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> 37d6606683SLinus Torvalds 3863551ae0SDavid Gibson #include <asm/page.h> 39ca15ca40SMike Rapoport #include <asm/pgalloc.h> 4024669e58SAneesh Kumar K.V #include <asm/tlb.h> 4163551ae0SDavid Gibson 4224669e58SAneesh Kumar K.V #include <linux/io.h> 4363551ae0SDavid Gibson #include <linux/hugetlb.h> 449dd540e2SAneesh Kumar K.V #include <linux/hugetlb_cgroup.h> 459a305230SLee Schermerhorn #include <linux/node.h> 46ab5ac90aSMichal Hocko #include <linux/page_owner.h> 477835e98bSNick Piggin #include "internal.h" 48f41f2ed4SMuchun Song #include "hugetlb_vmemmap.h" 491da177e4SLinus Torvalds 50c3f38a38SAneesh Kumar K.V int hugetlb_max_hstate __read_mostly; 51e5ff2159SAndi Kleen unsigned int default_hstate_idx; 52e5ff2159SAndi Kleen struct hstate hstates[HUGE_MAX_HSTATE]; 53cf11e85fSRoman Gushchin 54dbda8feaSBarry Song #ifdef CONFIG_CMA 55cf11e85fSRoman Gushchin static struct cma *hugetlb_cma[MAX_NUMNODES]; 5638e719abSBaolin Wang static unsigned long hugetlb_cma_size_in_node[MAX_NUMNODES] __initdata; 572f6c57d6SSidhartha Kumar static bool hugetlb_cma_folio(struct folio *folio, unsigned int order) 58a01f4390SMike Kravetz { 592f6c57d6SSidhartha Kumar return cma_pages_valid(hugetlb_cma[folio_nid(folio)], &folio->page, 60a01f4390SMike Kravetz 1 << order); 61a01f4390SMike Kravetz } 62a01f4390SMike Kravetz #else 632f6c57d6SSidhartha Kumar static bool hugetlb_cma_folio(struct folio *folio, unsigned int order) 64a01f4390SMike Kravetz { 65a01f4390SMike Kravetz return false; 66a01f4390SMike Kravetz } 67dbda8feaSBarry Song #endif 68dbda8feaSBarry Song static unsigned long hugetlb_cma_size __initdata; 69cf11e85fSRoman Gushchin 7053ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 7153ba51d2SJon Tollefson 72e5ff2159SAndi Kleen /* for command line parsing */ 73e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 74e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 759fee021dSVaishali Thakkar static bool __initdata parsed_valid_hugepagesz = true; 76282f4214SMike Kravetz static bool __initdata parsed_default_hugepagesz; 77b5389086SZhenguo Yao static unsigned int default_hugepages_in_node[MAX_NUMNODES] __initdata; 78e5ff2159SAndi Kleen 793935baa9SDavid Gibson /* 8031caf665SNaoya Horiguchi * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, 8131caf665SNaoya Horiguchi * free_huge_pages, and surplus_huge_pages. 823935baa9SDavid Gibson */ 83c3f38a38SAneesh Kumar K.V DEFINE_SPINLOCK(hugetlb_lock); 840bd0f9fbSEric Paris 858382d914SDavidlohr Bueso /* 868382d914SDavidlohr Bueso * Serializes faults on the same logical page. This is used to 878382d914SDavidlohr Bueso * prevent spurious OOMs when the hugepage pool is fully utilized. 888382d914SDavidlohr Bueso */ 898382d914SDavidlohr Bueso static int num_fault_mutexes; 90c672c7f2SMike Kravetz struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; 918382d914SDavidlohr Bueso 927ca02d0aSMike Kravetz /* Forward declaration */ 937ca02d0aSMike Kravetz static int hugetlb_acct_memory(struct hstate *h, long delta); 948d9bfb26SMike Kravetz static void hugetlb_vma_lock_free(struct vm_area_struct *vma); 958d9bfb26SMike Kravetz static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma); 96ecfbd733SMike Kravetz static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma); 977ca02d0aSMike Kravetz 981d88433bSMiaohe Lin static inline bool subpool_is_free(struct hugepage_subpool *spool) 991d88433bSMiaohe Lin { 1001d88433bSMiaohe Lin if (spool->count) 1011d88433bSMiaohe Lin return false; 1021d88433bSMiaohe Lin if (spool->max_hpages != -1) 1031d88433bSMiaohe Lin return spool->used_hpages == 0; 1041d88433bSMiaohe Lin if (spool->min_hpages != -1) 1051d88433bSMiaohe Lin return spool->rsv_hpages == spool->min_hpages; 1061d88433bSMiaohe Lin 1071d88433bSMiaohe Lin return true; 1081d88433bSMiaohe Lin } 1091d88433bSMiaohe Lin 110db71ef79SMike Kravetz static inline void unlock_or_release_subpool(struct hugepage_subpool *spool, 111db71ef79SMike Kravetz unsigned long irq_flags) 11290481622SDavid Gibson { 113db71ef79SMike Kravetz spin_unlock_irqrestore(&spool->lock, irq_flags); 11490481622SDavid Gibson 11590481622SDavid Gibson /* If no pages are used, and no other handles to the subpool 1167c8de358SEthon Paul * remain, give up any reservations based on minimum size and 1177ca02d0aSMike Kravetz * free the subpool */ 1181d88433bSMiaohe Lin if (subpool_is_free(spool)) { 1197ca02d0aSMike Kravetz if (spool->min_hpages != -1) 1207ca02d0aSMike Kravetz hugetlb_acct_memory(spool->hstate, 1217ca02d0aSMike Kravetz -spool->min_hpages); 12290481622SDavid Gibson kfree(spool); 12390481622SDavid Gibson } 1247ca02d0aSMike Kravetz } 12590481622SDavid Gibson 1267ca02d0aSMike Kravetz struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, 1277ca02d0aSMike Kravetz long min_hpages) 12890481622SDavid Gibson { 12990481622SDavid Gibson struct hugepage_subpool *spool; 13090481622SDavid Gibson 131c6a91820SMike Kravetz spool = kzalloc(sizeof(*spool), GFP_KERNEL); 13290481622SDavid Gibson if (!spool) 13390481622SDavid Gibson return NULL; 13490481622SDavid Gibson 13590481622SDavid Gibson spin_lock_init(&spool->lock); 13690481622SDavid Gibson spool->count = 1; 1377ca02d0aSMike Kravetz spool->max_hpages = max_hpages; 1387ca02d0aSMike Kravetz spool->hstate = h; 1397ca02d0aSMike Kravetz spool->min_hpages = min_hpages; 1407ca02d0aSMike Kravetz 1417ca02d0aSMike Kravetz if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { 1427ca02d0aSMike Kravetz kfree(spool); 1437ca02d0aSMike Kravetz return NULL; 1447ca02d0aSMike Kravetz } 1457ca02d0aSMike Kravetz spool->rsv_hpages = min_hpages; 14690481622SDavid Gibson 14790481622SDavid Gibson return spool; 14890481622SDavid Gibson } 14990481622SDavid Gibson 15090481622SDavid Gibson void hugepage_put_subpool(struct hugepage_subpool *spool) 15190481622SDavid Gibson { 152db71ef79SMike Kravetz unsigned long flags; 153db71ef79SMike Kravetz 154db71ef79SMike Kravetz spin_lock_irqsave(&spool->lock, flags); 15590481622SDavid Gibson BUG_ON(!spool->count); 15690481622SDavid Gibson spool->count--; 157db71ef79SMike Kravetz unlock_or_release_subpool(spool, flags); 15890481622SDavid Gibson } 15990481622SDavid Gibson 1601c5ecae3SMike Kravetz /* 1611c5ecae3SMike Kravetz * Subpool accounting for allocating and reserving pages. 1621c5ecae3SMike Kravetz * Return -ENOMEM if there are not enough resources to satisfy the 1639e7ee400SRandy Dunlap * request. Otherwise, return the number of pages by which the 1641c5ecae3SMike Kravetz * global pools must be adjusted (upward). The returned value may 1651c5ecae3SMike Kravetz * only be different than the passed value (delta) in the case where 1667c8de358SEthon Paul * a subpool minimum size must be maintained. 1671c5ecae3SMike Kravetz */ 1681c5ecae3SMike Kravetz static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, 16990481622SDavid Gibson long delta) 17090481622SDavid Gibson { 1711c5ecae3SMike Kravetz long ret = delta; 17290481622SDavid Gibson 17390481622SDavid Gibson if (!spool) 1741c5ecae3SMike Kravetz return ret; 17590481622SDavid Gibson 176db71ef79SMike Kravetz spin_lock_irq(&spool->lock); 17790481622SDavid Gibson 1781c5ecae3SMike Kravetz if (spool->max_hpages != -1) { /* maximum size accounting */ 1791c5ecae3SMike Kravetz if ((spool->used_hpages + delta) <= spool->max_hpages) 1801c5ecae3SMike Kravetz spool->used_hpages += delta; 1811c5ecae3SMike Kravetz else { 1821c5ecae3SMike Kravetz ret = -ENOMEM; 1831c5ecae3SMike Kravetz goto unlock_ret; 1841c5ecae3SMike Kravetz } 1851c5ecae3SMike Kravetz } 1861c5ecae3SMike Kravetz 18709a95e29SMike Kravetz /* minimum size accounting */ 18809a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->rsv_hpages) { 1891c5ecae3SMike Kravetz if (delta > spool->rsv_hpages) { 1901c5ecae3SMike Kravetz /* 1911c5ecae3SMike Kravetz * Asking for more reserves than those already taken on 1921c5ecae3SMike Kravetz * behalf of subpool. Return difference. 1931c5ecae3SMike Kravetz */ 1941c5ecae3SMike Kravetz ret = delta - spool->rsv_hpages; 1951c5ecae3SMike Kravetz spool->rsv_hpages = 0; 1961c5ecae3SMike Kravetz } else { 1971c5ecae3SMike Kravetz ret = 0; /* reserves already accounted for */ 1981c5ecae3SMike Kravetz spool->rsv_hpages -= delta; 1991c5ecae3SMike Kravetz } 2001c5ecae3SMike Kravetz } 2011c5ecae3SMike Kravetz 2021c5ecae3SMike Kravetz unlock_ret: 203db71ef79SMike Kravetz spin_unlock_irq(&spool->lock); 20490481622SDavid Gibson return ret; 20590481622SDavid Gibson } 20690481622SDavid Gibson 2071c5ecae3SMike Kravetz /* 2081c5ecae3SMike Kravetz * Subpool accounting for freeing and unreserving pages. 2091c5ecae3SMike Kravetz * Return the number of global page reservations that must be dropped. 2101c5ecae3SMike Kravetz * The return value may only be different than the passed value (delta) 2111c5ecae3SMike Kravetz * in the case where a subpool minimum size must be maintained. 2121c5ecae3SMike Kravetz */ 2131c5ecae3SMike Kravetz static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, 21490481622SDavid Gibson long delta) 21590481622SDavid Gibson { 2161c5ecae3SMike Kravetz long ret = delta; 217db71ef79SMike Kravetz unsigned long flags; 2181c5ecae3SMike Kravetz 21990481622SDavid Gibson if (!spool) 2201c5ecae3SMike Kravetz return delta; 22190481622SDavid Gibson 222db71ef79SMike Kravetz spin_lock_irqsave(&spool->lock, flags); 2231c5ecae3SMike Kravetz 2241c5ecae3SMike Kravetz if (spool->max_hpages != -1) /* maximum size accounting */ 22590481622SDavid Gibson spool->used_hpages -= delta; 2261c5ecae3SMike Kravetz 22709a95e29SMike Kravetz /* minimum size accounting */ 22809a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { 2291c5ecae3SMike Kravetz if (spool->rsv_hpages + delta <= spool->min_hpages) 2301c5ecae3SMike Kravetz ret = 0; 2311c5ecae3SMike Kravetz else 2321c5ecae3SMike Kravetz ret = spool->rsv_hpages + delta - spool->min_hpages; 2331c5ecae3SMike Kravetz 2341c5ecae3SMike Kravetz spool->rsv_hpages += delta; 2351c5ecae3SMike Kravetz if (spool->rsv_hpages > spool->min_hpages) 2361c5ecae3SMike Kravetz spool->rsv_hpages = spool->min_hpages; 2371c5ecae3SMike Kravetz } 2381c5ecae3SMike Kravetz 2391c5ecae3SMike Kravetz /* 2401c5ecae3SMike Kravetz * If hugetlbfs_put_super couldn't free spool due to an outstanding 2411c5ecae3SMike Kravetz * quota reference, free it now. 2421c5ecae3SMike Kravetz */ 243db71ef79SMike Kravetz unlock_or_release_subpool(spool, flags); 2441c5ecae3SMike Kravetz 2451c5ecae3SMike Kravetz return ret; 24690481622SDavid Gibson } 24790481622SDavid Gibson 24890481622SDavid Gibson static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 24990481622SDavid Gibson { 25090481622SDavid Gibson return HUGETLBFS_SB(inode->i_sb)->spool; 25190481622SDavid Gibson } 25290481622SDavid Gibson 25390481622SDavid Gibson static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 25490481622SDavid Gibson { 255496ad9aaSAl Viro return subpool_inode(file_inode(vma->vm_file)); 25690481622SDavid Gibson } 25790481622SDavid Gibson 258e700898fSMike Kravetz /* 259e700898fSMike Kravetz * hugetlb vma_lock helper routines 260e700898fSMike Kravetz */ 261e700898fSMike Kravetz static bool __vma_shareable_lock(struct vm_area_struct *vma) 262e700898fSMike Kravetz { 263e700898fSMike Kravetz return vma->vm_flags & (VM_MAYSHARE | VM_SHARED) && 264e700898fSMike Kravetz vma->vm_private_data; 265e700898fSMike Kravetz } 266e700898fSMike Kravetz 267e700898fSMike Kravetz void hugetlb_vma_lock_read(struct vm_area_struct *vma) 268e700898fSMike Kravetz { 269e700898fSMike Kravetz if (__vma_shareable_lock(vma)) { 270e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 271e700898fSMike Kravetz 272e700898fSMike Kravetz down_read(&vma_lock->rw_sema); 273e700898fSMike Kravetz } 274e700898fSMike Kravetz } 275e700898fSMike Kravetz 276e700898fSMike Kravetz void hugetlb_vma_unlock_read(struct vm_area_struct *vma) 277e700898fSMike Kravetz { 278e700898fSMike Kravetz if (__vma_shareable_lock(vma)) { 279e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 280e700898fSMike Kravetz 281e700898fSMike Kravetz up_read(&vma_lock->rw_sema); 282e700898fSMike Kravetz } 283e700898fSMike Kravetz } 284e700898fSMike Kravetz 285e700898fSMike Kravetz void hugetlb_vma_lock_write(struct vm_area_struct *vma) 286e700898fSMike Kravetz { 287e700898fSMike Kravetz if (__vma_shareable_lock(vma)) { 288e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 289e700898fSMike Kravetz 290e700898fSMike Kravetz down_write(&vma_lock->rw_sema); 291e700898fSMike Kravetz } 292e700898fSMike Kravetz } 293e700898fSMike Kravetz 294e700898fSMike Kravetz void hugetlb_vma_unlock_write(struct vm_area_struct *vma) 295e700898fSMike Kravetz { 296e700898fSMike Kravetz if (__vma_shareable_lock(vma)) { 297e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 298e700898fSMike Kravetz 299e700898fSMike Kravetz up_write(&vma_lock->rw_sema); 300e700898fSMike Kravetz } 301e700898fSMike Kravetz } 302e700898fSMike Kravetz 303e700898fSMike Kravetz int hugetlb_vma_trylock_write(struct vm_area_struct *vma) 304e700898fSMike Kravetz { 305e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 306e700898fSMike Kravetz 307e700898fSMike Kravetz if (!__vma_shareable_lock(vma)) 308e700898fSMike Kravetz return 1; 309e700898fSMike Kravetz 310e700898fSMike Kravetz return down_write_trylock(&vma_lock->rw_sema); 311e700898fSMike Kravetz } 312e700898fSMike Kravetz 313e700898fSMike Kravetz void hugetlb_vma_assert_locked(struct vm_area_struct *vma) 314e700898fSMike Kravetz { 315e700898fSMike Kravetz if (__vma_shareable_lock(vma)) { 316e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 317e700898fSMike Kravetz 318e700898fSMike Kravetz lockdep_assert_held(&vma_lock->rw_sema); 319e700898fSMike Kravetz } 320e700898fSMike Kravetz } 321e700898fSMike Kravetz 322e700898fSMike Kravetz void hugetlb_vma_lock_release(struct kref *kref) 323e700898fSMike Kravetz { 324e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = container_of(kref, 325e700898fSMike Kravetz struct hugetlb_vma_lock, refs); 326e700898fSMike Kravetz 327e700898fSMike Kravetz kfree(vma_lock); 328e700898fSMike Kravetz } 329e700898fSMike Kravetz 330e700898fSMike Kravetz static void __hugetlb_vma_unlock_write_put(struct hugetlb_vma_lock *vma_lock) 331e700898fSMike Kravetz { 332e700898fSMike Kravetz struct vm_area_struct *vma = vma_lock->vma; 333e700898fSMike Kravetz 334e700898fSMike Kravetz /* 335e700898fSMike Kravetz * vma_lock structure may or not be released as a result of put, 336e700898fSMike Kravetz * it certainly will no longer be attached to vma so clear pointer. 337e700898fSMike Kravetz * Semaphore synchronizes access to vma_lock->vma field. 338e700898fSMike Kravetz */ 339e700898fSMike Kravetz vma_lock->vma = NULL; 340e700898fSMike Kravetz vma->vm_private_data = NULL; 341e700898fSMike Kravetz up_write(&vma_lock->rw_sema); 342e700898fSMike Kravetz kref_put(&vma_lock->refs, hugetlb_vma_lock_release); 343e700898fSMike Kravetz } 344e700898fSMike Kravetz 345e700898fSMike Kravetz static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma) 346e700898fSMike Kravetz { 347e700898fSMike Kravetz if (__vma_shareable_lock(vma)) { 348e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 349e700898fSMike Kravetz 350e700898fSMike Kravetz __hugetlb_vma_unlock_write_put(vma_lock); 351e700898fSMike Kravetz } 352e700898fSMike Kravetz } 353e700898fSMike Kravetz 354e700898fSMike Kravetz static void hugetlb_vma_lock_free(struct vm_area_struct *vma) 355e700898fSMike Kravetz { 356e700898fSMike Kravetz /* 357e700898fSMike Kravetz * Only present in sharable vmas. 358e700898fSMike Kravetz */ 359e700898fSMike Kravetz if (!vma || !__vma_shareable_lock(vma)) 360e700898fSMike Kravetz return; 361e700898fSMike Kravetz 362e700898fSMike Kravetz if (vma->vm_private_data) { 363e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 364e700898fSMike Kravetz 365e700898fSMike Kravetz down_write(&vma_lock->rw_sema); 366e700898fSMike Kravetz __hugetlb_vma_unlock_write_put(vma_lock); 367e700898fSMike Kravetz } 368e700898fSMike Kravetz } 369e700898fSMike Kravetz 370e700898fSMike Kravetz static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma) 371e700898fSMike Kravetz { 372e700898fSMike Kravetz struct hugetlb_vma_lock *vma_lock; 373e700898fSMike Kravetz 374e700898fSMike Kravetz /* Only establish in (flags) sharable vmas */ 375e700898fSMike Kravetz if (!vma || !(vma->vm_flags & VM_MAYSHARE)) 376e700898fSMike Kravetz return; 377e700898fSMike Kravetz 378e700898fSMike Kravetz /* Should never get here with non-NULL vm_private_data */ 379e700898fSMike Kravetz if (vma->vm_private_data) 380e700898fSMike Kravetz return; 381e700898fSMike Kravetz 382e700898fSMike Kravetz vma_lock = kmalloc(sizeof(*vma_lock), GFP_KERNEL); 383e700898fSMike Kravetz if (!vma_lock) { 384e700898fSMike Kravetz /* 385e700898fSMike Kravetz * If we can not allocate structure, then vma can not 386e700898fSMike Kravetz * participate in pmd sharing. This is only a possible 387e700898fSMike Kravetz * performance enhancement and memory saving issue. 388e700898fSMike Kravetz * However, the lock is also used to synchronize page 389e700898fSMike Kravetz * faults with truncation. If the lock is not present, 390e700898fSMike Kravetz * unlikely races could leave pages in a file past i_size 391e700898fSMike Kravetz * until the file is removed. Warn in the unlikely case of 392e700898fSMike Kravetz * allocation failure. 393e700898fSMike Kravetz */ 394e700898fSMike Kravetz pr_warn_once("HugeTLB: unable to allocate vma specific lock\n"); 395e700898fSMike Kravetz return; 396e700898fSMike Kravetz } 397e700898fSMike Kravetz 398e700898fSMike Kravetz kref_init(&vma_lock->refs); 399e700898fSMike Kravetz init_rwsem(&vma_lock->rw_sema); 400e700898fSMike Kravetz vma_lock->vma = vma; 401e700898fSMike Kravetz vma->vm_private_data = vma_lock; 402e700898fSMike Kravetz } 403e700898fSMike Kravetz 4040db9d74eSMina Almasry /* Helper that removes a struct file_region from the resv_map cache and returns 4050db9d74eSMina Almasry * it for use. 4060db9d74eSMina Almasry */ 4070db9d74eSMina Almasry static struct file_region * 4080db9d74eSMina Almasry get_file_region_entry_from_cache(struct resv_map *resv, long from, long to) 4090db9d74eSMina Almasry { 4103259914fSXU pengfei struct file_region *nrg; 4110db9d74eSMina Almasry 4120db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count <= 0); 4130db9d74eSMina Almasry 4140db9d74eSMina Almasry resv->region_cache_count--; 4150db9d74eSMina Almasry nrg = list_first_entry(&resv->region_cache, struct file_region, link); 4160db9d74eSMina Almasry list_del(&nrg->link); 4170db9d74eSMina Almasry 4180db9d74eSMina Almasry nrg->from = from; 4190db9d74eSMina Almasry nrg->to = to; 4200db9d74eSMina Almasry 4210db9d74eSMina Almasry return nrg; 4220db9d74eSMina Almasry } 4230db9d74eSMina Almasry 424075a61d0SMina Almasry static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, 425075a61d0SMina Almasry struct file_region *rg) 426075a61d0SMina Almasry { 427075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 428075a61d0SMina Almasry nrg->reservation_counter = rg->reservation_counter; 429075a61d0SMina Almasry nrg->css = rg->css; 430075a61d0SMina Almasry if (rg->css) 431075a61d0SMina Almasry css_get(rg->css); 432075a61d0SMina Almasry #endif 433075a61d0SMina Almasry } 434075a61d0SMina Almasry 435075a61d0SMina Almasry /* Helper that records hugetlb_cgroup uncharge info. */ 436075a61d0SMina Almasry static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, 437075a61d0SMina Almasry struct hstate *h, 438075a61d0SMina Almasry struct resv_map *resv, 439075a61d0SMina Almasry struct file_region *nrg) 440075a61d0SMina Almasry { 441075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 442075a61d0SMina Almasry if (h_cg) { 443075a61d0SMina Almasry nrg->reservation_counter = 444075a61d0SMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 445075a61d0SMina Almasry nrg->css = &h_cg->css; 446d85aecf2SMiaohe Lin /* 447d85aecf2SMiaohe Lin * The caller will hold exactly one h_cg->css reference for the 448d85aecf2SMiaohe Lin * whole contiguous reservation region. But this area might be 449d85aecf2SMiaohe Lin * scattered when there are already some file_regions reside in 450d85aecf2SMiaohe Lin * it. As a result, many file_regions may share only one css 451d85aecf2SMiaohe Lin * reference. In order to ensure that one file_region must hold 452d85aecf2SMiaohe Lin * exactly one h_cg->css reference, we should do css_get for 453d85aecf2SMiaohe Lin * each file_region and leave the reference held by caller 454d85aecf2SMiaohe Lin * untouched. 455d85aecf2SMiaohe Lin */ 456d85aecf2SMiaohe Lin css_get(&h_cg->css); 457075a61d0SMina Almasry if (!resv->pages_per_hpage) 458075a61d0SMina Almasry resv->pages_per_hpage = pages_per_huge_page(h); 459075a61d0SMina Almasry /* pages_per_hpage should be the same for all entries in 460075a61d0SMina Almasry * a resv_map. 461075a61d0SMina Almasry */ 462075a61d0SMina Almasry VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); 463075a61d0SMina Almasry } else { 464075a61d0SMina Almasry nrg->reservation_counter = NULL; 465075a61d0SMina Almasry nrg->css = NULL; 466075a61d0SMina Almasry } 467075a61d0SMina Almasry #endif 468075a61d0SMina Almasry } 469075a61d0SMina Almasry 470d85aecf2SMiaohe Lin static void put_uncharge_info(struct file_region *rg) 471d85aecf2SMiaohe Lin { 472d85aecf2SMiaohe Lin #ifdef CONFIG_CGROUP_HUGETLB 473d85aecf2SMiaohe Lin if (rg->css) 474d85aecf2SMiaohe Lin css_put(rg->css); 475d85aecf2SMiaohe Lin #endif 476d85aecf2SMiaohe Lin } 477d85aecf2SMiaohe Lin 478a9b3f867SMina Almasry static bool has_same_uncharge_info(struct file_region *rg, 479a9b3f867SMina Almasry struct file_region *org) 480a9b3f867SMina Almasry { 481a9b3f867SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 4820739eb43SBaolin Wang return rg->reservation_counter == org->reservation_counter && 483a9b3f867SMina Almasry rg->css == org->css; 484a9b3f867SMina Almasry 485a9b3f867SMina Almasry #else 486a9b3f867SMina Almasry return true; 487a9b3f867SMina Almasry #endif 488a9b3f867SMina Almasry } 489a9b3f867SMina Almasry 490a9b3f867SMina Almasry static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) 491a9b3f867SMina Almasry { 4923259914fSXU pengfei struct file_region *nrg, *prg; 493a9b3f867SMina Almasry 494a9b3f867SMina Almasry prg = list_prev_entry(rg, link); 495a9b3f867SMina Almasry if (&prg->link != &resv->regions && prg->to == rg->from && 496a9b3f867SMina Almasry has_same_uncharge_info(prg, rg)) { 497a9b3f867SMina Almasry prg->to = rg->to; 498a9b3f867SMina Almasry 499a9b3f867SMina Almasry list_del(&rg->link); 500d85aecf2SMiaohe Lin put_uncharge_info(rg); 501a9b3f867SMina Almasry kfree(rg); 502a9b3f867SMina Almasry 5037db5e7b6SWei Yang rg = prg; 504a9b3f867SMina Almasry } 505a9b3f867SMina Almasry 506a9b3f867SMina Almasry nrg = list_next_entry(rg, link); 507a9b3f867SMina Almasry if (&nrg->link != &resv->regions && nrg->from == rg->to && 508a9b3f867SMina Almasry has_same_uncharge_info(nrg, rg)) { 509a9b3f867SMina Almasry nrg->from = rg->from; 510a9b3f867SMina Almasry 511a9b3f867SMina Almasry list_del(&rg->link); 512d85aecf2SMiaohe Lin put_uncharge_info(rg); 513a9b3f867SMina Almasry kfree(rg); 514a9b3f867SMina Almasry } 515a9b3f867SMina Almasry } 516a9b3f867SMina Almasry 5172103cf9cSPeter Xu static inline long 51884448c8eSJakob Koschel hugetlb_resv_map_add(struct resv_map *map, struct list_head *rg, long from, 5192103cf9cSPeter Xu long to, struct hstate *h, struct hugetlb_cgroup *cg, 5202103cf9cSPeter Xu long *regions_needed) 5212103cf9cSPeter Xu { 5222103cf9cSPeter Xu struct file_region *nrg; 5232103cf9cSPeter Xu 5242103cf9cSPeter Xu if (!regions_needed) { 5252103cf9cSPeter Xu nrg = get_file_region_entry_from_cache(map, from, to); 5262103cf9cSPeter Xu record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg); 52784448c8eSJakob Koschel list_add(&nrg->link, rg); 5282103cf9cSPeter Xu coalesce_file_region(map, nrg); 5292103cf9cSPeter Xu } else 5302103cf9cSPeter Xu *regions_needed += 1; 5312103cf9cSPeter Xu 5322103cf9cSPeter Xu return to - from; 5332103cf9cSPeter Xu } 5342103cf9cSPeter Xu 535972a3da3SWei Yang /* 536972a3da3SWei Yang * Must be called with resv->lock held. 537972a3da3SWei Yang * 538972a3da3SWei Yang * Calling this with regions_needed != NULL will count the number of pages 539972a3da3SWei Yang * to be added but will not modify the linked list. And regions_needed will 540972a3da3SWei Yang * indicate the number of file_regions needed in the cache to carry out to add 541972a3da3SWei Yang * the regions for this range. 542d75c6af9SMina Almasry */ 543d75c6af9SMina Almasry static long add_reservation_in_range(struct resv_map *resv, long f, long t, 544075a61d0SMina Almasry struct hugetlb_cgroup *h_cg, 545972a3da3SWei Yang struct hstate *h, long *regions_needed) 546d75c6af9SMina Almasry { 5470db9d74eSMina Almasry long add = 0; 548d75c6af9SMina Almasry struct list_head *head = &resv->regions; 5490db9d74eSMina Almasry long last_accounted_offset = f; 55084448c8eSJakob Koschel struct file_region *iter, *trg = NULL; 55184448c8eSJakob Koschel struct list_head *rg = NULL; 552d75c6af9SMina Almasry 5530db9d74eSMina Almasry if (regions_needed) 5540db9d74eSMina Almasry *regions_needed = 0; 555d75c6af9SMina Almasry 5560db9d74eSMina Almasry /* In this loop, we essentially handle an entry for the range 55784448c8eSJakob Koschel * [last_accounted_offset, iter->from), at every iteration, with some 5580db9d74eSMina Almasry * bounds checking. 5590db9d74eSMina Almasry */ 56084448c8eSJakob Koschel list_for_each_entry_safe(iter, trg, head, link) { 5610db9d74eSMina Almasry /* Skip irrelevant regions that start before our range. */ 56284448c8eSJakob Koschel if (iter->from < f) { 5630db9d74eSMina Almasry /* If this region ends after the last accounted offset, 5640db9d74eSMina Almasry * then we need to update last_accounted_offset. 5650db9d74eSMina Almasry */ 56684448c8eSJakob Koschel if (iter->to > last_accounted_offset) 56784448c8eSJakob Koschel last_accounted_offset = iter->to; 5680db9d74eSMina Almasry continue; 5690db9d74eSMina Almasry } 570d75c6af9SMina Almasry 5710db9d74eSMina Almasry /* When we find a region that starts beyond our range, we've 5720db9d74eSMina Almasry * finished. 5730db9d74eSMina Almasry */ 57484448c8eSJakob Koschel if (iter->from >= t) { 57584448c8eSJakob Koschel rg = iter->link.prev; 576d75c6af9SMina Almasry break; 57784448c8eSJakob Koschel } 578d75c6af9SMina Almasry 57984448c8eSJakob Koschel /* Add an entry for last_accounted_offset -> iter->from, and 5800db9d74eSMina Almasry * update last_accounted_offset. 581d75c6af9SMina Almasry */ 58284448c8eSJakob Koschel if (iter->from > last_accounted_offset) 58384448c8eSJakob Koschel add += hugetlb_resv_map_add(resv, iter->link.prev, 5842103cf9cSPeter Xu last_accounted_offset, 58584448c8eSJakob Koschel iter->from, h, h_cg, 5862103cf9cSPeter Xu regions_needed); 587d75c6af9SMina Almasry 58884448c8eSJakob Koschel last_accounted_offset = iter->to; 5890db9d74eSMina Almasry } 5900db9d74eSMina Almasry 5910db9d74eSMina Almasry /* Handle the case where our range extends beyond 5920db9d74eSMina Almasry * last_accounted_offset. 5930db9d74eSMina Almasry */ 59484448c8eSJakob Koschel if (!rg) 59584448c8eSJakob Koschel rg = head->prev; 5962103cf9cSPeter Xu if (last_accounted_offset < t) 5972103cf9cSPeter Xu add += hugetlb_resv_map_add(resv, rg, last_accounted_offset, 5982103cf9cSPeter Xu t, h, h_cg, regions_needed); 5990db9d74eSMina Almasry 6000db9d74eSMina Almasry return add; 6010db9d74eSMina Almasry } 6020db9d74eSMina Almasry 6030db9d74eSMina Almasry /* Must be called with resv->lock acquired. Will drop lock to allocate entries. 6040db9d74eSMina Almasry */ 6050db9d74eSMina Almasry static int allocate_file_region_entries(struct resv_map *resv, 6060db9d74eSMina Almasry int regions_needed) 6070db9d74eSMina Almasry __must_hold(&resv->lock) 6080db9d74eSMina Almasry { 60934665341SMiaohe Lin LIST_HEAD(allocated_regions); 6100db9d74eSMina Almasry int to_allocate = 0, i = 0; 6110db9d74eSMina Almasry struct file_region *trg = NULL, *rg = NULL; 6120db9d74eSMina Almasry 6130db9d74eSMina Almasry VM_BUG_ON(regions_needed < 0); 6140db9d74eSMina Almasry 6150db9d74eSMina Almasry /* 6160db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 6170db9d74eSMina Almasry * the number of in progress add operations plus regions_needed. 6180db9d74eSMina Almasry * 6190db9d74eSMina Almasry * This is a while loop because when we drop the lock, some other call 6200db9d74eSMina Almasry * to region_add or region_del may have consumed some region_entries, 6210db9d74eSMina Almasry * so we keep looping here until we finally have enough entries for 6220db9d74eSMina Almasry * (adds_in_progress + regions_needed). 6230db9d74eSMina Almasry */ 6240db9d74eSMina Almasry while (resv->region_cache_count < 6250db9d74eSMina Almasry (resv->adds_in_progress + regions_needed)) { 6260db9d74eSMina Almasry to_allocate = resv->adds_in_progress + regions_needed - 6270db9d74eSMina Almasry resv->region_cache_count; 6280db9d74eSMina Almasry 6290db9d74eSMina Almasry /* At this point, we should have enough entries in the cache 630f0953a1bSIngo Molnar * for all the existing adds_in_progress. We should only be 6310db9d74eSMina Almasry * needing to allocate for regions_needed. 6320db9d74eSMina Almasry */ 6330db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); 6340db9d74eSMina Almasry 6350db9d74eSMina Almasry spin_unlock(&resv->lock); 6360db9d74eSMina Almasry for (i = 0; i < to_allocate; i++) { 6370db9d74eSMina Almasry trg = kmalloc(sizeof(*trg), GFP_KERNEL); 6380db9d74eSMina Almasry if (!trg) 6390db9d74eSMina Almasry goto out_of_memory; 6400db9d74eSMina Almasry list_add(&trg->link, &allocated_regions); 6410db9d74eSMina Almasry } 6420db9d74eSMina Almasry 6430db9d74eSMina Almasry spin_lock(&resv->lock); 6440db9d74eSMina Almasry 645d3ec7b6eSWei Yang list_splice(&allocated_regions, &resv->region_cache); 646d3ec7b6eSWei Yang resv->region_cache_count += to_allocate; 6470db9d74eSMina Almasry } 6480db9d74eSMina Almasry 6490db9d74eSMina Almasry return 0; 6500db9d74eSMina Almasry 6510db9d74eSMina Almasry out_of_memory: 6520db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, &allocated_regions, link) { 653d75c6af9SMina Almasry list_del(&rg->link); 654d75c6af9SMina Almasry kfree(rg); 655d75c6af9SMina Almasry } 6560db9d74eSMina Almasry return -ENOMEM; 657d75c6af9SMina Almasry } 658d75c6af9SMina Almasry 6591dd308a7SMike Kravetz /* 6601dd308a7SMike Kravetz * Add the huge page range represented by [f, t) to the reserve 6610db9d74eSMina Almasry * map. Regions will be taken from the cache to fill in this range. 6620db9d74eSMina Almasry * Sufficient regions should exist in the cache due to the previous 6630db9d74eSMina Almasry * call to region_chg with the same range, but in some cases the cache will not 6640db9d74eSMina Almasry * have sufficient entries due to races with other code doing region_add or 6650db9d74eSMina Almasry * region_del. The extra needed entries will be allocated. 666cf3ad20bSMike Kravetz * 6670db9d74eSMina Almasry * regions_needed is the out value provided by a previous call to region_chg. 6680db9d74eSMina Almasry * 6690db9d74eSMina Almasry * Return the number of new huge pages added to the map. This number is greater 6700db9d74eSMina Almasry * than or equal to zero. If file_region entries needed to be allocated for 6717c8de358SEthon Paul * this operation and we were not able to allocate, it returns -ENOMEM. 6720db9d74eSMina Almasry * region_add of regions of length 1 never allocate file_regions and cannot 6730db9d74eSMina Almasry * fail; region_chg will always allocate at least 1 entry and a region_add for 6740db9d74eSMina Almasry * 1 page will only require at most 1 entry. 6751dd308a7SMike Kravetz */ 6760db9d74eSMina Almasry static long region_add(struct resv_map *resv, long f, long t, 677075a61d0SMina Almasry long in_regions_needed, struct hstate *h, 678075a61d0SMina Almasry struct hugetlb_cgroup *h_cg) 67996822904SAndy Whitcroft { 6800db9d74eSMina Almasry long add = 0, actual_regions_needed = 0; 68196822904SAndy Whitcroft 6827b24d861SDavidlohr Bueso spin_lock(&resv->lock); 6830db9d74eSMina Almasry retry: 6840db9d74eSMina Almasry 6850db9d74eSMina Almasry /* Count how many regions are actually needed to execute this add. */ 686972a3da3SWei Yang add_reservation_in_range(resv, f, t, NULL, NULL, 687972a3da3SWei Yang &actual_regions_needed); 68896822904SAndy Whitcroft 6895e911373SMike Kravetz /* 6900db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 6910db9d74eSMina Almasry * this add operation. Note that actual_regions_needed may be greater 6920db9d74eSMina Almasry * than in_regions_needed, as the resv_map may have been modified since 6930db9d74eSMina Almasry * the region_chg call. In this case, we need to make sure that we 6940db9d74eSMina Almasry * allocate extra entries, such that we have enough for all the 6950db9d74eSMina Almasry * existing adds_in_progress, plus the excess needed for this 6960db9d74eSMina Almasry * operation. 6975e911373SMike Kravetz */ 6980db9d74eSMina Almasry if (actual_regions_needed > in_regions_needed && 6990db9d74eSMina Almasry resv->region_cache_count < 7000db9d74eSMina Almasry resv->adds_in_progress + 7010db9d74eSMina Almasry (actual_regions_needed - in_regions_needed)) { 7020db9d74eSMina Almasry /* region_add operation of range 1 should never need to 7030db9d74eSMina Almasry * allocate file_region entries. 7040db9d74eSMina Almasry */ 7050db9d74eSMina Almasry VM_BUG_ON(t - f <= 1); 7065e911373SMike Kravetz 7070db9d74eSMina Almasry if (allocate_file_region_entries( 7080db9d74eSMina Almasry resv, actual_regions_needed - in_regions_needed)) { 7090db9d74eSMina Almasry return -ENOMEM; 7105e911373SMike Kravetz } 7115e911373SMike Kravetz 7120db9d74eSMina Almasry goto retry; 7130db9d74eSMina Almasry } 714cf3ad20bSMike Kravetz 715972a3da3SWei Yang add = add_reservation_in_range(resv, f, t, h_cg, h, NULL); 7160db9d74eSMina Almasry 7170db9d74eSMina Almasry resv->adds_in_progress -= in_regions_needed; 7180db9d74eSMina Almasry 7197b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 720cf3ad20bSMike Kravetz return add; 72196822904SAndy Whitcroft } 72296822904SAndy Whitcroft 7231dd308a7SMike Kravetz /* 7241dd308a7SMike Kravetz * Examine the existing reserve map and determine how many 7251dd308a7SMike Kravetz * huge pages in the specified range [f, t) are NOT currently 7261dd308a7SMike Kravetz * represented. This routine is called before a subsequent 7271dd308a7SMike Kravetz * call to region_add that will actually modify the reserve 7281dd308a7SMike Kravetz * map to add the specified range [f, t). region_chg does 7291dd308a7SMike Kravetz * not change the number of huge pages represented by the 7300db9d74eSMina Almasry * map. A number of new file_region structures is added to the cache as a 7310db9d74eSMina Almasry * placeholder, for the subsequent region_add call to use. At least 1 7320db9d74eSMina Almasry * file_region structure is added. 7330db9d74eSMina Almasry * 7340db9d74eSMina Almasry * out_regions_needed is the number of regions added to the 7350db9d74eSMina Almasry * resv->adds_in_progress. This value needs to be provided to a follow up call 7360db9d74eSMina Almasry * to region_add or region_abort for proper accounting. 7375e911373SMike Kravetz * 7385e911373SMike Kravetz * Returns the number of huge pages that need to be added to the existing 7395e911373SMike Kravetz * reservation map for the range [f, t). This number is greater or equal to 7405e911373SMike Kravetz * zero. -ENOMEM is returned if a new file_region structure or cache entry 7415e911373SMike Kravetz * is needed and can not be allocated. 7421dd308a7SMike Kravetz */ 7430db9d74eSMina Almasry static long region_chg(struct resv_map *resv, long f, long t, 7440db9d74eSMina Almasry long *out_regions_needed) 74596822904SAndy Whitcroft { 74696822904SAndy Whitcroft long chg = 0; 74796822904SAndy Whitcroft 7487b24d861SDavidlohr Bueso spin_lock(&resv->lock); 7495e911373SMike Kravetz 750972a3da3SWei Yang /* Count how many hugepages in this range are NOT represented. */ 751075a61d0SMina Almasry chg = add_reservation_in_range(resv, f, t, NULL, NULL, 752972a3da3SWei Yang out_regions_needed); 7535e911373SMike Kravetz 7540db9d74eSMina Almasry if (*out_regions_needed == 0) 7550db9d74eSMina Almasry *out_regions_needed = 1; 7565e911373SMike Kravetz 7570db9d74eSMina Almasry if (allocate_file_region_entries(resv, *out_regions_needed)) 7585e911373SMike Kravetz return -ENOMEM; 7595e911373SMike Kravetz 7600db9d74eSMina Almasry resv->adds_in_progress += *out_regions_needed; 76196822904SAndy Whitcroft 7627b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 76396822904SAndy Whitcroft return chg; 76496822904SAndy Whitcroft } 76596822904SAndy Whitcroft 7661dd308a7SMike Kravetz /* 7675e911373SMike Kravetz * Abort the in progress add operation. The adds_in_progress field 7685e911373SMike Kravetz * of the resv_map keeps track of the operations in progress between 7695e911373SMike Kravetz * calls to region_chg and region_add. Operations are sometimes 7705e911373SMike Kravetz * aborted after the call to region_chg. In such cases, region_abort 7710db9d74eSMina Almasry * is called to decrement the adds_in_progress counter. regions_needed 7720db9d74eSMina Almasry * is the value returned by the region_chg call, it is used to decrement 7730db9d74eSMina Almasry * the adds_in_progress counter. 7745e911373SMike Kravetz * 7755e911373SMike Kravetz * NOTE: The range arguments [f, t) are not needed or used in this 7765e911373SMike Kravetz * routine. They are kept to make reading the calling code easier as 7775e911373SMike Kravetz * arguments will match the associated region_chg call. 7785e911373SMike Kravetz */ 7790db9d74eSMina Almasry static void region_abort(struct resv_map *resv, long f, long t, 7800db9d74eSMina Almasry long regions_needed) 7815e911373SMike Kravetz { 7825e911373SMike Kravetz spin_lock(&resv->lock); 7835e911373SMike Kravetz VM_BUG_ON(!resv->region_cache_count); 7840db9d74eSMina Almasry resv->adds_in_progress -= regions_needed; 7855e911373SMike Kravetz spin_unlock(&resv->lock); 7865e911373SMike Kravetz } 7875e911373SMike Kravetz 7885e911373SMike Kravetz /* 789feba16e2SMike Kravetz * Delete the specified range [f, t) from the reserve map. If the 790feba16e2SMike Kravetz * t parameter is LONG_MAX, this indicates that ALL regions after f 791feba16e2SMike Kravetz * should be deleted. Locate the regions which intersect [f, t) 792feba16e2SMike Kravetz * and either trim, delete or split the existing regions. 793feba16e2SMike Kravetz * 794feba16e2SMike Kravetz * Returns the number of huge pages deleted from the reserve map. 795feba16e2SMike Kravetz * In the normal case, the return value is zero or more. In the 796feba16e2SMike Kravetz * case where a region must be split, a new region descriptor must 797feba16e2SMike Kravetz * be allocated. If the allocation fails, -ENOMEM will be returned. 798feba16e2SMike Kravetz * NOTE: If the parameter t == LONG_MAX, then we will never split 799feba16e2SMike Kravetz * a region and possibly return -ENOMEM. Callers specifying 800feba16e2SMike Kravetz * t == LONG_MAX do not need to check for -ENOMEM error. 8011dd308a7SMike Kravetz */ 802feba16e2SMike Kravetz static long region_del(struct resv_map *resv, long f, long t) 80396822904SAndy Whitcroft { 8041406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 80596822904SAndy Whitcroft struct file_region *rg, *trg; 806feba16e2SMike Kravetz struct file_region *nrg = NULL; 807feba16e2SMike Kravetz long del = 0; 80896822904SAndy Whitcroft 809feba16e2SMike Kravetz retry: 8107b24d861SDavidlohr Bueso spin_lock(&resv->lock); 811feba16e2SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 812dbe409e4SMike Kravetz /* 813dbe409e4SMike Kravetz * Skip regions before the range to be deleted. file_region 814dbe409e4SMike Kravetz * ranges are normally of the form [from, to). However, there 815dbe409e4SMike Kravetz * may be a "placeholder" entry in the map which is of the form 816dbe409e4SMike Kravetz * (from, to) with from == to. Check for placeholder entries 817dbe409e4SMike Kravetz * at the beginning of the range to be deleted. 818dbe409e4SMike Kravetz */ 819dbe409e4SMike Kravetz if (rg->to <= f && (rg->to != rg->from || rg->to != f)) 820feba16e2SMike Kravetz continue; 821dbe409e4SMike Kravetz 822feba16e2SMike Kravetz if (rg->from >= t) 82396822904SAndy Whitcroft break; 82496822904SAndy Whitcroft 825feba16e2SMike Kravetz if (f > rg->from && t < rg->to) { /* Must split region */ 826feba16e2SMike Kravetz /* 827feba16e2SMike Kravetz * Check for an entry in the cache before dropping 828feba16e2SMike Kravetz * lock and attempting allocation. 829feba16e2SMike Kravetz */ 830feba16e2SMike Kravetz if (!nrg && 831feba16e2SMike Kravetz resv->region_cache_count > resv->adds_in_progress) { 832feba16e2SMike Kravetz nrg = list_first_entry(&resv->region_cache, 833feba16e2SMike Kravetz struct file_region, 834feba16e2SMike Kravetz link); 835feba16e2SMike Kravetz list_del(&nrg->link); 836feba16e2SMike Kravetz resv->region_cache_count--; 83796822904SAndy Whitcroft } 83896822904SAndy Whitcroft 839feba16e2SMike Kravetz if (!nrg) { 840feba16e2SMike Kravetz spin_unlock(&resv->lock); 841feba16e2SMike Kravetz nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 842feba16e2SMike Kravetz if (!nrg) 843feba16e2SMike Kravetz return -ENOMEM; 844feba16e2SMike Kravetz goto retry; 845feba16e2SMike Kravetz } 846feba16e2SMike Kravetz 847feba16e2SMike Kravetz del += t - f; 84879aa925bSMike Kravetz hugetlb_cgroup_uncharge_file_region( 849d85aecf2SMiaohe Lin resv, rg, t - f, false); 850feba16e2SMike Kravetz 851feba16e2SMike Kravetz /* New entry for end of split region */ 852feba16e2SMike Kravetz nrg->from = t; 853feba16e2SMike Kravetz nrg->to = rg->to; 854075a61d0SMina Almasry 855075a61d0SMina Almasry copy_hugetlb_cgroup_uncharge_info(nrg, rg); 856075a61d0SMina Almasry 857feba16e2SMike Kravetz INIT_LIST_HEAD(&nrg->link); 858feba16e2SMike Kravetz 859feba16e2SMike Kravetz /* Original entry is trimmed */ 860feba16e2SMike Kravetz rg->to = f; 861feba16e2SMike Kravetz 862feba16e2SMike Kravetz list_add(&nrg->link, &rg->link); 863feba16e2SMike Kravetz nrg = NULL; 86496822904SAndy Whitcroft break; 865feba16e2SMike Kravetz } 866feba16e2SMike Kravetz 867feba16e2SMike Kravetz if (f <= rg->from && t >= rg->to) { /* Remove entire region */ 868feba16e2SMike Kravetz del += rg->to - rg->from; 869075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 870d85aecf2SMiaohe Lin rg->to - rg->from, true); 87196822904SAndy Whitcroft list_del(&rg->link); 87296822904SAndy Whitcroft kfree(rg); 873feba16e2SMike Kravetz continue; 87496822904SAndy Whitcroft } 8757b24d861SDavidlohr Bueso 876feba16e2SMike Kravetz if (f <= rg->from) { /* Trim beginning of region */ 877075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 878d85aecf2SMiaohe Lin t - rg->from, false); 879075a61d0SMina Almasry 88079aa925bSMike Kravetz del += t - rg->from; 88179aa925bSMike Kravetz rg->from = t; 88279aa925bSMike Kravetz } else { /* Trim end of region */ 883075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 884d85aecf2SMiaohe Lin rg->to - f, false); 88579aa925bSMike Kravetz 88679aa925bSMike Kravetz del += rg->to - f; 88779aa925bSMike Kravetz rg->to = f; 888feba16e2SMike Kravetz } 889feba16e2SMike Kravetz } 890feba16e2SMike Kravetz 8917b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 892feba16e2SMike Kravetz kfree(nrg); 893feba16e2SMike Kravetz return del; 89496822904SAndy Whitcroft } 89596822904SAndy Whitcroft 8961dd308a7SMike Kravetz /* 897b5cec28dSMike Kravetz * A rare out of memory error was encountered which prevented removal of 898b5cec28dSMike Kravetz * the reserve map region for a page. The huge page itself was free'ed 899b5cec28dSMike Kravetz * and removed from the page cache. This routine will adjust the subpool 900b5cec28dSMike Kravetz * usage count, and the global reserve count if needed. By incrementing 901b5cec28dSMike Kravetz * these counts, the reserve map entry which could not be deleted will 902b5cec28dSMike Kravetz * appear as a "reserved" entry instead of simply dangling with incorrect 903b5cec28dSMike Kravetz * counts. 904b5cec28dSMike Kravetz */ 90572e2936cSzhong jiang void hugetlb_fix_reserve_counts(struct inode *inode) 906b5cec28dSMike Kravetz { 907b5cec28dSMike Kravetz struct hugepage_subpool *spool = subpool_inode(inode); 908b5cec28dSMike Kravetz long rsv_adjust; 909da56388cSMiaohe Lin bool reserved = false; 910b5cec28dSMike Kravetz 911b5cec28dSMike Kravetz rsv_adjust = hugepage_subpool_get_pages(spool, 1); 912da56388cSMiaohe Lin if (rsv_adjust > 0) { 913b5cec28dSMike Kravetz struct hstate *h = hstate_inode(inode); 914b5cec28dSMike Kravetz 915da56388cSMiaohe Lin if (!hugetlb_acct_memory(h, 1)) 916da56388cSMiaohe Lin reserved = true; 917da56388cSMiaohe Lin } else if (!rsv_adjust) { 918da56388cSMiaohe Lin reserved = true; 919b5cec28dSMike Kravetz } 920da56388cSMiaohe Lin 921da56388cSMiaohe Lin if (!reserved) 922da56388cSMiaohe Lin pr_warn("hugetlb: Huge Page Reserved count may go negative.\n"); 923b5cec28dSMike Kravetz } 924b5cec28dSMike Kravetz 925b5cec28dSMike Kravetz /* 9261dd308a7SMike Kravetz * Count and return the number of huge pages in the reserve map 9271dd308a7SMike Kravetz * that intersect with the range [f, t). 9281dd308a7SMike Kravetz */ 9291406ec9bSJoonsoo Kim static long region_count(struct resv_map *resv, long f, long t) 93084afd99bSAndy Whitcroft { 9311406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 93284afd99bSAndy Whitcroft struct file_region *rg; 93384afd99bSAndy Whitcroft long chg = 0; 93484afd99bSAndy Whitcroft 9357b24d861SDavidlohr Bueso spin_lock(&resv->lock); 93684afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 93784afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 938f2135a4aSWang Sheng-Hui long seg_from; 939f2135a4aSWang Sheng-Hui long seg_to; 94084afd99bSAndy Whitcroft 94184afd99bSAndy Whitcroft if (rg->to <= f) 94284afd99bSAndy Whitcroft continue; 94384afd99bSAndy Whitcroft if (rg->from >= t) 94484afd99bSAndy Whitcroft break; 94584afd99bSAndy Whitcroft 94684afd99bSAndy Whitcroft seg_from = max(rg->from, f); 94784afd99bSAndy Whitcroft seg_to = min(rg->to, t); 94884afd99bSAndy Whitcroft 94984afd99bSAndy Whitcroft chg += seg_to - seg_from; 95084afd99bSAndy Whitcroft } 9517b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 95284afd99bSAndy Whitcroft 95384afd99bSAndy Whitcroft return chg; 95484afd99bSAndy Whitcroft } 95584afd99bSAndy Whitcroft 95696822904SAndy Whitcroft /* 957e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 958e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 959e7c4b0bfSAndy Whitcroft */ 960a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 961a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 962e7c4b0bfSAndy Whitcroft { 963a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 964a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 965e7c4b0bfSAndy Whitcroft } 966e7c4b0bfSAndy Whitcroft 9670fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 9680fe6e20bSNaoya Horiguchi unsigned long address) 9690fe6e20bSNaoya Horiguchi { 9700fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 9710fe6e20bSNaoya Horiguchi } 972dee41079SDan Williams EXPORT_SYMBOL_GPL(linear_hugepage_index); 9730fe6e20bSNaoya Horiguchi 97484afd99bSAndy Whitcroft /* 97508fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 97608fba699SMel Gorman * cases this will be same size as used by the page table entries. 97708fba699SMel Gorman */ 97808fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 97908fba699SMel Gorman { 98005ea8860SDan Williams if (vma->vm_ops && vma->vm_ops->pagesize) 98105ea8860SDan Williams return vma->vm_ops->pagesize(vma); 98208fba699SMel Gorman return PAGE_SIZE; 98308fba699SMel Gorman } 984f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 98508fba699SMel Gorman 98608fba699SMel Gorman /* 9873340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 9883340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 98909135cc5SDan Williams * architectures where it differs, an architecture-specific 'strong' 99009135cc5SDan Williams * version of this symbol is required. 9913340289dSMel Gorman */ 99209135cc5SDan Williams __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 9933340289dSMel Gorman { 9943340289dSMel Gorman return vma_kernel_pagesize(vma); 9953340289dSMel Gorman } 9963340289dSMel Gorman 9973340289dSMel Gorman /* 99884afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 99984afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 100084afd99bSAndy Whitcroft * alignment. 100184afd99bSAndy Whitcroft */ 100284afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 100384afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 100404f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 100584afd99bSAndy Whitcroft 1006a1e78772SMel Gorman /* 1007a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 1008a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 1009a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 1010a1e78772SMel Gorman * 10118d9bfb26SMike Kravetz * With the exception of hugetlb_dup_vma_private() which is called at fork(), 1012a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 1013a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 1014a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 101584afd99bSAndy Whitcroft * 101684afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 101784afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 101884afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 101984afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 102084afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 102184afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 102284afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 102384afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 1024a1e78772SMel Gorman */ 1025e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 1026e7c4b0bfSAndy Whitcroft { 1027e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 1028e7c4b0bfSAndy Whitcroft } 1029e7c4b0bfSAndy Whitcroft 1030e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 1031e7c4b0bfSAndy Whitcroft unsigned long value) 1032e7c4b0bfSAndy Whitcroft { 1033e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 1034e7c4b0bfSAndy Whitcroft } 1035e7c4b0bfSAndy Whitcroft 1036e9fe92aeSMina Almasry static void 1037e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, 1038e9fe92aeSMina Almasry struct hugetlb_cgroup *h_cg, 1039e9fe92aeSMina Almasry struct hstate *h) 1040e9fe92aeSMina Almasry { 1041e9fe92aeSMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 1042e9fe92aeSMina Almasry if (!h_cg || !h) { 1043e9fe92aeSMina Almasry resv_map->reservation_counter = NULL; 1044e9fe92aeSMina Almasry resv_map->pages_per_hpage = 0; 1045e9fe92aeSMina Almasry resv_map->css = NULL; 1046e9fe92aeSMina Almasry } else { 1047e9fe92aeSMina Almasry resv_map->reservation_counter = 1048e9fe92aeSMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 1049e9fe92aeSMina Almasry resv_map->pages_per_hpage = pages_per_huge_page(h); 1050e9fe92aeSMina Almasry resv_map->css = &h_cg->css; 1051e9fe92aeSMina Almasry } 1052e9fe92aeSMina Almasry #endif 1053e9fe92aeSMina Almasry } 1054e9fe92aeSMina Almasry 10559119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 105684afd99bSAndy Whitcroft { 105784afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 10585e911373SMike Kravetz struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 10595e911373SMike Kravetz 10605e911373SMike Kravetz if (!resv_map || !rg) { 10615e911373SMike Kravetz kfree(resv_map); 10625e911373SMike Kravetz kfree(rg); 106384afd99bSAndy Whitcroft return NULL; 10645e911373SMike Kravetz } 106584afd99bSAndy Whitcroft 106684afd99bSAndy Whitcroft kref_init(&resv_map->refs); 10677b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 106884afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 106984afd99bSAndy Whitcroft 10705e911373SMike Kravetz resv_map->adds_in_progress = 0; 1071e9fe92aeSMina Almasry /* 1072e9fe92aeSMina Almasry * Initialize these to 0. On shared mappings, 0's here indicate these 1073e9fe92aeSMina Almasry * fields don't do cgroup accounting. On private mappings, these will be 1074e9fe92aeSMina Almasry * re-initialized to the proper values, to indicate that hugetlb cgroup 1075e9fe92aeSMina Almasry * reservations are to be un-charged from here. 1076e9fe92aeSMina Almasry */ 1077e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); 10785e911373SMike Kravetz 10795e911373SMike Kravetz INIT_LIST_HEAD(&resv_map->region_cache); 10805e911373SMike Kravetz list_add(&rg->link, &resv_map->region_cache); 10815e911373SMike Kravetz resv_map->region_cache_count = 1; 10825e911373SMike Kravetz 108384afd99bSAndy Whitcroft return resv_map; 108484afd99bSAndy Whitcroft } 108584afd99bSAndy Whitcroft 10869119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 108784afd99bSAndy Whitcroft { 108884afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 10895e911373SMike Kravetz struct list_head *head = &resv_map->region_cache; 10905e911373SMike Kravetz struct file_region *rg, *trg; 109184afd99bSAndy Whitcroft 109284afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 1093feba16e2SMike Kravetz region_del(resv_map, 0, LONG_MAX); 10945e911373SMike Kravetz 10955e911373SMike Kravetz /* ... and any entries left in the cache */ 10965e911373SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 10975e911373SMike Kravetz list_del(&rg->link); 10985e911373SMike Kravetz kfree(rg); 10995e911373SMike Kravetz } 11005e911373SMike Kravetz 11015e911373SMike Kravetz VM_BUG_ON(resv_map->adds_in_progress); 11025e911373SMike Kravetz 110384afd99bSAndy Whitcroft kfree(resv_map); 110484afd99bSAndy Whitcroft } 110584afd99bSAndy Whitcroft 11064e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 11074e35f483SJoonsoo Kim { 1108f27a5136SMike Kravetz /* 1109f27a5136SMike Kravetz * At inode evict time, i_mapping may not point to the original 1110f27a5136SMike Kravetz * address space within the inode. This original address space 1111f27a5136SMike Kravetz * contains the pointer to the resv_map. So, always use the 1112f27a5136SMike Kravetz * address space embedded within the inode. 1113f27a5136SMike Kravetz * The VERY common case is inode->mapping == &inode->i_data but, 1114f27a5136SMike Kravetz * this may not be true for device special inodes. 1115f27a5136SMike Kravetz */ 1116f27a5136SMike Kravetz return (struct resv_map *)(&inode->i_data)->private_data; 11174e35f483SJoonsoo Kim } 11184e35f483SJoonsoo Kim 111984afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 1120a1e78772SMel Gorman { 112181d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 11224e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 11234e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 11244e35f483SJoonsoo Kim struct inode *inode = mapping->host; 11254e35f483SJoonsoo Kim 11264e35f483SJoonsoo Kim return inode_resv_map(inode); 11274e35f483SJoonsoo Kim 11284e35f483SJoonsoo Kim } else { 112984afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 113084afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 11314e35f483SJoonsoo Kim } 1132a1e78772SMel Gorman } 1133a1e78772SMel Gorman 113484afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 1135a1e78772SMel Gorman { 113681d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 113781d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 1138a1e78772SMel Gorman 113984afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 114084afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 114104f2cbe3SMel Gorman } 114204f2cbe3SMel Gorman 114304f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 114404f2cbe3SMel Gorman { 114581d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 114681d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 1147e7c4b0bfSAndy Whitcroft 1148e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 114904f2cbe3SMel Gorman } 115004f2cbe3SMel Gorman 115104f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 115204f2cbe3SMel Gorman { 115381d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 1154e7c4b0bfSAndy Whitcroft 1155e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 1156a1e78772SMel Gorman } 1157a1e78772SMel Gorman 11588d9bfb26SMike Kravetz void hugetlb_dup_vma_private(struct vm_area_struct *vma) 1159a1e78772SMel Gorman { 116081d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 11618d9bfb26SMike Kravetz /* 11628d9bfb26SMike Kravetz * Clear vm_private_data 1163612b8a31SMike Kravetz * - For shared mappings this is a per-vma semaphore that may be 1164612b8a31SMike Kravetz * allocated in a subsequent call to hugetlb_vm_op_open. 1165612b8a31SMike Kravetz * Before clearing, make sure pointer is not associated with vma 1166612b8a31SMike Kravetz * as this will leak the structure. This is the case when called 1167612b8a31SMike Kravetz * via clear_vma_resv_huge_pages() and hugetlb_vm_op_open has already 1168612b8a31SMike Kravetz * been called to allocate a new structure. 11698d9bfb26SMike Kravetz * - For MAP_PRIVATE mappings, this is the reserve map which does 11708d9bfb26SMike Kravetz * not apply to children. Faults generated by the children are 11718d9bfb26SMike Kravetz * not guaranteed to succeed, even if read-only. 11728d9bfb26SMike Kravetz */ 1173612b8a31SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 1174612b8a31SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 1175612b8a31SMike Kravetz 1176612b8a31SMike Kravetz if (vma_lock && vma_lock->vma != vma) 1177612b8a31SMike Kravetz vma->vm_private_data = NULL; 1178612b8a31SMike Kravetz } else 1179612b8a31SMike Kravetz vma->vm_private_data = NULL; 1180a1e78772SMel Gorman } 1181a1e78772SMel Gorman 1182550a7d60SMina Almasry /* 1183550a7d60SMina Almasry * Reset and decrement one ref on hugepage private reservation. 1184550a7d60SMina Almasry * Called with mm->mmap_sem writer semaphore held. 1185550a7d60SMina Almasry * This function should be only used by move_vma() and operate on 1186550a7d60SMina Almasry * same sized vma. It should never come here with last ref on the 1187550a7d60SMina Almasry * reservation. 1188550a7d60SMina Almasry */ 1189550a7d60SMina Almasry void clear_vma_resv_huge_pages(struct vm_area_struct *vma) 1190550a7d60SMina Almasry { 1191550a7d60SMina Almasry /* 1192550a7d60SMina Almasry * Clear the old hugetlb private page reservation. 1193550a7d60SMina Almasry * It has already been transferred to new_vma. 1194550a7d60SMina Almasry * 1195550a7d60SMina Almasry * During a mremap() operation of a hugetlb vma we call move_vma() 1196550a7d60SMina Almasry * which copies vma into new_vma and unmaps vma. After the copy 1197550a7d60SMina Almasry * operation both new_vma and vma share a reference to the resv_map 1198550a7d60SMina Almasry * struct, and at that point vma is about to be unmapped. We don't 1199550a7d60SMina Almasry * want to return the reservation to the pool at unmap of vma because 1200550a7d60SMina Almasry * the reservation still lives on in new_vma, so simply decrement the 1201550a7d60SMina Almasry * ref here and remove the resv_map reference from this vma. 1202550a7d60SMina Almasry */ 1203550a7d60SMina Almasry struct resv_map *reservations = vma_resv_map(vma); 1204550a7d60SMina Almasry 1205afe041c2SBui Quang Minh if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 1206afe041c2SBui Quang Minh resv_map_put_hugetlb_cgroup_uncharge_info(reservations); 1207550a7d60SMina Almasry kref_put(&reservations->refs, resv_map_release); 1208afe041c2SBui Quang Minh } 1209550a7d60SMina Almasry 12108d9bfb26SMike Kravetz hugetlb_dup_vma_private(vma); 1211550a7d60SMina Almasry } 1212550a7d60SMina Almasry 1213a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 1214559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 1215a1e78772SMel Gorman { 1216af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 1217af0ed73eSJoonsoo Kim /* 1218af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 1219af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 1220af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 1221af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 1222af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 1223af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 1224af0ed73eSJoonsoo Kim * properly, so add work-around here. 1225af0ed73eSJoonsoo Kim */ 1226af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 1227559ec2f8SNicholas Krause return true; 1228af0ed73eSJoonsoo Kim else 1229559ec2f8SNicholas Krause return false; 1230af0ed73eSJoonsoo Kim } 1231a63884e9SJoonsoo Kim 1232a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 12331fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 12341fb1b0e9SMike Kravetz /* 12351fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 12361fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 12371fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 12387c8de358SEthon Paul * fallocate. In this case, there really are no reserves to 12391fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 12401fb1b0e9SMike Kravetz */ 12411fb1b0e9SMike Kravetz if (chg) 12421fb1b0e9SMike Kravetz return false; 12431fb1b0e9SMike Kravetz else 1244559ec2f8SNicholas Krause return true; 12451fb1b0e9SMike Kravetz } 1246a63884e9SJoonsoo Kim 1247a63884e9SJoonsoo Kim /* 1248a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 1249a63884e9SJoonsoo Kim * private mappings. 1250a63884e9SJoonsoo Kim */ 125167961f9dSMike Kravetz if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 125267961f9dSMike Kravetz /* 125367961f9dSMike Kravetz * Like the shared case above, a hole punch or truncate 125467961f9dSMike Kravetz * could have been performed on the private mapping. 125567961f9dSMike Kravetz * Examine the value of chg to determine if reserves 125667961f9dSMike Kravetz * actually exist or were previously consumed. 125767961f9dSMike Kravetz * Very Subtle - The value of chg comes from a previous 125867961f9dSMike Kravetz * call to vma_needs_reserves(). The reserve map for 125967961f9dSMike Kravetz * private mappings has different (opposite) semantics 126067961f9dSMike Kravetz * than that of shared mappings. vma_needs_reserves() 126167961f9dSMike Kravetz * has already taken this difference in semantics into 126267961f9dSMike Kravetz * account. Therefore, the meaning of chg is the same 126367961f9dSMike Kravetz * as in the shared case above. Code could easily be 126467961f9dSMike Kravetz * combined, but keeping it separate draws attention to 126567961f9dSMike Kravetz * subtle differences. 126667961f9dSMike Kravetz */ 126767961f9dSMike Kravetz if (chg) 126867961f9dSMike Kravetz return false; 126967961f9dSMike Kravetz else 1270559ec2f8SNicholas Krause return true; 127167961f9dSMike Kravetz } 1272a63884e9SJoonsoo Kim 1273559ec2f8SNicholas Krause return false; 1274a1e78772SMel Gorman } 1275a1e78772SMel Gorman 1276240d67a8SSidhartha Kumar static void enqueue_hugetlb_folio(struct hstate *h, struct folio *folio) 12771da177e4SLinus Torvalds { 1278240d67a8SSidhartha Kumar int nid = folio_nid(folio); 12799487ca60SMike Kravetz 12809487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1281240d67a8SSidhartha Kumar VM_BUG_ON_FOLIO(folio_ref_count(folio), folio); 1282b65a4edaSMike Kravetz 1283240d67a8SSidhartha Kumar list_move(&folio->lru, &h->hugepage_freelists[nid]); 1284a5516438SAndi Kleen h->free_huge_pages++; 1285a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 1286240d67a8SSidhartha Kumar folio_set_hugetlb_freed(folio); 12871da177e4SLinus Torvalds } 12881da177e4SLinus Torvalds 128994310cbcSAnshuman Khandual static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) 1290bf50bab2SNaoya Horiguchi { 1291bf50bab2SNaoya Horiguchi struct page *page; 12921a08ae36SPavel Tatashin bool pin = !!(current->flags & PF_MEMALLOC_PIN); 1293bf50bab2SNaoya Horiguchi 12949487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1295bbe88753SJoonsoo Kim list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { 12966077c943SAlex Sierra if (pin && !is_longterm_pinnable_page(page)) 1297bbe88753SJoonsoo Kim continue; 1298bbe88753SJoonsoo Kim 12996664bfc8SWei Yang if (PageHWPoison(page)) 13006664bfc8SWei Yang continue; 1301bbe88753SJoonsoo Kim 13020edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 1303a9869b83SNaoya Horiguchi set_page_refcounted(page); 13046c037149SMike Kravetz ClearHPageFreed(page); 1305bf50bab2SNaoya Horiguchi h->free_huge_pages--; 1306bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 1307bf50bab2SNaoya Horiguchi return page; 1308bf50bab2SNaoya Horiguchi } 1309bf50bab2SNaoya Horiguchi 13106664bfc8SWei Yang return NULL; 13116664bfc8SWei Yang } 13126664bfc8SWei Yang 13133e59fcb0SMichal Hocko static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, 13143e59fcb0SMichal Hocko nodemask_t *nmask) 131594310cbcSAnshuman Khandual { 13163e59fcb0SMichal Hocko unsigned int cpuset_mems_cookie; 13173e59fcb0SMichal Hocko struct zonelist *zonelist; 13183e59fcb0SMichal Hocko struct zone *zone; 13193e59fcb0SMichal Hocko struct zoneref *z; 132098fa15f3SAnshuman Khandual int node = NUMA_NO_NODE; 13213e59fcb0SMichal Hocko 13223e59fcb0SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 13233e59fcb0SMichal Hocko 13243e59fcb0SMichal Hocko retry_cpuset: 13253e59fcb0SMichal Hocko cpuset_mems_cookie = read_mems_allowed_begin(); 13263e59fcb0SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 132794310cbcSAnshuman Khandual struct page *page; 132894310cbcSAnshuman Khandual 13293e59fcb0SMichal Hocko if (!cpuset_zone_allowed(zone, gfp_mask)) 13303e59fcb0SMichal Hocko continue; 13313e59fcb0SMichal Hocko /* 13323e59fcb0SMichal Hocko * no need to ask again on the same node. Pool is node rather than 13333e59fcb0SMichal Hocko * zone aware 13343e59fcb0SMichal Hocko */ 13353e59fcb0SMichal Hocko if (zone_to_nid(zone) == node) 13363e59fcb0SMichal Hocko continue; 13373e59fcb0SMichal Hocko node = zone_to_nid(zone); 133894310cbcSAnshuman Khandual 133994310cbcSAnshuman Khandual page = dequeue_huge_page_node_exact(h, node); 134094310cbcSAnshuman Khandual if (page) 134194310cbcSAnshuman Khandual return page; 134294310cbcSAnshuman Khandual } 13433e59fcb0SMichal Hocko if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 13443e59fcb0SMichal Hocko goto retry_cpuset; 13453e59fcb0SMichal Hocko 134694310cbcSAnshuman Khandual return NULL; 134794310cbcSAnshuman Khandual } 134894310cbcSAnshuman Khandual 13498346d69dSXin Hao static unsigned long available_huge_pages(struct hstate *h) 13508346d69dSXin Hao { 13518346d69dSXin Hao return h->free_huge_pages - h->resv_huge_pages; 13528346d69dSXin Hao } 13538346d69dSXin Hao 1354a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 1355a5516438SAndi Kleen struct vm_area_struct *vma, 1356af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 1357af0ed73eSJoonsoo Kim long chg) 13581da177e4SLinus Torvalds { 1359cfcaa66fSBen Widawsky struct page *page = NULL; 1360480eccf9SLee Schermerhorn struct mempolicy *mpol; 136104ec6264SVlastimil Babka gfp_t gfp_mask; 13623e59fcb0SMichal Hocko nodemask_t *nodemask; 136304ec6264SVlastimil Babka int nid; 13641da177e4SLinus Torvalds 1365a1e78772SMel Gorman /* 1366a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 1367a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 1368a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 1369a1e78772SMel Gorman */ 13708346d69dSXin Hao if (!vma_has_reserves(vma, chg) && !available_huge_pages(h)) 1371c0ff7453SMiao Xie goto err; 1372a1e78772SMel Gorman 137304f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 13748346d69dSXin Hao if (avoid_reserve && !available_huge_pages(h)) 13756eab04a8SJustin P. Mattock goto err; 137604f2cbe3SMel Gorman 137704ec6264SVlastimil Babka gfp_mask = htlb_alloc_mask(h); 137804ec6264SVlastimil Babka nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 1379cfcaa66fSBen Widawsky 1380cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 13813e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 1382cfcaa66fSBen Widawsky 1383cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 1384cfcaa66fSBen Widawsky nodemask = NULL; 1385cfcaa66fSBen Widawsky } 1386cfcaa66fSBen Widawsky 1387cfcaa66fSBen Widawsky if (!page) 1388cfcaa66fSBen Widawsky page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 1389cfcaa66fSBen Widawsky 13903e59fcb0SMichal Hocko if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { 1391d6995da3SMike Kravetz SetHPageRestoreReserve(page); 1392a63884e9SJoonsoo Kim h->resv_huge_pages--; 1393bf50bab2SNaoya Horiguchi } 1394cc9a6c87SMel Gorman 1395cc9a6c87SMel Gorman mpol_cond_put(mpol); 1396cc9a6c87SMel Gorman return page; 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 1481911565b8SSidhartha Kumar atomic_set(folio_mapcount_ptr(folio), 0); 1482911565b8SSidhartha Kumar atomic_set(folio_subpages_mapcount_ptr(folio), 0); 1483911565b8SSidhartha Kumar atomic_set(folio_pincount_ptr(folio), 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_set_compound_order(folio, 0); 1494911565b8SSidhartha Kumar __folio_clear_head(folio); 1495944d9fecSLuiz Capitulino } 1496944d9fecSLuiz Capitulino 1497911565b8SSidhartha Kumar static void destroy_compound_hugetlb_folio_for_demote(struct folio *folio, 14988531fc6fSMike Kravetz unsigned int order) 14998531fc6fSMike Kravetz { 1500911565b8SSidhartha Kumar __destroy_compound_gigantic_folio(folio, order, true); 15018531fc6fSMike Kravetz } 15028531fc6fSMike Kravetz 15038531fc6fSMike Kravetz #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 1504911565b8SSidhartha Kumar static void destroy_compound_gigantic_folio(struct folio *folio, 150534d9e35bSMike Kravetz unsigned int order) 150634d9e35bSMike Kravetz { 1507911565b8SSidhartha Kumar __destroy_compound_gigantic_folio(folio, order, false); 150834d9e35bSMike Kravetz } 150934d9e35bSMike Kravetz 15107f325a8dSSidhartha Kumar static void free_gigantic_folio(struct folio *folio, unsigned int order) 1511944d9fecSLuiz Capitulino { 1512cf11e85fSRoman Gushchin /* 1513cf11e85fSRoman Gushchin * If the page isn't allocated using the cma allocator, 1514cf11e85fSRoman Gushchin * cma_release() returns false. 1515cf11e85fSRoman Gushchin */ 1516dbda8feaSBarry Song #ifdef CONFIG_CMA 15177f325a8dSSidhartha Kumar int nid = folio_nid(folio); 15187f325a8dSSidhartha Kumar 15197f325a8dSSidhartha Kumar if (cma_release(hugetlb_cma[nid], &folio->page, 1 << order)) 1520cf11e85fSRoman Gushchin return; 1521dbda8feaSBarry Song #endif 1522cf11e85fSRoman Gushchin 15237f325a8dSSidhartha Kumar free_contig_range(folio_pfn(folio), 1 << order); 1524944d9fecSLuiz Capitulino } 1525944d9fecSLuiz Capitulino 15264eb0716eSAlexandre Ghiti #ifdef CONFIG_CONTIG_ALLOC 152719fc1a7eSSidhartha Kumar static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask, 1528d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) 1529944d9fecSLuiz Capitulino { 153019fc1a7eSSidhartha Kumar struct page *page; 153104adbc3fSMiaohe Lin unsigned long nr_pages = pages_per_huge_page(h); 1532953f064aSLi Xinhai if (nid == NUMA_NO_NODE) 1533953f064aSLi Xinhai nid = numa_mem_id(); 1534944d9fecSLuiz Capitulino 1535dbda8feaSBarry Song #ifdef CONFIG_CMA 1536dbda8feaSBarry Song { 1537cf11e85fSRoman Gushchin int node; 1538cf11e85fSRoman Gushchin 1539953f064aSLi Xinhai if (hugetlb_cma[nid]) { 1540953f064aSLi Xinhai page = cma_alloc(hugetlb_cma[nid], nr_pages, 1541953f064aSLi Xinhai huge_page_order(h), true); 1542953f064aSLi Xinhai if (page) 154319fc1a7eSSidhartha Kumar return page_folio(page); 1544953f064aSLi Xinhai } 1545953f064aSLi Xinhai 1546953f064aSLi Xinhai if (!(gfp_mask & __GFP_THISNODE)) { 1547cf11e85fSRoman Gushchin for_each_node_mask(node, *nodemask) { 1548953f064aSLi Xinhai if (node == nid || !hugetlb_cma[node]) 1549cf11e85fSRoman Gushchin continue; 1550cf11e85fSRoman Gushchin 1551cf11e85fSRoman Gushchin page = cma_alloc(hugetlb_cma[node], nr_pages, 1552cf11e85fSRoman Gushchin huge_page_order(h), true); 1553cf11e85fSRoman Gushchin if (page) 155419fc1a7eSSidhartha Kumar return page_folio(page); 1555cf11e85fSRoman Gushchin } 1556cf11e85fSRoman Gushchin } 1557953f064aSLi Xinhai } 1558dbda8feaSBarry Song #endif 1559cf11e85fSRoman Gushchin 156019fc1a7eSSidhartha Kumar page = alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); 156119fc1a7eSSidhartha Kumar return page ? page_folio(page) : NULL; 1562944d9fecSLuiz Capitulino } 1563944d9fecSLuiz Capitulino 15644eb0716eSAlexandre Ghiti #else /* !CONFIG_CONTIG_ALLOC */ 156519fc1a7eSSidhartha Kumar static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask, 15664eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 15674eb0716eSAlexandre Ghiti { 15684eb0716eSAlexandre Ghiti return NULL; 15694eb0716eSAlexandre Ghiti } 15704eb0716eSAlexandre Ghiti #endif /* CONFIG_CONTIG_ALLOC */ 1571944d9fecSLuiz Capitulino 1572e1073d1eSAneesh Kumar K.V #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 157319fc1a7eSSidhartha Kumar static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask, 15744eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 15754eb0716eSAlexandre Ghiti { 15764eb0716eSAlexandre Ghiti return NULL; 15774eb0716eSAlexandre Ghiti } 15787f325a8dSSidhartha Kumar static inline void free_gigantic_folio(struct folio *folio, 15797f325a8dSSidhartha Kumar unsigned int order) { } 1580911565b8SSidhartha Kumar static inline void destroy_compound_gigantic_folio(struct folio *folio, 1581d00181b9SKirill A. Shutemov unsigned int order) { } 1582944d9fecSLuiz Capitulino #endif 1583944d9fecSLuiz Capitulino 15846eb4e88aSMike Kravetz /* 1585cfd5082bSSidhartha Kumar * Remove hugetlb folio from lists, and update dtor so that the folio appears 158634d9e35bSMike Kravetz * as just a compound page. 158734d9e35bSMike Kravetz * 1588cfd5082bSSidhartha Kumar * A reference is held on the folio, except in the case of demote. 15896eb4e88aSMike Kravetz * 15906eb4e88aSMike Kravetz * Must be called with hugetlb lock held. 15916eb4e88aSMike Kravetz */ 1592cfd5082bSSidhartha Kumar static void __remove_hugetlb_folio(struct hstate *h, struct folio *folio, 159334d9e35bSMike Kravetz bool adjust_surplus, 159434d9e35bSMike Kravetz bool demote) 15956eb4e88aSMike Kravetz { 1596cfd5082bSSidhartha Kumar int nid = folio_nid(folio); 15976eb4e88aSMike Kravetz 1598f074732dSSidhartha Kumar VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio(folio), folio); 1599f074732dSSidhartha Kumar VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio_rsvd(folio), folio); 16006eb4e88aSMike Kravetz 16019487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 16026eb4e88aSMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 16036eb4e88aSMike Kravetz return; 16046eb4e88aSMike Kravetz 1605cfd5082bSSidhartha Kumar list_del(&folio->lru); 16066eb4e88aSMike Kravetz 1607cfd5082bSSidhartha Kumar if (folio_test_hugetlb_freed(folio)) { 16086eb4e88aSMike Kravetz h->free_huge_pages--; 16096eb4e88aSMike Kravetz h->free_huge_pages_node[nid]--; 16106eb4e88aSMike Kravetz } 16116eb4e88aSMike Kravetz if (adjust_surplus) { 16126eb4e88aSMike Kravetz h->surplus_huge_pages--; 16136eb4e88aSMike Kravetz h->surplus_huge_pages_node[nid]--; 16146eb4e88aSMike Kravetz } 16156eb4e88aSMike Kravetz 1616e32d20c0SMike Kravetz /* 1617e32d20c0SMike Kravetz * Very subtle 1618e32d20c0SMike Kravetz * 1619e32d20c0SMike Kravetz * For non-gigantic pages set the destructor to the normal compound 1620e32d20c0SMike Kravetz * page dtor. This is needed in case someone takes an additional 1621e32d20c0SMike Kravetz * temporary ref to the page, and freeing is delayed until they drop 1622e32d20c0SMike Kravetz * their reference. 1623e32d20c0SMike Kravetz * 1624e32d20c0SMike Kravetz * For gigantic pages set the destructor to the null dtor. This 1625e32d20c0SMike Kravetz * destructor will never be called. Before freeing the gigantic 1626911565b8SSidhartha Kumar * page destroy_compound_gigantic_folio will turn the folio into a 1627911565b8SSidhartha Kumar * simple group of pages. After this the destructor does not 1628e32d20c0SMike Kravetz * apply. 1629e32d20c0SMike Kravetz * 1630e32d20c0SMike Kravetz * This handles the case where more than one ref is held when and 1631d6ef19e2SSidhartha Kumar * after update_and_free_hugetlb_folio is called. 163234d9e35bSMike Kravetz * 163334d9e35bSMike Kravetz * In the case of demote we do not ref count the page as it will soon 163434d9e35bSMike Kravetz * be turned into a page of smaller size. 1635e32d20c0SMike Kravetz */ 163634d9e35bSMike Kravetz if (!demote) 1637cfd5082bSSidhartha Kumar folio_ref_unfreeze(folio, 1); 1638e32d20c0SMike Kravetz if (hstate_is_gigantic(h)) 1639cfd5082bSSidhartha Kumar folio_set_compound_dtor(folio, NULL_COMPOUND_DTOR); 1640e32d20c0SMike Kravetz else 1641cfd5082bSSidhartha Kumar folio_set_compound_dtor(folio, COMPOUND_PAGE_DTOR); 16426eb4e88aSMike Kravetz 16436eb4e88aSMike Kravetz h->nr_huge_pages--; 16446eb4e88aSMike Kravetz h->nr_huge_pages_node[nid]--; 16456eb4e88aSMike Kravetz } 16466eb4e88aSMike Kravetz 1647cfd5082bSSidhartha Kumar static void remove_hugetlb_folio(struct hstate *h, struct folio *folio, 164834d9e35bSMike Kravetz bool adjust_surplus) 164934d9e35bSMike Kravetz { 1650cfd5082bSSidhartha Kumar __remove_hugetlb_folio(h, folio, adjust_surplus, false); 165134d9e35bSMike Kravetz } 165234d9e35bSMike Kravetz 1653cfd5082bSSidhartha Kumar static void remove_hugetlb_folio_for_demote(struct hstate *h, struct folio *folio, 16548531fc6fSMike Kravetz bool adjust_surplus) 16558531fc6fSMike Kravetz { 1656cfd5082bSSidhartha Kumar __remove_hugetlb_folio(h, folio, adjust_surplus, true); 16578531fc6fSMike Kravetz } 16588531fc6fSMike Kravetz 16592f6c57d6SSidhartha Kumar static void add_hugetlb_folio(struct hstate *h, struct folio *folio, 1660ad2fa371SMuchun Song bool adjust_surplus) 1661ad2fa371SMuchun Song { 1662ad2fa371SMuchun Song int zeroed; 16632f6c57d6SSidhartha Kumar int nid = folio_nid(folio); 1664ad2fa371SMuchun Song 16652f6c57d6SSidhartha Kumar VM_BUG_ON_FOLIO(!folio_test_hugetlb_vmemmap_optimized(folio), folio); 1666ad2fa371SMuchun Song 1667ad2fa371SMuchun Song lockdep_assert_held(&hugetlb_lock); 1668ad2fa371SMuchun Song 16692f6c57d6SSidhartha Kumar INIT_LIST_HEAD(&folio->lru); 1670ad2fa371SMuchun Song h->nr_huge_pages++; 1671ad2fa371SMuchun Song h->nr_huge_pages_node[nid]++; 1672ad2fa371SMuchun Song 1673ad2fa371SMuchun Song if (adjust_surplus) { 1674ad2fa371SMuchun Song h->surplus_huge_pages++; 1675ad2fa371SMuchun Song h->surplus_huge_pages_node[nid]++; 1676ad2fa371SMuchun Song } 1677ad2fa371SMuchun Song 16782f6c57d6SSidhartha Kumar folio_set_compound_dtor(folio, HUGETLB_PAGE_DTOR); 16792f6c57d6SSidhartha Kumar folio_change_private(folio, NULL); 1680a9e1eab2SMiaohe Lin /* 16812f6c57d6SSidhartha Kumar * We have to set hugetlb_vmemmap_optimized again as above 16822f6c57d6SSidhartha Kumar * folio_change_private(folio, NULL) cleared it. 1683a9e1eab2SMiaohe Lin */ 16842f6c57d6SSidhartha Kumar folio_set_hugetlb_vmemmap_optimized(folio); 1685ad2fa371SMuchun Song 1686ad2fa371SMuchun Song /* 16872f6c57d6SSidhartha Kumar * This folio is about to be managed by the hugetlb allocator and 1688b65a4edaSMike Kravetz * should have no users. Drop our reference, and check for others 1689b65a4edaSMike Kravetz * just in case. 1690ad2fa371SMuchun Song */ 16912f6c57d6SSidhartha Kumar zeroed = folio_put_testzero(folio); 16922f6c57d6SSidhartha Kumar if (unlikely(!zeroed)) 1693b65a4edaSMike Kravetz /* 1694b65a4edaSMike Kravetz * It is VERY unlikely soneone else has taken a ref on 1695b65a4edaSMike Kravetz * the page. In this case, we simply return as the 1696b65a4edaSMike Kravetz * hugetlb destructor (free_huge_page) will be called 1697b65a4edaSMike Kravetz * when this other ref is dropped. 1698b65a4edaSMike Kravetz */ 1699b65a4edaSMike Kravetz return; 1700b65a4edaSMike Kravetz 17012f6c57d6SSidhartha Kumar arch_clear_hugepage_flags(&folio->page); 1702240d67a8SSidhartha Kumar enqueue_hugetlb_folio(h, folio); 1703ad2fa371SMuchun Song } 1704ad2fa371SMuchun Song 1705b65d4adbSMuchun Song static void __update_and_free_page(struct hstate *h, struct page *page) 17066af2acb6SAdam Litke { 17076af2acb6SAdam Litke int i; 1708911565b8SSidhartha Kumar struct folio *folio = page_folio(page); 170914455eabSCheng Li struct page *subpage; 1710a5516438SAndi Kleen 17114eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1712944d9fecSLuiz Capitulino return; 171318229df5SAndy Whitcroft 1714161df60eSNaoya Horiguchi /* 1715161df60eSNaoya Horiguchi * If we don't know which subpages are hwpoisoned, we can't free 1716161df60eSNaoya Horiguchi * the hugepage, so it's leaked intentionally. 1717161df60eSNaoya Horiguchi */ 17187f325a8dSSidhartha Kumar if (folio_test_hugetlb_raw_hwp_unreliable(folio)) 1719161df60eSNaoya Horiguchi return; 1720161df60eSNaoya Horiguchi 17216213834cSMuchun Song if (hugetlb_vmemmap_restore(h, page)) { 1722ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 1723ad2fa371SMuchun Song /* 1724ad2fa371SMuchun Song * If we cannot allocate vmemmap pages, just refuse to free the 1725ad2fa371SMuchun Song * page and put the page back on the hugetlb free list and treat 1726ad2fa371SMuchun Song * as a surplus page. 1727ad2fa371SMuchun Song */ 17287f325a8dSSidhartha Kumar add_hugetlb_folio(h, folio, true); 1729ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 1730ad2fa371SMuchun Song return; 1731ad2fa371SMuchun Song } 1732ad2fa371SMuchun Song 1733161df60eSNaoya Horiguchi /* 1734161df60eSNaoya Horiguchi * Move PageHWPoison flag from head page to the raw error pages, 1735161df60eSNaoya Horiguchi * which makes any healthy subpages reusable. 1736161df60eSNaoya Horiguchi */ 1737911565b8SSidhartha Kumar if (unlikely(folio_test_hwpoison(folio))) 1738911565b8SSidhartha Kumar hugetlb_clear_page_hwpoison(&folio->page); 1739161df60eSNaoya Horiguchi 174014455eabSCheng Li for (i = 0; i < pages_per_huge_page(h); i++) { 17417f325a8dSSidhartha Kumar subpage = folio_page(folio, i); 1742dbfee5aeSMike Kravetz subpage->flags &= ~(1 << PG_locked | 1 << PG_error | 174332f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1744a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1745a7407a27SLuiz Capitulino 1 << PG_writeback); 17466af2acb6SAdam Litke } 1747a01f4390SMike Kravetz 1748a01f4390SMike Kravetz /* 1749a01f4390SMike Kravetz * Non-gigantic pages demoted from CMA allocated gigantic pages 17507f325a8dSSidhartha Kumar * need to be given back to CMA in free_gigantic_folio. 1751a01f4390SMike Kravetz */ 1752a01f4390SMike Kravetz if (hstate_is_gigantic(h) || 17532f6c57d6SSidhartha Kumar hugetlb_cma_folio(folio, huge_page_order(h))) { 1754911565b8SSidhartha Kumar destroy_compound_gigantic_folio(folio, huge_page_order(h)); 17557f325a8dSSidhartha Kumar free_gigantic_folio(folio, huge_page_order(h)); 1756944d9fecSLuiz Capitulino } else { 1757a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 17586af2acb6SAdam Litke } 1759944d9fecSLuiz Capitulino } 17606af2acb6SAdam Litke 1761b65d4adbSMuchun Song /* 1762d6ef19e2SSidhartha Kumar * As update_and_free_hugetlb_folio() can be called under any context, so we cannot 1763b65d4adbSMuchun Song * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the 1764b65d4adbSMuchun Song * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate 1765b65d4adbSMuchun Song * the vmemmap pages. 1766b65d4adbSMuchun Song * 1767b65d4adbSMuchun Song * free_hpage_workfn() locklessly retrieves the linked list of pages to be 1768b65d4adbSMuchun Song * freed and frees them one-by-one. As the page->mapping pointer is going 1769b65d4adbSMuchun Song * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node 1770b65d4adbSMuchun Song * structure of a lockless linked list of huge pages to be freed. 1771b65d4adbSMuchun Song */ 1772b65d4adbSMuchun Song static LLIST_HEAD(hpage_freelist); 1773b65d4adbSMuchun Song 1774b65d4adbSMuchun Song static void free_hpage_workfn(struct work_struct *work) 1775b65d4adbSMuchun Song { 1776b65d4adbSMuchun Song struct llist_node *node; 1777b65d4adbSMuchun Song 1778b65d4adbSMuchun Song node = llist_del_all(&hpage_freelist); 1779b65d4adbSMuchun Song 1780b65d4adbSMuchun Song while (node) { 1781b65d4adbSMuchun Song struct page *page; 1782b65d4adbSMuchun Song struct hstate *h; 1783b65d4adbSMuchun Song 1784b65d4adbSMuchun Song page = container_of((struct address_space **)node, 1785b65d4adbSMuchun Song struct page, mapping); 1786b65d4adbSMuchun Song node = node->next; 1787b65d4adbSMuchun Song page->mapping = NULL; 1788b65d4adbSMuchun Song /* 1789b65d4adbSMuchun Song * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate() 1790b65d4adbSMuchun Song * is going to trigger because a previous call to 1791cfd5082bSSidhartha Kumar * remove_hugetlb_folio() will call folio_set_compound_dtor 1792cfd5082bSSidhartha Kumar * (folio, NULL_COMPOUND_DTOR), so do not use page_hstate() 1793cfd5082bSSidhartha Kumar * directly. 1794b65d4adbSMuchun Song */ 1795b65d4adbSMuchun Song h = size_to_hstate(page_size(page)); 1796b65d4adbSMuchun Song 1797b65d4adbSMuchun Song __update_and_free_page(h, page); 1798b65d4adbSMuchun Song 1799b65d4adbSMuchun Song cond_resched(); 1800b65d4adbSMuchun Song } 1801b65d4adbSMuchun Song } 1802b65d4adbSMuchun Song static DECLARE_WORK(free_hpage_work, free_hpage_workfn); 1803b65d4adbSMuchun Song 1804b65d4adbSMuchun Song static inline void flush_free_hpage_work(struct hstate *h) 1805b65d4adbSMuchun Song { 18066213834cSMuchun Song if (hugetlb_vmemmap_optimizable(h)) 1807b65d4adbSMuchun Song flush_work(&free_hpage_work); 1808b65d4adbSMuchun Song } 1809b65d4adbSMuchun Song 1810d6ef19e2SSidhartha Kumar static void update_and_free_hugetlb_folio(struct hstate *h, struct folio *folio, 1811b65d4adbSMuchun Song bool atomic) 1812b65d4adbSMuchun Song { 1813d6ef19e2SSidhartha Kumar if (!folio_test_hugetlb_vmemmap_optimized(folio) || !atomic) { 1814d6ef19e2SSidhartha Kumar __update_and_free_page(h, &folio->page); 1815b65d4adbSMuchun Song return; 1816b65d4adbSMuchun Song } 1817b65d4adbSMuchun Song 1818b65d4adbSMuchun Song /* 1819b65d4adbSMuchun Song * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages. 1820b65d4adbSMuchun Song * 1821b65d4adbSMuchun Song * Only call schedule_work() if hpage_freelist is previously 1822b65d4adbSMuchun Song * empty. Otherwise, schedule_work() had been called but the workfn 1823b65d4adbSMuchun Song * hasn't retrieved the list yet. 1824b65d4adbSMuchun Song */ 1825d6ef19e2SSidhartha Kumar if (llist_add((struct llist_node *)&folio->mapping, &hpage_freelist)) 1826b65d4adbSMuchun Song schedule_work(&free_hpage_work); 1827b65d4adbSMuchun Song } 1828b65d4adbSMuchun Song 182910c6ec49SMike Kravetz static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list) 183010c6ec49SMike Kravetz { 183110c6ec49SMike Kravetz struct page *page, *t_page; 1832d6ef19e2SSidhartha Kumar struct folio *folio; 183310c6ec49SMike Kravetz 183410c6ec49SMike Kravetz list_for_each_entry_safe(page, t_page, list, lru) { 1835d6ef19e2SSidhartha Kumar folio = page_folio(page); 1836d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, folio, false); 183710c6ec49SMike Kravetz cond_resched(); 183810c6ec49SMike Kravetz } 183910c6ec49SMike Kravetz } 184010c6ec49SMike Kravetz 1841e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1842e5ff2159SAndi Kleen { 1843e5ff2159SAndi Kleen struct hstate *h; 1844e5ff2159SAndi Kleen 1845e5ff2159SAndi Kleen for_each_hstate(h) { 1846e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1847e5ff2159SAndi Kleen return h; 1848e5ff2159SAndi Kleen } 1849e5ff2159SAndi Kleen return NULL; 1850e5ff2159SAndi Kleen } 1851e5ff2159SAndi Kleen 1852db71ef79SMike Kravetz void free_huge_page(struct page *page) 185327a85ef1SDavid Gibson { 1854a5516438SAndi Kleen /* 1855a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1856a5516438SAndi Kleen * compound page destructor. 1857a5516438SAndi Kleen */ 18580356c4b9SSidhartha Kumar struct folio *folio = page_folio(page); 18590356c4b9SSidhartha Kumar struct hstate *h = folio_hstate(folio); 18600356c4b9SSidhartha Kumar int nid = folio_nid(folio); 18610356c4b9SSidhartha Kumar struct hugepage_subpool *spool = hugetlb_folio_subpool(folio); 186207443a85SJoonsoo Kim bool restore_reserve; 1863db71ef79SMike Kravetz unsigned long flags; 186427a85ef1SDavid Gibson 18650356c4b9SSidhartha Kumar VM_BUG_ON_FOLIO(folio_ref_count(folio), folio); 18660356c4b9SSidhartha Kumar VM_BUG_ON_FOLIO(folio_mapcount(folio), folio); 18678ace22bcSYongkai Wu 18680356c4b9SSidhartha Kumar hugetlb_set_folio_subpool(folio, NULL); 18690356c4b9SSidhartha Kumar if (folio_test_anon(folio)) 18700356c4b9SSidhartha Kumar __ClearPageAnonExclusive(&folio->page); 18710356c4b9SSidhartha Kumar folio->mapping = NULL; 18720356c4b9SSidhartha Kumar restore_reserve = folio_test_hugetlb_restore_reserve(folio); 18730356c4b9SSidhartha Kumar folio_clear_hugetlb_restore_reserve(folio); 187427a85ef1SDavid Gibson 18751c5ecae3SMike Kravetz /* 1876d6995da3SMike Kravetz * If HPageRestoreReserve was set on page, page allocation consumed a 18770919e1b6SMike Kravetz * reservation. If the page was associated with a subpool, there 18780919e1b6SMike Kravetz * would have been a page reserved in the subpool before allocation 18790919e1b6SMike Kravetz * via hugepage_subpool_get_pages(). Since we are 'restoring' the 18806c26d310SMiaohe Lin * reservation, do not call hugepage_subpool_put_pages() as this will 18810919e1b6SMike Kravetz * remove the reserved page from the subpool. 18820919e1b6SMike Kravetz */ 18830919e1b6SMike Kravetz if (!restore_reserve) { 18840919e1b6SMike Kravetz /* 18850919e1b6SMike Kravetz * A return code of zero implies that the subpool will be 18860919e1b6SMike Kravetz * under its minimum size if the reservation is not restored 18870919e1b6SMike Kravetz * after page is free. Therefore, force restore_reserve 18880919e1b6SMike Kravetz * operation. 18891c5ecae3SMike Kravetz */ 18901c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 18911c5ecae3SMike Kravetz restore_reserve = true; 18920919e1b6SMike Kravetz } 18931c5ecae3SMike Kravetz 1894db71ef79SMike Kravetz spin_lock_irqsave(&hugetlb_lock, flags); 18950356c4b9SSidhartha Kumar folio_clear_hugetlb_migratable(folio); 1896d4ab0316SSidhartha Kumar hugetlb_cgroup_uncharge_folio(hstate_index(h), 1897d4ab0316SSidhartha Kumar pages_per_huge_page(h), folio); 1898d4ab0316SSidhartha Kumar hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h), 1899d4ab0316SSidhartha Kumar pages_per_huge_page(h), folio); 190007443a85SJoonsoo Kim if (restore_reserve) 190107443a85SJoonsoo Kim h->resv_huge_pages++; 190207443a85SJoonsoo Kim 19030356c4b9SSidhartha Kumar if (folio_test_hugetlb_temporary(folio)) { 1904cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, false); 1905db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1906d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, folio, true); 1907ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 19080edaecfaSAneesh Kumar K.V /* remove the page from active list */ 1909cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, true); 1910db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1911d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, folio, true); 19127893d1d5SAdam Litke } else { 19135d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1914240d67a8SSidhartha Kumar enqueue_hugetlb_folio(h, folio); 1915db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 191627a85ef1SDavid Gibson } 19171121828aSMike Kravetz } 191827a85ef1SDavid Gibson 1919d3d99fccSOscar Salvador /* 1920d3d99fccSOscar Salvador * Must be called with the hugetlb lock held 1921d3d99fccSOscar Salvador */ 1922d3d99fccSOscar Salvador static void __prep_account_new_huge_page(struct hstate *h, int nid) 1923d3d99fccSOscar Salvador { 1924d3d99fccSOscar Salvador lockdep_assert_held(&hugetlb_lock); 1925d3d99fccSOscar Salvador h->nr_huge_pages++; 1926d3d99fccSOscar Salvador h->nr_huge_pages_node[nid]++; 1927d3d99fccSOscar Salvador } 1928d3d99fccSOscar Salvador 1929de656ed3SSidhartha Kumar static void __prep_new_hugetlb_folio(struct hstate *h, struct folio *folio) 1930b7ba30c6SAndi Kleen { 1931de656ed3SSidhartha Kumar hugetlb_vmemmap_optimize(h, &folio->page); 1932de656ed3SSidhartha Kumar INIT_LIST_HEAD(&folio->lru); 19339fd33058SSidhartha Kumar folio_set_compound_dtor(folio, HUGETLB_PAGE_DTOR); 1934de656ed3SSidhartha Kumar hugetlb_set_folio_subpool(folio, NULL); 1935de656ed3SSidhartha Kumar set_hugetlb_cgroup(folio, NULL); 1936de656ed3SSidhartha Kumar set_hugetlb_cgroup_rsvd(folio, NULL); 1937d3d99fccSOscar Salvador } 1938d3d99fccSOscar Salvador 1939d1c60955SSidhartha Kumar static void prep_new_hugetlb_folio(struct hstate *h, struct folio *folio, int nid) 1940d3d99fccSOscar Salvador { 1941de656ed3SSidhartha Kumar __prep_new_hugetlb_folio(h, folio); 1942db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 1943d3d99fccSOscar Salvador __prep_account_new_huge_page(h, nid); 1944db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1945b7ba30c6SAndi Kleen } 1946b7ba30c6SAndi Kleen 1947d1c60955SSidhartha Kumar static bool __prep_compound_gigantic_folio(struct folio *folio, 1948d1c60955SSidhartha Kumar unsigned int order, bool demote) 194920a0307cSWu Fengguang { 19507118fc29SMike Kravetz int i, j; 195120a0307cSWu Fengguang int nr_pages = 1 << order; 195214455eabSCheng Li struct page *p; 195320a0307cSWu Fengguang 1954d1c60955SSidhartha Kumar __folio_clear_reserved(folio); 1955d1c60955SSidhartha Kumar __folio_set_head(folio); 1956c45bc55aSSidhartha Kumar /* we rely on prep_new_hugetlb_folio to set the destructor */ 1957c45bc55aSSidhartha Kumar folio_set_compound_order(folio, order); 19582b21624fSMike Kravetz for (i = 0; i < nr_pages; i++) { 1959d1c60955SSidhartha Kumar p = folio_page(folio, i); 196014455eabSCheng Li 1961ef5a22beSAndrea Arcangeli /* 1962ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1963ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1964ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 19657c8de358SEthon Paul * too. Otherwise drivers using get_user_pages() to access tail 1966ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1967ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1968ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1969ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1970ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1971ef5a22beSAndrea Arcangeli * after get_user_pages(). 1972ef5a22beSAndrea Arcangeli */ 19737fb0728aSMike Kravetz if (i != 0) /* head page cleared above */ 1974ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 19757118fc29SMike Kravetz /* 19767118fc29SMike Kravetz * Subtle and very unlikely 19777118fc29SMike Kravetz * 19787118fc29SMike Kravetz * Gigantic 'page allocators' such as memblock or cma will 19797118fc29SMike Kravetz * return a set of pages with each page ref counted. We need 19807118fc29SMike Kravetz * to turn this set of pages into a compound page with tail 19817118fc29SMike Kravetz * page ref counts set to zero. Code such as speculative page 19827118fc29SMike Kravetz * cache adding could take a ref on a 'to be' tail page. 19837118fc29SMike Kravetz * We need to respect any increased ref count, and only set 19847118fc29SMike Kravetz * the ref count to zero if count is currently 1. If count 1985416d85edSMike Kravetz * is not 1, we return an error. An error return indicates 1986416d85edSMike Kravetz * the set of pages can not be converted to a gigantic page. 1987416d85edSMike Kravetz * The caller who allocated the pages should then discard the 1988416d85edSMike Kravetz * pages using the appropriate free interface. 198934d9e35bSMike Kravetz * 199034d9e35bSMike Kravetz * In the case of demote, the ref count will be zero. 19917118fc29SMike Kravetz */ 199234d9e35bSMike Kravetz if (!demote) { 19937118fc29SMike Kravetz if (!page_ref_freeze(p, 1)) { 1994416d85edSMike Kravetz pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n"); 19957118fc29SMike Kravetz goto out_error; 19967118fc29SMike Kravetz } 199734d9e35bSMike Kravetz } else { 199834d9e35bSMike Kravetz VM_BUG_ON_PAGE(page_count(p), p); 199934d9e35bSMike Kravetz } 20002b21624fSMike Kravetz if (i != 0) 2001d1c60955SSidhartha Kumar set_compound_head(p, &folio->page); 200220a0307cSWu Fengguang } 2003d1c60955SSidhartha Kumar atomic_set(folio_mapcount_ptr(folio), -1); 2004d1c60955SSidhartha Kumar atomic_set(folio_subpages_mapcount_ptr(folio), 0); 2005d1c60955SSidhartha Kumar atomic_set(folio_pincount_ptr(folio), 0); 20067118fc29SMike Kravetz return true; 20077118fc29SMike Kravetz 20087118fc29SMike Kravetz out_error: 20092b21624fSMike Kravetz /* undo page modifications made above */ 20102b21624fSMike Kravetz for (j = 0; j < i; j++) { 2011d1c60955SSidhartha Kumar p = folio_page(folio, j); 20122b21624fSMike Kravetz if (j != 0) 20137118fc29SMike Kravetz clear_compound_head(p); 20147118fc29SMike Kravetz set_page_refcounted(p); 20157118fc29SMike Kravetz } 20167118fc29SMike Kravetz /* need to clear PG_reserved on remaining tail pages */ 201714455eabSCheng Li for (; j < nr_pages; j++) { 2018d1c60955SSidhartha Kumar p = folio_page(folio, j); 20197118fc29SMike Kravetz __ClearPageReserved(p); 202014455eabSCheng Li } 2021d1c60955SSidhartha Kumar folio_set_compound_order(folio, 0); 2022d1c60955SSidhartha Kumar __folio_clear_head(folio); 20237118fc29SMike Kravetz return false; 202420a0307cSWu Fengguang } 202520a0307cSWu Fengguang 2026d1c60955SSidhartha Kumar static bool prep_compound_gigantic_folio(struct folio *folio, 20278531fc6fSMike Kravetz unsigned int order) 20288531fc6fSMike Kravetz { 2029d1c60955SSidhartha Kumar return __prep_compound_gigantic_folio(folio, order, false); 2030d1c60955SSidhartha Kumar } 2031d1c60955SSidhartha Kumar 2032d1c60955SSidhartha Kumar static bool prep_compound_gigantic_folio_for_demote(struct folio *folio, 2033d1c60955SSidhartha Kumar unsigned int order) 2034d1c60955SSidhartha Kumar { 2035d1c60955SSidhartha Kumar return __prep_compound_gigantic_folio(folio, order, true); 20368531fc6fSMike Kravetz } 20378531fc6fSMike Kravetz 20387795912cSAndrew Morton /* 20397795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 20407795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 20417795912cSAndrew Morton * details. 20427795912cSAndrew Morton */ 204320a0307cSWu Fengguang int PageHuge(struct page *page) 204420a0307cSWu Fengguang { 204520a0307cSWu Fengguang if (!PageCompound(page)) 204620a0307cSWu Fengguang return 0; 204720a0307cSWu Fengguang 204820a0307cSWu Fengguang page = compound_head(page); 2049f1e61557SKirill A. Shutemov return page[1].compound_dtor == HUGETLB_PAGE_DTOR; 205020a0307cSWu Fengguang } 205143131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 205243131e14SNaoya Horiguchi 205327c73ae7SAndrea Arcangeli /* 205427c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 205527c73ae7SAndrea Arcangeli * normal or transparent huge pages. 205627c73ae7SAndrea Arcangeli */ 205727c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 205827c73ae7SAndrea Arcangeli { 205927c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 206027c73ae7SAndrea Arcangeli return 0; 206127c73ae7SAndrea Arcangeli 2062d4af73e3SVlastimil Babka return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; 206327c73ae7SAndrea Arcangeli } 20644e936eccSDavid Howells EXPORT_SYMBOL_GPL(PageHeadHuge); 206527c73ae7SAndrea Arcangeli 2066c0d0381aSMike Kravetz /* 2067c0d0381aSMike Kravetz * Find and lock address space (mapping) in write mode. 2068c0d0381aSMike Kravetz * 2069336bf30eSMike Kravetz * Upon entry, the page is locked which means that page_mapping() is 2070336bf30eSMike Kravetz * stable. Due to locking order, we can only trylock_write. If we can 2071336bf30eSMike Kravetz * not get the lock, simply return NULL to caller. 2072c0d0381aSMike Kravetz */ 2073c0d0381aSMike Kravetz struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) 2074c0d0381aSMike Kravetz { 2075336bf30eSMike Kravetz struct address_space *mapping = page_mapping(hpage); 2076c0d0381aSMike Kravetz 2077c0d0381aSMike Kravetz if (!mapping) 2078c0d0381aSMike Kravetz return mapping; 2079c0d0381aSMike Kravetz 2080c0d0381aSMike Kravetz if (i_mmap_trylock_write(mapping)) 2081c0d0381aSMike Kravetz return mapping; 2082c0d0381aSMike Kravetz 2083c0d0381aSMike Kravetz return NULL; 2084c0d0381aSMike Kravetz } 2085c0d0381aSMike Kravetz 2086fe19bd3dSHugh Dickins pgoff_t hugetlb_basepage_index(struct page *page) 208713d60f4bSZhang Yi { 208813d60f4bSZhang Yi struct page *page_head = compound_head(page); 208913d60f4bSZhang Yi pgoff_t index = page_index(page_head); 209013d60f4bSZhang Yi unsigned long compound_idx; 209113d60f4bSZhang Yi 209213d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 209313d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 209413d60f4bSZhang Yi else 209513d60f4bSZhang Yi compound_idx = page - page_head; 209613d60f4bSZhang Yi 209713d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 209813d60f4bSZhang Yi } 209913d60f4bSZhang Yi 210019fc1a7eSSidhartha Kumar static struct folio *alloc_buddy_hugetlb_folio(struct hstate *h, 2101f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 2102f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 21031da177e4SLinus Torvalds { 2104af0fb9dfSMichal Hocko int order = huge_page_order(h); 21051da177e4SLinus Torvalds struct page *page; 2106f60858f9SMike Kravetz bool alloc_try_hard = true; 21072b21624fSMike Kravetz bool retry = true; 2108f96efd58SJoe Jin 2109f60858f9SMike Kravetz /* 2110f60858f9SMike Kravetz * By default we always try hard to allocate the page with 2111f60858f9SMike Kravetz * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in 2112f60858f9SMike Kravetz * a loop (to adjust global huge page counts) and previous allocation 2113f60858f9SMike Kravetz * failed, do not continue to try hard on the same node. Use the 2114f60858f9SMike Kravetz * node_alloc_noretry bitmap to manage this state information. 2115f60858f9SMike Kravetz */ 2116f60858f9SMike Kravetz if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) 2117f60858f9SMike Kravetz alloc_try_hard = false; 2118f60858f9SMike Kravetz gfp_mask |= __GFP_COMP|__GFP_NOWARN; 2119f60858f9SMike Kravetz if (alloc_try_hard) 2120f60858f9SMike Kravetz gfp_mask |= __GFP_RETRY_MAYFAIL; 2121af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 2122af0fb9dfSMichal Hocko nid = numa_mem_id(); 21232b21624fSMike Kravetz retry: 212484172f4bSMatthew Wilcox (Oracle) page = __alloc_pages(gfp_mask, order, nid, nmask); 21252b21624fSMike Kravetz 21262b21624fSMike Kravetz /* Freeze head page */ 21272b21624fSMike Kravetz if (page && !page_ref_freeze(page, 1)) { 21282b21624fSMike Kravetz __free_pages(page, order); 21292b21624fSMike Kravetz if (retry) { /* retry once */ 21302b21624fSMike Kravetz retry = false; 21312b21624fSMike Kravetz goto retry; 21322b21624fSMike Kravetz } 21332b21624fSMike Kravetz /* WOW! twice in a row. */ 21342b21624fSMike Kravetz pr_warn("HugeTLB head page unexpected inflated ref count\n"); 21352b21624fSMike Kravetz page = NULL; 21362b21624fSMike Kravetz } 21372b21624fSMike Kravetz 2138f60858f9SMike Kravetz /* 2139f60858f9SMike Kravetz * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this 2140f60858f9SMike Kravetz * indicates an overall state change. Clear bit so that we resume 2141f60858f9SMike Kravetz * normal 'try hard' allocations. 2142f60858f9SMike Kravetz */ 2143f60858f9SMike Kravetz if (node_alloc_noretry && page && !alloc_try_hard) 2144f60858f9SMike Kravetz node_clear(nid, *node_alloc_noretry); 2145f60858f9SMike Kravetz 2146f60858f9SMike Kravetz /* 2147f60858f9SMike Kravetz * If we tried hard to get a page but failed, set bit so that 2148f60858f9SMike Kravetz * subsequent attempts will not try as hard until there is an 2149f60858f9SMike Kravetz * overall state change. 2150f60858f9SMike Kravetz */ 2151f60858f9SMike Kravetz if (node_alloc_noretry && !page && alloc_try_hard) 2152f60858f9SMike Kravetz node_set(nid, *node_alloc_noretry); 2153f60858f9SMike Kravetz 215419fc1a7eSSidhartha Kumar if (!page) { 215519fc1a7eSSidhartha Kumar __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 215619fc1a7eSSidhartha Kumar return NULL; 215719fc1a7eSSidhartha Kumar } 215819fc1a7eSSidhartha Kumar 215919fc1a7eSSidhartha Kumar __count_vm_event(HTLB_BUDDY_PGALLOC); 216019fc1a7eSSidhartha Kumar return page_folio(page); 216163b4613cSNishanth Aravamudan } 216263b4613cSNishanth Aravamudan 2163af0fb9dfSMichal Hocko /* 21640c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 21650c397daeSMichal Hocko * should use this function to get new hugetlb pages 21662b21624fSMike Kravetz * 21672b21624fSMike Kravetz * Note that returned page is 'frozen': ref count of head page and all tail 21682b21624fSMike Kravetz * pages is zero. 21690c397daeSMichal Hocko */ 217019fc1a7eSSidhartha Kumar static struct folio *alloc_fresh_hugetlb_folio(struct hstate *h, 2171f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 2172f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 21730c397daeSMichal Hocko { 21747f325a8dSSidhartha Kumar struct folio *folio; 21757118fc29SMike Kravetz bool retry = false; 21760c397daeSMichal Hocko 21777118fc29SMike Kravetz retry: 21780c397daeSMichal Hocko if (hstate_is_gigantic(h)) 217919fc1a7eSSidhartha Kumar folio = alloc_gigantic_folio(h, gfp_mask, nid, nmask); 21800c397daeSMichal Hocko else 218119fc1a7eSSidhartha Kumar folio = alloc_buddy_hugetlb_folio(h, gfp_mask, 2182f60858f9SMike Kravetz nid, nmask, node_alloc_noretry); 218319fc1a7eSSidhartha Kumar if (!folio) 21840c397daeSMichal Hocko return NULL; 21857118fc29SMike Kravetz if (hstate_is_gigantic(h)) { 2186d1c60955SSidhartha Kumar if (!prep_compound_gigantic_folio(folio, huge_page_order(h))) { 21877118fc29SMike Kravetz /* 21887118fc29SMike Kravetz * Rare failure to convert pages to compound page. 21897118fc29SMike Kravetz * Free pages and try again - ONCE! 21907118fc29SMike Kravetz */ 21917f325a8dSSidhartha Kumar free_gigantic_folio(folio, huge_page_order(h)); 21927118fc29SMike Kravetz if (!retry) { 21937118fc29SMike Kravetz retry = true; 21947118fc29SMike Kravetz goto retry; 21957118fc29SMike Kravetz } 21967118fc29SMike Kravetz return NULL; 21977118fc29SMike Kravetz } 21987118fc29SMike Kravetz } 2199d1c60955SSidhartha Kumar prep_new_hugetlb_folio(h, folio, folio_nid(folio)); 22000c397daeSMichal Hocko 220119fc1a7eSSidhartha Kumar return folio; 22020c397daeSMichal Hocko } 22030c397daeSMichal Hocko 22040c397daeSMichal Hocko /* 2205af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 2206af0fb9dfSMichal Hocko * manner. 2207af0fb9dfSMichal Hocko */ 2208f60858f9SMike Kravetz static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 2209f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 2210b2261026SJoonsoo Kim { 221119fc1a7eSSidhartha Kumar struct folio *folio; 2212b2261026SJoonsoo Kim int nr_nodes, node; 2213af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2214b2261026SJoonsoo Kim 2215b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 221619fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, node, 221719fc1a7eSSidhartha Kumar nodes_allowed, node_alloc_noretry); 221819fc1a7eSSidhartha Kumar if (folio) { 221919fc1a7eSSidhartha Kumar free_huge_page(&folio->page); /* free it into the hugepage allocator */ 222019fc1a7eSSidhartha Kumar return 1; 222119fc1a7eSSidhartha Kumar } 2222b2261026SJoonsoo Kim } 2223b2261026SJoonsoo Kim 2224af0fb9dfSMichal Hocko return 0; 2225b2261026SJoonsoo Kim } 2226b2261026SJoonsoo Kim 2227e8c5c824SLee Schermerhorn /* 222810c6ec49SMike Kravetz * Remove huge page from pool from next node to free. Attempt to keep 222910c6ec49SMike Kravetz * persistent huge pages more or less balanced over allowed nodes. 223010c6ec49SMike Kravetz * This routine only 'removes' the hugetlb page. The caller must make 223110c6ec49SMike Kravetz * an additional call to free the page to low level allocators. 2232e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 2233e8c5c824SLee Schermerhorn */ 223410c6ec49SMike Kravetz static struct page *remove_pool_huge_page(struct hstate *h, 223510c6ec49SMike Kravetz nodemask_t *nodes_allowed, 22366ae11b27SLee Schermerhorn bool acct_surplus) 2237e8c5c824SLee Schermerhorn { 2238b2261026SJoonsoo Kim int nr_nodes, node; 223910c6ec49SMike Kravetz struct page *page = NULL; 2240cfd5082bSSidhartha Kumar struct folio *folio; 2241e8c5c824SLee Schermerhorn 22429487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2243b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2244685f3457SLee Schermerhorn /* 2245685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 2246685f3457SLee Schermerhorn * nodes with surplus pages. 2247685f3457SLee Schermerhorn */ 2248b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 2249b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 225010c6ec49SMike Kravetz page = list_entry(h->hugepage_freelists[node].next, 2251e8c5c824SLee Schermerhorn struct page, lru); 2252cfd5082bSSidhartha Kumar folio = page_folio(page); 2253cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, acct_surplus); 22549a76db09SLee Schermerhorn break; 2255e8c5c824SLee Schermerhorn } 2256b2261026SJoonsoo Kim } 2257e8c5c824SLee Schermerhorn 225810c6ec49SMike Kravetz return page; 2259e8c5c824SLee Schermerhorn } 2260e8c5c824SLee Schermerhorn 2261c8721bbbSNaoya Horiguchi /* 2262c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 2263faf53defSNaoya Horiguchi * nothing for in-use hugepages and non-hugepages. 2264faf53defSNaoya Horiguchi * This function returns values like below: 2265faf53defSNaoya Horiguchi * 2266ad2fa371SMuchun Song * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages 2267ad2fa371SMuchun Song * when the system is under memory pressure and the feature of 2268ad2fa371SMuchun Song * freeing unused vmemmap pages associated with each hugetlb page 2269ad2fa371SMuchun Song * is enabled. 2270faf53defSNaoya Horiguchi * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 2271faf53defSNaoya Horiguchi * (allocated or reserved.) 2272faf53defSNaoya Horiguchi * 0: successfully dissolved free hugepages or the page is not a 2273faf53defSNaoya Horiguchi * hugepage (considered as already dissolved) 2274c8721bbbSNaoya Horiguchi */ 2275c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 2276c8721bbbSNaoya Horiguchi { 22776bc9b564SNaoya Horiguchi int rc = -EBUSY; 22781a7cdab5SSidhartha Kumar struct folio *folio = page_folio(page); 2279082d5b6bSGerald Schaefer 22807ffddd49SMuchun Song retry: 2281faf53defSNaoya Horiguchi /* Not to disrupt normal path by vainly holding hugetlb_lock */ 22821a7cdab5SSidhartha Kumar if (!folio_test_hugetlb(folio)) 2283faf53defSNaoya Horiguchi return 0; 2284faf53defSNaoya Horiguchi 2285db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 22861a7cdab5SSidhartha Kumar if (!folio_test_hugetlb(folio)) { 2287faf53defSNaoya Horiguchi rc = 0; 2288faf53defSNaoya Horiguchi goto out; 2289faf53defSNaoya Horiguchi } 2290faf53defSNaoya Horiguchi 22911a7cdab5SSidhartha Kumar if (!folio_ref_count(folio)) { 22921a7cdab5SSidhartha Kumar struct hstate *h = folio_hstate(folio); 22938346d69dSXin Hao if (!available_huge_pages(h)) 2294082d5b6bSGerald Schaefer goto out; 22957ffddd49SMuchun Song 22967ffddd49SMuchun Song /* 22977ffddd49SMuchun Song * We should make sure that the page is already on the free list 22987ffddd49SMuchun Song * when it is dissolved. 22997ffddd49SMuchun Song */ 23001a7cdab5SSidhartha Kumar if (unlikely(!folio_test_hugetlb_freed(folio))) { 2301db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 23027ffddd49SMuchun Song cond_resched(); 23037ffddd49SMuchun Song 23047ffddd49SMuchun Song /* 23057ffddd49SMuchun Song * Theoretically, we should return -EBUSY when we 23067ffddd49SMuchun Song * encounter this race. In fact, we have a chance 23077ffddd49SMuchun Song * to successfully dissolve the page if we do a 23087ffddd49SMuchun Song * retry. Because the race window is quite small. 23097ffddd49SMuchun Song * If we seize this opportunity, it is an optimization 23107ffddd49SMuchun Song * for increasing the success rate of dissolving page. 23117ffddd49SMuchun Song */ 23127ffddd49SMuchun Song goto retry; 23137ffddd49SMuchun Song } 23147ffddd49SMuchun Song 2315cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, folio, false); 2316ad2fa371SMuchun Song h->max_huge_pages--; 2317ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2318ad2fa371SMuchun Song 2319c3114a84SAnshuman Khandual /* 2320d6ef19e2SSidhartha Kumar * Normally update_and_free_hugtlb_folio will allocate required vmemmmap 2321d6ef19e2SSidhartha Kumar * before freeing the page. update_and_free_hugtlb_folio will fail to 2322ad2fa371SMuchun Song * free the page if it can not allocate required vmemmap. We 2323ad2fa371SMuchun Song * need to adjust max_huge_pages if the page is not freed. 2324ad2fa371SMuchun Song * Attempt to allocate vmemmmap here so that we can take 2325ad2fa371SMuchun Song * appropriate action on failure. 2326ad2fa371SMuchun Song */ 23271a7cdab5SSidhartha Kumar rc = hugetlb_vmemmap_restore(h, &folio->page); 2328ad2fa371SMuchun Song if (!rc) { 2329d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, folio, false); 2330ad2fa371SMuchun Song } else { 2331ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 23322f6c57d6SSidhartha Kumar add_hugetlb_folio(h, folio, false); 2333ad2fa371SMuchun Song h->max_huge_pages++; 2334ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2335ad2fa371SMuchun Song } 2336ad2fa371SMuchun Song 2337ad2fa371SMuchun Song return rc; 2338c8721bbbSNaoya Horiguchi } 2339082d5b6bSGerald Schaefer out: 2340db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2341082d5b6bSGerald Schaefer return rc; 2342c8721bbbSNaoya Horiguchi } 2343c8721bbbSNaoya Horiguchi 2344c8721bbbSNaoya Horiguchi /* 2345c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 2346c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 23472247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 23482247bb33SGerald Schaefer * part of it lies within the given range. 2349082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 2350082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 2351c8721bbbSNaoya Horiguchi */ 2352082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 2353c8721bbbSNaoya Horiguchi { 2354c8721bbbSNaoya Horiguchi unsigned long pfn; 2355eb03aa00SGerald Schaefer struct page *page; 2356082d5b6bSGerald Schaefer int rc = 0; 2357dc2628f3SMuchun Song unsigned int order; 2358dc2628f3SMuchun Song struct hstate *h; 2359c8721bbbSNaoya Horiguchi 2360d0177639SLi Zhong if (!hugepages_supported()) 2361082d5b6bSGerald Schaefer return rc; 2362d0177639SLi Zhong 2363dc2628f3SMuchun Song order = huge_page_order(&default_hstate); 2364dc2628f3SMuchun Song for_each_hstate(h) 2365dc2628f3SMuchun Song order = min(order, huge_page_order(h)); 2366dc2628f3SMuchun Song 2367dc2628f3SMuchun Song for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) { 2368eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 2369eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 2370eb03aa00SGerald Schaefer if (rc) 2371082d5b6bSGerald Schaefer break; 2372eb03aa00SGerald Schaefer } 2373082d5b6bSGerald Schaefer 2374082d5b6bSGerald Schaefer return rc; 2375c8721bbbSNaoya Horiguchi } 2376c8721bbbSNaoya Horiguchi 2377ab5ac90aSMichal Hocko /* 2378ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 2379ab5ac90aSMichal Hocko */ 23800c397daeSMichal Hocko static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, 23812b21624fSMike Kravetz int nid, nodemask_t *nmask) 23827893d1d5SAdam Litke { 238319fc1a7eSSidhartha Kumar struct folio *folio = NULL; 23847893d1d5SAdam Litke 2385bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2386aa888a74SAndi Kleen return NULL; 2387aa888a74SAndi Kleen 2388db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 23899980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 23909980d744SMichal Hocko goto out_unlock; 2391db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2392d1c3fb1fSNishanth Aravamudan 239319fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL); 239419fc1a7eSSidhartha Kumar if (!folio) 23950c397daeSMichal Hocko return NULL; 2396d1c3fb1fSNishanth Aravamudan 2397db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 23989980d744SMichal Hocko /* 23999980d744SMichal Hocko * We could have raced with the pool size change. 24009980d744SMichal Hocko * Double check that and simply deallocate the new page 24019980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 24029980d744SMichal Hocko * temporary page to workaround the nasty free_huge_page 24039980d744SMichal Hocko * codeflow 24049980d744SMichal Hocko */ 24059980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 240619fc1a7eSSidhartha Kumar folio_set_hugetlb_temporary(folio); 2407db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 240819fc1a7eSSidhartha Kumar free_huge_page(&folio->page); 24092bf753e6SKai Shen return NULL; 2410b65a4edaSMike Kravetz } 2411b65a4edaSMike Kravetz 24129980d744SMichal Hocko h->surplus_huge_pages++; 241319fc1a7eSSidhartha Kumar h->surplus_huge_pages_node[folio_nid(folio)]++; 24149980d744SMichal Hocko 24159980d744SMichal Hocko out_unlock: 2416db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 24177893d1d5SAdam Litke 241819fc1a7eSSidhartha Kumar return &folio->page; 24197893d1d5SAdam Litke } 24207893d1d5SAdam Litke 2421bbe88753SJoonsoo Kim static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, 2422ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 2423ab5ac90aSMichal Hocko { 242419fc1a7eSSidhartha Kumar struct folio *folio; 2425ab5ac90aSMichal Hocko 2426ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 2427ab5ac90aSMichal Hocko return NULL; 2428ab5ac90aSMichal Hocko 242919fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL); 243019fc1a7eSSidhartha Kumar if (!folio) 2431ab5ac90aSMichal Hocko return NULL; 2432ab5ac90aSMichal Hocko 24332b21624fSMike Kravetz /* fresh huge pages are frozen */ 243419fc1a7eSSidhartha Kumar folio_ref_unfreeze(folio, 1); 2435ab5ac90aSMichal Hocko /* 2436ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 2437ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 2438ab5ac90aSMichal Hocko */ 243919fc1a7eSSidhartha Kumar folio_set_hugetlb_temporary(folio); 2440ab5ac90aSMichal Hocko 244119fc1a7eSSidhartha Kumar return &folio->page; 2442ab5ac90aSMichal Hocko } 2443ab5ac90aSMichal Hocko 2444e4e574b7SAdam Litke /* 2445099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 2446099730d6SDave Hansen */ 2447e0ec90eeSDave Hansen static 24480c397daeSMichal Hocko struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, 2449099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 2450099730d6SDave Hansen { 2451cfcaa66fSBen Widawsky struct page *page = NULL; 2452aaf14e40SMichal Hocko struct mempolicy *mpol; 2453aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 2454aaf14e40SMichal Hocko int nid; 2455aaf14e40SMichal Hocko nodemask_t *nodemask; 2456aaf14e40SMichal Hocko 2457aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 2458cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 2459cfcaa66fSBen Widawsky gfp_t gfp = gfp_mask | __GFP_NOWARN; 2460cfcaa66fSBen Widawsky 2461cfcaa66fSBen Widawsky gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL); 24622b21624fSMike Kravetz page = alloc_surplus_huge_page(h, gfp, nid, nodemask); 2463cfcaa66fSBen Widawsky 2464cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 2465cfcaa66fSBen Widawsky nodemask = NULL; 2466cfcaa66fSBen Widawsky } 2467cfcaa66fSBen Widawsky 2468cfcaa66fSBen Widawsky if (!page) 24692b21624fSMike Kravetz page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); 2470aaf14e40SMichal Hocko mpol_cond_put(mpol); 2471aaf14e40SMichal Hocko return page; 2472099730d6SDave Hansen } 2473099730d6SDave Hansen 2474ab5ac90aSMichal Hocko /* page migration callback function */ 24753e59fcb0SMichal Hocko struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, 2476d92bbc27SJoonsoo Kim nodemask_t *nmask, gfp_t gfp_mask) 24774db9b2efSMichal Hocko { 2478db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 24798346d69dSXin Hao if (available_huge_pages(h)) { 24803e59fcb0SMichal Hocko struct page *page; 24813e59fcb0SMichal Hocko 24823e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); 24833e59fcb0SMichal Hocko if (page) { 2484db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 24853e59fcb0SMichal Hocko return page; 24864db9b2efSMichal Hocko } 24874db9b2efSMichal Hocko } 2488db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 24894db9b2efSMichal Hocko 24900c397daeSMichal Hocko return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); 24914db9b2efSMichal Hocko } 24924db9b2efSMichal Hocko 2493ebd63723SMichal Hocko /* mempolicy aware migration callback */ 2494389c8178SMichal Hocko struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, 2495389c8178SMichal Hocko unsigned long address) 2496ebd63723SMichal Hocko { 2497ebd63723SMichal Hocko struct mempolicy *mpol; 2498ebd63723SMichal Hocko nodemask_t *nodemask; 2499ebd63723SMichal Hocko struct page *page; 2500ebd63723SMichal Hocko gfp_t gfp_mask; 2501ebd63723SMichal Hocko int node; 2502ebd63723SMichal Hocko 2503ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 2504ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 2505d92bbc27SJoonsoo Kim page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); 2506ebd63723SMichal Hocko mpol_cond_put(mpol); 2507ebd63723SMichal Hocko 2508ebd63723SMichal Hocko return page; 2509ebd63723SMichal Hocko } 2510ebd63723SMichal Hocko 2511bf50bab2SNaoya Horiguchi /* 251225985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 2513e4e574b7SAdam Litke * of size 'delta'. 2514e4e574b7SAdam Litke */ 25150a4f3d1bSLiu Xiang static int gather_surplus_pages(struct hstate *h, long delta) 25161b2a1e7bSJules Irenge __must_hold(&hugetlb_lock) 2517e4e574b7SAdam Litke { 251834665341SMiaohe Lin LIST_HEAD(surplus_list); 2519e4e574b7SAdam Litke struct page *page, *tmp; 25200a4f3d1bSLiu Xiang int ret; 25210a4f3d1bSLiu Xiang long i; 25220a4f3d1bSLiu Xiang long needed, allocated; 252328073b02SHillf Danton bool alloc_ok = true; 2524e4e574b7SAdam Litke 25259487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2526a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 2527ac09b3a1SAdam Litke if (needed <= 0) { 2528a5516438SAndi Kleen h->resv_huge_pages += delta; 2529e4e574b7SAdam Litke return 0; 2530ac09b3a1SAdam Litke } 2531e4e574b7SAdam Litke 2532e4e574b7SAdam Litke allocated = 0; 2533e4e574b7SAdam Litke 2534e4e574b7SAdam Litke ret = -ENOMEM; 2535e4e574b7SAdam Litke retry: 2536db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2537e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 25380c397daeSMichal Hocko page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), 25392b21624fSMike Kravetz NUMA_NO_NODE, NULL); 254028073b02SHillf Danton if (!page) { 254128073b02SHillf Danton alloc_ok = false; 254228073b02SHillf Danton break; 254328073b02SHillf Danton } 2544e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 254569ed779aSDavid Rientjes cond_resched(); 2546e4e574b7SAdam Litke } 254728073b02SHillf Danton allocated += i; 2548e4e574b7SAdam Litke 2549e4e574b7SAdam Litke /* 2550e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 2551e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 2552e4e574b7SAdam Litke */ 2553db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2554a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 2555a5516438SAndi Kleen (h->free_huge_pages + allocated); 255628073b02SHillf Danton if (needed > 0) { 255728073b02SHillf Danton if (alloc_ok) 2558e4e574b7SAdam Litke goto retry; 255928073b02SHillf Danton /* 256028073b02SHillf Danton * We were not able to allocate enough pages to 256128073b02SHillf Danton * satisfy the entire reservation so we free what 256228073b02SHillf Danton * we've allocated so far. 256328073b02SHillf Danton */ 256428073b02SHillf Danton goto free; 256528073b02SHillf Danton } 2566e4e574b7SAdam Litke /* 2567e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 256825985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 2569e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 2570ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 2571ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 2572ac09b3a1SAdam Litke * before they are reserved. 2573e4e574b7SAdam Litke */ 2574e4e574b7SAdam Litke needed += allocated; 2575a5516438SAndi Kleen h->resv_huge_pages += delta; 2576e4e574b7SAdam Litke ret = 0; 2577a9869b83SNaoya Horiguchi 257819fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 257919fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 258019fc3f0aSAdam Litke if ((--needed) < 0) 258119fc3f0aSAdam Litke break; 2582b65a4edaSMike Kravetz /* Add the page to the hugetlb allocator */ 2583240d67a8SSidhartha Kumar enqueue_hugetlb_folio(h, page_folio(page)); 258419fc3f0aSAdam Litke } 258528073b02SHillf Danton free: 2586db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 258719fc3f0aSAdam Litke 2588b65a4edaSMike Kravetz /* 2589b65a4edaSMike Kravetz * Free unnecessary surplus pages to the buddy allocator. 2590b65a4edaSMike Kravetz * Pages have no ref count, call free_huge_page directly. 2591b65a4edaSMike Kravetz */ 2592c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 2593b65a4edaSMike Kravetz free_huge_page(page); 2594db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2595e4e574b7SAdam Litke 2596e4e574b7SAdam Litke return ret; 2597e4e574b7SAdam Litke } 2598e4e574b7SAdam Litke 2599e4e574b7SAdam Litke /* 2600e5bbc8a6SMike Kravetz * This routine has two main purposes: 2601e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 2602e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 2603e5bbc8a6SMike Kravetz * to the associated reservation map. 2604e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 2605e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 2606e4e574b7SAdam Litke */ 2607a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 2608a5516438SAndi Kleen unsigned long unused_resv_pages) 2609e4e574b7SAdam Litke { 2610e4e574b7SAdam Litke unsigned long nr_pages; 261110c6ec49SMike Kravetz struct page *page; 261210c6ec49SMike Kravetz LIST_HEAD(page_list); 261310c6ec49SMike Kravetz 26149487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 261510c6ec49SMike Kravetz /* Uncommit the reservation */ 261610c6ec49SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 2617e4e574b7SAdam Litke 2618c0531714SNaoya Horiguchi if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 2619e5bbc8a6SMike Kravetz goto out; 2620aa888a74SAndi Kleen 2621e5bbc8a6SMike Kravetz /* 2622e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 2623e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 2624e5bbc8a6SMike Kravetz */ 2625a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 2626e4e574b7SAdam Litke 2627685f3457SLee Schermerhorn /* 2628685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 26299b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 26309b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 26319b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 263210c6ec49SMike Kravetz * remove_pool_huge_page() will balance the freed pages across the 26339b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 2634685f3457SLee Schermerhorn */ 2635685f3457SLee Schermerhorn while (nr_pages--) { 263610c6ec49SMike Kravetz page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1); 263710c6ec49SMike Kravetz if (!page) 2638e5bbc8a6SMike Kravetz goto out; 263910c6ec49SMike Kravetz 264010c6ec49SMike Kravetz list_add(&page->lru, &page_list); 2641e4e574b7SAdam Litke } 2642e5bbc8a6SMike Kravetz 2643e5bbc8a6SMike Kravetz out: 2644db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 264510c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 2646db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2647e4e574b7SAdam Litke } 2648e4e574b7SAdam Litke 26495e911373SMike Kravetz 2650c37f9fb1SAndy Whitcroft /* 2651feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 26525e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 2653cf3ad20bSMike Kravetz * 2654cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 2655cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 2656cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 2657cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 2658cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 2659feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 2660feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 2661feba16e2SMike Kravetz * is called in such cases. 2662cf3ad20bSMike Kravetz * 2663cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 2664cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 2665cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 2666cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 2667cf3ad20bSMike Kravetz * take appropriate action. 266896b96a96SMike Kravetz * 266996b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 267096b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 267196b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 267296b96a96SMike Kravetz * reservation exists. 2673846be085SMike Kravetz * 2674846be085SMike Kravetz * vma_del_reservation is used in error paths where an entry in the reserve 2675846be085SMike Kravetz * map was created during huge page allocation and must be removed. It is to 2676846be085SMike Kravetz * be called after calling vma_needs_reservation to determine if a reservation 2677846be085SMike Kravetz * exists. 2678c37f9fb1SAndy Whitcroft */ 26795e911373SMike Kravetz enum vma_resv_mode { 26805e911373SMike Kravetz VMA_NEEDS_RESV, 26815e911373SMike Kravetz VMA_COMMIT_RESV, 2682feba16e2SMike Kravetz VMA_END_RESV, 268396b96a96SMike Kravetz VMA_ADD_RESV, 2684846be085SMike Kravetz VMA_DEL_RESV, 26855e911373SMike Kravetz }; 2686cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 2687cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 26885e911373SMike Kravetz enum vma_resv_mode mode) 2689c37f9fb1SAndy Whitcroft { 26904e35f483SJoonsoo Kim struct resv_map *resv; 26914e35f483SJoonsoo Kim pgoff_t idx; 2692cf3ad20bSMike Kravetz long ret; 26930db9d74eSMina Almasry long dummy_out_regions_needed; 2694c37f9fb1SAndy Whitcroft 26954e35f483SJoonsoo Kim resv = vma_resv_map(vma); 26964e35f483SJoonsoo Kim if (!resv) 2697c37f9fb1SAndy Whitcroft return 1; 2698c37f9fb1SAndy Whitcroft 26994e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 27005e911373SMike Kravetz switch (mode) { 27015e911373SMike Kravetz case VMA_NEEDS_RESV: 27020db9d74eSMina Almasry ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); 27030db9d74eSMina Almasry /* We assume that vma_reservation_* routines always operate on 27040db9d74eSMina Almasry * 1 page, and that adding to resv map a 1 page entry can only 27050db9d74eSMina Almasry * ever require 1 region. 27060db9d74eSMina Almasry */ 27070db9d74eSMina Almasry VM_BUG_ON(dummy_out_regions_needed != 1); 27085e911373SMike Kravetz break; 27095e911373SMike Kravetz case VMA_COMMIT_RESV: 2710075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 27110db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 27120db9d74eSMina Almasry VM_BUG_ON(ret < 0); 27135e911373SMike Kravetz break; 2714feba16e2SMike Kravetz case VMA_END_RESV: 27150db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 27165e911373SMike Kravetz ret = 0; 27175e911373SMike Kravetz break; 271896b96a96SMike Kravetz case VMA_ADD_RESV: 27190db9d74eSMina Almasry if (vma->vm_flags & VM_MAYSHARE) { 2720075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 27210db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 27220db9d74eSMina Almasry VM_BUG_ON(ret < 0); 27230db9d74eSMina Almasry } else { 27240db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 272596b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 272696b96a96SMike Kravetz } 272796b96a96SMike Kravetz break; 2728846be085SMike Kravetz case VMA_DEL_RESV: 2729846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 2730846be085SMike Kravetz region_abort(resv, idx, idx + 1, 1); 2731846be085SMike Kravetz ret = region_del(resv, idx, idx + 1); 2732846be085SMike Kravetz } else { 2733846be085SMike Kravetz ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 2734846be085SMike Kravetz /* region_add calls of range 1 should never fail. */ 2735846be085SMike Kravetz VM_BUG_ON(ret < 0); 2736846be085SMike Kravetz } 2737846be085SMike Kravetz break; 27385e911373SMike Kravetz default: 27395e911373SMike Kravetz BUG(); 27405e911373SMike Kravetz } 274184afd99bSAndy Whitcroft 2742846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV) 2743cf3ad20bSMike Kravetz return ret; 274467961f9dSMike Kravetz /* 2745bf3d12b9SMiaohe Lin * We know private mapping must have HPAGE_RESV_OWNER set. 2746bf3d12b9SMiaohe Lin * 274767961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 274867961f9dSMike Kravetz * However, a file associated with mapping could have been 274967961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 275067961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 275167961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 275267961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 275367961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 275467961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 275567961f9dSMike Kravetz * reservation has already been consumed. As a result, the 275667961f9dSMike Kravetz * return value of this routine is the opposite of the 275767961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 275867961f9dSMike Kravetz */ 2759bf3d12b9SMiaohe Lin if (ret > 0) 276067961f9dSMike Kravetz return 0; 2761bf3d12b9SMiaohe Lin if (ret == 0) 276267961f9dSMike Kravetz return 1; 2763bf3d12b9SMiaohe Lin return ret; 276484afd99bSAndy Whitcroft } 2765cf3ad20bSMike Kravetz 2766cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 2767a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 2768c37f9fb1SAndy Whitcroft { 27695e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 2770cf3ad20bSMike Kravetz } 2771c37f9fb1SAndy Whitcroft 2772cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 2773cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 2774cf3ad20bSMike Kravetz { 27755e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 27765e911373SMike Kravetz } 27775e911373SMike Kravetz 2778feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 27795e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 27805e911373SMike Kravetz { 2781feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 2782c37f9fb1SAndy Whitcroft } 2783c37f9fb1SAndy Whitcroft 278496b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 278596b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 278696b96a96SMike Kravetz { 278796b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 278896b96a96SMike Kravetz } 278996b96a96SMike Kravetz 2790846be085SMike Kravetz static long vma_del_reservation(struct hstate *h, 2791846be085SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 279296b96a96SMike Kravetz { 2793846be085SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV); 2794846be085SMike Kravetz } 2795846be085SMike Kravetz 2796846be085SMike Kravetz /* 2797846be085SMike Kravetz * This routine is called to restore reservation information on error paths. 2798846be085SMike Kravetz * It should ONLY be called for pages allocated via alloc_huge_page(), and 2799846be085SMike Kravetz * the hugetlb mutex should remain held when calling this routine. 2800846be085SMike Kravetz * 2801846be085SMike Kravetz * It handles two specific cases: 2802846be085SMike Kravetz * 1) A reservation was in place and the page consumed the reservation. 2803846be085SMike Kravetz * HPageRestoreReserve is set in the page. 2804846be085SMike Kravetz * 2) No reservation was in place for the page, so HPageRestoreReserve is 2805846be085SMike Kravetz * not set. However, alloc_huge_page always updates the reserve map. 2806846be085SMike Kravetz * 2807846be085SMike Kravetz * In case 1, free_huge_page later in the error path will increment the 2808846be085SMike Kravetz * global reserve count. But, free_huge_page does not have enough context 2809846be085SMike Kravetz * to adjust the reservation map. This case deals primarily with private 2810846be085SMike Kravetz * mappings. Adjust the reserve map here to be consistent with global 2811846be085SMike Kravetz * reserve count adjustments to be made by free_huge_page. Make sure the 2812846be085SMike Kravetz * reserve map indicates there is a reservation present. 2813846be085SMike Kravetz * 2814846be085SMike Kravetz * In case 2, simply undo reserve map modifications done by alloc_huge_page. 2815846be085SMike Kravetz */ 2816846be085SMike Kravetz void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma, 2817846be085SMike Kravetz unsigned long address, struct page *page) 2818846be085SMike Kravetz { 281996b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 282096b96a96SMike Kravetz 2821846be085SMike Kravetz if (HPageRestoreReserve(page)) { 2822846be085SMike Kravetz if (unlikely(rc < 0)) 282396b96a96SMike Kravetz /* 282496b96a96SMike Kravetz * Rare out of memory condition in reserve map 2825d6995da3SMike Kravetz * manipulation. Clear HPageRestoreReserve so that 282696b96a96SMike Kravetz * global reserve count will not be incremented 282796b96a96SMike Kravetz * by free_huge_page. This will make it appear 282896b96a96SMike Kravetz * as though the reservation for this page was 282996b96a96SMike Kravetz * consumed. This may prevent the task from 283096b96a96SMike Kravetz * faulting in the page at a later time. This 283196b96a96SMike Kravetz * is better than inconsistent global huge page 283296b96a96SMike Kravetz * accounting of reserve counts. 283396b96a96SMike Kravetz */ 2834d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 2835846be085SMike Kravetz else if (rc) 2836846be085SMike Kravetz (void)vma_add_reservation(h, vma, address); 2837846be085SMike Kravetz else 2838846be085SMike Kravetz vma_end_reservation(h, vma, address); 2839846be085SMike Kravetz } else { 2840846be085SMike Kravetz if (!rc) { 284196b96a96SMike Kravetz /* 2842846be085SMike Kravetz * This indicates there is an entry in the reserve map 2843c7b1850dSMike Kravetz * not added by alloc_huge_page. We know it was added 2844846be085SMike Kravetz * before the alloc_huge_page call, otherwise 2845846be085SMike Kravetz * HPageRestoreReserve would be set on the page. 2846846be085SMike Kravetz * Remove the entry so that a subsequent allocation 2847846be085SMike Kravetz * does not consume a reservation. 284896b96a96SMike Kravetz */ 2849846be085SMike Kravetz rc = vma_del_reservation(h, vma, address); 2850846be085SMike Kravetz if (rc < 0) 2851846be085SMike Kravetz /* 2852846be085SMike Kravetz * VERY rare out of memory condition. Since 2853846be085SMike Kravetz * we can not delete the entry, set 2854846be085SMike Kravetz * HPageRestoreReserve so that the reserve 2855846be085SMike Kravetz * count will be incremented when the page 2856846be085SMike Kravetz * is freed. This reserve will be consumed 2857846be085SMike Kravetz * on a subsequent allocation. 2858846be085SMike Kravetz */ 2859846be085SMike Kravetz SetHPageRestoreReserve(page); 2860846be085SMike Kravetz } else if (rc < 0) { 2861846be085SMike Kravetz /* 2862846be085SMike Kravetz * Rare out of memory condition from 2863846be085SMike Kravetz * vma_needs_reservation call. Memory allocation is 2864846be085SMike Kravetz * only attempted if a new entry is needed. Therefore, 2865846be085SMike Kravetz * this implies there is not an entry in the 2866846be085SMike Kravetz * reserve map. 2867846be085SMike Kravetz * 2868846be085SMike Kravetz * For shared mappings, no entry in the map indicates 2869846be085SMike Kravetz * no reservation. We are done. 2870846be085SMike Kravetz */ 2871846be085SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 2872846be085SMike Kravetz /* 2873846be085SMike Kravetz * For private mappings, no entry indicates 2874846be085SMike Kravetz * a reservation is present. Since we can 2875846be085SMike Kravetz * not add an entry, set SetHPageRestoreReserve 2876846be085SMike Kravetz * on the page so reserve count will be 2877846be085SMike Kravetz * incremented when freed. This reserve will 2878846be085SMike Kravetz * be consumed on a subsequent allocation. 2879846be085SMike Kravetz */ 2880846be085SMike Kravetz SetHPageRestoreReserve(page); 288196b96a96SMike Kravetz } else 2882846be085SMike Kravetz /* 2883846be085SMike Kravetz * No reservation present, do nothing 2884846be085SMike Kravetz */ 288596b96a96SMike Kravetz vma_end_reservation(h, vma, address); 288696b96a96SMike Kravetz } 288796b96a96SMike Kravetz } 288896b96a96SMike Kravetz 2889369fa227SOscar Salvador /* 289019fc1a7eSSidhartha Kumar * alloc_and_dissolve_hugetlb_folio - Allocate a new folio and dissolve 289119fc1a7eSSidhartha Kumar * the old one 2892369fa227SOscar Salvador * @h: struct hstate old page belongs to 289319fc1a7eSSidhartha Kumar * @old_folio: Old folio to dissolve 2894ae37c7ffSOscar Salvador * @list: List to isolate the page in case we need to 2895369fa227SOscar Salvador * Returns 0 on success, otherwise negated error. 2896369fa227SOscar Salvador */ 289719fc1a7eSSidhartha Kumar static int alloc_and_dissolve_hugetlb_folio(struct hstate *h, 289819fc1a7eSSidhartha Kumar struct folio *old_folio, struct list_head *list) 2899369fa227SOscar Salvador { 2900369fa227SOscar Salvador gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2901de656ed3SSidhartha Kumar int nid = folio_nid(old_folio); 2902de656ed3SSidhartha Kumar struct folio *new_folio; 2903369fa227SOscar Salvador int ret = 0; 2904369fa227SOscar Salvador 2905369fa227SOscar Salvador /* 290619fc1a7eSSidhartha Kumar * Before dissolving the folio, we need to allocate a new one for the 290719fc1a7eSSidhartha Kumar * pool to remain stable. Here, we allocate the folio and 'prep' it 2908f41f2ed4SMuchun Song * by doing everything but actually updating counters and adding to 2909f41f2ed4SMuchun Song * the pool. This simplifies and let us do most of the processing 2910f41f2ed4SMuchun Song * under the lock. 2911369fa227SOscar Salvador */ 291219fc1a7eSSidhartha Kumar new_folio = alloc_buddy_hugetlb_folio(h, gfp_mask, nid, NULL, NULL); 291319fc1a7eSSidhartha Kumar if (!new_folio) 2914369fa227SOscar Salvador return -ENOMEM; 2915de656ed3SSidhartha Kumar __prep_new_hugetlb_folio(h, new_folio); 2916369fa227SOscar Salvador 2917369fa227SOscar Salvador retry: 2918369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2919de656ed3SSidhartha Kumar if (!folio_test_hugetlb(old_folio)) { 2920369fa227SOscar Salvador /* 292119fc1a7eSSidhartha Kumar * Freed from under us. Drop new_folio too. 2922369fa227SOscar Salvador */ 2923369fa227SOscar Salvador goto free_new; 2924de656ed3SSidhartha Kumar } else if (folio_ref_count(old_folio)) { 2925369fa227SOscar Salvador /* 292619fc1a7eSSidhartha Kumar * Someone has grabbed the folio, try to isolate it here. 2927ae37c7ffSOscar Salvador * Fail with -EBUSY if not possible. 2928369fa227SOscar Salvador */ 2929ae37c7ffSOscar Salvador spin_unlock_irq(&hugetlb_lock); 293019fc1a7eSSidhartha Kumar ret = isolate_hugetlb(&old_folio->page, list); 2931ae37c7ffSOscar Salvador spin_lock_irq(&hugetlb_lock); 2932369fa227SOscar Salvador goto free_new; 2933de656ed3SSidhartha Kumar } else if (!folio_test_hugetlb_freed(old_folio)) { 2934369fa227SOscar Salvador /* 293519fc1a7eSSidhartha Kumar * Folio's refcount is 0 but it has not been enqueued in the 2936369fa227SOscar Salvador * freelist yet. Race window is small, so we can succeed here if 2937369fa227SOscar Salvador * we retry. 2938369fa227SOscar Salvador */ 2939369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2940369fa227SOscar Salvador cond_resched(); 2941369fa227SOscar Salvador goto retry; 2942369fa227SOscar Salvador } else { 2943369fa227SOscar Salvador /* 294419fc1a7eSSidhartha Kumar * Ok, old_folio is still a genuine free hugepage. Remove it from 2945369fa227SOscar Salvador * the freelist and decrease the counters. These will be 2946369fa227SOscar Salvador * incremented again when calling __prep_account_new_huge_page() 2947240d67a8SSidhartha Kumar * and enqueue_hugetlb_folio() for new_folio. The counters will 2948240d67a8SSidhartha Kumar * remain stable since this happens under the lock. 2949369fa227SOscar Salvador */ 2950cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, old_folio, false); 2951369fa227SOscar Salvador 2952369fa227SOscar Salvador /* 295319fc1a7eSSidhartha Kumar * Ref count on new_folio is already zero as it was dropped 2954b65a4edaSMike Kravetz * earlier. It can be directly added to the pool free list. 2955369fa227SOscar Salvador */ 2956369fa227SOscar Salvador __prep_account_new_huge_page(h, nid); 2957240d67a8SSidhartha Kumar enqueue_hugetlb_folio(h, new_folio); 2958369fa227SOscar Salvador 2959369fa227SOscar Salvador /* 296019fc1a7eSSidhartha Kumar * Folio has been replaced, we can safely free the old one. 2961369fa227SOscar Salvador */ 2962369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2963d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, old_folio, false); 2964369fa227SOscar Salvador } 2965369fa227SOscar Salvador 2966369fa227SOscar Salvador return ret; 2967369fa227SOscar Salvador 2968369fa227SOscar Salvador free_new: 2969369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 297019fc1a7eSSidhartha Kumar /* Folio has a zero ref count, but needs a ref to be freed */ 2971de656ed3SSidhartha Kumar folio_ref_unfreeze(new_folio, 1); 2972d6ef19e2SSidhartha Kumar update_and_free_hugetlb_folio(h, new_folio, false); 2973369fa227SOscar Salvador 2974369fa227SOscar Salvador return ret; 2975369fa227SOscar Salvador } 2976369fa227SOscar Salvador 2977ae37c7ffSOscar Salvador int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list) 2978369fa227SOscar Salvador { 2979369fa227SOscar Salvador struct hstate *h; 2980d5e33bd8SSidhartha Kumar struct folio *folio = page_folio(page); 2981ae37c7ffSOscar Salvador int ret = -EBUSY; 2982369fa227SOscar Salvador 2983369fa227SOscar Salvador /* 2984369fa227SOscar Salvador * The page might have been dissolved from under our feet, so make sure 2985369fa227SOscar Salvador * to carefully check the state under the lock. 2986369fa227SOscar Salvador * Return success when racing as if we dissolved the page ourselves. 2987369fa227SOscar Salvador */ 2988369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2989d5e33bd8SSidhartha Kumar if (folio_test_hugetlb(folio)) { 2990d5e33bd8SSidhartha Kumar h = folio_hstate(folio); 2991369fa227SOscar Salvador } else { 2992369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2993369fa227SOscar Salvador return 0; 2994369fa227SOscar Salvador } 2995369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2996369fa227SOscar Salvador 2997369fa227SOscar Salvador /* 2998369fa227SOscar Salvador * Fence off gigantic pages as there is a cyclic dependency between 2999369fa227SOscar Salvador * alloc_contig_range and them. Return -ENOMEM as this has the effect 3000369fa227SOscar Salvador * of bailing out right away without further retrying. 3001369fa227SOscar Salvador */ 3002369fa227SOscar Salvador if (hstate_is_gigantic(h)) 3003369fa227SOscar Salvador return -ENOMEM; 3004369fa227SOscar Salvador 3005d5e33bd8SSidhartha Kumar if (folio_ref_count(folio) && !isolate_hugetlb(&folio->page, list)) 3006ae37c7ffSOscar Salvador ret = 0; 3007d5e33bd8SSidhartha Kumar else if (!folio_ref_count(folio)) 300819fc1a7eSSidhartha Kumar ret = alloc_and_dissolve_hugetlb_folio(h, folio, list); 3009ae37c7ffSOscar Salvador 3010ae37c7ffSOscar Salvador return ret; 3011369fa227SOscar Salvador } 3012369fa227SOscar Salvador 301370c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 301404f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 3015348ea204SAdam Litke { 301690481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 3017a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3018348ea204SAdam Litke struct page *page; 3019d4ab0316SSidhartha Kumar struct folio *folio; 3020d85f69b0SMike Kravetz long map_chg, map_commit; 3021d85f69b0SMike Kravetz long gbl_chg; 30226d76dcf4SAneesh Kumar K.V int ret, idx; 30236d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 302408cf9fafSMina Almasry bool deferred_reserve; 30252fc39cecSAdam Litke 30266d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 3027a1e78772SMel Gorman /* 3028d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 3029d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 3030d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 3031a1e78772SMel Gorman */ 3032d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 3033d85f69b0SMike Kravetz if (map_chg < 0) 303476dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 3035d85f69b0SMike Kravetz 3036d85f69b0SMike Kravetz /* 3037d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 3038d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 3039d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 3040d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 3041d85f69b0SMike Kravetz * checked against any subpool limit. 3042d85f69b0SMike Kravetz */ 3043d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 3044d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 3045d85f69b0SMike Kravetz if (gbl_chg < 0) { 3046feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 304776dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 30485e911373SMike Kravetz } 304990d8b7e6SAdam Litke 3050d85f69b0SMike Kravetz /* 3051d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 3052d85f69b0SMike Kravetz * map, there could be reservations associated with the 3053d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 3054d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 3055d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 3056d85f69b0SMike Kravetz * the subpool reservations. 3057d85f69b0SMike Kravetz */ 3058d85f69b0SMike Kravetz if (avoid_reserve) 3059d85f69b0SMike Kravetz gbl_chg = 1; 3060d85f69b0SMike Kravetz } 3061d85f69b0SMike Kravetz 306208cf9fafSMina Almasry /* If this allocation is not consuming a reservation, charge it now. 306308cf9fafSMina Almasry */ 30646501fe5fSMiaohe Lin deferred_reserve = map_chg || avoid_reserve; 306508cf9fafSMina Almasry if (deferred_reserve) { 306608cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 306708cf9fafSMina Almasry idx, pages_per_huge_page(h), &h_cg); 30688f34af6fSJianyu Zhan if (ret) 30698f34af6fSJianyu Zhan goto out_subpool_put; 307008cf9fafSMina Almasry } 307108cf9fafSMina Almasry 307208cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 307308cf9fafSMina Almasry if (ret) 307408cf9fafSMina Almasry goto out_uncharge_cgroup_reservation; 30758f34af6fSJianyu Zhan 3076db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3077d85f69b0SMike Kravetz /* 3078d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 3079d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 3080d85f69b0SMike Kravetz * a reservation exists for the allocation. 3081d85f69b0SMike Kravetz */ 3082d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 308381a6fcaeSJoonsoo Kim if (!page) { 3084db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 30850c397daeSMichal Hocko page = alloc_buddy_huge_page_with_mpol(h, vma, addr); 30868f34af6fSJianyu Zhan if (!page) 30878f34af6fSJianyu Zhan goto out_uncharge_cgroup; 308812df140fSRik van Riel spin_lock_irq(&hugetlb_lock); 3089a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 3090d6995da3SMike Kravetz SetHPageRestoreReserve(page); 3091a88c7695SNaoya Horiguchi h->resv_huge_pages--; 3092a88c7695SNaoya Horiguchi } 309315a8d68eSWei Yang list_add(&page->lru, &h->hugepage_activelist); 30942b21624fSMike Kravetz set_page_refcounted(page); 309581a6fcaeSJoonsoo Kim /* Fall through */ 3096a1e78772SMel Gorman } 3097d4ab0316SSidhartha Kumar folio = page_folio(page); 309881a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 309908cf9fafSMina Almasry /* If allocation is not consuming a reservation, also store the 310008cf9fafSMina Almasry * hugetlb_cgroup pointer on the page. 310108cf9fafSMina Almasry */ 310208cf9fafSMina Almasry if (deferred_reserve) { 310308cf9fafSMina Almasry hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), 310408cf9fafSMina Almasry h_cg, page); 310508cf9fafSMina Almasry } 310608cf9fafSMina Almasry 3107db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3108a1e78772SMel Gorman 3109d6995da3SMike Kravetz hugetlb_set_page_subpool(page, spool); 3110a1e78772SMel Gorman 3111d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 3112d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 311333039678SMike Kravetz /* 311433039678SMike Kravetz * The page was added to the reservation map between 311533039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 311633039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 311733039678SMike Kravetz * Adjust for the subpool count incremented above AND 311833039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 311933039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 312033039678SMike Kravetz * no longer applies. 312133039678SMike Kravetz */ 312233039678SMike Kravetz long rsv_adjust; 312333039678SMike Kravetz 312433039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 312533039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 312679aa925bSMike Kravetz if (deferred_reserve) 3127d4ab0316SSidhartha Kumar hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h), 3128d4ab0316SSidhartha Kumar pages_per_huge_page(h), folio); 312933039678SMike Kravetz } 31307893d1d5SAdam Litke return page; 31318f34af6fSJianyu Zhan 31328f34af6fSJianyu Zhan out_uncharge_cgroup: 31338f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 313408cf9fafSMina Almasry out_uncharge_cgroup_reservation: 313508cf9fafSMina Almasry if (deferred_reserve) 313608cf9fafSMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), 313708cf9fafSMina Almasry h_cg); 31388f34af6fSJianyu Zhan out_subpool_put: 3139d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 31408f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 3141feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 31428f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 3143b45b5bd6SDavid Gibson } 3144b45b5bd6SDavid Gibson 3145b5389086SZhenguo Yao int alloc_bootmem_huge_page(struct hstate *h, int nid) 3146e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 3147b5389086SZhenguo Yao int __alloc_bootmem_huge_page(struct hstate *h, int nid) 3148aa888a74SAndi Kleen { 3149b5389086SZhenguo Yao struct huge_bootmem_page *m = NULL; /* initialize for clang */ 3150b2261026SJoonsoo Kim int nr_nodes, node; 3151aa888a74SAndi Kleen 3152b5389086SZhenguo Yao /* do node specific alloc */ 3153b5389086SZhenguo Yao if (nid != NUMA_NO_NODE) { 3154b5389086SZhenguo Yao m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h), 3155b5389086SZhenguo Yao 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid); 3156b5389086SZhenguo Yao if (!m) 3157b5389086SZhenguo Yao return 0; 3158b5389086SZhenguo Yao goto found; 3159b5389086SZhenguo Yao } 3160b5389086SZhenguo Yao /* allocate from next node when distributing huge pages */ 3161b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 3162b5389086SZhenguo Yao m = memblock_alloc_try_nid_raw( 31638b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 316497ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 3165aa888a74SAndi Kleen /* 3166aa888a74SAndi Kleen * Use the beginning of the huge page to store the 3167aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 3168aa888a74SAndi Kleen * puts them into the mem_map). 3169aa888a74SAndi Kleen */ 3170b5389086SZhenguo Yao if (!m) 3171b5389086SZhenguo Yao return 0; 3172aa888a74SAndi Kleen goto found; 3173aa888a74SAndi Kleen } 3174aa888a74SAndi Kleen 3175aa888a74SAndi Kleen found: 3176aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 3177330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 3178aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 3179aa888a74SAndi Kleen m->hstate = h; 3180aa888a74SAndi Kleen return 1; 3181aa888a74SAndi Kleen } 3182aa888a74SAndi Kleen 318348b8d744SMike Kravetz /* 318448b8d744SMike Kravetz * Put bootmem huge pages into the standard lists after mem_map is up. 318548b8d744SMike Kravetz * Note: This only applies to gigantic (order > MAX_ORDER) pages. 318648b8d744SMike Kravetz */ 3187aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 3188aa888a74SAndi Kleen { 3189aa888a74SAndi Kleen struct huge_bootmem_page *m; 3190aa888a74SAndi Kleen 3191aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 319240d18ebfSMike Kravetz struct page *page = virt_to_page(m); 31937f325a8dSSidhartha Kumar struct folio *folio = page_folio(page); 3194aa888a74SAndi Kleen struct hstate *h = m->hstate; 3195ee8f248dSBecky Bruce 319648b8d744SMike Kravetz VM_BUG_ON(!hstate_is_gigantic(h)); 3197d1c60955SSidhartha Kumar WARN_ON(folio_ref_count(folio) != 1); 3198d1c60955SSidhartha Kumar if (prep_compound_gigantic_folio(folio, huge_page_order(h))) { 3199d1c60955SSidhartha Kumar WARN_ON(folio_test_reserved(folio)); 3200d1c60955SSidhartha Kumar prep_new_hugetlb_folio(h, folio, folio_nid(folio)); 32012b21624fSMike Kravetz free_huge_page(page); /* add to the hugepage allocator */ 32027118fc29SMike Kravetz } else { 3203416d85edSMike Kravetz /* VERY unlikely inflated ref count on a tail page */ 32047f325a8dSSidhartha Kumar free_gigantic_folio(folio, huge_page_order(h)); 32057118fc29SMike Kravetz } 3206af0fb9dfSMichal Hocko 3207b0320c7bSRafael Aquini /* 320848b8d744SMike Kravetz * We need to restore the 'stolen' pages to totalram_pages 320948b8d744SMike Kravetz * in order to fix confusing memory reports from free(1) and 321048b8d744SMike Kravetz * other side-effects, like CommitLimit going negative. 3211b0320c7bSRafael Aquini */ 3212c78a7f36SMiaohe Lin adjust_managed_page_count(page, pages_per_huge_page(h)); 3213520495feSCannon Matthews cond_resched(); 3214aa888a74SAndi Kleen } 3215aa888a74SAndi Kleen } 3216b5389086SZhenguo Yao static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid) 3217b5389086SZhenguo Yao { 3218b5389086SZhenguo Yao unsigned long i; 3219b5389086SZhenguo Yao char buf[32]; 3220b5389086SZhenguo Yao 3221b5389086SZhenguo Yao for (i = 0; i < h->max_huge_pages_node[nid]; ++i) { 3222b5389086SZhenguo Yao if (hstate_is_gigantic(h)) { 3223b5389086SZhenguo Yao if (!alloc_bootmem_huge_page(h, nid)) 3224b5389086SZhenguo Yao break; 3225b5389086SZhenguo Yao } else { 322619fc1a7eSSidhartha Kumar struct folio *folio; 3227b5389086SZhenguo Yao gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 3228b5389086SZhenguo Yao 322919fc1a7eSSidhartha Kumar folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, 3230b5389086SZhenguo Yao &node_states[N_MEMORY], NULL); 323119fc1a7eSSidhartha Kumar if (!folio) 3232b5389086SZhenguo Yao break; 323319fc1a7eSSidhartha Kumar free_huge_page(&folio->page); /* free it into the hugepage allocator */ 3234b5389086SZhenguo Yao } 3235b5389086SZhenguo Yao cond_resched(); 3236b5389086SZhenguo Yao } 3237b5389086SZhenguo Yao if (i == h->max_huge_pages_node[nid]) 3238b5389086SZhenguo Yao return; 3239b5389086SZhenguo Yao 3240b5389086SZhenguo Yao string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3241b5389086SZhenguo Yao pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n", 3242b5389086SZhenguo Yao h->max_huge_pages_node[nid], buf, nid, i); 3243b5389086SZhenguo Yao h->max_huge_pages -= (h->max_huge_pages_node[nid] - i); 3244b5389086SZhenguo Yao h->max_huge_pages_node[nid] = i; 3245b5389086SZhenguo Yao } 3246aa888a74SAndi Kleen 32478faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 32481da177e4SLinus Torvalds { 32491da177e4SLinus Torvalds unsigned long i; 3250f60858f9SMike Kravetz nodemask_t *node_alloc_noretry; 3251b5389086SZhenguo Yao bool node_specific_alloc = false; 3252f60858f9SMike Kravetz 3253b5389086SZhenguo Yao /* skip gigantic hugepages allocation if hugetlb_cma enabled */ 3254b5389086SZhenguo Yao if (hstate_is_gigantic(h) && hugetlb_cma_size) { 3255b5389086SZhenguo Yao pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); 3256b5389086SZhenguo Yao return; 3257b5389086SZhenguo Yao } 3258b5389086SZhenguo Yao 3259b5389086SZhenguo Yao /* do node specific alloc */ 32600a7a0f6fSPeng Liu for_each_online_node(i) { 3261b5389086SZhenguo Yao if (h->max_huge_pages_node[i] > 0) { 3262b5389086SZhenguo Yao hugetlb_hstate_alloc_pages_onenode(h, i); 3263b5389086SZhenguo Yao node_specific_alloc = true; 3264b5389086SZhenguo Yao } 3265b5389086SZhenguo Yao } 3266b5389086SZhenguo Yao 3267b5389086SZhenguo Yao if (node_specific_alloc) 3268b5389086SZhenguo Yao return; 3269b5389086SZhenguo Yao 3270b5389086SZhenguo Yao /* below will do all node balanced alloc */ 3271f60858f9SMike Kravetz if (!hstate_is_gigantic(h)) { 3272f60858f9SMike Kravetz /* 3273f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3274f60858f9SMike Kravetz * Ignore errors as lower level routines can deal with 3275f60858f9SMike Kravetz * node_alloc_noretry == NULL. If this kmalloc fails at boot 3276f60858f9SMike Kravetz * time, we are likely in bigger trouble. 3277f60858f9SMike Kravetz */ 3278f60858f9SMike Kravetz node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), 3279f60858f9SMike Kravetz GFP_KERNEL); 3280f60858f9SMike Kravetz } else { 3281f60858f9SMike Kravetz /* allocations done at boot time */ 3282f60858f9SMike Kravetz node_alloc_noretry = NULL; 3283f60858f9SMike Kravetz } 3284f60858f9SMike Kravetz 3285f60858f9SMike Kravetz /* bit mask controlling how hard we retry per-node allocations */ 3286f60858f9SMike Kravetz if (node_alloc_noretry) 3287f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 32881da177e4SLinus Torvalds 3289e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 3290bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 3291b5389086SZhenguo Yao if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE)) 3292aa888a74SAndi Kleen break; 32930c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 3294f60858f9SMike Kravetz &node_states[N_MEMORY], 3295f60858f9SMike Kravetz node_alloc_noretry)) 32961da177e4SLinus Torvalds break; 329769ed779aSDavid Rientjes cond_resched(); 32981da177e4SLinus Torvalds } 3299d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 3300d715cf80SLiam R. Howlett char buf[32]; 3301d715cf80SLiam R. Howlett 3302c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3303d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 3304d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 33058faa8b07SAndi Kleen h->max_huge_pages = i; 3306e5ff2159SAndi Kleen } 3307f60858f9SMike Kravetz kfree(node_alloc_noretry); 3308d715cf80SLiam R. Howlett } 3309e5ff2159SAndi Kleen 3310e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 3311e5ff2159SAndi Kleen { 331279dfc695SMike Kravetz struct hstate *h, *h2; 3313e5ff2159SAndi Kleen 3314e5ff2159SAndi Kleen for_each_hstate(h) { 33158faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 3316bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 33178faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 331879dfc695SMike Kravetz 331979dfc695SMike Kravetz /* 332079dfc695SMike Kravetz * Set demote order for each hstate. Note that 332179dfc695SMike Kravetz * h->demote_order is initially 0. 332279dfc695SMike Kravetz * - We can not demote gigantic pages if runtime freeing 332379dfc695SMike Kravetz * is not supported, so skip this. 3324a01f4390SMike Kravetz * - If CMA allocation is possible, we can not demote 3325a01f4390SMike Kravetz * HUGETLB_PAGE_ORDER or smaller size pages. 332679dfc695SMike Kravetz */ 332779dfc695SMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 332879dfc695SMike Kravetz continue; 3329a01f4390SMike Kravetz if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER) 3330a01f4390SMike Kravetz continue; 333179dfc695SMike Kravetz for_each_hstate(h2) { 333279dfc695SMike Kravetz if (h2 == h) 333379dfc695SMike Kravetz continue; 333479dfc695SMike Kravetz if (h2->order < h->order && 333579dfc695SMike Kravetz h2->order > h->demote_order) 333679dfc695SMike Kravetz h->demote_order = h2->order; 333779dfc695SMike Kravetz } 3338e5ff2159SAndi Kleen } 3339e5ff2159SAndi Kleen } 3340e5ff2159SAndi Kleen 3341e5ff2159SAndi Kleen static void __init report_hugepages(void) 3342e5ff2159SAndi Kleen { 3343e5ff2159SAndi Kleen struct hstate *h; 3344e5ff2159SAndi Kleen 3345e5ff2159SAndi Kleen for_each_hstate(h) { 33464abd32dbSAndi Kleen char buf[32]; 3347c6247f72SMatthew Wilcox 3348c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 33496213834cSMuchun Song pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n", 3350c6247f72SMatthew Wilcox buf, h->free_huge_pages); 33516213834cSMuchun Song pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n", 33526213834cSMuchun Song hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf); 3353e5ff2159SAndi Kleen } 3354e5ff2159SAndi Kleen } 3355e5ff2159SAndi Kleen 33561da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 33576ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 33586ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 33591da177e4SLinus Torvalds { 33604415cc8dSChristoph Lameter int i; 33611121828aSMike Kravetz LIST_HEAD(page_list); 33624415cc8dSChristoph Lameter 33639487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 3364bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3365aa888a74SAndi Kleen return; 3366aa888a74SAndi Kleen 33671121828aSMike Kravetz /* 33681121828aSMike Kravetz * Collect pages to be freed on a list, and free after dropping lock 33691121828aSMike Kravetz */ 33706ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 337110c6ec49SMike Kravetz struct page *page, *next; 3372a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 3373a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 3374a5516438SAndi Kleen if (count >= h->nr_huge_pages) 33751121828aSMike Kravetz goto out; 33761da177e4SLinus Torvalds if (PageHighMem(page)) 33771da177e4SLinus Torvalds continue; 3378cfd5082bSSidhartha Kumar remove_hugetlb_folio(h, page_folio(page), false); 33791121828aSMike Kravetz list_add(&page->lru, &page_list); 33801121828aSMike Kravetz } 33811121828aSMike Kravetz } 33821121828aSMike Kravetz 33831121828aSMike Kravetz out: 3384db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 338510c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3386db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 33871da177e4SLinus Torvalds } 33881da177e4SLinus Torvalds #else 33896ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 33906ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 33911da177e4SLinus Torvalds { 33921da177e4SLinus Torvalds } 33931da177e4SLinus Torvalds #endif 33941da177e4SLinus Torvalds 339520a0307cSWu Fengguang /* 339620a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 339720a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 339820a0307cSWu Fengguang * Returns 1 if an adjustment was made. 339920a0307cSWu Fengguang */ 34006ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 34016ae11b27SLee Schermerhorn int delta) 340220a0307cSWu Fengguang { 3403b2261026SJoonsoo Kim int nr_nodes, node; 340420a0307cSWu Fengguang 34059487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 340620a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 340720a0307cSWu Fengguang 3408e8c5c824SLee Schermerhorn if (delta < 0) { 3409b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 3410b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 3411b2261026SJoonsoo Kim goto found; 3412b2261026SJoonsoo Kim } 3413b2261026SJoonsoo Kim } else { 3414b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 3415b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 3416b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 3417b2261026SJoonsoo Kim goto found; 3418e8c5c824SLee Schermerhorn } 34199a76db09SLee Schermerhorn } 3420b2261026SJoonsoo Kim return 0; 342120a0307cSWu Fengguang 3422b2261026SJoonsoo Kim found: 342320a0307cSWu Fengguang h->surplus_huge_pages += delta; 3424b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 3425b2261026SJoonsoo Kim return 1; 342620a0307cSWu Fengguang } 342720a0307cSWu Fengguang 3428a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 3429fd875dcaSMike Kravetz static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 34306ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 34311da177e4SLinus Torvalds { 34327893d1d5SAdam Litke unsigned long min_count, ret; 343310c6ec49SMike Kravetz struct page *page; 343410c6ec49SMike Kravetz LIST_HEAD(page_list); 3435f60858f9SMike Kravetz NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); 3436f60858f9SMike Kravetz 3437f60858f9SMike Kravetz /* 3438f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3439f60858f9SMike Kravetz * If we can not allocate the bit mask, do not attempt to allocate 3440f60858f9SMike Kravetz * the requested huge pages. 3441f60858f9SMike Kravetz */ 3442f60858f9SMike Kravetz if (node_alloc_noretry) 3443f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 3444f60858f9SMike Kravetz else 3445f60858f9SMike Kravetz return -ENOMEM; 34461da177e4SLinus Torvalds 344729383967SMike Kravetz /* 344829383967SMike Kravetz * resize_lock mutex prevents concurrent adjustments to number of 344929383967SMike Kravetz * pages in hstate via the proc/sysfs interfaces. 345029383967SMike Kravetz */ 345129383967SMike Kravetz mutex_lock(&h->resize_lock); 3452b65d4adbSMuchun Song flush_free_hpage_work(h); 3453db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 34544eb0716eSAlexandre Ghiti 34554eb0716eSAlexandre Ghiti /* 3456fd875dcaSMike Kravetz * Check for a node specific request. 3457fd875dcaSMike Kravetz * Changing node specific huge page count may require a corresponding 3458fd875dcaSMike Kravetz * change to the global count. In any case, the passed node mask 3459fd875dcaSMike Kravetz * (nodes_allowed) will restrict alloc/free to the specified node. 3460fd875dcaSMike Kravetz */ 3461fd875dcaSMike Kravetz if (nid != NUMA_NO_NODE) { 3462fd875dcaSMike Kravetz unsigned long old_count = count; 3463fd875dcaSMike Kravetz 3464fd875dcaSMike Kravetz count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 3465fd875dcaSMike Kravetz /* 3466fd875dcaSMike Kravetz * User may have specified a large count value which caused the 3467fd875dcaSMike Kravetz * above calculation to overflow. In this case, they wanted 3468fd875dcaSMike Kravetz * to allocate as many huge pages as possible. Set count to 3469fd875dcaSMike Kravetz * largest possible value to align with their intention. 3470fd875dcaSMike Kravetz */ 3471fd875dcaSMike Kravetz if (count < old_count) 3472fd875dcaSMike Kravetz count = ULONG_MAX; 3473fd875dcaSMike Kravetz } 3474fd875dcaSMike Kravetz 3475fd875dcaSMike Kravetz /* 34764eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 34774eb0716eSAlexandre Ghiti * page range allocation. 34784eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 34794eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 34804eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 34814eb0716eSAlexandre Ghiti */ 34824eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 34834eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 3484db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 348529383967SMike Kravetz mutex_unlock(&h->resize_lock); 3486f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 34874eb0716eSAlexandre Ghiti return -EINVAL; 34884eb0716eSAlexandre Ghiti } 34894eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 34904eb0716eSAlexandre Ghiti } 3491aa888a74SAndi Kleen 34927893d1d5SAdam Litke /* 34937893d1d5SAdam Litke * Increase the pool size 34947893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 34957893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 3496d1c3fb1fSNishanth Aravamudan * 34970c397daeSMichal Hocko * We might race with alloc_surplus_huge_page() here and be unable 3498d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 3499d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 3500d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 3501d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 35027893d1d5SAdam Litke */ 3503a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 35046ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 35057893d1d5SAdam Litke break; 35067893d1d5SAdam Litke } 35077893d1d5SAdam Litke 3508a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 35097893d1d5SAdam Litke /* 35107893d1d5SAdam Litke * If this allocation races such that we no longer need the 35117893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 35127893d1d5SAdam Litke * and reducing the surplus. 35137893d1d5SAdam Litke */ 3514db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3515649920c6SJia He 3516649920c6SJia He /* yield cpu to avoid soft lockup */ 3517649920c6SJia He cond_resched(); 3518649920c6SJia He 3519f60858f9SMike Kravetz ret = alloc_pool_huge_page(h, nodes_allowed, 3520f60858f9SMike Kravetz node_alloc_noretry); 3521db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 35227893d1d5SAdam Litke if (!ret) 35237893d1d5SAdam Litke goto out; 35247893d1d5SAdam Litke 3525536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 3526536240f2SMel Gorman if (signal_pending(current)) 3527536240f2SMel Gorman goto out; 35287893d1d5SAdam Litke } 35297893d1d5SAdam Litke 35307893d1d5SAdam Litke /* 35317893d1d5SAdam Litke * Decrease the pool size 35327893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 35337893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 35347893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 35357893d1d5SAdam Litke * to the desired size as pages become free. 3536d1c3fb1fSNishanth Aravamudan * 3537d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 3538d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 3539d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 35400c397daeSMichal Hocko * alloc_surplus_huge_page() is checking the global counter, 3541d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 3542d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 3543d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 35447893d1d5SAdam Litke */ 3545a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 35466b0c880dSAdam Litke min_count = max(count, min_count); 35476ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 354810c6ec49SMike Kravetz 354910c6ec49SMike Kravetz /* 355010c6ec49SMike Kravetz * Collect pages to be removed on list without dropping lock 355110c6ec49SMike Kravetz */ 3552a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 355310c6ec49SMike Kravetz page = remove_pool_huge_page(h, nodes_allowed, 0); 355410c6ec49SMike Kravetz if (!page) 35551da177e4SLinus Torvalds break; 355610c6ec49SMike Kravetz 355710c6ec49SMike Kravetz list_add(&page->lru, &page_list); 35581da177e4SLinus Torvalds } 355910c6ec49SMike Kravetz /* free the pages after dropping lock */ 3560db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 356110c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3562b65d4adbSMuchun Song flush_free_hpage_work(h); 3563db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 356410c6ec49SMike Kravetz 3565a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 35666ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 35677893d1d5SAdam Litke break; 35687893d1d5SAdam Litke } 35697893d1d5SAdam Litke out: 35704eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 3571db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 357229383967SMike Kravetz mutex_unlock(&h->resize_lock); 35734eb0716eSAlexandre Ghiti 3574f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 3575f60858f9SMike Kravetz 35764eb0716eSAlexandre Ghiti return 0; 35771da177e4SLinus Torvalds } 35781da177e4SLinus Torvalds 35798531fc6fSMike Kravetz static int demote_free_huge_page(struct hstate *h, struct page *page) 35808531fc6fSMike Kravetz { 35818531fc6fSMike Kravetz int i, nid = page_to_nid(page); 35828531fc6fSMike Kravetz struct hstate *target_hstate; 3583911565b8SSidhartha Kumar struct folio *folio = page_folio(page); 358431731452SDoug Berger struct page *subpage; 35858531fc6fSMike Kravetz int rc = 0; 35868531fc6fSMike Kravetz 35878531fc6fSMike Kravetz target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order); 35888531fc6fSMike Kravetz 3589cfd5082bSSidhartha Kumar remove_hugetlb_folio_for_demote(h, folio, false); 35908531fc6fSMike Kravetz spin_unlock_irq(&hugetlb_lock); 35918531fc6fSMike Kravetz 35926213834cSMuchun Song rc = hugetlb_vmemmap_restore(h, page); 35938531fc6fSMike Kravetz if (rc) { 35948531fc6fSMike Kravetz /* Allocation of vmemmmap failed, we can not demote page */ 35958531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 35968531fc6fSMike Kravetz set_page_refcounted(page); 35972f6c57d6SSidhartha Kumar add_hugetlb_folio(h, page_folio(page), false); 35988531fc6fSMike Kravetz return rc; 35998531fc6fSMike Kravetz } 36008531fc6fSMike Kravetz 36018531fc6fSMike Kravetz /* 3602911565b8SSidhartha Kumar * Use destroy_compound_hugetlb_folio_for_demote for all huge page 36038531fc6fSMike Kravetz * sizes as it will not ref count pages. 36048531fc6fSMike Kravetz */ 3605911565b8SSidhartha Kumar destroy_compound_hugetlb_folio_for_demote(folio, huge_page_order(h)); 36068531fc6fSMike Kravetz 36078531fc6fSMike Kravetz /* 36088531fc6fSMike Kravetz * Taking target hstate mutex synchronizes with set_max_huge_pages. 36098531fc6fSMike Kravetz * Without the mutex, pages added to target hstate could be marked 36108531fc6fSMike Kravetz * as surplus. 36118531fc6fSMike Kravetz * 36128531fc6fSMike Kravetz * Note that we already hold h->resize_lock. To prevent deadlock, 36138531fc6fSMike Kravetz * use the convention of always taking larger size hstate mutex first. 36148531fc6fSMike Kravetz */ 36158531fc6fSMike Kravetz mutex_lock(&target_hstate->resize_lock); 36168531fc6fSMike Kravetz for (i = 0; i < pages_per_huge_page(h); 36178531fc6fSMike Kravetz i += pages_per_huge_page(target_hstate)) { 361831731452SDoug Berger subpage = nth_page(page, i); 3619d1c60955SSidhartha Kumar folio = page_folio(subpage); 36208531fc6fSMike Kravetz if (hstate_is_gigantic(target_hstate)) 3621d1c60955SSidhartha Kumar prep_compound_gigantic_folio_for_demote(folio, 36228531fc6fSMike Kravetz target_hstate->order); 36238531fc6fSMike Kravetz else 362431731452SDoug Berger prep_compound_page(subpage, target_hstate->order); 362531731452SDoug Berger set_page_private(subpage, 0); 3626d1c60955SSidhartha Kumar prep_new_hugetlb_folio(target_hstate, folio, nid); 36272b21624fSMike Kravetz free_huge_page(subpage); 36288531fc6fSMike Kravetz } 36298531fc6fSMike Kravetz mutex_unlock(&target_hstate->resize_lock); 36308531fc6fSMike Kravetz 36318531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 36328531fc6fSMike Kravetz 36338531fc6fSMike Kravetz /* 36348531fc6fSMike Kravetz * Not absolutely necessary, but for consistency update max_huge_pages 36358531fc6fSMike Kravetz * based on pool changes for the demoted page. 36368531fc6fSMike Kravetz */ 36378531fc6fSMike Kravetz h->max_huge_pages--; 3638a43a83c7SMiaohe Lin target_hstate->max_huge_pages += 3639a43a83c7SMiaohe Lin pages_per_huge_page(h) / pages_per_huge_page(target_hstate); 36408531fc6fSMike Kravetz 36418531fc6fSMike Kravetz return rc; 36428531fc6fSMike Kravetz } 36438531fc6fSMike Kravetz 364479dfc695SMike Kravetz static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 364579dfc695SMike Kravetz __must_hold(&hugetlb_lock) 364679dfc695SMike Kravetz { 36478531fc6fSMike Kravetz int nr_nodes, node; 36488531fc6fSMike Kravetz struct page *page; 364979dfc695SMike Kravetz 365079dfc695SMike Kravetz lockdep_assert_held(&hugetlb_lock); 365179dfc695SMike Kravetz 365279dfc695SMike Kravetz /* We should never get here if no demote order */ 365379dfc695SMike Kravetz if (!h->demote_order) { 365479dfc695SMike Kravetz pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n"); 365579dfc695SMike Kravetz return -EINVAL; /* internal error */ 365679dfc695SMike Kravetz } 365779dfc695SMike Kravetz 36588531fc6fSMike Kravetz for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 36595a317412SMike Kravetz list_for_each_entry(page, &h->hugepage_freelists[node], lru) { 36605a317412SMike Kravetz if (PageHWPoison(page)) 36615a317412SMike Kravetz continue; 36625a317412SMike Kravetz 36635a317412SMike Kravetz return demote_free_huge_page(h, page); 36648531fc6fSMike Kravetz } 36658531fc6fSMike Kravetz } 36668531fc6fSMike Kravetz 36675a317412SMike Kravetz /* 36685a317412SMike Kravetz * Only way to get here is if all pages on free lists are poisoned. 36695a317412SMike Kravetz * Return -EBUSY so that caller will not retry. 36705a317412SMike Kravetz */ 36715a317412SMike Kravetz return -EBUSY; 367279dfc695SMike Kravetz } 367379dfc695SMike Kravetz 3674a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 3675a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 3676a3437870SNishanth Aravamudan 367779dfc695SMike Kravetz #define HSTATE_ATTR_WO(_name) \ 367879dfc695SMike Kravetz static struct kobj_attribute _name##_attr = __ATTR_WO(_name) 367979dfc695SMike Kravetz 3680a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 368198bc26acSMiaohe Lin static struct kobj_attribute _name##_attr = __ATTR_RW(_name) 3682a3437870SNishanth Aravamudan 3683a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 3684a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 3685a3437870SNishanth Aravamudan 36869a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 36879a305230SLee Schermerhorn 36889a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 3689a3437870SNishanth Aravamudan { 3690a3437870SNishanth Aravamudan int i; 36919a305230SLee Schermerhorn 3692a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 36939a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 36949a305230SLee Schermerhorn if (nidp) 36959a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 3696a3437870SNishanth Aravamudan return &hstates[i]; 36979a305230SLee Schermerhorn } 36989a305230SLee Schermerhorn 36999a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 3700a3437870SNishanth Aravamudan } 3701a3437870SNishanth Aravamudan 370206808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 3703a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3704a3437870SNishanth Aravamudan { 37059a305230SLee Schermerhorn struct hstate *h; 37069a305230SLee Schermerhorn unsigned long nr_huge_pages; 37079a305230SLee Schermerhorn int nid; 37089a305230SLee Schermerhorn 37099a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 37109a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 37119a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 37129a305230SLee Schermerhorn else 37139a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 37149a305230SLee Schermerhorn 3715ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", nr_huge_pages); 3716a3437870SNishanth Aravamudan } 3717adbe8726SEric B Munson 3718238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 3719238d3c13SDavid Rientjes struct hstate *h, int nid, 3720238d3c13SDavid Rientjes unsigned long count, size_t len) 3721a3437870SNishanth Aravamudan { 3722a3437870SNishanth Aravamudan int err; 37232d0adf7eSOscar Salvador nodemask_t nodes_allowed, *n_mask; 3724a3437870SNishanth Aravamudan 37252d0adf7eSOscar Salvador if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 37262d0adf7eSOscar Salvador return -EINVAL; 3727adbe8726SEric B Munson 37289a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 37299a305230SLee Schermerhorn /* 37309a305230SLee Schermerhorn * global hstate attribute 37319a305230SLee Schermerhorn */ 37329a305230SLee Schermerhorn if (!(obey_mempolicy && 37332d0adf7eSOscar Salvador init_nodemask_of_mempolicy(&nodes_allowed))) 37342d0adf7eSOscar Salvador n_mask = &node_states[N_MEMORY]; 37352d0adf7eSOscar Salvador else 37362d0adf7eSOscar Salvador n_mask = &nodes_allowed; 37372d0adf7eSOscar Salvador } else { 37389a305230SLee Schermerhorn /* 3739fd875dcaSMike Kravetz * Node specific request. count adjustment happens in 3740fd875dcaSMike Kravetz * set_max_huge_pages() after acquiring hugetlb_lock. 37419a305230SLee Schermerhorn */ 37422d0adf7eSOscar Salvador init_nodemask_of_node(&nodes_allowed, nid); 37432d0adf7eSOscar Salvador n_mask = &nodes_allowed; 3744fd875dcaSMike Kravetz } 37459a305230SLee Schermerhorn 37462d0adf7eSOscar Salvador err = set_max_huge_pages(h, count, nid, n_mask); 374706808b08SLee Schermerhorn 37484eb0716eSAlexandre Ghiti return err ? err : len; 374906808b08SLee Schermerhorn } 375006808b08SLee Schermerhorn 3751238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 3752238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 3753238d3c13SDavid Rientjes size_t len) 3754238d3c13SDavid Rientjes { 3755238d3c13SDavid Rientjes struct hstate *h; 3756238d3c13SDavid Rientjes unsigned long count; 3757238d3c13SDavid Rientjes int nid; 3758238d3c13SDavid Rientjes int err; 3759238d3c13SDavid Rientjes 3760238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 3761238d3c13SDavid Rientjes if (err) 3762238d3c13SDavid Rientjes return err; 3763238d3c13SDavid Rientjes 3764238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 3765238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 3766238d3c13SDavid Rientjes } 3767238d3c13SDavid Rientjes 376806808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 376906808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 377006808b08SLee Schermerhorn { 377106808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 377206808b08SLee Schermerhorn } 377306808b08SLee Schermerhorn 377406808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 377506808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 377606808b08SLee Schermerhorn { 3777238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 3778a3437870SNishanth Aravamudan } 3779a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 3780a3437870SNishanth Aravamudan 378106808b08SLee Schermerhorn #ifdef CONFIG_NUMA 378206808b08SLee Schermerhorn 378306808b08SLee Schermerhorn /* 378406808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 378506808b08SLee Schermerhorn * huge page alloc/free. 378606808b08SLee Schermerhorn */ 378706808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 3788ae7a927dSJoe Perches struct kobj_attribute *attr, 3789ae7a927dSJoe Perches char *buf) 379006808b08SLee Schermerhorn { 379106808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 379206808b08SLee Schermerhorn } 379306808b08SLee Schermerhorn 379406808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 379506808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 379606808b08SLee Schermerhorn { 3797238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 379806808b08SLee Schermerhorn } 379906808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 380006808b08SLee Schermerhorn #endif 380106808b08SLee Schermerhorn 380206808b08SLee Schermerhorn 3803a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 3804a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3805a3437870SNishanth Aravamudan { 38069a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3807ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); 3808a3437870SNishanth Aravamudan } 3809adbe8726SEric B Munson 3810a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 3811a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 3812a3437870SNishanth Aravamudan { 3813a3437870SNishanth Aravamudan int err; 3814a3437870SNishanth Aravamudan unsigned long input; 38159a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3816a3437870SNishanth Aravamudan 3817bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3818adbe8726SEric B Munson return -EINVAL; 3819adbe8726SEric B Munson 38203dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 3821a3437870SNishanth Aravamudan if (err) 382273ae31e5SEric B Munson return err; 3823a3437870SNishanth Aravamudan 3824db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3825a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 3826db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3827a3437870SNishanth Aravamudan 3828a3437870SNishanth Aravamudan return count; 3829a3437870SNishanth Aravamudan } 3830a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 3831a3437870SNishanth Aravamudan 3832a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 3833a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3834a3437870SNishanth Aravamudan { 38359a305230SLee Schermerhorn struct hstate *h; 38369a305230SLee Schermerhorn unsigned long free_huge_pages; 38379a305230SLee Schermerhorn int nid; 38389a305230SLee Schermerhorn 38399a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 38409a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 38419a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 38429a305230SLee Schermerhorn else 38439a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 38449a305230SLee Schermerhorn 3845ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", free_huge_pages); 3846a3437870SNishanth Aravamudan } 3847a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 3848a3437870SNishanth Aravamudan 3849a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 3850a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3851a3437870SNishanth Aravamudan { 38529a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3853ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); 3854a3437870SNishanth Aravamudan } 3855a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 3856a3437870SNishanth Aravamudan 3857a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 3858a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3859a3437870SNishanth Aravamudan { 38609a305230SLee Schermerhorn struct hstate *h; 38619a305230SLee Schermerhorn unsigned long surplus_huge_pages; 38629a305230SLee Schermerhorn int nid; 38639a305230SLee Schermerhorn 38649a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 38659a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 38669a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 38679a305230SLee Schermerhorn else 38689a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 38699a305230SLee Schermerhorn 3870ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", surplus_huge_pages); 3871a3437870SNishanth Aravamudan } 3872a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 3873a3437870SNishanth Aravamudan 387479dfc695SMike Kravetz static ssize_t demote_store(struct kobject *kobj, 387579dfc695SMike Kravetz struct kobj_attribute *attr, const char *buf, size_t len) 387679dfc695SMike Kravetz { 387779dfc695SMike Kravetz unsigned long nr_demote; 387879dfc695SMike Kravetz unsigned long nr_available; 387979dfc695SMike Kravetz nodemask_t nodes_allowed, *n_mask; 388079dfc695SMike Kravetz struct hstate *h; 38818eeda55fSLi zeming int err; 388279dfc695SMike Kravetz int nid; 388379dfc695SMike Kravetz 388479dfc695SMike Kravetz err = kstrtoul(buf, 10, &nr_demote); 388579dfc695SMike Kravetz if (err) 388679dfc695SMike Kravetz return err; 388779dfc695SMike Kravetz h = kobj_to_hstate(kobj, &nid); 388879dfc695SMike Kravetz 388979dfc695SMike Kravetz if (nid != NUMA_NO_NODE) { 389079dfc695SMike Kravetz init_nodemask_of_node(&nodes_allowed, nid); 389179dfc695SMike Kravetz n_mask = &nodes_allowed; 389279dfc695SMike Kravetz } else { 389379dfc695SMike Kravetz n_mask = &node_states[N_MEMORY]; 389479dfc695SMike Kravetz } 389579dfc695SMike Kravetz 389679dfc695SMike Kravetz /* Synchronize with other sysfs operations modifying huge pages */ 389779dfc695SMike Kravetz mutex_lock(&h->resize_lock); 389879dfc695SMike Kravetz spin_lock_irq(&hugetlb_lock); 389979dfc695SMike Kravetz 390079dfc695SMike Kravetz while (nr_demote) { 390179dfc695SMike Kravetz /* 390279dfc695SMike Kravetz * Check for available pages to demote each time thorough the 390379dfc695SMike Kravetz * loop as demote_pool_huge_page will drop hugetlb_lock. 390479dfc695SMike Kravetz */ 390579dfc695SMike Kravetz if (nid != NUMA_NO_NODE) 390679dfc695SMike Kravetz nr_available = h->free_huge_pages_node[nid]; 390779dfc695SMike Kravetz else 390879dfc695SMike Kravetz nr_available = h->free_huge_pages; 390979dfc695SMike Kravetz nr_available -= h->resv_huge_pages; 391079dfc695SMike Kravetz if (!nr_available) 391179dfc695SMike Kravetz break; 391279dfc695SMike Kravetz 391379dfc695SMike Kravetz err = demote_pool_huge_page(h, n_mask); 391479dfc695SMike Kravetz if (err) 391579dfc695SMike Kravetz break; 391679dfc695SMike Kravetz 391779dfc695SMike Kravetz nr_demote--; 391879dfc695SMike Kravetz } 391979dfc695SMike Kravetz 392079dfc695SMike Kravetz spin_unlock_irq(&hugetlb_lock); 392179dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 392279dfc695SMike Kravetz 392379dfc695SMike Kravetz if (err) 392479dfc695SMike Kravetz return err; 392579dfc695SMike Kravetz return len; 392679dfc695SMike Kravetz } 392779dfc695SMike Kravetz HSTATE_ATTR_WO(demote); 392879dfc695SMike Kravetz 392979dfc695SMike Kravetz static ssize_t demote_size_show(struct kobject *kobj, 393079dfc695SMike Kravetz struct kobj_attribute *attr, char *buf) 393179dfc695SMike Kravetz { 393212658abfSMiaohe Lin struct hstate *h = kobj_to_hstate(kobj, NULL); 393379dfc695SMike Kravetz unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K; 393479dfc695SMike Kravetz 393579dfc695SMike Kravetz return sysfs_emit(buf, "%lukB\n", demote_size); 393679dfc695SMike Kravetz } 393779dfc695SMike Kravetz 393879dfc695SMike Kravetz static ssize_t demote_size_store(struct kobject *kobj, 393979dfc695SMike Kravetz struct kobj_attribute *attr, 394079dfc695SMike Kravetz const char *buf, size_t count) 394179dfc695SMike Kravetz { 394279dfc695SMike Kravetz struct hstate *h, *demote_hstate; 394379dfc695SMike Kravetz unsigned long demote_size; 394479dfc695SMike Kravetz unsigned int demote_order; 394579dfc695SMike Kravetz 394679dfc695SMike Kravetz demote_size = (unsigned long)memparse(buf, NULL); 394779dfc695SMike Kravetz 394879dfc695SMike Kravetz demote_hstate = size_to_hstate(demote_size); 394979dfc695SMike Kravetz if (!demote_hstate) 395079dfc695SMike Kravetz return -EINVAL; 395179dfc695SMike Kravetz demote_order = demote_hstate->order; 3952a01f4390SMike Kravetz if (demote_order < HUGETLB_PAGE_ORDER) 3953a01f4390SMike Kravetz return -EINVAL; 395479dfc695SMike Kravetz 395579dfc695SMike Kravetz /* demote order must be smaller than hstate order */ 395612658abfSMiaohe Lin h = kobj_to_hstate(kobj, NULL); 395779dfc695SMike Kravetz if (demote_order >= h->order) 395879dfc695SMike Kravetz return -EINVAL; 395979dfc695SMike Kravetz 396079dfc695SMike Kravetz /* resize_lock synchronizes access to demote size and writes */ 396179dfc695SMike Kravetz mutex_lock(&h->resize_lock); 396279dfc695SMike Kravetz h->demote_order = demote_order; 396379dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 396479dfc695SMike Kravetz 396579dfc695SMike Kravetz return count; 396679dfc695SMike Kravetz } 396779dfc695SMike Kravetz HSTATE_ATTR(demote_size); 396879dfc695SMike Kravetz 3969a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 3970a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 3971a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 3972a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 3973a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 3974a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 397506808b08SLee Schermerhorn #ifdef CONFIG_NUMA 397606808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 397706808b08SLee Schermerhorn #endif 3978a3437870SNishanth Aravamudan NULL, 3979a3437870SNishanth Aravamudan }; 3980a3437870SNishanth Aravamudan 398167e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 3982a3437870SNishanth Aravamudan .attrs = hstate_attrs, 3983a3437870SNishanth Aravamudan }; 3984a3437870SNishanth Aravamudan 398579dfc695SMike Kravetz static struct attribute *hstate_demote_attrs[] = { 398679dfc695SMike Kravetz &demote_size_attr.attr, 398779dfc695SMike Kravetz &demote_attr.attr, 398879dfc695SMike Kravetz NULL, 398979dfc695SMike Kravetz }; 399079dfc695SMike Kravetz 399179dfc695SMike Kravetz static const struct attribute_group hstate_demote_attr_group = { 399279dfc695SMike Kravetz .attrs = hstate_demote_attrs, 399379dfc695SMike Kravetz }; 399479dfc695SMike Kravetz 3995094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 39969a305230SLee Schermerhorn struct kobject **hstate_kobjs, 399767e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 3998a3437870SNishanth Aravamudan { 3999a3437870SNishanth Aravamudan int retval; 4000972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 4001a3437870SNishanth Aravamudan 40029a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 40039a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 4004a3437870SNishanth Aravamudan return -ENOMEM; 4005a3437870SNishanth Aravamudan 40069a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 4007cc2205a6SMiaohe Lin if (retval) { 40089a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 4009cc2205a6SMiaohe Lin hstate_kobjs[hi] = NULL; 40103a6bdda0SMiaohe Lin return retval; 4011cc2205a6SMiaohe Lin } 4012a3437870SNishanth Aravamudan 401379dfc695SMike Kravetz if (h->demote_order) { 401401088a60SMiaohe Lin retval = sysfs_create_group(hstate_kobjs[hi], 401501088a60SMiaohe Lin &hstate_demote_attr_group); 401601088a60SMiaohe Lin if (retval) { 401779dfc695SMike Kravetz pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name); 401801088a60SMiaohe Lin sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group); 401901088a60SMiaohe Lin kobject_put(hstate_kobjs[hi]); 402001088a60SMiaohe Lin hstate_kobjs[hi] = NULL; 402101088a60SMiaohe Lin return retval; 402201088a60SMiaohe Lin } 402379dfc695SMike Kravetz } 402479dfc695SMike Kravetz 402501088a60SMiaohe Lin return 0; 4026a3437870SNishanth Aravamudan } 4027a3437870SNishanth Aravamudan 40289a305230SLee Schermerhorn #ifdef CONFIG_NUMA 4029a4a00b45SMuchun Song static bool hugetlb_sysfs_initialized __ro_after_init; 40309a305230SLee Schermerhorn 40319a305230SLee Schermerhorn /* 40329a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 403310fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 403410fbcf4cSKay Sievers * index of a node device or _hstate == node id. 403510fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 40369a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 40379a305230SLee Schermerhorn */ 40389a305230SLee Schermerhorn struct node_hstate { 40399a305230SLee Schermerhorn struct kobject *hugepages_kobj; 40409a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 40419a305230SLee Schermerhorn }; 4042b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 40439a305230SLee Schermerhorn 40449a305230SLee Schermerhorn /* 404510fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 40469a305230SLee Schermerhorn */ 40479a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 40489a305230SLee Schermerhorn &nr_hugepages_attr.attr, 40499a305230SLee Schermerhorn &free_hugepages_attr.attr, 40509a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 40519a305230SLee Schermerhorn NULL, 40529a305230SLee Schermerhorn }; 40539a305230SLee Schermerhorn 405467e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 40559a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 40569a305230SLee Schermerhorn }; 40579a305230SLee Schermerhorn 40589a305230SLee Schermerhorn /* 405910fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 40609a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 40619a305230SLee Schermerhorn */ 40629a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 40639a305230SLee Schermerhorn { 40649a305230SLee Schermerhorn int nid; 40659a305230SLee Schermerhorn 40669a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 40679a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 40689a305230SLee Schermerhorn int i; 40699a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 40709a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 40719a305230SLee Schermerhorn if (nidp) 40729a305230SLee Schermerhorn *nidp = nid; 40739a305230SLee Schermerhorn return &hstates[i]; 40749a305230SLee Schermerhorn } 40759a305230SLee Schermerhorn } 40769a305230SLee Schermerhorn 40779a305230SLee Schermerhorn BUG(); 40789a305230SLee Schermerhorn return NULL; 40799a305230SLee Schermerhorn } 40809a305230SLee Schermerhorn 40819a305230SLee Schermerhorn /* 408210fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 40839a305230SLee Schermerhorn * No-op if no hstate attributes attached. 40849a305230SLee Schermerhorn */ 4085a4a00b45SMuchun Song void hugetlb_unregister_node(struct node *node) 40869a305230SLee Schermerhorn { 40879a305230SLee Schermerhorn struct hstate *h; 408810fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 40899a305230SLee Schermerhorn 40909a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 40919b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 40929a305230SLee Schermerhorn 4093972dc4deSAneesh Kumar K.V for_each_hstate(h) { 4094972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 409501088a60SMiaohe Lin struct kobject *hstate_kobj = nhs->hstate_kobjs[idx]; 409601088a60SMiaohe Lin 409701088a60SMiaohe Lin if (!hstate_kobj) 409801088a60SMiaohe Lin continue; 409901088a60SMiaohe Lin if (h->demote_order) 410001088a60SMiaohe Lin sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group); 410101088a60SMiaohe Lin sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group); 410201088a60SMiaohe Lin kobject_put(hstate_kobj); 4103972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 4104972dc4deSAneesh Kumar K.V } 41059a305230SLee Schermerhorn 41069a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 41079a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 41089a305230SLee Schermerhorn } 41099a305230SLee Schermerhorn 41109a305230SLee Schermerhorn 41119a305230SLee Schermerhorn /* 411210fbcf4cSKay Sievers * Register hstate attributes for a single node device. 41139a305230SLee Schermerhorn * No-op if attributes already registered. 41149a305230SLee Schermerhorn */ 4115a4a00b45SMuchun Song void hugetlb_register_node(struct node *node) 41169a305230SLee Schermerhorn { 41179a305230SLee Schermerhorn struct hstate *h; 411810fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 41199a305230SLee Schermerhorn int err; 41209a305230SLee Schermerhorn 4121a4a00b45SMuchun Song if (!hugetlb_sysfs_initialized) 4122a4a00b45SMuchun Song return; 4123a4a00b45SMuchun Song 41249a305230SLee Schermerhorn if (nhs->hugepages_kobj) 41259a305230SLee Schermerhorn return; /* already allocated */ 41269a305230SLee Schermerhorn 41279a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 412810fbcf4cSKay Sievers &node->dev.kobj); 41299a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 41309a305230SLee Schermerhorn return; 41319a305230SLee Schermerhorn 41329a305230SLee Schermerhorn for_each_hstate(h) { 41339a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 41349a305230SLee Schermerhorn nhs->hstate_kobjs, 41359a305230SLee Schermerhorn &per_node_hstate_attr_group); 41369a305230SLee Schermerhorn if (err) { 4137282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s for node %d\n", 413810fbcf4cSKay Sievers h->name, node->dev.id); 41399a305230SLee Schermerhorn hugetlb_unregister_node(node); 41409a305230SLee Schermerhorn break; 41419a305230SLee Schermerhorn } 41429a305230SLee Schermerhorn } 41439a305230SLee Schermerhorn } 41449a305230SLee Schermerhorn 41459a305230SLee Schermerhorn /* 41469b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 414710fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 414810fbcf4cSKay Sievers * registered their associated device by this time. 41499a305230SLee Schermerhorn */ 41507d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 41519a305230SLee Schermerhorn { 41529a305230SLee Schermerhorn int nid; 41539a305230SLee Schermerhorn 4154a4a00b45SMuchun Song for_each_online_node(nid) 4155b958d4d0SMuchun Song hugetlb_register_node(node_devices[nid]); 41569a305230SLee Schermerhorn } 41579a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 41589a305230SLee Schermerhorn 41599a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 41609a305230SLee Schermerhorn { 41619a305230SLee Schermerhorn BUG(); 41629a305230SLee Schermerhorn if (nidp) 41639a305230SLee Schermerhorn *nidp = -1; 41649a305230SLee Schermerhorn return NULL; 41659a305230SLee Schermerhorn } 41669a305230SLee Schermerhorn 41679a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 41689a305230SLee Schermerhorn 41699a305230SLee Schermerhorn #endif 41709a305230SLee Schermerhorn 4171263b8998SMiaohe Lin #ifdef CONFIG_CMA 4172263b8998SMiaohe Lin static void __init hugetlb_cma_check(void); 4173263b8998SMiaohe Lin #else 4174263b8998SMiaohe Lin static inline __init void hugetlb_cma_check(void) 4175263b8998SMiaohe Lin { 4176263b8998SMiaohe Lin } 4177263b8998SMiaohe Lin #endif 4178263b8998SMiaohe Lin 4179a4a00b45SMuchun Song static void __init hugetlb_sysfs_init(void) 4180a4a00b45SMuchun Song { 4181a4a00b45SMuchun Song struct hstate *h; 4182a4a00b45SMuchun Song int err; 4183a4a00b45SMuchun Song 4184a4a00b45SMuchun Song hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 4185a4a00b45SMuchun Song if (!hugepages_kobj) 4186a4a00b45SMuchun Song return; 4187a4a00b45SMuchun Song 4188a4a00b45SMuchun Song for_each_hstate(h) { 4189a4a00b45SMuchun Song err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 4190a4a00b45SMuchun Song hstate_kobjs, &hstate_attr_group); 4191a4a00b45SMuchun Song if (err) 4192a4a00b45SMuchun Song pr_err("HugeTLB: Unable to add hstate %s", h->name); 4193a4a00b45SMuchun Song } 4194a4a00b45SMuchun Song 4195a4a00b45SMuchun Song #ifdef CONFIG_NUMA 4196a4a00b45SMuchun Song hugetlb_sysfs_initialized = true; 4197a4a00b45SMuchun Song #endif 4198a4a00b45SMuchun Song hugetlb_register_all_nodes(); 4199a4a00b45SMuchun Song } 4200a4a00b45SMuchun Song 4201a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 4202a3437870SNishanth Aravamudan { 42038382d914SDavidlohr Bueso int i; 42048382d914SDavidlohr Bueso 4205d6995da3SMike Kravetz BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < 4206d6995da3SMike Kravetz __NR_HPAGEFLAGS); 4207d6995da3SMike Kravetz 4208c2833a5bSMike Kravetz if (!hugepages_supported()) { 4209c2833a5bSMike Kravetz if (hugetlb_max_hstate || default_hstate_max_huge_pages) 4210c2833a5bSMike Kravetz pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); 42110ef89d25SBenjamin Herrenschmidt return 0; 4212d715cf80SLiam R. Howlett } 4213d715cf80SLiam R. Howlett 4214282f4214SMike Kravetz /* 4215282f4214SMike Kravetz * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some 4216282f4214SMike Kravetz * architectures depend on setup being done here. 4217282f4214SMike Kravetz */ 4218a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 4219282f4214SMike Kravetz if (!parsed_default_hugepagesz) { 4220282f4214SMike Kravetz /* 4221282f4214SMike Kravetz * If we did not parse a default huge page size, set 4222282f4214SMike Kravetz * default_hstate_idx to HPAGE_SIZE hstate. And, if the 4223282f4214SMike Kravetz * number of huge pages for this default size was implicitly 4224282f4214SMike Kravetz * specified, set that here as well. 4225282f4214SMike Kravetz * Note that the implicit setting will overwrite an explicit 4226282f4214SMike Kravetz * setting. A warning will be printed in this case. 4227282f4214SMike Kravetz */ 4228282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); 4229f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 4230282f4214SMike Kravetz if (default_hstate.max_huge_pages) { 4231282f4214SMike Kravetz char buf[32]; 4232282f4214SMike Kravetz 4233282f4214SMike Kravetz string_get_size(huge_page_size(&default_hstate), 4234282f4214SMike Kravetz 1, STRING_UNITS_2, buf, 32); 4235282f4214SMike Kravetz pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", 4236282f4214SMike Kravetz default_hstate.max_huge_pages, buf); 4237282f4214SMike Kravetz pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", 4238282f4214SMike Kravetz default_hstate_max_huge_pages); 4239282f4214SMike Kravetz } 4240282f4214SMike Kravetz default_hstate.max_huge_pages = 4241282f4214SMike Kravetz default_hstate_max_huge_pages; 4242b5389086SZhenguo Yao 42430a7a0f6fSPeng Liu for_each_online_node(i) 4244b5389086SZhenguo Yao default_hstate.max_huge_pages_node[i] = 4245b5389086SZhenguo Yao default_hugepages_in_node[i]; 4246282f4214SMike Kravetz } 4247f8b74815SVaishali Thakkar } 4248a3437870SNishanth Aravamudan 4249cf11e85fSRoman Gushchin hugetlb_cma_check(); 4250a3437870SNishanth Aravamudan hugetlb_init_hstates(); 4251aa888a74SAndi Kleen gather_bootmem_prealloc(); 4252a3437870SNishanth Aravamudan report_hugepages(); 4253a3437870SNishanth Aravamudan 4254a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 42557179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 42569a305230SLee Schermerhorn 42578382d914SDavidlohr Bueso #ifdef CONFIG_SMP 42588382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 42598382d914SDavidlohr Bueso #else 42608382d914SDavidlohr Bueso num_fault_mutexes = 1; 42618382d914SDavidlohr Bueso #endif 4262c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 42636da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 42646da2ec56SKees Cook GFP_KERNEL); 4265c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 42668382d914SDavidlohr Bueso 42678382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 4268c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 4269a3437870SNishanth Aravamudan return 0; 4270a3437870SNishanth Aravamudan } 42713e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 4272a3437870SNishanth Aravamudan 4273ae94da89SMike Kravetz /* Overwritten by architectures with more huge page sizes */ 4274ae94da89SMike Kravetz bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) 42759fee021dSVaishali Thakkar { 4276ae94da89SMike Kravetz return size == HPAGE_SIZE; 42779fee021dSVaishali Thakkar } 42789fee021dSVaishali Thakkar 4279d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 4280a3437870SNishanth Aravamudan { 4281a3437870SNishanth Aravamudan struct hstate *h; 42828faa8b07SAndi Kleen unsigned long i; 42838faa8b07SAndi Kleen 4284a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 4285a3437870SNishanth Aravamudan return; 4286a3437870SNishanth Aravamudan } 428747d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 4288a3437870SNishanth Aravamudan BUG_ON(order == 0); 428947d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 429029383967SMike Kravetz mutex_init(&h->resize_lock); 4291a3437870SNishanth Aravamudan h->order = order; 4292aca78307SMiaohe Lin h->mask = ~(huge_page_size(h) - 1); 42938faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 42948faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 42950edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 429654f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 429754f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 4298a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 4299c2c3a60aSMiaohe Lin huge_page_size(h)/SZ_1K); 43008faa8b07SAndi Kleen 4301a3437870SNishanth Aravamudan parsed_hstate = h; 4302a3437870SNishanth Aravamudan } 4303a3437870SNishanth Aravamudan 4304b5389086SZhenguo Yao bool __init __weak hugetlb_node_alloc_supported(void) 4305b5389086SZhenguo Yao { 4306b5389086SZhenguo Yao return true; 4307b5389086SZhenguo Yao } 4308f87442f4SPeng Liu 4309f87442f4SPeng Liu static void __init hugepages_clear_pages_in_node(void) 4310f87442f4SPeng Liu { 4311f87442f4SPeng Liu if (!hugetlb_max_hstate) { 4312f87442f4SPeng Liu default_hstate_max_huge_pages = 0; 4313f87442f4SPeng Liu memset(default_hugepages_in_node, 0, 431410395680SMiaohe Lin sizeof(default_hugepages_in_node)); 4315f87442f4SPeng Liu } else { 4316f87442f4SPeng Liu parsed_hstate->max_huge_pages = 0; 4317f87442f4SPeng Liu memset(parsed_hstate->max_huge_pages_node, 0, 431810395680SMiaohe Lin sizeof(parsed_hstate->max_huge_pages_node)); 4319f87442f4SPeng Liu } 4320f87442f4SPeng Liu } 4321f87442f4SPeng Liu 4322282f4214SMike Kravetz /* 4323282f4214SMike Kravetz * hugepages command line processing 4324282f4214SMike Kravetz * hugepages normally follows a valid hugepagsz or default_hugepagsz 4325282f4214SMike Kravetz * specification. If not, ignore the hugepages value. hugepages can also 4326282f4214SMike Kravetz * be the first huge page command line option in which case it implicitly 4327282f4214SMike Kravetz * specifies the number of huge pages for the default size. 4328282f4214SMike Kravetz */ 4329282f4214SMike Kravetz static int __init hugepages_setup(char *s) 4330a3437870SNishanth Aravamudan { 4331a3437870SNishanth Aravamudan unsigned long *mhp; 43328faa8b07SAndi Kleen static unsigned long *last_mhp; 4333b5389086SZhenguo Yao int node = NUMA_NO_NODE; 4334b5389086SZhenguo Yao int count; 4335b5389086SZhenguo Yao unsigned long tmp; 4336b5389086SZhenguo Yao char *p = s; 4337a3437870SNishanth Aravamudan 43389fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 4339282f4214SMike Kravetz pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); 43409fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 4341f81f6e4bSPeng Liu return 1; 43429fee021dSVaishali Thakkar } 4343282f4214SMike Kravetz 4344a3437870SNishanth Aravamudan /* 4345282f4214SMike Kravetz * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter 4346282f4214SMike Kravetz * yet, so this hugepages= parameter goes to the "default hstate". 4347282f4214SMike Kravetz * Otherwise, it goes with the previously parsed hugepagesz or 4348282f4214SMike Kravetz * default_hugepagesz. 4349a3437870SNishanth Aravamudan */ 43509fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 4351a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 4352a3437870SNishanth Aravamudan else 4353a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 4354a3437870SNishanth Aravamudan 43558faa8b07SAndi Kleen if (mhp == last_mhp) { 4356282f4214SMike Kravetz pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); 4357f81f6e4bSPeng Liu return 1; 43588faa8b07SAndi Kleen } 43598faa8b07SAndi Kleen 4360b5389086SZhenguo Yao while (*p) { 4361b5389086SZhenguo Yao count = 0; 4362b5389086SZhenguo Yao if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4363b5389086SZhenguo Yao goto invalid; 4364b5389086SZhenguo Yao /* Parameter is node format */ 4365b5389086SZhenguo Yao if (p[count] == ':') { 4366b5389086SZhenguo Yao if (!hugetlb_node_alloc_supported()) { 4367b5389086SZhenguo Yao pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n"); 4368f81f6e4bSPeng Liu return 1; 4369b5389086SZhenguo Yao } 43700a7a0f6fSPeng Liu if (tmp >= MAX_NUMNODES || !node_online(tmp)) 4371e79ce983SLiu Yuntao goto invalid; 43720a7a0f6fSPeng Liu node = array_index_nospec(tmp, MAX_NUMNODES); 4373b5389086SZhenguo Yao p += count + 1; 4374b5389086SZhenguo Yao /* Parse hugepages */ 4375b5389086SZhenguo Yao if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4376b5389086SZhenguo Yao goto invalid; 4377b5389086SZhenguo Yao if (!hugetlb_max_hstate) 4378b5389086SZhenguo Yao default_hugepages_in_node[node] = tmp; 4379b5389086SZhenguo Yao else 4380b5389086SZhenguo Yao parsed_hstate->max_huge_pages_node[node] = tmp; 4381b5389086SZhenguo Yao *mhp += tmp; 4382b5389086SZhenguo Yao /* Go to parse next node*/ 4383b5389086SZhenguo Yao if (p[count] == ',') 4384b5389086SZhenguo Yao p += count + 1; 4385b5389086SZhenguo Yao else 4386b5389086SZhenguo Yao break; 4387b5389086SZhenguo Yao } else { 4388b5389086SZhenguo Yao if (p != s) 4389b5389086SZhenguo Yao goto invalid; 4390b5389086SZhenguo Yao *mhp = tmp; 4391b5389086SZhenguo Yao break; 4392b5389086SZhenguo Yao } 4393b5389086SZhenguo Yao } 4394a3437870SNishanth Aravamudan 43958faa8b07SAndi Kleen /* 43968faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 439704adbc3fSMiaohe Lin * But we need to allocate gigantic hstates here early to still 43988faa8b07SAndi Kleen * use the bootmem allocator. 43998faa8b07SAndi Kleen */ 440004adbc3fSMiaohe Lin if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) 44018faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 44028faa8b07SAndi Kleen 44038faa8b07SAndi Kleen last_mhp = mhp; 44048faa8b07SAndi Kleen 4405a3437870SNishanth Aravamudan return 1; 4406b5389086SZhenguo Yao 4407b5389086SZhenguo Yao invalid: 4408b5389086SZhenguo Yao pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p); 4409f87442f4SPeng Liu hugepages_clear_pages_in_node(); 4410f81f6e4bSPeng Liu return 1; 4411a3437870SNishanth Aravamudan } 4412282f4214SMike Kravetz __setup("hugepages=", hugepages_setup); 4413e11bfbfcSNick Piggin 4414282f4214SMike Kravetz /* 4415282f4214SMike Kravetz * hugepagesz command line processing 4416282f4214SMike Kravetz * A specific huge page size can only be specified once with hugepagesz. 4417282f4214SMike Kravetz * hugepagesz is followed by hugepages on the command line. The global 4418282f4214SMike Kravetz * variable 'parsed_valid_hugepagesz' is used to determine if prior 4419282f4214SMike Kravetz * hugepagesz argument was valid. 4420282f4214SMike Kravetz */ 4421359f2544SMike Kravetz static int __init hugepagesz_setup(char *s) 4422e11bfbfcSNick Piggin { 4423359f2544SMike Kravetz unsigned long size; 4424282f4214SMike Kravetz struct hstate *h; 4425282f4214SMike Kravetz 4426282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4427359f2544SMike Kravetz size = (unsigned long)memparse(s, NULL); 4428359f2544SMike Kravetz 4429359f2544SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4430282f4214SMike Kravetz pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); 4431f81f6e4bSPeng Liu return 1; 4432359f2544SMike Kravetz } 4433359f2544SMike Kravetz 4434282f4214SMike Kravetz h = size_to_hstate(size); 4435282f4214SMike Kravetz if (h) { 4436282f4214SMike Kravetz /* 4437282f4214SMike Kravetz * hstate for this size already exists. This is normally 4438282f4214SMike Kravetz * an error, but is allowed if the existing hstate is the 4439282f4214SMike Kravetz * default hstate. More specifically, it is only allowed if 4440282f4214SMike Kravetz * the number of huge pages for the default hstate was not 4441282f4214SMike Kravetz * previously specified. 4442282f4214SMike Kravetz */ 4443282f4214SMike Kravetz if (!parsed_default_hugepagesz || h != &default_hstate || 4444282f4214SMike Kravetz default_hstate.max_huge_pages) { 4445282f4214SMike Kravetz pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); 4446f81f6e4bSPeng Liu return 1; 444738237830SMike Kravetz } 444838237830SMike Kravetz 4449282f4214SMike Kravetz /* 4450282f4214SMike Kravetz * No need to call hugetlb_add_hstate() as hstate already 4451282f4214SMike Kravetz * exists. But, do set parsed_hstate so that a following 4452282f4214SMike Kravetz * hugepages= parameter will be applied to this hstate. 4453282f4214SMike Kravetz */ 4454282f4214SMike Kravetz parsed_hstate = h; 4455282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4456e11bfbfcSNick Piggin return 1; 4457e11bfbfcSNick Piggin } 4458282f4214SMike Kravetz 4459359f2544SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4460282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4461359f2544SMike Kravetz return 1; 4462359f2544SMike Kravetz } 4463359f2544SMike Kravetz __setup("hugepagesz=", hugepagesz_setup); 4464359f2544SMike Kravetz 4465282f4214SMike Kravetz /* 4466282f4214SMike Kravetz * default_hugepagesz command line input 4467282f4214SMike Kravetz * Only one instance of default_hugepagesz allowed on command line. 4468282f4214SMike Kravetz */ 4469ae94da89SMike Kravetz static int __init default_hugepagesz_setup(char *s) 4470e11bfbfcSNick Piggin { 4471ae94da89SMike Kravetz unsigned long size; 4472b5389086SZhenguo Yao int i; 4473ae94da89SMike Kravetz 4474282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4475282f4214SMike Kravetz if (parsed_default_hugepagesz) { 4476282f4214SMike Kravetz pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); 4477f81f6e4bSPeng Liu return 1; 4478282f4214SMike Kravetz } 4479282f4214SMike Kravetz 4480282f4214SMike Kravetz size = (unsigned long)memparse(s, NULL); 4481282f4214SMike Kravetz 4482282f4214SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4483282f4214SMike Kravetz pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); 4484f81f6e4bSPeng Liu return 1; 4485282f4214SMike Kravetz } 4486282f4214SMike Kravetz 4487282f4214SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4488282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4489282f4214SMike Kravetz parsed_default_hugepagesz = true; 4490282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(size)); 4491282f4214SMike Kravetz 4492282f4214SMike Kravetz /* 4493282f4214SMike Kravetz * The number of default huge pages (for this size) could have been 4494282f4214SMike Kravetz * specified as the first hugetlb parameter: hugepages=X. If so, 4495282f4214SMike Kravetz * then default_hstate_max_huge_pages is set. If the default huge 4496282f4214SMike Kravetz * page size is gigantic (>= MAX_ORDER), then the pages must be 4497282f4214SMike Kravetz * allocated here from bootmem allocator. 4498282f4214SMike Kravetz */ 4499282f4214SMike Kravetz if (default_hstate_max_huge_pages) { 4500282f4214SMike Kravetz default_hstate.max_huge_pages = default_hstate_max_huge_pages; 45010a7a0f6fSPeng Liu for_each_online_node(i) 4502b5389086SZhenguo Yao default_hstate.max_huge_pages_node[i] = 4503b5389086SZhenguo Yao default_hugepages_in_node[i]; 4504282f4214SMike Kravetz if (hstate_is_gigantic(&default_hstate)) 4505282f4214SMike Kravetz hugetlb_hstate_alloc_pages(&default_hstate); 4506282f4214SMike Kravetz default_hstate_max_huge_pages = 0; 4507282f4214SMike Kravetz } 4508282f4214SMike Kravetz 4509e11bfbfcSNick Piggin return 1; 4510e11bfbfcSNick Piggin } 4511ae94da89SMike Kravetz __setup("default_hugepagesz=", default_hugepagesz_setup); 4512a3437870SNishanth Aravamudan 4513d2226ebdSFeng Tang static nodemask_t *policy_mbind_nodemask(gfp_t gfp) 4514d2226ebdSFeng Tang { 4515d2226ebdSFeng Tang #ifdef CONFIG_NUMA 4516d2226ebdSFeng Tang struct mempolicy *mpol = get_task_policy(current); 4517d2226ebdSFeng Tang 4518d2226ebdSFeng Tang /* 4519d2226ebdSFeng Tang * Only enforce MPOL_BIND policy which overlaps with cpuset policy 4520d2226ebdSFeng Tang * (from policy_nodemask) specifically for hugetlb case 4521d2226ebdSFeng Tang */ 4522d2226ebdSFeng Tang if (mpol->mode == MPOL_BIND && 4523d2226ebdSFeng Tang (apply_policy_zone(mpol, gfp_zone(gfp)) && 4524d2226ebdSFeng Tang cpuset_nodemask_valid_mems_allowed(&mpol->nodes))) 4525d2226ebdSFeng Tang return &mpol->nodes; 4526d2226ebdSFeng Tang #endif 4527d2226ebdSFeng Tang return NULL; 4528d2226ebdSFeng Tang } 4529d2226ebdSFeng Tang 45308ca39e68SMuchun Song static unsigned int allowed_mems_nr(struct hstate *h) 45318a213460SNishanth Aravamudan { 45328a213460SNishanth Aravamudan int node; 45338a213460SNishanth Aravamudan unsigned int nr = 0; 4534d2226ebdSFeng Tang nodemask_t *mbind_nodemask; 45358ca39e68SMuchun Song unsigned int *array = h->free_huge_pages_node; 45368ca39e68SMuchun Song gfp_t gfp_mask = htlb_alloc_mask(h); 45378a213460SNishanth Aravamudan 4538d2226ebdSFeng Tang mbind_nodemask = policy_mbind_nodemask(gfp_mask); 45398ca39e68SMuchun Song for_each_node_mask(node, cpuset_current_mems_allowed) { 4540d2226ebdSFeng Tang if (!mbind_nodemask || node_isset(node, *mbind_nodemask)) 45418a213460SNishanth Aravamudan nr += array[node]; 45428ca39e68SMuchun Song } 45438a213460SNishanth Aravamudan 45448a213460SNishanth Aravamudan return nr; 45458a213460SNishanth Aravamudan } 45468a213460SNishanth Aravamudan 45478a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 454817743798SMuchun Song static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, 454917743798SMuchun Song void *buffer, size_t *length, 455017743798SMuchun Song loff_t *ppos, unsigned long *out) 455117743798SMuchun Song { 455217743798SMuchun Song struct ctl_table dup_table; 455317743798SMuchun Song 455417743798SMuchun Song /* 455517743798SMuchun Song * In order to avoid races with __do_proc_doulongvec_minmax(), we 455617743798SMuchun Song * can duplicate the @table and alter the duplicate of it. 455717743798SMuchun Song */ 455817743798SMuchun Song dup_table = *table; 455917743798SMuchun Song dup_table.data = out; 456017743798SMuchun Song 456117743798SMuchun Song return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); 456217743798SMuchun Song } 456317743798SMuchun Song 456406808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 456506808b08SLee Schermerhorn struct ctl_table *table, int write, 456632927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 45671da177e4SLinus Torvalds { 4568e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 4569238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 457008d4a246SMichal Hocko int ret; 4571e5ff2159SAndi Kleen 4572457c1b27SNishanth Aravamudan if (!hugepages_supported()) 457386613628SJan Stancek return -EOPNOTSUPP; 4574457c1b27SNishanth Aravamudan 457517743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 457617743798SMuchun Song &tmp); 457708d4a246SMichal Hocko if (ret) 457808d4a246SMichal Hocko goto out; 4579e5ff2159SAndi Kleen 4580238d3c13SDavid Rientjes if (write) 4581238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 4582238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 458308d4a246SMichal Hocko out: 458408d4a246SMichal Hocko return ret; 45851da177e4SLinus Torvalds } 4586396faf03SMel Gorman 458706808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 458832927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 458906808b08SLee Schermerhorn { 459006808b08SLee Schermerhorn 459106808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 459206808b08SLee Schermerhorn buffer, length, ppos); 459306808b08SLee Schermerhorn } 459406808b08SLee Schermerhorn 459506808b08SLee Schermerhorn #ifdef CONFIG_NUMA 459606808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 459732927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 459806808b08SLee Schermerhorn { 459906808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 460006808b08SLee Schermerhorn buffer, length, ppos); 460106808b08SLee Schermerhorn } 460206808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 460306808b08SLee Schermerhorn 4604a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 460532927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 4606a3d0c6aaSNishanth Aravamudan { 4607a5516438SAndi Kleen struct hstate *h = &default_hstate; 4608e5ff2159SAndi Kleen unsigned long tmp; 460908d4a246SMichal Hocko int ret; 4610e5ff2159SAndi Kleen 4611457c1b27SNishanth Aravamudan if (!hugepages_supported()) 461286613628SJan Stancek return -EOPNOTSUPP; 4613457c1b27SNishanth Aravamudan 4614e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 4615e5ff2159SAndi Kleen 4616bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 4617adbe8726SEric B Munson return -EINVAL; 4618adbe8726SEric B Munson 461917743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 462017743798SMuchun Song &tmp); 462108d4a246SMichal Hocko if (ret) 462208d4a246SMichal Hocko goto out; 4623e5ff2159SAndi Kleen 4624e5ff2159SAndi Kleen if (write) { 4625db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4626e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 4627db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4628e5ff2159SAndi Kleen } 462908d4a246SMichal Hocko out: 463008d4a246SMichal Hocko return ret; 4631a3d0c6aaSNishanth Aravamudan } 4632a3d0c6aaSNishanth Aravamudan 46331da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 46341da177e4SLinus Torvalds 4635e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 46361da177e4SLinus Torvalds { 4637fcb2b0c5SRoman Gushchin struct hstate *h; 4638fcb2b0c5SRoman Gushchin unsigned long total = 0; 4639fcb2b0c5SRoman Gushchin 4640457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4641457c1b27SNishanth Aravamudan return; 4642fcb2b0c5SRoman Gushchin 4643fcb2b0c5SRoman Gushchin for_each_hstate(h) { 4644fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 4645fcb2b0c5SRoman Gushchin 4646aca78307SMiaohe Lin total += huge_page_size(h) * count; 4647fcb2b0c5SRoman Gushchin 4648fcb2b0c5SRoman Gushchin if (h == &default_hstate) 4649e1759c21SAlexey Dobriyan seq_printf(m, 46501da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 46511da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 4652b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 46537893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 46544f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 4655fcb2b0c5SRoman Gushchin count, 4656a5516438SAndi Kleen h->free_huge_pages, 4657a5516438SAndi Kleen h->resv_huge_pages, 4658a5516438SAndi Kleen h->surplus_huge_pages, 4659aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4660fcb2b0c5SRoman Gushchin } 4661fcb2b0c5SRoman Gushchin 4662aca78307SMiaohe Lin seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); 46631da177e4SLinus Torvalds } 46641da177e4SLinus Torvalds 46657981593bSJoe Perches int hugetlb_report_node_meminfo(char *buf, int len, int nid) 46661da177e4SLinus Torvalds { 4667a5516438SAndi Kleen struct hstate *h = &default_hstate; 46687981593bSJoe Perches 4669457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4670457c1b27SNishanth Aravamudan return 0; 46717981593bSJoe Perches 46727981593bSJoe Perches return sysfs_emit_at(buf, len, 46731da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 4674a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 4675a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 4676a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 4677a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 4678a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 46791da177e4SLinus Torvalds } 46801da177e4SLinus Torvalds 4681dcadcf1cSGang Li void hugetlb_show_meminfo_node(int nid) 4682949f7ec5SDavid Rientjes { 4683949f7ec5SDavid Rientjes struct hstate *h; 4684949f7ec5SDavid Rientjes 4685457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4686457c1b27SNishanth Aravamudan return; 4687457c1b27SNishanth Aravamudan 4688949f7ec5SDavid Rientjes for_each_hstate(h) 4689dcadcf1cSGang Li printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 4690949f7ec5SDavid Rientjes nid, 4691949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 4692949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 4693949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 4694aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4695949f7ec5SDavid Rientjes } 4696949f7ec5SDavid Rientjes 46975d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 46985d317b2bSNaoya Horiguchi { 46995d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 47005d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 47015d317b2bSNaoya Horiguchi } 47025d317b2bSNaoya Horiguchi 47031da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 47041da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 47051da177e4SLinus Torvalds { 4706d0028588SWanpeng Li struct hstate *h; 4707d0028588SWanpeng Li unsigned long nr_total_pages = 0; 4708d0028588SWanpeng Li 4709d0028588SWanpeng Li for_each_hstate(h) 4710d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 4711d0028588SWanpeng Li return nr_total_pages; 47121da177e4SLinus Torvalds } 47131da177e4SLinus Torvalds 4714a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 4715fc1b8a73SMel Gorman { 4716fc1b8a73SMel Gorman int ret = -ENOMEM; 4717fc1b8a73SMel Gorman 47180aa7f354SMiaohe Lin if (!delta) 47190aa7f354SMiaohe Lin return 0; 47200aa7f354SMiaohe Lin 4721db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4722fc1b8a73SMel Gorman /* 4723fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 4724fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 4725fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 4726fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 4727fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 4728fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 4729fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 4730fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 4731fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 4732fc1b8a73SMel Gorman * 4733fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 4734fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 4735fc1b8a73SMel Gorman * we fall back to check against current free page availability as 4736fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 4737fc1b8a73SMel Gorman * semantics that cpuset has. 47388ca39e68SMuchun Song * 47398ca39e68SMuchun Song * Apart from cpuset, we also have memory policy mechanism that 47408ca39e68SMuchun Song * also determines from which node the kernel will allocate memory 47418ca39e68SMuchun Song * in a NUMA system. So similar to cpuset, we also should consider 47428ca39e68SMuchun Song * the memory policy of the current task. Similar to the description 47438ca39e68SMuchun Song * above. 4744fc1b8a73SMel Gorman */ 4745fc1b8a73SMel Gorman if (delta > 0) { 4746a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 4747fc1b8a73SMel Gorman goto out; 4748fc1b8a73SMel Gorman 47498ca39e68SMuchun Song if (delta > allowed_mems_nr(h)) { 4750a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 4751fc1b8a73SMel Gorman goto out; 4752fc1b8a73SMel Gorman } 4753fc1b8a73SMel Gorman } 4754fc1b8a73SMel Gorman 4755fc1b8a73SMel Gorman ret = 0; 4756fc1b8a73SMel Gorman if (delta < 0) 4757a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 4758fc1b8a73SMel Gorman 4759fc1b8a73SMel Gorman out: 4760db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4761fc1b8a73SMel Gorman return ret; 4762fc1b8a73SMel Gorman } 4763fc1b8a73SMel Gorman 476484afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 476584afd99bSAndy Whitcroft { 4766f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 476784afd99bSAndy Whitcroft 476884afd99bSAndy Whitcroft /* 4769612b8a31SMike Kravetz * HPAGE_RESV_OWNER indicates a private mapping. 477084afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 477184afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 477284afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 477325985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 477484afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 477584afd99bSAndy Whitcroft * new reference here without additional locking. 477684afd99bSAndy Whitcroft */ 477709a26e83SMike Kravetz if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 477809a26e83SMike Kravetz resv_map_dup_hugetlb_cgroup_uncharge_info(resv); 4779f522c3acSJoonsoo Kim kref_get(&resv->refs); 478084afd99bSAndy Whitcroft } 47818d9bfb26SMike Kravetz 4782131a79b4SMike Kravetz /* 4783131a79b4SMike Kravetz * vma_lock structure for sharable mappings is vma specific. 4784612b8a31SMike Kravetz * Clear old pointer (if copied via vm_area_dup) and allocate 4785612b8a31SMike Kravetz * new structure. Before clearing, make sure vma_lock is not 4786612b8a31SMike Kravetz * for this vma. 4787131a79b4SMike Kravetz */ 4788131a79b4SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 4789612b8a31SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 4790612b8a31SMike Kravetz 4791612b8a31SMike Kravetz if (vma_lock) { 4792612b8a31SMike Kravetz if (vma_lock->vma != vma) { 4793131a79b4SMike Kravetz vma->vm_private_data = NULL; 47948d9bfb26SMike Kravetz hugetlb_vma_lock_alloc(vma); 4795612b8a31SMike Kravetz } else 4796612b8a31SMike Kravetz pr_warn("HugeTLB: vma_lock already exists in %s.\n", __func__); 4797612b8a31SMike Kravetz } else 4798612b8a31SMike Kravetz hugetlb_vma_lock_alloc(vma); 479909a26e83SMike Kravetz } 4800131a79b4SMike Kravetz } 480184afd99bSAndy Whitcroft 4802a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 4803a1e78772SMel Gorman { 4804a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 48058d9bfb26SMike Kravetz struct resv_map *resv; 480690481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 48074e35f483SJoonsoo Kim unsigned long reserve, start, end; 48081c5ecae3SMike Kravetz long gbl_reserve; 480984afd99bSAndy Whitcroft 48108d9bfb26SMike Kravetz hugetlb_vma_lock_free(vma); 48118d9bfb26SMike Kravetz 48128d9bfb26SMike Kravetz resv = vma_resv_map(vma); 48134e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 48144e35f483SJoonsoo Kim return; 48154e35f483SJoonsoo Kim 4816a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 4817a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 481884afd99bSAndy Whitcroft 48194e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 4820e9fe92aeSMina Almasry hugetlb_cgroup_uncharge_counter(resv, start, end); 48217251ff78SAdam Litke if (reserve) { 48221c5ecae3SMike Kravetz /* 48231c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 48241c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 48251c5ecae3SMike Kravetz */ 48261c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 48271c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 48287251ff78SAdam Litke } 4829e9fe92aeSMina Almasry 4830e9fe92aeSMina Almasry kref_put(&resv->refs, resv_map_release); 4831a1e78772SMel Gorman } 4832a1e78772SMel Gorman 483331383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 483431383c68SDan Williams { 483531383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 483631383c68SDan Williams return -EINVAL; 483731383c68SDan Williams return 0; 483831383c68SDan Williams } 483931383c68SDan Williams 484005ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 484105ea8860SDan Williams { 4842aca78307SMiaohe Lin return huge_page_size(hstate_vma(vma)); 484305ea8860SDan Williams } 484405ea8860SDan Williams 48451da177e4SLinus Torvalds /* 48461da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 48471da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 48486c26d310SMiaohe Lin * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get 48491da177e4SLinus Torvalds * this far. 48501da177e4SLinus Torvalds */ 4851b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 48521da177e4SLinus Torvalds { 48531da177e4SLinus Torvalds BUG(); 4854d0217ac0SNick Piggin return 0; 48551da177e4SLinus Torvalds } 48561da177e4SLinus Torvalds 4857eec3636aSJane Chu /* 4858eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 4859eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 4860eec3636aSJane Chu * This is because under System V memory model, mappings created via 4861eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 4862eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 4863eec3636aSJane Chu */ 4864f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 4865d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 486684afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 4867a1e78772SMel Gorman .close = hugetlb_vm_op_close, 4868dd3b614fSDmitry Safonov .may_split = hugetlb_vm_op_split, 486905ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 48701da177e4SLinus Torvalds }; 48711da177e4SLinus Torvalds 48721e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 48731e8f889bSDavid Gibson int writable) 487463551ae0SDavid Gibson { 487563551ae0SDavid Gibson pte_t entry; 487679c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 487763551ae0SDavid Gibson 48781e8f889bSDavid Gibson if (writable) { 4879106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 4880106c992aSGerald Schaefer vma->vm_page_prot))); 488163551ae0SDavid Gibson } else { 4882106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 4883106c992aSGerald Schaefer vma->vm_page_prot)); 488463551ae0SDavid Gibson } 488563551ae0SDavid Gibson entry = pte_mkyoung(entry); 488679c1c594SChristophe Leroy entry = arch_make_huge_pte(entry, shift, vma->vm_flags); 488763551ae0SDavid Gibson 488863551ae0SDavid Gibson return entry; 488963551ae0SDavid Gibson } 489063551ae0SDavid Gibson 48911e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 48921e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 48931e8f889bSDavid Gibson { 48941e8f889bSDavid Gibson pte_t entry; 48951e8f889bSDavid Gibson 4896106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 489732f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 48984b3073e1SRussell King update_mmu_cache(vma, address, ptep); 48991e8f889bSDavid Gibson } 49001e8f889bSDavid Gibson 4901d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 49024a705fefSNaoya Horiguchi { 49034a705fefSNaoya Horiguchi swp_entry_t swp; 49044a705fefSNaoya Horiguchi 49054a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 4906d5ed7444SAneesh Kumar K.V return false; 49074a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4908d79d176aSBaoquan He if (is_migration_entry(swp)) 4909d5ed7444SAneesh Kumar K.V return true; 49104a705fefSNaoya Horiguchi else 4911d5ed7444SAneesh Kumar K.V return false; 49124a705fefSNaoya Horiguchi } 49134a705fefSNaoya Horiguchi 49143e5c3600SBaoquan He static bool is_hugetlb_entry_hwpoisoned(pte_t pte) 49154a705fefSNaoya Horiguchi { 49164a705fefSNaoya Horiguchi swp_entry_t swp; 49174a705fefSNaoya Horiguchi 49184a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 49193e5c3600SBaoquan He return false; 49204a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4921d79d176aSBaoquan He if (is_hwpoison_entry(swp)) 49223e5c3600SBaoquan He return true; 49234a705fefSNaoya Horiguchi else 49243e5c3600SBaoquan He return false; 49254a705fefSNaoya Horiguchi } 49261e8f889bSDavid Gibson 49274eae4efaSPeter Xu static void 49284eae4efaSPeter Xu hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, 49294eae4efaSPeter Xu struct page *new_page) 49304eae4efaSPeter Xu { 49314eae4efaSPeter Xu __SetPageUptodate(new_page); 49324eae4efaSPeter Xu hugepage_add_new_anon_rmap(new_page, vma, addr); 49331eba86c0SPasha Tatashin set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1)); 49344eae4efaSPeter Xu hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); 49354eae4efaSPeter Xu SetHPageMigratable(new_page); 49364eae4efaSPeter Xu } 49374eae4efaSPeter Xu 493863551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 4939bc70fbf2SPeter Xu struct vm_area_struct *dst_vma, 4940bc70fbf2SPeter Xu struct vm_area_struct *src_vma) 494163551ae0SDavid Gibson { 49423aa4ed80SMiaohe Lin pte_t *src_pte, *dst_pte, entry; 494363551ae0SDavid Gibson struct page *ptepage; 49441c59827dSHugh Dickins unsigned long addr; 4945bc70fbf2SPeter Xu bool cow = is_cow_mapping(src_vma->vm_flags); 4946bc70fbf2SPeter Xu struct hstate *h = hstate_vma(src_vma); 4947a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 49484eae4efaSPeter Xu unsigned long npages = pages_per_huge_page(h); 4949ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 4950e95a9851SMike Kravetz unsigned long last_addr_mask; 4951e8569dd2SAndreas Sandberg int ret = 0; 49521e8f889bSDavid Gibson 4953ac46d4f3SJérôme Glisse if (cow) { 4954bc70fbf2SPeter Xu mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src_vma, src, 4955bc70fbf2SPeter Xu src_vma->vm_start, 4956bc70fbf2SPeter Xu src_vma->vm_end); 4957ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4958623a1ddfSDavid Hildenbrand mmap_assert_write_locked(src); 4959623a1ddfSDavid Hildenbrand raw_write_seqcount_begin(&src->write_protect_seq); 496040549ba8SMike Kravetz } else { 496140549ba8SMike Kravetz /* 496240549ba8SMike Kravetz * For shared mappings the vma lock must be held before 496340549ba8SMike Kravetz * calling huge_pte_offset in the src vma. Otherwise, the 496440549ba8SMike Kravetz * returned ptep could go away if part of a shared pmd and 496540549ba8SMike Kravetz * another thread calls huge_pmd_unshare. 496640549ba8SMike Kravetz */ 496740549ba8SMike Kravetz hugetlb_vma_lock_read(src_vma); 4968ac46d4f3SJérôme Glisse } 4969e8569dd2SAndreas Sandberg 4970e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 4971bc70fbf2SPeter Xu for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) { 4972cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 49737868a208SPunit Agrawal src_pte = huge_pte_offset(src, addr, sz); 4974e95a9851SMike Kravetz if (!src_pte) { 4975e95a9851SMike Kravetz addr |= last_addr_mask; 4976c74df32cSHugh Dickins continue; 4977e95a9851SMike Kravetz } 4978bc70fbf2SPeter Xu dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz); 4979e8569dd2SAndreas Sandberg if (!dst_pte) { 4980e8569dd2SAndreas Sandberg ret = -ENOMEM; 4981e8569dd2SAndreas Sandberg break; 4982e8569dd2SAndreas Sandberg } 4983c5c99429SLarry Woodman 49845e41540cSMike Kravetz /* 49855e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 49865e41540cSMike Kravetz * 49873aa4ed80SMiaohe Lin * dst_pte == src_pte is the common case of src/dest sharing. 49885e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 49893aa4ed80SMiaohe Lin * another vma. So page_count of ptep page is checked instead 49903aa4ed80SMiaohe Lin * to reliably determine whether pte is shared. 49915e41540cSMike Kravetz */ 49923aa4ed80SMiaohe Lin if (page_count(virt_to_page(dst_pte)) > 1) { 4993e95a9851SMike Kravetz addr |= last_addr_mask; 4994c5c99429SLarry Woodman continue; 4995e95a9851SMike Kravetz } 4996c5c99429SLarry Woodman 4997cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 4998cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 4999cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 50004a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 50014eae4efaSPeter Xu again: 50023aa4ed80SMiaohe Lin if (huge_pte_none(entry)) { 50035e41540cSMike Kravetz /* 50043aa4ed80SMiaohe Lin * Skip if src entry none. 50055e41540cSMike Kravetz */ 50064a705fefSNaoya Horiguchi ; 5007c2cb0dccSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) { 5008c2cb0dccSNaoya Horiguchi bool uffd_wp = huge_pte_uffd_wp(entry); 5009c2cb0dccSNaoya Horiguchi 5010c2cb0dccSNaoya Horiguchi if (!userfaultfd_wp(dst_vma) && uffd_wp) 5011c2cb0dccSNaoya Horiguchi entry = huge_pte_clear_uffd_wp(entry); 5012c2cb0dccSNaoya Horiguchi set_huge_pte_at(dst, addr, dst_pte, entry); 5013c2cb0dccSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry))) { 50144a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 5015bc70fbf2SPeter Xu bool uffd_wp = huge_pte_uffd_wp(entry); 50164a705fefSNaoya Horiguchi 50176c287605SDavid Hildenbrand if (!is_readable_migration_entry(swp_entry) && cow) { 50184a705fefSNaoya Horiguchi /* 50194a705fefSNaoya Horiguchi * COW mappings require pages in both 50204a705fefSNaoya Horiguchi * parent and child to be set to read. 50214a705fefSNaoya Horiguchi */ 50224dd845b5SAlistair Popple swp_entry = make_readable_migration_entry( 50234dd845b5SAlistair Popple swp_offset(swp_entry)); 50244a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 5025bc70fbf2SPeter Xu if (userfaultfd_wp(src_vma) && uffd_wp) 5026bc70fbf2SPeter Xu entry = huge_pte_mkuffd_wp(entry); 502718f39629SQi Zheng set_huge_pte_at(src, addr, src_pte, entry); 50284a705fefSNaoya Horiguchi } 5029bc70fbf2SPeter Xu if (!userfaultfd_wp(dst_vma) && uffd_wp) 5030bc70fbf2SPeter Xu entry = huge_pte_clear_uffd_wp(entry); 503118f39629SQi Zheng set_huge_pte_at(dst, addr, dst_pte, entry); 5032bc70fbf2SPeter Xu } else if (unlikely(is_pte_marker(entry))) { 5033bc70fbf2SPeter Xu /* 5034bc70fbf2SPeter Xu * We copy the pte marker only if the dst vma has 5035bc70fbf2SPeter Xu * uffd-wp enabled. 5036bc70fbf2SPeter Xu */ 5037bc70fbf2SPeter Xu if (userfaultfd_wp(dst_vma)) 5038bc70fbf2SPeter Xu set_huge_pte_at(dst, addr, dst_pte, entry); 50394a705fefSNaoya Horiguchi } else { 50404eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 50414eae4efaSPeter Xu ptepage = pte_page(entry); 50424eae4efaSPeter Xu get_page(ptepage); 50434eae4efaSPeter Xu 50444eae4efaSPeter Xu /* 5045fb3d824dSDavid Hildenbrand * Failing to duplicate the anon rmap is a rare case 5046fb3d824dSDavid Hildenbrand * where we see pinned hugetlb pages while they're 5047fb3d824dSDavid Hildenbrand * prone to COW. We need to do the COW earlier during 5048fb3d824dSDavid Hildenbrand * fork. 50494eae4efaSPeter Xu * 50504eae4efaSPeter Xu * When pre-allocating the page or copying data, we 50514eae4efaSPeter Xu * need to be without the pgtable locks since we could 50524eae4efaSPeter Xu * sleep during the process. 50534eae4efaSPeter Xu */ 5054fb3d824dSDavid Hildenbrand if (!PageAnon(ptepage)) { 5055fb3d824dSDavid Hildenbrand page_dup_file_rmap(ptepage, true); 5056bc70fbf2SPeter Xu } else if (page_try_dup_anon_rmap(ptepage, true, 5057bc70fbf2SPeter Xu src_vma)) { 50584eae4efaSPeter Xu pte_t src_pte_old = entry; 50594eae4efaSPeter Xu struct page *new; 50604eae4efaSPeter Xu 50614eae4efaSPeter Xu spin_unlock(src_ptl); 50624eae4efaSPeter Xu spin_unlock(dst_ptl); 50634eae4efaSPeter Xu /* Do not use reserve as it's private owned */ 5064bc70fbf2SPeter Xu new = alloc_huge_page(dst_vma, addr, 1); 50654eae4efaSPeter Xu if (IS_ERR(new)) { 50664eae4efaSPeter Xu put_page(ptepage); 50674eae4efaSPeter Xu ret = PTR_ERR(new); 50684eae4efaSPeter Xu break; 50694eae4efaSPeter Xu } 5070bc70fbf2SPeter Xu copy_user_huge_page(new, ptepage, addr, dst_vma, 50714eae4efaSPeter Xu npages); 50724eae4efaSPeter Xu put_page(ptepage); 50734eae4efaSPeter Xu 50744eae4efaSPeter Xu /* Install the new huge page if src pte stable */ 50754eae4efaSPeter Xu dst_ptl = huge_pte_lock(h, dst, dst_pte); 50764eae4efaSPeter Xu src_ptl = huge_pte_lockptr(h, src, src_pte); 50774eae4efaSPeter Xu spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 50784eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 50794eae4efaSPeter Xu if (!pte_same(src_pte_old, entry)) { 5080bc70fbf2SPeter Xu restore_reserve_on_error(h, dst_vma, addr, 5081846be085SMike Kravetz new); 50824eae4efaSPeter Xu put_page(new); 50833aa4ed80SMiaohe Lin /* huge_ptep of dst_pte won't change as in child */ 50844eae4efaSPeter Xu goto again; 50854eae4efaSPeter Xu } 5086bc70fbf2SPeter Xu hugetlb_install_page(dst_vma, dst_pte, addr, new); 50874eae4efaSPeter Xu spin_unlock(src_ptl); 50884eae4efaSPeter Xu spin_unlock(dst_ptl); 50894eae4efaSPeter Xu continue; 50904eae4efaSPeter Xu } 50914eae4efaSPeter Xu 509234ee645eSJoerg Roedel if (cow) { 50930f10851eSJérôme Glisse /* 50940f10851eSJérôme Glisse * No need to notify as we are downgrading page 50950f10851eSJérôme Glisse * table protection not changing it to point 50960f10851eSJérôme Glisse * to a new page. 50970f10851eSJérôme Glisse * 5098ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 50990f10851eSJérôme Glisse */ 51007f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 510184894e1cSPeter Xu entry = huge_pte_wrprotect(entry); 510234ee645eSJoerg Roedel } 51034eae4efaSPeter Xu 510463551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 51054eae4efaSPeter Xu hugetlb_count_add(npages, dst); 51061c59827dSHugh Dickins } 5107cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 5108cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 510963551ae0SDavid Gibson } 511063551ae0SDavid Gibson 5111623a1ddfSDavid Hildenbrand if (cow) { 5112623a1ddfSDavid Hildenbrand raw_write_seqcount_end(&src->write_protect_seq); 5113ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 511440549ba8SMike Kravetz } else { 511540549ba8SMike Kravetz hugetlb_vma_unlock_read(src_vma); 5116623a1ddfSDavid Hildenbrand } 5117e8569dd2SAndreas Sandberg 5118e8569dd2SAndreas Sandberg return ret; 511963551ae0SDavid Gibson } 512063551ae0SDavid Gibson 5121550a7d60SMina Almasry static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr, 5122db110a99SAneesh Kumar K.V unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte) 5123550a7d60SMina Almasry { 5124550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 5125550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 5126550a7d60SMina Almasry spinlock_t *src_ptl, *dst_ptl; 5127db110a99SAneesh Kumar K.V pte_t pte; 5128550a7d60SMina Almasry 5129550a7d60SMina Almasry dst_ptl = huge_pte_lock(h, mm, dst_pte); 5130550a7d60SMina Almasry src_ptl = huge_pte_lockptr(h, mm, src_pte); 5131550a7d60SMina Almasry 5132550a7d60SMina Almasry /* 5133550a7d60SMina Almasry * We don't have to worry about the ordering of src and dst ptlocks 5134550a7d60SMina Almasry * because exclusive mmap_sem (or the i_mmap_lock) prevents deadlock. 5135550a7d60SMina Almasry */ 5136550a7d60SMina Almasry if (src_ptl != dst_ptl) 5137550a7d60SMina Almasry spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 5138550a7d60SMina Almasry 5139550a7d60SMina Almasry pte = huge_ptep_get_and_clear(mm, old_addr, src_pte); 5140550a7d60SMina Almasry set_huge_pte_at(mm, new_addr, dst_pte, pte); 5141550a7d60SMina Almasry 5142550a7d60SMina Almasry if (src_ptl != dst_ptl) 5143550a7d60SMina Almasry spin_unlock(src_ptl); 5144550a7d60SMina Almasry spin_unlock(dst_ptl); 5145550a7d60SMina Almasry } 5146550a7d60SMina Almasry 5147550a7d60SMina Almasry int move_hugetlb_page_tables(struct vm_area_struct *vma, 5148550a7d60SMina Almasry struct vm_area_struct *new_vma, 5149550a7d60SMina Almasry unsigned long old_addr, unsigned long new_addr, 5150550a7d60SMina Almasry unsigned long len) 5151550a7d60SMina Almasry { 5152550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 5153550a7d60SMina Almasry struct address_space *mapping = vma->vm_file->f_mapping; 5154550a7d60SMina Almasry unsigned long sz = huge_page_size(h); 5155550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 5156550a7d60SMina Almasry unsigned long old_end = old_addr + len; 5157e95a9851SMike Kravetz unsigned long last_addr_mask; 5158550a7d60SMina Almasry pte_t *src_pte, *dst_pte; 5159550a7d60SMina Almasry struct mmu_notifier_range range; 51603d0b95cdSBaolin Wang bool shared_pmd = false; 5161550a7d60SMina Almasry 5162550a7d60SMina Almasry mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, old_addr, 5163550a7d60SMina Almasry old_end); 5164550a7d60SMina Almasry adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 51653d0b95cdSBaolin Wang /* 51663d0b95cdSBaolin Wang * In case of shared PMDs, we should cover the maximum possible 51673d0b95cdSBaolin Wang * range. 51683d0b95cdSBaolin Wang */ 51693d0b95cdSBaolin Wang flush_cache_range(vma, range.start, range.end); 51703d0b95cdSBaolin Wang 5171550a7d60SMina Almasry mmu_notifier_invalidate_range_start(&range); 5172e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 5173550a7d60SMina Almasry /* Prevent race with file truncation */ 517440549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 5175550a7d60SMina Almasry i_mmap_lock_write(mapping); 5176550a7d60SMina Almasry for (; old_addr < old_end; old_addr += sz, new_addr += sz) { 5177550a7d60SMina Almasry src_pte = huge_pte_offset(mm, old_addr, sz); 5178e95a9851SMike Kravetz if (!src_pte) { 5179e95a9851SMike Kravetz old_addr |= last_addr_mask; 5180e95a9851SMike Kravetz new_addr |= last_addr_mask; 5181550a7d60SMina Almasry continue; 5182e95a9851SMike Kravetz } 5183550a7d60SMina Almasry if (huge_pte_none(huge_ptep_get(src_pte))) 5184550a7d60SMina Almasry continue; 5185550a7d60SMina Almasry 51864ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) { 51873d0b95cdSBaolin Wang shared_pmd = true; 51884ddb4d91SMike Kravetz old_addr |= last_addr_mask; 51894ddb4d91SMike Kravetz new_addr |= last_addr_mask; 5190550a7d60SMina Almasry continue; 51913d0b95cdSBaolin Wang } 5192550a7d60SMina Almasry 5193550a7d60SMina Almasry dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz); 5194550a7d60SMina Almasry if (!dst_pte) 5195550a7d60SMina Almasry break; 5196550a7d60SMina Almasry 5197db110a99SAneesh Kumar K.V move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte); 5198550a7d60SMina Almasry } 51993d0b95cdSBaolin Wang 52003d0b95cdSBaolin Wang if (shared_pmd) 52013d0b95cdSBaolin Wang flush_tlb_range(vma, range.start, range.end); 52023d0b95cdSBaolin Wang else 5203550a7d60SMina Almasry flush_tlb_range(vma, old_end - len, old_end); 5204550a7d60SMina Almasry mmu_notifier_invalidate_range_end(&range); 520513e4ad2cSNadav Amit i_mmap_unlock_write(mapping); 520640549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 5207550a7d60SMina Almasry 5208550a7d60SMina Almasry return len + old_addr - old_end; 5209550a7d60SMina Almasry } 5210550a7d60SMina Almasry 521173c54763SPeter Xu static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 521224669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 521305e90bd0SPeter Xu struct page *ref_page, zap_flags_t zap_flags) 521463551ae0SDavid Gibson { 521563551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 521663551ae0SDavid Gibson unsigned long address; 5217c7546f8fSDavid Gibson pte_t *ptep; 521863551ae0SDavid Gibson pte_t pte; 5219cb900f41SKirill A. Shutemov spinlock_t *ptl; 522063551ae0SDavid Gibson struct page *page; 5221a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 5222a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 5223e95a9851SMike Kravetz unsigned long last_addr_mask; 5224a4a118f2SNadav Amit bool force_flush = false; 5225a5516438SAndi Kleen 522663551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 5227a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 5228a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 522963551ae0SDavid Gibson 523007e32661SAneesh Kumar K.V /* 523107e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 523207e32661SAneesh Kumar K.V * to huge page. 523307e32661SAneesh Kumar K.V */ 5234ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 523524669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 5236dff11abeSMike Kravetz 5237e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 5238569f48b8SHillf Danton address = start; 5239569f48b8SHillf Danton for (; address < end; address += sz) { 52407868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, sz); 5241e95a9851SMike Kravetz if (!ptep) { 5242e95a9851SMike Kravetz address |= last_addr_mask; 5243c7546f8fSDavid Gibson continue; 5244e95a9851SMike Kravetz } 5245c7546f8fSDavid Gibson 5246cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 52474ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, address, ptep)) { 524831d49da5SAneesh Kumar K.V spin_unlock(ptl); 5249a4a118f2SNadav Amit tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE); 5250a4a118f2SNadav Amit force_flush = true; 52514ddb4d91SMike Kravetz address |= last_addr_mask; 525231d49da5SAneesh Kumar K.V continue; 525331d49da5SAneesh Kumar K.V } 525439dde65cSChen, Kenneth W 52556629326bSHillf Danton pte = huge_ptep_get(ptep); 525631d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 525731d49da5SAneesh Kumar K.V spin_unlock(ptl); 525831d49da5SAneesh Kumar K.V continue; 525931d49da5SAneesh Kumar K.V } 52606629326bSHillf Danton 52616629326bSHillf Danton /* 52629fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 52639fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 52646629326bSHillf Danton */ 52659fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 526605e90bd0SPeter Xu /* 526705e90bd0SPeter Xu * If the pte was wr-protected by uffd-wp in any of the 526805e90bd0SPeter Xu * swap forms, meanwhile the caller does not want to 526905e90bd0SPeter Xu * drop the uffd-wp bit in this zap, then replace the 527005e90bd0SPeter Xu * pte with a marker. 527105e90bd0SPeter Xu */ 527205e90bd0SPeter Xu if (pte_swp_uffd_wp_any(pte) && 527305e90bd0SPeter Xu !(zap_flags & ZAP_FLAG_DROP_MARKER)) 527405e90bd0SPeter Xu set_huge_pte_at(mm, address, ptep, 527505e90bd0SPeter Xu make_pte_marker(PTE_MARKER_UFFD_WP)); 527605e90bd0SPeter Xu else 52779386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 527831d49da5SAneesh Kumar K.V spin_unlock(ptl); 527931d49da5SAneesh Kumar K.V continue; 52808c4894c6SNaoya Horiguchi } 52816629326bSHillf Danton 52826629326bSHillf Danton page = pte_page(pte); 528304f2cbe3SMel Gorman /* 528404f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 528504f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 528604f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 528704f2cbe3SMel Gorman */ 528804f2cbe3SMel Gorman if (ref_page) { 528931d49da5SAneesh Kumar K.V if (page != ref_page) { 529031d49da5SAneesh Kumar K.V spin_unlock(ptl); 529131d49da5SAneesh Kumar K.V continue; 529231d49da5SAneesh Kumar K.V } 529304f2cbe3SMel Gorman /* 529404f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 529504f2cbe3SMel Gorman * future faults in this VMA will fail rather than 529604f2cbe3SMel Gorman * looking like data was lost 529704f2cbe3SMel Gorman */ 529804f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 529904f2cbe3SMel Gorman } 530004f2cbe3SMel Gorman 5301c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 5302b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 5303106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 53046649a386SKen Chen set_page_dirty(page); 530505e90bd0SPeter Xu /* Leave a uffd-wp pte marker if needed */ 530605e90bd0SPeter Xu if (huge_pte_uffd_wp(pte) && 530705e90bd0SPeter Xu !(zap_flags & ZAP_FLAG_DROP_MARKER)) 530805e90bd0SPeter Xu set_huge_pte_at(mm, address, ptep, 530905e90bd0SPeter Xu make_pte_marker(PTE_MARKER_UFFD_WP)); 53105d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 5311cea86fe2SHugh Dickins page_remove_rmap(page, vma, true); 531231d49da5SAneesh Kumar K.V 5313cb900f41SKirill A. Shutemov spin_unlock(ptl); 5314e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 531524669e58SAneesh Kumar K.V /* 531631d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 531724669e58SAneesh Kumar K.V */ 531831d49da5SAneesh Kumar K.V if (ref_page) 531931d49da5SAneesh Kumar K.V break; 5320fe1668aeSChen, Kenneth W } 532124669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 5322a4a118f2SNadav Amit 5323a4a118f2SNadav Amit /* 5324a4a118f2SNadav Amit * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We 5325a4a118f2SNadav Amit * could defer the flush until now, since by holding i_mmap_rwsem we 5326a4a118f2SNadav Amit * guaranteed that the last refernece would not be dropped. But we must 5327a4a118f2SNadav Amit * do the flushing before we return, as otherwise i_mmap_rwsem will be 5328a4a118f2SNadav Amit * dropped and the last reference to the shared PMDs page might be 5329a4a118f2SNadav Amit * dropped as well. 5330a4a118f2SNadav Amit * 5331a4a118f2SNadav Amit * In theory we could defer the freeing of the PMD pages as well, but 5332a4a118f2SNadav Amit * huge_pmd_unshare() relies on the exact page_count for the PMD page to 5333a4a118f2SNadav Amit * detect sharing, so we cannot defer the release of the page either. 5334a4a118f2SNadav Amit * Instead, do flush now. 5335a4a118f2SNadav Amit */ 5336a4a118f2SNadav Amit if (force_flush) 5337a4a118f2SNadav Amit tlb_flush_mmu_tlbonly(tlb); 53381da177e4SLinus Torvalds } 533963551ae0SDavid Gibson 5340d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 5341d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 534205e90bd0SPeter Xu unsigned long end, struct page *ref_page, 534305e90bd0SPeter Xu zap_flags_t zap_flags) 5344d833352aSMel Gorman { 5345131a79b4SMike Kravetz hugetlb_vma_lock_write(vma); 5346131a79b4SMike Kravetz i_mmap_lock_write(vma->vm_file->f_mapping); 5347131a79b4SMike Kravetz 5348369258ceSMike Kravetz /* mmu notification performed in caller */ 534905e90bd0SPeter Xu __unmap_hugepage_range(tlb, vma, start, end, ref_page, zap_flags); 5350d833352aSMel Gorman 535104ada095SMike Kravetz if (zap_flags & ZAP_FLAG_UNMAP) { /* final unmap */ 5352d833352aSMel Gorman /* 535304ada095SMike Kravetz * Unlock and free the vma lock before releasing i_mmap_rwsem. 535404ada095SMike Kravetz * When the vma_lock is freed, this makes the vma ineligible 535504ada095SMike Kravetz * for pmd sharing. And, i_mmap_rwsem is required to set up 535604ada095SMike Kravetz * pmd sharing. This is important as page tables for this 535704ada095SMike Kravetz * unmapped range will be asynchrously deleted. If the page 535804ada095SMike Kravetz * tables are shared, there will be issues when accessed by 535904ada095SMike Kravetz * someone else. 5360d833352aSMel Gorman */ 5361ecfbd733SMike Kravetz __hugetlb_vma_unlock_write_free(vma); 5362131a79b4SMike Kravetz i_mmap_unlock_write(vma->vm_file->f_mapping); 536304ada095SMike Kravetz } else { 536404ada095SMike Kravetz i_mmap_unlock_write(vma->vm_file->f_mapping); 536504ada095SMike Kravetz hugetlb_vma_unlock_write(vma); 536604ada095SMike Kravetz } 5367d833352aSMel Gorman } 5368d833352aSMel Gorman 5369502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 537005e90bd0SPeter Xu unsigned long end, struct page *ref_page, 537105e90bd0SPeter Xu zap_flags_t zap_flags) 5372502717f4SChen, Kenneth W { 5373369258ceSMike Kravetz struct mmu_notifier_range range; 537424669e58SAneesh Kumar K.V struct mmu_gather tlb; 5375dff11abeSMike Kravetz 5376369258ceSMike Kravetz mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm, 5377369258ceSMike Kravetz start, end); 5378369258ceSMike Kravetz adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 5379369258ceSMike Kravetz mmu_notifier_invalidate_range_start(&range); 5380a72afd87SWill Deacon tlb_gather_mmu(&tlb, vma->vm_mm); 5381369258ceSMike Kravetz 538205e90bd0SPeter Xu __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags); 5383369258ceSMike Kravetz 5384369258ceSMike Kravetz mmu_notifier_invalidate_range_end(&range); 5385ae8eba8bSWill Deacon tlb_finish_mmu(&tlb); 5386502717f4SChen, Kenneth W } 5387502717f4SChen, Kenneth W 538804f2cbe3SMel Gorman /* 538904f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 5390578b7725SZhiyuan Dai * mapping it owns the reserve page for. The intention is to unmap the page 539104f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 539204f2cbe3SMel Gorman * same region. 539304f2cbe3SMel Gorman */ 53942f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 53952a4b3dedSHarvey Harrison struct page *page, unsigned long address) 539604f2cbe3SMel Gorman { 53977526674dSAdam Litke struct hstate *h = hstate_vma(vma); 539804f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 539904f2cbe3SMel Gorman struct address_space *mapping; 540004f2cbe3SMel Gorman pgoff_t pgoff; 540104f2cbe3SMel Gorman 540204f2cbe3SMel Gorman /* 540304f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 540404f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 540504f2cbe3SMel Gorman */ 54067526674dSAdam Litke address = address & huge_page_mask(h); 540736e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 540836e4f20aSMichal Hocko vma->vm_pgoff; 540993c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 541004f2cbe3SMel Gorman 54114eb2b1dcSMel Gorman /* 54124eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 54134eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 54144eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 54154eb2b1dcSMel Gorman */ 541683cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 54176b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 541804f2cbe3SMel Gorman /* Do not unmap the current VMA */ 541904f2cbe3SMel Gorman if (iter_vma == vma) 542004f2cbe3SMel Gorman continue; 542104f2cbe3SMel Gorman 542204f2cbe3SMel Gorman /* 54232f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 54242f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 54252f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 54262f84a899SMel Gorman */ 54272f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 54282f84a899SMel Gorman continue; 54292f84a899SMel Gorman 54302f84a899SMel Gorman /* 543104f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 543204f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 543304f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 543404f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 543504f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 543604f2cbe3SMel Gorman */ 543704f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 543824669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 543905e90bd0SPeter Xu address + huge_page_size(h), page, 0); 544004f2cbe3SMel Gorman } 544183cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 544204f2cbe3SMel Gorman } 544304f2cbe3SMel Gorman 54440fe6e20bSNaoya Horiguchi /* 5445c89357e2SDavid Hildenbrand * hugetlb_wp() should be called with page lock of the original hugepage held. 5446aa6d2e8cSBaolin Wang * Called with hugetlb_fault_mutex_table held and pte_page locked so we 5447ef009b25SMichal Hocko * cannot race with other handlers or page migration. 5448ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 54490fe6e20bSNaoya Horiguchi */ 5450c89357e2SDavid Hildenbrand static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma, 5451c89357e2SDavid Hildenbrand unsigned long address, pte_t *ptep, unsigned int flags, 5452cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 54531e8f889bSDavid Gibson { 5454c89357e2SDavid Hildenbrand const bool unshare = flags & FAULT_FLAG_UNSHARE; 54553999f52eSAneesh Kumar K.V pte_t pte; 5456a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 54571e8f889bSDavid Gibson struct page *old_page, *new_page; 54582b740303SSouptick Joarder int outside_reserve = 0; 54592b740303SSouptick Joarder vm_fault_t ret = 0; 5460974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 5461ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 54621e8f889bSDavid Gibson 54631d8d1464SDavid Hildenbrand /* 54641d8d1464SDavid Hildenbrand * hugetlb does not support FOLL_FORCE-style write faults that keep the 54651d8d1464SDavid Hildenbrand * PTE mapped R/O such as maybe_mkwrite() would do. 54661d8d1464SDavid Hildenbrand */ 54671d8d1464SDavid Hildenbrand if (WARN_ON_ONCE(!unshare && !(vma->vm_flags & VM_WRITE))) 54681d8d1464SDavid Hildenbrand return VM_FAULT_SIGSEGV; 54691d8d1464SDavid Hildenbrand 54701d8d1464SDavid Hildenbrand /* Let's take out MAP_SHARED mappings first. */ 54711d8d1464SDavid Hildenbrand if (vma->vm_flags & VM_MAYSHARE) { 54721d8d1464SDavid Hildenbrand set_huge_ptep_writable(vma, haddr, ptep); 54731d8d1464SDavid Hildenbrand return 0; 54741d8d1464SDavid Hildenbrand } 54751d8d1464SDavid Hildenbrand 54763999f52eSAneesh Kumar K.V pte = huge_ptep_get(ptep); 54771e8f889bSDavid Gibson old_page = pte_page(pte); 54781e8f889bSDavid Gibson 5479662ce1dcSYang Yang delayacct_wpcopy_start(); 5480662ce1dcSYang Yang 548104f2cbe3SMel Gorman retry_avoidcopy: 5482c89357e2SDavid Hildenbrand /* 5483c89357e2SDavid Hildenbrand * If no-one else is actually using this page, we're the exclusive 5484c89357e2SDavid Hildenbrand * owner and can reuse this page. 5485c89357e2SDavid Hildenbrand */ 548637a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 5487c89357e2SDavid Hildenbrand if (!PageAnonExclusive(old_page)) 54885a49973dSHugh Dickins page_move_anon_rmap(old_page, vma); 5489c89357e2SDavid Hildenbrand if (likely(!unshare)) 54905b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 5491662ce1dcSYang Yang 5492662ce1dcSYang Yang delayacct_wpcopy_end(); 549383c54070SNick Piggin return 0; 54941e8f889bSDavid Gibson } 54956c287605SDavid Hildenbrand VM_BUG_ON_PAGE(PageAnon(old_page) && PageAnonExclusive(old_page), 54966c287605SDavid Hildenbrand old_page); 54971e8f889bSDavid Gibson 549804f2cbe3SMel Gorman /* 549904f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 550004f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 550104f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 550204f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 550304f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 550404f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 550504f2cbe3SMel Gorman * of the full address range. 550604f2cbe3SMel Gorman */ 55075944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 550804f2cbe3SMel Gorman old_page != pagecache_page) 550904f2cbe3SMel Gorman outside_reserve = 1; 551004f2cbe3SMel Gorman 551109cbfeafSKirill A. Shutemov get_page(old_page); 5512b76c8cfbSLarry Woodman 5513ad4404a2SDavidlohr Bueso /* 5514ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 5515ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 5516ad4404a2SDavidlohr Bueso */ 5517cb900f41SKirill A. Shutemov spin_unlock(ptl); 55185b7a1d40SHuang Ying new_page = alloc_huge_page(vma, haddr, outside_reserve); 55191e8f889bSDavid Gibson 55202fc39cecSAdam Litke if (IS_ERR(new_page)) { 552104f2cbe3SMel Gorman /* 552204f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 552304f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 552404f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 552504f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 552604f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 552704f2cbe3SMel Gorman */ 552804f2cbe3SMel Gorman if (outside_reserve) { 552940549ba8SMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 553040549ba8SMike Kravetz pgoff_t idx; 553140549ba8SMike Kravetz u32 hash; 553240549ba8SMike Kravetz 553309cbfeafSKirill A. Shutemov put_page(old_page); 553440549ba8SMike Kravetz /* 553540549ba8SMike Kravetz * Drop hugetlb_fault_mutex and vma_lock before 553640549ba8SMike Kravetz * unmapping. unmapping needs to hold vma_lock 553740549ba8SMike Kravetz * in write mode. Dropping vma_lock in read mode 553840549ba8SMike Kravetz * here is OK as COW mappings do not interact with 553940549ba8SMike Kravetz * PMD sharing. 554040549ba8SMike Kravetz * 554140549ba8SMike Kravetz * Reacquire both after unmap operation. 554240549ba8SMike Kravetz */ 554340549ba8SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 554440549ba8SMike Kravetz hash = hugetlb_fault_mutex_hash(mapping, idx); 554540549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 554640549ba8SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 554740549ba8SMike Kravetz 55485b7a1d40SHuang Ying unmap_ref_private(mm, vma, old_page, haddr); 554940549ba8SMike Kravetz 555040549ba8SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 555140549ba8SMike Kravetz hugetlb_vma_lock_read(vma); 5552cb900f41SKirill A. Shutemov spin_lock(ptl); 55535b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 5554a9af0c5dSNaoya Horiguchi if (likely(ptep && 5555a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 555604f2cbe3SMel Gorman goto retry_avoidcopy; 5557a734bcc8SHillf Danton /* 5558cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 5559cb900f41SKirill A. Shutemov * lock, and our job is done. 5560a734bcc8SHillf Danton */ 5561662ce1dcSYang Yang delayacct_wpcopy_end(); 5562a734bcc8SHillf Danton return 0; 556304f2cbe3SMel Gorman } 556404f2cbe3SMel Gorman 55652b740303SSouptick Joarder ret = vmf_error(PTR_ERR(new_page)); 5566ad4404a2SDavidlohr Bueso goto out_release_old; 55671e8f889bSDavid Gibson } 55681e8f889bSDavid Gibson 55690fe6e20bSNaoya Horiguchi /* 55700fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 55710fe6e20bSNaoya Horiguchi * anon_vma prepared. 55720fe6e20bSNaoya Horiguchi */ 557344e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 5574ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 5575ad4404a2SDavidlohr Bueso goto out_release_all; 557644e2aa93SDean Nelson } 55770fe6e20bSNaoya Horiguchi 5578974e6d66SHuang Ying copy_user_huge_page(new_page, old_page, address, vma, 557947ad8475SAndrea Arcangeli pages_per_huge_page(h)); 55800ed361deSNick Piggin __SetPageUptodate(new_page); 55811e8f889bSDavid Gibson 55827269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, 55836f4f13e8SJérôme Glisse haddr + huge_page_size(h)); 5584ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 5585ad4404a2SDavidlohr Bueso 5586b76c8cfbSLarry Woodman /* 5587cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 5588b76c8cfbSLarry Woodman * before the page tables are altered 5589b76c8cfbSLarry Woodman */ 5590cb900f41SKirill A. Shutemov spin_lock(ptl); 55915b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 5592a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 5593c89357e2SDavid Hildenbrand /* Break COW or unshare */ 55945b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 5595ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, range.end); 5596cea86fe2SHugh Dickins page_remove_rmap(old_page, vma, true); 55975b7a1d40SHuang Ying hugepage_add_new_anon_rmap(new_page, vma, haddr); 55981eba86c0SPasha Tatashin set_huge_pte_at(mm, haddr, ptep, 5599c89357e2SDavid Hildenbrand make_huge_pte(vma, new_page, !unshare)); 56008f251a3dSMike Kravetz SetHPageMigratable(new_page); 56011e8f889bSDavid Gibson /* Make the old page be freed below */ 56021e8f889bSDavid Gibson new_page = old_page; 56031e8f889bSDavid Gibson } 5604cb900f41SKirill A. Shutemov spin_unlock(ptl); 5605ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 5606ad4404a2SDavidlohr Bueso out_release_all: 5607c89357e2SDavid Hildenbrand /* 5608c89357e2SDavid Hildenbrand * No restore in case of successful pagetable update (Break COW or 5609c89357e2SDavid Hildenbrand * unshare) 5610c89357e2SDavid Hildenbrand */ 5611c7b1850dSMike Kravetz if (new_page != old_page) 56125b7a1d40SHuang Ying restore_reserve_on_error(h, vma, haddr, new_page); 561309cbfeafSKirill A. Shutemov put_page(new_page); 5614ad4404a2SDavidlohr Bueso out_release_old: 561509cbfeafSKirill A. Shutemov put_page(old_page); 56168312034fSJoonsoo Kim 5617ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 5618662ce1dcSYang Yang 5619662ce1dcSYang Yang delayacct_wpcopy_end(); 5620ad4404a2SDavidlohr Bueso return ret; 56211e8f889bSDavid Gibson } 56221e8f889bSDavid Gibson 56233ae77f43SHugh Dickins /* 56243ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 56253ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 56263ae77f43SHugh Dickins */ 56273ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 56282a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 56292a15efc9SHugh Dickins { 56302a15efc9SHugh Dickins struct address_space *mapping; 56312a15efc9SHugh Dickins pgoff_t idx; 56322a15efc9SHugh Dickins struct page *page; 56332a15efc9SHugh Dickins 56342a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 56352a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 56362a15efc9SHugh Dickins 56372a15efc9SHugh Dickins page = find_get_page(mapping, idx); 56382a15efc9SHugh Dickins if (page) 56392a15efc9SHugh Dickins put_page(page); 56402a15efc9SHugh Dickins return page != NULL; 56412a15efc9SHugh Dickins } 56422a15efc9SHugh Dickins 56437e1813d4SMike Kravetz int hugetlb_add_to_page_cache(struct page *page, struct address_space *mapping, 5644ab76ad54SMike Kravetz pgoff_t idx) 5645ab76ad54SMike Kravetz { 5646d9ef44deSMatthew Wilcox (Oracle) struct folio *folio = page_folio(page); 5647ab76ad54SMike Kravetz struct inode *inode = mapping->host; 5648ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 5649d9ef44deSMatthew Wilcox (Oracle) int err; 5650ab76ad54SMike Kravetz 5651d9ef44deSMatthew Wilcox (Oracle) __folio_set_locked(folio); 5652d9ef44deSMatthew Wilcox (Oracle) err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL); 5653d9ef44deSMatthew Wilcox (Oracle) 5654d9ef44deSMatthew Wilcox (Oracle) if (unlikely(err)) { 5655d9ef44deSMatthew Wilcox (Oracle) __folio_clear_locked(folio); 5656ab76ad54SMike Kravetz return err; 5657d9ef44deSMatthew Wilcox (Oracle) } 5658d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 5659ab76ad54SMike Kravetz 566022146c3cSMike Kravetz /* 5661d9ef44deSMatthew Wilcox (Oracle) * mark folio dirty so that it will not be removed from cache/file 566222146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 566322146c3cSMike Kravetz */ 5664d9ef44deSMatthew Wilcox (Oracle) folio_mark_dirty(folio); 566522146c3cSMike Kravetz 5666ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 5667ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 5668ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 5669ab76ad54SMike Kravetz return 0; 5670ab76ad54SMike Kravetz } 5671ab76ad54SMike Kravetz 56727677f7fdSAxel Rasmussen static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma, 56737677f7fdSAxel Rasmussen struct address_space *mapping, 56747677f7fdSAxel Rasmussen pgoff_t idx, 56757677f7fdSAxel Rasmussen unsigned int flags, 56767677f7fdSAxel Rasmussen unsigned long haddr, 5677824ddc60SNadav Amit unsigned long addr, 56787677f7fdSAxel Rasmussen unsigned long reason) 56797677f7fdSAxel Rasmussen { 56807677f7fdSAxel Rasmussen u32 hash; 56817677f7fdSAxel Rasmussen struct vm_fault vmf = { 56827677f7fdSAxel Rasmussen .vma = vma, 56837677f7fdSAxel Rasmussen .address = haddr, 5684824ddc60SNadav Amit .real_address = addr, 56857677f7fdSAxel Rasmussen .flags = flags, 56867677f7fdSAxel Rasmussen 56877677f7fdSAxel Rasmussen /* 56887677f7fdSAxel Rasmussen * Hard to debug if it ends up being 56897677f7fdSAxel Rasmussen * used by a callee that assumes 56907677f7fdSAxel Rasmussen * something about the other 56917677f7fdSAxel Rasmussen * uninitialized fields... same as in 56927677f7fdSAxel Rasmussen * memory.c 56937677f7fdSAxel Rasmussen */ 56947677f7fdSAxel Rasmussen }; 56957677f7fdSAxel Rasmussen 56967677f7fdSAxel Rasmussen /* 5697958f32ceSLiu Shixin * vma_lock and hugetlb_fault_mutex must be dropped before handling 5698958f32ceSLiu Shixin * userfault. Also mmap_lock could be dropped due to handling 5699958f32ceSLiu Shixin * userfault, any vma operation should be careful from here. 57007677f7fdSAxel Rasmussen */ 570140549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 57027677f7fdSAxel Rasmussen hash = hugetlb_fault_mutex_hash(mapping, idx); 57037677f7fdSAxel Rasmussen mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5704958f32ceSLiu Shixin return handle_userfault(&vmf, reason); 57057677f7fdSAxel Rasmussen } 57067677f7fdSAxel Rasmussen 57072ea7ff1eSPeter Xu /* 57082ea7ff1eSPeter Xu * Recheck pte with pgtable lock. Returns true if pte didn't change, or 57092ea7ff1eSPeter Xu * false if pte changed or is changing. 57102ea7ff1eSPeter Xu */ 57112ea7ff1eSPeter Xu static bool hugetlb_pte_stable(struct hstate *h, struct mm_struct *mm, 57122ea7ff1eSPeter Xu pte_t *ptep, pte_t old_pte) 57132ea7ff1eSPeter Xu { 57142ea7ff1eSPeter Xu spinlock_t *ptl; 57152ea7ff1eSPeter Xu bool same; 57162ea7ff1eSPeter Xu 57172ea7ff1eSPeter Xu ptl = huge_pte_lock(h, mm, ptep); 57182ea7ff1eSPeter Xu same = pte_same(huge_ptep_get(ptep), old_pte); 57192ea7ff1eSPeter Xu spin_unlock(ptl); 57202ea7ff1eSPeter Xu 57212ea7ff1eSPeter Xu return same; 57222ea7ff1eSPeter Xu } 57232ea7ff1eSPeter Xu 57242b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 57252b740303SSouptick Joarder struct vm_area_struct *vma, 57268382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 5727c64e912cSPeter Xu unsigned long address, pte_t *ptep, 5728c64e912cSPeter Xu pte_t old_pte, unsigned int flags) 5729ac9b9c66SHugh Dickins { 5730a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 57312b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 5732409eb8c2SHillf Danton int anon_rmap = 0; 57334c887265SAdam Litke unsigned long size; 57344c887265SAdam Litke struct page *page; 57351e8f889bSDavid Gibson pte_t new_pte; 5736cb900f41SKirill A. Shutemov spinlock_t *ptl; 5737285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 5738c7b1850dSMike Kravetz bool new_page, new_pagecache_page = false; 5739958f32ceSLiu Shixin u32 hash = hugetlb_fault_mutex_hash(mapping, idx); 57404c887265SAdam Litke 574104f2cbe3SMel Gorman /* 574204f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 574304f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 5744c89357e2SDavid Hildenbrand * COW/unsharing. Warn that such a situation has occurred as it may not 5745c89357e2SDavid Hildenbrand * be obvious. 574604f2cbe3SMel Gorman */ 574704f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 5748910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 574904f2cbe3SMel Gorman current->pid); 5750958f32ceSLiu Shixin goto out; 575104f2cbe3SMel Gorman } 575204f2cbe3SMel Gorman 57534c887265SAdam Litke /* 5754188a3972SMike Kravetz * Use page lock to guard against racing truncation 5755188a3972SMike Kravetz * before we get page_table_lock. 57564c887265SAdam Litke */ 5757c7b1850dSMike Kravetz new_page = false; 575887bf91d3SMike Kravetz page = find_lock_page(mapping, idx); 575987bf91d3SMike Kravetz if (!page) { 5760188a3972SMike Kravetz size = i_size_read(mapping->host) >> huge_page_shift(h); 5761188a3972SMike Kravetz if (idx >= size) 5762188a3972SMike Kravetz goto out; 57637677f7fdSAxel Rasmussen /* Check for page in userfault range */ 57642ea7ff1eSPeter Xu if (userfaultfd_missing(vma)) { 57652ea7ff1eSPeter Xu /* 57662ea7ff1eSPeter Xu * Since hugetlb_no_page() was examining pte 57672ea7ff1eSPeter Xu * without pgtable lock, we need to re-test under 57682ea7ff1eSPeter Xu * lock because the pte may not be stable and could 57692ea7ff1eSPeter Xu * have changed from under us. Try to detect 57702ea7ff1eSPeter Xu * either changed or during-changing ptes and retry 57712ea7ff1eSPeter Xu * properly when needed. 57722ea7ff1eSPeter Xu * 57732ea7ff1eSPeter Xu * Note that userfaultfd is actually fine with 57742ea7ff1eSPeter Xu * false positives (e.g. caused by pte changed), 57752ea7ff1eSPeter Xu * but not wrong logical events (e.g. caused by 57762ea7ff1eSPeter Xu * reading a pte during changing). The latter can 57772ea7ff1eSPeter Xu * confuse the userspace, so the strictness is very 57782ea7ff1eSPeter Xu * much preferred. E.g., MISSING event should 57792ea7ff1eSPeter Xu * never happen on the page after UFFDIO_COPY has 57802ea7ff1eSPeter Xu * correctly installed the page and returned. 57812ea7ff1eSPeter Xu */ 57822ea7ff1eSPeter Xu if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) { 57832ea7ff1eSPeter Xu ret = 0; 57842ea7ff1eSPeter Xu goto out; 57852ea7ff1eSPeter Xu } 57862ea7ff1eSPeter Xu 57872ea7ff1eSPeter Xu return hugetlb_handle_userfault(vma, mapping, idx, flags, 57882ea7ff1eSPeter Xu haddr, address, 57897677f7fdSAxel Rasmussen VM_UFFD_MISSING); 57902ea7ff1eSPeter Xu } 57911a1aad8aSMike Kravetz 5792285b8dcaSHuang Ying page = alloc_huge_page(vma, haddr, 0); 57932fc39cecSAdam Litke if (IS_ERR(page)) { 57944643d67eSMike Kravetz /* 57954643d67eSMike Kravetz * Returning error will result in faulting task being 57964643d67eSMike Kravetz * sent SIGBUS. The hugetlb fault mutex prevents two 57974643d67eSMike Kravetz * tasks from racing to fault in the same page which 57984643d67eSMike Kravetz * could result in false unable to allocate errors. 57994643d67eSMike Kravetz * Page migration does not take the fault mutex, but 58004643d67eSMike Kravetz * does a clear then write of pte's under page table 58014643d67eSMike Kravetz * lock. Page fault code could race with migration, 58024643d67eSMike Kravetz * notice the clear pte and try to allocate a page 58034643d67eSMike Kravetz * here. Before returning error, get ptl and make 58044643d67eSMike Kravetz * sure there really is no pte entry. 58054643d67eSMike Kravetz */ 5806f9bf6c03SPeter Xu if (hugetlb_pte_stable(h, mm, ptep, old_pte)) 58072b740303SSouptick Joarder ret = vmf_error(PTR_ERR(page)); 5808f9bf6c03SPeter Xu else 5809f9bf6c03SPeter Xu ret = 0; 58106bda666aSChristoph Lameter goto out; 58116bda666aSChristoph Lameter } 581247ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 58130ed361deSNick Piggin __SetPageUptodate(page); 5814cb6acd01SMike Kravetz new_page = true; 5815ac9b9c66SHugh Dickins 5816f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 58177e1813d4SMike Kravetz int err = hugetlb_add_to_page_cache(page, mapping, idx); 58186bda666aSChristoph Lameter if (err) { 58193a5497a2SMiaohe Lin /* 58203a5497a2SMiaohe Lin * err can't be -EEXIST which implies someone 58213a5497a2SMiaohe Lin * else consumed the reservation since hugetlb 58223a5497a2SMiaohe Lin * fault mutex is held when add a hugetlb page 58233a5497a2SMiaohe Lin * to the page cache. So it's safe to call 58243a5497a2SMiaohe Lin * restore_reserve_on_error() here. 58253a5497a2SMiaohe Lin */ 58263a5497a2SMiaohe Lin restore_reserve_on_error(h, vma, haddr, page); 58276bda666aSChristoph Lameter put_page(page); 58286bda666aSChristoph Lameter goto out; 58296bda666aSChristoph Lameter } 5830c7b1850dSMike Kravetz new_pagecache_page = true; 583123be7468SMel Gorman } else { 58326bda666aSChristoph Lameter lock_page(page); 58330fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 58340fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 58350fe6e20bSNaoya Horiguchi goto backout_unlocked; 583623be7468SMel Gorman } 5837409eb8c2SHillf Danton anon_rmap = 1; 58380fe6e20bSNaoya Horiguchi } 58390fe6e20bSNaoya Horiguchi } else { 584057303d80SAndy Whitcroft /* 5841998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 5842998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 5843998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 5844fd6a03edSNaoya Horiguchi */ 5845fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 58460eb98f15SMiaohe Lin ret = VM_FAULT_HWPOISON_LARGE | 5847972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 5848fd6a03edSNaoya Horiguchi goto backout_unlocked; 58496bda666aSChristoph Lameter } 58507677f7fdSAxel Rasmussen 58517677f7fdSAxel Rasmussen /* Check for page in userfault range. */ 58527677f7fdSAxel Rasmussen if (userfaultfd_minor(vma)) { 58537677f7fdSAxel Rasmussen unlock_page(page); 58547677f7fdSAxel Rasmussen put_page(page); 58552ea7ff1eSPeter Xu /* See comment in userfaultfd_missing() block above */ 58562ea7ff1eSPeter Xu if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) { 58572ea7ff1eSPeter Xu ret = 0; 58582ea7ff1eSPeter Xu goto out; 58592ea7ff1eSPeter Xu } 58602ea7ff1eSPeter Xu return hugetlb_handle_userfault(vma, mapping, idx, flags, 58612ea7ff1eSPeter Xu haddr, address, 58627677f7fdSAxel Rasmussen VM_UFFD_MINOR); 58637677f7fdSAxel Rasmussen } 5864998b4382SNaoya Horiguchi } 58651e8f889bSDavid Gibson 586657303d80SAndy Whitcroft /* 586757303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 586857303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 586957303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 587057303d80SAndy Whitcroft * the spinlock. 587157303d80SAndy Whitcroft */ 58725e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 5873285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 58742b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 58752b26736cSAndy Whitcroft goto backout_unlocked; 58762b26736cSAndy Whitcroft } 58775e911373SMike Kravetz /* Just decrements count, does not deallocate */ 5878285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 58795e911373SMike Kravetz } 588057303d80SAndy Whitcroft 58818bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 588283c54070SNick Piggin ret = 0; 5883c64e912cSPeter Xu /* If pte changed from under us, retry */ 5884c64e912cSPeter Xu if (!pte_same(huge_ptep_get(ptep), old_pte)) 58854c887265SAdam Litke goto backout; 58864c887265SAdam Litke 58874781593dSPeter Xu if (anon_rmap) 5888285b8dcaSHuang Ying hugepage_add_new_anon_rmap(page, vma, haddr); 58894781593dSPeter Xu else 5890fb3d824dSDavid Hildenbrand page_dup_file_rmap(page, true); 58911e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 58921e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 5893c64e912cSPeter Xu /* 5894c64e912cSPeter Xu * If this pte was previously wr-protected, keep it wr-protected even 5895c64e912cSPeter Xu * if populated. 5896c64e912cSPeter Xu */ 5897c64e912cSPeter Xu if (unlikely(pte_marker_uffd_wp(old_pte))) 5898c64e912cSPeter Xu new_pte = huge_pte_wrprotect(huge_pte_mkuffd_wp(new_pte)); 5899285b8dcaSHuang Ying set_huge_pte_at(mm, haddr, ptep, new_pte); 59001e8f889bSDavid Gibson 59015d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 5902788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 59031e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 5904c89357e2SDavid Hildenbrand ret = hugetlb_wp(mm, vma, address, ptep, flags, page, ptl); 59051e8f889bSDavid Gibson } 59061e8f889bSDavid Gibson 5907cb900f41SKirill A. Shutemov spin_unlock(ptl); 5908cb6acd01SMike Kravetz 5909cb6acd01SMike Kravetz /* 59108f251a3dSMike Kravetz * Only set HPageMigratable in newly allocated pages. Existing pages 59118f251a3dSMike Kravetz * found in the pagecache may not have HPageMigratableset if they have 59128f251a3dSMike Kravetz * been isolated for migration. 5913cb6acd01SMike Kravetz */ 5914cb6acd01SMike Kravetz if (new_page) 59158f251a3dSMike Kravetz SetHPageMigratable(page); 5916cb6acd01SMike Kravetz 59174c887265SAdam Litke unlock_page(page); 59184c887265SAdam Litke out: 5919958f32ceSLiu Shixin hugetlb_vma_unlock_read(vma); 5920958f32ceSLiu Shixin mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5921ac9b9c66SHugh Dickins return ret; 59224c887265SAdam Litke 59234c887265SAdam Litke backout: 5924cb900f41SKirill A. Shutemov spin_unlock(ptl); 59252b26736cSAndy Whitcroft backout_unlocked: 5926c7b1850dSMike Kravetz if (new_page && !new_pagecache_page) 5927285b8dcaSHuang Ying restore_reserve_on_error(h, vma, haddr, page); 5928fa27759aSMike Kravetz 5929fa27759aSMike Kravetz unlock_page(page); 59304c887265SAdam Litke put_page(page); 59314c887265SAdam Litke goto out; 5932ac9b9c66SHugh Dickins } 5933ac9b9c66SHugh Dickins 59348382d914SDavidlohr Bueso #ifdef CONFIG_SMP 5935188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 59368382d914SDavidlohr Bueso { 59378382d914SDavidlohr Bueso unsigned long key[2]; 59388382d914SDavidlohr Bueso u32 hash; 59398382d914SDavidlohr Bueso 59408382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 59418382d914SDavidlohr Bueso key[1] = idx; 59428382d914SDavidlohr Bueso 594355254636SMike Kravetz hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); 59448382d914SDavidlohr Bueso 59458382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 59468382d914SDavidlohr Bueso } 59478382d914SDavidlohr Bueso #else 59488382d914SDavidlohr Bueso /* 59496c26d310SMiaohe Lin * For uniprocessor systems we always use a single mutex, so just 59508382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 59518382d914SDavidlohr Bueso */ 5952188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 59538382d914SDavidlohr Bueso { 59548382d914SDavidlohr Bueso return 0; 59558382d914SDavidlohr Bueso } 59568382d914SDavidlohr Bueso #endif 59578382d914SDavidlohr Bueso 59582b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 5959788c7df4SHugh Dickins unsigned long address, unsigned int flags) 596086e5216fSAdam Litke { 59618382d914SDavidlohr Bueso pte_t *ptep, entry; 5962cb900f41SKirill A. Shutemov spinlock_t *ptl; 59632b740303SSouptick Joarder vm_fault_t ret; 59648382d914SDavidlohr Bueso u32 hash; 59658382d914SDavidlohr Bueso pgoff_t idx; 59660fe6e20bSNaoya Horiguchi struct page *page = NULL; 596757303d80SAndy Whitcroft struct page *pagecache_page = NULL; 5968a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 59698382d914SDavidlohr Bueso struct address_space *mapping; 59700f792cf9SNaoya Horiguchi int need_wait_lock = 0; 5971285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 597286e5216fSAdam Litke 5973285b8dcaSHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 5974fd6a03edSNaoya Horiguchi if (ptep) { 597540549ba8SMike Kravetz /* 597640549ba8SMike Kravetz * Since we hold no locks, ptep could be stale. That is 597740549ba8SMike Kravetz * OK as we are only making decisions based on content and 597840549ba8SMike Kravetz * not actually modifying content here. 597940549ba8SMike Kravetz */ 5980fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 5981290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 5982ad1ac596SMiaohe Lin migration_entry_wait_huge(vma, ptep); 5983290408d4SNaoya Horiguchi return 0; 5984290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 5985aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 5986972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 5987c0d0381aSMike Kravetz } 5988ddeaab32SMike Kravetz 59893935baa9SDavid Gibson /* 59903935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 59913935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 59923935baa9SDavid Gibson * the same page in the page cache. 59933935baa9SDavid Gibson */ 599440549ba8SMike Kravetz mapping = vma->vm_file->f_mapping; 599540549ba8SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 5996188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 5997c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 59988382d914SDavidlohr Bueso 599940549ba8SMike Kravetz /* 600040549ba8SMike Kravetz * Acquire vma lock before calling huge_pte_alloc and hold 600140549ba8SMike Kravetz * until finished with ptep. This prevents huge_pmd_unshare from 600240549ba8SMike Kravetz * being called elsewhere and making the ptep no longer valid. 600340549ba8SMike Kravetz * 600440549ba8SMike Kravetz * ptep could have already be assigned via huge_pte_offset. That 600540549ba8SMike Kravetz * is OK, as huge_pte_alloc will return the same value unless 600640549ba8SMike Kravetz * something has changed. 600740549ba8SMike Kravetz */ 600840549ba8SMike Kravetz hugetlb_vma_lock_read(vma); 600940549ba8SMike Kravetz ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); 601040549ba8SMike Kravetz if (!ptep) { 601140549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 601240549ba8SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 601340549ba8SMike Kravetz return VM_FAULT_OOM; 601440549ba8SMike Kravetz } 601540549ba8SMike Kravetz 60167f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 6017c64e912cSPeter Xu /* PTE markers should be handled the same way as none pte */ 6018958f32ceSLiu Shixin if (huge_pte_none_mostly(entry)) 6019958f32ceSLiu Shixin /* 6020958f32ceSLiu Shixin * hugetlb_no_page will drop vma lock and hugetlb fault 6021958f32ceSLiu Shixin * mutex internally, which make us return immediately. 6022958f32ceSLiu Shixin */ 6023958f32ceSLiu Shixin return hugetlb_no_page(mm, vma, mapping, idx, address, ptep, 6024c64e912cSPeter Xu entry, flags); 602586e5216fSAdam Litke 602683c54070SNick Piggin ret = 0; 60271e8f889bSDavid Gibson 602857303d80SAndy Whitcroft /* 60290f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 60300f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 60317c8de358SEthon Paul * an active hugepage in pagecache. This goto expects the 2nd page 60327c8de358SEthon Paul * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will 60337c8de358SEthon Paul * properly handle it. 60340f792cf9SNaoya Horiguchi */ 60350f792cf9SNaoya Horiguchi if (!pte_present(entry)) 60360f792cf9SNaoya Horiguchi goto out_mutex; 60370f792cf9SNaoya Horiguchi 60380f792cf9SNaoya Horiguchi /* 6039c89357e2SDavid Hildenbrand * If we are going to COW/unshare the mapping later, we examine the 6040c89357e2SDavid Hildenbrand * pending reservations for this page now. This will ensure that any 604157303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 60421d8d1464SDavid Hildenbrand * spinlock. Also lookup the pagecache page now as it is used to 60431d8d1464SDavid Hildenbrand * determine if a reservation has been consumed. 604457303d80SAndy Whitcroft */ 6045c89357e2SDavid Hildenbrand if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) && 60461d8d1464SDavid Hildenbrand !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(entry)) { 6047285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 60482b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 6049b4d1d99fSDavid Gibson goto out_mutex; 60502b26736cSAndy Whitcroft } 60515e911373SMike Kravetz /* Just decrements count, does not deallocate */ 6052285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 605357303d80SAndy Whitcroft 605429be8426SMiaohe Lin pagecache_page = find_lock_page(mapping, idx); 605557303d80SAndy Whitcroft } 605657303d80SAndy Whitcroft 60570f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 60580fe6e20bSNaoya Horiguchi 6059c89357e2SDavid Hildenbrand /* Check for a racing update before calling hugetlb_wp() */ 6060b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 6061cb900f41SKirill A. Shutemov goto out_ptl; 6062b4d1d99fSDavid Gibson 6063166f3eccSPeter Xu /* Handle userfault-wp first, before trying to lock more pages */ 6064166f3eccSPeter Xu if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) && 6065166f3eccSPeter Xu (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 6066166f3eccSPeter Xu struct vm_fault vmf = { 6067166f3eccSPeter Xu .vma = vma, 6068166f3eccSPeter Xu .address = haddr, 6069166f3eccSPeter Xu .real_address = address, 6070166f3eccSPeter Xu .flags = flags, 6071166f3eccSPeter Xu }; 6072166f3eccSPeter Xu 6073166f3eccSPeter Xu spin_unlock(ptl); 6074166f3eccSPeter Xu if (pagecache_page) { 6075166f3eccSPeter Xu unlock_page(pagecache_page); 6076166f3eccSPeter Xu put_page(pagecache_page); 6077166f3eccSPeter Xu } 607840549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 6079166f3eccSPeter Xu mutex_unlock(&hugetlb_fault_mutex_table[hash]); 6080166f3eccSPeter Xu return handle_userfault(&vmf, VM_UFFD_WP); 6081166f3eccSPeter Xu } 6082166f3eccSPeter Xu 60830f792cf9SNaoya Horiguchi /* 6084c89357e2SDavid Hildenbrand * hugetlb_wp() requires page locks of pte_page(entry) and 60850f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 60860f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 60870f792cf9SNaoya Horiguchi */ 60880f792cf9SNaoya Horiguchi page = pte_page(entry); 60890f792cf9SNaoya Horiguchi if (page != pagecache_page) 60900f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 60910f792cf9SNaoya Horiguchi need_wait_lock = 1; 60920f792cf9SNaoya Horiguchi goto out_ptl; 60930f792cf9SNaoya Horiguchi } 60940f792cf9SNaoya Horiguchi 60950f792cf9SNaoya Horiguchi get_page(page); 6096b4d1d99fSDavid Gibson 6097c89357e2SDavid Hildenbrand if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) { 6098106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 6099c89357e2SDavid Hildenbrand ret = hugetlb_wp(mm, vma, address, ptep, flags, 6100cb900f41SKirill A. Shutemov pagecache_page, ptl); 61010f792cf9SNaoya Horiguchi goto out_put_page; 6102c89357e2SDavid Hildenbrand } else if (likely(flags & FAULT_FLAG_WRITE)) { 6103106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 6104b4d1d99fSDavid Gibson } 6105c89357e2SDavid Hildenbrand } 6106b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 6107285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 6108788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 6109285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 61100f792cf9SNaoya Horiguchi out_put_page: 61110f792cf9SNaoya Horiguchi if (page != pagecache_page) 61120f792cf9SNaoya Horiguchi unlock_page(page); 61130f792cf9SNaoya Horiguchi put_page(page); 6114cb900f41SKirill A. Shutemov out_ptl: 6115cb900f41SKirill A. Shutemov spin_unlock(ptl); 611657303d80SAndy Whitcroft 611757303d80SAndy Whitcroft if (pagecache_page) { 611857303d80SAndy Whitcroft unlock_page(pagecache_page); 611957303d80SAndy Whitcroft put_page(pagecache_page); 612057303d80SAndy Whitcroft } 6121b4d1d99fSDavid Gibson out_mutex: 612240549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 6123c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 61240f792cf9SNaoya Horiguchi /* 61250f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 61260f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 61270f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 61280f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 61290f792cf9SNaoya Horiguchi * here without taking refcount. 61300f792cf9SNaoya Horiguchi */ 61310f792cf9SNaoya Horiguchi if (need_wait_lock) 61320f792cf9SNaoya Horiguchi wait_on_page_locked(page); 61331e8f889bSDavid Gibson return ret; 613486e5216fSAdam Litke } 613586e5216fSAdam Litke 6136714c1891SAxel Rasmussen #ifdef CONFIG_USERFAULTFD 61378fb5debcSMike Kravetz /* 61388fb5debcSMike Kravetz * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with 61398fb5debcSMike Kravetz * modifications for huge pages. 61408fb5debcSMike Kravetz */ 61418fb5debcSMike Kravetz int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, 61428fb5debcSMike Kravetz pte_t *dst_pte, 61438fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 61448fb5debcSMike Kravetz unsigned long dst_addr, 61458fb5debcSMike Kravetz unsigned long src_addr, 6146f6191471SAxel Rasmussen enum mcopy_atomic_mode mode, 61476041c691SPeter Xu struct page **pagep, 61486041c691SPeter Xu bool wp_copy) 61498fb5debcSMike Kravetz { 6150f6191471SAxel Rasmussen bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE); 61518cc5fcbbSMina Almasry struct hstate *h = hstate_vma(dst_vma); 61528cc5fcbbSMina Almasry struct address_space *mapping = dst_vma->vm_file->f_mapping; 61538cc5fcbbSMina Almasry pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr); 61541e392147SAndrea Arcangeli unsigned long size; 61551c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 61568fb5debcSMike Kravetz pte_t _dst_pte; 61578fb5debcSMike Kravetz spinlock_t *ptl; 61588cc5fcbbSMina Almasry int ret = -ENOMEM; 61598fb5debcSMike Kravetz struct page *page; 6160f6191471SAxel Rasmussen int writable; 6161cc30042dSMina Almasry bool page_in_pagecache = false; 61628fb5debcSMike Kravetz 6163f6191471SAxel Rasmussen if (is_continue) { 6164f6191471SAxel Rasmussen ret = -EFAULT; 6165f6191471SAxel Rasmussen page = find_lock_page(mapping, idx); 6166f6191471SAxel Rasmussen if (!page) 6167f6191471SAxel Rasmussen goto out; 6168cc30042dSMina Almasry page_in_pagecache = true; 6169f6191471SAxel Rasmussen } else if (!*pagep) { 6170d84cf06eSMina Almasry /* If a page already exists, then it's UFFDIO_COPY for 6171d84cf06eSMina Almasry * a non-missing case. Return -EEXIST. 6172d84cf06eSMina Almasry */ 6173d84cf06eSMina Almasry if (vm_shared && 6174d84cf06eSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 6175d84cf06eSMina Almasry ret = -EEXIST; 61768fb5debcSMike Kravetz goto out; 6177d84cf06eSMina Almasry } 6178d84cf06eSMina Almasry 6179d84cf06eSMina Almasry page = alloc_huge_page(dst_vma, dst_addr, 0); 6180d84cf06eSMina Almasry if (IS_ERR(page)) { 6181d84cf06eSMina Almasry ret = -ENOMEM; 6182d84cf06eSMina Almasry goto out; 6183d84cf06eSMina Almasry } 61848fb5debcSMike Kravetz 61858fb5debcSMike Kravetz ret = copy_huge_page_from_user(page, 61868fb5debcSMike Kravetz (const void __user *) src_addr, 6187810a56b9SMike Kravetz pages_per_huge_page(h), false); 61888fb5debcSMike Kravetz 6189c1e8d7c6SMichel Lespinasse /* fallback to copy_from_user outside mmap_lock */ 61908fb5debcSMike Kravetz if (unlikely(ret)) { 61919e368259SAndrea Arcangeli ret = -ENOENT; 61928cc5fcbbSMina Almasry /* Free the allocated page which may have 61938cc5fcbbSMina Almasry * consumed a reservation. 61948cc5fcbbSMina Almasry */ 61958cc5fcbbSMina Almasry restore_reserve_on_error(h, dst_vma, dst_addr, page); 61968cc5fcbbSMina Almasry put_page(page); 61978cc5fcbbSMina Almasry 61988cc5fcbbSMina Almasry /* Allocate a temporary page to hold the copied 61998cc5fcbbSMina Almasry * contents. 62008cc5fcbbSMina Almasry */ 62018cc5fcbbSMina Almasry page = alloc_huge_page_vma(h, dst_vma, dst_addr); 62028cc5fcbbSMina Almasry if (!page) { 62038cc5fcbbSMina Almasry ret = -ENOMEM; 62048cc5fcbbSMina Almasry goto out; 62058cc5fcbbSMina Almasry } 62068fb5debcSMike Kravetz *pagep = page; 62078cc5fcbbSMina Almasry /* Set the outparam pagep and return to the caller to 62088cc5fcbbSMina Almasry * copy the contents outside the lock. Don't free the 62098cc5fcbbSMina Almasry * page. 62108cc5fcbbSMina Almasry */ 62118fb5debcSMike Kravetz goto out; 62128fb5debcSMike Kravetz } 62138fb5debcSMike Kravetz } else { 62148cc5fcbbSMina Almasry if (vm_shared && 62158cc5fcbbSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 62168cc5fcbbSMina Almasry put_page(*pagep); 62178cc5fcbbSMina Almasry ret = -EEXIST; 62188cc5fcbbSMina Almasry *pagep = NULL; 62198cc5fcbbSMina Almasry goto out; 62208cc5fcbbSMina Almasry } 62218cc5fcbbSMina Almasry 62228cc5fcbbSMina Almasry page = alloc_huge_page(dst_vma, dst_addr, 0); 62238cc5fcbbSMina Almasry if (IS_ERR(page)) { 6224da9a298fSMiaohe Lin put_page(*pagep); 62258cc5fcbbSMina Almasry ret = -ENOMEM; 62268cc5fcbbSMina Almasry *pagep = NULL; 62278cc5fcbbSMina Almasry goto out; 62288cc5fcbbSMina Almasry } 622934892366SMuchun Song copy_user_huge_page(page, *pagep, dst_addr, dst_vma, 623034892366SMuchun Song pages_per_huge_page(h)); 62318cc5fcbbSMina Almasry put_page(*pagep); 62328fb5debcSMike Kravetz *pagep = NULL; 62338fb5debcSMike Kravetz } 62348fb5debcSMike Kravetz 62358fb5debcSMike Kravetz /* 62368fb5debcSMike Kravetz * The memory barrier inside __SetPageUptodate makes sure that 62378fb5debcSMike Kravetz * preceding stores to the page contents become visible before 62388fb5debcSMike Kravetz * the set_pte_at() write. 62398fb5debcSMike Kravetz */ 62408fb5debcSMike Kravetz __SetPageUptodate(page); 62418fb5debcSMike Kravetz 6242f6191471SAxel Rasmussen /* Add shared, newly allocated pages to the page cache. */ 6243f6191471SAxel Rasmussen if (vm_shared && !is_continue) { 62441e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 62451e392147SAndrea Arcangeli ret = -EFAULT; 62461e392147SAndrea Arcangeli if (idx >= size) 62471e392147SAndrea Arcangeli goto out_release_nounlock; 62481c9e8defSMike Kravetz 62491e392147SAndrea Arcangeli /* 62501e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 62517e1813d4SMike Kravetz * hugetlb_add_to_page_cache() below happens through the 62521e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 62531e392147SAndrea Arcangeli * the caller. 62541e392147SAndrea Arcangeli */ 62557e1813d4SMike Kravetz ret = hugetlb_add_to_page_cache(page, mapping, idx); 62561c9e8defSMike Kravetz if (ret) 62571c9e8defSMike Kravetz goto out_release_nounlock; 6258cc30042dSMina Almasry page_in_pagecache = true; 62591c9e8defSMike Kravetz } 62601c9e8defSMike Kravetz 6261bcc66543SMiaohe Lin ptl = huge_pte_lock(h, dst_mm, dst_pte); 62628fb5debcSMike Kravetz 62638625147cSJames Houghton ret = -EIO; 62648625147cSJames Houghton if (PageHWPoison(page)) 62658625147cSJames Houghton goto out_release_unlock; 62668625147cSJames Houghton 62671e392147SAndrea Arcangeli /* 62686041c691SPeter Xu * We allow to overwrite a pte marker: consider when both MISSING|WP 62696041c691SPeter Xu * registered, we firstly wr-protect a none pte which has no page cache 62706041c691SPeter Xu * page backing it, then access the page. 62716041c691SPeter Xu */ 6272fa27759aSMike Kravetz ret = -EEXIST; 62736041c691SPeter Xu if (!huge_pte_none_mostly(huge_ptep_get(dst_pte))) 62748fb5debcSMike Kravetz goto out_release_unlock; 62758fb5debcSMike Kravetz 62764781593dSPeter Xu if (page_in_pagecache) 6277fb3d824dSDavid Hildenbrand page_dup_file_rmap(page, true); 62784781593dSPeter Xu else 62798fb5debcSMike Kravetz hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); 62808fb5debcSMike Kravetz 62816041c691SPeter Xu /* 62826041c691SPeter Xu * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY 62836041c691SPeter Xu * with wp flag set, don't set pte write bit. 62846041c691SPeter Xu */ 62856041c691SPeter Xu if (wp_copy || (is_continue && !vm_shared)) 6286f6191471SAxel Rasmussen writable = 0; 6287f6191471SAxel Rasmussen else 6288f6191471SAxel Rasmussen writable = dst_vma->vm_flags & VM_WRITE; 6289f6191471SAxel Rasmussen 6290f6191471SAxel Rasmussen _dst_pte = make_huge_pte(dst_vma, page, writable); 62916041c691SPeter Xu /* 62926041c691SPeter Xu * Always mark UFFDIO_COPY page dirty; note that this may not be 62936041c691SPeter Xu * extremely important for hugetlbfs for now since swapping is not 62946041c691SPeter Xu * supported, but we should still be clear in that this page cannot be 62956041c691SPeter Xu * thrown away at will, even if write bit not set. 62966041c691SPeter Xu */ 62978fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 62988fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 62998fb5debcSMike Kravetz 63006041c691SPeter Xu if (wp_copy) 63016041c691SPeter Xu _dst_pte = huge_pte_mkuffd_wp(_dst_pte); 63026041c691SPeter Xu 63038fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 63048fb5debcSMike Kravetz 63058fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 63068fb5debcSMike Kravetz 63078fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 63088fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 63098fb5debcSMike Kravetz 63108fb5debcSMike Kravetz spin_unlock(ptl); 6311f6191471SAxel Rasmussen if (!is_continue) 63128f251a3dSMike Kravetz SetHPageMigratable(page); 6313f6191471SAxel Rasmussen if (vm_shared || is_continue) 63141c9e8defSMike Kravetz unlock_page(page); 63158fb5debcSMike Kravetz ret = 0; 63168fb5debcSMike Kravetz out: 63178fb5debcSMike Kravetz return ret; 63188fb5debcSMike Kravetz out_release_unlock: 63198fb5debcSMike Kravetz spin_unlock(ptl); 6320f6191471SAxel Rasmussen if (vm_shared || is_continue) 63211c9e8defSMike Kravetz unlock_page(page); 63225af10dfdSAndrea Arcangeli out_release_nounlock: 6323cc30042dSMina Almasry if (!page_in_pagecache) 6324846be085SMike Kravetz restore_reserve_on_error(h, dst_vma, dst_addr, page); 63258fb5debcSMike Kravetz put_page(page); 63268fb5debcSMike Kravetz goto out; 63278fb5debcSMike Kravetz } 6328714c1891SAxel Rasmussen #endif /* CONFIG_USERFAULTFD */ 63298fb5debcSMike Kravetz 633082e5d378SJoao Martins static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma, 633182e5d378SJoao Martins int refs, struct page **pages, 633282e5d378SJoao Martins struct vm_area_struct **vmas) 633382e5d378SJoao Martins { 633482e5d378SJoao Martins int nr; 633582e5d378SJoao Martins 633682e5d378SJoao Martins for (nr = 0; nr < refs; nr++) { 633782e5d378SJoao Martins if (likely(pages)) 633814455eabSCheng Li pages[nr] = nth_page(page, nr); 633982e5d378SJoao Martins if (vmas) 634082e5d378SJoao Martins vmas[nr] = vma; 634182e5d378SJoao Martins } 634282e5d378SJoao Martins } 634382e5d378SJoao Martins 634484209e87SDavid Hildenbrand static inline bool __follow_hugetlb_must_fault(struct vm_area_struct *vma, 634584209e87SDavid Hildenbrand unsigned int flags, pte_t *pte, 6346a7f22660SDavid Hildenbrand bool *unshare) 6347a7f22660SDavid Hildenbrand { 6348a7f22660SDavid Hildenbrand pte_t pteval = huge_ptep_get(pte); 6349a7f22660SDavid Hildenbrand 6350a7f22660SDavid Hildenbrand *unshare = false; 6351a7f22660SDavid Hildenbrand if (is_swap_pte(pteval)) 6352a7f22660SDavid Hildenbrand return true; 6353a7f22660SDavid Hildenbrand if (huge_pte_write(pteval)) 6354a7f22660SDavid Hildenbrand return false; 6355a7f22660SDavid Hildenbrand if (flags & FOLL_WRITE) 6356a7f22660SDavid Hildenbrand return true; 635784209e87SDavid Hildenbrand if (gup_must_unshare(vma, flags, pte_page(pteval))) { 6358a7f22660SDavid Hildenbrand *unshare = true; 6359a7f22660SDavid Hildenbrand return true; 6360a7f22660SDavid Hildenbrand } 6361a7f22660SDavid Hildenbrand return false; 6362a7f22660SDavid Hildenbrand } 6363a7f22660SDavid Hildenbrand 636457a196a5SMike Kravetz struct page *hugetlb_follow_page_mask(struct vm_area_struct *vma, 636557a196a5SMike Kravetz unsigned long address, unsigned int flags) 636657a196a5SMike Kravetz { 636757a196a5SMike Kravetz struct hstate *h = hstate_vma(vma); 636857a196a5SMike Kravetz struct mm_struct *mm = vma->vm_mm; 636957a196a5SMike Kravetz unsigned long haddr = address & huge_page_mask(h); 637057a196a5SMike Kravetz struct page *page = NULL; 637157a196a5SMike Kravetz spinlock_t *ptl; 637257a196a5SMike Kravetz pte_t *pte, entry; 637357a196a5SMike Kravetz 637457a196a5SMike Kravetz /* 637557a196a5SMike Kravetz * FOLL_PIN is not supported for follow_page(). Ordinary GUP goes via 637657a196a5SMike Kravetz * follow_hugetlb_page(). 637757a196a5SMike Kravetz */ 637857a196a5SMike Kravetz if (WARN_ON_ONCE(flags & FOLL_PIN)) 637957a196a5SMike Kravetz return NULL; 638057a196a5SMike Kravetz 638157a196a5SMike Kravetz retry: 638257a196a5SMike Kravetz pte = huge_pte_offset(mm, haddr, huge_page_size(h)); 638357a196a5SMike Kravetz if (!pte) 638457a196a5SMike Kravetz return NULL; 638557a196a5SMike Kravetz 638657a196a5SMike Kravetz ptl = huge_pte_lock(h, mm, pte); 638757a196a5SMike Kravetz entry = huge_ptep_get(pte); 638857a196a5SMike Kravetz if (pte_present(entry)) { 638957a196a5SMike Kravetz page = pte_page(entry) + 639057a196a5SMike Kravetz ((address & ~huge_page_mask(h)) >> PAGE_SHIFT); 639157a196a5SMike Kravetz /* 639257a196a5SMike Kravetz * Note that page may be a sub-page, and with vmemmap 639357a196a5SMike Kravetz * optimizations the page struct may be read only. 639457a196a5SMike Kravetz * try_grab_page() will increase the ref count on the 639557a196a5SMike Kravetz * head page, so this will be OK. 639657a196a5SMike Kravetz * 6397e2ca6ba6SLinus Torvalds * try_grab_page() should always be able to get the page here, 6398e2ca6ba6SLinus Torvalds * because we hold the ptl lock and have verified pte_present(). 639957a196a5SMike Kravetz */ 6400e2ca6ba6SLinus Torvalds if (try_grab_page(page, flags)) { 640157a196a5SMike Kravetz page = NULL; 640257a196a5SMike Kravetz goto out; 640357a196a5SMike Kravetz } 640457a196a5SMike Kravetz } else { 640557a196a5SMike Kravetz if (is_hugetlb_entry_migration(entry)) { 640657a196a5SMike Kravetz spin_unlock(ptl); 640757a196a5SMike Kravetz __migration_entry_wait_huge(pte, ptl); 640857a196a5SMike Kravetz goto retry; 640957a196a5SMike Kravetz } 641057a196a5SMike Kravetz /* 641157a196a5SMike Kravetz * hwpoisoned entry is treated as no_page_table in 641257a196a5SMike Kravetz * follow_page_mask(). 641357a196a5SMike Kravetz */ 641457a196a5SMike Kravetz } 641557a196a5SMike Kravetz out: 641657a196a5SMike Kravetz spin_unlock(ptl); 641757a196a5SMike Kravetz return page; 641857a196a5SMike Kravetz } 641957a196a5SMike Kravetz 642028a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 642163551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 642228a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 64234f6da934SPeter Xu long i, unsigned int flags, int *locked) 642463551ae0SDavid Gibson { 6425d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 6426d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 642728a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 6428a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 64290fa5bc40SJoao Martins int err = -EFAULT, refs; 643063551ae0SDavid Gibson 643163551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 643263551ae0SDavid Gibson pte_t *pte; 6433cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 6434a7f22660SDavid Hildenbrand bool unshare = false; 64352a15efc9SHugh Dickins int absent; 643663551ae0SDavid Gibson struct page *page; 643763551ae0SDavid Gibson 64384c887265SAdam Litke /* 643902057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 644002057967SDavid Rientjes * potentially allocating memory. 644102057967SDavid Rientjes */ 6442fa45f116SDavidlohr Bueso if (fatal_signal_pending(current)) { 644302057967SDavid Rientjes remainder = 0; 644402057967SDavid Rientjes break; 644502057967SDavid Rientjes } 644602057967SDavid Rientjes 644702057967SDavid Rientjes /* 64484c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 64492a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 64504c887265SAdam Litke * first, for the page indexing below to work. 6451cb900f41SKirill A. Shutemov * 6452cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 64534c887265SAdam Litke */ 64547868a208SPunit Agrawal pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), 64557868a208SPunit Agrawal huge_page_size(h)); 6456cb900f41SKirill A. Shutemov if (pte) 6457cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 64582a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 645963551ae0SDavid Gibson 64602a15efc9SHugh Dickins /* 64612a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 64623ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 64633ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 64643ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 64653ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 64662a15efc9SHugh Dickins */ 64673ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 64683ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 6469cb900f41SKirill A. Shutemov if (pte) 6470cb900f41SKirill A. Shutemov spin_unlock(ptl); 64712a15efc9SHugh Dickins remainder = 0; 64722a15efc9SHugh Dickins break; 64732a15efc9SHugh Dickins } 64742a15efc9SHugh Dickins 64759cc3a5bdSNaoya Horiguchi /* 64769cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 64779cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 64789cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 64799cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 64809cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 64819cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 64829cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 64839cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 64849cc3a5bdSNaoya Horiguchi */ 6485a7f22660SDavid Hildenbrand if (absent || 648684209e87SDavid Hildenbrand __follow_hugetlb_must_fault(vma, flags, pte, &unshare)) { 64872b740303SSouptick Joarder vm_fault_t ret; 648887ffc118SAndrea Arcangeli unsigned int fault_flags = 0; 64894c887265SAdam Litke 6490cb900f41SKirill A. Shutemov if (pte) 6491cb900f41SKirill A. Shutemov spin_unlock(ptl); 649287ffc118SAndrea Arcangeli if (flags & FOLL_WRITE) 649387ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_WRITE; 6494a7f22660SDavid Hildenbrand else if (unshare) 6495a7f22660SDavid Hildenbrand fault_flags |= FAULT_FLAG_UNSHARE; 649693c5c61dSPeter Xu if (locked) { 649771335f37SPeter Xu fault_flags |= FAULT_FLAG_ALLOW_RETRY | 649871335f37SPeter Xu FAULT_FLAG_KILLABLE; 649993c5c61dSPeter Xu if (flags & FOLL_INTERRUPTIBLE) 650093c5c61dSPeter Xu fault_flags |= FAULT_FLAG_INTERRUPTIBLE; 650193c5c61dSPeter Xu } 650287ffc118SAndrea Arcangeli if (flags & FOLL_NOWAIT) 650387ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY | 650487ffc118SAndrea Arcangeli FAULT_FLAG_RETRY_NOWAIT; 650587ffc118SAndrea Arcangeli if (flags & FOLL_TRIED) { 65064426e945SPeter Xu /* 65074426e945SPeter Xu * Note: FAULT_FLAG_ALLOW_RETRY and 65084426e945SPeter Xu * FAULT_FLAG_TRIED can co-exist 65094426e945SPeter Xu */ 651087ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_TRIED; 651187ffc118SAndrea Arcangeli } 651287ffc118SAndrea Arcangeli ret = hugetlb_fault(mm, vma, vaddr, fault_flags); 651387ffc118SAndrea Arcangeli if (ret & VM_FAULT_ERROR) { 65142be7cfedSDaniel Jordan err = vm_fault_to_errno(ret, flags); 65151c59827dSHugh Dickins remainder = 0; 65161c59827dSHugh Dickins break; 65171c59827dSHugh Dickins } 651887ffc118SAndrea Arcangeli if (ret & VM_FAULT_RETRY) { 65194f6da934SPeter Xu if (locked && 65201ac25013SAndrea Arcangeli !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) 65214f6da934SPeter Xu *locked = 0; 652287ffc118SAndrea Arcangeli *nr_pages = 0; 652387ffc118SAndrea Arcangeli /* 652487ffc118SAndrea Arcangeli * VM_FAULT_RETRY must not return an 652587ffc118SAndrea Arcangeli * error, it will return zero 652687ffc118SAndrea Arcangeli * instead. 652787ffc118SAndrea Arcangeli * 652887ffc118SAndrea Arcangeli * No need to update "position" as the 652987ffc118SAndrea Arcangeli * caller will not check it after 653087ffc118SAndrea Arcangeli * *nr_pages is set to 0. 653187ffc118SAndrea Arcangeli */ 653287ffc118SAndrea Arcangeli return i; 653387ffc118SAndrea Arcangeli } 653487ffc118SAndrea Arcangeli continue; 653587ffc118SAndrea Arcangeli } 653663551ae0SDavid Gibson 6537a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 65387f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 65398fde12caSLinus Torvalds 6540b6a2619cSDavid Hildenbrand VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) && 6541b6a2619cSDavid Hildenbrand !PageAnonExclusive(page), page); 6542b6a2619cSDavid Hildenbrand 65438fde12caSLinus Torvalds /* 6544acbfb087SZhigang Lu * If subpage information not requested, update counters 6545acbfb087SZhigang Lu * and skip the same_page loop below. 6546acbfb087SZhigang Lu */ 6547acbfb087SZhigang Lu if (!pages && !vmas && !pfn_offset && 6548acbfb087SZhigang Lu (vaddr + huge_page_size(h) < vma->vm_end) && 6549acbfb087SZhigang Lu (remainder >= pages_per_huge_page(h))) { 6550acbfb087SZhigang Lu vaddr += huge_page_size(h); 6551acbfb087SZhigang Lu remainder -= pages_per_huge_page(h); 6552acbfb087SZhigang Lu i += pages_per_huge_page(h); 6553acbfb087SZhigang Lu spin_unlock(ptl); 6554acbfb087SZhigang Lu continue; 6555acbfb087SZhigang Lu } 6556acbfb087SZhigang Lu 6557d08af0a5SJoao Martins /* vaddr may not be aligned to PAGE_SIZE */ 6558d08af0a5SJoao Martins refs = min3(pages_per_huge_page(h) - pfn_offset, remainder, 6559d08af0a5SJoao Martins (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT); 65600fa5bc40SJoao Martins 656182e5d378SJoao Martins if (pages || vmas) 656214455eabSCheng Li record_subpages_vmas(nth_page(page, pfn_offset), 656382e5d378SJoao Martins vma, refs, 656482e5d378SJoao Martins likely(pages) ? pages + i : NULL, 656582e5d378SJoao Martins vmas ? vmas + i : NULL); 656663551ae0SDavid Gibson 656782e5d378SJoao Martins if (pages) { 65680fa5bc40SJoao Martins /* 6569822951d8SMatthew Wilcox (Oracle) * try_grab_folio() should always succeed here, 65700fa5bc40SJoao Martins * because: a) we hold the ptl lock, and b) we've just 65710fa5bc40SJoao Martins * checked that the huge page is present in the page 65720fa5bc40SJoao Martins * tables. If the huge page is present, then the tail 65730fa5bc40SJoao Martins * pages must also be present. The ptl prevents the 65740fa5bc40SJoao Martins * head page and tail pages from being rearranged in 65754003f107SLogan Gunthorpe * any way. As this is hugetlb, the pages will never 65764003f107SLogan Gunthorpe * be p2pdma or not longterm pinable. So this page 65774003f107SLogan Gunthorpe * must be available at this point, unless the page 65784003f107SLogan Gunthorpe * refcount overflowed: 65790fa5bc40SJoao Martins */ 6580822951d8SMatthew Wilcox (Oracle) if (WARN_ON_ONCE(!try_grab_folio(pages[i], refs, 65810fa5bc40SJoao Martins flags))) { 65820fa5bc40SJoao Martins spin_unlock(ptl); 65830fa5bc40SJoao Martins remainder = 0; 65840fa5bc40SJoao Martins err = -ENOMEM; 65850fa5bc40SJoao Martins break; 65860fa5bc40SJoao Martins } 6587d5d4b0aaSChen, Kenneth W } 658882e5d378SJoao Martins 658982e5d378SJoao Martins vaddr += (refs << PAGE_SHIFT); 659082e5d378SJoao Martins remainder -= refs; 659182e5d378SJoao Martins i += refs; 659282e5d378SJoao Martins 6593cb900f41SKirill A. Shutemov spin_unlock(ptl); 659463551ae0SDavid Gibson } 659528a35716SMichel Lespinasse *nr_pages = remainder; 659687ffc118SAndrea Arcangeli /* 659787ffc118SAndrea Arcangeli * setting position is actually required only if remainder is 659887ffc118SAndrea Arcangeli * not zero but it's faster not to add a "if (remainder)" 659987ffc118SAndrea Arcangeli * branch. 660087ffc118SAndrea Arcangeli */ 660163551ae0SDavid Gibson *position = vaddr; 660263551ae0SDavid Gibson 66032be7cfedSDaniel Jordan return i ? i : err; 660463551ae0SDavid Gibson } 66058f860591SZhang, Yanmin 66067da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 66075a90d5a1SPeter Xu unsigned long address, unsigned long end, 66085a90d5a1SPeter Xu pgprot_t newprot, unsigned long cp_flags) 66098f860591SZhang, Yanmin { 66108f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 66118f860591SZhang, Yanmin unsigned long start = address; 66128f860591SZhang, Yanmin pte_t *ptep; 66138f860591SZhang, Yanmin pte_t pte; 6614a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 661560dfaad6SPeter Xu unsigned long pages = 0, psize = huge_page_size(h); 6616dff11abeSMike Kravetz bool shared_pmd = false; 6617ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 6618e95a9851SMike Kravetz unsigned long last_addr_mask; 66195a90d5a1SPeter Xu bool uffd_wp = cp_flags & MM_CP_UFFD_WP; 66205a90d5a1SPeter Xu bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE; 6621dff11abeSMike Kravetz 6622dff11abeSMike Kravetz /* 6623dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 6624ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 6625dff11abeSMike Kravetz * range if PMD sharing is possible. 6626dff11abeSMike Kravetz */ 66277269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 66287269f999SJérôme Glisse 0, vma, mm, start, end); 6629ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 66308f860591SZhang, Yanmin 66318f860591SZhang, Yanmin BUG_ON(address >= end); 6632ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 66338f860591SZhang, Yanmin 6634ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 663540549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 663683cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 663740549ba8SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 663860dfaad6SPeter Xu for (; address < end; address += psize) { 6639cb900f41SKirill A. Shutemov spinlock_t *ptl; 664060dfaad6SPeter Xu ptep = huge_pte_offset(mm, address, psize); 6641e95a9851SMike Kravetz if (!ptep) { 6642e95a9851SMike Kravetz address |= last_addr_mask; 66438f860591SZhang, Yanmin continue; 6644e95a9851SMike Kravetz } 6645cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 66464ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, address, ptep)) { 664760dfaad6SPeter Xu /* 664860dfaad6SPeter Xu * When uffd-wp is enabled on the vma, unshare 664960dfaad6SPeter Xu * shouldn't happen at all. Warn about it if it 665060dfaad6SPeter Xu * happened due to some reason. 665160dfaad6SPeter Xu */ 665260dfaad6SPeter Xu WARN_ON_ONCE(uffd_wp || uffd_wp_resolve); 66537da4d641SPeter Zijlstra pages++; 6654cb900f41SKirill A. Shutemov spin_unlock(ptl); 6655dff11abeSMike Kravetz shared_pmd = true; 66564ddb4d91SMike Kravetz address |= last_addr_mask; 665739dde65cSChen, Kenneth W continue; 66587da4d641SPeter Zijlstra } 6659a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 6660a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 66610e678153SDavid Hildenbrand /* Nothing to do. */ 66620e678153SDavid Hildenbrand } else if (unlikely(is_hugetlb_entry_migration(pte))) { 6663a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 66646c287605SDavid Hildenbrand struct page *page = pfn_swap_entry_to_page(entry); 6665*44f86392SDavid Hildenbrand pte_t newpte = pte; 6666a8bda28dSNaoya Horiguchi 6667*44f86392SDavid Hildenbrand if (is_writable_migration_entry(entry)) { 66686c287605SDavid Hildenbrand if (PageAnon(page)) 66696c287605SDavid Hildenbrand entry = make_readable_exclusive_migration_entry( 66706c287605SDavid Hildenbrand swp_offset(entry)); 66716c287605SDavid Hildenbrand else 66724dd845b5SAlistair Popple entry = make_readable_migration_entry( 66734dd845b5SAlistair Popple swp_offset(entry)); 6674a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 6675*44f86392SDavid Hildenbrand pages++; 6676*44f86392SDavid Hildenbrand } 6677*44f86392SDavid Hildenbrand 66785a90d5a1SPeter Xu if (uffd_wp) 66795a90d5a1SPeter Xu newpte = pte_swp_mkuffd_wp(newpte); 66805a90d5a1SPeter Xu else if (uffd_wp_resolve) 66815a90d5a1SPeter Xu newpte = pte_swp_clear_uffd_wp(newpte); 6682*44f86392SDavid Hildenbrand if (!pte_same(pte, newpte)) 668318f39629SQi Zheng set_huge_pte_at(mm, address, ptep, newpte); 66840e678153SDavid Hildenbrand } else if (unlikely(is_pte_marker(pte))) { 66850e678153SDavid Hildenbrand /* No other markers apply for now. */ 66860e678153SDavid Hildenbrand WARN_ON_ONCE(!pte_marker_uffd_wp(pte)); 668760dfaad6SPeter Xu if (uffd_wp_resolve) 66880e678153SDavid Hildenbrand /* Safe to modify directly (non-present->none). */ 668960dfaad6SPeter Xu huge_pte_clear(mm, address, ptep, psize); 66900e678153SDavid Hildenbrand } else if (!huge_pte_none(pte)) { 6691023bdd00SAneesh Kumar K.V pte_t old_pte; 669279c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 6693023bdd00SAneesh Kumar K.V 6694023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 669516785bd7SAnshuman Khandual pte = huge_pte_modify(old_pte, newprot); 669679c1c594SChristophe Leroy pte = arch_make_huge_pte(pte, shift, vma->vm_flags); 66975a90d5a1SPeter Xu if (uffd_wp) 66985a90d5a1SPeter Xu pte = huge_pte_mkuffd_wp(huge_pte_wrprotect(pte)); 66995a90d5a1SPeter Xu else if (uffd_wp_resolve) 67005a90d5a1SPeter Xu pte = huge_pte_clear_uffd_wp(pte); 6701023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 67027da4d641SPeter Zijlstra pages++; 670360dfaad6SPeter Xu } else { 670460dfaad6SPeter Xu /* None pte */ 670560dfaad6SPeter Xu if (unlikely(uffd_wp)) 670660dfaad6SPeter Xu /* Safe to modify directly (none->non-present). */ 670760dfaad6SPeter Xu set_huge_pte_at(mm, address, ptep, 670860dfaad6SPeter Xu make_pte_marker(PTE_MARKER_UFFD_WP)); 67098f860591SZhang, Yanmin } 6710cb900f41SKirill A. Shutemov spin_unlock(ptl); 67118f860591SZhang, Yanmin } 6712d833352aSMel Gorman /* 6713c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 6714d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 6715c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 6716dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 6717dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 6718d833352aSMel Gorman */ 6719dff11abeSMike Kravetz if (shared_pmd) 6720ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 6721dff11abeSMike Kravetz else 67225491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 67230f10851eSJérôme Glisse /* 67240f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 67250f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 67260f10851eSJérôme Glisse * 6727ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 67280f10851eSJérôme Glisse */ 672983cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 673040549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 6731ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 67327da4d641SPeter Zijlstra 67337da4d641SPeter Zijlstra return pages << h->order; 67348f860591SZhang, Yanmin } 67358f860591SZhang, Yanmin 673633b8f84aSMike Kravetz /* Return true if reservation was successful, false otherwise. */ 673733b8f84aSMike Kravetz bool hugetlb_reserve_pages(struct inode *inode, 6738a1e78772SMel Gorman long from, long to, 67395a6fe125SMel Gorman struct vm_area_struct *vma, 6740ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 6741e4e574b7SAdam Litke { 674233b8f84aSMike Kravetz long chg, add = -1; 6743a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 674490481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 67459119a41eSJoonsoo Kim struct resv_map *resv_map; 6746075a61d0SMina Almasry struct hugetlb_cgroup *h_cg = NULL; 67470db9d74eSMina Almasry long gbl_reserve, regions_needed = 0; 6748e4e574b7SAdam Litke 674963489f8eSMike Kravetz /* This should never happen */ 675063489f8eSMike Kravetz if (from > to) { 675163489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 675233b8f84aSMike Kravetz return false; 675363489f8eSMike Kravetz } 675463489f8eSMike Kravetz 6755a1e78772SMel Gorman /* 6756e700898fSMike Kravetz * vma specific semaphore used for pmd sharing and fault/truncation 6757e700898fSMike Kravetz * synchronization 67588d9bfb26SMike Kravetz */ 67598d9bfb26SMike Kravetz hugetlb_vma_lock_alloc(vma); 67608d9bfb26SMike Kravetz 67618d9bfb26SMike Kravetz /* 676217c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 676317c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 676490481622SDavid Gibson * without using reserves 676517c9d12eSMel Gorman */ 6766ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 676733b8f84aSMike Kravetz return true; 676817c9d12eSMel Gorman 676917c9d12eSMel Gorman /* 6770a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 6771a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 6772a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 6773a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 6774a1e78772SMel Gorman */ 67759119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 6776f27a5136SMike Kravetz /* 6777f27a5136SMike Kravetz * resv_map can not be NULL as hugetlb_reserve_pages is only 6778f27a5136SMike Kravetz * called for inodes for which resv_maps were created (see 6779f27a5136SMike Kravetz * hugetlbfs_get_inode). 6780f27a5136SMike Kravetz */ 67814e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 67829119a41eSJoonsoo Kim 67830db9d74eSMina Almasry chg = region_chg(resv_map, from, to, ®ions_needed); 67849119a41eSJoonsoo Kim } else { 6785e9fe92aeSMina Almasry /* Private mapping. */ 67869119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 67875a6fe125SMel Gorman if (!resv_map) 67888d9bfb26SMike Kravetz goto out_err; 67895a6fe125SMel Gorman 679017c9d12eSMel Gorman chg = to - from; 679117c9d12eSMel Gorman 67925a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 67935a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 67945a6fe125SMel Gorman } 67955a6fe125SMel Gorman 679633b8f84aSMike Kravetz if (chg < 0) 6797c50ac050SDave Hansen goto out_err; 679817c9d12eSMel Gorman 679933b8f84aSMike Kravetz if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), 680033b8f84aSMike Kravetz chg * pages_per_huge_page(h), &h_cg) < 0) 6801075a61d0SMina Almasry goto out_err; 6802075a61d0SMina Almasry 6803075a61d0SMina Almasry if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { 6804075a61d0SMina Almasry /* For private mappings, the hugetlb_cgroup uncharge info hangs 6805075a61d0SMina Almasry * of the resv_map. 6806075a61d0SMina Almasry */ 6807075a61d0SMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); 6808075a61d0SMina Almasry } 6809075a61d0SMina Almasry 68101c5ecae3SMike Kravetz /* 68111c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 68121c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 68131c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 68141c5ecae3SMike Kravetz */ 68151c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 681633b8f84aSMike Kravetz if (gbl_reserve < 0) 6817075a61d0SMina Almasry goto out_uncharge_cgroup; 681817c9d12eSMel Gorman 681917c9d12eSMel Gorman /* 682017c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 682190481622SDavid Gibson * Hand the pages back to the subpool if there are not 682217c9d12eSMel Gorman */ 682333b8f84aSMike Kravetz if (hugetlb_acct_memory(h, gbl_reserve) < 0) 6824075a61d0SMina Almasry goto out_put_pages; 682517c9d12eSMel Gorman 682617c9d12eSMel Gorman /* 682717c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 682817c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 682917c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 683017c9d12eSMel Gorman * the reservation was and the page cache tells how much of 683117c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 683217c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 683317c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 683417c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 683517c9d12eSMel Gorman * else has to be done for private mappings here 683617c9d12eSMel Gorman */ 683733039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 6838075a61d0SMina Almasry add = region_add(resv_map, from, to, regions_needed, h, h_cg); 683933039678SMike Kravetz 68400db9d74eSMina Almasry if (unlikely(add < 0)) { 68410db9d74eSMina Almasry hugetlb_acct_memory(h, -gbl_reserve); 6842075a61d0SMina Almasry goto out_put_pages; 68430db9d74eSMina Almasry } else if (unlikely(chg > add)) { 684433039678SMike Kravetz /* 684533039678SMike Kravetz * pages in this range were added to the reserve 684633039678SMike Kravetz * map between region_chg and region_add. This 684733039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 684833039678SMike Kravetz * the subpool and reserve counts modified above 684933039678SMike Kravetz * based on the difference. 685033039678SMike Kravetz */ 685133039678SMike Kravetz long rsv_adjust; 685233039678SMike Kravetz 6853d85aecf2SMiaohe Lin /* 6854d85aecf2SMiaohe Lin * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the 6855d85aecf2SMiaohe Lin * reference to h_cg->css. See comment below for detail. 6856d85aecf2SMiaohe Lin */ 6857075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd( 6858075a61d0SMina Almasry hstate_index(h), 6859075a61d0SMina Almasry (chg - add) * pages_per_huge_page(h), h_cg); 6860075a61d0SMina Almasry 686133039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 686233039678SMike Kravetz chg - add); 686333039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 6864d85aecf2SMiaohe Lin } else if (h_cg) { 6865d85aecf2SMiaohe Lin /* 6866d85aecf2SMiaohe Lin * The file_regions will hold their own reference to 6867d85aecf2SMiaohe Lin * h_cg->css. So we should release the reference held 6868d85aecf2SMiaohe Lin * via hugetlb_cgroup_charge_cgroup_rsvd() when we are 6869d85aecf2SMiaohe Lin * done. 6870d85aecf2SMiaohe Lin */ 6871d85aecf2SMiaohe Lin hugetlb_cgroup_put_rsvd_cgroup(h_cg); 687233039678SMike Kravetz } 687333039678SMike Kravetz } 687433b8f84aSMike Kravetz return true; 687533b8f84aSMike Kravetz 6876075a61d0SMina Almasry out_put_pages: 6877075a61d0SMina Almasry /* put back original number of pages, chg */ 6878075a61d0SMina Almasry (void)hugepage_subpool_put_pages(spool, chg); 6879075a61d0SMina Almasry out_uncharge_cgroup: 6880075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), 6881075a61d0SMina Almasry chg * pages_per_huge_page(h), h_cg); 6882c50ac050SDave Hansen out_err: 68838d9bfb26SMike Kravetz hugetlb_vma_lock_free(vma); 68845e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 68850db9d74eSMina Almasry /* Only call region_abort if the region_chg succeeded but the 68860db9d74eSMina Almasry * region_add failed or didn't run. 68870db9d74eSMina Almasry */ 68880db9d74eSMina Almasry if (chg >= 0 && add < 0) 68890db9d74eSMina Almasry region_abort(resv_map, from, to, regions_needed); 6890f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 6891f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 689233b8f84aSMike Kravetz return false; 6893a43a8c39SChen, Kenneth W } 6894a43a8c39SChen, Kenneth W 6895b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 6896b5cec28dSMike Kravetz long freed) 6897a43a8c39SChen, Kenneth W { 6898a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 68994e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 69009119a41eSJoonsoo Kim long chg = 0; 690190481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 69021c5ecae3SMike Kravetz long gbl_reserve; 690345c682a6SKen Chen 6904f27a5136SMike Kravetz /* 6905f27a5136SMike Kravetz * Since this routine can be called in the evict inode path for all 6906f27a5136SMike Kravetz * hugetlbfs inodes, resv_map could be NULL. 6907f27a5136SMike Kravetz */ 6908b5cec28dSMike Kravetz if (resv_map) { 6909b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 6910b5cec28dSMike Kravetz /* 6911b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 6912b5cec28dSMike Kravetz * must be split and another region descriptor can not be 6913b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 6914b5cec28dSMike Kravetz */ 6915b5cec28dSMike Kravetz if (chg < 0) 6916b5cec28dSMike Kravetz return chg; 6917b5cec28dSMike Kravetz } 6918b5cec28dSMike Kravetz 691945c682a6SKen Chen spin_lock(&inode->i_lock); 6920e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 692145c682a6SKen Chen spin_unlock(&inode->i_lock); 692245c682a6SKen Chen 69231c5ecae3SMike Kravetz /* 69241c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 69251c5ecae3SMike Kravetz * reservations to be released may be adjusted. 6926dddf31a4SMiaohe Lin * 6927dddf31a4SMiaohe Lin * Note that !resv_map implies freed == 0. So (chg - freed) 6928dddf31a4SMiaohe Lin * won't go negative. 69291c5ecae3SMike Kravetz */ 69301c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 69311c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 6932b5cec28dSMike Kravetz 6933b5cec28dSMike Kravetz return 0; 6934a43a8c39SChen, Kenneth W } 693593f70f90SNaoya Horiguchi 69363212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 69373212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 69383212b535SSteve Capper struct vm_area_struct *vma, 69393212b535SSteve Capper unsigned long addr, pgoff_t idx) 69403212b535SSteve Capper { 69413212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 69423212b535SSteve Capper svma->vm_start; 69433212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 69443212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 69453212b535SSteve Capper 69463212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 6947de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 6948de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 69493212b535SSteve Capper 69503212b535SSteve Capper /* 69513212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 69523212b535SSteve Capper * page table page. 6953131a79b4SMike Kravetz * 6954131a79b4SMike Kravetz * Also, vma_lock (vm_private_data) is required for sharing. 69553212b535SSteve Capper */ 69563212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 69573212b535SSteve Capper vm_flags != svm_flags || 6958131a79b4SMike Kravetz !range_in_vma(svma, sbase, s_end) || 6959131a79b4SMike Kravetz !svma->vm_private_data) 69603212b535SSteve Capper return 0; 69613212b535SSteve Capper 69623212b535SSteve Capper return saddr; 69633212b535SSteve Capper } 69643212b535SSteve Capper 6965bbff39ccSMike Kravetz bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 69663212b535SSteve Capper { 6967bbff39ccSMike Kravetz unsigned long start = addr & PUD_MASK; 6968bbff39ccSMike Kravetz unsigned long end = start + PUD_SIZE; 6969bbff39ccSMike Kravetz 69708d9bfb26SMike Kravetz #ifdef CONFIG_USERFAULTFD 69718d9bfb26SMike Kravetz if (uffd_disable_huge_pmd_share(vma)) 69728d9bfb26SMike Kravetz return false; 69738d9bfb26SMike Kravetz #endif 69743212b535SSteve Capper /* 69753212b535SSteve Capper * check on proper vm_flags and page table alignment 69763212b535SSteve Capper */ 69778d9bfb26SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 697831aafb45SNicholas Krause return false; 6979bbff39ccSMike Kravetz if (!vma->vm_private_data) /* vma lock required for sharing */ 69808d9bfb26SMike Kravetz return false; 69818d9bfb26SMike Kravetz if (!range_in_vma(vma, start, end)) 69828d9bfb26SMike Kravetz return false; 69838d9bfb26SMike Kravetz return true; 69848d9bfb26SMike Kravetz } 69858d9bfb26SMike Kravetz 69863212b535SSteve Capper /* 6987017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 6988017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 6989017b1660SMike Kravetz * shared pmd mappings. 6990017b1660SMike Kravetz */ 6991017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 6992017b1660SMike Kravetz unsigned long *start, unsigned long *end) 6993017b1660SMike Kravetz { 6994a1ba9da8SLi Xinhai unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), 6995a1ba9da8SLi Xinhai v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 6996017b1660SMike Kravetz 6997a1ba9da8SLi Xinhai /* 6998f0953a1bSIngo Molnar * vma needs to span at least one aligned PUD size, and the range 6999f0953a1bSIngo Molnar * must be at least partially within in. 7000a1ba9da8SLi Xinhai */ 7001a1ba9da8SLi Xinhai if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || 7002a1ba9da8SLi Xinhai (*end <= v_start) || (*start >= v_end)) 7003017b1660SMike Kravetz return; 7004017b1660SMike Kravetz 700575802ca6SPeter Xu /* Extend the range to be PUD aligned for a worst case scenario */ 7006a1ba9da8SLi Xinhai if (*start > v_start) 7007a1ba9da8SLi Xinhai *start = ALIGN_DOWN(*start, PUD_SIZE); 7008017b1660SMike Kravetz 7009a1ba9da8SLi Xinhai if (*end < v_end) 7010a1ba9da8SLi Xinhai *end = ALIGN(*end, PUD_SIZE); 7011017b1660SMike Kravetz } 7012017b1660SMike Kravetz 7013017b1660SMike Kravetz /* 70143212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 70153212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 70163212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 70173a47c54fSMike Kravetz * code much cleaner. pmd allocation is essential for the shared case because 70183a47c54fSMike Kravetz * pud has to be populated inside the same i_mmap_rwsem section - otherwise 70193a47c54fSMike Kravetz * racing tasks could either miss the sharing (see huge_pte_offset) or select a 70203a47c54fSMike Kravetz * bad pmd for sharing. 70213212b535SSteve Capper */ 7022aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 7023aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 70243212b535SSteve Capper { 70253212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 70263212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 70273212b535SSteve Capper vma->vm_pgoff; 70283212b535SSteve Capper struct vm_area_struct *svma; 70293212b535SSteve Capper unsigned long saddr; 70303212b535SSteve Capper pte_t *spte = NULL; 70313212b535SSteve Capper pte_t *pte; 7032cb900f41SKirill A. Shutemov spinlock_t *ptl; 70333212b535SSteve Capper 70343a47c54fSMike Kravetz i_mmap_lock_read(mapping); 70353212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 70363212b535SSteve Capper if (svma == vma) 70373212b535SSteve Capper continue; 70383212b535SSteve Capper 70393212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 70403212b535SSteve Capper if (saddr) { 70417868a208SPunit Agrawal spte = huge_pte_offset(svma->vm_mm, saddr, 70427868a208SPunit Agrawal vma_mmu_pagesize(svma)); 70433212b535SSteve Capper if (spte) { 70443212b535SSteve Capper get_page(virt_to_page(spte)); 70453212b535SSteve Capper break; 70463212b535SSteve Capper } 70473212b535SSteve Capper } 70483212b535SSteve Capper } 70493212b535SSteve Capper 70503212b535SSteve Capper if (!spte) 70513212b535SSteve Capper goto out; 70523212b535SSteve Capper 70538bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 7054dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 70553212b535SSteve Capper pud_populate(mm, pud, 70563212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 7057c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 7058dc6c9a35SKirill A. Shutemov } else { 70593212b535SSteve Capper put_page(virt_to_page(spte)); 7060dc6c9a35SKirill A. Shutemov } 7061cb900f41SKirill A. Shutemov spin_unlock(ptl); 70623212b535SSteve Capper out: 70633212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 70643a47c54fSMike Kravetz i_mmap_unlock_read(mapping); 70653212b535SSteve Capper return pte; 70663212b535SSteve Capper } 70673212b535SSteve Capper 70683212b535SSteve Capper /* 70693212b535SSteve Capper * unmap huge page backed by shared pte. 70703212b535SSteve Capper * 70713212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 70723212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 70733212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 70743212b535SSteve Capper * 70753a47c54fSMike Kravetz * Called with page table lock held. 70763212b535SSteve Capper * 70773212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 70783212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 70793212b535SSteve Capper */ 708034ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 70814ddb4d91SMike Kravetz unsigned long addr, pte_t *ptep) 70823212b535SSteve Capper { 70834ddb4d91SMike Kravetz pgd_t *pgd = pgd_offset(mm, addr); 70844ddb4d91SMike Kravetz p4d_t *p4d = p4d_offset(pgd, addr); 70854ddb4d91SMike Kravetz pud_t *pud = pud_offset(p4d, addr); 70863212b535SSteve Capper 708734ae204fSMike Kravetz i_mmap_assert_write_locked(vma->vm_file->f_mapping); 708840549ba8SMike Kravetz hugetlb_vma_assert_locked(vma); 70893212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 70903212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 70913212b535SSteve Capper return 0; 70923212b535SSteve Capper 70933212b535SSteve Capper pud_clear(pud); 70943212b535SSteve Capper put_page(virt_to_page(ptep)); 7095dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 70963212b535SSteve Capper return 1; 70973212b535SSteve Capper } 7098c1991e07SPeter Xu 70999e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 71008d9bfb26SMike Kravetz 7101aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 7102aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 71039e5fc74cSSteve Capper { 71049e5fc74cSSteve Capper return NULL; 71059e5fc74cSSteve Capper } 7106e81f2d22SZhang Zhen 710734ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 71084ddb4d91SMike Kravetz unsigned long addr, pte_t *ptep) 7109e81f2d22SZhang Zhen { 7110e81f2d22SZhang Zhen return 0; 7111e81f2d22SZhang Zhen } 7112017b1660SMike Kravetz 7113017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 7114017b1660SMike Kravetz unsigned long *start, unsigned long *end) 7115017b1660SMike Kravetz { 7116017b1660SMike Kravetz } 7117c1991e07SPeter Xu 7118c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 7119c1991e07SPeter Xu { 7120c1991e07SPeter Xu return false; 7121c1991e07SPeter Xu } 71223212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 71233212b535SSteve Capper 71249e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 7125aec44e0fSPeter Xu pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, 71269e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 71279e5fc74cSSteve Capper { 71289e5fc74cSSteve Capper pgd_t *pgd; 7129c2febafcSKirill A. Shutemov p4d_t *p4d; 71309e5fc74cSSteve Capper pud_t *pud; 71319e5fc74cSSteve Capper pte_t *pte = NULL; 71329e5fc74cSSteve Capper 71339e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 7134f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 7135f4f0a3d8SKirill A. Shutemov if (!p4d) 7136f4f0a3d8SKirill A. Shutemov return NULL; 7137c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 71389e5fc74cSSteve Capper if (pud) { 71399e5fc74cSSteve Capper if (sz == PUD_SIZE) { 71409e5fc74cSSteve Capper pte = (pte_t *)pud; 71419e5fc74cSSteve Capper } else { 71429e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 7143c1991e07SPeter Xu if (want_pmd_share(vma, addr) && pud_none(*pud)) 7144aec44e0fSPeter Xu pte = huge_pmd_share(mm, vma, addr, pud); 71459e5fc74cSSteve Capper else 71469e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 71479e5fc74cSSteve Capper } 71489e5fc74cSSteve Capper } 71494e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 71509e5fc74cSSteve Capper 71519e5fc74cSSteve Capper return pte; 71529e5fc74cSSteve Capper } 71539e5fc74cSSteve Capper 71549b19df29SPunit Agrawal /* 71559b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 71569b19df29SPunit Agrawal * entry at address @addr 71579b19df29SPunit Agrawal * 71588ac0b81aSLi Xinhai * Return: Pointer to page table entry (PUD or PMD) for 71598ac0b81aSLi Xinhai * address @addr, or NULL if a !p*d_present() entry is encountered and the 71609b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 71619b19df29SPunit Agrawal * table. 71629b19df29SPunit Agrawal */ 71637868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 71647868a208SPunit Agrawal unsigned long addr, unsigned long sz) 71659e5fc74cSSteve Capper { 71669e5fc74cSSteve Capper pgd_t *pgd; 7167c2febafcSKirill A. Shutemov p4d_t *p4d; 71688ac0b81aSLi Xinhai pud_t *pud; 71698ac0b81aSLi Xinhai pmd_t *pmd; 71709e5fc74cSSteve Capper 71719e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 7172c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 7173c2febafcSKirill A. Shutemov return NULL; 7174c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 7175c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 7176c2febafcSKirill A. Shutemov return NULL; 71779b19df29SPunit Agrawal 7178c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 71798ac0b81aSLi Xinhai if (sz == PUD_SIZE) 71808ac0b81aSLi Xinhai /* must be pud huge, non-present or none */ 71819e5fc74cSSteve Capper return (pte_t *)pud; 71828ac0b81aSLi Xinhai if (!pud_present(*pud)) 71838ac0b81aSLi Xinhai return NULL; 71848ac0b81aSLi Xinhai /* must have a valid entry and size to go further */ 71859b19df29SPunit Agrawal 71869e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 71878ac0b81aSLi Xinhai /* must be pmd huge, non-present or none */ 71889e5fc74cSSteve Capper return (pte_t *)pmd; 71899e5fc74cSSteve Capper } 71909e5fc74cSSteve Capper 7191e95a9851SMike Kravetz /* 7192e95a9851SMike Kravetz * Return a mask that can be used to update an address to the last huge 7193e95a9851SMike Kravetz * page in a page table page mapping size. Used to skip non-present 7194e95a9851SMike Kravetz * page table entries when linearly scanning address ranges. Architectures 7195e95a9851SMike Kravetz * with unique huge page to page table relationships can define their own 7196e95a9851SMike Kravetz * version of this routine. 7197e95a9851SMike Kravetz */ 7198e95a9851SMike Kravetz unsigned long hugetlb_mask_last_page(struct hstate *h) 7199e95a9851SMike Kravetz { 7200e95a9851SMike Kravetz unsigned long hp_size = huge_page_size(h); 7201e95a9851SMike Kravetz 7202e95a9851SMike Kravetz if (hp_size == PUD_SIZE) 7203e95a9851SMike Kravetz return P4D_SIZE - PUD_SIZE; 7204e95a9851SMike Kravetz else if (hp_size == PMD_SIZE) 7205e95a9851SMike Kravetz return PUD_SIZE - PMD_SIZE; 7206e95a9851SMike Kravetz else 7207e95a9851SMike Kravetz return 0UL; 7208e95a9851SMike Kravetz } 7209e95a9851SMike Kravetz 7210e95a9851SMike Kravetz #else 7211e95a9851SMike Kravetz 7212e95a9851SMike Kravetz /* See description above. Architectures can provide their own version. */ 7213e95a9851SMike Kravetz __weak unsigned long hugetlb_mask_last_page(struct hstate *h) 7214e95a9851SMike Kravetz { 72154ddb4d91SMike Kravetz #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 72164ddb4d91SMike Kravetz if (huge_page_size(h) == PMD_SIZE) 72174ddb4d91SMike Kravetz return PUD_SIZE - PMD_SIZE; 72184ddb4d91SMike Kravetz #endif 7219e95a9851SMike Kravetz return 0UL; 7220e95a9851SMike Kravetz } 7221e95a9851SMike Kravetz 722261f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 722361f77edaSNaoya Horiguchi 722461f77edaSNaoya Horiguchi /* 722561f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 722661f77edaSNaoya Horiguchi * behavior. 722761f77edaSNaoya Horiguchi */ 72287ce82f4cSMiaohe Lin int isolate_hugetlb(struct page *page, struct list_head *list) 722931caf665SNaoya Horiguchi { 72307ce82f4cSMiaohe Lin int ret = 0; 7231bcc54222SNaoya Horiguchi 7232db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 72338f251a3dSMike Kravetz if (!PageHeadHuge(page) || 72348f251a3dSMike Kravetz !HPageMigratable(page) || 72350eb2df2bSMuchun Song !get_page_unless_zero(page)) { 72367ce82f4cSMiaohe Lin ret = -EBUSY; 7237bcc54222SNaoya Horiguchi goto unlock; 7238bcc54222SNaoya Horiguchi } 72398f251a3dSMike Kravetz ClearHPageMigratable(page); 724031caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 7241bcc54222SNaoya Horiguchi unlock: 7242db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7243bcc54222SNaoya Horiguchi return ret; 724431caf665SNaoya Horiguchi } 724531caf665SNaoya Horiguchi 7246e591ef7dSNaoya Horiguchi int get_hwpoison_huge_page(struct page *page, bool *hugetlb, bool unpoison) 724725182f05SNaoya Horiguchi { 724825182f05SNaoya Horiguchi int ret = 0; 724925182f05SNaoya Horiguchi 725025182f05SNaoya Horiguchi *hugetlb = false; 725125182f05SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 725225182f05SNaoya Horiguchi if (PageHeadHuge(page)) { 725325182f05SNaoya Horiguchi *hugetlb = true; 7254b283d983SNaoya Horiguchi if (HPageFreed(page)) 7255b283d983SNaoya Horiguchi ret = 0; 7256e591ef7dSNaoya Horiguchi else if (HPageMigratable(page) || unpoison) 725725182f05SNaoya Horiguchi ret = get_page_unless_zero(page); 72580ed950d1SNaoya Horiguchi else 72590ed950d1SNaoya Horiguchi ret = -EBUSY; 726025182f05SNaoya Horiguchi } 726125182f05SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 726225182f05SNaoya Horiguchi return ret; 726325182f05SNaoya Horiguchi } 726425182f05SNaoya Horiguchi 7265e591ef7dSNaoya Horiguchi int get_huge_page_for_hwpoison(unsigned long pfn, int flags, 7266e591ef7dSNaoya Horiguchi bool *migratable_cleared) 7267405ce051SNaoya Horiguchi { 7268405ce051SNaoya Horiguchi int ret; 7269405ce051SNaoya Horiguchi 7270405ce051SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 7271e591ef7dSNaoya Horiguchi ret = __get_huge_page_for_hwpoison(pfn, flags, migratable_cleared); 7272405ce051SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 7273405ce051SNaoya Horiguchi return ret; 7274405ce051SNaoya Horiguchi } 7275405ce051SNaoya Horiguchi 727631caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 727731caf665SNaoya Horiguchi { 7278db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 72798f251a3dSMike Kravetz SetHPageMigratable(page); 728031caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 7281db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 728231caf665SNaoya Horiguchi put_page(page); 728331caf665SNaoya Horiguchi } 7284ab5ac90aSMichal Hocko 7285345c62d1SSidhartha Kumar void move_hugetlb_state(struct folio *old_folio, struct folio *new_folio, int reason) 7286ab5ac90aSMichal Hocko { 7287345c62d1SSidhartha Kumar struct hstate *h = folio_hstate(old_folio); 7288ab5ac90aSMichal Hocko 7289345c62d1SSidhartha Kumar hugetlb_cgroup_migrate(old_folio, new_folio); 7290345c62d1SSidhartha Kumar set_page_owner_migrate_reason(&new_folio->page, reason); 7291ab5ac90aSMichal Hocko 7292ab5ac90aSMichal Hocko /* 7293345c62d1SSidhartha Kumar * transfer temporary state of the new hugetlb folio. This is 7294ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 7295ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 7296ab5ac90aSMichal Hocko * the temporary status. 7297ab5ac90aSMichal Hocko * 7298ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 7299ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 7300ab5ac90aSMichal Hocko * the per-node's. 7301ab5ac90aSMichal Hocko */ 7302345c62d1SSidhartha Kumar if (folio_test_hugetlb_temporary(new_folio)) { 7303345c62d1SSidhartha Kumar int old_nid = folio_nid(old_folio); 7304345c62d1SSidhartha Kumar int new_nid = folio_nid(new_folio); 7305ab5ac90aSMichal Hocko 7306345c62d1SSidhartha Kumar folio_set_hugetlb_temporary(old_folio); 7307345c62d1SSidhartha Kumar folio_clear_hugetlb_temporary(new_folio); 7308345c62d1SSidhartha Kumar 7309ab5ac90aSMichal Hocko 73105af1ab1dSMiaohe Lin /* 73115af1ab1dSMiaohe Lin * There is no need to transfer the per-node surplus state 73125af1ab1dSMiaohe Lin * when we do not cross the node. 73135af1ab1dSMiaohe Lin */ 73145af1ab1dSMiaohe Lin if (new_nid == old_nid) 73155af1ab1dSMiaohe Lin return; 7316db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 7317ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 7318ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 7319ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 7320ab5ac90aSMichal Hocko } 7321db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7322ab5ac90aSMichal Hocko } 7323ab5ac90aSMichal Hocko } 7324cf11e85fSRoman Gushchin 73256dfeaff9SPeter Xu /* 73266dfeaff9SPeter Xu * This function will unconditionally remove all the shared pmd pgtable entries 73276dfeaff9SPeter Xu * within the specific vma for a hugetlbfs memory range. 73286dfeaff9SPeter Xu */ 73296dfeaff9SPeter Xu void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) 73306dfeaff9SPeter Xu { 73316dfeaff9SPeter Xu struct hstate *h = hstate_vma(vma); 73326dfeaff9SPeter Xu unsigned long sz = huge_page_size(h); 73336dfeaff9SPeter Xu struct mm_struct *mm = vma->vm_mm; 73346dfeaff9SPeter Xu struct mmu_notifier_range range; 73356dfeaff9SPeter Xu unsigned long address, start, end; 73366dfeaff9SPeter Xu spinlock_t *ptl; 73376dfeaff9SPeter Xu pte_t *ptep; 73386dfeaff9SPeter Xu 73396dfeaff9SPeter Xu if (!(vma->vm_flags & VM_MAYSHARE)) 73406dfeaff9SPeter Xu return; 73416dfeaff9SPeter Xu 73426dfeaff9SPeter Xu start = ALIGN(vma->vm_start, PUD_SIZE); 73436dfeaff9SPeter Xu end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 73446dfeaff9SPeter Xu 73456dfeaff9SPeter Xu if (start >= end) 73466dfeaff9SPeter Xu return; 73476dfeaff9SPeter Xu 73489c8bbfacSBaolin Wang flush_cache_range(vma, start, end); 73496dfeaff9SPeter Xu /* 73506dfeaff9SPeter Xu * No need to call adjust_range_if_pmd_sharing_possible(), because 73516dfeaff9SPeter Xu * we have already done the PUD_SIZE alignment. 73526dfeaff9SPeter Xu */ 73536dfeaff9SPeter Xu mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, 73546dfeaff9SPeter Xu start, end); 73556dfeaff9SPeter Xu mmu_notifier_invalidate_range_start(&range); 735640549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 73576dfeaff9SPeter Xu i_mmap_lock_write(vma->vm_file->f_mapping); 73586dfeaff9SPeter Xu for (address = start; address < end; address += PUD_SIZE) { 73596dfeaff9SPeter Xu ptep = huge_pte_offset(mm, address, sz); 73606dfeaff9SPeter Xu if (!ptep) 73616dfeaff9SPeter Xu continue; 73626dfeaff9SPeter Xu ptl = huge_pte_lock(h, mm, ptep); 73634ddb4d91SMike Kravetz huge_pmd_unshare(mm, vma, address, ptep); 73646dfeaff9SPeter Xu spin_unlock(ptl); 73656dfeaff9SPeter Xu } 73666dfeaff9SPeter Xu flush_hugetlb_tlb_range(vma, start, end); 73676dfeaff9SPeter Xu i_mmap_unlock_write(vma->vm_file->f_mapping); 736840549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 73696dfeaff9SPeter Xu /* 73706dfeaff9SPeter Xu * No need to call mmu_notifier_invalidate_range(), see 7371ee65728eSMike Rapoport * Documentation/mm/mmu_notifier.rst. 73726dfeaff9SPeter Xu */ 73736dfeaff9SPeter Xu mmu_notifier_invalidate_range_end(&range); 73746dfeaff9SPeter Xu } 73756dfeaff9SPeter Xu 7376cf11e85fSRoman Gushchin #ifdef CONFIG_CMA 7377cf11e85fSRoman Gushchin static bool cma_reserve_called __initdata; 7378cf11e85fSRoman Gushchin 7379cf11e85fSRoman Gushchin static int __init cmdline_parse_hugetlb_cma(char *p) 7380cf11e85fSRoman Gushchin { 738138e719abSBaolin Wang int nid, count = 0; 738238e719abSBaolin Wang unsigned long tmp; 738338e719abSBaolin Wang char *s = p; 738438e719abSBaolin Wang 738538e719abSBaolin Wang while (*s) { 738638e719abSBaolin Wang if (sscanf(s, "%lu%n", &tmp, &count) != 1) 738738e719abSBaolin Wang break; 738838e719abSBaolin Wang 738938e719abSBaolin Wang if (s[count] == ':') { 7390f9317f77SMike Kravetz if (tmp >= MAX_NUMNODES) 739138e719abSBaolin Wang break; 7392f9317f77SMike Kravetz nid = array_index_nospec(tmp, MAX_NUMNODES); 739338e719abSBaolin Wang 739438e719abSBaolin Wang s += count + 1; 739538e719abSBaolin Wang tmp = memparse(s, &s); 739638e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = tmp; 739738e719abSBaolin Wang hugetlb_cma_size += tmp; 739838e719abSBaolin Wang 739938e719abSBaolin Wang /* 740038e719abSBaolin Wang * Skip the separator if have one, otherwise 740138e719abSBaolin Wang * break the parsing. 740238e719abSBaolin Wang */ 740338e719abSBaolin Wang if (*s == ',') 740438e719abSBaolin Wang s++; 740538e719abSBaolin Wang else 740638e719abSBaolin Wang break; 740738e719abSBaolin Wang } else { 7408cf11e85fSRoman Gushchin hugetlb_cma_size = memparse(p, &p); 740938e719abSBaolin Wang break; 741038e719abSBaolin Wang } 741138e719abSBaolin Wang } 741238e719abSBaolin Wang 7413cf11e85fSRoman Gushchin return 0; 7414cf11e85fSRoman Gushchin } 7415cf11e85fSRoman Gushchin 7416cf11e85fSRoman Gushchin early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); 7417cf11e85fSRoman Gushchin 7418cf11e85fSRoman Gushchin void __init hugetlb_cma_reserve(int order) 7419cf11e85fSRoman Gushchin { 7420cf11e85fSRoman Gushchin unsigned long size, reserved, per_node; 742138e719abSBaolin Wang bool node_specific_cma_alloc = false; 7422cf11e85fSRoman Gushchin int nid; 7423cf11e85fSRoman Gushchin 7424cf11e85fSRoman Gushchin cma_reserve_called = true; 7425cf11e85fSRoman Gushchin 7426cf11e85fSRoman Gushchin if (!hugetlb_cma_size) 7427cf11e85fSRoman Gushchin return; 7428cf11e85fSRoman Gushchin 742938e719abSBaolin Wang for (nid = 0; nid < MAX_NUMNODES; nid++) { 743038e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 743138e719abSBaolin Wang continue; 743238e719abSBaolin Wang 743330a51400SPeng Liu if (!node_online(nid)) { 743438e719abSBaolin Wang pr_warn("hugetlb_cma: invalid node %d specified\n", nid); 743538e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 743638e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 743738e719abSBaolin Wang continue; 743838e719abSBaolin Wang } 743938e719abSBaolin Wang 744038e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) { 744138e719abSBaolin Wang pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n", 744238e719abSBaolin Wang nid, (PAGE_SIZE << order) / SZ_1M); 744338e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 744438e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 744538e719abSBaolin Wang } else { 744638e719abSBaolin Wang node_specific_cma_alloc = true; 744738e719abSBaolin Wang } 744838e719abSBaolin Wang } 744938e719abSBaolin Wang 745038e719abSBaolin Wang /* Validate the CMA size again in case some invalid nodes specified. */ 745138e719abSBaolin Wang if (!hugetlb_cma_size) 745238e719abSBaolin Wang return; 745338e719abSBaolin Wang 7454cf11e85fSRoman Gushchin if (hugetlb_cma_size < (PAGE_SIZE << order)) { 7455cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", 7456cf11e85fSRoman Gushchin (PAGE_SIZE << order) / SZ_1M); 7457a01f4390SMike Kravetz hugetlb_cma_size = 0; 7458cf11e85fSRoman Gushchin return; 7459cf11e85fSRoman Gushchin } 7460cf11e85fSRoman Gushchin 746138e719abSBaolin Wang if (!node_specific_cma_alloc) { 7462cf11e85fSRoman Gushchin /* 7463cf11e85fSRoman Gushchin * If 3 GB area is requested on a machine with 4 numa nodes, 7464cf11e85fSRoman Gushchin * let's allocate 1 GB on first three nodes and ignore the last one. 7465cf11e85fSRoman Gushchin */ 7466cf11e85fSRoman Gushchin per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); 7467cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", 7468cf11e85fSRoman Gushchin hugetlb_cma_size / SZ_1M, per_node / SZ_1M); 746938e719abSBaolin Wang } 7470cf11e85fSRoman Gushchin 7471cf11e85fSRoman Gushchin reserved = 0; 747230a51400SPeng Liu for_each_online_node(nid) { 7473cf11e85fSRoman Gushchin int res; 74742281f797SBarry Song char name[CMA_MAX_NAME]; 7475cf11e85fSRoman Gushchin 747638e719abSBaolin Wang if (node_specific_cma_alloc) { 747738e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 747838e719abSBaolin Wang continue; 747938e719abSBaolin Wang 748038e719abSBaolin Wang size = hugetlb_cma_size_in_node[nid]; 748138e719abSBaolin Wang } else { 7482cf11e85fSRoman Gushchin size = min(per_node, hugetlb_cma_size - reserved); 748338e719abSBaolin Wang } 748438e719abSBaolin Wang 7485cf11e85fSRoman Gushchin size = round_up(size, PAGE_SIZE << order); 7486cf11e85fSRoman Gushchin 74872281f797SBarry Song snprintf(name, sizeof(name), "hugetlb%d", nid); 7488a01f4390SMike Kravetz /* 7489a01f4390SMike Kravetz * Note that 'order per bit' is based on smallest size that 7490a01f4390SMike Kravetz * may be returned to CMA allocator in the case of 7491a01f4390SMike Kravetz * huge page demotion. 7492a01f4390SMike Kravetz */ 7493a01f4390SMike Kravetz res = cma_declare_contiguous_nid(0, size, 0, 7494a01f4390SMike Kravetz PAGE_SIZE << HUGETLB_PAGE_ORDER, 749529d0f41dSBarry Song 0, false, name, 7496cf11e85fSRoman Gushchin &hugetlb_cma[nid], nid); 7497cf11e85fSRoman Gushchin if (res) { 7498cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: reservation failed: err %d, node %d", 7499cf11e85fSRoman Gushchin res, nid); 7500cf11e85fSRoman Gushchin continue; 7501cf11e85fSRoman Gushchin } 7502cf11e85fSRoman Gushchin 7503cf11e85fSRoman Gushchin reserved += size; 7504cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", 7505cf11e85fSRoman Gushchin size / SZ_1M, nid); 7506cf11e85fSRoman Gushchin 7507cf11e85fSRoman Gushchin if (reserved >= hugetlb_cma_size) 7508cf11e85fSRoman Gushchin break; 7509cf11e85fSRoman Gushchin } 7510a01f4390SMike Kravetz 7511a01f4390SMike Kravetz if (!reserved) 7512a01f4390SMike Kravetz /* 7513a01f4390SMike Kravetz * hugetlb_cma_size is used to determine if allocations from 7514a01f4390SMike Kravetz * cma are possible. Set to zero if no cma regions are set up. 7515a01f4390SMike Kravetz */ 7516a01f4390SMike Kravetz hugetlb_cma_size = 0; 7517cf11e85fSRoman Gushchin } 7518cf11e85fSRoman Gushchin 7519263b8998SMiaohe Lin static void __init hugetlb_cma_check(void) 7520cf11e85fSRoman Gushchin { 7521cf11e85fSRoman Gushchin if (!hugetlb_cma_size || cma_reserve_called) 7522cf11e85fSRoman Gushchin return; 7523cf11e85fSRoman Gushchin 7524cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); 7525cf11e85fSRoman Gushchin } 7526cf11e85fSRoman Gushchin 7527cf11e85fSRoman Gushchin #endif /* CONFIG_CMA */ 7528