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> 30d6606683SLinus Torvalds 3163551ae0SDavid Gibson #include <asm/page.h> 3263551ae0SDavid Gibson #include <asm/pgtable.h> 3324669e58SAneesh Kumar K.V #include <asm/tlb.h> 3463551ae0SDavid Gibson 3524669e58SAneesh Kumar K.V #include <linux/io.h> 3663551ae0SDavid Gibson #include <linux/hugetlb.h> 379dd540e2SAneesh Kumar K.V #include <linux/hugetlb_cgroup.h> 389a305230SLee Schermerhorn #include <linux/node.h> 391a1aad8aSMike Kravetz #include <linux/userfaultfd_k.h> 40ab5ac90aSMichal Hocko #include <linux/page_owner.h> 417835e98bSNick Piggin #include "internal.h" 421da177e4SLinus Torvalds 43c3f38a38SAneesh Kumar K.V int hugetlb_max_hstate __read_mostly; 44e5ff2159SAndi Kleen unsigned int default_hstate_idx; 45e5ff2159SAndi Kleen struct hstate hstates[HUGE_MAX_HSTATE]; 46641844f5SNaoya Horiguchi /* 47641844f5SNaoya Horiguchi * Minimum page order among possible hugepage sizes, set to a proper value 48641844f5SNaoya Horiguchi * at boot time. 49641844f5SNaoya Horiguchi */ 50641844f5SNaoya Horiguchi static unsigned int minimum_order __read_mostly = UINT_MAX; 51e5ff2159SAndi Kleen 5253ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 5353ba51d2SJon Tollefson 54e5ff2159SAndi Kleen /* for command line parsing */ 55e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 56e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 57e11bfbfcSNick Piggin static unsigned long __initdata default_hstate_size; 589fee021dSVaishali Thakkar static bool __initdata parsed_valid_hugepagesz = true; 59e5ff2159SAndi Kleen 603935baa9SDavid Gibson /* 6131caf665SNaoya Horiguchi * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, 6231caf665SNaoya Horiguchi * free_huge_pages, and surplus_huge_pages. 633935baa9SDavid Gibson */ 64c3f38a38SAneesh Kumar K.V DEFINE_SPINLOCK(hugetlb_lock); 650bd0f9fbSEric Paris 668382d914SDavidlohr Bueso /* 678382d914SDavidlohr Bueso * Serializes faults on the same logical page. This is used to 688382d914SDavidlohr Bueso * prevent spurious OOMs when the hugepage pool is fully utilized. 698382d914SDavidlohr Bueso */ 708382d914SDavidlohr Bueso static int num_fault_mutexes; 71c672c7f2SMike Kravetz struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; 728382d914SDavidlohr Bueso 737ca02d0aSMike Kravetz /* Forward declaration */ 747ca02d0aSMike Kravetz static int hugetlb_acct_memory(struct hstate *h, long delta); 757ca02d0aSMike Kravetz 7690481622SDavid Gibson static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) 7790481622SDavid Gibson { 7890481622SDavid Gibson bool free = (spool->count == 0) && (spool->used_hpages == 0); 7990481622SDavid Gibson 8090481622SDavid Gibson spin_unlock(&spool->lock); 8190481622SDavid Gibson 8290481622SDavid Gibson /* If no pages are used, and no other handles to the subpool 837ca02d0aSMike Kravetz * remain, give up any reservations mased on minimum size and 847ca02d0aSMike Kravetz * free the subpool */ 857ca02d0aSMike Kravetz if (free) { 867ca02d0aSMike Kravetz if (spool->min_hpages != -1) 877ca02d0aSMike Kravetz hugetlb_acct_memory(spool->hstate, 887ca02d0aSMike Kravetz -spool->min_hpages); 8990481622SDavid Gibson kfree(spool); 9090481622SDavid Gibson } 917ca02d0aSMike Kravetz } 9290481622SDavid Gibson 937ca02d0aSMike Kravetz struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, 947ca02d0aSMike Kravetz long min_hpages) 9590481622SDavid Gibson { 9690481622SDavid Gibson struct hugepage_subpool *spool; 9790481622SDavid Gibson 98c6a91820SMike Kravetz spool = kzalloc(sizeof(*spool), GFP_KERNEL); 9990481622SDavid Gibson if (!spool) 10090481622SDavid Gibson return NULL; 10190481622SDavid Gibson 10290481622SDavid Gibson spin_lock_init(&spool->lock); 10390481622SDavid Gibson spool->count = 1; 1047ca02d0aSMike Kravetz spool->max_hpages = max_hpages; 1057ca02d0aSMike Kravetz spool->hstate = h; 1067ca02d0aSMike Kravetz spool->min_hpages = min_hpages; 1077ca02d0aSMike Kravetz 1087ca02d0aSMike Kravetz if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { 1097ca02d0aSMike Kravetz kfree(spool); 1107ca02d0aSMike Kravetz return NULL; 1117ca02d0aSMike Kravetz } 1127ca02d0aSMike Kravetz spool->rsv_hpages = min_hpages; 11390481622SDavid Gibson 11490481622SDavid Gibson return spool; 11590481622SDavid Gibson } 11690481622SDavid Gibson 11790481622SDavid Gibson void hugepage_put_subpool(struct hugepage_subpool *spool) 11890481622SDavid Gibson { 11990481622SDavid Gibson spin_lock(&spool->lock); 12090481622SDavid Gibson BUG_ON(!spool->count); 12190481622SDavid Gibson spool->count--; 12290481622SDavid Gibson unlock_or_release_subpool(spool); 12390481622SDavid Gibson } 12490481622SDavid Gibson 1251c5ecae3SMike Kravetz /* 1261c5ecae3SMike Kravetz * Subpool accounting for allocating and reserving pages. 1271c5ecae3SMike Kravetz * Return -ENOMEM if there are not enough resources to satisfy the 1281c5ecae3SMike Kravetz * the request. Otherwise, return the number of pages by which the 1291c5ecae3SMike Kravetz * global pools must be adjusted (upward). The returned value may 1301c5ecae3SMike Kravetz * only be different than the passed value (delta) in the case where 1311c5ecae3SMike Kravetz * a subpool minimum size must be manitained. 1321c5ecae3SMike Kravetz */ 1331c5ecae3SMike Kravetz static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, 13490481622SDavid Gibson long delta) 13590481622SDavid Gibson { 1361c5ecae3SMike Kravetz long ret = delta; 13790481622SDavid Gibson 13890481622SDavid Gibson if (!spool) 1391c5ecae3SMike Kravetz return ret; 14090481622SDavid Gibson 14190481622SDavid Gibson spin_lock(&spool->lock); 14290481622SDavid Gibson 1431c5ecae3SMike Kravetz if (spool->max_hpages != -1) { /* maximum size accounting */ 1441c5ecae3SMike Kravetz if ((spool->used_hpages + delta) <= spool->max_hpages) 1451c5ecae3SMike Kravetz spool->used_hpages += delta; 1461c5ecae3SMike Kravetz else { 1471c5ecae3SMike Kravetz ret = -ENOMEM; 1481c5ecae3SMike Kravetz goto unlock_ret; 1491c5ecae3SMike Kravetz } 1501c5ecae3SMike Kravetz } 1511c5ecae3SMike Kravetz 15209a95e29SMike Kravetz /* minimum size accounting */ 15309a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->rsv_hpages) { 1541c5ecae3SMike Kravetz if (delta > spool->rsv_hpages) { 1551c5ecae3SMike Kravetz /* 1561c5ecae3SMike Kravetz * Asking for more reserves than those already taken on 1571c5ecae3SMike Kravetz * behalf of subpool. Return difference. 1581c5ecae3SMike Kravetz */ 1591c5ecae3SMike Kravetz ret = delta - spool->rsv_hpages; 1601c5ecae3SMike Kravetz spool->rsv_hpages = 0; 1611c5ecae3SMike Kravetz } else { 1621c5ecae3SMike Kravetz ret = 0; /* reserves already accounted for */ 1631c5ecae3SMike Kravetz spool->rsv_hpages -= delta; 1641c5ecae3SMike Kravetz } 1651c5ecae3SMike Kravetz } 1661c5ecae3SMike Kravetz 1671c5ecae3SMike Kravetz unlock_ret: 1681c5ecae3SMike Kravetz spin_unlock(&spool->lock); 16990481622SDavid Gibson return ret; 17090481622SDavid Gibson } 17190481622SDavid Gibson 1721c5ecae3SMike Kravetz /* 1731c5ecae3SMike Kravetz * Subpool accounting for freeing and unreserving pages. 1741c5ecae3SMike Kravetz * Return the number of global page reservations that must be dropped. 1751c5ecae3SMike Kravetz * The return value may only be different than the passed value (delta) 1761c5ecae3SMike Kravetz * in the case where a subpool minimum size must be maintained. 1771c5ecae3SMike Kravetz */ 1781c5ecae3SMike Kravetz static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, 17990481622SDavid Gibson long delta) 18090481622SDavid Gibson { 1811c5ecae3SMike Kravetz long ret = delta; 1821c5ecae3SMike Kravetz 18390481622SDavid Gibson if (!spool) 1841c5ecae3SMike Kravetz return delta; 18590481622SDavid Gibson 18690481622SDavid Gibson spin_lock(&spool->lock); 1871c5ecae3SMike Kravetz 1881c5ecae3SMike Kravetz if (spool->max_hpages != -1) /* maximum size accounting */ 18990481622SDavid Gibson spool->used_hpages -= delta; 1901c5ecae3SMike Kravetz 19109a95e29SMike Kravetz /* minimum size accounting */ 19209a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { 1931c5ecae3SMike Kravetz if (spool->rsv_hpages + delta <= spool->min_hpages) 1941c5ecae3SMike Kravetz ret = 0; 1951c5ecae3SMike Kravetz else 1961c5ecae3SMike Kravetz ret = spool->rsv_hpages + delta - spool->min_hpages; 1971c5ecae3SMike Kravetz 1981c5ecae3SMike Kravetz spool->rsv_hpages += delta; 1991c5ecae3SMike Kravetz if (spool->rsv_hpages > spool->min_hpages) 2001c5ecae3SMike Kravetz spool->rsv_hpages = spool->min_hpages; 2011c5ecae3SMike Kravetz } 2021c5ecae3SMike Kravetz 2031c5ecae3SMike Kravetz /* 2041c5ecae3SMike Kravetz * If hugetlbfs_put_super couldn't free spool due to an outstanding 2051c5ecae3SMike Kravetz * quota reference, free it now. 2061c5ecae3SMike Kravetz */ 20790481622SDavid Gibson unlock_or_release_subpool(spool); 2081c5ecae3SMike Kravetz 2091c5ecae3SMike Kravetz return ret; 21090481622SDavid Gibson } 21190481622SDavid Gibson 21290481622SDavid Gibson static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 21390481622SDavid Gibson { 21490481622SDavid Gibson return HUGETLBFS_SB(inode->i_sb)->spool; 21590481622SDavid Gibson } 21690481622SDavid Gibson 21790481622SDavid Gibson static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 21890481622SDavid Gibson { 219496ad9aaSAl Viro return subpool_inode(file_inode(vma->vm_file)); 22090481622SDavid Gibson } 22190481622SDavid Gibson 222e7c4b0bfSAndy Whitcroft /* 22396822904SAndy Whitcroft * Region tracking -- allows tracking of reservations and instantiated pages 22496822904SAndy Whitcroft * across the pages in a mapping. 22584afd99bSAndy Whitcroft * 2261dd308a7SMike Kravetz * The region data structures are embedded into a resv_map and protected 2271dd308a7SMike Kravetz * by a resv_map's lock. The set of regions within the resv_map represent 2281dd308a7SMike Kravetz * reservations for huge pages, or huge pages that have already been 2291dd308a7SMike Kravetz * instantiated within the map. The from and to elements are huge page 2301dd308a7SMike Kravetz * indicies into the associated mapping. from indicates the starting index 2311dd308a7SMike Kravetz * of the region. to represents the first index past the end of the region. 2321dd308a7SMike Kravetz * 2331dd308a7SMike Kravetz * For example, a file region structure with from == 0 and to == 4 represents 2341dd308a7SMike Kravetz * four huge pages in a mapping. It is important to note that the to element 2351dd308a7SMike Kravetz * represents the first element past the end of the region. This is used in 2361dd308a7SMike Kravetz * arithmetic as 4(to) - 0(from) = 4 huge pages in the region. 2371dd308a7SMike Kravetz * 2381dd308a7SMike Kravetz * Interval notation of the form [from, to) will be used to indicate that 2391dd308a7SMike Kravetz * the endpoint from is inclusive and to is exclusive. 24096822904SAndy Whitcroft */ 24196822904SAndy Whitcroft struct file_region { 24296822904SAndy Whitcroft struct list_head link; 24396822904SAndy Whitcroft long from; 24496822904SAndy Whitcroft long to; 24596822904SAndy Whitcroft }; 24696822904SAndy Whitcroft 2471dd308a7SMike Kravetz /* 2481dd308a7SMike Kravetz * Add the huge page range represented by [f, t) to the reserve 2495e911373SMike Kravetz * map. In the normal case, existing regions will be expanded 2505e911373SMike Kravetz * to accommodate the specified range. Sufficient regions should 2515e911373SMike Kravetz * exist for expansion due to the previous call to region_chg 2525e911373SMike Kravetz * with the same range. However, it is possible that region_del 2535e911373SMike Kravetz * could have been called after region_chg and modifed the map 2545e911373SMike Kravetz * in such a way that no region exists to be expanded. In this 2555e911373SMike Kravetz * case, pull a region descriptor from the cache associated with 2565e911373SMike Kravetz * the map and use that for the new range. 257cf3ad20bSMike Kravetz * 258cf3ad20bSMike Kravetz * Return the number of new huge pages added to the map. This 259cf3ad20bSMike Kravetz * number is greater than or equal to zero. 2601dd308a7SMike Kravetz */ 2611406ec9bSJoonsoo Kim static long region_add(struct resv_map *resv, long f, long t) 26296822904SAndy Whitcroft { 2631406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 26496822904SAndy Whitcroft struct file_region *rg, *nrg, *trg; 265cf3ad20bSMike Kravetz long add = 0; 26696822904SAndy Whitcroft 2677b24d861SDavidlohr Bueso spin_lock(&resv->lock); 26896822904SAndy Whitcroft /* Locate the region we are either in or before. */ 26996822904SAndy Whitcroft list_for_each_entry(rg, head, link) 27096822904SAndy Whitcroft if (f <= rg->to) 27196822904SAndy Whitcroft break; 27296822904SAndy Whitcroft 2735e911373SMike Kravetz /* 2745e911373SMike Kravetz * If no region exists which can be expanded to include the 2755e911373SMike Kravetz * specified range, the list must have been modified by an 2765e911373SMike Kravetz * interleving call to region_del(). Pull a region descriptor 2775e911373SMike Kravetz * from the cache and use it for this range. 2785e911373SMike Kravetz */ 2795e911373SMike Kravetz if (&rg->link == head || t < rg->from) { 2805e911373SMike Kravetz VM_BUG_ON(resv->region_cache_count <= 0); 2815e911373SMike Kravetz 2825e911373SMike Kravetz resv->region_cache_count--; 2835e911373SMike Kravetz nrg = list_first_entry(&resv->region_cache, struct file_region, 2845e911373SMike Kravetz link); 2855e911373SMike Kravetz list_del(&nrg->link); 2865e911373SMike Kravetz 2875e911373SMike Kravetz nrg->from = f; 2885e911373SMike Kravetz nrg->to = t; 2895e911373SMike Kravetz list_add(&nrg->link, rg->link.prev); 2905e911373SMike Kravetz 2915e911373SMike Kravetz add += t - f; 2925e911373SMike Kravetz goto out_locked; 2935e911373SMike Kravetz } 2945e911373SMike Kravetz 29596822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 29696822904SAndy Whitcroft if (f > rg->from) 29796822904SAndy Whitcroft f = rg->from; 29896822904SAndy Whitcroft 29996822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 30096822904SAndy Whitcroft nrg = rg; 30196822904SAndy Whitcroft list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 30296822904SAndy Whitcroft if (&rg->link == head) 30396822904SAndy Whitcroft break; 30496822904SAndy Whitcroft if (rg->from > t) 30596822904SAndy Whitcroft break; 30696822904SAndy Whitcroft 30796822904SAndy Whitcroft /* If this area reaches higher then extend our area to 30896822904SAndy Whitcroft * include it completely. If this is not the first area 30996822904SAndy Whitcroft * which we intend to reuse, free it. */ 31096822904SAndy Whitcroft if (rg->to > t) 31196822904SAndy Whitcroft t = rg->to; 31296822904SAndy Whitcroft if (rg != nrg) { 313cf3ad20bSMike Kravetz /* Decrement return value by the deleted range. 314cf3ad20bSMike Kravetz * Another range will span this area so that by 315cf3ad20bSMike Kravetz * end of routine add will be >= zero 316cf3ad20bSMike Kravetz */ 317cf3ad20bSMike Kravetz add -= (rg->to - rg->from); 31896822904SAndy Whitcroft list_del(&rg->link); 31996822904SAndy Whitcroft kfree(rg); 32096822904SAndy Whitcroft } 32196822904SAndy Whitcroft } 322cf3ad20bSMike Kravetz 323cf3ad20bSMike Kravetz add += (nrg->from - f); /* Added to beginning of region */ 32496822904SAndy Whitcroft nrg->from = f; 325cf3ad20bSMike Kravetz add += t - nrg->to; /* Added to end of region */ 32696822904SAndy Whitcroft nrg->to = t; 327cf3ad20bSMike Kravetz 3285e911373SMike Kravetz out_locked: 3295e911373SMike Kravetz resv->adds_in_progress--; 3307b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 331cf3ad20bSMike Kravetz VM_BUG_ON(add < 0); 332cf3ad20bSMike Kravetz return add; 33396822904SAndy Whitcroft } 33496822904SAndy Whitcroft 3351dd308a7SMike Kravetz /* 3361dd308a7SMike Kravetz * Examine the existing reserve map and determine how many 3371dd308a7SMike Kravetz * huge pages in the specified range [f, t) are NOT currently 3381dd308a7SMike Kravetz * represented. This routine is called before a subsequent 3391dd308a7SMike Kravetz * call to region_add that will actually modify the reserve 3401dd308a7SMike Kravetz * map to add the specified range [f, t). region_chg does 3411dd308a7SMike Kravetz * not change the number of huge pages represented by the 3421dd308a7SMike Kravetz * map. However, if the existing regions in the map can not 3431dd308a7SMike Kravetz * be expanded to represent the new range, a new file_region 3441dd308a7SMike Kravetz * structure is added to the map as a placeholder. This is 3451dd308a7SMike Kravetz * so that the subsequent region_add call will have all the 3461dd308a7SMike Kravetz * regions it needs and will not fail. 3471dd308a7SMike Kravetz * 3485e911373SMike Kravetz * Upon entry, region_chg will also examine the cache of region descriptors 3495e911373SMike Kravetz * associated with the map. If there are not enough descriptors cached, one 3505e911373SMike Kravetz * will be allocated for the in progress add operation. 3515e911373SMike Kravetz * 3525e911373SMike Kravetz * Returns the number of huge pages that need to be added to the existing 3535e911373SMike Kravetz * reservation map for the range [f, t). This number is greater or equal to 3545e911373SMike Kravetz * zero. -ENOMEM is returned if a new file_region structure or cache entry 3555e911373SMike Kravetz * is needed and can not be allocated. 3561dd308a7SMike Kravetz */ 3571406ec9bSJoonsoo Kim static long region_chg(struct resv_map *resv, long f, long t) 35896822904SAndy Whitcroft { 3591406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 3607b24d861SDavidlohr Bueso struct file_region *rg, *nrg = NULL; 36196822904SAndy Whitcroft long chg = 0; 36296822904SAndy Whitcroft 3637b24d861SDavidlohr Bueso retry: 3647b24d861SDavidlohr Bueso spin_lock(&resv->lock); 3655e911373SMike Kravetz retry_locked: 3665e911373SMike Kravetz resv->adds_in_progress++; 3675e911373SMike Kravetz 3685e911373SMike Kravetz /* 3695e911373SMike Kravetz * Check for sufficient descriptors in the cache to accommodate 3705e911373SMike Kravetz * the number of in progress add operations. 3715e911373SMike Kravetz */ 3725e911373SMike Kravetz if (resv->adds_in_progress > resv->region_cache_count) { 3735e911373SMike Kravetz struct file_region *trg; 3745e911373SMike Kravetz 3755e911373SMike Kravetz VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1); 3765e911373SMike Kravetz /* Must drop lock to allocate a new descriptor. */ 3775e911373SMike Kravetz resv->adds_in_progress--; 3785e911373SMike Kravetz spin_unlock(&resv->lock); 3795e911373SMike Kravetz 3805e911373SMike Kravetz trg = kmalloc(sizeof(*trg), GFP_KERNEL); 381dbe409e4SMike Kravetz if (!trg) { 382dbe409e4SMike Kravetz kfree(nrg); 3835e911373SMike Kravetz return -ENOMEM; 384dbe409e4SMike Kravetz } 3855e911373SMike Kravetz 3865e911373SMike Kravetz spin_lock(&resv->lock); 3875e911373SMike Kravetz list_add(&trg->link, &resv->region_cache); 3885e911373SMike Kravetz resv->region_cache_count++; 3895e911373SMike Kravetz goto retry_locked; 3905e911373SMike Kravetz } 3915e911373SMike Kravetz 39296822904SAndy Whitcroft /* Locate the region we are before or in. */ 39396822904SAndy Whitcroft list_for_each_entry(rg, head, link) 39496822904SAndy Whitcroft if (f <= rg->to) 39596822904SAndy Whitcroft break; 39696822904SAndy Whitcroft 39796822904SAndy Whitcroft /* If we are below the current region then a new region is required. 39896822904SAndy Whitcroft * Subtle, allocate a new region at the position but make it zero 39996822904SAndy Whitcroft * size such that we can guarantee to record the reservation. */ 40096822904SAndy Whitcroft if (&rg->link == head || t < rg->from) { 4017b24d861SDavidlohr Bueso if (!nrg) { 4025e911373SMike Kravetz resv->adds_in_progress--; 4037b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 40496822904SAndy Whitcroft nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 40596822904SAndy Whitcroft if (!nrg) 40696822904SAndy Whitcroft return -ENOMEM; 4077b24d861SDavidlohr Bueso 40896822904SAndy Whitcroft nrg->from = f; 40996822904SAndy Whitcroft nrg->to = f; 41096822904SAndy Whitcroft INIT_LIST_HEAD(&nrg->link); 4117b24d861SDavidlohr Bueso goto retry; 4127b24d861SDavidlohr Bueso } 41396822904SAndy Whitcroft 4147b24d861SDavidlohr Bueso list_add(&nrg->link, rg->link.prev); 4157b24d861SDavidlohr Bueso chg = t - f; 4167b24d861SDavidlohr Bueso goto out_nrg; 41796822904SAndy Whitcroft } 41896822904SAndy Whitcroft 41996822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 42096822904SAndy Whitcroft if (f > rg->from) 42196822904SAndy Whitcroft f = rg->from; 42296822904SAndy Whitcroft chg = t - f; 42396822904SAndy Whitcroft 42496822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 42596822904SAndy Whitcroft list_for_each_entry(rg, rg->link.prev, link) { 42696822904SAndy Whitcroft if (&rg->link == head) 42796822904SAndy Whitcroft break; 42896822904SAndy Whitcroft if (rg->from > t) 4297b24d861SDavidlohr Bueso goto out; 43096822904SAndy Whitcroft 43125985edcSLucas De Marchi /* We overlap with this area, if it extends further than 43296822904SAndy Whitcroft * us then we must extend ourselves. Account for its 43396822904SAndy Whitcroft * existing reservation. */ 43496822904SAndy Whitcroft if (rg->to > t) { 43596822904SAndy Whitcroft chg += rg->to - t; 43696822904SAndy Whitcroft t = rg->to; 43796822904SAndy Whitcroft } 43896822904SAndy Whitcroft chg -= rg->to - rg->from; 43996822904SAndy Whitcroft } 4407b24d861SDavidlohr Bueso 4417b24d861SDavidlohr Bueso out: 4427b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 4437b24d861SDavidlohr Bueso /* We already know we raced and no longer need the new region */ 4447b24d861SDavidlohr Bueso kfree(nrg); 4457b24d861SDavidlohr Bueso return chg; 4467b24d861SDavidlohr Bueso out_nrg: 4477b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 44896822904SAndy Whitcroft return chg; 44996822904SAndy Whitcroft } 45096822904SAndy Whitcroft 4511dd308a7SMike Kravetz /* 4525e911373SMike Kravetz * Abort the in progress add operation. The adds_in_progress field 4535e911373SMike Kravetz * of the resv_map keeps track of the operations in progress between 4545e911373SMike Kravetz * calls to region_chg and region_add. Operations are sometimes 4555e911373SMike Kravetz * aborted after the call to region_chg. In such cases, region_abort 4565e911373SMike Kravetz * is called to decrement the adds_in_progress counter. 4575e911373SMike Kravetz * 4585e911373SMike Kravetz * NOTE: The range arguments [f, t) are not needed or used in this 4595e911373SMike Kravetz * routine. They are kept to make reading the calling code easier as 4605e911373SMike Kravetz * arguments will match the associated region_chg call. 4615e911373SMike Kravetz */ 4625e911373SMike Kravetz static void region_abort(struct resv_map *resv, long f, long t) 4635e911373SMike Kravetz { 4645e911373SMike Kravetz spin_lock(&resv->lock); 4655e911373SMike Kravetz VM_BUG_ON(!resv->region_cache_count); 4665e911373SMike Kravetz resv->adds_in_progress--; 4675e911373SMike Kravetz spin_unlock(&resv->lock); 4685e911373SMike Kravetz } 4695e911373SMike Kravetz 4705e911373SMike Kravetz /* 471feba16e2SMike Kravetz * Delete the specified range [f, t) from the reserve map. If the 472feba16e2SMike Kravetz * t parameter is LONG_MAX, this indicates that ALL regions after f 473feba16e2SMike Kravetz * should be deleted. Locate the regions which intersect [f, t) 474feba16e2SMike Kravetz * and either trim, delete or split the existing regions. 475feba16e2SMike Kravetz * 476feba16e2SMike Kravetz * Returns the number of huge pages deleted from the reserve map. 477feba16e2SMike Kravetz * In the normal case, the return value is zero or more. In the 478feba16e2SMike Kravetz * case where a region must be split, a new region descriptor must 479feba16e2SMike Kravetz * be allocated. If the allocation fails, -ENOMEM will be returned. 480feba16e2SMike Kravetz * NOTE: If the parameter t == LONG_MAX, then we will never split 481feba16e2SMike Kravetz * a region and possibly return -ENOMEM. Callers specifying 482feba16e2SMike Kravetz * t == LONG_MAX do not need to check for -ENOMEM error. 4831dd308a7SMike Kravetz */ 484feba16e2SMike Kravetz static long region_del(struct resv_map *resv, long f, long t) 48596822904SAndy Whitcroft { 4861406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 48796822904SAndy Whitcroft struct file_region *rg, *trg; 488feba16e2SMike Kravetz struct file_region *nrg = NULL; 489feba16e2SMike Kravetz long del = 0; 49096822904SAndy Whitcroft 491feba16e2SMike Kravetz retry: 4927b24d861SDavidlohr Bueso spin_lock(&resv->lock); 493feba16e2SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 494dbe409e4SMike Kravetz /* 495dbe409e4SMike Kravetz * Skip regions before the range to be deleted. file_region 496dbe409e4SMike Kravetz * ranges are normally of the form [from, to). However, there 497dbe409e4SMike Kravetz * may be a "placeholder" entry in the map which is of the form 498dbe409e4SMike Kravetz * (from, to) with from == to. Check for placeholder entries 499dbe409e4SMike Kravetz * at the beginning of the range to be deleted. 500dbe409e4SMike Kravetz */ 501dbe409e4SMike Kravetz if (rg->to <= f && (rg->to != rg->from || rg->to != f)) 502feba16e2SMike Kravetz continue; 503dbe409e4SMike Kravetz 504feba16e2SMike Kravetz if (rg->from >= t) 50596822904SAndy Whitcroft break; 50696822904SAndy Whitcroft 507feba16e2SMike Kravetz if (f > rg->from && t < rg->to) { /* Must split region */ 508feba16e2SMike Kravetz /* 509feba16e2SMike Kravetz * Check for an entry in the cache before dropping 510feba16e2SMike Kravetz * lock and attempting allocation. 511feba16e2SMike Kravetz */ 512feba16e2SMike Kravetz if (!nrg && 513feba16e2SMike Kravetz resv->region_cache_count > resv->adds_in_progress) { 514feba16e2SMike Kravetz nrg = list_first_entry(&resv->region_cache, 515feba16e2SMike Kravetz struct file_region, 516feba16e2SMike Kravetz link); 517feba16e2SMike Kravetz list_del(&nrg->link); 518feba16e2SMike Kravetz resv->region_cache_count--; 51996822904SAndy Whitcroft } 52096822904SAndy Whitcroft 521feba16e2SMike Kravetz if (!nrg) { 522feba16e2SMike Kravetz spin_unlock(&resv->lock); 523feba16e2SMike Kravetz nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 524feba16e2SMike Kravetz if (!nrg) 525feba16e2SMike Kravetz return -ENOMEM; 526feba16e2SMike Kravetz goto retry; 527feba16e2SMike Kravetz } 528feba16e2SMike Kravetz 529feba16e2SMike Kravetz del += t - f; 530feba16e2SMike Kravetz 531feba16e2SMike Kravetz /* New entry for end of split region */ 532feba16e2SMike Kravetz nrg->from = t; 533feba16e2SMike Kravetz nrg->to = rg->to; 534feba16e2SMike Kravetz INIT_LIST_HEAD(&nrg->link); 535feba16e2SMike Kravetz 536feba16e2SMike Kravetz /* Original entry is trimmed */ 537feba16e2SMike Kravetz rg->to = f; 538feba16e2SMike Kravetz 539feba16e2SMike Kravetz list_add(&nrg->link, &rg->link); 540feba16e2SMike Kravetz nrg = NULL; 54196822904SAndy Whitcroft break; 542feba16e2SMike Kravetz } 543feba16e2SMike Kravetz 544feba16e2SMike Kravetz if (f <= rg->from && t >= rg->to) { /* Remove entire region */ 545feba16e2SMike Kravetz del += rg->to - rg->from; 54696822904SAndy Whitcroft list_del(&rg->link); 54796822904SAndy Whitcroft kfree(rg); 548feba16e2SMike Kravetz continue; 54996822904SAndy Whitcroft } 5507b24d861SDavidlohr Bueso 551feba16e2SMike Kravetz if (f <= rg->from) { /* Trim beginning of region */ 552feba16e2SMike Kravetz del += t - rg->from; 553feba16e2SMike Kravetz rg->from = t; 554feba16e2SMike Kravetz } else { /* Trim end of region */ 555feba16e2SMike Kravetz del += rg->to - f; 556feba16e2SMike Kravetz rg->to = f; 557feba16e2SMike Kravetz } 558feba16e2SMike Kravetz } 559feba16e2SMike Kravetz 5607b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 561feba16e2SMike Kravetz kfree(nrg); 562feba16e2SMike Kravetz return del; 56396822904SAndy Whitcroft } 56496822904SAndy Whitcroft 5651dd308a7SMike Kravetz /* 566b5cec28dSMike Kravetz * A rare out of memory error was encountered which prevented removal of 567b5cec28dSMike Kravetz * the reserve map region for a page. The huge page itself was free'ed 568b5cec28dSMike Kravetz * and removed from the page cache. This routine will adjust the subpool 569b5cec28dSMike Kravetz * usage count, and the global reserve count if needed. By incrementing 570b5cec28dSMike Kravetz * these counts, the reserve map entry which could not be deleted will 571b5cec28dSMike Kravetz * appear as a "reserved" entry instead of simply dangling with incorrect 572b5cec28dSMike Kravetz * counts. 573b5cec28dSMike Kravetz */ 57472e2936cSzhong jiang void hugetlb_fix_reserve_counts(struct inode *inode) 575b5cec28dSMike Kravetz { 576b5cec28dSMike Kravetz struct hugepage_subpool *spool = subpool_inode(inode); 577b5cec28dSMike Kravetz long rsv_adjust; 578b5cec28dSMike Kravetz 579b5cec28dSMike Kravetz rsv_adjust = hugepage_subpool_get_pages(spool, 1); 58072e2936cSzhong jiang if (rsv_adjust) { 581b5cec28dSMike Kravetz struct hstate *h = hstate_inode(inode); 582b5cec28dSMike Kravetz 583b5cec28dSMike Kravetz hugetlb_acct_memory(h, 1); 584b5cec28dSMike Kravetz } 585b5cec28dSMike Kravetz } 586b5cec28dSMike Kravetz 587b5cec28dSMike Kravetz /* 5881dd308a7SMike Kravetz * Count and return the number of huge pages in the reserve map 5891dd308a7SMike Kravetz * that intersect with the range [f, t). 5901dd308a7SMike Kravetz */ 5911406ec9bSJoonsoo Kim static long region_count(struct resv_map *resv, long f, long t) 59284afd99bSAndy Whitcroft { 5931406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 59484afd99bSAndy Whitcroft struct file_region *rg; 59584afd99bSAndy Whitcroft long chg = 0; 59684afd99bSAndy Whitcroft 5977b24d861SDavidlohr Bueso spin_lock(&resv->lock); 59884afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 59984afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 600f2135a4aSWang Sheng-Hui long seg_from; 601f2135a4aSWang Sheng-Hui long seg_to; 60284afd99bSAndy Whitcroft 60384afd99bSAndy Whitcroft if (rg->to <= f) 60484afd99bSAndy Whitcroft continue; 60584afd99bSAndy Whitcroft if (rg->from >= t) 60684afd99bSAndy Whitcroft break; 60784afd99bSAndy Whitcroft 60884afd99bSAndy Whitcroft seg_from = max(rg->from, f); 60984afd99bSAndy Whitcroft seg_to = min(rg->to, t); 61084afd99bSAndy Whitcroft 61184afd99bSAndy Whitcroft chg += seg_to - seg_from; 61284afd99bSAndy Whitcroft } 6137b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 61484afd99bSAndy Whitcroft 61584afd99bSAndy Whitcroft return chg; 61684afd99bSAndy Whitcroft } 61784afd99bSAndy Whitcroft 61896822904SAndy Whitcroft /* 619e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 620e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 621e7c4b0bfSAndy Whitcroft */ 622a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 623a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 624e7c4b0bfSAndy Whitcroft { 625a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 626a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 627e7c4b0bfSAndy Whitcroft } 628e7c4b0bfSAndy Whitcroft 6290fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 6300fe6e20bSNaoya Horiguchi unsigned long address) 6310fe6e20bSNaoya Horiguchi { 6320fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 6330fe6e20bSNaoya Horiguchi } 634dee41079SDan Williams EXPORT_SYMBOL_GPL(linear_hugepage_index); 6350fe6e20bSNaoya Horiguchi 63684afd99bSAndy Whitcroft /* 63708fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 63808fba699SMel Gorman * cases this will be same size as used by the page table entries. 63908fba699SMel Gorman */ 64008fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 64108fba699SMel Gorman { 64205ea8860SDan Williams if (vma->vm_ops && vma->vm_ops->pagesize) 64305ea8860SDan Williams return vma->vm_ops->pagesize(vma); 64408fba699SMel Gorman return PAGE_SIZE; 64508fba699SMel Gorman } 646f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 64708fba699SMel Gorman 64808fba699SMel Gorman /* 6493340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 6503340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 65109135cc5SDan Williams * architectures where it differs, an architecture-specific 'strong' 65209135cc5SDan Williams * version of this symbol is required. 6533340289dSMel Gorman */ 65409135cc5SDan Williams __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 6553340289dSMel Gorman { 6563340289dSMel Gorman return vma_kernel_pagesize(vma); 6573340289dSMel Gorman } 6583340289dSMel Gorman 6593340289dSMel Gorman /* 66084afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 66184afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 66284afd99bSAndy Whitcroft * alignment. 66384afd99bSAndy Whitcroft */ 66484afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 66584afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 66604f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 66784afd99bSAndy Whitcroft 668a1e78772SMel Gorman /* 669a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 670a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 671a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 672a1e78772SMel Gorman * 673a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 674a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 675a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 676a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 67784afd99bSAndy Whitcroft * 67884afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 67984afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 68084afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 68184afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 68284afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 68384afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 68484afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 68584afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 686a1e78772SMel Gorman */ 687e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 688e7c4b0bfSAndy Whitcroft { 689e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 690e7c4b0bfSAndy Whitcroft } 691e7c4b0bfSAndy Whitcroft 692e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 693e7c4b0bfSAndy Whitcroft unsigned long value) 694e7c4b0bfSAndy Whitcroft { 695e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 696e7c4b0bfSAndy Whitcroft } 697e7c4b0bfSAndy Whitcroft 6989119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 69984afd99bSAndy Whitcroft { 70084afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 7015e911373SMike Kravetz struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 7025e911373SMike Kravetz 7035e911373SMike Kravetz if (!resv_map || !rg) { 7045e911373SMike Kravetz kfree(resv_map); 7055e911373SMike Kravetz kfree(rg); 70684afd99bSAndy Whitcroft return NULL; 7075e911373SMike Kravetz } 70884afd99bSAndy Whitcroft 70984afd99bSAndy Whitcroft kref_init(&resv_map->refs); 7107b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 71184afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 71284afd99bSAndy Whitcroft 7135e911373SMike Kravetz resv_map->adds_in_progress = 0; 7145e911373SMike Kravetz 7155e911373SMike Kravetz INIT_LIST_HEAD(&resv_map->region_cache); 7165e911373SMike Kravetz list_add(&rg->link, &resv_map->region_cache); 7175e911373SMike Kravetz resv_map->region_cache_count = 1; 7185e911373SMike Kravetz 71984afd99bSAndy Whitcroft return resv_map; 72084afd99bSAndy Whitcroft } 72184afd99bSAndy Whitcroft 7229119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 72384afd99bSAndy Whitcroft { 72484afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 7255e911373SMike Kravetz struct list_head *head = &resv_map->region_cache; 7265e911373SMike Kravetz struct file_region *rg, *trg; 72784afd99bSAndy Whitcroft 72884afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 729feba16e2SMike Kravetz region_del(resv_map, 0, LONG_MAX); 7305e911373SMike Kravetz 7315e911373SMike Kravetz /* ... and any entries left in the cache */ 7325e911373SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 7335e911373SMike Kravetz list_del(&rg->link); 7345e911373SMike Kravetz kfree(rg); 7355e911373SMike Kravetz } 7365e911373SMike Kravetz 7375e911373SMike Kravetz VM_BUG_ON(resv_map->adds_in_progress); 7385e911373SMike Kravetz 73984afd99bSAndy Whitcroft kfree(resv_map); 74084afd99bSAndy Whitcroft } 74184afd99bSAndy Whitcroft 7424e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 7434e35f483SJoonsoo Kim { 744f27a5136SMike Kravetz /* 745f27a5136SMike Kravetz * At inode evict time, i_mapping may not point to the original 746f27a5136SMike Kravetz * address space within the inode. This original address space 747f27a5136SMike Kravetz * contains the pointer to the resv_map. So, always use the 748f27a5136SMike Kravetz * address space embedded within the inode. 749f27a5136SMike Kravetz * The VERY common case is inode->mapping == &inode->i_data but, 750f27a5136SMike Kravetz * this may not be true for device special inodes. 751f27a5136SMike Kravetz */ 752f27a5136SMike Kravetz return (struct resv_map *)(&inode->i_data)->private_data; 7534e35f483SJoonsoo Kim } 7544e35f483SJoonsoo Kim 75584afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 756a1e78772SMel Gorman { 75781d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 7584e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 7594e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 7604e35f483SJoonsoo Kim struct inode *inode = mapping->host; 7614e35f483SJoonsoo Kim 7624e35f483SJoonsoo Kim return inode_resv_map(inode); 7634e35f483SJoonsoo Kim 7644e35f483SJoonsoo Kim } else { 76584afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 76684afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 7674e35f483SJoonsoo Kim } 768a1e78772SMel Gorman } 769a1e78772SMel Gorman 77084afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 771a1e78772SMel Gorman { 77281d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 77381d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 774a1e78772SMel Gorman 77584afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 77684afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 77704f2cbe3SMel Gorman } 77804f2cbe3SMel Gorman 77904f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 78004f2cbe3SMel Gorman { 78181d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 78281d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 783e7c4b0bfSAndy Whitcroft 784e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 78504f2cbe3SMel Gorman } 78604f2cbe3SMel Gorman 78704f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 78804f2cbe3SMel Gorman { 78981d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 790e7c4b0bfSAndy Whitcroft 791e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 792a1e78772SMel Gorman } 793a1e78772SMel Gorman 79404f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 795a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 796a1e78772SMel Gorman { 79781d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 798f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 799a1e78772SMel Gorman vma->vm_private_data = (void *)0; 800a1e78772SMel Gorman } 801a1e78772SMel Gorman 802a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 803559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 804a1e78772SMel Gorman { 805af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 806af0ed73eSJoonsoo Kim /* 807af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 808af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 809af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 810af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 811af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 812af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 813af0ed73eSJoonsoo Kim * properly, so add work-around here. 814af0ed73eSJoonsoo Kim */ 815af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 816559ec2f8SNicholas Krause return true; 817af0ed73eSJoonsoo Kim else 818559ec2f8SNicholas Krause return false; 819af0ed73eSJoonsoo Kim } 820a63884e9SJoonsoo Kim 821a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 8221fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 8231fb1b0e9SMike Kravetz /* 8241fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 8251fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 8261fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 8271fb1b0e9SMike Kravetz * fallocate. In this case, there really are no reverves to 8281fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 8291fb1b0e9SMike Kravetz */ 8301fb1b0e9SMike Kravetz if (chg) 8311fb1b0e9SMike Kravetz return false; 8321fb1b0e9SMike Kravetz else 833559ec2f8SNicholas Krause return true; 8341fb1b0e9SMike Kravetz } 835a63884e9SJoonsoo Kim 836a63884e9SJoonsoo Kim /* 837a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 838a63884e9SJoonsoo Kim * private mappings. 839a63884e9SJoonsoo Kim */ 84067961f9dSMike Kravetz if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 84167961f9dSMike Kravetz /* 84267961f9dSMike Kravetz * Like the shared case above, a hole punch or truncate 84367961f9dSMike Kravetz * could have been performed on the private mapping. 84467961f9dSMike Kravetz * Examine the value of chg to determine if reserves 84567961f9dSMike Kravetz * actually exist or were previously consumed. 84667961f9dSMike Kravetz * Very Subtle - The value of chg comes from a previous 84767961f9dSMike Kravetz * call to vma_needs_reserves(). The reserve map for 84867961f9dSMike Kravetz * private mappings has different (opposite) semantics 84967961f9dSMike Kravetz * than that of shared mappings. vma_needs_reserves() 85067961f9dSMike Kravetz * has already taken this difference in semantics into 85167961f9dSMike Kravetz * account. Therefore, the meaning of chg is the same 85267961f9dSMike Kravetz * as in the shared case above. Code could easily be 85367961f9dSMike Kravetz * combined, but keeping it separate draws attention to 85467961f9dSMike Kravetz * subtle differences. 85567961f9dSMike Kravetz */ 85667961f9dSMike Kravetz if (chg) 85767961f9dSMike Kravetz return false; 85867961f9dSMike Kravetz else 859559ec2f8SNicholas Krause return true; 86067961f9dSMike Kravetz } 861a63884e9SJoonsoo Kim 862559ec2f8SNicholas Krause return false; 863a1e78772SMel Gorman } 864a1e78772SMel Gorman 865a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 8661da177e4SLinus Torvalds { 8671da177e4SLinus Torvalds int nid = page_to_nid(page); 8680edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 869a5516438SAndi Kleen h->free_huge_pages++; 870a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 8711da177e4SLinus Torvalds } 8721da177e4SLinus Torvalds 87394310cbcSAnshuman Khandual static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) 874bf50bab2SNaoya Horiguchi { 875bf50bab2SNaoya Horiguchi struct page *page; 876bf50bab2SNaoya Horiguchi 877c8721bbbSNaoya Horiguchi list_for_each_entry(page, &h->hugepage_freelists[nid], lru) 878243abd5bSNaoya Horiguchi if (!PageHWPoison(page)) 879c8721bbbSNaoya Horiguchi break; 880c8721bbbSNaoya Horiguchi /* 881c8721bbbSNaoya Horiguchi * if 'non-isolated free hugepage' not found on the list, 882c8721bbbSNaoya Horiguchi * the allocation fails. 883c8721bbbSNaoya Horiguchi */ 884c8721bbbSNaoya Horiguchi if (&h->hugepage_freelists[nid] == &page->lru) 885bf50bab2SNaoya Horiguchi return NULL; 8860edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 887a9869b83SNaoya Horiguchi set_page_refcounted(page); 888bf50bab2SNaoya Horiguchi h->free_huge_pages--; 889bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 890bf50bab2SNaoya Horiguchi return page; 891bf50bab2SNaoya Horiguchi } 892bf50bab2SNaoya Horiguchi 8933e59fcb0SMichal Hocko static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, 8943e59fcb0SMichal Hocko nodemask_t *nmask) 89594310cbcSAnshuman Khandual { 8963e59fcb0SMichal Hocko unsigned int cpuset_mems_cookie; 8973e59fcb0SMichal Hocko struct zonelist *zonelist; 8983e59fcb0SMichal Hocko struct zone *zone; 8993e59fcb0SMichal Hocko struct zoneref *z; 90098fa15f3SAnshuman Khandual int node = NUMA_NO_NODE; 9013e59fcb0SMichal Hocko 9023e59fcb0SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 9033e59fcb0SMichal Hocko 9043e59fcb0SMichal Hocko retry_cpuset: 9053e59fcb0SMichal Hocko cpuset_mems_cookie = read_mems_allowed_begin(); 9063e59fcb0SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 90794310cbcSAnshuman Khandual struct page *page; 90894310cbcSAnshuman Khandual 9093e59fcb0SMichal Hocko if (!cpuset_zone_allowed(zone, gfp_mask)) 9103e59fcb0SMichal Hocko continue; 9113e59fcb0SMichal Hocko /* 9123e59fcb0SMichal Hocko * no need to ask again on the same node. Pool is node rather than 9133e59fcb0SMichal Hocko * zone aware 9143e59fcb0SMichal Hocko */ 9153e59fcb0SMichal Hocko if (zone_to_nid(zone) == node) 9163e59fcb0SMichal Hocko continue; 9173e59fcb0SMichal Hocko node = zone_to_nid(zone); 91894310cbcSAnshuman Khandual 91994310cbcSAnshuman Khandual page = dequeue_huge_page_node_exact(h, node); 92094310cbcSAnshuman Khandual if (page) 92194310cbcSAnshuman Khandual return page; 92294310cbcSAnshuman Khandual } 9233e59fcb0SMichal Hocko if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 9243e59fcb0SMichal Hocko goto retry_cpuset; 9253e59fcb0SMichal Hocko 92694310cbcSAnshuman Khandual return NULL; 92794310cbcSAnshuman Khandual } 92894310cbcSAnshuman Khandual 92986cdb465SNaoya Horiguchi /* Movability of hugepages depends on migration support. */ 93086cdb465SNaoya Horiguchi static inline gfp_t htlb_alloc_mask(struct hstate *h) 93186cdb465SNaoya Horiguchi { 9327ed2c31dSAnshuman Khandual if (hugepage_movable_supported(h)) 93386cdb465SNaoya Horiguchi return GFP_HIGHUSER_MOVABLE; 93486cdb465SNaoya Horiguchi else 93586cdb465SNaoya Horiguchi return GFP_HIGHUSER; 93686cdb465SNaoya Horiguchi } 93786cdb465SNaoya Horiguchi 938a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 939a5516438SAndi Kleen struct vm_area_struct *vma, 940af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 941af0ed73eSJoonsoo Kim long chg) 9421da177e4SLinus Torvalds { 9433e59fcb0SMichal Hocko struct page *page; 944480eccf9SLee Schermerhorn struct mempolicy *mpol; 94504ec6264SVlastimil Babka gfp_t gfp_mask; 9463e59fcb0SMichal Hocko nodemask_t *nodemask; 94704ec6264SVlastimil Babka int nid; 9481da177e4SLinus Torvalds 949a1e78772SMel Gorman /* 950a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 951a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 952a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 953a1e78772SMel Gorman */ 954af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg) && 955a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 956c0ff7453SMiao Xie goto err; 957a1e78772SMel Gorman 95804f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 959a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 9606eab04a8SJustin P. Mattock goto err; 96104f2cbe3SMel Gorman 96204ec6264SVlastimil Babka gfp_mask = htlb_alloc_mask(h); 96304ec6264SVlastimil Babka nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 9643e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 9653e59fcb0SMichal Hocko if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { 96607443a85SJoonsoo Kim SetPagePrivate(page); 967a63884e9SJoonsoo Kim h->resv_huge_pages--; 968bf50bab2SNaoya Horiguchi } 969cc9a6c87SMel Gorman 970cc9a6c87SMel Gorman mpol_cond_put(mpol); 971cc9a6c87SMel Gorman return page; 972cc9a6c87SMel Gorman 973c0ff7453SMiao Xie err: 974cc9a6c87SMel Gorman return NULL; 9751da177e4SLinus Torvalds } 9761da177e4SLinus Torvalds 9771cac6f2cSLuiz Capitulino /* 9781cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 9791cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 9801cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 9811cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 9821cac6f2cSLuiz Capitulino * node for alloc or free. 9831cac6f2cSLuiz Capitulino */ 9841cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 9851cac6f2cSLuiz Capitulino { 9860edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 9871cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 9881cac6f2cSLuiz Capitulino 9891cac6f2cSLuiz Capitulino return nid; 9901cac6f2cSLuiz Capitulino } 9911cac6f2cSLuiz Capitulino 9921cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 9931cac6f2cSLuiz Capitulino { 9941cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 9951cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 9961cac6f2cSLuiz Capitulino return nid; 9971cac6f2cSLuiz Capitulino } 9981cac6f2cSLuiz Capitulino 9991cac6f2cSLuiz Capitulino /* 10001cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 10011cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 10021cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 10031cac6f2cSLuiz Capitulino * mask. 10041cac6f2cSLuiz Capitulino */ 10051cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 10061cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 10071cac6f2cSLuiz Capitulino { 10081cac6f2cSLuiz Capitulino int nid; 10091cac6f2cSLuiz Capitulino 10101cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 10111cac6f2cSLuiz Capitulino 10121cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 10131cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 10141cac6f2cSLuiz Capitulino 10151cac6f2cSLuiz Capitulino return nid; 10161cac6f2cSLuiz Capitulino } 10171cac6f2cSLuiz Capitulino 10181cac6f2cSLuiz Capitulino /* 10191cac6f2cSLuiz Capitulino * helper for free_pool_huge_page() - return the previously saved 10201cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 10211cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 10221cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 10231cac6f2cSLuiz Capitulino */ 10241cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 10251cac6f2cSLuiz Capitulino { 10261cac6f2cSLuiz Capitulino int nid; 10271cac6f2cSLuiz Capitulino 10281cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 10291cac6f2cSLuiz Capitulino 10301cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 10311cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 10321cac6f2cSLuiz Capitulino 10331cac6f2cSLuiz Capitulino return nid; 10341cac6f2cSLuiz Capitulino } 10351cac6f2cSLuiz Capitulino 10361cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 10371cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 10381cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 10391cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 10401cac6f2cSLuiz Capitulino nr_nodes--) 10411cac6f2cSLuiz Capitulino 10421cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 10431cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 10441cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 10451cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 10461cac6f2cSLuiz Capitulino nr_nodes--) 10471cac6f2cSLuiz Capitulino 1048e1073d1eSAneesh Kumar K.V #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 1049944d9fecSLuiz Capitulino static void destroy_compound_gigantic_page(struct page *page, 1050d00181b9SKirill A. Shutemov unsigned int order) 1051944d9fecSLuiz Capitulino { 1052944d9fecSLuiz Capitulino int i; 1053944d9fecSLuiz Capitulino int nr_pages = 1 << order; 1054944d9fecSLuiz Capitulino struct page *p = page + 1; 1055944d9fecSLuiz Capitulino 1056c8cc708aSGerald Schaefer atomic_set(compound_mapcount_ptr(page), 0); 1057944d9fecSLuiz Capitulino for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 10581d798ca3SKirill A. Shutemov clear_compound_head(p); 1059944d9fecSLuiz Capitulino set_page_refcounted(p); 1060944d9fecSLuiz Capitulino } 1061944d9fecSLuiz Capitulino 1062944d9fecSLuiz Capitulino set_compound_order(page, 0); 1063944d9fecSLuiz Capitulino __ClearPageHead(page); 1064944d9fecSLuiz Capitulino } 1065944d9fecSLuiz Capitulino 1066d00181b9SKirill A. Shutemov static void free_gigantic_page(struct page *page, unsigned int order) 1067944d9fecSLuiz Capitulino { 1068944d9fecSLuiz Capitulino free_contig_range(page_to_pfn(page), 1 << order); 1069944d9fecSLuiz Capitulino } 1070944d9fecSLuiz Capitulino 10714eb0716eSAlexandre Ghiti #ifdef CONFIG_CONTIG_ALLOC 1072944d9fecSLuiz Capitulino static int __alloc_gigantic_page(unsigned long start_pfn, 107379b63f12SMichal Hocko unsigned long nr_pages, gfp_t gfp_mask) 1074944d9fecSLuiz Capitulino { 1075944d9fecSLuiz Capitulino unsigned long end_pfn = start_pfn + nr_pages; 1076ca96b625SLucas Stach return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE, 107779b63f12SMichal Hocko gfp_mask); 1078944d9fecSLuiz Capitulino } 1079944d9fecSLuiz Capitulino 1080f44b2ddaSJoonsoo Kim static bool pfn_range_valid_gigantic(struct zone *z, 1081f44b2ddaSJoonsoo Kim unsigned long start_pfn, unsigned long nr_pages) 1082944d9fecSLuiz Capitulino { 1083944d9fecSLuiz Capitulino unsigned long i, end_pfn = start_pfn + nr_pages; 1084944d9fecSLuiz Capitulino struct page *page; 1085944d9fecSLuiz Capitulino 1086944d9fecSLuiz Capitulino for (i = start_pfn; i < end_pfn; i++) { 1087944d9fecSLuiz Capitulino if (!pfn_valid(i)) 1088944d9fecSLuiz Capitulino return false; 1089944d9fecSLuiz Capitulino 1090944d9fecSLuiz Capitulino page = pfn_to_page(i); 1091944d9fecSLuiz Capitulino 1092f44b2ddaSJoonsoo Kim if (page_zone(page) != z) 1093f44b2ddaSJoonsoo Kim return false; 1094f44b2ddaSJoonsoo Kim 1095944d9fecSLuiz Capitulino if (PageReserved(page)) 1096944d9fecSLuiz Capitulino return false; 1097944d9fecSLuiz Capitulino 1098944d9fecSLuiz Capitulino if (page_count(page) > 0) 1099944d9fecSLuiz Capitulino return false; 1100944d9fecSLuiz Capitulino 1101944d9fecSLuiz Capitulino if (PageHuge(page)) 1102944d9fecSLuiz Capitulino return false; 1103944d9fecSLuiz Capitulino } 1104944d9fecSLuiz Capitulino 1105944d9fecSLuiz Capitulino return true; 1106944d9fecSLuiz Capitulino } 1107944d9fecSLuiz Capitulino 1108944d9fecSLuiz Capitulino static bool zone_spans_last_pfn(const struct zone *zone, 1109944d9fecSLuiz Capitulino unsigned long start_pfn, unsigned long nr_pages) 1110944d9fecSLuiz Capitulino { 1111944d9fecSLuiz Capitulino unsigned long last_pfn = start_pfn + nr_pages - 1; 1112944d9fecSLuiz Capitulino return zone_spans_pfn(zone, last_pfn); 1113944d9fecSLuiz Capitulino } 1114944d9fecSLuiz Capitulino 1115d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 1116d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) 1117944d9fecSLuiz Capitulino { 111879b63f12SMichal Hocko unsigned int order = huge_page_order(h); 1119944d9fecSLuiz Capitulino unsigned long nr_pages = 1 << order; 1120944d9fecSLuiz Capitulino unsigned long ret, pfn, flags; 112179b63f12SMichal Hocko struct zonelist *zonelist; 112279b63f12SMichal Hocko struct zone *zone; 112379b63f12SMichal Hocko struct zoneref *z; 1124944d9fecSLuiz Capitulino 112579b63f12SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 1126d9cc948fSMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nodemask) { 112779b63f12SMichal Hocko spin_lock_irqsave(&zone->lock, flags); 1128944d9fecSLuiz Capitulino 112979b63f12SMichal Hocko pfn = ALIGN(zone->zone_start_pfn, nr_pages); 113079b63f12SMichal Hocko while (zone_spans_last_pfn(zone, pfn, nr_pages)) { 113179b63f12SMichal Hocko if (pfn_range_valid_gigantic(zone, pfn, nr_pages)) { 1132944d9fecSLuiz Capitulino /* 1133944d9fecSLuiz Capitulino * We release the zone lock here because 1134944d9fecSLuiz Capitulino * alloc_contig_range() will also lock the zone 1135944d9fecSLuiz Capitulino * at some point. If there's an allocation 1136944d9fecSLuiz Capitulino * spinning on this lock, it may win the race 1137944d9fecSLuiz Capitulino * and cause alloc_contig_range() to fail... 1138944d9fecSLuiz Capitulino */ 113979b63f12SMichal Hocko spin_unlock_irqrestore(&zone->lock, flags); 114079b63f12SMichal Hocko ret = __alloc_gigantic_page(pfn, nr_pages, gfp_mask); 1141944d9fecSLuiz Capitulino if (!ret) 1142944d9fecSLuiz Capitulino return pfn_to_page(pfn); 114379b63f12SMichal Hocko spin_lock_irqsave(&zone->lock, flags); 1144944d9fecSLuiz Capitulino } 1145944d9fecSLuiz Capitulino pfn += nr_pages; 1146944d9fecSLuiz Capitulino } 1147944d9fecSLuiz Capitulino 114879b63f12SMichal Hocko spin_unlock_irqrestore(&zone->lock, flags); 1149944d9fecSLuiz Capitulino } 1150944d9fecSLuiz Capitulino 1151944d9fecSLuiz Capitulino return NULL; 1152944d9fecSLuiz Capitulino } 1153944d9fecSLuiz Capitulino 1154944d9fecSLuiz Capitulino static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); 1155d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order); 11564eb0716eSAlexandre Ghiti #else /* !CONFIG_CONTIG_ALLOC */ 11574eb0716eSAlexandre Ghiti static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 11584eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 11594eb0716eSAlexandre Ghiti { 11604eb0716eSAlexandre Ghiti return NULL; 11614eb0716eSAlexandre Ghiti } 11624eb0716eSAlexandre Ghiti #endif /* CONFIG_CONTIG_ALLOC */ 1163944d9fecSLuiz Capitulino 1164e1073d1eSAneesh Kumar K.V #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 1165d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 11664eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 11674eb0716eSAlexandre Ghiti { 11684eb0716eSAlexandre Ghiti return NULL; 11694eb0716eSAlexandre Ghiti } 1170d00181b9SKirill A. Shutemov static inline void free_gigantic_page(struct page *page, unsigned int order) { } 1171944d9fecSLuiz Capitulino static inline void destroy_compound_gigantic_page(struct page *page, 1172d00181b9SKirill A. Shutemov unsigned int order) { } 1173944d9fecSLuiz Capitulino #endif 1174944d9fecSLuiz Capitulino 1175a5516438SAndi Kleen static void update_and_free_page(struct hstate *h, struct page *page) 11766af2acb6SAdam Litke { 11776af2acb6SAdam Litke int i; 1178a5516438SAndi Kleen 11794eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1180944d9fecSLuiz Capitulino return; 118118229df5SAndy Whitcroft 1182a5516438SAndi Kleen h->nr_huge_pages--; 1183a5516438SAndi Kleen h->nr_huge_pages_node[page_to_nid(page)]--; 1184a5516438SAndi Kleen for (i = 0; i < pages_per_huge_page(h); i++) { 118532f84528SChris Forbes page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 118632f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1187a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1188a7407a27SLuiz Capitulino 1 << PG_writeback); 11896af2acb6SAdam Litke } 1190309381feSSasha Levin VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 1191f1e61557SKirill A. Shutemov set_compound_page_dtor(page, NULL_COMPOUND_DTOR); 11926af2acb6SAdam Litke set_page_refcounted(page); 1193944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) { 1194944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 1195944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 1196944d9fecSLuiz Capitulino } else { 1197a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 11986af2acb6SAdam Litke } 1199944d9fecSLuiz Capitulino } 12006af2acb6SAdam Litke 1201e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1202e5ff2159SAndi Kleen { 1203e5ff2159SAndi Kleen struct hstate *h; 1204e5ff2159SAndi Kleen 1205e5ff2159SAndi Kleen for_each_hstate(h) { 1206e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1207e5ff2159SAndi Kleen return h; 1208e5ff2159SAndi Kleen } 1209e5ff2159SAndi Kleen return NULL; 1210e5ff2159SAndi Kleen } 1211e5ff2159SAndi Kleen 1212bcc54222SNaoya Horiguchi /* 1213bcc54222SNaoya Horiguchi * Test to determine whether the hugepage is "active/in-use" (i.e. being linked 1214bcc54222SNaoya Horiguchi * to hstate->hugepage_activelist.) 1215bcc54222SNaoya Horiguchi * 1216bcc54222SNaoya Horiguchi * This function can be called for tail pages, but never returns true for them. 1217bcc54222SNaoya Horiguchi */ 1218bcc54222SNaoya Horiguchi bool page_huge_active(struct page *page) 1219bcc54222SNaoya Horiguchi { 1220bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHuge(page), page); 1221bcc54222SNaoya Horiguchi return PageHead(page) && PagePrivate(&page[1]); 1222bcc54222SNaoya Horiguchi } 1223bcc54222SNaoya Horiguchi 1224bcc54222SNaoya Horiguchi /* never called for tail page */ 1225bcc54222SNaoya Horiguchi static void set_page_huge_active(struct page *page) 1226bcc54222SNaoya Horiguchi { 1227bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 1228bcc54222SNaoya Horiguchi SetPagePrivate(&page[1]); 1229bcc54222SNaoya Horiguchi } 1230bcc54222SNaoya Horiguchi 1231bcc54222SNaoya Horiguchi static void clear_page_huge_active(struct page *page) 1232bcc54222SNaoya Horiguchi { 1233bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 1234bcc54222SNaoya Horiguchi ClearPagePrivate(&page[1]); 1235bcc54222SNaoya Horiguchi } 1236bcc54222SNaoya Horiguchi 1237ab5ac90aSMichal Hocko /* 1238ab5ac90aSMichal Hocko * Internal hugetlb specific page flag. Do not use outside of the hugetlb 1239ab5ac90aSMichal Hocko * code 1240ab5ac90aSMichal Hocko */ 1241ab5ac90aSMichal Hocko static inline bool PageHugeTemporary(struct page *page) 1242ab5ac90aSMichal Hocko { 1243ab5ac90aSMichal Hocko if (!PageHuge(page)) 1244ab5ac90aSMichal Hocko return false; 1245ab5ac90aSMichal Hocko 1246ab5ac90aSMichal Hocko return (unsigned long)page[2].mapping == -1U; 1247ab5ac90aSMichal Hocko } 1248ab5ac90aSMichal Hocko 1249ab5ac90aSMichal Hocko static inline void SetPageHugeTemporary(struct page *page) 1250ab5ac90aSMichal Hocko { 1251ab5ac90aSMichal Hocko page[2].mapping = (void *)-1U; 1252ab5ac90aSMichal Hocko } 1253ab5ac90aSMichal Hocko 1254ab5ac90aSMichal Hocko static inline void ClearPageHugeTemporary(struct page *page) 1255ab5ac90aSMichal Hocko { 1256ab5ac90aSMichal Hocko page[2].mapping = NULL; 1257ab5ac90aSMichal Hocko } 1258ab5ac90aSMichal Hocko 12598f1d26d0SAtsushi Kumagai void free_huge_page(struct page *page) 126027a85ef1SDavid Gibson { 1261a5516438SAndi Kleen /* 1262a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1263a5516438SAndi Kleen * compound page destructor. 1264a5516438SAndi Kleen */ 1265e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 12667893d1d5SAdam Litke int nid = page_to_nid(page); 126790481622SDavid Gibson struct hugepage_subpool *spool = 126890481622SDavid Gibson (struct hugepage_subpool *)page_private(page); 126907443a85SJoonsoo Kim bool restore_reserve; 127027a85ef1SDavid Gibson 1271b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1272b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_mapcount(page), page); 12738ace22bcSYongkai Wu 12748ace22bcSYongkai Wu set_page_private(page, 0); 12758ace22bcSYongkai Wu page->mapping = NULL; 127607443a85SJoonsoo Kim restore_reserve = PagePrivate(page); 127716c794b4SJoonsoo Kim ClearPagePrivate(page); 127827a85ef1SDavid Gibson 12791c5ecae3SMike Kravetz /* 12800919e1b6SMike Kravetz * If PagePrivate() was set on page, page allocation consumed a 12810919e1b6SMike Kravetz * reservation. If the page was associated with a subpool, there 12820919e1b6SMike Kravetz * would have been a page reserved in the subpool before allocation 12830919e1b6SMike Kravetz * via hugepage_subpool_get_pages(). Since we are 'restoring' the 12840919e1b6SMike Kravetz * reservtion, do not call hugepage_subpool_put_pages() as this will 12850919e1b6SMike Kravetz * remove the reserved page from the subpool. 12860919e1b6SMike Kravetz */ 12870919e1b6SMike Kravetz if (!restore_reserve) { 12880919e1b6SMike Kravetz /* 12890919e1b6SMike Kravetz * A return code of zero implies that the subpool will be 12900919e1b6SMike Kravetz * under its minimum size if the reservation is not restored 12910919e1b6SMike Kravetz * after page is free. Therefore, force restore_reserve 12920919e1b6SMike Kravetz * operation. 12931c5ecae3SMike Kravetz */ 12941c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 12951c5ecae3SMike Kravetz restore_reserve = true; 12960919e1b6SMike Kravetz } 12971c5ecae3SMike Kravetz 129827a85ef1SDavid Gibson spin_lock(&hugetlb_lock); 1299bcc54222SNaoya Horiguchi clear_page_huge_active(page); 13006d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 13016d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 130207443a85SJoonsoo Kim if (restore_reserve) 130307443a85SJoonsoo Kim h->resv_huge_pages++; 130407443a85SJoonsoo Kim 1305ab5ac90aSMichal Hocko if (PageHugeTemporary(page)) { 1306ab5ac90aSMichal Hocko list_del(&page->lru); 1307ab5ac90aSMichal Hocko ClearPageHugeTemporary(page); 1308ab5ac90aSMichal Hocko update_and_free_page(h, page); 1309ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 13100edaecfaSAneesh Kumar K.V /* remove the page from active list */ 13110edaecfaSAneesh Kumar K.V list_del(&page->lru); 1312a5516438SAndi Kleen update_and_free_page(h, page); 1313a5516438SAndi Kleen h->surplus_huge_pages--; 1314a5516438SAndi Kleen h->surplus_huge_pages_node[nid]--; 13157893d1d5SAdam Litke } else { 13165d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1317a5516438SAndi Kleen enqueue_huge_page(h, page); 13187893d1d5SAdam Litke } 131927a85ef1SDavid Gibson spin_unlock(&hugetlb_lock); 132027a85ef1SDavid Gibson } 132127a85ef1SDavid Gibson 1322a5516438SAndi Kleen static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1323b7ba30c6SAndi Kleen { 13240edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1325f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1326b7ba30c6SAndi Kleen spin_lock(&hugetlb_lock); 13279dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 1328a5516438SAndi Kleen h->nr_huge_pages++; 1329a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 1330b7ba30c6SAndi Kleen spin_unlock(&hugetlb_lock); 1331b7ba30c6SAndi Kleen } 1332b7ba30c6SAndi Kleen 1333d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order) 133420a0307cSWu Fengguang { 133520a0307cSWu Fengguang int i; 133620a0307cSWu Fengguang int nr_pages = 1 << order; 133720a0307cSWu Fengguang struct page *p = page + 1; 133820a0307cSWu Fengguang 133920a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 134020a0307cSWu Fengguang set_compound_order(page, order); 1341ef5a22beSAndrea Arcangeli __ClearPageReserved(page); 1342de09d31dSKirill A. Shutemov __SetPageHead(page); 134320a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1344ef5a22beSAndrea Arcangeli /* 1345ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1346ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1347ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 1348ef5a22beSAndrea Arcangeli * too. Otherwse drivers using get_user_pages() to access tail 1349ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1350ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1351ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1352ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1353ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1354ef5a22beSAndrea Arcangeli * after get_user_pages(). 1355ef5a22beSAndrea Arcangeli */ 1356ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 135758a84aa9SYouquan Song set_page_count(p, 0); 13581d798ca3SKirill A. Shutemov set_compound_head(p, page); 135920a0307cSWu Fengguang } 1360b4330afbSMike Kravetz atomic_set(compound_mapcount_ptr(page), -1); 136120a0307cSWu Fengguang } 136220a0307cSWu Fengguang 13637795912cSAndrew Morton /* 13647795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 13657795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 13667795912cSAndrew Morton * details. 13677795912cSAndrew Morton */ 136820a0307cSWu Fengguang int PageHuge(struct page *page) 136920a0307cSWu Fengguang { 137020a0307cSWu Fengguang if (!PageCompound(page)) 137120a0307cSWu Fengguang return 0; 137220a0307cSWu Fengguang 137320a0307cSWu Fengguang page = compound_head(page); 1374f1e61557SKirill A. Shutemov return page[1].compound_dtor == HUGETLB_PAGE_DTOR; 137520a0307cSWu Fengguang } 137643131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 137743131e14SNaoya Horiguchi 137827c73ae7SAndrea Arcangeli /* 137927c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 138027c73ae7SAndrea Arcangeli * normal or transparent huge pages. 138127c73ae7SAndrea Arcangeli */ 138227c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 138327c73ae7SAndrea Arcangeli { 138427c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 138527c73ae7SAndrea Arcangeli return 0; 138627c73ae7SAndrea Arcangeli 1387758f66a2SAndrew Morton return get_compound_page_dtor(page_head) == free_huge_page; 138827c73ae7SAndrea Arcangeli } 138927c73ae7SAndrea Arcangeli 139013d60f4bSZhang Yi pgoff_t __basepage_index(struct page *page) 139113d60f4bSZhang Yi { 139213d60f4bSZhang Yi struct page *page_head = compound_head(page); 139313d60f4bSZhang Yi pgoff_t index = page_index(page_head); 139413d60f4bSZhang Yi unsigned long compound_idx; 139513d60f4bSZhang Yi 139613d60f4bSZhang Yi if (!PageHuge(page_head)) 139713d60f4bSZhang Yi return page_index(page); 139813d60f4bSZhang Yi 139913d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 140013d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 140113d60f4bSZhang Yi else 140213d60f4bSZhang Yi compound_idx = page - page_head; 140313d60f4bSZhang Yi 140413d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 140513d60f4bSZhang Yi } 140613d60f4bSZhang Yi 14070c397daeSMichal Hocko static struct page *alloc_buddy_huge_page(struct hstate *h, 1408af0fb9dfSMichal Hocko gfp_t gfp_mask, int nid, nodemask_t *nmask) 14091da177e4SLinus Torvalds { 1410af0fb9dfSMichal Hocko int order = huge_page_order(h); 14111da177e4SLinus Torvalds struct page *page; 1412f96efd58SJoe Jin 1413af0fb9dfSMichal Hocko gfp_mask |= __GFP_COMP|__GFP_RETRY_MAYFAIL|__GFP_NOWARN; 1414af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 1415af0fb9dfSMichal Hocko nid = numa_mem_id(); 1416af0fb9dfSMichal Hocko page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask); 1417af0fb9dfSMichal Hocko if (page) 1418af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC); 1419af0fb9dfSMichal Hocko else 1420af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 142163b4613cSNishanth Aravamudan 142263b4613cSNishanth Aravamudan return page; 142363b4613cSNishanth Aravamudan } 142463b4613cSNishanth Aravamudan 1425af0fb9dfSMichal Hocko /* 14260c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 14270c397daeSMichal Hocko * should use this function to get new hugetlb pages 14280c397daeSMichal Hocko */ 14290c397daeSMichal Hocko static struct page *alloc_fresh_huge_page(struct hstate *h, 14300c397daeSMichal Hocko gfp_t gfp_mask, int nid, nodemask_t *nmask) 14310c397daeSMichal Hocko { 14320c397daeSMichal Hocko struct page *page; 14330c397daeSMichal Hocko 14340c397daeSMichal Hocko if (hstate_is_gigantic(h)) 14350c397daeSMichal Hocko page = alloc_gigantic_page(h, gfp_mask, nid, nmask); 14360c397daeSMichal Hocko else 14370c397daeSMichal Hocko page = alloc_buddy_huge_page(h, gfp_mask, 14380c397daeSMichal Hocko nid, nmask); 14390c397daeSMichal Hocko if (!page) 14400c397daeSMichal Hocko return NULL; 14410c397daeSMichal Hocko 14420c397daeSMichal Hocko if (hstate_is_gigantic(h)) 14430c397daeSMichal Hocko prep_compound_gigantic_page(page, huge_page_order(h)); 14440c397daeSMichal Hocko prep_new_huge_page(h, page, page_to_nid(page)); 14450c397daeSMichal Hocko 14460c397daeSMichal Hocko return page; 14470c397daeSMichal Hocko } 14480c397daeSMichal Hocko 14490c397daeSMichal Hocko /* 1450af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 1451af0fb9dfSMichal Hocko * manner. 1452af0fb9dfSMichal Hocko */ 14530c397daeSMichal Hocko static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 1454b2261026SJoonsoo Kim { 1455b2261026SJoonsoo Kim struct page *page; 1456b2261026SJoonsoo Kim int nr_nodes, node; 1457af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 1458b2261026SJoonsoo Kim 1459b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 14600c397daeSMichal Hocko page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed); 1461af0fb9dfSMichal Hocko if (page) 1462b2261026SJoonsoo Kim break; 1463b2261026SJoonsoo Kim } 1464b2261026SJoonsoo Kim 1465af0fb9dfSMichal Hocko if (!page) 1466af0fb9dfSMichal Hocko return 0; 1467b2261026SJoonsoo Kim 1468af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 1469af0fb9dfSMichal Hocko 1470af0fb9dfSMichal Hocko return 1; 1471b2261026SJoonsoo Kim } 1472b2261026SJoonsoo Kim 1473e8c5c824SLee Schermerhorn /* 1474e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 1475e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 1476e8c5c824SLee Schermerhorn * balanced over allowed nodes. 1477e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 1478e8c5c824SLee Schermerhorn */ 14796ae11b27SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 14806ae11b27SLee Schermerhorn bool acct_surplus) 1481e8c5c824SLee Schermerhorn { 1482b2261026SJoonsoo Kim int nr_nodes, node; 1483e8c5c824SLee Schermerhorn int ret = 0; 1484e8c5c824SLee Schermerhorn 1485b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1486685f3457SLee Schermerhorn /* 1487685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 1488685f3457SLee Schermerhorn * nodes with surplus pages. 1489685f3457SLee Schermerhorn */ 1490b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 1491b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 1492e8c5c824SLee Schermerhorn struct page *page = 1493b2261026SJoonsoo Kim list_entry(h->hugepage_freelists[node].next, 1494e8c5c824SLee Schermerhorn struct page, lru); 1495e8c5c824SLee Schermerhorn list_del(&page->lru); 1496e8c5c824SLee Schermerhorn h->free_huge_pages--; 1497b2261026SJoonsoo Kim h->free_huge_pages_node[node]--; 1498685f3457SLee Schermerhorn if (acct_surplus) { 1499685f3457SLee Schermerhorn h->surplus_huge_pages--; 1500b2261026SJoonsoo Kim h->surplus_huge_pages_node[node]--; 1501685f3457SLee Schermerhorn } 1502e8c5c824SLee Schermerhorn update_and_free_page(h, page); 1503e8c5c824SLee Schermerhorn ret = 1; 15049a76db09SLee Schermerhorn break; 1505e8c5c824SLee Schermerhorn } 1506b2261026SJoonsoo Kim } 1507e8c5c824SLee Schermerhorn 1508e8c5c824SLee Schermerhorn return ret; 1509e8c5c824SLee Schermerhorn } 1510e8c5c824SLee Schermerhorn 1511c8721bbbSNaoya Horiguchi /* 1512c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 1513faf53defSNaoya Horiguchi * nothing for in-use hugepages and non-hugepages. 1514faf53defSNaoya Horiguchi * This function returns values like below: 1515faf53defSNaoya Horiguchi * 1516faf53defSNaoya Horiguchi * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 1517faf53defSNaoya Horiguchi * (allocated or reserved.) 1518faf53defSNaoya Horiguchi * 0: successfully dissolved free hugepages or the page is not a 1519faf53defSNaoya Horiguchi * hugepage (considered as already dissolved) 1520c8721bbbSNaoya Horiguchi */ 1521c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 1522c8721bbbSNaoya Horiguchi { 15236bc9b564SNaoya Horiguchi int rc = -EBUSY; 1524082d5b6bSGerald Schaefer 1525faf53defSNaoya Horiguchi /* Not to disrupt normal path by vainly holding hugetlb_lock */ 1526faf53defSNaoya Horiguchi if (!PageHuge(page)) 1527faf53defSNaoya Horiguchi return 0; 1528faf53defSNaoya Horiguchi 1529c8721bbbSNaoya Horiguchi spin_lock(&hugetlb_lock); 1530faf53defSNaoya Horiguchi if (!PageHuge(page)) { 1531faf53defSNaoya Horiguchi rc = 0; 1532faf53defSNaoya Horiguchi goto out; 1533faf53defSNaoya Horiguchi } 1534faf53defSNaoya Horiguchi 1535faf53defSNaoya Horiguchi if (!page_count(page)) { 15362247bb33SGerald Schaefer struct page *head = compound_head(page); 15372247bb33SGerald Schaefer struct hstate *h = page_hstate(head); 15382247bb33SGerald Schaefer int nid = page_to_nid(head); 15396bc9b564SNaoya Horiguchi if (h->free_huge_pages - h->resv_huge_pages == 0) 1540082d5b6bSGerald Schaefer goto out; 1541c3114a84SAnshuman Khandual /* 1542c3114a84SAnshuman Khandual * Move PageHWPoison flag from head page to the raw error page, 1543c3114a84SAnshuman Khandual * which makes any subpages rather than the error page reusable. 1544c3114a84SAnshuman Khandual */ 1545c3114a84SAnshuman Khandual if (PageHWPoison(head) && page != head) { 1546c3114a84SAnshuman Khandual SetPageHWPoison(page); 1547c3114a84SAnshuman Khandual ClearPageHWPoison(head); 1548c3114a84SAnshuman Khandual } 15492247bb33SGerald Schaefer list_del(&head->lru); 1550c8721bbbSNaoya Horiguchi h->free_huge_pages--; 1551c8721bbbSNaoya Horiguchi h->free_huge_pages_node[nid]--; 1552c1470b33Szhong jiang h->max_huge_pages--; 15532247bb33SGerald Schaefer update_and_free_page(h, head); 15546bc9b564SNaoya Horiguchi rc = 0; 1555c8721bbbSNaoya Horiguchi } 1556082d5b6bSGerald Schaefer out: 1557c8721bbbSNaoya Horiguchi spin_unlock(&hugetlb_lock); 1558082d5b6bSGerald Schaefer return rc; 1559c8721bbbSNaoya Horiguchi } 1560c8721bbbSNaoya Horiguchi 1561c8721bbbSNaoya Horiguchi /* 1562c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 1563c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 15642247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 15652247bb33SGerald Schaefer * part of it lies within the given range. 1566082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 1567082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 1568c8721bbbSNaoya Horiguchi */ 1569082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 1570c8721bbbSNaoya Horiguchi { 1571c8721bbbSNaoya Horiguchi unsigned long pfn; 1572eb03aa00SGerald Schaefer struct page *page; 1573082d5b6bSGerald Schaefer int rc = 0; 1574c8721bbbSNaoya Horiguchi 1575d0177639SLi Zhong if (!hugepages_supported()) 1576082d5b6bSGerald Schaefer return rc; 1577d0177639SLi Zhong 1578eb03aa00SGerald Schaefer for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { 1579eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 1580eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 1581eb03aa00SGerald Schaefer if (rc) 1582082d5b6bSGerald Schaefer break; 1583eb03aa00SGerald Schaefer } 1584082d5b6bSGerald Schaefer 1585082d5b6bSGerald Schaefer return rc; 1586c8721bbbSNaoya Horiguchi } 1587c8721bbbSNaoya Horiguchi 1588ab5ac90aSMichal Hocko /* 1589ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 1590ab5ac90aSMichal Hocko */ 15910c397daeSMichal Hocko static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, 1592aaf14e40SMichal Hocko int nid, nodemask_t *nmask) 15937893d1d5SAdam Litke { 15949980d744SMichal Hocko struct page *page = NULL; 15957893d1d5SAdam Litke 1596bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1597aa888a74SAndi Kleen return NULL; 1598aa888a74SAndi Kleen 1599d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 16009980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 16019980d744SMichal Hocko goto out_unlock; 1602d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1603d1c3fb1fSNishanth Aravamudan 16040c397daeSMichal Hocko page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask); 16059980d744SMichal Hocko if (!page) 16060c397daeSMichal Hocko return NULL; 1607d1c3fb1fSNishanth Aravamudan 16087893d1d5SAdam Litke spin_lock(&hugetlb_lock); 16099980d744SMichal Hocko /* 16109980d744SMichal Hocko * We could have raced with the pool size change. 16119980d744SMichal Hocko * Double check that and simply deallocate the new page 16129980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 16139980d744SMichal Hocko * temporary page to workaround the nasty free_huge_page 16149980d744SMichal Hocko * codeflow 16159980d744SMichal Hocko */ 16169980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 16179980d744SMichal Hocko SetPageHugeTemporary(page); 16182bf753e6SKai Shen spin_unlock(&hugetlb_lock); 16199980d744SMichal Hocko put_page(page); 16202bf753e6SKai Shen return NULL; 16219980d744SMichal Hocko } else { 16229980d744SMichal Hocko h->surplus_huge_pages++; 16234704dea3SMichal Hocko h->surplus_huge_pages_node[page_to_nid(page)]++; 16247893d1d5SAdam Litke } 16259980d744SMichal Hocko 16269980d744SMichal Hocko out_unlock: 1627d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 16287893d1d5SAdam Litke 16297893d1d5SAdam Litke return page; 16307893d1d5SAdam Litke } 16317893d1d5SAdam Litke 16329a4e9f3bSAneesh Kumar K.V struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, 1633ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 1634ab5ac90aSMichal Hocko { 1635ab5ac90aSMichal Hocko struct page *page; 1636ab5ac90aSMichal Hocko 1637ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 1638ab5ac90aSMichal Hocko return NULL; 1639ab5ac90aSMichal Hocko 16400c397daeSMichal Hocko page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask); 1641ab5ac90aSMichal Hocko if (!page) 1642ab5ac90aSMichal Hocko return NULL; 1643ab5ac90aSMichal Hocko 1644ab5ac90aSMichal Hocko /* 1645ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 1646ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 1647ab5ac90aSMichal Hocko */ 1648ab5ac90aSMichal Hocko SetPageHugeTemporary(page); 1649ab5ac90aSMichal Hocko 1650ab5ac90aSMichal Hocko return page; 1651ab5ac90aSMichal Hocko } 1652ab5ac90aSMichal Hocko 1653e4e574b7SAdam Litke /* 1654099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 1655099730d6SDave Hansen */ 1656e0ec90eeSDave Hansen static 16570c397daeSMichal Hocko struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, 1658099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 1659099730d6SDave Hansen { 1660aaf14e40SMichal Hocko struct page *page; 1661aaf14e40SMichal Hocko struct mempolicy *mpol; 1662aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 1663aaf14e40SMichal Hocko int nid; 1664aaf14e40SMichal Hocko nodemask_t *nodemask; 1665aaf14e40SMichal Hocko 1666aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 16670c397daeSMichal Hocko page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); 1668aaf14e40SMichal Hocko mpol_cond_put(mpol); 1669aaf14e40SMichal Hocko 1670aaf14e40SMichal Hocko return page; 1671099730d6SDave Hansen } 1672099730d6SDave Hansen 1673ab5ac90aSMichal Hocko /* page migration callback function */ 1674bf50bab2SNaoya Horiguchi struct page *alloc_huge_page_node(struct hstate *h, int nid) 1675bf50bab2SNaoya Horiguchi { 1676aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 16774ef91848SJoonsoo Kim struct page *page = NULL; 1678bf50bab2SNaoya Horiguchi 1679aaf14e40SMichal Hocko if (nid != NUMA_NO_NODE) 1680aaf14e40SMichal Hocko gfp_mask |= __GFP_THISNODE; 1681aaf14e40SMichal Hocko 1682bf50bab2SNaoya Horiguchi spin_lock(&hugetlb_lock); 16834ef91848SJoonsoo Kim if (h->free_huge_pages - h->resv_huge_pages > 0) 16843e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL); 1685bf50bab2SNaoya Horiguchi spin_unlock(&hugetlb_lock); 1686bf50bab2SNaoya Horiguchi 168794ae8ba7SAneesh Kumar K.V if (!page) 16880c397daeSMichal Hocko page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL); 1689bf50bab2SNaoya Horiguchi 1690bf50bab2SNaoya Horiguchi return page; 1691bf50bab2SNaoya Horiguchi } 1692bf50bab2SNaoya Horiguchi 1693ab5ac90aSMichal Hocko /* page migration callback function */ 16943e59fcb0SMichal Hocko struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, 16953e59fcb0SMichal Hocko nodemask_t *nmask) 16964db9b2efSMichal Hocko { 1697aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 16984db9b2efSMichal Hocko 16994db9b2efSMichal Hocko spin_lock(&hugetlb_lock); 17004db9b2efSMichal Hocko if (h->free_huge_pages - h->resv_huge_pages > 0) { 17013e59fcb0SMichal Hocko struct page *page; 17023e59fcb0SMichal Hocko 17033e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); 17043e59fcb0SMichal Hocko if (page) { 17053e59fcb0SMichal Hocko spin_unlock(&hugetlb_lock); 17063e59fcb0SMichal Hocko return page; 17074db9b2efSMichal Hocko } 17084db9b2efSMichal Hocko } 17094db9b2efSMichal Hocko spin_unlock(&hugetlb_lock); 17104db9b2efSMichal Hocko 17110c397daeSMichal Hocko return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); 17124db9b2efSMichal Hocko } 17134db9b2efSMichal Hocko 1714ebd63723SMichal Hocko /* mempolicy aware migration callback */ 1715389c8178SMichal Hocko struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, 1716389c8178SMichal Hocko unsigned long address) 1717ebd63723SMichal Hocko { 1718ebd63723SMichal Hocko struct mempolicy *mpol; 1719ebd63723SMichal Hocko nodemask_t *nodemask; 1720ebd63723SMichal Hocko struct page *page; 1721ebd63723SMichal Hocko gfp_t gfp_mask; 1722ebd63723SMichal Hocko int node; 1723ebd63723SMichal Hocko 1724ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 1725ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 1726ebd63723SMichal Hocko page = alloc_huge_page_nodemask(h, node, nodemask); 1727ebd63723SMichal Hocko mpol_cond_put(mpol); 1728ebd63723SMichal Hocko 1729ebd63723SMichal Hocko return page; 1730ebd63723SMichal Hocko } 1731ebd63723SMichal Hocko 1732bf50bab2SNaoya Horiguchi /* 173325985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 1734e4e574b7SAdam Litke * of size 'delta'. 1735e4e574b7SAdam Litke */ 1736a5516438SAndi Kleen static int gather_surplus_pages(struct hstate *h, int delta) 1737e4e574b7SAdam Litke { 1738e4e574b7SAdam Litke struct list_head surplus_list; 1739e4e574b7SAdam Litke struct page *page, *tmp; 1740e4e574b7SAdam Litke int ret, i; 1741e4e574b7SAdam Litke int needed, allocated; 174228073b02SHillf Danton bool alloc_ok = true; 1743e4e574b7SAdam Litke 1744a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 1745ac09b3a1SAdam Litke if (needed <= 0) { 1746a5516438SAndi Kleen h->resv_huge_pages += delta; 1747e4e574b7SAdam Litke return 0; 1748ac09b3a1SAdam Litke } 1749e4e574b7SAdam Litke 1750e4e574b7SAdam Litke allocated = 0; 1751e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 1752e4e574b7SAdam Litke 1753e4e574b7SAdam Litke ret = -ENOMEM; 1754e4e574b7SAdam Litke retry: 1755e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 1756e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 17570c397daeSMichal Hocko page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), 1758aaf14e40SMichal Hocko NUMA_NO_NODE, NULL); 175928073b02SHillf Danton if (!page) { 176028073b02SHillf Danton alloc_ok = false; 176128073b02SHillf Danton break; 176228073b02SHillf Danton } 1763e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 176469ed779aSDavid Rientjes cond_resched(); 1765e4e574b7SAdam Litke } 176628073b02SHillf Danton allocated += i; 1767e4e574b7SAdam Litke 1768e4e574b7SAdam Litke /* 1769e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 1770e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 1771e4e574b7SAdam Litke */ 1772e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 1773a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 1774a5516438SAndi Kleen (h->free_huge_pages + allocated); 177528073b02SHillf Danton if (needed > 0) { 177628073b02SHillf Danton if (alloc_ok) 1777e4e574b7SAdam Litke goto retry; 177828073b02SHillf Danton /* 177928073b02SHillf Danton * We were not able to allocate enough pages to 178028073b02SHillf Danton * satisfy the entire reservation so we free what 178128073b02SHillf Danton * we've allocated so far. 178228073b02SHillf Danton */ 178328073b02SHillf Danton goto free; 178428073b02SHillf Danton } 1785e4e574b7SAdam Litke /* 1786e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 178725985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 1788e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 1789ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 1790ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 1791ac09b3a1SAdam Litke * before they are reserved. 1792e4e574b7SAdam Litke */ 1793e4e574b7SAdam Litke needed += allocated; 1794a5516438SAndi Kleen h->resv_huge_pages += delta; 1795e4e574b7SAdam Litke ret = 0; 1796a9869b83SNaoya Horiguchi 179719fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 179819fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 179919fc3f0aSAdam Litke if ((--needed) < 0) 180019fc3f0aSAdam Litke break; 1801a9869b83SNaoya Horiguchi /* 1802a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 1803a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 1804a9869b83SNaoya Horiguchi */ 1805a9869b83SNaoya Horiguchi put_page_testzero(page); 1806309381feSSasha Levin VM_BUG_ON_PAGE(page_count(page), page); 1807a5516438SAndi Kleen enqueue_huge_page(h, page); 180819fc3f0aSAdam Litke } 180928073b02SHillf Danton free: 1810b0365c8dSHillf Danton spin_unlock(&hugetlb_lock); 181119fc3f0aSAdam Litke 181219fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 1813c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 1814a9869b83SNaoya Horiguchi put_page(page); 181519fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 1816e4e574b7SAdam Litke 1817e4e574b7SAdam Litke return ret; 1818e4e574b7SAdam Litke } 1819e4e574b7SAdam Litke 1820e4e574b7SAdam Litke /* 1821e5bbc8a6SMike Kravetz * This routine has two main purposes: 1822e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 1823e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 1824e5bbc8a6SMike Kravetz * to the associated reservation map. 1825e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 1826e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 1827e5bbc8a6SMike Kravetz * 1828e5bbc8a6SMike Kravetz * Called with hugetlb_lock held. However, the lock could be dropped (and 1829e5bbc8a6SMike Kravetz * reacquired) during calls to cond_resched_lock. Whenever dropping the lock, 1830e5bbc8a6SMike Kravetz * we must make sure nobody else can claim pages we are in the process of 1831e5bbc8a6SMike Kravetz * freeing. Do this by ensuring resv_huge_page always is greater than the 1832e5bbc8a6SMike Kravetz * number of huge pages we plan to free when dropping the lock. 1833e4e574b7SAdam Litke */ 1834a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 1835a5516438SAndi Kleen unsigned long unused_resv_pages) 1836e4e574b7SAdam Litke { 1837e4e574b7SAdam Litke unsigned long nr_pages; 1838e4e574b7SAdam Litke 1839aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 1840bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1841e5bbc8a6SMike Kravetz goto out; 1842aa888a74SAndi Kleen 1843e5bbc8a6SMike Kravetz /* 1844e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 1845e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 1846e5bbc8a6SMike Kravetz */ 1847a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 1848e4e574b7SAdam Litke 1849685f3457SLee Schermerhorn /* 1850685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 18519b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 18529b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 18539b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 18549b5e5d0fSLee Schermerhorn * free_pool_huge_page() will balance the the freed pages across the 18559b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 1856e5bbc8a6SMike Kravetz * 1857e5bbc8a6SMike Kravetz * Note that we decrement resv_huge_pages as we free the pages. If 1858e5bbc8a6SMike Kravetz * we drop the lock, resv_huge_pages will still be sufficiently large 1859e5bbc8a6SMike Kravetz * to cover subsequent pages we may free. 1860685f3457SLee Schermerhorn */ 1861685f3457SLee Schermerhorn while (nr_pages--) { 1862e5bbc8a6SMike Kravetz h->resv_huge_pages--; 1863e5bbc8a6SMike Kravetz unused_resv_pages--; 18648cebfcd0SLai Jiangshan if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) 1865e5bbc8a6SMike Kravetz goto out; 18667848a4bfSMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 1867e4e574b7SAdam Litke } 1868e5bbc8a6SMike Kravetz 1869e5bbc8a6SMike Kravetz out: 1870e5bbc8a6SMike Kravetz /* Fully uncommit the reservation */ 1871e5bbc8a6SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 1872e4e574b7SAdam Litke } 1873e4e574b7SAdam Litke 18745e911373SMike Kravetz 1875c37f9fb1SAndy Whitcroft /* 1876feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 18775e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 1878cf3ad20bSMike Kravetz * 1879cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 1880cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 1881cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 1882cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 1883cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 1884feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 1885feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 1886feba16e2SMike Kravetz * is called in such cases. 1887cf3ad20bSMike Kravetz * 1888cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 1889cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 1890cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 1891cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 1892cf3ad20bSMike Kravetz * take appropriate action. 189396b96a96SMike Kravetz * 189496b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 189596b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 189696b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 189796b96a96SMike Kravetz * reservation exists. 1898c37f9fb1SAndy Whitcroft */ 18995e911373SMike Kravetz enum vma_resv_mode { 19005e911373SMike Kravetz VMA_NEEDS_RESV, 19015e911373SMike Kravetz VMA_COMMIT_RESV, 1902feba16e2SMike Kravetz VMA_END_RESV, 190396b96a96SMike Kravetz VMA_ADD_RESV, 19045e911373SMike Kravetz }; 1905cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 1906cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 19075e911373SMike Kravetz enum vma_resv_mode mode) 1908c37f9fb1SAndy Whitcroft { 19094e35f483SJoonsoo Kim struct resv_map *resv; 19104e35f483SJoonsoo Kim pgoff_t idx; 1911cf3ad20bSMike Kravetz long ret; 1912c37f9fb1SAndy Whitcroft 19134e35f483SJoonsoo Kim resv = vma_resv_map(vma); 19144e35f483SJoonsoo Kim if (!resv) 1915c37f9fb1SAndy Whitcroft return 1; 1916c37f9fb1SAndy Whitcroft 19174e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 19185e911373SMike Kravetz switch (mode) { 19195e911373SMike Kravetz case VMA_NEEDS_RESV: 1920cf3ad20bSMike Kravetz ret = region_chg(resv, idx, idx + 1); 19215e911373SMike Kravetz break; 19225e911373SMike Kravetz case VMA_COMMIT_RESV: 19235e911373SMike Kravetz ret = region_add(resv, idx, idx + 1); 19245e911373SMike Kravetz break; 1925feba16e2SMike Kravetz case VMA_END_RESV: 19265e911373SMike Kravetz region_abort(resv, idx, idx + 1); 19275e911373SMike Kravetz ret = 0; 19285e911373SMike Kravetz break; 192996b96a96SMike Kravetz case VMA_ADD_RESV: 193096b96a96SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) 193196b96a96SMike Kravetz ret = region_add(resv, idx, idx + 1); 193296b96a96SMike Kravetz else { 193396b96a96SMike Kravetz region_abort(resv, idx, idx + 1); 193496b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 193596b96a96SMike Kravetz } 193696b96a96SMike Kravetz break; 19375e911373SMike Kravetz default: 19385e911373SMike Kravetz BUG(); 19395e911373SMike Kravetz } 194084afd99bSAndy Whitcroft 19414e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) 1942cf3ad20bSMike Kravetz return ret; 194367961f9dSMike Kravetz else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { 194467961f9dSMike Kravetz /* 194567961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 194667961f9dSMike Kravetz * However, a file associated with mapping could have been 194767961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 194867961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 194967961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 195067961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 195167961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 195267961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 195367961f9dSMike Kravetz * reservation has already been consumed. As a result, the 195467961f9dSMike Kravetz * return value of this routine is the opposite of the 195567961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 195667961f9dSMike Kravetz */ 195767961f9dSMike Kravetz if (ret) 195867961f9dSMike Kravetz return 0; 195967961f9dSMike Kravetz else 196067961f9dSMike Kravetz return 1; 196167961f9dSMike Kravetz } 19624e35f483SJoonsoo Kim else 1963cf3ad20bSMike Kravetz return ret < 0 ? ret : 0; 196484afd99bSAndy Whitcroft } 1965cf3ad20bSMike Kravetz 1966cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 1967a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 1968c37f9fb1SAndy Whitcroft { 19695e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 1970cf3ad20bSMike Kravetz } 1971c37f9fb1SAndy Whitcroft 1972cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 1973cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 1974cf3ad20bSMike Kravetz { 19755e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 19765e911373SMike Kravetz } 19775e911373SMike Kravetz 1978feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 19795e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 19805e911373SMike Kravetz { 1981feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 1982c37f9fb1SAndy Whitcroft } 1983c37f9fb1SAndy Whitcroft 198496b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 198596b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 198696b96a96SMike Kravetz { 198796b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 198896b96a96SMike Kravetz } 198996b96a96SMike Kravetz 199096b96a96SMike Kravetz /* 199196b96a96SMike Kravetz * This routine is called to restore a reservation on error paths. In the 199296b96a96SMike Kravetz * specific error paths, a huge page was allocated (via alloc_huge_page) 199396b96a96SMike Kravetz * and is about to be freed. If a reservation for the page existed, 199496b96a96SMike Kravetz * alloc_huge_page would have consumed the reservation and set PagePrivate 199596b96a96SMike Kravetz * in the newly allocated page. When the page is freed via free_huge_page, 199696b96a96SMike Kravetz * the global reservation count will be incremented if PagePrivate is set. 199796b96a96SMike Kravetz * However, free_huge_page can not adjust the reserve map. Adjust the 199896b96a96SMike Kravetz * reserve map here to be consistent with global reserve count adjustments 199996b96a96SMike Kravetz * to be made by free_huge_page. 200096b96a96SMike Kravetz */ 200196b96a96SMike Kravetz static void restore_reserve_on_error(struct hstate *h, 200296b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long address, 200396b96a96SMike Kravetz struct page *page) 200496b96a96SMike Kravetz { 200596b96a96SMike Kravetz if (unlikely(PagePrivate(page))) { 200696b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 200796b96a96SMike Kravetz 200896b96a96SMike Kravetz if (unlikely(rc < 0)) { 200996b96a96SMike Kravetz /* 201096b96a96SMike Kravetz * Rare out of memory condition in reserve map 201196b96a96SMike Kravetz * manipulation. Clear PagePrivate so that 201296b96a96SMike Kravetz * global reserve count will not be incremented 201396b96a96SMike Kravetz * by free_huge_page. This will make it appear 201496b96a96SMike Kravetz * as though the reservation for this page was 201596b96a96SMike Kravetz * consumed. This may prevent the task from 201696b96a96SMike Kravetz * faulting in the page at a later time. This 201796b96a96SMike Kravetz * is better than inconsistent global huge page 201896b96a96SMike Kravetz * accounting of reserve counts. 201996b96a96SMike Kravetz */ 202096b96a96SMike Kravetz ClearPagePrivate(page); 202196b96a96SMike Kravetz } else if (rc) { 202296b96a96SMike Kravetz rc = vma_add_reservation(h, vma, address); 202396b96a96SMike Kravetz if (unlikely(rc < 0)) 202496b96a96SMike Kravetz /* 202596b96a96SMike Kravetz * See above comment about rare out of 202696b96a96SMike Kravetz * memory condition. 202796b96a96SMike Kravetz */ 202896b96a96SMike Kravetz ClearPagePrivate(page); 202996b96a96SMike Kravetz } else 203096b96a96SMike Kravetz vma_end_reservation(h, vma, address); 203196b96a96SMike Kravetz } 203296b96a96SMike Kravetz } 203396b96a96SMike Kravetz 203470c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 203504f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 2036348ea204SAdam Litke { 203790481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 2038a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2039348ea204SAdam Litke struct page *page; 2040d85f69b0SMike Kravetz long map_chg, map_commit; 2041d85f69b0SMike Kravetz long gbl_chg; 20426d76dcf4SAneesh Kumar K.V int ret, idx; 20436d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 20442fc39cecSAdam Litke 20456d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 2046a1e78772SMel Gorman /* 2047d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 2048d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 2049d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 2050a1e78772SMel Gorman */ 2051d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 2052d85f69b0SMike Kravetz if (map_chg < 0) 205376dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 2054d85f69b0SMike Kravetz 2055d85f69b0SMike Kravetz /* 2056d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 2057d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 2058d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 2059d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 2060d85f69b0SMike Kravetz * checked against any subpool limit. 2061d85f69b0SMike Kravetz */ 2062d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 2063d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 2064d85f69b0SMike Kravetz if (gbl_chg < 0) { 2065feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 206676dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 20675e911373SMike Kravetz } 206890d8b7e6SAdam Litke 2069d85f69b0SMike Kravetz /* 2070d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 2071d85f69b0SMike Kravetz * map, there could be reservations associated with the 2072d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 2073d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 2074d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 2075d85f69b0SMike Kravetz * the subpool reservations. 2076d85f69b0SMike Kravetz */ 2077d85f69b0SMike Kravetz if (avoid_reserve) 2078d85f69b0SMike Kravetz gbl_chg = 1; 2079d85f69b0SMike Kravetz } 2080d85f69b0SMike Kravetz 20816d76dcf4SAneesh Kumar K.V ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 20828f34af6fSJianyu Zhan if (ret) 20838f34af6fSJianyu Zhan goto out_subpool_put; 20848f34af6fSJianyu Zhan 2085a1e78772SMel Gorman spin_lock(&hugetlb_lock); 2086d85f69b0SMike Kravetz /* 2087d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 2088d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 2089d85f69b0SMike Kravetz * a reservation exists for the allocation. 2090d85f69b0SMike Kravetz */ 2091d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 209281a6fcaeSJoonsoo Kim if (!page) { 209394ae8ba7SAneesh Kumar K.V spin_unlock(&hugetlb_lock); 20940c397daeSMichal Hocko page = alloc_buddy_huge_page_with_mpol(h, vma, addr); 20958f34af6fSJianyu Zhan if (!page) 20968f34af6fSJianyu Zhan goto out_uncharge_cgroup; 2097a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 2098a88c7695SNaoya Horiguchi SetPagePrivate(page); 2099a88c7695SNaoya Horiguchi h->resv_huge_pages--; 2100a88c7695SNaoya Horiguchi } 210179dbb236SAneesh Kumar K.V spin_lock(&hugetlb_lock); 210279dbb236SAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 210381a6fcaeSJoonsoo Kim /* Fall through */ 2104a1e78772SMel Gorman } 210581a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 210681a6fcaeSJoonsoo Kim spin_unlock(&hugetlb_lock); 2107a1e78772SMel Gorman 210890481622SDavid Gibson set_page_private(page, (unsigned long)spool); 2109a1e78772SMel Gorman 2110d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 2111d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 211233039678SMike Kravetz /* 211333039678SMike Kravetz * The page was added to the reservation map between 211433039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 211533039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 211633039678SMike Kravetz * Adjust for the subpool count incremented above AND 211733039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 211833039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 211933039678SMike Kravetz * no longer applies. 212033039678SMike Kravetz */ 212133039678SMike Kravetz long rsv_adjust; 212233039678SMike Kravetz 212333039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 212433039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 212533039678SMike Kravetz } 21267893d1d5SAdam Litke return page; 21278f34af6fSJianyu Zhan 21288f34af6fSJianyu Zhan out_uncharge_cgroup: 21298f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 21308f34af6fSJianyu Zhan out_subpool_put: 2131d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 21328f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 2133feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 21348f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 2135b45b5bd6SDavid Gibson } 2136b45b5bd6SDavid Gibson 2137e24a1307SAneesh Kumar K.V int alloc_bootmem_huge_page(struct hstate *h) 2138e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 2139e24a1307SAneesh Kumar K.V int __alloc_bootmem_huge_page(struct hstate *h) 2140aa888a74SAndi Kleen { 2141aa888a74SAndi Kleen struct huge_bootmem_page *m; 2142b2261026SJoonsoo Kim int nr_nodes, node; 2143aa888a74SAndi Kleen 2144b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 2145aa888a74SAndi Kleen void *addr; 2146aa888a74SAndi Kleen 2147eb31d559SMike Rapoport addr = memblock_alloc_try_nid_raw( 21488b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 214997ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 2150aa888a74SAndi Kleen if (addr) { 2151aa888a74SAndi Kleen /* 2152aa888a74SAndi Kleen * Use the beginning of the huge page to store the 2153aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 2154aa888a74SAndi Kleen * puts them into the mem_map). 2155aa888a74SAndi Kleen */ 2156aa888a74SAndi Kleen m = addr; 2157aa888a74SAndi Kleen goto found; 2158aa888a74SAndi Kleen } 2159aa888a74SAndi Kleen } 2160aa888a74SAndi Kleen return 0; 2161aa888a74SAndi Kleen 2162aa888a74SAndi Kleen found: 2163df994eadSLuiz Capitulino BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); 2164aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 2165330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 2166aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 2167aa888a74SAndi Kleen m->hstate = h; 2168aa888a74SAndi Kleen return 1; 2169aa888a74SAndi Kleen } 2170aa888a74SAndi Kleen 2171d00181b9SKirill A. Shutemov static void __init prep_compound_huge_page(struct page *page, 2172d00181b9SKirill A. Shutemov unsigned int order) 217318229df5SAndy Whitcroft { 217418229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 217518229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 217618229df5SAndy Whitcroft else 217718229df5SAndy Whitcroft prep_compound_page(page, order); 217818229df5SAndy Whitcroft } 217918229df5SAndy Whitcroft 2180aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 2181aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 2182aa888a74SAndi Kleen { 2183aa888a74SAndi Kleen struct huge_bootmem_page *m; 2184aa888a74SAndi Kleen 2185aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 218640d18ebfSMike Kravetz struct page *page = virt_to_page(m); 2187aa888a74SAndi Kleen struct hstate *h = m->hstate; 2188ee8f248dSBecky Bruce 2189aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 219018229df5SAndy Whitcroft prep_compound_huge_page(page, h->order); 2191ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 2192aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 2193af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 2194af0fb9dfSMichal Hocko 2195b0320c7bSRafael Aquini /* 2196b0320c7bSRafael Aquini * If we had gigantic hugepages allocated at boot time, we need 2197b0320c7bSRafael Aquini * to restore the 'stolen' pages to totalram_pages in order to 2198b0320c7bSRafael Aquini * fix confusing memory reports from free(1) and another 2199b0320c7bSRafael Aquini * side-effects, like CommitLimit going negative. 2200b0320c7bSRafael Aquini */ 2201bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 22023dcc0571SJiang Liu adjust_managed_page_count(page, 1 << h->order); 2203520495feSCannon Matthews cond_resched(); 2204aa888a74SAndi Kleen } 2205aa888a74SAndi Kleen } 2206aa888a74SAndi Kleen 22078faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 22081da177e4SLinus Torvalds { 22091da177e4SLinus Torvalds unsigned long i; 22101da177e4SLinus Torvalds 2211e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 2212bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 2213aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 2214aa888a74SAndi Kleen break; 22150c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 22168cebfcd0SLai Jiangshan &node_states[N_MEMORY])) 22171da177e4SLinus Torvalds break; 221869ed779aSDavid Rientjes cond_resched(); 22191da177e4SLinus Torvalds } 2220d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 2221d715cf80SLiam R. Howlett char buf[32]; 2222d715cf80SLiam R. Howlett 2223c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2224d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 2225d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 22268faa8b07SAndi Kleen h->max_huge_pages = i; 2227e5ff2159SAndi Kleen } 2228d715cf80SLiam R. Howlett } 2229e5ff2159SAndi Kleen 2230e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 2231e5ff2159SAndi Kleen { 2232e5ff2159SAndi Kleen struct hstate *h; 2233e5ff2159SAndi Kleen 2234e5ff2159SAndi Kleen for_each_hstate(h) { 2235641844f5SNaoya Horiguchi if (minimum_order > huge_page_order(h)) 2236641844f5SNaoya Horiguchi minimum_order = huge_page_order(h); 2237641844f5SNaoya Horiguchi 22388faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 2239bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 22408faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 2241e5ff2159SAndi Kleen } 2242641844f5SNaoya Horiguchi VM_BUG_ON(minimum_order == UINT_MAX); 2243e5ff2159SAndi Kleen } 2244e5ff2159SAndi Kleen 2245e5ff2159SAndi Kleen static void __init report_hugepages(void) 2246e5ff2159SAndi Kleen { 2247e5ff2159SAndi Kleen struct hstate *h; 2248e5ff2159SAndi Kleen 2249e5ff2159SAndi Kleen for_each_hstate(h) { 22504abd32dbSAndi Kleen char buf[32]; 2251c6247f72SMatthew Wilcox 2252c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2253ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 2254c6247f72SMatthew Wilcox buf, h->free_huge_pages); 2255e5ff2159SAndi Kleen } 2256e5ff2159SAndi Kleen } 2257e5ff2159SAndi Kleen 22581da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 22596ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 22606ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 22611da177e4SLinus Torvalds { 22624415cc8dSChristoph Lameter int i; 22634415cc8dSChristoph Lameter 2264bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2265aa888a74SAndi Kleen return; 2266aa888a74SAndi Kleen 22676ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 22681da177e4SLinus Torvalds struct page *page, *next; 2269a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 2270a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 2271a5516438SAndi Kleen if (count >= h->nr_huge_pages) 22726b0c880dSAdam Litke return; 22731da177e4SLinus Torvalds if (PageHighMem(page)) 22741da177e4SLinus Torvalds continue; 22751da177e4SLinus Torvalds list_del(&page->lru); 2276e5ff2159SAndi Kleen update_and_free_page(h, page); 2277a5516438SAndi Kleen h->free_huge_pages--; 2278a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 22791da177e4SLinus Torvalds } 22801da177e4SLinus Torvalds } 22811da177e4SLinus Torvalds } 22821da177e4SLinus Torvalds #else 22836ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 22846ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 22851da177e4SLinus Torvalds { 22861da177e4SLinus Torvalds } 22871da177e4SLinus Torvalds #endif 22881da177e4SLinus Torvalds 228920a0307cSWu Fengguang /* 229020a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 229120a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 229220a0307cSWu Fengguang * Returns 1 if an adjustment was made. 229320a0307cSWu Fengguang */ 22946ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 22956ae11b27SLee Schermerhorn int delta) 229620a0307cSWu Fengguang { 2297b2261026SJoonsoo Kim int nr_nodes, node; 229820a0307cSWu Fengguang 229920a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 230020a0307cSWu Fengguang 2301e8c5c824SLee Schermerhorn if (delta < 0) { 2302b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 2303b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 2304b2261026SJoonsoo Kim goto found; 2305b2261026SJoonsoo Kim } 2306b2261026SJoonsoo Kim } else { 2307b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2308b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 2309b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 2310b2261026SJoonsoo Kim goto found; 2311e8c5c824SLee Schermerhorn } 23129a76db09SLee Schermerhorn } 2313b2261026SJoonsoo Kim return 0; 231420a0307cSWu Fengguang 2315b2261026SJoonsoo Kim found: 231620a0307cSWu Fengguang h->surplus_huge_pages += delta; 2317b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 2318b2261026SJoonsoo Kim return 1; 231920a0307cSWu Fengguang } 232020a0307cSWu Fengguang 2321a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 2322fd875dcaSMike Kravetz static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 23236ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 23241da177e4SLinus Torvalds { 23257893d1d5SAdam Litke unsigned long min_count, ret; 23261da177e4SLinus Torvalds 23274eb0716eSAlexandre Ghiti spin_lock(&hugetlb_lock); 23284eb0716eSAlexandre Ghiti 23294eb0716eSAlexandre Ghiti /* 2330fd875dcaSMike Kravetz * Check for a node specific request. 2331fd875dcaSMike Kravetz * Changing node specific huge page count may require a corresponding 2332fd875dcaSMike Kravetz * change to the global count. In any case, the passed node mask 2333fd875dcaSMike Kravetz * (nodes_allowed) will restrict alloc/free to the specified node. 2334fd875dcaSMike Kravetz */ 2335fd875dcaSMike Kravetz if (nid != NUMA_NO_NODE) { 2336fd875dcaSMike Kravetz unsigned long old_count = count; 2337fd875dcaSMike Kravetz 2338fd875dcaSMike Kravetz count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 2339fd875dcaSMike Kravetz /* 2340fd875dcaSMike Kravetz * User may have specified a large count value which caused the 2341fd875dcaSMike Kravetz * above calculation to overflow. In this case, they wanted 2342fd875dcaSMike Kravetz * to allocate as many huge pages as possible. Set count to 2343fd875dcaSMike Kravetz * largest possible value to align with their intention. 2344fd875dcaSMike Kravetz */ 2345fd875dcaSMike Kravetz if (count < old_count) 2346fd875dcaSMike Kravetz count = ULONG_MAX; 2347fd875dcaSMike Kravetz } 2348fd875dcaSMike Kravetz 2349fd875dcaSMike Kravetz /* 23504eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 23514eb0716eSAlexandre Ghiti * page range allocation. 23524eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 23534eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 23544eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 23554eb0716eSAlexandre Ghiti */ 23564eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 23574eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 23584eb0716eSAlexandre Ghiti spin_unlock(&hugetlb_lock); 23594eb0716eSAlexandre Ghiti return -EINVAL; 23604eb0716eSAlexandre Ghiti } 23614eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 23624eb0716eSAlexandre Ghiti } 2363aa888a74SAndi Kleen 23647893d1d5SAdam Litke /* 23657893d1d5SAdam Litke * Increase the pool size 23667893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 23677893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 2368d1c3fb1fSNishanth Aravamudan * 23690c397daeSMichal Hocko * We might race with alloc_surplus_huge_page() here and be unable 2370d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 2371d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 2372d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 2373d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 23747893d1d5SAdam Litke */ 2375a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 23766ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 23777893d1d5SAdam Litke break; 23787893d1d5SAdam Litke } 23797893d1d5SAdam Litke 2380a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 23817893d1d5SAdam Litke /* 23827893d1d5SAdam Litke * If this allocation races such that we no longer need the 23837893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 23847893d1d5SAdam Litke * and reducing the surplus. 23857893d1d5SAdam Litke */ 23867893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 2387649920c6SJia He 2388649920c6SJia He /* yield cpu to avoid soft lockup */ 2389649920c6SJia He cond_resched(); 2390649920c6SJia He 23910c397daeSMichal Hocko ret = alloc_pool_huge_page(h, nodes_allowed); 23927893d1d5SAdam Litke spin_lock(&hugetlb_lock); 23937893d1d5SAdam Litke if (!ret) 23947893d1d5SAdam Litke goto out; 23957893d1d5SAdam Litke 2396536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 2397536240f2SMel Gorman if (signal_pending(current)) 2398536240f2SMel Gorman goto out; 23997893d1d5SAdam Litke } 24007893d1d5SAdam Litke 24017893d1d5SAdam Litke /* 24027893d1d5SAdam Litke * Decrease the pool size 24037893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 24047893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 24057893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 24067893d1d5SAdam Litke * to the desired size as pages become free. 2407d1c3fb1fSNishanth Aravamudan * 2408d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 2409d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 2410d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 24110c397daeSMichal Hocko * alloc_surplus_huge_page() is checking the global counter, 2412d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 2413d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 2414d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 24157893d1d5SAdam Litke */ 2416a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 24176b0c880dSAdam Litke min_count = max(count, min_count); 24186ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 2419a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 24206ae11b27SLee Schermerhorn if (!free_pool_huge_page(h, nodes_allowed, 0)) 24211da177e4SLinus Torvalds break; 242255f67141SMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 24231da177e4SLinus Torvalds } 2424a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 24256ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 24267893d1d5SAdam Litke break; 24277893d1d5SAdam Litke } 24287893d1d5SAdam Litke out: 24294eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 24301da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 24314eb0716eSAlexandre Ghiti 24324eb0716eSAlexandre Ghiti return 0; 24331da177e4SLinus Torvalds } 24341da177e4SLinus Torvalds 2435a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 2436a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 2437a3437870SNishanth Aravamudan 2438a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 2439a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 2440a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 2441a3437870SNishanth Aravamudan 2442a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 2443a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 2444a3437870SNishanth Aravamudan 24459a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 24469a305230SLee Schermerhorn 24479a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 2448a3437870SNishanth Aravamudan { 2449a3437870SNishanth Aravamudan int i; 24509a305230SLee Schermerhorn 2451a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 24529a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 24539a305230SLee Schermerhorn if (nidp) 24549a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 2455a3437870SNishanth Aravamudan return &hstates[i]; 24569a305230SLee Schermerhorn } 24579a305230SLee Schermerhorn 24589a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 2459a3437870SNishanth Aravamudan } 2460a3437870SNishanth Aravamudan 246106808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 2462a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2463a3437870SNishanth Aravamudan { 24649a305230SLee Schermerhorn struct hstate *h; 24659a305230SLee Schermerhorn unsigned long nr_huge_pages; 24669a305230SLee Schermerhorn int nid; 24679a305230SLee Schermerhorn 24689a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 24699a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 24709a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 24719a305230SLee Schermerhorn else 24729a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 24739a305230SLee Schermerhorn 24749a305230SLee Schermerhorn return sprintf(buf, "%lu\n", nr_huge_pages); 2475a3437870SNishanth Aravamudan } 2476adbe8726SEric B Munson 2477238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 2478238d3c13SDavid Rientjes struct hstate *h, int nid, 2479238d3c13SDavid Rientjes unsigned long count, size_t len) 2480a3437870SNishanth Aravamudan { 2481a3437870SNishanth Aravamudan int err; 24822d0adf7eSOscar Salvador nodemask_t nodes_allowed, *n_mask; 2483a3437870SNishanth Aravamudan 24842d0adf7eSOscar Salvador if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 24852d0adf7eSOscar Salvador return -EINVAL; 2486adbe8726SEric B Munson 24879a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 24889a305230SLee Schermerhorn /* 24899a305230SLee Schermerhorn * global hstate attribute 24909a305230SLee Schermerhorn */ 24919a305230SLee Schermerhorn if (!(obey_mempolicy && 24922d0adf7eSOscar Salvador init_nodemask_of_mempolicy(&nodes_allowed))) 24932d0adf7eSOscar Salvador n_mask = &node_states[N_MEMORY]; 24942d0adf7eSOscar Salvador else 24952d0adf7eSOscar Salvador n_mask = &nodes_allowed; 24962d0adf7eSOscar Salvador } else { 24979a305230SLee Schermerhorn /* 2498fd875dcaSMike Kravetz * Node specific request. count adjustment happens in 2499fd875dcaSMike Kravetz * set_max_huge_pages() after acquiring hugetlb_lock. 25009a305230SLee Schermerhorn */ 25012d0adf7eSOscar Salvador init_nodemask_of_node(&nodes_allowed, nid); 25022d0adf7eSOscar Salvador n_mask = &nodes_allowed; 2503fd875dcaSMike Kravetz } 25049a305230SLee Schermerhorn 25052d0adf7eSOscar Salvador err = set_max_huge_pages(h, count, nid, n_mask); 250606808b08SLee Schermerhorn 25074eb0716eSAlexandre Ghiti return err ? err : len; 250806808b08SLee Schermerhorn } 250906808b08SLee Schermerhorn 2510238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 2511238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 2512238d3c13SDavid Rientjes size_t len) 2513238d3c13SDavid Rientjes { 2514238d3c13SDavid Rientjes struct hstate *h; 2515238d3c13SDavid Rientjes unsigned long count; 2516238d3c13SDavid Rientjes int nid; 2517238d3c13SDavid Rientjes int err; 2518238d3c13SDavid Rientjes 2519238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 2520238d3c13SDavid Rientjes if (err) 2521238d3c13SDavid Rientjes return err; 2522238d3c13SDavid Rientjes 2523238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 2524238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 2525238d3c13SDavid Rientjes } 2526238d3c13SDavid Rientjes 252706808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 252806808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 252906808b08SLee Schermerhorn { 253006808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 253106808b08SLee Schermerhorn } 253206808b08SLee Schermerhorn 253306808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 253406808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 253506808b08SLee Schermerhorn { 2536238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 2537a3437870SNishanth Aravamudan } 2538a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 2539a3437870SNishanth Aravamudan 254006808b08SLee Schermerhorn #ifdef CONFIG_NUMA 254106808b08SLee Schermerhorn 254206808b08SLee Schermerhorn /* 254306808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 254406808b08SLee Schermerhorn * huge page alloc/free. 254506808b08SLee Schermerhorn */ 254606808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 254706808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 254806808b08SLee Schermerhorn { 254906808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 255006808b08SLee Schermerhorn } 255106808b08SLee Schermerhorn 255206808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 255306808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 255406808b08SLee Schermerhorn { 2555238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 255606808b08SLee Schermerhorn } 255706808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 255806808b08SLee Schermerhorn #endif 255906808b08SLee Schermerhorn 256006808b08SLee Schermerhorn 2561a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 2562a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2563a3437870SNishanth Aravamudan { 25649a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2565a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 2566a3437870SNishanth Aravamudan } 2567adbe8726SEric B Munson 2568a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 2569a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 2570a3437870SNishanth Aravamudan { 2571a3437870SNishanth Aravamudan int err; 2572a3437870SNishanth Aravamudan unsigned long input; 25739a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2574a3437870SNishanth Aravamudan 2575bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2576adbe8726SEric B Munson return -EINVAL; 2577adbe8726SEric B Munson 25783dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 2579a3437870SNishanth Aravamudan if (err) 258073ae31e5SEric B Munson return err; 2581a3437870SNishanth Aravamudan 2582a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 2583a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 2584a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 2585a3437870SNishanth Aravamudan 2586a3437870SNishanth Aravamudan return count; 2587a3437870SNishanth Aravamudan } 2588a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 2589a3437870SNishanth Aravamudan 2590a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 2591a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2592a3437870SNishanth Aravamudan { 25939a305230SLee Schermerhorn struct hstate *h; 25949a305230SLee Schermerhorn unsigned long free_huge_pages; 25959a305230SLee Schermerhorn int nid; 25969a305230SLee Schermerhorn 25979a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 25989a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 25999a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 26009a305230SLee Schermerhorn else 26019a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 26029a305230SLee Schermerhorn 26039a305230SLee Schermerhorn return sprintf(buf, "%lu\n", free_huge_pages); 2604a3437870SNishanth Aravamudan } 2605a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 2606a3437870SNishanth Aravamudan 2607a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 2608a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2609a3437870SNishanth Aravamudan { 26109a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2611a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->resv_huge_pages); 2612a3437870SNishanth Aravamudan } 2613a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 2614a3437870SNishanth Aravamudan 2615a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 2616a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2617a3437870SNishanth Aravamudan { 26189a305230SLee Schermerhorn struct hstate *h; 26199a305230SLee Schermerhorn unsigned long surplus_huge_pages; 26209a305230SLee Schermerhorn int nid; 26219a305230SLee Schermerhorn 26229a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 26239a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 26249a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 26259a305230SLee Schermerhorn else 26269a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 26279a305230SLee Schermerhorn 26289a305230SLee Schermerhorn return sprintf(buf, "%lu\n", surplus_huge_pages); 2629a3437870SNishanth Aravamudan } 2630a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 2631a3437870SNishanth Aravamudan 2632a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 2633a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 2634a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 2635a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 2636a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 2637a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 263806808b08SLee Schermerhorn #ifdef CONFIG_NUMA 263906808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 264006808b08SLee Schermerhorn #endif 2641a3437870SNishanth Aravamudan NULL, 2642a3437870SNishanth Aravamudan }; 2643a3437870SNishanth Aravamudan 264467e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 2645a3437870SNishanth Aravamudan .attrs = hstate_attrs, 2646a3437870SNishanth Aravamudan }; 2647a3437870SNishanth Aravamudan 2648094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 26499a305230SLee Schermerhorn struct kobject **hstate_kobjs, 265067e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 2651a3437870SNishanth Aravamudan { 2652a3437870SNishanth Aravamudan int retval; 2653972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 2654a3437870SNishanth Aravamudan 26559a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 26569a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 2657a3437870SNishanth Aravamudan return -ENOMEM; 2658a3437870SNishanth Aravamudan 26599a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 2660a3437870SNishanth Aravamudan if (retval) 26619a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 2662a3437870SNishanth Aravamudan 2663a3437870SNishanth Aravamudan return retval; 2664a3437870SNishanth Aravamudan } 2665a3437870SNishanth Aravamudan 2666a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 2667a3437870SNishanth Aravamudan { 2668a3437870SNishanth Aravamudan struct hstate *h; 2669a3437870SNishanth Aravamudan int err; 2670a3437870SNishanth Aravamudan 2671a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 2672a3437870SNishanth Aravamudan if (!hugepages_kobj) 2673a3437870SNishanth Aravamudan return; 2674a3437870SNishanth Aravamudan 2675a3437870SNishanth Aravamudan for_each_hstate(h) { 26769a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 26779a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 2678a3437870SNishanth Aravamudan if (err) 2679ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s", h->name); 2680a3437870SNishanth Aravamudan } 2681a3437870SNishanth Aravamudan } 2682a3437870SNishanth Aravamudan 26839a305230SLee Schermerhorn #ifdef CONFIG_NUMA 26849a305230SLee Schermerhorn 26859a305230SLee Schermerhorn /* 26869a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 268710fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 268810fbcf4cSKay Sievers * index of a node device or _hstate == node id. 268910fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 26909a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 26919a305230SLee Schermerhorn */ 26929a305230SLee Schermerhorn struct node_hstate { 26939a305230SLee Schermerhorn struct kobject *hugepages_kobj; 26949a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 26959a305230SLee Schermerhorn }; 2696b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 26979a305230SLee Schermerhorn 26989a305230SLee Schermerhorn /* 269910fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 27009a305230SLee Schermerhorn */ 27019a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 27029a305230SLee Schermerhorn &nr_hugepages_attr.attr, 27039a305230SLee Schermerhorn &free_hugepages_attr.attr, 27049a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 27059a305230SLee Schermerhorn NULL, 27069a305230SLee Schermerhorn }; 27079a305230SLee Schermerhorn 270867e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 27099a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 27109a305230SLee Schermerhorn }; 27119a305230SLee Schermerhorn 27129a305230SLee Schermerhorn /* 271310fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 27149a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 27159a305230SLee Schermerhorn */ 27169a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 27179a305230SLee Schermerhorn { 27189a305230SLee Schermerhorn int nid; 27199a305230SLee Schermerhorn 27209a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 27219a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 27229a305230SLee Schermerhorn int i; 27239a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 27249a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 27259a305230SLee Schermerhorn if (nidp) 27269a305230SLee Schermerhorn *nidp = nid; 27279a305230SLee Schermerhorn return &hstates[i]; 27289a305230SLee Schermerhorn } 27299a305230SLee Schermerhorn } 27309a305230SLee Schermerhorn 27319a305230SLee Schermerhorn BUG(); 27329a305230SLee Schermerhorn return NULL; 27339a305230SLee Schermerhorn } 27349a305230SLee Schermerhorn 27359a305230SLee Schermerhorn /* 273610fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 27379a305230SLee Schermerhorn * No-op if no hstate attributes attached. 27389a305230SLee Schermerhorn */ 27393cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 27409a305230SLee Schermerhorn { 27419a305230SLee Schermerhorn struct hstate *h; 274210fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 27439a305230SLee Schermerhorn 27449a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 27459b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 27469a305230SLee Schermerhorn 2747972dc4deSAneesh Kumar K.V for_each_hstate(h) { 2748972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 2749972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 2750972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 2751972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 2752972dc4deSAneesh Kumar K.V } 27539a305230SLee Schermerhorn } 27549a305230SLee Schermerhorn 27559a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 27569a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 27579a305230SLee Schermerhorn } 27589a305230SLee Schermerhorn 27599a305230SLee Schermerhorn 27609a305230SLee Schermerhorn /* 276110fbcf4cSKay Sievers * Register hstate attributes for a single node device. 27629a305230SLee Schermerhorn * No-op if attributes already registered. 27639a305230SLee Schermerhorn */ 27643cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 27659a305230SLee Schermerhorn { 27669a305230SLee Schermerhorn struct hstate *h; 276710fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 27689a305230SLee Schermerhorn int err; 27699a305230SLee Schermerhorn 27709a305230SLee Schermerhorn if (nhs->hugepages_kobj) 27719a305230SLee Schermerhorn return; /* already allocated */ 27729a305230SLee Schermerhorn 27739a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 277410fbcf4cSKay Sievers &node->dev.kobj); 27759a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 27769a305230SLee Schermerhorn return; 27779a305230SLee Schermerhorn 27789a305230SLee Schermerhorn for_each_hstate(h) { 27799a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 27809a305230SLee Schermerhorn nhs->hstate_kobjs, 27819a305230SLee Schermerhorn &per_node_hstate_attr_group); 27829a305230SLee Schermerhorn if (err) { 2783ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s for node %d\n", 278410fbcf4cSKay Sievers h->name, node->dev.id); 27859a305230SLee Schermerhorn hugetlb_unregister_node(node); 27869a305230SLee Schermerhorn break; 27879a305230SLee Schermerhorn } 27889a305230SLee Schermerhorn } 27899a305230SLee Schermerhorn } 27909a305230SLee Schermerhorn 27919a305230SLee Schermerhorn /* 27929b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 279310fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 279410fbcf4cSKay Sievers * registered their associated device by this time. 27959a305230SLee Schermerhorn */ 27967d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 27979a305230SLee Schermerhorn { 27989a305230SLee Schermerhorn int nid; 27999a305230SLee Schermerhorn 28008cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 28018732794bSWen Congyang struct node *node = node_devices[nid]; 280210fbcf4cSKay Sievers if (node->dev.id == nid) 28039a305230SLee Schermerhorn hugetlb_register_node(node); 28049a305230SLee Schermerhorn } 28059a305230SLee Schermerhorn 28069a305230SLee Schermerhorn /* 280710fbcf4cSKay Sievers * Let the node device driver know we're here so it can 28089a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 28099a305230SLee Schermerhorn */ 28109a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 28119a305230SLee Schermerhorn hugetlb_unregister_node); 28129a305230SLee Schermerhorn } 28139a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 28149a305230SLee Schermerhorn 28159a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 28169a305230SLee Schermerhorn { 28179a305230SLee Schermerhorn BUG(); 28189a305230SLee Schermerhorn if (nidp) 28199a305230SLee Schermerhorn *nidp = -1; 28209a305230SLee Schermerhorn return NULL; 28219a305230SLee Schermerhorn } 28229a305230SLee Schermerhorn 28239a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 28249a305230SLee Schermerhorn 28259a305230SLee Schermerhorn #endif 28269a305230SLee Schermerhorn 2827a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 2828a3437870SNishanth Aravamudan { 28298382d914SDavidlohr Bueso int i; 28308382d914SDavidlohr Bueso 2831457c1b27SNishanth Aravamudan if (!hugepages_supported()) 28320ef89d25SBenjamin Herrenschmidt return 0; 2833a3437870SNishanth Aravamudan 2834e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) { 2835d715cf80SLiam R. Howlett if (default_hstate_size != 0) { 2836d715cf80SLiam R. Howlett pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n", 2837d715cf80SLiam R. Howlett default_hstate_size, HPAGE_SIZE); 2838d715cf80SLiam R. Howlett } 2839d715cf80SLiam R. Howlett 2840e11bfbfcSNick Piggin default_hstate_size = HPAGE_SIZE; 2841e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) 2842a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 2843a3437870SNishanth Aravamudan } 2844972dc4deSAneesh Kumar K.V default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); 2845f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 2846f8b74815SVaishali Thakkar if (!default_hstate.max_huge_pages) 2847e11bfbfcSNick Piggin default_hstate.max_huge_pages = default_hstate_max_huge_pages; 2848f8b74815SVaishali Thakkar } 2849a3437870SNishanth Aravamudan 2850a3437870SNishanth Aravamudan hugetlb_init_hstates(); 2851aa888a74SAndi Kleen gather_bootmem_prealloc(); 2852a3437870SNishanth Aravamudan report_hugepages(); 2853a3437870SNishanth Aravamudan 2854a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 28559a305230SLee Schermerhorn hugetlb_register_all_nodes(); 28567179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 28579a305230SLee Schermerhorn 28588382d914SDavidlohr Bueso #ifdef CONFIG_SMP 28598382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 28608382d914SDavidlohr Bueso #else 28618382d914SDavidlohr Bueso num_fault_mutexes = 1; 28628382d914SDavidlohr Bueso #endif 2863c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 28646da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 28656da2ec56SKees Cook GFP_KERNEL); 2866c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 28678382d914SDavidlohr Bueso 28688382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 2869c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 2870a3437870SNishanth Aravamudan return 0; 2871a3437870SNishanth Aravamudan } 28723e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 2873a3437870SNishanth Aravamudan 2874a3437870SNishanth Aravamudan /* Should be called on processing a hugepagesz=... option */ 28759fee021dSVaishali Thakkar void __init hugetlb_bad_size(void) 28769fee021dSVaishali Thakkar { 28779fee021dSVaishali Thakkar parsed_valid_hugepagesz = false; 28789fee021dSVaishali Thakkar } 28799fee021dSVaishali Thakkar 2880d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 2881a3437870SNishanth Aravamudan { 2882a3437870SNishanth Aravamudan struct hstate *h; 28838faa8b07SAndi Kleen unsigned long i; 28848faa8b07SAndi Kleen 2885a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 2886598d8091SJoe Perches pr_warn("hugepagesz= specified twice, ignoring\n"); 2887a3437870SNishanth Aravamudan return; 2888a3437870SNishanth Aravamudan } 288947d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 2890a3437870SNishanth Aravamudan BUG_ON(order == 0); 289147d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 2892a3437870SNishanth Aravamudan h->order = order; 2893a3437870SNishanth Aravamudan h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 28948faa8b07SAndi Kleen h->nr_huge_pages = 0; 28958faa8b07SAndi Kleen h->free_huge_pages = 0; 28968faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 28978faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 28980edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 289954f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 290054f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 2901a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 2902a3437870SNishanth Aravamudan huge_page_size(h)/1024); 29038faa8b07SAndi Kleen 2904a3437870SNishanth Aravamudan parsed_hstate = h; 2905a3437870SNishanth Aravamudan } 2906a3437870SNishanth Aravamudan 2907e11bfbfcSNick Piggin static int __init hugetlb_nrpages_setup(char *s) 2908a3437870SNishanth Aravamudan { 2909a3437870SNishanth Aravamudan unsigned long *mhp; 29108faa8b07SAndi Kleen static unsigned long *last_mhp; 2911a3437870SNishanth Aravamudan 29129fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 29139fee021dSVaishali Thakkar pr_warn("hugepages = %s preceded by " 29149fee021dSVaishali Thakkar "an unsupported hugepagesz, ignoring\n", s); 29159fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 29169fee021dSVaishali Thakkar return 1; 29179fee021dSVaishali Thakkar } 2918a3437870SNishanth Aravamudan /* 291947d38344SAneesh Kumar K.V * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, 2920a3437870SNishanth Aravamudan * so this hugepages= parameter goes to the "default hstate". 2921a3437870SNishanth Aravamudan */ 29229fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 2923a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 2924a3437870SNishanth Aravamudan else 2925a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 2926a3437870SNishanth Aravamudan 29278faa8b07SAndi Kleen if (mhp == last_mhp) { 2928598d8091SJoe Perches pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n"); 29298faa8b07SAndi Kleen return 1; 29308faa8b07SAndi Kleen } 29318faa8b07SAndi Kleen 2932a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 2933a3437870SNishanth Aravamudan *mhp = 0; 2934a3437870SNishanth Aravamudan 29358faa8b07SAndi Kleen /* 29368faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 29378faa8b07SAndi Kleen * But we need to allocate >= MAX_ORDER hstates here early to still 29388faa8b07SAndi Kleen * use the bootmem allocator. 29398faa8b07SAndi Kleen */ 294047d38344SAneesh Kumar K.V if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) 29418faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 29428faa8b07SAndi Kleen 29438faa8b07SAndi Kleen last_mhp = mhp; 29448faa8b07SAndi Kleen 2945a3437870SNishanth Aravamudan return 1; 2946a3437870SNishanth Aravamudan } 2947e11bfbfcSNick Piggin __setup("hugepages=", hugetlb_nrpages_setup); 2948e11bfbfcSNick Piggin 2949e11bfbfcSNick Piggin static int __init hugetlb_default_setup(char *s) 2950e11bfbfcSNick Piggin { 2951e11bfbfcSNick Piggin default_hstate_size = memparse(s, &s); 2952e11bfbfcSNick Piggin return 1; 2953e11bfbfcSNick Piggin } 2954e11bfbfcSNick Piggin __setup("default_hugepagesz=", hugetlb_default_setup); 2955a3437870SNishanth Aravamudan 29568a213460SNishanth Aravamudan static unsigned int cpuset_mems_nr(unsigned int *array) 29578a213460SNishanth Aravamudan { 29588a213460SNishanth Aravamudan int node; 29598a213460SNishanth Aravamudan unsigned int nr = 0; 29608a213460SNishanth Aravamudan 29618a213460SNishanth Aravamudan for_each_node_mask(node, cpuset_current_mems_allowed) 29628a213460SNishanth Aravamudan nr += array[node]; 29638a213460SNishanth Aravamudan 29648a213460SNishanth Aravamudan return nr; 29658a213460SNishanth Aravamudan } 29668a213460SNishanth Aravamudan 29678a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 296806808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 296906808b08SLee Schermerhorn struct ctl_table *table, int write, 297006808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 29711da177e4SLinus Torvalds { 2972e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 2973238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 297408d4a246SMichal Hocko int ret; 2975e5ff2159SAndi Kleen 2976457c1b27SNishanth Aravamudan if (!hugepages_supported()) 297786613628SJan Stancek return -EOPNOTSUPP; 2978457c1b27SNishanth Aravamudan 2979e5ff2159SAndi Kleen table->data = &tmp; 2980e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 298108d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 298208d4a246SMichal Hocko if (ret) 298308d4a246SMichal Hocko goto out; 2984e5ff2159SAndi Kleen 2985238d3c13SDavid Rientjes if (write) 2986238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 2987238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 298808d4a246SMichal Hocko out: 298908d4a246SMichal Hocko return ret; 29901da177e4SLinus Torvalds } 2991396faf03SMel Gorman 299206808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 299306808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 299406808b08SLee Schermerhorn { 299506808b08SLee Schermerhorn 299606808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 299706808b08SLee Schermerhorn buffer, length, ppos); 299806808b08SLee Schermerhorn } 299906808b08SLee Schermerhorn 300006808b08SLee Schermerhorn #ifdef CONFIG_NUMA 300106808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 300206808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 300306808b08SLee Schermerhorn { 300406808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 300506808b08SLee Schermerhorn buffer, length, ppos); 300606808b08SLee Schermerhorn } 300706808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 300806808b08SLee Schermerhorn 3009a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 30108d65af78SAlexey Dobriyan void __user *buffer, 3011a3d0c6aaSNishanth Aravamudan size_t *length, loff_t *ppos) 3012a3d0c6aaSNishanth Aravamudan { 3013a5516438SAndi Kleen struct hstate *h = &default_hstate; 3014e5ff2159SAndi Kleen unsigned long tmp; 301508d4a246SMichal Hocko int ret; 3016e5ff2159SAndi Kleen 3017457c1b27SNishanth Aravamudan if (!hugepages_supported()) 301886613628SJan Stancek return -EOPNOTSUPP; 3019457c1b27SNishanth Aravamudan 3020e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 3021e5ff2159SAndi Kleen 3022bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 3023adbe8726SEric B Munson return -EINVAL; 3024adbe8726SEric B Munson 3025e5ff2159SAndi Kleen table->data = &tmp; 3026e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 302708d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 302808d4a246SMichal Hocko if (ret) 302908d4a246SMichal Hocko goto out; 3030e5ff2159SAndi Kleen 3031e5ff2159SAndi Kleen if (write) { 3032064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 3033e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 3034a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 3035e5ff2159SAndi Kleen } 303608d4a246SMichal Hocko out: 303708d4a246SMichal Hocko return ret; 3038a3d0c6aaSNishanth Aravamudan } 3039a3d0c6aaSNishanth Aravamudan 30401da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 30411da177e4SLinus Torvalds 3042e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 30431da177e4SLinus Torvalds { 3044fcb2b0c5SRoman Gushchin struct hstate *h; 3045fcb2b0c5SRoman Gushchin unsigned long total = 0; 3046fcb2b0c5SRoman Gushchin 3047457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3048457c1b27SNishanth Aravamudan return; 3049fcb2b0c5SRoman Gushchin 3050fcb2b0c5SRoman Gushchin for_each_hstate(h) { 3051fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 3052fcb2b0c5SRoman Gushchin 3053fcb2b0c5SRoman Gushchin total += (PAGE_SIZE << huge_page_order(h)) * count; 3054fcb2b0c5SRoman Gushchin 3055fcb2b0c5SRoman Gushchin if (h == &default_hstate) 3056e1759c21SAlexey Dobriyan seq_printf(m, 30571da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 30581da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 3059b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 30607893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 30614f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 3062fcb2b0c5SRoman Gushchin count, 3063a5516438SAndi Kleen h->free_huge_pages, 3064a5516438SAndi Kleen h->resv_huge_pages, 3065a5516438SAndi Kleen h->surplus_huge_pages, 3066fcb2b0c5SRoman Gushchin (PAGE_SIZE << huge_page_order(h)) / 1024); 3067fcb2b0c5SRoman Gushchin } 3068fcb2b0c5SRoman Gushchin 3069fcb2b0c5SRoman Gushchin seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024); 30701da177e4SLinus Torvalds } 30711da177e4SLinus Torvalds 30721da177e4SLinus Torvalds int hugetlb_report_node_meminfo(int nid, char *buf) 30731da177e4SLinus Torvalds { 3074a5516438SAndi Kleen struct hstate *h = &default_hstate; 3075457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3076457c1b27SNishanth Aravamudan return 0; 30771da177e4SLinus Torvalds return sprintf(buf, 30781da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 3079a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 3080a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 3081a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 3082a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 3083a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 30841da177e4SLinus Torvalds } 30851da177e4SLinus Torvalds 3086949f7ec5SDavid Rientjes void hugetlb_show_meminfo(void) 3087949f7ec5SDavid Rientjes { 3088949f7ec5SDavid Rientjes struct hstate *h; 3089949f7ec5SDavid Rientjes int nid; 3090949f7ec5SDavid Rientjes 3091457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3092457c1b27SNishanth Aravamudan return; 3093457c1b27SNishanth Aravamudan 3094949f7ec5SDavid Rientjes for_each_node_state(nid, N_MEMORY) 3095949f7ec5SDavid Rientjes for_each_hstate(h) 3096949f7ec5SDavid Rientjes pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 3097949f7ec5SDavid Rientjes nid, 3098949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 3099949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 3100949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 3101949f7ec5SDavid Rientjes 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 3102949f7ec5SDavid Rientjes } 3103949f7ec5SDavid Rientjes 31045d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 31055d317b2bSNaoya Horiguchi { 31065d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 31075d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 31085d317b2bSNaoya Horiguchi } 31095d317b2bSNaoya Horiguchi 31101da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 31111da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 31121da177e4SLinus Torvalds { 3113d0028588SWanpeng Li struct hstate *h; 3114d0028588SWanpeng Li unsigned long nr_total_pages = 0; 3115d0028588SWanpeng Li 3116d0028588SWanpeng Li for_each_hstate(h) 3117d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 3118d0028588SWanpeng Li return nr_total_pages; 31191da177e4SLinus Torvalds } 31201da177e4SLinus Torvalds 3121a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 3122fc1b8a73SMel Gorman { 3123fc1b8a73SMel Gorman int ret = -ENOMEM; 3124fc1b8a73SMel Gorman 3125fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 3126fc1b8a73SMel Gorman /* 3127fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 3128fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 3129fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 3130fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 3131fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 3132fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 3133fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 3134fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 3135fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 3136fc1b8a73SMel Gorman * 3137fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 3138fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 3139fc1b8a73SMel Gorman * we fall back to check against current free page availability as 3140fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 3141fc1b8a73SMel Gorman * semantics that cpuset has. 3142fc1b8a73SMel Gorman */ 3143fc1b8a73SMel Gorman if (delta > 0) { 3144a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 3145fc1b8a73SMel Gorman goto out; 3146fc1b8a73SMel Gorman 3147a5516438SAndi Kleen if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 3148a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 3149fc1b8a73SMel Gorman goto out; 3150fc1b8a73SMel Gorman } 3151fc1b8a73SMel Gorman } 3152fc1b8a73SMel Gorman 3153fc1b8a73SMel Gorman ret = 0; 3154fc1b8a73SMel Gorman if (delta < 0) 3155a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 3156fc1b8a73SMel Gorman 3157fc1b8a73SMel Gorman out: 3158fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 3159fc1b8a73SMel Gorman return ret; 3160fc1b8a73SMel Gorman } 3161fc1b8a73SMel Gorman 316284afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 316384afd99bSAndy Whitcroft { 3164f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 316584afd99bSAndy Whitcroft 316684afd99bSAndy Whitcroft /* 316784afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 316884afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 316984afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 317025985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 317184afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 317284afd99bSAndy Whitcroft * new reference here without additional locking. 317384afd99bSAndy Whitcroft */ 31744e35f483SJoonsoo Kim if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 3175f522c3acSJoonsoo Kim kref_get(&resv->refs); 317684afd99bSAndy Whitcroft } 317784afd99bSAndy Whitcroft 3178a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 3179a1e78772SMel Gorman { 3180a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3181f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 318290481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 31834e35f483SJoonsoo Kim unsigned long reserve, start, end; 31841c5ecae3SMike Kravetz long gbl_reserve; 318584afd99bSAndy Whitcroft 31864e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 31874e35f483SJoonsoo Kim return; 31884e35f483SJoonsoo Kim 3189a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 3190a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 319184afd99bSAndy Whitcroft 31924e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 319384afd99bSAndy Whitcroft 3194f031dd27SJoonsoo Kim kref_put(&resv->refs, resv_map_release); 319584afd99bSAndy Whitcroft 31967251ff78SAdam Litke if (reserve) { 31971c5ecae3SMike Kravetz /* 31981c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 31991c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 32001c5ecae3SMike Kravetz */ 32011c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 32021c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 32037251ff78SAdam Litke } 3204a1e78772SMel Gorman } 3205a1e78772SMel Gorman 320631383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 320731383c68SDan Williams { 320831383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 320931383c68SDan Williams return -EINVAL; 321031383c68SDan Williams return 0; 321131383c68SDan Williams } 321231383c68SDan Williams 321305ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 321405ea8860SDan Williams { 321505ea8860SDan Williams struct hstate *hstate = hstate_vma(vma); 321605ea8860SDan Williams 321705ea8860SDan Williams return 1UL << huge_page_shift(hstate); 321805ea8860SDan Williams } 321905ea8860SDan Williams 32201da177e4SLinus Torvalds /* 32211da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 32221da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 32231da177e4SLinus Torvalds * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 32241da177e4SLinus Torvalds * this far. 32251da177e4SLinus Torvalds */ 3226b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 32271da177e4SLinus Torvalds { 32281da177e4SLinus Torvalds BUG(); 3229d0217ac0SNick Piggin return 0; 32301da177e4SLinus Torvalds } 32311da177e4SLinus Torvalds 3232eec3636aSJane Chu /* 3233eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 3234eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 3235eec3636aSJane Chu * This is because under System V memory model, mappings created via 3236eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 3237eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 3238eec3636aSJane Chu */ 3239f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 3240d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 324184afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 3242a1e78772SMel Gorman .close = hugetlb_vm_op_close, 324331383c68SDan Williams .split = hugetlb_vm_op_split, 324405ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 32451da177e4SLinus Torvalds }; 32461da177e4SLinus Torvalds 32471e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 32481e8f889bSDavid Gibson int writable) 324963551ae0SDavid Gibson { 325063551ae0SDavid Gibson pte_t entry; 325163551ae0SDavid Gibson 32521e8f889bSDavid Gibson if (writable) { 3253106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 3254106c992aSGerald Schaefer vma->vm_page_prot))); 325563551ae0SDavid Gibson } else { 3256106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 3257106c992aSGerald Schaefer vma->vm_page_prot)); 325863551ae0SDavid Gibson } 325963551ae0SDavid Gibson entry = pte_mkyoung(entry); 326063551ae0SDavid Gibson entry = pte_mkhuge(entry); 3261d9ed9faaSChris Metcalf entry = arch_make_huge_pte(entry, vma, page, writable); 326263551ae0SDavid Gibson 326363551ae0SDavid Gibson return entry; 326463551ae0SDavid Gibson } 326563551ae0SDavid Gibson 32661e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 32671e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 32681e8f889bSDavid Gibson { 32691e8f889bSDavid Gibson pte_t entry; 32701e8f889bSDavid Gibson 3271106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 327232f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 32734b3073e1SRussell King update_mmu_cache(vma, address, ptep); 32741e8f889bSDavid Gibson } 32751e8f889bSDavid Gibson 3276d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 32774a705fefSNaoya Horiguchi { 32784a705fefSNaoya Horiguchi swp_entry_t swp; 32794a705fefSNaoya Horiguchi 32804a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 3281d5ed7444SAneesh Kumar K.V return false; 32824a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 32834a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_migration_entry(swp)) 3284d5ed7444SAneesh Kumar K.V return true; 32854a705fefSNaoya Horiguchi else 3286d5ed7444SAneesh Kumar K.V return false; 32874a705fefSNaoya Horiguchi } 32884a705fefSNaoya Horiguchi 32894a705fefSNaoya Horiguchi static int is_hugetlb_entry_hwpoisoned(pte_t pte) 32904a705fefSNaoya Horiguchi { 32914a705fefSNaoya Horiguchi swp_entry_t swp; 32924a705fefSNaoya Horiguchi 32934a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 32944a705fefSNaoya Horiguchi return 0; 32954a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 32964a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_hwpoison_entry(swp)) 32974a705fefSNaoya Horiguchi return 1; 32984a705fefSNaoya Horiguchi else 32994a705fefSNaoya Horiguchi return 0; 33004a705fefSNaoya Horiguchi } 33011e8f889bSDavid Gibson 330263551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 330363551ae0SDavid Gibson struct vm_area_struct *vma) 330463551ae0SDavid Gibson { 33055e41540cSMike Kravetz pte_t *src_pte, *dst_pte, entry, dst_entry; 330663551ae0SDavid Gibson struct page *ptepage; 33071c59827dSHugh Dickins unsigned long addr; 33081e8f889bSDavid Gibson int cow; 3309a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3310a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 3311ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 3312e8569dd2SAndreas Sandberg int ret = 0; 33131e8f889bSDavid Gibson 33141e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 331563551ae0SDavid Gibson 3316ac46d4f3SJérôme Glisse if (cow) { 33177269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src, 33186f4f13e8SJérôme Glisse vma->vm_start, 3319ac46d4f3SJérôme Glisse vma->vm_end); 3320ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 3321ac46d4f3SJérôme Glisse } 3322e8569dd2SAndreas Sandberg 3323a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 3324cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 33257868a208SPunit Agrawal src_pte = huge_pte_offset(src, addr, sz); 3326c74df32cSHugh Dickins if (!src_pte) 3327c74df32cSHugh Dickins continue; 3328a5516438SAndi Kleen dst_pte = huge_pte_alloc(dst, addr, sz); 3329e8569dd2SAndreas Sandberg if (!dst_pte) { 3330e8569dd2SAndreas Sandberg ret = -ENOMEM; 3331e8569dd2SAndreas Sandberg break; 3332e8569dd2SAndreas Sandberg } 3333c5c99429SLarry Woodman 33345e41540cSMike Kravetz /* 33355e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 33365e41540cSMike Kravetz * dst_pte == src_pte is the common case of src/dest sharing. 33375e41540cSMike Kravetz * 33385e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 33395e41540cSMike Kravetz * another vma. If dst_pte !none, this implies sharing. 33405e41540cSMike Kravetz * Check here before taking page table lock, and once again 33415e41540cSMike Kravetz * after taking the lock below. 33425e41540cSMike Kravetz */ 33435e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 33445e41540cSMike Kravetz if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) 3345c5c99429SLarry Woodman continue; 3346c5c99429SLarry Woodman 3347cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 3348cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 3349cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 33504a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 33515e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 33525e41540cSMike Kravetz if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { 33535e41540cSMike Kravetz /* 33545e41540cSMike Kravetz * Skip if src entry none. Also, skip in the 33555e41540cSMike Kravetz * unlikely case dst entry !none as this implies 33565e41540cSMike Kravetz * sharing with another vma. 33575e41540cSMike Kravetz */ 33584a705fefSNaoya Horiguchi ; 33594a705fefSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry) || 33604a705fefSNaoya Horiguchi is_hugetlb_entry_hwpoisoned(entry))) { 33614a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 33624a705fefSNaoya Horiguchi 33634a705fefSNaoya Horiguchi if (is_write_migration_entry(swp_entry) && cow) { 33644a705fefSNaoya Horiguchi /* 33654a705fefSNaoya Horiguchi * COW mappings require pages in both 33664a705fefSNaoya Horiguchi * parent and child to be set to read. 33674a705fefSNaoya Horiguchi */ 33684a705fefSNaoya Horiguchi make_migration_entry_read(&swp_entry); 33694a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 3370e5251fd4SPunit Agrawal set_huge_swap_pte_at(src, addr, src_pte, 3371e5251fd4SPunit Agrawal entry, sz); 33724a705fefSNaoya Horiguchi } 3373e5251fd4SPunit Agrawal set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); 33744a705fefSNaoya Horiguchi } else { 337534ee645eSJoerg Roedel if (cow) { 33760f10851eSJérôme Glisse /* 33770f10851eSJérôme Glisse * No need to notify as we are downgrading page 33780f10851eSJérôme Glisse * table protection not changing it to point 33790f10851eSJérôme Glisse * to a new page. 33800f10851eSJérôme Glisse * 3381ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 33820f10851eSJérôme Glisse */ 33837f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 338434ee645eSJoerg Roedel } 33850253d634SNaoya Horiguchi entry = huge_ptep_get(src_pte); 338663551ae0SDavid Gibson ptepage = pte_page(entry); 338763551ae0SDavid Gibson get_page(ptepage); 338853f9263bSKirill A. Shutemov page_dup_rmap(ptepage, true); 338963551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 33905d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), dst); 33911c59827dSHugh Dickins } 3392cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 3393cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 339463551ae0SDavid Gibson } 339563551ae0SDavid Gibson 3396e8569dd2SAndreas Sandberg if (cow) 3397ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 3398e8569dd2SAndreas Sandberg 3399e8569dd2SAndreas Sandberg return ret; 340063551ae0SDavid Gibson } 340163551ae0SDavid Gibson 340224669e58SAneesh Kumar K.V void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 340324669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 340424669e58SAneesh Kumar K.V struct page *ref_page) 340563551ae0SDavid Gibson { 340663551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 340763551ae0SDavid Gibson unsigned long address; 3408c7546f8fSDavid Gibson pte_t *ptep; 340963551ae0SDavid Gibson pte_t pte; 3410cb900f41SKirill A. Shutemov spinlock_t *ptl; 341163551ae0SDavid Gibson struct page *page; 3412a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3413a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 3414ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 3415a5516438SAndi Kleen 341663551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 3417a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 3418a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 341963551ae0SDavid Gibson 342007e32661SAneesh Kumar K.V /* 342107e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 342207e32661SAneesh Kumar K.V * to huge page. 342307e32661SAneesh Kumar K.V */ 3424ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 342524669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 3426dff11abeSMike Kravetz 3427dff11abeSMike Kravetz /* 3428dff11abeSMike Kravetz * If sharing possible, alert mmu notifiers of worst case. 3429dff11abeSMike Kravetz */ 34306f4f13e8SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, 34316f4f13e8SJérôme Glisse end); 3432ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 3433ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 3434569f48b8SHillf Danton address = start; 3435569f48b8SHillf Danton for (; address < end; address += sz) { 34367868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, sz); 3437c7546f8fSDavid Gibson if (!ptep) 3438c7546f8fSDavid Gibson continue; 3439c7546f8fSDavid Gibson 3440cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 344131d49da5SAneesh Kumar K.V if (huge_pmd_unshare(mm, &address, ptep)) { 344231d49da5SAneesh Kumar K.V spin_unlock(ptl); 3443dff11abeSMike Kravetz /* 3444dff11abeSMike Kravetz * We just unmapped a page of PMDs by clearing a PUD. 3445dff11abeSMike Kravetz * The caller's TLB flush range should cover this area. 3446dff11abeSMike Kravetz */ 344731d49da5SAneesh Kumar K.V continue; 344831d49da5SAneesh Kumar K.V } 344939dde65cSChen, Kenneth W 34506629326bSHillf Danton pte = huge_ptep_get(ptep); 345131d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 345231d49da5SAneesh Kumar K.V spin_unlock(ptl); 345331d49da5SAneesh Kumar K.V continue; 345431d49da5SAneesh Kumar K.V } 34556629326bSHillf Danton 34566629326bSHillf Danton /* 34579fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 34589fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 34596629326bSHillf Danton */ 34609fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 34619386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 346231d49da5SAneesh Kumar K.V spin_unlock(ptl); 346331d49da5SAneesh Kumar K.V continue; 34648c4894c6SNaoya Horiguchi } 34656629326bSHillf Danton 34666629326bSHillf Danton page = pte_page(pte); 346704f2cbe3SMel Gorman /* 346804f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 346904f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 347004f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 347104f2cbe3SMel Gorman */ 347204f2cbe3SMel Gorman if (ref_page) { 347331d49da5SAneesh Kumar K.V if (page != ref_page) { 347431d49da5SAneesh Kumar K.V spin_unlock(ptl); 347531d49da5SAneesh Kumar K.V continue; 347631d49da5SAneesh Kumar K.V } 347704f2cbe3SMel Gorman /* 347804f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 347904f2cbe3SMel Gorman * future faults in this VMA will fail rather than 348004f2cbe3SMel Gorman * looking like data was lost 348104f2cbe3SMel Gorman */ 348204f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 348304f2cbe3SMel Gorman } 348404f2cbe3SMel Gorman 3485c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 3486b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 3487106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 34886649a386SKen Chen set_page_dirty(page); 34899e81130bSHillf Danton 34905d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 3491d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 349231d49da5SAneesh Kumar K.V 3493cb900f41SKirill A. Shutemov spin_unlock(ptl); 3494e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 349524669e58SAneesh Kumar K.V /* 349631d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 349724669e58SAneesh Kumar K.V */ 349831d49da5SAneesh Kumar K.V if (ref_page) 349931d49da5SAneesh Kumar K.V break; 3500fe1668aeSChen, Kenneth W } 3501ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 350224669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 35031da177e4SLinus Torvalds } 350463551ae0SDavid Gibson 3505d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 3506d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 3507d833352aSMel Gorman unsigned long end, struct page *ref_page) 3508d833352aSMel Gorman { 3509d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 3510d833352aSMel Gorman 3511d833352aSMel Gorman /* 3512d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 3513d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 3514d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 3515d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 3516d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 3517c8c06efaSDavidlohr Bueso * is to clear it before releasing the i_mmap_rwsem. This works 3518d833352aSMel Gorman * because in the context this is called, the VMA is about to be 3519c8c06efaSDavidlohr Bueso * destroyed and the i_mmap_rwsem is held. 3520d833352aSMel Gorman */ 3521d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 3522d833352aSMel Gorman } 3523d833352aSMel Gorman 3524502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 352504f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 3526502717f4SChen, Kenneth W { 352724669e58SAneesh Kumar K.V struct mm_struct *mm; 352824669e58SAneesh Kumar K.V struct mmu_gather tlb; 3529dff11abeSMike Kravetz unsigned long tlb_start = start; 3530dff11abeSMike Kravetz unsigned long tlb_end = end; 3531dff11abeSMike Kravetz 3532dff11abeSMike Kravetz /* 3533dff11abeSMike Kravetz * If shared PMDs were possibly used within this vma range, adjust 3534dff11abeSMike Kravetz * start/end for worst case tlb flushing. 3535dff11abeSMike Kravetz * Note that we can not be sure if PMDs are shared until we try to 3536dff11abeSMike Kravetz * unmap pages. However, we want to make sure TLB flushing covers 3537dff11abeSMike Kravetz * the largest possible range. 3538dff11abeSMike Kravetz */ 3539dff11abeSMike Kravetz adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end); 354024669e58SAneesh Kumar K.V 354124669e58SAneesh Kumar K.V mm = vma->vm_mm; 354224669e58SAneesh Kumar K.V 3543dff11abeSMike Kravetz tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end); 354424669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 3545dff11abeSMike Kravetz tlb_finish_mmu(&tlb, tlb_start, tlb_end); 3546502717f4SChen, Kenneth W } 3547502717f4SChen, Kenneth W 354804f2cbe3SMel Gorman /* 354904f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 355004f2cbe3SMel Gorman * mappping it owns the reserve page for. The intention is to unmap the page 355104f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 355204f2cbe3SMel Gorman * same region. 355304f2cbe3SMel Gorman */ 35542f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 35552a4b3dedSHarvey Harrison struct page *page, unsigned long address) 355604f2cbe3SMel Gorman { 35577526674dSAdam Litke struct hstate *h = hstate_vma(vma); 355804f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 355904f2cbe3SMel Gorman struct address_space *mapping; 356004f2cbe3SMel Gorman pgoff_t pgoff; 356104f2cbe3SMel Gorman 356204f2cbe3SMel Gorman /* 356304f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 356404f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 356504f2cbe3SMel Gorman */ 35667526674dSAdam Litke address = address & huge_page_mask(h); 356736e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 356836e4f20aSMichal Hocko vma->vm_pgoff; 356993c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 357004f2cbe3SMel Gorman 35714eb2b1dcSMel Gorman /* 35724eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 35734eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 35744eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 35754eb2b1dcSMel Gorman */ 357683cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 35776b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 357804f2cbe3SMel Gorman /* Do not unmap the current VMA */ 357904f2cbe3SMel Gorman if (iter_vma == vma) 358004f2cbe3SMel Gorman continue; 358104f2cbe3SMel Gorman 358204f2cbe3SMel Gorman /* 35832f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 35842f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 35852f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 35862f84a899SMel Gorman */ 35872f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 35882f84a899SMel Gorman continue; 35892f84a899SMel Gorman 35902f84a899SMel Gorman /* 359104f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 359204f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 359304f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 359404f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 359504f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 359604f2cbe3SMel Gorman */ 359704f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 359824669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 359924669e58SAneesh Kumar K.V address + huge_page_size(h), page); 360004f2cbe3SMel Gorman } 360183cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 360204f2cbe3SMel Gorman } 360304f2cbe3SMel Gorman 36040fe6e20bSNaoya Horiguchi /* 36050fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 3606ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 3607ef009b25SMichal Hocko * cannot race with other handlers or page migration. 3608ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 36090fe6e20bSNaoya Horiguchi */ 36102b740303SSouptick Joarder static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 3611974e6d66SHuang Ying unsigned long address, pte_t *ptep, 3612cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 36131e8f889bSDavid Gibson { 36143999f52eSAneesh Kumar K.V pte_t pte; 3615a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 36161e8f889bSDavid Gibson struct page *old_page, *new_page; 36172b740303SSouptick Joarder int outside_reserve = 0; 36182b740303SSouptick Joarder vm_fault_t ret = 0; 3619974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 3620ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 36211e8f889bSDavid Gibson 36223999f52eSAneesh Kumar K.V pte = huge_ptep_get(ptep); 36231e8f889bSDavid Gibson old_page = pte_page(pte); 36241e8f889bSDavid Gibson 362504f2cbe3SMel Gorman retry_avoidcopy: 36261e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 36271e8f889bSDavid Gibson * and just make the page writable */ 362837a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 36295a49973dSHugh Dickins page_move_anon_rmap(old_page, vma); 36305b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 363183c54070SNick Piggin return 0; 36321e8f889bSDavid Gibson } 36331e8f889bSDavid Gibson 363404f2cbe3SMel Gorman /* 363504f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 363604f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 363704f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 363804f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 363904f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 364004f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 364104f2cbe3SMel Gorman * of the full address range. 364204f2cbe3SMel Gorman */ 36435944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 364404f2cbe3SMel Gorman old_page != pagecache_page) 364504f2cbe3SMel Gorman outside_reserve = 1; 364604f2cbe3SMel Gorman 364709cbfeafSKirill A. Shutemov get_page(old_page); 3648b76c8cfbSLarry Woodman 3649ad4404a2SDavidlohr Bueso /* 3650ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 3651ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 3652ad4404a2SDavidlohr Bueso */ 3653cb900f41SKirill A. Shutemov spin_unlock(ptl); 36545b7a1d40SHuang Ying new_page = alloc_huge_page(vma, haddr, outside_reserve); 36551e8f889bSDavid Gibson 36562fc39cecSAdam Litke if (IS_ERR(new_page)) { 365704f2cbe3SMel Gorman /* 365804f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 365904f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 366004f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 366104f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 366204f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 366304f2cbe3SMel Gorman */ 366404f2cbe3SMel Gorman if (outside_reserve) { 366509cbfeafSKirill A. Shutemov put_page(old_page); 366604f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 36675b7a1d40SHuang Ying unmap_ref_private(mm, vma, old_page, haddr); 366804f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 3669cb900f41SKirill A. Shutemov spin_lock(ptl); 36705b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 3671a9af0c5dSNaoya Horiguchi if (likely(ptep && 3672a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 367304f2cbe3SMel Gorman goto retry_avoidcopy; 3674a734bcc8SHillf Danton /* 3675cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 3676cb900f41SKirill A. Shutemov * lock, and our job is done. 3677a734bcc8SHillf Danton */ 3678a734bcc8SHillf Danton return 0; 367904f2cbe3SMel Gorman } 368004f2cbe3SMel Gorman 36812b740303SSouptick Joarder ret = vmf_error(PTR_ERR(new_page)); 3682ad4404a2SDavidlohr Bueso goto out_release_old; 36831e8f889bSDavid Gibson } 36841e8f889bSDavid Gibson 36850fe6e20bSNaoya Horiguchi /* 36860fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 36870fe6e20bSNaoya Horiguchi * anon_vma prepared. 36880fe6e20bSNaoya Horiguchi */ 368944e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 3690ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 3691ad4404a2SDavidlohr Bueso goto out_release_all; 369244e2aa93SDean Nelson } 36930fe6e20bSNaoya Horiguchi 3694974e6d66SHuang Ying copy_user_huge_page(new_page, old_page, address, vma, 369547ad8475SAndrea Arcangeli pages_per_huge_page(h)); 36960ed361deSNick Piggin __SetPageUptodate(new_page); 36971e8f889bSDavid Gibson 36987269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, 36996f4f13e8SJérôme Glisse haddr + huge_page_size(h)); 3700ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 3701ad4404a2SDavidlohr Bueso 3702b76c8cfbSLarry Woodman /* 3703cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 3704b76c8cfbSLarry Woodman * before the page tables are altered 3705b76c8cfbSLarry Woodman */ 3706cb900f41SKirill A. Shutemov spin_lock(ptl); 37075b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 3708a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 370907443a85SJoonsoo Kim ClearPagePrivate(new_page); 371007443a85SJoonsoo Kim 37111e8f889bSDavid Gibson /* Break COW */ 37125b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 3713ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, range.end); 37145b7a1d40SHuang Ying set_huge_pte_at(mm, haddr, ptep, 37151e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 3716d281ee61SKirill A. Shutemov page_remove_rmap(old_page, true); 37175b7a1d40SHuang Ying hugepage_add_new_anon_rmap(new_page, vma, haddr); 3718cb6acd01SMike Kravetz set_page_huge_active(new_page); 37191e8f889bSDavid Gibson /* Make the old page be freed below */ 37201e8f889bSDavid Gibson new_page = old_page; 37211e8f889bSDavid Gibson } 3722cb900f41SKirill A. Shutemov spin_unlock(ptl); 3723ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 3724ad4404a2SDavidlohr Bueso out_release_all: 37255b7a1d40SHuang Ying restore_reserve_on_error(h, vma, haddr, new_page); 372609cbfeafSKirill A. Shutemov put_page(new_page); 3727ad4404a2SDavidlohr Bueso out_release_old: 372809cbfeafSKirill A. Shutemov put_page(old_page); 37298312034fSJoonsoo Kim 3730ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 3731ad4404a2SDavidlohr Bueso return ret; 37321e8f889bSDavid Gibson } 37331e8f889bSDavid Gibson 373404f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 3735a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 3736a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 373704f2cbe3SMel Gorman { 373804f2cbe3SMel Gorman struct address_space *mapping; 3739e7c4b0bfSAndy Whitcroft pgoff_t idx; 374004f2cbe3SMel Gorman 374104f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 3742a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 374304f2cbe3SMel Gorman 374404f2cbe3SMel Gorman return find_lock_page(mapping, idx); 374504f2cbe3SMel Gorman } 374604f2cbe3SMel Gorman 37473ae77f43SHugh Dickins /* 37483ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 37493ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 37503ae77f43SHugh Dickins */ 37513ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 37522a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 37532a15efc9SHugh Dickins { 37542a15efc9SHugh Dickins struct address_space *mapping; 37552a15efc9SHugh Dickins pgoff_t idx; 37562a15efc9SHugh Dickins struct page *page; 37572a15efc9SHugh Dickins 37582a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 37592a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 37602a15efc9SHugh Dickins 37612a15efc9SHugh Dickins page = find_get_page(mapping, idx); 37622a15efc9SHugh Dickins if (page) 37632a15efc9SHugh Dickins put_page(page); 37642a15efc9SHugh Dickins return page != NULL; 37652a15efc9SHugh Dickins } 37662a15efc9SHugh Dickins 3767ab76ad54SMike Kravetz int huge_add_to_page_cache(struct page *page, struct address_space *mapping, 3768ab76ad54SMike Kravetz pgoff_t idx) 3769ab76ad54SMike Kravetz { 3770ab76ad54SMike Kravetz struct inode *inode = mapping->host; 3771ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 3772ab76ad54SMike Kravetz int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 3773ab76ad54SMike Kravetz 3774ab76ad54SMike Kravetz if (err) 3775ab76ad54SMike Kravetz return err; 3776ab76ad54SMike Kravetz ClearPagePrivate(page); 3777ab76ad54SMike Kravetz 377822146c3cSMike Kravetz /* 377922146c3cSMike Kravetz * set page dirty so that it will not be removed from cache/file 378022146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 378122146c3cSMike Kravetz */ 378222146c3cSMike Kravetz set_page_dirty(page); 378322146c3cSMike Kravetz 3784ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 3785ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 3786ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 3787ab76ad54SMike Kravetz return 0; 3788ab76ad54SMike Kravetz } 3789ab76ad54SMike Kravetz 37902b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 37912b740303SSouptick Joarder struct vm_area_struct *vma, 37928382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 3793788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 3794ac9b9c66SHugh Dickins { 3795a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 37962b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 3797409eb8c2SHillf Danton int anon_rmap = 0; 37984c887265SAdam Litke unsigned long size; 37994c887265SAdam Litke struct page *page; 38001e8f889bSDavid Gibson pte_t new_pte; 3801cb900f41SKirill A. Shutemov spinlock_t *ptl; 3802285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 3803cb6acd01SMike Kravetz bool new_page = false; 38044c887265SAdam Litke 380504f2cbe3SMel Gorman /* 380604f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 380704f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 380825985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 380904f2cbe3SMel Gorman */ 381004f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 3811910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 381204f2cbe3SMel Gorman current->pid); 381304f2cbe3SMel Gorman return ret; 381404f2cbe3SMel Gorman } 381504f2cbe3SMel Gorman 38164c887265SAdam Litke /* 3817e7c58097SMike Kravetz * Use page lock to guard against racing truncation 3818e7c58097SMike Kravetz * before we get page_table_lock. 38194c887265SAdam Litke */ 3820e7c58097SMike Kravetz retry: 3821e7c58097SMike Kravetz page = find_lock_page(mapping, idx); 3822e7c58097SMike Kravetz if (!page) { 3823a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 3824ebed4bfcSHugh Dickins if (idx >= size) 3825ebed4bfcSHugh Dickins goto out; 38261a1aad8aSMike Kravetz 38271a1aad8aSMike Kravetz /* 38281a1aad8aSMike Kravetz * Check for page in userfault range 38291a1aad8aSMike Kravetz */ 38301a1aad8aSMike Kravetz if (userfaultfd_missing(vma)) { 38311a1aad8aSMike Kravetz u32 hash; 38321a1aad8aSMike Kravetz struct vm_fault vmf = { 38331a1aad8aSMike Kravetz .vma = vma, 3834285b8dcaSHuang Ying .address = haddr, 38351a1aad8aSMike Kravetz .flags = flags, 38361a1aad8aSMike Kravetz /* 38371a1aad8aSMike Kravetz * Hard to debug if it ends up being 38381a1aad8aSMike Kravetz * used by a callee that assumes 38391a1aad8aSMike Kravetz * something about the other 38401a1aad8aSMike Kravetz * uninitialized fields... same as in 38411a1aad8aSMike Kravetz * memory.c 38421a1aad8aSMike Kravetz */ 38431a1aad8aSMike Kravetz }; 38441a1aad8aSMike Kravetz 38451a1aad8aSMike Kravetz /* 3846ddeaab32SMike Kravetz * hugetlb_fault_mutex must be dropped before 3847ddeaab32SMike Kravetz * handling userfault. Reacquire after handling 3848ddeaab32SMike Kravetz * fault to make calling code simpler. 38491a1aad8aSMike Kravetz */ 38501b426bacSMike Kravetz hash = hugetlb_fault_mutex_hash(h, mapping, idx, haddr); 38511a1aad8aSMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 38521a1aad8aSMike Kravetz ret = handle_userfault(&vmf, VM_UFFD_MISSING); 38531a1aad8aSMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 38541a1aad8aSMike Kravetz goto out; 38551a1aad8aSMike Kravetz } 38561a1aad8aSMike Kravetz 3857285b8dcaSHuang Ying page = alloc_huge_page(vma, haddr, 0); 38582fc39cecSAdam Litke if (IS_ERR(page)) { 38592b740303SSouptick Joarder ret = vmf_error(PTR_ERR(page)); 38606bda666aSChristoph Lameter goto out; 38616bda666aSChristoph Lameter } 386247ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 38630ed361deSNick Piggin __SetPageUptodate(page); 3864cb6acd01SMike Kravetz new_page = true; 3865ac9b9c66SHugh Dickins 3866f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 3867ab76ad54SMike Kravetz int err = huge_add_to_page_cache(page, mapping, idx); 38686bda666aSChristoph Lameter if (err) { 38696bda666aSChristoph Lameter put_page(page); 38706bda666aSChristoph Lameter if (err == -EEXIST) 38716bda666aSChristoph Lameter goto retry; 38726bda666aSChristoph Lameter goto out; 38736bda666aSChristoph Lameter } 387423be7468SMel Gorman } else { 38756bda666aSChristoph Lameter lock_page(page); 38760fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 38770fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 38780fe6e20bSNaoya Horiguchi goto backout_unlocked; 387923be7468SMel Gorman } 3880409eb8c2SHillf Danton anon_rmap = 1; 38810fe6e20bSNaoya Horiguchi } 38820fe6e20bSNaoya Horiguchi } else { 388357303d80SAndy Whitcroft /* 3884998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 3885998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 3886998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 3887fd6a03edSNaoya Horiguchi */ 3888fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 3889aa50d3a7SAndi Kleen ret = VM_FAULT_HWPOISON | 3890972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 3891fd6a03edSNaoya Horiguchi goto backout_unlocked; 38926bda666aSChristoph Lameter } 3893998b4382SNaoya Horiguchi } 38941e8f889bSDavid Gibson 389557303d80SAndy Whitcroft /* 389657303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 389757303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 389857303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 389957303d80SAndy Whitcroft * the spinlock. 390057303d80SAndy Whitcroft */ 39015e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 3902285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 39032b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 39042b26736cSAndy Whitcroft goto backout_unlocked; 39052b26736cSAndy Whitcroft } 39065e911373SMike Kravetz /* Just decrements count, does not deallocate */ 3907285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 39085e911373SMike Kravetz } 390957303d80SAndy Whitcroft 39108bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 3911e7c58097SMike Kravetz size = i_size_read(mapping->host) >> huge_page_shift(h); 3912e7c58097SMike Kravetz if (idx >= size) 3913e7c58097SMike Kravetz goto backout; 39144c887265SAdam Litke 391583c54070SNick Piggin ret = 0; 39167f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 39174c887265SAdam Litke goto backout; 39184c887265SAdam Litke 391907443a85SJoonsoo Kim if (anon_rmap) { 392007443a85SJoonsoo Kim ClearPagePrivate(page); 3921285b8dcaSHuang Ying hugepage_add_new_anon_rmap(page, vma, haddr); 3922ac714904SChoi Gi-yong } else 392353f9263bSKirill A. Shutemov page_dup_rmap(page, true); 39241e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 39251e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 3926285b8dcaSHuang Ying set_huge_pte_at(mm, haddr, ptep, new_pte); 39271e8f889bSDavid Gibson 39285d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 3929788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 39301e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 3931974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); 39321e8f889bSDavid Gibson } 39331e8f889bSDavid Gibson 3934cb900f41SKirill A. Shutemov spin_unlock(ptl); 3935cb6acd01SMike Kravetz 3936cb6acd01SMike Kravetz /* 3937cb6acd01SMike Kravetz * Only make newly allocated pages active. Existing pages found 3938cb6acd01SMike Kravetz * in the pagecache could be !page_huge_active() if they have been 3939cb6acd01SMike Kravetz * isolated for migration. 3940cb6acd01SMike Kravetz */ 3941cb6acd01SMike Kravetz if (new_page) 3942cb6acd01SMike Kravetz set_page_huge_active(page); 3943cb6acd01SMike Kravetz 39444c887265SAdam Litke unlock_page(page); 39454c887265SAdam Litke out: 3946ac9b9c66SHugh Dickins return ret; 39474c887265SAdam Litke 39484c887265SAdam Litke backout: 3949cb900f41SKirill A. Shutemov spin_unlock(ptl); 39502b26736cSAndy Whitcroft backout_unlocked: 39514c887265SAdam Litke unlock_page(page); 3952285b8dcaSHuang Ying restore_reserve_on_error(h, vma, haddr, page); 39534c887265SAdam Litke put_page(page); 39544c887265SAdam Litke goto out; 3955ac9b9c66SHugh Dickins } 3956ac9b9c66SHugh Dickins 39578382d914SDavidlohr Bueso #ifdef CONFIG_SMP 39581b426bacSMike Kravetz u32 hugetlb_fault_mutex_hash(struct hstate *h, struct address_space *mapping, 39598382d914SDavidlohr Bueso pgoff_t idx, unsigned long address) 39608382d914SDavidlohr Bueso { 39618382d914SDavidlohr Bueso unsigned long key[2]; 39628382d914SDavidlohr Bueso u32 hash; 39638382d914SDavidlohr Bueso 39648382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 39658382d914SDavidlohr Bueso key[1] = idx; 39668382d914SDavidlohr Bueso 39678382d914SDavidlohr Bueso hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0); 39688382d914SDavidlohr Bueso 39698382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 39708382d914SDavidlohr Bueso } 39718382d914SDavidlohr Bueso #else 39728382d914SDavidlohr Bueso /* 39738382d914SDavidlohr Bueso * For uniprocesor systems we always use a single mutex, so just 39748382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 39758382d914SDavidlohr Bueso */ 39761b426bacSMike Kravetz u32 hugetlb_fault_mutex_hash(struct hstate *h, struct address_space *mapping, 39778382d914SDavidlohr Bueso pgoff_t idx, unsigned long address) 39788382d914SDavidlohr Bueso { 39798382d914SDavidlohr Bueso return 0; 39808382d914SDavidlohr Bueso } 39818382d914SDavidlohr Bueso #endif 39828382d914SDavidlohr Bueso 39832b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 3984788c7df4SHugh Dickins unsigned long address, unsigned int flags) 398586e5216fSAdam Litke { 39868382d914SDavidlohr Bueso pte_t *ptep, entry; 3987cb900f41SKirill A. Shutemov spinlock_t *ptl; 39882b740303SSouptick Joarder vm_fault_t ret; 39898382d914SDavidlohr Bueso u32 hash; 39908382d914SDavidlohr Bueso pgoff_t idx; 39910fe6e20bSNaoya Horiguchi struct page *page = NULL; 399257303d80SAndy Whitcroft struct page *pagecache_page = NULL; 3993a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 39948382d914SDavidlohr Bueso struct address_space *mapping; 39950f792cf9SNaoya Horiguchi int need_wait_lock = 0; 3996285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 399786e5216fSAdam Litke 3998285b8dcaSHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 3999fd6a03edSNaoya Horiguchi if (ptep) { 4000fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 4001290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 4002cb900f41SKirill A. Shutemov migration_entry_wait_huge(vma, mm, ptep); 4003290408d4SNaoya Horiguchi return 0; 4004290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 4005aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 4006972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 4007ddeaab32SMike Kravetz } else { 4008b43a9990SMike Kravetz ptep = huge_pte_alloc(mm, haddr, huge_page_size(h)); 4009ddeaab32SMike Kravetz if (!ptep) 4010b43a9990SMike Kravetz return VM_FAULT_OOM; 4011b43a9990SMike Kravetz } 40128382d914SDavidlohr Bueso 4013ddeaab32SMike Kravetz mapping = vma->vm_file->f_mapping; 4014ddeaab32SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 4015ddeaab32SMike Kravetz 40163935baa9SDavid Gibson /* 40173935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 40183935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 40193935baa9SDavid Gibson * the same page in the page cache. 40203935baa9SDavid Gibson */ 40211b426bacSMike Kravetz hash = hugetlb_fault_mutex_hash(h, mapping, idx, haddr); 4022c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 40238382d914SDavidlohr Bueso 40247f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 40257f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 40268382d914SDavidlohr Bueso ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 4027b4d1d99fSDavid Gibson goto out_mutex; 40283935baa9SDavid Gibson } 402986e5216fSAdam Litke 403083c54070SNick Piggin ret = 0; 40311e8f889bSDavid Gibson 403257303d80SAndy Whitcroft /* 40330f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 40340f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 40350f792cf9SNaoya Horiguchi * a active hugepage in pagecache. This goto expects the 2nd page fault, 40360f792cf9SNaoya Horiguchi * and is_hugetlb_entry_(migration|hwpoisoned) check will properly 40370f792cf9SNaoya Horiguchi * handle it. 40380f792cf9SNaoya Horiguchi */ 40390f792cf9SNaoya Horiguchi if (!pte_present(entry)) 40400f792cf9SNaoya Horiguchi goto out_mutex; 40410f792cf9SNaoya Horiguchi 40420f792cf9SNaoya Horiguchi /* 404357303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 404457303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 404557303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 404657303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 404757303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 404857303d80SAndy Whitcroft * consumed. 404957303d80SAndy Whitcroft */ 4050106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 4051285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 40522b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 4053b4d1d99fSDavid Gibson goto out_mutex; 40542b26736cSAndy Whitcroft } 40555e911373SMike Kravetz /* Just decrements count, does not deallocate */ 4056285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 405757303d80SAndy Whitcroft 4058f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 405957303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 4060285b8dcaSHuang Ying vma, haddr); 406157303d80SAndy Whitcroft } 406257303d80SAndy Whitcroft 40630f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 40640fe6e20bSNaoya Horiguchi 40651e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 4066b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 4067cb900f41SKirill A. Shutemov goto out_ptl; 4068b4d1d99fSDavid Gibson 40690f792cf9SNaoya Horiguchi /* 40700f792cf9SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 40710f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 40720f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 40730f792cf9SNaoya Horiguchi */ 40740f792cf9SNaoya Horiguchi page = pte_page(entry); 40750f792cf9SNaoya Horiguchi if (page != pagecache_page) 40760f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 40770f792cf9SNaoya Horiguchi need_wait_lock = 1; 40780f792cf9SNaoya Horiguchi goto out_ptl; 40790f792cf9SNaoya Horiguchi } 40800f792cf9SNaoya Horiguchi 40810f792cf9SNaoya Horiguchi get_page(page); 4082b4d1d99fSDavid Gibson 4083788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 4084106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 4085974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, 4086cb900f41SKirill A. Shutemov pagecache_page, ptl); 40870f792cf9SNaoya Horiguchi goto out_put_page; 4088b4d1d99fSDavid Gibson } 4089106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 4090b4d1d99fSDavid Gibson } 4091b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 4092285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 4093788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 4094285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 40950f792cf9SNaoya Horiguchi out_put_page: 40960f792cf9SNaoya Horiguchi if (page != pagecache_page) 40970f792cf9SNaoya Horiguchi unlock_page(page); 40980f792cf9SNaoya Horiguchi put_page(page); 4099cb900f41SKirill A. Shutemov out_ptl: 4100cb900f41SKirill A. Shutemov spin_unlock(ptl); 410157303d80SAndy Whitcroft 410257303d80SAndy Whitcroft if (pagecache_page) { 410357303d80SAndy Whitcroft unlock_page(pagecache_page); 410457303d80SAndy Whitcroft put_page(pagecache_page); 410557303d80SAndy Whitcroft } 4106b4d1d99fSDavid Gibson out_mutex: 4107c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 41080f792cf9SNaoya Horiguchi /* 41090f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 41100f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 41110f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 41120f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 41130f792cf9SNaoya Horiguchi * here without taking refcount. 41140f792cf9SNaoya Horiguchi */ 41150f792cf9SNaoya Horiguchi if (need_wait_lock) 41160f792cf9SNaoya Horiguchi wait_on_page_locked(page); 41171e8f889bSDavid Gibson return ret; 411886e5216fSAdam Litke } 411986e5216fSAdam Litke 41208fb5debcSMike Kravetz /* 41218fb5debcSMike Kravetz * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with 41228fb5debcSMike Kravetz * modifications for huge pages. 41238fb5debcSMike Kravetz */ 41248fb5debcSMike Kravetz int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, 41258fb5debcSMike Kravetz pte_t *dst_pte, 41268fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 41278fb5debcSMike Kravetz unsigned long dst_addr, 41288fb5debcSMike Kravetz unsigned long src_addr, 41298fb5debcSMike Kravetz struct page **pagep) 41308fb5debcSMike Kravetz { 41311e392147SAndrea Arcangeli struct address_space *mapping; 41321e392147SAndrea Arcangeli pgoff_t idx; 41331e392147SAndrea Arcangeli unsigned long size; 41341c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 41358fb5debcSMike Kravetz struct hstate *h = hstate_vma(dst_vma); 41368fb5debcSMike Kravetz pte_t _dst_pte; 41378fb5debcSMike Kravetz spinlock_t *ptl; 41388fb5debcSMike Kravetz int ret; 41398fb5debcSMike Kravetz struct page *page; 41408fb5debcSMike Kravetz 41418fb5debcSMike Kravetz if (!*pagep) { 41428fb5debcSMike Kravetz ret = -ENOMEM; 41438fb5debcSMike Kravetz page = alloc_huge_page(dst_vma, dst_addr, 0); 41448fb5debcSMike Kravetz if (IS_ERR(page)) 41458fb5debcSMike Kravetz goto out; 41468fb5debcSMike Kravetz 41478fb5debcSMike Kravetz ret = copy_huge_page_from_user(page, 41488fb5debcSMike Kravetz (const void __user *) src_addr, 4149810a56b9SMike Kravetz pages_per_huge_page(h), false); 41508fb5debcSMike Kravetz 41518fb5debcSMike Kravetz /* fallback to copy_from_user outside mmap_sem */ 41528fb5debcSMike Kravetz if (unlikely(ret)) { 41539e368259SAndrea Arcangeli ret = -ENOENT; 41548fb5debcSMike Kravetz *pagep = page; 41558fb5debcSMike Kravetz /* don't free the page */ 41568fb5debcSMike Kravetz goto out; 41578fb5debcSMike Kravetz } 41588fb5debcSMike Kravetz } else { 41598fb5debcSMike Kravetz page = *pagep; 41608fb5debcSMike Kravetz *pagep = NULL; 41618fb5debcSMike Kravetz } 41628fb5debcSMike Kravetz 41638fb5debcSMike Kravetz /* 41648fb5debcSMike Kravetz * The memory barrier inside __SetPageUptodate makes sure that 41658fb5debcSMike Kravetz * preceding stores to the page contents become visible before 41668fb5debcSMike Kravetz * the set_pte_at() write. 41678fb5debcSMike Kravetz */ 41688fb5debcSMike Kravetz __SetPageUptodate(page); 41698fb5debcSMike Kravetz 41701e392147SAndrea Arcangeli mapping = dst_vma->vm_file->f_mapping; 41711e392147SAndrea Arcangeli idx = vma_hugecache_offset(h, dst_vma, dst_addr); 41721e392147SAndrea Arcangeli 41731c9e8defSMike Kravetz /* 41741c9e8defSMike Kravetz * If shared, add to page cache 41751c9e8defSMike Kravetz */ 41761c9e8defSMike Kravetz if (vm_shared) { 41771e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 41781e392147SAndrea Arcangeli ret = -EFAULT; 41791e392147SAndrea Arcangeli if (idx >= size) 41801e392147SAndrea Arcangeli goto out_release_nounlock; 41811c9e8defSMike Kravetz 41821e392147SAndrea Arcangeli /* 41831e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 41841e392147SAndrea Arcangeli * huge_add_to_page_cache() below happens through the 41851e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 41861e392147SAndrea Arcangeli * the caller. 41871e392147SAndrea Arcangeli */ 41881c9e8defSMike Kravetz ret = huge_add_to_page_cache(page, mapping, idx); 41891c9e8defSMike Kravetz if (ret) 41901c9e8defSMike Kravetz goto out_release_nounlock; 41911c9e8defSMike Kravetz } 41921c9e8defSMike Kravetz 41938fb5debcSMike Kravetz ptl = huge_pte_lockptr(h, dst_mm, dst_pte); 41948fb5debcSMike Kravetz spin_lock(ptl); 41958fb5debcSMike Kravetz 41961e392147SAndrea Arcangeli /* 41971e392147SAndrea Arcangeli * Recheck the i_size after holding PT lock to make sure not 41981e392147SAndrea Arcangeli * to leave any page mapped (as page_mapped()) beyond the end 41991e392147SAndrea Arcangeli * of the i_size (remove_inode_hugepages() is strict about 42001e392147SAndrea Arcangeli * enforcing that). If we bail out here, we'll also leave a 42011e392147SAndrea Arcangeli * page in the radix tree in the vm_shared case beyond the end 42021e392147SAndrea Arcangeli * of the i_size, but remove_inode_hugepages() will take care 42031e392147SAndrea Arcangeli * of it as soon as we drop the hugetlb_fault_mutex_table. 42041e392147SAndrea Arcangeli */ 42051e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 42061e392147SAndrea Arcangeli ret = -EFAULT; 42071e392147SAndrea Arcangeli if (idx >= size) 42081e392147SAndrea Arcangeli goto out_release_unlock; 42091e392147SAndrea Arcangeli 42108fb5debcSMike Kravetz ret = -EEXIST; 42118fb5debcSMike Kravetz if (!huge_pte_none(huge_ptep_get(dst_pte))) 42128fb5debcSMike Kravetz goto out_release_unlock; 42138fb5debcSMike Kravetz 42141c9e8defSMike Kravetz if (vm_shared) { 42151c9e8defSMike Kravetz page_dup_rmap(page, true); 42161c9e8defSMike Kravetz } else { 42178fb5debcSMike Kravetz ClearPagePrivate(page); 42188fb5debcSMike Kravetz hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); 42191c9e8defSMike Kravetz } 42208fb5debcSMike Kravetz 42218fb5debcSMike Kravetz _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); 42228fb5debcSMike Kravetz if (dst_vma->vm_flags & VM_WRITE) 42238fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 42248fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 42258fb5debcSMike Kravetz 42268fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 42278fb5debcSMike Kravetz 42288fb5debcSMike Kravetz (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, 42298fb5debcSMike Kravetz dst_vma->vm_flags & VM_WRITE); 42308fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 42318fb5debcSMike Kravetz 42328fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 42338fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 42348fb5debcSMike Kravetz 42358fb5debcSMike Kravetz spin_unlock(ptl); 4236cb6acd01SMike Kravetz set_page_huge_active(page); 42371c9e8defSMike Kravetz if (vm_shared) 42381c9e8defSMike Kravetz unlock_page(page); 42398fb5debcSMike Kravetz ret = 0; 42408fb5debcSMike Kravetz out: 42418fb5debcSMike Kravetz return ret; 42428fb5debcSMike Kravetz out_release_unlock: 42438fb5debcSMike Kravetz spin_unlock(ptl); 42441c9e8defSMike Kravetz if (vm_shared) 42451c9e8defSMike Kravetz unlock_page(page); 42465af10dfdSAndrea Arcangeli out_release_nounlock: 42478fb5debcSMike Kravetz put_page(page); 42488fb5debcSMike Kravetz goto out; 42498fb5debcSMike Kravetz } 42508fb5debcSMike Kravetz 425128a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 425263551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 425328a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 425487ffc118SAndrea Arcangeli long i, unsigned int flags, int *nonblocking) 425563551ae0SDavid Gibson { 4256d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 4257d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 425828a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 4259a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 42602be7cfedSDaniel Jordan int err = -EFAULT; 426163551ae0SDavid Gibson 426263551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 426363551ae0SDavid Gibson pte_t *pte; 4264cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 42652a15efc9SHugh Dickins int absent; 426663551ae0SDavid Gibson struct page *page; 426763551ae0SDavid Gibson 42684c887265SAdam Litke /* 426902057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 427002057967SDavid Rientjes * potentially allocating memory. 427102057967SDavid Rientjes */ 4272fa45f116SDavidlohr Bueso if (fatal_signal_pending(current)) { 427302057967SDavid Rientjes remainder = 0; 427402057967SDavid Rientjes break; 427502057967SDavid Rientjes } 427602057967SDavid Rientjes 427702057967SDavid Rientjes /* 42784c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 42792a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 42804c887265SAdam Litke * first, for the page indexing below to work. 4281cb900f41SKirill A. Shutemov * 4282cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 42834c887265SAdam Litke */ 42847868a208SPunit Agrawal pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), 42857868a208SPunit Agrawal huge_page_size(h)); 4286cb900f41SKirill A. Shutemov if (pte) 4287cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 42882a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 428963551ae0SDavid Gibson 42902a15efc9SHugh Dickins /* 42912a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 42923ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 42933ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 42943ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 42953ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 42962a15efc9SHugh Dickins */ 42973ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 42983ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 4299cb900f41SKirill A. Shutemov if (pte) 4300cb900f41SKirill A. Shutemov spin_unlock(ptl); 43012a15efc9SHugh Dickins remainder = 0; 43022a15efc9SHugh Dickins break; 43032a15efc9SHugh Dickins } 43042a15efc9SHugh Dickins 43059cc3a5bdSNaoya Horiguchi /* 43069cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 43079cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 43089cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 43099cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 43109cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 43119cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 43129cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 43139cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 43149cc3a5bdSNaoya Horiguchi */ 43159cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 4316106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 4317106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 43182b740303SSouptick Joarder vm_fault_t ret; 431987ffc118SAndrea Arcangeli unsigned int fault_flags = 0; 43204c887265SAdam Litke 4321cb900f41SKirill A. Shutemov if (pte) 4322cb900f41SKirill A. Shutemov spin_unlock(ptl); 432387ffc118SAndrea Arcangeli if (flags & FOLL_WRITE) 432487ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_WRITE; 432587ffc118SAndrea Arcangeli if (nonblocking) 432687ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY; 432787ffc118SAndrea Arcangeli if (flags & FOLL_NOWAIT) 432887ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY | 432987ffc118SAndrea Arcangeli FAULT_FLAG_RETRY_NOWAIT; 433087ffc118SAndrea Arcangeli if (flags & FOLL_TRIED) { 433187ffc118SAndrea Arcangeli VM_WARN_ON_ONCE(fault_flags & 433287ffc118SAndrea Arcangeli FAULT_FLAG_ALLOW_RETRY); 433387ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_TRIED; 433487ffc118SAndrea Arcangeli } 433587ffc118SAndrea Arcangeli ret = hugetlb_fault(mm, vma, vaddr, fault_flags); 433687ffc118SAndrea Arcangeli if (ret & VM_FAULT_ERROR) { 43372be7cfedSDaniel Jordan err = vm_fault_to_errno(ret, flags); 43381c59827dSHugh Dickins remainder = 0; 43391c59827dSHugh Dickins break; 43401c59827dSHugh Dickins } 434187ffc118SAndrea Arcangeli if (ret & VM_FAULT_RETRY) { 43421ac25013SAndrea Arcangeli if (nonblocking && 43431ac25013SAndrea Arcangeli !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) 434487ffc118SAndrea Arcangeli *nonblocking = 0; 434587ffc118SAndrea Arcangeli *nr_pages = 0; 434687ffc118SAndrea Arcangeli /* 434787ffc118SAndrea Arcangeli * VM_FAULT_RETRY must not return an 434887ffc118SAndrea Arcangeli * error, it will return zero 434987ffc118SAndrea Arcangeli * instead. 435087ffc118SAndrea Arcangeli * 435187ffc118SAndrea Arcangeli * No need to update "position" as the 435287ffc118SAndrea Arcangeli * caller will not check it after 435387ffc118SAndrea Arcangeli * *nr_pages is set to 0. 435487ffc118SAndrea Arcangeli */ 435587ffc118SAndrea Arcangeli return i; 435687ffc118SAndrea Arcangeli } 435787ffc118SAndrea Arcangeli continue; 435887ffc118SAndrea Arcangeli } 435963551ae0SDavid Gibson 4360a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 43617f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 43628fde12caSLinus Torvalds 43638fde12caSLinus Torvalds /* 43648fde12caSLinus Torvalds * Instead of doing 'try_get_page()' below in the same_page 43658fde12caSLinus Torvalds * loop, just check the count once here. 43668fde12caSLinus Torvalds */ 43678fde12caSLinus Torvalds if (unlikely(page_count(page) <= 0)) { 43688fde12caSLinus Torvalds if (pages) { 43698fde12caSLinus Torvalds spin_unlock(ptl); 43708fde12caSLinus Torvalds remainder = 0; 43718fde12caSLinus Torvalds err = -ENOMEM; 43728fde12caSLinus Torvalds break; 43738fde12caSLinus Torvalds } 43748fde12caSLinus Torvalds } 4375d5d4b0aaSChen, Kenneth W same_page: 4376d6692183SChen, Kenneth W if (pages) { 437769d177c2SAndy Whitcroft pages[i] = mem_map_offset(page, pfn_offset); 4378ddc58f27SKirill A. Shutemov get_page(pages[i]); 4379d6692183SChen, Kenneth W } 438063551ae0SDavid Gibson 438163551ae0SDavid Gibson if (vmas) 438263551ae0SDavid Gibson vmas[i] = vma; 438363551ae0SDavid Gibson 438463551ae0SDavid Gibson vaddr += PAGE_SIZE; 4385d5d4b0aaSChen, Kenneth W ++pfn_offset; 438663551ae0SDavid Gibson --remainder; 438763551ae0SDavid Gibson ++i; 4388d5d4b0aaSChen, Kenneth W if (vaddr < vma->vm_end && remainder && 4389a5516438SAndi Kleen pfn_offset < pages_per_huge_page(h)) { 4390d5d4b0aaSChen, Kenneth W /* 4391d5d4b0aaSChen, Kenneth W * We use pfn_offset to avoid touching the pageframes 4392d5d4b0aaSChen, Kenneth W * of this compound page. 4393d5d4b0aaSChen, Kenneth W */ 4394d5d4b0aaSChen, Kenneth W goto same_page; 4395d5d4b0aaSChen, Kenneth W } 4396cb900f41SKirill A. Shutemov spin_unlock(ptl); 439763551ae0SDavid Gibson } 439828a35716SMichel Lespinasse *nr_pages = remainder; 439987ffc118SAndrea Arcangeli /* 440087ffc118SAndrea Arcangeli * setting position is actually required only if remainder is 440187ffc118SAndrea Arcangeli * not zero but it's faster not to add a "if (remainder)" 440287ffc118SAndrea Arcangeli * branch. 440387ffc118SAndrea Arcangeli */ 440463551ae0SDavid Gibson *position = vaddr; 440563551ae0SDavid Gibson 44062be7cfedSDaniel Jordan return i ? i : err; 440763551ae0SDavid Gibson } 44088f860591SZhang, Yanmin 44095491ae7bSAneesh Kumar K.V #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE 44105491ae7bSAneesh Kumar K.V /* 44115491ae7bSAneesh Kumar K.V * ARCHes with special requirements for evicting HUGETLB backing TLB entries can 44125491ae7bSAneesh Kumar K.V * implement this. 44135491ae7bSAneesh Kumar K.V */ 44145491ae7bSAneesh Kumar K.V #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) 44155491ae7bSAneesh Kumar K.V #endif 44165491ae7bSAneesh Kumar K.V 44177da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 44188f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 44198f860591SZhang, Yanmin { 44208f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 44218f860591SZhang, Yanmin unsigned long start = address; 44228f860591SZhang, Yanmin pte_t *ptep; 44238f860591SZhang, Yanmin pte_t pte; 4424a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 44257da4d641SPeter Zijlstra unsigned long pages = 0; 4426dff11abeSMike Kravetz bool shared_pmd = false; 4427ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 4428dff11abeSMike Kravetz 4429dff11abeSMike Kravetz /* 4430dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 4431ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 4432dff11abeSMike Kravetz * range if PMD sharing is possible. 4433dff11abeSMike Kravetz */ 44347269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 44357269f999SJérôme Glisse 0, vma, mm, start, end); 4436ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 44378f860591SZhang, Yanmin 44388f860591SZhang, Yanmin BUG_ON(address >= end); 4439ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 44408f860591SZhang, Yanmin 4441ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 444283cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 4443a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 4444cb900f41SKirill A. Shutemov spinlock_t *ptl; 44457868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, huge_page_size(h)); 44468f860591SZhang, Yanmin if (!ptep) 44478f860591SZhang, Yanmin continue; 4448cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 44497da4d641SPeter Zijlstra if (huge_pmd_unshare(mm, &address, ptep)) { 44507da4d641SPeter Zijlstra pages++; 4451cb900f41SKirill A. Shutemov spin_unlock(ptl); 4452dff11abeSMike Kravetz shared_pmd = true; 445339dde65cSChen, Kenneth W continue; 44547da4d641SPeter Zijlstra } 4455a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 4456a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 4457a8bda28dSNaoya Horiguchi spin_unlock(ptl); 4458a8bda28dSNaoya Horiguchi continue; 4459a8bda28dSNaoya Horiguchi } 4460a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 4461a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 4462a8bda28dSNaoya Horiguchi 4463a8bda28dSNaoya Horiguchi if (is_write_migration_entry(entry)) { 4464a8bda28dSNaoya Horiguchi pte_t newpte; 4465a8bda28dSNaoya Horiguchi 4466a8bda28dSNaoya Horiguchi make_migration_entry_read(&entry); 4467a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 4468e5251fd4SPunit Agrawal set_huge_swap_pte_at(mm, address, ptep, 4469e5251fd4SPunit Agrawal newpte, huge_page_size(h)); 4470a8bda28dSNaoya Horiguchi pages++; 4471a8bda28dSNaoya Horiguchi } 4472a8bda28dSNaoya Horiguchi spin_unlock(ptl); 4473a8bda28dSNaoya Horiguchi continue; 4474a8bda28dSNaoya Horiguchi } 4475a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 4476023bdd00SAneesh Kumar K.V pte_t old_pte; 4477023bdd00SAneesh Kumar K.V 4478023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 4479023bdd00SAneesh Kumar K.V pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); 4480be7517d6STony Lu pte = arch_make_huge_pte(pte, vma, NULL, 0); 4481023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 44827da4d641SPeter Zijlstra pages++; 44838f860591SZhang, Yanmin } 4484cb900f41SKirill A. Shutemov spin_unlock(ptl); 44858f860591SZhang, Yanmin } 4486d833352aSMel Gorman /* 4487c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 4488d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 4489c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 4490dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 4491dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 4492d833352aSMel Gorman */ 4493dff11abeSMike Kravetz if (shared_pmd) 4494ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 4495dff11abeSMike Kravetz else 44965491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 44970f10851eSJérôme Glisse /* 44980f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 44990f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 45000f10851eSJérôme Glisse * 4501ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 45020f10851eSJérôme Glisse */ 450383cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 4504ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 45057da4d641SPeter Zijlstra 45067da4d641SPeter Zijlstra return pages << h->order; 45078f860591SZhang, Yanmin } 45088f860591SZhang, Yanmin 4509a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 4510a1e78772SMel Gorman long from, long to, 45115a6fe125SMel Gorman struct vm_area_struct *vma, 4512ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 4513e4e574b7SAdam Litke { 451417c9d12eSMel Gorman long ret, chg; 4515a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 451690481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 45179119a41eSJoonsoo Kim struct resv_map *resv_map; 45181c5ecae3SMike Kravetz long gbl_reserve; 4519e4e574b7SAdam Litke 452063489f8eSMike Kravetz /* This should never happen */ 452163489f8eSMike Kravetz if (from > to) { 452263489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 452363489f8eSMike Kravetz return -EINVAL; 452463489f8eSMike Kravetz } 452563489f8eSMike Kravetz 4526a1e78772SMel Gorman /* 452717c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 452817c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 452990481622SDavid Gibson * without using reserves 453017c9d12eSMel Gorman */ 4531ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 453217c9d12eSMel Gorman return 0; 453317c9d12eSMel Gorman 453417c9d12eSMel Gorman /* 4535a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 4536a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 4537a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 4538a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 4539a1e78772SMel Gorman */ 45409119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 4541f27a5136SMike Kravetz /* 4542f27a5136SMike Kravetz * resv_map can not be NULL as hugetlb_reserve_pages is only 4543f27a5136SMike Kravetz * called for inodes for which resv_maps were created (see 4544f27a5136SMike Kravetz * hugetlbfs_get_inode). 4545f27a5136SMike Kravetz */ 45464e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 45479119a41eSJoonsoo Kim 45481406ec9bSJoonsoo Kim chg = region_chg(resv_map, from, to); 45499119a41eSJoonsoo Kim 45509119a41eSJoonsoo Kim } else { 45519119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 45525a6fe125SMel Gorman if (!resv_map) 45535a6fe125SMel Gorman return -ENOMEM; 45545a6fe125SMel Gorman 455517c9d12eSMel Gorman chg = to - from; 455617c9d12eSMel Gorman 45575a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 45585a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 45595a6fe125SMel Gorman } 45605a6fe125SMel Gorman 4561c50ac050SDave Hansen if (chg < 0) { 4562c50ac050SDave Hansen ret = chg; 4563c50ac050SDave Hansen goto out_err; 4564c50ac050SDave Hansen } 456517c9d12eSMel Gorman 45661c5ecae3SMike Kravetz /* 45671c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 45681c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 45691c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 45701c5ecae3SMike Kravetz */ 45711c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 45721c5ecae3SMike Kravetz if (gbl_reserve < 0) { 4573c50ac050SDave Hansen ret = -ENOSPC; 4574c50ac050SDave Hansen goto out_err; 4575c50ac050SDave Hansen } 457617c9d12eSMel Gorman 457717c9d12eSMel Gorman /* 457817c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 457990481622SDavid Gibson * Hand the pages back to the subpool if there are not 458017c9d12eSMel Gorman */ 45811c5ecae3SMike Kravetz ret = hugetlb_acct_memory(h, gbl_reserve); 458217c9d12eSMel Gorman if (ret < 0) { 45831c5ecae3SMike Kravetz /* put back original number of pages, chg */ 45841c5ecae3SMike Kravetz (void)hugepage_subpool_put_pages(spool, chg); 4585c50ac050SDave Hansen goto out_err; 458617c9d12eSMel Gorman } 458717c9d12eSMel Gorman 458817c9d12eSMel Gorman /* 458917c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 459017c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 459117c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 459217c9d12eSMel Gorman * the reservation was and the page cache tells how much of 459317c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 459417c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 459517c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 459617c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 459717c9d12eSMel Gorman * else has to be done for private mappings here 459817c9d12eSMel Gorman */ 459933039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 460033039678SMike Kravetz long add = region_add(resv_map, from, to); 460133039678SMike Kravetz 460233039678SMike Kravetz if (unlikely(chg > add)) { 460333039678SMike Kravetz /* 460433039678SMike Kravetz * pages in this range were added to the reserve 460533039678SMike Kravetz * map between region_chg and region_add. This 460633039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 460733039678SMike Kravetz * the subpool and reserve counts modified above 460833039678SMike Kravetz * based on the difference. 460933039678SMike Kravetz */ 461033039678SMike Kravetz long rsv_adjust; 461133039678SMike Kravetz 461233039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 461333039678SMike Kravetz chg - add); 461433039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 461533039678SMike Kravetz } 461633039678SMike Kravetz } 4617a43a8c39SChen, Kenneth W return 0; 4618c50ac050SDave Hansen out_err: 46195e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 4620ff8c0c53SMike Kravetz /* Don't call region_abort if region_chg failed */ 4621ff8c0c53SMike Kravetz if (chg >= 0) 46225e911373SMike Kravetz region_abort(resv_map, from, to); 4623f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 4624f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 4625c50ac050SDave Hansen return ret; 4626a43a8c39SChen, Kenneth W } 4627a43a8c39SChen, Kenneth W 4628b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 4629b5cec28dSMike Kravetz long freed) 4630a43a8c39SChen, Kenneth W { 4631a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 46324e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 46339119a41eSJoonsoo Kim long chg = 0; 463490481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 46351c5ecae3SMike Kravetz long gbl_reserve; 463645c682a6SKen Chen 4637f27a5136SMike Kravetz /* 4638f27a5136SMike Kravetz * Since this routine can be called in the evict inode path for all 4639f27a5136SMike Kravetz * hugetlbfs inodes, resv_map could be NULL. 4640f27a5136SMike Kravetz */ 4641b5cec28dSMike Kravetz if (resv_map) { 4642b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 4643b5cec28dSMike Kravetz /* 4644b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 4645b5cec28dSMike Kravetz * must be split and another region descriptor can not be 4646b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 4647b5cec28dSMike Kravetz */ 4648b5cec28dSMike Kravetz if (chg < 0) 4649b5cec28dSMike Kravetz return chg; 4650b5cec28dSMike Kravetz } 4651b5cec28dSMike Kravetz 465245c682a6SKen Chen spin_lock(&inode->i_lock); 4653e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 465445c682a6SKen Chen spin_unlock(&inode->i_lock); 465545c682a6SKen Chen 46561c5ecae3SMike Kravetz /* 46571c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 46581c5ecae3SMike Kravetz * reservations to be released may be adjusted. 46591c5ecae3SMike Kravetz */ 46601c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 46611c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 4662b5cec28dSMike Kravetz 4663b5cec28dSMike Kravetz return 0; 4664a43a8c39SChen, Kenneth W } 466593f70f90SNaoya Horiguchi 46663212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 46673212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 46683212b535SSteve Capper struct vm_area_struct *vma, 46693212b535SSteve Capper unsigned long addr, pgoff_t idx) 46703212b535SSteve Capper { 46713212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 46723212b535SSteve Capper svma->vm_start; 46733212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 46743212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 46753212b535SSteve Capper 46763212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 4677de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 4678de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 46793212b535SSteve Capper 46803212b535SSteve Capper /* 46813212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 46823212b535SSteve Capper * page table page. 46833212b535SSteve Capper */ 46843212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 46853212b535SSteve Capper vm_flags != svm_flags || 46863212b535SSteve Capper sbase < svma->vm_start || svma->vm_end < s_end) 46873212b535SSteve Capper return 0; 46883212b535SSteve Capper 46893212b535SSteve Capper return saddr; 46903212b535SSteve Capper } 46913212b535SSteve Capper 469231aafb45SNicholas Krause static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) 46933212b535SSteve Capper { 46943212b535SSteve Capper unsigned long base = addr & PUD_MASK; 46953212b535SSteve Capper unsigned long end = base + PUD_SIZE; 46963212b535SSteve Capper 46973212b535SSteve Capper /* 46983212b535SSteve Capper * check on proper vm_flags and page table alignment 46993212b535SSteve Capper */ 4700017b1660SMike Kravetz if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) 470131aafb45SNicholas Krause return true; 470231aafb45SNicholas Krause return false; 47033212b535SSteve Capper } 47043212b535SSteve Capper 47053212b535SSteve Capper /* 4706017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 4707017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 4708017b1660SMike Kravetz * shared pmd mappings. 4709017b1660SMike Kravetz */ 4710017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 4711017b1660SMike Kravetz unsigned long *start, unsigned long *end) 4712017b1660SMike Kravetz { 4713017b1660SMike Kravetz unsigned long check_addr = *start; 4714017b1660SMike Kravetz 4715017b1660SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 4716017b1660SMike Kravetz return; 4717017b1660SMike Kravetz 4718017b1660SMike Kravetz for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) { 4719017b1660SMike Kravetz unsigned long a_start = check_addr & PUD_MASK; 4720017b1660SMike Kravetz unsigned long a_end = a_start + PUD_SIZE; 4721017b1660SMike Kravetz 4722017b1660SMike Kravetz /* 4723017b1660SMike Kravetz * If sharing is possible, adjust start/end if necessary. 4724017b1660SMike Kravetz */ 4725017b1660SMike Kravetz if (range_in_vma(vma, a_start, a_end)) { 4726017b1660SMike Kravetz if (a_start < *start) 4727017b1660SMike Kravetz *start = a_start; 4728017b1660SMike Kravetz if (a_end > *end) 4729017b1660SMike Kravetz *end = a_end; 4730017b1660SMike Kravetz } 4731017b1660SMike Kravetz } 4732017b1660SMike Kravetz } 4733017b1660SMike Kravetz 4734017b1660SMike Kravetz /* 47353212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 47363212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 47373212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 4738ddeaab32SMike Kravetz * code much cleaner. pmd allocation is essential for the shared case because 4739ddeaab32SMike Kravetz * pud has to be populated inside the same i_mmap_rwsem section - otherwise 4740ddeaab32SMike Kravetz * racing tasks could either miss the sharing (see huge_pte_offset) or select a 4741ddeaab32SMike Kravetz * bad pmd for sharing. 47423212b535SSteve Capper */ 47433212b535SSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 47443212b535SSteve Capper { 47453212b535SSteve Capper struct vm_area_struct *vma = find_vma(mm, addr); 47463212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 47473212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 47483212b535SSteve Capper vma->vm_pgoff; 47493212b535SSteve Capper struct vm_area_struct *svma; 47503212b535SSteve Capper unsigned long saddr; 47513212b535SSteve Capper pte_t *spte = NULL; 47523212b535SSteve Capper pte_t *pte; 4753cb900f41SKirill A. Shutemov spinlock_t *ptl; 47543212b535SSteve Capper 47553212b535SSteve Capper if (!vma_shareable(vma, addr)) 47563212b535SSteve Capper return (pte_t *)pmd_alloc(mm, pud, addr); 47573212b535SSteve Capper 4758ddeaab32SMike Kravetz i_mmap_lock_write(mapping); 47593212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 47603212b535SSteve Capper if (svma == vma) 47613212b535SSteve Capper continue; 47623212b535SSteve Capper 47633212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 47643212b535SSteve Capper if (saddr) { 47657868a208SPunit Agrawal spte = huge_pte_offset(svma->vm_mm, saddr, 47667868a208SPunit Agrawal vma_mmu_pagesize(svma)); 47673212b535SSteve Capper if (spte) { 47683212b535SSteve Capper get_page(virt_to_page(spte)); 47693212b535SSteve Capper break; 47703212b535SSteve Capper } 47713212b535SSteve Capper } 47723212b535SSteve Capper } 47733212b535SSteve Capper 47743212b535SSteve Capper if (!spte) 47753212b535SSteve Capper goto out; 47763212b535SSteve Capper 47778bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 4778dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 47793212b535SSteve Capper pud_populate(mm, pud, 47803212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 4781c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 4782dc6c9a35SKirill A. Shutemov } else { 47833212b535SSteve Capper put_page(virt_to_page(spte)); 4784dc6c9a35SKirill A. Shutemov } 4785cb900f41SKirill A. Shutemov spin_unlock(ptl); 47863212b535SSteve Capper out: 47873212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 4788ddeaab32SMike Kravetz i_mmap_unlock_write(mapping); 47893212b535SSteve Capper return pte; 47903212b535SSteve Capper } 47913212b535SSteve Capper 47923212b535SSteve Capper /* 47933212b535SSteve Capper * unmap huge page backed by shared pte. 47943212b535SSteve Capper * 47953212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 47963212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 47973212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 47983212b535SSteve Capper * 4799ddeaab32SMike Kravetz * called with page table lock held. 48003212b535SSteve Capper * 48013212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 48023212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 48033212b535SSteve Capper */ 48043212b535SSteve Capper int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 48053212b535SSteve Capper { 48063212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 4807c2febafcSKirill A. Shutemov p4d_t *p4d = p4d_offset(pgd, *addr); 4808c2febafcSKirill A. Shutemov pud_t *pud = pud_offset(p4d, *addr); 48093212b535SSteve Capper 48103212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 48113212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 48123212b535SSteve Capper return 0; 48133212b535SSteve Capper 48143212b535SSteve Capper pud_clear(pud); 48153212b535SSteve Capper put_page(virt_to_page(ptep)); 4816dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 48173212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 48183212b535SSteve Capper return 1; 48193212b535SSteve Capper } 48209e5fc74cSSteve Capper #define want_pmd_share() (1) 48219e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 48229e5fc74cSSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 48239e5fc74cSSteve Capper { 48249e5fc74cSSteve Capper return NULL; 48259e5fc74cSSteve Capper } 4826e81f2d22SZhang Zhen 4827e81f2d22SZhang Zhen int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 4828e81f2d22SZhang Zhen { 4829e81f2d22SZhang Zhen return 0; 4830e81f2d22SZhang Zhen } 4831017b1660SMike Kravetz 4832017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 4833017b1660SMike Kravetz unsigned long *start, unsigned long *end) 4834017b1660SMike Kravetz { 4835017b1660SMike Kravetz } 48369e5fc74cSSteve Capper #define want_pmd_share() (0) 48373212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 48383212b535SSteve Capper 48399e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 48409e5fc74cSSteve Capper pte_t *huge_pte_alloc(struct mm_struct *mm, 48419e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 48429e5fc74cSSteve Capper { 48439e5fc74cSSteve Capper pgd_t *pgd; 4844c2febafcSKirill A. Shutemov p4d_t *p4d; 48459e5fc74cSSteve Capper pud_t *pud; 48469e5fc74cSSteve Capper pte_t *pte = NULL; 48479e5fc74cSSteve Capper 48489e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 4849f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 4850f4f0a3d8SKirill A. Shutemov if (!p4d) 4851f4f0a3d8SKirill A. Shutemov return NULL; 4852c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 48539e5fc74cSSteve Capper if (pud) { 48549e5fc74cSSteve Capper if (sz == PUD_SIZE) { 48559e5fc74cSSteve Capper pte = (pte_t *)pud; 48569e5fc74cSSteve Capper } else { 48579e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 48589e5fc74cSSteve Capper if (want_pmd_share() && pud_none(*pud)) 48599e5fc74cSSteve Capper pte = huge_pmd_share(mm, addr, pud); 48609e5fc74cSSteve Capper else 48619e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 48629e5fc74cSSteve Capper } 48639e5fc74cSSteve Capper } 48644e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 48659e5fc74cSSteve Capper 48669e5fc74cSSteve Capper return pte; 48679e5fc74cSSteve Capper } 48689e5fc74cSSteve Capper 48699b19df29SPunit Agrawal /* 48709b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 48719b19df29SPunit Agrawal * entry at address @addr 48729b19df29SPunit Agrawal * 48739b19df29SPunit Agrawal * Return: Pointer to page table or swap entry (PUD or PMD) for 48749b19df29SPunit Agrawal * address @addr, or NULL if a p*d_none() entry is encountered and the 48759b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 48769b19df29SPunit Agrawal * table. 48779b19df29SPunit Agrawal */ 48787868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 48797868a208SPunit Agrawal unsigned long addr, unsigned long sz) 48809e5fc74cSSteve Capper { 48819e5fc74cSSteve Capper pgd_t *pgd; 4882c2febafcSKirill A. Shutemov p4d_t *p4d; 48839e5fc74cSSteve Capper pud_t *pud; 4884c2febafcSKirill A. Shutemov pmd_t *pmd; 48859e5fc74cSSteve Capper 48869e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 4887c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 4888c2febafcSKirill A. Shutemov return NULL; 4889c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 4890c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 4891c2febafcSKirill A. Shutemov return NULL; 48929b19df29SPunit Agrawal 4893c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 48949b19df29SPunit Agrawal if (sz != PUD_SIZE && pud_none(*pud)) 4895c2febafcSKirill A. Shutemov return NULL; 48969b19df29SPunit Agrawal /* hugepage or swap? */ 48979b19df29SPunit Agrawal if (pud_huge(*pud) || !pud_present(*pud)) 48989e5fc74cSSteve Capper return (pte_t *)pud; 48999b19df29SPunit Agrawal 49009e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 49019b19df29SPunit Agrawal if (sz != PMD_SIZE && pmd_none(*pmd)) 49029b19df29SPunit Agrawal return NULL; 49039b19df29SPunit Agrawal /* hugepage or swap? */ 49049b19df29SPunit Agrawal if (pmd_huge(*pmd) || !pmd_present(*pmd)) 49059e5fc74cSSteve Capper return (pte_t *)pmd; 49069b19df29SPunit Agrawal 49079b19df29SPunit Agrawal return NULL; 49089e5fc74cSSteve Capper } 49099e5fc74cSSteve Capper 491061f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 491161f77edaSNaoya Horiguchi 491261f77edaSNaoya Horiguchi /* 491361f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 491461f77edaSNaoya Horiguchi * behavior. 491561f77edaSNaoya Horiguchi */ 491661f77edaSNaoya Horiguchi struct page * __weak 491761f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 491861f77edaSNaoya Horiguchi int write) 491961f77edaSNaoya Horiguchi { 492061f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 492161f77edaSNaoya Horiguchi } 492261f77edaSNaoya Horiguchi 492361f77edaSNaoya Horiguchi struct page * __weak 49244dc71451SAneesh Kumar K.V follow_huge_pd(struct vm_area_struct *vma, 49254dc71451SAneesh Kumar K.V unsigned long address, hugepd_t hpd, int flags, int pdshift) 49264dc71451SAneesh Kumar K.V { 49274dc71451SAneesh Kumar K.V WARN(1, "hugepd follow called with no support for hugepage directory format\n"); 49284dc71451SAneesh Kumar K.V return NULL; 49294dc71451SAneesh Kumar K.V } 49304dc71451SAneesh Kumar K.V 49314dc71451SAneesh Kumar K.V struct page * __weak 49329e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 4933e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 49349e5fc74cSSteve Capper { 4935e66f17ffSNaoya Horiguchi struct page *page = NULL; 4936e66f17ffSNaoya Horiguchi spinlock_t *ptl; 4937c9d398faSNaoya Horiguchi pte_t pte; 4938e66f17ffSNaoya Horiguchi retry: 4939e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 4940e66f17ffSNaoya Horiguchi spin_lock(ptl); 4941e66f17ffSNaoya Horiguchi /* 4942e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 4943e66f17ffSNaoya Horiguchi * unmapped from other threads. 4944e66f17ffSNaoya Horiguchi */ 4945e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 4946e66f17ffSNaoya Horiguchi goto out; 4947c9d398faSNaoya Horiguchi pte = huge_ptep_get((pte_t *)pmd); 4948c9d398faSNaoya Horiguchi if (pte_present(pte)) { 494997534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 4950e66f17ffSNaoya Horiguchi if (flags & FOLL_GET) 4951e66f17ffSNaoya Horiguchi get_page(page); 4952e66f17ffSNaoya Horiguchi } else { 4953c9d398faSNaoya Horiguchi if (is_hugetlb_entry_migration(pte)) { 4954e66f17ffSNaoya Horiguchi spin_unlock(ptl); 4955e66f17ffSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pmd, ptl); 4956e66f17ffSNaoya Horiguchi goto retry; 4957e66f17ffSNaoya Horiguchi } 4958e66f17ffSNaoya Horiguchi /* 4959e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 4960e66f17ffSNaoya Horiguchi * follow_page_mask(). 4961e66f17ffSNaoya Horiguchi */ 4962e66f17ffSNaoya Horiguchi } 4963e66f17ffSNaoya Horiguchi out: 4964e66f17ffSNaoya Horiguchi spin_unlock(ptl); 49659e5fc74cSSteve Capper return page; 49669e5fc74cSSteve Capper } 49679e5fc74cSSteve Capper 496861f77edaSNaoya Horiguchi struct page * __weak 49699e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 4970e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 49719e5fc74cSSteve Capper { 4972e66f17ffSNaoya Horiguchi if (flags & FOLL_GET) 4973e66f17ffSNaoya Horiguchi return NULL; 49749e5fc74cSSteve Capper 4975e66f17ffSNaoya Horiguchi return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 49769e5fc74cSSteve Capper } 49779e5fc74cSSteve Capper 4978faaa5b62SAnshuman Khandual struct page * __weak 4979faaa5b62SAnshuman Khandual follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) 4980faaa5b62SAnshuman Khandual { 4981faaa5b62SAnshuman Khandual if (flags & FOLL_GET) 4982faaa5b62SAnshuman Khandual return NULL; 4983faaa5b62SAnshuman Khandual 4984faaa5b62SAnshuman Khandual return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); 4985faaa5b62SAnshuman Khandual } 4986faaa5b62SAnshuman Khandual 498731caf665SNaoya Horiguchi bool isolate_huge_page(struct page *page, struct list_head *list) 498831caf665SNaoya Horiguchi { 4989bcc54222SNaoya Horiguchi bool ret = true; 4990bcc54222SNaoya Horiguchi 4991309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 499231caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 4993bcc54222SNaoya Horiguchi if (!page_huge_active(page) || !get_page_unless_zero(page)) { 4994bcc54222SNaoya Horiguchi ret = false; 4995bcc54222SNaoya Horiguchi goto unlock; 4996bcc54222SNaoya Horiguchi } 4997bcc54222SNaoya Horiguchi clear_page_huge_active(page); 499831caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 4999bcc54222SNaoya Horiguchi unlock: 500031caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 5001bcc54222SNaoya Horiguchi return ret; 500231caf665SNaoya Horiguchi } 500331caf665SNaoya Horiguchi 500431caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 500531caf665SNaoya Horiguchi { 5006309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 500731caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 5008bcc54222SNaoya Horiguchi set_page_huge_active(page); 500931caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 501031caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 501131caf665SNaoya Horiguchi put_page(page); 501231caf665SNaoya Horiguchi } 5013ab5ac90aSMichal Hocko 5014ab5ac90aSMichal Hocko void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) 5015ab5ac90aSMichal Hocko { 5016ab5ac90aSMichal Hocko struct hstate *h = page_hstate(oldpage); 5017ab5ac90aSMichal Hocko 5018ab5ac90aSMichal Hocko hugetlb_cgroup_migrate(oldpage, newpage); 5019ab5ac90aSMichal Hocko set_page_owner_migrate_reason(newpage, reason); 5020ab5ac90aSMichal Hocko 5021ab5ac90aSMichal Hocko /* 5022ab5ac90aSMichal Hocko * transfer temporary state of the new huge page. This is 5023ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 5024ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 5025ab5ac90aSMichal Hocko * the temporary status. 5026ab5ac90aSMichal Hocko * 5027ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 5028ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 5029ab5ac90aSMichal Hocko * the per-node's. 5030ab5ac90aSMichal Hocko */ 5031ab5ac90aSMichal Hocko if (PageHugeTemporary(newpage)) { 5032ab5ac90aSMichal Hocko int old_nid = page_to_nid(oldpage); 5033ab5ac90aSMichal Hocko int new_nid = page_to_nid(newpage); 5034ab5ac90aSMichal Hocko 5035ab5ac90aSMichal Hocko SetPageHugeTemporary(oldpage); 5036ab5ac90aSMichal Hocko ClearPageHugeTemporary(newpage); 5037ab5ac90aSMichal Hocko 5038ab5ac90aSMichal Hocko spin_lock(&hugetlb_lock); 5039ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 5040ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 5041ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 5042ab5ac90aSMichal Hocko } 5043ab5ac90aSMichal Hocko spin_unlock(&hugetlb_lock); 5044ab5ac90aSMichal Hocko } 5045ab5ac90aSMichal Hocko } 5046