_pages(struct inode *inode, 6702a1e78772SMel Gorman long from, long to, 67035a6fe125SMel Gorman struct vm_area_struct *vma, 6704ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 6705e4e574b7SAdam Litke { 6706c5094ec7SMike Kravetz long chg = -1, add = -1; 6707a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 670890481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 67099119a41eSJoonsoo Kim struct resv_map *resv_map; 6710075a61d0SMina Almasry struct hugetlb_cgroup *h_cg = NULL; 67110db9d74eSMina Almasry long gbl_reserve, regions_needed = 0; 6712e4e574b7SAdam Litke 671363489f8eSMike Kravetz /* This should never happen */ 671463489f8eSMike Kravetz if (from > to) { 671563489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 671633b8f84aSMike Kravetz return false; 671763489f8eSMike Kravetz } 671863489f8eSMike Kravetz 6719a1e78772SMel Gorman /* 6720e700898fSMike Kravetz * vma specific semaphore used for pmd sharing and fault/truncation 6721e700898fSMike Kravetz * synchronization 67228d9bfb26SMike Kravetz */ 67238d9bfb26SMike Kravetz hugetlb_vma_lock_alloc(vma); 67248d9bfb26SMike Kravetz 67258d9bfb26SMike Kravetz /* 672617c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 672717c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 672890481622SDavid Gibson * without using reserves 672917c9d12eSMel Gorman */ 6730ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 673133b8f84aSMike Kravetz return true; 673217c9d12eSMel Gorman 673317c9d12eSMel Gorman /* 6734a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 6735a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 6736a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 6737a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 6738a1e78772SMel Gorman */ 67399119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 6740f27a5136SMike Kravetz /* 6741f27a5136SMike Kravetz * resv_map can not be NULL as hugetlb_reserve_pages is only 6742f27a5136SMike Kravetz * called for inodes for which resv_maps were created (see 6743f27a5136SMike Kravetz * hugetlbfs_get_inode). 6744f27a5136SMike Kravetz */ 67454e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 67469119a41eSJoonsoo Kim 67470db9d74eSMina Almasry chg = region_chg(resv_map, from, to, ®ions_needed); 67489119a41eSJoonsoo Kim } else { 6749e9fe92aeSMina Almasry /* Private mapping. */ 67509119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 67515a6fe125SMel Gorman if (!resv_map) 67528d9bfb26SMike Kravetz goto out_err; 67535a6fe125SMel Gorman 675417c9d12eSMel Gorman chg = to - from; 675517c9d12eSMel Gorman 67565a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 67575a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 67585a6fe125SMel Gorman } 67595a6fe125SMel Gorman 676033b8f84aSMike Kravetz if (chg < 0) 6761c50ac050SDave Hansen goto out_err; 676217c9d12eSMel Gorman 676333b8f84aSMike Kravetz if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), 676433b8f84aSMike Kravetz chg * pages_per_huge_page(h), &h_cg) < 0) 6765075a61d0SMina Almasry goto out_err; 6766075a61d0SMina Almasry 6767075a61d0SMina Almasry if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { 6768075a61d0SMina Almasry /* For private mappings, the hugetlb_cgroup uncharge info hangs 6769075a61d0SMina Almasry * of the resv_map. 6770075a61d0SMina Almasry */ 6771075a61d0SMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); 6772075a61d0SMina Almasry } 6773075a61d0SMina Almasry 67741c5ecae3SMike Kravetz /* 67751c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 67761c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 67771c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 67781c5ecae3SMike Kravetz */ 67791c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 678033b8f84aSMike Kravetz if (gbl_reserve < 0) 6781075a61d0SMina Almasry goto out_uncharge_cgroup; 678217c9d12eSMel Gorman 678317c9d12eSMel Gorman /* 678417c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 678590481622SDavid Gibson * Hand the pages back to the subpool if there are not 678617c9d12eSMel Gorman */ 678733b8f84aSMike Kravetz if (hugetlb_acct_memory(h, gbl_reserve) < 0) 6788075a61d0SMina Almasry goto out_put_pages; 678917c9d12eSMel Gorman 679017c9d12eSMel Gorman /* 679117c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 679217c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 679317c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 679417c9d12eSMel Gorman * the reservation was and the page cache tells how much of 679517c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 679617c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 679717c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 679817c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 679917c9d12eSMel Gorman * else has to be done for private mappings here 680017c9d12eSMel Gorman */ 680133039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 6802075a61d0SMina Almasry add = region_add(resv_map, from, to, regions_needed, h, h_cg); 680333039678SMike Kravetz 68040db9d74eSMina Almasry if (unlikely(add < 0)) { 68050db9d74eSMina Almasry hugetlb_acct_memory(h, -gbl_reserve); 6806075a61d0SMina Almasry goto out_put_pages; 68070db9d74eSMina Almasry } else if (unlikely(chg > add)) { 680833039678SMike Kravetz /* 680933039678SMike Kravetz * pages in this range were added to the reserve 681033039678SMike Kravetz * map between region_chg and region_add. This 6811d0ce0e47SSidhartha Kumar * indicates a race with alloc_hugetlb_folio. Adjust 681233039678SMike Kravetz * the subpool and reserve counts modified above 681333039678SMike Kravetz * based on the difference. 681433039678SMike Kravetz */ 681533039678SMike Kravetz long rsv_adjust; 681633039678SMike Kravetz 6817d85aecf2SMiaohe Lin /* 6818d85aecf2SMiaohe Lin * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the 6819d85aecf2SMiaohe Lin * reference to h_cg->css. See comment below for detail. 6820d85aecf2SMiaohe Lin */ 6821075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd( 6822075a61d0SMina Almasry hstate_index(h), 6823075a61d0SMina Almasry (chg - add) * pages_per_huge_page(h), h_cg); 6824075a61d0SMina Almasry 682533039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 682633039678SMike Kravetz chg - add); 682733039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 6828d85aecf2SMiaohe Lin } else if (h_cg) { 6829d85aecf2SMiaohe Lin /* 6830d85aecf2SMiaohe Lin * The file_regions will hold their own reference to 6831d85aecf2SMiaohe Lin * h_cg->css. So we should release the reference held 6832d85aecf2SMiaohe Lin * via hugetlb_cgroup_charge_cgroup_rsvd() when we are 6833d85aecf2SMiaohe Lin * done. 6834d85aecf2SMiaohe Lin */ 6835d85aecf2SMiaohe Lin hugetlb_cgroup_put_rsvd_cgroup(h_cg); 683633039678SMike Kravetz } 683733039678SMike Kravetz } 683833b8f84aSMike Kravetz return true; 683933b8f84aSMike Kravetz 6840075a61d0SMina Almasry out_put_pages: 6841075a61d0SMina Almasry /* put back original number of pages, chg */ 6842075a61d0SMina Almasry (void)hugepage_subpool_put_pages(spool, chg); 6843075a61d0SMina Almasry out_uncharge_cgroup: 6844075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), 6845075a61d0SMina Almasry chg * pages_per_huge_page(h), h_cg); 6846c50ac050SDave Hansen out_err: 68478d9bfb26SMike Kravetz hugetlb_vma_lock_free(vma); 68485e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 68490db9d74eSMina Almasry /* Only call region_abort if the region_chg succeeded but the 68500db9d74eSMina Almasry * region_add failed or didn't run. 68510db9d74eSMina Almasry */ 68520db9d74eSMina Almasry if (chg >= 0 && add < 0) 68530db9d74eSMina Almasry region_abort(resv_map, from, to, regions_needed); 685492fe9dcbSRik van Riel if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 6855f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 685692fe9dcbSRik van Riel set_vma_resv_map(vma, NULL); 685792fe9dcbSRik van Riel } 685833b8f84aSMike Kravetz return false; 6859a43a8c39SChen, Kenneth W } 6860a43a8c39SChen, Kenneth W 6861b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 6862b5cec28dSMike Kravetz long freed) 6863a43a8c39SChen, Kenneth W { 6864a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 68654e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 68669119a41eSJoonsoo Kim long chg = 0; 686790481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 68681c5ecae3SMike Kravetz long gbl_reserve; 686945c682a6SKen Chen 6870f27a5136SMike Kravetz /* 6871f27a5136SMike Kravetz * Since this routine can be called in the evict inode path for all 6872f27a5136SMike Kravetz * hugetlbfs inodes, resv_map could be NULL. 6873f27a5136SMike Kravetz */ 6874b5cec28dSMike Kravetz if (resv_map) { 6875b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 6876b5cec28dSMike Kravetz /* 6877b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 6878b5cec28dSMike Kravetz * must be split and another region descriptor can not be 6879b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 6880b5cec28dSMike Kravetz */ 6881b5cec28dSMike Kravetz if (chg < 0) 6882b5cec28dSMike Kravetz return chg; 6883b5cec28dSMike Kravetz } 6884b5cec28dSMike Kravetz 688545c682a6SKen Chen spin_lock(&inode->i_lock); 6886e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 688745c682a6SKen Chen spin_unlock(&inode->i_lock); 688845c682a6SKen Chen 68891c5ecae3SMike Kravetz /* 68901c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 68911c5ecae3SMike Kravetz * reservations to be released may be adjusted. 6892dddf31a4SMiaohe Lin * 6893dddf31a4SMiaohe Lin * Note that !resv_map implies freed == 0. So (chg - freed) 6894dddf31a4SMiaohe Lin * won't go negative. 68951c5ecae3SMike Kravetz */ 68961c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 68971c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 6898b5cec28dSMike Kravetz 6899b5cec28dSMike Kravetz return 0; 6900a43a8c39SChen, Kenneth W } 690193f70f90SNaoya Horiguchi 69023212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 69033212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 69043212b535SSteve Capper struct vm_area_struct *vma, 69053212b535SSteve Capper unsigned long addr, pgoff_t idx) 69063212b535SSteve Capper { 69073212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 69083212b535SSteve Capper svma->vm_start; 69093212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 69103212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 69113212b535SSteve Capper 69123212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 6913e430a95aSSuren Baghdasaryan unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED_MASK; 6914e430a95aSSuren Baghdasaryan unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED_MASK; 69153212b535SSteve Capper 69163212b535SSteve Capper /* 69173212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 69183212b535SSteve Capper * page table page. 6919131a79b4SMike Kravetz * 6920131a79b4SMike Kravetz * Also, vma_lock (vm_private_data) is required for sharing. 69213212b535SSteve Capper */ 69223212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 69233212b535SSteve Capper vm_flags != svm_flags || 6924131a79b4SMike Kravetz !range_in_vma(svma, sbase, s_end) || 6925131a79b4SMike Kravetz !svma->vm_private_data) 69263212b535SSteve Capper return 0; 69273212b535SSteve Capper 69283212b535SSteve Capper return saddr; 69293212b535SSteve Capper } 69303212b535SSteve Capper 6931bbff39ccSMike Kravetz bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 69323212b535SSteve Capper { 6933bbff39ccSMike Kravetz unsigned long start = addr & PUD_MASK; 6934bbff39ccSMike Kravetz unsigned long end = start + PUD_SIZE; 6935bbff39ccSMike Kravetz 69368d9bfb26SMike Kravetz #ifdef CONFIG_USERFAULTFD 69378d9bfb26SMike Kravetz if (uffd_disable_huge_pmd_share(vma)) 69388d9bfb26SMike Kravetz return false; 69398d9bfb26SMike Kravetz #endif 69403212b535SSteve Capper /* 69413212b535SSteve Capper * check on proper vm_flags and page table alignment 69423212b535SSteve Capper */ 69438d9bfb26SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 694431aafb45SNicholas Krause return false; 6945bbff39ccSMike Kravetz if (!vma->vm_private_data) /* vma lock required for sharing */ 69468d9bfb26SMike Kravetz return false; 69478d9bfb26SMike Kravetz if (!range_in_vma(vma, start, end)) 69488d9bfb26SMike Kravetz return false; 69498d9bfb26SMike Kravetz return true; 69508d9bfb26SMike Kravetz } 69518d9bfb26SMike Kravetz 69523212b535SSteve Capper /* 6953017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 6954017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 6955017b1660SMike Kravetz * shared pmd mappings. 6956017b1660SMike Kravetz */ 6957017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 6958017b1660SMike Kravetz unsigned long *start, unsigned long *end) 6959017b1660SMike Kravetz { 6960a1ba9da8SLi Xinhai unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), 6961a1ba9da8SLi Xinhai v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 6962017b1660SMike Kravetz 6963a1ba9da8SLi Xinhai /* 6964f0953a1bSIngo Molnar * vma needs to span at least one aligned PUD size, and the range 6965f0953a1bSIngo Molnar * must be at least partially within in. 6966a1ba9da8SLi Xinhai */ 6967a1ba9da8SLi Xinhai if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || 6968a1ba9da8SLi Xinhai (*end <= v_start) || (*start >= v_end)) 6969017b1660SMike Kravetz return; 6970017b1660SMike Kravetz 697175802ca6SPeter Xu /* Extend the range to be PUD aligned for a worst case scenario */ 6972a1ba9da8SLi Xinhai if (*start > v_start) 6973a1ba9da8SLi Xinhai *start = ALIGN_DOWN(*start, PUD_SIZE); 6974017b1660SMike Kravetz 6975a1ba9da8SLi Xinhai if (*end < v_end) 6976a1ba9da8SLi Xinhai *end = ALIGN(*end, PUD_SIZE); 6977017b1660SMike Kravetz } 6978017b1660SMike Kravetz 6979017b1660SMike Kravetz /* 69803212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 69813212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 69823212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 69833a47c54fSMike Kravetz * code much cleaner. pmd allocation is essential for the shared case because 69843a47c54fSMike Kravetz * pud has to be populated inside the same i_mmap_rwsem section - otherwise 69853a47c54fSMike Kravetz * racing tasks could either miss the sharing (see huge_pte_offset) or select a 69863a47c54fSMike Kravetz * bad pmd for sharing. 69873212b535SSteve Capper */ 6988aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 6989aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 69903212b535SSteve Capper { 69913212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 69923212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 69933212b535SSteve Capper vma->vm_pgoff; 69943212b535SSteve Capper struct vm_area_struct *svma; 69953212b535SSteve Capper unsigned long saddr; 69963212b535SSteve Capper pte_t *spte = NULL; 69973212b535SSteve Capper pte_t *pte; 69983212b535SSteve Capper 69993a47c54fSMike Kravetz i_mmap_lock_read(mapping); 70003212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 70013212b535SSteve Capper if (svma == vma) 70023212b535SSteve Capper continue; 70033212b535SSteve Capper 70043212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 70053212b535SSteve Capper if (saddr) { 70069c67a207SPeter Xu spte = hugetlb_walk(svma, saddr, 70077868a208SPunit Agrawal vma_mmu_pagesize(svma)); 70083212b535SSteve Capper if (spte) { 70093212b535SSteve Capper get_page(virt_to_page(spte)); 70103212b535SSteve Capper break; 70113212b535SSteve Capper } 70123212b535SSteve Capper } 70133212b535SSteve Capper } 70143212b535SSteve Capper 70153212b535SSteve Capper if (!spte) 70163212b535SSteve Capper goto out; 70173212b535SSteve Capper 7018349d1670SPeter Xu spin_lock(&mm->page_table_lock); 7019dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 70203212b535SSteve Capper pud_populate(mm, pud, 70213212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 7022c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 7023dc6c9a35SKirill A. Shutemov } else { 70243212b535SSteve Capper put_page(virt_to_page(spte)); 7025dc6c9a35SKirill A. Shutemov } 7026349d1670SPeter Xu spin_unlock(&mm->page_table_lock); 70273212b535SSteve Capper out: 70283212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 70293a47c54fSMike Kravetz i_mmap_unlock_read(mapping); 70303212b535SSteve Capper return pte; 70313212b535SSteve Capper } 70323212b535SSteve Capper 70333212b535SSteve Capper /* 70343212b535SSteve Capper * unmap huge page backed by shared pte. 70353212b535SSteve Capper * 70363212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 70373212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 70383212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 70393212b535SSteve Capper * 70403a47c54fSMike Kravetz * Called with page table lock held. 70413212b535SSteve Capper * 70423212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 70433212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 70443212b535SSteve Capper */ 704534ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 70464ddb4d91SMike Kravetz unsigned long addr, pte_t *ptep) 70473212b535SSteve Capper { 70484ddb4d91SMike Kravetz pgd_t *pgd = pgd_offset(mm, addr); 70494ddb4d91SMike Kravetz p4d_t *p4d = p4d_offset(pgd, addr); 70504ddb4d91SMike Kravetz pud_t *pud = pud_offset(p4d, addr); 70513212b535SSteve Capper 705234ae204fSMike Kravetz i_mmap_assert_write_locked(vma->vm_file->f_mapping); 705340549ba8SMike Kravetz hugetlb_vma_assert_locked(vma); 70543212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 70553212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 70563212b535SSteve Capper return 0; 70573212b535SSteve Capper 70583212b535SSteve Capper pud_clear(pud); 70593212b535SSteve Capper put_page(virt_to_page(ptep)); 7060dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 70613212b535SSteve Capper return 1; 70623212b535SSteve Capper } 7063c1991e07SPeter Xu 70649e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 70658d9bfb26SMike Kravetz 7066aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 7067aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 70689e5fc74cSSteve Capper { 70699e5fc74cSSteve Capper return NULL; 70709e5fc74cSSteve Capper } 7071e81f2d22SZhang Zhen 707234ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 70734ddb4d91SMike Kravetz unsigned long addr, pte_t *ptep) 7074e81f2d22SZhang Zhen { 7075e81f2d22SZhang Zhen return 0; 7076e81f2d22SZhang Zhen } 7077017b1660SMike Kravetz 7078017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 7079017b1660SMike Kravetz unsigned long *start, unsigned long *end) 7080017b1660SMike Kravetz { 7081017b1660SMike Kravetz } 7082c1991e07SPeter Xu 7083c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 7084c1991e07SPeter Xu { 7085c1991e07SPeter Xu return false; 7086c1991e07SPeter Xu } 70873212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 70883212b535SSteve Capper 70899e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 7090aec44e0fSPeter Xu pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, 70919e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 70929e5fc74cSSteve Capper { 70939e5fc74cSSteve Capper pgd_t *pgd; 7094c2febafcSKirill A. Shutemov p4d_t *p4d; 70959e5fc74cSSteve Capper pud_t *pud; 70969e5fc74cSSteve Capper pte_t *pte = NULL; 70979e5fc74cSSteve Capper 70989e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 7099f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 7100f4f0a3d8SKirill A. Shutemov if (!p4d) 7101f4f0a3d8SKirill A. Shutemov return NULL; 7102c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 71039e5fc74cSSteve Capper if (pud) { 71049e5fc74cSSteve Capper if (sz == PUD_SIZE) { 71059e5fc74cSSteve Capper pte = (pte_t *)pud; 71069e5fc74cSSteve Capper } else { 71079e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 7108c1991e07SPeter Xu if (want_pmd_share(vma, addr) && pud_none(*pud)) 7109aec44e0fSPeter Xu pte = huge_pmd_share(mm, vma, addr, pud); 71109e5fc74cSSteve Capper else 71119e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 71129e5fc74cSSteve Capper } 71139e5fc74cSSteve Capper } 7114191fcdb6SJohn Hubbard 7115191fcdb6SJohn Hubbard if (pte) { 7116191fcdb6SJohn Hubbard pte_t pteval = ptep_get_lockless(pte); 7117191fcdb6SJohn Hubbard 7118191fcdb6SJohn Hubbard BUG_ON(pte_present(pteval) && !pte_huge(pteval)); 7119191fcdb6SJohn Hubbard } 71209e5fc74cSSteve Capper 71219e5fc74cSSteve Capper return pte; 71229e5fc74cSSteve Capper } 71239e5fc74cSSteve Capper 71249b19df29SPunit Agrawal /* 71259b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 71269b19df29SPunit Agrawal * entry at address @addr 71279b19df29SPunit Agrawal * 71288ac0b81aSLi Xinhai * Return: Pointer to page table entry (PUD or PMD) for 71298ac0b81aSLi Xinhai * address @addr, or NULL if a !p*d_present() entry is encountered and the 71309b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 71319b19df29SPunit Agrawal * table. 71329b19df29SPunit Agrawal */ 71337868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 71347868a208SPunit Agrawal unsigned long addr, unsigned long sz) 71359e5fc74cSSteve Capper { 71369e5fc74cSSteve Capper pgd_t *pgd; 7137c2febafcSKirill A. Shutemov p4d_t *p4d; 71388ac0b81aSLi Xinhai pud_t *pud; 71398ac0b81aSLi Xinhai pmd_t *pmd; 71409e5fc74cSSteve Capper 71419e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 7142c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 7143c2febafcSKirill A. Shutemov return NULL; 7144c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 7145c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 7146c2febafcSKirill A. Shutemov return NULL; 71479b19df29SPunit Agrawal 7148c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 71498ac0b81aSLi Xinhai if (sz == PUD_SIZE) 71508ac0b81aSLi Xinhai /* must be pud huge, non-present or none */ 71519e5fc74cSSteve Capper return (pte_t *)pud; 71528ac0b81aSLi Xinhai if (!pud_present(*pud)) 71538ac0b81aSLi Xinhai return NULL; 71548ac0b81aSLi Xinhai /* must have a valid entry and size to go further */ 71559b19df29SPunit Agrawal 71569e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 71578ac0b81aSLi Xinhai /* must be pmd huge, non-present or none */ 71589e5fc74cSSteve Capper return (pte_t *)pmd; 71599e5fc74cSSteve Capper } 71609e5fc74cSSteve Capper 7161e95a9851SMike Kravetz /* 7162e95a9851SMike Kravetz * Return a mask that can be used to update an address to the last huge 7163e95a9851SMike Kravetz * page in a page table page mapping size. Used to skip non-present 7164e95a9851SMike Kravetz * page table entries when linearly scanning address ranges. Architectures 7165e95a9851SMike Kravetz * with unique huge page to page table relationships can define their own 7166e95a9851SMike Kravetz * version of this routine. 7167e95a9851SMike Kravetz */ 7168e95a9851SMike Kravetz unsigned long hugetlb_mask_last_page(struct hstate *h) 7169e95a9851SMike Kravetz { 7170e95a9851SMike Kravetz unsigned long hp_size = huge_page_size(h); 7171e95a9851SMike Kravetz 7172e95a9851SMike Kravetz if (hp_size == PUD_SIZE) 7173e95a9851SMike Kravetz return P4D_SIZE - PUD_SIZE; 7174e95a9851SMike Kravetz else if (hp_size == PMD_SIZE) 7175e95a9851SMike Kravetz return PUD_SIZE - PMD_SIZE; 7176e95a9851SMike Kravetz else 7177e95a9851SMike Kravetz return 0UL; 7178e95a9851SMike Kravetz } 7179e95a9851SMike Kravetz 7180e95a9851SMike Kravetz #else 7181e95a9851SMike Kravetz 7182e95a9851SMike Kravetz /* See description above. Architectures can provide their own version. */ 7183e95a9851SMike Kravetz __weak unsigned long hugetlb_mask_last_page(struct hstate *h) 7184e95a9851SMike Kravetz { 71854ddb4d91SMike Kravetz #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 71864ddb4d91SMike Kravetz if (huge_page_size(h) == PMD_SIZE) 71874ddb4d91SMike Kravetz return PUD_SIZE - PMD_SIZE; 71884ddb4d91SMike Kravetz #endif 7189e95a9851SMike Kravetz return 0UL; 7190e95a9851SMike Kravetz } 7191e95a9851SMike Kravetz 719261f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 719361f77edaSNaoya Horiguchi 719461f77edaSNaoya Horiguchi /* 719561f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 719661f77edaSNaoya Horiguchi * behavior. 719761f77edaSNaoya Horiguchi */ 71989747b9e9SBaolin Wang bool isolate_hugetlb(struct folio *folio, struct list_head *list) 719931caf665SNaoya Horiguchi { 72009747b9e9SBaolin Wang bool ret = true; 7201bcc54222SNaoya Horiguchi 7202db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 72036aa3a920SSidhartha Kumar if (!folio_test_hugetlb(folio) || 72046aa3a920SSidhartha Kumar !folio_test_hugetlb_migratable(folio) || 72056aa3a920SSidhartha Kumar !folio_try_get(folio)) { 72069747b9e9SBaolin Wang ret = false; 7207bcc54222SNaoya Horiguchi goto unlock; 7208bcc54222SNaoya Horiguchi } 72096aa3a920SSidhartha Kumar folio_clear_hugetlb_migratable(folio); 72106aa3a920SSidhartha Kumar list_move_tail(&folio->lru, list); 7211bcc54222SNaoya Horiguchi unlock: 7212db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7213bcc54222SNaoya Horiguchi return ret; 721431caf665SNaoya Horiguchi } 721531caf665SNaoya Horiguchi 721604bac040SSidhartha Kumar int get_hwpoison_hugetlb_folio(struct folio *folio, bool *hugetlb, bool unpoison) 721725182f05SNaoya Horiguchi { 721825182f05SNaoya Horiguchi int ret = 0; 721925182f05SNaoya Horiguchi 722025182f05SNaoya Horiguchi *hugetlb = false; 722125182f05SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 722204bac040SSidhartha Kumar if (folio_test_hugetlb(folio)) { 722325182f05SNaoya Horiguchi *hugetlb = true; 722404bac040SSidhartha Kumar if (folio_test_hugetlb_freed(folio)) 7225b283d983SNaoya Horiguchi ret = 0; 722604bac040SSidhartha Kumar else if (folio_test_hugetlb_migratable(folio) || unpoison) 722704bac040SSidhartha Kumar ret = folio_try_get(folio); 72280ed950d1SNaoya Horiguchi else 72290ed950d1SNaoya Horiguchi ret = -EBUSY; 723025182f05SNaoya Horiguchi } 723125182f05SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 723225182f05SNaoya Horiguchi return ret; 723325182f05SNaoya Horiguchi } 723425182f05SNaoya Horiguchi 7235e591ef7dSNaoya Horiguchi int get_huge_page_for_hwpoison(unsigned long pfn, int flags, 7236e591ef7dSNaoya Horiguchi bool *migratable_cleared) 7237405ce051SNaoya Horiguchi { 7238405ce051SNaoya Horiguchi int ret; 7239405ce051SNaoya Horiguchi 7240405ce051SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 7241e591ef7dSNaoya Horiguchi ret = __get_huge_page_for_hwpoison(pfn, flags, migratable_cleared); 7242405ce051SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 7243405ce051SNaoya Horiguchi return ret; 7244405ce051SNaoya Horiguchi } 7245405ce051SNaoya Horiguchi 7246ea8e72f4SSidhartha Kumar void folio_putback_active_hugetlb(struct folio *folio) 724731caf665SNaoya Horiguchi { 7248db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 7249ea8e72f4SSidhartha Kumar folio_set_hugetlb_migratable(folio); 7250ea8e72f4SSidhartha Kumar list_move_tail(&folio->lru, &(folio_hstate(folio))->hugepage_activelist); 7251db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7252ea8e72f4SSidhartha Kumar folio_put(folio); 725331caf665SNaoya Horiguchi } 7254ab5ac90aSMichal Hocko 7255345c62d1SSidhartha Kumar void move_hugetlb_state(struct folio *old_folio, struct folio *new_folio, int reason) 7256ab5ac90aSMichal Hocko { 7257345c62d1SSidhartha Kumar struct hstate *h = folio_hstate(old_folio); 7258ab5ac90aSMichal Hocko 7259345c62d1SSidhartha Kumar hugetlb_cgroup_migrate(old_folio, new_folio); 7260345c62d1SSidhartha Kumar set_page_owner_migrate_reason(&new_folio->page, reason); 7261ab5ac90aSMichal Hocko 7262ab5ac90aSMichal Hocko /* 7263345c62d1SSidhartha Kumar * transfer temporary state of the new hugetlb folio. This is 7264ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 7265ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 7266ab5ac90aSMichal Hocko * the temporary status. 7267ab5ac90aSMichal Hocko * 7268ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 7269ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 7270ab5ac90aSMichal Hocko * the per-node's. 7271ab5ac90aSMichal Hocko */ 7272345c62d1SSidhartha Kumar if (folio_test_hugetlb_temporary(new_folio)) { 7273345c62d1SSidhartha Kumar int old_nid = folio_nid(old_folio); 7274345c62d1SSidhartha Kumar int new_nid = folio_nid(new_folio); 7275ab5ac90aSMichal Hocko 7276345c62d1SSidhartha Kumar folio_set_hugetlb_temporary(old_folio); 7277345c62d1SSidhartha Kumar folio_clear_hugetlb_temporary(new_folio); 7278345c62d1SSidhartha Kumar 7279ab5ac90aSMichal Hocko 72805af1ab1dSMiaohe Lin /* 72815af1ab1dSMiaohe Lin * There is no need to transfer the per-node surplus state 72825af1ab1dSMiaohe Lin * when we do not cross the node. 72835af1ab1dSMiaohe Lin */ 72845af1ab1dSMiaohe Lin if (new_nid == old_nid) 72855af1ab1dSMiaohe Lin return; 7286db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 7287ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 7288ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 7289ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 7290ab5ac90aSMichal Hocko } 7291db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7292ab5ac90aSMichal Hocko } 7293ab5ac90aSMichal Hocko } 7294cf11e85fSRoman Gushchin 7295b30c14cdSJames Houghton static void hugetlb_unshare_pmds(struct vm_area_struct *vma, 7296b30c14cdSJames Houghton unsigned long start, 7297b30c14cdSJames Houghton unsigned long end) 72986dfeaff9SPeter Xu { 72996dfeaff9SPeter Xu struct hstate *h = hstate_vma(vma); 73006dfeaff9SPeter Xu unsigned long sz = huge_page_size(h); 73016dfeaff9SPeter Xu struct mm_struct *mm = vma->vm_mm; 73026dfeaff9SPeter Xu struct mmu_notifier_range range; 7303b30c14cdSJames Houghton unsigned long address; 73046dfeaff9SPeter Xu spinlock_t *ptl; 73056dfeaff9SPeter Xu pte_t *ptep; 73066dfeaff9SPeter Xu 73076dfeaff9SPeter Xu if (!(vma->vm_flags & VM_MAYSHARE)) 73086dfeaff9SPeter Xu return; 73096dfeaff9SPeter Xu 73106dfeaff9SPeter Xu if (start >= end) 73116dfeaff9SPeter Xu return; 73126dfeaff9SPeter Xu 73139c8bbfacSBaolin Wang flush_cache_range(vma, start, end); 73146dfeaff9SPeter Xu /* 73156dfeaff9SPeter Xu * No need to call adjust_range_if_pmd_sharing_possible(), because 73166dfeaff9SPeter Xu * we have already done the PUD_SIZE alignment. 73176dfeaff9SPeter Xu */ 73187d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, 73196dfeaff9SPeter Xu start, end); 73206dfeaff9SPeter Xu mmu_notifier_invalidate_range_start(&range); 732140549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 73226dfeaff9SPeter Xu i_mmap_lock_write(vma->vm_file->f_mapping); 73236dfeaff9SPeter Xu for (address = start; address < end; address += PUD_SIZE) { 73249c67a207SPeter Xu ptep = hugetlb_walk(vma, address, sz); 73256dfeaff9SPeter Xu if (!ptep) 73266dfeaff9SPeter Xu continue; 73276dfeaff9SPeter Xu ptl = huge_pte_lock(h, mm, ptep); 73284ddb4d91SMike Kravetz huge_pmd_unshare(mm, vma, address, ptep); 73296dfeaff9SPeter Xu spin_unlock(ptl); 73306dfeaff9SPeter Xu } 73316dfeaff9SPeter Xu flush_hugetlb_tlb_range(vma, start, end); 73326dfeaff9SPeter Xu i_mmap_unlock_write(vma->vm_file->f_mapping); 733340549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 73346dfeaff9SPeter Xu /* 73351af5a810SAlistair Popple * No need to call mmu_notifier_arch_invalidate_secondary_tlbs(), see 7336ee65728eSMike Rapoport * Documentation/mm/mmu_notifier.rst. 73376dfeaff9SPeter Xu */ 73386dfeaff9SPeter Xu mmu_notifier_invalidate_range_end(&range); 73396dfeaff9SPeter Xu } 73406dfeaff9SPeter Xu 7341b30c14cdSJames Houghton /* 7342b30c14cdSJames Houghton * This function will unconditionally remove all the shared pmd pgtable entries 7343b30c14cdSJames Houghton * within the specific vma for a hugetlbfs memory range. 7344b30c14cdSJames Houghton */ 7345b30c14cdSJames Houghton void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) 7346b30c14cdSJames Houghton { 7347b30c14cdSJames Houghton hugetlb_unshare_pmds(vma, ALIGN(vma->vm_start, PUD_SIZE), 7348b30c14cdSJames Houghton ALIGN_DOWN(vma->vm_end, PUD_SIZE)); 7349b30c14cdSJames Houghton } 7350b30c14cdSJames Houghton 7351cf11e85fSRoman Gushchin #ifdef CONFIG_CMA 7352cf11e85fSRoman Gushchin static bool cma_reserve_called __initdata; 7353cf11e85fSRoman Gushchin 7354cf11e85fSRoman Gushchin static int __init cmdline_parse_hugetlb_cma(char *p) 7355cf11e85fSRoman Gushchin { 735638e719abSBaolin Wang int nid, count = 0; 735738e719abSBaolin Wang unsigned long tmp; 735838e719abSBaolin Wang char *s = p; 735938e719abSBaolin Wang 736038e719abSBaolin Wang while (*s) { 736138e719abSBaolin Wang if (sscanf(s, "%lu%n", &tmp, &count) != 1) 736238e719abSBaolin Wang break; 736338e719abSBaolin Wang 736438e719abSBaolin Wang if (s[count] == ':') { 7365f9317f77SMike Kravetz if (tmp >= MAX_NUMNODES) 736638e719abSBaolin Wang break; 7367f9317f77SMike Kravetz nid = array_index_nospec(tmp, MAX_NUMNODES); 736838e719abSBaolin Wang 736938e719abSBaolin Wang s += count + 1; 737038e719abSBaolin Wang tmp = memparse(s, &s); 737138e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = tmp; 737238e719abSBaolin Wang hugetlb_cma_size += tmp; 737338e719abSBaolin Wang 737438e719abSBaolin Wang /* 737538e719abSBaolin Wang * Skip the separator if have one, otherwise 737638e719abSBaolin Wang * break the parsing. 737738e719abSBaolin Wang */ 737838e719abSBaolin Wang if (*s == ',') 737938e719abSBaolin Wang s++; 738038e719abSBaolin Wang else 738138e719abSBaolin Wang break; 738238e719abSBaolin Wang } else { 7383cf11e85fSRoman Gushchin hugetlb_cma_size = memparse(p, &p); 738438e719abSBaolin Wang break; 738538e719abSBaolin Wang } 738638e719abSBaolin Wang } 738738e719abSBaolin Wang 7388cf11e85fSRoman Gushchin return 0; 7389cf11e85fSRoman Gushchin } 7390cf11e85fSRoman Gushchin 7391cf11e85fSRoman Gushchin early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); 7392cf11e85fSRoman Gushchin 7393cf11e85fSRoman Gushchin void __init hugetlb_cma_reserve(int order) 7394cf11e85fSRoman Gushchin { 7395cf11e85fSRoman Gushchin unsigned long size, reserved, per_node; 739638e719abSBaolin Wang bool node_specific_cma_alloc = false; 7397cf11e85fSRoman Gushchin int nid; 7398cf11e85fSRoman Gushchin 7399cf11e85fSRoman Gushchin cma_reserve_called = true; 7400cf11e85fSRoman Gushchin 7401cf11e85fSRoman Gushchin if (!hugetlb_cma_size) 7402cf11e85fSRoman Gushchin return; 7403cf11e85fSRoman Gushchin 740438e719abSBaolin Wang for (nid = 0; nid < MAX_NUMNODES; nid++) { 740538e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 740638e719abSBaolin Wang continue; 740738e719abSBaolin Wang 740830a51400SPeng Liu if (!node_online(nid)) { 740938e719abSBaolin Wang pr_warn("hugetlb_cma: invalid node %d specified\n", nid); 741038e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 741138e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 741238e719abSBaolin Wang continue; 741338e719abSBaolin Wang } 741438e719abSBaolin Wang 741538e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) { 741638e719abSBaolin Wang pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n", 741738e719abSBaolin Wang nid, (PAGE_SIZE << order) / SZ_1M); 741838e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 741938e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 742038e719abSBaolin Wang } else { 742138e719abSBaolin Wang node_specific_cma_alloc = true; 742238e719abSBaolin Wang } 742338e719abSBaolin Wang } 742438e719abSBaolin Wang 742538e719abSBaolin Wang /* Validate the CMA size again in case some invalid nodes specified. */ 742638e719abSBaolin Wang if (!hugetlb_cma_size) 742738e719abSBaolin Wang return; 742838e719abSBaolin Wang 7429cf11e85fSRoman Gushchin if (hugetlb_cma_size < (PAGE_SIZE << order)) { 7430cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", 7431cf11e85fSRoman Gushchin (PAGE_SIZE << order) / SZ_1M); 7432a01f4390SMike Kravetz hugetlb_cma_size = 0; 7433cf11e85fSRoman Gushchin return; 7434cf11e85fSRoman Gushchin } 7435cf11e85fSRoman Gushchin 743638e719abSBaolin Wang if (!node_specific_cma_alloc) { 7437cf11e85fSRoman Gushchin /* 7438cf11e85fSRoman Gushchin * If 3 GB area is requested on a machine with 4 numa nodes, 7439cf11e85fSRoman Gushchin * let's allocate 1 GB on first three nodes and ignore the last one. 7440cf11e85fSRoman Gushchin */ 7441cf11e85fSRoman Gushchin per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); 7442cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", 7443cf11e85fSRoman Gushchin hugetlb_cma_size / SZ_1M, per_node / SZ_1M); 744438e719abSBaolin Wang } 7445cf11e85fSRoman Gushchin 7446cf11e85fSRoman Gushchin reserved = 0; 744730a51400SPeng Liu for_each_online_node(nid) { 7448cf11e85fSRoman Gushchin int res; 74492281f797SBarry Song char name[CMA_MAX_NAME]; 7450cf11e85fSRoman Gushchin 745138e719abSBaolin Wang if (node_specific_cma_alloc) { 745238e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 745338e719abSBaolin Wang continue; 745438e719abSBaolin Wang 745538e719abSBaolin Wang size = hugetlb_cma_size_in_node[nid]; 745638e719abSBaolin Wang } else { 7457cf11e85fSRoman Gushchin size = min(per_node, hugetlb_cma_size - reserved); 745838e719abSBaolin Wang } 745938e719abSBaolin Wang 7460cf11e85fSRoman Gushchin size = round_up(size, PAGE_SIZE << order); 7461cf11e85fSRoman Gushchin 74622281f797SBarry Song snprintf(name, sizeof(name), "hugetlb%d", nid); 7463a01f4390SMike Kravetz /* 7464a01f4390SMike Kravetz * Note that 'order per bit' is based on smallest size that 7465a01f4390SMike Kravetz * may be returned to CMA allocator in the case of 7466a01f4390SMike Kravetz * huge page demotion. 7467a01f4390SMike Kravetz */ 7468a01f4390SMike Kravetz res = cma_declare_contiguous_nid(0, size, 0, 7469a01f4390SMike Kravetz PAGE_SIZE << HUGETLB_PAGE_ORDER, 747029d0f41dSBarry Song 0, false, name, 7471cf11e85fSRoman Gushchin &hugetlb_cma[nid], nid); 7472cf11e85fSRoman Gushchin if (res) { 7473cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: reservation failed: err %d, node %d", 7474cf11e85fSRoman Gushchin res, nid); 7475cf11e85fSRoman Gushchin continue; 7476cf11e85fSRoman Gushchin } 7477cf11e85fSRoman Gushchin 7478cf11e85fSRoman Gushchin reserved += size; 7479cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", 7480cf11e85fSRoman Gushchin size / SZ_1M, nid); 7481cf11e85fSRoman Gushchin 7482cf11e85fSRoman Gushchin if (reserved >= hugetlb_cma_size) 7483cf11e85fSRoman Gushchin break; 7484cf11e85fSRoman Gushchin } 7485a01f4390SMike Kravetz 7486a01f4390SMike Kravetz if (!reserved) 7487a01f4390SMike Kravetz /* 7488a01f4390SMike Kravetz * hugetlb_cma_size is used to determine if allocations from 7489a01f4390SMike Kravetz * cma are possible. Set to zero if no cma regions are set up. 7490a01f4390SMike Kravetz */ 7491a01f4390SMike Kravetz hugetlb_cma_size = 0; 7492cf11e85fSRoman Gushchin } 7493cf11e85fSRoman Gushchin 7494263b8998SMiaohe Lin static void __init hugetlb_cma_check(void) 7495cf11e85fSRoman Gushchin { 7496cf11e85fSRoman Gushchin if (!hugetlb_cma_size || cma_reserve_called) 7497cf11e85fSRoman Gushchin return; 7498cf11e85fSRoman Gushchin 7499cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); 7500cf11e85fSRoman Gushchin } 7501cf11e85fSRoman Gushchin 7502cf11e85fSRoman Gushchin #endif /* CONFIG_CMA */ 7503