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> 21174cd4b1SIngo Molnar #include <linux/sched/signal.h> 220fe6e20bSNaoya Horiguchi #include <linux/rmap.h> 23c6247f72SMatthew Wilcox #include <linux/string_helpers.h> 24fd6a03edSNaoya Horiguchi #include <linux/swap.h> 25fd6a03edSNaoya Horiguchi #include <linux/swapops.h> 268382d914SDavidlohr Bueso #include <linux/jhash.h> 27d6606683SLinus Torvalds 2863551ae0SDavid Gibson #include <asm/page.h> 2963551ae0SDavid Gibson #include <asm/pgtable.h> 3024669e58SAneesh Kumar K.V #include <asm/tlb.h> 3163551ae0SDavid Gibson 3224669e58SAneesh Kumar K.V #include <linux/io.h> 3363551ae0SDavid Gibson #include <linux/hugetlb.h> 349dd540e2SAneesh Kumar K.V #include <linux/hugetlb_cgroup.h> 359a305230SLee Schermerhorn #include <linux/node.h> 361a1aad8aSMike Kravetz #include <linux/userfaultfd_k.h> 37ab5ac90aSMichal Hocko #include <linux/page_owner.h> 387835e98bSNick Piggin #include "internal.h" 391da177e4SLinus Torvalds 40c3f38a38SAneesh Kumar K.V int hugetlb_max_hstate __read_mostly; 41e5ff2159SAndi Kleen unsigned int default_hstate_idx; 42e5ff2159SAndi Kleen struct hstate hstates[HUGE_MAX_HSTATE]; 43641844f5SNaoya Horiguchi /* 44641844f5SNaoya Horiguchi * Minimum page order among possible hugepage sizes, set to a proper value 45641844f5SNaoya Horiguchi * at boot time. 46641844f5SNaoya Horiguchi */ 47641844f5SNaoya Horiguchi static unsigned int minimum_order __read_mostly = UINT_MAX; 48e5ff2159SAndi Kleen 4953ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 5053ba51d2SJon Tollefson 51e5ff2159SAndi Kleen /* for command line parsing */ 52e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 53e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 54e11bfbfcSNick Piggin static unsigned long __initdata default_hstate_size; 559fee021dSVaishali Thakkar static bool __initdata parsed_valid_hugepagesz = true; 56e5ff2159SAndi Kleen 573935baa9SDavid Gibson /* 5831caf665SNaoya Horiguchi * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, 5931caf665SNaoya Horiguchi * free_huge_pages, and surplus_huge_pages. 603935baa9SDavid Gibson */ 61c3f38a38SAneesh Kumar K.V DEFINE_SPINLOCK(hugetlb_lock); 620bd0f9fbSEric Paris 638382d914SDavidlohr Bueso /* 648382d914SDavidlohr Bueso * Serializes faults on the same logical page. This is used to 658382d914SDavidlohr Bueso * prevent spurious OOMs when the hugepage pool is fully utilized. 668382d914SDavidlohr Bueso */ 678382d914SDavidlohr Bueso static int num_fault_mutexes; 68c672c7f2SMike Kravetz struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; 698382d914SDavidlohr Bueso 707ca02d0aSMike Kravetz /* Forward declaration */ 717ca02d0aSMike Kravetz static int hugetlb_acct_memory(struct hstate *h, long delta); 727ca02d0aSMike Kravetz 7390481622SDavid Gibson static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) 7490481622SDavid Gibson { 7590481622SDavid Gibson bool free = (spool->count == 0) && (spool->used_hpages == 0); 7690481622SDavid Gibson 7790481622SDavid Gibson spin_unlock(&spool->lock); 7890481622SDavid Gibson 7990481622SDavid Gibson /* If no pages are used, and no other handles to the subpool 807ca02d0aSMike Kravetz * remain, give up any reservations mased on minimum size and 817ca02d0aSMike Kravetz * free the subpool */ 827ca02d0aSMike Kravetz if (free) { 837ca02d0aSMike Kravetz if (spool->min_hpages != -1) 847ca02d0aSMike Kravetz hugetlb_acct_memory(spool->hstate, 857ca02d0aSMike Kravetz -spool->min_hpages); 8690481622SDavid Gibson kfree(spool); 8790481622SDavid Gibson } 887ca02d0aSMike Kravetz } 8990481622SDavid Gibson 907ca02d0aSMike Kravetz struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, 917ca02d0aSMike Kravetz long min_hpages) 9290481622SDavid Gibson { 9390481622SDavid Gibson struct hugepage_subpool *spool; 9490481622SDavid Gibson 95c6a91820SMike Kravetz spool = kzalloc(sizeof(*spool), GFP_KERNEL); 9690481622SDavid Gibson if (!spool) 9790481622SDavid Gibson return NULL; 9890481622SDavid Gibson 9990481622SDavid Gibson spin_lock_init(&spool->lock); 10090481622SDavid Gibson spool->count = 1; 1017ca02d0aSMike Kravetz spool->max_hpages = max_hpages; 1027ca02d0aSMike Kravetz spool->hstate = h; 1037ca02d0aSMike Kravetz spool->min_hpages = min_hpages; 1047ca02d0aSMike Kravetz 1057ca02d0aSMike Kravetz if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { 1067ca02d0aSMike Kravetz kfree(spool); 1077ca02d0aSMike Kravetz return NULL; 1087ca02d0aSMike Kravetz } 1097ca02d0aSMike Kravetz spool->rsv_hpages = min_hpages; 11090481622SDavid Gibson 11190481622SDavid Gibson return spool; 11290481622SDavid Gibson } 11390481622SDavid Gibson 11490481622SDavid Gibson void hugepage_put_subpool(struct hugepage_subpool *spool) 11590481622SDavid Gibson { 11690481622SDavid Gibson spin_lock(&spool->lock); 11790481622SDavid Gibson BUG_ON(!spool->count); 11890481622SDavid Gibson spool->count--; 11990481622SDavid Gibson unlock_or_release_subpool(spool); 12090481622SDavid Gibson } 12190481622SDavid Gibson 1221c5ecae3SMike Kravetz /* 1231c5ecae3SMike Kravetz * Subpool accounting for allocating and reserving pages. 1241c5ecae3SMike Kravetz * Return -ENOMEM if there are not enough resources to satisfy the 1251c5ecae3SMike Kravetz * the request. Otherwise, return the number of pages by which the 1261c5ecae3SMike Kravetz * global pools must be adjusted (upward). The returned value may 1271c5ecae3SMike Kravetz * only be different than the passed value (delta) in the case where 1281c5ecae3SMike Kravetz * a subpool minimum size must be manitained. 1291c5ecae3SMike Kravetz */ 1301c5ecae3SMike Kravetz static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, 13190481622SDavid Gibson long delta) 13290481622SDavid Gibson { 1331c5ecae3SMike Kravetz long ret = delta; 13490481622SDavid Gibson 13590481622SDavid Gibson if (!spool) 1361c5ecae3SMike Kravetz return ret; 13790481622SDavid Gibson 13890481622SDavid Gibson spin_lock(&spool->lock); 13990481622SDavid Gibson 1401c5ecae3SMike Kravetz if (spool->max_hpages != -1) { /* maximum size accounting */ 1411c5ecae3SMike Kravetz if ((spool->used_hpages + delta) <= spool->max_hpages) 1421c5ecae3SMike Kravetz spool->used_hpages += delta; 1431c5ecae3SMike Kravetz else { 1441c5ecae3SMike Kravetz ret = -ENOMEM; 1451c5ecae3SMike Kravetz goto unlock_ret; 1461c5ecae3SMike Kravetz } 1471c5ecae3SMike Kravetz } 1481c5ecae3SMike Kravetz 14909a95e29SMike Kravetz /* minimum size accounting */ 15009a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->rsv_hpages) { 1511c5ecae3SMike Kravetz if (delta > spool->rsv_hpages) { 1521c5ecae3SMike Kravetz /* 1531c5ecae3SMike Kravetz * Asking for more reserves than those already taken on 1541c5ecae3SMike Kravetz * behalf of subpool. Return difference. 1551c5ecae3SMike Kravetz */ 1561c5ecae3SMike Kravetz ret = delta - spool->rsv_hpages; 1571c5ecae3SMike Kravetz spool->rsv_hpages = 0; 1581c5ecae3SMike Kravetz } else { 1591c5ecae3SMike Kravetz ret = 0; /* reserves already accounted for */ 1601c5ecae3SMike Kravetz spool->rsv_hpages -= delta; 1611c5ecae3SMike Kravetz } 1621c5ecae3SMike Kravetz } 1631c5ecae3SMike Kravetz 1641c5ecae3SMike Kravetz unlock_ret: 1651c5ecae3SMike Kravetz spin_unlock(&spool->lock); 16690481622SDavid Gibson return ret; 16790481622SDavid Gibson } 16890481622SDavid Gibson 1691c5ecae3SMike Kravetz /* 1701c5ecae3SMike Kravetz * Subpool accounting for freeing and unreserving pages. 1711c5ecae3SMike Kravetz * Return the number of global page reservations that must be dropped. 1721c5ecae3SMike Kravetz * The return value may only be different than the passed value (delta) 1731c5ecae3SMike Kravetz * in the case where a subpool minimum size must be maintained. 1741c5ecae3SMike Kravetz */ 1751c5ecae3SMike Kravetz static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, 17690481622SDavid Gibson long delta) 17790481622SDavid Gibson { 1781c5ecae3SMike Kravetz long ret = delta; 1791c5ecae3SMike Kravetz 18090481622SDavid Gibson if (!spool) 1811c5ecae3SMike Kravetz return delta; 18290481622SDavid Gibson 18390481622SDavid Gibson spin_lock(&spool->lock); 1841c5ecae3SMike Kravetz 1851c5ecae3SMike Kravetz if (spool->max_hpages != -1) /* maximum size accounting */ 18690481622SDavid Gibson spool->used_hpages -= delta; 1871c5ecae3SMike Kravetz 18809a95e29SMike Kravetz /* minimum size accounting */ 18909a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { 1901c5ecae3SMike Kravetz if (spool->rsv_hpages + delta <= spool->min_hpages) 1911c5ecae3SMike Kravetz ret = 0; 1921c5ecae3SMike Kravetz else 1931c5ecae3SMike Kravetz ret = spool->rsv_hpages + delta - spool->min_hpages; 1941c5ecae3SMike Kravetz 1951c5ecae3SMike Kravetz spool->rsv_hpages += delta; 1961c5ecae3SMike Kravetz if (spool->rsv_hpages > spool->min_hpages) 1971c5ecae3SMike Kravetz spool->rsv_hpages = spool->min_hpages; 1981c5ecae3SMike Kravetz } 1991c5ecae3SMike Kravetz 2001c5ecae3SMike Kravetz /* 2011c5ecae3SMike Kravetz * If hugetlbfs_put_super couldn't free spool due to an outstanding 2021c5ecae3SMike Kravetz * quota reference, free it now. 2031c5ecae3SMike Kravetz */ 20490481622SDavid Gibson unlock_or_release_subpool(spool); 2051c5ecae3SMike Kravetz 2061c5ecae3SMike Kravetz return ret; 20790481622SDavid Gibson } 20890481622SDavid Gibson 20990481622SDavid Gibson static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 21090481622SDavid Gibson { 21190481622SDavid Gibson return HUGETLBFS_SB(inode->i_sb)->spool; 21290481622SDavid Gibson } 21390481622SDavid Gibson 21490481622SDavid Gibson static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 21590481622SDavid Gibson { 216496ad9aaSAl Viro return subpool_inode(file_inode(vma->vm_file)); 21790481622SDavid Gibson } 21890481622SDavid Gibson 219e7c4b0bfSAndy Whitcroft /* 22096822904SAndy Whitcroft * Region tracking -- allows tracking of reservations and instantiated pages 22196822904SAndy Whitcroft * across the pages in a mapping. 22284afd99bSAndy Whitcroft * 2231dd308a7SMike Kravetz * The region data structures are embedded into a resv_map and protected 2241dd308a7SMike Kravetz * by a resv_map's lock. The set of regions within the resv_map represent 2251dd308a7SMike Kravetz * reservations for huge pages, or huge pages that have already been 2261dd308a7SMike Kravetz * instantiated within the map. The from and to elements are huge page 2271dd308a7SMike Kravetz * indicies into the associated mapping. from indicates the starting index 2281dd308a7SMike Kravetz * of the region. to represents the first index past the end of the region. 2291dd308a7SMike Kravetz * 2301dd308a7SMike Kravetz * For example, a file region structure with from == 0 and to == 4 represents 2311dd308a7SMike Kravetz * four huge pages in a mapping. It is important to note that the to element 2321dd308a7SMike Kravetz * represents the first element past the end of the region. This is used in 2331dd308a7SMike Kravetz * arithmetic as 4(to) - 0(from) = 4 huge pages in the region. 2341dd308a7SMike Kravetz * 2351dd308a7SMike Kravetz * Interval notation of the form [from, to) will be used to indicate that 2361dd308a7SMike Kravetz * the endpoint from is inclusive and to is exclusive. 23796822904SAndy Whitcroft */ 23896822904SAndy Whitcroft struct file_region { 23996822904SAndy Whitcroft struct list_head link; 24096822904SAndy Whitcroft long from; 24196822904SAndy Whitcroft long to; 24296822904SAndy Whitcroft }; 24396822904SAndy Whitcroft 2441dd308a7SMike Kravetz /* 2451dd308a7SMike Kravetz * Add the huge page range represented by [f, t) to the reserve 2465e911373SMike Kravetz * map. In the normal case, existing regions will be expanded 2475e911373SMike Kravetz * to accommodate the specified range. Sufficient regions should 2485e911373SMike Kravetz * exist for expansion due to the previous call to region_chg 2495e911373SMike Kravetz * with the same range. However, it is possible that region_del 2505e911373SMike Kravetz * could have been called after region_chg and modifed the map 2515e911373SMike Kravetz * in such a way that no region exists to be expanded. In this 2525e911373SMike Kravetz * case, pull a region descriptor from the cache associated with 2535e911373SMike Kravetz * the map and use that for the new range. 254cf3ad20bSMike Kravetz * 255cf3ad20bSMike Kravetz * Return the number of new huge pages added to the map. This 256cf3ad20bSMike Kravetz * number is greater than or equal to zero. 2571dd308a7SMike Kravetz */ 2581406ec9bSJoonsoo Kim static long region_add(struct resv_map *resv, long f, long t) 25996822904SAndy Whitcroft { 2601406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 26196822904SAndy Whitcroft struct file_region *rg, *nrg, *trg; 262cf3ad20bSMike Kravetz long add = 0; 26396822904SAndy Whitcroft 2647b24d861SDavidlohr Bueso spin_lock(&resv->lock); 26596822904SAndy Whitcroft /* Locate the region we are either in or before. */ 26696822904SAndy Whitcroft list_for_each_entry(rg, head, link) 26796822904SAndy Whitcroft if (f <= rg->to) 26896822904SAndy Whitcroft break; 26996822904SAndy Whitcroft 2705e911373SMike Kravetz /* 2715e911373SMike Kravetz * If no region exists which can be expanded to include the 2725e911373SMike Kravetz * specified range, the list must have been modified by an 2735e911373SMike Kravetz * interleving call to region_del(). Pull a region descriptor 2745e911373SMike Kravetz * from the cache and use it for this range. 2755e911373SMike Kravetz */ 2765e911373SMike Kravetz if (&rg->link == head || t < rg->from) { 2775e911373SMike Kravetz VM_BUG_ON(resv->region_cache_count <= 0); 2785e911373SMike Kravetz 2795e911373SMike Kravetz resv->region_cache_count--; 2805e911373SMike Kravetz nrg = list_first_entry(&resv->region_cache, struct file_region, 2815e911373SMike Kravetz link); 2825e911373SMike Kravetz list_del(&nrg->link); 2835e911373SMike Kravetz 2845e911373SMike Kravetz nrg->from = f; 2855e911373SMike Kravetz nrg->to = t; 2865e911373SMike Kravetz list_add(&nrg->link, rg->link.prev); 2875e911373SMike Kravetz 2885e911373SMike Kravetz add += t - f; 2895e911373SMike Kravetz goto out_locked; 2905e911373SMike Kravetz } 2915e911373SMike Kravetz 29296822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 29396822904SAndy Whitcroft if (f > rg->from) 29496822904SAndy Whitcroft f = rg->from; 29596822904SAndy Whitcroft 29696822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 29796822904SAndy Whitcroft nrg = rg; 29896822904SAndy Whitcroft list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 29996822904SAndy Whitcroft if (&rg->link == head) 30096822904SAndy Whitcroft break; 30196822904SAndy Whitcroft if (rg->from > t) 30296822904SAndy Whitcroft break; 30396822904SAndy Whitcroft 30496822904SAndy Whitcroft /* If this area reaches higher then extend our area to 30596822904SAndy Whitcroft * include it completely. If this is not the first area 30696822904SAndy Whitcroft * which we intend to reuse, free it. */ 30796822904SAndy Whitcroft if (rg->to > t) 30896822904SAndy Whitcroft t = rg->to; 30996822904SAndy Whitcroft if (rg != nrg) { 310cf3ad20bSMike Kravetz /* Decrement return value by the deleted range. 311cf3ad20bSMike Kravetz * Another range will span this area so that by 312cf3ad20bSMike Kravetz * end of routine add will be >= zero 313cf3ad20bSMike Kravetz */ 314cf3ad20bSMike Kravetz add -= (rg->to - rg->from); 31596822904SAndy Whitcroft list_del(&rg->link); 31696822904SAndy Whitcroft kfree(rg); 31796822904SAndy Whitcroft } 31896822904SAndy Whitcroft } 319cf3ad20bSMike Kravetz 320cf3ad20bSMike Kravetz add += (nrg->from - f); /* Added to beginning of region */ 32196822904SAndy Whitcroft nrg->from = f; 322cf3ad20bSMike Kravetz add += t - nrg->to; /* Added to end of region */ 32396822904SAndy Whitcroft nrg->to = t; 324cf3ad20bSMike Kravetz 3255e911373SMike Kravetz out_locked: 3265e911373SMike Kravetz resv->adds_in_progress--; 3277b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 328cf3ad20bSMike Kravetz VM_BUG_ON(add < 0); 329cf3ad20bSMike Kravetz return add; 33096822904SAndy Whitcroft } 33196822904SAndy Whitcroft 3321dd308a7SMike Kravetz /* 3331dd308a7SMike Kravetz * Examine the existing reserve map and determine how many 3341dd308a7SMike Kravetz * huge pages in the specified range [f, t) are NOT currently 3351dd308a7SMike Kravetz * represented. This routine is called before a subsequent 3361dd308a7SMike Kravetz * call to region_add that will actually modify the reserve 3371dd308a7SMike Kravetz * map to add the specified range [f, t). region_chg does 3381dd308a7SMike Kravetz * not change the number of huge pages represented by the 3391dd308a7SMike Kravetz * map. However, if the existing regions in the map can not 3401dd308a7SMike Kravetz * be expanded to represent the new range, a new file_region 3411dd308a7SMike Kravetz * structure is added to the map as a placeholder. This is 3421dd308a7SMike Kravetz * so that the subsequent region_add call will have all the 3431dd308a7SMike Kravetz * regions it needs and will not fail. 3441dd308a7SMike Kravetz * 3455e911373SMike Kravetz * Upon entry, region_chg will also examine the cache of region descriptors 3465e911373SMike Kravetz * associated with the map. If there are not enough descriptors cached, one 3475e911373SMike Kravetz * will be allocated for the in progress add operation. 3485e911373SMike Kravetz * 3495e911373SMike Kravetz * Returns the number of huge pages that need to be added to the existing 3505e911373SMike Kravetz * reservation map for the range [f, t). This number is greater or equal to 3515e911373SMike Kravetz * zero. -ENOMEM is returned if a new file_region structure or cache entry 3525e911373SMike Kravetz * is needed and can not be allocated. 3531dd308a7SMike Kravetz */ 3541406ec9bSJoonsoo Kim static long region_chg(struct resv_map *resv, long f, long t) 35596822904SAndy Whitcroft { 3561406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 3577b24d861SDavidlohr Bueso struct file_region *rg, *nrg = NULL; 35896822904SAndy Whitcroft long chg = 0; 35996822904SAndy Whitcroft 3607b24d861SDavidlohr Bueso retry: 3617b24d861SDavidlohr Bueso spin_lock(&resv->lock); 3625e911373SMike Kravetz retry_locked: 3635e911373SMike Kravetz resv->adds_in_progress++; 3645e911373SMike Kravetz 3655e911373SMike Kravetz /* 3665e911373SMike Kravetz * Check for sufficient descriptors in the cache to accommodate 3675e911373SMike Kravetz * the number of in progress add operations. 3685e911373SMike Kravetz */ 3695e911373SMike Kravetz if (resv->adds_in_progress > resv->region_cache_count) { 3705e911373SMike Kravetz struct file_region *trg; 3715e911373SMike Kravetz 3725e911373SMike Kravetz VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1); 3735e911373SMike Kravetz /* Must drop lock to allocate a new descriptor. */ 3745e911373SMike Kravetz resv->adds_in_progress--; 3755e911373SMike Kravetz spin_unlock(&resv->lock); 3765e911373SMike Kravetz 3775e911373SMike Kravetz trg = kmalloc(sizeof(*trg), GFP_KERNEL); 378dbe409e4SMike Kravetz if (!trg) { 379dbe409e4SMike Kravetz kfree(nrg); 3805e911373SMike Kravetz return -ENOMEM; 381dbe409e4SMike Kravetz } 3825e911373SMike Kravetz 3835e911373SMike Kravetz spin_lock(&resv->lock); 3845e911373SMike Kravetz list_add(&trg->link, &resv->region_cache); 3855e911373SMike Kravetz resv->region_cache_count++; 3865e911373SMike Kravetz goto retry_locked; 3875e911373SMike Kravetz } 3885e911373SMike Kravetz 38996822904SAndy Whitcroft /* Locate the region we are before or in. */ 39096822904SAndy Whitcroft list_for_each_entry(rg, head, link) 39196822904SAndy Whitcroft if (f <= rg->to) 39296822904SAndy Whitcroft break; 39396822904SAndy Whitcroft 39496822904SAndy Whitcroft /* If we are below the current region then a new region is required. 39596822904SAndy Whitcroft * Subtle, allocate a new region at the position but make it zero 39696822904SAndy Whitcroft * size such that we can guarantee to record the reservation. */ 39796822904SAndy Whitcroft if (&rg->link == head || t < rg->from) { 3987b24d861SDavidlohr Bueso if (!nrg) { 3995e911373SMike Kravetz resv->adds_in_progress--; 4007b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 40196822904SAndy Whitcroft nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 40296822904SAndy Whitcroft if (!nrg) 40396822904SAndy Whitcroft return -ENOMEM; 4047b24d861SDavidlohr Bueso 40596822904SAndy Whitcroft nrg->from = f; 40696822904SAndy Whitcroft nrg->to = f; 40796822904SAndy Whitcroft INIT_LIST_HEAD(&nrg->link); 4087b24d861SDavidlohr Bueso goto retry; 4097b24d861SDavidlohr Bueso } 41096822904SAndy Whitcroft 4117b24d861SDavidlohr Bueso list_add(&nrg->link, rg->link.prev); 4127b24d861SDavidlohr Bueso chg = t - f; 4137b24d861SDavidlohr Bueso goto out_nrg; 41496822904SAndy Whitcroft } 41596822904SAndy Whitcroft 41696822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 41796822904SAndy Whitcroft if (f > rg->from) 41896822904SAndy Whitcroft f = rg->from; 41996822904SAndy Whitcroft chg = t - f; 42096822904SAndy Whitcroft 42196822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 42296822904SAndy Whitcroft list_for_each_entry(rg, rg->link.prev, link) { 42396822904SAndy Whitcroft if (&rg->link == head) 42496822904SAndy Whitcroft break; 42596822904SAndy Whitcroft if (rg->from > t) 4267b24d861SDavidlohr Bueso goto out; 42796822904SAndy Whitcroft 42825985edcSLucas De Marchi /* We overlap with this area, if it extends further than 42996822904SAndy Whitcroft * us then we must extend ourselves. Account for its 43096822904SAndy Whitcroft * existing reservation. */ 43196822904SAndy Whitcroft if (rg->to > t) { 43296822904SAndy Whitcroft chg += rg->to - t; 43396822904SAndy Whitcroft t = rg->to; 43496822904SAndy Whitcroft } 43596822904SAndy Whitcroft chg -= rg->to - rg->from; 43696822904SAndy Whitcroft } 4377b24d861SDavidlohr Bueso 4387b24d861SDavidlohr Bueso out: 4397b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 4407b24d861SDavidlohr Bueso /* We already know we raced and no longer need the new region */ 4417b24d861SDavidlohr Bueso kfree(nrg); 4427b24d861SDavidlohr Bueso return chg; 4437b24d861SDavidlohr Bueso out_nrg: 4447b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 44596822904SAndy Whitcroft return chg; 44696822904SAndy Whitcroft } 44796822904SAndy Whitcroft 4481dd308a7SMike Kravetz /* 4495e911373SMike Kravetz * Abort the in progress add operation. The adds_in_progress field 4505e911373SMike Kravetz * of the resv_map keeps track of the operations in progress between 4515e911373SMike Kravetz * calls to region_chg and region_add. Operations are sometimes 4525e911373SMike Kravetz * aborted after the call to region_chg. In such cases, region_abort 4535e911373SMike Kravetz * is called to decrement the adds_in_progress counter. 4545e911373SMike Kravetz * 4555e911373SMike Kravetz * NOTE: The range arguments [f, t) are not needed or used in this 4565e911373SMike Kravetz * routine. They are kept to make reading the calling code easier as 4575e911373SMike Kravetz * arguments will match the associated region_chg call. 4585e911373SMike Kravetz */ 4595e911373SMike Kravetz static void region_abort(struct resv_map *resv, long f, long t) 4605e911373SMike Kravetz { 4615e911373SMike Kravetz spin_lock(&resv->lock); 4625e911373SMike Kravetz VM_BUG_ON(!resv->region_cache_count); 4635e911373SMike Kravetz resv->adds_in_progress--; 4645e911373SMike Kravetz spin_unlock(&resv->lock); 4655e911373SMike Kravetz } 4665e911373SMike Kravetz 4675e911373SMike Kravetz /* 468feba16e2SMike Kravetz * Delete the specified range [f, t) from the reserve map. If the 469feba16e2SMike Kravetz * t parameter is LONG_MAX, this indicates that ALL regions after f 470feba16e2SMike Kravetz * should be deleted. Locate the regions which intersect [f, t) 471feba16e2SMike Kravetz * and either trim, delete or split the existing regions. 472feba16e2SMike Kravetz * 473feba16e2SMike Kravetz * Returns the number of huge pages deleted from the reserve map. 474feba16e2SMike Kravetz * In the normal case, the return value is zero or more. In the 475feba16e2SMike Kravetz * case where a region must be split, a new region descriptor must 476feba16e2SMike Kravetz * be allocated. If the allocation fails, -ENOMEM will be returned. 477feba16e2SMike Kravetz * NOTE: If the parameter t == LONG_MAX, then we will never split 478feba16e2SMike Kravetz * a region and possibly return -ENOMEM. Callers specifying 479feba16e2SMike Kravetz * t == LONG_MAX do not need to check for -ENOMEM error. 4801dd308a7SMike Kravetz */ 481feba16e2SMike Kravetz static long region_del(struct resv_map *resv, long f, long t) 48296822904SAndy Whitcroft { 4831406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 48496822904SAndy Whitcroft struct file_region *rg, *trg; 485feba16e2SMike Kravetz struct file_region *nrg = NULL; 486feba16e2SMike Kravetz long del = 0; 48796822904SAndy Whitcroft 488feba16e2SMike Kravetz retry: 4897b24d861SDavidlohr Bueso spin_lock(&resv->lock); 490feba16e2SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 491dbe409e4SMike Kravetz /* 492dbe409e4SMike Kravetz * Skip regions before the range to be deleted. file_region 493dbe409e4SMike Kravetz * ranges are normally of the form [from, to). However, there 494dbe409e4SMike Kravetz * may be a "placeholder" entry in the map which is of the form 495dbe409e4SMike Kravetz * (from, to) with from == to. Check for placeholder entries 496dbe409e4SMike Kravetz * at the beginning of the range to be deleted. 497dbe409e4SMike Kravetz */ 498dbe409e4SMike Kravetz if (rg->to <= f && (rg->to != rg->from || rg->to != f)) 499feba16e2SMike Kravetz continue; 500dbe409e4SMike Kravetz 501feba16e2SMike Kravetz if (rg->from >= t) 50296822904SAndy Whitcroft break; 50396822904SAndy Whitcroft 504feba16e2SMike Kravetz if (f > rg->from && t < rg->to) { /* Must split region */ 505feba16e2SMike Kravetz /* 506feba16e2SMike Kravetz * Check for an entry in the cache before dropping 507feba16e2SMike Kravetz * lock and attempting allocation. 508feba16e2SMike Kravetz */ 509feba16e2SMike Kravetz if (!nrg && 510feba16e2SMike Kravetz resv->region_cache_count > resv->adds_in_progress) { 511feba16e2SMike Kravetz nrg = list_first_entry(&resv->region_cache, 512feba16e2SMike Kravetz struct file_region, 513feba16e2SMike Kravetz link); 514feba16e2SMike Kravetz list_del(&nrg->link); 515feba16e2SMike Kravetz resv->region_cache_count--; 51696822904SAndy Whitcroft } 51796822904SAndy Whitcroft 518feba16e2SMike Kravetz if (!nrg) { 519feba16e2SMike Kravetz spin_unlock(&resv->lock); 520feba16e2SMike Kravetz nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 521feba16e2SMike Kravetz if (!nrg) 522feba16e2SMike Kravetz return -ENOMEM; 523feba16e2SMike Kravetz goto retry; 524feba16e2SMike Kravetz } 525feba16e2SMike Kravetz 526feba16e2SMike Kravetz del += t - f; 527feba16e2SMike Kravetz 528feba16e2SMike Kravetz /* New entry for end of split region */ 529feba16e2SMike Kravetz nrg->from = t; 530feba16e2SMike Kravetz nrg->to = rg->to; 531feba16e2SMike Kravetz INIT_LIST_HEAD(&nrg->link); 532feba16e2SMike Kravetz 533feba16e2SMike Kravetz /* Original entry is trimmed */ 534feba16e2SMike Kravetz rg->to = f; 535feba16e2SMike Kravetz 536feba16e2SMike Kravetz list_add(&nrg->link, &rg->link); 537feba16e2SMike Kravetz nrg = NULL; 53896822904SAndy Whitcroft break; 539feba16e2SMike Kravetz } 540feba16e2SMike Kravetz 541feba16e2SMike Kravetz if (f <= rg->from && t >= rg->to) { /* Remove entire region */ 542feba16e2SMike Kravetz del += rg->to - rg->from; 54396822904SAndy Whitcroft list_del(&rg->link); 54496822904SAndy Whitcroft kfree(rg); 545feba16e2SMike Kravetz continue; 54696822904SAndy Whitcroft } 5477b24d861SDavidlohr Bueso 548feba16e2SMike Kravetz if (f <= rg->from) { /* Trim beginning of region */ 549feba16e2SMike Kravetz del += t - rg->from; 550feba16e2SMike Kravetz rg->from = t; 551feba16e2SMike Kravetz } else { /* Trim end of region */ 552feba16e2SMike Kravetz del += rg->to - f; 553feba16e2SMike Kravetz rg->to = f; 554feba16e2SMike Kravetz } 555feba16e2SMike Kravetz } 556feba16e2SMike Kravetz 5577b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 558feba16e2SMike Kravetz kfree(nrg); 559feba16e2SMike Kravetz return del; 56096822904SAndy Whitcroft } 56196822904SAndy Whitcroft 5621dd308a7SMike Kravetz /* 563b5cec28dSMike Kravetz * A rare out of memory error was encountered which prevented removal of 564b5cec28dSMike Kravetz * the reserve map region for a page. The huge page itself was free'ed 565b5cec28dSMike Kravetz * and removed from the page cache. This routine will adjust the subpool 566b5cec28dSMike Kravetz * usage count, and the global reserve count if needed. By incrementing 567b5cec28dSMike Kravetz * these counts, the reserve map entry which could not be deleted will 568b5cec28dSMike Kravetz * appear as a "reserved" entry instead of simply dangling with incorrect 569b5cec28dSMike Kravetz * counts. 570b5cec28dSMike Kravetz */ 57172e2936cSzhong jiang void hugetlb_fix_reserve_counts(struct inode *inode) 572b5cec28dSMike Kravetz { 573b5cec28dSMike Kravetz struct hugepage_subpool *spool = subpool_inode(inode); 574b5cec28dSMike Kravetz long rsv_adjust; 575b5cec28dSMike Kravetz 576b5cec28dSMike Kravetz rsv_adjust = hugepage_subpool_get_pages(spool, 1); 57772e2936cSzhong jiang if (rsv_adjust) { 578b5cec28dSMike Kravetz struct hstate *h = hstate_inode(inode); 579b5cec28dSMike Kravetz 580b5cec28dSMike Kravetz hugetlb_acct_memory(h, 1); 581b5cec28dSMike Kravetz } 582b5cec28dSMike Kravetz } 583b5cec28dSMike Kravetz 584b5cec28dSMike Kravetz /* 5851dd308a7SMike Kravetz * Count and return the number of huge pages in the reserve map 5861dd308a7SMike Kravetz * that intersect with the range [f, t). 5871dd308a7SMike Kravetz */ 5881406ec9bSJoonsoo Kim static long region_count(struct resv_map *resv, long f, long t) 58984afd99bSAndy Whitcroft { 5901406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 59184afd99bSAndy Whitcroft struct file_region *rg; 59284afd99bSAndy Whitcroft long chg = 0; 59384afd99bSAndy Whitcroft 5947b24d861SDavidlohr Bueso spin_lock(&resv->lock); 59584afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 59684afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 597f2135a4aSWang Sheng-Hui long seg_from; 598f2135a4aSWang Sheng-Hui long seg_to; 59984afd99bSAndy Whitcroft 60084afd99bSAndy Whitcroft if (rg->to <= f) 60184afd99bSAndy Whitcroft continue; 60284afd99bSAndy Whitcroft if (rg->from >= t) 60384afd99bSAndy Whitcroft break; 60484afd99bSAndy Whitcroft 60584afd99bSAndy Whitcroft seg_from = max(rg->from, f); 60684afd99bSAndy Whitcroft seg_to = min(rg->to, t); 60784afd99bSAndy Whitcroft 60884afd99bSAndy Whitcroft chg += seg_to - seg_from; 60984afd99bSAndy Whitcroft } 6107b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 61184afd99bSAndy Whitcroft 61284afd99bSAndy Whitcroft return chg; 61384afd99bSAndy Whitcroft } 61484afd99bSAndy Whitcroft 61596822904SAndy Whitcroft /* 616e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 617e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 618e7c4b0bfSAndy Whitcroft */ 619a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 620a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 621e7c4b0bfSAndy Whitcroft { 622a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 623a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 624e7c4b0bfSAndy Whitcroft } 625e7c4b0bfSAndy Whitcroft 6260fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 6270fe6e20bSNaoya Horiguchi unsigned long address) 6280fe6e20bSNaoya Horiguchi { 6290fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 6300fe6e20bSNaoya Horiguchi } 631dee41079SDan Williams EXPORT_SYMBOL_GPL(linear_hugepage_index); 6320fe6e20bSNaoya Horiguchi 63384afd99bSAndy Whitcroft /* 63408fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 63508fba699SMel Gorman * cases this will be same size as used by the page table entries. 63608fba699SMel Gorman */ 63708fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 63808fba699SMel Gorman { 63908fba699SMel Gorman struct hstate *hstate; 64008fba699SMel Gorman 64108fba699SMel Gorman if (!is_vm_hugetlb_page(vma)) 64208fba699SMel Gorman return PAGE_SIZE; 64308fba699SMel Gorman 64408fba699SMel Gorman hstate = hstate_vma(vma); 64508fba699SMel Gorman 6462415cf12SWanpeng Li return 1UL << huge_page_shift(hstate); 64708fba699SMel Gorman } 648f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 64908fba699SMel Gorman 65008fba699SMel Gorman /* 6513340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 6523340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 6533340289dSMel Gorman * architectures where it differs, an architecture-specific version of this 6543340289dSMel Gorman * function is required. 6553340289dSMel Gorman */ 6563340289dSMel Gorman #ifndef vma_mmu_pagesize 6573340289dSMel Gorman unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 6583340289dSMel Gorman { 6593340289dSMel Gorman return vma_kernel_pagesize(vma); 6603340289dSMel Gorman } 6613340289dSMel Gorman #endif 6623340289dSMel Gorman 6633340289dSMel Gorman /* 66484afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 66584afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 66684afd99bSAndy Whitcroft * alignment. 66784afd99bSAndy Whitcroft */ 66884afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 66984afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 67004f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 67184afd99bSAndy Whitcroft 672a1e78772SMel Gorman /* 673a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 674a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 675a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 676a1e78772SMel Gorman * 677a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 678a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 679a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 680a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 68184afd99bSAndy Whitcroft * 68284afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 68384afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 68484afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 68584afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 68684afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 68784afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 68884afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 68984afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 690a1e78772SMel Gorman */ 691e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 692e7c4b0bfSAndy Whitcroft { 693e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 694e7c4b0bfSAndy Whitcroft } 695e7c4b0bfSAndy Whitcroft 696e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 697e7c4b0bfSAndy Whitcroft unsigned long value) 698e7c4b0bfSAndy Whitcroft { 699e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 700e7c4b0bfSAndy Whitcroft } 701e7c4b0bfSAndy Whitcroft 7029119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 70384afd99bSAndy Whitcroft { 70484afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 7055e911373SMike Kravetz struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 7065e911373SMike Kravetz 7075e911373SMike Kravetz if (!resv_map || !rg) { 7085e911373SMike Kravetz kfree(resv_map); 7095e911373SMike Kravetz kfree(rg); 71084afd99bSAndy Whitcroft return NULL; 7115e911373SMike Kravetz } 71284afd99bSAndy Whitcroft 71384afd99bSAndy Whitcroft kref_init(&resv_map->refs); 7147b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 71584afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 71684afd99bSAndy Whitcroft 7175e911373SMike Kravetz resv_map->adds_in_progress = 0; 7185e911373SMike Kravetz 7195e911373SMike Kravetz INIT_LIST_HEAD(&resv_map->region_cache); 7205e911373SMike Kravetz list_add(&rg->link, &resv_map->region_cache); 7215e911373SMike Kravetz resv_map->region_cache_count = 1; 7225e911373SMike Kravetz 72384afd99bSAndy Whitcroft return resv_map; 72484afd99bSAndy Whitcroft } 72584afd99bSAndy Whitcroft 7269119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 72784afd99bSAndy Whitcroft { 72884afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 7295e911373SMike Kravetz struct list_head *head = &resv_map->region_cache; 7305e911373SMike Kravetz struct file_region *rg, *trg; 73184afd99bSAndy Whitcroft 73284afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 733feba16e2SMike Kravetz region_del(resv_map, 0, LONG_MAX); 7345e911373SMike Kravetz 7355e911373SMike Kravetz /* ... and any entries left in the cache */ 7365e911373SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 7375e911373SMike Kravetz list_del(&rg->link); 7385e911373SMike Kravetz kfree(rg); 7395e911373SMike Kravetz } 7405e911373SMike Kravetz 7415e911373SMike Kravetz VM_BUG_ON(resv_map->adds_in_progress); 7425e911373SMike Kravetz 74384afd99bSAndy Whitcroft kfree(resv_map); 74484afd99bSAndy Whitcroft } 74584afd99bSAndy Whitcroft 7464e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 7474e35f483SJoonsoo Kim { 7484e35f483SJoonsoo Kim return inode->i_mapping->private_data; 7494e35f483SJoonsoo Kim } 7504e35f483SJoonsoo Kim 75184afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 752a1e78772SMel Gorman { 75381d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 7544e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 7554e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 7564e35f483SJoonsoo Kim struct inode *inode = mapping->host; 7574e35f483SJoonsoo Kim 7584e35f483SJoonsoo Kim return inode_resv_map(inode); 7594e35f483SJoonsoo Kim 7604e35f483SJoonsoo Kim } else { 76184afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 76284afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 7634e35f483SJoonsoo Kim } 764a1e78772SMel Gorman } 765a1e78772SMel Gorman 76684afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 767a1e78772SMel Gorman { 76881d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 76981d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 770a1e78772SMel Gorman 77184afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 77284afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 77304f2cbe3SMel Gorman } 77404f2cbe3SMel Gorman 77504f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 77604f2cbe3SMel Gorman { 77781d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 77881d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 779e7c4b0bfSAndy Whitcroft 780e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 78104f2cbe3SMel Gorman } 78204f2cbe3SMel Gorman 78304f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 78404f2cbe3SMel Gorman { 78581d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 786e7c4b0bfSAndy Whitcroft 787e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 788a1e78772SMel Gorman } 789a1e78772SMel Gorman 79004f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 791a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 792a1e78772SMel Gorman { 79381d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 794f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 795a1e78772SMel Gorman vma->vm_private_data = (void *)0; 796a1e78772SMel Gorman } 797a1e78772SMel Gorman 798a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 799559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 800a1e78772SMel Gorman { 801af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 802af0ed73eSJoonsoo Kim /* 803af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 804af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 805af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 806af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 807af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 808af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 809af0ed73eSJoonsoo Kim * properly, so add work-around here. 810af0ed73eSJoonsoo Kim */ 811af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 812559ec2f8SNicholas Krause return true; 813af0ed73eSJoonsoo Kim else 814559ec2f8SNicholas Krause return false; 815af0ed73eSJoonsoo Kim } 816a63884e9SJoonsoo Kim 817a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 8181fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 8191fb1b0e9SMike Kravetz /* 8201fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 8211fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 8221fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 8231fb1b0e9SMike Kravetz * fallocate. In this case, there really are no reverves to 8241fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 8251fb1b0e9SMike Kravetz */ 8261fb1b0e9SMike Kravetz if (chg) 8271fb1b0e9SMike Kravetz return false; 8281fb1b0e9SMike Kravetz else 829559ec2f8SNicholas Krause return true; 8301fb1b0e9SMike Kravetz } 831a63884e9SJoonsoo Kim 832a63884e9SJoonsoo Kim /* 833a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 834a63884e9SJoonsoo Kim * private mappings. 835a63884e9SJoonsoo Kim */ 83667961f9dSMike Kravetz if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 83767961f9dSMike Kravetz /* 83867961f9dSMike Kravetz * Like the shared case above, a hole punch or truncate 83967961f9dSMike Kravetz * could have been performed on the private mapping. 84067961f9dSMike Kravetz * Examine the value of chg to determine if reserves 84167961f9dSMike Kravetz * actually exist or were previously consumed. 84267961f9dSMike Kravetz * Very Subtle - The value of chg comes from a previous 84367961f9dSMike Kravetz * call to vma_needs_reserves(). The reserve map for 84467961f9dSMike Kravetz * private mappings has different (opposite) semantics 84567961f9dSMike Kravetz * than that of shared mappings. vma_needs_reserves() 84667961f9dSMike Kravetz * has already taken this difference in semantics into 84767961f9dSMike Kravetz * account. Therefore, the meaning of chg is the same 84867961f9dSMike Kravetz * as in the shared case above. Code could easily be 84967961f9dSMike Kravetz * combined, but keeping it separate draws attention to 85067961f9dSMike Kravetz * subtle differences. 85167961f9dSMike Kravetz */ 85267961f9dSMike Kravetz if (chg) 85367961f9dSMike Kravetz return false; 85467961f9dSMike Kravetz else 855559ec2f8SNicholas Krause return true; 85667961f9dSMike Kravetz } 857a63884e9SJoonsoo Kim 858559ec2f8SNicholas Krause return false; 859a1e78772SMel Gorman } 860a1e78772SMel Gorman 861a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 8621da177e4SLinus Torvalds { 8631da177e4SLinus Torvalds int nid = page_to_nid(page); 8640edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 865a5516438SAndi Kleen h->free_huge_pages++; 866a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 8671da177e4SLinus Torvalds } 8681da177e4SLinus Torvalds 86994310cbcSAnshuman Khandual static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) 870bf50bab2SNaoya Horiguchi { 871bf50bab2SNaoya Horiguchi struct page *page; 872bf50bab2SNaoya Horiguchi 873c8721bbbSNaoya Horiguchi list_for_each_entry(page, &h->hugepage_freelists[nid], lru) 874243abd5bSNaoya Horiguchi if (!PageHWPoison(page)) 875c8721bbbSNaoya Horiguchi break; 876c8721bbbSNaoya Horiguchi /* 877c8721bbbSNaoya Horiguchi * if 'non-isolated free hugepage' not found on the list, 878c8721bbbSNaoya Horiguchi * the allocation fails. 879c8721bbbSNaoya Horiguchi */ 880c8721bbbSNaoya Horiguchi if (&h->hugepage_freelists[nid] == &page->lru) 881bf50bab2SNaoya Horiguchi return NULL; 8820edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 883a9869b83SNaoya Horiguchi set_page_refcounted(page); 884bf50bab2SNaoya Horiguchi h->free_huge_pages--; 885bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 886bf50bab2SNaoya Horiguchi return page; 887bf50bab2SNaoya Horiguchi } 888bf50bab2SNaoya Horiguchi 8893e59fcb0SMichal Hocko static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, 8903e59fcb0SMichal Hocko nodemask_t *nmask) 89194310cbcSAnshuman Khandual { 8923e59fcb0SMichal Hocko unsigned int cpuset_mems_cookie; 8933e59fcb0SMichal Hocko struct zonelist *zonelist; 8943e59fcb0SMichal Hocko struct zone *zone; 8953e59fcb0SMichal Hocko struct zoneref *z; 8963e59fcb0SMichal Hocko int node = -1; 8973e59fcb0SMichal Hocko 8983e59fcb0SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 8993e59fcb0SMichal Hocko 9003e59fcb0SMichal Hocko retry_cpuset: 9013e59fcb0SMichal Hocko cpuset_mems_cookie = read_mems_allowed_begin(); 9023e59fcb0SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 90394310cbcSAnshuman Khandual struct page *page; 90494310cbcSAnshuman Khandual 9053e59fcb0SMichal Hocko if (!cpuset_zone_allowed(zone, gfp_mask)) 9063e59fcb0SMichal Hocko continue; 9073e59fcb0SMichal Hocko /* 9083e59fcb0SMichal Hocko * no need to ask again on the same node. Pool is node rather than 9093e59fcb0SMichal Hocko * zone aware 9103e59fcb0SMichal Hocko */ 9113e59fcb0SMichal Hocko if (zone_to_nid(zone) == node) 9123e59fcb0SMichal Hocko continue; 9133e59fcb0SMichal Hocko node = zone_to_nid(zone); 91494310cbcSAnshuman Khandual 91594310cbcSAnshuman Khandual page = dequeue_huge_page_node_exact(h, node); 91694310cbcSAnshuman Khandual if (page) 91794310cbcSAnshuman Khandual return page; 91894310cbcSAnshuman Khandual } 9193e59fcb0SMichal Hocko if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 9203e59fcb0SMichal Hocko goto retry_cpuset; 9213e59fcb0SMichal Hocko 92294310cbcSAnshuman Khandual return NULL; 92394310cbcSAnshuman Khandual } 92494310cbcSAnshuman Khandual 92586cdb465SNaoya Horiguchi /* Movability of hugepages depends on migration support. */ 92686cdb465SNaoya Horiguchi static inline gfp_t htlb_alloc_mask(struct hstate *h) 92786cdb465SNaoya Horiguchi { 928d6cb41ccSMichal Hocko if (hugepage_migration_supported(h)) 92986cdb465SNaoya Horiguchi return GFP_HIGHUSER_MOVABLE; 93086cdb465SNaoya Horiguchi else 93186cdb465SNaoya Horiguchi return GFP_HIGHUSER; 93286cdb465SNaoya Horiguchi } 93386cdb465SNaoya Horiguchi 934a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 935a5516438SAndi Kleen struct vm_area_struct *vma, 936af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 937af0ed73eSJoonsoo Kim long chg) 9381da177e4SLinus Torvalds { 9393e59fcb0SMichal Hocko struct page *page; 940480eccf9SLee Schermerhorn struct mempolicy *mpol; 94104ec6264SVlastimil Babka gfp_t gfp_mask; 9423e59fcb0SMichal Hocko nodemask_t *nodemask; 94304ec6264SVlastimil Babka int nid; 9441da177e4SLinus Torvalds 945a1e78772SMel Gorman /* 946a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 947a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 948a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 949a1e78772SMel Gorman */ 950af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg) && 951a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 952c0ff7453SMiao Xie goto err; 953a1e78772SMel Gorman 95404f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 955a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 9566eab04a8SJustin P. Mattock goto err; 95704f2cbe3SMel Gorman 95804ec6264SVlastimil Babka gfp_mask = htlb_alloc_mask(h); 95904ec6264SVlastimil Babka nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 9603e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 9613e59fcb0SMichal Hocko if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { 96207443a85SJoonsoo Kim SetPagePrivate(page); 963a63884e9SJoonsoo Kim h->resv_huge_pages--; 964bf50bab2SNaoya Horiguchi } 965cc9a6c87SMel Gorman 966cc9a6c87SMel Gorman mpol_cond_put(mpol); 967cc9a6c87SMel Gorman return page; 968cc9a6c87SMel Gorman 969c0ff7453SMiao Xie err: 970cc9a6c87SMel Gorman return NULL; 9711da177e4SLinus Torvalds } 9721da177e4SLinus Torvalds 9731cac6f2cSLuiz Capitulino /* 9741cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 9751cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 9761cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 9771cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 9781cac6f2cSLuiz Capitulino * node for alloc or free. 9791cac6f2cSLuiz Capitulino */ 9801cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 9811cac6f2cSLuiz Capitulino { 9820edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 9831cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 9841cac6f2cSLuiz Capitulino 9851cac6f2cSLuiz Capitulino return nid; 9861cac6f2cSLuiz Capitulino } 9871cac6f2cSLuiz Capitulino 9881cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 9891cac6f2cSLuiz Capitulino { 9901cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 9911cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 9921cac6f2cSLuiz Capitulino return nid; 9931cac6f2cSLuiz Capitulino } 9941cac6f2cSLuiz Capitulino 9951cac6f2cSLuiz Capitulino /* 9961cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 9971cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 9981cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 9991cac6f2cSLuiz Capitulino * mask. 10001cac6f2cSLuiz Capitulino */ 10011cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 10021cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 10031cac6f2cSLuiz Capitulino { 10041cac6f2cSLuiz Capitulino int nid; 10051cac6f2cSLuiz Capitulino 10061cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 10071cac6f2cSLuiz Capitulino 10081cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 10091cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 10101cac6f2cSLuiz Capitulino 10111cac6f2cSLuiz Capitulino return nid; 10121cac6f2cSLuiz Capitulino } 10131cac6f2cSLuiz Capitulino 10141cac6f2cSLuiz Capitulino /* 10151cac6f2cSLuiz Capitulino * helper for free_pool_huge_page() - return the previously saved 10161cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 10171cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 10181cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 10191cac6f2cSLuiz Capitulino */ 10201cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 10211cac6f2cSLuiz Capitulino { 10221cac6f2cSLuiz Capitulino int nid; 10231cac6f2cSLuiz Capitulino 10241cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 10251cac6f2cSLuiz Capitulino 10261cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 10271cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 10281cac6f2cSLuiz Capitulino 10291cac6f2cSLuiz Capitulino return nid; 10301cac6f2cSLuiz Capitulino } 10311cac6f2cSLuiz Capitulino 10321cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 10331cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 10341cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 10351cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 10361cac6f2cSLuiz Capitulino nr_nodes--) 10371cac6f2cSLuiz Capitulino 10381cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 10391cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 10401cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 10411cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 10421cac6f2cSLuiz Capitulino nr_nodes--) 10431cac6f2cSLuiz Capitulino 1044e1073d1eSAneesh Kumar K.V #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 1045944d9fecSLuiz Capitulino static void destroy_compound_gigantic_page(struct page *page, 1046d00181b9SKirill A. Shutemov unsigned int order) 1047944d9fecSLuiz Capitulino { 1048944d9fecSLuiz Capitulino int i; 1049944d9fecSLuiz Capitulino int nr_pages = 1 << order; 1050944d9fecSLuiz Capitulino struct page *p = page + 1; 1051944d9fecSLuiz Capitulino 1052c8cc708aSGerald Schaefer atomic_set(compound_mapcount_ptr(page), 0); 1053944d9fecSLuiz Capitulino for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 10541d798ca3SKirill A. Shutemov clear_compound_head(p); 1055944d9fecSLuiz Capitulino set_page_refcounted(p); 1056944d9fecSLuiz Capitulino } 1057944d9fecSLuiz Capitulino 1058944d9fecSLuiz Capitulino set_compound_order(page, 0); 1059944d9fecSLuiz Capitulino __ClearPageHead(page); 1060944d9fecSLuiz Capitulino } 1061944d9fecSLuiz Capitulino 1062d00181b9SKirill A. Shutemov static void free_gigantic_page(struct page *page, unsigned int order) 1063944d9fecSLuiz Capitulino { 1064944d9fecSLuiz Capitulino free_contig_range(page_to_pfn(page), 1 << order); 1065944d9fecSLuiz Capitulino } 1066944d9fecSLuiz Capitulino 1067944d9fecSLuiz Capitulino static int __alloc_gigantic_page(unsigned long start_pfn, 106879b63f12SMichal Hocko unsigned long nr_pages, gfp_t gfp_mask) 1069944d9fecSLuiz Capitulino { 1070944d9fecSLuiz Capitulino unsigned long end_pfn = start_pfn + nr_pages; 1071ca96b625SLucas Stach return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE, 107279b63f12SMichal Hocko gfp_mask); 1073944d9fecSLuiz Capitulino } 1074944d9fecSLuiz Capitulino 1075f44b2ddaSJoonsoo Kim static bool pfn_range_valid_gigantic(struct zone *z, 1076f44b2ddaSJoonsoo Kim unsigned long start_pfn, unsigned long nr_pages) 1077944d9fecSLuiz Capitulino { 1078944d9fecSLuiz Capitulino unsigned long i, end_pfn = start_pfn + nr_pages; 1079944d9fecSLuiz Capitulino struct page *page; 1080944d9fecSLuiz Capitulino 1081944d9fecSLuiz Capitulino for (i = start_pfn; i < end_pfn; i++) { 1082944d9fecSLuiz Capitulino if (!pfn_valid(i)) 1083944d9fecSLuiz Capitulino return false; 1084944d9fecSLuiz Capitulino 1085944d9fecSLuiz Capitulino page = pfn_to_page(i); 1086944d9fecSLuiz Capitulino 1087f44b2ddaSJoonsoo Kim if (page_zone(page) != z) 1088f44b2ddaSJoonsoo Kim return false; 1089f44b2ddaSJoonsoo Kim 1090944d9fecSLuiz Capitulino if (PageReserved(page)) 1091944d9fecSLuiz Capitulino return false; 1092944d9fecSLuiz Capitulino 1093944d9fecSLuiz Capitulino if (page_count(page) > 0) 1094944d9fecSLuiz Capitulino return false; 1095944d9fecSLuiz Capitulino 1096944d9fecSLuiz Capitulino if (PageHuge(page)) 1097944d9fecSLuiz Capitulino return false; 1098944d9fecSLuiz Capitulino } 1099944d9fecSLuiz Capitulino 1100944d9fecSLuiz Capitulino return true; 1101944d9fecSLuiz Capitulino } 1102944d9fecSLuiz Capitulino 1103944d9fecSLuiz Capitulino static bool zone_spans_last_pfn(const struct zone *zone, 1104944d9fecSLuiz Capitulino unsigned long start_pfn, unsigned long nr_pages) 1105944d9fecSLuiz Capitulino { 1106944d9fecSLuiz Capitulino unsigned long last_pfn = start_pfn + nr_pages - 1; 1107944d9fecSLuiz Capitulino return zone_spans_pfn(zone, last_pfn); 1108944d9fecSLuiz Capitulino } 1109944d9fecSLuiz Capitulino 1110d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 1111d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) 1112944d9fecSLuiz Capitulino { 111379b63f12SMichal Hocko unsigned int order = huge_page_order(h); 1114944d9fecSLuiz Capitulino unsigned long nr_pages = 1 << order; 1115944d9fecSLuiz Capitulino unsigned long ret, pfn, flags; 111679b63f12SMichal Hocko struct zonelist *zonelist; 111779b63f12SMichal Hocko struct zone *zone; 111879b63f12SMichal Hocko struct zoneref *z; 1119944d9fecSLuiz Capitulino 112079b63f12SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 1121d9cc948fSMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nodemask) { 112279b63f12SMichal Hocko spin_lock_irqsave(&zone->lock, flags); 1123944d9fecSLuiz Capitulino 112479b63f12SMichal Hocko pfn = ALIGN(zone->zone_start_pfn, nr_pages); 112579b63f12SMichal Hocko while (zone_spans_last_pfn(zone, pfn, nr_pages)) { 112679b63f12SMichal Hocko if (pfn_range_valid_gigantic(zone, pfn, nr_pages)) { 1127944d9fecSLuiz Capitulino /* 1128944d9fecSLuiz Capitulino * We release the zone lock here because 1129944d9fecSLuiz Capitulino * alloc_contig_range() will also lock the zone 1130944d9fecSLuiz Capitulino * at some point. If there's an allocation 1131944d9fecSLuiz Capitulino * spinning on this lock, it may win the race 1132944d9fecSLuiz Capitulino * and cause alloc_contig_range() to fail... 1133944d9fecSLuiz Capitulino */ 113479b63f12SMichal Hocko spin_unlock_irqrestore(&zone->lock, flags); 113579b63f12SMichal Hocko ret = __alloc_gigantic_page(pfn, nr_pages, gfp_mask); 1136944d9fecSLuiz Capitulino if (!ret) 1137944d9fecSLuiz Capitulino return pfn_to_page(pfn); 113879b63f12SMichal Hocko spin_lock_irqsave(&zone->lock, flags); 1139944d9fecSLuiz Capitulino } 1140944d9fecSLuiz Capitulino pfn += nr_pages; 1141944d9fecSLuiz Capitulino } 1142944d9fecSLuiz Capitulino 114379b63f12SMichal Hocko spin_unlock_irqrestore(&zone->lock, flags); 1144944d9fecSLuiz Capitulino } 1145944d9fecSLuiz Capitulino 1146944d9fecSLuiz Capitulino return NULL; 1147944d9fecSLuiz Capitulino } 1148944d9fecSLuiz Capitulino 1149944d9fecSLuiz Capitulino static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); 1150d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order); 1151944d9fecSLuiz Capitulino 1152e1073d1eSAneesh Kumar K.V #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 1153944d9fecSLuiz Capitulino static inline bool gigantic_page_supported(void) { return false; } 1154d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 1155d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) { return NULL; } 1156d00181b9SKirill A. Shutemov static inline void free_gigantic_page(struct page *page, unsigned int order) { } 1157944d9fecSLuiz Capitulino static inline void destroy_compound_gigantic_page(struct page *page, 1158d00181b9SKirill A. Shutemov unsigned int order) { } 1159944d9fecSLuiz Capitulino #endif 1160944d9fecSLuiz Capitulino 1161a5516438SAndi Kleen static void update_and_free_page(struct hstate *h, struct page *page) 11626af2acb6SAdam Litke { 11636af2acb6SAdam Litke int i; 1164a5516438SAndi Kleen 1165944d9fecSLuiz Capitulino if (hstate_is_gigantic(h) && !gigantic_page_supported()) 1166944d9fecSLuiz Capitulino return; 116718229df5SAndy Whitcroft 1168a5516438SAndi Kleen h->nr_huge_pages--; 1169a5516438SAndi Kleen h->nr_huge_pages_node[page_to_nid(page)]--; 1170a5516438SAndi Kleen for (i = 0; i < pages_per_huge_page(h); i++) { 117132f84528SChris Forbes page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 117232f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1173a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1174a7407a27SLuiz Capitulino 1 << PG_writeback); 11756af2acb6SAdam Litke } 1176309381feSSasha Levin VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 1177f1e61557SKirill A. Shutemov set_compound_page_dtor(page, NULL_COMPOUND_DTOR); 11786af2acb6SAdam Litke set_page_refcounted(page); 1179944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) { 1180944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 1181944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 1182944d9fecSLuiz Capitulino } else { 1183a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 11846af2acb6SAdam Litke } 1185944d9fecSLuiz Capitulino } 11866af2acb6SAdam Litke 1187e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1188e5ff2159SAndi Kleen { 1189e5ff2159SAndi Kleen struct hstate *h; 1190e5ff2159SAndi Kleen 1191e5ff2159SAndi Kleen for_each_hstate(h) { 1192e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1193e5ff2159SAndi Kleen return h; 1194e5ff2159SAndi Kleen } 1195e5ff2159SAndi Kleen return NULL; 1196e5ff2159SAndi Kleen } 1197e5ff2159SAndi Kleen 1198bcc54222SNaoya Horiguchi /* 1199bcc54222SNaoya Horiguchi * Test to determine whether the hugepage is "active/in-use" (i.e. being linked 1200bcc54222SNaoya Horiguchi * to hstate->hugepage_activelist.) 1201bcc54222SNaoya Horiguchi * 1202bcc54222SNaoya Horiguchi * This function can be called for tail pages, but never returns true for them. 1203bcc54222SNaoya Horiguchi */ 1204bcc54222SNaoya Horiguchi bool page_huge_active(struct page *page) 1205bcc54222SNaoya Horiguchi { 1206bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHuge(page), page); 1207bcc54222SNaoya Horiguchi return PageHead(page) && PagePrivate(&page[1]); 1208bcc54222SNaoya Horiguchi } 1209bcc54222SNaoya Horiguchi 1210bcc54222SNaoya Horiguchi /* never called for tail page */ 1211bcc54222SNaoya Horiguchi static void set_page_huge_active(struct page *page) 1212bcc54222SNaoya Horiguchi { 1213bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 1214bcc54222SNaoya Horiguchi SetPagePrivate(&page[1]); 1215bcc54222SNaoya Horiguchi } 1216bcc54222SNaoya Horiguchi 1217bcc54222SNaoya Horiguchi static void clear_page_huge_active(struct page *page) 1218bcc54222SNaoya Horiguchi { 1219bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 1220bcc54222SNaoya Horiguchi ClearPagePrivate(&page[1]); 1221bcc54222SNaoya Horiguchi } 1222bcc54222SNaoya Horiguchi 1223ab5ac90aSMichal Hocko /* 1224ab5ac90aSMichal Hocko * Internal hugetlb specific page flag. Do not use outside of the hugetlb 1225ab5ac90aSMichal Hocko * code 1226ab5ac90aSMichal Hocko */ 1227ab5ac90aSMichal Hocko static inline bool PageHugeTemporary(struct page *page) 1228ab5ac90aSMichal Hocko { 1229ab5ac90aSMichal Hocko if (!PageHuge(page)) 1230ab5ac90aSMichal Hocko return false; 1231ab5ac90aSMichal Hocko 1232ab5ac90aSMichal Hocko return (unsigned long)page[2].mapping == -1U; 1233ab5ac90aSMichal Hocko } 1234ab5ac90aSMichal Hocko 1235ab5ac90aSMichal Hocko static inline void SetPageHugeTemporary(struct page *page) 1236ab5ac90aSMichal Hocko { 1237ab5ac90aSMichal Hocko page[2].mapping = (void *)-1U; 1238ab5ac90aSMichal Hocko } 1239ab5ac90aSMichal Hocko 1240ab5ac90aSMichal Hocko static inline void ClearPageHugeTemporary(struct page *page) 1241ab5ac90aSMichal Hocko { 1242ab5ac90aSMichal Hocko page[2].mapping = NULL; 1243ab5ac90aSMichal Hocko } 1244ab5ac90aSMichal Hocko 12458f1d26d0SAtsushi Kumagai void free_huge_page(struct page *page) 124627a85ef1SDavid Gibson { 1247a5516438SAndi Kleen /* 1248a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1249a5516438SAndi Kleen * compound page destructor. 1250a5516438SAndi Kleen */ 1251e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 12527893d1d5SAdam Litke int nid = page_to_nid(page); 125390481622SDavid Gibson struct hugepage_subpool *spool = 125490481622SDavid Gibson (struct hugepage_subpool *)page_private(page); 125507443a85SJoonsoo Kim bool restore_reserve; 125627a85ef1SDavid Gibson 1257e5df70abSAndy Whitcroft set_page_private(page, 0); 125823be7468SMel Gorman page->mapping = NULL; 1259b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1260b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_mapcount(page), page); 126107443a85SJoonsoo Kim restore_reserve = PagePrivate(page); 126216c794b4SJoonsoo Kim ClearPagePrivate(page); 126327a85ef1SDavid Gibson 12641c5ecae3SMike Kravetz /* 12651c5ecae3SMike Kravetz * A return code of zero implies that the subpool will be under its 12661c5ecae3SMike Kravetz * minimum size if the reservation is not restored after page is free. 12671c5ecae3SMike Kravetz * Therefore, force restore_reserve operation. 12681c5ecae3SMike Kravetz */ 12691c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 12701c5ecae3SMike Kravetz restore_reserve = true; 12711c5ecae3SMike Kravetz 127227a85ef1SDavid Gibson spin_lock(&hugetlb_lock); 1273bcc54222SNaoya Horiguchi clear_page_huge_active(page); 12746d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 12756d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 127607443a85SJoonsoo Kim if (restore_reserve) 127707443a85SJoonsoo Kim h->resv_huge_pages++; 127807443a85SJoonsoo Kim 1279ab5ac90aSMichal Hocko if (PageHugeTemporary(page)) { 1280ab5ac90aSMichal Hocko list_del(&page->lru); 1281ab5ac90aSMichal Hocko ClearPageHugeTemporary(page); 1282ab5ac90aSMichal Hocko update_and_free_page(h, page); 1283ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 12840edaecfaSAneesh Kumar K.V /* remove the page from active list */ 12850edaecfaSAneesh Kumar K.V list_del(&page->lru); 1286a5516438SAndi Kleen update_and_free_page(h, page); 1287a5516438SAndi Kleen h->surplus_huge_pages--; 1288a5516438SAndi Kleen h->surplus_huge_pages_node[nid]--; 12897893d1d5SAdam Litke } else { 12905d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1291a5516438SAndi Kleen enqueue_huge_page(h, page); 12927893d1d5SAdam Litke } 129327a85ef1SDavid Gibson spin_unlock(&hugetlb_lock); 129427a85ef1SDavid Gibson } 129527a85ef1SDavid Gibson 1296a5516438SAndi Kleen static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1297b7ba30c6SAndi Kleen { 12980edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1299f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1300b7ba30c6SAndi Kleen spin_lock(&hugetlb_lock); 13019dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 1302a5516438SAndi Kleen h->nr_huge_pages++; 1303a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 1304b7ba30c6SAndi Kleen spin_unlock(&hugetlb_lock); 1305b7ba30c6SAndi Kleen } 1306b7ba30c6SAndi Kleen 1307d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order) 130820a0307cSWu Fengguang { 130920a0307cSWu Fengguang int i; 131020a0307cSWu Fengguang int nr_pages = 1 << order; 131120a0307cSWu Fengguang struct page *p = page + 1; 131220a0307cSWu Fengguang 131320a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 131420a0307cSWu Fengguang set_compound_order(page, order); 1315ef5a22beSAndrea Arcangeli __ClearPageReserved(page); 1316de09d31dSKirill A. Shutemov __SetPageHead(page); 131720a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1318ef5a22beSAndrea Arcangeli /* 1319ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1320ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1321ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 1322ef5a22beSAndrea Arcangeli * too. Otherwse drivers using get_user_pages() to access tail 1323ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1324ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1325ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1326ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1327ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1328ef5a22beSAndrea Arcangeli * after get_user_pages(). 1329ef5a22beSAndrea Arcangeli */ 1330ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 133158a84aa9SYouquan Song set_page_count(p, 0); 13321d798ca3SKirill A. Shutemov set_compound_head(p, page); 133320a0307cSWu Fengguang } 1334b4330afbSMike Kravetz atomic_set(compound_mapcount_ptr(page), -1); 133520a0307cSWu Fengguang } 133620a0307cSWu Fengguang 13377795912cSAndrew Morton /* 13387795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 13397795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 13407795912cSAndrew Morton * details. 13417795912cSAndrew Morton */ 134220a0307cSWu Fengguang int PageHuge(struct page *page) 134320a0307cSWu Fengguang { 134420a0307cSWu Fengguang if (!PageCompound(page)) 134520a0307cSWu Fengguang return 0; 134620a0307cSWu Fengguang 134720a0307cSWu Fengguang page = compound_head(page); 1348f1e61557SKirill A. Shutemov return page[1].compound_dtor == HUGETLB_PAGE_DTOR; 134920a0307cSWu Fengguang } 135043131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 135143131e14SNaoya Horiguchi 135227c73ae7SAndrea Arcangeli /* 135327c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 135427c73ae7SAndrea Arcangeli * normal or transparent huge pages. 135527c73ae7SAndrea Arcangeli */ 135627c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 135727c73ae7SAndrea Arcangeli { 135827c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 135927c73ae7SAndrea Arcangeli return 0; 136027c73ae7SAndrea Arcangeli 1361758f66a2SAndrew Morton return get_compound_page_dtor(page_head) == free_huge_page; 136227c73ae7SAndrea Arcangeli } 136327c73ae7SAndrea Arcangeli 136413d60f4bSZhang Yi pgoff_t __basepage_index(struct page *page) 136513d60f4bSZhang Yi { 136613d60f4bSZhang Yi struct page *page_head = compound_head(page); 136713d60f4bSZhang Yi pgoff_t index = page_index(page_head); 136813d60f4bSZhang Yi unsigned long compound_idx; 136913d60f4bSZhang Yi 137013d60f4bSZhang Yi if (!PageHuge(page_head)) 137113d60f4bSZhang Yi return page_index(page); 137213d60f4bSZhang Yi 137313d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 137413d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 137513d60f4bSZhang Yi else 137613d60f4bSZhang Yi compound_idx = page - page_head; 137713d60f4bSZhang Yi 137813d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 137913d60f4bSZhang Yi } 138013d60f4bSZhang Yi 1381af0fb9dfSMichal Hocko static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h, 1382af0fb9dfSMichal Hocko gfp_t gfp_mask, int nid, nodemask_t *nmask) 13831da177e4SLinus Torvalds { 1384af0fb9dfSMichal Hocko int order = huge_page_order(h); 13851da177e4SLinus Torvalds struct page *page; 1386f96efd58SJoe Jin 1387af0fb9dfSMichal Hocko gfp_mask |= __GFP_COMP|__GFP_RETRY_MAYFAIL|__GFP_NOWARN; 1388af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 1389af0fb9dfSMichal Hocko nid = numa_mem_id(); 1390af0fb9dfSMichal Hocko page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask); 1391af0fb9dfSMichal Hocko if (page) 1392af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC); 1393af0fb9dfSMichal Hocko else 1394af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 139563b4613cSNishanth Aravamudan 139663b4613cSNishanth Aravamudan return page; 139763b4613cSNishanth Aravamudan } 139863b4613cSNishanth Aravamudan 1399af0fb9dfSMichal Hocko /* 1400af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 1401af0fb9dfSMichal Hocko * manner. 1402af0fb9dfSMichal Hocko */ 1403b2261026SJoonsoo Kim static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 1404b2261026SJoonsoo Kim { 1405b2261026SJoonsoo Kim struct page *page; 1406b2261026SJoonsoo Kim int nr_nodes, node; 1407af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 1408b2261026SJoonsoo Kim 1409b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1410d9cc948fSMichal Hocko if (hstate_is_gigantic(h)) 1411d9cc948fSMichal Hocko page = alloc_gigantic_page(h, gfp_mask, 1412d9cc948fSMichal Hocko node, nodes_allowed); 1413d9cc948fSMichal Hocko else 1414af0fb9dfSMichal Hocko page = __hugetlb_alloc_buddy_huge_page(h, gfp_mask, 1415af0fb9dfSMichal Hocko node, nodes_allowed); 1416af0fb9dfSMichal Hocko if (page) 1417b2261026SJoonsoo Kim break; 1418af0fb9dfSMichal Hocko 1419b2261026SJoonsoo Kim } 1420b2261026SJoonsoo Kim 1421af0fb9dfSMichal Hocko if (!page) 1422af0fb9dfSMichal Hocko return 0; 1423b2261026SJoonsoo Kim 1424d9cc948fSMichal Hocko if (hstate_is_gigantic(h)) 1425d9cc948fSMichal Hocko prep_compound_gigantic_page(page, huge_page_order(h)); 1426af0fb9dfSMichal Hocko prep_new_huge_page(h, page, page_to_nid(page)); 1427af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 1428af0fb9dfSMichal Hocko 1429af0fb9dfSMichal Hocko return 1; 1430b2261026SJoonsoo Kim } 1431b2261026SJoonsoo Kim 1432e8c5c824SLee Schermerhorn /* 1433e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 1434e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 1435e8c5c824SLee Schermerhorn * balanced over allowed nodes. 1436e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 1437e8c5c824SLee Schermerhorn */ 14386ae11b27SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 14396ae11b27SLee Schermerhorn bool acct_surplus) 1440e8c5c824SLee Schermerhorn { 1441b2261026SJoonsoo Kim int nr_nodes, node; 1442e8c5c824SLee Schermerhorn int ret = 0; 1443e8c5c824SLee Schermerhorn 1444b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1445685f3457SLee Schermerhorn /* 1446685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 1447685f3457SLee Schermerhorn * nodes with surplus pages. 1448685f3457SLee Schermerhorn */ 1449b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 1450b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 1451e8c5c824SLee Schermerhorn struct page *page = 1452b2261026SJoonsoo Kim list_entry(h->hugepage_freelists[node].next, 1453e8c5c824SLee Schermerhorn struct page, lru); 1454e8c5c824SLee Schermerhorn list_del(&page->lru); 1455e8c5c824SLee Schermerhorn h->free_huge_pages--; 1456b2261026SJoonsoo Kim h->free_huge_pages_node[node]--; 1457685f3457SLee Schermerhorn if (acct_surplus) { 1458685f3457SLee Schermerhorn h->surplus_huge_pages--; 1459b2261026SJoonsoo Kim h->surplus_huge_pages_node[node]--; 1460685f3457SLee Schermerhorn } 1461e8c5c824SLee Schermerhorn update_and_free_page(h, page); 1462e8c5c824SLee Schermerhorn ret = 1; 14639a76db09SLee Schermerhorn break; 1464e8c5c824SLee Schermerhorn } 1465b2261026SJoonsoo Kim } 1466e8c5c824SLee Schermerhorn 1467e8c5c824SLee Schermerhorn return ret; 1468e8c5c824SLee Schermerhorn } 1469e8c5c824SLee Schermerhorn 1470c8721bbbSNaoya Horiguchi /* 1471c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 1472082d5b6bSGerald Schaefer * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the 1473082d5b6bSGerald Schaefer * number of free hugepages would be reduced below the number of reserved 1474082d5b6bSGerald Schaefer * hugepages. 1475c8721bbbSNaoya Horiguchi */ 1476c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 1477c8721bbbSNaoya Horiguchi { 1478082d5b6bSGerald Schaefer int rc = 0; 1479082d5b6bSGerald Schaefer 1480c8721bbbSNaoya Horiguchi spin_lock(&hugetlb_lock); 1481c8721bbbSNaoya Horiguchi if (PageHuge(page) && !page_count(page)) { 14822247bb33SGerald Schaefer struct page *head = compound_head(page); 14832247bb33SGerald Schaefer struct hstate *h = page_hstate(head); 14842247bb33SGerald Schaefer int nid = page_to_nid(head); 1485082d5b6bSGerald Schaefer if (h->free_huge_pages - h->resv_huge_pages == 0) { 1486082d5b6bSGerald Schaefer rc = -EBUSY; 1487082d5b6bSGerald Schaefer goto out; 1488082d5b6bSGerald Schaefer } 1489c3114a84SAnshuman Khandual /* 1490c3114a84SAnshuman Khandual * Move PageHWPoison flag from head page to the raw error page, 1491c3114a84SAnshuman Khandual * which makes any subpages rather than the error page reusable. 1492c3114a84SAnshuman Khandual */ 1493c3114a84SAnshuman Khandual if (PageHWPoison(head) && page != head) { 1494c3114a84SAnshuman Khandual SetPageHWPoison(page); 1495c3114a84SAnshuman Khandual ClearPageHWPoison(head); 1496c3114a84SAnshuman Khandual } 14972247bb33SGerald Schaefer list_del(&head->lru); 1498c8721bbbSNaoya Horiguchi h->free_huge_pages--; 1499c8721bbbSNaoya Horiguchi h->free_huge_pages_node[nid]--; 1500c1470b33Szhong jiang h->max_huge_pages--; 15012247bb33SGerald Schaefer update_and_free_page(h, head); 1502c8721bbbSNaoya Horiguchi } 1503082d5b6bSGerald Schaefer out: 1504c8721bbbSNaoya Horiguchi spin_unlock(&hugetlb_lock); 1505082d5b6bSGerald Schaefer return rc; 1506c8721bbbSNaoya Horiguchi } 1507c8721bbbSNaoya Horiguchi 1508c8721bbbSNaoya Horiguchi /* 1509c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 1510c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 15112247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 15122247bb33SGerald Schaefer * part of it lies within the given range. 1513082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 1514082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 1515c8721bbbSNaoya Horiguchi */ 1516082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 1517c8721bbbSNaoya Horiguchi { 1518c8721bbbSNaoya Horiguchi unsigned long pfn; 1519eb03aa00SGerald Schaefer struct page *page; 1520082d5b6bSGerald Schaefer int rc = 0; 1521c8721bbbSNaoya Horiguchi 1522d0177639SLi Zhong if (!hugepages_supported()) 1523082d5b6bSGerald Schaefer return rc; 1524d0177639SLi Zhong 1525eb03aa00SGerald Schaefer for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { 1526eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 1527eb03aa00SGerald Schaefer if (PageHuge(page) && !page_count(page)) { 1528eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 1529eb03aa00SGerald Schaefer if (rc) 1530082d5b6bSGerald Schaefer break; 1531eb03aa00SGerald Schaefer } 1532eb03aa00SGerald Schaefer } 1533082d5b6bSGerald Schaefer 1534082d5b6bSGerald Schaefer return rc; 1535c8721bbbSNaoya Horiguchi } 1536c8721bbbSNaoya Horiguchi 1537ab5ac90aSMichal Hocko /* 1538ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 1539ab5ac90aSMichal Hocko */ 1540ab5ac90aSMichal Hocko static struct page *__alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, 1541aaf14e40SMichal Hocko int nid, nodemask_t *nmask) 15427893d1d5SAdam Litke { 15437893d1d5SAdam Litke struct page *page; 1544bf50bab2SNaoya Horiguchi unsigned int r_nid; 15457893d1d5SAdam Litke 1546bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1547aa888a74SAndi Kleen return NULL; 1548aa888a74SAndi Kleen 1549d1c3fb1fSNishanth Aravamudan /* 1550d1c3fb1fSNishanth Aravamudan * Assume we will successfully allocate the surplus page to 1551d1c3fb1fSNishanth Aravamudan * prevent racing processes from causing the surplus to exceed 1552d1c3fb1fSNishanth Aravamudan * overcommit 1553d1c3fb1fSNishanth Aravamudan * 1554d1c3fb1fSNishanth Aravamudan * This however introduces a different race, where a process B 1555d1c3fb1fSNishanth Aravamudan * tries to grow the static hugepage pool while alloc_pages() is 1556d1c3fb1fSNishanth Aravamudan * called by process A. B will only examine the per-node 1557d1c3fb1fSNishanth Aravamudan * counters in determining if surplus huge pages can be 1558d1c3fb1fSNishanth Aravamudan * converted to normal huge pages in adjust_pool_surplus(). A 1559d1c3fb1fSNishanth Aravamudan * won't be able to increment the per-node counter, until the 1560d1c3fb1fSNishanth Aravamudan * lock is dropped by B, but B doesn't drop hugetlb_lock until 1561d1c3fb1fSNishanth Aravamudan * no more huge pages can be converted from surplus to normal 1562d1c3fb1fSNishanth Aravamudan * state (and doesn't try to convert again). Thus, we have a 1563d1c3fb1fSNishanth Aravamudan * case where a surplus huge page exists, the pool is grown, and 1564d1c3fb1fSNishanth Aravamudan * the surplus huge page still exists after, even though it 1565d1c3fb1fSNishanth Aravamudan * should just have been converted to a normal huge page. This 1566d1c3fb1fSNishanth Aravamudan * does not leak memory, though, as the hugepage will be freed 1567d1c3fb1fSNishanth Aravamudan * once it is out of use. It also does not allow the counters to 1568d1c3fb1fSNishanth Aravamudan * go out of whack in adjust_pool_surplus() as we don't modify 1569d1c3fb1fSNishanth Aravamudan * the node values until we've gotten the hugepage and only the 1570d1c3fb1fSNishanth Aravamudan * per-node value is checked there. 1571d1c3fb1fSNishanth Aravamudan */ 1572d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 1573a5516438SAndi Kleen if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 1574d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1575d1c3fb1fSNishanth Aravamudan return NULL; 1576d1c3fb1fSNishanth Aravamudan } else { 1577a5516438SAndi Kleen h->nr_huge_pages++; 1578a5516438SAndi Kleen h->surplus_huge_pages++; 1579d1c3fb1fSNishanth Aravamudan } 1580d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1581d1c3fb1fSNishanth Aravamudan 1582aaf14e40SMichal Hocko page = __hugetlb_alloc_buddy_huge_page(h, gfp_mask, nid, nmask); 1583d1c3fb1fSNishanth Aravamudan 15847893d1d5SAdam Litke spin_lock(&hugetlb_lock); 1585d1c3fb1fSNishanth Aravamudan if (page) { 15860edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1587bf50bab2SNaoya Horiguchi r_nid = page_to_nid(page); 1588f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 15899dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 1590d1c3fb1fSNishanth Aravamudan /* 1591d1c3fb1fSNishanth Aravamudan * We incremented the global counters already 1592d1c3fb1fSNishanth Aravamudan */ 1593bf50bab2SNaoya Horiguchi h->nr_huge_pages_node[r_nid]++; 1594bf50bab2SNaoya Horiguchi h->surplus_huge_pages_node[r_nid]++; 1595d1c3fb1fSNishanth Aravamudan } else { 1596a5516438SAndi Kleen h->nr_huge_pages--; 1597a5516438SAndi Kleen h->surplus_huge_pages--; 15987893d1d5SAdam Litke } 1599d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 16007893d1d5SAdam Litke 16017893d1d5SAdam Litke return page; 16027893d1d5SAdam Litke } 16037893d1d5SAdam Litke 1604ab5ac90aSMichal Hocko static struct page *__alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, 1605ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 1606ab5ac90aSMichal Hocko { 1607ab5ac90aSMichal Hocko struct page *page; 1608ab5ac90aSMichal Hocko 1609ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 1610ab5ac90aSMichal Hocko return NULL; 1611ab5ac90aSMichal Hocko 1612ab5ac90aSMichal Hocko page = __hugetlb_alloc_buddy_huge_page(h, gfp_mask, nid, nmask); 1613ab5ac90aSMichal Hocko if (!page) 1614ab5ac90aSMichal Hocko return NULL; 1615ab5ac90aSMichal Hocko 1616ab5ac90aSMichal Hocko /* 1617ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 1618ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 1619ab5ac90aSMichal Hocko */ 1620ab5ac90aSMichal Hocko prep_new_huge_page(h, page, page_to_nid(page)); 1621ab5ac90aSMichal Hocko SetPageHugeTemporary(page); 1622ab5ac90aSMichal Hocko 1623ab5ac90aSMichal Hocko return page; 1624ab5ac90aSMichal Hocko } 1625ab5ac90aSMichal Hocko 1626e4e574b7SAdam Litke /* 1627099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 1628099730d6SDave Hansen */ 1629e0ec90eeSDave Hansen static 1630099730d6SDave Hansen struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h, 1631099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 1632099730d6SDave Hansen { 1633aaf14e40SMichal Hocko struct page *page; 1634aaf14e40SMichal Hocko struct mempolicy *mpol; 1635aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 1636aaf14e40SMichal Hocko int nid; 1637aaf14e40SMichal Hocko nodemask_t *nodemask; 1638aaf14e40SMichal Hocko 1639aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 1640ab5ac90aSMichal Hocko page = __alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); 1641aaf14e40SMichal Hocko mpol_cond_put(mpol); 1642aaf14e40SMichal Hocko 1643aaf14e40SMichal Hocko return page; 1644099730d6SDave Hansen } 1645099730d6SDave Hansen 1646ab5ac90aSMichal Hocko /* page migration callback function */ 1647bf50bab2SNaoya Horiguchi struct page *alloc_huge_page_node(struct hstate *h, int nid) 1648bf50bab2SNaoya Horiguchi { 1649aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 16504ef91848SJoonsoo Kim struct page *page = NULL; 1651bf50bab2SNaoya Horiguchi 1652aaf14e40SMichal Hocko if (nid != NUMA_NO_NODE) 1653aaf14e40SMichal Hocko gfp_mask |= __GFP_THISNODE; 1654aaf14e40SMichal Hocko 1655bf50bab2SNaoya Horiguchi spin_lock(&hugetlb_lock); 16564ef91848SJoonsoo Kim if (h->free_huge_pages - h->resv_huge_pages > 0) 16573e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL); 1658bf50bab2SNaoya Horiguchi spin_unlock(&hugetlb_lock); 1659bf50bab2SNaoya Horiguchi 166094ae8ba7SAneesh Kumar K.V if (!page) 1661ab5ac90aSMichal Hocko page = __alloc_migrate_huge_page(h, gfp_mask, nid, NULL); 1662bf50bab2SNaoya Horiguchi 1663bf50bab2SNaoya Horiguchi return page; 1664bf50bab2SNaoya Horiguchi } 1665bf50bab2SNaoya Horiguchi 1666ab5ac90aSMichal Hocko /* page migration callback function */ 16673e59fcb0SMichal Hocko struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, 16683e59fcb0SMichal Hocko nodemask_t *nmask) 16694db9b2efSMichal Hocko { 1670aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 16714db9b2efSMichal Hocko 16724db9b2efSMichal Hocko spin_lock(&hugetlb_lock); 16734db9b2efSMichal Hocko if (h->free_huge_pages - h->resv_huge_pages > 0) { 16743e59fcb0SMichal Hocko struct page *page; 16753e59fcb0SMichal Hocko 16763e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); 16773e59fcb0SMichal Hocko if (page) { 16783e59fcb0SMichal Hocko spin_unlock(&hugetlb_lock); 16793e59fcb0SMichal Hocko return page; 16804db9b2efSMichal Hocko } 16814db9b2efSMichal Hocko } 16824db9b2efSMichal Hocko spin_unlock(&hugetlb_lock); 16834db9b2efSMichal Hocko 1684ab5ac90aSMichal Hocko return __alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); 16854db9b2efSMichal Hocko } 16864db9b2efSMichal Hocko 1687bf50bab2SNaoya Horiguchi /* 168825985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 1689e4e574b7SAdam Litke * of size 'delta'. 1690e4e574b7SAdam Litke */ 1691a5516438SAndi Kleen static int gather_surplus_pages(struct hstate *h, int delta) 1692e4e574b7SAdam Litke { 1693e4e574b7SAdam Litke struct list_head surplus_list; 1694e4e574b7SAdam Litke struct page *page, *tmp; 1695e4e574b7SAdam Litke int ret, i; 1696e4e574b7SAdam Litke int needed, allocated; 169728073b02SHillf Danton bool alloc_ok = true; 1698e4e574b7SAdam Litke 1699a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 1700ac09b3a1SAdam Litke if (needed <= 0) { 1701a5516438SAndi Kleen h->resv_huge_pages += delta; 1702e4e574b7SAdam Litke return 0; 1703ac09b3a1SAdam Litke } 1704e4e574b7SAdam Litke 1705e4e574b7SAdam Litke allocated = 0; 1706e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 1707e4e574b7SAdam Litke 1708e4e574b7SAdam Litke ret = -ENOMEM; 1709e4e574b7SAdam Litke retry: 1710e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 1711e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 1712ab5ac90aSMichal Hocko page = __alloc_surplus_huge_page(h, htlb_alloc_mask(h), 1713aaf14e40SMichal Hocko NUMA_NO_NODE, NULL); 171428073b02SHillf Danton if (!page) { 171528073b02SHillf Danton alloc_ok = false; 171628073b02SHillf Danton break; 171728073b02SHillf Danton } 1718e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 171969ed779aSDavid Rientjes cond_resched(); 1720e4e574b7SAdam Litke } 172128073b02SHillf Danton allocated += i; 1722e4e574b7SAdam Litke 1723e4e574b7SAdam Litke /* 1724e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 1725e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 1726e4e574b7SAdam Litke */ 1727e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 1728a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 1729a5516438SAndi Kleen (h->free_huge_pages + allocated); 173028073b02SHillf Danton if (needed > 0) { 173128073b02SHillf Danton if (alloc_ok) 1732e4e574b7SAdam Litke goto retry; 173328073b02SHillf Danton /* 173428073b02SHillf Danton * We were not able to allocate enough pages to 173528073b02SHillf Danton * satisfy the entire reservation so we free what 173628073b02SHillf Danton * we've allocated so far. 173728073b02SHillf Danton */ 173828073b02SHillf Danton goto free; 173928073b02SHillf Danton } 1740e4e574b7SAdam Litke /* 1741e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 174225985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 1743e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 1744ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 1745ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 1746ac09b3a1SAdam Litke * before they are reserved. 1747e4e574b7SAdam Litke */ 1748e4e574b7SAdam Litke needed += allocated; 1749a5516438SAndi Kleen h->resv_huge_pages += delta; 1750e4e574b7SAdam Litke ret = 0; 1751a9869b83SNaoya Horiguchi 175219fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 175319fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 175419fc3f0aSAdam Litke if ((--needed) < 0) 175519fc3f0aSAdam Litke break; 1756a9869b83SNaoya Horiguchi /* 1757a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 1758a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 1759a9869b83SNaoya Horiguchi */ 1760a9869b83SNaoya Horiguchi put_page_testzero(page); 1761309381feSSasha Levin VM_BUG_ON_PAGE(page_count(page), page); 1762a5516438SAndi Kleen enqueue_huge_page(h, page); 176319fc3f0aSAdam Litke } 176428073b02SHillf Danton free: 1765b0365c8dSHillf Danton spin_unlock(&hugetlb_lock); 176619fc3f0aSAdam Litke 176719fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 1768c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 1769a9869b83SNaoya Horiguchi put_page(page); 177019fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 1771e4e574b7SAdam Litke 1772e4e574b7SAdam Litke return ret; 1773e4e574b7SAdam Litke } 1774e4e574b7SAdam Litke 1775e4e574b7SAdam Litke /* 1776e5bbc8a6SMike Kravetz * This routine has two main purposes: 1777e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 1778e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 1779e5bbc8a6SMike Kravetz * to the associated reservation map. 1780e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 1781e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 1782e5bbc8a6SMike Kravetz * 1783e5bbc8a6SMike Kravetz * Called with hugetlb_lock held. However, the lock could be dropped (and 1784e5bbc8a6SMike Kravetz * reacquired) during calls to cond_resched_lock. Whenever dropping the lock, 1785e5bbc8a6SMike Kravetz * we must make sure nobody else can claim pages we are in the process of 1786e5bbc8a6SMike Kravetz * freeing. Do this by ensuring resv_huge_page always is greater than the 1787e5bbc8a6SMike Kravetz * number of huge pages we plan to free when dropping the lock. 1788e4e574b7SAdam Litke */ 1789a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 1790a5516438SAndi Kleen unsigned long unused_resv_pages) 1791e4e574b7SAdam Litke { 1792e4e574b7SAdam Litke unsigned long nr_pages; 1793e4e574b7SAdam Litke 1794aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 1795bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1796e5bbc8a6SMike Kravetz goto out; 1797aa888a74SAndi Kleen 1798e5bbc8a6SMike Kravetz /* 1799e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 1800e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 1801e5bbc8a6SMike Kravetz */ 1802a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 1803e4e574b7SAdam Litke 1804685f3457SLee Schermerhorn /* 1805685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 18069b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 18079b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 18089b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 18099b5e5d0fSLee Schermerhorn * free_pool_huge_page() will balance the the freed pages across the 18109b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 1811e5bbc8a6SMike Kravetz * 1812e5bbc8a6SMike Kravetz * Note that we decrement resv_huge_pages as we free the pages. If 1813e5bbc8a6SMike Kravetz * we drop the lock, resv_huge_pages will still be sufficiently large 1814e5bbc8a6SMike Kravetz * to cover subsequent pages we may free. 1815685f3457SLee Schermerhorn */ 1816685f3457SLee Schermerhorn while (nr_pages--) { 1817e5bbc8a6SMike Kravetz h->resv_huge_pages--; 1818e5bbc8a6SMike Kravetz unused_resv_pages--; 18198cebfcd0SLai Jiangshan if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) 1820e5bbc8a6SMike Kravetz goto out; 18217848a4bfSMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 1822e4e574b7SAdam Litke } 1823e5bbc8a6SMike Kravetz 1824e5bbc8a6SMike Kravetz out: 1825e5bbc8a6SMike Kravetz /* Fully uncommit the reservation */ 1826e5bbc8a6SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 1827e4e574b7SAdam Litke } 1828e4e574b7SAdam Litke 18295e911373SMike Kravetz 1830c37f9fb1SAndy Whitcroft /* 1831feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 18325e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 1833cf3ad20bSMike Kravetz * 1834cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 1835cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 1836cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 1837cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 1838cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 1839feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 1840feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 1841feba16e2SMike Kravetz * is called in such cases. 1842cf3ad20bSMike Kravetz * 1843cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 1844cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 1845cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 1846cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 1847cf3ad20bSMike Kravetz * take appropriate action. 184896b96a96SMike Kravetz * 184996b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 185096b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 185196b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 185296b96a96SMike Kravetz * reservation exists. 1853c37f9fb1SAndy Whitcroft */ 18545e911373SMike Kravetz enum vma_resv_mode { 18555e911373SMike Kravetz VMA_NEEDS_RESV, 18565e911373SMike Kravetz VMA_COMMIT_RESV, 1857feba16e2SMike Kravetz VMA_END_RESV, 185896b96a96SMike Kravetz VMA_ADD_RESV, 18595e911373SMike Kravetz }; 1860cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 1861cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 18625e911373SMike Kravetz enum vma_resv_mode mode) 1863c37f9fb1SAndy Whitcroft { 18644e35f483SJoonsoo Kim struct resv_map *resv; 18654e35f483SJoonsoo Kim pgoff_t idx; 1866cf3ad20bSMike Kravetz long ret; 1867c37f9fb1SAndy Whitcroft 18684e35f483SJoonsoo Kim resv = vma_resv_map(vma); 18694e35f483SJoonsoo Kim if (!resv) 1870c37f9fb1SAndy Whitcroft return 1; 1871c37f9fb1SAndy Whitcroft 18724e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 18735e911373SMike Kravetz switch (mode) { 18745e911373SMike Kravetz case VMA_NEEDS_RESV: 1875cf3ad20bSMike Kravetz ret = region_chg(resv, idx, idx + 1); 18765e911373SMike Kravetz break; 18775e911373SMike Kravetz case VMA_COMMIT_RESV: 18785e911373SMike Kravetz ret = region_add(resv, idx, idx + 1); 18795e911373SMike Kravetz break; 1880feba16e2SMike Kravetz case VMA_END_RESV: 18815e911373SMike Kravetz region_abort(resv, idx, idx + 1); 18825e911373SMike Kravetz ret = 0; 18835e911373SMike Kravetz break; 188496b96a96SMike Kravetz case VMA_ADD_RESV: 188596b96a96SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) 188696b96a96SMike Kravetz ret = region_add(resv, idx, idx + 1); 188796b96a96SMike Kravetz else { 188896b96a96SMike Kravetz region_abort(resv, idx, idx + 1); 188996b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 189096b96a96SMike Kravetz } 189196b96a96SMike Kravetz break; 18925e911373SMike Kravetz default: 18935e911373SMike Kravetz BUG(); 18945e911373SMike Kravetz } 189584afd99bSAndy Whitcroft 18964e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) 1897cf3ad20bSMike Kravetz return ret; 189867961f9dSMike Kravetz else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { 189967961f9dSMike Kravetz /* 190067961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 190167961f9dSMike Kravetz * However, a file associated with mapping could have been 190267961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 190367961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 190467961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 190567961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 190667961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 190767961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 190867961f9dSMike Kravetz * reservation has already been consumed. As a result, the 190967961f9dSMike Kravetz * return value of this routine is the opposite of the 191067961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 191167961f9dSMike Kravetz */ 191267961f9dSMike Kravetz if (ret) 191367961f9dSMike Kravetz return 0; 191467961f9dSMike Kravetz else 191567961f9dSMike Kravetz return 1; 191667961f9dSMike Kravetz } 19174e35f483SJoonsoo Kim else 1918cf3ad20bSMike Kravetz return ret < 0 ? ret : 0; 191984afd99bSAndy Whitcroft } 1920cf3ad20bSMike Kravetz 1921cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 1922a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 1923c37f9fb1SAndy Whitcroft { 19245e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 1925cf3ad20bSMike Kravetz } 1926c37f9fb1SAndy Whitcroft 1927cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 1928cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 1929cf3ad20bSMike Kravetz { 19305e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 19315e911373SMike Kravetz } 19325e911373SMike Kravetz 1933feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 19345e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 19355e911373SMike Kravetz { 1936feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 1937c37f9fb1SAndy Whitcroft } 1938c37f9fb1SAndy Whitcroft 193996b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 194096b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 194196b96a96SMike Kravetz { 194296b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 194396b96a96SMike Kravetz } 194496b96a96SMike Kravetz 194596b96a96SMike Kravetz /* 194696b96a96SMike Kravetz * This routine is called to restore a reservation on error paths. In the 194796b96a96SMike Kravetz * specific error paths, a huge page was allocated (via alloc_huge_page) 194896b96a96SMike Kravetz * and is about to be freed. If a reservation for the page existed, 194996b96a96SMike Kravetz * alloc_huge_page would have consumed the reservation and set PagePrivate 195096b96a96SMike Kravetz * in the newly allocated page. When the page is freed via free_huge_page, 195196b96a96SMike Kravetz * the global reservation count will be incremented if PagePrivate is set. 195296b96a96SMike Kravetz * However, free_huge_page can not adjust the reserve map. Adjust the 195396b96a96SMike Kravetz * reserve map here to be consistent with global reserve count adjustments 195496b96a96SMike Kravetz * to be made by free_huge_page. 195596b96a96SMike Kravetz */ 195696b96a96SMike Kravetz static void restore_reserve_on_error(struct hstate *h, 195796b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long address, 195896b96a96SMike Kravetz struct page *page) 195996b96a96SMike Kravetz { 196096b96a96SMike Kravetz if (unlikely(PagePrivate(page))) { 196196b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 196296b96a96SMike Kravetz 196396b96a96SMike Kravetz if (unlikely(rc < 0)) { 196496b96a96SMike Kravetz /* 196596b96a96SMike Kravetz * Rare out of memory condition in reserve map 196696b96a96SMike Kravetz * manipulation. Clear PagePrivate so that 196796b96a96SMike Kravetz * global reserve count will not be incremented 196896b96a96SMike Kravetz * by free_huge_page. This will make it appear 196996b96a96SMike Kravetz * as though the reservation for this page was 197096b96a96SMike Kravetz * consumed. This may prevent the task from 197196b96a96SMike Kravetz * faulting in the page at a later time. This 197296b96a96SMike Kravetz * is better than inconsistent global huge page 197396b96a96SMike Kravetz * accounting of reserve counts. 197496b96a96SMike Kravetz */ 197596b96a96SMike Kravetz ClearPagePrivate(page); 197696b96a96SMike Kravetz } else if (rc) { 197796b96a96SMike Kravetz rc = vma_add_reservation(h, vma, address); 197896b96a96SMike Kravetz if (unlikely(rc < 0)) 197996b96a96SMike Kravetz /* 198096b96a96SMike Kravetz * See above comment about rare out of 198196b96a96SMike Kravetz * memory condition. 198296b96a96SMike Kravetz */ 198396b96a96SMike Kravetz ClearPagePrivate(page); 198496b96a96SMike Kravetz } else 198596b96a96SMike Kravetz vma_end_reservation(h, vma, address); 198696b96a96SMike Kravetz } 198796b96a96SMike Kravetz } 198896b96a96SMike Kravetz 198970c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 199004f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 1991348ea204SAdam Litke { 199290481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 1993a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 1994348ea204SAdam Litke struct page *page; 1995d85f69b0SMike Kravetz long map_chg, map_commit; 1996d85f69b0SMike Kravetz long gbl_chg; 19976d76dcf4SAneesh Kumar K.V int ret, idx; 19986d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 19992fc39cecSAdam Litke 20006d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 2001a1e78772SMel Gorman /* 2002d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 2003d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 2004d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 2005a1e78772SMel Gorman */ 2006d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 2007d85f69b0SMike Kravetz if (map_chg < 0) 200876dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 2009d85f69b0SMike Kravetz 2010d85f69b0SMike Kravetz /* 2011d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 2012d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 2013d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 2014d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 2015d85f69b0SMike Kravetz * checked against any subpool limit. 2016d85f69b0SMike Kravetz */ 2017d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 2018d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 2019d85f69b0SMike Kravetz if (gbl_chg < 0) { 2020feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 202176dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 20225e911373SMike Kravetz } 202390d8b7e6SAdam Litke 2024d85f69b0SMike Kravetz /* 2025d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 2026d85f69b0SMike Kravetz * map, there could be reservations associated with the 2027d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 2028d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 2029d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 2030d85f69b0SMike Kravetz * the subpool reservations. 2031d85f69b0SMike Kravetz */ 2032d85f69b0SMike Kravetz if (avoid_reserve) 2033d85f69b0SMike Kravetz gbl_chg = 1; 2034d85f69b0SMike Kravetz } 2035d85f69b0SMike Kravetz 20366d76dcf4SAneesh Kumar K.V ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 20378f34af6fSJianyu Zhan if (ret) 20388f34af6fSJianyu Zhan goto out_subpool_put; 20398f34af6fSJianyu Zhan 2040a1e78772SMel Gorman spin_lock(&hugetlb_lock); 2041d85f69b0SMike Kravetz /* 2042d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 2043d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 2044d85f69b0SMike Kravetz * a reservation exists for the allocation. 2045d85f69b0SMike Kravetz */ 2046d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 204781a6fcaeSJoonsoo Kim if (!page) { 204894ae8ba7SAneesh Kumar K.V spin_unlock(&hugetlb_lock); 2049099730d6SDave Hansen page = __alloc_buddy_huge_page_with_mpol(h, vma, addr); 20508f34af6fSJianyu Zhan if (!page) 20518f34af6fSJianyu Zhan goto out_uncharge_cgroup; 2052a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 2053a88c7695SNaoya Horiguchi SetPagePrivate(page); 2054a88c7695SNaoya Horiguchi h->resv_huge_pages--; 2055a88c7695SNaoya Horiguchi } 205679dbb236SAneesh Kumar K.V spin_lock(&hugetlb_lock); 205779dbb236SAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 205881a6fcaeSJoonsoo Kim /* Fall through */ 2059a1e78772SMel Gorman } 206081a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 206181a6fcaeSJoonsoo Kim spin_unlock(&hugetlb_lock); 2062a1e78772SMel Gorman 206390481622SDavid Gibson set_page_private(page, (unsigned long)spool); 2064a1e78772SMel Gorman 2065d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 2066d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 206733039678SMike Kravetz /* 206833039678SMike Kravetz * The page was added to the reservation map between 206933039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 207033039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 207133039678SMike Kravetz * Adjust for the subpool count incremented above AND 207233039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 207333039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 207433039678SMike Kravetz * no longer applies. 207533039678SMike Kravetz */ 207633039678SMike Kravetz long rsv_adjust; 207733039678SMike Kravetz 207833039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 207933039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 208033039678SMike Kravetz } 20817893d1d5SAdam Litke return page; 20828f34af6fSJianyu Zhan 20838f34af6fSJianyu Zhan out_uncharge_cgroup: 20848f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 20858f34af6fSJianyu Zhan out_subpool_put: 2086d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 20878f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 2088feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 20898f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 2090b45b5bd6SDavid Gibson } 2091b45b5bd6SDavid Gibson 209274060e4dSNaoya Horiguchi /* 209374060e4dSNaoya Horiguchi * alloc_huge_page()'s wrapper which simply returns the page if allocation 209474060e4dSNaoya Horiguchi * succeeds, otherwise NULL. This function is called from new_vma_page(), 209574060e4dSNaoya Horiguchi * where no ERR_VALUE is expected to be returned. 209674060e4dSNaoya Horiguchi */ 209774060e4dSNaoya Horiguchi struct page *alloc_huge_page_noerr(struct vm_area_struct *vma, 209874060e4dSNaoya Horiguchi unsigned long addr, int avoid_reserve) 209974060e4dSNaoya Horiguchi { 210074060e4dSNaoya Horiguchi struct page *page = alloc_huge_page(vma, addr, avoid_reserve); 210174060e4dSNaoya Horiguchi if (IS_ERR(page)) 210274060e4dSNaoya Horiguchi page = NULL; 210374060e4dSNaoya Horiguchi return page; 210474060e4dSNaoya Horiguchi } 210574060e4dSNaoya Horiguchi 2106e24a1307SAneesh Kumar K.V int alloc_bootmem_huge_page(struct hstate *h) 2107e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 2108e24a1307SAneesh Kumar K.V int __alloc_bootmem_huge_page(struct hstate *h) 2109aa888a74SAndi Kleen { 2110aa888a74SAndi Kleen struct huge_bootmem_page *m; 2111b2261026SJoonsoo Kim int nr_nodes, node; 2112aa888a74SAndi Kleen 2113b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 2114aa888a74SAndi Kleen void *addr; 2115aa888a74SAndi Kleen 21168b89a116SGrygorii Strashko addr = memblock_virt_alloc_try_nid_nopanic( 21178b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 21188b89a116SGrygorii Strashko 0, BOOTMEM_ALLOC_ACCESSIBLE, node); 2119aa888a74SAndi Kleen if (addr) { 2120aa888a74SAndi Kleen /* 2121aa888a74SAndi Kleen * Use the beginning of the huge page to store the 2122aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 2123aa888a74SAndi Kleen * puts them into the mem_map). 2124aa888a74SAndi Kleen */ 2125aa888a74SAndi Kleen m = addr; 2126aa888a74SAndi Kleen goto found; 2127aa888a74SAndi Kleen } 2128aa888a74SAndi Kleen } 2129aa888a74SAndi Kleen return 0; 2130aa888a74SAndi Kleen 2131aa888a74SAndi Kleen found: 2132df994eadSLuiz Capitulino BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); 2133aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 2134aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 2135aa888a74SAndi Kleen m->hstate = h; 2136aa888a74SAndi Kleen return 1; 2137aa888a74SAndi Kleen } 2138aa888a74SAndi Kleen 2139d00181b9SKirill A. Shutemov static void __init prep_compound_huge_page(struct page *page, 2140d00181b9SKirill A. Shutemov unsigned int order) 214118229df5SAndy Whitcroft { 214218229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 214318229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 214418229df5SAndy Whitcroft else 214518229df5SAndy Whitcroft prep_compound_page(page, order); 214618229df5SAndy Whitcroft } 214718229df5SAndy Whitcroft 2148aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 2149aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 2150aa888a74SAndi Kleen { 2151aa888a74SAndi Kleen struct huge_bootmem_page *m; 2152aa888a74SAndi Kleen 2153aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 2154aa888a74SAndi Kleen struct hstate *h = m->hstate; 2155ee8f248dSBecky Bruce struct page *page; 2156ee8f248dSBecky Bruce 2157ee8f248dSBecky Bruce #ifdef CONFIG_HIGHMEM 2158ee8f248dSBecky Bruce page = pfn_to_page(m->phys >> PAGE_SHIFT); 21598b89a116SGrygorii Strashko memblock_free_late(__pa(m), 2160ee8f248dSBecky Bruce sizeof(struct huge_bootmem_page)); 2161ee8f248dSBecky Bruce #else 2162ee8f248dSBecky Bruce page = virt_to_page(m); 2163ee8f248dSBecky Bruce #endif 2164aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 216518229df5SAndy Whitcroft prep_compound_huge_page(page, h->order); 2166ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 2167aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 2168af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 2169af0fb9dfSMichal Hocko 2170b0320c7bSRafael Aquini /* 2171b0320c7bSRafael Aquini * If we had gigantic hugepages allocated at boot time, we need 2172b0320c7bSRafael Aquini * to restore the 'stolen' pages to totalram_pages in order to 2173b0320c7bSRafael Aquini * fix confusing memory reports from free(1) and another 2174b0320c7bSRafael Aquini * side-effects, like CommitLimit going negative. 2175b0320c7bSRafael Aquini */ 2176bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 21773dcc0571SJiang Liu adjust_managed_page_count(page, 1 << h->order); 2178aa888a74SAndi Kleen } 2179aa888a74SAndi Kleen } 2180aa888a74SAndi Kleen 21818faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 21821da177e4SLinus Torvalds { 21831da177e4SLinus Torvalds unsigned long i; 21841da177e4SLinus Torvalds 2185e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 2186bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 2187aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 2188aa888a74SAndi Kleen break; 21899b5e5d0fSLee Schermerhorn } else if (!alloc_fresh_huge_page(h, 21908cebfcd0SLai Jiangshan &node_states[N_MEMORY])) 21911da177e4SLinus Torvalds break; 219269ed779aSDavid Rientjes cond_resched(); 21931da177e4SLinus Torvalds } 2194d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 2195d715cf80SLiam R. Howlett char buf[32]; 2196d715cf80SLiam R. Howlett 2197c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2198d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 2199d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 22008faa8b07SAndi Kleen h->max_huge_pages = i; 2201e5ff2159SAndi Kleen } 2202d715cf80SLiam R. Howlett } 2203e5ff2159SAndi Kleen 2204e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 2205e5ff2159SAndi Kleen { 2206e5ff2159SAndi Kleen struct hstate *h; 2207e5ff2159SAndi Kleen 2208e5ff2159SAndi Kleen for_each_hstate(h) { 2209641844f5SNaoya Horiguchi if (minimum_order > huge_page_order(h)) 2210641844f5SNaoya Horiguchi minimum_order = huge_page_order(h); 2211641844f5SNaoya Horiguchi 22128faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 2213bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 22148faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 2215e5ff2159SAndi Kleen } 2216641844f5SNaoya Horiguchi VM_BUG_ON(minimum_order == UINT_MAX); 2217e5ff2159SAndi Kleen } 2218e5ff2159SAndi Kleen 2219e5ff2159SAndi Kleen static void __init report_hugepages(void) 2220e5ff2159SAndi Kleen { 2221e5ff2159SAndi Kleen struct hstate *h; 2222e5ff2159SAndi Kleen 2223e5ff2159SAndi Kleen for_each_hstate(h) { 22244abd32dbSAndi Kleen char buf[32]; 2225c6247f72SMatthew Wilcox 2226c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2227ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 2228c6247f72SMatthew Wilcox buf, h->free_huge_pages); 2229e5ff2159SAndi Kleen } 2230e5ff2159SAndi Kleen } 2231e5ff2159SAndi Kleen 22321da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 22336ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 22346ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 22351da177e4SLinus Torvalds { 22364415cc8dSChristoph Lameter int i; 22374415cc8dSChristoph Lameter 2238bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2239aa888a74SAndi Kleen return; 2240aa888a74SAndi Kleen 22416ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 22421da177e4SLinus Torvalds struct page *page, *next; 2243a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 2244a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 2245a5516438SAndi Kleen if (count >= h->nr_huge_pages) 22466b0c880dSAdam Litke return; 22471da177e4SLinus Torvalds if (PageHighMem(page)) 22481da177e4SLinus Torvalds continue; 22491da177e4SLinus Torvalds list_del(&page->lru); 2250e5ff2159SAndi Kleen update_and_free_page(h, page); 2251a5516438SAndi Kleen h->free_huge_pages--; 2252a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 22531da177e4SLinus Torvalds } 22541da177e4SLinus Torvalds } 22551da177e4SLinus Torvalds } 22561da177e4SLinus Torvalds #else 22576ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 22586ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 22591da177e4SLinus Torvalds { 22601da177e4SLinus Torvalds } 22611da177e4SLinus Torvalds #endif 22621da177e4SLinus Torvalds 226320a0307cSWu Fengguang /* 226420a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 226520a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 226620a0307cSWu Fengguang * Returns 1 if an adjustment was made. 226720a0307cSWu Fengguang */ 22686ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 22696ae11b27SLee Schermerhorn int delta) 227020a0307cSWu Fengguang { 2271b2261026SJoonsoo Kim int nr_nodes, node; 227220a0307cSWu Fengguang 227320a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 227420a0307cSWu Fengguang 2275e8c5c824SLee Schermerhorn if (delta < 0) { 2276b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 2277b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 2278b2261026SJoonsoo Kim goto found; 2279b2261026SJoonsoo Kim } 2280b2261026SJoonsoo Kim } else { 2281b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2282b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 2283b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 2284b2261026SJoonsoo Kim goto found; 2285e8c5c824SLee Schermerhorn } 22869a76db09SLee Schermerhorn } 2287b2261026SJoonsoo Kim return 0; 228820a0307cSWu Fengguang 2289b2261026SJoonsoo Kim found: 229020a0307cSWu Fengguang h->surplus_huge_pages += delta; 2291b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 2292b2261026SJoonsoo Kim return 1; 229320a0307cSWu Fengguang } 229420a0307cSWu Fengguang 2295a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 22966ae11b27SLee Schermerhorn static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, 22976ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 22981da177e4SLinus Torvalds { 22997893d1d5SAdam Litke unsigned long min_count, ret; 23001da177e4SLinus Torvalds 2301944d9fecSLuiz Capitulino if (hstate_is_gigantic(h) && !gigantic_page_supported()) 2302aa888a74SAndi Kleen return h->max_huge_pages; 2303aa888a74SAndi Kleen 23047893d1d5SAdam Litke /* 23057893d1d5SAdam Litke * Increase the pool size 23067893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 23077893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 2308d1c3fb1fSNishanth Aravamudan * 2309ab5ac90aSMichal Hocko * We might race with __alloc_surplus_huge_page() here and be unable 2310d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 2311d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 2312d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 2313d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 23147893d1d5SAdam Litke */ 23151da177e4SLinus Torvalds spin_lock(&hugetlb_lock); 2316a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 23176ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 23187893d1d5SAdam Litke break; 23197893d1d5SAdam Litke } 23207893d1d5SAdam Litke 2321a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 23227893d1d5SAdam Litke /* 23237893d1d5SAdam Litke * If this allocation races such that we no longer need the 23247893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 23257893d1d5SAdam Litke * and reducing the surplus. 23267893d1d5SAdam Litke */ 23277893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 2328649920c6SJia He 2329649920c6SJia He /* yield cpu to avoid soft lockup */ 2330649920c6SJia He cond_resched(); 2331649920c6SJia He 23326ae11b27SLee Schermerhorn ret = alloc_fresh_huge_page(h, nodes_allowed); 23337893d1d5SAdam Litke spin_lock(&hugetlb_lock); 23347893d1d5SAdam Litke if (!ret) 23357893d1d5SAdam Litke goto out; 23367893d1d5SAdam Litke 2337536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 2338536240f2SMel Gorman if (signal_pending(current)) 2339536240f2SMel Gorman goto out; 23407893d1d5SAdam Litke } 23417893d1d5SAdam Litke 23427893d1d5SAdam Litke /* 23437893d1d5SAdam Litke * Decrease the pool size 23447893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 23457893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 23467893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 23477893d1d5SAdam Litke * to the desired size as pages become free. 2348d1c3fb1fSNishanth Aravamudan * 2349d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 2350d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 2351d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 2352ab5ac90aSMichal Hocko * __alloc_surplus_huge_page() is checking the global counter, 2353d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 2354d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 2355d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 23567893d1d5SAdam Litke */ 2357a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 23586b0c880dSAdam Litke min_count = max(count, min_count); 23596ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 2360a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 23616ae11b27SLee Schermerhorn if (!free_pool_huge_page(h, nodes_allowed, 0)) 23621da177e4SLinus Torvalds break; 236355f67141SMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 23641da177e4SLinus Torvalds } 2365a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 23666ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 23677893d1d5SAdam Litke break; 23687893d1d5SAdam Litke } 23697893d1d5SAdam Litke out: 2370a5516438SAndi Kleen ret = persistent_huge_pages(h); 23711da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 23727893d1d5SAdam Litke return ret; 23731da177e4SLinus Torvalds } 23741da177e4SLinus Torvalds 2375a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 2376a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 2377a3437870SNishanth Aravamudan 2378a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 2379a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 2380a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 2381a3437870SNishanth Aravamudan 2382a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 2383a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 2384a3437870SNishanth Aravamudan 23859a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 23869a305230SLee Schermerhorn 23879a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 2388a3437870SNishanth Aravamudan { 2389a3437870SNishanth Aravamudan int i; 23909a305230SLee Schermerhorn 2391a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 23929a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 23939a305230SLee Schermerhorn if (nidp) 23949a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 2395a3437870SNishanth Aravamudan return &hstates[i]; 23969a305230SLee Schermerhorn } 23979a305230SLee Schermerhorn 23989a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 2399a3437870SNishanth Aravamudan } 2400a3437870SNishanth Aravamudan 240106808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 2402a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2403a3437870SNishanth Aravamudan { 24049a305230SLee Schermerhorn struct hstate *h; 24059a305230SLee Schermerhorn unsigned long nr_huge_pages; 24069a305230SLee Schermerhorn int nid; 24079a305230SLee Schermerhorn 24089a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 24099a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 24109a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 24119a305230SLee Schermerhorn else 24129a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 24139a305230SLee Schermerhorn 24149a305230SLee Schermerhorn return sprintf(buf, "%lu\n", nr_huge_pages); 2415a3437870SNishanth Aravamudan } 2416adbe8726SEric B Munson 2417238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 2418238d3c13SDavid Rientjes struct hstate *h, int nid, 2419238d3c13SDavid Rientjes unsigned long count, size_t len) 2420a3437870SNishanth Aravamudan { 2421a3437870SNishanth Aravamudan int err; 2422bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); 2423a3437870SNishanth Aravamudan 2424944d9fecSLuiz Capitulino if (hstate_is_gigantic(h) && !gigantic_page_supported()) { 2425adbe8726SEric B Munson err = -EINVAL; 2426adbe8726SEric B Munson goto out; 2427adbe8726SEric B Munson } 2428adbe8726SEric B Munson 24299a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 24309a305230SLee Schermerhorn /* 24319a305230SLee Schermerhorn * global hstate attribute 24329a305230SLee Schermerhorn */ 24339a305230SLee Schermerhorn if (!(obey_mempolicy && 24349a305230SLee Schermerhorn init_nodemask_of_mempolicy(nodes_allowed))) { 243506808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 24368cebfcd0SLai Jiangshan nodes_allowed = &node_states[N_MEMORY]; 243706808b08SLee Schermerhorn } 24389a305230SLee Schermerhorn } else if (nodes_allowed) { 24399a305230SLee Schermerhorn /* 24409a305230SLee Schermerhorn * per node hstate attribute: adjust count to global, 24419a305230SLee Schermerhorn * but restrict alloc/free to the specified node. 24429a305230SLee Schermerhorn */ 24439a305230SLee Schermerhorn count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 24449a305230SLee Schermerhorn init_nodemask_of_node(nodes_allowed, nid); 24459a305230SLee Schermerhorn } else 24468cebfcd0SLai Jiangshan nodes_allowed = &node_states[N_MEMORY]; 24479a305230SLee Schermerhorn 244806808b08SLee Schermerhorn h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); 2449a3437870SNishanth Aravamudan 24508cebfcd0SLai Jiangshan if (nodes_allowed != &node_states[N_MEMORY]) 245106808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 245206808b08SLee Schermerhorn 245306808b08SLee Schermerhorn return len; 2454adbe8726SEric B Munson out: 2455adbe8726SEric B Munson NODEMASK_FREE(nodes_allowed); 2456adbe8726SEric B Munson return err; 245706808b08SLee Schermerhorn } 245806808b08SLee Schermerhorn 2459238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 2460238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 2461238d3c13SDavid Rientjes size_t len) 2462238d3c13SDavid Rientjes { 2463238d3c13SDavid Rientjes struct hstate *h; 2464238d3c13SDavid Rientjes unsigned long count; 2465238d3c13SDavid Rientjes int nid; 2466238d3c13SDavid Rientjes int err; 2467238d3c13SDavid Rientjes 2468238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 2469238d3c13SDavid Rientjes if (err) 2470238d3c13SDavid Rientjes return err; 2471238d3c13SDavid Rientjes 2472238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 2473238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 2474238d3c13SDavid Rientjes } 2475238d3c13SDavid Rientjes 247606808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 247706808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 247806808b08SLee Schermerhorn { 247906808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 248006808b08SLee Schermerhorn } 248106808b08SLee Schermerhorn 248206808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 248306808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 248406808b08SLee Schermerhorn { 2485238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 2486a3437870SNishanth Aravamudan } 2487a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 2488a3437870SNishanth Aravamudan 248906808b08SLee Schermerhorn #ifdef CONFIG_NUMA 249006808b08SLee Schermerhorn 249106808b08SLee Schermerhorn /* 249206808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 249306808b08SLee Schermerhorn * huge page alloc/free. 249406808b08SLee Schermerhorn */ 249506808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 249606808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 249706808b08SLee Schermerhorn { 249806808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 249906808b08SLee Schermerhorn } 250006808b08SLee Schermerhorn 250106808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 250206808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 250306808b08SLee Schermerhorn { 2504238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 250506808b08SLee Schermerhorn } 250606808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 250706808b08SLee Schermerhorn #endif 250806808b08SLee Schermerhorn 250906808b08SLee Schermerhorn 2510a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 2511a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2512a3437870SNishanth Aravamudan { 25139a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2514a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 2515a3437870SNishanth Aravamudan } 2516adbe8726SEric B Munson 2517a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 2518a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 2519a3437870SNishanth Aravamudan { 2520a3437870SNishanth Aravamudan int err; 2521a3437870SNishanth Aravamudan unsigned long input; 25229a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2523a3437870SNishanth Aravamudan 2524bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2525adbe8726SEric B Munson return -EINVAL; 2526adbe8726SEric B Munson 25273dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 2528a3437870SNishanth Aravamudan if (err) 252973ae31e5SEric B Munson return err; 2530a3437870SNishanth Aravamudan 2531a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 2532a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 2533a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 2534a3437870SNishanth Aravamudan 2535a3437870SNishanth Aravamudan return count; 2536a3437870SNishanth Aravamudan } 2537a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 2538a3437870SNishanth Aravamudan 2539a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 2540a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2541a3437870SNishanth Aravamudan { 25429a305230SLee Schermerhorn struct hstate *h; 25439a305230SLee Schermerhorn unsigned long free_huge_pages; 25449a305230SLee Schermerhorn int nid; 25459a305230SLee Schermerhorn 25469a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 25479a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 25489a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 25499a305230SLee Schermerhorn else 25509a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 25519a305230SLee Schermerhorn 25529a305230SLee Schermerhorn return sprintf(buf, "%lu\n", free_huge_pages); 2553a3437870SNishanth Aravamudan } 2554a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 2555a3437870SNishanth Aravamudan 2556a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 2557a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2558a3437870SNishanth Aravamudan { 25599a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2560a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->resv_huge_pages); 2561a3437870SNishanth Aravamudan } 2562a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 2563a3437870SNishanth Aravamudan 2564a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 2565a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2566a3437870SNishanth Aravamudan { 25679a305230SLee Schermerhorn struct hstate *h; 25689a305230SLee Schermerhorn unsigned long surplus_huge_pages; 25699a305230SLee Schermerhorn int nid; 25709a305230SLee Schermerhorn 25719a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 25729a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 25739a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 25749a305230SLee Schermerhorn else 25759a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 25769a305230SLee Schermerhorn 25779a305230SLee Schermerhorn return sprintf(buf, "%lu\n", surplus_huge_pages); 2578a3437870SNishanth Aravamudan } 2579a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 2580a3437870SNishanth Aravamudan 2581a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 2582a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 2583a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 2584a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 2585a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 2586a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 258706808b08SLee Schermerhorn #ifdef CONFIG_NUMA 258806808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 258906808b08SLee Schermerhorn #endif 2590a3437870SNishanth Aravamudan NULL, 2591a3437870SNishanth Aravamudan }; 2592a3437870SNishanth Aravamudan 259367e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 2594a3437870SNishanth Aravamudan .attrs = hstate_attrs, 2595a3437870SNishanth Aravamudan }; 2596a3437870SNishanth Aravamudan 2597094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 25989a305230SLee Schermerhorn struct kobject **hstate_kobjs, 259967e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 2600a3437870SNishanth Aravamudan { 2601a3437870SNishanth Aravamudan int retval; 2602972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 2603a3437870SNishanth Aravamudan 26049a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 26059a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 2606a3437870SNishanth Aravamudan return -ENOMEM; 2607a3437870SNishanth Aravamudan 26089a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 2609a3437870SNishanth Aravamudan if (retval) 26109a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 2611a3437870SNishanth Aravamudan 2612a3437870SNishanth Aravamudan return retval; 2613a3437870SNishanth Aravamudan } 2614a3437870SNishanth Aravamudan 2615a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 2616a3437870SNishanth Aravamudan { 2617a3437870SNishanth Aravamudan struct hstate *h; 2618a3437870SNishanth Aravamudan int err; 2619a3437870SNishanth Aravamudan 2620a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 2621a3437870SNishanth Aravamudan if (!hugepages_kobj) 2622a3437870SNishanth Aravamudan return; 2623a3437870SNishanth Aravamudan 2624a3437870SNishanth Aravamudan for_each_hstate(h) { 26259a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 26269a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 2627a3437870SNishanth Aravamudan if (err) 2628ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s", h->name); 2629a3437870SNishanth Aravamudan } 2630a3437870SNishanth Aravamudan } 2631a3437870SNishanth Aravamudan 26329a305230SLee Schermerhorn #ifdef CONFIG_NUMA 26339a305230SLee Schermerhorn 26349a305230SLee Schermerhorn /* 26359a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 263610fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 263710fbcf4cSKay Sievers * index of a node device or _hstate == node id. 263810fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 26399a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 26409a305230SLee Schermerhorn */ 26419a305230SLee Schermerhorn struct node_hstate { 26429a305230SLee Schermerhorn struct kobject *hugepages_kobj; 26439a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 26449a305230SLee Schermerhorn }; 2645b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 26469a305230SLee Schermerhorn 26479a305230SLee Schermerhorn /* 264810fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 26499a305230SLee Schermerhorn */ 26509a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 26519a305230SLee Schermerhorn &nr_hugepages_attr.attr, 26529a305230SLee Schermerhorn &free_hugepages_attr.attr, 26539a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 26549a305230SLee Schermerhorn NULL, 26559a305230SLee Schermerhorn }; 26569a305230SLee Schermerhorn 265767e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 26589a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 26599a305230SLee Schermerhorn }; 26609a305230SLee Schermerhorn 26619a305230SLee Schermerhorn /* 266210fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 26639a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 26649a305230SLee Schermerhorn */ 26659a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 26669a305230SLee Schermerhorn { 26679a305230SLee Schermerhorn int nid; 26689a305230SLee Schermerhorn 26699a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 26709a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 26719a305230SLee Schermerhorn int i; 26729a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 26739a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 26749a305230SLee Schermerhorn if (nidp) 26759a305230SLee Schermerhorn *nidp = nid; 26769a305230SLee Schermerhorn return &hstates[i]; 26779a305230SLee Schermerhorn } 26789a305230SLee Schermerhorn } 26799a305230SLee Schermerhorn 26809a305230SLee Schermerhorn BUG(); 26819a305230SLee Schermerhorn return NULL; 26829a305230SLee Schermerhorn } 26839a305230SLee Schermerhorn 26849a305230SLee Schermerhorn /* 268510fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 26869a305230SLee Schermerhorn * No-op if no hstate attributes attached. 26879a305230SLee Schermerhorn */ 26883cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 26899a305230SLee Schermerhorn { 26909a305230SLee Schermerhorn struct hstate *h; 269110fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 26929a305230SLee Schermerhorn 26939a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 26949b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 26959a305230SLee Schermerhorn 2696972dc4deSAneesh Kumar K.V for_each_hstate(h) { 2697972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 2698972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 2699972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 2700972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 2701972dc4deSAneesh Kumar K.V } 27029a305230SLee Schermerhorn } 27039a305230SLee Schermerhorn 27049a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 27059a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 27069a305230SLee Schermerhorn } 27079a305230SLee Schermerhorn 27089a305230SLee Schermerhorn 27099a305230SLee Schermerhorn /* 271010fbcf4cSKay Sievers * Register hstate attributes for a single node device. 27119a305230SLee Schermerhorn * No-op if attributes already registered. 27129a305230SLee Schermerhorn */ 27133cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 27149a305230SLee Schermerhorn { 27159a305230SLee Schermerhorn struct hstate *h; 271610fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 27179a305230SLee Schermerhorn int err; 27189a305230SLee Schermerhorn 27199a305230SLee Schermerhorn if (nhs->hugepages_kobj) 27209a305230SLee Schermerhorn return; /* already allocated */ 27219a305230SLee Schermerhorn 27229a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 272310fbcf4cSKay Sievers &node->dev.kobj); 27249a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 27259a305230SLee Schermerhorn return; 27269a305230SLee Schermerhorn 27279a305230SLee Schermerhorn for_each_hstate(h) { 27289a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 27299a305230SLee Schermerhorn nhs->hstate_kobjs, 27309a305230SLee Schermerhorn &per_node_hstate_attr_group); 27319a305230SLee Schermerhorn if (err) { 2732ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s for node %d\n", 273310fbcf4cSKay Sievers h->name, node->dev.id); 27349a305230SLee Schermerhorn hugetlb_unregister_node(node); 27359a305230SLee Schermerhorn break; 27369a305230SLee Schermerhorn } 27379a305230SLee Schermerhorn } 27389a305230SLee Schermerhorn } 27399a305230SLee Schermerhorn 27409a305230SLee Schermerhorn /* 27419b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 274210fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 274310fbcf4cSKay Sievers * registered their associated device by this time. 27449a305230SLee Schermerhorn */ 27457d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 27469a305230SLee Schermerhorn { 27479a305230SLee Schermerhorn int nid; 27489a305230SLee Schermerhorn 27498cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 27508732794bSWen Congyang struct node *node = node_devices[nid]; 275110fbcf4cSKay Sievers if (node->dev.id == nid) 27529a305230SLee Schermerhorn hugetlb_register_node(node); 27539a305230SLee Schermerhorn } 27549a305230SLee Schermerhorn 27559a305230SLee Schermerhorn /* 275610fbcf4cSKay Sievers * Let the node device driver know we're here so it can 27579a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 27589a305230SLee Schermerhorn */ 27599a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 27609a305230SLee Schermerhorn hugetlb_unregister_node); 27619a305230SLee Schermerhorn } 27629a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 27639a305230SLee Schermerhorn 27649a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 27659a305230SLee Schermerhorn { 27669a305230SLee Schermerhorn BUG(); 27679a305230SLee Schermerhorn if (nidp) 27689a305230SLee Schermerhorn *nidp = -1; 27699a305230SLee Schermerhorn return NULL; 27709a305230SLee Schermerhorn } 27719a305230SLee Schermerhorn 27729a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 27739a305230SLee Schermerhorn 27749a305230SLee Schermerhorn #endif 27759a305230SLee Schermerhorn 2776a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 2777a3437870SNishanth Aravamudan { 27788382d914SDavidlohr Bueso int i; 27798382d914SDavidlohr Bueso 2780457c1b27SNishanth Aravamudan if (!hugepages_supported()) 27810ef89d25SBenjamin Herrenschmidt return 0; 2782a3437870SNishanth Aravamudan 2783e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) { 2784d715cf80SLiam R. Howlett if (default_hstate_size != 0) { 2785d715cf80SLiam R. Howlett pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n", 2786d715cf80SLiam R. Howlett default_hstate_size, HPAGE_SIZE); 2787d715cf80SLiam R. Howlett } 2788d715cf80SLiam R. Howlett 2789e11bfbfcSNick Piggin default_hstate_size = HPAGE_SIZE; 2790e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) 2791a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 2792a3437870SNishanth Aravamudan } 2793972dc4deSAneesh Kumar K.V default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); 2794f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 2795f8b74815SVaishali Thakkar if (!default_hstate.max_huge_pages) 2796e11bfbfcSNick Piggin default_hstate.max_huge_pages = default_hstate_max_huge_pages; 2797f8b74815SVaishali Thakkar } 2798a3437870SNishanth Aravamudan 2799a3437870SNishanth Aravamudan hugetlb_init_hstates(); 2800aa888a74SAndi Kleen gather_bootmem_prealloc(); 2801a3437870SNishanth Aravamudan report_hugepages(); 2802a3437870SNishanth Aravamudan 2803a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 28049a305230SLee Schermerhorn hugetlb_register_all_nodes(); 28057179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 28069a305230SLee Schermerhorn 28078382d914SDavidlohr Bueso #ifdef CONFIG_SMP 28088382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 28098382d914SDavidlohr Bueso #else 28108382d914SDavidlohr Bueso num_fault_mutexes = 1; 28118382d914SDavidlohr Bueso #endif 2812c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 28138382d914SDavidlohr Bueso kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL); 2814c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 28158382d914SDavidlohr Bueso 28168382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 2817c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 2818a3437870SNishanth Aravamudan return 0; 2819a3437870SNishanth Aravamudan } 28203e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 2821a3437870SNishanth Aravamudan 2822a3437870SNishanth Aravamudan /* Should be called on processing a hugepagesz=... option */ 28239fee021dSVaishali Thakkar void __init hugetlb_bad_size(void) 28249fee021dSVaishali Thakkar { 28259fee021dSVaishali Thakkar parsed_valid_hugepagesz = false; 28269fee021dSVaishali Thakkar } 28279fee021dSVaishali Thakkar 2828d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 2829a3437870SNishanth Aravamudan { 2830a3437870SNishanth Aravamudan struct hstate *h; 28318faa8b07SAndi Kleen unsigned long i; 28328faa8b07SAndi Kleen 2833a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 2834598d8091SJoe Perches pr_warn("hugepagesz= specified twice, ignoring\n"); 2835a3437870SNishanth Aravamudan return; 2836a3437870SNishanth Aravamudan } 283747d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 2838a3437870SNishanth Aravamudan BUG_ON(order == 0); 283947d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 2840a3437870SNishanth Aravamudan h->order = order; 2841a3437870SNishanth Aravamudan h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 28428faa8b07SAndi Kleen h->nr_huge_pages = 0; 28438faa8b07SAndi Kleen h->free_huge_pages = 0; 28448faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 28458faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 28460edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 284754f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 284854f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 2849a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 2850a3437870SNishanth Aravamudan huge_page_size(h)/1024); 28518faa8b07SAndi Kleen 2852a3437870SNishanth Aravamudan parsed_hstate = h; 2853a3437870SNishanth Aravamudan } 2854a3437870SNishanth Aravamudan 2855e11bfbfcSNick Piggin static int __init hugetlb_nrpages_setup(char *s) 2856a3437870SNishanth Aravamudan { 2857a3437870SNishanth Aravamudan unsigned long *mhp; 28588faa8b07SAndi Kleen static unsigned long *last_mhp; 2859a3437870SNishanth Aravamudan 28609fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 28619fee021dSVaishali Thakkar pr_warn("hugepages = %s preceded by " 28629fee021dSVaishali Thakkar "an unsupported hugepagesz, ignoring\n", s); 28639fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 28649fee021dSVaishali Thakkar return 1; 28659fee021dSVaishali Thakkar } 2866a3437870SNishanth Aravamudan /* 286747d38344SAneesh Kumar K.V * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, 2868a3437870SNishanth Aravamudan * so this hugepages= parameter goes to the "default hstate". 2869a3437870SNishanth Aravamudan */ 28709fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 2871a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 2872a3437870SNishanth Aravamudan else 2873a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 2874a3437870SNishanth Aravamudan 28758faa8b07SAndi Kleen if (mhp == last_mhp) { 2876598d8091SJoe Perches pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n"); 28778faa8b07SAndi Kleen return 1; 28788faa8b07SAndi Kleen } 28798faa8b07SAndi Kleen 2880a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 2881a3437870SNishanth Aravamudan *mhp = 0; 2882a3437870SNishanth Aravamudan 28838faa8b07SAndi Kleen /* 28848faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 28858faa8b07SAndi Kleen * But we need to allocate >= MAX_ORDER hstates here early to still 28868faa8b07SAndi Kleen * use the bootmem allocator. 28878faa8b07SAndi Kleen */ 288847d38344SAneesh Kumar K.V if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) 28898faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 28908faa8b07SAndi Kleen 28918faa8b07SAndi Kleen last_mhp = mhp; 28928faa8b07SAndi Kleen 2893a3437870SNishanth Aravamudan return 1; 2894a3437870SNishanth Aravamudan } 2895e11bfbfcSNick Piggin __setup("hugepages=", hugetlb_nrpages_setup); 2896e11bfbfcSNick Piggin 2897e11bfbfcSNick Piggin static int __init hugetlb_default_setup(char *s) 2898e11bfbfcSNick Piggin { 2899e11bfbfcSNick Piggin default_hstate_size = memparse(s, &s); 2900e11bfbfcSNick Piggin return 1; 2901e11bfbfcSNick Piggin } 2902e11bfbfcSNick Piggin __setup("default_hugepagesz=", hugetlb_default_setup); 2903a3437870SNishanth Aravamudan 29048a213460SNishanth Aravamudan static unsigned int cpuset_mems_nr(unsigned int *array) 29058a213460SNishanth Aravamudan { 29068a213460SNishanth Aravamudan int node; 29078a213460SNishanth Aravamudan unsigned int nr = 0; 29088a213460SNishanth Aravamudan 29098a213460SNishanth Aravamudan for_each_node_mask(node, cpuset_current_mems_allowed) 29108a213460SNishanth Aravamudan nr += array[node]; 29118a213460SNishanth Aravamudan 29128a213460SNishanth Aravamudan return nr; 29138a213460SNishanth Aravamudan } 29148a213460SNishanth Aravamudan 29158a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 291606808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 291706808b08SLee Schermerhorn struct ctl_table *table, int write, 291806808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 29191da177e4SLinus Torvalds { 2920e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 2921238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 292208d4a246SMichal Hocko int ret; 2923e5ff2159SAndi Kleen 2924457c1b27SNishanth Aravamudan if (!hugepages_supported()) 292586613628SJan Stancek return -EOPNOTSUPP; 2926457c1b27SNishanth Aravamudan 2927e5ff2159SAndi Kleen table->data = &tmp; 2928e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 292908d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 293008d4a246SMichal Hocko if (ret) 293108d4a246SMichal Hocko goto out; 2932e5ff2159SAndi Kleen 2933238d3c13SDavid Rientjes if (write) 2934238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 2935238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 293608d4a246SMichal Hocko out: 293708d4a246SMichal Hocko return ret; 29381da177e4SLinus Torvalds } 2939396faf03SMel Gorman 294006808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 294106808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 294206808b08SLee Schermerhorn { 294306808b08SLee Schermerhorn 294406808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 294506808b08SLee Schermerhorn buffer, length, ppos); 294606808b08SLee Schermerhorn } 294706808b08SLee Schermerhorn 294806808b08SLee Schermerhorn #ifdef CONFIG_NUMA 294906808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 295006808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 295106808b08SLee Schermerhorn { 295206808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 295306808b08SLee Schermerhorn buffer, length, ppos); 295406808b08SLee Schermerhorn } 295506808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 295606808b08SLee Schermerhorn 2957a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 29588d65af78SAlexey Dobriyan void __user *buffer, 2959a3d0c6aaSNishanth Aravamudan size_t *length, loff_t *ppos) 2960a3d0c6aaSNishanth Aravamudan { 2961a5516438SAndi Kleen struct hstate *h = &default_hstate; 2962e5ff2159SAndi Kleen unsigned long tmp; 296308d4a246SMichal Hocko int ret; 2964e5ff2159SAndi Kleen 2965457c1b27SNishanth Aravamudan if (!hugepages_supported()) 296686613628SJan Stancek return -EOPNOTSUPP; 2967457c1b27SNishanth Aravamudan 2968e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 2969e5ff2159SAndi Kleen 2970bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 2971adbe8726SEric B Munson return -EINVAL; 2972adbe8726SEric B Munson 2973e5ff2159SAndi Kleen table->data = &tmp; 2974e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 297508d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 297608d4a246SMichal Hocko if (ret) 297708d4a246SMichal Hocko goto out; 2978e5ff2159SAndi Kleen 2979e5ff2159SAndi Kleen if (write) { 2980064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 2981e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 2982a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 2983e5ff2159SAndi Kleen } 298408d4a246SMichal Hocko out: 298508d4a246SMichal Hocko return ret; 2986a3d0c6aaSNishanth Aravamudan } 2987a3d0c6aaSNishanth Aravamudan 29881da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 29891da177e4SLinus Torvalds 2990e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 29911da177e4SLinus Torvalds { 2992fcb2b0c5SRoman Gushchin struct hstate *h; 2993fcb2b0c5SRoman Gushchin unsigned long total = 0; 2994fcb2b0c5SRoman Gushchin 2995457c1b27SNishanth Aravamudan if (!hugepages_supported()) 2996457c1b27SNishanth Aravamudan return; 2997fcb2b0c5SRoman Gushchin 2998fcb2b0c5SRoman Gushchin for_each_hstate(h) { 2999fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 3000fcb2b0c5SRoman Gushchin 3001fcb2b0c5SRoman Gushchin total += (PAGE_SIZE << huge_page_order(h)) * count; 3002fcb2b0c5SRoman Gushchin 3003fcb2b0c5SRoman Gushchin if (h == &default_hstate) 3004e1759c21SAlexey Dobriyan seq_printf(m, 30051da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 30061da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 3007b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 30087893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 30094f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 3010fcb2b0c5SRoman Gushchin count, 3011a5516438SAndi Kleen h->free_huge_pages, 3012a5516438SAndi Kleen h->resv_huge_pages, 3013a5516438SAndi Kleen h->surplus_huge_pages, 3014fcb2b0c5SRoman Gushchin (PAGE_SIZE << huge_page_order(h)) / 1024); 3015fcb2b0c5SRoman Gushchin } 3016fcb2b0c5SRoman Gushchin 3017fcb2b0c5SRoman Gushchin seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024); 30181da177e4SLinus Torvalds } 30191da177e4SLinus Torvalds 30201da177e4SLinus Torvalds int hugetlb_report_node_meminfo(int nid, char *buf) 30211da177e4SLinus Torvalds { 3022a5516438SAndi Kleen struct hstate *h = &default_hstate; 3023457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3024457c1b27SNishanth Aravamudan return 0; 30251da177e4SLinus Torvalds return sprintf(buf, 30261da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 3027a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 3028a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 3029a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 3030a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 3031a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 30321da177e4SLinus Torvalds } 30331da177e4SLinus Torvalds 3034949f7ec5SDavid Rientjes void hugetlb_show_meminfo(void) 3035949f7ec5SDavid Rientjes { 3036949f7ec5SDavid Rientjes struct hstate *h; 3037949f7ec5SDavid Rientjes int nid; 3038949f7ec5SDavid Rientjes 3039457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3040457c1b27SNishanth Aravamudan return; 3041457c1b27SNishanth Aravamudan 3042949f7ec5SDavid Rientjes for_each_node_state(nid, N_MEMORY) 3043949f7ec5SDavid Rientjes for_each_hstate(h) 3044949f7ec5SDavid Rientjes pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 3045949f7ec5SDavid Rientjes nid, 3046949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 3047949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 3048949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 3049949f7ec5SDavid Rientjes 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 3050949f7ec5SDavid Rientjes } 3051949f7ec5SDavid Rientjes 30525d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 30535d317b2bSNaoya Horiguchi { 30545d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 30555d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 30565d317b2bSNaoya Horiguchi } 30575d317b2bSNaoya Horiguchi 30581da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 30591da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 30601da177e4SLinus Torvalds { 3061d0028588SWanpeng Li struct hstate *h; 3062d0028588SWanpeng Li unsigned long nr_total_pages = 0; 3063d0028588SWanpeng Li 3064d0028588SWanpeng Li for_each_hstate(h) 3065d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 3066d0028588SWanpeng Li return nr_total_pages; 30671da177e4SLinus Torvalds } 30681da177e4SLinus Torvalds 3069a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 3070fc1b8a73SMel Gorman { 3071fc1b8a73SMel Gorman int ret = -ENOMEM; 3072fc1b8a73SMel Gorman 3073fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 3074fc1b8a73SMel Gorman /* 3075fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 3076fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 3077fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 3078fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 3079fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 3080fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 3081fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 3082fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 3083fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 3084fc1b8a73SMel Gorman * 3085fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 3086fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 3087fc1b8a73SMel Gorman * we fall back to check against current free page availability as 3088fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 3089fc1b8a73SMel Gorman * semantics that cpuset has. 3090fc1b8a73SMel Gorman */ 3091fc1b8a73SMel Gorman if (delta > 0) { 3092a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 3093fc1b8a73SMel Gorman goto out; 3094fc1b8a73SMel Gorman 3095a5516438SAndi Kleen if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 3096a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 3097fc1b8a73SMel Gorman goto out; 3098fc1b8a73SMel Gorman } 3099fc1b8a73SMel Gorman } 3100fc1b8a73SMel Gorman 3101fc1b8a73SMel Gorman ret = 0; 3102fc1b8a73SMel Gorman if (delta < 0) 3103a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 3104fc1b8a73SMel Gorman 3105fc1b8a73SMel Gorman out: 3106fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 3107fc1b8a73SMel Gorman return ret; 3108fc1b8a73SMel Gorman } 3109fc1b8a73SMel Gorman 311084afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 311184afd99bSAndy Whitcroft { 3112f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 311384afd99bSAndy Whitcroft 311484afd99bSAndy Whitcroft /* 311584afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 311684afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 311784afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 311825985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 311984afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 312084afd99bSAndy Whitcroft * new reference here without additional locking. 312184afd99bSAndy Whitcroft */ 31224e35f483SJoonsoo Kim if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 3123f522c3acSJoonsoo Kim kref_get(&resv->refs); 312484afd99bSAndy Whitcroft } 312584afd99bSAndy Whitcroft 3126a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 3127a1e78772SMel Gorman { 3128a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3129f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 313090481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 31314e35f483SJoonsoo Kim unsigned long reserve, start, end; 31321c5ecae3SMike Kravetz long gbl_reserve; 313384afd99bSAndy Whitcroft 31344e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 31354e35f483SJoonsoo Kim return; 31364e35f483SJoonsoo Kim 3137a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 3138a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 313984afd99bSAndy Whitcroft 31404e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 314184afd99bSAndy Whitcroft 3142f031dd27SJoonsoo Kim kref_put(&resv->refs, resv_map_release); 314384afd99bSAndy Whitcroft 31447251ff78SAdam Litke if (reserve) { 31451c5ecae3SMike Kravetz /* 31461c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 31471c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 31481c5ecae3SMike Kravetz */ 31491c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 31501c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 31517251ff78SAdam Litke } 3152a1e78772SMel Gorman } 3153a1e78772SMel Gorman 315431383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 315531383c68SDan Williams { 315631383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 315731383c68SDan Williams return -EINVAL; 315831383c68SDan Williams return 0; 315931383c68SDan Williams } 316031383c68SDan Williams 31611da177e4SLinus Torvalds /* 31621da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 31631da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 31641da177e4SLinus Torvalds * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 31651da177e4SLinus Torvalds * this far. 31661da177e4SLinus Torvalds */ 316711bac800SDave Jiang static int hugetlb_vm_op_fault(struct vm_fault *vmf) 31681da177e4SLinus Torvalds { 31691da177e4SLinus Torvalds BUG(); 3170d0217ac0SNick Piggin return 0; 31711da177e4SLinus Torvalds } 31721da177e4SLinus Torvalds 3173f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 3174d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 317584afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 3176a1e78772SMel Gorman .close = hugetlb_vm_op_close, 317731383c68SDan Williams .split = hugetlb_vm_op_split, 31781da177e4SLinus Torvalds }; 31791da177e4SLinus Torvalds 31801e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 31811e8f889bSDavid Gibson int writable) 318263551ae0SDavid Gibson { 318363551ae0SDavid Gibson pte_t entry; 318463551ae0SDavid Gibson 31851e8f889bSDavid Gibson if (writable) { 3186106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 3187106c992aSGerald Schaefer vma->vm_page_prot))); 318863551ae0SDavid Gibson } else { 3189106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 3190106c992aSGerald Schaefer vma->vm_page_prot)); 319163551ae0SDavid Gibson } 319263551ae0SDavid Gibson entry = pte_mkyoung(entry); 319363551ae0SDavid Gibson entry = pte_mkhuge(entry); 3194d9ed9faaSChris Metcalf entry = arch_make_huge_pte(entry, vma, page, writable); 319563551ae0SDavid Gibson 319663551ae0SDavid Gibson return entry; 319763551ae0SDavid Gibson } 319863551ae0SDavid Gibson 31991e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 32001e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 32011e8f889bSDavid Gibson { 32021e8f889bSDavid Gibson pte_t entry; 32031e8f889bSDavid Gibson 3204106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 320532f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 32064b3073e1SRussell King update_mmu_cache(vma, address, ptep); 32071e8f889bSDavid Gibson } 32081e8f889bSDavid Gibson 3209d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 32104a705fefSNaoya Horiguchi { 32114a705fefSNaoya Horiguchi swp_entry_t swp; 32124a705fefSNaoya Horiguchi 32134a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 3214d5ed7444SAneesh Kumar K.V return false; 32154a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 32164a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_migration_entry(swp)) 3217d5ed7444SAneesh Kumar K.V return true; 32184a705fefSNaoya Horiguchi else 3219d5ed7444SAneesh Kumar K.V return false; 32204a705fefSNaoya Horiguchi } 32214a705fefSNaoya Horiguchi 32224a705fefSNaoya Horiguchi static int is_hugetlb_entry_hwpoisoned(pte_t pte) 32234a705fefSNaoya Horiguchi { 32244a705fefSNaoya Horiguchi swp_entry_t swp; 32254a705fefSNaoya Horiguchi 32264a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 32274a705fefSNaoya Horiguchi return 0; 32284a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 32294a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_hwpoison_entry(swp)) 32304a705fefSNaoya Horiguchi return 1; 32314a705fefSNaoya Horiguchi else 32324a705fefSNaoya Horiguchi return 0; 32334a705fefSNaoya Horiguchi } 32341e8f889bSDavid Gibson 323563551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 323663551ae0SDavid Gibson struct vm_area_struct *vma) 323763551ae0SDavid Gibson { 323863551ae0SDavid Gibson pte_t *src_pte, *dst_pte, entry; 323963551ae0SDavid Gibson struct page *ptepage; 32401c59827dSHugh Dickins unsigned long addr; 32411e8f889bSDavid Gibson int cow; 3242a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3243a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 3244e8569dd2SAndreas Sandberg unsigned long mmun_start; /* For mmu_notifiers */ 3245e8569dd2SAndreas Sandberg unsigned long mmun_end; /* For mmu_notifiers */ 3246e8569dd2SAndreas Sandberg int ret = 0; 32471e8f889bSDavid Gibson 32481e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 324963551ae0SDavid Gibson 3250e8569dd2SAndreas Sandberg mmun_start = vma->vm_start; 3251e8569dd2SAndreas Sandberg mmun_end = vma->vm_end; 3252e8569dd2SAndreas Sandberg if (cow) 3253e8569dd2SAndreas Sandberg mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end); 3254e8569dd2SAndreas Sandberg 3255a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 3256cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 32577868a208SPunit Agrawal src_pte = huge_pte_offset(src, addr, sz); 3258c74df32cSHugh Dickins if (!src_pte) 3259c74df32cSHugh Dickins continue; 3260a5516438SAndi Kleen dst_pte = huge_pte_alloc(dst, addr, sz); 3261e8569dd2SAndreas Sandberg if (!dst_pte) { 3262e8569dd2SAndreas Sandberg ret = -ENOMEM; 3263e8569dd2SAndreas Sandberg break; 3264e8569dd2SAndreas Sandberg } 3265c5c99429SLarry Woodman 3266c5c99429SLarry Woodman /* If the pagetables are shared don't copy or take references */ 3267c5c99429SLarry Woodman if (dst_pte == src_pte) 3268c5c99429SLarry Woodman continue; 3269c5c99429SLarry Woodman 3270cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 3271cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 3272cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 32734a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 32744a705fefSNaoya Horiguchi if (huge_pte_none(entry)) { /* skip none entry */ 32754a705fefSNaoya Horiguchi ; 32764a705fefSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry) || 32774a705fefSNaoya Horiguchi is_hugetlb_entry_hwpoisoned(entry))) { 32784a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 32794a705fefSNaoya Horiguchi 32804a705fefSNaoya Horiguchi if (is_write_migration_entry(swp_entry) && cow) { 32814a705fefSNaoya Horiguchi /* 32824a705fefSNaoya Horiguchi * COW mappings require pages in both 32834a705fefSNaoya Horiguchi * parent and child to be set to read. 32844a705fefSNaoya Horiguchi */ 32854a705fefSNaoya Horiguchi make_migration_entry_read(&swp_entry); 32864a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 3287e5251fd4SPunit Agrawal set_huge_swap_pte_at(src, addr, src_pte, 3288e5251fd4SPunit Agrawal entry, sz); 32894a705fefSNaoya Horiguchi } 3290e5251fd4SPunit Agrawal set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); 32914a705fefSNaoya Horiguchi } else { 329234ee645eSJoerg Roedel if (cow) { 32930f10851eSJérôme Glisse /* 32940f10851eSJérôme Glisse * No need to notify as we are downgrading page 32950f10851eSJérôme Glisse * table protection not changing it to point 32960f10851eSJérôme Glisse * to a new page. 32970f10851eSJérôme Glisse * 32980f10851eSJérôme Glisse * See Documentation/vm/mmu_notifier.txt 32990f10851eSJérôme Glisse */ 33007f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 330134ee645eSJoerg Roedel } 33020253d634SNaoya Horiguchi entry = huge_ptep_get(src_pte); 330363551ae0SDavid Gibson ptepage = pte_page(entry); 330463551ae0SDavid Gibson get_page(ptepage); 330553f9263bSKirill A. Shutemov page_dup_rmap(ptepage, true); 330663551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 33075d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), dst); 33081c59827dSHugh Dickins } 3309cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 3310cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 331163551ae0SDavid Gibson } 331263551ae0SDavid Gibson 3313e8569dd2SAndreas Sandberg if (cow) 3314e8569dd2SAndreas Sandberg mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end); 3315e8569dd2SAndreas Sandberg 3316e8569dd2SAndreas Sandberg return ret; 331763551ae0SDavid Gibson } 331863551ae0SDavid Gibson 331924669e58SAneesh Kumar K.V void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 332024669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 332124669e58SAneesh Kumar K.V struct page *ref_page) 332263551ae0SDavid Gibson { 332363551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 332463551ae0SDavid Gibson unsigned long address; 3325c7546f8fSDavid Gibson pte_t *ptep; 332663551ae0SDavid Gibson pte_t pte; 3327cb900f41SKirill A. Shutemov spinlock_t *ptl; 332863551ae0SDavid Gibson struct page *page; 3329a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3330a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 33312ec74c3eSSagi Grimberg const unsigned long mmun_start = start; /* For mmu_notifiers */ 33322ec74c3eSSagi Grimberg const unsigned long mmun_end = end; /* For mmu_notifiers */ 3333a5516438SAndi Kleen 333463551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 3335a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 3336a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 333763551ae0SDavid Gibson 333807e32661SAneesh Kumar K.V /* 333907e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 334007e32661SAneesh Kumar K.V * to huge page. 334107e32661SAneesh Kumar K.V */ 334207e32661SAneesh Kumar K.V tlb_remove_check_page_size_change(tlb, sz); 334324669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 33442ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 3345569f48b8SHillf Danton address = start; 3346569f48b8SHillf Danton for (; address < end; address += sz) { 33477868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, sz); 3348c7546f8fSDavid Gibson if (!ptep) 3349c7546f8fSDavid Gibson continue; 3350c7546f8fSDavid Gibson 3351cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 335231d49da5SAneesh Kumar K.V if (huge_pmd_unshare(mm, &address, ptep)) { 335331d49da5SAneesh Kumar K.V spin_unlock(ptl); 335431d49da5SAneesh Kumar K.V continue; 335531d49da5SAneesh Kumar K.V } 335639dde65cSChen, Kenneth W 33576629326bSHillf Danton pte = huge_ptep_get(ptep); 335831d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 335931d49da5SAneesh Kumar K.V spin_unlock(ptl); 336031d49da5SAneesh Kumar K.V continue; 336131d49da5SAneesh Kumar K.V } 33626629326bSHillf Danton 33636629326bSHillf Danton /* 33649fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 33659fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 33666629326bSHillf Danton */ 33679fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 33689386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 336931d49da5SAneesh Kumar K.V spin_unlock(ptl); 337031d49da5SAneesh Kumar K.V continue; 33718c4894c6SNaoya Horiguchi } 33726629326bSHillf Danton 33736629326bSHillf Danton page = pte_page(pte); 337404f2cbe3SMel Gorman /* 337504f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 337604f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 337704f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 337804f2cbe3SMel Gorman */ 337904f2cbe3SMel Gorman if (ref_page) { 338031d49da5SAneesh Kumar K.V if (page != ref_page) { 338131d49da5SAneesh Kumar K.V spin_unlock(ptl); 338231d49da5SAneesh Kumar K.V continue; 338331d49da5SAneesh Kumar K.V } 338404f2cbe3SMel Gorman /* 338504f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 338604f2cbe3SMel Gorman * future faults in this VMA will fail rather than 338704f2cbe3SMel Gorman * looking like data was lost 338804f2cbe3SMel Gorman */ 338904f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 339004f2cbe3SMel Gorman } 339104f2cbe3SMel Gorman 3392c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 3393b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 3394106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 33956649a386SKen Chen set_page_dirty(page); 33969e81130bSHillf Danton 33975d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 3398d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 339931d49da5SAneesh Kumar K.V 3400cb900f41SKirill A. Shutemov spin_unlock(ptl); 3401e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 340224669e58SAneesh Kumar K.V /* 340331d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 340424669e58SAneesh Kumar K.V */ 340531d49da5SAneesh Kumar K.V if (ref_page) 340631d49da5SAneesh Kumar K.V break; 3407fe1668aeSChen, Kenneth W } 34082ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 340924669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 34101da177e4SLinus Torvalds } 341163551ae0SDavid Gibson 3412d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 3413d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 3414d833352aSMel Gorman unsigned long end, struct page *ref_page) 3415d833352aSMel Gorman { 3416d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 3417d833352aSMel Gorman 3418d833352aSMel Gorman /* 3419d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 3420d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 3421d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 3422d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 3423d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 3424c8c06efaSDavidlohr Bueso * is to clear it before releasing the i_mmap_rwsem. This works 3425d833352aSMel Gorman * because in the context this is called, the VMA is about to be 3426c8c06efaSDavidlohr Bueso * destroyed and the i_mmap_rwsem is held. 3427d833352aSMel Gorman */ 3428d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 3429d833352aSMel Gorman } 3430d833352aSMel Gorman 3431502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 343204f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 3433502717f4SChen, Kenneth W { 343424669e58SAneesh Kumar K.V struct mm_struct *mm; 343524669e58SAneesh Kumar K.V struct mmu_gather tlb; 343624669e58SAneesh Kumar K.V 343724669e58SAneesh Kumar K.V mm = vma->vm_mm; 343824669e58SAneesh Kumar K.V 34392b047252SLinus Torvalds tlb_gather_mmu(&tlb, mm, start, end); 344024669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 344124669e58SAneesh Kumar K.V tlb_finish_mmu(&tlb, start, end); 3442502717f4SChen, Kenneth W } 3443502717f4SChen, Kenneth W 344404f2cbe3SMel Gorman /* 344504f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 344604f2cbe3SMel Gorman * mappping it owns the reserve page for. The intention is to unmap the page 344704f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 344804f2cbe3SMel Gorman * same region. 344904f2cbe3SMel Gorman */ 34502f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 34512a4b3dedSHarvey Harrison struct page *page, unsigned long address) 345204f2cbe3SMel Gorman { 34537526674dSAdam Litke struct hstate *h = hstate_vma(vma); 345404f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 345504f2cbe3SMel Gorman struct address_space *mapping; 345604f2cbe3SMel Gorman pgoff_t pgoff; 345704f2cbe3SMel Gorman 345804f2cbe3SMel Gorman /* 345904f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 346004f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 346104f2cbe3SMel Gorman */ 34627526674dSAdam Litke address = address & huge_page_mask(h); 346336e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 346436e4f20aSMichal Hocko vma->vm_pgoff; 346593c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 346604f2cbe3SMel Gorman 34674eb2b1dcSMel Gorman /* 34684eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 34694eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 34704eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 34714eb2b1dcSMel Gorman */ 347283cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 34736b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 347404f2cbe3SMel Gorman /* Do not unmap the current VMA */ 347504f2cbe3SMel Gorman if (iter_vma == vma) 347604f2cbe3SMel Gorman continue; 347704f2cbe3SMel Gorman 347804f2cbe3SMel Gorman /* 34792f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 34802f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 34812f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 34822f84a899SMel Gorman */ 34832f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 34842f84a899SMel Gorman continue; 34852f84a899SMel Gorman 34862f84a899SMel Gorman /* 348704f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 348804f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 348904f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 349004f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 349104f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 349204f2cbe3SMel Gorman */ 349304f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 349424669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 349524669e58SAneesh Kumar K.V address + huge_page_size(h), page); 349604f2cbe3SMel Gorman } 349783cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 349804f2cbe3SMel Gorman } 349904f2cbe3SMel Gorman 35000fe6e20bSNaoya Horiguchi /* 35010fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 3502ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 3503ef009b25SMichal Hocko * cannot race with other handlers or page migration. 3504ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 35050fe6e20bSNaoya Horiguchi */ 35061e8f889bSDavid Gibson static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 35073999f52eSAneesh Kumar K.V unsigned long address, pte_t *ptep, 3508cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 35091e8f889bSDavid Gibson { 35103999f52eSAneesh Kumar K.V pte_t pte; 3511a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 35121e8f889bSDavid Gibson struct page *old_page, *new_page; 3513ad4404a2SDavidlohr Bueso int ret = 0, outside_reserve = 0; 35142ec74c3eSSagi Grimberg unsigned long mmun_start; /* For mmu_notifiers */ 35152ec74c3eSSagi Grimberg unsigned long mmun_end; /* For mmu_notifiers */ 35161e8f889bSDavid Gibson 35173999f52eSAneesh Kumar K.V pte = huge_ptep_get(ptep); 35181e8f889bSDavid Gibson old_page = pte_page(pte); 35191e8f889bSDavid Gibson 352004f2cbe3SMel Gorman retry_avoidcopy: 35211e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 35221e8f889bSDavid Gibson * and just make the page writable */ 352337a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 35245a49973dSHugh Dickins page_move_anon_rmap(old_page, vma); 35251e8f889bSDavid Gibson set_huge_ptep_writable(vma, address, ptep); 352683c54070SNick Piggin return 0; 35271e8f889bSDavid Gibson } 35281e8f889bSDavid Gibson 352904f2cbe3SMel Gorman /* 353004f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 353104f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 353204f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 353304f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 353404f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 353504f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 353604f2cbe3SMel Gorman * of the full address range. 353704f2cbe3SMel Gorman */ 35385944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 353904f2cbe3SMel Gorman old_page != pagecache_page) 354004f2cbe3SMel Gorman outside_reserve = 1; 354104f2cbe3SMel Gorman 354209cbfeafSKirill A. Shutemov get_page(old_page); 3543b76c8cfbSLarry Woodman 3544ad4404a2SDavidlohr Bueso /* 3545ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 3546ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 3547ad4404a2SDavidlohr Bueso */ 3548cb900f41SKirill A. Shutemov spin_unlock(ptl); 354904f2cbe3SMel Gorman new_page = alloc_huge_page(vma, address, outside_reserve); 35501e8f889bSDavid Gibson 35512fc39cecSAdam Litke if (IS_ERR(new_page)) { 355204f2cbe3SMel Gorman /* 355304f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 355404f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 355504f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 355604f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 355704f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 355804f2cbe3SMel Gorman */ 355904f2cbe3SMel Gorman if (outside_reserve) { 356009cbfeafSKirill A. Shutemov put_page(old_page); 356104f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 35622f4612afSDavidlohr Bueso unmap_ref_private(mm, vma, old_page, address); 356304f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 3564cb900f41SKirill A. Shutemov spin_lock(ptl); 35657868a208SPunit Agrawal ptep = huge_pte_offset(mm, address & huge_page_mask(h), 35667868a208SPunit Agrawal huge_page_size(h)); 3567a9af0c5dSNaoya Horiguchi if (likely(ptep && 3568a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 356904f2cbe3SMel Gorman goto retry_avoidcopy; 3570a734bcc8SHillf Danton /* 3571cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 3572cb900f41SKirill A. Shutemov * lock, and our job is done. 3573a734bcc8SHillf Danton */ 3574a734bcc8SHillf Danton return 0; 357504f2cbe3SMel Gorman } 357604f2cbe3SMel Gorman 3577ad4404a2SDavidlohr Bueso ret = (PTR_ERR(new_page) == -ENOMEM) ? 3578ad4404a2SDavidlohr Bueso VM_FAULT_OOM : VM_FAULT_SIGBUS; 3579ad4404a2SDavidlohr Bueso goto out_release_old; 35801e8f889bSDavid Gibson } 35811e8f889bSDavid Gibson 35820fe6e20bSNaoya Horiguchi /* 35830fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 35840fe6e20bSNaoya Horiguchi * anon_vma prepared. 35850fe6e20bSNaoya Horiguchi */ 358644e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 3587ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 3588ad4404a2SDavidlohr Bueso goto out_release_all; 358944e2aa93SDean Nelson } 35900fe6e20bSNaoya Horiguchi 359147ad8475SAndrea Arcangeli copy_user_huge_page(new_page, old_page, address, vma, 359247ad8475SAndrea Arcangeli pages_per_huge_page(h)); 35930ed361deSNick Piggin __SetPageUptodate(new_page); 3594bcc54222SNaoya Horiguchi set_page_huge_active(new_page); 35951e8f889bSDavid Gibson 35962ec74c3eSSagi Grimberg mmun_start = address & huge_page_mask(h); 35972ec74c3eSSagi Grimberg mmun_end = mmun_start + huge_page_size(h); 35982ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 3599ad4404a2SDavidlohr Bueso 3600b76c8cfbSLarry Woodman /* 3601cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 3602b76c8cfbSLarry Woodman * before the page tables are altered 3603b76c8cfbSLarry Woodman */ 3604cb900f41SKirill A. Shutemov spin_lock(ptl); 36057868a208SPunit Agrawal ptep = huge_pte_offset(mm, address & huge_page_mask(h), 36067868a208SPunit Agrawal huge_page_size(h)); 3607a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 360807443a85SJoonsoo Kim ClearPagePrivate(new_page); 360907443a85SJoonsoo Kim 36101e8f889bSDavid Gibson /* Break COW */ 36118fe627ecSGerald Schaefer huge_ptep_clear_flush(vma, address, ptep); 361234ee645eSJoerg Roedel mmu_notifier_invalidate_range(mm, mmun_start, mmun_end); 36131e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, 36141e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 3615d281ee61SKirill A. Shutemov page_remove_rmap(old_page, true); 3616cd67f0d2SNaoya Horiguchi hugepage_add_new_anon_rmap(new_page, vma, address); 36171e8f889bSDavid Gibson /* Make the old page be freed below */ 36181e8f889bSDavid Gibson new_page = old_page; 36191e8f889bSDavid Gibson } 3620cb900f41SKirill A. Shutemov spin_unlock(ptl); 36212ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 3622ad4404a2SDavidlohr Bueso out_release_all: 362396b96a96SMike Kravetz restore_reserve_on_error(h, vma, address, new_page); 362409cbfeafSKirill A. Shutemov put_page(new_page); 3625ad4404a2SDavidlohr Bueso out_release_old: 362609cbfeafSKirill A. Shutemov put_page(old_page); 36278312034fSJoonsoo Kim 3628ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 3629ad4404a2SDavidlohr Bueso return ret; 36301e8f889bSDavid Gibson } 36311e8f889bSDavid Gibson 363204f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 3633a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 3634a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 363504f2cbe3SMel Gorman { 363604f2cbe3SMel Gorman struct address_space *mapping; 3637e7c4b0bfSAndy Whitcroft pgoff_t idx; 363804f2cbe3SMel Gorman 363904f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 3640a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 364104f2cbe3SMel Gorman 364204f2cbe3SMel Gorman return find_lock_page(mapping, idx); 364304f2cbe3SMel Gorman } 364404f2cbe3SMel Gorman 36453ae77f43SHugh Dickins /* 36463ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 36473ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 36483ae77f43SHugh Dickins */ 36493ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 36502a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 36512a15efc9SHugh Dickins { 36522a15efc9SHugh Dickins struct address_space *mapping; 36532a15efc9SHugh Dickins pgoff_t idx; 36542a15efc9SHugh Dickins struct page *page; 36552a15efc9SHugh Dickins 36562a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 36572a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 36582a15efc9SHugh Dickins 36592a15efc9SHugh Dickins page = find_get_page(mapping, idx); 36602a15efc9SHugh Dickins if (page) 36612a15efc9SHugh Dickins put_page(page); 36622a15efc9SHugh Dickins return page != NULL; 36632a15efc9SHugh Dickins } 36642a15efc9SHugh Dickins 3665ab76ad54SMike Kravetz int huge_add_to_page_cache(struct page *page, struct address_space *mapping, 3666ab76ad54SMike Kravetz pgoff_t idx) 3667ab76ad54SMike Kravetz { 3668ab76ad54SMike Kravetz struct inode *inode = mapping->host; 3669ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 3670ab76ad54SMike Kravetz int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 3671ab76ad54SMike Kravetz 3672ab76ad54SMike Kravetz if (err) 3673ab76ad54SMike Kravetz return err; 3674ab76ad54SMike Kravetz ClearPagePrivate(page); 3675ab76ad54SMike Kravetz 3676ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 3677ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 3678ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 3679ab76ad54SMike Kravetz return 0; 3680ab76ad54SMike Kravetz } 3681ab76ad54SMike Kravetz 3682a1ed3ddaSRobert P. J. Day static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, 36838382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 3684788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 3685ac9b9c66SHugh Dickins { 3686a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3687ac9b9c66SHugh Dickins int ret = VM_FAULT_SIGBUS; 3688409eb8c2SHillf Danton int anon_rmap = 0; 36894c887265SAdam Litke unsigned long size; 36904c887265SAdam Litke struct page *page; 36911e8f889bSDavid Gibson pte_t new_pte; 3692cb900f41SKirill A. Shutemov spinlock_t *ptl; 36934c887265SAdam Litke 369404f2cbe3SMel Gorman /* 369504f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 369604f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 369725985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 369804f2cbe3SMel Gorman */ 369904f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 3700910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 370104f2cbe3SMel Gorman current->pid); 370204f2cbe3SMel Gorman return ret; 370304f2cbe3SMel Gorman } 370404f2cbe3SMel Gorman 37054c887265SAdam Litke /* 37064c887265SAdam Litke * Use page lock to guard against racing truncation 37074c887265SAdam Litke * before we get page_table_lock. 37084c887265SAdam Litke */ 37096bda666aSChristoph Lameter retry: 37106bda666aSChristoph Lameter page = find_lock_page(mapping, idx); 37116bda666aSChristoph Lameter if (!page) { 3712a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 3713ebed4bfcSHugh Dickins if (idx >= size) 3714ebed4bfcSHugh Dickins goto out; 37151a1aad8aSMike Kravetz 37161a1aad8aSMike Kravetz /* 37171a1aad8aSMike Kravetz * Check for page in userfault range 37181a1aad8aSMike Kravetz */ 37191a1aad8aSMike Kravetz if (userfaultfd_missing(vma)) { 37201a1aad8aSMike Kravetz u32 hash; 37211a1aad8aSMike Kravetz struct vm_fault vmf = { 37221a1aad8aSMike Kravetz .vma = vma, 37231a1aad8aSMike Kravetz .address = address, 37241a1aad8aSMike Kravetz .flags = flags, 37251a1aad8aSMike Kravetz /* 37261a1aad8aSMike Kravetz * Hard to debug if it ends up being 37271a1aad8aSMike Kravetz * used by a callee that assumes 37281a1aad8aSMike Kravetz * something about the other 37291a1aad8aSMike Kravetz * uninitialized fields... same as in 37301a1aad8aSMike Kravetz * memory.c 37311a1aad8aSMike Kravetz */ 37321a1aad8aSMike Kravetz }; 37331a1aad8aSMike Kravetz 37341a1aad8aSMike Kravetz /* 37351a1aad8aSMike Kravetz * hugetlb_fault_mutex must be dropped before 37361a1aad8aSMike Kravetz * handling userfault. Reacquire after handling 37371a1aad8aSMike Kravetz * fault to make calling code simpler. 37381a1aad8aSMike Kravetz */ 37391a1aad8aSMike Kravetz hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, 37401a1aad8aSMike Kravetz idx, address); 37411a1aad8aSMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 37421a1aad8aSMike Kravetz ret = handle_userfault(&vmf, VM_UFFD_MISSING); 37431a1aad8aSMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 37441a1aad8aSMike Kravetz goto out; 37451a1aad8aSMike Kravetz } 37461a1aad8aSMike Kravetz 374704f2cbe3SMel Gorman page = alloc_huge_page(vma, address, 0); 37482fc39cecSAdam Litke if (IS_ERR(page)) { 374976dcee75SAneesh Kumar K.V ret = PTR_ERR(page); 375076dcee75SAneesh Kumar K.V if (ret == -ENOMEM) 375176dcee75SAneesh Kumar K.V ret = VM_FAULT_OOM; 375276dcee75SAneesh Kumar K.V else 375376dcee75SAneesh Kumar K.V ret = VM_FAULT_SIGBUS; 37546bda666aSChristoph Lameter goto out; 37556bda666aSChristoph Lameter } 375647ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 37570ed361deSNick Piggin __SetPageUptodate(page); 3758bcc54222SNaoya Horiguchi set_page_huge_active(page); 3759ac9b9c66SHugh Dickins 3760f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 3761ab76ad54SMike Kravetz int err = huge_add_to_page_cache(page, mapping, idx); 37626bda666aSChristoph Lameter if (err) { 37636bda666aSChristoph Lameter put_page(page); 37646bda666aSChristoph Lameter if (err == -EEXIST) 37656bda666aSChristoph Lameter goto retry; 37666bda666aSChristoph Lameter goto out; 37676bda666aSChristoph Lameter } 376823be7468SMel Gorman } else { 37696bda666aSChristoph Lameter lock_page(page); 37700fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 37710fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 37720fe6e20bSNaoya Horiguchi goto backout_unlocked; 377323be7468SMel Gorman } 3774409eb8c2SHillf Danton anon_rmap = 1; 37750fe6e20bSNaoya Horiguchi } 37760fe6e20bSNaoya Horiguchi } else { 377757303d80SAndy Whitcroft /* 3778998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 3779998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 3780998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 3781fd6a03edSNaoya Horiguchi */ 3782fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 3783aa50d3a7SAndi Kleen ret = VM_FAULT_HWPOISON | 3784972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 3785fd6a03edSNaoya Horiguchi goto backout_unlocked; 37866bda666aSChristoph Lameter } 3787998b4382SNaoya Horiguchi } 37881e8f889bSDavid Gibson 378957303d80SAndy Whitcroft /* 379057303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 379157303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 379257303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 379357303d80SAndy Whitcroft * the spinlock. 379457303d80SAndy Whitcroft */ 37955e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 37962b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 37972b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 37982b26736cSAndy Whitcroft goto backout_unlocked; 37992b26736cSAndy Whitcroft } 38005e911373SMike Kravetz /* Just decrements count, does not deallocate */ 3801feba16e2SMike Kravetz vma_end_reservation(h, vma, address); 38025e911373SMike Kravetz } 380357303d80SAndy Whitcroft 38048bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 3805a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 38064c887265SAdam Litke if (idx >= size) 38074c887265SAdam Litke goto backout; 38084c887265SAdam Litke 380983c54070SNick Piggin ret = 0; 38107f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 38114c887265SAdam Litke goto backout; 38124c887265SAdam Litke 381307443a85SJoonsoo Kim if (anon_rmap) { 381407443a85SJoonsoo Kim ClearPagePrivate(page); 3815409eb8c2SHillf Danton hugepage_add_new_anon_rmap(page, vma, address); 3816ac714904SChoi Gi-yong } else 381753f9263bSKirill A. Shutemov page_dup_rmap(page, true); 38181e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 38191e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 38201e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, new_pte); 38211e8f889bSDavid Gibson 38225d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 3823788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 38241e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 38253999f52eSAneesh Kumar K.V ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); 38261e8f889bSDavid Gibson } 38271e8f889bSDavid Gibson 3828cb900f41SKirill A. Shutemov spin_unlock(ptl); 38294c887265SAdam Litke unlock_page(page); 38304c887265SAdam Litke out: 3831ac9b9c66SHugh Dickins return ret; 38324c887265SAdam Litke 38334c887265SAdam Litke backout: 3834cb900f41SKirill A. Shutemov spin_unlock(ptl); 38352b26736cSAndy Whitcroft backout_unlocked: 38364c887265SAdam Litke unlock_page(page); 383796b96a96SMike Kravetz restore_reserve_on_error(h, vma, address, page); 38384c887265SAdam Litke put_page(page); 38394c887265SAdam Litke goto out; 3840ac9b9c66SHugh Dickins } 3841ac9b9c66SHugh Dickins 38428382d914SDavidlohr Bueso #ifdef CONFIG_SMP 3843c672c7f2SMike Kravetz u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, 38448382d914SDavidlohr Bueso struct vm_area_struct *vma, 38458382d914SDavidlohr Bueso struct address_space *mapping, 38468382d914SDavidlohr Bueso pgoff_t idx, unsigned long address) 38478382d914SDavidlohr Bueso { 38488382d914SDavidlohr Bueso unsigned long key[2]; 38498382d914SDavidlohr Bueso u32 hash; 38508382d914SDavidlohr Bueso 38518382d914SDavidlohr Bueso if (vma->vm_flags & VM_SHARED) { 38528382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 38538382d914SDavidlohr Bueso key[1] = idx; 38548382d914SDavidlohr Bueso } else { 38558382d914SDavidlohr Bueso key[0] = (unsigned long) mm; 38568382d914SDavidlohr Bueso key[1] = address >> huge_page_shift(h); 38578382d914SDavidlohr Bueso } 38588382d914SDavidlohr Bueso 38598382d914SDavidlohr Bueso hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0); 38608382d914SDavidlohr Bueso 38618382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 38628382d914SDavidlohr Bueso } 38638382d914SDavidlohr Bueso #else 38648382d914SDavidlohr Bueso /* 38658382d914SDavidlohr Bueso * For uniprocesor systems we always use a single mutex, so just 38668382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 38678382d914SDavidlohr Bueso */ 3868c672c7f2SMike Kravetz u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, 38698382d914SDavidlohr Bueso struct vm_area_struct *vma, 38708382d914SDavidlohr Bueso struct address_space *mapping, 38718382d914SDavidlohr Bueso pgoff_t idx, unsigned long address) 38728382d914SDavidlohr Bueso { 38738382d914SDavidlohr Bueso return 0; 38748382d914SDavidlohr Bueso } 38758382d914SDavidlohr Bueso #endif 38768382d914SDavidlohr Bueso 387786e5216fSAdam Litke int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 3878788c7df4SHugh Dickins unsigned long address, unsigned int flags) 387986e5216fSAdam Litke { 38808382d914SDavidlohr Bueso pte_t *ptep, entry; 3881cb900f41SKirill A. Shutemov spinlock_t *ptl; 38821e8f889bSDavid Gibson int ret; 38838382d914SDavidlohr Bueso u32 hash; 38848382d914SDavidlohr Bueso pgoff_t idx; 38850fe6e20bSNaoya Horiguchi struct page *page = NULL; 388657303d80SAndy Whitcroft struct page *pagecache_page = NULL; 3887a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 38888382d914SDavidlohr Bueso struct address_space *mapping; 38890f792cf9SNaoya Horiguchi int need_wait_lock = 0; 389086e5216fSAdam Litke 38911e16a539SKAMEZAWA Hiroyuki address &= huge_page_mask(h); 38921e16a539SKAMEZAWA Hiroyuki 38937868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, huge_page_size(h)); 3894fd6a03edSNaoya Horiguchi if (ptep) { 3895fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 3896290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 3897cb900f41SKirill A. Shutemov migration_entry_wait_huge(vma, mm, ptep); 3898290408d4SNaoya Horiguchi return 0; 3899290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 3900aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 3901972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 39020d777df5SNaoya Horiguchi } else { 3903a5516438SAndi Kleen ptep = huge_pte_alloc(mm, address, huge_page_size(h)); 390486e5216fSAdam Litke if (!ptep) 390586e5216fSAdam Litke return VM_FAULT_OOM; 39060d777df5SNaoya Horiguchi } 390786e5216fSAdam Litke 39088382d914SDavidlohr Bueso mapping = vma->vm_file->f_mapping; 39098382d914SDavidlohr Bueso idx = vma_hugecache_offset(h, vma, address); 39108382d914SDavidlohr Bueso 39113935baa9SDavid Gibson /* 39123935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 39133935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 39143935baa9SDavid Gibson * the same page in the page cache. 39153935baa9SDavid Gibson */ 3916c672c7f2SMike Kravetz hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address); 3917c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 39188382d914SDavidlohr Bueso 39197f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 39207f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 39218382d914SDavidlohr Bueso ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 3922b4d1d99fSDavid Gibson goto out_mutex; 39233935baa9SDavid Gibson } 392486e5216fSAdam Litke 392583c54070SNick Piggin ret = 0; 39261e8f889bSDavid Gibson 392757303d80SAndy Whitcroft /* 39280f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 39290f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 39300f792cf9SNaoya Horiguchi * a active hugepage in pagecache. This goto expects the 2nd page fault, 39310f792cf9SNaoya Horiguchi * and is_hugetlb_entry_(migration|hwpoisoned) check will properly 39320f792cf9SNaoya Horiguchi * handle it. 39330f792cf9SNaoya Horiguchi */ 39340f792cf9SNaoya Horiguchi if (!pte_present(entry)) 39350f792cf9SNaoya Horiguchi goto out_mutex; 39360f792cf9SNaoya Horiguchi 39370f792cf9SNaoya Horiguchi /* 393857303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 393957303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 394057303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 394157303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 394257303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 394357303d80SAndy Whitcroft * consumed. 394457303d80SAndy Whitcroft */ 3945106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 39462b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 39472b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 3948b4d1d99fSDavid Gibson goto out_mutex; 39492b26736cSAndy Whitcroft } 39505e911373SMike Kravetz /* Just decrements count, does not deallocate */ 3951feba16e2SMike Kravetz vma_end_reservation(h, vma, address); 395257303d80SAndy Whitcroft 3953f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 395457303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 395557303d80SAndy Whitcroft vma, address); 395657303d80SAndy Whitcroft } 395757303d80SAndy Whitcroft 39580f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 39590fe6e20bSNaoya Horiguchi 39601e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 3961b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 3962cb900f41SKirill A. Shutemov goto out_ptl; 3963b4d1d99fSDavid Gibson 39640f792cf9SNaoya Horiguchi /* 39650f792cf9SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 39660f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 39670f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 39680f792cf9SNaoya Horiguchi */ 39690f792cf9SNaoya Horiguchi page = pte_page(entry); 39700f792cf9SNaoya Horiguchi if (page != pagecache_page) 39710f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 39720f792cf9SNaoya Horiguchi need_wait_lock = 1; 39730f792cf9SNaoya Horiguchi goto out_ptl; 39740f792cf9SNaoya Horiguchi } 39750f792cf9SNaoya Horiguchi 39760f792cf9SNaoya Horiguchi get_page(page); 3977b4d1d99fSDavid Gibson 3978788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 3979106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 39803999f52eSAneesh Kumar K.V ret = hugetlb_cow(mm, vma, address, ptep, 3981cb900f41SKirill A. Shutemov pagecache_page, ptl); 39820f792cf9SNaoya Horiguchi goto out_put_page; 3983b4d1d99fSDavid Gibson } 3984106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 3985b4d1d99fSDavid Gibson } 3986b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 3987788c7df4SHugh Dickins if (huge_ptep_set_access_flags(vma, address, ptep, entry, 3988788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 39894b3073e1SRussell King update_mmu_cache(vma, address, ptep); 39900f792cf9SNaoya Horiguchi out_put_page: 39910f792cf9SNaoya Horiguchi if (page != pagecache_page) 39920f792cf9SNaoya Horiguchi unlock_page(page); 39930f792cf9SNaoya Horiguchi put_page(page); 3994cb900f41SKirill A. Shutemov out_ptl: 3995cb900f41SKirill A. Shutemov spin_unlock(ptl); 399657303d80SAndy Whitcroft 399757303d80SAndy Whitcroft if (pagecache_page) { 399857303d80SAndy Whitcroft unlock_page(pagecache_page); 399957303d80SAndy Whitcroft put_page(pagecache_page); 400057303d80SAndy Whitcroft } 4001b4d1d99fSDavid Gibson out_mutex: 4002c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 40030f792cf9SNaoya Horiguchi /* 40040f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 40050f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 40060f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 40070f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 40080f792cf9SNaoya Horiguchi * here without taking refcount. 40090f792cf9SNaoya Horiguchi */ 40100f792cf9SNaoya Horiguchi if (need_wait_lock) 40110f792cf9SNaoya Horiguchi wait_on_page_locked(page); 40121e8f889bSDavid Gibson return ret; 401386e5216fSAdam Litke } 401486e5216fSAdam Litke 40158fb5debcSMike Kravetz /* 40168fb5debcSMike Kravetz * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with 40178fb5debcSMike Kravetz * modifications for huge pages. 40188fb5debcSMike Kravetz */ 40198fb5debcSMike Kravetz int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, 40208fb5debcSMike Kravetz pte_t *dst_pte, 40218fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 40228fb5debcSMike Kravetz unsigned long dst_addr, 40238fb5debcSMike Kravetz unsigned long src_addr, 40248fb5debcSMike Kravetz struct page **pagep) 40258fb5debcSMike Kravetz { 40261e392147SAndrea Arcangeli struct address_space *mapping; 40271e392147SAndrea Arcangeli pgoff_t idx; 40281e392147SAndrea Arcangeli unsigned long size; 40291c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 40308fb5debcSMike Kravetz struct hstate *h = hstate_vma(dst_vma); 40318fb5debcSMike Kravetz pte_t _dst_pte; 40328fb5debcSMike Kravetz spinlock_t *ptl; 40338fb5debcSMike Kravetz int ret; 40348fb5debcSMike Kravetz struct page *page; 40358fb5debcSMike Kravetz 40368fb5debcSMike Kravetz if (!*pagep) { 40378fb5debcSMike Kravetz ret = -ENOMEM; 40388fb5debcSMike Kravetz page = alloc_huge_page(dst_vma, dst_addr, 0); 40398fb5debcSMike Kravetz if (IS_ERR(page)) 40408fb5debcSMike Kravetz goto out; 40418fb5debcSMike Kravetz 40428fb5debcSMike Kravetz ret = copy_huge_page_from_user(page, 40438fb5debcSMike Kravetz (const void __user *) src_addr, 4044810a56b9SMike Kravetz pages_per_huge_page(h), false); 40458fb5debcSMike Kravetz 40468fb5debcSMike Kravetz /* fallback to copy_from_user outside mmap_sem */ 40478fb5debcSMike Kravetz if (unlikely(ret)) { 40488fb5debcSMike Kravetz ret = -EFAULT; 40498fb5debcSMike Kravetz *pagep = page; 40508fb5debcSMike Kravetz /* don't free the page */ 40518fb5debcSMike Kravetz goto out; 40528fb5debcSMike Kravetz } 40538fb5debcSMike Kravetz } else { 40548fb5debcSMike Kravetz page = *pagep; 40558fb5debcSMike Kravetz *pagep = NULL; 40568fb5debcSMike Kravetz } 40578fb5debcSMike Kravetz 40588fb5debcSMike Kravetz /* 40598fb5debcSMike Kravetz * The memory barrier inside __SetPageUptodate makes sure that 40608fb5debcSMike Kravetz * preceding stores to the page contents become visible before 40618fb5debcSMike Kravetz * the set_pte_at() write. 40628fb5debcSMike Kravetz */ 40638fb5debcSMike Kravetz __SetPageUptodate(page); 40648fb5debcSMike Kravetz set_page_huge_active(page); 40658fb5debcSMike Kravetz 40661e392147SAndrea Arcangeli mapping = dst_vma->vm_file->f_mapping; 40671e392147SAndrea Arcangeli idx = vma_hugecache_offset(h, dst_vma, dst_addr); 40681e392147SAndrea Arcangeli 40691c9e8defSMike Kravetz /* 40701c9e8defSMike Kravetz * If shared, add to page cache 40711c9e8defSMike Kravetz */ 40721c9e8defSMike Kravetz if (vm_shared) { 40731e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 40741e392147SAndrea Arcangeli ret = -EFAULT; 40751e392147SAndrea Arcangeli if (idx >= size) 40761e392147SAndrea Arcangeli goto out_release_nounlock; 40771c9e8defSMike Kravetz 40781e392147SAndrea Arcangeli /* 40791e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 40801e392147SAndrea Arcangeli * huge_add_to_page_cache() below happens through the 40811e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 40821e392147SAndrea Arcangeli * the caller. 40831e392147SAndrea Arcangeli */ 40841c9e8defSMike Kravetz ret = huge_add_to_page_cache(page, mapping, idx); 40851c9e8defSMike Kravetz if (ret) 40861c9e8defSMike Kravetz goto out_release_nounlock; 40871c9e8defSMike Kravetz } 40881c9e8defSMike Kravetz 40898fb5debcSMike Kravetz ptl = huge_pte_lockptr(h, dst_mm, dst_pte); 40908fb5debcSMike Kravetz spin_lock(ptl); 40918fb5debcSMike Kravetz 40921e392147SAndrea Arcangeli /* 40931e392147SAndrea Arcangeli * Recheck the i_size after holding PT lock to make sure not 40941e392147SAndrea Arcangeli * to leave any page mapped (as page_mapped()) beyond the end 40951e392147SAndrea Arcangeli * of the i_size (remove_inode_hugepages() is strict about 40961e392147SAndrea Arcangeli * enforcing that). If we bail out here, we'll also leave a 40971e392147SAndrea Arcangeli * page in the radix tree in the vm_shared case beyond the end 40981e392147SAndrea Arcangeli * of the i_size, but remove_inode_hugepages() will take care 40991e392147SAndrea Arcangeli * of it as soon as we drop the hugetlb_fault_mutex_table. 41001e392147SAndrea Arcangeli */ 41011e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 41021e392147SAndrea Arcangeli ret = -EFAULT; 41031e392147SAndrea Arcangeli if (idx >= size) 41041e392147SAndrea Arcangeli goto out_release_unlock; 41051e392147SAndrea Arcangeli 41068fb5debcSMike Kravetz ret = -EEXIST; 41078fb5debcSMike Kravetz if (!huge_pte_none(huge_ptep_get(dst_pte))) 41088fb5debcSMike Kravetz goto out_release_unlock; 41098fb5debcSMike Kravetz 41101c9e8defSMike Kravetz if (vm_shared) { 41111c9e8defSMike Kravetz page_dup_rmap(page, true); 41121c9e8defSMike Kravetz } else { 41138fb5debcSMike Kravetz ClearPagePrivate(page); 41148fb5debcSMike Kravetz hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); 41151c9e8defSMike Kravetz } 41168fb5debcSMike Kravetz 41178fb5debcSMike Kravetz _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); 41188fb5debcSMike Kravetz if (dst_vma->vm_flags & VM_WRITE) 41198fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 41208fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 41218fb5debcSMike Kravetz 41228fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 41238fb5debcSMike Kravetz 41248fb5debcSMike Kravetz (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, 41258fb5debcSMike Kravetz dst_vma->vm_flags & VM_WRITE); 41268fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 41278fb5debcSMike Kravetz 41288fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 41298fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 41308fb5debcSMike Kravetz 41318fb5debcSMike Kravetz spin_unlock(ptl); 41321c9e8defSMike Kravetz if (vm_shared) 41331c9e8defSMike Kravetz unlock_page(page); 41348fb5debcSMike Kravetz ret = 0; 41358fb5debcSMike Kravetz out: 41368fb5debcSMike Kravetz return ret; 41378fb5debcSMike Kravetz out_release_unlock: 41388fb5debcSMike Kravetz spin_unlock(ptl); 41391c9e8defSMike Kravetz if (vm_shared) 41401c9e8defSMike Kravetz unlock_page(page); 41415af10dfdSAndrea Arcangeli out_release_nounlock: 41428fb5debcSMike Kravetz put_page(page); 41438fb5debcSMike Kravetz goto out; 41448fb5debcSMike Kravetz } 41458fb5debcSMike Kravetz 414628a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 414763551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 414828a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 414987ffc118SAndrea Arcangeli long i, unsigned int flags, int *nonblocking) 415063551ae0SDavid Gibson { 4151d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 4152d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 415328a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 4154a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 41552be7cfedSDaniel Jordan int err = -EFAULT; 415663551ae0SDavid Gibson 415763551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 415863551ae0SDavid Gibson pte_t *pte; 4159cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 41602a15efc9SHugh Dickins int absent; 416163551ae0SDavid Gibson struct page *page; 416263551ae0SDavid Gibson 41634c887265SAdam Litke /* 416402057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 416502057967SDavid Rientjes * potentially allocating memory. 416602057967SDavid Rientjes */ 416702057967SDavid Rientjes if (unlikely(fatal_signal_pending(current))) { 416802057967SDavid Rientjes remainder = 0; 416902057967SDavid Rientjes break; 417002057967SDavid Rientjes } 417102057967SDavid Rientjes 417202057967SDavid Rientjes /* 41734c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 41742a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 41754c887265SAdam Litke * first, for the page indexing below to work. 4176cb900f41SKirill A. Shutemov * 4177cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 41784c887265SAdam Litke */ 41797868a208SPunit Agrawal pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), 41807868a208SPunit Agrawal huge_page_size(h)); 4181cb900f41SKirill A. Shutemov if (pte) 4182cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 41832a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 418463551ae0SDavid Gibson 41852a15efc9SHugh Dickins /* 41862a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 41873ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 41883ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 41893ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 41903ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 41912a15efc9SHugh Dickins */ 41923ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 41933ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 4194cb900f41SKirill A. Shutemov if (pte) 4195cb900f41SKirill A. Shutemov spin_unlock(ptl); 41962a15efc9SHugh Dickins remainder = 0; 41972a15efc9SHugh Dickins break; 41982a15efc9SHugh Dickins } 41992a15efc9SHugh Dickins 42009cc3a5bdSNaoya Horiguchi /* 42019cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 42029cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 42039cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 42049cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 42059cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 42069cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 42079cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 42089cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 42099cc3a5bdSNaoya Horiguchi */ 42109cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 4211106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 4212106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 42134c887265SAdam Litke int ret; 421487ffc118SAndrea Arcangeli unsigned int fault_flags = 0; 42154c887265SAdam Litke 4216cb900f41SKirill A. Shutemov if (pte) 4217cb900f41SKirill A. Shutemov spin_unlock(ptl); 421887ffc118SAndrea Arcangeli if (flags & FOLL_WRITE) 421987ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_WRITE; 422087ffc118SAndrea Arcangeli if (nonblocking) 422187ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY; 422287ffc118SAndrea Arcangeli if (flags & FOLL_NOWAIT) 422387ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY | 422487ffc118SAndrea Arcangeli FAULT_FLAG_RETRY_NOWAIT; 422587ffc118SAndrea Arcangeli if (flags & FOLL_TRIED) { 422687ffc118SAndrea Arcangeli VM_WARN_ON_ONCE(fault_flags & 422787ffc118SAndrea Arcangeli FAULT_FLAG_ALLOW_RETRY); 422887ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_TRIED; 422987ffc118SAndrea Arcangeli } 423087ffc118SAndrea Arcangeli ret = hugetlb_fault(mm, vma, vaddr, fault_flags); 423187ffc118SAndrea Arcangeli if (ret & VM_FAULT_ERROR) { 42322be7cfedSDaniel Jordan err = vm_fault_to_errno(ret, flags); 42331c59827dSHugh Dickins remainder = 0; 42341c59827dSHugh Dickins break; 42351c59827dSHugh Dickins } 423687ffc118SAndrea Arcangeli if (ret & VM_FAULT_RETRY) { 423787ffc118SAndrea Arcangeli if (nonblocking) 423887ffc118SAndrea Arcangeli *nonblocking = 0; 423987ffc118SAndrea Arcangeli *nr_pages = 0; 424087ffc118SAndrea Arcangeli /* 424187ffc118SAndrea Arcangeli * VM_FAULT_RETRY must not return an 424287ffc118SAndrea Arcangeli * error, it will return zero 424387ffc118SAndrea Arcangeli * instead. 424487ffc118SAndrea Arcangeli * 424587ffc118SAndrea Arcangeli * No need to update "position" as the 424687ffc118SAndrea Arcangeli * caller will not check it after 424787ffc118SAndrea Arcangeli * *nr_pages is set to 0. 424887ffc118SAndrea Arcangeli */ 424987ffc118SAndrea Arcangeli return i; 425087ffc118SAndrea Arcangeli } 425187ffc118SAndrea Arcangeli continue; 425287ffc118SAndrea Arcangeli } 425363551ae0SDavid Gibson 4254a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 42557f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 4256d5d4b0aaSChen, Kenneth W same_page: 4257d6692183SChen, Kenneth W if (pages) { 425869d177c2SAndy Whitcroft pages[i] = mem_map_offset(page, pfn_offset); 4259ddc58f27SKirill A. Shutemov get_page(pages[i]); 4260d6692183SChen, Kenneth W } 426163551ae0SDavid Gibson 426263551ae0SDavid Gibson if (vmas) 426363551ae0SDavid Gibson vmas[i] = vma; 426463551ae0SDavid Gibson 426563551ae0SDavid Gibson vaddr += PAGE_SIZE; 4266d5d4b0aaSChen, Kenneth W ++pfn_offset; 426763551ae0SDavid Gibson --remainder; 426863551ae0SDavid Gibson ++i; 4269d5d4b0aaSChen, Kenneth W if (vaddr < vma->vm_end && remainder && 4270a5516438SAndi Kleen pfn_offset < pages_per_huge_page(h)) { 4271d5d4b0aaSChen, Kenneth W /* 4272d5d4b0aaSChen, Kenneth W * We use pfn_offset to avoid touching the pageframes 4273d5d4b0aaSChen, Kenneth W * of this compound page. 4274d5d4b0aaSChen, Kenneth W */ 4275d5d4b0aaSChen, Kenneth W goto same_page; 4276d5d4b0aaSChen, Kenneth W } 4277cb900f41SKirill A. Shutemov spin_unlock(ptl); 427863551ae0SDavid Gibson } 427928a35716SMichel Lespinasse *nr_pages = remainder; 428087ffc118SAndrea Arcangeli /* 428187ffc118SAndrea Arcangeli * setting position is actually required only if remainder is 428287ffc118SAndrea Arcangeli * not zero but it's faster not to add a "if (remainder)" 428387ffc118SAndrea Arcangeli * branch. 428487ffc118SAndrea Arcangeli */ 428563551ae0SDavid Gibson *position = vaddr; 428663551ae0SDavid Gibson 42872be7cfedSDaniel Jordan return i ? i : err; 428863551ae0SDavid Gibson } 42898f860591SZhang, Yanmin 42905491ae7bSAneesh Kumar K.V #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE 42915491ae7bSAneesh Kumar K.V /* 42925491ae7bSAneesh Kumar K.V * ARCHes with special requirements for evicting HUGETLB backing TLB entries can 42935491ae7bSAneesh Kumar K.V * implement this. 42945491ae7bSAneesh Kumar K.V */ 42955491ae7bSAneesh Kumar K.V #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) 42965491ae7bSAneesh Kumar K.V #endif 42975491ae7bSAneesh Kumar K.V 42987da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 42998f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 43008f860591SZhang, Yanmin { 43018f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 43028f860591SZhang, Yanmin unsigned long start = address; 43038f860591SZhang, Yanmin pte_t *ptep; 43048f860591SZhang, Yanmin pte_t pte; 4305a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 43067da4d641SPeter Zijlstra unsigned long pages = 0; 43078f860591SZhang, Yanmin 43088f860591SZhang, Yanmin BUG_ON(address >= end); 43098f860591SZhang, Yanmin flush_cache_range(vma, address, end); 43108f860591SZhang, Yanmin 4311a5338093SRik van Riel mmu_notifier_invalidate_range_start(mm, start, end); 431283cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 4313a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 4314cb900f41SKirill A. Shutemov spinlock_t *ptl; 43157868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, huge_page_size(h)); 43168f860591SZhang, Yanmin if (!ptep) 43178f860591SZhang, Yanmin continue; 4318cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 43197da4d641SPeter Zijlstra if (huge_pmd_unshare(mm, &address, ptep)) { 43207da4d641SPeter Zijlstra pages++; 4321cb900f41SKirill A. Shutemov spin_unlock(ptl); 432239dde65cSChen, Kenneth W continue; 43237da4d641SPeter Zijlstra } 4324a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 4325a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 4326a8bda28dSNaoya Horiguchi spin_unlock(ptl); 4327a8bda28dSNaoya Horiguchi continue; 4328a8bda28dSNaoya Horiguchi } 4329a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 4330a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 4331a8bda28dSNaoya Horiguchi 4332a8bda28dSNaoya Horiguchi if (is_write_migration_entry(entry)) { 4333a8bda28dSNaoya Horiguchi pte_t newpte; 4334a8bda28dSNaoya Horiguchi 4335a8bda28dSNaoya Horiguchi make_migration_entry_read(&entry); 4336a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 4337e5251fd4SPunit Agrawal set_huge_swap_pte_at(mm, address, ptep, 4338e5251fd4SPunit Agrawal newpte, huge_page_size(h)); 4339a8bda28dSNaoya Horiguchi pages++; 4340a8bda28dSNaoya Horiguchi } 4341a8bda28dSNaoya Horiguchi spin_unlock(ptl); 4342a8bda28dSNaoya Horiguchi continue; 4343a8bda28dSNaoya Horiguchi } 4344a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 43458f860591SZhang, Yanmin pte = huge_ptep_get_and_clear(mm, address, ptep); 4346106c992aSGerald Schaefer pte = pte_mkhuge(huge_pte_modify(pte, newprot)); 4347be7517d6STony Lu pte = arch_make_huge_pte(pte, vma, NULL, 0); 43488f860591SZhang, Yanmin set_huge_pte_at(mm, address, ptep, pte); 43497da4d641SPeter Zijlstra pages++; 43508f860591SZhang, Yanmin } 4351cb900f41SKirill A. Shutemov spin_unlock(ptl); 43528f860591SZhang, Yanmin } 4353d833352aSMel Gorman /* 4354c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 4355d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 4356c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 4357d833352aSMel Gorman * and that page table be reused and filled with junk. 4358d833352aSMel Gorman */ 43595491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 43600f10851eSJérôme Glisse /* 43610f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 43620f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 43630f10851eSJérôme Glisse * 43640f10851eSJérôme Glisse * See Documentation/vm/mmu_notifier.txt 43650f10851eSJérôme Glisse */ 436683cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 4367a5338093SRik van Riel mmu_notifier_invalidate_range_end(mm, start, end); 43687da4d641SPeter Zijlstra 43697da4d641SPeter Zijlstra return pages << h->order; 43708f860591SZhang, Yanmin } 43718f860591SZhang, Yanmin 4372a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 4373a1e78772SMel Gorman long from, long to, 43745a6fe125SMel Gorman struct vm_area_struct *vma, 4375ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 4376e4e574b7SAdam Litke { 437717c9d12eSMel Gorman long ret, chg; 4378a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 437990481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 43809119a41eSJoonsoo Kim struct resv_map *resv_map; 43811c5ecae3SMike Kravetz long gbl_reserve; 4382e4e574b7SAdam Litke 4383a1e78772SMel Gorman /* 438417c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 438517c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 438690481622SDavid Gibson * without using reserves 438717c9d12eSMel Gorman */ 4388ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 438917c9d12eSMel Gorman return 0; 439017c9d12eSMel Gorman 439117c9d12eSMel Gorman /* 4392a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 4393a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 4394a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 4395a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 4396a1e78772SMel Gorman */ 43979119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 43984e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 43999119a41eSJoonsoo Kim 44001406ec9bSJoonsoo Kim chg = region_chg(resv_map, from, to); 44019119a41eSJoonsoo Kim 44029119a41eSJoonsoo Kim } else { 44039119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 44045a6fe125SMel Gorman if (!resv_map) 44055a6fe125SMel Gorman return -ENOMEM; 44065a6fe125SMel Gorman 440717c9d12eSMel Gorman chg = to - from; 440817c9d12eSMel Gorman 44095a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 44105a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 44115a6fe125SMel Gorman } 44125a6fe125SMel Gorman 4413c50ac050SDave Hansen if (chg < 0) { 4414c50ac050SDave Hansen ret = chg; 4415c50ac050SDave Hansen goto out_err; 4416c50ac050SDave Hansen } 441717c9d12eSMel Gorman 44181c5ecae3SMike Kravetz /* 44191c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 44201c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 44211c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 44221c5ecae3SMike Kravetz */ 44231c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 44241c5ecae3SMike Kravetz if (gbl_reserve < 0) { 4425c50ac050SDave Hansen ret = -ENOSPC; 4426c50ac050SDave Hansen goto out_err; 4427c50ac050SDave Hansen } 442817c9d12eSMel Gorman 442917c9d12eSMel Gorman /* 443017c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 443190481622SDavid Gibson * Hand the pages back to the subpool if there are not 443217c9d12eSMel Gorman */ 44331c5ecae3SMike Kravetz ret = hugetlb_acct_memory(h, gbl_reserve); 443417c9d12eSMel Gorman if (ret < 0) { 44351c5ecae3SMike Kravetz /* put back original number of pages, chg */ 44361c5ecae3SMike Kravetz (void)hugepage_subpool_put_pages(spool, chg); 4437c50ac050SDave Hansen goto out_err; 443817c9d12eSMel Gorman } 443917c9d12eSMel Gorman 444017c9d12eSMel Gorman /* 444117c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 444217c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 444317c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 444417c9d12eSMel Gorman * the reservation was and the page cache tells how much of 444517c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 444617c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 444717c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 444817c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 444917c9d12eSMel Gorman * else has to be done for private mappings here 445017c9d12eSMel Gorman */ 445133039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 445233039678SMike Kravetz long add = region_add(resv_map, from, to); 445333039678SMike Kravetz 445433039678SMike Kravetz if (unlikely(chg > add)) { 445533039678SMike Kravetz /* 445633039678SMike Kravetz * pages in this range were added to the reserve 445733039678SMike Kravetz * map between region_chg and region_add. This 445833039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 445933039678SMike Kravetz * the subpool and reserve counts modified above 446033039678SMike Kravetz * based on the difference. 446133039678SMike Kravetz */ 446233039678SMike Kravetz long rsv_adjust; 446333039678SMike Kravetz 446433039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 446533039678SMike Kravetz chg - add); 446633039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 446733039678SMike Kravetz } 446833039678SMike Kravetz } 4469a43a8c39SChen, Kenneth W return 0; 4470c50ac050SDave Hansen out_err: 44715e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 4472ff8c0c53SMike Kravetz /* Don't call region_abort if region_chg failed */ 4473ff8c0c53SMike Kravetz if (chg >= 0) 44745e911373SMike Kravetz region_abort(resv_map, from, to); 4475f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 4476f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 4477c50ac050SDave Hansen return ret; 4478a43a8c39SChen, Kenneth W } 4479a43a8c39SChen, Kenneth W 4480b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 4481b5cec28dSMike Kravetz long freed) 4482a43a8c39SChen, Kenneth W { 4483a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 44844e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 44859119a41eSJoonsoo Kim long chg = 0; 448690481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 44871c5ecae3SMike Kravetz long gbl_reserve; 448845c682a6SKen Chen 4489b5cec28dSMike Kravetz if (resv_map) { 4490b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 4491b5cec28dSMike Kravetz /* 4492b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 4493b5cec28dSMike Kravetz * must be split and another region descriptor can not be 4494b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 4495b5cec28dSMike Kravetz */ 4496b5cec28dSMike Kravetz if (chg < 0) 4497b5cec28dSMike Kravetz return chg; 4498b5cec28dSMike Kravetz } 4499b5cec28dSMike Kravetz 450045c682a6SKen Chen spin_lock(&inode->i_lock); 4501e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 450245c682a6SKen Chen spin_unlock(&inode->i_lock); 450345c682a6SKen Chen 45041c5ecae3SMike Kravetz /* 45051c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 45061c5ecae3SMike Kravetz * reservations to be released may be adjusted. 45071c5ecae3SMike Kravetz */ 45081c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 45091c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 4510b5cec28dSMike Kravetz 4511b5cec28dSMike Kravetz return 0; 4512a43a8c39SChen, Kenneth W } 451393f70f90SNaoya Horiguchi 45143212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 45153212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 45163212b535SSteve Capper struct vm_area_struct *vma, 45173212b535SSteve Capper unsigned long addr, pgoff_t idx) 45183212b535SSteve Capper { 45193212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 45203212b535SSteve Capper svma->vm_start; 45213212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 45223212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 45233212b535SSteve Capper 45243212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 4525de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 4526de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 45273212b535SSteve Capper 45283212b535SSteve Capper /* 45293212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 45303212b535SSteve Capper * page table page. 45313212b535SSteve Capper */ 45323212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 45333212b535SSteve Capper vm_flags != svm_flags || 45343212b535SSteve Capper sbase < svma->vm_start || svma->vm_end < s_end) 45353212b535SSteve Capper return 0; 45363212b535SSteve Capper 45373212b535SSteve Capper return saddr; 45383212b535SSteve Capper } 45393212b535SSteve Capper 454031aafb45SNicholas Krause static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) 45413212b535SSteve Capper { 45423212b535SSteve Capper unsigned long base = addr & PUD_MASK; 45433212b535SSteve Capper unsigned long end = base + PUD_SIZE; 45443212b535SSteve Capper 45453212b535SSteve Capper /* 45463212b535SSteve Capper * check on proper vm_flags and page table alignment 45473212b535SSteve Capper */ 45483212b535SSteve Capper if (vma->vm_flags & VM_MAYSHARE && 45493212b535SSteve Capper vma->vm_start <= base && end <= vma->vm_end) 455031aafb45SNicholas Krause return true; 455131aafb45SNicholas Krause return false; 45523212b535SSteve Capper } 45533212b535SSteve Capper 45543212b535SSteve Capper /* 45553212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 45563212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 45573212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 45583212b535SSteve Capper * code much cleaner. pmd allocation is essential for the shared case because 4559c8c06efaSDavidlohr Bueso * pud has to be populated inside the same i_mmap_rwsem section - otherwise 45603212b535SSteve Capper * racing tasks could either miss the sharing (see huge_pte_offset) or select a 45613212b535SSteve Capper * bad pmd for sharing. 45623212b535SSteve Capper */ 45633212b535SSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 45643212b535SSteve Capper { 45653212b535SSteve Capper struct vm_area_struct *vma = find_vma(mm, addr); 45663212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 45673212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 45683212b535SSteve Capper vma->vm_pgoff; 45693212b535SSteve Capper struct vm_area_struct *svma; 45703212b535SSteve Capper unsigned long saddr; 45713212b535SSteve Capper pte_t *spte = NULL; 45723212b535SSteve Capper pte_t *pte; 4573cb900f41SKirill A. Shutemov spinlock_t *ptl; 45743212b535SSteve Capper 45753212b535SSteve Capper if (!vma_shareable(vma, addr)) 45763212b535SSteve Capper return (pte_t *)pmd_alloc(mm, pud, addr); 45773212b535SSteve Capper 457883cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 45793212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 45803212b535SSteve Capper if (svma == vma) 45813212b535SSteve Capper continue; 45823212b535SSteve Capper 45833212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 45843212b535SSteve Capper if (saddr) { 45857868a208SPunit Agrawal spte = huge_pte_offset(svma->vm_mm, saddr, 45867868a208SPunit Agrawal vma_mmu_pagesize(svma)); 45873212b535SSteve Capper if (spte) { 45883212b535SSteve Capper get_page(virt_to_page(spte)); 45893212b535SSteve Capper break; 45903212b535SSteve Capper } 45913212b535SSteve Capper } 45923212b535SSteve Capper } 45933212b535SSteve Capper 45943212b535SSteve Capper if (!spte) 45953212b535SSteve Capper goto out; 45963212b535SSteve Capper 45978bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 4598dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 45993212b535SSteve Capper pud_populate(mm, pud, 46003212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 4601c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 4602dc6c9a35SKirill A. Shutemov } else { 46033212b535SSteve Capper put_page(virt_to_page(spte)); 4604dc6c9a35SKirill A. Shutemov } 4605cb900f41SKirill A. Shutemov spin_unlock(ptl); 46063212b535SSteve Capper out: 46073212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 460883cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 46093212b535SSteve Capper return pte; 46103212b535SSteve Capper } 46113212b535SSteve Capper 46123212b535SSteve Capper /* 46133212b535SSteve Capper * unmap huge page backed by shared pte. 46143212b535SSteve Capper * 46153212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 46163212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 46173212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 46183212b535SSteve Capper * 4619cb900f41SKirill A. Shutemov * called with page table lock held. 46203212b535SSteve Capper * 46213212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 46223212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 46233212b535SSteve Capper */ 46243212b535SSteve Capper int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 46253212b535SSteve Capper { 46263212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 4627c2febafcSKirill A. Shutemov p4d_t *p4d = p4d_offset(pgd, *addr); 4628c2febafcSKirill A. Shutemov pud_t *pud = pud_offset(p4d, *addr); 46293212b535SSteve Capper 46303212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 46313212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 46323212b535SSteve Capper return 0; 46333212b535SSteve Capper 46343212b535SSteve Capper pud_clear(pud); 46353212b535SSteve Capper put_page(virt_to_page(ptep)); 4636dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 46373212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 46383212b535SSteve Capper return 1; 46393212b535SSteve Capper } 46409e5fc74cSSteve Capper #define want_pmd_share() (1) 46419e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 46429e5fc74cSSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 46439e5fc74cSSteve Capper { 46449e5fc74cSSteve Capper return NULL; 46459e5fc74cSSteve Capper } 4646e81f2d22SZhang Zhen 4647e81f2d22SZhang Zhen int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 4648e81f2d22SZhang Zhen { 4649e81f2d22SZhang Zhen return 0; 4650e81f2d22SZhang Zhen } 46519e5fc74cSSteve Capper #define want_pmd_share() (0) 46523212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 46533212b535SSteve Capper 46549e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 46559e5fc74cSSteve Capper pte_t *huge_pte_alloc(struct mm_struct *mm, 46569e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 46579e5fc74cSSteve Capper { 46589e5fc74cSSteve Capper pgd_t *pgd; 4659c2febafcSKirill A. Shutemov p4d_t *p4d; 46609e5fc74cSSteve Capper pud_t *pud; 46619e5fc74cSSteve Capper pte_t *pte = NULL; 46629e5fc74cSSteve Capper 46639e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 4664f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 4665f4f0a3d8SKirill A. Shutemov if (!p4d) 4666f4f0a3d8SKirill A. Shutemov return NULL; 4667c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 46689e5fc74cSSteve Capper if (pud) { 46699e5fc74cSSteve Capper if (sz == PUD_SIZE) { 46709e5fc74cSSteve Capper pte = (pte_t *)pud; 46719e5fc74cSSteve Capper } else { 46729e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 46739e5fc74cSSteve Capper if (want_pmd_share() && pud_none(*pud)) 46749e5fc74cSSteve Capper pte = huge_pmd_share(mm, addr, pud); 46759e5fc74cSSteve Capper else 46769e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 46779e5fc74cSSteve Capper } 46789e5fc74cSSteve Capper } 46794e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 46809e5fc74cSSteve Capper 46819e5fc74cSSteve Capper return pte; 46829e5fc74cSSteve Capper } 46839e5fc74cSSteve Capper 46849b19df29SPunit Agrawal /* 46859b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 46869b19df29SPunit Agrawal * entry at address @addr 46879b19df29SPunit Agrawal * 46889b19df29SPunit Agrawal * Return: Pointer to page table or swap entry (PUD or PMD) for 46899b19df29SPunit Agrawal * address @addr, or NULL if a p*d_none() entry is encountered and the 46909b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 46919b19df29SPunit Agrawal * table. 46929b19df29SPunit Agrawal */ 46937868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 46947868a208SPunit Agrawal unsigned long addr, unsigned long sz) 46959e5fc74cSSteve Capper { 46969e5fc74cSSteve Capper pgd_t *pgd; 4697c2febafcSKirill A. Shutemov p4d_t *p4d; 46989e5fc74cSSteve Capper pud_t *pud; 4699c2febafcSKirill A. Shutemov pmd_t *pmd; 47009e5fc74cSSteve Capper 47019e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 4702c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 4703c2febafcSKirill A. Shutemov return NULL; 4704c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 4705c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 4706c2febafcSKirill A. Shutemov return NULL; 47079b19df29SPunit Agrawal 4708c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 47099b19df29SPunit Agrawal if (sz != PUD_SIZE && pud_none(*pud)) 4710c2febafcSKirill A. Shutemov return NULL; 47119b19df29SPunit Agrawal /* hugepage or swap? */ 47129b19df29SPunit Agrawal if (pud_huge(*pud) || !pud_present(*pud)) 47139e5fc74cSSteve Capper return (pte_t *)pud; 47149b19df29SPunit Agrawal 47159e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 47169b19df29SPunit Agrawal if (sz != PMD_SIZE && pmd_none(*pmd)) 47179b19df29SPunit Agrawal return NULL; 47189b19df29SPunit Agrawal /* hugepage or swap? */ 47199b19df29SPunit Agrawal if (pmd_huge(*pmd) || !pmd_present(*pmd)) 47209e5fc74cSSteve Capper return (pte_t *)pmd; 47219b19df29SPunit Agrawal 47229b19df29SPunit Agrawal return NULL; 47239e5fc74cSSteve Capper } 47249e5fc74cSSteve Capper 472561f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 472661f77edaSNaoya Horiguchi 472761f77edaSNaoya Horiguchi /* 472861f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 472961f77edaSNaoya Horiguchi * behavior. 473061f77edaSNaoya Horiguchi */ 473161f77edaSNaoya Horiguchi struct page * __weak 473261f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 473361f77edaSNaoya Horiguchi int write) 473461f77edaSNaoya Horiguchi { 473561f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 473661f77edaSNaoya Horiguchi } 473761f77edaSNaoya Horiguchi 473861f77edaSNaoya Horiguchi struct page * __weak 47394dc71451SAneesh Kumar K.V follow_huge_pd(struct vm_area_struct *vma, 47404dc71451SAneesh Kumar K.V unsigned long address, hugepd_t hpd, int flags, int pdshift) 47414dc71451SAneesh Kumar K.V { 47424dc71451SAneesh Kumar K.V WARN(1, "hugepd follow called with no support for hugepage directory format\n"); 47434dc71451SAneesh Kumar K.V return NULL; 47444dc71451SAneesh Kumar K.V } 47454dc71451SAneesh Kumar K.V 47464dc71451SAneesh Kumar K.V struct page * __weak 47479e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 4748e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 47499e5fc74cSSteve Capper { 4750e66f17ffSNaoya Horiguchi struct page *page = NULL; 4751e66f17ffSNaoya Horiguchi spinlock_t *ptl; 4752c9d398faSNaoya Horiguchi pte_t pte; 4753e66f17ffSNaoya Horiguchi retry: 4754e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 4755e66f17ffSNaoya Horiguchi spin_lock(ptl); 4756e66f17ffSNaoya Horiguchi /* 4757e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 4758e66f17ffSNaoya Horiguchi * unmapped from other threads. 4759e66f17ffSNaoya Horiguchi */ 4760e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 4761e66f17ffSNaoya Horiguchi goto out; 4762c9d398faSNaoya Horiguchi pte = huge_ptep_get((pte_t *)pmd); 4763c9d398faSNaoya Horiguchi if (pte_present(pte)) { 476497534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 4765e66f17ffSNaoya Horiguchi if (flags & FOLL_GET) 4766e66f17ffSNaoya Horiguchi get_page(page); 4767e66f17ffSNaoya Horiguchi } else { 4768c9d398faSNaoya Horiguchi if (is_hugetlb_entry_migration(pte)) { 4769e66f17ffSNaoya Horiguchi spin_unlock(ptl); 4770e66f17ffSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pmd, ptl); 4771e66f17ffSNaoya Horiguchi goto retry; 4772e66f17ffSNaoya Horiguchi } 4773e66f17ffSNaoya Horiguchi /* 4774e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 4775e66f17ffSNaoya Horiguchi * follow_page_mask(). 4776e66f17ffSNaoya Horiguchi */ 4777e66f17ffSNaoya Horiguchi } 4778e66f17ffSNaoya Horiguchi out: 4779e66f17ffSNaoya Horiguchi spin_unlock(ptl); 47809e5fc74cSSteve Capper return page; 47819e5fc74cSSteve Capper } 47829e5fc74cSSteve Capper 478361f77edaSNaoya Horiguchi struct page * __weak 47849e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 4785e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 47869e5fc74cSSteve Capper { 4787e66f17ffSNaoya Horiguchi if (flags & FOLL_GET) 4788e66f17ffSNaoya Horiguchi return NULL; 47899e5fc74cSSteve Capper 4790e66f17ffSNaoya Horiguchi return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 47919e5fc74cSSteve Capper } 47929e5fc74cSSteve Capper 4793faaa5b62SAnshuman Khandual struct page * __weak 4794faaa5b62SAnshuman Khandual follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) 4795faaa5b62SAnshuman Khandual { 4796faaa5b62SAnshuman Khandual if (flags & FOLL_GET) 4797faaa5b62SAnshuman Khandual return NULL; 4798faaa5b62SAnshuman Khandual 4799faaa5b62SAnshuman Khandual return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); 4800faaa5b62SAnshuman Khandual } 4801faaa5b62SAnshuman Khandual 480231caf665SNaoya Horiguchi bool isolate_huge_page(struct page *page, struct list_head *list) 480331caf665SNaoya Horiguchi { 4804bcc54222SNaoya Horiguchi bool ret = true; 4805bcc54222SNaoya Horiguchi 4806309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 480731caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 4808bcc54222SNaoya Horiguchi if (!page_huge_active(page) || !get_page_unless_zero(page)) { 4809bcc54222SNaoya Horiguchi ret = false; 4810bcc54222SNaoya Horiguchi goto unlock; 4811bcc54222SNaoya Horiguchi } 4812bcc54222SNaoya Horiguchi clear_page_huge_active(page); 481331caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 4814bcc54222SNaoya Horiguchi unlock: 481531caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 4816bcc54222SNaoya Horiguchi return ret; 481731caf665SNaoya Horiguchi } 481831caf665SNaoya Horiguchi 481931caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 482031caf665SNaoya Horiguchi { 4821309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 482231caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 4823bcc54222SNaoya Horiguchi set_page_huge_active(page); 482431caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 482531caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 482631caf665SNaoya Horiguchi put_page(page); 482731caf665SNaoya Horiguchi } 4828ab5ac90aSMichal Hocko 4829ab5ac90aSMichal Hocko void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) 4830ab5ac90aSMichal Hocko { 4831ab5ac90aSMichal Hocko struct hstate *h = page_hstate(oldpage); 4832ab5ac90aSMichal Hocko 4833ab5ac90aSMichal Hocko hugetlb_cgroup_migrate(oldpage, newpage); 4834ab5ac90aSMichal Hocko set_page_owner_migrate_reason(newpage, reason); 4835ab5ac90aSMichal Hocko 4836ab5ac90aSMichal Hocko /* 4837ab5ac90aSMichal Hocko * transfer temporary state of the new huge page. This is 4838ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 4839ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 4840ab5ac90aSMichal Hocko * the temporary status. 4841ab5ac90aSMichal Hocko * 4842ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 4843ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 4844ab5ac90aSMichal Hocko * the per-node's. 4845ab5ac90aSMichal Hocko */ 4846ab5ac90aSMichal Hocko if (PageHugeTemporary(newpage)) { 4847ab5ac90aSMichal Hocko int old_nid = page_to_nid(oldpage); 4848ab5ac90aSMichal Hocko int new_nid = page_to_nid(newpage); 4849ab5ac90aSMichal Hocko 4850ab5ac90aSMichal Hocko SetPageHugeTemporary(oldpage); 4851ab5ac90aSMichal Hocko ClearPageHugeTemporary(newpage); 4852ab5ac90aSMichal Hocko 4853ab5ac90aSMichal Hocko spin_lock(&hugetlb_lock); 4854ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 4855ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 4856ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 4857ab5ac90aSMichal Hocko } 4858ab5ac90aSMichal Hocko spin_unlock(&hugetlb_lock); 4859ab5ac90aSMichal Hocko } 4860ab5ac90aSMichal Hocko } 4861