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 */ 83567961f9dSMike Kravetz if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 83667961f9dSMike Kravetz /* 83767961f9dSMike Kravetz * Like the shared case above, a hole punch or truncate 83867961f9dSMike Kravetz * could have been performed on the private mapping. 83967961f9dSMike Kravetz * Examine the value of chg to determine if reserves 84067961f9dSMike Kravetz * actually exist or were previously consumed. 84167961f9dSMike Kravetz * Very Subtle - The value of chg comes from a previous 84267961f9dSMike Kravetz * call to vma_needs_reserves(). The reserve map for 84367961f9dSMike Kravetz * private mappings has different (opposite) semantics 84467961f9dSMike Kravetz * than that of shared mappings. vma_needs_reserves() 84567961f9dSMike Kravetz * has already taken this difference in semantics into 84667961f9dSMike Kravetz * account. Therefore, the meaning of chg is the same 84767961f9dSMike Kravetz * as in the shared case above. Code could easily be 84867961f9dSMike Kravetz * combined, but keeping it separate draws attention to 84967961f9dSMike Kravetz * subtle differences. 85067961f9dSMike Kravetz */ 85167961f9dSMike Kravetz if (chg) 85267961f9dSMike Kravetz return false; 85367961f9dSMike Kravetz else 854559ec2f8SNicholas Krause return true; 85567961f9dSMike Kravetz } 856a63884e9SJoonsoo Kim 857559ec2f8SNicholas Krause return false; 858a1e78772SMel Gorman } 859a1e78772SMel Gorman 860a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 8611da177e4SLinus Torvalds { 8621da177e4SLinus Torvalds int nid = page_to_nid(page); 8630edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 864a5516438SAndi Kleen h->free_huge_pages++; 865a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 8661da177e4SLinus Torvalds } 8671da177e4SLinus Torvalds 868bf50bab2SNaoya Horiguchi static struct page *dequeue_huge_page_node(struct hstate *h, int nid) 869bf50bab2SNaoya Horiguchi { 870bf50bab2SNaoya Horiguchi struct page *page; 871bf50bab2SNaoya Horiguchi 872c8721bbbSNaoya Horiguchi list_for_each_entry(page, &h->hugepage_freelists[nid], lru) 873c8721bbbSNaoya Horiguchi if (!is_migrate_isolate_page(page)) 874c8721bbbSNaoya Horiguchi break; 875c8721bbbSNaoya Horiguchi /* 876c8721bbbSNaoya Horiguchi * if 'non-isolated free hugepage' not found on the list, 877c8721bbbSNaoya Horiguchi * the allocation fails. 878c8721bbbSNaoya Horiguchi */ 879c8721bbbSNaoya Horiguchi if (&h->hugepage_freelists[nid] == &page->lru) 880bf50bab2SNaoya Horiguchi return NULL; 8810edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 882a9869b83SNaoya Horiguchi set_page_refcounted(page); 883bf50bab2SNaoya Horiguchi h->free_huge_pages--; 884bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 885bf50bab2SNaoya Horiguchi return page; 886bf50bab2SNaoya Horiguchi } 887bf50bab2SNaoya Horiguchi 88886cdb465SNaoya Horiguchi /* Movability of hugepages depends on migration support. */ 88986cdb465SNaoya Horiguchi static inline gfp_t htlb_alloc_mask(struct hstate *h) 89086cdb465SNaoya Horiguchi { 891100873d7SNaoya Horiguchi if (hugepages_treat_as_movable || hugepage_migration_supported(h)) 89286cdb465SNaoya Horiguchi return GFP_HIGHUSER_MOVABLE; 89386cdb465SNaoya Horiguchi else 89486cdb465SNaoya Horiguchi return GFP_HIGHUSER; 89586cdb465SNaoya Horiguchi } 89686cdb465SNaoya Horiguchi 897a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 898a5516438SAndi Kleen struct vm_area_struct *vma, 899af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 900af0ed73eSJoonsoo Kim long chg) 9011da177e4SLinus Torvalds { 902b1c12cbcSKonstantin Khlebnikov struct page *page = NULL; 903480eccf9SLee Schermerhorn struct mempolicy *mpol; 90419770b32SMel Gorman nodemask_t *nodemask; 905c0ff7453SMiao Xie struct zonelist *zonelist; 906dd1a239fSMel Gorman struct zone *zone; 907dd1a239fSMel Gorman struct zoneref *z; 908cc9a6c87SMel Gorman unsigned int cpuset_mems_cookie; 9091da177e4SLinus Torvalds 910a1e78772SMel Gorman /* 911a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 912a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 913a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 914a1e78772SMel Gorman */ 915af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg) && 916a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 917c0ff7453SMiao Xie goto err; 918a1e78772SMel Gorman 91904f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 920a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 9216eab04a8SJustin P. Mattock goto err; 92204f2cbe3SMel Gorman 9239966c4bbSJoonsoo Kim retry_cpuset: 924d26914d1SMel Gorman cpuset_mems_cookie = read_mems_allowed_begin(); 9259966c4bbSJoonsoo Kim zonelist = huge_zonelist(vma, address, 92686cdb465SNaoya Horiguchi htlb_alloc_mask(h), &mpol, &nodemask); 9279966c4bbSJoonsoo Kim 92819770b32SMel Gorman for_each_zone_zonelist_nodemask(zone, z, zonelist, 92919770b32SMel Gorman MAX_NR_ZONES - 1, nodemask) { 930344736f2SVladimir Davydov if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) { 931bf50bab2SNaoya Horiguchi page = dequeue_huge_page_node(h, zone_to_nid(zone)); 932bf50bab2SNaoya Horiguchi if (page) { 933af0ed73eSJoonsoo Kim if (avoid_reserve) 934af0ed73eSJoonsoo Kim break; 935af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg)) 936af0ed73eSJoonsoo Kim break; 937af0ed73eSJoonsoo Kim 93807443a85SJoonsoo Kim SetPagePrivate(page); 939a63884e9SJoonsoo Kim h->resv_huge_pages--; 9405ab3ee7bSKen Chen break; 9411da177e4SLinus Torvalds } 9423abf7afdSAndrew Morton } 943bf50bab2SNaoya Horiguchi } 944cc9a6c87SMel Gorman 945cc9a6c87SMel Gorman mpol_cond_put(mpol); 946d26914d1SMel Gorman if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) 947cc9a6c87SMel Gorman goto retry_cpuset; 948cc9a6c87SMel Gorman return page; 949cc9a6c87SMel Gorman 950c0ff7453SMiao Xie err: 951cc9a6c87SMel Gorman return NULL; 9521da177e4SLinus Torvalds } 9531da177e4SLinus Torvalds 9541cac6f2cSLuiz Capitulino /* 9551cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 9561cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 9571cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 9581cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 9591cac6f2cSLuiz Capitulino * node for alloc or free. 9601cac6f2cSLuiz Capitulino */ 9611cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 9621cac6f2cSLuiz Capitulino { 9630edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 9641cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 9651cac6f2cSLuiz Capitulino 9661cac6f2cSLuiz Capitulino return nid; 9671cac6f2cSLuiz Capitulino } 9681cac6f2cSLuiz Capitulino 9691cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 9701cac6f2cSLuiz Capitulino { 9711cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 9721cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 9731cac6f2cSLuiz Capitulino return nid; 9741cac6f2cSLuiz Capitulino } 9751cac6f2cSLuiz Capitulino 9761cac6f2cSLuiz Capitulino /* 9771cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 9781cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 9791cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 9801cac6f2cSLuiz Capitulino * mask. 9811cac6f2cSLuiz Capitulino */ 9821cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 9831cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 9841cac6f2cSLuiz Capitulino { 9851cac6f2cSLuiz Capitulino int nid; 9861cac6f2cSLuiz Capitulino 9871cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 9881cac6f2cSLuiz Capitulino 9891cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 9901cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 9911cac6f2cSLuiz Capitulino 9921cac6f2cSLuiz Capitulino return nid; 9931cac6f2cSLuiz Capitulino } 9941cac6f2cSLuiz Capitulino 9951cac6f2cSLuiz Capitulino /* 9961cac6f2cSLuiz Capitulino * helper for free_pool_huge_page() - return the previously saved 9971cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 9981cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 9991cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 10001cac6f2cSLuiz Capitulino */ 10011cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 10021cac6f2cSLuiz Capitulino { 10031cac6f2cSLuiz Capitulino int nid; 10041cac6f2cSLuiz Capitulino 10051cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 10061cac6f2cSLuiz Capitulino 10071cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 10081cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 10091cac6f2cSLuiz Capitulino 10101cac6f2cSLuiz Capitulino return nid; 10111cac6f2cSLuiz Capitulino } 10121cac6f2cSLuiz Capitulino 10131cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 10141cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 10151cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 10161cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 10171cac6f2cSLuiz Capitulino nr_nodes--) 10181cac6f2cSLuiz Capitulino 10191cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 10201cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 10211cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 10221cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 10231cac6f2cSLuiz Capitulino nr_nodes--) 10241cac6f2cSLuiz Capitulino 1025080fe206SVlastimil Babka #if defined(CONFIG_X86_64) && ((defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)) 1026944d9fecSLuiz Capitulino static void destroy_compound_gigantic_page(struct page *page, 1027d00181b9SKirill A. Shutemov unsigned int order) 1028944d9fecSLuiz Capitulino { 1029944d9fecSLuiz Capitulino int i; 1030944d9fecSLuiz Capitulino int nr_pages = 1 << order; 1031944d9fecSLuiz Capitulino struct page *p = page + 1; 1032944d9fecSLuiz Capitulino 1033944d9fecSLuiz Capitulino for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 10341d798ca3SKirill A. Shutemov clear_compound_head(p); 1035944d9fecSLuiz Capitulino set_page_refcounted(p); 1036944d9fecSLuiz Capitulino } 1037944d9fecSLuiz Capitulino 1038944d9fecSLuiz Capitulino set_compound_order(page, 0); 1039944d9fecSLuiz Capitulino __ClearPageHead(page); 1040944d9fecSLuiz Capitulino } 1041944d9fecSLuiz Capitulino 1042d00181b9SKirill A. Shutemov static void free_gigantic_page(struct page *page, unsigned int order) 1043944d9fecSLuiz Capitulino { 1044944d9fecSLuiz Capitulino free_contig_range(page_to_pfn(page), 1 << order); 1045944d9fecSLuiz Capitulino } 1046944d9fecSLuiz Capitulino 1047944d9fecSLuiz Capitulino static int __alloc_gigantic_page(unsigned long start_pfn, 1048944d9fecSLuiz Capitulino unsigned long nr_pages) 1049944d9fecSLuiz Capitulino { 1050944d9fecSLuiz Capitulino unsigned long end_pfn = start_pfn + nr_pages; 1051944d9fecSLuiz Capitulino return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE); 1052944d9fecSLuiz Capitulino } 1053944d9fecSLuiz Capitulino 1054f44b2ddaSJoonsoo Kim static bool pfn_range_valid_gigantic(struct zone *z, 1055f44b2ddaSJoonsoo Kim unsigned long start_pfn, unsigned long nr_pages) 1056944d9fecSLuiz Capitulino { 1057944d9fecSLuiz Capitulino unsigned long i, end_pfn = start_pfn + nr_pages; 1058944d9fecSLuiz Capitulino struct page *page; 1059944d9fecSLuiz Capitulino 1060944d9fecSLuiz Capitulino for (i = start_pfn; i < end_pfn; i++) { 1061944d9fecSLuiz Capitulino if (!pfn_valid(i)) 1062944d9fecSLuiz Capitulino return false; 1063944d9fecSLuiz Capitulino 1064944d9fecSLuiz Capitulino page = pfn_to_page(i); 1065944d9fecSLuiz Capitulino 1066f44b2ddaSJoonsoo Kim if (page_zone(page) != z) 1067f44b2ddaSJoonsoo Kim return false; 1068f44b2ddaSJoonsoo Kim 1069944d9fecSLuiz Capitulino if (PageReserved(page)) 1070944d9fecSLuiz Capitulino return false; 1071944d9fecSLuiz Capitulino 1072944d9fecSLuiz Capitulino if (page_count(page) > 0) 1073944d9fecSLuiz Capitulino return false; 1074944d9fecSLuiz Capitulino 1075944d9fecSLuiz Capitulino if (PageHuge(page)) 1076944d9fecSLuiz Capitulino return false; 1077944d9fecSLuiz Capitulino } 1078944d9fecSLuiz Capitulino 1079944d9fecSLuiz Capitulino return true; 1080944d9fecSLuiz Capitulino } 1081944d9fecSLuiz Capitulino 1082944d9fecSLuiz Capitulino static bool zone_spans_last_pfn(const struct zone *zone, 1083944d9fecSLuiz Capitulino unsigned long start_pfn, unsigned long nr_pages) 1084944d9fecSLuiz Capitulino { 1085944d9fecSLuiz Capitulino unsigned long last_pfn = start_pfn + nr_pages - 1; 1086944d9fecSLuiz Capitulino return zone_spans_pfn(zone, last_pfn); 1087944d9fecSLuiz Capitulino } 1088944d9fecSLuiz Capitulino 1089d00181b9SKirill A. Shutemov static struct page *alloc_gigantic_page(int nid, unsigned int order) 1090944d9fecSLuiz Capitulino { 1091944d9fecSLuiz Capitulino unsigned long nr_pages = 1 << order; 1092944d9fecSLuiz Capitulino unsigned long ret, pfn, flags; 1093944d9fecSLuiz Capitulino struct zone *z; 1094944d9fecSLuiz Capitulino 1095944d9fecSLuiz Capitulino z = NODE_DATA(nid)->node_zones; 1096944d9fecSLuiz Capitulino for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) { 1097944d9fecSLuiz Capitulino spin_lock_irqsave(&z->lock, flags); 1098944d9fecSLuiz Capitulino 1099944d9fecSLuiz Capitulino pfn = ALIGN(z->zone_start_pfn, nr_pages); 1100944d9fecSLuiz Capitulino while (zone_spans_last_pfn(z, pfn, nr_pages)) { 1101f44b2ddaSJoonsoo Kim if (pfn_range_valid_gigantic(z, pfn, nr_pages)) { 1102944d9fecSLuiz Capitulino /* 1103944d9fecSLuiz Capitulino * We release the zone lock here because 1104944d9fecSLuiz Capitulino * alloc_contig_range() will also lock the zone 1105944d9fecSLuiz Capitulino * at some point. If there's an allocation 1106944d9fecSLuiz Capitulino * spinning on this lock, it may win the race 1107944d9fecSLuiz Capitulino * and cause alloc_contig_range() to fail... 1108944d9fecSLuiz Capitulino */ 1109944d9fecSLuiz Capitulino spin_unlock_irqrestore(&z->lock, flags); 1110944d9fecSLuiz Capitulino ret = __alloc_gigantic_page(pfn, nr_pages); 1111944d9fecSLuiz Capitulino if (!ret) 1112944d9fecSLuiz Capitulino return pfn_to_page(pfn); 1113944d9fecSLuiz Capitulino spin_lock_irqsave(&z->lock, flags); 1114944d9fecSLuiz Capitulino } 1115944d9fecSLuiz Capitulino pfn += nr_pages; 1116944d9fecSLuiz Capitulino } 1117944d9fecSLuiz Capitulino 1118944d9fecSLuiz Capitulino spin_unlock_irqrestore(&z->lock, flags); 1119944d9fecSLuiz Capitulino } 1120944d9fecSLuiz Capitulino 1121944d9fecSLuiz Capitulino return NULL; 1122944d9fecSLuiz Capitulino } 1123944d9fecSLuiz Capitulino 1124944d9fecSLuiz Capitulino static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); 1125d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order); 1126944d9fecSLuiz Capitulino 1127944d9fecSLuiz Capitulino static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid) 1128944d9fecSLuiz Capitulino { 1129944d9fecSLuiz Capitulino struct page *page; 1130944d9fecSLuiz Capitulino 1131944d9fecSLuiz Capitulino page = alloc_gigantic_page(nid, huge_page_order(h)); 1132944d9fecSLuiz Capitulino if (page) { 1133944d9fecSLuiz Capitulino prep_compound_gigantic_page(page, huge_page_order(h)); 1134944d9fecSLuiz Capitulino prep_new_huge_page(h, page, nid); 1135944d9fecSLuiz Capitulino } 1136944d9fecSLuiz Capitulino 1137944d9fecSLuiz Capitulino return page; 1138944d9fecSLuiz Capitulino } 1139944d9fecSLuiz Capitulino 1140944d9fecSLuiz Capitulino static int alloc_fresh_gigantic_page(struct hstate *h, 1141944d9fecSLuiz Capitulino nodemask_t *nodes_allowed) 1142944d9fecSLuiz Capitulino { 1143944d9fecSLuiz Capitulino struct page *page = NULL; 1144944d9fecSLuiz Capitulino int nr_nodes, node; 1145944d9fecSLuiz Capitulino 1146944d9fecSLuiz Capitulino for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1147944d9fecSLuiz Capitulino page = alloc_fresh_gigantic_page_node(h, node); 1148944d9fecSLuiz Capitulino if (page) 1149944d9fecSLuiz Capitulino return 1; 1150944d9fecSLuiz Capitulino } 1151944d9fecSLuiz Capitulino 1152944d9fecSLuiz Capitulino return 0; 1153944d9fecSLuiz Capitulino } 1154944d9fecSLuiz Capitulino 1155944d9fecSLuiz Capitulino static inline bool gigantic_page_supported(void) { return true; } 1156944d9fecSLuiz Capitulino #else 1157944d9fecSLuiz Capitulino static inline bool gigantic_page_supported(void) { return false; } 1158d00181b9SKirill A. Shutemov static inline void free_gigantic_page(struct page *page, unsigned int order) { } 1159944d9fecSLuiz Capitulino static inline void destroy_compound_gigantic_page(struct page *page, 1160d00181b9SKirill A. Shutemov unsigned int order) { } 1161944d9fecSLuiz Capitulino static inline int alloc_fresh_gigantic_page(struct hstate *h, 1162944d9fecSLuiz Capitulino nodemask_t *nodes_allowed) { return 0; } 1163944d9fecSLuiz Capitulino #endif 1164944d9fecSLuiz Capitulino 1165a5516438SAndi Kleen static void update_and_free_page(struct hstate *h, struct page *page) 11666af2acb6SAdam Litke { 11676af2acb6SAdam Litke int i; 1168a5516438SAndi Kleen 1169944d9fecSLuiz Capitulino if (hstate_is_gigantic(h) && !gigantic_page_supported()) 1170944d9fecSLuiz Capitulino return; 117118229df5SAndy Whitcroft 1172a5516438SAndi Kleen h->nr_huge_pages--; 1173a5516438SAndi Kleen h->nr_huge_pages_node[page_to_nid(page)]--; 1174a5516438SAndi Kleen for (i = 0; i < pages_per_huge_page(h); i++) { 117532f84528SChris Forbes page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 117632f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1177a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1178a7407a27SLuiz Capitulino 1 << PG_writeback); 11796af2acb6SAdam Litke } 1180309381feSSasha Levin VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 1181f1e61557SKirill A. Shutemov set_compound_page_dtor(page, NULL_COMPOUND_DTOR); 11826af2acb6SAdam Litke set_page_refcounted(page); 1183944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) { 1184944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 1185944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 1186944d9fecSLuiz Capitulino } else { 1187a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 11886af2acb6SAdam Litke } 1189944d9fecSLuiz Capitulino } 11906af2acb6SAdam Litke 1191e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1192e5ff2159SAndi Kleen { 1193e5ff2159SAndi Kleen struct hstate *h; 1194e5ff2159SAndi Kleen 1195e5ff2159SAndi Kleen for_each_hstate(h) { 1196e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1197e5ff2159SAndi Kleen return h; 1198e5ff2159SAndi Kleen } 1199e5ff2159SAndi Kleen return NULL; 1200e5ff2159SAndi Kleen } 1201e5ff2159SAndi Kleen 1202bcc54222SNaoya Horiguchi /* 1203bcc54222SNaoya Horiguchi * Test to determine whether the hugepage is "active/in-use" (i.e. being linked 1204bcc54222SNaoya Horiguchi * to hstate->hugepage_activelist.) 1205bcc54222SNaoya Horiguchi * 1206bcc54222SNaoya Horiguchi * This function can be called for tail pages, but never returns true for them. 1207bcc54222SNaoya Horiguchi */ 1208bcc54222SNaoya Horiguchi bool page_huge_active(struct page *page) 1209bcc54222SNaoya Horiguchi { 1210bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHuge(page), page); 1211bcc54222SNaoya Horiguchi return PageHead(page) && PagePrivate(&page[1]); 1212bcc54222SNaoya Horiguchi } 1213bcc54222SNaoya Horiguchi 1214bcc54222SNaoya Horiguchi /* never called for tail page */ 1215bcc54222SNaoya Horiguchi static void set_page_huge_active(struct page *page) 1216bcc54222SNaoya Horiguchi { 1217bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 1218bcc54222SNaoya Horiguchi SetPagePrivate(&page[1]); 1219bcc54222SNaoya Horiguchi } 1220bcc54222SNaoya Horiguchi 1221bcc54222SNaoya Horiguchi static void clear_page_huge_active(struct page *page) 1222bcc54222SNaoya Horiguchi { 1223bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 1224bcc54222SNaoya Horiguchi ClearPagePrivate(&page[1]); 1225bcc54222SNaoya Horiguchi } 1226bcc54222SNaoya Horiguchi 12278f1d26d0SAtsushi Kumagai void free_huge_page(struct page *page) 122827a85ef1SDavid Gibson { 1229a5516438SAndi Kleen /* 1230a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1231a5516438SAndi Kleen * compound page destructor. 1232a5516438SAndi Kleen */ 1233e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 12347893d1d5SAdam Litke int nid = page_to_nid(page); 123590481622SDavid Gibson struct hugepage_subpool *spool = 123690481622SDavid Gibson (struct hugepage_subpool *)page_private(page); 123707443a85SJoonsoo Kim bool restore_reserve; 123827a85ef1SDavid Gibson 1239e5df70abSAndy Whitcroft set_page_private(page, 0); 124023be7468SMel Gorman page->mapping = NULL; 1241b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1242b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_mapcount(page), page); 124307443a85SJoonsoo Kim restore_reserve = PagePrivate(page); 124416c794b4SJoonsoo Kim ClearPagePrivate(page); 124527a85ef1SDavid Gibson 12461c5ecae3SMike Kravetz /* 12471c5ecae3SMike Kravetz * A return code of zero implies that the subpool will be under its 12481c5ecae3SMike Kravetz * minimum size if the reservation is not restored after page is free. 12491c5ecae3SMike Kravetz * Therefore, force restore_reserve operation. 12501c5ecae3SMike Kravetz */ 12511c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 12521c5ecae3SMike Kravetz restore_reserve = true; 12531c5ecae3SMike Kravetz 125427a85ef1SDavid Gibson spin_lock(&hugetlb_lock); 1255bcc54222SNaoya Horiguchi clear_page_huge_active(page); 12566d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 12576d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 125807443a85SJoonsoo Kim if (restore_reserve) 125907443a85SJoonsoo Kim h->resv_huge_pages++; 126007443a85SJoonsoo Kim 1261944d9fecSLuiz Capitulino if (h->surplus_huge_pages_node[nid]) { 12620edaecfaSAneesh Kumar K.V /* remove the page from active list */ 12630edaecfaSAneesh Kumar K.V list_del(&page->lru); 1264a5516438SAndi Kleen update_and_free_page(h, page); 1265a5516438SAndi Kleen h->surplus_huge_pages--; 1266a5516438SAndi Kleen h->surplus_huge_pages_node[nid]--; 12677893d1d5SAdam Litke } else { 12685d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1269a5516438SAndi Kleen enqueue_huge_page(h, page); 12707893d1d5SAdam Litke } 127127a85ef1SDavid Gibson spin_unlock(&hugetlb_lock); 127227a85ef1SDavid Gibson } 127327a85ef1SDavid Gibson 1274a5516438SAndi Kleen static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1275b7ba30c6SAndi Kleen { 12760edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1277f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1278b7ba30c6SAndi Kleen spin_lock(&hugetlb_lock); 12799dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 1280a5516438SAndi Kleen h->nr_huge_pages++; 1281a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 1282b7ba30c6SAndi Kleen spin_unlock(&hugetlb_lock); 1283b7ba30c6SAndi Kleen put_page(page); /* free it into the hugepage allocator */ 1284b7ba30c6SAndi Kleen } 1285b7ba30c6SAndi Kleen 1286d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order) 128720a0307cSWu Fengguang { 128820a0307cSWu Fengguang int i; 128920a0307cSWu Fengguang int nr_pages = 1 << order; 129020a0307cSWu Fengguang struct page *p = page + 1; 129120a0307cSWu Fengguang 129220a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 129320a0307cSWu Fengguang set_compound_order(page, order); 1294ef5a22beSAndrea Arcangeli __ClearPageReserved(page); 1295de09d31dSKirill A. Shutemov __SetPageHead(page); 129620a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1297ef5a22beSAndrea Arcangeli /* 1298ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1299ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1300ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 1301ef5a22beSAndrea Arcangeli * too. Otherwse drivers using get_user_pages() to access tail 1302ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1303ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1304ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1305ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1306ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1307ef5a22beSAndrea Arcangeli * after get_user_pages(). 1308ef5a22beSAndrea Arcangeli */ 1309ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 131058a84aa9SYouquan Song set_page_count(p, 0); 13111d798ca3SKirill A. Shutemov set_compound_head(p, page); 131220a0307cSWu Fengguang } 1313b4330afbSMike Kravetz atomic_set(compound_mapcount_ptr(page), -1); 131420a0307cSWu Fengguang } 131520a0307cSWu Fengguang 13167795912cSAndrew Morton /* 13177795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 13187795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 13197795912cSAndrew Morton * details. 13207795912cSAndrew Morton */ 132120a0307cSWu Fengguang int PageHuge(struct page *page) 132220a0307cSWu Fengguang { 132320a0307cSWu Fengguang if (!PageCompound(page)) 132420a0307cSWu Fengguang return 0; 132520a0307cSWu Fengguang 132620a0307cSWu Fengguang page = compound_head(page); 1327f1e61557SKirill A. Shutemov return page[1].compound_dtor == HUGETLB_PAGE_DTOR; 132820a0307cSWu Fengguang } 132943131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 133043131e14SNaoya Horiguchi 133127c73ae7SAndrea Arcangeli /* 133227c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 133327c73ae7SAndrea Arcangeli * normal or transparent huge pages. 133427c73ae7SAndrea Arcangeli */ 133527c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 133627c73ae7SAndrea Arcangeli { 133727c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 133827c73ae7SAndrea Arcangeli return 0; 133927c73ae7SAndrea Arcangeli 1340758f66a2SAndrew Morton return get_compound_page_dtor(page_head) == free_huge_page; 134127c73ae7SAndrea Arcangeli } 134227c73ae7SAndrea Arcangeli 134313d60f4bSZhang Yi pgoff_t __basepage_index(struct page *page) 134413d60f4bSZhang Yi { 134513d60f4bSZhang Yi struct page *page_head = compound_head(page); 134613d60f4bSZhang Yi pgoff_t index = page_index(page_head); 134713d60f4bSZhang Yi unsigned long compound_idx; 134813d60f4bSZhang Yi 134913d60f4bSZhang Yi if (!PageHuge(page_head)) 135013d60f4bSZhang Yi return page_index(page); 135113d60f4bSZhang Yi 135213d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 135313d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 135413d60f4bSZhang Yi else 135513d60f4bSZhang Yi compound_idx = page - page_head; 135613d60f4bSZhang Yi 135713d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 135813d60f4bSZhang Yi } 135913d60f4bSZhang Yi 1360a5516438SAndi Kleen static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) 13611da177e4SLinus Torvalds { 13621da177e4SLinus Torvalds struct page *page; 1363f96efd58SJoe Jin 136496db800fSVlastimil Babka page = __alloc_pages_node(nid, 136586cdb465SNaoya Horiguchi htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| 1366551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 1367a5516438SAndi Kleen huge_page_order(h)); 13681da177e4SLinus Torvalds if (page) { 1369a5516438SAndi Kleen prep_new_huge_page(h, page, nid); 13701da177e4SLinus Torvalds } 137163b4613cSNishanth Aravamudan 137263b4613cSNishanth Aravamudan return page; 137363b4613cSNishanth Aravamudan } 137463b4613cSNishanth Aravamudan 1375b2261026SJoonsoo Kim static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 1376b2261026SJoonsoo Kim { 1377b2261026SJoonsoo Kim struct page *page; 1378b2261026SJoonsoo Kim int nr_nodes, node; 1379b2261026SJoonsoo Kim int ret = 0; 1380b2261026SJoonsoo Kim 1381b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1382b2261026SJoonsoo Kim page = alloc_fresh_huge_page_node(h, node); 1383b2261026SJoonsoo Kim if (page) { 1384b2261026SJoonsoo Kim ret = 1; 1385b2261026SJoonsoo Kim break; 1386b2261026SJoonsoo Kim } 1387b2261026SJoonsoo Kim } 1388b2261026SJoonsoo Kim 1389b2261026SJoonsoo Kim if (ret) 1390b2261026SJoonsoo Kim count_vm_event(HTLB_BUDDY_PGALLOC); 1391b2261026SJoonsoo Kim else 1392b2261026SJoonsoo Kim count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 1393b2261026SJoonsoo Kim 1394b2261026SJoonsoo Kim return ret; 1395b2261026SJoonsoo Kim } 1396b2261026SJoonsoo Kim 1397e8c5c824SLee Schermerhorn /* 1398e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 1399e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 1400e8c5c824SLee Schermerhorn * balanced over allowed nodes. 1401e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 1402e8c5c824SLee Schermerhorn */ 14036ae11b27SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 14046ae11b27SLee Schermerhorn bool acct_surplus) 1405e8c5c824SLee Schermerhorn { 1406b2261026SJoonsoo Kim int nr_nodes, node; 1407e8c5c824SLee Schermerhorn int ret = 0; 1408e8c5c824SLee Schermerhorn 1409b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1410685f3457SLee Schermerhorn /* 1411685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 1412685f3457SLee Schermerhorn * nodes with surplus pages. 1413685f3457SLee Schermerhorn */ 1414b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 1415b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 1416e8c5c824SLee Schermerhorn struct page *page = 1417b2261026SJoonsoo Kim list_entry(h->hugepage_freelists[node].next, 1418e8c5c824SLee Schermerhorn struct page, lru); 1419e8c5c824SLee Schermerhorn list_del(&page->lru); 1420e8c5c824SLee Schermerhorn h->free_huge_pages--; 1421b2261026SJoonsoo Kim h->free_huge_pages_node[node]--; 1422685f3457SLee Schermerhorn if (acct_surplus) { 1423685f3457SLee Schermerhorn h->surplus_huge_pages--; 1424b2261026SJoonsoo Kim h->surplus_huge_pages_node[node]--; 1425685f3457SLee Schermerhorn } 1426e8c5c824SLee Schermerhorn update_and_free_page(h, page); 1427e8c5c824SLee Schermerhorn ret = 1; 14289a76db09SLee Schermerhorn break; 1429e8c5c824SLee Schermerhorn } 1430b2261026SJoonsoo Kim } 1431e8c5c824SLee Schermerhorn 1432e8c5c824SLee Schermerhorn return ret; 1433e8c5c824SLee Schermerhorn } 1434e8c5c824SLee Schermerhorn 1435c8721bbbSNaoya Horiguchi /* 1436c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 1437c8721bbbSNaoya Horiguchi * nothing for in-use (including surplus) hugepages. 1438c8721bbbSNaoya Horiguchi */ 1439c8721bbbSNaoya Horiguchi static void dissolve_free_huge_page(struct page *page) 1440c8721bbbSNaoya Horiguchi { 1441c8721bbbSNaoya Horiguchi spin_lock(&hugetlb_lock); 1442c8721bbbSNaoya Horiguchi if (PageHuge(page) && !page_count(page)) { 1443c8721bbbSNaoya Horiguchi struct hstate *h = page_hstate(page); 1444c8721bbbSNaoya Horiguchi int nid = page_to_nid(page); 1445c8721bbbSNaoya Horiguchi list_del(&page->lru); 1446c8721bbbSNaoya Horiguchi h->free_huge_pages--; 1447c8721bbbSNaoya Horiguchi h->free_huge_pages_node[nid]--; 1448c8721bbbSNaoya Horiguchi update_and_free_page(h, page); 1449c8721bbbSNaoya Horiguchi } 1450c8721bbbSNaoya Horiguchi spin_unlock(&hugetlb_lock); 1451c8721bbbSNaoya Horiguchi } 1452c8721bbbSNaoya Horiguchi 1453c8721bbbSNaoya Horiguchi /* 1454c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 1455c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 1456c8721bbbSNaoya Horiguchi * Note that start_pfn should aligned with (minimum) hugepage size. 1457c8721bbbSNaoya Horiguchi */ 1458c8721bbbSNaoya Horiguchi void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 1459c8721bbbSNaoya Horiguchi { 1460c8721bbbSNaoya Horiguchi unsigned long pfn; 1461c8721bbbSNaoya Horiguchi 1462d0177639SLi Zhong if (!hugepages_supported()) 1463d0177639SLi Zhong return; 1464d0177639SLi Zhong 1465641844f5SNaoya Horiguchi VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << minimum_order)); 1466641844f5SNaoya Horiguchi for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) 1467c8721bbbSNaoya Horiguchi dissolve_free_huge_page(pfn_to_page(pfn)); 1468c8721bbbSNaoya Horiguchi } 1469c8721bbbSNaoya Horiguchi 1470099730d6SDave Hansen /* 1471099730d6SDave Hansen * There are 3 ways this can get called: 1472099730d6SDave Hansen * 1. With vma+addr: we use the VMA's memory policy 1473099730d6SDave Hansen * 2. With !vma, but nid=NUMA_NO_NODE: We try to allocate a huge 1474099730d6SDave Hansen * page from any node, and let the buddy allocator itself figure 1475099730d6SDave Hansen * it out. 1476099730d6SDave Hansen * 3. With !vma, but nid!=NUMA_NO_NODE. We allocate a huge page 1477099730d6SDave Hansen * strictly from 'nid' 1478099730d6SDave Hansen */ 1479099730d6SDave Hansen static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h, 1480099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr, int nid) 1481099730d6SDave Hansen { 1482099730d6SDave Hansen int order = huge_page_order(h); 1483099730d6SDave Hansen gfp_t gfp = htlb_alloc_mask(h)|__GFP_COMP|__GFP_REPEAT|__GFP_NOWARN; 1484099730d6SDave Hansen unsigned int cpuset_mems_cookie; 1485099730d6SDave Hansen 1486099730d6SDave Hansen /* 1487099730d6SDave Hansen * We need a VMA to get a memory policy. If we do not 1488e0ec90eeSDave Hansen * have one, we use the 'nid' argument. 1489e0ec90eeSDave Hansen * 1490e0ec90eeSDave Hansen * The mempolicy stuff below has some non-inlined bits 1491e0ec90eeSDave Hansen * and calls ->vm_ops. That makes it hard to optimize at 1492e0ec90eeSDave Hansen * compile-time, even when NUMA is off and it does 1493e0ec90eeSDave Hansen * nothing. This helps the compiler optimize it out. 1494099730d6SDave Hansen */ 1495e0ec90eeSDave Hansen if (!IS_ENABLED(CONFIG_NUMA) || !vma) { 1496099730d6SDave Hansen /* 1497099730d6SDave Hansen * If a specific node is requested, make sure to 1498099730d6SDave Hansen * get memory from there, but only when a node 1499099730d6SDave Hansen * is explicitly specified. 1500099730d6SDave Hansen */ 1501099730d6SDave Hansen if (nid != NUMA_NO_NODE) 1502099730d6SDave Hansen gfp |= __GFP_THISNODE; 1503099730d6SDave Hansen /* 1504099730d6SDave Hansen * Make sure to call something that can handle 1505099730d6SDave Hansen * nid=NUMA_NO_NODE 1506099730d6SDave Hansen */ 1507099730d6SDave Hansen return alloc_pages_node(nid, gfp, order); 1508099730d6SDave Hansen } 1509099730d6SDave Hansen 1510099730d6SDave Hansen /* 1511099730d6SDave Hansen * OK, so we have a VMA. Fetch the mempolicy and try to 1512e0ec90eeSDave Hansen * allocate a huge page with it. We will only reach this 1513e0ec90eeSDave Hansen * when CONFIG_NUMA=y. 1514099730d6SDave Hansen */ 1515099730d6SDave Hansen do { 1516099730d6SDave Hansen struct page *page; 1517099730d6SDave Hansen struct mempolicy *mpol; 1518099730d6SDave Hansen struct zonelist *zl; 1519099730d6SDave Hansen nodemask_t *nodemask; 1520099730d6SDave Hansen 1521099730d6SDave Hansen cpuset_mems_cookie = read_mems_allowed_begin(); 1522099730d6SDave Hansen zl = huge_zonelist(vma, addr, gfp, &mpol, &nodemask); 1523099730d6SDave Hansen mpol_cond_put(mpol); 1524099730d6SDave Hansen page = __alloc_pages_nodemask(gfp, order, zl, nodemask); 1525099730d6SDave Hansen if (page) 1526099730d6SDave Hansen return page; 1527099730d6SDave Hansen } while (read_mems_allowed_retry(cpuset_mems_cookie)); 1528099730d6SDave Hansen 1529099730d6SDave Hansen return NULL; 1530099730d6SDave Hansen } 1531099730d6SDave Hansen 1532099730d6SDave Hansen /* 1533099730d6SDave Hansen * There are two ways to allocate a huge page: 1534099730d6SDave Hansen * 1. When you have a VMA and an address (like a fault) 1535099730d6SDave Hansen * 2. When you have no VMA (like when setting /proc/.../nr_hugepages) 1536099730d6SDave Hansen * 1537099730d6SDave Hansen * 'vma' and 'addr' are only for (1). 'nid' is always NUMA_NO_NODE in 1538099730d6SDave Hansen * this case which signifies that the allocation should be done with 1539099730d6SDave Hansen * respect for the VMA's memory policy. 1540099730d6SDave Hansen * 1541099730d6SDave Hansen * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This 1542099730d6SDave Hansen * implies that memory policies will not be taken in to account. 1543099730d6SDave Hansen */ 1544099730d6SDave Hansen static struct page *__alloc_buddy_huge_page(struct hstate *h, 1545099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr, int nid) 15467893d1d5SAdam Litke { 15477893d1d5SAdam Litke struct page *page; 1548bf50bab2SNaoya Horiguchi unsigned int r_nid; 15497893d1d5SAdam Litke 1550bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1551aa888a74SAndi Kleen return NULL; 1552aa888a74SAndi Kleen 1553d1c3fb1fSNishanth Aravamudan /* 1554099730d6SDave Hansen * Make sure that anyone specifying 'nid' is not also specifying a VMA. 1555099730d6SDave Hansen * This makes sure the caller is picking _one_ of the modes with which 1556099730d6SDave Hansen * we can call this function, not both. 1557099730d6SDave Hansen */ 1558099730d6SDave Hansen if (vma || (addr != -1)) { 1559e0ec90eeSDave Hansen VM_WARN_ON_ONCE(addr == -1); 1560e0ec90eeSDave Hansen VM_WARN_ON_ONCE(nid != NUMA_NO_NODE); 1561099730d6SDave Hansen } 1562099730d6SDave Hansen /* 1563d1c3fb1fSNishanth Aravamudan * Assume we will successfully allocate the surplus page to 1564d1c3fb1fSNishanth Aravamudan * prevent racing processes from causing the surplus to exceed 1565d1c3fb1fSNishanth Aravamudan * overcommit 1566d1c3fb1fSNishanth Aravamudan * 1567d1c3fb1fSNishanth Aravamudan * This however introduces a different race, where a process B 1568d1c3fb1fSNishanth Aravamudan * tries to grow the static hugepage pool while alloc_pages() is 1569d1c3fb1fSNishanth Aravamudan * called by process A. B will only examine the per-node 1570d1c3fb1fSNishanth Aravamudan * counters in determining if surplus huge pages can be 1571d1c3fb1fSNishanth Aravamudan * converted to normal huge pages in adjust_pool_surplus(). A 1572d1c3fb1fSNishanth Aravamudan * won't be able to increment the per-node counter, until the 1573d1c3fb1fSNishanth Aravamudan * lock is dropped by B, but B doesn't drop hugetlb_lock until 1574d1c3fb1fSNishanth Aravamudan * no more huge pages can be converted from surplus to normal 1575d1c3fb1fSNishanth Aravamudan * state (and doesn't try to convert again). Thus, we have a 1576d1c3fb1fSNishanth Aravamudan * case where a surplus huge page exists, the pool is grown, and 1577d1c3fb1fSNishanth Aravamudan * the surplus huge page still exists after, even though it 1578d1c3fb1fSNishanth Aravamudan * should just have been converted to a normal huge page. This 1579d1c3fb1fSNishanth Aravamudan * does not leak memory, though, as the hugepage will be freed 1580d1c3fb1fSNishanth Aravamudan * once it is out of use. It also does not allow the counters to 1581d1c3fb1fSNishanth Aravamudan * go out of whack in adjust_pool_surplus() as we don't modify 1582d1c3fb1fSNishanth Aravamudan * the node values until we've gotten the hugepage and only the 1583d1c3fb1fSNishanth Aravamudan * per-node value is checked there. 1584d1c3fb1fSNishanth Aravamudan */ 1585d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 1586a5516438SAndi Kleen if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 1587d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1588d1c3fb1fSNishanth Aravamudan return NULL; 1589d1c3fb1fSNishanth Aravamudan } else { 1590a5516438SAndi Kleen h->nr_huge_pages++; 1591a5516438SAndi Kleen h->surplus_huge_pages++; 1592d1c3fb1fSNishanth Aravamudan } 1593d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1594d1c3fb1fSNishanth Aravamudan 1595099730d6SDave Hansen page = __hugetlb_alloc_buddy_huge_page(h, vma, addr, nid); 1596d1c3fb1fSNishanth Aravamudan 15977893d1d5SAdam Litke spin_lock(&hugetlb_lock); 1598d1c3fb1fSNishanth Aravamudan if (page) { 15990edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1600bf50bab2SNaoya Horiguchi r_nid = page_to_nid(page); 1601f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 16029dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 1603d1c3fb1fSNishanth Aravamudan /* 1604d1c3fb1fSNishanth Aravamudan * We incremented the global counters already 1605d1c3fb1fSNishanth Aravamudan */ 1606bf50bab2SNaoya Horiguchi h->nr_huge_pages_node[r_nid]++; 1607bf50bab2SNaoya Horiguchi h->surplus_huge_pages_node[r_nid]++; 16083b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC); 1609d1c3fb1fSNishanth Aravamudan } else { 1610a5516438SAndi Kleen h->nr_huge_pages--; 1611a5516438SAndi Kleen h->surplus_huge_pages--; 16123b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 16137893d1d5SAdam Litke } 1614d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 16157893d1d5SAdam Litke 16167893d1d5SAdam Litke return page; 16177893d1d5SAdam Litke } 16187893d1d5SAdam Litke 1619e4e574b7SAdam Litke /* 1620099730d6SDave Hansen * Allocate a huge page from 'nid'. Note, 'nid' may be 1621099730d6SDave Hansen * NUMA_NO_NODE, which means that it may be allocated 1622099730d6SDave Hansen * anywhere. 1623099730d6SDave Hansen */ 1624e0ec90eeSDave Hansen static 1625099730d6SDave Hansen struct page *__alloc_buddy_huge_page_no_mpol(struct hstate *h, int nid) 1626099730d6SDave Hansen { 1627099730d6SDave Hansen unsigned long addr = -1; 1628099730d6SDave Hansen 1629099730d6SDave Hansen return __alloc_buddy_huge_page(h, NULL, addr, nid); 1630099730d6SDave Hansen } 1631099730d6SDave Hansen 1632099730d6SDave Hansen /* 1633099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 1634099730d6SDave Hansen */ 1635e0ec90eeSDave Hansen static 1636099730d6SDave Hansen struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h, 1637099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 1638099730d6SDave Hansen { 1639099730d6SDave Hansen return __alloc_buddy_huge_page(h, vma, addr, NUMA_NO_NODE); 1640099730d6SDave Hansen } 1641099730d6SDave Hansen 1642099730d6SDave Hansen /* 1643bf50bab2SNaoya Horiguchi * This allocation function is useful in the context where vma is irrelevant. 1644bf50bab2SNaoya Horiguchi * E.g. soft-offlining uses this function because it only cares physical 1645bf50bab2SNaoya Horiguchi * address of error page. 1646bf50bab2SNaoya Horiguchi */ 1647bf50bab2SNaoya Horiguchi struct page *alloc_huge_page_node(struct hstate *h, int nid) 1648bf50bab2SNaoya Horiguchi { 16494ef91848SJoonsoo Kim struct page *page = NULL; 1650bf50bab2SNaoya Horiguchi 1651bf50bab2SNaoya Horiguchi spin_lock(&hugetlb_lock); 16524ef91848SJoonsoo Kim if (h->free_huge_pages - h->resv_huge_pages > 0) 1653bf50bab2SNaoya Horiguchi page = dequeue_huge_page_node(h, nid); 1654bf50bab2SNaoya Horiguchi spin_unlock(&hugetlb_lock); 1655bf50bab2SNaoya Horiguchi 165694ae8ba7SAneesh Kumar K.V if (!page) 1657099730d6SDave Hansen page = __alloc_buddy_huge_page_no_mpol(h, nid); 1658bf50bab2SNaoya Horiguchi 1659bf50bab2SNaoya Horiguchi return page; 1660bf50bab2SNaoya Horiguchi } 1661bf50bab2SNaoya Horiguchi 1662bf50bab2SNaoya Horiguchi /* 166325985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 1664e4e574b7SAdam Litke * of size 'delta'. 1665e4e574b7SAdam Litke */ 1666a5516438SAndi Kleen static int gather_surplus_pages(struct hstate *h, int delta) 1667e4e574b7SAdam Litke { 1668e4e574b7SAdam Litke struct list_head surplus_list; 1669e4e574b7SAdam Litke struct page *page, *tmp; 1670e4e574b7SAdam Litke int ret, i; 1671e4e574b7SAdam Litke int needed, allocated; 167228073b02SHillf Danton bool alloc_ok = true; 1673e4e574b7SAdam Litke 1674a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 1675ac09b3a1SAdam Litke if (needed <= 0) { 1676a5516438SAndi Kleen h->resv_huge_pages += delta; 1677e4e574b7SAdam Litke return 0; 1678ac09b3a1SAdam Litke } 1679e4e574b7SAdam Litke 1680e4e574b7SAdam Litke allocated = 0; 1681e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 1682e4e574b7SAdam Litke 1683e4e574b7SAdam Litke ret = -ENOMEM; 1684e4e574b7SAdam Litke retry: 1685e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 1686e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 1687099730d6SDave Hansen page = __alloc_buddy_huge_page_no_mpol(h, NUMA_NO_NODE); 168828073b02SHillf Danton if (!page) { 168928073b02SHillf Danton alloc_ok = false; 169028073b02SHillf Danton break; 169128073b02SHillf Danton } 1692e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 1693e4e574b7SAdam Litke } 169428073b02SHillf Danton allocated += i; 1695e4e574b7SAdam Litke 1696e4e574b7SAdam Litke /* 1697e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 1698e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 1699e4e574b7SAdam Litke */ 1700e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 1701a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 1702a5516438SAndi Kleen (h->free_huge_pages + allocated); 170328073b02SHillf Danton if (needed > 0) { 170428073b02SHillf Danton if (alloc_ok) 1705e4e574b7SAdam Litke goto retry; 170628073b02SHillf Danton /* 170728073b02SHillf Danton * We were not able to allocate enough pages to 170828073b02SHillf Danton * satisfy the entire reservation so we free what 170928073b02SHillf Danton * we've allocated so far. 171028073b02SHillf Danton */ 171128073b02SHillf Danton goto free; 171228073b02SHillf Danton } 1713e4e574b7SAdam Litke /* 1714e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 171525985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 1716e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 1717ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 1718ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 1719ac09b3a1SAdam Litke * before they are reserved. 1720e4e574b7SAdam Litke */ 1721e4e574b7SAdam Litke needed += allocated; 1722a5516438SAndi Kleen h->resv_huge_pages += delta; 1723e4e574b7SAdam Litke ret = 0; 1724a9869b83SNaoya Horiguchi 172519fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 172619fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 172719fc3f0aSAdam Litke if ((--needed) < 0) 172819fc3f0aSAdam Litke break; 1729a9869b83SNaoya Horiguchi /* 1730a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 1731a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 1732a9869b83SNaoya Horiguchi */ 1733a9869b83SNaoya Horiguchi put_page_testzero(page); 1734309381feSSasha Levin VM_BUG_ON_PAGE(page_count(page), page); 1735a5516438SAndi Kleen enqueue_huge_page(h, page); 173619fc3f0aSAdam Litke } 173728073b02SHillf Danton free: 1738b0365c8dSHillf Danton spin_unlock(&hugetlb_lock); 173919fc3f0aSAdam Litke 174019fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 1741c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 1742a9869b83SNaoya Horiguchi put_page(page); 174319fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 1744e4e574b7SAdam Litke 1745e4e574b7SAdam Litke return ret; 1746e4e574b7SAdam Litke } 1747e4e574b7SAdam Litke 1748e4e574b7SAdam Litke /* 1749e4e574b7SAdam Litke * When releasing a hugetlb pool reservation, any surplus pages that were 1750e4e574b7SAdam Litke * allocated to satisfy the reservation must be explicitly freed if they were 1751e4e574b7SAdam Litke * never used. 1752685f3457SLee Schermerhorn * Called with hugetlb_lock held. 1753e4e574b7SAdam Litke */ 1754a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 1755a5516438SAndi Kleen unsigned long unused_resv_pages) 1756e4e574b7SAdam Litke { 1757e4e574b7SAdam Litke unsigned long nr_pages; 1758e4e574b7SAdam Litke 1759ac09b3a1SAdam Litke /* Uncommit the reservation */ 1760a5516438SAndi Kleen h->resv_huge_pages -= unused_resv_pages; 1761ac09b3a1SAdam Litke 1762aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 1763bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1764aa888a74SAndi Kleen return; 1765aa888a74SAndi Kleen 1766a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 1767e4e574b7SAdam Litke 1768685f3457SLee Schermerhorn /* 1769685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 17709b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 17719b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 17729b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 17739b5e5d0fSLee Schermerhorn * free_pool_huge_page() will balance the the freed pages across the 17749b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 1775685f3457SLee Schermerhorn */ 1776685f3457SLee Schermerhorn while (nr_pages--) { 17778cebfcd0SLai Jiangshan if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) 1778685f3457SLee Schermerhorn break; 17797848a4bfSMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 1780e4e574b7SAdam Litke } 1781e4e574b7SAdam Litke } 1782e4e574b7SAdam Litke 17835e911373SMike Kravetz 1784c37f9fb1SAndy Whitcroft /* 1785feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 17865e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 1787cf3ad20bSMike Kravetz * 1788cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 1789cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 1790cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 1791cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 1792cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 1793feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 1794feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 1795feba16e2SMike Kravetz * is called in such cases. 1796cf3ad20bSMike Kravetz * 1797cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 1798cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 1799cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 1800cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 1801cf3ad20bSMike Kravetz * take appropriate action. 1802c37f9fb1SAndy Whitcroft */ 18035e911373SMike Kravetz enum vma_resv_mode { 18045e911373SMike Kravetz VMA_NEEDS_RESV, 18055e911373SMike Kravetz VMA_COMMIT_RESV, 1806feba16e2SMike Kravetz VMA_END_RESV, 18075e911373SMike Kravetz }; 1808cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 1809cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 18105e911373SMike Kravetz enum vma_resv_mode mode) 1811c37f9fb1SAndy Whitcroft { 18124e35f483SJoonsoo Kim struct resv_map *resv; 18134e35f483SJoonsoo Kim pgoff_t idx; 1814cf3ad20bSMike Kravetz long ret; 1815c37f9fb1SAndy Whitcroft 18164e35f483SJoonsoo Kim resv = vma_resv_map(vma); 18174e35f483SJoonsoo Kim if (!resv) 1818c37f9fb1SAndy Whitcroft return 1; 1819c37f9fb1SAndy Whitcroft 18204e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 18215e911373SMike Kravetz switch (mode) { 18225e911373SMike Kravetz case VMA_NEEDS_RESV: 1823cf3ad20bSMike Kravetz ret = region_chg(resv, idx, idx + 1); 18245e911373SMike Kravetz break; 18255e911373SMike Kravetz case VMA_COMMIT_RESV: 18265e911373SMike Kravetz ret = region_add(resv, idx, idx + 1); 18275e911373SMike Kravetz break; 1828feba16e2SMike Kravetz case VMA_END_RESV: 18295e911373SMike Kravetz region_abort(resv, idx, idx + 1); 18305e911373SMike Kravetz ret = 0; 18315e911373SMike Kravetz break; 18325e911373SMike Kravetz default: 18335e911373SMike Kravetz BUG(); 18345e911373SMike Kravetz } 183584afd99bSAndy Whitcroft 18364e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) 1837cf3ad20bSMike Kravetz return ret; 183867961f9dSMike Kravetz else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { 183967961f9dSMike Kravetz /* 184067961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 184167961f9dSMike Kravetz * However, a file associated with mapping could have been 184267961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 184367961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 184467961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 184567961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 184667961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 184767961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 184867961f9dSMike Kravetz * reservation has already been consumed. As a result, the 184967961f9dSMike Kravetz * return value of this routine is the opposite of the 185067961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 185167961f9dSMike Kravetz */ 185267961f9dSMike Kravetz if (ret) 185367961f9dSMike Kravetz return 0; 185467961f9dSMike Kravetz else 185567961f9dSMike Kravetz return 1; 185667961f9dSMike Kravetz } 18574e35f483SJoonsoo Kim else 1858cf3ad20bSMike Kravetz return ret < 0 ? ret : 0; 185984afd99bSAndy Whitcroft } 1860cf3ad20bSMike Kravetz 1861cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 1862a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 1863c37f9fb1SAndy Whitcroft { 18645e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 1865cf3ad20bSMike Kravetz } 1866c37f9fb1SAndy Whitcroft 1867cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 1868cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 1869cf3ad20bSMike Kravetz { 18705e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 18715e911373SMike Kravetz } 18725e911373SMike Kravetz 1873feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 18745e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 18755e911373SMike Kravetz { 1876feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 1877c37f9fb1SAndy Whitcroft } 1878c37f9fb1SAndy Whitcroft 187970c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 188004f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 1881348ea204SAdam Litke { 188290481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 1883a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 1884348ea204SAdam Litke struct page *page; 1885d85f69b0SMike Kravetz long map_chg, map_commit; 1886d85f69b0SMike Kravetz long gbl_chg; 18876d76dcf4SAneesh Kumar K.V int ret, idx; 18886d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 18892fc39cecSAdam Litke 18906d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 1891a1e78772SMel Gorman /* 1892d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 1893d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 1894d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 1895a1e78772SMel Gorman */ 1896d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 1897d85f69b0SMike Kravetz if (map_chg < 0) 189876dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 1899d85f69b0SMike Kravetz 1900d85f69b0SMike Kravetz /* 1901d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 1902d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 1903d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 1904d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 1905d85f69b0SMike Kravetz * checked against any subpool limit. 1906d85f69b0SMike Kravetz */ 1907d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 1908d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 1909d85f69b0SMike Kravetz if (gbl_chg < 0) { 1910feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 191176dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 19125e911373SMike Kravetz } 191390d8b7e6SAdam Litke 1914d85f69b0SMike Kravetz /* 1915d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 1916d85f69b0SMike Kravetz * map, there could be reservations associated with the 1917d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 1918d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 1919d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 1920d85f69b0SMike Kravetz * the subpool reservations. 1921d85f69b0SMike Kravetz */ 1922d85f69b0SMike Kravetz if (avoid_reserve) 1923d85f69b0SMike Kravetz gbl_chg = 1; 1924d85f69b0SMike Kravetz } 1925d85f69b0SMike Kravetz 19266d76dcf4SAneesh Kumar K.V ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 19278f34af6fSJianyu Zhan if (ret) 19288f34af6fSJianyu Zhan goto out_subpool_put; 19298f34af6fSJianyu Zhan 1930a1e78772SMel Gorman spin_lock(&hugetlb_lock); 1931d85f69b0SMike Kravetz /* 1932d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 1933d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 1934d85f69b0SMike Kravetz * a reservation exists for the allocation. 1935d85f69b0SMike Kravetz */ 1936d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 193781a6fcaeSJoonsoo Kim if (!page) { 193894ae8ba7SAneesh Kumar K.V spin_unlock(&hugetlb_lock); 1939099730d6SDave Hansen page = __alloc_buddy_huge_page_with_mpol(h, vma, addr); 19408f34af6fSJianyu Zhan if (!page) 19418f34af6fSJianyu Zhan goto out_uncharge_cgroup; 1942a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 1943a88c7695SNaoya Horiguchi SetPagePrivate(page); 1944a88c7695SNaoya Horiguchi h->resv_huge_pages--; 1945a88c7695SNaoya Horiguchi } 194679dbb236SAneesh Kumar K.V spin_lock(&hugetlb_lock); 194779dbb236SAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 194881a6fcaeSJoonsoo Kim /* Fall through */ 1949a1e78772SMel Gorman } 195081a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 195181a6fcaeSJoonsoo Kim spin_unlock(&hugetlb_lock); 1952a1e78772SMel Gorman 195390481622SDavid Gibson set_page_private(page, (unsigned long)spool); 1954a1e78772SMel Gorman 1955d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 1956d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 195733039678SMike Kravetz /* 195833039678SMike Kravetz * The page was added to the reservation map between 195933039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 196033039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 196133039678SMike Kravetz * Adjust for the subpool count incremented above AND 196233039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 196333039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 196433039678SMike Kravetz * no longer applies. 196533039678SMike Kravetz */ 196633039678SMike Kravetz long rsv_adjust; 196733039678SMike Kravetz 196833039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 196933039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 197033039678SMike Kravetz } 19717893d1d5SAdam Litke return page; 19728f34af6fSJianyu Zhan 19738f34af6fSJianyu Zhan out_uncharge_cgroup: 19748f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 19758f34af6fSJianyu Zhan out_subpool_put: 1976d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 19778f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 1978feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 19798f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 1980b45b5bd6SDavid Gibson } 1981b45b5bd6SDavid Gibson 198274060e4dSNaoya Horiguchi /* 198374060e4dSNaoya Horiguchi * alloc_huge_page()'s wrapper which simply returns the page if allocation 198474060e4dSNaoya Horiguchi * succeeds, otherwise NULL. This function is called from new_vma_page(), 198574060e4dSNaoya Horiguchi * where no ERR_VALUE is expected to be returned. 198674060e4dSNaoya Horiguchi */ 198774060e4dSNaoya Horiguchi struct page *alloc_huge_page_noerr(struct vm_area_struct *vma, 198874060e4dSNaoya Horiguchi unsigned long addr, int avoid_reserve) 198974060e4dSNaoya Horiguchi { 199074060e4dSNaoya Horiguchi struct page *page = alloc_huge_page(vma, addr, avoid_reserve); 199174060e4dSNaoya Horiguchi if (IS_ERR(page)) 199274060e4dSNaoya Horiguchi page = NULL; 199374060e4dSNaoya Horiguchi return page; 199474060e4dSNaoya Horiguchi } 199574060e4dSNaoya Horiguchi 199691f47662SCyrill Gorcunov int __weak alloc_bootmem_huge_page(struct hstate *h) 1997aa888a74SAndi Kleen { 1998aa888a74SAndi Kleen struct huge_bootmem_page *m; 1999b2261026SJoonsoo Kim int nr_nodes, node; 2000aa888a74SAndi Kleen 2001b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 2002aa888a74SAndi Kleen void *addr; 2003aa888a74SAndi Kleen 20048b89a116SGrygorii Strashko addr = memblock_virt_alloc_try_nid_nopanic( 20058b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 20068b89a116SGrygorii Strashko 0, BOOTMEM_ALLOC_ACCESSIBLE, node); 2007aa888a74SAndi Kleen if (addr) { 2008aa888a74SAndi Kleen /* 2009aa888a74SAndi Kleen * Use the beginning of the huge page to store the 2010aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 2011aa888a74SAndi Kleen * puts them into the mem_map). 2012aa888a74SAndi Kleen */ 2013aa888a74SAndi Kleen m = addr; 2014aa888a74SAndi Kleen goto found; 2015aa888a74SAndi Kleen } 2016aa888a74SAndi Kleen } 2017aa888a74SAndi Kleen return 0; 2018aa888a74SAndi Kleen 2019aa888a74SAndi Kleen found: 2020df994eadSLuiz Capitulino BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); 2021aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 2022aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 2023aa888a74SAndi Kleen m->hstate = h; 2024aa888a74SAndi Kleen return 1; 2025aa888a74SAndi Kleen } 2026aa888a74SAndi Kleen 2027d00181b9SKirill A. Shutemov static void __init prep_compound_huge_page(struct page *page, 2028d00181b9SKirill A. Shutemov unsigned int order) 202918229df5SAndy Whitcroft { 203018229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 203118229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 203218229df5SAndy Whitcroft else 203318229df5SAndy Whitcroft prep_compound_page(page, order); 203418229df5SAndy Whitcroft } 203518229df5SAndy Whitcroft 2036aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 2037aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 2038aa888a74SAndi Kleen { 2039aa888a74SAndi Kleen struct huge_bootmem_page *m; 2040aa888a74SAndi Kleen 2041aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 2042aa888a74SAndi Kleen struct hstate *h = m->hstate; 2043ee8f248dSBecky Bruce struct page *page; 2044ee8f248dSBecky Bruce 2045ee8f248dSBecky Bruce #ifdef CONFIG_HIGHMEM 2046ee8f248dSBecky Bruce page = pfn_to_page(m->phys >> PAGE_SHIFT); 20478b89a116SGrygorii Strashko memblock_free_late(__pa(m), 2048ee8f248dSBecky Bruce sizeof(struct huge_bootmem_page)); 2049ee8f248dSBecky Bruce #else 2050ee8f248dSBecky Bruce page = virt_to_page(m); 2051ee8f248dSBecky Bruce #endif 2052aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 205318229df5SAndy Whitcroft prep_compound_huge_page(page, h->order); 2054ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 2055aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 2056b0320c7bSRafael Aquini /* 2057b0320c7bSRafael Aquini * If we had gigantic hugepages allocated at boot time, we need 2058b0320c7bSRafael Aquini * to restore the 'stolen' pages to totalram_pages in order to 2059b0320c7bSRafael Aquini * fix confusing memory reports from free(1) and another 2060b0320c7bSRafael Aquini * side-effects, like CommitLimit going negative. 2061b0320c7bSRafael Aquini */ 2062bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 20633dcc0571SJiang Liu adjust_managed_page_count(page, 1 << h->order); 2064aa888a74SAndi Kleen } 2065aa888a74SAndi Kleen } 2066aa888a74SAndi Kleen 20678faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 20681da177e4SLinus Torvalds { 20691da177e4SLinus Torvalds unsigned long i; 20701da177e4SLinus Torvalds 2071e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 2072bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 2073aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 2074aa888a74SAndi Kleen break; 20759b5e5d0fSLee Schermerhorn } else if (!alloc_fresh_huge_page(h, 20768cebfcd0SLai Jiangshan &node_states[N_MEMORY])) 20771da177e4SLinus Torvalds break; 20781da177e4SLinus Torvalds } 20798faa8b07SAndi Kleen h->max_huge_pages = i; 2080e5ff2159SAndi Kleen } 2081e5ff2159SAndi Kleen 2082e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 2083e5ff2159SAndi Kleen { 2084e5ff2159SAndi Kleen struct hstate *h; 2085e5ff2159SAndi Kleen 2086e5ff2159SAndi Kleen for_each_hstate(h) { 2087641844f5SNaoya Horiguchi if (minimum_order > huge_page_order(h)) 2088641844f5SNaoya Horiguchi minimum_order = huge_page_order(h); 2089641844f5SNaoya Horiguchi 20908faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 2091bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 20928faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 2093e5ff2159SAndi Kleen } 2094641844f5SNaoya Horiguchi VM_BUG_ON(minimum_order == UINT_MAX); 2095e5ff2159SAndi Kleen } 2096e5ff2159SAndi Kleen 20974abd32dbSAndi Kleen static char * __init memfmt(char *buf, unsigned long n) 20984abd32dbSAndi Kleen { 20994abd32dbSAndi Kleen if (n >= (1UL << 30)) 21004abd32dbSAndi Kleen sprintf(buf, "%lu GB", n >> 30); 21014abd32dbSAndi Kleen else if (n >= (1UL << 20)) 21024abd32dbSAndi Kleen sprintf(buf, "%lu MB", n >> 20); 21034abd32dbSAndi Kleen else 21044abd32dbSAndi Kleen sprintf(buf, "%lu KB", n >> 10); 21054abd32dbSAndi Kleen return buf; 21064abd32dbSAndi Kleen } 21074abd32dbSAndi Kleen 2108e5ff2159SAndi Kleen static void __init report_hugepages(void) 2109e5ff2159SAndi Kleen { 2110e5ff2159SAndi Kleen struct hstate *h; 2111e5ff2159SAndi Kleen 2112e5ff2159SAndi Kleen for_each_hstate(h) { 21134abd32dbSAndi Kleen char buf[32]; 2114ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 21154abd32dbSAndi Kleen memfmt(buf, huge_page_size(h)), 21164abd32dbSAndi Kleen h->free_huge_pages); 2117e5ff2159SAndi Kleen } 2118e5ff2159SAndi Kleen } 2119e5ff2159SAndi Kleen 21201da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 21216ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 21226ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 21231da177e4SLinus Torvalds { 21244415cc8dSChristoph Lameter int i; 21254415cc8dSChristoph Lameter 2126bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2127aa888a74SAndi Kleen return; 2128aa888a74SAndi Kleen 21296ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 21301da177e4SLinus Torvalds struct page *page, *next; 2131a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 2132a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 2133a5516438SAndi Kleen if (count >= h->nr_huge_pages) 21346b0c880dSAdam Litke return; 21351da177e4SLinus Torvalds if (PageHighMem(page)) 21361da177e4SLinus Torvalds continue; 21371da177e4SLinus Torvalds list_del(&page->lru); 2138e5ff2159SAndi Kleen update_and_free_page(h, page); 2139a5516438SAndi Kleen h->free_huge_pages--; 2140a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 21411da177e4SLinus Torvalds } 21421da177e4SLinus Torvalds } 21431da177e4SLinus Torvalds } 21441da177e4SLinus Torvalds #else 21456ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 21466ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 21471da177e4SLinus Torvalds { 21481da177e4SLinus Torvalds } 21491da177e4SLinus Torvalds #endif 21501da177e4SLinus Torvalds 215120a0307cSWu Fengguang /* 215220a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 215320a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 215420a0307cSWu Fengguang * Returns 1 if an adjustment was made. 215520a0307cSWu Fengguang */ 21566ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 21576ae11b27SLee Schermerhorn int delta) 215820a0307cSWu Fengguang { 2159b2261026SJoonsoo Kim int nr_nodes, node; 216020a0307cSWu Fengguang 216120a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 216220a0307cSWu Fengguang 2163e8c5c824SLee Schermerhorn if (delta < 0) { 2164b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 2165b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 2166b2261026SJoonsoo Kim goto found; 2167b2261026SJoonsoo Kim } 2168b2261026SJoonsoo Kim } else { 2169b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2170b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 2171b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 2172b2261026SJoonsoo Kim goto found; 2173e8c5c824SLee Schermerhorn } 21749a76db09SLee Schermerhorn } 2175b2261026SJoonsoo Kim return 0; 217620a0307cSWu Fengguang 2177b2261026SJoonsoo Kim found: 217820a0307cSWu Fengguang h->surplus_huge_pages += delta; 2179b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 2180b2261026SJoonsoo Kim return 1; 218120a0307cSWu Fengguang } 218220a0307cSWu Fengguang 2183a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 21846ae11b27SLee Schermerhorn static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, 21856ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 21861da177e4SLinus Torvalds { 21877893d1d5SAdam Litke unsigned long min_count, ret; 21881da177e4SLinus Torvalds 2189944d9fecSLuiz Capitulino if (hstate_is_gigantic(h) && !gigantic_page_supported()) 2190aa888a74SAndi Kleen return h->max_huge_pages; 2191aa888a74SAndi Kleen 21927893d1d5SAdam Litke /* 21937893d1d5SAdam Litke * Increase the pool size 21947893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 21957893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 2196d1c3fb1fSNishanth Aravamudan * 2197d15c7c09SNaoya Horiguchi * We might race with __alloc_buddy_huge_page() here and be unable 2198d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 2199d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 2200d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 2201d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 22027893d1d5SAdam Litke */ 22031da177e4SLinus Torvalds spin_lock(&hugetlb_lock); 2204a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 22056ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 22067893d1d5SAdam Litke break; 22077893d1d5SAdam Litke } 22087893d1d5SAdam Litke 2209a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 22107893d1d5SAdam Litke /* 22117893d1d5SAdam Litke * If this allocation races such that we no longer need the 22127893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 22137893d1d5SAdam Litke * and reducing the surplus. 22147893d1d5SAdam Litke */ 22157893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 2216944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) 2217944d9fecSLuiz Capitulino ret = alloc_fresh_gigantic_page(h, nodes_allowed); 2218944d9fecSLuiz Capitulino else 22196ae11b27SLee Schermerhorn ret = alloc_fresh_huge_page(h, nodes_allowed); 22207893d1d5SAdam Litke spin_lock(&hugetlb_lock); 22217893d1d5SAdam Litke if (!ret) 22227893d1d5SAdam Litke goto out; 22237893d1d5SAdam Litke 2224536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 2225536240f2SMel Gorman if (signal_pending(current)) 2226536240f2SMel Gorman goto out; 22277893d1d5SAdam Litke } 22287893d1d5SAdam Litke 22297893d1d5SAdam Litke /* 22307893d1d5SAdam Litke * Decrease the pool size 22317893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 22327893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 22337893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 22347893d1d5SAdam Litke * to the desired size as pages become free. 2235d1c3fb1fSNishanth Aravamudan * 2236d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 2237d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 2238d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 2239d15c7c09SNaoya Horiguchi * __alloc_buddy_huge_page() is checking the global counter, 2240d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 2241d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 2242d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 22437893d1d5SAdam Litke */ 2244a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 22456b0c880dSAdam Litke min_count = max(count, min_count); 22466ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 2247a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 22486ae11b27SLee Schermerhorn if (!free_pool_huge_page(h, nodes_allowed, 0)) 22491da177e4SLinus Torvalds break; 225055f67141SMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 22511da177e4SLinus Torvalds } 2252a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 22536ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 22547893d1d5SAdam Litke break; 22557893d1d5SAdam Litke } 22567893d1d5SAdam Litke out: 2257a5516438SAndi Kleen ret = persistent_huge_pages(h); 22581da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 22597893d1d5SAdam Litke return ret; 22601da177e4SLinus Torvalds } 22611da177e4SLinus Torvalds 2262a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 2263a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 2264a3437870SNishanth Aravamudan 2265a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 2266a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 2267a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 2268a3437870SNishanth Aravamudan 2269a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 2270a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 2271a3437870SNishanth Aravamudan 22729a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 22739a305230SLee Schermerhorn 22749a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 2275a3437870SNishanth Aravamudan { 2276a3437870SNishanth Aravamudan int i; 22779a305230SLee Schermerhorn 2278a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 22799a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 22809a305230SLee Schermerhorn if (nidp) 22819a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 2282a3437870SNishanth Aravamudan return &hstates[i]; 22839a305230SLee Schermerhorn } 22849a305230SLee Schermerhorn 22859a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 2286a3437870SNishanth Aravamudan } 2287a3437870SNishanth Aravamudan 228806808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 2289a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2290a3437870SNishanth Aravamudan { 22919a305230SLee Schermerhorn struct hstate *h; 22929a305230SLee Schermerhorn unsigned long nr_huge_pages; 22939a305230SLee Schermerhorn int nid; 22949a305230SLee Schermerhorn 22959a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 22969a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 22979a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 22989a305230SLee Schermerhorn else 22999a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 23009a305230SLee Schermerhorn 23019a305230SLee Schermerhorn return sprintf(buf, "%lu\n", nr_huge_pages); 2302a3437870SNishanth Aravamudan } 2303adbe8726SEric B Munson 2304238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 2305238d3c13SDavid Rientjes struct hstate *h, int nid, 2306238d3c13SDavid Rientjes unsigned long count, size_t len) 2307a3437870SNishanth Aravamudan { 2308a3437870SNishanth Aravamudan int err; 2309bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); 2310a3437870SNishanth Aravamudan 2311944d9fecSLuiz Capitulino if (hstate_is_gigantic(h) && !gigantic_page_supported()) { 2312adbe8726SEric B Munson err = -EINVAL; 2313adbe8726SEric B Munson goto out; 2314adbe8726SEric B Munson } 2315adbe8726SEric B Munson 23169a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 23179a305230SLee Schermerhorn /* 23189a305230SLee Schermerhorn * global hstate attribute 23199a305230SLee Schermerhorn */ 23209a305230SLee Schermerhorn if (!(obey_mempolicy && 23219a305230SLee Schermerhorn init_nodemask_of_mempolicy(nodes_allowed))) { 232206808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 23238cebfcd0SLai Jiangshan nodes_allowed = &node_states[N_MEMORY]; 232406808b08SLee Schermerhorn } 23259a305230SLee Schermerhorn } else if (nodes_allowed) { 23269a305230SLee Schermerhorn /* 23279a305230SLee Schermerhorn * per node hstate attribute: adjust count to global, 23289a305230SLee Schermerhorn * but restrict alloc/free to the specified node. 23299a305230SLee Schermerhorn */ 23309a305230SLee Schermerhorn count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 23319a305230SLee Schermerhorn init_nodemask_of_node(nodes_allowed, nid); 23329a305230SLee Schermerhorn } else 23338cebfcd0SLai Jiangshan nodes_allowed = &node_states[N_MEMORY]; 23349a305230SLee Schermerhorn 233506808b08SLee Schermerhorn h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); 2336a3437870SNishanth Aravamudan 23378cebfcd0SLai Jiangshan if (nodes_allowed != &node_states[N_MEMORY]) 233806808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 233906808b08SLee Schermerhorn 234006808b08SLee Schermerhorn return len; 2341adbe8726SEric B Munson out: 2342adbe8726SEric B Munson NODEMASK_FREE(nodes_allowed); 2343adbe8726SEric B Munson return err; 234406808b08SLee Schermerhorn } 234506808b08SLee Schermerhorn 2346238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 2347238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 2348238d3c13SDavid Rientjes size_t len) 2349238d3c13SDavid Rientjes { 2350238d3c13SDavid Rientjes struct hstate *h; 2351238d3c13SDavid Rientjes unsigned long count; 2352238d3c13SDavid Rientjes int nid; 2353238d3c13SDavid Rientjes int err; 2354238d3c13SDavid Rientjes 2355238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 2356238d3c13SDavid Rientjes if (err) 2357238d3c13SDavid Rientjes return err; 2358238d3c13SDavid Rientjes 2359238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 2360238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 2361238d3c13SDavid Rientjes } 2362238d3c13SDavid Rientjes 236306808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 236406808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 236506808b08SLee Schermerhorn { 236606808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 236706808b08SLee Schermerhorn } 236806808b08SLee Schermerhorn 236906808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 237006808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 237106808b08SLee Schermerhorn { 2372238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 2373a3437870SNishanth Aravamudan } 2374a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 2375a3437870SNishanth Aravamudan 237606808b08SLee Schermerhorn #ifdef CONFIG_NUMA 237706808b08SLee Schermerhorn 237806808b08SLee Schermerhorn /* 237906808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 238006808b08SLee Schermerhorn * huge page alloc/free. 238106808b08SLee Schermerhorn */ 238206808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 238306808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 238406808b08SLee Schermerhorn { 238506808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 238606808b08SLee Schermerhorn } 238706808b08SLee Schermerhorn 238806808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 238906808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 239006808b08SLee Schermerhorn { 2391238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 239206808b08SLee Schermerhorn } 239306808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 239406808b08SLee Schermerhorn #endif 239506808b08SLee Schermerhorn 239606808b08SLee Schermerhorn 2397a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 2398a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2399a3437870SNishanth Aravamudan { 24009a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2401a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 2402a3437870SNishanth Aravamudan } 2403adbe8726SEric B Munson 2404a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 2405a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 2406a3437870SNishanth Aravamudan { 2407a3437870SNishanth Aravamudan int err; 2408a3437870SNishanth Aravamudan unsigned long input; 24099a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2410a3437870SNishanth Aravamudan 2411bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2412adbe8726SEric B Munson return -EINVAL; 2413adbe8726SEric B Munson 24143dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 2415a3437870SNishanth Aravamudan if (err) 241673ae31e5SEric B Munson return err; 2417a3437870SNishanth Aravamudan 2418a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 2419a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 2420a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 2421a3437870SNishanth Aravamudan 2422a3437870SNishanth Aravamudan return count; 2423a3437870SNishanth Aravamudan } 2424a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 2425a3437870SNishanth Aravamudan 2426a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 2427a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2428a3437870SNishanth Aravamudan { 24299a305230SLee Schermerhorn struct hstate *h; 24309a305230SLee Schermerhorn unsigned long free_huge_pages; 24319a305230SLee Schermerhorn int nid; 24329a305230SLee Schermerhorn 24339a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 24349a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 24359a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 24369a305230SLee Schermerhorn else 24379a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 24389a305230SLee Schermerhorn 24399a305230SLee Schermerhorn return sprintf(buf, "%lu\n", free_huge_pages); 2440a3437870SNishanth Aravamudan } 2441a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 2442a3437870SNishanth Aravamudan 2443a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 2444a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2445a3437870SNishanth Aravamudan { 24469a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2447a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->resv_huge_pages); 2448a3437870SNishanth Aravamudan } 2449a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 2450a3437870SNishanth Aravamudan 2451a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 2452a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2453a3437870SNishanth Aravamudan { 24549a305230SLee Schermerhorn struct hstate *h; 24559a305230SLee Schermerhorn unsigned long surplus_huge_pages; 24569a305230SLee Schermerhorn int nid; 24579a305230SLee Schermerhorn 24589a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 24599a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 24609a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 24619a305230SLee Schermerhorn else 24629a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 24639a305230SLee Schermerhorn 24649a305230SLee Schermerhorn return sprintf(buf, "%lu\n", surplus_huge_pages); 2465a3437870SNishanth Aravamudan } 2466a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 2467a3437870SNishanth Aravamudan 2468a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 2469a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 2470a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 2471a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 2472a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 2473a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 247406808b08SLee Schermerhorn #ifdef CONFIG_NUMA 247506808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 247606808b08SLee Schermerhorn #endif 2477a3437870SNishanth Aravamudan NULL, 2478a3437870SNishanth Aravamudan }; 2479a3437870SNishanth Aravamudan 2480a3437870SNishanth Aravamudan static struct attribute_group hstate_attr_group = { 2481a3437870SNishanth Aravamudan .attrs = hstate_attrs, 2482a3437870SNishanth Aravamudan }; 2483a3437870SNishanth Aravamudan 2484094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 24859a305230SLee Schermerhorn struct kobject **hstate_kobjs, 24869a305230SLee Schermerhorn struct attribute_group *hstate_attr_group) 2487a3437870SNishanth Aravamudan { 2488a3437870SNishanth Aravamudan int retval; 2489972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 2490a3437870SNishanth Aravamudan 24919a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 24929a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 2493a3437870SNishanth Aravamudan return -ENOMEM; 2494a3437870SNishanth Aravamudan 24959a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 2496a3437870SNishanth Aravamudan if (retval) 24979a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 2498a3437870SNishanth Aravamudan 2499a3437870SNishanth Aravamudan return retval; 2500a3437870SNishanth Aravamudan } 2501a3437870SNishanth Aravamudan 2502a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 2503a3437870SNishanth Aravamudan { 2504a3437870SNishanth Aravamudan struct hstate *h; 2505a3437870SNishanth Aravamudan int err; 2506a3437870SNishanth Aravamudan 2507a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 2508a3437870SNishanth Aravamudan if (!hugepages_kobj) 2509a3437870SNishanth Aravamudan return; 2510a3437870SNishanth Aravamudan 2511a3437870SNishanth Aravamudan for_each_hstate(h) { 25129a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 25139a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 2514a3437870SNishanth Aravamudan if (err) 2515ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s", h->name); 2516a3437870SNishanth Aravamudan } 2517a3437870SNishanth Aravamudan } 2518a3437870SNishanth Aravamudan 25199a305230SLee Schermerhorn #ifdef CONFIG_NUMA 25209a305230SLee Schermerhorn 25219a305230SLee Schermerhorn /* 25229a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 252310fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 252410fbcf4cSKay Sievers * index of a node device or _hstate == node id. 252510fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 25269a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 25279a305230SLee Schermerhorn */ 25289a305230SLee Schermerhorn struct node_hstate { 25299a305230SLee Schermerhorn struct kobject *hugepages_kobj; 25309a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 25319a305230SLee Schermerhorn }; 2532b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 25339a305230SLee Schermerhorn 25349a305230SLee Schermerhorn /* 253510fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 25369a305230SLee Schermerhorn */ 25379a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 25389a305230SLee Schermerhorn &nr_hugepages_attr.attr, 25399a305230SLee Schermerhorn &free_hugepages_attr.attr, 25409a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 25419a305230SLee Schermerhorn NULL, 25429a305230SLee Schermerhorn }; 25439a305230SLee Schermerhorn 25449a305230SLee Schermerhorn static struct attribute_group per_node_hstate_attr_group = { 25459a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 25469a305230SLee Schermerhorn }; 25479a305230SLee Schermerhorn 25489a305230SLee Schermerhorn /* 254910fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 25509a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 25519a305230SLee Schermerhorn */ 25529a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 25539a305230SLee Schermerhorn { 25549a305230SLee Schermerhorn int nid; 25559a305230SLee Schermerhorn 25569a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 25579a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 25589a305230SLee Schermerhorn int i; 25599a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 25609a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 25619a305230SLee Schermerhorn if (nidp) 25629a305230SLee Schermerhorn *nidp = nid; 25639a305230SLee Schermerhorn return &hstates[i]; 25649a305230SLee Schermerhorn } 25659a305230SLee Schermerhorn } 25669a305230SLee Schermerhorn 25679a305230SLee Schermerhorn BUG(); 25689a305230SLee Schermerhorn return NULL; 25699a305230SLee Schermerhorn } 25709a305230SLee Schermerhorn 25719a305230SLee Schermerhorn /* 257210fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 25739a305230SLee Schermerhorn * No-op if no hstate attributes attached. 25749a305230SLee Schermerhorn */ 25753cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 25769a305230SLee Schermerhorn { 25779a305230SLee Schermerhorn struct hstate *h; 257810fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 25799a305230SLee Schermerhorn 25809a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 25819b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 25829a305230SLee Schermerhorn 2583972dc4deSAneesh Kumar K.V for_each_hstate(h) { 2584972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 2585972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 2586972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 2587972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 2588972dc4deSAneesh Kumar K.V } 25899a305230SLee Schermerhorn } 25909a305230SLee Schermerhorn 25919a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 25929a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 25939a305230SLee Schermerhorn } 25949a305230SLee Schermerhorn 25959a305230SLee Schermerhorn 25969a305230SLee Schermerhorn /* 259710fbcf4cSKay Sievers * Register hstate attributes for a single node device. 25989a305230SLee Schermerhorn * No-op if attributes already registered. 25999a305230SLee Schermerhorn */ 26003cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 26019a305230SLee Schermerhorn { 26029a305230SLee Schermerhorn struct hstate *h; 260310fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 26049a305230SLee Schermerhorn int err; 26059a305230SLee Schermerhorn 26069a305230SLee Schermerhorn if (nhs->hugepages_kobj) 26079a305230SLee Schermerhorn return; /* already allocated */ 26089a305230SLee Schermerhorn 26099a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 261010fbcf4cSKay Sievers &node->dev.kobj); 26119a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 26129a305230SLee Schermerhorn return; 26139a305230SLee Schermerhorn 26149a305230SLee Schermerhorn for_each_hstate(h) { 26159a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 26169a305230SLee Schermerhorn nhs->hstate_kobjs, 26179a305230SLee Schermerhorn &per_node_hstate_attr_group); 26189a305230SLee Schermerhorn if (err) { 2619ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s for node %d\n", 262010fbcf4cSKay Sievers h->name, node->dev.id); 26219a305230SLee Schermerhorn hugetlb_unregister_node(node); 26229a305230SLee Schermerhorn break; 26239a305230SLee Schermerhorn } 26249a305230SLee Schermerhorn } 26259a305230SLee Schermerhorn } 26269a305230SLee Schermerhorn 26279a305230SLee Schermerhorn /* 26289b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 262910fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 263010fbcf4cSKay Sievers * registered their associated device by this time. 26319a305230SLee Schermerhorn */ 26327d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 26339a305230SLee Schermerhorn { 26349a305230SLee Schermerhorn int nid; 26359a305230SLee Schermerhorn 26368cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 26378732794bSWen Congyang struct node *node = node_devices[nid]; 263810fbcf4cSKay Sievers if (node->dev.id == nid) 26399a305230SLee Schermerhorn hugetlb_register_node(node); 26409a305230SLee Schermerhorn } 26419a305230SLee Schermerhorn 26429a305230SLee Schermerhorn /* 264310fbcf4cSKay Sievers * Let the node device driver know we're here so it can 26449a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 26459a305230SLee Schermerhorn */ 26469a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 26479a305230SLee Schermerhorn hugetlb_unregister_node); 26489a305230SLee Schermerhorn } 26499a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 26509a305230SLee Schermerhorn 26519a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 26529a305230SLee Schermerhorn { 26539a305230SLee Schermerhorn BUG(); 26549a305230SLee Schermerhorn if (nidp) 26559a305230SLee Schermerhorn *nidp = -1; 26569a305230SLee Schermerhorn return NULL; 26579a305230SLee Schermerhorn } 26589a305230SLee Schermerhorn 26599a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 26609a305230SLee Schermerhorn 26619a305230SLee Schermerhorn #endif 26629a305230SLee Schermerhorn 2663a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 2664a3437870SNishanth Aravamudan { 26658382d914SDavidlohr Bueso int i; 26668382d914SDavidlohr Bueso 2667457c1b27SNishanth Aravamudan if (!hugepages_supported()) 26680ef89d25SBenjamin Herrenschmidt return 0; 2669a3437870SNishanth Aravamudan 2670e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) { 2671e11bfbfcSNick Piggin default_hstate_size = HPAGE_SIZE; 2672e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) 2673a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 2674a3437870SNishanth Aravamudan } 2675972dc4deSAneesh Kumar K.V default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); 2676f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 2677f8b74815SVaishali Thakkar if (!default_hstate.max_huge_pages) 2678e11bfbfcSNick Piggin default_hstate.max_huge_pages = default_hstate_max_huge_pages; 2679f8b74815SVaishali Thakkar } 2680a3437870SNishanth Aravamudan 2681a3437870SNishanth Aravamudan hugetlb_init_hstates(); 2682aa888a74SAndi Kleen gather_bootmem_prealloc(); 2683a3437870SNishanth Aravamudan report_hugepages(); 2684a3437870SNishanth Aravamudan 2685a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 26869a305230SLee Schermerhorn hugetlb_register_all_nodes(); 26877179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 26889a305230SLee Schermerhorn 26898382d914SDavidlohr Bueso #ifdef CONFIG_SMP 26908382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 26918382d914SDavidlohr Bueso #else 26928382d914SDavidlohr Bueso num_fault_mutexes = 1; 26938382d914SDavidlohr Bueso #endif 2694c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 26958382d914SDavidlohr Bueso kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL); 2696c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 26978382d914SDavidlohr Bueso 26988382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 2699c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 2700a3437870SNishanth Aravamudan return 0; 2701a3437870SNishanth Aravamudan } 27023e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 2703a3437870SNishanth Aravamudan 2704a3437870SNishanth Aravamudan /* Should be called on processing a hugepagesz=... option */ 27059fee021dSVaishali Thakkar void __init hugetlb_bad_size(void) 27069fee021dSVaishali Thakkar { 27079fee021dSVaishali Thakkar parsed_valid_hugepagesz = false; 27089fee021dSVaishali Thakkar } 27099fee021dSVaishali Thakkar 2710d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 2711a3437870SNishanth Aravamudan { 2712a3437870SNishanth Aravamudan struct hstate *h; 27138faa8b07SAndi Kleen unsigned long i; 27148faa8b07SAndi Kleen 2715a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 2716598d8091SJoe Perches pr_warn("hugepagesz= specified twice, ignoring\n"); 2717a3437870SNishanth Aravamudan return; 2718a3437870SNishanth Aravamudan } 271947d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 2720a3437870SNishanth Aravamudan BUG_ON(order == 0); 272147d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 2722a3437870SNishanth Aravamudan h->order = order; 2723a3437870SNishanth Aravamudan h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 27248faa8b07SAndi Kleen h->nr_huge_pages = 0; 27258faa8b07SAndi Kleen h->free_huge_pages = 0; 27268faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 27278faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 27280edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 272954f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 273054f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 2731a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 2732a3437870SNishanth Aravamudan huge_page_size(h)/1024); 27338faa8b07SAndi Kleen 2734a3437870SNishanth Aravamudan parsed_hstate = h; 2735a3437870SNishanth Aravamudan } 2736a3437870SNishanth Aravamudan 2737e11bfbfcSNick Piggin static int __init hugetlb_nrpages_setup(char *s) 2738a3437870SNishanth Aravamudan { 2739a3437870SNishanth Aravamudan unsigned long *mhp; 27408faa8b07SAndi Kleen static unsigned long *last_mhp; 2741a3437870SNishanth Aravamudan 27429fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 27439fee021dSVaishali Thakkar pr_warn("hugepages = %s preceded by " 27449fee021dSVaishali Thakkar "an unsupported hugepagesz, ignoring\n", s); 27459fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 27469fee021dSVaishali Thakkar return 1; 27479fee021dSVaishali Thakkar } 2748a3437870SNishanth Aravamudan /* 274947d38344SAneesh Kumar K.V * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, 2750a3437870SNishanth Aravamudan * so this hugepages= parameter goes to the "default hstate". 2751a3437870SNishanth Aravamudan */ 27529fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 2753a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 2754a3437870SNishanth Aravamudan else 2755a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 2756a3437870SNishanth Aravamudan 27578faa8b07SAndi Kleen if (mhp == last_mhp) { 2758598d8091SJoe Perches pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n"); 27598faa8b07SAndi Kleen return 1; 27608faa8b07SAndi Kleen } 27618faa8b07SAndi Kleen 2762a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 2763a3437870SNishanth Aravamudan *mhp = 0; 2764a3437870SNishanth Aravamudan 27658faa8b07SAndi Kleen /* 27668faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 27678faa8b07SAndi Kleen * But we need to allocate >= MAX_ORDER hstates here early to still 27688faa8b07SAndi Kleen * use the bootmem allocator. 27698faa8b07SAndi Kleen */ 277047d38344SAneesh Kumar K.V if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) 27718faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 27728faa8b07SAndi Kleen 27738faa8b07SAndi Kleen last_mhp = mhp; 27748faa8b07SAndi Kleen 2775a3437870SNishanth Aravamudan return 1; 2776a3437870SNishanth Aravamudan } 2777e11bfbfcSNick Piggin __setup("hugepages=", hugetlb_nrpages_setup); 2778e11bfbfcSNick Piggin 2779e11bfbfcSNick Piggin static int __init hugetlb_default_setup(char *s) 2780e11bfbfcSNick Piggin { 2781e11bfbfcSNick Piggin default_hstate_size = memparse(s, &s); 2782e11bfbfcSNick Piggin return 1; 2783e11bfbfcSNick Piggin } 2784e11bfbfcSNick Piggin __setup("default_hugepagesz=", hugetlb_default_setup); 2785a3437870SNishanth Aravamudan 27868a213460SNishanth Aravamudan static unsigned int cpuset_mems_nr(unsigned int *array) 27878a213460SNishanth Aravamudan { 27888a213460SNishanth Aravamudan int node; 27898a213460SNishanth Aravamudan unsigned int nr = 0; 27908a213460SNishanth Aravamudan 27918a213460SNishanth Aravamudan for_each_node_mask(node, cpuset_current_mems_allowed) 27928a213460SNishanth Aravamudan nr += array[node]; 27938a213460SNishanth Aravamudan 27948a213460SNishanth Aravamudan return nr; 27958a213460SNishanth Aravamudan } 27968a213460SNishanth Aravamudan 27978a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 279806808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 279906808b08SLee Schermerhorn struct ctl_table *table, int write, 280006808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 28011da177e4SLinus Torvalds { 2802e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 2803238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 280408d4a246SMichal Hocko int ret; 2805e5ff2159SAndi Kleen 2806457c1b27SNishanth Aravamudan if (!hugepages_supported()) 280786613628SJan Stancek return -EOPNOTSUPP; 2808457c1b27SNishanth Aravamudan 2809e5ff2159SAndi Kleen table->data = &tmp; 2810e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 281108d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 281208d4a246SMichal Hocko if (ret) 281308d4a246SMichal Hocko goto out; 2814e5ff2159SAndi Kleen 2815238d3c13SDavid Rientjes if (write) 2816238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 2817238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 281808d4a246SMichal Hocko out: 281908d4a246SMichal Hocko return ret; 28201da177e4SLinus Torvalds } 2821396faf03SMel Gorman 282206808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 282306808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 282406808b08SLee Schermerhorn { 282506808b08SLee Schermerhorn 282606808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 282706808b08SLee Schermerhorn buffer, length, ppos); 282806808b08SLee Schermerhorn } 282906808b08SLee Schermerhorn 283006808b08SLee Schermerhorn #ifdef CONFIG_NUMA 283106808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 283206808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 283306808b08SLee Schermerhorn { 283406808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 283506808b08SLee Schermerhorn buffer, length, ppos); 283606808b08SLee Schermerhorn } 283706808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 283806808b08SLee Schermerhorn 2839a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 28408d65af78SAlexey Dobriyan void __user *buffer, 2841a3d0c6aaSNishanth Aravamudan size_t *length, loff_t *ppos) 2842a3d0c6aaSNishanth Aravamudan { 2843a5516438SAndi Kleen struct hstate *h = &default_hstate; 2844e5ff2159SAndi Kleen unsigned long tmp; 284508d4a246SMichal Hocko int ret; 2846e5ff2159SAndi Kleen 2847457c1b27SNishanth Aravamudan if (!hugepages_supported()) 284886613628SJan Stancek return -EOPNOTSUPP; 2849457c1b27SNishanth Aravamudan 2850e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 2851e5ff2159SAndi Kleen 2852bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 2853adbe8726SEric B Munson return -EINVAL; 2854adbe8726SEric B Munson 2855e5ff2159SAndi Kleen table->data = &tmp; 2856e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 285708d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 285808d4a246SMichal Hocko if (ret) 285908d4a246SMichal Hocko goto out; 2860e5ff2159SAndi Kleen 2861e5ff2159SAndi Kleen if (write) { 2862064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 2863e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 2864a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 2865e5ff2159SAndi Kleen } 286608d4a246SMichal Hocko out: 286708d4a246SMichal Hocko return ret; 2868a3d0c6aaSNishanth Aravamudan } 2869a3d0c6aaSNishanth Aravamudan 28701da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 28711da177e4SLinus Torvalds 2872e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 28731da177e4SLinus Torvalds { 2874a5516438SAndi Kleen struct hstate *h = &default_hstate; 2875457c1b27SNishanth Aravamudan if (!hugepages_supported()) 2876457c1b27SNishanth Aravamudan return; 2877e1759c21SAlexey Dobriyan seq_printf(m, 28781da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 28791da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 2880b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 28817893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 28824f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 2883a5516438SAndi Kleen h->nr_huge_pages, 2884a5516438SAndi Kleen h->free_huge_pages, 2885a5516438SAndi Kleen h->resv_huge_pages, 2886a5516438SAndi Kleen h->surplus_huge_pages, 2887a5516438SAndi Kleen 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 28881da177e4SLinus Torvalds } 28891da177e4SLinus Torvalds 28901da177e4SLinus Torvalds int hugetlb_report_node_meminfo(int nid, char *buf) 28911da177e4SLinus Torvalds { 2892a5516438SAndi Kleen struct hstate *h = &default_hstate; 2893457c1b27SNishanth Aravamudan if (!hugepages_supported()) 2894457c1b27SNishanth Aravamudan return 0; 28951da177e4SLinus Torvalds return sprintf(buf, 28961da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 2897a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 2898a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 2899a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 2900a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 2901a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 29021da177e4SLinus Torvalds } 29031da177e4SLinus Torvalds 2904949f7ec5SDavid Rientjes void hugetlb_show_meminfo(void) 2905949f7ec5SDavid Rientjes { 2906949f7ec5SDavid Rientjes struct hstate *h; 2907949f7ec5SDavid Rientjes int nid; 2908949f7ec5SDavid Rientjes 2909457c1b27SNishanth Aravamudan if (!hugepages_supported()) 2910457c1b27SNishanth Aravamudan return; 2911457c1b27SNishanth Aravamudan 2912949f7ec5SDavid Rientjes for_each_node_state(nid, N_MEMORY) 2913949f7ec5SDavid Rientjes for_each_hstate(h) 2914949f7ec5SDavid Rientjes pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 2915949f7ec5SDavid Rientjes nid, 2916949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 2917949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 2918949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 2919949f7ec5SDavid Rientjes 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 2920949f7ec5SDavid Rientjes } 2921949f7ec5SDavid Rientjes 29225d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 29235d317b2bSNaoya Horiguchi { 29245d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 29255d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 29265d317b2bSNaoya Horiguchi } 29275d317b2bSNaoya Horiguchi 29281da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 29291da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 29301da177e4SLinus Torvalds { 2931d0028588SWanpeng Li struct hstate *h; 2932d0028588SWanpeng Li unsigned long nr_total_pages = 0; 2933d0028588SWanpeng Li 2934d0028588SWanpeng Li for_each_hstate(h) 2935d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 2936d0028588SWanpeng Li return nr_total_pages; 29371da177e4SLinus Torvalds } 29381da177e4SLinus Torvalds 2939a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 2940fc1b8a73SMel Gorman { 2941fc1b8a73SMel Gorman int ret = -ENOMEM; 2942fc1b8a73SMel Gorman 2943fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 2944fc1b8a73SMel Gorman /* 2945fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 2946fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 2947fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 2948fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 2949fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 2950fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 2951fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 2952fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 2953fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 2954fc1b8a73SMel Gorman * 2955fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 2956fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 2957fc1b8a73SMel Gorman * we fall back to check against current free page availability as 2958fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 2959fc1b8a73SMel Gorman * semantics that cpuset has. 2960fc1b8a73SMel Gorman */ 2961fc1b8a73SMel Gorman if (delta > 0) { 2962a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 2963fc1b8a73SMel Gorman goto out; 2964fc1b8a73SMel Gorman 2965a5516438SAndi Kleen if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 2966a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 2967fc1b8a73SMel Gorman goto out; 2968fc1b8a73SMel Gorman } 2969fc1b8a73SMel Gorman } 2970fc1b8a73SMel Gorman 2971fc1b8a73SMel Gorman ret = 0; 2972fc1b8a73SMel Gorman if (delta < 0) 2973a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 2974fc1b8a73SMel Gorman 2975fc1b8a73SMel Gorman out: 2976fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 2977fc1b8a73SMel Gorman return ret; 2978fc1b8a73SMel Gorman } 2979fc1b8a73SMel Gorman 298084afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 298184afd99bSAndy Whitcroft { 2982f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 298384afd99bSAndy Whitcroft 298484afd99bSAndy Whitcroft /* 298584afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 298684afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 298784afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 298825985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 298984afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 299084afd99bSAndy Whitcroft * new reference here without additional locking. 299184afd99bSAndy Whitcroft */ 29924e35f483SJoonsoo Kim if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 2993f522c3acSJoonsoo Kim kref_get(&resv->refs); 299484afd99bSAndy Whitcroft } 299584afd99bSAndy Whitcroft 2996a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 2997a1e78772SMel Gorman { 2998a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2999f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 300090481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 30014e35f483SJoonsoo Kim unsigned long reserve, start, end; 30021c5ecae3SMike Kravetz long gbl_reserve; 300384afd99bSAndy Whitcroft 30044e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 30054e35f483SJoonsoo Kim return; 30064e35f483SJoonsoo Kim 3007a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 3008a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 300984afd99bSAndy Whitcroft 30104e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 301184afd99bSAndy Whitcroft 3012f031dd27SJoonsoo Kim kref_put(&resv->refs, resv_map_release); 301384afd99bSAndy Whitcroft 30147251ff78SAdam Litke if (reserve) { 30151c5ecae3SMike Kravetz /* 30161c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 30171c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 30181c5ecae3SMike Kravetz */ 30191c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 30201c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 30217251ff78SAdam Litke } 3022a1e78772SMel Gorman } 3023a1e78772SMel Gorman 30241da177e4SLinus Torvalds /* 30251da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 30261da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 30271da177e4SLinus Torvalds * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 30281da177e4SLinus Torvalds * this far. 30291da177e4SLinus Torvalds */ 3030d0217ac0SNick Piggin static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) 30311da177e4SLinus Torvalds { 30321da177e4SLinus Torvalds BUG(); 3033d0217ac0SNick Piggin return 0; 30341da177e4SLinus Torvalds } 30351da177e4SLinus Torvalds 3036f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 3037d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 303884afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 3039a1e78772SMel Gorman .close = hugetlb_vm_op_close, 30401da177e4SLinus Torvalds }; 30411da177e4SLinus Torvalds 30421e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 30431e8f889bSDavid Gibson int writable) 304463551ae0SDavid Gibson { 304563551ae0SDavid Gibson pte_t entry; 304663551ae0SDavid Gibson 30471e8f889bSDavid Gibson if (writable) { 3048106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 3049106c992aSGerald Schaefer vma->vm_page_prot))); 305063551ae0SDavid Gibson } else { 3051106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 3052106c992aSGerald Schaefer vma->vm_page_prot)); 305363551ae0SDavid Gibson } 305463551ae0SDavid Gibson entry = pte_mkyoung(entry); 305563551ae0SDavid Gibson entry = pte_mkhuge(entry); 3056d9ed9faaSChris Metcalf entry = arch_make_huge_pte(entry, vma, page, writable); 305763551ae0SDavid Gibson 305863551ae0SDavid Gibson return entry; 305963551ae0SDavid Gibson } 306063551ae0SDavid Gibson 30611e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 30621e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 30631e8f889bSDavid Gibson { 30641e8f889bSDavid Gibson pte_t entry; 30651e8f889bSDavid Gibson 3066106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 306732f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 30684b3073e1SRussell King update_mmu_cache(vma, address, ptep); 30691e8f889bSDavid Gibson } 30701e8f889bSDavid Gibson 30714a705fefSNaoya Horiguchi static int is_hugetlb_entry_migration(pte_t pte) 30724a705fefSNaoya Horiguchi { 30734a705fefSNaoya Horiguchi swp_entry_t swp; 30744a705fefSNaoya Horiguchi 30754a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 30764a705fefSNaoya Horiguchi return 0; 30774a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 30784a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_migration_entry(swp)) 30794a705fefSNaoya Horiguchi return 1; 30804a705fefSNaoya Horiguchi else 30814a705fefSNaoya Horiguchi return 0; 30824a705fefSNaoya Horiguchi } 30834a705fefSNaoya Horiguchi 30844a705fefSNaoya Horiguchi static int is_hugetlb_entry_hwpoisoned(pte_t pte) 30854a705fefSNaoya Horiguchi { 30864a705fefSNaoya Horiguchi swp_entry_t swp; 30874a705fefSNaoya Horiguchi 30884a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 30894a705fefSNaoya Horiguchi return 0; 30904a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 30914a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_hwpoison_entry(swp)) 30924a705fefSNaoya Horiguchi return 1; 30934a705fefSNaoya Horiguchi else 30944a705fefSNaoya Horiguchi return 0; 30954a705fefSNaoya Horiguchi } 30961e8f889bSDavid Gibson 309763551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 309863551ae0SDavid Gibson struct vm_area_struct *vma) 309963551ae0SDavid Gibson { 310063551ae0SDavid Gibson pte_t *src_pte, *dst_pte, entry; 310163551ae0SDavid Gibson struct page *ptepage; 31021c59827dSHugh Dickins unsigned long addr; 31031e8f889bSDavid Gibson int cow; 3104a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3105a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 3106e8569dd2SAndreas Sandberg unsigned long mmun_start; /* For mmu_notifiers */ 3107e8569dd2SAndreas Sandberg unsigned long mmun_end; /* For mmu_notifiers */ 3108e8569dd2SAndreas Sandberg int ret = 0; 31091e8f889bSDavid Gibson 31101e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 311163551ae0SDavid Gibson 3112e8569dd2SAndreas Sandberg mmun_start = vma->vm_start; 3113e8569dd2SAndreas Sandberg mmun_end = vma->vm_end; 3114e8569dd2SAndreas Sandberg if (cow) 3115e8569dd2SAndreas Sandberg mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end); 3116e8569dd2SAndreas Sandberg 3117a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 3118cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 3119c74df32cSHugh Dickins src_pte = huge_pte_offset(src, addr); 3120c74df32cSHugh Dickins if (!src_pte) 3121c74df32cSHugh Dickins continue; 3122a5516438SAndi Kleen dst_pte = huge_pte_alloc(dst, addr, sz); 3123e8569dd2SAndreas Sandberg if (!dst_pte) { 3124e8569dd2SAndreas Sandberg ret = -ENOMEM; 3125e8569dd2SAndreas Sandberg break; 3126e8569dd2SAndreas Sandberg } 3127c5c99429SLarry Woodman 3128c5c99429SLarry Woodman /* If the pagetables are shared don't copy or take references */ 3129c5c99429SLarry Woodman if (dst_pte == src_pte) 3130c5c99429SLarry Woodman continue; 3131c5c99429SLarry Woodman 3132cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 3133cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 3134cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 31354a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 31364a705fefSNaoya Horiguchi if (huge_pte_none(entry)) { /* skip none entry */ 31374a705fefSNaoya Horiguchi ; 31384a705fefSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry) || 31394a705fefSNaoya Horiguchi is_hugetlb_entry_hwpoisoned(entry))) { 31404a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 31414a705fefSNaoya Horiguchi 31424a705fefSNaoya Horiguchi if (is_write_migration_entry(swp_entry) && cow) { 31434a705fefSNaoya Horiguchi /* 31444a705fefSNaoya Horiguchi * COW mappings require pages in both 31454a705fefSNaoya Horiguchi * parent and child to be set to read. 31464a705fefSNaoya Horiguchi */ 31474a705fefSNaoya Horiguchi make_migration_entry_read(&swp_entry); 31484a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 31494a705fefSNaoya Horiguchi set_huge_pte_at(src, addr, src_pte, entry); 31504a705fefSNaoya Horiguchi } 31514a705fefSNaoya Horiguchi set_huge_pte_at(dst, addr, dst_pte, entry); 31524a705fefSNaoya Horiguchi } else { 315334ee645eSJoerg Roedel if (cow) { 31547f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 315534ee645eSJoerg Roedel mmu_notifier_invalidate_range(src, mmun_start, 315634ee645eSJoerg Roedel mmun_end); 315734ee645eSJoerg Roedel } 31580253d634SNaoya Horiguchi entry = huge_ptep_get(src_pte); 315963551ae0SDavid Gibson ptepage = pte_page(entry); 316063551ae0SDavid Gibson get_page(ptepage); 316153f9263bSKirill A. Shutemov page_dup_rmap(ptepage, true); 316263551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 31635d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), dst); 31641c59827dSHugh Dickins } 3165cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 3166cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 316763551ae0SDavid Gibson } 316863551ae0SDavid Gibson 3169e8569dd2SAndreas Sandberg if (cow) 3170e8569dd2SAndreas Sandberg mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end); 3171e8569dd2SAndreas Sandberg 3172e8569dd2SAndreas Sandberg return ret; 317363551ae0SDavid Gibson } 317463551ae0SDavid Gibson 317524669e58SAneesh Kumar K.V void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 317624669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 317724669e58SAneesh Kumar K.V struct page *ref_page) 317863551ae0SDavid Gibson { 317924669e58SAneesh Kumar K.V int force_flush = 0; 318063551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 318163551ae0SDavid Gibson unsigned long address; 3182c7546f8fSDavid Gibson pte_t *ptep; 318363551ae0SDavid Gibson pte_t pte; 3184cb900f41SKirill A. Shutemov spinlock_t *ptl; 318563551ae0SDavid Gibson struct page *page; 3186a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3187a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 31882ec74c3eSSagi Grimberg const unsigned long mmun_start = start; /* For mmu_notifiers */ 31892ec74c3eSSagi Grimberg const unsigned long mmun_end = end; /* For mmu_notifiers */ 3190a5516438SAndi Kleen 319163551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 3192a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 3193a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 319463551ae0SDavid Gibson 319524669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 31962ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 3197569f48b8SHillf Danton address = start; 319824669e58SAneesh Kumar K.V again: 3199569f48b8SHillf Danton for (; address < end; address += sz) { 3200c7546f8fSDavid Gibson ptep = huge_pte_offset(mm, address); 3201c7546f8fSDavid Gibson if (!ptep) 3202c7546f8fSDavid Gibson continue; 3203c7546f8fSDavid Gibson 3204cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 320539dde65cSChen, Kenneth W if (huge_pmd_unshare(mm, &address, ptep)) 3206cb900f41SKirill A. Shutemov goto unlock; 320739dde65cSChen, Kenneth W 32086629326bSHillf Danton pte = huge_ptep_get(ptep); 32096629326bSHillf Danton if (huge_pte_none(pte)) 3210cb900f41SKirill A. Shutemov goto unlock; 32116629326bSHillf Danton 32126629326bSHillf Danton /* 32139fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 32149fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 32156629326bSHillf Danton */ 32169fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 3217106c992aSGerald Schaefer huge_pte_clear(mm, address, ptep); 3218cb900f41SKirill A. Shutemov goto unlock; 32198c4894c6SNaoya Horiguchi } 32206629326bSHillf Danton 32216629326bSHillf Danton page = pte_page(pte); 322204f2cbe3SMel Gorman /* 322304f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 322404f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 322504f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 322604f2cbe3SMel Gorman */ 322704f2cbe3SMel Gorman if (ref_page) { 322804f2cbe3SMel Gorman if (page != ref_page) 3229cb900f41SKirill A. Shutemov goto unlock; 323004f2cbe3SMel Gorman 323104f2cbe3SMel Gorman /* 323204f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 323304f2cbe3SMel Gorman * future faults in this VMA will fail rather than 323404f2cbe3SMel Gorman * looking like data was lost 323504f2cbe3SMel Gorman */ 323604f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 323704f2cbe3SMel Gorman } 323804f2cbe3SMel Gorman 3239c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 324024669e58SAneesh Kumar K.V tlb_remove_tlb_entry(tlb, ptep, address); 3241106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 32426649a386SKen Chen set_page_dirty(page); 32439e81130bSHillf Danton 32445d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 3245d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 324624669e58SAneesh Kumar K.V force_flush = !__tlb_remove_page(tlb, page); 3247cb900f41SKirill A. Shutemov if (force_flush) { 3248569f48b8SHillf Danton address += sz; 3249cb900f41SKirill A. Shutemov spin_unlock(ptl); 32509e81130bSHillf Danton break; 325163551ae0SDavid Gibson } 3252cb900f41SKirill A. Shutemov /* Bail out after unmapping reference page if supplied */ 3253cb900f41SKirill A. Shutemov if (ref_page) { 3254cb900f41SKirill A. Shutemov spin_unlock(ptl); 3255cb900f41SKirill A. Shutemov break; 3256cb900f41SKirill A. Shutemov } 3257cb900f41SKirill A. Shutemov unlock: 3258cb900f41SKirill A. Shutemov spin_unlock(ptl); 3259cb900f41SKirill A. Shutemov } 326024669e58SAneesh Kumar K.V /* 326124669e58SAneesh Kumar K.V * mmu_gather ran out of room to batch pages, we break out of 326224669e58SAneesh Kumar K.V * the PTE lock to avoid doing the potential expensive TLB invalidate 326324669e58SAneesh Kumar K.V * and page-free while holding it. 326424669e58SAneesh Kumar K.V */ 326524669e58SAneesh Kumar K.V if (force_flush) { 326624669e58SAneesh Kumar K.V force_flush = 0; 326724669e58SAneesh Kumar K.V tlb_flush_mmu(tlb); 326824669e58SAneesh Kumar K.V if (address < end && !ref_page) 326924669e58SAneesh Kumar K.V goto again; 3270fe1668aeSChen, Kenneth W } 32712ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 327224669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 32731da177e4SLinus Torvalds } 327463551ae0SDavid Gibson 3275d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 3276d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 3277d833352aSMel Gorman unsigned long end, struct page *ref_page) 3278d833352aSMel Gorman { 3279d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 3280d833352aSMel Gorman 3281d833352aSMel Gorman /* 3282d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 3283d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 3284d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 3285d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 3286d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 3287c8c06efaSDavidlohr Bueso * is to clear it before releasing the i_mmap_rwsem. This works 3288d833352aSMel Gorman * because in the context this is called, the VMA is about to be 3289c8c06efaSDavidlohr Bueso * destroyed and the i_mmap_rwsem is held. 3290d833352aSMel Gorman */ 3291d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 3292d833352aSMel Gorman } 3293d833352aSMel Gorman 3294502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 329504f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 3296502717f4SChen, Kenneth W { 329724669e58SAneesh Kumar K.V struct mm_struct *mm; 329824669e58SAneesh Kumar K.V struct mmu_gather tlb; 329924669e58SAneesh Kumar K.V 330024669e58SAneesh Kumar K.V mm = vma->vm_mm; 330124669e58SAneesh Kumar K.V 33022b047252SLinus Torvalds tlb_gather_mmu(&tlb, mm, start, end); 330324669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 330424669e58SAneesh Kumar K.V tlb_finish_mmu(&tlb, start, end); 3305502717f4SChen, Kenneth W } 3306502717f4SChen, Kenneth W 330704f2cbe3SMel Gorman /* 330804f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 330904f2cbe3SMel Gorman * mappping it owns the reserve page for. The intention is to unmap the page 331004f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 331104f2cbe3SMel Gorman * same region. 331204f2cbe3SMel Gorman */ 33132f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 33142a4b3dedSHarvey Harrison struct page *page, unsigned long address) 331504f2cbe3SMel Gorman { 33167526674dSAdam Litke struct hstate *h = hstate_vma(vma); 331704f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 331804f2cbe3SMel Gorman struct address_space *mapping; 331904f2cbe3SMel Gorman pgoff_t pgoff; 332004f2cbe3SMel Gorman 332104f2cbe3SMel Gorman /* 332204f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 332304f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 332404f2cbe3SMel Gorman */ 33257526674dSAdam Litke address = address & huge_page_mask(h); 332636e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 332736e4f20aSMichal Hocko vma->vm_pgoff; 3328496ad9aaSAl Viro mapping = file_inode(vma->vm_file)->i_mapping; 332904f2cbe3SMel Gorman 33304eb2b1dcSMel Gorman /* 33314eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 33324eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 33334eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 33344eb2b1dcSMel Gorman */ 333583cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 33366b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 333704f2cbe3SMel Gorman /* Do not unmap the current VMA */ 333804f2cbe3SMel Gorman if (iter_vma == vma) 333904f2cbe3SMel Gorman continue; 334004f2cbe3SMel Gorman 334104f2cbe3SMel Gorman /* 33422f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 33432f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 33442f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 33452f84a899SMel Gorman */ 33462f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 33472f84a899SMel Gorman continue; 33482f84a899SMel Gorman 33492f84a899SMel Gorman /* 335004f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 335104f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 335204f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 335304f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 335404f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 335504f2cbe3SMel Gorman */ 335604f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 335724669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 335824669e58SAneesh Kumar K.V address + huge_page_size(h), page); 335904f2cbe3SMel Gorman } 336083cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 336104f2cbe3SMel Gorman } 336204f2cbe3SMel Gorman 33630fe6e20bSNaoya Horiguchi /* 33640fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 3365ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 3366ef009b25SMichal Hocko * cannot race with other handlers or page migration. 3367ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 33680fe6e20bSNaoya Horiguchi */ 33691e8f889bSDavid Gibson static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 337004f2cbe3SMel Gorman unsigned long address, pte_t *ptep, pte_t pte, 3371cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 33721e8f889bSDavid Gibson { 3373a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 33741e8f889bSDavid Gibson struct page *old_page, *new_page; 3375ad4404a2SDavidlohr Bueso int ret = 0, outside_reserve = 0; 33762ec74c3eSSagi Grimberg unsigned long mmun_start; /* For mmu_notifiers */ 33772ec74c3eSSagi Grimberg unsigned long mmun_end; /* For mmu_notifiers */ 33781e8f889bSDavid Gibson 33791e8f889bSDavid Gibson old_page = pte_page(pte); 33801e8f889bSDavid Gibson 338104f2cbe3SMel Gorman retry_avoidcopy: 33821e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 33831e8f889bSDavid Gibson * and just make the page writable */ 338437a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 33850fe6e20bSNaoya Horiguchi page_move_anon_rmap(old_page, vma, address); 33861e8f889bSDavid Gibson set_huge_ptep_writable(vma, address, ptep); 338783c54070SNick Piggin return 0; 33881e8f889bSDavid Gibson } 33891e8f889bSDavid Gibson 339004f2cbe3SMel Gorman /* 339104f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 339204f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 339304f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 339404f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 339504f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 339604f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 339704f2cbe3SMel Gorman * of the full address range. 339804f2cbe3SMel Gorman */ 33995944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 340004f2cbe3SMel Gorman old_page != pagecache_page) 340104f2cbe3SMel Gorman outside_reserve = 1; 340204f2cbe3SMel Gorman 340309cbfeafSKirill A. Shutemov get_page(old_page); 3404b76c8cfbSLarry Woodman 3405ad4404a2SDavidlohr Bueso /* 3406ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 3407ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 3408ad4404a2SDavidlohr Bueso */ 3409cb900f41SKirill A. Shutemov spin_unlock(ptl); 341004f2cbe3SMel Gorman new_page = alloc_huge_page(vma, address, outside_reserve); 34111e8f889bSDavid Gibson 34122fc39cecSAdam Litke if (IS_ERR(new_page)) { 341304f2cbe3SMel Gorman /* 341404f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 341504f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 341604f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 341704f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 341804f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 341904f2cbe3SMel Gorman */ 342004f2cbe3SMel Gorman if (outside_reserve) { 342109cbfeafSKirill A. Shutemov put_page(old_page); 342204f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 34232f4612afSDavidlohr Bueso unmap_ref_private(mm, vma, old_page, address); 342404f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 3425cb900f41SKirill A. Shutemov spin_lock(ptl); 3426a734bcc8SHillf Danton ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 3427a9af0c5dSNaoya Horiguchi if (likely(ptep && 3428a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 342904f2cbe3SMel Gorman goto retry_avoidcopy; 3430a734bcc8SHillf Danton /* 3431cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 3432cb900f41SKirill A. Shutemov * lock, and our job is done. 3433a734bcc8SHillf Danton */ 3434a734bcc8SHillf Danton return 0; 343504f2cbe3SMel Gorman } 343604f2cbe3SMel Gorman 3437ad4404a2SDavidlohr Bueso ret = (PTR_ERR(new_page) == -ENOMEM) ? 3438ad4404a2SDavidlohr Bueso VM_FAULT_OOM : VM_FAULT_SIGBUS; 3439ad4404a2SDavidlohr Bueso goto out_release_old; 34401e8f889bSDavid Gibson } 34411e8f889bSDavid Gibson 34420fe6e20bSNaoya Horiguchi /* 34430fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 34440fe6e20bSNaoya Horiguchi * anon_vma prepared. 34450fe6e20bSNaoya Horiguchi */ 344644e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 3447ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 3448ad4404a2SDavidlohr Bueso goto out_release_all; 344944e2aa93SDean Nelson } 34500fe6e20bSNaoya Horiguchi 345147ad8475SAndrea Arcangeli copy_user_huge_page(new_page, old_page, address, vma, 345247ad8475SAndrea Arcangeli pages_per_huge_page(h)); 34530ed361deSNick Piggin __SetPageUptodate(new_page); 3454bcc54222SNaoya Horiguchi set_page_huge_active(new_page); 34551e8f889bSDavid Gibson 34562ec74c3eSSagi Grimberg mmun_start = address & huge_page_mask(h); 34572ec74c3eSSagi Grimberg mmun_end = mmun_start + huge_page_size(h); 34582ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 3459ad4404a2SDavidlohr Bueso 3460b76c8cfbSLarry Woodman /* 3461cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 3462b76c8cfbSLarry Woodman * before the page tables are altered 3463b76c8cfbSLarry Woodman */ 3464cb900f41SKirill A. Shutemov spin_lock(ptl); 3465a5516438SAndi Kleen ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 3466a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 346707443a85SJoonsoo Kim ClearPagePrivate(new_page); 346807443a85SJoonsoo Kim 34691e8f889bSDavid Gibson /* Break COW */ 34708fe627ecSGerald Schaefer huge_ptep_clear_flush(vma, address, ptep); 347134ee645eSJoerg Roedel mmu_notifier_invalidate_range(mm, mmun_start, mmun_end); 34721e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, 34731e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 3474d281ee61SKirill A. Shutemov page_remove_rmap(old_page, true); 3475cd67f0d2SNaoya Horiguchi hugepage_add_new_anon_rmap(new_page, vma, address); 34761e8f889bSDavid Gibson /* Make the old page be freed below */ 34771e8f889bSDavid Gibson new_page = old_page; 34781e8f889bSDavid Gibson } 3479cb900f41SKirill A. Shutemov spin_unlock(ptl); 34802ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 3481ad4404a2SDavidlohr Bueso out_release_all: 348209cbfeafSKirill A. Shutemov put_page(new_page); 3483ad4404a2SDavidlohr Bueso out_release_old: 348409cbfeafSKirill A. Shutemov put_page(old_page); 34858312034fSJoonsoo Kim 3486ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 3487ad4404a2SDavidlohr Bueso return ret; 34881e8f889bSDavid Gibson } 34891e8f889bSDavid Gibson 349004f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 3491a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 3492a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 349304f2cbe3SMel Gorman { 349404f2cbe3SMel Gorman struct address_space *mapping; 3495e7c4b0bfSAndy Whitcroft pgoff_t idx; 349604f2cbe3SMel Gorman 349704f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 3498a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 349904f2cbe3SMel Gorman 350004f2cbe3SMel Gorman return find_lock_page(mapping, idx); 350104f2cbe3SMel Gorman } 350204f2cbe3SMel Gorman 35033ae77f43SHugh Dickins /* 35043ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 35053ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 35063ae77f43SHugh Dickins */ 35073ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 35082a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 35092a15efc9SHugh Dickins { 35102a15efc9SHugh Dickins struct address_space *mapping; 35112a15efc9SHugh Dickins pgoff_t idx; 35122a15efc9SHugh Dickins struct page *page; 35132a15efc9SHugh Dickins 35142a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 35152a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 35162a15efc9SHugh Dickins 35172a15efc9SHugh Dickins page = find_get_page(mapping, idx); 35182a15efc9SHugh Dickins if (page) 35192a15efc9SHugh Dickins put_page(page); 35202a15efc9SHugh Dickins return page != NULL; 35212a15efc9SHugh Dickins } 35222a15efc9SHugh Dickins 3523ab76ad54SMike Kravetz int huge_add_to_page_cache(struct page *page, struct address_space *mapping, 3524ab76ad54SMike Kravetz pgoff_t idx) 3525ab76ad54SMike Kravetz { 3526ab76ad54SMike Kravetz struct inode *inode = mapping->host; 3527ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 3528ab76ad54SMike Kravetz int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 3529ab76ad54SMike Kravetz 3530ab76ad54SMike Kravetz if (err) 3531ab76ad54SMike Kravetz return err; 3532ab76ad54SMike Kravetz ClearPagePrivate(page); 3533ab76ad54SMike Kravetz 3534ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 3535ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 3536ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 3537ab76ad54SMike Kravetz return 0; 3538ab76ad54SMike Kravetz } 3539ab76ad54SMike Kravetz 3540a1ed3ddaSRobert P. J. Day static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, 35418382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 3542788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 3543ac9b9c66SHugh Dickins { 3544a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3545ac9b9c66SHugh Dickins int ret = VM_FAULT_SIGBUS; 3546409eb8c2SHillf Danton int anon_rmap = 0; 35474c887265SAdam Litke unsigned long size; 35484c887265SAdam Litke struct page *page; 35491e8f889bSDavid Gibson pte_t new_pte; 3550cb900f41SKirill A. Shutemov spinlock_t *ptl; 35514c887265SAdam Litke 355204f2cbe3SMel Gorman /* 355304f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 355404f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 355525985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 355604f2cbe3SMel Gorman */ 355704f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 3558910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 355904f2cbe3SMel Gorman current->pid); 356004f2cbe3SMel Gorman return ret; 356104f2cbe3SMel Gorman } 356204f2cbe3SMel Gorman 35634c887265SAdam Litke /* 35644c887265SAdam Litke * Use page lock to guard against racing truncation 35654c887265SAdam Litke * before we get page_table_lock. 35664c887265SAdam Litke */ 35676bda666aSChristoph Lameter retry: 35686bda666aSChristoph Lameter page = find_lock_page(mapping, idx); 35696bda666aSChristoph Lameter if (!page) { 3570a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 3571ebed4bfcSHugh Dickins if (idx >= size) 3572ebed4bfcSHugh Dickins goto out; 357304f2cbe3SMel Gorman page = alloc_huge_page(vma, address, 0); 35742fc39cecSAdam Litke if (IS_ERR(page)) { 357576dcee75SAneesh Kumar K.V ret = PTR_ERR(page); 357676dcee75SAneesh Kumar K.V if (ret == -ENOMEM) 357776dcee75SAneesh Kumar K.V ret = VM_FAULT_OOM; 357876dcee75SAneesh Kumar K.V else 357976dcee75SAneesh Kumar K.V ret = VM_FAULT_SIGBUS; 35806bda666aSChristoph Lameter goto out; 35816bda666aSChristoph Lameter } 358247ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 35830ed361deSNick Piggin __SetPageUptodate(page); 3584bcc54222SNaoya Horiguchi set_page_huge_active(page); 3585ac9b9c66SHugh Dickins 3586f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 3587ab76ad54SMike Kravetz int err = huge_add_to_page_cache(page, mapping, idx); 35886bda666aSChristoph Lameter if (err) { 35896bda666aSChristoph Lameter put_page(page); 35906bda666aSChristoph Lameter if (err == -EEXIST) 35916bda666aSChristoph Lameter goto retry; 35926bda666aSChristoph Lameter goto out; 35936bda666aSChristoph Lameter } 359423be7468SMel Gorman } else { 35956bda666aSChristoph Lameter lock_page(page); 35960fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 35970fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 35980fe6e20bSNaoya Horiguchi goto backout_unlocked; 359923be7468SMel Gorman } 3600409eb8c2SHillf Danton anon_rmap = 1; 36010fe6e20bSNaoya Horiguchi } 36020fe6e20bSNaoya Horiguchi } else { 360357303d80SAndy Whitcroft /* 3604998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 3605998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 3606998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 3607fd6a03edSNaoya Horiguchi */ 3608fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 3609aa50d3a7SAndi Kleen ret = VM_FAULT_HWPOISON | 3610972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 3611fd6a03edSNaoya Horiguchi goto backout_unlocked; 36126bda666aSChristoph Lameter } 3613998b4382SNaoya Horiguchi } 36141e8f889bSDavid Gibson 361557303d80SAndy Whitcroft /* 361657303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 361757303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 361857303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 361957303d80SAndy Whitcroft * the spinlock. 362057303d80SAndy Whitcroft */ 36215e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 36222b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 36232b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 36242b26736cSAndy Whitcroft goto backout_unlocked; 36252b26736cSAndy Whitcroft } 36265e911373SMike Kravetz /* Just decrements count, does not deallocate */ 3627feba16e2SMike Kravetz vma_end_reservation(h, vma, address); 36285e911373SMike Kravetz } 362957303d80SAndy Whitcroft 3630cb900f41SKirill A. Shutemov ptl = huge_pte_lockptr(h, mm, ptep); 3631cb900f41SKirill A. Shutemov spin_lock(ptl); 3632a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 36334c887265SAdam Litke if (idx >= size) 36344c887265SAdam Litke goto backout; 36354c887265SAdam Litke 363683c54070SNick Piggin ret = 0; 36377f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 36384c887265SAdam Litke goto backout; 36394c887265SAdam Litke 364007443a85SJoonsoo Kim if (anon_rmap) { 364107443a85SJoonsoo Kim ClearPagePrivate(page); 3642409eb8c2SHillf Danton hugepage_add_new_anon_rmap(page, vma, address); 3643ac714904SChoi Gi-yong } else 364453f9263bSKirill A. Shutemov page_dup_rmap(page, true); 36451e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 36461e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 36471e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, new_pte); 36481e8f889bSDavid Gibson 36495d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 3650788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 36511e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 3652cb900f41SKirill A. Shutemov ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl); 36531e8f889bSDavid Gibson } 36541e8f889bSDavid Gibson 3655cb900f41SKirill A. Shutemov spin_unlock(ptl); 36564c887265SAdam Litke unlock_page(page); 36574c887265SAdam Litke out: 3658ac9b9c66SHugh Dickins return ret; 36594c887265SAdam Litke 36604c887265SAdam Litke backout: 3661cb900f41SKirill A. Shutemov spin_unlock(ptl); 36622b26736cSAndy Whitcroft backout_unlocked: 36634c887265SAdam Litke unlock_page(page); 36644c887265SAdam Litke put_page(page); 36654c887265SAdam Litke goto out; 3666ac9b9c66SHugh Dickins } 3667ac9b9c66SHugh Dickins 36688382d914SDavidlohr Bueso #ifdef CONFIG_SMP 3669c672c7f2SMike Kravetz u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, 36708382d914SDavidlohr Bueso struct vm_area_struct *vma, 36718382d914SDavidlohr Bueso struct address_space *mapping, 36728382d914SDavidlohr Bueso pgoff_t idx, unsigned long address) 36738382d914SDavidlohr Bueso { 36748382d914SDavidlohr Bueso unsigned long key[2]; 36758382d914SDavidlohr Bueso u32 hash; 36768382d914SDavidlohr Bueso 36778382d914SDavidlohr Bueso if (vma->vm_flags & VM_SHARED) { 36788382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 36798382d914SDavidlohr Bueso key[1] = idx; 36808382d914SDavidlohr Bueso } else { 36818382d914SDavidlohr Bueso key[0] = (unsigned long) mm; 36828382d914SDavidlohr Bueso key[1] = address >> huge_page_shift(h); 36838382d914SDavidlohr Bueso } 36848382d914SDavidlohr Bueso 36858382d914SDavidlohr Bueso hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0); 36868382d914SDavidlohr Bueso 36878382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 36888382d914SDavidlohr Bueso } 36898382d914SDavidlohr Bueso #else 36908382d914SDavidlohr Bueso /* 36918382d914SDavidlohr Bueso * For uniprocesor systems we always use a single mutex, so just 36928382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 36938382d914SDavidlohr Bueso */ 3694c672c7f2SMike Kravetz u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, 36958382d914SDavidlohr Bueso struct vm_area_struct *vma, 36968382d914SDavidlohr Bueso struct address_space *mapping, 36978382d914SDavidlohr Bueso pgoff_t idx, unsigned long address) 36988382d914SDavidlohr Bueso { 36998382d914SDavidlohr Bueso return 0; 37008382d914SDavidlohr Bueso } 37018382d914SDavidlohr Bueso #endif 37028382d914SDavidlohr Bueso 370386e5216fSAdam Litke int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 3704788c7df4SHugh Dickins unsigned long address, unsigned int flags) 370586e5216fSAdam Litke { 37068382d914SDavidlohr Bueso pte_t *ptep, entry; 3707cb900f41SKirill A. Shutemov spinlock_t *ptl; 37081e8f889bSDavid Gibson int ret; 37098382d914SDavidlohr Bueso u32 hash; 37108382d914SDavidlohr Bueso pgoff_t idx; 37110fe6e20bSNaoya Horiguchi struct page *page = NULL; 371257303d80SAndy Whitcroft struct page *pagecache_page = NULL; 3713a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 37148382d914SDavidlohr Bueso struct address_space *mapping; 37150f792cf9SNaoya Horiguchi int need_wait_lock = 0; 371686e5216fSAdam Litke 37171e16a539SKAMEZAWA Hiroyuki address &= huge_page_mask(h); 37181e16a539SKAMEZAWA Hiroyuki 3719fd6a03edSNaoya Horiguchi ptep = huge_pte_offset(mm, address); 3720fd6a03edSNaoya Horiguchi if (ptep) { 3721fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 3722290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 3723cb900f41SKirill A. Shutemov migration_entry_wait_huge(vma, mm, ptep); 3724290408d4SNaoya Horiguchi return 0; 3725290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 3726aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 3727972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 37280d777df5SNaoya Horiguchi } else { 3729a5516438SAndi Kleen ptep = huge_pte_alloc(mm, address, huge_page_size(h)); 373086e5216fSAdam Litke if (!ptep) 373186e5216fSAdam Litke return VM_FAULT_OOM; 37320d777df5SNaoya Horiguchi } 373386e5216fSAdam Litke 37348382d914SDavidlohr Bueso mapping = vma->vm_file->f_mapping; 37358382d914SDavidlohr Bueso idx = vma_hugecache_offset(h, vma, address); 37368382d914SDavidlohr Bueso 37373935baa9SDavid Gibson /* 37383935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 37393935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 37403935baa9SDavid Gibson * the same page in the page cache. 37413935baa9SDavid Gibson */ 3742c672c7f2SMike Kravetz hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address); 3743c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 37448382d914SDavidlohr Bueso 37457f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 37467f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 37478382d914SDavidlohr Bueso ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 3748b4d1d99fSDavid Gibson goto out_mutex; 37493935baa9SDavid Gibson } 375086e5216fSAdam Litke 375183c54070SNick Piggin ret = 0; 37521e8f889bSDavid Gibson 375357303d80SAndy Whitcroft /* 37540f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 37550f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 37560f792cf9SNaoya Horiguchi * a active hugepage in pagecache. This goto expects the 2nd page fault, 37570f792cf9SNaoya Horiguchi * and is_hugetlb_entry_(migration|hwpoisoned) check will properly 37580f792cf9SNaoya Horiguchi * handle it. 37590f792cf9SNaoya Horiguchi */ 37600f792cf9SNaoya Horiguchi if (!pte_present(entry)) 37610f792cf9SNaoya Horiguchi goto out_mutex; 37620f792cf9SNaoya Horiguchi 37630f792cf9SNaoya Horiguchi /* 376457303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 376557303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 376657303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 376757303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 376857303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 376957303d80SAndy Whitcroft * consumed. 377057303d80SAndy Whitcroft */ 3771106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 37722b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 37732b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 3774b4d1d99fSDavid Gibson goto out_mutex; 37752b26736cSAndy Whitcroft } 37765e911373SMike Kravetz /* Just decrements count, does not deallocate */ 3777feba16e2SMike Kravetz vma_end_reservation(h, vma, address); 377857303d80SAndy Whitcroft 3779f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 378057303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 378157303d80SAndy Whitcroft vma, address); 378257303d80SAndy Whitcroft } 378357303d80SAndy Whitcroft 37840f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 37850fe6e20bSNaoya Horiguchi 37861e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 3787b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 3788cb900f41SKirill A. Shutemov goto out_ptl; 3789b4d1d99fSDavid Gibson 37900f792cf9SNaoya Horiguchi /* 37910f792cf9SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 37920f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 37930f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 37940f792cf9SNaoya Horiguchi */ 37950f792cf9SNaoya Horiguchi page = pte_page(entry); 37960f792cf9SNaoya Horiguchi if (page != pagecache_page) 37970f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 37980f792cf9SNaoya Horiguchi need_wait_lock = 1; 37990f792cf9SNaoya Horiguchi goto out_ptl; 38000f792cf9SNaoya Horiguchi } 38010f792cf9SNaoya Horiguchi 38020f792cf9SNaoya Horiguchi get_page(page); 3803b4d1d99fSDavid Gibson 3804788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 3805106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 380657303d80SAndy Whitcroft ret = hugetlb_cow(mm, vma, address, ptep, entry, 3807cb900f41SKirill A. Shutemov pagecache_page, ptl); 38080f792cf9SNaoya Horiguchi goto out_put_page; 3809b4d1d99fSDavid Gibson } 3810106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 3811b4d1d99fSDavid Gibson } 3812b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 3813788c7df4SHugh Dickins if (huge_ptep_set_access_flags(vma, address, ptep, entry, 3814788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 38154b3073e1SRussell King update_mmu_cache(vma, address, ptep); 38160f792cf9SNaoya Horiguchi out_put_page: 38170f792cf9SNaoya Horiguchi if (page != pagecache_page) 38180f792cf9SNaoya Horiguchi unlock_page(page); 38190f792cf9SNaoya Horiguchi put_page(page); 3820cb900f41SKirill A. Shutemov out_ptl: 3821cb900f41SKirill A. Shutemov spin_unlock(ptl); 382257303d80SAndy Whitcroft 382357303d80SAndy Whitcroft if (pagecache_page) { 382457303d80SAndy Whitcroft unlock_page(pagecache_page); 382557303d80SAndy Whitcroft put_page(pagecache_page); 382657303d80SAndy Whitcroft } 3827b4d1d99fSDavid Gibson out_mutex: 3828c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 38290f792cf9SNaoya Horiguchi /* 38300f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 38310f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 38320f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 38330f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 38340f792cf9SNaoya Horiguchi * here without taking refcount. 38350f792cf9SNaoya Horiguchi */ 38360f792cf9SNaoya Horiguchi if (need_wait_lock) 38370f792cf9SNaoya Horiguchi wait_on_page_locked(page); 38381e8f889bSDavid Gibson return ret; 383986e5216fSAdam Litke } 384086e5216fSAdam Litke 384128a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 384263551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 384328a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 384428a35716SMichel Lespinasse long i, unsigned int flags) 384563551ae0SDavid Gibson { 3846d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 3847d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 384828a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 3849a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 385063551ae0SDavid Gibson 385163551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 385263551ae0SDavid Gibson pte_t *pte; 3853cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 38542a15efc9SHugh Dickins int absent; 385563551ae0SDavid Gibson struct page *page; 385663551ae0SDavid Gibson 38574c887265SAdam Litke /* 385802057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 385902057967SDavid Rientjes * potentially allocating memory. 386002057967SDavid Rientjes */ 386102057967SDavid Rientjes if (unlikely(fatal_signal_pending(current))) { 386202057967SDavid Rientjes remainder = 0; 386302057967SDavid Rientjes break; 386402057967SDavid Rientjes } 386502057967SDavid Rientjes 386602057967SDavid Rientjes /* 38674c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 38682a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 38694c887265SAdam Litke * first, for the page indexing below to work. 3870cb900f41SKirill A. Shutemov * 3871cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 38724c887265SAdam Litke */ 3873a5516438SAndi Kleen pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); 3874cb900f41SKirill A. Shutemov if (pte) 3875cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 38762a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 387763551ae0SDavid Gibson 38782a15efc9SHugh Dickins /* 38792a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 38803ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 38813ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 38823ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 38833ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 38842a15efc9SHugh Dickins */ 38853ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 38863ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 3887cb900f41SKirill A. Shutemov if (pte) 3888cb900f41SKirill A. Shutemov spin_unlock(ptl); 38892a15efc9SHugh Dickins remainder = 0; 38902a15efc9SHugh Dickins break; 38912a15efc9SHugh Dickins } 38922a15efc9SHugh Dickins 38939cc3a5bdSNaoya Horiguchi /* 38949cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 38959cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 38969cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 38979cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 38989cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 38999cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 39009cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 39019cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 39029cc3a5bdSNaoya Horiguchi */ 39039cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 3904106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 3905106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 39064c887265SAdam Litke int ret; 39074c887265SAdam Litke 3908cb900f41SKirill A. Shutemov if (pte) 3909cb900f41SKirill A. Shutemov spin_unlock(ptl); 39102a15efc9SHugh Dickins ret = hugetlb_fault(mm, vma, vaddr, 39112a15efc9SHugh Dickins (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); 3912a89182c7SAdam Litke if (!(ret & VM_FAULT_ERROR)) 39134c887265SAdam Litke continue; 39144c887265SAdam Litke 39151c59827dSHugh Dickins remainder = 0; 39161c59827dSHugh Dickins break; 39171c59827dSHugh Dickins } 391863551ae0SDavid Gibson 3919a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 39207f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 3921d5d4b0aaSChen, Kenneth W same_page: 3922d6692183SChen, Kenneth W if (pages) { 392369d177c2SAndy Whitcroft pages[i] = mem_map_offset(page, pfn_offset); 3924ddc58f27SKirill A. Shutemov get_page(pages[i]); 3925d6692183SChen, Kenneth W } 392663551ae0SDavid Gibson 392763551ae0SDavid Gibson if (vmas) 392863551ae0SDavid Gibson vmas[i] = vma; 392963551ae0SDavid Gibson 393063551ae0SDavid Gibson vaddr += PAGE_SIZE; 3931d5d4b0aaSChen, Kenneth W ++pfn_offset; 393263551ae0SDavid Gibson --remainder; 393363551ae0SDavid Gibson ++i; 3934d5d4b0aaSChen, Kenneth W if (vaddr < vma->vm_end && remainder && 3935a5516438SAndi Kleen pfn_offset < pages_per_huge_page(h)) { 3936d5d4b0aaSChen, Kenneth W /* 3937d5d4b0aaSChen, Kenneth W * We use pfn_offset to avoid touching the pageframes 3938d5d4b0aaSChen, Kenneth W * of this compound page. 3939d5d4b0aaSChen, Kenneth W */ 3940d5d4b0aaSChen, Kenneth W goto same_page; 3941d5d4b0aaSChen, Kenneth W } 3942cb900f41SKirill A. Shutemov spin_unlock(ptl); 394363551ae0SDavid Gibson } 394428a35716SMichel Lespinasse *nr_pages = remainder; 394563551ae0SDavid Gibson *position = vaddr; 394663551ae0SDavid Gibson 39472a15efc9SHugh Dickins return i ? i : -EFAULT; 394863551ae0SDavid Gibson } 39498f860591SZhang, Yanmin 39507da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 39518f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 39528f860591SZhang, Yanmin { 39538f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 39548f860591SZhang, Yanmin unsigned long start = address; 39558f860591SZhang, Yanmin pte_t *ptep; 39568f860591SZhang, Yanmin pte_t pte; 3957a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 39587da4d641SPeter Zijlstra unsigned long pages = 0; 39598f860591SZhang, Yanmin 39608f860591SZhang, Yanmin BUG_ON(address >= end); 39618f860591SZhang, Yanmin flush_cache_range(vma, address, end); 39628f860591SZhang, Yanmin 3963a5338093SRik van Riel mmu_notifier_invalidate_range_start(mm, start, end); 396483cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 3965a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 3966cb900f41SKirill A. Shutemov spinlock_t *ptl; 39678f860591SZhang, Yanmin ptep = huge_pte_offset(mm, address); 39688f860591SZhang, Yanmin if (!ptep) 39698f860591SZhang, Yanmin continue; 3970cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 39717da4d641SPeter Zijlstra if (huge_pmd_unshare(mm, &address, ptep)) { 39727da4d641SPeter Zijlstra pages++; 3973cb900f41SKirill A. Shutemov spin_unlock(ptl); 397439dde65cSChen, Kenneth W continue; 39757da4d641SPeter Zijlstra } 3976a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 3977a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 3978a8bda28dSNaoya Horiguchi spin_unlock(ptl); 3979a8bda28dSNaoya Horiguchi continue; 3980a8bda28dSNaoya Horiguchi } 3981a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 3982a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 3983a8bda28dSNaoya Horiguchi 3984a8bda28dSNaoya Horiguchi if (is_write_migration_entry(entry)) { 3985a8bda28dSNaoya Horiguchi pte_t newpte; 3986a8bda28dSNaoya Horiguchi 3987a8bda28dSNaoya Horiguchi make_migration_entry_read(&entry); 3988a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 3989a8bda28dSNaoya Horiguchi set_huge_pte_at(mm, address, ptep, newpte); 3990a8bda28dSNaoya Horiguchi pages++; 3991a8bda28dSNaoya Horiguchi } 3992a8bda28dSNaoya Horiguchi spin_unlock(ptl); 3993a8bda28dSNaoya Horiguchi continue; 3994a8bda28dSNaoya Horiguchi } 3995a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 39968f860591SZhang, Yanmin pte = huge_ptep_get_and_clear(mm, address, ptep); 3997106c992aSGerald Schaefer pte = pte_mkhuge(huge_pte_modify(pte, newprot)); 3998be7517d6STony Lu pte = arch_make_huge_pte(pte, vma, NULL, 0); 39998f860591SZhang, Yanmin set_huge_pte_at(mm, address, ptep, pte); 40007da4d641SPeter Zijlstra pages++; 40018f860591SZhang, Yanmin } 4002cb900f41SKirill A. Shutemov spin_unlock(ptl); 40038f860591SZhang, Yanmin } 4004d833352aSMel Gorman /* 4005c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 4006d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 4007c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 4008d833352aSMel Gorman * and that page table be reused and filled with junk. 4009d833352aSMel Gorman */ 40108f860591SZhang, Yanmin flush_tlb_range(vma, start, end); 401134ee645eSJoerg Roedel mmu_notifier_invalidate_range(mm, start, end); 401283cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 4013a5338093SRik van Riel mmu_notifier_invalidate_range_end(mm, start, end); 40147da4d641SPeter Zijlstra 40157da4d641SPeter Zijlstra return pages << h->order; 40168f860591SZhang, Yanmin } 40178f860591SZhang, Yanmin 4018a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 4019a1e78772SMel Gorman long from, long to, 40205a6fe125SMel Gorman struct vm_area_struct *vma, 4021ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 4022e4e574b7SAdam Litke { 402317c9d12eSMel Gorman long ret, chg; 4024a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 402590481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 40269119a41eSJoonsoo Kim struct resv_map *resv_map; 40271c5ecae3SMike Kravetz long gbl_reserve; 4028e4e574b7SAdam Litke 4029a1e78772SMel Gorman /* 403017c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 403117c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 403290481622SDavid Gibson * without using reserves 403317c9d12eSMel Gorman */ 4034ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 403517c9d12eSMel Gorman return 0; 403617c9d12eSMel Gorman 403717c9d12eSMel Gorman /* 4038a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 4039a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 4040a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 4041a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 4042a1e78772SMel Gorman */ 40439119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 40444e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 40459119a41eSJoonsoo Kim 40461406ec9bSJoonsoo Kim chg = region_chg(resv_map, from, to); 40479119a41eSJoonsoo Kim 40489119a41eSJoonsoo Kim } else { 40499119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 40505a6fe125SMel Gorman if (!resv_map) 40515a6fe125SMel Gorman return -ENOMEM; 40525a6fe125SMel Gorman 405317c9d12eSMel Gorman chg = to - from; 405417c9d12eSMel Gorman 40555a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 40565a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 40575a6fe125SMel Gorman } 40585a6fe125SMel Gorman 4059c50ac050SDave Hansen if (chg < 0) { 4060c50ac050SDave Hansen ret = chg; 4061c50ac050SDave Hansen goto out_err; 4062c50ac050SDave Hansen } 406317c9d12eSMel Gorman 40641c5ecae3SMike Kravetz /* 40651c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 40661c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 40671c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 40681c5ecae3SMike Kravetz */ 40691c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 40701c5ecae3SMike Kravetz if (gbl_reserve < 0) { 4071c50ac050SDave Hansen ret = -ENOSPC; 4072c50ac050SDave Hansen goto out_err; 4073c50ac050SDave Hansen } 407417c9d12eSMel Gorman 407517c9d12eSMel Gorman /* 407617c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 407790481622SDavid Gibson * Hand the pages back to the subpool if there are not 407817c9d12eSMel Gorman */ 40791c5ecae3SMike Kravetz ret = hugetlb_acct_memory(h, gbl_reserve); 408017c9d12eSMel Gorman if (ret < 0) { 40811c5ecae3SMike Kravetz /* put back original number of pages, chg */ 40821c5ecae3SMike Kravetz (void)hugepage_subpool_put_pages(spool, chg); 4083c50ac050SDave Hansen goto out_err; 408417c9d12eSMel Gorman } 408517c9d12eSMel Gorman 408617c9d12eSMel Gorman /* 408717c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 408817c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 408917c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 409017c9d12eSMel Gorman * the reservation was and the page cache tells how much of 409117c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 409217c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 409317c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 409417c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 409517c9d12eSMel Gorman * else has to be done for private mappings here 409617c9d12eSMel Gorman */ 409733039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 409833039678SMike Kravetz long add = region_add(resv_map, from, to); 409933039678SMike Kravetz 410033039678SMike Kravetz if (unlikely(chg > add)) { 410133039678SMike Kravetz /* 410233039678SMike Kravetz * pages in this range were added to the reserve 410333039678SMike Kravetz * map between region_chg and region_add. This 410433039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 410533039678SMike Kravetz * the subpool and reserve counts modified above 410633039678SMike Kravetz * based on the difference. 410733039678SMike Kravetz */ 410833039678SMike Kravetz long rsv_adjust; 410933039678SMike Kravetz 411033039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 411133039678SMike Kravetz chg - add); 411233039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 411333039678SMike Kravetz } 411433039678SMike Kravetz } 4115a43a8c39SChen, Kenneth W return 0; 4116c50ac050SDave Hansen out_err: 41175e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 41185e911373SMike Kravetz region_abort(resv_map, from, to); 4119f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 4120f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 4121c50ac050SDave Hansen return ret; 4122a43a8c39SChen, Kenneth W } 4123a43a8c39SChen, Kenneth W 4124b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 4125b5cec28dSMike Kravetz long freed) 4126a43a8c39SChen, Kenneth W { 4127a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 41284e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 41299119a41eSJoonsoo Kim long chg = 0; 413090481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 41311c5ecae3SMike Kravetz long gbl_reserve; 413245c682a6SKen Chen 4133b5cec28dSMike Kravetz if (resv_map) { 4134b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 4135b5cec28dSMike Kravetz /* 4136b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 4137b5cec28dSMike Kravetz * must be split and another region descriptor can not be 4138b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 4139b5cec28dSMike Kravetz */ 4140b5cec28dSMike Kravetz if (chg < 0) 4141b5cec28dSMike Kravetz return chg; 4142b5cec28dSMike Kravetz } 4143b5cec28dSMike Kravetz 414445c682a6SKen Chen spin_lock(&inode->i_lock); 4145e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 414645c682a6SKen Chen spin_unlock(&inode->i_lock); 414745c682a6SKen Chen 41481c5ecae3SMike Kravetz /* 41491c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 41501c5ecae3SMike Kravetz * reservations to be released may be adjusted. 41511c5ecae3SMike Kravetz */ 41521c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 41531c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 4154b5cec28dSMike Kravetz 4155b5cec28dSMike Kravetz return 0; 4156a43a8c39SChen, Kenneth W } 415793f70f90SNaoya Horiguchi 41583212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 41593212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 41603212b535SSteve Capper struct vm_area_struct *vma, 41613212b535SSteve Capper unsigned long addr, pgoff_t idx) 41623212b535SSteve Capper { 41633212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 41643212b535SSteve Capper svma->vm_start; 41653212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 41663212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 41673212b535SSteve Capper 41683212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 4169de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 4170de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 41713212b535SSteve Capper 41723212b535SSteve Capper /* 41733212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 41743212b535SSteve Capper * page table page. 41753212b535SSteve Capper */ 41763212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 41773212b535SSteve Capper vm_flags != svm_flags || 41783212b535SSteve Capper sbase < svma->vm_start || svma->vm_end < s_end) 41793212b535SSteve Capper return 0; 41803212b535SSteve Capper 41813212b535SSteve Capper return saddr; 41823212b535SSteve Capper } 41833212b535SSteve Capper 418431aafb45SNicholas Krause static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) 41853212b535SSteve Capper { 41863212b535SSteve Capper unsigned long base = addr & PUD_MASK; 41873212b535SSteve Capper unsigned long end = base + PUD_SIZE; 41883212b535SSteve Capper 41893212b535SSteve Capper /* 41903212b535SSteve Capper * check on proper vm_flags and page table alignment 41913212b535SSteve Capper */ 41923212b535SSteve Capper if (vma->vm_flags & VM_MAYSHARE && 41933212b535SSteve Capper vma->vm_start <= base && end <= vma->vm_end) 419431aafb45SNicholas Krause return true; 419531aafb45SNicholas Krause return false; 41963212b535SSteve Capper } 41973212b535SSteve Capper 41983212b535SSteve Capper /* 41993212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 42003212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 42013212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 42023212b535SSteve Capper * code much cleaner. pmd allocation is essential for the shared case because 4203c8c06efaSDavidlohr Bueso * pud has to be populated inside the same i_mmap_rwsem section - otherwise 42043212b535SSteve Capper * racing tasks could either miss the sharing (see huge_pte_offset) or select a 42053212b535SSteve Capper * bad pmd for sharing. 42063212b535SSteve Capper */ 42073212b535SSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 42083212b535SSteve Capper { 42093212b535SSteve Capper struct vm_area_struct *vma = find_vma(mm, addr); 42103212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 42113212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 42123212b535SSteve Capper vma->vm_pgoff; 42133212b535SSteve Capper struct vm_area_struct *svma; 42143212b535SSteve Capper unsigned long saddr; 42153212b535SSteve Capper pte_t *spte = NULL; 42163212b535SSteve Capper pte_t *pte; 4217cb900f41SKirill A. Shutemov spinlock_t *ptl; 42183212b535SSteve Capper 42193212b535SSteve Capper if (!vma_shareable(vma, addr)) 42203212b535SSteve Capper return (pte_t *)pmd_alloc(mm, pud, addr); 42213212b535SSteve Capper 422283cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 42233212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 42243212b535SSteve Capper if (svma == vma) 42253212b535SSteve Capper continue; 42263212b535SSteve Capper 42273212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 42283212b535SSteve Capper if (saddr) { 42293212b535SSteve Capper spte = huge_pte_offset(svma->vm_mm, saddr); 42303212b535SSteve Capper if (spte) { 4231dc6c9a35SKirill A. Shutemov mm_inc_nr_pmds(mm); 42323212b535SSteve Capper get_page(virt_to_page(spte)); 42333212b535SSteve Capper break; 42343212b535SSteve Capper } 42353212b535SSteve Capper } 42363212b535SSteve Capper } 42373212b535SSteve Capper 42383212b535SSteve Capper if (!spte) 42393212b535SSteve Capper goto out; 42403212b535SSteve Capper 4241cb900f41SKirill A. Shutemov ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte); 4242cb900f41SKirill A. Shutemov spin_lock(ptl); 4243dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 42443212b535SSteve Capper pud_populate(mm, pud, 42453212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 4246dc6c9a35SKirill A. Shutemov } else { 42473212b535SSteve Capper put_page(virt_to_page(spte)); 4248dc6c9a35SKirill A. Shutemov mm_inc_nr_pmds(mm); 4249dc6c9a35SKirill A. Shutemov } 4250cb900f41SKirill A. Shutemov spin_unlock(ptl); 42513212b535SSteve Capper out: 42523212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 425383cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 42543212b535SSteve Capper return pte; 42553212b535SSteve Capper } 42563212b535SSteve Capper 42573212b535SSteve Capper /* 42583212b535SSteve Capper * unmap huge page backed by shared pte. 42593212b535SSteve Capper * 42603212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 42613212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 42623212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 42633212b535SSteve Capper * 4264cb900f41SKirill A. Shutemov * called with page table lock held. 42653212b535SSteve Capper * 42663212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 42673212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 42683212b535SSteve Capper */ 42693212b535SSteve Capper int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 42703212b535SSteve Capper { 42713212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 42723212b535SSteve Capper pud_t *pud = pud_offset(pgd, *addr); 42733212b535SSteve Capper 42743212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 42753212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 42763212b535SSteve Capper return 0; 42773212b535SSteve Capper 42783212b535SSteve Capper pud_clear(pud); 42793212b535SSteve Capper put_page(virt_to_page(ptep)); 4280dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 42813212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 42823212b535SSteve Capper return 1; 42833212b535SSteve Capper } 42849e5fc74cSSteve Capper #define want_pmd_share() (1) 42859e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 42869e5fc74cSSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 42879e5fc74cSSteve Capper { 42889e5fc74cSSteve Capper return NULL; 42899e5fc74cSSteve Capper } 4290e81f2d22SZhang Zhen 4291e81f2d22SZhang Zhen int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 4292e81f2d22SZhang Zhen { 4293e81f2d22SZhang Zhen return 0; 4294e81f2d22SZhang Zhen } 42959e5fc74cSSteve Capper #define want_pmd_share() (0) 42963212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 42973212b535SSteve Capper 42989e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 42999e5fc74cSSteve Capper pte_t *huge_pte_alloc(struct mm_struct *mm, 43009e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 43019e5fc74cSSteve Capper { 43029e5fc74cSSteve Capper pgd_t *pgd; 43039e5fc74cSSteve Capper pud_t *pud; 43049e5fc74cSSteve Capper pte_t *pte = NULL; 43059e5fc74cSSteve Capper 43069e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 43079e5fc74cSSteve Capper pud = pud_alloc(mm, pgd, addr); 43089e5fc74cSSteve Capper if (pud) { 43099e5fc74cSSteve Capper if (sz == PUD_SIZE) { 43109e5fc74cSSteve Capper pte = (pte_t *)pud; 43119e5fc74cSSteve Capper } else { 43129e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 43139e5fc74cSSteve Capper if (want_pmd_share() && pud_none(*pud)) 43149e5fc74cSSteve Capper pte = huge_pmd_share(mm, addr, pud); 43159e5fc74cSSteve Capper else 43169e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 43179e5fc74cSSteve Capper } 43189e5fc74cSSteve Capper } 43199e5fc74cSSteve Capper BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte)); 43209e5fc74cSSteve Capper 43219e5fc74cSSteve Capper return pte; 43229e5fc74cSSteve Capper } 43239e5fc74cSSteve Capper 43249e5fc74cSSteve Capper pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) 43259e5fc74cSSteve Capper { 43269e5fc74cSSteve Capper pgd_t *pgd; 43279e5fc74cSSteve Capper pud_t *pud; 43289e5fc74cSSteve Capper pmd_t *pmd = NULL; 43299e5fc74cSSteve Capper 43309e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 43319e5fc74cSSteve Capper if (pgd_present(*pgd)) { 43329e5fc74cSSteve Capper pud = pud_offset(pgd, addr); 43339e5fc74cSSteve Capper if (pud_present(*pud)) { 43349e5fc74cSSteve Capper if (pud_huge(*pud)) 43359e5fc74cSSteve Capper return (pte_t *)pud; 43369e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 43379e5fc74cSSteve Capper } 43389e5fc74cSSteve Capper } 43399e5fc74cSSteve Capper return (pte_t *) pmd; 43409e5fc74cSSteve Capper } 43419e5fc74cSSteve Capper 434261f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 434361f77edaSNaoya Horiguchi 434461f77edaSNaoya Horiguchi /* 434561f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 434661f77edaSNaoya Horiguchi * behavior. 434761f77edaSNaoya Horiguchi */ 434861f77edaSNaoya Horiguchi struct page * __weak 434961f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 435061f77edaSNaoya Horiguchi int write) 435161f77edaSNaoya Horiguchi { 435261f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 435361f77edaSNaoya Horiguchi } 435461f77edaSNaoya Horiguchi 435561f77edaSNaoya Horiguchi struct page * __weak 43569e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 4357e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 43589e5fc74cSSteve Capper { 4359e66f17ffSNaoya Horiguchi struct page *page = NULL; 4360e66f17ffSNaoya Horiguchi spinlock_t *ptl; 4361e66f17ffSNaoya Horiguchi retry: 4362e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 4363e66f17ffSNaoya Horiguchi spin_lock(ptl); 4364e66f17ffSNaoya Horiguchi /* 4365e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 4366e66f17ffSNaoya Horiguchi * unmapped from other threads. 4367e66f17ffSNaoya Horiguchi */ 4368e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 4369e66f17ffSNaoya Horiguchi goto out; 4370e66f17ffSNaoya Horiguchi if (pmd_present(*pmd)) { 437197534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 4372e66f17ffSNaoya Horiguchi if (flags & FOLL_GET) 4373e66f17ffSNaoya Horiguchi get_page(page); 4374e66f17ffSNaoya Horiguchi } else { 4375e66f17ffSNaoya Horiguchi if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) { 4376e66f17ffSNaoya Horiguchi spin_unlock(ptl); 4377e66f17ffSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pmd, ptl); 4378e66f17ffSNaoya Horiguchi goto retry; 4379e66f17ffSNaoya Horiguchi } 4380e66f17ffSNaoya Horiguchi /* 4381e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 4382e66f17ffSNaoya Horiguchi * follow_page_mask(). 4383e66f17ffSNaoya Horiguchi */ 4384e66f17ffSNaoya Horiguchi } 4385e66f17ffSNaoya Horiguchi out: 4386e66f17ffSNaoya Horiguchi spin_unlock(ptl); 43879e5fc74cSSteve Capper return page; 43889e5fc74cSSteve Capper } 43899e5fc74cSSteve Capper 439061f77edaSNaoya Horiguchi struct page * __weak 43919e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 4392e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 43939e5fc74cSSteve Capper { 4394e66f17ffSNaoya Horiguchi if (flags & FOLL_GET) 4395e66f17ffSNaoya Horiguchi return NULL; 43969e5fc74cSSteve Capper 4397e66f17ffSNaoya Horiguchi return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 43989e5fc74cSSteve Capper } 43999e5fc74cSSteve Capper 4400d5bd9106SAndi Kleen #ifdef CONFIG_MEMORY_FAILURE 4401d5bd9106SAndi Kleen 440293f70f90SNaoya Horiguchi /* 440393f70f90SNaoya Horiguchi * This function is called from memory failure code. 440493f70f90SNaoya Horiguchi * Assume the caller holds page lock of the head page. 440593f70f90SNaoya Horiguchi */ 44066de2b1aaSNaoya Horiguchi int dequeue_hwpoisoned_huge_page(struct page *hpage) 440793f70f90SNaoya Horiguchi { 440893f70f90SNaoya Horiguchi struct hstate *h = page_hstate(hpage); 440993f70f90SNaoya Horiguchi int nid = page_to_nid(hpage); 44106de2b1aaSNaoya Horiguchi int ret = -EBUSY; 441193f70f90SNaoya Horiguchi 441293f70f90SNaoya Horiguchi spin_lock(&hugetlb_lock); 44137e1f049eSNaoya Horiguchi /* 44147e1f049eSNaoya Horiguchi * Just checking !page_huge_active is not enough, because that could be 44157e1f049eSNaoya Horiguchi * an isolated/hwpoisoned hugepage (which have >0 refcount). 44167e1f049eSNaoya Horiguchi */ 44177e1f049eSNaoya Horiguchi if (!page_huge_active(hpage) && !page_count(hpage)) { 441856f2fb14SNaoya Horiguchi /* 441956f2fb14SNaoya Horiguchi * Hwpoisoned hugepage isn't linked to activelist or freelist, 442056f2fb14SNaoya Horiguchi * but dangling hpage->lru can trigger list-debug warnings 442156f2fb14SNaoya Horiguchi * (this happens when we call unpoison_memory() on it), 442256f2fb14SNaoya Horiguchi * so let it point to itself with list_del_init(). 442356f2fb14SNaoya Horiguchi */ 442456f2fb14SNaoya Horiguchi list_del_init(&hpage->lru); 44258c6c2ecbSNaoya Horiguchi set_page_refcounted(hpage); 442693f70f90SNaoya Horiguchi h->free_huge_pages--; 442793f70f90SNaoya Horiguchi h->free_huge_pages_node[nid]--; 44286de2b1aaSNaoya Horiguchi ret = 0; 442993f70f90SNaoya Horiguchi } 44306de2b1aaSNaoya Horiguchi spin_unlock(&hugetlb_lock); 44316de2b1aaSNaoya Horiguchi return ret; 44326de2b1aaSNaoya Horiguchi } 44336de2b1aaSNaoya Horiguchi #endif 443431caf665SNaoya Horiguchi 443531caf665SNaoya Horiguchi bool isolate_huge_page(struct page *page, struct list_head *list) 443631caf665SNaoya Horiguchi { 4437bcc54222SNaoya Horiguchi bool ret = true; 4438bcc54222SNaoya Horiguchi 4439309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 444031caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 4441bcc54222SNaoya Horiguchi if (!page_huge_active(page) || !get_page_unless_zero(page)) { 4442bcc54222SNaoya Horiguchi ret = false; 4443bcc54222SNaoya Horiguchi goto unlock; 4444bcc54222SNaoya Horiguchi } 4445bcc54222SNaoya Horiguchi clear_page_huge_active(page); 444631caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 4447bcc54222SNaoya Horiguchi unlock: 444831caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 4449bcc54222SNaoya Horiguchi return ret; 445031caf665SNaoya Horiguchi } 445131caf665SNaoya Horiguchi 445231caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 445331caf665SNaoya Horiguchi { 4454309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 445531caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 4456bcc54222SNaoya Horiguchi set_page_huge_active(page); 445731caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 445831caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 445931caf665SNaoya Horiguchi put_page(page); 446031caf665SNaoya Horiguchi } 4461