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> 2263489f8eSMike Kravetz #include <linux/mmdebug.h> 23174cd4b1SIngo Molnar #include <linux/sched/signal.h> 240fe6e20bSNaoya Horiguchi #include <linux/rmap.h> 25c6247f72SMatthew Wilcox #include <linux/string_helpers.h> 26fd6a03edSNaoya Horiguchi #include <linux/swap.h> 27fd6a03edSNaoya Horiguchi #include <linux/swapops.h> 288382d914SDavidlohr Bueso #include <linux/jhash.h> 2998fa15f3SAnshuman Khandual #include <linux/numa.h> 30c77c0a8aSWaiman Long #include <linux/llist.h> 31d6606683SLinus Torvalds 3263551ae0SDavid Gibson #include <asm/page.h> 3363551ae0SDavid Gibson #include <asm/pgtable.h> 3424669e58SAneesh Kumar K.V #include <asm/tlb.h> 3563551ae0SDavid Gibson 3624669e58SAneesh Kumar K.V #include <linux/io.h> 3763551ae0SDavid Gibson #include <linux/hugetlb.h> 389dd540e2SAneesh Kumar K.V #include <linux/hugetlb_cgroup.h> 399a305230SLee Schermerhorn #include <linux/node.h> 401a1aad8aSMike Kravetz #include <linux/userfaultfd_k.h> 41ab5ac90aSMichal Hocko #include <linux/page_owner.h> 427835e98bSNick Piggin #include "internal.h" 431da177e4SLinus Torvalds 44c3f38a38SAneesh Kumar K.V int hugetlb_max_hstate __read_mostly; 45e5ff2159SAndi Kleen unsigned int default_hstate_idx; 46e5ff2159SAndi Kleen struct hstate hstates[HUGE_MAX_HSTATE]; 47641844f5SNaoya Horiguchi /* 48641844f5SNaoya Horiguchi * Minimum page order among possible hugepage sizes, set to a proper value 49641844f5SNaoya Horiguchi * at boot time. 50641844f5SNaoya Horiguchi */ 51641844f5SNaoya Horiguchi static unsigned int minimum_order __read_mostly = UINT_MAX; 52e5ff2159SAndi Kleen 5353ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 5453ba51d2SJon Tollefson 55e5ff2159SAndi Kleen /* for command line parsing */ 56e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 57e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 58e11bfbfcSNick Piggin static unsigned long __initdata default_hstate_size; 599fee021dSVaishali Thakkar static bool __initdata parsed_valid_hugepagesz = true; 60e5ff2159SAndi Kleen 613935baa9SDavid Gibson /* 6231caf665SNaoya Horiguchi * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, 6331caf665SNaoya Horiguchi * free_huge_pages, and surplus_huge_pages. 643935baa9SDavid Gibson */ 65c3f38a38SAneesh Kumar K.V DEFINE_SPINLOCK(hugetlb_lock); 660bd0f9fbSEric Paris 678382d914SDavidlohr Bueso /* 688382d914SDavidlohr Bueso * Serializes faults on the same logical page. This is used to 698382d914SDavidlohr Bueso * prevent spurious OOMs when the hugepage pool is fully utilized. 708382d914SDavidlohr Bueso */ 718382d914SDavidlohr Bueso static int num_fault_mutexes; 72c672c7f2SMike Kravetz struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; 738382d914SDavidlohr Bueso 747ca02d0aSMike Kravetz /* Forward declaration */ 757ca02d0aSMike Kravetz static int hugetlb_acct_memory(struct hstate *h, long delta); 767ca02d0aSMike Kravetz 7790481622SDavid Gibson static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) 7890481622SDavid Gibson { 7990481622SDavid Gibson bool free = (spool->count == 0) && (spool->used_hpages == 0); 8090481622SDavid Gibson 8190481622SDavid Gibson spin_unlock(&spool->lock); 8290481622SDavid Gibson 8390481622SDavid Gibson /* If no pages are used, and no other handles to the subpool 847ca02d0aSMike Kravetz * remain, give up any reservations mased on minimum size and 857ca02d0aSMike Kravetz * free the subpool */ 867ca02d0aSMike Kravetz if (free) { 877ca02d0aSMike Kravetz if (spool->min_hpages != -1) 887ca02d0aSMike Kravetz hugetlb_acct_memory(spool->hstate, 897ca02d0aSMike Kravetz -spool->min_hpages); 9090481622SDavid Gibson kfree(spool); 9190481622SDavid Gibson } 927ca02d0aSMike Kravetz } 9390481622SDavid Gibson 947ca02d0aSMike Kravetz struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, 957ca02d0aSMike Kravetz long min_hpages) 9690481622SDavid Gibson { 9790481622SDavid Gibson struct hugepage_subpool *spool; 9890481622SDavid Gibson 99c6a91820SMike Kravetz spool = kzalloc(sizeof(*spool), GFP_KERNEL); 10090481622SDavid Gibson if (!spool) 10190481622SDavid Gibson return NULL; 10290481622SDavid Gibson 10390481622SDavid Gibson spin_lock_init(&spool->lock); 10490481622SDavid Gibson spool->count = 1; 1057ca02d0aSMike Kravetz spool->max_hpages = max_hpages; 1067ca02d0aSMike Kravetz spool->hstate = h; 1077ca02d0aSMike Kravetz spool->min_hpages = min_hpages; 1087ca02d0aSMike Kravetz 1097ca02d0aSMike Kravetz if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { 1107ca02d0aSMike Kravetz kfree(spool); 1117ca02d0aSMike Kravetz return NULL; 1127ca02d0aSMike Kravetz } 1137ca02d0aSMike Kravetz spool->rsv_hpages = min_hpages; 11490481622SDavid Gibson 11590481622SDavid Gibson return spool; 11690481622SDavid Gibson } 11790481622SDavid Gibson 11890481622SDavid Gibson void hugepage_put_subpool(struct hugepage_subpool *spool) 11990481622SDavid Gibson { 12090481622SDavid Gibson spin_lock(&spool->lock); 12190481622SDavid Gibson BUG_ON(!spool->count); 12290481622SDavid Gibson spool->count--; 12390481622SDavid Gibson unlock_or_release_subpool(spool); 12490481622SDavid Gibson } 12590481622SDavid Gibson 1261c5ecae3SMike Kravetz /* 1271c5ecae3SMike Kravetz * Subpool accounting for allocating and reserving pages. 1281c5ecae3SMike Kravetz * Return -ENOMEM if there are not enough resources to satisfy the 1291c5ecae3SMike Kravetz * the request. Otherwise, return the number of pages by which the 1301c5ecae3SMike Kravetz * global pools must be adjusted (upward). The returned value may 1311c5ecae3SMike Kravetz * only be different than the passed value (delta) in the case where 1321c5ecae3SMike Kravetz * a subpool minimum size must be manitained. 1331c5ecae3SMike Kravetz */ 1341c5ecae3SMike Kravetz static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, 13590481622SDavid Gibson long delta) 13690481622SDavid Gibson { 1371c5ecae3SMike Kravetz long ret = delta; 13890481622SDavid Gibson 13990481622SDavid Gibson if (!spool) 1401c5ecae3SMike Kravetz return ret; 14190481622SDavid Gibson 14290481622SDavid Gibson spin_lock(&spool->lock); 14390481622SDavid Gibson 1441c5ecae3SMike Kravetz if (spool->max_hpages != -1) { /* maximum size accounting */ 1451c5ecae3SMike Kravetz if ((spool->used_hpages + delta) <= spool->max_hpages) 1461c5ecae3SMike Kravetz spool->used_hpages += delta; 1471c5ecae3SMike Kravetz else { 1481c5ecae3SMike Kravetz ret = -ENOMEM; 1491c5ecae3SMike Kravetz goto unlock_ret; 1501c5ecae3SMike Kravetz } 1511c5ecae3SMike Kravetz } 1521c5ecae3SMike Kravetz 15309a95e29SMike Kravetz /* minimum size accounting */ 15409a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->rsv_hpages) { 1551c5ecae3SMike Kravetz if (delta > spool->rsv_hpages) { 1561c5ecae3SMike Kravetz /* 1571c5ecae3SMike Kravetz * Asking for more reserves than those already taken on 1581c5ecae3SMike Kravetz * behalf of subpool. Return difference. 1591c5ecae3SMike Kravetz */ 1601c5ecae3SMike Kravetz ret = delta - spool->rsv_hpages; 1611c5ecae3SMike Kravetz spool->rsv_hpages = 0; 1621c5ecae3SMike Kravetz } else { 1631c5ecae3SMike Kravetz ret = 0; /* reserves already accounted for */ 1641c5ecae3SMike Kravetz spool->rsv_hpages -= delta; 1651c5ecae3SMike Kravetz } 1661c5ecae3SMike Kravetz } 1671c5ecae3SMike Kravetz 1681c5ecae3SMike Kravetz unlock_ret: 1691c5ecae3SMike Kravetz spin_unlock(&spool->lock); 17090481622SDavid Gibson return ret; 17190481622SDavid Gibson } 17290481622SDavid Gibson 1731c5ecae3SMike Kravetz /* 1741c5ecae3SMike Kravetz * Subpool accounting for freeing and unreserving pages. 1751c5ecae3SMike Kravetz * Return the number of global page reservations that must be dropped. 1761c5ecae3SMike Kravetz * The return value may only be different than the passed value (delta) 1771c5ecae3SMike Kravetz * in the case where a subpool minimum size must be maintained. 1781c5ecae3SMike Kravetz */ 1791c5ecae3SMike Kravetz static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, 18090481622SDavid Gibson long delta) 18190481622SDavid Gibson { 1821c5ecae3SMike Kravetz long ret = delta; 1831c5ecae3SMike Kravetz 18490481622SDavid Gibson if (!spool) 1851c5ecae3SMike Kravetz return delta; 18690481622SDavid Gibson 18790481622SDavid Gibson spin_lock(&spool->lock); 1881c5ecae3SMike Kravetz 1891c5ecae3SMike Kravetz if (spool->max_hpages != -1) /* maximum size accounting */ 19090481622SDavid Gibson spool->used_hpages -= delta; 1911c5ecae3SMike Kravetz 19209a95e29SMike Kravetz /* minimum size accounting */ 19309a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { 1941c5ecae3SMike Kravetz if (spool->rsv_hpages + delta <= spool->min_hpages) 1951c5ecae3SMike Kravetz ret = 0; 1961c5ecae3SMike Kravetz else 1971c5ecae3SMike Kravetz ret = spool->rsv_hpages + delta - spool->min_hpages; 1981c5ecae3SMike Kravetz 1991c5ecae3SMike Kravetz spool->rsv_hpages += delta; 2001c5ecae3SMike Kravetz if (spool->rsv_hpages > spool->min_hpages) 2011c5ecae3SMike Kravetz spool->rsv_hpages = spool->min_hpages; 2021c5ecae3SMike Kravetz } 2031c5ecae3SMike Kravetz 2041c5ecae3SMike Kravetz /* 2051c5ecae3SMike Kravetz * If hugetlbfs_put_super couldn't free spool due to an outstanding 2061c5ecae3SMike Kravetz * quota reference, free it now. 2071c5ecae3SMike Kravetz */ 20890481622SDavid Gibson unlock_or_release_subpool(spool); 2091c5ecae3SMike Kravetz 2101c5ecae3SMike Kravetz return ret; 21190481622SDavid Gibson } 21290481622SDavid Gibson 21390481622SDavid Gibson static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 21490481622SDavid Gibson { 21590481622SDavid Gibson return HUGETLBFS_SB(inode->i_sb)->spool; 21690481622SDavid Gibson } 21790481622SDavid Gibson 21890481622SDavid Gibson static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 21990481622SDavid Gibson { 220496ad9aaSAl Viro return subpool_inode(file_inode(vma->vm_file)); 22190481622SDavid Gibson } 22290481622SDavid Gibson 223e7c4b0bfSAndy Whitcroft /* 22496822904SAndy Whitcroft * Region tracking -- allows tracking of reservations and instantiated pages 22596822904SAndy Whitcroft * across the pages in a mapping. 22684afd99bSAndy Whitcroft * 2271dd308a7SMike Kravetz * The region data structures are embedded into a resv_map and protected 2281dd308a7SMike Kravetz * by a resv_map's lock. The set of regions within the resv_map represent 2291dd308a7SMike Kravetz * reservations for huge pages, or huge pages that have already been 2301dd308a7SMike Kravetz * instantiated within the map. The from and to elements are huge page 2311dd308a7SMike Kravetz * indicies into the associated mapping. from indicates the starting index 2321dd308a7SMike Kravetz * of the region. to represents the first index past the end of the region. 2331dd308a7SMike Kravetz * 2341dd308a7SMike Kravetz * For example, a file region structure with from == 0 and to == 4 represents 2351dd308a7SMike Kravetz * four huge pages in a mapping. It is important to note that the to element 2361dd308a7SMike Kravetz * represents the first element past the end of the region. This is used in 2371dd308a7SMike Kravetz * arithmetic as 4(to) - 0(from) = 4 huge pages in the region. 2381dd308a7SMike Kravetz * 2391dd308a7SMike Kravetz * Interval notation of the form [from, to) will be used to indicate that 2401dd308a7SMike Kravetz * the endpoint from is inclusive and to is exclusive. 24196822904SAndy Whitcroft */ 24296822904SAndy Whitcroft struct file_region { 24396822904SAndy Whitcroft struct list_head link; 24496822904SAndy Whitcroft long from; 24596822904SAndy Whitcroft long to; 24696822904SAndy Whitcroft }; 24796822904SAndy Whitcroft 248d75c6af9SMina Almasry /* Must be called with resv->lock held. Calling this with count_only == true 249d75c6af9SMina Almasry * will count the number of pages to be added but will not modify the linked 250d75c6af9SMina Almasry * list. 251d75c6af9SMina Almasry */ 252d75c6af9SMina Almasry static long add_reservation_in_range(struct resv_map *resv, long f, long t, 253d75c6af9SMina Almasry bool count_only) 254d75c6af9SMina Almasry { 255d75c6af9SMina Almasry long chg = 0; 256d75c6af9SMina Almasry struct list_head *head = &resv->regions; 257d75c6af9SMina Almasry struct file_region *rg = NULL, *trg = NULL, *nrg = NULL; 258d75c6af9SMina Almasry 259d75c6af9SMina Almasry /* Locate the region we are before or in. */ 260d75c6af9SMina Almasry list_for_each_entry(rg, head, link) 261d75c6af9SMina Almasry if (f <= rg->to) 262d75c6af9SMina Almasry break; 263d75c6af9SMina Almasry 264d75c6af9SMina Almasry /* Round our left edge to the current segment if it encloses us. */ 265d75c6af9SMina Almasry if (f > rg->from) 266d75c6af9SMina Almasry f = rg->from; 267d75c6af9SMina Almasry 268d75c6af9SMina Almasry chg = t - f; 269d75c6af9SMina Almasry 270d75c6af9SMina Almasry /* Check for and consume any regions we now overlap with. */ 271d75c6af9SMina Almasry nrg = rg; 272d75c6af9SMina Almasry list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 273d75c6af9SMina Almasry if (&rg->link == head) 274d75c6af9SMina Almasry break; 275d75c6af9SMina Almasry if (rg->from > t) 276d75c6af9SMina Almasry break; 277d75c6af9SMina Almasry 278d75c6af9SMina Almasry /* We overlap with this area, if it extends further than 279d75c6af9SMina Almasry * us then we must extend ourselves. Account for its 280d75c6af9SMina Almasry * existing reservation. 281d75c6af9SMina Almasry */ 282d75c6af9SMina Almasry if (rg->to > t) { 283d75c6af9SMina Almasry chg += rg->to - t; 284d75c6af9SMina Almasry t = rg->to; 285d75c6af9SMina Almasry } 286d75c6af9SMina Almasry chg -= rg->to - rg->from; 287d75c6af9SMina Almasry 288d75c6af9SMina Almasry if (!count_only && rg != nrg) { 289d75c6af9SMina Almasry list_del(&rg->link); 290d75c6af9SMina Almasry kfree(rg); 291d75c6af9SMina Almasry } 292d75c6af9SMina Almasry } 293d75c6af9SMina Almasry 294d75c6af9SMina Almasry if (!count_only) { 295d75c6af9SMina Almasry nrg->from = f; 296d75c6af9SMina Almasry nrg->to = t; 297d75c6af9SMina Almasry } 298d75c6af9SMina Almasry 299d75c6af9SMina Almasry return chg; 300d75c6af9SMina Almasry } 301d75c6af9SMina Almasry 3021dd308a7SMike Kravetz /* 3031dd308a7SMike Kravetz * Add the huge page range represented by [f, t) to the reserve 3045c911954SMina Almasry * map. Existing regions will be expanded to accommodate the specified 3055c911954SMina Almasry * range, or a region will be taken from the cache. Sufficient regions 3065c911954SMina Almasry * must exist in the cache due to the previous call to region_chg with 3075c911954SMina Almasry * the same range. 308cf3ad20bSMike Kravetz * 309cf3ad20bSMike Kravetz * Return the number of new huge pages added to the map. This 310cf3ad20bSMike Kravetz * number is greater than or equal to zero. 3111dd308a7SMike Kravetz */ 3121406ec9bSJoonsoo Kim static long region_add(struct resv_map *resv, long f, long t) 31396822904SAndy Whitcroft { 3141406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 315d75c6af9SMina Almasry struct file_region *rg, *nrg; 316cf3ad20bSMike Kravetz long add = 0; 31796822904SAndy Whitcroft 3187b24d861SDavidlohr Bueso spin_lock(&resv->lock); 31996822904SAndy Whitcroft /* Locate the region we are either in or before. */ 32096822904SAndy Whitcroft list_for_each_entry(rg, head, link) 32196822904SAndy Whitcroft if (f <= rg->to) 32296822904SAndy Whitcroft break; 32396822904SAndy Whitcroft 3245e911373SMike Kravetz /* 3255e911373SMike Kravetz * If no region exists which can be expanded to include the 3265c911954SMina Almasry * specified range, pull a region descriptor from the cache 3275c911954SMina Almasry * and use it for this range. 3285e911373SMike Kravetz */ 3295e911373SMike Kravetz if (&rg->link == head || t < rg->from) { 3305e911373SMike Kravetz VM_BUG_ON(resv->region_cache_count <= 0); 3315e911373SMike Kravetz 3325e911373SMike Kravetz resv->region_cache_count--; 3335e911373SMike Kravetz nrg = list_first_entry(&resv->region_cache, struct file_region, 3345e911373SMike Kravetz link); 3355e911373SMike Kravetz list_del(&nrg->link); 3365e911373SMike Kravetz 3375e911373SMike Kravetz nrg->from = f; 3385e911373SMike Kravetz nrg->to = t; 3395e911373SMike Kravetz list_add(&nrg->link, rg->link.prev); 3405e911373SMike Kravetz 3415e911373SMike Kravetz add += t - f; 3425e911373SMike Kravetz goto out_locked; 3435e911373SMike Kravetz } 3445e911373SMike Kravetz 345d75c6af9SMina Almasry add = add_reservation_in_range(resv, f, t, false); 346cf3ad20bSMike Kravetz 3475e911373SMike Kravetz out_locked: 3485e911373SMike Kravetz resv->adds_in_progress--; 3497b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 350cf3ad20bSMike Kravetz VM_BUG_ON(add < 0); 351cf3ad20bSMike Kravetz return add; 35296822904SAndy Whitcroft } 35396822904SAndy Whitcroft 3541dd308a7SMike Kravetz /* 3551dd308a7SMike Kravetz * Examine the existing reserve map and determine how many 3561dd308a7SMike Kravetz * huge pages in the specified range [f, t) are NOT currently 3571dd308a7SMike Kravetz * represented. This routine is called before a subsequent 3581dd308a7SMike Kravetz * call to region_add that will actually modify the reserve 3591dd308a7SMike Kravetz * map to add the specified range [f, t). region_chg does 3601dd308a7SMike Kravetz * not change the number of huge pages represented by the 3615c911954SMina Almasry * map. A new file_region structure is added to the cache 3625c911954SMina Almasry * as a placeholder, so that the subsequent region_add 3635c911954SMina Almasry * call will have all the regions it needs and will not fail. 3645e911373SMike Kravetz * 3655e911373SMike Kravetz * Returns the number of huge pages that need to be added to the existing 3665e911373SMike Kravetz * reservation map for the range [f, t). This number is greater or equal to 3675e911373SMike Kravetz * zero. -ENOMEM is returned if a new file_region structure or cache entry 3685e911373SMike Kravetz * is needed and can not be allocated. 3691dd308a7SMike Kravetz */ 3701406ec9bSJoonsoo Kim static long region_chg(struct resv_map *resv, long f, long t) 37196822904SAndy Whitcroft { 37296822904SAndy Whitcroft long chg = 0; 37396822904SAndy Whitcroft 3747b24d861SDavidlohr Bueso spin_lock(&resv->lock); 3755e911373SMike Kravetz retry_locked: 3765e911373SMike Kravetz resv->adds_in_progress++; 3775e911373SMike Kravetz 3785e911373SMike Kravetz /* 3795e911373SMike Kravetz * Check for sufficient descriptors in the cache to accommodate 3805e911373SMike Kravetz * the number of in progress add operations. 3815e911373SMike Kravetz */ 3825e911373SMike Kravetz if (resv->adds_in_progress > resv->region_cache_count) { 3835e911373SMike Kravetz struct file_region *trg; 3845e911373SMike Kravetz 3855e911373SMike Kravetz VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1); 3865e911373SMike Kravetz /* Must drop lock to allocate a new descriptor. */ 3875e911373SMike Kravetz resv->adds_in_progress--; 3885e911373SMike Kravetz spin_unlock(&resv->lock); 3895e911373SMike Kravetz 3905e911373SMike Kravetz trg = kmalloc(sizeof(*trg), GFP_KERNEL); 3915c911954SMina Almasry if (!trg) 3925e911373SMike Kravetz return -ENOMEM; 3935e911373SMike Kravetz 3945e911373SMike Kravetz spin_lock(&resv->lock); 3955e911373SMike Kravetz list_add(&trg->link, &resv->region_cache); 3965e911373SMike Kravetz resv->region_cache_count++; 3975e911373SMike Kravetz goto retry_locked; 3985e911373SMike Kravetz } 3995e911373SMike Kravetz 400d75c6af9SMina Almasry chg = add_reservation_in_range(resv, f, t, true); 40196822904SAndy Whitcroft 4027b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 40396822904SAndy Whitcroft return chg; 40496822904SAndy Whitcroft } 40596822904SAndy Whitcroft 4061dd308a7SMike Kravetz /* 4075e911373SMike Kravetz * Abort the in progress add operation. The adds_in_progress field 4085e911373SMike Kravetz * of the resv_map keeps track of the operations in progress between 4095e911373SMike Kravetz * calls to region_chg and region_add. Operations are sometimes 4105e911373SMike Kravetz * aborted after the call to region_chg. In such cases, region_abort 4115e911373SMike Kravetz * is called to decrement the adds_in_progress counter. 4125e911373SMike Kravetz * 4135e911373SMike Kravetz * NOTE: The range arguments [f, t) are not needed or used in this 4145e911373SMike Kravetz * routine. They are kept to make reading the calling code easier as 4155e911373SMike Kravetz * arguments will match the associated region_chg call. 4165e911373SMike Kravetz */ 4175e911373SMike Kravetz static void region_abort(struct resv_map *resv, long f, long t) 4185e911373SMike Kravetz { 4195e911373SMike Kravetz spin_lock(&resv->lock); 4205e911373SMike Kravetz VM_BUG_ON(!resv->region_cache_count); 4215e911373SMike Kravetz resv->adds_in_progress--; 4225e911373SMike Kravetz spin_unlock(&resv->lock); 4235e911373SMike Kravetz } 4245e911373SMike Kravetz 4255e911373SMike Kravetz /* 426feba16e2SMike Kravetz * Delete the specified range [f, t) from the reserve map. If the 427feba16e2SMike Kravetz * t parameter is LONG_MAX, this indicates that ALL regions after f 428feba16e2SMike Kravetz * should be deleted. Locate the regions which intersect [f, t) 429feba16e2SMike Kravetz * and either trim, delete or split the existing regions. 430feba16e2SMike Kravetz * 431feba16e2SMike Kravetz * Returns the number of huge pages deleted from the reserve map. 432feba16e2SMike Kravetz * In the normal case, the return value is zero or more. In the 433feba16e2SMike Kravetz * case where a region must be split, a new region descriptor must 434feba16e2SMike Kravetz * be allocated. If the allocation fails, -ENOMEM will be returned. 435feba16e2SMike Kravetz * NOTE: If the parameter t == LONG_MAX, then we will never split 436feba16e2SMike Kravetz * a region and possibly return -ENOMEM. Callers specifying 437feba16e2SMike Kravetz * t == LONG_MAX do not need to check for -ENOMEM error. 4381dd308a7SMike Kravetz */ 439feba16e2SMike Kravetz static long region_del(struct resv_map *resv, long f, long t) 44096822904SAndy Whitcroft { 4411406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 44296822904SAndy Whitcroft struct file_region *rg, *trg; 443feba16e2SMike Kravetz struct file_region *nrg = NULL; 444feba16e2SMike Kravetz long del = 0; 44596822904SAndy Whitcroft 446feba16e2SMike Kravetz retry: 4477b24d861SDavidlohr Bueso spin_lock(&resv->lock); 448feba16e2SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 449dbe409e4SMike Kravetz /* 450dbe409e4SMike Kravetz * Skip regions before the range to be deleted. file_region 451dbe409e4SMike Kravetz * ranges are normally of the form [from, to). However, there 452dbe409e4SMike Kravetz * may be a "placeholder" entry in the map which is of the form 453dbe409e4SMike Kravetz * (from, to) with from == to. Check for placeholder entries 454dbe409e4SMike Kravetz * at the beginning of the range to be deleted. 455dbe409e4SMike Kravetz */ 456dbe409e4SMike Kravetz if (rg->to <= f && (rg->to != rg->from || rg->to != f)) 457feba16e2SMike Kravetz continue; 458dbe409e4SMike Kravetz 459feba16e2SMike Kravetz if (rg->from >= t) 46096822904SAndy Whitcroft break; 46196822904SAndy Whitcroft 462feba16e2SMike Kravetz if (f > rg->from && t < rg->to) { /* Must split region */ 463feba16e2SMike Kravetz /* 464feba16e2SMike Kravetz * Check for an entry in the cache before dropping 465feba16e2SMike Kravetz * lock and attempting allocation. 466feba16e2SMike Kravetz */ 467feba16e2SMike Kravetz if (!nrg && 468feba16e2SMike Kravetz resv->region_cache_count > resv->adds_in_progress) { 469feba16e2SMike Kravetz nrg = list_first_entry(&resv->region_cache, 470feba16e2SMike Kravetz struct file_region, 471feba16e2SMike Kravetz link); 472feba16e2SMike Kravetz list_del(&nrg->link); 473feba16e2SMike Kravetz resv->region_cache_count--; 47496822904SAndy Whitcroft } 47596822904SAndy Whitcroft 476feba16e2SMike Kravetz if (!nrg) { 477feba16e2SMike Kravetz spin_unlock(&resv->lock); 478feba16e2SMike Kravetz nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 479feba16e2SMike Kravetz if (!nrg) 480feba16e2SMike Kravetz return -ENOMEM; 481feba16e2SMike Kravetz goto retry; 482feba16e2SMike Kravetz } 483feba16e2SMike Kravetz 484feba16e2SMike Kravetz del += t - f; 485feba16e2SMike Kravetz 486feba16e2SMike Kravetz /* New entry for end of split region */ 487feba16e2SMike Kravetz nrg->from = t; 488feba16e2SMike Kravetz nrg->to = rg->to; 489feba16e2SMike Kravetz INIT_LIST_HEAD(&nrg->link); 490feba16e2SMike Kravetz 491feba16e2SMike Kravetz /* Original entry is trimmed */ 492feba16e2SMike Kravetz rg->to = f; 493feba16e2SMike Kravetz 494feba16e2SMike Kravetz list_add(&nrg->link, &rg->link); 495feba16e2SMike Kravetz nrg = NULL; 49696822904SAndy Whitcroft break; 497feba16e2SMike Kravetz } 498feba16e2SMike Kravetz 499feba16e2SMike Kravetz if (f <= rg->from && t >= rg->to) { /* Remove entire region */ 500feba16e2SMike Kravetz del += rg->to - rg->from; 50196822904SAndy Whitcroft list_del(&rg->link); 50296822904SAndy Whitcroft kfree(rg); 503feba16e2SMike Kravetz continue; 50496822904SAndy Whitcroft } 5057b24d861SDavidlohr Bueso 506feba16e2SMike Kravetz if (f <= rg->from) { /* Trim beginning of region */ 507feba16e2SMike Kravetz del += t - rg->from; 508feba16e2SMike Kravetz rg->from = t; 509feba16e2SMike Kravetz } else { /* Trim end of region */ 510feba16e2SMike Kravetz del += rg->to - f; 511feba16e2SMike Kravetz rg->to = f; 512feba16e2SMike Kravetz } 513feba16e2SMike Kravetz } 514feba16e2SMike Kravetz 5157b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 516feba16e2SMike Kravetz kfree(nrg); 517feba16e2SMike Kravetz return del; 51896822904SAndy Whitcroft } 51996822904SAndy Whitcroft 5201dd308a7SMike Kravetz /* 521b5cec28dSMike Kravetz * A rare out of memory error was encountered which prevented removal of 522b5cec28dSMike Kravetz * the reserve map region for a page. The huge page itself was free'ed 523b5cec28dSMike Kravetz * and removed from the page cache. This routine will adjust the subpool 524b5cec28dSMike Kravetz * usage count, and the global reserve count if needed. By incrementing 525b5cec28dSMike Kravetz * these counts, the reserve map entry which could not be deleted will 526b5cec28dSMike Kravetz * appear as a "reserved" entry instead of simply dangling with incorrect 527b5cec28dSMike Kravetz * counts. 528b5cec28dSMike Kravetz */ 52972e2936cSzhong jiang void hugetlb_fix_reserve_counts(struct inode *inode) 530b5cec28dSMike Kravetz { 531b5cec28dSMike Kravetz struct hugepage_subpool *spool = subpool_inode(inode); 532b5cec28dSMike Kravetz long rsv_adjust; 533b5cec28dSMike Kravetz 534b5cec28dSMike Kravetz rsv_adjust = hugepage_subpool_get_pages(spool, 1); 53572e2936cSzhong jiang if (rsv_adjust) { 536b5cec28dSMike Kravetz struct hstate *h = hstate_inode(inode); 537b5cec28dSMike Kravetz 538b5cec28dSMike Kravetz hugetlb_acct_memory(h, 1); 539b5cec28dSMike Kravetz } 540b5cec28dSMike Kravetz } 541b5cec28dSMike Kravetz 542b5cec28dSMike Kravetz /* 5431dd308a7SMike Kravetz * Count and return the number of huge pages in the reserve map 5441dd308a7SMike Kravetz * that intersect with the range [f, t). 5451dd308a7SMike Kravetz */ 5461406ec9bSJoonsoo Kim static long region_count(struct resv_map *resv, long f, long t) 54784afd99bSAndy Whitcroft { 5481406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 54984afd99bSAndy Whitcroft struct file_region *rg; 55084afd99bSAndy Whitcroft long chg = 0; 55184afd99bSAndy Whitcroft 5527b24d861SDavidlohr Bueso spin_lock(&resv->lock); 55384afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 55484afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 555f2135a4aSWang Sheng-Hui long seg_from; 556f2135a4aSWang Sheng-Hui long seg_to; 55784afd99bSAndy Whitcroft 55884afd99bSAndy Whitcroft if (rg->to <= f) 55984afd99bSAndy Whitcroft continue; 56084afd99bSAndy Whitcroft if (rg->from >= t) 56184afd99bSAndy Whitcroft break; 56284afd99bSAndy Whitcroft 56384afd99bSAndy Whitcroft seg_from = max(rg->from, f); 56484afd99bSAndy Whitcroft seg_to = min(rg->to, t); 56584afd99bSAndy Whitcroft 56684afd99bSAndy Whitcroft chg += seg_to - seg_from; 56784afd99bSAndy Whitcroft } 5687b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 56984afd99bSAndy Whitcroft 57084afd99bSAndy Whitcroft return chg; 57184afd99bSAndy Whitcroft } 57284afd99bSAndy Whitcroft 57396822904SAndy Whitcroft /* 574e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 575e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 576e7c4b0bfSAndy Whitcroft */ 577a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 578a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 579e7c4b0bfSAndy Whitcroft { 580a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 581a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 582e7c4b0bfSAndy Whitcroft } 583e7c4b0bfSAndy Whitcroft 5840fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 5850fe6e20bSNaoya Horiguchi unsigned long address) 5860fe6e20bSNaoya Horiguchi { 5870fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 5880fe6e20bSNaoya Horiguchi } 589dee41079SDan Williams EXPORT_SYMBOL_GPL(linear_hugepage_index); 5900fe6e20bSNaoya Horiguchi 59184afd99bSAndy Whitcroft /* 59208fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 59308fba699SMel Gorman * cases this will be same size as used by the page table entries. 59408fba699SMel Gorman */ 59508fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 59608fba699SMel Gorman { 59705ea8860SDan Williams if (vma->vm_ops && vma->vm_ops->pagesize) 59805ea8860SDan Williams return vma->vm_ops->pagesize(vma); 59908fba699SMel Gorman return PAGE_SIZE; 60008fba699SMel Gorman } 601f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 60208fba699SMel Gorman 60308fba699SMel Gorman /* 6043340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 6053340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 60609135cc5SDan Williams * architectures where it differs, an architecture-specific 'strong' 60709135cc5SDan Williams * version of this symbol is required. 6083340289dSMel Gorman */ 60909135cc5SDan Williams __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 6103340289dSMel Gorman { 6113340289dSMel Gorman return vma_kernel_pagesize(vma); 6123340289dSMel Gorman } 6133340289dSMel Gorman 6143340289dSMel Gorman /* 61584afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 61684afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 61784afd99bSAndy Whitcroft * alignment. 61884afd99bSAndy Whitcroft */ 61984afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 62084afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 62104f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 62284afd99bSAndy Whitcroft 623a1e78772SMel Gorman /* 624a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 625a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 626a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 627a1e78772SMel Gorman * 628a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 629a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 630a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 631a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 63284afd99bSAndy Whitcroft * 63384afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 63484afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 63584afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 63684afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 63784afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 63884afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 63984afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 64084afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 641a1e78772SMel Gorman */ 642e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 643e7c4b0bfSAndy Whitcroft { 644e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 645e7c4b0bfSAndy Whitcroft } 646e7c4b0bfSAndy Whitcroft 647e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 648e7c4b0bfSAndy Whitcroft unsigned long value) 649e7c4b0bfSAndy Whitcroft { 650e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 651e7c4b0bfSAndy Whitcroft } 652e7c4b0bfSAndy Whitcroft 6539119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 65484afd99bSAndy Whitcroft { 65584afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 6565e911373SMike Kravetz struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 6575e911373SMike Kravetz 6585e911373SMike Kravetz if (!resv_map || !rg) { 6595e911373SMike Kravetz kfree(resv_map); 6605e911373SMike Kravetz kfree(rg); 66184afd99bSAndy Whitcroft return NULL; 6625e911373SMike Kravetz } 66384afd99bSAndy Whitcroft 66484afd99bSAndy Whitcroft kref_init(&resv_map->refs); 6657b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 66684afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 66784afd99bSAndy Whitcroft 6685e911373SMike Kravetz resv_map->adds_in_progress = 0; 6695e911373SMike Kravetz 6705e911373SMike Kravetz INIT_LIST_HEAD(&resv_map->region_cache); 6715e911373SMike Kravetz list_add(&rg->link, &resv_map->region_cache); 6725e911373SMike Kravetz resv_map->region_cache_count = 1; 6735e911373SMike Kravetz 67484afd99bSAndy Whitcroft return resv_map; 67584afd99bSAndy Whitcroft } 67684afd99bSAndy Whitcroft 6779119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 67884afd99bSAndy Whitcroft { 67984afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 6805e911373SMike Kravetz struct list_head *head = &resv_map->region_cache; 6815e911373SMike Kravetz struct file_region *rg, *trg; 68284afd99bSAndy Whitcroft 68384afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 684feba16e2SMike Kravetz region_del(resv_map, 0, LONG_MAX); 6855e911373SMike Kravetz 6865e911373SMike Kravetz /* ... and any entries left in the cache */ 6875e911373SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 6885e911373SMike Kravetz list_del(&rg->link); 6895e911373SMike Kravetz kfree(rg); 6905e911373SMike Kravetz } 6915e911373SMike Kravetz 6925e911373SMike Kravetz VM_BUG_ON(resv_map->adds_in_progress); 6935e911373SMike Kravetz 69484afd99bSAndy Whitcroft kfree(resv_map); 69584afd99bSAndy Whitcroft } 69684afd99bSAndy Whitcroft 6974e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 6984e35f483SJoonsoo Kim { 699f27a5136SMike Kravetz /* 700f27a5136SMike Kravetz * At inode evict time, i_mapping may not point to the original 701f27a5136SMike Kravetz * address space within the inode. This original address space 702f27a5136SMike Kravetz * contains the pointer to the resv_map. So, always use the 703f27a5136SMike Kravetz * address space embedded within the inode. 704f27a5136SMike Kravetz * The VERY common case is inode->mapping == &inode->i_data but, 705f27a5136SMike Kravetz * this may not be true for device special inodes. 706f27a5136SMike Kravetz */ 707f27a5136SMike Kravetz return (struct resv_map *)(&inode->i_data)->private_data; 7084e35f483SJoonsoo Kim } 7094e35f483SJoonsoo Kim 71084afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 711a1e78772SMel Gorman { 71281d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 7134e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 7144e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 7154e35f483SJoonsoo Kim struct inode *inode = mapping->host; 7164e35f483SJoonsoo Kim 7174e35f483SJoonsoo Kim return inode_resv_map(inode); 7184e35f483SJoonsoo Kim 7194e35f483SJoonsoo Kim } else { 72084afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 72184afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 7224e35f483SJoonsoo Kim } 723a1e78772SMel Gorman } 724a1e78772SMel Gorman 72584afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 726a1e78772SMel Gorman { 72781d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 72881d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 729a1e78772SMel Gorman 73084afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 73184afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 73204f2cbe3SMel Gorman } 73304f2cbe3SMel Gorman 73404f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 73504f2cbe3SMel Gorman { 73681d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 73781d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 738e7c4b0bfSAndy Whitcroft 739e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 74004f2cbe3SMel Gorman } 74104f2cbe3SMel Gorman 74204f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 74304f2cbe3SMel Gorman { 74481d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 745e7c4b0bfSAndy Whitcroft 746e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 747a1e78772SMel Gorman } 748a1e78772SMel Gorman 74904f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 750a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 751a1e78772SMel Gorman { 75281d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 753f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 754a1e78772SMel Gorman vma->vm_private_data = (void *)0; 755a1e78772SMel Gorman } 756a1e78772SMel Gorman 757a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 758559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 759a1e78772SMel Gorman { 760af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 761af0ed73eSJoonsoo Kim /* 762af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 763af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 764af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 765af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 766af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 767af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 768af0ed73eSJoonsoo Kim * properly, so add work-around here. 769af0ed73eSJoonsoo Kim */ 770af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 771559ec2f8SNicholas Krause return true; 772af0ed73eSJoonsoo Kim else 773559ec2f8SNicholas Krause return false; 774af0ed73eSJoonsoo Kim } 775a63884e9SJoonsoo Kim 776a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 7771fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 7781fb1b0e9SMike Kravetz /* 7791fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 7801fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 7811fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 7821fb1b0e9SMike Kravetz * fallocate. In this case, there really are no reverves to 7831fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 7841fb1b0e9SMike Kravetz */ 7851fb1b0e9SMike Kravetz if (chg) 7861fb1b0e9SMike Kravetz return false; 7871fb1b0e9SMike Kravetz else 788559ec2f8SNicholas Krause return true; 7891fb1b0e9SMike Kravetz } 790a63884e9SJoonsoo Kim 791a63884e9SJoonsoo Kim /* 792a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 793a63884e9SJoonsoo Kim * private mappings. 794a63884e9SJoonsoo Kim */ 79567961f9dSMike Kravetz if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 79667961f9dSMike Kravetz /* 79767961f9dSMike Kravetz * Like the shared case above, a hole punch or truncate 79867961f9dSMike Kravetz * could have been performed on the private mapping. 79967961f9dSMike Kravetz * Examine the value of chg to determine if reserves 80067961f9dSMike Kravetz * actually exist or were previously consumed. 80167961f9dSMike Kravetz * Very Subtle - The value of chg comes from a previous 80267961f9dSMike Kravetz * call to vma_needs_reserves(). The reserve map for 80367961f9dSMike Kravetz * private mappings has different (opposite) semantics 80467961f9dSMike Kravetz * than that of shared mappings. vma_needs_reserves() 80567961f9dSMike Kravetz * has already taken this difference in semantics into 80667961f9dSMike Kravetz * account. Therefore, the meaning of chg is the same 80767961f9dSMike Kravetz * as in the shared case above. Code could easily be 80867961f9dSMike Kravetz * combined, but keeping it separate draws attention to 80967961f9dSMike Kravetz * subtle differences. 81067961f9dSMike Kravetz */ 81167961f9dSMike Kravetz if (chg) 81267961f9dSMike Kravetz return false; 81367961f9dSMike Kravetz else 814559ec2f8SNicholas Krause return true; 81567961f9dSMike Kravetz } 816a63884e9SJoonsoo Kim 817559ec2f8SNicholas Krause return false; 818a1e78772SMel Gorman } 819a1e78772SMel Gorman 820a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 8211da177e4SLinus Torvalds { 8221da177e4SLinus Torvalds int nid = page_to_nid(page); 8230edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 824a5516438SAndi Kleen h->free_huge_pages++; 825a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 8261da177e4SLinus Torvalds } 8271da177e4SLinus Torvalds 82894310cbcSAnshuman Khandual static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) 829bf50bab2SNaoya Horiguchi { 830bf50bab2SNaoya Horiguchi struct page *page; 831bf50bab2SNaoya Horiguchi 832c8721bbbSNaoya Horiguchi list_for_each_entry(page, &h->hugepage_freelists[nid], lru) 833243abd5bSNaoya Horiguchi if (!PageHWPoison(page)) 834c8721bbbSNaoya Horiguchi break; 835c8721bbbSNaoya Horiguchi /* 836c8721bbbSNaoya Horiguchi * if 'non-isolated free hugepage' not found on the list, 837c8721bbbSNaoya Horiguchi * the allocation fails. 838c8721bbbSNaoya Horiguchi */ 839c8721bbbSNaoya Horiguchi if (&h->hugepage_freelists[nid] == &page->lru) 840bf50bab2SNaoya Horiguchi return NULL; 8410edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 842a9869b83SNaoya Horiguchi set_page_refcounted(page); 843bf50bab2SNaoya Horiguchi h->free_huge_pages--; 844bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 845bf50bab2SNaoya Horiguchi return page; 846bf50bab2SNaoya Horiguchi } 847bf50bab2SNaoya Horiguchi 8483e59fcb0SMichal Hocko static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, 8493e59fcb0SMichal Hocko nodemask_t *nmask) 85094310cbcSAnshuman Khandual { 8513e59fcb0SMichal Hocko unsigned int cpuset_mems_cookie; 8523e59fcb0SMichal Hocko struct zonelist *zonelist; 8533e59fcb0SMichal Hocko struct zone *zone; 8543e59fcb0SMichal Hocko struct zoneref *z; 85598fa15f3SAnshuman Khandual int node = NUMA_NO_NODE; 8563e59fcb0SMichal Hocko 8573e59fcb0SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 8583e59fcb0SMichal Hocko 8593e59fcb0SMichal Hocko retry_cpuset: 8603e59fcb0SMichal Hocko cpuset_mems_cookie = read_mems_allowed_begin(); 8613e59fcb0SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 86294310cbcSAnshuman Khandual struct page *page; 86394310cbcSAnshuman Khandual 8643e59fcb0SMichal Hocko if (!cpuset_zone_allowed(zone, gfp_mask)) 8653e59fcb0SMichal Hocko continue; 8663e59fcb0SMichal Hocko /* 8673e59fcb0SMichal Hocko * no need to ask again on the same node. Pool is node rather than 8683e59fcb0SMichal Hocko * zone aware 8693e59fcb0SMichal Hocko */ 8703e59fcb0SMichal Hocko if (zone_to_nid(zone) == node) 8713e59fcb0SMichal Hocko continue; 8723e59fcb0SMichal Hocko node = zone_to_nid(zone); 87394310cbcSAnshuman Khandual 87494310cbcSAnshuman Khandual page = dequeue_huge_page_node_exact(h, node); 87594310cbcSAnshuman Khandual if (page) 87694310cbcSAnshuman Khandual return page; 87794310cbcSAnshuman Khandual } 8783e59fcb0SMichal Hocko if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 8793e59fcb0SMichal Hocko goto retry_cpuset; 8803e59fcb0SMichal Hocko 88194310cbcSAnshuman Khandual return NULL; 88294310cbcSAnshuman Khandual } 88394310cbcSAnshuman Khandual 88486cdb465SNaoya Horiguchi /* Movability of hugepages depends on migration support. */ 88586cdb465SNaoya Horiguchi static inline gfp_t htlb_alloc_mask(struct hstate *h) 88686cdb465SNaoya Horiguchi { 8877ed2c31dSAnshuman Khandual if (hugepage_movable_supported(h)) 88886cdb465SNaoya Horiguchi return GFP_HIGHUSER_MOVABLE; 88986cdb465SNaoya Horiguchi else 89086cdb465SNaoya Horiguchi return GFP_HIGHUSER; 89186cdb465SNaoya Horiguchi } 89286cdb465SNaoya Horiguchi 893a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 894a5516438SAndi Kleen struct vm_area_struct *vma, 895af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 896af0ed73eSJoonsoo Kim long chg) 8971da177e4SLinus Torvalds { 8983e59fcb0SMichal Hocko struct page *page; 899480eccf9SLee Schermerhorn struct mempolicy *mpol; 90004ec6264SVlastimil Babka gfp_t gfp_mask; 9013e59fcb0SMichal Hocko nodemask_t *nodemask; 90204ec6264SVlastimil Babka int nid; 9031da177e4SLinus Torvalds 904a1e78772SMel Gorman /* 905a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 906a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 907a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 908a1e78772SMel Gorman */ 909af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg) && 910a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 911c0ff7453SMiao Xie goto err; 912a1e78772SMel Gorman 91304f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 914a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 9156eab04a8SJustin P. Mattock goto err; 91604f2cbe3SMel Gorman 91704ec6264SVlastimil Babka gfp_mask = htlb_alloc_mask(h); 91804ec6264SVlastimil Babka nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 9193e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 9203e59fcb0SMichal Hocko if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { 92107443a85SJoonsoo Kim SetPagePrivate(page); 922a63884e9SJoonsoo Kim h->resv_huge_pages--; 923bf50bab2SNaoya Horiguchi } 924cc9a6c87SMel Gorman 925cc9a6c87SMel Gorman mpol_cond_put(mpol); 926cc9a6c87SMel Gorman return page; 927cc9a6c87SMel Gorman 928c0ff7453SMiao Xie err: 929cc9a6c87SMel Gorman return NULL; 9301da177e4SLinus Torvalds } 9311da177e4SLinus Torvalds 9321cac6f2cSLuiz Capitulino /* 9331cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 9341cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 9351cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 9361cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 9371cac6f2cSLuiz Capitulino * node for alloc or free. 9381cac6f2cSLuiz Capitulino */ 9391cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 9401cac6f2cSLuiz Capitulino { 9410edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 9421cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 9431cac6f2cSLuiz Capitulino 9441cac6f2cSLuiz Capitulino return nid; 9451cac6f2cSLuiz Capitulino } 9461cac6f2cSLuiz Capitulino 9471cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 9481cac6f2cSLuiz Capitulino { 9491cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 9501cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 9511cac6f2cSLuiz Capitulino return nid; 9521cac6f2cSLuiz Capitulino } 9531cac6f2cSLuiz Capitulino 9541cac6f2cSLuiz Capitulino /* 9551cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 9561cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 9571cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 9581cac6f2cSLuiz Capitulino * mask. 9591cac6f2cSLuiz Capitulino */ 9601cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 9611cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 9621cac6f2cSLuiz Capitulino { 9631cac6f2cSLuiz Capitulino int nid; 9641cac6f2cSLuiz Capitulino 9651cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 9661cac6f2cSLuiz Capitulino 9671cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 9681cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 9691cac6f2cSLuiz Capitulino 9701cac6f2cSLuiz Capitulino return nid; 9711cac6f2cSLuiz Capitulino } 9721cac6f2cSLuiz Capitulino 9731cac6f2cSLuiz Capitulino /* 9741cac6f2cSLuiz Capitulino * helper for free_pool_huge_page() - return the previously saved 9751cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 9761cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 9771cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 9781cac6f2cSLuiz Capitulino */ 9791cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 9801cac6f2cSLuiz Capitulino { 9811cac6f2cSLuiz Capitulino int nid; 9821cac6f2cSLuiz Capitulino 9831cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 9841cac6f2cSLuiz Capitulino 9851cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 9861cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 9871cac6f2cSLuiz Capitulino 9881cac6f2cSLuiz Capitulino return nid; 9891cac6f2cSLuiz Capitulino } 9901cac6f2cSLuiz Capitulino 9911cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 9921cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 9931cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 9941cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 9951cac6f2cSLuiz Capitulino nr_nodes--) 9961cac6f2cSLuiz Capitulino 9971cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 9981cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 9991cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 10001cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 10011cac6f2cSLuiz Capitulino nr_nodes--) 10021cac6f2cSLuiz Capitulino 1003e1073d1eSAneesh Kumar K.V #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 1004944d9fecSLuiz Capitulino static void destroy_compound_gigantic_page(struct page *page, 1005d00181b9SKirill A. Shutemov unsigned int order) 1006944d9fecSLuiz Capitulino { 1007944d9fecSLuiz Capitulino int i; 1008944d9fecSLuiz Capitulino int nr_pages = 1 << order; 1009944d9fecSLuiz Capitulino struct page *p = page + 1; 1010944d9fecSLuiz Capitulino 1011c8cc708aSGerald Schaefer atomic_set(compound_mapcount_ptr(page), 0); 1012944d9fecSLuiz Capitulino for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 10131d798ca3SKirill A. Shutemov clear_compound_head(p); 1014944d9fecSLuiz Capitulino set_page_refcounted(p); 1015944d9fecSLuiz Capitulino } 1016944d9fecSLuiz Capitulino 1017944d9fecSLuiz Capitulino set_compound_order(page, 0); 1018944d9fecSLuiz Capitulino __ClearPageHead(page); 1019944d9fecSLuiz Capitulino } 1020944d9fecSLuiz Capitulino 1021d00181b9SKirill A. Shutemov static void free_gigantic_page(struct page *page, unsigned int order) 1022944d9fecSLuiz Capitulino { 1023944d9fecSLuiz Capitulino free_contig_range(page_to_pfn(page), 1 << order); 1024944d9fecSLuiz Capitulino } 1025944d9fecSLuiz Capitulino 10264eb0716eSAlexandre Ghiti #ifdef CONFIG_CONTIG_ALLOC 1027d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 1028d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) 1029944d9fecSLuiz Capitulino { 10305e27a2dfSAnshuman Khandual unsigned long nr_pages = 1UL << huge_page_order(h); 1031944d9fecSLuiz Capitulino 10325e27a2dfSAnshuman Khandual return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); 1033944d9fecSLuiz Capitulino } 1034944d9fecSLuiz Capitulino 1035944d9fecSLuiz Capitulino static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); 1036d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order); 10374eb0716eSAlexandre Ghiti #else /* !CONFIG_CONTIG_ALLOC */ 10384eb0716eSAlexandre Ghiti static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 10394eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 10404eb0716eSAlexandre Ghiti { 10414eb0716eSAlexandre Ghiti return NULL; 10424eb0716eSAlexandre Ghiti } 10434eb0716eSAlexandre Ghiti #endif /* CONFIG_CONTIG_ALLOC */ 1044944d9fecSLuiz Capitulino 1045e1073d1eSAneesh Kumar K.V #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 1046d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 10474eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 10484eb0716eSAlexandre Ghiti { 10494eb0716eSAlexandre Ghiti return NULL; 10504eb0716eSAlexandre Ghiti } 1051d00181b9SKirill A. Shutemov static inline void free_gigantic_page(struct page *page, unsigned int order) { } 1052944d9fecSLuiz Capitulino static inline void destroy_compound_gigantic_page(struct page *page, 1053d00181b9SKirill A. Shutemov unsigned int order) { } 1054944d9fecSLuiz Capitulino #endif 1055944d9fecSLuiz Capitulino 1056a5516438SAndi Kleen static void update_and_free_page(struct hstate *h, struct page *page) 10576af2acb6SAdam Litke { 10586af2acb6SAdam Litke int i; 1059a5516438SAndi Kleen 10604eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1061944d9fecSLuiz Capitulino return; 106218229df5SAndy Whitcroft 1063a5516438SAndi Kleen h->nr_huge_pages--; 1064a5516438SAndi Kleen h->nr_huge_pages_node[page_to_nid(page)]--; 1065a5516438SAndi Kleen for (i = 0; i < pages_per_huge_page(h); i++) { 106632f84528SChris Forbes page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 106732f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1068a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1069a7407a27SLuiz Capitulino 1 << PG_writeback); 10706af2acb6SAdam Litke } 1071309381feSSasha Levin VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 1072f1e61557SKirill A. Shutemov set_compound_page_dtor(page, NULL_COMPOUND_DTOR); 10736af2acb6SAdam Litke set_page_refcounted(page); 1074944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) { 1075944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 1076944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 1077944d9fecSLuiz Capitulino } else { 1078a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 10796af2acb6SAdam Litke } 1080944d9fecSLuiz Capitulino } 10816af2acb6SAdam Litke 1082e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1083e5ff2159SAndi Kleen { 1084e5ff2159SAndi Kleen struct hstate *h; 1085e5ff2159SAndi Kleen 1086e5ff2159SAndi Kleen for_each_hstate(h) { 1087e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1088e5ff2159SAndi Kleen return h; 1089e5ff2159SAndi Kleen } 1090e5ff2159SAndi Kleen return NULL; 1091e5ff2159SAndi Kleen } 1092e5ff2159SAndi Kleen 1093bcc54222SNaoya Horiguchi /* 1094bcc54222SNaoya Horiguchi * Test to determine whether the hugepage is "active/in-use" (i.e. being linked 1095bcc54222SNaoya Horiguchi * to hstate->hugepage_activelist.) 1096bcc54222SNaoya Horiguchi * 1097bcc54222SNaoya Horiguchi * This function can be called for tail pages, but never returns true for them. 1098bcc54222SNaoya Horiguchi */ 1099bcc54222SNaoya Horiguchi bool page_huge_active(struct page *page) 1100bcc54222SNaoya Horiguchi { 1101bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHuge(page), page); 1102bcc54222SNaoya Horiguchi return PageHead(page) && PagePrivate(&page[1]); 1103bcc54222SNaoya Horiguchi } 1104bcc54222SNaoya Horiguchi 1105bcc54222SNaoya Horiguchi /* never called for tail page */ 1106bcc54222SNaoya Horiguchi static void set_page_huge_active(struct page *page) 1107bcc54222SNaoya Horiguchi { 1108bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 1109bcc54222SNaoya Horiguchi SetPagePrivate(&page[1]); 1110bcc54222SNaoya Horiguchi } 1111bcc54222SNaoya Horiguchi 1112bcc54222SNaoya Horiguchi static void clear_page_huge_active(struct page *page) 1113bcc54222SNaoya Horiguchi { 1114bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 1115bcc54222SNaoya Horiguchi ClearPagePrivate(&page[1]); 1116bcc54222SNaoya Horiguchi } 1117bcc54222SNaoya Horiguchi 1118ab5ac90aSMichal Hocko /* 1119ab5ac90aSMichal Hocko * Internal hugetlb specific page flag. Do not use outside of the hugetlb 1120ab5ac90aSMichal Hocko * code 1121ab5ac90aSMichal Hocko */ 1122ab5ac90aSMichal Hocko static inline bool PageHugeTemporary(struct page *page) 1123ab5ac90aSMichal Hocko { 1124ab5ac90aSMichal Hocko if (!PageHuge(page)) 1125ab5ac90aSMichal Hocko return false; 1126ab5ac90aSMichal Hocko 1127ab5ac90aSMichal Hocko return (unsigned long)page[2].mapping == -1U; 1128ab5ac90aSMichal Hocko } 1129ab5ac90aSMichal Hocko 1130ab5ac90aSMichal Hocko static inline void SetPageHugeTemporary(struct page *page) 1131ab5ac90aSMichal Hocko { 1132ab5ac90aSMichal Hocko page[2].mapping = (void *)-1U; 1133ab5ac90aSMichal Hocko } 1134ab5ac90aSMichal Hocko 1135ab5ac90aSMichal Hocko static inline void ClearPageHugeTemporary(struct page *page) 1136ab5ac90aSMichal Hocko { 1137ab5ac90aSMichal Hocko page[2].mapping = NULL; 1138ab5ac90aSMichal Hocko } 1139ab5ac90aSMichal Hocko 1140c77c0a8aSWaiman Long static void __free_huge_page(struct page *page) 114127a85ef1SDavid Gibson { 1142a5516438SAndi Kleen /* 1143a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1144a5516438SAndi Kleen * compound page destructor. 1145a5516438SAndi Kleen */ 1146e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 11477893d1d5SAdam Litke int nid = page_to_nid(page); 114890481622SDavid Gibson struct hugepage_subpool *spool = 114990481622SDavid Gibson (struct hugepage_subpool *)page_private(page); 115007443a85SJoonsoo Kim bool restore_reserve; 115127a85ef1SDavid Gibson 1152b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1153b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_mapcount(page), page); 11548ace22bcSYongkai Wu 11558ace22bcSYongkai Wu set_page_private(page, 0); 11568ace22bcSYongkai Wu page->mapping = NULL; 115707443a85SJoonsoo Kim restore_reserve = PagePrivate(page); 115816c794b4SJoonsoo Kim ClearPagePrivate(page); 115927a85ef1SDavid Gibson 11601c5ecae3SMike Kravetz /* 11610919e1b6SMike Kravetz * If PagePrivate() was set on page, page allocation consumed a 11620919e1b6SMike Kravetz * reservation. If the page was associated with a subpool, there 11630919e1b6SMike Kravetz * would have been a page reserved in the subpool before allocation 11640919e1b6SMike Kravetz * via hugepage_subpool_get_pages(). Since we are 'restoring' the 11650919e1b6SMike Kravetz * reservtion, do not call hugepage_subpool_put_pages() as this will 11660919e1b6SMike Kravetz * remove the reserved page from the subpool. 11670919e1b6SMike Kravetz */ 11680919e1b6SMike Kravetz if (!restore_reserve) { 11690919e1b6SMike Kravetz /* 11700919e1b6SMike Kravetz * A return code of zero implies that the subpool will be 11710919e1b6SMike Kravetz * under its minimum size if the reservation is not restored 11720919e1b6SMike Kravetz * after page is free. Therefore, force restore_reserve 11730919e1b6SMike Kravetz * operation. 11741c5ecae3SMike Kravetz */ 11751c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 11761c5ecae3SMike Kravetz restore_reserve = true; 11770919e1b6SMike Kravetz } 11781c5ecae3SMike Kravetz 117927a85ef1SDavid Gibson spin_lock(&hugetlb_lock); 1180bcc54222SNaoya Horiguchi clear_page_huge_active(page); 11816d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 11826d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 118307443a85SJoonsoo Kim if (restore_reserve) 118407443a85SJoonsoo Kim h->resv_huge_pages++; 118507443a85SJoonsoo Kim 1186ab5ac90aSMichal Hocko if (PageHugeTemporary(page)) { 1187ab5ac90aSMichal Hocko list_del(&page->lru); 1188ab5ac90aSMichal Hocko ClearPageHugeTemporary(page); 1189ab5ac90aSMichal Hocko update_and_free_page(h, page); 1190ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 11910edaecfaSAneesh Kumar K.V /* remove the page from active list */ 11920edaecfaSAneesh Kumar K.V list_del(&page->lru); 1193a5516438SAndi Kleen update_and_free_page(h, page); 1194a5516438SAndi Kleen h->surplus_huge_pages--; 1195a5516438SAndi Kleen h->surplus_huge_pages_node[nid]--; 11967893d1d5SAdam Litke } else { 11975d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1198a5516438SAndi Kleen enqueue_huge_page(h, page); 11997893d1d5SAdam Litke } 120027a85ef1SDavid Gibson spin_unlock(&hugetlb_lock); 120127a85ef1SDavid Gibson } 120227a85ef1SDavid Gibson 1203c77c0a8aSWaiman Long /* 1204c77c0a8aSWaiman Long * As free_huge_page() can be called from a non-task context, we have 1205c77c0a8aSWaiman Long * to defer the actual freeing in a workqueue to prevent potential 1206c77c0a8aSWaiman Long * hugetlb_lock deadlock. 1207c77c0a8aSWaiman Long * 1208c77c0a8aSWaiman Long * free_hpage_workfn() locklessly retrieves the linked list of pages to 1209c77c0a8aSWaiman Long * be freed and frees them one-by-one. As the page->mapping pointer is 1210c77c0a8aSWaiman Long * going to be cleared in __free_huge_page() anyway, it is reused as the 1211c77c0a8aSWaiman Long * llist_node structure of a lockless linked list of huge pages to be freed. 1212c77c0a8aSWaiman Long */ 1213c77c0a8aSWaiman Long static LLIST_HEAD(hpage_freelist); 1214c77c0a8aSWaiman Long 1215c77c0a8aSWaiman Long static void free_hpage_workfn(struct work_struct *work) 1216c77c0a8aSWaiman Long { 1217c77c0a8aSWaiman Long struct llist_node *node; 1218c77c0a8aSWaiman Long struct page *page; 1219c77c0a8aSWaiman Long 1220c77c0a8aSWaiman Long node = llist_del_all(&hpage_freelist); 1221c77c0a8aSWaiman Long 1222c77c0a8aSWaiman Long while (node) { 1223c77c0a8aSWaiman Long page = container_of((struct address_space **)node, 1224c77c0a8aSWaiman Long struct page, mapping); 1225c77c0a8aSWaiman Long node = node->next; 1226c77c0a8aSWaiman Long __free_huge_page(page); 1227c77c0a8aSWaiman Long } 1228c77c0a8aSWaiman Long } 1229c77c0a8aSWaiman Long static DECLARE_WORK(free_hpage_work, free_hpage_workfn); 1230c77c0a8aSWaiman Long 1231c77c0a8aSWaiman Long void free_huge_page(struct page *page) 1232c77c0a8aSWaiman Long { 1233c77c0a8aSWaiman Long /* 1234c77c0a8aSWaiman Long * Defer freeing if in non-task context to avoid hugetlb_lock deadlock. 1235c77c0a8aSWaiman Long */ 1236c77c0a8aSWaiman Long if (!in_task()) { 1237c77c0a8aSWaiman Long /* 1238c77c0a8aSWaiman Long * Only call schedule_work() if hpage_freelist is previously 1239c77c0a8aSWaiman Long * empty. Otherwise, schedule_work() had been called but the 1240c77c0a8aSWaiman Long * workfn hasn't retrieved the list yet. 1241c77c0a8aSWaiman Long */ 1242c77c0a8aSWaiman Long if (llist_add((struct llist_node *)&page->mapping, 1243c77c0a8aSWaiman Long &hpage_freelist)) 1244c77c0a8aSWaiman Long schedule_work(&free_hpage_work); 1245c77c0a8aSWaiman Long return; 1246c77c0a8aSWaiman Long } 1247c77c0a8aSWaiman Long 1248c77c0a8aSWaiman Long __free_huge_page(page); 1249c77c0a8aSWaiman Long } 1250c77c0a8aSWaiman Long 1251a5516438SAndi Kleen static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1252b7ba30c6SAndi Kleen { 12530edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1254f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1255b7ba30c6SAndi Kleen spin_lock(&hugetlb_lock); 12569dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 1257a5516438SAndi Kleen h->nr_huge_pages++; 1258a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 1259b7ba30c6SAndi Kleen spin_unlock(&hugetlb_lock); 1260b7ba30c6SAndi Kleen } 1261b7ba30c6SAndi Kleen 1262d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order) 126320a0307cSWu Fengguang { 126420a0307cSWu Fengguang int i; 126520a0307cSWu Fengguang int nr_pages = 1 << order; 126620a0307cSWu Fengguang struct page *p = page + 1; 126720a0307cSWu Fengguang 126820a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 126920a0307cSWu Fengguang set_compound_order(page, order); 1270ef5a22beSAndrea Arcangeli __ClearPageReserved(page); 1271de09d31dSKirill A. Shutemov __SetPageHead(page); 127220a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1273ef5a22beSAndrea Arcangeli /* 1274ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1275ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1276ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 1277ef5a22beSAndrea Arcangeli * too. Otherwse drivers using get_user_pages() to access tail 1278ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1279ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1280ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1281ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1282ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1283ef5a22beSAndrea Arcangeli * after get_user_pages(). 1284ef5a22beSAndrea Arcangeli */ 1285ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 128658a84aa9SYouquan Song set_page_count(p, 0); 12871d798ca3SKirill A. Shutemov set_compound_head(p, page); 128820a0307cSWu Fengguang } 1289b4330afbSMike Kravetz atomic_set(compound_mapcount_ptr(page), -1); 129020a0307cSWu Fengguang } 129120a0307cSWu Fengguang 12927795912cSAndrew Morton /* 12937795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 12947795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 12957795912cSAndrew Morton * details. 12967795912cSAndrew Morton */ 129720a0307cSWu Fengguang int PageHuge(struct page *page) 129820a0307cSWu Fengguang { 129920a0307cSWu Fengguang if (!PageCompound(page)) 130020a0307cSWu Fengguang return 0; 130120a0307cSWu Fengguang 130220a0307cSWu Fengguang page = compound_head(page); 1303f1e61557SKirill A. Shutemov return page[1].compound_dtor == HUGETLB_PAGE_DTOR; 130420a0307cSWu Fengguang } 130543131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 130643131e14SNaoya Horiguchi 130727c73ae7SAndrea Arcangeli /* 130827c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 130927c73ae7SAndrea Arcangeli * normal or transparent huge pages. 131027c73ae7SAndrea Arcangeli */ 131127c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 131227c73ae7SAndrea Arcangeli { 131327c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 131427c73ae7SAndrea Arcangeli return 0; 131527c73ae7SAndrea Arcangeli 1316758f66a2SAndrew Morton return get_compound_page_dtor(page_head) == free_huge_page; 131727c73ae7SAndrea Arcangeli } 131827c73ae7SAndrea Arcangeli 131913d60f4bSZhang Yi pgoff_t __basepage_index(struct page *page) 132013d60f4bSZhang Yi { 132113d60f4bSZhang Yi struct page *page_head = compound_head(page); 132213d60f4bSZhang Yi pgoff_t index = page_index(page_head); 132313d60f4bSZhang Yi unsigned long compound_idx; 132413d60f4bSZhang Yi 132513d60f4bSZhang Yi if (!PageHuge(page_head)) 132613d60f4bSZhang Yi return page_index(page); 132713d60f4bSZhang Yi 132813d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 132913d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 133013d60f4bSZhang Yi else 133113d60f4bSZhang Yi compound_idx = page - page_head; 133213d60f4bSZhang Yi 133313d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 133413d60f4bSZhang Yi } 133513d60f4bSZhang Yi 13360c397daeSMichal Hocko static struct page *alloc_buddy_huge_page(struct hstate *h, 1337f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1338f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 13391da177e4SLinus Torvalds { 1340af0fb9dfSMichal Hocko int order = huge_page_order(h); 13411da177e4SLinus Torvalds struct page *page; 1342f60858f9SMike Kravetz bool alloc_try_hard = true; 1343f96efd58SJoe Jin 1344f60858f9SMike Kravetz /* 1345f60858f9SMike Kravetz * By default we always try hard to allocate the page with 1346f60858f9SMike Kravetz * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in 1347f60858f9SMike Kravetz * a loop (to adjust global huge page counts) and previous allocation 1348f60858f9SMike Kravetz * failed, do not continue to try hard on the same node. Use the 1349f60858f9SMike Kravetz * node_alloc_noretry bitmap to manage this state information. 1350f60858f9SMike Kravetz */ 1351f60858f9SMike Kravetz if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) 1352f60858f9SMike Kravetz alloc_try_hard = false; 1353f60858f9SMike Kravetz gfp_mask |= __GFP_COMP|__GFP_NOWARN; 1354f60858f9SMike Kravetz if (alloc_try_hard) 1355f60858f9SMike Kravetz gfp_mask |= __GFP_RETRY_MAYFAIL; 1356af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 1357af0fb9dfSMichal Hocko nid = numa_mem_id(); 1358af0fb9dfSMichal Hocko page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask); 1359af0fb9dfSMichal Hocko if (page) 1360af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC); 1361af0fb9dfSMichal Hocko else 1362af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 136363b4613cSNishanth Aravamudan 1364f60858f9SMike Kravetz /* 1365f60858f9SMike Kravetz * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this 1366f60858f9SMike Kravetz * indicates an overall state change. Clear bit so that we resume 1367f60858f9SMike Kravetz * normal 'try hard' allocations. 1368f60858f9SMike Kravetz */ 1369f60858f9SMike Kravetz if (node_alloc_noretry && page && !alloc_try_hard) 1370f60858f9SMike Kravetz node_clear(nid, *node_alloc_noretry); 1371f60858f9SMike Kravetz 1372f60858f9SMike Kravetz /* 1373f60858f9SMike Kravetz * If we tried hard to get a page but failed, set bit so that 1374f60858f9SMike Kravetz * subsequent attempts will not try as hard until there is an 1375f60858f9SMike Kravetz * overall state change. 1376f60858f9SMike Kravetz */ 1377f60858f9SMike Kravetz if (node_alloc_noretry && !page && alloc_try_hard) 1378f60858f9SMike Kravetz node_set(nid, *node_alloc_noretry); 1379f60858f9SMike Kravetz 138063b4613cSNishanth Aravamudan return page; 138163b4613cSNishanth Aravamudan } 138263b4613cSNishanth Aravamudan 1383af0fb9dfSMichal Hocko /* 13840c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 13850c397daeSMichal Hocko * should use this function to get new hugetlb pages 13860c397daeSMichal Hocko */ 13870c397daeSMichal Hocko static struct page *alloc_fresh_huge_page(struct hstate *h, 1388f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1389f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 13900c397daeSMichal Hocko { 13910c397daeSMichal Hocko struct page *page; 13920c397daeSMichal Hocko 13930c397daeSMichal Hocko if (hstate_is_gigantic(h)) 13940c397daeSMichal Hocko page = alloc_gigantic_page(h, gfp_mask, nid, nmask); 13950c397daeSMichal Hocko else 13960c397daeSMichal Hocko page = alloc_buddy_huge_page(h, gfp_mask, 1397f60858f9SMike Kravetz nid, nmask, node_alloc_noretry); 13980c397daeSMichal Hocko if (!page) 13990c397daeSMichal Hocko return NULL; 14000c397daeSMichal Hocko 14010c397daeSMichal Hocko if (hstate_is_gigantic(h)) 14020c397daeSMichal Hocko prep_compound_gigantic_page(page, huge_page_order(h)); 14030c397daeSMichal Hocko prep_new_huge_page(h, page, page_to_nid(page)); 14040c397daeSMichal Hocko 14050c397daeSMichal Hocko return page; 14060c397daeSMichal Hocko } 14070c397daeSMichal Hocko 14080c397daeSMichal Hocko /* 1409af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 1410af0fb9dfSMichal Hocko * manner. 1411af0fb9dfSMichal Hocko */ 1412f60858f9SMike Kravetz static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 1413f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 1414b2261026SJoonsoo Kim { 1415b2261026SJoonsoo Kim struct page *page; 1416b2261026SJoonsoo Kim int nr_nodes, node; 1417af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 1418b2261026SJoonsoo Kim 1419b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1420f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, 1421f60858f9SMike Kravetz node_alloc_noretry); 1422af0fb9dfSMichal Hocko if (page) 1423b2261026SJoonsoo Kim break; 1424b2261026SJoonsoo Kim } 1425b2261026SJoonsoo Kim 1426af0fb9dfSMichal Hocko if (!page) 1427af0fb9dfSMichal Hocko return 0; 1428b2261026SJoonsoo Kim 1429af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 1430af0fb9dfSMichal Hocko 1431af0fb9dfSMichal Hocko return 1; 1432b2261026SJoonsoo Kim } 1433b2261026SJoonsoo Kim 1434e8c5c824SLee Schermerhorn /* 1435e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 1436e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 1437e8c5c824SLee Schermerhorn * balanced over allowed nodes. 1438e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 1439e8c5c824SLee Schermerhorn */ 14406ae11b27SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 14416ae11b27SLee Schermerhorn bool acct_surplus) 1442e8c5c824SLee Schermerhorn { 1443b2261026SJoonsoo Kim int nr_nodes, node; 1444e8c5c824SLee Schermerhorn int ret = 0; 1445e8c5c824SLee Schermerhorn 1446b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1447685f3457SLee Schermerhorn /* 1448685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 1449685f3457SLee Schermerhorn * nodes with surplus pages. 1450685f3457SLee Schermerhorn */ 1451b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 1452b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 1453e8c5c824SLee Schermerhorn struct page *page = 1454b2261026SJoonsoo Kim list_entry(h->hugepage_freelists[node].next, 1455e8c5c824SLee Schermerhorn struct page, lru); 1456e8c5c824SLee Schermerhorn list_del(&page->lru); 1457e8c5c824SLee Schermerhorn h->free_huge_pages--; 1458b2261026SJoonsoo Kim h->free_huge_pages_node[node]--; 1459685f3457SLee Schermerhorn if (acct_surplus) { 1460685f3457SLee Schermerhorn h->surplus_huge_pages--; 1461b2261026SJoonsoo Kim h->surplus_huge_pages_node[node]--; 1462685f3457SLee Schermerhorn } 1463e8c5c824SLee Schermerhorn update_and_free_page(h, page); 1464e8c5c824SLee Schermerhorn ret = 1; 14659a76db09SLee Schermerhorn break; 1466e8c5c824SLee Schermerhorn } 1467b2261026SJoonsoo Kim } 1468e8c5c824SLee Schermerhorn 1469e8c5c824SLee Schermerhorn return ret; 1470e8c5c824SLee Schermerhorn } 1471e8c5c824SLee Schermerhorn 1472c8721bbbSNaoya Horiguchi /* 1473c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 1474faf53defSNaoya Horiguchi * nothing for in-use hugepages and non-hugepages. 1475faf53defSNaoya Horiguchi * This function returns values like below: 1476faf53defSNaoya Horiguchi * 1477faf53defSNaoya Horiguchi * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 1478faf53defSNaoya Horiguchi * (allocated or reserved.) 1479faf53defSNaoya Horiguchi * 0: successfully dissolved free hugepages or the page is not a 1480faf53defSNaoya Horiguchi * hugepage (considered as already dissolved) 1481c8721bbbSNaoya Horiguchi */ 1482c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 1483c8721bbbSNaoya Horiguchi { 14846bc9b564SNaoya Horiguchi int rc = -EBUSY; 1485082d5b6bSGerald Schaefer 1486faf53defSNaoya Horiguchi /* Not to disrupt normal path by vainly holding hugetlb_lock */ 1487faf53defSNaoya Horiguchi if (!PageHuge(page)) 1488faf53defSNaoya Horiguchi return 0; 1489faf53defSNaoya Horiguchi 1490c8721bbbSNaoya Horiguchi spin_lock(&hugetlb_lock); 1491faf53defSNaoya Horiguchi if (!PageHuge(page)) { 1492faf53defSNaoya Horiguchi rc = 0; 1493faf53defSNaoya Horiguchi goto out; 1494faf53defSNaoya Horiguchi } 1495faf53defSNaoya Horiguchi 1496faf53defSNaoya Horiguchi if (!page_count(page)) { 14972247bb33SGerald Schaefer struct page *head = compound_head(page); 14982247bb33SGerald Schaefer struct hstate *h = page_hstate(head); 14992247bb33SGerald Schaefer int nid = page_to_nid(head); 15006bc9b564SNaoya Horiguchi if (h->free_huge_pages - h->resv_huge_pages == 0) 1501082d5b6bSGerald Schaefer goto out; 1502c3114a84SAnshuman Khandual /* 1503c3114a84SAnshuman Khandual * Move PageHWPoison flag from head page to the raw error page, 1504c3114a84SAnshuman Khandual * which makes any subpages rather than the error page reusable. 1505c3114a84SAnshuman Khandual */ 1506c3114a84SAnshuman Khandual if (PageHWPoison(head) && page != head) { 1507c3114a84SAnshuman Khandual SetPageHWPoison(page); 1508c3114a84SAnshuman Khandual ClearPageHWPoison(head); 1509c3114a84SAnshuman Khandual } 15102247bb33SGerald Schaefer list_del(&head->lru); 1511c8721bbbSNaoya Horiguchi h->free_huge_pages--; 1512c8721bbbSNaoya Horiguchi h->free_huge_pages_node[nid]--; 1513c1470b33Szhong jiang h->max_huge_pages--; 15142247bb33SGerald Schaefer update_and_free_page(h, head); 15156bc9b564SNaoya Horiguchi rc = 0; 1516c8721bbbSNaoya Horiguchi } 1517082d5b6bSGerald Schaefer out: 1518c8721bbbSNaoya Horiguchi spin_unlock(&hugetlb_lock); 1519082d5b6bSGerald Schaefer return rc; 1520c8721bbbSNaoya Horiguchi } 1521c8721bbbSNaoya Horiguchi 1522c8721bbbSNaoya Horiguchi /* 1523c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 1524c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 15252247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 15262247bb33SGerald Schaefer * part of it lies within the given range. 1527082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 1528082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 1529c8721bbbSNaoya Horiguchi */ 1530082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 1531c8721bbbSNaoya Horiguchi { 1532c8721bbbSNaoya Horiguchi unsigned long pfn; 1533eb03aa00SGerald Schaefer struct page *page; 1534082d5b6bSGerald Schaefer int rc = 0; 1535c8721bbbSNaoya Horiguchi 1536d0177639SLi Zhong if (!hugepages_supported()) 1537082d5b6bSGerald Schaefer return rc; 1538d0177639SLi Zhong 1539eb03aa00SGerald Schaefer for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { 1540eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 1541eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 1542eb03aa00SGerald Schaefer if (rc) 1543082d5b6bSGerald Schaefer break; 1544eb03aa00SGerald Schaefer } 1545082d5b6bSGerald Schaefer 1546082d5b6bSGerald Schaefer return rc; 1547c8721bbbSNaoya Horiguchi } 1548c8721bbbSNaoya Horiguchi 1549ab5ac90aSMichal Hocko /* 1550ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 1551ab5ac90aSMichal Hocko */ 15520c397daeSMichal Hocko static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, 1553aaf14e40SMichal Hocko int nid, nodemask_t *nmask) 15547893d1d5SAdam Litke { 15559980d744SMichal Hocko struct page *page = NULL; 15567893d1d5SAdam Litke 1557bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1558aa888a74SAndi Kleen return NULL; 1559aa888a74SAndi Kleen 1560d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 15619980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 15629980d744SMichal Hocko goto out_unlock; 1563d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1564d1c3fb1fSNishanth Aravamudan 1565f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 15669980d744SMichal Hocko if (!page) 15670c397daeSMichal Hocko return NULL; 1568d1c3fb1fSNishanth Aravamudan 15697893d1d5SAdam Litke spin_lock(&hugetlb_lock); 15709980d744SMichal Hocko /* 15719980d744SMichal Hocko * We could have raced with the pool size change. 15729980d744SMichal Hocko * Double check that and simply deallocate the new page 15739980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 15749980d744SMichal Hocko * temporary page to workaround the nasty free_huge_page 15759980d744SMichal Hocko * codeflow 15769980d744SMichal Hocko */ 15779980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 15789980d744SMichal Hocko SetPageHugeTemporary(page); 15792bf753e6SKai Shen spin_unlock(&hugetlb_lock); 15809980d744SMichal Hocko put_page(page); 15812bf753e6SKai Shen return NULL; 15829980d744SMichal Hocko } else { 15839980d744SMichal Hocko h->surplus_huge_pages++; 15844704dea3SMichal Hocko h->surplus_huge_pages_node[page_to_nid(page)]++; 15857893d1d5SAdam Litke } 15869980d744SMichal Hocko 15879980d744SMichal Hocko out_unlock: 1588d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 15897893d1d5SAdam Litke 15907893d1d5SAdam Litke return page; 15917893d1d5SAdam Litke } 15927893d1d5SAdam Litke 15939a4e9f3bSAneesh Kumar K.V struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, 1594ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 1595ab5ac90aSMichal Hocko { 1596ab5ac90aSMichal Hocko struct page *page; 1597ab5ac90aSMichal Hocko 1598ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 1599ab5ac90aSMichal Hocko return NULL; 1600ab5ac90aSMichal Hocko 1601f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 1602ab5ac90aSMichal Hocko if (!page) 1603ab5ac90aSMichal Hocko return NULL; 1604ab5ac90aSMichal Hocko 1605ab5ac90aSMichal Hocko /* 1606ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 1607ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 1608ab5ac90aSMichal Hocko */ 1609ab5ac90aSMichal Hocko SetPageHugeTemporary(page); 1610ab5ac90aSMichal Hocko 1611ab5ac90aSMichal Hocko return page; 1612ab5ac90aSMichal Hocko } 1613ab5ac90aSMichal Hocko 1614e4e574b7SAdam Litke /* 1615099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 1616099730d6SDave Hansen */ 1617e0ec90eeSDave Hansen static 16180c397daeSMichal Hocko struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, 1619099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 1620099730d6SDave Hansen { 1621aaf14e40SMichal Hocko struct page *page; 1622aaf14e40SMichal Hocko struct mempolicy *mpol; 1623aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 1624aaf14e40SMichal Hocko int nid; 1625aaf14e40SMichal Hocko nodemask_t *nodemask; 1626aaf14e40SMichal Hocko 1627aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 16280c397daeSMichal Hocko page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); 1629aaf14e40SMichal Hocko mpol_cond_put(mpol); 1630aaf14e40SMichal Hocko 1631aaf14e40SMichal Hocko return page; 1632099730d6SDave Hansen } 1633099730d6SDave Hansen 1634ab5ac90aSMichal Hocko /* page migration callback function */ 1635bf50bab2SNaoya Horiguchi struct page *alloc_huge_page_node(struct hstate *h, int nid) 1636bf50bab2SNaoya Horiguchi { 1637aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 16384ef91848SJoonsoo Kim struct page *page = NULL; 1639bf50bab2SNaoya Horiguchi 1640aaf14e40SMichal Hocko if (nid != NUMA_NO_NODE) 1641aaf14e40SMichal Hocko gfp_mask |= __GFP_THISNODE; 1642aaf14e40SMichal Hocko 1643bf50bab2SNaoya Horiguchi spin_lock(&hugetlb_lock); 16444ef91848SJoonsoo Kim if (h->free_huge_pages - h->resv_huge_pages > 0) 16453e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL); 1646bf50bab2SNaoya Horiguchi spin_unlock(&hugetlb_lock); 1647bf50bab2SNaoya Horiguchi 164894ae8ba7SAneesh Kumar K.V if (!page) 16490c397daeSMichal Hocko page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL); 1650bf50bab2SNaoya Horiguchi 1651bf50bab2SNaoya Horiguchi return page; 1652bf50bab2SNaoya Horiguchi } 1653bf50bab2SNaoya Horiguchi 1654ab5ac90aSMichal Hocko /* page migration callback function */ 16553e59fcb0SMichal Hocko struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, 16563e59fcb0SMichal Hocko nodemask_t *nmask) 16574db9b2efSMichal Hocko { 1658aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 16594db9b2efSMichal Hocko 16604db9b2efSMichal Hocko spin_lock(&hugetlb_lock); 16614db9b2efSMichal Hocko if (h->free_huge_pages - h->resv_huge_pages > 0) { 16623e59fcb0SMichal Hocko struct page *page; 16633e59fcb0SMichal Hocko 16643e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); 16653e59fcb0SMichal Hocko if (page) { 16663e59fcb0SMichal Hocko spin_unlock(&hugetlb_lock); 16673e59fcb0SMichal Hocko return page; 16684db9b2efSMichal Hocko } 16694db9b2efSMichal Hocko } 16704db9b2efSMichal Hocko spin_unlock(&hugetlb_lock); 16714db9b2efSMichal Hocko 16720c397daeSMichal Hocko return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); 16734db9b2efSMichal Hocko } 16744db9b2efSMichal Hocko 1675ebd63723SMichal Hocko /* mempolicy aware migration callback */ 1676389c8178SMichal Hocko struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, 1677389c8178SMichal Hocko unsigned long address) 1678ebd63723SMichal Hocko { 1679ebd63723SMichal Hocko struct mempolicy *mpol; 1680ebd63723SMichal Hocko nodemask_t *nodemask; 1681ebd63723SMichal Hocko struct page *page; 1682ebd63723SMichal Hocko gfp_t gfp_mask; 1683ebd63723SMichal Hocko int node; 1684ebd63723SMichal Hocko 1685ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 1686ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 1687ebd63723SMichal Hocko page = alloc_huge_page_nodemask(h, node, nodemask); 1688ebd63723SMichal Hocko mpol_cond_put(mpol); 1689ebd63723SMichal Hocko 1690ebd63723SMichal Hocko return page; 1691ebd63723SMichal Hocko } 1692ebd63723SMichal Hocko 1693bf50bab2SNaoya Horiguchi /* 169425985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 1695e4e574b7SAdam Litke * of size 'delta'. 1696e4e574b7SAdam Litke */ 1697a5516438SAndi Kleen static int gather_surplus_pages(struct hstate *h, int delta) 1698e4e574b7SAdam Litke { 1699e4e574b7SAdam Litke struct list_head surplus_list; 1700e4e574b7SAdam Litke struct page *page, *tmp; 1701e4e574b7SAdam Litke int ret, i; 1702e4e574b7SAdam Litke int needed, allocated; 170328073b02SHillf Danton bool alloc_ok = true; 1704e4e574b7SAdam Litke 1705a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 1706ac09b3a1SAdam Litke if (needed <= 0) { 1707a5516438SAndi Kleen h->resv_huge_pages += delta; 1708e4e574b7SAdam Litke return 0; 1709ac09b3a1SAdam Litke } 1710e4e574b7SAdam Litke 1711e4e574b7SAdam Litke allocated = 0; 1712e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 1713e4e574b7SAdam Litke 1714e4e574b7SAdam Litke ret = -ENOMEM; 1715e4e574b7SAdam Litke retry: 1716e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 1717e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 17180c397daeSMichal Hocko page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), 1719aaf14e40SMichal Hocko NUMA_NO_NODE, NULL); 172028073b02SHillf Danton if (!page) { 172128073b02SHillf Danton alloc_ok = false; 172228073b02SHillf Danton break; 172328073b02SHillf Danton } 1724e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 172569ed779aSDavid Rientjes cond_resched(); 1726e4e574b7SAdam Litke } 172728073b02SHillf Danton allocated += i; 1728e4e574b7SAdam Litke 1729e4e574b7SAdam Litke /* 1730e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 1731e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 1732e4e574b7SAdam Litke */ 1733e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 1734a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 1735a5516438SAndi Kleen (h->free_huge_pages + allocated); 173628073b02SHillf Danton if (needed > 0) { 173728073b02SHillf Danton if (alloc_ok) 1738e4e574b7SAdam Litke goto retry; 173928073b02SHillf Danton /* 174028073b02SHillf Danton * We were not able to allocate enough pages to 174128073b02SHillf Danton * satisfy the entire reservation so we free what 174228073b02SHillf Danton * we've allocated so far. 174328073b02SHillf Danton */ 174428073b02SHillf Danton goto free; 174528073b02SHillf Danton } 1746e4e574b7SAdam Litke /* 1747e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 174825985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 1749e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 1750ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 1751ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 1752ac09b3a1SAdam Litke * before they are reserved. 1753e4e574b7SAdam Litke */ 1754e4e574b7SAdam Litke needed += allocated; 1755a5516438SAndi Kleen h->resv_huge_pages += delta; 1756e4e574b7SAdam Litke ret = 0; 1757a9869b83SNaoya Horiguchi 175819fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 175919fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 176019fc3f0aSAdam Litke if ((--needed) < 0) 176119fc3f0aSAdam Litke break; 1762a9869b83SNaoya Horiguchi /* 1763a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 1764a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 1765a9869b83SNaoya Horiguchi */ 1766a9869b83SNaoya Horiguchi put_page_testzero(page); 1767309381feSSasha Levin VM_BUG_ON_PAGE(page_count(page), page); 1768a5516438SAndi Kleen enqueue_huge_page(h, page); 176919fc3f0aSAdam Litke } 177028073b02SHillf Danton free: 1771b0365c8dSHillf Danton spin_unlock(&hugetlb_lock); 177219fc3f0aSAdam Litke 177319fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 1774c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 1775a9869b83SNaoya Horiguchi put_page(page); 177619fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 1777e4e574b7SAdam Litke 1778e4e574b7SAdam Litke return ret; 1779e4e574b7SAdam Litke } 1780e4e574b7SAdam Litke 1781e4e574b7SAdam Litke /* 1782e5bbc8a6SMike Kravetz * This routine has two main purposes: 1783e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 1784e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 1785e5bbc8a6SMike Kravetz * to the associated reservation map. 1786e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 1787e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 1788e5bbc8a6SMike Kravetz * 1789e5bbc8a6SMike Kravetz * Called with hugetlb_lock held. However, the lock could be dropped (and 1790e5bbc8a6SMike Kravetz * reacquired) during calls to cond_resched_lock. Whenever dropping the lock, 1791e5bbc8a6SMike Kravetz * we must make sure nobody else can claim pages we are in the process of 1792e5bbc8a6SMike Kravetz * freeing. Do this by ensuring resv_huge_page always is greater than the 1793e5bbc8a6SMike Kravetz * number of huge pages we plan to free when dropping the lock. 1794e4e574b7SAdam Litke */ 1795a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 1796a5516438SAndi Kleen unsigned long unused_resv_pages) 1797e4e574b7SAdam Litke { 1798e4e574b7SAdam Litke unsigned long nr_pages; 1799e4e574b7SAdam Litke 1800aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 1801bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1802e5bbc8a6SMike Kravetz goto out; 1803aa888a74SAndi Kleen 1804e5bbc8a6SMike Kravetz /* 1805e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 1806e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 1807e5bbc8a6SMike Kravetz */ 1808a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 1809e4e574b7SAdam Litke 1810685f3457SLee Schermerhorn /* 1811685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 18129b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 18139b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 18149b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 18159b5e5d0fSLee Schermerhorn * free_pool_huge_page() will balance the the freed pages across the 18169b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 1817e5bbc8a6SMike Kravetz * 1818e5bbc8a6SMike Kravetz * Note that we decrement resv_huge_pages as we free the pages. If 1819e5bbc8a6SMike Kravetz * we drop the lock, resv_huge_pages will still be sufficiently large 1820e5bbc8a6SMike Kravetz * to cover subsequent pages we may free. 1821685f3457SLee Schermerhorn */ 1822685f3457SLee Schermerhorn while (nr_pages--) { 1823e5bbc8a6SMike Kravetz h->resv_huge_pages--; 1824e5bbc8a6SMike Kravetz unused_resv_pages--; 18258cebfcd0SLai Jiangshan if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) 1826e5bbc8a6SMike Kravetz goto out; 18277848a4bfSMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 1828e4e574b7SAdam Litke } 1829e5bbc8a6SMike Kravetz 1830e5bbc8a6SMike Kravetz out: 1831e5bbc8a6SMike Kravetz /* Fully uncommit the reservation */ 1832e5bbc8a6SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 1833e4e574b7SAdam Litke } 1834e4e574b7SAdam Litke 18355e911373SMike Kravetz 1836c37f9fb1SAndy Whitcroft /* 1837feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 18385e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 1839cf3ad20bSMike Kravetz * 1840cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 1841cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 1842cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 1843cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 1844cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 1845feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 1846feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 1847feba16e2SMike Kravetz * is called in such cases. 1848cf3ad20bSMike Kravetz * 1849cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 1850cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 1851cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 1852cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 1853cf3ad20bSMike Kravetz * take appropriate action. 185496b96a96SMike Kravetz * 185596b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 185696b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 185796b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 185896b96a96SMike Kravetz * reservation exists. 1859c37f9fb1SAndy Whitcroft */ 18605e911373SMike Kravetz enum vma_resv_mode { 18615e911373SMike Kravetz VMA_NEEDS_RESV, 18625e911373SMike Kravetz VMA_COMMIT_RESV, 1863feba16e2SMike Kravetz VMA_END_RESV, 186496b96a96SMike Kravetz VMA_ADD_RESV, 18655e911373SMike Kravetz }; 1866cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 1867cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 18685e911373SMike Kravetz enum vma_resv_mode mode) 1869c37f9fb1SAndy Whitcroft { 18704e35f483SJoonsoo Kim struct resv_map *resv; 18714e35f483SJoonsoo Kim pgoff_t idx; 1872cf3ad20bSMike Kravetz long ret; 1873c37f9fb1SAndy Whitcroft 18744e35f483SJoonsoo Kim resv = vma_resv_map(vma); 18754e35f483SJoonsoo Kim if (!resv) 1876c37f9fb1SAndy Whitcroft return 1; 1877c37f9fb1SAndy Whitcroft 18784e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 18795e911373SMike Kravetz switch (mode) { 18805e911373SMike Kravetz case VMA_NEEDS_RESV: 1881cf3ad20bSMike Kravetz ret = region_chg(resv, idx, idx + 1); 18825e911373SMike Kravetz break; 18835e911373SMike Kravetz case VMA_COMMIT_RESV: 18845e911373SMike Kravetz ret = region_add(resv, idx, idx + 1); 18855e911373SMike Kravetz break; 1886feba16e2SMike Kravetz case VMA_END_RESV: 18875e911373SMike Kravetz region_abort(resv, idx, idx + 1); 18885e911373SMike Kravetz ret = 0; 18895e911373SMike Kravetz break; 189096b96a96SMike Kravetz case VMA_ADD_RESV: 189196b96a96SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) 189296b96a96SMike Kravetz ret = region_add(resv, idx, idx + 1); 189396b96a96SMike Kravetz else { 189496b96a96SMike Kravetz region_abort(resv, idx, idx + 1); 189596b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 189696b96a96SMike Kravetz } 189796b96a96SMike Kravetz break; 18985e911373SMike Kravetz default: 18995e911373SMike Kravetz BUG(); 19005e911373SMike Kravetz } 190184afd99bSAndy Whitcroft 19024e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) 1903cf3ad20bSMike Kravetz return ret; 190467961f9dSMike Kravetz else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { 190567961f9dSMike Kravetz /* 190667961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 190767961f9dSMike Kravetz * However, a file associated with mapping could have been 190867961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 190967961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 191067961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 191167961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 191267961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 191367961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 191467961f9dSMike Kravetz * reservation has already been consumed. As a result, the 191567961f9dSMike Kravetz * return value of this routine is the opposite of the 191667961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 191767961f9dSMike Kravetz */ 191867961f9dSMike Kravetz if (ret) 191967961f9dSMike Kravetz return 0; 192067961f9dSMike Kravetz else 192167961f9dSMike Kravetz return 1; 192267961f9dSMike Kravetz } 19234e35f483SJoonsoo Kim else 1924cf3ad20bSMike Kravetz return ret < 0 ? ret : 0; 192584afd99bSAndy Whitcroft } 1926cf3ad20bSMike Kravetz 1927cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 1928a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 1929c37f9fb1SAndy Whitcroft { 19305e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 1931cf3ad20bSMike Kravetz } 1932c37f9fb1SAndy Whitcroft 1933cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 1934cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 1935cf3ad20bSMike Kravetz { 19365e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 19375e911373SMike Kravetz } 19385e911373SMike Kravetz 1939feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 19405e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 19415e911373SMike Kravetz { 1942feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 1943c37f9fb1SAndy Whitcroft } 1944c37f9fb1SAndy Whitcroft 194596b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 194696b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 194796b96a96SMike Kravetz { 194896b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 194996b96a96SMike Kravetz } 195096b96a96SMike Kravetz 195196b96a96SMike Kravetz /* 195296b96a96SMike Kravetz * This routine is called to restore a reservation on error paths. In the 195396b96a96SMike Kravetz * specific error paths, a huge page was allocated (via alloc_huge_page) 195496b96a96SMike Kravetz * and is about to be freed. If a reservation for the page existed, 195596b96a96SMike Kravetz * alloc_huge_page would have consumed the reservation and set PagePrivate 195696b96a96SMike Kravetz * in the newly allocated page. When the page is freed via free_huge_page, 195796b96a96SMike Kravetz * the global reservation count will be incremented if PagePrivate is set. 195896b96a96SMike Kravetz * However, free_huge_page can not adjust the reserve map. Adjust the 195996b96a96SMike Kravetz * reserve map here to be consistent with global reserve count adjustments 196096b96a96SMike Kravetz * to be made by free_huge_page. 196196b96a96SMike Kravetz */ 196296b96a96SMike Kravetz static void restore_reserve_on_error(struct hstate *h, 196396b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long address, 196496b96a96SMike Kravetz struct page *page) 196596b96a96SMike Kravetz { 196696b96a96SMike Kravetz if (unlikely(PagePrivate(page))) { 196796b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 196896b96a96SMike Kravetz 196996b96a96SMike Kravetz if (unlikely(rc < 0)) { 197096b96a96SMike Kravetz /* 197196b96a96SMike Kravetz * Rare out of memory condition in reserve map 197296b96a96SMike Kravetz * manipulation. Clear PagePrivate so that 197396b96a96SMike Kravetz * global reserve count will not be incremented 197496b96a96SMike Kravetz * by free_huge_page. This will make it appear 197596b96a96SMike Kravetz * as though the reservation for this page was 197696b96a96SMike Kravetz * consumed. This may prevent the task from 197796b96a96SMike Kravetz * faulting in the page at a later time. This 197896b96a96SMike Kravetz * is better than inconsistent global huge page 197996b96a96SMike Kravetz * accounting of reserve counts. 198096b96a96SMike Kravetz */ 198196b96a96SMike Kravetz ClearPagePrivate(page); 198296b96a96SMike Kravetz } else if (rc) { 198396b96a96SMike Kravetz rc = vma_add_reservation(h, vma, address); 198496b96a96SMike Kravetz if (unlikely(rc < 0)) 198596b96a96SMike Kravetz /* 198696b96a96SMike Kravetz * See above comment about rare out of 198796b96a96SMike Kravetz * memory condition. 198896b96a96SMike Kravetz */ 198996b96a96SMike Kravetz ClearPagePrivate(page); 199096b96a96SMike Kravetz } else 199196b96a96SMike Kravetz vma_end_reservation(h, vma, address); 199296b96a96SMike Kravetz } 199396b96a96SMike Kravetz } 199496b96a96SMike Kravetz 199570c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 199604f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 1997348ea204SAdam Litke { 199890481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 1999a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2000348ea204SAdam Litke struct page *page; 2001d85f69b0SMike Kravetz long map_chg, map_commit; 2002d85f69b0SMike Kravetz long gbl_chg; 20036d76dcf4SAneesh Kumar K.V int ret, idx; 20046d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 20052fc39cecSAdam Litke 20066d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 2007a1e78772SMel Gorman /* 2008d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 2009d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 2010d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 2011a1e78772SMel Gorman */ 2012d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 2013d85f69b0SMike Kravetz if (map_chg < 0) 201476dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 2015d85f69b0SMike Kravetz 2016d85f69b0SMike Kravetz /* 2017d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 2018d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 2019d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 2020d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 2021d85f69b0SMike Kravetz * checked against any subpool limit. 2022d85f69b0SMike Kravetz */ 2023d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 2024d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 2025d85f69b0SMike Kravetz if (gbl_chg < 0) { 2026feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 202776dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 20285e911373SMike Kravetz } 202990d8b7e6SAdam Litke 2030d85f69b0SMike Kravetz /* 2031d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 2032d85f69b0SMike Kravetz * map, there could be reservations associated with the 2033d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 2034d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 2035d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 2036d85f69b0SMike Kravetz * the subpool reservations. 2037d85f69b0SMike Kravetz */ 2038d85f69b0SMike Kravetz if (avoid_reserve) 2039d85f69b0SMike Kravetz gbl_chg = 1; 2040d85f69b0SMike Kravetz } 2041d85f69b0SMike Kravetz 20426d76dcf4SAneesh Kumar K.V ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 20438f34af6fSJianyu Zhan if (ret) 20448f34af6fSJianyu Zhan goto out_subpool_put; 20458f34af6fSJianyu Zhan 2046a1e78772SMel Gorman spin_lock(&hugetlb_lock); 2047d85f69b0SMike Kravetz /* 2048d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 2049d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 2050d85f69b0SMike Kravetz * a reservation exists for the allocation. 2051d85f69b0SMike Kravetz */ 2052d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 205381a6fcaeSJoonsoo Kim if (!page) { 205494ae8ba7SAneesh Kumar K.V spin_unlock(&hugetlb_lock); 20550c397daeSMichal Hocko page = alloc_buddy_huge_page_with_mpol(h, vma, addr); 20568f34af6fSJianyu Zhan if (!page) 20578f34af6fSJianyu Zhan goto out_uncharge_cgroup; 2058a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 2059a88c7695SNaoya Horiguchi SetPagePrivate(page); 2060a88c7695SNaoya Horiguchi h->resv_huge_pages--; 2061a88c7695SNaoya Horiguchi } 206279dbb236SAneesh Kumar K.V spin_lock(&hugetlb_lock); 206379dbb236SAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 206481a6fcaeSJoonsoo Kim /* Fall through */ 2065a1e78772SMel Gorman } 206681a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 206781a6fcaeSJoonsoo Kim spin_unlock(&hugetlb_lock); 2068a1e78772SMel Gorman 206990481622SDavid Gibson set_page_private(page, (unsigned long)spool); 2070a1e78772SMel Gorman 2071d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 2072d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 207333039678SMike Kravetz /* 207433039678SMike Kravetz * The page was added to the reservation map between 207533039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 207633039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 207733039678SMike Kravetz * Adjust for the subpool count incremented above AND 207833039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 207933039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 208033039678SMike Kravetz * no longer applies. 208133039678SMike Kravetz */ 208233039678SMike Kravetz long rsv_adjust; 208333039678SMike Kravetz 208433039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 208533039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 208633039678SMike Kravetz } 20877893d1d5SAdam Litke return page; 20888f34af6fSJianyu Zhan 20898f34af6fSJianyu Zhan out_uncharge_cgroup: 20908f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 20918f34af6fSJianyu Zhan out_subpool_put: 2092d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 20938f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 2094feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 20958f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 2096b45b5bd6SDavid Gibson } 2097b45b5bd6SDavid Gibson 2098e24a1307SAneesh Kumar K.V int alloc_bootmem_huge_page(struct hstate *h) 2099e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 2100e24a1307SAneesh Kumar K.V int __alloc_bootmem_huge_page(struct hstate *h) 2101aa888a74SAndi Kleen { 2102aa888a74SAndi Kleen struct huge_bootmem_page *m; 2103b2261026SJoonsoo Kim int nr_nodes, node; 2104aa888a74SAndi Kleen 2105b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 2106aa888a74SAndi Kleen void *addr; 2107aa888a74SAndi Kleen 2108eb31d559SMike Rapoport addr = memblock_alloc_try_nid_raw( 21098b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 211097ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 2111aa888a74SAndi Kleen if (addr) { 2112aa888a74SAndi Kleen /* 2113aa888a74SAndi Kleen * Use the beginning of the huge page to store the 2114aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 2115aa888a74SAndi Kleen * puts them into the mem_map). 2116aa888a74SAndi Kleen */ 2117aa888a74SAndi Kleen m = addr; 2118aa888a74SAndi Kleen goto found; 2119aa888a74SAndi Kleen } 2120aa888a74SAndi Kleen } 2121aa888a74SAndi Kleen return 0; 2122aa888a74SAndi Kleen 2123aa888a74SAndi Kleen found: 2124df994eadSLuiz Capitulino BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); 2125aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 2126330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 2127aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 2128aa888a74SAndi Kleen m->hstate = h; 2129aa888a74SAndi Kleen return 1; 2130aa888a74SAndi Kleen } 2131aa888a74SAndi Kleen 2132d00181b9SKirill A. Shutemov static void __init prep_compound_huge_page(struct page *page, 2133d00181b9SKirill A. Shutemov unsigned int order) 213418229df5SAndy Whitcroft { 213518229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 213618229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 213718229df5SAndy Whitcroft else 213818229df5SAndy Whitcroft prep_compound_page(page, order); 213918229df5SAndy Whitcroft } 214018229df5SAndy Whitcroft 2141aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 2142aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 2143aa888a74SAndi Kleen { 2144aa888a74SAndi Kleen struct huge_bootmem_page *m; 2145aa888a74SAndi Kleen 2146aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 214740d18ebfSMike Kravetz struct page *page = virt_to_page(m); 2148aa888a74SAndi Kleen struct hstate *h = m->hstate; 2149ee8f248dSBecky Bruce 2150aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 215118229df5SAndy Whitcroft prep_compound_huge_page(page, h->order); 2152ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 2153aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 2154af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 2155af0fb9dfSMichal Hocko 2156b0320c7bSRafael Aquini /* 2157b0320c7bSRafael Aquini * If we had gigantic hugepages allocated at boot time, we need 2158b0320c7bSRafael Aquini * to restore the 'stolen' pages to totalram_pages in order to 2159b0320c7bSRafael Aquini * fix confusing memory reports from free(1) and another 2160b0320c7bSRafael Aquini * side-effects, like CommitLimit going negative. 2161b0320c7bSRafael Aquini */ 2162bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 21633dcc0571SJiang Liu adjust_managed_page_count(page, 1 << h->order); 2164520495feSCannon Matthews cond_resched(); 2165aa888a74SAndi Kleen } 2166aa888a74SAndi Kleen } 2167aa888a74SAndi Kleen 21688faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 21691da177e4SLinus Torvalds { 21701da177e4SLinus Torvalds unsigned long i; 2171f60858f9SMike Kravetz nodemask_t *node_alloc_noretry; 2172f60858f9SMike Kravetz 2173f60858f9SMike Kravetz if (!hstate_is_gigantic(h)) { 2174f60858f9SMike Kravetz /* 2175f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 2176f60858f9SMike Kravetz * Ignore errors as lower level routines can deal with 2177f60858f9SMike Kravetz * node_alloc_noretry == NULL. If this kmalloc fails at boot 2178f60858f9SMike Kravetz * time, we are likely in bigger trouble. 2179f60858f9SMike Kravetz */ 2180f60858f9SMike Kravetz node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), 2181f60858f9SMike Kravetz GFP_KERNEL); 2182f60858f9SMike Kravetz } else { 2183f60858f9SMike Kravetz /* allocations done at boot time */ 2184f60858f9SMike Kravetz node_alloc_noretry = NULL; 2185f60858f9SMike Kravetz } 2186f60858f9SMike Kravetz 2187f60858f9SMike Kravetz /* bit mask controlling how hard we retry per-node allocations */ 2188f60858f9SMike Kravetz if (node_alloc_noretry) 2189f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 21901da177e4SLinus Torvalds 2191e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 2192bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 2193aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 2194aa888a74SAndi Kleen break; 21950c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 2196f60858f9SMike Kravetz &node_states[N_MEMORY], 2197f60858f9SMike Kravetz node_alloc_noretry)) 21981da177e4SLinus Torvalds break; 219969ed779aSDavid Rientjes cond_resched(); 22001da177e4SLinus Torvalds } 2201d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 2202d715cf80SLiam R. Howlett char buf[32]; 2203d715cf80SLiam R. Howlett 2204c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2205d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 2206d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 22078faa8b07SAndi Kleen h->max_huge_pages = i; 2208e5ff2159SAndi Kleen } 2209f60858f9SMike Kravetz 2210f60858f9SMike Kravetz kfree(node_alloc_noretry); 2211d715cf80SLiam R. Howlett } 2212e5ff2159SAndi Kleen 2213e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 2214e5ff2159SAndi Kleen { 2215e5ff2159SAndi Kleen struct hstate *h; 2216e5ff2159SAndi Kleen 2217e5ff2159SAndi Kleen for_each_hstate(h) { 2218641844f5SNaoya Horiguchi if (minimum_order > huge_page_order(h)) 2219641844f5SNaoya Horiguchi minimum_order = huge_page_order(h); 2220641844f5SNaoya Horiguchi 22218faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 2222bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 22238faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 2224e5ff2159SAndi Kleen } 2225641844f5SNaoya Horiguchi VM_BUG_ON(minimum_order == UINT_MAX); 2226e5ff2159SAndi Kleen } 2227e5ff2159SAndi Kleen 2228e5ff2159SAndi Kleen static void __init report_hugepages(void) 2229e5ff2159SAndi Kleen { 2230e5ff2159SAndi Kleen struct hstate *h; 2231e5ff2159SAndi Kleen 2232e5ff2159SAndi Kleen for_each_hstate(h) { 22334abd32dbSAndi Kleen char buf[32]; 2234c6247f72SMatthew Wilcox 2235c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2236ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 2237c6247f72SMatthew Wilcox buf, h->free_huge_pages); 2238e5ff2159SAndi Kleen } 2239e5ff2159SAndi Kleen } 2240e5ff2159SAndi Kleen 22411da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 22426ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 22436ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 22441da177e4SLinus Torvalds { 22454415cc8dSChristoph Lameter int i; 22464415cc8dSChristoph Lameter 2247bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2248aa888a74SAndi Kleen return; 2249aa888a74SAndi Kleen 22506ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 22511da177e4SLinus Torvalds struct page *page, *next; 2252a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 2253a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 2254a5516438SAndi Kleen if (count >= h->nr_huge_pages) 22556b0c880dSAdam Litke return; 22561da177e4SLinus Torvalds if (PageHighMem(page)) 22571da177e4SLinus Torvalds continue; 22581da177e4SLinus Torvalds list_del(&page->lru); 2259e5ff2159SAndi Kleen update_and_free_page(h, page); 2260a5516438SAndi Kleen h->free_huge_pages--; 2261a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 22621da177e4SLinus Torvalds } 22631da177e4SLinus Torvalds } 22641da177e4SLinus Torvalds } 22651da177e4SLinus Torvalds #else 22666ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 22676ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 22681da177e4SLinus Torvalds { 22691da177e4SLinus Torvalds } 22701da177e4SLinus Torvalds #endif 22711da177e4SLinus Torvalds 227220a0307cSWu Fengguang /* 227320a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 227420a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 227520a0307cSWu Fengguang * Returns 1 if an adjustment was made. 227620a0307cSWu Fengguang */ 22776ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 22786ae11b27SLee Schermerhorn int delta) 227920a0307cSWu Fengguang { 2280b2261026SJoonsoo Kim int nr_nodes, node; 228120a0307cSWu Fengguang 228220a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 228320a0307cSWu Fengguang 2284e8c5c824SLee Schermerhorn if (delta < 0) { 2285b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 2286b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 2287b2261026SJoonsoo Kim goto found; 2288b2261026SJoonsoo Kim } 2289b2261026SJoonsoo Kim } else { 2290b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2291b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 2292b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 2293b2261026SJoonsoo Kim goto found; 2294e8c5c824SLee Schermerhorn } 22959a76db09SLee Schermerhorn } 2296b2261026SJoonsoo Kim return 0; 229720a0307cSWu Fengguang 2298b2261026SJoonsoo Kim found: 229920a0307cSWu Fengguang h->surplus_huge_pages += delta; 2300b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 2301b2261026SJoonsoo Kim return 1; 230220a0307cSWu Fengguang } 230320a0307cSWu Fengguang 2304a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 2305fd875dcaSMike Kravetz static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 23066ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 23071da177e4SLinus Torvalds { 23087893d1d5SAdam Litke unsigned long min_count, ret; 2309f60858f9SMike Kravetz NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); 2310f60858f9SMike Kravetz 2311f60858f9SMike Kravetz /* 2312f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 2313f60858f9SMike Kravetz * If we can not allocate the bit mask, do not attempt to allocate 2314f60858f9SMike Kravetz * the requested huge pages. 2315f60858f9SMike Kravetz */ 2316f60858f9SMike Kravetz if (node_alloc_noretry) 2317f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 2318f60858f9SMike Kravetz else 2319f60858f9SMike Kravetz return -ENOMEM; 23201da177e4SLinus Torvalds 23214eb0716eSAlexandre Ghiti spin_lock(&hugetlb_lock); 23224eb0716eSAlexandre Ghiti 23234eb0716eSAlexandre Ghiti /* 2324fd875dcaSMike Kravetz * Check for a node specific request. 2325fd875dcaSMike Kravetz * Changing node specific huge page count may require a corresponding 2326fd875dcaSMike Kravetz * change to the global count. In any case, the passed node mask 2327fd875dcaSMike Kravetz * (nodes_allowed) will restrict alloc/free to the specified node. 2328fd875dcaSMike Kravetz */ 2329fd875dcaSMike Kravetz if (nid != NUMA_NO_NODE) { 2330fd875dcaSMike Kravetz unsigned long old_count = count; 2331fd875dcaSMike Kravetz 2332fd875dcaSMike Kravetz count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 2333fd875dcaSMike Kravetz /* 2334fd875dcaSMike Kravetz * User may have specified a large count value which caused the 2335fd875dcaSMike Kravetz * above calculation to overflow. In this case, they wanted 2336fd875dcaSMike Kravetz * to allocate as many huge pages as possible. Set count to 2337fd875dcaSMike Kravetz * largest possible value to align with their intention. 2338fd875dcaSMike Kravetz */ 2339fd875dcaSMike Kravetz if (count < old_count) 2340fd875dcaSMike Kravetz count = ULONG_MAX; 2341fd875dcaSMike Kravetz } 2342fd875dcaSMike Kravetz 2343fd875dcaSMike Kravetz /* 23444eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 23454eb0716eSAlexandre Ghiti * page range allocation. 23464eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 23474eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 23484eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 23494eb0716eSAlexandre Ghiti */ 23504eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 23514eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 23524eb0716eSAlexandre Ghiti spin_unlock(&hugetlb_lock); 2353f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 23544eb0716eSAlexandre Ghiti return -EINVAL; 23554eb0716eSAlexandre Ghiti } 23564eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 23574eb0716eSAlexandre Ghiti } 2358aa888a74SAndi Kleen 23597893d1d5SAdam Litke /* 23607893d1d5SAdam Litke * Increase the pool size 23617893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 23627893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 2363d1c3fb1fSNishanth Aravamudan * 23640c397daeSMichal Hocko * We might race with alloc_surplus_huge_page() here and be unable 2365d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 2366d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 2367d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 2368d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 23697893d1d5SAdam Litke */ 2370a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 23716ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 23727893d1d5SAdam Litke break; 23737893d1d5SAdam Litke } 23747893d1d5SAdam Litke 2375a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 23767893d1d5SAdam Litke /* 23777893d1d5SAdam Litke * If this allocation races such that we no longer need the 23787893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 23797893d1d5SAdam Litke * and reducing the surplus. 23807893d1d5SAdam Litke */ 23817893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 2382649920c6SJia He 2383649920c6SJia He /* yield cpu to avoid soft lockup */ 2384649920c6SJia He cond_resched(); 2385649920c6SJia He 2386f60858f9SMike Kravetz ret = alloc_pool_huge_page(h, nodes_allowed, 2387f60858f9SMike Kravetz node_alloc_noretry); 23887893d1d5SAdam Litke spin_lock(&hugetlb_lock); 23897893d1d5SAdam Litke if (!ret) 23907893d1d5SAdam Litke goto out; 23917893d1d5SAdam Litke 2392536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 2393536240f2SMel Gorman if (signal_pending(current)) 2394536240f2SMel Gorman goto out; 23957893d1d5SAdam Litke } 23967893d1d5SAdam Litke 23977893d1d5SAdam Litke /* 23987893d1d5SAdam Litke * Decrease the pool size 23997893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 24007893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 24017893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 24027893d1d5SAdam Litke * to the desired size as pages become free. 2403d1c3fb1fSNishanth Aravamudan * 2404d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 2405d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 2406d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 24070c397daeSMichal Hocko * alloc_surplus_huge_page() is checking the global counter, 2408d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 2409d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 2410d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 24117893d1d5SAdam Litke */ 2412a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 24136b0c880dSAdam Litke min_count = max(count, min_count); 24146ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 2415a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 24166ae11b27SLee Schermerhorn if (!free_pool_huge_page(h, nodes_allowed, 0)) 24171da177e4SLinus Torvalds break; 241855f67141SMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 24191da177e4SLinus Torvalds } 2420a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 24216ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 24227893d1d5SAdam Litke break; 24237893d1d5SAdam Litke } 24247893d1d5SAdam Litke out: 24254eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 24261da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 24274eb0716eSAlexandre Ghiti 2428f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 2429f60858f9SMike Kravetz 24304eb0716eSAlexandre Ghiti return 0; 24311da177e4SLinus Torvalds } 24321da177e4SLinus Torvalds 2433a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 2434a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 2435a3437870SNishanth Aravamudan 2436a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 2437a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 2438a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 2439a3437870SNishanth Aravamudan 2440a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 2441a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 2442a3437870SNishanth Aravamudan 24439a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 24449a305230SLee Schermerhorn 24459a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 2446a3437870SNishanth Aravamudan { 2447a3437870SNishanth Aravamudan int i; 24489a305230SLee Schermerhorn 2449a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 24509a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 24519a305230SLee Schermerhorn if (nidp) 24529a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 2453a3437870SNishanth Aravamudan return &hstates[i]; 24549a305230SLee Schermerhorn } 24559a305230SLee Schermerhorn 24569a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 2457a3437870SNishanth Aravamudan } 2458a3437870SNishanth Aravamudan 245906808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 2460a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2461a3437870SNishanth Aravamudan { 24629a305230SLee Schermerhorn struct hstate *h; 24639a305230SLee Schermerhorn unsigned long nr_huge_pages; 24649a305230SLee Schermerhorn int nid; 24659a305230SLee Schermerhorn 24669a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 24679a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 24689a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 24699a305230SLee Schermerhorn else 24709a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 24719a305230SLee Schermerhorn 24729a305230SLee Schermerhorn return sprintf(buf, "%lu\n", nr_huge_pages); 2473a3437870SNishanth Aravamudan } 2474adbe8726SEric B Munson 2475238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 2476238d3c13SDavid Rientjes struct hstate *h, int nid, 2477238d3c13SDavid Rientjes unsigned long count, size_t len) 2478a3437870SNishanth Aravamudan { 2479a3437870SNishanth Aravamudan int err; 24802d0adf7eSOscar Salvador nodemask_t nodes_allowed, *n_mask; 2481a3437870SNishanth Aravamudan 24822d0adf7eSOscar Salvador if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 24832d0adf7eSOscar Salvador return -EINVAL; 2484adbe8726SEric B Munson 24859a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 24869a305230SLee Schermerhorn /* 24879a305230SLee Schermerhorn * global hstate attribute 24889a305230SLee Schermerhorn */ 24899a305230SLee Schermerhorn if (!(obey_mempolicy && 24902d0adf7eSOscar Salvador init_nodemask_of_mempolicy(&nodes_allowed))) 24912d0adf7eSOscar Salvador n_mask = &node_states[N_MEMORY]; 24922d0adf7eSOscar Salvador else 24932d0adf7eSOscar Salvador n_mask = &nodes_allowed; 24942d0adf7eSOscar Salvador } else { 24959a305230SLee Schermerhorn /* 2496fd875dcaSMike Kravetz * Node specific request. count adjustment happens in 2497fd875dcaSMike Kravetz * set_max_huge_pages() after acquiring hugetlb_lock. 24989a305230SLee Schermerhorn */ 24992d0adf7eSOscar Salvador init_nodemask_of_node(&nodes_allowed, nid); 25002d0adf7eSOscar Salvador n_mask = &nodes_allowed; 2501fd875dcaSMike Kravetz } 25029a305230SLee Schermerhorn 25032d0adf7eSOscar Salvador err = set_max_huge_pages(h, count, nid, n_mask); 250406808b08SLee Schermerhorn 25054eb0716eSAlexandre Ghiti return err ? err : len; 250606808b08SLee Schermerhorn } 250706808b08SLee Schermerhorn 2508238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 2509238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 2510238d3c13SDavid Rientjes size_t len) 2511238d3c13SDavid Rientjes { 2512238d3c13SDavid Rientjes struct hstate *h; 2513238d3c13SDavid Rientjes unsigned long count; 2514238d3c13SDavid Rientjes int nid; 2515238d3c13SDavid Rientjes int err; 2516238d3c13SDavid Rientjes 2517238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 2518238d3c13SDavid Rientjes if (err) 2519238d3c13SDavid Rientjes return err; 2520238d3c13SDavid Rientjes 2521238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 2522238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 2523238d3c13SDavid Rientjes } 2524238d3c13SDavid Rientjes 252506808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 252606808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 252706808b08SLee Schermerhorn { 252806808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 252906808b08SLee Schermerhorn } 253006808b08SLee Schermerhorn 253106808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 253206808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 253306808b08SLee Schermerhorn { 2534238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 2535a3437870SNishanth Aravamudan } 2536a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 2537a3437870SNishanth Aravamudan 253806808b08SLee Schermerhorn #ifdef CONFIG_NUMA 253906808b08SLee Schermerhorn 254006808b08SLee Schermerhorn /* 254106808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 254206808b08SLee Schermerhorn * huge page alloc/free. 254306808b08SLee Schermerhorn */ 254406808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 254506808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 254606808b08SLee Schermerhorn { 254706808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 254806808b08SLee Schermerhorn } 254906808b08SLee Schermerhorn 255006808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 255106808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 255206808b08SLee Schermerhorn { 2553238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 255406808b08SLee Schermerhorn } 255506808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 255606808b08SLee Schermerhorn #endif 255706808b08SLee Schermerhorn 255806808b08SLee Schermerhorn 2559a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 2560a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2561a3437870SNishanth Aravamudan { 25629a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2563a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 2564a3437870SNishanth Aravamudan } 2565adbe8726SEric B Munson 2566a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 2567a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 2568a3437870SNishanth Aravamudan { 2569a3437870SNishanth Aravamudan int err; 2570a3437870SNishanth Aravamudan unsigned long input; 25719a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2572a3437870SNishanth Aravamudan 2573bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2574adbe8726SEric B Munson return -EINVAL; 2575adbe8726SEric B Munson 25763dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 2577a3437870SNishanth Aravamudan if (err) 257873ae31e5SEric B Munson return err; 2579a3437870SNishanth Aravamudan 2580a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 2581a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 2582a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 2583a3437870SNishanth Aravamudan 2584a3437870SNishanth Aravamudan return count; 2585a3437870SNishanth Aravamudan } 2586a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 2587a3437870SNishanth Aravamudan 2588a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 2589a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2590a3437870SNishanth Aravamudan { 25919a305230SLee Schermerhorn struct hstate *h; 25929a305230SLee Schermerhorn unsigned long free_huge_pages; 25939a305230SLee Schermerhorn int nid; 25949a305230SLee Schermerhorn 25959a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 25969a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 25979a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 25989a305230SLee Schermerhorn else 25999a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 26009a305230SLee Schermerhorn 26019a305230SLee Schermerhorn return sprintf(buf, "%lu\n", free_huge_pages); 2602a3437870SNishanth Aravamudan } 2603a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 2604a3437870SNishanth Aravamudan 2605a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 2606a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2607a3437870SNishanth Aravamudan { 26089a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2609a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->resv_huge_pages); 2610a3437870SNishanth Aravamudan } 2611a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 2612a3437870SNishanth Aravamudan 2613a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 2614a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2615a3437870SNishanth Aravamudan { 26169a305230SLee Schermerhorn struct hstate *h; 26179a305230SLee Schermerhorn unsigned long surplus_huge_pages; 26189a305230SLee Schermerhorn int nid; 26199a305230SLee Schermerhorn 26209a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 26219a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 26229a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 26239a305230SLee Schermerhorn else 26249a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 26259a305230SLee Schermerhorn 26269a305230SLee Schermerhorn return sprintf(buf, "%lu\n", surplus_huge_pages); 2627a3437870SNishanth Aravamudan } 2628a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 2629a3437870SNishanth Aravamudan 2630a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 2631a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 2632a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 2633a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 2634a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 2635a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 263606808b08SLee Schermerhorn #ifdef CONFIG_NUMA 263706808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 263806808b08SLee Schermerhorn #endif 2639a3437870SNishanth Aravamudan NULL, 2640a3437870SNishanth Aravamudan }; 2641a3437870SNishanth Aravamudan 264267e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 2643a3437870SNishanth Aravamudan .attrs = hstate_attrs, 2644a3437870SNishanth Aravamudan }; 2645a3437870SNishanth Aravamudan 2646094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 26479a305230SLee Schermerhorn struct kobject **hstate_kobjs, 264867e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 2649a3437870SNishanth Aravamudan { 2650a3437870SNishanth Aravamudan int retval; 2651972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 2652a3437870SNishanth Aravamudan 26539a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 26549a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 2655a3437870SNishanth Aravamudan return -ENOMEM; 2656a3437870SNishanth Aravamudan 26579a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 2658a3437870SNishanth Aravamudan if (retval) 26599a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 2660a3437870SNishanth Aravamudan 2661a3437870SNishanth Aravamudan return retval; 2662a3437870SNishanth Aravamudan } 2663a3437870SNishanth Aravamudan 2664a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 2665a3437870SNishanth Aravamudan { 2666a3437870SNishanth Aravamudan struct hstate *h; 2667a3437870SNishanth Aravamudan int err; 2668a3437870SNishanth Aravamudan 2669a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 2670a3437870SNishanth Aravamudan if (!hugepages_kobj) 2671a3437870SNishanth Aravamudan return; 2672a3437870SNishanth Aravamudan 2673a3437870SNishanth Aravamudan for_each_hstate(h) { 26749a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 26759a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 2676a3437870SNishanth Aravamudan if (err) 2677ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s", h->name); 2678a3437870SNishanth Aravamudan } 2679a3437870SNishanth Aravamudan } 2680a3437870SNishanth Aravamudan 26819a305230SLee Schermerhorn #ifdef CONFIG_NUMA 26829a305230SLee Schermerhorn 26839a305230SLee Schermerhorn /* 26849a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 268510fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 268610fbcf4cSKay Sievers * index of a node device or _hstate == node id. 268710fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 26889a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 26899a305230SLee Schermerhorn */ 26909a305230SLee Schermerhorn struct node_hstate { 26919a305230SLee Schermerhorn struct kobject *hugepages_kobj; 26929a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 26939a305230SLee Schermerhorn }; 2694b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 26959a305230SLee Schermerhorn 26969a305230SLee Schermerhorn /* 269710fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 26989a305230SLee Schermerhorn */ 26999a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 27009a305230SLee Schermerhorn &nr_hugepages_attr.attr, 27019a305230SLee Schermerhorn &free_hugepages_attr.attr, 27029a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 27039a305230SLee Schermerhorn NULL, 27049a305230SLee Schermerhorn }; 27059a305230SLee Schermerhorn 270667e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 27079a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 27089a305230SLee Schermerhorn }; 27099a305230SLee Schermerhorn 27109a305230SLee Schermerhorn /* 271110fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 27129a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 27139a305230SLee Schermerhorn */ 27149a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 27159a305230SLee Schermerhorn { 27169a305230SLee Schermerhorn int nid; 27179a305230SLee Schermerhorn 27189a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 27199a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 27209a305230SLee Schermerhorn int i; 27219a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 27229a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 27239a305230SLee Schermerhorn if (nidp) 27249a305230SLee Schermerhorn *nidp = nid; 27259a305230SLee Schermerhorn return &hstates[i]; 27269a305230SLee Schermerhorn } 27279a305230SLee Schermerhorn } 27289a305230SLee Schermerhorn 27299a305230SLee Schermerhorn BUG(); 27309a305230SLee Schermerhorn return NULL; 27319a305230SLee Schermerhorn } 27329a305230SLee Schermerhorn 27339a305230SLee Schermerhorn /* 273410fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 27359a305230SLee Schermerhorn * No-op if no hstate attributes attached. 27369a305230SLee Schermerhorn */ 27373cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 27389a305230SLee Schermerhorn { 27399a305230SLee Schermerhorn struct hstate *h; 274010fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 27419a305230SLee Schermerhorn 27429a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 27439b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 27449a305230SLee Schermerhorn 2745972dc4deSAneesh Kumar K.V for_each_hstate(h) { 2746972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 2747972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 2748972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 2749972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 2750972dc4deSAneesh Kumar K.V } 27519a305230SLee Schermerhorn } 27529a305230SLee Schermerhorn 27539a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 27549a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 27559a305230SLee Schermerhorn } 27569a305230SLee Schermerhorn 27579a305230SLee Schermerhorn 27589a305230SLee Schermerhorn /* 275910fbcf4cSKay Sievers * Register hstate attributes for a single node device. 27609a305230SLee Schermerhorn * No-op if attributes already registered. 27619a305230SLee Schermerhorn */ 27623cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 27639a305230SLee Schermerhorn { 27649a305230SLee Schermerhorn struct hstate *h; 276510fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 27669a305230SLee Schermerhorn int err; 27679a305230SLee Schermerhorn 27689a305230SLee Schermerhorn if (nhs->hugepages_kobj) 27699a305230SLee Schermerhorn return; /* already allocated */ 27709a305230SLee Schermerhorn 27719a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 277210fbcf4cSKay Sievers &node->dev.kobj); 27739a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 27749a305230SLee Schermerhorn return; 27759a305230SLee Schermerhorn 27769a305230SLee Schermerhorn for_each_hstate(h) { 27779a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 27789a305230SLee Schermerhorn nhs->hstate_kobjs, 27799a305230SLee Schermerhorn &per_node_hstate_attr_group); 27809a305230SLee Schermerhorn if (err) { 2781ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s for node %d\n", 278210fbcf4cSKay Sievers h->name, node->dev.id); 27839a305230SLee Schermerhorn hugetlb_unregister_node(node); 27849a305230SLee Schermerhorn break; 27859a305230SLee Schermerhorn } 27869a305230SLee Schermerhorn } 27879a305230SLee Schermerhorn } 27889a305230SLee Schermerhorn 27899a305230SLee Schermerhorn /* 27909b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 279110fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 279210fbcf4cSKay Sievers * registered their associated device by this time. 27939a305230SLee Schermerhorn */ 27947d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 27959a305230SLee Schermerhorn { 27969a305230SLee Schermerhorn int nid; 27979a305230SLee Schermerhorn 27988cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 27998732794bSWen Congyang struct node *node = node_devices[nid]; 280010fbcf4cSKay Sievers if (node->dev.id == nid) 28019a305230SLee Schermerhorn hugetlb_register_node(node); 28029a305230SLee Schermerhorn } 28039a305230SLee Schermerhorn 28049a305230SLee Schermerhorn /* 280510fbcf4cSKay Sievers * Let the node device driver know we're here so it can 28069a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 28079a305230SLee Schermerhorn */ 28089a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 28099a305230SLee Schermerhorn hugetlb_unregister_node); 28109a305230SLee Schermerhorn } 28119a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 28129a305230SLee Schermerhorn 28139a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 28149a305230SLee Schermerhorn { 28159a305230SLee Schermerhorn BUG(); 28169a305230SLee Schermerhorn if (nidp) 28179a305230SLee Schermerhorn *nidp = -1; 28189a305230SLee Schermerhorn return NULL; 28199a305230SLee Schermerhorn } 28209a305230SLee Schermerhorn 28219a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 28229a305230SLee Schermerhorn 28239a305230SLee Schermerhorn #endif 28249a305230SLee Schermerhorn 2825a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 2826a3437870SNishanth Aravamudan { 28278382d914SDavidlohr Bueso int i; 28288382d914SDavidlohr Bueso 2829457c1b27SNishanth Aravamudan if (!hugepages_supported()) 28300ef89d25SBenjamin Herrenschmidt return 0; 2831a3437870SNishanth Aravamudan 2832e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) { 2833d715cf80SLiam R. Howlett if (default_hstate_size != 0) { 2834d715cf80SLiam R. Howlett pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n", 2835d715cf80SLiam R. Howlett default_hstate_size, HPAGE_SIZE); 2836d715cf80SLiam R. Howlett } 2837d715cf80SLiam R. Howlett 2838e11bfbfcSNick Piggin default_hstate_size = HPAGE_SIZE; 2839e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) 2840a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 2841a3437870SNishanth Aravamudan } 2842972dc4deSAneesh Kumar K.V default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); 2843f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 2844f8b74815SVaishali Thakkar if (!default_hstate.max_huge_pages) 2845e11bfbfcSNick Piggin default_hstate.max_huge_pages = default_hstate_max_huge_pages; 2846f8b74815SVaishali Thakkar } 2847a3437870SNishanth Aravamudan 2848a3437870SNishanth Aravamudan hugetlb_init_hstates(); 2849aa888a74SAndi Kleen gather_bootmem_prealloc(); 2850a3437870SNishanth Aravamudan report_hugepages(); 2851a3437870SNishanth Aravamudan 2852a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 28539a305230SLee Schermerhorn hugetlb_register_all_nodes(); 28547179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 28559a305230SLee Schermerhorn 28568382d914SDavidlohr Bueso #ifdef CONFIG_SMP 28578382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 28588382d914SDavidlohr Bueso #else 28598382d914SDavidlohr Bueso num_fault_mutexes = 1; 28608382d914SDavidlohr Bueso #endif 2861c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 28626da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 28636da2ec56SKees Cook GFP_KERNEL); 2864c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 28658382d914SDavidlohr Bueso 28668382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 2867c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 2868a3437870SNishanth Aravamudan return 0; 2869a3437870SNishanth Aravamudan } 28703e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 2871a3437870SNishanth Aravamudan 2872a3437870SNishanth Aravamudan /* Should be called on processing a hugepagesz=... option */ 28739fee021dSVaishali Thakkar void __init hugetlb_bad_size(void) 28749fee021dSVaishali Thakkar { 28759fee021dSVaishali Thakkar parsed_valid_hugepagesz = false; 28769fee021dSVaishali Thakkar } 28779fee021dSVaishali Thakkar 2878d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 2879a3437870SNishanth Aravamudan { 2880a3437870SNishanth Aravamudan struct hstate *h; 28818faa8b07SAndi Kleen unsigned long i; 28828faa8b07SAndi Kleen 2883a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 2884598d8091SJoe Perches pr_warn("hugepagesz= specified twice, ignoring\n"); 2885a3437870SNishanth Aravamudan return; 2886a3437870SNishanth Aravamudan } 288747d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 2888a3437870SNishanth Aravamudan BUG_ON(order == 0); 288947d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 2890a3437870SNishanth Aravamudan h->order = order; 2891a3437870SNishanth Aravamudan h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 28928faa8b07SAndi Kleen h->nr_huge_pages = 0; 28938faa8b07SAndi Kleen h->free_huge_pages = 0; 28948faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 28958faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 28960edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 289754f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 289854f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 2899a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 2900a3437870SNishanth Aravamudan huge_page_size(h)/1024); 29018faa8b07SAndi Kleen 2902a3437870SNishanth Aravamudan parsed_hstate = h; 2903a3437870SNishanth Aravamudan } 2904a3437870SNishanth Aravamudan 2905e11bfbfcSNick Piggin static int __init hugetlb_nrpages_setup(char *s) 2906a3437870SNishanth Aravamudan { 2907a3437870SNishanth Aravamudan unsigned long *mhp; 29088faa8b07SAndi Kleen static unsigned long *last_mhp; 2909a3437870SNishanth Aravamudan 29109fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 29119fee021dSVaishali Thakkar pr_warn("hugepages = %s preceded by " 29129fee021dSVaishali Thakkar "an unsupported hugepagesz, ignoring\n", s); 29139fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 29149fee021dSVaishali Thakkar return 1; 29159fee021dSVaishali Thakkar } 2916a3437870SNishanth Aravamudan /* 291747d38344SAneesh Kumar K.V * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, 2918a3437870SNishanth Aravamudan * so this hugepages= parameter goes to the "default hstate". 2919a3437870SNishanth Aravamudan */ 29209fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 2921a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 2922a3437870SNishanth Aravamudan else 2923a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 2924a3437870SNishanth Aravamudan 29258faa8b07SAndi Kleen if (mhp == last_mhp) { 2926598d8091SJoe Perches pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n"); 29278faa8b07SAndi Kleen return 1; 29288faa8b07SAndi Kleen } 29298faa8b07SAndi Kleen 2930a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 2931a3437870SNishanth Aravamudan *mhp = 0; 2932a3437870SNishanth Aravamudan 29338faa8b07SAndi Kleen /* 29348faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 29358faa8b07SAndi Kleen * But we need to allocate >= MAX_ORDER hstates here early to still 29368faa8b07SAndi Kleen * use the bootmem allocator. 29378faa8b07SAndi Kleen */ 293847d38344SAneesh Kumar K.V if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) 29398faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 29408faa8b07SAndi Kleen 29418faa8b07SAndi Kleen last_mhp = mhp; 29428faa8b07SAndi Kleen 2943a3437870SNishanth Aravamudan return 1; 2944a3437870SNishanth Aravamudan } 2945e11bfbfcSNick Piggin __setup("hugepages=", hugetlb_nrpages_setup); 2946e11bfbfcSNick Piggin 2947e11bfbfcSNick Piggin static int __init hugetlb_default_setup(char *s) 2948e11bfbfcSNick Piggin { 2949e11bfbfcSNick Piggin default_hstate_size = memparse(s, &s); 2950e11bfbfcSNick Piggin return 1; 2951e11bfbfcSNick Piggin } 2952e11bfbfcSNick Piggin __setup("default_hugepagesz=", hugetlb_default_setup); 2953a3437870SNishanth Aravamudan 29548a213460SNishanth Aravamudan static unsigned int cpuset_mems_nr(unsigned int *array) 29558a213460SNishanth Aravamudan { 29568a213460SNishanth Aravamudan int node; 29578a213460SNishanth Aravamudan unsigned int nr = 0; 29588a213460SNishanth Aravamudan 29598a213460SNishanth Aravamudan for_each_node_mask(node, cpuset_current_mems_allowed) 29608a213460SNishanth Aravamudan nr += array[node]; 29618a213460SNishanth Aravamudan 29628a213460SNishanth Aravamudan return nr; 29638a213460SNishanth Aravamudan } 29648a213460SNishanth Aravamudan 29658a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 296606808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 296706808b08SLee Schermerhorn struct ctl_table *table, int write, 296806808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 29691da177e4SLinus Torvalds { 2970e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 2971238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 297208d4a246SMichal Hocko int ret; 2973e5ff2159SAndi Kleen 2974457c1b27SNishanth Aravamudan if (!hugepages_supported()) 297586613628SJan Stancek return -EOPNOTSUPP; 2976457c1b27SNishanth Aravamudan 2977e5ff2159SAndi Kleen table->data = &tmp; 2978e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 297908d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 298008d4a246SMichal Hocko if (ret) 298108d4a246SMichal Hocko goto out; 2982e5ff2159SAndi Kleen 2983238d3c13SDavid Rientjes if (write) 2984238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 2985238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 298608d4a246SMichal Hocko out: 298708d4a246SMichal Hocko return ret; 29881da177e4SLinus Torvalds } 2989396faf03SMel Gorman 299006808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 299106808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 299206808b08SLee Schermerhorn { 299306808b08SLee Schermerhorn 299406808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 299506808b08SLee Schermerhorn buffer, length, ppos); 299606808b08SLee Schermerhorn } 299706808b08SLee Schermerhorn 299806808b08SLee Schermerhorn #ifdef CONFIG_NUMA 299906808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 300006808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 300106808b08SLee Schermerhorn { 300206808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 300306808b08SLee Schermerhorn buffer, length, ppos); 300406808b08SLee Schermerhorn } 300506808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 300606808b08SLee Schermerhorn 3007a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 30088d65af78SAlexey Dobriyan void __user *buffer, 3009a3d0c6aaSNishanth Aravamudan size_t *length, loff_t *ppos) 3010a3d0c6aaSNishanth Aravamudan { 3011a5516438SAndi Kleen struct hstate *h = &default_hstate; 3012e5ff2159SAndi Kleen unsigned long tmp; 301308d4a246SMichal Hocko int ret; 3014e5ff2159SAndi Kleen 3015457c1b27SNishanth Aravamudan if (!hugepages_supported()) 301686613628SJan Stancek return -EOPNOTSUPP; 3017457c1b27SNishanth Aravamudan 3018e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 3019e5ff2159SAndi Kleen 3020bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 3021adbe8726SEric B Munson return -EINVAL; 3022adbe8726SEric B Munson 3023e5ff2159SAndi Kleen table->data = &tmp; 3024e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 302508d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 302608d4a246SMichal Hocko if (ret) 302708d4a246SMichal Hocko goto out; 3028e5ff2159SAndi Kleen 3029e5ff2159SAndi Kleen if (write) { 3030064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 3031e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 3032a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 3033e5ff2159SAndi Kleen } 303408d4a246SMichal Hocko out: 303508d4a246SMichal Hocko return ret; 3036a3d0c6aaSNishanth Aravamudan } 3037a3d0c6aaSNishanth Aravamudan 30381da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 30391da177e4SLinus Torvalds 3040e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 30411da177e4SLinus Torvalds { 3042fcb2b0c5SRoman Gushchin struct hstate *h; 3043fcb2b0c5SRoman Gushchin unsigned long total = 0; 3044fcb2b0c5SRoman Gushchin 3045457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3046457c1b27SNishanth Aravamudan return; 3047fcb2b0c5SRoman Gushchin 3048fcb2b0c5SRoman Gushchin for_each_hstate(h) { 3049fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 3050fcb2b0c5SRoman Gushchin 3051fcb2b0c5SRoman Gushchin total += (PAGE_SIZE << huge_page_order(h)) * count; 3052fcb2b0c5SRoman Gushchin 3053fcb2b0c5SRoman Gushchin if (h == &default_hstate) 3054e1759c21SAlexey Dobriyan seq_printf(m, 30551da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 30561da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 3057b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 30587893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 30594f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 3060fcb2b0c5SRoman Gushchin count, 3061a5516438SAndi Kleen h->free_huge_pages, 3062a5516438SAndi Kleen h->resv_huge_pages, 3063a5516438SAndi Kleen h->surplus_huge_pages, 3064fcb2b0c5SRoman Gushchin (PAGE_SIZE << huge_page_order(h)) / 1024); 3065fcb2b0c5SRoman Gushchin } 3066fcb2b0c5SRoman Gushchin 3067fcb2b0c5SRoman Gushchin seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024); 30681da177e4SLinus Torvalds } 30691da177e4SLinus Torvalds 30701da177e4SLinus Torvalds int hugetlb_report_node_meminfo(int nid, char *buf) 30711da177e4SLinus Torvalds { 3072a5516438SAndi Kleen struct hstate *h = &default_hstate; 3073457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3074457c1b27SNishanth Aravamudan return 0; 30751da177e4SLinus Torvalds return sprintf(buf, 30761da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 3077a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 3078a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 3079a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 3080a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 3081a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 30821da177e4SLinus Torvalds } 30831da177e4SLinus Torvalds 3084949f7ec5SDavid Rientjes void hugetlb_show_meminfo(void) 3085949f7ec5SDavid Rientjes { 3086949f7ec5SDavid Rientjes struct hstate *h; 3087949f7ec5SDavid Rientjes int nid; 3088949f7ec5SDavid Rientjes 3089457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3090457c1b27SNishanth Aravamudan return; 3091457c1b27SNishanth Aravamudan 3092949f7ec5SDavid Rientjes for_each_node_state(nid, N_MEMORY) 3093949f7ec5SDavid Rientjes for_each_hstate(h) 3094949f7ec5SDavid Rientjes pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 3095949f7ec5SDavid Rientjes nid, 3096949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 3097949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 3098949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 3099949f7ec5SDavid Rientjes 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 3100949f7ec5SDavid Rientjes } 3101949f7ec5SDavid Rientjes 31025d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 31035d317b2bSNaoya Horiguchi { 31045d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 31055d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 31065d317b2bSNaoya Horiguchi } 31075d317b2bSNaoya Horiguchi 31081da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 31091da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 31101da177e4SLinus Torvalds { 3111d0028588SWanpeng Li struct hstate *h; 3112d0028588SWanpeng Li unsigned long nr_total_pages = 0; 3113d0028588SWanpeng Li 3114d0028588SWanpeng Li for_each_hstate(h) 3115d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 3116d0028588SWanpeng Li return nr_total_pages; 31171da177e4SLinus Torvalds } 31181da177e4SLinus Torvalds 3119a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 3120fc1b8a73SMel Gorman { 3121fc1b8a73SMel Gorman int ret = -ENOMEM; 3122fc1b8a73SMel Gorman 3123fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 3124fc1b8a73SMel Gorman /* 3125fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 3126fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 3127fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 3128fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 3129fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 3130fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 3131fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 3132fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 3133fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 3134fc1b8a73SMel Gorman * 3135fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 3136fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 3137fc1b8a73SMel Gorman * we fall back to check against current free page availability as 3138fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 3139fc1b8a73SMel Gorman * semantics that cpuset has. 3140fc1b8a73SMel Gorman */ 3141fc1b8a73SMel Gorman if (delta > 0) { 3142a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 3143fc1b8a73SMel Gorman goto out; 3144fc1b8a73SMel Gorman 3145a5516438SAndi Kleen if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 3146a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 3147fc1b8a73SMel Gorman goto out; 3148fc1b8a73SMel Gorman } 3149fc1b8a73SMel Gorman } 3150fc1b8a73SMel Gorman 3151fc1b8a73SMel Gorman ret = 0; 3152fc1b8a73SMel Gorman if (delta < 0) 3153a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 3154fc1b8a73SMel Gorman 3155fc1b8a73SMel Gorman out: 3156fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 3157fc1b8a73SMel Gorman return ret; 3158fc1b8a73SMel Gorman } 3159fc1b8a73SMel Gorman 316084afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 316184afd99bSAndy Whitcroft { 3162f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 316384afd99bSAndy Whitcroft 316484afd99bSAndy Whitcroft /* 316584afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 316684afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 316784afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 316825985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 316984afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 317084afd99bSAndy Whitcroft * new reference here without additional locking. 317184afd99bSAndy Whitcroft */ 31724e35f483SJoonsoo Kim if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 3173f522c3acSJoonsoo Kim kref_get(&resv->refs); 317484afd99bSAndy Whitcroft } 317584afd99bSAndy Whitcroft 3176a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 3177a1e78772SMel Gorman { 3178a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3179f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 318090481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 31814e35f483SJoonsoo Kim unsigned long reserve, start, end; 31821c5ecae3SMike Kravetz long gbl_reserve; 318384afd99bSAndy Whitcroft 31844e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 31854e35f483SJoonsoo Kim return; 31864e35f483SJoonsoo Kim 3187a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 3188a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 318984afd99bSAndy Whitcroft 31904e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 319184afd99bSAndy Whitcroft 3192f031dd27SJoonsoo Kim kref_put(&resv->refs, resv_map_release); 319384afd99bSAndy Whitcroft 31947251ff78SAdam Litke if (reserve) { 31951c5ecae3SMike Kravetz /* 31961c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 31971c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 31981c5ecae3SMike Kravetz */ 31991c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 32001c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 32017251ff78SAdam Litke } 3202a1e78772SMel Gorman } 3203a1e78772SMel Gorman 320431383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 320531383c68SDan Williams { 320631383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 320731383c68SDan Williams return -EINVAL; 320831383c68SDan Williams return 0; 320931383c68SDan Williams } 321031383c68SDan Williams 321105ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 321205ea8860SDan Williams { 321305ea8860SDan Williams struct hstate *hstate = hstate_vma(vma); 321405ea8860SDan Williams 321505ea8860SDan Williams return 1UL << huge_page_shift(hstate); 321605ea8860SDan Williams } 321705ea8860SDan Williams 32181da177e4SLinus Torvalds /* 32191da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 32201da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 32211da177e4SLinus Torvalds * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 32221da177e4SLinus Torvalds * this far. 32231da177e4SLinus Torvalds */ 3224b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 32251da177e4SLinus Torvalds { 32261da177e4SLinus Torvalds BUG(); 3227d0217ac0SNick Piggin return 0; 32281da177e4SLinus Torvalds } 32291da177e4SLinus Torvalds 3230eec3636aSJane Chu /* 3231eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 3232eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 3233eec3636aSJane Chu * This is because under System V memory model, mappings created via 3234eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 3235eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 3236eec3636aSJane Chu */ 3237f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 3238d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 323984afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 3240a1e78772SMel Gorman .close = hugetlb_vm_op_close, 324131383c68SDan Williams .split = hugetlb_vm_op_split, 324205ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 32431da177e4SLinus Torvalds }; 32441da177e4SLinus Torvalds 32451e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 32461e8f889bSDavid Gibson int writable) 324763551ae0SDavid Gibson { 324863551ae0SDavid Gibson pte_t entry; 324963551ae0SDavid Gibson 32501e8f889bSDavid Gibson if (writable) { 3251106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 3252106c992aSGerald Schaefer vma->vm_page_prot))); 325363551ae0SDavid Gibson } else { 3254106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 3255106c992aSGerald Schaefer vma->vm_page_prot)); 325663551ae0SDavid Gibson } 325763551ae0SDavid Gibson entry = pte_mkyoung(entry); 325863551ae0SDavid Gibson entry = pte_mkhuge(entry); 3259d9ed9faaSChris Metcalf entry = arch_make_huge_pte(entry, vma, page, writable); 326063551ae0SDavid Gibson 326163551ae0SDavid Gibson return entry; 326263551ae0SDavid Gibson } 326363551ae0SDavid Gibson 32641e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 32651e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 32661e8f889bSDavid Gibson { 32671e8f889bSDavid Gibson pte_t entry; 32681e8f889bSDavid Gibson 3269106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 327032f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 32714b3073e1SRussell King update_mmu_cache(vma, address, ptep); 32721e8f889bSDavid Gibson } 32731e8f889bSDavid Gibson 3274d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 32754a705fefSNaoya Horiguchi { 32764a705fefSNaoya Horiguchi swp_entry_t swp; 32774a705fefSNaoya Horiguchi 32784a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 3279d5ed7444SAneesh Kumar K.V return false; 32804a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 32814a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_migration_entry(swp)) 3282d5ed7444SAneesh Kumar K.V return true; 32834a705fefSNaoya Horiguchi else 3284d5ed7444SAneesh Kumar K.V return false; 32854a705fefSNaoya Horiguchi } 32864a705fefSNaoya Horiguchi 32874a705fefSNaoya Horiguchi static int is_hugetlb_entry_hwpoisoned(pte_t pte) 32884a705fefSNaoya Horiguchi { 32894a705fefSNaoya Horiguchi swp_entry_t swp; 32904a705fefSNaoya Horiguchi 32914a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 32924a705fefSNaoya Horiguchi return 0; 32934a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 32944a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_hwpoison_entry(swp)) 32954a705fefSNaoya Horiguchi return 1; 32964a705fefSNaoya Horiguchi else 32974a705fefSNaoya Horiguchi return 0; 32984a705fefSNaoya Horiguchi } 32991e8f889bSDavid Gibson 330063551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 330163551ae0SDavid Gibson struct vm_area_struct *vma) 330263551ae0SDavid Gibson { 33035e41540cSMike Kravetz pte_t *src_pte, *dst_pte, entry, dst_entry; 330463551ae0SDavid Gibson struct page *ptepage; 33051c59827dSHugh Dickins unsigned long addr; 33061e8f889bSDavid Gibson int cow; 3307a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3308a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 3309ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 3310e8569dd2SAndreas Sandberg int ret = 0; 33111e8f889bSDavid Gibson 33121e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 331363551ae0SDavid Gibson 3314ac46d4f3SJérôme Glisse if (cow) { 33157269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src, 33166f4f13e8SJérôme Glisse vma->vm_start, 3317ac46d4f3SJérôme Glisse vma->vm_end); 3318ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 3319ac46d4f3SJérôme Glisse } 3320e8569dd2SAndreas Sandberg 3321a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 3322cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 33237868a208SPunit Agrawal src_pte = huge_pte_offset(src, addr, sz); 3324c74df32cSHugh Dickins if (!src_pte) 3325c74df32cSHugh Dickins continue; 3326a5516438SAndi Kleen dst_pte = huge_pte_alloc(dst, addr, sz); 3327e8569dd2SAndreas Sandberg if (!dst_pte) { 3328e8569dd2SAndreas Sandberg ret = -ENOMEM; 3329e8569dd2SAndreas Sandberg break; 3330e8569dd2SAndreas Sandberg } 3331c5c99429SLarry Woodman 33325e41540cSMike Kravetz /* 33335e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 33345e41540cSMike Kravetz * dst_pte == src_pte is the common case of src/dest sharing. 33355e41540cSMike Kravetz * 33365e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 33375e41540cSMike Kravetz * another vma. If dst_pte !none, this implies sharing. 33385e41540cSMike Kravetz * Check here before taking page table lock, and once again 33395e41540cSMike Kravetz * after taking the lock below. 33405e41540cSMike Kravetz */ 33415e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 33425e41540cSMike Kravetz if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) 3343c5c99429SLarry Woodman continue; 3344c5c99429SLarry Woodman 3345cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 3346cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 3347cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 33484a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 33495e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 33505e41540cSMike Kravetz if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { 33515e41540cSMike Kravetz /* 33525e41540cSMike Kravetz * Skip if src entry none. Also, skip in the 33535e41540cSMike Kravetz * unlikely case dst entry !none as this implies 33545e41540cSMike Kravetz * sharing with another vma. 33555e41540cSMike Kravetz */ 33564a705fefSNaoya Horiguchi ; 33574a705fefSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry) || 33584a705fefSNaoya Horiguchi is_hugetlb_entry_hwpoisoned(entry))) { 33594a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 33604a705fefSNaoya Horiguchi 33614a705fefSNaoya Horiguchi if (is_write_migration_entry(swp_entry) && cow) { 33624a705fefSNaoya Horiguchi /* 33634a705fefSNaoya Horiguchi * COW mappings require pages in both 33644a705fefSNaoya Horiguchi * parent and child to be set to read. 33654a705fefSNaoya Horiguchi */ 33664a705fefSNaoya Horiguchi make_migration_entry_read(&swp_entry); 33674a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 3368e5251fd4SPunit Agrawal set_huge_swap_pte_at(src, addr, src_pte, 3369e5251fd4SPunit Agrawal entry, sz); 33704a705fefSNaoya Horiguchi } 3371e5251fd4SPunit Agrawal set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); 33724a705fefSNaoya Horiguchi } else { 337334ee645eSJoerg Roedel if (cow) { 33740f10851eSJérôme Glisse /* 33750f10851eSJérôme Glisse * No need to notify as we are downgrading page 33760f10851eSJérôme Glisse * table protection not changing it to point 33770f10851eSJérôme Glisse * to a new page. 33780f10851eSJérôme Glisse * 3379ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 33800f10851eSJérôme Glisse */ 33817f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 338234ee645eSJoerg Roedel } 33830253d634SNaoya Horiguchi entry = huge_ptep_get(src_pte); 338463551ae0SDavid Gibson ptepage = pte_page(entry); 338563551ae0SDavid Gibson get_page(ptepage); 338653f9263bSKirill A. Shutemov page_dup_rmap(ptepage, true); 338763551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 33885d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), dst); 33891c59827dSHugh Dickins } 3390cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 3391cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 339263551ae0SDavid Gibson } 339363551ae0SDavid Gibson 3394e8569dd2SAndreas Sandberg if (cow) 3395ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 3396e8569dd2SAndreas Sandberg 3397e8569dd2SAndreas Sandberg return ret; 339863551ae0SDavid Gibson } 339963551ae0SDavid Gibson 340024669e58SAneesh Kumar K.V void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 340124669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 340224669e58SAneesh Kumar K.V struct page *ref_page) 340363551ae0SDavid Gibson { 340463551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 340563551ae0SDavid Gibson unsigned long address; 3406c7546f8fSDavid Gibson pte_t *ptep; 340763551ae0SDavid Gibson pte_t pte; 3408cb900f41SKirill A. Shutemov spinlock_t *ptl; 340963551ae0SDavid Gibson struct page *page; 3410a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3411a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 3412ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 3413a5516438SAndi Kleen 341463551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 3415a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 3416a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 341763551ae0SDavid Gibson 341807e32661SAneesh Kumar K.V /* 341907e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 342007e32661SAneesh Kumar K.V * to huge page. 342107e32661SAneesh Kumar K.V */ 3422ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 342324669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 3424dff11abeSMike Kravetz 3425dff11abeSMike Kravetz /* 3426dff11abeSMike Kravetz * If sharing possible, alert mmu notifiers of worst case. 3427dff11abeSMike Kravetz */ 34286f4f13e8SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, 34296f4f13e8SJérôme Glisse end); 3430ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 3431ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 3432569f48b8SHillf Danton address = start; 3433569f48b8SHillf Danton for (; address < end; address += sz) { 34347868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, sz); 3435c7546f8fSDavid Gibson if (!ptep) 3436c7546f8fSDavid Gibson continue; 3437c7546f8fSDavid Gibson 3438cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 343931d49da5SAneesh Kumar K.V if (huge_pmd_unshare(mm, &address, ptep)) { 344031d49da5SAneesh Kumar K.V spin_unlock(ptl); 3441dff11abeSMike Kravetz /* 3442dff11abeSMike Kravetz * We just unmapped a page of PMDs by clearing a PUD. 3443dff11abeSMike Kravetz * The caller's TLB flush range should cover this area. 3444dff11abeSMike Kravetz */ 344531d49da5SAneesh Kumar K.V continue; 344631d49da5SAneesh Kumar K.V } 344739dde65cSChen, Kenneth W 34486629326bSHillf Danton pte = huge_ptep_get(ptep); 344931d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 345031d49da5SAneesh Kumar K.V spin_unlock(ptl); 345131d49da5SAneesh Kumar K.V continue; 345231d49da5SAneesh Kumar K.V } 34536629326bSHillf Danton 34546629326bSHillf Danton /* 34559fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 34569fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 34576629326bSHillf Danton */ 34589fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 34599386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 346031d49da5SAneesh Kumar K.V spin_unlock(ptl); 346131d49da5SAneesh Kumar K.V continue; 34628c4894c6SNaoya Horiguchi } 34636629326bSHillf Danton 34646629326bSHillf Danton page = pte_page(pte); 346504f2cbe3SMel Gorman /* 346604f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 346704f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 346804f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 346904f2cbe3SMel Gorman */ 347004f2cbe3SMel Gorman if (ref_page) { 347131d49da5SAneesh Kumar K.V if (page != ref_page) { 347231d49da5SAneesh Kumar K.V spin_unlock(ptl); 347331d49da5SAneesh Kumar K.V continue; 347431d49da5SAneesh Kumar K.V } 347504f2cbe3SMel Gorman /* 347604f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 347704f2cbe3SMel Gorman * future faults in this VMA will fail rather than 347804f2cbe3SMel Gorman * looking like data was lost 347904f2cbe3SMel Gorman */ 348004f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 348104f2cbe3SMel Gorman } 348204f2cbe3SMel Gorman 3483c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 3484b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 3485106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 34866649a386SKen Chen set_page_dirty(page); 34879e81130bSHillf Danton 34885d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 3489d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 349031d49da5SAneesh Kumar K.V 3491cb900f41SKirill A. Shutemov spin_unlock(ptl); 3492e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 349324669e58SAneesh Kumar K.V /* 349431d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 349524669e58SAneesh Kumar K.V */ 349631d49da5SAneesh Kumar K.V if (ref_page) 349731d49da5SAneesh Kumar K.V break; 3498fe1668aeSChen, Kenneth W } 3499ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 350024669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 35011da177e4SLinus Torvalds } 350263551ae0SDavid Gibson 3503d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 3504d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 3505d833352aSMel Gorman unsigned long end, struct page *ref_page) 3506d833352aSMel Gorman { 3507d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 3508d833352aSMel Gorman 3509d833352aSMel Gorman /* 3510d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 3511d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 3512d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 3513d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 3514d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 3515c8c06efaSDavidlohr Bueso * is to clear it before releasing the i_mmap_rwsem. This works 3516d833352aSMel Gorman * because in the context this is called, the VMA is about to be 3517c8c06efaSDavidlohr Bueso * destroyed and the i_mmap_rwsem is held. 3518d833352aSMel Gorman */ 3519d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 3520d833352aSMel Gorman } 3521d833352aSMel Gorman 3522502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 352304f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 3524502717f4SChen, Kenneth W { 352524669e58SAneesh Kumar K.V struct mm_struct *mm; 352624669e58SAneesh Kumar K.V struct mmu_gather tlb; 3527dff11abeSMike Kravetz unsigned long tlb_start = start; 3528dff11abeSMike Kravetz unsigned long tlb_end = end; 3529dff11abeSMike Kravetz 3530dff11abeSMike Kravetz /* 3531dff11abeSMike Kravetz * If shared PMDs were possibly used within this vma range, adjust 3532dff11abeSMike Kravetz * start/end for worst case tlb flushing. 3533dff11abeSMike Kravetz * Note that we can not be sure if PMDs are shared until we try to 3534dff11abeSMike Kravetz * unmap pages. However, we want to make sure TLB flushing covers 3535dff11abeSMike Kravetz * the largest possible range. 3536dff11abeSMike Kravetz */ 3537dff11abeSMike Kravetz adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end); 353824669e58SAneesh Kumar K.V 353924669e58SAneesh Kumar K.V mm = vma->vm_mm; 354024669e58SAneesh Kumar K.V 3541dff11abeSMike Kravetz tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end); 354224669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 3543dff11abeSMike Kravetz tlb_finish_mmu(&tlb, tlb_start, tlb_end); 3544502717f4SChen, Kenneth W } 3545502717f4SChen, Kenneth W 354604f2cbe3SMel Gorman /* 354704f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 354804f2cbe3SMel Gorman * mappping it owns the reserve page for. The intention is to unmap the page 354904f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 355004f2cbe3SMel Gorman * same region. 355104f2cbe3SMel Gorman */ 35522f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 35532a4b3dedSHarvey Harrison struct page *page, unsigned long address) 355404f2cbe3SMel Gorman { 35557526674dSAdam Litke struct hstate *h = hstate_vma(vma); 355604f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 355704f2cbe3SMel Gorman struct address_space *mapping; 355804f2cbe3SMel Gorman pgoff_t pgoff; 355904f2cbe3SMel Gorman 356004f2cbe3SMel Gorman /* 356104f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 356204f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 356304f2cbe3SMel Gorman */ 35647526674dSAdam Litke address = address & huge_page_mask(h); 356536e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 356636e4f20aSMichal Hocko vma->vm_pgoff; 356793c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 356804f2cbe3SMel Gorman 35694eb2b1dcSMel Gorman /* 35704eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 35714eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 35724eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 35734eb2b1dcSMel Gorman */ 357483cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 35756b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 357604f2cbe3SMel Gorman /* Do not unmap the current VMA */ 357704f2cbe3SMel Gorman if (iter_vma == vma) 357804f2cbe3SMel Gorman continue; 357904f2cbe3SMel Gorman 358004f2cbe3SMel Gorman /* 35812f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 35822f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 35832f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 35842f84a899SMel Gorman */ 35852f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 35862f84a899SMel Gorman continue; 35872f84a899SMel Gorman 35882f84a899SMel Gorman /* 358904f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 359004f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 359104f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 359204f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 359304f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 359404f2cbe3SMel Gorman */ 359504f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 359624669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 359724669e58SAneesh Kumar K.V address + huge_page_size(h), page); 359804f2cbe3SMel Gorman } 359983cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 360004f2cbe3SMel Gorman } 360104f2cbe3SMel Gorman 36020fe6e20bSNaoya Horiguchi /* 36030fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 3604ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 3605ef009b25SMichal Hocko * cannot race with other handlers or page migration. 3606ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 36070fe6e20bSNaoya Horiguchi */ 36082b740303SSouptick Joarder static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 3609974e6d66SHuang Ying unsigned long address, pte_t *ptep, 3610cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 36111e8f889bSDavid Gibson { 36123999f52eSAneesh Kumar K.V pte_t pte; 3613a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 36141e8f889bSDavid Gibson struct page *old_page, *new_page; 36152b740303SSouptick Joarder int outside_reserve = 0; 36162b740303SSouptick Joarder vm_fault_t ret = 0; 3617974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 3618ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 36191e8f889bSDavid Gibson 36203999f52eSAneesh Kumar K.V pte = huge_ptep_get(ptep); 36211e8f889bSDavid Gibson old_page = pte_page(pte); 36221e8f889bSDavid Gibson 362304f2cbe3SMel Gorman retry_avoidcopy: 36241e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 36251e8f889bSDavid Gibson * and just make the page writable */ 362637a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 36275a49973dSHugh Dickins page_move_anon_rmap(old_page, vma); 36285b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 362983c54070SNick Piggin return 0; 36301e8f889bSDavid Gibson } 36311e8f889bSDavid Gibson 363204f2cbe3SMel Gorman /* 363304f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 363404f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 363504f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 363604f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 363704f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 363804f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 363904f2cbe3SMel Gorman * of the full address range. 364004f2cbe3SMel Gorman */ 36415944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 364204f2cbe3SMel Gorman old_page != pagecache_page) 364304f2cbe3SMel Gorman outside_reserve = 1; 364404f2cbe3SMel Gorman 364509cbfeafSKirill A. Shutemov get_page(old_page); 3646b76c8cfbSLarry Woodman 3647ad4404a2SDavidlohr Bueso /* 3648ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 3649ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 3650ad4404a2SDavidlohr Bueso */ 3651cb900f41SKirill A. Shutemov spin_unlock(ptl); 36525b7a1d40SHuang Ying new_page = alloc_huge_page(vma, haddr, outside_reserve); 36531e8f889bSDavid Gibson 36542fc39cecSAdam Litke if (IS_ERR(new_page)) { 365504f2cbe3SMel Gorman /* 365604f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 365704f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 365804f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 365904f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 366004f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 366104f2cbe3SMel Gorman */ 366204f2cbe3SMel Gorman if (outside_reserve) { 366309cbfeafSKirill A. Shutemov put_page(old_page); 366404f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 36655b7a1d40SHuang Ying unmap_ref_private(mm, vma, old_page, haddr); 366604f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 3667cb900f41SKirill A. Shutemov spin_lock(ptl); 36685b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 3669a9af0c5dSNaoya Horiguchi if (likely(ptep && 3670a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 367104f2cbe3SMel Gorman goto retry_avoidcopy; 3672a734bcc8SHillf Danton /* 3673cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 3674cb900f41SKirill A. Shutemov * lock, and our job is done. 3675a734bcc8SHillf Danton */ 3676a734bcc8SHillf Danton return 0; 367704f2cbe3SMel Gorman } 367804f2cbe3SMel Gorman 36792b740303SSouptick Joarder ret = vmf_error(PTR_ERR(new_page)); 3680ad4404a2SDavidlohr Bueso goto out_release_old; 36811e8f889bSDavid Gibson } 36821e8f889bSDavid Gibson 36830fe6e20bSNaoya Horiguchi /* 36840fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 36850fe6e20bSNaoya Horiguchi * anon_vma prepared. 36860fe6e20bSNaoya Horiguchi */ 368744e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 3688ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 3689ad4404a2SDavidlohr Bueso goto out_release_all; 369044e2aa93SDean Nelson } 36910fe6e20bSNaoya Horiguchi 3692974e6d66SHuang Ying copy_user_huge_page(new_page, old_page, address, vma, 369347ad8475SAndrea Arcangeli pages_per_huge_page(h)); 36940ed361deSNick Piggin __SetPageUptodate(new_page); 36951e8f889bSDavid Gibson 36967269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, 36976f4f13e8SJérôme Glisse haddr + huge_page_size(h)); 3698ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 3699ad4404a2SDavidlohr Bueso 3700b76c8cfbSLarry Woodman /* 3701cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 3702b76c8cfbSLarry Woodman * before the page tables are altered 3703b76c8cfbSLarry Woodman */ 3704cb900f41SKirill A. Shutemov spin_lock(ptl); 37055b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 3706a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 370707443a85SJoonsoo Kim ClearPagePrivate(new_page); 370807443a85SJoonsoo Kim 37091e8f889bSDavid Gibson /* Break COW */ 37105b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 3711ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, range.end); 37125b7a1d40SHuang Ying set_huge_pte_at(mm, haddr, ptep, 37131e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 3714d281ee61SKirill A. Shutemov page_remove_rmap(old_page, true); 37155b7a1d40SHuang Ying hugepage_add_new_anon_rmap(new_page, vma, haddr); 3716cb6acd01SMike Kravetz set_page_huge_active(new_page); 37171e8f889bSDavid Gibson /* Make the old page be freed below */ 37181e8f889bSDavid Gibson new_page = old_page; 37191e8f889bSDavid Gibson } 3720cb900f41SKirill A. Shutemov spin_unlock(ptl); 3721ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 3722ad4404a2SDavidlohr Bueso out_release_all: 37235b7a1d40SHuang Ying restore_reserve_on_error(h, vma, haddr, new_page); 372409cbfeafSKirill A. Shutemov put_page(new_page); 3725ad4404a2SDavidlohr Bueso out_release_old: 372609cbfeafSKirill A. Shutemov put_page(old_page); 37278312034fSJoonsoo Kim 3728ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 3729ad4404a2SDavidlohr Bueso return ret; 37301e8f889bSDavid Gibson } 37311e8f889bSDavid Gibson 373204f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 3733a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 3734a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 373504f2cbe3SMel Gorman { 373604f2cbe3SMel Gorman struct address_space *mapping; 3737e7c4b0bfSAndy Whitcroft pgoff_t idx; 373804f2cbe3SMel Gorman 373904f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 3740a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 374104f2cbe3SMel Gorman 374204f2cbe3SMel Gorman return find_lock_page(mapping, idx); 374304f2cbe3SMel Gorman } 374404f2cbe3SMel Gorman 37453ae77f43SHugh Dickins /* 37463ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 37473ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 37483ae77f43SHugh Dickins */ 37493ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 37502a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 37512a15efc9SHugh Dickins { 37522a15efc9SHugh Dickins struct address_space *mapping; 37532a15efc9SHugh Dickins pgoff_t idx; 37542a15efc9SHugh Dickins struct page *page; 37552a15efc9SHugh Dickins 37562a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 37572a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 37582a15efc9SHugh Dickins 37592a15efc9SHugh Dickins page = find_get_page(mapping, idx); 37602a15efc9SHugh Dickins if (page) 37612a15efc9SHugh Dickins put_page(page); 37622a15efc9SHugh Dickins return page != NULL; 37632a15efc9SHugh Dickins } 37642a15efc9SHugh Dickins 3765ab76ad54SMike Kravetz int huge_add_to_page_cache(struct page *page, struct address_space *mapping, 3766ab76ad54SMike Kravetz pgoff_t idx) 3767ab76ad54SMike Kravetz { 3768ab76ad54SMike Kravetz struct inode *inode = mapping->host; 3769ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 3770ab76ad54SMike Kravetz int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 3771ab76ad54SMike Kravetz 3772ab76ad54SMike Kravetz if (err) 3773ab76ad54SMike Kravetz return err; 3774ab76ad54SMike Kravetz ClearPagePrivate(page); 3775ab76ad54SMike Kravetz 377622146c3cSMike Kravetz /* 377722146c3cSMike Kravetz * set page dirty so that it will not be removed from cache/file 377822146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 377922146c3cSMike Kravetz */ 378022146c3cSMike Kravetz set_page_dirty(page); 378122146c3cSMike Kravetz 3782ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 3783ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 3784ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 3785ab76ad54SMike Kravetz return 0; 3786ab76ad54SMike Kravetz } 3787ab76ad54SMike Kravetz 37882b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 37892b740303SSouptick Joarder struct vm_area_struct *vma, 37908382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 3791788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 3792ac9b9c66SHugh Dickins { 3793a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 37942b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 3795409eb8c2SHillf Danton int anon_rmap = 0; 37964c887265SAdam Litke unsigned long size; 37974c887265SAdam Litke struct page *page; 37981e8f889bSDavid Gibson pte_t new_pte; 3799cb900f41SKirill A. Shutemov spinlock_t *ptl; 3800285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 3801cb6acd01SMike Kravetz bool new_page = false; 38024c887265SAdam Litke 380304f2cbe3SMel Gorman /* 380404f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 380504f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 380625985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 380704f2cbe3SMel Gorman */ 380804f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 3809910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 381004f2cbe3SMel Gorman current->pid); 381104f2cbe3SMel Gorman return ret; 381204f2cbe3SMel Gorman } 381304f2cbe3SMel Gorman 38144c887265SAdam Litke /* 3815e7c58097SMike Kravetz * Use page lock to guard against racing truncation 3816e7c58097SMike Kravetz * before we get page_table_lock. 38174c887265SAdam Litke */ 3818e7c58097SMike Kravetz retry: 3819e7c58097SMike Kravetz page = find_lock_page(mapping, idx); 3820e7c58097SMike Kravetz if (!page) { 3821a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 3822ebed4bfcSHugh Dickins if (idx >= size) 3823ebed4bfcSHugh Dickins goto out; 38241a1aad8aSMike Kravetz 38251a1aad8aSMike Kravetz /* 38261a1aad8aSMike Kravetz * Check for page in userfault range 38271a1aad8aSMike Kravetz */ 38281a1aad8aSMike Kravetz if (userfaultfd_missing(vma)) { 38291a1aad8aSMike Kravetz u32 hash; 38301a1aad8aSMike Kravetz struct vm_fault vmf = { 38311a1aad8aSMike Kravetz .vma = vma, 3832285b8dcaSHuang Ying .address = haddr, 38331a1aad8aSMike Kravetz .flags = flags, 38341a1aad8aSMike Kravetz /* 38351a1aad8aSMike Kravetz * Hard to debug if it ends up being 38361a1aad8aSMike Kravetz * used by a callee that assumes 38371a1aad8aSMike Kravetz * something about the other 38381a1aad8aSMike Kravetz * uninitialized fields... same as in 38391a1aad8aSMike Kravetz * memory.c 38401a1aad8aSMike Kravetz */ 38411a1aad8aSMike Kravetz }; 38421a1aad8aSMike Kravetz 38431a1aad8aSMike Kravetz /* 3844ddeaab32SMike Kravetz * hugetlb_fault_mutex must be dropped before 3845ddeaab32SMike Kravetz * handling userfault. Reacquire after handling 3846ddeaab32SMike Kravetz * fault to make calling code simpler. 38471a1aad8aSMike Kravetz */ 3848188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 38491a1aad8aSMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 38501a1aad8aSMike Kravetz ret = handle_userfault(&vmf, VM_UFFD_MISSING); 38511a1aad8aSMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 38521a1aad8aSMike Kravetz goto out; 38531a1aad8aSMike Kravetz } 38541a1aad8aSMike Kravetz 3855285b8dcaSHuang Ying page = alloc_huge_page(vma, haddr, 0); 38562fc39cecSAdam Litke if (IS_ERR(page)) { 38574643d67eSMike Kravetz /* 38584643d67eSMike Kravetz * Returning error will result in faulting task being 38594643d67eSMike Kravetz * sent SIGBUS. The hugetlb fault mutex prevents two 38604643d67eSMike Kravetz * tasks from racing to fault in the same page which 38614643d67eSMike Kravetz * could result in false unable to allocate errors. 38624643d67eSMike Kravetz * Page migration does not take the fault mutex, but 38634643d67eSMike Kravetz * does a clear then write of pte's under page table 38644643d67eSMike Kravetz * lock. Page fault code could race with migration, 38654643d67eSMike Kravetz * notice the clear pte and try to allocate a page 38664643d67eSMike Kravetz * here. Before returning error, get ptl and make 38674643d67eSMike Kravetz * sure there really is no pte entry. 38684643d67eSMike Kravetz */ 38694643d67eSMike Kravetz ptl = huge_pte_lock(h, mm, ptep); 38704643d67eSMike Kravetz if (!huge_pte_none(huge_ptep_get(ptep))) { 38714643d67eSMike Kravetz ret = 0; 38724643d67eSMike Kravetz spin_unlock(ptl); 38734643d67eSMike Kravetz goto out; 38744643d67eSMike Kravetz } 38754643d67eSMike Kravetz spin_unlock(ptl); 38762b740303SSouptick Joarder ret = vmf_error(PTR_ERR(page)); 38776bda666aSChristoph Lameter goto out; 38786bda666aSChristoph Lameter } 387947ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 38800ed361deSNick Piggin __SetPageUptodate(page); 3881cb6acd01SMike Kravetz new_page = true; 3882ac9b9c66SHugh Dickins 3883f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 3884ab76ad54SMike Kravetz int err = huge_add_to_page_cache(page, mapping, idx); 38856bda666aSChristoph Lameter if (err) { 38866bda666aSChristoph Lameter put_page(page); 38876bda666aSChristoph Lameter if (err == -EEXIST) 38886bda666aSChristoph Lameter goto retry; 38896bda666aSChristoph Lameter goto out; 38906bda666aSChristoph Lameter } 389123be7468SMel Gorman } else { 38926bda666aSChristoph Lameter lock_page(page); 38930fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 38940fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 38950fe6e20bSNaoya Horiguchi goto backout_unlocked; 389623be7468SMel Gorman } 3897409eb8c2SHillf Danton anon_rmap = 1; 38980fe6e20bSNaoya Horiguchi } 38990fe6e20bSNaoya Horiguchi } else { 390057303d80SAndy Whitcroft /* 3901998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 3902998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 3903998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 3904fd6a03edSNaoya Horiguchi */ 3905fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 3906aa50d3a7SAndi Kleen ret = VM_FAULT_HWPOISON | 3907972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 3908fd6a03edSNaoya Horiguchi goto backout_unlocked; 39096bda666aSChristoph Lameter } 3910998b4382SNaoya Horiguchi } 39111e8f889bSDavid Gibson 391257303d80SAndy Whitcroft /* 391357303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 391457303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 391557303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 391657303d80SAndy Whitcroft * the spinlock. 391757303d80SAndy Whitcroft */ 39185e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 3919285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 39202b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 39212b26736cSAndy Whitcroft goto backout_unlocked; 39222b26736cSAndy Whitcroft } 39235e911373SMike Kravetz /* Just decrements count, does not deallocate */ 3924285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 39255e911373SMike Kravetz } 392657303d80SAndy Whitcroft 39278bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 3928e7c58097SMike Kravetz size = i_size_read(mapping->host) >> huge_page_shift(h); 3929e7c58097SMike Kravetz if (idx >= size) 3930e7c58097SMike Kravetz goto backout; 39314c887265SAdam Litke 393283c54070SNick Piggin ret = 0; 39337f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 39344c887265SAdam Litke goto backout; 39354c887265SAdam Litke 393607443a85SJoonsoo Kim if (anon_rmap) { 393707443a85SJoonsoo Kim ClearPagePrivate(page); 3938285b8dcaSHuang Ying hugepage_add_new_anon_rmap(page, vma, haddr); 3939ac714904SChoi Gi-yong } else 394053f9263bSKirill A. Shutemov page_dup_rmap(page, true); 39411e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 39421e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 3943285b8dcaSHuang Ying set_huge_pte_at(mm, haddr, ptep, new_pte); 39441e8f889bSDavid Gibson 39455d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 3946788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 39471e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 3948974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); 39491e8f889bSDavid Gibson } 39501e8f889bSDavid Gibson 3951cb900f41SKirill A. Shutemov spin_unlock(ptl); 3952cb6acd01SMike Kravetz 3953cb6acd01SMike Kravetz /* 3954cb6acd01SMike Kravetz * Only make newly allocated pages active. Existing pages found 3955cb6acd01SMike Kravetz * in the pagecache could be !page_huge_active() if they have been 3956cb6acd01SMike Kravetz * isolated for migration. 3957cb6acd01SMike Kravetz */ 3958cb6acd01SMike Kravetz if (new_page) 3959cb6acd01SMike Kravetz set_page_huge_active(page); 3960cb6acd01SMike Kravetz 39614c887265SAdam Litke unlock_page(page); 39624c887265SAdam Litke out: 3963ac9b9c66SHugh Dickins return ret; 39644c887265SAdam Litke 39654c887265SAdam Litke backout: 3966cb900f41SKirill A. Shutemov spin_unlock(ptl); 39672b26736cSAndy Whitcroft backout_unlocked: 39684c887265SAdam Litke unlock_page(page); 3969285b8dcaSHuang Ying restore_reserve_on_error(h, vma, haddr, page); 39704c887265SAdam Litke put_page(page); 39714c887265SAdam Litke goto out; 3972ac9b9c66SHugh Dickins } 3973ac9b9c66SHugh Dickins 39748382d914SDavidlohr Bueso #ifdef CONFIG_SMP 3975188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 39768382d914SDavidlohr Bueso { 39778382d914SDavidlohr Bueso unsigned long key[2]; 39788382d914SDavidlohr Bueso u32 hash; 39798382d914SDavidlohr Bueso 39808382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 39818382d914SDavidlohr Bueso key[1] = idx; 39828382d914SDavidlohr Bueso 398355254636SMike Kravetz hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); 39848382d914SDavidlohr Bueso 39858382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 39868382d914SDavidlohr Bueso } 39878382d914SDavidlohr Bueso #else 39888382d914SDavidlohr Bueso /* 39898382d914SDavidlohr Bueso * For uniprocesor systems we always use a single mutex, so just 39908382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 39918382d914SDavidlohr Bueso */ 3992188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 39938382d914SDavidlohr Bueso { 39948382d914SDavidlohr Bueso return 0; 39958382d914SDavidlohr Bueso } 39968382d914SDavidlohr Bueso #endif 39978382d914SDavidlohr Bueso 39982b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 3999788c7df4SHugh Dickins unsigned long address, unsigned int flags) 400086e5216fSAdam Litke { 40018382d914SDavidlohr Bueso pte_t *ptep, entry; 4002cb900f41SKirill A. Shutemov spinlock_t *ptl; 40032b740303SSouptick Joarder vm_fault_t ret; 40048382d914SDavidlohr Bueso u32 hash; 40058382d914SDavidlohr Bueso pgoff_t idx; 40060fe6e20bSNaoya Horiguchi struct page *page = NULL; 400757303d80SAndy Whitcroft struct page *pagecache_page = NULL; 4008a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 40098382d914SDavidlohr Bueso struct address_space *mapping; 40100f792cf9SNaoya Horiguchi int need_wait_lock = 0; 4011285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 401286e5216fSAdam Litke 4013285b8dcaSHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4014fd6a03edSNaoya Horiguchi if (ptep) { 4015fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 4016290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 4017cb900f41SKirill A. Shutemov migration_entry_wait_huge(vma, mm, ptep); 4018290408d4SNaoya Horiguchi return 0; 4019290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 4020aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 4021972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 4022ddeaab32SMike Kravetz } else { 4023b43a9990SMike Kravetz ptep = huge_pte_alloc(mm, haddr, huge_page_size(h)); 4024ddeaab32SMike Kravetz if (!ptep) 4025b43a9990SMike Kravetz return VM_FAULT_OOM; 4026b43a9990SMike Kravetz } 40278382d914SDavidlohr Bueso 4028ddeaab32SMike Kravetz mapping = vma->vm_file->f_mapping; 4029ddeaab32SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 4030ddeaab32SMike Kravetz 40313935baa9SDavid Gibson /* 40323935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 40333935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 40343935baa9SDavid Gibson * the same page in the page cache. 40353935baa9SDavid Gibson */ 4036188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 4037c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 40388382d914SDavidlohr Bueso 40397f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 40407f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 40418382d914SDavidlohr Bueso ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 4042b4d1d99fSDavid Gibson goto out_mutex; 40433935baa9SDavid Gibson } 404486e5216fSAdam Litke 404583c54070SNick Piggin ret = 0; 40461e8f889bSDavid Gibson 404757303d80SAndy Whitcroft /* 40480f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 40490f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 40500f792cf9SNaoya Horiguchi * a active hugepage in pagecache. This goto expects the 2nd page fault, 40510f792cf9SNaoya Horiguchi * and is_hugetlb_entry_(migration|hwpoisoned) check will properly 40520f792cf9SNaoya Horiguchi * handle it. 40530f792cf9SNaoya Horiguchi */ 40540f792cf9SNaoya Horiguchi if (!pte_present(entry)) 40550f792cf9SNaoya Horiguchi goto out_mutex; 40560f792cf9SNaoya Horiguchi 40570f792cf9SNaoya Horiguchi /* 405857303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 405957303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 406057303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 406157303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 406257303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 406357303d80SAndy Whitcroft * consumed. 406457303d80SAndy Whitcroft */ 4065106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 4066285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 40672b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 4068b4d1d99fSDavid Gibson goto out_mutex; 40692b26736cSAndy Whitcroft } 40705e911373SMike Kravetz /* Just decrements count, does not deallocate */ 4071285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 407257303d80SAndy Whitcroft 4073f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 407457303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 4075285b8dcaSHuang Ying vma, haddr); 407657303d80SAndy Whitcroft } 407757303d80SAndy Whitcroft 40780f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 40790fe6e20bSNaoya Horiguchi 40801e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 4081b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 4082cb900f41SKirill A. Shutemov goto out_ptl; 4083b4d1d99fSDavid Gibson 40840f792cf9SNaoya Horiguchi /* 40850f792cf9SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 40860f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 40870f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 40880f792cf9SNaoya Horiguchi */ 40890f792cf9SNaoya Horiguchi page = pte_page(entry); 40900f792cf9SNaoya Horiguchi if (page != pagecache_page) 40910f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 40920f792cf9SNaoya Horiguchi need_wait_lock = 1; 40930f792cf9SNaoya Horiguchi goto out_ptl; 40940f792cf9SNaoya Horiguchi } 40950f792cf9SNaoya Horiguchi 40960f792cf9SNaoya Horiguchi get_page(page); 4097b4d1d99fSDavid Gibson 4098788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 4099106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 4100974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, 4101cb900f41SKirill A. Shutemov pagecache_page, ptl); 41020f792cf9SNaoya Horiguchi goto out_put_page; 4103b4d1d99fSDavid Gibson } 4104106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 4105b4d1d99fSDavid Gibson } 4106b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 4107285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 4108788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 4109285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 41100f792cf9SNaoya Horiguchi out_put_page: 41110f792cf9SNaoya Horiguchi if (page != pagecache_page) 41120f792cf9SNaoya Horiguchi unlock_page(page); 41130f792cf9SNaoya Horiguchi put_page(page); 4114cb900f41SKirill A. Shutemov out_ptl: 4115cb900f41SKirill A. Shutemov spin_unlock(ptl); 411657303d80SAndy Whitcroft 411757303d80SAndy Whitcroft if (pagecache_page) { 411857303d80SAndy Whitcroft unlock_page(pagecache_page); 411957303d80SAndy Whitcroft put_page(pagecache_page); 412057303d80SAndy Whitcroft } 4121b4d1d99fSDavid Gibson out_mutex: 4122c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 41230f792cf9SNaoya Horiguchi /* 41240f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 41250f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 41260f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 41270f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 41280f792cf9SNaoya Horiguchi * here without taking refcount. 41290f792cf9SNaoya Horiguchi */ 41300f792cf9SNaoya Horiguchi if (need_wait_lock) 41310f792cf9SNaoya Horiguchi wait_on_page_locked(page); 41321e8f889bSDavid Gibson return ret; 413386e5216fSAdam Litke } 413486e5216fSAdam Litke 41358fb5debcSMike Kravetz /* 41368fb5debcSMike Kravetz * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with 41378fb5debcSMike Kravetz * modifications for huge pages. 41388fb5debcSMike Kravetz */ 41398fb5debcSMike Kravetz int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, 41408fb5debcSMike Kravetz pte_t *dst_pte, 41418fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 41428fb5debcSMike Kravetz unsigned long dst_addr, 41438fb5debcSMike Kravetz unsigned long src_addr, 41448fb5debcSMike Kravetz struct page **pagep) 41458fb5debcSMike Kravetz { 41461e392147SAndrea Arcangeli struct address_space *mapping; 41471e392147SAndrea Arcangeli pgoff_t idx; 41481e392147SAndrea Arcangeli unsigned long size; 41491c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 41508fb5debcSMike Kravetz struct hstate *h = hstate_vma(dst_vma); 41518fb5debcSMike Kravetz pte_t _dst_pte; 41528fb5debcSMike Kravetz spinlock_t *ptl; 41538fb5debcSMike Kravetz int ret; 41548fb5debcSMike Kravetz struct page *page; 41558fb5debcSMike Kravetz 41568fb5debcSMike Kravetz if (!*pagep) { 41578fb5debcSMike Kravetz ret = -ENOMEM; 41588fb5debcSMike Kravetz page = alloc_huge_page(dst_vma, dst_addr, 0); 41598fb5debcSMike Kravetz if (IS_ERR(page)) 41608fb5debcSMike Kravetz goto out; 41618fb5debcSMike Kravetz 41628fb5debcSMike Kravetz ret = copy_huge_page_from_user(page, 41638fb5debcSMike Kravetz (const void __user *) src_addr, 4164810a56b9SMike Kravetz pages_per_huge_page(h), false); 41658fb5debcSMike Kravetz 41668fb5debcSMike Kravetz /* fallback to copy_from_user outside mmap_sem */ 41678fb5debcSMike Kravetz if (unlikely(ret)) { 41689e368259SAndrea Arcangeli ret = -ENOENT; 41698fb5debcSMike Kravetz *pagep = page; 41708fb5debcSMike Kravetz /* don't free the page */ 41718fb5debcSMike Kravetz goto out; 41728fb5debcSMike Kravetz } 41738fb5debcSMike Kravetz } else { 41748fb5debcSMike Kravetz page = *pagep; 41758fb5debcSMike Kravetz *pagep = NULL; 41768fb5debcSMike Kravetz } 41778fb5debcSMike Kravetz 41788fb5debcSMike Kravetz /* 41798fb5debcSMike Kravetz * The memory barrier inside __SetPageUptodate makes sure that 41808fb5debcSMike Kravetz * preceding stores to the page contents become visible before 41818fb5debcSMike Kravetz * the set_pte_at() write. 41828fb5debcSMike Kravetz */ 41838fb5debcSMike Kravetz __SetPageUptodate(page); 41848fb5debcSMike Kravetz 41851e392147SAndrea Arcangeli mapping = dst_vma->vm_file->f_mapping; 41861e392147SAndrea Arcangeli idx = vma_hugecache_offset(h, dst_vma, dst_addr); 41871e392147SAndrea Arcangeli 41881c9e8defSMike Kravetz /* 41891c9e8defSMike Kravetz * If shared, add to page cache 41901c9e8defSMike Kravetz */ 41911c9e8defSMike Kravetz if (vm_shared) { 41921e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 41931e392147SAndrea Arcangeli ret = -EFAULT; 41941e392147SAndrea Arcangeli if (idx >= size) 41951e392147SAndrea Arcangeli goto out_release_nounlock; 41961c9e8defSMike Kravetz 41971e392147SAndrea Arcangeli /* 41981e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 41991e392147SAndrea Arcangeli * huge_add_to_page_cache() below happens through the 42001e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 42011e392147SAndrea Arcangeli * the caller. 42021e392147SAndrea Arcangeli */ 42031c9e8defSMike Kravetz ret = huge_add_to_page_cache(page, mapping, idx); 42041c9e8defSMike Kravetz if (ret) 42051c9e8defSMike Kravetz goto out_release_nounlock; 42061c9e8defSMike Kravetz } 42071c9e8defSMike Kravetz 42088fb5debcSMike Kravetz ptl = huge_pte_lockptr(h, dst_mm, dst_pte); 42098fb5debcSMike Kravetz spin_lock(ptl); 42108fb5debcSMike Kravetz 42111e392147SAndrea Arcangeli /* 42121e392147SAndrea Arcangeli * Recheck the i_size after holding PT lock to make sure not 42131e392147SAndrea Arcangeli * to leave any page mapped (as page_mapped()) beyond the end 42141e392147SAndrea Arcangeli * of the i_size (remove_inode_hugepages() is strict about 42151e392147SAndrea Arcangeli * enforcing that). If we bail out here, we'll also leave a 42161e392147SAndrea Arcangeli * page in the radix tree in the vm_shared case beyond the end 42171e392147SAndrea Arcangeli * of the i_size, but remove_inode_hugepages() will take care 42181e392147SAndrea Arcangeli * of it as soon as we drop the hugetlb_fault_mutex_table. 42191e392147SAndrea Arcangeli */ 42201e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 42211e392147SAndrea Arcangeli ret = -EFAULT; 42221e392147SAndrea Arcangeli if (idx >= size) 42231e392147SAndrea Arcangeli goto out_release_unlock; 42241e392147SAndrea Arcangeli 42258fb5debcSMike Kravetz ret = -EEXIST; 42268fb5debcSMike Kravetz if (!huge_pte_none(huge_ptep_get(dst_pte))) 42278fb5debcSMike Kravetz goto out_release_unlock; 42288fb5debcSMike Kravetz 42291c9e8defSMike Kravetz if (vm_shared) { 42301c9e8defSMike Kravetz page_dup_rmap(page, true); 42311c9e8defSMike Kravetz } else { 42328fb5debcSMike Kravetz ClearPagePrivate(page); 42338fb5debcSMike Kravetz hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); 42341c9e8defSMike Kravetz } 42358fb5debcSMike Kravetz 42368fb5debcSMike Kravetz _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); 42378fb5debcSMike Kravetz if (dst_vma->vm_flags & VM_WRITE) 42388fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 42398fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 42408fb5debcSMike Kravetz 42418fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 42428fb5debcSMike Kravetz 42438fb5debcSMike Kravetz (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, 42448fb5debcSMike Kravetz dst_vma->vm_flags & VM_WRITE); 42458fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 42468fb5debcSMike Kravetz 42478fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 42488fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 42498fb5debcSMike Kravetz 42508fb5debcSMike Kravetz spin_unlock(ptl); 4251cb6acd01SMike Kravetz set_page_huge_active(page); 42521c9e8defSMike Kravetz if (vm_shared) 42531c9e8defSMike Kravetz unlock_page(page); 42548fb5debcSMike Kravetz ret = 0; 42558fb5debcSMike Kravetz out: 42568fb5debcSMike Kravetz return ret; 42578fb5debcSMike Kravetz out_release_unlock: 42588fb5debcSMike Kravetz spin_unlock(ptl); 42591c9e8defSMike Kravetz if (vm_shared) 42601c9e8defSMike Kravetz unlock_page(page); 42615af10dfdSAndrea Arcangeli out_release_nounlock: 42628fb5debcSMike Kravetz put_page(page); 42638fb5debcSMike Kravetz goto out; 42648fb5debcSMike Kravetz } 42658fb5debcSMike Kravetz 426628a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 426763551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 426828a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 426987ffc118SAndrea Arcangeli long i, unsigned int flags, int *nonblocking) 427063551ae0SDavid Gibson { 4271d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 4272d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 427328a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 4274a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 42752be7cfedSDaniel Jordan int err = -EFAULT; 427663551ae0SDavid Gibson 427763551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 427863551ae0SDavid Gibson pte_t *pte; 4279cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 42802a15efc9SHugh Dickins int absent; 428163551ae0SDavid Gibson struct page *page; 428263551ae0SDavid Gibson 42834c887265SAdam Litke /* 428402057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 428502057967SDavid Rientjes * potentially allocating memory. 428602057967SDavid Rientjes */ 4287fa45f116SDavidlohr Bueso if (fatal_signal_pending(current)) { 428802057967SDavid Rientjes remainder = 0; 428902057967SDavid Rientjes break; 429002057967SDavid Rientjes } 429102057967SDavid Rientjes 429202057967SDavid Rientjes /* 42934c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 42942a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 42954c887265SAdam Litke * first, for the page indexing below to work. 4296cb900f41SKirill A. Shutemov * 4297cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 42984c887265SAdam Litke */ 42997868a208SPunit Agrawal pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), 43007868a208SPunit Agrawal huge_page_size(h)); 4301cb900f41SKirill A. Shutemov if (pte) 4302cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 43032a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 430463551ae0SDavid Gibson 43052a15efc9SHugh Dickins /* 43062a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 43073ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 43083ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 43093ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 43103ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 43112a15efc9SHugh Dickins */ 43123ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 43133ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 4314cb900f41SKirill A. Shutemov if (pte) 4315cb900f41SKirill A. Shutemov spin_unlock(ptl); 43162a15efc9SHugh Dickins remainder = 0; 43172a15efc9SHugh Dickins break; 43182a15efc9SHugh Dickins } 43192a15efc9SHugh Dickins 43209cc3a5bdSNaoya Horiguchi /* 43219cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 43229cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 43239cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 43249cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 43259cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 43269cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 43279cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 43289cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 43299cc3a5bdSNaoya Horiguchi */ 43309cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 4331106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 4332106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 43332b740303SSouptick Joarder vm_fault_t ret; 433487ffc118SAndrea Arcangeli unsigned int fault_flags = 0; 43354c887265SAdam Litke 4336cb900f41SKirill A. Shutemov if (pte) 4337cb900f41SKirill A. Shutemov spin_unlock(ptl); 433887ffc118SAndrea Arcangeli if (flags & FOLL_WRITE) 433987ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_WRITE; 434087ffc118SAndrea Arcangeli if (nonblocking) 434187ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY; 434287ffc118SAndrea Arcangeli if (flags & FOLL_NOWAIT) 434387ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY | 434487ffc118SAndrea Arcangeli FAULT_FLAG_RETRY_NOWAIT; 434587ffc118SAndrea Arcangeli if (flags & FOLL_TRIED) { 434687ffc118SAndrea Arcangeli VM_WARN_ON_ONCE(fault_flags & 434787ffc118SAndrea Arcangeli FAULT_FLAG_ALLOW_RETRY); 434887ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_TRIED; 434987ffc118SAndrea Arcangeli } 435087ffc118SAndrea Arcangeli ret = hugetlb_fault(mm, vma, vaddr, fault_flags); 435187ffc118SAndrea Arcangeli if (ret & VM_FAULT_ERROR) { 43522be7cfedSDaniel Jordan err = vm_fault_to_errno(ret, flags); 43531c59827dSHugh Dickins remainder = 0; 43541c59827dSHugh Dickins break; 43551c59827dSHugh Dickins } 435687ffc118SAndrea Arcangeli if (ret & VM_FAULT_RETRY) { 43571ac25013SAndrea Arcangeli if (nonblocking && 43581ac25013SAndrea Arcangeli !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) 435987ffc118SAndrea Arcangeli *nonblocking = 0; 436087ffc118SAndrea Arcangeli *nr_pages = 0; 436187ffc118SAndrea Arcangeli /* 436287ffc118SAndrea Arcangeli * VM_FAULT_RETRY must not return an 436387ffc118SAndrea Arcangeli * error, it will return zero 436487ffc118SAndrea Arcangeli * instead. 436587ffc118SAndrea Arcangeli * 436687ffc118SAndrea Arcangeli * No need to update "position" as the 436787ffc118SAndrea Arcangeli * caller will not check it after 436887ffc118SAndrea Arcangeli * *nr_pages is set to 0. 436987ffc118SAndrea Arcangeli */ 437087ffc118SAndrea Arcangeli return i; 437187ffc118SAndrea Arcangeli } 437287ffc118SAndrea Arcangeli continue; 437387ffc118SAndrea Arcangeli } 437463551ae0SDavid Gibson 4375a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 43767f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 43778fde12caSLinus Torvalds 43788fde12caSLinus Torvalds /* 43798fde12caSLinus Torvalds * Instead of doing 'try_get_page()' below in the same_page 43808fde12caSLinus Torvalds * loop, just check the count once here. 43818fde12caSLinus Torvalds */ 43828fde12caSLinus Torvalds if (unlikely(page_count(page) <= 0)) { 43838fde12caSLinus Torvalds if (pages) { 43848fde12caSLinus Torvalds spin_unlock(ptl); 43858fde12caSLinus Torvalds remainder = 0; 43868fde12caSLinus Torvalds err = -ENOMEM; 43878fde12caSLinus Torvalds break; 43888fde12caSLinus Torvalds } 43898fde12caSLinus Torvalds } 4390acbfb087SZhigang Lu 4391acbfb087SZhigang Lu /* 4392acbfb087SZhigang Lu * If subpage information not requested, update counters 4393acbfb087SZhigang Lu * and skip the same_page loop below. 4394acbfb087SZhigang Lu */ 4395acbfb087SZhigang Lu if (!pages && !vmas && !pfn_offset && 4396acbfb087SZhigang Lu (vaddr + huge_page_size(h) < vma->vm_end) && 4397acbfb087SZhigang Lu (remainder >= pages_per_huge_page(h))) { 4398acbfb087SZhigang Lu vaddr += huge_page_size(h); 4399acbfb087SZhigang Lu remainder -= pages_per_huge_page(h); 4400acbfb087SZhigang Lu i += pages_per_huge_page(h); 4401acbfb087SZhigang Lu spin_unlock(ptl); 4402acbfb087SZhigang Lu continue; 4403acbfb087SZhigang Lu } 4404acbfb087SZhigang Lu 4405d5d4b0aaSChen, Kenneth W same_page: 4406d6692183SChen, Kenneth W if (pages) { 440769d177c2SAndy Whitcroft pages[i] = mem_map_offset(page, pfn_offset); 4408ddc58f27SKirill A. Shutemov get_page(pages[i]); 4409d6692183SChen, Kenneth W } 441063551ae0SDavid Gibson 441163551ae0SDavid Gibson if (vmas) 441263551ae0SDavid Gibson vmas[i] = vma; 441363551ae0SDavid Gibson 441463551ae0SDavid Gibson vaddr += PAGE_SIZE; 4415d5d4b0aaSChen, Kenneth W ++pfn_offset; 441663551ae0SDavid Gibson --remainder; 441763551ae0SDavid Gibson ++i; 4418d5d4b0aaSChen, Kenneth W if (vaddr < vma->vm_end && remainder && 4419a5516438SAndi Kleen pfn_offset < pages_per_huge_page(h)) { 4420d5d4b0aaSChen, Kenneth W /* 4421d5d4b0aaSChen, Kenneth W * We use pfn_offset to avoid touching the pageframes 4422d5d4b0aaSChen, Kenneth W * of this compound page. 4423d5d4b0aaSChen, Kenneth W */ 4424d5d4b0aaSChen, Kenneth W goto same_page; 4425d5d4b0aaSChen, Kenneth W } 4426cb900f41SKirill A. Shutemov spin_unlock(ptl); 442763551ae0SDavid Gibson } 442828a35716SMichel Lespinasse *nr_pages = remainder; 442987ffc118SAndrea Arcangeli /* 443087ffc118SAndrea Arcangeli * setting position is actually required only if remainder is 443187ffc118SAndrea Arcangeli * not zero but it's faster not to add a "if (remainder)" 443287ffc118SAndrea Arcangeli * branch. 443387ffc118SAndrea Arcangeli */ 443463551ae0SDavid Gibson *position = vaddr; 443563551ae0SDavid Gibson 44362be7cfedSDaniel Jordan return i ? i : err; 443763551ae0SDavid Gibson } 44388f860591SZhang, Yanmin 44395491ae7bSAneesh Kumar K.V #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE 44405491ae7bSAneesh Kumar K.V /* 44415491ae7bSAneesh Kumar K.V * ARCHes with special requirements for evicting HUGETLB backing TLB entries can 44425491ae7bSAneesh Kumar K.V * implement this. 44435491ae7bSAneesh Kumar K.V */ 44445491ae7bSAneesh Kumar K.V #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) 44455491ae7bSAneesh Kumar K.V #endif 44465491ae7bSAneesh Kumar K.V 44477da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 44488f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 44498f860591SZhang, Yanmin { 44508f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 44518f860591SZhang, Yanmin unsigned long start = address; 44528f860591SZhang, Yanmin pte_t *ptep; 44538f860591SZhang, Yanmin pte_t pte; 4454a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 44557da4d641SPeter Zijlstra unsigned long pages = 0; 4456dff11abeSMike Kravetz bool shared_pmd = false; 4457ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 4458dff11abeSMike Kravetz 4459dff11abeSMike Kravetz /* 4460dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 4461ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 4462dff11abeSMike Kravetz * range if PMD sharing is possible. 4463dff11abeSMike Kravetz */ 44647269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 44657269f999SJérôme Glisse 0, vma, mm, start, end); 4466ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 44678f860591SZhang, Yanmin 44688f860591SZhang, Yanmin BUG_ON(address >= end); 4469ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 44708f860591SZhang, Yanmin 4471ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 447283cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 4473a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 4474cb900f41SKirill A. Shutemov spinlock_t *ptl; 44757868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, huge_page_size(h)); 44768f860591SZhang, Yanmin if (!ptep) 44778f860591SZhang, Yanmin continue; 4478cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 44797da4d641SPeter Zijlstra if (huge_pmd_unshare(mm, &address, ptep)) { 44807da4d641SPeter Zijlstra pages++; 4481cb900f41SKirill A. Shutemov spin_unlock(ptl); 4482dff11abeSMike Kravetz shared_pmd = true; 448339dde65cSChen, Kenneth W continue; 44847da4d641SPeter Zijlstra } 4485a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 4486a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 4487a8bda28dSNaoya Horiguchi spin_unlock(ptl); 4488a8bda28dSNaoya Horiguchi continue; 4489a8bda28dSNaoya Horiguchi } 4490a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 4491a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 4492a8bda28dSNaoya Horiguchi 4493a8bda28dSNaoya Horiguchi if (is_write_migration_entry(entry)) { 4494a8bda28dSNaoya Horiguchi pte_t newpte; 4495a8bda28dSNaoya Horiguchi 4496a8bda28dSNaoya Horiguchi make_migration_entry_read(&entry); 4497a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 4498e5251fd4SPunit Agrawal set_huge_swap_pte_at(mm, address, ptep, 4499e5251fd4SPunit Agrawal newpte, huge_page_size(h)); 4500a8bda28dSNaoya Horiguchi pages++; 4501a8bda28dSNaoya Horiguchi } 4502a8bda28dSNaoya Horiguchi spin_unlock(ptl); 4503a8bda28dSNaoya Horiguchi continue; 4504a8bda28dSNaoya Horiguchi } 4505a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 4506023bdd00SAneesh Kumar K.V pte_t old_pte; 4507023bdd00SAneesh Kumar K.V 4508023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 4509023bdd00SAneesh Kumar K.V pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); 4510be7517d6STony Lu pte = arch_make_huge_pte(pte, vma, NULL, 0); 4511023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 45127da4d641SPeter Zijlstra pages++; 45138f860591SZhang, Yanmin } 4514cb900f41SKirill A. Shutemov spin_unlock(ptl); 45158f860591SZhang, Yanmin } 4516d833352aSMel Gorman /* 4517c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 4518d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 4519c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 4520dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 4521dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 4522d833352aSMel Gorman */ 4523dff11abeSMike Kravetz if (shared_pmd) 4524ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 4525dff11abeSMike Kravetz else 45265491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 45270f10851eSJérôme Glisse /* 45280f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 45290f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 45300f10851eSJérôme Glisse * 4531ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 45320f10851eSJérôme Glisse */ 453383cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 4534ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 45357da4d641SPeter Zijlstra 45367da4d641SPeter Zijlstra return pages << h->order; 45378f860591SZhang, Yanmin } 45388f860591SZhang, Yanmin 4539a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 4540a1e78772SMel Gorman long from, long to, 45415a6fe125SMel Gorman struct vm_area_struct *vma, 4542ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 4543e4e574b7SAdam Litke { 454417c9d12eSMel Gorman long ret, chg; 4545a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 454690481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 45479119a41eSJoonsoo Kim struct resv_map *resv_map; 45481c5ecae3SMike Kravetz long gbl_reserve; 4549e4e574b7SAdam Litke 455063489f8eSMike Kravetz /* This should never happen */ 455163489f8eSMike Kravetz if (from > to) { 455263489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 455363489f8eSMike Kravetz return -EINVAL; 455463489f8eSMike Kravetz } 455563489f8eSMike Kravetz 4556a1e78772SMel Gorman /* 455717c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 455817c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 455990481622SDavid Gibson * without using reserves 456017c9d12eSMel Gorman */ 4561ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 456217c9d12eSMel Gorman return 0; 456317c9d12eSMel Gorman 456417c9d12eSMel Gorman /* 4565a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 4566a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 4567a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 4568a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 4569a1e78772SMel Gorman */ 45709119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 4571f27a5136SMike Kravetz /* 4572f27a5136SMike Kravetz * resv_map can not be NULL as hugetlb_reserve_pages is only 4573f27a5136SMike Kravetz * called for inodes for which resv_maps were created (see 4574f27a5136SMike Kravetz * hugetlbfs_get_inode). 4575f27a5136SMike Kravetz */ 45764e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 45779119a41eSJoonsoo Kim 45781406ec9bSJoonsoo Kim chg = region_chg(resv_map, from, to); 45799119a41eSJoonsoo Kim 45809119a41eSJoonsoo Kim } else { 45819119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 45825a6fe125SMel Gorman if (!resv_map) 45835a6fe125SMel Gorman return -ENOMEM; 45845a6fe125SMel Gorman 458517c9d12eSMel Gorman chg = to - from; 458617c9d12eSMel Gorman 45875a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 45885a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 45895a6fe125SMel Gorman } 45905a6fe125SMel Gorman 4591c50ac050SDave Hansen if (chg < 0) { 4592c50ac050SDave Hansen ret = chg; 4593c50ac050SDave Hansen goto out_err; 4594c50ac050SDave Hansen } 459517c9d12eSMel Gorman 45961c5ecae3SMike Kravetz /* 45971c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 45981c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 45991c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 46001c5ecae3SMike Kravetz */ 46011c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 46021c5ecae3SMike Kravetz if (gbl_reserve < 0) { 4603c50ac050SDave Hansen ret = -ENOSPC; 4604c50ac050SDave Hansen goto out_err; 4605c50ac050SDave Hansen } 460617c9d12eSMel Gorman 460717c9d12eSMel Gorman /* 460817c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 460990481622SDavid Gibson * Hand the pages back to the subpool if there are not 461017c9d12eSMel Gorman */ 46111c5ecae3SMike Kravetz ret = hugetlb_acct_memory(h, gbl_reserve); 461217c9d12eSMel Gorman if (ret < 0) { 46131c5ecae3SMike Kravetz /* put back original number of pages, chg */ 46141c5ecae3SMike Kravetz (void)hugepage_subpool_put_pages(spool, chg); 4615c50ac050SDave Hansen goto out_err; 461617c9d12eSMel Gorman } 461717c9d12eSMel Gorman 461817c9d12eSMel Gorman /* 461917c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 462017c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 462117c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 462217c9d12eSMel Gorman * the reservation was and the page cache tells how much of 462317c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 462417c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 462517c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 462617c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 462717c9d12eSMel Gorman * else has to be done for private mappings here 462817c9d12eSMel Gorman */ 462933039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 463033039678SMike Kravetz long add = region_add(resv_map, from, to); 463133039678SMike Kravetz 463233039678SMike Kravetz if (unlikely(chg > add)) { 463333039678SMike Kravetz /* 463433039678SMike Kravetz * pages in this range were added to the reserve 463533039678SMike Kravetz * map between region_chg and region_add. This 463633039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 463733039678SMike Kravetz * the subpool and reserve counts modified above 463833039678SMike Kravetz * based on the difference. 463933039678SMike Kravetz */ 464033039678SMike Kravetz long rsv_adjust; 464133039678SMike Kravetz 464233039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 464333039678SMike Kravetz chg - add); 464433039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 464533039678SMike Kravetz } 464633039678SMike Kravetz } 4647a43a8c39SChen, Kenneth W return 0; 4648c50ac050SDave Hansen out_err: 46495e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 4650ff8c0c53SMike Kravetz /* Don't call region_abort if region_chg failed */ 4651ff8c0c53SMike Kravetz if (chg >= 0) 46525e911373SMike Kravetz region_abort(resv_map, from, to); 4653f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 4654f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 4655c50ac050SDave Hansen return ret; 4656a43a8c39SChen, Kenneth W } 4657a43a8c39SChen, Kenneth W 4658b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 4659b5cec28dSMike Kravetz long freed) 4660a43a8c39SChen, Kenneth W { 4661a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 46624e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 46639119a41eSJoonsoo Kim long chg = 0; 466490481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 46651c5ecae3SMike Kravetz long gbl_reserve; 466645c682a6SKen Chen 4667f27a5136SMike Kravetz /* 4668f27a5136SMike Kravetz * Since this routine can be called in the evict inode path for all 4669f27a5136SMike Kravetz * hugetlbfs inodes, resv_map could be NULL. 4670f27a5136SMike Kravetz */ 4671b5cec28dSMike Kravetz if (resv_map) { 4672b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 4673b5cec28dSMike Kravetz /* 4674b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 4675b5cec28dSMike Kravetz * must be split and another region descriptor can not be 4676b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 4677b5cec28dSMike Kravetz */ 4678b5cec28dSMike Kravetz if (chg < 0) 4679b5cec28dSMike Kravetz return chg; 4680b5cec28dSMike Kravetz } 4681b5cec28dSMike Kravetz 468245c682a6SKen Chen spin_lock(&inode->i_lock); 4683e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 468445c682a6SKen Chen spin_unlock(&inode->i_lock); 468545c682a6SKen Chen 46861c5ecae3SMike Kravetz /* 46871c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 46881c5ecae3SMike Kravetz * reservations to be released may be adjusted. 46891c5ecae3SMike Kravetz */ 46901c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 46911c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 4692b5cec28dSMike Kravetz 4693b5cec28dSMike Kravetz return 0; 4694a43a8c39SChen, Kenneth W } 469593f70f90SNaoya Horiguchi 46963212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 46973212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 46983212b535SSteve Capper struct vm_area_struct *vma, 46993212b535SSteve Capper unsigned long addr, pgoff_t idx) 47003212b535SSteve Capper { 47013212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 47023212b535SSteve Capper svma->vm_start; 47033212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 47043212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 47053212b535SSteve Capper 47063212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 4707de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 4708de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 47093212b535SSteve Capper 47103212b535SSteve Capper /* 47113212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 47123212b535SSteve Capper * page table page. 47133212b535SSteve Capper */ 47143212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 47153212b535SSteve Capper vm_flags != svm_flags || 47163212b535SSteve Capper sbase < svma->vm_start || svma->vm_end < s_end) 47173212b535SSteve Capper return 0; 47183212b535SSteve Capper 47193212b535SSteve Capper return saddr; 47203212b535SSteve Capper } 47213212b535SSteve Capper 472231aafb45SNicholas Krause static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) 47233212b535SSteve Capper { 47243212b535SSteve Capper unsigned long base = addr & PUD_MASK; 47253212b535SSteve Capper unsigned long end = base + PUD_SIZE; 47263212b535SSteve Capper 47273212b535SSteve Capper /* 47283212b535SSteve Capper * check on proper vm_flags and page table alignment 47293212b535SSteve Capper */ 4730017b1660SMike Kravetz if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) 473131aafb45SNicholas Krause return true; 473231aafb45SNicholas Krause return false; 47333212b535SSteve Capper } 47343212b535SSteve Capper 47353212b535SSteve Capper /* 4736017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 4737017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 4738017b1660SMike Kravetz * shared pmd mappings. 4739017b1660SMike Kravetz */ 4740017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 4741017b1660SMike Kravetz unsigned long *start, unsigned long *end) 4742017b1660SMike Kravetz { 4743017b1660SMike Kravetz unsigned long check_addr = *start; 4744017b1660SMike Kravetz 4745017b1660SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 4746017b1660SMike Kravetz return; 4747017b1660SMike Kravetz 4748017b1660SMike Kravetz for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) { 4749017b1660SMike Kravetz unsigned long a_start = check_addr & PUD_MASK; 4750017b1660SMike Kravetz unsigned long a_end = a_start + PUD_SIZE; 4751017b1660SMike Kravetz 4752017b1660SMike Kravetz /* 4753017b1660SMike Kravetz * If sharing is possible, adjust start/end if necessary. 4754017b1660SMike Kravetz */ 4755017b1660SMike Kravetz if (range_in_vma(vma, a_start, a_end)) { 4756017b1660SMike Kravetz if (a_start < *start) 4757017b1660SMike Kravetz *start = a_start; 4758017b1660SMike Kravetz if (a_end > *end) 4759017b1660SMike Kravetz *end = a_end; 4760017b1660SMike Kravetz } 4761017b1660SMike Kravetz } 4762017b1660SMike Kravetz } 4763017b1660SMike Kravetz 4764017b1660SMike Kravetz /* 47653212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 47663212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 47673212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 4768ddeaab32SMike Kravetz * code much cleaner. pmd allocation is essential for the shared case because 4769ddeaab32SMike Kravetz * pud has to be populated inside the same i_mmap_rwsem section - otherwise 4770ddeaab32SMike Kravetz * racing tasks could either miss the sharing (see huge_pte_offset) or select a 4771ddeaab32SMike Kravetz * bad pmd for sharing. 47723212b535SSteve Capper */ 47733212b535SSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 47743212b535SSteve Capper { 47753212b535SSteve Capper struct vm_area_struct *vma = find_vma(mm, addr); 47763212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 47773212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 47783212b535SSteve Capper vma->vm_pgoff; 47793212b535SSteve Capper struct vm_area_struct *svma; 47803212b535SSteve Capper unsigned long saddr; 47813212b535SSteve Capper pte_t *spte = NULL; 47823212b535SSteve Capper pte_t *pte; 4783cb900f41SKirill A. Shutemov spinlock_t *ptl; 47843212b535SSteve Capper 47853212b535SSteve Capper if (!vma_shareable(vma, addr)) 47863212b535SSteve Capper return (pte_t *)pmd_alloc(mm, pud, addr); 47873212b535SSteve Capper 4788930668c3SWaiman Long i_mmap_lock_read(mapping); 47893212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 47903212b535SSteve Capper if (svma == vma) 47913212b535SSteve Capper continue; 47923212b535SSteve Capper 47933212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 47943212b535SSteve Capper if (saddr) { 47957868a208SPunit Agrawal spte = huge_pte_offset(svma->vm_mm, saddr, 47967868a208SPunit Agrawal vma_mmu_pagesize(svma)); 47973212b535SSteve Capper if (spte) { 47983212b535SSteve Capper get_page(virt_to_page(spte)); 47993212b535SSteve Capper break; 48003212b535SSteve Capper } 48013212b535SSteve Capper } 48023212b535SSteve Capper } 48033212b535SSteve Capper 48043212b535SSteve Capper if (!spte) 48053212b535SSteve Capper goto out; 48063212b535SSteve Capper 48078bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 4808dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 48093212b535SSteve Capper pud_populate(mm, pud, 48103212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 4811c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 4812dc6c9a35SKirill A. Shutemov } else { 48133212b535SSteve Capper put_page(virt_to_page(spte)); 4814dc6c9a35SKirill A. Shutemov } 4815cb900f41SKirill A. Shutemov spin_unlock(ptl); 48163212b535SSteve Capper out: 48173212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 4818930668c3SWaiman Long i_mmap_unlock_read(mapping); 48193212b535SSteve Capper return pte; 48203212b535SSteve Capper } 48213212b535SSteve Capper 48223212b535SSteve Capper /* 48233212b535SSteve Capper * unmap huge page backed by shared pte. 48243212b535SSteve Capper * 48253212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 48263212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 48273212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 48283212b535SSteve Capper * 4829ddeaab32SMike Kravetz * called with page table lock held. 48303212b535SSteve Capper * 48313212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 48323212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 48333212b535SSteve Capper */ 48343212b535SSteve Capper int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 48353212b535SSteve Capper { 48363212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 4837c2febafcSKirill A. Shutemov p4d_t *p4d = p4d_offset(pgd, *addr); 4838c2febafcSKirill A. Shutemov pud_t *pud = pud_offset(p4d, *addr); 48393212b535SSteve Capper 48403212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 48413212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 48423212b535SSteve Capper return 0; 48433212b535SSteve Capper 48443212b535SSteve Capper pud_clear(pud); 48453212b535SSteve Capper put_page(virt_to_page(ptep)); 4846dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 48473212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 48483212b535SSteve Capper return 1; 48493212b535SSteve Capper } 48509e5fc74cSSteve Capper #define want_pmd_share() (1) 48519e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 48529e5fc74cSSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 48539e5fc74cSSteve Capper { 48549e5fc74cSSteve Capper return NULL; 48559e5fc74cSSteve Capper } 4856e81f2d22SZhang Zhen 4857e81f2d22SZhang Zhen int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 4858e81f2d22SZhang Zhen { 4859e81f2d22SZhang Zhen return 0; 4860e81f2d22SZhang Zhen } 4861017b1660SMike Kravetz 4862017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 4863017b1660SMike Kravetz unsigned long *start, unsigned long *end) 4864017b1660SMike Kravetz { 4865017b1660SMike Kravetz } 48669e5fc74cSSteve Capper #define want_pmd_share() (0) 48673212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 48683212b535SSteve Capper 48699e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 48709e5fc74cSSteve Capper pte_t *huge_pte_alloc(struct mm_struct *mm, 48719e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 48729e5fc74cSSteve Capper { 48739e5fc74cSSteve Capper pgd_t *pgd; 4874c2febafcSKirill A. Shutemov p4d_t *p4d; 48759e5fc74cSSteve Capper pud_t *pud; 48769e5fc74cSSteve Capper pte_t *pte = NULL; 48779e5fc74cSSteve Capper 48789e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 4879f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 4880f4f0a3d8SKirill A. Shutemov if (!p4d) 4881f4f0a3d8SKirill A. Shutemov return NULL; 4882c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 48839e5fc74cSSteve Capper if (pud) { 48849e5fc74cSSteve Capper if (sz == PUD_SIZE) { 48859e5fc74cSSteve Capper pte = (pte_t *)pud; 48869e5fc74cSSteve Capper } else { 48879e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 48889e5fc74cSSteve Capper if (want_pmd_share() && pud_none(*pud)) 48899e5fc74cSSteve Capper pte = huge_pmd_share(mm, addr, pud); 48909e5fc74cSSteve Capper else 48919e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 48929e5fc74cSSteve Capper } 48939e5fc74cSSteve Capper } 48944e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 48959e5fc74cSSteve Capper 48969e5fc74cSSteve Capper return pte; 48979e5fc74cSSteve Capper } 48989e5fc74cSSteve Capper 48999b19df29SPunit Agrawal /* 49009b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 49019b19df29SPunit Agrawal * entry at address @addr 49029b19df29SPunit Agrawal * 49039b19df29SPunit Agrawal * Return: Pointer to page table or swap entry (PUD or PMD) for 49049b19df29SPunit Agrawal * address @addr, or NULL if a p*d_none() entry is encountered and the 49059b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 49069b19df29SPunit Agrawal * table. 49079b19df29SPunit Agrawal */ 49087868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 49097868a208SPunit Agrawal unsigned long addr, unsigned long sz) 49109e5fc74cSSteve Capper { 49119e5fc74cSSteve Capper pgd_t *pgd; 4912c2febafcSKirill A. Shutemov p4d_t *p4d; 49139e5fc74cSSteve Capper pud_t *pud; 4914c2febafcSKirill A. Shutemov pmd_t *pmd; 49159e5fc74cSSteve Capper 49169e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 4917c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 4918c2febafcSKirill A. Shutemov return NULL; 4919c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 4920c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 4921c2febafcSKirill A. Shutemov return NULL; 49229b19df29SPunit Agrawal 4923c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 49249b19df29SPunit Agrawal if (sz != PUD_SIZE && pud_none(*pud)) 4925c2febafcSKirill A. Shutemov return NULL; 49269b19df29SPunit Agrawal /* hugepage or swap? */ 49279b19df29SPunit Agrawal if (pud_huge(*pud) || !pud_present(*pud)) 49289e5fc74cSSteve Capper return (pte_t *)pud; 49299b19df29SPunit Agrawal 49309e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 49319b19df29SPunit Agrawal if (sz != PMD_SIZE && pmd_none(*pmd)) 49329b19df29SPunit Agrawal return NULL; 49339b19df29SPunit Agrawal /* hugepage or swap? */ 49349b19df29SPunit Agrawal if (pmd_huge(*pmd) || !pmd_present(*pmd)) 49359e5fc74cSSteve Capper return (pte_t *)pmd; 49369b19df29SPunit Agrawal 49379b19df29SPunit Agrawal return NULL; 49389e5fc74cSSteve Capper } 49399e5fc74cSSteve Capper 494061f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 494161f77edaSNaoya Horiguchi 494261f77edaSNaoya Horiguchi /* 494361f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 494461f77edaSNaoya Horiguchi * behavior. 494561f77edaSNaoya Horiguchi */ 494661f77edaSNaoya Horiguchi struct page * __weak 494761f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 494861f77edaSNaoya Horiguchi int write) 494961f77edaSNaoya Horiguchi { 495061f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 495161f77edaSNaoya Horiguchi } 495261f77edaSNaoya Horiguchi 495361f77edaSNaoya Horiguchi struct page * __weak 49544dc71451SAneesh Kumar K.V follow_huge_pd(struct vm_area_struct *vma, 49554dc71451SAneesh Kumar K.V unsigned long address, hugepd_t hpd, int flags, int pdshift) 49564dc71451SAneesh Kumar K.V { 49574dc71451SAneesh Kumar K.V WARN(1, "hugepd follow called with no support for hugepage directory format\n"); 49584dc71451SAneesh Kumar K.V return NULL; 49594dc71451SAneesh Kumar K.V } 49604dc71451SAneesh Kumar K.V 49614dc71451SAneesh Kumar K.V struct page * __weak 49629e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 4963e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 49649e5fc74cSSteve Capper { 4965e66f17ffSNaoya Horiguchi struct page *page = NULL; 4966e66f17ffSNaoya Horiguchi spinlock_t *ptl; 4967c9d398faSNaoya Horiguchi pte_t pte; 4968e66f17ffSNaoya Horiguchi retry: 4969e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 4970e66f17ffSNaoya Horiguchi spin_lock(ptl); 4971e66f17ffSNaoya Horiguchi /* 4972e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 4973e66f17ffSNaoya Horiguchi * unmapped from other threads. 4974e66f17ffSNaoya Horiguchi */ 4975e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 4976e66f17ffSNaoya Horiguchi goto out; 4977c9d398faSNaoya Horiguchi pte = huge_ptep_get((pte_t *)pmd); 4978c9d398faSNaoya Horiguchi if (pte_present(pte)) { 497997534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 4980e66f17ffSNaoya Horiguchi if (flags & FOLL_GET) 4981e66f17ffSNaoya Horiguchi get_page(page); 4982e66f17ffSNaoya Horiguchi } else { 4983c9d398faSNaoya Horiguchi if (is_hugetlb_entry_migration(pte)) { 4984e66f17ffSNaoya Horiguchi spin_unlock(ptl); 4985e66f17ffSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pmd, ptl); 4986e66f17ffSNaoya Horiguchi goto retry; 4987e66f17ffSNaoya Horiguchi } 4988e66f17ffSNaoya Horiguchi /* 4989e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 4990e66f17ffSNaoya Horiguchi * follow_page_mask(). 4991e66f17ffSNaoya Horiguchi */ 4992e66f17ffSNaoya Horiguchi } 4993e66f17ffSNaoya Horiguchi out: 4994e66f17ffSNaoya Horiguchi spin_unlock(ptl); 49959e5fc74cSSteve Capper return page; 49969e5fc74cSSteve Capper } 49979e5fc74cSSteve Capper 499861f77edaSNaoya Horiguchi struct page * __weak 49999e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 5000e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 50019e5fc74cSSteve Capper { 5002e66f17ffSNaoya Horiguchi if (flags & FOLL_GET) 5003e66f17ffSNaoya Horiguchi return NULL; 50049e5fc74cSSteve Capper 5005e66f17ffSNaoya Horiguchi return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 50069e5fc74cSSteve Capper } 50079e5fc74cSSteve Capper 5008faaa5b62SAnshuman Khandual struct page * __weak 5009faaa5b62SAnshuman Khandual follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) 5010faaa5b62SAnshuman Khandual { 5011faaa5b62SAnshuman Khandual if (flags & FOLL_GET) 5012faaa5b62SAnshuman Khandual return NULL; 5013faaa5b62SAnshuman Khandual 5014faaa5b62SAnshuman Khandual return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); 5015faaa5b62SAnshuman Khandual } 5016faaa5b62SAnshuman Khandual 501731caf665SNaoya Horiguchi bool isolate_huge_page(struct page *page, struct list_head *list) 501831caf665SNaoya Horiguchi { 5019bcc54222SNaoya Horiguchi bool ret = true; 5020bcc54222SNaoya Horiguchi 5021309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 502231caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 5023bcc54222SNaoya Horiguchi if (!page_huge_active(page) || !get_page_unless_zero(page)) { 5024bcc54222SNaoya Horiguchi ret = false; 5025bcc54222SNaoya Horiguchi goto unlock; 5026bcc54222SNaoya Horiguchi } 5027bcc54222SNaoya Horiguchi clear_page_huge_active(page); 502831caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 5029bcc54222SNaoya Horiguchi unlock: 503031caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 5031bcc54222SNaoya Horiguchi return ret; 503231caf665SNaoya Horiguchi } 503331caf665SNaoya Horiguchi 503431caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 503531caf665SNaoya Horiguchi { 5036309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 503731caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 5038bcc54222SNaoya Horiguchi set_page_huge_active(page); 503931caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 504031caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 504131caf665SNaoya Horiguchi put_page(page); 504231caf665SNaoya Horiguchi } 5043ab5ac90aSMichal Hocko 5044ab5ac90aSMichal Hocko void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) 5045ab5ac90aSMichal Hocko { 5046ab5ac90aSMichal Hocko struct hstate *h = page_hstate(oldpage); 5047ab5ac90aSMichal Hocko 5048ab5ac90aSMichal Hocko hugetlb_cgroup_migrate(oldpage, newpage); 5049ab5ac90aSMichal Hocko set_page_owner_migrate_reason(newpage, reason); 5050ab5ac90aSMichal Hocko 5051ab5ac90aSMichal Hocko /* 5052ab5ac90aSMichal Hocko * transfer temporary state of the new huge page. This is 5053ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 5054ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 5055ab5ac90aSMichal Hocko * the temporary status. 5056ab5ac90aSMichal Hocko * 5057ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 5058ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 5059ab5ac90aSMichal Hocko * the per-node's. 5060ab5ac90aSMichal Hocko */ 5061ab5ac90aSMichal Hocko if (PageHugeTemporary(newpage)) { 5062ab5ac90aSMichal Hocko int old_nid = page_to_nid(oldpage); 5063ab5ac90aSMichal Hocko int new_nid = page_to_nid(newpage); 5064ab5ac90aSMichal Hocko 5065ab5ac90aSMichal Hocko SetPageHugeTemporary(oldpage); 5066ab5ac90aSMichal Hocko ClearPageHugeTemporary(newpage); 5067ab5ac90aSMichal Hocko 5068ab5ac90aSMichal Hocko spin_lock(&hugetlb_lock); 5069ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 5070ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 5071ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 5072ab5ac90aSMichal Hocko } 5073ab5ac90aSMichal Hocko spin_unlock(&hugetlb_lock); 5074ab5ac90aSMichal Hocko } 5075ab5ac90aSMichal Hocko } 5076