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 97184afd99bSAndy Whitcroft /* 97208fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 97308fba699SMel Gorman * cases this will be same size as used by the page table entries. 97408fba699SMel Gorman */ 97508fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 97608fba699SMel Gorman { 97705ea8860SDan Williams if (vma->vm_ops && vma->vm_ops->pagesize) 97805ea8860SDan Williams return vma->vm_ops->pagesize(vma); 97908fba699SMel Gorman return PAGE_SIZE; 98008fba699SMel Gorman } 981f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 98208fba699SMel Gorman 98308fba699SMel Gorman /* 9843340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 9853340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 98609135cc5SDan Williams * architectures where it differs, an architecture-specific 'strong' 98709135cc5SDan Williams * version of this symbol is required. 9883340289dSMel Gorman */ 98909135cc5SDan Williams __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 9903340289dSMel Gorman { 9913340289dSMel Gorman return vma_kernel_pagesize(vma); 9923340289dSMel Gorman } 9933340289dSMel Gorman 9943340289dSMel Gorman /* 99584afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 99684afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 99784afd99bSAndy Whitcroft * alignment. 99884afd99bSAndy Whitcroft */ 99984afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 100084afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 100104f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 100284afd99bSAndy Whitcroft 1003a1e78772SMel Gorman /* 1004a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 1005a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 1006a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 1007a1e78772SMel Gorman * 10088d9bfb26SMike Kravetz * With the exception of hugetlb_dup_vma_private() which is called at fork(), 1009a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 1010a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 1011a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 101284afd99bSAndy Whitcroft * 101384afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 101484afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 101584afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 101684afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 101784afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 101884afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 101984afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 102084afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 1021a1e78772SMel Gorman */ 1022e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 1023e7c4b0bfSAndy Whitcroft { 1024e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 1025e7c4b0bfSAndy Whitcroft } 1026e7c4b0bfSAndy Whitcroft 1027e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 1028e7c4b0bfSAndy Whitcroft unsigned long value) 1029e7c4b0bfSAndy Whitcroft { 1030e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 1031e7c4b0bfSAndy Whitcroft } 1032e7c4b0bfSAndy Whitcroft 1033e9fe92aeSMina Almasry static void 1034e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, 1035e9fe92aeSMina Almasry struct hugetlb_cgroup *h_cg, 1036e9fe92aeSMina Almasry struct hstate *h) 1037e9fe92aeSMina Almasry { 1038e9fe92aeSMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 1039e9fe92aeSMina Almasry if (!h_cg || !h) { 1040e9fe92aeSMina Almasry resv_map->reservation_counter = NULL; 1041e9fe92aeSMina Almasry resv_map->pages_per_hpage = 0; 1042e9fe92aeSMina Almasry resv_map->css = NULL; 1043e9fe92aeSMina Almasry } else { 1044e9fe92aeSMina Almasry resv_map->reservation_counter = 1045e9fe92aeSMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 1046e9fe92aeSMina Almasry resv_map->pages_per_hpage = pages_per_huge_page(h); 1047e9fe92aeSMina Almasry resv_map->css = &h_cg->css; 1048e9fe92aeSMina Almasry } 1049e9fe92aeSMina Almasry #endif 1050e9fe92aeSMina Almasry } 1051e9fe92aeSMina Almasry 10529119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 105384afd99bSAndy Whitcroft { 105484afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 10555e911373SMike Kravetz struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 10565e911373SMike Kravetz 10575e911373SMike Kravetz if (!resv_map || !rg) { 10585e911373SMike Kravetz kfree(resv_map); 10595e911373SMike Kravetz kfree(rg); 106084afd99bSAndy Whitcroft return NULL; 10615e911373SMike Kravetz } 106284afd99bSAndy Whitcroft 106384afd99bSAndy Whitcroft kref_init(&resv_map->refs); 10647b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 106584afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 106684afd99bSAndy Whitcroft 10675e911373SMike Kravetz resv_map->adds_in_progress = 0; 1068e9fe92aeSMina Almasry /* 1069e9fe92aeSMina Almasry * Initialize these to 0. On shared mappings, 0's here indicate these 1070e9fe92aeSMina Almasry * fields don't do cgroup accounting. On private mappings, these will be 1071e9fe92aeSMina Almasry * re-initialized to the proper values, to indicate that hugetlb cgroup 1072e9fe92aeSMina Almasry * reservations are to be un-charged from here. 1073e9fe92aeSMina Almasry */ 1074e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); 10755e911373SMike Kravetz 10765e911373SMike Kravetz INIT_LIST_HEAD(&resv_map->region_cache); 10775e911373SMike Kravetz list_add(&rg->link, &resv_map->region_cache); 10785e911373SMike Kravetz resv_map->region_cache_count = 1; 10795e911373SMike Kravetz 108084afd99bSAndy Whitcroft return resv_map; 108184afd99bSAndy Whitcroft } 108284afd99bSAndy Whitcroft 10839119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 108484afd99bSAndy Whitcroft { 108584afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 10865e911373SMike Kravetz struct list_head *head = &resv_map->region_cache; 10875e911373SMike Kravetz struct file_region *rg, *trg; 108884afd99bSAndy Whitcroft 108984afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 1090feba16e2SMike Kravetz region_del(resv_map, 0, LONG_MAX); 10915e911373SMike Kravetz 10925e911373SMike Kravetz /* ... and any entries left in the cache */ 10935e911373SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 10945e911373SMike Kravetz list_del(&rg->link); 10955e911373SMike Kravetz kfree(rg); 10965e911373SMike Kravetz } 10975e911373SMike Kravetz 10985e911373SMike Kravetz VM_BUG_ON(resv_map->adds_in_progress); 10995e911373SMike Kravetz 110084afd99bSAndy Whitcroft kfree(resv_map); 110184afd99bSAndy Whitcroft } 110284afd99bSAndy Whitcroft 11034e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 11044e35f483SJoonsoo Kim { 1105f27a5136SMike Kravetz /* 1106f27a5136SMike Kravetz * At inode evict time, i_mapping may not point to the original 1107f27a5136SMike Kravetz * address space within the inode. This original address space 1108f27a5136SMike Kravetz * contains the pointer to the resv_map. So, always use the 1109f27a5136SMike Kravetz * address space embedded within the inode. 1110f27a5136SMike Kravetz * The VERY common case is inode->mapping == &inode->i_data but, 1111f27a5136SMike Kravetz * this may not be true for device special inodes. 1112f27a5136SMike Kravetz */ 1113f27a5136SMike Kravetz return (struct resv_map *)(&inode->i_data)->private_data; 11144e35f483SJoonsoo Kim } 11154e35f483SJoonsoo Kim 111684afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 1117a1e78772SMel Gorman { 111881d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 11194e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 11204e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 11214e35f483SJoonsoo Kim struct inode *inode = mapping->host; 11224e35f483SJoonsoo Kim 11234e35f483SJoonsoo Kim return inode_resv_map(inode); 11244e35f483SJoonsoo Kim 11254e35f483SJoonsoo Kim } else { 112684afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 112784afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 11284e35f483SJoonsoo Kim } 1129a1e78772SMel Gorman } 1130a1e78772SMel Gorman 113184afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 1132a1e78772SMel Gorman { 113381d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 113481d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 1135a1e78772SMel Gorman 113684afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 113784afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 113804f2cbe3SMel Gorman } 113904f2cbe3SMel Gorman 114004f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 114104f2cbe3SMel Gorman { 114281d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 114381d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 1144e7c4b0bfSAndy Whitcroft 1145e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 114604f2cbe3SMel Gorman } 114704f2cbe3SMel Gorman 114804f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 114904f2cbe3SMel Gorman { 115081d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 1151e7c4b0bfSAndy Whitcroft 1152e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 1153a1e78772SMel Gorman } 1154a1e78772SMel Gorman 11558d9bfb26SMike Kravetz void hugetlb_dup_vma_private(struct vm_area_struct *vma) 1156a1e78772SMel Gorman { 115781d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 11588d9bfb26SMike Kravetz /* 11598d9bfb26SMike Kravetz * Clear vm_private_data 1160612b8a31SMike Kravetz * - For shared mappings this is a per-vma semaphore that may be 1161612b8a31SMike Kravetz * allocated in a subsequent call to hugetlb_vm_op_open. 1162612b8a31SMike Kravetz * Before clearing, make sure pointer is not associated with vma 1163612b8a31SMike Kravetz * as this will leak the structure. This is the case when called 1164612b8a31SMike Kravetz * via clear_vma_resv_huge_pages() and hugetlb_vm_op_open has already 1165612b8a31SMike Kravetz * been called to allocate a new structure. 11668d9bfb26SMike Kravetz * - For MAP_PRIVATE mappings, this is the reserve map which does 11678d9bfb26SMike Kravetz * not apply to children. Faults generated by the children are 11688d9bfb26SMike Kravetz * not guaranteed to succeed, even if read-only. 11698d9bfb26SMike Kravetz */ 1170612b8a31SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 1171612b8a31SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 1172612b8a31SMike Kravetz 1173612b8a31SMike Kravetz if (vma_lock && vma_lock->vma != vma) 1174612b8a31SMike Kravetz vma->vm_private_data = NULL; 1175612b8a31SMike Kravetz } else 1176612b8a31SMike Kravetz vma->vm_private_data = NULL; 1177a1e78772SMel Gorman } 1178a1e78772SMel Gorman 1179550a7d60SMina Almasry /* 1180550a7d60SMina Almasry * Reset and decrement one ref on hugepage private reservation. 11818651a137SLorenzo Stoakes * Called with mm->mmap_lock writer semaphore held. 1182550a7d60SMina Almasry * This function should be only used by move_vma() and operate on 1183550a7d60SMina Almasry * same sized vma. It should never come here with last ref on the 1184550a7d60SMina Almasry * reservation. 1185550a7d60SMina Almasry */ 1186550a7d60SMina Almasry void clear_vma_resv_huge_pages(struct vm_area_struct *vma) 1187550a7d60SMina Almasry { 1188550a7d60SMina Almasry /* 1189550a7d60SMina Almasry * Clear the old hugetlb private page reservation. 1190550a7d60SMina Almasry * It has already been transferred to new_vma. 1191550a7d60SMina Almasry * 1192550a7d60SMina Almasry * During a mremap() operation of a hugetlb vma we call move_vma() 1193550a7d60SMina Almasry * which copies vma into new_vma and unmaps vma. After the copy 1194550a7d60SMina Almasry * operation both new_vma and vma share a reference to the resv_map 1195550a7d60SMina Almasry * struct, and at that point vma is about to be unmapped. We don't 1196550a7d60SMina Almasry * want to return the reservation to the pool at unmap of vma because 1197550a7d60SMina Almasry * the reservation still lives on in new_vma, so simply decrement the 1198550a7d60SMina Almasry * ref here and remove the resv_map reference from this vma. 1199550a7d60SMina Almasry */ 1200550a7d60SMina Almasry struct resv_map *reservations = vma_resv_map(vma); 1201550a7d60SMina Almasry 1202afe041c2SBui Quang Minh if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 1203afe041c2SBui Quang Minh resv_map_put_hugetlb_cgroup_uncharge_info(reservations); 1204550a7d60SMina Almasry kref_put(&reservations->refs, resv_map_release); 1205afe041c2SBui Quang Minh } 1206550a7d60SMina Almasry 12078d9bfb26SMike Kravetz hugetlb_dup_vma_private(vma); 1208550a7d60SMina Almasry } 1209550a7d60SMina Almasry 1210a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 1211559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 1212a1e78772SMel Gorman { 1213af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 1214af0ed73eSJoonsoo Kim /* 1215af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 1216af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 1217af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 1218af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 1219af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 1220af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 1221af0ed73eSJoonsoo Kim * properly, so add work-around here. 1222af0ed73eSJoonsoo Kim */ 1223af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 1224559ec2f8SNicholas Krause return true; 1225af0ed73eSJoonsoo Kim else 1226559ec2f8SNicholas Krause return false; 1227af0ed73eSJoonsoo Kim } 1228a63884e9SJoonsoo Kim 1229a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 12301fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 12311fb1b0e9SMike Kravetz /* 12321fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 12331fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 12341fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 12357c8de358SEthon Paul * fallocate. In this case, there really are no reserves to 12361fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 12371fb1b0e9SMike Kravetz */ 12381fb1b0e9SMike Kravetz if (chg) 12391fb1b0e9SMike Kravetz return false; 12401fb1b0e9SMike Kravetz else 1241559ec2f8SNicholas Krause return true; 12421fb1b0e9SMike Kravetz } 1243a63884e9SJoonsoo Kim 1244a63884e9SJoonsoo Kim /* 1245a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 1246a63884e9SJoonsoo Kim * private mappings. 1247a63884e9SJoonsoo Kim */ 124867961f9dSMike Kravetz if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 124967961f9dSMike Kravetz /* 125067961f9dSMike Kravetz * Like the shared case above, a hole punch or truncate 125167961f9dSMike Kravetz * could have been performed on the private mapping. 125267961f9dSMike Kravetz * Examine the value of chg to determine if reserves 125367961f9dSMike Kravetz * actually exist or were previously consumed. 125467961f9dSMike Kravetz * Very Subtle - The value of chg comes from a previous 125567961f9dSMike Kravetz * call to vma_needs_reserves(). The reserve map for 125667961f9dSMike Kravetz * private mappings has different (opposite) semantics 125767961f9dSMike Kravetz * than that of shared mappings. vma_needs_reserves() 125867961f9dSMike Kravetz * has already taken this difference in semantics into 125967961f9dSMike Kravetz * account. Therefore, the meaning of chg is the same 126067961f9dSMike Kravetz * as in the shared case above. Code could easily be 126167961f9dSMike Kravetz * combined, but keeping it separate draws attention to 126267961f9dSMike Kravetz * subtle differences. 126367961f9dSMike Kravetz */ 126467961f9dSMike Kravetz if (chg) 126567961f9dSMike Kravetz return false; 126667961f9dSMike Kravetz else 1267559ec2f8SNicholas Krause return true; 126867961f9dSMike Kravetz } 1269a63884e9SJoonsoo Kim 1270559ec2f8SNicholas Krause return false; 1271a1e78772SMel Gorman } 1272a1e78772SMel Gorman 1273240d67a8SSidhartha Kumar static void enqueue_hugetlb_folio(struct hstate *h, struct folio *folio) 12741da177e4SLinus Torvalds { 1275240d67a8SSidhartha Kumar int nid = folio_nid(folio); 12769487ca60SMike Kravetz 12779487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1278240d67a8SSidhartha Kumar VM_BUG_ON_FOLIO(folio_ref_count(folio), folio); 1279b65a4edaSMike Kravetz 1280240d67a8SSidhartha Kumar list_move(&folio->lru, &h->hugepage_freelists[nid]); 1281a5516438SAndi Kleen h->free_huge_pages++; 1282a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 1283240d67a8SSidhartha Kumar folio_set_hugetlb_freed(folio); 12841da177e4SLinus Torvalds } 12851da177e4SLinus Torvalds 1286a36f1e90SSidhartha Kumar static struct folio *dequeue_hugetlb_folio_node_exact(struct hstate *h, 1287a36f1e90SSidhartha Kumar int nid) 1288bf50bab2SNaoya Horiguchi { 1289a36f1e90SSidhartha Kumar struct folio *folio; 12901a08ae36SPavel Tatashin bool pin = !!(current->flags & PF_MEMALLOC_PIN); 1291bf50bab2SNaoya Horiguchi 12929487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1293a36f1e90SSidhartha Kumar list_for_each_entry(folio, &h->hugepage_freelists[nid], lru) { 1294a36f1e90SSidhartha Kumar if (pin && !folio_is_longterm_pinnable(folio)) 1295bbe88753SJoonsoo Kim continue; 1296bbe88753SJoonsoo Kim 1297a36f1e90SSidhartha Kumar if (folio_test_hwpoison(folio)) 12986664bfc8SWei Yang continue; 1299bbe88753SJoonsoo Kim 1300a36f1e90SSidhartha Kumar list_move(&folio->lru, &h->hugepage_activelist); 1301a36f1e90SSidhartha Kumar folio_ref_unfreeze(folio, 1); 1302a36f1e90SSidhartha Kumar folio_clear_hugetlb_freed(folio); 1303bf50bab2SNaoya Horiguchi h->free_huge_pages--; 1304bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 1305a36f1e90SSidhartha Kumar return folio; 1306bf50bab2SNaoya Horiguchi } 1307bf50bab2SNaoya Horiguchi 13086664bfc8SWei Yang return NULL; 13096664bfc8SWei Yang } 13106664bfc8SWei Yang 1311a36f1e90SSidhartha Kumar static struct folio *dequeue_hugetlb_folio_nodemask(struct hstate *h, gfp_t gfp_mask, 1312a36f1e90SSidhartha Kumar int nid, nodemask_t *nmask) 131394310cbcSAnshuman Khandual { 13143e59fcb0SMichal Hocko unsigned int cpuset_mems_cookie; 13153e59fcb0SMichal Hocko struct zonelist *zonelist; 13163e59fcb0SMichal Hocko struct zone *zone; 13173e59fcb0SMichal Hocko struct zoneref *z; 131898fa15f3SAnshuman Khandual int node = NUMA_NO_NODE; 13193e59fcb0SMichal Hocko 13203e59fcb0SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 13213e59fcb0SMichal Hocko 13223e59fcb0SMichal Hocko retry_cpuset: 13233e59fcb0SMichal Hocko cpuset_mems_cookie = read_mems_allowed_begin(); 13243e59fcb0SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 1325a36f1e90SSidhartha Kumar struct folio *folio; 132694310cbcSAnshuman Khandual 13273e59fcb0SMichal Hocko if (!cpuset_zone_allowed(zone, gfp_mask)) 13283e59fcb0SMichal Hocko continue; 13293e59fcb0SMichal Hocko /* 13303e59fcb0SMichal Hocko * no need to ask again on the same node. Pool is node rather than 13313e59fcb0SMichal Hocko * zone aware 13323e59fcb0SMichal Hocko */ 13333e59fcb0SMichal Hocko if (zone_to_nid(zone) == node) 13343e59fcb0SMichal Hocko continue; 13353e59fcb0SMichal Hocko node = zone_to_nid(zone); 133694310cbcSAnshuman Khandual 1337a36f1e90SSidhartha Kumar folio = dequeue_hugetlb_folio_node_exact(h, node); 1338a36f1e90SSidhartha Kumar if (folio) 1339a36f1e90SSidhartha Kumar return folio; 134094310cbcSAnshuman Khandual } 13413e59fcb0SMichal Hocko if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 13423e59fcb0SMichal Hocko goto retry_cpuset; 13433e59fcb0SMichal Hocko 134494310cbcSAnshuman Khandual return NULL; 134594310cbcSAnshuman Khandual } 134694310cbcSAnshuman Khandual 13478346d69dSXin Hao static unsigned long available_huge_pages(struct hstate *h) 13488346d69dSXin Hao { 13498346d69dSXin Hao return h->free_huge_pages - h->resv_huge_pages; 13508346d69dSXin Hao } 13518346d69dSXin Hao 1352ff7d853bSSidhartha Kumar static struct folio *dequeue_hugetlb_folio_vma(struct hstate *h, 1353a5516438SAndi Kleen struct vm_area_struct *vma, 1354af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 1355af0ed73eSJoonsoo Kim long chg) 13561da177e4SLinus Torvalds { 1357a36f1e90SSidhartha Kumar struct folio *folio = NULL; 1358480eccf9SLee Schermerhorn struct mempolicy *mpol; 135904ec6264SVlastimil Babka gfp_t gfp_mask; 13603e59fcb0SMichal Hocko nodemask_t *nodemask; 136104ec6264SVlastimil Babka int nid; 13621da177e4SLinus Torvalds 1363a1e78772SMel Gorman /* 1364a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 1365a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 1366a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 1367a1e78772SMel Gorman */ 13688346d69dSXin Hao if (!vma_has_reserves(vma, chg) && !available_huge_pages(h)) 1369c0ff7453SMiao Xie goto err; 1370a1e78772SMel Gorman 137104f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 13728346d69dSXin Hao if (avoid_reserve && !available_huge_pages(h)) 13736eab04a8SJustin P. Mattock goto err; 137404f2cbe3SMel Gorman 137504ec6264SVlastimil Babka gfp_mask = htlb_alloc_mask(h); 137604ec6264SVlastimil Babka nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 1377cfcaa66fSBen Widawsky 1378cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 1379a36f1e90SSidhartha Kumar folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask, 1380a36f1e90SSidhartha Kumar nid, nodemask); 1381cfcaa66fSBen Widawsky 1382cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 1383cfcaa66fSBen Widawsky nodemask = NULL; 1384cfcaa66fSBen Widawsky } 1385cfcaa66fSBen Widawsky 1386a36f1e90SSidhartha Kumar if (!folio) 1387a36f1e90SSidhartha Kumar folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask, 1388a36f1e90SSidhartha Kumar nid, nodemask); 1389cfcaa66fSBen Widawsky 1390a36f1e90SSidhartha Kumar if (folio && !avoid_reserve && vma_has_reserves(vma, chg)) { 1391a36f1e90SSidhartha Kumar folio_set_hugetlb_restore_reserve(folio); 1392a63884e9SJoonsoo Kim h->resv_huge_pages--; 1393bf50bab2SNaoya Horiguchi } 1394cc9a6c87SMel Gorman 1395cc9a6c87SMel Gorman mpol_cond_put(mpol); 1396ff7d853bSSidhartha Kumar return folio; 1397cc9a6c87SMel Gorman 1398c0ff7453SMiao Xie err: 1399cc9a6c87SMel Gorman return NULL; 14001da177e4SLinus Torvalds } 14011da177e4SLinus Torvalds 14021cac6f2cSLuiz Capitulino /* 14031cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 14041cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 14051cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 14061cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 14071cac6f2cSLuiz Capitulino * node for alloc or free. 14081cac6f2cSLuiz Capitulino */ 14091cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 14101cac6f2cSLuiz Capitulino { 14110edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 14121cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 14131cac6f2cSLuiz Capitulino 14141cac6f2cSLuiz Capitulino return nid; 14151cac6f2cSLuiz Capitulino } 14161cac6f2cSLuiz Capitulino 14171cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 14181cac6f2cSLuiz Capitulino { 14191cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 14201cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 14211cac6f2cSLuiz Capitulino return nid; 14221cac6f2cSLuiz Capitulino } 14231cac6f2cSLuiz Capitulino 14241cac6f2cSLuiz Capitulino /* 14251cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 14261cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 14271cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 14281cac6f2cSLuiz Capitulino * mask. 14291cac6f2cSLuiz Capitulino */ 14301cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 14311cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 14321cac6f2cSLuiz Capitulino { 14331cac6f2cSLuiz Capitulino int nid; 14341cac6f2cSLuiz Capitulino 14351cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 14361cac6f2cSLuiz Capitulino 14371cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 14381cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 14391cac6f2cSLuiz Capitulino 14401cac6f2cSLuiz Capitulino return nid; 14411cac6f2cSLuiz Capitulino } 14421cac6f2cSLuiz Capitulino 14431cac6f2cSLuiz Capitulino /* 144410c6ec49SMike Kravetz * helper for remove_pool_huge_page() - return the previously saved 14451cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 14461cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 14471cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 14481cac6f2cSLuiz Capitulino */ 14491cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 14501cac6f2cSLuiz Capitulino { 14511cac6f2cSLuiz Capitulino int nid; 14521cac6f2cSLuiz Capitulino 14531cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 14541cac6f2cSLuiz Capitulino 14551cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 14561cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 14571cac6f2cSLuiz Capitulino 14581cac6f2cSLuiz Capitulino return nid; 14591cac6f2cSLuiz Capitulino } 14601cac6f2cSLuiz Capitulino 14611cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 14621cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 14631cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 14641cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 14651cac6f2cSLuiz Capitulino nr_nodes--) 14661cac6f2cSLuiz Capitulino 14671cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 14681cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 14691cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 14701cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 14711cac6f2cSLuiz Capitulino nr_nodes--) 14721cac6f2cSLuiz Capitulino 14738531fc6fSMike Kravetz /* used to demote non-gigantic_huge pages as well */ 1474911565b8SSidhartha Kumar static void __destroy_compound_gigantic_folio(struct folio *folio, 147534d9e35bSMike Kravetz unsigned int order, bool demote) 1476944d9fecSLuiz Capitulino { 1477944d9fecSLuiz Capitulino int i; 1478944d9fecSLuiz Capitulino int nr_pages = 1 << order; 147914455eabSCheng Li struct page *p; 1480944d9fecSLuiz Capitulino 148146f27228SMatthew Wilcox (Oracle) atomic_set(&folio->_entire_mapcount, 0); 1482eec20426SMatthew Wilcox (Oracle) atomic_set(&folio->_nr_pages_mapped, 0); 148394688e8eSMatthew Wilcox (Oracle) atomic_set(&folio->_pincount, 0); 148447e29d32SJohn Hubbard 148514455eabSCheng Li for (i = 1; i < nr_pages; i++) { 1486911565b8SSidhartha Kumar p = folio_page(folio, i); 1487a01f4390SMike Kravetz p->mapping = NULL; 14881d798ca3SKirill A. Shutemov clear_compound_head(p); 148934d9e35bSMike Kravetz if (!demote) 1490944d9fecSLuiz Capitulino set_page_refcounted(p); 1491944d9fecSLuiz Capitulino } 1492944d9fecSLuiz Capitulino 1493911565b8SSidhartha Kumar __folio_clear_head(folio); 1494944d9fecSLuiz Capitulino } 1495944d9fecSLuiz Capitulino 1496911565b8SSidhartha Kumar static void destroy_compound_hugetlb_folio_for_demote(struct folio *folio, 14978531fc6fSMike Kravetz unsigned int order) 14988531fc6fSMike Kravetz { 1499911565b8SSidhartha Kumar __destroy_compound_gigantic_folio(folio, order, true); 15008531fc6fSMike Kravetz } 15018531fc6fSMike Kravetz 15028531fc6fSMike Kravetz #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 1503911565b8SSidhartha Kumar static void destroy_compound_gigantic_folio(struct folio *folio, 150434d9e35bSMike Kravetz unsigned int order) 150534d9e35bSMike Kravetz { 1506911565b8SSidhartha Kumar __destroy_compound_gigantic_folio(folio, order, false); 150734d9e35bSMike Kravetz } 150834d9e35bSMike Kravetz 15097f325a8dSSidhartha Kumar static void free_gigantic_folio(struct folio *folio, unsigned int order) 1510944d9fecSLuiz Capitulino { 1511cf11e85fSRoman Gushchin /* 1512cf11e85fSRoman Gushchin * If the page isn't allocated using the cma allocator, 1513cf11e85fSRoman Gushchin * cma_release() returns false. 1514cf11e85fSRoman Gushchin */ 1515dbda8feaSBarry Song #ifdef CONFIG_CMA 15167f325a8dSSidhartha Kumar int nid = folio_nid(folio); 15177f325a8dSSidhartha Kumar 15187f325a8dSSidhartha Kumar if (cma_release(hugetlb_cma[nid], &folio->page, 1 << order)) 1519cf11e85fSRoman Gushchin return; 1520dbda8feaSBarry Song #endif 1521cf11e85fSRoman Gushchin 15227f325a8dSSidhartha Kumar free_contig_range(folio_pfn(folio), 1 << order); 1523944d9fecSLuiz Capitulino } 1524944d9fecSLuiz Capitulino 15254eb0716eSAlexandre Ghiti #ifdef CONFIG_CONTIG_ALLOC 152619fc1a7eSSidhartha Kumar static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask, 1527d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) 1528944d9fecSLuiz Capitulino { 152919fc1a7eSSidhartha Kumar struct page *page; 153004adbc3fSMiaohe Lin unsigned long nr_pages = pages_per_huge_page(h); 1531953f064aSLi Xinhai if (nid == NUMA_NO_NODE) 1532953f064aSLi Xinhai nid = numa_mem_id(); 1533944d9fecSLuiz Capitulino 1534dbda8feaSBarry Song #ifdef CONFIG_CMA 1535dbda8feaSBarry Song { 1536cf11e85fSRoman Gushchin int node; 1537cf11e85fSRoman Gushchin 1538953f064aSLi Xinhai if (hugetlb_cma[nid]) { 1539953f064aSLi Xinhai page = cma_alloc(hugetlb_cma[nid], nr_pages, 1540953f064aSLi Xinhai huge_page_order(h), true); 1541953f064aSLi Xinhai if (page) 154219fc1a7eSSidhartha Kumar return page_folio(page); 1543953f064aSLi Xinhai } 1544953f064aSLi Xinhai 1545953f064aSLi Xinhai if (!(gfp_mask & __GFP_THISNODE)) { 1546cf11e85fSRoman Gushchin for_each_node_mask(node, *nodemask) { 1547953f064aSLi Xinhai if (node == nid || !hugetlb_cma[node]) 1548cf11e85fSRoman Gushchin continue; 1549cf11e85fSRoman Gushchin 1550cf11e85fSRoman Gushchin page = cma_alloc(hugetlb_cma[node], nr_pages, 1551cf11e85fSRoman Gushchin huge_page_order(h), true); 1552cf11e85fSRoman Gushchin if (page) 155319fc1a7eSSidhartha Kumar return page_folio(page); 1554cf11e85fSRoman Gushchin } 1555cf11e85fSRoman Gushchin } 1556953f064aSLi Xinhai } 1557dbda8feaSBarry Song #endif 1558cf11e85fSRoman Gushchin 155919fc1a7eSSidhartha Kumar page = alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); 156019fc1a7eSSidhartha Kumar return page ? page_folio(page) : NULL; 1561944d9fecSLuiz Capitulino } 1562944d9fecSLuiz Capitulino 15634eb0716eSAlexandre Ghiti #else /* !CONFIG_CONTIG_ALLOC */ 156419fc1a7eSSidhartha Kumar static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask, 15654eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 15664eb0716eSAlexandre Ghiti { 15674eb0716eSAlexandre Ghiti return NULL; 15684eb0716eSAlexandre Ghiti } 15694eb0716eSAlexandre Ghiti #endif /* CONFIG_CONTIG_ALLOC */ 1570944d9fecSLuiz Capitulino 1571e1073d1eSAneesh Kumar K.V #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 157219fc1a7eSSidhartha Kumar static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask, 15734eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 15744eb0716eSAlexandre Ghiti { 15754eb0716eSAlexandre Ghiti return NULL; 15764eb0716eSAlexandre Ghiti } 15777f325a8dSSidhartha Kumar static inline void free_gigantic_folio(struct folio *folio, 15787f325a8dSSidhartha Kumar unsigned int order) { } 1579911565b8SSidhartha Kumar static inline void destroy_compound_gigantic_folio(struct folio *folio, 1580d00181b9SKirill A. Shutemov unsigned int order) { } 1581944d9fecSLuiz Capitulino #endif 1582944d9fecSLuiz Capitulino 15836eb4e88aSMike Kravetz /* 1584cfd5082bSSidhartha Kumar * Remove hugetlb folio from lists, and update dtor so that the folio appears 158534d9e35bSMike Kravetz * as just a compound page. 158634d9e35bSMike Kravetz * 1587cfd5082bSSidhartha Kumar * A reference is held on the folio, except in the case of demote. 15886eb4e88aSMike Kravetz * 15896eb4e88aSMike Kravetz * Must be called with hugetlb lock held. 15906eb4e88aSMike Kravetz */ 1591cfd5082bSSidhartha Kumar static void __remove_hugetlb_folio(struct hstate *h, struct folio *folio, 159234d9e35bSMike Kravetz bool adjust_surplus, 159334d9e35bSMike Kravetz bool demote) 15946eb4e88aSMike Kravetz { 1595cfd5082bSSidhartha Kumar int nid = folio_nid(folio); 15966eb4e88aSMike Kravetz 1597f074732dSSidhartha Kumar VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio(folio), folio); 1598f074732dSSidhartha Kumar VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio_rsvd(folio), folio); 15996eb4e88aSMike Kravetz 16009487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 16016eb4e88aSMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 16026eb4e88aSMike Kravetz return; 16036eb4e88aSMike Kravetz 1604cfd5082bSSidhartha Kumar list_del(&folio->lru); 16056eb4e88aSMike Kravetz 1606cfd5082bSSidhartha Kumar if (folio_test_hugetlb_freed(folio)) { 16076eb4e88aSMike Kravetz h->free_huge_pages--; 16086eb4e88aSMike Kravetz h->free_huge_pages_node[nid]--; 16096eb4e88aSMike Kravetz } 16106eb4e88aSMike Kravetz if (adjust_surplus) { 16116eb4e88aSMike Kravetz h->surplus_huge_pages--; 16126eb4e88aSMike Kravetz h->surplus_huge_pages_node[nid]--; 16136eb4e88aSMike Kravetz } 16146eb4e88aSMike Kravetz 1615e32d20c0SMike Kravetz /* 1616e32d20c0SMike Kravetz * Very subtle 1617e32d20c0SMike Kravetz * 1618e32d20c0SMike Kravetz * For non-gigantic pages set the destructor to the normal compound 1619e32d20c0SMike Kravetz * page dtor. This is needed in case someone takes an additional 1620e32d20c0SMike Kravetz * temporary ref to the page, and freeing is delayed until they drop 1621e32d20c0SMike Kravetz * their reference. 1622e32d20c0SMike Kravetz * 1623e32d20c0SMike Kravetz * For gigantic pages set the destructor to the null dtor. This 1624e32d20c0SMike Kravetz * destructor will never be called. Before freeing the gigantic 1625911565b8SSidhartha Kumar * page destroy_compound_gigantic_folio will turn the folio into a 1626911565b8SSidhartha Kumar * simple group of pages. After this the destructor does not 1627e32d20c0SMike Kravetz * apply. 1628e32d20c0SMike Kravetz * 1629e32d20c0SMike Kravetz * This handles the case where more than one ref is held when and 1630d6ef19e2SSidhartha Kumar * after update_and_free_hugetlb_folio is called. 163134d9e35bSMike Kravetz * 163234d9e35bSMike Kravetz * In the case of demote we do not ref count the page as it will soon 163334d9e35bSMike Kravetz * be turned into a page of smaller size. 1634e32d20c0SMike Kravetz */ 163534d9e35bSMike Kravetz if (!demote) 1636cfd5082bSSidhartha Kumar folio_ref_unfreeze(folio, 1); 1637e32d20c0SMike Kravetz if (hstate_is_gigantic(h)) 1638cfd5082bSSidhartha Kumar folio_set_compound_dtor(folio, NULL_COMPOUND_DTOR); 1639e32d20c0SMike Kravetz else 1640cfd5082bSSidhartha Kumar folio_set_compound_dtor(folio, COMPOUND_PAGE_DTOR); 16416eb4e88aSMike Kravetz 16426eb4e88aSMike Kravetz h->nr_huge_pages--; 16436eb4e88aSMike Kravetz h->nr_huge_pages_node[nid]--; 16446eb4e88aSMike Kravetz } 16456eb4e88aSMike Kravetz 1646cfd5082bSSidhartha Kumar static void remove_hugetlb_folio(struct hstate *h, struct folio *folio, 164734d9e35bSMike Kravetz bool adjust_surplus) 164834d9e35bSMike Kravetz { 1649cfd5082bSSidhartha Kumar __remove_hugetlb_folio(h, folio, adjust_surplus, false); 165034d9e35bSMike Kravetz } 165134d9e35bSMike Kravetz 1652cfd5082bSSidhartha Kumar static void remove_hugetlb_folio_for_demote(struct hstate *h, struct folio *folio, 16538531fc6fSMike Kravetz bool adjust_surplus) 16548531fc6fSMike Kravetz { 1655cfd5082bSSidhartha Kumar __remove_hugetlb_folio(h, folio, adjust_surplus, true); 16568531fc6fSMike Kravetz } 16578531fc6fSMike Kravetz 16582f6c57d6SSidhartha Kumar static void add_hugetlb_folio(struct hstate *h, struct folio *folio, 1659ad2fa371SMuchun Song bool adjust_surplus) 1660ad2fa371SMuchun Song { 1661ad2fa371SMuchun Song int zeroed; 16622f6c57d6SSidhartha Kumar int nid = folio_nid(folio); 1663ad2fa371SMuchun Song 16642f6c57d6SSidhartha Kumar VM_BUG_ON_FOLIO(!folio_test_hugetlb_vmemmap_optimized(folio), folio); 1665ad2fa371SMuchun Song 1666ad2fa371SMuchun Song lockdep_assert_held(&hugetlb_lock); 1667ad2fa371SMuchun Song 16682f6c57d6SSidhartha Kumar INIT_LIST_HEAD(&folio->lru); 1669ad2fa371SMuchun Song h->nr_huge_pages++; 1670ad2fa371SMuchun Song h->nr_huge_pages_node[nid]++; 1671ad2fa371SMuchun Song 1672ad2fa371SMuchun Song if (adjust_surplus) { 1673ad2fa371SMuchun Song h->surplus_huge_pages++; 1674ad2fa371SMuchun Song h->surplus_huge_pages_node[nid]++; 1675ad2fa371SMuchun Song } 1676ad2fa371SMuchun Song 16772f6c57d6SSidhartha Kumar folio_set_compound_dtor(folio, HUGETLB_PAGE_DTOR); 16782f6c57d6SSidhartha Kumar folio_change_private(folio, NULL); 1679a9e1eab2SMiaohe Lin /* 16802f6c57d6SSidhartha Kumar * We have to set hugetlb_vmemmap_optimized again as above 16812f6c57d6SSidhartha Kumar * folio_change_private(folio, NULL) cleared it. 1682a9e1eab2SMiaohe Lin */ 16832f6c57d6SSidhartha Kumar folio_set_hugetlb_vmemmap_optimized(folio); 1684ad2fa371SMuchun Song 1685ad2fa371SMuchun Song /* 16862f6c57d6SSidhartha Kumar * This folio is about to be managed by the hugetlb allocator and 1687b65a4edaSMike Kravetz * should have no users. Drop our reference, and check for others 1688b65a4edaSMike Kravetz * just in case. 1689ad2fa371SMuchun Song */ 16902f6c57d6SSidhartha Kumar zeroed = folio_put_testzero(folio); 16912f6c57d6SSidhartha Kumar if (unlikely(!zeroed)) 1692b65a4edaSMike Kravetz /* 1693b65a4edaSMike Kravetz * It is VERY unlikely soneone else has taken a ref on 1694b65a4edaSMike Kravetz * the page. In this case, we simply return as the 1695b65a4edaSMike Kravetz * hugetlb destructor (free_huge_page) will be called 1696b65a4edaSMike Kravetz * when this other ref is dropped. 1697b65a4edaSMike Kravetz */ 1698b65a4edaSMike Kravetz return; 1699b65a4edaSMike Kravetz 17002f6c57d6SSidhartha Kumar arch_clear_hugepage_flags(&folio->page); 1701240d67a8SSidhartha Kumar enqueue_hugetlb_folio(h, folio); 1702ad2fa371SMuchun Song } 1703ad2fa371SMuchun Song 17046f6956cfSSidhartha Kumar static void __update_and_free_hugetlb_folio(struct hstate *h, 17056f6956cfSSidhartha Kumar struct folio *folio) 17066af2acb6SAdam Litke { 17076af2acb6SAdam Litke int i; 170814455eabSCheng Li struct page *subpage; 1709a5516438SAndi Kleen 17104eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1711944d9fecSLuiz Capitulino return; 171218229df5SAndy Whitcroft 1713161df60eSNaoya Horiguchi /* 1714161df60eSNaoya Horiguchi * If we don't know which subpages are hwpoisoned, we can't free 1715161df60eSNaoya Horiguchi * the hugepage, so it's leaked intentionally. 1716161df60eSNaoya Horiguchi */ 17177f325a8dSSidhartha Kumar if (folio_test_hugetlb_raw_hwp_unreliable(folio)) 1718161df60eSNaoya Horiguchi return; 1719161df60eSNaoya Horiguchi 17206f6956cfSSidhartha Kumar if (hugetlb_vmemmap_restore(h, &folio->page)) { 1721ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 1722ad2fa371SMuchun Song /* 1723ad2fa371SMuchun Song * If we cannot allocate vmemmap pages, just refuse to free the 1724ad2fa371SMuchun Song * page and put the page back on the hugetlb free list and treat 1725ad2fa371SMuchun Song * as a surplus page. 1726ad2fa371SMuchun Song */ 17277f325a8dSSidhartha Kumar add_hugetlb_folio(h, folio, true); 1728ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 1729ad2fa371SMuchun Song return; 1730ad2fa371SMuchun Song } 1731ad2fa371SMuchun Song 1732161df60eSNaoya Horiguchi /* 1733161df60eSNaoya Horiguchi * Move PageHWPoison flag from head page to the raw error pages, 1734161df60eSNaoya Horiguchi * which makes any healthy subpages reusable. 1735161df60eSNaoya Horiguchi */ 1736911565b8SSidhartha Kumar if (unlikely(folio_test_hwpoison(folio))) 17372ff6ceceSSidhartha Kumar folio_clear_hugetlb_hwpoison(folio); 1738161df60eSNaoya Horiguchi 173914455eabSCheng Li for (i = 0; i < pages_per_huge_page(h); i++) { 17407f325a8dSSidhartha Kumar subpage = folio_page(folio, i); 1741dbfee5aeSMike Kravetz subpage->flags &= ~(1 << PG_locked | 1 << PG_error | 174232f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1743a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1744a7407a27SLuiz Capitulino 1 << PG_writeback); 17456af2acb6SAdam Litke } 1746a01f4390SMike Kravetz 1747a01f4390SMike Kravetz /* 1748a01f4390SMike Kravetz * Non-gigantic pages demoted from CMA allocated gigantic pages 17497f325a8dSSidhartha Kumar * need to be given back to CMA in free_gigantic_folio. 1750a01f4390SMike Kravetz */ 1751a01f4390SMike Kravetz if (hstate_is_gigantic(h) || 17522f6c57d6SSidhartha Kumar hugetlb_cma_folio(folio, huge_page_order(h))) { 1753911565b8SSidhartha Kumar destroy_compound_gigantic_folio(folio, huge_page_order(h)); 17547f325a8dSSidhartha Kumar free_gigantic_folio(folio, huge_page_order(h)); 1755944d9fecSLuiz Capitulino } else { 17566f6956cfSSidhartha Kumar __free_pages(&folio->page, huge_page_order(h)); 17576af2acb6SAdam Litke } 1758944d9fecSLuiz Capitulino } 17596af2acb6SAdam Litke 1760b65d4adbSMuchun Song /* 1761d6ef19e2SSidhartha Kumar * As update_and_free_hugetlb_folio() can be called under any context, so we cannot 1762b65d4adbSMuchun Song * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the 1763b65d4adbSMuchun Song * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate 1764b65d4adbSMuchun Song * the vmemmap pages. 1765b65d4adbSMuchun Song * 1766b65d4adbSMuchun Song * free_hpage_workfn() locklessly retrieves the linked list of pages to be 1767b65d4adbSMuchun Song * freed and frees them one-by-one. As the page->mapping pointer is going 1768b65d4adbSMuchun Song * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node 1769b65d4adbSMuchun Song * structure of a lockless linked list of huge pages to be freed. 1770b65d4adbSMuchun Song */ 1771b65d4adbSMuchun Song static LLIST_HEAD(hpage_freelist); 1772b65d4adbSMuchun Song 1773b65d4adbSMuchun Song static void free_hpage_workfn(struct work_struct *work) 1774b65d4adbSMuchun Song { 1775b65d4adbSMuchun Song struct llist_node *node; 1776b65d4adbSMuchun Song 1777b65d4adbSMuchun Song node = llist_del_all(&hpage_freelist); 1778b65d4adbSMuchun Song 1779b65d4adbSMuchun Song while (node) { 1780b65d4adbSMuchun Song struct page *page; 1781b65d4adbSMuchun Song struct hstate *h; 1782b65d4adbSMuchun Song 1783b65d4adbSMuchun Song page = container_of((struct address_space **)node, 1784b65d4adbSMuchun Song struct page, mapping); 1785b65d4adbSMuchun Song node = node->next; 1786b65d4adbSMuchun Song page->mapping = NULL; 1787b65d4adbSMuchun Song /* 1788b65d4adbSMuchun Song * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate() 1789b65d4adbSMuchun Song * is going to trigger because a previous call to 1790cfd5082bSSidhartha Kumar * remove_hugetlb_folio() will call folio_set_compound_dtor 1791cfd5082bSSidhartha Kumar * (folio, NULL_COMPOUND_DTOR), so do not use page_hstate() 1792cfd5082bSSidhartha Kumar * directly. 1793b65d4adbSMuchun Song */ 1794b65d4adbSMuchun Song h = size_to_hstate(page_size(page)); 1795b65d4adbSMuchun Song 17966f6956cfSSidhartha Kumar __update_and_free_hugetlb_folio(h, page_folio(page)); 1797b65d4adbSMuchun Song 1798b65d4adbSMuchun Song cond_resched(); 1799b65d4adbSMuchun Song } 1800b65d4adbSMuchun Song } 1801b65d4adbSMuchun Song static DECLARE_WORK(free_hpage_work, free_hpage_workfn); 1802b65d4adbSMuchun Song 1803b65d4adbSMuchun Song static inline void flush_free_hpage_work(struct hstate *h) 1804b65d4adbSMuchun Song { 18056213834cSMuchun Song if (hugetlb_vmemmap_optimizable(h)) 1806b65d4adbSMuchun Song flush_work(&free_hpage_work); 1807b65d4adbSMuchun Song } 1808b65d4adbSMuchun Song 1809d6ef19e2SSidhartha Kumar static void update_and_free_hugetlb_folio(struct hstate *h, struct folio *folio, 1810b65d4adbSMuchun Song bool atomic) 1811b65d4adbSMuchun Song { 1812d6ef19e2SSidhartha Kumar if (!folio_test_hugetlb_vmemmap_optimized(folio) || !atomic) { 18136f6956cfSSidhartha Kumar __update_and_free_hugetlb_folio(h, folio); 1814b65d4adbSMuchun Song return; 1815b65d4adbSMuchun Song } 1816b65d4adbSMuchun Song 1817b65d4adbSMuchun Song /* 1818b65d4adbSMuchun Song * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages. 1819b65d4adbSMuchun Song * 1820b65d4adbSMuchun Song * Only call schedule_work() if hpage_freelist is previously 1821b65d4adbSMuchun Song * empty. Otherwise, schedule_work() had been called but the workfn 1822b65d4adbSMuchun Song * hasn't retrieved the list yet. 1823b65d4adbSMuchun Song */ 1824d6ef19e2SSidhartha Kumar if (llist_add((struct llist_node *)&folio->mapping, &hpage_freelist)) 1825b65d4adbSMuchun Song schedule_work(&free_hpage_work); 1826b65d4adbSMuchun Song } 1827b65d4adbSMuchun Song 182810c6ec49SMike Kravetz static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list) 182910c6ec49SMike Kravetz { 183010c6ec49SMike Kravetz struct page *page, *t_page; 1831d6ef19e2SSidhartha Kumar struct folio *folio; 183210c6ec49SMike Kravetz 183310c6ec49SMike Kravetz list_for_each_entry_safe(page, t_page, list, lru) { 1834d6ef19e2SSidhartha Kumar folio = page_folio(page); 1835d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, folio, false); 183610c6ec49SMike Kravetz cond_resched(); 183710c6ec49SMike Kravetz } 183810c6ec49SMike Kravetz } 183910c6ec49SMike Kravetz 1840e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1841e5ff2159SAndi Kleen { 1842e5ff2159SAndi Kleen struct hstate *h; 1843e5ff2159SAndi Kleen 1844e5ff2159SAndi Kleen for_each_hstate(h) { 1845e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1846e5ff2159SAndi Kleen return h; 1847e5ff2159SAndi Kleen } 1848e5ff2159SAndi Kleen return NULL; 1849e5ff2159SAndi Kleen } 1850e5ff2159SAndi Kleen 1851db71ef79SMike Kravetz void free_huge_page(struct page *page) 185227a85ef1SDavid Gibson { 1853a5516438SAndi Kleen /* 1854a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1855a5516438SAndi Kleen * compound page destructor. 1856a5516438SAndi Kleen */ 18570356c4b9SSidhartha Kumar struct folio *folio = page_folio(page); 18580356c4b9SSidhartha Kumar struct hstate *h = folio_hstate(folio); 18590356c4b9SSidhartha Kumar int nid = folio_nid(folio); 18600356c4b9SSidhartha Kumar struct hugepage_subpool *spool = hugetlb_folio_subpool(folio); 186107443a85SJoonsoo Kim bool restore_reserve; 1862db71ef79SMike Kravetz unsigned long flags; 186327a85ef1SDavid Gibson 18640356c4b9SSidhartha Kumar VM_BUG_ON_FOLIO(folio_ref_count(folio), folio); 18650356c4b9SSidhartha Kumar VM_BUG_ON_FOLIO(folio_mapcount(folio), folio); 18668ace22bcSYongkai Wu 18670356c4b9SSidhartha Kumar hugetlb_set_folio_subpool(folio, NULL); 18680356c4b9SSidhartha Kumar if (folio_test_anon(folio)) 18690356c4b9SSidhartha Kumar __ClearPageAnonExclusive(&folio->page); 18700356c4b9SSidhartha Kumar folio->mapping = NULL; 18710356c4b9SSidhartha Kumar restore_reserve = folio_test_hugetlb_restore_reserve(folio); 18720356c4b9SSidhartha Kumar folio_clear_hugetlb_restore_reserve(folio); 187327a85ef1SDavid Gibson 18741c5ecae3SMike Kravetz /* 1875d6995da3SMike Kravetz * If HPageRestoreReserve was set on page, page allocation consumed a 18760919e1b6SMike Kravetz * reservation. If the page was associated with a subpool, there 18770919e1b6SMike Kravetz * would have been a page reserved in the subpool before allocation 18780919e1b6SMike Kravetz * via hugepage_subpool_get_pages(). Since we are 'restoring' the 18796c26d310SMiaohe Lin * reservation, do not call hugepage_subpool_put_pages() as this will 18800919e1b6SMike Kravetz * remove the reserved page from the subpool. 18810919e1b6SMike Kravetz */ 18820919e1b6SMike Kravetz if (!restore_reserve) { 18830919e1b6SMike Kravetz /* 18840919e1b6SMike Kravetz * A return code of zero implies that the subpool will be 18850919e1b6SMike Kravetz * under its minimum size if the reservation is not restored 18860919e1b6SMike Kravetz * after page is free. Therefore, force restore_reserve 18870919e1b6SMike Kravetz * operation. 18881c5ecae3SMike Kravetz */ 18891c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 18901c5ecae3SMike Kravetz restore_reserve = true; 18910919e1b6SMike Kravetz } 18921c5ecae3SMike Kravetz 1893db71ef79SMike Kravetz spin_lock_irqsave(&hugetlb_lock, flags); 18940356c4b9SSidhartha Kumar folio_clear_hugetlb_migratable(folio); 1895d4ab0316SSidhartha Kumar hugetlb_cgroup_uncharge_folio(hstate_index(h), 1896d4ab0316SSidhartha Kumar pages_per_huge_page(h), folio); 1897d4ab0316SSidhartha Kumar hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h), 1898d4ab0316SSidhartha Kumar pages_per_huge_page(h), folio); 189907443a85SJoonsoo Kim if (restore_reserve) 190007443a85SJoonsoo Kim h->resv_huge_pages++; 190107443a85SJoonsoo Kim 19020356c4b9SSidhartha Kumar if (folio_test_hugetlb_temporary(folio)) { 1903cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, false); 1904db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1905d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, folio, true); 1906ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 19070edaecfaSAneesh Kumar K.V /* remove the page from active list */ 1908cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, true); 1909db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1910d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, folio, true); 19117893d1d5SAdam Litke } else { 19125d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1913240d67a8SSidhartha Kumar enqueue_hugetlb_folio(h, folio); 1914db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 191527a85ef1SDavid Gibson } 19161121828aSMike Kravetz } 191727a85ef1SDavid Gibson 1918d3d99fccSOscar Salvador /* 1919d3d99fccSOscar Salvador * Must be called with the hugetlb lock held 1920d3d99fccSOscar Salvador */ 1921d3d99fccSOscar Salvador static void __prep_account_new_huge_page(struct hstate *h, int nid) 1922d3d99fccSOscar Salvador { 1923d3d99fccSOscar Salvador lockdep_assert_held(&hugetlb_lock); 1924d3d99fccSOscar Salvador h->nr_huge_pages++; 1925d3d99fccSOscar Salvador h->nr_huge_pages_node[nid]++; 1926d3d99fccSOscar Salvador } 1927d3d99fccSOscar Salvador 1928de656ed3SSidhartha Kumar static void __prep_new_hugetlb_folio(struct hstate *h, struct folio *folio) 1929b7ba30c6SAndi Kleen { 1930de656ed3SSidhartha Kumar hugetlb_vmemmap_optimize(h, &folio->page); 1931de656ed3SSidhartha Kumar INIT_LIST_HEAD(&folio->lru); 19329fd33058SSidhartha Kumar folio_set_compound_dtor(folio, HUGETLB_PAGE_DTOR); 1933de656ed3SSidhartha Kumar hugetlb_set_folio_subpool(folio, NULL); 1934de656ed3SSidhartha Kumar set_hugetlb_cgroup(folio, NULL); 1935de656ed3SSidhartha Kumar set_hugetlb_cgroup_rsvd(folio, NULL); 1936d3d99fccSOscar Salvador } 1937d3d99fccSOscar Salvador 1938d1c60955SSidhartha Kumar static void prep_new_hugetlb_folio(struct hstate *h, struct folio *folio, int nid) 1939d3d99fccSOscar Salvador { 1940de656ed3SSidhartha Kumar __prep_new_hugetlb_folio(h, folio); 1941db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 1942d3d99fccSOscar Salvador __prep_account_new_huge_page(h, nid); 1943db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1944b7ba30c6SAndi Kleen } 1945b7ba30c6SAndi Kleen 1946d1c60955SSidhartha Kumar static bool __prep_compound_gigantic_folio(struct folio *folio, 1947d1c60955SSidhartha Kumar unsigned int order, bool demote) 194820a0307cSWu Fengguang { 19497118fc29SMike Kravetz int i, j; 195020a0307cSWu Fengguang int nr_pages = 1 << order; 195114455eabSCheng Li struct page *p; 195220a0307cSWu Fengguang 1953d1c60955SSidhartha Kumar __folio_clear_reserved(folio); 19542b21624fSMike Kravetz for (i = 0; i < nr_pages; i++) { 1955d1c60955SSidhartha Kumar p = folio_page(folio, i); 195614455eabSCheng Li 1957ef5a22beSAndrea Arcangeli /* 1958ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1959ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1960ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 19617c8de358SEthon Paul * too. Otherwise drivers using get_user_pages() to access tail 1962ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1963ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1964ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1965ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1966ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1967ef5a22beSAndrea Arcangeli * after get_user_pages(). 1968ef5a22beSAndrea Arcangeli */ 19697fb0728aSMike Kravetz if (i != 0) /* head page cleared above */ 1970ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 19717118fc29SMike Kravetz /* 19727118fc29SMike Kravetz * Subtle and very unlikely 19737118fc29SMike Kravetz * 19747118fc29SMike Kravetz * Gigantic 'page allocators' such as memblock or cma will 19757118fc29SMike Kravetz * return a set of pages with each page ref counted. We need 19767118fc29SMike Kravetz * to turn this set of pages into a compound page with tail 19777118fc29SMike Kravetz * page ref counts set to zero. Code such as speculative page 19787118fc29SMike Kravetz * cache adding could take a ref on a 'to be' tail page. 19797118fc29SMike Kravetz * We need to respect any increased ref count, and only set 19807118fc29SMike Kravetz * the ref count to zero if count is currently 1. If count 1981416d85edSMike Kravetz * is not 1, we return an error. An error return indicates 1982416d85edSMike Kravetz * the set of pages can not be converted to a gigantic page. 1983416d85edSMike Kravetz * The caller who allocated the pages should then discard the 1984416d85edSMike Kravetz * pages using the appropriate free interface. 198534d9e35bSMike Kravetz * 198634d9e35bSMike Kravetz * In the case of demote, the ref count will be zero. 19877118fc29SMike Kravetz */ 198834d9e35bSMike Kravetz if (!demote) { 19897118fc29SMike Kravetz if (!page_ref_freeze(p, 1)) { 1990416d85edSMike Kravetz pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n"); 19917118fc29SMike Kravetz goto out_error; 19927118fc29SMike Kravetz } 199334d9e35bSMike Kravetz } else { 199434d9e35bSMike Kravetz VM_BUG_ON_PAGE(page_count(p), p); 199534d9e35bSMike Kravetz } 19962b21624fSMike Kravetz if (i != 0) 1997d1c60955SSidhartha Kumar set_compound_head(p, &folio->page); 199820a0307cSWu Fengguang } 1999e3b7bf97STarun Sahu __folio_set_head(folio); 2000e3b7bf97STarun Sahu /* we rely on prep_new_hugetlb_folio to set the destructor */ 2001e3b7bf97STarun Sahu folio_set_order(folio, order); 200246f27228SMatthew Wilcox (Oracle) atomic_set(&folio->_entire_mapcount, -1); 2003eec20426SMatthew Wilcox (Oracle) atomic_set(&folio->_nr_pages_mapped, 0); 200494688e8eSMatthew Wilcox (Oracle) atomic_set(&folio->_pincount, 0); 20057118fc29SMike Kravetz return true; 20067118fc29SMike Kravetz 20077118fc29SMike Kravetz out_error: 20082b21624fSMike Kravetz /* undo page modifications made above */ 20092b21624fSMike Kravetz for (j = 0; j < i; j++) { 2010d1c60955SSidhartha Kumar p = folio_page(folio, j); 20112b21624fSMike Kravetz if (j != 0) 20127118fc29SMike Kravetz clear_compound_head(p); 20137118fc29SMike Kravetz set_page_refcounted(p); 20147118fc29SMike Kravetz } 20157118fc29SMike Kravetz /* need to clear PG_reserved on remaining tail pages */ 201614455eabSCheng Li for (; j < nr_pages; j++) { 2017d1c60955SSidhartha Kumar p = folio_page(folio, j); 20187118fc29SMike Kravetz __ClearPageReserved(p); 201914455eabSCheng Li } 20207118fc29SMike Kravetz return false; 202120a0307cSWu Fengguang } 202220a0307cSWu Fengguang 2023d1c60955SSidhartha Kumar static bool prep_compound_gigantic_folio(struct folio *folio, 20248531fc6fSMike Kravetz unsigned int order) 20258531fc6fSMike Kravetz { 2026d1c60955SSidhartha Kumar return __prep_compound_gigantic_folio(folio, order, false); 2027d1c60955SSidhartha Kumar } 2028d1c60955SSidhartha Kumar 2029d1c60955SSidhartha Kumar static bool prep_compound_gigantic_folio_for_demote(struct folio *folio, 2030d1c60955SSidhartha Kumar unsigned int order) 2031d1c60955SSidhartha Kumar { 2032d1c60955SSidhartha Kumar return __prep_compound_gigantic_folio(folio, order, true); 20338531fc6fSMike Kravetz } 20348531fc6fSMike Kravetz 20357795912cSAndrew Morton /* 20367795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 20377795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 20387795912cSAndrew Morton * details. 20397795912cSAndrew Morton */ 204020a0307cSWu Fengguang int PageHuge(struct page *page) 204120a0307cSWu Fengguang { 20422d678c64SMatthew Wilcox (Oracle) struct folio *folio; 20432d678c64SMatthew Wilcox (Oracle) 204420a0307cSWu Fengguang if (!PageCompound(page)) 204520a0307cSWu Fengguang return 0; 20462d678c64SMatthew Wilcox (Oracle) folio = page_folio(page); 20472d678c64SMatthew Wilcox (Oracle) return folio->_folio_dtor == HUGETLB_PAGE_DTOR; 204820a0307cSWu Fengguang } 204943131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 205043131e14SNaoya Horiguchi 2051957ebbdfSMatthew Wilcox (Oracle) /** 2052957ebbdfSMatthew Wilcox (Oracle) * folio_test_hugetlb - Determine if the folio belongs to hugetlbfs 2053957ebbdfSMatthew Wilcox (Oracle) * @folio: The folio to test. 2054957ebbdfSMatthew Wilcox (Oracle) * 2055957ebbdfSMatthew Wilcox (Oracle) * Context: Any context. Caller should have a reference on the folio to 2056957ebbdfSMatthew Wilcox (Oracle) * prevent it from being turned into a tail page. 2057957ebbdfSMatthew Wilcox (Oracle) * Return: True for hugetlbfs folios, false for anon folios or folios 2058957ebbdfSMatthew Wilcox (Oracle) * belonging to other filesystems. 205927c73ae7SAndrea Arcangeli */ 2060957ebbdfSMatthew Wilcox (Oracle) bool folio_test_hugetlb(struct folio *folio) 206127c73ae7SAndrea Arcangeli { 20622d678c64SMatthew Wilcox (Oracle) if (!folio_test_large(folio)) 2063957ebbdfSMatthew Wilcox (Oracle) return false; 206427c73ae7SAndrea Arcangeli 20652d678c64SMatthew Wilcox (Oracle) return folio->_folio_dtor == HUGETLB_PAGE_DTOR; 206627c73ae7SAndrea Arcangeli } 2067957ebbdfSMatthew Wilcox (Oracle) EXPORT_SYMBOL_GPL(folio_test_hugetlb); 206827c73ae7SAndrea Arcangeli 2069c0d0381aSMike Kravetz /* 2070c0d0381aSMike Kravetz * Find and lock address space (mapping) in write mode. 2071c0d0381aSMike Kravetz * 2072336bf30eSMike Kravetz * Upon entry, the page is locked which means that page_mapping() is 2073336bf30eSMike Kravetz * stable. Due to locking order, we can only trylock_write. If we can 2074336bf30eSMike Kravetz * not get the lock, simply return NULL to caller. 2075c0d0381aSMike Kravetz */ 2076c0d0381aSMike Kravetz struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) 2077c0d0381aSMike Kravetz { 2078336bf30eSMike Kravetz struct address_space *mapping = page_mapping(hpage); 2079c0d0381aSMike Kravetz 2080c0d0381aSMike Kravetz if (!mapping) 2081c0d0381aSMike Kravetz return mapping; 2082c0d0381aSMike Kravetz 2083c0d0381aSMike Kravetz if (i_mmap_trylock_write(mapping)) 2084c0d0381aSMike Kravetz return mapping; 2085c0d0381aSMike Kravetz 2086c0d0381aSMike Kravetz return NULL; 2087c0d0381aSMike Kravetz } 2088c0d0381aSMike Kravetz 2089fe19bd3dSHugh Dickins pgoff_t hugetlb_basepage_index(struct page *page) 209013d60f4bSZhang Yi { 209113d60f4bSZhang Yi struct page *page_head = compound_head(page); 209213d60f4bSZhang Yi pgoff_t index = page_index(page_head); 209313d60f4bSZhang Yi unsigned long compound_idx; 209413d60f4bSZhang Yi 209523baf831SKirill A. Shutemov if (compound_order(page_head) > MAX_ORDER) 209613d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 209713d60f4bSZhang Yi else 209813d60f4bSZhang Yi compound_idx = page - page_head; 209913d60f4bSZhang Yi 210013d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 210113d60f4bSZhang Yi } 210213d60f4bSZhang Yi 210319fc1a7eSSidhartha Kumar static struct folio *alloc_buddy_hugetlb_folio(struct hstate *h, 2104f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 2105f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 21061da177e4SLinus Torvalds { 2107af0fb9dfSMichal Hocko int order = huge_page_order(h); 21081da177e4SLinus Torvalds struct page *page; 2109f60858f9SMike Kravetz bool alloc_try_hard = true; 21102b21624fSMike Kravetz bool retry = true; 2111f96efd58SJoe Jin 2112f60858f9SMike Kravetz /* 2113f60858f9SMike Kravetz * By default we always try hard to allocate the page with 2114f60858f9SMike Kravetz * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in 2115f60858f9SMike Kravetz * a loop (to adjust global huge page counts) and previous allocation 2116f60858f9SMike Kravetz * failed, do not continue to try hard on the same node. Use the 2117f60858f9SMike Kravetz * node_alloc_noretry bitmap to manage this state information. 2118f60858f9SMike Kravetz */ 2119f60858f9SMike Kravetz if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) 2120f60858f9SMike Kravetz alloc_try_hard = false; 2121f60858f9SMike Kravetz gfp_mask |= __GFP_COMP|__GFP_NOWARN; 2122f60858f9SMike Kravetz if (alloc_try_hard) 2123f60858f9SMike Kravetz gfp_mask |= __GFP_RETRY_MAYFAIL; 2124af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 2125af0fb9dfSMichal Hocko nid = numa_mem_id(); 21262b21624fSMike Kravetz retry: 212784172f4bSMatthew Wilcox (Oracle) page = __alloc_pages(gfp_mask, order, nid, nmask); 21282b21624fSMike Kravetz 21292b21624fSMike Kravetz /* Freeze head page */ 21302b21624fSMike Kravetz if (page && !page_ref_freeze(page, 1)) { 21312b21624fSMike Kravetz __free_pages(page, order); 21322b21624fSMike Kravetz if (retry) { /* retry once */ 21332b21624fSMike Kravetz retry = false; 21342b21624fSMike Kravetz goto retry; 21352b21624fSMike Kravetz } 21362b21624fSMike Kravetz /* WOW! twice in a row. */ 21372b21624fSMike Kravetz pr_warn("HugeTLB head page unexpected inflated ref count\n"); 21382b21624fSMike Kravetz page = NULL; 21392b21624fSMike Kravetz } 21402b21624fSMike Kravetz 2141f60858f9SMike Kravetz /* 2142f60858f9SMike Kravetz * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this 2143f60858f9SMike Kravetz * indicates an overall state change. Clear bit so that we resume 2144f60858f9SMike Kravetz * normal 'try hard' allocations. 2145f60858f9SMike Kravetz */ 2146f60858f9SMike Kravetz if (node_alloc_noretry && page && !alloc_try_hard) 2147f60858f9SMike Kravetz node_clear(nid, *node_alloc_noretry); 2148f60858f9SMike Kravetz 2149f60858f9SMike Kravetz /* 2150f60858f9SMike Kravetz * If we tried hard to get a page but failed, set bit so that 2151f60858f9SMike Kravetz * subsequent attempts will not try as hard until there is an 2152f60858f9SMike Kravetz * overall state change. 2153f60858f9SMike Kravetz */ 2154f60858f9SMike Kravetz if (node_alloc_noretry && !page && alloc_try_hard) 2155f60858f9SMike Kravetz node_set(nid, *node_alloc_noretry); 2156f60858f9SMike Kravetz 215719fc1a7eSSidhartha Kumar if (!page) { 215819fc1a7eSSidhartha Kumar __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 215919fc1a7eSSidhartha Kumar return NULL; 216019fc1a7eSSidhartha Kumar } 216119fc1a7eSSidhartha Kumar 216219fc1a7eSSidhartha Kumar __count_vm_event(HTLB_BUDDY_PGALLOC); 216319fc1a7eSSidhartha Kumar return page_folio(page); 216463b4613cSNishanth Aravamudan } 216563b4613cSNishanth Aravamudan 2166af0fb9dfSMichal Hocko /* 21670c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 21680c397daeSMichal Hocko * should use this function to get new hugetlb pages 21692b21624fSMike Kravetz * 21702b21624fSMike Kravetz * Note that returned page is 'frozen': ref count of head page and all tail 21712b21624fSMike Kravetz * pages is zero. 21720c397daeSMichal Hocko */ 217319fc1a7eSSidhartha Kumar static struct folio *alloc_fresh_hugetlb_folio(struct hstate *h, 2174f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 2175f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 21760c397daeSMichal Hocko { 21777f325a8dSSidhartha Kumar struct folio *folio; 21787118fc29SMike Kravetz bool retry = false; 21790c397daeSMichal Hocko 21807118fc29SMike Kravetz retry: 21810c397daeSMichal Hocko if (hstate_is_gigantic(h)) 218219fc1a7eSSidhartha Kumar folio = alloc_gigantic_folio(h, gfp_mask, nid, nmask); 21830c397daeSMichal Hocko else 218419fc1a7eSSidhartha Kumar folio = alloc_buddy_hugetlb_folio(h, gfp_mask, 2185f60858f9SMike Kravetz nid, nmask, node_alloc_noretry); 218619fc1a7eSSidhartha Kumar if (!folio) 21870c397daeSMichal Hocko return NULL; 21887118fc29SMike Kravetz if (hstate_is_gigantic(h)) { 2189d1c60955SSidhartha Kumar if (!prep_compound_gigantic_folio(folio, huge_page_order(h))) { 21907118fc29SMike Kravetz /* 21917118fc29SMike Kravetz * Rare failure to convert pages to compound page. 21927118fc29SMike Kravetz * Free pages and try again - ONCE! 21937118fc29SMike Kravetz */ 21947f325a8dSSidhartha Kumar free_gigantic_folio(folio, huge_page_order(h)); 21957118fc29SMike Kravetz if (!retry) { 21967118fc29SMike Kravetz retry = true; 21977118fc29SMike Kravetz goto retry; 21987118fc29SMike Kravetz } 21997118fc29SMike Kravetz return NULL; 22007118fc29SMike Kravetz } 22017118fc29SMike Kravetz } 2202d1c60955SSidhartha Kumar prep_new_hugetlb_folio(h, folio, folio_nid(folio)); 22030c397daeSMichal Hocko 220419fc1a7eSSidhartha Kumar return folio; 22050c397daeSMichal Hocko } 22060c397daeSMichal Hocko 22070c397daeSMichal Hocko /* 2208af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 2209af0fb9dfSMichal Hocko * manner. 2210af0fb9dfSMichal Hocko */ 2211f60858f9SMike Kravetz static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 2212f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 2213b2261026SJoonsoo Kim { 221419fc1a7eSSidhartha Kumar struct folio *folio; 2215b2261026SJoonsoo Kim int nr_nodes, node; 2216af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2217b2261026SJoonsoo Kim 2218b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 221919fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, node, 222019fc1a7eSSidhartha Kumar nodes_allowed, node_alloc_noretry); 222119fc1a7eSSidhartha Kumar if (folio) { 222219fc1a7eSSidhartha Kumar free_huge_page(&folio->page); /* free it into the hugepage allocator */ 222319fc1a7eSSidhartha Kumar return 1; 222419fc1a7eSSidhartha Kumar } 2225b2261026SJoonsoo Kim } 2226b2261026SJoonsoo Kim 2227af0fb9dfSMichal Hocko return 0; 2228b2261026SJoonsoo Kim } 2229b2261026SJoonsoo Kim 2230e8c5c824SLee Schermerhorn /* 223110c6ec49SMike Kravetz * Remove huge page from pool from next node to free. Attempt to keep 223210c6ec49SMike Kravetz * persistent huge pages more or less balanced over allowed nodes. 223310c6ec49SMike Kravetz * This routine only 'removes' the hugetlb page. The caller must make 223410c6ec49SMike Kravetz * an additional call to free the page to low level allocators. 2235e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 2236e8c5c824SLee Schermerhorn */ 223710c6ec49SMike Kravetz static struct page *remove_pool_huge_page(struct hstate *h, 223810c6ec49SMike Kravetz nodemask_t *nodes_allowed, 22396ae11b27SLee Schermerhorn bool acct_surplus) 2240e8c5c824SLee Schermerhorn { 2241b2261026SJoonsoo Kim int nr_nodes, node; 224210c6ec49SMike Kravetz struct page *page = NULL; 2243cfd5082bSSidhartha Kumar struct folio *folio; 2244e8c5c824SLee Schermerhorn 22459487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2246b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2247685f3457SLee Schermerhorn /* 2248685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 2249685f3457SLee Schermerhorn * nodes with surplus pages. 2250685f3457SLee Schermerhorn */ 2251b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 2252b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 225310c6ec49SMike Kravetz page = list_entry(h->hugepage_freelists[node].next, 2254e8c5c824SLee Schermerhorn struct page, lru); 2255cfd5082bSSidhartha Kumar folio = page_folio(page); 2256cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, acct_surplus); 22579a76db09SLee Schermerhorn break; 2258e8c5c824SLee Schermerhorn } 2259b2261026SJoonsoo Kim } 2260e8c5c824SLee Schermerhorn 226110c6ec49SMike Kravetz return page; 2262e8c5c824SLee Schermerhorn } 2263e8c5c824SLee Schermerhorn 2264c8721bbbSNaoya Horiguchi /* 2265c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 2266faf53defSNaoya Horiguchi * nothing for in-use hugepages and non-hugepages. 2267faf53defSNaoya Horiguchi * This function returns values like below: 2268faf53defSNaoya Horiguchi * 2269ad2fa371SMuchun Song * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages 2270ad2fa371SMuchun Song * when the system is under memory pressure and the feature of 2271ad2fa371SMuchun Song * freeing unused vmemmap pages associated with each hugetlb page 2272ad2fa371SMuchun Song * is enabled. 2273faf53defSNaoya Horiguchi * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 2274faf53defSNaoya Horiguchi * (allocated or reserved.) 2275faf53defSNaoya Horiguchi * 0: successfully dissolved free hugepages or the page is not a 2276faf53defSNaoya Horiguchi * hugepage (considered as already dissolved) 2277c8721bbbSNaoya Horiguchi */ 2278c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 2279c8721bbbSNaoya Horiguchi { 22806bc9b564SNaoya Horiguchi int rc = -EBUSY; 22811a7cdab5SSidhartha Kumar struct folio *folio = page_folio(page); 2282082d5b6bSGerald Schaefer 22837ffddd49SMuchun Song retry: 2284faf53defSNaoya Horiguchi /* Not to disrupt normal path by vainly holding hugetlb_lock */ 22851a7cdab5SSidhartha Kumar if (!folio_test_hugetlb(folio)) 2286faf53defSNaoya Horiguchi return 0; 2287faf53defSNaoya Horiguchi 2288db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 22891a7cdab5SSidhartha Kumar if (!folio_test_hugetlb(folio)) { 2290faf53defSNaoya Horiguchi rc = 0; 2291faf53defSNaoya Horiguchi goto out; 2292faf53defSNaoya Horiguchi } 2293faf53defSNaoya Horiguchi 22941a7cdab5SSidhartha Kumar if (!folio_ref_count(folio)) { 22951a7cdab5SSidhartha Kumar struct hstate *h = folio_hstate(folio); 22968346d69dSXin Hao if (!available_huge_pages(h)) 2297082d5b6bSGerald Schaefer goto out; 22987ffddd49SMuchun Song 22997ffddd49SMuchun Song /* 23007ffddd49SMuchun Song * We should make sure that the page is already on the free list 23017ffddd49SMuchun Song * when it is dissolved. 23027ffddd49SMuchun Song */ 23031a7cdab5SSidhartha Kumar if (unlikely(!folio_test_hugetlb_freed(folio))) { 2304db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 23057ffddd49SMuchun Song cond_resched(); 23067ffddd49SMuchun Song 23077ffddd49SMuchun Song /* 23087ffddd49SMuchun Song * Theoretically, we should return -EBUSY when we 23097ffddd49SMuchun Song * encounter this race. In fact, we have a chance 23107ffddd49SMuchun Song * to successfully dissolve the page if we do a 23117ffddd49SMuchun Song * retry. Because the race window is quite small. 23127ffddd49SMuchun Song * If we seize this opportunity, it is an optimization 23137ffddd49SMuchun Song * for increasing the success rate of dissolving page. 23147ffddd49SMuchun Song */ 23157ffddd49SMuchun Song goto retry; 23167ffddd49SMuchun Song } 23177ffddd49SMuchun Song 2318cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, false); 2319ad2fa371SMuchun Song h->max_huge_pages--; 2320ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2321ad2fa371SMuchun Song 2322c3114a84SAnshuman Khandual /* 2323d6ef19e2SSidhartha Kumar * Normally update_and_free_hugtlb_folio will allocate required vmemmmap 2324d6ef19e2SSidhartha Kumar * before freeing the page. update_and_free_hugtlb_folio will fail to 2325ad2fa371SMuchun Song * free the page if it can not allocate required vmemmap. We 2326ad2fa371SMuchun Song * need to adjust max_huge_pages if the page is not freed. 2327ad2fa371SMuchun Song * Attempt to allocate vmemmmap here so that we can take 2328ad2fa371SMuchun Song * appropriate action on failure. 2329ad2fa371SMuchun Song */ 23301a7cdab5SSidhartha Kumar rc = hugetlb_vmemmap_restore(h, &folio->page); 2331ad2fa371SMuchun Song if (!rc) { 2332d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, folio, false); 2333ad2fa371SMuchun Song } else { 2334ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 23352f6c57d6SSidhartha Kumar add_hugetlb_folio(h, folio, false); 2336ad2fa371SMuchun Song h->max_huge_pages++; 2337ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2338ad2fa371SMuchun Song } 2339ad2fa371SMuchun Song 2340ad2fa371SMuchun Song return rc; 2341c8721bbbSNaoya Horiguchi } 2342082d5b6bSGerald Schaefer out: 2343db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2344082d5b6bSGerald Schaefer return rc; 2345c8721bbbSNaoya Horiguchi } 2346c8721bbbSNaoya Horiguchi 2347c8721bbbSNaoya Horiguchi /* 2348c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 2349c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 23502247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 23512247bb33SGerald Schaefer * part of it lies within the given range. 2352082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 2353082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 2354c8721bbbSNaoya Horiguchi */ 2355082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 2356c8721bbbSNaoya Horiguchi { 2357c8721bbbSNaoya Horiguchi unsigned long pfn; 2358eb03aa00SGerald Schaefer struct page *page; 2359082d5b6bSGerald Schaefer int rc = 0; 2360dc2628f3SMuchun Song unsigned int order; 2361dc2628f3SMuchun Song struct hstate *h; 2362c8721bbbSNaoya Horiguchi 2363d0177639SLi Zhong if (!hugepages_supported()) 2364082d5b6bSGerald Schaefer return rc; 2365d0177639SLi Zhong 2366dc2628f3SMuchun Song order = huge_page_order(&default_hstate); 2367dc2628f3SMuchun Song for_each_hstate(h) 2368dc2628f3SMuchun Song order = min(order, huge_page_order(h)); 2369dc2628f3SMuchun Song 2370dc2628f3SMuchun Song for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) { 2371eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 2372eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 2373eb03aa00SGerald Schaefer if (rc) 2374082d5b6bSGerald Schaefer break; 2375eb03aa00SGerald Schaefer } 2376082d5b6bSGerald Schaefer 2377082d5b6bSGerald Schaefer return rc; 2378c8721bbbSNaoya Horiguchi } 2379c8721bbbSNaoya Horiguchi 2380ab5ac90aSMichal Hocko /* 2381ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 2382ab5ac90aSMichal Hocko */ 23833a740e8bSSidhartha Kumar static struct folio *alloc_surplus_hugetlb_folio(struct hstate *h, 23843a740e8bSSidhartha Kumar gfp_t gfp_mask, int nid, nodemask_t *nmask) 23857893d1d5SAdam Litke { 238619fc1a7eSSidhartha Kumar struct folio *folio = NULL; 23877893d1d5SAdam Litke 2388bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2389aa888a74SAndi Kleen return NULL; 2390aa888a74SAndi Kleen 2391db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 23929980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 23939980d744SMichal Hocko goto out_unlock; 2394db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2395d1c3fb1fSNishanth Aravamudan 239619fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL); 239719fc1a7eSSidhartha Kumar if (!folio) 23980c397daeSMichal Hocko return NULL; 2399d1c3fb1fSNishanth Aravamudan 2400db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 24019980d744SMichal Hocko /* 24029980d744SMichal Hocko * We could have raced with the pool size change. 24039980d744SMichal Hocko * Double check that and simply deallocate the new page 24049980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 24059980d744SMichal Hocko * temporary page to workaround the nasty free_huge_page 24069980d744SMichal Hocko * codeflow 24079980d744SMichal Hocko */ 24089980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 240919fc1a7eSSidhartha Kumar folio_set_hugetlb_temporary(folio); 2410db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 241119fc1a7eSSidhartha Kumar free_huge_page(&folio->page); 24122bf753e6SKai Shen return NULL; 2413b65a4edaSMike Kravetz } 2414b65a4edaSMike Kravetz 24159980d744SMichal Hocko h->surplus_huge_pages++; 241619fc1a7eSSidhartha Kumar h->surplus_huge_pages_node[folio_nid(folio)]++; 24179980d744SMichal Hocko 24189980d744SMichal Hocko out_unlock: 2419db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 24207893d1d5SAdam Litke 24213a740e8bSSidhartha Kumar return folio; 24227893d1d5SAdam Litke } 24237893d1d5SAdam Litke 2424e37d3e83SSidhartha Kumar static struct folio *alloc_migrate_hugetlb_folio(struct hstate *h, gfp_t gfp_mask, 2425ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 2426ab5ac90aSMichal Hocko { 242719fc1a7eSSidhartha Kumar struct folio *folio; 2428ab5ac90aSMichal Hocko 2429ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 2430ab5ac90aSMichal Hocko return NULL; 2431ab5ac90aSMichal Hocko 243219fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL); 243319fc1a7eSSidhartha Kumar if (!folio) 2434ab5ac90aSMichal Hocko return NULL; 2435ab5ac90aSMichal Hocko 24362b21624fSMike Kravetz /* fresh huge pages are frozen */ 243719fc1a7eSSidhartha Kumar folio_ref_unfreeze(folio, 1); 2438ab5ac90aSMichal Hocko /* 2439ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 2440ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 2441ab5ac90aSMichal Hocko */ 244219fc1a7eSSidhartha Kumar folio_set_hugetlb_temporary(folio); 2443ab5ac90aSMichal Hocko 2444e37d3e83SSidhartha Kumar return folio; 2445ab5ac90aSMichal Hocko } 2446ab5ac90aSMichal Hocko 2447e4e574b7SAdam Litke /* 2448099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 2449099730d6SDave Hansen */ 2450e0ec90eeSDave Hansen static 2451ff7d853bSSidhartha Kumar struct folio *alloc_buddy_hugetlb_folio_with_mpol(struct hstate *h, 2452099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 2453099730d6SDave Hansen { 24543a740e8bSSidhartha Kumar struct folio *folio = NULL; 2455aaf14e40SMichal Hocko struct mempolicy *mpol; 2456aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 2457aaf14e40SMichal Hocko int nid; 2458aaf14e40SMichal Hocko nodemask_t *nodemask; 2459aaf14e40SMichal Hocko 2460aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 2461cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 2462cfcaa66fSBen Widawsky gfp_t gfp = gfp_mask | __GFP_NOWARN; 2463cfcaa66fSBen Widawsky 2464cfcaa66fSBen Widawsky gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL); 24653a740e8bSSidhartha Kumar folio = alloc_surplus_hugetlb_folio(h, gfp, nid, nodemask); 2466cfcaa66fSBen Widawsky 2467cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 2468cfcaa66fSBen Widawsky nodemask = NULL; 2469cfcaa66fSBen Widawsky } 2470cfcaa66fSBen Widawsky 24713a740e8bSSidhartha Kumar if (!folio) 24723a740e8bSSidhartha Kumar folio = alloc_surplus_hugetlb_folio(h, gfp_mask, nid, nodemask); 2473aaf14e40SMichal Hocko mpol_cond_put(mpol); 2474ff7d853bSSidhartha Kumar return folio; 2475099730d6SDave Hansen } 2476099730d6SDave Hansen 2477e37d3e83SSidhartha Kumar /* folio migration callback function */ 2478e37d3e83SSidhartha Kumar struct folio *alloc_hugetlb_folio_nodemask(struct hstate *h, int preferred_nid, 2479d92bbc27SJoonsoo Kim nodemask_t *nmask, gfp_t gfp_mask) 24804db9b2efSMichal Hocko { 2481db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 24828346d69dSXin Hao if (available_huge_pages(h)) { 2483a36f1e90SSidhartha Kumar struct folio *folio; 24843e59fcb0SMichal Hocko 2485a36f1e90SSidhartha Kumar folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask, 2486a36f1e90SSidhartha Kumar preferred_nid, nmask); 2487a36f1e90SSidhartha Kumar if (folio) { 2488db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2489e37d3e83SSidhartha Kumar return folio; 24904db9b2efSMichal Hocko } 24914db9b2efSMichal Hocko } 2492db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 24934db9b2efSMichal Hocko 2494e37d3e83SSidhartha Kumar return alloc_migrate_hugetlb_folio(h, gfp_mask, preferred_nid, nmask); 24954db9b2efSMichal Hocko } 24964db9b2efSMichal Hocko 2497ebd63723SMichal Hocko /* mempolicy aware migration callback */ 2498d0ce0e47SSidhartha Kumar struct folio *alloc_hugetlb_folio_vma(struct hstate *h, struct vm_area_struct *vma, 2499389c8178SMichal Hocko unsigned long address) 2500ebd63723SMichal Hocko { 2501ebd63723SMichal Hocko struct mempolicy *mpol; 2502ebd63723SMichal Hocko nodemask_t *nodemask; 2503e37d3e83SSidhartha Kumar struct folio *folio; 2504ebd63723SMichal Hocko gfp_t gfp_mask; 2505ebd63723SMichal Hocko int node; 2506ebd63723SMichal Hocko 2507ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 2508ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 2509e37d3e83SSidhartha Kumar folio = alloc_hugetlb_folio_nodemask(h, node, nodemask, gfp_mask); 2510ebd63723SMichal Hocko mpol_cond_put(mpol); 2511ebd63723SMichal Hocko 2512d0ce0e47SSidhartha Kumar return folio; 2513ebd63723SMichal Hocko } 2514ebd63723SMichal Hocko 2515bf50bab2SNaoya Horiguchi /* 251625985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 2517e4e574b7SAdam Litke * of size 'delta'. 2518e4e574b7SAdam Litke */ 25190a4f3d1bSLiu Xiang static int gather_surplus_pages(struct hstate *h, long delta) 25201b2a1e7bSJules Irenge __must_hold(&hugetlb_lock) 2521e4e574b7SAdam Litke { 252234665341SMiaohe Lin LIST_HEAD(surplus_list); 25233a740e8bSSidhartha Kumar struct folio *folio; 2524e4e574b7SAdam Litke struct page *page, *tmp; 25250a4f3d1bSLiu Xiang int ret; 25260a4f3d1bSLiu Xiang long i; 25270a4f3d1bSLiu Xiang long needed, allocated; 252828073b02SHillf Danton bool alloc_ok = true; 2529e4e574b7SAdam Litke 25309487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2531a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 2532ac09b3a1SAdam Litke if (needed <= 0) { 2533a5516438SAndi Kleen h->resv_huge_pages += delta; 2534e4e574b7SAdam Litke return 0; 2535ac09b3a1SAdam Litke } 2536e4e574b7SAdam Litke 2537e4e574b7SAdam Litke allocated = 0; 2538e4e574b7SAdam Litke 2539e4e574b7SAdam Litke ret = -ENOMEM; 2540e4e574b7SAdam Litke retry: 2541db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2542e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 25433a740e8bSSidhartha Kumar folio = alloc_surplus_hugetlb_folio(h, htlb_alloc_mask(h), 25442b21624fSMike Kravetz NUMA_NO_NODE, NULL); 25453a740e8bSSidhartha Kumar if (!folio) { 254628073b02SHillf Danton alloc_ok = false; 254728073b02SHillf Danton break; 254828073b02SHillf Danton } 25493a740e8bSSidhartha Kumar list_add(&folio->lru, &surplus_list); 255069ed779aSDavid Rientjes cond_resched(); 2551e4e574b7SAdam Litke } 255228073b02SHillf Danton allocated += i; 2553e4e574b7SAdam Litke 2554e4e574b7SAdam Litke /* 2555e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 2556e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 2557e4e574b7SAdam Litke */ 2558db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2559a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 2560a5516438SAndi Kleen (h->free_huge_pages + allocated); 256128073b02SHillf Danton if (needed > 0) { 256228073b02SHillf Danton if (alloc_ok) 2563e4e574b7SAdam Litke goto retry; 256428073b02SHillf Danton /* 256528073b02SHillf Danton * We were not able to allocate enough pages to 256628073b02SHillf Danton * satisfy the entire reservation so we free what 256728073b02SHillf Danton * we've allocated so far. 256828073b02SHillf Danton */ 256928073b02SHillf Danton goto free; 257028073b02SHillf Danton } 2571e4e574b7SAdam Litke /* 2572e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 257325985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 2574e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 2575ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 2576ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 2577ac09b3a1SAdam Litke * before they are reserved. 2578e4e574b7SAdam Litke */ 2579e4e574b7SAdam Litke needed += allocated; 2580a5516438SAndi Kleen h->resv_huge_pages += delta; 2581e4e574b7SAdam Litke ret = 0; 2582a9869b83SNaoya Horiguchi 258319fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 258419fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 258519fc3f0aSAdam Litke if ((--needed) < 0) 258619fc3f0aSAdam Litke break; 2587b65a4edaSMike Kravetz /* Add the page to the hugetlb allocator */ 2588240d67a8SSidhartha Kumar enqueue_hugetlb_folio(h, page_folio(page)); 258919fc3f0aSAdam Litke } 259028073b02SHillf Danton free: 2591db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 259219fc3f0aSAdam Litke 2593b65a4edaSMike Kravetz /* 2594b65a4edaSMike Kravetz * Free unnecessary surplus pages to the buddy allocator. 2595b65a4edaSMike Kravetz * Pages have no ref count, call free_huge_page directly. 2596b65a4edaSMike Kravetz */ 2597c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 2598b65a4edaSMike Kravetz free_huge_page(page); 2599db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2600e4e574b7SAdam Litke 2601e4e574b7SAdam Litke return ret; 2602e4e574b7SAdam Litke } 2603e4e574b7SAdam Litke 2604e4e574b7SAdam Litke /* 2605e5bbc8a6SMike Kravetz * This routine has two main purposes: 2606e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 2607e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 2608e5bbc8a6SMike Kravetz * to the associated reservation map. 2609e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 2610e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 2611e4e574b7SAdam Litke */ 2612a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 2613a5516438SAndi Kleen unsigned long unused_resv_pages) 2614e4e574b7SAdam Litke { 2615e4e574b7SAdam Litke unsigned long nr_pages; 261610c6ec49SMike Kravetz struct page *page; 261710c6ec49SMike Kravetz LIST_HEAD(page_list); 261810c6ec49SMike Kravetz 26199487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 262010c6ec49SMike Kravetz /* Uncommit the reservation */ 262110c6ec49SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 2622e4e574b7SAdam Litke 2623c0531714SNaoya Horiguchi if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 2624e5bbc8a6SMike Kravetz goto out; 2625aa888a74SAndi Kleen 2626e5bbc8a6SMike Kravetz /* 2627e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 2628e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 2629e5bbc8a6SMike Kravetz */ 2630a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 2631e4e574b7SAdam Litke 2632685f3457SLee Schermerhorn /* 2633685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 26349b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 26359b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 26369b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 263710c6ec49SMike Kravetz * remove_pool_huge_page() will balance the freed pages across the 26389b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 2639685f3457SLee Schermerhorn */ 2640685f3457SLee Schermerhorn while (nr_pages--) { 264110c6ec49SMike Kravetz page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1); 264210c6ec49SMike Kravetz if (!page) 2643e5bbc8a6SMike Kravetz goto out; 264410c6ec49SMike Kravetz 264510c6ec49SMike Kravetz list_add(&page->lru, &page_list); 2646e4e574b7SAdam Litke } 2647e5bbc8a6SMike Kravetz 2648e5bbc8a6SMike Kravetz out: 2649db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 265010c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 2651db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2652e4e574b7SAdam Litke } 2653e4e574b7SAdam Litke 26545e911373SMike Kravetz 2655c37f9fb1SAndy Whitcroft /* 2656feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 26575e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 2658cf3ad20bSMike Kravetz * 2659cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 2660cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 2661cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 2662cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 2663cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 2664feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 2665feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 2666feba16e2SMike Kravetz * is called in such cases. 2667cf3ad20bSMike Kravetz * 2668cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 2669cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 2670cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 2671cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 2672cf3ad20bSMike Kravetz * take appropriate action. 267396b96a96SMike Kravetz * 267496b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 267596b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 267696b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 267796b96a96SMike Kravetz * reservation exists. 2678846be085SMike Kravetz * 2679846be085SMike Kravetz * vma_del_reservation is used in error paths where an entry in the reserve 2680846be085SMike Kravetz * map was created during huge page allocation and must be removed. It is to 2681846be085SMike Kravetz * be called after calling vma_needs_reservation to determine if a reservation 2682846be085SMike Kravetz * exists. 2683c37f9fb1SAndy Whitcroft */ 26845e911373SMike Kravetz enum vma_resv_mode { 26855e911373SMike Kravetz VMA_NEEDS_RESV, 26865e911373SMike Kravetz VMA_COMMIT_RESV, 2687feba16e2SMike Kravetz VMA_END_RESV, 268896b96a96SMike Kravetz VMA_ADD_RESV, 2689846be085SMike Kravetz VMA_DEL_RESV, 26905e911373SMike Kravetz }; 2691cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 2692cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 26935e911373SMike Kravetz enum vma_resv_mode mode) 2694c37f9fb1SAndy Whitcroft { 26954e35f483SJoonsoo Kim struct resv_map *resv; 26964e35f483SJoonsoo Kim pgoff_t idx; 2697cf3ad20bSMike Kravetz long ret; 26980db9d74eSMina Almasry long dummy_out_regions_needed; 2699c37f9fb1SAndy Whitcroft 27004e35f483SJoonsoo Kim resv = vma_resv_map(vma); 27014e35f483SJoonsoo Kim if (!resv) 2702c37f9fb1SAndy Whitcroft return 1; 2703c37f9fb1SAndy Whitcroft 27044e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 27055e911373SMike Kravetz switch (mode) { 27065e911373SMike Kravetz case VMA_NEEDS_RESV: 27070db9d74eSMina Almasry ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); 27080db9d74eSMina Almasry /* We assume that vma_reservation_* routines always operate on 27090db9d74eSMina Almasry * 1 page, and that adding to resv map a 1 page entry can only 27100db9d74eSMina Almasry * ever require 1 region. 27110db9d74eSMina Almasry */ 27120db9d74eSMina Almasry VM_BUG_ON(dummy_out_regions_needed != 1); 27135e911373SMike Kravetz break; 27145e911373SMike Kravetz case VMA_COMMIT_RESV: 2715075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 27160db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 27170db9d74eSMina Almasry VM_BUG_ON(ret < 0); 27185e911373SMike Kravetz break; 2719feba16e2SMike Kravetz case VMA_END_RESV: 27200db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 27215e911373SMike Kravetz ret = 0; 27225e911373SMike Kravetz break; 272396b96a96SMike Kravetz case VMA_ADD_RESV: 27240db9d74eSMina Almasry if (vma->vm_flags & VM_MAYSHARE) { 2725075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 27260db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 27270db9d74eSMina Almasry VM_BUG_ON(ret < 0); 27280db9d74eSMina Almasry } else { 27290db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 273096b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 273196b96a96SMike Kravetz } 273296b96a96SMike Kravetz break; 2733846be085SMike Kravetz case VMA_DEL_RESV: 2734846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 2735846be085SMike Kravetz region_abort(resv, idx, idx + 1, 1); 2736846be085SMike Kravetz ret = region_del(resv, idx, idx + 1); 2737846be085SMike Kravetz } else { 2738846be085SMike Kravetz ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 2739846be085SMike Kravetz /* region_add calls of range 1 should never fail. */ 2740846be085SMike Kravetz VM_BUG_ON(ret < 0); 2741846be085SMike Kravetz } 2742846be085SMike Kravetz break; 27435e911373SMike Kravetz default: 27445e911373SMike Kravetz BUG(); 27455e911373SMike Kravetz } 274684afd99bSAndy Whitcroft 2747846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV) 2748cf3ad20bSMike Kravetz return ret; 274967961f9dSMike Kravetz /* 2750bf3d12b9SMiaohe Lin * We know private mapping must have HPAGE_RESV_OWNER set. 2751bf3d12b9SMiaohe Lin * 275267961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 275367961f9dSMike Kravetz * However, a file associated with mapping could have been 275467961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 275567961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 275667961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 275767961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 275867961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 275967961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 276067961f9dSMike Kravetz * reservation has already been consumed. As a result, the 276167961f9dSMike Kravetz * return value of this routine is the opposite of the 276267961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 276367961f9dSMike Kravetz */ 2764bf3d12b9SMiaohe Lin if (ret > 0) 276567961f9dSMike Kravetz return 0; 2766bf3d12b9SMiaohe Lin if (ret == 0) 276767961f9dSMike Kravetz return 1; 2768bf3d12b9SMiaohe Lin return ret; 276984afd99bSAndy Whitcroft } 2770cf3ad20bSMike Kravetz 2771cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 2772a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 2773c37f9fb1SAndy Whitcroft { 27745e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 2775cf3ad20bSMike Kravetz } 2776c37f9fb1SAndy Whitcroft 2777cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 2778cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 2779cf3ad20bSMike Kravetz { 27805e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 27815e911373SMike Kravetz } 27825e911373SMike Kravetz 2783feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 27845e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 27855e911373SMike Kravetz { 2786feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 2787c37f9fb1SAndy Whitcroft } 2788c37f9fb1SAndy Whitcroft 278996b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 279096b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 279196b96a96SMike Kravetz { 279296b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 279396b96a96SMike Kravetz } 279496b96a96SMike Kravetz 2795846be085SMike Kravetz static long vma_del_reservation(struct hstate *h, 2796846be085SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 279796b96a96SMike Kravetz { 2798846be085SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV); 2799846be085SMike Kravetz } 2800846be085SMike Kravetz 2801846be085SMike Kravetz /* 2802846be085SMike Kravetz * This routine is called to restore reservation information on error paths. 2803d0ce0e47SSidhartha Kumar * It should ONLY be called for folios allocated via alloc_hugetlb_folio(), 2804d0ce0e47SSidhartha Kumar * and the hugetlb mutex should remain held when calling this routine. 2805846be085SMike Kravetz * 2806846be085SMike Kravetz * It handles two specific cases: 2807d2d7bb44SSidhartha Kumar * 1) A reservation was in place and the folio consumed the reservation. 2808d2d7bb44SSidhartha Kumar * hugetlb_restore_reserve is set in the folio. 2809d2d7bb44SSidhartha Kumar * 2) No reservation was in place for the page, so hugetlb_restore_reserve is 2810d0ce0e47SSidhartha Kumar * not set. However, alloc_hugetlb_folio always updates the reserve map. 2811846be085SMike Kravetz * 2812846be085SMike Kravetz * In case 1, free_huge_page later in the error path will increment the 2813846be085SMike Kravetz * global reserve count. But, free_huge_page does not have enough context 2814846be085SMike Kravetz * to adjust the reservation map. This case deals primarily with private 2815846be085SMike Kravetz * mappings. Adjust the reserve map here to be consistent with global 2816846be085SMike Kravetz * reserve count adjustments to be made by free_huge_page. Make sure the 2817846be085SMike Kravetz * reserve map indicates there is a reservation present. 2818846be085SMike Kravetz * 2819d0ce0e47SSidhartha Kumar * In case 2, simply undo reserve map modifications done by alloc_hugetlb_folio. 2820846be085SMike Kravetz */ 2821846be085SMike Kravetz void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma, 2822d2d7bb44SSidhartha Kumar unsigned long address, struct folio *folio) 2823846be085SMike Kravetz { 282496b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 282596b96a96SMike Kravetz 28260ffdc38eSSidhartha Kumar if (folio_test_hugetlb_restore_reserve(folio)) { 2827846be085SMike Kravetz if (unlikely(rc < 0)) 282896b96a96SMike Kravetz /* 282996b96a96SMike Kravetz * Rare out of memory condition in reserve map 28300ffdc38eSSidhartha Kumar * manipulation. Clear hugetlb_restore_reserve so 28310ffdc38eSSidhartha Kumar * that global reserve count will not be incremented 283296b96a96SMike Kravetz * by free_huge_page. This will make it appear 28330ffdc38eSSidhartha Kumar * as though the reservation for this folio was 283496b96a96SMike Kravetz * consumed. This may prevent the task from 28350ffdc38eSSidhartha Kumar * faulting in the folio at a later time. This 283696b96a96SMike Kravetz * is better than inconsistent global huge page 283796b96a96SMike Kravetz * accounting of reserve counts. 283896b96a96SMike Kravetz */ 28390ffdc38eSSidhartha Kumar folio_clear_hugetlb_restore_reserve(folio); 2840846be085SMike Kravetz else if (rc) 2841846be085SMike Kravetz (void)vma_add_reservation(h, vma, address); 2842846be085SMike Kravetz else 2843846be085SMike Kravetz vma_end_reservation(h, vma, address); 2844846be085SMike Kravetz } else { 2845846be085SMike Kravetz if (!rc) { 284696b96a96SMike Kravetz /* 2847846be085SMike Kravetz * This indicates there is an entry in the reserve map 2848d0ce0e47SSidhartha Kumar * not added by alloc_hugetlb_folio. We know it was added 2849d0ce0e47SSidhartha Kumar * before the alloc_hugetlb_folio call, otherwise 28500ffdc38eSSidhartha Kumar * hugetlb_restore_reserve would be set on the folio. 2851846be085SMike Kravetz * Remove the entry so that a subsequent allocation 2852846be085SMike Kravetz * does not consume a reservation. 285396b96a96SMike Kravetz */ 2854846be085SMike Kravetz rc = vma_del_reservation(h, vma, address); 2855846be085SMike Kravetz if (rc < 0) 2856846be085SMike Kravetz /* 2857846be085SMike Kravetz * VERY rare out of memory condition. Since 2858846be085SMike Kravetz * we can not delete the entry, set 28590ffdc38eSSidhartha Kumar * hugetlb_restore_reserve so that the reserve 28600ffdc38eSSidhartha Kumar * count will be incremented when the folio 2861846be085SMike Kravetz * is freed. This reserve will be consumed 2862846be085SMike Kravetz * on a subsequent allocation. 2863846be085SMike Kravetz */ 28640ffdc38eSSidhartha Kumar folio_set_hugetlb_restore_reserve(folio); 2865846be085SMike Kravetz } else if (rc < 0) { 2866846be085SMike Kravetz /* 2867846be085SMike Kravetz * Rare out of memory condition from 2868846be085SMike Kravetz * vma_needs_reservation call. Memory allocation is 2869846be085SMike Kravetz * only attempted if a new entry is needed. Therefore, 2870846be085SMike Kravetz * this implies there is not an entry in the 2871846be085SMike Kravetz * reserve map. 2872846be085SMike Kravetz * 2873846be085SMike Kravetz * For shared mappings, no entry in the map indicates 2874846be085SMike Kravetz * no reservation. We are done. 2875846be085SMike Kravetz */ 2876846be085SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 2877846be085SMike Kravetz /* 2878846be085SMike Kravetz * For private mappings, no entry indicates 2879846be085SMike Kravetz * a reservation is present. Since we can 28800ffdc38eSSidhartha Kumar * not add an entry, set hugetlb_restore_reserve 28810ffdc38eSSidhartha Kumar * on the folio so reserve count will be 2882846be085SMike Kravetz * incremented when freed. This reserve will 2883846be085SMike Kravetz * be consumed on a subsequent allocation. 2884846be085SMike Kravetz */ 28850ffdc38eSSidhartha Kumar folio_set_hugetlb_restore_reserve(folio); 288696b96a96SMike Kravetz } else 2887846be085SMike Kravetz /* 2888846be085SMike Kravetz * No reservation present, do nothing 2889846be085SMike Kravetz */ 289096b96a96SMike Kravetz vma_end_reservation(h, vma, address); 289196b96a96SMike Kravetz } 289296b96a96SMike Kravetz } 289396b96a96SMike Kravetz 2894369fa227SOscar Salvador /* 289519fc1a7eSSidhartha Kumar * alloc_and_dissolve_hugetlb_folio - Allocate a new folio and dissolve 289619fc1a7eSSidhartha Kumar * the old one 2897369fa227SOscar Salvador * @h: struct hstate old page belongs to 289819fc1a7eSSidhartha Kumar * @old_folio: Old folio to dissolve 2899ae37c7ffSOscar Salvador * @list: List to isolate the page in case we need to 2900369fa227SOscar Salvador * Returns 0 on success, otherwise negated error. 2901369fa227SOscar Salvador */ 290219fc1a7eSSidhartha Kumar static int alloc_and_dissolve_hugetlb_folio(struct hstate *h, 290319fc1a7eSSidhartha Kumar struct folio *old_folio, struct list_head *list) 2904369fa227SOscar Salvador { 2905369fa227SOscar Salvador gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2906de656ed3SSidhartha Kumar int nid = folio_nid(old_folio); 2907de656ed3SSidhartha Kumar struct folio *new_folio; 2908369fa227SOscar Salvador int ret = 0; 2909369fa227SOscar Salvador 2910369fa227SOscar Salvador /* 291119fc1a7eSSidhartha Kumar * Before dissolving the folio, we need to allocate a new one for the 291219fc1a7eSSidhartha Kumar * pool to remain stable. Here, we allocate the folio and 'prep' it 2913f41f2ed4SMuchun Song * by doing everything but actually updating counters and adding to 2914f41f2ed4SMuchun Song * the pool. This simplifies and let us do most of the processing 2915f41f2ed4SMuchun Song * under the lock. 2916369fa227SOscar Salvador */ 291719fc1a7eSSidhartha Kumar new_folio = alloc_buddy_hugetlb_folio(h, gfp_mask, nid, NULL, NULL); 291819fc1a7eSSidhartha Kumar if (!new_folio) 2919369fa227SOscar Salvador return -ENOMEM; 2920de656ed3SSidhartha Kumar __prep_new_hugetlb_folio(h, new_folio); 2921369fa227SOscar Salvador 2922369fa227SOscar Salvador retry: 2923369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2924de656ed3SSidhartha Kumar if (!folio_test_hugetlb(old_folio)) { 2925369fa227SOscar Salvador /* 292619fc1a7eSSidhartha Kumar * Freed from under us. Drop new_folio too. 2927369fa227SOscar Salvador */ 2928369fa227SOscar Salvador goto free_new; 2929de656ed3SSidhartha Kumar } else if (folio_ref_count(old_folio)) { 29309747b9e9SBaolin Wang bool isolated; 29319747b9e9SBaolin Wang 2932369fa227SOscar Salvador /* 293319fc1a7eSSidhartha Kumar * Someone has grabbed the folio, try to isolate it here. 2934ae37c7ffSOscar Salvador * Fail with -EBUSY if not possible. 2935369fa227SOscar Salvador */ 2936ae37c7ffSOscar Salvador spin_unlock_irq(&hugetlb_lock); 29379747b9e9SBaolin Wang isolated = isolate_hugetlb(old_folio, list); 29389747b9e9SBaolin Wang ret = isolated ? 0 : -EBUSY; 2939ae37c7ffSOscar Salvador spin_lock_irq(&hugetlb_lock); 2940369fa227SOscar Salvador goto free_new; 2941de656ed3SSidhartha Kumar } else if (!folio_test_hugetlb_freed(old_folio)) { 2942369fa227SOscar Salvador /* 294319fc1a7eSSidhartha Kumar * Folio's refcount is 0 but it has not been enqueued in the 2944369fa227SOscar Salvador * freelist yet. Race window is small, so we can succeed here if 2945369fa227SOscar Salvador * we retry. 2946369fa227SOscar Salvador */ 2947369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2948369fa227SOscar Salvador cond_resched(); 2949369fa227SOscar Salvador goto retry; 2950369fa227SOscar Salvador } else { 2951369fa227SOscar Salvador /* 295219fc1a7eSSidhartha Kumar * Ok, old_folio is still a genuine free hugepage. Remove it from 2953369fa227SOscar Salvador * the freelist and decrease the counters. These will be 2954369fa227SOscar Salvador * incremented again when calling __prep_account_new_huge_page() 2955240d67a8SSidhartha Kumar * and enqueue_hugetlb_folio() for new_folio. The counters will 2956240d67a8SSidhartha Kumar * remain stable since this happens under the lock. 2957369fa227SOscar Salvador */ 2958cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, old_folio, false); 2959369fa227SOscar Salvador 2960369fa227SOscar Salvador /* 296119fc1a7eSSidhartha Kumar * Ref count on new_folio is already zero as it was dropped 2962b65a4edaSMike Kravetz * earlier. It can be directly added to the pool free list. 2963369fa227SOscar Salvador */ 2964369fa227SOscar Salvador __prep_account_new_huge_page(h, nid); 2965240d67a8SSidhartha Kumar enqueue_hugetlb_folio(h, new_folio); 2966369fa227SOscar Salvador 2967369fa227SOscar Salvador /* 296819fc1a7eSSidhartha Kumar * Folio has been replaced, we can safely free the old one. 2969369fa227SOscar Salvador */ 2970369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2971d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, old_folio, false); 2972369fa227SOscar Salvador } 2973369fa227SOscar Salvador 2974369fa227SOscar Salvador return ret; 2975369fa227SOscar Salvador 2976369fa227SOscar Salvador free_new: 2977369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 297819fc1a7eSSidhartha Kumar /* Folio has a zero ref count, but needs a ref to be freed */ 2979de656ed3SSidhartha Kumar folio_ref_unfreeze(new_folio, 1); 2980d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, new_folio, false); 2981369fa227SOscar Salvador 2982369fa227SOscar Salvador return ret; 2983369fa227SOscar Salvador } 2984369fa227SOscar Salvador 2985ae37c7ffSOscar Salvador int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list) 2986369fa227SOscar Salvador { 2987369fa227SOscar Salvador struct hstate *h; 2988d5e33bd8SSidhartha Kumar struct folio *folio = page_folio(page); 2989ae37c7ffSOscar Salvador int ret = -EBUSY; 2990369fa227SOscar Salvador 2991369fa227SOscar Salvador /* 2992369fa227SOscar Salvador * The page might have been dissolved from under our feet, so make sure 2993369fa227SOscar Salvador * to carefully check the state under the lock. 2994369fa227SOscar Salvador * Return success when racing as if we dissolved the page ourselves. 2995369fa227SOscar Salvador */ 2996369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2997d5e33bd8SSidhartha Kumar if (folio_test_hugetlb(folio)) { 2998d5e33bd8SSidhartha Kumar h = folio_hstate(folio); 2999369fa227SOscar Salvador } else { 3000369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 3001369fa227SOscar Salvador return 0; 3002369fa227SOscar Salvador } 3003369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 3004369fa227SOscar Salvador 3005369fa227SOscar Salvador /* 3006369fa227SOscar Salvador * Fence off gigantic pages as there is a cyclic dependency between 3007369fa227SOscar Salvador * alloc_contig_range and them. Return -ENOMEM as this has the effect 3008369fa227SOscar Salvador * of bailing out right away without further retrying. 3009369fa227SOscar Salvador */ 3010369fa227SOscar Salvador if (hstate_is_gigantic(h)) 3011369fa227SOscar Salvador return -ENOMEM; 3012369fa227SOscar Salvador 30139747b9e9SBaolin Wang if (folio_ref_count(folio) && isolate_hugetlb(folio, list)) 3014ae37c7ffSOscar Salvador ret = 0; 3015d5e33bd8SSidhartha Kumar else if (!folio_ref_count(folio)) 301619fc1a7eSSidhartha Kumar ret = alloc_and_dissolve_hugetlb_folio(h, folio, list); 3017ae37c7ffSOscar Salvador 3018ae37c7ffSOscar Salvador return ret; 3019369fa227SOscar Salvador } 3020369fa227SOscar Salvador 3021d0ce0e47SSidhartha Kumar struct folio *alloc_hugetlb_folio(struct vm_area_struct *vma, 302204f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 3023348ea204SAdam Litke { 302490481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 3025a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3026d4ab0316SSidhartha Kumar struct folio *folio; 3027d85f69b0SMike Kravetz long map_chg, map_commit; 3028d85f69b0SMike Kravetz long gbl_chg; 30296d76dcf4SAneesh Kumar K.V int ret, idx; 3030d0ce0e47SSidhartha Kumar struct hugetlb_cgroup *h_cg = NULL; 303108cf9fafSMina Almasry bool deferred_reserve; 30322fc39cecSAdam Litke 30336d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 3034a1e78772SMel Gorman /* 3035d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 3036d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 3037d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 3038a1e78772SMel Gorman */ 3039d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 3040d85f69b0SMike Kravetz if (map_chg < 0) 304176dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 3042d85f69b0SMike Kravetz 3043d85f69b0SMike Kravetz /* 3044d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 3045d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 3046d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 3047d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 3048d85f69b0SMike Kravetz * checked against any subpool limit. 3049d85f69b0SMike Kravetz */ 3050d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 3051d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 3052d85f69b0SMike Kravetz if (gbl_chg < 0) { 3053feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 305476dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 30555e911373SMike Kravetz } 305690d8b7e6SAdam Litke 3057d85f69b0SMike Kravetz /* 3058d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 3059d85f69b0SMike Kravetz * map, there could be reservations associated with the 3060d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 3061d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 3062d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 3063d85f69b0SMike Kravetz * the subpool reservations. 3064d85f69b0SMike Kravetz */ 3065d85f69b0SMike Kravetz if (avoid_reserve) 3066d85f69b0SMike Kravetz gbl_chg = 1; 3067d85f69b0SMike Kravetz } 3068d85f69b0SMike Kravetz 306908cf9fafSMina Almasry /* If this allocation is not consuming a reservation, charge it now. 307008cf9fafSMina Almasry */ 30716501fe5fSMiaohe Lin deferred_reserve = map_chg || avoid_reserve; 307208cf9fafSMina Almasry if (deferred_reserve) { 307308cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 307408cf9fafSMina Almasry idx, pages_per_huge_page(h), &h_cg); 30758f34af6fSJianyu Zhan if (ret) 30768f34af6fSJianyu Zhan goto out_subpool_put; 307708cf9fafSMina Almasry } 307808cf9fafSMina Almasry 307908cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 308008cf9fafSMina Almasry if (ret) 308108cf9fafSMina Almasry goto out_uncharge_cgroup_reservation; 30828f34af6fSJianyu Zhan 3083db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3084d85f69b0SMike Kravetz /* 3085d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 3086d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 3087d85f69b0SMike Kravetz * a reservation exists for the allocation. 3088d85f69b0SMike Kravetz */ 3089ff7d853bSSidhartha Kumar folio = dequeue_hugetlb_folio_vma(h, vma, addr, avoid_reserve, gbl_chg); 3090ff7d853bSSidhartha Kumar if (!folio) { 3091db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3092ff7d853bSSidhartha Kumar folio = alloc_buddy_hugetlb_folio_with_mpol(h, vma, addr); 3093ff7d853bSSidhartha Kumar if (!folio) 30948f34af6fSJianyu Zhan goto out_uncharge_cgroup; 309512df140fSRik van Riel spin_lock_irq(&hugetlb_lock); 3096a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 3097ff7d853bSSidhartha Kumar folio_set_hugetlb_restore_reserve(folio); 3098a88c7695SNaoya Horiguchi h->resv_huge_pages--; 3099a88c7695SNaoya Horiguchi } 3100ff7d853bSSidhartha Kumar list_add(&folio->lru, &h->hugepage_activelist); 3101ff7d853bSSidhartha Kumar folio_ref_unfreeze(folio, 1); 310281a6fcaeSJoonsoo Kim /* Fall through */ 3103a1e78772SMel Gorman } 3104ff7d853bSSidhartha Kumar 3105ff7d853bSSidhartha Kumar hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, folio); 310608cf9fafSMina Almasry /* If allocation is not consuming a reservation, also store the 310708cf9fafSMina Almasry * hugetlb_cgroup pointer on the page. 310808cf9fafSMina Almasry */ 310908cf9fafSMina Almasry if (deferred_reserve) { 311008cf9fafSMina Almasry hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), 3111ff7d853bSSidhartha Kumar h_cg, folio); 311208cf9fafSMina Almasry } 311308cf9fafSMina Almasry 3114db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3115a1e78772SMel Gorman 3116ff7d853bSSidhartha Kumar hugetlb_set_folio_subpool(folio, spool); 3117a1e78772SMel Gorman 3118d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 3119d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 312033039678SMike Kravetz /* 312133039678SMike Kravetz * The page was added to the reservation map between 312233039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 312333039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 312433039678SMike Kravetz * Adjust for the subpool count incremented above AND 312533039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 312633039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 312733039678SMike Kravetz * no longer applies. 312833039678SMike Kravetz */ 312933039678SMike Kravetz long rsv_adjust; 313033039678SMike Kravetz 313133039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 313233039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 313379aa925bSMike Kravetz if (deferred_reserve) 3134d4ab0316SSidhartha Kumar hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h), 3135d4ab0316SSidhartha Kumar pages_per_huge_page(h), folio); 313633039678SMike Kravetz } 3137d0ce0e47SSidhartha Kumar return folio; 31388f34af6fSJianyu Zhan 31398f34af6fSJianyu Zhan out_uncharge_cgroup: 31408f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 314108cf9fafSMina Almasry out_uncharge_cgroup_reservation: 314208cf9fafSMina Almasry if (deferred_reserve) 314308cf9fafSMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), 314408cf9fafSMina Almasry h_cg); 31458f34af6fSJianyu Zhan out_subpool_put: 3146d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 31478f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 3148feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 31498f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 3150b45b5bd6SDavid Gibson } 3151b45b5bd6SDavid Gibson 3152b5389086SZhenguo Yao int alloc_bootmem_huge_page(struct hstate *h, int nid) 3153e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 3154b5389086SZhenguo Yao int __alloc_bootmem_huge_page(struct hstate *h, int nid) 3155aa888a74SAndi Kleen { 3156b5389086SZhenguo Yao struct huge_bootmem_page *m = NULL; /* initialize for clang */ 3157b2261026SJoonsoo Kim int nr_nodes, node; 3158aa888a74SAndi Kleen 3159b5389086SZhenguo Yao /* do node specific alloc */ 3160b5389086SZhenguo Yao if (nid != NUMA_NO_NODE) { 3161b5389086SZhenguo Yao m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h), 3162b5389086SZhenguo Yao 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid); 3163b5389086SZhenguo Yao if (!m) 3164b5389086SZhenguo Yao return 0; 3165b5389086SZhenguo Yao goto found; 3166b5389086SZhenguo Yao } 3167b5389086SZhenguo Yao /* allocate from next node when distributing huge pages */ 3168b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 3169b5389086SZhenguo Yao m = memblock_alloc_try_nid_raw( 31708b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 317197ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 3172aa888a74SAndi Kleen /* 3173aa888a74SAndi Kleen * Use the beginning of the huge page to store the 3174aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 3175aa888a74SAndi Kleen * puts them into the mem_map). 3176aa888a74SAndi Kleen */ 3177b5389086SZhenguo Yao if (!m) 3178b5389086SZhenguo Yao return 0; 3179aa888a74SAndi Kleen goto found; 3180aa888a74SAndi Kleen } 3181aa888a74SAndi Kleen 3182aa888a74SAndi Kleen found: 3183aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 3184330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 3185aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 3186aa888a74SAndi Kleen m->hstate = h; 3187aa888a74SAndi Kleen return 1; 3188aa888a74SAndi Kleen } 3189aa888a74SAndi Kleen 319048b8d744SMike Kravetz /* 319148b8d744SMike Kravetz * Put bootmem huge pages into the standard lists after mem_map is up. 319248b8d744SMike Kravetz * Note: This only applies to gigantic (order > MAX_ORDER) pages. 319348b8d744SMike Kravetz */ 3194aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 3195aa888a74SAndi Kleen { 3196aa888a74SAndi Kleen struct huge_bootmem_page *m; 3197aa888a74SAndi Kleen 3198aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 319940d18ebfSMike Kravetz struct page *page = virt_to_page(m); 32007f325a8dSSidhartha Kumar struct folio *folio = page_folio(page); 3201aa888a74SAndi Kleen struct hstate *h = m->hstate; 3202ee8f248dSBecky Bruce 320348b8d744SMike Kravetz VM_BUG_ON(!hstate_is_gigantic(h)); 3204d1c60955SSidhartha Kumar WARN_ON(folio_ref_count(folio) != 1); 3205d1c60955SSidhartha Kumar if (prep_compound_gigantic_folio(folio, huge_page_order(h))) { 3206d1c60955SSidhartha Kumar WARN_ON(folio_test_reserved(folio)); 3207d1c60955SSidhartha Kumar prep_new_hugetlb_folio(h, folio, folio_nid(folio)); 32082b21624fSMike Kravetz free_huge_page(page); /* add to the hugepage allocator */ 32097118fc29SMike Kravetz } else { 3210416d85edSMike Kravetz /* VERY unlikely inflated ref count on a tail page */ 32117f325a8dSSidhartha Kumar free_gigantic_folio(folio, huge_page_order(h)); 32127118fc29SMike Kravetz } 3213af0fb9dfSMichal Hocko 3214b0320c7bSRafael Aquini /* 321548b8d744SMike Kravetz * We need to restore the 'stolen' pages to totalram_pages 321648b8d744SMike Kravetz * in order to fix confusing memory reports from free(1) and 321748b8d744SMike Kravetz * other side-effects, like CommitLimit going negative. 3218b0320c7bSRafael Aquini */ 3219c78a7f36SMiaohe Lin adjust_managed_page_count(page, pages_per_huge_page(h)); 3220520495feSCannon Matthews cond_resched(); 3221aa888a74SAndi Kleen } 3222aa888a74SAndi Kleen } 3223b5389086SZhenguo Yao static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid) 3224b5389086SZhenguo Yao { 3225b5389086SZhenguo Yao unsigned long i; 3226b5389086SZhenguo Yao char buf[32]; 3227b5389086SZhenguo Yao 3228b5389086SZhenguo Yao for (i = 0; i < h->max_huge_pages_node[nid]; ++i) { 3229b5389086SZhenguo Yao if (hstate_is_gigantic(h)) { 3230b5389086SZhenguo Yao if (!alloc_bootmem_huge_page(h, nid)) 3231b5389086SZhenguo Yao break; 3232b5389086SZhenguo Yao } else { 323319fc1a7eSSidhartha Kumar struct folio *folio; 3234b5389086SZhenguo Yao gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 3235b5389086SZhenguo Yao 323619fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, 3237b5389086SZhenguo Yao &node_states[N_MEMORY], NULL); 323819fc1a7eSSidhartha Kumar if (!folio) 3239b5389086SZhenguo Yao break; 324019fc1a7eSSidhartha Kumar free_huge_page(&folio->page); /* free it into the hugepage allocator */ 3241b5389086SZhenguo Yao } 3242b5389086SZhenguo Yao cond_resched(); 3243b5389086SZhenguo Yao } 3244b5389086SZhenguo Yao if (i == h->max_huge_pages_node[nid]) 3245b5389086SZhenguo Yao return; 3246b5389086SZhenguo Yao 3247b5389086SZhenguo Yao string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3248b5389086SZhenguo Yao pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n", 3249b5389086SZhenguo Yao h->max_huge_pages_node[nid], buf, nid, i); 3250b5389086SZhenguo Yao h->max_huge_pages -= (h->max_huge_pages_node[nid] - i); 3251b5389086SZhenguo Yao h->max_huge_pages_node[nid] = i; 3252b5389086SZhenguo Yao } 3253aa888a74SAndi Kleen 32548faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 32551da177e4SLinus Torvalds { 32561da177e4SLinus Torvalds unsigned long i; 3257f60858f9SMike Kravetz nodemask_t *node_alloc_noretry; 3258b5389086SZhenguo Yao bool node_specific_alloc = false; 3259f60858f9SMike Kravetz 3260b5389086SZhenguo Yao /* skip gigantic hugepages allocation if hugetlb_cma enabled */ 3261b5389086SZhenguo Yao if (hstate_is_gigantic(h) && hugetlb_cma_size) { 3262b5389086SZhenguo Yao pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); 3263b5389086SZhenguo Yao return; 3264b5389086SZhenguo Yao } 3265b5389086SZhenguo Yao 3266b5389086SZhenguo Yao /* do node specific alloc */ 32670a7a0f6fSPeng Liu for_each_online_node(i) { 3268b5389086SZhenguo Yao if (h->max_huge_pages_node[i] > 0) { 3269b5389086SZhenguo Yao hugetlb_hstate_alloc_pages_onenode(h, i); 3270b5389086SZhenguo Yao node_specific_alloc = true; 3271b5389086SZhenguo Yao } 3272b5389086SZhenguo Yao } 3273b5389086SZhenguo Yao 3274b5389086SZhenguo Yao if (node_specific_alloc) 3275b5389086SZhenguo Yao return; 3276b5389086SZhenguo Yao 3277b5389086SZhenguo Yao /* below will do all node balanced alloc */ 3278f60858f9SMike Kravetz if (!hstate_is_gigantic(h)) { 3279f60858f9SMike Kravetz /* 3280f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3281f60858f9SMike Kravetz * Ignore errors as lower level routines can deal with 3282f60858f9SMike Kravetz * node_alloc_noretry == NULL. If this kmalloc fails at boot 3283f60858f9SMike Kravetz * time, we are likely in bigger trouble. 3284f60858f9SMike Kravetz */ 3285f60858f9SMike Kravetz node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), 3286f60858f9SMike Kravetz GFP_KERNEL); 3287f60858f9SMike Kravetz } else { 3288f60858f9SMike Kravetz /* allocations done at boot time */ 3289f60858f9SMike Kravetz node_alloc_noretry = NULL; 3290f60858f9SMike Kravetz } 3291f60858f9SMike Kravetz 3292f60858f9SMike Kravetz /* bit mask controlling how hard we retry per-node allocations */ 3293f60858f9SMike Kravetz if (node_alloc_noretry) 3294f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 32951da177e4SLinus Torvalds 3296e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 3297bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 3298b5389086SZhenguo Yao if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE)) 3299aa888a74SAndi Kleen break; 33000c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 3301f60858f9SMike Kravetz &node_states[N_MEMORY], 3302f60858f9SMike Kravetz node_alloc_noretry)) 33031da177e4SLinus Torvalds break; 330469ed779aSDavid Rientjes cond_resched(); 33051da177e4SLinus Torvalds } 3306d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 3307d715cf80SLiam R. Howlett char buf[32]; 3308d715cf80SLiam R. Howlett 3309c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3310d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 3311d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 33128faa8b07SAndi Kleen h->max_huge_pages = i; 3313e5ff2159SAndi Kleen } 3314f60858f9SMike Kravetz kfree(node_alloc_noretry); 3315d715cf80SLiam R. Howlett } 3316e5ff2159SAndi Kleen 3317e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 3318e5ff2159SAndi Kleen { 331979dfc695SMike Kravetz struct hstate *h, *h2; 3320e5ff2159SAndi Kleen 3321e5ff2159SAndi Kleen for_each_hstate(h) { 33228faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 3323bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 33248faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 332579dfc695SMike Kravetz 332679dfc695SMike Kravetz /* 332779dfc695SMike Kravetz * Set demote order for each hstate. Note that 332879dfc695SMike Kravetz * h->demote_order is initially 0. 332979dfc695SMike Kravetz * - We can not demote gigantic pages if runtime freeing 333079dfc695SMike Kravetz * is not supported, so skip this. 3331a01f4390SMike Kravetz * - If CMA allocation is possible, we can not demote 3332a01f4390SMike Kravetz * HUGETLB_PAGE_ORDER or smaller size pages. 333379dfc695SMike Kravetz */ 333479dfc695SMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 333579dfc695SMike Kravetz continue; 3336a01f4390SMike Kravetz if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER) 3337a01f4390SMike Kravetz continue; 333879dfc695SMike Kravetz for_each_hstate(h2) { 333979dfc695SMike Kravetz if (h2 == h) 334079dfc695SMike Kravetz continue; 334179dfc695SMike Kravetz if (h2->order < h->order && 334279dfc695SMike Kravetz h2->order > h->demote_order) 334379dfc695SMike Kravetz h->demote_order = h2->order; 334479dfc695SMike Kravetz } 3345e5ff2159SAndi Kleen } 3346e5ff2159SAndi Kleen } 3347e5ff2159SAndi Kleen 3348e5ff2159SAndi Kleen static void __init report_hugepages(void) 3349e5ff2159SAndi Kleen { 3350e5ff2159SAndi Kleen struct hstate *h; 3351e5ff2159SAndi Kleen 3352e5ff2159SAndi Kleen for_each_hstate(h) { 33534abd32dbSAndi Kleen char buf[32]; 3354c6247f72SMatthew Wilcox 3355c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 33566213834cSMuchun Song pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n", 3357c6247f72SMatthew Wilcox buf, h->free_huge_pages); 33586213834cSMuchun Song pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n", 33596213834cSMuchun Song hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf); 3360e5ff2159SAndi Kleen } 3361e5ff2159SAndi Kleen } 3362e5ff2159SAndi Kleen 33631da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 33646ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 33656ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 33661da177e4SLinus Torvalds { 33674415cc8dSChristoph Lameter int i; 33681121828aSMike Kravetz LIST_HEAD(page_list); 33694415cc8dSChristoph Lameter 33709487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 3371bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3372aa888a74SAndi Kleen return; 3373aa888a74SAndi Kleen 33741121828aSMike Kravetz /* 33751121828aSMike Kravetz * Collect pages to be freed on a list, and free after dropping lock 33761121828aSMike Kravetz */ 33776ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 337810c6ec49SMike Kravetz struct page *page, *next; 3379a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 3380a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 3381a5516438SAndi Kleen if (count >= h->nr_huge_pages) 33821121828aSMike Kravetz goto out; 33831da177e4SLinus Torvalds if (PageHighMem(page)) 33841da177e4SLinus Torvalds continue; 3385cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, page_folio(page), false); 33861121828aSMike Kravetz list_add(&page->lru, &page_list); 33871121828aSMike Kravetz } 33881121828aSMike Kravetz } 33891121828aSMike Kravetz 33901121828aSMike Kravetz out: 3391db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 339210c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3393db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 33941da177e4SLinus Torvalds } 33951da177e4SLinus Torvalds #else 33966ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 33976ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 33981da177e4SLinus Torvalds { 33991da177e4SLinus Torvalds } 34001da177e4SLinus Torvalds #endif 34011da177e4SLinus Torvalds 340220a0307cSWu Fengguang /* 340320a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 340420a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 340520a0307cSWu Fengguang * Returns 1 if an adjustment was made. 340620a0307cSWu Fengguang */ 34076ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 34086ae11b27SLee Schermerhorn int delta) 340920a0307cSWu Fengguang { 3410b2261026SJoonsoo Kim int nr_nodes, node; 341120a0307cSWu Fengguang 34129487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 341320a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 341420a0307cSWu Fengguang 3415e8c5c824SLee Schermerhorn if (delta < 0) { 3416b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 3417b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 3418b2261026SJoonsoo Kim goto found; 3419b2261026SJoonsoo Kim } 3420b2261026SJoonsoo Kim } else { 3421b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 3422b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 3423b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 3424b2261026SJoonsoo Kim goto found; 3425e8c5c824SLee Schermerhorn } 34269a76db09SLee Schermerhorn } 3427b2261026SJoonsoo Kim return 0; 342820a0307cSWu Fengguang 3429b2261026SJoonsoo Kim found: 343020a0307cSWu Fengguang h->surplus_huge_pages += delta; 3431b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 3432b2261026SJoonsoo Kim return 1; 343320a0307cSWu Fengguang } 343420a0307cSWu Fengguang 3435a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 3436fd875dcaSMike Kravetz static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 34376ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 34381da177e4SLinus Torvalds { 34397893d1d5SAdam Litke unsigned long min_count, ret; 344010c6ec49SMike Kravetz struct page *page; 344110c6ec49SMike Kravetz LIST_HEAD(page_list); 3442f60858f9SMike Kravetz NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); 3443f60858f9SMike Kravetz 3444f60858f9SMike Kravetz /* 3445f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3446f60858f9SMike Kravetz * If we can not allocate the bit mask, do not attempt to allocate 3447f60858f9SMike Kravetz * the requested huge pages. 3448f60858f9SMike Kravetz */ 3449f60858f9SMike Kravetz if (node_alloc_noretry) 3450f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 3451f60858f9SMike Kravetz else 3452f60858f9SMike Kravetz return -ENOMEM; 34531da177e4SLinus Torvalds 345429383967SMike Kravetz /* 345529383967SMike Kravetz * resize_lock mutex prevents concurrent adjustments to number of 345629383967SMike Kravetz * pages in hstate via the proc/sysfs interfaces. 345729383967SMike Kravetz */ 345829383967SMike Kravetz mutex_lock(&h->resize_lock); 3459b65d4adbSMuchun Song flush_free_hpage_work(h); 3460db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 34614eb0716eSAlexandre Ghiti 34624eb0716eSAlexandre Ghiti /* 3463fd875dcaSMike Kravetz * Check for a node specific request. 3464fd875dcaSMike Kravetz * Changing node specific huge page count may require a corresponding 3465fd875dcaSMike Kravetz * change to the global count. In any case, the passed node mask 3466fd875dcaSMike Kravetz * (nodes_allowed) will restrict alloc/free to the specified node. 3467fd875dcaSMike Kravetz */ 3468fd875dcaSMike Kravetz if (nid != NUMA_NO_NODE) { 3469fd875dcaSMike Kravetz unsigned long old_count = count; 3470fd875dcaSMike Kravetz 3471fd875dcaSMike Kravetz count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 3472fd875dcaSMike Kravetz /* 3473fd875dcaSMike Kravetz * User may have specified a large count value which caused the 3474fd875dcaSMike Kravetz * above calculation to overflow. In this case, they wanted 3475fd875dcaSMike Kravetz * to allocate as many huge pages as possible. Set count to 3476fd875dcaSMike Kravetz * largest possible value to align with their intention. 3477fd875dcaSMike Kravetz */ 3478fd875dcaSMike Kravetz if (count < old_count) 3479fd875dcaSMike Kravetz count = ULONG_MAX; 3480fd875dcaSMike Kravetz } 3481fd875dcaSMike Kravetz 3482fd875dcaSMike Kravetz /* 34834eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 34844eb0716eSAlexandre Ghiti * page range allocation. 34854eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 34864eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 34874eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 34884eb0716eSAlexandre Ghiti */ 34894eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 34904eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 3491db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 349229383967SMike Kravetz mutex_unlock(&h->resize_lock); 3493f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 34944eb0716eSAlexandre Ghiti return -EINVAL; 34954eb0716eSAlexandre Ghiti } 34964eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 34974eb0716eSAlexandre Ghiti } 3498aa888a74SAndi Kleen 34997893d1d5SAdam Litke /* 35007893d1d5SAdam Litke * Increase the pool size 35017893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 35027893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 3503d1c3fb1fSNishanth Aravamudan * 35043a740e8bSSidhartha Kumar * We might race with alloc_surplus_hugetlb_folio() here and be unable 3505d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 3506d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 3507d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 3508d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 35097893d1d5SAdam Litke */ 3510a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 35116ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 35127893d1d5SAdam Litke break; 35137893d1d5SAdam Litke } 35147893d1d5SAdam Litke 3515a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 35167893d1d5SAdam Litke /* 35177893d1d5SAdam Litke * If this allocation races such that we no longer need the 35187893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 35197893d1d5SAdam Litke * and reducing the surplus. 35207893d1d5SAdam Litke */ 3521db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3522649920c6SJia He 3523649920c6SJia He /* yield cpu to avoid soft lockup */ 3524649920c6SJia He cond_resched(); 3525649920c6SJia He 3526f60858f9SMike Kravetz ret = alloc_pool_huge_page(h, nodes_allowed, 3527f60858f9SMike Kravetz node_alloc_noretry); 3528db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 35297893d1d5SAdam Litke if (!ret) 35307893d1d5SAdam Litke goto out; 35317893d1d5SAdam Litke 3532536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 3533536240f2SMel Gorman if (signal_pending(current)) 3534536240f2SMel Gorman goto out; 35357893d1d5SAdam Litke } 35367893d1d5SAdam Litke 35377893d1d5SAdam Litke /* 35387893d1d5SAdam Litke * Decrease the pool size 35397893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 35407893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 35417893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 35427893d1d5SAdam Litke * to the desired size as pages become free. 3543d1c3fb1fSNishanth Aravamudan * 3544d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 3545d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 3546d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 35473a740e8bSSidhartha Kumar * alloc_surplus_hugetlb_folio() is checking the global counter, 3548d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 3549d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 3550d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 35517893d1d5SAdam Litke */ 3552a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 35536b0c880dSAdam Litke min_count = max(count, min_count); 35546ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 355510c6ec49SMike Kravetz 355610c6ec49SMike Kravetz /* 355710c6ec49SMike Kravetz * Collect pages to be removed on list without dropping lock 355810c6ec49SMike Kravetz */ 3559a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 356010c6ec49SMike Kravetz page = remove_pool_huge_page(h, nodes_allowed, 0); 356110c6ec49SMike Kravetz if (!page) 35621da177e4SLinus Torvalds break; 356310c6ec49SMike Kravetz 356410c6ec49SMike Kravetz list_add(&page->lru, &page_list); 35651da177e4SLinus Torvalds } 356610c6ec49SMike Kravetz /* free the pages after dropping lock */ 3567db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 356810c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3569b65d4adbSMuchun Song flush_free_hpage_work(h); 3570db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 357110c6ec49SMike Kravetz 3572a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 35736ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 35747893d1d5SAdam Litke break; 35757893d1d5SAdam Litke } 35767893d1d5SAdam Litke out: 35774eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 3578db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 357929383967SMike Kravetz mutex_unlock(&h->resize_lock); 35804eb0716eSAlexandre Ghiti 3581f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 3582f60858f9SMike Kravetz 35834eb0716eSAlexandre Ghiti return 0; 35841da177e4SLinus Torvalds } 35851da177e4SLinus Torvalds 3586bdd7be07SSidhartha Kumar static int demote_free_hugetlb_folio(struct hstate *h, struct folio *folio) 35878531fc6fSMike Kravetz { 3588bdd7be07SSidhartha Kumar int i, nid = folio_nid(folio); 35898531fc6fSMike Kravetz struct hstate *target_hstate; 359031731452SDoug Berger struct page *subpage; 3591bdd7be07SSidhartha Kumar struct folio *inner_folio; 35928531fc6fSMike Kravetz int rc = 0; 35938531fc6fSMike Kravetz 35948531fc6fSMike Kravetz target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order); 35958531fc6fSMike Kravetz 3596cfd5082bSSidhartha Kumar remove_hugetlb_folio_for_demote(h, folio, false); 35978531fc6fSMike Kravetz spin_unlock_irq(&hugetlb_lock); 35988531fc6fSMike Kravetz 3599bdd7be07SSidhartha Kumar rc = hugetlb_vmemmap_restore(h, &folio->page); 36008531fc6fSMike Kravetz if (rc) { 3601bdd7be07SSidhartha Kumar /* Allocation of vmemmmap failed, we can not demote folio */ 36028531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 3603bdd7be07SSidhartha Kumar folio_ref_unfreeze(folio, 1); 3604bdd7be07SSidhartha Kumar add_hugetlb_folio(h, folio, false); 36058531fc6fSMike Kravetz return rc; 36068531fc6fSMike Kravetz } 36078531fc6fSMike Kravetz 36088531fc6fSMike Kravetz /* 3609911565b8SSidhartha Kumar * Use destroy_compound_hugetlb_folio_for_demote for all huge page 3610bdd7be07SSidhartha Kumar * sizes as it will not ref count folios. 36118531fc6fSMike Kravetz */ 3612911565b8SSidhartha Kumar destroy_compound_hugetlb_folio_for_demote(folio, huge_page_order(h)); 36138531fc6fSMike Kravetz 36148531fc6fSMike Kravetz /* 36158531fc6fSMike Kravetz * Taking target hstate mutex synchronizes with set_max_huge_pages. 36168531fc6fSMike Kravetz * Without the mutex, pages added to target hstate could be marked 36178531fc6fSMike Kravetz * as surplus. 36188531fc6fSMike Kravetz * 36198531fc6fSMike Kravetz * Note that we already hold h->resize_lock. To prevent deadlock, 36208531fc6fSMike Kravetz * use the convention of always taking larger size hstate mutex first. 36218531fc6fSMike Kravetz */ 36228531fc6fSMike Kravetz mutex_lock(&target_hstate->resize_lock); 36238531fc6fSMike Kravetz for (i = 0; i < pages_per_huge_page(h); 36248531fc6fSMike Kravetz i += pages_per_huge_page(target_hstate)) { 3625bdd7be07SSidhartha Kumar subpage = folio_page(folio, i); 3626bdd7be07SSidhartha Kumar inner_folio = page_folio(subpage); 36278531fc6fSMike Kravetz if (hstate_is_gigantic(target_hstate)) 3628bdd7be07SSidhartha Kumar prep_compound_gigantic_folio_for_demote(inner_folio, 36298531fc6fSMike Kravetz target_hstate->order); 36308531fc6fSMike Kravetz else 363131731452SDoug Berger prep_compound_page(subpage, target_hstate->order); 3632bdd7be07SSidhartha Kumar folio_change_private(inner_folio, NULL); 3633bdd7be07SSidhartha Kumar prep_new_hugetlb_folio(target_hstate, inner_folio, nid); 36342b21624fSMike Kravetz free_huge_page(subpage); 36358531fc6fSMike Kravetz } 36368531fc6fSMike Kravetz mutex_unlock(&target_hstate->resize_lock); 36378531fc6fSMike Kravetz 36388531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 36398531fc6fSMike Kravetz 36408531fc6fSMike Kravetz /* 36418531fc6fSMike Kravetz * Not absolutely necessary, but for consistency update max_huge_pages 36428531fc6fSMike Kravetz * based on pool changes for the demoted page. 36438531fc6fSMike Kravetz */ 36448531fc6fSMike Kravetz h->max_huge_pages--; 3645a43a83c7SMiaohe Lin target_hstate->max_huge_pages += 3646a43a83c7SMiaohe Lin pages_per_huge_page(h) / pages_per_huge_page(target_hstate); 36478531fc6fSMike Kravetz 36488531fc6fSMike Kravetz return rc; 36498531fc6fSMike Kravetz } 36508531fc6fSMike Kravetz 365179dfc695SMike Kravetz static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 365279dfc695SMike Kravetz __must_hold(&hugetlb_lock) 365379dfc695SMike Kravetz { 36548531fc6fSMike Kravetz int nr_nodes, node; 3655bdd7be07SSidhartha Kumar struct folio *folio; 365679dfc695SMike Kravetz 365779dfc695SMike Kravetz lockdep_assert_held(&hugetlb_lock); 365879dfc695SMike Kravetz 365979dfc695SMike Kravetz /* We should never get here if no demote order */ 366079dfc695SMike Kravetz if (!h->demote_order) { 366179dfc695SMike Kravetz pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n"); 366279dfc695SMike Kravetz return -EINVAL; /* internal error */ 366379dfc695SMike Kravetz } 366479dfc695SMike Kravetz 36658531fc6fSMike Kravetz for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 3666bdd7be07SSidhartha Kumar list_for_each_entry(folio, &h->hugepage_freelists[node], lru) { 3667bdd7be07SSidhartha Kumar if (folio_test_hwpoison(folio)) 36685a317412SMike Kravetz continue; 3669bdd7be07SSidhartha Kumar return demote_free_hugetlb_folio(h, folio); 36708531fc6fSMike Kravetz } 36718531fc6fSMike Kravetz } 36728531fc6fSMike Kravetz 36735a317412SMike Kravetz /* 36745a317412SMike Kravetz * Only way to get here is if all pages on free lists are poisoned. 36755a317412SMike Kravetz * Return -EBUSY so that caller will not retry. 36765a317412SMike Kravetz */ 36775a317412SMike Kravetz return -EBUSY; 367879dfc695SMike Kravetz } 367979dfc695SMike Kravetz 3680a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 3681a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 3682a3437870SNishanth Aravamudan 368379dfc695SMike Kravetz #define HSTATE_ATTR_WO(_name) \ 368479dfc695SMike Kravetz static struct kobj_attribute _name##_attr = __ATTR_WO(_name) 368579dfc695SMike Kravetz 3686a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 368798bc26acSMiaohe Lin static struct kobj_attribute _name##_attr = __ATTR_RW(_name) 3688a3437870SNishanth Aravamudan 3689a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 3690a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 3691a3437870SNishanth Aravamudan 36929a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 36939a305230SLee Schermerhorn 36949a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 3695a3437870SNishanth Aravamudan { 3696a3437870SNishanth Aravamudan int i; 36979a305230SLee Schermerhorn 3698a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 36999a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 37009a305230SLee Schermerhorn if (nidp) 37019a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 3702a3437870SNishanth Aravamudan return &hstates[i]; 37039a305230SLee Schermerhorn } 37049a305230SLee Schermerhorn 37059a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 3706a3437870SNishanth Aravamudan } 3707a3437870SNishanth Aravamudan 370806808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 3709a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3710a3437870SNishanth Aravamudan { 37119a305230SLee Schermerhorn struct hstate *h; 37129a305230SLee Schermerhorn unsigned long nr_huge_pages; 37139a305230SLee Schermerhorn int nid; 37149a305230SLee Schermerhorn 37159a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 37169a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 37179a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 37189a305230SLee Schermerhorn else 37199a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 37209a305230SLee Schermerhorn 3721ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", nr_huge_pages); 3722a3437870SNishanth Aravamudan } 3723adbe8726SEric B Munson 3724238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 3725238d3c13SDavid Rientjes struct hstate *h, int nid, 3726238d3c13SDavid Rientjes unsigned long count, size_t len) 3727a3437870SNishanth Aravamudan { 3728a3437870SNishanth Aravamudan int err; 37292d0adf7eSOscar Salvador nodemask_t nodes_allowed, *n_mask; 3730a3437870SNishanth Aravamudan 37312d0adf7eSOscar Salvador if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 37322d0adf7eSOscar Salvador return -EINVAL; 3733adbe8726SEric B Munson 37349a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 37359a305230SLee Schermerhorn /* 37369a305230SLee Schermerhorn * global hstate attribute 37379a305230SLee Schermerhorn */ 37389a305230SLee Schermerhorn if (!(obey_mempolicy && 37392d0adf7eSOscar Salvador init_nodemask_of_mempolicy(&nodes_allowed))) 37402d0adf7eSOscar Salvador n_mask = &node_states[N_MEMORY]; 37412d0adf7eSOscar Salvador else 37422d0adf7eSOscar Salvador n_mask = &nodes_allowed; 37432d0adf7eSOscar Salvador } else { 37449a305230SLee Schermerhorn /* 3745fd875dcaSMike Kravetz * Node specific request. count adjustment happens in 3746fd875dcaSMike Kravetz * set_max_huge_pages() after acquiring hugetlb_lock. 37479a305230SLee Schermerhorn */ 37482d0adf7eSOscar Salvador init_nodemask_of_node(&nodes_allowed, nid); 37492d0adf7eSOscar Salvador n_mask = &nodes_allowed; 3750fd875dcaSMike Kravetz } 37519a305230SLee Schermerhorn 37522d0adf7eSOscar Salvador err = set_max_huge_pages(h, count, nid, n_mask); 375306808b08SLee Schermerhorn 37544eb0716eSAlexandre Ghiti return err ? err : len; 375506808b08SLee Schermerhorn } 375606808b08SLee Schermerhorn 3757238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 3758238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 3759238d3c13SDavid Rientjes size_t len) 3760238d3c13SDavid Rientjes { 3761238d3c13SDavid Rientjes struct hstate *h; 3762238d3c13SDavid Rientjes unsigned long count; 3763238d3c13SDavid Rientjes int nid; 3764238d3c13SDavid Rientjes int err; 3765238d3c13SDavid Rientjes 3766238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 3767238d3c13SDavid Rientjes if (err) 3768238d3c13SDavid Rientjes return err; 3769238d3c13SDavid Rientjes 3770238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 3771238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 3772238d3c13SDavid Rientjes } 3773238d3c13SDavid Rientjes 377406808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 377506808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 377606808b08SLee Schermerhorn { 377706808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 377806808b08SLee Schermerhorn } 377906808b08SLee Schermerhorn 378006808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 378106808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 378206808b08SLee Schermerhorn { 3783238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 3784a3437870SNishanth Aravamudan } 3785a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 3786a3437870SNishanth Aravamudan 378706808b08SLee Schermerhorn #ifdef CONFIG_NUMA 378806808b08SLee Schermerhorn 378906808b08SLee Schermerhorn /* 379006808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 379106808b08SLee Schermerhorn * huge page alloc/free. 379206808b08SLee Schermerhorn */ 379306808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 3794ae7a927dSJoe Perches struct kobj_attribute *attr, 3795ae7a927dSJoe Perches char *buf) 379606808b08SLee Schermerhorn { 379706808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 379806808b08SLee Schermerhorn } 379906808b08SLee Schermerhorn 380006808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 380106808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 380206808b08SLee Schermerhorn { 3803238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 380406808b08SLee Schermerhorn } 380506808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 380606808b08SLee Schermerhorn #endif 380706808b08SLee Schermerhorn 380806808b08SLee Schermerhorn 3809a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 3810a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3811a3437870SNishanth Aravamudan { 38129a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3813ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); 3814a3437870SNishanth Aravamudan } 3815adbe8726SEric B Munson 3816a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 3817a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 3818a3437870SNishanth Aravamudan { 3819a3437870SNishanth Aravamudan int err; 3820a3437870SNishanth Aravamudan unsigned long input; 38219a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3822a3437870SNishanth Aravamudan 3823bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3824adbe8726SEric B Munson return -EINVAL; 3825adbe8726SEric B Munson 38263dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 3827a3437870SNishanth Aravamudan if (err) 382873ae31e5SEric B Munson return err; 3829a3437870SNishanth Aravamudan 3830db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3831a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 3832db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3833a3437870SNishanth Aravamudan 3834a3437870SNishanth Aravamudan return count; 3835a3437870SNishanth Aravamudan } 3836a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 3837a3437870SNishanth Aravamudan 3838a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 3839a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3840a3437870SNishanth Aravamudan { 38419a305230SLee Schermerhorn struct hstate *h; 38429a305230SLee Schermerhorn unsigned long free_huge_pages; 38439a305230SLee Schermerhorn int nid; 38449a305230SLee Schermerhorn 38459a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 38469a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 38479a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 38489a305230SLee Schermerhorn else 38499a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 38509a305230SLee Schermerhorn 3851ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", free_huge_pages); 3852a3437870SNishanth Aravamudan } 3853a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 3854a3437870SNishanth Aravamudan 3855a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 3856a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3857a3437870SNishanth Aravamudan { 38589a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3859ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); 3860a3437870SNishanth Aravamudan } 3861a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 3862a3437870SNishanth Aravamudan 3863a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 3864a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3865a3437870SNishanth Aravamudan { 38669a305230SLee Schermerhorn struct hstate *h; 38679a305230SLee Schermerhorn unsigned long surplus_huge_pages; 38689a305230SLee Schermerhorn int nid; 38699a305230SLee Schermerhorn 38709a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 38719a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 38729a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 38739a305230SLee Schermerhorn else 38749a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 38759a305230SLee Schermerhorn 3876ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", surplus_huge_pages); 3877a3437870SNishanth Aravamudan } 3878a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 3879a3437870SNishanth Aravamudan 388079dfc695SMike Kravetz static ssize_t demote_store(struct kobject *kobj, 388179dfc695SMike Kravetz struct kobj_attribute *attr, const char *buf, size_t len) 388279dfc695SMike Kravetz { 388379dfc695SMike Kravetz unsigned long nr_demote; 388479dfc695SMike Kravetz unsigned long nr_available; 388579dfc695SMike Kravetz nodemask_t nodes_allowed, *n_mask; 388679dfc695SMike Kravetz struct hstate *h; 38878eeda55fSLi zeming int err; 388879dfc695SMike Kravetz int nid; 388979dfc695SMike Kravetz 389079dfc695SMike Kravetz err = kstrtoul(buf, 10, &nr_demote); 389179dfc695SMike Kravetz if (err) 389279dfc695SMike Kravetz return err; 389379dfc695SMike Kravetz h = kobj_to_hstate(kobj, &nid); 389479dfc695SMike Kravetz 389579dfc695SMike Kravetz if (nid != NUMA_NO_NODE) { 389679dfc695SMike Kravetz init_nodemask_of_node(&nodes_allowed, nid); 389779dfc695SMike Kravetz n_mask = &nodes_allowed; 389879dfc695SMike Kravetz } else { 389979dfc695SMike Kravetz n_mask = &node_states[N_MEMORY]; 390079dfc695SMike Kravetz } 390179dfc695SMike Kravetz 390279dfc695SMike Kravetz /* Synchronize with other sysfs operations modifying huge pages */ 390379dfc695SMike Kravetz mutex_lock(&h->resize_lock); 390479dfc695SMike Kravetz spin_lock_irq(&hugetlb_lock); 390579dfc695SMike Kravetz 390679dfc695SMike Kravetz while (nr_demote) { 390779dfc695SMike Kravetz /* 390879dfc695SMike Kravetz * Check for available pages to demote each time thorough the 390979dfc695SMike Kravetz * loop as demote_pool_huge_page will drop hugetlb_lock. 391079dfc695SMike Kravetz */ 391179dfc695SMike Kravetz if (nid != NUMA_NO_NODE) 391279dfc695SMike Kravetz nr_available = h->free_huge_pages_node[nid]; 391379dfc695SMike Kravetz else 391479dfc695SMike Kravetz nr_available = h->free_huge_pages; 391579dfc695SMike Kravetz nr_available -= h->resv_huge_pages; 391679dfc695SMike Kravetz if (!nr_available) 391779dfc695SMike Kravetz break; 391879dfc695SMike Kravetz 391979dfc695SMike Kravetz err = demote_pool_huge_page(h, n_mask); 392079dfc695SMike Kravetz if (err) 392179dfc695SMike Kravetz break; 392279dfc695SMike Kravetz 392379dfc695SMike Kravetz nr_demote--; 392479dfc695SMike Kravetz } 392579dfc695SMike Kravetz 392679dfc695SMike Kravetz spin_unlock_irq(&hugetlb_lock); 392779dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 392879dfc695SMike Kravetz 392979dfc695SMike Kravetz if (err) 393079dfc695SMike Kravetz return err; 393179dfc695SMike Kravetz return len; 393279dfc695SMike Kravetz } 393379dfc695SMike Kravetz HSTATE_ATTR_WO(demote); 393479dfc695SMike Kravetz 393579dfc695SMike Kravetz static ssize_t demote_size_show(struct kobject *kobj, 393679dfc695SMike Kravetz struct kobj_attribute *attr, char *buf) 393779dfc695SMike Kravetz { 393812658abfSMiaohe Lin struct hstate *h = kobj_to_hstate(kobj, NULL); 393979dfc695SMike Kravetz unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K; 394079dfc695SMike Kravetz 394179dfc695SMike Kravetz return sysfs_emit(buf, "%lukB\n", demote_size); 394279dfc695SMike Kravetz } 394379dfc695SMike Kravetz 394479dfc695SMike Kravetz static ssize_t demote_size_store(struct kobject *kobj, 394579dfc695SMike Kravetz struct kobj_attribute *attr, 394679dfc695SMike Kravetz const char *buf, size_t count) 394779dfc695SMike Kravetz { 394879dfc695SMike Kravetz struct hstate *h, *demote_hstate; 394979dfc695SMike Kravetz unsigned long demote_size; 395079dfc695SMike Kravetz unsigned int demote_order; 395179dfc695SMike Kravetz 395279dfc695SMike Kravetz demote_size = (unsigned long)memparse(buf, NULL); 395379dfc695SMike Kravetz 395479dfc695SMike Kravetz demote_hstate = size_to_hstate(demote_size); 395579dfc695SMike Kravetz if (!demote_hstate) 395679dfc695SMike Kravetz return -EINVAL; 395779dfc695SMike Kravetz demote_order = demote_hstate->order; 3958a01f4390SMike Kravetz if (demote_order < HUGETLB_PAGE_ORDER) 3959a01f4390SMike Kravetz return -EINVAL; 396079dfc695SMike Kravetz 396179dfc695SMike Kravetz /* demote order must be smaller than hstate order */ 396212658abfSMiaohe Lin h = kobj_to_hstate(kobj, NULL); 396379dfc695SMike Kravetz if (demote_order >= h->order) 396479dfc695SMike Kravetz return -EINVAL; 396579dfc695SMike Kravetz 396679dfc695SMike Kravetz /* resize_lock synchronizes access to demote size and writes */ 396779dfc695SMike Kravetz mutex_lock(&h->resize_lock); 396879dfc695SMike Kravetz h->demote_order = demote_order; 396979dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 397079dfc695SMike Kravetz 397179dfc695SMike Kravetz return count; 397279dfc695SMike Kravetz } 397379dfc695SMike Kravetz HSTATE_ATTR(demote_size); 397479dfc695SMike Kravetz 3975a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 3976a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 3977a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 3978a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 3979a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 3980a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 398106808b08SLee Schermerhorn #ifdef CONFIG_NUMA 398206808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 398306808b08SLee Schermerhorn #endif 3984a3437870SNishanth Aravamudan NULL, 3985a3437870SNishanth Aravamudan }; 3986a3437870SNishanth Aravamudan 398767e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 3988a3437870SNishanth Aravamudan .attrs = hstate_attrs, 3989a3437870SNishanth Aravamudan }; 3990a3437870SNishanth Aravamudan 399179dfc695SMike Kravetz static struct attribute *hstate_demote_attrs[] = { 399279dfc695SMike Kravetz &demote_size_attr.attr, 399379dfc695SMike Kravetz &demote_attr.attr, 399479dfc695SMike Kravetz NULL, 399579dfc695SMike Kravetz }; 399679dfc695SMike Kravetz 399779dfc695SMike Kravetz static const struct attribute_group hstate_demote_attr_group = { 399879dfc695SMike Kravetz .attrs = hstate_demote_attrs, 399979dfc695SMike Kravetz }; 400079dfc695SMike Kravetz 4001094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 40029a305230SLee Schermerhorn struct kobject **hstate_kobjs, 400367e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 4004a3437870SNishanth Aravamudan { 4005a3437870SNishanth Aravamudan int retval; 4006972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 4007a3437870SNishanth Aravamudan 40089a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 40099a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 4010a3437870SNishanth Aravamudan return -ENOMEM; 4011a3437870SNishanth Aravamudan 40129a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 4013cc2205a6SMiaohe Lin if (retval) { 40149a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 4015cc2205a6SMiaohe Lin hstate_kobjs[hi] = NULL; 40163a6bdda0SMiaohe Lin return retval; 4017cc2205a6SMiaohe Lin } 4018a3437870SNishanth Aravamudan 401979dfc695SMike Kravetz if (h->demote_order) { 402001088a60SMiaohe Lin retval = sysfs_create_group(hstate_kobjs[hi], 402101088a60SMiaohe Lin &hstate_demote_attr_group); 402201088a60SMiaohe Lin if (retval) { 402379dfc695SMike Kravetz pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name); 402401088a60SMiaohe Lin sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group); 402501088a60SMiaohe Lin kobject_put(hstate_kobjs[hi]); 402601088a60SMiaohe Lin hstate_kobjs[hi] = NULL; 402701088a60SMiaohe Lin return retval; 402801088a60SMiaohe Lin } 402979dfc695SMike Kravetz } 403079dfc695SMike Kravetz 403101088a60SMiaohe Lin return 0; 4032a3437870SNishanth Aravamudan } 4033a3437870SNishanth Aravamudan 40349a305230SLee Schermerhorn #ifdef CONFIG_NUMA 4035a4a00b45SMuchun Song static bool hugetlb_sysfs_initialized __ro_after_init; 40369a305230SLee Schermerhorn 40379a305230SLee Schermerhorn /* 40389a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 403910fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 404010fbcf4cSKay Sievers * index of a node device or _hstate == node id. 404110fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 40429a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 40439a305230SLee Schermerhorn */ 40449a305230SLee Schermerhorn struct node_hstate { 40459a305230SLee Schermerhorn struct kobject *hugepages_kobj; 40469a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 40479a305230SLee Schermerhorn }; 4048b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 40499a305230SLee Schermerhorn 40509a305230SLee Schermerhorn /* 405110fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 40529a305230SLee Schermerhorn */ 40539a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 40549a305230SLee Schermerhorn &nr_hugepages_attr.attr, 40559a305230SLee Schermerhorn &free_hugepages_attr.attr, 40569a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 40579a305230SLee Schermerhorn NULL, 40589a305230SLee Schermerhorn }; 40599a305230SLee Schermerhorn 406067e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 40619a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 40629a305230SLee Schermerhorn }; 40639a305230SLee Schermerhorn 40649a305230SLee Schermerhorn /* 406510fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 40669a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 40679a305230SLee Schermerhorn */ 40689a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 40699a305230SLee Schermerhorn { 40709a305230SLee Schermerhorn int nid; 40719a305230SLee Schermerhorn 40729a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 40739a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 40749a305230SLee Schermerhorn int i; 40759a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 40769a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 40779a305230SLee Schermerhorn if (nidp) 40789a305230SLee Schermerhorn *nidp = nid; 40799a305230SLee Schermerhorn return &hstates[i]; 40809a305230SLee Schermerhorn } 40819a305230SLee Schermerhorn } 40829a305230SLee Schermerhorn 40839a305230SLee Schermerhorn BUG(); 40849a305230SLee Schermerhorn return NULL; 40859a305230SLee Schermerhorn } 40869a305230SLee Schermerhorn 40879a305230SLee Schermerhorn /* 408810fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 40899a305230SLee Schermerhorn * No-op if no hstate attributes attached. 40909a305230SLee Schermerhorn */ 4091a4a00b45SMuchun Song void hugetlb_unregister_node(struct node *node) 40929a305230SLee Schermerhorn { 40939a305230SLee Schermerhorn struct hstate *h; 409410fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 40959a305230SLee Schermerhorn 40969a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 40979b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 40989a305230SLee Schermerhorn 4099972dc4deSAneesh Kumar K.V for_each_hstate(h) { 4100972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 410101088a60SMiaohe Lin struct kobject *hstate_kobj = nhs->hstate_kobjs[idx]; 410201088a60SMiaohe Lin 410301088a60SMiaohe Lin if (!hstate_kobj) 410401088a60SMiaohe Lin continue; 410501088a60SMiaohe Lin if (h->demote_order) 410601088a60SMiaohe Lin sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group); 410701088a60SMiaohe Lin sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group); 410801088a60SMiaohe Lin kobject_put(hstate_kobj); 4109972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 4110972dc4deSAneesh Kumar K.V } 41119a305230SLee Schermerhorn 41129a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 41139a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 41149a305230SLee Schermerhorn } 41159a305230SLee Schermerhorn 41169a305230SLee Schermerhorn 41179a305230SLee Schermerhorn /* 411810fbcf4cSKay Sievers * Register hstate attributes for a single node device. 41199a305230SLee Schermerhorn * No-op if attributes already registered. 41209a305230SLee Schermerhorn */ 4121a4a00b45SMuchun Song void hugetlb_register_node(struct node *node) 41229a305230SLee Schermerhorn { 41239a305230SLee Schermerhorn struct hstate *h; 412410fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 41259a305230SLee Schermerhorn int err; 41269a305230SLee Schermerhorn 4127a4a00b45SMuchun Song if (!hugetlb_sysfs_initialized) 4128a4a00b45SMuchun Song return; 4129a4a00b45SMuchun Song 41309a305230SLee Schermerhorn if (nhs->hugepages_kobj) 41319a305230SLee Schermerhorn return; /* already allocated */ 41329a305230SLee Schermerhorn 41339a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 413410fbcf4cSKay Sievers &node->dev.kobj); 41359a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 41369a305230SLee Schermerhorn return; 41379a305230SLee Schermerhorn 41389a305230SLee Schermerhorn for_each_hstate(h) { 41399a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 41409a305230SLee Schermerhorn nhs->hstate_kobjs, 41419a305230SLee Schermerhorn &per_node_hstate_attr_group); 41429a305230SLee Schermerhorn if (err) { 4143282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s for node %d\n", 414410fbcf4cSKay Sievers h->name, node->dev.id); 41459a305230SLee Schermerhorn hugetlb_unregister_node(node); 41469a305230SLee Schermerhorn break; 41479a305230SLee Schermerhorn } 41489a305230SLee Schermerhorn } 41499a305230SLee Schermerhorn } 41509a305230SLee Schermerhorn 41519a305230SLee Schermerhorn /* 41529b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 415310fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 415410fbcf4cSKay Sievers * registered their associated device by this time. 41559a305230SLee Schermerhorn */ 41567d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 41579a305230SLee Schermerhorn { 41589a305230SLee Schermerhorn int nid; 41599a305230SLee Schermerhorn 4160a4a00b45SMuchun Song for_each_online_node(nid) 4161b958d4d0SMuchun Song hugetlb_register_node(node_devices[nid]); 41629a305230SLee Schermerhorn } 41639a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 41649a305230SLee Schermerhorn 41659a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 41669a305230SLee Schermerhorn { 41679a305230SLee Schermerhorn BUG(); 41689a305230SLee Schermerhorn if (nidp) 41699a305230SLee Schermerhorn *nidp = -1; 41709a305230SLee Schermerhorn return NULL; 41719a305230SLee Schermerhorn } 41729a305230SLee Schermerhorn 41739a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 41749a305230SLee Schermerhorn 41759a305230SLee Schermerhorn #endif 41769a305230SLee Schermerhorn 4177263b8998SMiaohe Lin #ifdef CONFIG_CMA 4178263b8998SMiaohe Lin static void __init hugetlb_cma_check(void); 4179263b8998SMiaohe Lin #else 4180263b8998SMiaohe Lin static inline __init void hugetlb_cma_check(void) 4181263b8998SMiaohe Lin { 4182263b8998SMiaohe Lin } 4183263b8998SMiaohe Lin #endif 4184263b8998SMiaohe Lin 4185a4a00b45SMuchun Song static void __init hugetlb_sysfs_init(void) 4186a4a00b45SMuchun Song { 4187a4a00b45SMuchun Song struct hstate *h; 4188a4a00b45SMuchun Song int err; 4189a4a00b45SMuchun Song 4190a4a00b45SMuchun Song hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 4191a4a00b45SMuchun Song if (!hugepages_kobj) 4192a4a00b45SMuchun Song return; 4193a4a00b45SMuchun Song 4194a4a00b45SMuchun Song for_each_hstate(h) { 4195a4a00b45SMuchun Song err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 4196a4a00b45SMuchun Song hstate_kobjs, &hstate_attr_group); 4197a4a00b45SMuchun Song if (err) 4198a4a00b45SMuchun Song pr_err("HugeTLB: Unable to add hstate %s", h->name); 4199a4a00b45SMuchun Song } 4200a4a00b45SMuchun Song 4201a4a00b45SMuchun Song #ifdef CONFIG_NUMA 4202a4a00b45SMuchun Song hugetlb_sysfs_initialized = true; 4203a4a00b45SMuchun Song #endif 4204a4a00b45SMuchun Song hugetlb_register_all_nodes(); 4205a4a00b45SMuchun Song } 4206a4a00b45SMuchun Song 4207962de548SKefeng Wang #ifdef CONFIG_SYSCTL 4208962de548SKefeng Wang static void hugetlb_sysctl_init(void); 4209962de548SKefeng Wang #else 4210962de548SKefeng Wang static inline void hugetlb_sysctl_init(void) { } 4211962de548SKefeng Wang #endif 4212962de548SKefeng Wang 4213a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 4214a3437870SNishanth Aravamudan { 42158382d914SDavidlohr Bueso int i; 42168382d914SDavidlohr Bueso 4217d6995da3SMike Kravetz BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < 4218d6995da3SMike Kravetz __NR_HPAGEFLAGS); 4219d6995da3SMike Kravetz 4220c2833a5bSMike Kravetz if (!hugepages_supported()) { 4221c2833a5bSMike Kravetz if (hugetlb_max_hstate || default_hstate_max_huge_pages) 4222c2833a5bSMike Kravetz pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); 42230ef89d25SBenjamin Herrenschmidt return 0; 4224d715cf80SLiam R. Howlett } 4225d715cf80SLiam R. Howlett 4226282f4214SMike Kravetz /* 4227282f4214SMike Kravetz * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some 4228282f4214SMike Kravetz * architectures depend on setup being done here. 4229282f4214SMike Kravetz */ 4230a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 4231282f4214SMike Kravetz if (!parsed_default_hugepagesz) { 4232282f4214SMike Kravetz /* 4233282f4214SMike Kravetz * If we did not parse a default huge page size, set 4234282f4214SMike Kravetz * default_hstate_idx to HPAGE_SIZE hstate. And, if the 4235282f4214SMike Kravetz * number of huge pages for this default size was implicitly 4236282f4214SMike Kravetz * specified, set that here as well. 4237282f4214SMike Kravetz * Note that the implicit setting will overwrite an explicit 4238282f4214SMike Kravetz * setting. A warning will be printed in this case. 4239282f4214SMike Kravetz */ 4240282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); 4241f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 4242282f4214SMike Kravetz if (default_hstate.max_huge_pages) { 4243282f4214SMike Kravetz char buf[32]; 4244282f4214SMike Kravetz 4245282f4214SMike Kravetz string_get_size(huge_page_size(&default_hstate), 4246282f4214SMike Kravetz 1, STRING_UNITS_2, buf, 32); 4247282f4214SMike Kravetz pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", 4248282f4214SMike Kravetz default_hstate.max_huge_pages, buf); 4249282f4214SMike Kravetz pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", 4250282f4214SMike Kravetz default_hstate_max_huge_pages); 4251282f4214SMike Kravetz } 4252282f4214SMike Kravetz default_hstate.max_huge_pages = 4253282f4214SMike Kravetz default_hstate_max_huge_pages; 4254b5389086SZhenguo Yao 42550a7a0f6fSPeng Liu for_each_online_node(i) 4256b5389086SZhenguo Yao default_hstate.max_huge_pages_node[i] = 4257b5389086SZhenguo Yao default_hugepages_in_node[i]; 4258282f4214SMike Kravetz } 4259f8b74815SVaishali Thakkar } 4260a3437870SNishanth Aravamudan 4261cf11e85fSRoman Gushchin hugetlb_cma_check(); 4262a3437870SNishanth Aravamudan hugetlb_init_hstates(); 4263aa888a74SAndi Kleen gather_bootmem_prealloc(); 4264a3437870SNishanth Aravamudan report_hugepages(); 4265a3437870SNishanth Aravamudan 4266a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 42677179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 4268962de548SKefeng Wang hugetlb_sysctl_init(); 42699a305230SLee Schermerhorn 42708382d914SDavidlohr Bueso #ifdef CONFIG_SMP 42718382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 42728382d914SDavidlohr Bueso #else 42738382d914SDavidlohr Bueso num_fault_mutexes = 1; 42748382d914SDavidlohr Bueso #endif 4275c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 42766da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 42776da2ec56SKees Cook GFP_KERNEL); 4278c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 42798382d914SDavidlohr Bueso 42808382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 4281c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 4282a3437870SNishanth Aravamudan return 0; 4283a3437870SNishanth Aravamudan } 42843e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 4285a3437870SNishanth Aravamudan 4286ae94da89SMike Kravetz /* Overwritten by architectures with more huge page sizes */ 4287ae94da89SMike Kravetz bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) 42889fee021dSVaishali Thakkar { 4289ae94da89SMike Kravetz return size == HPAGE_SIZE; 42909fee021dSVaishali Thakkar } 42919fee021dSVaishali Thakkar 4292d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 4293a3437870SNishanth Aravamudan { 4294a3437870SNishanth Aravamudan struct hstate *h; 42958faa8b07SAndi Kleen unsigned long i; 42968faa8b07SAndi Kleen 4297a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 4298a3437870SNishanth Aravamudan return; 4299a3437870SNishanth Aravamudan } 430047d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 4301a3437870SNishanth Aravamudan BUG_ON(order == 0); 430247d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 430329383967SMike Kravetz mutex_init(&h->resize_lock); 4304a3437870SNishanth Aravamudan h->order = order; 4305aca78307SMiaohe Lin h->mask = ~(huge_page_size(h) - 1); 43068faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 43078faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 43080edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 430954f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 431054f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 4311a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 4312c2c3a60aSMiaohe Lin huge_page_size(h)/SZ_1K); 43138faa8b07SAndi Kleen 4314a3437870SNishanth Aravamudan parsed_hstate = h; 4315a3437870SNishanth Aravamudan } 4316a3437870SNishanth Aravamudan 4317b5389086SZhenguo Yao bool __init __weak hugetlb_node_alloc_supported(void) 4318b5389086SZhenguo Yao { 4319b5389086SZhenguo Yao return true; 4320b5389086SZhenguo Yao } 4321f87442f4SPeng Liu 4322f87442f4SPeng Liu static void __init hugepages_clear_pages_in_node(void) 4323f87442f4SPeng Liu { 4324f87442f4SPeng Liu if (!hugetlb_max_hstate) { 4325f87442f4SPeng Liu default_hstate_max_huge_pages = 0; 4326f87442f4SPeng Liu memset(default_hugepages_in_node, 0, 432710395680SMiaohe Lin sizeof(default_hugepages_in_node)); 4328f87442f4SPeng Liu } else { 4329f87442f4SPeng Liu parsed_hstate->max_huge_pages = 0; 4330f87442f4SPeng Liu memset(parsed_hstate->max_huge_pages_node, 0, 433110395680SMiaohe Lin sizeof(parsed_hstate->max_huge_pages_node)); 4332f87442f4SPeng Liu } 4333f87442f4SPeng Liu } 4334f87442f4SPeng Liu 4335282f4214SMike Kravetz /* 4336282f4214SMike Kravetz * hugepages command line processing 4337282f4214SMike Kravetz * hugepages normally follows a valid hugepagsz or default_hugepagsz 4338282f4214SMike Kravetz * specification. If not, ignore the hugepages value. hugepages can also 4339282f4214SMike Kravetz * be the first huge page command line option in which case it implicitly 4340282f4214SMike Kravetz * specifies the number of huge pages for the default size. 4341282f4214SMike Kravetz */ 4342282f4214SMike Kravetz static int __init hugepages_setup(char *s) 4343a3437870SNishanth Aravamudan { 4344a3437870SNishanth Aravamudan unsigned long *mhp; 43458faa8b07SAndi Kleen static unsigned long *last_mhp; 4346b5389086SZhenguo Yao int node = NUMA_NO_NODE; 4347b5389086SZhenguo Yao int count; 4348b5389086SZhenguo Yao unsigned long tmp; 4349b5389086SZhenguo Yao char *p = s; 4350a3437870SNishanth Aravamudan 43519fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 4352282f4214SMike Kravetz pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); 43539fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 4354f81f6e4bSPeng Liu return 1; 43559fee021dSVaishali Thakkar } 4356282f4214SMike Kravetz 4357a3437870SNishanth Aravamudan /* 4358282f4214SMike Kravetz * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter 4359282f4214SMike Kravetz * yet, so this hugepages= parameter goes to the "default hstate". 4360282f4214SMike Kravetz * Otherwise, it goes with the previously parsed hugepagesz or 4361282f4214SMike Kravetz * default_hugepagesz. 4362a3437870SNishanth Aravamudan */ 43639fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 4364a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 4365a3437870SNishanth Aravamudan else 4366a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 4367a3437870SNishanth Aravamudan 43688faa8b07SAndi Kleen if (mhp == last_mhp) { 4369282f4214SMike Kravetz pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); 4370f81f6e4bSPeng Liu return 1; 43718faa8b07SAndi Kleen } 43728faa8b07SAndi Kleen 4373b5389086SZhenguo Yao while (*p) { 4374b5389086SZhenguo Yao count = 0; 4375b5389086SZhenguo Yao if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4376b5389086SZhenguo Yao goto invalid; 4377b5389086SZhenguo Yao /* Parameter is node format */ 4378b5389086SZhenguo Yao if (p[count] == ':') { 4379b5389086SZhenguo Yao if (!hugetlb_node_alloc_supported()) { 4380b5389086SZhenguo Yao pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n"); 4381f81f6e4bSPeng Liu return 1; 4382b5389086SZhenguo Yao } 43830a7a0f6fSPeng Liu if (tmp >= MAX_NUMNODES || !node_online(tmp)) 4384e79ce983SLiu Yuntao goto invalid; 43850a7a0f6fSPeng Liu node = array_index_nospec(tmp, MAX_NUMNODES); 4386b5389086SZhenguo Yao p += count + 1; 4387b5389086SZhenguo Yao /* Parse hugepages */ 4388b5389086SZhenguo Yao if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4389b5389086SZhenguo Yao goto invalid; 4390b5389086SZhenguo Yao if (!hugetlb_max_hstate) 4391b5389086SZhenguo Yao default_hugepages_in_node[node] = tmp; 4392b5389086SZhenguo Yao else 4393b5389086SZhenguo Yao parsed_hstate->max_huge_pages_node[node] = tmp; 4394b5389086SZhenguo Yao *mhp += tmp; 4395b5389086SZhenguo Yao /* Go to parse next node*/ 4396b5389086SZhenguo Yao if (p[count] == ',') 4397b5389086SZhenguo Yao p += count + 1; 4398b5389086SZhenguo Yao else 4399b5389086SZhenguo Yao break; 4400b5389086SZhenguo Yao } else { 4401b5389086SZhenguo Yao if (p != s) 4402b5389086SZhenguo Yao goto invalid; 4403b5389086SZhenguo Yao *mhp = tmp; 4404b5389086SZhenguo Yao break; 4405b5389086SZhenguo Yao } 4406b5389086SZhenguo Yao } 4407a3437870SNishanth Aravamudan 44088faa8b07SAndi Kleen /* 44098faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 441004adbc3fSMiaohe Lin * But we need to allocate gigantic hstates here early to still 44118faa8b07SAndi Kleen * use the bootmem allocator. 44128faa8b07SAndi Kleen */ 441304adbc3fSMiaohe Lin if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) 44148faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 44158faa8b07SAndi Kleen 44168faa8b07SAndi Kleen last_mhp = mhp; 44178faa8b07SAndi Kleen 4418a3437870SNishanth Aravamudan return 1; 4419b5389086SZhenguo Yao 4420b5389086SZhenguo Yao invalid: 4421b5389086SZhenguo Yao pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p); 4422f87442f4SPeng Liu hugepages_clear_pages_in_node(); 4423f81f6e4bSPeng Liu return 1; 4424a3437870SNishanth Aravamudan } 4425282f4214SMike Kravetz __setup("hugepages=", hugepages_setup); 4426e11bfbfcSNick Piggin 4427282f4214SMike Kravetz /* 4428282f4214SMike Kravetz * hugepagesz command line processing 4429282f4214SMike Kravetz * A specific huge page size can only be specified once with hugepagesz. 4430282f4214SMike Kravetz * hugepagesz is followed by hugepages on the command line. The global 4431282f4214SMike Kravetz * variable 'parsed_valid_hugepagesz' is used to determine if prior 4432282f4214SMike Kravetz * hugepagesz argument was valid. 4433282f4214SMike Kravetz */ 4434359f2544SMike Kravetz static int __init hugepagesz_setup(char *s) 4435e11bfbfcSNick Piggin { 4436359f2544SMike Kravetz unsigned long size; 4437282f4214SMike Kravetz struct hstate *h; 4438282f4214SMike Kravetz 4439282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4440359f2544SMike Kravetz size = (unsigned long)memparse(s, NULL); 4441359f2544SMike Kravetz 4442359f2544SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4443282f4214SMike Kravetz pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); 4444f81f6e4bSPeng Liu return 1; 4445359f2544SMike Kravetz } 4446359f2544SMike Kravetz 4447282f4214SMike Kravetz h = size_to_hstate(size); 4448282f4214SMike Kravetz if (h) { 4449282f4214SMike Kravetz /* 4450282f4214SMike Kravetz * hstate for this size already exists. This is normally 4451282f4214SMike Kravetz * an error, but is allowed if the existing hstate is the 4452282f4214SMike Kravetz * default hstate. More specifically, it is only allowed if 4453282f4214SMike Kravetz * the number of huge pages for the default hstate was not 4454282f4214SMike Kravetz * previously specified. 4455282f4214SMike Kravetz */ 4456282f4214SMike Kravetz if (!parsed_default_hugepagesz || h != &default_hstate || 4457282f4214SMike Kravetz default_hstate.max_huge_pages) { 4458282f4214SMike Kravetz pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); 4459f81f6e4bSPeng Liu return 1; 446038237830SMike Kravetz } 446138237830SMike Kravetz 4462282f4214SMike Kravetz /* 4463282f4214SMike Kravetz * No need to call hugetlb_add_hstate() as hstate already 4464282f4214SMike Kravetz * exists. But, do set parsed_hstate so that a following 4465282f4214SMike Kravetz * hugepages= parameter will be applied to this hstate. 4466282f4214SMike Kravetz */ 4467282f4214SMike Kravetz parsed_hstate = h; 4468282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4469e11bfbfcSNick Piggin return 1; 4470e11bfbfcSNick Piggin } 4471282f4214SMike Kravetz 4472359f2544SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4473282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4474359f2544SMike Kravetz return 1; 4475359f2544SMike Kravetz } 4476359f2544SMike Kravetz __setup("hugepagesz=", hugepagesz_setup); 4477359f2544SMike Kravetz 4478282f4214SMike Kravetz /* 4479282f4214SMike Kravetz * default_hugepagesz command line input 4480282f4214SMike Kravetz * Only one instance of default_hugepagesz allowed on command line. 4481282f4214SMike Kravetz */ 4482ae94da89SMike Kravetz static int __init default_hugepagesz_setup(char *s) 4483e11bfbfcSNick Piggin { 4484ae94da89SMike Kravetz unsigned long size; 4485b5389086SZhenguo Yao int i; 4486ae94da89SMike Kravetz 4487282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4488282f4214SMike Kravetz if (parsed_default_hugepagesz) { 4489282f4214SMike Kravetz pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); 4490f81f6e4bSPeng Liu return 1; 4491282f4214SMike Kravetz } 4492282f4214SMike Kravetz 4493282f4214SMike Kravetz size = (unsigned long)memparse(s, NULL); 4494282f4214SMike Kravetz 4495282f4214SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4496282f4214SMike Kravetz pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); 4497f81f6e4bSPeng Liu return 1; 4498282f4214SMike Kravetz } 4499282f4214SMike Kravetz 4500282f4214SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4501282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4502282f4214SMike Kravetz parsed_default_hugepagesz = true; 4503282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(size)); 4504282f4214SMike Kravetz 4505282f4214SMike Kravetz /* 4506282f4214SMike Kravetz * The number of default huge pages (for this size) could have been 4507282f4214SMike Kravetz * specified as the first hugetlb parameter: hugepages=X. If so, 4508282f4214SMike Kravetz * then default_hstate_max_huge_pages is set. If the default huge 450923baf831SKirill A. Shutemov * page size is gigantic (> MAX_ORDER), then the pages must be 4510282f4214SMike Kravetz * allocated here from bootmem allocator. 4511282f4214SMike Kravetz */ 4512282f4214SMike Kravetz if (default_hstate_max_huge_pages) { 4513282f4214SMike Kravetz default_hstate.max_huge_pages = default_hstate_max_huge_pages; 45140a7a0f6fSPeng Liu for_each_online_node(i) 4515b5389086SZhenguo Yao default_hstate.max_huge_pages_node[i] = 4516b5389086SZhenguo Yao default_hugepages_in_node[i]; 4517282f4214SMike Kravetz if (hstate_is_gigantic(&default_hstate)) 4518282f4214SMike Kravetz hugetlb_hstate_alloc_pages(&default_hstate); 4519282f4214SMike Kravetz default_hstate_max_huge_pages = 0; 4520282f4214SMike Kravetz } 4521282f4214SMike Kravetz 4522e11bfbfcSNick Piggin return 1; 4523e11bfbfcSNick Piggin } 4524ae94da89SMike Kravetz __setup("default_hugepagesz=", default_hugepagesz_setup); 4525a3437870SNishanth Aravamudan 4526d2226ebdSFeng Tang static nodemask_t *policy_mbind_nodemask(gfp_t gfp) 4527d2226ebdSFeng Tang { 4528d2226ebdSFeng Tang #ifdef CONFIG_NUMA 4529d2226ebdSFeng Tang struct mempolicy *mpol = get_task_policy(current); 4530d2226ebdSFeng Tang 4531d2226ebdSFeng Tang /* 4532d2226ebdSFeng Tang * Only enforce MPOL_BIND policy which overlaps with cpuset policy 4533d2226ebdSFeng Tang * (from policy_nodemask) specifically for hugetlb case 4534d2226ebdSFeng Tang */ 4535d2226ebdSFeng Tang if (mpol->mode == MPOL_BIND && 4536d2226ebdSFeng Tang (apply_policy_zone(mpol, gfp_zone(gfp)) && 4537d2226ebdSFeng Tang cpuset_nodemask_valid_mems_allowed(&mpol->nodes))) 4538d2226ebdSFeng Tang return &mpol->nodes; 4539d2226ebdSFeng Tang #endif 4540d2226ebdSFeng Tang return NULL; 4541d2226ebdSFeng Tang } 4542d2226ebdSFeng Tang 45438ca39e68SMuchun Song static unsigned int allowed_mems_nr(struct hstate *h) 45448a213460SNishanth Aravamudan { 45458a213460SNishanth Aravamudan int node; 45468a213460SNishanth Aravamudan unsigned int nr = 0; 4547d2226ebdSFeng Tang nodemask_t *mbind_nodemask; 45488ca39e68SMuchun Song unsigned int *array = h->free_huge_pages_node; 45498ca39e68SMuchun Song gfp_t gfp_mask = htlb_alloc_mask(h); 45508a213460SNishanth Aravamudan 4551d2226ebdSFeng Tang mbind_nodemask = policy_mbind_nodemask(gfp_mask); 45528ca39e68SMuchun Song for_each_node_mask(node, cpuset_current_mems_allowed) { 4553d2226ebdSFeng Tang if (!mbind_nodemask || node_isset(node, *mbind_nodemask)) 45548a213460SNishanth Aravamudan nr += array[node]; 45558ca39e68SMuchun Song } 45568a213460SNishanth Aravamudan 45578a213460SNishanth Aravamudan return nr; 45588a213460SNishanth Aravamudan } 45598a213460SNishanth Aravamudan 45608a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 456117743798SMuchun Song static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, 456217743798SMuchun Song void *buffer, size_t *length, 456317743798SMuchun Song loff_t *ppos, unsigned long *out) 456417743798SMuchun Song { 456517743798SMuchun Song struct ctl_table dup_table; 456617743798SMuchun Song 456717743798SMuchun Song /* 456817743798SMuchun Song * In order to avoid races with __do_proc_doulongvec_minmax(), we 456917743798SMuchun Song * can duplicate the @table and alter the duplicate of it. 457017743798SMuchun Song */ 457117743798SMuchun Song dup_table = *table; 457217743798SMuchun Song dup_table.data = out; 457317743798SMuchun Song 457417743798SMuchun Song return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); 457517743798SMuchun Song } 457617743798SMuchun Song 457706808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 457806808b08SLee Schermerhorn struct ctl_table *table, int write, 457932927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 45801da177e4SLinus Torvalds { 4581e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 4582238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 458308d4a246SMichal Hocko int ret; 4584e5ff2159SAndi Kleen 4585457c1b27SNishanth Aravamudan if (!hugepages_supported()) 458686613628SJan Stancek return -EOPNOTSUPP; 4587457c1b27SNishanth Aravamudan 458817743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 458917743798SMuchun Song &tmp); 459008d4a246SMichal Hocko if (ret) 459108d4a246SMichal Hocko goto out; 4592e5ff2159SAndi Kleen 4593238d3c13SDavid Rientjes if (write) 4594238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 4595238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 459608d4a246SMichal Hocko out: 459708d4a246SMichal Hocko return ret; 45981da177e4SLinus Torvalds } 4599396faf03SMel Gorman 4600962de548SKefeng Wang static int hugetlb_sysctl_handler(struct ctl_table *table, int write, 460132927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 460206808b08SLee Schermerhorn { 460306808b08SLee Schermerhorn 460406808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 460506808b08SLee Schermerhorn buffer, length, ppos); 460606808b08SLee Schermerhorn } 460706808b08SLee Schermerhorn 460806808b08SLee Schermerhorn #ifdef CONFIG_NUMA 4609962de548SKefeng Wang static int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 461032927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 461106808b08SLee Schermerhorn { 461206808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 461306808b08SLee Schermerhorn buffer, length, ppos); 461406808b08SLee Schermerhorn } 461506808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 461606808b08SLee Schermerhorn 4617962de548SKefeng Wang static int hugetlb_overcommit_handler(struct ctl_table *table, int write, 461832927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 4619a3d0c6aaSNishanth Aravamudan { 4620a5516438SAndi Kleen struct hstate *h = &default_hstate; 4621e5ff2159SAndi Kleen unsigned long tmp; 462208d4a246SMichal Hocko int ret; 4623e5ff2159SAndi Kleen 4624457c1b27SNishanth Aravamudan if (!hugepages_supported()) 462586613628SJan Stancek return -EOPNOTSUPP; 4626457c1b27SNishanth Aravamudan 4627e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 4628e5ff2159SAndi Kleen 4629bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 4630adbe8726SEric B Munson return -EINVAL; 4631adbe8726SEric B Munson 463217743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 463317743798SMuchun Song &tmp); 463408d4a246SMichal Hocko if (ret) 463508d4a246SMichal Hocko goto out; 4636e5ff2159SAndi Kleen 4637e5ff2159SAndi Kleen if (write) { 4638db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4639e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 4640db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4641e5ff2159SAndi Kleen } 464208d4a246SMichal Hocko out: 464308d4a246SMichal Hocko return ret; 4644a3d0c6aaSNishanth Aravamudan } 4645a3d0c6aaSNishanth Aravamudan 4646962de548SKefeng Wang static struct ctl_table hugetlb_table[] = { 4647962de548SKefeng Wang { 4648962de548SKefeng Wang .procname = "nr_hugepages", 4649962de548SKefeng Wang .data = NULL, 4650962de548SKefeng Wang .maxlen = sizeof(unsigned long), 4651962de548SKefeng Wang .mode = 0644, 4652962de548SKefeng Wang .proc_handler = hugetlb_sysctl_handler, 4653962de548SKefeng Wang }, 4654962de548SKefeng Wang #ifdef CONFIG_NUMA 4655962de548SKefeng Wang { 4656962de548SKefeng Wang .procname = "nr_hugepages_mempolicy", 4657962de548SKefeng Wang .data = NULL, 4658962de548SKefeng Wang .maxlen = sizeof(unsigned long), 4659962de548SKefeng Wang .mode = 0644, 4660962de548SKefeng Wang .proc_handler = &hugetlb_mempolicy_sysctl_handler, 4661962de548SKefeng Wang }, 4662962de548SKefeng Wang #endif 4663962de548SKefeng Wang { 4664962de548SKefeng Wang .procname = "hugetlb_shm_group", 4665962de548SKefeng Wang .data = &sysctl_hugetlb_shm_group, 4666962de548SKefeng Wang .maxlen = sizeof(gid_t), 4667962de548SKefeng Wang .mode = 0644, 4668962de548SKefeng Wang .proc_handler = proc_dointvec, 4669962de548SKefeng Wang }, 4670962de548SKefeng Wang { 4671962de548SKefeng Wang .procname = "nr_overcommit_hugepages", 4672962de548SKefeng Wang .data = NULL, 4673962de548SKefeng Wang .maxlen = sizeof(unsigned long), 4674962de548SKefeng Wang .mode = 0644, 4675962de548SKefeng Wang .proc_handler = hugetlb_overcommit_handler, 4676962de548SKefeng Wang }, 4677962de548SKefeng Wang { } 4678962de548SKefeng Wang }; 4679962de548SKefeng Wang 4680962de548SKefeng Wang static void hugetlb_sysctl_init(void) 4681962de548SKefeng Wang { 4682962de548SKefeng Wang register_sysctl_init("vm", hugetlb_table); 4683962de548SKefeng Wang } 46841da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 46851da177e4SLinus Torvalds 4686e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 46871da177e4SLinus Torvalds { 4688fcb2b0c5SRoman Gushchin struct hstate *h; 4689fcb2b0c5SRoman Gushchin unsigned long total = 0; 4690fcb2b0c5SRoman Gushchin 4691457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4692457c1b27SNishanth Aravamudan return; 4693fcb2b0c5SRoman Gushchin 4694fcb2b0c5SRoman Gushchin for_each_hstate(h) { 4695fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 4696fcb2b0c5SRoman Gushchin 4697aca78307SMiaohe Lin total += huge_page_size(h) * count; 4698fcb2b0c5SRoman Gushchin 4699fcb2b0c5SRoman Gushchin if (h == &default_hstate) 4700e1759c21SAlexey Dobriyan seq_printf(m, 47011da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 47021da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 4703b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 47047893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 47054f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 4706fcb2b0c5SRoman Gushchin count, 4707a5516438SAndi Kleen h->free_huge_pages, 4708a5516438SAndi Kleen h->resv_huge_pages, 4709a5516438SAndi Kleen h->surplus_huge_pages, 4710aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4711fcb2b0c5SRoman Gushchin } 4712fcb2b0c5SRoman Gushchin 4713aca78307SMiaohe Lin seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); 47141da177e4SLinus Torvalds } 47151da177e4SLinus Torvalds 47167981593bSJoe Perches int hugetlb_report_node_meminfo(char *buf, int len, int nid) 47171da177e4SLinus Torvalds { 4718a5516438SAndi Kleen struct hstate *h = &default_hstate; 47197981593bSJoe Perches 4720457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4721457c1b27SNishanth Aravamudan return 0; 47227981593bSJoe Perches 47237981593bSJoe Perches return sysfs_emit_at(buf, len, 47241da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 4725a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 4726a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 4727a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 4728a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 4729a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 47301da177e4SLinus Torvalds } 47311da177e4SLinus Torvalds 4732dcadcf1cSGang Li void hugetlb_show_meminfo_node(int nid) 4733949f7ec5SDavid Rientjes { 4734949f7ec5SDavid Rientjes struct hstate *h; 4735949f7ec5SDavid Rientjes 4736457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4737457c1b27SNishanth Aravamudan return; 4738457c1b27SNishanth Aravamudan 4739949f7ec5SDavid Rientjes for_each_hstate(h) 4740dcadcf1cSGang Li printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 4741949f7ec5SDavid Rientjes nid, 4742949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 4743949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 4744949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 4745aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4746949f7ec5SDavid Rientjes } 4747949f7ec5SDavid Rientjes 47485d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 47495d317b2bSNaoya Horiguchi { 47505d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 47515d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 47525d317b2bSNaoya Horiguchi } 47535d317b2bSNaoya Horiguchi 47541da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 47551da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 47561da177e4SLinus Torvalds { 4757d0028588SWanpeng Li struct hstate *h; 4758d0028588SWanpeng Li unsigned long nr_total_pages = 0; 4759d0028588SWanpeng Li 4760d0028588SWanpeng Li for_each_hstate(h) 4761d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 4762d0028588SWanpeng Li return nr_total_pages; 47631da177e4SLinus Torvalds } 47641da177e4SLinus Torvalds 4765a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 4766fc1b8a73SMel Gorman { 4767fc1b8a73SMel Gorman int ret = -ENOMEM; 4768fc1b8a73SMel Gorman 47690aa7f354SMiaohe Lin if (!delta) 47700aa7f354SMiaohe Lin return 0; 47710aa7f354SMiaohe Lin 4772db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4773fc1b8a73SMel Gorman /* 4774fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 4775fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 4776fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 4777fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 4778fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 4779fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 4780fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 4781fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 4782fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 4783fc1b8a73SMel Gorman * 4784fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 4785fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 4786fc1b8a73SMel Gorman * we fall back to check against current free page availability as 4787fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 4788fc1b8a73SMel Gorman * semantics that cpuset has. 47898ca39e68SMuchun Song * 47908ca39e68SMuchun Song * Apart from cpuset, we also have memory policy mechanism that 47918ca39e68SMuchun Song * also determines from which node the kernel will allocate memory 47928ca39e68SMuchun Song * in a NUMA system. So similar to cpuset, we also should consider 47938ca39e68SMuchun Song * the memory policy of the current task. Similar to the description 47948ca39e68SMuchun Song * above. 4795fc1b8a73SMel Gorman */ 4796fc1b8a73SMel Gorman if (delta > 0) { 4797a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 4798fc1b8a73SMel Gorman goto out; 4799fc1b8a73SMel Gorman 48008ca39e68SMuchun Song if (delta > allowed_mems_nr(h)) { 4801a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 4802fc1b8a73SMel Gorman goto out; 4803fc1b8a73SMel Gorman } 4804fc1b8a73SMel Gorman } 4805fc1b8a73SMel Gorman 4806fc1b8a73SMel Gorman ret = 0; 4807fc1b8a73SMel Gorman if (delta < 0) 4808a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 4809fc1b8a73SMel Gorman 4810fc1b8a73SMel Gorman out: 4811db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4812fc1b8a73SMel Gorman return ret; 4813fc1b8a73SMel Gorman } 4814fc1b8a73SMel Gorman 481584afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 481684afd99bSAndy Whitcroft { 4817f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 481884afd99bSAndy Whitcroft 481984afd99bSAndy Whitcroft /* 4820612b8a31SMike Kravetz * HPAGE_RESV_OWNER indicates a private mapping. 482184afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 482284afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 482384afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 482425985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 482584afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 482684afd99bSAndy Whitcroft * new reference here without additional locking. 482784afd99bSAndy Whitcroft */ 482809a26e83SMike Kravetz if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 482909a26e83SMike Kravetz resv_map_dup_hugetlb_cgroup_uncharge_info(resv); 4830f522c3acSJoonsoo Kim kref_get(&resv->refs); 483184afd99bSAndy Whitcroft } 48328d9bfb26SMike Kravetz 4833131a79b4SMike Kravetz /* 4834131a79b4SMike Kravetz * vma_lock structure for sharable mappings is vma specific. 4835612b8a31SMike Kravetz * Clear old pointer (if copied via vm_area_dup) and allocate 4836612b8a31SMike Kravetz * new structure. Before clearing, make sure vma_lock is not 4837612b8a31SMike Kravetz * for this vma. 4838131a79b4SMike Kravetz */ 4839131a79b4SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 4840612b8a31SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 4841612b8a31SMike Kravetz 4842612b8a31SMike Kravetz if (vma_lock) { 4843612b8a31SMike Kravetz if (vma_lock->vma != vma) { 4844131a79b4SMike Kravetz vma->vm_private_data = NULL; 48458d9bfb26SMike Kravetz hugetlb_vma_lock_alloc(vma); 4846612b8a31SMike Kravetz } else 4847612b8a31SMike Kravetz pr_warn("HugeTLB: vma_lock already exists in %s.\n", __func__); 4848612b8a31SMike Kravetz } else 4849612b8a31SMike Kravetz hugetlb_vma_lock_alloc(vma); 485009a26e83SMike Kravetz } 4851131a79b4SMike Kravetz } 485284afd99bSAndy Whitcroft 4853a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 4854a1e78772SMel Gorman { 4855a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 48568d9bfb26SMike Kravetz struct resv_map *resv; 485790481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 48584e35f483SJoonsoo Kim unsigned long reserve, start, end; 48591c5ecae3SMike Kravetz long gbl_reserve; 486084afd99bSAndy Whitcroft 48618d9bfb26SMike Kravetz hugetlb_vma_lock_free(vma); 48628d9bfb26SMike Kravetz 48638d9bfb26SMike Kravetz resv = vma_resv_map(vma); 48644e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 48654e35f483SJoonsoo Kim return; 48664e35f483SJoonsoo Kim 4867a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 4868a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 486984afd99bSAndy Whitcroft 48704e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 4871e9fe92aeSMina Almasry hugetlb_cgroup_uncharge_counter(resv, start, end); 48727251ff78SAdam Litke if (reserve) { 48731c5ecae3SMike Kravetz /* 48741c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 48751c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 48761c5ecae3SMike Kravetz */ 48771c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 48781c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 48797251ff78SAdam Litke } 4880e9fe92aeSMina Almasry 4881e9fe92aeSMina Almasry kref_put(&resv->refs, resv_map_release); 4882a1e78772SMel Gorman } 4883a1e78772SMel Gorman 488431383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 488531383c68SDan Williams { 488631383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 488731383c68SDan Williams return -EINVAL; 4888b30c14cdSJames Houghton 4889b30c14cdSJames Houghton /* 4890b30c14cdSJames Houghton * PMD sharing is only possible for PUD_SIZE-aligned address ranges 4891b30c14cdSJames Houghton * in HugeTLB VMAs. If we will lose PUD_SIZE alignment due to this 4892b30c14cdSJames Houghton * split, unshare PMDs in the PUD_SIZE interval surrounding addr now. 4893b30c14cdSJames Houghton */ 4894b30c14cdSJames Houghton if (addr & ~PUD_MASK) { 4895b30c14cdSJames Houghton /* 4896b30c14cdSJames Houghton * hugetlb_vm_op_split is called right before we attempt to 4897b30c14cdSJames Houghton * split the VMA. We will need to unshare PMDs in the old and 4898b30c14cdSJames Houghton * new VMAs, so let's unshare before we split. 4899b30c14cdSJames Houghton */ 4900b30c14cdSJames Houghton unsigned long floor = addr & PUD_MASK; 4901b30c14cdSJames Houghton unsigned long ceil = floor + PUD_SIZE; 4902b30c14cdSJames Houghton 4903b30c14cdSJames Houghton if (floor >= vma->vm_start && ceil <= vma->vm_end) 4904b30c14cdSJames Houghton hugetlb_unshare_pmds(vma, floor, ceil); 4905b30c14cdSJames Houghton } 4906b30c14cdSJames Houghton 490731383c68SDan Williams return 0; 490831383c68SDan Williams } 490931383c68SDan Williams 491005ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 491105ea8860SDan Williams { 4912aca78307SMiaohe Lin return huge_page_size(hstate_vma(vma)); 491305ea8860SDan Williams } 491405ea8860SDan Williams 49151da177e4SLinus Torvalds /* 49161da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 49171da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 49186c26d310SMiaohe Lin * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get 49191da177e4SLinus Torvalds * this far. 49201da177e4SLinus Torvalds */ 4921b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 49221da177e4SLinus Torvalds { 49231da177e4SLinus Torvalds BUG(); 4924d0217ac0SNick Piggin return 0; 49251da177e4SLinus Torvalds } 49261da177e4SLinus Torvalds 4927eec3636aSJane Chu /* 4928eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 4929eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 4930eec3636aSJane Chu * This is because under System V memory model, mappings created via 4931eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 4932eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 4933eec3636aSJane Chu */ 4934f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 4935d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 493684afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 4937a1e78772SMel Gorman .close = hugetlb_vm_op_close, 4938dd3b614fSDmitry Safonov .may_split = hugetlb_vm_op_split, 493905ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 49401da177e4SLinus Torvalds }; 49411da177e4SLinus Torvalds 49421e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 49431e8f889bSDavid Gibson int writable) 494463551ae0SDavid Gibson { 494563551ae0SDavid Gibson pte_t entry; 494679c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 494763551ae0SDavid Gibson 49481e8f889bSDavid Gibson if (writable) { 4949106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 4950106c992aSGerald Schaefer vma->vm_page_prot))); 495163551ae0SDavid Gibson } else { 4952106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 4953106c992aSGerald Schaefer vma->vm_page_prot)); 495463551ae0SDavid Gibson } 495563551ae0SDavid Gibson entry = pte_mkyoung(entry); 495679c1c594SChristophe Leroy entry = arch_make_huge_pte(entry, shift, vma->vm_flags); 495763551ae0SDavid Gibson 495863551ae0SDavid Gibson return entry; 495963551ae0SDavid Gibson } 496063551ae0SDavid Gibson 49611e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 49621e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 49631e8f889bSDavid Gibson { 49641e8f889bSDavid Gibson pte_t entry; 49651e8f889bSDavid Gibson 4966106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 496732f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 49684b3073e1SRussell King update_mmu_cache(vma, address, ptep); 49691e8f889bSDavid Gibson } 49701e8f889bSDavid Gibson 4971d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 49724a705fefSNaoya Horiguchi { 49734a705fefSNaoya Horiguchi swp_entry_t swp; 49744a705fefSNaoya Horiguchi 49754a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 4976d5ed7444SAneesh Kumar K.V return false; 49774a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4978d79d176aSBaoquan He if (is_migration_entry(swp)) 4979d5ed7444SAneesh Kumar K.V return true; 49804a705fefSNaoya Horiguchi else 4981d5ed7444SAneesh Kumar K.V return false; 49824a705fefSNaoya Horiguchi } 49834a705fefSNaoya Horiguchi 49843e5c3600SBaoquan He static bool is_hugetlb_entry_hwpoisoned(pte_t pte) 49854a705fefSNaoya Horiguchi { 49864a705fefSNaoya Horiguchi swp_entry_t swp; 49874a705fefSNaoya Horiguchi 49884a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 49893e5c3600SBaoquan He return false; 49904a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4991d79d176aSBaoquan He if (is_hwpoison_entry(swp)) 49923e5c3600SBaoquan He return true; 49934a705fefSNaoya Horiguchi else 49943e5c3600SBaoquan He return false; 49954a705fefSNaoya Horiguchi } 49961e8f889bSDavid Gibson 49974eae4efaSPeter Xu static void 4998ea4c353dSSidhartha Kumar hugetlb_install_folio(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, 49995a2f8d22SPeter Xu struct folio *new_folio, pte_t old) 50004eae4efaSPeter Xu { 50015a2f8d22SPeter Xu pte_t newpte = make_huge_pte(vma, &new_folio->page, 1); 50025a2f8d22SPeter Xu 5003ea4c353dSSidhartha Kumar __folio_mark_uptodate(new_folio); 5004d0ce0e47SSidhartha Kumar hugepage_add_new_anon_rmap(new_folio, vma, addr); 50055a2f8d22SPeter Xu if (userfaultfd_wp(vma) && huge_pte_uffd_wp(old)) 50065a2f8d22SPeter Xu newpte = huge_pte_mkuffd_wp(newpte); 50075a2f8d22SPeter Xu set_huge_pte_at(vma->vm_mm, addr, ptep, newpte); 50084eae4efaSPeter Xu hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); 5009ea4c353dSSidhartha Kumar folio_set_hugetlb_migratable(new_folio); 50104eae4efaSPeter Xu } 50114eae4efaSPeter Xu 501263551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 5013bc70fbf2SPeter Xu struct vm_area_struct *dst_vma, 5014bc70fbf2SPeter Xu struct vm_area_struct *src_vma) 501563551ae0SDavid Gibson { 50163aa4ed80SMiaohe Lin pte_t *src_pte, *dst_pte, entry; 5017ad27ce20SZhangPeng struct folio *pte_folio; 50181c59827dSHugh Dickins unsigned long addr; 5019bc70fbf2SPeter Xu bool cow = is_cow_mapping(src_vma->vm_flags); 5020bc70fbf2SPeter Xu struct hstate *h = hstate_vma(src_vma); 5021a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 50224eae4efaSPeter Xu unsigned long npages = pages_per_huge_page(h); 5023ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 5024e95a9851SMike Kravetz unsigned long last_addr_mask; 5025e8569dd2SAndreas Sandberg int ret = 0; 50261e8f889bSDavid Gibson 5027ac46d4f3SJérôme Glisse if (cow) { 50287d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src, 5029bc70fbf2SPeter Xu src_vma->vm_start, 5030bc70fbf2SPeter Xu src_vma->vm_end); 5031ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 5032623a1ddfSDavid Hildenbrand mmap_assert_write_locked(src); 5033623a1ddfSDavid Hildenbrand raw_write_seqcount_begin(&src->write_protect_seq); 503440549ba8SMike Kravetz } else { 503540549ba8SMike Kravetz /* 503640549ba8SMike Kravetz * For shared mappings the vma lock must be held before 50379c67a207SPeter Xu * calling hugetlb_walk() in the src vma. Otherwise, the 503840549ba8SMike Kravetz * returned ptep could go away if part of a shared pmd and 503940549ba8SMike Kravetz * another thread calls huge_pmd_unshare. 504040549ba8SMike Kravetz */ 504140549ba8SMike Kravetz hugetlb_vma_lock_read(src_vma); 5042ac46d4f3SJérôme Glisse } 5043e8569dd2SAndreas Sandberg 5044e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 5045bc70fbf2SPeter Xu for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) { 5046cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 50479c67a207SPeter Xu src_pte = hugetlb_walk(src_vma, addr, sz); 5048e95a9851SMike Kravetz if (!src_pte) { 5049e95a9851SMike Kravetz addr |= last_addr_mask; 5050c74df32cSHugh Dickins continue; 5051e95a9851SMike Kravetz } 5052bc70fbf2SPeter Xu dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz); 5053e8569dd2SAndreas Sandberg if (!dst_pte) { 5054e8569dd2SAndreas Sandberg ret = -ENOMEM; 5055e8569dd2SAndreas Sandberg break; 5056e8569dd2SAndreas Sandberg } 5057c5c99429SLarry Woodman 50585e41540cSMike Kravetz /* 50595e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 50605e41540cSMike Kravetz * 50613aa4ed80SMiaohe Lin * dst_pte == src_pte is the common case of src/dest sharing. 50625e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 50633aa4ed80SMiaohe Lin * another vma. So page_count of ptep page is checked instead 50643aa4ed80SMiaohe Lin * to reliably determine whether pte is shared. 50655e41540cSMike Kravetz */ 50663aa4ed80SMiaohe Lin if (page_count(virt_to_page(dst_pte)) > 1) { 5067e95a9851SMike Kravetz addr |= last_addr_mask; 5068c5c99429SLarry Woodman continue; 5069e95a9851SMike Kravetz } 5070c5c99429SLarry Woodman 5071cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 5072cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 5073cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 50744a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 50754eae4efaSPeter Xu again: 50763aa4ed80SMiaohe Lin if (huge_pte_none(entry)) { 50775e41540cSMike Kravetz /* 50783aa4ed80SMiaohe Lin * Skip if src entry none. 50795e41540cSMike Kravetz */ 50804a705fefSNaoya Horiguchi ; 5081c2cb0dccSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) { 50825a2f8d22SPeter Xu if (!userfaultfd_wp(dst_vma)) 5083c2cb0dccSNaoya Horiguchi entry = huge_pte_clear_uffd_wp(entry); 5084c2cb0dccSNaoya Horiguchi set_huge_pte_at(dst, addr, dst_pte, entry); 5085c2cb0dccSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry))) { 50864a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 50875a2f8d22SPeter Xu bool uffd_wp = pte_swp_uffd_wp(entry); 50884a705fefSNaoya Horiguchi 50896c287605SDavid Hildenbrand if (!is_readable_migration_entry(swp_entry) && cow) { 50904a705fefSNaoya Horiguchi /* 50914a705fefSNaoya Horiguchi * COW mappings require pages in both 50924a705fefSNaoya Horiguchi * parent and child to be set to read. 50934a705fefSNaoya Horiguchi */ 50944dd845b5SAlistair Popple swp_entry = make_readable_migration_entry( 50954dd845b5SAlistair Popple swp_offset(swp_entry)); 50964a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 5097bc70fbf2SPeter Xu if (userfaultfd_wp(src_vma) && uffd_wp) 50985a2f8d22SPeter Xu entry = pte_swp_mkuffd_wp(entry); 509918f39629SQi Zheng set_huge_pte_at(src, addr, src_pte, entry); 51004a705fefSNaoya Horiguchi } 51015a2f8d22SPeter Xu if (!userfaultfd_wp(dst_vma)) 5102bc70fbf2SPeter Xu entry = huge_pte_clear_uffd_wp(entry); 510318f39629SQi Zheng set_huge_pte_at(dst, addr, dst_pte, entry); 5104bc70fbf2SPeter Xu } else if (unlikely(is_pte_marker(entry))) { 5105af19487fSAxel Rasmussen pte_marker marker = copy_pte_marker( 5106af19487fSAxel Rasmussen pte_to_swp_entry(entry), dst_vma); 5107af19487fSAxel Rasmussen 5108af19487fSAxel Rasmussen if (marker) 5109af19487fSAxel Rasmussen set_huge_pte_at(dst, addr, dst_pte, 5110af19487fSAxel Rasmussen make_pte_marker(marker)); 51114a705fefSNaoya Horiguchi } else { 51124eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 5113ad27ce20SZhangPeng pte_folio = page_folio(pte_page(entry)); 5114ad27ce20SZhangPeng folio_get(pte_folio); 51154eae4efaSPeter Xu 51164eae4efaSPeter Xu /* 5117fb3d824dSDavid Hildenbrand * Failing to duplicate the anon rmap is a rare case 5118fb3d824dSDavid Hildenbrand * where we see pinned hugetlb pages while they're 5119fb3d824dSDavid Hildenbrand * prone to COW. We need to do the COW earlier during 5120fb3d824dSDavid Hildenbrand * fork. 51214eae4efaSPeter Xu * 51224eae4efaSPeter Xu * When pre-allocating the page or copying data, we 51234eae4efaSPeter Xu * need to be without the pgtable locks since we could 51244eae4efaSPeter Xu * sleep during the process. 51254eae4efaSPeter Xu */ 5126ad27ce20SZhangPeng if (!folio_test_anon(pte_folio)) { 5127ad27ce20SZhangPeng page_dup_file_rmap(&pte_folio->page, true); 5128ad27ce20SZhangPeng } else if (page_try_dup_anon_rmap(&pte_folio->page, 5129ad27ce20SZhangPeng true, src_vma)) { 51304eae4efaSPeter Xu pte_t src_pte_old = entry; 5131d0ce0e47SSidhartha Kumar struct folio *new_folio; 51324eae4efaSPeter Xu 51334eae4efaSPeter Xu spin_unlock(src_ptl); 51344eae4efaSPeter Xu spin_unlock(dst_ptl); 51354eae4efaSPeter Xu /* Do not use reserve as it's private owned */ 5136d0ce0e47SSidhartha Kumar new_folio = alloc_hugetlb_folio(dst_vma, addr, 1); 5137d0ce0e47SSidhartha Kumar if (IS_ERR(new_folio)) { 5138ad27ce20SZhangPeng folio_put(pte_folio); 5139d0ce0e47SSidhartha Kumar ret = PTR_ERR(new_folio); 51404eae4efaSPeter Xu break; 51414eae4efaSPeter Xu } 51421cb9dc4bSLiu Shixin ret = copy_user_large_folio(new_folio, 5143ad27ce20SZhangPeng pte_folio, 5144c0e8150eSZhangPeng addr, dst_vma); 5145ad27ce20SZhangPeng folio_put(pte_folio); 51461cb9dc4bSLiu Shixin if (ret) { 51471cb9dc4bSLiu Shixin folio_put(new_folio); 51481cb9dc4bSLiu Shixin break; 51491cb9dc4bSLiu Shixin } 51504eae4efaSPeter Xu 5151d0ce0e47SSidhartha Kumar /* Install the new hugetlb folio if src pte stable */ 51524eae4efaSPeter Xu dst_ptl = huge_pte_lock(h, dst, dst_pte); 51534eae4efaSPeter Xu src_ptl = huge_pte_lockptr(h, src, src_pte); 51544eae4efaSPeter Xu spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 51554eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 51564eae4efaSPeter Xu if (!pte_same(src_pte_old, entry)) { 5157bc70fbf2SPeter Xu restore_reserve_on_error(h, dst_vma, addr, 5158d2d7bb44SSidhartha Kumar new_folio); 5159d0ce0e47SSidhartha Kumar folio_put(new_folio); 51603aa4ed80SMiaohe Lin /* huge_ptep of dst_pte won't change as in child */ 51614eae4efaSPeter Xu goto again; 51624eae4efaSPeter Xu } 51635a2f8d22SPeter Xu hugetlb_install_folio(dst_vma, dst_pte, addr, 51645a2f8d22SPeter Xu new_folio, src_pte_old); 51654eae4efaSPeter Xu spin_unlock(src_ptl); 51664eae4efaSPeter Xu spin_unlock(dst_ptl); 51674eae4efaSPeter Xu continue; 51684eae4efaSPeter Xu } 51694eae4efaSPeter Xu 517034ee645eSJoerg Roedel if (cow) { 51710f10851eSJérôme Glisse /* 51720f10851eSJérôme Glisse * No need to notify as we are downgrading page 51730f10851eSJérôme Glisse * table protection not changing it to point 51740f10851eSJérôme Glisse * to a new page. 51750f10851eSJérôme Glisse * 5176ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 51770f10851eSJérôme Glisse */ 51787f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 517984894e1cSPeter Xu entry = huge_pte_wrprotect(entry); 518034ee645eSJoerg Roedel } 51814eae4efaSPeter Xu 51825a2f8d22SPeter Xu if (!userfaultfd_wp(dst_vma)) 51835a2f8d22SPeter Xu entry = huge_pte_clear_uffd_wp(entry); 51845a2f8d22SPeter Xu 518563551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 51864eae4efaSPeter Xu hugetlb_count_add(npages, dst); 51871c59827dSHugh Dickins } 5188cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 5189cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 519063551ae0SDavid Gibson } 519163551ae0SDavid Gibson 5192623a1ddfSDavid Hildenbrand if (cow) { 5193623a1ddfSDavid Hildenbrand raw_write_seqcount_end(&src->write_protect_seq); 5194ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 519540549ba8SMike Kravetz } else { 519640549ba8SMike Kravetz hugetlb_vma_unlock_read(src_vma); 5197623a1ddfSDavid Hildenbrand } 5198e8569dd2SAndreas Sandberg 5199e8569dd2SAndreas Sandberg return ret; 520063551ae0SDavid Gibson } 520163551ae0SDavid Gibson 5202550a7d60SMina Almasry static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr, 5203db110a99SAneesh Kumar K.V unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte) 5204550a7d60SMina Almasry { 5205550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 5206550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 5207550a7d60SMina Almasry spinlock_t *src_ptl, *dst_ptl; 5208db110a99SAneesh Kumar K.V pte_t pte; 5209550a7d60SMina Almasry 5210550a7d60SMina Almasry dst_ptl = huge_pte_lock(h, mm, dst_pte); 5211550a7d60SMina Almasry src_ptl = huge_pte_lockptr(h, mm, src_pte); 5212550a7d60SMina Almasry 5213550a7d60SMina Almasry /* 5214550a7d60SMina Almasry * We don't have to worry about the ordering of src and dst ptlocks 52158651a137SLorenzo Stoakes * because exclusive mmap_lock (or the i_mmap_lock) prevents deadlock. 5216550a7d60SMina Almasry */ 5217550a7d60SMina Almasry if (src_ptl != dst_ptl) 5218550a7d60SMina Almasry spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 5219550a7d60SMina Almasry 5220550a7d60SMina Almasry pte = huge_ptep_get_and_clear(mm, old_addr, src_pte); 5221550a7d60SMina Almasry set_huge_pte_at(mm, new_addr, dst_pte, pte); 5222550a7d60SMina Almasry 5223550a7d60SMina Almasry if (src_ptl != dst_ptl) 5224550a7d60SMina Almasry spin_unlock(src_ptl); 5225550a7d60SMina Almasry spin_unlock(dst_ptl); 5226550a7d60SMina Almasry } 5227550a7d60SMina Almasry 5228550a7d60SMina Almasry int move_hugetlb_page_tables(struct vm_area_struct *vma, 5229550a7d60SMina Almasry struct vm_area_struct *new_vma, 5230550a7d60SMina Almasry unsigned long old_addr, unsigned long new_addr, 5231550a7d60SMina Almasry unsigned long len) 5232550a7d60SMina Almasry { 5233550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 5234550a7d60SMina Almasry struct address_space *mapping = vma->vm_file->f_mapping; 5235550a7d60SMina Almasry unsigned long sz = huge_page_size(h); 5236550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 5237550a7d60SMina Almasry unsigned long old_end = old_addr + len; 5238e95a9851SMike Kravetz unsigned long last_addr_mask; 5239550a7d60SMina Almasry pte_t *src_pte, *dst_pte; 5240550a7d60SMina Almasry struct mmu_notifier_range range; 52413d0b95cdSBaolin Wang bool shared_pmd = false; 5242550a7d60SMina Almasry 52437d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, old_addr, 5244550a7d60SMina Almasry old_end); 5245550a7d60SMina Almasry adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 52463d0b95cdSBaolin Wang /* 52473d0b95cdSBaolin Wang * In case of shared PMDs, we should cover the maximum possible 52483d0b95cdSBaolin Wang * range. 52493d0b95cdSBaolin Wang */ 52503d0b95cdSBaolin Wang flush_cache_range(vma, range.start, range.end); 52513d0b95cdSBaolin Wang 5252550a7d60SMina Almasry mmu_notifier_invalidate_range_start(&range); 5253e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 5254550a7d60SMina Almasry /* Prevent race with file truncation */ 525540549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 5256550a7d60SMina Almasry i_mmap_lock_write(mapping); 5257550a7d60SMina Almasry for (; old_addr < old_end; old_addr += sz, new_addr += sz) { 52589c67a207SPeter Xu src_pte = hugetlb_walk(vma, old_addr, sz); 5259e95a9851SMike Kravetz if (!src_pte) { 5260e95a9851SMike Kravetz old_addr |= last_addr_mask; 5261e95a9851SMike Kravetz new_addr |= last_addr_mask; 5262550a7d60SMina Almasry continue; 5263e95a9851SMike Kravetz } 5264550a7d60SMina Almasry if (huge_pte_none(huge_ptep_get(src_pte))) 5265550a7d60SMina Almasry continue; 5266550a7d60SMina Almasry 52674ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) { 52683d0b95cdSBaolin Wang shared_pmd = true; 52694ddb4d91SMike Kravetz old_addr |= last_addr_mask; 52704ddb4d91SMike Kravetz new_addr |= last_addr_mask; 5271550a7d60SMina Almasry continue; 52723d0b95cdSBaolin Wang } 5273550a7d60SMina Almasry 5274550a7d60SMina Almasry dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz); 5275550a7d60SMina Almasry if (!dst_pte) 5276550a7d60SMina Almasry break; 5277550a7d60SMina Almasry 5278db110a99SAneesh Kumar K.V move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte); 5279550a7d60SMina Almasry } 52803d0b95cdSBaolin Wang 52813d0b95cdSBaolin Wang if (shared_pmd) 52823d0b95cdSBaolin Wang flush_tlb_range(vma, range.start, range.end); 52833d0b95cdSBaolin Wang else 5284550a7d60SMina Almasry flush_tlb_range(vma, old_end - len, old_end); 5285550a7d60SMina Almasry mmu_notifier_invalidate_range_end(&range); 528613e4ad2cSNadav Amit i_mmap_unlock_write(mapping); 528740549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 5288550a7d60SMina Almasry 5289550a7d60SMina Almasry return len + old_addr - old_end; 5290550a7d60SMina Almasry } 5291550a7d60SMina Almasry 529273c54763SPeter Xu static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 529324669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 529405e90bd0SPeter Xu struct page *ref_page, zap_flags_t zap_flags) 529563551ae0SDavid Gibson { 529663551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 529763551ae0SDavid Gibson unsigned long address; 5298c7546f8fSDavid Gibson pte_t *ptep; 529963551ae0SDavid Gibson pte_t pte; 5300cb900f41SKirill A. Shutemov spinlock_t *ptl; 530163551ae0SDavid Gibson struct page *page; 5302a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 5303a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 5304e95a9851SMike Kravetz unsigned long last_addr_mask; 5305a4a118f2SNadav Amit bool force_flush = false; 5306a5516438SAndi Kleen 530763551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 5308a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 5309a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 531063551ae0SDavid Gibson 531107e32661SAneesh Kumar K.V /* 531207e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 531307e32661SAneesh Kumar K.V * to huge page. 531407e32661SAneesh Kumar K.V */ 5315ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 531624669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 5317dff11abeSMike Kravetz 5318e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 5319569f48b8SHillf Danton address = start; 5320569f48b8SHillf Danton for (; address < end; address += sz) { 53219c67a207SPeter Xu ptep = hugetlb_walk(vma, address, sz); 5322e95a9851SMike Kravetz if (!ptep) { 5323e95a9851SMike Kravetz address |= last_addr_mask; 5324c7546f8fSDavid Gibson continue; 5325e95a9851SMike Kravetz } 5326c7546f8fSDavid Gibson 5327cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 53284ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, address, ptep)) { 532931d49da5SAneesh Kumar K.V spin_unlock(ptl); 5330a4a118f2SNadav Amit tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE); 5331a4a118f2SNadav Amit force_flush = true; 53324ddb4d91SMike Kravetz address |= last_addr_mask; 533331d49da5SAneesh Kumar K.V continue; 533431d49da5SAneesh Kumar K.V } 533539dde65cSChen, Kenneth W 53366629326bSHillf Danton pte = huge_ptep_get(ptep); 533731d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 533831d49da5SAneesh Kumar K.V spin_unlock(ptl); 533931d49da5SAneesh Kumar K.V continue; 534031d49da5SAneesh Kumar K.V } 53416629326bSHillf Danton 53426629326bSHillf Danton /* 53439fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 53449fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 53456629326bSHillf Danton */ 53469fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 534705e90bd0SPeter Xu /* 534805e90bd0SPeter Xu * If the pte was wr-protected by uffd-wp in any of the 534905e90bd0SPeter Xu * swap forms, meanwhile the caller does not want to 535005e90bd0SPeter Xu * drop the uffd-wp bit in this zap, then replace the 535105e90bd0SPeter Xu * pte with a marker. 535205e90bd0SPeter Xu */ 535305e90bd0SPeter Xu if (pte_swp_uffd_wp_any(pte) && 535405e90bd0SPeter Xu !(zap_flags & ZAP_FLAG_DROP_MARKER)) 535505e90bd0SPeter Xu set_huge_pte_at(mm, address, ptep, 535605e90bd0SPeter Xu make_pte_marker(PTE_MARKER_UFFD_WP)); 535705e90bd0SPeter Xu else 53589386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 535931d49da5SAneesh Kumar K.V spin_unlock(ptl); 536031d49da5SAneesh Kumar K.V continue; 53618c4894c6SNaoya Horiguchi } 53626629326bSHillf Danton 53636629326bSHillf Danton page = pte_page(pte); 536404f2cbe3SMel Gorman /* 536504f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 536604f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 536704f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 536804f2cbe3SMel Gorman */ 536904f2cbe3SMel Gorman if (ref_page) { 537031d49da5SAneesh Kumar K.V if (page != ref_page) { 537131d49da5SAneesh Kumar K.V spin_unlock(ptl); 537231d49da5SAneesh Kumar K.V continue; 537331d49da5SAneesh Kumar K.V } 537404f2cbe3SMel Gorman /* 537504f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 537604f2cbe3SMel Gorman * future faults in this VMA will fail rather than 537704f2cbe3SMel Gorman * looking like data was lost 537804f2cbe3SMel Gorman */ 537904f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 538004f2cbe3SMel Gorman } 538104f2cbe3SMel Gorman 5382c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 5383b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 5384106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 53856649a386SKen Chen set_page_dirty(page); 538605e90bd0SPeter Xu /* Leave a uffd-wp pte marker if needed */ 538705e90bd0SPeter Xu if (huge_pte_uffd_wp(pte) && 538805e90bd0SPeter Xu !(zap_flags & ZAP_FLAG_DROP_MARKER)) 538905e90bd0SPeter Xu set_huge_pte_at(mm, address, ptep, 539005e90bd0SPeter Xu make_pte_marker(PTE_MARKER_UFFD_WP)); 53915d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 5392cea86fe2SHugh Dickins page_remove_rmap(page, vma, true); 539331d49da5SAneesh Kumar K.V 5394cb900f41SKirill A. Shutemov spin_unlock(ptl); 5395e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 539624669e58SAneesh Kumar K.V /* 539731d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 539824669e58SAneesh Kumar K.V */ 539931d49da5SAneesh Kumar K.V if (ref_page) 540031d49da5SAneesh Kumar K.V break; 5401fe1668aeSChen, Kenneth W } 540224669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 5403a4a118f2SNadav Amit 5404a4a118f2SNadav Amit /* 5405a4a118f2SNadav Amit * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We 5406a4a118f2SNadav Amit * could defer the flush until now, since by holding i_mmap_rwsem we 5407a4a118f2SNadav Amit * guaranteed that the last refernece would not be dropped. But we must 5408a4a118f2SNadav Amit * do the flushing before we return, as otherwise i_mmap_rwsem will be 5409a4a118f2SNadav Amit * dropped and the last reference to the shared PMDs page might be 5410a4a118f2SNadav Amit * dropped as well. 5411a4a118f2SNadav Amit * 5412a4a118f2SNadav Amit * In theory we could defer the freeing of the PMD pages as well, but 5413a4a118f2SNadav Amit * huge_pmd_unshare() relies on the exact page_count for the PMD page to 5414a4a118f2SNadav Amit * detect sharing, so we cannot defer the release of the page either. 5415a4a118f2SNadav Amit * Instead, do flush now. 5416a4a118f2SNadav Amit */ 5417a4a118f2SNadav Amit if (force_flush) 5418a4a118f2SNadav Amit tlb_flush_mmu_tlbonly(tlb); 54191da177e4SLinus Torvalds } 542063551ae0SDavid Gibson 5421d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 5422d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 542305e90bd0SPeter Xu unsigned long end, struct page *ref_page, 542405e90bd0SPeter Xu zap_flags_t zap_flags) 5425d833352aSMel Gorman { 5426131a79b4SMike Kravetz hugetlb_vma_lock_write(vma); 5427131a79b4SMike Kravetz i_mmap_lock_write(vma->vm_file->f_mapping); 5428131a79b4SMike Kravetz 5429369258ceSMike Kravetz /* mmu notification performed in caller */ 543005e90bd0SPeter Xu __unmap_hugepage_range(tlb, vma, start, end, ref_page, zap_flags); 5431d833352aSMel Gorman 543204ada095SMike Kravetz if (zap_flags & ZAP_FLAG_UNMAP) { /* final unmap */ 5433d833352aSMel Gorman /* 543404ada095SMike Kravetz * Unlock and free the vma lock before releasing i_mmap_rwsem. 543504ada095SMike Kravetz * When the vma_lock is freed, this makes the vma ineligible 543604ada095SMike Kravetz * for pmd sharing. And, i_mmap_rwsem is required to set up 543704ada095SMike Kravetz * pmd sharing. This is important as page tables for this 543804ada095SMike Kravetz * unmapped range will be asynchrously deleted. If the page 543904ada095SMike Kravetz * tables are shared, there will be issues when accessed by 544004ada095SMike Kravetz * someone else. 5441d833352aSMel Gorman */ 5442ecfbd733SMike Kravetz __hugetlb_vma_unlock_write_free(vma); 5443131a79b4SMike Kravetz i_mmap_unlock_write(vma->vm_file->f_mapping); 544404ada095SMike Kravetz } else { 544504ada095SMike Kravetz i_mmap_unlock_write(vma->vm_file->f_mapping); 544604ada095SMike Kravetz hugetlb_vma_unlock_write(vma); 544704ada095SMike Kravetz } 5448d833352aSMel Gorman } 5449d833352aSMel Gorman 5450502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 545105e90bd0SPeter Xu unsigned long end, struct page *ref_page, 545205e90bd0SPeter Xu zap_flags_t zap_flags) 5453502717f4SChen, Kenneth W { 5454369258ceSMike Kravetz struct mmu_notifier_range range; 545524669e58SAneesh Kumar K.V struct mmu_gather tlb; 5456dff11abeSMike Kravetz 54577d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm, 5458369258ceSMike Kravetz start, end); 5459369258ceSMike Kravetz adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 5460369258ceSMike Kravetz mmu_notifier_invalidate_range_start(&range); 5461a72afd87SWill Deacon tlb_gather_mmu(&tlb, vma->vm_mm); 5462369258ceSMike Kravetz 546305e90bd0SPeter Xu __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags); 5464369258ceSMike Kravetz 5465369258ceSMike Kravetz mmu_notifier_invalidate_range_end(&range); 5466ae8eba8bSWill Deacon tlb_finish_mmu(&tlb); 5467502717f4SChen, Kenneth W } 5468502717f4SChen, Kenneth W 546904f2cbe3SMel Gorman /* 547004f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 5471578b7725SZhiyuan Dai * mapping it owns the reserve page for. The intention is to unmap the page 547204f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 547304f2cbe3SMel Gorman * same region. 547404f2cbe3SMel Gorman */ 54752f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 54762a4b3dedSHarvey Harrison struct page *page, unsigned long address) 547704f2cbe3SMel Gorman { 54787526674dSAdam Litke struct hstate *h = hstate_vma(vma); 547904f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 548004f2cbe3SMel Gorman struct address_space *mapping; 548104f2cbe3SMel Gorman pgoff_t pgoff; 548204f2cbe3SMel Gorman 548304f2cbe3SMel Gorman /* 548404f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 548504f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 548604f2cbe3SMel Gorman */ 54877526674dSAdam Litke address = address & huge_page_mask(h); 548836e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 548936e4f20aSMichal Hocko vma->vm_pgoff; 549093c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 549104f2cbe3SMel Gorman 54924eb2b1dcSMel Gorman /* 54934eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 54944eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 54954eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 54964eb2b1dcSMel Gorman */ 549783cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 54986b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 549904f2cbe3SMel Gorman /* Do not unmap the current VMA */ 550004f2cbe3SMel Gorman if (iter_vma == vma) 550104f2cbe3SMel Gorman continue; 550204f2cbe3SMel Gorman 550304f2cbe3SMel Gorman /* 55042f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 55052f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 55062f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 55072f84a899SMel Gorman */ 55082f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 55092f84a899SMel Gorman continue; 55102f84a899SMel Gorman 55112f84a899SMel Gorman /* 551204f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 551304f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 551404f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 551504f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 551604f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 551704f2cbe3SMel Gorman */ 551804f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 551924669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 552005e90bd0SPeter Xu address + huge_page_size(h), page, 0); 552104f2cbe3SMel Gorman } 552283cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 552304f2cbe3SMel Gorman } 552404f2cbe3SMel Gorman 55250fe6e20bSNaoya Horiguchi /* 5526c89357e2SDavid Hildenbrand * hugetlb_wp() should be called with page lock of the original hugepage held. 5527aa6d2e8cSBaolin Wang * Called with hugetlb_fault_mutex_table held and pte_page locked so we 5528ef009b25SMichal Hocko * cannot race with other handlers or page migration. 5529ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 55300fe6e20bSNaoya Horiguchi */ 5531c89357e2SDavid Hildenbrand static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma, 5532c89357e2SDavid Hildenbrand unsigned long address, pte_t *ptep, unsigned int flags, 5533371607a3SSidhartha Kumar struct folio *pagecache_folio, spinlock_t *ptl) 55341e8f889bSDavid Gibson { 5535c89357e2SDavid Hildenbrand const bool unshare = flags & FAULT_FLAG_UNSHARE; 553660d5b473SPeter Xu pte_t pte = huge_ptep_get(ptep); 5537a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 5538959a78b6SZhangPeng struct folio *old_folio; 5539d0ce0e47SSidhartha Kumar struct folio *new_folio; 55402b740303SSouptick Joarder int outside_reserve = 0; 55412b740303SSouptick Joarder vm_fault_t ret = 0; 5542974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 5543ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 55441e8f889bSDavid Gibson 55451d8d1464SDavid Hildenbrand /* 554660d5b473SPeter Xu * Never handle CoW for uffd-wp protected pages. It should be only 554760d5b473SPeter Xu * handled when the uffd-wp protection is removed. 554860d5b473SPeter Xu * 554960d5b473SPeter Xu * Note that only the CoW optimization path (in hugetlb_no_page()) 555060d5b473SPeter Xu * can trigger this, because hugetlb_fault() will always resolve 555160d5b473SPeter Xu * uffd-wp bit first. 555260d5b473SPeter Xu */ 555360d5b473SPeter Xu if (!unshare && huge_pte_uffd_wp(pte)) 555460d5b473SPeter Xu return 0; 555560d5b473SPeter Xu 555660d5b473SPeter Xu /* 55571d8d1464SDavid Hildenbrand * hugetlb does not support FOLL_FORCE-style write faults that keep the 55581d8d1464SDavid Hildenbrand * PTE mapped R/O such as maybe_mkwrite() would do. 55591d8d1464SDavid Hildenbrand */ 55601d8d1464SDavid Hildenbrand if (WARN_ON_ONCE(!unshare && !(vma->vm_flags & VM_WRITE))) 55611d8d1464SDavid Hildenbrand return VM_FAULT_SIGSEGV; 55621d8d1464SDavid Hildenbrand 55631d8d1464SDavid Hildenbrand /* Let's take out MAP_SHARED mappings first. */ 55641d8d1464SDavid Hildenbrand if (vma->vm_flags & VM_MAYSHARE) { 55651d8d1464SDavid Hildenbrand set_huge_ptep_writable(vma, haddr, ptep); 55661d8d1464SDavid Hildenbrand return 0; 55671d8d1464SDavid Hildenbrand } 55681d8d1464SDavid Hildenbrand 5569959a78b6SZhangPeng old_folio = page_folio(pte_page(pte)); 55701e8f889bSDavid Gibson 5571662ce1dcSYang Yang delayacct_wpcopy_start(); 5572662ce1dcSYang Yang 557304f2cbe3SMel Gorman retry_avoidcopy: 5574c89357e2SDavid Hildenbrand /* 5575c89357e2SDavid Hildenbrand * If no-one else is actually using this page, we're the exclusive 5576c89357e2SDavid Hildenbrand * owner and can reuse this page. 5577c89357e2SDavid Hildenbrand */ 5578959a78b6SZhangPeng if (folio_mapcount(old_folio) == 1 && folio_test_anon(old_folio)) { 5579959a78b6SZhangPeng if (!PageAnonExclusive(&old_folio->page)) 5580959a78b6SZhangPeng page_move_anon_rmap(&old_folio->page, vma); 5581c89357e2SDavid Hildenbrand if (likely(!unshare)) 55825b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 5583662ce1dcSYang Yang 5584662ce1dcSYang Yang delayacct_wpcopy_end(); 558583c54070SNick Piggin return 0; 55861e8f889bSDavid Gibson } 5587959a78b6SZhangPeng VM_BUG_ON_PAGE(folio_test_anon(old_folio) && 5588959a78b6SZhangPeng PageAnonExclusive(&old_folio->page), &old_folio->page); 55891e8f889bSDavid Gibson 559004f2cbe3SMel Gorman /* 559104f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 559204f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 559304f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 559404f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 559504f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 559604f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 559704f2cbe3SMel Gorman * of the full address range. 559804f2cbe3SMel Gorman */ 55995944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 5600959a78b6SZhangPeng old_folio != pagecache_folio) 560104f2cbe3SMel Gorman outside_reserve = 1; 560204f2cbe3SMel Gorman 5603959a78b6SZhangPeng folio_get(old_folio); 5604b76c8cfbSLarry Woodman 5605ad4404a2SDavidlohr Bueso /* 5606ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 5607ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 5608ad4404a2SDavidlohr Bueso */ 5609cb900f41SKirill A. Shutemov spin_unlock(ptl); 5610d0ce0e47SSidhartha Kumar new_folio = alloc_hugetlb_folio(vma, haddr, outside_reserve); 56111e8f889bSDavid Gibson 5612d0ce0e47SSidhartha Kumar if (IS_ERR(new_folio)) { 561304f2cbe3SMel Gorman /* 561404f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 561504f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 561604f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 561704f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 561804f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 561904f2cbe3SMel Gorman */ 562004f2cbe3SMel Gorman if (outside_reserve) { 562140549ba8SMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 562240549ba8SMike Kravetz pgoff_t idx; 562340549ba8SMike Kravetz u32 hash; 562440549ba8SMike Kravetz 5625959a78b6SZhangPeng folio_put(old_folio); 562640549ba8SMike Kravetz /* 562740549ba8SMike Kravetz * Drop hugetlb_fault_mutex and vma_lock before 562840549ba8SMike Kravetz * unmapping. unmapping needs to hold vma_lock 562940549ba8SMike Kravetz * in write mode. Dropping vma_lock in read mode 563040549ba8SMike Kravetz * here is OK as COW mappings do not interact with 563140549ba8SMike Kravetz * PMD sharing. 563240549ba8SMike Kravetz * 563340549ba8SMike Kravetz * Reacquire both after unmap operation. 563440549ba8SMike Kravetz */ 563540549ba8SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 563640549ba8SMike Kravetz hash = hugetlb_fault_mutex_hash(mapping, idx); 563740549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 563840549ba8SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 563940549ba8SMike Kravetz 5640959a78b6SZhangPeng unmap_ref_private(mm, vma, &old_folio->page, haddr); 564140549ba8SMike Kravetz 564240549ba8SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 564340549ba8SMike Kravetz hugetlb_vma_lock_read(vma); 5644cb900f41SKirill A. Shutemov spin_lock(ptl); 56459c67a207SPeter Xu ptep = hugetlb_walk(vma, haddr, huge_page_size(h)); 5646a9af0c5dSNaoya Horiguchi if (likely(ptep && 5647a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 564804f2cbe3SMel Gorman goto retry_avoidcopy; 5649a734bcc8SHillf Danton /* 5650cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 5651cb900f41SKirill A. Shutemov * lock, and our job is done. 5652a734bcc8SHillf Danton */ 5653662ce1dcSYang Yang delayacct_wpcopy_end(); 5654a734bcc8SHillf Danton return 0; 565504f2cbe3SMel Gorman } 565604f2cbe3SMel Gorman 5657d0ce0e47SSidhartha Kumar ret = vmf_error(PTR_ERR(new_folio)); 5658ad4404a2SDavidlohr Bueso goto out_release_old; 56591e8f889bSDavid Gibson } 56601e8f889bSDavid Gibson 56610fe6e20bSNaoya Horiguchi /* 56620fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 56630fe6e20bSNaoya Horiguchi * anon_vma prepared. 56640fe6e20bSNaoya Horiguchi */ 566544e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 5666ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 5667ad4404a2SDavidlohr Bueso goto out_release_all; 566844e2aa93SDean Nelson } 56690fe6e20bSNaoya Horiguchi 5670959a78b6SZhangPeng if (copy_user_large_folio(new_folio, old_folio, address, vma)) { 56711cb9dc4bSLiu Shixin ret = VM_FAULT_HWPOISON_LARGE; 56721cb9dc4bSLiu Shixin goto out_release_all; 56731cb9dc4bSLiu Shixin } 5674d0ce0e47SSidhartha Kumar __folio_mark_uptodate(new_folio); 56751e8f889bSDavid Gibson 56767d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, haddr, 56776f4f13e8SJérôme Glisse haddr + huge_page_size(h)); 5678ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 5679ad4404a2SDavidlohr Bueso 5680b76c8cfbSLarry Woodman /* 5681cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 5682b76c8cfbSLarry Woodman * before the page tables are altered 5683b76c8cfbSLarry Woodman */ 5684cb900f41SKirill A. Shutemov spin_lock(ptl); 56859c67a207SPeter Xu ptep = hugetlb_walk(vma, haddr, huge_page_size(h)); 5686a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 56870f230bc2SPeter Xu pte_t newpte = make_huge_pte(vma, &new_folio->page, !unshare); 56880f230bc2SPeter Xu 5689c89357e2SDavid Hildenbrand /* Break COW or unshare */ 56905b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 5691ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, range.end); 5692959a78b6SZhangPeng page_remove_rmap(&old_folio->page, vma, true); 5693d0ce0e47SSidhartha Kumar hugepage_add_new_anon_rmap(new_folio, vma, haddr); 56940f230bc2SPeter Xu if (huge_pte_uffd_wp(pte)) 56950f230bc2SPeter Xu newpte = huge_pte_mkuffd_wp(newpte); 56960f230bc2SPeter Xu set_huge_pte_at(mm, haddr, ptep, newpte); 5697d0ce0e47SSidhartha Kumar folio_set_hugetlb_migratable(new_folio); 56981e8f889bSDavid Gibson /* Make the old page be freed below */ 5699959a78b6SZhangPeng new_folio = old_folio; 57001e8f889bSDavid Gibson } 5701cb900f41SKirill A. Shutemov spin_unlock(ptl); 5702ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 5703ad4404a2SDavidlohr Bueso out_release_all: 5704c89357e2SDavid Hildenbrand /* 5705c89357e2SDavid Hildenbrand * No restore in case of successful pagetable update (Break COW or 5706c89357e2SDavid Hildenbrand * unshare) 5707c89357e2SDavid Hildenbrand */ 5708959a78b6SZhangPeng if (new_folio != old_folio) 5709d2d7bb44SSidhartha Kumar restore_reserve_on_error(h, vma, haddr, new_folio); 5710d0ce0e47SSidhartha Kumar folio_put(new_folio); 5711ad4404a2SDavidlohr Bueso out_release_old: 5712959a78b6SZhangPeng folio_put(old_folio); 57138312034fSJoonsoo Kim 5714ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 5715662ce1dcSYang Yang 5716662ce1dcSYang Yang delayacct_wpcopy_end(); 5717ad4404a2SDavidlohr Bueso return ret; 57181e8f889bSDavid Gibson } 57191e8f889bSDavid Gibson 57203ae77f43SHugh Dickins /* 57213ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 57223ae77f43SHugh Dickins */ 57233ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 57242a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 57252a15efc9SHugh Dickins { 572691a2fb95SSidhartha Kumar struct address_space *mapping = vma->vm_file->f_mapping; 572791a2fb95SSidhartha Kumar pgoff_t idx = vma_hugecache_offset(h, vma, address); 5728fd4aed8dSMike Kravetz struct folio *folio; 57292a15efc9SHugh Dickins 5730fd4aed8dSMike Kravetz folio = filemap_get_folio(mapping, idx); 5731fd4aed8dSMike Kravetz if (IS_ERR(folio)) 5732fd4aed8dSMike Kravetz return false; 5733fd4aed8dSMike Kravetz folio_put(folio); 5734fd4aed8dSMike Kravetz return true; 57352a15efc9SHugh Dickins } 57362a15efc9SHugh Dickins 57379b91c0e2SSidhartha Kumar int hugetlb_add_to_page_cache(struct folio *folio, struct address_space *mapping, 5738ab76ad54SMike Kravetz pgoff_t idx) 5739ab76ad54SMike Kravetz { 5740ab76ad54SMike Kravetz struct inode *inode = mapping->host; 5741ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 5742d9ef44deSMatthew Wilcox (Oracle) int err; 5743ab76ad54SMike Kravetz 5744d9ef44deSMatthew Wilcox (Oracle) __folio_set_locked(folio); 5745d9ef44deSMatthew Wilcox (Oracle) err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL); 5746d9ef44deSMatthew Wilcox (Oracle) 5747d9ef44deSMatthew Wilcox (Oracle) if (unlikely(err)) { 5748d9ef44deSMatthew Wilcox (Oracle) __folio_clear_locked(folio); 5749ab76ad54SMike Kravetz return err; 5750d9ef44deSMatthew Wilcox (Oracle) } 57519b91c0e2SSidhartha Kumar folio_clear_hugetlb_restore_reserve(folio); 5752ab76ad54SMike Kravetz 575322146c3cSMike Kravetz /* 5754d9ef44deSMatthew Wilcox (Oracle) * mark folio dirty so that it will not be removed from cache/file 575522146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 575622146c3cSMike Kravetz */ 5757d9ef44deSMatthew Wilcox (Oracle) folio_mark_dirty(folio); 575822146c3cSMike Kravetz 5759ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 5760ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 5761ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 5762ab76ad54SMike Kravetz return 0; 5763ab76ad54SMike Kravetz } 5764ab76ad54SMike Kravetz 57657677f7fdSAxel Rasmussen static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma, 57667677f7fdSAxel Rasmussen struct address_space *mapping, 57677677f7fdSAxel Rasmussen pgoff_t idx, 57687677f7fdSAxel Rasmussen unsigned int flags, 57697677f7fdSAxel Rasmussen unsigned long haddr, 5770824ddc60SNadav Amit unsigned long addr, 57717677f7fdSAxel Rasmussen unsigned long reason) 57727677f7fdSAxel Rasmussen { 57737677f7fdSAxel Rasmussen u32 hash; 57747677f7fdSAxel Rasmussen struct vm_fault vmf = { 57757677f7fdSAxel Rasmussen .vma = vma, 57767677f7fdSAxel Rasmussen .address = haddr, 5777824ddc60SNadav Amit .real_address = addr, 57787677f7fdSAxel Rasmussen .flags = flags, 57797677f7fdSAxel Rasmussen 57807677f7fdSAxel Rasmussen /* 57817677f7fdSAxel Rasmussen * Hard to debug if it ends up being 57827677f7fdSAxel Rasmussen * used by a callee that assumes 57837677f7fdSAxel Rasmussen * something about the other 57847677f7fdSAxel Rasmussen * uninitialized fields... same as in 57857677f7fdSAxel Rasmussen * memory.c 57867677f7fdSAxel Rasmussen */ 57877677f7fdSAxel Rasmussen }; 57887677f7fdSAxel Rasmussen 57897677f7fdSAxel Rasmussen /* 5790958f32ceSLiu Shixin * vma_lock and hugetlb_fault_mutex must be dropped before handling 5791958f32ceSLiu Shixin * userfault. Also mmap_lock could be dropped due to handling 5792958f32ceSLiu Shixin * userfault, any vma operation should be careful from here. 57937677f7fdSAxel Rasmussen */ 579440549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 57957677f7fdSAxel Rasmussen hash = hugetlb_fault_mutex_hash(mapping, idx); 57967677f7fdSAxel Rasmussen mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5797958f32ceSLiu Shixin return handle_userfault(&vmf, reason); 57987677f7fdSAxel Rasmussen } 57997677f7fdSAxel Rasmussen 58002ea7ff1eSPeter Xu /* 58012ea7ff1eSPeter Xu * Recheck pte with pgtable lock. Returns true if pte didn't change, or 58022ea7ff1eSPeter Xu * false if pte changed or is changing. 58032ea7ff1eSPeter Xu */ 58042ea7ff1eSPeter Xu static bool hugetlb_pte_stable(struct hstate *h, struct mm_struct *mm, 58052ea7ff1eSPeter Xu pte_t *ptep, pte_t old_pte) 58062ea7ff1eSPeter Xu { 58072ea7ff1eSPeter Xu spinlock_t *ptl; 58082ea7ff1eSPeter Xu bool same; 58092ea7ff1eSPeter Xu 58102ea7ff1eSPeter Xu ptl = huge_pte_lock(h, mm, ptep); 58112ea7ff1eSPeter Xu same = pte_same(huge_ptep_get(ptep), old_pte); 58122ea7ff1eSPeter Xu spin_unlock(ptl); 58132ea7ff1eSPeter Xu 58142ea7ff1eSPeter Xu return same; 58152ea7ff1eSPeter Xu } 58162ea7ff1eSPeter Xu 58172b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 58182b740303SSouptick Joarder struct vm_area_struct *vma, 58198382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 5820c64e912cSPeter Xu unsigned long address, pte_t *ptep, 5821c64e912cSPeter Xu pte_t old_pte, unsigned int flags) 5822ac9b9c66SHugh Dickins { 5823a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 58242b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 5825409eb8c2SHillf Danton int anon_rmap = 0; 58264c887265SAdam Litke unsigned long size; 5827d0ce0e47SSidhartha Kumar struct folio *folio; 58281e8f889bSDavid Gibson pte_t new_pte; 5829cb900f41SKirill A. Shutemov spinlock_t *ptl; 5830285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 5831d0ce0e47SSidhartha Kumar bool new_folio, new_pagecache_folio = false; 5832958f32ceSLiu Shixin u32 hash = hugetlb_fault_mutex_hash(mapping, idx); 58334c887265SAdam Litke 583404f2cbe3SMel Gorman /* 583504f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 583604f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 5837c89357e2SDavid Hildenbrand * COW/unsharing. Warn that such a situation has occurred as it may not 5838c89357e2SDavid Hildenbrand * be obvious. 583904f2cbe3SMel Gorman */ 584004f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 5841910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 584204f2cbe3SMel Gorman current->pid); 5843958f32ceSLiu Shixin goto out; 584404f2cbe3SMel Gorman } 584504f2cbe3SMel Gorman 58464c887265SAdam Litke /* 5847188a3972SMike Kravetz * Use page lock to guard against racing truncation 5848188a3972SMike Kravetz * before we get page_table_lock. 58494c887265SAdam Litke */ 5850d0ce0e47SSidhartha Kumar new_folio = false; 5851d0ce0e47SSidhartha Kumar folio = filemap_lock_folio(mapping, idx); 585266dabbb6SChristoph Hellwig if (IS_ERR(folio)) { 5853188a3972SMike Kravetz size = i_size_read(mapping->host) >> huge_page_shift(h); 5854188a3972SMike Kravetz if (idx >= size) 5855188a3972SMike Kravetz goto out; 58567677f7fdSAxel Rasmussen /* Check for page in userfault range */ 58572ea7ff1eSPeter Xu if (userfaultfd_missing(vma)) { 58582ea7ff1eSPeter Xu /* 58592ea7ff1eSPeter Xu * Since hugetlb_no_page() was examining pte 58602ea7ff1eSPeter Xu * without pgtable lock, we need to re-test under 58612ea7ff1eSPeter Xu * lock because the pte may not be stable and could 58622ea7ff1eSPeter Xu * have changed from under us. Try to detect 58632ea7ff1eSPeter Xu * either changed or during-changing ptes and retry 58642ea7ff1eSPeter Xu * properly when needed. 58652ea7ff1eSPeter Xu * 58662ea7ff1eSPeter Xu * Note that userfaultfd is actually fine with 58672ea7ff1eSPeter Xu * false positives (e.g. caused by pte changed), 58682ea7ff1eSPeter Xu * but not wrong logical events (e.g. caused by 58692ea7ff1eSPeter Xu * reading a pte during changing). The latter can 58702ea7ff1eSPeter Xu * confuse the userspace, so the strictness is very 58712ea7ff1eSPeter Xu * much preferred. E.g., MISSING event should 58722ea7ff1eSPeter Xu * never happen on the page after UFFDIO_COPY has 58732ea7ff1eSPeter Xu * correctly installed the page and returned. 58742ea7ff1eSPeter Xu */ 58752ea7ff1eSPeter Xu if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) { 58762ea7ff1eSPeter Xu ret = 0; 58772ea7ff1eSPeter Xu goto out; 58782ea7ff1eSPeter Xu } 58792ea7ff1eSPeter Xu 58802ea7ff1eSPeter Xu return hugetlb_handle_userfault(vma, mapping, idx, flags, 58812ea7ff1eSPeter Xu haddr, address, 58827677f7fdSAxel Rasmussen VM_UFFD_MISSING); 58832ea7ff1eSPeter Xu } 58841a1aad8aSMike Kravetz 5885d0ce0e47SSidhartha Kumar folio = alloc_hugetlb_folio(vma, haddr, 0); 5886d0ce0e47SSidhartha Kumar if (IS_ERR(folio)) { 58874643d67eSMike Kravetz /* 58884643d67eSMike Kravetz * Returning error will result in faulting task being 58894643d67eSMike Kravetz * sent SIGBUS. The hugetlb fault mutex prevents two 58904643d67eSMike Kravetz * tasks from racing to fault in the same page which 58914643d67eSMike Kravetz * could result in false unable to allocate errors. 58924643d67eSMike Kravetz * Page migration does not take the fault mutex, but 58934643d67eSMike Kravetz * does a clear then write of pte's under page table 58944643d67eSMike Kravetz * lock. Page fault code could race with migration, 58954643d67eSMike Kravetz * notice the clear pte and try to allocate a page 58964643d67eSMike Kravetz * here. Before returning error, get ptl and make 58974643d67eSMike Kravetz * sure there really is no pte entry. 58984643d67eSMike Kravetz */ 5899f9bf6c03SPeter Xu if (hugetlb_pte_stable(h, mm, ptep, old_pte)) 5900d0ce0e47SSidhartha Kumar ret = vmf_error(PTR_ERR(folio)); 5901f9bf6c03SPeter Xu else 5902f9bf6c03SPeter Xu ret = 0; 59036bda666aSChristoph Lameter goto out; 59046bda666aSChristoph Lameter } 5905d0ce0e47SSidhartha Kumar clear_huge_page(&folio->page, address, pages_per_huge_page(h)); 5906d0ce0e47SSidhartha Kumar __folio_mark_uptodate(folio); 5907d0ce0e47SSidhartha Kumar new_folio = true; 5908ac9b9c66SHugh Dickins 5909f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 59109b91c0e2SSidhartha Kumar int err = hugetlb_add_to_page_cache(folio, mapping, idx); 59116bda666aSChristoph Lameter if (err) { 59123a5497a2SMiaohe Lin /* 59133a5497a2SMiaohe Lin * err can't be -EEXIST which implies someone 59143a5497a2SMiaohe Lin * else consumed the reservation since hugetlb 59153a5497a2SMiaohe Lin * fault mutex is held when add a hugetlb page 59163a5497a2SMiaohe Lin * to the page cache. So it's safe to call 59173a5497a2SMiaohe Lin * restore_reserve_on_error() here. 59183a5497a2SMiaohe Lin */ 5919d2d7bb44SSidhartha Kumar restore_reserve_on_error(h, vma, haddr, folio); 5920d0ce0e47SSidhartha Kumar folio_put(folio); 59216bda666aSChristoph Lameter goto out; 59226bda666aSChristoph Lameter } 5923d0ce0e47SSidhartha Kumar new_pagecache_folio = true; 592423be7468SMel Gorman } else { 5925d0ce0e47SSidhartha Kumar folio_lock(folio); 59260fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 59270fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 59280fe6e20bSNaoya Horiguchi goto backout_unlocked; 592923be7468SMel Gorman } 5930409eb8c2SHillf Danton anon_rmap = 1; 59310fe6e20bSNaoya Horiguchi } 59320fe6e20bSNaoya Horiguchi } else { 593357303d80SAndy Whitcroft /* 5934998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 5935998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 5936998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 5937fd6a03edSNaoya Horiguchi */ 5938d0ce0e47SSidhartha Kumar if (unlikely(folio_test_hwpoison(folio))) { 59390eb98f15SMiaohe Lin ret = VM_FAULT_HWPOISON_LARGE | 5940972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 5941fd6a03edSNaoya Horiguchi goto backout_unlocked; 59426bda666aSChristoph Lameter } 59437677f7fdSAxel Rasmussen 59447677f7fdSAxel Rasmussen /* Check for page in userfault range. */ 59457677f7fdSAxel Rasmussen if (userfaultfd_minor(vma)) { 5946d0ce0e47SSidhartha Kumar folio_unlock(folio); 5947d0ce0e47SSidhartha Kumar folio_put(folio); 59482ea7ff1eSPeter Xu /* See comment in userfaultfd_missing() block above */ 59492ea7ff1eSPeter Xu if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) { 59502ea7ff1eSPeter Xu ret = 0; 59512ea7ff1eSPeter Xu goto out; 59522ea7ff1eSPeter Xu } 59532ea7ff1eSPeter Xu return hugetlb_handle_userfault(vma, mapping, idx, flags, 59542ea7ff1eSPeter Xu haddr, address, 59557677f7fdSAxel Rasmussen VM_UFFD_MINOR); 59567677f7fdSAxel Rasmussen } 5957998b4382SNaoya Horiguchi } 59581e8f889bSDavid Gibson 595957303d80SAndy Whitcroft /* 596057303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 596157303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 596257303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 596357303d80SAndy Whitcroft * the spinlock. 596457303d80SAndy Whitcroft */ 59655e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 5966285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 59672b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 59682b26736cSAndy Whitcroft goto backout_unlocked; 59692b26736cSAndy Whitcroft } 59705e911373SMike Kravetz /* Just decrements count, does not deallocate */ 5971285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 59725e911373SMike Kravetz } 597357303d80SAndy Whitcroft 59748bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 597583c54070SNick Piggin ret = 0; 5976c64e912cSPeter Xu /* If pte changed from under us, retry */ 5977c64e912cSPeter Xu if (!pte_same(huge_ptep_get(ptep), old_pte)) 59784c887265SAdam Litke goto backout; 59794c887265SAdam Litke 59804781593dSPeter Xu if (anon_rmap) 5981d0ce0e47SSidhartha Kumar hugepage_add_new_anon_rmap(folio, vma, haddr); 59824781593dSPeter Xu else 5983d0ce0e47SSidhartha Kumar page_dup_file_rmap(&folio->page, true); 5984d0ce0e47SSidhartha Kumar new_pte = make_huge_pte(vma, &folio->page, ((vma->vm_flags & VM_WRITE) 59851e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 5986c64e912cSPeter Xu /* 5987c64e912cSPeter Xu * If this pte was previously wr-protected, keep it wr-protected even 5988c64e912cSPeter Xu * if populated. 5989c64e912cSPeter Xu */ 5990c64e912cSPeter Xu if (unlikely(pte_marker_uffd_wp(old_pte))) 5991f1eb1bacSPeter Xu new_pte = huge_pte_mkuffd_wp(new_pte); 5992285b8dcaSHuang Ying set_huge_pte_at(mm, haddr, ptep, new_pte); 59931e8f889bSDavid Gibson 59945d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 5995788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 59961e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 5997371607a3SSidhartha Kumar ret = hugetlb_wp(mm, vma, address, ptep, flags, folio, ptl); 59981e8f889bSDavid Gibson } 59991e8f889bSDavid Gibson 6000cb900f41SKirill A. Shutemov spin_unlock(ptl); 6001cb6acd01SMike Kravetz 6002cb6acd01SMike Kravetz /* 6003d0ce0e47SSidhartha Kumar * Only set hugetlb_migratable in newly allocated pages. Existing pages 6004d0ce0e47SSidhartha Kumar * found in the pagecache may not have hugetlb_migratable if they have 60058f251a3dSMike Kravetz * been isolated for migration. 6006cb6acd01SMike Kravetz */ 6007d0ce0e47SSidhartha Kumar if (new_folio) 6008d0ce0e47SSidhartha Kumar folio_set_hugetlb_migratable(folio); 6009cb6acd01SMike Kravetz 6010d0ce0e47SSidhartha Kumar folio_unlock(folio); 60114c887265SAdam Litke out: 6012958f32ceSLiu Shixin hugetlb_vma_unlock_read(vma); 6013958f32ceSLiu Shixin mutex_unlock(&hugetlb_fault_mutex_table[hash]); 6014ac9b9c66SHugh Dickins return ret; 60154c887265SAdam Litke 60164c887265SAdam Litke backout: 6017cb900f41SKirill A. Shutemov spin_unlock(ptl); 60182b26736cSAndy Whitcroft backout_unlocked: 6019d0ce0e47SSidhartha Kumar if (new_folio && !new_pagecache_folio) 6020d2d7bb44SSidhartha Kumar restore_reserve_on_error(h, vma, haddr, folio); 6021fa27759aSMike Kravetz 6022d0ce0e47SSidhartha Kumar folio_unlock(folio); 6023d0ce0e47SSidhartha Kumar folio_put(folio); 60244c887265SAdam Litke goto out; 6025ac9b9c66SHugh Dickins } 6026ac9b9c66SHugh Dickins 60278382d914SDavidlohr Bueso #ifdef CONFIG_SMP 6028188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 60298382d914SDavidlohr Bueso { 60308382d914SDavidlohr Bueso unsigned long key[2]; 60318382d914SDavidlohr Bueso u32 hash; 60328382d914SDavidlohr Bueso 60338382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 60348382d914SDavidlohr Bueso key[1] = idx; 60358382d914SDavidlohr Bueso 603655254636SMike Kravetz hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); 60378382d914SDavidlohr Bueso 60388382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 60398382d914SDavidlohr Bueso } 60408382d914SDavidlohr Bueso #else 60418382d914SDavidlohr Bueso /* 60426c26d310SMiaohe Lin * For uniprocessor systems we always use a single mutex, so just 60438382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 60448382d914SDavidlohr Bueso */ 6045188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 60468382d914SDavidlohr Bueso { 60478382d914SDavidlohr Bueso return 0; 60488382d914SDavidlohr Bueso } 60498382d914SDavidlohr Bueso #endif 60508382d914SDavidlohr Bueso 60512b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 6052788c7df4SHugh Dickins unsigned long address, unsigned int flags) 605386e5216fSAdam Litke { 60548382d914SDavidlohr Bueso pte_t *ptep, entry; 6055cb900f41SKirill A. Shutemov spinlock_t *ptl; 60562b740303SSouptick Joarder vm_fault_t ret; 60578382d914SDavidlohr Bueso u32 hash; 60588382d914SDavidlohr Bueso pgoff_t idx; 6059061e62e8SZhangPeng struct folio *folio = NULL; 6060371607a3SSidhartha Kumar struct folio *pagecache_folio = NULL; 6061a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 60628382d914SDavidlohr Bueso struct address_space *mapping; 60630f792cf9SNaoya Horiguchi int need_wait_lock = 0; 6064285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 606586e5216fSAdam Litke 60663935baa9SDavid Gibson /* 60673935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 60683935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 60693935baa9SDavid Gibson * the same page in the page cache. 60703935baa9SDavid Gibson */ 607140549ba8SMike Kravetz mapping = vma->vm_file->f_mapping; 607240549ba8SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 6073188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 6074c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 60758382d914SDavidlohr Bueso 607640549ba8SMike Kravetz /* 607740549ba8SMike Kravetz * Acquire vma lock before calling huge_pte_alloc and hold 607840549ba8SMike Kravetz * until finished with ptep. This prevents huge_pmd_unshare from 607940549ba8SMike Kravetz * being called elsewhere and making the ptep no longer valid. 608040549ba8SMike Kravetz */ 608140549ba8SMike Kravetz hugetlb_vma_lock_read(vma); 608240549ba8SMike Kravetz ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); 608340549ba8SMike Kravetz if (!ptep) { 608440549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 608540549ba8SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 608640549ba8SMike Kravetz return VM_FAULT_OOM; 608740549ba8SMike Kravetz } 608840549ba8SMike Kravetz 60897f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 6090af19487fSAxel Rasmussen if (huge_pte_none_mostly(entry)) { 6091af19487fSAxel Rasmussen if (is_pte_marker(entry)) { 6092af19487fSAxel Rasmussen pte_marker marker = 6093af19487fSAxel Rasmussen pte_marker_get(pte_to_swp_entry(entry)); 6094af19487fSAxel Rasmussen 6095af19487fSAxel Rasmussen if (marker & PTE_MARKER_POISONED) { 6096af19487fSAxel Rasmussen ret = VM_FAULT_HWPOISON_LARGE; 6097af19487fSAxel Rasmussen goto out_mutex; 6098af19487fSAxel Rasmussen } 6099af19487fSAxel Rasmussen } 6100af19487fSAxel Rasmussen 6101958f32ceSLiu Shixin /* 6102af19487fSAxel Rasmussen * Other PTE markers should be handled the same way as none PTE. 6103af19487fSAxel Rasmussen * 6104958f32ceSLiu Shixin * hugetlb_no_page will drop vma lock and hugetlb fault 6105958f32ceSLiu Shixin * mutex internally, which make us return immediately. 6106958f32ceSLiu Shixin */ 6107958f32ceSLiu Shixin return hugetlb_no_page(mm, vma, mapping, idx, address, ptep, 6108c64e912cSPeter Xu entry, flags); 6109af19487fSAxel Rasmussen } 611086e5216fSAdam Litke 611183c54070SNick Piggin ret = 0; 61121e8f889bSDavid Gibson 611357303d80SAndy Whitcroft /* 61140f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 61150f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 61167c8de358SEthon Paul * an active hugepage in pagecache. This goto expects the 2nd page 61177c8de358SEthon Paul * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will 61187c8de358SEthon Paul * properly handle it. 61190f792cf9SNaoya Horiguchi */ 6120fcd48540SPeter Xu if (!pte_present(entry)) { 6121fcd48540SPeter Xu if (unlikely(is_hugetlb_entry_migration(entry))) { 6122fcd48540SPeter Xu /* 6123fcd48540SPeter Xu * Release the hugetlb fault lock now, but retain 6124fcd48540SPeter Xu * the vma lock, because it is needed to guard the 6125fcd48540SPeter Xu * huge_pte_lockptr() later in 6126fcd48540SPeter Xu * migration_entry_wait_huge(). The vma lock will 6127fcd48540SPeter Xu * be released there. 6128fcd48540SPeter Xu */ 6129fcd48540SPeter Xu mutex_unlock(&hugetlb_fault_mutex_table[hash]); 6130fcd48540SPeter Xu migration_entry_wait_huge(vma, ptep); 6131fcd48540SPeter Xu return 0; 6132fcd48540SPeter Xu } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 6133fcd48540SPeter Xu ret = VM_FAULT_HWPOISON_LARGE | 6134fcd48540SPeter Xu VM_FAULT_SET_HINDEX(hstate_index(h)); 61350f792cf9SNaoya Horiguchi goto out_mutex; 6136fcd48540SPeter Xu } 61370f792cf9SNaoya Horiguchi 61380f792cf9SNaoya Horiguchi /* 6139c89357e2SDavid Hildenbrand * If we are going to COW/unshare the mapping later, we examine the 6140c89357e2SDavid Hildenbrand * pending reservations for this page now. This will ensure that any 614157303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 61421d8d1464SDavid Hildenbrand * spinlock. Also lookup the pagecache page now as it is used to 61431d8d1464SDavid Hildenbrand * determine if a reservation has been consumed. 614457303d80SAndy Whitcroft */ 6145c89357e2SDavid Hildenbrand if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) && 61461d8d1464SDavid Hildenbrand !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(entry)) { 6147285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 61482b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 6149b4d1d99fSDavid Gibson goto out_mutex; 61502b26736cSAndy Whitcroft } 61515e911373SMike Kravetz /* Just decrements count, does not deallocate */ 6152285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 615357303d80SAndy Whitcroft 6154371607a3SSidhartha Kumar pagecache_folio = filemap_lock_folio(mapping, idx); 615566dabbb6SChristoph Hellwig if (IS_ERR(pagecache_folio)) 615666dabbb6SChristoph Hellwig pagecache_folio = NULL; 615757303d80SAndy Whitcroft } 615857303d80SAndy Whitcroft 61590f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 61600fe6e20bSNaoya Horiguchi 6161c89357e2SDavid Hildenbrand /* Check for a racing update before calling hugetlb_wp() */ 6162b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 6163cb900f41SKirill A. Shutemov goto out_ptl; 6164b4d1d99fSDavid Gibson 6165166f3eccSPeter Xu /* Handle userfault-wp first, before trying to lock more pages */ 6166166f3eccSPeter Xu if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) && 6167166f3eccSPeter Xu (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 6168166f3eccSPeter Xu struct vm_fault vmf = { 6169166f3eccSPeter Xu .vma = vma, 6170166f3eccSPeter Xu .address = haddr, 6171166f3eccSPeter Xu .real_address = address, 6172166f3eccSPeter Xu .flags = flags, 6173166f3eccSPeter Xu }; 6174166f3eccSPeter Xu 6175166f3eccSPeter Xu spin_unlock(ptl); 6176371607a3SSidhartha Kumar if (pagecache_folio) { 6177371607a3SSidhartha Kumar folio_unlock(pagecache_folio); 6178371607a3SSidhartha Kumar folio_put(pagecache_folio); 6179166f3eccSPeter Xu } 618040549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 6181166f3eccSPeter Xu mutex_unlock(&hugetlb_fault_mutex_table[hash]); 6182166f3eccSPeter Xu return handle_userfault(&vmf, VM_UFFD_WP); 6183166f3eccSPeter Xu } 6184166f3eccSPeter Xu 61850f792cf9SNaoya Horiguchi /* 6186c89357e2SDavid Hildenbrand * hugetlb_wp() requires page locks of pte_page(entry) and 6187371607a3SSidhartha Kumar * pagecache_folio, so here we need take the former one 6188061e62e8SZhangPeng * when folio != pagecache_folio or !pagecache_folio. 61890f792cf9SNaoya Horiguchi */ 6190061e62e8SZhangPeng folio = page_folio(pte_page(entry)); 6191061e62e8SZhangPeng if (folio != pagecache_folio) 6192061e62e8SZhangPeng if (!folio_trylock(folio)) { 61930f792cf9SNaoya Horiguchi need_wait_lock = 1; 61940f792cf9SNaoya Horiguchi goto out_ptl; 61950f792cf9SNaoya Horiguchi } 61960f792cf9SNaoya Horiguchi 6197061e62e8SZhangPeng folio_get(folio); 6198b4d1d99fSDavid Gibson 6199c89357e2SDavid Hildenbrand if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) { 6200106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 6201c89357e2SDavid Hildenbrand ret = hugetlb_wp(mm, vma, address, ptep, flags, 6202371607a3SSidhartha Kumar pagecache_folio, ptl); 62030f792cf9SNaoya Horiguchi goto out_put_page; 6204c89357e2SDavid Hildenbrand } else if (likely(flags & FAULT_FLAG_WRITE)) { 6205106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 6206b4d1d99fSDavid Gibson } 6207c89357e2SDavid Hildenbrand } 6208b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 6209285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 6210788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 6211285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 62120f792cf9SNaoya Horiguchi out_put_page: 6213061e62e8SZhangPeng if (folio != pagecache_folio) 6214061e62e8SZhangPeng folio_unlock(folio); 6215061e62e8SZhangPeng folio_put(folio); 6216cb900f41SKirill A. Shutemov out_ptl: 6217cb900f41SKirill A. Shutemov spin_unlock(ptl); 621857303d80SAndy Whitcroft 6219371607a3SSidhartha Kumar if (pagecache_folio) { 6220371607a3SSidhartha Kumar folio_unlock(pagecache_folio); 6221371607a3SSidhartha Kumar folio_put(pagecache_folio); 622257303d80SAndy Whitcroft } 6223b4d1d99fSDavid Gibson out_mutex: 622440549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 6225c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 62260f792cf9SNaoya Horiguchi /* 62270f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 62280f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 62290f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 62300f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 62310f792cf9SNaoya Horiguchi * here without taking refcount. 62320f792cf9SNaoya Horiguchi */ 62330f792cf9SNaoya Horiguchi if (need_wait_lock) 6234061e62e8SZhangPeng folio_wait_locked(folio); 62351e8f889bSDavid Gibson return ret; 623686e5216fSAdam Litke } 623786e5216fSAdam Litke 6238714c1891SAxel Rasmussen #ifdef CONFIG_USERFAULTFD 62398fb5debcSMike Kravetz /* 6240a734991cSAxel Rasmussen * Used by userfaultfd UFFDIO_* ioctls. Based on userfaultfd's mfill_atomic_pte 6241a734991cSAxel Rasmussen * with modifications for hugetlb pages. 62428fb5debcSMike Kravetz */ 624361c50040SAxel Rasmussen int hugetlb_mfill_atomic_pte(pte_t *dst_pte, 62448fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 62458fb5debcSMike Kravetz unsigned long dst_addr, 62468fb5debcSMike Kravetz unsigned long src_addr, 6247d9712937SAxel Rasmussen uffd_flags_t flags, 62480169fd51SZhangPeng struct folio **foliop) 62498fb5debcSMike Kravetz { 625061c50040SAxel Rasmussen struct mm_struct *dst_mm = dst_vma->vm_mm; 6251d9712937SAxel Rasmussen bool is_continue = uffd_flags_mode_is(flags, MFILL_ATOMIC_CONTINUE); 6252d9712937SAxel Rasmussen bool wp_enabled = (flags & MFILL_ATOMIC_WP); 62538cc5fcbbSMina Almasry struct hstate *h = hstate_vma(dst_vma); 62548cc5fcbbSMina Almasry struct address_space *mapping = dst_vma->vm_file->f_mapping; 62558cc5fcbbSMina Almasry pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr); 62561e392147SAndrea Arcangeli unsigned long size; 62571c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 62588fb5debcSMike Kravetz pte_t _dst_pte; 62598fb5debcSMike Kravetz spinlock_t *ptl; 62608cc5fcbbSMina Almasry int ret = -ENOMEM; 6261d0ce0e47SSidhartha Kumar struct folio *folio; 6262f6191471SAxel Rasmussen int writable; 6263d0ce0e47SSidhartha Kumar bool folio_in_pagecache = false; 62648fb5debcSMike Kravetz 6265*8a13897fSAxel Rasmussen if (uffd_flags_mode_is(flags, MFILL_ATOMIC_POISON)) { 6266*8a13897fSAxel Rasmussen ptl = huge_pte_lock(h, dst_mm, dst_pte); 6267*8a13897fSAxel Rasmussen 6268*8a13897fSAxel Rasmussen /* Don't overwrite any existing PTEs (even markers) */ 6269*8a13897fSAxel Rasmussen if (!huge_pte_none(huge_ptep_get(dst_pte))) { 6270*8a13897fSAxel Rasmussen spin_unlock(ptl); 6271*8a13897fSAxel Rasmussen return -EEXIST; 6272*8a13897fSAxel Rasmussen } 6273*8a13897fSAxel Rasmussen 6274*8a13897fSAxel Rasmussen _dst_pte = make_pte_marker(PTE_MARKER_POISONED); 6275*8a13897fSAxel Rasmussen set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 6276*8a13897fSAxel Rasmussen 6277*8a13897fSAxel Rasmussen /* No need to invalidate - it was non-present before */ 6278*8a13897fSAxel Rasmussen update_mmu_cache(dst_vma, dst_addr, dst_pte); 6279*8a13897fSAxel Rasmussen 6280*8a13897fSAxel Rasmussen spin_unlock(ptl); 6281*8a13897fSAxel Rasmussen return 0; 6282*8a13897fSAxel Rasmussen } 6283*8a13897fSAxel Rasmussen 6284f6191471SAxel Rasmussen if (is_continue) { 6285f6191471SAxel Rasmussen ret = -EFAULT; 6286d0ce0e47SSidhartha Kumar folio = filemap_lock_folio(mapping, idx); 628766dabbb6SChristoph Hellwig if (IS_ERR(folio)) 6288f6191471SAxel Rasmussen goto out; 6289d0ce0e47SSidhartha Kumar folio_in_pagecache = true; 62900169fd51SZhangPeng } else if (!*foliop) { 62910169fd51SZhangPeng /* If a folio already exists, then it's UFFDIO_COPY for 6292d84cf06eSMina Almasry * a non-missing case. Return -EEXIST. 6293d84cf06eSMina Almasry */ 6294d84cf06eSMina Almasry if (vm_shared && 6295d84cf06eSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 6296d84cf06eSMina Almasry ret = -EEXIST; 62978fb5debcSMike Kravetz goto out; 6298d84cf06eSMina Almasry } 6299d84cf06eSMina Almasry 6300d0ce0e47SSidhartha Kumar folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0); 6301d0ce0e47SSidhartha Kumar if (IS_ERR(folio)) { 6302d84cf06eSMina Almasry ret = -ENOMEM; 6303d84cf06eSMina Almasry goto out; 6304d84cf06eSMina Almasry } 63058fb5debcSMike Kravetz 6306e87340caSZhangPeng ret = copy_folio_from_user(folio, (const void __user *) src_addr, 6307e87340caSZhangPeng false); 63088fb5debcSMike Kravetz 6309c1e8d7c6SMichel Lespinasse /* fallback to copy_from_user outside mmap_lock */ 63108fb5debcSMike Kravetz if (unlikely(ret)) { 63119e368259SAndrea Arcangeli ret = -ENOENT; 6312d0ce0e47SSidhartha Kumar /* Free the allocated folio which may have 63138cc5fcbbSMina Almasry * consumed a reservation. 63148cc5fcbbSMina Almasry */ 6315d2d7bb44SSidhartha Kumar restore_reserve_on_error(h, dst_vma, dst_addr, folio); 6316d0ce0e47SSidhartha Kumar folio_put(folio); 63178cc5fcbbSMina Almasry 6318d0ce0e47SSidhartha Kumar /* Allocate a temporary folio to hold the copied 63198cc5fcbbSMina Almasry * contents. 63208cc5fcbbSMina Almasry */ 6321d0ce0e47SSidhartha Kumar folio = alloc_hugetlb_folio_vma(h, dst_vma, dst_addr); 6322d0ce0e47SSidhartha Kumar if (!folio) { 63238cc5fcbbSMina Almasry ret = -ENOMEM; 63248cc5fcbbSMina Almasry goto out; 63258cc5fcbbSMina Almasry } 63260169fd51SZhangPeng *foliop = folio; 63270169fd51SZhangPeng /* Set the outparam foliop and return to the caller to 63288cc5fcbbSMina Almasry * copy the contents outside the lock. Don't free the 63290169fd51SZhangPeng * folio. 63308cc5fcbbSMina Almasry */ 63318fb5debcSMike Kravetz goto out; 63328fb5debcSMike Kravetz } 63338fb5debcSMike Kravetz } else { 63348cc5fcbbSMina Almasry if (vm_shared && 63358cc5fcbbSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 63360169fd51SZhangPeng folio_put(*foliop); 63378cc5fcbbSMina Almasry ret = -EEXIST; 63380169fd51SZhangPeng *foliop = NULL; 63398cc5fcbbSMina Almasry goto out; 63408cc5fcbbSMina Almasry } 63418cc5fcbbSMina Almasry 6342d0ce0e47SSidhartha Kumar folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0); 6343d0ce0e47SSidhartha Kumar if (IS_ERR(folio)) { 63440169fd51SZhangPeng folio_put(*foliop); 63458cc5fcbbSMina Almasry ret = -ENOMEM; 63460169fd51SZhangPeng *foliop = NULL; 63478cc5fcbbSMina Almasry goto out; 63488cc5fcbbSMina Almasry } 63491cb9dc4bSLiu Shixin ret = copy_user_large_folio(folio, *foliop, dst_addr, dst_vma); 63500169fd51SZhangPeng folio_put(*foliop); 63510169fd51SZhangPeng *foliop = NULL; 63521cb9dc4bSLiu Shixin if (ret) { 63531cb9dc4bSLiu Shixin folio_put(folio); 63541cb9dc4bSLiu Shixin goto out; 63551cb9dc4bSLiu Shixin } 63568fb5debcSMike Kravetz } 63578fb5debcSMike Kravetz 63588fb5debcSMike Kravetz /* 6359d0ce0e47SSidhartha Kumar * The memory barrier inside __folio_mark_uptodate makes sure that 63608fb5debcSMike Kravetz * preceding stores to the page contents become visible before 63618fb5debcSMike Kravetz * the set_pte_at() write. 63628fb5debcSMike Kravetz */ 6363d0ce0e47SSidhartha Kumar __folio_mark_uptodate(folio); 63648fb5debcSMike Kravetz 6365f6191471SAxel Rasmussen /* Add shared, newly allocated pages to the page cache. */ 6366f6191471SAxel Rasmussen if (vm_shared && !is_continue) { 63671e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 63681e392147SAndrea Arcangeli ret = -EFAULT; 63691e392147SAndrea Arcangeli if (idx >= size) 63701e392147SAndrea Arcangeli goto out_release_nounlock; 63711c9e8defSMike Kravetz 63721e392147SAndrea Arcangeli /* 63731e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 63747e1813d4SMike Kravetz * hugetlb_add_to_page_cache() below happens through the 63751e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 63761e392147SAndrea Arcangeli * the caller. 63771e392147SAndrea Arcangeli */ 63789b91c0e2SSidhartha Kumar ret = hugetlb_add_to_page_cache(folio, mapping, idx); 63791c9e8defSMike Kravetz if (ret) 63801c9e8defSMike Kravetz goto out_release_nounlock; 6381d0ce0e47SSidhartha Kumar folio_in_pagecache = true; 63821c9e8defSMike Kravetz } 63831c9e8defSMike Kravetz 6384bcc66543SMiaohe Lin ptl = huge_pte_lock(h, dst_mm, dst_pte); 63858fb5debcSMike Kravetz 63868625147cSJames Houghton ret = -EIO; 6387d0ce0e47SSidhartha Kumar if (folio_test_hwpoison(folio)) 63888625147cSJames Houghton goto out_release_unlock; 63898625147cSJames Houghton 63901e392147SAndrea Arcangeli /* 63916041c691SPeter Xu * We allow to overwrite a pte marker: consider when both MISSING|WP 63926041c691SPeter Xu * registered, we firstly wr-protect a none pte which has no page cache 63936041c691SPeter Xu * page backing it, then access the page. 63946041c691SPeter Xu */ 6395fa27759aSMike Kravetz ret = -EEXIST; 63966041c691SPeter Xu if (!huge_pte_none_mostly(huge_ptep_get(dst_pte))) 63978fb5debcSMike Kravetz goto out_release_unlock; 63988fb5debcSMike Kravetz 6399d0ce0e47SSidhartha Kumar if (folio_in_pagecache) 6400d0ce0e47SSidhartha Kumar page_dup_file_rmap(&folio->page, true); 64014781593dSPeter Xu else 6402d0ce0e47SSidhartha Kumar hugepage_add_new_anon_rmap(folio, dst_vma, dst_addr); 64038fb5debcSMike Kravetz 64046041c691SPeter Xu /* 64056041c691SPeter Xu * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY 64066041c691SPeter Xu * with wp flag set, don't set pte write bit. 64076041c691SPeter Xu */ 6408d9712937SAxel Rasmussen if (wp_enabled || (is_continue && !vm_shared)) 6409f6191471SAxel Rasmussen writable = 0; 6410f6191471SAxel Rasmussen else 6411f6191471SAxel Rasmussen writable = dst_vma->vm_flags & VM_WRITE; 6412f6191471SAxel Rasmussen 6413d0ce0e47SSidhartha Kumar _dst_pte = make_huge_pte(dst_vma, &folio->page, writable); 64146041c691SPeter Xu /* 64156041c691SPeter Xu * Always mark UFFDIO_COPY page dirty; note that this may not be 64166041c691SPeter Xu * extremely important for hugetlbfs for now since swapping is not 64176041c691SPeter Xu * supported, but we should still be clear in that this page cannot be 64186041c691SPeter Xu * thrown away at will, even if write bit not set. 64196041c691SPeter Xu */ 64208fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 64218fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 64228fb5debcSMike Kravetz 6423d9712937SAxel Rasmussen if (wp_enabled) 64246041c691SPeter Xu _dst_pte = huge_pte_mkuffd_wp(_dst_pte); 64256041c691SPeter Xu 64268fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 64278fb5debcSMike Kravetz 64288fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 64298fb5debcSMike Kravetz 64308fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 64318fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 64328fb5debcSMike Kravetz 64338fb5debcSMike Kravetz spin_unlock(ptl); 6434f6191471SAxel Rasmussen if (!is_continue) 6435d0ce0e47SSidhartha Kumar folio_set_hugetlb_migratable(folio); 6436f6191471SAxel Rasmussen if (vm_shared || is_continue) 6437d0ce0e47SSidhartha Kumar folio_unlock(folio); 64388fb5debcSMike Kravetz ret = 0; 64398fb5debcSMike Kravetz out: 64408fb5debcSMike Kravetz return ret; 64418fb5debcSMike Kravetz out_release_unlock: 64428fb5debcSMike Kravetz spin_unlock(ptl); 6443f6191471SAxel Rasmussen if (vm_shared || is_continue) 6444d0ce0e47SSidhartha Kumar folio_unlock(folio); 64455af10dfdSAndrea Arcangeli out_release_nounlock: 6446d0ce0e47SSidhartha Kumar if (!folio_in_pagecache) 6447d2d7bb44SSidhartha Kumar restore_reserve_on_error(h, dst_vma, dst_addr, folio); 6448d0ce0e47SSidhartha Kumar folio_put(folio); 64498fb5debcSMike Kravetz goto out; 64508fb5debcSMike Kravetz } 6451714c1891SAxel Rasmussen #endif /* CONFIG_USERFAULTFD */ 64528fb5debcSMike Kravetz 645357a196a5SMike Kravetz struct page *hugetlb_follow_page_mask(struct vm_area_struct *vma, 64545502ea44SPeter Xu unsigned long address, unsigned int flags, 64555502ea44SPeter Xu unsigned int *page_mask) 645657a196a5SMike Kravetz { 645757a196a5SMike Kravetz struct hstate *h = hstate_vma(vma); 645857a196a5SMike Kravetz struct mm_struct *mm = vma->vm_mm; 645957a196a5SMike Kravetz unsigned long haddr = address & huge_page_mask(h); 646057a196a5SMike Kravetz struct page *page = NULL; 646157a196a5SMike Kravetz spinlock_t *ptl; 646257a196a5SMike Kravetz pte_t *pte, entry; 6463458568c9SPeter Xu int ret; 646457a196a5SMike Kravetz 64657d049f3aSPeter Xu hugetlb_vma_lock_read(vma); 64669c67a207SPeter Xu pte = hugetlb_walk(vma, haddr, huge_page_size(h)); 646757a196a5SMike Kravetz if (!pte) 64687d049f3aSPeter Xu goto out_unlock; 646957a196a5SMike Kravetz 647057a196a5SMike Kravetz ptl = huge_pte_lock(h, mm, pte); 647157a196a5SMike Kravetz entry = huge_ptep_get(pte); 647257a196a5SMike Kravetz if (pte_present(entry)) { 6473458568c9SPeter Xu page = pte_page(entry); 6474458568c9SPeter Xu 6475458568c9SPeter Xu if (!huge_pte_write(entry)) { 6476458568c9SPeter Xu if (flags & FOLL_WRITE) { 6477458568c9SPeter Xu page = NULL; 6478458568c9SPeter Xu goto out; 6479458568c9SPeter Xu } 6480458568c9SPeter Xu 6481458568c9SPeter Xu if (gup_must_unshare(vma, flags, page)) { 6482458568c9SPeter Xu /* Tell the caller to do unsharing */ 6483458568c9SPeter Xu page = ERR_PTR(-EMLINK); 6484458568c9SPeter Xu goto out; 6485458568c9SPeter Xu } 6486458568c9SPeter Xu } 6487458568c9SPeter Xu 6488458568c9SPeter Xu page += ((address & ~huge_page_mask(h)) >> PAGE_SHIFT); 6489458568c9SPeter Xu 649057a196a5SMike Kravetz /* 649157a196a5SMike Kravetz * Note that page may be a sub-page, and with vmemmap 649257a196a5SMike Kravetz * optimizations the page struct may be read only. 649357a196a5SMike Kravetz * try_grab_page() will increase the ref count on the 649457a196a5SMike Kravetz * head page, so this will be OK. 649557a196a5SMike Kravetz * 6496e2ca6ba6SLinus Torvalds * try_grab_page() should always be able to get the page here, 6497e2ca6ba6SLinus Torvalds * because we hold the ptl lock and have verified pte_present(). 649857a196a5SMike Kravetz */ 6499458568c9SPeter Xu ret = try_grab_page(page, flags); 6500458568c9SPeter Xu 6501458568c9SPeter Xu if (WARN_ON_ONCE(ret)) { 6502458568c9SPeter Xu page = ERR_PTR(ret); 650357a196a5SMike Kravetz goto out; 650457a196a5SMike Kravetz } 65055502ea44SPeter Xu 65065502ea44SPeter Xu *page_mask = (1U << huge_page_order(h)) - 1; 650757a196a5SMike Kravetz } 650857a196a5SMike Kravetz out: 650957a196a5SMike Kravetz spin_unlock(ptl); 65107d049f3aSPeter Xu out_unlock: 65117d049f3aSPeter Xu hugetlb_vma_unlock_read(vma); 6512dd767aaaSPeter Xu 6513dd767aaaSPeter Xu /* 6514dd767aaaSPeter Xu * Fixup retval for dump requests: if pagecache doesn't exist, 6515dd767aaaSPeter Xu * don't try to allocate a new page but just skip it. 6516dd767aaaSPeter Xu */ 6517dd767aaaSPeter Xu if (!page && (flags & FOLL_DUMP) && 6518dd767aaaSPeter Xu !hugetlbfs_pagecache_present(h, vma, address)) 6519dd767aaaSPeter Xu page = ERR_PTR(-EFAULT); 6520dd767aaaSPeter Xu 652157a196a5SMike Kravetz return page; 652257a196a5SMike Kravetz } 652357a196a5SMike Kravetz 6524a79390f5SPeter Xu long hugetlb_change_protection(struct vm_area_struct *vma, 65255a90d5a1SPeter Xu unsigned long address, unsigned long end, 65265a90d5a1SPeter Xu pgprot_t newprot, unsigned long cp_flags) 65278f860591SZhang, Yanmin { 65288f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 65298f860591SZhang, Yanmin unsigned long start = address; 65308f860591SZhang, Yanmin pte_t *ptep; 65318f860591SZhang, Yanmin pte_t pte; 6532a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 6533a79390f5SPeter Xu long pages = 0, psize = huge_page_size(h); 6534dff11abeSMike Kravetz bool shared_pmd = false; 6535ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 6536e95a9851SMike Kravetz unsigned long last_addr_mask; 65375a90d5a1SPeter Xu bool uffd_wp = cp_flags & MM_CP_UFFD_WP; 65385a90d5a1SPeter Xu bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE; 6539dff11abeSMike Kravetz 6540dff11abeSMike Kravetz /* 6541dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 6542ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 6543dff11abeSMike Kravetz * range if PMD sharing is possible. 6544dff11abeSMike Kravetz */ 65457269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 65467d4a8be0SAlistair Popple 0, mm, start, end); 6547ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 65488f860591SZhang, Yanmin 65498f860591SZhang, Yanmin BUG_ON(address >= end); 6550ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 65518f860591SZhang, Yanmin 6552ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 655340549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 655483cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 655540549ba8SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 655660dfaad6SPeter Xu for (; address < end; address += psize) { 6557cb900f41SKirill A. Shutemov spinlock_t *ptl; 65589c67a207SPeter Xu ptep = hugetlb_walk(vma, address, psize); 6559e95a9851SMike Kravetz if (!ptep) { 6560fed15f13SPeter Xu if (!uffd_wp) { 6561e95a9851SMike Kravetz address |= last_addr_mask; 65628f860591SZhang, Yanmin continue; 6563e95a9851SMike Kravetz } 6564fed15f13SPeter Xu /* 6565fed15f13SPeter Xu * Userfaultfd wr-protect requires pgtable 6566fed15f13SPeter Xu * pre-allocations to install pte markers. 6567fed15f13SPeter Xu */ 6568fed15f13SPeter Xu ptep = huge_pte_alloc(mm, vma, address, psize); 6569d1751118SPeter Xu if (!ptep) { 6570d1751118SPeter Xu pages = -ENOMEM; 6571fed15f13SPeter Xu break; 6572fed15f13SPeter Xu } 6573d1751118SPeter Xu } 6574cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 65754ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, address, ptep)) { 657660dfaad6SPeter Xu /* 657760dfaad6SPeter Xu * When uffd-wp is enabled on the vma, unshare 657860dfaad6SPeter Xu * shouldn't happen at all. Warn about it if it 657960dfaad6SPeter Xu * happened due to some reason. 658060dfaad6SPeter Xu */ 658160dfaad6SPeter Xu WARN_ON_ONCE(uffd_wp || uffd_wp_resolve); 65827da4d641SPeter Zijlstra pages++; 6583cb900f41SKirill A. Shutemov spin_unlock(ptl); 6584dff11abeSMike Kravetz shared_pmd = true; 65854ddb4d91SMike Kravetz address |= last_addr_mask; 658639dde65cSChen, Kenneth W continue; 65877da4d641SPeter Zijlstra } 6588a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 6589a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 65900e678153SDavid Hildenbrand /* Nothing to do. */ 65910e678153SDavid Hildenbrand } else if (unlikely(is_hugetlb_entry_migration(pte))) { 6592a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 65936c287605SDavid Hildenbrand struct page *page = pfn_swap_entry_to_page(entry); 659444f86392SDavid Hildenbrand pte_t newpte = pte; 6595a8bda28dSNaoya Horiguchi 659644f86392SDavid Hildenbrand if (is_writable_migration_entry(entry)) { 65976c287605SDavid Hildenbrand if (PageAnon(page)) 65986c287605SDavid Hildenbrand entry = make_readable_exclusive_migration_entry( 65996c287605SDavid Hildenbrand swp_offset(entry)); 66006c287605SDavid Hildenbrand else 66014dd845b5SAlistair Popple entry = make_readable_migration_entry( 66024dd845b5SAlistair Popple swp_offset(entry)); 6603a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 660444f86392SDavid Hildenbrand pages++; 660544f86392SDavid Hildenbrand } 660644f86392SDavid Hildenbrand 66075a90d5a1SPeter Xu if (uffd_wp) 66085a90d5a1SPeter Xu newpte = pte_swp_mkuffd_wp(newpte); 66095a90d5a1SPeter Xu else if (uffd_wp_resolve) 66105a90d5a1SPeter Xu newpte = pte_swp_clear_uffd_wp(newpte); 661144f86392SDavid Hildenbrand if (!pte_same(pte, newpte)) 661218f39629SQi Zheng set_huge_pte_at(mm, address, ptep, newpte); 66130e678153SDavid Hildenbrand } else if (unlikely(is_pte_marker(pte))) { 66140e678153SDavid Hildenbrand /* No other markers apply for now. */ 66150e678153SDavid Hildenbrand WARN_ON_ONCE(!pte_marker_uffd_wp(pte)); 661660dfaad6SPeter Xu if (uffd_wp_resolve) 66170e678153SDavid Hildenbrand /* Safe to modify directly (non-present->none). */ 661860dfaad6SPeter Xu huge_pte_clear(mm, address, ptep, psize); 66190e678153SDavid Hildenbrand } else if (!huge_pte_none(pte)) { 6620023bdd00SAneesh Kumar K.V pte_t old_pte; 662179c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 6622023bdd00SAneesh Kumar K.V 6623023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 662416785bd7SAnshuman Khandual pte = huge_pte_modify(old_pte, newprot); 662579c1c594SChristophe Leroy pte = arch_make_huge_pte(pte, shift, vma->vm_flags); 66265a90d5a1SPeter Xu if (uffd_wp) 6627f1eb1bacSPeter Xu pte = huge_pte_mkuffd_wp(pte); 66285a90d5a1SPeter Xu else if (uffd_wp_resolve) 66295a90d5a1SPeter Xu pte = huge_pte_clear_uffd_wp(pte); 6630023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 66317da4d641SPeter Zijlstra pages++; 663260dfaad6SPeter Xu } else { 663360dfaad6SPeter Xu /* None pte */ 663460dfaad6SPeter Xu if (unlikely(uffd_wp)) 663560dfaad6SPeter Xu /* Safe to modify directly (none->non-present). */ 663660dfaad6SPeter Xu set_huge_pte_at(mm, address, ptep, 663760dfaad6SPeter Xu make_pte_marker(PTE_MARKER_UFFD_WP)); 66388f860591SZhang, Yanmin } 6639cb900f41SKirill A. Shutemov spin_unlock(ptl); 66408f860591SZhang, Yanmin } 6641d833352aSMel Gorman /* 6642c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 6643d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 6644c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 6645dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 6646dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 6647d833352aSMel Gorman */ 6648dff11abeSMike Kravetz if (shared_pmd) 6649ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 6650dff11abeSMike Kravetz else 66515491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 66520f10851eSJérôme Glisse /* 66530f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 66540f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 66550f10851eSJérôme Glisse * 6656ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 66570f10851eSJérôme Glisse */ 665883cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 665940549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 6660ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 66617da4d641SPeter Zijlstra 6662d1751118SPeter Xu return pages > 0 ? (pages << h->order) : pages; 66638f860591SZhang, Yanmin } 66648f860591SZhang, Yanmin 666533b8f84aSMike Kravetz /* Return true if reservation was successful, false otherwise. */ 666633b8f84aSMike Kravetz bool hugetlb_reserve_pages(struct inode *inode, 6667a1e78772SMel Gorman long from, long to, 66685a6fe125SMel Gorman struct vm_area_struct *vma, 6669ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 6670e4e574b7SAdam Litke { 6671c5094ec7SMike Kravetz long chg = -1, add = -1; 6672a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 667390481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 66749119a41eSJoonsoo Kim struct resv_map *resv_map; 6675075a61d0SMina Almasry struct hugetlb_cgroup *h_cg = NULL; 66760db9d74eSMina Almasry long gbl_reserve, regions_needed = 0; 6677e4e574b7SAdam Litke 667863489f8eSMike Kravetz /* This should never happen */ 667963489f8eSMike Kravetz if (from > to) { 668063489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 668133b8f84aSMike Kravetz return false; 668263489f8eSMike Kravetz } 668363489f8eSMike Kravetz 6684a1e78772SMel Gorman /* 6685e700898fSMike Kravetz * vma specific semaphore used for pmd sharing and fault/truncation 6686e700898fSMike Kravetz * synchronization 66878d9bfb26SMike Kravetz */ 66888d9bfb26SMike Kravetz hugetlb_vma_lock_alloc(vma); 66898d9bfb26SMike Kravetz 66908d9bfb26SMike Kravetz /* 669117c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 669217c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 669390481622SDavid Gibson * without using reserves 669417c9d12eSMel Gorman */ 6695ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 669633b8f84aSMike Kravetz return true; 669717c9d12eSMel Gorman 669817c9d12eSMel Gorman /* 6699a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 6700a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 6701a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 6702a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 6703a1e78772SMel Gorman */ 67049119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 6705f27a5136SMike Kravetz /* 6706f27a5136SMike Kravetz * resv_map can not be NULL as hugetlb_reserve_pages is only 6707f27a5136SMike Kravetz * called for inodes for which resv_maps were created (see 6708f27a5136SMike Kravetz * hugetlbfs_get_inode). 6709f27a5136SMike Kravetz */ 67104e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 67119119a41eSJoonsoo Kim 67120db9d74eSMina Almasry chg = region_chg(resv_map, from, to, ®ions_needed); 67139119a41eSJoonsoo Kim } else { 6714e9fe92aeSMina Almasry /* Private mapping. */ 67159119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 67165a6fe125SMel Gorman if (!resv_map) 67178d9bfb26SMike Kravetz goto out_err; 67185a6fe125SMel Gorman 671917c9d12eSMel Gorman chg = to - from; 672017c9d12eSMel Gorman 67215a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 67225a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 67235a6fe125SMel Gorman } 67245a6fe125SMel Gorman 672533b8f84aSMike Kravetz if (chg < 0) 6726c50ac050SDave Hansen goto out_err; 672717c9d12eSMel Gorman 672833b8f84aSMike Kravetz if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), 672933b8f84aSMike Kravetz chg * pages_per_huge_page(h), &h_cg) < 0) 6730075a61d0SMina Almasry goto out_err; 6731075a61d0SMina Almasry 6732075a61d0SMina Almasry if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { 6733075a61d0SMina Almasry /* For private mappings, the hugetlb_cgroup uncharge info hangs 6734075a61d0SMina Almasry * of the resv_map. 6735075a61d0SMina Almasry */ 6736075a61d0SMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); 6737075a61d0SMina Almasry } 6738075a61d0SMina Almasry 67391c5ecae3SMike Kravetz /* 67401c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 67411c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 67421c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 67431c5ecae3SMike Kravetz */ 67441c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 674533b8f84aSMike Kravetz if (gbl_reserve < 0) 6746075a61d0SMina Almasry goto out_uncharge_cgroup; 674717c9d12eSMel Gorman 674817c9d12eSMel Gorman /* 674917c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 675090481622SDavid Gibson * Hand the pages back to the subpool if there are not 675117c9d12eSMel Gorman */ 675233b8f84aSMike Kravetz if (hugetlb_acct_memory(h, gbl_reserve) < 0) 6753075a61d0SMina Almasry goto out_put_pages; 675417c9d12eSMel Gorman 675517c9d12eSMel Gorman /* 675617c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 675717c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 675817c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 675917c9d12eSMel Gorman * the reservation was and the page cache tells how much of 676017c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 676117c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 676217c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 676317c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 676417c9d12eSMel Gorman * else has to be done for private mappings here 676517c9d12eSMel Gorman */ 676633039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 6767075a61d0SMina Almasry add = region_add(resv_map, from, to, regions_needed, h, h_cg); 676833039678SMike Kravetz 67690db9d74eSMina Almasry if (unlikely(add < 0)) { 67700db9d74eSMina Almasry hugetlb_acct_memory(h, -gbl_reserve); 6771075a61d0SMina Almasry goto out_put_pages; 67720db9d74eSMina Almasry } else if (unlikely(chg > add)) { 677333039678SMike Kravetz /* 677433039678SMike Kravetz * pages in this range were added to the reserve 677533039678SMike Kravetz * map between region_chg and region_add. This 6776d0ce0e47SSidhartha Kumar * indicates a race with alloc_hugetlb_folio. Adjust 677733039678SMike Kravetz * the subpool and reserve counts modified above 677833039678SMike Kravetz * based on the difference. 677933039678SMike Kravetz */ 678033039678SMike Kravetz long rsv_adjust; 678133039678SMike Kravetz 6782d85aecf2SMiaohe Lin /* 6783d85aecf2SMiaohe Lin * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the 6784d85aecf2SMiaohe Lin * reference to h_cg->css. See comment below for detail. 6785d85aecf2SMiaohe Lin */ 6786075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd( 6787075a61d0SMina Almasry hstate_index(h), 6788075a61d0SMina Almasry (chg - add) * pages_per_huge_page(h), h_cg); 6789075a61d0SMina Almasry 679033039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 679133039678SMike Kravetz chg - add); 679233039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 6793d85aecf2SMiaohe Lin } else if (h_cg) { 6794d85aecf2SMiaohe Lin /* 6795d85aecf2SMiaohe Lin * The file_regions will hold their own reference to 6796d85aecf2SMiaohe Lin * h_cg->css. So we should release the reference held 6797d85aecf2SMiaohe Lin * via hugetlb_cgroup_charge_cgroup_rsvd() when we are 6798d85aecf2SMiaohe Lin * done. 6799d85aecf2SMiaohe Lin */ 6800d85aecf2SMiaohe Lin hugetlb_cgroup_put_rsvd_cgroup(h_cg); 680133039678SMike Kravetz } 680233039678SMike Kravetz } 680333b8f84aSMike Kravetz return true; 680433b8f84aSMike Kravetz 6805075a61d0SMina Almasry out_put_pages: 6806075a61d0SMina Almasry /* put back original number of pages, chg */ 6807075a61d0SMina Almasry (void)hugepage_subpool_put_pages(spool, chg); 6808075a61d0SMina Almasry out_uncharge_cgroup: 6809075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), 6810075a61d0SMina Almasry chg * pages_per_huge_page(h), h_cg); 6811c50ac050SDave Hansen out_err: 68128d9bfb26SMike Kravetz hugetlb_vma_lock_free(vma); 68135e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 68140db9d74eSMina Almasry /* Only call region_abort if the region_chg succeeded but the 68150db9d74eSMina Almasry * region_add failed or didn't run. 68160db9d74eSMina Almasry */ 68170db9d74eSMina Almasry if (chg >= 0 && add < 0) 68180db9d74eSMina Almasry region_abort(resv_map, from, to, regions_needed); 6819f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 6820f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 682133b8f84aSMike Kravetz return false; 6822a43a8c39SChen, Kenneth W } 6823a43a8c39SChen, Kenneth W 6824b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 6825b5cec28dSMike Kravetz long freed) 6826a43a8c39SChen, Kenneth W { 6827a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 68284e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 68299119a41eSJoonsoo Kim long chg = 0; 683090481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 68311c5ecae3SMike Kravetz long gbl_reserve; 683245c682a6SKen Chen 6833f27a5136SMike Kravetz /* 6834f27a5136SMike Kravetz * Since this routine can be called in the evict inode path for all 6835f27a5136SMike Kravetz * hugetlbfs inodes, resv_map could be NULL. 6836f27a5136SMike Kravetz */ 6837b5cec28dSMike Kravetz if (resv_map) { 6838b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 6839b5cec28dSMike Kravetz /* 6840b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 6841b5cec28dSMike Kravetz * must be split and another region descriptor can not be 6842b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 6843b5cec28dSMike Kravetz */ 6844b5cec28dSMike Kravetz if (chg < 0) 6845b5cec28dSMike Kravetz return chg; 6846b5cec28dSMike Kravetz } 6847b5cec28dSMike Kravetz 684845c682a6SKen Chen spin_lock(&inode->i_lock); 6849e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 685045c682a6SKen Chen spin_unlock(&inode->i_lock); 685145c682a6SKen Chen 68521c5ecae3SMike Kravetz /* 68531c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 68541c5ecae3SMike Kravetz * reservations to be released may be adjusted. 6855dddf31a4SMiaohe Lin * 6856dddf31a4SMiaohe Lin * Note that !resv_map implies freed == 0. So (chg - freed) 6857dddf31a4SMiaohe Lin * won't go negative. 68581c5ecae3SMike Kravetz */ 68591c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 68601c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 6861b5cec28dSMike Kravetz 6862b5cec28dSMike Kravetz return 0; 6863a43a8c39SChen, Kenneth W } 686493f70f90SNaoya Horiguchi 68653212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 68663212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 68673212b535SSteve Capper struct vm_area_struct *vma, 68683212b535SSteve Capper unsigned long addr, pgoff_t idx) 68693212b535SSteve Capper { 68703212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 68713212b535SSteve Capper svma->vm_start; 68723212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 68733212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 68743212b535SSteve Capper 68753212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 6876e430a95aSSuren Baghdasaryan unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED_MASK; 6877e430a95aSSuren Baghdasaryan unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED_MASK; 68783212b535SSteve Capper 68793212b535SSteve Capper /* 68803212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 68813212b535SSteve Capper * page table page. 6882131a79b4SMike Kravetz * 6883131a79b4SMike Kravetz * Also, vma_lock (vm_private_data) is required for sharing. 68843212b535SSteve Capper */ 68853212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 68863212b535SSteve Capper vm_flags != svm_flags || 6887131a79b4SMike Kravetz !range_in_vma(svma, sbase, s_end) || 6888131a79b4SMike Kravetz !svma->vm_private_data) 68893212b535SSteve Capper return 0; 68903212b535SSteve Capper 68913212b535SSteve Capper return saddr; 68923212b535SSteve Capper } 68933212b535SSteve Capper 6894bbff39ccSMike Kravetz bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 68953212b535SSteve Capper { 6896bbff39ccSMike Kravetz unsigned long start = addr & PUD_MASK; 6897bbff39ccSMike Kravetz unsigned long end = start + PUD_SIZE; 6898bbff39ccSMike Kravetz 68998d9bfb26SMike Kravetz #ifdef CONFIG_USERFAULTFD 69008d9bfb26SMike Kravetz if (uffd_disable_huge_pmd_share(vma)) 69018d9bfb26SMike Kravetz return false; 69028d9bfb26SMike Kravetz #endif 69033212b535SSteve Capper /* 69043212b535SSteve Capper * check on proper vm_flags and page table alignment 69053212b535SSteve Capper */ 69068d9bfb26SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 690731aafb45SNicholas Krause return false; 6908bbff39ccSMike Kravetz if (!vma->vm_private_data) /* vma lock required for sharing */ 69098d9bfb26SMike Kravetz return false; 69108d9bfb26SMike Kravetz if (!range_in_vma(vma, start, end)) 69118d9bfb26SMike Kravetz return false; 69128d9bfb26SMike Kravetz return true; 69138d9bfb26SMike Kravetz } 69148d9bfb26SMike Kravetz 69153212b535SSteve Capper /* 6916017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 6917017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 6918017b1660SMike Kravetz * shared pmd mappings. 6919017b1660SMike Kravetz */ 6920017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 6921017b1660SMike Kravetz unsigned long *start, unsigned long *end) 6922017b1660SMike Kravetz { 6923a1ba9da8SLi Xinhai unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), 6924a1ba9da8SLi Xinhai v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 6925017b1660SMike Kravetz 6926a1ba9da8SLi Xinhai /* 6927f0953a1bSIngo Molnar * vma needs to span at least one aligned PUD size, and the range 6928f0953a1bSIngo Molnar * must be at least partially within in. 6929a1ba9da8SLi Xinhai */ 6930a1ba9da8SLi Xinhai if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || 6931a1ba9da8SLi Xinhai (*end <= v_start) || (*start >= v_end)) 6932017b1660SMike Kravetz return; 6933017b1660SMike Kravetz 693475802ca6SPeter Xu /* Extend the range to be PUD aligned for a worst case scenario */ 6935a1ba9da8SLi Xinhai if (*start > v_start) 6936a1ba9da8SLi Xinhai *start = ALIGN_DOWN(*start, PUD_SIZE); 6937017b1660SMike Kravetz 6938a1ba9da8SLi Xinhai if (*end < v_end) 6939a1ba9da8SLi Xinhai *end = ALIGN(*end, PUD_SIZE); 6940017b1660SMike Kravetz } 6941017b1660SMike Kravetz 6942017b1660SMike Kravetz /* 69433212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 69443212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 69453212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 69463a47c54fSMike Kravetz * code much cleaner. pmd allocation is essential for the shared case because 69473a47c54fSMike Kravetz * pud has to be populated inside the same i_mmap_rwsem section - otherwise 69483a47c54fSMike Kravetz * racing tasks could either miss the sharing (see huge_pte_offset) or select a 69493a47c54fSMike Kravetz * bad pmd for sharing. 69503212b535SSteve Capper */ 6951aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 6952aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 69533212b535SSteve Capper { 69543212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 69553212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 69563212b535SSteve Capper vma->vm_pgoff; 69573212b535SSteve Capper struct vm_area_struct *svma; 69583212b535SSteve Capper unsigned long saddr; 69593212b535SSteve Capper pte_t *spte = NULL; 69603212b535SSteve Capper pte_t *pte; 69613212b535SSteve Capper 69623a47c54fSMike Kravetz i_mmap_lock_read(mapping); 69633212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 69643212b535SSteve Capper if (svma == vma) 69653212b535SSteve Capper continue; 69663212b535SSteve Capper 69673212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 69683212b535SSteve Capper if (saddr) { 69699c67a207SPeter Xu spte = hugetlb_walk(svma, saddr, 69707868a208SPunit Agrawal vma_mmu_pagesize(svma)); 69713212b535SSteve Capper if (spte) { 69723212b535SSteve Capper get_page(virt_to_page(spte)); 69733212b535SSteve Capper break; 69743212b535SSteve Capper } 69753212b535SSteve Capper } 69763212b535SSteve Capper } 69773212b535SSteve Capper 69783212b535SSteve Capper if (!spte) 69793212b535SSteve Capper goto out; 69803212b535SSteve Capper 6981349d1670SPeter Xu spin_lock(&mm->page_table_lock); 6982dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 69833212b535SSteve Capper pud_populate(mm, pud, 69843212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 6985c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 6986dc6c9a35SKirill A. Shutemov } else { 69873212b535SSteve Capper put_page(virt_to_page(spte)); 6988dc6c9a35SKirill A. Shutemov } 6989349d1670SPeter Xu spin_unlock(&mm->page_table_lock); 69903212b535SSteve Capper out: 69913212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 69923a47c54fSMike Kravetz i_mmap_unlock_read(mapping); 69933212b535SSteve Capper return pte; 69943212b535SSteve Capper } 69953212b535SSteve Capper 69963212b535SSteve Capper /* 69973212b535SSteve Capper * unmap huge page backed by shared pte. 69983212b535SSteve Capper * 69993212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 70003212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 70013212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 70023212b535SSteve Capper * 70033a47c54fSMike Kravetz * Called with page table lock held. 70043212b535SSteve Capper * 70053212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 70063212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 70073212b535SSteve Capper */ 700834ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 70094ddb4d91SMike Kravetz unsigned long addr, pte_t *ptep) 70103212b535SSteve Capper { 70114ddb4d91SMike Kravetz pgd_t *pgd = pgd_offset(mm, addr); 70124ddb4d91SMike Kravetz p4d_t *p4d = p4d_offset(pgd, addr); 70134ddb4d91SMike Kravetz pud_t *pud = pud_offset(p4d, addr); 70143212b535SSteve Capper 701534ae204fSMike Kravetz i_mmap_assert_write_locked(vma->vm_file->f_mapping); 701640549ba8SMike Kravetz hugetlb_vma_assert_locked(vma); 70173212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 70183212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 70193212b535SSteve Capper return 0; 70203212b535SSteve Capper 70213212b535SSteve Capper pud_clear(pud); 70223212b535SSteve Capper put_page(virt_to_page(ptep)); 7023dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 70243212b535SSteve Capper return 1; 70253212b535SSteve Capper } 7026c1991e07SPeter Xu 70279e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 70288d9bfb26SMike Kravetz 7029aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 7030aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 70319e5fc74cSSteve Capper { 70329e5fc74cSSteve Capper return NULL; 70339e5fc74cSSteve Capper } 7034e81f2d22SZhang Zhen 703534ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 70364ddb4d91SMike Kravetz unsigned long addr, pte_t *ptep) 7037e81f2d22SZhang Zhen { 7038e81f2d22SZhang Zhen return 0; 7039e81f2d22SZhang Zhen } 7040017b1660SMike Kravetz 7041017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 7042017b1660SMike Kravetz unsigned long *start, unsigned long *end) 7043017b1660SMike Kravetz { 7044017b1660SMike Kravetz } 7045c1991e07SPeter Xu 7046c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 7047c1991e07SPeter Xu { 7048c1991e07SPeter Xu return false; 7049c1991e07SPeter Xu } 70503212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 70513212b535SSteve Capper 70529e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 7053aec44e0fSPeter Xu pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, 70549e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 70559e5fc74cSSteve Capper { 70569e5fc74cSSteve Capper pgd_t *pgd; 7057c2febafcSKirill A. Shutemov p4d_t *p4d; 70589e5fc74cSSteve Capper pud_t *pud; 70599e5fc74cSSteve Capper pte_t *pte = NULL; 70609e5fc74cSSteve Capper 70619e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 7062f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 7063f4f0a3d8SKirill A. Shutemov if (!p4d) 7064f4f0a3d8SKirill A. Shutemov return NULL; 7065c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 70669e5fc74cSSteve Capper if (pud) { 70679e5fc74cSSteve Capper if (sz == PUD_SIZE) { 70689e5fc74cSSteve Capper pte = (pte_t *)pud; 70699e5fc74cSSteve Capper } else { 70709e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 7071c1991e07SPeter Xu if (want_pmd_share(vma, addr) && pud_none(*pud)) 7072aec44e0fSPeter Xu pte = huge_pmd_share(mm, vma, addr, pud); 70739e5fc74cSSteve Capper else 70749e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 70759e5fc74cSSteve Capper } 70769e5fc74cSSteve Capper } 7077191fcdb6SJohn Hubbard 7078191fcdb6SJohn Hubbard if (pte) { 7079191fcdb6SJohn Hubbard pte_t pteval = ptep_get_lockless(pte); 7080191fcdb6SJohn Hubbard 7081191fcdb6SJohn Hubbard BUG_ON(pte_present(pteval) && !pte_huge(pteval)); 7082191fcdb6SJohn Hubbard } 70839e5fc74cSSteve Capper 70849e5fc74cSSteve Capper return pte; 70859e5fc74cSSteve Capper } 70869e5fc74cSSteve Capper 70879b19df29SPunit Agrawal /* 70889b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 70899b19df29SPunit Agrawal * entry at address @addr 70909b19df29SPunit Agrawal * 70918ac0b81aSLi Xinhai * Return: Pointer to page table entry (PUD or PMD) for 70928ac0b81aSLi Xinhai * address @addr, or NULL if a !p*d_present() entry is encountered and the 70939b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 70949b19df29SPunit Agrawal * table. 70959b19df29SPunit Agrawal */ 70967868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 70977868a208SPunit Agrawal unsigned long addr, unsigned long sz) 70989e5fc74cSSteve Capper { 70999e5fc74cSSteve Capper pgd_t *pgd; 7100c2febafcSKirill A. Shutemov p4d_t *p4d; 71018ac0b81aSLi Xinhai pud_t *pud; 71028ac0b81aSLi Xinhai pmd_t *pmd; 71039e5fc74cSSteve Capper 71049e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 7105c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 7106c2febafcSKirill A. Shutemov return NULL; 7107c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 7108c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 7109c2febafcSKirill A. Shutemov return NULL; 71109b19df29SPunit Agrawal 7111c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 71128ac0b81aSLi Xinhai if (sz == PUD_SIZE) 71138ac0b81aSLi Xinhai /* must be pud huge, non-present or none */ 71149e5fc74cSSteve Capper return (pte_t *)pud; 71158ac0b81aSLi Xinhai if (!pud_present(*pud)) 71168ac0b81aSLi Xinhai return NULL; 71178ac0b81aSLi Xinhai /* must have a valid entry and size to go further */ 71189b19df29SPunit Agrawal 71199e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 71208ac0b81aSLi Xinhai /* must be pmd huge, non-present or none */ 71219e5fc74cSSteve Capper return (pte_t *)pmd; 71229e5fc74cSSteve Capper } 71239e5fc74cSSteve Capper 7124e95a9851SMike Kravetz /* 7125e95a9851SMike Kravetz * Return a mask that can be used to update an address to the last huge 7126e95a9851SMike Kravetz * page in a page table page mapping size. Used to skip non-present 7127e95a9851SMike Kravetz * page table entries when linearly scanning address ranges. Architectures 7128e95a9851SMike Kravetz * with unique huge page to page table relationships can define their own 7129e95a9851SMike Kravetz * version of this routine. 7130e95a9851SMike Kravetz */ 7131e95a9851SMike Kravetz unsigned long hugetlb_mask_last_page(struct hstate *h) 7132e95a9851SMike Kravetz { 7133e95a9851SMike Kravetz unsigned long hp_size = huge_page_size(h); 7134e95a9851SMike Kravetz 7135e95a9851SMike Kravetz if (hp_size == PUD_SIZE) 7136e95a9851SMike Kravetz return P4D_SIZE - PUD_SIZE; 7137e95a9851SMike Kravetz else if (hp_size == PMD_SIZE) 7138e95a9851SMike Kravetz return PUD_SIZE - PMD_SIZE; 7139e95a9851SMike Kravetz else 7140e95a9851SMike Kravetz return 0UL; 7141e95a9851SMike Kravetz } 7142e95a9851SMike Kravetz 7143e95a9851SMike Kravetz #else 7144e95a9851SMike Kravetz 7145e95a9851SMike Kravetz /* See description above. Architectures can provide their own version. */ 7146e95a9851SMike Kravetz __weak unsigned long hugetlb_mask_last_page(struct hstate *h) 7147e95a9851SMike Kravetz { 71484ddb4d91SMike Kravetz #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 71494ddb4d91SMike Kravetz if (huge_page_size(h) == PMD_SIZE) 71504ddb4d91SMike Kravetz return PUD_SIZE - PMD_SIZE; 71514ddb4d91SMike Kravetz #endif 7152e95a9851SMike Kravetz return 0UL; 7153e95a9851SMike Kravetz } 7154e95a9851SMike Kravetz 715561f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 715661f77edaSNaoya Horiguchi 715761f77edaSNaoya Horiguchi /* 715861f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 715961f77edaSNaoya Horiguchi * behavior. 716061f77edaSNaoya Horiguchi */ 71619747b9e9SBaolin Wang bool isolate_hugetlb(struct folio *folio, struct list_head *list) 716231caf665SNaoya Horiguchi { 71639747b9e9SBaolin Wang bool ret = true; 7164bcc54222SNaoya Horiguchi 7165db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 71666aa3a920SSidhartha Kumar if (!folio_test_hugetlb(folio) || 71676aa3a920SSidhartha Kumar !folio_test_hugetlb_migratable(folio) || 71686aa3a920SSidhartha Kumar !folio_try_get(folio)) { 71699747b9e9SBaolin Wang ret = false; 7170bcc54222SNaoya Horiguchi goto unlock; 7171bcc54222SNaoya Horiguchi } 71726aa3a920SSidhartha Kumar folio_clear_hugetlb_migratable(folio); 71736aa3a920SSidhartha Kumar list_move_tail(&folio->lru, list); 7174bcc54222SNaoya Horiguchi unlock: 7175db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7176bcc54222SNaoya Horiguchi return ret; 717731caf665SNaoya Horiguchi } 717831caf665SNaoya Horiguchi 717904bac040SSidhartha Kumar int get_hwpoison_hugetlb_folio(struct folio *folio, bool *hugetlb, bool unpoison) 718025182f05SNaoya Horiguchi { 718125182f05SNaoya Horiguchi int ret = 0; 718225182f05SNaoya Horiguchi 718325182f05SNaoya Horiguchi *hugetlb = false; 718425182f05SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 718504bac040SSidhartha Kumar if (folio_test_hugetlb(folio)) { 718625182f05SNaoya Horiguchi *hugetlb = true; 718704bac040SSidhartha Kumar if (folio_test_hugetlb_freed(folio)) 7188b283d983SNaoya Horiguchi ret = 0; 718904bac040SSidhartha Kumar else if (folio_test_hugetlb_migratable(folio) || unpoison) 719004bac040SSidhartha Kumar ret = folio_try_get(folio); 71910ed950d1SNaoya Horiguchi else 71920ed950d1SNaoya Horiguchi ret = -EBUSY; 719325182f05SNaoya Horiguchi } 719425182f05SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 719525182f05SNaoya Horiguchi return ret; 719625182f05SNaoya Horiguchi } 719725182f05SNaoya Horiguchi 7198e591ef7dSNaoya Horiguchi int get_huge_page_for_hwpoison(unsigned long pfn, int flags, 7199e591ef7dSNaoya Horiguchi bool *migratable_cleared) 7200405ce051SNaoya Horiguchi { 7201405ce051SNaoya Horiguchi int ret; 7202405ce051SNaoya Horiguchi 7203405ce051SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 7204e591ef7dSNaoya Horiguchi ret = __get_huge_page_for_hwpoison(pfn, flags, migratable_cleared); 7205405ce051SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 7206405ce051SNaoya Horiguchi return ret; 7207405ce051SNaoya Horiguchi } 7208405ce051SNaoya Horiguchi 7209ea8e72f4SSidhartha Kumar void folio_putback_active_hugetlb(struct folio *folio) 721031caf665SNaoya Horiguchi { 7211db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 7212ea8e72f4SSidhartha Kumar folio_set_hugetlb_migratable(folio); 7213ea8e72f4SSidhartha Kumar list_move_tail(&folio->lru, &(folio_hstate(folio))->hugepage_activelist); 7214db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7215ea8e72f4SSidhartha Kumar folio_put(folio); 721631caf665SNaoya Horiguchi } 7217ab5ac90aSMichal Hocko 7218345c62d1SSidhartha Kumar void move_hugetlb_state(struct folio *old_folio, struct folio *new_folio, int reason) 7219ab5ac90aSMichal Hocko { 7220345c62d1SSidhartha Kumar struct hstate *h = folio_hstate(old_folio); 7221ab5ac90aSMichal Hocko 7222345c62d1SSidhartha Kumar hugetlb_cgroup_migrate(old_folio, new_folio); 7223345c62d1SSidhartha Kumar set_page_owner_migrate_reason(&new_folio->page, reason); 7224ab5ac90aSMichal Hocko 7225ab5ac90aSMichal Hocko /* 7226345c62d1SSidhartha Kumar * transfer temporary state of the new hugetlb folio. This is 7227ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 7228ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 7229ab5ac90aSMichal Hocko * the temporary status. 7230ab5ac90aSMichal Hocko * 7231ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 7232ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 7233ab5ac90aSMichal Hocko * the per-node's. 7234ab5ac90aSMichal Hocko */ 7235345c62d1SSidhartha Kumar if (folio_test_hugetlb_temporary(new_folio)) { 7236345c62d1SSidhartha Kumar int old_nid = folio_nid(old_folio); 7237345c62d1SSidhartha Kumar int new_nid = folio_nid(new_folio); 7238ab5ac90aSMichal Hocko 7239345c62d1SSidhartha Kumar folio_set_hugetlb_temporary(old_folio); 7240345c62d1SSidhartha Kumar folio_clear_hugetlb_temporary(new_folio); 7241345c62d1SSidhartha Kumar 7242ab5ac90aSMichal Hocko 72435af1ab1dSMiaohe Lin /* 72445af1ab1dSMiaohe Lin * There is no need to transfer the per-node surplus state 72455af1ab1dSMiaohe Lin * when we do not cross the node. 72465af1ab1dSMiaohe Lin */ 72475af1ab1dSMiaohe Lin if (new_nid == old_nid) 72485af1ab1dSMiaohe Lin return; 7249db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 7250ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 7251ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 7252ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 7253ab5ac90aSMichal Hocko } 7254db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7255ab5ac90aSMichal Hocko } 7256ab5ac90aSMichal Hocko } 7257cf11e85fSRoman Gushchin 7258b30c14cdSJames Houghton static void hugetlb_unshare_pmds(struct vm_area_struct *vma, 7259b30c14cdSJames Houghton unsigned long start, 7260b30c14cdSJames Houghton unsigned long end) 72616dfeaff9SPeter Xu { 72626dfeaff9SPeter Xu struct hstate *h = hstate_vma(vma); 72636dfeaff9SPeter Xu unsigned long sz = huge_page_size(h); 72646dfeaff9SPeter Xu struct mm_struct *mm = vma->vm_mm; 72656dfeaff9SPeter Xu struct mmu_notifier_range range; 7266b30c14cdSJames Houghton unsigned long address; 72676dfeaff9SPeter Xu spinlock_t *ptl; 72686dfeaff9SPeter Xu pte_t *ptep; 72696dfeaff9SPeter Xu 72706dfeaff9SPeter Xu if (!(vma->vm_flags & VM_MAYSHARE)) 72716dfeaff9SPeter Xu return; 72726dfeaff9SPeter Xu 72736dfeaff9SPeter Xu if (start >= end) 72746dfeaff9SPeter Xu return; 72756dfeaff9SPeter Xu 72769c8bbfacSBaolin Wang flush_cache_range(vma, start, end); 72776dfeaff9SPeter Xu /* 72786dfeaff9SPeter Xu * No need to call adjust_range_if_pmd_sharing_possible(), because 72796dfeaff9SPeter Xu * we have already done the PUD_SIZE alignment. 72806dfeaff9SPeter Xu */ 72817d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, 72826dfeaff9SPeter Xu start, end); 72836dfeaff9SPeter Xu mmu_notifier_invalidate_range_start(&range); 728440549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 72856dfeaff9SPeter Xu i_mmap_lock_write(vma->vm_file->f_mapping); 72866dfeaff9SPeter Xu for (address = start; address < end; address += PUD_SIZE) { 72879c67a207SPeter Xu ptep = hugetlb_walk(vma, address, sz); 72886dfeaff9SPeter Xu if (!ptep) 72896dfeaff9SPeter Xu continue; 72906dfeaff9SPeter Xu ptl = huge_pte_lock(h, mm, ptep); 72914ddb4d91SMike Kravetz huge_pmd_unshare(mm, vma, address, ptep); 72926dfeaff9SPeter Xu spin_unlock(ptl); 72936dfeaff9SPeter Xu } 72946dfeaff9SPeter Xu flush_hugetlb_tlb_range(vma, start, end); 72956dfeaff9SPeter Xu i_mmap_unlock_write(vma->vm_file->f_mapping); 729640549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 72976dfeaff9SPeter Xu /* 72986dfeaff9SPeter Xu * No need to call mmu_notifier_invalidate_range(), see 7299ee65728eSMike Rapoport * Documentation/mm/mmu_notifier.rst. 73006dfeaff9SPeter Xu */ 73016dfeaff9SPeter Xu mmu_notifier_invalidate_range_end(&range); 73026dfeaff9SPeter Xu } 73036dfeaff9SPeter Xu 7304b30c14cdSJames Houghton /* 7305b30c14cdSJames Houghton * This function will unconditionally remove all the shared pmd pgtable entries 7306b30c14cdSJames Houghton * within the specific vma for a hugetlbfs memory range. 7307b30c14cdSJames Houghton */ 7308b30c14cdSJames Houghton void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) 7309b30c14cdSJames Houghton { 7310b30c14cdSJames Houghton hugetlb_unshare_pmds(vma, ALIGN(vma->vm_start, PUD_SIZE), 7311b30c14cdSJames Houghton ALIGN_DOWN(vma->vm_end, PUD_SIZE)); 7312b30c14cdSJames Houghton } 7313b30c14cdSJames Houghton 7314cf11e85fSRoman Gushchin #ifdef CONFIG_CMA 7315cf11e85fSRoman Gushchin static bool cma_reserve_called __initdata; 7316cf11e85fSRoman Gushchin 7317cf11e85fSRoman Gushchin static int __init cmdline_parse_hugetlb_cma(char *p) 7318cf11e85fSRoman Gushchin { 731938e719abSBaolin Wang int nid, count = 0; 732038e719abSBaolin Wang unsigned long tmp; 732138e719abSBaolin Wang char *s = p; 732238e719abSBaolin Wang 732338e719abSBaolin Wang while (*s) { 732438e719abSBaolin Wang if (sscanf(s, "%lu%n", &tmp, &count) != 1) 732538e719abSBaolin Wang break; 732638e719abSBaolin Wang 732738e719abSBaolin Wang if (s[count] == ':') { 7328f9317f77SMike Kravetz if (tmp >= MAX_NUMNODES) 732938e719abSBaolin Wang break; 7330f9317f77SMike Kravetz nid = array_index_nospec(tmp, MAX_NUMNODES); 733138e719abSBaolin Wang 733238e719abSBaolin Wang s += count + 1; 733338e719abSBaolin Wang tmp = memparse(s, &s); 733438e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = tmp; 733538e719abSBaolin Wang hugetlb_cma_size += tmp; 733638e719abSBaolin Wang 733738e719abSBaolin Wang /* 733838e719abSBaolin Wang * Skip the separator if have one, otherwise 733938e719abSBaolin Wang * break the parsing. 734038e719abSBaolin Wang */ 734138e719abSBaolin Wang if (*s == ',') 734238e719abSBaolin Wang s++; 734338e719abSBaolin Wang else 734438e719abSBaolin Wang break; 734538e719abSBaolin Wang } else { 7346cf11e85fSRoman Gushchin hugetlb_cma_size = memparse(p, &p); 734738e719abSBaolin Wang break; 734838e719abSBaolin Wang } 734938e719abSBaolin Wang } 735038e719abSBaolin Wang 7351cf11e85fSRoman Gushchin return 0; 7352cf11e85fSRoman Gushchin } 7353cf11e85fSRoman Gushchin 7354cf11e85fSRoman Gushchin early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); 7355cf11e85fSRoman Gushchin 7356cf11e85fSRoman Gushchin void __init hugetlb_cma_reserve(int order) 7357cf11e85fSRoman Gushchin { 7358cf11e85fSRoman Gushchin unsigned long size, reserved, per_node; 735938e719abSBaolin Wang bool node_specific_cma_alloc = false; 7360cf11e85fSRoman Gushchin int nid; 7361cf11e85fSRoman Gushchin 7362cf11e85fSRoman Gushchin cma_reserve_called = true; 7363cf11e85fSRoman Gushchin 7364cf11e85fSRoman Gushchin if (!hugetlb_cma_size) 7365cf11e85fSRoman Gushchin return; 7366cf11e85fSRoman Gushchin 736738e719abSBaolin Wang for (nid = 0; nid < MAX_NUMNODES; nid++) { 736838e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 736938e719abSBaolin Wang continue; 737038e719abSBaolin Wang 737130a51400SPeng Liu if (!node_online(nid)) { 737238e719abSBaolin Wang pr_warn("hugetlb_cma: invalid node %d specified\n", nid); 737338e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 737438e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 737538e719abSBaolin Wang continue; 737638e719abSBaolin Wang } 737738e719abSBaolin Wang 737838e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) { 737938e719abSBaolin Wang pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n", 738038e719abSBaolin Wang nid, (PAGE_SIZE << order) / SZ_1M); 738138e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 738238e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 738338e719abSBaolin Wang } else { 738438e719abSBaolin Wang node_specific_cma_alloc = true; 738538e719abSBaolin Wang } 738638e719abSBaolin Wang } 738738e719abSBaolin Wang 738838e719abSBaolin Wang /* Validate the CMA size again in case some invalid nodes specified. */ 738938e719abSBaolin Wang if (!hugetlb_cma_size) 739038e719abSBaolin Wang return; 739138e719abSBaolin Wang 7392cf11e85fSRoman Gushchin if (hugetlb_cma_size < (PAGE_SIZE << order)) { 7393cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", 7394cf11e85fSRoman Gushchin (PAGE_SIZE << order) / SZ_1M); 7395a01f4390SMike Kravetz hugetlb_cma_size = 0; 7396cf11e85fSRoman Gushchin return; 7397cf11e85fSRoman Gushchin } 7398cf11e85fSRoman Gushchin 739938e719abSBaolin Wang if (!node_specific_cma_alloc) { 7400cf11e85fSRoman Gushchin /* 7401cf11e85fSRoman Gushchin * If 3 GB area is requested on a machine with 4 numa nodes, 7402cf11e85fSRoman Gushchin * let's allocate 1 GB on first three nodes and ignore the last one. 7403cf11e85fSRoman Gushchin */ 7404cf11e85fSRoman Gushchin per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); 7405cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", 7406cf11e85fSRoman Gushchin hugetlb_cma_size / SZ_1M, per_node / SZ_1M); 740738e719abSBaolin Wang } 7408cf11e85fSRoman Gushchin 7409cf11e85fSRoman Gushchin reserved = 0; 741030a51400SPeng Liu for_each_online_node(nid) { 7411cf11e85fSRoman Gushchin int res; 74122281f797SBarry Song char name[CMA_MAX_NAME]; 7413cf11e85fSRoman Gushchin 741438e719abSBaolin Wang if (node_specific_cma_alloc) { 741538e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 741638e719abSBaolin Wang continue; 741738e719abSBaolin Wang 741838e719abSBaolin Wang size = hugetlb_cma_size_in_node[nid]; 741938e719abSBaolin Wang } else { 7420cf11e85fSRoman Gushchin size = min(per_node, hugetlb_cma_size - reserved); 742138e719abSBaolin Wang } 742238e719abSBaolin Wang 7423cf11e85fSRoman Gushchin size = round_up(size, PAGE_SIZE << order); 7424cf11e85fSRoman Gushchin 74252281f797SBarry Song snprintf(name, sizeof(name), "hugetlb%d", nid); 7426a01f4390SMike Kravetz /* 7427a01f4390SMike Kravetz * Note that 'order per bit' is based on smallest size that 7428a01f4390SMike Kravetz * may be returned to CMA allocator in the case of 7429a01f4390SMike Kravetz * huge page demotion. 7430a01f4390SMike Kravetz */ 7431a01f4390SMike Kravetz res = cma_declare_contiguous_nid(0, size, 0, 7432a01f4390SMike Kravetz PAGE_SIZE << HUGETLB_PAGE_ORDER, 743329d0f41dSBarry Song 0, false, name, 7434cf11e85fSRoman Gushchin &hugetlb_cma[nid], nid); 7435cf11e85fSRoman Gushchin if (res) { 7436cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: reservation failed: err %d, node %d", 7437cf11e85fSRoman Gushchin res, nid); 7438cf11e85fSRoman Gushchin continue; 7439cf11e85fSRoman Gushchin } 7440cf11e85fSRoman Gushchin 7441cf11e85fSRoman Gushchin reserved += size; 7442cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", 7443cf11e85fSRoman Gushchin size / SZ_1M, nid); 7444cf11e85fSRoman Gushchin 7445cf11e85fSRoman Gushchin if (reserved >= hugetlb_cma_size) 7446cf11e85fSRoman Gushchin break; 7447cf11e85fSRoman Gushchin } 7448a01f4390SMike Kravetz 7449a01f4390SMike Kravetz if (!reserved) 7450a01f4390SMike Kravetz /* 7451a01f4390SMike Kravetz * hugetlb_cma_size is used to determine if allocations from 7452a01f4390SMike Kravetz * cma are possible. Set to zero if no cma regions are set up. 7453a01f4390SMike Kravetz */ 7454a01f4390SMike Kravetz hugetlb_cma_size = 0; 7455cf11e85fSRoman Gushchin } 7456cf11e85fSRoman Gushchin 7457263b8998SMiaohe Lin static void __init hugetlb_cma_check(void) 7458cf11e85fSRoman Gushchin { 7459cf11e85fSRoman Gushchin if (!hugetlb_cma_size || cma_reserve_called) 7460cf11e85fSRoman Gushchin return; 7461cf11e85fSRoman Gushchin 7462cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); 7463cf11e85fSRoman Gushchin } 7464cf11e85fSRoman Gushchin 7465cf11e85fSRoman Gushchin #endif /* CONFIG_CMA */ 7466