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); 1007ca02d0aSMike Kravetz 1011d88433bSMiaohe Lin static inline bool subpool_is_free(struct hugepage_subpool *spool) 1021d88433bSMiaohe Lin { 1031d88433bSMiaohe Lin if (spool->count) 1041d88433bSMiaohe Lin return false; 1051d88433bSMiaohe Lin if (spool->max_hpages != -1) 1061d88433bSMiaohe Lin return spool->used_hpages == 0; 1071d88433bSMiaohe Lin if (spool->min_hpages != -1) 1081d88433bSMiaohe Lin return spool->rsv_hpages == spool->min_hpages; 1091d88433bSMiaohe Lin 1101d88433bSMiaohe Lin return true; 1111d88433bSMiaohe Lin } 1121d88433bSMiaohe Lin 113db71ef79SMike Kravetz static inline void unlock_or_release_subpool(struct hugepage_subpool *spool, 114db71ef79SMike Kravetz unsigned long irq_flags) 11590481622SDavid Gibson { 116db71ef79SMike Kravetz spin_unlock_irqrestore(&spool->lock, irq_flags); 11790481622SDavid Gibson 11890481622SDavid Gibson /* If no pages are used, and no other handles to the subpool 1197c8de358SEthon Paul * remain, give up any reservations based on minimum size and 1207ca02d0aSMike Kravetz * free the subpool */ 1211d88433bSMiaohe Lin if (subpool_is_free(spool)) { 1227ca02d0aSMike Kravetz if (spool->min_hpages != -1) 1237ca02d0aSMike Kravetz hugetlb_acct_memory(spool->hstate, 1247ca02d0aSMike Kravetz -spool->min_hpages); 12590481622SDavid Gibson kfree(spool); 12690481622SDavid Gibson } 1277ca02d0aSMike Kravetz } 12890481622SDavid Gibson 1297ca02d0aSMike Kravetz struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, 1307ca02d0aSMike Kravetz long min_hpages) 13190481622SDavid Gibson { 13290481622SDavid Gibson struct hugepage_subpool *spool; 13390481622SDavid Gibson 134c6a91820SMike Kravetz spool = kzalloc(sizeof(*spool), GFP_KERNEL); 13590481622SDavid Gibson if (!spool) 13690481622SDavid Gibson return NULL; 13790481622SDavid Gibson 13890481622SDavid Gibson spin_lock_init(&spool->lock); 13990481622SDavid Gibson spool->count = 1; 1407ca02d0aSMike Kravetz spool->max_hpages = max_hpages; 1417ca02d0aSMike Kravetz spool->hstate = h; 1427ca02d0aSMike Kravetz spool->min_hpages = min_hpages; 1437ca02d0aSMike Kravetz 1447ca02d0aSMike Kravetz if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { 1457ca02d0aSMike Kravetz kfree(spool); 1467ca02d0aSMike Kravetz return NULL; 1477ca02d0aSMike Kravetz } 1487ca02d0aSMike Kravetz spool->rsv_hpages = min_hpages; 14990481622SDavid Gibson 15090481622SDavid Gibson return spool; 15190481622SDavid Gibson } 15290481622SDavid Gibson 15390481622SDavid Gibson void hugepage_put_subpool(struct hugepage_subpool *spool) 15490481622SDavid Gibson { 155db71ef79SMike Kravetz unsigned long flags; 156db71ef79SMike Kravetz 157db71ef79SMike Kravetz spin_lock_irqsave(&spool->lock, flags); 15890481622SDavid Gibson BUG_ON(!spool->count); 15990481622SDavid Gibson spool->count--; 160db71ef79SMike Kravetz unlock_or_release_subpool(spool, flags); 16190481622SDavid Gibson } 16290481622SDavid Gibson 1631c5ecae3SMike Kravetz /* 1641c5ecae3SMike Kravetz * Subpool accounting for allocating and reserving pages. 1651c5ecae3SMike Kravetz * Return -ENOMEM if there are not enough resources to satisfy the 1669e7ee400SRandy Dunlap * request. Otherwise, return the number of pages by which the 1671c5ecae3SMike Kravetz * global pools must be adjusted (upward). The returned value may 1681c5ecae3SMike Kravetz * only be different than the passed value (delta) in the case where 1697c8de358SEthon Paul * a subpool minimum size must be maintained. 1701c5ecae3SMike Kravetz */ 1711c5ecae3SMike Kravetz static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, 17290481622SDavid Gibson long delta) 17390481622SDavid Gibson { 1741c5ecae3SMike Kravetz long ret = delta; 17590481622SDavid Gibson 17690481622SDavid Gibson if (!spool) 1771c5ecae3SMike Kravetz return ret; 17890481622SDavid Gibson 179db71ef79SMike Kravetz spin_lock_irq(&spool->lock); 18090481622SDavid Gibson 1811c5ecae3SMike Kravetz if (spool->max_hpages != -1) { /* maximum size accounting */ 1821c5ecae3SMike Kravetz if ((spool->used_hpages + delta) <= spool->max_hpages) 1831c5ecae3SMike Kravetz spool->used_hpages += delta; 1841c5ecae3SMike Kravetz else { 1851c5ecae3SMike Kravetz ret = -ENOMEM; 1861c5ecae3SMike Kravetz goto unlock_ret; 1871c5ecae3SMike Kravetz } 1881c5ecae3SMike Kravetz } 1891c5ecae3SMike Kravetz 19009a95e29SMike Kravetz /* minimum size accounting */ 19109a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->rsv_hpages) { 1921c5ecae3SMike Kravetz if (delta > spool->rsv_hpages) { 1931c5ecae3SMike Kravetz /* 1941c5ecae3SMike Kravetz * Asking for more reserves than those already taken on 1951c5ecae3SMike Kravetz * behalf of subpool. Return difference. 1961c5ecae3SMike Kravetz */ 1971c5ecae3SMike Kravetz ret = delta - spool->rsv_hpages; 1981c5ecae3SMike Kravetz spool->rsv_hpages = 0; 1991c5ecae3SMike Kravetz } else { 2001c5ecae3SMike Kravetz ret = 0; /* reserves already accounted for */ 2011c5ecae3SMike Kravetz spool->rsv_hpages -= delta; 2021c5ecae3SMike Kravetz } 2031c5ecae3SMike Kravetz } 2041c5ecae3SMike Kravetz 2051c5ecae3SMike Kravetz unlock_ret: 206db71ef79SMike Kravetz spin_unlock_irq(&spool->lock); 20790481622SDavid Gibson return ret; 20890481622SDavid Gibson } 20990481622SDavid Gibson 2101c5ecae3SMike Kravetz /* 2111c5ecae3SMike Kravetz * Subpool accounting for freeing and unreserving pages. 2121c5ecae3SMike Kravetz * Return the number of global page reservations that must be dropped. 2131c5ecae3SMike Kravetz * The return value may only be different than the passed value (delta) 2141c5ecae3SMike Kravetz * in the case where a subpool minimum size must be maintained. 2151c5ecae3SMike Kravetz */ 2161c5ecae3SMike Kravetz static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, 21790481622SDavid Gibson long delta) 21890481622SDavid Gibson { 2191c5ecae3SMike Kravetz long ret = delta; 220db71ef79SMike Kravetz unsigned long flags; 2211c5ecae3SMike Kravetz 22290481622SDavid Gibson if (!spool) 2231c5ecae3SMike Kravetz return delta; 22490481622SDavid Gibson 225db71ef79SMike Kravetz spin_lock_irqsave(&spool->lock, flags); 2261c5ecae3SMike Kravetz 2271c5ecae3SMike Kravetz if (spool->max_hpages != -1) /* maximum size accounting */ 22890481622SDavid Gibson spool->used_hpages -= delta; 2291c5ecae3SMike Kravetz 23009a95e29SMike Kravetz /* minimum size accounting */ 23109a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { 2321c5ecae3SMike Kravetz if (spool->rsv_hpages + delta <= spool->min_hpages) 2331c5ecae3SMike Kravetz ret = 0; 2341c5ecae3SMike Kravetz else 2351c5ecae3SMike Kravetz ret = spool->rsv_hpages + delta - spool->min_hpages; 2361c5ecae3SMike Kravetz 2371c5ecae3SMike Kravetz spool->rsv_hpages += delta; 2381c5ecae3SMike Kravetz if (spool->rsv_hpages > spool->min_hpages) 2391c5ecae3SMike Kravetz spool->rsv_hpages = spool->min_hpages; 2401c5ecae3SMike Kravetz } 2411c5ecae3SMike Kravetz 2421c5ecae3SMike Kravetz /* 2431c5ecae3SMike Kravetz * If hugetlbfs_put_super couldn't free spool due to an outstanding 2441c5ecae3SMike Kravetz * quota reference, free it now. 2451c5ecae3SMike Kravetz */ 246db71ef79SMike Kravetz unlock_or_release_subpool(spool, flags); 2471c5ecae3SMike Kravetz 2481c5ecae3SMike Kravetz return ret; 24990481622SDavid Gibson } 25090481622SDavid Gibson 25190481622SDavid Gibson static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 25290481622SDavid Gibson { 25390481622SDavid Gibson return HUGETLBFS_SB(inode->i_sb)->spool; 25490481622SDavid Gibson } 25590481622SDavid Gibson 25690481622SDavid Gibson static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 25790481622SDavid Gibson { 258496ad9aaSAl Viro return subpool_inode(file_inode(vma->vm_file)); 25990481622SDavid Gibson } 26090481622SDavid Gibson 261e700898fSMike Kravetz /* 262e700898fSMike Kravetz * hugetlb vma_lock helper routines 263e700898fSMike Kravetz */ 264e700898fSMike Kravetz void hugetlb_vma_lock_read(struct vm_area_struct *vma) 265e700898fSMike Kravetz { 266e700898fSMike Kravetz if (__vma_shareable_lock(vma)) { 267e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 268e700898fSMike Kravetz 269e700898fSMike Kravetz down_read(&vma_lock->rw_sema); 270e700898fSMike Kravetz } 271e700898fSMike Kravetz } 272e700898fSMike Kravetz 273e700898fSMike Kravetz void hugetlb_vma_unlock_read(struct vm_area_struct *vma) 274e700898fSMike Kravetz { 275e700898fSMike Kravetz if (__vma_shareable_lock(vma)) { 276e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 277e700898fSMike Kravetz 278e700898fSMike Kravetz up_read(&vma_lock->rw_sema); 279e700898fSMike Kravetz } 280e700898fSMike Kravetz } 281e700898fSMike Kravetz 282e700898fSMike Kravetz void hugetlb_vma_lock_write(struct vm_area_struct *vma) 283e700898fSMike Kravetz { 284e700898fSMike Kravetz if (__vma_shareable_lock(vma)) { 285e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 286e700898fSMike Kravetz 287e700898fSMike Kravetz down_write(&vma_lock->rw_sema); 288e700898fSMike Kravetz } 289e700898fSMike Kravetz } 290e700898fSMike Kravetz 291e700898fSMike Kravetz void hugetlb_vma_unlock_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 up_write(&vma_lock->rw_sema); 297e700898fSMike Kravetz } 298e700898fSMike Kravetz } 299e700898fSMike Kravetz 300e700898fSMike Kravetz int hugetlb_vma_trylock_write(struct vm_area_struct *vma) 301e700898fSMike Kravetz { 302e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 303e700898fSMike Kravetz 304e700898fSMike Kravetz if (!__vma_shareable_lock(vma)) 305e700898fSMike Kravetz return 1; 306e700898fSMike Kravetz 307e700898fSMike Kravetz return down_write_trylock(&vma_lock->rw_sema); 308e700898fSMike Kravetz } 309e700898fSMike Kravetz 310e700898fSMike Kravetz void hugetlb_vma_assert_locked(struct vm_area_struct *vma) 311e700898fSMike Kravetz { 312e700898fSMike Kravetz if (__vma_shareable_lock(vma)) { 313e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 314e700898fSMike Kravetz 315e700898fSMike Kravetz lockdep_assert_held(&vma_lock->rw_sema); 316e700898fSMike Kravetz } 317e700898fSMike Kravetz } 318e700898fSMike Kravetz 319e700898fSMike Kravetz void hugetlb_vma_lock_release(struct kref *kref) 320e700898fSMike Kravetz { 321e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = container_of(kref, 322e700898fSMike Kravetz struct hugetlb_vma_lock, refs); 323e700898fSMike Kravetz 324e700898fSMike Kravetz kfree(vma_lock); 325e700898fSMike Kravetz } 326e700898fSMike Kravetz 327e700898fSMike Kravetz static void __hugetlb_vma_unlock_write_put(struct hugetlb_vma_lock *vma_lock) 328e700898fSMike Kravetz { 329e700898fSMike Kravetz struct vm_area_struct *vma = vma_lock->vma; 330e700898fSMike Kravetz 331e700898fSMike Kravetz /* 332e700898fSMike Kravetz * vma_lock structure may or not be released as a result of put, 333e700898fSMike Kravetz * it certainly will no longer be attached to vma so clear pointer. 334e700898fSMike Kravetz * Semaphore synchronizes access to vma_lock->vma field. 335e700898fSMike Kravetz */ 336e700898fSMike Kravetz vma_lock->vma = NULL; 337e700898fSMike Kravetz vma->vm_private_data = NULL; 338e700898fSMike Kravetz up_write(&vma_lock->rw_sema); 339e700898fSMike Kravetz kref_put(&vma_lock->refs, hugetlb_vma_lock_release); 340e700898fSMike Kravetz } 341e700898fSMike Kravetz 342e700898fSMike Kravetz static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma) 343e700898fSMike Kravetz { 344e700898fSMike Kravetz if (__vma_shareable_lock(vma)) { 345e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 346e700898fSMike Kravetz 347e700898fSMike Kravetz __hugetlb_vma_unlock_write_put(vma_lock); 348e700898fSMike Kravetz } 349e700898fSMike Kravetz } 350e700898fSMike Kravetz 351e700898fSMike Kravetz static void hugetlb_vma_lock_free(struct vm_area_struct *vma) 352e700898fSMike Kravetz { 353e700898fSMike Kravetz /* 354e700898fSMike Kravetz * Only present in sharable vmas. 355e700898fSMike Kravetz */ 356e700898fSMike Kravetz if (!vma || !__vma_shareable_lock(vma)) 357e700898fSMike Kravetz return; 358e700898fSMike Kravetz 359e700898fSMike Kravetz if (vma->vm_private_data) { 360e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 361e700898fSMike Kravetz 362e700898fSMike Kravetz down_write(&vma_lock->rw_sema); 363e700898fSMike Kravetz __hugetlb_vma_unlock_write_put(vma_lock); 364e700898fSMike Kravetz } 365e700898fSMike Kravetz } 366e700898fSMike Kravetz 367e700898fSMike Kravetz static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma) 368e700898fSMike Kravetz { 369e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock; 370e700898fSMike Kravetz 371e700898fSMike Kravetz /* Only establish in (flags) sharable vmas */ 372e700898fSMike Kravetz if (!vma || !(vma->vm_flags & VM_MAYSHARE)) 373e700898fSMike Kravetz return; 374e700898fSMike Kravetz 375e700898fSMike Kravetz /* Should never get here with non-NULL vm_private_data */ 376e700898fSMike Kravetz if (vma->vm_private_data) 377e700898fSMike Kravetz return; 378e700898fSMike Kravetz 379e700898fSMike Kravetz vma_lock = kmalloc(sizeof(*vma_lock), GFP_KERNEL); 380e700898fSMike Kravetz if (!vma_lock) { 381e700898fSMike Kravetz /* 382e700898fSMike Kravetz * If we can not allocate structure, then vma can not 383e700898fSMike Kravetz * participate in pmd sharing. This is only a possible 384e700898fSMike Kravetz * performance enhancement and memory saving issue. 385e700898fSMike Kravetz * However, the lock is also used to synchronize page 386e700898fSMike Kravetz * faults with truncation. If the lock is not present, 387e700898fSMike Kravetz * unlikely races could leave pages in a file past i_size 388e700898fSMike Kravetz * until the file is removed. Warn in the unlikely case of 389e700898fSMike Kravetz * allocation failure. 390e700898fSMike Kravetz */ 391e700898fSMike Kravetz pr_warn_once("HugeTLB: unable to allocate vma specific lock\n"); 392e700898fSMike Kravetz return; 393e700898fSMike Kravetz } 394e700898fSMike Kravetz 395e700898fSMike Kravetz kref_init(&vma_lock->refs); 396e700898fSMike Kravetz init_rwsem(&vma_lock->rw_sema); 397e700898fSMike Kravetz vma_lock->vma = vma; 398e700898fSMike Kravetz vma->vm_private_data = vma_lock; 399e700898fSMike Kravetz } 400e700898fSMike Kravetz 4010db9d74eSMina Almasry /* Helper that removes a struct file_region from the resv_map cache and returns 4020db9d74eSMina Almasry * it for use. 4030db9d74eSMina Almasry */ 4040db9d74eSMina Almasry static struct file_region * 4050db9d74eSMina Almasry get_file_region_entry_from_cache(struct resv_map *resv, long from, long to) 4060db9d74eSMina Almasry { 4073259914fSXU pengfei struct file_region *nrg; 4080db9d74eSMina Almasry 4090db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count <= 0); 4100db9d74eSMina Almasry 4110db9d74eSMina Almasry resv->region_cache_count--; 4120db9d74eSMina Almasry nrg = list_first_entry(&resv->region_cache, struct file_region, link); 4130db9d74eSMina Almasry list_del(&nrg->link); 4140db9d74eSMina Almasry 4150db9d74eSMina Almasry nrg->from = from; 4160db9d74eSMina Almasry nrg->to = to; 4170db9d74eSMina Almasry 4180db9d74eSMina Almasry return nrg; 4190db9d74eSMina Almasry } 4200db9d74eSMina Almasry 421075a61d0SMina Almasry static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, 422075a61d0SMina Almasry struct file_region *rg) 423075a61d0SMina Almasry { 424075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 425075a61d0SMina Almasry nrg->reservation_counter = rg->reservation_counter; 426075a61d0SMina Almasry nrg->css = rg->css; 427075a61d0SMina Almasry if (rg->css) 428075a61d0SMina Almasry css_get(rg->css); 429075a61d0SMina Almasry #endif 430075a61d0SMina Almasry } 431075a61d0SMina Almasry 432075a61d0SMina Almasry /* Helper that records hugetlb_cgroup uncharge info. */ 433075a61d0SMina Almasry static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, 434075a61d0SMina Almasry struct hstate *h, 435075a61d0SMina Almasry struct resv_map *resv, 436075a61d0SMina Almasry struct file_region *nrg) 437075a61d0SMina Almasry { 438075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 439075a61d0SMina Almasry if (h_cg) { 440075a61d0SMina Almasry nrg->reservation_counter = 441075a61d0SMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 442075a61d0SMina Almasry nrg->css = &h_cg->css; 443d85aecf2SMiaohe Lin /* 444d85aecf2SMiaohe Lin * The caller will hold exactly one h_cg->css reference for the 445d85aecf2SMiaohe Lin * whole contiguous reservation region. But this area might be 446d85aecf2SMiaohe Lin * scattered when there are already some file_regions reside in 447d85aecf2SMiaohe Lin * it. As a result, many file_regions may share only one css 448d85aecf2SMiaohe Lin * reference. In order to ensure that one file_region must hold 449d85aecf2SMiaohe Lin * exactly one h_cg->css reference, we should do css_get for 450d85aecf2SMiaohe Lin * each file_region and leave the reference held by caller 451d85aecf2SMiaohe Lin * untouched. 452d85aecf2SMiaohe Lin */ 453d85aecf2SMiaohe Lin css_get(&h_cg->css); 454075a61d0SMina Almasry if (!resv->pages_per_hpage) 455075a61d0SMina Almasry resv->pages_per_hpage = pages_per_huge_page(h); 456075a61d0SMina Almasry /* pages_per_hpage should be the same for all entries in 457075a61d0SMina Almasry * a resv_map. 458075a61d0SMina Almasry */ 459075a61d0SMina Almasry VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); 460075a61d0SMina Almasry } else { 461075a61d0SMina Almasry nrg->reservation_counter = NULL; 462075a61d0SMina Almasry nrg->css = NULL; 463075a61d0SMina Almasry } 464075a61d0SMina Almasry #endif 465075a61d0SMina Almasry } 466075a61d0SMina Almasry 467d85aecf2SMiaohe Lin static void put_uncharge_info(struct file_region *rg) 468d85aecf2SMiaohe Lin { 469d85aecf2SMiaohe Lin #ifdef CONFIG_CGROUP_HUGETLB 470d85aecf2SMiaohe Lin if (rg->css) 471d85aecf2SMiaohe Lin css_put(rg->css); 472d85aecf2SMiaohe Lin #endif 473d85aecf2SMiaohe Lin } 474d85aecf2SMiaohe Lin 475a9b3f867SMina Almasry static bool has_same_uncharge_info(struct file_region *rg, 476a9b3f867SMina Almasry struct file_region *org) 477a9b3f867SMina Almasry { 478a9b3f867SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 4790739eb43SBaolin Wang return rg->reservation_counter == org->reservation_counter && 480a9b3f867SMina Almasry rg->css == org->css; 481a9b3f867SMina Almasry 482a9b3f867SMina Almasry #else 483a9b3f867SMina Almasry return true; 484a9b3f867SMina Almasry #endif 485a9b3f867SMina Almasry } 486a9b3f867SMina Almasry 487a9b3f867SMina Almasry static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) 488a9b3f867SMina Almasry { 4893259914fSXU pengfei struct file_region *nrg, *prg; 490a9b3f867SMina Almasry 491a9b3f867SMina Almasry prg = list_prev_entry(rg, link); 492a9b3f867SMina Almasry if (&prg->link != &resv->regions && prg->to == rg->from && 493a9b3f867SMina Almasry has_same_uncharge_info(prg, rg)) { 494a9b3f867SMina Almasry prg->to = rg->to; 495a9b3f867SMina Almasry 496a9b3f867SMina Almasry list_del(&rg->link); 497d85aecf2SMiaohe Lin put_uncharge_info(rg); 498a9b3f867SMina Almasry kfree(rg); 499a9b3f867SMina Almasry 5007db5e7b6SWei Yang rg = prg; 501a9b3f867SMina Almasry } 502a9b3f867SMina Almasry 503a9b3f867SMina Almasry nrg = list_next_entry(rg, link); 504a9b3f867SMina Almasry if (&nrg->link != &resv->regions && nrg->from == rg->to && 505a9b3f867SMina Almasry has_same_uncharge_info(nrg, rg)) { 506a9b3f867SMina Almasry nrg->from = rg->from; 507a9b3f867SMina Almasry 508a9b3f867SMina Almasry list_del(&rg->link); 509d85aecf2SMiaohe Lin put_uncharge_info(rg); 510a9b3f867SMina Almasry kfree(rg); 511a9b3f867SMina Almasry } 512a9b3f867SMina Almasry } 513a9b3f867SMina Almasry 5142103cf9cSPeter Xu static inline long 51584448c8eSJakob Koschel hugetlb_resv_map_add(struct resv_map *map, struct list_head *rg, long from, 5162103cf9cSPeter Xu long to, struct hstate *h, struct hugetlb_cgroup *cg, 5172103cf9cSPeter Xu long *regions_needed) 5182103cf9cSPeter Xu { 5192103cf9cSPeter Xu struct file_region *nrg; 5202103cf9cSPeter Xu 5212103cf9cSPeter Xu if (!regions_needed) { 5222103cf9cSPeter Xu nrg = get_file_region_entry_from_cache(map, from, to); 5232103cf9cSPeter Xu record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg); 52484448c8eSJakob Koschel list_add(&nrg->link, rg); 5252103cf9cSPeter Xu coalesce_file_region(map, nrg); 5262103cf9cSPeter Xu } else 5272103cf9cSPeter Xu *regions_needed += 1; 5282103cf9cSPeter Xu 5292103cf9cSPeter Xu return to - from; 5302103cf9cSPeter Xu } 5312103cf9cSPeter Xu 532972a3da3SWei Yang /* 533972a3da3SWei Yang * Must be called with resv->lock held. 534972a3da3SWei Yang * 535972a3da3SWei Yang * Calling this with regions_needed != NULL will count the number of pages 536972a3da3SWei Yang * to be added but will not modify the linked list. And regions_needed will 537972a3da3SWei Yang * indicate the number of file_regions needed in the cache to carry out to add 538972a3da3SWei Yang * the regions for this range. 539d75c6af9SMina Almasry */ 540d75c6af9SMina Almasry static long add_reservation_in_range(struct resv_map *resv, long f, long t, 541075a61d0SMina Almasry struct hugetlb_cgroup *h_cg, 542972a3da3SWei Yang struct hstate *h, long *regions_needed) 543d75c6af9SMina Almasry { 5440db9d74eSMina Almasry long add = 0; 545d75c6af9SMina Almasry struct list_head *head = &resv->regions; 5460db9d74eSMina Almasry long last_accounted_offset = f; 54784448c8eSJakob Koschel struct file_region *iter, *trg = NULL; 54884448c8eSJakob Koschel struct list_head *rg = NULL; 549d75c6af9SMina Almasry 5500db9d74eSMina Almasry if (regions_needed) 5510db9d74eSMina Almasry *regions_needed = 0; 552d75c6af9SMina Almasry 5530db9d74eSMina Almasry /* In this loop, we essentially handle an entry for the range 55484448c8eSJakob Koschel * [last_accounted_offset, iter->from), at every iteration, with some 5550db9d74eSMina Almasry * bounds checking. 5560db9d74eSMina Almasry */ 55784448c8eSJakob Koschel list_for_each_entry_safe(iter, trg, head, link) { 5580db9d74eSMina Almasry /* Skip irrelevant regions that start before our range. */ 55984448c8eSJakob Koschel if (iter->from < f) { 5600db9d74eSMina Almasry /* If this region ends after the last accounted offset, 5610db9d74eSMina Almasry * then we need to update last_accounted_offset. 5620db9d74eSMina Almasry */ 56384448c8eSJakob Koschel if (iter->to > last_accounted_offset) 56484448c8eSJakob Koschel last_accounted_offset = iter->to; 5650db9d74eSMina Almasry continue; 5660db9d74eSMina Almasry } 567d75c6af9SMina Almasry 5680db9d74eSMina Almasry /* When we find a region that starts beyond our range, we've 5690db9d74eSMina Almasry * finished. 5700db9d74eSMina Almasry */ 57184448c8eSJakob Koschel if (iter->from >= t) { 57284448c8eSJakob Koschel rg = iter->link.prev; 573d75c6af9SMina Almasry break; 57484448c8eSJakob Koschel } 575d75c6af9SMina Almasry 57684448c8eSJakob Koschel /* Add an entry for last_accounted_offset -> iter->from, and 5770db9d74eSMina Almasry * update last_accounted_offset. 578d75c6af9SMina Almasry */ 57984448c8eSJakob Koschel if (iter->from > last_accounted_offset) 58084448c8eSJakob Koschel add += hugetlb_resv_map_add(resv, iter->link.prev, 5812103cf9cSPeter Xu last_accounted_offset, 58284448c8eSJakob Koschel iter->from, h, h_cg, 5832103cf9cSPeter Xu regions_needed); 584d75c6af9SMina Almasry 58584448c8eSJakob Koschel last_accounted_offset = iter->to; 5860db9d74eSMina Almasry } 5870db9d74eSMina Almasry 5880db9d74eSMina Almasry /* Handle the case where our range extends beyond 5890db9d74eSMina Almasry * last_accounted_offset. 5900db9d74eSMina Almasry */ 59184448c8eSJakob Koschel if (!rg) 59284448c8eSJakob Koschel rg = head->prev; 5932103cf9cSPeter Xu if (last_accounted_offset < t) 5942103cf9cSPeter Xu add += hugetlb_resv_map_add(resv, rg, last_accounted_offset, 5952103cf9cSPeter Xu t, h, h_cg, regions_needed); 5960db9d74eSMina Almasry 5970db9d74eSMina Almasry return add; 5980db9d74eSMina Almasry } 5990db9d74eSMina Almasry 6000db9d74eSMina Almasry /* Must be called with resv->lock acquired. Will drop lock to allocate entries. 6010db9d74eSMina Almasry */ 6020db9d74eSMina Almasry static int allocate_file_region_entries(struct resv_map *resv, 6030db9d74eSMina Almasry int regions_needed) 6040db9d74eSMina Almasry __must_hold(&resv->lock) 6050db9d74eSMina Almasry { 60634665341SMiaohe Lin LIST_HEAD(allocated_regions); 6070db9d74eSMina Almasry int to_allocate = 0, i = 0; 6080db9d74eSMina Almasry struct file_region *trg = NULL, *rg = NULL; 6090db9d74eSMina Almasry 6100db9d74eSMina Almasry VM_BUG_ON(regions_needed < 0); 6110db9d74eSMina Almasry 6120db9d74eSMina Almasry /* 6130db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 6140db9d74eSMina Almasry * the number of in progress add operations plus regions_needed. 6150db9d74eSMina Almasry * 6160db9d74eSMina Almasry * This is a while loop because when we drop the lock, some other call 6170db9d74eSMina Almasry * to region_add or region_del may have consumed some region_entries, 6180db9d74eSMina Almasry * so we keep looping here until we finally have enough entries for 6190db9d74eSMina Almasry * (adds_in_progress + regions_needed). 6200db9d74eSMina Almasry */ 6210db9d74eSMina Almasry while (resv->region_cache_count < 6220db9d74eSMina Almasry (resv->adds_in_progress + regions_needed)) { 6230db9d74eSMina Almasry to_allocate = resv->adds_in_progress + regions_needed - 6240db9d74eSMina Almasry resv->region_cache_count; 6250db9d74eSMina Almasry 6260db9d74eSMina Almasry /* At this point, we should have enough entries in the cache 627f0953a1bSIngo Molnar * for all the existing adds_in_progress. We should only be 6280db9d74eSMina Almasry * needing to allocate for regions_needed. 6290db9d74eSMina Almasry */ 6300db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); 6310db9d74eSMina Almasry 6320db9d74eSMina Almasry spin_unlock(&resv->lock); 6330db9d74eSMina Almasry for (i = 0; i < to_allocate; i++) { 6340db9d74eSMina Almasry trg = kmalloc(sizeof(*trg), GFP_KERNEL); 6350db9d74eSMina Almasry if (!trg) 6360db9d74eSMina Almasry goto out_of_memory; 6370db9d74eSMina Almasry list_add(&trg->link, &allocated_regions); 6380db9d74eSMina Almasry } 6390db9d74eSMina Almasry 6400db9d74eSMina Almasry spin_lock(&resv->lock); 6410db9d74eSMina Almasry 642d3ec7b6eSWei Yang list_splice(&allocated_regions, &resv->region_cache); 643d3ec7b6eSWei Yang resv->region_cache_count += to_allocate; 6440db9d74eSMina Almasry } 6450db9d74eSMina Almasry 6460db9d74eSMina Almasry return 0; 6470db9d74eSMina Almasry 6480db9d74eSMina Almasry out_of_memory: 6490db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, &allocated_regions, link) { 650d75c6af9SMina Almasry list_del(&rg->link); 651d75c6af9SMina Almasry kfree(rg); 652d75c6af9SMina Almasry } 6530db9d74eSMina Almasry return -ENOMEM; 654d75c6af9SMina Almasry } 655d75c6af9SMina Almasry 6561dd308a7SMike Kravetz /* 6571dd308a7SMike Kravetz * Add the huge page range represented by [f, t) to the reserve 6580db9d74eSMina Almasry * map. Regions will be taken from the cache to fill in this range. 6590db9d74eSMina Almasry * Sufficient regions should exist in the cache due to the previous 6600db9d74eSMina Almasry * call to region_chg with the same range, but in some cases the cache will not 6610db9d74eSMina Almasry * have sufficient entries due to races with other code doing region_add or 6620db9d74eSMina Almasry * region_del. The extra needed entries will be allocated. 663cf3ad20bSMike Kravetz * 6640db9d74eSMina Almasry * regions_needed is the out value provided by a previous call to region_chg. 6650db9d74eSMina Almasry * 6660db9d74eSMina Almasry * Return the number of new huge pages added to the map. This number is greater 6670db9d74eSMina Almasry * than or equal to zero. If file_region entries needed to be allocated for 6687c8de358SEthon Paul * this operation and we were not able to allocate, it returns -ENOMEM. 6690db9d74eSMina Almasry * region_add of regions of length 1 never allocate file_regions and cannot 6700db9d74eSMina Almasry * fail; region_chg will always allocate at least 1 entry and a region_add for 6710db9d74eSMina Almasry * 1 page will only require at most 1 entry. 6721dd308a7SMike Kravetz */ 6730db9d74eSMina Almasry static long region_add(struct resv_map *resv, long f, long t, 674075a61d0SMina Almasry long in_regions_needed, struct hstate *h, 675075a61d0SMina Almasry struct hugetlb_cgroup *h_cg) 67696822904SAndy Whitcroft { 6770db9d74eSMina Almasry long add = 0, actual_regions_needed = 0; 67896822904SAndy Whitcroft 6797b24d861SDavidlohr Bueso spin_lock(&resv->lock); 6800db9d74eSMina Almasry retry: 6810db9d74eSMina Almasry 6820db9d74eSMina Almasry /* Count how many regions are actually needed to execute this add. */ 683972a3da3SWei Yang add_reservation_in_range(resv, f, t, NULL, NULL, 684972a3da3SWei Yang &actual_regions_needed); 68596822904SAndy Whitcroft 6865e911373SMike Kravetz /* 6870db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 6880db9d74eSMina Almasry * this add operation. Note that actual_regions_needed may be greater 6890db9d74eSMina Almasry * than in_regions_needed, as the resv_map may have been modified since 6900db9d74eSMina Almasry * the region_chg call. In this case, we need to make sure that we 6910db9d74eSMina Almasry * allocate extra entries, such that we have enough for all the 6920db9d74eSMina Almasry * existing adds_in_progress, plus the excess needed for this 6930db9d74eSMina Almasry * operation. 6945e911373SMike Kravetz */ 6950db9d74eSMina Almasry if (actual_regions_needed > in_regions_needed && 6960db9d74eSMina Almasry resv->region_cache_count < 6970db9d74eSMina Almasry resv->adds_in_progress + 6980db9d74eSMina Almasry (actual_regions_needed - in_regions_needed)) { 6990db9d74eSMina Almasry /* region_add operation of range 1 should never need to 7000db9d74eSMina Almasry * allocate file_region entries. 7010db9d74eSMina Almasry */ 7020db9d74eSMina Almasry VM_BUG_ON(t - f <= 1); 7035e911373SMike Kravetz 7040db9d74eSMina Almasry if (allocate_file_region_entries( 7050db9d74eSMina Almasry resv, actual_regions_needed - in_regions_needed)) { 7060db9d74eSMina Almasry return -ENOMEM; 7075e911373SMike Kravetz } 7085e911373SMike Kravetz 7090db9d74eSMina Almasry goto retry; 7100db9d74eSMina Almasry } 711cf3ad20bSMike Kravetz 712972a3da3SWei Yang add = add_reservation_in_range(resv, f, t, h_cg, h, NULL); 7130db9d74eSMina Almasry 7140db9d74eSMina Almasry resv->adds_in_progress -= in_regions_needed; 7150db9d74eSMina Almasry 7167b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 717cf3ad20bSMike Kravetz return add; 71896822904SAndy Whitcroft } 71996822904SAndy Whitcroft 7201dd308a7SMike Kravetz /* 7211dd308a7SMike Kravetz * Examine the existing reserve map and determine how many 7221dd308a7SMike Kravetz * huge pages in the specified range [f, t) are NOT currently 7231dd308a7SMike Kravetz * represented. This routine is called before a subsequent 7241dd308a7SMike Kravetz * call to region_add that will actually modify the reserve 7251dd308a7SMike Kravetz * map to add the specified range [f, t). region_chg does 7261dd308a7SMike Kravetz * not change the number of huge pages represented by the 7270db9d74eSMina Almasry * map. A number of new file_region structures is added to the cache as a 7280db9d74eSMina Almasry * placeholder, for the subsequent region_add call to use. At least 1 7290db9d74eSMina Almasry * file_region structure is added. 7300db9d74eSMina Almasry * 7310db9d74eSMina Almasry * out_regions_needed is the number of regions added to the 7320db9d74eSMina Almasry * resv->adds_in_progress. This value needs to be provided to a follow up call 7330db9d74eSMina Almasry * to region_add or region_abort for proper accounting. 7345e911373SMike Kravetz * 7355e911373SMike Kravetz * Returns the number of huge pages that need to be added to the existing 7365e911373SMike Kravetz * reservation map for the range [f, t). This number is greater or equal to 7375e911373SMike Kravetz * zero. -ENOMEM is returned if a new file_region structure or cache entry 7385e911373SMike Kravetz * is needed and can not be allocated. 7391dd308a7SMike Kravetz */ 7400db9d74eSMina Almasry static long region_chg(struct resv_map *resv, long f, long t, 7410db9d74eSMina Almasry long *out_regions_needed) 74296822904SAndy Whitcroft { 74396822904SAndy Whitcroft long chg = 0; 74496822904SAndy Whitcroft 7457b24d861SDavidlohr Bueso spin_lock(&resv->lock); 7465e911373SMike Kravetz 747972a3da3SWei Yang /* Count how many hugepages in this range are NOT represented. */ 748075a61d0SMina Almasry chg = add_reservation_in_range(resv, f, t, NULL, NULL, 749972a3da3SWei Yang out_regions_needed); 7505e911373SMike Kravetz 7510db9d74eSMina Almasry if (*out_regions_needed == 0) 7520db9d74eSMina Almasry *out_regions_needed = 1; 7535e911373SMike Kravetz 7540db9d74eSMina Almasry if (allocate_file_region_entries(resv, *out_regions_needed)) 7555e911373SMike Kravetz return -ENOMEM; 7565e911373SMike Kravetz 7570db9d74eSMina Almasry resv->adds_in_progress += *out_regions_needed; 75896822904SAndy Whitcroft 7597b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 76096822904SAndy Whitcroft return chg; 76196822904SAndy Whitcroft } 76296822904SAndy Whitcroft 7631dd308a7SMike Kravetz /* 7645e911373SMike Kravetz * Abort the in progress add operation. The adds_in_progress field 7655e911373SMike Kravetz * of the resv_map keeps track of the operations in progress between 7665e911373SMike Kravetz * calls to region_chg and region_add. Operations are sometimes 7675e911373SMike Kravetz * aborted after the call to region_chg. In such cases, region_abort 7680db9d74eSMina Almasry * is called to decrement the adds_in_progress counter. regions_needed 7690db9d74eSMina Almasry * is the value returned by the region_chg call, it is used to decrement 7700db9d74eSMina Almasry * the adds_in_progress counter. 7715e911373SMike Kravetz * 7725e911373SMike Kravetz * NOTE: The range arguments [f, t) are not needed or used in this 7735e911373SMike Kravetz * routine. They are kept to make reading the calling code easier as 7745e911373SMike Kravetz * arguments will match the associated region_chg call. 7755e911373SMike Kravetz */ 7760db9d74eSMina Almasry static void region_abort(struct resv_map *resv, long f, long t, 7770db9d74eSMina Almasry long regions_needed) 7785e911373SMike Kravetz { 7795e911373SMike Kravetz spin_lock(&resv->lock); 7805e911373SMike Kravetz VM_BUG_ON(!resv->region_cache_count); 7810db9d74eSMina Almasry resv->adds_in_progress -= regions_needed; 7825e911373SMike Kravetz spin_unlock(&resv->lock); 7835e911373SMike Kravetz } 7845e911373SMike Kravetz 7855e911373SMike Kravetz /* 786feba16e2SMike Kravetz * Delete the specified range [f, t) from the reserve map. If the 787feba16e2SMike Kravetz * t parameter is LONG_MAX, this indicates that ALL regions after f 788feba16e2SMike Kravetz * should be deleted. Locate the regions which intersect [f, t) 789feba16e2SMike Kravetz * and either trim, delete or split the existing regions. 790feba16e2SMike Kravetz * 791feba16e2SMike Kravetz * Returns the number of huge pages deleted from the reserve map. 792feba16e2SMike Kravetz * In the normal case, the return value is zero or more. In the 793feba16e2SMike Kravetz * case where a region must be split, a new region descriptor must 794feba16e2SMike Kravetz * be allocated. If the allocation fails, -ENOMEM will be returned. 795feba16e2SMike Kravetz * NOTE: If the parameter t == LONG_MAX, then we will never split 796feba16e2SMike Kravetz * a region and possibly return -ENOMEM. Callers specifying 797feba16e2SMike Kravetz * t == LONG_MAX do not need to check for -ENOMEM error. 7981dd308a7SMike Kravetz */ 799feba16e2SMike Kravetz static long region_del(struct resv_map *resv, long f, long t) 80096822904SAndy Whitcroft { 8011406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 80296822904SAndy Whitcroft struct file_region *rg, *trg; 803feba16e2SMike Kravetz struct file_region *nrg = NULL; 804feba16e2SMike Kravetz long del = 0; 80596822904SAndy Whitcroft 806feba16e2SMike Kravetz retry: 8077b24d861SDavidlohr Bueso spin_lock(&resv->lock); 808feba16e2SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 809dbe409e4SMike Kravetz /* 810dbe409e4SMike Kravetz * Skip regions before the range to be deleted. file_region 811dbe409e4SMike Kravetz * ranges are normally of the form [from, to). However, there 812dbe409e4SMike Kravetz * may be a "placeholder" entry in the map which is of the form 813dbe409e4SMike Kravetz * (from, to) with from == to. Check for placeholder entries 814dbe409e4SMike Kravetz * at the beginning of the range to be deleted. 815dbe409e4SMike Kravetz */ 816dbe409e4SMike Kravetz if (rg->to <= f && (rg->to != rg->from || rg->to != f)) 817feba16e2SMike Kravetz continue; 818dbe409e4SMike Kravetz 819feba16e2SMike Kravetz if (rg->from >= t) 82096822904SAndy Whitcroft break; 82196822904SAndy Whitcroft 822feba16e2SMike Kravetz if (f > rg->from && t < rg->to) { /* Must split region */ 823feba16e2SMike Kravetz /* 824feba16e2SMike Kravetz * Check for an entry in the cache before dropping 825feba16e2SMike Kravetz * lock and attempting allocation. 826feba16e2SMike Kravetz */ 827feba16e2SMike Kravetz if (!nrg && 828feba16e2SMike Kravetz resv->region_cache_count > resv->adds_in_progress) { 829feba16e2SMike Kravetz nrg = list_first_entry(&resv->region_cache, 830feba16e2SMike Kravetz struct file_region, 831feba16e2SMike Kravetz link); 832feba16e2SMike Kravetz list_del(&nrg->link); 833feba16e2SMike Kravetz resv->region_cache_count--; 83496822904SAndy Whitcroft } 83596822904SAndy Whitcroft 836feba16e2SMike Kravetz if (!nrg) { 837feba16e2SMike Kravetz spin_unlock(&resv->lock); 838feba16e2SMike Kravetz nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 839feba16e2SMike Kravetz if (!nrg) 840feba16e2SMike Kravetz return -ENOMEM; 841feba16e2SMike Kravetz goto retry; 842feba16e2SMike Kravetz } 843feba16e2SMike Kravetz 844feba16e2SMike Kravetz del += t - f; 84579aa925bSMike Kravetz hugetlb_cgroup_uncharge_file_region( 846d85aecf2SMiaohe Lin resv, rg, t - f, false); 847feba16e2SMike Kravetz 848feba16e2SMike Kravetz /* New entry for end of split region */ 849feba16e2SMike Kravetz nrg->from = t; 850feba16e2SMike Kravetz nrg->to = rg->to; 851075a61d0SMina Almasry 852075a61d0SMina Almasry copy_hugetlb_cgroup_uncharge_info(nrg, rg); 853075a61d0SMina Almasry 854feba16e2SMike Kravetz INIT_LIST_HEAD(&nrg->link); 855feba16e2SMike Kravetz 856feba16e2SMike Kravetz /* Original entry is trimmed */ 857feba16e2SMike Kravetz rg->to = f; 858feba16e2SMike Kravetz 859feba16e2SMike Kravetz list_add(&nrg->link, &rg->link); 860feba16e2SMike Kravetz nrg = NULL; 86196822904SAndy Whitcroft break; 862feba16e2SMike Kravetz } 863feba16e2SMike Kravetz 864feba16e2SMike Kravetz if (f <= rg->from && t >= rg->to) { /* Remove entire region */ 865feba16e2SMike Kravetz del += rg->to - rg->from; 866075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 867d85aecf2SMiaohe Lin rg->to - rg->from, true); 86896822904SAndy Whitcroft list_del(&rg->link); 86996822904SAndy Whitcroft kfree(rg); 870feba16e2SMike Kravetz continue; 87196822904SAndy Whitcroft } 8727b24d861SDavidlohr Bueso 873feba16e2SMike Kravetz if (f <= rg->from) { /* Trim beginning of region */ 874075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 875d85aecf2SMiaohe Lin t - rg->from, false); 876075a61d0SMina Almasry 87779aa925bSMike Kravetz del += t - rg->from; 87879aa925bSMike Kravetz rg->from = t; 87979aa925bSMike Kravetz } else { /* Trim end of region */ 880075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 881d85aecf2SMiaohe Lin rg->to - f, false); 88279aa925bSMike Kravetz 88379aa925bSMike Kravetz del += rg->to - f; 88479aa925bSMike Kravetz rg->to = f; 885feba16e2SMike Kravetz } 886feba16e2SMike Kravetz } 887feba16e2SMike Kravetz 8887b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 889feba16e2SMike Kravetz kfree(nrg); 890feba16e2SMike Kravetz return del; 89196822904SAndy Whitcroft } 89296822904SAndy Whitcroft 8931dd308a7SMike Kravetz /* 894b5cec28dSMike Kravetz * A rare out of memory error was encountered which prevented removal of 895b5cec28dSMike Kravetz * the reserve map region for a page. The huge page itself was free'ed 896b5cec28dSMike Kravetz * and removed from the page cache. This routine will adjust the subpool 897b5cec28dSMike Kravetz * usage count, and the global reserve count if needed. By incrementing 898b5cec28dSMike Kravetz * these counts, the reserve map entry which could not be deleted will 899b5cec28dSMike Kravetz * appear as a "reserved" entry instead of simply dangling with incorrect 900b5cec28dSMike Kravetz * counts. 901b5cec28dSMike Kravetz */ 90272e2936cSzhong jiang void hugetlb_fix_reserve_counts(struct inode *inode) 903b5cec28dSMike Kravetz { 904b5cec28dSMike Kravetz struct hugepage_subpool *spool = subpool_inode(inode); 905b5cec28dSMike Kravetz long rsv_adjust; 906da56388cSMiaohe Lin bool reserved = false; 907b5cec28dSMike Kravetz 908b5cec28dSMike Kravetz rsv_adjust = hugepage_subpool_get_pages(spool, 1); 909da56388cSMiaohe Lin if (rsv_adjust > 0) { 910b5cec28dSMike Kravetz struct hstate *h = hstate_inode(inode); 911b5cec28dSMike Kravetz 912da56388cSMiaohe Lin if (!hugetlb_acct_memory(h, 1)) 913da56388cSMiaohe Lin reserved = true; 914da56388cSMiaohe Lin } else if (!rsv_adjust) { 915da56388cSMiaohe Lin reserved = true; 916b5cec28dSMike Kravetz } 917da56388cSMiaohe Lin 918da56388cSMiaohe Lin if (!reserved) 919da56388cSMiaohe Lin pr_warn("hugetlb: Huge Page Reserved count may go negative.\n"); 920b5cec28dSMike Kravetz } 921b5cec28dSMike Kravetz 922b5cec28dSMike Kravetz /* 9231dd308a7SMike Kravetz * Count and return the number of huge pages in the reserve map 9241dd308a7SMike Kravetz * that intersect with the range [f, t). 9251dd308a7SMike Kravetz */ 9261406ec9bSJoonsoo Kim static long region_count(struct resv_map *resv, long f, long t) 92784afd99bSAndy Whitcroft { 9281406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 92984afd99bSAndy Whitcroft struct file_region *rg; 93084afd99bSAndy Whitcroft long chg = 0; 93184afd99bSAndy Whitcroft 9327b24d861SDavidlohr Bueso spin_lock(&resv->lock); 93384afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 93484afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 935f2135a4aSWang Sheng-Hui long seg_from; 936f2135a4aSWang Sheng-Hui long seg_to; 93784afd99bSAndy Whitcroft 93884afd99bSAndy Whitcroft if (rg->to <= f) 93984afd99bSAndy Whitcroft continue; 94084afd99bSAndy Whitcroft if (rg->from >= t) 94184afd99bSAndy Whitcroft break; 94284afd99bSAndy Whitcroft 94384afd99bSAndy Whitcroft seg_from = max(rg->from, f); 94484afd99bSAndy Whitcroft seg_to = min(rg->to, t); 94584afd99bSAndy Whitcroft 94684afd99bSAndy Whitcroft chg += seg_to - seg_from; 94784afd99bSAndy Whitcroft } 9487b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 94984afd99bSAndy Whitcroft 95084afd99bSAndy Whitcroft return chg; 95184afd99bSAndy Whitcroft } 95284afd99bSAndy Whitcroft 95396822904SAndy Whitcroft /* 954e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 955e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 956e7c4b0bfSAndy Whitcroft */ 957a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 958a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 959e7c4b0bfSAndy Whitcroft { 960a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 961a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 962e7c4b0bfSAndy Whitcroft } 963e7c4b0bfSAndy Whitcroft 9640fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 9650fe6e20bSNaoya Horiguchi unsigned long address) 9660fe6e20bSNaoya Horiguchi { 9670fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 9680fe6e20bSNaoya Horiguchi } 969dee41079SDan Williams EXPORT_SYMBOL_GPL(linear_hugepage_index); 9700fe6e20bSNaoya Horiguchi 9718cfd014eSMatthew Wilcox (Oracle) /** 9728cfd014eSMatthew Wilcox (Oracle) * vma_kernel_pagesize - Page size granularity for this VMA. 9738cfd014eSMatthew Wilcox (Oracle) * @vma: The user mapping. 9748cfd014eSMatthew Wilcox (Oracle) * 9758cfd014eSMatthew Wilcox (Oracle) * Folios in this VMA will be aligned to, and at least the size of the 9768cfd014eSMatthew Wilcox (Oracle) * number of bytes returned by this function. 9778cfd014eSMatthew Wilcox (Oracle) * 9788cfd014eSMatthew Wilcox (Oracle) * Return: The default size of the folios allocated when backing a VMA. 97908fba699SMel Gorman */ 98008fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 98108fba699SMel Gorman { 98205ea8860SDan Williams if (vma->vm_ops && vma->vm_ops->pagesize) 98305ea8860SDan Williams return vma->vm_ops->pagesize(vma); 98408fba699SMel Gorman return PAGE_SIZE; 98508fba699SMel Gorman } 986f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 98708fba699SMel Gorman 98808fba699SMel Gorman /* 9893340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 9903340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 99109135cc5SDan Williams * architectures where it differs, an architecture-specific 'strong' 99209135cc5SDan Williams * version of this symbol is required. 9933340289dSMel Gorman */ 99409135cc5SDan Williams __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 9953340289dSMel Gorman { 9963340289dSMel Gorman return vma_kernel_pagesize(vma); 9973340289dSMel Gorman } 9983340289dSMel Gorman 9993340289dSMel Gorman /* 100084afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 100184afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 100284afd99bSAndy Whitcroft * alignment. 100384afd99bSAndy Whitcroft */ 100484afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 100584afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 100604f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 100784afd99bSAndy Whitcroft 1008a1e78772SMel Gorman /* 1009a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 1010a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 1011a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 1012a1e78772SMel Gorman * 10138d9bfb26SMike Kravetz * With the exception of hugetlb_dup_vma_private() which is called at fork(), 1014a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 1015a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 1016a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 101784afd99bSAndy Whitcroft * 101884afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 101984afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 102084afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 102184afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 102284afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 102384afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 102484afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 102584afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 1026a1e78772SMel Gorman */ 1027e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 1028e7c4b0bfSAndy Whitcroft { 1029e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 1030e7c4b0bfSAndy Whitcroft } 1031e7c4b0bfSAndy Whitcroft 1032e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 1033e7c4b0bfSAndy Whitcroft unsigned long value) 1034e7c4b0bfSAndy Whitcroft { 1035e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 1036e7c4b0bfSAndy Whitcroft } 1037e7c4b0bfSAndy Whitcroft 1038e9fe92aeSMina Almasry static void 1039e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, 1040e9fe92aeSMina Almasry struct hugetlb_cgroup *h_cg, 1041e9fe92aeSMina Almasry struct hstate *h) 1042e9fe92aeSMina Almasry { 1043e9fe92aeSMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 1044e9fe92aeSMina Almasry if (!h_cg || !h) { 1045e9fe92aeSMina Almasry resv_map->reservation_counter = NULL; 1046e9fe92aeSMina Almasry resv_map->pages_per_hpage = 0; 1047e9fe92aeSMina Almasry resv_map->css = NULL; 1048e9fe92aeSMina Almasry } else { 1049e9fe92aeSMina Almasry resv_map->reservation_counter = 1050e9fe92aeSMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 1051e9fe92aeSMina Almasry resv_map->pages_per_hpage = pages_per_huge_page(h); 1052e9fe92aeSMina Almasry resv_map->css = &h_cg->css; 1053e9fe92aeSMina Almasry } 1054e9fe92aeSMina Almasry #endif 1055e9fe92aeSMina Almasry } 1056e9fe92aeSMina Almasry 10579119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 105884afd99bSAndy Whitcroft { 105984afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 10605e911373SMike Kravetz struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 10615e911373SMike Kravetz 10625e911373SMike Kravetz if (!resv_map || !rg) { 10635e911373SMike Kravetz kfree(resv_map); 10645e911373SMike Kravetz kfree(rg); 106584afd99bSAndy Whitcroft return NULL; 10665e911373SMike Kravetz } 106784afd99bSAndy Whitcroft 106884afd99bSAndy Whitcroft kref_init(&resv_map->refs); 10697b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 107084afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 107184afd99bSAndy Whitcroft 10725e911373SMike Kravetz resv_map->adds_in_progress = 0; 1073e9fe92aeSMina Almasry /* 1074e9fe92aeSMina Almasry * Initialize these to 0. On shared mappings, 0's here indicate these 1075e9fe92aeSMina Almasry * fields don't do cgroup accounting. On private mappings, these will be 1076e9fe92aeSMina Almasry * re-initialized to the proper values, to indicate that hugetlb cgroup 1077e9fe92aeSMina Almasry * reservations are to be un-charged from here. 1078e9fe92aeSMina Almasry */ 1079e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); 10805e911373SMike Kravetz 10815e911373SMike Kravetz INIT_LIST_HEAD(&resv_map->region_cache); 10825e911373SMike Kravetz list_add(&rg->link, &resv_map->region_cache); 10835e911373SMike Kravetz resv_map->region_cache_count = 1; 10845e911373SMike Kravetz 108584afd99bSAndy Whitcroft return resv_map; 108684afd99bSAndy Whitcroft } 108784afd99bSAndy Whitcroft 10889119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 108984afd99bSAndy Whitcroft { 109084afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 10915e911373SMike Kravetz struct list_head *head = &resv_map->region_cache; 10925e911373SMike Kravetz struct file_region *rg, *trg; 109384afd99bSAndy Whitcroft 109484afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 1095feba16e2SMike Kravetz region_del(resv_map, 0, LONG_MAX); 10965e911373SMike Kravetz 10975e911373SMike Kravetz /* ... and any entries left in the cache */ 10985e911373SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 10995e911373SMike Kravetz list_del(&rg->link); 11005e911373SMike Kravetz kfree(rg); 11015e911373SMike Kravetz } 11025e911373SMike Kravetz 11035e911373SMike Kravetz VM_BUG_ON(resv_map->adds_in_progress); 11045e911373SMike Kravetz 110584afd99bSAndy Whitcroft kfree(resv_map); 110684afd99bSAndy Whitcroft } 110784afd99bSAndy Whitcroft 11084e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 11094e35f483SJoonsoo Kim { 1110f27a5136SMike Kravetz /* 1111f27a5136SMike Kravetz * At inode evict time, i_mapping may not point to the original 1112f27a5136SMike Kravetz * address space within the inode. This original address space 1113f27a5136SMike Kravetz * contains the pointer to the resv_map. So, always use the 1114f27a5136SMike Kravetz * address space embedded within the inode. 1115f27a5136SMike Kravetz * The VERY common case is inode->mapping == &inode->i_data but, 1116f27a5136SMike Kravetz * this may not be true for device special inodes. 1117f27a5136SMike Kravetz */ 1118f27a5136SMike Kravetz return (struct resv_map *)(&inode->i_data)->private_data; 11194e35f483SJoonsoo Kim } 11204e35f483SJoonsoo Kim 112184afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 1122a1e78772SMel Gorman { 112381d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 11244e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 11254e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 11264e35f483SJoonsoo Kim struct inode *inode = mapping->host; 11274e35f483SJoonsoo Kim 11284e35f483SJoonsoo Kim return inode_resv_map(inode); 11294e35f483SJoonsoo Kim 11304e35f483SJoonsoo Kim } else { 113184afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 113284afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 11334e35f483SJoonsoo Kim } 1134a1e78772SMel Gorman } 1135a1e78772SMel Gorman 113684afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 1137a1e78772SMel Gorman { 113881d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 113981d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 1140a1e78772SMel Gorman 1141*92fe9dcbSRik van Riel set_vma_private_data(vma, (unsigned long)map); 114204f2cbe3SMel Gorman } 114304f2cbe3SMel Gorman 114404f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 114504f2cbe3SMel Gorman { 114681d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 114781d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 1148e7c4b0bfSAndy Whitcroft 1149e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 115004f2cbe3SMel Gorman } 115104f2cbe3SMel Gorman 115204f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 115304f2cbe3SMel Gorman { 115481d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 1155e7c4b0bfSAndy Whitcroft 1156e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 1157a1e78772SMel Gorman } 1158a1e78772SMel Gorman 11598d9bfb26SMike Kravetz void hugetlb_dup_vma_private(struct vm_area_struct *vma) 1160a1e78772SMel Gorman { 116181d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 11628d9bfb26SMike Kravetz /* 11638d9bfb26SMike Kravetz * Clear vm_private_data 1164612b8a31SMike Kravetz * - For shared mappings this is a per-vma semaphore that may be 1165612b8a31SMike Kravetz * allocated in a subsequent call to hugetlb_vm_op_open. 1166612b8a31SMike Kravetz * Before clearing, make sure pointer is not associated with vma 1167612b8a31SMike Kravetz * as this will leak the structure. This is the case when called 1168612b8a31SMike Kravetz * via clear_vma_resv_huge_pages() and hugetlb_vm_op_open has already 1169612b8a31SMike Kravetz * been called to allocate a new structure. 11708d9bfb26SMike Kravetz * - For MAP_PRIVATE mappings, this is the reserve map which does 11718d9bfb26SMike Kravetz * not apply to children. Faults generated by the children are 11728d9bfb26SMike Kravetz * not guaranteed to succeed, even if read-only. 11738d9bfb26SMike Kravetz */ 1174612b8a31SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 1175612b8a31SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 1176612b8a31SMike Kravetz 1177612b8a31SMike Kravetz if (vma_lock && vma_lock->vma != vma) 1178612b8a31SMike Kravetz vma->vm_private_data = NULL; 1179612b8a31SMike Kravetz } else 1180612b8a31SMike Kravetz vma->vm_private_data = NULL; 1181a1e78772SMel Gorman } 1182a1e78772SMel Gorman 1183550a7d60SMina Almasry /* 1184550a7d60SMina Almasry * Reset and decrement one ref on hugepage private reservation. 11858651a137SLorenzo Stoakes * Called with mm->mmap_lock writer semaphore held. 1186550a7d60SMina Almasry * This function should be only used by move_vma() and operate on 1187550a7d60SMina Almasry * same sized vma. It should never come here with last ref on the 1188550a7d60SMina Almasry * reservation. 1189550a7d60SMina Almasry */ 1190550a7d60SMina Almasry void clear_vma_resv_huge_pages(struct vm_area_struct *vma) 1191550a7d60SMina Almasry { 1192550a7d60SMina Almasry /* 1193550a7d60SMina Almasry * Clear the old hugetlb private page reservation. 1194550a7d60SMina Almasry * It has already been transferred to new_vma. 1195550a7d60SMina Almasry * 1196550a7d60SMina Almasry * During a mremap() operation of a hugetlb vma we call move_vma() 1197550a7d60SMina Almasry * which copies vma into new_vma and unmaps vma. After the copy 1198550a7d60SMina Almasry * operation both new_vma and vma share a reference to the resv_map 1199550a7d60SMina Almasry * struct, and at that point vma is about to be unmapped. We don't 1200550a7d60SMina Almasry * want to return the reservation to the pool at unmap of vma because 1201550a7d60SMina Almasry * the reservation still lives on in new_vma, so simply decrement the 1202550a7d60SMina Almasry * ref here and remove the resv_map reference from this vma. 1203550a7d60SMina Almasry */ 1204550a7d60SMina Almasry struct resv_map *reservations = vma_resv_map(vma); 1205550a7d60SMina Almasry 1206afe041c2SBui Quang Minh if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 1207afe041c2SBui Quang Minh resv_map_put_hugetlb_cgroup_uncharge_info(reservations); 1208550a7d60SMina Almasry kref_put(&reservations->refs, resv_map_release); 1209afe041c2SBui Quang Minh } 1210550a7d60SMina Almasry 12118d9bfb26SMike Kravetz hugetlb_dup_vma_private(vma); 1212550a7d60SMina Almasry } 1213550a7d60SMina Almasry 1214a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 1215559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 1216a1e78772SMel Gorman { 1217af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 1218af0ed73eSJoonsoo Kim /* 1219af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 1220af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 1221af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 1222af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 1223af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 1224af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 1225af0ed73eSJoonsoo Kim * properly, so add work-around here. 1226af0ed73eSJoonsoo Kim */ 1227af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 1228559ec2f8SNicholas Krause return true; 1229af0ed73eSJoonsoo Kim else 1230559ec2f8SNicholas Krause return false; 1231af0ed73eSJoonsoo Kim } 1232a63884e9SJoonsoo Kim 1233a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 12341fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 12351fb1b0e9SMike Kravetz /* 12361fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 12371fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 12381fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 12397c8de358SEthon Paul * fallocate. In this case, there really are no reserves to 12401fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 12411fb1b0e9SMike Kravetz */ 12421fb1b0e9SMike Kravetz if (chg) 12431fb1b0e9SMike Kravetz return false; 12441fb1b0e9SMike Kravetz else 1245559ec2f8SNicholas Krause return true; 12461fb1b0e9SMike Kravetz } 1247a63884e9SJoonsoo Kim 1248a63884e9SJoonsoo Kim /* 1249a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 1250a63884e9SJoonsoo Kim * private mappings. 1251a63884e9SJoonsoo Kim */ 125267961f9dSMike Kravetz if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 125367961f9dSMike Kravetz /* 125467961f9dSMike Kravetz * Like the shared case above, a hole punch or truncate 125567961f9dSMike Kravetz * could have been performed on the private mapping. 125667961f9dSMike Kravetz * Examine the value of chg to determine if reserves 125767961f9dSMike Kravetz * actually exist or were previously consumed. 125867961f9dSMike Kravetz * Very Subtle - The value of chg comes from a previous 125967961f9dSMike Kravetz * call to vma_needs_reserves(). The reserve map for 126067961f9dSMike Kravetz * private mappings has different (opposite) semantics 126167961f9dSMike Kravetz * than that of shared mappings. vma_needs_reserves() 126267961f9dSMike Kravetz * has already taken this difference in semantics into 126367961f9dSMike Kravetz * account. Therefore, the meaning of chg is the same 126467961f9dSMike Kravetz * as in the shared case above. Code could easily be 126567961f9dSMike Kravetz * combined, but keeping it separate draws attention to 126667961f9dSMike Kravetz * subtle differences. 126767961f9dSMike Kravetz */ 126867961f9dSMike Kravetz if (chg) 126967961f9dSMike Kravetz return false; 127067961f9dSMike Kravetz else 1271559ec2f8SNicholas Krause return true; 127267961f9dSMike Kravetz } 1273a63884e9SJoonsoo Kim 1274559ec2f8SNicholas Krause return false; 1275a1e78772SMel Gorman } 1276a1e78772SMel Gorman 1277240d67a8SSidhartha Kumar static void enqueue_hugetlb_folio(struct hstate *h, struct folio *folio) 12781da177e4SLinus Torvalds { 1279240d67a8SSidhartha Kumar int nid = folio_nid(folio); 12809487ca60SMike Kravetz 12819487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1282240d67a8SSidhartha Kumar VM_BUG_ON_FOLIO(folio_ref_count(folio), folio); 1283b65a4edaSMike Kravetz 1284240d67a8SSidhartha Kumar list_move(&folio->lru, &h->hugepage_freelists[nid]); 1285a5516438SAndi Kleen h->free_huge_pages++; 1286a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 1287240d67a8SSidhartha Kumar folio_set_hugetlb_freed(folio); 12881da177e4SLinus Torvalds } 12891da177e4SLinus Torvalds 1290a36f1e90SSidhartha Kumar static struct folio *dequeue_hugetlb_folio_node_exact(struct hstate *h, 1291a36f1e90SSidhartha Kumar int nid) 1292bf50bab2SNaoya Horiguchi { 1293a36f1e90SSidhartha Kumar struct folio *folio; 12941a08ae36SPavel Tatashin bool pin = !!(current->flags & PF_MEMALLOC_PIN); 1295bf50bab2SNaoya Horiguchi 12969487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1297a36f1e90SSidhartha Kumar list_for_each_entry(folio, &h->hugepage_freelists[nid], lru) { 1298a36f1e90SSidhartha Kumar if (pin && !folio_is_longterm_pinnable(folio)) 1299bbe88753SJoonsoo Kim continue; 1300bbe88753SJoonsoo Kim 1301a36f1e90SSidhartha Kumar if (folio_test_hwpoison(folio)) 13026664bfc8SWei Yang continue; 1303bbe88753SJoonsoo Kim 1304a36f1e90SSidhartha Kumar list_move(&folio->lru, &h->hugepage_activelist); 1305a36f1e90SSidhartha Kumar folio_ref_unfreeze(folio, 1); 1306a36f1e90SSidhartha Kumar folio_clear_hugetlb_freed(folio); 1307bf50bab2SNaoya Horiguchi h->free_huge_pages--; 1308bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 1309a36f1e90SSidhartha Kumar return folio; 1310bf50bab2SNaoya Horiguchi } 1311bf50bab2SNaoya Horiguchi 13126664bfc8SWei Yang return NULL; 13136664bfc8SWei Yang } 13146664bfc8SWei Yang 1315a36f1e90SSidhartha Kumar static struct folio *dequeue_hugetlb_folio_nodemask(struct hstate *h, gfp_t gfp_mask, 1316a36f1e90SSidhartha Kumar int nid, nodemask_t *nmask) 131794310cbcSAnshuman Khandual { 13183e59fcb0SMichal Hocko unsigned int cpuset_mems_cookie; 13193e59fcb0SMichal Hocko struct zonelist *zonelist; 13203e59fcb0SMichal Hocko struct zone *zone; 13213e59fcb0SMichal Hocko struct zoneref *z; 132298fa15f3SAnshuman Khandual int node = NUMA_NO_NODE; 13233e59fcb0SMichal Hocko 13243e59fcb0SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 13253e59fcb0SMichal Hocko 13263e59fcb0SMichal Hocko retry_cpuset: 13273e59fcb0SMichal Hocko cpuset_mems_cookie = read_mems_allowed_begin(); 13283e59fcb0SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 1329a36f1e90SSidhartha Kumar struct folio *folio; 133094310cbcSAnshuman Khandual 13313e59fcb0SMichal Hocko if (!cpuset_zone_allowed(zone, gfp_mask)) 13323e59fcb0SMichal Hocko continue; 13333e59fcb0SMichal Hocko /* 13343e59fcb0SMichal Hocko * no need to ask again on the same node. Pool is node rather than 13353e59fcb0SMichal Hocko * zone aware 13363e59fcb0SMichal Hocko */ 13373e59fcb0SMichal Hocko if (zone_to_nid(zone) == node) 13383e59fcb0SMichal Hocko continue; 13393e59fcb0SMichal Hocko node = zone_to_nid(zone); 134094310cbcSAnshuman Khandual 1341a36f1e90SSidhartha Kumar folio = dequeue_hugetlb_folio_node_exact(h, node); 1342a36f1e90SSidhartha Kumar if (folio) 1343a36f1e90SSidhartha Kumar return folio; 134494310cbcSAnshuman Khandual } 13453e59fcb0SMichal Hocko if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 13463e59fcb0SMichal Hocko goto retry_cpuset; 13473e59fcb0SMichal Hocko 134894310cbcSAnshuman Khandual return NULL; 134994310cbcSAnshuman Khandual } 135094310cbcSAnshuman Khandual 13518346d69dSXin Hao static unsigned long available_huge_pages(struct hstate *h) 13528346d69dSXin Hao { 13538346d69dSXin Hao return h->free_huge_pages - h->resv_huge_pages; 13548346d69dSXin Hao } 13558346d69dSXin Hao 1356ff7d853bSSidhartha Kumar static struct folio *dequeue_hugetlb_folio_vma(struct hstate *h, 1357a5516438SAndi Kleen struct vm_area_struct *vma, 1358af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 1359af0ed73eSJoonsoo Kim long chg) 13601da177e4SLinus Torvalds { 1361a36f1e90SSidhartha Kumar struct folio *folio = NULL; 1362480eccf9SLee Schermerhorn struct mempolicy *mpol; 136304ec6264SVlastimil Babka gfp_t gfp_mask; 13643e59fcb0SMichal Hocko nodemask_t *nodemask; 136504ec6264SVlastimil Babka int nid; 13661da177e4SLinus Torvalds 1367a1e78772SMel Gorman /* 1368a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 1369a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 1370a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 1371a1e78772SMel Gorman */ 13728346d69dSXin Hao if (!vma_has_reserves(vma, chg) && !available_huge_pages(h)) 1373c0ff7453SMiao Xie goto err; 1374a1e78772SMel Gorman 137504f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 13768346d69dSXin Hao if (avoid_reserve && !available_huge_pages(h)) 13776eab04a8SJustin P. Mattock goto err; 137804f2cbe3SMel Gorman 137904ec6264SVlastimil Babka gfp_mask = htlb_alloc_mask(h); 138004ec6264SVlastimil Babka nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 1381cfcaa66fSBen Widawsky 1382cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 1383a36f1e90SSidhartha Kumar folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask, 1384a36f1e90SSidhartha Kumar nid, nodemask); 1385cfcaa66fSBen Widawsky 1386cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 1387cfcaa66fSBen Widawsky nodemask = NULL; 1388cfcaa66fSBen Widawsky } 1389cfcaa66fSBen Widawsky 1390a36f1e90SSidhartha Kumar if (!folio) 1391a36f1e90SSidhartha Kumar folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask, 1392a36f1e90SSidhartha Kumar nid, nodemask); 1393cfcaa66fSBen Widawsky 1394a36f1e90SSidhartha Kumar if (folio && !avoid_reserve && vma_has_reserves(vma, chg)) { 1395a36f1e90SSidhartha Kumar folio_set_hugetlb_restore_reserve(folio); 1396a63884e9SJoonsoo Kim h->resv_huge_pages--; 1397bf50bab2SNaoya Horiguchi } 1398cc9a6c87SMel Gorman 1399cc9a6c87SMel Gorman mpol_cond_put(mpol); 1400ff7d853bSSidhartha Kumar return folio; 1401cc9a6c87SMel Gorman 1402c0ff7453SMiao Xie err: 1403cc9a6c87SMel Gorman return NULL; 14041da177e4SLinus Torvalds } 14051da177e4SLinus Torvalds 14061cac6f2cSLuiz Capitulino /* 14071cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 14081cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 14091cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 14101cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 14111cac6f2cSLuiz Capitulino * node for alloc or free. 14121cac6f2cSLuiz Capitulino */ 14131cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 14141cac6f2cSLuiz Capitulino { 14150edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 14161cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 14171cac6f2cSLuiz Capitulino 14181cac6f2cSLuiz Capitulino return nid; 14191cac6f2cSLuiz Capitulino } 14201cac6f2cSLuiz Capitulino 14211cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 14221cac6f2cSLuiz Capitulino { 14231cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 14241cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 14251cac6f2cSLuiz Capitulino return nid; 14261cac6f2cSLuiz Capitulino } 14271cac6f2cSLuiz Capitulino 14281cac6f2cSLuiz Capitulino /* 14291cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 14301cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 14311cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 14321cac6f2cSLuiz Capitulino * mask. 14331cac6f2cSLuiz Capitulino */ 14341cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 14351cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 14361cac6f2cSLuiz Capitulino { 14371cac6f2cSLuiz Capitulino int nid; 14381cac6f2cSLuiz Capitulino 14391cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 14401cac6f2cSLuiz Capitulino 14411cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 14421cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 14431cac6f2cSLuiz Capitulino 14441cac6f2cSLuiz Capitulino return nid; 14451cac6f2cSLuiz Capitulino } 14461cac6f2cSLuiz Capitulino 14471cac6f2cSLuiz Capitulino /* 144810c6ec49SMike Kravetz * helper for remove_pool_huge_page() - return the previously saved 14491cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 14501cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 14511cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 14521cac6f2cSLuiz Capitulino */ 14531cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 14541cac6f2cSLuiz Capitulino { 14551cac6f2cSLuiz Capitulino int nid; 14561cac6f2cSLuiz Capitulino 14571cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 14581cac6f2cSLuiz Capitulino 14591cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 14601cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 14611cac6f2cSLuiz Capitulino 14621cac6f2cSLuiz Capitulino return nid; 14631cac6f2cSLuiz Capitulino } 14641cac6f2cSLuiz Capitulino 14651cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 14661cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 14671cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 14681cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 14691cac6f2cSLuiz Capitulino nr_nodes--) 14701cac6f2cSLuiz Capitulino 14711cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 14721cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 14731cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 14741cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 14751cac6f2cSLuiz Capitulino nr_nodes--) 14761cac6f2cSLuiz Capitulino 14778531fc6fSMike Kravetz /* used to demote non-gigantic_huge pages as well */ 1478911565b8SSidhartha Kumar static void __destroy_compound_gigantic_folio(struct folio *folio, 147934d9e35bSMike Kravetz unsigned int order, bool demote) 1480944d9fecSLuiz Capitulino { 1481944d9fecSLuiz Capitulino int i; 1482944d9fecSLuiz Capitulino int nr_pages = 1 << order; 148314455eabSCheng Li struct page *p; 1484944d9fecSLuiz Capitulino 148546f27228SMatthew Wilcox (Oracle) atomic_set(&folio->_entire_mapcount, 0); 1486eec20426SMatthew Wilcox (Oracle) atomic_set(&folio->_nr_pages_mapped, 0); 148794688e8eSMatthew Wilcox (Oracle) atomic_set(&folio->_pincount, 0); 148847e29d32SJohn Hubbard 148914455eabSCheng Li for (i = 1; i < nr_pages; i++) { 1490911565b8SSidhartha Kumar p = folio_page(folio, i); 14916c141973SMike Kravetz p->flags &= ~PAGE_FLAGS_CHECK_AT_FREE; 1492a01f4390SMike Kravetz p->mapping = NULL; 14931d798ca3SKirill A. Shutemov clear_compound_head(p); 149434d9e35bSMike Kravetz if (!demote) 1495944d9fecSLuiz Capitulino set_page_refcounted(p); 1496944d9fecSLuiz Capitulino } 1497944d9fecSLuiz Capitulino 1498911565b8SSidhartha Kumar __folio_clear_head(folio); 1499944d9fecSLuiz Capitulino } 1500944d9fecSLuiz Capitulino 1501911565b8SSidhartha Kumar static void destroy_compound_hugetlb_folio_for_demote(struct folio *folio, 15028531fc6fSMike Kravetz unsigned int order) 15038531fc6fSMike Kravetz { 1504911565b8SSidhartha Kumar __destroy_compound_gigantic_folio(folio, order, true); 15058531fc6fSMike Kravetz } 15068531fc6fSMike Kravetz 15078531fc6fSMike Kravetz #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 1508911565b8SSidhartha Kumar static void destroy_compound_gigantic_folio(struct folio *folio, 150934d9e35bSMike Kravetz unsigned int order) 151034d9e35bSMike Kravetz { 1511911565b8SSidhartha Kumar __destroy_compound_gigantic_folio(folio, order, false); 151234d9e35bSMike Kravetz } 151334d9e35bSMike Kravetz 15147f325a8dSSidhartha Kumar static void free_gigantic_folio(struct folio *folio, unsigned int order) 1515944d9fecSLuiz Capitulino { 1516cf11e85fSRoman Gushchin /* 1517cf11e85fSRoman Gushchin * If the page isn't allocated using the cma allocator, 1518cf11e85fSRoman Gushchin * cma_release() returns false. 1519cf11e85fSRoman Gushchin */ 1520dbda8feaSBarry Song #ifdef CONFIG_CMA 15217f325a8dSSidhartha Kumar int nid = folio_nid(folio); 15227f325a8dSSidhartha Kumar 15237f325a8dSSidhartha Kumar if (cma_release(hugetlb_cma[nid], &folio->page, 1 << order)) 1524cf11e85fSRoman Gushchin return; 1525dbda8feaSBarry Song #endif 1526cf11e85fSRoman Gushchin 15277f325a8dSSidhartha Kumar free_contig_range(folio_pfn(folio), 1 << order); 1528944d9fecSLuiz Capitulino } 1529944d9fecSLuiz Capitulino 15304eb0716eSAlexandre Ghiti #ifdef CONFIG_CONTIG_ALLOC 153119fc1a7eSSidhartha Kumar static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask, 1532d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) 1533944d9fecSLuiz Capitulino { 153419fc1a7eSSidhartha Kumar struct page *page; 153504adbc3fSMiaohe Lin unsigned long nr_pages = pages_per_huge_page(h); 1536953f064aSLi Xinhai if (nid == NUMA_NO_NODE) 1537953f064aSLi Xinhai nid = numa_mem_id(); 1538944d9fecSLuiz Capitulino 1539dbda8feaSBarry Song #ifdef CONFIG_CMA 1540dbda8feaSBarry Song { 1541cf11e85fSRoman Gushchin int node; 1542cf11e85fSRoman Gushchin 1543953f064aSLi Xinhai if (hugetlb_cma[nid]) { 1544953f064aSLi Xinhai page = cma_alloc(hugetlb_cma[nid], nr_pages, 1545953f064aSLi Xinhai huge_page_order(h), true); 1546953f064aSLi Xinhai if (page) 154719fc1a7eSSidhartha Kumar return page_folio(page); 1548953f064aSLi Xinhai } 1549953f064aSLi Xinhai 1550953f064aSLi Xinhai if (!(gfp_mask & __GFP_THISNODE)) { 1551cf11e85fSRoman Gushchin for_each_node_mask(node, *nodemask) { 1552953f064aSLi Xinhai if (node == nid || !hugetlb_cma[node]) 1553cf11e85fSRoman Gushchin continue; 1554cf11e85fSRoman Gushchin 1555cf11e85fSRoman Gushchin page = cma_alloc(hugetlb_cma[node], nr_pages, 1556cf11e85fSRoman Gushchin huge_page_order(h), true); 1557cf11e85fSRoman Gushchin if (page) 155819fc1a7eSSidhartha Kumar return page_folio(page); 1559cf11e85fSRoman Gushchin } 1560cf11e85fSRoman Gushchin } 1561953f064aSLi Xinhai } 1562dbda8feaSBarry Song #endif 1563cf11e85fSRoman Gushchin 156419fc1a7eSSidhartha Kumar page = alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); 156519fc1a7eSSidhartha Kumar return page ? page_folio(page) : NULL; 1566944d9fecSLuiz Capitulino } 1567944d9fecSLuiz Capitulino 15684eb0716eSAlexandre Ghiti #else /* !CONFIG_CONTIG_ALLOC */ 156919fc1a7eSSidhartha Kumar static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask, 15704eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 15714eb0716eSAlexandre Ghiti { 15724eb0716eSAlexandre Ghiti return NULL; 15734eb0716eSAlexandre Ghiti } 15744eb0716eSAlexandre Ghiti #endif /* CONFIG_CONTIG_ALLOC */ 1575944d9fecSLuiz Capitulino 1576e1073d1eSAneesh Kumar K.V #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 157719fc1a7eSSidhartha Kumar static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask, 15784eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 15794eb0716eSAlexandre Ghiti { 15804eb0716eSAlexandre Ghiti return NULL; 15814eb0716eSAlexandre Ghiti } 15827f325a8dSSidhartha Kumar static inline void free_gigantic_folio(struct folio *folio, 15837f325a8dSSidhartha Kumar unsigned int order) { } 1584911565b8SSidhartha Kumar static inline void destroy_compound_gigantic_folio(struct folio *folio, 1585d00181b9SKirill A. Shutemov unsigned int order) { } 1586944d9fecSLuiz Capitulino #endif 1587944d9fecSLuiz Capitulino 158832c87719SMike Kravetz static inline void __clear_hugetlb_destructor(struct hstate *h, 158932c87719SMike Kravetz struct folio *folio) 159032c87719SMike Kravetz { 159132c87719SMike Kravetz lockdep_assert_held(&hugetlb_lock); 159232c87719SMike Kravetz 15939c5ccf2dSMatthew Wilcox (Oracle) folio_clear_hugetlb(folio); 159432c87719SMike Kravetz } 159532c87719SMike Kravetz 159632c87719SMike Kravetz /* 159732c87719SMike Kravetz * Remove hugetlb folio from lists. 159832c87719SMike Kravetz * If vmemmap exists for the folio, update dtor so that the folio appears 159932c87719SMike Kravetz * as just a compound page. Otherwise, wait until after allocating vmemmap 160032c87719SMike Kravetz * to update dtor. 160134d9e35bSMike Kravetz * 1602cfd5082bSSidhartha Kumar * A reference is held on the folio, except in the case of demote. 16036eb4e88aSMike Kravetz * 16046eb4e88aSMike Kravetz * Must be called with hugetlb lock held. 16056eb4e88aSMike Kravetz */ 1606cfd5082bSSidhartha Kumar static void __remove_hugetlb_folio(struct hstate *h, struct folio *folio, 160734d9e35bSMike Kravetz bool adjust_surplus, 160834d9e35bSMike Kravetz bool demote) 16096eb4e88aSMike Kravetz { 1610cfd5082bSSidhartha Kumar int nid = folio_nid(folio); 16116eb4e88aSMike Kravetz 1612f074732dSSidhartha Kumar VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio(folio), folio); 1613f074732dSSidhartha Kumar VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio_rsvd(folio), folio); 16146eb4e88aSMike Kravetz 16159487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 16166eb4e88aSMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 16176eb4e88aSMike Kravetz return; 16186eb4e88aSMike Kravetz 1619cfd5082bSSidhartha Kumar list_del(&folio->lru); 16206eb4e88aSMike Kravetz 1621cfd5082bSSidhartha Kumar if (folio_test_hugetlb_freed(folio)) { 16226eb4e88aSMike Kravetz h->free_huge_pages--; 16236eb4e88aSMike Kravetz h->free_huge_pages_node[nid]--; 16246eb4e88aSMike Kravetz } 16256eb4e88aSMike Kravetz if (adjust_surplus) { 16266eb4e88aSMike Kravetz h->surplus_huge_pages--; 16276eb4e88aSMike Kravetz h->surplus_huge_pages_node[nid]--; 16286eb4e88aSMike Kravetz } 16296eb4e88aSMike Kravetz 1630e32d20c0SMike Kravetz /* 163132c87719SMike Kravetz * We can only clear the hugetlb destructor after allocating vmemmap 163232c87719SMike Kravetz * pages. Otherwise, someone (memory error handling) may try to write 163332c87719SMike Kravetz * to tail struct pages. 163432c87719SMike Kravetz */ 163532c87719SMike Kravetz if (!folio_test_hugetlb_vmemmap_optimized(folio)) 163632c87719SMike Kravetz __clear_hugetlb_destructor(h, folio); 163732c87719SMike Kravetz 163832c87719SMike Kravetz /* 163934d9e35bSMike Kravetz * In the case of demote we do not ref count the page as it will soon 164034d9e35bSMike Kravetz * be turned into a page of smaller size. 1641e32d20c0SMike Kravetz */ 164234d9e35bSMike Kravetz if (!demote) 1643cfd5082bSSidhartha Kumar folio_ref_unfreeze(folio, 1); 16446eb4e88aSMike Kravetz 16456eb4e88aSMike Kravetz h->nr_huge_pages--; 16466eb4e88aSMike Kravetz h->nr_huge_pages_node[nid]--; 16476eb4e88aSMike Kravetz } 16486eb4e88aSMike Kravetz 1649cfd5082bSSidhartha Kumar static void remove_hugetlb_folio(struct hstate *h, struct folio *folio, 165034d9e35bSMike Kravetz bool adjust_surplus) 165134d9e35bSMike Kravetz { 1652cfd5082bSSidhartha Kumar __remove_hugetlb_folio(h, folio, adjust_surplus, false); 165334d9e35bSMike Kravetz } 165434d9e35bSMike Kravetz 1655cfd5082bSSidhartha Kumar static void remove_hugetlb_folio_for_demote(struct hstate *h, struct folio *folio, 16568531fc6fSMike Kravetz bool adjust_surplus) 16578531fc6fSMike Kravetz { 1658cfd5082bSSidhartha Kumar __remove_hugetlb_folio(h, folio, adjust_surplus, true); 16598531fc6fSMike Kravetz } 16608531fc6fSMike Kravetz 16612f6c57d6SSidhartha Kumar static void add_hugetlb_folio(struct hstate *h, struct folio *folio, 1662ad2fa371SMuchun Song bool adjust_surplus) 1663ad2fa371SMuchun Song { 1664ad2fa371SMuchun Song int zeroed; 16652f6c57d6SSidhartha Kumar int nid = folio_nid(folio); 1666ad2fa371SMuchun Song 16672f6c57d6SSidhartha Kumar VM_BUG_ON_FOLIO(!folio_test_hugetlb_vmemmap_optimized(folio), folio); 1668ad2fa371SMuchun Song 1669ad2fa371SMuchun Song lockdep_assert_held(&hugetlb_lock); 1670ad2fa371SMuchun Song 16712f6c57d6SSidhartha Kumar INIT_LIST_HEAD(&folio->lru); 1672ad2fa371SMuchun Song h->nr_huge_pages++; 1673ad2fa371SMuchun Song h->nr_huge_pages_node[nid]++; 1674ad2fa371SMuchun Song 1675ad2fa371SMuchun Song if (adjust_surplus) { 1676ad2fa371SMuchun Song h->surplus_huge_pages++; 1677ad2fa371SMuchun Song h->surplus_huge_pages_node[nid]++; 1678ad2fa371SMuchun Song } 1679ad2fa371SMuchun Song 16809c5ccf2dSMatthew Wilcox (Oracle) folio_set_hugetlb(folio); 16812f6c57d6SSidhartha Kumar folio_change_private(folio, NULL); 1682a9e1eab2SMiaohe Lin /* 16832f6c57d6SSidhartha Kumar * We have to set hugetlb_vmemmap_optimized again as above 16842f6c57d6SSidhartha Kumar * folio_change_private(folio, NULL) cleared it. 1685a9e1eab2SMiaohe Lin */ 16862f6c57d6SSidhartha Kumar folio_set_hugetlb_vmemmap_optimized(folio); 1687ad2fa371SMuchun Song 1688ad2fa371SMuchun Song /* 16892f6c57d6SSidhartha Kumar * This folio is about to be managed by the hugetlb allocator and 1690b65a4edaSMike Kravetz * should have no users. Drop our reference, and check for others 1691b65a4edaSMike Kravetz * just in case. 1692ad2fa371SMuchun Song */ 16932f6c57d6SSidhartha Kumar zeroed = folio_put_testzero(folio); 16942f6c57d6SSidhartha Kumar if (unlikely(!zeroed)) 1695b65a4edaSMike Kravetz /* 1696454a00c4SMatthew Wilcox (Oracle) * It is VERY unlikely soneone else has taken a ref 1697454a00c4SMatthew Wilcox (Oracle) * on the folio. In this case, we simply return as 1698454a00c4SMatthew Wilcox (Oracle) * free_huge_folio() will be called when this other ref 1699454a00c4SMatthew Wilcox (Oracle) * is dropped. 1700b65a4edaSMike Kravetz */ 1701b65a4edaSMike Kravetz return; 1702b65a4edaSMike Kravetz 17032f6c57d6SSidhartha Kumar arch_clear_hugepage_flags(&folio->page); 1704240d67a8SSidhartha Kumar enqueue_hugetlb_folio(h, folio); 1705ad2fa371SMuchun Song } 1706ad2fa371SMuchun Song 17076f6956cfSSidhartha Kumar static void __update_and_free_hugetlb_folio(struct hstate *h, 17086f6956cfSSidhartha Kumar struct folio *folio) 17096af2acb6SAdam Litke { 171032c87719SMike Kravetz bool clear_dtor = folio_test_hugetlb_vmemmap_optimized(folio); 1711a5516438SAndi Kleen 17124eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1713944d9fecSLuiz Capitulino return; 171418229df5SAndy Whitcroft 1715161df60eSNaoya Horiguchi /* 1716161df60eSNaoya Horiguchi * If we don't know which subpages are hwpoisoned, we can't free 1717161df60eSNaoya Horiguchi * the hugepage, so it's leaked intentionally. 1718161df60eSNaoya Horiguchi */ 17197f325a8dSSidhartha Kumar if (folio_test_hugetlb_raw_hwp_unreliable(folio)) 1720161df60eSNaoya Horiguchi return; 1721161df60eSNaoya Horiguchi 17226f6956cfSSidhartha Kumar if (hugetlb_vmemmap_restore(h, &folio->page)) { 1723ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 1724ad2fa371SMuchun Song /* 1725ad2fa371SMuchun Song * If we cannot allocate vmemmap pages, just refuse to free the 1726ad2fa371SMuchun Song * page and put the page back on the hugetlb free list and treat 1727ad2fa371SMuchun Song * as a surplus page. 1728ad2fa371SMuchun Song */ 17297f325a8dSSidhartha Kumar add_hugetlb_folio(h, folio, true); 1730ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 1731ad2fa371SMuchun Song return; 1732ad2fa371SMuchun Song } 1733ad2fa371SMuchun Song 1734161df60eSNaoya Horiguchi /* 1735161df60eSNaoya Horiguchi * Move PageHWPoison flag from head page to the raw error pages, 1736161df60eSNaoya Horiguchi * which makes any healthy subpages reusable. 1737161df60eSNaoya Horiguchi */ 1738911565b8SSidhartha Kumar if (unlikely(folio_test_hwpoison(folio))) 17392ff6ceceSSidhartha Kumar folio_clear_hugetlb_hwpoison(folio); 1740161df60eSNaoya Horiguchi 174132c87719SMike Kravetz /* 174232c87719SMike Kravetz * If vmemmap pages were allocated above, then we need to clear the 174332c87719SMike Kravetz * hugetlb destructor under the hugetlb lock. 174432c87719SMike Kravetz */ 174532c87719SMike Kravetz if (clear_dtor) { 174632c87719SMike Kravetz spin_lock_irq(&hugetlb_lock); 174732c87719SMike Kravetz __clear_hugetlb_destructor(h, folio); 174832c87719SMike Kravetz spin_unlock_irq(&hugetlb_lock); 174932c87719SMike Kravetz } 175032c87719SMike Kravetz 1751a01f4390SMike Kravetz /* 1752a01f4390SMike Kravetz * Non-gigantic pages demoted from CMA allocated gigantic pages 17537f325a8dSSidhartha Kumar * need to be given back to CMA in free_gigantic_folio. 1754a01f4390SMike Kravetz */ 1755a01f4390SMike Kravetz if (hstate_is_gigantic(h) || 17562f6c57d6SSidhartha Kumar hugetlb_cma_folio(folio, huge_page_order(h))) { 1757911565b8SSidhartha Kumar destroy_compound_gigantic_folio(folio, huge_page_order(h)); 17587f325a8dSSidhartha Kumar free_gigantic_folio(folio, huge_page_order(h)); 1759944d9fecSLuiz Capitulino } else { 17606f6956cfSSidhartha Kumar __free_pages(&folio->page, huge_page_order(h)); 17616af2acb6SAdam Litke } 1762944d9fecSLuiz Capitulino } 17636af2acb6SAdam Litke 1764b65d4adbSMuchun Song /* 1765d6ef19e2SSidhartha Kumar * As update_and_free_hugetlb_folio() can be called under any context, so we cannot 1766b65d4adbSMuchun Song * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the 1767b65d4adbSMuchun Song * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate 1768b65d4adbSMuchun Song * the vmemmap pages. 1769b65d4adbSMuchun Song * 1770b65d4adbSMuchun Song * free_hpage_workfn() locklessly retrieves the linked list of pages to be 1771b65d4adbSMuchun Song * freed and frees them one-by-one. As the page->mapping pointer is going 1772b65d4adbSMuchun Song * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node 1773b65d4adbSMuchun Song * structure of a lockless linked list of huge pages to be freed. 1774b65d4adbSMuchun Song */ 1775b65d4adbSMuchun Song static LLIST_HEAD(hpage_freelist); 1776b65d4adbSMuchun Song 1777b65d4adbSMuchun Song static void free_hpage_workfn(struct work_struct *work) 1778b65d4adbSMuchun Song { 1779b65d4adbSMuchun Song struct llist_node *node; 1780b65d4adbSMuchun Song 1781b65d4adbSMuchun Song node = llist_del_all(&hpage_freelist); 1782b65d4adbSMuchun Song 1783b65d4adbSMuchun Song while (node) { 1784b65d4adbSMuchun Song struct page *page; 1785b65d4adbSMuchun Song struct hstate *h; 1786b65d4adbSMuchun Song 1787b65d4adbSMuchun Song page = container_of((struct address_space **)node, 1788b65d4adbSMuchun Song struct page, mapping); 1789b65d4adbSMuchun Song node = node->next; 1790b65d4adbSMuchun Song page->mapping = NULL; 1791b65d4adbSMuchun Song /* 1792affd26b1SSidhartha Kumar * The VM_BUG_ON_FOLIO(!folio_test_hugetlb(folio), folio) in 1793affd26b1SSidhartha Kumar * folio_hstate() is going to trigger because a previous call to 17949c5ccf2dSMatthew Wilcox (Oracle) * remove_hugetlb_folio() will clear the hugetlb bit, so do 17959c5ccf2dSMatthew Wilcox (Oracle) * not use folio_hstate() directly. 1796b65d4adbSMuchun Song */ 1797b65d4adbSMuchun Song h = size_to_hstate(page_size(page)); 1798b65d4adbSMuchun Song 17996f6956cfSSidhartha Kumar __update_and_free_hugetlb_folio(h, page_folio(page)); 1800b65d4adbSMuchun Song 1801b65d4adbSMuchun Song cond_resched(); 1802b65d4adbSMuchun Song } 1803b65d4adbSMuchun Song } 1804b65d4adbSMuchun Song static DECLARE_WORK(free_hpage_work, free_hpage_workfn); 1805b65d4adbSMuchun Song 1806b65d4adbSMuchun Song static inline void flush_free_hpage_work(struct hstate *h) 1807b65d4adbSMuchun Song { 18086213834cSMuchun Song if (hugetlb_vmemmap_optimizable(h)) 1809b65d4adbSMuchun Song flush_work(&free_hpage_work); 1810b65d4adbSMuchun Song } 1811b65d4adbSMuchun Song 1812d6ef19e2SSidhartha Kumar static void update_and_free_hugetlb_folio(struct hstate *h, struct folio *folio, 1813b65d4adbSMuchun Song bool atomic) 1814b65d4adbSMuchun Song { 1815d6ef19e2SSidhartha Kumar if (!folio_test_hugetlb_vmemmap_optimized(folio) || !atomic) { 18166f6956cfSSidhartha Kumar __update_and_free_hugetlb_folio(h, folio); 1817b65d4adbSMuchun Song return; 1818b65d4adbSMuchun Song } 1819b65d4adbSMuchun Song 1820b65d4adbSMuchun Song /* 1821b65d4adbSMuchun Song * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages. 1822b65d4adbSMuchun Song * 1823b65d4adbSMuchun Song * Only call schedule_work() if hpage_freelist is previously 1824b65d4adbSMuchun Song * empty. Otherwise, schedule_work() had been called but the workfn 1825b65d4adbSMuchun Song * hasn't retrieved the list yet. 1826b65d4adbSMuchun Song */ 1827d6ef19e2SSidhartha Kumar if (llist_add((struct llist_node *)&folio->mapping, &hpage_freelist)) 1828b65d4adbSMuchun Song schedule_work(&free_hpage_work); 1829b65d4adbSMuchun Song } 1830b65d4adbSMuchun Song 183110c6ec49SMike Kravetz static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list) 183210c6ec49SMike Kravetz { 183310c6ec49SMike Kravetz struct page *page, *t_page; 1834d6ef19e2SSidhartha Kumar struct folio *folio; 183510c6ec49SMike Kravetz 183610c6ec49SMike Kravetz list_for_each_entry_safe(page, t_page, list, lru) { 1837d6ef19e2SSidhartha Kumar folio = page_folio(page); 1838d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, folio, false); 183910c6ec49SMike Kravetz cond_resched(); 184010c6ec49SMike Kravetz } 184110c6ec49SMike Kravetz } 184210c6ec49SMike Kravetz 1843e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1844e5ff2159SAndi Kleen { 1845e5ff2159SAndi Kleen struct hstate *h; 1846e5ff2159SAndi Kleen 1847e5ff2159SAndi Kleen for_each_hstate(h) { 1848e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1849e5ff2159SAndi Kleen return h; 1850e5ff2159SAndi Kleen } 1851e5ff2159SAndi Kleen return NULL; 1852e5ff2159SAndi Kleen } 1853e5ff2159SAndi Kleen 1854454a00c4SMatthew Wilcox (Oracle) void free_huge_folio(struct folio *folio) 185527a85ef1SDavid Gibson { 1856a5516438SAndi Kleen /* 1857a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1858a5516438SAndi Kleen * compound page destructor. 1859a5516438SAndi Kleen */ 18600356c4b9SSidhartha Kumar struct hstate *h = folio_hstate(folio); 18610356c4b9SSidhartha Kumar int nid = folio_nid(folio); 18620356c4b9SSidhartha Kumar struct hugepage_subpool *spool = hugetlb_folio_subpool(folio); 186307443a85SJoonsoo Kim bool restore_reserve; 1864db71ef79SMike Kravetz unsigned long flags; 186527a85ef1SDavid Gibson 18660356c4b9SSidhartha Kumar VM_BUG_ON_FOLIO(folio_ref_count(folio), folio); 18670356c4b9SSidhartha Kumar VM_BUG_ON_FOLIO(folio_mapcount(folio), folio); 18688ace22bcSYongkai Wu 18690356c4b9SSidhartha Kumar hugetlb_set_folio_subpool(folio, NULL); 18700356c4b9SSidhartha Kumar if (folio_test_anon(folio)) 18710356c4b9SSidhartha Kumar __ClearPageAnonExclusive(&folio->page); 18720356c4b9SSidhartha Kumar folio->mapping = NULL; 18730356c4b9SSidhartha Kumar restore_reserve = folio_test_hugetlb_restore_reserve(folio); 18740356c4b9SSidhartha Kumar folio_clear_hugetlb_restore_reserve(folio); 187527a85ef1SDavid Gibson 18761c5ecae3SMike Kravetz /* 1877d6995da3SMike Kravetz * If HPageRestoreReserve was set on page, page allocation consumed a 18780919e1b6SMike Kravetz * reservation. If the page was associated with a subpool, there 18790919e1b6SMike Kravetz * would have been a page reserved in the subpool before allocation 18800919e1b6SMike Kravetz * via hugepage_subpool_get_pages(). Since we are 'restoring' the 18816c26d310SMiaohe Lin * reservation, do not call hugepage_subpool_put_pages() as this will 18820919e1b6SMike Kravetz * remove the reserved page from the subpool. 18830919e1b6SMike Kravetz */ 18840919e1b6SMike Kravetz if (!restore_reserve) { 18850919e1b6SMike Kravetz /* 18860919e1b6SMike Kravetz * A return code of zero implies that the subpool will be 18870919e1b6SMike Kravetz * under its minimum size if the reservation is not restored 18880919e1b6SMike Kravetz * after page is free. Therefore, force restore_reserve 18890919e1b6SMike Kravetz * operation. 18901c5ecae3SMike Kravetz */ 18911c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 18921c5ecae3SMike Kravetz restore_reserve = true; 18930919e1b6SMike Kravetz } 18941c5ecae3SMike Kravetz 1895db71ef79SMike Kravetz spin_lock_irqsave(&hugetlb_lock, flags); 18960356c4b9SSidhartha Kumar folio_clear_hugetlb_migratable(folio); 1897d4ab0316SSidhartha Kumar hugetlb_cgroup_uncharge_folio(hstate_index(h), 1898d4ab0316SSidhartha Kumar pages_per_huge_page(h), folio); 1899d4ab0316SSidhartha Kumar hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h), 1900d4ab0316SSidhartha Kumar pages_per_huge_page(h), folio); 190107443a85SJoonsoo Kim if (restore_reserve) 190207443a85SJoonsoo Kim h->resv_huge_pages++; 190307443a85SJoonsoo Kim 19040356c4b9SSidhartha Kumar if (folio_test_hugetlb_temporary(folio)) { 1905cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, false); 1906db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1907d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, folio, true); 1908ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 19090edaecfaSAneesh Kumar K.V /* remove the page from active list */ 1910cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, true); 1911db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1912d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, folio, true); 19137893d1d5SAdam Litke } else { 1914454a00c4SMatthew Wilcox (Oracle) arch_clear_hugepage_flags(&folio->page); 1915240d67a8SSidhartha Kumar enqueue_hugetlb_folio(h, folio); 1916db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 191727a85ef1SDavid Gibson } 19181121828aSMike Kravetz } 191927a85ef1SDavid Gibson 1920d3d99fccSOscar Salvador /* 1921d3d99fccSOscar Salvador * Must be called with the hugetlb lock held 1922d3d99fccSOscar Salvador */ 1923d3d99fccSOscar Salvador static void __prep_account_new_huge_page(struct hstate *h, int nid) 1924d3d99fccSOscar Salvador { 1925d3d99fccSOscar Salvador lockdep_assert_held(&hugetlb_lock); 1926d3d99fccSOscar Salvador h->nr_huge_pages++; 1927d3d99fccSOscar Salvador h->nr_huge_pages_node[nid]++; 1928d3d99fccSOscar Salvador } 1929d3d99fccSOscar Salvador 1930de656ed3SSidhartha Kumar static void __prep_new_hugetlb_folio(struct hstate *h, struct folio *folio) 1931b7ba30c6SAndi Kleen { 1932de656ed3SSidhartha Kumar hugetlb_vmemmap_optimize(h, &folio->page); 1933de656ed3SSidhartha Kumar INIT_LIST_HEAD(&folio->lru); 19349c5ccf2dSMatthew Wilcox (Oracle) folio_set_hugetlb(folio); 1935de656ed3SSidhartha Kumar hugetlb_set_folio_subpool(folio, NULL); 1936de656ed3SSidhartha Kumar set_hugetlb_cgroup(folio, NULL); 1937de656ed3SSidhartha Kumar set_hugetlb_cgroup_rsvd(folio, NULL); 1938d3d99fccSOscar Salvador } 1939d3d99fccSOscar Salvador 1940d1c60955SSidhartha Kumar static void prep_new_hugetlb_folio(struct hstate *h, struct folio *folio, int nid) 1941d3d99fccSOscar Salvador { 1942de656ed3SSidhartha Kumar __prep_new_hugetlb_folio(h, folio); 1943db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 1944d3d99fccSOscar Salvador __prep_account_new_huge_page(h, nid); 1945db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1946b7ba30c6SAndi Kleen } 1947b7ba30c6SAndi Kleen 1948d1c60955SSidhartha Kumar static bool __prep_compound_gigantic_folio(struct folio *folio, 1949d1c60955SSidhartha Kumar unsigned int order, bool demote) 195020a0307cSWu Fengguang { 19517118fc29SMike Kravetz int i, j; 195220a0307cSWu Fengguang int nr_pages = 1 << order; 195314455eabSCheng Li struct page *p; 195420a0307cSWu Fengguang 1955d1c60955SSidhartha Kumar __folio_clear_reserved(folio); 19562b21624fSMike Kravetz for (i = 0; i < nr_pages; i++) { 1957d1c60955SSidhartha Kumar p = folio_page(folio, i); 195814455eabSCheng Li 1959ef5a22beSAndrea Arcangeli /* 1960ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1961ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1962ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 19637c8de358SEthon Paul * too. Otherwise drivers using get_user_pages() to access tail 1964ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1965ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1966ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1967ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1968ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1969ef5a22beSAndrea Arcangeli * after get_user_pages(). 1970ef5a22beSAndrea Arcangeli */ 19717fb0728aSMike Kravetz if (i != 0) /* head page cleared above */ 1972ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 19737118fc29SMike Kravetz /* 19747118fc29SMike Kravetz * Subtle and very unlikely 19757118fc29SMike Kravetz * 19767118fc29SMike Kravetz * Gigantic 'page allocators' such as memblock or cma will 19777118fc29SMike Kravetz * return a set of pages with each page ref counted. We need 19787118fc29SMike Kravetz * to turn this set of pages into a compound page with tail 19797118fc29SMike Kravetz * page ref counts set to zero. Code such as speculative page 19807118fc29SMike Kravetz * cache adding could take a ref on a 'to be' tail page. 19817118fc29SMike Kravetz * We need to respect any increased ref count, and only set 19827118fc29SMike Kravetz * the ref count to zero if count is currently 1. If count 1983416d85edSMike Kravetz * is not 1, we return an error. An error return indicates 1984416d85edSMike Kravetz * the set of pages can not be converted to a gigantic page. 1985416d85edSMike Kravetz * The caller who allocated the pages should then discard the 1986416d85edSMike Kravetz * pages using the appropriate free interface. 198734d9e35bSMike Kravetz * 198834d9e35bSMike Kravetz * In the case of demote, the ref count will be zero. 19897118fc29SMike Kravetz */ 199034d9e35bSMike Kravetz if (!demote) { 19917118fc29SMike Kravetz if (!page_ref_freeze(p, 1)) { 1992416d85edSMike Kravetz pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n"); 19937118fc29SMike Kravetz goto out_error; 19947118fc29SMike Kravetz } 199534d9e35bSMike Kravetz } else { 199634d9e35bSMike Kravetz VM_BUG_ON_PAGE(page_count(p), p); 199734d9e35bSMike Kravetz } 19982b21624fSMike Kravetz if (i != 0) 1999d1c60955SSidhartha Kumar set_compound_head(p, &folio->page); 200020a0307cSWu Fengguang } 2001e3b7bf97STarun Sahu __folio_set_head(folio); 2002e3b7bf97STarun Sahu /* we rely on prep_new_hugetlb_folio to set the destructor */ 2003e3b7bf97STarun Sahu folio_set_order(folio, order); 200446f27228SMatthew Wilcox (Oracle) atomic_set(&folio->_entire_mapcount, -1); 2005eec20426SMatthew Wilcox (Oracle) atomic_set(&folio->_nr_pages_mapped, 0); 200694688e8eSMatthew Wilcox (Oracle) atomic_set(&folio->_pincount, 0); 20077118fc29SMike Kravetz return true; 20087118fc29SMike Kravetz 20097118fc29SMike Kravetz out_error: 20102b21624fSMike Kravetz /* undo page modifications made above */ 20112b21624fSMike Kravetz for (j = 0; j < i; j++) { 2012d1c60955SSidhartha Kumar p = folio_page(folio, j); 20132b21624fSMike Kravetz if (j != 0) 20147118fc29SMike Kravetz clear_compound_head(p); 20157118fc29SMike Kravetz set_page_refcounted(p); 20167118fc29SMike Kravetz } 20177118fc29SMike Kravetz /* need to clear PG_reserved on remaining tail pages */ 201814455eabSCheng Li for (; j < nr_pages; j++) { 2019d1c60955SSidhartha Kumar p = folio_page(folio, j); 20207118fc29SMike Kravetz __ClearPageReserved(p); 202114455eabSCheng Li } 20227118fc29SMike Kravetz return false; 202320a0307cSWu Fengguang } 202420a0307cSWu Fengguang 2025d1c60955SSidhartha Kumar static bool prep_compound_gigantic_folio(struct folio *folio, 20268531fc6fSMike Kravetz unsigned int order) 20278531fc6fSMike Kravetz { 2028d1c60955SSidhartha Kumar return __prep_compound_gigantic_folio(folio, order, false); 2029d1c60955SSidhartha Kumar } 2030d1c60955SSidhartha Kumar 2031d1c60955SSidhartha Kumar static bool prep_compound_gigantic_folio_for_demote(struct folio *folio, 2032d1c60955SSidhartha Kumar unsigned int order) 2033d1c60955SSidhartha Kumar { 2034d1c60955SSidhartha Kumar return __prep_compound_gigantic_folio(folio, order, true); 20358531fc6fSMike Kravetz } 20368531fc6fSMike Kravetz 20377795912cSAndrew Morton /* 20387795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 20397795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 20407795912cSAndrew Morton * details. 20417795912cSAndrew Morton */ 204220a0307cSWu Fengguang int PageHuge(struct page *page) 204320a0307cSWu Fengguang { 20442d678c64SMatthew Wilcox (Oracle) struct folio *folio; 20452d678c64SMatthew Wilcox (Oracle) 204620a0307cSWu Fengguang if (!PageCompound(page)) 204720a0307cSWu Fengguang return 0; 20482d678c64SMatthew Wilcox (Oracle) folio = page_folio(page); 20499c5ccf2dSMatthew Wilcox (Oracle) return folio_test_hugetlb(folio); 205020a0307cSWu Fengguang } 205143131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 205243131e14SNaoya Horiguchi 2053c0d0381aSMike Kravetz /* 2054c0d0381aSMike Kravetz * Find and lock address space (mapping) in write mode. 2055c0d0381aSMike Kravetz * 2056336bf30eSMike Kravetz * Upon entry, the page is locked which means that page_mapping() is 2057336bf30eSMike Kravetz * stable. Due to locking order, we can only trylock_write. If we can 2058336bf30eSMike Kravetz * not get the lock, simply return NULL to caller. 2059c0d0381aSMike Kravetz */ 2060c0d0381aSMike Kravetz struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) 2061c0d0381aSMike Kravetz { 2062336bf30eSMike Kravetz struct address_space *mapping = page_mapping(hpage); 2063c0d0381aSMike Kravetz 2064c0d0381aSMike Kravetz if (!mapping) 2065c0d0381aSMike Kravetz return mapping; 2066c0d0381aSMike Kravetz 2067c0d0381aSMike Kravetz if (i_mmap_trylock_write(mapping)) 2068c0d0381aSMike Kravetz return mapping; 2069c0d0381aSMike Kravetz 2070c0d0381aSMike Kravetz return NULL; 2071c0d0381aSMike Kravetz } 2072c0d0381aSMike Kravetz 2073fe19bd3dSHugh Dickins pgoff_t hugetlb_basepage_index(struct page *page) 207413d60f4bSZhang Yi { 207513d60f4bSZhang Yi struct page *page_head = compound_head(page); 207613d60f4bSZhang Yi pgoff_t index = page_index(page_head); 207713d60f4bSZhang Yi unsigned long compound_idx; 207813d60f4bSZhang Yi 207923baf831SKirill A. Shutemov if (compound_order(page_head) > MAX_ORDER) 208013d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 208113d60f4bSZhang Yi else 208213d60f4bSZhang Yi compound_idx = page - page_head; 208313d60f4bSZhang Yi 208413d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 208513d60f4bSZhang Yi } 208613d60f4bSZhang Yi 208719fc1a7eSSidhartha Kumar static struct folio *alloc_buddy_hugetlb_folio(struct hstate *h, 2088f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 2089f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 20901da177e4SLinus Torvalds { 2091af0fb9dfSMichal Hocko int order = huge_page_order(h); 20921da177e4SLinus Torvalds struct page *page; 2093f60858f9SMike Kravetz bool alloc_try_hard = true; 20942b21624fSMike Kravetz bool retry = true; 2095f96efd58SJoe Jin 2096f60858f9SMike Kravetz /* 2097f60858f9SMike Kravetz * By default we always try hard to allocate the page with 2098f60858f9SMike Kravetz * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in 2099f60858f9SMike Kravetz * a loop (to adjust global huge page counts) and previous allocation 2100f60858f9SMike Kravetz * failed, do not continue to try hard on the same node. Use the 2101f60858f9SMike Kravetz * node_alloc_noretry bitmap to manage this state information. 2102f60858f9SMike Kravetz */ 2103f60858f9SMike Kravetz if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) 2104f60858f9SMike Kravetz alloc_try_hard = false; 2105f60858f9SMike Kravetz gfp_mask |= __GFP_COMP|__GFP_NOWARN; 2106f60858f9SMike Kravetz if (alloc_try_hard) 2107f60858f9SMike Kravetz gfp_mask |= __GFP_RETRY_MAYFAIL; 2108af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 2109af0fb9dfSMichal Hocko nid = numa_mem_id(); 21102b21624fSMike Kravetz retry: 211184172f4bSMatthew Wilcox (Oracle) page = __alloc_pages(gfp_mask, order, nid, nmask); 21122b21624fSMike Kravetz 21132b21624fSMike Kravetz /* Freeze head page */ 21142b21624fSMike Kravetz if (page && !page_ref_freeze(page, 1)) { 21152b21624fSMike Kravetz __free_pages(page, order); 21162b21624fSMike Kravetz if (retry) { /* retry once */ 21172b21624fSMike Kravetz retry = false; 21182b21624fSMike Kravetz goto retry; 21192b21624fSMike Kravetz } 21202b21624fSMike Kravetz /* WOW! twice in a row. */ 21212b21624fSMike Kravetz pr_warn("HugeTLB head page unexpected inflated ref count\n"); 21222b21624fSMike Kravetz page = NULL; 21232b21624fSMike Kravetz } 21242b21624fSMike Kravetz 2125f60858f9SMike Kravetz /* 2126f60858f9SMike Kravetz * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this 2127f60858f9SMike Kravetz * indicates an overall state change. Clear bit so that we resume 2128f60858f9SMike Kravetz * normal 'try hard' allocations. 2129f60858f9SMike Kravetz */ 2130f60858f9SMike Kravetz if (node_alloc_noretry && page && !alloc_try_hard) 2131f60858f9SMike Kravetz node_clear(nid, *node_alloc_noretry); 2132f60858f9SMike Kravetz 2133f60858f9SMike Kravetz /* 2134f60858f9SMike Kravetz * If we tried hard to get a page but failed, set bit so that 2135f60858f9SMike Kravetz * subsequent attempts will not try as hard until there is an 2136f60858f9SMike Kravetz * overall state change. 2137f60858f9SMike Kravetz */ 2138f60858f9SMike Kravetz if (node_alloc_noretry && !page && alloc_try_hard) 2139f60858f9SMike Kravetz node_set(nid, *node_alloc_noretry); 2140f60858f9SMike Kravetz 214119fc1a7eSSidhartha Kumar if (!page) { 214219fc1a7eSSidhartha Kumar __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 214319fc1a7eSSidhartha Kumar return NULL; 214419fc1a7eSSidhartha Kumar } 214519fc1a7eSSidhartha Kumar 214619fc1a7eSSidhartha Kumar __count_vm_event(HTLB_BUDDY_PGALLOC); 214719fc1a7eSSidhartha Kumar return page_folio(page); 214863b4613cSNishanth Aravamudan } 214963b4613cSNishanth Aravamudan 2150af0fb9dfSMichal Hocko /* 21510c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 21520c397daeSMichal Hocko * should use this function to get new hugetlb pages 21532b21624fSMike Kravetz * 21542b21624fSMike Kravetz * Note that returned page is 'frozen': ref count of head page and all tail 21552b21624fSMike Kravetz * pages is zero. 21560c397daeSMichal Hocko */ 215719fc1a7eSSidhartha Kumar static struct folio *alloc_fresh_hugetlb_folio(struct hstate *h, 2158f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 2159f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 21600c397daeSMichal Hocko { 21617f325a8dSSidhartha Kumar struct folio *folio; 21627118fc29SMike Kravetz bool retry = false; 21630c397daeSMichal Hocko 21647118fc29SMike Kravetz retry: 21650c397daeSMichal Hocko if (hstate_is_gigantic(h)) 216619fc1a7eSSidhartha Kumar folio = alloc_gigantic_folio(h, gfp_mask, nid, nmask); 21670c397daeSMichal Hocko else 216819fc1a7eSSidhartha Kumar folio = alloc_buddy_hugetlb_folio(h, gfp_mask, 2169f60858f9SMike Kravetz nid, nmask, node_alloc_noretry); 217019fc1a7eSSidhartha Kumar if (!folio) 21710c397daeSMichal Hocko return NULL; 21727118fc29SMike Kravetz if (hstate_is_gigantic(h)) { 2173d1c60955SSidhartha Kumar if (!prep_compound_gigantic_folio(folio, huge_page_order(h))) { 21747118fc29SMike Kravetz /* 21757118fc29SMike Kravetz * Rare failure to convert pages to compound page. 21767118fc29SMike Kravetz * Free pages and try again - ONCE! 21777118fc29SMike Kravetz */ 21787f325a8dSSidhartha Kumar free_gigantic_folio(folio, huge_page_order(h)); 21797118fc29SMike Kravetz if (!retry) { 21807118fc29SMike Kravetz retry = true; 21817118fc29SMike Kravetz goto retry; 21827118fc29SMike Kravetz } 21837118fc29SMike Kravetz return NULL; 21847118fc29SMike Kravetz } 21857118fc29SMike Kravetz } 2186d1c60955SSidhartha Kumar prep_new_hugetlb_folio(h, folio, folio_nid(folio)); 21870c397daeSMichal Hocko 218819fc1a7eSSidhartha Kumar return folio; 21890c397daeSMichal Hocko } 21900c397daeSMichal Hocko 21910c397daeSMichal Hocko /* 2192af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 2193af0fb9dfSMichal Hocko * manner. 2194af0fb9dfSMichal Hocko */ 2195f60858f9SMike Kravetz static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 2196f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 2197b2261026SJoonsoo Kim { 219819fc1a7eSSidhartha Kumar struct folio *folio; 2199b2261026SJoonsoo Kim int nr_nodes, node; 2200af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2201b2261026SJoonsoo Kim 2202b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 220319fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, node, 220419fc1a7eSSidhartha Kumar nodes_allowed, node_alloc_noretry); 220519fc1a7eSSidhartha Kumar if (folio) { 2206454a00c4SMatthew Wilcox (Oracle) free_huge_folio(folio); /* free it into the hugepage allocator */ 220719fc1a7eSSidhartha Kumar return 1; 220819fc1a7eSSidhartha Kumar } 2209b2261026SJoonsoo Kim } 2210b2261026SJoonsoo Kim 2211af0fb9dfSMichal Hocko return 0; 2212b2261026SJoonsoo Kim } 2213b2261026SJoonsoo Kim 2214e8c5c824SLee Schermerhorn /* 221510c6ec49SMike Kravetz * Remove huge page from pool from next node to free. Attempt to keep 221610c6ec49SMike Kravetz * persistent huge pages more or less balanced over allowed nodes. 221710c6ec49SMike Kravetz * This routine only 'removes' the hugetlb page. The caller must make 221810c6ec49SMike Kravetz * an additional call to free the page to low level allocators. 2219e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 2220e8c5c824SLee Schermerhorn */ 222110c6ec49SMike Kravetz static struct page *remove_pool_huge_page(struct hstate *h, 222210c6ec49SMike Kravetz nodemask_t *nodes_allowed, 22236ae11b27SLee Schermerhorn bool acct_surplus) 2224e8c5c824SLee Schermerhorn { 2225b2261026SJoonsoo Kim int nr_nodes, node; 222610c6ec49SMike Kravetz struct page *page = NULL; 2227cfd5082bSSidhartha Kumar struct folio *folio; 2228e8c5c824SLee Schermerhorn 22299487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2230b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2231685f3457SLee Schermerhorn /* 2232685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 2233685f3457SLee Schermerhorn * nodes with surplus pages. 2234685f3457SLee Schermerhorn */ 2235b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 2236b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 223710c6ec49SMike Kravetz page = list_entry(h->hugepage_freelists[node].next, 2238e8c5c824SLee Schermerhorn struct page, lru); 2239cfd5082bSSidhartha Kumar folio = page_folio(page); 2240cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, acct_surplus); 22419a76db09SLee Schermerhorn break; 2242e8c5c824SLee Schermerhorn } 2243b2261026SJoonsoo Kim } 2244e8c5c824SLee Schermerhorn 224510c6ec49SMike Kravetz return page; 2246e8c5c824SLee Schermerhorn } 2247e8c5c824SLee Schermerhorn 2248c8721bbbSNaoya Horiguchi /* 2249c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 2250faf53defSNaoya Horiguchi * nothing for in-use hugepages and non-hugepages. 2251faf53defSNaoya Horiguchi * This function returns values like below: 2252faf53defSNaoya Horiguchi * 2253ad2fa371SMuchun Song * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages 2254ad2fa371SMuchun Song * when the system is under memory pressure and the feature of 2255ad2fa371SMuchun Song * freeing unused vmemmap pages associated with each hugetlb page 2256ad2fa371SMuchun Song * is enabled. 2257faf53defSNaoya Horiguchi * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 2258faf53defSNaoya Horiguchi * (allocated or reserved.) 2259faf53defSNaoya Horiguchi * 0: successfully dissolved free hugepages or the page is not a 2260faf53defSNaoya Horiguchi * hugepage (considered as already dissolved) 2261c8721bbbSNaoya Horiguchi */ 2262c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 2263c8721bbbSNaoya Horiguchi { 22646bc9b564SNaoya Horiguchi int rc = -EBUSY; 22651a7cdab5SSidhartha Kumar struct folio *folio = page_folio(page); 2266082d5b6bSGerald Schaefer 22677ffddd49SMuchun Song retry: 2268faf53defSNaoya Horiguchi /* Not to disrupt normal path by vainly holding hugetlb_lock */ 22691a7cdab5SSidhartha Kumar if (!folio_test_hugetlb(folio)) 2270faf53defSNaoya Horiguchi return 0; 2271faf53defSNaoya Horiguchi 2272db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 22731a7cdab5SSidhartha Kumar if (!folio_test_hugetlb(folio)) { 2274faf53defSNaoya Horiguchi rc = 0; 2275faf53defSNaoya Horiguchi goto out; 2276faf53defSNaoya Horiguchi } 2277faf53defSNaoya Horiguchi 22781a7cdab5SSidhartha Kumar if (!folio_ref_count(folio)) { 22791a7cdab5SSidhartha Kumar struct hstate *h = folio_hstate(folio); 22808346d69dSXin Hao if (!available_huge_pages(h)) 2281082d5b6bSGerald Schaefer goto out; 22827ffddd49SMuchun Song 22837ffddd49SMuchun Song /* 22847ffddd49SMuchun Song * We should make sure that the page is already on the free list 22857ffddd49SMuchun Song * when it is dissolved. 22867ffddd49SMuchun Song */ 22871a7cdab5SSidhartha Kumar if (unlikely(!folio_test_hugetlb_freed(folio))) { 2288db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 22897ffddd49SMuchun Song cond_resched(); 22907ffddd49SMuchun Song 22917ffddd49SMuchun Song /* 22927ffddd49SMuchun Song * Theoretically, we should return -EBUSY when we 22937ffddd49SMuchun Song * encounter this race. In fact, we have a chance 22947ffddd49SMuchun Song * to successfully dissolve the page if we do a 22957ffddd49SMuchun Song * retry. Because the race window is quite small. 22967ffddd49SMuchun Song * If we seize this opportunity, it is an optimization 22977ffddd49SMuchun Song * for increasing the success rate of dissolving page. 22987ffddd49SMuchun Song */ 22997ffddd49SMuchun Song goto retry; 23007ffddd49SMuchun Song } 23017ffddd49SMuchun Song 2302cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, false); 2303ad2fa371SMuchun Song h->max_huge_pages--; 2304ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2305ad2fa371SMuchun Song 2306c3114a84SAnshuman Khandual /* 2307d6ef19e2SSidhartha Kumar * Normally update_and_free_hugtlb_folio will allocate required vmemmmap 2308d6ef19e2SSidhartha Kumar * before freeing the page. update_and_free_hugtlb_folio will fail to 2309ad2fa371SMuchun Song * free the page if it can not allocate required vmemmap. We 2310ad2fa371SMuchun Song * need to adjust max_huge_pages if the page is not freed. 2311ad2fa371SMuchun Song * Attempt to allocate vmemmmap here so that we can take 2312ad2fa371SMuchun Song * appropriate action on failure. 2313ad2fa371SMuchun Song */ 23141a7cdab5SSidhartha Kumar rc = hugetlb_vmemmap_restore(h, &folio->page); 2315ad2fa371SMuchun Song if (!rc) { 2316d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, folio, false); 2317ad2fa371SMuchun Song } else { 2318ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 23192f6c57d6SSidhartha Kumar add_hugetlb_folio(h, folio, false); 2320ad2fa371SMuchun Song h->max_huge_pages++; 2321ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2322ad2fa371SMuchun Song } 2323ad2fa371SMuchun Song 2324ad2fa371SMuchun Song return rc; 2325c8721bbbSNaoya Horiguchi } 2326082d5b6bSGerald Schaefer out: 2327db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2328082d5b6bSGerald Schaefer return rc; 2329c8721bbbSNaoya Horiguchi } 2330c8721bbbSNaoya Horiguchi 2331c8721bbbSNaoya Horiguchi /* 2332c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 2333c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 23342247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 23352247bb33SGerald Schaefer * part of it lies within the given range. 2336082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 2337082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 2338c8721bbbSNaoya Horiguchi */ 2339082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 2340c8721bbbSNaoya Horiguchi { 2341c8721bbbSNaoya Horiguchi unsigned long pfn; 2342eb03aa00SGerald Schaefer struct page *page; 2343082d5b6bSGerald Schaefer int rc = 0; 2344dc2628f3SMuchun Song unsigned int order; 2345dc2628f3SMuchun Song struct hstate *h; 2346c8721bbbSNaoya Horiguchi 2347d0177639SLi Zhong if (!hugepages_supported()) 2348082d5b6bSGerald Schaefer return rc; 2349d0177639SLi Zhong 2350dc2628f3SMuchun Song order = huge_page_order(&default_hstate); 2351dc2628f3SMuchun Song for_each_hstate(h) 2352dc2628f3SMuchun Song order = min(order, huge_page_order(h)); 2353dc2628f3SMuchun Song 2354dc2628f3SMuchun Song for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) { 2355eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 2356eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 2357eb03aa00SGerald Schaefer if (rc) 2358082d5b6bSGerald Schaefer break; 2359eb03aa00SGerald Schaefer } 2360082d5b6bSGerald Schaefer 2361082d5b6bSGerald Schaefer return rc; 2362c8721bbbSNaoya Horiguchi } 2363c8721bbbSNaoya Horiguchi 2364ab5ac90aSMichal Hocko /* 2365ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 2366ab5ac90aSMichal Hocko */ 23673a740e8bSSidhartha Kumar static struct folio *alloc_surplus_hugetlb_folio(struct hstate *h, 23683a740e8bSSidhartha Kumar gfp_t gfp_mask, int nid, nodemask_t *nmask) 23697893d1d5SAdam Litke { 237019fc1a7eSSidhartha Kumar struct folio *folio = NULL; 23717893d1d5SAdam Litke 2372bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2373aa888a74SAndi Kleen return NULL; 2374aa888a74SAndi Kleen 2375db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 23769980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 23779980d744SMichal Hocko goto out_unlock; 2378db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2379d1c3fb1fSNishanth Aravamudan 238019fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL); 238119fc1a7eSSidhartha Kumar if (!folio) 23820c397daeSMichal Hocko return NULL; 2383d1c3fb1fSNishanth Aravamudan 2384db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 23859980d744SMichal Hocko /* 23869980d744SMichal Hocko * We could have raced with the pool size change. 23879980d744SMichal Hocko * Double check that and simply deallocate the new page 23889980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 2389454a00c4SMatthew Wilcox (Oracle) * temporary page to workaround the nasty free_huge_folio 23909980d744SMichal Hocko * codeflow 23919980d744SMichal Hocko */ 23929980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 239319fc1a7eSSidhartha Kumar folio_set_hugetlb_temporary(folio); 2394db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2395454a00c4SMatthew Wilcox (Oracle) free_huge_folio(folio); 23962bf753e6SKai Shen return NULL; 2397b65a4edaSMike Kravetz } 2398b65a4edaSMike Kravetz 23999980d744SMichal Hocko h->surplus_huge_pages++; 240019fc1a7eSSidhartha Kumar h->surplus_huge_pages_node[folio_nid(folio)]++; 24019980d744SMichal Hocko 24029980d744SMichal Hocko out_unlock: 2403db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 24047893d1d5SAdam Litke 24053a740e8bSSidhartha Kumar return folio; 24067893d1d5SAdam Litke } 24077893d1d5SAdam Litke 2408e37d3e83SSidhartha Kumar static struct folio *alloc_migrate_hugetlb_folio(struct hstate *h, gfp_t gfp_mask, 2409ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 2410ab5ac90aSMichal Hocko { 241119fc1a7eSSidhartha Kumar struct folio *folio; 2412ab5ac90aSMichal Hocko 2413ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 2414ab5ac90aSMichal Hocko return NULL; 2415ab5ac90aSMichal Hocko 241619fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL); 241719fc1a7eSSidhartha Kumar if (!folio) 2418ab5ac90aSMichal Hocko return NULL; 2419ab5ac90aSMichal Hocko 24202b21624fSMike Kravetz /* fresh huge pages are frozen */ 242119fc1a7eSSidhartha Kumar folio_ref_unfreeze(folio, 1); 2422ab5ac90aSMichal Hocko /* 2423ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 2424ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 2425ab5ac90aSMichal Hocko */ 242619fc1a7eSSidhartha Kumar folio_set_hugetlb_temporary(folio); 2427ab5ac90aSMichal Hocko 2428e37d3e83SSidhartha Kumar return folio; 2429ab5ac90aSMichal Hocko } 2430ab5ac90aSMichal Hocko 2431e4e574b7SAdam Litke /* 2432099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 2433099730d6SDave Hansen */ 2434e0ec90eeSDave Hansen static 2435ff7d853bSSidhartha Kumar struct folio *alloc_buddy_hugetlb_folio_with_mpol(struct hstate *h, 2436099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 2437099730d6SDave Hansen { 24383a740e8bSSidhartha Kumar struct folio *folio = NULL; 2439aaf14e40SMichal Hocko struct mempolicy *mpol; 2440aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 2441aaf14e40SMichal Hocko int nid; 2442aaf14e40SMichal Hocko nodemask_t *nodemask; 2443aaf14e40SMichal Hocko 2444aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 2445cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 2446cfcaa66fSBen Widawsky gfp_t gfp = gfp_mask | __GFP_NOWARN; 2447cfcaa66fSBen Widawsky 2448cfcaa66fSBen Widawsky gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL); 24493a740e8bSSidhartha Kumar folio = alloc_surplus_hugetlb_folio(h, gfp, nid, nodemask); 2450cfcaa66fSBen Widawsky 2451cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 2452cfcaa66fSBen Widawsky nodemask = NULL; 2453cfcaa66fSBen Widawsky } 2454cfcaa66fSBen Widawsky 24553a740e8bSSidhartha Kumar if (!folio) 24563a740e8bSSidhartha Kumar folio = alloc_surplus_hugetlb_folio(h, gfp_mask, nid, nodemask); 2457aaf14e40SMichal Hocko mpol_cond_put(mpol); 2458ff7d853bSSidhartha Kumar return folio; 2459099730d6SDave Hansen } 2460099730d6SDave Hansen 2461e37d3e83SSidhartha Kumar /* folio migration callback function */ 2462e37d3e83SSidhartha Kumar struct folio *alloc_hugetlb_folio_nodemask(struct hstate *h, int preferred_nid, 2463d92bbc27SJoonsoo Kim nodemask_t *nmask, gfp_t gfp_mask) 24644db9b2efSMichal Hocko { 2465db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 24668346d69dSXin Hao if (available_huge_pages(h)) { 2467a36f1e90SSidhartha Kumar struct folio *folio; 24683e59fcb0SMichal Hocko 2469a36f1e90SSidhartha Kumar folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask, 2470a36f1e90SSidhartha Kumar preferred_nid, nmask); 2471a36f1e90SSidhartha Kumar if (folio) { 2472db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2473e37d3e83SSidhartha Kumar return folio; 24744db9b2efSMichal Hocko } 24754db9b2efSMichal Hocko } 2476db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 24774db9b2efSMichal Hocko 2478e37d3e83SSidhartha Kumar return alloc_migrate_hugetlb_folio(h, gfp_mask, preferred_nid, nmask); 24794db9b2efSMichal Hocko } 24804db9b2efSMichal Hocko 2481ebd63723SMichal Hocko /* mempolicy aware migration callback */ 2482d0ce0e47SSidhartha Kumar struct folio *alloc_hugetlb_folio_vma(struct hstate *h, struct vm_area_struct *vma, 2483389c8178SMichal Hocko unsigned long address) 2484ebd63723SMichal Hocko { 2485ebd63723SMichal Hocko struct mempolicy *mpol; 2486ebd63723SMichal Hocko nodemask_t *nodemask; 2487e37d3e83SSidhartha Kumar struct folio *folio; 2488ebd63723SMichal Hocko gfp_t gfp_mask; 2489ebd63723SMichal Hocko int node; 2490ebd63723SMichal Hocko 2491ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 2492ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 2493e37d3e83SSidhartha Kumar folio = alloc_hugetlb_folio_nodemask(h, node, nodemask, gfp_mask); 2494ebd63723SMichal Hocko mpol_cond_put(mpol); 2495ebd63723SMichal Hocko 2496d0ce0e47SSidhartha Kumar return folio; 2497ebd63723SMichal Hocko } 2498ebd63723SMichal Hocko 2499bf50bab2SNaoya Horiguchi /* 250025985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 2501e4e574b7SAdam Litke * of size 'delta'. 2502e4e574b7SAdam Litke */ 25030a4f3d1bSLiu Xiang static int gather_surplus_pages(struct hstate *h, long delta) 25041b2a1e7bSJules Irenge __must_hold(&hugetlb_lock) 2505e4e574b7SAdam Litke { 250634665341SMiaohe Lin LIST_HEAD(surplus_list); 2507454a00c4SMatthew Wilcox (Oracle) struct folio *folio, *tmp; 25080a4f3d1bSLiu Xiang int ret; 25090a4f3d1bSLiu Xiang long i; 25100a4f3d1bSLiu Xiang long needed, allocated; 251128073b02SHillf Danton bool alloc_ok = true; 2512e4e574b7SAdam Litke 25139487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2514a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 2515ac09b3a1SAdam Litke if (needed <= 0) { 2516a5516438SAndi Kleen h->resv_huge_pages += delta; 2517e4e574b7SAdam Litke return 0; 2518ac09b3a1SAdam Litke } 2519e4e574b7SAdam Litke 2520e4e574b7SAdam Litke allocated = 0; 2521e4e574b7SAdam Litke 2522e4e574b7SAdam Litke ret = -ENOMEM; 2523e4e574b7SAdam Litke retry: 2524db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2525e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 25263a740e8bSSidhartha Kumar folio = alloc_surplus_hugetlb_folio(h, htlb_alloc_mask(h), 25272b21624fSMike Kravetz NUMA_NO_NODE, NULL); 25283a740e8bSSidhartha Kumar if (!folio) { 252928073b02SHillf Danton alloc_ok = false; 253028073b02SHillf Danton break; 253128073b02SHillf Danton } 25323a740e8bSSidhartha Kumar list_add(&folio->lru, &surplus_list); 253369ed779aSDavid Rientjes cond_resched(); 2534e4e574b7SAdam Litke } 253528073b02SHillf Danton allocated += i; 2536e4e574b7SAdam Litke 2537e4e574b7SAdam Litke /* 2538e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 2539e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 2540e4e574b7SAdam Litke */ 2541db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2542a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 2543a5516438SAndi Kleen (h->free_huge_pages + allocated); 254428073b02SHillf Danton if (needed > 0) { 254528073b02SHillf Danton if (alloc_ok) 2546e4e574b7SAdam Litke goto retry; 254728073b02SHillf Danton /* 254828073b02SHillf Danton * We were not able to allocate enough pages to 254928073b02SHillf Danton * satisfy the entire reservation so we free what 255028073b02SHillf Danton * we've allocated so far. 255128073b02SHillf Danton */ 255228073b02SHillf Danton goto free; 255328073b02SHillf Danton } 2554e4e574b7SAdam Litke /* 2555e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 255625985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 2557e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 2558ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 2559ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 2560ac09b3a1SAdam Litke * before they are reserved. 2561e4e574b7SAdam Litke */ 2562e4e574b7SAdam Litke needed += allocated; 2563a5516438SAndi Kleen h->resv_huge_pages += delta; 2564e4e574b7SAdam Litke ret = 0; 2565a9869b83SNaoya Horiguchi 256619fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 2567454a00c4SMatthew Wilcox (Oracle) list_for_each_entry_safe(folio, tmp, &surplus_list, lru) { 256819fc3f0aSAdam Litke if ((--needed) < 0) 256919fc3f0aSAdam Litke break; 2570b65a4edaSMike Kravetz /* Add the page to the hugetlb allocator */ 2571454a00c4SMatthew Wilcox (Oracle) enqueue_hugetlb_folio(h, folio); 257219fc3f0aSAdam Litke } 257328073b02SHillf Danton free: 2574db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 257519fc3f0aSAdam Litke 2576b65a4edaSMike Kravetz /* 2577b65a4edaSMike Kravetz * Free unnecessary surplus pages to the buddy allocator. 2578454a00c4SMatthew Wilcox (Oracle) * Pages have no ref count, call free_huge_folio directly. 2579b65a4edaSMike Kravetz */ 2580454a00c4SMatthew Wilcox (Oracle) list_for_each_entry_safe(folio, tmp, &surplus_list, lru) 2581454a00c4SMatthew Wilcox (Oracle) free_huge_folio(folio); 2582db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2583e4e574b7SAdam Litke 2584e4e574b7SAdam Litke return ret; 2585e4e574b7SAdam Litke } 2586e4e574b7SAdam Litke 2587e4e574b7SAdam Litke /* 2588e5bbc8a6SMike Kravetz * This routine has two main purposes: 2589e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 2590e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 2591e5bbc8a6SMike Kravetz * to the associated reservation map. 2592e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 2593e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 2594e4e574b7SAdam Litke */ 2595a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 2596a5516438SAndi Kleen unsigned long unused_resv_pages) 2597e4e574b7SAdam Litke { 2598e4e574b7SAdam Litke unsigned long nr_pages; 259910c6ec49SMike Kravetz struct page *page; 260010c6ec49SMike Kravetz LIST_HEAD(page_list); 260110c6ec49SMike Kravetz 26029487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 260310c6ec49SMike Kravetz /* Uncommit the reservation */ 260410c6ec49SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 2605e4e574b7SAdam Litke 2606c0531714SNaoya Horiguchi if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 2607e5bbc8a6SMike Kravetz goto out; 2608aa888a74SAndi Kleen 2609e5bbc8a6SMike Kravetz /* 2610e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 2611e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 2612e5bbc8a6SMike Kravetz */ 2613a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 2614e4e574b7SAdam Litke 2615685f3457SLee Schermerhorn /* 2616685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 26179b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 26189b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 26199b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 262010c6ec49SMike Kravetz * remove_pool_huge_page() will balance the freed pages across the 26219b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 2622685f3457SLee Schermerhorn */ 2623685f3457SLee Schermerhorn while (nr_pages--) { 262410c6ec49SMike Kravetz page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1); 262510c6ec49SMike Kravetz if (!page) 2626e5bbc8a6SMike Kravetz goto out; 262710c6ec49SMike Kravetz 262810c6ec49SMike Kravetz list_add(&page->lru, &page_list); 2629e4e574b7SAdam Litke } 2630e5bbc8a6SMike Kravetz 2631e5bbc8a6SMike Kravetz out: 2632db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 263310c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 2634db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2635e4e574b7SAdam Litke } 2636e4e574b7SAdam Litke 26375e911373SMike Kravetz 2638c37f9fb1SAndy Whitcroft /* 2639feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 26405e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 2641cf3ad20bSMike Kravetz * 2642cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 2643cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 2644cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 2645cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 2646cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 2647feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 2648feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 2649feba16e2SMike Kravetz * is called in such cases. 2650cf3ad20bSMike Kravetz * 2651cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 2652cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 2653cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 2654cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 2655cf3ad20bSMike Kravetz * take appropriate action. 265696b96a96SMike Kravetz * 265796b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 265896b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 265996b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 266096b96a96SMike Kravetz * reservation exists. 2661846be085SMike Kravetz * 2662846be085SMike Kravetz * vma_del_reservation is used in error paths where an entry in the reserve 2663846be085SMike Kravetz * map was created during huge page allocation and must be removed. It is to 2664846be085SMike Kravetz * be called after calling vma_needs_reservation to determine if a reservation 2665846be085SMike Kravetz * exists. 2666c37f9fb1SAndy Whitcroft */ 26675e911373SMike Kravetz enum vma_resv_mode { 26685e911373SMike Kravetz VMA_NEEDS_RESV, 26695e911373SMike Kravetz VMA_COMMIT_RESV, 2670feba16e2SMike Kravetz VMA_END_RESV, 267196b96a96SMike Kravetz VMA_ADD_RESV, 2672846be085SMike Kravetz VMA_DEL_RESV, 26735e911373SMike Kravetz }; 2674cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 2675cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 26765e911373SMike Kravetz enum vma_resv_mode mode) 2677c37f9fb1SAndy Whitcroft { 26784e35f483SJoonsoo Kim struct resv_map *resv; 26794e35f483SJoonsoo Kim pgoff_t idx; 2680cf3ad20bSMike Kravetz long ret; 26810db9d74eSMina Almasry long dummy_out_regions_needed; 2682c37f9fb1SAndy Whitcroft 26834e35f483SJoonsoo Kim resv = vma_resv_map(vma); 26844e35f483SJoonsoo Kim if (!resv) 2685c37f9fb1SAndy Whitcroft return 1; 2686c37f9fb1SAndy Whitcroft 26874e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 26885e911373SMike Kravetz switch (mode) { 26895e911373SMike Kravetz case VMA_NEEDS_RESV: 26900db9d74eSMina Almasry ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); 26910db9d74eSMina Almasry /* We assume that vma_reservation_* routines always operate on 26920db9d74eSMina Almasry * 1 page, and that adding to resv map a 1 page entry can only 26930db9d74eSMina Almasry * ever require 1 region. 26940db9d74eSMina Almasry */ 26950db9d74eSMina Almasry VM_BUG_ON(dummy_out_regions_needed != 1); 26965e911373SMike Kravetz break; 26975e911373SMike Kravetz case VMA_COMMIT_RESV: 2698075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 26990db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 27000db9d74eSMina Almasry VM_BUG_ON(ret < 0); 27015e911373SMike Kravetz break; 2702feba16e2SMike Kravetz case VMA_END_RESV: 27030db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 27045e911373SMike Kravetz ret = 0; 27055e911373SMike Kravetz break; 270696b96a96SMike Kravetz case VMA_ADD_RESV: 27070db9d74eSMina Almasry if (vma->vm_flags & VM_MAYSHARE) { 2708075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 27090db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 27100db9d74eSMina Almasry VM_BUG_ON(ret < 0); 27110db9d74eSMina Almasry } else { 27120db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 271396b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 271496b96a96SMike Kravetz } 271596b96a96SMike Kravetz break; 2716846be085SMike Kravetz case VMA_DEL_RESV: 2717846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 2718846be085SMike Kravetz region_abort(resv, idx, idx + 1, 1); 2719846be085SMike Kravetz ret = region_del(resv, idx, idx + 1); 2720846be085SMike Kravetz } else { 2721846be085SMike Kravetz ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 2722846be085SMike Kravetz /* region_add calls of range 1 should never fail. */ 2723846be085SMike Kravetz VM_BUG_ON(ret < 0); 2724846be085SMike Kravetz } 2725846be085SMike Kravetz break; 27265e911373SMike Kravetz default: 27275e911373SMike Kravetz BUG(); 27285e911373SMike Kravetz } 272984afd99bSAndy Whitcroft 2730846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV) 2731cf3ad20bSMike Kravetz return ret; 273267961f9dSMike Kravetz /* 2733bf3d12b9SMiaohe Lin * We know private mapping must have HPAGE_RESV_OWNER set. 2734bf3d12b9SMiaohe Lin * 273567961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 273667961f9dSMike Kravetz * However, a file associated with mapping could have been 273767961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 273867961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 273967961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 274067961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 274167961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 274267961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 274367961f9dSMike Kravetz * reservation has already been consumed. As a result, the 274467961f9dSMike Kravetz * return value of this routine is the opposite of the 274567961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 274667961f9dSMike Kravetz */ 2747bf3d12b9SMiaohe Lin if (ret > 0) 274867961f9dSMike Kravetz return 0; 2749bf3d12b9SMiaohe Lin if (ret == 0) 275067961f9dSMike Kravetz return 1; 2751bf3d12b9SMiaohe Lin return ret; 275284afd99bSAndy Whitcroft } 2753cf3ad20bSMike Kravetz 2754cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 2755a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 2756c37f9fb1SAndy Whitcroft { 27575e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 2758cf3ad20bSMike Kravetz } 2759c37f9fb1SAndy Whitcroft 2760cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 2761cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 2762cf3ad20bSMike Kravetz { 27635e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 27645e911373SMike Kravetz } 27655e911373SMike Kravetz 2766feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 27675e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 27685e911373SMike Kravetz { 2769feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 2770c37f9fb1SAndy Whitcroft } 2771c37f9fb1SAndy Whitcroft 277296b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 277396b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 277496b96a96SMike Kravetz { 277596b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 277696b96a96SMike Kravetz } 277796b96a96SMike Kravetz 2778846be085SMike Kravetz static long vma_del_reservation(struct hstate *h, 2779846be085SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 278096b96a96SMike Kravetz { 2781846be085SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV); 2782846be085SMike Kravetz } 2783846be085SMike Kravetz 2784846be085SMike Kravetz /* 2785846be085SMike Kravetz * This routine is called to restore reservation information on error paths. 2786d0ce0e47SSidhartha Kumar * It should ONLY be called for folios allocated via alloc_hugetlb_folio(), 2787d0ce0e47SSidhartha Kumar * and the hugetlb mutex should remain held when calling this routine. 2788846be085SMike Kravetz * 2789846be085SMike Kravetz * It handles two specific cases: 2790d2d7bb44SSidhartha Kumar * 1) A reservation was in place and the folio consumed the reservation. 2791d2d7bb44SSidhartha Kumar * hugetlb_restore_reserve is set in the folio. 2792d2d7bb44SSidhartha Kumar * 2) No reservation was in place for the page, so hugetlb_restore_reserve is 2793d0ce0e47SSidhartha Kumar * not set. However, alloc_hugetlb_folio always updates the reserve map. 2794846be085SMike Kravetz * 2795454a00c4SMatthew Wilcox (Oracle) * In case 1, free_huge_folio later in the error path will increment the 2796454a00c4SMatthew Wilcox (Oracle) * global reserve count. But, free_huge_folio does not have enough context 2797846be085SMike Kravetz * to adjust the reservation map. This case deals primarily with private 2798846be085SMike Kravetz * mappings. Adjust the reserve map here to be consistent with global 2799454a00c4SMatthew Wilcox (Oracle) * reserve count adjustments to be made by free_huge_folio. Make sure the 2800846be085SMike Kravetz * reserve map indicates there is a reservation present. 2801846be085SMike Kravetz * 2802d0ce0e47SSidhartha Kumar * In case 2, simply undo reserve map modifications done by alloc_hugetlb_folio. 2803846be085SMike Kravetz */ 2804846be085SMike Kravetz void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma, 2805d2d7bb44SSidhartha Kumar unsigned long address, struct folio *folio) 2806846be085SMike Kravetz { 280796b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 280896b96a96SMike Kravetz 28090ffdc38eSSidhartha Kumar if (folio_test_hugetlb_restore_reserve(folio)) { 2810846be085SMike Kravetz if (unlikely(rc < 0)) 281196b96a96SMike Kravetz /* 281296b96a96SMike Kravetz * Rare out of memory condition in reserve map 28130ffdc38eSSidhartha Kumar * manipulation. Clear hugetlb_restore_reserve so 28140ffdc38eSSidhartha Kumar * that global reserve count will not be incremented 2815454a00c4SMatthew Wilcox (Oracle) * by free_huge_folio. This will make it appear 28160ffdc38eSSidhartha Kumar * as though the reservation for this folio was 281796b96a96SMike Kravetz * consumed. This may prevent the task from 28180ffdc38eSSidhartha Kumar * faulting in the folio at a later time. This 281996b96a96SMike Kravetz * is better than inconsistent global huge page 282096b96a96SMike Kravetz * accounting of reserve counts. 282196b96a96SMike Kravetz */ 28220ffdc38eSSidhartha Kumar folio_clear_hugetlb_restore_reserve(folio); 2823846be085SMike Kravetz else if (rc) 2824846be085SMike Kravetz (void)vma_add_reservation(h, vma, address); 2825846be085SMike Kravetz else 2826846be085SMike Kravetz vma_end_reservation(h, vma, address); 2827846be085SMike Kravetz } else { 2828846be085SMike Kravetz if (!rc) { 282996b96a96SMike Kravetz /* 2830846be085SMike Kravetz * This indicates there is an entry in the reserve map 2831d0ce0e47SSidhartha Kumar * not added by alloc_hugetlb_folio. We know it was added 2832d0ce0e47SSidhartha Kumar * before the alloc_hugetlb_folio call, otherwise 28330ffdc38eSSidhartha Kumar * hugetlb_restore_reserve would be set on the folio. 2834846be085SMike Kravetz * Remove the entry so that a subsequent allocation 2835846be085SMike Kravetz * does not consume a reservation. 283696b96a96SMike Kravetz */ 2837846be085SMike Kravetz rc = vma_del_reservation(h, vma, address); 2838846be085SMike Kravetz if (rc < 0) 2839846be085SMike Kravetz /* 2840846be085SMike Kravetz * VERY rare out of memory condition. Since 2841846be085SMike Kravetz * we can not delete the entry, set 28420ffdc38eSSidhartha Kumar * hugetlb_restore_reserve so that the reserve 28430ffdc38eSSidhartha Kumar * count will be incremented when the folio 2844846be085SMike Kravetz * is freed. This reserve will be consumed 2845846be085SMike Kravetz * on a subsequent allocation. 2846846be085SMike Kravetz */ 28470ffdc38eSSidhartha Kumar folio_set_hugetlb_restore_reserve(folio); 2848846be085SMike Kravetz } else if (rc < 0) { 2849846be085SMike Kravetz /* 2850846be085SMike Kravetz * Rare out of memory condition from 2851846be085SMike Kravetz * vma_needs_reservation call. Memory allocation is 2852846be085SMike Kravetz * only attempted if a new entry is needed. Therefore, 2853846be085SMike Kravetz * this implies there is not an entry in the 2854846be085SMike Kravetz * reserve map. 2855846be085SMike Kravetz * 2856846be085SMike Kravetz * For shared mappings, no entry in the map indicates 2857846be085SMike Kravetz * no reservation. We are done. 2858846be085SMike Kravetz */ 2859846be085SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 2860846be085SMike Kravetz /* 2861846be085SMike Kravetz * For private mappings, no entry indicates 2862846be085SMike Kravetz * a reservation is present. Since we can 28630ffdc38eSSidhartha Kumar * not add an entry, set hugetlb_restore_reserve 28640ffdc38eSSidhartha Kumar * on the folio so reserve count will be 2865846be085SMike Kravetz * incremented when freed. This reserve will 2866846be085SMike Kravetz * be consumed on a subsequent allocation. 2867846be085SMike Kravetz */ 28680ffdc38eSSidhartha Kumar folio_set_hugetlb_restore_reserve(folio); 286996b96a96SMike Kravetz } else 2870846be085SMike Kravetz /* 2871846be085SMike Kravetz * No reservation present, do nothing 2872846be085SMike Kravetz */ 287396b96a96SMike Kravetz vma_end_reservation(h, vma, address); 287496b96a96SMike Kravetz } 287596b96a96SMike Kravetz } 287696b96a96SMike Kravetz 2877369fa227SOscar Salvador /* 287819fc1a7eSSidhartha Kumar * alloc_and_dissolve_hugetlb_folio - Allocate a new folio and dissolve 287919fc1a7eSSidhartha Kumar * the old one 2880369fa227SOscar Salvador * @h: struct hstate old page belongs to 288119fc1a7eSSidhartha Kumar * @old_folio: Old folio to dissolve 2882ae37c7ffSOscar Salvador * @list: List to isolate the page in case we need to 2883369fa227SOscar Salvador * Returns 0 on success, otherwise negated error. 2884369fa227SOscar Salvador */ 288519fc1a7eSSidhartha Kumar static int alloc_and_dissolve_hugetlb_folio(struct hstate *h, 288619fc1a7eSSidhartha Kumar struct folio *old_folio, struct list_head *list) 2887369fa227SOscar Salvador { 2888369fa227SOscar Salvador gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2889de656ed3SSidhartha Kumar int nid = folio_nid(old_folio); 2890de656ed3SSidhartha Kumar struct folio *new_folio; 2891369fa227SOscar Salvador int ret = 0; 2892369fa227SOscar Salvador 2893369fa227SOscar Salvador /* 289419fc1a7eSSidhartha Kumar * Before dissolving the folio, we need to allocate a new one for the 289519fc1a7eSSidhartha Kumar * pool to remain stable. Here, we allocate the folio and 'prep' it 2896f41f2ed4SMuchun Song * by doing everything but actually updating counters and adding to 2897f41f2ed4SMuchun Song * the pool. This simplifies and let us do most of the processing 2898f41f2ed4SMuchun Song * under the lock. 2899369fa227SOscar Salvador */ 290019fc1a7eSSidhartha Kumar new_folio = alloc_buddy_hugetlb_folio(h, gfp_mask, nid, NULL, NULL); 290119fc1a7eSSidhartha Kumar if (!new_folio) 2902369fa227SOscar Salvador return -ENOMEM; 2903de656ed3SSidhartha Kumar __prep_new_hugetlb_folio(h, new_folio); 2904369fa227SOscar Salvador 2905369fa227SOscar Salvador retry: 2906369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2907de656ed3SSidhartha Kumar if (!folio_test_hugetlb(old_folio)) { 2908369fa227SOscar Salvador /* 290919fc1a7eSSidhartha Kumar * Freed from under us. Drop new_folio too. 2910369fa227SOscar Salvador */ 2911369fa227SOscar Salvador goto free_new; 2912de656ed3SSidhartha Kumar } else if (folio_ref_count(old_folio)) { 29139747b9e9SBaolin Wang bool isolated; 29149747b9e9SBaolin Wang 2915369fa227SOscar Salvador /* 291619fc1a7eSSidhartha Kumar * Someone has grabbed the folio, try to isolate it here. 2917ae37c7ffSOscar Salvador * Fail with -EBUSY if not possible. 2918369fa227SOscar Salvador */ 2919ae37c7ffSOscar Salvador spin_unlock_irq(&hugetlb_lock); 29209747b9e9SBaolin Wang isolated = isolate_hugetlb(old_folio, list); 29219747b9e9SBaolin Wang ret = isolated ? 0 : -EBUSY; 2922ae37c7ffSOscar Salvador spin_lock_irq(&hugetlb_lock); 2923369fa227SOscar Salvador goto free_new; 2924de656ed3SSidhartha Kumar } else if (!folio_test_hugetlb_freed(old_folio)) { 2925369fa227SOscar Salvador /* 292619fc1a7eSSidhartha Kumar * Folio's refcount is 0 but it has not been enqueued in the 2927369fa227SOscar Salvador * freelist yet. Race window is small, so we can succeed here if 2928369fa227SOscar Salvador * we retry. 2929369fa227SOscar Salvador */ 2930369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2931369fa227SOscar Salvador cond_resched(); 2932369fa227SOscar Salvador goto retry; 2933369fa227SOscar Salvador } else { 2934369fa227SOscar Salvador /* 293519fc1a7eSSidhartha Kumar * Ok, old_folio is still a genuine free hugepage. Remove it from 2936369fa227SOscar Salvador * the freelist and decrease the counters. These will be 2937369fa227SOscar Salvador * incremented again when calling __prep_account_new_huge_page() 2938240d67a8SSidhartha Kumar * and enqueue_hugetlb_folio() for new_folio. The counters will 2939240d67a8SSidhartha Kumar * remain stable since this happens under the lock. 2940369fa227SOscar Salvador */ 2941cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, old_folio, false); 2942369fa227SOscar Salvador 2943369fa227SOscar Salvador /* 294419fc1a7eSSidhartha Kumar * Ref count on new_folio is already zero as it was dropped 2945b65a4edaSMike Kravetz * earlier. It can be directly added to the pool free list. 2946369fa227SOscar Salvador */ 2947369fa227SOscar Salvador __prep_account_new_huge_page(h, nid); 2948240d67a8SSidhartha Kumar enqueue_hugetlb_folio(h, new_folio); 2949369fa227SOscar Salvador 2950369fa227SOscar Salvador /* 295119fc1a7eSSidhartha Kumar * Folio has been replaced, we can safely free the old one. 2952369fa227SOscar Salvador */ 2953369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2954d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, old_folio, false); 2955369fa227SOscar Salvador } 2956369fa227SOscar Salvador 2957369fa227SOscar Salvador return ret; 2958369fa227SOscar Salvador 2959369fa227SOscar Salvador free_new: 2960369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 296119fc1a7eSSidhartha Kumar /* Folio has a zero ref count, but needs a ref to be freed */ 2962de656ed3SSidhartha Kumar folio_ref_unfreeze(new_folio, 1); 2963d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, new_folio, false); 2964369fa227SOscar Salvador 2965369fa227SOscar Salvador return ret; 2966369fa227SOscar Salvador } 2967369fa227SOscar Salvador 2968ae37c7ffSOscar Salvador int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list) 2969369fa227SOscar Salvador { 2970369fa227SOscar Salvador struct hstate *h; 2971d5e33bd8SSidhartha Kumar struct folio *folio = page_folio(page); 2972ae37c7ffSOscar Salvador int ret = -EBUSY; 2973369fa227SOscar Salvador 2974369fa227SOscar Salvador /* 2975369fa227SOscar Salvador * The page might have been dissolved from under our feet, so make sure 2976369fa227SOscar Salvador * to carefully check the state under the lock. 2977369fa227SOscar Salvador * Return success when racing as if we dissolved the page ourselves. 2978369fa227SOscar Salvador */ 2979369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2980d5e33bd8SSidhartha Kumar if (folio_test_hugetlb(folio)) { 2981d5e33bd8SSidhartha Kumar h = folio_hstate(folio); 2982369fa227SOscar Salvador } else { 2983369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2984369fa227SOscar Salvador return 0; 2985369fa227SOscar Salvador } 2986369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2987369fa227SOscar Salvador 2988369fa227SOscar Salvador /* 2989369fa227SOscar Salvador * Fence off gigantic pages as there is a cyclic dependency between 2990369fa227SOscar Salvador * alloc_contig_range and them. Return -ENOMEM as this has the effect 2991369fa227SOscar Salvador * of bailing out right away without further retrying. 2992369fa227SOscar Salvador */ 2993369fa227SOscar Salvador if (hstate_is_gigantic(h)) 2994369fa227SOscar Salvador return -ENOMEM; 2995369fa227SOscar Salvador 29969747b9e9SBaolin Wang if (folio_ref_count(folio) && isolate_hugetlb(folio, list)) 2997ae37c7ffSOscar Salvador ret = 0; 2998d5e33bd8SSidhartha Kumar else if (!folio_ref_count(folio)) 299919fc1a7eSSidhartha Kumar ret = alloc_and_dissolve_hugetlb_folio(h, folio, list); 3000ae37c7ffSOscar Salvador 3001ae37c7ffSOscar Salvador return ret; 3002369fa227SOscar Salvador } 3003369fa227SOscar Salvador 3004d0ce0e47SSidhartha Kumar struct folio *alloc_hugetlb_folio(struct vm_area_struct *vma, 300504f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 3006348ea204SAdam Litke { 300790481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 3008a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3009d4ab0316SSidhartha Kumar struct folio *folio; 3010d85f69b0SMike Kravetz long map_chg, map_commit; 3011d85f69b0SMike Kravetz long gbl_chg; 30126d76dcf4SAneesh Kumar K.V int ret, idx; 3013d0ce0e47SSidhartha Kumar struct hugetlb_cgroup *h_cg = NULL; 301408cf9fafSMina Almasry bool deferred_reserve; 30152fc39cecSAdam Litke 30166d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 3017a1e78772SMel Gorman /* 3018d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 3019d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 3020d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 3021a1e78772SMel Gorman */ 3022d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 3023d85f69b0SMike Kravetz if (map_chg < 0) 302476dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 3025d85f69b0SMike Kravetz 3026d85f69b0SMike Kravetz /* 3027d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 3028d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 3029d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 3030d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 3031d85f69b0SMike Kravetz * checked against any subpool limit. 3032d85f69b0SMike Kravetz */ 3033d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 3034d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 3035d85f69b0SMike Kravetz if (gbl_chg < 0) { 3036feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 303776dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 30385e911373SMike Kravetz } 303990d8b7e6SAdam Litke 3040d85f69b0SMike Kravetz /* 3041d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 3042d85f69b0SMike Kravetz * map, there could be reservations associated with the 3043d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 3044d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 3045d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 3046d85f69b0SMike Kravetz * the subpool reservations. 3047d85f69b0SMike Kravetz */ 3048d85f69b0SMike Kravetz if (avoid_reserve) 3049d85f69b0SMike Kravetz gbl_chg = 1; 3050d85f69b0SMike Kravetz } 3051d85f69b0SMike Kravetz 305208cf9fafSMina Almasry /* If this allocation is not consuming a reservation, charge it now. 305308cf9fafSMina Almasry */ 30546501fe5fSMiaohe Lin deferred_reserve = map_chg || avoid_reserve; 305508cf9fafSMina Almasry if (deferred_reserve) { 305608cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 305708cf9fafSMina Almasry idx, pages_per_huge_page(h), &h_cg); 30588f34af6fSJianyu Zhan if (ret) 30598f34af6fSJianyu Zhan goto out_subpool_put; 306008cf9fafSMina Almasry } 306108cf9fafSMina Almasry 306208cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 306308cf9fafSMina Almasry if (ret) 306408cf9fafSMina Almasry goto out_uncharge_cgroup_reservation; 30658f34af6fSJianyu Zhan 3066db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3067d85f69b0SMike Kravetz /* 3068d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 3069d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 3070d85f69b0SMike Kravetz * a reservation exists for the allocation. 3071d85f69b0SMike Kravetz */ 3072ff7d853bSSidhartha Kumar folio = dequeue_hugetlb_folio_vma(h, vma, addr, avoid_reserve, gbl_chg); 3073ff7d853bSSidhartha Kumar if (!folio) { 3074db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3075ff7d853bSSidhartha Kumar folio = alloc_buddy_hugetlb_folio_with_mpol(h, vma, addr); 3076ff7d853bSSidhartha Kumar if (!folio) 30778f34af6fSJianyu Zhan goto out_uncharge_cgroup; 307812df140fSRik van Riel spin_lock_irq(&hugetlb_lock); 3079a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 3080ff7d853bSSidhartha Kumar folio_set_hugetlb_restore_reserve(folio); 3081a88c7695SNaoya Horiguchi h->resv_huge_pages--; 3082a88c7695SNaoya Horiguchi } 3083ff7d853bSSidhartha Kumar list_add(&folio->lru, &h->hugepage_activelist); 3084ff7d853bSSidhartha Kumar folio_ref_unfreeze(folio, 1); 308581a6fcaeSJoonsoo Kim /* Fall through */ 3086a1e78772SMel Gorman } 3087ff7d853bSSidhartha Kumar 3088ff7d853bSSidhartha Kumar hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, folio); 308908cf9fafSMina Almasry /* If allocation is not consuming a reservation, also store the 309008cf9fafSMina Almasry * hugetlb_cgroup pointer on the page. 309108cf9fafSMina Almasry */ 309208cf9fafSMina Almasry if (deferred_reserve) { 309308cf9fafSMina Almasry hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), 3094ff7d853bSSidhartha Kumar h_cg, folio); 309508cf9fafSMina Almasry } 309608cf9fafSMina Almasry 3097db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3098a1e78772SMel Gorman 3099ff7d853bSSidhartha Kumar hugetlb_set_folio_subpool(folio, spool); 3100a1e78772SMel Gorman 3101d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 3102d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 310333039678SMike Kravetz /* 310433039678SMike Kravetz * The page was added to the reservation map between 310533039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 310633039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 310733039678SMike Kravetz * Adjust for the subpool count incremented above AND 310833039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 310933039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 311033039678SMike Kravetz * no longer applies. 311133039678SMike Kravetz */ 311233039678SMike Kravetz long rsv_adjust; 311333039678SMike Kravetz 311433039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 311533039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 311679aa925bSMike Kravetz if (deferred_reserve) 3117d4ab0316SSidhartha Kumar hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h), 3118d4ab0316SSidhartha Kumar pages_per_huge_page(h), folio); 311933039678SMike Kravetz } 3120d0ce0e47SSidhartha Kumar return folio; 31218f34af6fSJianyu Zhan 31228f34af6fSJianyu Zhan out_uncharge_cgroup: 31238f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 312408cf9fafSMina Almasry out_uncharge_cgroup_reservation: 312508cf9fafSMina Almasry if (deferred_reserve) 312608cf9fafSMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), 312708cf9fafSMina Almasry h_cg); 31288f34af6fSJianyu Zhan out_subpool_put: 3129d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 31308f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 3131feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 31328f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 3133b45b5bd6SDavid Gibson } 3134b45b5bd6SDavid Gibson 3135b5389086SZhenguo Yao int alloc_bootmem_huge_page(struct hstate *h, int nid) 3136e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 3137b5389086SZhenguo Yao int __alloc_bootmem_huge_page(struct hstate *h, int nid) 3138aa888a74SAndi Kleen { 3139b5389086SZhenguo Yao struct huge_bootmem_page *m = NULL; /* initialize for clang */ 3140b2261026SJoonsoo Kim int nr_nodes, node; 3141aa888a74SAndi Kleen 3142b5389086SZhenguo Yao /* do node specific alloc */ 3143b5389086SZhenguo Yao if (nid != NUMA_NO_NODE) { 3144b5389086SZhenguo Yao m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h), 3145b5389086SZhenguo Yao 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid); 3146b5389086SZhenguo Yao if (!m) 3147b5389086SZhenguo Yao return 0; 3148b5389086SZhenguo Yao goto found; 3149b5389086SZhenguo Yao } 3150b5389086SZhenguo Yao /* allocate from next node when distributing huge pages */ 3151b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 3152b5389086SZhenguo Yao m = memblock_alloc_try_nid_raw( 31538b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 315497ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 3155aa888a74SAndi Kleen /* 3156aa888a74SAndi Kleen * Use the beginning of the huge page to store the 3157aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 3158aa888a74SAndi Kleen * puts them into the mem_map). 3159aa888a74SAndi Kleen */ 3160b5389086SZhenguo Yao if (!m) 3161b5389086SZhenguo Yao return 0; 3162aa888a74SAndi Kleen goto found; 3163aa888a74SAndi Kleen } 3164aa888a74SAndi Kleen 3165aa888a74SAndi Kleen found: 3166aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 3167330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 3168aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 3169aa888a74SAndi Kleen m->hstate = h; 3170aa888a74SAndi Kleen return 1; 3171aa888a74SAndi Kleen } 3172aa888a74SAndi Kleen 317348b8d744SMike Kravetz /* 317448b8d744SMike Kravetz * Put bootmem huge pages into the standard lists after mem_map is up. 317548b8d744SMike Kravetz * Note: This only applies to gigantic (order > MAX_ORDER) pages. 317648b8d744SMike Kravetz */ 3177aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 3178aa888a74SAndi Kleen { 3179aa888a74SAndi Kleen struct huge_bootmem_page *m; 3180aa888a74SAndi Kleen 3181aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 318240d18ebfSMike Kravetz struct page *page = virt_to_page(m); 31837f325a8dSSidhartha Kumar struct folio *folio = page_folio(page); 3184aa888a74SAndi Kleen struct hstate *h = m->hstate; 3185ee8f248dSBecky Bruce 318648b8d744SMike Kravetz VM_BUG_ON(!hstate_is_gigantic(h)); 3187d1c60955SSidhartha Kumar WARN_ON(folio_ref_count(folio) != 1); 3188d1c60955SSidhartha Kumar if (prep_compound_gigantic_folio(folio, huge_page_order(h))) { 3189d1c60955SSidhartha Kumar WARN_ON(folio_test_reserved(folio)); 3190d1c60955SSidhartha Kumar prep_new_hugetlb_folio(h, folio, folio_nid(folio)); 3191454a00c4SMatthew Wilcox (Oracle) free_huge_folio(folio); /* add to the hugepage allocator */ 31927118fc29SMike Kravetz } else { 3193416d85edSMike Kravetz /* VERY unlikely inflated ref count on a tail page */ 31947f325a8dSSidhartha Kumar free_gigantic_folio(folio, huge_page_order(h)); 31957118fc29SMike Kravetz } 3196af0fb9dfSMichal Hocko 3197b0320c7bSRafael Aquini /* 319848b8d744SMike Kravetz * We need to restore the 'stolen' pages to totalram_pages 319948b8d744SMike Kravetz * in order to fix confusing memory reports from free(1) and 320048b8d744SMike Kravetz * other side-effects, like CommitLimit going negative. 3201b0320c7bSRafael Aquini */ 3202c78a7f36SMiaohe Lin adjust_managed_page_count(page, pages_per_huge_page(h)); 3203520495feSCannon Matthews cond_resched(); 3204aa888a74SAndi Kleen } 3205aa888a74SAndi Kleen } 3206b5389086SZhenguo Yao static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid) 3207b5389086SZhenguo Yao { 3208b5389086SZhenguo Yao unsigned long i; 3209b5389086SZhenguo Yao char buf[32]; 3210b5389086SZhenguo Yao 3211b5389086SZhenguo Yao for (i = 0; i < h->max_huge_pages_node[nid]; ++i) { 3212b5389086SZhenguo Yao if (hstate_is_gigantic(h)) { 3213b5389086SZhenguo Yao if (!alloc_bootmem_huge_page(h, nid)) 3214b5389086SZhenguo Yao break; 3215b5389086SZhenguo Yao } else { 321619fc1a7eSSidhartha Kumar struct folio *folio; 3217b5389086SZhenguo Yao gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 3218b5389086SZhenguo Yao 321919fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, 3220b5389086SZhenguo Yao &node_states[N_MEMORY], NULL); 322119fc1a7eSSidhartha Kumar if (!folio) 3222b5389086SZhenguo Yao break; 3223454a00c4SMatthew Wilcox (Oracle) free_huge_folio(folio); /* free it into the hugepage allocator */ 3224b5389086SZhenguo Yao } 3225b5389086SZhenguo Yao cond_resched(); 3226b5389086SZhenguo Yao } 3227b5389086SZhenguo Yao if (i == h->max_huge_pages_node[nid]) 3228b5389086SZhenguo Yao return; 3229b5389086SZhenguo Yao 3230b5389086SZhenguo Yao string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3231b5389086SZhenguo Yao pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n", 3232b5389086SZhenguo Yao h->max_huge_pages_node[nid], buf, nid, i); 3233b5389086SZhenguo Yao h->max_huge_pages -= (h->max_huge_pages_node[nid] - i); 3234b5389086SZhenguo Yao h->max_huge_pages_node[nid] = i; 3235b5389086SZhenguo Yao } 3236aa888a74SAndi Kleen 32378faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 32381da177e4SLinus Torvalds { 32391da177e4SLinus Torvalds unsigned long i; 3240f60858f9SMike Kravetz nodemask_t *node_alloc_noretry; 3241b5389086SZhenguo Yao bool node_specific_alloc = false; 3242f60858f9SMike Kravetz 3243b5389086SZhenguo Yao /* skip gigantic hugepages allocation if hugetlb_cma enabled */ 3244b5389086SZhenguo Yao if (hstate_is_gigantic(h) && hugetlb_cma_size) { 3245b5389086SZhenguo Yao pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); 3246b5389086SZhenguo Yao return; 3247b5389086SZhenguo Yao } 3248b5389086SZhenguo Yao 3249b5389086SZhenguo Yao /* do node specific alloc */ 32500a7a0f6fSPeng Liu for_each_online_node(i) { 3251b5389086SZhenguo Yao if (h->max_huge_pages_node[i] > 0) { 3252b5389086SZhenguo Yao hugetlb_hstate_alloc_pages_onenode(h, i); 3253b5389086SZhenguo Yao node_specific_alloc = true; 3254b5389086SZhenguo Yao } 3255b5389086SZhenguo Yao } 3256b5389086SZhenguo Yao 3257b5389086SZhenguo Yao if (node_specific_alloc) 3258b5389086SZhenguo Yao return; 3259b5389086SZhenguo Yao 3260b5389086SZhenguo Yao /* below will do all node balanced alloc */ 3261f60858f9SMike Kravetz if (!hstate_is_gigantic(h)) { 3262f60858f9SMike Kravetz /* 3263f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3264f60858f9SMike Kravetz * Ignore errors as lower level routines can deal with 3265f60858f9SMike Kravetz * node_alloc_noretry == NULL. If this kmalloc fails at boot 3266f60858f9SMike Kravetz * time, we are likely in bigger trouble. 3267f60858f9SMike Kravetz */ 3268f60858f9SMike Kravetz node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), 3269f60858f9SMike Kravetz GFP_KERNEL); 3270f60858f9SMike Kravetz } else { 3271f60858f9SMike Kravetz /* allocations done at boot time */ 3272f60858f9SMike Kravetz node_alloc_noretry = NULL; 3273f60858f9SMike Kravetz } 3274f60858f9SMike Kravetz 3275f60858f9SMike Kravetz /* bit mask controlling how hard we retry per-node allocations */ 3276f60858f9SMike Kravetz if (node_alloc_noretry) 3277f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 32781da177e4SLinus Torvalds 3279e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 3280bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 3281b5389086SZhenguo Yao if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE)) 3282aa888a74SAndi Kleen break; 32830c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 3284f60858f9SMike Kravetz &node_states[N_MEMORY], 3285f60858f9SMike Kravetz node_alloc_noretry)) 32861da177e4SLinus Torvalds break; 328769ed779aSDavid Rientjes cond_resched(); 32881da177e4SLinus Torvalds } 3289d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 3290d715cf80SLiam R. Howlett char buf[32]; 3291d715cf80SLiam R. Howlett 3292c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3293d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 3294d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 32958faa8b07SAndi Kleen h->max_huge_pages = i; 3296e5ff2159SAndi Kleen } 3297f60858f9SMike Kravetz kfree(node_alloc_noretry); 3298d715cf80SLiam R. Howlett } 3299e5ff2159SAndi Kleen 3300e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 3301e5ff2159SAndi Kleen { 330279dfc695SMike Kravetz struct hstate *h, *h2; 3303e5ff2159SAndi Kleen 3304e5ff2159SAndi Kleen for_each_hstate(h) { 33058faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 3306bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 33078faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 330879dfc695SMike Kravetz 330979dfc695SMike Kravetz /* 331079dfc695SMike Kravetz * Set demote order for each hstate. Note that 331179dfc695SMike Kravetz * h->demote_order is initially 0. 331279dfc695SMike Kravetz * - We can not demote gigantic pages if runtime freeing 331379dfc695SMike Kravetz * is not supported, so skip this. 3314a01f4390SMike Kravetz * - If CMA allocation is possible, we can not demote 3315a01f4390SMike Kravetz * HUGETLB_PAGE_ORDER or smaller size pages. 331679dfc695SMike Kravetz */ 331779dfc695SMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 331879dfc695SMike Kravetz continue; 3319a01f4390SMike Kravetz if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER) 3320a01f4390SMike Kravetz continue; 332179dfc695SMike Kravetz for_each_hstate(h2) { 332279dfc695SMike Kravetz if (h2 == h) 332379dfc695SMike Kravetz continue; 332479dfc695SMike Kravetz if (h2->order < h->order && 332579dfc695SMike Kravetz h2->order > h->demote_order) 332679dfc695SMike Kravetz h->demote_order = h2->order; 332779dfc695SMike Kravetz } 3328e5ff2159SAndi Kleen } 3329e5ff2159SAndi Kleen } 3330e5ff2159SAndi Kleen 3331e5ff2159SAndi Kleen static void __init report_hugepages(void) 3332e5ff2159SAndi Kleen { 3333e5ff2159SAndi Kleen struct hstate *h; 3334e5ff2159SAndi Kleen 3335e5ff2159SAndi Kleen for_each_hstate(h) { 33364abd32dbSAndi Kleen char buf[32]; 3337c6247f72SMatthew Wilcox 3338c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 33396213834cSMuchun Song pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n", 3340c6247f72SMatthew Wilcox buf, h->free_huge_pages); 33416213834cSMuchun Song pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n", 33426213834cSMuchun Song hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf); 3343e5ff2159SAndi Kleen } 3344e5ff2159SAndi Kleen } 3345e5ff2159SAndi Kleen 33461da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 33476ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 33486ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 33491da177e4SLinus Torvalds { 33504415cc8dSChristoph Lameter int i; 33511121828aSMike Kravetz LIST_HEAD(page_list); 33524415cc8dSChristoph Lameter 33539487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 3354bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3355aa888a74SAndi Kleen return; 3356aa888a74SAndi Kleen 33571121828aSMike Kravetz /* 33581121828aSMike Kravetz * Collect pages to be freed on a list, and free after dropping lock 33591121828aSMike Kravetz */ 33606ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 336110c6ec49SMike Kravetz struct page *page, *next; 3362a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 3363a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 3364a5516438SAndi Kleen if (count >= h->nr_huge_pages) 33651121828aSMike Kravetz goto out; 33661da177e4SLinus Torvalds if (PageHighMem(page)) 33671da177e4SLinus Torvalds continue; 3368cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, page_folio(page), false); 33691121828aSMike Kravetz list_add(&page->lru, &page_list); 33701121828aSMike Kravetz } 33711121828aSMike Kravetz } 33721121828aSMike Kravetz 33731121828aSMike Kravetz out: 3374db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 337510c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3376db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 33771da177e4SLinus Torvalds } 33781da177e4SLinus Torvalds #else 33796ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 33806ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 33811da177e4SLinus Torvalds { 33821da177e4SLinus Torvalds } 33831da177e4SLinus Torvalds #endif 33841da177e4SLinus Torvalds 338520a0307cSWu Fengguang /* 338620a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 338720a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 338820a0307cSWu Fengguang * Returns 1 if an adjustment was made. 338920a0307cSWu Fengguang */ 33906ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 33916ae11b27SLee Schermerhorn int delta) 339220a0307cSWu Fengguang { 3393b2261026SJoonsoo Kim int nr_nodes, node; 339420a0307cSWu Fengguang 33959487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 339620a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 339720a0307cSWu Fengguang 3398e8c5c824SLee Schermerhorn if (delta < 0) { 3399b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 3400b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 3401b2261026SJoonsoo Kim goto found; 3402b2261026SJoonsoo Kim } 3403b2261026SJoonsoo Kim } else { 3404b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 3405b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 3406b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 3407b2261026SJoonsoo Kim goto found; 3408e8c5c824SLee Schermerhorn } 34099a76db09SLee Schermerhorn } 3410b2261026SJoonsoo Kim return 0; 341120a0307cSWu Fengguang 3412b2261026SJoonsoo Kim found: 341320a0307cSWu Fengguang h->surplus_huge_pages += delta; 3414b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 3415b2261026SJoonsoo Kim return 1; 341620a0307cSWu Fengguang } 341720a0307cSWu Fengguang 3418a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 3419fd875dcaSMike Kravetz static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 34206ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 34211da177e4SLinus Torvalds { 34227893d1d5SAdam Litke unsigned long min_count, ret; 342310c6ec49SMike Kravetz struct page *page; 342410c6ec49SMike Kravetz LIST_HEAD(page_list); 3425f60858f9SMike Kravetz NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); 3426f60858f9SMike Kravetz 3427f60858f9SMike Kravetz /* 3428f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3429f60858f9SMike Kravetz * If we can not allocate the bit mask, do not attempt to allocate 3430f60858f9SMike Kravetz * the requested huge pages. 3431f60858f9SMike Kravetz */ 3432f60858f9SMike Kravetz if (node_alloc_noretry) 3433f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 3434f60858f9SMike Kravetz else 3435f60858f9SMike Kravetz return -ENOMEM; 34361da177e4SLinus Torvalds 343729383967SMike Kravetz /* 343829383967SMike Kravetz * resize_lock mutex prevents concurrent adjustments to number of 343929383967SMike Kravetz * pages in hstate via the proc/sysfs interfaces. 344029383967SMike Kravetz */ 344129383967SMike Kravetz mutex_lock(&h->resize_lock); 3442b65d4adbSMuchun Song flush_free_hpage_work(h); 3443db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 34444eb0716eSAlexandre Ghiti 34454eb0716eSAlexandre Ghiti /* 3446fd875dcaSMike Kravetz * Check for a node specific request. 3447fd875dcaSMike Kravetz * Changing node specific huge page count may require a corresponding 3448fd875dcaSMike Kravetz * change to the global count. In any case, the passed node mask 3449fd875dcaSMike Kravetz * (nodes_allowed) will restrict alloc/free to the specified node. 3450fd875dcaSMike Kravetz */ 3451fd875dcaSMike Kravetz if (nid != NUMA_NO_NODE) { 3452fd875dcaSMike Kravetz unsigned long old_count = count; 3453fd875dcaSMike Kravetz 3454fd875dcaSMike Kravetz count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 3455fd875dcaSMike Kravetz /* 3456fd875dcaSMike Kravetz * User may have specified a large count value which caused the 3457fd875dcaSMike Kravetz * above calculation to overflow. In this case, they wanted 3458fd875dcaSMike Kravetz * to allocate as many huge pages as possible. Set count to 3459fd875dcaSMike Kravetz * largest possible value to align with their intention. 3460fd875dcaSMike Kravetz */ 3461fd875dcaSMike Kravetz if (count < old_count) 3462fd875dcaSMike Kravetz count = ULONG_MAX; 3463fd875dcaSMike Kravetz } 3464fd875dcaSMike Kravetz 3465fd875dcaSMike Kravetz /* 34664eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 34674eb0716eSAlexandre Ghiti * page range allocation. 34684eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 34694eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 34704eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 34714eb0716eSAlexandre Ghiti */ 34724eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 34734eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 3474db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 347529383967SMike Kravetz mutex_unlock(&h->resize_lock); 3476f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 34774eb0716eSAlexandre Ghiti return -EINVAL; 34784eb0716eSAlexandre Ghiti } 34794eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 34804eb0716eSAlexandre Ghiti } 3481aa888a74SAndi Kleen 34827893d1d5SAdam Litke /* 34837893d1d5SAdam Litke * Increase the pool size 34847893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 34857893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 3486d1c3fb1fSNishanth Aravamudan * 34873a740e8bSSidhartha Kumar * We might race with alloc_surplus_hugetlb_folio() here and be unable 3488d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 3489d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 3490d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 3491d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 34927893d1d5SAdam Litke */ 3493a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 34946ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 34957893d1d5SAdam Litke break; 34967893d1d5SAdam Litke } 34977893d1d5SAdam Litke 3498a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 34997893d1d5SAdam Litke /* 35007893d1d5SAdam Litke * If this allocation races such that we no longer need the 3501454a00c4SMatthew Wilcox (Oracle) * page, free_huge_folio will handle it by freeing the page 35027893d1d5SAdam Litke * and reducing the surplus. 35037893d1d5SAdam Litke */ 3504db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3505649920c6SJia He 3506649920c6SJia He /* yield cpu to avoid soft lockup */ 3507649920c6SJia He cond_resched(); 3508649920c6SJia He 3509f60858f9SMike Kravetz ret = alloc_pool_huge_page(h, nodes_allowed, 3510f60858f9SMike Kravetz node_alloc_noretry); 3511db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 35127893d1d5SAdam Litke if (!ret) 35137893d1d5SAdam Litke goto out; 35147893d1d5SAdam Litke 3515536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 3516536240f2SMel Gorman if (signal_pending(current)) 3517536240f2SMel Gorman goto out; 35187893d1d5SAdam Litke } 35197893d1d5SAdam Litke 35207893d1d5SAdam Litke /* 35217893d1d5SAdam Litke * Decrease the pool size 35227893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 35237893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 35247893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 35257893d1d5SAdam Litke * to the desired size as pages become free. 3526d1c3fb1fSNishanth Aravamudan * 3527d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 3528d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 3529d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 35303a740e8bSSidhartha Kumar * alloc_surplus_hugetlb_folio() is checking the global counter, 3531d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 3532d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 3533d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 35347893d1d5SAdam Litke */ 3535a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 35366b0c880dSAdam Litke min_count = max(count, min_count); 35376ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 353810c6ec49SMike Kravetz 353910c6ec49SMike Kravetz /* 354010c6ec49SMike Kravetz * Collect pages to be removed on list without dropping lock 354110c6ec49SMike Kravetz */ 3542a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 354310c6ec49SMike Kravetz page = remove_pool_huge_page(h, nodes_allowed, 0); 354410c6ec49SMike Kravetz if (!page) 35451da177e4SLinus Torvalds break; 354610c6ec49SMike Kravetz 354710c6ec49SMike Kravetz list_add(&page->lru, &page_list); 35481da177e4SLinus Torvalds } 354910c6ec49SMike Kravetz /* free the pages after dropping lock */ 3550db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 355110c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3552b65d4adbSMuchun Song flush_free_hpage_work(h); 3553db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 355410c6ec49SMike Kravetz 3555a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 35566ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 35577893d1d5SAdam Litke break; 35587893d1d5SAdam Litke } 35597893d1d5SAdam Litke out: 35604eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 3561db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 356229383967SMike Kravetz mutex_unlock(&h->resize_lock); 35634eb0716eSAlexandre Ghiti 3564f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 3565f60858f9SMike Kravetz 35664eb0716eSAlexandre Ghiti return 0; 35671da177e4SLinus Torvalds } 35681da177e4SLinus Torvalds 3569bdd7be07SSidhartha Kumar static int demote_free_hugetlb_folio(struct hstate *h, struct folio *folio) 35708531fc6fSMike Kravetz { 3571bdd7be07SSidhartha Kumar int i, nid = folio_nid(folio); 35728531fc6fSMike Kravetz struct hstate *target_hstate; 357331731452SDoug Berger struct page *subpage; 3574bdd7be07SSidhartha Kumar struct folio *inner_folio; 35758531fc6fSMike Kravetz int rc = 0; 35768531fc6fSMike Kravetz 35778531fc6fSMike Kravetz target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order); 35788531fc6fSMike Kravetz 3579cfd5082bSSidhartha Kumar remove_hugetlb_folio_for_demote(h, folio, false); 35808531fc6fSMike Kravetz spin_unlock_irq(&hugetlb_lock); 35818531fc6fSMike Kravetz 3582bdd7be07SSidhartha Kumar rc = hugetlb_vmemmap_restore(h, &folio->page); 35838531fc6fSMike Kravetz if (rc) { 3584bdd7be07SSidhartha Kumar /* Allocation of vmemmmap failed, we can not demote folio */ 35858531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 3586bdd7be07SSidhartha Kumar folio_ref_unfreeze(folio, 1); 3587bdd7be07SSidhartha Kumar add_hugetlb_folio(h, folio, false); 35888531fc6fSMike Kravetz return rc; 35898531fc6fSMike Kravetz } 35908531fc6fSMike Kravetz 35918531fc6fSMike Kravetz /* 3592911565b8SSidhartha Kumar * Use destroy_compound_hugetlb_folio_for_demote for all huge page 3593bdd7be07SSidhartha Kumar * sizes as it will not ref count folios. 35948531fc6fSMike Kravetz */ 3595911565b8SSidhartha Kumar destroy_compound_hugetlb_folio_for_demote(folio, huge_page_order(h)); 35968531fc6fSMike Kravetz 35978531fc6fSMike Kravetz /* 35988531fc6fSMike Kravetz * Taking target hstate mutex synchronizes with set_max_huge_pages. 35998531fc6fSMike Kravetz * Without the mutex, pages added to target hstate could be marked 36008531fc6fSMike Kravetz * as surplus. 36018531fc6fSMike Kravetz * 36028531fc6fSMike Kravetz * Note that we already hold h->resize_lock. To prevent deadlock, 36038531fc6fSMike Kravetz * use the convention of always taking larger size hstate mutex first. 36048531fc6fSMike Kravetz */ 36058531fc6fSMike Kravetz mutex_lock(&target_hstate->resize_lock); 36068531fc6fSMike Kravetz for (i = 0; i < pages_per_huge_page(h); 36078531fc6fSMike Kravetz i += pages_per_huge_page(target_hstate)) { 3608bdd7be07SSidhartha Kumar subpage = folio_page(folio, i); 3609bdd7be07SSidhartha Kumar inner_folio = page_folio(subpage); 36108531fc6fSMike Kravetz if (hstate_is_gigantic(target_hstate)) 3611bdd7be07SSidhartha Kumar prep_compound_gigantic_folio_for_demote(inner_folio, 36128531fc6fSMike Kravetz target_hstate->order); 36138531fc6fSMike Kravetz else 361431731452SDoug Berger prep_compound_page(subpage, target_hstate->order); 3615bdd7be07SSidhartha Kumar folio_change_private(inner_folio, NULL); 3616bdd7be07SSidhartha Kumar prep_new_hugetlb_folio(target_hstate, inner_folio, nid); 3617454a00c4SMatthew Wilcox (Oracle) free_huge_folio(inner_folio); 36188531fc6fSMike Kravetz } 36198531fc6fSMike Kravetz mutex_unlock(&target_hstate->resize_lock); 36208531fc6fSMike Kravetz 36218531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 36228531fc6fSMike Kravetz 36238531fc6fSMike Kravetz /* 36248531fc6fSMike Kravetz * Not absolutely necessary, but for consistency update max_huge_pages 36258531fc6fSMike Kravetz * based on pool changes for the demoted page. 36268531fc6fSMike Kravetz */ 36278531fc6fSMike Kravetz h->max_huge_pages--; 3628a43a83c7SMiaohe Lin target_hstate->max_huge_pages += 3629a43a83c7SMiaohe Lin pages_per_huge_page(h) / pages_per_huge_page(target_hstate); 36308531fc6fSMike Kravetz 36318531fc6fSMike Kravetz return rc; 36328531fc6fSMike Kravetz } 36338531fc6fSMike Kravetz 363479dfc695SMike Kravetz static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 363579dfc695SMike Kravetz __must_hold(&hugetlb_lock) 363679dfc695SMike Kravetz { 36378531fc6fSMike Kravetz int nr_nodes, node; 3638bdd7be07SSidhartha Kumar struct folio *folio; 363979dfc695SMike Kravetz 364079dfc695SMike Kravetz lockdep_assert_held(&hugetlb_lock); 364179dfc695SMike Kravetz 364279dfc695SMike Kravetz /* We should never get here if no demote order */ 364379dfc695SMike Kravetz if (!h->demote_order) { 364479dfc695SMike Kravetz pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n"); 364579dfc695SMike Kravetz return -EINVAL; /* internal error */ 364679dfc695SMike Kravetz } 364779dfc695SMike Kravetz 36488531fc6fSMike Kravetz for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 3649bdd7be07SSidhartha Kumar list_for_each_entry(folio, &h->hugepage_freelists[node], lru) { 3650bdd7be07SSidhartha Kumar if (folio_test_hwpoison(folio)) 36515a317412SMike Kravetz continue; 3652bdd7be07SSidhartha Kumar return demote_free_hugetlb_folio(h, folio); 36538531fc6fSMike Kravetz } 36548531fc6fSMike Kravetz } 36558531fc6fSMike Kravetz 36565a317412SMike Kravetz /* 36575a317412SMike Kravetz * Only way to get here is if all pages on free lists are poisoned. 36585a317412SMike Kravetz * Return -EBUSY so that caller will not retry. 36595a317412SMike Kravetz */ 36605a317412SMike Kravetz return -EBUSY; 366179dfc695SMike Kravetz } 366279dfc695SMike Kravetz 3663a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 3664a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 3665a3437870SNishanth Aravamudan 366679dfc695SMike Kravetz #define HSTATE_ATTR_WO(_name) \ 366779dfc695SMike Kravetz static struct kobj_attribute _name##_attr = __ATTR_WO(_name) 366879dfc695SMike Kravetz 3669a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 367098bc26acSMiaohe Lin static struct kobj_attribute _name##_attr = __ATTR_RW(_name) 3671a3437870SNishanth Aravamudan 3672a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 3673a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 3674a3437870SNishanth Aravamudan 36759a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 36769a305230SLee Schermerhorn 36779a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 3678a3437870SNishanth Aravamudan { 3679a3437870SNishanth Aravamudan int i; 36809a305230SLee Schermerhorn 3681a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 36829a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 36839a305230SLee Schermerhorn if (nidp) 36849a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 3685a3437870SNishanth Aravamudan return &hstates[i]; 36869a305230SLee Schermerhorn } 36879a305230SLee Schermerhorn 36889a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 3689a3437870SNishanth Aravamudan } 3690a3437870SNishanth Aravamudan 369106808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 3692a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3693a3437870SNishanth Aravamudan { 36949a305230SLee Schermerhorn struct hstate *h; 36959a305230SLee Schermerhorn unsigned long nr_huge_pages; 36969a305230SLee Schermerhorn int nid; 36979a305230SLee Schermerhorn 36989a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 36999a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 37009a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 37019a305230SLee Schermerhorn else 37029a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 37039a305230SLee Schermerhorn 3704ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", nr_huge_pages); 3705a3437870SNishanth Aravamudan } 3706adbe8726SEric B Munson 3707238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 3708238d3c13SDavid Rientjes struct hstate *h, int nid, 3709238d3c13SDavid Rientjes unsigned long count, size_t len) 3710a3437870SNishanth Aravamudan { 3711a3437870SNishanth Aravamudan int err; 37122d0adf7eSOscar Salvador nodemask_t nodes_allowed, *n_mask; 3713a3437870SNishanth Aravamudan 37142d0adf7eSOscar Salvador if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 37152d0adf7eSOscar Salvador return -EINVAL; 3716adbe8726SEric B Munson 37179a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 37189a305230SLee Schermerhorn /* 37199a305230SLee Schermerhorn * global hstate attribute 37209a305230SLee Schermerhorn */ 37219a305230SLee Schermerhorn if (!(obey_mempolicy && 37222d0adf7eSOscar Salvador init_nodemask_of_mempolicy(&nodes_allowed))) 37232d0adf7eSOscar Salvador n_mask = &node_states[N_MEMORY]; 37242d0adf7eSOscar Salvador else 37252d0adf7eSOscar Salvador n_mask = &nodes_allowed; 37262d0adf7eSOscar Salvador } else { 37279a305230SLee Schermerhorn /* 3728fd875dcaSMike Kravetz * Node specific request. count adjustment happens in 3729fd875dcaSMike Kravetz * set_max_huge_pages() after acquiring hugetlb_lock. 37309a305230SLee Schermerhorn */ 37312d0adf7eSOscar Salvador init_nodemask_of_node(&nodes_allowed, nid); 37322d0adf7eSOscar Salvador n_mask = &nodes_allowed; 3733fd875dcaSMike Kravetz } 37349a305230SLee Schermerhorn 37352d0adf7eSOscar Salvador err = set_max_huge_pages(h, count, nid, n_mask); 373606808b08SLee Schermerhorn 37374eb0716eSAlexandre Ghiti return err ? err : len; 373806808b08SLee Schermerhorn } 373906808b08SLee Schermerhorn 3740238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 3741238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 3742238d3c13SDavid Rientjes size_t len) 3743238d3c13SDavid Rientjes { 3744238d3c13SDavid Rientjes struct hstate *h; 3745238d3c13SDavid Rientjes unsigned long count; 3746238d3c13SDavid Rientjes int nid; 3747238d3c13SDavid Rientjes int err; 3748238d3c13SDavid Rientjes 3749238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 3750238d3c13SDavid Rientjes if (err) 3751238d3c13SDavid Rientjes return err; 3752238d3c13SDavid Rientjes 3753238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 3754238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 3755238d3c13SDavid Rientjes } 3756238d3c13SDavid Rientjes 375706808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 375806808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 375906808b08SLee Schermerhorn { 376006808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 376106808b08SLee Schermerhorn } 376206808b08SLee Schermerhorn 376306808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 376406808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 376506808b08SLee Schermerhorn { 3766238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 3767a3437870SNishanth Aravamudan } 3768a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 3769a3437870SNishanth Aravamudan 377006808b08SLee Schermerhorn #ifdef CONFIG_NUMA 377106808b08SLee Schermerhorn 377206808b08SLee Schermerhorn /* 377306808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 377406808b08SLee Schermerhorn * huge page alloc/free. 377506808b08SLee Schermerhorn */ 377606808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 3777ae7a927dSJoe Perches struct kobj_attribute *attr, 3778ae7a927dSJoe Perches char *buf) 377906808b08SLee Schermerhorn { 378006808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 378106808b08SLee Schermerhorn } 378206808b08SLee Schermerhorn 378306808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 378406808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 378506808b08SLee Schermerhorn { 3786238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 378706808b08SLee Schermerhorn } 378806808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 378906808b08SLee Schermerhorn #endif 379006808b08SLee Schermerhorn 379106808b08SLee Schermerhorn 3792a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 3793a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3794a3437870SNishanth Aravamudan { 37959a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3796ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); 3797a3437870SNishanth Aravamudan } 3798adbe8726SEric B Munson 3799a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 3800a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 3801a3437870SNishanth Aravamudan { 3802a3437870SNishanth Aravamudan int err; 3803a3437870SNishanth Aravamudan unsigned long input; 38049a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3805a3437870SNishanth Aravamudan 3806bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3807adbe8726SEric B Munson return -EINVAL; 3808adbe8726SEric B Munson 38093dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 3810a3437870SNishanth Aravamudan if (err) 381173ae31e5SEric B Munson return err; 3812a3437870SNishanth Aravamudan 3813db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3814a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 3815db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3816a3437870SNishanth Aravamudan 3817a3437870SNishanth Aravamudan return count; 3818a3437870SNishanth Aravamudan } 3819a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 3820a3437870SNishanth Aravamudan 3821a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 3822a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3823a3437870SNishanth Aravamudan { 38249a305230SLee Schermerhorn struct hstate *h; 38259a305230SLee Schermerhorn unsigned long free_huge_pages; 38269a305230SLee Schermerhorn int nid; 38279a305230SLee Schermerhorn 38289a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 38299a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 38309a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 38319a305230SLee Schermerhorn else 38329a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 38339a305230SLee Schermerhorn 3834ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", free_huge_pages); 3835a3437870SNishanth Aravamudan } 3836a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 3837a3437870SNishanth Aravamudan 3838a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 3839a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3840a3437870SNishanth Aravamudan { 38419a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3842ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); 3843a3437870SNishanth Aravamudan } 3844a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 3845a3437870SNishanth Aravamudan 3846a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 3847a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3848a3437870SNishanth Aravamudan { 38499a305230SLee Schermerhorn struct hstate *h; 38509a305230SLee Schermerhorn unsigned long surplus_huge_pages; 38519a305230SLee Schermerhorn int nid; 38529a305230SLee Schermerhorn 38539a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 38549a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 38559a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 38569a305230SLee Schermerhorn else 38579a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 38589a305230SLee Schermerhorn 3859ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", surplus_huge_pages); 3860a3437870SNishanth Aravamudan } 3861a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 3862a3437870SNishanth Aravamudan 386379dfc695SMike Kravetz static ssize_t demote_store(struct kobject *kobj, 386479dfc695SMike Kravetz struct kobj_attribute *attr, const char *buf, size_t len) 386579dfc695SMike Kravetz { 386679dfc695SMike Kravetz unsigned long nr_demote; 386779dfc695SMike Kravetz unsigned long nr_available; 386879dfc695SMike Kravetz nodemask_t nodes_allowed, *n_mask; 386979dfc695SMike Kravetz struct hstate *h; 38708eeda55fSLi zeming int err; 387179dfc695SMike Kravetz int nid; 387279dfc695SMike Kravetz 387379dfc695SMike Kravetz err = kstrtoul(buf, 10, &nr_demote); 387479dfc695SMike Kravetz if (err) 387579dfc695SMike Kravetz return err; 387679dfc695SMike Kravetz h = kobj_to_hstate(kobj, &nid); 387779dfc695SMike Kravetz 387879dfc695SMike Kravetz if (nid != NUMA_NO_NODE) { 387979dfc695SMike Kravetz init_nodemask_of_node(&nodes_allowed, nid); 388079dfc695SMike Kravetz n_mask = &nodes_allowed; 388179dfc695SMike Kravetz } else { 388279dfc695SMike Kravetz n_mask = &node_states[N_MEMORY]; 388379dfc695SMike Kravetz } 388479dfc695SMike Kravetz 388579dfc695SMike Kravetz /* Synchronize with other sysfs operations modifying huge pages */ 388679dfc695SMike Kravetz mutex_lock(&h->resize_lock); 388779dfc695SMike Kravetz spin_lock_irq(&hugetlb_lock); 388879dfc695SMike Kravetz 388979dfc695SMike Kravetz while (nr_demote) { 389079dfc695SMike Kravetz /* 389179dfc695SMike Kravetz * Check for available pages to demote each time thorough the 389279dfc695SMike Kravetz * loop as demote_pool_huge_page will drop hugetlb_lock. 389379dfc695SMike Kravetz */ 389479dfc695SMike Kravetz if (nid != NUMA_NO_NODE) 389579dfc695SMike Kravetz nr_available = h->free_huge_pages_node[nid]; 389679dfc695SMike Kravetz else 389779dfc695SMike Kravetz nr_available = h->free_huge_pages; 389879dfc695SMike Kravetz nr_available -= h->resv_huge_pages; 389979dfc695SMike Kravetz if (!nr_available) 390079dfc695SMike Kravetz break; 390179dfc695SMike Kravetz 390279dfc695SMike Kravetz err = demote_pool_huge_page(h, n_mask); 390379dfc695SMike Kravetz if (err) 390479dfc695SMike Kravetz break; 390579dfc695SMike Kravetz 390679dfc695SMike Kravetz nr_demote--; 390779dfc695SMike Kravetz } 390879dfc695SMike Kravetz 390979dfc695SMike Kravetz spin_unlock_irq(&hugetlb_lock); 391079dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 391179dfc695SMike Kravetz 391279dfc695SMike Kravetz if (err) 391379dfc695SMike Kravetz return err; 391479dfc695SMike Kravetz return len; 391579dfc695SMike Kravetz } 391679dfc695SMike Kravetz HSTATE_ATTR_WO(demote); 391779dfc695SMike Kravetz 391879dfc695SMike Kravetz static ssize_t demote_size_show(struct kobject *kobj, 391979dfc695SMike Kravetz struct kobj_attribute *attr, char *buf) 392079dfc695SMike Kravetz { 392112658abfSMiaohe Lin struct hstate *h = kobj_to_hstate(kobj, NULL); 392279dfc695SMike Kravetz unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K; 392379dfc695SMike Kravetz 392479dfc695SMike Kravetz return sysfs_emit(buf, "%lukB\n", demote_size); 392579dfc695SMike Kravetz } 392679dfc695SMike Kravetz 392779dfc695SMike Kravetz static ssize_t demote_size_store(struct kobject *kobj, 392879dfc695SMike Kravetz struct kobj_attribute *attr, 392979dfc695SMike Kravetz const char *buf, size_t count) 393079dfc695SMike Kravetz { 393179dfc695SMike Kravetz struct hstate *h, *demote_hstate; 393279dfc695SMike Kravetz unsigned long demote_size; 393379dfc695SMike Kravetz unsigned int demote_order; 393479dfc695SMike Kravetz 393579dfc695SMike Kravetz demote_size = (unsigned long)memparse(buf, NULL); 393679dfc695SMike Kravetz 393779dfc695SMike Kravetz demote_hstate = size_to_hstate(demote_size); 393879dfc695SMike Kravetz if (!demote_hstate) 393979dfc695SMike Kravetz return -EINVAL; 394079dfc695SMike Kravetz demote_order = demote_hstate->order; 3941a01f4390SMike Kravetz if (demote_order < HUGETLB_PAGE_ORDER) 3942a01f4390SMike Kravetz return -EINVAL; 394379dfc695SMike Kravetz 394479dfc695SMike Kravetz /* demote order must be smaller than hstate order */ 394512658abfSMiaohe Lin h = kobj_to_hstate(kobj, NULL); 394679dfc695SMike Kravetz if (demote_order >= h->order) 394779dfc695SMike Kravetz return -EINVAL; 394879dfc695SMike Kravetz 394979dfc695SMike Kravetz /* resize_lock synchronizes access to demote size and writes */ 395079dfc695SMike Kravetz mutex_lock(&h->resize_lock); 395179dfc695SMike Kravetz h->demote_order = demote_order; 395279dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 395379dfc695SMike Kravetz 395479dfc695SMike Kravetz return count; 395579dfc695SMike Kravetz } 395679dfc695SMike Kravetz HSTATE_ATTR(demote_size); 395779dfc695SMike Kravetz 3958a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 3959a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 3960a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 3961a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 3962a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 3963a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 396406808b08SLee Schermerhorn #ifdef CONFIG_NUMA 396506808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 396606808b08SLee Schermerhorn #endif 3967a3437870SNishanth Aravamudan NULL, 3968a3437870SNishanth Aravamudan }; 3969a3437870SNishanth Aravamudan 397067e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 3971a3437870SNishanth Aravamudan .attrs = hstate_attrs, 3972a3437870SNishanth Aravamudan }; 3973a3437870SNishanth Aravamudan 397479dfc695SMike Kravetz static struct attribute *hstate_demote_attrs[] = { 397579dfc695SMike Kravetz &demote_size_attr.attr, 397679dfc695SMike Kravetz &demote_attr.attr, 397779dfc695SMike Kravetz NULL, 397879dfc695SMike Kravetz }; 397979dfc695SMike Kravetz 398079dfc695SMike Kravetz static const struct attribute_group hstate_demote_attr_group = { 398179dfc695SMike Kravetz .attrs = hstate_demote_attrs, 398279dfc695SMike Kravetz }; 398379dfc695SMike Kravetz 3984094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 39859a305230SLee Schermerhorn struct kobject **hstate_kobjs, 398667e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 3987a3437870SNishanth Aravamudan { 3988a3437870SNishanth Aravamudan int retval; 3989972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 3990a3437870SNishanth Aravamudan 39919a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 39929a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 3993a3437870SNishanth Aravamudan return -ENOMEM; 3994a3437870SNishanth Aravamudan 39959a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 3996cc2205a6SMiaohe Lin if (retval) { 39979a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 3998cc2205a6SMiaohe Lin hstate_kobjs[hi] = NULL; 39993a6bdda0SMiaohe Lin return retval; 4000cc2205a6SMiaohe Lin } 4001a3437870SNishanth Aravamudan 400279dfc695SMike Kravetz if (h->demote_order) { 400301088a60SMiaohe Lin retval = sysfs_create_group(hstate_kobjs[hi], 400401088a60SMiaohe Lin &hstate_demote_attr_group); 400501088a60SMiaohe Lin if (retval) { 400679dfc695SMike Kravetz pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name); 400701088a60SMiaohe Lin sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group); 400801088a60SMiaohe Lin kobject_put(hstate_kobjs[hi]); 400901088a60SMiaohe Lin hstate_kobjs[hi] = NULL; 401001088a60SMiaohe Lin return retval; 401101088a60SMiaohe Lin } 401279dfc695SMike Kravetz } 401379dfc695SMike Kravetz 401401088a60SMiaohe Lin return 0; 4015a3437870SNishanth Aravamudan } 4016a3437870SNishanth Aravamudan 40179a305230SLee Schermerhorn #ifdef CONFIG_NUMA 4018a4a00b45SMuchun Song static bool hugetlb_sysfs_initialized __ro_after_init; 40199a305230SLee Schermerhorn 40209a305230SLee Schermerhorn /* 40219a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 402210fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 402310fbcf4cSKay Sievers * index of a node device or _hstate == node id. 402410fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 40259a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 40269a305230SLee Schermerhorn */ 40279a305230SLee Schermerhorn struct node_hstate { 40289a305230SLee Schermerhorn struct kobject *hugepages_kobj; 40299a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 40309a305230SLee Schermerhorn }; 4031b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 40329a305230SLee Schermerhorn 40339a305230SLee Schermerhorn /* 403410fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 40359a305230SLee Schermerhorn */ 40369a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 40379a305230SLee Schermerhorn &nr_hugepages_attr.attr, 40389a305230SLee Schermerhorn &free_hugepages_attr.attr, 40399a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 40409a305230SLee Schermerhorn NULL, 40419a305230SLee Schermerhorn }; 40429a305230SLee Schermerhorn 404367e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 40449a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 40459a305230SLee Schermerhorn }; 40469a305230SLee Schermerhorn 40479a305230SLee Schermerhorn /* 404810fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 40499a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 40509a305230SLee Schermerhorn */ 40519a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 40529a305230SLee Schermerhorn { 40539a305230SLee Schermerhorn int nid; 40549a305230SLee Schermerhorn 40559a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 40569a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 40579a305230SLee Schermerhorn int i; 40589a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 40599a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 40609a305230SLee Schermerhorn if (nidp) 40619a305230SLee Schermerhorn *nidp = nid; 40629a305230SLee Schermerhorn return &hstates[i]; 40639a305230SLee Schermerhorn } 40649a305230SLee Schermerhorn } 40659a305230SLee Schermerhorn 40669a305230SLee Schermerhorn BUG(); 40679a305230SLee Schermerhorn return NULL; 40689a305230SLee Schermerhorn } 40699a305230SLee Schermerhorn 40709a305230SLee Schermerhorn /* 407110fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 40729a305230SLee Schermerhorn * No-op if no hstate attributes attached. 40739a305230SLee Schermerhorn */ 4074a4a00b45SMuchun Song void hugetlb_unregister_node(struct node *node) 40759a305230SLee Schermerhorn { 40769a305230SLee Schermerhorn struct hstate *h; 407710fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 40789a305230SLee Schermerhorn 40799a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 40809b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 40819a305230SLee Schermerhorn 4082972dc4deSAneesh Kumar K.V for_each_hstate(h) { 4083972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 408401088a60SMiaohe Lin struct kobject *hstate_kobj = nhs->hstate_kobjs[idx]; 408501088a60SMiaohe Lin 408601088a60SMiaohe Lin if (!hstate_kobj) 408701088a60SMiaohe Lin continue; 408801088a60SMiaohe Lin if (h->demote_order) 408901088a60SMiaohe Lin sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group); 409001088a60SMiaohe Lin sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group); 409101088a60SMiaohe Lin kobject_put(hstate_kobj); 4092972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 4093972dc4deSAneesh Kumar K.V } 40949a305230SLee Schermerhorn 40959a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 40969a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 40979a305230SLee Schermerhorn } 40989a305230SLee Schermerhorn 40999a305230SLee Schermerhorn 41009a305230SLee Schermerhorn /* 410110fbcf4cSKay Sievers * Register hstate attributes for a single node device. 41029a305230SLee Schermerhorn * No-op if attributes already registered. 41039a305230SLee Schermerhorn */ 4104a4a00b45SMuchun Song void hugetlb_register_node(struct node *node) 41059a305230SLee Schermerhorn { 41069a305230SLee Schermerhorn struct hstate *h; 410710fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 41089a305230SLee Schermerhorn int err; 41099a305230SLee Schermerhorn 4110a4a00b45SMuchun Song if (!hugetlb_sysfs_initialized) 4111a4a00b45SMuchun Song return; 4112a4a00b45SMuchun Song 41139a305230SLee Schermerhorn if (nhs->hugepages_kobj) 41149a305230SLee Schermerhorn return; /* already allocated */ 41159a305230SLee Schermerhorn 41169a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 411710fbcf4cSKay Sievers &node->dev.kobj); 41189a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 41199a305230SLee Schermerhorn return; 41209a305230SLee Schermerhorn 41219a305230SLee Schermerhorn for_each_hstate(h) { 41229a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 41239a305230SLee Schermerhorn nhs->hstate_kobjs, 41249a305230SLee Schermerhorn &per_node_hstate_attr_group); 41259a305230SLee Schermerhorn if (err) { 4126282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s for node %d\n", 412710fbcf4cSKay Sievers h->name, node->dev.id); 41289a305230SLee Schermerhorn hugetlb_unregister_node(node); 41299a305230SLee Schermerhorn break; 41309a305230SLee Schermerhorn } 41319a305230SLee Schermerhorn } 41329a305230SLee Schermerhorn } 41339a305230SLee Schermerhorn 41349a305230SLee Schermerhorn /* 41359b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 413610fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 413710fbcf4cSKay Sievers * registered their associated device by this time. 41389a305230SLee Schermerhorn */ 41397d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 41409a305230SLee Schermerhorn { 41419a305230SLee Schermerhorn int nid; 41429a305230SLee Schermerhorn 4143a4a00b45SMuchun Song for_each_online_node(nid) 4144b958d4d0SMuchun Song hugetlb_register_node(node_devices[nid]); 41459a305230SLee Schermerhorn } 41469a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 41479a305230SLee Schermerhorn 41489a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 41499a305230SLee Schermerhorn { 41509a305230SLee Schermerhorn BUG(); 41519a305230SLee Schermerhorn if (nidp) 41529a305230SLee Schermerhorn *nidp = -1; 41539a305230SLee Schermerhorn return NULL; 41549a305230SLee Schermerhorn } 41559a305230SLee Schermerhorn 41569a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 41579a305230SLee Schermerhorn 41589a305230SLee Schermerhorn #endif 41599a305230SLee Schermerhorn 4160263b8998SMiaohe Lin #ifdef CONFIG_CMA 4161263b8998SMiaohe Lin static void __init hugetlb_cma_check(void); 4162263b8998SMiaohe Lin #else 4163263b8998SMiaohe Lin static inline __init void hugetlb_cma_check(void) 4164263b8998SMiaohe Lin { 4165263b8998SMiaohe Lin } 4166263b8998SMiaohe Lin #endif 4167263b8998SMiaohe Lin 4168a4a00b45SMuchun Song static void __init hugetlb_sysfs_init(void) 4169a4a00b45SMuchun Song { 4170a4a00b45SMuchun Song struct hstate *h; 4171a4a00b45SMuchun Song int err; 4172a4a00b45SMuchun Song 4173a4a00b45SMuchun Song hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 4174a4a00b45SMuchun Song if (!hugepages_kobj) 4175a4a00b45SMuchun Song return; 4176a4a00b45SMuchun Song 4177a4a00b45SMuchun Song for_each_hstate(h) { 4178a4a00b45SMuchun Song err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 4179a4a00b45SMuchun Song hstate_kobjs, &hstate_attr_group); 4180a4a00b45SMuchun Song if (err) 4181a4a00b45SMuchun Song pr_err("HugeTLB: Unable to add hstate %s", h->name); 4182a4a00b45SMuchun Song } 4183a4a00b45SMuchun Song 4184a4a00b45SMuchun Song #ifdef CONFIG_NUMA 4185a4a00b45SMuchun Song hugetlb_sysfs_initialized = true; 4186a4a00b45SMuchun Song #endif 4187a4a00b45SMuchun Song hugetlb_register_all_nodes(); 4188a4a00b45SMuchun Song } 4189a4a00b45SMuchun Song 4190962de548SKefeng Wang #ifdef CONFIG_SYSCTL 4191962de548SKefeng Wang static void hugetlb_sysctl_init(void); 4192962de548SKefeng Wang #else 4193962de548SKefeng Wang static inline void hugetlb_sysctl_init(void) { } 4194962de548SKefeng Wang #endif 4195962de548SKefeng Wang 4196a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 4197a3437870SNishanth Aravamudan { 41988382d914SDavidlohr Bueso int i; 41998382d914SDavidlohr Bueso 4200d6995da3SMike Kravetz BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < 4201d6995da3SMike Kravetz __NR_HPAGEFLAGS); 4202d6995da3SMike Kravetz 4203c2833a5bSMike Kravetz if (!hugepages_supported()) { 4204c2833a5bSMike Kravetz if (hugetlb_max_hstate || default_hstate_max_huge_pages) 4205c2833a5bSMike Kravetz pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); 42060ef89d25SBenjamin Herrenschmidt return 0; 4207d715cf80SLiam R. Howlett } 4208d715cf80SLiam R. Howlett 4209282f4214SMike Kravetz /* 4210282f4214SMike Kravetz * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some 4211282f4214SMike Kravetz * architectures depend on setup being done here. 4212282f4214SMike Kravetz */ 4213a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 4214282f4214SMike Kravetz if (!parsed_default_hugepagesz) { 4215282f4214SMike Kravetz /* 4216282f4214SMike Kravetz * If we did not parse a default huge page size, set 4217282f4214SMike Kravetz * default_hstate_idx to HPAGE_SIZE hstate. And, if the 4218282f4214SMike Kravetz * number of huge pages for this default size was implicitly 4219282f4214SMike Kravetz * specified, set that here as well. 4220282f4214SMike Kravetz * Note that the implicit setting will overwrite an explicit 4221282f4214SMike Kravetz * setting. A warning will be printed in this case. 4222282f4214SMike Kravetz */ 4223282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); 4224f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 4225282f4214SMike Kravetz if (default_hstate.max_huge_pages) { 4226282f4214SMike Kravetz char buf[32]; 4227282f4214SMike Kravetz 4228282f4214SMike Kravetz string_get_size(huge_page_size(&default_hstate), 4229282f4214SMike Kravetz 1, STRING_UNITS_2, buf, 32); 4230282f4214SMike Kravetz pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", 4231282f4214SMike Kravetz default_hstate.max_huge_pages, buf); 4232282f4214SMike Kravetz pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", 4233282f4214SMike Kravetz default_hstate_max_huge_pages); 4234282f4214SMike Kravetz } 4235282f4214SMike Kravetz default_hstate.max_huge_pages = 4236282f4214SMike Kravetz default_hstate_max_huge_pages; 4237b5389086SZhenguo Yao 42380a7a0f6fSPeng Liu for_each_online_node(i) 4239b5389086SZhenguo Yao default_hstate.max_huge_pages_node[i] = 4240b5389086SZhenguo Yao default_hugepages_in_node[i]; 4241282f4214SMike Kravetz } 4242f8b74815SVaishali Thakkar } 4243a3437870SNishanth Aravamudan 4244cf11e85fSRoman Gushchin hugetlb_cma_check(); 4245a3437870SNishanth Aravamudan hugetlb_init_hstates(); 4246aa888a74SAndi Kleen gather_bootmem_prealloc(); 4247a3437870SNishanth Aravamudan report_hugepages(); 4248a3437870SNishanth Aravamudan 4249a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 42507179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 4251962de548SKefeng Wang hugetlb_sysctl_init(); 42529a305230SLee Schermerhorn 42538382d914SDavidlohr Bueso #ifdef CONFIG_SMP 42548382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 42558382d914SDavidlohr Bueso #else 42568382d914SDavidlohr Bueso num_fault_mutexes = 1; 42578382d914SDavidlohr Bueso #endif 4258c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 42596da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 42606da2ec56SKees Cook GFP_KERNEL); 4261c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 42628382d914SDavidlohr Bueso 42638382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 4264c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 4265a3437870SNishanth Aravamudan return 0; 4266a3437870SNishanth Aravamudan } 42673e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 4268a3437870SNishanth Aravamudan 4269ae94da89SMike Kravetz /* Overwritten by architectures with more huge page sizes */ 4270ae94da89SMike Kravetz bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) 42719fee021dSVaishali Thakkar { 4272ae94da89SMike Kravetz return size == HPAGE_SIZE; 42739fee021dSVaishali Thakkar } 42749fee021dSVaishali Thakkar 4275d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 4276a3437870SNishanth Aravamudan { 4277a3437870SNishanth Aravamudan struct hstate *h; 42788faa8b07SAndi Kleen unsigned long i; 42798faa8b07SAndi Kleen 4280a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 4281a3437870SNishanth Aravamudan return; 4282a3437870SNishanth Aravamudan } 428347d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 4284a3437870SNishanth Aravamudan BUG_ON(order == 0); 428547d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 428629383967SMike Kravetz mutex_init(&h->resize_lock); 4287a3437870SNishanth Aravamudan h->order = order; 4288aca78307SMiaohe Lin h->mask = ~(huge_page_size(h) - 1); 42898faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 42908faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 42910edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 429254f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 429354f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 4294a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 4295c2c3a60aSMiaohe Lin huge_page_size(h)/SZ_1K); 42968faa8b07SAndi Kleen 4297a3437870SNishanth Aravamudan parsed_hstate = h; 4298a3437870SNishanth Aravamudan } 4299a3437870SNishanth Aravamudan 4300b5389086SZhenguo Yao bool __init __weak hugetlb_node_alloc_supported(void) 4301b5389086SZhenguo Yao { 4302b5389086SZhenguo Yao return true; 4303b5389086SZhenguo Yao } 4304f87442f4SPeng Liu 4305f87442f4SPeng Liu static void __init hugepages_clear_pages_in_node(void) 4306f87442f4SPeng Liu { 4307f87442f4SPeng Liu if (!hugetlb_max_hstate) { 4308f87442f4SPeng Liu default_hstate_max_huge_pages = 0; 4309f87442f4SPeng Liu memset(default_hugepages_in_node, 0, 431010395680SMiaohe Lin sizeof(default_hugepages_in_node)); 4311f87442f4SPeng Liu } else { 4312f87442f4SPeng Liu parsed_hstate->max_huge_pages = 0; 4313f87442f4SPeng Liu memset(parsed_hstate->max_huge_pages_node, 0, 431410395680SMiaohe Lin sizeof(parsed_hstate->max_huge_pages_node)); 4315f87442f4SPeng Liu } 4316f87442f4SPeng Liu } 4317f87442f4SPeng Liu 4318282f4214SMike Kravetz /* 4319282f4214SMike Kravetz * hugepages command line processing 4320282f4214SMike Kravetz * hugepages normally follows a valid hugepagsz or default_hugepagsz 4321282f4214SMike Kravetz * specification. If not, ignore the hugepages value. hugepages can also 4322282f4214SMike Kravetz * be the first huge page command line option in which case it implicitly 4323282f4214SMike Kravetz * specifies the number of huge pages for the default size. 4324282f4214SMike Kravetz */ 4325282f4214SMike Kravetz static int __init hugepages_setup(char *s) 4326a3437870SNishanth Aravamudan { 4327a3437870SNishanth Aravamudan unsigned long *mhp; 43288faa8b07SAndi Kleen static unsigned long *last_mhp; 4329b5389086SZhenguo Yao int node = NUMA_NO_NODE; 4330b5389086SZhenguo Yao int count; 4331b5389086SZhenguo Yao unsigned long tmp; 4332b5389086SZhenguo Yao char *p = s; 4333a3437870SNishanth Aravamudan 43349fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 4335282f4214SMike Kravetz pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); 43369fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 4337f81f6e4bSPeng Liu return 1; 43389fee021dSVaishali Thakkar } 4339282f4214SMike Kravetz 4340a3437870SNishanth Aravamudan /* 4341282f4214SMike Kravetz * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter 4342282f4214SMike Kravetz * yet, so this hugepages= parameter goes to the "default hstate". 4343282f4214SMike Kravetz * Otherwise, it goes with the previously parsed hugepagesz or 4344282f4214SMike Kravetz * default_hugepagesz. 4345a3437870SNishanth Aravamudan */ 43469fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 4347a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 4348a3437870SNishanth Aravamudan else 4349a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 4350a3437870SNishanth Aravamudan 43518faa8b07SAndi Kleen if (mhp == last_mhp) { 4352282f4214SMike Kravetz pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); 4353f81f6e4bSPeng Liu return 1; 43548faa8b07SAndi Kleen } 43558faa8b07SAndi Kleen 4356b5389086SZhenguo Yao while (*p) { 4357b5389086SZhenguo Yao count = 0; 4358b5389086SZhenguo Yao if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4359b5389086SZhenguo Yao goto invalid; 4360b5389086SZhenguo Yao /* Parameter is node format */ 4361b5389086SZhenguo Yao if (p[count] == ':') { 4362b5389086SZhenguo Yao if (!hugetlb_node_alloc_supported()) { 4363b5389086SZhenguo Yao pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n"); 4364f81f6e4bSPeng Liu return 1; 4365b5389086SZhenguo Yao } 43660a7a0f6fSPeng Liu if (tmp >= MAX_NUMNODES || !node_online(tmp)) 4367e79ce983SLiu Yuntao goto invalid; 43680a7a0f6fSPeng Liu node = array_index_nospec(tmp, MAX_NUMNODES); 4369b5389086SZhenguo Yao p += count + 1; 4370b5389086SZhenguo Yao /* Parse hugepages */ 4371b5389086SZhenguo Yao if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4372b5389086SZhenguo Yao goto invalid; 4373b5389086SZhenguo Yao if (!hugetlb_max_hstate) 4374b5389086SZhenguo Yao default_hugepages_in_node[node] = tmp; 4375b5389086SZhenguo Yao else 4376b5389086SZhenguo Yao parsed_hstate->max_huge_pages_node[node] = tmp; 4377b5389086SZhenguo Yao *mhp += tmp; 4378b5389086SZhenguo Yao /* Go to parse next node*/ 4379b5389086SZhenguo Yao if (p[count] == ',') 4380b5389086SZhenguo Yao p += count + 1; 4381b5389086SZhenguo Yao else 4382b5389086SZhenguo Yao break; 4383b5389086SZhenguo Yao } else { 4384b5389086SZhenguo Yao if (p != s) 4385b5389086SZhenguo Yao goto invalid; 4386b5389086SZhenguo Yao *mhp = tmp; 4387b5389086SZhenguo Yao break; 4388b5389086SZhenguo Yao } 4389b5389086SZhenguo Yao } 4390a3437870SNishanth Aravamudan 43918faa8b07SAndi Kleen /* 43928faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 439304adbc3fSMiaohe Lin * But we need to allocate gigantic hstates here early to still 43948faa8b07SAndi Kleen * use the bootmem allocator. 43958faa8b07SAndi Kleen */ 439604adbc3fSMiaohe Lin if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) 43978faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 43988faa8b07SAndi Kleen 43998faa8b07SAndi Kleen last_mhp = mhp; 44008faa8b07SAndi Kleen 4401a3437870SNishanth Aravamudan return 1; 4402b5389086SZhenguo Yao 4403b5389086SZhenguo Yao invalid: 4404b5389086SZhenguo Yao pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p); 4405f87442f4SPeng Liu hugepages_clear_pages_in_node(); 4406f81f6e4bSPeng Liu return 1; 4407a3437870SNishanth Aravamudan } 4408282f4214SMike Kravetz __setup("hugepages=", hugepages_setup); 4409e11bfbfcSNick Piggin 4410282f4214SMike Kravetz /* 4411282f4214SMike Kravetz * hugepagesz command line processing 4412282f4214SMike Kravetz * A specific huge page size can only be specified once with hugepagesz. 4413282f4214SMike Kravetz * hugepagesz is followed by hugepages on the command line. The global 4414282f4214SMike Kravetz * variable 'parsed_valid_hugepagesz' is used to determine if prior 4415282f4214SMike Kravetz * hugepagesz argument was valid. 4416282f4214SMike Kravetz */ 4417359f2544SMike Kravetz static int __init hugepagesz_setup(char *s) 4418e11bfbfcSNick Piggin { 4419359f2544SMike Kravetz unsigned long size; 4420282f4214SMike Kravetz struct hstate *h; 4421282f4214SMike Kravetz 4422282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4423359f2544SMike Kravetz size = (unsigned long)memparse(s, NULL); 4424359f2544SMike Kravetz 4425359f2544SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4426282f4214SMike Kravetz pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); 4427f81f6e4bSPeng Liu return 1; 4428359f2544SMike Kravetz } 4429359f2544SMike Kravetz 4430282f4214SMike Kravetz h = size_to_hstate(size); 4431282f4214SMike Kravetz if (h) { 4432282f4214SMike Kravetz /* 4433282f4214SMike Kravetz * hstate for this size already exists. This is normally 4434282f4214SMike Kravetz * an error, but is allowed if the existing hstate is the 4435282f4214SMike Kravetz * default hstate. More specifically, it is only allowed if 4436282f4214SMike Kravetz * the number of huge pages for the default hstate was not 4437282f4214SMike Kravetz * previously specified. 4438282f4214SMike Kravetz */ 4439282f4214SMike Kravetz if (!parsed_default_hugepagesz || h != &default_hstate || 4440282f4214SMike Kravetz default_hstate.max_huge_pages) { 4441282f4214SMike Kravetz pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); 4442f81f6e4bSPeng Liu return 1; 444338237830SMike Kravetz } 444438237830SMike Kravetz 4445282f4214SMike Kravetz /* 4446282f4214SMike Kravetz * No need to call hugetlb_add_hstate() as hstate already 4447282f4214SMike Kravetz * exists. But, do set parsed_hstate so that a following 4448282f4214SMike Kravetz * hugepages= parameter will be applied to this hstate. 4449282f4214SMike Kravetz */ 4450282f4214SMike Kravetz parsed_hstate = h; 4451282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4452e11bfbfcSNick Piggin return 1; 4453e11bfbfcSNick Piggin } 4454282f4214SMike Kravetz 4455359f2544SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4456282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4457359f2544SMike Kravetz return 1; 4458359f2544SMike Kravetz } 4459359f2544SMike Kravetz __setup("hugepagesz=", hugepagesz_setup); 4460359f2544SMike Kravetz 4461282f4214SMike Kravetz /* 4462282f4214SMike Kravetz * default_hugepagesz command line input 4463282f4214SMike Kravetz * Only one instance of default_hugepagesz allowed on command line. 4464282f4214SMike Kravetz */ 4465ae94da89SMike Kravetz static int __init default_hugepagesz_setup(char *s) 4466e11bfbfcSNick Piggin { 4467ae94da89SMike Kravetz unsigned long size; 4468b5389086SZhenguo Yao int i; 4469ae94da89SMike Kravetz 4470282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4471282f4214SMike Kravetz if (parsed_default_hugepagesz) { 4472282f4214SMike Kravetz pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); 4473f81f6e4bSPeng Liu return 1; 4474282f4214SMike Kravetz } 4475282f4214SMike Kravetz 4476282f4214SMike Kravetz size = (unsigned long)memparse(s, NULL); 4477282f4214SMike Kravetz 4478282f4214SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4479282f4214SMike Kravetz pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); 4480f81f6e4bSPeng Liu return 1; 4481282f4214SMike Kravetz } 4482282f4214SMike Kravetz 4483282f4214SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4484282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4485282f4214SMike Kravetz parsed_default_hugepagesz = true; 4486282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(size)); 4487282f4214SMike Kravetz 4488282f4214SMike Kravetz /* 4489282f4214SMike Kravetz * The number of default huge pages (for this size) could have been 4490282f4214SMike Kravetz * specified as the first hugetlb parameter: hugepages=X. If so, 4491282f4214SMike Kravetz * then default_hstate_max_huge_pages is set. If the default huge 449223baf831SKirill A. Shutemov * page size is gigantic (> MAX_ORDER), then the pages must be 4493282f4214SMike Kravetz * allocated here from bootmem allocator. 4494282f4214SMike Kravetz */ 4495282f4214SMike Kravetz if (default_hstate_max_huge_pages) { 4496282f4214SMike Kravetz default_hstate.max_huge_pages = default_hstate_max_huge_pages; 44970a7a0f6fSPeng Liu for_each_online_node(i) 4498b5389086SZhenguo Yao default_hstate.max_huge_pages_node[i] = 4499b5389086SZhenguo Yao default_hugepages_in_node[i]; 4500282f4214SMike Kravetz if (hstate_is_gigantic(&default_hstate)) 4501282f4214SMike Kravetz hugetlb_hstate_alloc_pages(&default_hstate); 4502282f4214SMike Kravetz default_hstate_max_huge_pages = 0; 4503282f4214SMike Kravetz } 4504282f4214SMike Kravetz 4505e11bfbfcSNick Piggin return 1; 4506e11bfbfcSNick Piggin } 4507ae94da89SMike Kravetz __setup("default_hugepagesz=", default_hugepagesz_setup); 4508a3437870SNishanth Aravamudan 4509d2226ebdSFeng Tang static nodemask_t *policy_mbind_nodemask(gfp_t gfp) 4510d2226ebdSFeng Tang { 4511d2226ebdSFeng Tang #ifdef CONFIG_NUMA 4512d2226ebdSFeng Tang struct mempolicy *mpol = get_task_policy(current); 4513d2226ebdSFeng Tang 4514d2226ebdSFeng Tang /* 4515d2226ebdSFeng Tang * Only enforce MPOL_BIND policy which overlaps with cpuset policy 4516d2226ebdSFeng Tang * (from policy_nodemask) specifically for hugetlb case 4517d2226ebdSFeng Tang */ 4518d2226ebdSFeng Tang if (mpol->mode == MPOL_BIND && 4519d2226ebdSFeng Tang (apply_policy_zone(mpol, gfp_zone(gfp)) && 4520d2226ebdSFeng Tang cpuset_nodemask_valid_mems_allowed(&mpol->nodes))) 4521d2226ebdSFeng Tang return &mpol->nodes; 4522d2226ebdSFeng Tang #endif 4523d2226ebdSFeng Tang return NULL; 4524d2226ebdSFeng Tang } 4525d2226ebdSFeng Tang 45268ca39e68SMuchun Song static unsigned int allowed_mems_nr(struct hstate *h) 45278a213460SNishanth Aravamudan { 45288a213460SNishanth Aravamudan int node; 45298a213460SNishanth Aravamudan unsigned int nr = 0; 4530d2226ebdSFeng Tang nodemask_t *mbind_nodemask; 45318ca39e68SMuchun Song unsigned int *array = h->free_huge_pages_node; 45328ca39e68SMuchun Song gfp_t gfp_mask = htlb_alloc_mask(h); 45338a213460SNishanth Aravamudan 4534d2226ebdSFeng Tang mbind_nodemask = policy_mbind_nodemask(gfp_mask); 45358ca39e68SMuchun Song for_each_node_mask(node, cpuset_current_mems_allowed) { 4536d2226ebdSFeng Tang if (!mbind_nodemask || node_isset(node, *mbind_nodemask)) 45378a213460SNishanth Aravamudan nr += array[node]; 45388ca39e68SMuchun Song } 45398a213460SNishanth Aravamudan 45408a213460SNishanth Aravamudan return nr; 45418a213460SNishanth Aravamudan } 45428a213460SNishanth Aravamudan 45438a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 454417743798SMuchun Song static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, 454517743798SMuchun Song void *buffer, size_t *length, 454617743798SMuchun Song loff_t *ppos, unsigned long *out) 454717743798SMuchun Song { 454817743798SMuchun Song struct ctl_table dup_table; 454917743798SMuchun Song 455017743798SMuchun Song /* 455117743798SMuchun Song * In order to avoid races with __do_proc_doulongvec_minmax(), we 455217743798SMuchun Song * can duplicate the @table and alter the duplicate of it. 455317743798SMuchun Song */ 455417743798SMuchun Song dup_table = *table; 455517743798SMuchun Song dup_table.data = out; 455617743798SMuchun Song 455717743798SMuchun Song return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); 455817743798SMuchun Song } 455917743798SMuchun Song 456006808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 456106808b08SLee Schermerhorn struct ctl_table *table, int write, 456232927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 45631da177e4SLinus Torvalds { 4564e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 4565238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 456608d4a246SMichal Hocko int ret; 4567e5ff2159SAndi Kleen 4568457c1b27SNishanth Aravamudan if (!hugepages_supported()) 456986613628SJan Stancek return -EOPNOTSUPP; 4570457c1b27SNishanth Aravamudan 457117743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 457217743798SMuchun Song &tmp); 457308d4a246SMichal Hocko if (ret) 457408d4a246SMichal Hocko goto out; 4575e5ff2159SAndi Kleen 4576238d3c13SDavid Rientjes if (write) 4577238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 4578238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 457908d4a246SMichal Hocko out: 458008d4a246SMichal Hocko return ret; 45811da177e4SLinus Torvalds } 4582396faf03SMel Gorman 4583962de548SKefeng Wang static int hugetlb_sysctl_handler(struct ctl_table *table, int write, 458432927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 458506808b08SLee Schermerhorn { 458606808b08SLee Schermerhorn 458706808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 458806808b08SLee Schermerhorn buffer, length, ppos); 458906808b08SLee Schermerhorn } 459006808b08SLee Schermerhorn 459106808b08SLee Schermerhorn #ifdef CONFIG_NUMA 4592962de548SKefeng Wang static int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 459332927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 459406808b08SLee Schermerhorn { 459506808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 459606808b08SLee Schermerhorn buffer, length, ppos); 459706808b08SLee Schermerhorn } 459806808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 459906808b08SLee Schermerhorn 4600962de548SKefeng Wang static int hugetlb_overcommit_handler(struct ctl_table *table, int write, 460132927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 4602a3d0c6aaSNishanth Aravamudan { 4603a5516438SAndi Kleen struct hstate *h = &default_hstate; 4604e5ff2159SAndi Kleen unsigned long tmp; 460508d4a246SMichal Hocko int ret; 4606e5ff2159SAndi Kleen 4607457c1b27SNishanth Aravamudan if (!hugepages_supported()) 460886613628SJan Stancek return -EOPNOTSUPP; 4609457c1b27SNishanth Aravamudan 4610e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 4611e5ff2159SAndi Kleen 4612bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 4613adbe8726SEric B Munson return -EINVAL; 4614adbe8726SEric B Munson 461517743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 461617743798SMuchun Song &tmp); 461708d4a246SMichal Hocko if (ret) 461808d4a246SMichal Hocko goto out; 4619e5ff2159SAndi Kleen 4620e5ff2159SAndi Kleen if (write) { 4621db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4622e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 4623db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4624e5ff2159SAndi Kleen } 462508d4a246SMichal Hocko out: 462608d4a246SMichal Hocko return ret; 4627a3d0c6aaSNishanth Aravamudan } 4628a3d0c6aaSNishanth Aravamudan 4629962de548SKefeng Wang static struct ctl_table hugetlb_table[] = { 4630962de548SKefeng Wang { 4631962de548SKefeng Wang .procname = "nr_hugepages", 4632962de548SKefeng Wang .data = NULL, 4633962de548SKefeng Wang .maxlen = sizeof(unsigned long), 4634962de548SKefeng Wang .mode = 0644, 4635962de548SKefeng Wang .proc_handler = hugetlb_sysctl_handler, 4636962de548SKefeng Wang }, 4637962de548SKefeng Wang #ifdef CONFIG_NUMA 4638962de548SKefeng Wang { 4639962de548SKefeng Wang .procname = "nr_hugepages_mempolicy", 4640962de548SKefeng Wang .data = NULL, 4641962de548SKefeng Wang .maxlen = sizeof(unsigned long), 4642962de548SKefeng Wang .mode = 0644, 4643962de548SKefeng Wang .proc_handler = &hugetlb_mempolicy_sysctl_handler, 4644962de548SKefeng Wang }, 4645962de548SKefeng Wang #endif 4646962de548SKefeng Wang { 4647962de548SKefeng Wang .procname = "hugetlb_shm_group", 4648962de548SKefeng Wang .data = &sysctl_hugetlb_shm_group, 4649962de548SKefeng Wang .maxlen = sizeof(gid_t), 4650962de548SKefeng Wang .mode = 0644, 4651962de548SKefeng Wang .proc_handler = proc_dointvec, 4652962de548SKefeng Wang }, 4653962de548SKefeng Wang { 4654962de548SKefeng Wang .procname = "nr_overcommit_hugepages", 4655962de548SKefeng Wang .data = NULL, 4656962de548SKefeng Wang .maxlen = sizeof(unsigned long), 4657962de548SKefeng Wang .mode = 0644, 4658962de548SKefeng Wang .proc_handler = hugetlb_overcommit_handler, 4659962de548SKefeng Wang }, 4660962de548SKefeng Wang { } 4661962de548SKefeng Wang }; 4662962de548SKefeng Wang 4663962de548SKefeng Wang static void hugetlb_sysctl_init(void) 4664962de548SKefeng Wang { 4665962de548SKefeng Wang register_sysctl_init("vm", hugetlb_table); 4666962de548SKefeng Wang } 46671da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 46681da177e4SLinus Torvalds 4669e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 46701da177e4SLinus Torvalds { 4671fcb2b0c5SRoman Gushchin struct hstate *h; 4672fcb2b0c5SRoman Gushchin unsigned long total = 0; 4673fcb2b0c5SRoman Gushchin 4674457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4675457c1b27SNishanth Aravamudan return; 4676fcb2b0c5SRoman Gushchin 4677fcb2b0c5SRoman Gushchin for_each_hstate(h) { 4678fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 4679fcb2b0c5SRoman Gushchin 4680aca78307SMiaohe Lin total += huge_page_size(h) * count; 4681fcb2b0c5SRoman Gushchin 4682fcb2b0c5SRoman Gushchin if (h == &default_hstate) 4683e1759c21SAlexey Dobriyan seq_printf(m, 46841da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 46851da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 4686b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 46877893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 46884f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 4689fcb2b0c5SRoman Gushchin count, 4690a5516438SAndi Kleen h->free_huge_pages, 4691a5516438SAndi Kleen h->resv_huge_pages, 4692a5516438SAndi Kleen h->surplus_huge_pages, 4693aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4694fcb2b0c5SRoman Gushchin } 4695fcb2b0c5SRoman Gushchin 4696aca78307SMiaohe Lin seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); 46971da177e4SLinus Torvalds } 46981da177e4SLinus Torvalds 46997981593bSJoe Perches int hugetlb_report_node_meminfo(char *buf, int len, int nid) 47001da177e4SLinus Torvalds { 4701a5516438SAndi Kleen struct hstate *h = &default_hstate; 47027981593bSJoe Perches 4703457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4704457c1b27SNishanth Aravamudan return 0; 47057981593bSJoe Perches 47067981593bSJoe Perches return sysfs_emit_at(buf, len, 47071da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 4708a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 4709a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 4710a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 4711a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 4712a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 47131da177e4SLinus Torvalds } 47141da177e4SLinus Torvalds 4715dcadcf1cSGang Li void hugetlb_show_meminfo_node(int nid) 4716949f7ec5SDavid Rientjes { 4717949f7ec5SDavid Rientjes struct hstate *h; 4718949f7ec5SDavid Rientjes 4719457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4720457c1b27SNishanth Aravamudan return; 4721457c1b27SNishanth Aravamudan 4722949f7ec5SDavid Rientjes for_each_hstate(h) 4723dcadcf1cSGang Li printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 4724949f7ec5SDavid Rientjes nid, 4725949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 4726949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 4727949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 4728aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4729949f7ec5SDavid Rientjes } 4730949f7ec5SDavid Rientjes 47315d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 47325d317b2bSNaoya Horiguchi { 47335d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 47346c1aa2d3SZhangPeng K(atomic_long_read(&mm->hugetlb_usage))); 47355d317b2bSNaoya Horiguchi } 47365d317b2bSNaoya Horiguchi 47371da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 47381da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 47391da177e4SLinus Torvalds { 4740d0028588SWanpeng Li struct hstate *h; 4741d0028588SWanpeng Li unsigned long nr_total_pages = 0; 4742d0028588SWanpeng Li 4743d0028588SWanpeng Li for_each_hstate(h) 4744d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 4745d0028588SWanpeng Li return nr_total_pages; 47461da177e4SLinus Torvalds } 47471da177e4SLinus Torvalds 4748a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 4749fc1b8a73SMel Gorman { 4750fc1b8a73SMel Gorman int ret = -ENOMEM; 4751fc1b8a73SMel Gorman 47520aa7f354SMiaohe Lin if (!delta) 47530aa7f354SMiaohe Lin return 0; 47540aa7f354SMiaohe Lin 4755db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4756fc1b8a73SMel Gorman /* 4757fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 4758fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 4759fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 4760fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 4761fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 4762fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 4763fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 4764fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 4765fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 4766fc1b8a73SMel Gorman * 4767fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 4768fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 4769fc1b8a73SMel Gorman * we fall back to check against current free page availability as 4770fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 4771fc1b8a73SMel Gorman * semantics that cpuset has. 47728ca39e68SMuchun Song * 47738ca39e68SMuchun Song * Apart from cpuset, we also have memory policy mechanism that 47748ca39e68SMuchun Song * also determines from which node the kernel will allocate memory 47758ca39e68SMuchun Song * in a NUMA system. So similar to cpuset, we also should consider 47768ca39e68SMuchun Song * the memory policy of the current task. Similar to the description 47778ca39e68SMuchun Song * above. 4778fc1b8a73SMel Gorman */ 4779fc1b8a73SMel Gorman if (delta > 0) { 4780a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 4781fc1b8a73SMel Gorman goto out; 4782fc1b8a73SMel Gorman 47838ca39e68SMuchun Song if (delta > allowed_mems_nr(h)) { 4784a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 4785fc1b8a73SMel Gorman goto out; 4786fc1b8a73SMel Gorman } 4787fc1b8a73SMel Gorman } 4788fc1b8a73SMel Gorman 4789fc1b8a73SMel Gorman ret = 0; 4790fc1b8a73SMel Gorman if (delta < 0) 4791a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 4792fc1b8a73SMel Gorman 4793fc1b8a73SMel Gorman out: 4794db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4795fc1b8a73SMel Gorman return ret; 4796fc1b8a73SMel Gorman } 4797fc1b8a73SMel Gorman 479884afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 479984afd99bSAndy Whitcroft { 4800f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 480184afd99bSAndy Whitcroft 480284afd99bSAndy Whitcroft /* 4803612b8a31SMike Kravetz * HPAGE_RESV_OWNER indicates a private mapping. 480484afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 480584afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 480684afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 480725985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 480884afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 480984afd99bSAndy Whitcroft * new reference here without additional locking. 481084afd99bSAndy Whitcroft */ 481109a26e83SMike Kravetz if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 481209a26e83SMike Kravetz resv_map_dup_hugetlb_cgroup_uncharge_info(resv); 4813f522c3acSJoonsoo Kim kref_get(&resv->refs); 481484afd99bSAndy Whitcroft } 48158d9bfb26SMike Kravetz 4816131a79b4SMike Kravetz /* 4817131a79b4SMike Kravetz * vma_lock structure for sharable mappings is vma specific. 4818612b8a31SMike Kravetz * Clear old pointer (if copied via vm_area_dup) and allocate 4819612b8a31SMike Kravetz * new structure. Before clearing, make sure vma_lock is not 4820612b8a31SMike Kravetz * for this vma. 4821131a79b4SMike Kravetz */ 4822131a79b4SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 4823612b8a31SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 4824612b8a31SMike Kravetz 4825612b8a31SMike Kravetz if (vma_lock) { 4826612b8a31SMike Kravetz if (vma_lock->vma != vma) { 4827131a79b4SMike Kravetz vma->vm_private_data = NULL; 48288d9bfb26SMike Kravetz hugetlb_vma_lock_alloc(vma); 4829612b8a31SMike Kravetz } else 4830612b8a31SMike Kravetz pr_warn("HugeTLB: vma_lock already exists in %s.\n", __func__); 4831612b8a31SMike Kravetz } else 4832612b8a31SMike Kravetz hugetlb_vma_lock_alloc(vma); 483309a26e83SMike Kravetz } 4834131a79b4SMike Kravetz } 483584afd99bSAndy Whitcroft 4836a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 4837a1e78772SMel Gorman { 4838a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 48398d9bfb26SMike Kravetz struct resv_map *resv; 484090481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 48414e35f483SJoonsoo Kim unsigned long reserve, start, end; 48421c5ecae3SMike Kravetz long gbl_reserve; 484384afd99bSAndy Whitcroft 48448d9bfb26SMike Kravetz hugetlb_vma_lock_free(vma); 48458d9bfb26SMike Kravetz 48468d9bfb26SMike Kravetz resv = vma_resv_map(vma); 48474e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 48484e35f483SJoonsoo Kim return; 48494e35f483SJoonsoo Kim 4850a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 4851a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 485284afd99bSAndy Whitcroft 48534e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 4854e9fe92aeSMina Almasry hugetlb_cgroup_uncharge_counter(resv, start, end); 48557251ff78SAdam Litke if (reserve) { 48561c5ecae3SMike Kravetz /* 48571c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 48581c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 48591c5ecae3SMike Kravetz */ 48601c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 48611c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 48627251ff78SAdam Litke } 4863e9fe92aeSMina Almasry 4864e9fe92aeSMina Almasry kref_put(&resv->refs, resv_map_release); 4865a1e78772SMel Gorman } 4866a1e78772SMel Gorman 486731383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 486831383c68SDan Williams { 486931383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 487031383c68SDan Williams return -EINVAL; 4871b30c14cdSJames Houghton 4872b30c14cdSJames Houghton /* 4873b30c14cdSJames Houghton * PMD sharing is only possible for PUD_SIZE-aligned address ranges 4874b30c14cdSJames Houghton * in HugeTLB VMAs. If we will lose PUD_SIZE alignment due to this 4875b30c14cdSJames Houghton * split, unshare PMDs in the PUD_SIZE interval surrounding addr now. 4876b30c14cdSJames Houghton */ 4877b30c14cdSJames Houghton if (addr & ~PUD_MASK) { 4878b30c14cdSJames Houghton /* 4879b30c14cdSJames Houghton * hugetlb_vm_op_split is called right before we attempt to 4880b30c14cdSJames Houghton * split the VMA. We will need to unshare PMDs in the old and 4881b30c14cdSJames Houghton * new VMAs, so let's unshare before we split. 4882b30c14cdSJames Houghton */ 4883b30c14cdSJames Houghton unsigned long floor = addr & PUD_MASK; 4884b30c14cdSJames Houghton unsigned long ceil = floor + PUD_SIZE; 4885b30c14cdSJames Houghton 4886b30c14cdSJames Houghton if (floor >= vma->vm_start && ceil <= vma->vm_end) 4887b30c14cdSJames Houghton hugetlb_unshare_pmds(vma, floor, ceil); 4888b30c14cdSJames Houghton } 4889b30c14cdSJames Houghton 489031383c68SDan Williams return 0; 489131383c68SDan Williams } 489231383c68SDan Williams 489305ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 489405ea8860SDan Williams { 4895aca78307SMiaohe Lin return huge_page_size(hstate_vma(vma)); 489605ea8860SDan Williams } 489705ea8860SDan Williams 48981da177e4SLinus Torvalds /* 48991da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 49001da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 49016c26d310SMiaohe Lin * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get 49021da177e4SLinus Torvalds * this far. 49031da177e4SLinus Torvalds */ 4904b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 49051da177e4SLinus Torvalds { 49061da177e4SLinus Torvalds BUG(); 4907d0217ac0SNick Piggin return 0; 49081da177e4SLinus Torvalds } 49091da177e4SLinus Torvalds 4910eec3636aSJane Chu /* 4911eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 4912eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 4913eec3636aSJane Chu * This is because under System V memory model, mappings created via 4914eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 4915eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 4916eec3636aSJane Chu */ 4917f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 4918d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 491984afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 4920a1e78772SMel Gorman .close = hugetlb_vm_op_close, 4921dd3b614fSDmitry Safonov .may_split = hugetlb_vm_op_split, 492205ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 49231da177e4SLinus Torvalds }; 49241da177e4SLinus Torvalds 49251e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 49261e8f889bSDavid Gibson int writable) 492763551ae0SDavid Gibson { 492863551ae0SDavid Gibson pte_t entry; 492979c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 493063551ae0SDavid Gibson 49311e8f889bSDavid Gibson if (writable) { 4932106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 4933106c992aSGerald Schaefer vma->vm_page_prot))); 493463551ae0SDavid Gibson } else { 4935106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 4936106c992aSGerald Schaefer vma->vm_page_prot)); 493763551ae0SDavid Gibson } 493863551ae0SDavid Gibson entry = pte_mkyoung(entry); 493979c1c594SChristophe Leroy entry = arch_make_huge_pte(entry, shift, vma->vm_flags); 494063551ae0SDavid Gibson 494163551ae0SDavid Gibson return entry; 494263551ae0SDavid Gibson } 494363551ae0SDavid Gibson 49441e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 49451e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 49461e8f889bSDavid Gibson { 49471e8f889bSDavid Gibson pte_t entry; 49481e8f889bSDavid Gibson 4949106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 495032f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 49514b3073e1SRussell King update_mmu_cache(vma, address, ptep); 49521e8f889bSDavid Gibson } 49531e8f889bSDavid Gibson 4954d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 49554a705fefSNaoya Horiguchi { 49564a705fefSNaoya Horiguchi swp_entry_t swp; 49574a705fefSNaoya Horiguchi 49584a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 4959d5ed7444SAneesh Kumar K.V return false; 49604a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4961d79d176aSBaoquan He if (is_migration_entry(swp)) 4962d5ed7444SAneesh Kumar K.V return true; 49634a705fefSNaoya Horiguchi else 4964d5ed7444SAneesh Kumar K.V return false; 49654a705fefSNaoya Horiguchi } 49664a705fefSNaoya Horiguchi 49673e5c3600SBaoquan He static bool is_hugetlb_entry_hwpoisoned(pte_t pte) 49684a705fefSNaoya Horiguchi { 49694a705fefSNaoya Horiguchi swp_entry_t swp; 49704a705fefSNaoya Horiguchi 49714a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 49723e5c3600SBaoquan He return false; 49734a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4974d79d176aSBaoquan He if (is_hwpoison_entry(swp)) 49753e5c3600SBaoquan He return true; 49764a705fefSNaoya Horiguchi else 49773e5c3600SBaoquan He return false; 49784a705fefSNaoya Horiguchi } 49791e8f889bSDavid Gibson 49804eae4efaSPeter Xu static void 4981ea4c353dSSidhartha Kumar hugetlb_install_folio(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, 4982935d4f0cSRyan Roberts struct folio *new_folio, pte_t old, unsigned long sz) 49834eae4efaSPeter Xu { 49845a2f8d22SPeter Xu pte_t newpte = make_huge_pte(vma, &new_folio->page, 1); 49855a2f8d22SPeter Xu 4986ea4c353dSSidhartha Kumar __folio_mark_uptodate(new_folio); 4987d0ce0e47SSidhartha Kumar hugepage_add_new_anon_rmap(new_folio, vma, addr); 49885a2f8d22SPeter Xu if (userfaultfd_wp(vma) && huge_pte_uffd_wp(old)) 49895a2f8d22SPeter Xu newpte = huge_pte_mkuffd_wp(newpte); 4990935d4f0cSRyan Roberts set_huge_pte_at(vma->vm_mm, addr, ptep, newpte, sz); 49914eae4efaSPeter Xu hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); 4992ea4c353dSSidhartha Kumar folio_set_hugetlb_migratable(new_folio); 49934eae4efaSPeter Xu } 49944eae4efaSPeter Xu 499563551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 4996bc70fbf2SPeter Xu struct vm_area_struct *dst_vma, 4997bc70fbf2SPeter Xu struct vm_area_struct *src_vma) 499863551ae0SDavid Gibson { 49993aa4ed80SMiaohe Lin pte_t *src_pte, *dst_pte, entry; 5000ad27ce20SZhangPeng struct folio *pte_folio; 50011c59827dSHugh Dickins unsigned long addr; 5002bc70fbf2SPeter Xu bool cow = is_cow_mapping(src_vma->vm_flags); 5003bc70fbf2SPeter Xu struct hstate *h = hstate_vma(src_vma); 5004a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 50054eae4efaSPeter Xu unsigned long npages = pages_per_huge_page(h); 5006ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 5007e95a9851SMike Kravetz unsigned long last_addr_mask; 5008e8569dd2SAndreas Sandberg int ret = 0; 50091e8f889bSDavid Gibson 5010ac46d4f3SJérôme Glisse if (cow) { 50117d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src, 5012bc70fbf2SPeter Xu src_vma->vm_start, 5013bc70fbf2SPeter Xu src_vma->vm_end); 5014ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 5015e727bfd5SSuren Baghdasaryan vma_assert_write_locked(src_vma); 5016623a1ddfSDavid Hildenbrand raw_write_seqcount_begin(&src->write_protect_seq); 501740549ba8SMike Kravetz } else { 501840549ba8SMike Kravetz /* 501940549ba8SMike Kravetz * For shared mappings the vma lock must be held before 50209c67a207SPeter Xu * calling hugetlb_walk() in the src vma. Otherwise, the 502140549ba8SMike Kravetz * returned ptep could go away if part of a shared pmd and 502240549ba8SMike Kravetz * another thread calls huge_pmd_unshare. 502340549ba8SMike Kravetz */ 502440549ba8SMike Kravetz hugetlb_vma_lock_read(src_vma); 5025ac46d4f3SJérôme Glisse } 5026e8569dd2SAndreas Sandberg 5027e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 5028bc70fbf2SPeter Xu for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) { 5029cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 50309c67a207SPeter Xu src_pte = hugetlb_walk(src_vma, addr, sz); 5031e95a9851SMike Kravetz if (!src_pte) { 5032e95a9851SMike Kravetz addr |= last_addr_mask; 5033c74df32cSHugh Dickins continue; 5034e95a9851SMike Kravetz } 5035bc70fbf2SPeter Xu dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz); 5036e8569dd2SAndreas Sandberg if (!dst_pte) { 5037e8569dd2SAndreas Sandberg ret = -ENOMEM; 5038e8569dd2SAndreas Sandberg break; 5039e8569dd2SAndreas Sandberg } 5040c5c99429SLarry Woodman 50415e41540cSMike Kravetz /* 50425e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 50435e41540cSMike Kravetz * 50443aa4ed80SMiaohe Lin * dst_pte == src_pte is the common case of src/dest sharing. 50455e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 50463aa4ed80SMiaohe Lin * another vma. So page_count of ptep page is checked instead 50473aa4ed80SMiaohe Lin * to reliably determine whether pte is shared. 50485e41540cSMike Kravetz */ 50493aa4ed80SMiaohe Lin if (page_count(virt_to_page(dst_pte)) > 1) { 5050e95a9851SMike Kravetz addr |= last_addr_mask; 5051c5c99429SLarry Woodman continue; 5052e95a9851SMike Kravetz } 5053c5c99429SLarry Woodman 5054cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 5055cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 5056cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 50574a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 50584eae4efaSPeter Xu again: 50593aa4ed80SMiaohe Lin if (huge_pte_none(entry)) { 50605e41540cSMike Kravetz /* 50613aa4ed80SMiaohe Lin * Skip if src entry none. 50625e41540cSMike Kravetz */ 50634a705fefSNaoya Horiguchi ; 5064c2cb0dccSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) { 50655a2f8d22SPeter Xu if (!userfaultfd_wp(dst_vma)) 5066c2cb0dccSNaoya Horiguchi entry = huge_pte_clear_uffd_wp(entry); 5067935d4f0cSRyan Roberts set_huge_pte_at(dst, addr, dst_pte, entry, sz); 5068c2cb0dccSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry))) { 50694a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 50705a2f8d22SPeter Xu bool uffd_wp = pte_swp_uffd_wp(entry); 50714a705fefSNaoya Horiguchi 50726c287605SDavid Hildenbrand if (!is_readable_migration_entry(swp_entry) && cow) { 50734a705fefSNaoya Horiguchi /* 50744a705fefSNaoya Horiguchi * COW mappings require pages in both 50754a705fefSNaoya Horiguchi * parent and child to be set to read. 50764a705fefSNaoya Horiguchi */ 50774dd845b5SAlistair Popple swp_entry = make_readable_migration_entry( 50784dd845b5SAlistair Popple swp_offset(swp_entry)); 50794a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 5080bc70fbf2SPeter Xu if (userfaultfd_wp(src_vma) && uffd_wp) 50815a2f8d22SPeter Xu entry = pte_swp_mkuffd_wp(entry); 5082935d4f0cSRyan Roberts set_huge_pte_at(src, addr, src_pte, entry, sz); 50834a705fefSNaoya Horiguchi } 50845a2f8d22SPeter Xu if (!userfaultfd_wp(dst_vma)) 5085bc70fbf2SPeter Xu entry = huge_pte_clear_uffd_wp(entry); 5086935d4f0cSRyan Roberts set_huge_pte_at(dst, addr, dst_pte, entry, sz); 5087bc70fbf2SPeter Xu } else if (unlikely(is_pte_marker(entry))) { 5088af19487fSAxel Rasmussen pte_marker marker = copy_pte_marker( 5089af19487fSAxel Rasmussen pte_to_swp_entry(entry), dst_vma); 5090af19487fSAxel Rasmussen 5091af19487fSAxel Rasmussen if (marker) 5092af19487fSAxel Rasmussen set_huge_pte_at(dst, addr, dst_pte, 5093935d4f0cSRyan Roberts make_pte_marker(marker), sz); 50944a705fefSNaoya Horiguchi } else { 50954eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 5096ad27ce20SZhangPeng pte_folio = page_folio(pte_page(entry)); 5097ad27ce20SZhangPeng folio_get(pte_folio); 50984eae4efaSPeter Xu 50994eae4efaSPeter Xu /* 5100fb3d824dSDavid Hildenbrand * Failing to duplicate the anon rmap is a rare case 5101fb3d824dSDavid Hildenbrand * where we see pinned hugetlb pages while they're 5102fb3d824dSDavid Hildenbrand * prone to COW. We need to do the COW earlier during 5103fb3d824dSDavid Hildenbrand * fork. 51044eae4efaSPeter Xu * 51054eae4efaSPeter Xu * When pre-allocating the page or copying data, we 51064eae4efaSPeter Xu * need to be without the pgtable locks since we could 51074eae4efaSPeter Xu * sleep during the process. 51084eae4efaSPeter Xu */ 5109ad27ce20SZhangPeng if (!folio_test_anon(pte_folio)) { 5110ad27ce20SZhangPeng page_dup_file_rmap(&pte_folio->page, true); 5111ad27ce20SZhangPeng } else if (page_try_dup_anon_rmap(&pte_folio->page, 5112ad27ce20SZhangPeng true, src_vma)) { 51134eae4efaSPeter Xu pte_t src_pte_old = entry; 5114d0ce0e47SSidhartha Kumar struct folio *new_folio; 51154eae4efaSPeter Xu 51164eae4efaSPeter Xu spin_unlock(src_ptl); 51174eae4efaSPeter Xu spin_unlock(dst_ptl); 51184eae4efaSPeter Xu /* Do not use reserve as it's private owned */ 5119d0ce0e47SSidhartha Kumar new_folio = alloc_hugetlb_folio(dst_vma, addr, 1); 5120d0ce0e47SSidhartha Kumar if (IS_ERR(new_folio)) { 5121ad27ce20SZhangPeng folio_put(pte_folio); 5122d0ce0e47SSidhartha Kumar ret = PTR_ERR(new_folio); 51234eae4efaSPeter Xu break; 51244eae4efaSPeter Xu } 51251cb9dc4bSLiu Shixin ret = copy_user_large_folio(new_folio, 5126ad27ce20SZhangPeng pte_folio, 5127c0e8150eSZhangPeng addr, dst_vma); 5128ad27ce20SZhangPeng folio_put(pte_folio); 51291cb9dc4bSLiu Shixin if (ret) { 51301cb9dc4bSLiu Shixin folio_put(new_folio); 51311cb9dc4bSLiu Shixin break; 51321cb9dc4bSLiu Shixin } 51334eae4efaSPeter Xu 5134d0ce0e47SSidhartha Kumar /* Install the new hugetlb folio if src pte stable */ 51354eae4efaSPeter Xu dst_ptl = huge_pte_lock(h, dst, dst_pte); 51364eae4efaSPeter Xu src_ptl = huge_pte_lockptr(h, src, src_pte); 51374eae4efaSPeter Xu spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 51384eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 51394eae4efaSPeter Xu if (!pte_same(src_pte_old, entry)) { 5140bc70fbf2SPeter Xu restore_reserve_on_error(h, dst_vma, addr, 5141d2d7bb44SSidhartha Kumar new_folio); 5142d0ce0e47SSidhartha Kumar folio_put(new_folio); 51433aa4ed80SMiaohe Lin /* huge_ptep of dst_pte won't change as in child */ 51444eae4efaSPeter Xu goto again; 51454eae4efaSPeter Xu } 51465a2f8d22SPeter Xu hugetlb_install_folio(dst_vma, dst_pte, addr, 5147935d4f0cSRyan Roberts new_folio, src_pte_old, sz); 51484eae4efaSPeter Xu spin_unlock(src_ptl); 51494eae4efaSPeter Xu spin_unlock(dst_ptl); 51504eae4efaSPeter Xu continue; 51514eae4efaSPeter Xu } 51524eae4efaSPeter Xu 515334ee645eSJoerg Roedel if (cow) { 51540f10851eSJérôme Glisse /* 51550f10851eSJérôme Glisse * No need to notify as we are downgrading page 51560f10851eSJérôme Glisse * table protection not changing it to point 51570f10851eSJérôme Glisse * to a new page. 51580f10851eSJérôme Glisse * 5159ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 51600f10851eSJérôme Glisse */ 51617f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 516284894e1cSPeter Xu entry = huge_pte_wrprotect(entry); 516334ee645eSJoerg Roedel } 51644eae4efaSPeter Xu 51655a2f8d22SPeter Xu if (!userfaultfd_wp(dst_vma)) 51665a2f8d22SPeter Xu entry = huge_pte_clear_uffd_wp(entry); 51675a2f8d22SPeter Xu 5168935d4f0cSRyan Roberts set_huge_pte_at(dst, addr, dst_pte, entry, sz); 51694eae4efaSPeter Xu hugetlb_count_add(npages, dst); 51701c59827dSHugh Dickins } 5171cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 5172cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 517363551ae0SDavid Gibson } 517463551ae0SDavid Gibson 5175623a1ddfSDavid Hildenbrand if (cow) { 5176623a1ddfSDavid Hildenbrand raw_write_seqcount_end(&src->write_protect_seq); 5177ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 517840549ba8SMike Kravetz } else { 517940549ba8SMike Kravetz hugetlb_vma_unlock_read(src_vma); 5180623a1ddfSDavid Hildenbrand } 5181e8569dd2SAndreas Sandberg 5182e8569dd2SAndreas Sandberg return ret; 518363551ae0SDavid Gibson } 518463551ae0SDavid Gibson 5185550a7d60SMina Almasry static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr, 5186935d4f0cSRyan Roberts unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte, 5187935d4f0cSRyan Roberts unsigned long sz) 5188550a7d60SMina Almasry { 5189550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 5190550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 5191550a7d60SMina Almasry spinlock_t *src_ptl, *dst_ptl; 5192db110a99SAneesh Kumar K.V pte_t pte; 5193550a7d60SMina Almasry 5194550a7d60SMina Almasry dst_ptl = huge_pte_lock(h, mm, dst_pte); 5195550a7d60SMina Almasry src_ptl = huge_pte_lockptr(h, mm, src_pte); 5196550a7d60SMina Almasry 5197550a7d60SMina Almasry /* 5198550a7d60SMina Almasry * We don't have to worry about the ordering of src and dst ptlocks 51998651a137SLorenzo Stoakes * because exclusive mmap_lock (or the i_mmap_lock) prevents deadlock. 5200550a7d60SMina Almasry */ 5201550a7d60SMina Almasry if (src_ptl != dst_ptl) 5202550a7d60SMina Almasry spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 5203550a7d60SMina Almasry 5204550a7d60SMina Almasry pte = huge_ptep_get_and_clear(mm, old_addr, src_pte); 5205935d4f0cSRyan Roberts set_huge_pte_at(mm, new_addr, dst_pte, pte, sz); 5206550a7d60SMina Almasry 5207550a7d60SMina Almasry if (src_ptl != dst_ptl) 5208550a7d60SMina Almasry spin_unlock(src_ptl); 5209550a7d60SMina Almasry spin_unlock(dst_ptl); 5210550a7d60SMina Almasry } 5211550a7d60SMina Almasry 5212550a7d60SMina Almasry int move_hugetlb_page_tables(struct vm_area_struct *vma, 5213550a7d60SMina Almasry struct vm_area_struct *new_vma, 5214550a7d60SMina Almasry unsigned long old_addr, unsigned long new_addr, 5215550a7d60SMina Almasry unsigned long len) 5216550a7d60SMina Almasry { 5217550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 5218550a7d60SMina Almasry struct address_space *mapping = vma->vm_file->f_mapping; 5219550a7d60SMina Almasry unsigned long sz = huge_page_size(h); 5220550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 5221550a7d60SMina Almasry unsigned long old_end = old_addr + len; 5222e95a9851SMike Kravetz unsigned long last_addr_mask; 5223550a7d60SMina Almasry pte_t *src_pte, *dst_pte; 5224550a7d60SMina Almasry struct mmu_notifier_range range; 52253d0b95cdSBaolin Wang bool shared_pmd = false; 5226550a7d60SMina Almasry 52277d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, old_addr, 5228550a7d60SMina Almasry old_end); 5229550a7d60SMina Almasry adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 52303d0b95cdSBaolin Wang /* 52313d0b95cdSBaolin Wang * In case of shared PMDs, we should cover the maximum possible 52323d0b95cdSBaolin Wang * range. 52333d0b95cdSBaolin Wang */ 52343d0b95cdSBaolin Wang flush_cache_range(vma, range.start, range.end); 52353d0b95cdSBaolin Wang 5236550a7d60SMina Almasry mmu_notifier_invalidate_range_start(&range); 5237e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 5238550a7d60SMina Almasry /* Prevent race with file truncation */ 523940549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 5240550a7d60SMina Almasry i_mmap_lock_write(mapping); 5241550a7d60SMina Almasry for (; old_addr < old_end; old_addr += sz, new_addr += sz) { 52429c67a207SPeter Xu src_pte = hugetlb_walk(vma, old_addr, sz); 5243e95a9851SMike Kravetz if (!src_pte) { 5244e95a9851SMike Kravetz old_addr |= last_addr_mask; 5245e95a9851SMike Kravetz new_addr |= last_addr_mask; 5246550a7d60SMina Almasry continue; 5247e95a9851SMike Kravetz } 5248550a7d60SMina Almasry if (huge_pte_none(huge_ptep_get(src_pte))) 5249550a7d60SMina Almasry continue; 5250550a7d60SMina Almasry 52514ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) { 52523d0b95cdSBaolin Wang shared_pmd = true; 52534ddb4d91SMike Kravetz old_addr |= last_addr_mask; 52544ddb4d91SMike Kravetz new_addr |= last_addr_mask; 5255550a7d60SMina Almasry continue; 52563d0b95cdSBaolin Wang } 5257550a7d60SMina Almasry 5258550a7d60SMina Almasry dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz); 5259550a7d60SMina Almasry if (!dst_pte) 5260550a7d60SMina Almasry break; 5261550a7d60SMina Almasry 5262935d4f0cSRyan Roberts move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte, sz); 5263550a7d60SMina Almasry } 52643d0b95cdSBaolin Wang 52653d0b95cdSBaolin Wang if (shared_pmd) 5266f720b471SKefeng Wang flush_hugetlb_tlb_range(vma, range.start, range.end); 52673d0b95cdSBaolin Wang else 5268f720b471SKefeng Wang flush_hugetlb_tlb_range(vma, old_end - len, old_end); 5269550a7d60SMina Almasry mmu_notifier_invalidate_range_end(&range); 527013e4ad2cSNadav Amit i_mmap_unlock_write(mapping); 527140549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 5272550a7d60SMina Almasry 5273550a7d60SMina Almasry return len + old_addr - old_end; 5274550a7d60SMina Almasry } 5275550a7d60SMina Almasry 527673c54763SPeter Xu static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 527724669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 527805e90bd0SPeter Xu struct page *ref_page, zap_flags_t zap_flags) 527963551ae0SDavid Gibson { 528063551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 528163551ae0SDavid Gibson unsigned long address; 5282c7546f8fSDavid Gibson pte_t *ptep; 528363551ae0SDavid Gibson pte_t pte; 5284cb900f41SKirill A. Shutemov spinlock_t *ptl; 528563551ae0SDavid Gibson struct page *page; 5286a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 5287a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 5288e95a9851SMike Kravetz unsigned long last_addr_mask; 5289a4a118f2SNadav Amit bool force_flush = false; 5290a5516438SAndi Kleen 529163551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 5292a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 5293a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 529463551ae0SDavid Gibson 529507e32661SAneesh Kumar K.V /* 529607e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 529707e32661SAneesh Kumar K.V * to huge page. 529807e32661SAneesh Kumar K.V */ 5299ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 530024669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 5301dff11abeSMike Kravetz 5302e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 5303569f48b8SHillf Danton address = start; 5304569f48b8SHillf Danton for (; address < end; address += sz) { 53059c67a207SPeter Xu ptep = hugetlb_walk(vma, address, sz); 5306e95a9851SMike Kravetz if (!ptep) { 5307e95a9851SMike Kravetz address |= last_addr_mask; 5308c7546f8fSDavid Gibson continue; 5309e95a9851SMike Kravetz } 5310c7546f8fSDavid Gibson 5311cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 53124ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, address, ptep)) { 531331d49da5SAneesh Kumar K.V spin_unlock(ptl); 5314a4a118f2SNadav Amit tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE); 5315a4a118f2SNadav Amit force_flush = true; 53164ddb4d91SMike Kravetz address |= last_addr_mask; 531731d49da5SAneesh Kumar K.V continue; 531831d49da5SAneesh Kumar K.V } 531939dde65cSChen, Kenneth W 53206629326bSHillf Danton pte = huge_ptep_get(ptep); 532131d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 532231d49da5SAneesh Kumar K.V spin_unlock(ptl); 532331d49da5SAneesh Kumar K.V continue; 532431d49da5SAneesh Kumar K.V } 53256629326bSHillf Danton 53266629326bSHillf Danton /* 53279fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 53289fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 53296629326bSHillf Danton */ 53309fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 533105e90bd0SPeter Xu /* 533205e90bd0SPeter Xu * If the pte was wr-protected by uffd-wp in any of the 533305e90bd0SPeter Xu * swap forms, meanwhile the caller does not want to 533405e90bd0SPeter Xu * drop the uffd-wp bit in this zap, then replace the 533505e90bd0SPeter Xu * pte with a marker. 533605e90bd0SPeter Xu */ 533705e90bd0SPeter Xu if (pte_swp_uffd_wp_any(pte) && 533805e90bd0SPeter Xu !(zap_flags & ZAP_FLAG_DROP_MARKER)) 533905e90bd0SPeter Xu set_huge_pte_at(mm, address, ptep, 5340935d4f0cSRyan Roberts make_pte_marker(PTE_MARKER_UFFD_WP), 5341935d4f0cSRyan Roberts sz); 534205e90bd0SPeter Xu else 53439386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 534431d49da5SAneesh Kumar K.V spin_unlock(ptl); 534531d49da5SAneesh Kumar K.V continue; 53468c4894c6SNaoya Horiguchi } 53476629326bSHillf Danton 53486629326bSHillf Danton page = pte_page(pte); 534904f2cbe3SMel Gorman /* 535004f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 535104f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 535204f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 535304f2cbe3SMel Gorman */ 535404f2cbe3SMel Gorman if (ref_page) { 535531d49da5SAneesh Kumar K.V if (page != ref_page) { 535631d49da5SAneesh Kumar K.V spin_unlock(ptl); 535731d49da5SAneesh Kumar K.V continue; 535831d49da5SAneesh Kumar K.V } 535904f2cbe3SMel Gorman /* 536004f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 536104f2cbe3SMel Gorman * future faults in this VMA will fail rather than 536204f2cbe3SMel Gorman * looking like data was lost 536304f2cbe3SMel Gorman */ 536404f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 536504f2cbe3SMel Gorman } 536604f2cbe3SMel Gorman 5367c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 5368b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 5369106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 53706649a386SKen Chen set_page_dirty(page); 537105e90bd0SPeter Xu /* Leave a uffd-wp pte marker if needed */ 537205e90bd0SPeter Xu if (huge_pte_uffd_wp(pte) && 537305e90bd0SPeter Xu !(zap_flags & ZAP_FLAG_DROP_MARKER)) 537405e90bd0SPeter Xu set_huge_pte_at(mm, address, ptep, 5375935d4f0cSRyan Roberts make_pte_marker(PTE_MARKER_UFFD_WP), 5376935d4f0cSRyan Roberts sz); 53775d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 5378cea86fe2SHugh Dickins page_remove_rmap(page, vma, true); 537931d49da5SAneesh Kumar K.V 5380cb900f41SKirill A. Shutemov spin_unlock(ptl); 5381e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 538224669e58SAneesh Kumar K.V /* 538331d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 538424669e58SAneesh Kumar K.V */ 538531d49da5SAneesh Kumar K.V if (ref_page) 538631d49da5SAneesh Kumar K.V break; 5387fe1668aeSChen, Kenneth W } 538824669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 5389a4a118f2SNadav Amit 5390a4a118f2SNadav Amit /* 5391a4a118f2SNadav Amit * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We 5392a4a118f2SNadav Amit * could defer the flush until now, since by holding i_mmap_rwsem we 5393a4a118f2SNadav Amit * guaranteed that the last refernece would not be dropped. But we must 5394a4a118f2SNadav Amit * do the flushing before we return, as otherwise i_mmap_rwsem will be 5395a4a118f2SNadav Amit * dropped and the last reference to the shared PMDs page might be 5396a4a118f2SNadav Amit * dropped as well. 5397a4a118f2SNadav Amit * 5398a4a118f2SNadav Amit * In theory we could defer the freeing of the PMD pages as well, but 5399a4a118f2SNadav Amit * huge_pmd_unshare() relies on the exact page_count for the PMD page to 5400a4a118f2SNadav Amit * detect sharing, so we cannot defer the release of the page either. 5401a4a118f2SNadav Amit * Instead, do flush now. 5402a4a118f2SNadav Amit */ 5403a4a118f2SNadav Amit if (force_flush) 5404a4a118f2SNadav Amit tlb_flush_mmu_tlbonly(tlb); 54051da177e4SLinus Torvalds } 540663551ae0SDavid Gibson 5407d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 5408d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 540905e90bd0SPeter Xu unsigned long end, struct page *ref_page, 541005e90bd0SPeter Xu zap_flags_t zap_flags) 5411d833352aSMel Gorman { 5412131a79b4SMike Kravetz hugetlb_vma_lock_write(vma); 5413131a79b4SMike Kravetz i_mmap_lock_write(vma->vm_file->f_mapping); 5414131a79b4SMike Kravetz 5415369258ceSMike Kravetz /* mmu notification performed in caller */ 541605e90bd0SPeter Xu __unmap_hugepage_range(tlb, vma, start, end, ref_page, zap_flags); 5417d833352aSMel Gorman 541804ada095SMike Kravetz if (zap_flags & ZAP_FLAG_UNMAP) { /* final unmap */ 5419d833352aSMel Gorman /* 542004ada095SMike Kravetz * Unlock and free the vma lock before releasing i_mmap_rwsem. 542104ada095SMike Kravetz * When the vma_lock is freed, this makes the vma ineligible 542204ada095SMike Kravetz * for pmd sharing. And, i_mmap_rwsem is required to set up 542304ada095SMike Kravetz * pmd sharing. This is important as page tables for this 542404ada095SMike Kravetz * unmapped range will be asynchrously deleted. If the page 542504ada095SMike Kravetz * tables are shared, there will be issues when accessed by 542604ada095SMike Kravetz * someone else. 5427d833352aSMel Gorman */ 5428ecfbd733SMike Kravetz __hugetlb_vma_unlock_write_free(vma); 5429131a79b4SMike Kravetz i_mmap_unlock_write(vma->vm_file->f_mapping); 543004ada095SMike Kravetz } else { 543104ada095SMike Kravetz i_mmap_unlock_write(vma->vm_file->f_mapping); 543204ada095SMike Kravetz hugetlb_vma_unlock_write(vma); 543304ada095SMike Kravetz } 5434d833352aSMel Gorman } 5435d833352aSMel Gorman 5436502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 543705e90bd0SPeter Xu unsigned long end, struct page *ref_page, 543805e90bd0SPeter Xu zap_flags_t zap_flags) 5439502717f4SChen, Kenneth W { 5440369258ceSMike Kravetz struct mmu_notifier_range range; 544124669e58SAneesh Kumar K.V struct mmu_gather tlb; 5442dff11abeSMike Kravetz 54437d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm, 5444369258ceSMike Kravetz start, end); 5445369258ceSMike Kravetz adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 5446369258ceSMike Kravetz mmu_notifier_invalidate_range_start(&range); 5447a72afd87SWill Deacon tlb_gather_mmu(&tlb, vma->vm_mm); 5448369258ceSMike Kravetz 544905e90bd0SPeter Xu __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags); 5450369258ceSMike Kravetz 5451369258ceSMike Kravetz mmu_notifier_invalidate_range_end(&range); 5452ae8eba8bSWill Deacon tlb_finish_mmu(&tlb); 5453502717f4SChen, Kenneth W } 5454502717f4SChen, Kenneth W 545504f2cbe3SMel Gorman /* 545604f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 5457578b7725SZhiyuan Dai * mapping it owns the reserve page for. The intention is to unmap the page 545804f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 545904f2cbe3SMel Gorman * same region. 546004f2cbe3SMel Gorman */ 54612f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 54622a4b3dedSHarvey Harrison struct page *page, unsigned long address) 546304f2cbe3SMel Gorman { 54647526674dSAdam Litke struct hstate *h = hstate_vma(vma); 546504f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 546604f2cbe3SMel Gorman struct address_space *mapping; 546704f2cbe3SMel Gorman pgoff_t pgoff; 546804f2cbe3SMel Gorman 546904f2cbe3SMel Gorman /* 547004f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 547104f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 547204f2cbe3SMel Gorman */ 54737526674dSAdam Litke address = address & huge_page_mask(h); 547436e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 547536e4f20aSMichal Hocko vma->vm_pgoff; 547693c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 547704f2cbe3SMel Gorman 54784eb2b1dcSMel Gorman /* 54794eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 54804eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 54814eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 54824eb2b1dcSMel Gorman */ 548383cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 54846b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 548504f2cbe3SMel Gorman /* Do not unmap the current VMA */ 548604f2cbe3SMel Gorman if (iter_vma == vma) 548704f2cbe3SMel Gorman continue; 548804f2cbe3SMel Gorman 548904f2cbe3SMel Gorman /* 54902f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 54912f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 54922f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 54932f84a899SMel Gorman */ 54942f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 54952f84a899SMel Gorman continue; 54962f84a899SMel Gorman 54972f84a899SMel Gorman /* 549804f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 549904f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 550004f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 550104f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 550204f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 550304f2cbe3SMel Gorman */ 550404f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 550524669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 550605e90bd0SPeter Xu address + huge_page_size(h), page, 0); 550704f2cbe3SMel Gorman } 550883cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 550904f2cbe3SMel Gorman } 551004f2cbe3SMel Gorman 55110fe6e20bSNaoya Horiguchi /* 5512c89357e2SDavid Hildenbrand * hugetlb_wp() should be called with page lock of the original hugepage held. 5513aa6d2e8cSBaolin Wang * Called with hugetlb_fault_mutex_table held and pte_page locked so we 5514ef009b25SMichal Hocko * cannot race with other handlers or page migration. 5515ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 55160fe6e20bSNaoya Horiguchi */ 5517c89357e2SDavid Hildenbrand static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma, 5518c89357e2SDavid Hildenbrand unsigned long address, pte_t *ptep, unsigned int flags, 5519371607a3SSidhartha Kumar struct folio *pagecache_folio, spinlock_t *ptl) 55201e8f889bSDavid Gibson { 5521c89357e2SDavid Hildenbrand const bool unshare = flags & FAULT_FLAG_UNSHARE; 552260d5b473SPeter Xu pte_t pte = huge_ptep_get(ptep); 5523a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 5524959a78b6SZhangPeng struct folio *old_folio; 5525d0ce0e47SSidhartha Kumar struct folio *new_folio; 55262b740303SSouptick Joarder int outside_reserve = 0; 55272b740303SSouptick Joarder vm_fault_t ret = 0; 5528974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 5529ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 55301e8f889bSDavid Gibson 55311d8d1464SDavid Hildenbrand /* 553260d5b473SPeter Xu * Never handle CoW for uffd-wp protected pages. It should be only 553360d5b473SPeter Xu * handled when the uffd-wp protection is removed. 553460d5b473SPeter Xu * 553560d5b473SPeter Xu * Note that only the CoW optimization path (in hugetlb_no_page()) 553660d5b473SPeter Xu * can trigger this, because hugetlb_fault() will always resolve 553760d5b473SPeter Xu * uffd-wp bit first. 553860d5b473SPeter Xu */ 553960d5b473SPeter Xu if (!unshare && huge_pte_uffd_wp(pte)) 554060d5b473SPeter Xu return 0; 554160d5b473SPeter Xu 554260d5b473SPeter Xu /* 55431d8d1464SDavid Hildenbrand * hugetlb does not support FOLL_FORCE-style write faults that keep the 55441d8d1464SDavid Hildenbrand * PTE mapped R/O such as maybe_mkwrite() would do. 55451d8d1464SDavid Hildenbrand */ 55461d8d1464SDavid Hildenbrand if (WARN_ON_ONCE(!unshare && !(vma->vm_flags & VM_WRITE))) 55471d8d1464SDavid Hildenbrand return VM_FAULT_SIGSEGV; 55481d8d1464SDavid Hildenbrand 55491d8d1464SDavid Hildenbrand /* Let's take out MAP_SHARED mappings first. */ 55501d8d1464SDavid Hildenbrand if (vma->vm_flags & VM_MAYSHARE) { 55511d8d1464SDavid Hildenbrand set_huge_ptep_writable(vma, haddr, ptep); 55521d8d1464SDavid Hildenbrand return 0; 55531d8d1464SDavid Hildenbrand } 55541d8d1464SDavid Hildenbrand 5555959a78b6SZhangPeng old_folio = page_folio(pte_page(pte)); 55561e8f889bSDavid Gibson 5557662ce1dcSYang Yang delayacct_wpcopy_start(); 5558662ce1dcSYang Yang 555904f2cbe3SMel Gorman retry_avoidcopy: 5560c89357e2SDavid Hildenbrand /* 5561c89357e2SDavid Hildenbrand * If no-one else is actually using this page, we're the exclusive 5562c89357e2SDavid Hildenbrand * owner and can reuse this page. 5563c89357e2SDavid Hildenbrand */ 5564959a78b6SZhangPeng if (folio_mapcount(old_folio) == 1 && folio_test_anon(old_folio)) { 5565959a78b6SZhangPeng if (!PageAnonExclusive(&old_folio->page)) 5566959a78b6SZhangPeng page_move_anon_rmap(&old_folio->page, vma); 5567c89357e2SDavid Hildenbrand if (likely(!unshare)) 55685b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 5569662ce1dcSYang Yang 5570662ce1dcSYang Yang delayacct_wpcopy_end(); 557183c54070SNick Piggin return 0; 55721e8f889bSDavid Gibson } 5573959a78b6SZhangPeng VM_BUG_ON_PAGE(folio_test_anon(old_folio) && 5574959a78b6SZhangPeng PageAnonExclusive(&old_folio->page), &old_folio->page); 55751e8f889bSDavid Gibson 557604f2cbe3SMel Gorman /* 557704f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 557804f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 557904f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 558004f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 558104f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 558204f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 558304f2cbe3SMel Gorman * of the full address range. 558404f2cbe3SMel Gorman */ 55855944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 5586959a78b6SZhangPeng old_folio != pagecache_folio) 558704f2cbe3SMel Gorman outside_reserve = 1; 558804f2cbe3SMel Gorman 5589959a78b6SZhangPeng folio_get(old_folio); 5590b76c8cfbSLarry Woodman 5591ad4404a2SDavidlohr Bueso /* 5592ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 5593ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 5594ad4404a2SDavidlohr Bueso */ 5595cb900f41SKirill A. Shutemov spin_unlock(ptl); 5596d0ce0e47SSidhartha Kumar new_folio = alloc_hugetlb_folio(vma, haddr, outside_reserve); 55971e8f889bSDavid Gibson 5598d0ce0e47SSidhartha Kumar if (IS_ERR(new_folio)) { 559904f2cbe3SMel Gorman /* 560004f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 560104f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 560204f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 560304f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 560404f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 560504f2cbe3SMel Gorman */ 560604f2cbe3SMel Gorman if (outside_reserve) { 560740549ba8SMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 560840549ba8SMike Kravetz pgoff_t idx; 560940549ba8SMike Kravetz u32 hash; 561040549ba8SMike Kravetz 5611959a78b6SZhangPeng folio_put(old_folio); 561240549ba8SMike Kravetz /* 561340549ba8SMike Kravetz * Drop hugetlb_fault_mutex and vma_lock before 561440549ba8SMike Kravetz * unmapping. unmapping needs to hold vma_lock 561540549ba8SMike Kravetz * in write mode. Dropping vma_lock in read mode 561640549ba8SMike Kravetz * here is OK as COW mappings do not interact with 561740549ba8SMike Kravetz * PMD sharing. 561840549ba8SMike Kravetz * 561940549ba8SMike Kravetz * Reacquire both after unmap operation. 562040549ba8SMike Kravetz */ 562140549ba8SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 562240549ba8SMike Kravetz hash = hugetlb_fault_mutex_hash(mapping, idx); 562340549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 562440549ba8SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 562540549ba8SMike Kravetz 5626959a78b6SZhangPeng unmap_ref_private(mm, vma, &old_folio->page, haddr); 562740549ba8SMike Kravetz 562840549ba8SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 562940549ba8SMike Kravetz hugetlb_vma_lock_read(vma); 5630cb900f41SKirill A. Shutemov spin_lock(ptl); 56319c67a207SPeter Xu ptep = hugetlb_walk(vma, haddr, huge_page_size(h)); 5632a9af0c5dSNaoya Horiguchi if (likely(ptep && 5633a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 563404f2cbe3SMel Gorman goto retry_avoidcopy; 5635a734bcc8SHillf Danton /* 5636cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 5637cb900f41SKirill A. Shutemov * lock, and our job is done. 5638a734bcc8SHillf Danton */ 5639662ce1dcSYang Yang delayacct_wpcopy_end(); 5640a734bcc8SHillf Danton return 0; 564104f2cbe3SMel Gorman } 564204f2cbe3SMel Gorman 5643d0ce0e47SSidhartha Kumar ret = vmf_error(PTR_ERR(new_folio)); 5644ad4404a2SDavidlohr Bueso goto out_release_old; 56451e8f889bSDavid Gibson } 56461e8f889bSDavid Gibson 56470fe6e20bSNaoya Horiguchi /* 56480fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 56490fe6e20bSNaoya Horiguchi * anon_vma prepared. 56500fe6e20bSNaoya Horiguchi */ 565144e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 5652ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 5653ad4404a2SDavidlohr Bueso goto out_release_all; 565444e2aa93SDean Nelson } 56550fe6e20bSNaoya Horiguchi 5656959a78b6SZhangPeng if (copy_user_large_folio(new_folio, old_folio, address, vma)) { 56571cb9dc4bSLiu Shixin ret = VM_FAULT_HWPOISON_LARGE; 56581cb9dc4bSLiu Shixin goto out_release_all; 56591cb9dc4bSLiu Shixin } 5660d0ce0e47SSidhartha Kumar __folio_mark_uptodate(new_folio); 56611e8f889bSDavid Gibson 56627d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, haddr, 56636f4f13e8SJérôme Glisse haddr + huge_page_size(h)); 5664ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 5665ad4404a2SDavidlohr Bueso 5666b76c8cfbSLarry Woodman /* 5667cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 5668b76c8cfbSLarry Woodman * before the page tables are altered 5669b76c8cfbSLarry Woodman */ 5670cb900f41SKirill A. Shutemov spin_lock(ptl); 56719c67a207SPeter Xu ptep = hugetlb_walk(vma, haddr, huge_page_size(h)); 5672a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 56730f230bc2SPeter Xu pte_t newpte = make_huge_pte(vma, &new_folio->page, !unshare); 56740f230bc2SPeter Xu 5675c89357e2SDavid Hildenbrand /* Break COW or unshare */ 56765b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 5677959a78b6SZhangPeng page_remove_rmap(&old_folio->page, vma, true); 5678d0ce0e47SSidhartha Kumar hugepage_add_new_anon_rmap(new_folio, vma, haddr); 56790f230bc2SPeter Xu if (huge_pte_uffd_wp(pte)) 56800f230bc2SPeter Xu newpte = huge_pte_mkuffd_wp(newpte); 5681935d4f0cSRyan Roberts set_huge_pte_at(mm, haddr, ptep, newpte, huge_page_size(h)); 5682d0ce0e47SSidhartha Kumar folio_set_hugetlb_migratable(new_folio); 56831e8f889bSDavid Gibson /* Make the old page be freed below */ 5684959a78b6SZhangPeng new_folio = old_folio; 56851e8f889bSDavid Gibson } 5686cb900f41SKirill A. Shutemov spin_unlock(ptl); 5687ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 5688ad4404a2SDavidlohr Bueso out_release_all: 5689c89357e2SDavid Hildenbrand /* 5690c89357e2SDavid Hildenbrand * No restore in case of successful pagetable update (Break COW or 5691c89357e2SDavid Hildenbrand * unshare) 5692c89357e2SDavid Hildenbrand */ 5693959a78b6SZhangPeng if (new_folio != old_folio) 5694d2d7bb44SSidhartha Kumar restore_reserve_on_error(h, vma, haddr, new_folio); 5695d0ce0e47SSidhartha Kumar folio_put(new_folio); 5696ad4404a2SDavidlohr Bueso out_release_old: 5697959a78b6SZhangPeng folio_put(old_folio); 56988312034fSJoonsoo Kim 5699ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 5700662ce1dcSYang Yang 5701662ce1dcSYang Yang delayacct_wpcopy_end(); 5702ad4404a2SDavidlohr Bueso return ret; 57031e8f889bSDavid Gibson } 57041e8f889bSDavid Gibson 57053ae77f43SHugh Dickins /* 57063ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 57073ae77f43SHugh Dickins */ 57083ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 57092a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 57102a15efc9SHugh Dickins { 571191a2fb95SSidhartha Kumar struct address_space *mapping = vma->vm_file->f_mapping; 571291a2fb95SSidhartha Kumar pgoff_t idx = vma_hugecache_offset(h, vma, address); 5713fd4aed8dSMike Kravetz struct folio *folio; 57142a15efc9SHugh Dickins 5715fd4aed8dSMike Kravetz folio = filemap_get_folio(mapping, idx); 5716fd4aed8dSMike Kravetz if (IS_ERR(folio)) 5717fd4aed8dSMike Kravetz return false; 5718fd4aed8dSMike Kravetz folio_put(folio); 5719fd4aed8dSMike Kravetz return true; 57202a15efc9SHugh Dickins } 57212a15efc9SHugh Dickins 57229b91c0e2SSidhartha Kumar int hugetlb_add_to_page_cache(struct folio *folio, struct address_space *mapping, 5723ab76ad54SMike Kravetz pgoff_t idx) 5724ab76ad54SMike Kravetz { 5725ab76ad54SMike Kravetz struct inode *inode = mapping->host; 5726ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 5727d9ef44deSMatthew Wilcox (Oracle) int err; 5728ab76ad54SMike Kravetz 5729d9ef44deSMatthew Wilcox (Oracle) __folio_set_locked(folio); 5730d9ef44deSMatthew Wilcox (Oracle) err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL); 5731d9ef44deSMatthew Wilcox (Oracle) 5732d9ef44deSMatthew Wilcox (Oracle) if (unlikely(err)) { 5733d9ef44deSMatthew Wilcox (Oracle) __folio_clear_locked(folio); 5734ab76ad54SMike Kravetz return err; 5735d9ef44deSMatthew Wilcox (Oracle) } 57369b91c0e2SSidhartha Kumar folio_clear_hugetlb_restore_reserve(folio); 5737ab76ad54SMike Kravetz 573822146c3cSMike Kravetz /* 5739d9ef44deSMatthew Wilcox (Oracle) * mark folio dirty so that it will not be removed from cache/file 574022146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 574122146c3cSMike Kravetz */ 5742d9ef44deSMatthew Wilcox (Oracle) folio_mark_dirty(folio); 574322146c3cSMike Kravetz 5744ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 5745ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 5746ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 5747ab76ad54SMike Kravetz return 0; 5748ab76ad54SMike Kravetz } 5749ab76ad54SMike Kravetz 57507677f7fdSAxel Rasmussen static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma, 57517677f7fdSAxel Rasmussen struct address_space *mapping, 57527677f7fdSAxel Rasmussen pgoff_t idx, 57537677f7fdSAxel Rasmussen unsigned int flags, 57547677f7fdSAxel Rasmussen unsigned long haddr, 5755824ddc60SNadav Amit unsigned long addr, 57567677f7fdSAxel Rasmussen unsigned long reason) 57577677f7fdSAxel Rasmussen { 57587677f7fdSAxel Rasmussen u32 hash; 57597677f7fdSAxel Rasmussen struct vm_fault vmf = { 57607677f7fdSAxel Rasmussen .vma = vma, 57617677f7fdSAxel Rasmussen .address = haddr, 5762824ddc60SNadav Amit .real_address = addr, 57637677f7fdSAxel Rasmussen .flags = flags, 57647677f7fdSAxel Rasmussen 57657677f7fdSAxel Rasmussen /* 57667677f7fdSAxel Rasmussen * Hard to debug if it ends up being 57677677f7fdSAxel Rasmussen * used by a callee that assumes 57687677f7fdSAxel Rasmussen * something about the other 57697677f7fdSAxel Rasmussen * uninitialized fields... same as in 57707677f7fdSAxel Rasmussen * memory.c 57717677f7fdSAxel Rasmussen */ 57727677f7fdSAxel Rasmussen }; 57737677f7fdSAxel Rasmussen 57747677f7fdSAxel Rasmussen /* 5775958f32ceSLiu Shixin * vma_lock and hugetlb_fault_mutex must be dropped before handling 5776958f32ceSLiu Shixin * userfault. Also mmap_lock could be dropped due to handling 5777958f32ceSLiu Shixin * userfault, any vma operation should be careful from here. 57787677f7fdSAxel Rasmussen */ 577940549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 57807677f7fdSAxel Rasmussen hash = hugetlb_fault_mutex_hash(mapping, idx); 57817677f7fdSAxel Rasmussen mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5782958f32ceSLiu Shixin return handle_userfault(&vmf, reason); 57837677f7fdSAxel Rasmussen } 57847677f7fdSAxel Rasmussen 57852ea7ff1eSPeter Xu /* 57862ea7ff1eSPeter Xu * Recheck pte with pgtable lock. Returns true if pte didn't change, or 57872ea7ff1eSPeter Xu * false if pte changed or is changing. 57882ea7ff1eSPeter Xu */ 57892ea7ff1eSPeter Xu static bool hugetlb_pte_stable(struct hstate *h, struct mm_struct *mm, 57902ea7ff1eSPeter Xu pte_t *ptep, pte_t old_pte) 57912ea7ff1eSPeter Xu { 57922ea7ff1eSPeter Xu spinlock_t *ptl; 57932ea7ff1eSPeter Xu bool same; 57942ea7ff1eSPeter Xu 57952ea7ff1eSPeter Xu ptl = huge_pte_lock(h, mm, ptep); 57962ea7ff1eSPeter Xu same = pte_same(huge_ptep_get(ptep), old_pte); 57972ea7ff1eSPeter Xu spin_unlock(ptl); 57982ea7ff1eSPeter Xu 57992ea7ff1eSPeter Xu return same; 58002ea7ff1eSPeter Xu } 58012ea7ff1eSPeter Xu 58022b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 58032b740303SSouptick Joarder struct vm_area_struct *vma, 58048382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 5805c64e912cSPeter Xu unsigned long address, pte_t *ptep, 5806c64e912cSPeter Xu pte_t old_pte, unsigned int flags) 5807ac9b9c66SHugh Dickins { 5808a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 58092b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 5810409eb8c2SHillf Danton int anon_rmap = 0; 58114c887265SAdam Litke unsigned long size; 5812d0ce0e47SSidhartha Kumar struct folio *folio; 58131e8f889bSDavid Gibson pte_t new_pte; 5814cb900f41SKirill A. Shutemov spinlock_t *ptl; 5815285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 5816d0ce0e47SSidhartha Kumar bool new_folio, new_pagecache_folio = false; 5817958f32ceSLiu Shixin u32 hash = hugetlb_fault_mutex_hash(mapping, idx); 58184c887265SAdam Litke 581904f2cbe3SMel Gorman /* 582004f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 582104f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 5822c89357e2SDavid Hildenbrand * COW/unsharing. Warn that such a situation has occurred as it may not 5823c89357e2SDavid Hildenbrand * be obvious. 582404f2cbe3SMel Gorman */ 582504f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 5826910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 582704f2cbe3SMel Gorman current->pid); 5828958f32ceSLiu Shixin goto out; 582904f2cbe3SMel Gorman } 583004f2cbe3SMel Gorman 58314c887265SAdam Litke /* 5832188a3972SMike Kravetz * Use page lock to guard against racing truncation 5833188a3972SMike Kravetz * before we get page_table_lock. 58344c887265SAdam Litke */ 5835d0ce0e47SSidhartha Kumar new_folio = false; 5836d0ce0e47SSidhartha Kumar folio = filemap_lock_folio(mapping, idx); 583766dabbb6SChristoph Hellwig if (IS_ERR(folio)) { 5838188a3972SMike Kravetz size = i_size_read(mapping->host) >> huge_page_shift(h); 5839188a3972SMike Kravetz if (idx >= size) 5840188a3972SMike Kravetz goto out; 58417677f7fdSAxel Rasmussen /* Check for page in userfault range */ 58422ea7ff1eSPeter Xu if (userfaultfd_missing(vma)) { 58432ea7ff1eSPeter Xu /* 58442ea7ff1eSPeter Xu * Since hugetlb_no_page() was examining pte 58452ea7ff1eSPeter Xu * without pgtable lock, we need to re-test under 58462ea7ff1eSPeter Xu * lock because the pte may not be stable and could 58472ea7ff1eSPeter Xu * have changed from under us. Try to detect 58482ea7ff1eSPeter Xu * either changed or during-changing ptes and retry 58492ea7ff1eSPeter Xu * properly when needed. 58502ea7ff1eSPeter Xu * 58512ea7ff1eSPeter Xu * Note that userfaultfd is actually fine with 58522ea7ff1eSPeter Xu * false positives (e.g. caused by pte changed), 58532ea7ff1eSPeter Xu * but not wrong logical events (e.g. caused by 58542ea7ff1eSPeter Xu * reading a pte during changing). The latter can 58552ea7ff1eSPeter Xu * confuse the userspace, so the strictness is very 58562ea7ff1eSPeter Xu * much preferred. E.g., MISSING event should 58572ea7ff1eSPeter Xu * never happen on the page after UFFDIO_COPY has 58582ea7ff1eSPeter Xu * correctly installed the page and returned. 58592ea7ff1eSPeter Xu */ 58602ea7ff1eSPeter Xu if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) { 58612ea7ff1eSPeter Xu ret = 0; 58622ea7ff1eSPeter Xu goto out; 58632ea7ff1eSPeter Xu } 58642ea7ff1eSPeter Xu 58652ea7ff1eSPeter Xu return hugetlb_handle_userfault(vma, mapping, idx, flags, 58662ea7ff1eSPeter Xu haddr, address, 58677677f7fdSAxel Rasmussen VM_UFFD_MISSING); 58682ea7ff1eSPeter Xu } 58691a1aad8aSMike Kravetz 5870d0ce0e47SSidhartha Kumar folio = alloc_hugetlb_folio(vma, haddr, 0); 5871d0ce0e47SSidhartha Kumar if (IS_ERR(folio)) { 58724643d67eSMike Kravetz /* 58734643d67eSMike Kravetz * Returning error will result in faulting task being 58744643d67eSMike Kravetz * sent SIGBUS. The hugetlb fault mutex prevents two 58754643d67eSMike Kravetz * tasks from racing to fault in the same page which 58764643d67eSMike Kravetz * could result in false unable to allocate errors. 58774643d67eSMike Kravetz * Page migration does not take the fault mutex, but 58784643d67eSMike Kravetz * does a clear then write of pte's under page table 58794643d67eSMike Kravetz * lock. Page fault code could race with migration, 58804643d67eSMike Kravetz * notice the clear pte and try to allocate a page 58814643d67eSMike Kravetz * here. Before returning error, get ptl and make 58824643d67eSMike Kravetz * sure there really is no pte entry. 58834643d67eSMike Kravetz */ 5884f9bf6c03SPeter Xu if (hugetlb_pte_stable(h, mm, ptep, old_pte)) 5885d0ce0e47SSidhartha Kumar ret = vmf_error(PTR_ERR(folio)); 5886f9bf6c03SPeter Xu else 5887f9bf6c03SPeter Xu ret = 0; 58886bda666aSChristoph Lameter goto out; 58896bda666aSChristoph Lameter } 5890d0ce0e47SSidhartha Kumar clear_huge_page(&folio->page, address, pages_per_huge_page(h)); 5891d0ce0e47SSidhartha Kumar __folio_mark_uptodate(folio); 5892d0ce0e47SSidhartha Kumar new_folio = true; 5893ac9b9c66SHugh Dickins 5894f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 58959b91c0e2SSidhartha Kumar int err = hugetlb_add_to_page_cache(folio, mapping, idx); 58966bda666aSChristoph Lameter if (err) { 58973a5497a2SMiaohe Lin /* 58983a5497a2SMiaohe Lin * err can't be -EEXIST which implies someone 58993a5497a2SMiaohe Lin * else consumed the reservation since hugetlb 59003a5497a2SMiaohe Lin * fault mutex is held when add a hugetlb page 59013a5497a2SMiaohe Lin * to the page cache. So it's safe to call 59023a5497a2SMiaohe Lin * restore_reserve_on_error() here. 59033a5497a2SMiaohe Lin */ 5904d2d7bb44SSidhartha Kumar restore_reserve_on_error(h, vma, haddr, folio); 5905d0ce0e47SSidhartha Kumar folio_put(folio); 59066bda666aSChristoph Lameter goto out; 59076bda666aSChristoph Lameter } 5908d0ce0e47SSidhartha Kumar new_pagecache_folio = true; 590923be7468SMel Gorman } else { 5910d0ce0e47SSidhartha Kumar folio_lock(folio); 59110fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 59120fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 59130fe6e20bSNaoya Horiguchi goto backout_unlocked; 591423be7468SMel Gorman } 5915409eb8c2SHillf Danton anon_rmap = 1; 59160fe6e20bSNaoya Horiguchi } 59170fe6e20bSNaoya Horiguchi } else { 591857303d80SAndy Whitcroft /* 5919998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 5920998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 5921998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 5922fd6a03edSNaoya Horiguchi */ 5923d0ce0e47SSidhartha Kumar if (unlikely(folio_test_hwpoison(folio))) { 59240eb98f15SMiaohe Lin ret = VM_FAULT_HWPOISON_LARGE | 5925972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 5926fd6a03edSNaoya Horiguchi goto backout_unlocked; 59276bda666aSChristoph Lameter } 59287677f7fdSAxel Rasmussen 59297677f7fdSAxel Rasmussen /* Check for page in userfault range. */ 59307677f7fdSAxel Rasmussen if (userfaultfd_minor(vma)) { 5931d0ce0e47SSidhartha Kumar folio_unlock(folio); 5932d0ce0e47SSidhartha Kumar folio_put(folio); 59332ea7ff1eSPeter Xu /* See comment in userfaultfd_missing() block above */ 59342ea7ff1eSPeter Xu if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) { 59352ea7ff1eSPeter Xu ret = 0; 59362ea7ff1eSPeter Xu goto out; 59372ea7ff1eSPeter Xu } 59382ea7ff1eSPeter Xu return hugetlb_handle_userfault(vma, mapping, idx, flags, 59392ea7ff1eSPeter Xu haddr, address, 59407677f7fdSAxel Rasmussen VM_UFFD_MINOR); 59417677f7fdSAxel Rasmussen } 5942998b4382SNaoya Horiguchi } 59431e8f889bSDavid Gibson 594457303d80SAndy Whitcroft /* 594557303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 594657303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 594757303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 594857303d80SAndy Whitcroft * the spinlock. 594957303d80SAndy Whitcroft */ 59505e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 5951285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 59522b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 59532b26736cSAndy Whitcroft goto backout_unlocked; 59542b26736cSAndy Whitcroft } 59555e911373SMike Kravetz /* Just decrements count, does not deallocate */ 5956285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 59575e911373SMike Kravetz } 595857303d80SAndy Whitcroft 59598bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 596083c54070SNick Piggin ret = 0; 5961c64e912cSPeter Xu /* If pte changed from under us, retry */ 5962c64e912cSPeter Xu if (!pte_same(huge_ptep_get(ptep), old_pte)) 59634c887265SAdam Litke goto backout; 59644c887265SAdam Litke 59654781593dSPeter Xu if (anon_rmap) 5966d0ce0e47SSidhartha Kumar hugepage_add_new_anon_rmap(folio, vma, haddr); 59674781593dSPeter Xu else 5968d0ce0e47SSidhartha Kumar page_dup_file_rmap(&folio->page, true); 5969d0ce0e47SSidhartha Kumar new_pte = make_huge_pte(vma, &folio->page, ((vma->vm_flags & VM_WRITE) 59701e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 5971c64e912cSPeter Xu /* 5972c64e912cSPeter Xu * If this pte was previously wr-protected, keep it wr-protected even 5973c64e912cSPeter Xu * if populated. 5974c64e912cSPeter Xu */ 5975c64e912cSPeter Xu if (unlikely(pte_marker_uffd_wp(old_pte))) 5976f1eb1bacSPeter Xu new_pte = huge_pte_mkuffd_wp(new_pte); 5977935d4f0cSRyan Roberts set_huge_pte_at(mm, haddr, ptep, new_pte, huge_page_size(h)); 59781e8f889bSDavid Gibson 59795d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 5980788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 59811e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 5982371607a3SSidhartha Kumar ret = hugetlb_wp(mm, vma, address, ptep, flags, folio, ptl); 59831e8f889bSDavid Gibson } 59841e8f889bSDavid Gibson 5985cb900f41SKirill A. Shutemov spin_unlock(ptl); 5986cb6acd01SMike Kravetz 5987cb6acd01SMike Kravetz /* 5988d0ce0e47SSidhartha Kumar * Only set hugetlb_migratable in newly allocated pages. Existing pages 5989d0ce0e47SSidhartha Kumar * found in the pagecache may not have hugetlb_migratable if they have 59908f251a3dSMike Kravetz * been isolated for migration. 5991cb6acd01SMike Kravetz */ 5992d0ce0e47SSidhartha Kumar if (new_folio) 5993d0ce0e47SSidhartha Kumar folio_set_hugetlb_migratable(folio); 5994cb6acd01SMike Kravetz 5995d0ce0e47SSidhartha Kumar folio_unlock(folio); 59964c887265SAdam Litke out: 5997958f32ceSLiu Shixin hugetlb_vma_unlock_read(vma); 5998958f32ceSLiu Shixin mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5999ac9b9c66SHugh Dickins return ret; 60004c887265SAdam Litke 60014c887265SAdam Litke backout: 6002cb900f41SKirill A. Shutemov spin_unlock(ptl); 60032b26736cSAndy Whitcroft backout_unlocked: 6004d0ce0e47SSidhartha Kumar if (new_folio && !new_pagecache_folio) 6005d2d7bb44SSidhartha Kumar restore_reserve_on_error(h, vma, haddr, folio); 6006fa27759aSMike Kravetz 6007d0ce0e47SSidhartha Kumar folio_unlock(folio); 6008d0ce0e47SSidhartha Kumar folio_put(folio); 60094c887265SAdam Litke goto out; 6010ac9b9c66SHugh Dickins } 6011ac9b9c66SHugh Dickins 60128382d914SDavidlohr Bueso #ifdef CONFIG_SMP 6013188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 60148382d914SDavidlohr Bueso { 60158382d914SDavidlohr Bueso unsigned long key[2]; 60168382d914SDavidlohr Bueso u32 hash; 60178382d914SDavidlohr Bueso 60188382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 60198382d914SDavidlohr Bueso key[1] = idx; 60208382d914SDavidlohr Bueso 602155254636SMike Kravetz hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); 60228382d914SDavidlohr Bueso 60238382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 60248382d914SDavidlohr Bueso } 60258382d914SDavidlohr Bueso #else 60268382d914SDavidlohr Bueso /* 60276c26d310SMiaohe Lin * For uniprocessor systems we always use a single mutex, so just 60288382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 60298382d914SDavidlohr Bueso */ 6030188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 60318382d914SDavidlohr Bueso { 60328382d914SDavidlohr Bueso return 0; 60338382d914SDavidlohr Bueso } 60348382d914SDavidlohr Bueso #endif 60358382d914SDavidlohr Bueso 60362b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 6037788c7df4SHugh Dickins unsigned long address, unsigned int flags) 603886e5216fSAdam Litke { 60398382d914SDavidlohr Bueso pte_t *ptep, entry; 6040cb900f41SKirill A. Shutemov spinlock_t *ptl; 60412b740303SSouptick Joarder vm_fault_t ret; 60428382d914SDavidlohr Bueso u32 hash; 60438382d914SDavidlohr Bueso pgoff_t idx; 6044061e62e8SZhangPeng struct folio *folio = NULL; 6045371607a3SSidhartha Kumar struct folio *pagecache_folio = NULL; 6046a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 60478382d914SDavidlohr Bueso struct address_space *mapping; 60480f792cf9SNaoya Horiguchi int need_wait_lock = 0; 6049285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 605086e5216fSAdam Litke 60514ec31152SMatthew Wilcox (Oracle) /* TODO: Handle faults under the VMA lock */ 60524ec31152SMatthew Wilcox (Oracle) if (flags & FAULT_FLAG_VMA_LOCK) { 60534ec31152SMatthew Wilcox (Oracle) vma_end_read(vma); 60544ec31152SMatthew Wilcox (Oracle) return VM_FAULT_RETRY; 60554ec31152SMatthew Wilcox (Oracle) } 60564ec31152SMatthew Wilcox (Oracle) 60573935baa9SDavid Gibson /* 60583935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 60593935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 60603935baa9SDavid Gibson * the same page in the page cache. 60613935baa9SDavid Gibson */ 606240549ba8SMike Kravetz mapping = vma->vm_file->f_mapping; 606340549ba8SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 6064188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 6065c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 60668382d914SDavidlohr Bueso 606740549ba8SMike Kravetz /* 606840549ba8SMike Kravetz * Acquire vma lock before calling huge_pte_alloc and hold 606940549ba8SMike Kravetz * until finished with ptep. This prevents huge_pmd_unshare from 607040549ba8SMike Kravetz * being called elsewhere and making the ptep no longer valid. 607140549ba8SMike Kravetz */ 607240549ba8SMike Kravetz hugetlb_vma_lock_read(vma); 607340549ba8SMike Kravetz ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); 607440549ba8SMike Kravetz if (!ptep) { 607540549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 607640549ba8SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 607740549ba8SMike Kravetz return VM_FAULT_OOM; 607840549ba8SMike Kravetz } 607940549ba8SMike Kravetz 60807f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 6081af19487fSAxel Rasmussen if (huge_pte_none_mostly(entry)) { 6082af19487fSAxel Rasmussen if (is_pte_marker(entry)) { 6083af19487fSAxel Rasmussen pte_marker marker = 6084af19487fSAxel Rasmussen pte_marker_get(pte_to_swp_entry(entry)); 6085af19487fSAxel Rasmussen 6086af19487fSAxel Rasmussen if (marker & PTE_MARKER_POISONED) { 6087af19487fSAxel Rasmussen ret = VM_FAULT_HWPOISON_LARGE; 6088af19487fSAxel Rasmussen goto out_mutex; 6089af19487fSAxel Rasmussen } 6090af19487fSAxel Rasmussen } 6091af19487fSAxel Rasmussen 6092958f32ceSLiu Shixin /* 6093af19487fSAxel Rasmussen * Other PTE markers should be handled the same way as none PTE. 6094af19487fSAxel Rasmussen * 6095958f32ceSLiu Shixin * hugetlb_no_page will drop vma lock and hugetlb fault 6096958f32ceSLiu Shixin * mutex internally, which make us return immediately. 6097958f32ceSLiu Shixin */ 6098958f32ceSLiu Shixin return hugetlb_no_page(mm, vma, mapping, idx, address, ptep, 6099c64e912cSPeter Xu entry, flags); 6100af19487fSAxel Rasmussen } 610186e5216fSAdam Litke 610283c54070SNick Piggin ret = 0; 61031e8f889bSDavid Gibson 610457303d80SAndy Whitcroft /* 61050f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 61060f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 61077c8de358SEthon Paul * an active hugepage in pagecache. This goto expects the 2nd page 61087c8de358SEthon Paul * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will 61097c8de358SEthon Paul * properly handle it. 61100f792cf9SNaoya Horiguchi */ 6111fcd48540SPeter Xu if (!pte_present(entry)) { 6112fcd48540SPeter Xu if (unlikely(is_hugetlb_entry_migration(entry))) { 6113fcd48540SPeter Xu /* 6114fcd48540SPeter Xu * Release the hugetlb fault lock now, but retain 6115fcd48540SPeter Xu * the vma lock, because it is needed to guard the 6116fcd48540SPeter Xu * huge_pte_lockptr() later in 6117fcd48540SPeter Xu * migration_entry_wait_huge(). The vma lock will 6118fcd48540SPeter Xu * be released there. 6119fcd48540SPeter Xu */ 6120fcd48540SPeter Xu mutex_unlock(&hugetlb_fault_mutex_table[hash]); 6121fcd48540SPeter Xu migration_entry_wait_huge(vma, ptep); 6122fcd48540SPeter Xu return 0; 6123fcd48540SPeter Xu } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 6124fcd48540SPeter Xu ret = VM_FAULT_HWPOISON_LARGE | 6125fcd48540SPeter Xu VM_FAULT_SET_HINDEX(hstate_index(h)); 61260f792cf9SNaoya Horiguchi goto out_mutex; 6127fcd48540SPeter Xu } 61280f792cf9SNaoya Horiguchi 61290f792cf9SNaoya Horiguchi /* 6130c89357e2SDavid Hildenbrand * If we are going to COW/unshare the mapping later, we examine the 6131c89357e2SDavid Hildenbrand * pending reservations for this page now. This will ensure that any 613257303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 61331d8d1464SDavid Hildenbrand * spinlock. Also lookup the pagecache page now as it is used to 61341d8d1464SDavid Hildenbrand * determine if a reservation has been consumed. 613557303d80SAndy Whitcroft */ 6136c89357e2SDavid Hildenbrand if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) && 61371d8d1464SDavid Hildenbrand !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(entry)) { 6138285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 61392b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 6140b4d1d99fSDavid Gibson goto out_mutex; 61412b26736cSAndy Whitcroft } 61425e911373SMike Kravetz /* Just decrements count, does not deallocate */ 6143285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 614457303d80SAndy Whitcroft 6145371607a3SSidhartha Kumar pagecache_folio = filemap_lock_folio(mapping, idx); 614666dabbb6SChristoph Hellwig if (IS_ERR(pagecache_folio)) 614766dabbb6SChristoph Hellwig pagecache_folio = NULL; 614857303d80SAndy Whitcroft } 614957303d80SAndy Whitcroft 61500f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 61510fe6e20bSNaoya Horiguchi 6152c89357e2SDavid Hildenbrand /* Check for a racing update before calling hugetlb_wp() */ 6153b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 6154cb900f41SKirill A. Shutemov goto out_ptl; 6155b4d1d99fSDavid Gibson 6156166f3eccSPeter Xu /* Handle userfault-wp first, before trying to lock more pages */ 6157166f3eccSPeter Xu if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) && 6158166f3eccSPeter Xu (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 6159166f3eccSPeter Xu struct vm_fault vmf = { 6160166f3eccSPeter Xu .vma = vma, 6161166f3eccSPeter Xu .address = haddr, 6162166f3eccSPeter Xu .real_address = address, 6163166f3eccSPeter Xu .flags = flags, 6164166f3eccSPeter Xu }; 6165166f3eccSPeter Xu 6166166f3eccSPeter Xu spin_unlock(ptl); 6167371607a3SSidhartha Kumar if (pagecache_folio) { 6168371607a3SSidhartha Kumar folio_unlock(pagecache_folio); 6169371607a3SSidhartha Kumar folio_put(pagecache_folio); 6170166f3eccSPeter Xu } 617140549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 6172166f3eccSPeter Xu mutex_unlock(&hugetlb_fault_mutex_table[hash]); 6173166f3eccSPeter Xu return handle_userfault(&vmf, VM_UFFD_WP); 6174166f3eccSPeter Xu } 6175166f3eccSPeter Xu 61760f792cf9SNaoya Horiguchi /* 6177c89357e2SDavid Hildenbrand * hugetlb_wp() requires page locks of pte_page(entry) and 6178371607a3SSidhartha Kumar * pagecache_folio, so here we need take the former one 6179061e62e8SZhangPeng * when folio != pagecache_folio or !pagecache_folio. 61800f792cf9SNaoya Horiguchi */ 6181061e62e8SZhangPeng folio = page_folio(pte_page(entry)); 6182061e62e8SZhangPeng if (folio != pagecache_folio) 6183061e62e8SZhangPeng if (!folio_trylock(folio)) { 61840f792cf9SNaoya Horiguchi need_wait_lock = 1; 61850f792cf9SNaoya Horiguchi goto out_ptl; 61860f792cf9SNaoya Horiguchi } 61870f792cf9SNaoya Horiguchi 6188061e62e8SZhangPeng folio_get(folio); 6189b4d1d99fSDavid Gibson 6190c89357e2SDavid Hildenbrand if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) { 6191106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 6192c89357e2SDavid Hildenbrand ret = hugetlb_wp(mm, vma, address, ptep, flags, 6193371607a3SSidhartha Kumar pagecache_folio, ptl); 61940f792cf9SNaoya Horiguchi goto out_put_page; 6195c89357e2SDavid Hildenbrand } else if (likely(flags & FAULT_FLAG_WRITE)) { 6196106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 6197b4d1d99fSDavid Gibson } 6198c89357e2SDavid Hildenbrand } 6199b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 6200285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 6201788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 6202285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 62030f792cf9SNaoya Horiguchi out_put_page: 6204061e62e8SZhangPeng if (folio != pagecache_folio) 6205061e62e8SZhangPeng folio_unlock(folio); 6206061e62e8SZhangPeng folio_put(folio); 6207cb900f41SKirill A. Shutemov out_ptl: 6208cb900f41SKirill A. Shutemov spin_unlock(ptl); 620957303d80SAndy Whitcroft 6210371607a3SSidhartha Kumar if (pagecache_folio) { 6211371607a3SSidhartha Kumar folio_unlock(pagecache_folio); 6212371607a3SSidhartha Kumar folio_put(pagecache_folio); 621357303d80SAndy Whitcroft } 6214b4d1d99fSDavid Gibson out_mutex: 621540549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 6216c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 62170f792cf9SNaoya Horiguchi /* 62180f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 62190f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 62200f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 62210f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 62220f792cf9SNaoya Horiguchi * here without taking refcount. 62230f792cf9SNaoya Horiguchi */ 62240f792cf9SNaoya Horiguchi if (need_wait_lock) 6225061e62e8SZhangPeng folio_wait_locked(folio); 62261e8f889bSDavid Gibson return ret; 622786e5216fSAdam Litke } 622886e5216fSAdam Litke 6229714c1891SAxel Rasmussen #ifdef CONFIG_USERFAULTFD 62308fb5debcSMike Kravetz /* 6231a734991cSAxel Rasmussen * Used by userfaultfd UFFDIO_* ioctls. Based on userfaultfd's mfill_atomic_pte 6232a734991cSAxel Rasmussen * with modifications for hugetlb pages. 62338fb5debcSMike Kravetz */ 623461c50040SAxel Rasmussen int hugetlb_mfill_atomic_pte(pte_t *dst_pte, 62358fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 62368fb5debcSMike Kravetz unsigned long dst_addr, 62378fb5debcSMike Kravetz unsigned long src_addr, 6238d9712937SAxel Rasmussen uffd_flags_t flags, 62390169fd51SZhangPeng struct folio **foliop) 62408fb5debcSMike Kravetz { 624161c50040SAxel Rasmussen struct mm_struct *dst_mm = dst_vma->vm_mm; 6242d9712937SAxel Rasmussen bool is_continue = uffd_flags_mode_is(flags, MFILL_ATOMIC_CONTINUE); 6243d9712937SAxel Rasmussen bool wp_enabled = (flags & MFILL_ATOMIC_WP); 62448cc5fcbbSMina Almasry struct hstate *h = hstate_vma(dst_vma); 62458cc5fcbbSMina Almasry struct address_space *mapping = dst_vma->vm_file->f_mapping; 62468cc5fcbbSMina Almasry pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr); 62471e392147SAndrea Arcangeli unsigned long size; 62481c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 62498fb5debcSMike Kravetz pte_t _dst_pte; 62508fb5debcSMike Kravetz spinlock_t *ptl; 62518cc5fcbbSMina Almasry int ret = -ENOMEM; 6252d0ce0e47SSidhartha Kumar struct folio *folio; 6253f6191471SAxel Rasmussen int writable; 6254d0ce0e47SSidhartha Kumar bool folio_in_pagecache = false; 62558fb5debcSMike Kravetz 62568a13897fSAxel Rasmussen if (uffd_flags_mode_is(flags, MFILL_ATOMIC_POISON)) { 62578a13897fSAxel Rasmussen ptl = huge_pte_lock(h, dst_mm, dst_pte); 62588a13897fSAxel Rasmussen 62598a13897fSAxel Rasmussen /* Don't overwrite any existing PTEs (even markers) */ 62608a13897fSAxel Rasmussen if (!huge_pte_none(huge_ptep_get(dst_pte))) { 62618a13897fSAxel Rasmussen spin_unlock(ptl); 62628a13897fSAxel Rasmussen return -EEXIST; 62638a13897fSAxel Rasmussen } 62648a13897fSAxel Rasmussen 62658a13897fSAxel Rasmussen _dst_pte = make_pte_marker(PTE_MARKER_POISONED); 6266935d4f0cSRyan Roberts set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte, 6267935d4f0cSRyan Roberts huge_page_size(h)); 62688a13897fSAxel Rasmussen 62698a13897fSAxel Rasmussen /* No need to invalidate - it was non-present before */ 62708a13897fSAxel Rasmussen update_mmu_cache(dst_vma, dst_addr, dst_pte); 62718a13897fSAxel Rasmussen 62728a13897fSAxel Rasmussen spin_unlock(ptl); 62738a13897fSAxel Rasmussen return 0; 62748a13897fSAxel Rasmussen } 62758a13897fSAxel Rasmussen 6276f6191471SAxel Rasmussen if (is_continue) { 6277f6191471SAxel Rasmussen ret = -EFAULT; 6278d0ce0e47SSidhartha Kumar folio = filemap_lock_folio(mapping, idx); 627966dabbb6SChristoph Hellwig if (IS_ERR(folio)) 6280f6191471SAxel Rasmussen goto out; 6281d0ce0e47SSidhartha Kumar folio_in_pagecache = true; 62820169fd51SZhangPeng } else if (!*foliop) { 62830169fd51SZhangPeng /* If a folio already exists, then it's UFFDIO_COPY for 6284d84cf06eSMina Almasry * a non-missing case. Return -EEXIST. 6285d84cf06eSMina Almasry */ 6286d84cf06eSMina Almasry if (vm_shared && 6287d84cf06eSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 6288d84cf06eSMina Almasry ret = -EEXIST; 62898fb5debcSMike Kravetz goto out; 6290d84cf06eSMina Almasry } 6291d84cf06eSMina Almasry 6292d0ce0e47SSidhartha Kumar folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0); 6293d0ce0e47SSidhartha Kumar if (IS_ERR(folio)) { 6294d84cf06eSMina Almasry ret = -ENOMEM; 6295d84cf06eSMina Almasry goto out; 6296d84cf06eSMina Almasry } 62978fb5debcSMike Kravetz 6298e87340caSZhangPeng ret = copy_folio_from_user(folio, (const void __user *) src_addr, 6299e87340caSZhangPeng false); 63008fb5debcSMike Kravetz 6301c1e8d7c6SMichel Lespinasse /* fallback to copy_from_user outside mmap_lock */ 63028fb5debcSMike Kravetz if (unlikely(ret)) { 63039e368259SAndrea Arcangeli ret = -ENOENT; 6304d0ce0e47SSidhartha Kumar /* Free the allocated folio which may have 63058cc5fcbbSMina Almasry * consumed a reservation. 63068cc5fcbbSMina Almasry */ 6307d2d7bb44SSidhartha Kumar restore_reserve_on_error(h, dst_vma, dst_addr, folio); 6308d0ce0e47SSidhartha Kumar folio_put(folio); 63098cc5fcbbSMina Almasry 6310d0ce0e47SSidhartha Kumar /* Allocate a temporary folio to hold the copied 63118cc5fcbbSMina Almasry * contents. 63128cc5fcbbSMina Almasry */ 6313d0ce0e47SSidhartha Kumar folio = alloc_hugetlb_folio_vma(h, dst_vma, dst_addr); 6314d0ce0e47SSidhartha Kumar if (!folio) { 63158cc5fcbbSMina Almasry ret = -ENOMEM; 63168cc5fcbbSMina Almasry goto out; 63178cc5fcbbSMina Almasry } 63180169fd51SZhangPeng *foliop = folio; 63190169fd51SZhangPeng /* Set the outparam foliop and return to the caller to 63208cc5fcbbSMina Almasry * copy the contents outside the lock. Don't free the 63210169fd51SZhangPeng * folio. 63228cc5fcbbSMina Almasry */ 63238fb5debcSMike Kravetz goto out; 63248fb5debcSMike Kravetz } 63258fb5debcSMike Kravetz } else { 63268cc5fcbbSMina Almasry if (vm_shared && 63278cc5fcbbSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 63280169fd51SZhangPeng folio_put(*foliop); 63298cc5fcbbSMina Almasry ret = -EEXIST; 63300169fd51SZhangPeng *foliop = NULL; 63318cc5fcbbSMina Almasry goto out; 63328cc5fcbbSMina Almasry } 63338cc5fcbbSMina Almasry 6334d0ce0e47SSidhartha Kumar folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0); 6335d0ce0e47SSidhartha Kumar if (IS_ERR(folio)) { 63360169fd51SZhangPeng folio_put(*foliop); 63378cc5fcbbSMina Almasry ret = -ENOMEM; 63380169fd51SZhangPeng *foliop = NULL; 63398cc5fcbbSMina Almasry goto out; 63408cc5fcbbSMina Almasry } 63411cb9dc4bSLiu Shixin ret = copy_user_large_folio(folio, *foliop, dst_addr, dst_vma); 63420169fd51SZhangPeng folio_put(*foliop); 63430169fd51SZhangPeng *foliop = NULL; 63441cb9dc4bSLiu Shixin if (ret) { 63451cb9dc4bSLiu Shixin folio_put(folio); 63461cb9dc4bSLiu Shixin goto out; 63471cb9dc4bSLiu Shixin } 63488fb5debcSMike Kravetz } 63498fb5debcSMike Kravetz 63508fb5debcSMike Kravetz /* 6351d0ce0e47SSidhartha Kumar * The memory barrier inside __folio_mark_uptodate makes sure that 63528fb5debcSMike Kravetz * preceding stores to the page contents become visible before 63538fb5debcSMike Kravetz * the set_pte_at() write. 63548fb5debcSMike Kravetz */ 6355d0ce0e47SSidhartha Kumar __folio_mark_uptodate(folio); 63568fb5debcSMike Kravetz 6357f6191471SAxel Rasmussen /* Add shared, newly allocated pages to the page cache. */ 6358f6191471SAxel Rasmussen if (vm_shared && !is_continue) { 63591e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 63601e392147SAndrea Arcangeli ret = -EFAULT; 63611e392147SAndrea Arcangeli if (idx >= size) 63621e392147SAndrea Arcangeli goto out_release_nounlock; 63631c9e8defSMike Kravetz 63641e392147SAndrea Arcangeli /* 63651e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 63667e1813d4SMike Kravetz * hugetlb_add_to_page_cache() below happens through the 63671e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 63681e392147SAndrea Arcangeli * the caller. 63691e392147SAndrea Arcangeli */ 63709b91c0e2SSidhartha Kumar ret = hugetlb_add_to_page_cache(folio, mapping, idx); 63711c9e8defSMike Kravetz if (ret) 63721c9e8defSMike Kravetz goto out_release_nounlock; 6373d0ce0e47SSidhartha Kumar folio_in_pagecache = true; 63741c9e8defSMike Kravetz } 63751c9e8defSMike Kravetz 6376bcc66543SMiaohe Lin ptl = huge_pte_lock(h, dst_mm, dst_pte); 63778fb5debcSMike Kravetz 63788625147cSJames Houghton ret = -EIO; 6379d0ce0e47SSidhartha Kumar if (folio_test_hwpoison(folio)) 63808625147cSJames Houghton goto out_release_unlock; 63818625147cSJames Houghton 63821e392147SAndrea Arcangeli /* 63836041c691SPeter Xu * We allow to overwrite a pte marker: consider when both MISSING|WP 63846041c691SPeter Xu * registered, we firstly wr-protect a none pte which has no page cache 63856041c691SPeter Xu * page backing it, then access the page. 63866041c691SPeter Xu */ 6387fa27759aSMike Kravetz ret = -EEXIST; 63886041c691SPeter Xu if (!huge_pte_none_mostly(huge_ptep_get(dst_pte))) 63898fb5debcSMike Kravetz goto out_release_unlock; 63908fb5debcSMike Kravetz 6391d0ce0e47SSidhartha Kumar if (folio_in_pagecache) 6392d0ce0e47SSidhartha Kumar page_dup_file_rmap(&folio->page, true); 63934781593dSPeter Xu else 6394d0ce0e47SSidhartha Kumar hugepage_add_new_anon_rmap(folio, dst_vma, dst_addr); 63958fb5debcSMike Kravetz 63966041c691SPeter Xu /* 63976041c691SPeter Xu * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY 63986041c691SPeter Xu * with wp flag set, don't set pte write bit. 63996041c691SPeter Xu */ 6400d9712937SAxel Rasmussen if (wp_enabled || (is_continue && !vm_shared)) 6401f6191471SAxel Rasmussen writable = 0; 6402f6191471SAxel Rasmussen else 6403f6191471SAxel Rasmussen writable = dst_vma->vm_flags & VM_WRITE; 6404f6191471SAxel Rasmussen 6405d0ce0e47SSidhartha Kumar _dst_pte = make_huge_pte(dst_vma, &folio->page, writable); 64066041c691SPeter Xu /* 64076041c691SPeter Xu * Always mark UFFDIO_COPY page dirty; note that this may not be 64086041c691SPeter Xu * extremely important for hugetlbfs for now since swapping is not 64096041c691SPeter Xu * supported, but we should still be clear in that this page cannot be 64106041c691SPeter Xu * thrown away at will, even if write bit not set. 64116041c691SPeter Xu */ 64128fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 64138fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 64148fb5debcSMike Kravetz 6415d9712937SAxel Rasmussen if (wp_enabled) 64166041c691SPeter Xu _dst_pte = huge_pte_mkuffd_wp(_dst_pte); 64176041c691SPeter Xu 6418935d4f0cSRyan Roberts set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte, huge_page_size(h)); 64198fb5debcSMike Kravetz 64208fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 64218fb5debcSMike Kravetz 64228fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 64238fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 64248fb5debcSMike Kravetz 64258fb5debcSMike Kravetz spin_unlock(ptl); 6426f6191471SAxel Rasmussen if (!is_continue) 6427d0ce0e47SSidhartha Kumar folio_set_hugetlb_migratable(folio); 6428f6191471SAxel Rasmussen if (vm_shared || is_continue) 6429d0ce0e47SSidhartha Kumar folio_unlock(folio); 64308fb5debcSMike Kravetz ret = 0; 64318fb5debcSMike Kravetz out: 64328fb5debcSMike Kravetz return ret; 64338fb5debcSMike Kravetz out_release_unlock: 64348fb5debcSMike Kravetz spin_unlock(ptl); 6435f6191471SAxel Rasmussen if (vm_shared || is_continue) 6436d0ce0e47SSidhartha Kumar folio_unlock(folio); 64375af10dfdSAndrea Arcangeli out_release_nounlock: 6438d0ce0e47SSidhartha Kumar if (!folio_in_pagecache) 6439d2d7bb44SSidhartha Kumar restore_reserve_on_error(h, dst_vma, dst_addr, folio); 6440d0ce0e47SSidhartha Kumar folio_put(folio); 64418fb5debcSMike Kravetz goto out; 64428fb5debcSMike Kravetz } 6443714c1891SAxel Rasmussen #endif /* CONFIG_USERFAULTFD */ 64448fb5debcSMike Kravetz 644557a196a5SMike Kravetz struct page *hugetlb_follow_page_mask(struct vm_area_struct *vma, 64465502ea44SPeter Xu unsigned long address, unsigned int flags, 64475502ea44SPeter Xu unsigned int *page_mask) 644857a196a5SMike Kravetz { 644957a196a5SMike Kravetz struct hstate *h = hstate_vma(vma); 645057a196a5SMike Kravetz struct mm_struct *mm = vma->vm_mm; 645157a196a5SMike Kravetz unsigned long haddr = address & huge_page_mask(h); 645257a196a5SMike Kravetz struct page *page = NULL; 645357a196a5SMike Kravetz spinlock_t *ptl; 645457a196a5SMike Kravetz pte_t *pte, entry; 6455458568c9SPeter Xu int ret; 645657a196a5SMike Kravetz 64577d049f3aSPeter Xu hugetlb_vma_lock_read(vma); 64589c67a207SPeter Xu pte = hugetlb_walk(vma, haddr, huge_page_size(h)); 645957a196a5SMike Kravetz if (!pte) 64607d049f3aSPeter Xu goto out_unlock; 646157a196a5SMike Kravetz 646257a196a5SMike Kravetz ptl = huge_pte_lock(h, mm, pte); 646357a196a5SMike Kravetz entry = huge_ptep_get(pte); 646457a196a5SMike Kravetz if (pte_present(entry)) { 6465458568c9SPeter Xu page = pte_page(entry); 6466458568c9SPeter Xu 6467458568c9SPeter Xu if (!huge_pte_write(entry)) { 6468458568c9SPeter Xu if (flags & FOLL_WRITE) { 6469458568c9SPeter Xu page = NULL; 6470458568c9SPeter Xu goto out; 6471458568c9SPeter Xu } 6472458568c9SPeter Xu 6473458568c9SPeter Xu if (gup_must_unshare(vma, flags, page)) { 6474458568c9SPeter Xu /* Tell the caller to do unsharing */ 6475458568c9SPeter Xu page = ERR_PTR(-EMLINK); 6476458568c9SPeter Xu goto out; 6477458568c9SPeter Xu } 6478458568c9SPeter Xu } 6479458568c9SPeter Xu 6480458568c9SPeter Xu page += ((address & ~huge_page_mask(h)) >> PAGE_SHIFT); 6481458568c9SPeter Xu 648257a196a5SMike Kravetz /* 648357a196a5SMike Kravetz * Note that page may be a sub-page, and with vmemmap 648457a196a5SMike Kravetz * optimizations the page struct may be read only. 648557a196a5SMike Kravetz * try_grab_page() will increase the ref count on the 648657a196a5SMike Kravetz * head page, so this will be OK. 648757a196a5SMike Kravetz * 6488e2ca6ba6SLinus Torvalds * try_grab_page() should always be able to get the page here, 6489e2ca6ba6SLinus Torvalds * because we hold the ptl lock and have verified pte_present(). 649057a196a5SMike Kravetz */ 6491458568c9SPeter Xu ret = try_grab_page(page, flags); 6492458568c9SPeter Xu 6493458568c9SPeter Xu if (WARN_ON_ONCE(ret)) { 6494458568c9SPeter Xu page = ERR_PTR(ret); 649557a196a5SMike Kravetz goto out; 649657a196a5SMike Kravetz } 64975502ea44SPeter Xu 64985502ea44SPeter Xu *page_mask = (1U << huge_page_order(h)) - 1; 649957a196a5SMike Kravetz } 650057a196a5SMike Kravetz out: 650157a196a5SMike Kravetz spin_unlock(ptl); 65027d049f3aSPeter Xu out_unlock: 65037d049f3aSPeter Xu hugetlb_vma_unlock_read(vma); 6504dd767aaaSPeter Xu 6505dd767aaaSPeter Xu /* 6506dd767aaaSPeter Xu * Fixup retval for dump requests: if pagecache doesn't exist, 6507dd767aaaSPeter Xu * don't try to allocate a new page but just skip it. 6508dd767aaaSPeter Xu */ 6509dd767aaaSPeter Xu if (!page && (flags & FOLL_DUMP) && 6510dd767aaaSPeter Xu !hugetlbfs_pagecache_present(h, vma, address)) 6511dd767aaaSPeter Xu page = ERR_PTR(-EFAULT); 6512dd767aaaSPeter Xu 651357a196a5SMike Kravetz return page; 651457a196a5SMike Kravetz } 651557a196a5SMike Kravetz 6516a79390f5SPeter Xu long hugetlb_change_protection(struct vm_area_struct *vma, 65175a90d5a1SPeter Xu unsigned long address, unsigned long end, 65185a90d5a1SPeter Xu pgprot_t newprot, unsigned long cp_flags) 65198f860591SZhang, Yanmin { 65208f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 65218f860591SZhang, Yanmin unsigned long start = address; 65228f860591SZhang, Yanmin pte_t *ptep; 65238f860591SZhang, Yanmin pte_t pte; 6524a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 6525a79390f5SPeter Xu long pages = 0, psize = huge_page_size(h); 6526dff11abeSMike Kravetz bool shared_pmd = false; 6527ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 6528e95a9851SMike Kravetz unsigned long last_addr_mask; 65295a90d5a1SPeter Xu bool uffd_wp = cp_flags & MM_CP_UFFD_WP; 65305a90d5a1SPeter Xu bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE; 6531dff11abeSMike Kravetz 6532dff11abeSMike Kravetz /* 6533dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 6534ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 6535dff11abeSMike Kravetz * range if PMD sharing is possible. 6536dff11abeSMike Kravetz */ 65377269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 65387d4a8be0SAlistair Popple 0, mm, start, end); 6539ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 65408f860591SZhang, Yanmin 65418f860591SZhang, Yanmin BUG_ON(address >= end); 6542ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 65438f860591SZhang, Yanmin 6544ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 654540549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 654683cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 654740549ba8SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 654860dfaad6SPeter Xu for (; address < end; address += psize) { 6549cb900f41SKirill A. Shutemov spinlock_t *ptl; 65509c67a207SPeter Xu ptep = hugetlb_walk(vma, address, psize); 6551e95a9851SMike Kravetz if (!ptep) { 6552fed15f13SPeter Xu if (!uffd_wp) { 6553e95a9851SMike Kravetz address |= last_addr_mask; 65548f860591SZhang, Yanmin continue; 6555e95a9851SMike Kravetz } 6556fed15f13SPeter Xu /* 6557fed15f13SPeter Xu * Userfaultfd wr-protect requires pgtable 6558fed15f13SPeter Xu * pre-allocations to install pte markers. 6559fed15f13SPeter Xu */ 6560fed15f13SPeter Xu ptep = huge_pte_alloc(mm, vma, address, psize); 6561d1751118SPeter Xu if (!ptep) { 6562d1751118SPeter Xu pages = -ENOMEM; 6563fed15f13SPeter Xu break; 6564fed15f13SPeter Xu } 6565d1751118SPeter Xu } 6566cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 65674ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, address, ptep)) { 656860dfaad6SPeter Xu /* 656960dfaad6SPeter Xu * When uffd-wp is enabled on the vma, unshare 657060dfaad6SPeter Xu * shouldn't happen at all. Warn about it if it 657160dfaad6SPeter Xu * happened due to some reason. 657260dfaad6SPeter Xu */ 657360dfaad6SPeter Xu WARN_ON_ONCE(uffd_wp || uffd_wp_resolve); 65747da4d641SPeter Zijlstra pages++; 6575cb900f41SKirill A. Shutemov spin_unlock(ptl); 6576dff11abeSMike Kravetz shared_pmd = true; 65774ddb4d91SMike Kravetz address |= last_addr_mask; 657839dde65cSChen, Kenneth W continue; 65797da4d641SPeter Zijlstra } 6580a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 6581a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 65820e678153SDavid Hildenbrand /* Nothing to do. */ 65830e678153SDavid Hildenbrand } else if (unlikely(is_hugetlb_entry_migration(pte))) { 6584a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 65856c287605SDavid Hildenbrand struct page *page = pfn_swap_entry_to_page(entry); 658644f86392SDavid Hildenbrand pte_t newpte = pte; 6587a8bda28dSNaoya Horiguchi 658844f86392SDavid Hildenbrand if (is_writable_migration_entry(entry)) { 65896c287605SDavid Hildenbrand if (PageAnon(page)) 65906c287605SDavid Hildenbrand entry = make_readable_exclusive_migration_entry( 65916c287605SDavid Hildenbrand swp_offset(entry)); 65926c287605SDavid Hildenbrand else 65934dd845b5SAlistair Popple entry = make_readable_migration_entry( 65944dd845b5SAlistair Popple swp_offset(entry)); 6595a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 659644f86392SDavid Hildenbrand pages++; 659744f86392SDavid Hildenbrand } 659844f86392SDavid Hildenbrand 65995a90d5a1SPeter Xu if (uffd_wp) 66005a90d5a1SPeter Xu newpte = pte_swp_mkuffd_wp(newpte); 66015a90d5a1SPeter Xu else if (uffd_wp_resolve) 66025a90d5a1SPeter Xu newpte = pte_swp_clear_uffd_wp(newpte); 660344f86392SDavid Hildenbrand if (!pte_same(pte, newpte)) 6604935d4f0cSRyan Roberts set_huge_pte_at(mm, address, ptep, newpte, psize); 66050e678153SDavid Hildenbrand } else if (unlikely(is_pte_marker(pte))) { 66060e678153SDavid Hildenbrand /* No other markers apply for now. */ 66070e678153SDavid Hildenbrand WARN_ON_ONCE(!pte_marker_uffd_wp(pte)); 660860dfaad6SPeter Xu if (uffd_wp_resolve) 66090e678153SDavid Hildenbrand /* Safe to modify directly (non-present->none). */ 661060dfaad6SPeter Xu huge_pte_clear(mm, address, ptep, psize); 66110e678153SDavid Hildenbrand } else if (!huge_pte_none(pte)) { 6612023bdd00SAneesh Kumar K.V pte_t old_pte; 661379c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 6614023bdd00SAneesh Kumar K.V 6615023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 661616785bd7SAnshuman Khandual pte = huge_pte_modify(old_pte, newprot); 661779c1c594SChristophe Leroy pte = arch_make_huge_pte(pte, shift, vma->vm_flags); 66185a90d5a1SPeter Xu if (uffd_wp) 6619f1eb1bacSPeter Xu pte = huge_pte_mkuffd_wp(pte); 66205a90d5a1SPeter Xu else if (uffd_wp_resolve) 66215a90d5a1SPeter Xu pte = huge_pte_clear_uffd_wp(pte); 6622023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 66237da4d641SPeter Zijlstra pages++; 662460dfaad6SPeter Xu } else { 662560dfaad6SPeter Xu /* None pte */ 662660dfaad6SPeter Xu if (unlikely(uffd_wp)) 662760dfaad6SPeter Xu /* Safe to modify directly (none->non-present). */ 662860dfaad6SPeter Xu set_huge_pte_at(mm, address, ptep, 6629935d4f0cSRyan Roberts make_pte_marker(PTE_MARKER_UFFD_WP), 6630935d4f0cSRyan Roberts psize); 66318f860591SZhang, Yanmin } 6632cb900f41SKirill A. Shutemov spin_unlock(ptl); 66338f860591SZhang, Yanmin } 6634d833352aSMel Gorman /* 6635c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 6636d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 6637c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 6638dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 6639dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 6640d833352aSMel Gorman */ 6641dff11abeSMike Kravetz if (shared_pmd) 6642ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 6643dff11abeSMike Kravetz else 66445491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 66450f10851eSJérôme Glisse /* 66461af5a810SAlistair Popple * No need to call mmu_notifier_arch_invalidate_secondary_tlbs() we are 66471af5a810SAlistair Popple * downgrading page table protection not changing it to point to a new 66481af5a810SAlistair Popple * page. 66490f10851eSJérôme Glisse * 6650ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 66510f10851eSJérôme Glisse */ 665283cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 665340549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 6654ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 66557da4d641SPeter Zijlstra 6656d1751118SPeter Xu return pages > 0 ? (pages << h->order) : pages; 66578f860591SZhang, Yanmin } 66588f860591SZhang, Yanmin 665933b8f84aSMike Kravetz /* Return true if reservation was successful, false otherwise. */ 666033b8f84aSMike Kravetz bool hugetlb_reserve_pages(struct inode *inode, 6661a1e78772SMel Gorman long from, long to, 66625a6fe125SMel Gorman struct vm_area_struct *vma, 6663ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 6664e4e574b7SAdam Litke { 6665c5094ec7SMike Kravetz long chg = -1, add = -1; 6666a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 666790481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 66689119a41eSJoonsoo Kim struct resv_map *resv_map; 6669075a61d0SMina Almasry struct hugetlb_cgroup *h_cg = NULL; 66700db9d74eSMina Almasry long gbl_reserve, regions_needed = 0; 6671e4e574b7SAdam Litke 667263489f8eSMike Kravetz /* This should never happen */ 667363489f8eSMike Kravetz if (from > to) { 667463489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 667533b8f84aSMike Kravetz return false; 667663489f8eSMike Kravetz } 667763489f8eSMike Kravetz 6678a1e78772SMel Gorman /* 6679e700898fSMike Kravetz * vma specific semaphore used for pmd sharing and fault/truncation 6680e700898fSMike Kravetz * synchronization 66818d9bfb26SMike Kravetz */ 66828d9bfb26SMike Kravetz hugetlb_vma_lock_alloc(vma); 66838d9bfb26SMike Kravetz 66848d9bfb26SMike Kravetz /* 668517c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 668617c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 668790481622SDavid Gibson * without using reserves 668817c9d12eSMel Gorman */ 6689ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 669033b8f84aSMike Kravetz return true; 669117c9d12eSMel Gorman 669217c9d12eSMel Gorman /* 6693a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 6694a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 6695a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 6696a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 6697a1e78772SMel Gorman */ 66989119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 6699f27a5136SMike Kravetz /* 6700f27a5136SMike Kravetz * resv_map can not be NULL as hugetlb_reserve_pages is only 6701f27a5136SMike Kravetz * called for inodes for which resv_maps were created (see 6702f27a5136SMike Kravetz * hugetlbfs_get_inode). 6703f27a5136SMike Kravetz */ 67044e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 67059119a41eSJoonsoo Kim 67060db9d74eSMina Almasry chg = region_chg(resv_map, from, to, ®ions_needed); 67079119a41eSJoonsoo Kim } else { 6708e9fe92aeSMina Almasry /* Private mapping. */ 67099119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 67105a6fe125SMel Gorman if (!resv_map) 67118d9bfb26SMike Kravetz goto out_err; 67125a6fe125SMel Gorman 671317c9d12eSMel Gorman chg = to - from; 671417c9d12eSMel Gorman 67155a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 67165a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 67175a6fe125SMel Gorman } 67185a6fe125SMel Gorman 671933b8f84aSMike Kravetz if (chg < 0) 6720c50ac050SDave Hansen goto out_err; 672117c9d12eSMel Gorman 672233b8f84aSMike Kravetz if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), 672333b8f84aSMike Kravetz chg * pages_per_huge_page(h), &h_cg) < 0) 6724075a61d0SMina Almasry goto out_err; 6725075a61d0SMina Almasry 6726075a61d0SMina Almasry if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { 6727075a61d0SMina Almasry /* For private mappings, the hugetlb_cgroup uncharge info hangs 6728075a61d0SMina Almasry * of the resv_map. 6729075a61d0SMina Almasry */ 6730075a61d0SMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); 6731075a61d0SMina Almasry } 6732075a61d0SMina Almasry 67331c5ecae3SMike Kravetz /* 67341c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 67351c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 67361c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 67371c5ecae3SMike Kravetz */ 67381c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 673933b8f84aSMike Kravetz if (gbl_reserve < 0) 6740075a61d0SMina Almasry goto out_uncharge_cgroup; 674117c9d12eSMel Gorman 674217c9d12eSMel Gorman /* 674317c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 674490481622SDavid Gibson * Hand the pages back to the subpool if there are not 674517c9d12eSMel Gorman */ 674633b8f84aSMike Kravetz if (hugetlb_acct_memory(h, gbl_reserve) < 0) 6747075a61d0SMina Almasry goto out_put_pages; 674817c9d12eSMel Gorman 674917c9d12eSMel Gorman /* 675017c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 675117c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 675217c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 675317c9d12eSMel Gorman * the reservation was and the page cache tells how much of 675417c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 675517c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 675617c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 675717c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 675817c9d12eSMel Gorman * else has to be done for private mappings here 675917c9d12eSMel Gorman */ 676033039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 6761075a61d0SMina Almasry add = region_add(resv_map, from, to, regions_needed, h, h_cg); 676233039678SMike Kravetz 67630db9d74eSMina Almasry if (unlikely(add < 0)) { 67640db9d74eSMina Almasry hugetlb_acct_memory(h, -gbl_reserve); 6765075a61d0SMina Almasry goto out_put_pages; 67660db9d74eSMina Almasry } else if (unlikely(chg > add)) { 676733039678SMike Kravetz /* 676833039678SMike Kravetz * pages in this range were added to the reserve 676933039678SMike Kravetz * map between region_chg and region_add. This 6770d0ce0e47SSidhartha Kumar * indicates a race with alloc_hugetlb_folio. Adjust 677133039678SMike Kravetz * the subpool and reserve counts modified above 677233039678SMike Kravetz * based on the difference. 677333039678SMike Kravetz */ 677433039678SMike Kravetz long rsv_adjust; 677533039678SMike Kravetz 6776d85aecf2SMiaohe Lin /* 6777d85aecf2SMiaohe Lin * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the 6778d85aecf2SMiaohe Lin * reference to h_cg->css. See comment below for detail. 6779d85aecf2SMiaohe Lin */ 6780075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd( 6781075a61d0SMina Almasry hstate_index(h), 6782075a61d0SMina Almasry (chg - add) * pages_per_huge_page(h), h_cg); 6783075a61d0SMina Almasry 678433039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 678533039678SMike Kravetz chg - add); 678633039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 6787d85aecf2SMiaohe Lin } else if (h_cg) { 6788d85aecf2SMiaohe Lin /* 6789d85aecf2SMiaohe Lin * The file_regions will hold their own reference to 6790d85aecf2SMiaohe Lin * h_cg->css. So we should release the reference held 6791d85aecf2SMiaohe Lin * via hugetlb_cgroup_charge_cgroup_rsvd() when we are 6792d85aecf2SMiaohe Lin * done. 6793d85aecf2SMiaohe Lin */ 6794d85aecf2SMiaohe Lin hugetlb_cgroup_put_rsvd_cgroup(h_cg); 679533039678SMike Kravetz } 679633039678SMike Kravetz } 679733b8f84aSMike Kravetz return true; 679833b8f84aSMike Kravetz 6799075a61d0SMina Almasry out_put_pages: 6800075a61d0SMina Almasry /* put back original number of pages, chg */ 6801075a61d0SMina Almasry (void)hugepage_subpool_put_pages(spool, chg); 6802075a61d0SMina Almasry out_uncharge_cgroup: 6803075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), 6804075a61d0SMina Almasry chg * pages_per_huge_page(h), h_cg); 6805c50ac050SDave Hansen out_err: 68068d9bfb26SMike Kravetz hugetlb_vma_lock_free(vma); 68075e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 68080db9d74eSMina Almasry /* Only call region_abort if the region_chg succeeded but the 68090db9d74eSMina Almasry * region_add failed or didn't run. 68100db9d74eSMina Almasry */ 68110db9d74eSMina Almasry if (chg >= 0 && add < 0) 68120db9d74eSMina Almasry region_abort(resv_map, from, to, regions_needed); 6813*92fe9dcbSRik van Riel if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 6814f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 6815*92fe9dcbSRik van Riel set_vma_resv_map(vma, NULL); 6816*92fe9dcbSRik van Riel } 681733b8f84aSMike Kravetz return false; 6818a43a8c39SChen, Kenneth W } 6819a43a8c39SChen, Kenneth W 6820b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 6821b5cec28dSMike Kravetz long freed) 6822a43a8c39SChen, Kenneth W { 6823a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 68244e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 68259119a41eSJoonsoo Kim long chg = 0; 682690481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 68271c5ecae3SMike Kravetz long gbl_reserve; 682845c682a6SKen Chen 6829f27a5136SMike Kravetz /* 6830f27a5136SMike Kravetz * Since this routine can be called in the evict inode path for all 6831f27a5136SMike Kravetz * hugetlbfs inodes, resv_map could be NULL. 6832f27a5136SMike Kravetz */ 6833b5cec28dSMike Kravetz if (resv_map) { 6834b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 6835b5cec28dSMike Kravetz /* 6836b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 6837b5cec28dSMike Kravetz * must be split and another region descriptor can not be 6838b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 6839b5cec28dSMike Kravetz */ 6840b5cec28dSMike Kravetz if (chg < 0) 6841b5cec28dSMike Kravetz return chg; 6842b5cec28dSMike Kravetz } 6843b5cec28dSMike Kravetz 684445c682a6SKen Chen spin_lock(&inode->i_lock); 6845e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 684645c682a6SKen Chen spin_unlock(&inode->i_lock); 684745c682a6SKen Chen 68481c5ecae3SMike Kravetz /* 68491c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 68501c5ecae3SMike Kravetz * reservations to be released may be adjusted. 6851dddf31a4SMiaohe Lin * 6852dddf31a4SMiaohe Lin * Note that !resv_map implies freed == 0. So (chg - freed) 6853dddf31a4SMiaohe Lin * won't go negative. 68541c5ecae3SMike Kravetz */ 68551c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 68561c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 6857b5cec28dSMike Kravetz 6858b5cec28dSMike Kravetz return 0; 6859a43a8c39SChen, Kenneth W } 686093f70f90SNaoya Horiguchi 68613212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 68623212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 68633212b535SSteve Capper struct vm_area_struct *vma, 68643212b535SSteve Capper unsigned long addr, pgoff_t idx) 68653212b535SSteve Capper { 68663212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 68673212b535SSteve Capper svma->vm_start; 68683212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 68693212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 68703212b535SSteve Capper 68713212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 6872e430a95aSSuren Baghdasaryan unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED_MASK; 6873e430a95aSSuren Baghdasaryan unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED_MASK; 68743212b535SSteve Capper 68753212b535SSteve Capper /* 68763212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 68773212b535SSteve Capper * page table page. 6878131a79b4SMike Kravetz * 6879131a79b4SMike Kravetz * Also, vma_lock (vm_private_data) is required for sharing. 68803212b535SSteve Capper */ 68813212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 68823212b535SSteve Capper vm_flags != svm_flags || 6883131a79b4SMike Kravetz !range_in_vma(svma, sbase, s_end) || 6884131a79b4SMike Kravetz !svma->vm_private_data) 68853212b535SSteve Capper return 0; 68863212b535SSteve Capper 68873212b535SSteve Capper return saddr; 68883212b535SSteve Capper } 68893212b535SSteve Capper 6890bbff39ccSMike Kravetz bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 68913212b535SSteve Capper { 6892bbff39ccSMike Kravetz unsigned long start = addr & PUD_MASK; 6893bbff39ccSMike Kravetz unsigned long end = start + PUD_SIZE; 6894bbff39ccSMike Kravetz 68958d9bfb26SMike Kravetz #ifdef CONFIG_USERFAULTFD 68968d9bfb26SMike Kravetz if (uffd_disable_huge_pmd_share(vma)) 68978d9bfb26SMike Kravetz return false; 68988d9bfb26SMike Kravetz #endif 68993212b535SSteve Capper /* 69003212b535SSteve Capper * check on proper vm_flags and page table alignment 69013212b535SSteve Capper */ 69028d9bfb26SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 690331aafb45SNicholas Krause return false; 6904bbff39ccSMike Kravetz if (!vma->vm_private_data) /* vma lock required for sharing */ 69058d9bfb26SMike Kravetz return false; 69068d9bfb26SMike Kravetz if (!range_in_vma(vma, start, end)) 69078d9bfb26SMike Kravetz return false; 69088d9bfb26SMike Kravetz return true; 69098d9bfb26SMike Kravetz } 69108d9bfb26SMike Kravetz 69113212b535SSteve Capper /* 6912017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 6913017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 6914017b1660SMike Kravetz * shared pmd mappings. 6915017b1660SMike Kravetz */ 6916017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 6917017b1660SMike Kravetz unsigned long *start, unsigned long *end) 6918017b1660SMike Kravetz { 6919a1ba9da8SLi Xinhai unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), 6920a1ba9da8SLi Xinhai v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 6921017b1660SMike Kravetz 6922a1ba9da8SLi Xinhai /* 6923f0953a1bSIngo Molnar * vma needs to span at least one aligned PUD size, and the range 6924f0953a1bSIngo Molnar * must be at least partially within in. 6925a1ba9da8SLi Xinhai */ 6926a1ba9da8SLi Xinhai if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || 6927a1ba9da8SLi Xinhai (*end <= v_start) || (*start >= v_end)) 6928017b1660SMike Kravetz return; 6929017b1660SMike Kravetz 693075802ca6SPeter Xu /* Extend the range to be PUD aligned for a worst case scenario */ 6931a1ba9da8SLi Xinhai if (*start > v_start) 6932a1ba9da8SLi Xinhai *start = ALIGN_DOWN(*start, PUD_SIZE); 6933017b1660SMike Kravetz 6934a1ba9da8SLi Xinhai if (*end < v_end) 6935a1ba9da8SLi Xinhai *end = ALIGN(*end, PUD_SIZE); 6936017b1660SMike Kravetz } 6937017b1660SMike Kravetz 6938017b1660SMike Kravetz /* 69393212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 69403212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 69413212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 69423a47c54fSMike Kravetz * code much cleaner. pmd allocation is essential for the shared case because 69433a47c54fSMike Kravetz * pud has to be populated inside the same i_mmap_rwsem section - otherwise 69443a47c54fSMike Kravetz * racing tasks could either miss the sharing (see huge_pte_offset) or select a 69453a47c54fSMike Kravetz * bad pmd for sharing. 69463212b535SSteve Capper */ 6947aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 6948aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 69493212b535SSteve Capper { 69503212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 69513212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 69523212b535SSteve Capper vma->vm_pgoff; 69533212b535SSteve Capper struct vm_area_struct *svma; 69543212b535SSteve Capper unsigned long saddr; 69553212b535SSteve Capper pte_t *spte = NULL; 69563212b535SSteve Capper pte_t *pte; 69573212b535SSteve Capper 69583a47c54fSMike Kravetz i_mmap_lock_read(mapping); 69593212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 69603212b535SSteve Capper if (svma == vma) 69613212b535SSteve Capper continue; 69623212b535SSteve Capper 69633212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 69643212b535SSteve Capper if (saddr) { 69659c67a207SPeter Xu spte = hugetlb_walk(svma, saddr, 69667868a208SPunit Agrawal vma_mmu_pagesize(svma)); 69673212b535SSteve Capper if (spte) { 69683212b535SSteve Capper get_page(virt_to_page(spte)); 69693212b535SSteve Capper break; 69703212b535SSteve Capper } 69713212b535SSteve Capper } 69723212b535SSteve Capper } 69733212b535SSteve Capper 69743212b535SSteve Capper if (!spte) 69753212b535SSteve Capper goto out; 69763212b535SSteve Capper 6977349d1670SPeter Xu spin_lock(&mm->page_table_lock); 6978dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 69793212b535SSteve Capper pud_populate(mm, pud, 69803212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 6981c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 6982dc6c9a35SKirill A. Shutemov } else { 69833212b535SSteve Capper put_page(virt_to_page(spte)); 6984dc6c9a35SKirill A. Shutemov } 6985349d1670SPeter Xu spin_unlock(&mm->page_table_lock); 69863212b535SSteve Capper out: 69873212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 69883a47c54fSMike Kravetz i_mmap_unlock_read(mapping); 69893212b535SSteve Capper return pte; 69903212b535SSteve Capper } 69913212b535SSteve Capper 69923212b535SSteve Capper /* 69933212b535SSteve Capper * unmap huge page backed by shared pte. 69943212b535SSteve Capper * 69953212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 69963212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 69973212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 69983212b535SSteve Capper * 69993a47c54fSMike Kravetz * Called with page table lock held. 70003212b535SSteve Capper * 70013212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 70023212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 70033212b535SSteve Capper */ 700434ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 70054ddb4d91SMike Kravetz unsigned long addr, pte_t *ptep) 70063212b535SSteve Capper { 70074ddb4d91SMike Kravetz pgd_t *pgd = pgd_offset(mm, addr); 70084ddb4d91SMike Kravetz p4d_t *p4d = p4d_offset(pgd, addr); 70094ddb4d91SMike Kravetz pud_t *pud = pud_offset(p4d, addr); 70103212b535SSteve Capper 701134ae204fSMike Kravetz i_mmap_assert_write_locked(vma->vm_file->f_mapping); 701240549ba8SMike Kravetz hugetlb_vma_assert_locked(vma); 70133212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 70143212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 70153212b535SSteve Capper return 0; 70163212b535SSteve Capper 70173212b535SSteve Capper pud_clear(pud); 70183212b535SSteve Capper put_page(virt_to_page(ptep)); 7019dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 70203212b535SSteve Capper return 1; 70213212b535SSteve Capper } 7022c1991e07SPeter Xu 70239e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 70248d9bfb26SMike Kravetz 7025aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 7026aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 70279e5fc74cSSteve Capper { 70289e5fc74cSSteve Capper return NULL; 70299e5fc74cSSteve Capper } 7030e81f2d22SZhang Zhen 703134ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 70324ddb4d91SMike Kravetz unsigned long addr, pte_t *ptep) 7033e81f2d22SZhang Zhen { 7034e81f2d22SZhang Zhen return 0; 7035e81f2d22SZhang Zhen } 7036017b1660SMike Kravetz 7037017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 7038017b1660SMike Kravetz unsigned long *start, unsigned long *end) 7039017b1660SMike Kravetz { 7040017b1660SMike Kravetz } 7041c1991e07SPeter Xu 7042c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 7043c1991e07SPeter Xu { 7044c1991e07SPeter Xu return false; 7045c1991e07SPeter Xu } 70463212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 70473212b535SSteve Capper 70489e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 7049aec44e0fSPeter Xu pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, 70509e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 70519e5fc74cSSteve Capper { 70529e5fc74cSSteve Capper pgd_t *pgd; 7053c2febafcSKirill A. Shutemov p4d_t *p4d; 70549e5fc74cSSteve Capper pud_t *pud; 70559e5fc74cSSteve Capper pte_t *pte = NULL; 70569e5fc74cSSteve Capper 70579e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 7058f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 7059f4f0a3d8SKirill A. Shutemov if (!p4d) 7060f4f0a3d8SKirill A. Shutemov return NULL; 7061c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 70629e5fc74cSSteve Capper if (pud) { 70639e5fc74cSSteve Capper if (sz == PUD_SIZE) { 70649e5fc74cSSteve Capper pte = (pte_t *)pud; 70659e5fc74cSSteve Capper } else { 70669e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 7067c1991e07SPeter Xu if (want_pmd_share(vma, addr) && pud_none(*pud)) 7068aec44e0fSPeter Xu pte = huge_pmd_share(mm, vma, addr, pud); 70699e5fc74cSSteve Capper else 70709e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 70719e5fc74cSSteve Capper } 70729e5fc74cSSteve Capper } 7073191fcdb6SJohn Hubbard 7074191fcdb6SJohn Hubbard if (pte) { 7075191fcdb6SJohn Hubbard pte_t pteval = ptep_get_lockless(pte); 7076191fcdb6SJohn Hubbard 7077191fcdb6SJohn Hubbard BUG_ON(pte_present(pteval) && !pte_huge(pteval)); 7078191fcdb6SJohn Hubbard } 70799e5fc74cSSteve Capper 70809e5fc74cSSteve Capper return pte; 70819e5fc74cSSteve Capper } 70829e5fc74cSSteve Capper 70839b19df29SPunit Agrawal /* 70849b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 70859b19df29SPunit Agrawal * entry at address @addr 70869b19df29SPunit Agrawal * 70878ac0b81aSLi Xinhai * Return: Pointer to page table entry (PUD or PMD) for 70888ac0b81aSLi Xinhai * address @addr, or NULL if a !p*d_present() entry is encountered and the 70899b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 70909b19df29SPunit Agrawal * table. 70919b19df29SPunit Agrawal */ 70927868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 70937868a208SPunit Agrawal unsigned long addr, unsigned long sz) 70949e5fc74cSSteve Capper { 70959e5fc74cSSteve Capper pgd_t *pgd; 7096c2febafcSKirill A. Shutemov p4d_t *p4d; 70978ac0b81aSLi Xinhai pud_t *pud; 70988ac0b81aSLi Xinhai pmd_t *pmd; 70999e5fc74cSSteve Capper 71009e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 7101c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 7102c2febafcSKirill A. Shutemov return NULL; 7103c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 7104c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 7105c2febafcSKirill A. Shutemov return NULL; 71069b19df29SPunit Agrawal 7107c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 71088ac0b81aSLi Xinhai if (sz == PUD_SIZE) 71098ac0b81aSLi Xinhai /* must be pud huge, non-present or none */ 71109e5fc74cSSteve Capper return (pte_t *)pud; 71118ac0b81aSLi Xinhai if (!pud_present(*pud)) 71128ac0b81aSLi Xinhai return NULL; 71138ac0b81aSLi Xinhai /* must have a valid entry and size to go further */ 71149b19df29SPunit Agrawal 71159e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 71168ac0b81aSLi Xinhai /* must be pmd huge, non-present or none */ 71179e5fc74cSSteve Capper return (pte_t *)pmd; 71189e5fc74cSSteve Capper } 71199e5fc74cSSteve Capper 7120e95a9851SMike Kravetz /* 7121e95a9851SMike Kravetz * Return a mask that can be used to update an address to the last huge 7122e95a9851SMike Kravetz * page in a page table page mapping size. Used to skip non-present 7123e95a9851SMike Kravetz * page table entries when linearly scanning address ranges. Architectures 7124e95a9851SMike Kravetz * with unique huge page to page table relationships can define their own 7125e95a9851SMike Kravetz * version of this routine. 7126e95a9851SMike Kravetz */ 7127e95a9851SMike Kravetz unsigned long hugetlb_mask_last_page(struct hstate *h) 7128e95a9851SMike Kravetz { 7129e95a9851SMike Kravetz unsigned long hp_size = huge_page_size(h); 7130e95a9851SMike Kravetz 7131e95a9851SMike Kravetz if (hp_size == PUD_SIZE) 7132e95a9851SMike Kravetz return P4D_SIZE - PUD_SIZE; 7133e95a9851SMike Kravetz else if (hp_size == PMD_SIZE) 7134e95a9851SMike Kravetz return PUD_SIZE - PMD_SIZE; 7135e95a9851SMike Kravetz else 7136e95a9851SMike Kravetz return 0UL; 7137e95a9851SMike Kravetz } 7138e95a9851SMike Kravetz 7139e95a9851SMike Kravetz #else 7140e95a9851SMike Kravetz 7141e95a9851SMike Kravetz /* See description above. Architectures can provide their own version. */ 7142e95a9851SMike Kravetz __weak unsigned long hugetlb_mask_last_page(struct hstate *h) 7143e95a9851SMike Kravetz { 71444ddb4d91SMike Kravetz #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 71454ddb4d91SMike Kravetz if (huge_page_size(h) == PMD_SIZE) 71464ddb4d91SMike Kravetz return PUD_SIZE - PMD_SIZE; 71474ddb4d91SMike Kravetz #endif 7148e95a9851SMike Kravetz return 0UL; 7149e95a9851SMike Kravetz } 7150e95a9851SMike Kravetz 715161f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 715261f77edaSNaoya Horiguchi 715361f77edaSNaoya Horiguchi /* 715461f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 715561f77edaSNaoya Horiguchi * behavior. 715661f77edaSNaoya Horiguchi */ 71579747b9e9SBaolin Wang bool isolate_hugetlb(struct folio *folio, struct list_head *list) 715831caf665SNaoya Horiguchi { 71599747b9e9SBaolin Wang bool ret = true; 7160bcc54222SNaoya Horiguchi 7161db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 71626aa3a920SSidhartha Kumar if (!folio_test_hugetlb(folio) || 71636aa3a920SSidhartha Kumar !folio_test_hugetlb_migratable(folio) || 71646aa3a920SSidhartha Kumar !folio_try_get(folio)) { 71659747b9e9SBaolin Wang ret = false; 7166bcc54222SNaoya Horiguchi goto unlock; 7167bcc54222SNaoya Horiguchi } 71686aa3a920SSidhartha Kumar folio_clear_hugetlb_migratable(folio); 71696aa3a920SSidhartha Kumar list_move_tail(&folio->lru, list); 7170bcc54222SNaoya Horiguchi unlock: 7171db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7172bcc54222SNaoya Horiguchi return ret; 717331caf665SNaoya Horiguchi } 717431caf665SNaoya Horiguchi 717504bac040SSidhartha Kumar int get_hwpoison_hugetlb_folio(struct folio *folio, bool *hugetlb, bool unpoison) 717625182f05SNaoya Horiguchi { 717725182f05SNaoya Horiguchi int ret = 0; 717825182f05SNaoya Horiguchi 717925182f05SNaoya Horiguchi *hugetlb = false; 718025182f05SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 718104bac040SSidhartha Kumar if (folio_test_hugetlb(folio)) { 718225182f05SNaoya Horiguchi *hugetlb = true; 718304bac040SSidhartha Kumar if (folio_test_hugetlb_freed(folio)) 7184b283d983SNaoya Horiguchi ret = 0; 718504bac040SSidhartha Kumar else if (folio_test_hugetlb_migratable(folio) || unpoison) 718604bac040SSidhartha Kumar ret = folio_try_get(folio); 71870ed950d1SNaoya Horiguchi else 71880ed950d1SNaoya Horiguchi ret = -EBUSY; 718925182f05SNaoya Horiguchi } 719025182f05SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 719125182f05SNaoya Horiguchi return ret; 719225182f05SNaoya Horiguchi } 719325182f05SNaoya Horiguchi 7194e591ef7dSNaoya Horiguchi int get_huge_page_for_hwpoison(unsigned long pfn, int flags, 7195e591ef7dSNaoya Horiguchi bool *migratable_cleared) 7196405ce051SNaoya Horiguchi { 7197405ce051SNaoya Horiguchi int ret; 7198405ce051SNaoya Horiguchi 7199405ce051SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 7200e591ef7dSNaoya Horiguchi ret = __get_huge_page_for_hwpoison(pfn, flags, migratable_cleared); 7201405ce051SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 7202405ce051SNaoya Horiguchi return ret; 7203405ce051SNaoya Horiguchi } 7204405ce051SNaoya Horiguchi 7205ea8e72f4SSidhartha Kumar void folio_putback_active_hugetlb(struct folio *folio) 720631caf665SNaoya Horiguchi { 7207db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 7208ea8e72f4SSidhartha Kumar folio_set_hugetlb_migratable(folio); 7209ea8e72f4SSidhartha Kumar list_move_tail(&folio->lru, &(folio_hstate(folio))->hugepage_activelist); 7210db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7211ea8e72f4SSidhartha Kumar folio_put(folio); 721231caf665SNaoya Horiguchi } 7213ab5ac90aSMichal Hocko 7214345c62d1SSidhartha Kumar void move_hugetlb_state(struct folio *old_folio, struct folio *new_folio, int reason) 7215ab5ac90aSMichal Hocko { 7216345c62d1SSidhartha Kumar struct hstate *h = folio_hstate(old_folio); 7217ab5ac90aSMichal Hocko 7218345c62d1SSidhartha Kumar hugetlb_cgroup_migrate(old_folio, new_folio); 7219345c62d1SSidhartha Kumar set_page_owner_migrate_reason(&new_folio->page, reason); 7220ab5ac90aSMichal Hocko 7221ab5ac90aSMichal Hocko /* 7222345c62d1SSidhartha Kumar * transfer temporary state of the new hugetlb folio. This is 7223ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 7224ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 7225ab5ac90aSMichal Hocko * the temporary status. 7226ab5ac90aSMichal Hocko * 7227ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 7228ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 7229ab5ac90aSMichal Hocko * the per-node's. 7230ab5ac90aSMichal Hocko */ 7231345c62d1SSidhartha Kumar if (folio_test_hugetlb_temporary(new_folio)) { 7232345c62d1SSidhartha Kumar int old_nid = folio_nid(old_folio); 7233345c62d1SSidhartha Kumar int new_nid = folio_nid(new_folio); 7234ab5ac90aSMichal Hocko 7235345c62d1SSidhartha Kumar folio_set_hugetlb_temporary(old_folio); 7236345c62d1SSidhartha Kumar folio_clear_hugetlb_temporary(new_folio); 7237345c62d1SSidhartha Kumar 7238ab5ac90aSMichal Hocko 72395af1ab1dSMiaohe Lin /* 72405af1ab1dSMiaohe Lin * There is no need to transfer the per-node surplus state 72415af1ab1dSMiaohe Lin * when we do not cross the node. 72425af1ab1dSMiaohe Lin */ 72435af1ab1dSMiaohe Lin if (new_nid == old_nid) 72445af1ab1dSMiaohe Lin return; 7245db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 7246ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 7247ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 7248ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 7249ab5ac90aSMichal Hocko } 7250db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7251ab5ac90aSMichal Hocko } 7252ab5ac90aSMichal Hocko } 7253cf11e85fSRoman Gushchin 7254b30c14cdSJames Houghton static void hugetlb_unshare_pmds(struct vm_area_struct *vma, 7255b30c14cdSJames Houghton unsigned long start, 7256b30c14cdSJames Houghton unsigned long end) 72576dfeaff9SPeter Xu { 72586dfeaff9SPeter Xu struct hstate *h = hstate_vma(vma); 72596dfeaff9SPeter Xu unsigned long sz = huge_page_size(h); 72606dfeaff9SPeter Xu struct mm_struct *mm = vma->vm_mm; 72616dfeaff9SPeter Xu struct mmu_notifier_range range; 7262b30c14cdSJames Houghton unsigned long address; 72636dfeaff9SPeter Xu spinlock_t *ptl; 72646dfeaff9SPeter Xu pte_t *ptep; 72656dfeaff9SPeter Xu 72666dfeaff9SPeter Xu if (!(vma->vm_flags & VM_MAYSHARE)) 72676dfeaff9SPeter Xu return; 72686dfeaff9SPeter Xu 72696dfeaff9SPeter Xu if (start >= end) 72706dfeaff9SPeter Xu return; 72716dfeaff9SPeter Xu 72729c8bbfacSBaolin Wang flush_cache_range(vma, start, end); 72736dfeaff9SPeter Xu /* 72746dfeaff9SPeter Xu * No need to call adjust_range_if_pmd_sharing_possible(), because 72756dfeaff9SPeter Xu * we have already done the PUD_SIZE alignment. 72766dfeaff9SPeter Xu */ 72777d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, 72786dfeaff9SPeter Xu start, end); 72796dfeaff9SPeter Xu mmu_notifier_invalidate_range_start(&range); 728040549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 72816dfeaff9SPeter Xu i_mmap_lock_write(vma->vm_file->f_mapping); 72826dfeaff9SPeter Xu for (address = start; address < end; address += PUD_SIZE) { 72839c67a207SPeter Xu ptep = hugetlb_walk(vma, address, sz); 72846dfeaff9SPeter Xu if (!ptep) 72856dfeaff9SPeter Xu continue; 72866dfeaff9SPeter Xu ptl = huge_pte_lock(h, mm, ptep); 72874ddb4d91SMike Kravetz huge_pmd_unshare(mm, vma, address, ptep); 72886dfeaff9SPeter Xu spin_unlock(ptl); 72896dfeaff9SPeter Xu } 72906dfeaff9SPeter Xu flush_hugetlb_tlb_range(vma, start, end); 72916dfeaff9SPeter Xu i_mmap_unlock_write(vma->vm_file->f_mapping); 729240549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 72936dfeaff9SPeter Xu /* 72941af5a810SAlistair Popple * No need to call mmu_notifier_arch_invalidate_secondary_tlbs(), see 7295ee65728eSMike Rapoport * Documentation/mm/mmu_notifier.rst. 72966dfeaff9SPeter Xu */ 72976dfeaff9SPeter Xu mmu_notifier_invalidate_range_end(&range); 72986dfeaff9SPeter Xu } 72996dfeaff9SPeter Xu 7300b30c14cdSJames Houghton /* 7301b30c14cdSJames Houghton * This function will unconditionally remove all the shared pmd pgtable entries 7302b30c14cdSJames Houghton * within the specific vma for a hugetlbfs memory range. 7303b30c14cdSJames Houghton */ 7304b30c14cdSJames Houghton void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) 7305b30c14cdSJames Houghton { 7306b30c14cdSJames Houghton hugetlb_unshare_pmds(vma, ALIGN(vma->vm_start, PUD_SIZE), 7307b30c14cdSJames Houghton ALIGN_DOWN(vma->vm_end, PUD_SIZE)); 7308b30c14cdSJames Houghton } 7309b30c14cdSJames Houghton 7310cf11e85fSRoman Gushchin #ifdef CONFIG_CMA 7311cf11e85fSRoman Gushchin static bool cma_reserve_called __initdata; 7312cf11e85fSRoman Gushchin 7313cf11e85fSRoman Gushchin static int __init cmdline_parse_hugetlb_cma(char *p) 7314cf11e85fSRoman Gushchin { 731538e719abSBaolin Wang int nid, count = 0; 731638e719abSBaolin Wang unsigned long tmp; 731738e719abSBaolin Wang char *s = p; 731838e719abSBaolin Wang 731938e719abSBaolin Wang while (*s) { 732038e719abSBaolin Wang if (sscanf(s, "%lu%n", &tmp, &count) != 1) 732138e719abSBaolin Wang break; 732238e719abSBaolin Wang 732338e719abSBaolin Wang if (s[count] == ':') { 7324f9317f77SMike Kravetz if (tmp >= MAX_NUMNODES) 732538e719abSBaolin Wang break; 7326f9317f77SMike Kravetz nid = array_index_nospec(tmp, MAX_NUMNODES); 732738e719abSBaolin Wang 732838e719abSBaolin Wang s += count + 1; 732938e719abSBaolin Wang tmp = memparse(s, &s); 733038e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = tmp; 733138e719abSBaolin Wang hugetlb_cma_size += tmp; 733238e719abSBaolin Wang 733338e719abSBaolin Wang /* 733438e719abSBaolin Wang * Skip the separator if have one, otherwise 733538e719abSBaolin Wang * break the parsing. 733638e719abSBaolin Wang */ 733738e719abSBaolin Wang if (*s == ',') 733838e719abSBaolin Wang s++; 733938e719abSBaolin Wang else 734038e719abSBaolin Wang break; 734138e719abSBaolin Wang } else { 7342cf11e85fSRoman Gushchin hugetlb_cma_size = memparse(p, &p); 734338e719abSBaolin Wang break; 734438e719abSBaolin Wang } 734538e719abSBaolin Wang } 734638e719abSBaolin Wang 7347cf11e85fSRoman Gushchin return 0; 7348cf11e85fSRoman Gushchin } 7349cf11e85fSRoman Gushchin 7350cf11e85fSRoman Gushchin early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); 7351cf11e85fSRoman Gushchin 7352cf11e85fSRoman Gushchin void __init hugetlb_cma_reserve(int order) 7353cf11e85fSRoman Gushchin { 7354cf11e85fSRoman Gushchin unsigned long size, reserved, per_node; 735538e719abSBaolin Wang bool node_specific_cma_alloc = false; 7356cf11e85fSRoman Gushchin int nid; 7357cf11e85fSRoman Gushchin 7358cf11e85fSRoman Gushchin cma_reserve_called = true; 7359cf11e85fSRoman Gushchin 7360cf11e85fSRoman Gushchin if (!hugetlb_cma_size) 7361cf11e85fSRoman Gushchin return; 7362cf11e85fSRoman Gushchin 736338e719abSBaolin Wang for (nid = 0; nid < MAX_NUMNODES; nid++) { 736438e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 736538e719abSBaolin Wang continue; 736638e719abSBaolin Wang 736730a51400SPeng Liu if (!node_online(nid)) { 736838e719abSBaolin Wang pr_warn("hugetlb_cma: invalid node %d specified\n", nid); 736938e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 737038e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 737138e719abSBaolin Wang continue; 737238e719abSBaolin Wang } 737338e719abSBaolin Wang 737438e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) { 737538e719abSBaolin Wang pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n", 737638e719abSBaolin Wang nid, (PAGE_SIZE << order) / SZ_1M); 737738e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 737838e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 737938e719abSBaolin Wang } else { 738038e719abSBaolin Wang node_specific_cma_alloc = true; 738138e719abSBaolin Wang } 738238e719abSBaolin Wang } 738338e719abSBaolin Wang 738438e719abSBaolin Wang /* Validate the CMA size again in case some invalid nodes specified. */ 738538e719abSBaolin Wang if (!hugetlb_cma_size) 738638e719abSBaolin Wang return; 738738e719abSBaolin Wang 7388cf11e85fSRoman Gushchin if (hugetlb_cma_size < (PAGE_SIZE << order)) { 7389cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", 7390cf11e85fSRoman Gushchin (PAGE_SIZE << order) / SZ_1M); 7391a01f4390SMike Kravetz hugetlb_cma_size = 0; 7392cf11e85fSRoman Gushchin return; 7393cf11e85fSRoman Gushchin } 7394cf11e85fSRoman Gushchin 739538e719abSBaolin Wang if (!node_specific_cma_alloc) { 7396cf11e85fSRoman Gushchin /* 7397cf11e85fSRoman Gushchin * If 3 GB area is requested on a machine with 4 numa nodes, 7398cf11e85fSRoman Gushchin * let's allocate 1 GB on first three nodes and ignore the last one. 7399cf11e85fSRoman Gushchin */ 7400cf11e85fSRoman Gushchin per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); 7401cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", 7402cf11e85fSRoman Gushchin hugetlb_cma_size / SZ_1M, per_node / SZ_1M); 740338e719abSBaolin Wang } 7404cf11e85fSRoman Gushchin 7405cf11e85fSRoman Gushchin reserved = 0; 740630a51400SPeng Liu for_each_online_node(nid) { 7407cf11e85fSRoman Gushchin int res; 74082281f797SBarry Song char name[CMA_MAX_NAME]; 7409cf11e85fSRoman Gushchin 741038e719abSBaolin Wang if (node_specific_cma_alloc) { 741138e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 741238e719abSBaolin Wang continue; 741338e719abSBaolin Wang 741438e719abSBaolin Wang size = hugetlb_cma_size_in_node[nid]; 741538e719abSBaolin Wang } else { 7416cf11e85fSRoman Gushchin size = min(per_node, hugetlb_cma_size - reserved); 741738e719abSBaolin Wang } 741838e719abSBaolin Wang 7419cf11e85fSRoman Gushchin size = round_up(size, PAGE_SIZE << order); 7420cf11e85fSRoman Gushchin 74212281f797SBarry Song snprintf(name, sizeof(name), "hugetlb%d", nid); 7422a01f4390SMike Kravetz /* 7423a01f4390SMike Kravetz * Note that 'order per bit' is based on smallest size that 7424a01f4390SMike Kravetz * may be returned to CMA allocator in the case of 7425a01f4390SMike Kravetz * huge page demotion. 7426a01f4390SMike Kravetz */ 7427a01f4390SMike Kravetz res = cma_declare_contiguous_nid(0, size, 0, 7428a01f4390SMike Kravetz PAGE_SIZE << HUGETLB_PAGE_ORDER, 742929d0f41dSBarry Song 0, false, name, 7430cf11e85fSRoman Gushchin &hugetlb_cma[nid], nid); 7431cf11e85fSRoman Gushchin if (res) { 7432cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: reservation failed: err %d, node %d", 7433cf11e85fSRoman Gushchin res, nid); 7434cf11e85fSRoman Gushchin continue; 7435cf11e85fSRoman Gushchin } 7436cf11e85fSRoman Gushchin 7437cf11e85fSRoman Gushchin reserved += size; 7438cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", 7439cf11e85fSRoman Gushchin size / SZ_1M, nid); 7440cf11e85fSRoman Gushchin 7441cf11e85fSRoman Gushchin if (reserved >= hugetlb_cma_size) 7442cf11e85fSRoman Gushchin break; 7443cf11e85fSRoman Gushchin } 7444a01f4390SMike Kravetz 7445a01f4390SMike Kravetz if (!reserved) 7446a01f4390SMike Kravetz /* 7447a01f4390SMike Kravetz * hugetlb_cma_size is used to determine if allocations from 7448a01f4390SMike Kravetz * cma are possible. Set to zero if no cma regions are set up. 7449a01f4390SMike Kravetz */ 7450a01f4390SMike Kravetz hugetlb_cma_size = 0; 7451cf11e85fSRoman Gushchin } 7452cf11e85fSRoman Gushchin 7453263b8998SMiaohe Lin static void __init hugetlb_cma_check(void) 7454cf11e85fSRoman Gushchin { 7455cf11e85fSRoman Gushchin if (!hugetlb_cma_size || cma_reserve_called) 7456cf11e85fSRoman Gushchin return; 7457cf11e85fSRoman Gushchin 7458cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); 7459cf11e85fSRoman Gushchin } 7460cf11e85fSRoman Gushchin 7461cf11e85fSRoman Gushchin #endif /* CONFIG_CMA */ 7462