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 158332c87719SMike Kravetz static inline void __clear_hugetlb_destructor(struct hstate *h, 158432c87719SMike Kravetz struct folio *folio) 158532c87719SMike Kravetz { 158632c87719SMike Kravetz lockdep_assert_held(&hugetlb_lock); 158732c87719SMike Kravetz 15886eb4e88aSMike Kravetz /* 158932c87719SMike Kravetz * Very subtle 159032c87719SMike Kravetz * 159132c87719SMike Kravetz * For non-gigantic pages set the destructor to the normal compound 159232c87719SMike Kravetz * page dtor. This is needed in case someone takes an additional 159332c87719SMike Kravetz * temporary ref to the page, and freeing is delayed until they drop 159432c87719SMike Kravetz * their reference. 159532c87719SMike Kravetz * 159632c87719SMike Kravetz * For gigantic pages set the destructor to the null dtor. This 159732c87719SMike Kravetz * destructor will never be called. Before freeing the gigantic 159832c87719SMike Kravetz * page destroy_compound_gigantic_folio will turn the folio into a 159932c87719SMike Kravetz * simple group of pages. After this the destructor does not 160032c87719SMike Kravetz * apply. 160132c87719SMike Kravetz * 160232c87719SMike Kravetz */ 160332c87719SMike Kravetz if (hstate_is_gigantic(h)) 160432c87719SMike Kravetz folio_set_compound_dtor(folio, NULL_COMPOUND_DTOR); 160532c87719SMike Kravetz else 160632c87719SMike Kravetz folio_set_compound_dtor(folio, COMPOUND_PAGE_DTOR); 160732c87719SMike Kravetz } 160832c87719SMike Kravetz 160932c87719SMike Kravetz /* 161032c87719SMike Kravetz * Remove hugetlb folio from lists. 161132c87719SMike Kravetz * If vmemmap exists for the folio, update dtor so that the folio appears 161232c87719SMike Kravetz * as just a compound page. Otherwise, wait until after allocating vmemmap 161332c87719SMike Kravetz * to update dtor. 161434d9e35bSMike Kravetz * 1615cfd5082bSSidhartha Kumar * A reference is held on the folio, except in the case of demote. 16166eb4e88aSMike Kravetz * 16176eb4e88aSMike Kravetz * Must be called with hugetlb lock held. 16186eb4e88aSMike Kravetz */ 1619cfd5082bSSidhartha Kumar static void __remove_hugetlb_folio(struct hstate *h, struct folio *folio, 162034d9e35bSMike Kravetz bool adjust_surplus, 162134d9e35bSMike Kravetz bool demote) 16226eb4e88aSMike Kravetz { 1623cfd5082bSSidhartha Kumar int nid = folio_nid(folio); 16246eb4e88aSMike Kravetz 1625f074732dSSidhartha Kumar VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio(folio), folio); 1626f074732dSSidhartha Kumar VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio_rsvd(folio), folio); 16276eb4e88aSMike Kravetz 16289487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 16296eb4e88aSMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 16306eb4e88aSMike Kravetz return; 16316eb4e88aSMike Kravetz 1632cfd5082bSSidhartha Kumar list_del(&folio->lru); 16336eb4e88aSMike Kravetz 1634cfd5082bSSidhartha Kumar if (folio_test_hugetlb_freed(folio)) { 16356eb4e88aSMike Kravetz h->free_huge_pages--; 16366eb4e88aSMike Kravetz h->free_huge_pages_node[nid]--; 16376eb4e88aSMike Kravetz } 16386eb4e88aSMike Kravetz if (adjust_surplus) { 16396eb4e88aSMike Kravetz h->surplus_huge_pages--; 16406eb4e88aSMike Kravetz h->surplus_huge_pages_node[nid]--; 16416eb4e88aSMike Kravetz } 16426eb4e88aSMike Kravetz 1643e32d20c0SMike Kravetz /* 164432c87719SMike Kravetz * We can only clear the hugetlb destructor after allocating vmemmap 164532c87719SMike Kravetz * pages. Otherwise, someone (memory error handling) may try to write 164632c87719SMike Kravetz * to tail struct pages. 164732c87719SMike Kravetz */ 164832c87719SMike Kravetz if (!folio_test_hugetlb_vmemmap_optimized(folio)) 164932c87719SMike Kravetz __clear_hugetlb_destructor(h, folio); 165032c87719SMike Kravetz 165132c87719SMike Kravetz /* 165234d9e35bSMike Kravetz * In the case of demote we do not ref count the page as it will soon 165334d9e35bSMike Kravetz * be turned into a page of smaller size. 1654e32d20c0SMike Kravetz */ 165534d9e35bSMike Kravetz if (!demote) 1656cfd5082bSSidhartha Kumar folio_ref_unfreeze(folio, 1); 16576eb4e88aSMike Kravetz 16586eb4e88aSMike Kravetz h->nr_huge_pages--; 16596eb4e88aSMike Kravetz h->nr_huge_pages_node[nid]--; 16606eb4e88aSMike Kravetz } 16616eb4e88aSMike Kravetz 1662cfd5082bSSidhartha Kumar static void remove_hugetlb_folio(struct hstate *h, struct folio *folio, 166334d9e35bSMike Kravetz bool adjust_surplus) 166434d9e35bSMike Kravetz { 1665cfd5082bSSidhartha Kumar __remove_hugetlb_folio(h, folio, adjust_surplus, false); 166634d9e35bSMike Kravetz } 166734d9e35bSMike Kravetz 1668cfd5082bSSidhartha Kumar static void remove_hugetlb_folio_for_demote(struct hstate *h, struct folio *folio, 16698531fc6fSMike Kravetz bool adjust_surplus) 16708531fc6fSMike Kravetz { 1671cfd5082bSSidhartha Kumar __remove_hugetlb_folio(h, folio, adjust_surplus, true); 16728531fc6fSMike Kravetz } 16738531fc6fSMike Kravetz 16742f6c57d6SSidhartha Kumar static void add_hugetlb_folio(struct hstate *h, struct folio *folio, 1675ad2fa371SMuchun Song bool adjust_surplus) 1676ad2fa371SMuchun Song { 1677ad2fa371SMuchun Song int zeroed; 16782f6c57d6SSidhartha Kumar int nid = folio_nid(folio); 1679ad2fa371SMuchun Song 16802f6c57d6SSidhartha Kumar VM_BUG_ON_FOLIO(!folio_test_hugetlb_vmemmap_optimized(folio), folio); 1681ad2fa371SMuchun Song 1682ad2fa371SMuchun Song lockdep_assert_held(&hugetlb_lock); 1683ad2fa371SMuchun Song 16842f6c57d6SSidhartha Kumar INIT_LIST_HEAD(&folio->lru); 1685ad2fa371SMuchun Song h->nr_huge_pages++; 1686ad2fa371SMuchun Song h->nr_huge_pages_node[nid]++; 1687ad2fa371SMuchun Song 1688ad2fa371SMuchun Song if (adjust_surplus) { 1689ad2fa371SMuchun Song h->surplus_huge_pages++; 1690ad2fa371SMuchun Song h->surplus_huge_pages_node[nid]++; 1691ad2fa371SMuchun Song } 1692ad2fa371SMuchun Song 16932f6c57d6SSidhartha Kumar folio_set_compound_dtor(folio, HUGETLB_PAGE_DTOR); 16942f6c57d6SSidhartha Kumar folio_change_private(folio, NULL); 1695a9e1eab2SMiaohe Lin /* 16962f6c57d6SSidhartha Kumar * We have to set hugetlb_vmemmap_optimized again as above 16972f6c57d6SSidhartha Kumar * folio_change_private(folio, NULL) cleared it. 1698a9e1eab2SMiaohe Lin */ 16992f6c57d6SSidhartha Kumar folio_set_hugetlb_vmemmap_optimized(folio); 1700ad2fa371SMuchun Song 1701ad2fa371SMuchun Song /* 17022f6c57d6SSidhartha Kumar * This folio is about to be managed by the hugetlb allocator and 1703b65a4edaSMike Kravetz * should have no users. Drop our reference, and check for others 1704b65a4edaSMike Kravetz * just in case. 1705ad2fa371SMuchun Song */ 17062f6c57d6SSidhartha Kumar zeroed = folio_put_testzero(folio); 17072f6c57d6SSidhartha Kumar if (unlikely(!zeroed)) 1708b65a4edaSMike Kravetz /* 1709*454a00c4SMatthew Wilcox (Oracle) * It is VERY unlikely soneone else has taken a ref 1710*454a00c4SMatthew Wilcox (Oracle) * on the folio. In this case, we simply return as 1711*454a00c4SMatthew Wilcox (Oracle) * free_huge_folio() will be called when this other ref 1712*454a00c4SMatthew Wilcox (Oracle) * is dropped. 1713b65a4edaSMike Kravetz */ 1714b65a4edaSMike Kravetz return; 1715b65a4edaSMike Kravetz 17162f6c57d6SSidhartha Kumar arch_clear_hugepage_flags(&folio->page); 1717240d67a8SSidhartha Kumar enqueue_hugetlb_folio(h, folio); 1718ad2fa371SMuchun Song } 1719ad2fa371SMuchun Song 17206f6956cfSSidhartha Kumar static void __update_and_free_hugetlb_folio(struct hstate *h, 17216f6956cfSSidhartha Kumar struct folio *folio) 17226af2acb6SAdam Litke { 17236af2acb6SAdam Litke int i; 172414455eabSCheng Li struct page *subpage; 172532c87719SMike Kravetz bool clear_dtor = folio_test_hugetlb_vmemmap_optimized(folio); 1726a5516438SAndi Kleen 17274eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1728944d9fecSLuiz Capitulino return; 172918229df5SAndy Whitcroft 1730161df60eSNaoya Horiguchi /* 1731161df60eSNaoya Horiguchi * If we don't know which subpages are hwpoisoned, we can't free 1732161df60eSNaoya Horiguchi * the hugepage, so it's leaked intentionally. 1733161df60eSNaoya Horiguchi */ 17347f325a8dSSidhartha Kumar if (folio_test_hugetlb_raw_hwp_unreliable(folio)) 1735161df60eSNaoya Horiguchi return; 1736161df60eSNaoya Horiguchi 17376f6956cfSSidhartha Kumar if (hugetlb_vmemmap_restore(h, &folio->page)) { 1738ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 1739ad2fa371SMuchun Song /* 1740ad2fa371SMuchun Song * If we cannot allocate vmemmap pages, just refuse to free the 1741ad2fa371SMuchun Song * page and put the page back on the hugetlb free list and treat 1742ad2fa371SMuchun Song * as a surplus page. 1743ad2fa371SMuchun Song */ 17447f325a8dSSidhartha Kumar add_hugetlb_folio(h, folio, true); 1745ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 1746ad2fa371SMuchun Song return; 1747ad2fa371SMuchun Song } 1748ad2fa371SMuchun Song 1749161df60eSNaoya Horiguchi /* 1750161df60eSNaoya Horiguchi * Move PageHWPoison flag from head page to the raw error pages, 1751161df60eSNaoya Horiguchi * which makes any healthy subpages reusable. 1752161df60eSNaoya Horiguchi */ 1753911565b8SSidhartha Kumar if (unlikely(folio_test_hwpoison(folio))) 17542ff6ceceSSidhartha Kumar folio_clear_hugetlb_hwpoison(folio); 1755161df60eSNaoya Horiguchi 175632c87719SMike Kravetz /* 175732c87719SMike Kravetz * If vmemmap pages were allocated above, then we need to clear the 175832c87719SMike Kravetz * hugetlb destructor under the hugetlb lock. 175932c87719SMike Kravetz */ 176032c87719SMike Kravetz if (clear_dtor) { 176132c87719SMike Kravetz spin_lock_irq(&hugetlb_lock); 176232c87719SMike Kravetz __clear_hugetlb_destructor(h, folio); 176332c87719SMike Kravetz spin_unlock_irq(&hugetlb_lock); 176432c87719SMike Kravetz } 176532c87719SMike Kravetz 176614455eabSCheng Li for (i = 0; i < pages_per_huge_page(h); i++) { 17677f325a8dSSidhartha Kumar subpage = folio_page(folio, i); 1768dbfee5aeSMike Kravetz subpage->flags &= ~(1 << PG_locked | 1 << PG_error | 176932f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1770a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1771a7407a27SLuiz Capitulino 1 << PG_writeback); 17726af2acb6SAdam Litke } 1773a01f4390SMike Kravetz 1774a01f4390SMike Kravetz /* 1775a01f4390SMike Kravetz * Non-gigantic pages demoted from CMA allocated gigantic pages 17767f325a8dSSidhartha Kumar * need to be given back to CMA in free_gigantic_folio. 1777a01f4390SMike Kravetz */ 1778a01f4390SMike Kravetz if (hstate_is_gigantic(h) || 17792f6c57d6SSidhartha Kumar hugetlb_cma_folio(folio, huge_page_order(h))) { 1780911565b8SSidhartha Kumar destroy_compound_gigantic_folio(folio, huge_page_order(h)); 17817f325a8dSSidhartha Kumar free_gigantic_folio(folio, huge_page_order(h)); 1782944d9fecSLuiz Capitulino } else { 17836f6956cfSSidhartha Kumar __free_pages(&folio->page, huge_page_order(h)); 17846af2acb6SAdam Litke } 1785944d9fecSLuiz Capitulino } 17866af2acb6SAdam Litke 1787b65d4adbSMuchun Song /* 1788d6ef19e2SSidhartha Kumar * As update_and_free_hugetlb_folio() can be called under any context, so we cannot 1789b65d4adbSMuchun Song * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the 1790b65d4adbSMuchun Song * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate 1791b65d4adbSMuchun Song * the vmemmap pages. 1792b65d4adbSMuchun Song * 1793b65d4adbSMuchun Song * free_hpage_workfn() locklessly retrieves the linked list of pages to be 1794b65d4adbSMuchun Song * freed and frees them one-by-one. As the page->mapping pointer is going 1795b65d4adbSMuchun Song * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node 1796b65d4adbSMuchun Song * structure of a lockless linked list of huge pages to be freed. 1797b65d4adbSMuchun Song */ 1798b65d4adbSMuchun Song static LLIST_HEAD(hpage_freelist); 1799b65d4adbSMuchun Song 1800b65d4adbSMuchun Song static void free_hpage_workfn(struct work_struct *work) 1801b65d4adbSMuchun Song { 1802b65d4adbSMuchun Song struct llist_node *node; 1803b65d4adbSMuchun Song 1804b65d4adbSMuchun Song node = llist_del_all(&hpage_freelist); 1805b65d4adbSMuchun Song 1806b65d4adbSMuchun Song while (node) { 1807b65d4adbSMuchun Song struct page *page; 1808b65d4adbSMuchun Song struct hstate *h; 1809b65d4adbSMuchun Song 1810b65d4adbSMuchun Song page = container_of((struct address_space **)node, 1811b65d4adbSMuchun Song struct page, mapping); 1812b65d4adbSMuchun Song node = node->next; 1813b65d4adbSMuchun Song page->mapping = NULL; 1814b65d4adbSMuchun Song /* 1815affd26b1SSidhartha Kumar * The VM_BUG_ON_FOLIO(!folio_test_hugetlb(folio), folio) in 1816affd26b1SSidhartha Kumar * folio_hstate() is going to trigger because a previous call to 1817cfd5082bSSidhartha Kumar * remove_hugetlb_folio() will call folio_set_compound_dtor 1818affd26b1SSidhartha Kumar * (folio, NULL_COMPOUND_DTOR), so do not use folio_hstate() 1819cfd5082bSSidhartha Kumar * directly. 1820b65d4adbSMuchun Song */ 1821b65d4adbSMuchun Song h = size_to_hstate(page_size(page)); 1822b65d4adbSMuchun Song 18236f6956cfSSidhartha Kumar __update_and_free_hugetlb_folio(h, page_folio(page)); 1824b65d4adbSMuchun Song 1825b65d4adbSMuchun Song cond_resched(); 1826b65d4adbSMuchun Song } 1827b65d4adbSMuchun Song } 1828b65d4adbSMuchun Song static DECLARE_WORK(free_hpage_work, free_hpage_workfn); 1829b65d4adbSMuchun Song 1830b65d4adbSMuchun Song static inline void flush_free_hpage_work(struct hstate *h) 1831b65d4adbSMuchun Song { 18326213834cSMuchun Song if (hugetlb_vmemmap_optimizable(h)) 1833b65d4adbSMuchun Song flush_work(&free_hpage_work); 1834b65d4adbSMuchun Song } 1835b65d4adbSMuchun Song 1836d6ef19e2SSidhartha Kumar static void update_and_free_hugetlb_folio(struct hstate *h, struct folio *folio, 1837b65d4adbSMuchun Song bool atomic) 1838b65d4adbSMuchun Song { 1839d6ef19e2SSidhartha Kumar if (!folio_test_hugetlb_vmemmap_optimized(folio) || !atomic) { 18406f6956cfSSidhartha Kumar __update_and_free_hugetlb_folio(h, folio); 1841b65d4adbSMuchun Song return; 1842b65d4adbSMuchun Song } 1843b65d4adbSMuchun Song 1844b65d4adbSMuchun Song /* 1845b65d4adbSMuchun Song * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages. 1846b65d4adbSMuchun Song * 1847b65d4adbSMuchun Song * Only call schedule_work() if hpage_freelist is previously 1848b65d4adbSMuchun Song * empty. Otherwise, schedule_work() had been called but the workfn 1849b65d4adbSMuchun Song * hasn't retrieved the list yet. 1850b65d4adbSMuchun Song */ 1851d6ef19e2SSidhartha Kumar if (llist_add((struct llist_node *)&folio->mapping, &hpage_freelist)) 1852b65d4adbSMuchun Song schedule_work(&free_hpage_work); 1853b65d4adbSMuchun Song } 1854b65d4adbSMuchun Song 185510c6ec49SMike Kravetz static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list) 185610c6ec49SMike Kravetz { 185710c6ec49SMike Kravetz struct page *page, *t_page; 1858d6ef19e2SSidhartha Kumar struct folio *folio; 185910c6ec49SMike Kravetz 186010c6ec49SMike Kravetz list_for_each_entry_safe(page, t_page, list, lru) { 1861d6ef19e2SSidhartha Kumar folio = page_folio(page); 1862d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, folio, false); 186310c6ec49SMike Kravetz cond_resched(); 186410c6ec49SMike Kravetz } 186510c6ec49SMike Kravetz } 186610c6ec49SMike Kravetz 1867e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1868e5ff2159SAndi Kleen { 1869e5ff2159SAndi Kleen struct hstate *h; 1870e5ff2159SAndi Kleen 1871e5ff2159SAndi Kleen for_each_hstate(h) { 1872e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1873e5ff2159SAndi Kleen return h; 1874e5ff2159SAndi Kleen } 1875e5ff2159SAndi Kleen return NULL; 1876e5ff2159SAndi Kleen } 1877e5ff2159SAndi Kleen 1878*454a00c4SMatthew Wilcox (Oracle) void free_huge_folio(struct folio *folio) 187927a85ef1SDavid Gibson { 1880a5516438SAndi Kleen /* 1881a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1882a5516438SAndi Kleen * compound page destructor. 1883a5516438SAndi Kleen */ 18840356c4b9SSidhartha Kumar struct hstate *h = folio_hstate(folio); 18850356c4b9SSidhartha Kumar int nid = folio_nid(folio); 18860356c4b9SSidhartha Kumar struct hugepage_subpool *spool = hugetlb_folio_subpool(folio); 188707443a85SJoonsoo Kim bool restore_reserve; 1888db71ef79SMike Kravetz unsigned long flags; 188927a85ef1SDavid Gibson 18900356c4b9SSidhartha Kumar VM_BUG_ON_FOLIO(folio_ref_count(folio), folio); 18910356c4b9SSidhartha Kumar VM_BUG_ON_FOLIO(folio_mapcount(folio), folio); 18928ace22bcSYongkai Wu 18930356c4b9SSidhartha Kumar hugetlb_set_folio_subpool(folio, NULL); 18940356c4b9SSidhartha Kumar if (folio_test_anon(folio)) 18950356c4b9SSidhartha Kumar __ClearPageAnonExclusive(&folio->page); 18960356c4b9SSidhartha Kumar folio->mapping = NULL; 18970356c4b9SSidhartha Kumar restore_reserve = folio_test_hugetlb_restore_reserve(folio); 18980356c4b9SSidhartha Kumar folio_clear_hugetlb_restore_reserve(folio); 189927a85ef1SDavid Gibson 19001c5ecae3SMike Kravetz /* 1901d6995da3SMike Kravetz * If HPageRestoreReserve was set on page, page allocation consumed a 19020919e1b6SMike Kravetz * reservation. If the page was associated with a subpool, there 19030919e1b6SMike Kravetz * would have been a page reserved in the subpool before allocation 19040919e1b6SMike Kravetz * via hugepage_subpool_get_pages(). Since we are 'restoring' the 19056c26d310SMiaohe Lin * reservation, do not call hugepage_subpool_put_pages() as this will 19060919e1b6SMike Kravetz * remove the reserved page from the subpool. 19070919e1b6SMike Kravetz */ 19080919e1b6SMike Kravetz if (!restore_reserve) { 19090919e1b6SMike Kravetz /* 19100919e1b6SMike Kravetz * A return code of zero implies that the subpool will be 19110919e1b6SMike Kravetz * under its minimum size if the reservation is not restored 19120919e1b6SMike Kravetz * after page is free. Therefore, force restore_reserve 19130919e1b6SMike Kravetz * operation. 19141c5ecae3SMike Kravetz */ 19151c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 19161c5ecae3SMike Kravetz restore_reserve = true; 19170919e1b6SMike Kravetz } 19181c5ecae3SMike Kravetz 1919db71ef79SMike Kravetz spin_lock_irqsave(&hugetlb_lock, flags); 19200356c4b9SSidhartha Kumar folio_clear_hugetlb_migratable(folio); 1921d4ab0316SSidhartha Kumar hugetlb_cgroup_uncharge_folio(hstate_index(h), 1922d4ab0316SSidhartha Kumar pages_per_huge_page(h), folio); 1923d4ab0316SSidhartha Kumar hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h), 1924d4ab0316SSidhartha Kumar pages_per_huge_page(h), folio); 192507443a85SJoonsoo Kim if (restore_reserve) 192607443a85SJoonsoo Kim h->resv_huge_pages++; 192707443a85SJoonsoo Kim 19280356c4b9SSidhartha Kumar if (folio_test_hugetlb_temporary(folio)) { 1929cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, false); 1930db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1931d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, folio, true); 1932ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 19330edaecfaSAneesh Kumar K.V /* remove the page from active list */ 1934cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, true); 1935db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1936d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, folio, true); 19377893d1d5SAdam Litke } else { 1938*454a00c4SMatthew Wilcox (Oracle) arch_clear_hugepage_flags(&folio->page); 1939240d67a8SSidhartha Kumar enqueue_hugetlb_folio(h, folio); 1940db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 194127a85ef1SDavid Gibson } 19421121828aSMike Kravetz } 194327a85ef1SDavid Gibson 1944d3d99fccSOscar Salvador /* 1945d3d99fccSOscar Salvador * Must be called with the hugetlb lock held 1946d3d99fccSOscar Salvador */ 1947d3d99fccSOscar Salvador static void __prep_account_new_huge_page(struct hstate *h, int nid) 1948d3d99fccSOscar Salvador { 1949d3d99fccSOscar Salvador lockdep_assert_held(&hugetlb_lock); 1950d3d99fccSOscar Salvador h->nr_huge_pages++; 1951d3d99fccSOscar Salvador h->nr_huge_pages_node[nid]++; 1952d3d99fccSOscar Salvador } 1953d3d99fccSOscar Salvador 1954de656ed3SSidhartha Kumar static void __prep_new_hugetlb_folio(struct hstate *h, struct folio *folio) 1955b7ba30c6SAndi Kleen { 1956de656ed3SSidhartha Kumar hugetlb_vmemmap_optimize(h, &folio->page); 1957de656ed3SSidhartha Kumar INIT_LIST_HEAD(&folio->lru); 19589fd33058SSidhartha Kumar folio_set_compound_dtor(folio, HUGETLB_PAGE_DTOR); 1959de656ed3SSidhartha Kumar hugetlb_set_folio_subpool(folio, NULL); 1960de656ed3SSidhartha Kumar set_hugetlb_cgroup(folio, NULL); 1961de656ed3SSidhartha Kumar set_hugetlb_cgroup_rsvd(folio, NULL); 1962d3d99fccSOscar Salvador } 1963d3d99fccSOscar Salvador 1964d1c60955SSidhartha Kumar static void prep_new_hugetlb_folio(struct hstate *h, struct folio *folio, int nid) 1965d3d99fccSOscar Salvador { 1966de656ed3SSidhartha Kumar __prep_new_hugetlb_folio(h, folio); 1967db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 1968d3d99fccSOscar Salvador __prep_account_new_huge_page(h, nid); 1969db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1970b7ba30c6SAndi Kleen } 1971b7ba30c6SAndi Kleen 1972d1c60955SSidhartha Kumar static bool __prep_compound_gigantic_folio(struct folio *folio, 1973d1c60955SSidhartha Kumar unsigned int order, bool demote) 197420a0307cSWu Fengguang { 19757118fc29SMike Kravetz int i, j; 197620a0307cSWu Fengguang int nr_pages = 1 << order; 197714455eabSCheng Li struct page *p; 197820a0307cSWu Fengguang 1979d1c60955SSidhartha Kumar __folio_clear_reserved(folio); 19802b21624fSMike Kravetz for (i = 0; i < nr_pages; i++) { 1981d1c60955SSidhartha Kumar p = folio_page(folio, i); 198214455eabSCheng Li 1983ef5a22beSAndrea Arcangeli /* 1984ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1985ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1986ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 19877c8de358SEthon Paul * too. Otherwise drivers using get_user_pages() to access tail 1988ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1989ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1990ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1991ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1992ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1993ef5a22beSAndrea Arcangeli * after get_user_pages(). 1994ef5a22beSAndrea Arcangeli */ 19957fb0728aSMike Kravetz if (i != 0) /* head page cleared above */ 1996ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 19977118fc29SMike Kravetz /* 19987118fc29SMike Kravetz * Subtle and very unlikely 19997118fc29SMike Kravetz * 20007118fc29SMike Kravetz * Gigantic 'page allocators' such as memblock or cma will 20017118fc29SMike Kravetz * return a set of pages with each page ref counted. We need 20027118fc29SMike Kravetz * to turn this set of pages into a compound page with tail 20037118fc29SMike Kravetz * page ref counts set to zero. Code such as speculative page 20047118fc29SMike Kravetz * cache adding could take a ref on a 'to be' tail page. 20057118fc29SMike Kravetz * We need to respect any increased ref count, and only set 20067118fc29SMike Kravetz * the ref count to zero if count is currently 1. If count 2007416d85edSMike Kravetz * is not 1, we return an error. An error return indicates 2008416d85edSMike Kravetz * the set of pages can not be converted to a gigantic page. 2009416d85edSMike Kravetz * The caller who allocated the pages should then discard the 2010416d85edSMike Kravetz * pages using the appropriate free interface. 201134d9e35bSMike Kravetz * 201234d9e35bSMike Kravetz * In the case of demote, the ref count will be zero. 20137118fc29SMike Kravetz */ 201434d9e35bSMike Kravetz if (!demote) { 20157118fc29SMike Kravetz if (!page_ref_freeze(p, 1)) { 2016416d85edSMike Kravetz pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n"); 20177118fc29SMike Kravetz goto out_error; 20187118fc29SMike Kravetz } 201934d9e35bSMike Kravetz } else { 202034d9e35bSMike Kravetz VM_BUG_ON_PAGE(page_count(p), p); 202134d9e35bSMike Kravetz } 20222b21624fSMike Kravetz if (i != 0) 2023d1c60955SSidhartha Kumar set_compound_head(p, &folio->page); 202420a0307cSWu Fengguang } 2025e3b7bf97STarun Sahu __folio_set_head(folio); 2026e3b7bf97STarun Sahu /* we rely on prep_new_hugetlb_folio to set the destructor */ 2027e3b7bf97STarun Sahu folio_set_order(folio, order); 202846f27228SMatthew Wilcox (Oracle) atomic_set(&folio->_entire_mapcount, -1); 2029eec20426SMatthew Wilcox (Oracle) atomic_set(&folio->_nr_pages_mapped, 0); 203094688e8eSMatthew Wilcox (Oracle) atomic_set(&folio->_pincount, 0); 20317118fc29SMike Kravetz return true; 20327118fc29SMike Kravetz 20337118fc29SMike Kravetz out_error: 20342b21624fSMike Kravetz /* undo page modifications made above */ 20352b21624fSMike Kravetz for (j = 0; j < i; j++) { 2036d1c60955SSidhartha Kumar p = folio_page(folio, j); 20372b21624fSMike Kravetz if (j != 0) 20387118fc29SMike Kravetz clear_compound_head(p); 20397118fc29SMike Kravetz set_page_refcounted(p); 20407118fc29SMike Kravetz } 20417118fc29SMike Kravetz /* need to clear PG_reserved on remaining tail pages */ 204214455eabSCheng Li for (; j < nr_pages; j++) { 2043d1c60955SSidhartha Kumar p = folio_page(folio, j); 20447118fc29SMike Kravetz __ClearPageReserved(p); 204514455eabSCheng Li } 20467118fc29SMike Kravetz return false; 204720a0307cSWu Fengguang } 204820a0307cSWu Fengguang 2049d1c60955SSidhartha Kumar static bool prep_compound_gigantic_folio(struct folio *folio, 20508531fc6fSMike Kravetz unsigned int order) 20518531fc6fSMike Kravetz { 2052d1c60955SSidhartha Kumar return __prep_compound_gigantic_folio(folio, order, false); 2053d1c60955SSidhartha Kumar } 2054d1c60955SSidhartha Kumar 2055d1c60955SSidhartha Kumar static bool prep_compound_gigantic_folio_for_demote(struct folio *folio, 2056d1c60955SSidhartha Kumar unsigned int order) 2057d1c60955SSidhartha Kumar { 2058d1c60955SSidhartha Kumar return __prep_compound_gigantic_folio(folio, order, true); 20598531fc6fSMike Kravetz } 20608531fc6fSMike Kravetz 20617795912cSAndrew Morton /* 20627795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 20637795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 20647795912cSAndrew Morton * details. 20657795912cSAndrew Morton */ 206620a0307cSWu Fengguang int PageHuge(struct page *page) 206720a0307cSWu Fengguang { 20682d678c64SMatthew Wilcox (Oracle) struct folio *folio; 20692d678c64SMatthew Wilcox (Oracle) 207020a0307cSWu Fengguang if (!PageCompound(page)) 207120a0307cSWu Fengguang return 0; 20722d678c64SMatthew Wilcox (Oracle) folio = page_folio(page); 20732d678c64SMatthew Wilcox (Oracle) return folio->_folio_dtor == HUGETLB_PAGE_DTOR; 207420a0307cSWu Fengguang } 207543131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 207643131e14SNaoya Horiguchi 2077957ebbdfSMatthew Wilcox (Oracle) /** 2078957ebbdfSMatthew Wilcox (Oracle) * folio_test_hugetlb - Determine if the folio belongs to hugetlbfs 2079957ebbdfSMatthew Wilcox (Oracle) * @folio: The folio to test. 2080957ebbdfSMatthew Wilcox (Oracle) * 2081957ebbdfSMatthew Wilcox (Oracle) * Context: Any context. Caller should have a reference on the folio to 2082957ebbdfSMatthew Wilcox (Oracle) * prevent it from being turned into a tail page. 2083957ebbdfSMatthew Wilcox (Oracle) * Return: True for hugetlbfs folios, false for anon folios or folios 2084957ebbdfSMatthew Wilcox (Oracle) * belonging to other filesystems. 208527c73ae7SAndrea Arcangeli */ 2086957ebbdfSMatthew Wilcox (Oracle) bool folio_test_hugetlb(struct folio *folio) 208727c73ae7SAndrea Arcangeli { 20882d678c64SMatthew Wilcox (Oracle) if (!folio_test_large(folio)) 2089957ebbdfSMatthew Wilcox (Oracle) return false; 209027c73ae7SAndrea Arcangeli 20912d678c64SMatthew Wilcox (Oracle) return folio->_folio_dtor == HUGETLB_PAGE_DTOR; 209227c73ae7SAndrea Arcangeli } 2093957ebbdfSMatthew Wilcox (Oracle) EXPORT_SYMBOL_GPL(folio_test_hugetlb); 209427c73ae7SAndrea Arcangeli 2095c0d0381aSMike Kravetz /* 2096c0d0381aSMike Kravetz * Find and lock address space (mapping) in write mode. 2097c0d0381aSMike Kravetz * 2098336bf30eSMike Kravetz * Upon entry, the page is locked which means that page_mapping() is 2099336bf30eSMike Kravetz * stable. Due to locking order, we can only trylock_write. If we can 2100336bf30eSMike Kravetz * not get the lock, simply return NULL to caller. 2101c0d0381aSMike Kravetz */ 2102c0d0381aSMike Kravetz struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) 2103c0d0381aSMike Kravetz { 2104336bf30eSMike Kravetz struct address_space *mapping = page_mapping(hpage); 2105c0d0381aSMike Kravetz 2106c0d0381aSMike Kravetz if (!mapping) 2107c0d0381aSMike Kravetz return mapping; 2108c0d0381aSMike Kravetz 2109c0d0381aSMike Kravetz if (i_mmap_trylock_write(mapping)) 2110c0d0381aSMike Kravetz return mapping; 2111c0d0381aSMike Kravetz 2112c0d0381aSMike Kravetz return NULL; 2113c0d0381aSMike Kravetz } 2114c0d0381aSMike Kravetz 2115fe19bd3dSHugh Dickins pgoff_t hugetlb_basepage_index(struct page *page) 211613d60f4bSZhang Yi { 211713d60f4bSZhang Yi struct page *page_head = compound_head(page); 211813d60f4bSZhang Yi pgoff_t index = page_index(page_head); 211913d60f4bSZhang Yi unsigned long compound_idx; 212013d60f4bSZhang Yi 212123baf831SKirill A. Shutemov if (compound_order(page_head) > MAX_ORDER) 212213d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 212313d60f4bSZhang Yi else 212413d60f4bSZhang Yi compound_idx = page - page_head; 212513d60f4bSZhang Yi 212613d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 212713d60f4bSZhang Yi } 212813d60f4bSZhang Yi 212919fc1a7eSSidhartha Kumar static struct folio *alloc_buddy_hugetlb_folio(struct hstate *h, 2130f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 2131f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 21321da177e4SLinus Torvalds { 2133af0fb9dfSMichal Hocko int order = huge_page_order(h); 21341da177e4SLinus Torvalds struct page *page; 2135f60858f9SMike Kravetz bool alloc_try_hard = true; 21362b21624fSMike Kravetz bool retry = true; 2137f96efd58SJoe Jin 2138f60858f9SMike Kravetz /* 2139f60858f9SMike Kravetz * By default we always try hard to allocate the page with 2140f60858f9SMike Kravetz * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in 2141f60858f9SMike Kravetz * a loop (to adjust global huge page counts) and previous allocation 2142f60858f9SMike Kravetz * failed, do not continue to try hard on the same node. Use the 2143f60858f9SMike Kravetz * node_alloc_noretry bitmap to manage this state information. 2144f60858f9SMike Kravetz */ 2145f60858f9SMike Kravetz if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) 2146f60858f9SMike Kravetz alloc_try_hard = false; 2147f60858f9SMike Kravetz gfp_mask |= __GFP_COMP|__GFP_NOWARN; 2148f60858f9SMike Kravetz if (alloc_try_hard) 2149f60858f9SMike Kravetz gfp_mask |= __GFP_RETRY_MAYFAIL; 2150af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 2151af0fb9dfSMichal Hocko nid = numa_mem_id(); 21522b21624fSMike Kravetz retry: 215384172f4bSMatthew Wilcox (Oracle) page = __alloc_pages(gfp_mask, order, nid, nmask); 21542b21624fSMike Kravetz 21552b21624fSMike Kravetz /* Freeze head page */ 21562b21624fSMike Kravetz if (page && !page_ref_freeze(page, 1)) { 21572b21624fSMike Kravetz __free_pages(page, order); 21582b21624fSMike Kravetz if (retry) { /* retry once */ 21592b21624fSMike Kravetz retry = false; 21602b21624fSMike Kravetz goto retry; 21612b21624fSMike Kravetz } 21622b21624fSMike Kravetz /* WOW! twice in a row. */ 21632b21624fSMike Kravetz pr_warn("HugeTLB head page unexpected inflated ref count\n"); 21642b21624fSMike Kravetz page = NULL; 21652b21624fSMike Kravetz } 21662b21624fSMike Kravetz 2167f60858f9SMike Kravetz /* 2168f60858f9SMike Kravetz * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this 2169f60858f9SMike Kravetz * indicates an overall state change. Clear bit so that we resume 2170f60858f9SMike Kravetz * normal 'try hard' allocations. 2171f60858f9SMike Kravetz */ 2172f60858f9SMike Kravetz if (node_alloc_noretry && page && !alloc_try_hard) 2173f60858f9SMike Kravetz node_clear(nid, *node_alloc_noretry); 2174f60858f9SMike Kravetz 2175f60858f9SMike Kravetz /* 2176f60858f9SMike Kravetz * If we tried hard to get a page but failed, set bit so that 2177f60858f9SMike Kravetz * subsequent attempts will not try as hard until there is an 2178f60858f9SMike Kravetz * overall state change. 2179f60858f9SMike Kravetz */ 2180f60858f9SMike Kravetz if (node_alloc_noretry && !page && alloc_try_hard) 2181f60858f9SMike Kravetz node_set(nid, *node_alloc_noretry); 2182f60858f9SMike Kravetz 218319fc1a7eSSidhartha Kumar if (!page) { 218419fc1a7eSSidhartha Kumar __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 218519fc1a7eSSidhartha Kumar return NULL; 218619fc1a7eSSidhartha Kumar } 218719fc1a7eSSidhartha Kumar 218819fc1a7eSSidhartha Kumar __count_vm_event(HTLB_BUDDY_PGALLOC); 218919fc1a7eSSidhartha Kumar return page_folio(page); 219063b4613cSNishanth Aravamudan } 219163b4613cSNishanth Aravamudan 2192af0fb9dfSMichal Hocko /* 21930c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 21940c397daeSMichal Hocko * should use this function to get new hugetlb pages 21952b21624fSMike Kravetz * 21962b21624fSMike Kravetz * Note that returned page is 'frozen': ref count of head page and all tail 21972b21624fSMike Kravetz * pages is zero. 21980c397daeSMichal Hocko */ 219919fc1a7eSSidhartha Kumar static struct folio *alloc_fresh_hugetlb_folio(struct hstate *h, 2200f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 2201f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 22020c397daeSMichal Hocko { 22037f325a8dSSidhartha Kumar struct folio *folio; 22047118fc29SMike Kravetz bool retry = false; 22050c397daeSMichal Hocko 22067118fc29SMike Kravetz retry: 22070c397daeSMichal Hocko if (hstate_is_gigantic(h)) 220819fc1a7eSSidhartha Kumar folio = alloc_gigantic_folio(h, gfp_mask, nid, nmask); 22090c397daeSMichal Hocko else 221019fc1a7eSSidhartha Kumar folio = alloc_buddy_hugetlb_folio(h, gfp_mask, 2211f60858f9SMike Kravetz nid, nmask, node_alloc_noretry); 221219fc1a7eSSidhartha Kumar if (!folio) 22130c397daeSMichal Hocko return NULL; 22147118fc29SMike Kravetz if (hstate_is_gigantic(h)) { 2215d1c60955SSidhartha Kumar if (!prep_compound_gigantic_folio(folio, huge_page_order(h))) { 22167118fc29SMike Kravetz /* 22177118fc29SMike Kravetz * Rare failure to convert pages to compound page. 22187118fc29SMike Kravetz * Free pages and try again - ONCE! 22197118fc29SMike Kravetz */ 22207f325a8dSSidhartha Kumar free_gigantic_folio(folio, huge_page_order(h)); 22217118fc29SMike Kravetz if (!retry) { 22227118fc29SMike Kravetz retry = true; 22237118fc29SMike Kravetz goto retry; 22247118fc29SMike Kravetz } 22257118fc29SMike Kravetz return NULL; 22267118fc29SMike Kravetz } 22277118fc29SMike Kravetz } 2228d1c60955SSidhartha Kumar prep_new_hugetlb_folio(h, folio, folio_nid(folio)); 22290c397daeSMichal Hocko 223019fc1a7eSSidhartha Kumar return folio; 22310c397daeSMichal Hocko } 22320c397daeSMichal Hocko 22330c397daeSMichal Hocko /* 2234af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 2235af0fb9dfSMichal Hocko * manner. 2236af0fb9dfSMichal Hocko */ 2237f60858f9SMike Kravetz static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 2238f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 2239b2261026SJoonsoo Kim { 224019fc1a7eSSidhartha Kumar struct folio *folio; 2241b2261026SJoonsoo Kim int nr_nodes, node; 2242af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2243b2261026SJoonsoo Kim 2244b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 224519fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, node, 224619fc1a7eSSidhartha Kumar nodes_allowed, node_alloc_noretry); 224719fc1a7eSSidhartha Kumar if (folio) { 2248*454a00c4SMatthew Wilcox (Oracle) free_huge_folio(folio); /* free it into the hugepage allocator */ 224919fc1a7eSSidhartha Kumar return 1; 225019fc1a7eSSidhartha Kumar } 2251b2261026SJoonsoo Kim } 2252b2261026SJoonsoo Kim 2253af0fb9dfSMichal Hocko return 0; 2254b2261026SJoonsoo Kim } 2255b2261026SJoonsoo Kim 2256e8c5c824SLee Schermerhorn /* 225710c6ec49SMike Kravetz * Remove huge page from pool from next node to free. Attempt to keep 225810c6ec49SMike Kravetz * persistent huge pages more or less balanced over allowed nodes. 225910c6ec49SMike Kravetz * This routine only 'removes' the hugetlb page. The caller must make 226010c6ec49SMike Kravetz * an additional call to free the page to low level allocators. 2261e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 2262e8c5c824SLee Schermerhorn */ 226310c6ec49SMike Kravetz static struct page *remove_pool_huge_page(struct hstate *h, 226410c6ec49SMike Kravetz nodemask_t *nodes_allowed, 22656ae11b27SLee Schermerhorn bool acct_surplus) 2266e8c5c824SLee Schermerhorn { 2267b2261026SJoonsoo Kim int nr_nodes, node; 226810c6ec49SMike Kravetz struct page *page = NULL; 2269cfd5082bSSidhartha Kumar struct folio *folio; 2270e8c5c824SLee Schermerhorn 22719487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2272b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2273685f3457SLee Schermerhorn /* 2274685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 2275685f3457SLee Schermerhorn * nodes with surplus pages. 2276685f3457SLee Schermerhorn */ 2277b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 2278b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 227910c6ec49SMike Kravetz page = list_entry(h->hugepage_freelists[node].next, 2280e8c5c824SLee Schermerhorn struct page, lru); 2281cfd5082bSSidhartha Kumar folio = page_folio(page); 2282cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, acct_surplus); 22839a76db09SLee Schermerhorn break; 2284e8c5c824SLee Schermerhorn } 2285b2261026SJoonsoo Kim } 2286e8c5c824SLee Schermerhorn 228710c6ec49SMike Kravetz return page; 2288e8c5c824SLee Schermerhorn } 2289e8c5c824SLee Schermerhorn 2290c8721bbbSNaoya Horiguchi /* 2291c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 2292faf53defSNaoya Horiguchi * nothing for in-use hugepages and non-hugepages. 2293faf53defSNaoya Horiguchi * This function returns values like below: 2294faf53defSNaoya Horiguchi * 2295ad2fa371SMuchun Song * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages 2296ad2fa371SMuchun Song * when the system is under memory pressure and the feature of 2297ad2fa371SMuchun Song * freeing unused vmemmap pages associated with each hugetlb page 2298ad2fa371SMuchun Song * is enabled. 2299faf53defSNaoya Horiguchi * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 2300faf53defSNaoya Horiguchi * (allocated or reserved.) 2301faf53defSNaoya Horiguchi * 0: successfully dissolved free hugepages or the page is not a 2302faf53defSNaoya Horiguchi * hugepage (considered as already dissolved) 2303c8721bbbSNaoya Horiguchi */ 2304c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 2305c8721bbbSNaoya Horiguchi { 23066bc9b564SNaoya Horiguchi int rc = -EBUSY; 23071a7cdab5SSidhartha Kumar struct folio *folio = page_folio(page); 2308082d5b6bSGerald Schaefer 23097ffddd49SMuchun Song retry: 2310faf53defSNaoya Horiguchi /* Not to disrupt normal path by vainly holding hugetlb_lock */ 23111a7cdab5SSidhartha Kumar if (!folio_test_hugetlb(folio)) 2312faf53defSNaoya Horiguchi return 0; 2313faf53defSNaoya Horiguchi 2314db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 23151a7cdab5SSidhartha Kumar if (!folio_test_hugetlb(folio)) { 2316faf53defSNaoya Horiguchi rc = 0; 2317faf53defSNaoya Horiguchi goto out; 2318faf53defSNaoya Horiguchi } 2319faf53defSNaoya Horiguchi 23201a7cdab5SSidhartha Kumar if (!folio_ref_count(folio)) { 23211a7cdab5SSidhartha Kumar struct hstate *h = folio_hstate(folio); 23228346d69dSXin Hao if (!available_huge_pages(h)) 2323082d5b6bSGerald Schaefer goto out; 23247ffddd49SMuchun Song 23257ffddd49SMuchun Song /* 23267ffddd49SMuchun Song * We should make sure that the page is already on the free list 23277ffddd49SMuchun Song * when it is dissolved. 23287ffddd49SMuchun Song */ 23291a7cdab5SSidhartha Kumar if (unlikely(!folio_test_hugetlb_freed(folio))) { 2330db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 23317ffddd49SMuchun Song cond_resched(); 23327ffddd49SMuchun Song 23337ffddd49SMuchun Song /* 23347ffddd49SMuchun Song * Theoretically, we should return -EBUSY when we 23357ffddd49SMuchun Song * encounter this race. In fact, we have a chance 23367ffddd49SMuchun Song * to successfully dissolve the page if we do a 23377ffddd49SMuchun Song * retry. Because the race window is quite small. 23387ffddd49SMuchun Song * If we seize this opportunity, it is an optimization 23397ffddd49SMuchun Song * for increasing the success rate of dissolving page. 23407ffddd49SMuchun Song */ 23417ffddd49SMuchun Song goto retry; 23427ffddd49SMuchun Song } 23437ffddd49SMuchun Song 2344cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, false); 2345ad2fa371SMuchun Song h->max_huge_pages--; 2346ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2347ad2fa371SMuchun Song 2348c3114a84SAnshuman Khandual /* 2349d6ef19e2SSidhartha Kumar * Normally update_and_free_hugtlb_folio will allocate required vmemmmap 2350d6ef19e2SSidhartha Kumar * before freeing the page. update_and_free_hugtlb_folio will fail to 2351ad2fa371SMuchun Song * free the page if it can not allocate required vmemmap. We 2352ad2fa371SMuchun Song * need to adjust max_huge_pages if the page is not freed. 2353ad2fa371SMuchun Song * Attempt to allocate vmemmmap here so that we can take 2354ad2fa371SMuchun Song * appropriate action on failure. 2355ad2fa371SMuchun Song */ 23561a7cdab5SSidhartha Kumar rc = hugetlb_vmemmap_restore(h, &folio->page); 2357ad2fa371SMuchun Song if (!rc) { 2358d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, folio, false); 2359ad2fa371SMuchun Song } else { 2360ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 23612f6c57d6SSidhartha Kumar add_hugetlb_folio(h, folio, false); 2362ad2fa371SMuchun Song h->max_huge_pages++; 2363ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2364ad2fa371SMuchun Song } 2365ad2fa371SMuchun Song 2366ad2fa371SMuchun Song return rc; 2367c8721bbbSNaoya Horiguchi } 2368082d5b6bSGerald Schaefer out: 2369db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2370082d5b6bSGerald Schaefer return rc; 2371c8721bbbSNaoya Horiguchi } 2372c8721bbbSNaoya Horiguchi 2373c8721bbbSNaoya Horiguchi /* 2374c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 2375c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 23762247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 23772247bb33SGerald Schaefer * part of it lies within the given range. 2378082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 2379082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 2380c8721bbbSNaoya Horiguchi */ 2381082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 2382c8721bbbSNaoya Horiguchi { 2383c8721bbbSNaoya Horiguchi unsigned long pfn; 2384eb03aa00SGerald Schaefer struct page *page; 2385082d5b6bSGerald Schaefer int rc = 0; 2386dc2628f3SMuchun Song unsigned int order; 2387dc2628f3SMuchun Song struct hstate *h; 2388c8721bbbSNaoya Horiguchi 2389d0177639SLi Zhong if (!hugepages_supported()) 2390082d5b6bSGerald Schaefer return rc; 2391d0177639SLi Zhong 2392dc2628f3SMuchun Song order = huge_page_order(&default_hstate); 2393dc2628f3SMuchun Song for_each_hstate(h) 2394dc2628f3SMuchun Song order = min(order, huge_page_order(h)); 2395dc2628f3SMuchun Song 2396dc2628f3SMuchun Song for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) { 2397eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 2398eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 2399eb03aa00SGerald Schaefer if (rc) 2400082d5b6bSGerald Schaefer break; 2401eb03aa00SGerald Schaefer } 2402082d5b6bSGerald Schaefer 2403082d5b6bSGerald Schaefer return rc; 2404c8721bbbSNaoya Horiguchi } 2405c8721bbbSNaoya Horiguchi 2406ab5ac90aSMichal Hocko /* 2407ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 2408ab5ac90aSMichal Hocko */ 24093a740e8bSSidhartha Kumar static struct folio *alloc_surplus_hugetlb_folio(struct hstate *h, 24103a740e8bSSidhartha Kumar gfp_t gfp_mask, int nid, nodemask_t *nmask) 24117893d1d5SAdam Litke { 241219fc1a7eSSidhartha Kumar struct folio *folio = NULL; 24137893d1d5SAdam Litke 2414bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2415aa888a74SAndi Kleen return NULL; 2416aa888a74SAndi Kleen 2417db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 24189980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 24199980d744SMichal Hocko goto out_unlock; 2420db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2421d1c3fb1fSNishanth Aravamudan 242219fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL); 242319fc1a7eSSidhartha Kumar if (!folio) 24240c397daeSMichal Hocko return NULL; 2425d1c3fb1fSNishanth Aravamudan 2426db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 24279980d744SMichal Hocko /* 24289980d744SMichal Hocko * We could have raced with the pool size change. 24299980d744SMichal Hocko * Double check that and simply deallocate the new page 24309980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 2431*454a00c4SMatthew Wilcox (Oracle) * temporary page to workaround the nasty free_huge_folio 24329980d744SMichal Hocko * codeflow 24339980d744SMichal Hocko */ 24349980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 243519fc1a7eSSidhartha Kumar folio_set_hugetlb_temporary(folio); 2436db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2437*454a00c4SMatthew Wilcox (Oracle) free_huge_folio(folio); 24382bf753e6SKai Shen return NULL; 2439b65a4edaSMike Kravetz } 2440b65a4edaSMike Kravetz 24419980d744SMichal Hocko h->surplus_huge_pages++; 244219fc1a7eSSidhartha Kumar h->surplus_huge_pages_node[folio_nid(folio)]++; 24439980d744SMichal Hocko 24449980d744SMichal Hocko out_unlock: 2445db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 24467893d1d5SAdam Litke 24473a740e8bSSidhartha Kumar return folio; 24487893d1d5SAdam Litke } 24497893d1d5SAdam Litke 2450e37d3e83SSidhartha Kumar static struct folio *alloc_migrate_hugetlb_folio(struct hstate *h, gfp_t gfp_mask, 2451ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 2452ab5ac90aSMichal Hocko { 245319fc1a7eSSidhartha Kumar struct folio *folio; 2454ab5ac90aSMichal Hocko 2455ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 2456ab5ac90aSMichal Hocko return NULL; 2457ab5ac90aSMichal Hocko 245819fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL); 245919fc1a7eSSidhartha Kumar if (!folio) 2460ab5ac90aSMichal Hocko return NULL; 2461ab5ac90aSMichal Hocko 24622b21624fSMike Kravetz /* fresh huge pages are frozen */ 246319fc1a7eSSidhartha Kumar folio_ref_unfreeze(folio, 1); 2464ab5ac90aSMichal Hocko /* 2465ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 2466ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 2467ab5ac90aSMichal Hocko */ 246819fc1a7eSSidhartha Kumar folio_set_hugetlb_temporary(folio); 2469ab5ac90aSMichal Hocko 2470e37d3e83SSidhartha Kumar return folio; 2471ab5ac90aSMichal Hocko } 2472ab5ac90aSMichal Hocko 2473e4e574b7SAdam Litke /* 2474099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 2475099730d6SDave Hansen */ 2476e0ec90eeSDave Hansen static 2477ff7d853bSSidhartha Kumar struct folio *alloc_buddy_hugetlb_folio_with_mpol(struct hstate *h, 2478099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 2479099730d6SDave Hansen { 24803a740e8bSSidhartha Kumar struct folio *folio = NULL; 2481aaf14e40SMichal Hocko struct mempolicy *mpol; 2482aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 2483aaf14e40SMichal Hocko int nid; 2484aaf14e40SMichal Hocko nodemask_t *nodemask; 2485aaf14e40SMichal Hocko 2486aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 2487cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 2488cfcaa66fSBen Widawsky gfp_t gfp = gfp_mask | __GFP_NOWARN; 2489cfcaa66fSBen Widawsky 2490cfcaa66fSBen Widawsky gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL); 24913a740e8bSSidhartha Kumar folio = alloc_surplus_hugetlb_folio(h, gfp, nid, nodemask); 2492cfcaa66fSBen Widawsky 2493cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 2494cfcaa66fSBen Widawsky nodemask = NULL; 2495cfcaa66fSBen Widawsky } 2496cfcaa66fSBen Widawsky 24973a740e8bSSidhartha Kumar if (!folio) 24983a740e8bSSidhartha Kumar folio = alloc_surplus_hugetlb_folio(h, gfp_mask, nid, nodemask); 2499aaf14e40SMichal Hocko mpol_cond_put(mpol); 2500ff7d853bSSidhartha Kumar return folio; 2501099730d6SDave Hansen } 2502099730d6SDave Hansen 2503e37d3e83SSidhartha Kumar /* folio migration callback function */ 2504e37d3e83SSidhartha Kumar struct folio *alloc_hugetlb_folio_nodemask(struct hstate *h, int preferred_nid, 2505d92bbc27SJoonsoo Kim nodemask_t *nmask, gfp_t gfp_mask) 25064db9b2efSMichal Hocko { 2507db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 25088346d69dSXin Hao if (available_huge_pages(h)) { 2509a36f1e90SSidhartha Kumar struct folio *folio; 25103e59fcb0SMichal Hocko 2511a36f1e90SSidhartha Kumar folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask, 2512a36f1e90SSidhartha Kumar preferred_nid, nmask); 2513a36f1e90SSidhartha Kumar if (folio) { 2514db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2515e37d3e83SSidhartha Kumar return folio; 25164db9b2efSMichal Hocko } 25174db9b2efSMichal Hocko } 2518db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 25194db9b2efSMichal Hocko 2520e37d3e83SSidhartha Kumar return alloc_migrate_hugetlb_folio(h, gfp_mask, preferred_nid, nmask); 25214db9b2efSMichal Hocko } 25224db9b2efSMichal Hocko 2523ebd63723SMichal Hocko /* mempolicy aware migration callback */ 2524d0ce0e47SSidhartha Kumar struct folio *alloc_hugetlb_folio_vma(struct hstate *h, struct vm_area_struct *vma, 2525389c8178SMichal Hocko unsigned long address) 2526ebd63723SMichal Hocko { 2527ebd63723SMichal Hocko struct mempolicy *mpol; 2528ebd63723SMichal Hocko nodemask_t *nodemask; 2529e37d3e83SSidhartha Kumar struct folio *folio; 2530ebd63723SMichal Hocko gfp_t gfp_mask; 2531ebd63723SMichal Hocko int node; 2532ebd63723SMichal Hocko 2533ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 2534ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 2535e37d3e83SSidhartha Kumar folio = alloc_hugetlb_folio_nodemask(h, node, nodemask, gfp_mask); 2536ebd63723SMichal Hocko mpol_cond_put(mpol); 2537ebd63723SMichal Hocko 2538d0ce0e47SSidhartha Kumar return folio; 2539ebd63723SMichal Hocko } 2540ebd63723SMichal Hocko 2541bf50bab2SNaoya Horiguchi /* 254225985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 2543e4e574b7SAdam Litke * of size 'delta'. 2544e4e574b7SAdam Litke */ 25450a4f3d1bSLiu Xiang static int gather_surplus_pages(struct hstate *h, long delta) 25461b2a1e7bSJules Irenge __must_hold(&hugetlb_lock) 2547e4e574b7SAdam Litke { 254834665341SMiaohe Lin LIST_HEAD(surplus_list); 2549*454a00c4SMatthew Wilcox (Oracle) struct folio *folio, *tmp; 25500a4f3d1bSLiu Xiang int ret; 25510a4f3d1bSLiu Xiang long i; 25520a4f3d1bSLiu Xiang long needed, allocated; 255328073b02SHillf Danton bool alloc_ok = true; 2554e4e574b7SAdam Litke 25559487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2556a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 2557ac09b3a1SAdam Litke if (needed <= 0) { 2558a5516438SAndi Kleen h->resv_huge_pages += delta; 2559e4e574b7SAdam Litke return 0; 2560ac09b3a1SAdam Litke } 2561e4e574b7SAdam Litke 2562e4e574b7SAdam Litke allocated = 0; 2563e4e574b7SAdam Litke 2564e4e574b7SAdam Litke ret = -ENOMEM; 2565e4e574b7SAdam Litke retry: 2566db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2567e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 25683a740e8bSSidhartha Kumar folio = alloc_surplus_hugetlb_folio(h, htlb_alloc_mask(h), 25692b21624fSMike Kravetz NUMA_NO_NODE, NULL); 25703a740e8bSSidhartha Kumar if (!folio) { 257128073b02SHillf Danton alloc_ok = false; 257228073b02SHillf Danton break; 257328073b02SHillf Danton } 25743a740e8bSSidhartha Kumar list_add(&folio->lru, &surplus_list); 257569ed779aSDavid Rientjes cond_resched(); 2576e4e574b7SAdam Litke } 257728073b02SHillf Danton allocated += i; 2578e4e574b7SAdam Litke 2579e4e574b7SAdam Litke /* 2580e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 2581e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 2582e4e574b7SAdam Litke */ 2583db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2584a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 2585a5516438SAndi Kleen (h->free_huge_pages + allocated); 258628073b02SHillf Danton if (needed > 0) { 258728073b02SHillf Danton if (alloc_ok) 2588e4e574b7SAdam Litke goto retry; 258928073b02SHillf Danton /* 259028073b02SHillf Danton * We were not able to allocate enough pages to 259128073b02SHillf Danton * satisfy the entire reservation so we free what 259228073b02SHillf Danton * we've allocated so far. 259328073b02SHillf Danton */ 259428073b02SHillf Danton goto free; 259528073b02SHillf Danton } 2596e4e574b7SAdam Litke /* 2597e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 259825985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 2599e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 2600ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 2601ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 2602ac09b3a1SAdam Litke * before they are reserved. 2603e4e574b7SAdam Litke */ 2604e4e574b7SAdam Litke needed += allocated; 2605a5516438SAndi Kleen h->resv_huge_pages += delta; 2606e4e574b7SAdam Litke ret = 0; 2607a9869b83SNaoya Horiguchi 260819fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 2609*454a00c4SMatthew Wilcox (Oracle) list_for_each_entry_safe(folio, tmp, &surplus_list, lru) { 261019fc3f0aSAdam Litke if ((--needed) < 0) 261119fc3f0aSAdam Litke break; 2612b65a4edaSMike Kravetz /* Add the page to the hugetlb allocator */ 2613*454a00c4SMatthew Wilcox (Oracle) enqueue_hugetlb_folio(h, folio); 261419fc3f0aSAdam Litke } 261528073b02SHillf Danton free: 2616db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 261719fc3f0aSAdam Litke 2618b65a4edaSMike Kravetz /* 2619b65a4edaSMike Kravetz * Free unnecessary surplus pages to the buddy allocator. 2620*454a00c4SMatthew Wilcox (Oracle) * Pages have no ref count, call free_huge_folio directly. 2621b65a4edaSMike Kravetz */ 2622*454a00c4SMatthew Wilcox (Oracle) list_for_each_entry_safe(folio, tmp, &surplus_list, lru) 2623*454a00c4SMatthew Wilcox (Oracle) free_huge_folio(folio); 2624db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2625e4e574b7SAdam Litke 2626e4e574b7SAdam Litke return ret; 2627e4e574b7SAdam Litke } 2628e4e574b7SAdam Litke 2629e4e574b7SAdam Litke /* 2630e5bbc8a6SMike Kravetz * This routine has two main purposes: 2631e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 2632e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 2633e5bbc8a6SMike Kravetz * to the associated reservation map. 2634e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 2635e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 2636e4e574b7SAdam Litke */ 2637a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 2638a5516438SAndi Kleen unsigned long unused_resv_pages) 2639e4e574b7SAdam Litke { 2640e4e574b7SAdam Litke unsigned long nr_pages; 264110c6ec49SMike Kravetz struct page *page; 264210c6ec49SMike Kravetz LIST_HEAD(page_list); 264310c6ec49SMike Kravetz 26449487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 264510c6ec49SMike Kravetz /* Uncommit the reservation */ 264610c6ec49SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 2647e4e574b7SAdam Litke 2648c0531714SNaoya Horiguchi if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 2649e5bbc8a6SMike Kravetz goto out; 2650aa888a74SAndi Kleen 2651e5bbc8a6SMike Kravetz /* 2652e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 2653e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 2654e5bbc8a6SMike Kravetz */ 2655a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 2656e4e574b7SAdam Litke 2657685f3457SLee Schermerhorn /* 2658685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 26599b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 26609b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 26619b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 266210c6ec49SMike Kravetz * remove_pool_huge_page() will balance the freed pages across the 26639b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 2664685f3457SLee Schermerhorn */ 2665685f3457SLee Schermerhorn while (nr_pages--) { 266610c6ec49SMike Kravetz page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1); 266710c6ec49SMike Kravetz if (!page) 2668e5bbc8a6SMike Kravetz goto out; 266910c6ec49SMike Kravetz 267010c6ec49SMike Kravetz list_add(&page->lru, &page_list); 2671e4e574b7SAdam Litke } 2672e5bbc8a6SMike Kravetz 2673e5bbc8a6SMike Kravetz out: 2674db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 267510c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 2676db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2677e4e574b7SAdam Litke } 2678e4e574b7SAdam Litke 26795e911373SMike Kravetz 2680c37f9fb1SAndy Whitcroft /* 2681feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 26825e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 2683cf3ad20bSMike Kravetz * 2684cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 2685cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 2686cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 2687cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 2688cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 2689feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 2690feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 2691feba16e2SMike Kravetz * is called in such cases. 2692cf3ad20bSMike Kravetz * 2693cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 2694cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 2695cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 2696cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 2697cf3ad20bSMike Kravetz * take appropriate action. 269896b96a96SMike Kravetz * 269996b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 270096b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 270196b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 270296b96a96SMike Kravetz * reservation exists. 2703846be085SMike Kravetz * 2704846be085SMike Kravetz * vma_del_reservation is used in error paths where an entry in the reserve 2705846be085SMike Kravetz * map was created during huge page allocation and must be removed. It is to 2706846be085SMike Kravetz * be called after calling vma_needs_reservation to determine if a reservation 2707846be085SMike Kravetz * exists. 2708c37f9fb1SAndy Whitcroft */ 27095e911373SMike Kravetz enum vma_resv_mode { 27105e911373SMike Kravetz VMA_NEEDS_RESV, 27115e911373SMike Kravetz VMA_COMMIT_RESV, 2712feba16e2SMike Kravetz VMA_END_RESV, 271396b96a96SMike Kravetz VMA_ADD_RESV, 2714846be085SMike Kravetz VMA_DEL_RESV, 27155e911373SMike Kravetz }; 2716cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 2717cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 27185e911373SMike Kravetz enum vma_resv_mode mode) 2719c37f9fb1SAndy Whitcroft { 27204e35f483SJoonsoo Kim struct resv_map *resv; 27214e35f483SJoonsoo Kim pgoff_t idx; 2722cf3ad20bSMike Kravetz long ret; 27230db9d74eSMina Almasry long dummy_out_regions_needed; 2724c37f9fb1SAndy Whitcroft 27254e35f483SJoonsoo Kim resv = vma_resv_map(vma); 27264e35f483SJoonsoo Kim if (!resv) 2727c37f9fb1SAndy Whitcroft return 1; 2728c37f9fb1SAndy Whitcroft 27294e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 27305e911373SMike Kravetz switch (mode) { 27315e911373SMike Kravetz case VMA_NEEDS_RESV: 27320db9d74eSMina Almasry ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); 27330db9d74eSMina Almasry /* We assume that vma_reservation_* routines always operate on 27340db9d74eSMina Almasry * 1 page, and that adding to resv map a 1 page entry can only 27350db9d74eSMina Almasry * ever require 1 region. 27360db9d74eSMina Almasry */ 27370db9d74eSMina Almasry VM_BUG_ON(dummy_out_regions_needed != 1); 27385e911373SMike Kravetz break; 27395e911373SMike Kravetz case VMA_COMMIT_RESV: 2740075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 27410db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 27420db9d74eSMina Almasry VM_BUG_ON(ret < 0); 27435e911373SMike Kravetz break; 2744feba16e2SMike Kravetz case VMA_END_RESV: 27450db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 27465e911373SMike Kravetz ret = 0; 27475e911373SMike Kravetz break; 274896b96a96SMike Kravetz case VMA_ADD_RESV: 27490db9d74eSMina Almasry if (vma->vm_flags & VM_MAYSHARE) { 2750075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 27510db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 27520db9d74eSMina Almasry VM_BUG_ON(ret < 0); 27530db9d74eSMina Almasry } else { 27540db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 275596b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 275696b96a96SMike Kravetz } 275796b96a96SMike Kravetz break; 2758846be085SMike Kravetz case VMA_DEL_RESV: 2759846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 2760846be085SMike Kravetz region_abort(resv, idx, idx + 1, 1); 2761846be085SMike Kravetz ret = region_del(resv, idx, idx + 1); 2762846be085SMike Kravetz } else { 2763846be085SMike Kravetz ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 2764846be085SMike Kravetz /* region_add calls of range 1 should never fail. */ 2765846be085SMike Kravetz VM_BUG_ON(ret < 0); 2766846be085SMike Kravetz } 2767846be085SMike Kravetz break; 27685e911373SMike Kravetz default: 27695e911373SMike Kravetz BUG(); 27705e911373SMike Kravetz } 277184afd99bSAndy Whitcroft 2772846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV) 2773cf3ad20bSMike Kravetz return ret; 277467961f9dSMike Kravetz /* 2775bf3d12b9SMiaohe Lin * We know private mapping must have HPAGE_RESV_OWNER set. 2776bf3d12b9SMiaohe Lin * 277767961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 277867961f9dSMike Kravetz * However, a file associated with mapping could have been 277967961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 278067961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 278167961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 278267961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 278367961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 278467961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 278567961f9dSMike Kravetz * reservation has already been consumed. As a result, the 278667961f9dSMike Kravetz * return value of this routine is the opposite of the 278767961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 278867961f9dSMike Kravetz */ 2789bf3d12b9SMiaohe Lin if (ret > 0) 279067961f9dSMike Kravetz return 0; 2791bf3d12b9SMiaohe Lin if (ret == 0) 279267961f9dSMike Kravetz return 1; 2793bf3d12b9SMiaohe Lin return ret; 279484afd99bSAndy Whitcroft } 2795cf3ad20bSMike Kravetz 2796cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 2797a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 2798c37f9fb1SAndy Whitcroft { 27995e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 2800cf3ad20bSMike Kravetz } 2801c37f9fb1SAndy Whitcroft 2802cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 2803cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 2804cf3ad20bSMike Kravetz { 28055e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 28065e911373SMike Kravetz } 28075e911373SMike Kravetz 2808feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 28095e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 28105e911373SMike Kravetz { 2811feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 2812c37f9fb1SAndy Whitcroft } 2813c37f9fb1SAndy Whitcroft 281496b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 281596b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 281696b96a96SMike Kravetz { 281796b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 281896b96a96SMike Kravetz } 281996b96a96SMike Kravetz 2820846be085SMike Kravetz static long vma_del_reservation(struct hstate *h, 2821846be085SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 282296b96a96SMike Kravetz { 2823846be085SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV); 2824846be085SMike Kravetz } 2825846be085SMike Kravetz 2826846be085SMike Kravetz /* 2827846be085SMike Kravetz * This routine is called to restore reservation information on error paths. 2828d0ce0e47SSidhartha Kumar * It should ONLY be called for folios allocated via alloc_hugetlb_folio(), 2829d0ce0e47SSidhartha Kumar * and the hugetlb mutex should remain held when calling this routine. 2830846be085SMike Kravetz * 2831846be085SMike Kravetz * It handles two specific cases: 2832d2d7bb44SSidhartha Kumar * 1) A reservation was in place and the folio consumed the reservation. 2833d2d7bb44SSidhartha Kumar * hugetlb_restore_reserve is set in the folio. 2834d2d7bb44SSidhartha Kumar * 2) No reservation was in place for the page, so hugetlb_restore_reserve is 2835d0ce0e47SSidhartha Kumar * not set. However, alloc_hugetlb_folio always updates the reserve map. 2836846be085SMike Kravetz * 2837*454a00c4SMatthew Wilcox (Oracle) * In case 1, free_huge_folio later in the error path will increment the 2838*454a00c4SMatthew Wilcox (Oracle) * global reserve count. But, free_huge_folio does not have enough context 2839846be085SMike Kravetz * to adjust the reservation map. This case deals primarily with private 2840846be085SMike Kravetz * mappings. Adjust the reserve map here to be consistent with global 2841*454a00c4SMatthew Wilcox (Oracle) * reserve count adjustments to be made by free_huge_folio. Make sure the 2842846be085SMike Kravetz * reserve map indicates there is a reservation present. 2843846be085SMike Kravetz * 2844d0ce0e47SSidhartha Kumar * In case 2, simply undo reserve map modifications done by alloc_hugetlb_folio. 2845846be085SMike Kravetz */ 2846846be085SMike Kravetz void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma, 2847d2d7bb44SSidhartha Kumar unsigned long address, struct folio *folio) 2848846be085SMike Kravetz { 284996b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 285096b96a96SMike Kravetz 28510ffdc38eSSidhartha Kumar if (folio_test_hugetlb_restore_reserve(folio)) { 2852846be085SMike Kravetz if (unlikely(rc < 0)) 285396b96a96SMike Kravetz /* 285496b96a96SMike Kravetz * Rare out of memory condition in reserve map 28550ffdc38eSSidhartha Kumar * manipulation. Clear hugetlb_restore_reserve so 28560ffdc38eSSidhartha Kumar * that global reserve count will not be incremented 2857*454a00c4SMatthew Wilcox (Oracle) * by free_huge_folio. This will make it appear 28580ffdc38eSSidhartha Kumar * as though the reservation for this folio was 285996b96a96SMike Kravetz * consumed. This may prevent the task from 28600ffdc38eSSidhartha Kumar * faulting in the folio at a later time. This 286196b96a96SMike Kravetz * is better than inconsistent global huge page 286296b96a96SMike Kravetz * accounting of reserve counts. 286396b96a96SMike Kravetz */ 28640ffdc38eSSidhartha Kumar folio_clear_hugetlb_restore_reserve(folio); 2865846be085SMike Kravetz else if (rc) 2866846be085SMike Kravetz (void)vma_add_reservation(h, vma, address); 2867846be085SMike Kravetz else 2868846be085SMike Kravetz vma_end_reservation(h, vma, address); 2869846be085SMike Kravetz } else { 2870846be085SMike Kravetz if (!rc) { 287196b96a96SMike Kravetz /* 2872846be085SMike Kravetz * This indicates there is an entry in the reserve map 2873d0ce0e47SSidhartha Kumar * not added by alloc_hugetlb_folio. We know it was added 2874d0ce0e47SSidhartha Kumar * before the alloc_hugetlb_folio call, otherwise 28750ffdc38eSSidhartha Kumar * hugetlb_restore_reserve would be set on the folio. 2876846be085SMike Kravetz * Remove the entry so that a subsequent allocation 2877846be085SMike Kravetz * does not consume a reservation. 287896b96a96SMike Kravetz */ 2879846be085SMike Kravetz rc = vma_del_reservation(h, vma, address); 2880846be085SMike Kravetz if (rc < 0) 2881846be085SMike Kravetz /* 2882846be085SMike Kravetz * VERY rare out of memory condition. Since 2883846be085SMike Kravetz * we can not delete the entry, set 28840ffdc38eSSidhartha Kumar * hugetlb_restore_reserve so that the reserve 28850ffdc38eSSidhartha Kumar * count will be incremented when the folio 2886846be085SMike Kravetz * is freed. This reserve will be consumed 2887846be085SMike Kravetz * on a subsequent allocation. 2888846be085SMike Kravetz */ 28890ffdc38eSSidhartha Kumar folio_set_hugetlb_restore_reserve(folio); 2890846be085SMike Kravetz } else if (rc < 0) { 2891846be085SMike Kravetz /* 2892846be085SMike Kravetz * Rare out of memory condition from 2893846be085SMike Kravetz * vma_needs_reservation call. Memory allocation is 2894846be085SMike Kravetz * only attempted if a new entry is needed. Therefore, 2895846be085SMike Kravetz * this implies there is not an entry in the 2896846be085SMike Kravetz * reserve map. 2897846be085SMike Kravetz * 2898846be085SMike Kravetz * For shared mappings, no entry in the map indicates 2899846be085SMike Kravetz * no reservation. We are done. 2900846be085SMike Kravetz */ 2901846be085SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 2902846be085SMike Kravetz /* 2903846be085SMike Kravetz * For private mappings, no entry indicates 2904846be085SMike Kravetz * a reservation is present. Since we can 29050ffdc38eSSidhartha Kumar * not add an entry, set hugetlb_restore_reserve 29060ffdc38eSSidhartha Kumar * on the folio so reserve count will be 2907846be085SMike Kravetz * incremented when freed. This reserve will 2908846be085SMike Kravetz * be consumed on a subsequent allocation. 2909846be085SMike Kravetz */ 29100ffdc38eSSidhartha Kumar folio_set_hugetlb_restore_reserve(folio); 291196b96a96SMike Kravetz } else 2912846be085SMike Kravetz /* 2913846be085SMike Kravetz * No reservation present, do nothing 2914846be085SMike Kravetz */ 291596b96a96SMike Kravetz vma_end_reservation(h, vma, address); 291696b96a96SMike Kravetz } 291796b96a96SMike Kravetz } 291896b96a96SMike Kravetz 2919369fa227SOscar Salvador /* 292019fc1a7eSSidhartha Kumar * alloc_and_dissolve_hugetlb_folio - Allocate a new folio and dissolve 292119fc1a7eSSidhartha Kumar * the old one 2922369fa227SOscar Salvador * @h: struct hstate old page belongs to 292319fc1a7eSSidhartha Kumar * @old_folio: Old folio to dissolve 2924ae37c7ffSOscar Salvador * @list: List to isolate the page in case we need to 2925369fa227SOscar Salvador * Returns 0 on success, otherwise negated error. 2926369fa227SOscar Salvador */ 292719fc1a7eSSidhartha Kumar static int alloc_and_dissolve_hugetlb_folio(struct hstate *h, 292819fc1a7eSSidhartha Kumar struct folio *old_folio, struct list_head *list) 2929369fa227SOscar Salvador { 2930369fa227SOscar Salvador gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2931de656ed3SSidhartha Kumar int nid = folio_nid(old_folio); 2932de656ed3SSidhartha Kumar struct folio *new_folio; 2933369fa227SOscar Salvador int ret = 0; 2934369fa227SOscar Salvador 2935369fa227SOscar Salvador /* 293619fc1a7eSSidhartha Kumar * Before dissolving the folio, we need to allocate a new one for the 293719fc1a7eSSidhartha Kumar * pool to remain stable. Here, we allocate the folio and 'prep' it 2938f41f2ed4SMuchun Song * by doing everything but actually updating counters and adding to 2939f41f2ed4SMuchun Song * the pool. This simplifies and let us do most of the processing 2940f41f2ed4SMuchun Song * under the lock. 2941369fa227SOscar Salvador */ 294219fc1a7eSSidhartha Kumar new_folio = alloc_buddy_hugetlb_folio(h, gfp_mask, nid, NULL, NULL); 294319fc1a7eSSidhartha Kumar if (!new_folio) 2944369fa227SOscar Salvador return -ENOMEM; 2945de656ed3SSidhartha Kumar __prep_new_hugetlb_folio(h, new_folio); 2946369fa227SOscar Salvador 2947369fa227SOscar Salvador retry: 2948369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2949de656ed3SSidhartha Kumar if (!folio_test_hugetlb(old_folio)) { 2950369fa227SOscar Salvador /* 295119fc1a7eSSidhartha Kumar * Freed from under us. Drop new_folio too. 2952369fa227SOscar Salvador */ 2953369fa227SOscar Salvador goto free_new; 2954de656ed3SSidhartha Kumar } else if (folio_ref_count(old_folio)) { 29559747b9e9SBaolin Wang bool isolated; 29569747b9e9SBaolin Wang 2957369fa227SOscar Salvador /* 295819fc1a7eSSidhartha Kumar * Someone has grabbed the folio, try to isolate it here. 2959ae37c7ffSOscar Salvador * Fail with -EBUSY if not possible. 2960369fa227SOscar Salvador */ 2961ae37c7ffSOscar Salvador spin_unlock_irq(&hugetlb_lock); 29629747b9e9SBaolin Wang isolated = isolate_hugetlb(old_folio, list); 29639747b9e9SBaolin Wang ret = isolated ? 0 : -EBUSY; 2964ae37c7ffSOscar Salvador spin_lock_irq(&hugetlb_lock); 2965369fa227SOscar Salvador goto free_new; 2966de656ed3SSidhartha Kumar } else if (!folio_test_hugetlb_freed(old_folio)) { 2967369fa227SOscar Salvador /* 296819fc1a7eSSidhartha Kumar * Folio's refcount is 0 but it has not been enqueued in the 2969369fa227SOscar Salvador * freelist yet. Race window is small, so we can succeed here if 2970369fa227SOscar Salvador * we retry. 2971369fa227SOscar Salvador */ 2972369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2973369fa227SOscar Salvador cond_resched(); 2974369fa227SOscar Salvador goto retry; 2975369fa227SOscar Salvador } else { 2976369fa227SOscar Salvador /* 297719fc1a7eSSidhartha Kumar * Ok, old_folio is still a genuine free hugepage. Remove it from 2978369fa227SOscar Salvador * the freelist and decrease the counters. These will be 2979369fa227SOscar Salvador * incremented again when calling __prep_account_new_huge_page() 2980240d67a8SSidhartha Kumar * and enqueue_hugetlb_folio() for new_folio. The counters will 2981240d67a8SSidhartha Kumar * remain stable since this happens under the lock. 2982369fa227SOscar Salvador */ 2983cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, old_folio, false); 2984369fa227SOscar Salvador 2985369fa227SOscar Salvador /* 298619fc1a7eSSidhartha Kumar * Ref count on new_folio is already zero as it was dropped 2987b65a4edaSMike Kravetz * earlier. It can be directly added to the pool free list. 2988369fa227SOscar Salvador */ 2989369fa227SOscar Salvador __prep_account_new_huge_page(h, nid); 2990240d67a8SSidhartha Kumar enqueue_hugetlb_folio(h, new_folio); 2991369fa227SOscar Salvador 2992369fa227SOscar Salvador /* 299319fc1a7eSSidhartha Kumar * Folio has been replaced, we can safely free the old one. 2994369fa227SOscar Salvador */ 2995369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2996d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, old_folio, false); 2997369fa227SOscar Salvador } 2998369fa227SOscar Salvador 2999369fa227SOscar Salvador return ret; 3000369fa227SOscar Salvador 3001369fa227SOscar Salvador free_new: 3002369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 300319fc1a7eSSidhartha Kumar /* Folio has a zero ref count, but needs a ref to be freed */ 3004de656ed3SSidhartha Kumar folio_ref_unfreeze(new_folio, 1); 3005d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, new_folio, false); 3006369fa227SOscar Salvador 3007369fa227SOscar Salvador return ret; 3008369fa227SOscar Salvador } 3009369fa227SOscar Salvador 3010ae37c7ffSOscar Salvador int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list) 3011369fa227SOscar Salvador { 3012369fa227SOscar Salvador struct hstate *h; 3013d5e33bd8SSidhartha Kumar struct folio *folio = page_folio(page); 3014ae37c7ffSOscar Salvador int ret = -EBUSY; 3015369fa227SOscar Salvador 3016369fa227SOscar Salvador /* 3017369fa227SOscar Salvador * The page might have been dissolved from under our feet, so make sure 3018369fa227SOscar Salvador * to carefully check the state under the lock. 3019369fa227SOscar Salvador * Return success when racing as if we dissolved the page ourselves. 3020369fa227SOscar Salvador */ 3021369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 3022d5e33bd8SSidhartha Kumar if (folio_test_hugetlb(folio)) { 3023d5e33bd8SSidhartha Kumar h = folio_hstate(folio); 3024369fa227SOscar Salvador } else { 3025369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 3026369fa227SOscar Salvador return 0; 3027369fa227SOscar Salvador } 3028369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 3029369fa227SOscar Salvador 3030369fa227SOscar Salvador /* 3031369fa227SOscar Salvador * Fence off gigantic pages as there is a cyclic dependency between 3032369fa227SOscar Salvador * alloc_contig_range and them. Return -ENOMEM as this has the effect 3033369fa227SOscar Salvador * of bailing out right away without further retrying. 3034369fa227SOscar Salvador */ 3035369fa227SOscar Salvador if (hstate_is_gigantic(h)) 3036369fa227SOscar Salvador return -ENOMEM; 3037369fa227SOscar Salvador 30389747b9e9SBaolin Wang if (folio_ref_count(folio) && isolate_hugetlb(folio, list)) 3039ae37c7ffSOscar Salvador ret = 0; 3040d5e33bd8SSidhartha Kumar else if (!folio_ref_count(folio)) 304119fc1a7eSSidhartha Kumar ret = alloc_and_dissolve_hugetlb_folio(h, folio, list); 3042ae37c7ffSOscar Salvador 3043ae37c7ffSOscar Salvador return ret; 3044369fa227SOscar Salvador } 3045369fa227SOscar Salvador 3046d0ce0e47SSidhartha Kumar struct folio *alloc_hugetlb_folio(struct vm_area_struct *vma, 304704f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 3048348ea204SAdam Litke { 304990481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 3050a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3051d4ab0316SSidhartha Kumar struct folio *folio; 3052d85f69b0SMike Kravetz long map_chg, map_commit; 3053d85f69b0SMike Kravetz long gbl_chg; 30546d76dcf4SAneesh Kumar K.V int ret, idx; 3055d0ce0e47SSidhartha Kumar struct hugetlb_cgroup *h_cg = NULL; 305608cf9fafSMina Almasry bool deferred_reserve; 30572fc39cecSAdam Litke 30586d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 3059a1e78772SMel Gorman /* 3060d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 3061d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 3062d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 3063a1e78772SMel Gorman */ 3064d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 3065d85f69b0SMike Kravetz if (map_chg < 0) 306676dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 3067d85f69b0SMike Kravetz 3068d85f69b0SMike Kravetz /* 3069d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 3070d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 3071d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 3072d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 3073d85f69b0SMike Kravetz * checked against any subpool limit. 3074d85f69b0SMike Kravetz */ 3075d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 3076d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 3077d85f69b0SMike Kravetz if (gbl_chg < 0) { 3078feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 307976dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 30805e911373SMike Kravetz } 308190d8b7e6SAdam Litke 3082d85f69b0SMike Kravetz /* 3083d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 3084d85f69b0SMike Kravetz * map, there could be reservations associated with the 3085d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 3086d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 3087d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 3088d85f69b0SMike Kravetz * the subpool reservations. 3089d85f69b0SMike Kravetz */ 3090d85f69b0SMike Kravetz if (avoid_reserve) 3091d85f69b0SMike Kravetz gbl_chg = 1; 3092d85f69b0SMike Kravetz } 3093d85f69b0SMike Kravetz 309408cf9fafSMina Almasry /* If this allocation is not consuming a reservation, charge it now. 309508cf9fafSMina Almasry */ 30966501fe5fSMiaohe Lin deferred_reserve = map_chg || avoid_reserve; 309708cf9fafSMina Almasry if (deferred_reserve) { 309808cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 309908cf9fafSMina Almasry idx, pages_per_huge_page(h), &h_cg); 31008f34af6fSJianyu Zhan if (ret) 31018f34af6fSJianyu Zhan goto out_subpool_put; 310208cf9fafSMina Almasry } 310308cf9fafSMina Almasry 310408cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 310508cf9fafSMina Almasry if (ret) 310608cf9fafSMina Almasry goto out_uncharge_cgroup_reservation; 31078f34af6fSJianyu Zhan 3108db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3109d85f69b0SMike Kravetz /* 3110d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 3111d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 3112d85f69b0SMike Kravetz * a reservation exists for the allocation. 3113d85f69b0SMike Kravetz */ 3114ff7d853bSSidhartha Kumar folio = dequeue_hugetlb_folio_vma(h, vma, addr, avoid_reserve, gbl_chg); 3115ff7d853bSSidhartha Kumar if (!folio) { 3116db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3117ff7d853bSSidhartha Kumar folio = alloc_buddy_hugetlb_folio_with_mpol(h, vma, addr); 3118ff7d853bSSidhartha Kumar if (!folio) 31198f34af6fSJianyu Zhan goto out_uncharge_cgroup; 312012df140fSRik van Riel spin_lock_irq(&hugetlb_lock); 3121a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 3122ff7d853bSSidhartha Kumar folio_set_hugetlb_restore_reserve(folio); 3123a88c7695SNaoya Horiguchi h->resv_huge_pages--; 3124a88c7695SNaoya Horiguchi } 3125ff7d853bSSidhartha Kumar list_add(&folio->lru, &h->hugepage_activelist); 3126ff7d853bSSidhartha Kumar folio_ref_unfreeze(folio, 1); 312781a6fcaeSJoonsoo Kim /* Fall through */ 3128a1e78772SMel Gorman } 3129ff7d853bSSidhartha Kumar 3130ff7d853bSSidhartha Kumar hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, folio); 313108cf9fafSMina Almasry /* If allocation is not consuming a reservation, also store the 313208cf9fafSMina Almasry * hugetlb_cgroup pointer on the page. 313308cf9fafSMina Almasry */ 313408cf9fafSMina Almasry if (deferred_reserve) { 313508cf9fafSMina Almasry hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), 3136ff7d853bSSidhartha Kumar h_cg, folio); 313708cf9fafSMina Almasry } 313808cf9fafSMina Almasry 3139db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3140a1e78772SMel Gorman 3141ff7d853bSSidhartha Kumar hugetlb_set_folio_subpool(folio, spool); 3142a1e78772SMel Gorman 3143d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 3144d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 314533039678SMike Kravetz /* 314633039678SMike Kravetz * The page was added to the reservation map between 314733039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 314833039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 314933039678SMike Kravetz * Adjust for the subpool count incremented above AND 315033039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 315133039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 315233039678SMike Kravetz * no longer applies. 315333039678SMike Kravetz */ 315433039678SMike Kravetz long rsv_adjust; 315533039678SMike Kravetz 315633039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 315733039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 315879aa925bSMike Kravetz if (deferred_reserve) 3159d4ab0316SSidhartha Kumar hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h), 3160d4ab0316SSidhartha Kumar pages_per_huge_page(h), folio); 316133039678SMike Kravetz } 3162d0ce0e47SSidhartha Kumar return folio; 31638f34af6fSJianyu Zhan 31648f34af6fSJianyu Zhan out_uncharge_cgroup: 31658f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 316608cf9fafSMina Almasry out_uncharge_cgroup_reservation: 316708cf9fafSMina Almasry if (deferred_reserve) 316808cf9fafSMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), 316908cf9fafSMina Almasry h_cg); 31708f34af6fSJianyu Zhan out_subpool_put: 3171d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 31728f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 3173feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 31748f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 3175b45b5bd6SDavid Gibson } 3176b45b5bd6SDavid Gibson 3177b5389086SZhenguo Yao int alloc_bootmem_huge_page(struct hstate *h, int nid) 3178e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 3179b5389086SZhenguo Yao int __alloc_bootmem_huge_page(struct hstate *h, int nid) 3180aa888a74SAndi Kleen { 3181b5389086SZhenguo Yao struct huge_bootmem_page *m = NULL; /* initialize for clang */ 3182b2261026SJoonsoo Kim int nr_nodes, node; 3183aa888a74SAndi Kleen 3184b5389086SZhenguo Yao /* do node specific alloc */ 3185b5389086SZhenguo Yao if (nid != NUMA_NO_NODE) { 3186b5389086SZhenguo Yao m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h), 3187b5389086SZhenguo Yao 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid); 3188b5389086SZhenguo Yao if (!m) 3189b5389086SZhenguo Yao return 0; 3190b5389086SZhenguo Yao goto found; 3191b5389086SZhenguo Yao } 3192b5389086SZhenguo Yao /* allocate from next node when distributing huge pages */ 3193b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 3194b5389086SZhenguo Yao m = memblock_alloc_try_nid_raw( 31958b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 319697ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 3197aa888a74SAndi Kleen /* 3198aa888a74SAndi Kleen * Use the beginning of the huge page to store the 3199aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 3200aa888a74SAndi Kleen * puts them into the mem_map). 3201aa888a74SAndi Kleen */ 3202b5389086SZhenguo Yao if (!m) 3203b5389086SZhenguo Yao return 0; 3204aa888a74SAndi Kleen goto found; 3205aa888a74SAndi Kleen } 3206aa888a74SAndi Kleen 3207aa888a74SAndi Kleen found: 3208aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 3209330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 3210aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 3211aa888a74SAndi Kleen m->hstate = h; 3212aa888a74SAndi Kleen return 1; 3213aa888a74SAndi Kleen } 3214aa888a74SAndi Kleen 321548b8d744SMike Kravetz /* 321648b8d744SMike Kravetz * Put bootmem huge pages into the standard lists after mem_map is up. 321748b8d744SMike Kravetz * Note: This only applies to gigantic (order > MAX_ORDER) pages. 321848b8d744SMike Kravetz */ 3219aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 3220aa888a74SAndi Kleen { 3221aa888a74SAndi Kleen struct huge_bootmem_page *m; 3222aa888a74SAndi Kleen 3223aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 322440d18ebfSMike Kravetz struct page *page = virt_to_page(m); 32257f325a8dSSidhartha Kumar struct folio *folio = page_folio(page); 3226aa888a74SAndi Kleen struct hstate *h = m->hstate; 3227ee8f248dSBecky Bruce 322848b8d744SMike Kravetz VM_BUG_ON(!hstate_is_gigantic(h)); 3229d1c60955SSidhartha Kumar WARN_ON(folio_ref_count(folio) != 1); 3230d1c60955SSidhartha Kumar if (prep_compound_gigantic_folio(folio, huge_page_order(h))) { 3231d1c60955SSidhartha Kumar WARN_ON(folio_test_reserved(folio)); 3232d1c60955SSidhartha Kumar prep_new_hugetlb_folio(h, folio, folio_nid(folio)); 3233*454a00c4SMatthew Wilcox (Oracle) free_huge_folio(folio); /* add to the hugepage allocator */ 32347118fc29SMike Kravetz } else { 3235416d85edSMike Kravetz /* VERY unlikely inflated ref count on a tail page */ 32367f325a8dSSidhartha Kumar free_gigantic_folio(folio, huge_page_order(h)); 32377118fc29SMike Kravetz } 3238af0fb9dfSMichal Hocko 3239b0320c7bSRafael Aquini /* 324048b8d744SMike Kravetz * We need to restore the 'stolen' pages to totalram_pages 324148b8d744SMike Kravetz * in order to fix confusing memory reports from free(1) and 324248b8d744SMike Kravetz * other side-effects, like CommitLimit going negative. 3243b0320c7bSRafael Aquini */ 3244c78a7f36SMiaohe Lin adjust_managed_page_count(page, pages_per_huge_page(h)); 3245520495feSCannon Matthews cond_resched(); 3246aa888a74SAndi Kleen } 3247aa888a74SAndi Kleen } 3248b5389086SZhenguo Yao static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid) 3249b5389086SZhenguo Yao { 3250b5389086SZhenguo Yao unsigned long i; 3251b5389086SZhenguo Yao char buf[32]; 3252b5389086SZhenguo Yao 3253b5389086SZhenguo Yao for (i = 0; i < h->max_huge_pages_node[nid]; ++i) { 3254b5389086SZhenguo Yao if (hstate_is_gigantic(h)) { 3255b5389086SZhenguo Yao if (!alloc_bootmem_huge_page(h, nid)) 3256b5389086SZhenguo Yao break; 3257b5389086SZhenguo Yao } else { 325819fc1a7eSSidhartha Kumar struct folio *folio; 3259b5389086SZhenguo Yao gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 3260b5389086SZhenguo Yao 326119fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, 3262b5389086SZhenguo Yao &node_states[N_MEMORY], NULL); 326319fc1a7eSSidhartha Kumar if (!folio) 3264b5389086SZhenguo Yao break; 3265*454a00c4SMatthew Wilcox (Oracle) free_huge_folio(folio); /* free it into the hugepage allocator */ 3266b5389086SZhenguo Yao } 3267b5389086SZhenguo Yao cond_resched(); 3268b5389086SZhenguo Yao } 3269b5389086SZhenguo Yao if (i == h->max_huge_pages_node[nid]) 3270b5389086SZhenguo Yao return; 3271b5389086SZhenguo Yao 3272b5389086SZhenguo Yao string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3273b5389086SZhenguo Yao pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n", 3274b5389086SZhenguo Yao h->max_huge_pages_node[nid], buf, nid, i); 3275b5389086SZhenguo Yao h->max_huge_pages -= (h->max_huge_pages_node[nid] - i); 3276b5389086SZhenguo Yao h->max_huge_pages_node[nid] = i; 3277b5389086SZhenguo Yao } 3278aa888a74SAndi Kleen 32798faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 32801da177e4SLinus Torvalds { 32811da177e4SLinus Torvalds unsigned long i; 3282f60858f9SMike Kravetz nodemask_t *node_alloc_noretry; 3283b5389086SZhenguo Yao bool node_specific_alloc = false; 3284f60858f9SMike Kravetz 3285b5389086SZhenguo Yao /* skip gigantic hugepages allocation if hugetlb_cma enabled */ 3286b5389086SZhenguo Yao if (hstate_is_gigantic(h) && hugetlb_cma_size) { 3287b5389086SZhenguo Yao pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); 3288b5389086SZhenguo Yao return; 3289b5389086SZhenguo Yao } 3290b5389086SZhenguo Yao 3291b5389086SZhenguo Yao /* do node specific alloc */ 32920a7a0f6fSPeng Liu for_each_online_node(i) { 3293b5389086SZhenguo Yao if (h->max_huge_pages_node[i] > 0) { 3294b5389086SZhenguo Yao hugetlb_hstate_alloc_pages_onenode(h, i); 3295b5389086SZhenguo Yao node_specific_alloc = true; 3296b5389086SZhenguo Yao } 3297b5389086SZhenguo Yao } 3298b5389086SZhenguo Yao 3299b5389086SZhenguo Yao if (node_specific_alloc) 3300b5389086SZhenguo Yao return; 3301b5389086SZhenguo Yao 3302b5389086SZhenguo Yao /* below will do all node balanced alloc */ 3303f60858f9SMike Kravetz if (!hstate_is_gigantic(h)) { 3304f60858f9SMike Kravetz /* 3305f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3306f60858f9SMike Kravetz * Ignore errors as lower level routines can deal with 3307f60858f9SMike Kravetz * node_alloc_noretry == NULL. If this kmalloc fails at boot 3308f60858f9SMike Kravetz * time, we are likely in bigger trouble. 3309f60858f9SMike Kravetz */ 3310f60858f9SMike Kravetz node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), 3311f60858f9SMike Kravetz GFP_KERNEL); 3312f60858f9SMike Kravetz } else { 3313f60858f9SMike Kravetz /* allocations done at boot time */ 3314f60858f9SMike Kravetz node_alloc_noretry = NULL; 3315f60858f9SMike Kravetz } 3316f60858f9SMike Kravetz 3317f60858f9SMike Kravetz /* bit mask controlling how hard we retry per-node allocations */ 3318f60858f9SMike Kravetz if (node_alloc_noretry) 3319f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 33201da177e4SLinus Torvalds 3321e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 3322bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 3323b5389086SZhenguo Yao if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE)) 3324aa888a74SAndi Kleen break; 33250c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 3326f60858f9SMike Kravetz &node_states[N_MEMORY], 3327f60858f9SMike Kravetz node_alloc_noretry)) 33281da177e4SLinus Torvalds break; 332969ed779aSDavid Rientjes cond_resched(); 33301da177e4SLinus Torvalds } 3331d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 3332d715cf80SLiam R. Howlett char buf[32]; 3333d715cf80SLiam R. Howlett 3334c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3335d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 3336d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 33378faa8b07SAndi Kleen h->max_huge_pages = i; 3338e5ff2159SAndi Kleen } 3339f60858f9SMike Kravetz kfree(node_alloc_noretry); 3340d715cf80SLiam R. Howlett } 3341e5ff2159SAndi Kleen 3342e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 3343e5ff2159SAndi Kleen { 334479dfc695SMike Kravetz struct hstate *h, *h2; 3345e5ff2159SAndi Kleen 3346e5ff2159SAndi Kleen for_each_hstate(h) { 33478faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 3348bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 33498faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 335079dfc695SMike Kravetz 335179dfc695SMike Kravetz /* 335279dfc695SMike Kravetz * Set demote order for each hstate. Note that 335379dfc695SMike Kravetz * h->demote_order is initially 0. 335479dfc695SMike Kravetz * - We can not demote gigantic pages if runtime freeing 335579dfc695SMike Kravetz * is not supported, so skip this. 3356a01f4390SMike Kravetz * - If CMA allocation is possible, we can not demote 3357a01f4390SMike Kravetz * HUGETLB_PAGE_ORDER or smaller size pages. 335879dfc695SMike Kravetz */ 335979dfc695SMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 336079dfc695SMike Kravetz continue; 3361a01f4390SMike Kravetz if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER) 3362a01f4390SMike Kravetz continue; 336379dfc695SMike Kravetz for_each_hstate(h2) { 336479dfc695SMike Kravetz if (h2 == h) 336579dfc695SMike Kravetz continue; 336679dfc695SMike Kravetz if (h2->order < h->order && 336779dfc695SMike Kravetz h2->order > h->demote_order) 336879dfc695SMike Kravetz h->demote_order = h2->order; 336979dfc695SMike Kravetz } 3370e5ff2159SAndi Kleen } 3371e5ff2159SAndi Kleen } 3372e5ff2159SAndi Kleen 3373e5ff2159SAndi Kleen static void __init report_hugepages(void) 3374e5ff2159SAndi Kleen { 3375e5ff2159SAndi Kleen struct hstate *h; 3376e5ff2159SAndi Kleen 3377e5ff2159SAndi Kleen for_each_hstate(h) { 33784abd32dbSAndi Kleen char buf[32]; 3379c6247f72SMatthew Wilcox 3380c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 33816213834cSMuchun Song pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n", 3382c6247f72SMatthew Wilcox buf, h->free_huge_pages); 33836213834cSMuchun Song pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n", 33846213834cSMuchun Song hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf); 3385e5ff2159SAndi Kleen } 3386e5ff2159SAndi Kleen } 3387e5ff2159SAndi Kleen 33881da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 33896ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 33906ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 33911da177e4SLinus Torvalds { 33924415cc8dSChristoph Lameter int i; 33931121828aSMike Kravetz LIST_HEAD(page_list); 33944415cc8dSChristoph Lameter 33959487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 3396bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3397aa888a74SAndi Kleen return; 3398aa888a74SAndi Kleen 33991121828aSMike Kravetz /* 34001121828aSMike Kravetz * Collect pages to be freed on a list, and free after dropping lock 34011121828aSMike Kravetz */ 34026ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 340310c6ec49SMike Kravetz struct page *page, *next; 3404a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 3405a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 3406a5516438SAndi Kleen if (count >= h->nr_huge_pages) 34071121828aSMike Kravetz goto out; 34081da177e4SLinus Torvalds if (PageHighMem(page)) 34091da177e4SLinus Torvalds continue; 3410cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, page_folio(page), false); 34111121828aSMike Kravetz list_add(&page->lru, &page_list); 34121121828aSMike Kravetz } 34131121828aSMike Kravetz } 34141121828aSMike Kravetz 34151121828aSMike Kravetz out: 3416db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 341710c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3418db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 34191da177e4SLinus Torvalds } 34201da177e4SLinus Torvalds #else 34216ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 34226ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 34231da177e4SLinus Torvalds { 34241da177e4SLinus Torvalds } 34251da177e4SLinus Torvalds #endif 34261da177e4SLinus Torvalds 342720a0307cSWu Fengguang /* 342820a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 342920a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 343020a0307cSWu Fengguang * Returns 1 if an adjustment was made. 343120a0307cSWu Fengguang */ 34326ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 34336ae11b27SLee Schermerhorn int delta) 343420a0307cSWu Fengguang { 3435b2261026SJoonsoo Kim int nr_nodes, node; 343620a0307cSWu Fengguang 34379487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 343820a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 343920a0307cSWu Fengguang 3440e8c5c824SLee Schermerhorn if (delta < 0) { 3441b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 3442b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 3443b2261026SJoonsoo Kim goto found; 3444b2261026SJoonsoo Kim } 3445b2261026SJoonsoo Kim } else { 3446b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 3447b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 3448b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 3449b2261026SJoonsoo Kim goto found; 3450e8c5c824SLee Schermerhorn } 34519a76db09SLee Schermerhorn } 3452b2261026SJoonsoo Kim return 0; 345320a0307cSWu Fengguang 3454b2261026SJoonsoo Kim found: 345520a0307cSWu Fengguang h->surplus_huge_pages += delta; 3456b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 3457b2261026SJoonsoo Kim return 1; 345820a0307cSWu Fengguang } 345920a0307cSWu Fengguang 3460a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 3461fd875dcaSMike Kravetz static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 34626ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 34631da177e4SLinus Torvalds { 34647893d1d5SAdam Litke unsigned long min_count, ret; 346510c6ec49SMike Kravetz struct page *page; 346610c6ec49SMike Kravetz LIST_HEAD(page_list); 3467f60858f9SMike Kravetz NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); 3468f60858f9SMike Kravetz 3469f60858f9SMike Kravetz /* 3470f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3471f60858f9SMike Kravetz * If we can not allocate the bit mask, do not attempt to allocate 3472f60858f9SMike Kravetz * the requested huge pages. 3473f60858f9SMike Kravetz */ 3474f60858f9SMike Kravetz if (node_alloc_noretry) 3475f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 3476f60858f9SMike Kravetz else 3477f60858f9SMike Kravetz return -ENOMEM; 34781da177e4SLinus Torvalds 347929383967SMike Kravetz /* 348029383967SMike Kravetz * resize_lock mutex prevents concurrent adjustments to number of 348129383967SMike Kravetz * pages in hstate via the proc/sysfs interfaces. 348229383967SMike Kravetz */ 348329383967SMike Kravetz mutex_lock(&h->resize_lock); 3484b65d4adbSMuchun Song flush_free_hpage_work(h); 3485db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 34864eb0716eSAlexandre Ghiti 34874eb0716eSAlexandre Ghiti /* 3488fd875dcaSMike Kravetz * Check for a node specific request. 3489fd875dcaSMike Kravetz * Changing node specific huge page count may require a corresponding 3490fd875dcaSMike Kravetz * change to the global count. In any case, the passed node mask 3491fd875dcaSMike Kravetz * (nodes_allowed) will restrict alloc/free to the specified node. 3492fd875dcaSMike Kravetz */ 3493fd875dcaSMike Kravetz if (nid != NUMA_NO_NODE) { 3494fd875dcaSMike Kravetz unsigned long old_count = count; 3495fd875dcaSMike Kravetz 3496fd875dcaSMike Kravetz count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 3497fd875dcaSMike Kravetz /* 3498fd875dcaSMike Kravetz * User may have specified a large count value which caused the 3499fd875dcaSMike Kravetz * above calculation to overflow. In this case, they wanted 3500fd875dcaSMike Kravetz * to allocate as many huge pages as possible. Set count to 3501fd875dcaSMike Kravetz * largest possible value to align with their intention. 3502fd875dcaSMike Kravetz */ 3503fd875dcaSMike Kravetz if (count < old_count) 3504fd875dcaSMike Kravetz count = ULONG_MAX; 3505fd875dcaSMike Kravetz } 3506fd875dcaSMike Kravetz 3507fd875dcaSMike Kravetz /* 35084eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 35094eb0716eSAlexandre Ghiti * page range allocation. 35104eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 35114eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 35124eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 35134eb0716eSAlexandre Ghiti */ 35144eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 35154eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 3516db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 351729383967SMike Kravetz mutex_unlock(&h->resize_lock); 3518f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 35194eb0716eSAlexandre Ghiti return -EINVAL; 35204eb0716eSAlexandre Ghiti } 35214eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 35224eb0716eSAlexandre Ghiti } 3523aa888a74SAndi Kleen 35247893d1d5SAdam Litke /* 35257893d1d5SAdam Litke * Increase the pool size 35267893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 35277893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 3528d1c3fb1fSNishanth Aravamudan * 35293a740e8bSSidhartha Kumar * We might race with alloc_surplus_hugetlb_folio() here and be unable 3530d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 3531d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 3532d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 3533d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 35347893d1d5SAdam Litke */ 3535a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 35366ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 35377893d1d5SAdam Litke break; 35387893d1d5SAdam Litke } 35397893d1d5SAdam Litke 3540a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 35417893d1d5SAdam Litke /* 35427893d1d5SAdam Litke * If this allocation races such that we no longer need the 3543*454a00c4SMatthew Wilcox (Oracle) * page, free_huge_folio will handle it by freeing the page 35447893d1d5SAdam Litke * and reducing the surplus. 35457893d1d5SAdam Litke */ 3546db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3547649920c6SJia He 3548649920c6SJia He /* yield cpu to avoid soft lockup */ 3549649920c6SJia He cond_resched(); 3550649920c6SJia He 3551f60858f9SMike Kravetz ret = alloc_pool_huge_page(h, nodes_allowed, 3552f60858f9SMike Kravetz node_alloc_noretry); 3553db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 35547893d1d5SAdam Litke if (!ret) 35557893d1d5SAdam Litke goto out; 35567893d1d5SAdam Litke 3557536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 3558536240f2SMel Gorman if (signal_pending(current)) 3559536240f2SMel Gorman goto out; 35607893d1d5SAdam Litke } 35617893d1d5SAdam Litke 35627893d1d5SAdam Litke /* 35637893d1d5SAdam Litke * Decrease the pool size 35647893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 35657893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 35667893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 35677893d1d5SAdam Litke * to the desired size as pages become free. 3568d1c3fb1fSNishanth Aravamudan * 3569d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 3570d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 3571d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 35723a740e8bSSidhartha Kumar * alloc_surplus_hugetlb_folio() is checking the global counter, 3573d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 3574d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 3575d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 35767893d1d5SAdam Litke */ 3577a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 35786b0c880dSAdam Litke min_count = max(count, min_count); 35796ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 358010c6ec49SMike Kravetz 358110c6ec49SMike Kravetz /* 358210c6ec49SMike Kravetz * Collect pages to be removed on list without dropping lock 358310c6ec49SMike Kravetz */ 3584a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 358510c6ec49SMike Kravetz page = remove_pool_huge_page(h, nodes_allowed, 0); 358610c6ec49SMike Kravetz if (!page) 35871da177e4SLinus Torvalds break; 358810c6ec49SMike Kravetz 358910c6ec49SMike Kravetz list_add(&page->lru, &page_list); 35901da177e4SLinus Torvalds } 359110c6ec49SMike Kravetz /* free the pages after dropping lock */ 3592db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 359310c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3594b65d4adbSMuchun Song flush_free_hpage_work(h); 3595db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 359610c6ec49SMike Kravetz 3597a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 35986ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 35997893d1d5SAdam Litke break; 36007893d1d5SAdam Litke } 36017893d1d5SAdam Litke out: 36024eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 3603db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 360429383967SMike Kravetz mutex_unlock(&h->resize_lock); 36054eb0716eSAlexandre Ghiti 3606f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 3607f60858f9SMike Kravetz 36084eb0716eSAlexandre Ghiti return 0; 36091da177e4SLinus Torvalds } 36101da177e4SLinus Torvalds 3611bdd7be07SSidhartha Kumar static int demote_free_hugetlb_folio(struct hstate *h, struct folio *folio) 36128531fc6fSMike Kravetz { 3613bdd7be07SSidhartha Kumar int i, nid = folio_nid(folio); 36148531fc6fSMike Kravetz struct hstate *target_hstate; 361531731452SDoug Berger struct page *subpage; 3616bdd7be07SSidhartha Kumar struct folio *inner_folio; 36178531fc6fSMike Kravetz int rc = 0; 36188531fc6fSMike Kravetz 36198531fc6fSMike Kravetz target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order); 36208531fc6fSMike Kravetz 3621cfd5082bSSidhartha Kumar remove_hugetlb_folio_for_demote(h, folio, false); 36228531fc6fSMike Kravetz spin_unlock_irq(&hugetlb_lock); 36238531fc6fSMike Kravetz 3624bdd7be07SSidhartha Kumar rc = hugetlb_vmemmap_restore(h, &folio->page); 36258531fc6fSMike Kravetz if (rc) { 3626bdd7be07SSidhartha Kumar /* Allocation of vmemmmap failed, we can not demote folio */ 36278531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 3628bdd7be07SSidhartha Kumar folio_ref_unfreeze(folio, 1); 3629bdd7be07SSidhartha Kumar add_hugetlb_folio(h, folio, false); 36308531fc6fSMike Kravetz return rc; 36318531fc6fSMike Kravetz } 36328531fc6fSMike Kravetz 36338531fc6fSMike Kravetz /* 3634911565b8SSidhartha Kumar * Use destroy_compound_hugetlb_folio_for_demote for all huge page 3635bdd7be07SSidhartha Kumar * sizes as it will not ref count folios. 36368531fc6fSMike Kravetz */ 3637911565b8SSidhartha Kumar destroy_compound_hugetlb_folio_for_demote(folio, huge_page_order(h)); 36388531fc6fSMike Kravetz 36398531fc6fSMike Kravetz /* 36408531fc6fSMike Kravetz * Taking target hstate mutex synchronizes with set_max_huge_pages. 36418531fc6fSMike Kravetz * Without the mutex, pages added to target hstate could be marked 36428531fc6fSMike Kravetz * as surplus. 36438531fc6fSMike Kravetz * 36448531fc6fSMike Kravetz * Note that we already hold h->resize_lock. To prevent deadlock, 36458531fc6fSMike Kravetz * use the convention of always taking larger size hstate mutex first. 36468531fc6fSMike Kravetz */ 36478531fc6fSMike Kravetz mutex_lock(&target_hstate->resize_lock); 36488531fc6fSMike Kravetz for (i = 0; i < pages_per_huge_page(h); 36498531fc6fSMike Kravetz i += pages_per_huge_page(target_hstate)) { 3650bdd7be07SSidhartha Kumar subpage = folio_page(folio, i); 3651bdd7be07SSidhartha Kumar inner_folio = page_folio(subpage); 36528531fc6fSMike Kravetz if (hstate_is_gigantic(target_hstate)) 3653bdd7be07SSidhartha Kumar prep_compound_gigantic_folio_for_demote(inner_folio, 36548531fc6fSMike Kravetz target_hstate->order); 36558531fc6fSMike Kravetz else 365631731452SDoug Berger prep_compound_page(subpage, target_hstate->order); 3657bdd7be07SSidhartha Kumar folio_change_private(inner_folio, NULL); 3658bdd7be07SSidhartha Kumar prep_new_hugetlb_folio(target_hstate, inner_folio, nid); 3659*454a00c4SMatthew Wilcox (Oracle) free_huge_folio(inner_folio); 36608531fc6fSMike Kravetz } 36618531fc6fSMike Kravetz mutex_unlock(&target_hstate->resize_lock); 36628531fc6fSMike Kravetz 36638531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 36648531fc6fSMike Kravetz 36658531fc6fSMike Kravetz /* 36668531fc6fSMike Kravetz * Not absolutely necessary, but for consistency update max_huge_pages 36678531fc6fSMike Kravetz * based on pool changes for the demoted page. 36688531fc6fSMike Kravetz */ 36698531fc6fSMike Kravetz h->max_huge_pages--; 3670a43a83c7SMiaohe Lin target_hstate->max_huge_pages += 3671a43a83c7SMiaohe Lin pages_per_huge_page(h) / pages_per_huge_page(target_hstate); 36728531fc6fSMike Kravetz 36738531fc6fSMike Kravetz return rc; 36748531fc6fSMike Kravetz } 36758531fc6fSMike Kravetz 367679dfc695SMike Kravetz static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 367779dfc695SMike Kravetz __must_hold(&hugetlb_lock) 367879dfc695SMike Kravetz { 36798531fc6fSMike Kravetz int nr_nodes, node; 3680bdd7be07SSidhartha Kumar struct folio *folio; 368179dfc695SMike Kravetz 368279dfc695SMike Kravetz lockdep_assert_held(&hugetlb_lock); 368379dfc695SMike Kravetz 368479dfc695SMike Kravetz /* We should never get here if no demote order */ 368579dfc695SMike Kravetz if (!h->demote_order) { 368679dfc695SMike Kravetz pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n"); 368779dfc695SMike Kravetz return -EINVAL; /* internal error */ 368879dfc695SMike Kravetz } 368979dfc695SMike Kravetz 36908531fc6fSMike Kravetz for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 3691bdd7be07SSidhartha Kumar list_for_each_entry(folio, &h->hugepage_freelists[node], lru) { 3692bdd7be07SSidhartha Kumar if (folio_test_hwpoison(folio)) 36935a317412SMike Kravetz continue; 3694bdd7be07SSidhartha Kumar return demote_free_hugetlb_folio(h, folio); 36958531fc6fSMike Kravetz } 36968531fc6fSMike Kravetz } 36978531fc6fSMike Kravetz 36985a317412SMike Kravetz /* 36995a317412SMike Kravetz * Only way to get here is if all pages on free lists are poisoned. 37005a317412SMike Kravetz * Return -EBUSY so that caller will not retry. 37015a317412SMike Kravetz */ 37025a317412SMike Kravetz return -EBUSY; 370379dfc695SMike Kravetz } 370479dfc695SMike Kravetz 3705a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 3706a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 3707a3437870SNishanth Aravamudan 370879dfc695SMike Kravetz #define HSTATE_ATTR_WO(_name) \ 370979dfc695SMike Kravetz static struct kobj_attribute _name##_attr = __ATTR_WO(_name) 371079dfc695SMike Kravetz 3711a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 371298bc26acSMiaohe Lin static struct kobj_attribute _name##_attr = __ATTR_RW(_name) 3713a3437870SNishanth Aravamudan 3714a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 3715a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 3716a3437870SNishanth Aravamudan 37179a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 37189a305230SLee Schermerhorn 37199a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 3720a3437870SNishanth Aravamudan { 3721a3437870SNishanth Aravamudan int i; 37229a305230SLee Schermerhorn 3723a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 37249a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 37259a305230SLee Schermerhorn if (nidp) 37269a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 3727a3437870SNishanth Aravamudan return &hstates[i]; 37289a305230SLee Schermerhorn } 37299a305230SLee Schermerhorn 37309a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 3731a3437870SNishanth Aravamudan } 3732a3437870SNishanth Aravamudan 373306808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 3734a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3735a3437870SNishanth Aravamudan { 37369a305230SLee Schermerhorn struct hstate *h; 37379a305230SLee Schermerhorn unsigned long nr_huge_pages; 37389a305230SLee Schermerhorn int nid; 37399a305230SLee Schermerhorn 37409a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 37419a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 37429a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 37439a305230SLee Schermerhorn else 37449a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 37459a305230SLee Schermerhorn 3746ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", nr_huge_pages); 3747a3437870SNishanth Aravamudan } 3748adbe8726SEric B Munson 3749238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 3750238d3c13SDavid Rientjes struct hstate *h, int nid, 3751238d3c13SDavid Rientjes unsigned long count, size_t len) 3752a3437870SNishanth Aravamudan { 3753a3437870SNishanth Aravamudan int err; 37542d0adf7eSOscar Salvador nodemask_t nodes_allowed, *n_mask; 3755a3437870SNishanth Aravamudan 37562d0adf7eSOscar Salvador if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 37572d0adf7eSOscar Salvador return -EINVAL; 3758adbe8726SEric B Munson 37599a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 37609a305230SLee Schermerhorn /* 37619a305230SLee Schermerhorn * global hstate attribute 37629a305230SLee Schermerhorn */ 37639a305230SLee Schermerhorn if (!(obey_mempolicy && 37642d0adf7eSOscar Salvador init_nodemask_of_mempolicy(&nodes_allowed))) 37652d0adf7eSOscar Salvador n_mask = &node_states[N_MEMORY]; 37662d0adf7eSOscar Salvador else 37672d0adf7eSOscar Salvador n_mask = &nodes_allowed; 37682d0adf7eSOscar Salvador } else { 37699a305230SLee Schermerhorn /* 3770fd875dcaSMike Kravetz * Node specific request. count adjustment happens in 3771fd875dcaSMike Kravetz * set_max_huge_pages() after acquiring hugetlb_lock. 37729a305230SLee Schermerhorn */ 37732d0adf7eSOscar Salvador init_nodemask_of_node(&nodes_allowed, nid); 37742d0adf7eSOscar Salvador n_mask = &nodes_allowed; 3775fd875dcaSMike Kravetz } 37769a305230SLee Schermerhorn 37772d0adf7eSOscar Salvador err = set_max_huge_pages(h, count, nid, n_mask); 377806808b08SLee Schermerhorn 37794eb0716eSAlexandre Ghiti return err ? err : len; 378006808b08SLee Schermerhorn } 378106808b08SLee Schermerhorn 3782238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 3783238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 3784238d3c13SDavid Rientjes size_t len) 3785238d3c13SDavid Rientjes { 3786238d3c13SDavid Rientjes struct hstate *h; 3787238d3c13SDavid Rientjes unsigned long count; 3788238d3c13SDavid Rientjes int nid; 3789238d3c13SDavid Rientjes int err; 3790238d3c13SDavid Rientjes 3791238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 3792238d3c13SDavid Rientjes if (err) 3793238d3c13SDavid Rientjes return err; 3794238d3c13SDavid Rientjes 3795238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 3796238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 3797238d3c13SDavid Rientjes } 3798238d3c13SDavid Rientjes 379906808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 380006808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 380106808b08SLee Schermerhorn { 380206808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 380306808b08SLee Schermerhorn } 380406808b08SLee Schermerhorn 380506808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 380606808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 380706808b08SLee Schermerhorn { 3808238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 3809a3437870SNishanth Aravamudan } 3810a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 3811a3437870SNishanth Aravamudan 381206808b08SLee Schermerhorn #ifdef CONFIG_NUMA 381306808b08SLee Schermerhorn 381406808b08SLee Schermerhorn /* 381506808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 381606808b08SLee Schermerhorn * huge page alloc/free. 381706808b08SLee Schermerhorn */ 381806808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 3819ae7a927dSJoe Perches struct kobj_attribute *attr, 3820ae7a927dSJoe Perches char *buf) 382106808b08SLee Schermerhorn { 382206808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 382306808b08SLee Schermerhorn } 382406808b08SLee Schermerhorn 382506808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 382606808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 382706808b08SLee Schermerhorn { 3828238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 382906808b08SLee Schermerhorn } 383006808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 383106808b08SLee Schermerhorn #endif 383206808b08SLee Schermerhorn 383306808b08SLee Schermerhorn 3834a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 3835a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3836a3437870SNishanth Aravamudan { 38379a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3838ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); 3839a3437870SNishanth Aravamudan } 3840adbe8726SEric B Munson 3841a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 3842a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 3843a3437870SNishanth Aravamudan { 3844a3437870SNishanth Aravamudan int err; 3845a3437870SNishanth Aravamudan unsigned long input; 38469a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3847a3437870SNishanth Aravamudan 3848bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3849adbe8726SEric B Munson return -EINVAL; 3850adbe8726SEric B Munson 38513dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 3852a3437870SNishanth Aravamudan if (err) 385373ae31e5SEric B Munson return err; 3854a3437870SNishanth Aravamudan 3855db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3856a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 3857db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3858a3437870SNishanth Aravamudan 3859a3437870SNishanth Aravamudan return count; 3860a3437870SNishanth Aravamudan } 3861a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 3862a3437870SNishanth Aravamudan 3863a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 3864a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3865a3437870SNishanth Aravamudan { 38669a305230SLee Schermerhorn struct hstate *h; 38679a305230SLee Schermerhorn unsigned long free_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 free_huge_pages = h->free_huge_pages; 38739a305230SLee Schermerhorn else 38749a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 38759a305230SLee Schermerhorn 3876ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", free_huge_pages); 3877a3437870SNishanth Aravamudan } 3878a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 3879a3437870SNishanth Aravamudan 3880a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 3881a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3882a3437870SNishanth Aravamudan { 38839a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3884ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); 3885a3437870SNishanth Aravamudan } 3886a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 3887a3437870SNishanth Aravamudan 3888a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 3889a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3890a3437870SNishanth Aravamudan { 38919a305230SLee Schermerhorn struct hstate *h; 38929a305230SLee Schermerhorn unsigned long surplus_huge_pages; 38939a305230SLee Schermerhorn int nid; 38949a305230SLee Schermerhorn 38959a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 38969a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 38979a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 38989a305230SLee Schermerhorn else 38999a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 39009a305230SLee Schermerhorn 3901ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", surplus_huge_pages); 3902a3437870SNishanth Aravamudan } 3903a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 3904a3437870SNishanth Aravamudan 390579dfc695SMike Kravetz static ssize_t demote_store(struct kobject *kobj, 390679dfc695SMike Kravetz struct kobj_attribute *attr, const char *buf, size_t len) 390779dfc695SMike Kravetz { 390879dfc695SMike Kravetz unsigned long nr_demote; 390979dfc695SMike Kravetz unsigned long nr_available; 391079dfc695SMike Kravetz nodemask_t nodes_allowed, *n_mask; 391179dfc695SMike Kravetz struct hstate *h; 39128eeda55fSLi zeming int err; 391379dfc695SMike Kravetz int nid; 391479dfc695SMike Kravetz 391579dfc695SMike Kravetz err = kstrtoul(buf, 10, &nr_demote); 391679dfc695SMike Kravetz if (err) 391779dfc695SMike Kravetz return err; 391879dfc695SMike Kravetz h = kobj_to_hstate(kobj, &nid); 391979dfc695SMike Kravetz 392079dfc695SMike Kravetz if (nid != NUMA_NO_NODE) { 392179dfc695SMike Kravetz init_nodemask_of_node(&nodes_allowed, nid); 392279dfc695SMike Kravetz n_mask = &nodes_allowed; 392379dfc695SMike Kravetz } else { 392479dfc695SMike Kravetz n_mask = &node_states[N_MEMORY]; 392579dfc695SMike Kravetz } 392679dfc695SMike Kravetz 392779dfc695SMike Kravetz /* Synchronize with other sysfs operations modifying huge pages */ 392879dfc695SMike Kravetz mutex_lock(&h->resize_lock); 392979dfc695SMike Kravetz spin_lock_irq(&hugetlb_lock); 393079dfc695SMike Kravetz 393179dfc695SMike Kravetz while (nr_demote) { 393279dfc695SMike Kravetz /* 393379dfc695SMike Kravetz * Check for available pages to demote each time thorough the 393479dfc695SMike Kravetz * loop as demote_pool_huge_page will drop hugetlb_lock. 393579dfc695SMike Kravetz */ 393679dfc695SMike Kravetz if (nid != NUMA_NO_NODE) 393779dfc695SMike Kravetz nr_available = h->free_huge_pages_node[nid]; 393879dfc695SMike Kravetz else 393979dfc695SMike Kravetz nr_available = h->free_huge_pages; 394079dfc695SMike Kravetz nr_available -= h->resv_huge_pages; 394179dfc695SMike Kravetz if (!nr_available) 394279dfc695SMike Kravetz break; 394379dfc695SMike Kravetz 394479dfc695SMike Kravetz err = demote_pool_huge_page(h, n_mask); 394579dfc695SMike Kravetz if (err) 394679dfc695SMike Kravetz break; 394779dfc695SMike Kravetz 394879dfc695SMike Kravetz nr_demote--; 394979dfc695SMike Kravetz } 395079dfc695SMike Kravetz 395179dfc695SMike Kravetz spin_unlock_irq(&hugetlb_lock); 395279dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 395379dfc695SMike Kravetz 395479dfc695SMike Kravetz if (err) 395579dfc695SMike Kravetz return err; 395679dfc695SMike Kravetz return len; 395779dfc695SMike Kravetz } 395879dfc695SMike Kravetz HSTATE_ATTR_WO(demote); 395979dfc695SMike Kravetz 396079dfc695SMike Kravetz static ssize_t demote_size_show(struct kobject *kobj, 396179dfc695SMike Kravetz struct kobj_attribute *attr, char *buf) 396279dfc695SMike Kravetz { 396312658abfSMiaohe Lin struct hstate *h = kobj_to_hstate(kobj, NULL); 396479dfc695SMike Kravetz unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K; 396579dfc695SMike Kravetz 396679dfc695SMike Kravetz return sysfs_emit(buf, "%lukB\n", demote_size); 396779dfc695SMike Kravetz } 396879dfc695SMike Kravetz 396979dfc695SMike Kravetz static ssize_t demote_size_store(struct kobject *kobj, 397079dfc695SMike Kravetz struct kobj_attribute *attr, 397179dfc695SMike Kravetz const char *buf, size_t count) 397279dfc695SMike Kravetz { 397379dfc695SMike Kravetz struct hstate *h, *demote_hstate; 397479dfc695SMike Kravetz unsigned long demote_size; 397579dfc695SMike Kravetz unsigned int demote_order; 397679dfc695SMike Kravetz 397779dfc695SMike Kravetz demote_size = (unsigned long)memparse(buf, NULL); 397879dfc695SMike Kravetz 397979dfc695SMike Kravetz demote_hstate = size_to_hstate(demote_size); 398079dfc695SMike Kravetz if (!demote_hstate) 398179dfc695SMike Kravetz return -EINVAL; 398279dfc695SMike Kravetz demote_order = demote_hstate->order; 3983a01f4390SMike Kravetz if (demote_order < HUGETLB_PAGE_ORDER) 3984a01f4390SMike Kravetz return -EINVAL; 398579dfc695SMike Kravetz 398679dfc695SMike Kravetz /* demote order must be smaller than hstate order */ 398712658abfSMiaohe Lin h = kobj_to_hstate(kobj, NULL); 398879dfc695SMike Kravetz if (demote_order >= h->order) 398979dfc695SMike Kravetz return -EINVAL; 399079dfc695SMike Kravetz 399179dfc695SMike Kravetz /* resize_lock synchronizes access to demote size and writes */ 399279dfc695SMike Kravetz mutex_lock(&h->resize_lock); 399379dfc695SMike Kravetz h->demote_order = demote_order; 399479dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 399579dfc695SMike Kravetz 399679dfc695SMike Kravetz return count; 399779dfc695SMike Kravetz } 399879dfc695SMike Kravetz HSTATE_ATTR(demote_size); 399979dfc695SMike Kravetz 4000a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 4001a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 4002a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 4003a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 4004a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 4005a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 400606808b08SLee Schermerhorn #ifdef CONFIG_NUMA 400706808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 400806808b08SLee Schermerhorn #endif 4009a3437870SNishanth Aravamudan NULL, 4010a3437870SNishanth Aravamudan }; 4011a3437870SNishanth Aravamudan 401267e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 4013a3437870SNishanth Aravamudan .attrs = hstate_attrs, 4014a3437870SNishanth Aravamudan }; 4015a3437870SNishanth Aravamudan 401679dfc695SMike Kravetz static struct attribute *hstate_demote_attrs[] = { 401779dfc695SMike Kravetz &demote_size_attr.attr, 401879dfc695SMike Kravetz &demote_attr.attr, 401979dfc695SMike Kravetz NULL, 402079dfc695SMike Kravetz }; 402179dfc695SMike Kravetz 402279dfc695SMike Kravetz static const struct attribute_group hstate_demote_attr_group = { 402379dfc695SMike Kravetz .attrs = hstate_demote_attrs, 402479dfc695SMike Kravetz }; 402579dfc695SMike Kravetz 4026094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 40279a305230SLee Schermerhorn struct kobject **hstate_kobjs, 402867e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 4029a3437870SNishanth Aravamudan { 4030a3437870SNishanth Aravamudan int retval; 4031972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 4032a3437870SNishanth Aravamudan 40339a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 40349a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 4035a3437870SNishanth Aravamudan return -ENOMEM; 4036a3437870SNishanth Aravamudan 40379a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 4038cc2205a6SMiaohe Lin if (retval) { 40399a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 4040cc2205a6SMiaohe Lin hstate_kobjs[hi] = NULL; 40413a6bdda0SMiaohe Lin return retval; 4042cc2205a6SMiaohe Lin } 4043a3437870SNishanth Aravamudan 404479dfc695SMike Kravetz if (h->demote_order) { 404501088a60SMiaohe Lin retval = sysfs_create_group(hstate_kobjs[hi], 404601088a60SMiaohe Lin &hstate_demote_attr_group); 404701088a60SMiaohe Lin if (retval) { 404879dfc695SMike Kravetz pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name); 404901088a60SMiaohe Lin sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group); 405001088a60SMiaohe Lin kobject_put(hstate_kobjs[hi]); 405101088a60SMiaohe Lin hstate_kobjs[hi] = NULL; 405201088a60SMiaohe Lin return retval; 405301088a60SMiaohe Lin } 405479dfc695SMike Kravetz } 405579dfc695SMike Kravetz 405601088a60SMiaohe Lin return 0; 4057a3437870SNishanth Aravamudan } 4058a3437870SNishanth Aravamudan 40599a305230SLee Schermerhorn #ifdef CONFIG_NUMA 4060a4a00b45SMuchun Song static bool hugetlb_sysfs_initialized __ro_after_init; 40619a305230SLee Schermerhorn 40629a305230SLee Schermerhorn /* 40639a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 406410fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 406510fbcf4cSKay Sievers * index of a node device or _hstate == node id. 406610fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 40679a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 40689a305230SLee Schermerhorn */ 40699a305230SLee Schermerhorn struct node_hstate { 40709a305230SLee Schermerhorn struct kobject *hugepages_kobj; 40719a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 40729a305230SLee Schermerhorn }; 4073b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 40749a305230SLee Schermerhorn 40759a305230SLee Schermerhorn /* 407610fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 40779a305230SLee Schermerhorn */ 40789a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 40799a305230SLee Schermerhorn &nr_hugepages_attr.attr, 40809a305230SLee Schermerhorn &free_hugepages_attr.attr, 40819a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 40829a305230SLee Schermerhorn NULL, 40839a305230SLee Schermerhorn }; 40849a305230SLee Schermerhorn 408567e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 40869a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 40879a305230SLee Schermerhorn }; 40889a305230SLee Schermerhorn 40899a305230SLee Schermerhorn /* 409010fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 40919a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 40929a305230SLee Schermerhorn */ 40939a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 40949a305230SLee Schermerhorn { 40959a305230SLee Schermerhorn int nid; 40969a305230SLee Schermerhorn 40979a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 40989a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 40999a305230SLee Schermerhorn int i; 41009a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 41019a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 41029a305230SLee Schermerhorn if (nidp) 41039a305230SLee Schermerhorn *nidp = nid; 41049a305230SLee Schermerhorn return &hstates[i]; 41059a305230SLee Schermerhorn } 41069a305230SLee Schermerhorn } 41079a305230SLee Schermerhorn 41089a305230SLee Schermerhorn BUG(); 41099a305230SLee Schermerhorn return NULL; 41109a305230SLee Schermerhorn } 41119a305230SLee Schermerhorn 41129a305230SLee Schermerhorn /* 411310fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 41149a305230SLee Schermerhorn * No-op if no hstate attributes attached. 41159a305230SLee Schermerhorn */ 4116a4a00b45SMuchun Song void hugetlb_unregister_node(struct node *node) 41179a305230SLee Schermerhorn { 41189a305230SLee Schermerhorn struct hstate *h; 411910fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 41209a305230SLee Schermerhorn 41219a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 41229b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 41239a305230SLee Schermerhorn 4124972dc4deSAneesh Kumar K.V for_each_hstate(h) { 4125972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 412601088a60SMiaohe Lin struct kobject *hstate_kobj = nhs->hstate_kobjs[idx]; 412701088a60SMiaohe Lin 412801088a60SMiaohe Lin if (!hstate_kobj) 412901088a60SMiaohe Lin continue; 413001088a60SMiaohe Lin if (h->demote_order) 413101088a60SMiaohe Lin sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group); 413201088a60SMiaohe Lin sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group); 413301088a60SMiaohe Lin kobject_put(hstate_kobj); 4134972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 4135972dc4deSAneesh Kumar K.V } 41369a305230SLee Schermerhorn 41379a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 41389a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 41399a305230SLee Schermerhorn } 41409a305230SLee Schermerhorn 41419a305230SLee Schermerhorn 41429a305230SLee Schermerhorn /* 414310fbcf4cSKay Sievers * Register hstate attributes for a single node device. 41449a305230SLee Schermerhorn * No-op if attributes already registered. 41459a305230SLee Schermerhorn */ 4146a4a00b45SMuchun Song void hugetlb_register_node(struct node *node) 41479a305230SLee Schermerhorn { 41489a305230SLee Schermerhorn struct hstate *h; 414910fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 41509a305230SLee Schermerhorn int err; 41519a305230SLee Schermerhorn 4152a4a00b45SMuchun Song if (!hugetlb_sysfs_initialized) 4153a4a00b45SMuchun Song return; 4154a4a00b45SMuchun Song 41559a305230SLee Schermerhorn if (nhs->hugepages_kobj) 41569a305230SLee Schermerhorn return; /* already allocated */ 41579a305230SLee Schermerhorn 41589a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 415910fbcf4cSKay Sievers &node->dev.kobj); 41609a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 41619a305230SLee Schermerhorn return; 41629a305230SLee Schermerhorn 41639a305230SLee Schermerhorn for_each_hstate(h) { 41649a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 41659a305230SLee Schermerhorn nhs->hstate_kobjs, 41669a305230SLee Schermerhorn &per_node_hstate_attr_group); 41679a305230SLee Schermerhorn if (err) { 4168282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s for node %d\n", 416910fbcf4cSKay Sievers h->name, node->dev.id); 41709a305230SLee Schermerhorn hugetlb_unregister_node(node); 41719a305230SLee Schermerhorn break; 41729a305230SLee Schermerhorn } 41739a305230SLee Schermerhorn } 41749a305230SLee Schermerhorn } 41759a305230SLee Schermerhorn 41769a305230SLee Schermerhorn /* 41779b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 417810fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 417910fbcf4cSKay Sievers * registered their associated device by this time. 41809a305230SLee Schermerhorn */ 41817d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 41829a305230SLee Schermerhorn { 41839a305230SLee Schermerhorn int nid; 41849a305230SLee Schermerhorn 4185a4a00b45SMuchun Song for_each_online_node(nid) 4186b958d4d0SMuchun Song hugetlb_register_node(node_devices[nid]); 41879a305230SLee Schermerhorn } 41889a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 41899a305230SLee Schermerhorn 41909a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 41919a305230SLee Schermerhorn { 41929a305230SLee Schermerhorn BUG(); 41939a305230SLee Schermerhorn if (nidp) 41949a305230SLee Schermerhorn *nidp = -1; 41959a305230SLee Schermerhorn return NULL; 41969a305230SLee Schermerhorn } 41979a305230SLee Schermerhorn 41989a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 41999a305230SLee Schermerhorn 42009a305230SLee Schermerhorn #endif 42019a305230SLee Schermerhorn 4202263b8998SMiaohe Lin #ifdef CONFIG_CMA 4203263b8998SMiaohe Lin static void __init hugetlb_cma_check(void); 4204263b8998SMiaohe Lin #else 4205263b8998SMiaohe Lin static inline __init void hugetlb_cma_check(void) 4206263b8998SMiaohe Lin { 4207263b8998SMiaohe Lin } 4208263b8998SMiaohe Lin #endif 4209263b8998SMiaohe Lin 4210a4a00b45SMuchun Song static void __init hugetlb_sysfs_init(void) 4211a4a00b45SMuchun Song { 4212a4a00b45SMuchun Song struct hstate *h; 4213a4a00b45SMuchun Song int err; 4214a4a00b45SMuchun Song 4215a4a00b45SMuchun Song hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 4216a4a00b45SMuchun Song if (!hugepages_kobj) 4217a4a00b45SMuchun Song return; 4218a4a00b45SMuchun Song 4219a4a00b45SMuchun Song for_each_hstate(h) { 4220a4a00b45SMuchun Song err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 4221a4a00b45SMuchun Song hstate_kobjs, &hstate_attr_group); 4222a4a00b45SMuchun Song if (err) 4223a4a00b45SMuchun Song pr_err("HugeTLB: Unable to add hstate %s", h->name); 4224a4a00b45SMuchun Song } 4225a4a00b45SMuchun Song 4226a4a00b45SMuchun Song #ifdef CONFIG_NUMA 4227a4a00b45SMuchun Song hugetlb_sysfs_initialized = true; 4228a4a00b45SMuchun Song #endif 4229a4a00b45SMuchun Song hugetlb_register_all_nodes(); 4230a4a00b45SMuchun Song } 4231a4a00b45SMuchun Song 4232962de548SKefeng Wang #ifdef CONFIG_SYSCTL 4233962de548SKefeng Wang static void hugetlb_sysctl_init(void); 4234962de548SKefeng Wang #else 4235962de548SKefeng Wang static inline void hugetlb_sysctl_init(void) { } 4236962de548SKefeng Wang #endif 4237962de548SKefeng Wang 4238a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 4239a3437870SNishanth Aravamudan { 42408382d914SDavidlohr Bueso int i; 42418382d914SDavidlohr Bueso 4242d6995da3SMike Kravetz BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < 4243d6995da3SMike Kravetz __NR_HPAGEFLAGS); 4244d6995da3SMike Kravetz 4245c2833a5bSMike Kravetz if (!hugepages_supported()) { 4246c2833a5bSMike Kravetz if (hugetlb_max_hstate || default_hstate_max_huge_pages) 4247c2833a5bSMike Kravetz pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); 42480ef89d25SBenjamin Herrenschmidt return 0; 4249d715cf80SLiam R. Howlett } 4250d715cf80SLiam R. Howlett 4251282f4214SMike Kravetz /* 4252282f4214SMike Kravetz * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some 4253282f4214SMike Kravetz * architectures depend on setup being done here. 4254282f4214SMike Kravetz */ 4255a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 4256282f4214SMike Kravetz if (!parsed_default_hugepagesz) { 4257282f4214SMike Kravetz /* 4258282f4214SMike Kravetz * If we did not parse a default huge page size, set 4259282f4214SMike Kravetz * default_hstate_idx to HPAGE_SIZE hstate. And, if the 4260282f4214SMike Kravetz * number of huge pages for this default size was implicitly 4261282f4214SMike Kravetz * specified, set that here as well. 4262282f4214SMike Kravetz * Note that the implicit setting will overwrite an explicit 4263282f4214SMike Kravetz * setting. A warning will be printed in this case. 4264282f4214SMike Kravetz */ 4265282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); 4266f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 4267282f4214SMike Kravetz if (default_hstate.max_huge_pages) { 4268282f4214SMike Kravetz char buf[32]; 4269282f4214SMike Kravetz 4270282f4214SMike Kravetz string_get_size(huge_page_size(&default_hstate), 4271282f4214SMike Kravetz 1, STRING_UNITS_2, buf, 32); 4272282f4214SMike Kravetz pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", 4273282f4214SMike Kravetz default_hstate.max_huge_pages, buf); 4274282f4214SMike Kravetz pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", 4275282f4214SMike Kravetz default_hstate_max_huge_pages); 4276282f4214SMike Kravetz } 4277282f4214SMike Kravetz default_hstate.max_huge_pages = 4278282f4214SMike Kravetz default_hstate_max_huge_pages; 4279b5389086SZhenguo Yao 42800a7a0f6fSPeng Liu for_each_online_node(i) 4281b5389086SZhenguo Yao default_hstate.max_huge_pages_node[i] = 4282b5389086SZhenguo Yao default_hugepages_in_node[i]; 4283282f4214SMike Kravetz } 4284f8b74815SVaishali Thakkar } 4285a3437870SNishanth Aravamudan 4286cf11e85fSRoman Gushchin hugetlb_cma_check(); 4287a3437870SNishanth Aravamudan hugetlb_init_hstates(); 4288aa888a74SAndi Kleen gather_bootmem_prealloc(); 4289a3437870SNishanth Aravamudan report_hugepages(); 4290a3437870SNishanth Aravamudan 4291a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 42927179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 4293962de548SKefeng Wang hugetlb_sysctl_init(); 42949a305230SLee Schermerhorn 42958382d914SDavidlohr Bueso #ifdef CONFIG_SMP 42968382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 42978382d914SDavidlohr Bueso #else 42988382d914SDavidlohr Bueso num_fault_mutexes = 1; 42998382d914SDavidlohr Bueso #endif 4300c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 43016da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 43026da2ec56SKees Cook GFP_KERNEL); 4303c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 43048382d914SDavidlohr Bueso 43058382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 4306c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 4307a3437870SNishanth Aravamudan return 0; 4308a3437870SNishanth Aravamudan } 43093e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 4310a3437870SNishanth Aravamudan 4311ae94da89SMike Kravetz /* Overwritten by architectures with more huge page sizes */ 4312ae94da89SMike Kravetz bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) 43139fee021dSVaishali Thakkar { 4314ae94da89SMike Kravetz return size == HPAGE_SIZE; 43159fee021dSVaishali Thakkar } 43169fee021dSVaishali Thakkar 4317d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 4318a3437870SNishanth Aravamudan { 4319a3437870SNishanth Aravamudan struct hstate *h; 43208faa8b07SAndi Kleen unsigned long i; 43218faa8b07SAndi Kleen 4322a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 4323a3437870SNishanth Aravamudan return; 4324a3437870SNishanth Aravamudan } 432547d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 4326a3437870SNishanth Aravamudan BUG_ON(order == 0); 432747d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 432829383967SMike Kravetz mutex_init(&h->resize_lock); 4329a3437870SNishanth Aravamudan h->order = order; 4330aca78307SMiaohe Lin h->mask = ~(huge_page_size(h) - 1); 43318faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 43328faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 43330edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 433454f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 433554f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 4336a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 4337c2c3a60aSMiaohe Lin huge_page_size(h)/SZ_1K); 43388faa8b07SAndi Kleen 4339a3437870SNishanth Aravamudan parsed_hstate = h; 4340a3437870SNishanth Aravamudan } 4341a3437870SNishanth Aravamudan 4342b5389086SZhenguo Yao bool __init __weak hugetlb_node_alloc_supported(void) 4343b5389086SZhenguo Yao { 4344b5389086SZhenguo Yao return true; 4345b5389086SZhenguo Yao } 4346f87442f4SPeng Liu 4347f87442f4SPeng Liu static void __init hugepages_clear_pages_in_node(void) 4348f87442f4SPeng Liu { 4349f87442f4SPeng Liu if (!hugetlb_max_hstate) { 4350f87442f4SPeng Liu default_hstate_max_huge_pages = 0; 4351f87442f4SPeng Liu memset(default_hugepages_in_node, 0, 435210395680SMiaohe Lin sizeof(default_hugepages_in_node)); 4353f87442f4SPeng Liu } else { 4354f87442f4SPeng Liu parsed_hstate->max_huge_pages = 0; 4355f87442f4SPeng Liu memset(parsed_hstate->max_huge_pages_node, 0, 435610395680SMiaohe Lin sizeof(parsed_hstate->max_huge_pages_node)); 4357f87442f4SPeng Liu } 4358f87442f4SPeng Liu } 4359f87442f4SPeng Liu 4360282f4214SMike Kravetz /* 4361282f4214SMike Kravetz * hugepages command line processing 4362282f4214SMike Kravetz * hugepages normally follows a valid hugepagsz or default_hugepagsz 4363282f4214SMike Kravetz * specification. If not, ignore the hugepages value. hugepages can also 4364282f4214SMike Kravetz * be the first huge page command line option in which case it implicitly 4365282f4214SMike Kravetz * specifies the number of huge pages for the default size. 4366282f4214SMike Kravetz */ 4367282f4214SMike Kravetz static int __init hugepages_setup(char *s) 4368a3437870SNishanth Aravamudan { 4369a3437870SNishanth Aravamudan unsigned long *mhp; 43708faa8b07SAndi Kleen static unsigned long *last_mhp; 4371b5389086SZhenguo Yao int node = NUMA_NO_NODE; 4372b5389086SZhenguo Yao int count; 4373b5389086SZhenguo Yao unsigned long tmp; 4374b5389086SZhenguo Yao char *p = s; 4375a3437870SNishanth Aravamudan 43769fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 4377282f4214SMike Kravetz pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); 43789fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 4379f81f6e4bSPeng Liu return 1; 43809fee021dSVaishali Thakkar } 4381282f4214SMike Kravetz 4382a3437870SNishanth Aravamudan /* 4383282f4214SMike Kravetz * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter 4384282f4214SMike Kravetz * yet, so this hugepages= parameter goes to the "default hstate". 4385282f4214SMike Kravetz * Otherwise, it goes with the previously parsed hugepagesz or 4386282f4214SMike Kravetz * default_hugepagesz. 4387a3437870SNishanth Aravamudan */ 43889fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 4389a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 4390a3437870SNishanth Aravamudan else 4391a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 4392a3437870SNishanth Aravamudan 43938faa8b07SAndi Kleen if (mhp == last_mhp) { 4394282f4214SMike Kravetz pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); 4395f81f6e4bSPeng Liu return 1; 43968faa8b07SAndi Kleen } 43978faa8b07SAndi Kleen 4398b5389086SZhenguo Yao while (*p) { 4399b5389086SZhenguo Yao count = 0; 4400b5389086SZhenguo Yao if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4401b5389086SZhenguo Yao goto invalid; 4402b5389086SZhenguo Yao /* Parameter is node format */ 4403b5389086SZhenguo Yao if (p[count] == ':') { 4404b5389086SZhenguo Yao if (!hugetlb_node_alloc_supported()) { 4405b5389086SZhenguo Yao pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n"); 4406f81f6e4bSPeng Liu return 1; 4407b5389086SZhenguo Yao } 44080a7a0f6fSPeng Liu if (tmp >= MAX_NUMNODES || !node_online(tmp)) 4409e79ce983SLiu Yuntao goto invalid; 44100a7a0f6fSPeng Liu node = array_index_nospec(tmp, MAX_NUMNODES); 4411b5389086SZhenguo Yao p += count + 1; 4412b5389086SZhenguo Yao /* Parse hugepages */ 4413b5389086SZhenguo Yao if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4414b5389086SZhenguo Yao goto invalid; 4415b5389086SZhenguo Yao if (!hugetlb_max_hstate) 4416b5389086SZhenguo Yao default_hugepages_in_node[node] = tmp; 4417b5389086SZhenguo Yao else 4418b5389086SZhenguo Yao parsed_hstate->max_huge_pages_node[node] = tmp; 4419b5389086SZhenguo Yao *mhp += tmp; 4420b5389086SZhenguo Yao /* Go to parse next node*/ 4421b5389086SZhenguo Yao if (p[count] == ',') 4422b5389086SZhenguo Yao p += count + 1; 4423b5389086SZhenguo Yao else 4424b5389086SZhenguo Yao break; 4425b5389086SZhenguo Yao } else { 4426b5389086SZhenguo Yao if (p != s) 4427b5389086SZhenguo Yao goto invalid; 4428b5389086SZhenguo Yao *mhp = tmp; 4429b5389086SZhenguo Yao break; 4430b5389086SZhenguo Yao } 4431b5389086SZhenguo Yao } 4432a3437870SNishanth Aravamudan 44338faa8b07SAndi Kleen /* 44348faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 443504adbc3fSMiaohe Lin * But we need to allocate gigantic hstates here early to still 44368faa8b07SAndi Kleen * use the bootmem allocator. 44378faa8b07SAndi Kleen */ 443804adbc3fSMiaohe Lin if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) 44398faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 44408faa8b07SAndi Kleen 44418faa8b07SAndi Kleen last_mhp = mhp; 44428faa8b07SAndi Kleen 4443a3437870SNishanth Aravamudan return 1; 4444b5389086SZhenguo Yao 4445b5389086SZhenguo Yao invalid: 4446b5389086SZhenguo Yao pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p); 4447f87442f4SPeng Liu hugepages_clear_pages_in_node(); 4448f81f6e4bSPeng Liu return 1; 4449a3437870SNishanth Aravamudan } 4450282f4214SMike Kravetz __setup("hugepages=", hugepages_setup); 4451e11bfbfcSNick Piggin 4452282f4214SMike Kravetz /* 4453282f4214SMike Kravetz * hugepagesz command line processing 4454282f4214SMike Kravetz * A specific huge page size can only be specified once with hugepagesz. 4455282f4214SMike Kravetz * hugepagesz is followed by hugepages on the command line. The global 4456282f4214SMike Kravetz * variable 'parsed_valid_hugepagesz' is used to determine if prior 4457282f4214SMike Kravetz * hugepagesz argument was valid. 4458282f4214SMike Kravetz */ 4459359f2544SMike Kravetz static int __init hugepagesz_setup(char *s) 4460e11bfbfcSNick Piggin { 4461359f2544SMike Kravetz unsigned long size; 4462282f4214SMike Kravetz struct hstate *h; 4463282f4214SMike Kravetz 4464282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4465359f2544SMike Kravetz size = (unsigned long)memparse(s, NULL); 4466359f2544SMike Kravetz 4467359f2544SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4468282f4214SMike Kravetz pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); 4469f81f6e4bSPeng Liu return 1; 4470359f2544SMike Kravetz } 4471359f2544SMike Kravetz 4472282f4214SMike Kravetz h = size_to_hstate(size); 4473282f4214SMike Kravetz if (h) { 4474282f4214SMike Kravetz /* 4475282f4214SMike Kravetz * hstate for this size already exists. This is normally 4476282f4214SMike Kravetz * an error, but is allowed if the existing hstate is the 4477282f4214SMike Kravetz * default hstate. More specifically, it is only allowed if 4478282f4214SMike Kravetz * the number of huge pages for the default hstate was not 4479282f4214SMike Kravetz * previously specified. 4480282f4214SMike Kravetz */ 4481282f4214SMike Kravetz if (!parsed_default_hugepagesz || h != &default_hstate || 4482282f4214SMike Kravetz default_hstate.max_huge_pages) { 4483282f4214SMike Kravetz pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); 4484f81f6e4bSPeng Liu return 1; 448538237830SMike Kravetz } 448638237830SMike Kravetz 4487282f4214SMike Kravetz /* 4488282f4214SMike Kravetz * No need to call hugetlb_add_hstate() as hstate already 4489282f4214SMike Kravetz * exists. But, do set parsed_hstate so that a following 4490282f4214SMike Kravetz * hugepages= parameter will be applied to this hstate. 4491282f4214SMike Kravetz */ 4492282f4214SMike Kravetz parsed_hstate = h; 4493282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4494e11bfbfcSNick Piggin return 1; 4495e11bfbfcSNick Piggin } 4496282f4214SMike Kravetz 4497359f2544SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4498282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4499359f2544SMike Kravetz return 1; 4500359f2544SMike Kravetz } 4501359f2544SMike Kravetz __setup("hugepagesz=", hugepagesz_setup); 4502359f2544SMike Kravetz 4503282f4214SMike Kravetz /* 4504282f4214SMike Kravetz * default_hugepagesz command line input 4505282f4214SMike Kravetz * Only one instance of default_hugepagesz allowed on command line. 4506282f4214SMike Kravetz */ 4507ae94da89SMike Kravetz static int __init default_hugepagesz_setup(char *s) 4508e11bfbfcSNick Piggin { 4509ae94da89SMike Kravetz unsigned long size; 4510b5389086SZhenguo Yao int i; 4511ae94da89SMike Kravetz 4512282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4513282f4214SMike Kravetz if (parsed_default_hugepagesz) { 4514282f4214SMike Kravetz pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); 4515f81f6e4bSPeng Liu return 1; 4516282f4214SMike Kravetz } 4517282f4214SMike Kravetz 4518282f4214SMike Kravetz size = (unsigned long)memparse(s, NULL); 4519282f4214SMike Kravetz 4520282f4214SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4521282f4214SMike Kravetz pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); 4522f81f6e4bSPeng Liu return 1; 4523282f4214SMike Kravetz } 4524282f4214SMike Kravetz 4525282f4214SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4526282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4527282f4214SMike Kravetz parsed_default_hugepagesz = true; 4528282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(size)); 4529282f4214SMike Kravetz 4530282f4214SMike Kravetz /* 4531282f4214SMike Kravetz * The number of default huge pages (for this size) could have been 4532282f4214SMike Kravetz * specified as the first hugetlb parameter: hugepages=X. If so, 4533282f4214SMike Kravetz * then default_hstate_max_huge_pages is set. If the default huge 453423baf831SKirill A. Shutemov * page size is gigantic (> MAX_ORDER), then the pages must be 4535282f4214SMike Kravetz * allocated here from bootmem allocator. 4536282f4214SMike Kravetz */ 4537282f4214SMike Kravetz if (default_hstate_max_huge_pages) { 4538282f4214SMike Kravetz default_hstate.max_huge_pages = default_hstate_max_huge_pages; 45390a7a0f6fSPeng Liu for_each_online_node(i) 4540b5389086SZhenguo Yao default_hstate.max_huge_pages_node[i] = 4541b5389086SZhenguo Yao default_hugepages_in_node[i]; 4542282f4214SMike Kravetz if (hstate_is_gigantic(&default_hstate)) 4543282f4214SMike Kravetz hugetlb_hstate_alloc_pages(&default_hstate); 4544282f4214SMike Kravetz default_hstate_max_huge_pages = 0; 4545282f4214SMike Kravetz } 4546282f4214SMike Kravetz 4547e11bfbfcSNick Piggin return 1; 4548e11bfbfcSNick Piggin } 4549ae94da89SMike Kravetz __setup("default_hugepagesz=", default_hugepagesz_setup); 4550a3437870SNishanth Aravamudan 4551d2226ebdSFeng Tang static nodemask_t *policy_mbind_nodemask(gfp_t gfp) 4552d2226ebdSFeng Tang { 4553d2226ebdSFeng Tang #ifdef CONFIG_NUMA 4554d2226ebdSFeng Tang struct mempolicy *mpol = get_task_policy(current); 4555d2226ebdSFeng Tang 4556d2226ebdSFeng Tang /* 4557d2226ebdSFeng Tang * Only enforce MPOL_BIND policy which overlaps with cpuset policy 4558d2226ebdSFeng Tang * (from policy_nodemask) specifically for hugetlb case 4559d2226ebdSFeng Tang */ 4560d2226ebdSFeng Tang if (mpol->mode == MPOL_BIND && 4561d2226ebdSFeng Tang (apply_policy_zone(mpol, gfp_zone(gfp)) && 4562d2226ebdSFeng Tang cpuset_nodemask_valid_mems_allowed(&mpol->nodes))) 4563d2226ebdSFeng Tang return &mpol->nodes; 4564d2226ebdSFeng Tang #endif 4565d2226ebdSFeng Tang return NULL; 4566d2226ebdSFeng Tang } 4567d2226ebdSFeng Tang 45688ca39e68SMuchun Song static unsigned int allowed_mems_nr(struct hstate *h) 45698a213460SNishanth Aravamudan { 45708a213460SNishanth Aravamudan int node; 45718a213460SNishanth Aravamudan unsigned int nr = 0; 4572d2226ebdSFeng Tang nodemask_t *mbind_nodemask; 45738ca39e68SMuchun Song unsigned int *array = h->free_huge_pages_node; 45748ca39e68SMuchun Song gfp_t gfp_mask = htlb_alloc_mask(h); 45758a213460SNishanth Aravamudan 4576d2226ebdSFeng Tang mbind_nodemask = policy_mbind_nodemask(gfp_mask); 45778ca39e68SMuchun Song for_each_node_mask(node, cpuset_current_mems_allowed) { 4578d2226ebdSFeng Tang if (!mbind_nodemask || node_isset(node, *mbind_nodemask)) 45798a213460SNishanth Aravamudan nr += array[node]; 45808ca39e68SMuchun Song } 45818a213460SNishanth Aravamudan 45828a213460SNishanth Aravamudan return nr; 45838a213460SNishanth Aravamudan } 45848a213460SNishanth Aravamudan 45858a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 458617743798SMuchun Song static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, 458717743798SMuchun Song void *buffer, size_t *length, 458817743798SMuchun Song loff_t *ppos, unsigned long *out) 458917743798SMuchun Song { 459017743798SMuchun Song struct ctl_table dup_table; 459117743798SMuchun Song 459217743798SMuchun Song /* 459317743798SMuchun Song * In order to avoid races with __do_proc_doulongvec_minmax(), we 459417743798SMuchun Song * can duplicate the @table and alter the duplicate of it. 459517743798SMuchun Song */ 459617743798SMuchun Song dup_table = *table; 459717743798SMuchun Song dup_table.data = out; 459817743798SMuchun Song 459917743798SMuchun Song return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); 460017743798SMuchun Song } 460117743798SMuchun Song 460206808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 460306808b08SLee Schermerhorn struct ctl_table *table, int write, 460432927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 46051da177e4SLinus Torvalds { 4606e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 4607238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 460808d4a246SMichal Hocko int ret; 4609e5ff2159SAndi Kleen 4610457c1b27SNishanth Aravamudan if (!hugepages_supported()) 461186613628SJan Stancek return -EOPNOTSUPP; 4612457c1b27SNishanth Aravamudan 461317743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 461417743798SMuchun Song &tmp); 461508d4a246SMichal Hocko if (ret) 461608d4a246SMichal Hocko goto out; 4617e5ff2159SAndi Kleen 4618238d3c13SDavid Rientjes if (write) 4619238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 4620238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 462108d4a246SMichal Hocko out: 462208d4a246SMichal Hocko return ret; 46231da177e4SLinus Torvalds } 4624396faf03SMel Gorman 4625962de548SKefeng Wang static int hugetlb_sysctl_handler(struct ctl_table *table, int write, 462632927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 462706808b08SLee Schermerhorn { 462806808b08SLee Schermerhorn 462906808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 463006808b08SLee Schermerhorn buffer, length, ppos); 463106808b08SLee Schermerhorn } 463206808b08SLee Schermerhorn 463306808b08SLee Schermerhorn #ifdef CONFIG_NUMA 4634962de548SKefeng Wang static int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 463532927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 463606808b08SLee Schermerhorn { 463706808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 463806808b08SLee Schermerhorn buffer, length, ppos); 463906808b08SLee Schermerhorn } 464006808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 464106808b08SLee Schermerhorn 4642962de548SKefeng Wang static int hugetlb_overcommit_handler(struct ctl_table *table, int write, 464332927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 4644a3d0c6aaSNishanth Aravamudan { 4645a5516438SAndi Kleen struct hstate *h = &default_hstate; 4646e5ff2159SAndi Kleen unsigned long tmp; 464708d4a246SMichal Hocko int ret; 4648e5ff2159SAndi Kleen 4649457c1b27SNishanth Aravamudan if (!hugepages_supported()) 465086613628SJan Stancek return -EOPNOTSUPP; 4651457c1b27SNishanth Aravamudan 4652e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 4653e5ff2159SAndi Kleen 4654bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 4655adbe8726SEric B Munson return -EINVAL; 4656adbe8726SEric B Munson 465717743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 465817743798SMuchun Song &tmp); 465908d4a246SMichal Hocko if (ret) 466008d4a246SMichal Hocko goto out; 4661e5ff2159SAndi Kleen 4662e5ff2159SAndi Kleen if (write) { 4663db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4664e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 4665db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4666e5ff2159SAndi Kleen } 466708d4a246SMichal Hocko out: 466808d4a246SMichal Hocko return ret; 4669a3d0c6aaSNishanth Aravamudan } 4670a3d0c6aaSNishanth Aravamudan 4671962de548SKefeng Wang static struct ctl_table hugetlb_table[] = { 4672962de548SKefeng Wang { 4673962de548SKefeng Wang .procname = "nr_hugepages", 4674962de548SKefeng Wang .data = NULL, 4675962de548SKefeng Wang .maxlen = sizeof(unsigned long), 4676962de548SKefeng Wang .mode = 0644, 4677962de548SKefeng Wang .proc_handler = hugetlb_sysctl_handler, 4678962de548SKefeng Wang }, 4679962de548SKefeng Wang #ifdef CONFIG_NUMA 4680962de548SKefeng Wang { 4681962de548SKefeng Wang .procname = "nr_hugepages_mempolicy", 4682962de548SKefeng Wang .data = NULL, 4683962de548SKefeng Wang .maxlen = sizeof(unsigned long), 4684962de548SKefeng Wang .mode = 0644, 4685962de548SKefeng Wang .proc_handler = &hugetlb_mempolicy_sysctl_handler, 4686962de548SKefeng Wang }, 4687962de548SKefeng Wang #endif 4688962de548SKefeng Wang { 4689962de548SKefeng Wang .procname = "hugetlb_shm_group", 4690962de548SKefeng Wang .data = &sysctl_hugetlb_shm_group, 4691962de548SKefeng Wang .maxlen = sizeof(gid_t), 4692962de548SKefeng Wang .mode = 0644, 4693962de548SKefeng Wang .proc_handler = proc_dointvec, 4694962de548SKefeng Wang }, 4695962de548SKefeng Wang { 4696962de548SKefeng Wang .procname = "nr_overcommit_hugepages", 4697962de548SKefeng Wang .data = NULL, 4698962de548SKefeng Wang .maxlen = sizeof(unsigned long), 4699962de548SKefeng Wang .mode = 0644, 4700962de548SKefeng Wang .proc_handler = hugetlb_overcommit_handler, 4701962de548SKefeng Wang }, 4702962de548SKefeng Wang { } 4703962de548SKefeng Wang }; 4704962de548SKefeng Wang 4705962de548SKefeng Wang static void hugetlb_sysctl_init(void) 4706962de548SKefeng Wang { 4707962de548SKefeng Wang register_sysctl_init("vm", hugetlb_table); 4708962de548SKefeng Wang } 47091da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 47101da177e4SLinus Torvalds 4711e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 47121da177e4SLinus Torvalds { 4713fcb2b0c5SRoman Gushchin struct hstate *h; 4714fcb2b0c5SRoman Gushchin unsigned long total = 0; 4715fcb2b0c5SRoman Gushchin 4716457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4717457c1b27SNishanth Aravamudan return; 4718fcb2b0c5SRoman Gushchin 4719fcb2b0c5SRoman Gushchin for_each_hstate(h) { 4720fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 4721fcb2b0c5SRoman Gushchin 4722aca78307SMiaohe Lin total += huge_page_size(h) * count; 4723fcb2b0c5SRoman Gushchin 4724fcb2b0c5SRoman Gushchin if (h == &default_hstate) 4725e1759c21SAlexey Dobriyan seq_printf(m, 47261da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 47271da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 4728b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 47297893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 47304f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 4731fcb2b0c5SRoman Gushchin count, 4732a5516438SAndi Kleen h->free_huge_pages, 4733a5516438SAndi Kleen h->resv_huge_pages, 4734a5516438SAndi Kleen h->surplus_huge_pages, 4735aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4736fcb2b0c5SRoman Gushchin } 4737fcb2b0c5SRoman Gushchin 4738aca78307SMiaohe Lin seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); 47391da177e4SLinus Torvalds } 47401da177e4SLinus Torvalds 47417981593bSJoe Perches int hugetlb_report_node_meminfo(char *buf, int len, int nid) 47421da177e4SLinus Torvalds { 4743a5516438SAndi Kleen struct hstate *h = &default_hstate; 47447981593bSJoe Perches 4745457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4746457c1b27SNishanth Aravamudan return 0; 47477981593bSJoe Perches 47487981593bSJoe Perches return sysfs_emit_at(buf, len, 47491da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 4750a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 4751a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 4752a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 4753a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 4754a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 47551da177e4SLinus Torvalds } 47561da177e4SLinus Torvalds 4757dcadcf1cSGang Li void hugetlb_show_meminfo_node(int nid) 4758949f7ec5SDavid Rientjes { 4759949f7ec5SDavid Rientjes struct hstate *h; 4760949f7ec5SDavid Rientjes 4761457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4762457c1b27SNishanth Aravamudan return; 4763457c1b27SNishanth Aravamudan 4764949f7ec5SDavid Rientjes for_each_hstate(h) 4765dcadcf1cSGang Li printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 4766949f7ec5SDavid Rientjes nid, 4767949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 4768949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 4769949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 4770aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4771949f7ec5SDavid Rientjes } 4772949f7ec5SDavid Rientjes 47735d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 47745d317b2bSNaoya Horiguchi { 47755d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 47766c1aa2d3SZhangPeng K(atomic_long_read(&mm->hugetlb_usage))); 47775d317b2bSNaoya Horiguchi } 47785d317b2bSNaoya Horiguchi 47791da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 47801da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 47811da177e4SLinus Torvalds { 4782d0028588SWanpeng Li struct hstate *h; 4783d0028588SWanpeng Li unsigned long nr_total_pages = 0; 4784d0028588SWanpeng Li 4785d0028588SWanpeng Li for_each_hstate(h) 4786d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 4787d0028588SWanpeng Li return nr_total_pages; 47881da177e4SLinus Torvalds } 47891da177e4SLinus Torvalds 4790a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 4791fc1b8a73SMel Gorman { 4792fc1b8a73SMel Gorman int ret = -ENOMEM; 4793fc1b8a73SMel Gorman 47940aa7f354SMiaohe Lin if (!delta) 47950aa7f354SMiaohe Lin return 0; 47960aa7f354SMiaohe Lin 4797db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4798fc1b8a73SMel Gorman /* 4799fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 4800fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 4801fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 4802fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 4803fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 4804fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 4805fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 4806fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 4807fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 4808fc1b8a73SMel Gorman * 4809fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 4810fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 4811fc1b8a73SMel Gorman * we fall back to check against current free page availability as 4812fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 4813fc1b8a73SMel Gorman * semantics that cpuset has. 48148ca39e68SMuchun Song * 48158ca39e68SMuchun Song * Apart from cpuset, we also have memory policy mechanism that 48168ca39e68SMuchun Song * also determines from which node the kernel will allocate memory 48178ca39e68SMuchun Song * in a NUMA system. So similar to cpuset, we also should consider 48188ca39e68SMuchun Song * the memory policy of the current task. Similar to the description 48198ca39e68SMuchun Song * above. 4820fc1b8a73SMel Gorman */ 4821fc1b8a73SMel Gorman if (delta > 0) { 4822a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 4823fc1b8a73SMel Gorman goto out; 4824fc1b8a73SMel Gorman 48258ca39e68SMuchun Song if (delta > allowed_mems_nr(h)) { 4826a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 4827fc1b8a73SMel Gorman goto out; 4828fc1b8a73SMel Gorman } 4829fc1b8a73SMel Gorman } 4830fc1b8a73SMel Gorman 4831fc1b8a73SMel Gorman ret = 0; 4832fc1b8a73SMel Gorman if (delta < 0) 4833a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 4834fc1b8a73SMel Gorman 4835fc1b8a73SMel Gorman out: 4836db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4837fc1b8a73SMel Gorman return ret; 4838fc1b8a73SMel Gorman } 4839fc1b8a73SMel Gorman 484084afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 484184afd99bSAndy Whitcroft { 4842f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 484384afd99bSAndy Whitcroft 484484afd99bSAndy Whitcroft /* 4845612b8a31SMike Kravetz * HPAGE_RESV_OWNER indicates a private mapping. 484684afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 484784afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 484884afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 484925985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 485084afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 485184afd99bSAndy Whitcroft * new reference here without additional locking. 485284afd99bSAndy Whitcroft */ 485309a26e83SMike Kravetz if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 485409a26e83SMike Kravetz resv_map_dup_hugetlb_cgroup_uncharge_info(resv); 4855f522c3acSJoonsoo Kim kref_get(&resv->refs); 485684afd99bSAndy Whitcroft } 48578d9bfb26SMike Kravetz 4858131a79b4SMike Kravetz /* 4859131a79b4SMike Kravetz * vma_lock structure for sharable mappings is vma specific. 4860612b8a31SMike Kravetz * Clear old pointer (if copied via vm_area_dup) and allocate 4861612b8a31SMike Kravetz * new structure. Before clearing, make sure vma_lock is not 4862612b8a31SMike Kravetz * for this vma. 4863131a79b4SMike Kravetz */ 4864131a79b4SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 4865612b8a31SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 4866612b8a31SMike Kravetz 4867612b8a31SMike Kravetz if (vma_lock) { 4868612b8a31SMike Kravetz if (vma_lock->vma != vma) { 4869131a79b4SMike Kravetz vma->vm_private_data = NULL; 48708d9bfb26SMike Kravetz hugetlb_vma_lock_alloc(vma); 4871612b8a31SMike Kravetz } else 4872612b8a31SMike Kravetz pr_warn("HugeTLB: vma_lock already exists in %s.\n", __func__); 4873612b8a31SMike Kravetz } else 4874612b8a31SMike Kravetz hugetlb_vma_lock_alloc(vma); 487509a26e83SMike Kravetz } 4876131a79b4SMike Kravetz } 487784afd99bSAndy Whitcroft 4878a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 4879a1e78772SMel Gorman { 4880a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 48818d9bfb26SMike Kravetz struct resv_map *resv; 488290481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 48834e35f483SJoonsoo Kim unsigned long reserve, start, end; 48841c5ecae3SMike Kravetz long gbl_reserve; 488584afd99bSAndy Whitcroft 48868d9bfb26SMike Kravetz hugetlb_vma_lock_free(vma); 48878d9bfb26SMike Kravetz 48888d9bfb26SMike Kravetz resv = vma_resv_map(vma); 48894e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 48904e35f483SJoonsoo Kim return; 48914e35f483SJoonsoo Kim 4892a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 4893a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 489484afd99bSAndy Whitcroft 48954e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 4896e9fe92aeSMina Almasry hugetlb_cgroup_uncharge_counter(resv, start, end); 48977251ff78SAdam Litke if (reserve) { 48981c5ecae3SMike Kravetz /* 48991c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 49001c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 49011c5ecae3SMike Kravetz */ 49021c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 49031c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 49047251ff78SAdam Litke } 4905e9fe92aeSMina Almasry 4906e9fe92aeSMina Almasry kref_put(&resv->refs, resv_map_release); 4907a1e78772SMel Gorman } 4908a1e78772SMel Gorman 490931383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 491031383c68SDan Williams { 491131383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 491231383c68SDan Williams return -EINVAL; 4913b30c14cdSJames Houghton 4914b30c14cdSJames Houghton /* 4915b30c14cdSJames Houghton * PMD sharing is only possible for PUD_SIZE-aligned address ranges 4916b30c14cdSJames Houghton * in HugeTLB VMAs. If we will lose PUD_SIZE alignment due to this 4917b30c14cdSJames Houghton * split, unshare PMDs in the PUD_SIZE interval surrounding addr now. 4918b30c14cdSJames Houghton */ 4919b30c14cdSJames Houghton if (addr & ~PUD_MASK) { 4920b30c14cdSJames Houghton /* 4921b30c14cdSJames Houghton * hugetlb_vm_op_split is called right before we attempt to 4922b30c14cdSJames Houghton * split the VMA. We will need to unshare PMDs in the old and 4923b30c14cdSJames Houghton * new VMAs, so let's unshare before we split. 4924b30c14cdSJames Houghton */ 4925b30c14cdSJames Houghton unsigned long floor = addr & PUD_MASK; 4926b30c14cdSJames Houghton unsigned long ceil = floor + PUD_SIZE; 4927b30c14cdSJames Houghton 4928b30c14cdSJames Houghton if (floor >= vma->vm_start && ceil <= vma->vm_end) 4929b30c14cdSJames Houghton hugetlb_unshare_pmds(vma, floor, ceil); 4930b30c14cdSJames Houghton } 4931b30c14cdSJames Houghton 493231383c68SDan Williams return 0; 493331383c68SDan Williams } 493431383c68SDan Williams 493505ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 493605ea8860SDan Williams { 4937aca78307SMiaohe Lin return huge_page_size(hstate_vma(vma)); 493805ea8860SDan Williams } 493905ea8860SDan Williams 49401da177e4SLinus Torvalds /* 49411da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 49421da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 49436c26d310SMiaohe Lin * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get 49441da177e4SLinus Torvalds * this far. 49451da177e4SLinus Torvalds */ 4946b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 49471da177e4SLinus Torvalds { 49481da177e4SLinus Torvalds BUG(); 4949d0217ac0SNick Piggin return 0; 49501da177e4SLinus Torvalds } 49511da177e4SLinus Torvalds 4952eec3636aSJane Chu /* 4953eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 4954eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 4955eec3636aSJane Chu * This is because under System V memory model, mappings created via 4956eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 4957eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 4958eec3636aSJane Chu */ 4959f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 4960d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 496184afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 4962a1e78772SMel Gorman .close = hugetlb_vm_op_close, 4963dd3b614fSDmitry Safonov .may_split = hugetlb_vm_op_split, 496405ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 49651da177e4SLinus Torvalds }; 49661da177e4SLinus Torvalds 49671e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 49681e8f889bSDavid Gibson int writable) 496963551ae0SDavid Gibson { 497063551ae0SDavid Gibson pte_t entry; 497179c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 497263551ae0SDavid Gibson 49731e8f889bSDavid Gibson if (writable) { 4974106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 4975106c992aSGerald Schaefer vma->vm_page_prot))); 497663551ae0SDavid Gibson } else { 4977106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 4978106c992aSGerald Schaefer vma->vm_page_prot)); 497963551ae0SDavid Gibson } 498063551ae0SDavid Gibson entry = pte_mkyoung(entry); 498179c1c594SChristophe Leroy entry = arch_make_huge_pte(entry, shift, vma->vm_flags); 498263551ae0SDavid Gibson 498363551ae0SDavid Gibson return entry; 498463551ae0SDavid Gibson } 498563551ae0SDavid Gibson 49861e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 49871e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 49881e8f889bSDavid Gibson { 49891e8f889bSDavid Gibson pte_t entry; 49901e8f889bSDavid Gibson 4991106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 499232f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 49934b3073e1SRussell King update_mmu_cache(vma, address, ptep); 49941e8f889bSDavid Gibson } 49951e8f889bSDavid Gibson 4996d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 49974a705fefSNaoya Horiguchi { 49984a705fefSNaoya Horiguchi swp_entry_t swp; 49994a705fefSNaoya Horiguchi 50004a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 5001d5ed7444SAneesh Kumar K.V return false; 50024a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 5003d79d176aSBaoquan He if (is_migration_entry(swp)) 5004d5ed7444SAneesh Kumar K.V return true; 50054a705fefSNaoya Horiguchi else 5006d5ed7444SAneesh Kumar K.V return false; 50074a705fefSNaoya Horiguchi } 50084a705fefSNaoya Horiguchi 50093e5c3600SBaoquan He static bool is_hugetlb_entry_hwpoisoned(pte_t pte) 50104a705fefSNaoya Horiguchi { 50114a705fefSNaoya Horiguchi swp_entry_t swp; 50124a705fefSNaoya Horiguchi 50134a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 50143e5c3600SBaoquan He return false; 50154a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 5016d79d176aSBaoquan He if (is_hwpoison_entry(swp)) 50173e5c3600SBaoquan He return true; 50184a705fefSNaoya Horiguchi else 50193e5c3600SBaoquan He return false; 50204a705fefSNaoya Horiguchi } 50211e8f889bSDavid Gibson 50224eae4efaSPeter Xu static void 5023ea4c353dSSidhartha Kumar hugetlb_install_folio(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, 50245a2f8d22SPeter Xu struct folio *new_folio, pte_t old) 50254eae4efaSPeter Xu { 50265a2f8d22SPeter Xu pte_t newpte = make_huge_pte(vma, &new_folio->page, 1); 50275a2f8d22SPeter Xu 5028ea4c353dSSidhartha Kumar __folio_mark_uptodate(new_folio); 5029d0ce0e47SSidhartha Kumar hugepage_add_new_anon_rmap(new_folio, vma, addr); 50305a2f8d22SPeter Xu if (userfaultfd_wp(vma) && huge_pte_uffd_wp(old)) 50315a2f8d22SPeter Xu newpte = huge_pte_mkuffd_wp(newpte); 50325a2f8d22SPeter Xu set_huge_pte_at(vma->vm_mm, addr, ptep, newpte); 50334eae4efaSPeter Xu hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); 5034ea4c353dSSidhartha Kumar folio_set_hugetlb_migratable(new_folio); 50354eae4efaSPeter Xu } 50364eae4efaSPeter Xu 503763551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 5038bc70fbf2SPeter Xu struct vm_area_struct *dst_vma, 5039bc70fbf2SPeter Xu struct vm_area_struct *src_vma) 504063551ae0SDavid Gibson { 50413aa4ed80SMiaohe Lin pte_t *src_pte, *dst_pte, entry; 5042ad27ce20SZhangPeng struct folio *pte_folio; 50431c59827dSHugh Dickins unsigned long addr; 5044bc70fbf2SPeter Xu bool cow = is_cow_mapping(src_vma->vm_flags); 5045bc70fbf2SPeter Xu struct hstate *h = hstate_vma(src_vma); 5046a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 50474eae4efaSPeter Xu unsigned long npages = pages_per_huge_page(h); 5048ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 5049e95a9851SMike Kravetz unsigned long last_addr_mask; 5050e8569dd2SAndreas Sandberg int ret = 0; 50511e8f889bSDavid Gibson 5052ac46d4f3SJérôme Glisse if (cow) { 50537d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src, 5054bc70fbf2SPeter Xu src_vma->vm_start, 5055bc70fbf2SPeter Xu src_vma->vm_end); 5056ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 5057e727bfd5SSuren Baghdasaryan vma_assert_write_locked(src_vma); 5058623a1ddfSDavid Hildenbrand raw_write_seqcount_begin(&src->write_protect_seq); 505940549ba8SMike Kravetz } else { 506040549ba8SMike Kravetz /* 506140549ba8SMike Kravetz * For shared mappings the vma lock must be held before 50629c67a207SPeter Xu * calling hugetlb_walk() in the src vma. Otherwise, the 506340549ba8SMike Kravetz * returned ptep could go away if part of a shared pmd and 506440549ba8SMike Kravetz * another thread calls huge_pmd_unshare. 506540549ba8SMike Kravetz */ 506640549ba8SMike Kravetz hugetlb_vma_lock_read(src_vma); 5067ac46d4f3SJérôme Glisse } 5068e8569dd2SAndreas Sandberg 5069e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 5070bc70fbf2SPeter Xu for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) { 5071cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 50729c67a207SPeter Xu src_pte = hugetlb_walk(src_vma, addr, sz); 5073e95a9851SMike Kravetz if (!src_pte) { 5074e95a9851SMike Kravetz addr |= last_addr_mask; 5075c74df32cSHugh Dickins continue; 5076e95a9851SMike Kravetz } 5077bc70fbf2SPeter Xu dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz); 5078e8569dd2SAndreas Sandberg if (!dst_pte) { 5079e8569dd2SAndreas Sandberg ret = -ENOMEM; 5080e8569dd2SAndreas Sandberg break; 5081e8569dd2SAndreas Sandberg } 5082c5c99429SLarry Woodman 50835e41540cSMike Kravetz /* 50845e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 50855e41540cSMike Kravetz * 50863aa4ed80SMiaohe Lin * dst_pte == src_pte is the common case of src/dest sharing. 50875e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 50883aa4ed80SMiaohe Lin * another vma. So page_count of ptep page is checked instead 50893aa4ed80SMiaohe Lin * to reliably determine whether pte is shared. 50905e41540cSMike Kravetz */ 50913aa4ed80SMiaohe Lin if (page_count(virt_to_page(dst_pte)) > 1) { 5092e95a9851SMike Kravetz addr |= last_addr_mask; 5093c5c99429SLarry Woodman continue; 5094e95a9851SMike Kravetz } 5095c5c99429SLarry Woodman 5096cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 5097cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 5098cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 50994a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 51004eae4efaSPeter Xu again: 51013aa4ed80SMiaohe Lin if (huge_pte_none(entry)) { 51025e41540cSMike Kravetz /* 51033aa4ed80SMiaohe Lin * Skip if src entry none. 51045e41540cSMike Kravetz */ 51054a705fefSNaoya Horiguchi ; 5106c2cb0dccSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) { 51075a2f8d22SPeter Xu if (!userfaultfd_wp(dst_vma)) 5108c2cb0dccSNaoya Horiguchi entry = huge_pte_clear_uffd_wp(entry); 5109c2cb0dccSNaoya Horiguchi set_huge_pte_at(dst, addr, dst_pte, entry); 5110c2cb0dccSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry))) { 51114a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 51125a2f8d22SPeter Xu bool uffd_wp = pte_swp_uffd_wp(entry); 51134a705fefSNaoya Horiguchi 51146c287605SDavid Hildenbrand if (!is_readable_migration_entry(swp_entry) && cow) { 51154a705fefSNaoya Horiguchi /* 51164a705fefSNaoya Horiguchi * COW mappings require pages in both 51174a705fefSNaoya Horiguchi * parent and child to be set to read. 51184a705fefSNaoya Horiguchi */ 51194dd845b5SAlistair Popple swp_entry = make_readable_migration_entry( 51204dd845b5SAlistair Popple swp_offset(swp_entry)); 51214a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 5122bc70fbf2SPeter Xu if (userfaultfd_wp(src_vma) && uffd_wp) 51235a2f8d22SPeter Xu entry = pte_swp_mkuffd_wp(entry); 512418f39629SQi Zheng set_huge_pte_at(src, addr, src_pte, entry); 51254a705fefSNaoya Horiguchi } 51265a2f8d22SPeter Xu if (!userfaultfd_wp(dst_vma)) 5127bc70fbf2SPeter Xu entry = huge_pte_clear_uffd_wp(entry); 512818f39629SQi Zheng set_huge_pte_at(dst, addr, dst_pte, entry); 5129bc70fbf2SPeter Xu } else if (unlikely(is_pte_marker(entry))) { 5130af19487fSAxel Rasmussen pte_marker marker = copy_pte_marker( 5131af19487fSAxel Rasmussen pte_to_swp_entry(entry), dst_vma); 5132af19487fSAxel Rasmussen 5133af19487fSAxel Rasmussen if (marker) 5134af19487fSAxel Rasmussen set_huge_pte_at(dst, addr, dst_pte, 5135af19487fSAxel Rasmussen make_pte_marker(marker)); 51364a705fefSNaoya Horiguchi } else { 51374eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 5138ad27ce20SZhangPeng pte_folio = page_folio(pte_page(entry)); 5139ad27ce20SZhangPeng folio_get(pte_folio); 51404eae4efaSPeter Xu 51414eae4efaSPeter Xu /* 5142fb3d824dSDavid Hildenbrand * Failing to duplicate the anon rmap is a rare case 5143fb3d824dSDavid Hildenbrand * where we see pinned hugetlb pages while they're 5144fb3d824dSDavid Hildenbrand * prone to COW. We need to do the COW earlier during 5145fb3d824dSDavid Hildenbrand * fork. 51464eae4efaSPeter Xu * 51474eae4efaSPeter Xu * When pre-allocating the page or copying data, we 51484eae4efaSPeter Xu * need to be without the pgtable locks since we could 51494eae4efaSPeter Xu * sleep during the process. 51504eae4efaSPeter Xu */ 5151ad27ce20SZhangPeng if (!folio_test_anon(pte_folio)) { 5152ad27ce20SZhangPeng page_dup_file_rmap(&pte_folio->page, true); 5153ad27ce20SZhangPeng } else if (page_try_dup_anon_rmap(&pte_folio->page, 5154ad27ce20SZhangPeng true, src_vma)) { 51554eae4efaSPeter Xu pte_t src_pte_old = entry; 5156d0ce0e47SSidhartha Kumar struct folio *new_folio; 51574eae4efaSPeter Xu 51584eae4efaSPeter Xu spin_unlock(src_ptl); 51594eae4efaSPeter Xu spin_unlock(dst_ptl); 51604eae4efaSPeter Xu /* Do not use reserve as it's private owned */ 5161d0ce0e47SSidhartha Kumar new_folio = alloc_hugetlb_folio(dst_vma, addr, 1); 5162d0ce0e47SSidhartha Kumar if (IS_ERR(new_folio)) { 5163ad27ce20SZhangPeng folio_put(pte_folio); 5164d0ce0e47SSidhartha Kumar ret = PTR_ERR(new_folio); 51654eae4efaSPeter Xu break; 51664eae4efaSPeter Xu } 51671cb9dc4bSLiu Shixin ret = copy_user_large_folio(new_folio, 5168ad27ce20SZhangPeng pte_folio, 5169c0e8150eSZhangPeng addr, dst_vma); 5170ad27ce20SZhangPeng folio_put(pte_folio); 51711cb9dc4bSLiu Shixin if (ret) { 51721cb9dc4bSLiu Shixin folio_put(new_folio); 51731cb9dc4bSLiu Shixin break; 51741cb9dc4bSLiu Shixin } 51754eae4efaSPeter Xu 5176d0ce0e47SSidhartha Kumar /* Install the new hugetlb folio if src pte stable */ 51774eae4efaSPeter Xu dst_ptl = huge_pte_lock(h, dst, dst_pte); 51784eae4efaSPeter Xu src_ptl = huge_pte_lockptr(h, src, src_pte); 51794eae4efaSPeter Xu spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 51804eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 51814eae4efaSPeter Xu if (!pte_same(src_pte_old, entry)) { 5182bc70fbf2SPeter Xu restore_reserve_on_error(h, dst_vma, addr, 5183d2d7bb44SSidhartha Kumar new_folio); 5184d0ce0e47SSidhartha Kumar folio_put(new_folio); 51853aa4ed80SMiaohe Lin /* huge_ptep of dst_pte won't change as in child */ 51864eae4efaSPeter Xu goto again; 51874eae4efaSPeter Xu } 51885a2f8d22SPeter Xu hugetlb_install_folio(dst_vma, dst_pte, addr, 51895a2f8d22SPeter Xu new_folio, src_pte_old); 51904eae4efaSPeter Xu spin_unlock(src_ptl); 51914eae4efaSPeter Xu spin_unlock(dst_ptl); 51924eae4efaSPeter Xu continue; 51934eae4efaSPeter Xu } 51944eae4efaSPeter Xu 519534ee645eSJoerg Roedel if (cow) { 51960f10851eSJérôme Glisse /* 51970f10851eSJérôme Glisse * No need to notify as we are downgrading page 51980f10851eSJérôme Glisse * table protection not changing it to point 51990f10851eSJérôme Glisse * to a new page. 52000f10851eSJérôme Glisse * 5201ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 52020f10851eSJérôme Glisse */ 52037f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 520484894e1cSPeter Xu entry = huge_pte_wrprotect(entry); 520534ee645eSJoerg Roedel } 52064eae4efaSPeter Xu 52075a2f8d22SPeter Xu if (!userfaultfd_wp(dst_vma)) 52085a2f8d22SPeter Xu entry = huge_pte_clear_uffd_wp(entry); 52095a2f8d22SPeter Xu 521063551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 52114eae4efaSPeter Xu hugetlb_count_add(npages, dst); 52121c59827dSHugh Dickins } 5213cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 5214cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 521563551ae0SDavid Gibson } 521663551ae0SDavid Gibson 5217623a1ddfSDavid Hildenbrand if (cow) { 5218623a1ddfSDavid Hildenbrand raw_write_seqcount_end(&src->write_protect_seq); 5219ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 522040549ba8SMike Kravetz } else { 522140549ba8SMike Kravetz hugetlb_vma_unlock_read(src_vma); 5222623a1ddfSDavid Hildenbrand } 5223e8569dd2SAndreas Sandberg 5224e8569dd2SAndreas Sandberg return ret; 522563551ae0SDavid Gibson } 522663551ae0SDavid Gibson 5227550a7d60SMina Almasry static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr, 5228db110a99SAneesh Kumar K.V unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte) 5229550a7d60SMina Almasry { 5230550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 5231550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 5232550a7d60SMina Almasry spinlock_t *src_ptl, *dst_ptl; 5233db110a99SAneesh Kumar K.V pte_t pte; 5234550a7d60SMina Almasry 5235550a7d60SMina Almasry dst_ptl = huge_pte_lock(h, mm, dst_pte); 5236550a7d60SMina Almasry src_ptl = huge_pte_lockptr(h, mm, src_pte); 5237550a7d60SMina Almasry 5238550a7d60SMina Almasry /* 5239550a7d60SMina Almasry * We don't have to worry about the ordering of src and dst ptlocks 52408651a137SLorenzo Stoakes * because exclusive mmap_lock (or the i_mmap_lock) prevents deadlock. 5241550a7d60SMina Almasry */ 5242550a7d60SMina Almasry if (src_ptl != dst_ptl) 5243550a7d60SMina Almasry spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 5244550a7d60SMina Almasry 5245550a7d60SMina Almasry pte = huge_ptep_get_and_clear(mm, old_addr, src_pte); 5246550a7d60SMina Almasry set_huge_pte_at(mm, new_addr, dst_pte, pte); 5247550a7d60SMina Almasry 5248550a7d60SMina Almasry if (src_ptl != dst_ptl) 5249550a7d60SMina Almasry spin_unlock(src_ptl); 5250550a7d60SMina Almasry spin_unlock(dst_ptl); 5251550a7d60SMina Almasry } 5252550a7d60SMina Almasry 5253550a7d60SMina Almasry int move_hugetlb_page_tables(struct vm_area_struct *vma, 5254550a7d60SMina Almasry struct vm_area_struct *new_vma, 5255550a7d60SMina Almasry unsigned long old_addr, unsigned long new_addr, 5256550a7d60SMina Almasry unsigned long len) 5257550a7d60SMina Almasry { 5258550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 5259550a7d60SMina Almasry struct address_space *mapping = vma->vm_file->f_mapping; 5260550a7d60SMina Almasry unsigned long sz = huge_page_size(h); 5261550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 5262550a7d60SMina Almasry unsigned long old_end = old_addr + len; 5263e95a9851SMike Kravetz unsigned long last_addr_mask; 5264550a7d60SMina Almasry pte_t *src_pte, *dst_pte; 5265550a7d60SMina Almasry struct mmu_notifier_range range; 52663d0b95cdSBaolin Wang bool shared_pmd = false; 5267550a7d60SMina Almasry 52687d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, old_addr, 5269550a7d60SMina Almasry old_end); 5270550a7d60SMina Almasry adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 52713d0b95cdSBaolin Wang /* 52723d0b95cdSBaolin Wang * In case of shared PMDs, we should cover the maximum possible 52733d0b95cdSBaolin Wang * range. 52743d0b95cdSBaolin Wang */ 52753d0b95cdSBaolin Wang flush_cache_range(vma, range.start, range.end); 52763d0b95cdSBaolin Wang 5277550a7d60SMina Almasry mmu_notifier_invalidate_range_start(&range); 5278e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 5279550a7d60SMina Almasry /* Prevent race with file truncation */ 528040549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 5281550a7d60SMina Almasry i_mmap_lock_write(mapping); 5282550a7d60SMina Almasry for (; old_addr < old_end; old_addr += sz, new_addr += sz) { 52839c67a207SPeter Xu src_pte = hugetlb_walk(vma, old_addr, sz); 5284e95a9851SMike Kravetz if (!src_pte) { 5285e95a9851SMike Kravetz old_addr |= last_addr_mask; 5286e95a9851SMike Kravetz new_addr |= last_addr_mask; 5287550a7d60SMina Almasry continue; 5288e95a9851SMike Kravetz } 5289550a7d60SMina Almasry if (huge_pte_none(huge_ptep_get(src_pte))) 5290550a7d60SMina Almasry continue; 5291550a7d60SMina Almasry 52924ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) { 52933d0b95cdSBaolin Wang shared_pmd = true; 52944ddb4d91SMike Kravetz old_addr |= last_addr_mask; 52954ddb4d91SMike Kravetz new_addr |= last_addr_mask; 5296550a7d60SMina Almasry continue; 52973d0b95cdSBaolin Wang } 5298550a7d60SMina Almasry 5299550a7d60SMina Almasry dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz); 5300550a7d60SMina Almasry if (!dst_pte) 5301550a7d60SMina Almasry break; 5302550a7d60SMina Almasry 5303db110a99SAneesh Kumar K.V move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte); 5304550a7d60SMina Almasry } 53053d0b95cdSBaolin Wang 53063d0b95cdSBaolin Wang if (shared_pmd) 5307f720b471SKefeng Wang flush_hugetlb_tlb_range(vma, range.start, range.end); 53083d0b95cdSBaolin Wang else 5309f720b471SKefeng Wang flush_hugetlb_tlb_range(vma, old_end - len, old_end); 5310550a7d60SMina Almasry mmu_notifier_invalidate_range_end(&range); 531113e4ad2cSNadav Amit i_mmap_unlock_write(mapping); 531240549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 5313550a7d60SMina Almasry 5314550a7d60SMina Almasry return len + old_addr - old_end; 5315550a7d60SMina Almasry } 5316550a7d60SMina Almasry 531773c54763SPeter Xu static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 531824669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 531905e90bd0SPeter Xu struct page *ref_page, zap_flags_t zap_flags) 532063551ae0SDavid Gibson { 532163551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 532263551ae0SDavid Gibson unsigned long address; 5323c7546f8fSDavid Gibson pte_t *ptep; 532463551ae0SDavid Gibson pte_t pte; 5325cb900f41SKirill A. Shutemov spinlock_t *ptl; 532663551ae0SDavid Gibson struct page *page; 5327a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 5328a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 5329e95a9851SMike Kravetz unsigned long last_addr_mask; 5330a4a118f2SNadav Amit bool force_flush = false; 5331a5516438SAndi Kleen 533263551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 5333a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 5334a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 533563551ae0SDavid Gibson 533607e32661SAneesh Kumar K.V /* 533707e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 533807e32661SAneesh Kumar K.V * to huge page. 533907e32661SAneesh Kumar K.V */ 5340ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 534124669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 5342dff11abeSMike Kravetz 5343e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 5344569f48b8SHillf Danton address = start; 5345569f48b8SHillf Danton for (; address < end; address += sz) { 53469c67a207SPeter Xu ptep = hugetlb_walk(vma, address, sz); 5347e95a9851SMike Kravetz if (!ptep) { 5348e95a9851SMike Kravetz address |= last_addr_mask; 5349c7546f8fSDavid Gibson continue; 5350e95a9851SMike Kravetz } 5351c7546f8fSDavid Gibson 5352cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 53534ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, address, ptep)) { 535431d49da5SAneesh Kumar K.V spin_unlock(ptl); 5355a4a118f2SNadav Amit tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE); 5356a4a118f2SNadav Amit force_flush = true; 53574ddb4d91SMike Kravetz address |= last_addr_mask; 535831d49da5SAneesh Kumar K.V continue; 535931d49da5SAneesh Kumar K.V } 536039dde65cSChen, Kenneth W 53616629326bSHillf Danton pte = huge_ptep_get(ptep); 536231d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 536331d49da5SAneesh Kumar K.V spin_unlock(ptl); 536431d49da5SAneesh Kumar K.V continue; 536531d49da5SAneesh Kumar K.V } 53666629326bSHillf Danton 53676629326bSHillf Danton /* 53689fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 53699fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 53706629326bSHillf Danton */ 53719fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 537205e90bd0SPeter Xu /* 537305e90bd0SPeter Xu * If the pte was wr-protected by uffd-wp in any of the 537405e90bd0SPeter Xu * swap forms, meanwhile the caller does not want to 537505e90bd0SPeter Xu * drop the uffd-wp bit in this zap, then replace the 537605e90bd0SPeter Xu * pte with a marker. 537705e90bd0SPeter Xu */ 537805e90bd0SPeter Xu if (pte_swp_uffd_wp_any(pte) && 537905e90bd0SPeter Xu !(zap_flags & ZAP_FLAG_DROP_MARKER)) 538005e90bd0SPeter Xu set_huge_pte_at(mm, address, ptep, 538105e90bd0SPeter Xu make_pte_marker(PTE_MARKER_UFFD_WP)); 538205e90bd0SPeter Xu else 53839386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 538431d49da5SAneesh Kumar K.V spin_unlock(ptl); 538531d49da5SAneesh Kumar K.V continue; 53868c4894c6SNaoya Horiguchi } 53876629326bSHillf Danton 53886629326bSHillf Danton page = pte_page(pte); 538904f2cbe3SMel Gorman /* 539004f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 539104f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 539204f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 539304f2cbe3SMel Gorman */ 539404f2cbe3SMel Gorman if (ref_page) { 539531d49da5SAneesh Kumar K.V if (page != ref_page) { 539631d49da5SAneesh Kumar K.V spin_unlock(ptl); 539731d49da5SAneesh Kumar K.V continue; 539831d49da5SAneesh Kumar K.V } 539904f2cbe3SMel Gorman /* 540004f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 540104f2cbe3SMel Gorman * future faults in this VMA will fail rather than 540204f2cbe3SMel Gorman * looking like data was lost 540304f2cbe3SMel Gorman */ 540404f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 540504f2cbe3SMel Gorman } 540604f2cbe3SMel Gorman 5407c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 5408b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 5409106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 54106649a386SKen Chen set_page_dirty(page); 541105e90bd0SPeter Xu /* Leave a uffd-wp pte marker if needed */ 541205e90bd0SPeter Xu if (huge_pte_uffd_wp(pte) && 541305e90bd0SPeter Xu !(zap_flags & ZAP_FLAG_DROP_MARKER)) 541405e90bd0SPeter Xu set_huge_pte_at(mm, address, ptep, 541505e90bd0SPeter Xu make_pte_marker(PTE_MARKER_UFFD_WP)); 54165d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 5417cea86fe2SHugh Dickins page_remove_rmap(page, vma, true); 541831d49da5SAneesh Kumar K.V 5419cb900f41SKirill A. Shutemov spin_unlock(ptl); 5420e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 542124669e58SAneesh Kumar K.V /* 542231d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 542324669e58SAneesh Kumar K.V */ 542431d49da5SAneesh Kumar K.V if (ref_page) 542531d49da5SAneesh Kumar K.V break; 5426fe1668aeSChen, Kenneth W } 542724669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 5428a4a118f2SNadav Amit 5429a4a118f2SNadav Amit /* 5430a4a118f2SNadav Amit * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We 5431a4a118f2SNadav Amit * could defer the flush until now, since by holding i_mmap_rwsem we 5432a4a118f2SNadav Amit * guaranteed that the last refernece would not be dropped. But we must 5433a4a118f2SNadav Amit * do the flushing before we return, as otherwise i_mmap_rwsem will be 5434a4a118f2SNadav Amit * dropped and the last reference to the shared PMDs page might be 5435a4a118f2SNadav Amit * dropped as well. 5436a4a118f2SNadav Amit * 5437a4a118f2SNadav Amit * In theory we could defer the freeing of the PMD pages as well, but 5438a4a118f2SNadav Amit * huge_pmd_unshare() relies on the exact page_count for the PMD page to 5439a4a118f2SNadav Amit * detect sharing, so we cannot defer the release of the page either. 5440a4a118f2SNadav Amit * Instead, do flush now. 5441a4a118f2SNadav Amit */ 5442a4a118f2SNadav Amit if (force_flush) 5443a4a118f2SNadav Amit tlb_flush_mmu_tlbonly(tlb); 54441da177e4SLinus Torvalds } 544563551ae0SDavid Gibson 5446d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 5447d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 544805e90bd0SPeter Xu unsigned long end, struct page *ref_page, 544905e90bd0SPeter Xu zap_flags_t zap_flags) 5450d833352aSMel Gorman { 5451131a79b4SMike Kravetz hugetlb_vma_lock_write(vma); 5452131a79b4SMike Kravetz i_mmap_lock_write(vma->vm_file->f_mapping); 5453131a79b4SMike Kravetz 5454369258ceSMike Kravetz /* mmu notification performed in caller */ 545505e90bd0SPeter Xu __unmap_hugepage_range(tlb, vma, start, end, ref_page, zap_flags); 5456d833352aSMel Gorman 545704ada095SMike Kravetz if (zap_flags & ZAP_FLAG_UNMAP) { /* final unmap */ 5458d833352aSMel Gorman /* 545904ada095SMike Kravetz * Unlock and free the vma lock before releasing i_mmap_rwsem. 546004ada095SMike Kravetz * When the vma_lock is freed, this makes the vma ineligible 546104ada095SMike Kravetz * for pmd sharing. And, i_mmap_rwsem is required to set up 546204ada095SMike Kravetz * pmd sharing. This is important as page tables for this 546304ada095SMike Kravetz * unmapped range will be asynchrously deleted. If the page 546404ada095SMike Kravetz * tables are shared, there will be issues when accessed by 546504ada095SMike Kravetz * someone else. 5466d833352aSMel Gorman */ 5467ecfbd733SMike Kravetz __hugetlb_vma_unlock_write_free(vma); 5468131a79b4SMike Kravetz i_mmap_unlock_write(vma->vm_file->f_mapping); 546904ada095SMike Kravetz } else { 547004ada095SMike Kravetz i_mmap_unlock_write(vma->vm_file->f_mapping); 547104ada095SMike Kravetz hugetlb_vma_unlock_write(vma); 547204ada095SMike Kravetz } 5473d833352aSMel Gorman } 5474d833352aSMel Gorman 5475502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 547605e90bd0SPeter Xu unsigned long end, struct page *ref_page, 547705e90bd0SPeter Xu zap_flags_t zap_flags) 5478502717f4SChen, Kenneth W { 5479369258ceSMike Kravetz struct mmu_notifier_range range; 548024669e58SAneesh Kumar K.V struct mmu_gather tlb; 5481dff11abeSMike Kravetz 54827d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm, 5483369258ceSMike Kravetz start, end); 5484369258ceSMike Kravetz adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 5485369258ceSMike Kravetz mmu_notifier_invalidate_range_start(&range); 5486a72afd87SWill Deacon tlb_gather_mmu(&tlb, vma->vm_mm); 5487369258ceSMike Kravetz 548805e90bd0SPeter Xu __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags); 5489369258ceSMike Kravetz 5490369258ceSMike Kravetz mmu_notifier_invalidate_range_end(&range); 5491ae8eba8bSWill Deacon tlb_finish_mmu(&tlb); 5492502717f4SChen, Kenneth W } 5493502717f4SChen, Kenneth W 549404f2cbe3SMel Gorman /* 549504f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 5496578b7725SZhiyuan Dai * mapping it owns the reserve page for. The intention is to unmap the page 549704f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 549804f2cbe3SMel Gorman * same region. 549904f2cbe3SMel Gorman */ 55002f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 55012a4b3dedSHarvey Harrison struct page *page, unsigned long address) 550204f2cbe3SMel Gorman { 55037526674dSAdam Litke struct hstate *h = hstate_vma(vma); 550404f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 550504f2cbe3SMel Gorman struct address_space *mapping; 550604f2cbe3SMel Gorman pgoff_t pgoff; 550704f2cbe3SMel Gorman 550804f2cbe3SMel Gorman /* 550904f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 551004f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 551104f2cbe3SMel Gorman */ 55127526674dSAdam Litke address = address & huge_page_mask(h); 551336e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 551436e4f20aSMichal Hocko vma->vm_pgoff; 551593c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 551604f2cbe3SMel Gorman 55174eb2b1dcSMel Gorman /* 55184eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 55194eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 55204eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 55214eb2b1dcSMel Gorman */ 552283cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 55236b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 552404f2cbe3SMel Gorman /* Do not unmap the current VMA */ 552504f2cbe3SMel Gorman if (iter_vma == vma) 552604f2cbe3SMel Gorman continue; 552704f2cbe3SMel Gorman 552804f2cbe3SMel Gorman /* 55292f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 55302f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 55312f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 55322f84a899SMel Gorman */ 55332f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 55342f84a899SMel Gorman continue; 55352f84a899SMel Gorman 55362f84a899SMel Gorman /* 553704f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 553804f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 553904f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 554004f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 554104f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 554204f2cbe3SMel Gorman */ 554304f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 554424669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 554505e90bd0SPeter Xu address + huge_page_size(h), page, 0); 554604f2cbe3SMel Gorman } 554783cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 554804f2cbe3SMel Gorman } 554904f2cbe3SMel Gorman 55500fe6e20bSNaoya Horiguchi /* 5551c89357e2SDavid Hildenbrand * hugetlb_wp() should be called with page lock of the original hugepage held. 5552aa6d2e8cSBaolin Wang * Called with hugetlb_fault_mutex_table held and pte_page locked so we 5553ef009b25SMichal Hocko * cannot race with other handlers or page migration. 5554ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 55550fe6e20bSNaoya Horiguchi */ 5556c89357e2SDavid Hildenbrand static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma, 5557c89357e2SDavid Hildenbrand unsigned long address, pte_t *ptep, unsigned int flags, 5558371607a3SSidhartha Kumar struct folio *pagecache_folio, spinlock_t *ptl) 55591e8f889bSDavid Gibson { 5560c89357e2SDavid Hildenbrand const bool unshare = flags & FAULT_FLAG_UNSHARE; 556160d5b473SPeter Xu pte_t pte = huge_ptep_get(ptep); 5562a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 5563959a78b6SZhangPeng struct folio *old_folio; 5564d0ce0e47SSidhartha Kumar struct folio *new_folio; 55652b740303SSouptick Joarder int outside_reserve = 0; 55662b740303SSouptick Joarder vm_fault_t ret = 0; 5567974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 5568ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 55691e8f889bSDavid Gibson 55701d8d1464SDavid Hildenbrand /* 557160d5b473SPeter Xu * Never handle CoW for uffd-wp protected pages. It should be only 557260d5b473SPeter Xu * handled when the uffd-wp protection is removed. 557360d5b473SPeter Xu * 557460d5b473SPeter Xu * Note that only the CoW optimization path (in hugetlb_no_page()) 557560d5b473SPeter Xu * can trigger this, because hugetlb_fault() will always resolve 557660d5b473SPeter Xu * uffd-wp bit first. 557760d5b473SPeter Xu */ 557860d5b473SPeter Xu if (!unshare && huge_pte_uffd_wp(pte)) 557960d5b473SPeter Xu return 0; 558060d5b473SPeter Xu 558160d5b473SPeter Xu /* 55821d8d1464SDavid Hildenbrand * hugetlb does not support FOLL_FORCE-style write faults that keep the 55831d8d1464SDavid Hildenbrand * PTE mapped R/O such as maybe_mkwrite() would do. 55841d8d1464SDavid Hildenbrand */ 55851d8d1464SDavid Hildenbrand if (WARN_ON_ONCE(!unshare && !(vma->vm_flags & VM_WRITE))) 55861d8d1464SDavid Hildenbrand return VM_FAULT_SIGSEGV; 55871d8d1464SDavid Hildenbrand 55881d8d1464SDavid Hildenbrand /* Let's take out MAP_SHARED mappings first. */ 55891d8d1464SDavid Hildenbrand if (vma->vm_flags & VM_MAYSHARE) { 55901d8d1464SDavid Hildenbrand set_huge_ptep_writable(vma, haddr, ptep); 55911d8d1464SDavid Hildenbrand return 0; 55921d8d1464SDavid Hildenbrand } 55931d8d1464SDavid Hildenbrand 5594959a78b6SZhangPeng old_folio = page_folio(pte_page(pte)); 55951e8f889bSDavid Gibson 5596662ce1dcSYang Yang delayacct_wpcopy_start(); 5597662ce1dcSYang Yang 559804f2cbe3SMel Gorman retry_avoidcopy: 5599c89357e2SDavid Hildenbrand /* 5600c89357e2SDavid Hildenbrand * If no-one else is actually using this page, we're the exclusive 5601c89357e2SDavid Hildenbrand * owner and can reuse this page. 5602c89357e2SDavid Hildenbrand */ 5603959a78b6SZhangPeng if (folio_mapcount(old_folio) == 1 && folio_test_anon(old_folio)) { 5604959a78b6SZhangPeng if (!PageAnonExclusive(&old_folio->page)) 5605959a78b6SZhangPeng page_move_anon_rmap(&old_folio->page, vma); 5606c89357e2SDavid Hildenbrand if (likely(!unshare)) 56075b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 5608662ce1dcSYang Yang 5609662ce1dcSYang Yang delayacct_wpcopy_end(); 561083c54070SNick Piggin return 0; 56111e8f889bSDavid Gibson } 5612959a78b6SZhangPeng VM_BUG_ON_PAGE(folio_test_anon(old_folio) && 5613959a78b6SZhangPeng PageAnonExclusive(&old_folio->page), &old_folio->page); 56141e8f889bSDavid Gibson 561504f2cbe3SMel Gorman /* 561604f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 561704f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 561804f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 561904f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 562004f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 562104f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 562204f2cbe3SMel Gorman * of the full address range. 562304f2cbe3SMel Gorman */ 56245944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 5625959a78b6SZhangPeng old_folio != pagecache_folio) 562604f2cbe3SMel Gorman outside_reserve = 1; 562704f2cbe3SMel Gorman 5628959a78b6SZhangPeng folio_get(old_folio); 5629b76c8cfbSLarry Woodman 5630ad4404a2SDavidlohr Bueso /* 5631ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 5632ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 5633ad4404a2SDavidlohr Bueso */ 5634cb900f41SKirill A. Shutemov spin_unlock(ptl); 5635d0ce0e47SSidhartha Kumar new_folio = alloc_hugetlb_folio(vma, haddr, outside_reserve); 56361e8f889bSDavid Gibson 5637d0ce0e47SSidhartha Kumar if (IS_ERR(new_folio)) { 563804f2cbe3SMel Gorman /* 563904f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 564004f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 564104f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 564204f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 564304f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 564404f2cbe3SMel Gorman */ 564504f2cbe3SMel Gorman if (outside_reserve) { 564640549ba8SMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 564740549ba8SMike Kravetz pgoff_t idx; 564840549ba8SMike Kravetz u32 hash; 564940549ba8SMike Kravetz 5650959a78b6SZhangPeng folio_put(old_folio); 565140549ba8SMike Kravetz /* 565240549ba8SMike Kravetz * Drop hugetlb_fault_mutex and vma_lock before 565340549ba8SMike Kravetz * unmapping. unmapping needs to hold vma_lock 565440549ba8SMike Kravetz * in write mode. Dropping vma_lock in read mode 565540549ba8SMike Kravetz * here is OK as COW mappings do not interact with 565640549ba8SMike Kravetz * PMD sharing. 565740549ba8SMike Kravetz * 565840549ba8SMike Kravetz * Reacquire both after unmap operation. 565940549ba8SMike Kravetz */ 566040549ba8SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 566140549ba8SMike Kravetz hash = hugetlb_fault_mutex_hash(mapping, idx); 566240549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 566340549ba8SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 566440549ba8SMike Kravetz 5665959a78b6SZhangPeng unmap_ref_private(mm, vma, &old_folio->page, haddr); 566640549ba8SMike Kravetz 566740549ba8SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 566840549ba8SMike Kravetz hugetlb_vma_lock_read(vma); 5669cb900f41SKirill A. Shutemov spin_lock(ptl); 56709c67a207SPeter Xu ptep = hugetlb_walk(vma, haddr, huge_page_size(h)); 5671a9af0c5dSNaoya Horiguchi if (likely(ptep && 5672a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 567304f2cbe3SMel Gorman goto retry_avoidcopy; 5674a734bcc8SHillf Danton /* 5675cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 5676cb900f41SKirill A. Shutemov * lock, and our job is done. 5677a734bcc8SHillf Danton */ 5678662ce1dcSYang Yang delayacct_wpcopy_end(); 5679a734bcc8SHillf Danton return 0; 568004f2cbe3SMel Gorman } 568104f2cbe3SMel Gorman 5682d0ce0e47SSidhartha Kumar ret = vmf_error(PTR_ERR(new_folio)); 5683ad4404a2SDavidlohr Bueso goto out_release_old; 56841e8f889bSDavid Gibson } 56851e8f889bSDavid Gibson 56860fe6e20bSNaoya Horiguchi /* 56870fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 56880fe6e20bSNaoya Horiguchi * anon_vma prepared. 56890fe6e20bSNaoya Horiguchi */ 569044e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 5691ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 5692ad4404a2SDavidlohr Bueso goto out_release_all; 569344e2aa93SDean Nelson } 56940fe6e20bSNaoya Horiguchi 5695959a78b6SZhangPeng if (copy_user_large_folio(new_folio, old_folio, address, vma)) { 56961cb9dc4bSLiu Shixin ret = VM_FAULT_HWPOISON_LARGE; 56971cb9dc4bSLiu Shixin goto out_release_all; 56981cb9dc4bSLiu Shixin } 5699d0ce0e47SSidhartha Kumar __folio_mark_uptodate(new_folio); 57001e8f889bSDavid Gibson 57017d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, haddr, 57026f4f13e8SJérôme Glisse haddr + huge_page_size(h)); 5703ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 5704ad4404a2SDavidlohr Bueso 5705b76c8cfbSLarry Woodman /* 5706cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 5707b76c8cfbSLarry Woodman * before the page tables are altered 5708b76c8cfbSLarry Woodman */ 5709cb900f41SKirill A. Shutemov spin_lock(ptl); 57109c67a207SPeter Xu ptep = hugetlb_walk(vma, haddr, huge_page_size(h)); 5711a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 57120f230bc2SPeter Xu pte_t newpte = make_huge_pte(vma, &new_folio->page, !unshare); 57130f230bc2SPeter Xu 5714c89357e2SDavid Hildenbrand /* Break COW or unshare */ 57155b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 5716959a78b6SZhangPeng page_remove_rmap(&old_folio->page, vma, true); 5717d0ce0e47SSidhartha Kumar hugepage_add_new_anon_rmap(new_folio, vma, haddr); 57180f230bc2SPeter Xu if (huge_pte_uffd_wp(pte)) 57190f230bc2SPeter Xu newpte = huge_pte_mkuffd_wp(newpte); 57200f230bc2SPeter Xu set_huge_pte_at(mm, haddr, ptep, newpte); 5721d0ce0e47SSidhartha Kumar folio_set_hugetlb_migratable(new_folio); 57221e8f889bSDavid Gibson /* Make the old page be freed below */ 5723959a78b6SZhangPeng new_folio = old_folio; 57241e8f889bSDavid Gibson } 5725cb900f41SKirill A. Shutemov spin_unlock(ptl); 5726ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 5727ad4404a2SDavidlohr Bueso out_release_all: 5728c89357e2SDavid Hildenbrand /* 5729c89357e2SDavid Hildenbrand * No restore in case of successful pagetable update (Break COW or 5730c89357e2SDavid Hildenbrand * unshare) 5731c89357e2SDavid Hildenbrand */ 5732959a78b6SZhangPeng if (new_folio != old_folio) 5733d2d7bb44SSidhartha Kumar restore_reserve_on_error(h, vma, haddr, new_folio); 5734d0ce0e47SSidhartha Kumar folio_put(new_folio); 5735ad4404a2SDavidlohr Bueso out_release_old: 5736959a78b6SZhangPeng folio_put(old_folio); 57378312034fSJoonsoo Kim 5738ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 5739662ce1dcSYang Yang 5740662ce1dcSYang Yang delayacct_wpcopy_end(); 5741ad4404a2SDavidlohr Bueso return ret; 57421e8f889bSDavid Gibson } 57431e8f889bSDavid Gibson 57443ae77f43SHugh Dickins /* 57453ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 57463ae77f43SHugh Dickins */ 57473ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 57482a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 57492a15efc9SHugh Dickins { 575091a2fb95SSidhartha Kumar struct address_space *mapping = vma->vm_file->f_mapping; 575191a2fb95SSidhartha Kumar pgoff_t idx = vma_hugecache_offset(h, vma, address); 5752fd4aed8dSMike Kravetz struct folio *folio; 57532a15efc9SHugh Dickins 5754fd4aed8dSMike Kravetz folio = filemap_get_folio(mapping, idx); 5755fd4aed8dSMike Kravetz if (IS_ERR(folio)) 5756fd4aed8dSMike Kravetz return false; 5757fd4aed8dSMike Kravetz folio_put(folio); 5758fd4aed8dSMike Kravetz return true; 57592a15efc9SHugh Dickins } 57602a15efc9SHugh Dickins 57619b91c0e2SSidhartha Kumar int hugetlb_add_to_page_cache(struct folio *folio, struct address_space *mapping, 5762ab76ad54SMike Kravetz pgoff_t idx) 5763ab76ad54SMike Kravetz { 5764ab76ad54SMike Kravetz struct inode *inode = mapping->host; 5765ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 5766d9ef44deSMatthew Wilcox (Oracle) int err; 5767ab76ad54SMike Kravetz 5768d9ef44deSMatthew Wilcox (Oracle) __folio_set_locked(folio); 5769d9ef44deSMatthew Wilcox (Oracle) err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL); 5770d9ef44deSMatthew Wilcox (Oracle) 5771d9ef44deSMatthew Wilcox (Oracle) if (unlikely(err)) { 5772d9ef44deSMatthew Wilcox (Oracle) __folio_clear_locked(folio); 5773ab76ad54SMike Kravetz return err; 5774d9ef44deSMatthew Wilcox (Oracle) } 57759b91c0e2SSidhartha Kumar folio_clear_hugetlb_restore_reserve(folio); 5776ab76ad54SMike Kravetz 577722146c3cSMike Kravetz /* 5778d9ef44deSMatthew Wilcox (Oracle) * mark folio dirty so that it will not be removed from cache/file 577922146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 578022146c3cSMike Kravetz */ 5781d9ef44deSMatthew Wilcox (Oracle) folio_mark_dirty(folio); 578222146c3cSMike Kravetz 5783ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 5784ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 5785ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 5786ab76ad54SMike Kravetz return 0; 5787ab76ad54SMike Kravetz } 5788ab76ad54SMike Kravetz 57897677f7fdSAxel Rasmussen static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma, 57907677f7fdSAxel Rasmussen struct address_space *mapping, 57917677f7fdSAxel Rasmussen pgoff_t idx, 57927677f7fdSAxel Rasmussen unsigned int flags, 57937677f7fdSAxel Rasmussen unsigned long haddr, 5794824ddc60SNadav Amit unsigned long addr, 57957677f7fdSAxel Rasmussen unsigned long reason) 57967677f7fdSAxel Rasmussen { 57977677f7fdSAxel Rasmussen u32 hash; 57987677f7fdSAxel Rasmussen struct vm_fault vmf = { 57997677f7fdSAxel Rasmussen .vma = vma, 58007677f7fdSAxel Rasmussen .address = haddr, 5801824ddc60SNadav Amit .real_address = addr, 58027677f7fdSAxel Rasmussen .flags = flags, 58037677f7fdSAxel Rasmussen 58047677f7fdSAxel Rasmussen /* 58057677f7fdSAxel Rasmussen * Hard to debug if it ends up being 58067677f7fdSAxel Rasmussen * used by a callee that assumes 58077677f7fdSAxel Rasmussen * something about the other 58087677f7fdSAxel Rasmussen * uninitialized fields... same as in 58097677f7fdSAxel Rasmussen * memory.c 58107677f7fdSAxel Rasmussen */ 58117677f7fdSAxel Rasmussen }; 58127677f7fdSAxel Rasmussen 58137677f7fdSAxel Rasmussen /* 5814958f32ceSLiu Shixin * vma_lock and hugetlb_fault_mutex must be dropped before handling 5815958f32ceSLiu Shixin * userfault. Also mmap_lock could be dropped due to handling 5816958f32ceSLiu Shixin * userfault, any vma operation should be careful from here. 58177677f7fdSAxel Rasmussen */ 581840549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 58197677f7fdSAxel Rasmussen hash = hugetlb_fault_mutex_hash(mapping, idx); 58207677f7fdSAxel Rasmussen mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5821958f32ceSLiu Shixin return handle_userfault(&vmf, reason); 58227677f7fdSAxel Rasmussen } 58237677f7fdSAxel Rasmussen 58242ea7ff1eSPeter Xu /* 58252ea7ff1eSPeter Xu * Recheck pte with pgtable lock. Returns true if pte didn't change, or 58262ea7ff1eSPeter Xu * false if pte changed or is changing. 58272ea7ff1eSPeter Xu */ 58282ea7ff1eSPeter Xu static bool hugetlb_pte_stable(struct hstate *h, struct mm_struct *mm, 58292ea7ff1eSPeter Xu pte_t *ptep, pte_t old_pte) 58302ea7ff1eSPeter Xu { 58312ea7ff1eSPeter Xu spinlock_t *ptl; 58322ea7ff1eSPeter Xu bool same; 58332ea7ff1eSPeter Xu 58342ea7ff1eSPeter Xu ptl = huge_pte_lock(h, mm, ptep); 58352ea7ff1eSPeter Xu same = pte_same(huge_ptep_get(ptep), old_pte); 58362ea7ff1eSPeter Xu spin_unlock(ptl); 58372ea7ff1eSPeter Xu 58382ea7ff1eSPeter Xu return same; 58392ea7ff1eSPeter Xu } 58402ea7ff1eSPeter Xu 58412b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 58422b740303SSouptick Joarder struct vm_area_struct *vma, 58438382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 5844c64e912cSPeter Xu unsigned long address, pte_t *ptep, 5845c64e912cSPeter Xu pte_t old_pte, unsigned int flags) 5846ac9b9c66SHugh Dickins { 5847a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 58482b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 5849409eb8c2SHillf Danton int anon_rmap = 0; 58504c887265SAdam Litke unsigned long size; 5851d0ce0e47SSidhartha Kumar struct folio *folio; 58521e8f889bSDavid Gibson pte_t new_pte; 5853cb900f41SKirill A. Shutemov spinlock_t *ptl; 5854285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 5855d0ce0e47SSidhartha Kumar bool new_folio, new_pagecache_folio = false; 5856958f32ceSLiu Shixin u32 hash = hugetlb_fault_mutex_hash(mapping, idx); 58574c887265SAdam Litke 585804f2cbe3SMel Gorman /* 585904f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 586004f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 5861c89357e2SDavid Hildenbrand * COW/unsharing. Warn that such a situation has occurred as it may not 5862c89357e2SDavid Hildenbrand * be obvious. 586304f2cbe3SMel Gorman */ 586404f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 5865910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 586604f2cbe3SMel Gorman current->pid); 5867958f32ceSLiu Shixin goto out; 586804f2cbe3SMel Gorman } 586904f2cbe3SMel Gorman 58704c887265SAdam Litke /* 5871188a3972SMike Kravetz * Use page lock to guard against racing truncation 5872188a3972SMike Kravetz * before we get page_table_lock. 58734c887265SAdam Litke */ 5874d0ce0e47SSidhartha Kumar new_folio = false; 5875d0ce0e47SSidhartha Kumar folio = filemap_lock_folio(mapping, idx); 587666dabbb6SChristoph Hellwig if (IS_ERR(folio)) { 5877188a3972SMike Kravetz size = i_size_read(mapping->host) >> huge_page_shift(h); 5878188a3972SMike Kravetz if (idx >= size) 5879188a3972SMike Kravetz goto out; 58807677f7fdSAxel Rasmussen /* Check for page in userfault range */ 58812ea7ff1eSPeter Xu if (userfaultfd_missing(vma)) { 58822ea7ff1eSPeter Xu /* 58832ea7ff1eSPeter Xu * Since hugetlb_no_page() was examining pte 58842ea7ff1eSPeter Xu * without pgtable lock, we need to re-test under 58852ea7ff1eSPeter Xu * lock because the pte may not be stable and could 58862ea7ff1eSPeter Xu * have changed from under us. Try to detect 58872ea7ff1eSPeter Xu * either changed or during-changing ptes and retry 58882ea7ff1eSPeter Xu * properly when needed. 58892ea7ff1eSPeter Xu * 58902ea7ff1eSPeter Xu * Note that userfaultfd is actually fine with 58912ea7ff1eSPeter Xu * false positives (e.g. caused by pte changed), 58922ea7ff1eSPeter Xu * but not wrong logical events (e.g. caused by 58932ea7ff1eSPeter Xu * reading a pte during changing). The latter can 58942ea7ff1eSPeter Xu * confuse the userspace, so the strictness is very 58952ea7ff1eSPeter Xu * much preferred. E.g., MISSING event should 58962ea7ff1eSPeter Xu * never happen on the page after UFFDIO_COPY has 58972ea7ff1eSPeter Xu * correctly installed the page and returned. 58982ea7ff1eSPeter Xu */ 58992ea7ff1eSPeter Xu if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) { 59002ea7ff1eSPeter Xu ret = 0; 59012ea7ff1eSPeter Xu goto out; 59022ea7ff1eSPeter Xu } 59032ea7ff1eSPeter Xu 59042ea7ff1eSPeter Xu return hugetlb_handle_userfault(vma, mapping, idx, flags, 59052ea7ff1eSPeter Xu haddr, address, 59067677f7fdSAxel Rasmussen VM_UFFD_MISSING); 59072ea7ff1eSPeter Xu } 59081a1aad8aSMike Kravetz 5909d0ce0e47SSidhartha Kumar folio = alloc_hugetlb_folio(vma, haddr, 0); 5910d0ce0e47SSidhartha Kumar if (IS_ERR(folio)) { 59114643d67eSMike Kravetz /* 59124643d67eSMike Kravetz * Returning error will result in faulting task being 59134643d67eSMike Kravetz * sent SIGBUS. The hugetlb fault mutex prevents two 59144643d67eSMike Kravetz * tasks from racing to fault in the same page which 59154643d67eSMike Kravetz * could result in false unable to allocate errors. 59164643d67eSMike Kravetz * Page migration does not take the fault mutex, but 59174643d67eSMike Kravetz * does a clear then write of pte's under page table 59184643d67eSMike Kravetz * lock. Page fault code could race with migration, 59194643d67eSMike Kravetz * notice the clear pte and try to allocate a page 59204643d67eSMike Kravetz * here. Before returning error, get ptl and make 59214643d67eSMike Kravetz * sure there really is no pte entry. 59224643d67eSMike Kravetz */ 5923f9bf6c03SPeter Xu if (hugetlb_pte_stable(h, mm, ptep, old_pte)) 5924d0ce0e47SSidhartha Kumar ret = vmf_error(PTR_ERR(folio)); 5925f9bf6c03SPeter Xu else 5926f9bf6c03SPeter Xu ret = 0; 59276bda666aSChristoph Lameter goto out; 59286bda666aSChristoph Lameter } 5929d0ce0e47SSidhartha Kumar clear_huge_page(&folio->page, address, pages_per_huge_page(h)); 5930d0ce0e47SSidhartha Kumar __folio_mark_uptodate(folio); 5931d0ce0e47SSidhartha Kumar new_folio = true; 5932ac9b9c66SHugh Dickins 5933f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 59349b91c0e2SSidhartha Kumar int err = hugetlb_add_to_page_cache(folio, mapping, idx); 59356bda666aSChristoph Lameter if (err) { 59363a5497a2SMiaohe Lin /* 59373a5497a2SMiaohe Lin * err can't be -EEXIST which implies someone 59383a5497a2SMiaohe Lin * else consumed the reservation since hugetlb 59393a5497a2SMiaohe Lin * fault mutex is held when add a hugetlb page 59403a5497a2SMiaohe Lin * to the page cache. So it's safe to call 59413a5497a2SMiaohe Lin * restore_reserve_on_error() here. 59423a5497a2SMiaohe Lin */ 5943d2d7bb44SSidhartha Kumar restore_reserve_on_error(h, vma, haddr, folio); 5944d0ce0e47SSidhartha Kumar folio_put(folio); 59456bda666aSChristoph Lameter goto out; 59466bda666aSChristoph Lameter } 5947d0ce0e47SSidhartha Kumar new_pagecache_folio = true; 594823be7468SMel Gorman } else { 5949d0ce0e47SSidhartha Kumar folio_lock(folio); 59500fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 59510fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 59520fe6e20bSNaoya Horiguchi goto backout_unlocked; 595323be7468SMel Gorman } 5954409eb8c2SHillf Danton anon_rmap = 1; 59550fe6e20bSNaoya Horiguchi } 59560fe6e20bSNaoya Horiguchi } else { 595757303d80SAndy Whitcroft /* 5958998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 5959998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 5960998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 5961fd6a03edSNaoya Horiguchi */ 5962d0ce0e47SSidhartha Kumar if (unlikely(folio_test_hwpoison(folio))) { 59630eb98f15SMiaohe Lin ret = VM_FAULT_HWPOISON_LARGE | 5964972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 5965fd6a03edSNaoya Horiguchi goto backout_unlocked; 59666bda666aSChristoph Lameter } 59677677f7fdSAxel Rasmussen 59687677f7fdSAxel Rasmussen /* Check for page in userfault range. */ 59697677f7fdSAxel Rasmussen if (userfaultfd_minor(vma)) { 5970d0ce0e47SSidhartha Kumar folio_unlock(folio); 5971d0ce0e47SSidhartha Kumar folio_put(folio); 59722ea7ff1eSPeter Xu /* See comment in userfaultfd_missing() block above */ 59732ea7ff1eSPeter Xu if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) { 59742ea7ff1eSPeter Xu ret = 0; 59752ea7ff1eSPeter Xu goto out; 59762ea7ff1eSPeter Xu } 59772ea7ff1eSPeter Xu return hugetlb_handle_userfault(vma, mapping, idx, flags, 59782ea7ff1eSPeter Xu haddr, address, 59797677f7fdSAxel Rasmussen VM_UFFD_MINOR); 59807677f7fdSAxel Rasmussen } 5981998b4382SNaoya Horiguchi } 59821e8f889bSDavid Gibson 598357303d80SAndy Whitcroft /* 598457303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 598557303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 598657303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 598757303d80SAndy Whitcroft * the spinlock. 598857303d80SAndy Whitcroft */ 59895e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 5990285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 59912b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 59922b26736cSAndy Whitcroft goto backout_unlocked; 59932b26736cSAndy Whitcroft } 59945e911373SMike Kravetz /* Just decrements count, does not deallocate */ 5995285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 59965e911373SMike Kravetz } 599757303d80SAndy Whitcroft 59988bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 599983c54070SNick Piggin ret = 0; 6000c64e912cSPeter Xu /* If pte changed from under us, retry */ 6001c64e912cSPeter Xu if (!pte_same(huge_ptep_get(ptep), old_pte)) 60024c887265SAdam Litke goto backout; 60034c887265SAdam Litke 60044781593dSPeter Xu if (anon_rmap) 6005d0ce0e47SSidhartha Kumar hugepage_add_new_anon_rmap(folio, vma, haddr); 60064781593dSPeter Xu else 6007d0ce0e47SSidhartha Kumar page_dup_file_rmap(&folio->page, true); 6008d0ce0e47SSidhartha Kumar new_pte = make_huge_pte(vma, &folio->page, ((vma->vm_flags & VM_WRITE) 60091e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 6010c64e912cSPeter Xu /* 6011c64e912cSPeter Xu * If this pte was previously wr-protected, keep it wr-protected even 6012c64e912cSPeter Xu * if populated. 6013c64e912cSPeter Xu */ 6014c64e912cSPeter Xu if (unlikely(pte_marker_uffd_wp(old_pte))) 6015f1eb1bacSPeter Xu new_pte = huge_pte_mkuffd_wp(new_pte); 6016285b8dcaSHuang Ying set_huge_pte_at(mm, haddr, ptep, new_pte); 60171e8f889bSDavid Gibson 60185d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 6019788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 60201e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 6021371607a3SSidhartha Kumar ret = hugetlb_wp(mm, vma, address, ptep, flags, folio, ptl); 60221e8f889bSDavid Gibson } 60231e8f889bSDavid Gibson 6024cb900f41SKirill A. Shutemov spin_unlock(ptl); 6025cb6acd01SMike Kravetz 6026cb6acd01SMike Kravetz /* 6027d0ce0e47SSidhartha Kumar * Only set hugetlb_migratable in newly allocated pages. Existing pages 6028d0ce0e47SSidhartha Kumar * found in the pagecache may not have hugetlb_migratable if they have 60298f251a3dSMike Kravetz * been isolated for migration. 6030cb6acd01SMike Kravetz */ 6031d0ce0e47SSidhartha Kumar if (new_folio) 6032d0ce0e47SSidhartha Kumar folio_set_hugetlb_migratable(folio); 6033cb6acd01SMike Kravetz 6034d0ce0e47SSidhartha Kumar folio_unlock(folio); 60354c887265SAdam Litke out: 6036958f32ceSLiu Shixin hugetlb_vma_unlock_read(vma); 6037958f32ceSLiu Shixin mutex_unlock(&hugetlb_fault_mutex_table[hash]); 6038ac9b9c66SHugh Dickins return ret; 60394c887265SAdam Litke 60404c887265SAdam Litke backout: 6041cb900f41SKirill A. Shutemov spin_unlock(ptl); 60422b26736cSAndy Whitcroft backout_unlocked: 6043d0ce0e47SSidhartha Kumar if (new_folio && !new_pagecache_folio) 6044d2d7bb44SSidhartha Kumar restore_reserve_on_error(h, vma, haddr, folio); 6045fa27759aSMike Kravetz 6046d0ce0e47SSidhartha Kumar folio_unlock(folio); 6047d0ce0e47SSidhartha Kumar folio_put(folio); 60484c887265SAdam Litke goto out; 6049ac9b9c66SHugh Dickins } 6050ac9b9c66SHugh Dickins 60518382d914SDavidlohr Bueso #ifdef CONFIG_SMP 6052188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 60538382d914SDavidlohr Bueso { 60548382d914SDavidlohr Bueso unsigned long key[2]; 60558382d914SDavidlohr Bueso u32 hash; 60568382d914SDavidlohr Bueso 60578382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 60588382d914SDavidlohr Bueso key[1] = idx; 60598382d914SDavidlohr Bueso 606055254636SMike Kravetz hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); 60618382d914SDavidlohr Bueso 60628382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 60638382d914SDavidlohr Bueso } 60648382d914SDavidlohr Bueso #else 60658382d914SDavidlohr Bueso /* 60666c26d310SMiaohe Lin * For uniprocessor systems we always use a single mutex, so just 60678382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 60688382d914SDavidlohr Bueso */ 6069188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 60708382d914SDavidlohr Bueso { 60718382d914SDavidlohr Bueso return 0; 60728382d914SDavidlohr Bueso } 60738382d914SDavidlohr Bueso #endif 60748382d914SDavidlohr Bueso 60752b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 6076788c7df4SHugh Dickins unsigned long address, unsigned int flags) 607786e5216fSAdam Litke { 60788382d914SDavidlohr Bueso pte_t *ptep, entry; 6079cb900f41SKirill A. Shutemov spinlock_t *ptl; 60802b740303SSouptick Joarder vm_fault_t ret; 60818382d914SDavidlohr Bueso u32 hash; 60828382d914SDavidlohr Bueso pgoff_t idx; 6083061e62e8SZhangPeng struct folio *folio = NULL; 6084371607a3SSidhartha Kumar struct folio *pagecache_folio = NULL; 6085a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 60868382d914SDavidlohr Bueso struct address_space *mapping; 60870f792cf9SNaoya Horiguchi int need_wait_lock = 0; 6088285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 608986e5216fSAdam Litke 60904ec31152SMatthew Wilcox (Oracle) /* TODO: Handle faults under the VMA lock */ 60914ec31152SMatthew Wilcox (Oracle) if (flags & FAULT_FLAG_VMA_LOCK) { 60924ec31152SMatthew Wilcox (Oracle) vma_end_read(vma); 60934ec31152SMatthew Wilcox (Oracle) return VM_FAULT_RETRY; 60944ec31152SMatthew Wilcox (Oracle) } 60954ec31152SMatthew Wilcox (Oracle) 60963935baa9SDavid Gibson /* 60973935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 60983935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 60993935baa9SDavid Gibson * the same page in the page cache. 61003935baa9SDavid Gibson */ 610140549ba8SMike Kravetz mapping = vma->vm_file->f_mapping; 610240549ba8SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 6103188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 6104c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 61058382d914SDavidlohr Bueso 610640549ba8SMike Kravetz /* 610740549ba8SMike Kravetz * Acquire vma lock before calling huge_pte_alloc and hold 610840549ba8SMike Kravetz * until finished with ptep. This prevents huge_pmd_unshare from 610940549ba8SMike Kravetz * being called elsewhere and making the ptep no longer valid. 611040549ba8SMike Kravetz */ 611140549ba8SMike Kravetz hugetlb_vma_lock_read(vma); 611240549ba8SMike Kravetz ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); 611340549ba8SMike Kravetz if (!ptep) { 611440549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 611540549ba8SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 611640549ba8SMike Kravetz return VM_FAULT_OOM; 611740549ba8SMike Kravetz } 611840549ba8SMike Kravetz 61197f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 6120af19487fSAxel Rasmussen if (huge_pte_none_mostly(entry)) { 6121af19487fSAxel Rasmussen if (is_pte_marker(entry)) { 6122af19487fSAxel Rasmussen pte_marker marker = 6123af19487fSAxel Rasmussen pte_marker_get(pte_to_swp_entry(entry)); 6124af19487fSAxel Rasmussen 6125af19487fSAxel Rasmussen if (marker & PTE_MARKER_POISONED) { 6126af19487fSAxel Rasmussen ret = VM_FAULT_HWPOISON_LARGE; 6127af19487fSAxel Rasmussen goto out_mutex; 6128af19487fSAxel Rasmussen } 6129af19487fSAxel Rasmussen } 6130af19487fSAxel Rasmussen 6131958f32ceSLiu Shixin /* 6132af19487fSAxel Rasmussen * Other PTE markers should be handled the same way as none PTE. 6133af19487fSAxel Rasmussen * 6134958f32ceSLiu Shixin * hugetlb_no_page will drop vma lock and hugetlb fault 6135958f32ceSLiu Shixin * mutex internally, which make us return immediately. 6136958f32ceSLiu Shixin */ 6137958f32ceSLiu Shixin return hugetlb_no_page(mm, vma, mapping, idx, address, ptep, 6138c64e912cSPeter Xu entry, flags); 6139af19487fSAxel Rasmussen } 614086e5216fSAdam Litke 614183c54070SNick Piggin ret = 0; 61421e8f889bSDavid Gibson 614357303d80SAndy Whitcroft /* 61440f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 61450f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 61467c8de358SEthon Paul * an active hugepage in pagecache. This goto expects the 2nd page 61477c8de358SEthon Paul * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will 61487c8de358SEthon Paul * properly handle it. 61490f792cf9SNaoya Horiguchi */ 6150fcd48540SPeter Xu if (!pte_present(entry)) { 6151fcd48540SPeter Xu if (unlikely(is_hugetlb_entry_migration(entry))) { 6152fcd48540SPeter Xu /* 6153fcd48540SPeter Xu * Release the hugetlb fault lock now, but retain 6154fcd48540SPeter Xu * the vma lock, because it is needed to guard the 6155fcd48540SPeter Xu * huge_pte_lockptr() later in 6156fcd48540SPeter Xu * migration_entry_wait_huge(). The vma lock will 6157fcd48540SPeter Xu * be released there. 6158fcd48540SPeter Xu */ 6159fcd48540SPeter Xu mutex_unlock(&hugetlb_fault_mutex_table[hash]); 6160fcd48540SPeter Xu migration_entry_wait_huge(vma, ptep); 6161fcd48540SPeter Xu return 0; 6162fcd48540SPeter Xu } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 6163fcd48540SPeter Xu ret = VM_FAULT_HWPOISON_LARGE | 6164fcd48540SPeter Xu VM_FAULT_SET_HINDEX(hstate_index(h)); 61650f792cf9SNaoya Horiguchi goto out_mutex; 6166fcd48540SPeter Xu } 61670f792cf9SNaoya Horiguchi 61680f792cf9SNaoya Horiguchi /* 6169c89357e2SDavid Hildenbrand * If we are going to COW/unshare the mapping later, we examine the 6170c89357e2SDavid Hildenbrand * pending reservations for this page now. This will ensure that any 617157303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 61721d8d1464SDavid Hildenbrand * spinlock. Also lookup the pagecache page now as it is used to 61731d8d1464SDavid Hildenbrand * determine if a reservation has been consumed. 617457303d80SAndy Whitcroft */ 6175c89357e2SDavid Hildenbrand if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) && 61761d8d1464SDavid Hildenbrand !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(entry)) { 6177285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 61782b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 6179b4d1d99fSDavid Gibson goto out_mutex; 61802b26736cSAndy Whitcroft } 61815e911373SMike Kravetz /* Just decrements count, does not deallocate */ 6182285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 618357303d80SAndy Whitcroft 6184371607a3SSidhartha Kumar pagecache_folio = filemap_lock_folio(mapping, idx); 618566dabbb6SChristoph Hellwig if (IS_ERR(pagecache_folio)) 618666dabbb6SChristoph Hellwig pagecache_folio = NULL; 618757303d80SAndy Whitcroft } 618857303d80SAndy Whitcroft 61890f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 61900fe6e20bSNaoya Horiguchi 6191c89357e2SDavid Hildenbrand /* Check for a racing update before calling hugetlb_wp() */ 6192b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 6193cb900f41SKirill A. Shutemov goto out_ptl; 6194b4d1d99fSDavid Gibson 6195166f3eccSPeter Xu /* Handle userfault-wp first, before trying to lock more pages */ 6196166f3eccSPeter Xu if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) && 6197166f3eccSPeter Xu (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 6198166f3eccSPeter Xu struct vm_fault vmf = { 6199166f3eccSPeter Xu .vma = vma, 6200166f3eccSPeter Xu .address = haddr, 6201166f3eccSPeter Xu .real_address = address, 6202166f3eccSPeter Xu .flags = flags, 6203166f3eccSPeter Xu }; 6204166f3eccSPeter Xu 6205166f3eccSPeter Xu spin_unlock(ptl); 6206371607a3SSidhartha Kumar if (pagecache_folio) { 6207371607a3SSidhartha Kumar folio_unlock(pagecache_folio); 6208371607a3SSidhartha Kumar folio_put(pagecache_folio); 6209166f3eccSPeter Xu } 621040549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 6211166f3eccSPeter Xu mutex_unlock(&hugetlb_fault_mutex_table[hash]); 6212166f3eccSPeter Xu return handle_userfault(&vmf, VM_UFFD_WP); 6213166f3eccSPeter Xu } 6214166f3eccSPeter Xu 62150f792cf9SNaoya Horiguchi /* 6216c89357e2SDavid Hildenbrand * hugetlb_wp() requires page locks of pte_page(entry) and 6217371607a3SSidhartha Kumar * pagecache_folio, so here we need take the former one 6218061e62e8SZhangPeng * when folio != pagecache_folio or !pagecache_folio. 62190f792cf9SNaoya Horiguchi */ 6220061e62e8SZhangPeng folio = page_folio(pte_page(entry)); 6221061e62e8SZhangPeng if (folio != pagecache_folio) 6222061e62e8SZhangPeng if (!folio_trylock(folio)) { 62230f792cf9SNaoya Horiguchi need_wait_lock = 1; 62240f792cf9SNaoya Horiguchi goto out_ptl; 62250f792cf9SNaoya Horiguchi } 62260f792cf9SNaoya Horiguchi 6227061e62e8SZhangPeng folio_get(folio); 6228b4d1d99fSDavid Gibson 6229c89357e2SDavid Hildenbrand if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) { 6230106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 6231c89357e2SDavid Hildenbrand ret = hugetlb_wp(mm, vma, address, ptep, flags, 6232371607a3SSidhartha Kumar pagecache_folio, ptl); 62330f792cf9SNaoya Horiguchi goto out_put_page; 6234c89357e2SDavid Hildenbrand } else if (likely(flags & FAULT_FLAG_WRITE)) { 6235106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 6236b4d1d99fSDavid Gibson } 6237c89357e2SDavid Hildenbrand } 6238b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 6239285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 6240788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 6241285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 62420f792cf9SNaoya Horiguchi out_put_page: 6243061e62e8SZhangPeng if (folio != pagecache_folio) 6244061e62e8SZhangPeng folio_unlock(folio); 6245061e62e8SZhangPeng folio_put(folio); 6246cb900f41SKirill A. Shutemov out_ptl: 6247cb900f41SKirill A. Shutemov spin_unlock(ptl); 624857303d80SAndy Whitcroft 6249371607a3SSidhartha Kumar if (pagecache_folio) { 6250371607a3SSidhartha Kumar folio_unlock(pagecache_folio); 6251371607a3SSidhartha Kumar folio_put(pagecache_folio); 625257303d80SAndy Whitcroft } 6253b4d1d99fSDavid Gibson out_mutex: 625440549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 6255c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 62560f792cf9SNaoya Horiguchi /* 62570f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 62580f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 62590f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 62600f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 62610f792cf9SNaoya Horiguchi * here without taking refcount. 62620f792cf9SNaoya Horiguchi */ 62630f792cf9SNaoya Horiguchi if (need_wait_lock) 6264061e62e8SZhangPeng folio_wait_locked(folio); 62651e8f889bSDavid Gibson return ret; 626686e5216fSAdam Litke } 626786e5216fSAdam Litke 6268714c1891SAxel Rasmussen #ifdef CONFIG_USERFAULTFD 62698fb5debcSMike Kravetz /* 6270a734991cSAxel Rasmussen * Used by userfaultfd UFFDIO_* ioctls. Based on userfaultfd's mfill_atomic_pte 6271a734991cSAxel Rasmussen * with modifications for hugetlb pages. 62728fb5debcSMike Kravetz */ 627361c50040SAxel Rasmussen int hugetlb_mfill_atomic_pte(pte_t *dst_pte, 62748fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 62758fb5debcSMike Kravetz unsigned long dst_addr, 62768fb5debcSMike Kravetz unsigned long src_addr, 6277d9712937SAxel Rasmussen uffd_flags_t flags, 62780169fd51SZhangPeng struct folio **foliop) 62798fb5debcSMike Kravetz { 628061c50040SAxel Rasmussen struct mm_struct *dst_mm = dst_vma->vm_mm; 6281d9712937SAxel Rasmussen bool is_continue = uffd_flags_mode_is(flags, MFILL_ATOMIC_CONTINUE); 6282d9712937SAxel Rasmussen bool wp_enabled = (flags & MFILL_ATOMIC_WP); 62838cc5fcbbSMina Almasry struct hstate *h = hstate_vma(dst_vma); 62848cc5fcbbSMina Almasry struct address_space *mapping = dst_vma->vm_file->f_mapping; 62858cc5fcbbSMina Almasry pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr); 62861e392147SAndrea Arcangeli unsigned long size; 62871c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 62888fb5debcSMike Kravetz pte_t _dst_pte; 62898fb5debcSMike Kravetz spinlock_t *ptl; 62908cc5fcbbSMina Almasry int ret = -ENOMEM; 6291d0ce0e47SSidhartha Kumar struct folio *folio; 6292f6191471SAxel Rasmussen int writable; 6293d0ce0e47SSidhartha Kumar bool folio_in_pagecache = false; 62948fb5debcSMike Kravetz 62958a13897fSAxel Rasmussen if (uffd_flags_mode_is(flags, MFILL_ATOMIC_POISON)) { 62968a13897fSAxel Rasmussen ptl = huge_pte_lock(h, dst_mm, dst_pte); 62978a13897fSAxel Rasmussen 62988a13897fSAxel Rasmussen /* Don't overwrite any existing PTEs (even markers) */ 62998a13897fSAxel Rasmussen if (!huge_pte_none(huge_ptep_get(dst_pte))) { 63008a13897fSAxel Rasmussen spin_unlock(ptl); 63018a13897fSAxel Rasmussen return -EEXIST; 63028a13897fSAxel Rasmussen } 63038a13897fSAxel Rasmussen 63048a13897fSAxel Rasmussen _dst_pte = make_pte_marker(PTE_MARKER_POISONED); 63058a13897fSAxel Rasmussen set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 63068a13897fSAxel Rasmussen 63078a13897fSAxel Rasmussen /* No need to invalidate - it was non-present before */ 63088a13897fSAxel Rasmussen update_mmu_cache(dst_vma, dst_addr, dst_pte); 63098a13897fSAxel Rasmussen 63108a13897fSAxel Rasmussen spin_unlock(ptl); 63118a13897fSAxel Rasmussen return 0; 63128a13897fSAxel Rasmussen } 63138a13897fSAxel Rasmussen 6314f6191471SAxel Rasmussen if (is_continue) { 6315f6191471SAxel Rasmussen ret = -EFAULT; 6316d0ce0e47SSidhartha Kumar folio = filemap_lock_folio(mapping, idx); 631766dabbb6SChristoph Hellwig if (IS_ERR(folio)) 6318f6191471SAxel Rasmussen goto out; 6319d0ce0e47SSidhartha Kumar folio_in_pagecache = true; 63200169fd51SZhangPeng } else if (!*foliop) { 63210169fd51SZhangPeng /* If a folio already exists, then it's UFFDIO_COPY for 6322d84cf06eSMina Almasry * a non-missing case. Return -EEXIST. 6323d84cf06eSMina Almasry */ 6324d84cf06eSMina Almasry if (vm_shared && 6325d84cf06eSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 6326d84cf06eSMina Almasry ret = -EEXIST; 63278fb5debcSMike Kravetz goto out; 6328d84cf06eSMina Almasry } 6329d84cf06eSMina Almasry 6330d0ce0e47SSidhartha Kumar folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0); 6331d0ce0e47SSidhartha Kumar if (IS_ERR(folio)) { 6332d84cf06eSMina Almasry ret = -ENOMEM; 6333d84cf06eSMina Almasry goto out; 6334d84cf06eSMina Almasry } 63358fb5debcSMike Kravetz 6336e87340caSZhangPeng ret = copy_folio_from_user(folio, (const void __user *) src_addr, 6337e87340caSZhangPeng false); 63388fb5debcSMike Kravetz 6339c1e8d7c6SMichel Lespinasse /* fallback to copy_from_user outside mmap_lock */ 63408fb5debcSMike Kravetz if (unlikely(ret)) { 63419e368259SAndrea Arcangeli ret = -ENOENT; 6342d0ce0e47SSidhartha Kumar /* Free the allocated folio which may have 63438cc5fcbbSMina Almasry * consumed a reservation. 63448cc5fcbbSMina Almasry */ 6345d2d7bb44SSidhartha Kumar restore_reserve_on_error(h, dst_vma, dst_addr, folio); 6346d0ce0e47SSidhartha Kumar folio_put(folio); 63478cc5fcbbSMina Almasry 6348d0ce0e47SSidhartha Kumar /* Allocate a temporary folio to hold the copied 63498cc5fcbbSMina Almasry * contents. 63508cc5fcbbSMina Almasry */ 6351d0ce0e47SSidhartha Kumar folio = alloc_hugetlb_folio_vma(h, dst_vma, dst_addr); 6352d0ce0e47SSidhartha Kumar if (!folio) { 63538cc5fcbbSMina Almasry ret = -ENOMEM; 63548cc5fcbbSMina Almasry goto out; 63558cc5fcbbSMina Almasry } 63560169fd51SZhangPeng *foliop = folio; 63570169fd51SZhangPeng /* Set the outparam foliop and return to the caller to 63588cc5fcbbSMina Almasry * copy the contents outside the lock. Don't free the 63590169fd51SZhangPeng * folio. 63608cc5fcbbSMina Almasry */ 63618fb5debcSMike Kravetz goto out; 63628fb5debcSMike Kravetz } 63638fb5debcSMike Kravetz } else { 63648cc5fcbbSMina Almasry if (vm_shared && 63658cc5fcbbSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 63660169fd51SZhangPeng folio_put(*foliop); 63678cc5fcbbSMina Almasry ret = -EEXIST; 63680169fd51SZhangPeng *foliop = NULL; 63698cc5fcbbSMina Almasry goto out; 63708cc5fcbbSMina Almasry } 63718cc5fcbbSMina Almasry 6372d0ce0e47SSidhartha Kumar folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0); 6373d0ce0e47SSidhartha Kumar if (IS_ERR(folio)) { 63740169fd51SZhangPeng folio_put(*foliop); 63758cc5fcbbSMina Almasry ret = -ENOMEM; 63760169fd51SZhangPeng *foliop = NULL; 63778cc5fcbbSMina Almasry goto out; 63788cc5fcbbSMina Almasry } 63791cb9dc4bSLiu Shixin ret = copy_user_large_folio(folio, *foliop, dst_addr, dst_vma); 63800169fd51SZhangPeng folio_put(*foliop); 63810169fd51SZhangPeng *foliop = NULL; 63821cb9dc4bSLiu Shixin if (ret) { 63831cb9dc4bSLiu Shixin folio_put(folio); 63841cb9dc4bSLiu Shixin goto out; 63851cb9dc4bSLiu Shixin } 63868fb5debcSMike Kravetz } 63878fb5debcSMike Kravetz 63888fb5debcSMike Kravetz /* 6389d0ce0e47SSidhartha Kumar * The memory barrier inside __folio_mark_uptodate makes sure that 63908fb5debcSMike Kravetz * preceding stores to the page contents become visible before 63918fb5debcSMike Kravetz * the set_pte_at() write. 63928fb5debcSMike Kravetz */ 6393d0ce0e47SSidhartha Kumar __folio_mark_uptodate(folio); 63948fb5debcSMike Kravetz 6395f6191471SAxel Rasmussen /* Add shared, newly allocated pages to the page cache. */ 6396f6191471SAxel Rasmussen if (vm_shared && !is_continue) { 63971e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 63981e392147SAndrea Arcangeli ret = -EFAULT; 63991e392147SAndrea Arcangeli if (idx >= size) 64001e392147SAndrea Arcangeli goto out_release_nounlock; 64011c9e8defSMike Kravetz 64021e392147SAndrea Arcangeli /* 64031e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 64047e1813d4SMike Kravetz * hugetlb_add_to_page_cache() below happens through the 64051e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 64061e392147SAndrea Arcangeli * the caller. 64071e392147SAndrea Arcangeli */ 64089b91c0e2SSidhartha Kumar ret = hugetlb_add_to_page_cache(folio, mapping, idx); 64091c9e8defSMike Kravetz if (ret) 64101c9e8defSMike Kravetz goto out_release_nounlock; 6411d0ce0e47SSidhartha Kumar folio_in_pagecache = true; 64121c9e8defSMike Kravetz } 64131c9e8defSMike Kravetz 6414bcc66543SMiaohe Lin ptl = huge_pte_lock(h, dst_mm, dst_pte); 64158fb5debcSMike Kravetz 64168625147cSJames Houghton ret = -EIO; 6417d0ce0e47SSidhartha Kumar if (folio_test_hwpoison(folio)) 64188625147cSJames Houghton goto out_release_unlock; 64198625147cSJames Houghton 64201e392147SAndrea Arcangeli /* 64216041c691SPeter Xu * We allow to overwrite a pte marker: consider when both MISSING|WP 64226041c691SPeter Xu * registered, we firstly wr-protect a none pte which has no page cache 64236041c691SPeter Xu * page backing it, then access the page. 64246041c691SPeter Xu */ 6425fa27759aSMike Kravetz ret = -EEXIST; 64266041c691SPeter Xu if (!huge_pte_none_mostly(huge_ptep_get(dst_pte))) 64278fb5debcSMike Kravetz goto out_release_unlock; 64288fb5debcSMike Kravetz 6429d0ce0e47SSidhartha Kumar if (folio_in_pagecache) 6430d0ce0e47SSidhartha Kumar page_dup_file_rmap(&folio->page, true); 64314781593dSPeter Xu else 6432d0ce0e47SSidhartha Kumar hugepage_add_new_anon_rmap(folio, dst_vma, dst_addr); 64338fb5debcSMike Kravetz 64346041c691SPeter Xu /* 64356041c691SPeter Xu * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY 64366041c691SPeter Xu * with wp flag set, don't set pte write bit. 64376041c691SPeter Xu */ 6438d9712937SAxel Rasmussen if (wp_enabled || (is_continue && !vm_shared)) 6439f6191471SAxel Rasmussen writable = 0; 6440f6191471SAxel Rasmussen else 6441f6191471SAxel Rasmussen writable = dst_vma->vm_flags & VM_WRITE; 6442f6191471SAxel Rasmussen 6443d0ce0e47SSidhartha Kumar _dst_pte = make_huge_pte(dst_vma, &folio->page, writable); 64446041c691SPeter Xu /* 64456041c691SPeter Xu * Always mark UFFDIO_COPY page dirty; note that this may not be 64466041c691SPeter Xu * extremely important for hugetlbfs for now since swapping is not 64476041c691SPeter Xu * supported, but we should still be clear in that this page cannot be 64486041c691SPeter Xu * thrown away at will, even if write bit not set. 64496041c691SPeter Xu */ 64508fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 64518fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 64528fb5debcSMike Kravetz 6453d9712937SAxel Rasmussen if (wp_enabled) 64546041c691SPeter Xu _dst_pte = huge_pte_mkuffd_wp(_dst_pte); 64556041c691SPeter Xu 64568fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 64578fb5debcSMike Kravetz 64588fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 64598fb5debcSMike Kravetz 64608fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 64618fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 64628fb5debcSMike Kravetz 64638fb5debcSMike Kravetz spin_unlock(ptl); 6464f6191471SAxel Rasmussen if (!is_continue) 6465d0ce0e47SSidhartha Kumar folio_set_hugetlb_migratable(folio); 6466f6191471SAxel Rasmussen if (vm_shared || is_continue) 6467d0ce0e47SSidhartha Kumar folio_unlock(folio); 64688fb5debcSMike Kravetz ret = 0; 64698fb5debcSMike Kravetz out: 64708fb5debcSMike Kravetz return ret; 64718fb5debcSMike Kravetz out_release_unlock: 64728fb5debcSMike Kravetz spin_unlock(ptl); 6473f6191471SAxel Rasmussen if (vm_shared || is_continue) 6474d0ce0e47SSidhartha Kumar folio_unlock(folio); 64755af10dfdSAndrea Arcangeli out_release_nounlock: 6476d0ce0e47SSidhartha Kumar if (!folio_in_pagecache) 6477d2d7bb44SSidhartha Kumar restore_reserve_on_error(h, dst_vma, dst_addr, folio); 6478d0ce0e47SSidhartha Kumar folio_put(folio); 64798fb5debcSMike Kravetz goto out; 64808fb5debcSMike Kravetz } 6481714c1891SAxel Rasmussen #endif /* CONFIG_USERFAULTFD */ 64828fb5debcSMike Kravetz 648357a196a5SMike Kravetz struct page *hugetlb_follow_page_mask(struct vm_area_struct *vma, 64845502ea44SPeter Xu unsigned long address, unsigned int flags, 64855502ea44SPeter Xu unsigned int *page_mask) 648657a196a5SMike Kravetz { 648757a196a5SMike Kravetz struct hstate *h = hstate_vma(vma); 648857a196a5SMike Kravetz struct mm_struct *mm = vma->vm_mm; 648957a196a5SMike Kravetz unsigned long haddr = address & huge_page_mask(h); 649057a196a5SMike Kravetz struct page *page = NULL; 649157a196a5SMike Kravetz spinlock_t *ptl; 649257a196a5SMike Kravetz pte_t *pte, entry; 6493458568c9SPeter Xu int ret; 649457a196a5SMike Kravetz 64957d049f3aSPeter Xu hugetlb_vma_lock_read(vma); 64969c67a207SPeter Xu pte = hugetlb_walk(vma, haddr, huge_page_size(h)); 649757a196a5SMike Kravetz if (!pte) 64987d049f3aSPeter Xu goto out_unlock; 649957a196a5SMike Kravetz 650057a196a5SMike Kravetz ptl = huge_pte_lock(h, mm, pte); 650157a196a5SMike Kravetz entry = huge_ptep_get(pte); 650257a196a5SMike Kravetz if (pte_present(entry)) { 6503458568c9SPeter Xu page = pte_page(entry); 6504458568c9SPeter Xu 6505458568c9SPeter Xu if (!huge_pte_write(entry)) { 6506458568c9SPeter Xu if (flags & FOLL_WRITE) { 6507458568c9SPeter Xu page = NULL; 6508458568c9SPeter Xu goto out; 6509458568c9SPeter Xu } 6510458568c9SPeter Xu 6511458568c9SPeter Xu if (gup_must_unshare(vma, flags, page)) { 6512458568c9SPeter Xu /* Tell the caller to do unsharing */ 6513458568c9SPeter Xu page = ERR_PTR(-EMLINK); 6514458568c9SPeter Xu goto out; 6515458568c9SPeter Xu } 6516458568c9SPeter Xu } 6517458568c9SPeter Xu 6518458568c9SPeter Xu page += ((address & ~huge_page_mask(h)) >> PAGE_SHIFT); 6519458568c9SPeter Xu 652057a196a5SMike Kravetz /* 652157a196a5SMike Kravetz * Note that page may be a sub-page, and with vmemmap 652257a196a5SMike Kravetz * optimizations the page struct may be read only. 652357a196a5SMike Kravetz * try_grab_page() will increase the ref count on the 652457a196a5SMike Kravetz * head page, so this will be OK. 652557a196a5SMike Kravetz * 6526e2ca6ba6SLinus Torvalds * try_grab_page() should always be able to get the page here, 6527e2ca6ba6SLinus Torvalds * because we hold the ptl lock and have verified pte_present(). 652857a196a5SMike Kravetz */ 6529458568c9SPeter Xu ret = try_grab_page(page, flags); 6530458568c9SPeter Xu 6531458568c9SPeter Xu if (WARN_ON_ONCE(ret)) { 6532458568c9SPeter Xu page = ERR_PTR(ret); 653357a196a5SMike Kravetz goto out; 653457a196a5SMike Kravetz } 65355502ea44SPeter Xu 65365502ea44SPeter Xu *page_mask = (1U << huge_page_order(h)) - 1; 653757a196a5SMike Kravetz } 653857a196a5SMike Kravetz out: 653957a196a5SMike Kravetz spin_unlock(ptl); 65407d049f3aSPeter Xu out_unlock: 65417d049f3aSPeter Xu hugetlb_vma_unlock_read(vma); 6542dd767aaaSPeter Xu 6543dd767aaaSPeter Xu /* 6544dd767aaaSPeter Xu * Fixup retval for dump requests: if pagecache doesn't exist, 6545dd767aaaSPeter Xu * don't try to allocate a new page but just skip it. 6546dd767aaaSPeter Xu */ 6547dd767aaaSPeter Xu if (!page && (flags & FOLL_DUMP) && 6548dd767aaaSPeter Xu !hugetlbfs_pagecache_present(h, vma, address)) 6549dd767aaaSPeter Xu page = ERR_PTR(-EFAULT); 6550dd767aaaSPeter Xu 655157a196a5SMike Kravetz return page; 655257a196a5SMike Kravetz } 655357a196a5SMike Kravetz 6554a79390f5SPeter Xu long hugetlb_change_protection(struct vm_area_struct *vma, 65555a90d5a1SPeter Xu unsigned long address, unsigned long end, 65565a90d5a1SPeter Xu pgprot_t newprot, unsigned long cp_flags) 65578f860591SZhang, Yanmin { 65588f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 65598f860591SZhang, Yanmin unsigned long start = address; 65608f860591SZhang, Yanmin pte_t *ptep; 65618f860591SZhang, Yanmin pte_t pte; 6562a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 6563a79390f5SPeter Xu long pages = 0, psize = huge_page_size(h); 6564dff11abeSMike Kravetz bool shared_pmd = false; 6565ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 6566e95a9851SMike Kravetz unsigned long last_addr_mask; 65675a90d5a1SPeter Xu bool uffd_wp = cp_flags & MM_CP_UFFD_WP; 65685a90d5a1SPeter Xu bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE; 6569dff11abeSMike Kravetz 6570dff11abeSMike Kravetz /* 6571dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 6572ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 6573dff11abeSMike Kravetz * range if PMD sharing is possible. 6574dff11abeSMike Kravetz */ 65757269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 65767d4a8be0SAlistair Popple 0, mm, start, end); 6577ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 65788f860591SZhang, Yanmin 65798f860591SZhang, Yanmin BUG_ON(address >= end); 6580ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 65818f860591SZhang, Yanmin 6582ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 658340549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 658483cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 658540549ba8SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 658660dfaad6SPeter Xu for (; address < end; address += psize) { 6587cb900f41SKirill A. Shutemov spinlock_t *ptl; 65889c67a207SPeter Xu ptep = hugetlb_walk(vma, address, psize); 6589e95a9851SMike Kravetz if (!ptep) { 6590fed15f13SPeter Xu if (!uffd_wp) { 6591e95a9851SMike Kravetz address |= last_addr_mask; 65928f860591SZhang, Yanmin continue; 6593e95a9851SMike Kravetz } 6594fed15f13SPeter Xu /* 6595fed15f13SPeter Xu * Userfaultfd wr-protect requires pgtable 6596fed15f13SPeter Xu * pre-allocations to install pte markers. 6597fed15f13SPeter Xu */ 6598fed15f13SPeter Xu ptep = huge_pte_alloc(mm, vma, address, psize); 6599d1751118SPeter Xu if (!ptep) { 6600d1751118SPeter Xu pages = -ENOMEM; 6601fed15f13SPeter Xu break; 6602fed15f13SPeter Xu } 6603d1751118SPeter Xu } 6604cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 66054ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, address, ptep)) { 660660dfaad6SPeter Xu /* 660760dfaad6SPeter Xu * When uffd-wp is enabled on the vma, unshare 660860dfaad6SPeter Xu * shouldn't happen at all. Warn about it if it 660960dfaad6SPeter Xu * happened due to some reason. 661060dfaad6SPeter Xu */ 661160dfaad6SPeter Xu WARN_ON_ONCE(uffd_wp || uffd_wp_resolve); 66127da4d641SPeter Zijlstra pages++; 6613cb900f41SKirill A. Shutemov spin_unlock(ptl); 6614dff11abeSMike Kravetz shared_pmd = true; 66154ddb4d91SMike Kravetz address |= last_addr_mask; 661639dde65cSChen, Kenneth W continue; 66177da4d641SPeter Zijlstra } 6618a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 6619a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 66200e678153SDavid Hildenbrand /* Nothing to do. */ 66210e678153SDavid Hildenbrand } else if (unlikely(is_hugetlb_entry_migration(pte))) { 6622a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 66236c287605SDavid Hildenbrand struct page *page = pfn_swap_entry_to_page(entry); 662444f86392SDavid Hildenbrand pte_t newpte = pte; 6625a8bda28dSNaoya Horiguchi 662644f86392SDavid Hildenbrand if (is_writable_migration_entry(entry)) { 66276c287605SDavid Hildenbrand if (PageAnon(page)) 66286c287605SDavid Hildenbrand entry = make_readable_exclusive_migration_entry( 66296c287605SDavid Hildenbrand swp_offset(entry)); 66306c287605SDavid Hildenbrand else 66314dd845b5SAlistair Popple entry = make_readable_migration_entry( 66324dd845b5SAlistair Popple swp_offset(entry)); 6633a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 663444f86392SDavid Hildenbrand pages++; 663544f86392SDavid Hildenbrand } 663644f86392SDavid Hildenbrand 66375a90d5a1SPeter Xu if (uffd_wp) 66385a90d5a1SPeter Xu newpte = pte_swp_mkuffd_wp(newpte); 66395a90d5a1SPeter Xu else if (uffd_wp_resolve) 66405a90d5a1SPeter Xu newpte = pte_swp_clear_uffd_wp(newpte); 664144f86392SDavid Hildenbrand if (!pte_same(pte, newpte)) 664218f39629SQi Zheng set_huge_pte_at(mm, address, ptep, newpte); 66430e678153SDavid Hildenbrand } else if (unlikely(is_pte_marker(pte))) { 66440e678153SDavid Hildenbrand /* No other markers apply for now. */ 66450e678153SDavid Hildenbrand WARN_ON_ONCE(!pte_marker_uffd_wp(pte)); 664660dfaad6SPeter Xu if (uffd_wp_resolve) 66470e678153SDavid Hildenbrand /* Safe to modify directly (non-present->none). */ 664860dfaad6SPeter Xu huge_pte_clear(mm, address, ptep, psize); 66490e678153SDavid Hildenbrand } else if (!huge_pte_none(pte)) { 6650023bdd00SAneesh Kumar K.V pte_t old_pte; 665179c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 6652023bdd00SAneesh Kumar K.V 6653023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 665416785bd7SAnshuman Khandual pte = huge_pte_modify(old_pte, newprot); 665579c1c594SChristophe Leroy pte = arch_make_huge_pte(pte, shift, vma->vm_flags); 66565a90d5a1SPeter Xu if (uffd_wp) 6657f1eb1bacSPeter Xu pte = huge_pte_mkuffd_wp(pte); 66585a90d5a1SPeter Xu else if (uffd_wp_resolve) 66595a90d5a1SPeter Xu pte = huge_pte_clear_uffd_wp(pte); 6660023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 66617da4d641SPeter Zijlstra pages++; 666260dfaad6SPeter Xu } else { 666360dfaad6SPeter Xu /* None pte */ 666460dfaad6SPeter Xu if (unlikely(uffd_wp)) 666560dfaad6SPeter Xu /* Safe to modify directly (none->non-present). */ 666660dfaad6SPeter Xu set_huge_pte_at(mm, address, ptep, 666760dfaad6SPeter Xu make_pte_marker(PTE_MARKER_UFFD_WP)); 66688f860591SZhang, Yanmin } 6669cb900f41SKirill A. Shutemov spin_unlock(ptl); 66708f860591SZhang, Yanmin } 6671d833352aSMel Gorman /* 6672c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 6673d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 6674c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 6675dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 6676dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 6677d833352aSMel Gorman */ 6678dff11abeSMike Kravetz if (shared_pmd) 6679ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 6680dff11abeSMike Kravetz else 66815491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 66820f10851eSJérôme Glisse /* 66831af5a810SAlistair Popple * No need to call mmu_notifier_arch_invalidate_secondary_tlbs() we are 66841af5a810SAlistair Popple * downgrading page table protection not changing it to point to a new 66851af5a810SAlistair Popple * page. 66860f10851eSJérôme Glisse * 6687ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 66880f10851eSJérôme Glisse */ 668983cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 669040549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 6691ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 66927da4d641SPeter Zijlstra 6693d1751118SPeter Xu return pages > 0 ? (pages << h->order) : pages; 66948f860591SZhang, Yanmin } 66958f860591SZhang, Yanmin 669633b8f84aSMike Kravetz /* Return true if reservation was successful, false otherwise. */ 669733b8f84aSMike Kravetz bool hugetlb_reserve_pages(struct inode *inode, 6698a1e78772SMel Gorman long from, long to, 66995a6fe125SMel Gorman struct vm_area_struct *vma, 6700ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 6701e4e574b7SAdam Litke { 6702c5094ec7SMike Kravetz long chg = -1, add = -1; 6703a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 670490481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 67059119a41eSJoonsoo Kim struct resv_map *resv_map; 6706075a61d0SMina Almasry struct hugetlb_cgroup *h_cg = NULL; 67070db9d74eSMina Almasry long gbl_reserve, regions_needed = 0; 6708e4e574b7SAdam Litke 670963489f8eSMike Kravetz /* This should never happen */ 671063489f8eSMike Kravetz if (from > to) { 671163489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 671233b8f84aSMike Kravetz return false; 671363489f8eSMike Kravetz } 671463489f8eSMike Kravetz 6715a1e78772SMel Gorman /* 6716e700898fSMike Kravetz * vma specific semaphore used for pmd sharing and fault/truncation 6717e700898fSMike Kravetz * synchronization 67188d9bfb26SMike Kravetz */ 67198d9bfb26SMike Kravetz hugetlb_vma_lock_alloc(vma); 67208d9bfb26SMike Kravetz 67218d9bfb26SMike Kravetz /* 672217c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 672317c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 672490481622SDavid Gibson * without using reserves 672517c9d12eSMel Gorman */ 6726ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 672733b8f84aSMike Kravetz return true; 672817c9d12eSMel Gorman 672917c9d12eSMel Gorman /* 6730a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 6731a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 6732a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 6733a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 6734a1e78772SMel Gorman */ 67359119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 6736f27a5136SMike Kravetz /* 6737f27a5136SMike Kravetz * resv_map can not be NULL as hugetlb_reserve_pages is only 6738f27a5136SMike Kravetz * called for inodes for which resv_maps were created (see 6739f27a5136SMike Kravetz * hugetlbfs_get_inode). 6740f27a5136SMike Kravetz */ 67414e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 67429119a41eSJoonsoo Kim 67430db9d74eSMina Almasry chg = region_chg(resv_map, from, to, ®ions_needed); 67449119a41eSJoonsoo Kim } else { 6745e9fe92aeSMina Almasry /* Private mapping. */ 67469119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 67475a6fe125SMel Gorman if (!resv_map) 67488d9bfb26SMike Kravetz goto out_err; 67495a6fe125SMel Gorman 675017c9d12eSMel Gorman chg = to - from; 675117c9d12eSMel Gorman 67525a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 67535a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 67545a6fe125SMel Gorman } 67555a6fe125SMel Gorman 675633b8f84aSMike Kravetz if (chg < 0) 6757c50ac050SDave Hansen goto out_err; 675817c9d12eSMel Gorman 675933b8f84aSMike Kravetz if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), 676033b8f84aSMike Kravetz chg * pages_per_huge_page(h), &h_cg) < 0) 6761075a61d0SMina Almasry goto out_err; 6762075a61d0SMina Almasry 6763075a61d0SMina Almasry if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { 6764075a61d0SMina Almasry /* For private mappings, the hugetlb_cgroup uncharge info hangs 6765075a61d0SMina Almasry * of the resv_map. 6766075a61d0SMina Almasry */ 6767075a61d0SMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); 6768075a61d0SMina Almasry } 6769075a61d0SMina Almasry 67701c5ecae3SMike Kravetz /* 67711c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 67721c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 67731c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 67741c5ecae3SMike Kravetz */ 67751c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 677633b8f84aSMike Kravetz if (gbl_reserve < 0) 6777075a61d0SMina Almasry goto out_uncharge_cgroup; 677817c9d12eSMel Gorman 677917c9d12eSMel Gorman /* 678017c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 678190481622SDavid Gibson * Hand the pages back to the subpool if there are not 678217c9d12eSMel Gorman */ 678333b8f84aSMike Kravetz if (hugetlb_acct_memory(h, gbl_reserve) < 0) 6784075a61d0SMina Almasry goto out_put_pages; 678517c9d12eSMel Gorman 678617c9d12eSMel Gorman /* 678717c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 678817c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 678917c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 679017c9d12eSMel Gorman * the reservation was and the page cache tells how much of 679117c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 679217c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 679317c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 679417c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 679517c9d12eSMel Gorman * else has to be done for private mappings here 679617c9d12eSMel Gorman */ 679733039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 6798075a61d0SMina Almasry add = region_add(resv_map, from, to, regions_needed, h, h_cg); 679933039678SMike Kravetz 68000db9d74eSMina Almasry if (unlikely(add < 0)) { 68010db9d74eSMina Almasry hugetlb_acct_memory(h, -gbl_reserve); 6802075a61d0SMina Almasry goto out_put_pages; 68030db9d74eSMina Almasry } else if (unlikely(chg > add)) { 680433039678SMike Kravetz /* 680533039678SMike Kravetz * pages in this range were added to the reserve 680633039678SMike Kravetz * map between region_chg and region_add. This 6807d0ce0e47SSidhartha Kumar * indicates a race with alloc_hugetlb_folio. Adjust 680833039678SMike Kravetz * the subpool and reserve counts modified above 680933039678SMike Kravetz * based on the difference. 681033039678SMike Kravetz */ 681133039678SMike Kravetz long rsv_adjust; 681233039678SMike Kravetz 6813d85aecf2SMiaohe Lin /* 6814d85aecf2SMiaohe Lin * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the 6815d85aecf2SMiaohe Lin * reference to h_cg->css. See comment below for detail. 6816d85aecf2SMiaohe Lin */ 6817075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd( 6818075a61d0SMina Almasry hstate_index(h), 6819075a61d0SMina Almasry (chg - add) * pages_per_huge_page(h), h_cg); 6820075a61d0SMina Almasry 682133039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 682233039678SMike Kravetz chg - add); 682333039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 6824d85aecf2SMiaohe Lin } else if (h_cg) { 6825d85aecf2SMiaohe Lin /* 6826d85aecf2SMiaohe Lin * The file_regions will hold their own reference to 6827d85aecf2SMiaohe Lin * h_cg->css. So we should release the reference held 6828d85aecf2SMiaohe Lin * via hugetlb_cgroup_charge_cgroup_rsvd() when we are 6829d85aecf2SMiaohe Lin * done. 6830d85aecf2SMiaohe Lin */ 6831d85aecf2SMiaohe Lin hugetlb_cgroup_put_rsvd_cgroup(h_cg); 683233039678SMike Kravetz } 683333039678SMike Kravetz } 683433b8f84aSMike Kravetz return true; 683533b8f84aSMike Kravetz 6836075a61d0SMina Almasry out_put_pages: 6837075a61d0SMina Almasry /* put back original number of pages, chg */ 6838075a61d0SMina Almasry (void)hugepage_subpool_put_pages(spool, chg); 6839075a61d0SMina Almasry out_uncharge_cgroup: 6840075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), 6841075a61d0SMina Almasry chg * pages_per_huge_page(h), h_cg); 6842c50ac050SDave Hansen out_err: 68438d9bfb26SMike Kravetz hugetlb_vma_lock_free(vma); 68445e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 68450db9d74eSMina Almasry /* Only call region_abort if the region_chg succeeded but the 68460db9d74eSMina Almasry * region_add failed or didn't run. 68470db9d74eSMina Almasry */ 68480db9d74eSMina Almasry if (chg >= 0 && add < 0) 68490db9d74eSMina Almasry region_abort(resv_map, from, to, regions_needed); 6850f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 6851f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 685233b8f84aSMike Kravetz return false; 6853a43a8c39SChen, Kenneth W } 6854a43a8c39SChen, Kenneth W 6855b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 6856b5cec28dSMike Kravetz long freed) 6857a43a8c39SChen, Kenneth W { 6858a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 68594e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 68609119a41eSJoonsoo Kim long chg = 0; 686190481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 68621c5ecae3SMike Kravetz long gbl_reserve; 686345c682a6SKen Chen 6864f27a5136SMike Kravetz /* 6865f27a5136SMike Kravetz * Since this routine can be called in the evict inode path for all 6866f27a5136SMike Kravetz * hugetlbfs inodes, resv_map could be NULL. 6867f27a5136SMike Kravetz */ 6868b5cec28dSMike Kravetz if (resv_map) { 6869b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 6870b5cec28dSMike Kravetz /* 6871b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 6872b5cec28dSMike Kravetz * must be split and another region descriptor can not be 6873b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 6874b5cec28dSMike Kravetz */ 6875b5cec28dSMike Kravetz if (chg < 0) 6876b5cec28dSMike Kravetz return chg; 6877b5cec28dSMike Kravetz } 6878b5cec28dSMike Kravetz 687945c682a6SKen Chen spin_lock(&inode->i_lock); 6880e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 688145c682a6SKen Chen spin_unlock(&inode->i_lock); 688245c682a6SKen Chen 68831c5ecae3SMike Kravetz /* 68841c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 68851c5ecae3SMike Kravetz * reservations to be released may be adjusted. 6886dddf31a4SMiaohe Lin * 6887dddf31a4SMiaohe Lin * Note that !resv_map implies freed == 0. So (chg - freed) 6888dddf31a4SMiaohe Lin * won't go negative. 68891c5ecae3SMike Kravetz */ 68901c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 68911c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 6892b5cec28dSMike Kravetz 6893b5cec28dSMike Kravetz return 0; 6894a43a8c39SChen, Kenneth W } 689593f70f90SNaoya Horiguchi 68963212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 68973212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 68983212b535SSteve Capper struct vm_area_struct *vma, 68993212b535SSteve Capper unsigned long addr, pgoff_t idx) 69003212b535SSteve Capper { 69013212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 69023212b535SSteve Capper svma->vm_start; 69033212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 69043212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 69053212b535SSteve Capper 69063212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 6907e430a95aSSuren Baghdasaryan unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED_MASK; 6908e430a95aSSuren Baghdasaryan unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED_MASK; 69093212b535SSteve Capper 69103212b535SSteve Capper /* 69113212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 69123212b535SSteve Capper * page table page. 6913131a79b4SMike Kravetz * 6914131a79b4SMike Kravetz * Also, vma_lock (vm_private_data) is required for sharing. 69153212b535SSteve Capper */ 69163212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 69173212b535SSteve Capper vm_flags != svm_flags || 6918131a79b4SMike Kravetz !range_in_vma(svma, sbase, s_end) || 6919131a79b4SMike Kravetz !svma->vm_private_data) 69203212b535SSteve Capper return 0; 69213212b535SSteve Capper 69223212b535SSteve Capper return saddr; 69233212b535SSteve Capper } 69243212b535SSteve Capper 6925bbff39ccSMike Kravetz bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 69263212b535SSteve Capper { 6927bbff39ccSMike Kravetz unsigned long start = addr & PUD_MASK; 6928bbff39ccSMike Kravetz unsigned long end = start + PUD_SIZE; 6929bbff39ccSMike Kravetz 69308d9bfb26SMike Kravetz #ifdef CONFIG_USERFAULTFD 69318d9bfb26SMike Kravetz if (uffd_disable_huge_pmd_share(vma)) 69328d9bfb26SMike Kravetz return false; 69338d9bfb26SMike Kravetz #endif 69343212b535SSteve Capper /* 69353212b535SSteve Capper * check on proper vm_flags and page table alignment 69363212b535SSteve Capper */ 69378d9bfb26SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 693831aafb45SNicholas Krause return false; 6939bbff39ccSMike Kravetz if (!vma->vm_private_data) /* vma lock required for sharing */ 69408d9bfb26SMike Kravetz return false; 69418d9bfb26SMike Kravetz if (!range_in_vma(vma, start, end)) 69428d9bfb26SMike Kravetz return false; 69438d9bfb26SMike Kravetz return true; 69448d9bfb26SMike Kravetz } 69458d9bfb26SMike Kravetz 69463212b535SSteve Capper /* 6947017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 6948017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 6949017b1660SMike Kravetz * shared pmd mappings. 6950017b1660SMike Kravetz */ 6951017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 6952017b1660SMike Kravetz unsigned long *start, unsigned long *end) 6953017b1660SMike Kravetz { 6954a1ba9da8SLi Xinhai unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), 6955a1ba9da8SLi Xinhai v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 6956017b1660SMike Kravetz 6957a1ba9da8SLi Xinhai /* 6958f0953a1bSIngo Molnar * vma needs to span at least one aligned PUD size, and the range 6959f0953a1bSIngo Molnar * must be at least partially within in. 6960a1ba9da8SLi Xinhai */ 6961a1ba9da8SLi Xinhai if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || 6962a1ba9da8SLi Xinhai (*end <= v_start) || (*start >= v_end)) 6963017b1660SMike Kravetz return; 6964017b1660SMike Kravetz 696575802ca6SPeter Xu /* Extend the range to be PUD aligned for a worst case scenario */ 6966a1ba9da8SLi Xinhai if (*start > v_start) 6967a1ba9da8SLi Xinhai *start = ALIGN_DOWN(*start, PUD_SIZE); 6968017b1660SMike Kravetz 6969a1ba9da8SLi Xinhai if (*end < v_end) 6970a1ba9da8SLi Xinhai *end = ALIGN(*end, PUD_SIZE); 6971017b1660SMike Kravetz } 6972017b1660SMike Kravetz 6973017b1660SMike Kravetz /* 69743212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 69753212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 69763212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 69773a47c54fSMike Kravetz * code much cleaner. pmd allocation is essential for the shared case because 69783a47c54fSMike Kravetz * pud has to be populated inside the same i_mmap_rwsem section - otherwise 69793a47c54fSMike Kravetz * racing tasks could either miss the sharing (see huge_pte_offset) or select a 69803a47c54fSMike Kravetz * bad pmd for sharing. 69813212b535SSteve Capper */ 6982aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 6983aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 69843212b535SSteve Capper { 69853212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 69863212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 69873212b535SSteve Capper vma->vm_pgoff; 69883212b535SSteve Capper struct vm_area_struct *svma; 69893212b535SSteve Capper unsigned long saddr; 69903212b535SSteve Capper pte_t *spte = NULL; 69913212b535SSteve Capper pte_t *pte; 69923212b535SSteve Capper 69933a47c54fSMike Kravetz i_mmap_lock_read(mapping); 69943212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 69953212b535SSteve Capper if (svma == vma) 69963212b535SSteve Capper continue; 69973212b535SSteve Capper 69983212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 69993212b535SSteve Capper if (saddr) { 70009c67a207SPeter Xu spte = hugetlb_walk(svma, saddr, 70017868a208SPunit Agrawal vma_mmu_pagesize(svma)); 70023212b535SSteve Capper if (spte) { 70033212b535SSteve Capper get_page(virt_to_page(spte)); 70043212b535SSteve Capper break; 70053212b535SSteve Capper } 70063212b535SSteve Capper } 70073212b535SSteve Capper } 70083212b535SSteve Capper 70093212b535SSteve Capper if (!spte) 70103212b535SSteve Capper goto out; 70113212b535SSteve Capper 7012349d1670SPeter Xu spin_lock(&mm->page_table_lock); 7013dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 70143212b535SSteve Capper pud_populate(mm, pud, 70153212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 7016c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 7017dc6c9a35SKirill A. Shutemov } else { 70183212b535SSteve Capper put_page(virt_to_page(spte)); 7019dc6c9a35SKirill A. Shutemov } 7020349d1670SPeter Xu spin_unlock(&mm->page_table_lock); 70213212b535SSteve Capper out: 70223212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 70233a47c54fSMike Kravetz i_mmap_unlock_read(mapping); 70243212b535SSteve Capper return pte; 70253212b535SSteve Capper } 70263212b535SSteve Capper 70273212b535SSteve Capper /* 70283212b535SSteve Capper * unmap huge page backed by shared pte. 70293212b535SSteve Capper * 70303212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 70313212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 70323212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 70333212b535SSteve Capper * 70343a47c54fSMike Kravetz * Called with page table lock held. 70353212b535SSteve Capper * 70363212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 70373212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 70383212b535SSteve Capper */ 703934ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 70404ddb4d91SMike Kravetz unsigned long addr, pte_t *ptep) 70413212b535SSteve Capper { 70424ddb4d91SMike Kravetz pgd_t *pgd = pgd_offset(mm, addr); 70434ddb4d91SMike Kravetz p4d_t *p4d = p4d_offset(pgd, addr); 70444ddb4d91SMike Kravetz pud_t *pud = pud_offset(p4d, addr); 70453212b535SSteve Capper 704634ae204fSMike Kravetz i_mmap_assert_write_locked(vma->vm_file->f_mapping); 704740549ba8SMike Kravetz hugetlb_vma_assert_locked(vma); 70483212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 70493212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 70503212b535SSteve Capper return 0; 70513212b535SSteve Capper 70523212b535SSteve Capper pud_clear(pud); 70533212b535SSteve Capper put_page(virt_to_page(ptep)); 7054dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 70553212b535SSteve Capper return 1; 70563212b535SSteve Capper } 7057c1991e07SPeter Xu 70589e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 70598d9bfb26SMike Kravetz 7060aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 7061aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 70629e5fc74cSSteve Capper { 70639e5fc74cSSteve Capper return NULL; 70649e5fc74cSSteve Capper } 7065e81f2d22SZhang Zhen 706634ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 70674ddb4d91SMike Kravetz unsigned long addr, pte_t *ptep) 7068e81f2d22SZhang Zhen { 7069e81f2d22SZhang Zhen return 0; 7070e81f2d22SZhang Zhen } 7071017b1660SMike Kravetz 7072017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 7073017b1660SMike Kravetz unsigned long *start, unsigned long *end) 7074017b1660SMike Kravetz { 7075017b1660SMike Kravetz } 7076c1991e07SPeter Xu 7077c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 7078c1991e07SPeter Xu { 7079c1991e07SPeter Xu return false; 7080c1991e07SPeter Xu } 70813212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 70823212b535SSteve Capper 70839e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 7084aec44e0fSPeter Xu pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, 70859e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 70869e5fc74cSSteve Capper { 70879e5fc74cSSteve Capper pgd_t *pgd; 7088c2febafcSKirill A. Shutemov p4d_t *p4d; 70899e5fc74cSSteve Capper pud_t *pud; 70909e5fc74cSSteve Capper pte_t *pte = NULL; 70919e5fc74cSSteve Capper 70929e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 7093f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 7094f4f0a3d8SKirill A. Shutemov if (!p4d) 7095f4f0a3d8SKirill A. Shutemov return NULL; 7096c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 70979e5fc74cSSteve Capper if (pud) { 70989e5fc74cSSteve Capper if (sz == PUD_SIZE) { 70999e5fc74cSSteve Capper pte = (pte_t *)pud; 71009e5fc74cSSteve Capper } else { 71019e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 7102c1991e07SPeter Xu if (want_pmd_share(vma, addr) && pud_none(*pud)) 7103aec44e0fSPeter Xu pte = huge_pmd_share(mm, vma, addr, pud); 71049e5fc74cSSteve Capper else 71059e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 71069e5fc74cSSteve Capper } 71079e5fc74cSSteve Capper } 7108191fcdb6SJohn Hubbard 7109191fcdb6SJohn Hubbard if (pte) { 7110191fcdb6SJohn Hubbard pte_t pteval = ptep_get_lockless(pte); 7111191fcdb6SJohn Hubbard 7112191fcdb6SJohn Hubbard BUG_ON(pte_present(pteval) && !pte_huge(pteval)); 7113191fcdb6SJohn Hubbard } 71149e5fc74cSSteve Capper 71159e5fc74cSSteve Capper return pte; 71169e5fc74cSSteve Capper } 71179e5fc74cSSteve Capper 71189b19df29SPunit Agrawal /* 71199b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 71209b19df29SPunit Agrawal * entry at address @addr 71219b19df29SPunit Agrawal * 71228ac0b81aSLi Xinhai * Return: Pointer to page table entry (PUD or PMD) for 71238ac0b81aSLi Xinhai * address @addr, or NULL if a !p*d_present() entry is encountered and the 71249b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 71259b19df29SPunit Agrawal * table. 71269b19df29SPunit Agrawal */ 71277868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 71287868a208SPunit Agrawal unsigned long addr, unsigned long sz) 71299e5fc74cSSteve Capper { 71309e5fc74cSSteve Capper pgd_t *pgd; 7131c2febafcSKirill A. Shutemov p4d_t *p4d; 71328ac0b81aSLi Xinhai pud_t *pud; 71338ac0b81aSLi Xinhai pmd_t *pmd; 71349e5fc74cSSteve Capper 71359e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 7136c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 7137c2febafcSKirill A. Shutemov return NULL; 7138c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 7139c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 7140c2febafcSKirill A. Shutemov return NULL; 71419b19df29SPunit Agrawal 7142c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 71438ac0b81aSLi Xinhai if (sz == PUD_SIZE) 71448ac0b81aSLi Xinhai /* must be pud huge, non-present or none */ 71459e5fc74cSSteve Capper return (pte_t *)pud; 71468ac0b81aSLi Xinhai if (!pud_present(*pud)) 71478ac0b81aSLi Xinhai return NULL; 71488ac0b81aSLi Xinhai /* must have a valid entry and size to go further */ 71499b19df29SPunit Agrawal 71509e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 71518ac0b81aSLi Xinhai /* must be pmd huge, non-present or none */ 71529e5fc74cSSteve Capper return (pte_t *)pmd; 71539e5fc74cSSteve Capper } 71549e5fc74cSSteve Capper 7155e95a9851SMike Kravetz /* 7156e95a9851SMike Kravetz * Return a mask that can be used to update an address to the last huge 7157e95a9851SMike Kravetz * page in a page table page mapping size. Used to skip non-present 7158e95a9851SMike Kravetz * page table entries when linearly scanning address ranges. Architectures 7159e95a9851SMike Kravetz * with unique huge page to page table relationships can define their own 7160e95a9851SMike Kravetz * version of this routine. 7161e95a9851SMike Kravetz */ 7162e95a9851SMike Kravetz unsigned long hugetlb_mask_last_page(struct hstate *h) 7163e95a9851SMike Kravetz { 7164e95a9851SMike Kravetz unsigned long hp_size = huge_page_size(h); 7165e95a9851SMike Kravetz 7166e95a9851SMike Kravetz if (hp_size == PUD_SIZE) 7167e95a9851SMike Kravetz return P4D_SIZE - PUD_SIZE; 7168e95a9851SMike Kravetz else if (hp_size == PMD_SIZE) 7169e95a9851SMike Kravetz return PUD_SIZE - PMD_SIZE; 7170e95a9851SMike Kravetz else 7171e95a9851SMike Kravetz return 0UL; 7172e95a9851SMike Kravetz } 7173e95a9851SMike Kravetz 7174e95a9851SMike Kravetz #else 7175e95a9851SMike Kravetz 7176e95a9851SMike Kravetz /* See description above. Architectures can provide their own version. */ 7177e95a9851SMike Kravetz __weak unsigned long hugetlb_mask_last_page(struct hstate *h) 7178e95a9851SMike Kravetz { 71794ddb4d91SMike Kravetz #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 71804ddb4d91SMike Kravetz if (huge_page_size(h) == PMD_SIZE) 71814ddb4d91SMike Kravetz return PUD_SIZE - PMD_SIZE; 71824ddb4d91SMike Kravetz #endif 7183e95a9851SMike Kravetz return 0UL; 7184e95a9851SMike Kravetz } 7185e95a9851SMike Kravetz 718661f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 718761f77edaSNaoya Horiguchi 718861f77edaSNaoya Horiguchi /* 718961f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 719061f77edaSNaoya Horiguchi * behavior. 719161f77edaSNaoya Horiguchi */ 71929747b9e9SBaolin Wang bool isolate_hugetlb(struct folio *folio, struct list_head *list) 719331caf665SNaoya Horiguchi { 71949747b9e9SBaolin Wang bool ret = true; 7195bcc54222SNaoya Horiguchi 7196db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 71976aa3a920SSidhartha Kumar if (!folio_test_hugetlb(folio) || 71986aa3a920SSidhartha Kumar !folio_test_hugetlb_migratable(folio) || 71996aa3a920SSidhartha Kumar !folio_try_get(folio)) { 72009747b9e9SBaolin Wang ret = false; 7201bcc54222SNaoya Horiguchi goto unlock; 7202bcc54222SNaoya Horiguchi } 72036aa3a920SSidhartha Kumar folio_clear_hugetlb_migratable(folio); 72046aa3a920SSidhartha Kumar list_move_tail(&folio->lru, list); 7205bcc54222SNaoya Horiguchi unlock: 7206db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7207bcc54222SNaoya Horiguchi return ret; 720831caf665SNaoya Horiguchi } 720931caf665SNaoya Horiguchi 721004bac040SSidhartha Kumar int get_hwpoison_hugetlb_folio(struct folio *folio, bool *hugetlb, bool unpoison) 721125182f05SNaoya Horiguchi { 721225182f05SNaoya Horiguchi int ret = 0; 721325182f05SNaoya Horiguchi 721425182f05SNaoya Horiguchi *hugetlb = false; 721525182f05SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 721604bac040SSidhartha Kumar if (folio_test_hugetlb(folio)) { 721725182f05SNaoya Horiguchi *hugetlb = true; 721804bac040SSidhartha Kumar if (folio_test_hugetlb_freed(folio)) 7219b283d983SNaoya Horiguchi ret = 0; 722004bac040SSidhartha Kumar else if (folio_test_hugetlb_migratable(folio) || unpoison) 722104bac040SSidhartha Kumar ret = folio_try_get(folio); 72220ed950d1SNaoya Horiguchi else 72230ed950d1SNaoya Horiguchi ret = -EBUSY; 722425182f05SNaoya Horiguchi } 722525182f05SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 722625182f05SNaoya Horiguchi return ret; 722725182f05SNaoya Horiguchi } 722825182f05SNaoya Horiguchi 7229e591ef7dSNaoya Horiguchi int get_huge_page_for_hwpoison(unsigned long pfn, int flags, 7230e591ef7dSNaoya Horiguchi bool *migratable_cleared) 7231405ce051SNaoya Horiguchi { 7232405ce051SNaoya Horiguchi int ret; 7233405ce051SNaoya Horiguchi 7234405ce051SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 7235e591ef7dSNaoya Horiguchi ret = __get_huge_page_for_hwpoison(pfn, flags, migratable_cleared); 7236405ce051SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 7237405ce051SNaoya Horiguchi return ret; 7238405ce051SNaoya Horiguchi } 7239405ce051SNaoya Horiguchi 7240ea8e72f4SSidhartha Kumar void folio_putback_active_hugetlb(struct folio *folio) 724131caf665SNaoya Horiguchi { 7242db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 7243ea8e72f4SSidhartha Kumar folio_set_hugetlb_migratable(folio); 7244ea8e72f4SSidhartha Kumar list_move_tail(&folio->lru, &(folio_hstate(folio))->hugepage_activelist); 7245db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7246ea8e72f4SSidhartha Kumar folio_put(folio); 724731caf665SNaoya Horiguchi } 7248ab5ac90aSMichal Hocko 7249345c62d1SSidhartha Kumar void move_hugetlb_state(struct folio *old_folio, struct folio *new_folio, int reason) 7250ab5ac90aSMichal Hocko { 7251345c62d1SSidhartha Kumar struct hstate *h = folio_hstate(old_folio); 7252ab5ac90aSMichal Hocko 7253345c62d1SSidhartha Kumar hugetlb_cgroup_migrate(old_folio, new_folio); 7254345c62d1SSidhartha Kumar set_page_owner_migrate_reason(&new_folio->page, reason); 7255ab5ac90aSMichal Hocko 7256ab5ac90aSMichal Hocko /* 7257345c62d1SSidhartha Kumar * transfer temporary state of the new hugetlb folio. This is 7258ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 7259ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 7260ab5ac90aSMichal Hocko * the temporary status. 7261ab5ac90aSMichal Hocko * 7262ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 7263ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 7264ab5ac90aSMichal Hocko * the per-node's. 7265ab5ac90aSMichal Hocko */ 7266345c62d1SSidhartha Kumar if (folio_test_hugetlb_temporary(new_folio)) { 7267345c62d1SSidhartha Kumar int old_nid = folio_nid(old_folio); 7268345c62d1SSidhartha Kumar int new_nid = folio_nid(new_folio); 7269ab5ac90aSMichal Hocko 7270345c62d1SSidhartha Kumar folio_set_hugetlb_temporary(old_folio); 7271345c62d1SSidhartha Kumar folio_clear_hugetlb_temporary(new_folio); 7272345c62d1SSidhartha Kumar 7273ab5ac90aSMichal Hocko 72745af1ab1dSMiaohe Lin /* 72755af1ab1dSMiaohe Lin * There is no need to transfer the per-node surplus state 72765af1ab1dSMiaohe Lin * when we do not cross the node. 72775af1ab1dSMiaohe Lin */ 72785af1ab1dSMiaohe Lin if (new_nid == old_nid) 72795af1ab1dSMiaohe Lin return; 7280db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 7281ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 7282ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 7283ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 7284ab5ac90aSMichal Hocko } 7285db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7286ab5ac90aSMichal Hocko } 7287ab5ac90aSMichal Hocko } 7288cf11e85fSRoman Gushchin 7289b30c14cdSJames Houghton static void hugetlb_unshare_pmds(struct vm_area_struct *vma, 7290b30c14cdSJames Houghton unsigned long start, 7291b30c14cdSJames Houghton unsigned long end) 72926dfeaff9SPeter Xu { 72936dfeaff9SPeter Xu struct hstate *h = hstate_vma(vma); 72946dfeaff9SPeter Xu unsigned long sz = huge_page_size(h); 72956dfeaff9SPeter Xu struct mm_struct *mm = vma->vm_mm; 72966dfeaff9SPeter Xu struct mmu_notifier_range range; 7297b30c14cdSJames Houghton unsigned long address; 72986dfeaff9SPeter Xu spinlock_t *ptl; 72996dfeaff9SPeter Xu pte_t *ptep; 73006dfeaff9SPeter Xu 73016dfeaff9SPeter Xu if (!(vma->vm_flags & VM_MAYSHARE)) 73026dfeaff9SPeter Xu return; 73036dfeaff9SPeter Xu 73046dfeaff9SPeter Xu if (start >= end) 73056dfeaff9SPeter Xu return; 73066dfeaff9SPeter Xu 73079c8bbfacSBaolin Wang flush_cache_range(vma, start, end); 73086dfeaff9SPeter Xu /* 73096dfeaff9SPeter Xu * No need to call adjust_range_if_pmd_sharing_possible(), because 73106dfeaff9SPeter Xu * we have already done the PUD_SIZE alignment. 73116dfeaff9SPeter Xu */ 73127d4a8be0SAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, 73136dfeaff9SPeter Xu start, end); 73146dfeaff9SPeter Xu mmu_notifier_invalidate_range_start(&range); 731540549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 73166dfeaff9SPeter Xu i_mmap_lock_write(vma->vm_file->f_mapping); 73176dfeaff9SPeter Xu for (address = start; address < end; address += PUD_SIZE) { 73189c67a207SPeter Xu ptep = hugetlb_walk(vma, address, sz); 73196dfeaff9SPeter Xu if (!ptep) 73206dfeaff9SPeter Xu continue; 73216dfeaff9SPeter Xu ptl = huge_pte_lock(h, mm, ptep); 73224ddb4d91SMike Kravetz huge_pmd_unshare(mm, vma, address, ptep); 73236dfeaff9SPeter Xu spin_unlock(ptl); 73246dfeaff9SPeter Xu } 73256dfeaff9SPeter Xu flush_hugetlb_tlb_range(vma, start, end); 73266dfeaff9SPeter Xu i_mmap_unlock_write(vma->vm_file->f_mapping); 732740549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 73286dfeaff9SPeter Xu /* 73291af5a810SAlistair Popple * No need to call mmu_notifier_arch_invalidate_secondary_tlbs(), see 7330ee65728eSMike Rapoport * Documentation/mm/mmu_notifier.rst. 73316dfeaff9SPeter Xu */ 73326dfeaff9SPeter Xu mmu_notifier_invalidate_range_end(&range); 73336dfeaff9SPeter Xu } 73346dfeaff9SPeter Xu 7335b30c14cdSJames Houghton /* 7336b30c14cdSJames Houghton * This function will unconditionally remove all the shared pmd pgtable entries 7337b30c14cdSJames Houghton * within the specific vma for a hugetlbfs memory range. 7338b30c14cdSJames Houghton */ 7339b30c14cdSJames Houghton void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) 7340b30c14cdSJames Houghton { 7341b30c14cdSJames Houghton hugetlb_unshare_pmds(vma, ALIGN(vma->vm_start, PUD_SIZE), 7342b30c14cdSJames Houghton ALIGN_DOWN(vma->vm_end, PUD_SIZE)); 7343b30c14cdSJames Houghton } 7344b30c14cdSJames Houghton 7345cf11e85fSRoman Gushchin #ifdef CONFIG_CMA 7346cf11e85fSRoman Gushchin static bool cma_reserve_called __initdata; 7347cf11e85fSRoman Gushchin 7348cf11e85fSRoman Gushchin static int __init cmdline_parse_hugetlb_cma(char *p) 7349cf11e85fSRoman Gushchin { 735038e719abSBaolin Wang int nid, count = 0; 735138e719abSBaolin Wang unsigned long tmp; 735238e719abSBaolin Wang char *s = p; 735338e719abSBaolin Wang 735438e719abSBaolin Wang while (*s) { 735538e719abSBaolin Wang if (sscanf(s, "%lu%n", &tmp, &count) != 1) 735638e719abSBaolin Wang break; 735738e719abSBaolin Wang 735838e719abSBaolin Wang if (s[count] == ':') { 7359f9317f77SMike Kravetz if (tmp >= MAX_NUMNODES) 736038e719abSBaolin Wang break; 7361f9317f77SMike Kravetz nid = array_index_nospec(tmp, MAX_NUMNODES); 736238e719abSBaolin Wang 736338e719abSBaolin Wang s += count + 1; 736438e719abSBaolin Wang tmp = memparse(s, &s); 736538e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = tmp; 736638e719abSBaolin Wang hugetlb_cma_size += tmp; 736738e719abSBaolin Wang 736838e719abSBaolin Wang /* 736938e719abSBaolin Wang * Skip the separator if have one, otherwise 737038e719abSBaolin Wang * break the parsing. 737138e719abSBaolin Wang */ 737238e719abSBaolin Wang if (*s == ',') 737338e719abSBaolin Wang s++; 737438e719abSBaolin Wang else 737538e719abSBaolin Wang break; 737638e719abSBaolin Wang } else { 7377cf11e85fSRoman Gushchin hugetlb_cma_size = memparse(p, &p); 737838e719abSBaolin Wang break; 737938e719abSBaolin Wang } 738038e719abSBaolin Wang } 738138e719abSBaolin Wang 7382cf11e85fSRoman Gushchin return 0; 7383cf11e85fSRoman Gushchin } 7384cf11e85fSRoman Gushchin 7385cf11e85fSRoman Gushchin early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); 7386cf11e85fSRoman Gushchin 7387cf11e85fSRoman Gushchin void __init hugetlb_cma_reserve(int order) 7388cf11e85fSRoman Gushchin { 7389cf11e85fSRoman Gushchin unsigned long size, reserved, per_node; 739038e719abSBaolin Wang bool node_specific_cma_alloc = false; 7391cf11e85fSRoman Gushchin int nid; 7392cf11e85fSRoman Gushchin 7393cf11e85fSRoman Gushchin cma_reserve_called = true; 7394cf11e85fSRoman Gushchin 7395cf11e85fSRoman Gushchin if (!hugetlb_cma_size) 7396cf11e85fSRoman Gushchin return; 7397cf11e85fSRoman Gushchin 739838e719abSBaolin Wang for (nid = 0; nid < MAX_NUMNODES; nid++) { 739938e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 740038e719abSBaolin Wang continue; 740138e719abSBaolin Wang 740230a51400SPeng Liu if (!node_online(nid)) { 740338e719abSBaolin Wang pr_warn("hugetlb_cma: invalid node %d specified\n", nid); 740438e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 740538e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 740638e719abSBaolin Wang continue; 740738e719abSBaolin Wang } 740838e719abSBaolin Wang 740938e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) { 741038e719abSBaolin Wang pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n", 741138e719abSBaolin Wang nid, (PAGE_SIZE << order) / SZ_1M); 741238e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 741338e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 741438e719abSBaolin Wang } else { 741538e719abSBaolin Wang node_specific_cma_alloc = true; 741638e719abSBaolin Wang } 741738e719abSBaolin Wang } 741838e719abSBaolin Wang 741938e719abSBaolin Wang /* Validate the CMA size again in case some invalid nodes specified. */ 742038e719abSBaolin Wang if (!hugetlb_cma_size) 742138e719abSBaolin Wang return; 742238e719abSBaolin Wang 7423cf11e85fSRoman Gushchin if (hugetlb_cma_size < (PAGE_SIZE << order)) { 7424cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", 7425cf11e85fSRoman Gushchin (PAGE_SIZE << order) / SZ_1M); 7426a01f4390SMike Kravetz hugetlb_cma_size = 0; 7427cf11e85fSRoman Gushchin return; 7428cf11e85fSRoman Gushchin } 7429cf11e85fSRoman Gushchin 743038e719abSBaolin Wang if (!node_specific_cma_alloc) { 7431cf11e85fSRoman Gushchin /* 7432cf11e85fSRoman Gushchin * If 3 GB area is requested on a machine with 4 numa nodes, 7433cf11e85fSRoman Gushchin * let's allocate 1 GB on first three nodes and ignore the last one. 7434cf11e85fSRoman Gushchin */ 7435cf11e85fSRoman Gushchin per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); 7436cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", 7437cf11e85fSRoman Gushchin hugetlb_cma_size / SZ_1M, per_node / SZ_1M); 743838e719abSBaolin Wang } 7439cf11e85fSRoman Gushchin 7440cf11e85fSRoman Gushchin reserved = 0; 744130a51400SPeng Liu for_each_online_node(nid) { 7442cf11e85fSRoman Gushchin int res; 74432281f797SBarry Song char name[CMA_MAX_NAME]; 7444cf11e85fSRoman Gushchin 744538e719abSBaolin Wang if (node_specific_cma_alloc) { 744638e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 744738e719abSBaolin Wang continue; 744838e719abSBaolin Wang 744938e719abSBaolin Wang size = hugetlb_cma_size_in_node[nid]; 745038e719abSBaolin Wang } else { 7451cf11e85fSRoman Gushchin size = min(per_node, hugetlb_cma_size - reserved); 745238e719abSBaolin Wang } 745338e719abSBaolin Wang 7454cf11e85fSRoman Gushchin size = round_up(size, PAGE_SIZE << order); 7455cf11e85fSRoman Gushchin 74562281f797SBarry Song snprintf(name, sizeof(name), "hugetlb%d", nid); 7457a01f4390SMike Kravetz /* 7458a01f4390SMike Kravetz * Note that 'order per bit' is based on smallest size that 7459a01f4390SMike Kravetz * may be returned to CMA allocator in the case of 7460a01f4390SMike Kravetz * huge page demotion. 7461a01f4390SMike Kravetz */ 7462a01f4390SMike Kravetz res = cma_declare_contiguous_nid(0, size, 0, 7463a01f4390SMike Kravetz PAGE_SIZE << HUGETLB_PAGE_ORDER, 746429d0f41dSBarry Song 0, false, name, 7465cf11e85fSRoman Gushchin &hugetlb_cma[nid], nid); 7466cf11e85fSRoman Gushchin if (res) { 7467cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: reservation failed: err %d, node %d", 7468cf11e85fSRoman Gushchin res, nid); 7469cf11e85fSRoman Gushchin continue; 7470cf11e85fSRoman Gushchin } 7471cf11e85fSRoman Gushchin 7472cf11e85fSRoman Gushchin reserved += size; 7473cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", 7474cf11e85fSRoman Gushchin size / SZ_1M, nid); 7475cf11e85fSRoman Gushchin 7476cf11e85fSRoman Gushchin if (reserved >= hugetlb_cma_size) 7477cf11e85fSRoman Gushchin break; 7478cf11e85fSRoman Gushchin } 7479a01f4390SMike Kravetz 7480a01f4390SMike Kravetz if (!reserved) 7481a01f4390SMike Kravetz /* 7482a01f4390SMike Kravetz * hugetlb_cma_size is used to determine if allocations from 7483a01f4390SMike Kravetz * cma are possible. Set to zero if no cma regions are set up. 7484a01f4390SMike Kravetz */ 7485a01f4390SMike Kravetz hugetlb_cma_size = 0; 7486cf11e85fSRoman Gushchin } 7487cf11e85fSRoman Gushchin 7488263b8998SMiaohe Lin static void __init hugetlb_cma_check(void) 7489cf11e85fSRoman Gushchin { 7490cf11e85fSRoman Gushchin if (!hugetlb_cma_size || cma_reserve_called) 7491cf11e85fSRoman Gushchin return; 7492cf11e85fSRoman Gushchin 7493cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); 7494cf11e85fSRoman Gushchin } 7495cf11e85fSRoman Gushchin 7496cf11e85fSRoman Gushchin #endif /* CONFIG_CMA */ 7497