11da177e4SLinus Torvalds /* 21da177e4SLinus Torvalds * Generic hugetlb support. 36d49e352SNadia Yvette Chambers * (C) Nadia Yvette Chambers, April 2004 41da177e4SLinus Torvalds */ 51da177e4SLinus Torvalds #include <linux/list.h> 61da177e4SLinus Torvalds #include <linux/init.h> 71da177e4SLinus Torvalds #include <linux/mm.h> 8e1759c21SAlexey Dobriyan #include <linux/seq_file.h> 91da177e4SLinus Torvalds #include <linux/sysctl.h> 101da177e4SLinus Torvalds #include <linux/highmem.h> 11cddb8a5cSAndrea Arcangeli #include <linux/mmu_notifier.h> 121da177e4SLinus Torvalds #include <linux/nodemask.h> 1363551ae0SDavid Gibson #include <linux/pagemap.h> 145da7ca86SChristoph Lameter #include <linux/mempolicy.h> 153b32123dSGideon Israel Dsouza #include <linux/compiler.h> 16aea47ff3SChristoph Lameter #include <linux/cpuset.h> 173935baa9SDavid Gibson #include <linux/mutex.h> 18aa888a74SAndi Kleen #include <linux/bootmem.h> 19a3437870SNishanth Aravamudan #include <linux/sysfs.h> 205a0e3ad6STejun Heo #include <linux/slab.h> 210fe6e20bSNaoya Horiguchi #include <linux/rmap.h> 22fd6a03edSNaoya Horiguchi #include <linux/swap.h> 23fd6a03edSNaoya Horiguchi #include <linux/swapops.h> 24c8721bbbSNaoya Horiguchi #include <linux/page-isolation.h> 258382d914SDavidlohr Bueso #include <linux/jhash.h> 26d6606683SLinus Torvalds 2763551ae0SDavid Gibson #include <asm/page.h> 2863551ae0SDavid Gibson #include <asm/pgtable.h> 2924669e58SAneesh Kumar K.V #include <asm/tlb.h> 3063551ae0SDavid Gibson 3124669e58SAneesh Kumar K.V #include <linux/io.h> 3263551ae0SDavid Gibson #include <linux/hugetlb.h> 339dd540e2SAneesh Kumar K.V #include <linux/hugetlb_cgroup.h> 349a305230SLee Schermerhorn #include <linux/node.h> 357835e98bSNick Piggin #include "internal.h" 361da177e4SLinus Torvalds 37753162cdSAndrey Ryabinin int hugepages_treat_as_movable; 38a5516438SAndi Kleen 39c3f38a38SAneesh Kumar K.V int hugetlb_max_hstate __read_mostly; 40e5ff2159SAndi Kleen unsigned int default_hstate_idx; 41e5ff2159SAndi Kleen struct hstate hstates[HUGE_MAX_HSTATE]; 42641844f5SNaoya Horiguchi /* 43641844f5SNaoya Horiguchi * Minimum page order among possible hugepage sizes, set to a proper value 44641844f5SNaoya Horiguchi * at boot time. 45641844f5SNaoya Horiguchi */ 46641844f5SNaoya Horiguchi static unsigned int minimum_order __read_mostly = UINT_MAX; 47e5ff2159SAndi Kleen 4853ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 4953ba51d2SJon Tollefson 50e5ff2159SAndi Kleen /* for command line parsing */ 51e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 52e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 53e11bfbfcSNick Piggin static unsigned long __initdata default_hstate_size; 549fee021dSVaishali Thakkar static bool __initdata parsed_valid_hugepagesz = true; 55e5ff2159SAndi Kleen 563935baa9SDavid Gibson /* 5731caf665SNaoya Horiguchi * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, 5831caf665SNaoya Horiguchi * free_huge_pages, and surplus_huge_pages. 593935baa9SDavid Gibson */ 60c3f38a38SAneesh Kumar K.V DEFINE_SPINLOCK(hugetlb_lock); 610bd0f9fbSEric Paris 628382d914SDavidlohr Bueso /* 638382d914SDavidlohr Bueso * Serializes faults on the same logical page. This is used to 648382d914SDavidlohr Bueso * prevent spurious OOMs when the hugepage pool is fully utilized. 658382d914SDavidlohr Bueso */ 668382d914SDavidlohr Bueso static int num_fault_mutexes; 67c672c7f2SMike Kravetz struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; 688382d914SDavidlohr Bueso 697ca02d0aSMike Kravetz /* Forward declaration */ 707ca02d0aSMike Kravetz static int hugetlb_acct_memory(struct hstate *h, long delta); 717ca02d0aSMike Kravetz 7290481622SDavid Gibson static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) 7390481622SDavid Gibson { 7490481622SDavid Gibson bool free = (spool->count == 0) && (spool->used_hpages == 0); 7590481622SDavid Gibson 7690481622SDavid Gibson spin_unlock(&spool->lock); 7790481622SDavid Gibson 7890481622SDavid Gibson /* If no pages are used, and no other handles to the subpool 797ca02d0aSMike Kravetz * remain, give up any reservations mased on minimum size and 807ca02d0aSMike Kravetz * free the subpool */ 817ca02d0aSMike Kravetz if (free) { 827ca02d0aSMike Kravetz if (spool->min_hpages != -1) 837ca02d0aSMike Kravetz hugetlb_acct_memory(spool->hstate, 847ca02d0aSMike Kravetz -spool->min_hpages); 8590481622SDavid Gibson kfree(spool); 8690481622SDavid Gibson } 877ca02d0aSMike Kravetz } 8890481622SDavid Gibson 897ca02d0aSMike Kravetz struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, 907ca02d0aSMike Kravetz long min_hpages) 9190481622SDavid Gibson { 9290481622SDavid Gibson struct hugepage_subpool *spool; 9390481622SDavid Gibson 94c6a91820SMike Kravetz spool = kzalloc(sizeof(*spool), GFP_KERNEL); 9590481622SDavid Gibson if (!spool) 9690481622SDavid Gibson return NULL; 9790481622SDavid Gibson 9890481622SDavid Gibson spin_lock_init(&spool->lock); 9990481622SDavid Gibson spool->count = 1; 1007ca02d0aSMike Kravetz spool->max_hpages = max_hpages; 1017ca02d0aSMike Kravetz spool->hstate = h; 1027ca02d0aSMike Kravetz spool->min_hpages = min_hpages; 1037ca02d0aSMike Kravetz 1047ca02d0aSMike Kravetz if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { 1057ca02d0aSMike Kravetz kfree(spool); 1067ca02d0aSMike Kravetz return NULL; 1077ca02d0aSMike Kravetz } 1087ca02d0aSMike Kravetz spool->rsv_hpages = min_hpages; 10990481622SDavid Gibson 11090481622SDavid Gibson return spool; 11190481622SDavid Gibson } 11290481622SDavid Gibson 11390481622SDavid Gibson void hugepage_put_subpool(struct hugepage_subpool *spool) 11490481622SDavid Gibson { 11590481622SDavid Gibson spin_lock(&spool->lock); 11690481622SDavid Gibson BUG_ON(!spool->count); 11790481622SDavid Gibson spool->count--; 11890481622SDavid Gibson unlock_or_release_subpool(spool); 11990481622SDavid Gibson } 12090481622SDavid Gibson 1211c5ecae3SMike Kravetz /* 1221c5ecae3SMike Kravetz * Subpool accounting for allocating and reserving pages. 1231c5ecae3SMike Kravetz * Return -ENOMEM if there are not enough resources to satisfy the 1241c5ecae3SMike Kravetz * the request. Otherwise, return the number of pages by which the 1251c5ecae3SMike Kravetz * global pools must be adjusted (upward). The returned value may 1261c5ecae3SMike Kravetz * only be different than the passed value (delta) in the case where 1271c5ecae3SMike Kravetz * a subpool minimum size must be manitained. 1281c5ecae3SMike Kravetz */ 1291c5ecae3SMike Kravetz static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, 13090481622SDavid Gibson long delta) 13190481622SDavid Gibson { 1321c5ecae3SMike Kravetz long ret = delta; 13390481622SDavid Gibson 13490481622SDavid Gibson if (!spool) 1351c5ecae3SMike Kravetz return ret; 13690481622SDavid Gibson 13790481622SDavid Gibson spin_lock(&spool->lock); 13890481622SDavid Gibson 1391c5ecae3SMike Kravetz if (spool->max_hpages != -1) { /* maximum size accounting */ 1401c5ecae3SMike Kravetz if ((spool->used_hpages + delta) <= spool->max_hpages) 1411c5ecae3SMike Kravetz spool->used_hpages += delta; 1421c5ecae3SMike Kravetz else { 1431c5ecae3SMike Kravetz ret = -ENOMEM; 1441c5ecae3SMike Kravetz goto unlock_ret; 1451c5ecae3SMike Kravetz } 1461c5ecae3SMike Kravetz } 1471c5ecae3SMike Kravetz 14809a95e29SMike Kravetz /* minimum size accounting */ 14909a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->rsv_hpages) { 1501c5ecae3SMike Kravetz if (delta > spool->rsv_hpages) { 1511c5ecae3SMike Kravetz /* 1521c5ecae3SMike Kravetz * Asking for more reserves than those already taken on 1531c5ecae3SMike Kravetz * behalf of subpool. Return difference. 1541c5ecae3SMike Kravetz */ 1551c5ecae3SMike Kravetz ret = delta - spool->rsv_hpages; 1561c5ecae3SMike Kravetz spool->rsv_hpages = 0; 1571c5ecae3SMike Kravetz } else { 1581c5ecae3SMike Kravetz ret = 0; /* reserves already accounted for */ 1591c5ecae3SMike Kravetz spool->rsv_hpages -= delta; 1601c5ecae3SMike Kravetz } 1611c5ecae3SMike Kravetz } 1621c5ecae3SMike Kravetz 1631c5ecae3SMike Kravetz unlock_ret: 1641c5ecae3SMike Kravetz spin_unlock(&spool->lock); 16590481622SDavid Gibson return ret; 16690481622SDavid Gibson } 16790481622SDavid Gibson 1681c5ecae3SMike Kravetz /* 1691c5ecae3SMike Kravetz * Subpool accounting for freeing and unreserving pages. 1701c5ecae3SMike Kravetz * Return the number of global page reservations that must be dropped. 1711c5ecae3SMike Kravetz * The return value may only be different than the passed value (delta) 1721c5ecae3SMike Kravetz * in the case where a subpool minimum size must be maintained. 1731c5ecae3SMike Kravetz */ 1741c5ecae3SMike Kravetz static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, 17590481622SDavid Gibson long delta) 17690481622SDavid Gibson { 1771c5ecae3SMike Kravetz long ret = delta; 1781c5ecae3SMike Kravetz 17990481622SDavid Gibson if (!spool) 1801c5ecae3SMike Kravetz return delta; 18190481622SDavid Gibson 18290481622SDavid Gibson spin_lock(&spool->lock); 1831c5ecae3SMike Kravetz 1841c5ecae3SMike Kravetz if (spool->max_hpages != -1) /* maximum size accounting */ 18590481622SDavid Gibson spool->used_hpages -= delta; 1861c5ecae3SMike Kravetz 18709a95e29SMike Kravetz /* minimum size accounting */ 18809a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { 1891c5ecae3SMike Kravetz if (spool->rsv_hpages + delta <= spool->min_hpages) 1901c5ecae3SMike Kravetz ret = 0; 1911c5ecae3SMike Kravetz else 1921c5ecae3SMike Kravetz ret = spool->rsv_hpages + delta - spool->min_hpages; 1931c5ecae3SMike Kravetz 1941c5ecae3SMike Kravetz spool->rsv_hpages += delta; 1951c5ecae3SMike Kravetz if (spool->rsv_hpages > spool->min_hpages) 1961c5ecae3SMike Kravetz spool->rsv_hpages = spool->min_hpages; 1971c5ecae3SMike Kravetz } 1981c5ecae3SMike Kravetz 1991c5ecae3SMike Kravetz /* 2001c5ecae3SMike Kravetz * If hugetlbfs_put_super couldn't free spool due to an outstanding 2011c5ecae3SMike Kravetz * quota reference, free it now. 2021c5ecae3SMike Kravetz */ 20390481622SDavid Gibson unlock_or_release_subpool(spool); 2041c5ecae3SMike Kravetz 2051c5ecae3SMike Kravetz return ret; 20690481622SDavid Gibson } 20790481622SDavid Gibson 20890481622SDavid Gibson static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 20990481622SDavid Gibson { 21090481622SDavid Gibson return HUGETLBFS_SB(inode->i_sb)->spool; 21190481622SDavid Gibson } 21290481622SDavid Gibson 21390481622SDavid Gibson static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 21490481622SDavid Gibson { 215496ad9aaSAl Viro return subpool_inode(file_inode(vma->vm_file)); 21690481622SDavid Gibson } 21790481622SDavid Gibson 218e7c4b0bfSAndy Whitcroft /* 21996822904SAndy Whitcroft * Region tracking -- allows tracking of reservations and instantiated pages 22096822904SAndy Whitcroft * across the pages in a mapping. 22184afd99bSAndy Whitcroft * 2221dd308a7SMike Kravetz * The region data structures are embedded into a resv_map and protected 2231dd308a7SMike Kravetz * by a resv_map's lock. The set of regions within the resv_map represent 2241dd308a7SMike Kravetz * reservations for huge pages, or huge pages that have already been 2251dd308a7SMike Kravetz * instantiated within the map. The from and to elements are huge page 2261dd308a7SMike Kravetz * indicies into the associated mapping. from indicates the starting index 2271dd308a7SMike Kravetz * of the region. to represents the first index past the end of the region. 2281dd308a7SMike Kravetz * 2291dd308a7SMike Kravetz * For example, a file region structure with from == 0 and to == 4 represents 2301dd308a7SMike Kravetz * four huge pages in a mapping. It is important to note that the to element 2311dd308a7SMike Kravetz * represents the first element past the end of the region. This is used in 2321dd308a7SMike Kravetz * arithmetic as 4(to) - 0(from) = 4 huge pages in the region. 2331dd308a7SMike Kravetz * 2341dd308a7SMike Kravetz * Interval notation of the form [from, to) will be used to indicate that 2351dd308a7SMike Kravetz * the endpoint from is inclusive and to is exclusive. 23696822904SAndy Whitcroft */ 23796822904SAndy Whitcroft struct file_region { 23896822904SAndy Whitcroft struct list_head link; 23996822904SAndy Whitcroft long from; 24096822904SAndy Whitcroft long to; 24196822904SAndy Whitcroft }; 24296822904SAndy Whitcroft 2431dd308a7SMike Kravetz /* 2441dd308a7SMike Kravetz * Add the huge page range represented by [f, t) to the reserve 2455e911373SMike Kravetz * map. In the normal case, existing regions will be expanded 2465e911373SMike Kravetz * to accommodate the specified range. Sufficient regions should 2475e911373SMike Kravetz * exist for expansion due to the previous call to region_chg 2485e911373SMike Kravetz * with the same range. However, it is possible that region_del 2495e911373SMike Kravetz * could have been called after region_chg and modifed the map 2505e911373SMike Kravetz * in such a way that no region exists to be expanded. In this 2515e911373SMike Kravetz * case, pull a region descriptor from the cache associated with 2525e911373SMike Kravetz * the map and use that for the new range. 253cf3ad20bSMike Kravetz * 254cf3ad20bSMike Kravetz * Return the number of new huge pages added to the map. This 255cf3ad20bSMike Kravetz * number is greater than or equal to zero. 2561dd308a7SMike Kravetz */ 2571406ec9bSJoonsoo Kim static long region_add(struct resv_map *resv, long f, long t) 25896822904SAndy Whitcroft { 2591406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 26096822904SAndy Whitcroft struct file_region *rg, *nrg, *trg; 261cf3ad20bSMike Kravetz long add = 0; 26296822904SAndy Whitcroft 2637b24d861SDavidlohr Bueso spin_lock(&resv->lock); 26496822904SAndy Whitcroft /* Locate the region we are either in or before. */ 26596822904SAndy Whitcroft list_for_each_entry(rg, head, link) 26696822904SAndy Whitcroft if (f <= rg->to) 26796822904SAndy Whitcroft break; 26896822904SAndy Whitcroft 2695e911373SMike Kravetz /* 2705e911373SMike Kravetz * If no region exists which can be expanded to include the 2715e911373SMike Kravetz * specified range, the list must have been modified by an 2725e911373SMike Kravetz * interleving call to region_del(). Pull a region descriptor 2735e911373SMike Kravetz * from the cache and use it for this range. 2745e911373SMike Kravetz */ 2755e911373SMike Kravetz if (&rg->link == head || t < rg->from) { 2765e911373SMike Kravetz VM_BUG_ON(resv->region_cache_count <= 0); 2775e911373SMike Kravetz 2785e911373SMike Kravetz resv->region_cache_count--; 2795e911373SMike Kravetz nrg = list_first_entry(&resv->region_cache, struct file_region, 2805e911373SMike Kravetz link); 2815e911373SMike Kravetz list_del(&nrg->link); 2825e911373SMike Kravetz 2835e911373SMike Kravetz nrg->from = f; 2845e911373SMike Kravetz nrg->to = t; 2855e911373SMike Kravetz list_add(&nrg->link, rg->link.prev); 2865e911373SMike Kravetz 2875e911373SMike Kravetz add += t - f; 2885e911373SMike Kravetz goto out_locked; 2895e911373SMike Kravetz } 2905e911373SMike Kravetz 29196822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 29296822904SAndy Whitcroft if (f > rg->from) 29396822904SAndy Whitcroft f = rg->from; 29496822904SAndy Whitcroft 29596822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 29696822904SAndy Whitcroft nrg = rg; 29796822904SAndy Whitcroft list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 29896822904SAndy Whitcroft if (&rg->link == head) 29996822904SAndy Whitcroft break; 30096822904SAndy Whitcroft if (rg->from > t) 30196822904SAndy Whitcroft break; 30296822904SAndy Whitcroft 30396822904SAndy Whitcroft /* If this area reaches higher then extend our area to 30496822904SAndy Whitcroft * include it completely. If this is not the first area 30596822904SAndy Whitcroft * which we intend to reuse, free it. */ 30696822904SAndy Whitcroft if (rg->to > t) 30796822904SAndy Whitcroft t = rg->to; 30896822904SAndy Whitcroft if (rg != nrg) { 309cf3ad20bSMike Kravetz /* Decrement return value by the deleted range. 310cf3ad20bSMike Kravetz * Another range will span this area so that by 311cf3ad20bSMike Kravetz * end of routine add will be >= zero 312cf3ad20bSMike Kravetz */ 313cf3ad20bSMike Kravetz add -= (rg->to - rg->from); 31496822904SAndy Whitcroft list_del(&rg->link); 31596822904SAndy Whitcroft kfree(rg); 31696822904SAndy Whitcroft } 31796822904SAndy Whitcroft } 318cf3ad20bSMike Kravetz 319cf3ad20bSMike Kravetz add += (nrg->from - f); /* Added to beginning of region */ 32096822904SAndy Whitcroft nrg->from = f; 321cf3ad20bSMike Kravetz add += t - nrg->to; /* Added to end of region */ 32296822904SAndy Whitcroft nrg->to = t; 323cf3ad20bSMike Kravetz 3245e911373SMike Kravetz out_locked: 3255e911373SMike Kravetz resv->adds_in_progress--; 3267b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 327cf3ad20bSMike Kravetz VM_BUG_ON(add < 0); 328cf3ad20bSMike Kravetz return add; 32996822904SAndy Whitcroft } 33096822904SAndy Whitcroft 3311dd308a7SMike Kravetz /* 3321dd308a7SMike Kravetz * Examine the existing reserve map and determine how many 3331dd308a7SMike Kravetz * huge pages in the specified range [f, t) are NOT currently 3341dd308a7SMike Kravetz * represented. This routine is called before a subsequent 3351dd308a7SMike Kravetz * call to region_add that will actually modify the reserve 3361dd308a7SMike Kravetz * map to add the specified range [f, t). region_chg does 3371dd308a7SMike Kravetz * not change the number of huge pages represented by the 3381dd308a7SMike Kravetz * map. However, if the existing regions in the map can not 3391dd308a7SMike Kravetz * be expanded to represent the new range, a new file_region 3401dd308a7SMike Kravetz * structure is added to the map as a placeholder. This is 3411dd308a7SMike Kravetz * so that the subsequent region_add call will have all the 3421dd308a7SMike Kravetz * regions it needs and will not fail. 3431dd308a7SMike Kravetz * 3445e911373SMike Kravetz * Upon entry, region_chg will also examine the cache of region descriptors 3455e911373SMike Kravetz * associated with the map. If there are not enough descriptors cached, one 3465e911373SMike Kravetz * will be allocated for the in progress add operation. 3475e911373SMike Kravetz * 3485e911373SMike Kravetz * Returns the number of huge pages that need to be added to the existing 3495e911373SMike Kravetz * reservation map for the range [f, t). This number is greater or equal to 3505e911373SMike Kravetz * zero. -ENOMEM is returned if a new file_region structure or cache entry 3515e911373SMike Kravetz * is needed and can not be allocated. 3521dd308a7SMike Kravetz */ 3531406ec9bSJoonsoo Kim static long region_chg(struct resv_map *resv, long f, long t) 35496822904SAndy Whitcroft { 3551406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 3567b24d861SDavidlohr Bueso struct file_region *rg, *nrg = NULL; 35796822904SAndy Whitcroft long chg = 0; 35896822904SAndy Whitcroft 3597b24d861SDavidlohr Bueso retry: 3607b24d861SDavidlohr Bueso spin_lock(&resv->lock); 3615e911373SMike Kravetz retry_locked: 3625e911373SMike Kravetz resv->adds_in_progress++; 3635e911373SMike Kravetz 3645e911373SMike Kravetz /* 3655e911373SMike Kravetz * Check for sufficient descriptors in the cache to accommodate 3665e911373SMike Kravetz * the number of in progress add operations. 3675e911373SMike Kravetz */ 3685e911373SMike Kravetz if (resv->adds_in_progress > resv->region_cache_count) { 3695e911373SMike Kravetz struct file_region *trg; 3705e911373SMike Kravetz 3715e911373SMike Kravetz VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1); 3725e911373SMike Kravetz /* Must drop lock to allocate a new descriptor. */ 3735e911373SMike Kravetz resv->adds_in_progress--; 3745e911373SMike Kravetz spin_unlock(&resv->lock); 3755e911373SMike Kravetz 3765e911373SMike Kravetz trg = kmalloc(sizeof(*trg), GFP_KERNEL); 377dbe409e4SMike Kravetz if (!trg) { 378dbe409e4SMike Kravetz kfree(nrg); 3795e911373SMike Kravetz return -ENOMEM; 380dbe409e4SMike Kravetz } 3815e911373SMike Kravetz 3825e911373SMike Kravetz spin_lock(&resv->lock); 3835e911373SMike Kravetz list_add(&trg->link, &resv->region_cache); 3845e911373SMike Kravetz resv->region_cache_count++; 3855e911373SMike Kravetz goto retry_locked; 3865e911373SMike Kravetz } 3875e911373SMike Kravetz 38896822904SAndy Whitcroft /* Locate the region we are before or in. */ 38996822904SAndy Whitcroft list_for_each_entry(rg, head, link) 39096822904SAndy Whitcroft if (f <= rg->to) 39196822904SAndy Whitcroft break; 39296822904SAndy Whitcroft 39396822904SAndy Whitcroft /* If we are below the current region then a new region is required. 39496822904SAndy Whitcroft * Subtle, allocate a new region at the position but make it zero 39596822904SAndy Whitcroft * size such that we can guarantee to record the reservation. */ 39696822904SAndy Whitcroft if (&rg->link == head || t < rg->from) { 3977b24d861SDavidlohr Bueso if (!nrg) { 3985e911373SMike Kravetz resv->adds_in_progress--; 3997b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 40096822904SAndy Whitcroft nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 40196822904SAndy Whitcroft if (!nrg) 40296822904SAndy Whitcroft return -ENOMEM; 4037b24d861SDavidlohr Bueso 40496822904SAndy Whitcroft nrg->from = f; 40596822904SAndy Whitcroft nrg->to = f; 40696822904SAndy Whitcroft INIT_LIST_HEAD(&nrg->link); 4077b24d861SDavidlohr Bueso goto retry; 4087b24d861SDavidlohr Bueso } 40996822904SAndy Whitcroft 4107b24d861SDavidlohr Bueso list_add(&nrg->link, rg->link.prev); 4117b24d861SDavidlohr Bueso chg = t - f; 4127b24d861SDavidlohr Bueso goto out_nrg; 41396822904SAndy Whitcroft } 41496822904SAndy Whitcroft 41596822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 41696822904SAndy Whitcroft if (f > rg->from) 41796822904SAndy Whitcroft f = rg->from; 41896822904SAndy Whitcroft chg = t - f; 41996822904SAndy Whitcroft 42096822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 42196822904SAndy Whitcroft list_for_each_entry(rg, rg->link.prev, link) { 42296822904SAndy Whitcroft if (&rg->link == head) 42396822904SAndy Whitcroft break; 42496822904SAndy Whitcroft if (rg->from > t) 4257b24d861SDavidlohr Bueso goto out; 42696822904SAndy Whitcroft 42725985edcSLucas De Marchi /* We overlap with this area, if it extends further than 42896822904SAndy Whitcroft * us then we must extend ourselves. Account for its 42996822904SAndy Whitcroft * existing reservation. */ 43096822904SAndy Whitcroft if (rg->to > t) { 43196822904SAndy Whitcroft chg += rg->to - t; 43296822904SAndy Whitcroft t = rg->to; 43396822904SAndy Whitcroft } 43496822904SAndy Whitcroft chg -= rg->to - rg->from; 43596822904SAndy Whitcroft } 4367b24d861SDavidlohr Bueso 4377b24d861SDavidlohr Bueso out: 4387b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 4397b24d861SDavidlohr Bueso /* We already know we raced and no longer need the new region */ 4407b24d861SDavidlohr Bueso kfree(nrg); 4417b24d861SDavidlohr Bueso return chg; 4427b24d861SDavidlohr Bueso out_nrg: 4437b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 44496822904SAndy Whitcroft return chg; 44596822904SAndy Whitcroft } 44696822904SAndy Whitcroft 4471dd308a7SMike Kravetz /* 4485e911373SMike Kravetz * Abort the in progress add operation. The adds_in_progress field 4495e911373SMike Kravetz * of the resv_map keeps track of the operations in progress between 4505e911373SMike Kravetz * calls to region_chg and region_add. Operations are sometimes 4515e911373SMike Kravetz * aborted after the call to region_chg. In such cases, region_abort 4525e911373SMike Kravetz * is called to decrement the adds_in_progress counter. 4535e911373SMike Kravetz * 4545e911373SMike Kravetz * NOTE: The range arguments [f, t) are not needed or used in this 4555e911373SMike Kravetz * routine. They are kept to make reading the calling code easier as 4565e911373SMike Kravetz * arguments will match the associated region_chg call. 4575e911373SMike Kravetz */ 4585e911373SMike Kravetz static void region_abort(struct resv_map *resv, long f, long t) 4595e911373SMike Kravetz { 4605e911373SMike Kravetz spin_lock(&resv->lock); 4615e911373SMike Kravetz VM_BUG_ON(!resv->region_cache_count); 4625e911373SMike Kravetz resv->adds_in_progress--; 4635e911373SMike Kravetz spin_unlock(&resv->lock); 4645e911373SMike Kravetz } 4655e911373SMike Kravetz 4665e911373SMike Kravetz /* 467feba16e2SMike Kravetz * Delete the specified range [f, t) from the reserve map. If the 468feba16e2SMike Kravetz * t parameter is LONG_MAX, this indicates that ALL regions after f 469feba16e2SMike Kravetz * should be deleted. Locate the regions which intersect [f, t) 470feba16e2SMike Kravetz * and either trim, delete or split the existing regions. 471feba16e2SMike Kravetz * 472feba16e2SMike Kravetz * Returns the number of huge pages deleted from the reserve map. 473feba16e2SMike Kravetz * In the normal case, the return value is zero or more. In the 474feba16e2SMike Kravetz * case where a region must be split, a new region descriptor must 475feba16e2SMike Kravetz * be allocated. If the allocation fails, -ENOMEM will be returned. 476feba16e2SMike Kravetz * NOTE: If the parameter t == LONG_MAX, then we will never split 477feba16e2SMike Kravetz * a region and possibly return -ENOMEM. Callers specifying 478feba16e2SMike Kravetz * t == LONG_MAX do not need to check for -ENOMEM error. 4791dd308a7SMike Kravetz */ 480feba16e2SMike Kravetz static long region_del(struct resv_map *resv, long f, long t) 48196822904SAndy Whitcroft { 4821406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 48396822904SAndy Whitcroft struct file_region *rg, *trg; 484feba16e2SMike Kravetz struct file_region *nrg = NULL; 485feba16e2SMike Kravetz long del = 0; 48696822904SAndy Whitcroft 487feba16e2SMike Kravetz retry: 4887b24d861SDavidlohr Bueso spin_lock(&resv->lock); 489feba16e2SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 490dbe409e4SMike Kravetz /* 491dbe409e4SMike Kravetz * Skip regions before the range to be deleted. file_region 492dbe409e4SMike Kravetz * ranges are normally of the form [from, to). However, there 493dbe409e4SMike Kravetz * may be a "placeholder" entry in the map which is of the form 494dbe409e4SMike Kravetz * (from, to) with from == to. Check for placeholder entries 495dbe409e4SMike Kravetz * at the beginning of the range to be deleted. 496dbe409e4SMike Kravetz */ 497dbe409e4SMike Kravetz if (rg->to <= f && (rg->to != rg->from || rg->to != f)) 498feba16e2SMike Kravetz continue; 499dbe409e4SMike Kravetz 500feba16e2SMike Kravetz if (rg->from >= t) 50196822904SAndy Whitcroft break; 50296822904SAndy Whitcroft 503feba16e2SMike Kravetz if (f > rg->from && t < rg->to) { /* Must split region */ 504feba16e2SMike Kravetz /* 505feba16e2SMike Kravetz * Check for an entry in the cache before dropping 506feba16e2SMike Kravetz * lock and attempting allocation. 507feba16e2SMike Kravetz */ 508feba16e2SMike Kravetz if (!nrg && 509feba16e2SMike Kravetz resv->region_cache_count > resv->adds_in_progress) { 510feba16e2SMike Kravetz nrg = list_first_entry(&resv->region_cache, 511feba16e2SMike Kravetz struct file_region, 512feba16e2SMike Kravetz link); 513feba16e2SMike Kravetz list_del(&nrg->link); 514feba16e2SMike Kravetz resv->region_cache_count--; 51596822904SAndy Whitcroft } 51696822904SAndy Whitcroft 517feba16e2SMike Kravetz if (!nrg) { 518feba16e2SMike Kravetz spin_unlock(&resv->lock); 519feba16e2SMike Kravetz nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 520feba16e2SMike Kravetz if (!nrg) 521feba16e2SMike Kravetz return -ENOMEM; 522feba16e2SMike Kravetz goto retry; 523feba16e2SMike Kravetz } 524feba16e2SMike Kravetz 525feba16e2SMike Kravetz del += t - f; 526feba16e2SMike Kravetz 527feba16e2SMike Kravetz /* New entry for end of split region */ 528feba16e2SMike Kravetz nrg->from = t; 529feba16e2SMike Kravetz nrg->to = rg->to; 530feba16e2SMike Kravetz INIT_LIST_HEAD(&nrg->link); 531feba16e2SMike Kravetz 532feba16e2SMike Kravetz /* Original entry is trimmed */ 533feba16e2SMike Kravetz rg->to = f; 534feba16e2SMike Kravetz 535feba16e2SMike Kravetz list_add(&nrg->link, &rg->link); 536feba16e2SMike Kravetz nrg = NULL; 53796822904SAndy Whitcroft break; 538feba16e2SMike Kravetz } 539feba16e2SMike Kravetz 540feba16e2SMike Kravetz if (f <= rg->from && t >= rg->to) { /* Remove entire region */ 541feba16e2SMike Kravetz del += rg->to - rg->from; 54296822904SAndy Whitcroft list_del(&rg->link); 54396822904SAndy Whitcroft kfree(rg); 544feba16e2SMike Kravetz continue; 54596822904SAndy Whitcroft } 5467b24d861SDavidlohr Bueso 547feba16e2SMike Kravetz if (f <= rg->from) { /* Trim beginning of region */ 548feba16e2SMike Kravetz del += t - rg->from; 549feba16e2SMike Kravetz rg->from = t; 550feba16e2SMike Kravetz } else { /* Trim end of region */ 551feba16e2SMike Kravetz del += rg->to - f; 552feba16e2SMike Kravetz rg->to = f; 553feba16e2SMike Kravetz } 554feba16e2SMike Kravetz } 555feba16e2SMike Kravetz 5567b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 557feba16e2SMike Kravetz kfree(nrg); 558feba16e2SMike Kravetz return del; 55996822904SAndy Whitcroft } 56096822904SAndy Whitcroft 5611dd308a7SMike Kravetz /* 562b5cec28dSMike Kravetz * A rare out of memory error was encountered which prevented removal of 563b5cec28dSMike Kravetz * the reserve map region for a page. The huge page itself was free'ed 564b5cec28dSMike Kravetz * and removed from the page cache. This routine will adjust the subpool 565b5cec28dSMike Kravetz * usage count, and the global reserve count if needed. By incrementing 566b5cec28dSMike Kravetz * these counts, the reserve map entry which could not be deleted will 567b5cec28dSMike Kravetz * appear as a "reserved" entry instead of simply dangling with incorrect 568b5cec28dSMike Kravetz * counts. 569b5cec28dSMike Kravetz */ 570b5cec28dSMike Kravetz void hugetlb_fix_reserve_counts(struct inode *inode, bool restore_reserve) 571b5cec28dSMike Kravetz { 572b5cec28dSMike Kravetz struct hugepage_subpool *spool = subpool_inode(inode); 573b5cec28dSMike Kravetz long rsv_adjust; 574b5cec28dSMike Kravetz 575b5cec28dSMike Kravetz rsv_adjust = hugepage_subpool_get_pages(spool, 1); 576b5cec28dSMike Kravetz if (restore_reserve && rsv_adjust) { 577b5cec28dSMike Kravetz struct hstate *h = hstate_inode(inode); 578b5cec28dSMike Kravetz 579b5cec28dSMike Kravetz hugetlb_acct_memory(h, 1); 580b5cec28dSMike Kravetz } 581b5cec28dSMike Kravetz } 582b5cec28dSMike Kravetz 583b5cec28dSMike Kravetz /* 5841dd308a7SMike Kravetz * Count and return the number of huge pages in the reserve map 5851dd308a7SMike Kravetz * that intersect with the range [f, t). 5861dd308a7SMike Kravetz */ 5871406ec9bSJoonsoo Kim static long region_count(struct resv_map *resv, long f, long t) 58884afd99bSAndy Whitcroft { 5891406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 59084afd99bSAndy Whitcroft struct file_region *rg; 59184afd99bSAndy Whitcroft long chg = 0; 59284afd99bSAndy Whitcroft 5937b24d861SDavidlohr Bueso spin_lock(&resv->lock); 59484afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 59584afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 596f2135a4aSWang Sheng-Hui long seg_from; 597f2135a4aSWang Sheng-Hui long seg_to; 59884afd99bSAndy Whitcroft 59984afd99bSAndy Whitcroft if (rg->to <= f) 60084afd99bSAndy Whitcroft continue; 60184afd99bSAndy Whitcroft if (rg->from >= t) 60284afd99bSAndy Whitcroft break; 60384afd99bSAndy Whitcroft 60484afd99bSAndy Whitcroft seg_from = max(rg->from, f); 60584afd99bSAndy Whitcroft seg_to = min(rg->to, t); 60684afd99bSAndy Whitcroft 60784afd99bSAndy Whitcroft chg += seg_to - seg_from; 60884afd99bSAndy Whitcroft } 6097b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 61084afd99bSAndy Whitcroft 61184afd99bSAndy Whitcroft return chg; 61284afd99bSAndy Whitcroft } 61384afd99bSAndy Whitcroft 61496822904SAndy Whitcroft /* 615e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 616e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 617e7c4b0bfSAndy Whitcroft */ 618a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 619a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 620e7c4b0bfSAndy Whitcroft { 621a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 622a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 623e7c4b0bfSAndy Whitcroft } 624e7c4b0bfSAndy Whitcroft 6250fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 6260fe6e20bSNaoya Horiguchi unsigned long address) 6270fe6e20bSNaoya Horiguchi { 6280fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 6290fe6e20bSNaoya Horiguchi } 630dee41079SDan Williams EXPORT_SYMBOL_GPL(linear_hugepage_index); 6310fe6e20bSNaoya Horiguchi 63284afd99bSAndy Whitcroft /* 63308fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 63408fba699SMel Gorman * cases this will be same size as used by the page table entries. 63508fba699SMel Gorman */ 63608fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 63708fba699SMel Gorman { 63808fba699SMel Gorman struct hstate *hstate; 63908fba699SMel Gorman 64008fba699SMel Gorman if (!is_vm_hugetlb_page(vma)) 64108fba699SMel Gorman return PAGE_SIZE; 64208fba699SMel Gorman 64308fba699SMel Gorman hstate = hstate_vma(vma); 64408fba699SMel Gorman 6452415cf12SWanpeng Li return 1UL << huge_page_shift(hstate); 64608fba699SMel Gorman } 647f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 64808fba699SMel Gorman 64908fba699SMel Gorman /* 6503340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 6513340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 6523340289dSMel Gorman * architectures where it differs, an architecture-specific version of this 6533340289dSMel Gorman * function is required. 6543340289dSMel Gorman */ 6553340289dSMel Gorman #ifndef vma_mmu_pagesize 6563340289dSMel Gorman unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 6573340289dSMel Gorman { 6583340289dSMel Gorman return vma_kernel_pagesize(vma); 6593340289dSMel Gorman } 6603340289dSMel Gorman #endif 6613340289dSMel Gorman 6623340289dSMel Gorman /* 66384afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 66484afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 66584afd99bSAndy Whitcroft * alignment. 66684afd99bSAndy Whitcroft */ 66784afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 66884afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 66904f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 67084afd99bSAndy Whitcroft 671a1e78772SMel Gorman /* 672a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 673a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 674a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 675a1e78772SMel Gorman * 676a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 677a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 678a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 679a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 68084afd99bSAndy Whitcroft * 68184afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 68284afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 68384afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 68484afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 68584afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 68684afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 68784afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 68884afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 689a1e78772SMel Gorman */ 690e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 691e7c4b0bfSAndy Whitcroft { 692e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 693e7c4b0bfSAndy Whitcroft } 694e7c4b0bfSAndy Whitcroft 695e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 696e7c4b0bfSAndy Whitcroft unsigned long value) 697e7c4b0bfSAndy Whitcroft { 698e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 699e7c4b0bfSAndy Whitcroft } 700e7c4b0bfSAndy Whitcroft 7019119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 70284afd99bSAndy Whitcroft { 70384afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 7045e911373SMike Kravetz struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 7055e911373SMike Kravetz 7065e911373SMike Kravetz if (!resv_map || !rg) { 7075e911373SMike Kravetz kfree(resv_map); 7085e911373SMike Kravetz kfree(rg); 70984afd99bSAndy Whitcroft return NULL; 7105e911373SMike Kravetz } 71184afd99bSAndy Whitcroft 71284afd99bSAndy Whitcroft kref_init(&resv_map->refs); 7137b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 71484afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 71584afd99bSAndy Whitcroft 7165e911373SMike Kravetz resv_map->adds_in_progress = 0; 7175e911373SMike Kravetz 7185e911373SMike Kravetz INIT_LIST_HEAD(&resv_map->region_cache); 7195e911373SMike Kravetz list_add(&rg->link, &resv_map->region_cache); 7205e911373SMike Kravetz resv_map->region_cache_count = 1; 7215e911373SMike Kravetz 72284afd99bSAndy Whitcroft return resv_map; 72384afd99bSAndy Whitcroft } 72484afd99bSAndy Whitcroft 7259119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 72684afd99bSAndy Whitcroft { 72784afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 7285e911373SMike Kravetz struct list_head *head = &resv_map->region_cache; 7295e911373SMike Kravetz struct file_region *rg, *trg; 73084afd99bSAndy Whitcroft 73184afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 732feba16e2SMike Kravetz region_del(resv_map, 0, LONG_MAX); 7335e911373SMike Kravetz 7345e911373SMike Kravetz /* ... and any entries left in the cache */ 7355e911373SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 7365e911373SMike Kravetz list_del(&rg->link); 7375e911373SMike Kravetz kfree(rg); 7385e911373SMike Kravetz } 7395e911373SMike Kravetz 7405e911373SMike Kravetz VM_BUG_ON(resv_map->adds_in_progress); 7415e911373SMike Kravetz 74284afd99bSAndy Whitcroft kfree(resv_map); 74384afd99bSAndy Whitcroft } 74484afd99bSAndy Whitcroft 7454e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 7464e35f483SJoonsoo Kim { 7474e35f483SJoonsoo Kim return inode->i_mapping->private_data; 7484e35f483SJoonsoo Kim } 7494e35f483SJoonsoo Kim 75084afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 751a1e78772SMel Gorman { 75281d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 7534e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 7544e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 7554e35f483SJoonsoo Kim struct inode *inode = mapping->host; 7564e35f483SJoonsoo Kim 7574e35f483SJoonsoo Kim return inode_resv_map(inode); 7584e35f483SJoonsoo Kim 7594e35f483SJoonsoo Kim } else { 76084afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 76184afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 7624e35f483SJoonsoo Kim } 763a1e78772SMel Gorman } 764a1e78772SMel Gorman 76584afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 766a1e78772SMel Gorman { 76781d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 76881d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 769a1e78772SMel Gorman 77084afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 77184afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 77204f2cbe3SMel Gorman } 77304f2cbe3SMel Gorman 77404f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 77504f2cbe3SMel Gorman { 77681d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 77781d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 778e7c4b0bfSAndy Whitcroft 779e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 78004f2cbe3SMel Gorman } 78104f2cbe3SMel Gorman 78204f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 78304f2cbe3SMel Gorman { 78481d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 785e7c4b0bfSAndy Whitcroft 786e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 787a1e78772SMel Gorman } 788a1e78772SMel Gorman 78904f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 790a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 791a1e78772SMel Gorman { 79281d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 793f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 794a1e78772SMel Gorman vma->vm_private_data = (void *)0; 795a1e78772SMel Gorman } 796a1e78772SMel Gorman 797a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 798559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 799a1e78772SMel Gorman { 800af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 801af0ed73eSJoonsoo Kim /* 802af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 803af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 804af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 805af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 806af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 807af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 808af0ed73eSJoonsoo Kim * properly, so add work-around here. 809af0ed73eSJoonsoo Kim */ 810af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 811559ec2f8SNicholas Krause return true; 812af0ed73eSJoonsoo Kim else 813559ec2f8SNicholas Krause return false; 814af0ed73eSJoonsoo Kim } 815a63884e9SJoonsoo Kim 816a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 8171fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 8181fb1b0e9SMike Kravetz /* 8191fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 8201fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 8211fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 8221fb1b0e9SMike Kravetz * fallocate. In this case, there really are no reverves to 8231fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 8241fb1b0e9SMike Kravetz */ 8251fb1b0e9SMike Kravetz if (chg) 8261fb1b0e9SMike Kravetz return false; 8271fb1b0e9SMike Kravetz else 828559ec2f8SNicholas Krause return true; 8291fb1b0e9SMike Kravetz } 830a63884e9SJoonsoo Kim 831a63884e9SJoonsoo Kim /* 832a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 833a63884e9SJoonsoo Kim * private mappings. 834a63884e9SJoonsoo Kim */ 8357f09ca51SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 836559ec2f8SNicholas Krause return true; 837a63884e9SJoonsoo Kim 838559ec2f8SNicholas Krause return false; 839a1e78772SMel Gorman } 840a1e78772SMel Gorman 841a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 8421da177e4SLinus Torvalds { 8431da177e4SLinus Torvalds int nid = page_to_nid(page); 8440edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 845a5516438SAndi Kleen h->free_huge_pages++; 846a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 8471da177e4SLinus Torvalds } 8481da177e4SLinus Torvalds 849bf50bab2SNaoya Horiguchi static struct page *dequeue_huge_page_node(struct hstate *h, int nid) 850bf50bab2SNaoya Horiguchi { 851bf50bab2SNaoya Horiguchi struct page *page; 852bf50bab2SNaoya Horiguchi 853c8721bbbSNaoya Horiguchi list_for_each_entry(page, &h->hugepage_freelists[nid], lru) 854c8721bbbSNaoya Horiguchi if (!is_migrate_isolate_page(page)) 855c8721bbbSNaoya Horiguchi break; 856c8721bbbSNaoya Horiguchi /* 857c8721bbbSNaoya Horiguchi * if 'non-isolated free hugepage' not found on the list, 858c8721bbbSNaoya Horiguchi * the allocation fails. 859c8721bbbSNaoya Horiguchi */ 860c8721bbbSNaoya Horiguchi if (&h->hugepage_freelists[nid] == &page->lru) 861bf50bab2SNaoya Horiguchi return NULL; 8620edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 863a9869b83SNaoya Horiguchi set_page_refcounted(page); 864bf50bab2SNaoya Horiguchi h->free_huge_pages--; 865bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 866bf50bab2SNaoya Horiguchi return page; 867bf50bab2SNaoya Horiguchi } 868bf50bab2SNaoya Horiguchi 86986cdb465SNaoya Horiguchi /* Movability of hugepages depends on migration support. */ 87086cdb465SNaoya Horiguchi static inline gfp_t htlb_alloc_mask(struct hstate *h) 87186cdb465SNaoya Horiguchi { 872100873d7SNaoya Horiguchi if (hugepages_treat_as_movable || hugepage_migration_supported(h)) 87386cdb465SNaoya Horiguchi return GFP_HIGHUSER_MOVABLE; 87486cdb465SNaoya Horiguchi else 87586cdb465SNaoya Horiguchi return GFP_HIGHUSER; 87686cdb465SNaoya Horiguchi } 87786cdb465SNaoya Horiguchi 878a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 879a5516438SAndi Kleen struct vm_area_struct *vma, 880af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 881af0ed73eSJoonsoo Kim long chg) 8821da177e4SLinus Torvalds { 883b1c12cbcSKonstantin Khlebnikov struct page *page = NULL; 884480eccf9SLee Schermerhorn struct mempolicy *mpol; 88519770b32SMel Gorman nodemask_t *nodemask; 886c0ff7453SMiao Xie struct zonelist *zonelist; 887dd1a239fSMel Gorman struct zone *zone; 888dd1a239fSMel Gorman struct zoneref *z; 889cc9a6c87SMel Gorman unsigned int cpuset_mems_cookie; 8901da177e4SLinus Torvalds 891a1e78772SMel Gorman /* 892a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 893a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 894a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 895a1e78772SMel Gorman */ 896af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg) && 897a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 898c0ff7453SMiao Xie goto err; 899a1e78772SMel Gorman 90004f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 901a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 9026eab04a8SJustin P. Mattock goto err; 90304f2cbe3SMel Gorman 9049966c4bbSJoonsoo Kim retry_cpuset: 905d26914d1SMel Gorman cpuset_mems_cookie = read_mems_allowed_begin(); 9069966c4bbSJoonsoo Kim zonelist = huge_zonelist(vma, address, 90786cdb465SNaoya Horiguchi htlb_alloc_mask(h), &mpol, &nodemask); 9089966c4bbSJoonsoo Kim 90919770b32SMel Gorman for_each_zone_zonelist_nodemask(zone, z, zonelist, 91019770b32SMel Gorman MAX_NR_ZONES - 1, nodemask) { 911344736f2SVladimir Davydov if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) { 912bf50bab2SNaoya Horiguchi page = dequeue_huge_page_node(h, zone_to_nid(zone)); 913bf50bab2SNaoya Horiguchi if (page) { 914af0ed73eSJoonsoo Kim if (avoid_reserve) 915af0ed73eSJoonsoo Kim break; 916af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg)) 917af0ed73eSJoonsoo Kim break; 918af0ed73eSJoonsoo Kim 91907443a85SJoonsoo Kim SetPagePrivate(page); 920a63884e9SJoonsoo Kim h->resv_huge_pages--; 9215ab3ee7bSKen Chen break; 9221da177e4SLinus Torvalds } 9233abf7afdSAndrew Morton } 924bf50bab2SNaoya Horiguchi } 925cc9a6c87SMel Gorman 926cc9a6c87SMel Gorman mpol_cond_put(mpol); 927d26914d1SMel Gorman if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) 928cc9a6c87SMel Gorman goto retry_cpuset; 929cc9a6c87SMel Gorman return page; 930cc9a6c87SMel Gorman 931c0ff7453SMiao Xie err: 932cc9a6c87SMel Gorman return NULL; 9331da177e4SLinus Torvalds } 9341da177e4SLinus Torvalds 9351cac6f2cSLuiz Capitulino /* 9361cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 9371cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 9381cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 9391cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 9401cac6f2cSLuiz Capitulino * node for alloc or free. 9411cac6f2cSLuiz Capitulino */ 9421cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 9431cac6f2cSLuiz Capitulino { 9440edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 9451cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 9461cac6f2cSLuiz Capitulino 9471cac6f2cSLuiz Capitulino return nid; 9481cac6f2cSLuiz Capitulino } 9491cac6f2cSLuiz Capitulino 9501cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 9511cac6f2cSLuiz Capitulino { 9521cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 9531cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 9541cac6f2cSLuiz Capitulino return nid; 9551cac6f2cSLuiz Capitulino } 9561cac6f2cSLuiz Capitulino 9571cac6f2cSLuiz Capitulino /* 9581cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 9591cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 9601cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 9611cac6f2cSLuiz Capitulino * mask. 9621cac6f2cSLuiz Capitulino */ 9631cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 9641cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 9651cac6f2cSLuiz Capitulino { 9661cac6f2cSLuiz Capitulino int nid; 9671cac6f2cSLuiz Capitulino 9681cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 9691cac6f2cSLuiz Capitulino 9701cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 9711cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 9721cac6f2cSLuiz Capitulino 9731cac6f2cSLuiz Capitulino return nid; 9741cac6f2cSLuiz Capitulino } 9751cac6f2cSLuiz Capitulino 9761cac6f2cSLuiz Capitulino /* 9771cac6f2cSLuiz Capitulino * helper for free_pool_huge_page() - return the previously saved 9781cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 9791cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 9801cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 9811cac6f2cSLuiz Capitulino */ 9821cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 9831cac6f2cSLuiz Capitulino { 9841cac6f2cSLuiz Capitulino int nid; 9851cac6f2cSLuiz Capitulino 9861cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 9871cac6f2cSLuiz Capitulino 9881cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 9891cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 9901cac6f2cSLuiz Capitulino 9911cac6f2cSLuiz Capitulino return nid; 9921cac6f2cSLuiz Capitulino } 9931cac6f2cSLuiz Capitulino 9941cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 9951cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 9961cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 9971cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 9981cac6f2cSLuiz Capitulino nr_nodes--) 9991cac6f2cSLuiz Capitulino 10001cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 10011cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 10021cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 10031cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 10041cac6f2cSLuiz Capitulino nr_nodes--) 10051cac6f2cSLuiz Capitulino 1006080fe206SVlastimil Babka #if defined(CONFIG_X86_64) && ((defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)) 1007944d9fecSLuiz Capitulino static void destroy_compound_gigantic_page(struct page *page, 1008d00181b9SKirill A. Shutemov unsigned int order) 1009944d9fecSLuiz Capitulino { 1010944d9fecSLuiz Capitulino int i; 1011944d9fecSLuiz Capitulino int nr_pages = 1 << order; 1012944d9fecSLuiz Capitulino struct page *p = page + 1; 1013944d9fecSLuiz Capitulino 1014944d9fecSLuiz Capitulino for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 10151d798ca3SKirill A. Shutemov clear_compound_head(p); 1016944d9fecSLuiz Capitulino set_page_refcounted(p); 1017944d9fecSLuiz Capitulino } 1018944d9fecSLuiz Capitulino 1019944d9fecSLuiz Capitulino set_compound_order(page, 0); 1020944d9fecSLuiz Capitulino __ClearPageHead(page); 1021944d9fecSLuiz Capitulino } 1022944d9fecSLuiz Capitulino 1023d00181b9SKirill A. Shutemov static void free_gigantic_page(struct page *page, unsigned int order) 1024944d9fecSLuiz Capitulino { 1025944d9fecSLuiz Capitulino free_contig_range(page_to_pfn(page), 1 << order); 1026944d9fecSLuiz Capitulino } 1027944d9fecSLuiz Capitulino 1028944d9fecSLuiz Capitulino static int __alloc_gigantic_page(unsigned long start_pfn, 1029944d9fecSLuiz Capitulino unsigned long nr_pages) 1030944d9fecSLuiz Capitulino { 1031944d9fecSLuiz Capitulino unsigned long end_pfn = start_pfn + nr_pages; 1032944d9fecSLuiz Capitulino return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE); 1033944d9fecSLuiz Capitulino } 1034944d9fecSLuiz Capitulino 1035f44b2ddaSJoonsoo Kim static bool pfn_range_valid_gigantic(struct zone *z, 1036f44b2ddaSJoonsoo Kim unsigned long start_pfn, unsigned long nr_pages) 1037944d9fecSLuiz Capitulino { 1038944d9fecSLuiz Capitulino unsigned long i, end_pfn = start_pfn + nr_pages; 1039944d9fecSLuiz Capitulino struct page *page; 1040944d9fecSLuiz Capitulino 1041944d9fecSLuiz Capitulino for (i = start_pfn; i < end_pfn; i++) { 1042944d9fecSLuiz Capitulino if (!pfn_valid(i)) 1043944d9fecSLuiz Capitulino return false; 1044944d9fecSLuiz Capitulino 1045944d9fecSLuiz Capitulino page = pfn_to_page(i); 1046944d9fecSLuiz Capitulino 1047f44b2ddaSJoonsoo Kim if (page_zone(page) != z) 1048f44b2ddaSJoonsoo Kim return false; 1049f44b2ddaSJoonsoo Kim 1050944d9fecSLuiz Capitulino if (PageReserved(page)) 1051944d9fecSLuiz Capitulino return false; 1052944d9fecSLuiz Capitulino 1053944d9fecSLuiz Capitulino if (page_count(page) > 0) 1054944d9fecSLuiz Capitulino return false; 1055944d9fecSLuiz Capitulino 1056944d9fecSLuiz Capitulino if (PageHuge(page)) 1057944d9fecSLuiz Capitulino return false; 1058944d9fecSLuiz Capitulino } 1059944d9fecSLuiz Capitulino 1060944d9fecSLuiz Capitulino return true; 1061944d9fecSLuiz Capitulino } 1062944d9fecSLuiz Capitulino 1063944d9fecSLuiz Capitulino static bool zone_spans_last_pfn(const struct zone *zone, 1064944d9fecSLuiz Capitulino unsigned long start_pfn, unsigned long nr_pages) 1065944d9fecSLuiz Capitulino { 1066944d9fecSLuiz Capitulino unsigned long last_pfn = start_pfn + nr_pages - 1; 1067944d9fecSLuiz Capitulino return zone_spans_pfn(zone, last_pfn); 1068944d9fecSLuiz Capitulino } 1069944d9fecSLuiz Capitulino 1070d00181b9SKirill A. Shutemov static struct page *alloc_gigantic_page(int nid, unsigned int order) 1071944d9fecSLuiz Capitulino { 1072944d9fecSLuiz Capitulino unsigned long nr_pages = 1 << order; 1073944d9fecSLuiz Capitulino unsigned long ret, pfn, flags; 1074944d9fecSLuiz Capitulino struct zone *z; 1075944d9fecSLuiz Capitulino 1076944d9fecSLuiz Capitulino z = NODE_DATA(nid)->node_zones; 1077944d9fecSLuiz Capitulino for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) { 1078944d9fecSLuiz Capitulino spin_lock_irqsave(&z->lock, flags); 1079944d9fecSLuiz Capitulino 1080944d9fecSLuiz Capitulino pfn = ALIGN(z->zone_start_pfn, nr_pages); 1081944d9fecSLuiz Capitulino while (zone_spans_last_pfn(z, pfn, nr_pages)) { 1082f44b2ddaSJoonsoo Kim if (pfn_range_valid_gigantic(z, pfn, nr_pages)) { 1083944d9fecSLuiz Capitulino /* 1084944d9fecSLuiz Capitulino * We release the zone lock here because 1085944d9fecSLuiz Capitulino * alloc_contig_range() will also lock the zone 1086944d9fecSLuiz Capitulino * at some point. If there's an allocation 1087944d9fecSLuiz Capitulino * spinning on this lock, it may win the race 1088944d9fecSLuiz Capitulino * and cause alloc_contig_range() to fail... 1089944d9fecSLuiz Capitulino */ 1090944d9fecSLuiz Capitulino spin_unlock_irqrestore(&z->lock, flags); 1091944d9fecSLuiz Capitulino ret = __alloc_gigantic_page(pfn, nr_pages); 1092944d9fecSLuiz Capitulino if (!ret) 1093944d9fecSLuiz Capitulino return pfn_to_page(pfn); 1094944d9fecSLuiz Capitulino spin_lock_irqsave(&z->lock, flags); 1095944d9fecSLuiz Capitulino } 1096944d9fecSLuiz Capitulino pfn += nr_pages; 1097944d9fecSLuiz Capitulino } 1098944d9fecSLuiz Capitulino 1099944d9fecSLuiz Capitulino spin_unlock_irqrestore(&z->lock, flags); 1100944d9fecSLuiz Capitulino } 1101944d9fecSLuiz Capitulino 1102944d9fecSLuiz Capitulino return NULL; 1103944d9fecSLuiz Capitulino } 1104944d9fecSLuiz Capitulino 1105944d9fecSLuiz Capitulino static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); 1106d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order); 1107944d9fecSLuiz Capitulino 1108944d9fecSLuiz Capitulino static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid) 1109944d9fecSLuiz Capitulino { 1110944d9fecSLuiz Capitulino struct page *page; 1111944d9fecSLuiz Capitulino 1112944d9fecSLuiz Capitulino page = alloc_gigantic_page(nid, huge_page_order(h)); 1113944d9fecSLuiz Capitulino if (page) { 1114944d9fecSLuiz Capitulino prep_compound_gigantic_page(page, huge_page_order(h)); 1115944d9fecSLuiz Capitulino prep_new_huge_page(h, page, nid); 1116944d9fecSLuiz Capitulino } 1117944d9fecSLuiz Capitulino 1118944d9fecSLuiz Capitulino return page; 1119944d9fecSLuiz Capitulino } 1120944d9fecSLuiz Capitulino 1121944d9fecSLuiz Capitulino static int alloc_fresh_gigantic_page(struct hstate *h, 1122944d9fecSLuiz Capitulino nodemask_t *nodes_allowed) 1123944d9fecSLuiz Capitulino { 1124944d9fecSLuiz Capitulino struct page *page = NULL; 1125944d9fecSLuiz Capitulino int nr_nodes, node; 1126944d9fecSLuiz Capitulino 1127944d9fecSLuiz Capitulino for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1128944d9fecSLuiz Capitulino page = alloc_fresh_gigantic_page_node(h, node); 1129944d9fecSLuiz Capitulino if (page) 1130944d9fecSLuiz Capitulino return 1; 1131944d9fecSLuiz Capitulino } 1132944d9fecSLuiz Capitulino 1133944d9fecSLuiz Capitulino return 0; 1134944d9fecSLuiz Capitulino } 1135944d9fecSLuiz Capitulino 1136944d9fecSLuiz Capitulino static inline bool gigantic_page_supported(void) { return true; } 1137944d9fecSLuiz Capitulino #else 1138944d9fecSLuiz Capitulino static inline bool gigantic_page_supported(void) { return false; } 1139d00181b9SKirill A. Shutemov static inline void free_gigantic_page(struct page *page, unsigned int order) { } 1140944d9fecSLuiz Capitulino static inline void destroy_compound_gigantic_page(struct page *page, 1141d00181b9SKirill A. Shutemov unsigned int order) { } 1142944d9fecSLuiz Capitulino static inline int alloc_fresh_gigantic_page(struct hstate *h, 1143944d9fecSLuiz Capitulino nodemask_t *nodes_allowed) { return 0; } 1144944d9fecSLuiz Capitulino #endif 1145944d9fecSLuiz Capitulino 1146a5516438SAndi Kleen static void update_and_free_page(struct hstate *h, struct page *page) 11476af2acb6SAdam Litke { 11486af2acb6SAdam Litke int i; 1149a5516438SAndi Kleen 1150944d9fecSLuiz Capitulino if (hstate_is_gigantic(h) && !gigantic_page_supported()) 1151944d9fecSLuiz Capitulino return; 115218229df5SAndy Whitcroft 1153a5516438SAndi Kleen h->nr_huge_pages--; 1154a5516438SAndi Kleen h->nr_huge_pages_node[page_to_nid(page)]--; 1155a5516438SAndi Kleen for (i = 0; i < pages_per_huge_page(h); i++) { 115632f84528SChris Forbes page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 115732f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1158a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1159a7407a27SLuiz Capitulino 1 << PG_writeback); 11606af2acb6SAdam Litke } 1161309381feSSasha Levin VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 1162f1e61557SKirill A. Shutemov set_compound_page_dtor(page, NULL_COMPOUND_DTOR); 11636af2acb6SAdam Litke set_page_refcounted(page); 1164944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) { 1165944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 1166944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 1167944d9fecSLuiz Capitulino } else { 1168a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 11696af2acb6SAdam Litke } 1170944d9fecSLuiz Capitulino } 11716af2acb6SAdam Litke 1172e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1173e5ff2159SAndi Kleen { 1174e5ff2159SAndi Kleen struct hstate *h; 1175e5ff2159SAndi Kleen 1176e5ff2159SAndi Kleen for_each_hstate(h) { 1177e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1178e5ff2159SAndi Kleen return h; 1179e5ff2159SAndi Kleen } 1180e5ff2159SAndi Kleen return NULL; 1181e5ff2159SAndi Kleen } 1182e5ff2159SAndi Kleen 1183bcc54222SNaoya Horiguchi /* 1184bcc54222SNaoya Horiguchi * Test to determine whether the hugepage is "active/in-use" (i.e. being linked 1185bcc54222SNaoya Horiguchi * to hstate->hugepage_activelist.) 1186bcc54222SNaoya Horiguchi * 1187bcc54222SNaoya Horiguchi * This function can be called for tail pages, but never returns true for them. 1188bcc54222SNaoya Horiguchi */ 1189bcc54222SNaoya Horiguchi bool page_huge_active(struct page *page) 1190bcc54222SNaoya Horiguchi { 1191bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHuge(page), page); 1192bcc54222SNaoya Horiguchi return PageHead(page) && PagePrivate(&page[1]); 1193bcc54222SNaoya Horiguchi } 1194bcc54222SNaoya Horiguchi 1195bcc54222SNaoya Horiguchi /* never called for tail page */ 1196bcc54222SNaoya Horiguchi static void set_page_huge_active(struct page *page) 1197bcc54222SNaoya Horiguchi { 1198bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 1199bcc54222SNaoya Horiguchi SetPagePrivate(&page[1]); 1200bcc54222SNaoya Horiguchi } 1201bcc54222SNaoya Horiguchi 1202bcc54222SNaoya Horiguchi static void clear_page_huge_active(struct page *page) 1203bcc54222SNaoya Horiguchi { 1204bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 1205bcc54222SNaoya Horiguchi ClearPagePrivate(&page[1]); 1206bcc54222SNaoya Horiguchi } 1207bcc54222SNaoya Horiguchi 12088f1d26d0SAtsushi Kumagai void free_huge_page(struct page *page) 120927a85ef1SDavid Gibson { 1210a5516438SAndi Kleen /* 1211a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1212a5516438SAndi Kleen * compound page destructor. 1213a5516438SAndi Kleen */ 1214e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 12157893d1d5SAdam Litke int nid = page_to_nid(page); 121690481622SDavid Gibson struct hugepage_subpool *spool = 121790481622SDavid Gibson (struct hugepage_subpool *)page_private(page); 121807443a85SJoonsoo Kim bool restore_reserve; 121927a85ef1SDavid Gibson 1220e5df70abSAndy Whitcroft set_page_private(page, 0); 122123be7468SMel Gorman page->mapping = NULL; 1222b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1223b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_mapcount(page), page); 122407443a85SJoonsoo Kim restore_reserve = PagePrivate(page); 122516c794b4SJoonsoo Kim ClearPagePrivate(page); 122627a85ef1SDavid Gibson 12271c5ecae3SMike Kravetz /* 12281c5ecae3SMike Kravetz * A return code of zero implies that the subpool will be under its 12291c5ecae3SMike Kravetz * minimum size if the reservation is not restored after page is free. 12301c5ecae3SMike Kravetz * Therefore, force restore_reserve operation. 12311c5ecae3SMike Kravetz */ 12321c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 12331c5ecae3SMike Kravetz restore_reserve = true; 12341c5ecae3SMike Kravetz 123527a85ef1SDavid Gibson spin_lock(&hugetlb_lock); 1236bcc54222SNaoya Horiguchi clear_page_huge_active(page); 12376d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 12386d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 123907443a85SJoonsoo Kim if (restore_reserve) 124007443a85SJoonsoo Kim h->resv_huge_pages++; 124107443a85SJoonsoo Kim 1242944d9fecSLuiz Capitulino if (h->surplus_huge_pages_node[nid]) { 12430edaecfaSAneesh Kumar K.V /* remove the page from active list */ 12440edaecfaSAneesh Kumar K.V list_del(&page->lru); 1245a5516438SAndi Kleen update_and_free_page(h, page); 1246a5516438SAndi Kleen h->surplus_huge_pages--; 1247a5516438SAndi Kleen h->surplus_huge_pages_node[nid]--; 12487893d1d5SAdam Litke } else { 12495d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1250a5516438SAndi Kleen enqueue_huge_page(h, page); 12517893d1d5SAdam Litke } 125227a85ef1SDavid Gibson spin_unlock(&hugetlb_lock); 125327a85ef1SDavid Gibson } 125427a85ef1SDavid Gibson 1255a5516438SAndi Kleen static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1256b7ba30c6SAndi Kleen { 12570edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1258f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1259b7ba30c6SAndi Kleen spin_lock(&hugetlb_lock); 12609dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 1261a5516438SAndi Kleen h->nr_huge_pages++; 1262a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 1263b7ba30c6SAndi Kleen spin_unlock(&hugetlb_lock); 1264b7ba30c6SAndi Kleen put_page(page); /* free it into the hugepage allocator */ 1265b7ba30c6SAndi Kleen } 1266b7ba30c6SAndi Kleen 1267d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order) 126820a0307cSWu Fengguang { 126920a0307cSWu Fengguang int i; 127020a0307cSWu Fengguang int nr_pages = 1 << order; 127120a0307cSWu Fengguang struct page *p = page + 1; 127220a0307cSWu Fengguang 127320a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 127420a0307cSWu Fengguang set_compound_order(page, order); 1275ef5a22beSAndrea Arcangeli __ClearPageReserved(page); 1276de09d31dSKirill A. Shutemov __SetPageHead(page); 127720a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1278ef5a22beSAndrea Arcangeli /* 1279ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1280ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1281ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 1282ef5a22beSAndrea Arcangeli * too. Otherwse drivers using get_user_pages() to access tail 1283ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1284ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1285ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1286ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1287ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1288ef5a22beSAndrea Arcangeli * after get_user_pages(). 1289ef5a22beSAndrea Arcangeli */ 1290ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 129158a84aa9SYouquan Song set_page_count(p, 0); 12921d798ca3SKirill A. Shutemov set_compound_head(p, page); 129320a0307cSWu Fengguang } 1294b4330afbSMike Kravetz atomic_set(compound_mapcount_ptr(page), -1); 129520a0307cSWu Fengguang } 129620a0307cSWu Fengguang 12977795912cSAndrew Morton /* 12987795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 12997795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 13007795912cSAndrew Morton * details. 13017795912cSAndrew Morton */ 130220a0307cSWu Fengguang int PageHuge(struct page *page) 130320a0307cSWu Fengguang { 130420a0307cSWu Fengguang if (!PageCompound(page)) 130520a0307cSWu Fengguang return 0; 130620a0307cSWu Fengguang 130720a0307cSWu Fengguang page = compound_head(page); 1308f1e61557SKirill A. Shutemov return page[1].compound_dtor == HUGETLB_PAGE_DTOR; 130920a0307cSWu Fengguang } 131043131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 131143131e14SNaoya Horiguchi 131227c73ae7SAndrea Arcangeli /* 131327c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 131427c73ae7SAndrea Arcangeli * normal or transparent huge pages. 131527c73ae7SAndrea Arcangeli */ 131627c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 131727c73ae7SAndrea Arcangeli { 131827c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 131927c73ae7SAndrea Arcangeli return 0; 132027c73ae7SAndrea Arcangeli 1321758f66a2SAndrew Morton return get_compound_page_dtor(page_head) == free_huge_page; 132227c73ae7SAndrea Arcangeli } 132327c73ae7SAndrea Arcangeli 132413d60f4bSZhang Yi pgoff_t __basepage_index(struct page *page) 132513d60f4bSZhang Yi { 132613d60f4bSZhang Yi struct page *page_head = compound_head(page); 132713d60f4bSZhang Yi pgoff_t index = page_index(page_head); 132813d60f4bSZhang Yi unsigned long compound_idx; 132913d60f4bSZhang Yi 133013d60f4bSZhang Yi if (!PageHuge(page_head)) 133113d60f4bSZhang Yi return page_index(page); 133213d60f4bSZhang Yi 133313d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 133413d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 133513d60f4bSZhang Yi else 133613d60f4bSZhang Yi compound_idx = page - page_head; 133713d60f4bSZhang Yi 133813d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 133913d60f4bSZhang Yi } 134013d60f4bSZhang Yi 1341a5516438SAndi Kleen static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) 13421da177e4SLinus Torvalds { 13431da177e4SLinus Torvalds struct page *page; 1344f96efd58SJoe Jin 134596db800fSVlastimil Babka page = __alloc_pages_node(nid, 134686cdb465SNaoya Horiguchi htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| 1347551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 1348a5516438SAndi Kleen huge_page_order(h)); 13491da177e4SLinus Torvalds if (page) { 1350a5516438SAndi Kleen prep_new_huge_page(h, page, nid); 13511da177e4SLinus Torvalds } 135263b4613cSNishanth Aravamudan 135363b4613cSNishanth Aravamudan return page; 135463b4613cSNishanth Aravamudan } 135563b4613cSNishanth Aravamudan 1356b2261026SJoonsoo Kim static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 1357b2261026SJoonsoo Kim { 1358b2261026SJoonsoo Kim struct page *page; 1359b2261026SJoonsoo Kim int nr_nodes, node; 1360b2261026SJoonsoo Kim int ret = 0; 1361b2261026SJoonsoo Kim 1362b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1363b2261026SJoonsoo Kim page = alloc_fresh_huge_page_node(h, node); 1364b2261026SJoonsoo Kim if (page) { 1365b2261026SJoonsoo Kim ret = 1; 1366b2261026SJoonsoo Kim break; 1367b2261026SJoonsoo Kim } 1368b2261026SJoonsoo Kim } 1369b2261026SJoonsoo Kim 1370b2261026SJoonsoo Kim if (ret) 1371b2261026SJoonsoo Kim count_vm_event(HTLB_BUDDY_PGALLOC); 1372b2261026SJoonsoo Kim else 1373b2261026SJoonsoo Kim count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 1374b2261026SJoonsoo Kim 1375b2261026SJoonsoo Kim return ret; 1376b2261026SJoonsoo Kim } 1377b2261026SJoonsoo Kim 1378e8c5c824SLee Schermerhorn /* 1379e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 1380e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 1381e8c5c824SLee Schermerhorn * balanced over allowed nodes. 1382e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 1383e8c5c824SLee Schermerhorn */ 13846ae11b27SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 13856ae11b27SLee Schermerhorn bool acct_surplus) 1386e8c5c824SLee Schermerhorn { 1387b2261026SJoonsoo Kim int nr_nodes, node; 1388e8c5c824SLee Schermerhorn int ret = 0; 1389e8c5c824SLee Schermerhorn 1390b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1391685f3457SLee Schermerhorn /* 1392685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 1393685f3457SLee Schermerhorn * nodes with surplus pages. 1394685f3457SLee Schermerhorn */ 1395b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 1396b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 1397e8c5c824SLee Schermerhorn struct page *page = 1398b2261026SJoonsoo Kim list_entry(h->hugepage_freelists[node].next, 1399e8c5c824SLee Schermerhorn struct page, lru); 1400e8c5c824SLee Schermerhorn list_del(&page->lru); 1401e8c5c824SLee Schermerhorn h->free_huge_pages--; 1402b2261026SJoonsoo Kim h->free_huge_pages_node[node]--; 1403685f3457SLee Schermerhorn if (acct_surplus) { 1404685f3457SLee Schermerhorn h->surplus_huge_pages--; 1405b2261026SJoonsoo Kim h->surplus_huge_pages_node[node]--; 1406685f3457SLee Schermerhorn } 1407e8c5c824SLee Schermerhorn update_and_free_page(h, page); 1408e8c5c824SLee Schermerhorn ret = 1; 14099a76db09SLee Schermerhorn break; 1410e8c5c824SLee Schermerhorn } 1411b2261026SJoonsoo Kim } 1412e8c5c824SLee Schermerhorn 1413e8c5c824SLee Schermerhorn return ret; 1414e8c5c824SLee Schermerhorn } 1415e8c5c824SLee Schermerhorn 1416c8721bbbSNaoya Horiguchi /* 1417c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 1418c8721bbbSNaoya Horiguchi * nothing for in-use (including surplus) hugepages. 1419c8721bbbSNaoya Horiguchi */ 1420c8721bbbSNaoya Horiguchi static void dissolve_free_huge_page(struct page *page) 1421c8721bbbSNaoya Horiguchi { 1422c8721bbbSNaoya Horiguchi spin_lock(&hugetlb_lock); 1423c8721bbbSNaoya Horiguchi if (PageHuge(page) && !page_count(page)) { 1424c8721bbbSNaoya Horiguchi struct hstate *h = page_hstate(page); 1425c8721bbbSNaoya Horiguchi int nid = page_to_nid(page); 1426c8721bbbSNaoya Horiguchi list_del(&page->lru); 1427c8721bbbSNaoya Horiguchi h->free_huge_pages--; 1428c8721bbbSNaoya Horiguchi h->free_huge_pages_node[nid]--; 1429c8721bbbSNaoya Horiguchi update_and_free_page(h, page); 1430c8721bbbSNaoya Horiguchi } 1431c8721bbbSNaoya Horiguchi spin_unlock(&hugetlb_lock); 1432c8721bbbSNaoya Horiguchi } 1433c8721bbbSNaoya Horiguchi 1434c8721bbbSNaoya Horiguchi /* 1435c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 1436c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 1437c8721bbbSNaoya Horiguchi * Note that start_pfn should aligned with (minimum) hugepage size. 1438c8721bbbSNaoya Horiguchi */ 1439c8721bbbSNaoya Horiguchi void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 1440c8721bbbSNaoya Horiguchi { 1441c8721bbbSNaoya Horiguchi unsigned long pfn; 1442c8721bbbSNaoya Horiguchi 1443d0177639SLi Zhong if (!hugepages_supported()) 1444d0177639SLi Zhong return; 1445d0177639SLi Zhong 1446641844f5SNaoya Horiguchi VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << minimum_order)); 1447641844f5SNaoya Horiguchi for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) 1448c8721bbbSNaoya Horiguchi dissolve_free_huge_page(pfn_to_page(pfn)); 1449c8721bbbSNaoya Horiguchi } 1450c8721bbbSNaoya Horiguchi 1451099730d6SDave Hansen /* 1452099730d6SDave Hansen * There are 3 ways this can get called: 1453099730d6SDave Hansen * 1. With vma+addr: we use the VMA's memory policy 1454099730d6SDave Hansen * 2. With !vma, but nid=NUMA_NO_NODE: We try to allocate a huge 1455099730d6SDave Hansen * page from any node, and let the buddy allocator itself figure 1456099730d6SDave Hansen * it out. 1457099730d6SDave Hansen * 3. With !vma, but nid!=NUMA_NO_NODE. We allocate a huge page 1458099730d6SDave Hansen * strictly from 'nid' 1459099730d6SDave Hansen */ 1460099730d6SDave Hansen static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h, 1461099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr, int nid) 1462099730d6SDave Hansen { 1463099730d6SDave Hansen int order = huge_page_order(h); 1464099730d6SDave Hansen gfp_t gfp = htlb_alloc_mask(h)|__GFP_COMP|__GFP_REPEAT|__GFP_NOWARN; 1465099730d6SDave Hansen unsigned int cpuset_mems_cookie; 1466099730d6SDave Hansen 1467099730d6SDave Hansen /* 1468099730d6SDave Hansen * We need a VMA to get a memory policy. If we do not 1469e0ec90eeSDave Hansen * have one, we use the 'nid' argument. 1470e0ec90eeSDave Hansen * 1471e0ec90eeSDave Hansen * The mempolicy stuff below has some non-inlined bits 1472e0ec90eeSDave Hansen * and calls ->vm_ops. That makes it hard to optimize at 1473e0ec90eeSDave Hansen * compile-time, even when NUMA is off and it does 1474e0ec90eeSDave Hansen * nothing. This helps the compiler optimize it out. 1475099730d6SDave Hansen */ 1476e0ec90eeSDave Hansen if (!IS_ENABLED(CONFIG_NUMA) || !vma) { 1477099730d6SDave Hansen /* 1478099730d6SDave Hansen * If a specific node is requested, make sure to 1479099730d6SDave Hansen * get memory from there, but only when a node 1480099730d6SDave Hansen * is explicitly specified. 1481099730d6SDave Hansen */ 1482099730d6SDave Hansen if (nid != NUMA_NO_NODE) 1483099730d6SDave Hansen gfp |= __GFP_THISNODE; 1484099730d6SDave Hansen /* 1485099730d6SDave Hansen * Make sure to call something that can handle 1486099730d6SDave Hansen * nid=NUMA_NO_NODE 1487099730d6SDave Hansen */ 1488099730d6SDave Hansen return alloc_pages_node(nid, gfp, order); 1489099730d6SDave Hansen } 1490099730d6SDave Hansen 1491099730d6SDave Hansen /* 1492099730d6SDave Hansen * OK, so we have a VMA. Fetch the mempolicy and try to 1493e0ec90eeSDave Hansen * allocate a huge page with it. We will only reach this 1494e0ec90eeSDave Hansen * when CONFIG_NUMA=y. 1495099730d6SDave Hansen */ 1496099730d6SDave Hansen do { 1497099730d6SDave Hansen struct page *page; 1498099730d6SDave Hansen struct mempolicy *mpol; 1499099730d6SDave Hansen struct zonelist *zl; 1500099730d6SDave Hansen nodemask_t *nodemask; 1501099730d6SDave Hansen 1502099730d6SDave Hansen cpuset_mems_cookie = read_mems_allowed_begin(); 1503099730d6SDave Hansen zl = huge_zonelist(vma, addr, gfp, &mpol, &nodemask); 1504099730d6SDave Hansen mpol_cond_put(mpol); 1505099730d6SDave Hansen page = __alloc_pages_nodemask(gfp, order, zl, nodemask); 1506099730d6SDave Hansen if (page) 1507099730d6SDave Hansen return page; 1508099730d6SDave Hansen } while (read_mems_allowed_retry(cpuset_mems_cookie)); 1509099730d6SDave Hansen 1510099730d6SDave Hansen return NULL; 1511099730d6SDave Hansen } 1512099730d6SDave Hansen 1513099730d6SDave Hansen /* 1514099730d6SDave Hansen * There are two ways to allocate a huge page: 1515099730d6SDave Hansen * 1. When you have a VMA and an address (like a fault) 1516099730d6SDave Hansen * 2. When you have no VMA (like when setting /proc/.../nr_hugepages) 1517099730d6SDave Hansen * 1518099730d6SDave Hansen * 'vma' and 'addr' are only for (1). 'nid' is always NUMA_NO_NODE in 1519099730d6SDave Hansen * this case which signifies that the allocation should be done with 1520099730d6SDave Hansen * respect for the VMA's memory policy. 1521099730d6SDave Hansen * 1522099730d6SDave Hansen * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This 1523099730d6SDave Hansen * implies that memory policies will not be taken in to account. 1524099730d6SDave Hansen */ 1525099730d6SDave Hansen static struct page *__alloc_buddy_huge_page(struct hstate *h, 1526099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr, int nid) 15277893d1d5SAdam Litke { 15287893d1d5SAdam Litke struct page *page; 1529bf50bab2SNaoya Horiguchi unsigned int r_nid; 15307893d1d5SAdam Litke 1531bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1532aa888a74SAndi Kleen return NULL; 1533aa888a74SAndi Kleen 1534d1c3fb1fSNishanth Aravamudan /* 1535099730d6SDave Hansen * Make sure that anyone specifying 'nid' is not also specifying a VMA. 1536099730d6SDave Hansen * This makes sure the caller is picking _one_ of the modes with which 1537099730d6SDave Hansen * we can call this function, not both. 1538099730d6SDave Hansen */ 1539099730d6SDave Hansen if (vma || (addr != -1)) { 1540e0ec90eeSDave Hansen VM_WARN_ON_ONCE(addr == -1); 1541e0ec90eeSDave Hansen VM_WARN_ON_ONCE(nid != NUMA_NO_NODE); 1542099730d6SDave Hansen } 1543099730d6SDave Hansen /* 1544d1c3fb1fSNishanth Aravamudan * Assume we will successfully allocate the surplus page to 1545d1c3fb1fSNishanth Aravamudan * prevent racing processes from causing the surplus to exceed 1546d1c3fb1fSNishanth Aravamudan * overcommit 1547d1c3fb1fSNishanth Aravamudan * 1548d1c3fb1fSNishanth Aravamudan * This however introduces a different race, where a process B 1549d1c3fb1fSNishanth Aravamudan * tries to grow the static hugepage pool while alloc_pages() is 1550d1c3fb1fSNishanth Aravamudan * called by process A. B will only examine the per-node 1551d1c3fb1fSNishanth Aravamudan * counters in determining if surplus huge pages can be 1552d1c3fb1fSNishanth Aravamudan * converted to normal huge pages in adjust_pool_surplus(). A 1553d1c3fb1fSNishanth Aravamudan * won't be able to increment the per-node counter, until the 1554d1c3fb1fSNishanth Aravamudan * lock is dropped by B, but B doesn't drop hugetlb_lock until 1555d1c3fb1fSNishanth Aravamudan * no more huge pages can be converted from surplus to normal 1556d1c3fb1fSNishanth Aravamudan * state (and doesn't try to convert again). Thus, we have a 1557d1c3fb1fSNishanth Aravamudan * case where a surplus huge page exists, the pool is grown, and 1558d1c3fb1fSNishanth Aravamudan * the surplus huge page still exists after, even though it 1559d1c3fb1fSNishanth Aravamudan * should just have been converted to a normal huge page. This 1560d1c3fb1fSNishanth Aravamudan * does not leak memory, though, as the hugepage will be freed 1561d1c3fb1fSNishanth Aravamudan * once it is out of use. It also does not allow the counters to 1562d1c3fb1fSNishanth Aravamudan * go out of whack in adjust_pool_surplus() as we don't modify 1563d1c3fb1fSNishanth Aravamudan * the node values until we've gotten the hugepage and only the 1564d1c3fb1fSNishanth Aravamudan * per-node value is checked there. 1565d1c3fb1fSNishanth Aravamudan */ 1566d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 1567a5516438SAndi Kleen if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 1568d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1569d1c3fb1fSNishanth Aravamudan return NULL; 1570d1c3fb1fSNishanth Aravamudan } else { 1571a5516438SAndi Kleen h->nr_huge_pages++; 1572a5516438SAndi Kleen h->surplus_huge_pages++; 1573d1c3fb1fSNishanth Aravamudan } 1574d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1575d1c3fb1fSNishanth Aravamudan 1576099730d6SDave Hansen page = __hugetlb_alloc_buddy_huge_page(h, vma, addr, nid); 1577d1c3fb1fSNishanth Aravamudan 15787893d1d5SAdam Litke spin_lock(&hugetlb_lock); 1579d1c3fb1fSNishanth Aravamudan if (page) { 15800edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1581bf50bab2SNaoya Horiguchi r_nid = page_to_nid(page); 1582f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 15839dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 1584d1c3fb1fSNishanth Aravamudan /* 1585d1c3fb1fSNishanth Aravamudan * We incremented the global counters already 1586d1c3fb1fSNishanth Aravamudan */ 1587bf50bab2SNaoya Horiguchi h->nr_huge_pages_node[r_nid]++; 1588bf50bab2SNaoya Horiguchi h->surplus_huge_pages_node[r_nid]++; 15893b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC); 1590d1c3fb1fSNishanth Aravamudan } else { 1591a5516438SAndi Kleen h->nr_huge_pages--; 1592a5516438SAndi Kleen h->surplus_huge_pages--; 15933b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 15947893d1d5SAdam Litke } 1595d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 15967893d1d5SAdam Litke 15977893d1d5SAdam Litke return page; 15987893d1d5SAdam Litke } 15997893d1d5SAdam Litke 1600e4e574b7SAdam Litke /* 1601099730d6SDave Hansen * Allocate a huge page from 'nid'. Note, 'nid' may be 1602099730d6SDave Hansen * NUMA_NO_NODE, which means that it may be allocated 1603099730d6SDave Hansen * anywhere. 1604099730d6SDave Hansen */ 1605e0ec90eeSDave Hansen static 1606099730d6SDave Hansen struct page *__alloc_buddy_huge_page_no_mpol(struct hstate *h, int nid) 1607099730d6SDave Hansen { 1608099730d6SDave Hansen unsigned long addr = -1; 1609099730d6SDave Hansen 1610099730d6SDave Hansen return __alloc_buddy_huge_page(h, NULL, addr, nid); 1611099730d6SDave Hansen } 1612099730d6SDave Hansen 1613099730d6SDave Hansen /* 1614099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 1615099730d6SDave Hansen */ 1616e0ec90eeSDave Hansen static 1617099730d6SDave Hansen struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h, 1618099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 1619099730d6SDave Hansen { 1620099730d6SDave Hansen return __alloc_buddy_huge_page(h, vma, addr, NUMA_NO_NODE); 1621099730d6SDave Hansen } 1622099730d6SDave Hansen 1623099730d6SDave Hansen /* 1624bf50bab2SNaoya Horiguchi * This allocation function is useful in the context where vma is irrelevant. 1625bf50bab2SNaoya Horiguchi * E.g. soft-offlining uses this function because it only cares physical 1626bf50bab2SNaoya Horiguchi * address of error page. 1627bf50bab2SNaoya Horiguchi */ 1628bf50bab2SNaoya Horiguchi struct page *alloc_huge_page_node(struct hstate *h, int nid) 1629bf50bab2SNaoya Horiguchi { 16304ef91848SJoonsoo Kim struct page *page = NULL; 1631bf50bab2SNaoya Horiguchi 1632bf50bab2SNaoya Horiguchi spin_lock(&hugetlb_lock); 16334ef91848SJoonsoo Kim if (h->free_huge_pages - h->resv_huge_pages > 0) 1634bf50bab2SNaoya Horiguchi page = dequeue_huge_page_node(h, nid); 1635bf50bab2SNaoya Horiguchi spin_unlock(&hugetlb_lock); 1636bf50bab2SNaoya Horiguchi 163794ae8ba7SAneesh Kumar K.V if (!page) 1638099730d6SDave Hansen page = __alloc_buddy_huge_page_no_mpol(h, nid); 1639bf50bab2SNaoya Horiguchi 1640bf50bab2SNaoya Horiguchi return page; 1641bf50bab2SNaoya Horiguchi } 1642bf50bab2SNaoya Horiguchi 1643bf50bab2SNaoya Horiguchi /* 164425985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 1645e4e574b7SAdam Litke * of size 'delta'. 1646e4e574b7SAdam Litke */ 1647a5516438SAndi Kleen static int gather_surplus_pages(struct hstate *h, int delta) 1648e4e574b7SAdam Litke { 1649e4e574b7SAdam Litke struct list_head surplus_list; 1650e4e574b7SAdam Litke struct page *page, *tmp; 1651e4e574b7SAdam Litke int ret, i; 1652e4e574b7SAdam Litke int needed, allocated; 165328073b02SHillf Danton bool alloc_ok = true; 1654e4e574b7SAdam Litke 1655a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 1656ac09b3a1SAdam Litke if (needed <= 0) { 1657a5516438SAndi Kleen h->resv_huge_pages += delta; 1658e4e574b7SAdam Litke return 0; 1659ac09b3a1SAdam Litke } 1660e4e574b7SAdam Litke 1661e4e574b7SAdam Litke allocated = 0; 1662e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 1663e4e574b7SAdam Litke 1664e4e574b7SAdam Litke ret = -ENOMEM; 1665e4e574b7SAdam Litke retry: 1666e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 1667e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 1668099730d6SDave Hansen page = __alloc_buddy_huge_page_no_mpol(h, NUMA_NO_NODE); 166928073b02SHillf Danton if (!page) { 167028073b02SHillf Danton alloc_ok = false; 167128073b02SHillf Danton break; 167228073b02SHillf Danton } 1673e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 1674e4e574b7SAdam Litke } 167528073b02SHillf Danton allocated += i; 1676e4e574b7SAdam Litke 1677e4e574b7SAdam Litke /* 1678e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 1679e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 1680e4e574b7SAdam Litke */ 1681e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 1682a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 1683a5516438SAndi Kleen (h->free_huge_pages + allocated); 168428073b02SHillf Danton if (needed > 0) { 168528073b02SHillf Danton if (alloc_ok) 1686e4e574b7SAdam Litke goto retry; 168728073b02SHillf Danton /* 168828073b02SHillf Danton * We were not able to allocate enough pages to 168928073b02SHillf Danton * satisfy the entire reservation so we free what 169028073b02SHillf Danton * we've allocated so far. 169128073b02SHillf Danton */ 169228073b02SHillf Danton goto free; 169328073b02SHillf Danton } 1694e4e574b7SAdam Litke /* 1695e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 169625985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 1697e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 1698ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 1699ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 1700ac09b3a1SAdam Litke * before they are reserved. 1701e4e574b7SAdam Litke */ 1702e4e574b7SAdam Litke needed += allocated; 1703a5516438SAndi Kleen h->resv_huge_pages += delta; 1704e4e574b7SAdam Litke ret = 0; 1705a9869b83SNaoya Horiguchi 170619fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 170719fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 170819fc3f0aSAdam Litke if ((--needed) < 0) 170919fc3f0aSAdam Litke break; 1710a9869b83SNaoya Horiguchi /* 1711a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 1712a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 1713a9869b83SNaoya Horiguchi */ 1714a9869b83SNaoya Horiguchi put_page_testzero(page); 1715309381feSSasha Levin VM_BUG_ON_PAGE(page_count(page), page); 1716a5516438SAndi Kleen enqueue_huge_page(h, page); 171719fc3f0aSAdam Litke } 171828073b02SHillf Danton free: 1719b0365c8dSHillf Danton spin_unlock(&hugetlb_lock); 172019fc3f0aSAdam Litke 172119fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 1722c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 1723a9869b83SNaoya Horiguchi put_page(page); 172419fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 1725e4e574b7SAdam Litke 1726e4e574b7SAdam Litke return ret; 1727e4e574b7SAdam Litke } 1728e4e574b7SAdam Litke 1729e4e574b7SAdam Litke /* 1730e4e574b7SAdam Litke * When releasing a hugetlb pool reservation, any surplus pages that were 1731e4e574b7SAdam Litke * allocated to satisfy the reservation must be explicitly freed if they were 1732e4e574b7SAdam Litke * never used. 1733685f3457SLee Schermerhorn * Called with hugetlb_lock held. 1734e4e574b7SAdam Litke */ 1735a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 1736a5516438SAndi Kleen unsigned long unused_resv_pages) 1737e4e574b7SAdam Litke { 1738e4e574b7SAdam Litke unsigned long nr_pages; 1739e4e574b7SAdam Litke 1740ac09b3a1SAdam Litke /* Uncommit the reservation */ 1741a5516438SAndi Kleen h->resv_huge_pages -= unused_resv_pages; 1742ac09b3a1SAdam Litke 1743aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 1744bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1745aa888a74SAndi Kleen return; 1746aa888a74SAndi Kleen 1747a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 1748e4e574b7SAdam Litke 1749685f3457SLee Schermerhorn /* 1750685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 17519b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 17529b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 17539b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 17549b5e5d0fSLee Schermerhorn * free_pool_huge_page() will balance the the freed pages across the 17559b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 1756685f3457SLee Schermerhorn */ 1757685f3457SLee Schermerhorn while (nr_pages--) { 17588cebfcd0SLai Jiangshan if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) 1759685f3457SLee Schermerhorn break; 17607848a4bfSMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 1761e4e574b7SAdam Litke } 1762e4e574b7SAdam Litke } 1763e4e574b7SAdam Litke 17645e911373SMike Kravetz 1765c37f9fb1SAndy Whitcroft /* 1766feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 17675e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 1768cf3ad20bSMike Kravetz * 1769cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 1770cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 1771cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 1772cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 1773cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 1774feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 1775feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 1776feba16e2SMike Kravetz * is called in such cases. 1777cf3ad20bSMike Kravetz * 1778cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 1779cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 1780cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 1781cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 1782cf3ad20bSMike Kravetz * take appropriate action. 1783c37f9fb1SAndy Whitcroft */ 17845e911373SMike Kravetz enum vma_resv_mode { 17855e911373SMike Kravetz VMA_NEEDS_RESV, 17865e911373SMike Kravetz VMA_COMMIT_RESV, 1787feba16e2SMike Kravetz VMA_END_RESV, 17885e911373SMike Kravetz }; 1789cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 1790cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 17915e911373SMike Kravetz enum vma_resv_mode mode) 1792c37f9fb1SAndy Whitcroft { 17934e35f483SJoonsoo Kim struct resv_map *resv; 17944e35f483SJoonsoo Kim pgoff_t idx; 1795cf3ad20bSMike Kravetz long ret; 1796c37f9fb1SAndy Whitcroft 17974e35f483SJoonsoo Kim resv = vma_resv_map(vma); 17984e35f483SJoonsoo Kim if (!resv) 1799c37f9fb1SAndy Whitcroft return 1; 1800c37f9fb1SAndy Whitcroft 18014e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 18025e911373SMike Kravetz switch (mode) { 18035e911373SMike Kravetz case VMA_NEEDS_RESV: 1804cf3ad20bSMike Kravetz ret = region_chg(resv, idx, idx + 1); 18055e911373SMike Kravetz break; 18065e911373SMike Kravetz case VMA_COMMIT_RESV: 18075e911373SMike Kravetz ret = region_add(resv, idx, idx + 1); 18085e911373SMike Kravetz break; 1809feba16e2SMike Kravetz case VMA_END_RESV: 18105e911373SMike Kravetz region_abort(resv, idx, idx + 1); 18115e911373SMike Kravetz ret = 0; 18125e911373SMike Kravetz break; 18135e911373SMike Kravetz default: 18145e911373SMike Kravetz BUG(); 18155e911373SMike Kravetz } 181684afd99bSAndy Whitcroft 18174e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) 1818cf3ad20bSMike Kravetz return ret; 18194e35f483SJoonsoo Kim else 1820cf3ad20bSMike Kravetz return ret < 0 ? ret : 0; 182184afd99bSAndy Whitcroft } 1822cf3ad20bSMike Kravetz 1823cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 1824a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 1825c37f9fb1SAndy Whitcroft { 18265e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 1827cf3ad20bSMike Kravetz } 1828c37f9fb1SAndy Whitcroft 1829cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 1830cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 1831cf3ad20bSMike Kravetz { 18325e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 18335e911373SMike Kravetz } 18345e911373SMike Kravetz 1835feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 18365e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 18375e911373SMike Kravetz { 1838feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 1839c37f9fb1SAndy Whitcroft } 1840c37f9fb1SAndy Whitcroft 184170c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 184204f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 1843348ea204SAdam Litke { 184490481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 1845a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 1846348ea204SAdam Litke struct page *page; 1847d85f69b0SMike Kravetz long map_chg, map_commit; 1848d85f69b0SMike Kravetz long gbl_chg; 18496d76dcf4SAneesh Kumar K.V int ret, idx; 18506d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 18512fc39cecSAdam Litke 18526d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 1853a1e78772SMel Gorman /* 1854d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 1855d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 1856d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 1857a1e78772SMel Gorman */ 1858d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 1859d85f69b0SMike Kravetz if (map_chg < 0) 186076dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 1861d85f69b0SMike Kravetz 1862d85f69b0SMike Kravetz /* 1863d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 1864d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 1865d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 1866d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 1867d85f69b0SMike Kravetz * checked against any subpool limit. 1868d85f69b0SMike Kravetz */ 1869d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 1870d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 1871d85f69b0SMike Kravetz if (gbl_chg < 0) { 1872feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 187376dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 18745e911373SMike Kravetz } 187590d8b7e6SAdam Litke 1876d85f69b0SMike Kravetz /* 1877d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 1878d85f69b0SMike Kravetz * map, there could be reservations associated with the 1879d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 1880d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 1881d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 1882d85f69b0SMike Kravetz * the subpool reservations. 1883d85f69b0SMike Kravetz */ 1884d85f69b0SMike Kravetz if (avoid_reserve) 1885d85f69b0SMike Kravetz gbl_chg = 1; 1886d85f69b0SMike Kravetz } 1887d85f69b0SMike Kravetz 18886d76dcf4SAneesh Kumar K.V ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 18898f34af6fSJianyu Zhan if (ret) 18908f34af6fSJianyu Zhan goto out_subpool_put; 18918f34af6fSJianyu Zhan 1892a1e78772SMel Gorman spin_lock(&hugetlb_lock); 1893d85f69b0SMike Kravetz /* 1894d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 1895d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 1896d85f69b0SMike Kravetz * a reservation exists for the allocation. 1897d85f69b0SMike Kravetz */ 1898d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 189981a6fcaeSJoonsoo Kim if (!page) { 190094ae8ba7SAneesh Kumar K.V spin_unlock(&hugetlb_lock); 1901099730d6SDave Hansen page = __alloc_buddy_huge_page_with_mpol(h, vma, addr); 19028f34af6fSJianyu Zhan if (!page) 19038f34af6fSJianyu Zhan goto out_uncharge_cgroup; 1904a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 1905a88c7695SNaoya Horiguchi SetPagePrivate(page); 1906a88c7695SNaoya Horiguchi h->resv_huge_pages--; 1907a88c7695SNaoya Horiguchi } 190879dbb236SAneesh Kumar K.V spin_lock(&hugetlb_lock); 190979dbb236SAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 191081a6fcaeSJoonsoo Kim /* Fall through */ 1911a1e78772SMel Gorman } 191281a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 191381a6fcaeSJoonsoo Kim spin_unlock(&hugetlb_lock); 1914a1e78772SMel Gorman 191590481622SDavid Gibson set_page_private(page, (unsigned long)spool); 1916a1e78772SMel Gorman 1917d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 1918d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 191933039678SMike Kravetz /* 192033039678SMike Kravetz * The page was added to the reservation map between 192133039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 192233039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 192333039678SMike Kravetz * Adjust for the subpool count incremented above AND 192433039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 192533039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 192633039678SMike Kravetz * no longer applies. 192733039678SMike Kravetz */ 192833039678SMike Kravetz long rsv_adjust; 192933039678SMike Kravetz 193033039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 193133039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 193233039678SMike Kravetz } 19337893d1d5SAdam Litke return page; 19348f34af6fSJianyu Zhan 19358f34af6fSJianyu Zhan out_uncharge_cgroup: 19368f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 19378f34af6fSJianyu Zhan out_subpool_put: 1938d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 19398f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 1940feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 19418f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 1942b45b5bd6SDavid Gibson } 1943b45b5bd6SDavid Gibson 194474060e4dSNaoya Horiguchi /* 194574060e4dSNaoya Horiguchi * alloc_huge_page()'s wrapper which simply returns the page if allocation 194674060e4dSNaoya Horiguchi * succeeds, otherwise NULL. This function is called from new_vma_page(), 194774060e4dSNaoya Horiguchi * where no ERR_VALUE is expected to be returned. 194874060e4dSNaoya Horiguchi */ 194974060e4dSNaoya Horiguchi struct page *alloc_huge_page_noerr(struct vm_area_struct *vma, 195074060e4dSNaoya Horiguchi unsigned long addr, int avoid_reserve) 195174060e4dSNaoya Horiguchi { 195274060e4dSNaoya Horiguchi struct page *page = alloc_huge_page(vma, addr, avoid_reserve); 195374060e4dSNaoya Horiguchi if (IS_ERR(page)) 195474060e4dSNaoya Horiguchi page = NULL; 195574060e4dSNaoya Horiguchi return page; 195674060e4dSNaoya Horiguchi } 195774060e4dSNaoya Horiguchi 195891f47662SCyrill Gorcunov int __weak alloc_bootmem_huge_page(struct hstate *h) 1959aa888a74SAndi Kleen { 1960aa888a74SAndi Kleen struct huge_bootmem_page *m; 1961b2261026SJoonsoo Kim int nr_nodes, node; 1962aa888a74SAndi Kleen 1963b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 1964aa888a74SAndi Kleen void *addr; 1965aa888a74SAndi Kleen 19668b89a116SGrygorii Strashko addr = memblock_virt_alloc_try_nid_nopanic( 19678b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 19688b89a116SGrygorii Strashko 0, BOOTMEM_ALLOC_ACCESSIBLE, node); 1969aa888a74SAndi Kleen if (addr) { 1970aa888a74SAndi Kleen /* 1971aa888a74SAndi Kleen * Use the beginning of the huge page to store the 1972aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 1973aa888a74SAndi Kleen * puts them into the mem_map). 1974aa888a74SAndi Kleen */ 1975aa888a74SAndi Kleen m = addr; 1976aa888a74SAndi Kleen goto found; 1977aa888a74SAndi Kleen } 1978aa888a74SAndi Kleen } 1979aa888a74SAndi Kleen return 0; 1980aa888a74SAndi Kleen 1981aa888a74SAndi Kleen found: 1982df994eadSLuiz Capitulino BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); 1983aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 1984aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 1985aa888a74SAndi Kleen m->hstate = h; 1986aa888a74SAndi Kleen return 1; 1987aa888a74SAndi Kleen } 1988aa888a74SAndi Kleen 1989d00181b9SKirill A. Shutemov static void __init prep_compound_huge_page(struct page *page, 1990d00181b9SKirill A. Shutemov unsigned int order) 199118229df5SAndy Whitcroft { 199218229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 199318229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 199418229df5SAndy Whitcroft else 199518229df5SAndy Whitcroft prep_compound_page(page, order); 199618229df5SAndy Whitcroft } 199718229df5SAndy Whitcroft 1998aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 1999aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 2000aa888a74SAndi Kleen { 2001aa888a74SAndi Kleen struct huge_bootmem_page *m; 2002aa888a74SAndi Kleen 2003aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 2004aa888a74SAndi Kleen struct hstate *h = m->hstate; 2005ee8f248dSBecky Bruce struct page *page; 2006ee8f248dSBecky Bruce 2007ee8f248dSBecky Bruce #ifdef CONFIG_HIGHMEM 2008ee8f248dSBecky Bruce page = pfn_to_page(m->phys >> PAGE_SHIFT); 20098b89a116SGrygorii Strashko memblock_free_late(__pa(m), 2010ee8f248dSBecky Bruce sizeof(struct huge_bootmem_page)); 2011ee8f248dSBecky Bruce #else 2012ee8f248dSBecky Bruce page = virt_to_page(m); 2013ee8f248dSBecky Bruce #endif 2014aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 201518229df5SAndy Whitcroft prep_compound_huge_page(page, h->order); 2016ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 2017aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 2018b0320c7bSRafael Aquini /* 2019b0320c7bSRafael Aquini * If we had gigantic hugepages allocated at boot time, we need 2020b0320c7bSRafael Aquini * to restore the 'stolen' pages to totalram_pages in order to 2021b0320c7bSRafael Aquini * fix confusing memory reports from free(1) and another 2022b0320c7bSRafael Aquini * side-effects, like CommitLimit going negative. 2023b0320c7bSRafael Aquini */ 2024bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 20253dcc0571SJiang Liu adjust_managed_page_count(page, 1 << h->order); 2026aa888a74SAndi Kleen } 2027aa888a74SAndi Kleen } 2028aa888a74SAndi Kleen 20298faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 20301da177e4SLinus Torvalds { 20311da177e4SLinus Torvalds unsigned long i; 20321da177e4SLinus Torvalds 2033e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 2034bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 2035aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 2036aa888a74SAndi Kleen break; 20379b5e5d0fSLee Schermerhorn } else if (!alloc_fresh_huge_page(h, 20388cebfcd0SLai Jiangshan &node_states[N_MEMORY])) 20391da177e4SLinus Torvalds break; 20401da177e4SLinus Torvalds } 20418faa8b07SAndi Kleen h->max_huge_pages = i; 2042e5ff2159SAndi Kleen } 2043e5ff2159SAndi Kleen 2044e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 2045e5ff2159SAndi Kleen { 2046e5ff2159SAndi Kleen struct hstate *h; 2047e5ff2159SAndi Kleen 2048e5ff2159SAndi Kleen for_each_hstate(h) { 2049641844f5SNaoya Horiguchi if (minimum_order > huge_page_order(h)) 2050641844f5SNaoya Horiguchi minimum_order = huge_page_order(h); 2051641844f5SNaoya Horiguchi 20528faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 2053bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 20548faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 2055e5ff2159SAndi Kleen } 2056641844f5SNaoya Horiguchi VM_BUG_ON(minimum_order == UINT_MAX); 2057e5ff2159SAndi Kleen } 2058e5ff2159SAndi Kleen 20594abd32dbSAndi Kleen static char * __init memfmt(char *buf, unsigned long n) 20604abd32dbSAndi Kleen { 20614abd32dbSAndi Kleen if (n >= (1UL << 30)) 20624abd32dbSAndi Kleen sprintf(buf, "%lu GB", n >> 30); 20634abd32dbSAndi Kleen else if (n >= (1UL << 20)) 20644abd32dbSAndi Kleen sprintf(buf, "%lu MB", n >> 20); 20654abd32dbSAndi Kleen else 20664abd32dbSAndi Kleen sprintf(buf, "%lu KB", n >> 10); 20674abd32dbSAndi Kleen return buf; 20684abd32dbSAndi Kleen } 20694abd32dbSAndi Kleen 2070e5ff2159SAndi Kleen static void __init report_hugepages(void) 2071e5ff2159SAndi Kleen { 2072e5ff2159SAndi Kleen struct hstate *h; 2073e5ff2159SAndi Kleen 2074e5ff2159SAndi Kleen for_each_hstate(h) { 20754abd32dbSAndi Kleen char buf[32]; 2076ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 20774abd32dbSAndi Kleen memfmt(buf, huge_page_size(h)), 20784abd32dbSAndi Kleen h->free_huge_pages); 2079e5ff2159SAndi Kleen } 2080e5ff2159SAndi Kleen } 2081e5ff2159SAndi Kleen 20821da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 20836ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 20846ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 20851da177e4SLinus Torvalds { 20864415cc8dSChristoph Lameter int i; 20874415cc8dSChristoph Lameter 2088bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2089aa888a74SAndi Kleen return; 2090aa888a74SAndi Kleen 20916ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 20921da177e4SLinus Torvalds struct page *page, *next; 2093a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 2094a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 2095a5516438SAndi Kleen if (count >= h->nr_huge_pages) 20966b0c880dSAdam Litke return; 20971da177e4SLinus Torvalds if (PageHighMem(page)) 20981da177e4SLinus Torvalds continue; 20991da177e4SLinus Torvalds list_del(&page->lru); 2100e5ff2159SAndi Kleen update_and_free_page(h, page); 2101a5516438SAndi Kleen h->free_huge_pages--; 2102a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 21031da177e4SLinus Torvalds } 21041da177e4SLinus Torvalds } 21051da177e4SLinus Torvalds } 21061da177e4SLinus Torvalds #else 21076ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 21086ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 21091da177e4SLinus Torvalds { 21101da177e4SLinus Torvalds } 21111da177e4SLinus Torvalds #endif 21121da177e4SLinus Torvalds 211320a0307cSWu Fengguang /* 211420a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 211520a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 211620a0307cSWu Fengguang * Returns 1 if an adjustment was made. 211720a0307cSWu Fengguang */ 21186ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 21196ae11b27SLee Schermerhorn int delta) 212020a0307cSWu Fengguang { 2121b2261026SJoonsoo Kim int nr_nodes, node; 212220a0307cSWu Fengguang 212320a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 212420a0307cSWu Fengguang 2125e8c5c824SLee Schermerhorn if (delta < 0) { 2126b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 2127b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 2128b2261026SJoonsoo Kim goto found; 2129b2261026SJoonsoo Kim } 2130b2261026SJoonsoo Kim } else { 2131b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2132b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 2133b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 2134b2261026SJoonsoo Kim goto found; 2135e8c5c824SLee Schermerhorn } 21369a76db09SLee Schermerhorn } 2137b2261026SJoonsoo Kim return 0; 213820a0307cSWu Fengguang 2139b2261026SJoonsoo Kim found: 214020a0307cSWu Fengguang h->surplus_huge_pages += delta; 2141b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 2142b2261026SJoonsoo Kim return 1; 214320a0307cSWu Fengguang } 214420a0307cSWu Fengguang 2145a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 21466ae11b27SLee Schermerhorn static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, 21476ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 21481da177e4SLinus Torvalds { 21497893d1d5SAdam Litke unsigned long min_count, ret; 21501da177e4SLinus Torvalds 2151944d9fecSLuiz Capitulino if (hstate_is_gigantic(h) && !gigantic_page_supported()) 2152aa888a74SAndi Kleen return h->max_huge_pages; 2153aa888a74SAndi Kleen 21547893d1d5SAdam Litke /* 21557893d1d5SAdam Litke * Increase the pool size 21567893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 21577893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 2158d1c3fb1fSNishanth Aravamudan * 2159d15c7c09SNaoya Horiguchi * We might race with __alloc_buddy_huge_page() here and be unable 2160d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 2161d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 2162d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 2163d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 21647893d1d5SAdam Litke */ 21651da177e4SLinus Torvalds spin_lock(&hugetlb_lock); 2166a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 21676ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 21687893d1d5SAdam Litke break; 21697893d1d5SAdam Litke } 21707893d1d5SAdam Litke 2171a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 21727893d1d5SAdam Litke /* 21737893d1d5SAdam Litke * If this allocation races such that we no longer need the 21747893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 21757893d1d5SAdam Litke * and reducing the surplus. 21767893d1d5SAdam Litke */ 21777893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 2178944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) 2179944d9fecSLuiz Capitulino ret = alloc_fresh_gigantic_page(h, nodes_allowed); 2180944d9fecSLuiz Capitulino else 21816ae11b27SLee Schermerhorn ret = alloc_fresh_huge_page(h, nodes_allowed); 21827893d1d5SAdam Litke spin_lock(&hugetlb_lock); 21837893d1d5SAdam Litke if (!ret) 21847893d1d5SAdam Litke goto out; 21857893d1d5SAdam Litke 2186536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 2187536240f2SMel Gorman if (signal_pending(current)) 2188536240f2SMel Gorman goto out; 21897893d1d5SAdam Litke } 21907893d1d5SAdam Litke 21917893d1d5SAdam Litke /* 21927893d1d5SAdam Litke * Decrease the pool size 21937893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 21947893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 21957893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 21967893d1d5SAdam Litke * to the desired size as pages become free. 2197d1c3fb1fSNishanth Aravamudan * 2198d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 2199d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 2200d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 2201d15c7c09SNaoya Horiguchi * __alloc_buddy_huge_page() is checking the global counter, 2202d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 2203d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 2204d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 22057893d1d5SAdam Litke */ 2206a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 22076b0c880dSAdam Litke min_count = max(count, min_count); 22086ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 2209a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 22106ae11b27SLee Schermerhorn if (!free_pool_huge_page(h, nodes_allowed, 0)) 22111da177e4SLinus Torvalds break; 221255f67141SMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 22131da177e4SLinus Torvalds } 2214a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 22156ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 22167893d1d5SAdam Litke break; 22177893d1d5SAdam Litke } 22187893d1d5SAdam Litke out: 2219a5516438SAndi Kleen ret = persistent_huge_pages(h); 22201da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 22217893d1d5SAdam Litke return ret; 22221da177e4SLinus Torvalds } 22231da177e4SLinus Torvalds 2224a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 2225a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 2226a3437870SNishanth Aravamudan 2227a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 2228a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 2229a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 2230a3437870SNishanth Aravamudan 2231a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 2232a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 2233a3437870SNishanth Aravamudan 22349a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 22359a305230SLee Schermerhorn 22369a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 2237a3437870SNishanth Aravamudan { 2238a3437870SNishanth Aravamudan int i; 22399a305230SLee Schermerhorn 2240a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 22419a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 22429a305230SLee Schermerhorn if (nidp) 22439a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 2244a3437870SNishanth Aravamudan return &hstates[i]; 22459a305230SLee Schermerhorn } 22469a305230SLee Schermerhorn 22479a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 2248a3437870SNishanth Aravamudan } 2249a3437870SNishanth Aravamudan 225006808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 2251a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2252a3437870SNishanth Aravamudan { 22539a305230SLee Schermerhorn struct hstate *h; 22549a305230SLee Schermerhorn unsigned long nr_huge_pages; 22559a305230SLee Schermerhorn int nid; 22569a305230SLee Schermerhorn 22579a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 22589a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 22599a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 22609a305230SLee Schermerhorn else 22619a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 22629a305230SLee Schermerhorn 22639a305230SLee Schermerhorn return sprintf(buf, "%lu\n", nr_huge_pages); 2264a3437870SNishanth Aravamudan } 2265adbe8726SEric B Munson 2266238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 2267238d3c13SDavid Rientjes struct hstate *h, int nid, 2268238d3c13SDavid Rientjes unsigned long count, size_t len) 2269a3437870SNishanth Aravamudan { 2270a3437870SNishanth Aravamudan int err; 2271bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); 2272a3437870SNishanth Aravamudan 2273944d9fecSLuiz Capitulino if (hstate_is_gigantic(h) && !gigantic_page_supported()) { 2274adbe8726SEric B Munson err = -EINVAL; 2275adbe8726SEric B Munson goto out; 2276adbe8726SEric B Munson } 2277adbe8726SEric B Munson 22789a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 22799a305230SLee Schermerhorn /* 22809a305230SLee Schermerhorn * global hstate attribute 22819a305230SLee Schermerhorn */ 22829a305230SLee Schermerhorn if (!(obey_mempolicy && 22839a305230SLee Schermerhorn init_nodemask_of_mempolicy(nodes_allowed))) { 228406808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 22858cebfcd0SLai Jiangshan nodes_allowed = &node_states[N_MEMORY]; 228606808b08SLee Schermerhorn } 22879a305230SLee Schermerhorn } else if (nodes_allowed) { 22889a305230SLee Schermerhorn /* 22899a305230SLee Schermerhorn * per node hstate attribute: adjust count to global, 22909a305230SLee Schermerhorn * but restrict alloc/free to the specified node. 22919a305230SLee Schermerhorn */ 22929a305230SLee Schermerhorn count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 22939a305230SLee Schermerhorn init_nodemask_of_node(nodes_allowed, nid); 22949a305230SLee Schermerhorn } else 22958cebfcd0SLai Jiangshan nodes_allowed = &node_states[N_MEMORY]; 22969a305230SLee Schermerhorn 229706808b08SLee Schermerhorn h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); 2298a3437870SNishanth Aravamudan 22998cebfcd0SLai Jiangshan if (nodes_allowed != &node_states[N_MEMORY]) 230006808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 230106808b08SLee Schermerhorn 230206808b08SLee Schermerhorn return len; 2303adbe8726SEric B Munson out: 2304adbe8726SEric B Munson NODEMASK_FREE(nodes_allowed); 2305adbe8726SEric B Munson return err; 230606808b08SLee Schermerhorn } 230706808b08SLee Schermerhorn 2308238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 2309238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 2310238d3c13SDavid Rientjes size_t len) 2311238d3c13SDavid Rientjes { 2312238d3c13SDavid Rientjes struct hstate *h; 2313238d3c13SDavid Rientjes unsigned long count; 2314238d3c13SDavid Rientjes int nid; 2315238d3c13SDavid Rientjes int err; 2316238d3c13SDavid Rientjes 2317238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 2318238d3c13SDavid Rientjes if (err) 2319238d3c13SDavid Rientjes return err; 2320238d3c13SDavid Rientjes 2321238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 2322238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 2323238d3c13SDavid Rientjes } 2324238d3c13SDavid Rientjes 232506808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 232606808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 232706808b08SLee Schermerhorn { 232806808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 232906808b08SLee Schermerhorn } 233006808b08SLee Schermerhorn 233106808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 233206808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 233306808b08SLee Schermerhorn { 2334238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 2335a3437870SNishanth Aravamudan } 2336a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 2337a3437870SNishanth Aravamudan 233806808b08SLee Schermerhorn #ifdef CONFIG_NUMA 233906808b08SLee Schermerhorn 234006808b08SLee Schermerhorn /* 234106808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 234206808b08SLee Schermerhorn * huge page alloc/free. 234306808b08SLee Schermerhorn */ 234406808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 234506808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 234606808b08SLee Schermerhorn { 234706808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 234806808b08SLee Schermerhorn } 234906808b08SLee Schermerhorn 235006808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 235106808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 235206808b08SLee Schermerhorn { 2353238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 235406808b08SLee Schermerhorn } 235506808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 235606808b08SLee Schermerhorn #endif 235706808b08SLee Schermerhorn 235806808b08SLee Schermerhorn 2359a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 2360a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2361a3437870SNishanth Aravamudan { 23629a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2363a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 2364a3437870SNishanth Aravamudan } 2365adbe8726SEric B Munson 2366a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 2367a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 2368a3437870SNishanth Aravamudan { 2369a3437870SNishanth Aravamudan int err; 2370a3437870SNishanth Aravamudan unsigned long input; 23719a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2372a3437870SNishanth Aravamudan 2373bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2374adbe8726SEric B Munson return -EINVAL; 2375adbe8726SEric B Munson 23763dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 2377a3437870SNishanth Aravamudan if (err) 237873ae31e5SEric B Munson return err; 2379a3437870SNishanth Aravamudan 2380a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 2381a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 2382a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 2383a3437870SNishanth Aravamudan 2384a3437870SNishanth Aravamudan return count; 2385a3437870SNishanth Aravamudan } 2386a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 2387a3437870SNishanth Aravamudan 2388a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 2389a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2390a3437870SNishanth Aravamudan { 23919a305230SLee Schermerhorn struct hstate *h; 23929a305230SLee Schermerhorn unsigned long free_huge_pages; 23939a305230SLee Schermerhorn int nid; 23949a305230SLee Schermerhorn 23959a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 23969a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 23979a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 23989a305230SLee Schermerhorn else 23999a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 24009a305230SLee Schermerhorn 24019a305230SLee Schermerhorn return sprintf(buf, "%lu\n", free_huge_pages); 2402a3437870SNishanth Aravamudan } 2403a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 2404a3437870SNishanth Aravamudan 2405a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 2406a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2407a3437870SNishanth Aravamudan { 24089a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2409a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->resv_huge_pages); 2410a3437870SNishanth Aravamudan } 2411a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 2412a3437870SNishanth Aravamudan 2413a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 2414a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2415a3437870SNishanth Aravamudan { 24169a305230SLee Schermerhorn struct hstate *h; 24179a305230SLee Schermerhorn unsigned long surplus_huge_pages; 24189a305230SLee Schermerhorn int nid; 24199a305230SLee Schermerhorn 24209a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 24219a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 24229a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 24239a305230SLee Schermerhorn else 24249a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 24259a305230SLee Schermerhorn 24269a305230SLee Schermerhorn return sprintf(buf, "%lu\n", surplus_huge_pages); 2427a3437870SNishanth Aravamudan } 2428a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 2429a3437870SNishanth Aravamudan 2430a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 2431a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 2432a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 2433a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 2434a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 2435a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 243606808b08SLee Schermerhorn #ifdef CONFIG_NUMA 243706808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 243806808b08SLee Schermerhorn #endif 2439a3437870SNishanth Aravamudan NULL, 2440a3437870SNishanth Aravamudan }; 2441a3437870SNishanth Aravamudan 2442a3437870SNishanth Aravamudan static struct attribute_group hstate_attr_group = { 2443a3437870SNishanth Aravamudan .attrs = hstate_attrs, 2444a3437870SNishanth Aravamudan }; 2445a3437870SNishanth Aravamudan 2446094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 24479a305230SLee Schermerhorn struct kobject **hstate_kobjs, 24489a305230SLee Schermerhorn struct attribute_group *hstate_attr_group) 2449a3437870SNishanth Aravamudan { 2450a3437870SNishanth Aravamudan int retval; 2451972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 2452a3437870SNishanth Aravamudan 24539a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 24549a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 2455a3437870SNishanth Aravamudan return -ENOMEM; 2456a3437870SNishanth Aravamudan 24579a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 2458a3437870SNishanth Aravamudan if (retval) 24599a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 2460a3437870SNishanth Aravamudan 2461a3437870SNishanth Aravamudan return retval; 2462a3437870SNishanth Aravamudan } 2463a3437870SNishanth Aravamudan 2464a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 2465a3437870SNishanth Aravamudan { 2466a3437870SNishanth Aravamudan struct hstate *h; 2467a3437870SNishanth Aravamudan int err; 2468a3437870SNishanth Aravamudan 2469a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 2470a3437870SNishanth Aravamudan if (!hugepages_kobj) 2471a3437870SNishanth Aravamudan return; 2472a3437870SNishanth Aravamudan 2473a3437870SNishanth Aravamudan for_each_hstate(h) { 24749a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 24759a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 2476a3437870SNishanth Aravamudan if (err) 2477ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s", h->name); 2478a3437870SNishanth Aravamudan } 2479a3437870SNishanth Aravamudan } 2480a3437870SNishanth Aravamudan 24819a305230SLee Schermerhorn #ifdef CONFIG_NUMA 24829a305230SLee Schermerhorn 24839a305230SLee Schermerhorn /* 24849a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 248510fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 248610fbcf4cSKay Sievers * index of a node device or _hstate == node id. 248710fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 24889a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 24899a305230SLee Schermerhorn */ 24909a305230SLee Schermerhorn struct node_hstate { 24919a305230SLee Schermerhorn struct kobject *hugepages_kobj; 24929a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 24939a305230SLee Schermerhorn }; 2494b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 24959a305230SLee Schermerhorn 24969a305230SLee Schermerhorn /* 249710fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 24989a305230SLee Schermerhorn */ 24999a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 25009a305230SLee Schermerhorn &nr_hugepages_attr.attr, 25019a305230SLee Schermerhorn &free_hugepages_attr.attr, 25029a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 25039a305230SLee Schermerhorn NULL, 25049a305230SLee Schermerhorn }; 25059a305230SLee Schermerhorn 25069a305230SLee Schermerhorn static struct attribute_group per_node_hstate_attr_group = { 25079a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 25089a305230SLee Schermerhorn }; 25099a305230SLee Schermerhorn 25109a305230SLee Schermerhorn /* 251110fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 25129a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 25139a305230SLee Schermerhorn */ 25149a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 25159a305230SLee Schermerhorn { 25169a305230SLee Schermerhorn int nid; 25179a305230SLee Schermerhorn 25189a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 25199a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 25209a305230SLee Schermerhorn int i; 25219a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 25229a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 25239a305230SLee Schermerhorn if (nidp) 25249a305230SLee Schermerhorn *nidp = nid; 25259a305230SLee Schermerhorn return &hstates[i]; 25269a305230SLee Schermerhorn } 25279a305230SLee Schermerhorn } 25289a305230SLee Schermerhorn 25299a305230SLee Schermerhorn BUG(); 25309a305230SLee Schermerhorn return NULL; 25319a305230SLee Schermerhorn } 25329a305230SLee Schermerhorn 25339a305230SLee Schermerhorn /* 253410fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 25359a305230SLee Schermerhorn * No-op if no hstate attributes attached. 25369a305230SLee Schermerhorn */ 25373cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 25389a305230SLee Schermerhorn { 25399a305230SLee Schermerhorn struct hstate *h; 254010fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 25419a305230SLee Schermerhorn 25429a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 25439b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 25449a305230SLee Schermerhorn 2545972dc4deSAneesh Kumar K.V for_each_hstate(h) { 2546972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 2547972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 2548972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 2549972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 2550972dc4deSAneesh Kumar K.V } 25519a305230SLee Schermerhorn } 25529a305230SLee Schermerhorn 25539a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 25549a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 25559a305230SLee Schermerhorn } 25569a305230SLee Schermerhorn 25579a305230SLee Schermerhorn 25589a305230SLee Schermerhorn /* 255910fbcf4cSKay Sievers * Register hstate attributes for a single node device. 25609a305230SLee Schermerhorn * No-op if attributes already registered. 25619a305230SLee Schermerhorn */ 25623cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 25639a305230SLee Schermerhorn { 25649a305230SLee Schermerhorn struct hstate *h; 256510fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 25669a305230SLee Schermerhorn int err; 25679a305230SLee Schermerhorn 25689a305230SLee Schermerhorn if (nhs->hugepages_kobj) 25699a305230SLee Schermerhorn return; /* already allocated */ 25709a305230SLee Schermerhorn 25719a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 257210fbcf4cSKay Sievers &node->dev.kobj); 25739a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 25749a305230SLee Schermerhorn return; 25759a305230SLee Schermerhorn 25769a305230SLee Schermerhorn for_each_hstate(h) { 25779a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 25789a305230SLee Schermerhorn nhs->hstate_kobjs, 25799a305230SLee Schermerhorn &per_node_hstate_attr_group); 25809a305230SLee Schermerhorn if (err) { 2581ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s for node %d\n", 258210fbcf4cSKay Sievers h->name, node->dev.id); 25839a305230SLee Schermerhorn hugetlb_unregister_node(node); 25849a305230SLee Schermerhorn break; 25859a305230SLee Schermerhorn } 25869a305230SLee Schermerhorn } 25879a305230SLee Schermerhorn } 25889a305230SLee Schermerhorn 25899a305230SLee Schermerhorn /* 25909b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 259110fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 259210fbcf4cSKay Sievers * registered their associated device by this time. 25939a305230SLee Schermerhorn */ 25947d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 25959a305230SLee Schermerhorn { 25969a305230SLee Schermerhorn int nid; 25979a305230SLee Schermerhorn 25988cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 25998732794bSWen Congyang struct node *node = node_devices[nid]; 260010fbcf4cSKay Sievers if (node->dev.id == nid) 26019a305230SLee Schermerhorn hugetlb_register_node(node); 26029a305230SLee Schermerhorn } 26039a305230SLee Schermerhorn 26049a305230SLee Schermerhorn /* 260510fbcf4cSKay Sievers * Let the node device driver know we're here so it can 26069a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 26079a305230SLee Schermerhorn */ 26089a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 26099a305230SLee Schermerhorn hugetlb_unregister_node); 26109a305230SLee Schermerhorn } 26119a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 26129a305230SLee Schermerhorn 26139a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 26149a305230SLee Schermerhorn { 26159a305230SLee Schermerhorn BUG(); 26169a305230SLee Schermerhorn if (nidp) 26179a305230SLee Schermerhorn *nidp = -1; 26189a305230SLee Schermerhorn return NULL; 26199a305230SLee Schermerhorn } 26209a305230SLee Schermerhorn 26219a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 26229a305230SLee Schermerhorn 26239a305230SLee Schermerhorn #endif 26249a305230SLee Schermerhorn 2625a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 2626a3437870SNishanth Aravamudan { 26278382d914SDavidlohr Bueso int i; 26288382d914SDavidlohr Bueso 2629457c1b27SNishanth Aravamudan if (!hugepages_supported()) 26300ef89d25SBenjamin Herrenschmidt return 0; 2631a3437870SNishanth Aravamudan 2632e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) { 2633e11bfbfcSNick Piggin default_hstate_size = HPAGE_SIZE; 2634e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) 2635a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 2636a3437870SNishanth Aravamudan } 2637972dc4deSAneesh Kumar K.V default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); 2638f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 2639f8b74815SVaishali Thakkar if (!default_hstate.max_huge_pages) 2640e11bfbfcSNick Piggin default_hstate.max_huge_pages = default_hstate_max_huge_pages; 2641f8b74815SVaishali Thakkar } 2642a3437870SNishanth Aravamudan 2643a3437870SNishanth Aravamudan hugetlb_init_hstates(); 2644aa888a74SAndi Kleen gather_bootmem_prealloc(); 2645a3437870SNishanth Aravamudan report_hugepages(); 2646a3437870SNishanth Aravamudan 2647a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 26489a305230SLee Schermerhorn hugetlb_register_all_nodes(); 26497179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 26509a305230SLee Schermerhorn 26518382d914SDavidlohr Bueso #ifdef CONFIG_SMP 26528382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 26538382d914SDavidlohr Bueso #else 26548382d914SDavidlohr Bueso num_fault_mutexes = 1; 26558382d914SDavidlohr Bueso #endif 2656c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 26578382d914SDavidlohr Bueso kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL); 2658c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 26598382d914SDavidlohr Bueso 26608382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 2661c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 2662a3437870SNishanth Aravamudan return 0; 2663a3437870SNishanth Aravamudan } 26643e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 2665a3437870SNishanth Aravamudan 2666a3437870SNishanth Aravamudan /* Should be called on processing a hugepagesz=... option */ 26679fee021dSVaishali Thakkar void __init hugetlb_bad_size(void) 26689fee021dSVaishali Thakkar { 26699fee021dSVaishali Thakkar parsed_valid_hugepagesz = false; 26709fee021dSVaishali Thakkar } 26719fee021dSVaishali Thakkar 2672d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 2673a3437870SNishanth Aravamudan { 2674a3437870SNishanth Aravamudan struct hstate *h; 26758faa8b07SAndi Kleen unsigned long i; 26768faa8b07SAndi Kleen 2677a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 2678598d8091SJoe Perches pr_warn("hugepagesz= specified twice, ignoring\n"); 2679a3437870SNishanth Aravamudan return; 2680a3437870SNishanth Aravamudan } 268147d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 2682a3437870SNishanth Aravamudan BUG_ON(order == 0); 268347d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 2684a3437870SNishanth Aravamudan h->order = order; 2685a3437870SNishanth Aravamudan h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 26868faa8b07SAndi Kleen h->nr_huge_pages = 0; 26878faa8b07SAndi Kleen h->free_huge_pages = 0; 26888faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 26898faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 26900edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 269154f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 269254f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 2693a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 2694a3437870SNishanth Aravamudan huge_page_size(h)/1024); 26958faa8b07SAndi Kleen 2696a3437870SNishanth Aravamudan parsed_hstate = h; 2697a3437870SNishanth Aravamudan } 2698a3437870SNishanth Aravamudan 2699e11bfbfcSNick Piggin static int __init hugetlb_nrpages_setup(char *s) 2700a3437870SNishanth Aravamudan { 2701a3437870SNishanth Aravamudan unsigned long *mhp; 27028faa8b07SAndi Kleen static unsigned long *last_mhp; 2703a3437870SNishanth Aravamudan 27049fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 27059fee021dSVaishali Thakkar pr_warn("hugepages = %s preceded by " 27069fee021dSVaishali Thakkar "an unsupported hugepagesz, ignoring\n", s); 27079fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 27089fee021dSVaishali Thakkar return 1; 27099fee021dSVaishali Thakkar } 2710a3437870SNishanth Aravamudan /* 271147d38344SAneesh Kumar K.V * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, 2712a3437870SNishanth Aravamudan * so this hugepages= parameter goes to the "default hstate". 2713a3437870SNishanth Aravamudan */ 27149fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 2715a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 2716a3437870SNishanth Aravamudan else 2717a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 2718a3437870SNishanth Aravamudan 27198faa8b07SAndi Kleen if (mhp == last_mhp) { 2720598d8091SJoe Perches pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n"); 27218faa8b07SAndi Kleen return 1; 27228faa8b07SAndi Kleen } 27238faa8b07SAndi Kleen 2724a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 2725a3437870SNishanth Aravamudan *mhp = 0; 2726a3437870SNishanth Aravamudan 27278faa8b07SAndi Kleen /* 27288faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 27298faa8b07SAndi Kleen * But we need to allocate >= MAX_ORDER hstates here early to still 27308faa8b07SAndi Kleen * use the bootmem allocator. 27318faa8b07SAndi Kleen */ 273247d38344SAneesh Kumar K.V if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) 27338faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 27348faa8b07SAndi Kleen 27358faa8b07SAndi Kleen last_mhp = mhp; 27368faa8b07SAndi Kleen 2737a3437870SNishanth Aravamudan return 1; 2738a3437870SNishanth Aravamudan } 2739e11bfbfcSNick Piggin __setup("hugepages=", hugetlb_nrpages_setup); 2740e11bfbfcSNick Piggin 2741e11bfbfcSNick Piggin static int __init hugetlb_default_setup(char *s) 2742e11bfbfcSNick Piggin { 2743e11bfbfcSNick Piggin default_hstate_size = memparse(s, &s); 2744e11bfbfcSNick Piggin return 1; 2745e11bfbfcSNick Piggin } 2746e11bfbfcSNick Piggin __setup("default_hugepagesz=", hugetlb_default_setup); 2747a3437870SNishanth Aravamudan 27488a213460SNishanth Aravamudan static unsigned int cpuset_mems_nr(unsigned int *array) 27498a213460SNishanth Aravamudan { 27508a213460SNishanth Aravamudan int node; 27518a213460SNishanth Aravamudan unsigned int nr = 0; 27528a213460SNishanth Aravamudan 27538a213460SNishanth Aravamudan for_each_node_mask(node, cpuset_current_mems_allowed) 27548a213460SNishanth Aravamudan nr += array[node]; 27558a213460SNishanth Aravamudan 27568a213460SNishanth Aravamudan return nr; 27578a213460SNishanth Aravamudan } 27588a213460SNishanth Aravamudan 27598a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 276006808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 276106808b08SLee Schermerhorn struct ctl_table *table, int write, 276206808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 27631da177e4SLinus Torvalds { 2764e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 2765238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 276608d4a246SMichal Hocko int ret; 2767e5ff2159SAndi Kleen 2768457c1b27SNishanth Aravamudan if (!hugepages_supported()) 276986613628SJan Stancek return -EOPNOTSUPP; 2770457c1b27SNishanth Aravamudan 2771e5ff2159SAndi Kleen table->data = &tmp; 2772e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 277308d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 277408d4a246SMichal Hocko if (ret) 277508d4a246SMichal Hocko goto out; 2776e5ff2159SAndi Kleen 2777238d3c13SDavid Rientjes if (write) 2778238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 2779238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 278008d4a246SMichal Hocko out: 278108d4a246SMichal Hocko return ret; 27821da177e4SLinus Torvalds } 2783396faf03SMel Gorman 278406808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 278506808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 278606808b08SLee Schermerhorn { 278706808b08SLee Schermerhorn 278806808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 278906808b08SLee Schermerhorn buffer, length, ppos); 279006808b08SLee Schermerhorn } 279106808b08SLee Schermerhorn 279206808b08SLee Schermerhorn #ifdef CONFIG_NUMA 279306808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 279406808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 279506808b08SLee Schermerhorn { 279606808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 279706808b08SLee Schermerhorn buffer, length, ppos); 279806808b08SLee Schermerhorn } 279906808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 280006808b08SLee Schermerhorn 2801a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 28028d65af78SAlexey Dobriyan void __user *buffer, 2803a3d0c6aaSNishanth Aravamudan size_t *length, loff_t *ppos) 2804a3d0c6aaSNishanth Aravamudan { 2805a5516438SAndi Kleen struct hstate *h = &default_hstate; 2806e5ff2159SAndi Kleen unsigned long tmp; 280708d4a246SMichal Hocko int ret; 2808e5ff2159SAndi Kleen 2809457c1b27SNishanth Aravamudan if (!hugepages_supported()) 281086613628SJan Stancek return -EOPNOTSUPP; 2811457c1b27SNishanth Aravamudan 2812e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 2813e5ff2159SAndi Kleen 2814bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 2815adbe8726SEric B Munson return -EINVAL; 2816adbe8726SEric B Munson 2817e5ff2159SAndi Kleen table->data = &tmp; 2818e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 281908d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 282008d4a246SMichal Hocko if (ret) 282108d4a246SMichal Hocko goto out; 2822e5ff2159SAndi Kleen 2823e5ff2159SAndi Kleen if (write) { 2824064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 2825e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 2826a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 2827e5ff2159SAndi Kleen } 282808d4a246SMichal Hocko out: 282908d4a246SMichal Hocko return ret; 2830a3d0c6aaSNishanth Aravamudan } 2831a3d0c6aaSNishanth Aravamudan 28321da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 28331da177e4SLinus Torvalds 2834e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 28351da177e4SLinus Torvalds { 2836a5516438SAndi Kleen struct hstate *h = &default_hstate; 2837457c1b27SNishanth Aravamudan if (!hugepages_supported()) 2838457c1b27SNishanth Aravamudan return; 2839e1759c21SAlexey Dobriyan seq_printf(m, 28401da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 28411da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 2842b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 28437893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 28444f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 2845a5516438SAndi Kleen h->nr_huge_pages, 2846a5516438SAndi Kleen h->free_huge_pages, 2847a5516438SAndi Kleen h->resv_huge_pages, 2848a5516438SAndi Kleen h->surplus_huge_pages, 2849a5516438SAndi Kleen 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 28501da177e4SLinus Torvalds } 28511da177e4SLinus Torvalds 28521da177e4SLinus Torvalds int hugetlb_report_node_meminfo(int nid, char *buf) 28531da177e4SLinus Torvalds { 2854a5516438SAndi Kleen struct hstate *h = &default_hstate; 2855457c1b27SNishanth Aravamudan if (!hugepages_supported()) 2856457c1b27SNishanth Aravamudan return 0; 28571da177e4SLinus Torvalds return sprintf(buf, 28581da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 2859a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 2860a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 2861a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 2862a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 2863a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 28641da177e4SLinus Torvalds } 28651da177e4SLinus Torvalds 2866949f7ec5SDavid Rientjes void hugetlb_show_meminfo(void) 2867949f7ec5SDavid Rientjes { 2868949f7ec5SDavid Rientjes struct hstate *h; 2869949f7ec5SDavid Rientjes int nid; 2870949f7ec5SDavid Rientjes 2871457c1b27SNishanth Aravamudan if (!hugepages_supported()) 2872457c1b27SNishanth Aravamudan return; 2873457c1b27SNishanth Aravamudan 2874949f7ec5SDavid Rientjes for_each_node_state(nid, N_MEMORY) 2875949f7ec5SDavid Rientjes for_each_hstate(h) 2876949f7ec5SDavid Rientjes pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 2877949f7ec5SDavid Rientjes nid, 2878949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 2879949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 2880949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 2881949f7ec5SDavid Rientjes 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 2882949f7ec5SDavid Rientjes } 2883949f7ec5SDavid Rientjes 28845d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 28855d317b2bSNaoya Horiguchi { 28865d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 28875d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 28885d317b2bSNaoya Horiguchi } 28895d317b2bSNaoya Horiguchi 28901da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 28911da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 28921da177e4SLinus Torvalds { 2893d0028588SWanpeng Li struct hstate *h; 2894d0028588SWanpeng Li unsigned long nr_total_pages = 0; 2895d0028588SWanpeng Li 2896d0028588SWanpeng Li for_each_hstate(h) 2897d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 2898d0028588SWanpeng Li return nr_total_pages; 28991da177e4SLinus Torvalds } 29001da177e4SLinus Torvalds 2901a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 2902fc1b8a73SMel Gorman { 2903fc1b8a73SMel Gorman int ret = -ENOMEM; 2904fc1b8a73SMel Gorman 2905fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 2906fc1b8a73SMel Gorman /* 2907fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 2908fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 2909fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 2910fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 2911fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 2912fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 2913fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 2914fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 2915fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 2916fc1b8a73SMel Gorman * 2917fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 2918fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 2919fc1b8a73SMel Gorman * we fall back to check against current free page availability as 2920fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 2921fc1b8a73SMel Gorman * semantics that cpuset has. 2922fc1b8a73SMel Gorman */ 2923fc1b8a73SMel Gorman if (delta > 0) { 2924a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 2925fc1b8a73SMel Gorman goto out; 2926fc1b8a73SMel Gorman 2927a5516438SAndi Kleen if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 2928a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 2929fc1b8a73SMel Gorman goto out; 2930fc1b8a73SMel Gorman } 2931fc1b8a73SMel Gorman } 2932fc1b8a73SMel Gorman 2933fc1b8a73SMel Gorman ret = 0; 2934fc1b8a73SMel Gorman if (delta < 0) 2935a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 2936fc1b8a73SMel Gorman 2937fc1b8a73SMel Gorman out: 2938fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 2939fc1b8a73SMel Gorman return ret; 2940fc1b8a73SMel Gorman } 2941fc1b8a73SMel Gorman 294284afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 294384afd99bSAndy Whitcroft { 2944f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 294584afd99bSAndy Whitcroft 294684afd99bSAndy Whitcroft /* 294784afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 294884afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 294984afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 295025985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 295184afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 295284afd99bSAndy Whitcroft * new reference here without additional locking. 295384afd99bSAndy Whitcroft */ 29544e35f483SJoonsoo Kim if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 2955f522c3acSJoonsoo Kim kref_get(&resv->refs); 295684afd99bSAndy Whitcroft } 295784afd99bSAndy Whitcroft 2958a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 2959a1e78772SMel Gorman { 2960a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2961f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 296290481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 29634e35f483SJoonsoo Kim unsigned long reserve, start, end; 29641c5ecae3SMike Kravetz long gbl_reserve; 296584afd99bSAndy Whitcroft 29664e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 29674e35f483SJoonsoo Kim return; 29684e35f483SJoonsoo Kim 2969a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 2970a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 297184afd99bSAndy Whitcroft 29724e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 297384afd99bSAndy Whitcroft 2974f031dd27SJoonsoo Kim kref_put(&resv->refs, resv_map_release); 297584afd99bSAndy Whitcroft 29767251ff78SAdam Litke if (reserve) { 29771c5ecae3SMike Kravetz /* 29781c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 29791c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 29801c5ecae3SMike Kravetz */ 29811c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 29821c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 29837251ff78SAdam Litke } 2984a1e78772SMel Gorman } 2985a1e78772SMel Gorman 29861da177e4SLinus Torvalds /* 29871da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 29881da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 29891da177e4SLinus Torvalds * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 29901da177e4SLinus Torvalds * this far. 29911da177e4SLinus Torvalds */ 2992d0217ac0SNick Piggin static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) 29931da177e4SLinus Torvalds { 29941da177e4SLinus Torvalds BUG(); 2995d0217ac0SNick Piggin return 0; 29961da177e4SLinus Torvalds } 29971da177e4SLinus Torvalds 2998f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 2999d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 300084afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 3001a1e78772SMel Gorman .close = hugetlb_vm_op_close, 30021da177e4SLinus Torvalds }; 30031da177e4SLinus Torvalds 30041e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 30051e8f889bSDavid Gibson int writable) 300663551ae0SDavid Gibson { 300763551ae0SDavid Gibson pte_t entry; 300863551ae0SDavid Gibson 30091e8f889bSDavid Gibson if (writable) { 3010106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 3011106c992aSGerald Schaefer vma->vm_page_prot))); 301263551ae0SDavid Gibson } else { 3013106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 3014106c992aSGerald Schaefer vma->vm_page_prot)); 301563551ae0SDavid Gibson } 301663551ae0SDavid Gibson entry = pte_mkyoung(entry); 301763551ae0SDavid Gibson entry = pte_mkhuge(entry); 3018d9ed9faaSChris Metcalf entry = arch_make_huge_pte(entry, vma, page, writable); 301963551ae0SDavid Gibson 302063551ae0SDavid Gibson return entry; 302163551ae0SDavid Gibson } 302263551ae0SDavid Gibson 30231e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 30241e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 30251e8f889bSDavid Gibson { 30261e8f889bSDavid Gibson pte_t entry; 30271e8f889bSDavid Gibson 3028106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 302932f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 30304b3073e1SRussell King update_mmu_cache(vma, address, ptep); 30311e8f889bSDavid Gibson } 30321e8f889bSDavid Gibson 30334a705fefSNaoya Horiguchi static int is_hugetlb_entry_migration(pte_t pte) 30344a705fefSNaoya Horiguchi { 30354a705fefSNaoya Horiguchi swp_entry_t swp; 30364a705fefSNaoya Horiguchi 30374a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 30384a705fefSNaoya Horiguchi return 0; 30394a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 30404a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_migration_entry(swp)) 30414a705fefSNaoya Horiguchi return 1; 30424a705fefSNaoya Horiguchi else 30434a705fefSNaoya Horiguchi return 0; 30444a705fefSNaoya Horiguchi } 30454a705fefSNaoya Horiguchi 30464a705fefSNaoya Horiguchi static int is_hugetlb_entry_hwpoisoned(pte_t pte) 30474a705fefSNaoya Horiguchi { 30484a705fefSNaoya Horiguchi swp_entry_t swp; 30494a705fefSNaoya Horiguchi 30504a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 30514a705fefSNaoya Horiguchi return 0; 30524a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 30534a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_hwpoison_entry(swp)) 30544a705fefSNaoya Horiguchi return 1; 30554a705fefSNaoya Horiguchi else 30564a705fefSNaoya Horiguchi return 0; 30574a705fefSNaoya Horiguchi } 30581e8f889bSDavid Gibson 305963551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 306063551ae0SDavid Gibson struct vm_area_struct *vma) 306163551ae0SDavid Gibson { 306263551ae0SDavid Gibson pte_t *src_pte, *dst_pte, entry; 306363551ae0SDavid Gibson struct page *ptepage; 30641c59827dSHugh Dickins unsigned long addr; 30651e8f889bSDavid Gibson int cow; 3066a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3067a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 3068e8569dd2SAndreas Sandberg unsigned long mmun_start; /* For mmu_notifiers */ 3069e8569dd2SAndreas Sandberg unsigned long mmun_end; /* For mmu_notifiers */ 3070e8569dd2SAndreas Sandberg int ret = 0; 30711e8f889bSDavid Gibson 30721e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 307363551ae0SDavid Gibson 3074e8569dd2SAndreas Sandberg mmun_start = vma->vm_start; 3075e8569dd2SAndreas Sandberg mmun_end = vma->vm_end; 3076e8569dd2SAndreas Sandberg if (cow) 3077e8569dd2SAndreas Sandberg mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end); 3078e8569dd2SAndreas Sandberg 3079a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 3080cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 3081c74df32cSHugh Dickins src_pte = huge_pte_offset(src, addr); 3082c74df32cSHugh Dickins if (!src_pte) 3083c74df32cSHugh Dickins continue; 3084a5516438SAndi Kleen dst_pte = huge_pte_alloc(dst, addr, sz); 3085e8569dd2SAndreas Sandberg if (!dst_pte) { 3086e8569dd2SAndreas Sandberg ret = -ENOMEM; 3087e8569dd2SAndreas Sandberg break; 3088e8569dd2SAndreas Sandberg } 3089c5c99429SLarry Woodman 3090c5c99429SLarry Woodman /* If the pagetables are shared don't copy or take references */ 3091c5c99429SLarry Woodman if (dst_pte == src_pte) 3092c5c99429SLarry Woodman continue; 3093c5c99429SLarry Woodman 3094cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 3095cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 3096cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 30974a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 30984a705fefSNaoya Horiguchi if (huge_pte_none(entry)) { /* skip none entry */ 30994a705fefSNaoya Horiguchi ; 31004a705fefSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry) || 31014a705fefSNaoya Horiguchi is_hugetlb_entry_hwpoisoned(entry))) { 31024a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 31034a705fefSNaoya Horiguchi 31044a705fefSNaoya Horiguchi if (is_write_migration_entry(swp_entry) && cow) { 31054a705fefSNaoya Horiguchi /* 31064a705fefSNaoya Horiguchi * COW mappings require pages in both 31074a705fefSNaoya Horiguchi * parent and child to be set to read. 31084a705fefSNaoya Horiguchi */ 31094a705fefSNaoya Horiguchi make_migration_entry_read(&swp_entry); 31104a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 31114a705fefSNaoya Horiguchi set_huge_pte_at(src, addr, src_pte, entry); 31124a705fefSNaoya Horiguchi } 31134a705fefSNaoya Horiguchi set_huge_pte_at(dst, addr, dst_pte, entry); 31144a705fefSNaoya Horiguchi } else { 311534ee645eSJoerg Roedel if (cow) { 31167f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 311734ee645eSJoerg Roedel mmu_notifier_invalidate_range(src, mmun_start, 311834ee645eSJoerg Roedel mmun_end); 311934ee645eSJoerg Roedel } 31200253d634SNaoya Horiguchi entry = huge_ptep_get(src_pte); 312163551ae0SDavid Gibson ptepage = pte_page(entry); 312263551ae0SDavid Gibson get_page(ptepage); 312353f9263bSKirill A. Shutemov page_dup_rmap(ptepage, true); 312463551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 31255d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), dst); 31261c59827dSHugh Dickins } 3127cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 3128cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 312963551ae0SDavid Gibson } 313063551ae0SDavid Gibson 3131e8569dd2SAndreas Sandberg if (cow) 3132e8569dd2SAndreas Sandberg mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end); 3133e8569dd2SAndreas Sandberg 3134e8569dd2SAndreas Sandberg return ret; 313563551ae0SDavid Gibson } 313663551ae0SDavid Gibson 313724669e58SAneesh Kumar K.V void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 313824669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 313924669e58SAneesh Kumar K.V struct page *ref_page) 314063551ae0SDavid Gibson { 314124669e58SAneesh Kumar K.V int force_flush = 0; 314263551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 314363551ae0SDavid Gibson unsigned long address; 3144c7546f8fSDavid Gibson pte_t *ptep; 314563551ae0SDavid Gibson pte_t pte; 3146cb900f41SKirill A. Shutemov spinlock_t *ptl; 314763551ae0SDavid Gibson struct page *page; 3148a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3149a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 31502ec74c3eSSagi Grimberg const unsigned long mmun_start = start; /* For mmu_notifiers */ 31512ec74c3eSSagi Grimberg const unsigned long mmun_end = end; /* For mmu_notifiers */ 3152a5516438SAndi Kleen 315363551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 3154a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 3155a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 315663551ae0SDavid Gibson 315724669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 31582ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 3159569f48b8SHillf Danton address = start; 316024669e58SAneesh Kumar K.V again: 3161569f48b8SHillf Danton for (; address < end; address += sz) { 3162c7546f8fSDavid Gibson ptep = huge_pte_offset(mm, address); 3163c7546f8fSDavid Gibson if (!ptep) 3164c7546f8fSDavid Gibson continue; 3165c7546f8fSDavid Gibson 3166cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 316739dde65cSChen, Kenneth W if (huge_pmd_unshare(mm, &address, ptep)) 3168cb900f41SKirill A. Shutemov goto unlock; 316939dde65cSChen, Kenneth W 31706629326bSHillf Danton pte = huge_ptep_get(ptep); 31716629326bSHillf Danton if (huge_pte_none(pte)) 3172cb900f41SKirill A. Shutemov goto unlock; 31736629326bSHillf Danton 31746629326bSHillf Danton /* 31759fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 31769fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 31776629326bSHillf Danton */ 31789fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 3179106c992aSGerald Schaefer huge_pte_clear(mm, address, ptep); 3180cb900f41SKirill A. Shutemov goto unlock; 31818c4894c6SNaoya Horiguchi } 31826629326bSHillf Danton 31836629326bSHillf Danton page = pte_page(pte); 318404f2cbe3SMel Gorman /* 318504f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 318604f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 318704f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 318804f2cbe3SMel Gorman */ 318904f2cbe3SMel Gorman if (ref_page) { 319004f2cbe3SMel Gorman if (page != ref_page) 3191cb900f41SKirill A. Shutemov goto unlock; 319204f2cbe3SMel Gorman 319304f2cbe3SMel Gorman /* 319404f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 319504f2cbe3SMel Gorman * future faults in this VMA will fail rather than 319604f2cbe3SMel Gorman * looking like data was lost 319704f2cbe3SMel Gorman */ 319804f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 319904f2cbe3SMel Gorman } 320004f2cbe3SMel Gorman 3201c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 320224669e58SAneesh Kumar K.V tlb_remove_tlb_entry(tlb, ptep, address); 3203106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 32046649a386SKen Chen set_page_dirty(page); 32059e81130bSHillf Danton 32065d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 3207d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 320824669e58SAneesh Kumar K.V force_flush = !__tlb_remove_page(tlb, page); 3209cb900f41SKirill A. Shutemov if (force_flush) { 3210569f48b8SHillf Danton address += sz; 3211cb900f41SKirill A. Shutemov spin_unlock(ptl); 32129e81130bSHillf Danton break; 321363551ae0SDavid Gibson } 3214cb900f41SKirill A. Shutemov /* Bail out after unmapping reference page if supplied */ 3215cb900f41SKirill A. Shutemov if (ref_page) { 3216cb900f41SKirill A. Shutemov spin_unlock(ptl); 3217cb900f41SKirill A. Shutemov break; 3218cb900f41SKirill A. Shutemov } 3219cb900f41SKirill A. Shutemov unlock: 3220cb900f41SKirill A. Shutemov spin_unlock(ptl); 3221cb900f41SKirill A. Shutemov } 322224669e58SAneesh Kumar K.V /* 322324669e58SAneesh Kumar K.V * mmu_gather ran out of room to batch pages, we break out of 322424669e58SAneesh Kumar K.V * the PTE lock to avoid doing the potential expensive TLB invalidate 322524669e58SAneesh Kumar K.V * and page-free while holding it. 322624669e58SAneesh Kumar K.V */ 322724669e58SAneesh Kumar K.V if (force_flush) { 322824669e58SAneesh Kumar K.V force_flush = 0; 322924669e58SAneesh Kumar K.V tlb_flush_mmu(tlb); 323024669e58SAneesh Kumar K.V if (address < end && !ref_page) 323124669e58SAneesh Kumar K.V goto again; 3232fe1668aeSChen, Kenneth W } 32332ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 323424669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 32351da177e4SLinus Torvalds } 323663551ae0SDavid Gibson 3237d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 3238d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 3239d833352aSMel Gorman unsigned long end, struct page *ref_page) 3240d833352aSMel Gorman { 3241d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 3242d833352aSMel Gorman 3243d833352aSMel Gorman /* 3244d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 3245d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 3246d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 3247d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 3248d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 3249c8c06efaSDavidlohr Bueso * is to clear it before releasing the i_mmap_rwsem. This works 3250d833352aSMel Gorman * because in the context this is called, the VMA is about to be 3251c8c06efaSDavidlohr Bueso * destroyed and the i_mmap_rwsem is held. 3252d833352aSMel Gorman */ 3253d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 3254d833352aSMel Gorman } 3255d833352aSMel Gorman 3256502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 325704f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 3258502717f4SChen, Kenneth W { 325924669e58SAneesh Kumar K.V struct mm_struct *mm; 326024669e58SAneesh Kumar K.V struct mmu_gather tlb; 326124669e58SAneesh Kumar K.V 326224669e58SAneesh Kumar K.V mm = vma->vm_mm; 326324669e58SAneesh Kumar K.V 32642b047252SLinus Torvalds tlb_gather_mmu(&tlb, mm, start, end); 326524669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 326624669e58SAneesh Kumar K.V tlb_finish_mmu(&tlb, start, end); 3267502717f4SChen, Kenneth W } 3268502717f4SChen, Kenneth W 326904f2cbe3SMel Gorman /* 327004f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 327104f2cbe3SMel Gorman * mappping it owns the reserve page for. The intention is to unmap the page 327204f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 327304f2cbe3SMel Gorman * same region. 327404f2cbe3SMel Gorman */ 32752f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 32762a4b3dedSHarvey Harrison struct page *page, unsigned long address) 327704f2cbe3SMel Gorman { 32787526674dSAdam Litke struct hstate *h = hstate_vma(vma); 327904f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 328004f2cbe3SMel Gorman struct address_space *mapping; 328104f2cbe3SMel Gorman pgoff_t pgoff; 328204f2cbe3SMel Gorman 328304f2cbe3SMel Gorman /* 328404f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 328504f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 328604f2cbe3SMel Gorman */ 32877526674dSAdam Litke address = address & huge_page_mask(h); 328836e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 328936e4f20aSMichal Hocko vma->vm_pgoff; 329093c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 329104f2cbe3SMel Gorman 32924eb2b1dcSMel Gorman /* 32934eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 32944eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 32954eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 32964eb2b1dcSMel Gorman */ 329783cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 32986b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 329904f2cbe3SMel Gorman /* Do not unmap the current VMA */ 330004f2cbe3SMel Gorman if (iter_vma == vma) 330104f2cbe3SMel Gorman continue; 330204f2cbe3SMel Gorman 330304f2cbe3SMel Gorman /* 33042f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 33052f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 33062f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 33072f84a899SMel Gorman */ 33082f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 33092f84a899SMel Gorman continue; 33102f84a899SMel Gorman 33112f84a899SMel Gorman /* 331204f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 331304f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 331404f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 331504f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 331604f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 331704f2cbe3SMel Gorman */ 331804f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 331924669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 332024669e58SAneesh Kumar K.V address + huge_page_size(h), page); 332104f2cbe3SMel Gorman } 332283cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 332304f2cbe3SMel Gorman } 332404f2cbe3SMel Gorman 33250fe6e20bSNaoya Horiguchi /* 33260fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 3327ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 3328ef009b25SMichal Hocko * cannot race with other handlers or page migration. 3329ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 33300fe6e20bSNaoya Horiguchi */ 33311e8f889bSDavid Gibson static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 333204f2cbe3SMel Gorman unsigned long address, pte_t *ptep, pte_t pte, 3333cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 33341e8f889bSDavid Gibson { 3335a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 33361e8f889bSDavid Gibson struct page *old_page, *new_page; 3337ad4404a2SDavidlohr Bueso int ret = 0, outside_reserve = 0; 33382ec74c3eSSagi Grimberg unsigned long mmun_start; /* For mmu_notifiers */ 33392ec74c3eSSagi Grimberg unsigned long mmun_end; /* For mmu_notifiers */ 33401e8f889bSDavid Gibson 33411e8f889bSDavid Gibson old_page = pte_page(pte); 33421e8f889bSDavid Gibson 334304f2cbe3SMel Gorman retry_avoidcopy: 33441e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 33451e8f889bSDavid Gibson * and just make the page writable */ 334637a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 33470fe6e20bSNaoya Horiguchi page_move_anon_rmap(old_page, vma, address); 33481e8f889bSDavid Gibson set_huge_ptep_writable(vma, address, ptep); 334983c54070SNick Piggin return 0; 33501e8f889bSDavid Gibson } 33511e8f889bSDavid Gibson 335204f2cbe3SMel Gorman /* 335304f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 335404f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 335504f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 335604f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 335704f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 335804f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 335904f2cbe3SMel Gorman * of the full address range. 336004f2cbe3SMel Gorman */ 33615944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 336204f2cbe3SMel Gorman old_page != pagecache_page) 336304f2cbe3SMel Gorman outside_reserve = 1; 336404f2cbe3SMel Gorman 336509cbfeafSKirill A. Shutemov get_page(old_page); 3366b76c8cfbSLarry Woodman 3367ad4404a2SDavidlohr Bueso /* 3368ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 3369ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 3370ad4404a2SDavidlohr Bueso */ 3371cb900f41SKirill A. Shutemov spin_unlock(ptl); 337204f2cbe3SMel Gorman new_page = alloc_huge_page(vma, address, outside_reserve); 33731e8f889bSDavid Gibson 33742fc39cecSAdam Litke if (IS_ERR(new_page)) { 337504f2cbe3SMel Gorman /* 337604f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 337704f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 337804f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 337904f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 338004f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 338104f2cbe3SMel Gorman */ 338204f2cbe3SMel Gorman if (outside_reserve) { 338309cbfeafSKirill A. Shutemov put_page(old_page); 338404f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 33852f4612afSDavidlohr Bueso unmap_ref_private(mm, vma, old_page, address); 338604f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 3387cb900f41SKirill A. Shutemov spin_lock(ptl); 3388a734bcc8SHillf Danton ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 3389a9af0c5dSNaoya Horiguchi if (likely(ptep && 3390a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 339104f2cbe3SMel Gorman goto retry_avoidcopy; 3392a734bcc8SHillf Danton /* 3393cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 3394cb900f41SKirill A. Shutemov * lock, and our job is done. 3395a734bcc8SHillf Danton */ 3396a734bcc8SHillf Danton return 0; 339704f2cbe3SMel Gorman } 339804f2cbe3SMel Gorman 3399ad4404a2SDavidlohr Bueso ret = (PTR_ERR(new_page) == -ENOMEM) ? 3400ad4404a2SDavidlohr Bueso VM_FAULT_OOM : VM_FAULT_SIGBUS; 3401ad4404a2SDavidlohr Bueso goto out_release_old; 34021e8f889bSDavid Gibson } 34031e8f889bSDavid Gibson 34040fe6e20bSNaoya Horiguchi /* 34050fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 34060fe6e20bSNaoya Horiguchi * anon_vma prepared. 34070fe6e20bSNaoya Horiguchi */ 340844e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 3409ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 3410ad4404a2SDavidlohr Bueso goto out_release_all; 341144e2aa93SDean Nelson } 34120fe6e20bSNaoya Horiguchi 341347ad8475SAndrea Arcangeli copy_user_huge_page(new_page, old_page, address, vma, 341447ad8475SAndrea Arcangeli pages_per_huge_page(h)); 34150ed361deSNick Piggin __SetPageUptodate(new_page); 3416bcc54222SNaoya Horiguchi set_page_huge_active(new_page); 34171e8f889bSDavid Gibson 34182ec74c3eSSagi Grimberg mmun_start = address & huge_page_mask(h); 34192ec74c3eSSagi Grimberg mmun_end = mmun_start + huge_page_size(h); 34202ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 3421ad4404a2SDavidlohr Bueso 3422b76c8cfbSLarry Woodman /* 3423cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 3424b76c8cfbSLarry Woodman * before the page tables are altered 3425b76c8cfbSLarry Woodman */ 3426cb900f41SKirill A. Shutemov spin_lock(ptl); 3427a5516438SAndi Kleen ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 3428a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 342907443a85SJoonsoo Kim ClearPagePrivate(new_page); 343007443a85SJoonsoo Kim 34311e8f889bSDavid Gibson /* Break COW */ 34328fe627ecSGerald Schaefer huge_ptep_clear_flush(vma, address, ptep); 343334ee645eSJoerg Roedel mmu_notifier_invalidate_range(mm, mmun_start, mmun_end); 34341e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, 34351e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 3436d281ee61SKirill A. Shutemov page_remove_rmap(old_page, true); 3437cd67f0d2SNaoya Horiguchi hugepage_add_new_anon_rmap(new_page, vma, address); 34381e8f889bSDavid Gibson /* Make the old page be freed below */ 34391e8f889bSDavid Gibson new_page = old_page; 34401e8f889bSDavid Gibson } 3441cb900f41SKirill A. Shutemov spin_unlock(ptl); 34422ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 3443ad4404a2SDavidlohr Bueso out_release_all: 344409cbfeafSKirill A. Shutemov put_page(new_page); 3445ad4404a2SDavidlohr Bueso out_release_old: 344609cbfeafSKirill A. Shutemov put_page(old_page); 34478312034fSJoonsoo Kim 3448ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 3449ad4404a2SDavidlohr Bueso return ret; 34501e8f889bSDavid Gibson } 34511e8f889bSDavid Gibson 345204f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 3453a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 3454a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 345504f2cbe3SMel Gorman { 345604f2cbe3SMel Gorman struct address_space *mapping; 3457e7c4b0bfSAndy Whitcroft pgoff_t idx; 345804f2cbe3SMel Gorman 345904f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 3460a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 346104f2cbe3SMel Gorman 346204f2cbe3SMel Gorman return find_lock_page(mapping, idx); 346304f2cbe3SMel Gorman } 346404f2cbe3SMel Gorman 34653ae77f43SHugh Dickins /* 34663ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 34673ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 34683ae77f43SHugh Dickins */ 34693ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 34702a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 34712a15efc9SHugh Dickins { 34722a15efc9SHugh Dickins struct address_space *mapping; 34732a15efc9SHugh Dickins pgoff_t idx; 34742a15efc9SHugh Dickins struct page *page; 34752a15efc9SHugh Dickins 34762a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 34772a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 34782a15efc9SHugh Dickins 34792a15efc9SHugh Dickins page = find_get_page(mapping, idx); 34802a15efc9SHugh Dickins if (page) 34812a15efc9SHugh Dickins put_page(page); 34822a15efc9SHugh Dickins return page != NULL; 34832a15efc9SHugh Dickins } 34842a15efc9SHugh Dickins 3485ab76ad54SMike Kravetz int huge_add_to_page_cache(struct page *page, struct address_space *mapping, 3486ab76ad54SMike Kravetz pgoff_t idx) 3487ab76ad54SMike Kravetz { 3488ab76ad54SMike Kravetz struct inode *inode = mapping->host; 3489ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 3490ab76ad54SMike Kravetz int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 3491ab76ad54SMike Kravetz 3492ab76ad54SMike Kravetz if (err) 3493ab76ad54SMike Kravetz return err; 3494ab76ad54SMike Kravetz ClearPagePrivate(page); 3495ab76ad54SMike Kravetz 3496ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 3497ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 3498ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 3499ab76ad54SMike Kravetz return 0; 3500ab76ad54SMike Kravetz } 3501ab76ad54SMike Kravetz 3502a1ed3ddaSRobert P. J. Day static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, 35038382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 3504788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 3505ac9b9c66SHugh Dickins { 3506a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3507ac9b9c66SHugh Dickins int ret = VM_FAULT_SIGBUS; 3508409eb8c2SHillf Danton int anon_rmap = 0; 35094c887265SAdam Litke unsigned long size; 35104c887265SAdam Litke struct page *page; 35111e8f889bSDavid Gibson pte_t new_pte; 3512cb900f41SKirill A. Shutemov spinlock_t *ptl; 35134c887265SAdam Litke 351404f2cbe3SMel Gorman /* 351504f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 351604f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 351725985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 351804f2cbe3SMel Gorman */ 351904f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 3520910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 352104f2cbe3SMel Gorman current->pid); 352204f2cbe3SMel Gorman return ret; 352304f2cbe3SMel Gorman } 352404f2cbe3SMel Gorman 35254c887265SAdam Litke /* 35264c887265SAdam Litke * Use page lock to guard against racing truncation 35274c887265SAdam Litke * before we get page_table_lock. 35284c887265SAdam Litke */ 35296bda666aSChristoph Lameter retry: 35306bda666aSChristoph Lameter page = find_lock_page(mapping, idx); 35316bda666aSChristoph Lameter if (!page) { 3532a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 3533ebed4bfcSHugh Dickins if (idx >= size) 3534ebed4bfcSHugh Dickins goto out; 353504f2cbe3SMel Gorman page = alloc_huge_page(vma, address, 0); 35362fc39cecSAdam Litke if (IS_ERR(page)) { 353776dcee75SAneesh Kumar K.V ret = PTR_ERR(page); 353876dcee75SAneesh Kumar K.V if (ret == -ENOMEM) 353976dcee75SAneesh Kumar K.V ret = VM_FAULT_OOM; 354076dcee75SAneesh Kumar K.V else 354176dcee75SAneesh Kumar K.V ret = VM_FAULT_SIGBUS; 35426bda666aSChristoph Lameter goto out; 35436bda666aSChristoph Lameter } 354447ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 35450ed361deSNick Piggin __SetPageUptodate(page); 3546bcc54222SNaoya Horiguchi set_page_huge_active(page); 3547ac9b9c66SHugh Dickins 3548f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 3549ab76ad54SMike Kravetz int err = huge_add_to_page_cache(page, mapping, idx); 35506bda666aSChristoph Lameter if (err) { 35516bda666aSChristoph Lameter put_page(page); 35526bda666aSChristoph Lameter if (err == -EEXIST) 35536bda666aSChristoph Lameter goto retry; 35546bda666aSChristoph Lameter goto out; 35556bda666aSChristoph Lameter } 355623be7468SMel Gorman } else { 35576bda666aSChristoph Lameter lock_page(page); 35580fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 35590fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 35600fe6e20bSNaoya Horiguchi goto backout_unlocked; 356123be7468SMel Gorman } 3562409eb8c2SHillf Danton anon_rmap = 1; 35630fe6e20bSNaoya Horiguchi } 35640fe6e20bSNaoya Horiguchi } else { 356557303d80SAndy Whitcroft /* 3566998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 3567998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 3568998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 3569fd6a03edSNaoya Horiguchi */ 3570fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 3571aa50d3a7SAndi Kleen ret = VM_FAULT_HWPOISON | 3572972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 3573fd6a03edSNaoya Horiguchi goto backout_unlocked; 35746bda666aSChristoph Lameter } 3575998b4382SNaoya Horiguchi } 35761e8f889bSDavid Gibson 357757303d80SAndy Whitcroft /* 357857303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 357957303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 358057303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 358157303d80SAndy Whitcroft * the spinlock. 358257303d80SAndy Whitcroft */ 35835e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 35842b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 35852b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 35862b26736cSAndy Whitcroft goto backout_unlocked; 35872b26736cSAndy Whitcroft } 35885e911373SMike Kravetz /* Just decrements count, does not deallocate */ 3589feba16e2SMike Kravetz vma_end_reservation(h, vma, address); 35905e911373SMike Kravetz } 359157303d80SAndy Whitcroft 3592cb900f41SKirill A. Shutemov ptl = huge_pte_lockptr(h, mm, ptep); 3593cb900f41SKirill A. Shutemov spin_lock(ptl); 3594a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 35954c887265SAdam Litke if (idx >= size) 35964c887265SAdam Litke goto backout; 35974c887265SAdam Litke 359883c54070SNick Piggin ret = 0; 35997f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 36004c887265SAdam Litke goto backout; 36014c887265SAdam Litke 360207443a85SJoonsoo Kim if (anon_rmap) { 360307443a85SJoonsoo Kim ClearPagePrivate(page); 3604409eb8c2SHillf Danton hugepage_add_new_anon_rmap(page, vma, address); 3605ac714904SChoi Gi-yong } else 360653f9263bSKirill A. Shutemov page_dup_rmap(page, true); 36071e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 36081e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 36091e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, new_pte); 36101e8f889bSDavid Gibson 36115d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 3612788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 36131e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 3614cb900f41SKirill A. Shutemov ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl); 36151e8f889bSDavid Gibson } 36161e8f889bSDavid Gibson 3617cb900f41SKirill A. Shutemov spin_unlock(ptl); 36184c887265SAdam Litke unlock_page(page); 36194c887265SAdam Litke out: 3620ac9b9c66SHugh Dickins return ret; 36214c887265SAdam Litke 36224c887265SAdam Litke backout: 3623cb900f41SKirill A. Shutemov spin_unlock(ptl); 36242b26736cSAndy Whitcroft backout_unlocked: 36254c887265SAdam Litke unlock_page(page); 36264c887265SAdam Litke put_page(page); 36274c887265SAdam Litke goto out; 3628ac9b9c66SHugh Dickins } 3629ac9b9c66SHugh Dickins 36308382d914SDavidlohr Bueso #ifdef CONFIG_SMP 3631c672c7f2SMike Kravetz u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, 36328382d914SDavidlohr Bueso struct vm_area_struct *vma, 36338382d914SDavidlohr Bueso struct address_space *mapping, 36348382d914SDavidlohr Bueso pgoff_t idx, unsigned long address) 36358382d914SDavidlohr Bueso { 36368382d914SDavidlohr Bueso unsigned long key[2]; 36378382d914SDavidlohr Bueso u32 hash; 36388382d914SDavidlohr Bueso 36398382d914SDavidlohr Bueso if (vma->vm_flags & VM_SHARED) { 36408382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 36418382d914SDavidlohr Bueso key[1] = idx; 36428382d914SDavidlohr Bueso } else { 36438382d914SDavidlohr Bueso key[0] = (unsigned long) mm; 36448382d914SDavidlohr Bueso key[1] = address >> huge_page_shift(h); 36458382d914SDavidlohr Bueso } 36468382d914SDavidlohr Bueso 36478382d914SDavidlohr Bueso hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0); 36488382d914SDavidlohr Bueso 36498382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 36508382d914SDavidlohr Bueso } 36518382d914SDavidlohr Bueso #else 36528382d914SDavidlohr Bueso /* 36538382d914SDavidlohr Bueso * For uniprocesor systems we always use a single mutex, so just 36548382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 36558382d914SDavidlohr Bueso */ 3656c672c7f2SMike Kravetz u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, 36578382d914SDavidlohr Bueso struct vm_area_struct *vma, 36588382d914SDavidlohr Bueso struct address_space *mapping, 36598382d914SDavidlohr Bueso pgoff_t idx, unsigned long address) 36608382d914SDavidlohr Bueso { 36618382d914SDavidlohr Bueso return 0; 36628382d914SDavidlohr Bueso } 36638382d914SDavidlohr Bueso #endif 36648382d914SDavidlohr Bueso 366586e5216fSAdam Litke int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 3666788c7df4SHugh Dickins unsigned long address, unsigned int flags) 366786e5216fSAdam Litke { 36688382d914SDavidlohr Bueso pte_t *ptep, entry; 3669cb900f41SKirill A. Shutemov spinlock_t *ptl; 36701e8f889bSDavid Gibson int ret; 36718382d914SDavidlohr Bueso u32 hash; 36728382d914SDavidlohr Bueso pgoff_t idx; 36730fe6e20bSNaoya Horiguchi struct page *page = NULL; 367457303d80SAndy Whitcroft struct page *pagecache_page = NULL; 3675a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 36768382d914SDavidlohr Bueso struct address_space *mapping; 36770f792cf9SNaoya Horiguchi int need_wait_lock = 0; 367886e5216fSAdam Litke 36791e16a539SKAMEZAWA Hiroyuki address &= huge_page_mask(h); 36801e16a539SKAMEZAWA Hiroyuki 3681fd6a03edSNaoya Horiguchi ptep = huge_pte_offset(mm, address); 3682fd6a03edSNaoya Horiguchi if (ptep) { 3683fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 3684290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 3685cb900f41SKirill A. Shutemov migration_entry_wait_huge(vma, mm, ptep); 3686290408d4SNaoya Horiguchi return 0; 3687290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 3688aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 3689972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 36900d777df5SNaoya Horiguchi } else { 3691a5516438SAndi Kleen ptep = huge_pte_alloc(mm, address, huge_page_size(h)); 369286e5216fSAdam Litke if (!ptep) 369386e5216fSAdam Litke return VM_FAULT_OOM; 36940d777df5SNaoya Horiguchi } 369586e5216fSAdam Litke 36968382d914SDavidlohr Bueso mapping = vma->vm_file->f_mapping; 36978382d914SDavidlohr Bueso idx = vma_hugecache_offset(h, vma, address); 36988382d914SDavidlohr Bueso 36993935baa9SDavid Gibson /* 37003935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 37013935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 37023935baa9SDavid Gibson * the same page in the page cache. 37033935baa9SDavid Gibson */ 3704c672c7f2SMike Kravetz hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address); 3705c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 37068382d914SDavidlohr Bueso 37077f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 37087f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 37098382d914SDavidlohr Bueso ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 3710b4d1d99fSDavid Gibson goto out_mutex; 37113935baa9SDavid Gibson } 371286e5216fSAdam Litke 371383c54070SNick Piggin ret = 0; 37141e8f889bSDavid Gibson 371557303d80SAndy Whitcroft /* 37160f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 37170f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 37180f792cf9SNaoya Horiguchi * a active hugepage in pagecache. This goto expects the 2nd page fault, 37190f792cf9SNaoya Horiguchi * and is_hugetlb_entry_(migration|hwpoisoned) check will properly 37200f792cf9SNaoya Horiguchi * handle it. 37210f792cf9SNaoya Horiguchi */ 37220f792cf9SNaoya Horiguchi if (!pte_present(entry)) 37230f792cf9SNaoya Horiguchi goto out_mutex; 37240f792cf9SNaoya Horiguchi 37250f792cf9SNaoya Horiguchi /* 372657303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 372757303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 372857303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 372957303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 373057303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 373157303d80SAndy Whitcroft * consumed. 373257303d80SAndy Whitcroft */ 3733106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 37342b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 37352b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 3736b4d1d99fSDavid Gibson goto out_mutex; 37372b26736cSAndy Whitcroft } 37385e911373SMike Kravetz /* Just decrements count, does not deallocate */ 3739feba16e2SMike Kravetz vma_end_reservation(h, vma, address); 374057303d80SAndy Whitcroft 3741f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 374257303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 374357303d80SAndy Whitcroft vma, address); 374457303d80SAndy Whitcroft } 374557303d80SAndy Whitcroft 37460f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 37470fe6e20bSNaoya Horiguchi 37481e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 3749b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 3750cb900f41SKirill A. Shutemov goto out_ptl; 3751b4d1d99fSDavid Gibson 37520f792cf9SNaoya Horiguchi /* 37530f792cf9SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 37540f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 37550f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 37560f792cf9SNaoya Horiguchi */ 37570f792cf9SNaoya Horiguchi page = pte_page(entry); 37580f792cf9SNaoya Horiguchi if (page != pagecache_page) 37590f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 37600f792cf9SNaoya Horiguchi need_wait_lock = 1; 37610f792cf9SNaoya Horiguchi goto out_ptl; 37620f792cf9SNaoya Horiguchi } 37630f792cf9SNaoya Horiguchi 37640f792cf9SNaoya Horiguchi get_page(page); 3765b4d1d99fSDavid Gibson 3766788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 3767106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 376857303d80SAndy Whitcroft ret = hugetlb_cow(mm, vma, address, ptep, entry, 3769cb900f41SKirill A. Shutemov pagecache_page, ptl); 37700f792cf9SNaoya Horiguchi goto out_put_page; 3771b4d1d99fSDavid Gibson } 3772106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 3773b4d1d99fSDavid Gibson } 3774b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 3775788c7df4SHugh Dickins if (huge_ptep_set_access_flags(vma, address, ptep, entry, 3776788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 37774b3073e1SRussell King update_mmu_cache(vma, address, ptep); 37780f792cf9SNaoya Horiguchi out_put_page: 37790f792cf9SNaoya Horiguchi if (page != pagecache_page) 37800f792cf9SNaoya Horiguchi unlock_page(page); 37810f792cf9SNaoya Horiguchi put_page(page); 3782cb900f41SKirill A. Shutemov out_ptl: 3783cb900f41SKirill A. Shutemov spin_unlock(ptl); 378457303d80SAndy Whitcroft 378557303d80SAndy Whitcroft if (pagecache_page) { 378657303d80SAndy Whitcroft unlock_page(pagecache_page); 378757303d80SAndy Whitcroft put_page(pagecache_page); 378857303d80SAndy Whitcroft } 3789b4d1d99fSDavid Gibson out_mutex: 3790c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 37910f792cf9SNaoya Horiguchi /* 37920f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 37930f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 37940f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 37950f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 37960f792cf9SNaoya Horiguchi * here without taking refcount. 37970f792cf9SNaoya Horiguchi */ 37980f792cf9SNaoya Horiguchi if (need_wait_lock) 37990f792cf9SNaoya Horiguchi wait_on_page_locked(page); 38001e8f889bSDavid Gibson return ret; 380186e5216fSAdam Litke } 380286e5216fSAdam Litke 380328a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 380463551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 380528a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 380628a35716SMichel Lespinasse long i, unsigned int flags) 380763551ae0SDavid Gibson { 3808d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 3809d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 381028a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 3811a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 381263551ae0SDavid Gibson 381363551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 381463551ae0SDavid Gibson pte_t *pte; 3815cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 38162a15efc9SHugh Dickins int absent; 381763551ae0SDavid Gibson struct page *page; 381863551ae0SDavid Gibson 38194c887265SAdam Litke /* 382002057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 382102057967SDavid Rientjes * potentially allocating memory. 382202057967SDavid Rientjes */ 382302057967SDavid Rientjes if (unlikely(fatal_signal_pending(current))) { 382402057967SDavid Rientjes remainder = 0; 382502057967SDavid Rientjes break; 382602057967SDavid Rientjes } 382702057967SDavid Rientjes 382802057967SDavid Rientjes /* 38294c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 38302a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 38314c887265SAdam Litke * first, for the page indexing below to work. 3832cb900f41SKirill A. Shutemov * 3833cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 38344c887265SAdam Litke */ 3835a5516438SAndi Kleen pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); 3836cb900f41SKirill A. Shutemov if (pte) 3837cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 38382a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 383963551ae0SDavid Gibson 38402a15efc9SHugh Dickins /* 38412a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 38423ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 38433ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 38443ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 38453ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 38462a15efc9SHugh Dickins */ 38473ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 38483ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 3849cb900f41SKirill A. Shutemov if (pte) 3850cb900f41SKirill A. Shutemov spin_unlock(ptl); 38512a15efc9SHugh Dickins remainder = 0; 38522a15efc9SHugh Dickins break; 38532a15efc9SHugh Dickins } 38542a15efc9SHugh Dickins 38559cc3a5bdSNaoya Horiguchi /* 38569cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 38579cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 38589cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 38599cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 38609cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 38619cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 38629cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 38639cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 38649cc3a5bdSNaoya Horiguchi */ 38659cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 3866106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 3867106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 38684c887265SAdam Litke int ret; 38694c887265SAdam Litke 3870cb900f41SKirill A. Shutemov if (pte) 3871cb900f41SKirill A. Shutemov spin_unlock(ptl); 38722a15efc9SHugh Dickins ret = hugetlb_fault(mm, vma, vaddr, 38732a15efc9SHugh Dickins (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); 3874a89182c7SAdam Litke if (!(ret & VM_FAULT_ERROR)) 38754c887265SAdam Litke continue; 38764c887265SAdam Litke 38771c59827dSHugh Dickins remainder = 0; 38781c59827dSHugh Dickins break; 38791c59827dSHugh Dickins } 388063551ae0SDavid Gibson 3881a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 38827f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 3883d5d4b0aaSChen, Kenneth W same_page: 3884d6692183SChen, Kenneth W if (pages) { 388569d177c2SAndy Whitcroft pages[i] = mem_map_offset(page, pfn_offset); 3886ddc58f27SKirill A. Shutemov get_page(pages[i]); 3887d6692183SChen, Kenneth W } 388863551ae0SDavid Gibson 388963551ae0SDavid Gibson if (vmas) 389063551ae0SDavid Gibson vmas[i] = vma; 389163551ae0SDavid Gibson 389263551ae0SDavid Gibson vaddr += PAGE_SIZE; 3893d5d4b0aaSChen, Kenneth W ++pfn_offset; 389463551ae0SDavid Gibson --remainder; 389563551ae0SDavid Gibson ++i; 3896d5d4b0aaSChen, Kenneth W if (vaddr < vma->vm_end && remainder && 3897a5516438SAndi Kleen pfn_offset < pages_per_huge_page(h)) { 3898d5d4b0aaSChen, Kenneth W /* 3899d5d4b0aaSChen, Kenneth W * We use pfn_offset to avoid touching the pageframes 3900d5d4b0aaSChen, Kenneth W * of this compound page. 3901d5d4b0aaSChen, Kenneth W */ 3902d5d4b0aaSChen, Kenneth W goto same_page; 3903d5d4b0aaSChen, Kenneth W } 3904cb900f41SKirill A. Shutemov spin_unlock(ptl); 390563551ae0SDavid Gibson } 390628a35716SMichel Lespinasse *nr_pages = remainder; 390763551ae0SDavid Gibson *position = vaddr; 390863551ae0SDavid Gibson 39092a15efc9SHugh Dickins return i ? i : -EFAULT; 391063551ae0SDavid Gibson } 39118f860591SZhang, Yanmin 39127da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 39138f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 39148f860591SZhang, Yanmin { 39158f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 39168f860591SZhang, Yanmin unsigned long start = address; 39178f860591SZhang, Yanmin pte_t *ptep; 39188f860591SZhang, Yanmin pte_t pte; 3919a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 39207da4d641SPeter Zijlstra unsigned long pages = 0; 39218f860591SZhang, Yanmin 39228f860591SZhang, Yanmin BUG_ON(address >= end); 39238f860591SZhang, Yanmin flush_cache_range(vma, address, end); 39248f860591SZhang, Yanmin 3925a5338093SRik van Riel mmu_notifier_invalidate_range_start(mm, start, end); 392683cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 3927a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 3928cb900f41SKirill A. Shutemov spinlock_t *ptl; 39298f860591SZhang, Yanmin ptep = huge_pte_offset(mm, address); 39308f860591SZhang, Yanmin if (!ptep) 39318f860591SZhang, Yanmin continue; 3932cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 39337da4d641SPeter Zijlstra if (huge_pmd_unshare(mm, &address, ptep)) { 39347da4d641SPeter Zijlstra pages++; 3935cb900f41SKirill A. Shutemov spin_unlock(ptl); 393639dde65cSChen, Kenneth W continue; 39377da4d641SPeter Zijlstra } 3938a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 3939a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 3940a8bda28dSNaoya Horiguchi spin_unlock(ptl); 3941a8bda28dSNaoya Horiguchi continue; 3942a8bda28dSNaoya Horiguchi } 3943a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 3944a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 3945a8bda28dSNaoya Horiguchi 3946a8bda28dSNaoya Horiguchi if (is_write_migration_entry(entry)) { 3947a8bda28dSNaoya Horiguchi pte_t newpte; 3948a8bda28dSNaoya Horiguchi 3949a8bda28dSNaoya Horiguchi make_migration_entry_read(&entry); 3950a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 3951a8bda28dSNaoya Horiguchi set_huge_pte_at(mm, address, ptep, newpte); 3952a8bda28dSNaoya Horiguchi pages++; 3953a8bda28dSNaoya Horiguchi } 3954a8bda28dSNaoya Horiguchi spin_unlock(ptl); 3955a8bda28dSNaoya Horiguchi continue; 3956a8bda28dSNaoya Horiguchi } 3957a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 39588f860591SZhang, Yanmin pte = huge_ptep_get_and_clear(mm, address, ptep); 3959106c992aSGerald Schaefer pte = pte_mkhuge(huge_pte_modify(pte, newprot)); 3960be7517d6STony Lu pte = arch_make_huge_pte(pte, vma, NULL, 0); 39618f860591SZhang, Yanmin set_huge_pte_at(mm, address, ptep, pte); 39627da4d641SPeter Zijlstra pages++; 39638f860591SZhang, Yanmin } 3964cb900f41SKirill A. Shutemov spin_unlock(ptl); 39658f860591SZhang, Yanmin } 3966d833352aSMel Gorman /* 3967c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 3968d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 3969c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 3970d833352aSMel Gorman * and that page table be reused and filled with junk. 3971d833352aSMel Gorman */ 39728f860591SZhang, Yanmin flush_tlb_range(vma, start, end); 397334ee645eSJoerg Roedel mmu_notifier_invalidate_range(mm, start, end); 397483cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 3975a5338093SRik van Riel mmu_notifier_invalidate_range_end(mm, start, end); 39767da4d641SPeter Zijlstra 39777da4d641SPeter Zijlstra return pages << h->order; 39788f860591SZhang, Yanmin } 39798f860591SZhang, Yanmin 3980a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 3981a1e78772SMel Gorman long from, long to, 39825a6fe125SMel Gorman struct vm_area_struct *vma, 3983ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 3984e4e574b7SAdam Litke { 398517c9d12eSMel Gorman long ret, chg; 3986a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 398790481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 39889119a41eSJoonsoo Kim struct resv_map *resv_map; 39891c5ecae3SMike Kravetz long gbl_reserve; 3990e4e574b7SAdam Litke 3991a1e78772SMel Gorman /* 399217c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 399317c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 399490481622SDavid Gibson * without using reserves 399517c9d12eSMel Gorman */ 3996ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 399717c9d12eSMel Gorman return 0; 399817c9d12eSMel Gorman 399917c9d12eSMel Gorman /* 4000a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 4001a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 4002a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 4003a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 4004a1e78772SMel Gorman */ 40059119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 40064e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 40079119a41eSJoonsoo Kim 40081406ec9bSJoonsoo Kim chg = region_chg(resv_map, from, to); 40099119a41eSJoonsoo Kim 40109119a41eSJoonsoo Kim } else { 40119119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 40125a6fe125SMel Gorman if (!resv_map) 40135a6fe125SMel Gorman return -ENOMEM; 40145a6fe125SMel Gorman 401517c9d12eSMel Gorman chg = to - from; 401617c9d12eSMel Gorman 40175a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 40185a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 40195a6fe125SMel Gorman } 40205a6fe125SMel Gorman 4021c50ac050SDave Hansen if (chg < 0) { 4022c50ac050SDave Hansen ret = chg; 4023c50ac050SDave Hansen goto out_err; 4024c50ac050SDave Hansen } 402517c9d12eSMel Gorman 40261c5ecae3SMike Kravetz /* 40271c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 40281c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 40291c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 40301c5ecae3SMike Kravetz */ 40311c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 40321c5ecae3SMike Kravetz if (gbl_reserve < 0) { 4033c50ac050SDave Hansen ret = -ENOSPC; 4034c50ac050SDave Hansen goto out_err; 4035c50ac050SDave Hansen } 403617c9d12eSMel Gorman 403717c9d12eSMel Gorman /* 403817c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 403990481622SDavid Gibson * Hand the pages back to the subpool if there are not 404017c9d12eSMel Gorman */ 40411c5ecae3SMike Kravetz ret = hugetlb_acct_memory(h, gbl_reserve); 404217c9d12eSMel Gorman if (ret < 0) { 40431c5ecae3SMike Kravetz /* put back original number of pages, chg */ 40441c5ecae3SMike Kravetz (void)hugepage_subpool_put_pages(spool, chg); 4045c50ac050SDave Hansen goto out_err; 404617c9d12eSMel Gorman } 404717c9d12eSMel Gorman 404817c9d12eSMel Gorman /* 404917c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 405017c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 405117c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 405217c9d12eSMel Gorman * the reservation was and the page cache tells how much of 405317c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 405417c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 405517c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 405617c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 405717c9d12eSMel Gorman * else has to be done for private mappings here 405817c9d12eSMel Gorman */ 405933039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 406033039678SMike Kravetz long add = region_add(resv_map, from, to); 406133039678SMike Kravetz 406233039678SMike Kravetz if (unlikely(chg > add)) { 406333039678SMike Kravetz /* 406433039678SMike Kravetz * pages in this range were added to the reserve 406533039678SMike Kravetz * map between region_chg and region_add. This 406633039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 406733039678SMike Kravetz * the subpool and reserve counts modified above 406833039678SMike Kravetz * based on the difference. 406933039678SMike Kravetz */ 407033039678SMike Kravetz long rsv_adjust; 407133039678SMike Kravetz 407233039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 407333039678SMike Kravetz chg - add); 407433039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 407533039678SMike Kravetz } 407633039678SMike Kravetz } 4077a43a8c39SChen, Kenneth W return 0; 4078c50ac050SDave Hansen out_err: 40795e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 40805e911373SMike Kravetz region_abort(resv_map, from, to); 4081f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 4082f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 4083c50ac050SDave Hansen return ret; 4084a43a8c39SChen, Kenneth W } 4085a43a8c39SChen, Kenneth W 4086b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 4087b5cec28dSMike Kravetz long freed) 4088a43a8c39SChen, Kenneth W { 4089a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 40904e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 40919119a41eSJoonsoo Kim long chg = 0; 409290481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 40931c5ecae3SMike Kravetz long gbl_reserve; 409445c682a6SKen Chen 4095b5cec28dSMike Kravetz if (resv_map) { 4096b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 4097b5cec28dSMike Kravetz /* 4098b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 4099b5cec28dSMike Kravetz * must be split and another region descriptor can not be 4100b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 4101b5cec28dSMike Kravetz */ 4102b5cec28dSMike Kravetz if (chg < 0) 4103b5cec28dSMike Kravetz return chg; 4104b5cec28dSMike Kravetz } 4105b5cec28dSMike Kravetz 410645c682a6SKen Chen spin_lock(&inode->i_lock); 4107e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 410845c682a6SKen Chen spin_unlock(&inode->i_lock); 410945c682a6SKen Chen 41101c5ecae3SMike Kravetz /* 41111c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 41121c5ecae3SMike Kravetz * reservations to be released may be adjusted. 41131c5ecae3SMike Kravetz */ 41141c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 41151c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 4116b5cec28dSMike Kravetz 4117b5cec28dSMike Kravetz return 0; 4118a43a8c39SChen, Kenneth W } 411993f70f90SNaoya Horiguchi 41203212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 41213212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 41223212b535SSteve Capper struct vm_area_struct *vma, 41233212b535SSteve Capper unsigned long addr, pgoff_t idx) 41243212b535SSteve Capper { 41253212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 41263212b535SSteve Capper svma->vm_start; 41273212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 41283212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 41293212b535SSteve Capper 41303212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 4131de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 4132de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 41333212b535SSteve Capper 41343212b535SSteve Capper /* 41353212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 41363212b535SSteve Capper * page table page. 41373212b535SSteve Capper */ 41383212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 41393212b535SSteve Capper vm_flags != svm_flags || 41403212b535SSteve Capper sbase < svma->vm_start || svma->vm_end < s_end) 41413212b535SSteve Capper return 0; 41423212b535SSteve Capper 41433212b535SSteve Capper return saddr; 41443212b535SSteve Capper } 41453212b535SSteve Capper 414631aafb45SNicholas Krause static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) 41473212b535SSteve Capper { 41483212b535SSteve Capper unsigned long base = addr & PUD_MASK; 41493212b535SSteve Capper unsigned long end = base + PUD_SIZE; 41503212b535SSteve Capper 41513212b535SSteve Capper /* 41523212b535SSteve Capper * check on proper vm_flags and page table alignment 41533212b535SSteve Capper */ 41543212b535SSteve Capper if (vma->vm_flags & VM_MAYSHARE && 41553212b535SSteve Capper vma->vm_start <= base && end <= vma->vm_end) 415631aafb45SNicholas Krause return true; 415731aafb45SNicholas Krause return false; 41583212b535SSteve Capper } 41593212b535SSteve Capper 41603212b535SSteve Capper /* 41613212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 41623212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 41633212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 41643212b535SSteve Capper * code much cleaner. pmd allocation is essential for the shared case because 4165c8c06efaSDavidlohr Bueso * pud has to be populated inside the same i_mmap_rwsem section - otherwise 41663212b535SSteve Capper * racing tasks could either miss the sharing (see huge_pte_offset) or select a 41673212b535SSteve Capper * bad pmd for sharing. 41683212b535SSteve Capper */ 41693212b535SSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 41703212b535SSteve Capper { 41713212b535SSteve Capper struct vm_area_struct *vma = find_vma(mm, addr); 41723212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 41733212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 41743212b535SSteve Capper vma->vm_pgoff; 41753212b535SSteve Capper struct vm_area_struct *svma; 41763212b535SSteve Capper unsigned long saddr; 41773212b535SSteve Capper pte_t *spte = NULL; 41783212b535SSteve Capper pte_t *pte; 4179cb900f41SKirill A. Shutemov spinlock_t *ptl; 41803212b535SSteve Capper 41813212b535SSteve Capper if (!vma_shareable(vma, addr)) 41823212b535SSteve Capper return (pte_t *)pmd_alloc(mm, pud, addr); 41833212b535SSteve Capper 418483cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 41853212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 41863212b535SSteve Capper if (svma == vma) 41873212b535SSteve Capper continue; 41883212b535SSteve Capper 41893212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 41903212b535SSteve Capper if (saddr) { 41913212b535SSteve Capper spte = huge_pte_offset(svma->vm_mm, saddr); 41923212b535SSteve Capper if (spte) { 4193dc6c9a35SKirill A. Shutemov mm_inc_nr_pmds(mm); 41943212b535SSteve Capper get_page(virt_to_page(spte)); 41953212b535SSteve Capper break; 41963212b535SSteve Capper } 41973212b535SSteve Capper } 41983212b535SSteve Capper } 41993212b535SSteve Capper 42003212b535SSteve Capper if (!spte) 42013212b535SSteve Capper goto out; 42023212b535SSteve Capper 4203cb900f41SKirill A. Shutemov ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte); 4204cb900f41SKirill A. Shutemov spin_lock(ptl); 4205dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 42063212b535SSteve Capper pud_populate(mm, pud, 42073212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 4208dc6c9a35SKirill A. Shutemov } else { 42093212b535SSteve Capper put_page(virt_to_page(spte)); 4210dc6c9a35SKirill A. Shutemov mm_inc_nr_pmds(mm); 4211dc6c9a35SKirill A. Shutemov } 4212cb900f41SKirill A. Shutemov spin_unlock(ptl); 42133212b535SSteve Capper out: 42143212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 421583cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 42163212b535SSteve Capper return pte; 42173212b535SSteve Capper } 42183212b535SSteve Capper 42193212b535SSteve Capper /* 42203212b535SSteve Capper * unmap huge page backed by shared pte. 42213212b535SSteve Capper * 42223212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 42233212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 42243212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 42253212b535SSteve Capper * 4226cb900f41SKirill A. Shutemov * called with page table lock held. 42273212b535SSteve Capper * 42283212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 42293212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 42303212b535SSteve Capper */ 42313212b535SSteve Capper int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 42323212b535SSteve Capper { 42333212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 42343212b535SSteve Capper pud_t *pud = pud_offset(pgd, *addr); 42353212b535SSteve Capper 42363212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 42373212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 42383212b535SSteve Capper return 0; 42393212b535SSteve Capper 42403212b535SSteve Capper pud_clear(pud); 42413212b535SSteve Capper put_page(virt_to_page(ptep)); 4242dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 42433212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 42443212b535SSteve Capper return 1; 42453212b535SSteve Capper } 42469e5fc74cSSteve Capper #define want_pmd_share() (1) 42479e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 42489e5fc74cSSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 42499e5fc74cSSteve Capper { 42509e5fc74cSSteve Capper return NULL; 42519e5fc74cSSteve Capper } 4252e81f2d22SZhang Zhen 4253e81f2d22SZhang Zhen int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 4254e81f2d22SZhang Zhen { 4255e81f2d22SZhang Zhen return 0; 4256e81f2d22SZhang Zhen } 42579e5fc74cSSteve Capper #define want_pmd_share() (0) 42583212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 42593212b535SSteve Capper 42609e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 42619e5fc74cSSteve Capper pte_t *huge_pte_alloc(struct mm_struct *mm, 42629e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 42639e5fc74cSSteve Capper { 42649e5fc74cSSteve Capper pgd_t *pgd; 42659e5fc74cSSteve Capper pud_t *pud; 42669e5fc74cSSteve Capper pte_t *pte = NULL; 42679e5fc74cSSteve Capper 42689e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 42699e5fc74cSSteve Capper pud = pud_alloc(mm, pgd, addr); 42709e5fc74cSSteve Capper if (pud) { 42719e5fc74cSSteve Capper if (sz == PUD_SIZE) { 42729e5fc74cSSteve Capper pte = (pte_t *)pud; 42739e5fc74cSSteve Capper } else { 42749e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 42759e5fc74cSSteve Capper if (want_pmd_share() && pud_none(*pud)) 42769e5fc74cSSteve Capper pte = huge_pmd_share(mm, addr, pud); 42779e5fc74cSSteve Capper else 42789e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 42799e5fc74cSSteve Capper } 42809e5fc74cSSteve Capper } 42819e5fc74cSSteve Capper BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte)); 42829e5fc74cSSteve Capper 42839e5fc74cSSteve Capper return pte; 42849e5fc74cSSteve Capper } 42859e5fc74cSSteve Capper 42869e5fc74cSSteve Capper pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) 42879e5fc74cSSteve Capper { 42889e5fc74cSSteve Capper pgd_t *pgd; 42899e5fc74cSSteve Capper pud_t *pud; 42909e5fc74cSSteve Capper pmd_t *pmd = NULL; 42919e5fc74cSSteve Capper 42929e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 42939e5fc74cSSteve Capper if (pgd_present(*pgd)) { 42949e5fc74cSSteve Capper pud = pud_offset(pgd, addr); 42959e5fc74cSSteve Capper if (pud_present(*pud)) { 42969e5fc74cSSteve Capper if (pud_huge(*pud)) 42979e5fc74cSSteve Capper return (pte_t *)pud; 42989e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 42999e5fc74cSSteve Capper } 43009e5fc74cSSteve Capper } 43019e5fc74cSSteve Capper return (pte_t *) pmd; 43029e5fc74cSSteve Capper } 43039e5fc74cSSteve Capper 430461f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 430561f77edaSNaoya Horiguchi 430661f77edaSNaoya Horiguchi /* 430761f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 430861f77edaSNaoya Horiguchi * behavior. 430961f77edaSNaoya Horiguchi */ 431061f77edaSNaoya Horiguchi struct page * __weak 431161f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 431261f77edaSNaoya Horiguchi int write) 431361f77edaSNaoya Horiguchi { 431461f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 431561f77edaSNaoya Horiguchi } 431661f77edaSNaoya Horiguchi 431761f77edaSNaoya Horiguchi struct page * __weak 43189e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 4319e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 43209e5fc74cSSteve Capper { 4321e66f17ffSNaoya Horiguchi struct page *page = NULL; 4322e66f17ffSNaoya Horiguchi spinlock_t *ptl; 4323e66f17ffSNaoya Horiguchi retry: 4324e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 4325e66f17ffSNaoya Horiguchi spin_lock(ptl); 4326e66f17ffSNaoya Horiguchi /* 4327e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 4328e66f17ffSNaoya Horiguchi * unmapped from other threads. 4329e66f17ffSNaoya Horiguchi */ 4330e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 4331e66f17ffSNaoya Horiguchi goto out; 4332e66f17ffSNaoya Horiguchi if (pmd_present(*pmd)) { 433397534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 4334e66f17ffSNaoya Horiguchi if (flags & FOLL_GET) 4335e66f17ffSNaoya Horiguchi get_page(page); 4336e66f17ffSNaoya Horiguchi } else { 4337e66f17ffSNaoya Horiguchi if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) { 4338e66f17ffSNaoya Horiguchi spin_unlock(ptl); 4339e66f17ffSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pmd, ptl); 4340e66f17ffSNaoya Horiguchi goto retry; 4341e66f17ffSNaoya Horiguchi } 4342e66f17ffSNaoya Horiguchi /* 4343e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 4344e66f17ffSNaoya Horiguchi * follow_page_mask(). 4345e66f17ffSNaoya Horiguchi */ 4346e66f17ffSNaoya Horiguchi } 4347e66f17ffSNaoya Horiguchi out: 4348e66f17ffSNaoya Horiguchi spin_unlock(ptl); 43499e5fc74cSSteve Capper return page; 43509e5fc74cSSteve Capper } 43519e5fc74cSSteve Capper 435261f77edaSNaoya Horiguchi struct page * __weak 43539e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 4354e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 43559e5fc74cSSteve Capper { 4356e66f17ffSNaoya Horiguchi if (flags & FOLL_GET) 4357e66f17ffSNaoya Horiguchi return NULL; 43589e5fc74cSSteve Capper 4359e66f17ffSNaoya Horiguchi return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 43609e5fc74cSSteve Capper } 43619e5fc74cSSteve Capper 4362d5bd9106SAndi Kleen #ifdef CONFIG_MEMORY_FAILURE 4363d5bd9106SAndi Kleen 436493f70f90SNaoya Horiguchi /* 436593f70f90SNaoya Horiguchi * This function is called from memory failure code. 436693f70f90SNaoya Horiguchi * Assume the caller holds page lock of the head page. 436793f70f90SNaoya Horiguchi */ 43686de2b1aaSNaoya Horiguchi int dequeue_hwpoisoned_huge_page(struct page *hpage) 436993f70f90SNaoya Horiguchi { 437093f70f90SNaoya Horiguchi struct hstate *h = page_hstate(hpage); 437193f70f90SNaoya Horiguchi int nid = page_to_nid(hpage); 43726de2b1aaSNaoya Horiguchi int ret = -EBUSY; 437393f70f90SNaoya Horiguchi 437493f70f90SNaoya Horiguchi spin_lock(&hugetlb_lock); 43757e1f049eSNaoya Horiguchi /* 43767e1f049eSNaoya Horiguchi * Just checking !page_huge_active is not enough, because that could be 43777e1f049eSNaoya Horiguchi * an isolated/hwpoisoned hugepage (which have >0 refcount). 43787e1f049eSNaoya Horiguchi */ 43797e1f049eSNaoya Horiguchi if (!page_huge_active(hpage) && !page_count(hpage)) { 438056f2fb14SNaoya Horiguchi /* 438156f2fb14SNaoya Horiguchi * Hwpoisoned hugepage isn't linked to activelist or freelist, 438256f2fb14SNaoya Horiguchi * but dangling hpage->lru can trigger list-debug warnings 438356f2fb14SNaoya Horiguchi * (this happens when we call unpoison_memory() on it), 438456f2fb14SNaoya Horiguchi * so let it point to itself with list_del_init(). 438556f2fb14SNaoya Horiguchi */ 438656f2fb14SNaoya Horiguchi list_del_init(&hpage->lru); 43878c6c2ecbSNaoya Horiguchi set_page_refcounted(hpage); 438893f70f90SNaoya Horiguchi h->free_huge_pages--; 438993f70f90SNaoya Horiguchi h->free_huge_pages_node[nid]--; 43906de2b1aaSNaoya Horiguchi ret = 0; 439193f70f90SNaoya Horiguchi } 43926de2b1aaSNaoya Horiguchi spin_unlock(&hugetlb_lock); 43936de2b1aaSNaoya Horiguchi return ret; 43946de2b1aaSNaoya Horiguchi } 43956de2b1aaSNaoya Horiguchi #endif 439631caf665SNaoya Horiguchi 439731caf665SNaoya Horiguchi bool isolate_huge_page(struct page *page, struct list_head *list) 439831caf665SNaoya Horiguchi { 4399bcc54222SNaoya Horiguchi bool ret = true; 4400bcc54222SNaoya Horiguchi 4401309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 440231caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 4403bcc54222SNaoya Horiguchi if (!page_huge_active(page) || !get_page_unless_zero(page)) { 4404bcc54222SNaoya Horiguchi ret = false; 4405bcc54222SNaoya Horiguchi goto unlock; 4406bcc54222SNaoya Horiguchi } 4407bcc54222SNaoya Horiguchi clear_page_huge_active(page); 440831caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 4409bcc54222SNaoya Horiguchi unlock: 441031caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 4411bcc54222SNaoya Horiguchi return ret; 441231caf665SNaoya Horiguchi } 441331caf665SNaoya Horiguchi 441431caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 441531caf665SNaoya Horiguchi { 4416309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 441731caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 4418bcc54222SNaoya Horiguchi set_page_huge_active(page); 441931caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 442031caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 442131caf665SNaoya Horiguchi put_page(page); 442231caf665SNaoya Horiguchi } 4423