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 653e9fe92aeSMina Almasry static void 654e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, 655e9fe92aeSMina Almasry struct hugetlb_cgroup *h_cg, 656e9fe92aeSMina Almasry struct hstate *h) 657e9fe92aeSMina Almasry { 658e9fe92aeSMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 659e9fe92aeSMina Almasry if (!h_cg || !h) { 660e9fe92aeSMina Almasry resv_map->reservation_counter = NULL; 661e9fe92aeSMina Almasry resv_map->pages_per_hpage = 0; 662e9fe92aeSMina Almasry resv_map->css = NULL; 663e9fe92aeSMina Almasry } else { 664e9fe92aeSMina Almasry resv_map->reservation_counter = 665e9fe92aeSMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 666e9fe92aeSMina Almasry resv_map->pages_per_hpage = pages_per_huge_page(h); 667e9fe92aeSMina Almasry resv_map->css = &h_cg->css; 668e9fe92aeSMina Almasry } 669e9fe92aeSMina Almasry #endif 670e9fe92aeSMina Almasry } 671e9fe92aeSMina Almasry 6729119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 67384afd99bSAndy Whitcroft { 67484afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 6755e911373SMike Kravetz struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 6765e911373SMike Kravetz 6775e911373SMike Kravetz if (!resv_map || !rg) { 6785e911373SMike Kravetz kfree(resv_map); 6795e911373SMike Kravetz kfree(rg); 68084afd99bSAndy Whitcroft return NULL; 6815e911373SMike Kravetz } 68284afd99bSAndy Whitcroft 68384afd99bSAndy Whitcroft kref_init(&resv_map->refs); 6847b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 68584afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 68684afd99bSAndy Whitcroft 6875e911373SMike Kravetz resv_map->adds_in_progress = 0; 688e9fe92aeSMina Almasry /* 689e9fe92aeSMina Almasry * Initialize these to 0. On shared mappings, 0's here indicate these 690e9fe92aeSMina Almasry * fields don't do cgroup accounting. On private mappings, these will be 691e9fe92aeSMina Almasry * re-initialized to the proper values, to indicate that hugetlb cgroup 692e9fe92aeSMina Almasry * reservations are to be un-charged from here. 693e9fe92aeSMina Almasry */ 694e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); 6955e911373SMike Kravetz 6965e911373SMike Kravetz INIT_LIST_HEAD(&resv_map->region_cache); 6975e911373SMike Kravetz list_add(&rg->link, &resv_map->region_cache); 6985e911373SMike Kravetz resv_map->region_cache_count = 1; 6995e911373SMike Kravetz 70084afd99bSAndy Whitcroft return resv_map; 70184afd99bSAndy Whitcroft } 70284afd99bSAndy Whitcroft 7039119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 70484afd99bSAndy Whitcroft { 70584afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 7065e911373SMike Kravetz struct list_head *head = &resv_map->region_cache; 7075e911373SMike Kravetz struct file_region *rg, *trg; 70884afd99bSAndy Whitcroft 70984afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 710feba16e2SMike Kravetz region_del(resv_map, 0, LONG_MAX); 7115e911373SMike Kravetz 7125e911373SMike Kravetz /* ... and any entries left in the cache */ 7135e911373SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 7145e911373SMike Kravetz list_del(&rg->link); 7155e911373SMike Kravetz kfree(rg); 7165e911373SMike Kravetz } 7175e911373SMike Kravetz 7185e911373SMike Kravetz VM_BUG_ON(resv_map->adds_in_progress); 7195e911373SMike Kravetz 72084afd99bSAndy Whitcroft kfree(resv_map); 72184afd99bSAndy Whitcroft } 72284afd99bSAndy Whitcroft 7234e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 7244e35f483SJoonsoo Kim { 725f27a5136SMike Kravetz /* 726f27a5136SMike Kravetz * At inode evict time, i_mapping may not point to the original 727f27a5136SMike Kravetz * address space within the inode. This original address space 728f27a5136SMike Kravetz * contains the pointer to the resv_map. So, always use the 729f27a5136SMike Kravetz * address space embedded within the inode. 730f27a5136SMike Kravetz * The VERY common case is inode->mapping == &inode->i_data but, 731f27a5136SMike Kravetz * this may not be true for device special inodes. 732f27a5136SMike Kravetz */ 733f27a5136SMike Kravetz return (struct resv_map *)(&inode->i_data)->private_data; 7344e35f483SJoonsoo Kim } 7354e35f483SJoonsoo Kim 73684afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 737a1e78772SMel Gorman { 73881d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 7394e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 7404e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 7414e35f483SJoonsoo Kim struct inode *inode = mapping->host; 7424e35f483SJoonsoo Kim 7434e35f483SJoonsoo Kim return inode_resv_map(inode); 7444e35f483SJoonsoo Kim 7454e35f483SJoonsoo Kim } else { 74684afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 74784afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 7484e35f483SJoonsoo Kim } 749a1e78772SMel Gorman } 750a1e78772SMel Gorman 75184afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 752a1e78772SMel Gorman { 75381d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 75481d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 755a1e78772SMel Gorman 75684afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 75784afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 75804f2cbe3SMel Gorman } 75904f2cbe3SMel Gorman 76004f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 76104f2cbe3SMel Gorman { 76281d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 76381d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 764e7c4b0bfSAndy Whitcroft 765e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 76604f2cbe3SMel Gorman } 76704f2cbe3SMel Gorman 76804f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 76904f2cbe3SMel Gorman { 77081d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 771e7c4b0bfSAndy Whitcroft 772e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 773a1e78772SMel Gorman } 774a1e78772SMel Gorman 77504f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 776a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 777a1e78772SMel Gorman { 77881d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 779f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 780a1e78772SMel Gorman vma->vm_private_data = (void *)0; 781a1e78772SMel Gorman } 782a1e78772SMel Gorman 783a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 784559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 785a1e78772SMel Gorman { 786af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 787af0ed73eSJoonsoo Kim /* 788af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 789af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 790af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 791af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 792af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 793af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 794af0ed73eSJoonsoo Kim * properly, so add work-around here. 795af0ed73eSJoonsoo Kim */ 796af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 797559ec2f8SNicholas Krause return true; 798af0ed73eSJoonsoo Kim else 799559ec2f8SNicholas Krause return false; 800af0ed73eSJoonsoo Kim } 801a63884e9SJoonsoo Kim 802a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 8031fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 8041fb1b0e9SMike Kravetz /* 8051fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 8061fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 8071fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 8081fb1b0e9SMike Kravetz * fallocate. In this case, there really are no reverves to 8091fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 8101fb1b0e9SMike Kravetz */ 8111fb1b0e9SMike Kravetz if (chg) 8121fb1b0e9SMike Kravetz return false; 8131fb1b0e9SMike Kravetz else 814559ec2f8SNicholas Krause return true; 8151fb1b0e9SMike Kravetz } 816a63884e9SJoonsoo Kim 817a63884e9SJoonsoo Kim /* 818a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 819a63884e9SJoonsoo Kim * private mappings. 820a63884e9SJoonsoo Kim */ 82167961f9dSMike Kravetz if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 82267961f9dSMike Kravetz /* 82367961f9dSMike Kravetz * Like the shared case above, a hole punch or truncate 82467961f9dSMike Kravetz * could have been performed on the private mapping. 82567961f9dSMike Kravetz * Examine the value of chg to determine if reserves 82667961f9dSMike Kravetz * actually exist or were previously consumed. 82767961f9dSMike Kravetz * Very Subtle - The value of chg comes from a previous 82867961f9dSMike Kravetz * call to vma_needs_reserves(). The reserve map for 82967961f9dSMike Kravetz * private mappings has different (opposite) semantics 83067961f9dSMike Kravetz * than that of shared mappings. vma_needs_reserves() 83167961f9dSMike Kravetz * has already taken this difference in semantics into 83267961f9dSMike Kravetz * account. Therefore, the meaning of chg is the same 83367961f9dSMike Kravetz * as in the shared case above. Code could easily be 83467961f9dSMike Kravetz * combined, but keeping it separate draws attention to 83567961f9dSMike Kravetz * subtle differences. 83667961f9dSMike Kravetz */ 83767961f9dSMike Kravetz if (chg) 83867961f9dSMike Kravetz return false; 83967961f9dSMike Kravetz else 840559ec2f8SNicholas Krause return true; 84167961f9dSMike Kravetz } 842a63884e9SJoonsoo Kim 843559ec2f8SNicholas Krause return false; 844a1e78772SMel Gorman } 845a1e78772SMel Gorman 846a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 8471da177e4SLinus Torvalds { 8481da177e4SLinus Torvalds int nid = page_to_nid(page); 8490edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 850a5516438SAndi Kleen h->free_huge_pages++; 851a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 8521da177e4SLinus Torvalds } 8531da177e4SLinus Torvalds 85494310cbcSAnshuman Khandual static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) 855bf50bab2SNaoya Horiguchi { 856bf50bab2SNaoya Horiguchi struct page *page; 857bf50bab2SNaoya Horiguchi 858c8721bbbSNaoya Horiguchi list_for_each_entry(page, &h->hugepage_freelists[nid], lru) 859243abd5bSNaoya Horiguchi if (!PageHWPoison(page)) 860c8721bbbSNaoya Horiguchi break; 861c8721bbbSNaoya Horiguchi /* 862c8721bbbSNaoya Horiguchi * if 'non-isolated free hugepage' not found on the list, 863c8721bbbSNaoya Horiguchi * the allocation fails. 864c8721bbbSNaoya Horiguchi */ 865c8721bbbSNaoya Horiguchi if (&h->hugepage_freelists[nid] == &page->lru) 866bf50bab2SNaoya Horiguchi return NULL; 8670edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 868a9869b83SNaoya Horiguchi set_page_refcounted(page); 869bf50bab2SNaoya Horiguchi h->free_huge_pages--; 870bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 871bf50bab2SNaoya Horiguchi return page; 872bf50bab2SNaoya Horiguchi } 873bf50bab2SNaoya Horiguchi 8743e59fcb0SMichal Hocko static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, 8753e59fcb0SMichal Hocko nodemask_t *nmask) 87694310cbcSAnshuman Khandual { 8773e59fcb0SMichal Hocko unsigned int cpuset_mems_cookie; 8783e59fcb0SMichal Hocko struct zonelist *zonelist; 8793e59fcb0SMichal Hocko struct zone *zone; 8803e59fcb0SMichal Hocko struct zoneref *z; 88198fa15f3SAnshuman Khandual int node = NUMA_NO_NODE; 8823e59fcb0SMichal Hocko 8833e59fcb0SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 8843e59fcb0SMichal Hocko 8853e59fcb0SMichal Hocko retry_cpuset: 8863e59fcb0SMichal Hocko cpuset_mems_cookie = read_mems_allowed_begin(); 8873e59fcb0SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 88894310cbcSAnshuman Khandual struct page *page; 88994310cbcSAnshuman Khandual 8903e59fcb0SMichal Hocko if (!cpuset_zone_allowed(zone, gfp_mask)) 8913e59fcb0SMichal Hocko continue; 8923e59fcb0SMichal Hocko /* 8933e59fcb0SMichal Hocko * no need to ask again on the same node. Pool is node rather than 8943e59fcb0SMichal Hocko * zone aware 8953e59fcb0SMichal Hocko */ 8963e59fcb0SMichal Hocko if (zone_to_nid(zone) == node) 8973e59fcb0SMichal Hocko continue; 8983e59fcb0SMichal Hocko node = zone_to_nid(zone); 89994310cbcSAnshuman Khandual 90094310cbcSAnshuman Khandual page = dequeue_huge_page_node_exact(h, node); 90194310cbcSAnshuman Khandual if (page) 90294310cbcSAnshuman Khandual return page; 90394310cbcSAnshuman Khandual } 9043e59fcb0SMichal Hocko if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 9053e59fcb0SMichal Hocko goto retry_cpuset; 9063e59fcb0SMichal Hocko 90794310cbcSAnshuman Khandual return NULL; 90894310cbcSAnshuman Khandual } 90994310cbcSAnshuman Khandual 91086cdb465SNaoya Horiguchi /* Movability of hugepages depends on migration support. */ 91186cdb465SNaoya Horiguchi static inline gfp_t htlb_alloc_mask(struct hstate *h) 91286cdb465SNaoya Horiguchi { 9137ed2c31dSAnshuman Khandual if (hugepage_movable_supported(h)) 91486cdb465SNaoya Horiguchi return GFP_HIGHUSER_MOVABLE; 91586cdb465SNaoya Horiguchi else 91686cdb465SNaoya Horiguchi return GFP_HIGHUSER; 91786cdb465SNaoya Horiguchi } 91886cdb465SNaoya Horiguchi 919a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 920a5516438SAndi Kleen struct vm_area_struct *vma, 921af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 922af0ed73eSJoonsoo Kim long chg) 9231da177e4SLinus Torvalds { 9243e59fcb0SMichal Hocko struct page *page; 925480eccf9SLee Schermerhorn struct mempolicy *mpol; 92604ec6264SVlastimil Babka gfp_t gfp_mask; 9273e59fcb0SMichal Hocko nodemask_t *nodemask; 92804ec6264SVlastimil Babka int nid; 9291da177e4SLinus Torvalds 930a1e78772SMel Gorman /* 931a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 932a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 933a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 934a1e78772SMel Gorman */ 935af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg) && 936a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 937c0ff7453SMiao Xie goto err; 938a1e78772SMel Gorman 93904f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 940a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 9416eab04a8SJustin P. Mattock goto err; 94204f2cbe3SMel Gorman 94304ec6264SVlastimil Babka gfp_mask = htlb_alloc_mask(h); 94404ec6264SVlastimil Babka nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 9453e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 9463e59fcb0SMichal Hocko if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { 94707443a85SJoonsoo Kim SetPagePrivate(page); 948a63884e9SJoonsoo Kim h->resv_huge_pages--; 949bf50bab2SNaoya Horiguchi } 950cc9a6c87SMel Gorman 951cc9a6c87SMel Gorman mpol_cond_put(mpol); 952cc9a6c87SMel Gorman return page; 953cc9a6c87SMel Gorman 954c0ff7453SMiao Xie err: 955cc9a6c87SMel Gorman return NULL; 9561da177e4SLinus Torvalds } 9571da177e4SLinus Torvalds 9581cac6f2cSLuiz Capitulino /* 9591cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 9601cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 9611cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 9621cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 9631cac6f2cSLuiz Capitulino * node for alloc or free. 9641cac6f2cSLuiz Capitulino */ 9651cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 9661cac6f2cSLuiz Capitulino { 9670edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 9681cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 9691cac6f2cSLuiz Capitulino 9701cac6f2cSLuiz Capitulino return nid; 9711cac6f2cSLuiz Capitulino } 9721cac6f2cSLuiz Capitulino 9731cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 9741cac6f2cSLuiz Capitulino { 9751cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 9761cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 9771cac6f2cSLuiz Capitulino return nid; 9781cac6f2cSLuiz Capitulino } 9791cac6f2cSLuiz Capitulino 9801cac6f2cSLuiz Capitulino /* 9811cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 9821cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 9831cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 9841cac6f2cSLuiz Capitulino * mask. 9851cac6f2cSLuiz Capitulino */ 9861cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 9871cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 9881cac6f2cSLuiz Capitulino { 9891cac6f2cSLuiz Capitulino int nid; 9901cac6f2cSLuiz Capitulino 9911cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 9921cac6f2cSLuiz Capitulino 9931cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 9941cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 9951cac6f2cSLuiz Capitulino 9961cac6f2cSLuiz Capitulino return nid; 9971cac6f2cSLuiz Capitulino } 9981cac6f2cSLuiz Capitulino 9991cac6f2cSLuiz Capitulino /* 10001cac6f2cSLuiz Capitulino * helper for free_pool_huge_page() - return the previously saved 10011cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 10021cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 10031cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 10041cac6f2cSLuiz Capitulino */ 10051cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 10061cac6f2cSLuiz Capitulino { 10071cac6f2cSLuiz Capitulino int nid; 10081cac6f2cSLuiz Capitulino 10091cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 10101cac6f2cSLuiz Capitulino 10111cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 10121cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 10131cac6f2cSLuiz Capitulino 10141cac6f2cSLuiz Capitulino return nid; 10151cac6f2cSLuiz Capitulino } 10161cac6f2cSLuiz Capitulino 10171cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 10181cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 10191cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 10201cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 10211cac6f2cSLuiz Capitulino nr_nodes--) 10221cac6f2cSLuiz Capitulino 10231cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 10241cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 10251cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 10261cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 10271cac6f2cSLuiz Capitulino nr_nodes--) 10281cac6f2cSLuiz Capitulino 1029e1073d1eSAneesh Kumar K.V #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 1030944d9fecSLuiz Capitulino static void destroy_compound_gigantic_page(struct page *page, 1031d00181b9SKirill A. Shutemov unsigned int order) 1032944d9fecSLuiz Capitulino { 1033944d9fecSLuiz Capitulino int i; 1034944d9fecSLuiz Capitulino int nr_pages = 1 << order; 1035944d9fecSLuiz Capitulino struct page *p = page + 1; 1036944d9fecSLuiz Capitulino 1037c8cc708aSGerald Schaefer atomic_set(compound_mapcount_ptr(page), 0); 103847e29d32SJohn Hubbard if (hpage_pincount_available(page)) 103947e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 104047e29d32SJohn Hubbard 1041944d9fecSLuiz Capitulino for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 10421d798ca3SKirill A. Shutemov clear_compound_head(p); 1043944d9fecSLuiz Capitulino set_page_refcounted(p); 1044944d9fecSLuiz Capitulino } 1045944d9fecSLuiz Capitulino 1046944d9fecSLuiz Capitulino set_compound_order(page, 0); 1047944d9fecSLuiz Capitulino __ClearPageHead(page); 1048944d9fecSLuiz Capitulino } 1049944d9fecSLuiz Capitulino 1050d00181b9SKirill A. Shutemov static void free_gigantic_page(struct page *page, unsigned int order) 1051944d9fecSLuiz Capitulino { 1052944d9fecSLuiz Capitulino free_contig_range(page_to_pfn(page), 1 << order); 1053944d9fecSLuiz Capitulino } 1054944d9fecSLuiz Capitulino 10554eb0716eSAlexandre Ghiti #ifdef CONFIG_CONTIG_ALLOC 1056d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 1057d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) 1058944d9fecSLuiz Capitulino { 10595e27a2dfSAnshuman Khandual unsigned long nr_pages = 1UL << huge_page_order(h); 1060944d9fecSLuiz Capitulino 10615e27a2dfSAnshuman Khandual return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); 1062944d9fecSLuiz Capitulino } 1063944d9fecSLuiz Capitulino 1064944d9fecSLuiz Capitulino static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); 1065d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order); 10664eb0716eSAlexandre Ghiti #else /* !CONFIG_CONTIG_ALLOC */ 10674eb0716eSAlexandre Ghiti static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 10684eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 10694eb0716eSAlexandre Ghiti { 10704eb0716eSAlexandre Ghiti return NULL; 10714eb0716eSAlexandre Ghiti } 10724eb0716eSAlexandre Ghiti #endif /* CONFIG_CONTIG_ALLOC */ 1073944d9fecSLuiz Capitulino 1074e1073d1eSAneesh Kumar K.V #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 1075d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 10764eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 10774eb0716eSAlexandre Ghiti { 10784eb0716eSAlexandre Ghiti return NULL; 10794eb0716eSAlexandre Ghiti } 1080d00181b9SKirill A. Shutemov static inline void free_gigantic_page(struct page *page, unsigned int order) { } 1081944d9fecSLuiz Capitulino static inline void destroy_compound_gigantic_page(struct page *page, 1082d00181b9SKirill A. Shutemov unsigned int order) { } 1083944d9fecSLuiz Capitulino #endif 1084944d9fecSLuiz Capitulino 1085a5516438SAndi Kleen static void update_and_free_page(struct hstate *h, struct page *page) 10866af2acb6SAdam Litke { 10876af2acb6SAdam Litke int i; 1088a5516438SAndi Kleen 10894eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1090944d9fecSLuiz Capitulino return; 109118229df5SAndy Whitcroft 1092a5516438SAndi Kleen h->nr_huge_pages--; 1093a5516438SAndi Kleen h->nr_huge_pages_node[page_to_nid(page)]--; 1094a5516438SAndi Kleen for (i = 0; i < pages_per_huge_page(h); i++) { 109532f84528SChris Forbes page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 109632f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1097a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1098a7407a27SLuiz Capitulino 1 << PG_writeback); 10996af2acb6SAdam Litke } 1100309381feSSasha Levin VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 11011adc4d41SMina Almasry VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page); 1102f1e61557SKirill A. Shutemov set_compound_page_dtor(page, NULL_COMPOUND_DTOR); 11036af2acb6SAdam Litke set_page_refcounted(page); 1104944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) { 1105944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 1106944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 1107944d9fecSLuiz Capitulino } else { 1108a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 11096af2acb6SAdam Litke } 1110944d9fecSLuiz Capitulino } 11116af2acb6SAdam Litke 1112e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1113e5ff2159SAndi Kleen { 1114e5ff2159SAndi Kleen struct hstate *h; 1115e5ff2159SAndi Kleen 1116e5ff2159SAndi Kleen for_each_hstate(h) { 1117e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1118e5ff2159SAndi Kleen return h; 1119e5ff2159SAndi Kleen } 1120e5ff2159SAndi Kleen return NULL; 1121e5ff2159SAndi Kleen } 1122e5ff2159SAndi Kleen 1123bcc54222SNaoya Horiguchi /* 1124bcc54222SNaoya Horiguchi * Test to determine whether the hugepage is "active/in-use" (i.e. being linked 1125bcc54222SNaoya Horiguchi * to hstate->hugepage_activelist.) 1126bcc54222SNaoya Horiguchi * 1127bcc54222SNaoya Horiguchi * This function can be called for tail pages, but never returns true for them. 1128bcc54222SNaoya Horiguchi */ 1129bcc54222SNaoya Horiguchi bool page_huge_active(struct page *page) 1130bcc54222SNaoya Horiguchi { 1131bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHuge(page), page); 1132bcc54222SNaoya Horiguchi return PageHead(page) && PagePrivate(&page[1]); 1133bcc54222SNaoya Horiguchi } 1134bcc54222SNaoya Horiguchi 1135bcc54222SNaoya Horiguchi /* never called for tail page */ 1136bcc54222SNaoya Horiguchi static void set_page_huge_active(struct page *page) 1137bcc54222SNaoya Horiguchi { 1138bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 1139bcc54222SNaoya Horiguchi SetPagePrivate(&page[1]); 1140bcc54222SNaoya Horiguchi } 1141bcc54222SNaoya Horiguchi 1142bcc54222SNaoya Horiguchi static void clear_page_huge_active(struct page *page) 1143bcc54222SNaoya Horiguchi { 1144bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 1145bcc54222SNaoya Horiguchi ClearPagePrivate(&page[1]); 1146bcc54222SNaoya Horiguchi } 1147bcc54222SNaoya Horiguchi 1148ab5ac90aSMichal Hocko /* 1149ab5ac90aSMichal Hocko * Internal hugetlb specific page flag. Do not use outside of the hugetlb 1150ab5ac90aSMichal Hocko * code 1151ab5ac90aSMichal Hocko */ 1152ab5ac90aSMichal Hocko static inline bool PageHugeTemporary(struct page *page) 1153ab5ac90aSMichal Hocko { 1154ab5ac90aSMichal Hocko if (!PageHuge(page)) 1155ab5ac90aSMichal Hocko return false; 1156ab5ac90aSMichal Hocko 1157ab5ac90aSMichal Hocko return (unsigned long)page[2].mapping == -1U; 1158ab5ac90aSMichal Hocko } 1159ab5ac90aSMichal Hocko 1160ab5ac90aSMichal Hocko static inline void SetPageHugeTemporary(struct page *page) 1161ab5ac90aSMichal Hocko { 1162ab5ac90aSMichal Hocko page[2].mapping = (void *)-1U; 1163ab5ac90aSMichal Hocko } 1164ab5ac90aSMichal Hocko 1165ab5ac90aSMichal Hocko static inline void ClearPageHugeTemporary(struct page *page) 1166ab5ac90aSMichal Hocko { 1167ab5ac90aSMichal Hocko page[2].mapping = NULL; 1168ab5ac90aSMichal Hocko } 1169ab5ac90aSMichal Hocko 1170c77c0a8aSWaiman Long static void __free_huge_page(struct page *page) 117127a85ef1SDavid Gibson { 1172a5516438SAndi Kleen /* 1173a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1174a5516438SAndi Kleen * compound page destructor. 1175a5516438SAndi Kleen */ 1176e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 11777893d1d5SAdam Litke int nid = page_to_nid(page); 117890481622SDavid Gibson struct hugepage_subpool *spool = 117990481622SDavid Gibson (struct hugepage_subpool *)page_private(page); 118007443a85SJoonsoo Kim bool restore_reserve; 118127a85ef1SDavid Gibson 1182b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1183b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_mapcount(page), page); 11848ace22bcSYongkai Wu 11858ace22bcSYongkai Wu set_page_private(page, 0); 11868ace22bcSYongkai Wu page->mapping = NULL; 118707443a85SJoonsoo Kim restore_reserve = PagePrivate(page); 118816c794b4SJoonsoo Kim ClearPagePrivate(page); 118927a85ef1SDavid Gibson 11901c5ecae3SMike Kravetz /* 11910919e1b6SMike Kravetz * If PagePrivate() was set on page, page allocation consumed a 11920919e1b6SMike Kravetz * reservation. If the page was associated with a subpool, there 11930919e1b6SMike Kravetz * would have been a page reserved in the subpool before allocation 11940919e1b6SMike Kravetz * via hugepage_subpool_get_pages(). Since we are 'restoring' the 11950919e1b6SMike Kravetz * reservtion, do not call hugepage_subpool_put_pages() as this will 11960919e1b6SMike Kravetz * remove the reserved page from the subpool. 11970919e1b6SMike Kravetz */ 11980919e1b6SMike Kravetz if (!restore_reserve) { 11990919e1b6SMike Kravetz /* 12000919e1b6SMike Kravetz * A return code of zero implies that the subpool will be 12010919e1b6SMike Kravetz * under its minimum size if the reservation is not restored 12020919e1b6SMike Kravetz * after page is free. Therefore, force restore_reserve 12030919e1b6SMike Kravetz * operation. 12041c5ecae3SMike Kravetz */ 12051c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 12061c5ecae3SMike Kravetz restore_reserve = true; 12070919e1b6SMike Kravetz } 12081c5ecae3SMike Kravetz 120927a85ef1SDavid Gibson spin_lock(&hugetlb_lock); 1210bcc54222SNaoya Horiguchi clear_page_huge_active(page); 12116d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 12126d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 121307443a85SJoonsoo Kim if (restore_reserve) 121407443a85SJoonsoo Kim h->resv_huge_pages++; 121507443a85SJoonsoo Kim 1216ab5ac90aSMichal Hocko if (PageHugeTemporary(page)) { 1217ab5ac90aSMichal Hocko list_del(&page->lru); 1218ab5ac90aSMichal Hocko ClearPageHugeTemporary(page); 1219ab5ac90aSMichal Hocko update_and_free_page(h, page); 1220ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 12210edaecfaSAneesh Kumar K.V /* remove the page from active list */ 12220edaecfaSAneesh Kumar K.V list_del(&page->lru); 1223a5516438SAndi Kleen update_and_free_page(h, page); 1224a5516438SAndi Kleen h->surplus_huge_pages--; 1225a5516438SAndi Kleen h->surplus_huge_pages_node[nid]--; 12267893d1d5SAdam Litke } else { 12275d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1228a5516438SAndi Kleen enqueue_huge_page(h, page); 12297893d1d5SAdam Litke } 123027a85ef1SDavid Gibson spin_unlock(&hugetlb_lock); 123127a85ef1SDavid Gibson } 123227a85ef1SDavid Gibson 1233c77c0a8aSWaiman Long /* 1234c77c0a8aSWaiman Long * As free_huge_page() can be called from a non-task context, we have 1235c77c0a8aSWaiman Long * to defer the actual freeing in a workqueue to prevent potential 1236c77c0a8aSWaiman Long * hugetlb_lock deadlock. 1237c77c0a8aSWaiman Long * 1238c77c0a8aSWaiman Long * free_hpage_workfn() locklessly retrieves the linked list of pages to 1239c77c0a8aSWaiman Long * be freed and frees them one-by-one. As the page->mapping pointer is 1240c77c0a8aSWaiman Long * going to be cleared in __free_huge_page() anyway, it is reused as the 1241c77c0a8aSWaiman Long * llist_node structure of a lockless linked list of huge pages to be freed. 1242c77c0a8aSWaiman Long */ 1243c77c0a8aSWaiman Long static LLIST_HEAD(hpage_freelist); 1244c77c0a8aSWaiman Long 1245c77c0a8aSWaiman Long static void free_hpage_workfn(struct work_struct *work) 1246c77c0a8aSWaiman Long { 1247c77c0a8aSWaiman Long struct llist_node *node; 1248c77c0a8aSWaiman Long struct page *page; 1249c77c0a8aSWaiman Long 1250c77c0a8aSWaiman Long node = llist_del_all(&hpage_freelist); 1251c77c0a8aSWaiman Long 1252c77c0a8aSWaiman Long while (node) { 1253c77c0a8aSWaiman Long page = container_of((struct address_space **)node, 1254c77c0a8aSWaiman Long struct page, mapping); 1255c77c0a8aSWaiman Long node = node->next; 1256c77c0a8aSWaiman Long __free_huge_page(page); 1257c77c0a8aSWaiman Long } 1258c77c0a8aSWaiman Long } 1259c77c0a8aSWaiman Long static DECLARE_WORK(free_hpage_work, free_hpage_workfn); 1260c77c0a8aSWaiman Long 1261c77c0a8aSWaiman Long void free_huge_page(struct page *page) 1262c77c0a8aSWaiman Long { 1263c77c0a8aSWaiman Long /* 1264c77c0a8aSWaiman Long * Defer freeing if in non-task context to avoid hugetlb_lock deadlock. 1265c77c0a8aSWaiman Long */ 1266c77c0a8aSWaiman Long if (!in_task()) { 1267c77c0a8aSWaiman Long /* 1268c77c0a8aSWaiman Long * Only call schedule_work() if hpage_freelist is previously 1269c77c0a8aSWaiman Long * empty. Otherwise, schedule_work() had been called but the 1270c77c0a8aSWaiman Long * workfn hasn't retrieved the list yet. 1271c77c0a8aSWaiman Long */ 1272c77c0a8aSWaiman Long if (llist_add((struct llist_node *)&page->mapping, 1273c77c0a8aSWaiman Long &hpage_freelist)) 1274c77c0a8aSWaiman Long schedule_work(&free_hpage_work); 1275c77c0a8aSWaiman Long return; 1276c77c0a8aSWaiman Long } 1277c77c0a8aSWaiman Long 1278c77c0a8aSWaiman Long __free_huge_page(page); 1279c77c0a8aSWaiman Long } 1280c77c0a8aSWaiman Long 1281a5516438SAndi Kleen static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1282b7ba30c6SAndi Kleen { 12830edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1284f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1285b7ba30c6SAndi Kleen spin_lock(&hugetlb_lock); 12869dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 12871adc4d41SMina Almasry set_hugetlb_cgroup_rsvd(page, NULL); 1288a5516438SAndi Kleen h->nr_huge_pages++; 1289a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 1290b7ba30c6SAndi Kleen spin_unlock(&hugetlb_lock); 1291b7ba30c6SAndi Kleen } 1292b7ba30c6SAndi Kleen 1293d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order) 129420a0307cSWu Fengguang { 129520a0307cSWu Fengguang int i; 129620a0307cSWu Fengguang int nr_pages = 1 << order; 129720a0307cSWu Fengguang struct page *p = page + 1; 129820a0307cSWu Fengguang 129920a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 130020a0307cSWu Fengguang set_compound_order(page, order); 1301ef5a22beSAndrea Arcangeli __ClearPageReserved(page); 1302de09d31dSKirill A. Shutemov __SetPageHead(page); 130320a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1304ef5a22beSAndrea Arcangeli /* 1305ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1306ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1307ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 1308ef5a22beSAndrea Arcangeli * too. Otherwse drivers using get_user_pages() to access tail 1309ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1310ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1311ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1312ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1313ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1314ef5a22beSAndrea Arcangeli * after get_user_pages(). 1315ef5a22beSAndrea Arcangeli */ 1316ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 131758a84aa9SYouquan Song set_page_count(p, 0); 13181d798ca3SKirill A. Shutemov set_compound_head(p, page); 131920a0307cSWu Fengguang } 1320b4330afbSMike Kravetz atomic_set(compound_mapcount_ptr(page), -1); 132147e29d32SJohn Hubbard 132247e29d32SJohn Hubbard if (hpage_pincount_available(page)) 132347e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 132420a0307cSWu Fengguang } 132520a0307cSWu Fengguang 13267795912cSAndrew Morton /* 13277795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 13287795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 13297795912cSAndrew Morton * details. 13307795912cSAndrew Morton */ 133120a0307cSWu Fengguang int PageHuge(struct page *page) 133220a0307cSWu Fengguang { 133320a0307cSWu Fengguang if (!PageCompound(page)) 133420a0307cSWu Fengguang return 0; 133520a0307cSWu Fengguang 133620a0307cSWu Fengguang page = compound_head(page); 1337f1e61557SKirill A. Shutemov return page[1].compound_dtor == HUGETLB_PAGE_DTOR; 133820a0307cSWu Fengguang } 133943131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 134043131e14SNaoya Horiguchi 134127c73ae7SAndrea Arcangeli /* 134227c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 134327c73ae7SAndrea Arcangeli * normal or transparent huge pages. 134427c73ae7SAndrea Arcangeli */ 134527c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 134627c73ae7SAndrea Arcangeli { 134727c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 134827c73ae7SAndrea Arcangeli return 0; 134927c73ae7SAndrea Arcangeli 1350758f66a2SAndrew Morton return get_compound_page_dtor(page_head) == free_huge_page; 135127c73ae7SAndrea Arcangeli } 135227c73ae7SAndrea Arcangeli 1353c0d0381aSMike Kravetz /* 1354c0d0381aSMike Kravetz * Find address_space associated with hugetlbfs page. 1355c0d0381aSMike Kravetz * Upon entry page is locked and page 'was' mapped although mapped state 1356c0d0381aSMike Kravetz * could change. If necessary, use anon_vma to find vma and associated 1357c0d0381aSMike Kravetz * address space. The returned mapping may be stale, but it can not be 1358c0d0381aSMike Kravetz * invalid as page lock (which is held) is required to destroy mapping. 1359c0d0381aSMike Kravetz */ 1360c0d0381aSMike Kravetz static struct address_space *_get_hugetlb_page_mapping(struct page *hpage) 1361c0d0381aSMike Kravetz { 1362c0d0381aSMike Kravetz struct anon_vma *anon_vma; 1363c0d0381aSMike Kravetz pgoff_t pgoff_start, pgoff_end; 1364c0d0381aSMike Kravetz struct anon_vma_chain *avc; 1365c0d0381aSMike Kravetz struct address_space *mapping = page_mapping(hpage); 1366c0d0381aSMike Kravetz 1367c0d0381aSMike Kravetz /* Simple file based mapping */ 1368c0d0381aSMike Kravetz if (mapping) 1369c0d0381aSMike Kravetz return mapping; 1370c0d0381aSMike Kravetz 1371c0d0381aSMike Kravetz /* 1372c0d0381aSMike Kravetz * Even anonymous hugetlbfs mappings are associated with an 1373c0d0381aSMike Kravetz * underlying hugetlbfs file (see hugetlb_file_setup in mmap 1374c0d0381aSMike Kravetz * code). Find a vma associated with the anonymous vma, and 1375c0d0381aSMike Kravetz * use the file pointer to get address_space. 1376c0d0381aSMike Kravetz */ 1377c0d0381aSMike Kravetz anon_vma = page_lock_anon_vma_read(hpage); 1378c0d0381aSMike Kravetz if (!anon_vma) 1379c0d0381aSMike Kravetz return mapping; /* NULL */ 1380c0d0381aSMike Kravetz 1381c0d0381aSMike Kravetz /* Use first found vma */ 1382c0d0381aSMike Kravetz pgoff_start = page_to_pgoff(hpage); 1383c0d0381aSMike Kravetz pgoff_end = pgoff_start + hpage_nr_pages(hpage) - 1; 1384c0d0381aSMike Kravetz anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, 1385c0d0381aSMike Kravetz pgoff_start, pgoff_end) { 1386c0d0381aSMike Kravetz struct vm_area_struct *vma = avc->vma; 1387c0d0381aSMike Kravetz 1388c0d0381aSMike Kravetz mapping = vma->vm_file->f_mapping; 1389c0d0381aSMike Kravetz break; 1390c0d0381aSMike Kravetz } 1391c0d0381aSMike Kravetz 1392c0d0381aSMike Kravetz anon_vma_unlock_read(anon_vma); 1393c0d0381aSMike Kravetz return mapping; 1394c0d0381aSMike Kravetz } 1395c0d0381aSMike Kravetz 1396c0d0381aSMike Kravetz /* 1397c0d0381aSMike Kravetz * Find and lock address space (mapping) in write mode. 1398c0d0381aSMike Kravetz * 1399c0d0381aSMike Kravetz * Upon entry, the page is locked which allows us to find the mapping 1400c0d0381aSMike Kravetz * even in the case of an anon page. However, locking order dictates 1401c0d0381aSMike Kravetz * the i_mmap_rwsem be acquired BEFORE the page lock. This is hugetlbfs 1402c0d0381aSMike Kravetz * specific. So, we first try to lock the sema while still holding the 1403c0d0381aSMike Kravetz * page lock. If this works, great! If not, then we need to drop the 1404c0d0381aSMike Kravetz * page lock and then acquire i_mmap_rwsem and reacquire page lock. Of 1405c0d0381aSMike Kravetz * course, need to revalidate state along the way. 1406c0d0381aSMike Kravetz */ 1407c0d0381aSMike Kravetz struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) 1408c0d0381aSMike Kravetz { 1409c0d0381aSMike Kravetz struct address_space *mapping, *mapping2; 1410c0d0381aSMike Kravetz 1411c0d0381aSMike Kravetz mapping = _get_hugetlb_page_mapping(hpage); 1412c0d0381aSMike Kravetz retry: 1413c0d0381aSMike Kravetz if (!mapping) 1414c0d0381aSMike Kravetz return mapping; 1415c0d0381aSMike Kravetz 1416c0d0381aSMike Kravetz /* 1417c0d0381aSMike Kravetz * If no contention, take lock and return 1418c0d0381aSMike Kravetz */ 1419c0d0381aSMike Kravetz if (i_mmap_trylock_write(mapping)) 1420c0d0381aSMike Kravetz return mapping; 1421c0d0381aSMike Kravetz 1422c0d0381aSMike Kravetz /* 1423c0d0381aSMike Kravetz * Must drop page lock and wait on mapping sema. 1424c0d0381aSMike Kravetz * Note: Once page lock is dropped, mapping could become invalid. 1425c0d0381aSMike Kravetz * As a hack, increase map count until we lock page again. 1426c0d0381aSMike Kravetz */ 1427c0d0381aSMike Kravetz atomic_inc(&hpage->_mapcount); 1428c0d0381aSMike Kravetz unlock_page(hpage); 1429c0d0381aSMike Kravetz i_mmap_lock_write(mapping); 1430c0d0381aSMike Kravetz lock_page(hpage); 1431c0d0381aSMike Kravetz atomic_add_negative(-1, &hpage->_mapcount); 1432c0d0381aSMike Kravetz 1433c0d0381aSMike Kravetz /* verify page is still mapped */ 1434c0d0381aSMike Kravetz if (!page_mapped(hpage)) { 1435c0d0381aSMike Kravetz i_mmap_unlock_write(mapping); 1436c0d0381aSMike Kravetz return NULL; 1437c0d0381aSMike Kravetz } 1438c0d0381aSMike Kravetz 1439c0d0381aSMike Kravetz /* 1440c0d0381aSMike Kravetz * Get address space again and verify it is the same one 1441c0d0381aSMike Kravetz * we locked. If not, drop lock and retry. 1442c0d0381aSMike Kravetz */ 1443c0d0381aSMike Kravetz mapping2 = _get_hugetlb_page_mapping(hpage); 1444c0d0381aSMike Kravetz if (mapping2 != mapping) { 1445c0d0381aSMike Kravetz i_mmap_unlock_write(mapping); 1446c0d0381aSMike Kravetz mapping = mapping2; 1447c0d0381aSMike Kravetz goto retry; 1448c0d0381aSMike Kravetz } 1449c0d0381aSMike Kravetz 1450c0d0381aSMike Kravetz return mapping; 1451c0d0381aSMike Kravetz } 1452c0d0381aSMike Kravetz 145313d60f4bSZhang Yi pgoff_t __basepage_index(struct page *page) 145413d60f4bSZhang Yi { 145513d60f4bSZhang Yi struct page *page_head = compound_head(page); 145613d60f4bSZhang Yi pgoff_t index = page_index(page_head); 145713d60f4bSZhang Yi unsigned long compound_idx; 145813d60f4bSZhang Yi 145913d60f4bSZhang Yi if (!PageHuge(page_head)) 146013d60f4bSZhang Yi return page_index(page); 146113d60f4bSZhang Yi 146213d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 146313d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 146413d60f4bSZhang Yi else 146513d60f4bSZhang Yi compound_idx = page - page_head; 146613d60f4bSZhang Yi 146713d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 146813d60f4bSZhang Yi } 146913d60f4bSZhang Yi 14700c397daeSMichal Hocko static struct page *alloc_buddy_huge_page(struct hstate *h, 1471f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1472f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 14731da177e4SLinus Torvalds { 1474af0fb9dfSMichal Hocko int order = huge_page_order(h); 14751da177e4SLinus Torvalds struct page *page; 1476f60858f9SMike Kravetz bool alloc_try_hard = true; 1477f96efd58SJoe Jin 1478f60858f9SMike Kravetz /* 1479f60858f9SMike Kravetz * By default we always try hard to allocate the page with 1480f60858f9SMike Kravetz * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in 1481f60858f9SMike Kravetz * a loop (to adjust global huge page counts) and previous allocation 1482f60858f9SMike Kravetz * failed, do not continue to try hard on the same node. Use the 1483f60858f9SMike Kravetz * node_alloc_noretry bitmap to manage this state information. 1484f60858f9SMike Kravetz */ 1485f60858f9SMike Kravetz if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) 1486f60858f9SMike Kravetz alloc_try_hard = false; 1487f60858f9SMike Kravetz gfp_mask |= __GFP_COMP|__GFP_NOWARN; 1488f60858f9SMike Kravetz if (alloc_try_hard) 1489f60858f9SMike Kravetz gfp_mask |= __GFP_RETRY_MAYFAIL; 1490af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 1491af0fb9dfSMichal Hocko nid = numa_mem_id(); 1492af0fb9dfSMichal Hocko page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask); 1493af0fb9dfSMichal Hocko if (page) 1494af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC); 1495af0fb9dfSMichal Hocko else 1496af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 149763b4613cSNishanth Aravamudan 1498f60858f9SMike Kravetz /* 1499f60858f9SMike Kravetz * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this 1500f60858f9SMike Kravetz * indicates an overall state change. Clear bit so that we resume 1501f60858f9SMike Kravetz * normal 'try hard' allocations. 1502f60858f9SMike Kravetz */ 1503f60858f9SMike Kravetz if (node_alloc_noretry && page && !alloc_try_hard) 1504f60858f9SMike Kravetz node_clear(nid, *node_alloc_noretry); 1505f60858f9SMike Kravetz 1506f60858f9SMike Kravetz /* 1507f60858f9SMike Kravetz * If we tried hard to get a page but failed, set bit so that 1508f60858f9SMike Kravetz * subsequent attempts will not try as hard until there is an 1509f60858f9SMike Kravetz * overall state change. 1510f60858f9SMike Kravetz */ 1511f60858f9SMike Kravetz if (node_alloc_noretry && !page && alloc_try_hard) 1512f60858f9SMike Kravetz node_set(nid, *node_alloc_noretry); 1513f60858f9SMike Kravetz 151463b4613cSNishanth Aravamudan return page; 151563b4613cSNishanth Aravamudan } 151663b4613cSNishanth Aravamudan 1517af0fb9dfSMichal Hocko /* 15180c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 15190c397daeSMichal Hocko * should use this function to get new hugetlb pages 15200c397daeSMichal Hocko */ 15210c397daeSMichal Hocko static struct page *alloc_fresh_huge_page(struct hstate *h, 1522f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1523f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 15240c397daeSMichal Hocko { 15250c397daeSMichal Hocko struct page *page; 15260c397daeSMichal Hocko 15270c397daeSMichal Hocko if (hstate_is_gigantic(h)) 15280c397daeSMichal Hocko page = alloc_gigantic_page(h, gfp_mask, nid, nmask); 15290c397daeSMichal Hocko else 15300c397daeSMichal Hocko page = alloc_buddy_huge_page(h, gfp_mask, 1531f60858f9SMike Kravetz nid, nmask, node_alloc_noretry); 15320c397daeSMichal Hocko if (!page) 15330c397daeSMichal Hocko return NULL; 15340c397daeSMichal Hocko 15350c397daeSMichal Hocko if (hstate_is_gigantic(h)) 15360c397daeSMichal Hocko prep_compound_gigantic_page(page, huge_page_order(h)); 15370c397daeSMichal Hocko prep_new_huge_page(h, page, page_to_nid(page)); 15380c397daeSMichal Hocko 15390c397daeSMichal Hocko return page; 15400c397daeSMichal Hocko } 15410c397daeSMichal Hocko 15420c397daeSMichal Hocko /* 1543af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 1544af0fb9dfSMichal Hocko * manner. 1545af0fb9dfSMichal Hocko */ 1546f60858f9SMike Kravetz static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 1547f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 1548b2261026SJoonsoo Kim { 1549b2261026SJoonsoo Kim struct page *page; 1550b2261026SJoonsoo Kim int nr_nodes, node; 1551af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 1552b2261026SJoonsoo Kim 1553b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1554f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, 1555f60858f9SMike Kravetz node_alloc_noretry); 1556af0fb9dfSMichal Hocko if (page) 1557b2261026SJoonsoo Kim break; 1558b2261026SJoonsoo Kim } 1559b2261026SJoonsoo Kim 1560af0fb9dfSMichal Hocko if (!page) 1561af0fb9dfSMichal Hocko return 0; 1562b2261026SJoonsoo Kim 1563af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 1564af0fb9dfSMichal Hocko 1565af0fb9dfSMichal Hocko return 1; 1566b2261026SJoonsoo Kim } 1567b2261026SJoonsoo Kim 1568e8c5c824SLee Schermerhorn /* 1569e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 1570e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 1571e8c5c824SLee Schermerhorn * balanced over allowed nodes. 1572e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 1573e8c5c824SLee Schermerhorn */ 15746ae11b27SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 15756ae11b27SLee Schermerhorn bool acct_surplus) 1576e8c5c824SLee Schermerhorn { 1577b2261026SJoonsoo Kim int nr_nodes, node; 1578e8c5c824SLee Schermerhorn int ret = 0; 1579e8c5c824SLee Schermerhorn 1580b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1581685f3457SLee Schermerhorn /* 1582685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 1583685f3457SLee Schermerhorn * nodes with surplus pages. 1584685f3457SLee Schermerhorn */ 1585b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 1586b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 1587e8c5c824SLee Schermerhorn struct page *page = 1588b2261026SJoonsoo Kim list_entry(h->hugepage_freelists[node].next, 1589e8c5c824SLee Schermerhorn struct page, lru); 1590e8c5c824SLee Schermerhorn list_del(&page->lru); 1591e8c5c824SLee Schermerhorn h->free_huge_pages--; 1592b2261026SJoonsoo Kim h->free_huge_pages_node[node]--; 1593685f3457SLee Schermerhorn if (acct_surplus) { 1594685f3457SLee Schermerhorn h->surplus_huge_pages--; 1595b2261026SJoonsoo Kim h->surplus_huge_pages_node[node]--; 1596685f3457SLee Schermerhorn } 1597e8c5c824SLee Schermerhorn update_and_free_page(h, page); 1598e8c5c824SLee Schermerhorn ret = 1; 15999a76db09SLee Schermerhorn break; 1600e8c5c824SLee Schermerhorn } 1601b2261026SJoonsoo Kim } 1602e8c5c824SLee Schermerhorn 1603e8c5c824SLee Schermerhorn return ret; 1604e8c5c824SLee Schermerhorn } 1605e8c5c824SLee Schermerhorn 1606c8721bbbSNaoya Horiguchi /* 1607c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 1608faf53defSNaoya Horiguchi * nothing for in-use hugepages and non-hugepages. 1609faf53defSNaoya Horiguchi * This function returns values like below: 1610faf53defSNaoya Horiguchi * 1611faf53defSNaoya Horiguchi * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 1612faf53defSNaoya Horiguchi * (allocated or reserved.) 1613faf53defSNaoya Horiguchi * 0: successfully dissolved free hugepages or the page is not a 1614faf53defSNaoya Horiguchi * hugepage (considered as already dissolved) 1615c8721bbbSNaoya Horiguchi */ 1616c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 1617c8721bbbSNaoya Horiguchi { 16186bc9b564SNaoya Horiguchi int rc = -EBUSY; 1619082d5b6bSGerald Schaefer 1620faf53defSNaoya Horiguchi /* Not to disrupt normal path by vainly holding hugetlb_lock */ 1621faf53defSNaoya Horiguchi if (!PageHuge(page)) 1622faf53defSNaoya Horiguchi return 0; 1623faf53defSNaoya Horiguchi 1624c8721bbbSNaoya Horiguchi spin_lock(&hugetlb_lock); 1625faf53defSNaoya Horiguchi if (!PageHuge(page)) { 1626faf53defSNaoya Horiguchi rc = 0; 1627faf53defSNaoya Horiguchi goto out; 1628faf53defSNaoya Horiguchi } 1629faf53defSNaoya Horiguchi 1630faf53defSNaoya Horiguchi if (!page_count(page)) { 16312247bb33SGerald Schaefer struct page *head = compound_head(page); 16322247bb33SGerald Schaefer struct hstate *h = page_hstate(head); 16332247bb33SGerald Schaefer int nid = page_to_nid(head); 16346bc9b564SNaoya Horiguchi if (h->free_huge_pages - h->resv_huge_pages == 0) 1635082d5b6bSGerald Schaefer goto out; 1636c3114a84SAnshuman Khandual /* 1637c3114a84SAnshuman Khandual * Move PageHWPoison flag from head page to the raw error page, 1638c3114a84SAnshuman Khandual * which makes any subpages rather than the error page reusable. 1639c3114a84SAnshuman Khandual */ 1640c3114a84SAnshuman Khandual if (PageHWPoison(head) && page != head) { 1641c3114a84SAnshuman Khandual SetPageHWPoison(page); 1642c3114a84SAnshuman Khandual ClearPageHWPoison(head); 1643c3114a84SAnshuman Khandual } 16442247bb33SGerald Schaefer list_del(&head->lru); 1645c8721bbbSNaoya Horiguchi h->free_huge_pages--; 1646c8721bbbSNaoya Horiguchi h->free_huge_pages_node[nid]--; 1647c1470b33Szhong jiang h->max_huge_pages--; 16482247bb33SGerald Schaefer update_and_free_page(h, head); 16496bc9b564SNaoya Horiguchi rc = 0; 1650c8721bbbSNaoya Horiguchi } 1651082d5b6bSGerald Schaefer out: 1652c8721bbbSNaoya Horiguchi spin_unlock(&hugetlb_lock); 1653082d5b6bSGerald Schaefer return rc; 1654c8721bbbSNaoya Horiguchi } 1655c8721bbbSNaoya Horiguchi 1656c8721bbbSNaoya Horiguchi /* 1657c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 1658c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 16592247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 16602247bb33SGerald Schaefer * part of it lies within the given range. 1661082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 1662082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 1663c8721bbbSNaoya Horiguchi */ 1664082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 1665c8721bbbSNaoya Horiguchi { 1666c8721bbbSNaoya Horiguchi unsigned long pfn; 1667eb03aa00SGerald Schaefer struct page *page; 1668082d5b6bSGerald Schaefer int rc = 0; 1669c8721bbbSNaoya Horiguchi 1670d0177639SLi Zhong if (!hugepages_supported()) 1671082d5b6bSGerald Schaefer return rc; 1672d0177639SLi Zhong 1673eb03aa00SGerald Schaefer for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { 1674eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 1675eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 1676eb03aa00SGerald Schaefer if (rc) 1677082d5b6bSGerald Schaefer break; 1678eb03aa00SGerald Schaefer } 1679082d5b6bSGerald Schaefer 1680082d5b6bSGerald Schaefer return rc; 1681c8721bbbSNaoya Horiguchi } 1682c8721bbbSNaoya Horiguchi 1683ab5ac90aSMichal Hocko /* 1684ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 1685ab5ac90aSMichal Hocko */ 16860c397daeSMichal Hocko static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, 1687aaf14e40SMichal Hocko int nid, nodemask_t *nmask) 16887893d1d5SAdam Litke { 16899980d744SMichal Hocko struct page *page = NULL; 16907893d1d5SAdam Litke 1691bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1692aa888a74SAndi Kleen return NULL; 1693aa888a74SAndi Kleen 1694d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 16959980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 16969980d744SMichal Hocko goto out_unlock; 1697d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1698d1c3fb1fSNishanth Aravamudan 1699f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 17009980d744SMichal Hocko if (!page) 17010c397daeSMichal Hocko return NULL; 1702d1c3fb1fSNishanth Aravamudan 17037893d1d5SAdam Litke spin_lock(&hugetlb_lock); 17049980d744SMichal Hocko /* 17059980d744SMichal Hocko * We could have raced with the pool size change. 17069980d744SMichal Hocko * Double check that and simply deallocate the new page 17079980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 17089980d744SMichal Hocko * temporary page to workaround the nasty free_huge_page 17099980d744SMichal Hocko * codeflow 17109980d744SMichal Hocko */ 17119980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 17129980d744SMichal Hocko SetPageHugeTemporary(page); 17132bf753e6SKai Shen spin_unlock(&hugetlb_lock); 17149980d744SMichal Hocko put_page(page); 17152bf753e6SKai Shen return NULL; 17169980d744SMichal Hocko } else { 17179980d744SMichal Hocko h->surplus_huge_pages++; 17184704dea3SMichal Hocko h->surplus_huge_pages_node[page_to_nid(page)]++; 17197893d1d5SAdam Litke } 17209980d744SMichal Hocko 17219980d744SMichal Hocko out_unlock: 1722d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 17237893d1d5SAdam Litke 17247893d1d5SAdam Litke return page; 17257893d1d5SAdam Litke } 17267893d1d5SAdam Litke 17279a4e9f3bSAneesh Kumar K.V struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, 1728ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 1729ab5ac90aSMichal Hocko { 1730ab5ac90aSMichal Hocko struct page *page; 1731ab5ac90aSMichal Hocko 1732ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 1733ab5ac90aSMichal Hocko return NULL; 1734ab5ac90aSMichal Hocko 1735f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 1736ab5ac90aSMichal Hocko if (!page) 1737ab5ac90aSMichal Hocko return NULL; 1738ab5ac90aSMichal Hocko 1739ab5ac90aSMichal Hocko /* 1740ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 1741ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 1742ab5ac90aSMichal Hocko */ 1743ab5ac90aSMichal Hocko SetPageHugeTemporary(page); 1744ab5ac90aSMichal Hocko 1745ab5ac90aSMichal Hocko return page; 1746ab5ac90aSMichal Hocko } 1747ab5ac90aSMichal Hocko 1748e4e574b7SAdam Litke /* 1749099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 1750099730d6SDave Hansen */ 1751e0ec90eeSDave Hansen static 17520c397daeSMichal Hocko struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, 1753099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 1754099730d6SDave Hansen { 1755aaf14e40SMichal Hocko struct page *page; 1756aaf14e40SMichal Hocko struct mempolicy *mpol; 1757aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 1758aaf14e40SMichal Hocko int nid; 1759aaf14e40SMichal Hocko nodemask_t *nodemask; 1760aaf14e40SMichal Hocko 1761aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 17620c397daeSMichal Hocko page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); 1763aaf14e40SMichal Hocko mpol_cond_put(mpol); 1764aaf14e40SMichal Hocko 1765aaf14e40SMichal Hocko return page; 1766099730d6SDave Hansen } 1767099730d6SDave Hansen 1768ab5ac90aSMichal Hocko /* page migration callback function */ 1769bf50bab2SNaoya Horiguchi struct page *alloc_huge_page_node(struct hstate *h, int nid) 1770bf50bab2SNaoya Horiguchi { 1771aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 17724ef91848SJoonsoo Kim struct page *page = NULL; 1773bf50bab2SNaoya Horiguchi 1774aaf14e40SMichal Hocko if (nid != NUMA_NO_NODE) 1775aaf14e40SMichal Hocko gfp_mask |= __GFP_THISNODE; 1776aaf14e40SMichal Hocko 1777bf50bab2SNaoya Horiguchi spin_lock(&hugetlb_lock); 17784ef91848SJoonsoo Kim if (h->free_huge_pages - h->resv_huge_pages > 0) 17793e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL); 1780bf50bab2SNaoya Horiguchi spin_unlock(&hugetlb_lock); 1781bf50bab2SNaoya Horiguchi 178294ae8ba7SAneesh Kumar K.V if (!page) 17830c397daeSMichal Hocko page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL); 1784bf50bab2SNaoya Horiguchi 1785bf50bab2SNaoya Horiguchi return page; 1786bf50bab2SNaoya Horiguchi } 1787bf50bab2SNaoya Horiguchi 1788ab5ac90aSMichal Hocko /* page migration callback function */ 17893e59fcb0SMichal Hocko struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, 17903e59fcb0SMichal Hocko nodemask_t *nmask) 17914db9b2efSMichal Hocko { 1792aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 17934db9b2efSMichal Hocko 17944db9b2efSMichal Hocko spin_lock(&hugetlb_lock); 17954db9b2efSMichal Hocko if (h->free_huge_pages - h->resv_huge_pages > 0) { 17963e59fcb0SMichal Hocko struct page *page; 17973e59fcb0SMichal Hocko 17983e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); 17993e59fcb0SMichal Hocko if (page) { 18003e59fcb0SMichal Hocko spin_unlock(&hugetlb_lock); 18013e59fcb0SMichal Hocko return page; 18024db9b2efSMichal Hocko } 18034db9b2efSMichal Hocko } 18044db9b2efSMichal Hocko spin_unlock(&hugetlb_lock); 18054db9b2efSMichal Hocko 18060c397daeSMichal Hocko return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); 18074db9b2efSMichal Hocko } 18084db9b2efSMichal Hocko 1809ebd63723SMichal Hocko /* mempolicy aware migration callback */ 1810389c8178SMichal Hocko struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, 1811389c8178SMichal Hocko unsigned long address) 1812ebd63723SMichal Hocko { 1813ebd63723SMichal Hocko struct mempolicy *mpol; 1814ebd63723SMichal Hocko nodemask_t *nodemask; 1815ebd63723SMichal Hocko struct page *page; 1816ebd63723SMichal Hocko gfp_t gfp_mask; 1817ebd63723SMichal Hocko int node; 1818ebd63723SMichal Hocko 1819ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 1820ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 1821ebd63723SMichal Hocko page = alloc_huge_page_nodemask(h, node, nodemask); 1822ebd63723SMichal Hocko mpol_cond_put(mpol); 1823ebd63723SMichal Hocko 1824ebd63723SMichal Hocko return page; 1825ebd63723SMichal Hocko } 1826ebd63723SMichal Hocko 1827bf50bab2SNaoya Horiguchi /* 182825985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 1829e4e574b7SAdam Litke * of size 'delta'. 1830e4e574b7SAdam Litke */ 1831a5516438SAndi Kleen static int gather_surplus_pages(struct hstate *h, int delta) 1832e4e574b7SAdam Litke { 1833e4e574b7SAdam Litke struct list_head surplus_list; 1834e4e574b7SAdam Litke struct page *page, *tmp; 1835e4e574b7SAdam Litke int ret, i; 1836e4e574b7SAdam Litke int needed, allocated; 183728073b02SHillf Danton bool alloc_ok = true; 1838e4e574b7SAdam Litke 1839a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 1840ac09b3a1SAdam Litke if (needed <= 0) { 1841a5516438SAndi Kleen h->resv_huge_pages += delta; 1842e4e574b7SAdam Litke return 0; 1843ac09b3a1SAdam Litke } 1844e4e574b7SAdam Litke 1845e4e574b7SAdam Litke allocated = 0; 1846e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 1847e4e574b7SAdam Litke 1848e4e574b7SAdam Litke ret = -ENOMEM; 1849e4e574b7SAdam Litke retry: 1850e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 1851e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 18520c397daeSMichal Hocko page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), 1853aaf14e40SMichal Hocko NUMA_NO_NODE, NULL); 185428073b02SHillf Danton if (!page) { 185528073b02SHillf Danton alloc_ok = false; 185628073b02SHillf Danton break; 185728073b02SHillf Danton } 1858e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 185969ed779aSDavid Rientjes cond_resched(); 1860e4e574b7SAdam Litke } 186128073b02SHillf Danton allocated += i; 1862e4e574b7SAdam Litke 1863e4e574b7SAdam Litke /* 1864e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 1865e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 1866e4e574b7SAdam Litke */ 1867e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 1868a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 1869a5516438SAndi Kleen (h->free_huge_pages + allocated); 187028073b02SHillf Danton if (needed > 0) { 187128073b02SHillf Danton if (alloc_ok) 1872e4e574b7SAdam Litke goto retry; 187328073b02SHillf Danton /* 187428073b02SHillf Danton * We were not able to allocate enough pages to 187528073b02SHillf Danton * satisfy the entire reservation so we free what 187628073b02SHillf Danton * we've allocated so far. 187728073b02SHillf Danton */ 187828073b02SHillf Danton goto free; 187928073b02SHillf Danton } 1880e4e574b7SAdam Litke /* 1881e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 188225985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 1883e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 1884ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 1885ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 1886ac09b3a1SAdam Litke * before they are reserved. 1887e4e574b7SAdam Litke */ 1888e4e574b7SAdam Litke needed += allocated; 1889a5516438SAndi Kleen h->resv_huge_pages += delta; 1890e4e574b7SAdam Litke ret = 0; 1891a9869b83SNaoya Horiguchi 189219fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 189319fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 189419fc3f0aSAdam Litke if ((--needed) < 0) 189519fc3f0aSAdam Litke break; 1896a9869b83SNaoya Horiguchi /* 1897a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 1898a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 1899a9869b83SNaoya Horiguchi */ 1900a9869b83SNaoya Horiguchi put_page_testzero(page); 1901309381feSSasha Levin VM_BUG_ON_PAGE(page_count(page), page); 1902a5516438SAndi Kleen enqueue_huge_page(h, page); 190319fc3f0aSAdam Litke } 190428073b02SHillf Danton free: 1905b0365c8dSHillf Danton spin_unlock(&hugetlb_lock); 190619fc3f0aSAdam Litke 190719fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 1908c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 1909a9869b83SNaoya Horiguchi put_page(page); 191019fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 1911e4e574b7SAdam Litke 1912e4e574b7SAdam Litke return ret; 1913e4e574b7SAdam Litke } 1914e4e574b7SAdam Litke 1915e4e574b7SAdam Litke /* 1916e5bbc8a6SMike Kravetz * This routine has two main purposes: 1917e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 1918e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 1919e5bbc8a6SMike Kravetz * to the associated reservation map. 1920e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 1921e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 1922e5bbc8a6SMike Kravetz * 1923e5bbc8a6SMike Kravetz * Called with hugetlb_lock held. However, the lock could be dropped (and 1924e5bbc8a6SMike Kravetz * reacquired) during calls to cond_resched_lock. Whenever dropping the lock, 1925e5bbc8a6SMike Kravetz * we must make sure nobody else can claim pages we are in the process of 1926e5bbc8a6SMike Kravetz * freeing. Do this by ensuring resv_huge_page always is greater than the 1927e5bbc8a6SMike Kravetz * number of huge pages we plan to free when dropping the lock. 1928e4e574b7SAdam Litke */ 1929a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 1930a5516438SAndi Kleen unsigned long unused_resv_pages) 1931e4e574b7SAdam Litke { 1932e4e574b7SAdam Litke unsigned long nr_pages; 1933e4e574b7SAdam Litke 1934aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 1935bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1936e5bbc8a6SMike Kravetz goto out; 1937aa888a74SAndi Kleen 1938e5bbc8a6SMike Kravetz /* 1939e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 1940e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 1941e5bbc8a6SMike Kravetz */ 1942a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 1943e4e574b7SAdam Litke 1944685f3457SLee Schermerhorn /* 1945685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 19469b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 19479b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 19489b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 19499b5e5d0fSLee Schermerhorn * free_pool_huge_page() will balance the the freed pages across the 19509b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 1951e5bbc8a6SMike Kravetz * 1952e5bbc8a6SMike Kravetz * Note that we decrement resv_huge_pages as we free the pages. If 1953e5bbc8a6SMike Kravetz * we drop the lock, resv_huge_pages will still be sufficiently large 1954e5bbc8a6SMike Kravetz * to cover subsequent pages we may free. 1955685f3457SLee Schermerhorn */ 1956685f3457SLee Schermerhorn while (nr_pages--) { 1957e5bbc8a6SMike Kravetz h->resv_huge_pages--; 1958e5bbc8a6SMike Kravetz unused_resv_pages--; 19598cebfcd0SLai Jiangshan if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) 1960e5bbc8a6SMike Kravetz goto out; 19617848a4bfSMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 1962e4e574b7SAdam Litke } 1963e5bbc8a6SMike Kravetz 1964e5bbc8a6SMike Kravetz out: 1965e5bbc8a6SMike Kravetz /* Fully uncommit the reservation */ 1966e5bbc8a6SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 1967e4e574b7SAdam Litke } 1968e4e574b7SAdam Litke 19695e911373SMike Kravetz 1970c37f9fb1SAndy Whitcroft /* 1971feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 19725e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 1973cf3ad20bSMike Kravetz * 1974cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 1975cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 1976cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 1977cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 1978cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 1979feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 1980feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 1981feba16e2SMike Kravetz * is called in such cases. 1982cf3ad20bSMike Kravetz * 1983cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 1984cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 1985cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 1986cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 1987cf3ad20bSMike Kravetz * take appropriate action. 198896b96a96SMike Kravetz * 198996b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 199096b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 199196b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 199296b96a96SMike Kravetz * reservation exists. 1993c37f9fb1SAndy Whitcroft */ 19945e911373SMike Kravetz enum vma_resv_mode { 19955e911373SMike Kravetz VMA_NEEDS_RESV, 19965e911373SMike Kravetz VMA_COMMIT_RESV, 1997feba16e2SMike Kravetz VMA_END_RESV, 199896b96a96SMike Kravetz VMA_ADD_RESV, 19995e911373SMike Kravetz }; 2000cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 2001cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 20025e911373SMike Kravetz enum vma_resv_mode mode) 2003c37f9fb1SAndy Whitcroft { 20044e35f483SJoonsoo Kim struct resv_map *resv; 20054e35f483SJoonsoo Kim pgoff_t idx; 2006cf3ad20bSMike Kravetz long ret; 2007c37f9fb1SAndy Whitcroft 20084e35f483SJoonsoo Kim resv = vma_resv_map(vma); 20094e35f483SJoonsoo Kim if (!resv) 2010c37f9fb1SAndy Whitcroft return 1; 2011c37f9fb1SAndy Whitcroft 20124e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 20135e911373SMike Kravetz switch (mode) { 20145e911373SMike Kravetz case VMA_NEEDS_RESV: 2015cf3ad20bSMike Kravetz ret = region_chg(resv, idx, idx + 1); 20165e911373SMike Kravetz break; 20175e911373SMike Kravetz case VMA_COMMIT_RESV: 20185e911373SMike Kravetz ret = region_add(resv, idx, idx + 1); 20195e911373SMike Kravetz break; 2020feba16e2SMike Kravetz case VMA_END_RESV: 20215e911373SMike Kravetz region_abort(resv, idx, idx + 1); 20225e911373SMike Kravetz ret = 0; 20235e911373SMike Kravetz break; 202496b96a96SMike Kravetz case VMA_ADD_RESV: 202596b96a96SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) 202696b96a96SMike Kravetz ret = region_add(resv, idx, idx + 1); 202796b96a96SMike Kravetz else { 202896b96a96SMike Kravetz region_abort(resv, idx, idx + 1); 202996b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 203096b96a96SMike Kravetz } 203196b96a96SMike Kravetz break; 20325e911373SMike Kravetz default: 20335e911373SMike Kravetz BUG(); 20345e911373SMike Kravetz } 203584afd99bSAndy Whitcroft 20364e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) 2037cf3ad20bSMike Kravetz return ret; 203867961f9dSMike Kravetz else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { 203967961f9dSMike Kravetz /* 204067961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 204167961f9dSMike Kravetz * However, a file associated with mapping could have been 204267961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 204367961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 204467961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 204567961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 204667961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 204767961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 204867961f9dSMike Kravetz * reservation has already been consumed. As a result, the 204967961f9dSMike Kravetz * return value of this routine is the opposite of the 205067961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 205167961f9dSMike Kravetz */ 205267961f9dSMike Kravetz if (ret) 205367961f9dSMike Kravetz return 0; 205467961f9dSMike Kravetz else 205567961f9dSMike Kravetz return 1; 205667961f9dSMike Kravetz } 20574e35f483SJoonsoo Kim else 2058cf3ad20bSMike Kravetz return ret < 0 ? ret : 0; 205984afd99bSAndy Whitcroft } 2060cf3ad20bSMike Kravetz 2061cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 2062a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 2063c37f9fb1SAndy Whitcroft { 20645e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 2065cf3ad20bSMike Kravetz } 2066c37f9fb1SAndy Whitcroft 2067cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 2068cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 2069cf3ad20bSMike Kravetz { 20705e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 20715e911373SMike Kravetz } 20725e911373SMike Kravetz 2073feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 20745e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 20755e911373SMike Kravetz { 2076feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 2077c37f9fb1SAndy Whitcroft } 2078c37f9fb1SAndy Whitcroft 207996b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 208096b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 208196b96a96SMike Kravetz { 208296b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 208396b96a96SMike Kravetz } 208496b96a96SMike Kravetz 208596b96a96SMike Kravetz /* 208696b96a96SMike Kravetz * This routine is called to restore a reservation on error paths. In the 208796b96a96SMike Kravetz * specific error paths, a huge page was allocated (via alloc_huge_page) 208896b96a96SMike Kravetz * and is about to be freed. If a reservation for the page existed, 208996b96a96SMike Kravetz * alloc_huge_page would have consumed the reservation and set PagePrivate 209096b96a96SMike Kravetz * in the newly allocated page. When the page is freed via free_huge_page, 209196b96a96SMike Kravetz * the global reservation count will be incremented if PagePrivate is set. 209296b96a96SMike Kravetz * However, free_huge_page can not adjust the reserve map. Adjust the 209396b96a96SMike Kravetz * reserve map here to be consistent with global reserve count adjustments 209496b96a96SMike Kravetz * to be made by free_huge_page. 209596b96a96SMike Kravetz */ 209696b96a96SMike Kravetz static void restore_reserve_on_error(struct hstate *h, 209796b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long address, 209896b96a96SMike Kravetz struct page *page) 209996b96a96SMike Kravetz { 210096b96a96SMike Kravetz if (unlikely(PagePrivate(page))) { 210196b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 210296b96a96SMike Kravetz 210396b96a96SMike Kravetz if (unlikely(rc < 0)) { 210496b96a96SMike Kravetz /* 210596b96a96SMike Kravetz * Rare out of memory condition in reserve map 210696b96a96SMike Kravetz * manipulation. Clear PagePrivate so that 210796b96a96SMike Kravetz * global reserve count will not be incremented 210896b96a96SMike Kravetz * by free_huge_page. This will make it appear 210996b96a96SMike Kravetz * as though the reservation for this page was 211096b96a96SMike Kravetz * consumed. This may prevent the task from 211196b96a96SMike Kravetz * faulting in the page at a later time. This 211296b96a96SMike Kravetz * is better than inconsistent global huge page 211396b96a96SMike Kravetz * accounting of reserve counts. 211496b96a96SMike Kravetz */ 211596b96a96SMike Kravetz ClearPagePrivate(page); 211696b96a96SMike Kravetz } else if (rc) { 211796b96a96SMike Kravetz rc = vma_add_reservation(h, vma, address); 211896b96a96SMike Kravetz if (unlikely(rc < 0)) 211996b96a96SMike Kravetz /* 212096b96a96SMike Kravetz * See above comment about rare out of 212196b96a96SMike Kravetz * memory condition. 212296b96a96SMike Kravetz */ 212396b96a96SMike Kravetz ClearPagePrivate(page); 212496b96a96SMike Kravetz } else 212596b96a96SMike Kravetz vma_end_reservation(h, vma, address); 212696b96a96SMike Kravetz } 212796b96a96SMike Kravetz } 212896b96a96SMike Kravetz 212970c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 213004f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 2131348ea204SAdam Litke { 213290481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 2133a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2134348ea204SAdam Litke struct page *page; 2135d85f69b0SMike Kravetz long map_chg, map_commit; 2136d85f69b0SMike Kravetz long gbl_chg; 21376d76dcf4SAneesh Kumar K.V int ret, idx; 21386d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 21392fc39cecSAdam Litke 21406d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 2141a1e78772SMel Gorman /* 2142d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 2143d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 2144d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 2145a1e78772SMel Gorman */ 2146d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 2147d85f69b0SMike Kravetz if (map_chg < 0) 214876dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 2149d85f69b0SMike Kravetz 2150d85f69b0SMike Kravetz /* 2151d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 2152d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 2153d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 2154d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 2155d85f69b0SMike Kravetz * checked against any subpool limit. 2156d85f69b0SMike Kravetz */ 2157d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 2158d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 2159d85f69b0SMike Kravetz if (gbl_chg < 0) { 2160feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 216176dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 21625e911373SMike Kravetz } 216390d8b7e6SAdam Litke 2164d85f69b0SMike Kravetz /* 2165d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 2166d85f69b0SMike Kravetz * map, there could be reservations associated with the 2167d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 2168d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 2169d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 2170d85f69b0SMike Kravetz * the subpool reservations. 2171d85f69b0SMike Kravetz */ 2172d85f69b0SMike Kravetz if (avoid_reserve) 2173d85f69b0SMike Kravetz gbl_chg = 1; 2174d85f69b0SMike Kravetz } 2175d85f69b0SMike Kravetz 21766d76dcf4SAneesh Kumar K.V ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 21778f34af6fSJianyu Zhan if (ret) 21788f34af6fSJianyu Zhan goto out_subpool_put; 21798f34af6fSJianyu Zhan 2180a1e78772SMel Gorman spin_lock(&hugetlb_lock); 2181d85f69b0SMike Kravetz /* 2182d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 2183d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 2184d85f69b0SMike Kravetz * a reservation exists for the allocation. 2185d85f69b0SMike Kravetz */ 2186d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 218781a6fcaeSJoonsoo Kim if (!page) { 218894ae8ba7SAneesh Kumar K.V spin_unlock(&hugetlb_lock); 21890c397daeSMichal Hocko page = alloc_buddy_huge_page_with_mpol(h, vma, addr); 21908f34af6fSJianyu Zhan if (!page) 21918f34af6fSJianyu Zhan goto out_uncharge_cgroup; 2192a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 2193a88c7695SNaoya Horiguchi SetPagePrivate(page); 2194a88c7695SNaoya Horiguchi h->resv_huge_pages--; 2195a88c7695SNaoya Horiguchi } 219679dbb236SAneesh Kumar K.V spin_lock(&hugetlb_lock); 219779dbb236SAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 219881a6fcaeSJoonsoo Kim /* Fall through */ 2199a1e78772SMel Gorman } 220081a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 220181a6fcaeSJoonsoo Kim spin_unlock(&hugetlb_lock); 2202a1e78772SMel Gorman 220390481622SDavid Gibson set_page_private(page, (unsigned long)spool); 2204a1e78772SMel Gorman 2205d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 2206d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 220733039678SMike Kravetz /* 220833039678SMike Kravetz * The page was added to the reservation map between 220933039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 221033039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 221133039678SMike Kravetz * Adjust for the subpool count incremented above AND 221233039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 221333039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 221433039678SMike Kravetz * no longer applies. 221533039678SMike Kravetz */ 221633039678SMike Kravetz long rsv_adjust; 221733039678SMike Kravetz 221833039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 221933039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 222033039678SMike Kravetz } 22217893d1d5SAdam Litke return page; 22228f34af6fSJianyu Zhan 22238f34af6fSJianyu Zhan out_uncharge_cgroup: 22248f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 22258f34af6fSJianyu Zhan out_subpool_put: 2226d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 22278f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 2228feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 22298f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 2230b45b5bd6SDavid Gibson } 2231b45b5bd6SDavid Gibson 2232e24a1307SAneesh Kumar K.V int alloc_bootmem_huge_page(struct hstate *h) 2233e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 2234e24a1307SAneesh Kumar K.V int __alloc_bootmem_huge_page(struct hstate *h) 2235aa888a74SAndi Kleen { 2236aa888a74SAndi Kleen struct huge_bootmem_page *m; 2237b2261026SJoonsoo Kim int nr_nodes, node; 2238aa888a74SAndi Kleen 2239b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 2240aa888a74SAndi Kleen void *addr; 2241aa888a74SAndi Kleen 2242eb31d559SMike Rapoport addr = memblock_alloc_try_nid_raw( 22438b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 224497ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 2245aa888a74SAndi Kleen if (addr) { 2246aa888a74SAndi Kleen /* 2247aa888a74SAndi Kleen * Use the beginning of the huge page to store the 2248aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 2249aa888a74SAndi Kleen * puts them into the mem_map). 2250aa888a74SAndi Kleen */ 2251aa888a74SAndi Kleen m = addr; 2252aa888a74SAndi Kleen goto found; 2253aa888a74SAndi Kleen } 2254aa888a74SAndi Kleen } 2255aa888a74SAndi Kleen return 0; 2256aa888a74SAndi Kleen 2257aa888a74SAndi Kleen found: 2258df994eadSLuiz Capitulino BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); 2259aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 2260330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 2261aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 2262aa888a74SAndi Kleen m->hstate = h; 2263aa888a74SAndi Kleen return 1; 2264aa888a74SAndi Kleen } 2265aa888a74SAndi Kleen 2266d00181b9SKirill A. Shutemov static void __init prep_compound_huge_page(struct page *page, 2267d00181b9SKirill A. Shutemov unsigned int order) 226818229df5SAndy Whitcroft { 226918229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 227018229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 227118229df5SAndy Whitcroft else 227218229df5SAndy Whitcroft prep_compound_page(page, order); 227318229df5SAndy Whitcroft } 227418229df5SAndy Whitcroft 2275aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 2276aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 2277aa888a74SAndi Kleen { 2278aa888a74SAndi Kleen struct huge_bootmem_page *m; 2279aa888a74SAndi Kleen 2280aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 228140d18ebfSMike Kravetz struct page *page = virt_to_page(m); 2282aa888a74SAndi Kleen struct hstate *h = m->hstate; 2283ee8f248dSBecky Bruce 2284aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 228518229df5SAndy Whitcroft prep_compound_huge_page(page, h->order); 2286ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 2287aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 2288af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 2289af0fb9dfSMichal Hocko 2290b0320c7bSRafael Aquini /* 2291b0320c7bSRafael Aquini * If we had gigantic hugepages allocated at boot time, we need 2292b0320c7bSRafael Aquini * to restore the 'stolen' pages to totalram_pages in order to 2293b0320c7bSRafael Aquini * fix confusing memory reports from free(1) and another 2294b0320c7bSRafael Aquini * side-effects, like CommitLimit going negative. 2295b0320c7bSRafael Aquini */ 2296bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 22973dcc0571SJiang Liu adjust_managed_page_count(page, 1 << h->order); 2298520495feSCannon Matthews cond_resched(); 2299aa888a74SAndi Kleen } 2300aa888a74SAndi Kleen } 2301aa888a74SAndi Kleen 23028faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 23031da177e4SLinus Torvalds { 23041da177e4SLinus Torvalds unsigned long i; 2305f60858f9SMike Kravetz nodemask_t *node_alloc_noretry; 2306f60858f9SMike Kravetz 2307f60858f9SMike Kravetz if (!hstate_is_gigantic(h)) { 2308f60858f9SMike Kravetz /* 2309f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 2310f60858f9SMike Kravetz * Ignore errors as lower level routines can deal with 2311f60858f9SMike Kravetz * node_alloc_noretry == NULL. If this kmalloc fails at boot 2312f60858f9SMike Kravetz * time, we are likely in bigger trouble. 2313f60858f9SMike Kravetz */ 2314f60858f9SMike Kravetz node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), 2315f60858f9SMike Kravetz GFP_KERNEL); 2316f60858f9SMike Kravetz } else { 2317f60858f9SMike Kravetz /* allocations done at boot time */ 2318f60858f9SMike Kravetz node_alloc_noretry = NULL; 2319f60858f9SMike Kravetz } 2320f60858f9SMike Kravetz 2321f60858f9SMike Kravetz /* bit mask controlling how hard we retry per-node allocations */ 2322f60858f9SMike Kravetz if (node_alloc_noretry) 2323f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 23241da177e4SLinus Torvalds 2325e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 2326bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 2327aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 2328aa888a74SAndi Kleen break; 23290c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 2330f60858f9SMike Kravetz &node_states[N_MEMORY], 2331f60858f9SMike Kravetz node_alloc_noretry)) 23321da177e4SLinus Torvalds break; 233369ed779aSDavid Rientjes cond_resched(); 23341da177e4SLinus Torvalds } 2335d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 2336d715cf80SLiam R. Howlett char buf[32]; 2337d715cf80SLiam R. Howlett 2338c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2339d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 2340d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 23418faa8b07SAndi Kleen h->max_huge_pages = i; 2342e5ff2159SAndi Kleen } 2343f60858f9SMike Kravetz 2344f60858f9SMike Kravetz kfree(node_alloc_noretry); 2345d715cf80SLiam R. Howlett } 2346e5ff2159SAndi Kleen 2347e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 2348e5ff2159SAndi Kleen { 2349e5ff2159SAndi Kleen struct hstate *h; 2350e5ff2159SAndi Kleen 2351e5ff2159SAndi Kleen for_each_hstate(h) { 2352641844f5SNaoya Horiguchi if (minimum_order > huge_page_order(h)) 2353641844f5SNaoya Horiguchi minimum_order = huge_page_order(h); 2354641844f5SNaoya Horiguchi 23558faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 2356bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 23578faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 2358e5ff2159SAndi Kleen } 2359641844f5SNaoya Horiguchi VM_BUG_ON(minimum_order == UINT_MAX); 2360e5ff2159SAndi Kleen } 2361e5ff2159SAndi Kleen 2362e5ff2159SAndi Kleen static void __init report_hugepages(void) 2363e5ff2159SAndi Kleen { 2364e5ff2159SAndi Kleen struct hstate *h; 2365e5ff2159SAndi Kleen 2366e5ff2159SAndi Kleen for_each_hstate(h) { 23674abd32dbSAndi Kleen char buf[32]; 2368c6247f72SMatthew Wilcox 2369c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2370ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 2371c6247f72SMatthew Wilcox buf, h->free_huge_pages); 2372e5ff2159SAndi Kleen } 2373e5ff2159SAndi Kleen } 2374e5ff2159SAndi Kleen 23751da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 23766ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 23776ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 23781da177e4SLinus Torvalds { 23794415cc8dSChristoph Lameter int i; 23804415cc8dSChristoph Lameter 2381bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2382aa888a74SAndi Kleen return; 2383aa888a74SAndi Kleen 23846ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 23851da177e4SLinus Torvalds struct page *page, *next; 2386a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 2387a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 2388a5516438SAndi Kleen if (count >= h->nr_huge_pages) 23896b0c880dSAdam Litke return; 23901da177e4SLinus Torvalds if (PageHighMem(page)) 23911da177e4SLinus Torvalds continue; 23921da177e4SLinus Torvalds list_del(&page->lru); 2393e5ff2159SAndi Kleen update_and_free_page(h, page); 2394a5516438SAndi Kleen h->free_huge_pages--; 2395a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 23961da177e4SLinus Torvalds } 23971da177e4SLinus Torvalds } 23981da177e4SLinus Torvalds } 23991da177e4SLinus Torvalds #else 24006ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 24016ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 24021da177e4SLinus Torvalds { 24031da177e4SLinus Torvalds } 24041da177e4SLinus Torvalds #endif 24051da177e4SLinus Torvalds 240620a0307cSWu Fengguang /* 240720a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 240820a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 240920a0307cSWu Fengguang * Returns 1 if an adjustment was made. 241020a0307cSWu Fengguang */ 24116ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 24126ae11b27SLee Schermerhorn int delta) 241320a0307cSWu Fengguang { 2414b2261026SJoonsoo Kim int nr_nodes, node; 241520a0307cSWu Fengguang 241620a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 241720a0307cSWu Fengguang 2418e8c5c824SLee Schermerhorn if (delta < 0) { 2419b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 2420b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 2421b2261026SJoonsoo Kim goto found; 2422b2261026SJoonsoo Kim } 2423b2261026SJoonsoo Kim } else { 2424b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2425b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 2426b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 2427b2261026SJoonsoo Kim goto found; 2428e8c5c824SLee Schermerhorn } 24299a76db09SLee Schermerhorn } 2430b2261026SJoonsoo Kim return 0; 243120a0307cSWu Fengguang 2432b2261026SJoonsoo Kim found: 243320a0307cSWu Fengguang h->surplus_huge_pages += delta; 2434b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 2435b2261026SJoonsoo Kim return 1; 243620a0307cSWu Fengguang } 243720a0307cSWu Fengguang 2438a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 2439fd875dcaSMike Kravetz static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 24406ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 24411da177e4SLinus Torvalds { 24427893d1d5SAdam Litke unsigned long min_count, ret; 2443f60858f9SMike Kravetz NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); 2444f60858f9SMike Kravetz 2445f60858f9SMike Kravetz /* 2446f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 2447f60858f9SMike Kravetz * If we can not allocate the bit mask, do not attempt to allocate 2448f60858f9SMike Kravetz * the requested huge pages. 2449f60858f9SMike Kravetz */ 2450f60858f9SMike Kravetz if (node_alloc_noretry) 2451f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 2452f60858f9SMike Kravetz else 2453f60858f9SMike Kravetz return -ENOMEM; 24541da177e4SLinus Torvalds 24554eb0716eSAlexandre Ghiti spin_lock(&hugetlb_lock); 24564eb0716eSAlexandre Ghiti 24574eb0716eSAlexandre Ghiti /* 2458fd875dcaSMike Kravetz * Check for a node specific request. 2459fd875dcaSMike Kravetz * Changing node specific huge page count may require a corresponding 2460fd875dcaSMike Kravetz * change to the global count. In any case, the passed node mask 2461fd875dcaSMike Kravetz * (nodes_allowed) will restrict alloc/free to the specified node. 2462fd875dcaSMike Kravetz */ 2463fd875dcaSMike Kravetz if (nid != NUMA_NO_NODE) { 2464fd875dcaSMike Kravetz unsigned long old_count = count; 2465fd875dcaSMike Kravetz 2466fd875dcaSMike Kravetz count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 2467fd875dcaSMike Kravetz /* 2468fd875dcaSMike Kravetz * User may have specified a large count value which caused the 2469fd875dcaSMike Kravetz * above calculation to overflow. In this case, they wanted 2470fd875dcaSMike Kravetz * to allocate as many huge pages as possible. Set count to 2471fd875dcaSMike Kravetz * largest possible value to align with their intention. 2472fd875dcaSMike Kravetz */ 2473fd875dcaSMike Kravetz if (count < old_count) 2474fd875dcaSMike Kravetz count = ULONG_MAX; 2475fd875dcaSMike Kravetz } 2476fd875dcaSMike Kravetz 2477fd875dcaSMike Kravetz /* 24784eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 24794eb0716eSAlexandre Ghiti * page range allocation. 24804eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 24814eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 24824eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 24834eb0716eSAlexandre Ghiti */ 24844eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 24854eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 24864eb0716eSAlexandre Ghiti spin_unlock(&hugetlb_lock); 2487f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 24884eb0716eSAlexandre Ghiti return -EINVAL; 24894eb0716eSAlexandre Ghiti } 24904eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 24914eb0716eSAlexandre Ghiti } 2492aa888a74SAndi Kleen 24937893d1d5SAdam Litke /* 24947893d1d5SAdam Litke * Increase the pool size 24957893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 24967893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 2497d1c3fb1fSNishanth Aravamudan * 24980c397daeSMichal Hocko * We might race with alloc_surplus_huge_page() here and be unable 2499d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 2500d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 2501d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 2502d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 25037893d1d5SAdam Litke */ 2504a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 25056ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 25067893d1d5SAdam Litke break; 25077893d1d5SAdam Litke } 25087893d1d5SAdam Litke 2509a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 25107893d1d5SAdam Litke /* 25117893d1d5SAdam Litke * If this allocation races such that we no longer need the 25127893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 25137893d1d5SAdam Litke * and reducing the surplus. 25147893d1d5SAdam Litke */ 25157893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 2516649920c6SJia He 2517649920c6SJia He /* yield cpu to avoid soft lockup */ 2518649920c6SJia He cond_resched(); 2519649920c6SJia He 2520f60858f9SMike Kravetz ret = alloc_pool_huge_page(h, nodes_allowed, 2521f60858f9SMike Kravetz node_alloc_noretry); 25227893d1d5SAdam Litke spin_lock(&hugetlb_lock); 25237893d1d5SAdam Litke if (!ret) 25247893d1d5SAdam Litke goto out; 25257893d1d5SAdam Litke 2526536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 2527536240f2SMel Gorman if (signal_pending(current)) 2528536240f2SMel Gorman goto out; 25297893d1d5SAdam Litke } 25307893d1d5SAdam Litke 25317893d1d5SAdam Litke /* 25327893d1d5SAdam Litke * Decrease the pool size 25337893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 25347893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 25357893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 25367893d1d5SAdam Litke * to the desired size as pages become free. 2537d1c3fb1fSNishanth Aravamudan * 2538d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 2539d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 2540d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 25410c397daeSMichal Hocko * alloc_surplus_huge_page() is checking the global counter, 2542d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 2543d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 2544d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 25457893d1d5SAdam Litke */ 2546a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 25476b0c880dSAdam Litke min_count = max(count, min_count); 25486ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 2549a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 25506ae11b27SLee Schermerhorn if (!free_pool_huge_page(h, nodes_allowed, 0)) 25511da177e4SLinus Torvalds break; 255255f67141SMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 25531da177e4SLinus Torvalds } 2554a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 25556ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 25567893d1d5SAdam Litke break; 25577893d1d5SAdam Litke } 25587893d1d5SAdam Litke out: 25594eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 25601da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 25614eb0716eSAlexandre Ghiti 2562f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 2563f60858f9SMike Kravetz 25644eb0716eSAlexandre Ghiti return 0; 25651da177e4SLinus Torvalds } 25661da177e4SLinus Torvalds 2567a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 2568a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 2569a3437870SNishanth Aravamudan 2570a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 2571a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 2572a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 2573a3437870SNishanth Aravamudan 2574a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 2575a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 2576a3437870SNishanth Aravamudan 25779a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 25789a305230SLee Schermerhorn 25799a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 2580a3437870SNishanth Aravamudan { 2581a3437870SNishanth Aravamudan int i; 25829a305230SLee Schermerhorn 2583a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 25849a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 25859a305230SLee Schermerhorn if (nidp) 25869a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 2587a3437870SNishanth Aravamudan return &hstates[i]; 25889a305230SLee Schermerhorn } 25899a305230SLee Schermerhorn 25909a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 2591a3437870SNishanth Aravamudan } 2592a3437870SNishanth Aravamudan 259306808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 2594a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2595a3437870SNishanth Aravamudan { 25969a305230SLee Schermerhorn struct hstate *h; 25979a305230SLee Schermerhorn unsigned long nr_huge_pages; 25989a305230SLee Schermerhorn int nid; 25999a305230SLee Schermerhorn 26009a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 26019a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 26029a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 26039a305230SLee Schermerhorn else 26049a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 26059a305230SLee Schermerhorn 26069a305230SLee Schermerhorn return sprintf(buf, "%lu\n", nr_huge_pages); 2607a3437870SNishanth Aravamudan } 2608adbe8726SEric B Munson 2609238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 2610238d3c13SDavid Rientjes struct hstate *h, int nid, 2611238d3c13SDavid Rientjes unsigned long count, size_t len) 2612a3437870SNishanth Aravamudan { 2613a3437870SNishanth Aravamudan int err; 26142d0adf7eSOscar Salvador nodemask_t nodes_allowed, *n_mask; 2615a3437870SNishanth Aravamudan 26162d0adf7eSOscar Salvador if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 26172d0adf7eSOscar Salvador return -EINVAL; 2618adbe8726SEric B Munson 26199a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 26209a305230SLee Schermerhorn /* 26219a305230SLee Schermerhorn * global hstate attribute 26229a305230SLee Schermerhorn */ 26239a305230SLee Schermerhorn if (!(obey_mempolicy && 26242d0adf7eSOscar Salvador init_nodemask_of_mempolicy(&nodes_allowed))) 26252d0adf7eSOscar Salvador n_mask = &node_states[N_MEMORY]; 26262d0adf7eSOscar Salvador else 26272d0adf7eSOscar Salvador n_mask = &nodes_allowed; 26282d0adf7eSOscar Salvador } else { 26299a305230SLee Schermerhorn /* 2630fd875dcaSMike Kravetz * Node specific request. count adjustment happens in 2631fd875dcaSMike Kravetz * set_max_huge_pages() after acquiring hugetlb_lock. 26329a305230SLee Schermerhorn */ 26332d0adf7eSOscar Salvador init_nodemask_of_node(&nodes_allowed, nid); 26342d0adf7eSOscar Salvador n_mask = &nodes_allowed; 2635fd875dcaSMike Kravetz } 26369a305230SLee Schermerhorn 26372d0adf7eSOscar Salvador err = set_max_huge_pages(h, count, nid, n_mask); 263806808b08SLee Schermerhorn 26394eb0716eSAlexandre Ghiti return err ? err : len; 264006808b08SLee Schermerhorn } 264106808b08SLee Schermerhorn 2642238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 2643238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 2644238d3c13SDavid Rientjes size_t len) 2645238d3c13SDavid Rientjes { 2646238d3c13SDavid Rientjes struct hstate *h; 2647238d3c13SDavid Rientjes unsigned long count; 2648238d3c13SDavid Rientjes int nid; 2649238d3c13SDavid Rientjes int err; 2650238d3c13SDavid Rientjes 2651238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 2652238d3c13SDavid Rientjes if (err) 2653238d3c13SDavid Rientjes return err; 2654238d3c13SDavid Rientjes 2655238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 2656238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 2657238d3c13SDavid Rientjes } 2658238d3c13SDavid Rientjes 265906808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 266006808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 266106808b08SLee Schermerhorn { 266206808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 266306808b08SLee Schermerhorn } 266406808b08SLee Schermerhorn 266506808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 266606808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 266706808b08SLee Schermerhorn { 2668238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 2669a3437870SNishanth Aravamudan } 2670a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 2671a3437870SNishanth Aravamudan 267206808b08SLee Schermerhorn #ifdef CONFIG_NUMA 267306808b08SLee Schermerhorn 267406808b08SLee Schermerhorn /* 267506808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 267606808b08SLee Schermerhorn * huge page alloc/free. 267706808b08SLee Schermerhorn */ 267806808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 267906808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 268006808b08SLee Schermerhorn { 268106808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 268206808b08SLee Schermerhorn } 268306808b08SLee Schermerhorn 268406808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 268506808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 268606808b08SLee Schermerhorn { 2687238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 268806808b08SLee Schermerhorn } 268906808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 269006808b08SLee Schermerhorn #endif 269106808b08SLee Schermerhorn 269206808b08SLee Schermerhorn 2693a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 2694a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2695a3437870SNishanth Aravamudan { 26969a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2697a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 2698a3437870SNishanth Aravamudan } 2699adbe8726SEric B Munson 2700a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 2701a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 2702a3437870SNishanth Aravamudan { 2703a3437870SNishanth Aravamudan int err; 2704a3437870SNishanth Aravamudan unsigned long input; 27059a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2706a3437870SNishanth Aravamudan 2707bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2708adbe8726SEric B Munson return -EINVAL; 2709adbe8726SEric B Munson 27103dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 2711a3437870SNishanth Aravamudan if (err) 271273ae31e5SEric B Munson return err; 2713a3437870SNishanth Aravamudan 2714a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 2715a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 2716a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 2717a3437870SNishanth Aravamudan 2718a3437870SNishanth Aravamudan return count; 2719a3437870SNishanth Aravamudan } 2720a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 2721a3437870SNishanth Aravamudan 2722a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 2723a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2724a3437870SNishanth Aravamudan { 27259a305230SLee Schermerhorn struct hstate *h; 27269a305230SLee Schermerhorn unsigned long free_huge_pages; 27279a305230SLee Schermerhorn int nid; 27289a305230SLee Schermerhorn 27299a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 27309a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 27319a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 27329a305230SLee Schermerhorn else 27339a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 27349a305230SLee Schermerhorn 27359a305230SLee Schermerhorn return sprintf(buf, "%lu\n", free_huge_pages); 2736a3437870SNishanth Aravamudan } 2737a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 2738a3437870SNishanth Aravamudan 2739a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 2740a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2741a3437870SNishanth Aravamudan { 27429a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2743a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->resv_huge_pages); 2744a3437870SNishanth Aravamudan } 2745a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 2746a3437870SNishanth Aravamudan 2747a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 2748a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2749a3437870SNishanth Aravamudan { 27509a305230SLee Schermerhorn struct hstate *h; 27519a305230SLee Schermerhorn unsigned long surplus_huge_pages; 27529a305230SLee Schermerhorn int nid; 27539a305230SLee Schermerhorn 27549a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 27559a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 27569a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 27579a305230SLee Schermerhorn else 27589a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 27599a305230SLee Schermerhorn 27609a305230SLee Schermerhorn return sprintf(buf, "%lu\n", surplus_huge_pages); 2761a3437870SNishanth Aravamudan } 2762a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 2763a3437870SNishanth Aravamudan 2764a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 2765a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 2766a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 2767a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 2768a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 2769a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 277006808b08SLee Schermerhorn #ifdef CONFIG_NUMA 277106808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 277206808b08SLee Schermerhorn #endif 2773a3437870SNishanth Aravamudan NULL, 2774a3437870SNishanth Aravamudan }; 2775a3437870SNishanth Aravamudan 277667e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 2777a3437870SNishanth Aravamudan .attrs = hstate_attrs, 2778a3437870SNishanth Aravamudan }; 2779a3437870SNishanth Aravamudan 2780094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 27819a305230SLee Schermerhorn struct kobject **hstate_kobjs, 278267e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 2783a3437870SNishanth Aravamudan { 2784a3437870SNishanth Aravamudan int retval; 2785972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 2786a3437870SNishanth Aravamudan 27879a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 27889a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 2789a3437870SNishanth Aravamudan return -ENOMEM; 2790a3437870SNishanth Aravamudan 27919a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 2792a3437870SNishanth Aravamudan if (retval) 27939a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 2794a3437870SNishanth Aravamudan 2795a3437870SNishanth Aravamudan return retval; 2796a3437870SNishanth Aravamudan } 2797a3437870SNishanth Aravamudan 2798a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 2799a3437870SNishanth Aravamudan { 2800a3437870SNishanth Aravamudan struct hstate *h; 2801a3437870SNishanth Aravamudan int err; 2802a3437870SNishanth Aravamudan 2803a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 2804a3437870SNishanth Aravamudan if (!hugepages_kobj) 2805a3437870SNishanth Aravamudan return; 2806a3437870SNishanth Aravamudan 2807a3437870SNishanth Aravamudan for_each_hstate(h) { 28089a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 28099a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 2810a3437870SNishanth Aravamudan if (err) 2811ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s", h->name); 2812a3437870SNishanth Aravamudan } 2813a3437870SNishanth Aravamudan } 2814a3437870SNishanth Aravamudan 28159a305230SLee Schermerhorn #ifdef CONFIG_NUMA 28169a305230SLee Schermerhorn 28179a305230SLee Schermerhorn /* 28189a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 281910fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 282010fbcf4cSKay Sievers * index of a node device or _hstate == node id. 282110fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 28229a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 28239a305230SLee Schermerhorn */ 28249a305230SLee Schermerhorn struct node_hstate { 28259a305230SLee Schermerhorn struct kobject *hugepages_kobj; 28269a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 28279a305230SLee Schermerhorn }; 2828b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 28299a305230SLee Schermerhorn 28309a305230SLee Schermerhorn /* 283110fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 28329a305230SLee Schermerhorn */ 28339a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 28349a305230SLee Schermerhorn &nr_hugepages_attr.attr, 28359a305230SLee Schermerhorn &free_hugepages_attr.attr, 28369a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 28379a305230SLee Schermerhorn NULL, 28389a305230SLee Schermerhorn }; 28399a305230SLee Schermerhorn 284067e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 28419a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 28429a305230SLee Schermerhorn }; 28439a305230SLee Schermerhorn 28449a305230SLee Schermerhorn /* 284510fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 28469a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 28479a305230SLee Schermerhorn */ 28489a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 28499a305230SLee Schermerhorn { 28509a305230SLee Schermerhorn int nid; 28519a305230SLee Schermerhorn 28529a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 28539a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 28549a305230SLee Schermerhorn int i; 28559a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 28569a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 28579a305230SLee Schermerhorn if (nidp) 28589a305230SLee Schermerhorn *nidp = nid; 28599a305230SLee Schermerhorn return &hstates[i]; 28609a305230SLee Schermerhorn } 28619a305230SLee Schermerhorn } 28629a305230SLee Schermerhorn 28639a305230SLee Schermerhorn BUG(); 28649a305230SLee Schermerhorn return NULL; 28659a305230SLee Schermerhorn } 28669a305230SLee Schermerhorn 28679a305230SLee Schermerhorn /* 286810fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 28699a305230SLee Schermerhorn * No-op if no hstate attributes attached. 28709a305230SLee Schermerhorn */ 28713cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 28729a305230SLee Schermerhorn { 28739a305230SLee Schermerhorn struct hstate *h; 287410fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 28759a305230SLee Schermerhorn 28769a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 28779b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 28789a305230SLee Schermerhorn 2879972dc4deSAneesh Kumar K.V for_each_hstate(h) { 2880972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 2881972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 2882972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 2883972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 2884972dc4deSAneesh Kumar K.V } 28859a305230SLee Schermerhorn } 28869a305230SLee Schermerhorn 28879a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 28889a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 28899a305230SLee Schermerhorn } 28909a305230SLee Schermerhorn 28919a305230SLee Schermerhorn 28929a305230SLee Schermerhorn /* 289310fbcf4cSKay Sievers * Register hstate attributes for a single node device. 28949a305230SLee Schermerhorn * No-op if attributes already registered. 28959a305230SLee Schermerhorn */ 28963cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 28979a305230SLee Schermerhorn { 28989a305230SLee Schermerhorn struct hstate *h; 289910fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 29009a305230SLee Schermerhorn int err; 29019a305230SLee Schermerhorn 29029a305230SLee Schermerhorn if (nhs->hugepages_kobj) 29039a305230SLee Schermerhorn return; /* already allocated */ 29049a305230SLee Schermerhorn 29059a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 290610fbcf4cSKay Sievers &node->dev.kobj); 29079a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 29089a305230SLee Schermerhorn return; 29099a305230SLee Schermerhorn 29109a305230SLee Schermerhorn for_each_hstate(h) { 29119a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 29129a305230SLee Schermerhorn nhs->hstate_kobjs, 29139a305230SLee Schermerhorn &per_node_hstate_attr_group); 29149a305230SLee Schermerhorn if (err) { 2915ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s for node %d\n", 291610fbcf4cSKay Sievers h->name, node->dev.id); 29179a305230SLee Schermerhorn hugetlb_unregister_node(node); 29189a305230SLee Schermerhorn break; 29199a305230SLee Schermerhorn } 29209a305230SLee Schermerhorn } 29219a305230SLee Schermerhorn } 29229a305230SLee Schermerhorn 29239a305230SLee Schermerhorn /* 29249b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 292510fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 292610fbcf4cSKay Sievers * registered their associated device by this time. 29279a305230SLee Schermerhorn */ 29287d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 29299a305230SLee Schermerhorn { 29309a305230SLee Schermerhorn int nid; 29319a305230SLee Schermerhorn 29328cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 29338732794bSWen Congyang struct node *node = node_devices[nid]; 293410fbcf4cSKay Sievers if (node->dev.id == nid) 29359a305230SLee Schermerhorn hugetlb_register_node(node); 29369a305230SLee Schermerhorn } 29379a305230SLee Schermerhorn 29389a305230SLee Schermerhorn /* 293910fbcf4cSKay Sievers * Let the node device driver know we're here so it can 29409a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 29419a305230SLee Schermerhorn */ 29429a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 29439a305230SLee Schermerhorn hugetlb_unregister_node); 29449a305230SLee Schermerhorn } 29459a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 29469a305230SLee Schermerhorn 29479a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 29489a305230SLee Schermerhorn { 29499a305230SLee Schermerhorn BUG(); 29509a305230SLee Schermerhorn if (nidp) 29519a305230SLee Schermerhorn *nidp = -1; 29529a305230SLee Schermerhorn return NULL; 29539a305230SLee Schermerhorn } 29549a305230SLee Schermerhorn 29559a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 29569a305230SLee Schermerhorn 29579a305230SLee Schermerhorn #endif 29589a305230SLee Schermerhorn 2959a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 2960a3437870SNishanth Aravamudan { 29618382d914SDavidlohr Bueso int i; 29628382d914SDavidlohr Bueso 2963457c1b27SNishanth Aravamudan if (!hugepages_supported()) 29640ef89d25SBenjamin Herrenschmidt return 0; 2965a3437870SNishanth Aravamudan 2966e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) { 2967d715cf80SLiam R. Howlett if (default_hstate_size != 0) { 2968d715cf80SLiam R. Howlett pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n", 2969d715cf80SLiam R. Howlett default_hstate_size, HPAGE_SIZE); 2970d715cf80SLiam R. Howlett } 2971d715cf80SLiam R. Howlett 2972e11bfbfcSNick Piggin default_hstate_size = HPAGE_SIZE; 2973e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) 2974a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 2975a3437870SNishanth Aravamudan } 2976972dc4deSAneesh Kumar K.V default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); 2977f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 2978f8b74815SVaishali Thakkar if (!default_hstate.max_huge_pages) 2979e11bfbfcSNick Piggin default_hstate.max_huge_pages = default_hstate_max_huge_pages; 2980f8b74815SVaishali Thakkar } 2981a3437870SNishanth Aravamudan 2982a3437870SNishanth Aravamudan hugetlb_init_hstates(); 2983aa888a74SAndi Kleen gather_bootmem_prealloc(); 2984a3437870SNishanth Aravamudan report_hugepages(); 2985a3437870SNishanth Aravamudan 2986a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 29879a305230SLee Schermerhorn hugetlb_register_all_nodes(); 29887179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 29899a305230SLee Schermerhorn 29908382d914SDavidlohr Bueso #ifdef CONFIG_SMP 29918382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 29928382d914SDavidlohr Bueso #else 29938382d914SDavidlohr Bueso num_fault_mutexes = 1; 29948382d914SDavidlohr Bueso #endif 2995c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 29966da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 29976da2ec56SKees Cook GFP_KERNEL); 2998c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 29998382d914SDavidlohr Bueso 30008382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 3001c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 3002a3437870SNishanth Aravamudan return 0; 3003a3437870SNishanth Aravamudan } 30043e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 3005a3437870SNishanth Aravamudan 3006a3437870SNishanth Aravamudan /* Should be called on processing a hugepagesz=... option */ 30079fee021dSVaishali Thakkar void __init hugetlb_bad_size(void) 30089fee021dSVaishali Thakkar { 30099fee021dSVaishali Thakkar parsed_valid_hugepagesz = false; 30109fee021dSVaishali Thakkar } 30119fee021dSVaishali Thakkar 3012d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 3013a3437870SNishanth Aravamudan { 3014a3437870SNishanth Aravamudan struct hstate *h; 30158faa8b07SAndi Kleen unsigned long i; 30168faa8b07SAndi Kleen 3017a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 3018598d8091SJoe Perches pr_warn("hugepagesz= specified twice, ignoring\n"); 3019a3437870SNishanth Aravamudan return; 3020a3437870SNishanth Aravamudan } 302147d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 3022a3437870SNishanth Aravamudan BUG_ON(order == 0); 302347d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 3024a3437870SNishanth Aravamudan h->order = order; 3025a3437870SNishanth Aravamudan h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 30268faa8b07SAndi Kleen h->nr_huge_pages = 0; 30278faa8b07SAndi Kleen h->free_huge_pages = 0; 30288faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 30298faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 30300edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 303154f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 303254f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 3033a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 3034a3437870SNishanth Aravamudan huge_page_size(h)/1024); 30358faa8b07SAndi Kleen 3036a3437870SNishanth Aravamudan parsed_hstate = h; 3037a3437870SNishanth Aravamudan } 3038a3437870SNishanth Aravamudan 3039e11bfbfcSNick Piggin static int __init hugetlb_nrpages_setup(char *s) 3040a3437870SNishanth Aravamudan { 3041a3437870SNishanth Aravamudan unsigned long *mhp; 30428faa8b07SAndi Kleen static unsigned long *last_mhp; 3043a3437870SNishanth Aravamudan 30449fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 30459fee021dSVaishali Thakkar pr_warn("hugepages = %s preceded by " 30469fee021dSVaishali Thakkar "an unsupported hugepagesz, ignoring\n", s); 30479fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 30489fee021dSVaishali Thakkar return 1; 30499fee021dSVaishali Thakkar } 3050a3437870SNishanth Aravamudan /* 305147d38344SAneesh Kumar K.V * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, 3052a3437870SNishanth Aravamudan * so this hugepages= parameter goes to the "default hstate". 3053a3437870SNishanth Aravamudan */ 30549fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 3055a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 3056a3437870SNishanth Aravamudan else 3057a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 3058a3437870SNishanth Aravamudan 30598faa8b07SAndi Kleen if (mhp == last_mhp) { 3060598d8091SJoe Perches pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n"); 30618faa8b07SAndi Kleen return 1; 30628faa8b07SAndi Kleen } 30638faa8b07SAndi Kleen 3064a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 3065a3437870SNishanth Aravamudan *mhp = 0; 3066a3437870SNishanth Aravamudan 30678faa8b07SAndi Kleen /* 30688faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 30698faa8b07SAndi Kleen * But we need to allocate >= MAX_ORDER hstates here early to still 30708faa8b07SAndi Kleen * use the bootmem allocator. 30718faa8b07SAndi Kleen */ 307247d38344SAneesh Kumar K.V if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) 30738faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 30748faa8b07SAndi Kleen 30758faa8b07SAndi Kleen last_mhp = mhp; 30768faa8b07SAndi Kleen 3077a3437870SNishanth Aravamudan return 1; 3078a3437870SNishanth Aravamudan } 3079e11bfbfcSNick Piggin __setup("hugepages=", hugetlb_nrpages_setup); 3080e11bfbfcSNick Piggin 3081e11bfbfcSNick Piggin static int __init hugetlb_default_setup(char *s) 3082e11bfbfcSNick Piggin { 3083e11bfbfcSNick Piggin default_hstate_size = memparse(s, &s); 3084e11bfbfcSNick Piggin return 1; 3085e11bfbfcSNick Piggin } 3086e11bfbfcSNick Piggin __setup("default_hugepagesz=", hugetlb_default_setup); 3087a3437870SNishanth Aravamudan 30888a213460SNishanth Aravamudan static unsigned int cpuset_mems_nr(unsigned int *array) 30898a213460SNishanth Aravamudan { 30908a213460SNishanth Aravamudan int node; 30918a213460SNishanth Aravamudan unsigned int nr = 0; 30928a213460SNishanth Aravamudan 30938a213460SNishanth Aravamudan for_each_node_mask(node, cpuset_current_mems_allowed) 30948a213460SNishanth Aravamudan nr += array[node]; 30958a213460SNishanth Aravamudan 30968a213460SNishanth Aravamudan return nr; 30978a213460SNishanth Aravamudan } 30988a213460SNishanth Aravamudan 30998a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 310006808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 310106808b08SLee Schermerhorn struct ctl_table *table, int write, 310206808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 31031da177e4SLinus Torvalds { 3104e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 3105238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 310608d4a246SMichal Hocko int ret; 3107e5ff2159SAndi Kleen 3108457c1b27SNishanth Aravamudan if (!hugepages_supported()) 310986613628SJan Stancek return -EOPNOTSUPP; 3110457c1b27SNishanth Aravamudan 3111e5ff2159SAndi Kleen table->data = &tmp; 3112e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 311308d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 311408d4a246SMichal Hocko if (ret) 311508d4a246SMichal Hocko goto out; 3116e5ff2159SAndi Kleen 3117238d3c13SDavid Rientjes if (write) 3118238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 3119238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 312008d4a246SMichal Hocko out: 312108d4a246SMichal Hocko return ret; 31221da177e4SLinus Torvalds } 3123396faf03SMel Gorman 312406808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 312506808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 312606808b08SLee Schermerhorn { 312706808b08SLee Schermerhorn 312806808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 312906808b08SLee Schermerhorn buffer, length, ppos); 313006808b08SLee Schermerhorn } 313106808b08SLee Schermerhorn 313206808b08SLee Schermerhorn #ifdef CONFIG_NUMA 313306808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 313406808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 313506808b08SLee Schermerhorn { 313606808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 313706808b08SLee Schermerhorn buffer, length, ppos); 313806808b08SLee Schermerhorn } 313906808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 314006808b08SLee Schermerhorn 3141a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 31428d65af78SAlexey Dobriyan void __user *buffer, 3143a3d0c6aaSNishanth Aravamudan size_t *length, loff_t *ppos) 3144a3d0c6aaSNishanth Aravamudan { 3145a5516438SAndi Kleen struct hstate *h = &default_hstate; 3146e5ff2159SAndi Kleen unsigned long tmp; 314708d4a246SMichal Hocko int ret; 3148e5ff2159SAndi Kleen 3149457c1b27SNishanth Aravamudan if (!hugepages_supported()) 315086613628SJan Stancek return -EOPNOTSUPP; 3151457c1b27SNishanth Aravamudan 3152e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 3153e5ff2159SAndi Kleen 3154bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 3155adbe8726SEric B Munson return -EINVAL; 3156adbe8726SEric B Munson 3157e5ff2159SAndi Kleen table->data = &tmp; 3158e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 315908d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 316008d4a246SMichal Hocko if (ret) 316108d4a246SMichal Hocko goto out; 3162e5ff2159SAndi Kleen 3163e5ff2159SAndi Kleen if (write) { 3164064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 3165e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 3166a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 3167e5ff2159SAndi Kleen } 316808d4a246SMichal Hocko out: 316908d4a246SMichal Hocko return ret; 3170a3d0c6aaSNishanth Aravamudan } 3171a3d0c6aaSNishanth Aravamudan 31721da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 31731da177e4SLinus Torvalds 3174e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 31751da177e4SLinus Torvalds { 3176fcb2b0c5SRoman Gushchin struct hstate *h; 3177fcb2b0c5SRoman Gushchin unsigned long total = 0; 3178fcb2b0c5SRoman Gushchin 3179457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3180457c1b27SNishanth Aravamudan return; 3181fcb2b0c5SRoman Gushchin 3182fcb2b0c5SRoman Gushchin for_each_hstate(h) { 3183fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 3184fcb2b0c5SRoman Gushchin 3185fcb2b0c5SRoman Gushchin total += (PAGE_SIZE << huge_page_order(h)) * count; 3186fcb2b0c5SRoman Gushchin 3187fcb2b0c5SRoman Gushchin if (h == &default_hstate) 3188e1759c21SAlexey Dobriyan seq_printf(m, 31891da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 31901da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 3191b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 31927893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 31934f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 3194fcb2b0c5SRoman Gushchin count, 3195a5516438SAndi Kleen h->free_huge_pages, 3196a5516438SAndi Kleen h->resv_huge_pages, 3197a5516438SAndi Kleen h->surplus_huge_pages, 3198fcb2b0c5SRoman Gushchin (PAGE_SIZE << huge_page_order(h)) / 1024); 3199fcb2b0c5SRoman Gushchin } 3200fcb2b0c5SRoman Gushchin 3201fcb2b0c5SRoman Gushchin seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024); 32021da177e4SLinus Torvalds } 32031da177e4SLinus Torvalds 32041da177e4SLinus Torvalds int hugetlb_report_node_meminfo(int nid, char *buf) 32051da177e4SLinus Torvalds { 3206a5516438SAndi Kleen struct hstate *h = &default_hstate; 3207457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3208457c1b27SNishanth Aravamudan return 0; 32091da177e4SLinus Torvalds return sprintf(buf, 32101da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 3211a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 3212a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 3213a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 3214a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 3215a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 32161da177e4SLinus Torvalds } 32171da177e4SLinus Torvalds 3218949f7ec5SDavid Rientjes void hugetlb_show_meminfo(void) 3219949f7ec5SDavid Rientjes { 3220949f7ec5SDavid Rientjes struct hstate *h; 3221949f7ec5SDavid Rientjes int nid; 3222949f7ec5SDavid Rientjes 3223457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3224457c1b27SNishanth Aravamudan return; 3225457c1b27SNishanth Aravamudan 3226949f7ec5SDavid Rientjes for_each_node_state(nid, N_MEMORY) 3227949f7ec5SDavid Rientjes for_each_hstate(h) 3228949f7ec5SDavid Rientjes pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 3229949f7ec5SDavid Rientjes nid, 3230949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 3231949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 3232949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 3233949f7ec5SDavid Rientjes 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 3234949f7ec5SDavid Rientjes } 3235949f7ec5SDavid Rientjes 32365d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 32375d317b2bSNaoya Horiguchi { 32385d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 32395d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 32405d317b2bSNaoya Horiguchi } 32415d317b2bSNaoya Horiguchi 32421da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 32431da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 32441da177e4SLinus Torvalds { 3245d0028588SWanpeng Li struct hstate *h; 3246d0028588SWanpeng Li unsigned long nr_total_pages = 0; 3247d0028588SWanpeng Li 3248d0028588SWanpeng Li for_each_hstate(h) 3249d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 3250d0028588SWanpeng Li return nr_total_pages; 32511da177e4SLinus Torvalds } 32521da177e4SLinus Torvalds 3253a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 3254fc1b8a73SMel Gorman { 3255fc1b8a73SMel Gorman int ret = -ENOMEM; 3256fc1b8a73SMel Gorman 3257fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 3258fc1b8a73SMel Gorman /* 3259fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 3260fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 3261fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 3262fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 3263fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 3264fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 3265fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 3266fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 3267fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 3268fc1b8a73SMel Gorman * 3269fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 3270fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 3271fc1b8a73SMel Gorman * we fall back to check against current free page availability as 3272fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 3273fc1b8a73SMel Gorman * semantics that cpuset has. 3274fc1b8a73SMel Gorman */ 3275fc1b8a73SMel Gorman if (delta > 0) { 3276a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 3277fc1b8a73SMel Gorman goto out; 3278fc1b8a73SMel Gorman 3279a5516438SAndi Kleen if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 3280a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 3281fc1b8a73SMel Gorman goto out; 3282fc1b8a73SMel Gorman } 3283fc1b8a73SMel Gorman } 3284fc1b8a73SMel Gorman 3285fc1b8a73SMel Gorman ret = 0; 3286fc1b8a73SMel Gorman if (delta < 0) 3287a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 3288fc1b8a73SMel Gorman 3289fc1b8a73SMel Gorman out: 3290fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 3291fc1b8a73SMel Gorman return ret; 3292fc1b8a73SMel Gorman } 3293fc1b8a73SMel Gorman 329484afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 329584afd99bSAndy Whitcroft { 3296f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 329784afd99bSAndy Whitcroft 329884afd99bSAndy Whitcroft /* 329984afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 330084afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 330184afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 330225985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 330384afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 330484afd99bSAndy Whitcroft * new reference here without additional locking. 330584afd99bSAndy Whitcroft */ 33064e35f483SJoonsoo Kim if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 3307f522c3acSJoonsoo Kim kref_get(&resv->refs); 330884afd99bSAndy Whitcroft } 330984afd99bSAndy Whitcroft 3310a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 3311a1e78772SMel Gorman { 3312a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3313f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 331490481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 33154e35f483SJoonsoo Kim unsigned long reserve, start, end; 33161c5ecae3SMike Kravetz long gbl_reserve; 331784afd99bSAndy Whitcroft 33184e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 33194e35f483SJoonsoo Kim return; 33204e35f483SJoonsoo Kim 3321a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 3322a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 332384afd99bSAndy Whitcroft 33244e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 3325e9fe92aeSMina Almasry hugetlb_cgroup_uncharge_counter(resv, start, end); 33267251ff78SAdam Litke if (reserve) { 33271c5ecae3SMike Kravetz /* 33281c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 33291c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 33301c5ecae3SMike Kravetz */ 33311c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 33321c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 33337251ff78SAdam Litke } 3334e9fe92aeSMina Almasry 3335e9fe92aeSMina Almasry kref_put(&resv->refs, resv_map_release); 3336a1e78772SMel Gorman } 3337a1e78772SMel Gorman 333831383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 333931383c68SDan Williams { 334031383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 334131383c68SDan Williams return -EINVAL; 334231383c68SDan Williams return 0; 334331383c68SDan Williams } 334431383c68SDan Williams 334505ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 334605ea8860SDan Williams { 334705ea8860SDan Williams struct hstate *hstate = hstate_vma(vma); 334805ea8860SDan Williams 334905ea8860SDan Williams return 1UL << huge_page_shift(hstate); 335005ea8860SDan Williams } 335105ea8860SDan Williams 33521da177e4SLinus Torvalds /* 33531da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 33541da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 33551da177e4SLinus Torvalds * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 33561da177e4SLinus Torvalds * this far. 33571da177e4SLinus Torvalds */ 3358b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 33591da177e4SLinus Torvalds { 33601da177e4SLinus Torvalds BUG(); 3361d0217ac0SNick Piggin return 0; 33621da177e4SLinus Torvalds } 33631da177e4SLinus Torvalds 3364eec3636aSJane Chu /* 3365eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 3366eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 3367eec3636aSJane Chu * This is because under System V memory model, mappings created via 3368eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 3369eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 3370eec3636aSJane Chu */ 3371f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 3372d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 337384afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 3374a1e78772SMel Gorman .close = hugetlb_vm_op_close, 337531383c68SDan Williams .split = hugetlb_vm_op_split, 337605ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 33771da177e4SLinus Torvalds }; 33781da177e4SLinus Torvalds 33791e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 33801e8f889bSDavid Gibson int writable) 338163551ae0SDavid Gibson { 338263551ae0SDavid Gibson pte_t entry; 338363551ae0SDavid Gibson 33841e8f889bSDavid Gibson if (writable) { 3385106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 3386106c992aSGerald Schaefer vma->vm_page_prot))); 338763551ae0SDavid Gibson } else { 3388106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 3389106c992aSGerald Schaefer vma->vm_page_prot)); 339063551ae0SDavid Gibson } 339163551ae0SDavid Gibson entry = pte_mkyoung(entry); 339263551ae0SDavid Gibson entry = pte_mkhuge(entry); 3393d9ed9faaSChris Metcalf entry = arch_make_huge_pte(entry, vma, page, writable); 339463551ae0SDavid Gibson 339563551ae0SDavid Gibson return entry; 339663551ae0SDavid Gibson } 339763551ae0SDavid Gibson 33981e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 33991e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 34001e8f889bSDavid Gibson { 34011e8f889bSDavid Gibson pte_t entry; 34021e8f889bSDavid Gibson 3403106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 340432f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 34054b3073e1SRussell King update_mmu_cache(vma, address, ptep); 34061e8f889bSDavid Gibson } 34071e8f889bSDavid Gibson 3408d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 34094a705fefSNaoya Horiguchi { 34104a705fefSNaoya Horiguchi swp_entry_t swp; 34114a705fefSNaoya Horiguchi 34124a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 3413d5ed7444SAneesh Kumar K.V return false; 34144a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 34154a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_migration_entry(swp)) 3416d5ed7444SAneesh Kumar K.V return true; 34174a705fefSNaoya Horiguchi else 3418d5ed7444SAneesh Kumar K.V return false; 34194a705fefSNaoya Horiguchi } 34204a705fefSNaoya Horiguchi 34214a705fefSNaoya Horiguchi static int is_hugetlb_entry_hwpoisoned(pte_t pte) 34224a705fefSNaoya Horiguchi { 34234a705fefSNaoya Horiguchi swp_entry_t swp; 34244a705fefSNaoya Horiguchi 34254a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 34264a705fefSNaoya Horiguchi return 0; 34274a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 34284a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_hwpoison_entry(swp)) 34294a705fefSNaoya Horiguchi return 1; 34304a705fefSNaoya Horiguchi else 34314a705fefSNaoya Horiguchi return 0; 34324a705fefSNaoya Horiguchi } 34331e8f889bSDavid Gibson 343463551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 343563551ae0SDavid Gibson struct vm_area_struct *vma) 343663551ae0SDavid Gibson { 34375e41540cSMike Kravetz pte_t *src_pte, *dst_pte, entry, dst_entry; 343863551ae0SDavid Gibson struct page *ptepage; 34391c59827dSHugh Dickins unsigned long addr; 34401e8f889bSDavid Gibson int cow; 3441a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3442a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 3443c0d0381aSMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 3444ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 3445e8569dd2SAndreas Sandberg int ret = 0; 34461e8f889bSDavid Gibson 34471e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 344863551ae0SDavid Gibson 3449ac46d4f3SJérôme Glisse if (cow) { 34507269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src, 34516f4f13e8SJérôme Glisse vma->vm_start, 3452ac46d4f3SJérôme Glisse vma->vm_end); 3453ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 3454c0d0381aSMike Kravetz } else { 3455c0d0381aSMike Kravetz /* 3456c0d0381aSMike Kravetz * For shared mappings i_mmap_rwsem must be held to call 3457c0d0381aSMike Kravetz * huge_pte_alloc, otherwise the returned ptep could go 3458c0d0381aSMike Kravetz * away if part of a shared pmd and another thread calls 3459c0d0381aSMike Kravetz * huge_pmd_unshare. 3460c0d0381aSMike Kravetz */ 3461c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 3462ac46d4f3SJérôme Glisse } 3463e8569dd2SAndreas Sandberg 3464a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 3465cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 34667868a208SPunit Agrawal src_pte = huge_pte_offset(src, addr, sz); 3467c74df32cSHugh Dickins if (!src_pte) 3468c74df32cSHugh Dickins continue; 3469a5516438SAndi Kleen dst_pte = huge_pte_alloc(dst, addr, sz); 3470e8569dd2SAndreas Sandberg if (!dst_pte) { 3471e8569dd2SAndreas Sandberg ret = -ENOMEM; 3472e8569dd2SAndreas Sandberg break; 3473e8569dd2SAndreas Sandberg } 3474c5c99429SLarry Woodman 34755e41540cSMike Kravetz /* 34765e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 34775e41540cSMike Kravetz * dst_pte == src_pte is the common case of src/dest sharing. 34785e41540cSMike Kravetz * 34795e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 34805e41540cSMike Kravetz * another vma. If dst_pte !none, this implies sharing. 34815e41540cSMike Kravetz * Check here before taking page table lock, and once again 34825e41540cSMike Kravetz * after taking the lock below. 34835e41540cSMike Kravetz */ 34845e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 34855e41540cSMike Kravetz if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) 3486c5c99429SLarry Woodman continue; 3487c5c99429SLarry Woodman 3488cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 3489cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 3490cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 34914a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 34925e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 34935e41540cSMike Kravetz if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { 34945e41540cSMike Kravetz /* 34955e41540cSMike Kravetz * Skip if src entry none. Also, skip in the 34965e41540cSMike Kravetz * unlikely case dst entry !none as this implies 34975e41540cSMike Kravetz * sharing with another vma. 34985e41540cSMike Kravetz */ 34994a705fefSNaoya Horiguchi ; 35004a705fefSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry) || 35014a705fefSNaoya Horiguchi is_hugetlb_entry_hwpoisoned(entry))) { 35024a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 35034a705fefSNaoya Horiguchi 35044a705fefSNaoya Horiguchi if (is_write_migration_entry(swp_entry) && cow) { 35054a705fefSNaoya Horiguchi /* 35064a705fefSNaoya Horiguchi * COW mappings require pages in both 35074a705fefSNaoya Horiguchi * parent and child to be set to read. 35084a705fefSNaoya Horiguchi */ 35094a705fefSNaoya Horiguchi make_migration_entry_read(&swp_entry); 35104a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 3511e5251fd4SPunit Agrawal set_huge_swap_pte_at(src, addr, src_pte, 3512e5251fd4SPunit Agrawal entry, sz); 35134a705fefSNaoya Horiguchi } 3514e5251fd4SPunit Agrawal set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); 35154a705fefSNaoya Horiguchi } else { 351634ee645eSJoerg Roedel if (cow) { 35170f10851eSJérôme Glisse /* 35180f10851eSJérôme Glisse * No need to notify as we are downgrading page 35190f10851eSJérôme Glisse * table protection not changing it to point 35200f10851eSJérôme Glisse * to a new page. 35210f10851eSJérôme Glisse * 3522ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 35230f10851eSJérôme Glisse */ 35247f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 352534ee645eSJoerg Roedel } 35260253d634SNaoya Horiguchi entry = huge_ptep_get(src_pte); 352763551ae0SDavid Gibson ptepage = pte_page(entry); 352863551ae0SDavid Gibson get_page(ptepage); 352953f9263bSKirill A. Shutemov page_dup_rmap(ptepage, true); 353063551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 35315d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), dst); 35321c59827dSHugh Dickins } 3533cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 3534cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 353563551ae0SDavid Gibson } 353663551ae0SDavid Gibson 3537e8569dd2SAndreas Sandberg if (cow) 3538ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 3539c0d0381aSMike Kravetz else 3540c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 3541e8569dd2SAndreas Sandberg 3542e8569dd2SAndreas Sandberg return ret; 354363551ae0SDavid Gibson } 354463551ae0SDavid Gibson 354524669e58SAneesh Kumar K.V void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 354624669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 354724669e58SAneesh Kumar K.V struct page *ref_page) 354863551ae0SDavid Gibson { 354963551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 355063551ae0SDavid Gibson unsigned long address; 3551c7546f8fSDavid Gibson pte_t *ptep; 355263551ae0SDavid Gibson pte_t pte; 3553cb900f41SKirill A. Shutemov spinlock_t *ptl; 355463551ae0SDavid Gibson struct page *page; 3555a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3556a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 3557ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 3558a5516438SAndi Kleen 355963551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 3560a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 3561a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 356263551ae0SDavid Gibson 356307e32661SAneesh Kumar K.V /* 356407e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 356507e32661SAneesh Kumar K.V * to huge page. 356607e32661SAneesh Kumar K.V */ 3567ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 356824669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 3569dff11abeSMike Kravetz 3570dff11abeSMike Kravetz /* 3571dff11abeSMike Kravetz * If sharing possible, alert mmu notifiers of worst case. 3572dff11abeSMike Kravetz */ 35736f4f13e8SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, 35746f4f13e8SJérôme Glisse end); 3575ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 3576ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 3577569f48b8SHillf Danton address = start; 3578569f48b8SHillf Danton for (; address < end; address += sz) { 35797868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, sz); 3580c7546f8fSDavid Gibson if (!ptep) 3581c7546f8fSDavid Gibson continue; 3582c7546f8fSDavid Gibson 3583cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 358431d49da5SAneesh Kumar K.V if (huge_pmd_unshare(mm, &address, ptep)) { 358531d49da5SAneesh Kumar K.V spin_unlock(ptl); 3586dff11abeSMike Kravetz /* 3587dff11abeSMike Kravetz * We just unmapped a page of PMDs by clearing a PUD. 3588dff11abeSMike Kravetz * The caller's TLB flush range should cover this area. 3589dff11abeSMike Kravetz */ 359031d49da5SAneesh Kumar K.V continue; 359131d49da5SAneesh Kumar K.V } 359239dde65cSChen, Kenneth W 35936629326bSHillf Danton pte = huge_ptep_get(ptep); 359431d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 359531d49da5SAneesh Kumar K.V spin_unlock(ptl); 359631d49da5SAneesh Kumar K.V continue; 359731d49da5SAneesh Kumar K.V } 35986629326bSHillf Danton 35996629326bSHillf Danton /* 36009fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 36019fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 36026629326bSHillf Danton */ 36039fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 36049386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 360531d49da5SAneesh Kumar K.V spin_unlock(ptl); 360631d49da5SAneesh Kumar K.V continue; 36078c4894c6SNaoya Horiguchi } 36086629326bSHillf Danton 36096629326bSHillf Danton page = pte_page(pte); 361004f2cbe3SMel Gorman /* 361104f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 361204f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 361304f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 361404f2cbe3SMel Gorman */ 361504f2cbe3SMel Gorman if (ref_page) { 361631d49da5SAneesh Kumar K.V if (page != ref_page) { 361731d49da5SAneesh Kumar K.V spin_unlock(ptl); 361831d49da5SAneesh Kumar K.V continue; 361931d49da5SAneesh Kumar K.V } 362004f2cbe3SMel Gorman /* 362104f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 362204f2cbe3SMel Gorman * future faults in this VMA will fail rather than 362304f2cbe3SMel Gorman * looking like data was lost 362404f2cbe3SMel Gorman */ 362504f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 362604f2cbe3SMel Gorman } 362704f2cbe3SMel Gorman 3628c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 3629b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 3630106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 36316649a386SKen Chen set_page_dirty(page); 36329e81130bSHillf Danton 36335d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 3634d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 363531d49da5SAneesh Kumar K.V 3636cb900f41SKirill A. Shutemov spin_unlock(ptl); 3637e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 363824669e58SAneesh Kumar K.V /* 363931d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 364024669e58SAneesh Kumar K.V */ 364131d49da5SAneesh Kumar K.V if (ref_page) 364231d49da5SAneesh Kumar K.V break; 3643fe1668aeSChen, Kenneth W } 3644ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 364524669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 36461da177e4SLinus Torvalds } 364763551ae0SDavid Gibson 3648d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 3649d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 3650d833352aSMel Gorman unsigned long end, struct page *ref_page) 3651d833352aSMel Gorman { 3652d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 3653d833352aSMel Gorman 3654d833352aSMel Gorman /* 3655d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 3656d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 3657d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 3658d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 3659d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 3660c8c06efaSDavidlohr Bueso * is to clear it before releasing the i_mmap_rwsem. This works 3661d833352aSMel Gorman * because in the context this is called, the VMA is about to be 3662c8c06efaSDavidlohr Bueso * destroyed and the i_mmap_rwsem is held. 3663d833352aSMel Gorman */ 3664d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 3665d833352aSMel Gorman } 3666d833352aSMel Gorman 3667502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 366804f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 3669502717f4SChen, Kenneth W { 367024669e58SAneesh Kumar K.V struct mm_struct *mm; 367124669e58SAneesh Kumar K.V struct mmu_gather tlb; 3672dff11abeSMike Kravetz unsigned long tlb_start = start; 3673dff11abeSMike Kravetz unsigned long tlb_end = end; 3674dff11abeSMike Kravetz 3675dff11abeSMike Kravetz /* 3676dff11abeSMike Kravetz * If shared PMDs were possibly used within this vma range, adjust 3677dff11abeSMike Kravetz * start/end for worst case tlb flushing. 3678dff11abeSMike Kravetz * Note that we can not be sure if PMDs are shared until we try to 3679dff11abeSMike Kravetz * unmap pages. However, we want to make sure TLB flushing covers 3680dff11abeSMike Kravetz * the largest possible range. 3681dff11abeSMike Kravetz */ 3682dff11abeSMike Kravetz adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end); 368324669e58SAneesh Kumar K.V 368424669e58SAneesh Kumar K.V mm = vma->vm_mm; 368524669e58SAneesh Kumar K.V 3686dff11abeSMike Kravetz tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end); 368724669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 3688dff11abeSMike Kravetz tlb_finish_mmu(&tlb, tlb_start, tlb_end); 3689502717f4SChen, Kenneth W } 3690502717f4SChen, Kenneth W 369104f2cbe3SMel Gorman /* 369204f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 369304f2cbe3SMel Gorman * mappping it owns the reserve page for. The intention is to unmap the page 369404f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 369504f2cbe3SMel Gorman * same region. 369604f2cbe3SMel Gorman */ 36972f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 36982a4b3dedSHarvey Harrison struct page *page, unsigned long address) 369904f2cbe3SMel Gorman { 37007526674dSAdam Litke struct hstate *h = hstate_vma(vma); 370104f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 370204f2cbe3SMel Gorman struct address_space *mapping; 370304f2cbe3SMel Gorman pgoff_t pgoff; 370404f2cbe3SMel Gorman 370504f2cbe3SMel Gorman /* 370604f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 370704f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 370804f2cbe3SMel Gorman */ 37097526674dSAdam Litke address = address & huge_page_mask(h); 371036e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 371136e4f20aSMichal Hocko vma->vm_pgoff; 371293c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 371304f2cbe3SMel Gorman 37144eb2b1dcSMel Gorman /* 37154eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 37164eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 37174eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 37184eb2b1dcSMel Gorman */ 371983cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 37206b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 372104f2cbe3SMel Gorman /* Do not unmap the current VMA */ 372204f2cbe3SMel Gorman if (iter_vma == vma) 372304f2cbe3SMel Gorman continue; 372404f2cbe3SMel Gorman 372504f2cbe3SMel Gorman /* 37262f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 37272f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 37282f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 37292f84a899SMel Gorman */ 37302f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 37312f84a899SMel Gorman continue; 37322f84a899SMel Gorman 37332f84a899SMel Gorman /* 373404f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 373504f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 373604f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 373704f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 373804f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 373904f2cbe3SMel Gorman */ 374004f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 374124669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 374224669e58SAneesh Kumar K.V address + huge_page_size(h), page); 374304f2cbe3SMel Gorman } 374483cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 374504f2cbe3SMel Gorman } 374604f2cbe3SMel Gorman 37470fe6e20bSNaoya Horiguchi /* 37480fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 3749ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 3750ef009b25SMichal Hocko * cannot race with other handlers or page migration. 3751ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 37520fe6e20bSNaoya Horiguchi */ 37532b740303SSouptick Joarder static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 3754974e6d66SHuang Ying unsigned long address, pte_t *ptep, 3755cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 37561e8f889bSDavid Gibson { 37573999f52eSAneesh Kumar K.V pte_t pte; 3758a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 37591e8f889bSDavid Gibson struct page *old_page, *new_page; 37602b740303SSouptick Joarder int outside_reserve = 0; 37612b740303SSouptick Joarder vm_fault_t ret = 0; 3762974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 3763ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 37641e8f889bSDavid Gibson 37653999f52eSAneesh Kumar K.V pte = huge_ptep_get(ptep); 37661e8f889bSDavid Gibson old_page = pte_page(pte); 37671e8f889bSDavid Gibson 376804f2cbe3SMel Gorman retry_avoidcopy: 37691e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 37701e8f889bSDavid Gibson * and just make the page writable */ 377137a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 37725a49973dSHugh Dickins page_move_anon_rmap(old_page, vma); 37735b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 377483c54070SNick Piggin return 0; 37751e8f889bSDavid Gibson } 37761e8f889bSDavid Gibson 377704f2cbe3SMel Gorman /* 377804f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 377904f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 378004f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 378104f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 378204f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 378304f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 378404f2cbe3SMel Gorman * of the full address range. 378504f2cbe3SMel Gorman */ 37865944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 378704f2cbe3SMel Gorman old_page != pagecache_page) 378804f2cbe3SMel Gorman outside_reserve = 1; 378904f2cbe3SMel Gorman 379009cbfeafSKirill A. Shutemov get_page(old_page); 3791b76c8cfbSLarry Woodman 3792ad4404a2SDavidlohr Bueso /* 3793ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 3794ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 3795ad4404a2SDavidlohr Bueso */ 3796cb900f41SKirill A. Shutemov spin_unlock(ptl); 37975b7a1d40SHuang Ying new_page = alloc_huge_page(vma, haddr, outside_reserve); 37981e8f889bSDavid Gibson 37992fc39cecSAdam Litke if (IS_ERR(new_page)) { 380004f2cbe3SMel Gorman /* 380104f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 380204f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 380304f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 380404f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 380504f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 380604f2cbe3SMel Gorman */ 380704f2cbe3SMel Gorman if (outside_reserve) { 380809cbfeafSKirill A. Shutemov put_page(old_page); 380904f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 38105b7a1d40SHuang Ying unmap_ref_private(mm, vma, old_page, haddr); 381104f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 3812cb900f41SKirill A. Shutemov spin_lock(ptl); 38135b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 3814a9af0c5dSNaoya Horiguchi if (likely(ptep && 3815a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 381604f2cbe3SMel Gorman goto retry_avoidcopy; 3817a734bcc8SHillf Danton /* 3818cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 3819cb900f41SKirill A. Shutemov * lock, and our job is done. 3820a734bcc8SHillf Danton */ 3821a734bcc8SHillf Danton return 0; 382204f2cbe3SMel Gorman } 382304f2cbe3SMel Gorman 38242b740303SSouptick Joarder ret = vmf_error(PTR_ERR(new_page)); 3825ad4404a2SDavidlohr Bueso goto out_release_old; 38261e8f889bSDavid Gibson } 38271e8f889bSDavid Gibson 38280fe6e20bSNaoya Horiguchi /* 38290fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 38300fe6e20bSNaoya Horiguchi * anon_vma prepared. 38310fe6e20bSNaoya Horiguchi */ 383244e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 3833ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 3834ad4404a2SDavidlohr Bueso goto out_release_all; 383544e2aa93SDean Nelson } 38360fe6e20bSNaoya Horiguchi 3837974e6d66SHuang Ying copy_user_huge_page(new_page, old_page, address, vma, 383847ad8475SAndrea Arcangeli pages_per_huge_page(h)); 38390ed361deSNick Piggin __SetPageUptodate(new_page); 38401e8f889bSDavid Gibson 38417269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, 38426f4f13e8SJérôme Glisse haddr + huge_page_size(h)); 3843ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 3844ad4404a2SDavidlohr Bueso 3845b76c8cfbSLarry Woodman /* 3846cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 3847b76c8cfbSLarry Woodman * before the page tables are altered 3848b76c8cfbSLarry Woodman */ 3849cb900f41SKirill A. Shutemov spin_lock(ptl); 38505b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 3851a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 385207443a85SJoonsoo Kim ClearPagePrivate(new_page); 385307443a85SJoonsoo Kim 38541e8f889bSDavid Gibson /* Break COW */ 38555b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 3856ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, range.end); 38575b7a1d40SHuang Ying set_huge_pte_at(mm, haddr, ptep, 38581e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 3859d281ee61SKirill A. Shutemov page_remove_rmap(old_page, true); 38605b7a1d40SHuang Ying hugepage_add_new_anon_rmap(new_page, vma, haddr); 3861cb6acd01SMike Kravetz set_page_huge_active(new_page); 38621e8f889bSDavid Gibson /* Make the old page be freed below */ 38631e8f889bSDavid Gibson new_page = old_page; 38641e8f889bSDavid Gibson } 3865cb900f41SKirill A. Shutemov spin_unlock(ptl); 3866ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 3867ad4404a2SDavidlohr Bueso out_release_all: 38685b7a1d40SHuang Ying restore_reserve_on_error(h, vma, haddr, new_page); 386909cbfeafSKirill A. Shutemov put_page(new_page); 3870ad4404a2SDavidlohr Bueso out_release_old: 387109cbfeafSKirill A. Shutemov put_page(old_page); 38728312034fSJoonsoo Kim 3873ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 3874ad4404a2SDavidlohr Bueso return ret; 38751e8f889bSDavid Gibson } 38761e8f889bSDavid Gibson 387704f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 3878a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 3879a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 388004f2cbe3SMel Gorman { 388104f2cbe3SMel Gorman struct address_space *mapping; 3882e7c4b0bfSAndy Whitcroft pgoff_t idx; 388304f2cbe3SMel Gorman 388404f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 3885a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 388604f2cbe3SMel Gorman 388704f2cbe3SMel Gorman return find_lock_page(mapping, idx); 388804f2cbe3SMel Gorman } 388904f2cbe3SMel Gorman 38903ae77f43SHugh Dickins /* 38913ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 38923ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 38933ae77f43SHugh Dickins */ 38943ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 38952a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 38962a15efc9SHugh Dickins { 38972a15efc9SHugh Dickins struct address_space *mapping; 38982a15efc9SHugh Dickins pgoff_t idx; 38992a15efc9SHugh Dickins struct page *page; 39002a15efc9SHugh Dickins 39012a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 39022a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 39032a15efc9SHugh Dickins 39042a15efc9SHugh Dickins page = find_get_page(mapping, idx); 39052a15efc9SHugh Dickins if (page) 39062a15efc9SHugh Dickins put_page(page); 39072a15efc9SHugh Dickins return page != NULL; 39082a15efc9SHugh Dickins } 39092a15efc9SHugh Dickins 3910ab76ad54SMike Kravetz int huge_add_to_page_cache(struct page *page, struct address_space *mapping, 3911ab76ad54SMike Kravetz pgoff_t idx) 3912ab76ad54SMike Kravetz { 3913ab76ad54SMike Kravetz struct inode *inode = mapping->host; 3914ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 3915ab76ad54SMike Kravetz int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 3916ab76ad54SMike Kravetz 3917ab76ad54SMike Kravetz if (err) 3918ab76ad54SMike Kravetz return err; 3919ab76ad54SMike Kravetz ClearPagePrivate(page); 3920ab76ad54SMike Kravetz 392122146c3cSMike Kravetz /* 392222146c3cSMike Kravetz * set page dirty so that it will not be removed from cache/file 392322146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 392422146c3cSMike Kravetz */ 392522146c3cSMike Kravetz set_page_dirty(page); 392622146c3cSMike Kravetz 3927ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 3928ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 3929ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 3930ab76ad54SMike Kravetz return 0; 3931ab76ad54SMike Kravetz } 3932ab76ad54SMike Kravetz 39332b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 39342b740303SSouptick Joarder struct vm_area_struct *vma, 39358382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 3936788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 3937ac9b9c66SHugh Dickins { 3938a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 39392b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 3940409eb8c2SHillf Danton int anon_rmap = 0; 39414c887265SAdam Litke unsigned long size; 39424c887265SAdam Litke struct page *page; 39431e8f889bSDavid Gibson pte_t new_pte; 3944cb900f41SKirill A. Shutemov spinlock_t *ptl; 3945285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 3946cb6acd01SMike Kravetz bool new_page = false; 39474c887265SAdam Litke 394804f2cbe3SMel Gorman /* 394904f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 395004f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 395125985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 395204f2cbe3SMel Gorman */ 395304f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 3954910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 395504f2cbe3SMel Gorman current->pid); 395604f2cbe3SMel Gorman return ret; 395704f2cbe3SMel Gorman } 395804f2cbe3SMel Gorman 39594c887265SAdam Litke /* 396087bf91d3SMike Kravetz * We can not race with truncation due to holding i_mmap_rwsem. 396187bf91d3SMike Kravetz * i_size is modified when holding i_mmap_rwsem, so check here 396287bf91d3SMike Kravetz * once for faults beyond end of file. 39634c887265SAdam Litke */ 3964a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 3965ebed4bfcSHugh Dickins if (idx >= size) 3966ebed4bfcSHugh Dickins goto out; 39671a1aad8aSMike Kravetz 396887bf91d3SMike Kravetz retry: 396987bf91d3SMike Kravetz page = find_lock_page(mapping, idx); 397087bf91d3SMike Kravetz if (!page) { 39711a1aad8aSMike Kravetz /* 39721a1aad8aSMike Kravetz * Check for page in userfault range 39731a1aad8aSMike Kravetz */ 39741a1aad8aSMike Kravetz if (userfaultfd_missing(vma)) { 39751a1aad8aSMike Kravetz u32 hash; 39761a1aad8aSMike Kravetz struct vm_fault vmf = { 39771a1aad8aSMike Kravetz .vma = vma, 3978285b8dcaSHuang Ying .address = haddr, 39791a1aad8aSMike Kravetz .flags = flags, 39801a1aad8aSMike Kravetz /* 39811a1aad8aSMike Kravetz * Hard to debug if it ends up being 39821a1aad8aSMike Kravetz * used by a callee that assumes 39831a1aad8aSMike Kravetz * something about the other 39841a1aad8aSMike Kravetz * uninitialized fields... same as in 39851a1aad8aSMike Kravetz * memory.c 39861a1aad8aSMike Kravetz */ 39871a1aad8aSMike Kravetz }; 39881a1aad8aSMike Kravetz 39891a1aad8aSMike Kravetz /* 3990c0d0381aSMike Kravetz * hugetlb_fault_mutex and i_mmap_rwsem must be 3991c0d0381aSMike Kravetz * dropped before handling userfault. Reacquire 3992c0d0381aSMike Kravetz * after handling fault to make calling code simpler. 39931a1aad8aSMike Kravetz */ 3994188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 39951a1aad8aSMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 3996c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 39971a1aad8aSMike Kravetz ret = handle_userfault(&vmf, VM_UFFD_MISSING); 3998c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 39991a1aad8aSMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 40001a1aad8aSMike Kravetz goto out; 40011a1aad8aSMike Kravetz } 40021a1aad8aSMike Kravetz 4003285b8dcaSHuang Ying page = alloc_huge_page(vma, haddr, 0); 40042fc39cecSAdam Litke if (IS_ERR(page)) { 40054643d67eSMike Kravetz /* 40064643d67eSMike Kravetz * Returning error will result in faulting task being 40074643d67eSMike Kravetz * sent SIGBUS. The hugetlb fault mutex prevents two 40084643d67eSMike Kravetz * tasks from racing to fault in the same page which 40094643d67eSMike Kravetz * could result in false unable to allocate errors. 40104643d67eSMike Kravetz * Page migration does not take the fault mutex, but 40114643d67eSMike Kravetz * does a clear then write of pte's under page table 40124643d67eSMike Kravetz * lock. Page fault code could race with migration, 40134643d67eSMike Kravetz * notice the clear pte and try to allocate a page 40144643d67eSMike Kravetz * here. Before returning error, get ptl and make 40154643d67eSMike Kravetz * sure there really is no pte entry. 40164643d67eSMike Kravetz */ 40174643d67eSMike Kravetz ptl = huge_pte_lock(h, mm, ptep); 40184643d67eSMike Kravetz if (!huge_pte_none(huge_ptep_get(ptep))) { 40194643d67eSMike Kravetz ret = 0; 40204643d67eSMike Kravetz spin_unlock(ptl); 40214643d67eSMike Kravetz goto out; 40224643d67eSMike Kravetz } 40234643d67eSMike Kravetz spin_unlock(ptl); 40242b740303SSouptick Joarder ret = vmf_error(PTR_ERR(page)); 40256bda666aSChristoph Lameter goto out; 40266bda666aSChristoph Lameter } 402747ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 40280ed361deSNick Piggin __SetPageUptodate(page); 4029cb6acd01SMike Kravetz new_page = true; 4030ac9b9c66SHugh Dickins 4031f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 4032ab76ad54SMike Kravetz int err = huge_add_to_page_cache(page, mapping, idx); 40336bda666aSChristoph Lameter if (err) { 40346bda666aSChristoph Lameter put_page(page); 40356bda666aSChristoph Lameter if (err == -EEXIST) 40366bda666aSChristoph Lameter goto retry; 40376bda666aSChristoph Lameter goto out; 40386bda666aSChristoph Lameter } 403923be7468SMel Gorman } else { 40406bda666aSChristoph Lameter lock_page(page); 40410fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 40420fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 40430fe6e20bSNaoya Horiguchi goto backout_unlocked; 404423be7468SMel Gorman } 4045409eb8c2SHillf Danton anon_rmap = 1; 40460fe6e20bSNaoya Horiguchi } 40470fe6e20bSNaoya Horiguchi } else { 404857303d80SAndy Whitcroft /* 4049998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 4050998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 4051998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 4052fd6a03edSNaoya Horiguchi */ 4053fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 4054aa50d3a7SAndi Kleen ret = VM_FAULT_HWPOISON | 4055972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 4056fd6a03edSNaoya Horiguchi goto backout_unlocked; 40576bda666aSChristoph Lameter } 4058998b4382SNaoya Horiguchi } 40591e8f889bSDavid Gibson 406057303d80SAndy Whitcroft /* 406157303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 406257303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 406357303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 406457303d80SAndy Whitcroft * the spinlock. 406557303d80SAndy Whitcroft */ 40665e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 4067285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 40682b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 40692b26736cSAndy Whitcroft goto backout_unlocked; 40702b26736cSAndy Whitcroft } 40715e911373SMike Kravetz /* Just decrements count, does not deallocate */ 4072285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 40735e911373SMike Kravetz } 407457303d80SAndy Whitcroft 40758bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 407683c54070SNick Piggin ret = 0; 40777f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 40784c887265SAdam Litke goto backout; 40794c887265SAdam Litke 408007443a85SJoonsoo Kim if (anon_rmap) { 408107443a85SJoonsoo Kim ClearPagePrivate(page); 4082285b8dcaSHuang Ying hugepage_add_new_anon_rmap(page, vma, haddr); 4083ac714904SChoi Gi-yong } else 408453f9263bSKirill A. Shutemov page_dup_rmap(page, true); 40851e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 40861e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 4087285b8dcaSHuang Ying set_huge_pte_at(mm, haddr, ptep, new_pte); 40881e8f889bSDavid Gibson 40895d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 4090788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 40911e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 4092974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); 40931e8f889bSDavid Gibson } 40941e8f889bSDavid Gibson 4095cb900f41SKirill A. Shutemov spin_unlock(ptl); 4096cb6acd01SMike Kravetz 4097cb6acd01SMike Kravetz /* 4098cb6acd01SMike Kravetz * Only make newly allocated pages active. Existing pages found 4099cb6acd01SMike Kravetz * in the pagecache could be !page_huge_active() if they have been 4100cb6acd01SMike Kravetz * isolated for migration. 4101cb6acd01SMike Kravetz */ 4102cb6acd01SMike Kravetz if (new_page) 4103cb6acd01SMike Kravetz set_page_huge_active(page); 4104cb6acd01SMike Kravetz 41054c887265SAdam Litke unlock_page(page); 41064c887265SAdam Litke out: 4107ac9b9c66SHugh Dickins return ret; 41084c887265SAdam Litke 41094c887265SAdam Litke backout: 4110cb900f41SKirill A. Shutemov spin_unlock(ptl); 41112b26736cSAndy Whitcroft backout_unlocked: 41124c887265SAdam Litke unlock_page(page); 4113285b8dcaSHuang Ying restore_reserve_on_error(h, vma, haddr, page); 41144c887265SAdam Litke put_page(page); 41154c887265SAdam Litke goto out; 4116ac9b9c66SHugh Dickins } 4117ac9b9c66SHugh Dickins 41188382d914SDavidlohr Bueso #ifdef CONFIG_SMP 4119188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 41208382d914SDavidlohr Bueso { 41218382d914SDavidlohr Bueso unsigned long key[2]; 41228382d914SDavidlohr Bueso u32 hash; 41238382d914SDavidlohr Bueso 41248382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 41258382d914SDavidlohr Bueso key[1] = idx; 41268382d914SDavidlohr Bueso 412755254636SMike Kravetz hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); 41288382d914SDavidlohr Bueso 41298382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 41308382d914SDavidlohr Bueso } 41318382d914SDavidlohr Bueso #else 41328382d914SDavidlohr Bueso /* 41338382d914SDavidlohr Bueso * For uniprocesor systems we always use a single mutex, so just 41348382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 41358382d914SDavidlohr Bueso */ 4136188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 41378382d914SDavidlohr Bueso { 41388382d914SDavidlohr Bueso return 0; 41398382d914SDavidlohr Bueso } 41408382d914SDavidlohr Bueso #endif 41418382d914SDavidlohr Bueso 41422b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 4143788c7df4SHugh Dickins unsigned long address, unsigned int flags) 414486e5216fSAdam Litke { 41458382d914SDavidlohr Bueso pte_t *ptep, entry; 4146cb900f41SKirill A. Shutemov spinlock_t *ptl; 41472b740303SSouptick Joarder vm_fault_t ret; 41488382d914SDavidlohr Bueso u32 hash; 41498382d914SDavidlohr Bueso pgoff_t idx; 41500fe6e20bSNaoya Horiguchi struct page *page = NULL; 415157303d80SAndy Whitcroft struct page *pagecache_page = NULL; 4152a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 41538382d914SDavidlohr Bueso struct address_space *mapping; 41540f792cf9SNaoya Horiguchi int need_wait_lock = 0; 4155285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 415686e5216fSAdam Litke 4157285b8dcaSHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4158fd6a03edSNaoya Horiguchi if (ptep) { 4159c0d0381aSMike Kravetz /* 4160c0d0381aSMike Kravetz * Since we hold no locks, ptep could be stale. That is 4161c0d0381aSMike Kravetz * OK as we are only making decisions based on content and 4162c0d0381aSMike Kravetz * not actually modifying content here. 4163c0d0381aSMike Kravetz */ 4164fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 4165290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 4166cb900f41SKirill A. Shutemov migration_entry_wait_huge(vma, mm, ptep); 4167290408d4SNaoya Horiguchi return 0; 4168290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 4169aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 4170972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 4171ddeaab32SMike Kravetz } else { 4172b43a9990SMike Kravetz ptep = huge_pte_alloc(mm, haddr, huge_page_size(h)); 4173ddeaab32SMike Kravetz if (!ptep) 4174b43a9990SMike Kravetz return VM_FAULT_OOM; 4175b43a9990SMike Kravetz } 41768382d914SDavidlohr Bueso 4177c0d0381aSMike Kravetz /* 4178c0d0381aSMike Kravetz * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold 417987bf91d3SMike Kravetz * until finished with ptep. This serves two purposes: 418087bf91d3SMike Kravetz * 1) It prevents huge_pmd_unshare from being called elsewhere 418187bf91d3SMike Kravetz * and making the ptep no longer valid. 418287bf91d3SMike Kravetz * 2) It synchronizes us with i_size modifications during truncation. 4183c0d0381aSMike Kravetz * 4184c0d0381aSMike Kravetz * ptep could have already be assigned via huge_pte_offset. That 4185c0d0381aSMike Kravetz * is OK, as huge_pte_alloc will return the same value unless 4186c0d0381aSMike Kravetz * something has changed. 4187c0d0381aSMike Kravetz */ 4188ddeaab32SMike Kravetz mapping = vma->vm_file->f_mapping; 4189c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 4190c0d0381aSMike Kravetz ptep = huge_pte_alloc(mm, haddr, huge_page_size(h)); 4191c0d0381aSMike Kravetz if (!ptep) { 4192c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 4193c0d0381aSMike Kravetz return VM_FAULT_OOM; 4194c0d0381aSMike Kravetz } 4195ddeaab32SMike Kravetz 41963935baa9SDavid Gibson /* 41973935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 41983935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 41993935baa9SDavid Gibson * the same page in the page cache. 42003935baa9SDavid Gibson */ 4201c0d0381aSMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 4202188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 4203c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 42048382d914SDavidlohr Bueso 42057f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 42067f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 42078382d914SDavidlohr Bueso ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 4208b4d1d99fSDavid Gibson goto out_mutex; 42093935baa9SDavid Gibson } 421086e5216fSAdam Litke 421183c54070SNick Piggin ret = 0; 42121e8f889bSDavid Gibson 421357303d80SAndy Whitcroft /* 42140f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 42150f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 42160f792cf9SNaoya Horiguchi * a active hugepage in pagecache. This goto expects the 2nd page fault, 42170f792cf9SNaoya Horiguchi * and is_hugetlb_entry_(migration|hwpoisoned) check will properly 42180f792cf9SNaoya Horiguchi * handle it. 42190f792cf9SNaoya Horiguchi */ 42200f792cf9SNaoya Horiguchi if (!pte_present(entry)) 42210f792cf9SNaoya Horiguchi goto out_mutex; 42220f792cf9SNaoya Horiguchi 42230f792cf9SNaoya Horiguchi /* 422457303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 422557303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 422657303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 422757303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 422857303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 422957303d80SAndy Whitcroft * consumed. 423057303d80SAndy Whitcroft */ 4231106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 4232285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 42332b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 4234b4d1d99fSDavid Gibson goto out_mutex; 42352b26736cSAndy Whitcroft } 42365e911373SMike Kravetz /* Just decrements count, does not deallocate */ 4237285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 423857303d80SAndy Whitcroft 4239f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 424057303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 4241285b8dcaSHuang Ying vma, haddr); 424257303d80SAndy Whitcroft } 424357303d80SAndy Whitcroft 42440f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 42450fe6e20bSNaoya Horiguchi 42461e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 4247b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 4248cb900f41SKirill A. Shutemov goto out_ptl; 4249b4d1d99fSDavid Gibson 42500f792cf9SNaoya Horiguchi /* 42510f792cf9SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 42520f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 42530f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 42540f792cf9SNaoya Horiguchi */ 42550f792cf9SNaoya Horiguchi page = pte_page(entry); 42560f792cf9SNaoya Horiguchi if (page != pagecache_page) 42570f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 42580f792cf9SNaoya Horiguchi need_wait_lock = 1; 42590f792cf9SNaoya Horiguchi goto out_ptl; 42600f792cf9SNaoya Horiguchi } 42610f792cf9SNaoya Horiguchi 42620f792cf9SNaoya Horiguchi get_page(page); 4263b4d1d99fSDavid Gibson 4264788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 4265106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 4266974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, 4267cb900f41SKirill A. Shutemov pagecache_page, ptl); 42680f792cf9SNaoya Horiguchi goto out_put_page; 4269b4d1d99fSDavid Gibson } 4270106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 4271b4d1d99fSDavid Gibson } 4272b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 4273285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 4274788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 4275285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 42760f792cf9SNaoya Horiguchi out_put_page: 42770f792cf9SNaoya Horiguchi if (page != pagecache_page) 42780f792cf9SNaoya Horiguchi unlock_page(page); 42790f792cf9SNaoya Horiguchi put_page(page); 4280cb900f41SKirill A. Shutemov out_ptl: 4281cb900f41SKirill A. Shutemov spin_unlock(ptl); 428257303d80SAndy Whitcroft 428357303d80SAndy Whitcroft if (pagecache_page) { 428457303d80SAndy Whitcroft unlock_page(pagecache_page); 428557303d80SAndy Whitcroft put_page(pagecache_page); 428657303d80SAndy Whitcroft } 4287b4d1d99fSDavid Gibson out_mutex: 4288c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 4289c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 42900f792cf9SNaoya Horiguchi /* 42910f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 42920f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 42930f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 42940f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 42950f792cf9SNaoya Horiguchi * here without taking refcount. 42960f792cf9SNaoya Horiguchi */ 42970f792cf9SNaoya Horiguchi if (need_wait_lock) 42980f792cf9SNaoya Horiguchi wait_on_page_locked(page); 42991e8f889bSDavid Gibson return ret; 430086e5216fSAdam Litke } 430186e5216fSAdam Litke 43028fb5debcSMike Kravetz /* 43038fb5debcSMike Kravetz * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with 43048fb5debcSMike Kravetz * modifications for huge pages. 43058fb5debcSMike Kravetz */ 43068fb5debcSMike Kravetz int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, 43078fb5debcSMike Kravetz pte_t *dst_pte, 43088fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 43098fb5debcSMike Kravetz unsigned long dst_addr, 43108fb5debcSMike Kravetz unsigned long src_addr, 43118fb5debcSMike Kravetz struct page **pagep) 43128fb5debcSMike Kravetz { 43131e392147SAndrea Arcangeli struct address_space *mapping; 43141e392147SAndrea Arcangeli pgoff_t idx; 43151e392147SAndrea Arcangeli unsigned long size; 43161c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 43178fb5debcSMike Kravetz struct hstate *h = hstate_vma(dst_vma); 43188fb5debcSMike Kravetz pte_t _dst_pte; 43198fb5debcSMike Kravetz spinlock_t *ptl; 43208fb5debcSMike Kravetz int ret; 43218fb5debcSMike Kravetz struct page *page; 43228fb5debcSMike Kravetz 43238fb5debcSMike Kravetz if (!*pagep) { 43248fb5debcSMike Kravetz ret = -ENOMEM; 43258fb5debcSMike Kravetz page = alloc_huge_page(dst_vma, dst_addr, 0); 43268fb5debcSMike Kravetz if (IS_ERR(page)) 43278fb5debcSMike Kravetz goto out; 43288fb5debcSMike Kravetz 43298fb5debcSMike Kravetz ret = copy_huge_page_from_user(page, 43308fb5debcSMike Kravetz (const void __user *) src_addr, 4331810a56b9SMike Kravetz pages_per_huge_page(h), false); 43328fb5debcSMike Kravetz 43338fb5debcSMike Kravetz /* fallback to copy_from_user outside mmap_sem */ 43348fb5debcSMike Kravetz if (unlikely(ret)) { 43359e368259SAndrea Arcangeli ret = -ENOENT; 43368fb5debcSMike Kravetz *pagep = page; 43378fb5debcSMike Kravetz /* don't free the page */ 43388fb5debcSMike Kravetz goto out; 43398fb5debcSMike Kravetz } 43408fb5debcSMike Kravetz } else { 43418fb5debcSMike Kravetz page = *pagep; 43428fb5debcSMike Kravetz *pagep = NULL; 43438fb5debcSMike Kravetz } 43448fb5debcSMike Kravetz 43458fb5debcSMike Kravetz /* 43468fb5debcSMike Kravetz * The memory barrier inside __SetPageUptodate makes sure that 43478fb5debcSMike Kravetz * preceding stores to the page contents become visible before 43488fb5debcSMike Kravetz * the set_pte_at() write. 43498fb5debcSMike Kravetz */ 43508fb5debcSMike Kravetz __SetPageUptodate(page); 43518fb5debcSMike Kravetz 43521e392147SAndrea Arcangeli mapping = dst_vma->vm_file->f_mapping; 43531e392147SAndrea Arcangeli idx = vma_hugecache_offset(h, dst_vma, dst_addr); 43541e392147SAndrea Arcangeli 43551c9e8defSMike Kravetz /* 43561c9e8defSMike Kravetz * If shared, add to page cache 43571c9e8defSMike Kravetz */ 43581c9e8defSMike Kravetz if (vm_shared) { 43591e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 43601e392147SAndrea Arcangeli ret = -EFAULT; 43611e392147SAndrea Arcangeli if (idx >= size) 43621e392147SAndrea Arcangeli goto out_release_nounlock; 43631c9e8defSMike Kravetz 43641e392147SAndrea Arcangeli /* 43651e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 43661e392147SAndrea Arcangeli * huge_add_to_page_cache() below happens through the 43671e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 43681e392147SAndrea Arcangeli * the caller. 43691e392147SAndrea Arcangeli */ 43701c9e8defSMike Kravetz ret = huge_add_to_page_cache(page, mapping, idx); 43711c9e8defSMike Kravetz if (ret) 43721c9e8defSMike Kravetz goto out_release_nounlock; 43731c9e8defSMike Kravetz } 43741c9e8defSMike Kravetz 43758fb5debcSMike Kravetz ptl = huge_pte_lockptr(h, dst_mm, dst_pte); 43768fb5debcSMike Kravetz spin_lock(ptl); 43778fb5debcSMike Kravetz 43781e392147SAndrea Arcangeli /* 43791e392147SAndrea Arcangeli * Recheck the i_size after holding PT lock to make sure not 43801e392147SAndrea Arcangeli * to leave any page mapped (as page_mapped()) beyond the end 43811e392147SAndrea Arcangeli * of the i_size (remove_inode_hugepages() is strict about 43821e392147SAndrea Arcangeli * enforcing that). If we bail out here, we'll also leave a 43831e392147SAndrea Arcangeli * page in the radix tree in the vm_shared case beyond the end 43841e392147SAndrea Arcangeli * of the i_size, but remove_inode_hugepages() will take care 43851e392147SAndrea Arcangeli * of it as soon as we drop the hugetlb_fault_mutex_table. 43861e392147SAndrea Arcangeli */ 43871e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 43881e392147SAndrea Arcangeli ret = -EFAULT; 43891e392147SAndrea Arcangeli if (idx >= size) 43901e392147SAndrea Arcangeli goto out_release_unlock; 43911e392147SAndrea Arcangeli 43928fb5debcSMike Kravetz ret = -EEXIST; 43938fb5debcSMike Kravetz if (!huge_pte_none(huge_ptep_get(dst_pte))) 43948fb5debcSMike Kravetz goto out_release_unlock; 43958fb5debcSMike Kravetz 43961c9e8defSMike Kravetz if (vm_shared) { 43971c9e8defSMike Kravetz page_dup_rmap(page, true); 43981c9e8defSMike Kravetz } else { 43998fb5debcSMike Kravetz ClearPagePrivate(page); 44008fb5debcSMike Kravetz hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); 44011c9e8defSMike Kravetz } 44028fb5debcSMike Kravetz 44038fb5debcSMike Kravetz _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); 44048fb5debcSMike Kravetz if (dst_vma->vm_flags & VM_WRITE) 44058fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 44068fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 44078fb5debcSMike Kravetz 44088fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 44098fb5debcSMike Kravetz 44108fb5debcSMike Kravetz (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, 44118fb5debcSMike Kravetz dst_vma->vm_flags & VM_WRITE); 44128fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 44138fb5debcSMike Kravetz 44148fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 44158fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 44168fb5debcSMike Kravetz 44178fb5debcSMike Kravetz spin_unlock(ptl); 4418cb6acd01SMike Kravetz set_page_huge_active(page); 44191c9e8defSMike Kravetz if (vm_shared) 44201c9e8defSMike Kravetz unlock_page(page); 44218fb5debcSMike Kravetz ret = 0; 44228fb5debcSMike Kravetz out: 44238fb5debcSMike Kravetz return ret; 44248fb5debcSMike Kravetz out_release_unlock: 44258fb5debcSMike Kravetz spin_unlock(ptl); 44261c9e8defSMike Kravetz if (vm_shared) 44271c9e8defSMike Kravetz unlock_page(page); 44285af10dfdSAndrea Arcangeli out_release_nounlock: 44298fb5debcSMike Kravetz put_page(page); 44308fb5debcSMike Kravetz goto out; 44318fb5debcSMike Kravetz } 44328fb5debcSMike Kravetz 443328a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 443463551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 443528a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 44364f6da934SPeter Xu long i, unsigned int flags, int *locked) 443763551ae0SDavid Gibson { 4438d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 4439d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 444028a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 4441a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 44422be7cfedSDaniel Jordan int err = -EFAULT; 444363551ae0SDavid Gibson 444463551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 444563551ae0SDavid Gibson pte_t *pte; 4446cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 44472a15efc9SHugh Dickins int absent; 444863551ae0SDavid Gibson struct page *page; 444963551ae0SDavid Gibson 44504c887265SAdam Litke /* 445102057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 445202057967SDavid Rientjes * potentially allocating memory. 445302057967SDavid Rientjes */ 4454fa45f116SDavidlohr Bueso if (fatal_signal_pending(current)) { 445502057967SDavid Rientjes remainder = 0; 445602057967SDavid Rientjes break; 445702057967SDavid Rientjes } 445802057967SDavid Rientjes 445902057967SDavid Rientjes /* 44604c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 44612a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 44624c887265SAdam Litke * first, for the page indexing below to work. 4463cb900f41SKirill A. Shutemov * 4464cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 44654c887265SAdam Litke */ 44667868a208SPunit Agrawal pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), 44677868a208SPunit Agrawal huge_page_size(h)); 4468cb900f41SKirill A. Shutemov if (pte) 4469cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 44702a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 447163551ae0SDavid Gibson 44722a15efc9SHugh Dickins /* 44732a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 44743ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 44753ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 44763ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 44773ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 44782a15efc9SHugh Dickins */ 44793ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 44803ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 4481cb900f41SKirill A. Shutemov if (pte) 4482cb900f41SKirill A. Shutemov spin_unlock(ptl); 44832a15efc9SHugh Dickins remainder = 0; 44842a15efc9SHugh Dickins break; 44852a15efc9SHugh Dickins } 44862a15efc9SHugh Dickins 44879cc3a5bdSNaoya Horiguchi /* 44889cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 44899cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 44909cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 44919cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 44929cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 44939cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 44949cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 44959cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 44969cc3a5bdSNaoya Horiguchi */ 44979cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 4498106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 4499106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 45002b740303SSouptick Joarder vm_fault_t ret; 450187ffc118SAndrea Arcangeli unsigned int fault_flags = 0; 45024c887265SAdam Litke 4503cb900f41SKirill A. Shutemov if (pte) 4504cb900f41SKirill A. Shutemov spin_unlock(ptl); 450587ffc118SAndrea Arcangeli if (flags & FOLL_WRITE) 450687ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_WRITE; 45074f6da934SPeter Xu if (locked) 450871335f37SPeter Xu fault_flags |= FAULT_FLAG_ALLOW_RETRY | 450971335f37SPeter Xu FAULT_FLAG_KILLABLE; 451087ffc118SAndrea Arcangeli if (flags & FOLL_NOWAIT) 451187ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY | 451287ffc118SAndrea Arcangeli FAULT_FLAG_RETRY_NOWAIT; 451387ffc118SAndrea Arcangeli if (flags & FOLL_TRIED) { 45144426e945SPeter Xu /* 45154426e945SPeter Xu * Note: FAULT_FLAG_ALLOW_RETRY and 45164426e945SPeter Xu * FAULT_FLAG_TRIED can co-exist 45174426e945SPeter Xu */ 451887ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_TRIED; 451987ffc118SAndrea Arcangeli } 452087ffc118SAndrea Arcangeli ret = hugetlb_fault(mm, vma, vaddr, fault_flags); 452187ffc118SAndrea Arcangeli if (ret & VM_FAULT_ERROR) { 45222be7cfedSDaniel Jordan err = vm_fault_to_errno(ret, flags); 45231c59827dSHugh Dickins remainder = 0; 45241c59827dSHugh Dickins break; 45251c59827dSHugh Dickins } 452687ffc118SAndrea Arcangeli if (ret & VM_FAULT_RETRY) { 45274f6da934SPeter Xu if (locked && 45281ac25013SAndrea Arcangeli !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) 45294f6da934SPeter Xu *locked = 0; 453087ffc118SAndrea Arcangeli *nr_pages = 0; 453187ffc118SAndrea Arcangeli /* 453287ffc118SAndrea Arcangeli * VM_FAULT_RETRY must not return an 453387ffc118SAndrea Arcangeli * error, it will return zero 453487ffc118SAndrea Arcangeli * instead. 453587ffc118SAndrea Arcangeli * 453687ffc118SAndrea Arcangeli * No need to update "position" as the 453787ffc118SAndrea Arcangeli * caller will not check it after 453887ffc118SAndrea Arcangeli * *nr_pages is set to 0. 453987ffc118SAndrea Arcangeli */ 454087ffc118SAndrea Arcangeli return i; 454187ffc118SAndrea Arcangeli } 454287ffc118SAndrea Arcangeli continue; 454387ffc118SAndrea Arcangeli } 454463551ae0SDavid Gibson 4545a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 45467f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 45478fde12caSLinus Torvalds 45488fde12caSLinus Torvalds /* 4549acbfb087SZhigang Lu * If subpage information not requested, update counters 4550acbfb087SZhigang Lu * and skip the same_page loop below. 4551acbfb087SZhigang Lu */ 4552acbfb087SZhigang Lu if (!pages && !vmas && !pfn_offset && 4553acbfb087SZhigang Lu (vaddr + huge_page_size(h) < vma->vm_end) && 4554acbfb087SZhigang Lu (remainder >= pages_per_huge_page(h))) { 4555acbfb087SZhigang Lu vaddr += huge_page_size(h); 4556acbfb087SZhigang Lu remainder -= pages_per_huge_page(h); 4557acbfb087SZhigang Lu i += pages_per_huge_page(h); 4558acbfb087SZhigang Lu spin_unlock(ptl); 4559acbfb087SZhigang Lu continue; 4560acbfb087SZhigang Lu } 4561acbfb087SZhigang Lu 4562d5d4b0aaSChen, Kenneth W same_page: 4563d6692183SChen, Kenneth W if (pages) { 456469d177c2SAndy Whitcroft pages[i] = mem_map_offset(page, pfn_offset); 45653faa52c0SJohn Hubbard /* 45663faa52c0SJohn Hubbard * try_grab_page() should always succeed here, because: 45673faa52c0SJohn Hubbard * a) we hold the ptl lock, and b) we've just checked 45683faa52c0SJohn Hubbard * that the huge page is present in the page tables. If 45693faa52c0SJohn Hubbard * the huge page is present, then the tail pages must 45703faa52c0SJohn Hubbard * also be present. The ptl prevents the head page and 45713faa52c0SJohn Hubbard * tail pages from being rearranged in any way. So this 45723faa52c0SJohn Hubbard * page must be available at this point, unless the page 45733faa52c0SJohn Hubbard * refcount overflowed: 45743faa52c0SJohn Hubbard */ 45753faa52c0SJohn Hubbard if (WARN_ON_ONCE(!try_grab_page(pages[i], flags))) { 45763faa52c0SJohn Hubbard spin_unlock(ptl); 45773faa52c0SJohn Hubbard remainder = 0; 45783faa52c0SJohn Hubbard err = -ENOMEM; 45793faa52c0SJohn Hubbard break; 45803faa52c0SJohn Hubbard } 4581d6692183SChen, Kenneth W } 458263551ae0SDavid Gibson 458363551ae0SDavid Gibson if (vmas) 458463551ae0SDavid Gibson vmas[i] = vma; 458563551ae0SDavid Gibson 458663551ae0SDavid Gibson vaddr += PAGE_SIZE; 4587d5d4b0aaSChen, Kenneth W ++pfn_offset; 458863551ae0SDavid Gibson --remainder; 458963551ae0SDavid Gibson ++i; 4590d5d4b0aaSChen, Kenneth W if (vaddr < vma->vm_end && remainder && 4591a5516438SAndi Kleen pfn_offset < pages_per_huge_page(h)) { 4592d5d4b0aaSChen, Kenneth W /* 4593d5d4b0aaSChen, Kenneth W * We use pfn_offset to avoid touching the pageframes 4594d5d4b0aaSChen, Kenneth W * of this compound page. 4595d5d4b0aaSChen, Kenneth W */ 4596d5d4b0aaSChen, Kenneth W goto same_page; 4597d5d4b0aaSChen, Kenneth W } 4598cb900f41SKirill A. Shutemov spin_unlock(ptl); 459963551ae0SDavid Gibson } 460028a35716SMichel Lespinasse *nr_pages = remainder; 460187ffc118SAndrea Arcangeli /* 460287ffc118SAndrea Arcangeli * setting position is actually required only if remainder is 460387ffc118SAndrea Arcangeli * not zero but it's faster not to add a "if (remainder)" 460487ffc118SAndrea Arcangeli * branch. 460587ffc118SAndrea Arcangeli */ 460663551ae0SDavid Gibson *position = vaddr; 460763551ae0SDavid Gibson 46082be7cfedSDaniel Jordan return i ? i : err; 460963551ae0SDavid Gibson } 46108f860591SZhang, Yanmin 46115491ae7bSAneesh Kumar K.V #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE 46125491ae7bSAneesh Kumar K.V /* 46135491ae7bSAneesh Kumar K.V * ARCHes with special requirements for evicting HUGETLB backing TLB entries can 46145491ae7bSAneesh Kumar K.V * implement this. 46155491ae7bSAneesh Kumar K.V */ 46165491ae7bSAneesh Kumar K.V #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) 46175491ae7bSAneesh Kumar K.V #endif 46185491ae7bSAneesh Kumar K.V 46197da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 46208f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 46218f860591SZhang, Yanmin { 46228f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 46238f860591SZhang, Yanmin unsigned long start = address; 46248f860591SZhang, Yanmin pte_t *ptep; 46258f860591SZhang, Yanmin pte_t pte; 4626a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 46277da4d641SPeter Zijlstra unsigned long pages = 0; 4628dff11abeSMike Kravetz bool shared_pmd = false; 4629ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 4630dff11abeSMike Kravetz 4631dff11abeSMike Kravetz /* 4632dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 4633ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 4634dff11abeSMike Kravetz * range if PMD sharing is possible. 4635dff11abeSMike Kravetz */ 46367269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 46377269f999SJérôme Glisse 0, vma, mm, start, end); 4638ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 46398f860591SZhang, Yanmin 46408f860591SZhang, Yanmin BUG_ON(address >= end); 4641ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 46428f860591SZhang, Yanmin 4643ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 464483cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 4645a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 4646cb900f41SKirill A. Shutemov spinlock_t *ptl; 46477868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, huge_page_size(h)); 46488f860591SZhang, Yanmin if (!ptep) 46498f860591SZhang, Yanmin continue; 4650cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 46517da4d641SPeter Zijlstra if (huge_pmd_unshare(mm, &address, ptep)) { 46527da4d641SPeter Zijlstra pages++; 4653cb900f41SKirill A. Shutemov spin_unlock(ptl); 4654dff11abeSMike Kravetz shared_pmd = true; 465539dde65cSChen, Kenneth W continue; 46567da4d641SPeter Zijlstra } 4657a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 4658a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 4659a8bda28dSNaoya Horiguchi spin_unlock(ptl); 4660a8bda28dSNaoya Horiguchi continue; 4661a8bda28dSNaoya Horiguchi } 4662a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 4663a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 4664a8bda28dSNaoya Horiguchi 4665a8bda28dSNaoya Horiguchi if (is_write_migration_entry(entry)) { 4666a8bda28dSNaoya Horiguchi pte_t newpte; 4667a8bda28dSNaoya Horiguchi 4668a8bda28dSNaoya Horiguchi make_migration_entry_read(&entry); 4669a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 4670e5251fd4SPunit Agrawal set_huge_swap_pte_at(mm, address, ptep, 4671e5251fd4SPunit Agrawal newpte, huge_page_size(h)); 4672a8bda28dSNaoya Horiguchi pages++; 4673a8bda28dSNaoya Horiguchi } 4674a8bda28dSNaoya Horiguchi spin_unlock(ptl); 4675a8bda28dSNaoya Horiguchi continue; 4676a8bda28dSNaoya Horiguchi } 4677a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 4678023bdd00SAneesh Kumar K.V pte_t old_pte; 4679023bdd00SAneesh Kumar K.V 4680023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 4681023bdd00SAneesh Kumar K.V pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); 4682be7517d6STony Lu pte = arch_make_huge_pte(pte, vma, NULL, 0); 4683023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 46847da4d641SPeter Zijlstra pages++; 46858f860591SZhang, Yanmin } 4686cb900f41SKirill A. Shutemov spin_unlock(ptl); 46878f860591SZhang, Yanmin } 4688d833352aSMel Gorman /* 4689c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 4690d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 4691c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 4692dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 4693dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 4694d833352aSMel Gorman */ 4695dff11abeSMike Kravetz if (shared_pmd) 4696ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 4697dff11abeSMike Kravetz else 46985491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 46990f10851eSJérôme Glisse /* 47000f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 47010f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 47020f10851eSJérôme Glisse * 4703ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 47040f10851eSJérôme Glisse */ 470583cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 4706ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 47077da4d641SPeter Zijlstra 47087da4d641SPeter Zijlstra return pages << h->order; 47098f860591SZhang, Yanmin } 47108f860591SZhang, Yanmin 4711a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 4712a1e78772SMel Gorman long from, long to, 47135a6fe125SMel Gorman struct vm_area_struct *vma, 4714ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 4715e4e574b7SAdam Litke { 471617c9d12eSMel Gorman long ret, chg; 4717a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 471890481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 47199119a41eSJoonsoo Kim struct resv_map *resv_map; 4720e9fe92aeSMina Almasry struct hugetlb_cgroup *h_cg; 47211c5ecae3SMike Kravetz long gbl_reserve; 4722e4e574b7SAdam Litke 472363489f8eSMike Kravetz /* This should never happen */ 472463489f8eSMike Kravetz if (from > to) { 472563489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 472663489f8eSMike Kravetz return -EINVAL; 472763489f8eSMike Kravetz } 472863489f8eSMike Kravetz 4729a1e78772SMel Gorman /* 473017c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 473117c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 473290481622SDavid Gibson * without using reserves 473317c9d12eSMel Gorman */ 4734ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 473517c9d12eSMel Gorman return 0; 473617c9d12eSMel Gorman 473717c9d12eSMel Gorman /* 4738a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 4739a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 4740a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 4741a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 4742a1e78772SMel Gorman */ 47439119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 4744f27a5136SMike Kravetz /* 4745f27a5136SMike Kravetz * resv_map can not be NULL as hugetlb_reserve_pages is only 4746f27a5136SMike Kravetz * called for inodes for which resv_maps were created (see 4747f27a5136SMike Kravetz * hugetlbfs_get_inode). 4748f27a5136SMike Kravetz */ 47494e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 47509119a41eSJoonsoo Kim 47511406ec9bSJoonsoo Kim chg = region_chg(resv_map, from, to); 47529119a41eSJoonsoo Kim 47539119a41eSJoonsoo Kim } else { 4754e9fe92aeSMina Almasry /* Private mapping. */ 47559119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 47565a6fe125SMel Gorman if (!resv_map) 47575a6fe125SMel Gorman return -ENOMEM; 47585a6fe125SMel Gorman 475917c9d12eSMel Gorman chg = to - from; 476017c9d12eSMel Gorman 4761e9fe92aeSMina Almasry if (hugetlb_cgroup_charge_cgroup_rsvd( 4762e9fe92aeSMina Almasry hstate_index(h), chg * pages_per_huge_page(h), 4763e9fe92aeSMina Almasry &h_cg)) { 4764e9fe92aeSMina Almasry kref_put(&resv_map->refs, resv_map_release); 4765e9fe92aeSMina Almasry return -ENOMEM; 4766e9fe92aeSMina Almasry } 4767e9fe92aeSMina Almasry 4768e9fe92aeSMina Almasry /* 4769e9fe92aeSMina Almasry * Since this branch handles private mappings, we attach the 4770e9fe92aeSMina Almasry * counter to uncharge for this reservation off resv_map. 4771e9fe92aeSMina Almasry */ 4772e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); 4773e9fe92aeSMina Almasry 47745a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 47755a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 47765a6fe125SMel Gorman } 47775a6fe125SMel Gorman 4778c50ac050SDave Hansen if (chg < 0) { 4779c50ac050SDave Hansen ret = chg; 4780c50ac050SDave Hansen goto out_err; 4781c50ac050SDave Hansen } 478217c9d12eSMel Gorman 47831c5ecae3SMike Kravetz /* 47841c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 47851c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 47861c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 47871c5ecae3SMike Kravetz */ 47881c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 47891c5ecae3SMike Kravetz if (gbl_reserve < 0) { 4790c50ac050SDave Hansen ret = -ENOSPC; 4791c50ac050SDave Hansen goto out_err; 4792c50ac050SDave Hansen } 479317c9d12eSMel Gorman 479417c9d12eSMel Gorman /* 479517c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 479690481622SDavid Gibson * Hand the pages back to the subpool if there are not 479717c9d12eSMel Gorman */ 47981c5ecae3SMike Kravetz ret = hugetlb_acct_memory(h, gbl_reserve); 479917c9d12eSMel Gorman if (ret < 0) { 48001c5ecae3SMike Kravetz /* put back original number of pages, chg */ 48011c5ecae3SMike Kravetz (void)hugepage_subpool_put_pages(spool, chg); 4802c50ac050SDave Hansen goto out_err; 480317c9d12eSMel Gorman } 480417c9d12eSMel Gorman 480517c9d12eSMel Gorman /* 480617c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 480717c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 480817c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 480917c9d12eSMel Gorman * the reservation was and the page cache tells how much of 481017c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 481117c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 481217c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 481317c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 481417c9d12eSMel Gorman * else has to be done for private mappings here 481517c9d12eSMel Gorman */ 481633039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 481733039678SMike Kravetz long add = region_add(resv_map, from, to); 481833039678SMike Kravetz 481933039678SMike Kravetz if (unlikely(chg > add)) { 482033039678SMike Kravetz /* 482133039678SMike Kravetz * pages in this range were added to the reserve 482233039678SMike Kravetz * map between region_chg and region_add. This 482333039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 482433039678SMike Kravetz * the subpool and reserve counts modified above 482533039678SMike Kravetz * based on the difference. 482633039678SMike Kravetz */ 482733039678SMike Kravetz long rsv_adjust; 482833039678SMike Kravetz 482933039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 483033039678SMike Kravetz chg - add); 483133039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 483233039678SMike Kravetz } 483333039678SMike Kravetz } 4834a43a8c39SChen, Kenneth W return 0; 4835c50ac050SDave Hansen out_err: 48365e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 4837ff8c0c53SMike Kravetz /* Don't call region_abort if region_chg failed */ 4838ff8c0c53SMike Kravetz if (chg >= 0) 48395e911373SMike Kravetz region_abort(resv_map, from, to); 4840f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 4841f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 4842c50ac050SDave Hansen return ret; 4843a43a8c39SChen, Kenneth W } 4844a43a8c39SChen, Kenneth W 4845b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 4846b5cec28dSMike Kravetz long freed) 4847a43a8c39SChen, Kenneth W { 4848a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 48494e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 48509119a41eSJoonsoo Kim long chg = 0; 485190481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 48521c5ecae3SMike Kravetz long gbl_reserve; 485345c682a6SKen Chen 4854f27a5136SMike Kravetz /* 4855f27a5136SMike Kravetz * Since this routine can be called in the evict inode path for all 4856f27a5136SMike Kravetz * hugetlbfs inodes, resv_map could be NULL. 4857f27a5136SMike Kravetz */ 4858b5cec28dSMike Kravetz if (resv_map) { 4859b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 4860b5cec28dSMike Kravetz /* 4861b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 4862b5cec28dSMike Kravetz * must be split and another region descriptor can not be 4863b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 4864b5cec28dSMike Kravetz */ 4865b5cec28dSMike Kravetz if (chg < 0) 4866b5cec28dSMike Kravetz return chg; 4867b5cec28dSMike Kravetz } 4868b5cec28dSMike Kravetz 486945c682a6SKen Chen spin_lock(&inode->i_lock); 4870e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 487145c682a6SKen Chen spin_unlock(&inode->i_lock); 487245c682a6SKen Chen 48731c5ecae3SMike Kravetz /* 48741c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 48751c5ecae3SMike Kravetz * reservations to be released may be adjusted. 48761c5ecae3SMike Kravetz */ 48771c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 48781c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 4879b5cec28dSMike Kravetz 4880b5cec28dSMike Kravetz return 0; 4881a43a8c39SChen, Kenneth W } 488293f70f90SNaoya Horiguchi 48833212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 48843212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 48853212b535SSteve Capper struct vm_area_struct *vma, 48863212b535SSteve Capper unsigned long addr, pgoff_t idx) 48873212b535SSteve Capper { 48883212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 48893212b535SSteve Capper svma->vm_start; 48903212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 48913212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 48923212b535SSteve Capper 48933212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 4894de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 4895de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 48963212b535SSteve Capper 48973212b535SSteve Capper /* 48983212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 48993212b535SSteve Capper * page table page. 49003212b535SSteve Capper */ 49013212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 49023212b535SSteve Capper vm_flags != svm_flags || 49033212b535SSteve Capper sbase < svma->vm_start || svma->vm_end < s_end) 49043212b535SSteve Capper return 0; 49053212b535SSteve Capper 49063212b535SSteve Capper return saddr; 49073212b535SSteve Capper } 49083212b535SSteve Capper 490931aafb45SNicholas Krause static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) 49103212b535SSteve Capper { 49113212b535SSteve Capper unsigned long base = addr & PUD_MASK; 49123212b535SSteve Capper unsigned long end = base + PUD_SIZE; 49133212b535SSteve Capper 49143212b535SSteve Capper /* 49153212b535SSteve Capper * check on proper vm_flags and page table alignment 49163212b535SSteve Capper */ 4917017b1660SMike Kravetz if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) 491831aafb45SNicholas Krause return true; 491931aafb45SNicholas Krause return false; 49203212b535SSteve Capper } 49213212b535SSteve Capper 49223212b535SSteve Capper /* 4923017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 4924017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 4925017b1660SMike Kravetz * shared pmd mappings. 4926017b1660SMike Kravetz */ 4927017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 4928017b1660SMike Kravetz unsigned long *start, unsigned long *end) 4929017b1660SMike Kravetz { 4930017b1660SMike Kravetz unsigned long check_addr = *start; 4931017b1660SMike Kravetz 4932017b1660SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 4933017b1660SMike Kravetz return; 4934017b1660SMike Kravetz 4935017b1660SMike Kravetz for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) { 4936017b1660SMike Kravetz unsigned long a_start = check_addr & PUD_MASK; 4937017b1660SMike Kravetz unsigned long a_end = a_start + PUD_SIZE; 4938017b1660SMike Kravetz 4939017b1660SMike Kravetz /* 4940017b1660SMike Kravetz * If sharing is possible, adjust start/end if necessary. 4941017b1660SMike Kravetz */ 4942017b1660SMike Kravetz if (range_in_vma(vma, a_start, a_end)) { 4943017b1660SMike Kravetz if (a_start < *start) 4944017b1660SMike Kravetz *start = a_start; 4945017b1660SMike Kravetz if (a_end > *end) 4946017b1660SMike Kravetz *end = a_end; 4947017b1660SMike Kravetz } 4948017b1660SMike Kravetz } 4949017b1660SMike Kravetz } 4950017b1660SMike Kravetz 4951017b1660SMike Kravetz /* 49523212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 49533212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 49543212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 4955c0d0381aSMike Kravetz * code much cleaner. 4956c0d0381aSMike Kravetz * 4957c0d0381aSMike Kravetz * This routine must be called with i_mmap_rwsem held in at least read mode. 4958c0d0381aSMike Kravetz * For hugetlbfs, this prevents removal of any page table entries associated 4959c0d0381aSMike Kravetz * with the address space. This is important as we are setting up sharing 4960c0d0381aSMike Kravetz * based on existing page table entries (mappings). 49613212b535SSteve Capper */ 49623212b535SSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 49633212b535SSteve Capper { 49643212b535SSteve Capper struct vm_area_struct *vma = find_vma(mm, addr); 49653212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 49663212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 49673212b535SSteve Capper vma->vm_pgoff; 49683212b535SSteve Capper struct vm_area_struct *svma; 49693212b535SSteve Capper unsigned long saddr; 49703212b535SSteve Capper pte_t *spte = NULL; 49713212b535SSteve Capper pte_t *pte; 4972cb900f41SKirill A. Shutemov spinlock_t *ptl; 49733212b535SSteve Capper 49743212b535SSteve Capper if (!vma_shareable(vma, addr)) 49753212b535SSteve Capper return (pte_t *)pmd_alloc(mm, pud, addr); 49763212b535SSteve Capper 49773212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 49783212b535SSteve Capper if (svma == vma) 49793212b535SSteve Capper continue; 49803212b535SSteve Capper 49813212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 49823212b535SSteve Capper if (saddr) { 49837868a208SPunit Agrawal spte = huge_pte_offset(svma->vm_mm, saddr, 49847868a208SPunit Agrawal vma_mmu_pagesize(svma)); 49853212b535SSteve Capper if (spte) { 49863212b535SSteve Capper get_page(virt_to_page(spte)); 49873212b535SSteve Capper break; 49883212b535SSteve Capper } 49893212b535SSteve Capper } 49903212b535SSteve Capper } 49913212b535SSteve Capper 49923212b535SSteve Capper if (!spte) 49933212b535SSteve Capper goto out; 49943212b535SSteve Capper 49958bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 4996dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 49973212b535SSteve Capper pud_populate(mm, pud, 49983212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 4999c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 5000dc6c9a35SKirill A. Shutemov } else { 50013212b535SSteve Capper put_page(virt_to_page(spte)); 5002dc6c9a35SKirill A. Shutemov } 5003cb900f41SKirill A. Shutemov spin_unlock(ptl); 50043212b535SSteve Capper out: 50053212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 50063212b535SSteve Capper return pte; 50073212b535SSteve Capper } 50083212b535SSteve Capper 50093212b535SSteve Capper /* 50103212b535SSteve Capper * unmap huge page backed by shared pte. 50113212b535SSteve Capper * 50123212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 50133212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 50143212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 50153212b535SSteve Capper * 5016c0d0381aSMike Kravetz * Called with page table lock held and i_mmap_rwsem held in write mode. 50173212b535SSteve Capper * 50183212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 50193212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 50203212b535SSteve Capper */ 50213212b535SSteve Capper int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 50223212b535SSteve Capper { 50233212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 5024c2febafcSKirill A. Shutemov p4d_t *p4d = p4d_offset(pgd, *addr); 5025c2febafcSKirill A. Shutemov pud_t *pud = pud_offset(p4d, *addr); 50263212b535SSteve Capper 50273212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 50283212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 50293212b535SSteve Capper return 0; 50303212b535SSteve Capper 50313212b535SSteve Capper pud_clear(pud); 50323212b535SSteve Capper put_page(virt_to_page(ptep)); 5033dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 50343212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 50353212b535SSteve Capper return 1; 50363212b535SSteve Capper } 50379e5fc74cSSteve Capper #define want_pmd_share() (1) 50389e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 50399e5fc74cSSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 50409e5fc74cSSteve Capper { 50419e5fc74cSSteve Capper return NULL; 50429e5fc74cSSteve Capper } 5043e81f2d22SZhang Zhen 5044e81f2d22SZhang Zhen int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 5045e81f2d22SZhang Zhen { 5046e81f2d22SZhang Zhen return 0; 5047e81f2d22SZhang Zhen } 5048017b1660SMike Kravetz 5049017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 5050017b1660SMike Kravetz unsigned long *start, unsigned long *end) 5051017b1660SMike Kravetz { 5052017b1660SMike Kravetz } 50539e5fc74cSSteve Capper #define want_pmd_share() (0) 50543212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 50553212b535SSteve Capper 50569e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 50579e5fc74cSSteve Capper pte_t *huge_pte_alloc(struct mm_struct *mm, 50589e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 50599e5fc74cSSteve Capper { 50609e5fc74cSSteve Capper pgd_t *pgd; 5061c2febafcSKirill A. Shutemov p4d_t *p4d; 50629e5fc74cSSteve Capper pud_t *pud; 50639e5fc74cSSteve Capper pte_t *pte = NULL; 50649e5fc74cSSteve Capper 50659e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 5066f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 5067f4f0a3d8SKirill A. Shutemov if (!p4d) 5068f4f0a3d8SKirill A. Shutemov return NULL; 5069c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 50709e5fc74cSSteve Capper if (pud) { 50719e5fc74cSSteve Capper if (sz == PUD_SIZE) { 50729e5fc74cSSteve Capper pte = (pte_t *)pud; 50739e5fc74cSSteve Capper } else { 50749e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 50759e5fc74cSSteve Capper if (want_pmd_share() && pud_none(*pud)) 50769e5fc74cSSteve Capper pte = huge_pmd_share(mm, addr, pud); 50779e5fc74cSSteve Capper else 50789e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 50799e5fc74cSSteve Capper } 50809e5fc74cSSteve Capper } 50814e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 50829e5fc74cSSteve Capper 50839e5fc74cSSteve Capper return pte; 50849e5fc74cSSteve Capper } 50859e5fc74cSSteve Capper 50869b19df29SPunit Agrawal /* 50879b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 50889b19df29SPunit Agrawal * entry at address @addr 50899b19df29SPunit Agrawal * 50909b19df29SPunit Agrawal * Return: Pointer to page table or swap entry (PUD or PMD) for 50919b19df29SPunit Agrawal * address @addr, or NULL if a p*d_none() entry is encountered and the 50929b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 50939b19df29SPunit Agrawal * table. 50949b19df29SPunit Agrawal */ 50957868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 50967868a208SPunit Agrawal unsigned long addr, unsigned long sz) 50979e5fc74cSSteve Capper { 50989e5fc74cSSteve Capper pgd_t *pgd; 5099c2febafcSKirill A. Shutemov p4d_t *p4d; 51009e5fc74cSSteve Capper pud_t *pud; 5101c2febafcSKirill A. Shutemov pmd_t *pmd; 51029e5fc74cSSteve Capper 51039e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 5104c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 5105c2febafcSKirill A. Shutemov return NULL; 5106c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 5107c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 5108c2febafcSKirill A. Shutemov return NULL; 51099b19df29SPunit Agrawal 5110c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 51119b19df29SPunit Agrawal if (sz != PUD_SIZE && pud_none(*pud)) 5112c2febafcSKirill A. Shutemov return NULL; 51139b19df29SPunit Agrawal /* hugepage or swap? */ 51149b19df29SPunit Agrawal if (pud_huge(*pud) || !pud_present(*pud)) 51159e5fc74cSSteve Capper return (pte_t *)pud; 51169b19df29SPunit Agrawal 51179e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 51189b19df29SPunit Agrawal if (sz != PMD_SIZE && pmd_none(*pmd)) 51199b19df29SPunit Agrawal return NULL; 51209b19df29SPunit Agrawal /* hugepage or swap? */ 51219b19df29SPunit Agrawal if (pmd_huge(*pmd) || !pmd_present(*pmd)) 51229e5fc74cSSteve Capper return (pte_t *)pmd; 51239b19df29SPunit Agrawal 51249b19df29SPunit Agrawal return NULL; 51259e5fc74cSSteve Capper } 51269e5fc74cSSteve Capper 512761f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 512861f77edaSNaoya Horiguchi 512961f77edaSNaoya Horiguchi /* 513061f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 513161f77edaSNaoya Horiguchi * behavior. 513261f77edaSNaoya Horiguchi */ 513361f77edaSNaoya Horiguchi struct page * __weak 513461f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 513561f77edaSNaoya Horiguchi int write) 513661f77edaSNaoya Horiguchi { 513761f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 513861f77edaSNaoya Horiguchi } 513961f77edaSNaoya Horiguchi 514061f77edaSNaoya Horiguchi struct page * __weak 51414dc71451SAneesh Kumar K.V follow_huge_pd(struct vm_area_struct *vma, 51424dc71451SAneesh Kumar K.V unsigned long address, hugepd_t hpd, int flags, int pdshift) 51434dc71451SAneesh Kumar K.V { 51444dc71451SAneesh Kumar K.V WARN(1, "hugepd follow called with no support for hugepage directory format\n"); 51454dc71451SAneesh Kumar K.V return NULL; 51464dc71451SAneesh Kumar K.V } 51474dc71451SAneesh Kumar K.V 51484dc71451SAneesh Kumar K.V struct page * __weak 51499e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 5150e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 51519e5fc74cSSteve Capper { 5152e66f17ffSNaoya Horiguchi struct page *page = NULL; 5153e66f17ffSNaoya Horiguchi spinlock_t *ptl; 5154c9d398faSNaoya Horiguchi pte_t pte; 51553faa52c0SJohn Hubbard 51563faa52c0SJohn Hubbard /* FOLL_GET and FOLL_PIN are mutually exclusive. */ 51573faa52c0SJohn Hubbard if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == 51583faa52c0SJohn Hubbard (FOLL_PIN | FOLL_GET))) 51593faa52c0SJohn Hubbard return NULL; 51603faa52c0SJohn Hubbard 5161e66f17ffSNaoya Horiguchi retry: 5162e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 5163e66f17ffSNaoya Horiguchi spin_lock(ptl); 5164e66f17ffSNaoya Horiguchi /* 5165e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 5166e66f17ffSNaoya Horiguchi * unmapped from other threads. 5167e66f17ffSNaoya Horiguchi */ 5168e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 5169e66f17ffSNaoya Horiguchi goto out; 5170c9d398faSNaoya Horiguchi pte = huge_ptep_get((pte_t *)pmd); 5171c9d398faSNaoya Horiguchi if (pte_present(pte)) { 517297534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 51733faa52c0SJohn Hubbard /* 51743faa52c0SJohn Hubbard * try_grab_page() should always succeed here, because: a) we 51753faa52c0SJohn Hubbard * hold the pmd (ptl) lock, and b) we've just checked that the 51763faa52c0SJohn Hubbard * huge pmd (head) page is present in the page tables. The ptl 51773faa52c0SJohn Hubbard * prevents the head page and tail pages from being rearranged 51783faa52c0SJohn Hubbard * in any way. So this page must be available at this point, 51793faa52c0SJohn Hubbard * unless the page refcount overflowed: 51803faa52c0SJohn Hubbard */ 51813faa52c0SJohn Hubbard if (WARN_ON_ONCE(!try_grab_page(page, flags))) { 51823faa52c0SJohn Hubbard page = NULL; 51833faa52c0SJohn Hubbard goto out; 51843faa52c0SJohn Hubbard } 5185e66f17ffSNaoya Horiguchi } else { 5186c9d398faSNaoya Horiguchi if (is_hugetlb_entry_migration(pte)) { 5187e66f17ffSNaoya Horiguchi spin_unlock(ptl); 5188e66f17ffSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pmd, ptl); 5189e66f17ffSNaoya Horiguchi goto retry; 5190e66f17ffSNaoya Horiguchi } 5191e66f17ffSNaoya Horiguchi /* 5192e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 5193e66f17ffSNaoya Horiguchi * follow_page_mask(). 5194e66f17ffSNaoya Horiguchi */ 5195e66f17ffSNaoya Horiguchi } 5196e66f17ffSNaoya Horiguchi out: 5197e66f17ffSNaoya Horiguchi spin_unlock(ptl); 51989e5fc74cSSteve Capper return page; 51999e5fc74cSSteve Capper } 52009e5fc74cSSteve Capper 520161f77edaSNaoya Horiguchi struct page * __weak 52029e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 5203e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 52049e5fc74cSSteve Capper { 52053faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 5206e66f17ffSNaoya Horiguchi return NULL; 52079e5fc74cSSteve Capper 5208e66f17ffSNaoya Horiguchi return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 52099e5fc74cSSteve Capper } 52109e5fc74cSSteve Capper 5211faaa5b62SAnshuman Khandual struct page * __weak 5212faaa5b62SAnshuman Khandual follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) 5213faaa5b62SAnshuman Khandual { 52143faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 5215faaa5b62SAnshuman Khandual return NULL; 5216faaa5b62SAnshuman Khandual 5217faaa5b62SAnshuman Khandual return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); 5218faaa5b62SAnshuman Khandual } 5219faaa5b62SAnshuman Khandual 522031caf665SNaoya Horiguchi bool isolate_huge_page(struct page *page, struct list_head *list) 522131caf665SNaoya Horiguchi { 5222bcc54222SNaoya Horiguchi bool ret = true; 5223bcc54222SNaoya Horiguchi 5224309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 522531caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 5226bcc54222SNaoya Horiguchi if (!page_huge_active(page) || !get_page_unless_zero(page)) { 5227bcc54222SNaoya Horiguchi ret = false; 5228bcc54222SNaoya Horiguchi goto unlock; 5229bcc54222SNaoya Horiguchi } 5230bcc54222SNaoya Horiguchi clear_page_huge_active(page); 523131caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 5232bcc54222SNaoya Horiguchi unlock: 523331caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 5234bcc54222SNaoya Horiguchi return ret; 523531caf665SNaoya Horiguchi } 523631caf665SNaoya Horiguchi 523731caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 523831caf665SNaoya Horiguchi { 5239309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 524031caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 5241bcc54222SNaoya Horiguchi set_page_huge_active(page); 524231caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 524331caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 524431caf665SNaoya Horiguchi put_page(page); 524531caf665SNaoya Horiguchi } 5246ab5ac90aSMichal Hocko 5247ab5ac90aSMichal Hocko void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) 5248ab5ac90aSMichal Hocko { 5249ab5ac90aSMichal Hocko struct hstate *h = page_hstate(oldpage); 5250ab5ac90aSMichal Hocko 5251ab5ac90aSMichal Hocko hugetlb_cgroup_migrate(oldpage, newpage); 5252ab5ac90aSMichal Hocko set_page_owner_migrate_reason(newpage, reason); 5253ab5ac90aSMichal Hocko 5254ab5ac90aSMichal Hocko /* 5255ab5ac90aSMichal Hocko * transfer temporary state of the new huge page. This is 5256ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 5257ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 5258ab5ac90aSMichal Hocko * the temporary status. 5259ab5ac90aSMichal Hocko * 5260ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 5261ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 5262ab5ac90aSMichal Hocko * the per-node's. 5263ab5ac90aSMichal Hocko */ 5264ab5ac90aSMichal Hocko if (PageHugeTemporary(newpage)) { 5265ab5ac90aSMichal Hocko int old_nid = page_to_nid(oldpage); 5266ab5ac90aSMichal Hocko int new_nid = page_to_nid(newpage); 5267ab5ac90aSMichal Hocko 5268ab5ac90aSMichal Hocko SetPageHugeTemporary(oldpage); 5269ab5ac90aSMichal Hocko ClearPageHugeTemporary(newpage); 5270ab5ac90aSMichal Hocko 5271ab5ac90aSMichal Hocko spin_lock(&hugetlb_lock); 5272ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 5273ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 5274ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 5275ab5ac90aSMichal Hocko } 5276ab5ac90aSMichal Hocko spin_unlock(&hugetlb_lock); 5277ab5ac90aSMichal Hocko } 5278ab5ac90aSMichal Hocko } 5279