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); 100*bf491692SRik 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); 271*bf491692SRik van Riel } else if (__vma_private_lock(vma)) { 272*bf491692SRik van Riel struct resv_map *resv_map = vma_resv_map(vma); 273*bf491692SRik van Riel 274*bf491692SRik 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); 284*bf491692SRik van Riel } else if (__vma_private_lock(vma)) { 285*bf491692SRik van Riel struct resv_map *resv_map = vma_resv_map(vma); 286*bf491692SRik van Riel 287*bf491692SRik 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); 297*bf491692SRik van Riel } else if (__vma_private_lock(vma)) { 298*bf491692SRik van Riel struct resv_map *resv_map = vma_resv_map(vma); 299*bf491692SRik van Riel 300*bf491692SRik 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); 310*bf491692SRik van Riel } else if (__vma_private_lock(vma)) { 311*bf491692SRik van Riel struct resv_map *resv_map = vma_resv_map(vma); 312*bf491692SRik van Riel 313*bf491692SRik 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 { 319*bf491692SRik van Riel 320*bf491692SRik 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); 324*bf491692SRik van Riel } else if (__vma_private_lock(vma)) { 325*bf491692SRik van Riel struct resv_map *resv_map = vma_resv_map(vma); 326*bf491692SRik van Riel 327*bf491692SRik van Riel return down_write_trylock(&resv_map->rw_sema); 328*bf491692SRik van Riel } 329*bf491692SRik van Riel 330*bf491692SRik 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); 339*bf491692SRik van Riel } else if (__vma_private_lock(vma)) { 340*bf491692SRik van Riel struct resv_map *resv_map = vma_resv_map(vma); 341*bf491692SRik van Riel 342*bf491692SRik 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); 375*bf491692SRik van Riel } else if (__vma_private_lock(vma)) { 376*bf491692SRik van Riel struct resv_map *resv_map = vma_resv_map(vma); 377*bf491692SRik van Riel 378*bf491692SRik van Riel /* no free for anon vmas, but still need to unlock */ 379*bf491692SRik 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); 1103*bf491692SRik 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 11928d9bfb26SMike Kravetz void hugetlb_dup_vma_private(struct vm_area_struct *vma) 1193a1e78772SMel Gorman { 119481d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 11958d9bfb26SMike Kravetz /* 11968d9bfb26SMike Kravetz * Clear vm_private_data 1197612b8a31SMike Kravetz * - For shared mappings this is a per-vma semaphore that may be 1198612b8a31SMike Kravetz * allocated in a subsequent call to hugetlb_vm_op_open. 1199612b8a31SMike Kravetz * Before clearing, make sure pointer is not associated with vma 1200612b8a31SMike Kravetz * as this will leak the structure. This is the case when called 1201612b8a31SMike Kravetz * via clear_vma_resv_huge_pages() and hugetlb_vm_op_open has already 1202612b8a31SMike Kravetz * been called to allocate a new structure. 12038d9bfb26SMike Kravetz * - For MAP_PRIVATE mappings, this is the reserve map which does 12048d9bfb26SMike Kravetz * not apply to children. Faults generated by the children are 12058d9bfb26SMike Kravetz * not guaranteed to succeed, even if read-only. 12068d9bfb26SMike Kravetz */ 1207612b8a31SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 1208612b8a31SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 1209612b8a31SMike Kravetz 1210612b8a31SMike Kravetz if (vma_lock && vma_lock->vma != vma) 1211612b8a31SMike Kravetz vma->vm_private_data = NULL; 1212612b8a31SMike Kravetz } else 1213612b8a31SMike Kravetz vma->vm_private_data = NULL; 1214a1e78772SMel Gorman } 1215a1e78772SMel Gorman 1216550a7d60SMina Almasry /* 1217550a7d60SMina Almasry * Reset and decrement one ref on hugepage private reservation. 12188651a137SLorenzo Stoakes * Called with mm->mmap_lock writer semaphore held. 1219550a7d60SMina Almasry * This function should be only used by move_vma() and operate on 1220550a7d60SMina Almasry * same sized vma. It should never come here with last ref on the 1221550a7d60SMina Almasry * reservation. 1222550a7d60SMina Almasry */ 1223550a7d60SMina Almasry void clear_vma_resv_huge_pages(struct vm_area_struct *vma) 1224550a7d60SMina Almasry { 1225550a7d60SMina Almasry /* 1226550a7d60SMina Almasry * Clear the old hugetlb private page reservation. 1227550a7d60SMina Almasry * It has already been transferred to new_vma. 1228550a7d60SMina Almasry * 1229550a7d60SMina Almasry * During a mremap() operation of a hugetlb vma we call move_vma() 1230550a7d60SMina Almasry * which copies vma into new_vma and unmaps vma. After the copy 1231550a7d60SMina Almasry * operation both new_vma and vma share a reference to the resv_map 1232550a7d60SMina Almasry * struct, and at that point vma is about to be unmapped. We don't 1233550a7d60SMina Almasry * want to return the reservation to the pool at unmap of vma because 1234550a7d60SMina Almasry * the reservation still lives on in new_vma, so simply decrement the 1235550a7d60SMina Almasry * ref here and remove the resv_map reference from this vma. 1236550a7d60SMina Almasry */ 1237550a7d60SMina Almasry struct resv_map *reservations = vma_resv_map(vma); 1238550a7d60SMina Almasry 1239afe041c2SBui Quang Minh if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 1240afe041c2SBui Quang Minh resv_map_put_hugetlb_cgroup_uncharge_info(reservations); 1241550a7d60SMina Almasry kref_put(&reservations->refs, resv_map_release); 1242afe041c2SBui Quang Minh } 1243550a7d60SMina Almasry 12448d9bfb26SMike Kravetz hugetlb_dup_vma_private(vma); 1245550a7d60SMina Almasry } 1246550a7d60SMina Almasry 1247a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 1248559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 1249a1e78772SMel Gorman { 1250af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 1251af0ed73eSJoonsoo Kim /* 1252af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 1253af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 1254af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 1255af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 1256af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 1257af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 1258af0ed73eSJoonsoo Kim * properly, so add work-around here. 1259af0ed73eSJoonsoo Kim */ 1260af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 1261559ec2f8SNicholas Krause return true; 1262af0ed73eSJoonsoo Kim else 1263559ec2f8SNicholas Krause return false; 1264af0ed73eSJoonsoo Kim } 1265a63884e9SJoonsoo Kim 1266a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 12671fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 12681fb1b0e9SMike Kravetz /* 12691fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 12701fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 12711fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 12727c8de358SEthon Paul * fallocate. In this case, there really are no reserves to 12731fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 12741fb1b0e9SMike Kravetz */ 12751fb1b0e9SMike Kravetz if (chg) 12761fb1b0e9SMike Kravetz return false; 12771fb1b0e9SMike Kravetz else 1278559ec2f8SNicholas Krause return true; 12791fb1b0e9SMike Kravetz } 1280a63884e9SJoonsoo Kim 1281a63884e9SJoonsoo Kim /* 1282a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 1283a63884e9SJoonsoo Kim * private mappings. 1284a63884e9SJoonsoo Kim */ 128567961f9dSMike Kravetz if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 128667961f9dSMike Kravetz /* 128767961f9dSMike Kravetz * Like the shared case above, a hole punch or truncate 128867961f9dSMike Kravetz * could have been performed on the private mapping. 128967961f9dSMike Kravetz * Examine the value of chg to determine if reserves 129067961f9dSMike Kravetz * actually exist or were previously consumed. 129167961f9dSMike Kravetz * Very Subtle - The value of chg comes from a previous 129267961f9dSMike Kravetz * call to vma_needs_reserves(). The reserve map for 129367961f9dSMike Kravetz * private mappings has different (opposite) semantics 129467961f9dSMike Kravetz * than that of shared mappings. vma_needs_reserves() 129567961f9dSMike Kravetz * has already taken this difference in semantics into 129667961f9dSMike Kravetz * account. Therefore, the meaning of chg is the same 129767961f9dSMike Kravetz * as in the shared case above. Code could easily be 129867961f9dSMike Kravetz * combined, but keeping it separate draws attention to 129967961f9dSMike Kravetz * subtle differences. 130067961f9dSMike Kravetz */ 130167961f9dSMike Kravetz if (chg) 130267961f9dSMike Kravetz return false; 130367961f9dSMike Kravetz else 1304559ec2f8SNicholas Krause return true; 130567961f9dSMike Kravetz } 1306a63884e9SJoonsoo Kim 1307559ec2f8SNicholas Krause return false; 1308a1e78772SMel Gorman } 1309a1e78772SMel Gorman 1310240d67a8SSidhartha Kumar static void enqueue_hugetlb_folio(struct hstate *h, struct folio *folio) 13111da177e4SLinus Torvalds { 1312240d67a8SSidhartha Kumar int nid = folio_nid(folio); 13139487ca60SMike Kravetz 13149487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1315240d67a8SSidhartha Kumar VM_BUG_ON_FOLIO(folio_ref_count(folio), folio); 1316b65a4edaSMike Kravetz 1317240d67a8SSidhartha Kumar list_move(&folio->lru, &h->hugepage_freelists[nid]); 1318a5516438SAndi Kleen h->free_huge_pages++; 1319a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 1320240d67a8SSidhartha Kumar folio_set_hugetlb_freed(folio); 13211da177e4SLinus Torvalds } 13221da177e4SLinus Torvalds 1323a36f1e90SSidhartha Kumar static struct folio *dequeue_hugetlb_folio_node_exact(struct hstate *h, 1324a36f1e90SSidhartha Kumar int nid) 1325bf50bab2SNaoya Horiguchi { 1326a36f1e90SSidhartha Kumar struct folio *folio; 13271a08ae36SPavel Tatashin bool pin = !!(current->flags & PF_MEMALLOC_PIN); 1328bf50bab2SNaoya Horiguchi 13299487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1330a36f1e90SSidhartha Kumar list_for_each_entry(folio, &h->hugepage_freelists[nid], lru) { 1331a36f1e90SSidhartha Kumar if (pin && !folio_is_longterm_pinnable(folio)) 1332bbe88753SJoonsoo Kim continue; 1333bbe88753SJoonsoo Kim 1334a36f1e90SSidhartha Kumar if (folio_test_hwpoison(folio)) 13356664bfc8SWei Yang continue; 1336bbe88753SJoonsoo Kim 1337a36f1e90SSidhartha Kumar list_move(&folio->lru, &h->hugepage_activelist); 1338a36f1e90SSidhartha Kumar folio_ref_unfreeze(folio, 1); 1339a36f1e90SSidhartha Kumar folio_clear_hugetlb_freed(folio); 1340bf50bab2SNaoya Horiguchi h->free_huge_pages--; 1341bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 1342a36f1e90SSidhartha Kumar return folio; 1343bf50bab2SNaoya Horiguchi } 1344bf50bab2SNaoya Horiguchi 13456664bfc8SWei Yang return NULL; 13466664bfc8SWei Yang } 13476664bfc8SWei Yang 1348a36f1e90SSidhartha Kumar static struct folio *dequeue_hugetlb_folio_nodemask(struct hstate *h, gfp_t gfp_mask, 1349a36f1e90SSidhartha Kumar int nid, nodemask_t *nmask) 135094310cbcSAnshuman Khandual { 13513e59fcb0SMichal Hocko unsigned int cpuset_mems_cookie; 13523e59fcb0SMichal Hocko struct zonelist *zonelist; 13533e59fcb0SMichal Hocko struct zone *zone; 13543e59fcb0SMichal Hocko struct zoneref *z; 135598fa15f3SAnshuman Khandual int node = NUMA_NO_NODE; 13563e59fcb0SMichal Hocko 13573e59fcb0SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 13583e59fcb0SMichal Hocko 13593e59fcb0SMichal Hocko retry_cpuset: 13603e59fcb0SMichal Hocko cpuset_mems_cookie = read_mems_allowed_begin(); 13613e59fcb0SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 1362a36f1e90SSidhartha Kumar struct folio *folio; 136394310cbcSAnshuman Khandual 13643e59fcb0SMichal Hocko if (!cpuset_zone_allowed(zone, gfp_mask)) 13653e59fcb0SMichal Hocko continue; 13663e59fcb0SMichal Hocko /* 13673e59fcb0SMichal Hocko * no need to ask again on the same node. Pool is node rather than 13683e59fcb0SMichal Hocko * zone aware 13693e59fcb0SMichal Hocko */ 13703e59fcb0SMichal Hocko if (zone_to_nid(zone) == node) 13713e59fcb0SMichal Hocko continue; 13723e59fcb0SMichal Hocko node = zone_to_nid(zone); 137394310cbcSAnshuman Khandual 1374a36f1e90SSidhartha Kumar folio = dequeue_hugetlb_folio_node_exact(h, node); 1375a36f1e90SSidhartha Kumar if (folio) 1376a36f1e90SSidhartha Kumar return folio; 137794310cbcSAnshuman Khandual } 13783e59fcb0SMichal Hocko if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 13793e59fcb0SMichal Hocko goto retry_cpuset; 13803e59fcb0SMichal Hocko 138194310cbcSAnshuman Khandual return NULL; 138294310cbcSAnshuman Khandual } 138394310cbcSAnshuman Khandual 13848346d69dSXin Hao static unsigned long available_huge_pages(struct hstate *h) 13858346d69dSXin Hao { 13868346d69dSXin Hao return h->free_huge_pages - h->resv_huge_pages; 13878346d69dSXin Hao } 13888346d69dSXin Hao 1389ff7d853bSSidhartha Kumar static struct folio *dequeue_hugetlb_folio_vma(struct hstate *h, 1390a5516438SAndi Kleen struct vm_area_struct *vma, 1391af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 1392af0ed73eSJoonsoo Kim long chg) 13931da177e4SLinus Torvalds { 1394a36f1e90SSidhartha Kumar struct folio *folio = NULL; 1395480eccf9SLee Schermerhorn struct mempolicy *mpol; 139604ec6264SVlastimil Babka gfp_t gfp_mask; 13973e59fcb0SMichal Hocko nodemask_t *nodemask; 139804ec6264SVlastimil Babka int nid; 13991da177e4SLinus Torvalds 1400a1e78772SMel Gorman /* 1401a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 1402a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 1403a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 1404a1e78772SMel Gorman */ 14058346d69dSXin Hao if (!vma_has_reserves(vma, chg) && !available_huge_pages(h)) 1406c0ff7453SMiao Xie goto err; 1407a1e78772SMel Gorman 140804f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 14098346d69dSXin Hao if (avoid_reserve && !available_huge_pages(h)) 14106eab04a8SJustin P. Mattock goto err; 141104f2cbe3SMel Gorman 141204ec6264SVlastimil Babka gfp_mask = htlb_alloc_mask(h); 141304ec6264SVlastimil Babka nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 1414cfcaa66fSBen Widawsky 1415cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 1416a36f1e90SSidhartha Kumar folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask, 1417a36f1e90SSidhartha Kumar nid, nodemask); 1418cfcaa66fSBen Widawsky 1419cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 1420cfcaa66fSBen Widawsky nodemask = NULL; 1421cfcaa66fSBen Widawsky } 1422cfcaa66fSBen Widawsky 1423a36f1e90SSidhartha Kumar if (!folio) 1424a36f1e90SSidhartha Kumar folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask, 1425a36f1e90SSidhartha Kumar nid, nodemask); 1426cfcaa66fSBen Widawsky 1427a36f1e90SSidhartha Kumar if (folio && !avoid_reserve && vma_has_reserves(vma, chg)) { 1428a36f1e90SSidhartha Kumar folio_set_hugetlb_restore_reserve(folio); 1429a63884e9SJoonsoo Kim h->resv_huge_pages--; 1430bf50bab2SNaoya Horiguchi } 1431cc9a6c87SMel Gorman 1432cc9a6c87SMel Gorman mpol_cond_put(mpol); 1433ff7d853bSSidhartha Kumar return folio; 1434cc9a6c87SMel Gorman 1435c0ff7453SMiao Xie err: 1436cc9a6c87SMel Gorman return NULL; 14371da177e4SLinus Torvalds } 14381da177e4SLinus Torvalds 14391cac6f2cSLuiz Capitulino /* 14401cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 14411cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 14421cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 14431cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 14441cac6f2cSLuiz Capitulino * node for alloc or free. 14451cac6f2cSLuiz Capitulino */ 14461cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 14471cac6f2cSLuiz Capitulino { 14480edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 14491cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 14501cac6f2cSLuiz Capitulino 14511cac6f2cSLuiz Capitulino return nid; 14521cac6f2cSLuiz Capitulino } 14531cac6f2cSLuiz Capitulino 14541cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 14551cac6f2cSLuiz Capitulino { 14561cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 14571cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 14581cac6f2cSLuiz Capitulino return nid; 14591cac6f2cSLuiz Capitulino } 14601cac6f2cSLuiz Capitulino 14611cac6f2cSLuiz Capitulino /* 14621cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 14631cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 14641cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 14651cac6f2cSLuiz Capitulino * mask. 14661cac6f2cSLuiz Capitulino */ 14671cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 14681cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 14691cac6f2cSLuiz Capitulino { 14701cac6f2cSLuiz Capitulino int nid; 14711cac6f2cSLuiz Capitulino 14721cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 14731cac6f2cSLuiz Capitulino 14741cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 14751cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 14761cac6f2cSLuiz Capitulino 14771cac6f2cSLuiz Capitulino return nid; 14781cac6f2cSLuiz Capitulino } 14791cac6f2cSLuiz Capitulino 14801cac6f2cSLuiz Capitulino /* 148110c6ec49SMike Kravetz * helper for remove_pool_huge_page() - return the previously saved 14821cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 14831cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 14841cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 14851cac6f2cSLuiz Capitulino */ 14861cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 14871cac6f2cSLuiz Capitulino { 14881cac6f2cSLuiz Capitulino int nid; 14891cac6f2cSLuiz Capitulino 14901cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 14911cac6f2cSLuiz Capitulino 14921cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 14931cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 14941cac6f2cSLuiz Capitulino 14951cac6f2cSLuiz Capitulino return nid; 14961cac6f2cSLuiz Capitulino } 14971cac6f2cSLuiz Capitulino 14981cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 14991cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 15001cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 15011cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 15021cac6f2cSLuiz Capitulino nr_nodes--) 15031cac6f2cSLuiz Capitulino 15041cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 15051cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 15061cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 15071cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 15081cac6f2cSLuiz Capitulino nr_nodes--) 15091cac6f2cSLuiz Capitulino 15108531fc6fSMike Kravetz /* used to demote non-gigantic_huge pages as well */ 1511911565b8SSidhartha Kumar static void __destroy_compound_gigantic_folio(struct folio *folio, 151234d9e35bSMike Kravetz unsigned int order, bool demote) 1513944d9fecSLuiz Capitulino { 1514944d9fecSLuiz Capitulino int i; 1515944d9fecSLuiz Capitulino int nr_pages = 1 << order; 151614455eabSCheng Li struct page *p; 1517944d9fecSLuiz Capitulino 151846f27228SMatthew Wilcox (Oracle) atomic_set(&folio->_entire_mapcount, 0); 1519eec20426SMatthew Wilcox (Oracle) atomic_set(&folio->_nr_pages_mapped, 0); 152094688e8eSMatthew Wilcox (Oracle) atomic_set(&folio->_pincount, 0); 152147e29d32SJohn Hubbard 152214455eabSCheng Li for (i = 1; i < nr_pages; i++) { 1523911565b8SSidhartha Kumar p = folio_page(folio, i); 15246c141973SMike Kravetz p->flags &= ~PAGE_FLAGS_CHECK_AT_FREE; 1525a01f4390SMike Kravetz p->mapping = NULL; 15261d798ca3SKirill A. Shutemov clear_compound_head(p); 152734d9e35bSMike Kravetz if (!demote) 1528944d9fecSLuiz Capitulino set_page_refcounted(p); 1529944d9fecSLuiz Capitulino } 1530944d9fecSLuiz Capitulino 1531911565b8SSidhartha Kumar __folio_clear_head(folio); 1532944d9fecSLuiz Capitulino } 1533944d9fecSLuiz Capitulino 1534911565b8SSidhartha Kumar static void destroy_compound_hugetlb_folio_for_demote(struct folio *folio, 15358531fc6fSMike Kravetz unsigned int order) 15368531fc6fSMike Kravetz { 1537911565b8SSidhartha Kumar __destroy_compound_gigantic_folio(folio, order, true); 15388531fc6fSMike Kravetz } 15398531fc6fSMike Kravetz 15408531fc6fSMike Kravetz #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 1541911565b8SSidhartha Kumar static void destroy_compound_gigantic_folio(struct folio *folio, 154234d9e35bSMike Kravetz unsigned int order) 154334d9e35bSMike Kravetz { 1544911565b8SSidhartha Kumar __destroy_compound_gigantic_folio(folio, order, false); 154534d9e35bSMike Kravetz } 154634d9e35bSMike Kravetz 15477f325a8dSSidhartha Kumar static void free_gigantic_folio(struct folio *folio, unsigned int order) 1548944d9fecSLuiz Capitulino { 1549cf11e85fSRoman Gushchin /* 1550cf11e85fSRoman Gushchin * If the page isn't allocated using the cma allocator, 1551cf11e85fSRoman Gushchin * cma_release() returns false. 1552cf11e85fSRoman Gushchin */ 1553dbda8feaSBarry Song #ifdef CONFIG_CMA 15547f325a8dSSidhartha Kumar int nid = folio_nid(folio); 15557f325a8dSSidhartha Kumar 15567f325a8dSSidhartha Kumar if (cma_release(hugetlb_cma[nid], &folio->page, 1 << order)) 1557cf11e85fSRoman Gushchin return; 1558dbda8feaSBarry Song #endif 1559cf11e85fSRoman Gushchin 15607f325a8dSSidhartha Kumar free_contig_range(folio_pfn(folio), 1 << order); 1561944d9fecSLuiz Capitulino } 1562944d9fecSLuiz Capitulino 15634eb0716eSAlexandre Ghiti #ifdef CONFIG_CONTIG_ALLOC 156419fc1a7eSSidhartha Kumar static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask, 1565d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) 1566944d9fecSLuiz Capitulino { 156719fc1a7eSSidhartha Kumar struct page *page; 156804adbc3fSMiaohe Lin unsigned long nr_pages = pages_per_huge_page(h); 1569953f064aSLi Xinhai if (nid == NUMA_NO_NODE) 1570953f064aSLi Xinhai nid = numa_mem_id(); 1571944d9fecSLuiz Capitulino 1572dbda8feaSBarry Song #ifdef CONFIG_CMA 1573dbda8feaSBarry Song { 1574cf11e85fSRoman Gushchin int node; 1575cf11e85fSRoman Gushchin 1576953f064aSLi Xinhai if (hugetlb_cma[nid]) { 1577953f064aSLi Xinhai page = cma_alloc(hugetlb_cma[nid], nr_pages, 1578953f064aSLi Xinhai huge_page_order(h), true); 1579953f064aSLi Xinhai if (page) 158019fc1a7eSSidhartha Kumar return page_folio(page); 1581953f064aSLi Xinhai } 1582953f064aSLi Xinhai 1583953f064aSLi Xinhai if (!(gfp_mask & __GFP_THISNODE)) { 1584cf11e85fSRoman Gushchin for_each_node_mask(node, *nodemask) { 1585953f064aSLi Xinhai if (node == nid || !hugetlb_cma[node]) 1586cf11e85fSRoman Gushchin continue; 1587cf11e85fSRoman Gushchin 1588cf11e85fSRoman Gushchin page = cma_alloc(hugetlb_cma[node], nr_pages, 1589cf11e85fSRoman Gushchin huge_page_order(h), true); 1590cf11e85fSRoman Gushchin if (page) 159119fc1a7eSSidhartha Kumar return page_folio(page); 1592cf11e85fSRoman Gushchin } 1593cf11e85fSRoman Gushchin } 1594953f064aSLi Xinhai } 1595dbda8feaSBarry Song #endif 1596cf11e85fSRoman Gushchin 159719fc1a7eSSidhartha Kumar page = alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); 159819fc1a7eSSidhartha Kumar return page ? page_folio(page) : NULL; 1599944d9fecSLuiz Capitulino } 1600944d9fecSLuiz Capitulino 16014eb0716eSAlexandre Ghiti #else /* !CONFIG_CONTIG_ALLOC */ 160219fc1a7eSSidhartha Kumar static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask, 16034eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 16044eb0716eSAlexandre Ghiti { 16054eb0716eSAlexandre Ghiti return NULL; 16064eb0716eSAlexandre Ghiti } 16074eb0716eSAlexandre Ghiti #endif /* CONFIG_CONTIG_ALLOC */ 1608944d9fecSLuiz Capitulino 1609e1073d1eSAneesh Kumar K.V #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 161019fc1a7eSSidhartha Kumar static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask, 16114eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 16124eb0716eSAlexandre Ghiti { 16134eb0716eSAlexandre Ghiti return NULL; 16144eb0716eSAlexandre Ghiti } 16157f325a8dSSidhartha Kumar static inline void free_gigantic_folio(struct folio *folio, 16167f325a8dSSidhartha Kumar unsigned int order) { } 1617911565b8SSidhartha Kumar static inline void destroy_compound_gigantic_folio(struct folio *folio, 1618d00181b9SKirill A. Shutemov unsigned int order) { } 1619944d9fecSLuiz Capitulino #endif 1620944d9fecSLuiz Capitulino 162132c87719SMike Kravetz static inline void __clear_hugetlb_destructor(struct hstate *h, 162232c87719SMike Kravetz struct folio *folio) 162332c87719SMike Kravetz { 162432c87719SMike Kravetz lockdep_assert_held(&hugetlb_lock); 162532c87719SMike Kravetz 16269c5ccf2dSMatthew Wilcox (Oracle) folio_clear_hugetlb(folio); 162732c87719SMike Kravetz } 162832c87719SMike Kravetz 162932c87719SMike Kravetz /* 163032c87719SMike Kravetz * Remove hugetlb folio from lists. 163132c87719SMike Kravetz * If vmemmap exists for the folio, update dtor so that the folio appears 163232c87719SMike Kravetz * as just a compound page. Otherwise, wait until after allocating vmemmap 163332c87719SMike Kravetz * to update dtor. 163434d9e35bSMike Kravetz * 1635cfd5082bSSidhartha Kumar * A reference is held on the folio, except in the case of demote. 16366eb4e88aSMike Kravetz * 16376eb4e88aSMike Kravetz * Must be called with hugetlb lock held. 16386eb4e88aSMike Kravetz */ 1639cfd5082bSSidhartha Kumar static void __remove_hugetlb_folio(struct hstate *h, struct folio *folio, 164034d9e35bSMike Kravetz bool adjust_surplus, 164134d9e35bSMike Kravetz bool demote) 16426eb4e88aSMike Kravetz { 1643cfd5082bSSidhartha Kumar int nid = folio_nid(folio); 16446eb4e88aSMike Kravetz 1645f074732dSSidhartha Kumar VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio(folio), folio); 1646f074732dSSidhartha Kumar VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio_rsvd(folio), folio); 16476eb4e88aSMike Kravetz 16489487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 16496eb4e88aSMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 16506eb4e88aSMike Kravetz return; 16516eb4e88aSMike Kravetz 1652cfd5082bSSidhartha Kumar list_del(&folio->lru); 16536eb4e88aSMike Kravetz 1654cfd5082bSSidhartha Kumar if (folio_test_hugetlb_freed(folio)) { 16556eb4e88aSMike Kravetz h->free_huge_pages--; 16566eb4e88aSMike Kravetz h->free_huge_pages_node[nid]--; 16576eb4e88aSMike Kravetz } 16586eb4e88aSMike Kravetz if (adjust_surplus) { 16596eb4e88aSMike Kravetz h->surplus_huge_pages--; 16606eb4e88aSMike Kravetz h->surplus_huge_pages_node[nid]--; 16616eb4e88aSMike Kravetz } 16626eb4e88aSMike Kravetz 1663e32d20c0SMike Kravetz /* 166432c87719SMike Kravetz * We can only clear the hugetlb destructor after allocating vmemmap 166532c87719SMike Kravetz * pages. Otherwise, someone (memory error handling) may try to write 166632c87719SMike Kravetz * to tail struct pages. 166732c87719SMike Kravetz */ 166832c87719SMike Kravetz if (!folio_test_hugetlb_vmemmap_optimized(folio)) 166932c87719SMike Kravetz __clear_hugetlb_destructor(h, folio); 167032c87719SMike Kravetz 167132c87719SMike Kravetz /* 167234d9e35bSMike Kravetz * In the case of demote we do not ref count the page as it will soon 167334d9e35bSMike Kravetz * be turned into a page of smaller size. 1674e32d20c0SMike Kravetz */ 167534d9e35bSMike Kravetz if (!demote) 1676cfd5082bSSidhartha Kumar folio_ref_unfreeze(folio, 1); 16776eb4e88aSMike Kravetz 16786eb4e88aSMike Kravetz h->nr_huge_pages--; 16796eb4e88aSMike Kravetz h->nr_huge_pages_node[nid]--; 16806eb4e88aSMike Kravetz } 16816eb4e88aSMike Kravetz 1682cfd5082bSSidhartha Kumar static void remove_hugetlb_folio(struct hstate *h, struct folio *folio, 168334d9e35bSMike Kravetz bool adjust_surplus) 168434d9e35bSMike Kravetz { 1685cfd5082bSSidhartha Kumar __remove_hugetlb_folio(h, folio, adjust_surplus, false); 168634d9e35bSMike Kravetz } 168734d9e35bSMike Kravetz 1688cfd5082bSSidhartha Kumar static void remove_hugetlb_folio_for_demote(struct hstate *h, struct folio *folio, 16898531fc6fSMike Kravetz bool adjust_surplus) 16908531fc6fSMike Kravetz { 1691cfd5082bSSidhartha Kumar __remove_hugetlb_folio(h, folio, adjust_surplus, true); 16928531fc6fSMike Kravetz } 16938531fc6fSMike Kravetz 16942f6c57d6SSidhartha Kumar static void add_hugetlb_folio(struct hstate *h, struct folio *folio, 1695ad2fa371SMuchun Song bool adjust_surplus) 1696ad2fa371SMuchun Song { 1697ad2fa371SMuchun Song int zeroed; 16982f6c57d6SSidhartha Kumar int nid = folio_nid(folio); 1699ad2fa371SMuchun Song 17002f6c57d6SSidhartha Kumar VM_BUG_ON_FOLIO(!folio_test_hugetlb_vmemmap_optimized(folio), folio); 1701ad2fa371SMuchun Song 1702ad2fa371SMuchun Song lockdep_assert_held(&hugetlb_lock); 1703ad2fa371SMuchun Song 17042f6c57d6SSidhartha Kumar INIT_LIST_HEAD(&folio->lru); 1705ad2fa371SMuchun Song h->nr_huge_pages++; 1706ad2fa371SMuchun Song h->nr_huge_pages_node[nid]++; 1707ad2fa371SMuchun Song 1708ad2fa371SMuchun Song if (adjust_surplus) { 1709ad2fa371SMuchun Song h->surplus_huge_pages++; 1710ad2fa371SMuchun Song h->surplus_huge_pages_node[nid]++; 1711ad2fa371SMuchun Song } 1712ad2fa371SMuchun Song 17139c5ccf2dSMatthew Wilcox (Oracle) folio_set_hugetlb(folio); 17142f6c57d6SSidhartha Kumar folio_change_private(folio, NULL); 1715a9e1eab2SMiaohe Lin /* 17162f6c57d6SSidhartha Kumar * We have to set hugetlb_vmemmap_optimized again as above 17172f6c57d6SSidhartha Kumar * folio_change_private(folio, NULL) cleared it. 1718a9e1eab2SMiaohe Lin */ 17192f6c57d6SSidhartha Kumar folio_set_hugetlb_vmemmap_optimized(folio); 1720ad2fa371SMuchun Song 1721ad2fa371SMuchun Song /* 17222f6c57d6SSidhartha Kumar * This folio is about to be managed by the hugetlb allocator and 1723b65a4edaSMike Kravetz * should have no users. Drop our reference, and check for others 1724b65a4edaSMike Kravetz * just in case. 1725ad2fa371SMuchun Song */ 17262f6c57d6SSidhartha Kumar zeroed = folio_put_testzero(folio); 17272f6c57d6SSidhartha Kumar if (unlikely(!zeroed)) 1728b65a4edaSMike Kravetz /* 1729454a00c4SMatthew Wilcox (Oracle) * It is VERY unlikely soneone else has taken a ref 1730454a00c4SMatthew Wilcox (Oracle) * on the folio. In this case, we simply return as 1731454a00c4SMatthew Wilcox (Oracle) * free_huge_folio() will be called when this other ref 1732454a00c4SMatthew Wilcox (Oracle) * is dropped. 1733b65a4edaSMike Kravetz */ 1734b65a4edaSMike Kravetz return; 1735b65a4edaSMike Kravetz 17362f6c57d6SSidhartha Kumar arch_clear_hugepage_flags(&folio->page); 1737240d67a8SSidhartha Kumar enqueue_hugetlb_folio(h, folio); 1738ad2fa371SMuchun Song } 1739ad2fa371SMuchun Song 17406f6956cfSSidhartha Kumar static void __update_and_free_hugetlb_folio(struct hstate *h, 17416f6956cfSSidhartha Kumar struct folio *folio) 17426af2acb6SAdam Litke { 174332c87719SMike Kravetz bool clear_dtor = folio_test_hugetlb_vmemmap_optimized(folio); 1744a5516438SAndi Kleen 17454eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1746944d9fecSLuiz Capitulino return; 174718229df5SAndy Whitcroft 1748161df60eSNaoya Horiguchi /* 1749161df60eSNaoya Horiguchi * If we don't know which subpages are hwpoisoned, we can't free 1750161df60eSNaoya Horiguchi * the hugepage, so it's leaked intentionally. 1751161df60eSNaoya Horiguchi */ 17527f325a8dSSidhartha Kumar if (folio_test_hugetlb_raw_hwp_unreliable(folio)) 1753161df60eSNaoya Horiguchi return; 1754161df60eSNaoya Horiguchi 17556f6956cfSSidhartha Kumar if (hugetlb_vmemmap_restore(h, &folio->page)) { 1756ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 1757ad2fa371SMuchun Song /* 1758ad2fa371SMuchun Song * If we cannot allocate vmemmap pages, just refuse to free the 1759ad2fa371SMuchun Song * page and put the page back on the hugetlb free list and treat 1760ad2fa371SMuchun Song * as a surplus page. 1761ad2fa371SMuchun Song */ 17627f325a8dSSidhartha Kumar add_hugetlb_folio(h, folio, true); 1763ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 1764ad2fa371SMuchun Song return; 1765ad2fa371SMuchun Song } 1766ad2fa371SMuchun Song 1767161df60eSNaoya Horiguchi /* 1768161df60eSNaoya Horiguchi * Move PageHWPoison flag from head page to the raw error pages, 1769161df60eSNaoya Horiguchi * which makes any healthy subpages reusable. 1770161df60eSNaoya Horiguchi */ 1771911565b8SSidhartha Kumar if (unlikely(folio_test_hwpoison(folio))) 17722ff6ceceSSidhartha Kumar folio_clear_hugetlb_hwpoison(folio); 1773161df60eSNaoya Horiguchi 177432c87719SMike Kravetz /* 177532c87719SMike Kravetz * If vmemmap pages were allocated above, then we need to clear the 177632c87719SMike Kravetz * hugetlb destructor under the hugetlb lock. 177732c87719SMike Kravetz */ 177832c87719SMike Kravetz if (clear_dtor) { 177932c87719SMike Kravetz spin_lock_irq(&hugetlb_lock); 178032c87719SMike Kravetz __clear_hugetlb_destructor(h, folio); 178132c87719SMike Kravetz spin_unlock_irq(&hugetlb_lock); 178232c87719SMike Kravetz } 178332c87719SMike Kravetz 1784a01f4390SMike Kravetz /* 1785a01f4390SMike Kravetz * Non-gigantic pages demoted from CMA allocated gigantic pages 17867f325a8dSSidhartha Kumar * need to be given back to CMA in free_gigantic_folio. 1787a01f4390SMike Kravetz */ 1788a01f4390SMike Kravetz if (hstate_is_gigantic(h) || 17892f6c57d6SSidhartha Kumar hugetlb_cma_folio(folio, huge_page_order(h))) { 1790911565b8SSidhartha Kumar destroy_compound_gigantic_folio(folio, huge_page_order(h)); 17917f325a8dSSidhartha Kumar free_gigantic_folio(folio, huge_page_order(h)); 1792944d9fecSLuiz Capitulino } else { 17936f6956cfSSidhartha Kumar __free_pages(&folio->page, huge_page_order(h)); 17946af2acb6SAdam Litke } 1795944d9fecSLuiz Capitulino } 17966af2acb6SAdam Litke 1797b65d4adbSMuchun Song /* 1798d6ef19e2SSidhartha Kumar * As update_and_free_hugetlb_folio() can be called under any context, so we cannot 1799b65d4adbSMuchun Song * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the 1800b65d4adbSMuchun Song * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate 1801b65d4adbSMuchun Song * the vmemmap pages. 1802b65d4adbSMuchun Song * 1803b65d4adbSMuchun Song * free_hpage_workfn() locklessly retrieves the linked list of pages to be 1804b65d4adbSMuchun Song * freed and frees them one-by-one. As the page->mapping pointer is going 1805b65d4adbSMuchun Song * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node 1806b65d4adbSMuchun Song * structure of a lockless linked list of huge pages to be freed. 1807b65d4adbSMuchun Song */ 1808b65d4adbSMuchun Song static LLIST_HEAD(hpage_freelist); 1809b65d4adbSMuchun Song 1810b65d4adbSMuchun Song static void free_hpage_workfn(struct work_struct *work) 1811b65d4adbSMuchun Song { 1812b65d4adbSMuchun Song struct llist_node *node; 1813b65d4adbSMuchun Song 1814b65d4adbSMuchun Song node = llist_del_all(&hpage_freelist); 1815b65d4adbSMuchun Song 1816b65d4adbSMuchun Song while (node) { 1817b65d4adbSMuchun Song struct page *page; 1818b65d4adbSMuchun Song struct hstate *h; 1819b65d4adbSMuchun Song 1820b65d4adbSMuchun Song page = container_of((struct address_space **)node, 1821b65d4adbSMuchun Song struct page, mapping); 1822b65d4adbSMuchun Song node = node->next; 1823b65d4adbSMuchun Song page->mapping = NULL; 1824b65d4adbSMuchun Song /* 1825affd26b1SSidhartha Kumar * The VM_BUG_ON_FOLIO(!folio_test_hugetlb(folio), folio) in 1826affd26b1SSidhartha Kumar * folio_hstate() is going to trigger because a previous call to 18279c5ccf2dSMatthew Wilcox (Oracle) * remove_hugetlb_folio() will clear the hugetlb bit, so do 18289c5ccf2dSMatthew Wilcox (Oracle) * not use folio_hstate() directly. 1829b65d4adbSMuchun Song */ 1830b65d4adbSMuchun Song h = size_to_hstate(page_size(page)); 1831b65d4adbSMuchun Song 18326f6956cfSSidhartha Kumar __update_and_free_hugetlb_folio(h, page_folio(page)); 1833b65d4adbSMuchun Song 1834b65d4adbSMuchun Song cond_resched(); 1835b65d4adbSMuchun Song } 1836b65d4adbSMuchun Song } 1837b65d4adbSMuchun Song static DECLARE_WORK(free_hpage_work, free_hpage_workfn); 1838b65d4adbSMuchun Song 1839b65d4adbSMuchun Song static inline void flush_free_hpage_work(struct hstate *h) 1840b65d4adbSMuchun Song { 18416213834cSMuchun Song if (hugetlb_vmemmap_optimizable(h)) 1842b65d4adbSMuchun Song flush_work(&free_hpage_work); 1843b65d4adbSMuchun Song } 1844b65d4adbSMuchun Song 1845d6ef19e2SSidhartha Kumar static void update_and_free_hugetlb_folio(struct hstate *h, struct folio *folio, 1846b65d4adbSMuchun Song bool atomic) 1847b65d4adbSMuchun Song { 1848d6ef19e2SSidhartha Kumar if (!folio_test_hugetlb_vmemmap_optimized(folio) || !atomic) { 18496f6956cfSSidhartha Kumar __update_and_free_hugetlb_folio(h, folio); 1850b65d4adbSMuchun Song return; 1851b65d4adbSMuchun Song } 1852b65d4adbSMuchun Song 1853b65d4adbSMuchun Song /* 1854b65d4adbSMuchun Song * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages. 1855b65d4adbSMuchun Song * 1856b65d4adbSMuchun Song * Only call schedule_work() if hpage_freelist is previously 1857b65d4adbSMuchun Song * empty. Otherwise, schedule_work() had been called but the workfn 1858b65d4adbSMuchun Song * hasn't retrieved the list yet. 1859b65d4adbSMuchun Song */ 1860d6ef19e2SSidhartha Kumar if (llist_add((struct llist_node *)&folio->mapping, &hpage_freelist)) 1861b65d4adbSMuchun Song schedule_work(&free_hpage_work); 1862b65d4adbSMuchun Song } 1863b65d4adbSMuchun Song 186410c6ec49SMike Kravetz static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list) 186510c6ec49SMike Kravetz { 186610c6ec49SMike Kravetz struct page *page, *t_page; 1867d6ef19e2SSidhartha Kumar struct folio *folio; 186810c6ec49SMike Kravetz 186910c6ec49SMike Kravetz list_for_each_entry_safe(page, t_page, list, lru) { 1870d6ef19e2SSidhartha Kumar folio = page_folio(page); 1871d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, folio, false); 187210c6ec49SMike Kravetz cond_resched(); 187310c6ec49SMike Kravetz } 187410c6ec49SMike Kravetz } 187510c6ec49SMike Kravetz 1876e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1877e5ff2159SAndi Kleen { 1878e5ff2159SAndi Kleen struct hstate *h; 1879e5ff2159SAndi Kleen 1880e5ff2159SAndi Kleen for_each_hstate(h) { 1881e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1882e5ff2159SAndi Kleen return h; 1883e5ff2159SAndi Kleen } 1884e5ff2159SAndi Kleen return NULL; 1885e5ff2159SAndi Kleen } 1886e5ff2159SAndi Kleen 1887454a00c4SMatthew Wilcox (Oracle) void free_huge_folio(struct folio *folio) 188827a85ef1SDavid Gibson { 1889a5516438SAndi Kleen /* 1890a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1891a5516438SAndi Kleen * compound page destructor. 1892a5516438SAndi Kleen */ 18930356c4b9SSidhartha Kumar struct hstate *h = folio_hstate(folio); 18940356c4b9SSidhartha Kumar int nid = folio_nid(folio); 18950356c4b9SSidhartha Kumar struct hugepage_subpool *spool = hugetlb_folio_subpool(folio); 189607443a85SJoonsoo Kim bool restore_reserve; 1897db71ef79SMike Kravetz unsigned long flags; 189827a85ef1SDavid Gibson 18990356c4b9SSidhartha Kumar VM_BUG_ON_FOLIO(folio_ref_count(folio), folio); 19000356c4b9SSidhartha Kumar VM_BUG_ON_FOLIO(folio_mapcount(folio), folio); 19018ace22bcSYongkai Wu 19020356c4b9SSidhartha Kumar hugetlb_set_folio_subpool(folio, NULL); 19030356c4b9SSidhartha Kumar if (folio_test_anon(folio)) 19040356c4b9SSidhartha Kumar __ClearPageAnonExclusive(&folio->page); 19050356c4b9SSidhartha Kumar folio->mapping = NULL; 19060356c4b9SSidhartha Kumar restore_reserve = folio_test_hugetlb_restore_reserve(folio); 19070356c4b9SSidhartha Kumar folio_clear_hugetlb_restore_reserve(folio); 190827a85ef1SDavid Gibson 19091c5ecae3SMike Kravetz /* 1910d6995da3SMike Kravetz * If HPageRestoreReserve was set on page, page allocation consumed a 19110919e1b6SMike Kravetz * reservation. If the page was associated with a subpool, there 19120919e1b6SMike Kravetz * would have been a page reserved in the subpool before allocation 19130919e1b6SMike Kravetz * via hugepage_subpool_get_pages(). Since we are 'restoring' the 19146c26d310SMiaohe Lin * reservation, do not call hugepage_subpool_put_pages() as this will 19150919e1b6SMike Kravetz * remove the reserved page from the subpool. 19160919e1b6SMike Kravetz */ 19170919e1b6SMike Kravetz if (!restore_reserve) { 19180919e1b6SMike Kravetz /* 19190919e1b6SMike Kravetz * A return code of zero implies that the subpool will be 19200919e1b6SMike Kravetz * under its minimum size if the reservation is not restored 19210919e1b6SMike Kravetz * after page is free. Therefore, force restore_reserve 19220919e1b6SMike Kravetz * operation. 19231c5ecae3SMike Kravetz */ 19241c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 19251c5ecae3SMike Kravetz restore_reserve = true; 19260919e1b6SMike Kravetz } 19271c5ecae3SMike Kravetz 1928db71ef79SMike Kravetz spin_lock_irqsave(&hugetlb_lock, flags); 19290356c4b9SSidhartha Kumar folio_clear_hugetlb_migratable(folio); 1930d4ab0316SSidhartha Kumar hugetlb_cgroup_uncharge_folio(hstate_index(h), 1931d4ab0316SSidhartha Kumar pages_per_huge_page(h), folio); 1932d4ab0316SSidhartha Kumar hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h), 1933d4ab0316SSidhartha Kumar pages_per_huge_page(h), folio); 193407443a85SJoonsoo Kim if (restore_reserve) 193507443a85SJoonsoo Kim h->resv_huge_pages++; 193607443a85SJoonsoo Kim 19370356c4b9SSidhartha Kumar if (folio_test_hugetlb_temporary(folio)) { 1938cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, false); 1939db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1940d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, folio, true); 1941ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 19420edaecfaSAneesh Kumar K.V /* remove the page from active list */ 1943cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, true); 1944db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1945d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, folio, true); 19467893d1d5SAdam Litke } else { 1947454a00c4SMatthew Wilcox (Oracle) arch_clear_hugepage_flags(&folio->page); 1948240d67a8SSidhartha Kumar enqueue_hugetlb_folio(h, folio); 1949db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 195027a85ef1SDavid Gibson } 19511121828aSMike Kravetz } 195227a85ef1SDavid Gibson 1953d3d99fccSOscar Salvador /* 1954d3d99fccSOscar Salvador * Must be called with the hugetlb lock held 1955d3d99fccSOscar Salvador */ 1956d3d99fccSOscar Salvador static void __prep_account_new_huge_page(struct hstate *h, int nid) 1957d3d99fccSOscar Salvador { 1958d3d99fccSOscar Salvador lockdep_assert_held(&hugetlb_lock); 1959d3d99fccSOscar Salvador h->nr_huge_pages++; 1960d3d99fccSOscar Salvador h->nr_huge_pages_node[nid]++; 1961d3d99fccSOscar Salvador } 1962d3d99fccSOscar Salvador 1963de656ed3SSidhartha Kumar static void __prep_new_hugetlb_folio(struct hstate *h, struct folio *folio) 1964b7ba30c6SAndi Kleen { 1965de656ed3SSidhartha Kumar hugetlb_vmemmap_optimize(h, &folio->page); 1966de656ed3SSidhartha Kumar INIT_LIST_HEAD(&folio->lru); 19679c5ccf2dSMatthew Wilcox (Oracle) folio_set_hugetlb(folio); 1968de656ed3SSidhartha Kumar hugetlb_set_folio_subpool(folio, NULL); 1969de656ed3SSidhartha Kumar set_hugetlb_cgroup(folio, NULL); 1970de656ed3SSidhartha Kumar set_hugetlb_cgroup_rsvd(folio, NULL); 1971d3d99fccSOscar Salvador } 1972d3d99fccSOscar Salvador 1973d1c60955SSidhartha Kumar static void prep_new_hugetlb_folio(struct hstate *h, struct folio *folio, int nid) 1974d3d99fccSOscar Salvador { 1975de656ed3SSidhartha Kumar __prep_new_hugetlb_folio(h, folio); 1976db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 1977d3d99fccSOscar Salvador __prep_account_new_huge_page(h, nid); 1978db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1979b7ba30c6SAndi Kleen } 1980b7ba30c6SAndi Kleen 1981d1c60955SSidhartha Kumar static bool __prep_compound_gigantic_folio(struct folio *folio, 1982d1c60955SSidhartha Kumar unsigned int order, bool demote) 198320a0307cSWu Fengguang { 19847118fc29SMike Kravetz int i, j; 198520a0307cSWu Fengguang int nr_pages = 1 << order; 198614455eabSCheng Li struct page *p; 198720a0307cSWu Fengguang 1988d1c60955SSidhartha Kumar __folio_clear_reserved(folio); 19892b21624fSMike Kravetz for (i = 0; i < nr_pages; i++) { 1990d1c60955SSidhartha Kumar p = folio_page(folio, i); 199114455eabSCheng Li 1992ef5a22beSAndrea Arcangeli /* 1993ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1994ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1995ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 19967c8de358SEthon Paul * too. Otherwise drivers using get_user_pages() to access tail 1997ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1998ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1999ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 2000ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 2001ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 2002ef5a22beSAndrea Arcangeli * after get_user_pages(). 2003ef5a22beSAndrea Arcangeli */ 20047fb0728aSMike Kravetz if (i != 0) /* head page cleared above */ 2005ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 20067118fc29SMike Kravetz /* 20077118fc29SMike Kravetz * Subtle and very unlikely 20087118fc29SMike Kravetz * 20097118fc29SMike Kravetz * Gigantic 'page allocators' such as memblock or cma will 20107118fc29SMike Kravetz * return a set of pages with each page ref counted. We need 20117118fc29SMike Kravetz * to turn this set of pages into a compound page with tail 20127118fc29SMike Kravetz * page ref counts set to zero. Code such as speculative page 20137118fc29SMike Kravetz * cache adding could take a ref on a 'to be' tail page. 20147118fc29SMike Kravetz * We need to respect any increased ref count, and only set 20157118fc29SMike Kravetz * the ref count to zero if count is currently 1. If count 2016416d85edSMike Kravetz * is not 1, we return an error. An error return indicates 2017416d85edSMike Kravetz * the set of pages can not be converted to a gigantic page. 2018416d85edSMike Kravetz * The caller who allocated the pages should then discard the 2019416d85edSMike Kravetz * pages using the appropriate free interface. 202034d9e35bSMike Kravetz * 202134d9e35bSMike Kravetz * In the case of demote, the ref count will be zero. 20227118fc29SMike Kravetz */ 202334d9e35bSMike Kravetz if (!demote) { 20247118fc29SMike Kravetz if (!page_ref_freeze(p, 1)) { 2025416d85edSMike Kravetz pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n"); 20267118fc29SMike Kravetz goto out_error; 20277118fc29SMike Kravetz } 202834d9e35bSMike Kravetz } else { 202934d9e35bSMike Kravetz VM_BUG_ON_PAGE(page_count(p), p); 203034d9e35bSMike Kravetz } 20312b21624fSMike Kravetz if (i != 0) 2032d1c60955SSidhartha Kumar set_compound_head(p, &folio->page); 203320a0307cSWu Fengguang } 2034e3b7bf97STarun Sahu __folio_set_head(folio); 2035e3b7bf97STarun Sahu /* we rely on prep_new_hugetlb_folio to set the destructor */ 2036e3b7bf97STarun Sahu folio_set_order(folio, order); 203746f27228SMatthew Wilcox (Oracle) atomic_set(&folio->_entire_mapcount, -1); 2038eec20426SMatthew Wilcox (Oracle) atomic_set(&folio->_nr_pages_mapped, 0); 203994688e8eSMatthew Wilcox (Oracle) atomic_set(&folio->_pincount, 0); 20407118fc29SMike Kravetz return true; 20417118fc29SMike Kravetz 20427118fc29SMike Kravetz out_error: 20432b21624fSMike Kravetz /* undo page modifications made above */ 20442b21624fSMike Kravetz for (j = 0; j < i; j++) { 2045d1c60955SSidhartha Kumar p = folio_page(folio, j); 20462b21624fSMike Kravetz if (j != 0) 20477118fc29SMike Kravetz clear_compound_head(p); 20487118fc29SMike Kravetz set_page_refcounted(p); 20497118fc29SMike Kravetz } 20507118fc29SMike Kravetz /* need to clear PG_reserved on remaining tail pages */ 205114455eabSCheng Li for (; j < nr_pages; j++) { 2052d1c60955SSidhartha Kumar p = folio_page(folio, j); 20537118fc29SMike Kravetz __ClearPageReserved(p); 205414455eabSCheng Li } 20557118fc29SMike Kravetz return false; 205620a0307cSWu Fengguang } 205720a0307cSWu Fengguang 2058d1c60955SSidhartha Kumar static bool prep_compound_gigantic_folio(struct folio *folio, 20598531fc6fSMike Kravetz unsigned int order) 20608531fc6fSMike Kravetz { 2061d1c60955SSidhartha Kumar return __prep_compound_gigantic_folio(folio, order, false); 2062d1c60955SSidhartha Kumar } 2063d1c60955SSidhartha Kumar 2064d1c60955SSidhartha Kumar static bool prep_compound_gigantic_folio_for_demote(struct folio *folio, 2065d1c60955SSidhartha Kumar unsigned int order) 2066d1c60955SSidhartha Kumar { 2067d1c60955SSidhartha Kumar return __prep_compound_gigantic_folio(folio, order, true); 20688531fc6fSMike Kravetz } 20698531fc6fSMike Kravetz 20707795912cSAndrew Morton /* 20717795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 20727795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 20737795912cSAndrew Morton * details. 20747795912cSAndrew Morton */ 207520a0307cSWu Fengguang int PageHuge(struct page *page) 207620a0307cSWu Fengguang { 20772d678c64SMatthew Wilcox (Oracle) struct folio *folio; 20782d678c64SMatthew Wilcox (Oracle) 207920a0307cSWu Fengguang if (!PageCompound(page)) 208020a0307cSWu Fengguang return 0; 20812d678c64SMatthew Wilcox (Oracle) folio = page_folio(page); 20829c5ccf2dSMatthew Wilcox (Oracle) return folio_test_hugetlb(folio); 208320a0307cSWu Fengguang } 208443131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 208543131e14SNaoya Horiguchi 2086c0d0381aSMike Kravetz /* 2087c0d0381aSMike Kravetz * Find and lock address space (mapping) in write mode. 2088c0d0381aSMike Kravetz * 2089336bf30eSMike Kravetz * Upon entry, the page is locked which means that page_mapping() is 2090336bf30eSMike Kravetz * stable. Due to locking order, we can only trylock_write. If we can 2091336bf30eSMike Kravetz * not get the lock, simply return NULL to caller. 2092c0d0381aSMike Kravetz */ 2093c0d0381aSMike Kravetz struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) 2094c0d0381aSMike Kravetz { 2095336bf30eSMike Kravetz struct address_space *mapping = page_mapping(hpage); 2096c0d0381aSMike Kravetz 2097c0d0381aSMike Kravetz if (!mapping) 2098c0d0381aSMike Kravetz return mapping; 2099c0d0381aSMike Kravetz 2100c0d0381aSMike Kravetz if (i_mmap_trylock_write(mapping)) 2101c0d0381aSMike Kravetz return mapping; 2102c0d0381aSMike Kravetz 2103c0d0381aSMike Kravetz return NULL; 2104c0d0381aSMike Kravetz } 2105c0d0381aSMike Kravetz 2106fe19bd3dSHugh Dickins pgoff_t hugetlb_basepage_index(struct page *page) 210713d60f4bSZhang Yi { 210813d60f4bSZhang Yi struct page *page_head = compound_head(page); 210913d60f4bSZhang Yi pgoff_t index = page_index(page_head); 211013d60f4bSZhang Yi unsigned long compound_idx; 211113d60f4bSZhang Yi 211223baf831SKirill A. Shutemov if (compound_order(page_head) > MAX_ORDER) 211313d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 211413d60f4bSZhang Yi else 211513d60f4bSZhang Yi compound_idx = page - page_head; 211613d60f4bSZhang Yi 211713d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 211813d60f4bSZhang Yi } 211913d60f4bSZhang Yi 212019fc1a7eSSidhartha Kumar static struct folio *alloc_buddy_hugetlb_folio(struct hstate *h, 2121f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 2122f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 21231da177e4SLinus Torvalds { 2124af0fb9dfSMichal Hocko int order = huge_page_order(h); 21251da177e4SLinus Torvalds struct page *page; 2126f60858f9SMike Kravetz bool alloc_try_hard = true; 21272b21624fSMike Kravetz bool retry = true; 2128f96efd58SJoe Jin 2129f60858f9SMike Kravetz /* 2130f60858f9SMike Kravetz * By default we always try hard to allocate the page with 2131f60858f9SMike Kravetz * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in 2132f60858f9SMike Kravetz * a loop (to adjust global huge page counts) and previous allocation 2133f60858f9SMike Kravetz * failed, do not continue to try hard on the same node. Use the 2134f60858f9SMike Kravetz * node_alloc_noretry bitmap to manage this state information. 2135f60858f9SMike Kravetz */ 2136f60858f9SMike Kravetz if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) 2137f60858f9SMike Kravetz alloc_try_hard = false; 2138f60858f9SMike Kravetz gfp_mask |= __GFP_COMP|__GFP_NOWARN; 2139f60858f9SMike Kravetz if (alloc_try_hard) 2140f60858f9SMike Kravetz gfp_mask |= __GFP_RETRY_MAYFAIL; 2141af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 2142af0fb9dfSMichal Hocko nid = numa_mem_id(); 21432b21624fSMike Kravetz retry: 214484172f4bSMatthew Wilcox (Oracle) page = __alloc_pages(gfp_mask, order, nid, nmask); 21452b21624fSMike Kravetz 21462b21624fSMike Kravetz /* Freeze head page */ 21472b21624fSMike Kravetz if (page && !page_ref_freeze(page, 1)) { 21482b21624fSMike Kravetz __free_pages(page, order); 21492b21624fSMike Kravetz if (retry) { /* retry once */ 21502b21624fSMike Kravetz retry = false; 21512b21624fSMike Kravetz goto retry; 21522b21624fSMike Kravetz } 21532b21624fSMike Kravetz /* WOW! twice in a row. */ 21542b21624fSMike Kravetz pr_warn("HugeTLB head page unexpected inflated ref count\n"); 21552b21624fSMike Kravetz page = NULL; 21562b21624fSMike Kravetz } 21572b21624fSMike Kravetz 2158f60858f9SMike Kravetz /* 2159f60858f9SMike Kravetz * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this 2160f60858f9SMike Kravetz * indicates an overall state change. Clear bit so that we resume 2161f60858f9SMike Kravetz * normal 'try hard' allocations. 2162f60858f9SMike Kravetz */ 2163f60858f9SMike Kravetz if (node_alloc_noretry && page && !alloc_try_hard) 2164f60858f9SMike Kravetz node_clear(nid, *node_alloc_noretry); 2165f60858f9SMike Kravetz 2166f60858f9SMike Kravetz /* 2167f60858f9SMike Kravetz * If we tried hard to get a page but failed, set bit so that 2168f60858f9SMike Kravetz * subsequent attempts will not try as hard until there is an 2169f60858f9SMike Kravetz * overall state change. 2170f60858f9SMike Kravetz */ 2171f60858f9SMike Kravetz if (node_alloc_noretry && !page && alloc_try_hard) 2172f60858f9SMike Kravetz node_set(nid, *node_alloc_noretry); 2173f60858f9SMike Kravetz 217419fc1a7eSSidhartha Kumar if (!page) { 217519fc1a7eSSidhartha Kumar __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 217619fc1a7eSSidhartha Kumar return NULL; 217719fc1a7eSSidhartha Kumar } 217819fc1a7eSSidhartha Kumar 217919fc1a7eSSidhartha Kumar __count_vm_event(HTLB_BUDDY_PGALLOC); 218019fc1a7eSSidhartha Kumar return page_folio(page); 218163b4613cSNishanth Aravamudan } 218263b4613cSNishanth Aravamudan 2183af0fb9dfSMichal Hocko /* 21840c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 21850c397daeSMichal Hocko * should use this function to get new hugetlb pages 21862b21624fSMike Kravetz * 21872b21624fSMike Kravetz * Note that returned page is 'frozen': ref count of head page and all tail 21882b21624fSMike Kravetz * pages is zero. 21890c397daeSMichal Hocko */ 219019fc1a7eSSidhartha Kumar static struct folio *alloc_fresh_hugetlb_folio(struct hstate *h, 2191f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 2192f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 21930c397daeSMichal Hocko { 21947f325a8dSSidhartha Kumar struct folio *folio; 21957118fc29SMike Kravetz bool retry = false; 21960c397daeSMichal Hocko 21977118fc29SMike Kravetz retry: 21980c397daeSMichal Hocko if (hstate_is_gigantic(h)) 219919fc1a7eSSidhartha Kumar folio = alloc_gigantic_folio(h, gfp_mask, nid, nmask); 22000c397daeSMichal Hocko else 220119fc1a7eSSidhartha Kumar folio = alloc_buddy_hugetlb_folio(h, gfp_mask, 2202f60858f9SMike Kravetz nid, nmask, node_alloc_noretry); 220319fc1a7eSSidhartha Kumar if (!folio) 22040c397daeSMichal Hocko return NULL; 22057118fc29SMike Kravetz if (hstate_is_gigantic(h)) { 2206d1c60955SSidhartha Kumar if (!prep_compound_gigantic_folio(folio, huge_page_order(h))) { 22077118fc29SMike Kravetz /* 22087118fc29SMike Kravetz * Rare failure to convert pages to compound page. 22097118fc29SMike Kravetz * Free pages and try again - ONCE! 22107118fc29SMike Kravetz */ 22117f325a8dSSidhartha Kumar free_gigantic_folio(folio, huge_page_order(h)); 22127118fc29SMike Kravetz if (!retry) { 22137118fc29SMike Kravetz retry = true; 22147118fc29SMike Kravetz goto retry; 22157118fc29SMike Kravetz } 22167118fc29SMike Kravetz return NULL; 22177118fc29SMike Kravetz } 22187118fc29SMike Kravetz } 2219d1c60955SSidhartha Kumar prep_new_hugetlb_folio(h, folio, folio_nid(folio)); 22200c397daeSMichal Hocko 222119fc1a7eSSidhartha Kumar return folio; 22220c397daeSMichal Hocko } 22230c397daeSMichal Hocko 22240c397daeSMichal Hocko /* 2225af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 2226af0fb9dfSMichal Hocko * manner. 2227af0fb9dfSMichal Hocko */ 2228f60858f9SMike Kravetz static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 2229f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 2230b2261026SJoonsoo Kim { 223119fc1a7eSSidhartha Kumar struct folio *folio; 2232b2261026SJoonsoo Kim int nr_nodes, node; 2233af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2234b2261026SJoonsoo Kim 2235b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 223619fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, node, 223719fc1a7eSSidhartha Kumar nodes_allowed, node_alloc_noretry); 223819fc1a7eSSidhartha Kumar if (folio) { 2239454a00c4SMatthew Wilcox (Oracle) free_huge_folio(folio); /* free it into the hugepage allocator */ 224019fc1a7eSSidhartha Kumar return 1; 224119fc1a7eSSidhartha Kumar } 2242b2261026SJoonsoo Kim } 2243b2261026SJoonsoo Kim 2244af0fb9dfSMichal Hocko return 0; 2245b2261026SJoonsoo Kim } 2246b2261026SJoonsoo Kim 2247e8c5c824SLee Schermerhorn /* 224810c6ec49SMike Kravetz * Remove huge page from pool from next node to free. Attempt to keep 224910c6ec49SMike Kravetz * persistent huge pages more or less balanced over allowed nodes. 225010c6ec49SMike Kravetz * This routine only 'removes' the hugetlb page. The caller must make 225110c6ec49SMike Kravetz * an additional call to free the page to low level allocators. 2252e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 2253e8c5c824SLee Schermerhorn */ 225410c6ec49SMike Kravetz static struct page *remove_pool_huge_page(struct hstate *h, 225510c6ec49SMike Kravetz nodemask_t *nodes_allowed, 22566ae11b27SLee Schermerhorn bool acct_surplus) 2257e8c5c824SLee Schermerhorn { 2258b2261026SJoonsoo Kim int nr_nodes, node; 225910c6ec49SMike Kravetz struct page *page = NULL; 2260cfd5082bSSidhartha Kumar struct folio *folio; 2261e8c5c824SLee Schermerhorn 22629487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2263b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2264685f3457SLee Schermerhorn /* 2265685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 2266685f3457SLee Schermerhorn * nodes with surplus pages. 2267685f3457SLee Schermerhorn */ 2268b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 2269b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 227010c6ec49SMike Kravetz page = list_entry(h->hugepage_freelists[node].next, 2271e8c5c824SLee Schermerhorn struct page, lru); 2272cfd5082bSSidhartha Kumar folio = page_folio(page); 2273cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, acct_surplus); 22749a76db09SLee Schermerhorn break; 2275e8c5c824SLee Schermerhorn } 2276b2261026SJoonsoo Kim } 2277e8c5c824SLee Schermerhorn 227810c6ec49SMike Kravetz return page; 2279e8c5c824SLee Schermerhorn } 2280e8c5c824SLee Schermerhorn 2281c8721bbbSNaoya Horiguchi /* 2282c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 2283faf53defSNaoya Horiguchi * nothing for in-use hugepages and non-hugepages. 2284faf53defSNaoya Horiguchi * This function returns values like below: 2285faf53defSNaoya Horiguchi * 2286ad2fa371SMuchun Song * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages 2287ad2fa371SMuchun Song * when the system is under memory pressure and the feature of 2288ad2fa371SMuchun Song * freeing unused vmemmap pages associated with each hugetlb page 2289ad2fa371SMuchun Song * is enabled. 2290faf53defSNaoya Horiguchi * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 2291faf53defSNaoya Horiguchi * (allocated or reserved.) 2292faf53defSNaoya Horiguchi * 0: successfully dissolved free hugepages or the page is not a 2293faf53defSNaoya Horiguchi * hugepage (considered as already dissolved) 2294c8721bbbSNaoya Horiguchi */ 2295c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 2296c8721bbbSNaoya Horiguchi { 22976bc9b564SNaoya Horiguchi int rc = -EBUSY; 22981a7cdab5SSidhartha Kumar struct folio *folio = page_folio(page); 2299082d5b6bSGerald Schaefer 23007ffddd49SMuchun Song retry: 2301faf53defSNaoya Horiguchi /* Not to disrupt normal path by vainly holding hugetlb_lock */ 23021a7cdab5SSidhartha Kumar if (!folio_test_hugetlb(folio)) 2303faf53defSNaoya Horiguchi return 0; 2304faf53defSNaoya Horiguchi 2305db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 23061a7cdab5SSidhartha Kumar if (!folio_test_hugetlb(folio)) { 2307faf53defSNaoya Horiguchi rc = 0; 2308faf53defSNaoya Horiguchi goto out; 2309faf53defSNaoya Horiguchi } 2310faf53defSNaoya Horiguchi 23111a7cdab5SSidhartha Kumar if (!folio_ref_count(folio)) { 23121a7cdab5SSidhartha Kumar struct hstate *h = folio_hstate(folio); 23138346d69dSXin Hao if (!available_huge_pages(h)) 2314082d5b6bSGerald Schaefer goto out; 23157ffddd49SMuchun Song 23167ffddd49SMuchun Song /* 23177ffddd49SMuchun Song * We should make sure that the page is already on the free list 23187ffddd49SMuchun Song * when it is dissolved. 23197ffddd49SMuchun Song */ 23201a7cdab5SSidhartha Kumar if (unlikely(!folio_test_hugetlb_freed(folio))) { 2321db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 23227ffddd49SMuchun Song cond_resched(); 23237ffddd49SMuchun Song 23247ffddd49SMuchun Song /* 23257ffddd49SMuchun Song * Theoretically, we should return -EBUSY when we 23267ffddd49SMuchun Song * encounter this race. In fact, we have a chance 23277ffddd49SMuchun Song * to successfully dissolve the page if we do a 23287ffddd49SMuchun Song * retry. Because the race window is quite small. 23297ffddd49SMuchun Song * If we seize this opportunity, it is an optimization 23307ffddd49SMuchun Song * for increasing the success rate of dissolving page. 23317ffddd49SMuchun Song */ 23327ffddd49SMuchun Song goto retry; 23337ffddd49SMuchun Song } 23347ffddd49SMuchun Song 2335cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, false); 2336ad2fa371SMuchun Song h->max_huge_pages--; 2337ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2338ad2fa371SMuchun Song 2339c3114a84SAnshuman Khandual /* 2340d6ef19e2SSidhartha Kumar * Normally update_and_free_hugtlb_folio will allocate required vmemmmap 2341d6ef19e2SSidhartha Kumar * before freeing the page. update_and_free_hugtlb_folio will fail to 2342ad2fa371SMuchun Song * free the page if it can not allocate required vmemmap. We 2343ad2fa371SMuchun Song * need to adjust max_huge_pages if the page is not freed. 2344ad2fa371SMuchun Song * Attempt to allocate vmemmmap here so that we can take 2345ad2fa371SMuchun Song * appropriate action on failure. 2346ad2fa371SMuchun Song */ 23471a7cdab5SSidhartha Kumar rc = hugetlb_vmemmap_restore(h, &folio->page); 2348ad2fa371SMuchun Song if (!rc) { 2349d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, folio, false); 2350ad2fa371SMuchun Song } else { 2351ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 23522f6c57d6SSidhartha Kumar add_hugetlb_folio(h, folio, false); 2353ad2fa371SMuchun Song h->max_huge_pages++; 2354ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2355ad2fa371SMuchun Song } 2356ad2fa371SMuchun Song 2357ad2fa371SMuchun Song return rc; 2358c8721bbbSNaoya Horiguchi } 2359082d5b6bSGerald Schaefer out: 2360db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2361082d5b6bSGerald Schaefer return rc; 2362c8721bbbSNaoya Horiguchi } 2363c8721bbbSNaoya Horiguchi 2364c8721bbbSNaoya Horiguchi /* 2365c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 2366c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 23672247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 23682247bb33SGerald Schaefer * part of it lies within the given range. 2369082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 2370082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 2371c8721bbbSNaoya Horiguchi */ 2372082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 2373c8721bbbSNaoya Horiguchi { 2374c8721bbbSNaoya Horiguchi unsigned long pfn; 2375eb03aa00SGerald Schaefer struct page *page; 2376082d5b6bSGerald Schaefer int rc = 0; 2377dc2628f3SMuchun Song unsigned int order; 2378dc2628f3SMuchun Song struct hstate *h; 2379c8721bbbSNaoya Horiguchi 2380d0177639SLi Zhong if (!hugepages_supported()) 2381082d5b6bSGerald Schaefer return rc; 2382d0177639SLi Zhong 2383dc2628f3SMuchun Song order = huge_page_order(&default_hstate); 2384dc2628f3SMuchun Song for_each_hstate(h) 2385dc2628f3SMuchun Song order = min(order, huge_page_order(h)); 2386dc2628f3SMuchun Song 2387dc2628f3SMuchun Song for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) { 2388eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 2389eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 2390eb03aa00SGerald Schaefer if (rc) 2391082d5b6bSGerald Schaefer break; 2392eb03aa00SGerald Schaefer } 2393082d5b6bSGerald Schaefer 2394082d5b6bSGerald Schaefer return rc; 2395c8721bbbSNaoya Horiguchi } 2396c8721bbbSNaoya Horiguchi 2397ab5ac90aSMichal Hocko /* 2398ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 2399ab5ac90aSMichal Hocko */ 24003a740e8bSSidhartha Kumar static struct folio *alloc_surplus_hugetlb_folio(struct hstate *h, 24013a740e8bSSidhartha Kumar gfp_t gfp_mask, int nid, nodemask_t *nmask) 24027893d1d5SAdam Litke { 240319fc1a7eSSidhartha Kumar struct folio *folio = NULL; 24047893d1d5SAdam Litke 2405bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2406aa888a74SAndi Kleen return NULL; 2407aa888a74SAndi Kleen 2408db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 24099980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 24109980d744SMichal Hocko goto out_unlock; 2411db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2412d1c3fb1fSNishanth Aravamudan 241319fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL); 241419fc1a7eSSidhartha Kumar if (!folio) 24150c397daeSMichal Hocko return NULL; 2416d1c3fb1fSNishanth Aravamudan 2417db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 24189980d744SMichal Hocko /* 24199980d744SMichal Hocko * We could have raced with the pool size change. 24209980d744SMichal Hocko * Double check that and simply deallocate the new page 24219980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 2422454a00c4SMatthew Wilcox (Oracle) * temporary page to workaround the nasty free_huge_folio 24239980d744SMichal Hocko * codeflow 24249980d744SMichal Hocko */ 24259980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 242619fc1a7eSSidhartha Kumar folio_set_hugetlb_temporary(folio); 2427db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2428454a00c4SMatthew Wilcox (Oracle) free_huge_folio(folio); 24292bf753e6SKai Shen return NULL; 2430b65a4edaSMike Kravetz } 2431b65a4edaSMike Kravetz 24329980d744SMichal Hocko h->surplus_huge_pages++; 243319fc1a7eSSidhartha Kumar h->surplus_huge_pages_node[folio_nid(folio)]++; 24349980d744SMichal Hocko 24359980d744SMichal Hocko out_unlock: 2436db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 24377893d1d5SAdam Litke 24383a740e8bSSidhartha Kumar return folio; 24397893d1d5SAdam Litke } 24407893d1d5SAdam Litke 2441e37d3e83SSidhartha Kumar static struct folio *alloc_migrate_hugetlb_folio(struct hstate *h, gfp_t gfp_mask, 2442ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 2443ab5ac90aSMichal Hocko { 244419fc1a7eSSidhartha Kumar struct folio *folio; 2445ab5ac90aSMichal Hocko 2446ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 2447ab5ac90aSMichal Hocko return NULL; 2448ab5ac90aSMichal Hocko 244919fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL); 245019fc1a7eSSidhartha Kumar if (!folio) 2451ab5ac90aSMichal Hocko return NULL; 2452ab5ac90aSMichal Hocko 24532b21624fSMike Kravetz /* fresh huge pages are frozen */ 245419fc1a7eSSidhartha Kumar folio_ref_unfreeze(folio, 1); 2455ab5ac90aSMichal Hocko /* 2456ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 2457ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 2458ab5ac90aSMichal Hocko */ 245919fc1a7eSSidhartha Kumar folio_set_hugetlb_temporary(folio); 2460ab5ac90aSMichal Hocko 2461e37d3e83SSidhartha Kumar return folio; 2462ab5ac90aSMichal Hocko } 2463ab5ac90aSMichal Hocko 2464e4e574b7SAdam Litke /* 2465099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 2466099730d6SDave Hansen */ 2467e0ec90eeSDave Hansen static 2468ff7d853bSSidhartha Kumar struct folio *alloc_buddy_hugetlb_folio_with_mpol(struct hstate *h, 2469099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 2470099730d6SDave Hansen { 24713a740e8bSSidhartha Kumar struct folio *folio = NULL; 2472aaf14e40SMichal Hocko struct mempolicy *mpol; 2473aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 2474aaf14e40SMichal Hocko int nid; 2475aaf14e40SMichal Hocko nodemask_t *nodemask; 2476aaf14e40SMichal Hocko 2477aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 2478cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 2479cfcaa66fSBen Widawsky gfp_t gfp = gfp_mask | __GFP_NOWARN; 2480cfcaa66fSBen Widawsky 2481cfcaa66fSBen Widawsky gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL); 24823a740e8bSSidhartha Kumar folio = alloc_surplus_hugetlb_folio(h, gfp, nid, nodemask); 2483cfcaa66fSBen Widawsky 2484cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 2485cfcaa66fSBen Widawsky nodemask = NULL; 2486cfcaa66fSBen Widawsky } 2487cfcaa66fSBen Widawsky 24883a740e8bSSidhartha Kumar if (!folio) 24893a740e8bSSidhartha Kumar folio = alloc_surplus_hugetlb_folio(h, gfp_mask, nid, nodemask); 2490aaf14e40SMichal Hocko mpol_cond_put(mpol); 2491ff7d853bSSidhartha Kumar return folio; 2492099730d6SDave Hansen } 2493099730d6SDave Hansen 2494e37d3e83SSidhartha Kumar /* folio migration callback function */ 2495e37d3e83SSidhartha Kumar struct folio *alloc_hugetlb_folio_nodemask(struct hstate *h, int preferred_nid, 2496d92bbc27SJoonsoo Kim nodemask_t *nmask, gfp_t gfp_mask) 24974db9b2efSMichal Hocko { 2498db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 24998346d69dSXin Hao if (available_huge_pages(h)) { 2500a36f1e90SSidhartha Kumar struct folio *folio; 25013e59fcb0SMichal Hocko 2502a36f1e90SSidhartha Kumar folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask, 2503a36f1e90SSidhartha Kumar preferred_nid, nmask); 2504a36f1e90SSidhartha Kumar if (folio) { 2505db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2506e37d3e83SSidhartha Kumar return folio; 25074db9b2efSMichal Hocko } 25084db9b2efSMichal Hocko } 2509db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 25104db9b2efSMichal Hocko 2511e37d3e83SSidhartha Kumar return alloc_migrate_hugetlb_folio(h, gfp_mask, preferred_nid, nmask); 25124db9b2efSMichal Hocko } 25134db9b2efSMichal Hocko 2514ebd63723SMichal Hocko /* mempolicy aware migration callback */ 2515d0ce0e47SSidhartha Kumar struct folio *alloc_hugetlb_folio_vma(struct hstate *h, struct vm_area_struct *vma, 2516389c8178SMichal Hocko unsigned long address) 2517ebd63723SMichal Hocko { 2518ebd63723SMichal Hocko struct mempolicy *mpol; 2519ebd63723SMichal Hocko nodemask_t *nodemask; 2520e37d3e83SSidhartha Kumar struct folio *folio; 2521ebd63723SMichal Hocko gfp_t gfp_mask; 2522ebd63723SMichal Hocko int node; 2523ebd63723SMichal Hocko 2524ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 2525ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 2526e37d3e83SSidhartha Kumar folio = alloc_hugetlb_folio_nodemask(h, node, nodemask, gfp_mask); 2527ebd63723SMichal Hocko mpol_cond_put(mpol); 2528ebd63723SMichal Hocko 2529d0ce0e47SSidhartha Kumar return folio; 2530ebd63723SMichal Hocko } 2531ebd63723SMichal Hocko 2532bf50bab2SNaoya Horiguchi /* 253325985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 2534e4e574b7SAdam Litke * of size 'delta'. 2535e4e574b7SAdam Litke */ 25360a4f3d1bSLiu Xiang static int gather_surplus_pages(struct hstate *h, long delta) 25371b2a1e7bSJules Irenge __must_hold(&hugetlb_lock) 2538e4e574b7SAdam Litke { 253934665341SMiaohe Lin LIST_HEAD(surplus_list); 2540454a00c4SMatthew Wilcox (Oracle) struct folio *folio, *tmp; 25410a4f3d1bSLiu Xiang int ret; 25420a4f3d1bSLiu Xiang long i; 25430a4f3d1bSLiu Xiang long needed, allocated; 254428073b02SHillf Danton bool alloc_ok = true; 2545e4e574b7SAdam Litke 25469487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2547a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 2548ac09b3a1SAdam Litke if (needed <= 0) { 2549a5516438SAndi Kleen h->resv_huge_pages += delta; 2550e4e574b7SAdam Litke return 0; 2551ac09b3a1SAdam Litke } 2552e4e574b7SAdam Litke 2553e4e574b7SAdam Litke allocated = 0; 2554e4e574b7SAdam Litke 2555e4e574b7SAdam Litke ret = -ENOMEM; 2556e4e574b7SAdam Litke retry: 2557db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2558e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 25593a740e8bSSidhartha Kumar folio = alloc_surplus_hugetlb_folio(h, htlb_alloc_mask(h), 25602b21624fSMike Kravetz NUMA_NO_NODE, NULL); 25613a740e8bSSidhartha Kumar if (!folio) { 256228073b02SHillf Danton alloc_ok = false; 256328073b02SHillf Danton break; 256428073b02SHillf Danton } 25653a740e8bSSidhartha Kumar list_add(&folio->lru, &surplus_list); 256669ed779aSDavid Rientjes cond_resched(); 2567e4e574b7SAdam Litke } 256828073b02SHillf Danton allocated += i; 2569e4e574b7SAdam Litke 2570e4e574b7SAdam Litke /* 2571e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 2572e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 2573e4e574b7SAdam Litke */ 2574db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2575a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 2576a5516438SAndi Kleen (h->free_huge_pages + allocated); 257728073b02SHillf Danton if (needed > 0) { 257828073b02SHillf Danton if (alloc_ok) 2579e4e574b7SAdam Litke goto retry; 258028073b02SHillf Danton /* 258128073b02SHillf Danton * We were not able to allocate enough pages to 258228073b02SHillf Danton * satisfy the entire reservation so we free what 258328073b02SHillf Danton * we've allocated so far. 258428073b02SHillf Danton */ 258528073b02SHillf Danton goto free; 258628073b02SHillf Danton } 2587e4e574b7SAdam Litke /* 2588e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 258925985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 2590e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 2591ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 2592ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 2593ac09b3a1SAdam Litke * before they are reserved. 2594e4e574b7SAdam Litke */ 2595e4e574b7SAdam Litke needed += allocated; 2596a5516438SAndi Kleen h->resv_huge_pages += delta; 2597e4e574b7SAdam Litke ret = 0; 2598a9869b83SNaoya Horiguchi 259919fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 2600454a00c4SMatthew Wilcox (Oracle) list_for_each_entry_safe(folio, tmp, &surplus_list, lru) { 260119fc3f0aSAdam Litke if ((--needed) < 0) 260219fc3f0aSAdam Litke break; 2603b65a4edaSMike Kravetz /* Add the page to the hugetlb allocator */ 2604454a00c4SMatthew Wilcox (Oracle) enqueue_hugetlb_folio(h, folio); 260519fc3f0aSAdam Litke } 260628073b02SHillf Danton free: 2607db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 260819fc3f0aSAdam Litke 2609b65a4edaSMike Kravetz /* 2610b65a4edaSMike Kravetz * Free unnecessary surplus pages to the buddy allocator. 2611454a00c4SMatthew Wilcox (Oracle) * Pages have no ref count, call free_huge_folio directly. 2612b65a4edaSMike Kravetz */ 2613454a00c4SMatthew Wilcox (Oracle) list_for_each_entry_safe(folio, tmp, &surplus_list, lru) 2614454a00c4SMatthew Wilcox (Oracle) free_huge_folio(folio); 2615db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2616e4e574b7SAdam Litke 2617e4e574b7SAdam Litke return ret; 2618e4e574b7SAdam Litke } 2619e4e574b7SAdam Litke 2620e4e574b7SAdam Litke /* 2621e5bbc8a6SMike Kravetz * This routine has two main purposes: 2622e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 2623e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 2624e5bbc8a6SMike Kravetz * to the associated reservation map. 2625e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 2626e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 2627e4e574b7SAdam Litke */ 2628a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 2629a5516438SAndi Kleen unsigned long unused_resv_pages) 2630e4e574b7SAdam Litke { 2631e4e574b7SAdam Litke unsigned long nr_pages; 263210c6ec49SMike Kravetz struct page *page; 263310c6ec49SMike Kravetz LIST_HEAD(page_list); 263410c6ec49SMike Kravetz 26359487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 263610c6ec49SMike Kravetz /* Uncommit the reservation */ 263710c6ec49SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 2638e4e574b7SAdam Litke 2639c0531714SNaoya Horiguchi if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 2640e5bbc8a6SMike Kravetz goto out; 2641aa888a74SAndi Kleen 2642e5bbc8a6SMike Kravetz /* 2643e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 2644e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 2645e5bbc8a6SMike Kravetz */ 2646a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 2647e4e574b7SAdam Litke 2648685f3457SLee Schermerhorn /* 2649685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 26509b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 26519b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 26529b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 265310c6ec49SMike Kravetz * remove_pool_huge_page() will balance the freed pages across the 26549b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 2655685f3457SLee Schermerhorn */ 2656685f3457SLee Schermerhorn while (nr_pages--) { 265710c6ec49SMike Kravetz page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1); 265810c6ec49SMike Kravetz if (!page) 2659e5bbc8a6SMike Kravetz goto out; 266010c6ec49SMike Kravetz 266110c6ec49SMike Kravetz list_add(&page->lru, &page_list); 2662e4e574b7SAdam Litke } 2663e5bbc8a6SMike Kravetz 2664e5bbc8a6SMike Kravetz out: 2665db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 266610c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 2667db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2668e4e574b7SAdam Litke } 2669e4e574b7SAdam Litke 26705e911373SMike Kravetz 2671c37f9fb1SAndy Whitcroft /* 2672feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 26735e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 2674cf3ad20bSMike Kravetz * 2675cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 2676cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 2677cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 2678cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 2679cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 2680feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 2681feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 2682feba16e2SMike Kravetz * is called in such cases. 2683cf3ad20bSMike Kravetz * 2684cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 2685cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 2686cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 2687cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 2688cf3ad20bSMike Kravetz * take appropriate action. 268996b96a96SMike Kravetz * 269096b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 269196b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 269296b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 269396b96a96SMike Kravetz * reservation exists. 2694846be085SMike Kravetz * 2695846be085SMike Kravetz * vma_del_reservation is used in error paths where an entry in the reserve 2696846be085SMike Kravetz * map was created during huge page allocation and must be removed. It is to 2697846be085SMike Kravetz * be called after calling vma_needs_reservation to determine if a reservation 2698846be085SMike Kravetz * exists. 2699c37f9fb1SAndy Whitcroft */ 27005e911373SMike Kravetz enum vma_resv_mode { 27015e911373SMike Kravetz VMA_NEEDS_RESV, 27025e911373SMike Kravetz VMA_COMMIT_RESV, 2703feba16e2SMike Kravetz VMA_END_RESV, 270496b96a96SMike Kravetz VMA_ADD_RESV, 2705846be085SMike Kravetz VMA_DEL_RESV, 27065e911373SMike Kravetz }; 2707cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 2708cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 27095e911373SMike Kravetz enum vma_resv_mode mode) 2710c37f9fb1SAndy Whitcroft { 27114e35f483SJoonsoo Kim struct resv_map *resv; 27124e35f483SJoonsoo Kim pgoff_t idx; 2713cf3ad20bSMike Kravetz long ret; 27140db9d74eSMina Almasry long dummy_out_regions_needed; 2715c37f9fb1SAndy Whitcroft 27164e35f483SJoonsoo Kim resv = vma_resv_map(vma); 27174e35f483SJoonsoo Kim if (!resv) 2718c37f9fb1SAndy Whitcroft return 1; 2719c37f9fb1SAndy Whitcroft 27204e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 27215e911373SMike Kravetz switch (mode) { 27225e911373SMike Kravetz case VMA_NEEDS_RESV: 27230db9d74eSMina Almasry ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); 27240db9d74eSMina Almasry /* We assume that vma_reservation_* routines always operate on 27250db9d74eSMina Almasry * 1 page, and that adding to resv map a 1 page entry can only 27260db9d74eSMina Almasry * ever require 1 region. 27270db9d74eSMina Almasry */ 27280db9d74eSMina Almasry VM_BUG_ON(dummy_out_regions_needed != 1); 27295e911373SMike Kravetz break; 27305e911373SMike Kravetz case VMA_COMMIT_RESV: 2731075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 27320db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 27330db9d74eSMina Almasry VM_BUG_ON(ret < 0); 27345e911373SMike Kravetz break; 2735feba16e2SMike Kravetz case VMA_END_RESV: 27360db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 27375e911373SMike Kravetz ret = 0; 27385e911373SMike Kravetz break; 273996b96a96SMike Kravetz case VMA_ADD_RESV: 27400db9d74eSMina Almasry if (vma->vm_flags & VM_MAYSHARE) { 2741075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 27420db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 27430db9d74eSMina Almasry VM_BUG_ON(ret < 0); 27440db9d74eSMina Almasry } else { 27450db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 274696b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 274796b96a96SMike Kravetz } 274896b96a96SMike Kravetz break; 2749846be085SMike Kravetz case VMA_DEL_RESV: 2750846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 2751846be085SMike Kravetz region_abort(resv, idx, idx + 1, 1); 2752846be085SMike Kravetz ret = region_del(resv, idx, idx + 1); 2753846be085SMike Kravetz } else { 2754846be085SMike Kravetz ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 2755846be085SMike Kravetz /* region_add calls of range 1 should never fail. */ 2756846be085SMike Kravetz VM_BUG_ON(ret < 0); 2757846be085SMike Kravetz } 2758846be085SMike Kravetz break; 27595e911373SMike Kravetz default: 27605e911373SMike Kravetz BUG(); 27615e911373SMike Kravetz } 276284afd99bSAndy Whitcroft 2763846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV) 2764cf3ad20bSMike Kravetz return ret; 276567961f9dSMike Kravetz /* 2766bf3d12b9SMiaohe Lin * We know private mapping must have HPAGE_RESV_OWNER set. 2767bf3d12b9SMiaohe Lin * 276867961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 276967961f9dSMike Kravetz * However, a file associated with mapping could have been 277067961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 277167961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 277267961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 277367961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 277467961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 277567961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 277667961f9dSMike Kravetz * reservation has already been consumed. As a result, the 277767961f9dSMike Kravetz * return value of this routine is the opposite of the 277867961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 277967961f9dSMike Kravetz */ 2780bf3d12b9SMiaohe Lin if (ret > 0) 278167961f9dSMike Kravetz return 0; 2782bf3d12b9SMiaohe Lin if (ret == 0) 278367961f9dSMike Kravetz return 1; 2784bf3d12b9SMiaohe Lin return ret; 278584afd99bSAndy Whitcroft } 2786cf3ad20bSMike Kravetz 2787cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 2788a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 2789c37f9fb1SAndy Whitcroft { 27905e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 2791cf3ad20bSMike Kravetz } 2792c37f9fb1SAndy Whitcroft 2793cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 2794cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 2795cf3ad20bSMike Kravetz { 27965e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 27975e911373SMike Kravetz } 27985e911373SMike Kravetz 2799feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 28005e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 28015e911373SMike Kravetz { 2802feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 2803c37f9fb1SAndy Whitcroft } 2804c37f9fb1SAndy Whitcroft 280596b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 280696b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 280796b96a96SMike Kravetz { 280896b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 280996b96a96SMike Kravetz } 281096b96a96SMike Kravetz 2811846be085SMike Kravetz static long vma_del_reservation(struct hstate *h, 2812846be085SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 281396b96a96SMike Kravetz { 2814846be085SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV); 2815846be085SMike Kravetz } 2816846be085SMike Kravetz 2817846be085SMike Kravetz /* 2818846be085SMike Kravetz * This routine is called to restore reservation information on error paths. 2819d0ce0e47SSidhartha Kumar * It should ONLY be called for folios allocated via alloc_hugetlb_folio(), 2820d0ce0e47SSidhartha Kumar * and the hugetlb mutex should remain held when calling this routine. 2821846be085SMike Kravetz * 2822846be085SMike Kravetz * It handles two specific cases: 2823d2d7bb44SSidhartha Kumar * 1) A reservation was in place and the folio consumed the reservation. 2824d2d7bb44SSidhartha Kumar * hugetlb_restore_reserve is set in the folio. 2825d2d7bb44SSidhartha Kumar * 2) No reservation was in place for the page, so hugetlb_restore_reserve is 2826d0ce0e47SSidhartha Kumar * not set. However, alloc_hugetlb_folio always updates the reserve map. 2827846be085SMike Kravetz * 2828454a00c4SMatthew Wilcox (Oracle) * In case 1, free_huge_folio later in the error path will increment the 2829454a00c4SMatthew Wilcox (Oracle) * global reserve count. But, free_huge_folio does not have enough context 2830846be085SMike Kravetz * to adjust the reservation map. This case deals primarily with private 2831846be085SMike Kravetz * mappings. Adjust the reserve map here to be consistent with global 2832454a00c4SMatthew Wilcox (Oracle) * reserve count adjustments to be made by free_huge_folio. Make sure the 2833846be085SMike Kravetz * reserve map indicates there is a reservation present. 2834846be085SMike Kravetz * 2835d0ce0e47SSidhartha Kumar * In case 2, simply undo reserve map modifications done by alloc_hugetlb_folio. 2836846be085SMike Kravetz */ 2837846be085SMike Kravetz void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma, 2838d2d7bb44SSidhartha Kumar unsigned long address, struct folio *folio) 2839846be085SMike Kravetz { 284096b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 284196b96a96SMike Kravetz 28420ffdc38eSSidhartha Kumar if (folio_test_hugetlb_restore_reserve(folio)) { 2843846be085SMike Kravetz if (unlikely(rc < 0)) 284496b96a96SMike Kravetz /* 284596b96a96SMike Kravetz * Rare out of memory condition in reserve map 28460ffdc38eSSidhartha Kumar * manipulation. Clear hugetlb_restore_reserve so 28470ffdc38eSSidhartha Kumar * that global reserve count will not be incremented 2848454a00c4SMatthew Wilcox (Oracle) * by free_huge_folio. This will make it appear 28490ffdc38eSSidhartha Kumar * as though the reservation for this folio was 285096b96a96SMike Kravetz * consumed. This may prevent the task from 28510ffdc38eSSidhartha Kumar * faulting in the folio at a later time. This 285296b96a96SMike Kravetz * is better than inconsistent global huge page 285396b96a96SMike Kravetz * accounting of reserve counts. 285496b96a96SMike Kravetz */ 28550ffdc38eSSidhartha Kumar folio_clear_hugetlb_restore_reserve(folio); 2856846be085SMike Kravetz else if (rc) 2857846be085SMike Kravetz (void)vma_add_reservation(h, vma, address); 2858846be085SMike Kravetz else 2859846be085SMike Kravetz vma_end_reservation(h, vma, address); 2860846be085SMike Kravetz } else { 2861846be085SMike Kravetz if (!rc) { 286296b96a96SMike Kravetz /* 2863846be085SMike Kravetz * This indicates there is an entry in the reserve map 2864d0ce0e47SSidhartha Kumar * not added by alloc_hugetlb_folio. We know it was added 2865d0ce0e47SSidhartha Kumar * before the alloc_hugetlb_folio call, otherwise 28660ffdc38eSSidhartha Kumar * hugetlb_restore_reserve would be set on the folio. 2867846be085SMike Kravetz * Remove the entry so that a subsequent allocation 2868846be085SMike Kravetz * does not consume a reservation. 286996b96a96SMike Kravetz */ 2870846be085SMike Kravetz rc = vma_del_reservation(h, vma, address); 2871846be085SMike Kravetz if (rc < 0) 2872846be085SMike Kravetz /* 2873846be085SMike Kravetz * VERY rare out of memory condition. Since 2874846be085SMike Kravetz * we can not delete the entry, set 28750ffdc38eSSidhartha Kumar * hugetlb_restore_reserve so that the reserve 28760ffdc38eSSidhartha Kumar * count will be incremented when the folio 2877846be085SMike Kravetz * is freed. This reserve will be consumed 2878846be085SMike Kravetz * on a subsequent allocation. 2879846be085SMike Kravetz */ 28800ffdc38eSSidhartha Kumar folio_set_hugetlb_restore_reserve(folio); 2881846be085SMike Kravetz } else if (rc < 0) { 2882846be085SMike Kravetz /* 2883846be085SMike Kravetz * Rare out of memory condition from 2884846be085SMike Kravetz * vma_needs_reservation call. Memory allocation is 2885846be085SMike Kravetz * only attempted if a new entry is needed. Therefore, 2886846be085SMike Kravetz * this implies there is not an entry in the 2887846be085SMike Kravetz * reserve map. 2888846be085SMike Kravetz * 2889846be085SMike Kravetz * For shared mappings, no entry in the map indicates 2890846be085SMike Kravetz * no reservation. We are done. 2891846be085SMike Kravetz */ 2892846be085SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 2893846be085SMike Kravetz /* 2894846be085SMike Kravetz * For private mappings, no entry indicates 2895846be085SMike Kravetz * a reservation is present. Since we can 28960ffdc38eSSidhartha Kumar * not add an entry, set hugetlb_restore_reserve 28970ffdc38eSSidhartha Kumar * on the folio so reserve count will be 2898846be085SMike Kravetz * incremented when freed. This reserve will 2899846be085SMike Kravetz * be consumed on a subsequent allocation. 2900846be085SMike Kravetz */ 29010ffdc38eSSidhartha Kumar folio_set_hugetlb_restore_reserve(folio); 290296b96a96SMike Kravetz } else 2903846be085SMike Kravetz /* 2904846be085SMike Kravetz * No reservation present, do nothing 2905846be085SMike Kravetz */ 290696b96a96SMike Kravetz vma_end_reservation(h, vma, address); 290796b96a96SMike Kravetz } 290896b96a96SMike Kravetz } 290996b96a96SMike Kravetz 2910369fa227SOscar Salvador /* 291119fc1a7eSSidhartha Kumar * alloc_and_dissolve_hugetlb_folio - Allocate a new folio and dissolve 291219fc1a7eSSidhartha Kumar * the old one 2913369fa227SOscar Salvador * @h: struct hstate old page belongs to 291419fc1a7eSSidhartha Kumar * @old_folio: Old folio to dissolve 2915ae37c7ffSOscar Salvador * @list: List to isolate the page in case we need to 2916369fa227SOscar Salvador * Returns 0 on success, otherwise negated error. 2917369fa227SOscar Salvador */ 291819fc1a7eSSidhartha Kumar static int alloc_and_dissolve_hugetlb_folio(struct hstate *h, 291919fc1a7eSSidhartha Kumar struct folio *old_folio, struct list_head *list) 2920369fa227SOscar Salvador { 2921369fa227SOscar Salvador gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2922de656ed3SSidhartha Kumar int nid = folio_nid(old_folio); 2923de656ed3SSidhartha Kumar struct folio *new_folio; 2924369fa227SOscar Salvador int ret = 0; 2925369fa227SOscar Salvador 2926369fa227SOscar Salvador /* 292719fc1a7eSSidhartha Kumar * Before dissolving the folio, we need to allocate a new one for the 292819fc1a7eSSidhartha Kumar * pool to remain stable. Here, we allocate the folio and 'prep' it 2929f41f2ed4SMuchun Song * by doing everything but actually updating counters and adding to 2930f41f2ed4SMuchun Song * the pool. This simplifies and let us do most of the processing 2931f41f2ed4SMuchun Song * under the lock. 2932369fa227SOscar Salvador */ 293319fc1a7eSSidhartha Kumar new_folio = alloc_buddy_hugetlb_folio(h, gfp_mask, nid, NULL, NULL); 293419fc1a7eSSidhartha Kumar if (!new_folio) 2935369fa227SOscar Salvador return -ENOMEM; 2936de656ed3SSidhartha Kumar __prep_new_hugetlb_folio(h, new_folio); 2937369fa227SOscar Salvador 2938369fa227SOscar Salvador retry: 2939369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2940de656ed3SSidhartha Kumar if (!folio_test_hugetlb(old_folio)) { 2941369fa227SOscar Salvador /* 294219fc1a7eSSidhartha Kumar * Freed from under us. Drop new_folio too. 2943369fa227SOscar Salvador */ 2944369fa227SOscar Salvador goto free_new; 2945de656ed3SSidhartha Kumar } else if (folio_ref_count(old_folio)) { 29469747b9e9SBaolin Wang bool isolated; 29479747b9e9SBaolin Wang 2948369fa227SOscar Salvador /* 294919fc1a7eSSidhartha Kumar * Someone has grabbed the folio, try to isolate it here. 2950ae37c7ffSOscar Salvador * Fail with -EBUSY if not possible. 2951369fa227SOscar Salvador */ 2952ae37c7ffSOscar Salvador spin_unlock_irq(&hugetlb_lock); 29539747b9e9SBaolin Wang isolated = isolate_hugetlb(old_folio, list); 29549747b9e9SBaolin Wang ret = isolated ? 0 : -EBUSY; 2955ae37c7ffSOscar Salvador spin_lock_irq(&hugetlb_lock); 2956369fa227SOscar Salvador goto free_new; 2957de656ed3SSidhartha Kumar } else if (!folio_test_hugetlb_freed(old_folio)) { 2958369fa227SOscar Salvador /* 295919fc1a7eSSidhartha Kumar * Folio's refcount is 0 but it has not been enqueued in the 2960369fa227SOscar Salvador * freelist yet. Race window is small, so we can succeed here if 2961369fa227SOscar Salvador * we retry. 2962369fa227SOscar Salvador */ 2963369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2964369fa227SOscar Salvador cond_resched(); 2965369fa227SOscar Salvador goto retry; 2966369fa227SOscar Salvador } else { 2967369fa227SOscar Salvador /* 296819fc1a7eSSidhartha Kumar * Ok, old_folio is still a genuine free hugepage. Remove it from 2969369fa227SOscar Salvador * the freelist and decrease the counters. These will be 2970369fa227SOscar Salvador * incremented again when calling __prep_account_new_huge_page() 2971240d67a8SSidhartha Kumar * and enqueue_hugetlb_folio() for new_folio. The counters will 2972240d67a8SSidhartha Kumar * remain stable since this happens under the lock. 2973369fa227SOscar Salvador */ 2974cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, old_folio, false); 2975369fa227SOscar Salvador 2976369fa227SOscar Salvador /* 297719fc1a7eSSidhartha Kumar * Ref count on new_folio is already zero as it was dropped 2978b65a4edaSMike Kravetz * earlier. It can be directly added to the pool free list. 2979369fa227SOscar Salvador */ 2980369fa227SOscar Salvador __prep_account_new_huge_page(h, nid); 2981240d67a8SSidhartha Kumar enqueue_hugetlb_folio(h, new_folio); 2982369fa227SOscar Salvador 2983369fa227SOscar Salvador /* 298419fc1a7eSSidhartha Kumar * Folio has been replaced, we can safely free the old one. 2985369fa227SOscar Salvador */ 2986369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2987d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, old_folio, false); 2988369fa227SOscar Salvador } 2989369fa227SOscar Salvador 2990369fa227SOscar Salvador return ret; 2991369fa227SOscar Salvador 2992369fa227SOscar Salvador free_new: 2993369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 299419fc1a7eSSidhartha Kumar /* Folio has a zero ref count, but needs a ref to be freed */ 2995de656ed3SSidhartha Kumar folio_ref_unfreeze(new_folio, 1); 2996d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, new_folio, false); 2997369fa227SOscar Salvador 2998369fa227SOscar Salvador return ret; 2999369fa227SOscar Salvador } 3000369fa227SOscar Salvador 3001ae37c7ffSOscar Salvador int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list) 3002369fa227SOscar Salvador { 3003369fa227SOscar Salvador struct hstate *h; 3004d5e33bd8SSidhartha Kumar struct folio *folio = page_folio(page); 3005ae37c7ffSOscar Salvador int ret = -EBUSY; 3006369fa227SOscar Salvador 3007369fa227SOscar Salvador /* 3008369fa227SOscar Salvador * The page might have been dissolved from under our feet, so make sure 3009369fa227SOscar Salvador * to carefully check the state under the lock. 3010369fa227SOscar Salvador * Return success when racing as if we dissolved the page ourselves. 3011369fa227SOscar Salvador */ 3012369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 3013d5e33bd8SSidhartha Kumar if (folio_test_hugetlb(folio)) { 3014d5e33bd8SSidhartha Kumar h = folio_hstate(folio); 3015369fa227SOscar Salvador } else { 3016369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 3017369fa227SOscar Salvador return 0; 3018369fa227SOscar Salvador } 3019369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 3020369fa227SOscar Salvador 3021369fa227SOscar Salvador /* 3022369fa227SOscar Salvador * Fence off gigantic pages as there is a cyclic dependency between 3023369fa227SOscar Salvador * alloc_contig_range and them. Return -ENOMEM as this has the effect 3024369fa227SOscar Salvador * of bailing out right away without further retrying. 3025369fa227SOscar Salvador */ 3026369fa227SOscar Salvador if (hstate_is_gigantic(h)) 3027369fa227SOscar Salvador return -ENOMEM; 3028369fa227SOscar Salvador 30299747b9e9SBaolin Wang if (folio_ref_count(folio) && isolate_hugetlb(folio, list)) 3030ae37c7ffSOscar Salvador ret = 0; 3031d5e33bd8SSidhartha Kumar else if (!folio_ref_count(folio)) 303219fc1a7eSSidhartha Kumar ret = alloc_and_dissolve_hugetlb_folio(h, folio, list); 3033ae37c7ffSOscar Salvador 3034ae37c7ffSOscar Salvador return ret; 3035369fa227SOscar Salvador } 3036369fa227SOscar Salvador 3037d0ce0e47SSidhartha Kumar struct folio *alloc_hugetlb_folio(struct vm_area_struct *vma, 303804f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 3039348ea204SAdam Litke { 304090481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 3041a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3042d4ab0316SSidhartha Kumar struct folio *folio; 3043d85f69b0SMike Kravetz long map_chg, map_commit; 3044d85f69b0SMike Kravetz long gbl_chg; 30456d76dcf4SAneesh Kumar K.V int ret, idx; 3046d0ce0e47SSidhartha Kumar struct hugetlb_cgroup *h_cg = NULL; 304708cf9fafSMina Almasry bool deferred_reserve; 30482fc39cecSAdam Litke 30496d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 3050a1e78772SMel Gorman /* 3051d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 3052d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 3053d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 3054a1e78772SMel Gorman */ 3055d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 3056d85f69b0SMike Kravetz if (map_chg < 0) 305776dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 3058d85f69b0SMike Kravetz 3059d85f69b0SMike Kravetz /* 3060d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 3061d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 3062d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 3063d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 3064d85f69b0SMike Kravetz * checked against any subpool limit. 3065d85f69b0SMike Kravetz */ 3066d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 3067d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 3068d85f69b0SMike Kravetz if (gbl_chg < 0) { 3069feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 307076dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 30715e911373SMike Kravetz } 307290d8b7e6SAdam Litke 3073d85f69b0SMike Kravetz /* 3074d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 3075d85f69b0SMike Kravetz * map, there could be reservations associated with the 3076d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 3077d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 3078d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 3079d85f69b0SMike Kravetz * the subpool reservations. 3080d85f69b0SMike Kravetz */ 3081d85f69b0SMike Kravetz if (avoid_reserve) 3082d85f69b0SMike Kravetz gbl_chg = 1; 3083d85f69b0SMike Kravetz } 3084d85f69b0SMike Kravetz 308508cf9fafSMina Almasry /* If this allocation is not consuming a reservation, charge it now. 308608cf9fafSMina Almasry */ 30876501fe5fSMiaohe Lin deferred_reserve = map_chg || avoid_reserve; 308808cf9fafSMina Almasry if (deferred_reserve) { 308908cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 309008cf9fafSMina Almasry idx, pages_per_huge_page(h), &h_cg); 30918f34af6fSJianyu Zhan if (ret) 30928f34af6fSJianyu Zhan goto out_subpool_put; 309308cf9fafSMina Almasry } 309408cf9fafSMina Almasry 309508cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 309608cf9fafSMina Almasry if (ret) 309708cf9fafSMina Almasry goto out_uncharge_cgroup_reservation; 30988f34af6fSJianyu Zhan 3099db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3100d85f69b0SMike Kravetz /* 3101d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 3102d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 3103d85f69b0SMike Kravetz * a reservation exists for the allocation. 3104d85f69b0SMike Kravetz */ 3105ff7d853bSSidhartha Kumar folio = dequeue_hugetlb_folio_vma(h, vma, addr, avoid_reserve, gbl_chg); 3106ff7d853bSSidhartha Kumar if (!folio) { 3107db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3108ff7d853bSSidhartha Kumar folio = alloc_buddy_hugetlb_folio_with_mpol(h, vma, addr); 3109ff7d853bSSidhartha Kumar if (!folio) 31108f34af6fSJianyu Zhan goto out_uncharge_cgroup; 311112df140fSRik van Riel spin_lock_irq(&hugetlb_lock); 3112a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 3113ff7d853bSSidhartha Kumar folio_set_hugetlb_restore_reserve(folio); 3114a88c7695SNaoya Horiguchi h->resv_huge_pages--; 3115a88c7695SNaoya Horiguchi } 3116ff7d853bSSidhartha Kumar list_add(&folio->lru, &h->hugepage_activelist); 3117ff7d853bSSidhartha Kumar folio_ref_unfreeze(folio, 1); 311881a6fcaeSJoonsoo Kim /* Fall through */ 3119a1e78772SMel Gorman } 3120ff7d853bSSidhartha Kumar 3121ff7d853bSSidhartha Kumar hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, folio); 312208cf9fafSMina Almasry /* If allocation is not consuming a reservation, also store the 312308cf9fafSMina Almasry * hugetlb_cgroup pointer on the page. 312408cf9fafSMina Almasry */ 312508cf9fafSMina Almasry if (deferred_reserve) { 312608cf9fafSMina Almasry hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), 3127ff7d853bSSidhartha Kumar h_cg, folio); 312808cf9fafSMina Almasry } 312908cf9fafSMina Almasry 3130db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3131a1e78772SMel Gorman 3132ff7d853bSSidhartha Kumar hugetlb_set_folio_subpool(folio, spool); 3133a1e78772SMel Gorman 3134d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 3135d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 313633039678SMike Kravetz /* 313733039678SMike Kravetz * The page was added to the reservation map between 313833039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 313933039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 314033039678SMike Kravetz * Adjust for the subpool count incremented above AND 314133039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 314233039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 314333039678SMike Kravetz * no longer applies. 314433039678SMike Kravetz */ 314533039678SMike Kravetz long rsv_adjust; 314633039678SMike Kravetz 314733039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 314833039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 314979aa925bSMike Kravetz if (deferred_reserve) 3150d4ab0316SSidhartha Kumar hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h), 3151d4ab0316SSidhartha Kumar pages_per_huge_page(h), folio); 315233039678SMike Kravetz } 3153d0ce0e47SSidhartha Kumar return folio; 31548f34af6fSJianyu Zhan 31558f34af6fSJianyu Zhan out_uncharge_cgroup: 31568f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 315708cf9fafSMina Almasry out_uncharge_cgroup_reservation: 315808cf9fafSMina Almasry if (deferred_reserve) 315908cf9fafSMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), 316008cf9fafSMina Almasry h_cg); 31618f34af6fSJianyu Zhan out_subpool_put: 3162d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 31638f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 3164feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 31658f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 3166b45b5bd6SDavid Gibson } 3167b45b5bd6SDavid Gibson 3168b5389086SZhenguo Yao int alloc_bootmem_huge_page(struct hstate *h, int nid) 3169e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 3170b5389086SZhenguo Yao int __alloc_bootmem_huge_page(struct hstate *h, int nid) 3171aa888a74SAndi Kleen { 3172b5389086SZhenguo Yao struct huge_bootmem_page *m = NULL; /* initialize for clang */ 3173b2261026SJoonsoo Kim int nr_nodes, node; 3174aa888a74SAndi Kleen 3175b5389086SZhenguo Yao /* do node specific alloc */ 3176b5389086SZhenguo Yao if (nid != NUMA_NO_NODE) { 3177b5389086SZhenguo Yao m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h), 3178b5389086SZhenguo Yao 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid); 3179b5389086SZhenguo Yao if (!m) 3180b5389086SZhenguo Yao return 0; 3181b5389086SZhenguo Yao goto found; 3182b5389086SZhenguo Yao } 3183b5389086SZhenguo Yao /* allocate from next node when distributing huge pages */ 3184b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 3185b5389086SZhenguo Yao m = memblock_alloc_try_nid_raw( 31868b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 318797ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 3188aa888a74SAndi Kleen /* 3189aa888a74SAndi Kleen * Use the beginning of the huge page to store the 3190aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 3191aa888a74SAndi Kleen * puts them into the mem_map). 3192aa888a74SAndi Kleen */ 3193b5389086SZhenguo Yao if (!m) 3194b5389086SZhenguo Yao return 0; 3195aa888a74SAndi Kleen goto found; 3196aa888a74SAndi Kleen } 3197aa888a74SAndi Kleen 3198aa888a74SAndi Kleen found: 3199aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 3200330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 3201aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 3202aa888a74SAndi Kleen m->hstate = h; 3203aa888a74SAndi Kleen return 1; 3204aa888a74SAndi Kleen } 3205aa888a74SAndi Kleen 320648b8d744SMike Kravetz /* 320748b8d744SMike Kravetz * Put bootmem huge pages into the standard lists after mem_map is up. 320848b8d744SMike Kravetz * Note: This only applies to gigantic (order > MAX_ORDER) pages. 320948b8d744SMike Kravetz */ 3210aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 3211aa888a74SAndi Kleen { 3212aa888a74SAndi Kleen struct huge_bootmem_page *m; 3213aa888a74SAndi Kleen 3214aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 321540d18ebfSMike Kravetz struct page *page = virt_to_page(m); 32167f325a8dSSidhartha Kumar struct folio *folio = page_folio(page); 3217aa888a74SAndi Kleen struct hstate *h = m->hstate; 3218ee8f248dSBecky Bruce 321948b8d744SMike Kravetz VM_BUG_ON(!hstate_is_gigantic(h)); 3220d1c60955SSidhartha Kumar WARN_ON(folio_ref_count(folio) != 1); 3221d1c60955SSidhartha Kumar if (prep_compound_gigantic_folio(folio, huge_page_order(h))) { 3222d1c60955SSidhartha Kumar WARN_ON(folio_test_reserved(folio)); 3223d1c60955SSidhartha Kumar prep_new_hugetlb_folio(h, folio, folio_nid(folio)); 3224454a00c4SMatthew Wilcox (Oracle) free_huge_folio(folio); /* add to the hugepage allocator */ 32257118fc29SMike Kravetz } else { 3226416d85edSMike Kravetz /* VERY unlikely inflated ref count on a tail page */ 32277f325a8dSSidhartha Kumar free_gigantic_folio(folio, huge_page_order(h)); 32287118fc29SMike Kravetz } 3229af0fb9dfSMichal Hocko 3230b0320c7bSRafael Aquini /* 323148b8d744SMike Kravetz * We need to restore the 'stolen' pages to totalram_pages 323248b8d744SMike Kravetz * in order to fix confusing memory reports from free(1) and 323348b8d744SMike Kravetz * other side-effects, like CommitLimit going negative. 3234b0320c7bSRafael Aquini */ 3235c78a7f36SMiaohe Lin adjust_managed_page_count(page, pages_per_huge_page(h)); 3236520495feSCannon Matthews cond_resched(); 3237aa888a74SAndi Kleen } 3238aa888a74SAndi Kleen } 3239b5389086SZhenguo Yao static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid) 3240b5389086SZhenguo Yao { 3241b5389086SZhenguo Yao unsigned long i; 3242b5389086SZhenguo Yao char buf[32]; 3243b5389086SZhenguo Yao 3244b5389086SZhenguo Yao for (i = 0; i < h->max_huge_pages_node[nid]; ++i) { 3245b5389086SZhenguo Yao if (hstate_is_gigantic(h)) { 3246b5389086SZhenguo Yao if (!alloc_bootmem_huge_page(h, nid)) 3247b5389086SZhenguo Yao break; 3248b5389086SZhenguo Yao } else { 324919fc1a7eSSidhartha Kumar struct folio *folio; 3250b5389086SZhenguo Yao gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 3251b5389086SZhenguo Yao 325219fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, 3253b5389086SZhenguo Yao &node_states[N_MEMORY], NULL); 325419fc1a7eSSidhartha Kumar if (!folio) 3255b5389086SZhenguo Yao break; 3256454a00c4SMatthew Wilcox (Oracle) free_huge_folio(folio); /* free it into the hugepage allocator */ 3257b5389086SZhenguo Yao } 3258b5389086SZhenguo Yao cond_resched(); 3259b5389086SZhenguo Yao } 3260b5389086SZhenguo Yao if (i == h->max_huge_pages_node[nid]) 3261b5389086SZhenguo Yao return; 3262b5389086SZhenguo Yao 3263b5389086SZhenguo Yao string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3264b5389086SZhenguo Yao pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n", 3265b5389086SZhenguo Yao h->max_huge_pages_node[nid], buf, nid, i); 3266b5389086SZhenguo Yao h->max_huge_pages -= (h->max_huge_pages_node[nid] - i); 3267b5389086SZhenguo Yao h->max_huge_pages_node[nid] = i; 3268b5389086SZhenguo Yao } 3269aa888a74SAndi Kleen 32708faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 32711da177e4SLinus Torvalds { 32721da177e4SLinus Torvalds unsigned long i; 3273f60858f9SMike Kravetz nodemask_t *node_alloc_noretry; 3274b5389086SZhenguo Yao bool node_specific_alloc = false; 3275f60858f9SMike Kravetz 3276b5389086SZhenguo Yao /* skip gigantic hugepages allocation if hugetlb_cma enabled */ 3277b5389086SZhenguo Yao if (hstate_is_gigantic(h) && hugetlb_cma_size) { 3278b5389086SZhenguo Yao pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); 3279b5389086SZhenguo Yao return; 3280b5389086SZhenguo Yao } 3281b5389086SZhenguo Yao 3282b5389086SZhenguo Yao /* do node specific alloc */ 32830a7a0f6fSPeng Liu for_each_online_node(i) { 3284b5389086SZhenguo Yao if (h->max_huge_pages_node[i] > 0) { 3285b5389086SZhenguo Yao hugetlb_hstate_alloc_pages_onenode(h, i); 3286b5389086SZhenguo Yao node_specific_alloc = true; 3287b5389086SZhenguo Yao } 3288b5389086SZhenguo Yao } 3289b5389086SZhenguo Yao 3290b5389086SZhenguo Yao if (node_specific_alloc) 3291b5389086SZhenguo Yao return; 3292b5389086SZhenguo Yao 3293b5389086SZhenguo Yao /* below will do all node balanced alloc */ 3294f60858f9SMike Kravetz if (!hstate_is_gigantic(h)) { 3295f60858f9SMike Kravetz /* 3296f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3297f60858f9SMike Kravetz * Ignore errors as lower level routines can deal with 3298f60858f9SMike Kravetz * node_alloc_noretry == NULL. If this kmalloc fails at boot 3299f60858f9SMike Kravetz * time, we are likely in bigger trouble. 3300f60858f9SMike Kravetz */ 3301f60858f9SMike Kravetz node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), 3302f60858f9SMike Kravetz GFP_KERNEL); 3303f60858f9SMike Kravetz } else { 3304f60858f9SMike Kravetz /* allocations done at boot time */ 3305f60858f9SMike Kravetz node_alloc_noretry = NULL; 3306f60858f9SMike Kravetz } 3307f60858f9SMike Kravetz 3308f60858f9SMike Kravetz /* bit mask controlling how hard we retry per-node allocations */ 3309f60858f9SMike Kravetz if (node_alloc_noretry) 3310f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 33111da177e4SLinus Torvalds 3312e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 3313bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 3314b5389086SZhenguo Yao if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE)) 3315aa888a74SAndi Kleen break; 33160c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 3317f60858f9SMike Kravetz &node_states[N_MEMORY], 3318f60858f9SMike Kravetz node_alloc_noretry)) 33191da177e4SLinus Torvalds break; 332069ed779aSDavid Rientjes cond_resched(); 33211da177e4SLinus Torvalds } 3322d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 3323d715cf80SLiam R. Howlett char buf[32]; 3324d715cf80SLiam R. Howlett 3325c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3326d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 3327d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 33288faa8b07SAndi Kleen h->max_huge_pages = i; 3329e5ff2159SAndi Kleen } 3330f60858f9SMike Kravetz kfree(node_alloc_noretry); 3331d715cf80SLiam R. Howlett } 3332e5ff2159SAndi Kleen 3333e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 3334e5ff2159SAndi Kleen { 333579dfc695SMike Kravetz struct hstate *h, *h2; 3336e5ff2159SAndi Kleen 3337e5ff2159SAndi Kleen for_each_hstate(h) { 33388faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 3339bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 33408faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 334179dfc695SMike Kravetz 334279dfc695SMike Kravetz /* 334379dfc695SMike Kravetz * Set demote order for each hstate. Note that 334479dfc695SMike Kravetz * h->demote_order is initially 0. 334579dfc695SMike Kravetz * - We can not demote gigantic pages if runtime freeing 334679dfc695SMike Kravetz * is not supported, so skip this. 3347a01f4390SMike Kravetz * - If CMA allocation is possible, we can not demote 3348a01f4390SMike Kravetz * HUGETLB_PAGE_ORDER or smaller size pages. 334979dfc695SMike Kravetz */ 335079dfc695SMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 335179dfc695SMike Kravetz continue; 3352a01f4390SMike Kravetz if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER) 3353a01f4390SMike Kravetz continue; 335479dfc695SMike Kravetz for_each_hstate(h2) { 335579dfc695SMike Kravetz if (h2 == h) 335679dfc695SMike Kravetz continue; 335779dfc695SMike Kravetz if (h2->order < h->order && 335879dfc695SMike Kravetz h2->order > h->demote_order) 335979dfc695SMike Kravetz h->demote_order = h2->order; 336079dfc695SMike Kravetz } 3361e5ff2159SAndi Kleen } 3362e5ff2159SAndi Kleen } 3363e5ff2159SAndi Kleen 3364e5ff2159SAndi Kleen static void __init report_hugepages(void) 3365e5ff2159SAndi Kleen { 3366e5ff2159SAndi Kleen struct hstate *h; 3367e5ff2159SAndi Kleen 3368e5ff2159SAndi Kleen for_each_hstate(h) { 33694abd32dbSAndi Kleen char buf[32]; 3370c6247f72SMatthew Wilcox 3371c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 33726213834cSMuchun Song pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n", 3373c6247f72SMatthew Wilcox buf, h->free_huge_pages); 33746213834cSMuchun Song pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n", 33756213834cSMuchun Song hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf); 3376e5ff2159SAndi Kleen } 3377e5ff2159SAndi Kleen } 3378e5ff2159SAndi Kleen 33791da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 33806ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 33816ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 33821da177e4SLinus Torvalds { 33834415cc8dSChristoph Lameter int i; 33841121828aSMike Kravetz LIST_HEAD(page_list); 33854415cc8dSChristoph Lameter 33869487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 3387bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3388aa888a74SAndi Kleen return; 3389aa888a74SAndi Kleen 33901121828aSMike Kravetz /* 33911121828aSMike Kravetz * Collect pages to be freed on a list, and free after dropping lock 33921121828aSMike Kravetz */ 33936ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 339410c6ec49SMike Kravetz struct page *page, *next; 3395a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 3396a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 3397a5516438SAndi Kleen if (count >= h->nr_huge_pages) 33981121828aSMike Kravetz goto out; 33991da177e4SLinus Torvalds if (PageHighMem(page)) 34001da177e4SLinus Torvalds continue; 3401cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, page_folio(page), false); 34021121828aSMike Kravetz list_add(&page->lru, &page_list); 34031121828aSMike Kravetz } 34041121828aSMike Kravetz } 34051121828aSMike Kravetz 34061121828aSMike Kravetz out: 3407db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 340810c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3409db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 34101da177e4SLinus Torvalds } 34111da177e4SLinus Torvalds #else 34126ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 34136ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 34141da177e4SLinus Torvalds { 34151da177e4SLinus Torvalds } 34161da177e4SLinus Torvalds #endif 34171da177e4SLinus Torvalds 341820a0307cSWu Fengguang /* 341920a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 342020a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 342120a0307cSWu Fengguang * Returns 1 if an adjustment was made. 342220a0307cSWu Fengguang */ 34236ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 34246ae11b27SLee Schermerhorn int delta) 342520a0307cSWu Fengguang { 3426b2261026SJoonsoo Kim int nr_nodes, node; 342720a0307cSWu Fengguang 34289487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 342920a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 343020a0307cSWu Fengguang 3431e8c5c824SLee Schermerhorn if (delta < 0) { 3432b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 3433b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 3434b2261026SJoonsoo Kim goto found; 3435b2261026SJoonsoo Kim } 3436b2261026SJoonsoo Kim } else { 3437b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 3438b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 3439b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 3440b2261026SJoonsoo Kim goto found; 3441e8c5c824SLee Schermerhorn } 34429a76db09SLee Schermerhorn } 3443b2261026SJoonsoo Kim return 0; 344420a0307cSWu Fengguang 3445b2261026SJoonsoo Kim found: 344620a0307cSWu Fengguang h->surplus_huge_pages += delta; 3447b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 3448b2261026SJoonsoo Kim return 1; 344920a0307cSWu Fengguang } 345020a0307cSWu Fengguang 3451a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 3452fd875dcaSMike Kravetz static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 34536ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 34541da177e4SLinus Torvalds { 34557893d1d5SAdam Litke unsigned long min_count, ret; 345610c6ec49SMike Kravetz struct page *page; 345710c6ec49SMike Kravetz LIST_HEAD(page_list); 3458f60858f9SMike Kravetz NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); 3459f60858f9SMike Kravetz 3460f60858f9SMike Kravetz /* 3461f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3462f60858f9SMike Kravetz * If we can not allocate the bit mask, do not attempt to allocate 3463f60858f9SMike Kravetz * the requested huge pages. 3464f60858f9SMike Kravetz */ 3465f60858f9SMike Kravetz if (node_alloc_noretry) 3466f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 3467f60858f9SMike Kravetz else 3468f60858f9SMike Kravetz return -ENOMEM; 34691da177e4SLinus Torvalds 347029383967SMike Kravetz /* 347129383967SMike Kravetz * resize_lock mutex prevents concurrent adjustments to number of 347229383967SMike Kravetz * pages in hstate via the proc/sysfs interfaces. 347329383967SMike Kravetz */ 347429383967SMike Kravetz mutex_lock(&h->resize_lock); 3475b65d4adbSMuchun Song flush_free_hpage_work(h); 3476db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 34774eb0716eSAlexandre Ghiti 34784eb0716eSAlexandre Ghiti /* 3479fd875dcaSMike Kravetz * Check for a node specific request. 3480fd875dcaSMike Kravetz * Changing node specific huge page count may require a corresponding 3481fd875dcaSMike Kravetz * change to the global count. In any case, the passed node mask 3482fd875dcaSMike Kravetz * (nodes_allowed) will restrict alloc/free to the specified node. 3483fd875dcaSMike Kravetz */ 3484fd875dcaSMike Kravetz if (nid != NUMA_NO_NODE) { 3485fd875dcaSMike Kravetz unsigned long old_count = count; 3486fd875dcaSMike Kravetz 3487fd875dcaSMike Kravetz count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 3488fd875dcaSMike Kravetz /* 3489fd875dcaSMike Kravetz * User may have specified a large count value which caused the 3490fd875dcaSMike Kravetz * above calculation to overflow. In this case, they wanted 3491fd875dcaSMike Kravetz * to allocate as many huge pages as possible. Set count to 3492fd875dcaSMike Kravetz * largest possible value to align with their intention. 3493fd875dcaSMike Kravetz */ 3494fd875dcaSMike Kravetz if (count < old_count) 3495fd875dcaSMike Kravetz count = ULONG_MAX; 3496fd875dcaSMike Kravetz } 3497fd875dcaSMike Kravetz 3498fd875dcaSMike Kravetz /* 34994eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 35004eb0716eSAlexandre Ghiti * page range allocation. 35014eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 35024eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 35034eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 35044eb0716eSAlexandre Ghiti */ 35054eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 35064eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 3507db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 350829383967SMike Kravetz mutex_unlock(&h->resize_lock); 3509f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 35104eb0716eSAlexandre Ghiti return -EINVAL; 35114eb0716eSAlexandre Ghiti } 35124eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 35134eb0716eSAlexandre Ghiti } 3514aa888a74SAndi Kleen 35157893d1d5SAdam Litke /* 35167893d1d5SAdam Litke * Increase the pool size 35177893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 35187893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 3519d1c3fb1fSNishanth Aravamudan * 35203a740e8bSSidhartha Kumar * We might race with alloc_surplus_hugetlb_folio() here and be unable 3521d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 3522d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 3523d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 3524d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 35257893d1d5SAdam Litke */ 3526a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 35276ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 35287893d1d5SAdam Litke break; 35297893d1d5SAdam Litke } 35307893d1d5SAdam Litke 3531a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 35327893d1d5SAdam Litke /* 35337893d1d5SAdam Litke * If this allocation races such that we no longer need the 3534454a00c4SMatthew Wilcox (Oracle) * page, free_huge_folio will handle it by freeing the page 35357893d1d5SAdam Litke * and reducing the surplus. 35367893d1d5SAdam Litke */ 3537db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3538649920c6SJia He 3539649920c6SJia He /* yield cpu to avoid soft lockup */ 3540649920c6SJia He cond_resched(); 3541649920c6SJia He 3542f60858f9SMike Kravetz ret = alloc_pool_huge_page(h, nodes_allowed, 3543f60858f9SMike Kravetz node_alloc_noretry); 3544db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 35457893d1d5SAdam Litke if (!ret) 35467893d1d5SAdam Litke goto out; 35477893d1d5SAdam Litke 3548536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 3549536240f2SMel Gorman if (signal_pending(current)) 3550536240f2SMel Gorman goto out; 35517893d1d5SAdam Litke } 35527893d1d5SAdam Litke 35537893d1d5SAdam Litke /* 35547893d1d5SAdam Litke * Decrease the pool size 35557893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 35567893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 35577893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 35587893d1d5SAdam Litke * to the desired size as pages become free. 3559d1c3fb1fSNishanth Aravamudan * 3560d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 3561d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 3562d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 35633a740e8bSSidhartha Kumar * alloc_surplus_hugetlb_folio() is checking the global counter, 3564d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 3565d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 3566d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 35677893d1d5SAdam Litke */ 3568a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 35696b0c880dSAdam Litke min_count = max(count, min_count); 35706ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 357110c6ec49SMike Kravetz 357210c6ec49SMike Kravetz /* 357310c6ec49SMike Kravetz * Collect pages to be removed on list without dropping lock 357410c6ec49SMike Kravetz */ 3575a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 357610c6ec49SMike Kravetz page = remove_pool_huge_page(h, nodes_allowed, 0); 357710c6ec49SMike Kravetz if (!page) 35781da177e4SLinus Torvalds break; 357910c6ec49SMike Kravetz 358010c6ec49SMike Kravetz list_add(&page->lru, &page_list); 35811da177e4SLinus Torvalds } 358210c6ec49SMike Kravetz /* free the pages after dropping lock */ 3583db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 358410c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3585b65d4adbSMuchun Song flush_free_hpage_work(h); 3586db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 358710c6ec49SMike Kravetz 3588a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 35896ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 35907893d1d5SAdam Litke break; 35917893d1d5SAdam Litke } 35927893d1d5SAdam Litke out: 35934eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 3594db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 359529383967SMike Kravetz mutex_unlock(&h->resize_lock); 35964eb0716eSAlexandre Ghiti 3597f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 3598f60858f9SMike Kravetz 35994eb0716eSAlexandre Ghiti return 0; 36001da177e4SLinus Torvalds } 36011da177e4SLinus Torvalds 3602bdd7be07SSidhartha Kumar static int demote_free_hugetlb_folio(struct hstate *h, struct folio *folio) 36038531fc6fSMike Kravetz { 3604bdd7be07SSidhartha Kumar int i, nid = folio_nid(folio); 36058531fc6fSMike Kravetz struct hstate *target_hstate; 360631731452SDoug Berger struct page *subpage; 3607bdd7be07SSidhartha Kumar struct folio *inner_folio; 36088531fc6fSMike Kravetz int rc = 0; 36098531fc6fSMike Kravetz 36108531fc6fSMike Kravetz target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order); 36118531fc6fSMike Kravetz 3612cfd5082bSSidhartha Kumar remove_hugetlb_folio_for_demote(h, folio, false); 36138531fc6fSMike Kravetz spin_unlock_irq(&hugetlb_lock); 36148531fc6fSMike Kravetz 3615bdd7be07SSidhartha Kumar rc = hugetlb_vmemmap_restore(h, &folio->page); 36168531fc6fSMike Kravetz if (rc) { 3617bdd7be07SSidhartha Kumar /* Allocation of vmemmmap failed, we can not demote folio */ 36188531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 3619bdd7be07SSidhartha Kumar folio_ref_unfreeze(folio, 1); 3620bdd7be07SSidhartha Kumar add_hugetlb_folio(h, folio, false); 36218531fc6fSMike Kravetz return rc; 36228531fc6fSMike Kravetz } 36238531fc6fSMike Kravetz 36248531fc6fSMike Kravetz /* 3625911565b8SSidhartha Kumar * Use destroy_compound_hugetlb_folio_for_demote for all huge page 3626bdd7be07SSidhartha Kumar * sizes as it will not ref count folios. 36278531fc6fSMike Kravetz */ 3628911565b8SSidhartha Kumar destroy_compound_hugetlb_folio_for_demote(folio, huge_page_order(h)); 36298531fc6fSMike Kravetz 36308531fc6fSMike Kravetz /* 36318531fc6fSMike Kravetz * Taking target hstate mutex synchronizes with set_max_huge_pages. 36328531fc6fSMike Kravetz * Without the mutex, pages added to target hstate could be marked 36338531fc6fSMike Kravetz * as surplus. 36348531fc6fSMike Kravetz * 36358531fc6fSMike Kravetz * Note that we already hold h->resize_lock. To prevent deadlock, 36368531fc6fSMike Kravetz * use the convention of always taking larger size hstate mutex first. 36378531fc6fSMike Kravetz */ 36388531fc6fSMike Kravetz mutex_lock(&target_hstate->resize_lock); 36398531fc6fSMike Kravetz for (i = 0; i < pages_per_huge_page(h); 36408531fc6fSMike Kravetz i += pages_per_huge_page(target_hstate)) { 3641bdd7be07SSidhartha Kumar subpage = folio_page(folio, i); 3642bdd7be07SSidhartha Kumar inner_folio = page_folio(subpage); 36438531fc6fSMike Kravetz if (hstate_is_gigantic(target_hstate)) 3644bdd7be07SSidhartha Kumar prep_compound_gigantic_folio_for_demote(inner_folio, 36458531fc6fSMike Kravetz target_hstate->order); 36468531fc6fSMike Kravetz else 364731731452SDoug Berger prep_compound_page(subpage, target_hstate->order); 3648bdd7be07SSidhartha Kumar folio_change_private(inner_folio, NULL); 3649bdd7be07SSidhartha Kumar prep_new_hugetlb_folio(target_hstate, inner_folio, nid); 3650454a00c4SMatthew Wilcox (Oracle) free_huge_folio(inner_folio); 36518531fc6fSMike Kravetz } 36528531fc6fSMike Kravetz mutex_unlock(&target_hstate->resize_lock); 36538531fc6fSMike Kravetz 36548531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 36558531fc6fSMike Kravetz 36568531fc6fSMike Kravetz /* 36578531fc6fSMike Kravetz * Not absolutely necessary, but for consistency update max_huge_pages 36588531fc6fSMike Kravetz * based on pool changes for the demoted page. 36598531fc6fSMike Kravetz */ 36608531fc6fSMike Kravetz h->max_huge_pages--; 3661a43a83c7SMiaohe Lin target_hstate->max_huge_pages += 3662a43a83c7SMiaohe Lin pages_per_huge_page(h) / pages_per_huge_page(target_hstate); 36638531fc6fSMike Kravetz 36648531fc6fSMike Kravetz return rc; 36658531fc6fSMike Kravetz } 36668531fc6fSMike Kravetz 366779dfc695SMike Kravetz static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 366879dfc695SMike Kravetz __must_hold(&hugetlb_lock) 366979dfc695SMike Kravetz { 36708531fc6fSMike Kravetz int nr_nodes, node; 3671bdd7be07SSidhartha Kumar struct folio *folio; 367279dfc695SMike Kravetz 367379dfc695SMike Kravetz lockdep_assert_held(&hugetlb_lock); 367479dfc695SMike Kravetz 367579dfc695SMike Kravetz /* We should never get here if no demote order */ 367679dfc695SMike Kravetz if (!h->demote_order) { 367779dfc695SMike Kravetz pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n"); 367879dfc695SMike Kravetz return -EINVAL; /* internal error */ 367979dfc695SMike Kravetz } 368079dfc695SMike Kravetz 36818531fc6fSMike Kravetz for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 3682bdd7be07SSidhartha Kumar list_for_each_entry(folio, &h->hugepage_freelists[node], lru) { 3683bdd7be07SSidhartha Kumar if (folio_test_hwpoison(folio)) 36845a317412SMike Kravetz continue; 3685bdd7be07SSidhartha Kumar return demote_free_hugetlb_folio(h, folio); 36868531fc6fSMike Kravetz } 36878531fc6fSMike Kravetz } 36888531fc6fSMike Kravetz 36895a317412SMike Kravetz /* 36905a317412SMike Kravetz * Only way to get here is if all pages on free lists are poisoned. 36915a317412SMike Kravetz * Return -EBUSY so that caller will not retry. 36925a317412SMike Kravetz */ 36935a317412SMike Kravetz return -EBUSY; 369479dfc695SMike Kravetz } 369579dfc695SMike Kravetz 3696a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 3697a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 3698a3437870SNishanth Aravamudan 369979dfc695SMike Kravetz #define HSTATE_ATTR_WO(_name) \ 370079dfc695SMike Kravetz static struct kobj_attribute _name##_attr = __ATTR_WO(_name) 370179dfc695SMike Kravetz 3702a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 370398bc26acSMiaohe Lin static struct kobj_attribute _name##_attr = __ATTR_RW(_name) 3704a3437870SNishanth Aravamudan 3705a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 3706a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 3707a3437870SNishanth Aravamudan 37089a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 37099a305230SLee Schermerhorn 37109a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 3711a3437870SNishanth Aravamudan { 3712a3437870SNishanth Aravamudan int i; 37139a305230SLee Schermerhorn 3714a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 37159a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 37169a305230SLee Schermerhorn if (nidp) 37179a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 3718a3437870SNishanth Aravamudan return &hstates[i]; 37199a305230SLee Schermerhorn } 37209a305230SLee Schermerhorn 37219a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 3722a3437870SNishanth Aravamudan } 3723a3437870SNishanth Aravamudan 372406808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 3725a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3726a3437870SNishanth Aravamudan { 37279a305230SLee Schermerhorn struct hstate *h; 37289a305230SLee Schermerhorn unsigned long nr_huge_pages; 37299a305230SLee Schermerhorn int nid; 37309a305230SLee Schermerhorn 37319a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 37329a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 37339a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 37349a305230SLee Schermerhorn else 37359a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 37369a305230SLee Schermerhorn 3737ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", nr_huge_pages); 3738a3437870SNishanth Aravamudan } 3739adbe8726SEric B Munson 3740238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 3741238d3c13SDavid Rientjes struct hstate *h, int nid, 3742238d3c13SDavid Rientjes unsigned long count, size_t len) 3743a3437870SNishanth Aravamudan { 3744a3437870SNishanth Aravamudan int err; 37452d0adf7eSOscar Salvador nodemask_t nodes_allowed, *n_mask; 3746a3437870SNishanth Aravamudan 37472d0adf7eSOscar Salvador if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 37482d0adf7eSOscar Salvador return -EINVAL; 3749adbe8726SEric B Munson 37509a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 37519a305230SLee Schermerhorn /* 37529a305230SLee Schermerhorn * global hstate attribute 37539a305230SLee Schermerhorn */ 37549a305230SLee Schermerhorn if (!(obey_mempolicy && 37552d0adf7eSOscar Salvador init_nodemask_of_mempolicy(&nodes_allowed))) 37562d0adf7eSOscar Salvador n_mask = &node_states[N_MEMORY]; 37572d0adf7eSOscar Salvador else 37582d0adf7eSOscar Salvador n_mask = &nodes_allowed; 37592d0adf7eSOscar Salvador } else { 37609a305230SLee Schermerhorn /* 3761fd875dcaSMike Kravetz * Node specific request. count adjustment happens in 3762fd875dcaSMike Kravetz * set_max_huge_pages() after acquiring hugetlb_lock. 37639a305230SLee Schermerhorn */ 37642d0adf7eSOscar Salvador init_nodemask_of_node(&nodes_allowed, nid); 37652d0adf7eSOscar Salvador n_mask = &nodes_allowed; 3766fd875dcaSMike Kravetz } 37679a305230SLee Schermerhorn 37682d0adf7eSOscar Salvador err = set_max_huge_pages(h, count, nid, n_mask); 376906808b08SLee Schermerhorn 37704eb0716eSAlexandre Ghiti return err ? err : len; 377106808b08SLee Schermerhorn } 377206808b08SLee Schermerhorn 3773238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 3774238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 3775238d3c13SDavid Rientjes size_t len) 3776238d3c13SDavid Rientjes { 3777238d3c13SDavid Rientjes struct hstate *h; 3778238d3c13SDavid Rientjes unsigned long count; 3779238d3c13SDavid Rientjes int nid; 3780238d3c13SDavid Rientjes int err; 3781238d3c13SDavid Rientjes 3782238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 3783238d3c13SDavid Rientjes if (err) 3784238d3c13SDavid Rientjes return err; 3785238d3c13SDavid Rientjes 3786238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 3787238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 3788238d3c13SDavid Rientjes } 3789238d3c13SDavid Rientjes 379006808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 379106808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 379206808b08SLee Schermerhorn { 379306808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 379406808b08SLee Schermerhorn } 379506808b08SLee Schermerhorn 379606808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 379706808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 379806808b08SLee Schermerhorn { 3799238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 3800a3437870SNishanth Aravamudan } 3801a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 3802a3437870SNishanth Aravamudan 380306808b08SLee Schermerhorn #ifdef CONFIG_NUMA 380406808b08SLee Schermerhorn 380506808b08SLee Schermerhorn /* 380606808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 380706808b08SLee Schermerhorn * huge page alloc/free. 380806808b08SLee Schermerhorn */ 380906808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 3810ae7a927dSJoe Perches struct kobj_attribute *attr, 3811ae7a927dSJoe Perches char *buf) 381206808b08SLee Schermerhorn { 381306808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 381406808b08SLee Schermerhorn } 381506808b08SLee Schermerhorn 381606808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 381706808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 381806808b08SLee Schermerhorn { 3819238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 382006808b08SLee Schermerhorn } 382106808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 382206808b08SLee Schermerhorn #endif 382306808b08SLee Schermerhorn 382406808b08SLee Schermerhorn 3825a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 3826a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3827a3437870SNishanth Aravamudan { 38289a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3829ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); 3830a3437870SNishanth Aravamudan } 3831adbe8726SEric B Munson 3832a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 3833a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 3834a3437870SNishanth Aravamudan { 3835a3437870SNishanth Aravamudan int err; 3836a3437870SNishanth Aravamudan unsigned long input; 38379a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3838a3437870SNishanth Aravamudan 3839bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3840adbe8726SEric B Munson return -EINVAL; 3841adbe8726SEric B Munson 38423dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 3843a3437870SNishanth Aravamudan if (err) 384473ae31e5SEric B Munson return err; 3845a3437870SNishanth Aravamudan 3846db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3847a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 3848db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3849a3437870SNishanth Aravamudan 3850a3437870SNishanth Aravamudan return count; 3851a3437870SNishanth Aravamudan } 3852a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 3853a3437870SNishanth Aravamudan 3854a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 3855a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3856a3437870SNishanth Aravamudan { 38579a305230SLee Schermerhorn struct hstate *h; 38589a305230SLee Schermerhorn unsigned long free_huge_pages; 38599a305230SLee Schermerhorn int nid; 38609a305230SLee Schermerhorn 38619a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 38629a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 38639a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 38649a305230SLee Schermerhorn else 38659a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 38669a305230SLee Schermerhorn 3867ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", free_huge_pages); 3868a3437870SNishanth Aravamudan } 3869a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 3870a3437870SNishanth Aravamudan 3871a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 3872a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3873a3437870SNishanth Aravamudan { 38749a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3875ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); 3876a3437870SNishanth Aravamudan } 3877a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 3878a3437870SNishanth Aravamudan 3879a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 3880a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3881a3437870SNishanth Aravamudan { 38829a305230SLee Schermerhorn struct hstate *h; 38839a305230SLee Schermerhorn unsigned long surplus_huge_pages; 38849a305230SLee Schermerhorn int nid; 38859a305230SLee Schermerhorn 38869a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 38879a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 38889a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 38899a305230SLee Schermerhorn else 38909a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 38919a305230SLee Schermerhorn 3892ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", surplus_huge_pages); 3893a3437870SNishanth Aravamudan } 3894a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 3895a3437870SNishanth Aravamudan 389679dfc695SMike Kravetz static ssize_t demote_store(struct kobject *kobj, 389779dfc695SMike Kravetz struct kobj_attribute *attr, const char *buf, size_t len) 389879dfc695SMike Kravetz { 389979dfc695SMike Kravetz unsigned long nr_demote; 390079dfc695SMike Kravetz unsigned long nr_available; 390179dfc695SMike Kravetz nodemask_t nodes_allowed, *n_mask; 390279dfc695SMike Kravetz struct hstate *h; 39038eeda55fSLi zeming int err; 390479dfc695SMike Kravetz int nid; 390579dfc695SMike Kravetz 390679dfc695SMike Kravetz err = kstrtoul(buf, 10, &nr_demote); 390779dfc695SMike Kravetz if (err) 390879dfc695SMike Kravetz return err; 390979dfc695SMike Kravetz h = kobj_to_hstate(kobj, &nid); 391079dfc695SMike Kravetz 391179dfc695SMike Kravetz if (nid != NUMA_NO_NODE) { 391279dfc695SMike Kravetz init_nodemask_of_node(&nodes_allowed, nid); 391379dfc695SMike Kravetz n_mask = &nodes_allowed; 391479dfc695SMike Kravetz } else { 391579dfc695SMike Kravetz n_mask = &node_states[N_MEMORY]; 391679dfc695SMike Kravetz } 391779dfc695SMike Kravetz 391879dfc695SMike Kravetz /* Synchronize with other sysfs operations modifying huge pages */ 391979dfc695SMike Kravetz mutex_lock(&h->resize_lock); 392079dfc695SMike Kravetz spin_lock_irq(&hugetlb_lock); 392179dfc695SMike Kravetz 392279dfc695SMike Kravetz while (nr_demote) { 392379dfc695SMike Kravetz /* 392479dfc695SMike Kravetz * Check for available pages to demote each time thorough the 392579dfc695SMike Kravetz * loop as demote_pool_huge_page will drop hugetlb_lock. 392679dfc695SMike Kravetz */ 392779dfc695SMike Kravetz if (nid != NUMA_NO_NODE) 392879dfc695SMike Kravetz nr_available = h->free_huge_pages_node[nid]; 392979dfc695SMike Kravetz else 393079dfc695SMike Kravetz nr_available = h->free_huge_pages; 393179dfc695SMike Kravetz nr_available -= h->resv_huge_pages; 393279dfc695SMike Kravetz if (!nr_available) 393379dfc695SMike Kravetz break; 393479dfc695SMike Kravetz 393579dfc695SMike Kravetz err = demote_pool_huge_page(h, n_mask); 393679dfc695SMike Kravetz if (err) 393779dfc695SMike Kravetz break; 393879dfc695SMike Kravetz 393979dfc695SMike Kravetz nr_demote--; 394079dfc695SMike Kravetz } 394179dfc695SMike Kravetz 394279dfc695SMike Kravetz spin_unlock_irq(&hugetlb_lock); 394379dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 394479dfc695SMike Kravetz 394579dfc695SMike Kravetz if (err) 394679dfc695SMike Kravetz return err; 394779dfc695SMike Kravetz return len; 394879dfc695SMike Kravetz } 394979dfc695SMike Kravetz HSTATE_ATTR_WO(demote); 395079dfc695SMike Kravetz 395179dfc695SMike Kravetz static ssize_t demote_size_show(struct kobject *kobj, 395279dfc695SMike Kravetz struct kobj_attribute *attr, char *buf) 395379dfc695SMike Kravetz { 395412658abfSMiaohe Lin struct hstate *h = kobj_to_hstate(kobj, NULL); 395579dfc695SMike Kravetz unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K; 395679dfc695SMike Kravetz 395779dfc695SMike Kravetz return sysfs_emit(buf, "%lukB\n", demote_size); 395879dfc695SMike Kravetz } 395979dfc695SMike Kravetz 396079dfc695SMike Kravetz static ssize_t demote_size_store(struct kobject *kobj, 396179dfc695SMike Kravetz struct kobj_attribute *attr, 396279dfc695SMike Kravetz const char *buf, size_t count) 396379dfc695SMike Kravetz { 396479dfc695SMike Kravetz struct hstate *h, *demote_hstate; 396579dfc695SMike Kravetz unsigned long demote_size; 396679dfc695SMike Kravetz unsigned int demote_order; 396779dfc695SMike Kravetz 396879dfc695SMike Kravetz demote_size = (unsigned long)memparse(buf, NULL); 396979dfc695SMike Kravetz 397079dfc695SMike Kravetz demote_hstate = size_to_hstate(demote_size); 397179dfc695SMike Kravetz if (!demote_hstate) 397279dfc695SMike Kravetz return -EINVAL; 397379dfc695SMike Kravetz demote_order = demote_hstate->order; 3974a01f4390SMike Kravetz if (demote_order < HUGETLB_PAGE_ORDER) 3975a01f4390SMike Kravetz return -EINVAL; 397679dfc695SMike Kravetz 397779dfc695SMike Kravetz /* demote order must be smaller than hstate order */ 397812658abfSMiaohe Lin h = kobj_to_hstate(kobj, NULL); 397979dfc695SMike Kravetz if (demote_order >= h->order) 398079dfc695SMike Kravetz return -EINVAL; 398179dfc695SMike Kravetz 398279dfc695SMike Kravetz /* resize_lock synchronizes access to demote size and writes */ 398379dfc695SMike Kravetz mutex_lock(&h->resize_lock); 398479dfc695SMike Kravetz h->demote_order = demote_order; 398579dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 398679dfc695SMike Kravetz 398779dfc695SMike Kravetz return count; 398879dfc695SMike Kravetz } 398979dfc695SMike Kravetz HSTATE_ATTR(demote_size); 399079dfc695SMike Kravetz 3991a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 3992a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 3993a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 3994a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 3995a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 3996a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 399706808b08SLee Schermerhorn #ifdef CONFIG_NUMA 399806808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 399906808b08SLee Schermerhorn #endif 4000a3437870SNishanth Aravamudan NULL, 4001a3437870SNishanth Aravamudan }; 4002a3437870SNishanth Aravamudan 400367e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 4004a3437870SNishanth Aravamudan .attrs = hstate_attrs, 4005a3437870SNishanth Aravamudan }; 4006a3437870SNishanth Aravamudan 400779dfc695SMike Kravetz static struct attribute *hstate_demote_attrs[] = { 400879dfc695SMike Kravetz &demote_size_attr.attr, 400979dfc695SMike Kravetz &demote_attr.attr, 401079dfc695SMike Kravetz NULL, 401179dfc695SMike Kravetz }; 401279dfc695SMike Kravetz 401379dfc695SMike Kravetz static const struct attribute_group hstate_demote_attr_group = { 401479dfc695SMike Kravetz .attrs = hstate_demote_attrs, 401579dfc695SMike Kravetz }; 401679dfc695SMike Kravetz 4017094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 40189a305230SLee Schermerhorn struct kobject **hstate_kobjs, 401967e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 4020a3437870SNishanth Aravamudan { 4021a3437870SNishanth Aravamudan int retval; 4022972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 4023a3437870SNishanth Aravamudan 40249a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 40259a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 4026a3437870SNishanth Aravamudan return -ENOMEM; 4027a3437870SNishanth Aravamudan 40289a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 4029cc2205a6SMiaohe Lin if (retval) { 40309a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 4031cc2205a6SMiaohe Lin hstate_kobjs[hi] = NULL; 40323a6bdda0SMiaohe Lin return retval; 4033cc2205a6SMiaohe Lin } 4034a3437870SNishanth Aravamudan 403579dfc695SMike Kravetz if (h->demote_order) { 403601088a60SMiaohe Lin retval = sysfs_create_group(hstate_kobjs[hi], 403701088a60SMiaohe Lin &hstate_demote_attr_group); 403801088a60SMiaohe Lin if (retval) { 403979dfc695SMike Kravetz pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name); 404001088a60SMiaohe Lin sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group); 404101088a60SMiaohe Lin kobject_put(hstate_kobjs[hi]); 404201088a60SMiaohe Lin hstate_kobjs[hi] = NULL; 404301088a60SMiaohe Lin return retval; 404401088a60SMiaohe Lin } 404579dfc695SMike Kravetz } 404679dfc695SMike Kravetz 404701088a60SMiaohe Lin return 0; 4048a3437870SNishanth Aravamudan } 4049a3437870SNishanth Aravamudan 40509a305230SLee Schermerhorn #ifdef CONFIG_NUMA 4051a4a00b45SMuchun Song static bool hugetlb_sysfs_initialized __ro_after_init; 40529a305230SLee Schermerhorn 40539a305230SLee Schermerhorn /* 40549a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 405510fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 405610fbcf4cSKay Sievers * index of a node device or _hstate == node id. 405710fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 40589a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 40599a305230SLee Schermerhorn */ 40609a305230SLee Schermerhorn struct node_hstate { 40619a305230SLee Schermerhorn struct kobject *hugepages_kobj; 40629a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 40639a305230SLee Schermerhorn }; 4064b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 40659a305230SLee Schermerhorn 40669a305230SLee Schermerhorn /* 406710fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 40689a305230SLee Schermerhorn */ 40699a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 40709a305230SLee Schermerhorn &nr_hugepages_attr.attr, 40719a305230SLee Schermerhorn &free_hugepages_attr.attr, 40729a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 40739a305230SLee Schermerhorn NULL, 40749a305230SLee Schermerhorn }; 40759a305230SLee Schermerhorn 407667e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 40779a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 40789a305230SLee Schermerhorn }; 40799a305230SLee Schermerhorn 40809a305230SLee Schermerhorn /* 408110fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 40829a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 40839a305230SLee Schermerhorn */ 40849a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 40859a305230SLee Schermerhorn { 40869a305230SLee Schermerhorn int nid; 40879a305230SLee Schermerhorn 40889a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 40899a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 40909a305230SLee Schermerhorn int i; 40919a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 40929a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 40939a305230SLee Schermerhorn if (nidp) 40949a305230SLee Schermerhorn *nidp = nid; 40959a305230SLee Schermerhorn return &hstates[i]; 40969a305230SLee Schermerhorn } 40979a305230SLee Schermerhorn } 40989a305230SLee Schermerhorn 40999a305230SLee Schermerhorn BUG(); 41009a305230SLee Schermerhorn return NULL; 41019a305230SLee Schermerhorn } 41029a305230SLee Schermerhorn 41039a305230SLee Schermerhorn /* 410410fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 41059a305230SLee Schermerhorn * No-op if no hstate attributes attached. 41069a305230SLee Schermerhorn */ 4107a4a00b45SMuchun Song void hugetlb_unregister_node(struct node *node) 41089a305230SLee Schermerhorn { 41099a305230SLee Schermerhorn struct hstate *h; 411010fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 41119a305230SLee Schermerhorn 41129a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 41139b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 41149a305230SLee Schermerhorn 4115972dc4deSAneesh Kumar K.V for_each_hstate(h) { 4116972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 411701088a60SMiaohe Lin struct kobject *hstate_kobj = nhs->hstate_kobjs[idx]; 411801088a60SMiaohe Lin 411901088a60SMiaohe Lin if (!hstate_kobj) 412001088a60SMiaohe Lin continue; 412101088a60SMiaohe Lin if (h->demote_order) 412201088a60SMiaohe Lin sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group); 412301088a60SMiaohe Lin sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group); 412401088a60SMiaohe Lin kobject_put(hstate_kobj); 4125972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 4126972dc4deSAneesh Kumar K.V } 41279a305230SLee Schermerhorn 41289a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 41299a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 41309a305230SLee Schermerhorn } 41319a305230SLee Schermerhorn 41329a305230SLee Schermerhorn 41339a305230SLee Schermerhorn /* 413410fbcf4cSKay Sievers * Register hstate attributes for a single node device. 41359a305230SLee Schermerhorn * No-op if attributes already registered. 41369a305230SLee Schermerhorn */ 4137a4a00b45SMuchun Song void hugetlb_register_node(struct node *node) 41389a305230SLee Schermerhorn { 41399a305230SLee Schermerhorn struct hstate *h; 414010fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 41419a305230SLee Schermerhorn int err; 41429a305230SLee Schermerhorn 4143a4a00b45SMuchun Song if (!hugetlb_sysfs_initialized) 4144a4a00b45SMuchun Song return; 4145a4a00b45SMuchun Song 41469a305230SLee Schermerhorn if (nhs->hugepages_kobj) 41479a305230SLee Schermerhorn return; /* already allocated */ 41489a305230SLee Schermerhorn 41499a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 415010fbcf4cSKay Sievers &node->dev.kobj); 41519a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 41529a305230SLee Schermerhorn return; 41539a305230SLee Schermerhorn 41549a305230SLee Schermerhorn for_each_hstate(h) { 41559a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 41569a305230SLee Schermerhorn nhs->hstate_kobjs, 41579a305230SLee Schermerhorn &per_node_hstate_attr_group); 41589a305230SLee Schermerhorn if (err) { 4159282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s for node %d\n", 416010fbcf4cSKay Sievers h->name, node->dev.id); 41619a305230SLee Schermerhorn hugetlb_unregister_node(node); 41629a305230SLee Schermerhorn break; 41639a305230SLee Schermerhorn } 41649a305230SLee Schermerhorn } 41659a305230SLee Schermerhorn } 41669a305230SLee Schermerhorn 41679a305230SLee Schermerhorn /* 41689b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 416910fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 417010fbcf4cSKay Sievers * registered their associated device by this time. 41719a305230SLee Schermerhorn */ 41727d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 41739a305230SLee Schermerhorn { 41749a305230SLee Schermerhorn int nid; 41759a305230SLee Schermerhorn 4176a4a00b45SMuchun Song for_each_online_node(nid) 4177b958d4d0SMuchun Song hugetlb_register_node(node_devices[nid]); 41789a305230SLee Schermerhorn } 41799a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 41809a305230SLee Schermerhorn 41819a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 41829a305230SLee Schermerhorn { 41839a305230SLee Schermerhorn BUG(); 41849a305230SLee Schermerhorn if (nidp) 41859a305230SLee Schermerhorn *nidp = -1; 41869a305230SLee Schermerhorn return NULL; 41879a305230SLee Schermerhorn } 41889a305230SLee Schermerhorn 41899a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 41909a305230SLee Schermerhorn 41919a305230SLee Schermerhorn #endif 41929a305230SLee Schermerhorn 4193263b8998SMiaohe Lin #ifdef CONFIG_CMA 4194263b8998SMiaohe Lin static void __init hugetlb_cma_check(void); 4195263b8998SMiaohe Lin #else 4196263b8998SMiaohe Lin static inline __init void hugetlb_cma_check(void) 4197263b8998SMiaohe Lin { 4198263b8998SMiaohe Lin } 4199263b8998SMiaohe Lin #endif 4200263b8998SMiaohe Lin 4201a4a00b45SMuchun Song static void __init hugetlb_sysfs_init(void) 4202a4a00b45SMuchun Song { 4203a4a00b45SMuchun Song struct hstate *h; 4204a4a00b45SMuchun Song int err; 4205a4a00b45SMuchun Song 4206a4a00b45SMuchun Song hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 4207a4a00b45SMuchun Song if (!hugepages_kobj) 4208a4a00b45SMuchun Song return; 4209a4a00b45SMuchun Song 4210a4a00b45SMuchun Song for_each_hstate(h) { 4211a4a00b45SMuchun Song err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 4212a4a00b45SMuchun Song hstate_kobjs, &hstate_attr_group); 4213a4a00b45SMuchun Song if (err) 4214a4a00b45SMuchun Song pr_err("HugeTLB: Unable to add hstate %s", h->name); 4215a4a00b45SMuchun Song } 4216a4a00b45SMuchun Song 4217a4a00b45SMuchun Song #ifdef CONFIG_NUMA 4218a4a00b45SMuchun Song hugetlb_sysfs_initialized = true; 4219a4a00b45SMuchun Song #endif 4220a4a00b45SMuchun Song hugetlb_register_all_nodes(); 4221a4a00b45SMuchun Song } 4222a4a00b45SMuchun Song 4223962de548SKefeng Wang #ifdef CONFIG_SYSCTL 4224962de548SKefeng Wang static void hugetlb_sysctl_init(void); 4225962de548SKefeng Wang #else 4226962de548SKefeng Wang static inline void hugetlb_sysctl_init(void) { } 4227962de548SKefeng Wang #endif 4228962de548SKefeng Wang 4229a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 4230a3437870SNishanth Aravamudan { 42318382d914SDavidlohr Bueso int i; 42328382d914SDavidlohr Bueso 4233d6995da3SMike Kravetz BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < 4234d6995da3SMike Kravetz __NR_HPAGEFLAGS); 4235d6995da3SMike Kravetz 4236c2833a5bSMike Kravetz if (!hugepages_supported()) { 4237c2833a5bSMike Kravetz if (hugetlb_max_hstate || default_hstate_max_huge_pages) 4238c2833a5bSMike Kravetz pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); 42390ef89d25SBenjamin Herrenschmidt return 0; 4240d715cf80SLiam R. Howlett } 4241d715cf80SLiam R. Howlett 4242282f4214SMike Kravetz /* 4243282f4214SMike Kravetz * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some 4244282f4214SMike Kravetz * architectures depend on setup being done here. 4245282f4214SMike Kravetz */ 4246a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 4247282f4214SMike Kravetz if (!parsed_default_hugepagesz) { 4248282f4214SMike Kravetz /* 4249282f4214SMike Kravetz * If we did not parse a default huge page size, set 4250282f4214SMike Kravetz * default_hstate_idx to HPAGE_SIZE hstate. And, if the 4251282f4214SMike Kravetz * number of huge pages for this default size was implicitly 4252282f4214SMike Kravetz * specified, set that here as well. 4253282f4214SMike Kravetz * Note that the implicit setting will overwrite an explicit 4254282f4214SMike Kravetz * setting. A warning will be printed in this case. 4255282f4214SMike Kravetz */ 4256282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); 4257f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 4258282f4214SMike Kravetz if (default_hstate.max_huge_pages) { 4259282f4214SMike Kravetz char buf[32]; 4260282f4214SMike Kravetz 4261282f4214SMike Kravetz string_get_size(huge_page_size(&default_hstate), 4262282f4214SMike Kravetz 1, STRING_UNITS_2, buf, 32); 4263282f4214SMike Kravetz pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", 4264282f4214SMike Kravetz default_hstate.max_huge_pages, buf); 4265282f4214SMike Kravetz pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", 4266282f4214SMike Kravetz default_hstate_max_huge_pages); 4267282f4214SMike Kravetz } 4268282f4214SMike Kravetz default_hstate.max_huge_pages = 4269282f4214SMike Kravetz default_hstate_max_huge_pages; 4270b5389086SZhenguo Yao 42710a7a0f6fSPeng Liu for_each_online_node(i) 4272b5389086SZhenguo Yao default_hstate.max_huge_pages_node[i] = 4273b5389086SZhenguo Yao default_hugepages_in_node[i]; 4274282f4214SMike Kravetz } 4275f8b74815SVaishali Thakkar } 4276a3437870SNishanth Aravamudan 4277cf11e85fSRoman Gushchin hugetlb_cma_check(); 4278a3437870SNishanth Aravamudan hugetlb_init_hstates(); 4279aa888a74SAndi Kleen gather_bootmem_prealloc(); 4280a3437870SNishanth Aravamudan report_hugepages(); 4281a3437870SNishanth Aravamudan 4282a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 42837179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 4284962de548SKefeng Wang hugetlb_sysctl_init(); 42859a305230SLee Schermerhorn 42868382d914SDavidlohr Bueso #ifdef CONFIG_SMP 42878382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 42888382d914SDavidlohr Bueso #else 42898382d914SDavidlohr Bueso num_fault_mutexes = 1; 42908382d914SDavidlohr Bueso #endif 4291c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 42926da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 42936da2ec56SKees Cook GFP_KERNEL); 4294c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 42958382d914SDavidlohr Bueso 42968382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 4297c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 4298a3437870SNishanth Aravamudan return 0; 4299a3437870SNishanth Aravamudan } 43003e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 4301a3437870SNishanth Aravamudan 4302ae94da89SMike Kravetz /* Overwritten by architectures with more huge page sizes */ 4303ae94da89SMike Kravetz bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) 43049fee021dSVaishali Thakkar { 4305ae94da89SMike Kravetz return size == HPAGE_SIZE; 43069fee021dSVaishali Thakkar } 43079fee021dSVaishali Thakkar 4308d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 4309a3437870SNishanth Aravamudan { 4310a3437870SNishanth Aravamudan struct hstate *h; 43118faa8b07SAndi Kleen unsigned long i; 43128faa8b07SAndi Kleen 4313a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 4314a3437870SNishanth Aravamudan return; 4315a3437870SNishanth Aravamudan } 431647d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 4317a3437870SNishanth Aravamudan BUG_ON(order == 0); 431847d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 431929383967SMike Kravetz mutex_init(&h->resize_lock); 4320a3437870SNishanth Aravamudan h->order = order; 4321aca78307SMiaohe Lin h->mask = ~(huge_page_size(h) - 1); 43228faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 43238faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 43240edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 432554f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 432654f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 4327a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 4328c2c3a60aSMiaohe Lin huge_page_size(h)/SZ_1K); 43298faa8b07SAndi Kleen 4330a3437870SNishanth Aravamudan parsed_hstate = h; 4331a3437870SNishanth Aravamudan } 4332a3437870SNishanth Aravamudan 4333b5389086SZhenguo Yao bool __init __weak hugetlb_node_alloc_supported(void) 4334b5389086SZhenguo Yao { 4335b5389086SZhenguo Yao return true; 4336b5389086SZhenguo Yao } 4337f87442f4SPeng Liu 4338f87442f4SPeng Liu static void __init hugepages_clear_pages_in_node(void) 4339f87442f4SPeng Liu { 4340f87442f4SPeng Liu if (!hugetlb_max_hstate) { 4341f87442f4SPeng Liu default_hstate_max_huge_pages = 0; 4342f87442f4SPeng Liu memset(default_hugepages_in_node, 0, 434310395680SMiaohe Lin sizeof(default_hugepages_in_node)); 4344f87442f4SPeng Liu } else { 4345f87442f4SPeng Liu parsed_hstate->max_huge_pages = 0; 4346f87442f4SPeng Liu memset(parsed_hstate->max_huge_pages_node, 0, 434710395680SMiaohe Lin sizeof(parsed_hstate->max_huge_pages_node)); 4348f87442f4SPeng Liu } 4349f87442f4SPeng Liu } 4350f87442f4SPeng Liu 4351282f4214SMike Kravetz /* 4352282f4214SMike Kravetz * hugepages command line processing 4353282f4214SMike Kravetz * hugepages normally follows a valid hugepagsz or default_hugepagsz 4354282f4214SMike Kravetz * specification. If not, ignore the hugepages value. hugepages can also 4355282f4214SMike Kravetz * be the first huge page command line option in which case it implicitly 4356282f4214SMike Kravetz * specifies the number of huge pages for the default size. 4357282f4214SMike Kravetz */ 4358282f4214SMike Kravetz static int __init hugepages_setup(char *s) 4359a3437870SNishanth Aravamudan { 4360a3437870SNishanth Aravamudan unsigned long *mhp; 43618faa8b07SAndi Kleen static unsigned long *last_mhp; 4362b5389086SZhenguo Yao int node = NUMA_NO_NODE; 4363b5389086SZhenguo Yao int count; 4364b5389086SZhenguo Yao unsigned long tmp; 4365b5389086SZhenguo Yao char *p = s; 4366a3437870SNishanth Aravamudan 43679fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 4368282f4214SMike Kravetz pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); 43699fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 4370f81f6e4bSPeng Liu return 1; 43719fee021dSVaishali Thakkar } 4372282f4214SMike Kravetz 4373a3437870SNishanth Aravamudan /* 4374282f4214SMike Kravetz * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter 4375282f4214SMike Kravetz * yet, so this hugepages= parameter goes to the "default hstate". 4376282f4214SMike Kravetz * Otherwise, it goes with the previously parsed hugepagesz or 4377282f4214SMike Kravetz * default_hugepagesz. 4378a3437870SNishanth Aravamudan */ 43799fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 4380a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 4381a3437870SNishanth Aravamudan else 4382a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 4383a3437870SNishanth Aravamudan 43848faa8b07SAndi Kleen if (mhp == last_mhp) { 4385282f4214SMike Kravetz pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); 4386f81f6e4bSPeng Liu return 1; 43878faa8b07SAndi Kleen } 43888faa8b07SAndi Kleen 4389b5389086SZhenguo Yao while (*p) { 4390b5389086SZhenguo Yao count = 0; 4391b5389086SZhenguo Yao if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4392b5389086SZhenguo Yao goto invalid; 4393b5389086SZhenguo Yao /* Parameter is node format */ 4394b5389086SZhenguo Yao if (p[count] == ':') { 4395b5389086SZhenguo Yao if (!hugetlb_node_alloc_supported()) { 4396b5389086SZhenguo Yao pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n"); 4397f81f6e4bSPeng Liu return 1; 4398b5389086SZhenguo Yao } 43990a7a0f6fSPeng Liu if (tmp >= MAX_NUMNODES || !node_online(tmp)) 4400e79ce983SLiu Yuntao goto invalid; 44010a7a0f6fSPeng Liu node = array_index_nospec(tmp, MAX_NUMNODES); 4402b5389086SZhenguo Yao p += count + 1; 4403b5389086SZhenguo Yao /* Parse hugepages */ 4404b5389086SZhenguo Yao if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4405b5389086SZhenguo Yao goto invalid; 4406b5389086SZhenguo Yao if (!hugetlb_max_hstate) 4407b5389086SZhenguo Yao default_hugepages_in_node[node] = tmp; 4408b5389086SZhenguo Yao else 4409b5389086SZhenguo Yao parsed_hstate->max_huge_pages_node[node] = tmp; 4410b5389086SZhenguo Yao *mhp += tmp; 4411b5389086SZhenguo Yao /* Go to parse next node*/ 4412b5389086SZhenguo Yao if (p[count] == ',') 4413b5389086SZhenguo Yao p += count + 1; 4414b5389086SZhenguo Yao else 4415b5389086SZhenguo Yao break; 4416b5389086SZhenguo Yao } else { 4417b5389086SZhenguo Yao if (p != s) 4418b5389086SZhenguo Yao goto invalid; 4419b5389086SZhenguo Yao *mhp = tmp; 4420b5389086SZhenguo Yao break; 4421b5389086SZhenguo Yao } 4422b5389086SZhenguo Yao } 4423a3437870SNishanth Aravamudan 44248faa8b07SAndi Kleen /* 44258faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 442604adbc3fSMiaohe Lin * But we need to allocate gigantic hstates here early to still 44278faa8b07SAndi Kleen * use the bootmem allocator. 44288faa8b07SAndi Kleen */ 442904adbc3fSMiaohe Lin if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) 44308faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 44318faa8b07SAndi Kleen 44328faa8b07SAndi Kleen last_mhp = mhp; 44338faa8b07SAndi Kleen 4434a3437870SNishanth Aravamudan return 1; 4435b5389086SZhenguo Yao 4436b5389086SZhenguo Yao invalid: 4437b5389086SZhenguo Yao pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p); 4438f87442f4SPeng Liu hugepages_clear_pages_in_node(); 4439f81f6e4bSPeng Liu return 1; 4440a3437870SNishanth Aravamudan } 4441282f4214SMike Kravetz __setup("hugepages=", hugepages_setup); 4442e11bfbfcSNick Piggin 4443282f4214SMike Kravetz /* 4444282f4214SMike Kravetz * hugepagesz command line processing 4445282f4214SMike Kravetz * A specific huge page size can only be specified once with hugepagesz. 4446282f4214SMike Kravetz * hugepagesz is followed by hugepages on the command line. The global 4447282f4214SMike Kravetz * variable 'parsed_valid_hugepagesz' is used to determine if prior 4448282f4214SMike Kravetz * hugepagesz argument was valid. 4449282f4214SMike Kravetz */ 4450359f2544SMike Kravetz static int __init hugepagesz_setup(char *s) 4451e11bfbfcSNick Piggin { 4452359f2544SMike Kravetz unsigned long size; 4453282f4214SMike Kravetz struct hstate *h; 4454282f4214SMike Kravetz 4455282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4456359f2544SMike Kravetz size = (unsigned long)memparse(s, NULL); 4457359f2544SMike Kravetz 4458359f2544SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4459282f4214SMike Kravetz pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); 4460f81f6e4bSPeng Liu return 1; 4461359f2544SMike Kravetz } 4462359f2544SMike Kravetz 4463282f4214SMike Kravetz h = size_to_hstate(size); 4464282f4214SMike Kravetz if (h) { 4465282f4214SMike Kravetz /* 4466282f4214SMike Kravetz * hstate for this size already exists. This is normally 4467282f4214SMike Kravetz * an error, but is allowed if the existing hstate is the 4468282f4214SMike Kravetz * default hstate. More specifically, it is only allowed if 4469282f4214SMike Kravetz * the number of huge pages for the default hstate was not 4470282f4214SMike Kravetz * previously specified. 4471282f4214SMike Kravetz */ 4472282f4214SMike Kravetz if (!parsed_default_hugepagesz || h != &default_hstate || 4473282f4214SMike Kravetz default_hstate.max_huge_pages) { 4474282f4214SMike Kravetz pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); 4475f81f6e4bSPeng Liu return 1; 447638237830SMike Kravetz } 447738237830SMike Kravetz 4478282f4214SMike Kravetz /* 4479282f4214SMike Kravetz * No need to call hugetlb_add_hstate() as hstate already 4480282f4214SMike Kravetz * exists. But, do set parsed_hstate so that a following 4481282f4214SMike Kravetz * hugepages= parameter will be applied to this hstate. 4482282f4214SMike Kravetz */ 4483282f4214SMike Kravetz parsed_hstate = h; 4484282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4485e11bfbfcSNick Piggin return 1; 4486e11bfbfcSNick Piggin } 4487282f4214SMike Kravetz 4488359f2544SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4489282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4490359f2544SMike Kravetz return 1; 4491359f2544SMike Kravetz } 4492359f2544SMike Kravetz __setup("hugepagesz=", hugepagesz_setup); 4493359f2544SMike Kravetz 4494282f4214SMike Kravetz /* 4495282f4214SMike Kravetz * default_hugepagesz command line input 4496282f4214SMike Kravetz * Only one instance of default_hugepagesz allowed on command line. 4497282f4214SMike Kravetz */ 4498ae94da89SMike Kravetz static int __init default_hugepagesz_setup(char *s) 4499e11bfbfcSNick Piggin { 4500ae94da89SMike Kravetz unsigned long size; 4501b5389086SZhenguo Yao int i; 4502ae94da89SMike Kravetz 4503282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4504282f4214SMike Kravetz if (parsed_default_hugepagesz) { 4505282f4214SMike Kravetz pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); 4506f81f6e4bSPeng Liu return 1; 4507282f4214SMike Kravetz } 4508282f4214SMike Kravetz 4509282f4214SMike Kravetz size = (unsigned long)memparse(s, NULL); 4510282f4214SMike Kravetz 4511282f4214SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4512282f4214SMike Kravetz pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); 4513f81f6e4bSPeng Liu return 1; 4514282f4214SMike Kravetz } 4515282f4214SMike Kravetz 4516282f4214SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4517282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4518282f4214SMike Kravetz parsed_default_hugepagesz = true; 4519282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(size)); 4520282f4214SMike Kravetz 4521282f4214SMike Kravetz /* 4522282f4214SMike Kravetz * The number of default huge pages (for this size) could have been 4523282f4214SMike Kravetz * specified as the first hugetlb parameter: hugepages=X. If so, 4524282f4214SMike Kravetz * then default_hstate_max_huge_pages is set. If the default huge 452523baf831SKirill A. Shutemov * page size is gigantic (> MAX_ORDER), then the pages must be 4526282f4214SMike Kravetz * allocated here from bootmem allocator. 4527282f4214SMike Kravetz */ 4528282f4214SMike Kravetz if (default_hstate_max_huge_pages) { 4529282f4214SMike Kravetz default_hstate.max_huge_pages = default_hstate_max_huge_pages; 45300a7a0f6fSPeng Liu for_each_online_node(i) 4531b5389086SZhenguo Yao default_hstate.max_huge_pages_node[i] = 4532b5389086SZhenguo Yao default_hugepages_in_node[i]; 4533282f4214SMike Kravetz if (hstate_is_gigantic(&default_hstate)) 4534282f4214SMike Kravetz hugetlb_hstate_alloc_pages(&default_hstate); 4535282f4214SMike Kravetz default_hstate_max_huge_pages = 0; 4536282f4214SMike Kravetz } 4537282f4214SMike Kravetz 4538e11bfbfcSNick Piggin return 1; 4539e11bfbfcSNick Piggin } 4540ae94da89SMike Kravetz __setup("default_hugepagesz=", default_hugepagesz_setup); 4541a3437870SNishanth Aravamudan 4542d2226ebdSFeng Tang static nodemask_t *policy_mbind_nodemask(gfp_t gfp) 4543d2226ebdSFeng Tang { 4544d2226ebdSFeng Tang #ifdef CONFIG_NUMA 4545d2226ebdSFeng Tang struct mempolicy *mpol = get_task_policy(current); 4546d2226ebdSFeng Tang 4547d2226ebdSFeng Tang /* 4548d2226ebdSFeng Tang * Only enforce MPOL_BIND policy which overlaps with cpuset policy 4549d2226ebdSFeng Tang * (from policy_nodemask) specifically for hugetlb case 4550d2226ebdSFeng Tang */ 4551d2226ebdSFeng Tang if (mpol->mode == MPOL_BIND && 4552d2226ebdSFeng Tang (apply_policy_zone(mpol, gfp_zone(gfp)) && 4553d2226ebdSFeng Tang cpuset_nodemask_valid_mems_allowed(&mpol->nodes))) 4554d2226ebdSFeng Tang return &mpol->nodes; 4555d2226ebdSFeng Tang #endif 4556d2226ebdSFeng Tang return NULL; 4557d2226ebdSFeng Tang } 4558d2226ebdSFeng Tang 45598ca39e68SMuchun Song static unsigned int allowed_mems_nr(struct hstate *h) 45608a213460SNishanth Aravamudan { 45618a213460SNishanth Aravamudan int node; 45628a213460SNishanth Aravamudan unsigned int nr = 0; 4563d2226ebdSFeng Tang nodemask_t *mbind_nodemask; 45648ca39e68SMuchun Song unsigned int *array = h->free_huge_pages_node; 45658ca39e68SMuchun Song gfp_t gfp_mask = htlb_alloc_mask(h); 45668a213460SNishanth Aravamudan 4567d2226ebdSFeng Tang mbind_nodemask = policy_mbind_nodemask(gfp_mask); 45688ca39e68SMuchun Song for_each_node_mask(node, cpuset_current_mems_allowed) { 4569d2226ebdSFeng Tang if (!mbind_nodemask || node_isset(node, *mbind_nodemask)) 45708a213460SNishanth Aravamudan nr += array[node]; 45718ca39e68SMuchun Song } 45728a213460SNishanth Aravamudan 45738a213460SNishanth Aravamudan return nr; 45748a213460SNishanth Aravamudan } 45758a213460SNishanth Aravamudan 45768a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 457717743798SMuchun Song static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, 457817743798SMuchun Song void *buffer, size_t *length, 457917743798SMuchun Song loff_t *ppos, unsigned long *out) 458017743798SMuchun Song { 458117743798SMuchun Song struct ctl_table dup_table; 458217743798SMuchun Song 458317743798SMuchun Song /* 458417743798SMuchun Song * In order to avoid races with __do_proc_doulongvec_minmax(), we 458517743798SMuchun Song * can duplicate the @table and alter the duplicate of it. 458617743798SMuchun Song */ 458717743798SMuchun Song dup_table = *table; 458817743798SMuchun Song dup_table.data = out; 458917743798SMuchun Song 459017743798SMuchun Song return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); 459117743798SMuchun Song } 459217743798SMuchun Song 459306808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 459406808b08SLee Schermerhorn struct ctl_table *table, int write, 459532927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 45961da177e4SLinus Torvalds { 4597e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 4598238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 459908d4a246SMichal Hocko int ret; 4600e5ff2159SAndi Kleen 4601457c1b27SNishanth Aravamudan if (!hugepages_supported()) 460286613628SJan Stancek return -EOPNOTSUPP; 4603457c1b27SNishanth Aravamudan 460417743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 460517743798SMuchun Song &tmp); 460608d4a246SMichal Hocko if (ret) 460708d4a246SMichal Hocko goto out; 4608e5ff2159SAndi Kleen 4609238d3c13SDavid Rientjes if (write) 4610238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 4611238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 461208d4a246SMichal Hocko out: 461308d4a246SMichal Hocko return ret; 46141da177e4SLinus Torvalds } 4615396faf03SMel Gorman 4616962de548SKefeng Wang static int hugetlb_sysctl_handler(struct ctl_table *table, int write, 461732927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 461806808b08SLee Schermerhorn { 461906808b08SLee Schermerhorn 462006808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 462106808b08SLee Schermerhorn buffer, length, ppos); 462206808b08SLee Schermerhorn } 462306808b08SLee Schermerhorn 462406808b08SLee Schermerhorn #ifdef CONFIG_NUMA 4625962de548SKefeng Wang static int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 462632927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 462706808b08SLee Schermerhorn { 462806808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 462906808b08SLee Schermerhorn buffer, length, ppos); 463006808b08SLee Schermerhorn } 463106808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 463206808b08SLee Schermerhorn 4633962de548SKefeng Wang static int hugetlb_overcommit_handler(struct ctl_table *table, int write, 463432927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 4635a3d0c6aaSNishanth Aravamudan { 4636a5516438SAndi Kleen struct hstate *h = &default_hstate; 4637e5ff2159SAndi Kleen unsigned long tmp; 463808d4a246SMichal Hocko int ret; 4639e5ff2159SAndi Kleen 4640457c1b27SNishanth Aravamudan if (!hugepages_supported()) 464186613628SJan Stancek return -EOPNOTSUPP; 4642457c1b27SNishanth Aravamudan 4643e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 4644e5ff2159SAndi Kleen 4645bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 4646adbe8726SEric B Munson return -EINVAL; 4647adbe8726SEric B Munson 464817743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 464917743798SMuchun Song &tmp); 465008d4a246SMichal Hocko if (ret) 465108d4a246SMichal Hocko goto out; 4652e5ff2159SAndi Kleen 4653e5ff2159SAndi Kleen if (write) { 4654db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4655e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 4656db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4657e5ff2159SAndi Kleen } 465808d4a246SMichal Hocko out: 465908d4a246SMichal Hocko return ret; 4660a3d0c6aaSNishanth Aravamudan } 4661a3d0c6aaSNishanth Aravamudan 4662962de548SKefeng Wang static struct ctl_table hugetlb_table[] = { 4663962de548SKefeng Wang { 4664962de548SKefeng Wang .procname = "nr_hugepages", 4665962de548SKefeng Wang .data = NULL, 4666962de548SKefeng Wang .maxlen = sizeof(unsigned long), 4667962de548SKefeng Wang .mode = 0644, 4668962de548SKefeng Wang .proc_handler = hugetlb_sysctl_handler, 4669962de548SKefeng Wang }, 4670962de548SKefeng Wang #ifdef CONFIG_NUMA 4671962de548SKefeng Wang { 4672962de548SKefeng Wang .procname = "nr_hugepages_mempolicy", 4673962de548SKefeng Wang .data = NULL, 4674962de548SKefeng Wang .maxlen = sizeof(unsigned long), 4675962de548SKefeng Wang .mode = 0644, 4676962de548SKefeng Wang .proc_handler = &hugetlb_mempolicy_sysctl_handler, 4677962de548SKefeng Wang }, 4678962de548SKefeng Wang #endif 4679962de548SKefeng Wang { 4680962de548SKefeng Wang .procname = "hugetlb_shm_group", 4681962de548SKefeng Wang .data = &sysctl_hugetlb_shm_group, 4682962de548SKefeng Wang .maxlen = sizeof(gid_t), 4683962de548SKefeng Wang .mode = 0644, 4684962de548SKefeng Wang .proc_handler = proc_dointvec, 4685962de548SKefeng Wang }, 4686962de548SKefeng Wang { 4687962de548SKefeng Wang .procname = "nr_overcommit_hugepages", 4688962de548SKefeng Wang .data = NULL, 4689962de548SKefeng Wang .maxlen = sizeof(unsigned long), 4690962de548SKefeng Wang .mode = 0644, 4691962de548SKefeng Wang .proc_handler = hugetlb_overcommit_handler, 4692962de548SKefeng Wang }, 4693962de548SKefeng Wang { } 4694962de548SKefeng Wang }; 4695962de548SKefeng Wang 4696962de548SKefeng Wang static void hugetlb_sysctl_init(void) 4697962de548SKefeng Wang { 4698962de548SKefeng Wang register_sysctl_init("vm", hugetlb_table); 4699962de548SKefeng Wang } 47001da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 47011da177e4SLinus Torvalds 4702e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 47031da177e4SLinus Torvalds { 4704fcb2b0c5SRoman Gushchin struct hstate *h; 4705fcb2b0c5SRoman Gushchin unsigned long total = 0; 4706fcb2b0c5SRoman Gushchin 4707457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4708457c1b27SNishanth Aravamudan return; 4709fcb2b0c5SRoman Gushchin 4710fcb2b0c5SRoman Gushchin for_each_hstate(h) { 4711fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 4712fcb2b0c5SRoman Gushchin 4713aca78307SMiaohe Lin total += huge_page_size(h) * count; 4714fcb2b0c5SRoman Gushchin 4715fcb2b0c5SRoman Gushchin if (h == &default_hstate) 4716e1759c21SAlexey Dobriyan seq_printf(m, 47171da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 47181da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 4719b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 47207893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 47214f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 4722fcb2b0c5SRoman Gushchin count, 4723a5516438SAndi Kleen h->free_huge_pages, 4724a5516438SAndi Kleen h->resv_huge_pages, 4725a5516438SAndi Kleen h->surplus_huge_pages, 4726aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4727fcb2b0c5SRoman Gushchin } 4728fcb2b0c5SRoman Gushchin 4729aca78307SMiaohe Lin seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); 47301da177e4SLinus Torvalds } 47311da177e4SLinus Torvalds 47327981593bSJoe Perches int hugetlb_report_node_meminfo(char *buf, int len, int nid) 47331da177e4SLinus Torvalds { 4734a5516438SAndi Kleen struct hstate *h = &default_hstate; 47357981593bSJoe Perches 4736457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4737457c1b27SNishanth Aravamudan return 0; 47387981593bSJoe Perches 47397981593bSJoe Perches return sysfs_emit_at(buf, len, 47401da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 4741a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 4742a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 4743a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 4744a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 4745a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 47461da177e4SLinus Torvalds } 47471da177e4SLinus Torvalds 4748dcadcf1cSGang Li void hugetlb_show_meminfo_node(int nid) 4749949f7ec5SDavid Rientjes { 4750949f7ec5SDavid Rientjes struct hstate *h; 4751949f7ec5SDavid Rientjes 4752457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4753457c1b27SNishanth Aravamudan return; 4754457c1b27SNishanth Aravamudan 4755949f7ec5SDavid Rientjes for_each_hstate(h) 4756dcadcf1cSGang Li printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 4757949f7ec5SDavid Rientjes nid, 4758949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 4759949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 4760949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 4761aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4762949f7ec5SDavid Rientjes } 4763949f7ec5SDavid Rientjes 47645d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 47655d317b2bSNaoya Horiguchi { 47665d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 47676c1aa2d3SZhangPeng K(atomic_long_read(&mm->hugetlb_usage))); 47685d317b2bSNaoya Horiguchi } 47695d317b2bSNaoya Horiguchi 47701da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 47711da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 47721da177e4SLinus Torvalds { 4773d0028588SWanpeng Li struct hstate *h; 4774d0028588SWanpeng Li unsigned long nr_total_pages = 0; 4775d0028588SWanpeng Li 4776d0028588SWanpeng Li for_each_hstate(h) 4777d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 4778d0028588SWanpeng Li return nr_total_pages; 47791da177e4SLinus Torvalds } 47801da177e4SLinus Torvalds 4781a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 4782fc1b8a73SMel Gorman { 4783fc1b8a73SMel Gorman int ret = -ENOMEM; 4784fc1b8a73SMel Gorman 47850aa7f354SMiaohe Lin if (!delta) 47860aa7f354SMiaohe Lin return 0; 47870aa7f354SMiaohe Lin 4788db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4789fc1b8a73SMel Gorman /* 4790fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 4791fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 4792fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 4793fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 4794fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 4795fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 4796fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 4797fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 4798fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 4799fc1b8a73SMel Gorman * 4800fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 4801fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 4802fc1b8a73SMel Gorman * we fall back to check against current free page availability as 4803fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 4804fc1b8a73SMel Gorman * semantics that cpuset has. 48058ca39e68SMuchun Song * 48068ca39e68SMuchun Song * Apart from cpuset, we also have memory policy mechanism that 48078ca39e68SMuchun Song * also determines from which node the kernel will allocate memory 48088ca39e68SMuchun Song * in a NUMA system. So similar to cpuset, we also should consider 48098ca39e68SMuchun Song * the memory policy of the current task. Similar to the description 48108ca39e68SMuchun Song * above. 4811fc1b8a73SMel Gorman */ 4812fc1b8a73SMel Gorman if (delta > 0) { 4813a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 4814fc1b8a73SMel Gorman goto out; 4815fc1b8a73SMel Gorman 48168ca39e68SMuchun Song if (delta > allowed_mems_nr(h)) { 4817a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 4818fc1b8a73SMel Gorman goto out; 4819fc1b8a73SMel Gorman } 4820fc1b8a73SMel Gorman } 4821fc1b8a73SMel Gorman 4822fc1b8a73SMel Gorman ret = 0; 4823fc1b8a73SMel Gorman if (delta < 0) 4824a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 4825fc1b8a73SMel Gorman 4826fc1b8a73SMel Gorman out: 4827db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4828fc1b8a73SMel Gorman return ret; 4829fc1b8a73SMel Gorman } 4830fc1b8a73SMel Gorman 483184afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 483284afd99bSAndy Whitcroft { 4833f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 483484afd99bSAndy Whitcroft 483584afd99bSAndy Whitcroft /* 4836612b8a31SMike Kravetz * HPAGE_RESV_OWNER indicates a private mapping. 483784afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 483884afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 483984afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 484025985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 484184afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 484284afd99bSAndy Whitcroft * new reference here without additional locking. 484384afd99bSAndy Whitcroft */ 484409a26e83SMike Kravetz if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 484509a26e83SMike Kravetz resv_map_dup_hugetlb_cgroup_uncharge_info(resv); 4846f522c3acSJoonsoo Kim kref_get(&resv->refs); 484784afd99bSAndy Whitcroft } 48488d9bfb26SMike Kravetz 4849131a79b4SMike Kravetz /* 4850131a79b4SMike Kravetz * vma_lock structure for sharable mappings is vma specific. 4851612b8a31SMike Kravetz * Clear old pointer (if copied via vm_area_dup) and allocate 4852612b8a31SMike Kravetz * new structure. Before clearing, make sure vma_lock is not 4853612b8a31SMike Kravetz * for this vma. 4854131a79b4SMike Kravetz */ 4855131a79b4SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 4856612b8a31SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 4857612b8a31SMike Kravetz 4858612b8a31SMike Kravetz if (vma_lock) { 4859612b8a31SMike Kravetz if (vma_lock->vma != vma) { 4860131a79b4SMike Kravetz vma->vm_private_data = NULL; 48618d9bfb26SMike Kravetz hugetlb_vma_lock_alloc(vma); 4862612b8a31SMike Kravetz } else 4863612b8a31SMike Kravetz pr_warn("HugeTLB: vma_lock already exists in %s.\n", __func__); 4864612b8a31SMike Kravetz } else 4865612b8a31SMike Kravetz hugetlb_vma_lock_alloc(vma); 486609a26e83SMike Kravetz } 4867131a79b4SMike Kravetz } 486884afd99bSAndy Whitcroft 4869a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 4870a1e78772SMel Gorman { 4871a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 48728d9bfb26SMike Kravetz struct resv_map *resv; 487390481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 48744e35f483SJoonsoo Kim unsigned long reserve, start, end; 48751c5ecae3SMike Kravetz long gbl_reserve; 487684afd99bSAndy Whitcroft 48778d9bfb26SMike Kravetz hugetlb_vma_lock_free(vma); 48788d9bfb26SMike Kravetz 48798d9bfb26SMike Kravetz resv = vma_resv_map(vma); 48804e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 48814e35f483SJoonsoo Kim return; 48824e35f483SJoonsoo Kim 4883a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 4884a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 488584afd99bSAndy Whitcroft 48864e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 4887e9fe92aeSMina Almasry hugetlb_cgroup_uncharge_counter(resv, start, end); 48887251ff78SAdam Litke if (reserve) { 48891c5ecae3SMike Kravetz /* 48901c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 48911c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 48921c5ecae3SMike Kravetz */ 48931c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 48941c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 48957251ff78SAdam Litke } 4896e9fe92aeSMina Almasry 4897e9fe92aeSMina Almasry kref_put(&resv->refs, resv_map_release); 4898a1e78772SMel Gorman } 4899a1e78772SMel Gorman 490031383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 490131383c68SDan Williams { 490231383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 490331383c68SDan Williams return -EINVAL; 4904b30c14cdSJames Houghton 4905b30c14cdSJames Houghton /* 4906b30c14cdSJames Houghton * PMD sharing is only possible for PUD_SIZE-aligned address ranges 4907b30c14cdSJames Houghton * in HugeTLB VMAs. If we will lose PUD_SIZE alignment due to this 4908b30c14cdSJames Houghton * split, unshare PMDs in the PUD_SIZE interval surrounding addr now. 4909b30c14cdSJames Houghton */ 4910b30c14cdSJames Houghton if (addr & ~PUD_MASK) { 4911b30c14cdSJames Houghton /* 4912b30c14cdSJames Houghton * hugetlb_vm_op_split is called right before we attempt to 4913b30c14cdSJames Houghton * split the VMA. We will need to unshare PMDs in the old and 4914b30c14cdSJames Houghton * new VMAs, so let's unshare before we split. 4915b30c14cdSJames Houghton */ 4916b30c14cdSJames Houghton unsigned long floor = addr & PUD_MASK; 4917b30c14cdSJames Houghton unsigned long ceil = floor + PUD_SIZE; 4918b30c14cdSJames Houghton 4919b30c14cdSJames Houghton if (floor >= vma->vm_start && ceil <= vma->vm_end) 4920b30c14cdSJames Houghton hugetlb_unshare_pmds(vma, floor, ceil); 4921b30c14cdSJames Houghton } 4922b30c14cdSJames Houghton 492331383c68SDan Williams return 0; 492431383c68SDan Williams } 492531383c68SDan Williams 492605ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 492705ea8860SDan Williams { 4928aca78307SMiaohe Lin return huge_page_size(hstate_vma(vma)); 492905ea8860SDan Williams } 493005ea8860SDan Williams 49311da177e4SLinus Torvalds /* 49321da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 49331da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 49346c26d310SMiaohe Lin * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get 49351da177e4SLinus Torvalds * this far. 49361da177e4SLinus Torvalds */ 4937b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 49381da177e4SLinus Torvalds { 49391da177e4SLinus Torvalds BUG(); 4940d0217ac0SNick Piggin return 0; 49411da177e4SLinus Torvalds } 49421da177e4SLinus Torvalds 4943eec3636aSJane Chu /* 4944eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 4945eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 4946eec3636aSJane Chu * This is because under System V memory model, mappings created via 4947eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 4948eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 4949eec3636aSJane Chu */ 4950f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 4951d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 495284afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 4953a1e78772SMel Gorman .close = hugetlb_vm_op_close, 4954dd3b614fSDmitry Safonov .may_split = hugetlb_vm_op_split, 495505ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 49561da177e4SLinus Torvalds }; 49571da177e4SLinus Torvalds 49581e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 49591e8f889bSDavid Gibson int writable) 496063551ae0SDavid Gibson { 496163551ae0SDavid Gibson pte_t entry; 496279c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 496363551ae0SDavid Gibson 49641e8f889bSDavid Gibson if (writable) { 4965106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 4966106c992aSGerald Schaefer vma->vm_page_prot))); 496763551ae0SDavid Gibson } else { 4968106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 4969106c992aSGerald Schaefer vma->vm_page_prot)); 497063551ae0SDavid Gibson } 497163551ae0SDavid Gibson entry = pte_mkyoung(entry); 497279c1c594SChristophe Leroy entry = arch_make_huge_pte(entry, shift, vma->vm_flags); 497363551ae0SDavid Gibson 497463551ae0SDavid Gibson return entry; 497563551ae0SDavid Gibson } 497663551ae0SDavid Gibson 49771e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 49781e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 49791e8f889bSDavid Gibson { 49801e8f889bSDavid Gibson pte_t entry; 49811e8f889bSDavid Gibson 4982106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 498332f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 49844b3073e1SRussell King update_mmu_cache(vma, address, ptep); 49851e8f889bSDavid Gibson } 49861e8f889bSDavid Gibson 4987d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 49884a705fefSNaoya Horiguchi { 49894a705fefSNaoya Horiguchi swp_entry_t swp; 49904a705fefSNaoya Horiguchi 49914a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 4992d5ed7444SAneesh Kumar K.V return false; 49934a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4994d79d176aSBaoquan He if (is_migration_entry(swp)) 4995d5ed7444SAneesh Kumar K.V return true; 49964a705fefSNaoya Horiguchi else 4997d5ed7444SAneesh Kumar K.V return false; 49984a705fefSNaoya Horiguchi } 49994a705fefSNaoya Horiguchi 50003e5c3600SBaoquan He static bool is_hugetlb_entry_hwpoisoned(pte_t pte) 50014a705fefSNaoya Horiguchi { 50024a705fefSNaoya Horiguchi swp_entry_t swp; 50034a705fefSNaoya Horiguchi 50044a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 50053e5c3600SBaoquan He return false; 50064a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 5007d79d176aSBaoquan He if (is_hwpoison_entry(swp)) 50083e5c3600SBaoquan He return true; 50094a705fefSNaoya Horiguchi else 50103e5c3600SBaoquan He return false; 50114a705fefSNaoya Horiguchi } 50121e8f889bSDavid Gibson 50134eae4efaSPeter Xu static void 5014ea4c353dSSidhartha Kumar hugetlb_install_folio(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, 5015935d4f0cSRyan Roberts struct folio *new_folio, pte_t old, unsigned long sz) 50164eae4efaSPeter Xu { 50175a2f8d22SPeter Xu pte_t newpte = make_huge_pte(vma, &new_folio->page, 1); 50185a2f8d22SPeter Xu 5019ea4c353dSSidhartha Kumar __folio_mark_uptodate(new_folio); 5020d0ce0e47SSidhartha Kumar hugepage_add_new_anon_rmap(new_folio, vma, addr); 50215a2f8d22SPeter Xu if (userfaultfd_wp(vma) && huge_pte_uffd_wp(old)) 50225a2f8d22SPeter Xu newpte = huge_pte_mkuffd_wp(newpte); 5023935d4f0cSRyan Roberts set_huge_pte_at(vma->vm_mm, addr, ptep, newpte, sz); 50244eae4efaSPeter Xu hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); 5025ea4c353dSSidhartha Kumar folio_set_hugetlb_migratable(new_folio); 50264eae4efaSPeter Xu } 50274eae4efaSPeter Xu 502863551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 5029bc70fbf2SPeter Xu struct vm_area_struct *dst_vma, 5030bc70fbf2SPeter Xu struct vm_area_struct *src_vma) 503163551ae0SDavid Gibson { 50323aa4ed80SMiaohe Lin pte_t *src_pte, *dst_pte, entry; 5033ad27ce20SZhangPeng struct folio *pte_folio; 50341c59827dSHugh Dickins unsigned long addr; 5035bc70fbf2SPeter Xu bool cow = is_cow_mapping(src_vma->vm_flags); 5036bc70fbf2SPeter Xu struct hstate *h = hstate_vma(src_vma); 5037a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 50384eae4efaSPeter Xu unsigned long npages = pages_per_huge_page(h); 5039ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 5040e95a9851SMike Kravetz unsigned long last_addr_mask; 5041e8569dd2SAndreas Sandberg int ret = 0; 50421e8f889bSDavid Gibson 5043ac46d4f3SJérôme Glisse if (cow) { 50447d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src, 5045bc70fbf2SPeter Xu src_vma->vm_start, 5046bc70fbf2SPeter Xu src_vma->vm_end); 5047ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 5048e727bfd5SSuren Baghdasaryan vma_assert_write_locked(src_vma); 5049623a1ddfSDavid Hildenbrand raw_write_seqcount_begin(&src->write_protect_seq); 505040549ba8SMike Kravetz } else { 505140549ba8SMike Kravetz /* 505240549ba8SMike Kravetz * For shared mappings the vma lock must be held before 50539c67a207SPeter Xu * calling hugetlb_walk() in the src vma. Otherwise, the 505440549ba8SMike Kravetz * returned ptep could go away if part of a shared pmd and 505540549ba8SMike Kravetz * another thread calls huge_pmd_unshare. 505640549ba8SMike Kravetz */ 505740549ba8SMike Kravetz hugetlb_vma_lock_read(src_vma); 5058ac46d4f3SJérôme Glisse } 5059e8569dd2SAndreas Sandberg 5060e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 5061bc70fbf2SPeter Xu for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) { 5062cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 50639c67a207SPeter Xu src_pte = hugetlb_walk(src_vma, addr, sz); 5064e95a9851SMike Kravetz if (!src_pte) { 5065e95a9851SMike Kravetz addr |= last_addr_mask; 5066c74df32cSHugh Dickins continue; 5067e95a9851SMike Kravetz } 5068bc70fbf2SPeter Xu dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz); 5069e8569dd2SAndreas Sandberg if (!dst_pte) { 5070e8569dd2SAndreas Sandberg ret = -ENOMEM; 5071e8569dd2SAndreas Sandberg break; 5072e8569dd2SAndreas Sandberg } 5073c5c99429SLarry Woodman 50745e41540cSMike Kravetz /* 50755e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 50765e41540cSMike Kravetz * 50773aa4ed80SMiaohe Lin * dst_pte == src_pte is the common case of src/dest sharing. 50785e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 50793aa4ed80SMiaohe Lin * another vma. So page_count of ptep page is checked instead 50803aa4ed80SMiaohe Lin * to reliably determine whether pte is shared. 50815e41540cSMike Kravetz */ 50823aa4ed80SMiaohe Lin if (page_count(virt_to_page(dst_pte)) > 1) { 5083e95a9851SMike Kravetz addr |= last_addr_mask; 5084c5c99429SLarry Woodman continue; 5085e95a9851SMike Kravetz } 5086c5c99429SLarry Woodman 5087cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 5088cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 5089cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 50904a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 50914eae4efaSPeter Xu again: 50923aa4ed80SMiaohe Lin if (huge_pte_none(entry)) { 50935e41540cSMike Kravetz /* 50943aa4ed80SMiaohe Lin * Skip if src entry none. 50955e41540cSMike Kravetz */ 50964a705fefSNaoya Horiguchi ; 5097c2cb0dccSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) { 50985a2f8d22SPeter Xu if (!userfaultfd_wp(dst_vma)) 5099c2cb0dccSNaoya Horiguchi entry = huge_pte_clear_uffd_wp(entry); 5100935d4f0cSRyan Roberts set_huge_pte_at(dst, addr, dst_pte, entry, sz); 5101c2cb0dccSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry))) { 51024a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 51035a2f8d22SPeter Xu bool uffd_wp = pte_swp_uffd_wp(entry); 51044a705fefSNaoya Horiguchi 51056c287605SDavid Hildenbrand if (!is_readable_migration_entry(swp_entry) && cow) { 51064a705fefSNaoya Horiguchi /* 51074a705fefSNaoya Horiguchi * COW mappings require pages in both 51084a705fefSNaoya Horiguchi * parent and child to be set to read. 51094a705fefSNaoya Horiguchi */ 51104dd845b5SAlistair Popple swp_entry = make_readable_migration_entry( 51114dd845b5SAlistair Popple swp_offset(swp_entry)); 51124a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 5113bc70fbf2SPeter Xu if (userfaultfd_wp(src_vma) && uffd_wp) 51145a2f8d22SPeter Xu entry = pte_swp_mkuffd_wp(entry); 5115935d4f0cSRyan Roberts set_huge_pte_at(src, addr, src_pte, entry, sz); 51164a705fefSNaoya Horiguchi } 51175a2f8d22SPeter Xu if (!userfaultfd_wp(dst_vma)) 5118bc70fbf2SPeter Xu entry = huge_pte_clear_uffd_wp(entry); 5119935d4f0cSRyan Roberts set_huge_pte_at(dst, addr, dst_pte, entry, sz); 5120bc70fbf2SPeter Xu } else if (unlikely(is_pte_marker(entry))) { 5121af19487fSAxel Rasmussen pte_marker marker = copy_pte_marker( 5122af19487fSAxel Rasmussen pte_to_swp_entry(entry), dst_vma); 5123af19487fSAxel Rasmussen 5124af19487fSAxel Rasmussen if (marker) 5125af19487fSAxel Rasmussen set_huge_pte_at(dst, addr, dst_pte, 5126935d4f0cSRyan Roberts make_pte_marker(marker), sz); 51274a705fefSNaoya Horiguchi } else { 51284eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 5129ad27ce20SZhangPeng pte_folio = page_folio(pte_page(entry)); 5130ad27ce20SZhangPeng folio_get(pte_folio); 51314eae4efaSPeter Xu 51324eae4efaSPeter Xu /* 5133fb3d824dSDavid Hildenbrand * Failing to duplicate the anon rmap is a rare case 5134fb3d824dSDavid Hildenbrand * where we see pinned hugetlb pages while they're 5135fb3d824dSDavid Hildenbrand * prone to COW. We need to do the COW earlier during 5136fb3d824dSDavid Hildenbrand * fork. 51374eae4efaSPeter Xu * 51384eae4efaSPeter Xu * When pre-allocating the page or copying data, we 51394eae4efaSPeter Xu * need to be without the pgtable locks since we could 51404eae4efaSPeter Xu * sleep during the process. 51414eae4efaSPeter Xu */ 5142ad27ce20SZhangPeng if (!folio_test_anon(pte_folio)) { 5143ad27ce20SZhangPeng page_dup_file_rmap(&pte_folio->page, true); 5144ad27ce20SZhangPeng } else if (page_try_dup_anon_rmap(&pte_folio->page, 5145ad27ce20SZhangPeng true, src_vma)) { 51464eae4efaSPeter Xu pte_t src_pte_old = entry; 5147d0ce0e47SSidhartha Kumar struct folio *new_folio; 51484eae4efaSPeter Xu 51494eae4efaSPeter Xu spin_unlock(src_ptl); 51504eae4efaSPeter Xu spin_unlock(dst_ptl); 51514eae4efaSPeter Xu /* Do not use reserve as it's private owned */ 5152d0ce0e47SSidhartha Kumar new_folio = alloc_hugetlb_folio(dst_vma, addr, 1); 5153d0ce0e47SSidhartha Kumar if (IS_ERR(new_folio)) { 5154ad27ce20SZhangPeng folio_put(pte_folio); 5155d0ce0e47SSidhartha Kumar ret = PTR_ERR(new_folio); 51564eae4efaSPeter Xu break; 51574eae4efaSPeter Xu } 51581cb9dc4bSLiu Shixin ret = copy_user_large_folio(new_folio, 5159ad27ce20SZhangPeng pte_folio, 5160c0e8150eSZhangPeng addr, dst_vma); 5161ad27ce20SZhangPeng folio_put(pte_folio); 51621cb9dc4bSLiu Shixin if (ret) { 51631cb9dc4bSLiu Shixin folio_put(new_folio); 51641cb9dc4bSLiu Shixin break; 51651cb9dc4bSLiu Shixin } 51664eae4efaSPeter Xu 5167d0ce0e47SSidhartha Kumar /* Install the new hugetlb folio if src pte stable */ 51684eae4efaSPeter Xu dst_ptl = huge_pte_lock(h, dst, dst_pte); 51694eae4efaSPeter Xu src_ptl = huge_pte_lockptr(h, src, src_pte); 51704eae4efaSPeter Xu spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 51714eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 51724eae4efaSPeter Xu if (!pte_same(src_pte_old, entry)) { 5173bc70fbf2SPeter Xu restore_reserve_on_error(h, dst_vma, addr, 5174d2d7bb44SSidhartha Kumar new_folio); 5175d0ce0e47SSidhartha Kumar folio_put(new_folio); 51763aa4ed80SMiaohe Lin /* huge_ptep of dst_pte won't change as in child */ 51774eae4efaSPeter Xu goto again; 51784eae4efaSPeter Xu } 51795a2f8d22SPeter Xu hugetlb_install_folio(dst_vma, dst_pte, addr, 5180935d4f0cSRyan Roberts new_folio, src_pte_old, sz); 51814eae4efaSPeter Xu spin_unlock(src_ptl); 51824eae4efaSPeter Xu spin_unlock(dst_ptl); 51834eae4efaSPeter Xu continue; 51844eae4efaSPeter Xu } 51854eae4efaSPeter Xu 518634ee645eSJoerg Roedel if (cow) { 51870f10851eSJérôme Glisse /* 51880f10851eSJérôme Glisse * No need to notify as we are downgrading page 51890f10851eSJérôme Glisse * table protection not changing it to point 51900f10851eSJérôme Glisse * to a new page. 51910f10851eSJérôme Glisse * 5192ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 51930f10851eSJérôme Glisse */ 51947f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 519584894e1cSPeter Xu entry = huge_pte_wrprotect(entry); 519634ee645eSJoerg Roedel } 51974eae4efaSPeter Xu 51985a2f8d22SPeter Xu if (!userfaultfd_wp(dst_vma)) 51995a2f8d22SPeter Xu entry = huge_pte_clear_uffd_wp(entry); 52005a2f8d22SPeter Xu 5201935d4f0cSRyan Roberts set_huge_pte_at(dst, addr, dst_pte, entry, sz); 52024eae4efaSPeter Xu hugetlb_count_add(npages, dst); 52031c59827dSHugh Dickins } 5204cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 5205cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 520663551ae0SDavid Gibson } 520763551ae0SDavid Gibson 5208623a1ddfSDavid Hildenbrand if (cow) { 5209623a1ddfSDavid Hildenbrand raw_write_seqcount_end(&src->write_protect_seq); 5210ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 521140549ba8SMike Kravetz } else { 521240549ba8SMike Kravetz hugetlb_vma_unlock_read(src_vma); 5213623a1ddfSDavid Hildenbrand } 5214e8569dd2SAndreas Sandberg 5215e8569dd2SAndreas Sandberg return ret; 521663551ae0SDavid Gibson } 521763551ae0SDavid Gibson 5218550a7d60SMina Almasry static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr, 5219935d4f0cSRyan Roberts unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte, 5220935d4f0cSRyan Roberts unsigned long sz) 5221550a7d60SMina Almasry { 5222550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 5223550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 5224550a7d60SMina Almasry spinlock_t *src_ptl, *dst_ptl; 5225db110a99SAneesh Kumar K.V pte_t pte; 5226550a7d60SMina Almasry 5227550a7d60SMina Almasry dst_ptl = huge_pte_lock(h, mm, dst_pte); 5228550a7d60SMina Almasry src_ptl = huge_pte_lockptr(h, mm, src_pte); 5229550a7d60SMina Almasry 5230550a7d60SMina Almasry /* 5231550a7d60SMina Almasry * We don't have to worry about the ordering of src and dst ptlocks 52328651a137SLorenzo Stoakes * because exclusive mmap_lock (or the i_mmap_lock) prevents deadlock. 5233550a7d60SMina Almasry */ 5234550a7d60SMina Almasry if (src_ptl != dst_ptl) 5235550a7d60SMina Almasry spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 5236550a7d60SMina Almasry 5237550a7d60SMina Almasry pte = huge_ptep_get_and_clear(mm, old_addr, src_pte); 5238935d4f0cSRyan Roberts set_huge_pte_at(mm, new_addr, dst_pte, pte, sz); 5239550a7d60SMina Almasry 5240550a7d60SMina Almasry if (src_ptl != dst_ptl) 5241550a7d60SMina Almasry spin_unlock(src_ptl); 5242550a7d60SMina Almasry spin_unlock(dst_ptl); 5243550a7d60SMina Almasry } 5244550a7d60SMina Almasry 5245550a7d60SMina Almasry int move_hugetlb_page_tables(struct vm_area_struct *vma, 5246550a7d60SMina Almasry struct vm_area_struct *new_vma, 5247550a7d60SMina Almasry unsigned long old_addr, unsigned long new_addr, 5248550a7d60SMina Almasry unsigned long len) 5249550a7d60SMina Almasry { 5250550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 5251550a7d60SMina Almasry struct address_space *mapping = vma->vm_file->f_mapping; 5252550a7d60SMina Almasry unsigned long sz = huge_page_size(h); 5253550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 5254550a7d60SMina Almasry unsigned long old_end = old_addr + len; 5255e95a9851SMike Kravetz unsigned long last_addr_mask; 5256550a7d60SMina Almasry pte_t *src_pte, *dst_pte; 5257550a7d60SMina Almasry struct mmu_notifier_range range; 52583d0b95cdSBaolin Wang bool shared_pmd = false; 5259550a7d60SMina Almasry 52607d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, old_addr, 5261550a7d60SMina Almasry old_end); 5262550a7d60SMina Almasry adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 52633d0b95cdSBaolin Wang /* 52643d0b95cdSBaolin Wang * In case of shared PMDs, we should cover the maximum possible 52653d0b95cdSBaolin Wang * range. 52663d0b95cdSBaolin Wang */ 52673d0b95cdSBaolin Wang flush_cache_range(vma, range.start, range.end); 52683d0b95cdSBaolin Wang 5269550a7d60SMina Almasry mmu_notifier_invalidate_range_start(&range); 5270e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 5271550a7d60SMina Almasry /* Prevent race with file truncation */ 527240549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 5273550a7d60SMina Almasry i_mmap_lock_write(mapping); 5274550a7d60SMina Almasry for (; old_addr < old_end; old_addr += sz, new_addr += sz) { 52759c67a207SPeter Xu src_pte = hugetlb_walk(vma, old_addr, sz); 5276e95a9851SMike Kravetz if (!src_pte) { 5277e95a9851SMike Kravetz old_addr |= last_addr_mask; 5278e95a9851SMike Kravetz new_addr |= last_addr_mask; 5279550a7d60SMina Almasry continue; 5280e95a9851SMike Kravetz } 5281550a7d60SMina Almasry if (huge_pte_none(huge_ptep_get(src_pte))) 5282550a7d60SMina Almasry continue; 5283550a7d60SMina Almasry 52844ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) { 52853d0b95cdSBaolin Wang shared_pmd = true; 52864ddb4d91SMike Kravetz old_addr |= last_addr_mask; 52874ddb4d91SMike Kravetz new_addr |= last_addr_mask; 5288550a7d60SMina Almasry continue; 52893d0b95cdSBaolin Wang } 5290550a7d60SMina Almasry 5291550a7d60SMina Almasry dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz); 5292550a7d60SMina Almasry if (!dst_pte) 5293550a7d60SMina Almasry break; 5294550a7d60SMina Almasry 5295935d4f0cSRyan Roberts move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte, sz); 5296550a7d60SMina Almasry } 52973d0b95cdSBaolin Wang 52983d0b95cdSBaolin Wang if (shared_pmd) 5299f720b471SKefeng Wang flush_hugetlb_tlb_range(vma, range.start, range.end); 53003d0b95cdSBaolin Wang else 5301f720b471SKefeng Wang flush_hugetlb_tlb_range(vma, old_end - len, old_end); 5302550a7d60SMina Almasry mmu_notifier_invalidate_range_end(&range); 530313e4ad2cSNadav Amit i_mmap_unlock_write(mapping); 530440549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 5305550a7d60SMina Almasry 5306550a7d60SMina Almasry return len + old_addr - old_end; 5307550a7d60SMina Almasry } 5308550a7d60SMina Almasry 530973c54763SPeter Xu static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 531024669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 531105e90bd0SPeter Xu struct page *ref_page, zap_flags_t zap_flags) 531263551ae0SDavid Gibson { 531363551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 531463551ae0SDavid Gibson unsigned long address; 5315c7546f8fSDavid Gibson pte_t *ptep; 531663551ae0SDavid Gibson pte_t pte; 5317cb900f41SKirill A. Shutemov spinlock_t *ptl; 531863551ae0SDavid Gibson struct page *page; 5319a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 5320a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 5321e95a9851SMike Kravetz unsigned long last_addr_mask; 5322a4a118f2SNadav Amit bool force_flush = false; 5323a5516438SAndi Kleen 532463551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 5325a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 5326a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 532763551ae0SDavid Gibson 532807e32661SAneesh Kumar K.V /* 532907e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 533007e32661SAneesh Kumar K.V * to huge page. 533107e32661SAneesh Kumar K.V */ 5332ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 533324669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 5334dff11abeSMike Kravetz 5335e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 5336569f48b8SHillf Danton address = start; 5337569f48b8SHillf Danton for (; address < end; address += sz) { 53389c67a207SPeter Xu ptep = hugetlb_walk(vma, address, sz); 5339e95a9851SMike Kravetz if (!ptep) { 5340e95a9851SMike Kravetz address |= last_addr_mask; 5341c7546f8fSDavid Gibson continue; 5342e95a9851SMike Kravetz } 5343c7546f8fSDavid Gibson 5344cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 53454ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, address, ptep)) { 534631d49da5SAneesh Kumar K.V spin_unlock(ptl); 5347a4a118f2SNadav Amit tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE); 5348a4a118f2SNadav Amit force_flush = true; 53494ddb4d91SMike Kravetz address |= last_addr_mask; 535031d49da5SAneesh Kumar K.V continue; 535131d49da5SAneesh Kumar K.V } 535239dde65cSChen, Kenneth W 53536629326bSHillf Danton pte = huge_ptep_get(ptep); 535431d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 535531d49da5SAneesh Kumar K.V spin_unlock(ptl); 535631d49da5SAneesh Kumar K.V continue; 535731d49da5SAneesh Kumar K.V } 53586629326bSHillf Danton 53596629326bSHillf Danton /* 53609fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 53619fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 53626629326bSHillf Danton */ 53639fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 536405e90bd0SPeter Xu /* 536505e90bd0SPeter Xu * If the pte was wr-protected by uffd-wp in any of the 536605e90bd0SPeter Xu * swap forms, meanwhile the caller does not want to 536705e90bd0SPeter Xu * drop the uffd-wp bit in this zap, then replace the 536805e90bd0SPeter Xu * pte with a marker. 536905e90bd0SPeter Xu */ 537005e90bd0SPeter Xu if (pte_swp_uffd_wp_any(pte) && 537105e90bd0SPeter Xu !(zap_flags & ZAP_FLAG_DROP_MARKER)) 537205e90bd0SPeter Xu set_huge_pte_at(mm, address, ptep, 5373935d4f0cSRyan Roberts make_pte_marker(PTE_MARKER_UFFD_WP), 5374935d4f0cSRyan Roberts sz); 537505e90bd0SPeter Xu else 53769386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 537731d49da5SAneesh Kumar K.V spin_unlock(ptl); 537831d49da5SAneesh Kumar K.V continue; 53798c4894c6SNaoya Horiguchi } 53806629326bSHillf Danton 53816629326bSHillf Danton page = pte_page(pte); 538204f2cbe3SMel Gorman /* 538304f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 538404f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 538504f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 538604f2cbe3SMel Gorman */ 538704f2cbe3SMel Gorman if (ref_page) { 538831d49da5SAneesh Kumar K.V if (page != ref_page) { 538931d49da5SAneesh Kumar K.V spin_unlock(ptl); 539031d49da5SAneesh Kumar K.V continue; 539131d49da5SAneesh Kumar K.V } 539204f2cbe3SMel Gorman /* 539304f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 539404f2cbe3SMel Gorman * future faults in this VMA will fail rather than 539504f2cbe3SMel Gorman * looking like data was lost 539604f2cbe3SMel Gorman */ 539704f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 539804f2cbe3SMel Gorman } 539904f2cbe3SMel Gorman 5400c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 5401b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 5402106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 54036649a386SKen Chen set_page_dirty(page); 540405e90bd0SPeter Xu /* Leave a uffd-wp pte marker if needed */ 540505e90bd0SPeter Xu if (huge_pte_uffd_wp(pte) && 540605e90bd0SPeter Xu !(zap_flags & ZAP_FLAG_DROP_MARKER)) 540705e90bd0SPeter Xu set_huge_pte_at(mm, address, ptep, 5408935d4f0cSRyan Roberts make_pte_marker(PTE_MARKER_UFFD_WP), 5409935d4f0cSRyan Roberts sz); 54105d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 5411cea86fe2SHugh Dickins page_remove_rmap(page, vma, true); 541231d49da5SAneesh Kumar K.V 5413cb900f41SKirill A. Shutemov spin_unlock(ptl); 5414e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 541524669e58SAneesh Kumar K.V /* 541631d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 541724669e58SAneesh Kumar K.V */ 541831d49da5SAneesh Kumar K.V if (ref_page) 541931d49da5SAneesh Kumar K.V break; 5420fe1668aeSChen, Kenneth W } 542124669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 5422a4a118f2SNadav Amit 5423a4a118f2SNadav Amit /* 5424a4a118f2SNadav Amit * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We 5425a4a118f2SNadav Amit * could defer the flush until now, since by holding i_mmap_rwsem we 5426a4a118f2SNadav Amit * guaranteed that the last refernece would not be dropped. But we must 5427a4a118f2SNadav Amit * do the flushing before we return, as otherwise i_mmap_rwsem will be 5428a4a118f2SNadav Amit * dropped and the last reference to the shared PMDs page might be 5429a4a118f2SNadav Amit * dropped as well. 5430a4a118f2SNadav Amit * 5431a4a118f2SNadav Amit * In theory we could defer the freeing of the PMD pages as well, but 5432a4a118f2SNadav Amit * huge_pmd_unshare() relies on the exact page_count for the PMD page to 5433a4a118f2SNadav Amit * detect sharing, so we cannot defer the release of the page either. 5434a4a118f2SNadav Amit * Instead, do flush now. 5435a4a118f2SNadav Amit */ 5436a4a118f2SNadav Amit if (force_flush) 5437a4a118f2SNadav Amit tlb_flush_mmu_tlbonly(tlb); 54381da177e4SLinus Torvalds } 543963551ae0SDavid Gibson 5440d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 5441d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 544205e90bd0SPeter Xu unsigned long end, struct page *ref_page, 544305e90bd0SPeter Xu zap_flags_t zap_flags) 5444d833352aSMel Gorman { 5445131a79b4SMike Kravetz hugetlb_vma_lock_write(vma); 5446131a79b4SMike Kravetz i_mmap_lock_write(vma->vm_file->f_mapping); 5447131a79b4SMike Kravetz 5448369258ceSMike Kravetz /* mmu notification performed in caller */ 544905e90bd0SPeter Xu __unmap_hugepage_range(tlb, vma, start, end, ref_page, zap_flags); 5450d833352aSMel Gorman 545104ada095SMike Kravetz if (zap_flags & ZAP_FLAG_UNMAP) { /* final unmap */ 5452d833352aSMel Gorman /* 545304ada095SMike Kravetz * Unlock and free the vma lock before releasing i_mmap_rwsem. 545404ada095SMike Kravetz * When the vma_lock is freed, this makes the vma ineligible 545504ada095SMike Kravetz * for pmd sharing. And, i_mmap_rwsem is required to set up 545604ada095SMike Kravetz * pmd sharing. This is important as page tables for this 545704ada095SMike Kravetz * unmapped range will be asynchrously deleted. If the page 545804ada095SMike Kravetz * tables are shared, there will be issues when accessed by 545904ada095SMike Kravetz * someone else. 5460d833352aSMel Gorman */ 5461ecfbd733SMike Kravetz __hugetlb_vma_unlock_write_free(vma); 5462131a79b4SMike Kravetz i_mmap_unlock_write(vma->vm_file->f_mapping); 546304ada095SMike Kravetz } else { 546404ada095SMike Kravetz i_mmap_unlock_write(vma->vm_file->f_mapping); 546504ada095SMike Kravetz hugetlb_vma_unlock_write(vma); 546604ada095SMike Kravetz } 5467d833352aSMel Gorman } 5468d833352aSMel Gorman 5469502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 547005e90bd0SPeter Xu unsigned long end, struct page *ref_page, 547105e90bd0SPeter Xu zap_flags_t zap_flags) 5472502717f4SChen, Kenneth W { 5473369258ceSMike Kravetz struct mmu_notifier_range range; 547424669e58SAneesh Kumar K.V struct mmu_gather tlb; 5475dff11abeSMike Kravetz 54767d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm, 5477369258ceSMike Kravetz start, end); 5478369258ceSMike Kravetz adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 5479369258ceSMike Kravetz mmu_notifier_invalidate_range_start(&range); 5480a72afd87SWill Deacon tlb_gather_mmu(&tlb, vma->vm_mm); 5481369258ceSMike Kravetz 548205e90bd0SPeter Xu __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags); 5483369258ceSMike Kravetz 5484369258ceSMike Kravetz mmu_notifier_invalidate_range_end(&range); 5485ae8eba8bSWill Deacon tlb_finish_mmu(&tlb); 5486502717f4SChen, Kenneth W } 5487502717f4SChen, Kenneth W 548804f2cbe3SMel Gorman /* 548904f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 5490578b7725SZhiyuan Dai * mapping it owns the reserve page for. The intention is to unmap the page 549104f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 549204f2cbe3SMel Gorman * same region. 549304f2cbe3SMel Gorman */ 54942f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 54952a4b3dedSHarvey Harrison struct page *page, unsigned long address) 549604f2cbe3SMel Gorman { 54977526674dSAdam Litke struct hstate *h = hstate_vma(vma); 549804f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 549904f2cbe3SMel Gorman struct address_space *mapping; 550004f2cbe3SMel Gorman pgoff_t pgoff; 550104f2cbe3SMel Gorman 550204f2cbe3SMel Gorman /* 550304f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 550404f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 550504f2cbe3SMel Gorman */ 55067526674dSAdam Litke address = address & huge_page_mask(h); 550736e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 550836e4f20aSMichal Hocko vma->vm_pgoff; 550993c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 551004f2cbe3SMel Gorman 55114eb2b1dcSMel Gorman /* 55124eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 55134eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 55144eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 55154eb2b1dcSMel Gorman */ 551683cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 55176b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 551804f2cbe3SMel Gorman /* Do not unmap the current VMA */ 551904f2cbe3SMel Gorman if (iter_vma == vma) 552004f2cbe3SMel Gorman continue; 552104f2cbe3SMel Gorman 552204f2cbe3SMel Gorman /* 55232f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 55242f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 55252f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 55262f84a899SMel Gorman */ 55272f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 55282f84a899SMel Gorman continue; 55292f84a899SMel Gorman 55302f84a899SMel Gorman /* 553104f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 553204f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 553304f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 553404f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 553504f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 553604f2cbe3SMel Gorman */ 553704f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 553824669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 553905e90bd0SPeter Xu address + huge_page_size(h), page, 0); 554004f2cbe3SMel Gorman } 554183cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 554204f2cbe3SMel Gorman } 554304f2cbe3SMel Gorman 55440fe6e20bSNaoya Horiguchi /* 5545c89357e2SDavid Hildenbrand * hugetlb_wp() should be called with page lock of the original hugepage held. 5546aa6d2e8cSBaolin Wang * Called with hugetlb_fault_mutex_table held and pte_page locked so we 5547ef009b25SMichal Hocko * cannot race with other handlers or page migration. 5548ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 55490fe6e20bSNaoya Horiguchi */ 5550c89357e2SDavid Hildenbrand static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma, 5551c89357e2SDavid Hildenbrand unsigned long address, pte_t *ptep, unsigned int flags, 5552371607a3SSidhartha Kumar struct folio *pagecache_folio, spinlock_t *ptl) 55531e8f889bSDavid Gibson { 5554c89357e2SDavid Hildenbrand const bool unshare = flags & FAULT_FLAG_UNSHARE; 555560d5b473SPeter Xu pte_t pte = huge_ptep_get(ptep); 5556a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 5557959a78b6SZhangPeng struct folio *old_folio; 5558d0ce0e47SSidhartha Kumar struct folio *new_folio; 55592b740303SSouptick Joarder int outside_reserve = 0; 55602b740303SSouptick Joarder vm_fault_t ret = 0; 5561974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 5562ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 55631e8f889bSDavid Gibson 55641d8d1464SDavid Hildenbrand /* 556560d5b473SPeter Xu * Never handle CoW for uffd-wp protected pages. It should be only 556660d5b473SPeter Xu * handled when the uffd-wp protection is removed. 556760d5b473SPeter Xu * 556860d5b473SPeter Xu * Note that only the CoW optimization path (in hugetlb_no_page()) 556960d5b473SPeter Xu * can trigger this, because hugetlb_fault() will always resolve 557060d5b473SPeter Xu * uffd-wp bit first. 557160d5b473SPeter Xu */ 557260d5b473SPeter Xu if (!unshare && huge_pte_uffd_wp(pte)) 557360d5b473SPeter Xu return 0; 557460d5b473SPeter Xu 557560d5b473SPeter Xu /* 55761d8d1464SDavid Hildenbrand * hugetlb does not support FOLL_FORCE-style write faults that keep the 55771d8d1464SDavid Hildenbrand * PTE mapped R/O such as maybe_mkwrite() would do. 55781d8d1464SDavid Hildenbrand */ 55791d8d1464SDavid Hildenbrand if (WARN_ON_ONCE(!unshare && !(vma->vm_flags & VM_WRITE))) 55801d8d1464SDavid Hildenbrand return VM_FAULT_SIGSEGV; 55811d8d1464SDavid Hildenbrand 55821d8d1464SDavid Hildenbrand /* Let's take out MAP_SHARED mappings first. */ 55831d8d1464SDavid Hildenbrand if (vma->vm_flags & VM_MAYSHARE) { 55841d8d1464SDavid Hildenbrand set_huge_ptep_writable(vma, haddr, ptep); 55851d8d1464SDavid Hildenbrand return 0; 55861d8d1464SDavid Hildenbrand } 55871d8d1464SDavid Hildenbrand 5588959a78b6SZhangPeng old_folio = page_folio(pte_page(pte)); 55891e8f889bSDavid Gibson 5590662ce1dcSYang Yang delayacct_wpcopy_start(); 5591662ce1dcSYang Yang 559204f2cbe3SMel Gorman retry_avoidcopy: 5593c89357e2SDavid Hildenbrand /* 5594c89357e2SDavid Hildenbrand * If no-one else is actually using this page, we're the exclusive 5595c89357e2SDavid Hildenbrand * owner and can reuse this page. 5596c89357e2SDavid Hildenbrand */ 5597959a78b6SZhangPeng if (folio_mapcount(old_folio) == 1 && folio_test_anon(old_folio)) { 5598959a78b6SZhangPeng if (!PageAnonExclusive(&old_folio->page)) 5599959a78b6SZhangPeng page_move_anon_rmap(&old_folio->page, vma); 5600c89357e2SDavid Hildenbrand if (likely(!unshare)) 56015b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 5602662ce1dcSYang Yang 5603662ce1dcSYang Yang delayacct_wpcopy_end(); 560483c54070SNick Piggin return 0; 56051e8f889bSDavid Gibson } 5606959a78b6SZhangPeng VM_BUG_ON_PAGE(folio_test_anon(old_folio) && 5607959a78b6SZhangPeng PageAnonExclusive(&old_folio->page), &old_folio->page); 56081e8f889bSDavid Gibson 560904f2cbe3SMel Gorman /* 561004f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 561104f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 561204f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 561304f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 561404f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 561504f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 561604f2cbe3SMel Gorman * of the full address range. 561704f2cbe3SMel Gorman */ 56185944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 5619959a78b6SZhangPeng old_folio != pagecache_folio) 562004f2cbe3SMel Gorman outside_reserve = 1; 562104f2cbe3SMel Gorman 5622959a78b6SZhangPeng folio_get(old_folio); 5623b76c8cfbSLarry Woodman 5624ad4404a2SDavidlohr Bueso /* 5625ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 5626ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 5627ad4404a2SDavidlohr Bueso */ 5628cb900f41SKirill A. Shutemov spin_unlock(ptl); 5629d0ce0e47SSidhartha Kumar new_folio = alloc_hugetlb_folio(vma, haddr, outside_reserve); 56301e8f889bSDavid Gibson 5631d0ce0e47SSidhartha Kumar if (IS_ERR(new_folio)) { 563204f2cbe3SMel Gorman /* 563304f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 563404f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 563504f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 563604f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 563704f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 563804f2cbe3SMel Gorman */ 563904f2cbe3SMel Gorman if (outside_reserve) { 564040549ba8SMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 564140549ba8SMike Kravetz pgoff_t idx; 564240549ba8SMike Kravetz u32 hash; 564340549ba8SMike Kravetz 5644959a78b6SZhangPeng folio_put(old_folio); 564540549ba8SMike Kravetz /* 564640549ba8SMike Kravetz * Drop hugetlb_fault_mutex and vma_lock before 564740549ba8SMike Kravetz * unmapping. unmapping needs to hold vma_lock 564840549ba8SMike Kravetz * in write mode. Dropping vma_lock in read mode 564940549ba8SMike Kravetz * here is OK as COW mappings do not interact with 565040549ba8SMike Kravetz * PMD sharing. 565140549ba8SMike Kravetz * 565240549ba8SMike Kravetz * Reacquire both after unmap operation. 565340549ba8SMike Kravetz */ 565440549ba8SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 565540549ba8SMike Kravetz hash = hugetlb_fault_mutex_hash(mapping, idx); 565640549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 565740549ba8SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 565840549ba8SMike Kravetz 5659959a78b6SZhangPeng unmap_ref_private(mm, vma, &old_folio->page, haddr); 566040549ba8SMike Kravetz 566140549ba8SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 566240549ba8SMike Kravetz hugetlb_vma_lock_read(vma); 5663cb900f41SKirill A. Shutemov spin_lock(ptl); 56649c67a207SPeter Xu ptep = hugetlb_walk(vma, haddr, huge_page_size(h)); 5665a9af0c5dSNaoya Horiguchi if (likely(ptep && 5666a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 566704f2cbe3SMel Gorman goto retry_avoidcopy; 5668a734bcc8SHillf Danton /* 5669cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 5670cb900f41SKirill A. Shutemov * lock, and our job is done. 5671a734bcc8SHillf Danton */ 5672662ce1dcSYang Yang delayacct_wpcopy_end(); 5673a734bcc8SHillf Danton return 0; 567404f2cbe3SMel Gorman } 567504f2cbe3SMel Gorman 5676d0ce0e47SSidhartha Kumar ret = vmf_error(PTR_ERR(new_folio)); 5677ad4404a2SDavidlohr Bueso goto out_release_old; 56781e8f889bSDavid Gibson } 56791e8f889bSDavid Gibson 56800fe6e20bSNaoya Horiguchi /* 56810fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 56820fe6e20bSNaoya Horiguchi * anon_vma prepared. 56830fe6e20bSNaoya Horiguchi */ 568444e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 5685ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 5686ad4404a2SDavidlohr Bueso goto out_release_all; 568744e2aa93SDean Nelson } 56880fe6e20bSNaoya Horiguchi 5689959a78b6SZhangPeng if (copy_user_large_folio(new_folio, old_folio, address, vma)) { 56901cb9dc4bSLiu Shixin ret = VM_FAULT_HWPOISON_LARGE; 56911cb9dc4bSLiu Shixin goto out_release_all; 56921cb9dc4bSLiu Shixin } 5693d0ce0e47SSidhartha Kumar __folio_mark_uptodate(new_folio); 56941e8f889bSDavid Gibson 56957d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, haddr, 56966f4f13e8SJérôme Glisse haddr + huge_page_size(h)); 5697ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 5698ad4404a2SDavidlohr Bueso 5699b76c8cfbSLarry Woodman /* 5700cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 5701b76c8cfbSLarry Woodman * before the page tables are altered 5702b76c8cfbSLarry Woodman */ 5703cb900f41SKirill A. Shutemov spin_lock(ptl); 57049c67a207SPeter Xu ptep = hugetlb_walk(vma, haddr, huge_page_size(h)); 5705a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 57060f230bc2SPeter Xu pte_t newpte = make_huge_pte(vma, &new_folio->page, !unshare); 57070f230bc2SPeter Xu 5708c89357e2SDavid Hildenbrand /* Break COW or unshare */ 57095b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 5710959a78b6SZhangPeng page_remove_rmap(&old_folio->page, vma, true); 5711d0ce0e47SSidhartha Kumar hugepage_add_new_anon_rmap(new_folio, vma, haddr); 57120f230bc2SPeter Xu if (huge_pte_uffd_wp(pte)) 57130f230bc2SPeter Xu newpte = huge_pte_mkuffd_wp(newpte); 5714935d4f0cSRyan Roberts set_huge_pte_at(mm, haddr, ptep, newpte, huge_page_size(h)); 5715d0ce0e47SSidhartha Kumar folio_set_hugetlb_migratable(new_folio); 57161e8f889bSDavid Gibson /* Make the old page be freed below */ 5717959a78b6SZhangPeng new_folio = old_folio; 57181e8f889bSDavid Gibson } 5719cb900f41SKirill A. Shutemov spin_unlock(ptl); 5720ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 5721ad4404a2SDavidlohr Bueso out_release_all: 5722c89357e2SDavid Hildenbrand /* 5723c89357e2SDavid Hildenbrand * No restore in case of successful pagetable update (Break COW or 5724c89357e2SDavid Hildenbrand * unshare) 5725c89357e2SDavid Hildenbrand */ 5726959a78b6SZhangPeng if (new_folio != old_folio) 5727d2d7bb44SSidhartha Kumar restore_reserve_on_error(h, vma, haddr, new_folio); 5728d0ce0e47SSidhartha Kumar folio_put(new_folio); 5729ad4404a2SDavidlohr Bueso out_release_old: 5730959a78b6SZhangPeng folio_put(old_folio); 57318312034fSJoonsoo Kim 5732ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 5733662ce1dcSYang Yang 5734662ce1dcSYang Yang delayacct_wpcopy_end(); 5735ad4404a2SDavidlohr Bueso return ret; 57361e8f889bSDavid Gibson } 57371e8f889bSDavid Gibson 57383ae77f43SHugh Dickins /* 57393ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 57403ae77f43SHugh Dickins */ 57413ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 57422a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 57432a15efc9SHugh Dickins { 574491a2fb95SSidhartha Kumar struct address_space *mapping = vma->vm_file->f_mapping; 574591a2fb95SSidhartha Kumar pgoff_t idx = vma_hugecache_offset(h, vma, address); 5746fd4aed8dSMike Kravetz struct folio *folio; 57472a15efc9SHugh Dickins 5748fd4aed8dSMike Kravetz folio = filemap_get_folio(mapping, idx); 5749fd4aed8dSMike Kravetz if (IS_ERR(folio)) 5750fd4aed8dSMike Kravetz return false; 5751fd4aed8dSMike Kravetz folio_put(folio); 5752fd4aed8dSMike Kravetz return true; 57532a15efc9SHugh Dickins } 57542a15efc9SHugh Dickins 57559b91c0e2SSidhartha Kumar int hugetlb_add_to_page_cache(struct folio *folio, struct address_space *mapping, 5756ab76ad54SMike Kravetz pgoff_t idx) 5757ab76ad54SMike Kravetz { 5758ab76ad54SMike Kravetz struct inode *inode = mapping->host; 5759ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 5760d9ef44deSMatthew Wilcox (Oracle) int err; 5761ab76ad54SMike Kravetz 5762d9ef44deSMatthew Wilcox (Oracle) __folio_set_locked(folio); 5763d9ef44deSMatthew Wilcox (Oracle) err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL); 5764d9ef44deSMatthew Wilcox (Oracle) 5765d9ef44deSMatthew Wilcox (Oracle) if (unlikely(err)) { 5766d9ef44deSMatthew Wilcox (Oracle) __folio_clear_locked(folio); 5767ab76ad54SMike Kravetz return err; 5768d9ef44deSMatthew Wilcox (Oracle) } 57699b91c0e2SSidhartha Kumar folio_clear_hugetlb_restore_reserve(folio); 5770ab76ad54SMike Kravetz 577122146c3cSMike Kravetz /* 5772d9ef44deSMatthew Wilcox (Oracle) * mark folio dirty so that it will not be removed from cache/file 577322146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 577422146c3cSMike Kravetz */ 5775d9ef44deSMatthew Wilcox (Oracle) folio_mark_dirty(folio); 577622146c3cSMike Kravetz 5777ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 5778ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 5779ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 5780ab76ad54SMike Kravetz return 0; 5781ab76ad54SMike Kravetz } 5782ab76ad54SMike Kravetz 57837677f7fdSAxel Rasmussen static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma, 57847677f7fdSAxel Rasmussen struct address_space *mapping, 57857677f7fdSAxel Rasmussen pgoff_t idx, 57867677f7fdSAxel Rasmussen unsigned int flags, 57877677f7fdSAxel Rasmussen unsigned long haddr, 5788824ddc60SNadav Amit unsigned long addr, 57897677f7fdSAxel Rasmussen unsigned long reason) 57907677f7fdSAxel Rasmussen { 57917677f7fdSAxel Rasmussen u32 hash; 57927677f7fdSAxel Rasmussen struct vm_fault vmf = { 57937677f7fdSAxel Rasmussen .vma = vma, 57947677f7fdSAxel Rasmussen .address = haddr, 5795824ddc60SNadav Amit .real_address = addr, 57967677f7fdSAxel Rasmussen .flags = flags, 57977677f7fdSAxel Rasmussen 57987677f7fdSAxel Rasmussen /* 57997677f7fdSAxel Rasmussen * Hard to debug if it ends up being 58007677f7fdSAxel Rasmussen * used by a callee that assumes 58017677f7fdSAxel Rasmussen * something about the other 58027677f7fdSAxel Rasmussen * uninitialized fields... same as in 58037677f7fdSAxel Rasmussen * memory.c 58047677f7fdSAxel Rasmussen */ 58057677f7fdSAxel Rasmussen }; 58067677f7fdSAxel Rasmussen 58077677f7fdSAxel Rasmussen /* 5808958f32ceSLiu Shixin * vma_lock and hugetlb_fault_mutex must be dropped before handling 5809958f32ceSLiu Shixin * userfault. Also mmap_lock could be dropped due to handling 5810958f32ceSLiu Shixin * userfault, any vma operation should be careful from here. 58117677f7fdSAxel Rasmussen */ 581240549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 58137677f7fdSAxel Rasmussen hash = hugetlb_fault_mutex_hash(mapping, idx); 58147677f7fdSAxel Rasmussen mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5815958f32ceSLiu Shixin return handle_userfault(&vmf, reason); 58167677f7fdSAxel Rasmussen } 58177677f7fdSAxel Rasmussen 58182ea7ff1eSPeter Xu /* 58192ea7ff1eSPeter Xu * Recheck pte with pgtable lock. Returns true if pte didn't change, or 58202ea7ff1eSPeter Xu * false if pte changed or is changing. 58212ea7ff1eSPeter Xu */ 58222ea7ff1eSPeter Xu static bool hugetlb_pte_stable(struct hstate *h, struct mm_struct *mm, 58232ea7ff1eSPeter Xu pte_t *ptep, pte_t old_pte) 58242ea7ff1eSPeter Xu { 58252ea7ff1eSPeter Xu spinlock_t *ptl; 58262ea7ff1eSPeter Xu bool same; 58272ea7ff1eSPeter Xu 58282ea7ff1eSPeter Xu ptl = huge_pte_lock(h, mm, ptep); 58292ea7ff1eSPeter Xu same = pte_same(huge_ptep_get(ptep), old_pte); 58302ea7ff1eSPeter Xu spin_unlock(ptl); 58312ea7ff1eSPeter Xu 58322ea7ff1eSPeter Xu return same; 58332ea7ff1eSPeter Xu } 58342ea7ff1eSPeter Xu 58352b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 58362b740303SSouptick Joarder struct vm_area_struct *vma, 58378382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 5838c64e912cSPeter Xu unsigned long address, pte_t *ptep, 5839c64e912cSPeter Xu pte_t old_pte, unsigned int flags) 5840ac9b9c66SHugh Dickins { 5841a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 58422b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 5843409eb8c2SHillf Danton int anon_rmap = 0; 58444c887265SAdam Litke unsigned long size; 5845d0ce0e47SSidhartha Kumar struct folio *folio; 58461e8f889bSDavid Gibson pte_t new_pte; 5847cb900f41SKirill A. Shutemov spinlock_t *ptl; 5848285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 5849d0ce0e47SSidhartha Kumar bool new_folio, new_pagecache_folio = false; 5850958f32ceSLiu Shixin u32 hash = hugetlb_fault_mutex_hash(mapping, idx); 58514c887265SAdam Litke 585204f2cbe3SMel Gorman /* 585304f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 585404f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 5855c89357e2SDavid Hildenbrand * COW/unsharing. Warn that such a situation has occurred as it may not 5856c89357e2SDavid Hildenbrand * be obvious. 585704f2cbe3SMel Gorman */ 585804f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 5859910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 586004f2cbe3SMel Gorman current->pid); 5861958f32ceSLiu Shixin goto out; 586204f2cbe3SMel Gorman } 586304f2cbe3SMel Gorman 58644c887265SAdam Litke /* 5865188a3972SMike Kravetz * Use page lock to guard against racing truncation 5866188a3972SMike Kravetz * before we get page_table_lock. 58674c887265SAdam Litke */ 5868d0ce0e47SSidhartha Kumar new_folio = false; 5869d0ce0e47SSidhartha Kumar folio = filemap_lock_folio(mapping, idx); 587066dabbb6SChristoph Hellwig if (IS_ERR(folio)) { 5871188a3972SMike Kravetz size = i_size_read(mapping->host) >> huge_page_shift(h); 5872188a3972SMike Kravetz if (idx >= size) 5873188a3972SMike Kravetz goto out; 58747677f7fdSAxel Rasmussen /* Check for page in userfault range */ 58752ea7ff1eSPeter Xu if (userfaultfd_missing(vma)) { 58762ea7ff1eSPeter Xu /* 58772ea7ff1eSPeter Xu * Since hugetlb_no_page() was examining pte 58782ea7ff1eSPeter Xu * without pgtable lock, we need to re-test under 58792ea7ff1eSPeter Xu * lock because the pte may not be stable and could 58802ea7ff1eSPeter Xu * have changed from under us. Try to detect 58812ea7ff1eSPeter Xu * either changed or during-changing ptes and retry 58822ea7ff1eSPeter Xu * properly when needed. 58832ea7ff1eSPeter Xu * 58842ea7ff1eSPeter Xu * Note that userfaultfd is actually fine with 58852ea7ff1eSPeter Xu * false positives (e.g. caused by pte changed), 58862ea7ff1eSPeter Xu * but not wrong logical events (e.g. caused by 58872ea7ff1eSPeter Xu * reading a pte during changing). The latter can 58882ea7ff1eSPeter Xu * confuse the userspace, so the strictness is very 58892ea7ff1eSPeter Xu * much preferred. E.g., MISSING event should 58902ea7ff1eSPeter Xu * never happen on the page after UFFDIO_COPY has 58912ea7ff1eSPeter Xu * correctly installed the page and returned. 58922ea7ff1eSPeter Xu */ 58932ea7ff1eSPeter Xu if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) { 58942ea7ff1eSPeter Xu ret = 0; 58952ea7ff1eSPeter Xu goto out; 58962ea7ff1eSPeter Xu } 58972ea7ff1eSPeter Xu 58982ea7ff1eSPeter Xu return hugetlb_handle_userfault(vma, mapping, idx, flags, 58992ea7ff1eSPeter Xu haddr, address, 59007677f7fdSAxel Rasmussen VM_UFFD_MISSING); 59012ea7ff1eSPeter Xu } 59021a1aad8aSMike Kravetz 5903d0ce0e47SSidhartha Kumar folio = alloc_hugetlb_folio(vma, haddr, 0); 5904d0ce0e47SSidhartha Kumar if (IS_ERR(folio)) { 59054643d67eSMike Kravetz /* 59064643d67eSMike Kravetz * Returning error will result in faulting task being 59074643d67eSMike Kravetz * sent SIGBUS. The hugetlb fault mutex prevents two 59084643d67eSMike Kravetz * tasks from racing to fault in the same page which 59094643d67eSMike Kravetz * could result in false unable to allocate errors. 59104643d67eSMike Kravetz * Page migration does not take the fault mutex, but 59114643d67eSMike Kravetz * does a clear then write of pte's under page table 59124643d67eSMike Kravetz * lock. Page fault code could race with migration, 59134643d67eSMike Kravetz * notice the clear pte and try to allocate a page 59144643d67eSMike Kravetz * here. Before returning error, get ptl and make 59154643d67eSMike Kravetz * sure there really is no pte entry. 59164643d67eSMike Kravetz */ 5917f9bf6c03SPeter Xu if (hugetlb_pte_stable(h, mm, ptep, old_pte)) 5918d0ce0e47SSidhartha Kumar ret = vmf_error(PTR_ERR(folio)); 5919f9bf6c03SPeter Xu else 5920f9bf6c03SPeter Xu ret = 0; 59216bda666aSChristoph Lameter goto out; 59226bda666aSChristoph Lameter } 5923d0ce0e47SSidhartha Kumar clear_huge_page(&folio->page, address, pages_per_huge_page(h)); 5924d0ce0e47SSidhartha Kumar __folio_mark_uptodate(folio); 5925d0ce0e47SSidhartha Kumar new_folio = true; 5926ac9b9c66SHugh Dickins 5927f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 59289b91c0e2SSidhartha Kumar int err = hugetlb_add_to_page_cache(folio, mapping, idx); 59296bda666aSChristoph Lameter if (err) { 59303a5497a2SMiaohe Lin /* 59313a5497a2SMiaohe Lin * err can't be -EEXIST which implies someone 59323a5497a2SMiaohe Lin * else consumed the reservation since hugetlb 59333a5497a2SMiaohe Lin * fault mutex is held when add a hugetlb page 59343a5497a2SMiaohe Lin * to the page cache. So it's safe to call 59353a5497a2SMiaohe Lin * restore_reserve_on_error() here. 59363a5497a2SMiaohe Lin */ 5937d2d7bb44SSidhartha Kumar restore_reserve_on_error(h, vma, haddr, folio); 5938d0ce0e47SSidhartha Kumar folio_put(folio); 59396bda666aSChristoph Lameter goto out; 59406bda666aSChristoph Lameter } 5941d0ce0e47SSidhartha Kumar new_pagecache_folio = true; 594223be7468SMel Gorman } else { 5943d0ce0e47SSidhartha Kumar folio_lock(folio); 59440fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 59450fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 59460fe6e20bSNaoya Horiguchi goto backout_unlocked; 594723be7468SMel Gorman } 5948409eb8c2SHillf Danton anon_rmap = 1; 59490fe6e20bSNaoya Horiguchi } 59500fe6e20bSNaoya Horiguchi } else { 595157303d80SAndy Whitcroft /* 5952998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 5953998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 5954998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 5955fd6a03edSNaoya Horiguchi */ 5956d0ce0e47SSidhartha Kumar if (unlikely(folio_test_hwpoison(folio))) { 59570eb98f15SMiaohe Lin ret = VM_FAULT_HWPOISON_LARGE | 5958972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 5959fd6a03edSNaoya Horiguchi goto backout_unlocked; 59606bda666aSChristoph Lameter } 59617677f7fdSAxel Rasmussen 59627677f7fdSAxel Rasmussen /* Check for page in userfault range. */ 59637677f7fdSAxel Rasmussen if (userfaultfd_minor(vma)) { 5964d0ce0e47SSidhartha Kumar folio_unlock(folio); 5965d0ce0e47SSidhartha Kumar folio_put(folio); 59662ea7ff1eSPeter Xu /* See comment in userfaultfd_missing() block above */ 59672ea7ff1eSPeter Xu if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) { 59682ea7ff1eSPeter Xu ret = 0; 59692ea7ff1eSPeter Xu goto out; 59702ea7ff1eSPeter Xu } 59712ea7ff1eSPeter Xu return hugetlb_handle_userfault(vma, mapping, idx, flags, 59722ea7ff1eSPeter Xu haddr, address, 59737677f7fdSAxel Rasmussen VM_UFFD_MINOR); 59747677f7fdSAxel Rasmussen } 5975998b4382SNaoya Horiguchi } 59761e8f889bSDavid Gibson 597757303d80SAndy Whitcroft /* 597857303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 597957303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 598057303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 598157303d80SAndy Whitcroft * the spinlock. 598257303d80SAndy Whitcroft */ 59835e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 5984285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 59852b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 59862b26736cSAndy Whitcroft goto backout_unlocked; 59872b26736cSAndy Whitcroft } 59885e911373SMike Kravetz /* Just decrements count, does not deallocate */ 5989285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 59905e911373SMike Kravetz } 599157303d80SAndy Whitcroft 59928bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 599383c54070SNick Piggin ret = 0; 5994c64e912cSPeter Xu /* If pte changed from under us, retry */ 5995c64e912cSPeter Xu if (!pte_same(huge_ptep_get(ptep), old_pte)) 59964c887265SAdam Litke goto backout; 59974c887265SAdam Litke 59984781593dSPeter Xu if (anon_rmap) 5999d0ce0e47SSidhartha Kumar hugepage_add_new_anon_rmap(folio, vma, haddr); 60004781593dSPeter Xu else 6001d0ce0e47SSidhartha Kumar page_dup_file_rmap(&folio->page, true); 6002d0ce0e47SSidhartha Kumar new_pte = make_huge_pte(vma, &folio->page, ((vma->vm_flags & VM_WRITE) 60031e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 6004c64e912cSPeter Xu /* 6005c64e912cSPeter Xu * If this pte was previously wr-protected, keep it wr-protected even 6006c64e912cSPeter Xu * if populated. 6007c64e912cSPeter Xu */ 6008c64e912cSPeter Xu if (unlikely(pte_marker_uffd_wp(old_pte))) 6009f1eb1bacSPeter Xu new_pte = huge_pte_mkuffd_wp(new_pte); 6010935d4f0cSRyan Roberts set_huge_pte_at(mm, haddr, ptep, new_pte, huge_page_size(h)); 60111e8f889bSDavid Gibson 60125d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 6013788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 60141e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 6015371607a3SSidhartha Kumar ret = hugetlb_wp(mm, vma, address, ptep, flags, folio, ptl); 60161e8f889bSDavid Gibson } 60171e8f889bSDavid Gibson 6018cb900f41SKirill A. Shutemov spin_unlock(ptl); 6019cb6acd01SMike Kravetz 6020cb6acd01SMike Kravetz /* 6021d0ce0e47SSidhartha Kumar * Only set hugetlb_migratable in newly allocated pages. Existing pages 6022d0ce0e47SSidhartha Kumar * found in the pagecache may not have hugetlb_migratable if they have 60238f251a3dSMike Kravetz * been isolated for migration. 6024cb6acd01SMike Kravetz */ 6025d0ce0e47SSidhartha Kumar if (new_folio) 6026d0ce0e47SSidhartha Kumar folio_set_hugetlb_migratable(folio); 6027cb6acd01SMike Kravetz 6028d0ce0e47SSidhartha Kumar folio_unlock(folio); 60294c887265SAdam Litke out: 6030958f32ceSLiu Shixin hugetlb_vma_unlock_read(vma); 6031958f32ceSLiu Shixin mutex_unlock(&hugetlb_fault_mutex_table[hash]); 6032ac9b9c66SHugh Dickins return ret; 60334c887265SAdam Litke 60344c887265SAdam Litke backout: 6035cb900f41SKirill A. Shutemov spin_unlock(ptl); 60362b26736cSAndy Whitcroft backout_unlocked: 6037d0ce0e47SSidhartha Kumar if (new_folio && !new_pagecache_folio) 6038d2d7bb44SSidhartha Kumar restore_reserve_on_error(h, vma, haddr, folio); 6039fa27759aSMike Kravetz 6040d0ce0e47SSidhartha Kumar folio_unlock(folio); 6041d0ce0e47SSidhartha Kumar folio_put(folio); 60424c887265SAdam Litke goto out; 6043ac9b9c66SHugh Dickins } 6044ac9b9c66SHugh Dickins 60458382d914SDavidlohr Bueso #ifdef CONFIG_SMP 6046188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 60478382d914SDavidlohr Bueso { 60488382d914SDavidlohr Bueso unsigned long key[2]; 60498382d914SDavidlohr Bueso u32 hash; 60508382d914SDavidlohr Bueso 60518382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 60528382d914SDavidlohr Bueso key[1] = idx; 60538382d914SDavidlohr Bueso 605455254636SMike Kravetz hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); 60558382d914SDavidlohr Bueso 60568382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 60578382d914SDavidlohr Bueso } 60588382d914SDavidlohr Bueso #else 60598382d914SDavidlohr Bueso /* 60606c26d310SMiaohe Lin * For uniprocessor systems we always use a single mutex, so just 60618382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 60628382d914SDavidlohr Bueso */ 6063188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 60648382d914SDavidlohr Bueso { 60658382d914SDavidlohr Bueso return 0; 60668382d914SDavidlohr Bueso } 60678382d914SDavidlohr Bueso #endif 60688382d914SDavidlohr Bueso 60692b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 6070788c7df4SHugh Dickins unsigned long address, unsigned int flags) 607186e5216fSAdam Litke { 60728382d914SDavidlohr Bueso pte_t *ptep, entry; 6073cb900f41SKirill A. Shutemov spinlock_t *ptl; 60742b740303SSouptick Joarder vm_fault_t ret; 60758382d914SDavidlohr Bueso u32 hash; 60768382d914SDavidlohr Bueso pgoff_t idx; 6077061e62e8SZhangPeng struct folio *folio = NULL; 6078371607a3SSidhartha Kumar struct folio *pagecache_folio = NULL; 6079a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 60808382d914SDavidlohr Bueso struct address_space *mapping; 60810f792cf9SNaoya Horiguchi int need_wait_lock = 0; 6082285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 608386e5216fSAdam Litke 60844ec31152SMatthew Wilcox (Oracle) /* TODO: Handle faults under the VMA lock */ 60854ec31152SMatthew Wilcox (Oracle) if (flags & FAULT_FLAG_VMA_LOCK) { 60864ec31152SMatthew Wilcox (Oracle) vma_end_read(vma); 60874ec31152SMatthew Wilcox (Oracle) return VM_FAULT_RETRY; 60884ec31152SMatthew Wilcox (Oracle) } 60894ec31152SMatthew Wilcox (Oracle) 60903935baa9SDavid Gibson /* 60913935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 60923935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 60933935baa9SDavid Gibson * the same page in the page cache. 60943935baa9SDavid Gibson */ 609540549ba8SMike Kravetz mapping = vma->vm_file->f_mapping; 609640549ba8SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 6097188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 6098c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 60998382d914SDavidlohr Bueso 610040549ba8SMike Kravetz /* 610140549ba8SMike Kravetz * Acquire vma lock before calling huge_pte_alloc and hold 610240549ba8SMike Kravetz * until finished with ptep. This prevents huge_pmd_unshare from 610340549ba8SMike Kravetz * being called elsewhere and making the ptep no longer valid. 610440549ba8SMike Kravetz */ 610540549ba8SMike Kravetz hugetlb_vma_lock_read(vma); 610640549ba8SMike Kravetz ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); 610740549ba8SMike Kravetz if (!ptep) { 610840549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 610940549ba8SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 611040549ba8SMike Kravetz return VM_FAULT_OOM; 611140549ba8SMike Kravetz } 611240549ba8SMike Kravetz 61137f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 6114af19487fSAxel Rasmussen if (huge_pte_none_mostly(entry)) { 6115af19487fSAxel Rasmussen if (is_pte_marker(entry)) { 6116af19487fSAxel Rasmussen pte_marker marker = 6117af19487fSAxel Rasmussen pte_marker_get(pte_to_swp_entry(entry)); 6118af19487fSAxel Rasmussen 6119af19487fSAxel Rasmussen if (marker & PTE_MARKER_POISONED) { 6120af19487fSAxel Rasmussen ret = VM_FAULT_HWPOISON_LARGE; 6121af19487fSAxel Rasmussen goto out_mutex; 6122af19487fSAxel Rasmussen } 6123af19487fSAxel Rasmussen } 6124af19487fSAxel Rasmussen 6125958f32ceSLiu Shixin /* 6126af19487fSAxel Rasmussen * Other PTE markers should be handled the same way as none PTE. 6127af19487fSAxel Rasmussen * 6128958f32ceSLiu Shixin * hugetlb_no_page will drop vma lock and hugetlb fault 6129958f32ceSLiu Shixin * mutex internally, which make us return immediately. 6130958f32ceSLiu Shixin */ 6131958f32ceSLiu Shixin return hugetlb_no_page(mm, vma, mapping, idx, address, ptep, 6132c64e912cSPeter Xu entry, flags); 6133af19487fSAxel Rasmussen } 613486e5216fSAdam Litke 613583c54070SNick Piggin ret = 0; 61361e8f889bSDavid Gibson 613757303d80SAndy Whitcroft /* 61380f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 61390f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 61407c8de358SEthon Paul * an active hugepage in pagecache. This goto expects the 2nd page 61417c8de358SEthon Paul * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will 61427c8de358SEthon Paul * properly handle it. 61430f792cf9SNaoya Horiguchi */ 6144fcd48540SPeter Xu if (!pte_present(entry)) { 6145fcd48540SPeter Xu if (unlikely(is_hugetlb_entry_migration(entry))) { 6146fcd48540SPeter Xu /* 6147fcd48540SPeter Xu * Release the hugetlb fault lock now, but retain 6148fcd48540SPeter Xu * the vma lock, because it is needed to guard the 6149fcd48540SPeter Xu * huge_pte_lockptr() later in 6150fcd48540SPeter Xu * migration_entry_wait_huge(). The vma lock will 6151fcd48540SPeter Xu * be released there. 6152fcd48540SPeter Xu */ 6153fcd48540SPeter Xu mutex_unlock(&hugetlb_fault_mutex_table[hash]); 6154fcd48540SPeter Xu migration_entry_wait_huge(vma, ptep); 6155fcd48540SPeter Xu return 0; 6156fcd48540SPeter Xu } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 6157fcd48540SPeter Xu ret = VM_FAULT_HWPOISON_LARGE | 6158fcd48540SPeter Xu VM_FAULT_SET_HINDEX(hstate_index(h)); 61590f792cf9SNaoya Horiguchi goto out_mutex; 6160fcd48540SPeter Xu } 61610f792cf9SNaoya Horiguchi 61620f792cf9SNaoya Horiguchi /* 6163c89357e2SDavid Hildenbrand * If we are going to COW/unshare the mapping later, we examine the 6164c89357e2SDavid Hildenbrand * pending reservations for this page now. This will ensure that any 616557303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 61661d8d1464SDavid Hildenbrand * spinlock. Also lookup the pagecache page now as it is used to 61671d8d1464SDavid Hildenbrand * determine if a reservation has been consumed. 616857303d80SAndy Whitcroft */ 6169c89357e2SDavid Hildenbrand if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) && 61701d8d1464SDavid Hildenbrand !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(entry)) { 6171285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 61722b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 6173b4d1d99fSDavid Gibson goto out_mutex; 61742b26736cSAndy Whitcroft } 61755e911373SMike Kravetz /* Just decrements count, does not deallocate */ 6176285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 617757303d80SAndy Whitcroft 6178371607a3SSidhartha Kumar pagecache_folio = filemap_lock_folio(mapping, idx); 617966dabbb6SChristoph Hellwig if (IS_ERR(pagecache_folio)) 618066dabbb6SChristoph Hellwig pagecache_folio = NULL; 618157303d80SAndy Whitcroft } 618257303d80SAndy Whitcroft 61830f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 61840fe6e20bSNaoya Horiguchi 6185c89357e2SDavid Hildenbrand /* Check for a racing update before calling hugetlb_wp() */ 6186b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 6187cb900f41SKirill A. Shutemov goto out_ptl; 6188b4d1d99fSDavid Gibson 6189166f3eccSPeter Xu /* Handle userfault-wp first, before trying to lock more pages */ 6190166f3eccSPeter Xu if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) && 6191166f3eccSPeter Xu (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 6192166f3eccSPeter Xu struct vm_fault vmf = { 6193166f3eccSPeter Xu .vma = vma, 6194166f3eccSPeter Xu .address = haddr, 6195166f3eccSPeter Xu .real_address = address, 6196166f3eccSPeter Xu .flags = flags, 6197166f3eccSPeter Xu }; 6198166f3eccSPeter Xu 6199166f3eccSPeter Xu spin_unlock(ptl); 6200371607a3SSidhartha Kumar if (pagecache_folio) { 6201371607a3SSidhartha Kumar folio_unlock(pagecache_folio); 6202371607a3SSidhartha Kumar folio_put(pagecache_folio); 6203166f3eccSPeter Xu } 620440549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 6205166f3eccSPeter Xu mutex_unlock(&hugetlb_fault_mutex_table[hash]); 6206166f3eccSPeter Xu return handle_userfault(&vmf, VM_UFFD_WP); 6207166f3eccSPeter Xu } 6208166f3eccSPeter Xu 62090f792cf9SNaoya Horiguchi /* 6210c89357e2SDavid Hildenbrand * hugetlb_wp() requires page locks of pte_page(entry) and 6211371607a3SSidhartha Kumar * pagecache_folio, so here we need take the former one 6212061e62e8SZhangPeng * when folio != pagecache_folio or !pagecache_folio. 62130f792cf9SNaoya Horiguchi */ 6214061e62e8SZhangPeng folio = page_folio(pte_page(entry)); 6215061e62e8SZhangPeng if (folio != pagecache_folio) 6216061e62e8SZhangPeng if (!folio_trylock(folio)) { 62170f792cf9SNaoya Horiguchi need_wait_lock = 1; 62180f792cf9SNaoya Horiguchi goto out_ptl; 62190f792cf9SNaoya Horiguchi } 62200f792cf9SNaoya Horiguchi 6221061e62e8SZhangPeng folio_get(folio); 6222b4d1d99fSDavid Gibson 6223c89357e2SDavid Hildenbrand if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) { 6224106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 6225c89357e2SDavid Hildenbrand ret = hugetlb_wp(mm, vma, address, ptep, flags, 6226371607a3SSidhartha Kumar pagecache_folio, ptl); 62270f792cf9SNaoya Horiguchi goto out_put_page; 6228c89357e2SDavid Hildenbrand } else if (likely(flags & FAULT_FLAG_WRITE)) { 6229106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 6230b4d1d99fSDavid Gibson } 6231c89357e2SDavid Hildenbrand } 6232b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 6233285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 6234788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 6235285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 62360f792cf9SNaoya Horiguchi out_put_page: 6237061e62e8SZhangPeng if (folio != pagecache_folio) 6238061e62e8SZhangPeng folio_unlock(folio); 6239061e62e8SZhangPeng folio_put(folio); 6240cb900f41SKirill A. Shutemov out_ptl: 6241cb900f41SKirill A. Shutemov spin_unlock(ptl); 624257303d80SAndy Whitcroft 6243371607a3SSidhartha Kumar if (pagecache_folio) { 6244371607a3SSidhartha Kumar folio_unlock(pagecache_folio); 6245371607a3SSidhartha Kumar folio_put(pagecache_folio); 624657303d80SAndy Whitcroft } 6247b4d1d99fSDavid Gibson out_mutex: 624840549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 6249c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 62500f792cf9SNaoya Horiguchi /* 62510f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 62520f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 62530f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 62540f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 62550f792cf9SNaoya Horiguchi * here without taking refcount. 62560f792cf9SNaoya Horiguchi */ 62570f792cf9SNaoya Horiguchi if (need_wait_lock) 6258061e62e8SZhangPeng folio_wait_locked(folio); 62591e8f889bSDavid Gibson return ret; 626086e5216fSAdam Litke } 626186e5216fSAdam Litke 6262714c1891SAxel Rasmussen #ifdef CONFIG_USERFAULTFD 62638fb5debcSMike Kravetz /* 6264a734991cSAxel Rasmussen * Used by userfaultfd UFFDIO_* ioctls. Based on userfaultfd's mfill_atomic_pte 6265a734991cSAxel Rasmussen * with modifications for hugetlb pages. 62668fb5debcSMike Kravetz */ 626761c50040SAxel Rasmussen int hugetlb_mfill_atomic_pte(pte_t *dst_pte, 62688fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 62698fb5debcSMike Kravetz unsigned long dst_addr, 62708fb5debcSMike Kravetz unsigned long src_addr, 6271d9712937SAxel Rasmussen uffd_flags_t flags, 62720169fd51SZhangPeng struct folio **foliop) 62738fb5debcSMike Kravetz { 627461c50040SAxel Rasmussen struct mm_struct *dst_mm = dst_vma->vm_mm; 6275d9712937SAxel Rasmussen bool is_continue = uffd_flags_mode_is(flags, MFILL_ATOMIC_CONTINUE); 6276d9712937SAxel Rasmussen bool wp_enabled = (flags & MFILL_ATOMIC_WP); 62778cc5fcbbSMina Almasry struct hstate *h = hstate_vma(dst_vma); 62788cc5fcbbSMina Almasry struct address_space *mapping = dst_vma->vm_file->f_mapping; 62798cc5fcbbSMina Almasry pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr); 62801e392147SAndrea Arcangeli unsigned long size; 62811c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 62828fb5debcSMike Kravetz pte_t _dst_pte; 62838fb5debcSMike Kravetz spinlock_t *ptl; 62848cc5fcbbSMina Almasry int ret = -ENOMEM; 6285d0ce0e47SSidhartha Kumar struct folio *folio; 6286f6191471SAxel Rasmussen int writable; 6287d0ce0e47SSidhartha Kumar bool folio_in_pagecache = false; 62888fb5debcSMike Kravetz 62898a13897fSAxel Rasmussen if (uffd_flags_mode_is(flags, MFILL_ATOMIC_POISON)) { 62908a13897fSAxel Rasmussen ptl = huge_pte_lock(h, dst_mm, dst_pte); 62918a13897fSAxel Rasmussen 62928a13897fSAxel Rasmussen /* Don't overwrite any existing PTEs (even markers) */ 62938a13897fSAxel Rasmussen if (!huge_pte_none(huge_ptep_get(dst_pte))) { 62948a13897fSAxel Rasmussen spin_unlock(ptl); 62958a13897fSAxel Rasmussen return -EEXIST; 62968a13897fSAxel Rasmussen } 62978a13897fSAxel Rasmussen 62988a13897fSAxel Rasmussen _dst_pte = make_pte_marker(PTE_MARKER_POISONED); 6299935d4f0cSRyan Roberts set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte, 6300935d4f0cSRyan Roberts huge_page_size(h)); 63018a13897fSAxel Rasmussen 63028a13897fSAxel Rasmussen /* No need to invalidate - it was non-present before */ 63038a13897fSAxel Rasmussen update_mmu_cache(dst_vma, dst_addr, dst_pte); 63048a13897fSAxel Rasmussen 63058a13897fSAxel Rasmussen spin_unlock(ptl); 63068a13897fSAxel Rasmussen return 0; 63078a13897fSAxel Rasmussen } 63088a13897fSAxel Rasmussen 6309f6191471SAxel Rasmussen if (is_continue) { 6310f6191471SAxel Rasmussen ret = -EFAULT; 6311d0ce0e47SSidhartha Kumar folio = filemap_lock_folio(mapping, idx); 631266dabbb6SChristoph Hellwig if (IS_ERR(folio)) 6313f6191471SAxel Rasmussen goto out; 6314d0ce0e47SSidhartha Kumar folio_in_pagecache = true; 63150169fd51SZhangPeng } else if (!*foliop) { 63160169fd51SZhangPeng /* If a folio already exists, then it's UFFDIO_COPY for 6317d84cf06eSMina Almasry * a non-missing case. Return -EEXIST. 6318d84cf06eSMina Almasry */ 6319d84cf06eSMina Almasry if (vm_shared && 6320d84cf06eSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 6321d84cf06eSMina Almasry ret = -EEXIST; 63228fb5debcSMike Kravetz goto out; 6323d84cf06eSMina Almasry } 6324d84cf06eSMina Almasry 6325d0ce0e47SSidhartha Kumar folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0); 6326d0ce0e47SSidhartha Kumar if (IS_ERR(folio)) { 6327d84cf06eSMina Almasry ret = -ENOMEM; 6328d84cf06eSMina Almasry goto out; 6329d84cf06eSMina Almasry } 63308fb5debcSMike Kravetz 6331e87340caSZhangPeng ret = copy_folio_from_user(folio, (const void __user *) src_addr, 6332e87340caSZhangPeng false); 63338fb5debcSMike Kravetz 6334c1e8d7c6SMichel Lespinasse /* fallback to copy_from_user outside mmap_lock */ 63358fb5debcSMike Kravetz if (unlikely(ret)) { 63369e368259SAndrea Arcangeli ret = -ENOENT; 6337d0ce0e47SSidhartha Kumar /* Free the allocated folio which may have 63388cc5fcbbSMina Almasry * consumed a reservation. 63398cc5fcbbSMina Almasry */ 6340d2d7bb44SSidhartha Kumar restore_reserve_on_error(h, dst_vma, dst_addr, folio); 6341d0ce0e47SSidhartha Kumar folio_put(folio); 63428cc5fcbbSMina Almasry 6343d0ce0e47SSidhartha Kumar /* Allocate a temporary folio to hold the copied 63448cc5fcbbSMina Almasry * contents. 63458cc5fcbbSMina Almasry */ 6346d0ce0e47SSidhartha Kumar folio = alloc_hugetlb_folio_vma(h, dst_vma, dst_addr); 6347d0ce0e47SSidhartha Kumar if (!folio) { 63488cc5fcbbSMina Almasry ret = -ENOMEM; 63498cc5fcbbSMina Almasry goto out; 63508cc5fcbbSMina Almasry } 63510169fd51SZhangPeng *foliop = folio; 63520169fd51SZhangPeng /* Set the outparam foliop and return to the caller to 63538cc5fcbbSMina Almasry * copy the contents outside the lock. Don't free the 63540169fd51SZhangPeng * folio. 63558cc5fcbbSMina Almasry */ 63568fb5debcSMike Kravetz goto out; 63578fb5debcSMike Kravetz } 63588fb5debcSMike Kravetz } else { 63598cc5fcbbSMina Almasry if (vm_shared && 63608cc5fcbbSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 63610169fd51SZhangPeng folio_put(*foliop); 63628cc5fcbbSMina Almasry ret = -EEXIST; 63630169fd51SZhangPeng *foliop = NULL; 63648cc5fcbbSMina Almasry goto out; 63658cc5fcbbSMina Almasry } 63668cc5fcbbSMina Almasry 6367d0ce0e47SSidhartha Kumar folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0); 6368d0ce0e47SSidhartha Kumar if (IS_ERR(folio)) { 63690169fd51SZhangPeng folio_put(*foliop); 63708cc5fcbbSMina Almasry ret = -ENOMEM; 63710169fd51SZhangPeng *foliop = NULL; 63728cc5fcbbSMina Almasry goto out; 63738cc5fcbbSMina Almasry } 63741cb9dc4bSLiu Shixin ret = copy_user_large_folio(folio, *foliop, dst_addr, dst_vma); 63750169fd51SZhangPeng folio_put(*foliop); 63760169fd51SZhangPeng *foliop = NULL; 63771cb9dc4bSLiu Shixin if (ret) { 63781cb9dc4bSLiu Shixin folio_put(folio); 63791cb9dc4bSLiu Shixin goto out; 63801cb9dc4bSLiu Shixin } 63818fb5debcSMike Kravetz } 63828fb5debcSMike Kravetz 63838fb5debcSMike Kravetz /* 6384d0ce0e47SSidhartha Kumar * The memory barrier inside __folio_mark_uptodate makes sure that 63858fb5debcSMike Kravetz * preceding stores to the page contents become visible before 63868fb5debcSMike Kravetz * the set_pte_at() write. 63878fb5debcSMike Kravetz */ 6388d0ce0e47SSidhartha Kumar __folio_mark_uptodate(folio); 63898fb5debcSMike Kravetz 6390f6191471SAxel Rasmussen /* Add shared, newly allocated pages to the page cache. */ 6391f6191471SAxel Rasmussen if (vm_shared && !is_continue) { 63921e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 63931e392147SAndrea Arcangeli ret = -EFAULT; 63941e392147SAndrea Arcangeli if (idx >= size) 63951e392147SAndrea Arcangeli goto out_release_nounlock; 63961c9e8defSMike Kravetz 63971e392147SAndrea Arcangeli /* 63981e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 63997e1813d4SMike Kravetz * hugetlb_add_to_page_cache() below happens through the 64001e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 64011e392147SAndrea Arcangeli * the caller. 64021e392147SAndrea Arcangeli */ 64039b91c0e2SSidhartha Kumar ret = hugetlb_add_to_page_cache(folio, mapping, idx); 64041c9e8defSMike Kravetz if (ret) 64051c9e8defSMike Kravetz goto out_release_nounlock; 6406d0ce0e47SSidhartha Kumar folio_in_pagecache = true; 64071c9e8defSMike Kravetz } 64081c9e8defSMike Kravetz 6409bcc66543SMiaohe Lin ptl = huge_pte_lock(h, dst_mm, dst_pte); 64108fb5debcSMike Kravetz 64118625147cSJames Houghton ret = -EIO; 6412d0ce0e47SSidhartha Kumar if (folio_test_hwpoison(folio)) 64138625147cSJames Houghton goto out_release_unlock; 64148625147cSJames Houghton 64151e392147SAndrea Arcangeli /* 64166041c691SPeter Xu * We allow to overwrite a pte marker: consider when both MISSING|WP 64176041c691SPeter Xu * registered, we firstly wr-protect a none pte which has no page cache 64186041c691SPeter Xu * page backing it, then access the page. 64196041c691SPeter Xu */ 6420fa27759aSMike Kravetz ret = -EEXIST; 64216041c691SPeter Xu if (!huge_pte_none_mostly(huge_ptep_get(dst_pte))) 64228fb5debcSMike Kravetz goto out_release_unlock; 64238fb5debcSMike Kravetz 6424d0ce0e47SSidhartha Kumar if (folio_in_pagecache) 6425d0ce0e47SSidhartha Kumar page_dup_file_rmap(&folio->page, true); 64264781593dSPeter Xu else 6427d0ce0e47SSidhartha Kumar hugepage_add_new_anon_rmap(folio, dst_vma, dst_addr); 64288fb5debcSMike Kravetz 64296041c691SPeter Xu /* 64306041c691SPeter Xu * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY 64316041c691SPeter Xu * with wp flag set, don't set pte write bit. 64326041c691SPeter Xu */ 6433d9712937SAxel Rasmussen if (wp_enabled || (is_continue && !vm_shared)) 6434f6191471SAxel Rasmussen writable = 0; 6435f6191471SAxel Rasmussen else 6436f6191471SAxel Rasmussen writable = dst_vma->vm_flags & VM_WRITE; 6437f6191471SAxel Rasmussen 6438d0ce0e47SSidhartha Kumar _dst_pte = make_huge_pte(dst_vma, &folio->page, writable); 64396041c691SPeter Xu /* 64406041c691SPeter Xu * Always mark UFFDIO_COPY page dirty; note that this may not be 64416041c691SPeter Xu * extremely important for hugetlbfs for now since swapping is not 64426041c691SPeter Xu * supported, but we should still be clear in that this page cannot be 64436041c691SPeter Xu * thrown away at will, even if write bit not set. 64446041c691SPeter Xu */ 64458fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 64468fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 64478fb5debcSMike Kravetz 6448d9712937SAxel Rasmussen if (wp_enabled) 64496041c691SPeter Xu _dst_pte = huge_pte_mkuffd_wp(_dst_pte); 64506041c691SPeter Xu 6451935d4f0cSRyan Roberts set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte, huge_page_size(h)); 64528fb5debcSMike Kravetz 64538fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 64548fb5debcSMike Kravetz 64558fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 64568fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 64578fb5debcSMike Kravetz 64588fb5debcSMike Kravetz spin_unlock(ptl); 6459f6191471SAxel Rasmussen if (!is_continue) 6460d0ce0e47SSidhartha Kumar folio_set_hugetlb_migratable(folio); 6461f6191471SAxel Rasmussen if (vm_shared || is_continue) 6462d0ce0e47SSidhartha Kumar folio_unlock(folio); 64638fb5debcSMike Kravetz ret = 0; 64648fb5debcSMike Kravetz out: 64658fb5debcSMike Kravetz return ret; 64668fb5debcSMike Kravetz out_release_unlock: 64678fb5debcSMike Kravetz spin_unlock(ptl); 6468f6191471SAxel Rasmussen if (vm_shared || is_continue) 6469d0ce0e47SSidhartha Kumar folio_unlock(folio); 64705af10dfdSAndrea Arcangeli out_release_nounlock: 6471d0ce0e47SSidhartha Kumar if (!folio_in_pagecache) 6472d2d7bb44SSidhartha Kumar restore_reserve_on_error(h, dst_vma, dst_addr, folio); 6473d0ce0e47SSidhartha Kumar folio_put(folio); 64748fb5debcSMike Kravetz goto out; 64758fb5debcSMike Kravetz } 6476714c1891SAxel Rasmussen #endif /* CONFIG_USERFAULTFD */ 64778fb5debcSMike Kravetz 647857a196a5SMike Kravetz struct page *hugetlb_follow_page_mask(struct vm_area_struct *vma, 64795502ea44SPeter Xu unsigned long address, unsigned int flags, 64805502ea44SPeter Xu unsigned int *page_mask) 648157a196a5SMike Kravetz { 648257a196a5SMike Kravetz struct hstate *h = hstate_vma(vma); 648357a196a5SMike Kravetz struct mm_struct *mm = vma->vm_mm; 648457a196a5SMike Kravetz unsigned long haddr = address & huge_page_mask(h); 648557a196a5SMike Kravetz struct page *page = NULL; 648657a196a5SMike Kravetz spinlock_t *ptl; 648757a196a5SMike Kravetz pte_t *pte, entry; 6488458568c9SPeter Xu int ret; 648957a196a5SMike Kravetz 64907d049f3aSPeter Xu hugetlb_vma_lock_read(vma); 64919c67a207SPeter Xu pte = hugetlb_walk(vma, haddr, huge_page_size(h)); 649257a196a5SMike Kravetz if (!pte) 64937d049f3aSPeter Xu goto out_unlock; 649457a196a5SMike Kravetz 649557a196a5SMike Kravetz ptl = huge_pte_lock(h, mm, pte); 649657a196a5SMike Kravetz entry = huge_ptep_get(pte); 649757a196a5SMike Kravetz if (pte_present(entry)) { 6498458568c9SPeter Xu page = pte_page(entry); 6499458568c9SPeter Xu 6500458568c9SPeter Xu if (!huge_pte_write(entry)) { 6501458568c9SPeter Xu if (flags & FOLL_WRITE) { 6502458568c9SPeter Xu page = NULL; 6503458568c9SPeter Xu goto out; 6504458568c9SPeter Xu } 6505458568c9SPeter Xu 6506458568c9SPeter Xu if (gup_must_unshare(vma, flags, page)) { 6507458568c9SPeter Xu /* Tell the caller to do unsharing */ 6508458568c9SPeter Xu page = ERR_PTR(-EMLINK); 6509458568c9SPeter Xu goto out; 6510458568c9SPeter Xu } 6511458568c9SPeter Xu } 6512458568c9SPeter Xu 6513458568c9SPeter Xu page += ((address & ~huge_page_mask(h)) >> PAGE_SHIFT); 6514458568c9SPeter Xu 651557a196a5SMike Kravetz /* 651657a196a5SMike Kravetz * Note that page may be a sub-page, and with vmemmap 651757a196a5SMike Kravetz * optimizations the page struct may be read only. 651857a196a5SMike Kravetz * try_grab_page() will increase the ref count on the 651957a196a5SMike Kravetz * head page, so this will be OK. 652057a196a5SMike Kravetz * 6521e2ca6ba6SLinus Torvalds * try_grab_page() should always be able to get the page here, 6522e2ca6ba6SLinus Torvalds * because we hold the ptl lock and have verified pte_present(). 652357a196a5SMike Kravetz */ 6524458568c9SPeter Xu ret = try_grab_page(page, flags); 6525458568c9SPeter Xu 6526458568c9SPeter Xu if (WARN_ON_ONCE(ret)) { 6527458568c9SPeter Xu page = ERR_PTR(ret); 652857a196a5SMike Kravetz goto out; 652957a196a5SMike Kravetz } 65305502ea44SPeter Xu 65315502ea44SPeter Xu *page_mask = (1U << huge_page_order(h)) - 1; 653257a196a5SMike Kravetz } 653357a196a5SMike Kravetz out: 653457a196a5SMike Kravetz spin_unlock(ptl); 65357d049f3aSPeter Xu out_unlock: 65367d049f3aSPeter Xu hugetlb_vma_unlock_read(vma); 6537dd767aaaSPeter Xu 6538dd767aaaSPeter Xu /* 6539dd767aaaSPeter Xu * Fixup retval for dump requests: if pagecache doesn't exist, 6540dd767aaaSPeter Xu * don't try to allocate a new page but just skip it. 6541dd767aaaSPeter Xu */ 6542dd767aaaSPeter Xu if (!page && (flags & FOLL_DUMP) && 6543dd767aaaSPeter Xu !hugetlbfs_pagecache_present(h, vma, address)) 6544dd767aaaSPeter Xu page = ERR_PTR(-EFAULT); 6545dd767aaaSPeter Xu 654657a196a5SMike Kravetz return page; 654757a196a5SMike Kravetz } 654857a196a5SMike Kravetz 6549a79390f5SPeter Xu long hugetlb_change_protection(struct vm_area_struct *vma, 65505a90d5a1SPeter Xu unsigned long address, unsigned long end, 65515a90d5a1SPeter Xu pgprot_t newprot, unsigned long cp_flags) 65528f860591SZhang, Yanmin { 65538f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 65548f860591SZhang, Yanmin unsigned long start = address; 65558f860591SZhang, Yanmin pte_t *ptep; 65568f860591SZhang, Yanmin pte_t pte; 6557a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 6558a79390f5SPeter Xu long pages = 0, psize = huge_page_size(h); 6559dff11abeSMike Kravetz bool shared_pmd = false; 6560ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 6561e95a9851SMike Kravetz unsigned long last_addr_mask; 65625a90d5a1SPeter Xu bool uffd_wp = cp_flags & MM_CP_UFFD_WP; 65635a90d5a1SPeter Xu bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE; 6564dff11abeSMike Kravetz 6565dff11abeSMike Kravetz /* 6566dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 6567ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 6568dff11abeSMike Kravetz * range if PMD sharing is possible. 6569dff11abeSMike Kravetz */ 65707269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 65717d4a8be0SAlistair Popple 0, mm, start, end); 6572ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 65738f860591SZhang, Yanmin 65748f860591SZhang, Yanmin BUG_ON(address >= end); 6575ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 65768f860591SZhang, Yanmin 6577ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 657840549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 657983cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 658040549ba8SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 658160dfaad6SPeter Xu for (; address < end; address += psize) { 6582cb900f41SKirill A. Shutemov spinlock_t *ptl; 65839c67a207SPeter Xu ptep = hugetlb_walk(vma, address, psize); 6584e95a9851SMike Kravetz if (!ptep) { 6585fed15f13SPeter Xu if (!uffd_wp) { 6586e95a9851SMike Kravetz address |= last_addr_mask; 65878f860591SZhang, Yanmin continue; 6588e95a9851SMike Kravetz } 6589fed15f13SPeter Xu /* 6590fed15f13SPeter Xu * Userfaultfd wr-protect requires pgtable 6591fed15f13SPeter Xu * pre-allocations to install pte markers. 6592fed15f13SPeter Xu */ 6593fed15f13SPeter Xu ptep = huge_pte_alloc(mm, vma, address, psize); 6594d1751118SPeter Xu if (!ptep) { 6595d1751118SPeter Xu pages = -ENOMEM; 6596fed15f13SPeter Xu break; 6597fed15f13SPeter Xu } 6598d1751118SPeter Xu } 6599cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 66004ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, address, ptep)) { 660160dfaad6SPeter Xu /* 660260dfaad6SPeter Xu * When uffd-wp is enabled on the vma, unshare 660360dfaad6SPeter Xu * shouldn't happen at all. Warn about it if it 660460dfaad6SPeter Xu * happened due to some reason. 660560dfaad6SPeter Xu */ 660660dfaad6SPeter Xu WARN_ON_ONCE(uffd_wp || uffd_wp_resolve); 66077da4d641SPeter Zijlstra pages++; 6608cb900f41SKirill A. Shutemov spin_unlock(ptl); 6609dff11abeSMike Kravetz shared_pmd = true; 66104ddb4d91SMike Kravetz address |= last_addr_mask; 661139dde65cSChen, Kenneth W continue; 66127da4d641SPeter Zijlstra } 6613a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 6614a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 66150e678153SDavid Hildenbrand /* Nothing to do. */ 66160e678153SDavid Hildenbrand } else if (unlikely(is_hugetlb_entry_migration(pte))) { 6617a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 66186c287605SDavid Hildenbrand struct page *page = pfn_swap_entry_to_page(entry); 661944f86392SDavid Hildenbrand pte_t newpte = pte; 6620a8bda28dSNaoya Horiguchi 662144f86392SDavid Hildenbrand if (is_writable_migration_entry(entry)) { 66226c287605SDavid Hildenbrand if (PageAnon(page)) 66236c287605SDavid Hildenbrand entry = make_readable_exclusive_migration_entry( 66246c287605SDavid Hildenbrand swp_offset(entry)); 66256c287605SDavid Hildenbrand else 66264dd845b5SAlistair Popple entry = make_readable_migration_entry( 66274dd845b5SAlistair Popple swp_offset(entry)); 6628a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 662944f86392SDavid Hildenbrand pages++; 663044f86392SDavid Hildenbrand } 663144f86392SDavid Hildenbrand 66325a90d5a1SPeter Xu if (uffd_wp) 66335a90d5a1SPeter Xu newpte = pte_swp_mkuffd_wp(newpte); 66345a90d5a1SPeter Xu else if (uffd_wp_resolve) 66355a90d5a1SPeter Xu newpte = pte_swp_clear_uffd_wp(newpte); 663644f86392SDavid Hildenbrand if (!pte_same(pte, newpte)) 6637935d4f0cSRyan Roberts set_huge_pte_at(mm, address, ptep, newpte, psize); 66380e678153SDavid Hildenbrand } else if (unlikely(is_pte_marker(pte))) { 66390e678153SDavid Hildenbrand /* No other markers apply for now. */ 66400e678153SDavid Hildenbrand WARN_ON_ONCE(!pte_marker_uffd_wp(pte)); 664160dfaad6SPeter Xu if (uffd_wp_resolve) 66420e678153SDavid Hildenbrand /* Safe to modify directly (non-present->none). */ 664360dfaad6SPeter Xu huge_pte_clear(mm, address, ptep, psize); 66440e678153SDavid Hildenbrand } else if (!huge_pte_none(pte)) { 6645023bdd00SAneesh Kumar K.V pte_t old_pte; 664679c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 6647023bdd00SAneesh Kumar K.V 6648023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 664916785bd7SAnshuman Khandual pte = huge_pte_modify(old_pte, newprot); 665079c1c594SChristophe Leroy pte = arch_make_huge_pte(pte, shift, vma->vm_flags); 66515a90d5a1SPeter Xu if (uffd_wp) 6652f1eb1bacSPeter Xu pte = huge_pte_mkuffd_wp(pte); 66535a90d5a1SPeter Xu else if (uffd_wp_resolve) 66545a90d5a1SPeter Xu pte = huge_pte_clear_uffd_wp(pte); 6655023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 66567da4d641SPeter Zijlstra pages++; 665760dfaad6SPeter Xu } else { 665860dfaad6SPeter Xu /* None pte */ 665960dfaad6SPeter Xu if (unlikely(uffd_wp)) 666060dfaad6SPeter Xu /* Safe to modify directly (none->non-present). */ 666160dfaad6SPeter Xu set_huge_pte_at(mm, address, ptep, 6662935d4f0cSRyan Roberts make_pte_marker(PTE_MARKER_UFFD_WP), 6663935d4f0cSRyan Roberts psize); 66648f860591SZhang, Yanmin } 6665cb900f41SKirill A. Shutemov spin_unlock(ptl); 66668f860591SZhang, Yanmin } 6667d833352aSMel Gorman /* 6668c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 6669d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 6670c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 6671dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 6672dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 6673d833352aSMel Gorman */ 6674dff11abeSMike Kravetz if (shared_pmd) 6675ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 6676dff11abeSMike Kravetz else 66775491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 66780f10851eSJérôme Glisse /* 66791af5a810SAlistair Popple * No need to call mmu_notifier_arch_invalidate_secondary_tlbs() we are 66801af5a810SAlistair Popple * downgrading page table protection not changing it to point to a new 66811af5a810SAlistair Popple * page. 66820f10851eSJérôme Glisse * 6683ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 66840f10851eSJérôme Glisse */ 668583cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 668640549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 6687ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 66887da4d641SPeter Zijlstra 6689d1751118SPeter Xu return pages > 0 ? (pages << h->order) : pages; 66908f860591SZhang, Yanmin } 66918f860591SZhang, Yanmin 669233b8f84aSMike Kravetz /* Return true if reservation was successful, false otherwise. */ 669333b8f84aSMike Kravetz bool hugetlb_reserve_pages(struct inode *inode, 6694a1e78772SMel Gorman long from, long to, 66955a6fe125SMel Gorman struct vm_area_struct *vma, 6696ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 6697e4e574b7SAdam Litke { 6698c5094ec7SMike Kravetz long chg = -1, add = -1; 6699a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 670090481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 67019119a41eSJoonsoo Kim struct resv_map *resv_map; 6702075a61d0SMina Almasry struct hugetlb_cgroup *h_cg = NULL; 67030db9d74eSMina Almasry long gbl_reserve, regions_needed = 0; 6704e4e574b7SAdam Litke 670563489f8eSMike Kravetz /* This should never happen */ 670663489f8eSMike Kravetz if (from > to) { 670763489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 670833b8f84aSMike Kravetz return false; 670963489f8eSMike Kravetz } 671063489f8eSMike Kravetz 6711a1e78772SMel Gorman /* 6712e700898fSMike Kravetz * vma specific semaphore used for pmd sharing and fault/truncation 6713e700898fSMike Kravetz * synchronization 67148d9bfb26SMike Kravetz */ 67158d9bfb26SMike Kravetz hugetlb_vma_lock_alloc(vma); 67168d9bfb26SMike Kravetz 67178d9bfb26SMike Kravetz /* 671817c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 671917c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 672090481622SDavid Gibson * without using reserves 672117c9d12eSMel Gorman */ 6722ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 672333b8f84aSMike Kravetz return true; 672417c9d12eSMel Gorman 672517c9d12eSMel Gorman /* 6726a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 6727a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 6728a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 6729a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 6730a1e78772SMel Gorman */ 67319119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 6732f27a5136SMike Kravetz /* 6733f27a5136SMike Kravetz * resv_map can not be NULL as hugetlb_reserve_pages is only 6734f27a5136SMike Kravetz * called for inodes for which resv_maps were created (see 6735f27a5136SMike Kravetz * hugetlbfs_get_inode). 6736f27a5136SMike Kravetz */ 67374e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 67389119a41eSJoonsoo Kim 67390db9d74eSMina Almasry chg = region_chg(resv_map, from, to, ®ions_needed); 67409119a41eSJoonsoo Kim } else { 6741e9fe92aeSMina Almasry /* Private mapping. */ 67429119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 67435a6fe125SMel Gorman if (!resv_map) 67448d9bfb26SMike Kravetz goto out_err; 67455a6fe125SMel Gorman 674617c9d12eSMel Gorman chg = to - from; 674717c9d12eSMel Gorman 67485a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 67495a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 67505a6fe125SMel Gorman } 67515a6fe125SMel Gorman 675233b8f84aSMike Kravetz if (chg < 0) 6753c50ac050SDave Hansen goto out_err; 675417c9d12eSMel Gorman 675533b8f84aSMike Kravetz if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), 675633b8f84aSMike Kravetz chg * pages_per_huge_page(h), &h_cg) < 0) 6757075a61d0SMina Almasry goto out_err; 6758075a61d0SMina Almasry 6759075a61d0SMina Almasry if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { 6760075a61d0SMina Almasry /* For private mappings, the hugetlb_cgroup uncharge info hangs 6761075a61d0SMina Almasry * of the resv_map. 6762075a61d0SMina Almasry */ 6763075a61d0SMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); 6764075a61d0SMina Almasry } 6765075a61d0SMina Almasry 67661c5ecae3SMike Kravetz /* 67671c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 67681c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 67691c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 67701c5ecae3SMike Kravetz */ 67711c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 677233b8f84aSMike Kravetz if (gbl_reserve < 0) 6773075a61d0SMina Almasry goto out_uncharge_cgroup; 677417c9d12eSMel Gorman 677517c9d12eSMel Gorman /* 677617c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 677790481622SDavid Gibson * Hand the pages back to the subpool if there are not 677817c9d12eSMel Gorman */ 677933b8f84aSMike Kravetz if (hugetlb_acct_memory(h, gbl_reserve) < 0) 6780075a61d0SMina Almasry goto out_put_pages; 678117c9d12eSMel Gorman 678217c9d12eSMel Gorman /* 678317c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 678417c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 678517c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 678617c9d12eSMel Gorman * the reservation was and the page cache tells how much of 678717c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 678817c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 678917c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 679017c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 679117c9d12eSMel Gorman * else has to be done for private mappings here 679217c9d12eSMel Gorman */ 679333039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 6794075a61d0SMina Almasry add = region_add(resv_map, from, to, regions_needed, h, h_cg); 679533039678SMike Kravetz 67960db9d74eSMina Almasry if (unlikely(add < 0)) { 67970db9d74eSMina Almasry hugetlb_acct_memory(h, -gbl_reserve); 6798075a61d0SMina Almasry goto out_put_pages; 67990db9d74eSMina Almasry } else if (unlikely(chg > add)) { 680033039678SMike Kravetz /* 680133039678SMike Kravetz * pages in this range were added to the reserve 680233039678SMike Kravetz * map between region_chg and region_add. This 6803d0ce0e47SSidhartha Kumar * indicates a race with alloc_hugetlb_folio. Adjust 680433039678SMike Kravetz * the subpool and reserve counts modified above 680533039678SMike Kravetz * based on the difference. 680633039678SMike Kravetz */ 680733039678SMike Kravetz long rsv_adjust; 680833039678SMike Kravetz 6809d85aecf2SMiaohe Lin /* 6810d85aecf2SMiaohe Lin * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the 6811d85aecf2SMiaohe Lin * reference to h_cg->css. See comment below for detail. 6812d85aecf2SMiaohe Lin */ 6813075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd( 6814075a61d0SMina Almasry hstate_index(h), 6815075a61d0SMina Almasry (chg - add) * pages_per_huge_page(h), h_cg); 6816075a61d0SMina Almasry 681733039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 681833039678SMike Kravetz chg - add); 681933039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 6820d85aecf2SMiaohe Lin } else if (h_cg) { 6821d85aecf2SMiaohe Lin /* 6822d85aecf2SMiaohe Lin * The file_regions will hold their own reference to 6823d85aecf2SMiaohe Lin * h_cg->css. So we should release the reference held 6824d85aecf2SMiaohe Lin * via hugetlb_cgroup_charge_cgroup_rsvd() when we are 6825d85aecf2SMiaohe Lin * done. 6826d85aecf2SMiaohe Lin */ 6827d85aecf2SMiaohe Lin hugetlb_cgroup_put_rsvd_cgroup(h_cg); 682833039678SMike Kravetz } 682933039678SMike Kravetz } 683033b8f84aSMike Kravetz return true; 683133b8f84aSMike Kravetz 6832075a61d0SMina Almasry out_put_pages: 6833075a61d0SMina Almasry /* put back original number of pages, chg */ 6834075a61d0SMina Almasry (void)hugepage_subpool_put_pages(spool, chg); 6835075a61d0SMina Almasry out_uncharge_cgroup: 6836075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), 6837075a61d0SMina Almasry chg * pages_per_huge_page(h), h_cg); 6838c50ac050SDave Hansen out_err: 68398d9bfb26SMike Kravetz hugetlb_vma_lock_free(vma); 68405e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 68410db9d74eSMina Almasry /* Only call region_abort if the region_chg succeeded but the 68420db9d74eSMina Almasry * region_add failed or didn't run. 68430db9d74eSMina Almasry */ 68440db9d74eSMina Almasry if (chg >= 0 && add < 0) 68450db9d74eSMina Almasry region_abort(resv_map, from, to, regions_needed); 684692fe9dcbSRik van Riel if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 6847f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 684892fe9dcbSRik van Riel set_vma_resv_map(vma, NULL); 684992fe9dcbSRik van Riel } 685033b8f84aSMike Kravetz return false; 6851a43a8c39SChen, Kenneth W } 6852a43a8c39SChen, Kenneth W 6853b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 6854b5cec28dSMike Kravetz long freed) 6855a43a8c39SChen, Kenneth W { 6856a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 68574e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 68589119a41eSJoonsoo Kim long chg = 0; 685990481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 68601c5ecae3SMike Kravetz long gbl_reserve; 686145c682a6SKen Chen 6862f27a5136SMike Kravetz /* 6863f27a5136SMike Kravetz * Since this routine can be called in the evict inode path for all 6864f27a5136SMike Kravetz * hugetlbfs inodes, resv_map could be NULL. 6865f27a5136SMike Kravetz */ 6866b5cec28dSMike Kravetz if (resv_map) { 6867b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 6868b5cec28dSMike Kravetz /* 6869b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 6870b5cec28dSMike Kravetz * must be split and another region descriptor can not be 6871b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 6872b5cec28dSMike Kravetz */ 6873b5cec28dSMike Kravetz if (chg < 0) 6874b5cec28dSMike Kravetz return chg; 6875b5cec28dSMike Kravetz } 6876b5cec28dSMike Kravetz 687745c682a6SKen Chen spin_lock(&inode->i_lock); 6878e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 687945c682a6SKen Chen spin_unlock(&inode->i_lock); 688045c682a6SKen Chen 68811c5ecae3SMike Kravetz /* 68821c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 68831c5ecae3SMike Kravetz * reservations to be released may be adjusted. 6884dddf31a4SMiaohe Lin * 6885dddf31a4SMiaohe Lin * Note that !resv_map implies freed == 0. So (chg - freed) 6886dddf31a4SMiaohe Lin * won't go negative. 68871c5ecae3SMike Kravetz */ 68881c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 68891c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 6890b5cec28dSMike Kravetz 6891b5cec28dSMike Kravetz return 0; 6892a43a8c39SChen, Kenneth W } 689393f70f90SNaoya Horiguchi 68943212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 68953212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 68963212b535SSteve Capper struct vm_area_struct *vma, 68973212b535SSteve Capper unsigned long addr, pgoff_t idx) 68983212b535SSteve Capper { 68993212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 69003212b535SSteve Capper svma->vm_start; 69013212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 69023212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 69033212b535SSteve Capper 69043212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 6905e430a95aSSuren Baghdasaryan unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED_MASK; 6906e430a95aSSuren Baghdasaryan unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED_MASK; 69073212b535SSteve Capper 69083212b535SSteve Capper /* 69093212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 69103212b535SSteve Capper * page table page. 6911131a79b4SMike Kravetz * 6912131a79b4SMike Kravetz * Also, vma_lock (vm_private_data) is required for sharing. 69133212b535SSteve Capper */ 69143212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 69153212b535SSteve Capper vm_flags != svm_flags || 6916131a79b4SMike Kravetz !range_in_vma(svma, sbase, s_end) || 6917131a79b4SMike Kravetz !svma->vm_private_data) 69183212b535SSteve Capper return 0; 69193212b535SSteve Capper 69203212b535SSteve Capper return saddr; 69213212b535SSteve Capper } 69223212b535SSteve Capper 6923bbff39ccSMike Kravetz bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 69243212b535SSteve Capper { 6925bbff39ccSMike Kravetz unsigned long start = addr & PUD_MASK; 6926bbff39ccSMike Kravetz unsigned long end = start + PUD_SIZE; 6927bbff39ccSMike Kravetz 69288d9bfb26SMike Kravetz #ifdef CONFIG_USERFAULTFD 69298d9bfb26SMike Kravetz if (uffd_disable_huge_pmd_share(vma)) 69308d9bfb26SMike Kravetz return false; 69318d9bfb26SMike Kravetz #endif 69323212b535SSteve Capper /* 69333212b535SSteve Capper * check on proper vm_flags and page table alignment 69343212b535SSteve Capper */ 69358d9bfb26SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 693631aafb45SNicholas Krause return false; 6937bbff39ccSMike Kravetz if (!vma->vm_private_data) /* vma lock required for sharing */ 69388d9bfb26SMike Kravetz return false; 69398d9bfb26SMike Kravetz if (!range_in_vma(vma, start, end)) 69408d9bfb26SMike Kravetz return false; 69418d9bfb26SMike Kravetz return true; 69428d9bfb26SMike Kravetz } 69438d9bfb26SMike Kravetz 69443212b535SSteve Capper /* 6945017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 6946017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 6947017b1660SMike Kravetz * shared pmd mappings. 6948017b1660SMike Kravetz */ 6949017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 6950017b1660SMike Kravetz unsigned long *start, unsigned long *end) 6951017b1660SMike Kravetz { 6952a1ba9da8SLi Xinhai unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), 6953a1ba9da8SLi Xinhai v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 6954017b1660SMike Kravetz 6955a1ba9da8SLi Xinhai /* 6956f0953a1bSIngo Molnar * vma needs to span at least one aligned PUD size, and the range 6957f0953a1bSIngo Molnar * must be at least partially within in. 6958a1ba9da8SLi Xinhai */ 6959a1ba9da8SLi Xinhai if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || 6960a1ba9da8SLi Xinhai (*end <= v_start) || (*start >= v_end)) 6961017b1660SMike Kravetz return; 6962017b1660SMike Kravetz 696375802ca6SPeter Xu /* Extend the range to be PUD aligned for a worst case scenario */ 6964a1ba9da8SLi Xinhai if (*start > v_start) 6965a1ba9da8SLi Xinhai *start = ALIGN_DOWN(*start, PUD_SIZE); 6966017b1660SMike Kravetz 6967a1ba9da8SLi Xinhai if (*end < v_end) 6968a1ba9da8SLi Xinhai *end = ALIGN(*end, PUD_SIZE); 6969017b1660SMike Kravetz } 6970017b1660SMike Kravetz 6971017b1660SMike Kravetz /* 69723212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 69733212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 69743212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 69753a47c54fSMike Kravetz * code much cleaner. pmd allocation is essential for the shared case because 69763a47c54fSMike Kravetz * pud has to be populated inside the same i_mmap_rwsem section - otherwise 69773a47c54fSMike Kravetz * racing tasks could either miss the sharing (see huge_pte_offset) or select a 69783a47c54fSMike Kravetz * bad pmd for sharing. 69793212b535SSteve Capper */ 6980aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 6981aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 69823212b535SSteve Capper { 69833212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 69843212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 69853212b535SSteve Capper vma->vm_pgoff; 69863212b535SSteve Capper struct vm_area_struct *svma; 69873212b535SSteve Capper unsigned long saddr; 69883212b535SSteve Capper pte_t *spte = NULL; 69893212b535SSteve Capper pte_t *pte; 69903212b535SSteve Capper 69913a47c54fSMike Kravetz i_mmap_lock_read(mapping); 69923212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 69933212b535SSteve Capper if (svma == vma) 69943212b535SSteve Capper continue; 69953212b535SSteve Capper 69963212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 69973212b535SSteve Capper if (saddr) { 69989c67a207SPeter Xu spte = hugetlb_walk(svma, saddr, 69997868a208SPunit Agrawal vma_mmu_pagesize(svma)); 70003212b535SSteve Capper if (spte) { 70013212b535SSteve Capper get_page(virt_to_page(spte)); 70023212b535SSteve Capper break; 70033212b535SSteve Capper } 70043212b535SSteve Capper } 70053212b535SSteve Capper } 70063212b535SSteve Capper 70073212b535SSteve Capper if (!spte) 70083212b535SSteve Capper goto out; 70093212b535SSteve Capper 7010349d1670SPeter Xu spin_lock(&mm->page_table_lock); 7011dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 70123212b535SSteve Capper pud_populate(mm, pud, 70133212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 7014c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 7015dc6c9a35SKirill A. Shutemov } else { 70163212b535SSteve Capper put_page(virt_to_page(spte)); 7017dc6c9a35SKirill A. Shutemov } 7018349d1670SPeter Xu spin_unlock(&mm->page_table_lock); 70193212b535SSteve Capper out: 70203212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 70213a47c54fSMike Kravetz i_mmap_unlock_read(mapping); 70223212b535SSteve Capper return pte; 70233212b535SSteve Capper } 70243212b535SSteve Capper 70253212b535SSteve Capper /* 70263212b535SSteve Capper * unmap huge page backed by shared pte. 70273212b535SSteve Capper * 70283212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 70293212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 70303212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 70313212b535SSteve Capper * 70323a47c54fSMike Kravetz * Called with page table lock held. 70333212b535SSteve Capper * 70343212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 70353212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 70363212b535SSteve Capper */ 703734ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 70384ddb4d91SMike Kravetz unsigned long addr, pte_t *ptep) 70393212b535SSteve Capper { 70404ddb4d91SMike Kravetz pgd_t *pgd = pgd_offset(mm, addr); 70414ddb4d91SMike Kravetz p4d_t *p4d = p4d_offset(pgd, addr); 70424ddb4d91SMike Kravetz pud_t *pud = pud_offset(p4d, addr); 70433212b535SSteve Capper 704434ae204fSMike Kravetz i_mmap_assert_write_locked(vma->vm_file->f_mapping); 704540549ba8SMike Kravetz hugetlb_vma_assert_locked(vma); 70463212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 70473212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 70483212b535SSteve Capper return 0; 70493212b535SSteve Capper 70503212b535SSteve Capper pud_clear(pud); 70513212b535SSteve Capper put_page(virt_to_page(ptep)); 7052dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 70533212b535SSteve Capper return 1; 70543212b535SSteve Capper } 7055c1991e07SPeter Xu 70569e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 70578d9bfb26SMike Kravetz 7058aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 7059aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 70609e5fc74cSSteve Capper { 70619e5fc74cSSteve Capper return NULL; 70629e5fc74cSSteve Capper } 7063e81f2d22SZhang Zhen 706434ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 70654ddb4d91SMike Kravetz unsigned long addr, pte_t *ptep) 7066e81f2d22SZhang Zhen { 7067e81f2d22SZhang Zhen return 0; 7068e81f2d22SZhang Zhen } 7069017b1660SMike Kravetz 7070017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 7071017b1660SMike Kravetz unsigned long *start, unsigned long *end) 7072017b1660SMike Kravetz { 7073017b1660SMike Kravetz } 7074c1991e07SPeter Xu 7075c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 7076c1991e07SPeter Xu { 7077c1991e07SPeter Xu return false; 7078c1991e07SPeter Xu } 70793212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 70803212b535SSteve Capper 70819e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 7082aec44e0fSPeter Xu pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, 70839e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 70849e5fc74cSSteve Capper { 70859e5fc74cSSteve Capper pgd_t *pgd; 7086c2febafcSKirill A. Shutemov p4d_t *p4d; 70879e5fc74cSSteve Capper pud_t *pud; 70889e5fc74cSSteve Capper pte_t *pte = NULL; 70899e5fc74cSSteve Capper 70909e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 7091f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 7092f4f0a3d8SKirill A. Shutemov if (!p4d) 7093f4f0a3d8SKirill A. Shutemov return NULL; 7094c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 70959e5fc74cSSteve Capper if (pud) { 70969e5fc74cSSteve Capper if (sz == PUD_SIZE) { 70979e5fc74cSSteve Capper pte = (pte_t *)pud; 70989e5fc74cSSteve Capper } else { 70999e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 7100c1991e07SPeter Xu if (want_pmd_share(vma, addr) && pud_none(*pud)) 7101aec44e0fSPeter Xu pte = huge_pmd_share(mm, vma, addr, pud); 71029e5fc74cSSteve Capper else 71039e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 71049e5fc74cSSteve Capper } 71059e5fc74cSSteve Capper } 7106191fcdb6SJohn Hubbard 7107191fcdb6SJohn Hubbard if (pte) { 7108191fcdb6SJohn Hubbard pte_t pteval = ptep_get_lockless(pte); 7109191fcdb6SJohn Hubbard 7110191fcdb6SJohn Hubbard BUG_ON(pte_present(pteval) && !pte_huge(pteval)); 7111191fcdb6SJohn Hubbard } 71129e5fc74cSSteve Capper 71139e5fc74cSSteve Capper return pte; 71149e5fc74cSSteve Capper } 71159e5fc74cSSteve Capper 71169b19df29SPunit Agrawal /* 71179b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 71189b19df29SPunit Agrawal * entry at address @addr 71199b19df29SPunit Agrawal * 71208ac0b81aSLi Xinhai * Return: Pointer to page table entry (PUD or PMD) for 71218ac0b81aSLi Xinhai * address @addr, or NULL if a !p*d_present() entry is encountered and the 71229b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 71239b19df29SPunit Agrawal * table. 71249b19df29SPunit Agrawal */ 71257868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 71267868a208SPunit Agrawal unsigned long addr, unsigned long sz) 71279e5fc74cSSteve Capper { 71289e5fc74cSSteve Capper pgd_t *pgd; 7129c2febafcSKirill A. Shutemov p4d_t *p4d; 71308ac0b81aSLi Xinhai pud_t *pud; 71318ac0b81aSLi Xinhai pmd_t *pmd; 71329e5fc74cSSteve Capper 71339e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 7134c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 7135c2febafcSKirill A. Shutemov return NULL; 7136c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 7137c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 7138c2febafcSKirill A. Shutemov return NULL; 71399b19df29SPunit Agrawal 7140c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 71418ac0b81aSLi Xinhai if (sz == PUD_SIZE) 71428ac0b81aSLi Xinhai /* must be pud huge, non-present or none */ 71439e5fc74cSSteve Capper return (pte_t *)pud; 71448ac0b81aSLi Xinhai if (!pud_present(*pud)) 71458ac0b81aSLi Xinhai return NULL; 71468ac0b81aSLi Xinhai /* must have a valid entry and size to go further */ 71479b19df29SPunit Agrawal 71489e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 71498ac0b81aSLi Xinhai /* must be pmd huge, non-present or none */ 71509e5fc74cSSteve Capper return (pte_t *)pmd; 71519e5fc74cSSteve Capper } 71529e5fc74cSSteve Capper 7153e95a9851SMike Kravetz /* 7154e95a9851SMike Kravetz * Return a mask that can be used to update an address to the last huge 7155e95a9851SMike Kravetz * page in a page table page mapping size. Used to skip non-present 7156e95a9851SMike Kravetz * page table entries when linearly scanning address ranges. Architectures 7157e95a9851SMike Kravetz * with unique huge page to page table relationships can define their own 7158e95a9851SMike Kravetz * version of this routine. 7159e95a9851SMike Kravetz */ 7160e95a9851SMike Kravetz unsigned long hugetlb_mask_last_page(struct hstate *h) 7161e95a9851SMike Kravetz { 7162e95a9851SMike Kravetz unsigned long hp_size = huge_page_size(h); 7163e95a9851SMike Kravetz 7164e95a9851SMike Kravetz if (hp_size == PUD_SIZE) 7165e95a9851SMike Kravetz return P4D_SIZE - PUD_SIZE; 7166e95a9851SMike Kravetz else if (hp_size == PMD_SIZE) 7167e95a9851SMike Kravetz return PUD_SIZE - PMD_SIZE; 7168e95a9851SMike Kravetz else 7169e95a9851SMike Kravetz return 0UL; 7170e95a9851SMike Kravetz } 7171e95a9851SMike Kravetz 7172e95a9851SMike Kravetz #else 7173e95a9851SMike Kravetz 7174e95a9851SMike Kravetz /* See description above. Architectures can provide their own version. */ 7175e95a9851SMike Kravetz __weak unsigned long hugetlb_mask_last_page(struct hstate *h) 7176e95a9851SMike Kravetz { 71774ddb4d91SMike Kravetz #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 71784ddb4d91SMike Kravetz if (huge_page_size(h) == PMD_SIZE) 71794ddb4d91SMike Kravetz return PUD_SIZE - PMD_SIZE; 71804ddb4d91SMike Kravetz #endif 7181e95a9851SMike Kravetz return 0UL; 7182e95a9851SMike Kravetz } 7183e95a9851SMike Kravetz 718461f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 718561f77edaSNaoya Horiguchi 718661f77edaSNaoya Horiguchi /* 718761f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 718861f77edaSNaoya Horiguchi * behavior. 718961f77edaSNaoya Horiguchi */ 71909747b9e9SBaolin Wang bool isolate_hugetlb(struct folio *folio, struct list_head *list) 719131caf665SNaoya Horiguchi { 71929747b9e9SBaolin Wang bool ret = true; 7193bcc54222SNaoya Horiguchi 7194db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 71956aa3a920SSidhartha Kumar if (!folio_test_hugetlb(folio) || 71966aa3a920SSidhartha Kumar !folio_test_hugetlb_migratable(folio) || 71976aa3a920SSidhartha Kumar !folio_try_get(folio)) { 71989747b9e9SBaolin Wang ret = false; 7199bcc54222SNaoya Horiguchi goto unlock; 7200bcc54222SNaoya Horiguchi } 72016aa3a920SSidhartha Kumar folio_clear_hugetlb_migratable(folio); 72026aa3a920SSidhartha Kumar list_move_tail(&folio->lru, list); 7203bcc54222SNaoya Horiguchi unlock: 7204db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7205bcc54222SNaoya Horiguchi return ret; 720631caf665SNaoya Horiguchi } 720731caf665SNaoya Horiguchi 720804bac040SSidhartha Kumar int get_hwpoison_hugetlb_folio(struct folio *folio, bool *hugetlb, bool unpoison) 720925182f05SNaoya Horiguchi { 721025182f05SNaoya Horiguchi int ret = 0; 721125182f05SNaoya Horiguchi 721225182f05SNaoya Horiguchi *hugetlb = false; 721325182f05SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 721404bac040SSidhartha Kumar if (folio_test_hugetlb(folio)) { 721525182f05SNaoya Horiguchi *hugetlb = true; 721604bac040SSidhartha Kumar if (folio_test_hugetlb_freed(folio)) 7217b283d983SNaoya Horiguchi ret = 0; 721804bac040SSidhartha Kumar else if (folio_test_hugetlb_migratable(folio) || unpoison) 721904bac040SSidhartha Kumar ret = folio_try_get(folio); 72200ed950d1SNaoya Horiguchi else 72210ed950d1SNaoya Horiguchi ret = -EBUSY; 722225182f05SNaoya Horiguchi } 722325182f05SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 722425182f05SNaoya Horiguchi return ret; 722525182f05SNaoya Horiguchi } 722625182f05SNaoya Horiguchi 7227e591ef7dSNaoya Horiguchi int get_huge_page_for_hwpoison(unsigned long pfn, int flags, 7228e591ef7dSNaoya Horiguchi bool *migratable_cleared) 7229405ce051SNaoya Horiguchi { 7230405ce051SNaoya Horiguchi int ret; 7231405ce051SNaoya Horiguchi 7232405ce051SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 7233e591ef7dSNaoya Horiguchi ret = __get_huge_page_for_hwpoison(pfn, flags, migratable_cleared); 7234405ce051SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 7235405ce051SNaoya Horiguchi return ret; 7236405ce051SNaoya Horiguchi } 7237405ce051SNaoya Horiguchi 7238ea8e72f4SSidhartha Kumar void folio_putback_active_hugetlb(struct folio *folio) 723931caf665SNaoya Horiguchi { 7240db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 7241ea8e72f4SSidhartha Kumar folio_set_hugetlb_migratable(folio); 7242ea8e72f4SSidhartha Kumar list_move_tail(&folio->lru, &(folio_hstate(folio))->hugepage_activelist); 7243db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7244ea8e72f4SSidhartha Kumar folio_put(folio); 724531caf665SNaoya Horiguchi } 7246ab5ac90aSMichal Hocko 7247345c62d1SSidhartha Kumar void move_hugetlb_state(struct folio *old_folio, struct folio *new_folio, int reason) 7248ab5ac90aSMichal Hocko { 7249345c62d1SSidhartha Kumar struct hstate *h = folio_hstate(old_folio); 7250ab5ac90aSMichal Hocko 7251345c62d1SSidhartha Kumar hugetlb_cgroup_migrate(old_folio, new_folio); 7252345c62d1SSidhartha Kumar set_page_owner_migrate_reason(&new_folio->page, reason); 7253ab5ac90aSMichal Hocko 7254ab5ac90aSMichal Hocko /* 7255345c62d1SSidhartha Kumar * transfer temporary state of the new hugetlb folio. This is 7256ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 7257ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 7258ab5ac90aSMichal Hocko * the temporary status. 7259ab5ac90aSMichal Hocko * 7260ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 7261ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 7262ab5ac90aSMichal Hocko * the per-node's. 7263ab5ac90aSMichal Hocko */ 7264345c62d1SSidhartha Kumar if (folio_test_hugetlb_temporary(new_folio)) { 7265345c62d1SSidhartha Kumar int old_nid = folio_nid(old_folio); 7266345c62d1SSidhartha Kumar int new_nid = folio_nid(new_folio); 7267ab5ac90aSMichal Hocko 7268345c62d1SSidhartha Kumar folio_set_hugetlb_temporary(old_folio); 7269345c62d1SSidhartha Kumar folio_clear_hugetlb_temporary(new_folio); 7270345c62d1SSidhartha Kumar 7271ab5ac90aSMichal Hocko 72725af1ab1dSMiaohe Lin /* 72735af1ab1dSMiaohe Lin * There is no need to transfer the per-node surplus state 72745af1ab1dSMiaohe Lin * when we do not cross the node. 72755af1ab1dSMiaohe Lin */ 72765af1ab1dSMiaohe Lin if (new_nid == old_nid) 72775af1ab1dSMiaohe Lin return; 7278db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 7279ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 7280ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 7281ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 7282ab5ac90aSMichal Hocko } 7283db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7284ab5ac90aSMichal Hocko } 7285ab5ac90aSMichal Hocko } 7286cf11e85fSRoman Gushchin 7287b30c14cdSJames Houghton static void hugetlb_unshare_pmds(struct vm_area_struct *vma, 7288b30c14cdSJames Houghton unsigned long start, 7289b30c14cdSJames Houghton unsigned long end) 72906dfeaff9SPeter Xu { 72916dfeaff9SPeter Xu struct hstate *h = hstate_vma(vma); 72926dfeaff9SPeter Xu unsigned long sz = huge_page_size(h); 72936dfeaff9SPeter Xu struct mm_struct *mm = vma->vm_mm; 72946dfeaff9SPeter Xu struct mmu_notifier_range range; 7295b30c14cdSJames Houghton unsigned long address; 72966dfeaff9SPeter Xu spinlock_t *ptl; 72976dfeaff9SPeter Xu pte_t *ptep; 72986dfeaff9SPeter Xu 72996dfeaff9SPeter Xu if (!(vma->vm_flags & VM_MAYSHARE)) 73006dfeaff9SPeter Xu return; 73016dfeaff9SPeter Xu 73026dfeaff9SPeter Xu if (start >= end) 73036dfeaff9SPeter Xu return; 73046dfeaff9SPeter Xu 73059c8bbfacSBaolin Wang flush_cache_range(vma, start, end); 73066dfeaff9SPeter Xu /* 73076dfeaff9SPeter Xu * No need to call adjust_range_if_pmd_sharing_possible(), because 73086dfeaff9SPeter Xu * we have already done the PUD_SIZE alignment. 73096dfeaff9SPeter Xu */ 73107d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, 73116dfeaff9SPeter Xu start, end); 73126dfeaff9SPeter Xu mmu_notifier_invalidate_range_start(&range); 731340549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 73146dfeaff9SPeter Xu i_mmap_lock_write(vma->vm_file->f_mapping); 73156dfeaff9SPeter Xu for (address = start; address < end; address += PUD_SIZE) { 73169c67a207SPeter Xu ptep = hugetlb_walk(vma, address, sz); 73176dfeaff9SPeter Xu if (!ptep) 73186dfeaff9SPeter Xu continue; 73196dfeaff9SPeter Xu ptl = huge_pte_lock(h, mm, ptep); 73204ddb4d91SMike Kravetz huge_pmd_unshare(mm, vma, address, ptep); 73216dfeaff9SPeter Xu spin_unlock(ptl); 73226dfeaff9SPeter Xu } 73236dfeaff9SPeter Xu flush_hugetlb_tlb_range(vma, start, end); 73246dfeaff9SPeter Xu i_mmap_unlock_write(vma->vm_file->f_mapping); 732540549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 73266dfeaff9SPeter Xu /* 73271af5a810SAlistair Popple * No need to call mmu_notifier_arch_invalidate_secondary_tlbs(), see 7328ee65728eSMike Rapoport * Documentation/mm/mmu_notifier.rst. 73296dfeaff9SPeter Xu */ 73306dfeaff9SPeter Xu mmu_notifier_invalidate_range_end(&range); 73316dfeaff9SPeter Xu } 73326dfeaff9SPeter Xu 7333b30c14cdSJames Houghton /* 7334b30c14cdSJames Houghton * This function will unconditionally remove all the shared pmd pgtable entries 7335b30c14cdSJames Houghton * within the specific vma for a hugetlbfs memory range. 7336b30c14cdSJames Houghton */ 7337b30c14cdSJames Houghton void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) 7338b30c14cdSJames Houghton { 7339b30c14cdSJames Houghton hugetlb_unshare_pmds(vma, ALIGN(vma->vm_start, PUD_SIZE), 7340b30c14cdSJames Houghton ALIGN_DOWN(vma->vm_end, PUD_SIZE)); 7341b30c14cdSJames Houghton } 7342b30c14cdSJames Houghton 7343cf11e85fSRoman Gushchin #ifdef CONFIG_CMA 7344cf11e85fSRoman Gushchin static bool cma_reserve_called __initdata; 7345cf11e85fSRoman Gushchin 7346cf11e85fSRoman Gushchin static int __init cmdline_parse_hugetlb_cma(char *p) 7347cf11e85fSRoman Gushchin { 734838e719abSBaolin Wang int nid, count = 0; 734938e719abSBaolin Wang unsigned long tmp; 735038e719abSBaolin Wang char *s = p; 735138e719abSBaolin Wang 735238e719abSBaolin Wang while (*s) { 735338e719abSBaolin Wang if (sscanf(s, "%lu%n", &tmp, &count) != 1) 735438e719abSBaolin Wang break; 735538e719abSBaolin Wang 735638e719abSBaolin Wang if (s[count] == ':') { 7357f9317f77SMike Kravetz if (tmp >= MAX_NUMNODES) 735838e719abSBaolin Wang break; 7359f9317f77SMike Kravetz nid = array_index_nospec(tmp, MAX_NUMNODES); 736038e719abSBaolin Wang 736138e719abSBaolin Wang s += count + 1; 736238e719abSBaolin Wang tmp = memparse(s, &s); 736338e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = tmp; 736438e719abSBaolin Wang hugetlb_cma_size += tmp; 736538e719abSBaolin Wang 736638e719abSBaolin Wang /* 736738e719abSBaolin Wang * Skip the separator if have one, otherwise 736838e719abSBaolin Wang * break the parsing. 736938e719abSBaolin Wang */ 737038e719abSBaolin Wang if (*s == ',') 737138e719abSBaolin Wang s++; 737238e719abSBaolin Wang else 737338e719abSBaolin Wang break; 737438e719abSBaolin Wang } else { 7375cf11e85fSRoman Gushchin hugetlb_cma_size = memparse(p, &p); 737638e719abSBaolin Wang break; 737738e719abSBaolin Wang } 737838e719abSBaolin Wang } 737938e719abSBaolin Wang 7380cf11e85fSRoman Gushchin return 0; 7381cf11e85fSRoman Gushchin } 7382cf11e85fSRoman Gushchin 7383cf11e85fSRoman Gushchin early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); 7384cf11e85fSRoman Gushchin 7385cf11e85fSRoman Gushchin void __init hugetlb_cma_reserve(int order) 7386cf11e85fSRoman Gushchin { 7387cf11e85fSRoman Gushchin unsigned long size, reserved, per_node; 738838e719abSBaolin Wang bool node_specific_cma_alloc = false; 7389cf11e85fSRoman Gushchin int nid; 7390cf11e85fSRoman Gushchin 7391cf11e85fSRoman Gushchin cma_reserve_called = true; 7392cf11e85fSRoman Gushchin 7393cf11e85fSRoman Gushchin if (!hugetlb_cma_size) 7394cf11e85fSRoman Gushchin return; 7395cf11e85fSRoman Gushchin 739638e719abSBaolin Wang for (nid = 0; nid < MAX_NUMNODES; nid++) { 739738e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 739838e719abSBaolin Wang continue; 739938e719abSBaolin Wang 740030a51400SPeng Liu if (!node_online(nid)) { 740138e719abSBaolin Wang pr_warn("hugetlb_cma: invalid node %d specified\n", nid); 740238e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 740338e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 740438e719abSBaolin Wang continue; 740538e719abSBaolin Wang } 740638e719abSBaolin Wang 740738e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) { 740838e719abSBaolin Wang pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n", 740938e719abSBaolin Wang nid, (PAGE_SIZE << order) / SZ_1M); 741038e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 741138e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 741238e719abSBaolin Wang } else { 741338e719abSBaolin Wang node_specific_cma_alloc = true; 741438e719abSBaolin Wang } 741538e719abSBaolin Wang } 741638e719abSBaolin Wang 741738e719abSBaolin Wang /* Validate the CMA size again in case some invalid nodes specified. */ 741838e719abSBaolin Wang if (!hugetlb_cma_size) 741938e719abSBaolin Wang return; 742038e719abSBaolin Wang 7421cf11e85fSRoman Gushchin if (hugetlb_cma_size < (PAGE_SIZE << order)) { 7422cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", 7423cf11e85fSRoman Gushchin (PAGE_SIZE << order) / SZ_1M); 7424a01f4390SMike Kravetz hugetlb_cma_size = 0; 7425cf11e85fSRoman Gushchin return; 7426cf11e85fSRoman Gushchin } 7427cf11e85fSRoman Gushchin 742838e719abSBaolin Wang if (!node_specific_cma_alloc) { 7429cf11e85fSRoman Gushchin /* 7430cf11e85fSRoman Gushchin * If 3 GB area is requested on a machine with 4 numa nodes, 7431cf11e85fSRoman Gushchin * let's allocate 1 GB on first three nodes and ignore the last one. 7432cf11e85fSRoman Gushchin */ 7433cf11e85fSRoman Gushchin per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); 7434cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", 7435cf11e85fSRoman Gushchin hugetlb_cma_size / SZ_1M, per_node / SZ_1M); 743638e719abSBaolin Wang } 7437cf11e85fSRoman Gushchin 7438cf11e85fSRoman Gushchin reserved = 0; 743930a51400SPeng Liu for_each_online_node(nid) { 7440cf11e85fSRoman Gushchin int res; 74412281f797SBarry Song char name[CMA_MAX_NAME]; 7442cf11e85fSRoman Gushchin 744338e719abSBaolin Wang if (node_specific_cma_alloc) { 744438e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 744538e719abSBaolin Wang continue; 744638e719abSBaolin Wang 744738e719abSBaolin Wang size = hugetlb_cma_size_in_node[nid]; 744838e719abSBaolin Wang } else { 7449cf11e85fSRoman Gushchin size = min(per_node, hugetlb_cma_size - reserved); 745038e719abSBaolin Wang } 745138e719abSBaolin Wang 7452cf11e85fSRoman Gushchin size = round_up(size, PAGE_SIZE << order); 7453cf11e85fSRoman Gushchin 74542281f797SBarry Song snprintf(name, sizeof(name), "hugetlb%d", nid); 7455a01f4390SMike Kravetz /* 7456a01f4390SMike Kravetz * Note that 'order per bit' is based on smallest size that 7457a01f4390SMike Kravetz * may be returned to CMA allocator in the case of 7458a01f4390SMike Kravetz * huge page demotion. 7459a01f4390SMike Kravetz */ 7460a01f4390SMike Kravetz res = cma_declare_contiguous_nid(0, size, 0, 7461a01f4390SMike Kravetz PAGE_SIZE << HUGETLB_PAGE_ORDER, 746229d0f41dSBarry Song 0, false, name, 7463cf11e85fSRoman Gushchin &hugetlb_cma[nid], nid); 7464cf11e85fSRoman Gushchin if (res) { 7465cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: reservation failed: err %d, node %d", 7466cf11e85fSRoman Gushchin res, nid); 7467cf11e85fSRoman Gushchin continue; 7468cf11e85fSRoman Gushchin } 7469cf11e85fSRoman Gushchin 7470cf11e85fSRoman Gushchin reserved += size; 7471cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", 7472cf11e85fSRoman Gushchin size / SZ_1M, nid); 7473cf11e85fSRoman Gushchin 7474cf11e85fSRoman Gushchin if (reserved >= hugetlb_cma_size) 7475cf11e85fSRoman Gushchin break; 7476cf11e85fSRoman Gushchin } 7477a01f4390SMike Kravetz 7478a01f4390SMike Kravetz if (!reserved) 7479a01f4390SMike Kravetz /* 7480a01f4390SMike Kravetz * hugetlb_cma_size is used to determine if allocations from 7481a01f4390SMike Kravetz * cma are possible. Set to zero if no cma regions are set up. 7482a01f4390SMike Kravetz */ 7483a01f4390SMike Kravetz hugetlb_cma_size = 0; 7484cf11e85fSRoman Gushchin } 7485cf11e85fSRoman Gushchin 7486263b8998SMiaohe Lin static void __init hugetlb_cma_check(void) 7487cf11e85fSRoman Gushchin { 7488cf11e85fSRoman Gushchin if (!hugetlb_cma_size || cma_reserve_called) 7489cf11e85fSRoman Gushchin return; 7490cf11e85fSRoman Gushchin 7491cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); 7492cf11e85fSRoman Gushchin } 7493cf11e85fSRoman Gushchin 7494cf11e85fSRoman Gushchin #endif /* CONFIG_CMA */ 7495