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 13810c397daeSMichal Hocko static struct page *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 /* 14000c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 14010c397daeSMichal Hocko * should use this function to get new hugetlb pages 14020c397daeSMichal Hocko */ 14030c397daeSMichal Hocko static struct page *alloc_fresh_huge_page(struct hstate *h, 14040c397daeSMichal Hocko gfp_t gfp_mask, int nid, nodemask_t *nmask) 14050c397daeSMichal Hocko { 14060c397daeSMichal Hocko struct page *page; 14070c397daeSMichal Hocko 14080c397daeSMichal Hocko if (hstate_is_gigantic(h)) 14090c397daeSMichal Hocko page = alloc_gigantic_page(h, gfp_mask, nid, nmask); 14100c397daeSMichal Hocko else 14110c397daeSMichal Hocko page = alloc_buddy_huge_page(h, gfp_mask, 14120c397daeSMichal Hocko nid, nmask); 14130c397daeSMichal Hocko if (!page) 14140c397daeSMichal Hocko return NULL; 14150c397daeSMichal Hocko 14160c397daeSMichal Hocko if (hstate_is_gigantic(h)) 14170c397daeSMichal Hocko prep_compound_gigantic_page(page, huge_page_order(h)); 14180c397daeSMichal Hocko prep_new_huge_page(h, page, page_to_nid(page)); 14190c397daeSMichal Hocko 14200c397daeSMichal Hocko return page; 14210c397daeSMichal Hocko } 14220c397daeSMichal Hocko 14230c397daeSMichal Hocko /* 1424af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 1425af0fb9dfSMichal Hocko * manner. 1426af0fb9dfSMichal Hocko */ 14270c397daeSMichal Hocko static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 1428b2261026SJoonsoo Kim { 1429b2261026SJoonsoo Kim struct page *page; 1430b2261026SJoonsoo Kim int nr_nodes, node; 1431af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 1432b2261026SJoonsoo Kim 1433b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 14340c397daeSMichal Hocko page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed); 1435af0fb9dfSMichal Hocko if (page) 1436b2261026SJoonsoo Kim break; 1437b2261026SJoonsoo Kim } 1438b2261026SJoonsoo Kim 1439af0fb9dfSMichal Hocko if (!page) 1440af0fb9dfSMichal Hocko return 0; 1441b2261026SJoonsoo Kim 1442af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 1443af0fb9dfSMichal Hocko 1444af0fb9dfSMichal Hocko return 1; 1445b2261026SJoonsoo Kim } 1446b2261026SJoonsoo Kim 1447e8c5c824SLee Schermerhorn /* 1448e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 1449e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 1450e8c5c824SLee Schermerhorn * balanced over allowed nodes. 1451e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 1452e8c5c824SLee Schermerhorn */ 14536ae11b27SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 14546ae11b27SLee Schermerhorn bool acct_surplus) 1455e8c5c824SLee Schermerhorn { 1456b2261026SJoonsoo Kim int nr_nodes, node; 1457e8c5c824SLee Schermerhorn int ret = 0; 1458e8c5c824SLee Schermerhorn 1459b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1460685f3457SLee Schermerhorn /* 1461685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 1462685f3457SLee Schermerhorn * nodes with surplus pages. 1463685f3457SLee Schermerhorn */ 1464b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 1465b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 1466e8c5c824SLee Schermerhorn struct page *page = 1467b2261026SJoonsoo Kim list_entry(h->hugepage_freelists[node].next, 1468e8c5c824SLee Schermerhorn struct page, lru); 1469e8c5c824SLee Schermerhorn list_del(&page->lru); 1470e8c5c824SLee Schermerhorn h->free_huge_pages--; 1471b2261026SJoonsoo Kim h->free_huge_pages_node[node]--; 1472685f3457SLee Schermerhorn if (acct_surplus) { 1473685f3457SLee Schermerhorn h->surplus_huge_pages--; 1474b2261026SJoonsoo Kim h->surplus_huge_pages_node[node]--; 1475685f3457SLee Schermerhorn } 1476e8c5c824SLee Schermerhorn update_and_free_page(h, page); 1477e8c5c824SLee Schermerhorn ret = 1; 14789a76db09SLee Schermerhorn break; 1479e8c5c824SLee Schermerhorn } 1480b2261026SJoonsoo Kim } 1481e8c5c824SLee Schermerhorn 1482e8c5c824SLee Schermerhorn return ret; 1483e8c5c824SLee Schermerhorn } 1484e8c5c824SLee Schermerhorn 1485c8721bbbSNaoya Horiguchi /* 1486c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 1487082d5b6bSGerald Schaefer * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the 1488082d5b6bSGerald Schaefer * number of free hugepages would be reduced below the number of reserved 1489082d5b6bSGerald Schaefer * hugepages. 1490c8721bbbSNaoya Horiguchi */ 1491c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 1492c8721bbbSNaoya Horiguchi { 1493082d5b6bSGerald Schaefer int rc = 0; 1494082d5b6bSGerald Schaefer 1495c8721bbbSNaoya Horiguchi spin_lock(&hugetlb_lock); 1496c8721bbbSNaoya Horiguchi if (PageHuge(page) && !page_count(page)) { 14972247bb33SGerald Schaefer struct page *head = compound_head(page); 14982247bb33SGerald Schaefer struct hstate *h = page_hstate(head); 14992247bb33SGerald Schaefer int nid = page_to_nid(head); 1500082d5b6bSGerald Schaefer if (h->free_huge_pages - h->resv_huge_pages == 0) { 1501082d5b6bSGerald Schaefer rc = -EBUSY; 1502082d5b6bSGerald Schaefer goto out; 1503082d5b6bSGerald Schaefer } 1504c3114a84SAnshuman Khandual /* 1505c3114a84SAnshuman Khandual * Move PageHWPoison flag from head page to the raw error page, 1506c3114a84SAnshuman Khandual * which makes any subpages rather than the error page reusable. 1507c3114a84SAnshuman Khandual */ 1508c3114a84SAnshuman Khandual if (PageHWPoison(head) && page != head) { 1509c3114a84SAnshuman Khandual SetPageHWPoison(page); 1510c3114a84SAnshuman Khandual ClearPageHWPoison(head); 1511c3114a84SAnshuman Khandual } 15122247bb33SGerald Schaefer list_del(&head->lru); 1513c8721bbbSNaoya Horiguchi h->free_huge_pages--; 1514c8721bbbSNaoya Horiguchi h->free_huge_pages_node[nid]--; 1515c1470b33Szhong jiang h->max_huge_pages--; 15162247bb33SGerald Schaefer update_and_free_page(h, head); 1517c8721bbbSNaoya Horiguchi } 1518082d5b6bSGerald Schaefer out: 1519c8721bbbSNaoya Horiguchi spin_unlock(&hugetlb_lock); 1520082d5b6bSGerald Schaefer return rc; 1521c8721bbbSNaoya Horiguchi } 1522c8721bbbSNaoya Horiguchi 1523c8721bbbSNaoya Horiguchi /* 1524c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 1525c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 15262247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 15272247bb33SGerald Schaefer * part of it lies within the given range. 1528082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 1529082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 1530c8721bbbSNaoya Horiguchi */ 1531082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 1532c8721bbbSNaoya Horiguchi { 1533c8721bbbSNaoya Horiguchi unsigned long pfn; 1534eb03aa00SGerald Schaefer struct page *page; 1535082d5b6bSGerald Schaefer int rc = 0; 1536c8721bbbSNaoya Horiguchi 1537d0177639SLi Zhong if (!hugepages_supported()) 1538082d5b6bSGerald Schaefer return rc; 1539d0177639SLi Zhong 1540eb03aa00SGerald Schaefer for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { 1541eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 1542eb03aa00SGerald Schaefer if (PageHuge(page) && !page_count(page)) { 1543eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 1544eb03aa00SGerald Schaefer if (rc) 1545082d5b6bSGerald Schaefer break; 1546eb03aa00SGerald Schaefer } 1547eb03aa00SGerald Schaefer } 1548082d5b6bSGerald Schaefer 1549082d5b6bSGerald Schaefer return rc; 1550c8721bbbSNaoya Horiguchi } 1551c8721bbbSNaoya Horiguchi 1552ab5ac90aSMichal Hocko /* 1553ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 1554ab5ac90aSMichal Hocko */ 15550c397daeSMichal Hocko static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, 1556aaf14e40SMichal Hocko int nid, nodemask_t *nmask) 15577893d1d5SAdam Litke { 15589980d744SMichal Hocko struct page *page = NULL; 15597893d1d5SAdam Litke 1560bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1561aa888a74SAndi Kleen return NULL; 1562aa888a74SAndi Kleen 1563d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 15649980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 15659980d744SMichal Hocko goto out_unlock; 1566d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1567d1c3fb1fSNishanth Aravamudan 15680c397daeSMichal Hocko page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask); 15699980d744SMichal Hocko if (!page) 15700c397daeSMichal Hocko return NULL; 1571d1c3fb1fSNishanth Aravamudan 15727893d1d5SAdam Litke spin_lock(&hugetlb_lock); 15739980d744SMichal Hocko /* 15749980d744SMichal Hocko * We could have raced with the pool size change. 15759980d744SMichal Hocko * Double check that and simply deallocate the new page 15769980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 15779980d744SMichal Hocko * temporary page to workaround the nasty free_huge_page 15789980d744SMichal Hocko * codeflow 15799980d744SMichal Hocko */ 15809980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 15819980d744SMichal Hocko SetPageHugeTemporary(page); 15829980d744SMichal Hocko put_page(page); 15839980d744SMichal Hocko page = NULL; 15849980d744SMichal Hocko } else { 15859980d744SMichal Hocko h->surplus_huge_pages++; 15860c397daeSMichal Hocko h->nr_huge_pages_node[page_to_nid(page)]++; 15877893d1d5SAdam Litke } 15889980d744SMichal Hocko 15899980d744SMichal Hocko out_unlock: 1590d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 15917893d1d5SAdam Litke 15927893d1d5SAdam Litke return page; 15937893d1d5SAdam Litke } 15947893d1d5SAdam Litke 15950c397daeSMichal Hocko static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, 1596ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 1597ab5ac90aSMichal Hocko { 1598ab5ac90aSMichal Hocko struct page *page; 1599ab5ac90aSMichal Hocko 1600ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 1601ab5ac90aSMichal Hocko return NULL; 1602ab5ac90aSMichal Hocko 16030c397daeSMichal Hocko page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask); 1604ab5ac90aSMichal Hocko if (!page) 1605ab5ac90aSMichal Hocko return NULL; 1606ab5ac90aSMichal Hocko 1607ab5ac90aSMichal Hocko /* 1608ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 1609ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 1610ab5ac90aSMichal Hocko */ 1611ab5ac90aSMichal Hocko SetPageHugeTemporary(page); 1612ab5ac90aSMichal Hocko 1613ab5ac90aSMichal Hocko return page; 1614ab5ac90aSMichal Hocko } 1615ab5ac90aSMichal Hocko 1616e4e574b7SAdam Litke /* 1617099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 1618099730d6SDave Hansen */ 1619e0ec90eeSDave Hansen static 16200c397daeSMichal Hocko struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, 1621099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 1622099730d6SDave Hansen { 1623aaf14e40SMichal Hocko struct page *page; 1624aaf14e40SMichal Hocko struct mempolicy *mpol; 1625aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 1626aaf14e40SMichal Hocko int nid; 1627aaf14e40SMichal Hocko nodemask_t *nodemask; 1628aaf14e40SMichal Hocko 1629aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 16300c397daeSMichal Hocko page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); 1631aaf14e40SMichal Hocko mpol_cond_put(mpol); 1632aaf14e40SMichal Hocko 1633aaf14e40SMichal Hocko return page; 1634099730d6SDave Hansen } 1635099730d6SDave Hansen 1636ab5ac90aSMichal Hocko /* page migration callback function */ 1637bf50bab2SNaoya Horiguchi struct page *alloc_huge_page_node(struct hstate *h, int nid) 1638bf50bab2SNaoya Horiguchi { 1639aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 16404ef91848SJoonsoo Kim struct page *page = NULL; 1641bf50bab2SNaoya Horiguchi 1642aaf14e40SMichal Hocko if (nid != NUMA_NO_NODE) 1643aaf14e40SMichal Hocko gfp_mask |= __GFP_THISNODE; 1644aaf14e40SMichal Hocko 1645bf50bab2SNaoya Horiguchi spin_lock(&hugetlb_lock); 16464ef91848SJoonsoo Kim if (h->free_huge_pages - h->resv_huge_pages > 0) 16473e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL); 1648bf50bab2SNaoya Horiguchi spin_unlock(&hugetlb_lock); 1649bf50bab2SNaoya Horiguchi 165094ae8ba7SAneesh Kumar K.V if (!page) 16510c397daeSMichal Hocko page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL); 1652bf50bab2SNaoya Horiguchi 1653bf50bab2SNaoya Horiguchi return page; 1654bf50bab2SNaoya Horiguchi } 1655bf50bab2SNaoya Horiguchi 1656ab5ac90aSMichal Hocko /* page migration callback function */ 16573e59fcb0SMichal Hocko struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, 16583e59fcb0SMichal Hocko nodemask_t *nmask) 16594db9b2efSMichal Hocko { 1660aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 16614db9b2efSMichal Hocko 16624db9b2efSMichal Hocko spin_lock(&hugetlb_lock); 16634db9b2efSMichal Hocko if (h->free_huge_pages - h->resv_huge_pages > 0) { 16643e59fcb0SMichal Hocko struct page *page; 16653e59fcb0SMichal Hocko 16663e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); 16673e59fcb0SMichal Hocko if (page) { 16683e59fcb0SMichal Hocko spin_unlock(&hugetlb_lock); 16693e59fcb0SMichal Hocko return page; 16704db9b2efSMichal Hocko } 16714db9b2efSMichal Hocko } 16724db9b2efSMichal Hocko spin_unlock(&hugetlb_lock); 16734db9b2efSMichal Hocko 16740c397daeSMichal Hocko return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); 16754db9b2efSMichal Hocko } 16764db9b2efSMichal Hocko 1677bf50bab2SNaoya Horiguchi /* 167825985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 1679e4e574b7SAdam Litke * of size 'delta'. 1680e4e574b7SAdam Litke */ 1681a5516438SAndi Kleen static int gather_surplus_pages(struct hstate *h, int delta) 1682e4e574b7SAdam Litke { 1683e4e574b7SAdam Litke struct list_head surplus_list; 1684e4e574b7SAdam Litke struct page *page, *tmp; 1685e4e574b7SAdam Litke int ret, i; 1686e4e574b7SAdam Litke int needed, allocated; 168728073b02SHillf Danton bool alloc_ok = true; 1688e4e574b7SAdam Litke 1689a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 1690ac09b3a1SAdam Litke if (needed <= 0) { 1691a5516438SAndi Kleen h->resv_huge_pages += delta; 1692e4e574b7SAdam Litke return 0; 1693ac09b3a1SAdam Litke } 1694e4e574b7SAdam Litke 1695e4e574b7SAdam Litke allocated = 0; 1696e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 1697e4e574b7SAdam Litke 1698e4e574b7SAdam Litke ret = -ENOMEM; 1699e4e574b7SAdam Litke retry: 1700e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 1701e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 17020c397daeSMichal Hocko page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), 1703aaf14e40SMichal Hocko NUMA_NO_NODE, NULL); 170428073b02SHillf Danton if (!page) { 170528073b02SHillf Danton alloc_ok = false; 170628073b02SHillf Danton break; 170728073b02SHillf Danton } 1708e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 170969ed779aSDavid Rientjes cond_resched(); 1710e4e574b7SAdam Litke } 171128073b02SHillf Danton allocated += i; 1712e4e574b7SAdam Litke 1713e4e574b7SAdam Litke /* 1714e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 1715e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 1716e4e574b7SAdam Litke */ 1717e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 1718a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 1719a5516438SAndi Kleen (h->free_huge_pages + allocated); 172028073b02SHillf Danton if (needed > 0) { 172128073b02SHillf Danton if (alloc_ok) 1722e4e574b7SAdam Litke goto retry; 172328073b02SHillf Danton /* 172428073b02SHillf Danton * We were not able to allocate enough pages to 172528073b02SHillf Danton * satisfy the entire reservation so we free what 172628073b02SHillf Danton * we've allocated so far. 172728073b02SHillf Danton */ 172828073b02SHillf Danton goto free; 172928073b02SHillf Danton } 1730e4e574b7SAdam Litke /* 1731e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 173225985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 1733e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 1734ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 1735ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 1736ac09b3a1SAdam Litke * before they are reserved. 1737e4e574b7SAdam Litke */ 1738e4e574b7SAdam Litke needed += allocated; 1739a5516438SAndi Kleen h->resv_huge_pages += delta; 1740e4e574b7SAdam Litke ret = 0; 1741a9869b83SNaoya Horiguchi 174219fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 174319fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 174419fc3f0aSAdam Litke if ((--needed) < 0) 174519fc3f0aSAdam Litke break; 1746a9869b83SNaoya Horiguchi /* 1747a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 1748a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 1749a9869b83SNaoya Horiguchi */ 1750a9869b83SNaoya Horiguchi put_page_testzero(page); 1751309381feSSasha Levin VM_BUG_ON_PAGE(page_count(page), page); 1752a5516438SAndi Kleen enqueue_huge_page(h, page); 175319fc3f0aSAdam Litke } 175428073b02SHillf Danton free: 1755b0365c8dSHillf Danton spin_unlock(&hugetlb_lock); 175619fc3f0aSAdam Litke 175719fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 1758c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 1759a9869b83SNaoya Horiguchi put_page(page); 176019fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 1761e4e574b7SAdam Litke 1762e4e574b7SAdam Litke return ret; 1763e4e574b7SAdam Litke } 1764e4e574b7SAdam Litke 1765e4e574b7SAdam Litke /* 1766e5bbc8a6SMike Kravetz * This routine has two main purposes: 1767e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 1768e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 1769e5bbc8a6SMike Kravetz * to the associated reservation map. 1770e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 1771e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 1772e5bbc8a6SMike Kravetz * 1773e5bbc8a6SMike Kravetz * Called with hugetlb_lock held. However, the lock could be dropped (and 1774e5bbc8a6SMike Kravetz * reacquired) during calls to cond_resched_lock. Whenever dropping the lock, 1775e5bbc8a6SMike Kravetz * we must make sure nobody else can claim pages we are in the process of 1776e5bbc8a6SMike Kravetz * freeing. Do this by ensuring resv_huge_page always is greater than the 1777e5bbc8a6SMike Kravetz * number of huge pages we plan to free when dropping the lock. 1778e4e574b7SAdam Litke */ 1779a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 1780a5516438SAndi Kleen unsigned long unused_resv_pages) 1781e4e574b7SAdam Litke { 1782e4e574b7SAdam Litke unsigned long nr_pages; 1783e4e574b7SAdam Litke 1784aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 1785bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1786e5bbc8a6SMike Kravetz goto out; 1787aa888a74SAndi Kleen 1788e5bbc8a6SMike Kravetz /* 1789e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 1790e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 1791e5bbc8a6SMike Kravetz */ 1792a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 1793e4e574b7SAdam Litke 1794685f3457SLee Schermerhorn /* 1795685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 17969b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 17979b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 17989b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 17999b5e5d0fSLee Schermerhorn * free_pool_huge_page() will balance the the freed pages across the 18009b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 1801e5bbc8a6SMike Kravetz * 1802e5bbc8a6SMike Kravetz * Note that we decrement resv_huge_pages as we free the pages. If 1803e5bbc8a6SMike Kravetz * we drop the lock, resv_huge_pages will still be sufficiently large 1804e5bbc8a6SMike Kravetz * to cover subsequent pages we may free. 1805685f3457SLee Schermerhorn */ 1806685f3457SLee Schermerhorn while (nr_pages--) { 1807e5bbc8a6SMike Kravetz h->resv_huge_pages--; 1808e5bbc8a6SMike Kravetz unused_resv_pages--; 18098cebfcd0SLai Jiangshan if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) 1810e5bbc8a6SMike Kravetz goto out; 18117848a4bfSMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 1812e4e574b7SAdam Litke } 1813e5bbc8a6SMike Kravetz 1814e5bbc8a6SMike Kravetz out: 1815e5bbc8a6SMike Kravetz /* Fully uncommit the reservation */ 1816e5bbc8a6SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 1817e4e574b7SAdam Litke } 1818e4e574b7SAdam Litke 18195e911373SMike Kravetz 1820c37f9fb1SAndy Whitcroft /* 1821feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 18225e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 1823cf3ad20bSMike Kravetz * 1824cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 1825cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 1826cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 1827cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 1828cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 1829feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 1830feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 1831feba16e2SMike Kravetz * is called in such cases. 1832cf3ad20bSMike Kravetz * 1833cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 1834cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 1835cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 1836cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 1837cf3ad20bSMike Kravetz * take appropriate action. 183896b96a96SMike Kravetz * 183996b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 184096b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 184196b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 184296b96a96SMike Kravetz * reservation exists. 1843c37f9fb1SAndy Whitcroft */ 18445e911373SMike Kravetz enum vma_resv_mode { 18455e911373SMike Kravetz VMA_NEEDS_RESV, 18465e911373SMike Kravetz VMA_COMMIT_RESV, 1847feba16e2SMike Kravetz VMA_END_RESV, 184896b96a96SMike Kravetz VMA_ADD_RESV, 18495e911373SMike Kravetz }; 1850cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 1851cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 18525e911373SMike Kravetz enum vma_resv_mode mode) 1853c37f9fb1SAndy Whitcroft { 18544e35f483SJoonsoo Kim struct resv_map *resv; 18554e35f483SJoonsoo Kim pgoff_t idx; 1856cf3ad20bSMike Kravetz long ret; 1857c37f9fb1SAndy Whitcroft 18584e35f483SJoonsoo Kim resv = vma_resv_map(vma); 18594e35f483SJoonsoo Kim if (!resv) 1860c37f9fb1SAndy Whitcroft return 1; 1861c37f9fb1SAndy Whitcroft 18624e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 18635e911373SMike Kravetz switch (mode) { 18645e911373SMike Kravetz case VMA_NEEDS_RESV: 1865cf3ad20bSMike Kravetz ret = region_chg(resv, idx, idx + 1); 18665e911373SMike Kravetz break; 18675e911373SMike Kravetz case VMA_COMMIT_RESV: 18685e911373SMike Kravetz ret = region_add(resv, idx, idx + 1); 18695e911373SMike Kravetz break; 1870feba16e2SMike Kravetz case VMA_END_RESV: 18715e911373SMike Kravetz region_abort(resv, idx, idx + 1); 18725e911373SMike Kravetz ret = 0; 18735e911373SMike Kravetz break; 187496b96a96SMike Kravetz case VMA_ADD_RESV: 187596b96a96SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) 187696b96a96SMike Kravetz ret = region_add(resv, idx, idx + 1); 187796b96a96SMike Kravetz else { 187896b96a96SMike Kravetz region_abort(resv, idx, idx + 1); 187996b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 188096b96a96SMike Kravetz } 188196b96a96SMike Kravetz break; 18825e911373SMike Kravetz default: 18835e911373SMike Kravetz BUG(); 18845e911373SMike Kravetz } 188584afd99bSAndy Whitcroft 18864e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) 1887cf3ad20bSMike Kravetz return ret; 188867961f9dSMike Kravetz else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { 188967961f9dSMike Kravetz /* 189067961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 189167961f9dSMike Kravetz * However, a file associated with mapping could have been 189267961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 189367961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 189467961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 189567961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 189667961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 189767961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 189867961f9dSMike Kravetz * reservation has already been consumed. As a result, the 189967961f9dSMike Kravetz * return value of this routine is the opposite of the 190067961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 190167961f9dSMike Kravetz */ 190267961f9dSMike Kravetz if (ret) 190367961f9dSMike Kravetz return 0; 190467961f9dSMike Kravetz else 190567961f9dSMike Kravetz return 1; 190667961f9dSMike Kravetz } 19074e35f483SJoonsoo Kim else 1908cf3ad20bSMike Kravetz return ret < 0 ? ret : 0; 190984afd99bSAndy Whitcroft } 1910cf3ad20bSMike Kravetz 1911cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 1912a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 1913c37f9fb1SAndy Whitcroft { 19145e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 1915cf3ad20bSMike Kravetz } 1916c37f9fb1SAndy Whitcroft 1917cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 1918cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 1919cf3ad20bSMike Kravetz { 19205e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 19215e911373SMike Kravetz } 19225e911373SMike Kravetz 1923feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 19245e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 19255e911373SMike Kravetz { 1926feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 1927c37f9fb1SAndy Whitcroft } 1928c37f9fb1SAndy Whitcroft 192996b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 193096b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 193196b96a96SMike Kravetz { 193296b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 193396b96a96SMike Kravetz } 193496b96a96SMike Kravetz 193596b96a96SMike Kravetz /* 193696b96a96SMike Kravetz * This routine is called to restore a reservation on error paths. In the 193796b96a96SMike Kravetz * specific error paths, a huge page was allocated (via alloc_huge_page) 193896b96a96SMike Kravetz * and is about to be freed. If a reservation for the page existed, 193996b96a96SMike Kravetz * alloc_huge_page would have consumed the reservation and set PagePrivate 194096b96a96SMike Kravetz * in the newly allocated page. When the page is freed via free_huge_page, 194196b96a96SMike Kravetz * the global reservation count will be incremented if PagePrivate is set. 194296b96a96SMike Kravetz * However, free_huge_page can not adjust the reserve map. Adjust the 194396b96a96SMike Kravetz * reserve map here to be consistent with global reserve count adjustments 194496b96a96SMike Kravetz * to be made by free_huge_page. 194596b96a96SMike Kravetz */ 194696b96a96SMike Kravetz static void restore_reserve_on_error(struct hstate *h, 194796b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long address, 194896b96a96SMike Kravetz struct page *page) 194996b96a96SMike Kravetz { 195096b96a96SMike Kravetz if (unlikely(PagePrivate(page))) { 195196b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 195296b96a96SMike Kravetz 195396b96a96SMike Kravetz if (unlikely(rc < 0)) { 195496b96a96SMike Kravetz /* 195596b96a96SMike Kravetz * Rare out of memory condition in reserve map 195696b96a96SMike Kravetz * manipulation. Clear PagePrivate so that 195796b96a96SMike Kravetz * global reserve count will not be incremented 195896b96a96SMike Kravetz * by free_huge_page. This will make it appear 195996b96a96SMike Kravetz * as though the reservation for this page was 196096b96a96SMike Kravetz * consumed. This may prevent the task from 196196b96a96SMike Kravetz * faulting in the page at a later time. This 196296b96a96SMike Kravetz * is better than inconsistent global huge page 196396b96a96SMike Kravetz * accounting of reserve counts. 196496b96a96SMike Kravetz */ 196596b96a96SMike Kravetz ClearPagePrivate(page); 196696b96a96SMike Kravetz } else if (rc) { 196796b96a96SMike Kravetz rc = vma_add_reservation(h, vma, address); 196896b96a96SMike Kravetz if (unlikely(rc < 0)) 196996b96a96SMike Kravetz /* 197096b96a96SMike Kravetz * See above comment about rare out of 197196b96a96SMike Kravetz * memory condition. 197296b96a96SMike Kravetz */ 197396b96a96SMike Kravetz ClearPagePrivate(page); 197496b96a96SMike Kravetz } else 197596b96a96SMike Kravetz vma_end_reservation(h, vma, address); 197696b96a96SMike Kravetz } 197796b96a96SMike Kravetz } 197896b96a96SMike Kravetz 197970c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 198004f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 1981348ea204SAdam Litke { 198290481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 1983a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 1984348ea204SAdam Litke struct page *page; 1985d85f69b0SMike Kravetz long map_chg, map_commit; 1986d85f69b0SMike Kravetz long gbl_chg; 19876d76dcf4SAneesh Kumar K.V int ret, idx; 19886d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 19892fc39cecSAdam Litke 19906d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 1991a1e78772SMel Gorman /* 1992d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 1993d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 1994d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 1995a1e78772SMel Gorman */ 1996d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 1997d85f69b0SMike Kravetz if (map_chg < 0) 199876dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 1999d85f69b0SMike Kravetz 2000d85f69b0SMike Kravetz /* 2001d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 2002d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 2003d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 2004d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 2005d85f69b0SMike Kravetz * checked against any subpool limit. 2006d85f69b0SMike Kravetz */ 2007d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 2008d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 2009d85f69b0SMike Kravetz if (gbl_chg < 0) { 2010feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 201176dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 20125e911373SMike Kravetz } 201390d8b7e6SAdam Litke 2014d85f69b0SMike Kravetz /* 2015d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 2016d85f69b0SMike Kravetz * map, there could be reservations associated with the 2017d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 2018d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 2019d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 2020d85f69b0SMike Kravetz * the subpool reservations. 2021d85f69b0SMike Kravetz */ 2022d85f69b0SMike Kravetz if (avoid_reserve) 2023d85f69b0SMike Kravetz gbl_chg = 1; 2024d85f69b0SMike Kravetz } 2025d85f69b0SMike Kravetz 20266d76dcf4SAneesh Kumar K.V ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 20278f34af6fSJianyu Zhan if (ret) 20288f34af6fSJianyu Zhan goto out_subpool_put; 20298f34af6fSJianyu Zhan 2030a1e78772SMel Gorman spin_lock(&hugetlb_lock); 2031d85f69b0SMike Kravetz /* 2032d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 2033d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 2034d85f69b0SMike Kravetz * a reservation exists for the allocation. 2035d85f69b0SMike Kravetz */ 2036d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 203781a6fcaeSJoonsoo Kim if (!page) { 203894ae8ba7SAneesh Kumar K.V spin_unlock(&hugetlb_lock); 20390c397daeSMichal Hocko page = alloc_buddy_huge_page_with_mpol(h, vma, addr); 20408f34af6fSJianyu Zhan if (!page) 20418f34af6fSJianyu Zhan goto out_uncharge_cgroup; 2042a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 2043a88c7695SNaoya Horiguchi SetPagePrivate(page); 2044a88c7695SNaoya Horiguchi h->resv_huge_pages--; 2045a88c7695SNaoya Horiguchi } 204679dbb236SAneesh Kumar K.V spin_lock(&hugetlb_lock); 204779dbb236SAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 204881a6fcaeSJoonsoo Kim /* Fall through */ 2049a1e78772SMel Gorman } 205081a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 205181a6fcaeSJoonsoo Kim spin_unlock(&hugetlb_lock); 2052a1e78772SMel Gorman 205390481622SDavid Gibson set_page_private(page, (unsigned long)spool); 2054a1e78772SMel Gorman 2055d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 2056d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 205733039678SMike Kravetz /* 205833039678SMike Kravetz * The page was added to the reservation map between 205933039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 206033039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 206133039678SMike Kravetz * Adjust for the subpool count incremented above AND 206233039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 206333039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 206433039678SMike Kravetz * no longer applies. 206533039678SMike Kravetz */ 206633039678SMike Kravetz long rsv_adjust; 206733039678SMike Kravetz 206833039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 206933039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 207033039678SMike Kravetz } 20717893d1d5SAdam Litke return page; 20728f34af6fSJianyu Zhan 20738f34af6fSJianyu Zhan out_uncharge_cgroup: 20748f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 20758f34af6fSJianyu Zhan out_subpool_put: 2076d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 20778f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 2078feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 20798f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 2080b45b5bd6SDavid Gibson } 2081b45b5bd6SDavid Gibson 208274060e4dSNaoya Horiguchi /* 208374060e4dSNaoya Horiguchi * alloc_huge_page()'s wrapper which simply returns the page if allocation 208474060e4dSNaoya Horiguchi * succeeds, otherwise NULL. This function is called from new_vma_page(), 208574060e4dSNaoya Horiguchi * where no ERR_VALUE is expected to be returned. 208674060e4dSNaoya Horiguchi */ 208774060e4dSNaoya Horiguchi struct page *alloc_huge_page_noerr(struct vm_area_struct *vma, 208874060e4dSNaoya Horiguchi unsigned long addr, int avoid_reserve) 208974060e4dSNaoya Horiguchi { 209074060e4dSNaoya Horiguchi struct page *page = alloc_huge_page(vma, addr, avoid_reserve); 209174060e4dSNaoya Horiguchi if (IS_ERR(page)) 209274060e4dSNaoya Horiguchi page = NULL; 209374060e4dSNaoya Horiguchi return page; 209474060e4dSNaoya Horiguchi } 209574060e4dSNaoya Horiguchi 2096e24a1307SAneesh Kumar K.V int alloc_bootmem_huge_page(struct hstate *h) 2097e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 2098e24a1307SAneesh Kumar K.V int __alloc_bootmem_huge_page(struct hstate *h) 2099aa888a74SAndi Kleen { 2100aa888a74SAndi Kleen struct huge_bootmem_page *m; 2101b2261026SJoonsoo Kim int nr_nodes, node; 2102aa888a74SAndi Kleen 2103b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 2104aa888a74SAndi Kleen void *addr; 2105aa888a74SAndi Kleen 21068b89a116SGrygorii Strashko addr = memblock_virt_alloc_try_nid_nopanic( 21078b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 21088b89a116SGrygorii Strashko 0, BOOTMEM_ALLOC_ACCESSIBLE, node); 2109aa888a74SAndi Kleen if (addr) { 2110aa888a74SAndi Kleen /* 2111aa888a74SAndi Kleen * Use the beginning of the huge page to store the 2112aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 2113aa888a74SAndi Kleen * puts them into the mem_map). 2114aa888a74SAndi Kleen */ 2115aa888a74SAndi Kleen m = addr; 2116aa888a74SAndi Kleen goto found; 2117aa888a74SAndi Kleen } 2118aa888a74SAndi Kleen } 2119aa888a74SAndi Kleen return 0; 2120aa888a74SAndi Kleen 2121aa888a74SAndi Kleen found: 2122df994eadSLuiz Capitulino BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); 2123aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 2124aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 2125aa888a74SAndi Kleen m->hstate = h; 2126aa888a74SAndi Kleen return 1; 2127aa888a74SAndi Kleen } 2128aa888a74SAndi Kleen 2129d00181b9SKirill A. Shutemov static void __init prep_compound_huge_page(struct page *page, 2130d00181b9SKirill A. Shutemov unsigned int order) 213118229df5SAndy Whitcroft { 213218229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 213318229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 213418229df5SAndy Whitcroft else 213518229df5SAndy Whitcroft prep_compound_page(page, order); 213618229df5SAndy Whitcroft } 213718229df5SAndy Whitcroft 2138aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 2139aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 2140aa888a74SAndi Kleen { 2141aa888a74SAndi Kleen struct huge_bootmem_page *m; 2142aa888a74SAndi Kleen 2143aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 2144aa888a74SAndi Kleen struct hstate *h = m->hstate; 2145ee8f248dSBecky Bruce struct page *page; 2146ee8f248dSBecky Bruce 2147ee8f248dSBecky Bruce #ifdef CONFIG_HIGHMEM 2148ee8f248dSBecky Bruce page = pfn_to_page(m->phys >> PAGE_SHIFT); 21498b89a116SGrygorii Strashko memblock_free_late(__pa(m), 2150ee8f248dSBecky Bruce sizeof(struct huge_bootmem_page)); 2151ee8f248dSBecky Bruce #else 2152ee8f248dSBecky Bruce page = virt_to_page(m); 2153ee8f248dSBecky Bruce #endif 2154aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 215518229df5SAndy Whitcroft prep_compound_huge_page(page, h->order); 2156ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 2157aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 2158af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 2159af0fb9dfSMichal Hocko 2160b0320c7bSRafael Aquini /* 2161b0320c7bSRafael Aquini * If we had gigantic hugepages allocated at boot time, we need 2162b0320c7bSRafael Aquini * to restore the 'stolen' pages to totalram_pages in order to 2163b0320c7bSRafael Aquini * fix confusing memory reports from free(1) and another 2164b0320c7bSRafael Aquini * side-effects, like CommitLimit going negative. 2165b0320c7bSRafael Aquini */ 2166bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 21673dcc0571SJiang Liu adjust_managed_page_count(page, 1 << h->order); 2168aa888a74SAndi Kleen } 2169aa888a74SAndi Kleen } 2170aa888a74SAndi Kleen 21718faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 21721da177e4SLinus Torvalds { 21731da177e4SLinus Torvalds unsigned long i; 21741da177e4SLinus Torvalds 2175e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 2176bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 2177aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 2178aa888a74SAndi Kleen break; 21790c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 21808cebfcd0SLai Jiangshan &node_states[N_MEMORY])) 21811da177e4SLinus Torvalds break; 218269ed779aSDavid Rientjes cond_resched(); 21831da177e4SLinus Torvalds } 2184d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 2185d715cf80SLiam R. Howlett char buf[32]; 2186d715cf80SLiam R. Howlett 2187c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2188d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 2189d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 21908faa8b07SAndi Kleen h->max_huge_pages = i; 2191e5ff2159SAndi Kleen } 2192d715cf80SLiam R. Howlett } 2193e5ff2159SAndi Kleen 2194e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 2195e5ff2159SAndi Kleen { 2196e5ff2159SAndi Kleen struct hstate *h; 2197e5ff2159SAndi Kleen 2198e5ff2159SAndi Kleen for_each_hstate(h) { 2199641844f5SNaoya Horiguchi if (minimum_order > huge_page_order(h)) 2200641844f5SNaoya Horiguchi minimum_order = huge_page_order(h); 2201641844f5SNaoya Horiguchi 22028faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 2203bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 22048faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 2205e5ff2159SAndi Kleen } 2206641844f5SNaoya Horiguchi VM_BUG_ON(minimum_order == UINT_MAX); 2207e5ff2159SAndi Kleen } 2208e5ff2159SAndi Kleen 2209e5ff2159SAndi Kleen static void __init report_hugepages(void) 2210e5ff2159SAndi Kleen { 2211e5ff2159SAndi Kleen struct hstate *h; 2212e5ff2159SAndi Kleen 2213e5ff2159SAndi Kleen for_each_hstate(h) { 22144abd32dbSAndi Kleen char buf[32]; 2215c6247f72SMatthew Wilcox 2216c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2217ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 2218c6247f72SMatthew Wilcox buf, h->free_huge_pages); 2219e5ff2159SAndi Kleen } 2220e5ff2159SAndi Kleen } 2221e5ff2159SAndi Kleen 22221da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 22236ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 22246ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 22251da177e4SLinus Torvalds { 22264415cc8dSChristoph Lameter int i; 22274415cc8dSChristoph Lameter 2228bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2229aa888a74SAndi Kleen return; 2230aa888a74SAndi Kleen 22316ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 22321da177e4SLinus Torvalds struct page *page, *next; 2233a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 2234a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 2235a5516438SAndi Kleen if (count >= h->nr_huge_pages) 22366b0c880dSAdam Litke return; 22371da177e4SLinus Torvalds if (PageHighMem(page)) 22381da177e4SLinus Torvalds continue; 22391da177e4SLinus Torvalds list_del(&page->lru); 2240e5ff2159SAndi Kleen update_and_free_page(h, page); 2241a5516438SAndi Kleen h->free_huge_pages--; 2242a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 22431da177e4SLinus Torvalds } 22441da177e4SLinus Torvalds } 22451da177e4SLinus Torvalds } 22461da177e4SLinus Torvalds #else 22476ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 22486ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 22491da177e4SLinus Torvalds { 22501da177e4SLinus Torvalds } 22511da177e4SLinus Torvalds #endif 22521da177e4SLinus Torvalds 225320a0307cSWu Fengguang /* 225420a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 225520a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 225620a0307cSWu Fengguang * Returns 1 if an adjustment was made. 225720a0307cSWu Fengguang */ 22586ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 22596ae11b27SLee Schermerhorn int delta) 226020a0307cSWu Fengguang { 2261b2261026SJoonsoo Kim int nr_nodes, node; 226220a0307cSWu Fengguang 226320a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 226420a0307cSWu Fengguang 2265e8c5c824SLee Schermerhorn if (delta < 0) { 2266b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 2267b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 2268b2261026SJoonsoo Kim goto found; 2269b2261026SJoonsoo Kim } 2270b2261026SJoonsoo Kim } else { 2271b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2272b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 2273b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 2274b2261026SJoonsoo Kim goto found; 2275e8c5c824SLee Schermerhorn } 22769a76db09SLee Schermerhorn } 2277b2261026SJoonsoo Kim return 0; 227820a0307cSWu Fengguang 2279b2261026SJoonsoo Kim found: 228020a0307cSWu Fengguang h->surplus_huge_pages += delta; 2281b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 2282b2261026SJoonsoo Kim return 1; 228320a0307cSWu Fengguang } 228420a0307cSWu Fengguang 2285a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 22866ae11b27SLee Schermerhorn static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, 22876ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 22881da177e4SLinus Torvalds { 22897893d1d5SAdam Litke unsigned long min_count, ret; 22901da177e4SLinus Torvalds 2291944d9fecSLuiz Capitulino if (hstate_is_gigantic(h) && !gigantic_page_supported()) 2292aa888a74SAndi Kleen return h->max_huge_pages; 2293aa888a74SAndi Kleen 22947893d1d5SAdam Litke /* 22957893d1d5SAdam Litke * Increase the pool size 22967893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 22977893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 2298d1c3fb1fSNishanth Aravamudan * 22990c397daeSMichal Hocko * We might race with alloc_surplus_huge_page() here and be unable 2300d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 2301d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 2302d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 2303d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 23047893d1d5SAdam Litke */ 23051da177e4SLinus Torvalds spin_lock(&hugetlb_lock); 2306a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 23076ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 23087893d1d5SAdam Litke break; 23097893d1d5SAdam Litke } 23107893d1d5SAdam Litke 2311a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 23127893d1d5SAdam Litke /* 23137893d1d5SAdam Litke * If this allocation races such that we no longer need the 23147893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 23157893d1d5SAdam Litke * and reducing the surplus. 23167893d1d5SAdam Litke */ 23177893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 2318649920c6SJia He 2319649920c6SJia He /* yield cpu to avoid soft lockup */ 2320649920c6SJia He cond_resched(); 2321649920c6SJia He 23220c397daeSMichal Hocko ret = alloc_pool_huge_page(h, nodes_allowed); 23237893d1d5SAdam Litke spin_lock(&hugetlb_lock); 23247893d1d5SAdam Litke if (!ret) 23257893d1d5SAdam Litke goto out; 23267893d1d5SAdam Litke 2327536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 2328536240f2SMel Gorman if (signal_pending(current)) 2329536240f2SMel Gorman goto out; 23307893d1d5SAdam Litke } 23317893d1d5SAdam Litke 23327893d1d5SAdam Litke /* 23337893d1d5SAdam Litke * Decrease the pool size 23347893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 23357893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 23367893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 23377893d1d5SAdam Litke * to the desired size as pages become free. 2338d1c3fb1fSNishanth Aravamudan * 2339d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 2340d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 2341d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 23420c397daeSMichal Hocko * alloc_surplus_huge_page() is checking the global counter, 2343d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 2344d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 2345d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 23467893d1d5SAdam Litke */ 2347a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 23486b0c880dSAdam Litke min_count = max(count, min_count); 23496ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 2350a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 23516ae11b27SLee Schermerhorn if (!free_pool_huge_page(h, nodes_allowed, 0)) 23521da177e4SLinus Torvalds break; 235355f67141SMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 23541da177e4SLinus Torvalds } 2355a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 23566ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 23577893d1d5SAdam Litke break; 23587893d1d5SAdam Litke } 23597893d1d5SAdam Litke out: 2360a5516438SAndi Kleen ret = persistent_huge_pages(h); 23611da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 23627893d1d5SAdam Litke return ret; 23631da177e4SLinus Torvalds } 23641da177e4SLinus Torvalds 2365a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 2366a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 2367a3437870SNishanth Aravamudan 2368a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 2369a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 2370a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 2371a3437870SNishanth Aravamudan 2372a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 2373a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 2374a3437870SNishanth Aravamudan 23759a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 23769a305230SLee Schermerhorn 23779a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 2378a3437870SNishanth Aravamudan { 2379a3437870SNishanth Aravamudan int i; 23809a305230SLee Schermerhorn 2381a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 23829a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 23839a305230SLee Schermerhorn if (nidp) 23849a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 2385a3437870SNishanth Aravamudan return &hstates[i]; 23869a305230SLee Schermerhorn } 23879a305230SLee Schermerhorn 23889a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 2389a3437870SNishanth Aravamudan } 2390a3437870SNishanth Aravamudan 239106808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 2392a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2393a3437870SNishanth Aravamudan { 23949a305230SLee Schermerhorn struct hstate *h; 23959a305230SLee Schermerhorn unsigned long nr_huge_pages; 23969a305230SLee Schermerhorn int nid; 23979a305230SLee Schermerhorn 23989a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 23999a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 24009a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 24019a305230SLee Schermerhorn else 24029a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 24039a305230SLee Schermerhorn 24049a305230SLee Schermerhorn return sprintf(buf, "%lu\n", nr_huge_pages); 2405a3437870SNishanth Aravamudan } 2406adbe8726SEric B Munson 2407238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 2408238d3c13SDavid Rientjes struct hstate *h, int nid, 2409238d3c13SDavid Rientjes unsigned long count, size_t len) 2410a3437870SNishanth Aravamudan { 2411a3437870SNishanth Aravamudan int err; 2412bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); 2413a3437870SNishanth Aravamudan 2414944d9fecSLuiz Capitulino if (hstate_is_gigantic(h) && !gigantic_page_supported()) { 2415adbe8726SEric B Munson err = -EINVAL; 2416adbe8726SEric B Munson goto out; 2417adbe8726SEric B Munson } 2418adbe8726SEric B Munson 24199a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 24209a305230SLee Schermerhorn /* 24219a305230SLee Schermerhorn * global hstate attribute 24229a305230SLee Schermerhorn */ 24239a305230SLee Schermerhorn if (!(obey_mempolicy && 24249a305230SLee Schermerhorn init_nodemask_of_mempolicy(nodes_allowed))) { 242506808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 24268cebfcd0SLai Jiangshan nodes_allowed = &node_states[N_MEMORY]; 242706808b08SLee Schermerhorn } 24289a305230SLee Schermerhorn } else if (nodes_allowed) { 24299a305230SLee Schermerhorn /* 24309a305230SLee Schermerhorn * per node hstate attribute: adjust count to global, 24319a305230SLee Schermerhorn * but restrict alloc/free to the specified node. 24329a305230SLee Schermerhorn */ 24339a305230SLee Schermerhorn count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 24349a305230SLee Schermerhorn init_nodemask_of_node(nodes_allowed, nid); 24359a305230SLee Schermerhorn } else 24368cebfcd0SLai Jiangshan nodes_allowed = &node_states[N_MEMORY]; 24379a305230SLee Schermerhorn 243806808b08SLee Schermerhorn h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); 2439a3437870SNishanth Aravamudan 24408cebfcd0SLai Jiangshan if (nodes_allowed != &node_states[N_MEMORY]) 244106808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 244206808b08SLee Schermerhorn 244306808b08SLee Schermerhorn return len; 2444adbe8726SEric B Munson out: 2445adbe8726SEric B Munson NODEMASK_FREE(nodes_allowed); 2446adbe8726SEric B Munson return err; 244706808b08SLee Schermerhorn } 244806808b08SLee Schermerhorn 2449238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 2450238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 2451238d3c13SDavid Rientjes size_t len) 2452238d3c13SDavid Rientjes { 2453238d3c13SDavid Rientjes struct hstate *h; 2454238d3c13SDavid Rientjes unsigned long count; 2455238d3c13SDavid Rientjes int nid; 2456238d3c13SDavid Rientjes int err; 2457238d3c13SDavid Rientjes 2458238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 2459238d3c13SDavid Rientjes if (err) 2460238d3c13SDavid Rientjes return err; 2461238d3c13SDavid Rientjes 2462238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 2463238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 2464238d3c13SDavid Rientjes } 2465238d3c13SDavid Rientjes 246606808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 246706808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 246806808b08SLee Schermerhorn { 246906808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 247006808b08SLee Schermerhorn } 247106808b08SLee Schermerhorn 247206808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 247306808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 247406808b08SLee Schermerhorn { 2475238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 2476a3437870SNishanth Aravamudan } 2477a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 2478a3437870SNishanth Aravamudan 247906808b08SLee Schermerhorn #ifdef CONFIG_NUMA 248006808b08SLee Schermerhorn 248106808b08SLee Schermerhorn /* 248206808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 248306808b08SLee Schermerhorn * huge page alloc/free. 248406808b08SLee Schermerhorn */ 248506808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 248606808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 248706808b08SLee Schermerhorn { 248806808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 248906808b08SLee Schermerhorn } 249006808b08SLee Schermerhorn 249106808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 249206808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 249306808b08SLee Schermerhorn { 2494238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 249506808b08SLee Schermerhorn } 249606808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 249706808b08SLee Schermerhorn #endif 249806808b08SLee Schermerhorn 249906808b08SLee Schermerhorn 2500a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 2501a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2502a3437870SNishanth Aravamudan { 25039a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2504a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 2505a3437870SNishanth Aravamudan } 2506adbe8726SEric B Munson 2507a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 2508a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 2509a3437870SNishanth Aravamudan { 2510a3437870SNishanth Aravamudan int err; 2511a3437870SNishanth Aravamudan unsigned long input; 25129a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2513a3437870SNishanth Aravamudan 2514bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2515adbe8726SEric B Munson return -EINVAL; 2516adbe8726SEric B Munson 25173dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 2518a3437870SNishanth Aravamudan if (err) 251973ae31e5SEric B Munson return err; 2520a3437870SNishanth Aravamudan 2521a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 2522a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 2523a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 2524a3437870SNishanth Aravamudan 2525a3437870SNishanth Aravamudan return count; 2526a3437870SNishanth Aravamudan } 2527a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 2528a3437870SNishanth Aravamudan 2529a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 2530a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2531a3437870SNishanth Aravamudan { 25329a305230SLee Schermerhorn struct hstate *h; 25339a305230SLee Schermerhorn unsigned long free_huge_pages; 25349a305230SLee Schermerhorn int nid; 25359a305230SLee Schermerhorn 25369a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 25379a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 25389a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 25399a305230SLee Schermerhorn else 25409a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 25419a305230SLee Schermerhorn 25429a305230SLee Schermerhorn return sprintf(buf, "%lu\n", free_huge_pages); 2543a3437870SNishanth Aravamudan } 2544a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 2545a3437870SNishanth Aravamudan 2546a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 2547a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2548a3437870SNishanth Aravamudan { 25499a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2550a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->resv_huge_pages); 2551a3437870SNishanth Aravamudan } 2552a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 2553a3437870SNishanth Aravamudan 2554a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 2555a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2556a3437870SNishanth Aravamudan { 25579a305230SLee Schermerhorn struct hstate *h; 25589a305230SLee Schermerhorn unsigned long surplus_huge_pages; 25599a305230SLee Schermerhorn int nid; 25609a305230SLee Schermerhorn 25619a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 25629a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 25639a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 25649a305230SLee Schermerhorn else 25659a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 25669a305230SLee Schermerhorn 25679a305230SLee Schermerhorn return sprintf(buf, "%lu\n", surplus_huge_pages); 2568a3437870SNishanth Aravamudan } 2569a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 2570a3437870SNishanth Aravamudan 2571a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 2572a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 2573a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 2574a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 2575a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 2576a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 257706808b08SLee Schermerhorn #ifdef CONFIG_NUMA 257806808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 257906808b08SLee Schermerhorn #endif 2580a3437870SNishanth Aravamudan NULL, 2581a3437870SNishanth Aravamudan }; 2582a3437870SNishanth Aravamudan 258367e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 2584a3437870SNishanth Aravamudan .attrs = hstate_attrs, 2585a3437870SNishanth Aravamudan }; 2586a3437870SNishanth Aravamudan 2587094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 25889a305230SLee Schermerhorn struct kobject **hstate_kobjs, 258967e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 2590a3437870SNishanth Aravamudan { 2591a3437870SNishanth Aravamudan int retval; 2592972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 2593a3437870SNishanth Aravamudan 25949a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 25959a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 2596a3437870SNishanth Aravamudan return -ENOMEM; 2597a3437870SNishanth Aravamudan 25989a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 2599a3437870SNishanth Aravamudan if (retval) 26009a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 2601a3437870SNishanth Aravamudan 2602a3437870SNishanth Aravamudan return retval; 2603a3437870SNishanth Aravamudan } 2604a3437870SNishanth Aravamudan 2605a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 2606a3437870SNishanth Aravamudan { 2607a3437870SNishanth Aravamudan struct hstate *h; 2608a3437870SNishanth Aravamudan int err; 2609a3437870SNishanth Aravamudan 2610a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 2611a3437870SNishanth Aravamudan if (!hugepages_kobj) 2612a3437870SNishanth Aravamudan return; 2613a3437870SNishanth Aravamudan 2614a3437870SNishanth Aravamudan for_each_hstate(h) { 26159a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 26169a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 2617a3437870SNishanth Aravamudan if (err) 2618ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s", h->name); 2619a3437870SNishanth Aravamudan } 2620a3437870SNishanth Aravamudan } 2621a3437870SNishanth Aravamudan 26229a305230SLee Schermerhorn #ifdef CONFIG_NUMA 26239a305230SLee Schermerhorn 26249a305230SLee Schermerhorn /* 26259a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 262610fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 262710fbcf4cSKay Sievers * index of a node device or _hstate == node id. 262810fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 26299a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 26309a305230SLee Schermerhorn */ 26319a305230SLee Schermerhorn struct node_hstate { 26329a305230SLee Schermerhorn struct kobject *hugepages_kobj; 26339a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 26349a305230SLee Schermerhorn }; 2635b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 26369a305230SLee Schermerhorn 26379a305230SLee Schermerhorn /* 263810fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 26399a305230SLee Schermerhorn */ 26409a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 26419a305230SLee Schermerhorn &nr_hugepages_attr.attr, 26429a305230SLee Schermerhorn &free_hugepages_attr.attr, 26439a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 26449a305230SLee Schermerhorn NULL, 26459a305230SLee Schermerhorn }; 26469a305230SLee Schermerhorn 264767e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 26489a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 26499a305230SLee Schermerhorn }; 26509a305230SLee Schermerhorn 26519a305230SLee Schermerhorn /* 265210fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 26539a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 26549a305230SLee Schermerhorn */ 26559a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 26569a305230SLee Schermerhorn { 26579a305230SLee Schermerhorn int nid; 26589a305230SLee Schermerhorn 26599a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 26609a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 26619a305230SLee Schermerhorn int i; 26629a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 26639a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 26649a305230SLee Schermerhorn if (nidp) 26659a305230SLee Schermerhorn *nidp = nid; 26669a305230SLee Schermerhorn return &hstates[i]; 26679a305230SLee Schermerhorn } 26689a305230SLee Schermerhorn } 26699a305230SLee Schermerhorn 26709a305230SLee Schermerhorn BUG(); 26719a305230SLee Schermerhorn return NULL; 26729a305230SLee Schermerhorn } 26739a305230SLee Schermerhorn 26749a305230SLee Schermerhorn /* 267510fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 26769a305230SLee Schermerhorn * No-op if no hstate attributes attached. 26779a305230SLee Schermerhorn */ 26783cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 26799a305230SLee Schermerhorn { 26809a305230SLee Schermerhorn struct hstate *h; 268110fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 26829a305230SLee Schermerhorn 26839a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 26849b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 26859a305230SLee Schermerhorn 2686972dc4deSAneesh Kumar K.V for_each_hstate(h) { 2687972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 2688972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 2689972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 2690972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 2691972dc4deSAneesh Kumar K.V } 26929a305230SLee Schermerhorn } 26939a305230SLee Schermerhorn 26949a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 26959a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 26969a305230SLee Schermerhorn } 26979a305230SLee Schermerhorn 26989a305230SLee Schermerhorn 26999a305230SLee Schermerhorn /* 270010fbcf4cSKay Sievers * Register hstate attributes for a single node device. 27019a305230SLee Schermerhorn * No-op if attributes already registered. 27029a305230SLee Schermerhorn */ 27033cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 27049a305230SLee Schermerhorn { 27059a305230SLee Schermerhorn struct hstate *h; 270610fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 27079a305230SLee Schermerhorn int err; 27089a305230SLee Schermerhorn 27099a305230SLee Schermerhorn if (nhs->hugepages_kobj) 27109a305230SLee Schermerhorn return; /* already allocated */ 27119a305230SLee Schermerhorn 27129a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 271310fbcf4cSKay Sievers &node->dev.kobj); 27149a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 27159a305230SLee Schermerhorn return; 27169a305230SLee Schermerhorn 27179a305230SLee Schermerhorn for_each_hstate(h) { 27189a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 27199a305230SLee Schermerhorn nhs->hstate_kobjs, 27209a305230SLee Schermerhorn &per_node_hstate_attr_group); 27219a305230SLee Schermerhorn if (err) { 2722ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s for node %d\n", 272310fbcf4cSKay Sievers h->name, node->dev.id); 27249a305230SLee Schermerhorn hugetlb_unregister_node(node); 27259a305230SLee Schermerhorn break; 27269a305230SLee Schermerhorn } 27279a305230SLee Schermerhorn } 27289a305230SLee Schermerhorn } 27299a305230SLee Schermerhorn 27309a305230SLee Schermerhorn /* 27319b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 273210fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 273310fbcf4cSKay Sievers * registered their associated device by this time. 27349a305230SLee Schermerhorn */ 27357d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 27369a305230SLee Schermerhorn { 27379a305230SLee Schermerhorn int nid; 27389a305230SLee Schermerhorn 27398cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 27408732794bSWen Congyang struct node *node = node_devices[nid]; 274110fbcf4cSKay Sievers if (node->dev.id == nid) 27429a305230SLee Schermerhorn hugetlb_register_node(node); 27439a305230SLee Schermerhorn } 27449a305230SLee Schermerhorn 27459a305230SLee Schermerhorn /* 274610fbcf4cSKay Sievers * Let the node device driver know we're here so it can 27479a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 27489a305230SLee Schermerhorn */ 27499a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 27509a305230SLee Schermerhorn hugetlb_unregister_node); 27519a305230SLee Schermerhorn } 27529a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 27539a305230SLee Schermerhorn 27549a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 27559a305230SLee Schermerhorn { 27569a305230SLee Schermerhorn BUG(); 27579a305230SLee Schermerhorn if (nidp) 27589a305230SLee Schermerhorn *nidp = -1; 27599a305230SLee Schermerhorn return NULL; 27609a305230SLee Schermerhorn } 27619a305230SLee Schermerhorn 27629a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 27639a305230SLee Schermerhorn 27649a305230SLee Schermerhorn #endif 27659a305230SLee Schermerhorn 2766a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 2767a3437870SNishanth Aravamudan { 27688382d914SDavidlohr Bueso int i; 27698382d914SDavidlohr Bueso 2770457c1b27SNishanth Aravamudan if (!hugepages_supported()) 27710ef89d25SBenjamin Herrenschmidt return 0; 2772a3437870SNishanth Aravamudan 2773e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) { 2774d715cf80SLiam R. Howlett if (default_hstate_size != 0) { 2775d715cf80SLiam R. Howlett pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n", 2776d715cf80SLiam R. Howlett default_hstate_size, HPAGE_SIZE); 2777d715cf80SLiam R. Howlett } 2778d715cf80SLiam R. Howlett 2779e11bfbfcSNick Piggin default_hstate_size = HPAGE_SIZE; 2780e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) 2781a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 2782a3437870SNishanth Aravamudan } 2783972dc4deSAneesh Kumar K.V default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); 2784f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 2785f8b74815SVaishali Thakkar if (!default_hstate.max_huge_pages) 2786e11bfbfcSNick Piggin default_hstate.max_huge_pages = default_hstate_max_huge_pages; 2787f8b74815SVaishali Thakkar } 2788a3437870SNishanth Aravamudan 2789a3437870SNishanth Aravamudan hugetlb_init_hstates(); 2790aa888a74SAndi Kleen gather_bootmem_prealloc(); 2791a3437870SNishanth Aravamudan report_hugepages(); 2792a3437870SNishanth Aravamudan 2793a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 27949a305230SLee Schermerhorn hugetlb_register_all_nodes(); 27957179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 27969a305230SLee Schermerhorn 27978382d914SDavidlohr Bueso #ifdef CONFIG_SMP 27988382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 27998382d914SDavidlohr Bueso #else 28008382d914SDavidlohr Bueso num_fault_mutexes = 1; 28018382d914SDavidlohr Bueso #endif 2802c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 28038382d914SDavidlohr Bueso kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL); 2804c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 28058382d914SDavidlohr Bueso 28068382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 2807c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 2808a3437870SNishanth Aravamudan return 0; 2809a3437870SNishanth Aravamudan } 28103e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 2811a3437870SNishanth Aravamudan 2812a3437870SNishanth Aravamudan /* Should be called on processing a hugepagesz=... option */ 28139fee021dSVaishali Thakkar void __init hugetlb_bad_size(void) 28149fee021dSVaishali Thakkar { 28159fee021dSVaishali Thakkar parsed_valid_hugepagesz = false; 28169fee021dSVaishali Thakkar } 28179fee021dSVaishali Thakkar 2818d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 2819a3437870SNishanth Aravamudan { 2820a3437870SNishanth Aravamudan struct hstate *h; 28218faa8b07SAndi Kleen unsigned long i; 28228faa8b07SAndi Kleen 2823a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 2824598d8091SJoe Perches pr_warn("hugepagesz= specified twice, ignoring\n"); 2825a3437870SNishanth Aravamudan return; 2826a3437870SNishanth Aravamudan } 282747d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 2828a3437870SNishanth Aravamudan BUG_ON(order == 0); 282947d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 2830a3437870SNishanth Aravamudan h->order = order; 2831a3437870SNishanth Aravamudan h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 28328faa8b07SAndi Kleen h->nr_huge_pages = 0; 28338faa8b07SAndi Kleen h->free_huge_pages = 0; 28348faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 28358faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 28360edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 283754f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 283854f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 2839a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 2840a3437870SNishanth Aravamudan huge_page_size(h)/1024); 28418faa8b07SAndi Kleen 2842a3437870SNishanth Aravamudan parsed_hstate = h; 2843a3437870SNishanth Aravamudan } 2844a3437870SNishanth Aravamudan 2845e11bfbfcSNick Piggin static int __init hugetlb_nrpages_setup(char *s) 2846a3437870SNishanth Aravamudan { 2847a3437870SNishanth Aravamudan unsigned long *mhp; 28488faa8b07SAndi Kleen static unsigned long *last_mhp; 2849a3437870SNishanth Aravamudan 28509fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 28519fee021dSVaishali Thakkar pr_warn("hugepages = %s preceded by " 28529fee021dSVaishali Thakkar "an unsupported hugepagesz, ignoring\n", s); 28539fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 28549fee021dSVaishali Thakkar return 1; 28559fee021dSVaishali Thakkar } 2856a3437870SNishanth Aravamudan /* 285747d38344SAneesh Kumar K.V * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, 2858a3437870SNishanth Aravamudan * so this hugepages= parameter goes to the "default hstate". 2859a3437870SNishanth Aravamudan */ 28609fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 2861a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 2862a3437870SNishanth Aravamudan else 2863a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 2864a3437870SNishanth Aravamudan 28658faa8b07SAndi Kleen if (mhp == last_mhp) { 2866598d8091SJoe Perches pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n"); 28678faa8b07SAndi Kleen return 1; 28688faa8b07SAndi Kleen } 28698faa8b07SAndi Kleen 2870a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 2871a3437870SNishanth Aravamudan *mhp = 0; 2872a3437870SNishanth Aravamudan 28738faa8b07SAndi Kleen /* 28748faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 28758faa8b07SAndi Kleen * But we need to allocate >= MAX_ORDER hstates here early to still 28768faa8b07SAndi Kleen * use the bootmem allocator. 28778faa8b07SAndi Kleen */ 287847d38344SAneesh Kumar K.V if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) 28798faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 28808faa8b07SAndi Kleen 28818faa8b07SAndi Kleen last_mhp = mhp; 28828faa8b07SAndi Kleen 2883a3437870SNishanth Aravamudan return 1; 2884a3437870SNishanth Aravamudan } 2885e11bfbfcSNick Piggin __setup("hugepages=", hugetlb_nrpages_setup); 2886e11bfbfcSNick Piggin 2887e11bfbfcSNick Piggin static int __init hugetlb_default_setup(char *s) 2888e11bfbfcSNick Piggin { 2889e11bfbfcSNick Piggin default_hstate_size = memparse(s, &s); 2890e11bfbfcSNick Piggin return 1; 2891e11bfbfcSNick Piggin } 2892e11bfbfcSNick Piggin __setup("default_hugepagesz=", hugetlb_default_setup); 2893a3437870SNishanth Aravamudan 28948a213460SNishanth Aravamudan static unsigned int cpuset_mems_nr(unsigned int *array) 28958a213460SNishanth Aravamudan { 28968a213460SNishanth Aravamudan int node; 28978a213460SNishanth Aravamudan unsigned int nr = 0; 28988a213460SNishanth Aravamudan 28998a213460SNishanth Aravamudan for_each_node_mask(node, cpuset_current_mems_allowed) 29008a213460SNishanth Aravamudan nr += array[node]; 29018a213460SNishanth Aravamudan 29028a213460SNishanth Aravamudan return nr; 29038a213460SNishanth Aravamudan } 29048a213460SNishanth Aravamudan 29058a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 290606808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 290706808b08SLee Schermerhorn struct ctl_table *table, int write, 290806808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 29091da177e4SLinus Torvalds { 2910e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 2911238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 291208d4a246SMichal Hocko int ret; 2913e5ff2159SAndi Kleen 2914457c1b27SNishanth Aravamudan if (!hugepages_supported()) 291586613628SJan Stancek return -EOPNOTSUPP; 2916457c1b27SNishanth Aravamudan 2917e5ff2159SAndi Kleen table->data = &tmp; 2918e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 291908d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 292008d4a246SMichal Hocko if (ret) 292108d4a246SMichal Hocko goto out; 2922e5ff2159SAndi Kleen 2923238d3c13SDavid Rientjes if (write) 2924238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 2925238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 292608d4a246SMichal Hocko out: 292708d4a246SMichal Hocko return ret; 29281da177e4SLinus Torvalds } 2929396faf03SMel Gorman 293006808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 293106808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 293206808b08SLee Schermerhorn { 293306808b08SLee Schermerhorn 293406808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 293506808b08SLee Schermerhorn buffer, length, ppos); 293606808b08SLee Schermerhorn } 293706808b08SLee Schermerhorn 293806808b08SLee Schermerhorn #ifdef CONFIG_NUMA 293906808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 294006808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 294106808b08SLee Schermerhorn { 294206808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 294306808b08SLee Schermerhorn buffer, length, ppos); 294406808b08SLee Schermerhorn } 294506808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 294606808b08SLee Schermerhorn 2947a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 29488d65af78SAlexey Dobriyan void __user *buffer, 2949a3d0c6aaSNishanth Aravamudan size_t *length, loff_t *ppos) 2950a3d0c6aaSNishanth Aravamudan { 2951a5516438SAndi Kleen struct hstate *h = &default_hstate; 2952e5ff2159SAndi Kleen unsigned long tmp; 295308d4a246SMichal Hocko int ret; 2954e5ff2159SAndi Kleen 2955457c1b27SNishanth Aravamudan if (!hugepages_supported()) 295686613628SJan Stancek return -EOPNOTSUPP; 2957457c1b27SNishanth Aravamudan 2958e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 2959e5ff2159SAndi Kleen 2960bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 2961adbe8726SEric B Munson return -EINVAL; 2962adbe8726SEric B Munson 2963e5ff2159SAndi Kleen table->data = &tmp; 2964e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 296508d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 296608d4a246SMichal Hocko if (ret) 296708d4a246SMichal Hocko goto out; 2968e5ff2159SAndi Kleen 2969e5ff2159SAndi Kleen if (write) { 2970064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 2971e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 2972a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 2973e5ff2159SAndi Kleen } 297408d4a246SMichal Hocko out: 297508d4a246SMichal Hocko return ret; 2976a3d0c6aaSNishanth Aravamudan } 2977a3d0c6aaSNishanth Aravamudan 29781da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 29791da177e4SLinus Torvalds 2980e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 29811da177e4SLinus Torvalds { 2982fcb2b0c5SRoman Gushchin struct hstate *h; 2983fcb2b0c5SRoman Gushchin unsigned long total = 0; 2984fcb2b0c5SRoman Gushchin 2985457c1b27SNishanth Aravamudan if (!hugepages_supported()) 2986457c1b27SNishanth Aravamudan return; 2987fcb2b0c5SRoman Gushchin 2988fcb2b0c5SRoman Gushchin for_each_hstate(h) { 2989fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 2990fcb2b0c5SRoman Gushchin 2991fcb2b0c5SRoman Gushchin total += (PAGE_SIZE << huge_page_order(h)) * count; 2992fcb2b0c5SRoman Gushchin 2993fcb2b0c5SRoman Gushchin if (h == &default_hstate) 2994e1759c21SAlexey Dobriyan seq_printf(m, 29951da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 29961da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 2997b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 29987893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 29994f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 3000fcb2b0c5SRoman Gushchin count, 3001a5516438SAndi Kleen h->free_huge_pages, 3002a5516438SAndi Kleen h->resv_huge_pages, 3003a5516438SAndi Kleen h->surplus_huge_pages, 3004fcb2b0c5SRoman Gushchin (PAGE_SIZE << huge_page_order(h)) / 1024); 3005fcb2b0c5SRoman Gushchin } 3006fcb2b0c5SRoman Gushchin 3007fcb2b0c5SRoman Gushchin seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024); 30081da177e4SLinus Torvalds } 30091da177e4SLinus Torvalds 30101da177e4SLinus Torvalds int hugetlb_report_node_meminfo(int nid, char *buf) 30111da177e4SLinus Torvalds { 3012a5516438SAndi Kleen struct hstate *h = &default_hstate; 3013457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3014457c1b27SNishanth Aravamudan return 0; 30151da177e4SLinus Torvalds return sprintf(buf, 30161da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 3017a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 3018a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 3019a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 3020a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 3021a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 30221da177e4SLinus Torvalds } 30231da177e4SLinus Torvalds 3024949f7ec5SDavid Rientjes void hugetlb_show_meminfo(void) 3025949f7ec5SDavid Rientjes { 3026949f7ec5SDavid Rientjes struct hstate *h; 3027949f7ec5SDavid Rientjes int nid; 3028949f7ec5SDavid Rientjes 3029457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3030457c1b27SNishanth Aravamudan return; 3031457c1b27SNishanth Aravamudan 3032949f7ec5SDavid Rientjes for_each_node_state(nid, N_MEMORY) 3033949f7ec5SDavid Rientjes for_each_hstate(h) 3034949f7ec5SDavid Rientjes pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 3035949f7ec5SDavid Rientjes nid, 3036949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 3037949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 3038949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 3039949f7ec5SDavid Rientjes 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 3040949f7ec5SDavid Rientjes } 3041949f7ec5SDavid Rientjes 30425d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 30435d317b2bSNaoya Horiguchi { 30445d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 30455d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 30465d317b2bSNaoya Horiguchi } 30475d317b2bSNaoya Horiguchi 30481da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 30491da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 30501da177e4SLinus Torvalds { 3051d0028588SWanpeng Li struct hstate *h; 3052d0028588SWanpeng Li unsigned long nr_total_pages = 0; 3053d0028588SWanpeng Li 3054d0028588SWanpeng Li for_each_hstate(h) 3055d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 3056d0028588SWanpeng Li return nr_total_pages; 30571da177e4SLinus Torvalds } 30581da177e4SLinus Torvalds 3059a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 3060fc1b8a73SMel Gorman { 3061fc1b8a73SMel Gorman int ret = -ENOMEM; 3062fc1b8a73SMel Gorman 3063fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 3064fc1b8a73SMel Gorman /* 3065fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 3066fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 3067fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 3068fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 3069fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 3070fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 3071fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 3072fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 3073fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 3074fc1b8a73SMel Gorman * 3075fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 3076fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 3077fc1b8a73SMel Gorman * we fall back to check against current free page availability as 3078fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 3079fc1b8a73SMel Gorman * semantics that cpuset has. 3080fc1b8a73SMel Gorman */ 3081fc1b8a73SMel Gorman if (delta > 0) { 3082a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 3083fc1b8a73SMel Gorman goto out; 3084fc1b8a73SMel Gorman 3085a5516438SAndi Kleen if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 3086a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 3087fc1b8a73SMel Gorman goto out; 3088fc1b8a73SMel Gorman } 3089fc1b8a73SMel Gorman } 3090fc1b8a73SMel Gorman 3091fc1b8a73SMel Gorman ret = 0; 3092fc1b8a73SMel Gorman if (delta < 0) 3093a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 3094fc1b8a73SMel Gorman 3095fc1b8a73SMel Gorman out: 3096fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 3097fc1b8a73SMel Gorman return ret; 3098fc1b8a73SMel Gorman } 3099fc1b8a73SMel Gorman 310084afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 310184afd99bSAndy Whitcroft { 3102f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 310384afd99bSAndy Whitcroft 310484afd99bSAndy Whitcroft /* 310584afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 310684afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 310784afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 310825985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 310984afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 311084afd99bSAndy Whitcroft * new reference here without additional locking. 311184afd99bSAndy Whitcroft */ 31124e35f483SJoonsoo Kim if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 3113f522c3acSJoonsoo Kim kref_get(&resv->refs); 311484afd99bSAndy Whitcroft } 311584afd99bSAndy Whitcroft 3116a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 3117a1e78772SMel Gorman { 3118a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3119f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 312090481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 31214e35f483SJoonsoo Kim unsigned long reserve, start, end; 31221c5ecae3SMike Kravetz long gbl_reserve; 312384afd99bSAndy Whitcroft 31244e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 31254e35f483SJoonsoo Kim return; 31264e35f483SJoonsoo Kim 3127a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 3128a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 312984afd99bSAndy Whitcroft 31304e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 313184afd99bSAndy Whitcroft 3132f031dd27SJoonsoo Kim kref_put(&resv->refs, resv_map_release); 313384afd99bSAndy Whitcroft 31347251ff78SAdam Litke if (reserve) { 31351c5ecae3SMike Kravetz /* 31361c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 31371c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 31381c5ecae3SMike Kravetz */ 31391c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 31401c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 31417251ff78SAdam Litke } 3142a1e78772SMel Gorman } 3143a1e78772SMel Gorman 314431383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 314531383c68SDan Williams { 314631383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 314731383c68SDan Williams return -EINVAL; 314831383c68SDan Williams return 0; 314931383c68SDan Williams } 315031383c68SDan Williams 31511da177e4SLinus Torvalds /* 31521da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 31531da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 31541da177e4SLinus Torvalds * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 31551da177e4SLinus Torvalds * this far. 31561da177e4SLinus Torvalds */ 315711bac800SDave Jiang static int hugetlb_vm_op_fault(struct vm_fault *vmf) 31581da177e4SLinus Torvalds { 31591da177e4SLinus Torvalds BUG(); 3160d0217ac0SNick Piggin return 0; 31611da177e4SLinus Torvalds } 31621da177e4SLinus Torvalds 3163f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 3164d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 316584afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 3166a1e78772SMel Gorman .close = hugetlb_vm_op_close, 316731383c68SDan Williams .split = hugetlb_vm_op_split, 31681da177e4SLinus Torvalds }; 31691da177e4SLinus Torvalds 31701e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 31711e8f889bSDavid Gibson int writable) 317263551ae0SDavid Gibson { 317363551ae0SDavid Gibson pte_t entry; 317463551ae0SDavid Gibson 31751e8f889bSDavid Gibson if (writable) { 3176106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 3177106c992aSGerald Schaefer vma->vm_page_prot))); 317863551ae0SDavid Gibson } else { 3179106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 3180106c992aSGerald Schaefer vma->vm_page_prot)); 318163551ae0SDavid Gibson } 318263551ae0SDavid Gibson entry = pte_mkyoung(entry); 318363551ae0SDavid Gibson entry = pte_mkhuge(entry); 3184d9ed9faaSChris Metcalf entry = arch_make_huge_pte(entry, vma, page, writable); 318563551ae0SDavid Gibson 318663551ae0SDavid Gibson return entry; 318763551ae0SDavid Gibson } 318863551ae0SDavid Gibson 31891e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 31901e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 31911e8f889bSDavid Gibson { 31921e8f889bSDavid Gibson pte_t entry; 31931e8f889bSDavid Gibson 3194106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 319532f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 31964b3073e1SRussell King update_mmu_cache(vma, address, ptep); 31971e8f889bSDavid Gibson } 31981e8f889bSDavid Gibson 3199d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 32004a705fefSNaoya Horiguchi { 32014a705fefSNaoya Horiguchi swp_entry_t swp; 32024a705fefSNaoya Horiguchi 32034a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 3204d5ed7444SAneesh Kumar K.V return false; 32054a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 32064a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_migration_entry(swp)) 3207d5ed7444SAneesh Kumar K.V return true; 32084a705fefSNaoya Horiguchi else 3209d5ed7444SAneesh Kumar K.V return false; 32104a705fefSNaoya Horiguchi } 32114a705fefSNaoya Horiguchi 32124a705fefSNaoya Horiguchi static int is_hugetlb_entry_hwpoisoned(pte_t pte) 32134a705fefSNaoya Horiguchi { 32144a705fefSNaoya Horiguchi swp_entry_t swp; 32154a705fefSNaoya Horiguchi 32164a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 32174a705fefSNaoya Horiguchi return 0; 32184a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 32194a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_hwpoison_entry(swp)) 32204a705fefSNaoya Horiguchi return 1; 32214a705fefSNaoya Horiguchi else 32224a705fefSNaoya Horiguchi return 0; 32234a705fefSNaoya Horiguchi } 32241e8f889bSDavid Gibson 322563551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 322663551ae0SDavid Gibson struct vm_area_struct *vma) 322763551ae0SDavid Gibson { 322863551ae0SDavid Gibson pte_t *src_pte, *dst_pte, entry; 322963551ae0SDavid Gibson struct page *ptepage; 32301c59827dSHugh Dickins unsigned long addr; 32311e8f889bSDavid Gibson int cow; 3232a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3233a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 3234e8569dd2SAndreas Sandberg unsigned long mmun_start; /* For mmu_notifiers */ 3235e8569dd2SAndreas Sandberg unsigned long mmun_end; /* For mmu_notifiers */ 3236e8569dd2SAndreas Sandberg int ret = 0; 32371e8f889bSDavid Gibson 32381e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 323963551ae0SDavid Gibson 3240e8569dd2SAndreas Sandberg mmun_start = vma->vm_start; 3241e8569dd2SAndreas Sandberg mmun_end = vma->vm_end; 3242e8569dd2SAndreas Sandberg if (cow) 3243e8569dd2SAndreas Sandberg mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end); 3244e8569dd2SAndreas Sandberg 3245a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 3246cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 32477868a208SPunit Agrawal src_pte = huge_pte_offset(src, addr, sz); 3248c74df32cSHugh Dickins if (!src_pte) 3249c74df32cSHugh Dickins continue; 3250a5516438SAndi Kleen dst_pte = huge_pte_alloc(dst, addr, sz); 3251e8569dd2SAndreas Sandberg if (!dst_pte) { 3252e8569dd2SAndreas Sandberg ret = -ENOMEM; 3253e8569dd2SAndreas Sandberg break; 3254e8569dd2SAndreas Sandberg } 3255c5c99429SLarry Woodman 3256c5c99429SLarry Woodman /* If the pagetables are shared don't copy or take references */ 3257c5c99429SLarry Woodman if (dst_pte == src_pte) 3258c5c99429SLarry Woodman continue; 3259c5c99429SLarry Woodman 3260cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 3261cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 3262cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 32634a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 32644a705fefSNaoya Horiguchi if (huge_pte_none(entry)) { /* skip none entry */ 32654a705fefSNaoya Horiguchi ; 32664a705fefSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry) || 32674a705fefSNaoya Horiguchi is_hugetlb_entry_hwpoisoned(entry))) { 32684a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 32694a705fefSNaoya Horiguchi 32704a705fefSNaoya Horiguchi if (is_write_migration_entry(swp_entry) && cow) { 32714a705fefSNaoya Horiguchi /* 32724a705fefSNaoya Horiguchi * COW mappings require pages in both 32734a705fefSNaoya Horiguchi * parent and child to be set to read. 32744a705fefSNaoya Horiguchi */ 32754a705fefSNaoya Horiguchi make_migration_entry_read(&swp_entry); 32764a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 3277e5251fd4SPunit Agrawal set_huge_swap_pte_at(src, addr, src_pte, 3278e5251fd4SPunit Agrawal entry, sz); 32794a705fefSNaoya Horiguchi } 3280e5251fd4SPunit Agrawal set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); 32814a705fefSNaoya Horiguchi } else { 328234ee645eSJoerg Roedel if (cow) { 32830f10851eSJérôme Glisse /* 32840f10851eSJérôme Glisse * No need to notify as we are downgrading page 32850f10851eSJérôme Glisse * table protection not changing it to point 32860f10851eSJérôme Glisse * to a new page. 32870f10851eSJérôme Glisse * 32880f10851eSJérôme Glisse * See Documentation/vm/mmu_notifier.txt 32890f10851eSJérôme Glisse */ 32907f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 329134ee645eSJoerg Roedel } 32920253d634SNaoya Horiguchi entry = huge_ptep_get(src_pte); 329363551ae0SDavid Gibson ptepage = pte_page(entry); 329463551ae0SDavid Gibson get_page(ptepage); 329553f9263bSKirill A. Shutemov page_dup_rmap(ptepage, true); 329663551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 32975d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), dst); 32981c59827dSHugh Dickins } 3299cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 3300cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 330163551ae0SDavid Gibson } 330263551ae0SDavid Gibson 3303e8569dd2SAndreas Sandberg if (cow) 3304e8569dd2SAndreas Sandberg mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end); 3305e8569dd2SAndreas Sandberg 3306e8569dd2SAndreas Sandberg return ret; 330763551ae0SDavid Gibson } 330863551ae0SDavid Gibson 330924669e58SAneesh Kumar K.V void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 331024669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 331124669e58SAneesh Kumar K.V struct page *ref_page) 331263551ae0SDavid Gibson { 331363551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 331463551ae0SDavid Gibson unsigned long address; 3315c7546f8fSDavid Gibson pte_t *ptep; 331663551ae0SDavid Gibson pte_t pte; 3317cb900f41SKirill A. Shutemov spinlock_t *ptl; 331863551ae0SDavid Gibson struct page *page; 3319a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3320a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 33212ec74c3eSSagi Grimberg const unsigned long mmun_start = start; /* For mmu_notifiers */ 33222ec74c3eSSagi Grimberg const unsigned long mmun_end = end; /* For mmu_notifiers */ 3323a5516438SAndi Kleen 332463551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 3325a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 3326a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 332763551ae0SDavid Gibson 332807e32661SAneesh Kumar K.V /* 332907e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 333007e32661SAneesh Kumar K.V * to huge page. 333107e32661SAneesh Kumar K.V */ 333207e32661SAneesh Kumar K.V tlb_remove_check_page_size_change(tlb, sz); 333324669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 33342ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 3335569f48b8SHillf Danton address = start; 3336569f48b8SHillf Danton for (; address < end; address += sz) { 33377868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, sz); 3338c7546f8fSDavid Gibson if (!ptep) 3339c7546f8fSDavid Gibson continue; 3340c7546f8fSDavid Gibson 3341cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 334231d49da5SAneesh Kumar K.V if (huge_pmd_unshare(mm, &address, ptep)) { 334331d49da5SAneesh Kumar K.V spin_unlock(ptl); 334431d49da5SAneesh Kumar K.V continue; 334531d49da5SAneesh Kumar K.V } 334639dde65cSChen, Kenneth W 33476629326bSHillf Danton pte = huge_ptep_get(ptep); 334831d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 334931d49da5SAneesh Kumar K.V spin_unlock(ptl); 335031d49da5SAneesh Kumar K.V continue; 335131d49da5SAneesh Kumar K.V } 33526629326bSHillf Danton 33536629326bSHillf Danton /* 33549fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 33559fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 33566629326bSHillf Danton */ 33579fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 33589386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 335931d49da5SAneesh Kumar K.V spin_unlock(ptl); 336031d49da5SAneesh Kumar K.V continue; 33618c4894c6SNaoya Horiguchi } 33626629326bSHillf Danton 33636629326bSHillf Danton page = pte_page(pte); 336404f2cbe3SMel Gorman /* 336504f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 336604f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 336704f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 336804f2cbe3SMel Gorman */ 336904f2cbe3SMel Gorman if (ref_page) { 337031d49da5SAneesh Kumar K.V if (page != ref_page) { 337131d49da5SAneesh Kumar K.V spin_unlock(ptl); 337231d49da5SAneesh Kumar K.V continue; 337331d49da5SAneesh Kumar K.V } 337404f2cbe3SMel Gorman /* 337504f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 337604f2cbe3SMel Gorman * future faults in this VMA will fail rather than 337704f2cbe3SMel Gorman * looking like data was lost 337804f2cbe3SMel Gorman */ 337904f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 338004f2cbe3SMel Gorman } 338104f2cbe3SMel Gorman 3382c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 3383b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 3384106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 33856649a386SKen Chen set_page_dirty(page); 33869e81130bSHillf Danton 33875d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 3388d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 338931d49da5SAneesh Kumar K.V 3390cb900f41SKirill A. Shutemov spin_unlock(ptl); 3391e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 339224669e58SAneesh Kumar K.V /* 339331d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 339424669e58SAneesh Kumar K.V */ 339531d49da5SAneesh Kumar K.V if (ref_page) 339631d49da5SAneesh Kumar K.V break; 3397fe1668aeSChen, Kenneth W } 33982ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 339924669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 34001da177e4SLinus Torvalds } 340163551ae0SDavid Gibson 3402d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 3403d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 3404d833352aSMel Gorman unsigned long end, struct page *ref_page) 3405d833352aSMel Gorman { 3406d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 3407d833352aSMel Gorman 3408d833352aSMel Gorman /* 3409d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 3410d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 3411d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 3412d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 3413d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 3414c8c06efaSDavidlohr Bueso * is to clear it before releasing the i_mmap_rwsem. This works 3415d833352aSMel Gorman * because in the context this is called, the VMA is about to be 3416c8c06efaSDavidlohr Bueso * destroyed and the i_mmap_rwsem is held. 3417d833352aSMel Gorman */ 3418d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 3419d833352aSMel Gorman } 3420d833352aSMel Gorman 3421502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 342204f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 3423502717f4SChen, Kenneth W { 342424669e58SAneesh Kumar K.V struct mm_struct *mm; 342524669e58SAneesh Kumar K.V struct mmu_gather tlb; 342624669e58SAneesh Kumar K.V 342724669e58SAneesh Kumar K.V mm = vma->vm_mm; 342824669e58SAneesh Kumar K.V 34292b047252SLinus Torvalds tlb_gather_mmu(&tlb, mm, start, end); 343024669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 343124669e58SAneesh Kumar K.V tlb_finish_mmu(&tlb, start, end); 3432502717f4SChen, Kenneth W } 3433502717f4SChen, Kenneth W 343404f2cbe3SMel Gorman /* 343504f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 343604f2cbe3SMel Gorman * mappping it owns the reserve page for. The intention is to unmap the page 343704f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 343804f2cbe3SMel Gorman * same region. 343904f2cbe3SMel Gorman */ 34402f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 34412a4b3dedSHarvey Harrison struct page *page, unsigned long address) 344204f2cbe3SMel Gorman { 34437526674dSAdam Litke struct hstate *h = hstate_vma(vma); 344404f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 344504f2cbe3SMel Gorman struct address_space *mapping; 344604f2cbe3SMel Gorman pgoff_t pgoff; 344704f2cbe3SMel Gorman 344804f2cbe3SMel Gorman /* 344904f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 345004f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 345104f2cbe3SMel Gorman */ 34527526674dSAdam Litke address = address & huge_page_mask(h); 345336e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 345436e4f20aSMichal Hocko vma->vm_pgoff; 345593c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 345604f2cbe3SMel Gorman 34574eb2b1dcSMel Gorman /* 34584eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 34594eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 34604eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 34614eb2b1dcSMel Gorman */ 346283cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 34636b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 346404f2cbe3SMel Gorman /* Do not unmap the current VMA */ 346504f2cbe3SMel Gorman if (iter_vma == vma) 346604f2cbe3SMel Gorman continue; 346704f2cbe3SMel Gorman 346804f2cbe3SMel Gorman /* 34692f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 34702f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 34712f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 34722f84a899SMel Gorman */ 34732f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 34742f84a899SMel Gorman continue; 34752f84a899SMel Gorman 34762f84a899SMel Gorman /* 347704f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 347804f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 347904f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 348004f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 348104f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 348204f2cbe3SMel Gorman */ 348304f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 348424669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 348524669e58SAneesh Kumar K.V address + huge_page_size(h), page); 348604f2cbe3SMel Gorman } 348783cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 348804f2cbe3SMel Gorman } 348904f2cbe3SMel Gorman 34900fe6e20bSNaoya Horiguchi /* 34910fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 3492ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 3493ef009b25SMichal Hocko * cannot race with other handlers or page migration. 3494ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 34950fe6e20bSNaoya Horiguchi */ 34961e8f889bSDavid Gibson static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 34973999f52eSAneesh Kumar K.V unsigned long address, pte_t *ptep, 3498cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 34991e8f889bSDavid Gibson { 35003999f52eSAneesh Kumar K.V pte_t pte; 3501a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 35021e8f889bSDavid Gibson struct page *old_page, *new_page; 3503ad4404a2SDavidlohr Bueso int ret = 0, outside_reserve = 0; 35042ec74c3eSSagi Grimberg unsigned long mmun_start; /* For mmu_notifiers */ 35052ec74c3eSSagi Grimberg unsigned long mmun_end; /* For mmu_notifiers */ 35061e8f889bSDavid Gibson 35073999f52eSAneesh Kumar K.V pte = huge_ptep_get(ptep); 35081e8f889bSDavid Gibson old_page = pte_page(pte); 35091e8f889bSDavid Gibson 351004f2cbe3SMel Gorman retry_avoidcopy: 35111e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 35121e8f889bSDavid Gibson * and just make the page writable */ 351337a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 35145a49973dSHugh Dickins page_move_anon_rmap(old_page, vma); 35151e8f889bSDavid Gibson set_huge_ptep_writable(vma, address, ptep); 351683c54070SNick Piggin return 0; 35171e8f889bSDavid Gibson } 35181e8f889bSDavid Gibson 351904f2cbe3SMel Gorman /* 352004f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 352104f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 352204f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 352304f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 352404f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 352504f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 352604f2cbe3SMel Gorman * of the full address range. 352704f2cbe3SMel Gorman */ 35285944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 352904f2cbe3SMel Gorman old_page != pagecache_page) 353004f2cbe3SMel Gorman outside_reserve = 1; 353104f2cbe3SMel Gorman 353209cbfeafSKirill A. Shutemov get_page(old_page); 3533b76c8cfbSLarry Woodman 3534ad4404a2SDavidlohr Bueso /* 3535ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 3536ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 3537ad4404a2SDavidlohr Bueso */ 3538cb900f41SKirill A. Shutemov spin_unlock(ptl); 353904f2cbe3SMel Gorman new_page = alloc_huge_page(vma, address, outside_reserve); 35401e8f889bSDavid Gibson 35412fc39cecSAdam Litke if (IS_ERR(new_page)) { 354204f2cbe3SMel Gorman /* 354304f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 354404f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 354504f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 354604f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 354704f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 354804f2cbe3SMel Gorman */ 354904f2cbe3SMel Gorman if (outside_reserve) { 355009cbfeafSKirill A. Shutemov put_page(old_page); 355104f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 35522f4612afSDavidlohr Bueso unmap_ref_private(mm, vma, old_page, address); 355304f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 3554cb900f41SKirill A. Shutemov spin_lock(ptl); 35557868a208SPunit Agrawal ptep = huge_pte_offset(mm, address & huge_page_mask(h), 35567868a208SPunit Agrawal huge_page_size(h)); 3557a9af0c5dSNaoya Horiguchi if (likely(ptep && 3558a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 355904f2cbe3SMel Gorman goto retry_avoidcopy; 3560a734bcc8SHillf Danton /* 3561cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 3562cb900f41SKirill A. Shutemov * lock, and our job is done. 3563a734bcc8SHillf Danton */ 3564a734bcc8SHillf Danton return 0; 356504f2cbe3SMel Gorman } 356604f2cbe3SMel Gorman 3567ad4404a2SDavidlohr Bueso ret = (PTR_ERR(new_page) == -ENOMEM) ? 3568ad4404a2SDavidlohr Bueso VM_FAULT_OOM : VM_FAULT_SIGBUS; 3569ad4404a2SDavidlohr Bueso goto out_release_old; 35701e8f889bSDavid Gibson } 35711e8f889bSDavid Gibson 35720fe6e20bSNaoya Horiguchi /* 35730fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 35740fe6e20bSNaoya Horiguchi * anon_vma prepared. 35750fe6e20bSNaoya Horiguchi */ 357644e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 3577ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 3578ad4404a2SDavidlohr Bueso goto out_release_all; 357944e2aa93SDean Nelson } 35800fe6e20bSNaoya Horiguchi 358147ad8475SAndrea Arcangeli copy_user_huge_page(new_page, old_page, address, vma, 358247ad8475SAndrea Arcangeli pages_per_huge_page(h)); 35830ed361deSNick Piggin __SetPageUptodate(new_page); 3584bcc54222SNaoya Horiguchi set_page_huge_active(new_page); 35851e8f889bSDavid Gibson 35862ec74c3eSSagi Grimberg mmun_start = address & huge_page_mask(h); 35872ec74c3eSSagi Grimberg mmun_end = mmun_start + huge_page_size(h); 35882ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 3589ad4404a2SDavidlohr Bueso 3590b76c8cfbSLarry Woodman /* 3591cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 3592b76c8cfbSLarry Woodman * before the page tables are altered 3593b76c8cfbSLarry Woodman */ 3594cb900f41SKirill A. Shutemov spin_lock(ptl); 35957868a208SPunit Agrawal ptep = huge_pte_offset(mm, address & huge_page_mask(h), 35967868a208SPunit Agrawal huge_page_size(h)); 3597a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 359807443a85SJoonsoo Kim ClearPagePrivate(new_page); 359907443a85SJoonsoo Kim 36001e8f889bSDavid Gibson /* Break COW */ 36018fe627ecSGerald Schaefer huge_ptep_clear_flush(vma, address, ptep); 360234ee645eSJoerg Roedel mmu_notifier_invalidate_range(mm, mmun_start, mmun_end); 36031e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, 36041e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 3605d281ee61SKirill A. Shutemov page_remove_rmap(old_page, true); 3606cd67f0d2SNaoya Horiguchi hugepage_add_new_anon_rmap(new_page, vma, address); 36071e8f889bSDavid Gibson /* Make the old page be freed below */ 36081e8f889bSDavid Gibson new_page = old_page; 36091e8f889bSDavid Gibson } 3610cb900f41SKirill A. Shutemov spin_unlock(ptl); 36112ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 3612ad4404a2SDavidlohr Bueso out_release_all: 361396b96a96SMike Kravetz restore_reserve_on_error(h, vma, address, new_page); 361409cbfeafSKirill A. Shutemov put_page(new_page); 3615ad4404a2SDavidlohr Bueso out_release_old: 361609cbfeafSKirill A. Shutemov put_page(old_page); 36178312034fSJoonsoo Kim 3618ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 3619ad4404a2SDavidlohr Bueso return ret; 36201e8f889bSDavid Gibson } 36211e8f889bSDavid Gibson 362204f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 3623a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 3624a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 362504f2cbe3SMel Gorman { 362604f2cbe3SMel Gorman struct address_space *mapping; 3627e7c4b0bfSAndy Whitcroft pgoff_t idx; 362804f2cbe3SMel Gorman 362904f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 3630a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 363104f2cbe3SMel Gorman 363204f2cbe3SMel Gorman return find_lock_page(mapping, idx); 363304f2cbe3SMel Gorman } 363404f2cbe3SMel Gorman 36353ae77f43SHugh Dickins /* 36363ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 36373ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 36383ae77f43SHugh Dickins */ 36393ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 36402a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 36412a15efc9SHugh Dickins { 36422a15efc9SHugh Dickins struct address_space *mapping; 36432a15efc9SHugh Dickins pgoff_t idx; 36442a15efc9SHugh Dickins struct page *page; 36452a15efc9SHugh Dickins 36462a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 36472a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 36482a15efc9SHugh Dickins 36492a15efc9SHugh Dickins page = find_get_page(mapping, idx); 36502a15efc9SHugh Dickins if (page) 36512a15efc9SHugh Dickins put_page(page); 36522a15efc9SHugh Dickins return page != NULL; 36532a15efc9SHugh Dickins } 36542a15efc9SHugh Dickins 3655ab76ad54SMike Kravetz int huge_add_to_page_cache(struct page *page, struct address_space *mapping, 3656ab76ad54SMike Kravetz pgoff_t idx) 3657ab76ad54SMike Kravetz { 3658ab76ad54SMike Kravetz struct inode *inode = mapping->host; 3659ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 3660ab76ad54SMike Kravetz int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 3661ab76ad54SMike Kravetz 3662ab76ad54SMike Kravetz if (err) 3663ab76ad54SMike Kravetz return err; 3664ab76ad54SMike Kravetz ClearPagePrivate(page); 3665ab76ad54SMike Kravetz 3666ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 3667ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 3668ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 3669ab76ad54SMike Kravetz return 0; 3670ab76ad54SMike Kravetz } 3671ab76ad54SMike Kravetz 3672a1ed3ddaSRobert P. J. Day static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, 36738382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 3674788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 3675ac9b9c66SHugh Dickins { 3676a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3677ac9b9c66SHugh Dickins int ret = VM_FAULT_SIGBUS; 3678409eb8c2SHillf Danton int anon_rmap = 0; 36794c887265SAdam Litke unsigned long size; 36804c887265SAdam Litke struct page *page; 36811e8f889bSDavid Gibson pte_t new_pte; 3682cb900f41SKirill A. Shutemov spinlock_t *ptl; 36834c887265SAdam Litke 368404f2cbe3SMel Gorman /* 368504f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 368604f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 368725985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 368804f2cbe3SMel Gorman */ 368904f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 3690910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 369104f2cbe3SMel Gorman current->pid); 369204f2cbe3SMel Gorman return ret; 369304f2cbe3SMel Gorman } 369404f2cbe3SMel Gorman 36954c887265SAdam Litke /* 36964c887265SAdam Litke * Use page lock to guard against racing truncation 36974c887265SAdam Litke * before we get page_table_lock. 36984c887265SAdam Litke */ 36996bda666aSChristoph Lameter retry: 37006bda666aSChristoph Lameter page = find_lock_page(mapping, idx); 37016bda666aSChristoph Lameter if (!page) { 3702a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 3703ebed4bfcSHugh Dickins if (idx >= size) 3704ebed4bfcSHugh Dickins goto out; 37051a1aad8aSMike Kravetz 37061a1aad8aSMike Kravetz /* 37071a1aad8aSMike Kravetz * Check for page in userfault range 37081a1aad8aSMike Kravetz */ 37091a1aad8aSMike Kravetz if (userfaultfd_missing(vma)) { 37101a1aad8aSMike Kravetz u32 hash; 37111a1aad8aSMike Kravetz struct vm_fault vmf = { 37121a1aad8aSMike Kravetz .vma = vma, 37131a1aad8aSMike Kravetz .address = address, 37141a1aad8aSMike Kravetz .flags = flags, 37151a1aad8aSMike Kravetz /* 37161a1aad8aSMike Kravetz * Hard to debug if it ends up being 37171a1aad8aSMike Kravetz * used by a callee that assumes 37181a1aad8aSMike Kravetz * something about the other 37191a1aad8aSMike Kravetz * uninitialized fields... same as in 37201a1aad8aSMike Kravetz * memory.c 37211a1aad8aSMike Kravetz */ 37221a1aad8aSMike Kravetz }; 37231a1aad8aSMike Kravetz 37241a1aad8aSMike Kravetz /* 37251a1aad8aSMike Kravetz * hugetlb_fault_mutex must be dropped before 37261a1aad8aSMike Kravetz * handling userfault. Reacquire after handling 37271a1aad8aSMike Kravetz * fault to make calling code simpler. 37281a1aad8aSMike Kravetz */ 37291a1aad8aSMike Kravetz hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, 37301a1aad8aSMike Kravetz idx, address); 37311a1aad8aSMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 37321a1aad8aSMike Kravetz ret = handle_userfault(&vmf, VM_UFFD_MISSING); 37331a1aad8aSMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 37341a1aad8aSMike Kravetz goto out; 37351a1aad8aSMike Kravetz } 37361a1aad8aSMike Kravetz 373704f2cbe3SMel Gorman page = alloc_huge_page(vma, address, 0); 37382fc39cecSAdam Litke if (IS_ERR(page)) { 373976dcee75SAneesh Kumar K.V ret = PTR_ERR(page); 374076dcee75SAneesh Kumar K.V if (ret == -ENOMEM) 374176dcee75SAneesh Kumar K.V ret = VM_FAULT_OOM; 374276dcee75SAneesh Kumar K.V else 374376dcee75SAneesh Kumar K.V ret = VM_FAULT_SIGBUS; 37446bda666aSChristoph Lameter goto out; 37456bda666aSChristoph Lameter } 374647ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 37470ed361deSNick Piggin __SetPageUptodate(page); 3748bcc54222SNaoya Horiguchi set_page_huge_active(page); 3749ac9b9c66SHugh Dickins 3750f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 3751ab76ad54SMike Kravetz int err = huge_add_to_page_cache(page, mapping, idx); 37526bda666aSChristoph Lameter if (err) { 37536bda666aSChristoph Lameter put_page(page); 37546bda666aSChristoph Lameter if (err == -EEXIST) 37556bda666aSChristoph Lameter goto retry; 37566bda666aSChristoph Lameter goto out; 37576bda666aSChristoph Lameter } 375823be7468SMel Gorman } else { 37596bda666aSChristoph Lameter lock_page(page); 37600fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 37610fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 37620fe6e20bSNaoya Horiguchi goto backout_unlocked; 376323be7468SMel Gorman } 3764409eb8c2SHillf Danton anon_rmap = 1; 37650fe6e20bSNaoya Horiguchi } 37660fe6e20bSNaoya Horiguchi } else { 376757303d80SAndy Whitcroft /* 3768998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 3769998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 3770998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 3771fd6a03edSNaoya Horiguchi */ 3772fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 3773aa50d3a7SAndi Kleen ret = VM_FAULT_HWPOISON | 3774972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 3775fd6a03edSNaoya Horiguchi goto backout_unlocked; 37766bda666aSChristoph Lameter } 3777998b4382SNaoya Horiguchi } 37781e8f889bSDavid Gibson 377957303d80SAndy Whitcroft /* 378057303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 378157303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 378257303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 378357303d80SAndy Whitcroft * the spinlock. 378457303d80SAndy Whitcroft */ 37855e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 37862b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 37872b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 37882b26736cSAndy Whitcroft goto backout_unlocked; 37892b26736cSAndy Whitcroft } 37905e911373SMike Kravetz /* Just decrements count, does not deallocate */ 3791feba16e2SMike Kravetz vma_end_reservation(h, vma, address); 37925e911373SMike Kravetz } 379357303d80SAndy Whitcroft 37948bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 3795a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 37964c887265SAdam Litke if (idx >= size) 37974c887265SAdam Litke goto backout; 37984c887265SAdam Litke 379983c54070SNick Piggin ret = 0; 38007f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 38014c887265SAdam Litke goto backout; 38024c887265SAdam Litke 380307443a85SJoonsoo Kim if (anon_rmap) { 380407443a85SJoonsoo Kim ClearPagePrivate(page); 3805409eb8c2SHillf Danton hugepage_add_new_anon_rmap(page, vma, address); 3806ac714904SChoi Gi-yong } else 380753f9263bSKirill A. Shutemov page_dup_rmap(page, true); 38081e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 38091e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 38101e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, new_pte); 38111e8f889bSDavid Gibson 38125d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 3813788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 38141e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 38153999f52eSAneesh Kumar K.V ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); 38161e8f889bSDavid Gibson } 38171e8f889bSDavid Gibson 3818cb900f41SKirill A. Shutemov spin_unlock(ptl); 38194c887265SAdam Litke unlock_page(page); 38204c887265SAdam Litke out: 3821ac9b9c66SHugh Dickins return ret; 38224c887265SAdam Litke 38234c887265SAdam Litke backout: 3824cb900f41SKirill A. Shutemov spin_unlock(ptl); 38252b26736cSAndy Whitcroft backout_unlocked: 38264c887265SAdam Litke unlock_page(page); 382796b96a96SMike Kravetz restore_reserve_on_error(h, vma, address, page); 38284c887265SAdam Litke put_page(page); 38294c887265SAdam Litke goto out; 3830ac9b9c66SHugh Dickins } 3831ac9b9c66SHugh Dickins 38328382d914SDavidlohr Bueso #ifdef CONFIG_SMP 3833c672c7f2SMike Kravetz u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, 38348382d914SDavidlohr Bueso struct vm_area_struct *vma, 38358382d914SDavidlohr Bueso struct address_space *mapping, 38368382d914SDavidlohr Bueso pgoff_t idx, unsigned long address) 38378382d914SDavidlohr Bueso { 38388382d914SDavidlohr Bueso unsigned long key[2]; 38398382d914SDavidlohr Bueso u32 hash; 38408382d914SDavidlohr Bueso 38418382d914SDavidlohr Bueso if (vma->vm_flags & VM_SHARED) { 38428382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 38438382d914SDavidlohr Bueso key[1] = idx; 38448382d914SDavidlohr Bueso } else { 38458382d914SDavidlohr Bueso key[0] = (unsigned long) mm; 38468382d914SDavidlohr Bueso key[1] = address >> huge_page_shift(h); 38478382d914SDavidlohr Bueso } 38488382d914SDavidlohr Bueso 38498382d914SDavidlohr Bueso hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0); 38508382d914SDavidlohr Bueso 38518382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 38528382d914SDavidlohr Bueso } 38538382d914SDavidlohr Bueso #else 38548382d914SDavidlohr Bueso /* 38558382d914SDavidlohr Bueso * For uniprocesor systems we always use a single mutex, so just 38568382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 38578382d914SDavidlohr Bueso */ 3858c672c7f2SMike Kravetz u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, 38598382d914SDavidlohr Bueso struct vm_area_struct *vma, 38608382d914SDavidlohr Bueso struct address_space *mapping, 38618382d914SDavidlohr Bueso pgoff_t idx, unsigned long address) 38628382d914SDavidlohr Bueso { 38638382d914SDavidlohr Bueso return 0; 38648382d914SDavidlohr Bueso } 38658382d914SDavidlohr Bueso #endif 38668382d914SDavidlohr Bueso 386786e5216fSAdam Litke int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 3868788c7df4SHugh Dickins unsigned long address, unsigned int flags) 386986e5216fSAdam Litke { 38708382d914SDavidlohr Bueso pte_t *ptep, entry; 3871cb900f41SKirill A. Shutemov spinlock_t *ptl; 38721e8f889bSDavid Gibson int ret; 38738382d914SDavidlohr Bueso u32 hash; 38748382d914SDavidlohr Bueso pgoff_t idx; 38750fe6e20bSNaoya Horiguchi struct page *page = NULL; 387657303d80SAndy Whitcroft struct page *pagecache_page = NULL; 3877a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 38788382d914SDavidlohr Bueso struct address_space *mapping; 38790f792cf9SNaoya Horiguchi int need_wait_lock = 0; 388086e5216fSAdam Litke 38811e16a539SKAMEZAWA Hiroyuki address &= huge_page_mask(h); 38821e16a539SKAMEZAWA Hiroyuki 38837868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, huge_page_size(h)); 3884fd6a03edSNaoya Horiguchi if (ptep) { 3885fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 3886290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 3887cb900f41SKirill A. Shutemov migration_entry_wait_huge(vma, mm, ptep); 3888290408d4SNaoya Horiguchi return 0; 3889290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 3890aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 3891972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 38920d777df5SNaoya Horiguchi } else { 3893a5516438SAndi Kleen ptep = huge_pte_alloc(mm, address, huge_page_size(h)); 389486e5216fSAdam Litke if (!ptep) 389586e5216fSAdam Litke return VM_FAULT_OOM; 38960d777df5SNaoya Horiguchi } 389786e5216fSAdam Litke 38988382d914SDavidlohr Bueso mapping = vma->vm_file->f_mapping; 38998382d914SDavidlohr Bueso idx = vma_hugecache_offset(h, vma, address); 39008382d914SDavidlohr Bueso 39013935baa9SDavid Gibson /* 39023935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 39033935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 39043935baa9SDavid Gibson * the same page in the page cache. 39053935baa9SDavid Gibson */ 3906c672c7f2SMike Kravetz hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address); 3907c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 39088382d914SDavidlohr Bueso 39097f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 39107f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 39118382d914SDavidlohr Bueso ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 3912b4d1d99fSDavid Gibson goto out_mutex; 39133935baa9SDavid Gibson } 391486e5216fSAdam Litke 391583c54070SNick Piggin ret = 0; 39161e8f889bSDavid Gibson 391757303d80SAndy Whitcroft /* 39180f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 39190f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 39200f792cf9SNaoya Horiguchi * a active hugepage in pagecache. This goto expects the 2nd page fault, 39210f792cf9SNaoya Horiguchi * and is_hugetlb_entry_(migration|hwpoisoned) check will properly 39220f792cf9SNaoya Horiguchi * handle it. 39230f792cf9SNaoya Horiguchi */ 39240f792cf9SNaoya Horiguchi if (!pte_present(entry)) 39250f792cf9SNaoya Horiguchi goto out_mutex; 39260f792cf9SNaoya Horiguchi 39270f792cf9SNaoya Horiguchi /* 392857303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 392957303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 393057303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 393157303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 393257303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 393357303d80SAndy Whitcroft * consumed. 393457303d80SAndy Whitcroft */ 3935106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 39362b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 39372b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 3938b4d1d99fSDavid Gibson goto out_mutex; 39392b26736cSAndy Whitcroft } 39405e911373SMike Kravetz /* Just decrements count, does not deallocate */ 3941feba16e2SMike Kravetz vma_end_reservation(h, vma, address); 394257303d80SAndy Whitcroft 3943f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 394457303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 394557303d80SAndy Whitcroft vma, address); 394657303d80SAndy Whitcroft } 394757303d80SAndy Whitcroft 39480f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 39490fe6e20bSNaoya Horiguchi 39501e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 3951b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 3952cb900f41SKirill A. Shutemov goto out_ptl; 3953b4d1d99fSDavid Gibson 39540f792cf9SNaoya Horiguchi /* 39550f792cf9SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 39560f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 39570f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 39580f792cf9SNaoya Horiguchi */ 39590f792cf9SNaoya Horiguchi page = pte_page(entry); 39600f792cf9SNaoya Horiguchi if (page != pagecache_page) 39610f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 39620f792cf9SNaoya Horiguchi need_wait_lock = 1; 39630f792cf9SNaoya Horiguchi goto out_ptl; 39640f792cf9SNaoya Horiguchi } 39650f792cf9SNaoya Horiguchi 39660f792cf9SNaoya Horiguchi get_page(page); 3967b4d1d99fSDavid Gibson 3968788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 3969106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 39703999f52eSAneesh Kumar K.V ret = hugetlb_cow(mm, vma, address, ptep, 3971cb900f41SKirill A. Shutemov pagecache_page, ptl); 39720f792cf9SNaoya Horiguchi goto out_put_page; 3973b4d1d99fSDavid Gibson } 3974106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 3975b4d1d99fSDavid Gibson } 3976b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 3977788c7df4SHugh Dickins if (huge_ptep_set_access_flags(vma, address, ptep, entry, 3978788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 39794b3073e1SRussell King update_mmu_cache(vma, address, ptep); 39800f792cf9SNaoya Horiguchi out_put_page: 39810f792cf9SNaoya Horiguchi if (page != pagecache_page) 39820f792cf9SNaoya Horiguchi unlock_page(page); 39830f792cf9SNaoya Horiguchi put_page(page); 3984cb900f41SKirill A. Shutemov out_ptl: 3985cb900f41SKirill A. Shutemov spin_unlock(ptl); 398657303d80SAndy Whitcroft 398757303d80SAndy Whitcroft if (pagecache_page) { 398857303d80SAndy Whitcroft unlock_page(pagecache_page); 398957303d80SAndy Whitcroft put_page(pagecache_page); 399057303d80SAndy Whitcroft } 3991b4d1d99fSDavid Gibson out_mutex: 3992c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 39930f792cf9SNaoya Horiguchi /* 39940f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 39950f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 39960f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 39970f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 39980f792cf9SNaoya Horiguchi * here without taking refcount. 39990f792cf9SNaoya Horiguchi */ 40000f792cf9SNaoya Horiguchi if (need_wait_lock) 40010f792cf9SNaoya Horiguchi wait_on_page_locked(page); 40021e8f889bSDavid Gibson return ret; 400386e5216fSAdam Litke } 400486e5216fSAdam Litke 40058fb5debcSMike Kravetz /* 40068fb5debcSMike Kravetz * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with 40078fb5debcSMike Kravetz * modifications for huge pages. 40088fb5debcSMike Kravetz */ 40098fb5debcSMike Kravetz int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, 40108fb5debcSMike Kravetz pte_t *dst_pte, 40118fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 40128fb5debcSMike Kravetz unsigned long dst_addr, 40138fb5debcSMike Kravetz unsigned long src_addr, 40148fb5debcSMike Kravetz struct page **pagep) 40158fb5debcSMike Kravetz { 40161e392147SAndrea Arcangeli struct address_space *mapping; 40171e392147SAndrea Arcangeli pgoff_t idx; 40181e392147SAndrea Arcangeli unsigned long size; 40191c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 40208fb5debcSMike Kravetz struct hstate *h = hstate_vma(dst_vma); 40218fb5debcSMike Kravetz pte_t _dst_pte; 40228fb5debcSMike Kravetz spinlock_t *ptl; 40238fb5debcSMike Kravetz int ret; 40248fb5debcSMike Kravetz struct page *page; 40258fb5debcSMike Kravetz 40268fb5debcSMike Kravetz if (!*pagep) { 40278fb5debcSMike Kravetz ret = -ENOMEM; 40288fb5debcSMike Kravetz page = alloc_huge_page(dst_vma, dst_addr, 0); 40298fb5debcSMike Kravetz if (IS_ERR(page)) 40308fb5debcSMike Kravetz goto out; 40318fb5debcSMike Kravetz 40328fb5debcSMike Kravetz ret = copy_huge_page_from_user(page, 40338fb5debcSMike Kravetz (const void __user *) src_addr, 4034810a56b9SMike Kravetz pages_per_huge_page(h), false); 40358fb5debcSMike Kravetz 40368fb5debcSMike Kravetz /* fallback to copy_from_user outside mmap_sem */ 40378fb5debcSMike Kravetz if (unlikely(ret)) { 40388fb5debcSMike Kravetz ret = -EFAULT; 40398fb5debcSMike Kravetz *pagep = page; 40408fb5debcSMike Kravetz /* don't free the page */ 40418fb5debcSMike Kravetz goto out; 40428fb5debcSMike Kravetz } 40438fb5debcSMike Kravetz } else { 40448fb5debcSMike Kravetz page = *pagep; 40458fb5debcSMike Kravetz *pagep = NULL; 40468fb5debcSMike Kravetz } 40478fb5debcSMike Kravetz 40488fb5debcSMike Kravetz /* 40498fb5debcSMike Kravetz * The memory barrier inside __SetPageUptodate makes sure that 40508fb5debcSMike Kravetz * preceding stores to the page contents become visible before 40518fb5debcSMike Kravetz * the set_pte_at() write. 40528fb5debcSMike Kravetz */ 40538fb5debcSMike Kravetz __SetPageUptodate(page); 40548fb5debcSMike Kravetz set_page_huge_active(page); 40558fb5debcSMike Kravetz 40561e392147SAndrea Arcangeli mapping = dst_vma->vm_file->f_mapping; 40571e392147SAndrea Arcangeli idx = vma_hugecache_offset(h, dst_vma, dst_addr); 40581e392147SAndrea Arcangeli 40591c9e8defSMike Kravetz /* 40601c9e8defSMike Kravetz * If shared, add to page cache 40611c9e8defSMike Kravetz */ 40621c9e8defSMike Kravetz if (vm_shared) { 40631e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 40641e392147SAndrea Arcangeli ret = -EFAULT; 40651e392147SAndrea Arcangeli if (idx >= size) 40661e392147SAndrea Arcangeli goto out_release_nounlock; 40671c9e8defSMike Kravetz 40681e392147SAndrea Arcangeli /* 40691e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 40701e392147SAndrea Arcangeli * huge_add_to_page_cache() below happens through the 40711e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 40721e392147SAndrea Arcangeli * the caller. 40731e392147SAndrea Arcangeli */ 40741c9e8defSMike Kravetz ret = huge_add_to_page_cache(page, mapping, idx); 40751c9e8defSMike Kravetz if (ret) 40761c9e8defSMike Kravetz goto out_release_nounlock; 40771c9e8defSMike Kravetz } 40781c9e8defSMike Kravetz 40798fb5debcSMike Kravetz ptl = huge_pte_lockptr(h, dst_mm, dst_pte); 40808fb5debcSMike Kravetz spin_lock(ptl); 40818fb5debcSMike Kravetz 40821e392147SAndrea Arcangeli /* 40831e392147SAndrea Arcangeli * Recheck the i_size after holding PT lock to make sure not 40841e392147SAndrea Arcangeli * to leave any page mapped (as page_mapped()) beyond the end 40851e392147SAndrea Arcangeli * of the i_size (remove_inode_hugepages() is strict about 40861e392147SAndrea Arcangeli * enforcing that). If we bail out here, we'll also leave a 40871e392147SAndrea Arcangeli * page in the radix tree in the vm_shared case beyond the end 40881e392147SAndrea Arcangeli * of the i_size, but remove_inode_hugepages() will take care 40891e392147SAndrea Arcangeli * of it as soon as we drop the hugetlb_fault_mutex_table. 40901e392147SAndrea Arcangeli */ 40911e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 40921e392147SAndrea Arcangeli ret = -EFAULT; 40931e392147SAndrea Arcangeli if (idx >= size) 40941e392147SAndrea Arcangeli goto out_release_unlock; 40951e392147SAndrea Arcangeli 40968fb5debcSMike Kravetz ret = -EEXIST; 40978fb5debcSMike Kravetz if (!huge_pte_none(huge_ptep_get(dst_pte))) 40988fb5debcSMike Kravetz goto out_release_unlock; 40998fb5debcSMike Kravetz 41001c9e8defSMike Kravetz if (vm_shared) { 41011c9e8defSMike Kravetz page_dup_rmap(page, true); 41021c9e8defSMike Kravetz } else { 41038fb5debcSMike Kravetz ClearPagePrivate(page); 41048fb5debcSMike Kravetz hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); 41051c9e8defSMike Kravetz } 41068fb5debcSMike Kravetz 41078fb5debcSMike Kravetz _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); 41088fb5debcSMike Kravetz if (dst_vma->vm_flags & VM_WRITE) 41098fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 41108fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 41118fb5debcSMike Kravetz 41128fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 41138fb5debcSMike Kravetz 41148fb5debcSMike Kravetz (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, 41158fb5debcSMike Kravetz dst_vma->vm_flags & VM_WRITE); 41168fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 41178fb5debcSMike Kravetz 41188fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 41198fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 41208fb5debcSMike Kravetz 41218fb5debcSMike Kravetz spin_unlock(ptl); 41221c9e8defSMike Kravetz if (vm_shared) 41231c9e8defSMike Kravetz unlock_page(page); 41248fb5debcSMike Kravetz ret = 0; 41258fb5debcSMike Kravetz out: 41268fb5debcSMike Kravetz return ret; 41278fb5debcSMike Kravetz out_release_unlock: 41288fb5debcSMike Kravetz spin_unlock(ptl); 41291c9e8defSMike Kravetz if (vm_shared) 41301c9e8defSMike Kravetz unlock_page(page); 41315af10dfdSAndrea Arcangeli out_release_nounlock: 41328fb5debcSMike Kravetz put_page(page); 41338fb5debcSMike Kravetz goto out; 41348fb5debcSMike Kravetz } 41358fb5debcSMike Kravetz 413628a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 413763551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 413828a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 413987ffc118SAndrea Arcangeli long i, unsigned int flags, int *nonblocking) 414063551ae0SDavid Gibson { 4141d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 4142d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 414328a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 4144a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 41452be7cfedSDaniel Jordan int err = -EFAULT; 414663551ae0SDavid Gibson 414763551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 414863551ae0SDavid Gibson pte_t *pte; 4149cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 41502a15efc9SHugh Dickins int absent; 415163551ae0SDavid Gibson struct page *page; 415263551ae0SDavid Gibson 41534c887265SAdam Litke /* 415402057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 415502057967SDavid Rientjes * potentially allocating memory. 415602057967SDavid Rientjes */ 415702057967SDavid Rientjes if (unlikely(fatal_signal_pending(current))) { 415802057967SDavid Rientjes remainder = 0; 415902057967SDavid Rientjes break; 416002057967SDavid Rientjes } 416102057967SDavid Rientjes 416202057967SDavid Rientjes /* 41634c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 41642a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 41654c887265SAdam Litke * first, for the page indexing below to work. 4166cb900f41SKirill A. Shutemov * 4167cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 41684c887265SAdam Litke */ 41697868a208SPunit Agrawal pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), 41707868a208SPunit Agrawal huge_page_size(h)); 4171cb900f41SKirill A. Shutemov if (pte) 4172cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 41732a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 417463551ae0SDavid Gibson 41752a15efc9SHugh Dickins /* 41762a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 41773ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 41783ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 41793ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 41803ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 41812a15efc9SHugh Dickins */ 41823ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 41833ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 4184cb900f41SKirill A. Shutemov if (pte) 4185cb900f41SKirill A. Shutemov spin_unlock(ptl); 41862a15efc9SHugh Dickins remainder = 0; 41872a15efc9SHugh Dickins break; 41882a15efc9SHugh Dickins } 41892a15efc9SHugh Dickins 41909cc3a5bdSNaoya Horiguchi /* 41919cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 41929cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 41939cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 41949cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 41959cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 41969cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 41979cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 41989cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 41999cc3a5bdSNaoya Horiguchi */ 42009cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 4201106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 4202106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 42034c887265SAdam Litke int ret; 420487ffc118SAndrea Arcangeli unsigned int fault_flags = 0; 42054c887265SAdam Litke 4206cb900f41SKirill A. Shutemov if (pte) 4207cb900f41SKirill A. Shutemov spin_unlock(ptl); 420887ffc118SAndrea Arcangeli if (flags & FOLL_WRITE) 420987ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_WRITE; 421087ffc118SAndrea Arcangeli if (nonblocking) 421187ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY; 421287ffc118SAndrea Arcangeli if (flags & FOLL_NOWAIT) 421387ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY | 421487ffc118SAndrea Arcangeli FAULT_FLAG_RETRY_NOWAIT; 421587ffc118SAndrea Arcangeli if (flags & FOLL_TRIED) { 421687ffc118SAndrea Arcangeli VM_WARN_ON_ONCE(fault_flags & 421787ffc118SAndrea Arcangeli FAULT_FLAG_ALLOW_RETRY); 421887ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_TRIED; 421987ffc118SAndrea Arcangeli } 422087ffc118SAndrea Arcangeli ret = hugetlb_fault(mm, vma, vaddr, fault_flags); 422187ffc118SAndrea Arcangeli if (ret & VM_FAULT_ERROR) { 42222be7cfedSDaniel Jordan err = vm_fault_to_errno(ret, flags); 42231c59827dSHugh Dickins remainder = 0; 42241c59827dSHugh Dickins break; 42251c59827dSHugh Dickins } 422687ffc118SAndrea Arcangeli if (ret & VM_FAULT_RETRY) { 422787ffc118SAndrea Arcangeli if (nonblocking) 422887ffc118SAndrea Arcangeli *nonblocking = 0; 422987ffc118SAndrea Arcangeli *nr_pages = 0; 423087ffc118SAndrea Arcangeli /* 423187ffc118SAndrea Arcangeli * VM_FAULT_RETRY must not return an 423287ffc118SAndrea Arcangeli * error, it will return zero 423387ffc118SAndrea Arcangeli * instead. 423487ffc118SAndrea Arcangeli * 423587ffc118SAndrea Arcangeli * No need to update "position" as the 423687ffc118SAndrea Arcangeli * caller will not check it after 423787ffc118SAndrea Arcangeli * *nr_pages is set to 0. 423887ffc118SAndrea Arcangeli */ 423987ffc118SAndrea Arcangeli return i; 424087ffc118SAndrea Arcangeli } 424187ffc118SAndrea Arcangeli continue; 424287ffc118SAndrea Arcangeli } 424363551ae0SDavid Gibson 4244a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 42457f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 4246d5d4b0aaSChen, Kenneth W same_page: 4247d6692183SChen, Kenneth W if (pages) { 424869d177c2SAndy Whitcroft pages[i] = mem_map_offset(page, pfn_offset); 4249ddc58f27SKirill A. Shutemov get_page(pages[i]); 4250d6692183SChen, Kenneth W } 425163551ae0SDavid Gibson 425263551ae0SDavid Gibson if (vmas) 425363551ae0SDavid Gibson vmas[i] = vma; 425463551ae0SDavid Gibson 425563551ae0SDavid Gibson vaddr += PAGE_SIZE; 4256d5d4b0aaSChen, Kenneth W ++pfn_offset; 425763551ae0SDavid Gibson --remainder; 425863551ae0SDavid Gibson ++i; 4259d5d4b0aaSChen, Kenneth W if (vaddr < vma->vm_end && remainder && 4260a5516438SAndi Kleen pfn_offset < pages_per_huge_page(h)) { 4261d5d4b0aaSChen, Kenneth W /* 4262d5d4b0aaSChen, Kenneth W * We use pfn_offset to avoid touching the pageframes 4263d5d4b0aaSChen, Kenneth W * of this compound page. 4264d5d4b0aaSChen, Kenneth W */ 4265d5d4b0aaSChen, Kenneth W goto same_page; 4266d5d4b0aaSChen, Kenneth W } 4267cb900f41SKirill A. Shutemov spin_unlock(ptl); 426863551ae0SDavid Gibson } 426928a35716SMichel Lespinasse *nr_pages = remainder; 427087ffc118SAndrea Arcangeli /* 427187ffc118SAndrea Arcangeli * setting position is actually required only if remainder is 427287ffc118SAndrea Arcangeli * not zero but it's faster not to add a "if (remainder)" 427387ffc118SAndrea Arcangeli * branch. 427487ffc118SAndrea Arcangeli */ 427563551ae0SDavid Gibson *position = vaddr; 427663551ae0SDavid Gibson 42772be7cfedSDaniel Jordan return i ? i : err; 427863551ae0SDavid Gibson } 42798f860591SZhang, Yanmin 42805491ae7bSAneesh Kumar K.V #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE 42815491ae7bSAneesh Kumar K.V /* 42825491ae7bSAneesh Kumar K.V * ARCHes with special requirements for evicting HUGETLB backing TLB entries can 42835491ae7bSAneesh Kumar K.V * implement this. 42845491ae7bSAneesh Kumar K.V */ 42855491ae7bSAneesh Kumar K.V #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) 42865491ae7bSAneesh Kumar K.V #endif 42875491ae7bSAneesh Kumar K.V 42887da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 42898f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 42908f860591SZhang, Yanmin { 42918f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 42928f860591SZhang, Yanmin unsigned long start = address; 42938f860591SZhang, Yanmin pte_t *ptep; 42948f860591SZhang, Yanmin pte_t pte; 4295a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 42967da4d641SPeter Zijlstra unsigned long pages = 0; 42978f860591SZhang, Yanmin 42988f860591SZhang, Yanmin BUG_ON(address >= end); 42998f860591SZhang, Yanmin flush_cache_range(vma, address, end); 43008f860591SZhang, Yanmin 4301a5338093SRik van Riel mmu_notifier_invalidate_range_start(mm, start, end); 430283cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 4303a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 4304cb900f41SKirill A. Shutemov spinlock_t *ptl; 43057868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, huge_page_size(h)); 43068f860591SZhang, Yanmin if (!ptep) 43078f860591SZhang, Yanmin continue; 4308cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 43097da4d641SPeter Zijlstra if (huge_pmd_unshare(mm, &address, ptep)) { 43107da4d641SPeter Zijlstra pages++; 4311cb900f41SKirill A. Shutemov spin_unlock(ptl); 431239dde65cSChen, Kenneth W continue; 43137da4d641SPeter Zijlstra } 4314a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 4315a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 4316a8bda28dSNaoya Horiguchi spin_unlock(ptl); 4317a8bda28dSNaoya Horiguchi continue; 4318a8bda28dSNaoya Horiguchi } 4319a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 4320a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 4321a8bda28dSNaoya Horiguchi 4322a8bda28dSNaoya Horiguchi if (is_write_migration_entry(entry)) { 4323a8bda28dSNaoya Horiguchi pte_t newpte; 4324a8bda28dSNaoya Horiguchi 4325a8bda28dSNaoya Horiguchi make_migration_entry_read(&entry); 4326a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 4327e5251fd4SPunit Agrawal set_huge_swap_pte_at(mm, address, ptep, 4328e5251fd4SPunit Agrawal newpte, huge_page_size(h)); 4329a8bda28dSNaoya Horiguchi pages++; 4330a8bda28dSNaoya Horiguchi } 4331a8bda28dSNaoya Horiguchi spin_unlock(ptl); 4332a8bda28dSNaoya Horiguchi continue; 4333a8bda28dSNaoya Horiguchi } 4334a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 43358f860591SZhang, Yanmin pte = huge_ptep_get_and_clear(mm, address, ptep); 4336106c992aSGerald Schaefer pte = pte_mkhuge(huge_pte_modify(pte, newprot)); 4337be7517d6STony Lu pte = arch_make_huge_pte(pte, vma, NULL, 0); 43388f860591SZhang, Yanmin set_huge_pte_at(mm, address, ptep, pte); 43397da4d641SPeter Zijlstra pages++; 43408f860591SZhang, Yanmin } 4341cb900f41SKirill A. Shutemov spin_unlock(ptl); 43428f860591SZhang, Yanmin } 4343d833352aSMel Gorman /* 4344c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 4345d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 4346c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 4347d833352aSMel Gorman * and that page table be reused and filled with junk. 4348d833352aSMel Gorman */ 43495491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 43500f10851eSJérôme Glisse /* 43510f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 43520f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 43530f10851eSJérôme Glisse * 43540f10851eSJérôme Glisse * See Documentation/vm/mmu_notifier.txt 43550f10851eSJérôme Glisse */ 435683cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 4357a5338093SRik van Riel mmu_notifier_invalidate_range_end(mm, start, end); 43587da4d641SPeter Zijlstra 43597da4d641SPeter Zijlstra return pages << h->order; 43608f860591SZhang, Yanmin } 43618f860591SZhang, Yanmin 4362a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 4363a1e78772SMel Gorman long from, long to, 43645a6fe125SMel Gorman struct vm_area_struct *vma, 4365ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 4366e4e574b7SAdam Litke { 436717c9d12eSMel Gorman long ret, chg; 4368a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 436990481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 43709119a41eSJoonsoo Kim struct resv_map *resv_map; 43711c5ecae3SMike Kravetz long gbl_reserve; 4372e4e574b7SAdam Litke 4373a1e78772SMel Gorman /* 437417c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 437517c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 437690481622SDavid Gibson * without using reserves 437717c9d12eSMel Gorman */ 4378ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 437917c9d12eSMel Gorman return 0; 438017c9d12eSMel Gorman 438117c9d12eSMel Gorman /* 4382a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 4383a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 4384a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 4385a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 4386a1e78772SMel Gorman */ 43879119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 43884e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 43899119a41eSJoonsoo Kim 43901406ec9bSJoonsoo Kim chg = region_chg(resv_map, from, to); 43919119a41eSJoonsoo Kim 43929119a41eSJoonsoo Kim } else { 43939119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 43945a6fe125SMel Gorman if (!resv_map) 43955a6fe125SMel Gorman return -ENOMEM; 43965a6fe125SMel Gorman 439717c9d12eSMel Gorman chg = to - from; 439817c9d12eSMel Gorman 43995a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 44005a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 44015a6fe125SMel Gorman } 44025a6fe125SMel Gorman 4403c50ac050SDave Hansen if (chg < 0) { 4404c50ac050SDave Hansen ret = chg; 4405c50ac050SDave Hansen goto out_err; 4406c50ac050SDave Hansen } 440717c9d12eSMel Gorman 44081c5ecae3SMike Kravetz /* 44091c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 44101c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 44111c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 44121c5ecae3SMike Kravetz */ 44131c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 44141c5ecae3SMike Kravetz if (gbl_reserve < 0) { 4415c50ac050SDave Hansen ret = -ENOSPC; 4416c50ac050SDave Hansen goto out_err; 4417c50ac050SDave Hansen } 441817c9d12eSMel Gorman 441917c9d12eSMel Gorman /* 442017c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 442190481622SDavid Gibson * Hand the pages back to the subpool if there are not 442217c9d12eSMel Gorman */ 44231c5ecae3SMike Kravetz ret = hugetlb_acct_memory(h, gbl_reserve); 442417c9d12eSMel Gorman if (ret < 0) { 44251c5ecae3SMike Kravetz /* put back original number of pages, chg */ 44261c5ecae3SMike Kravetz (void)hugepage_subpool_put_pages(spool, chg); 4427c50ac050SDave Hansen goto out_err; 442817c9d12eSMel Gorman } 442917c9d12eSMel Gorman 443017c9d12eSMel Gorman /* 443117c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 443217c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 443317c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 443417c9d12eSMel Gorman * the reservation was and the page cache tells how much of 443517c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 443617c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 443717c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 443817c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 443917c9d12eSMel Gorman * else has to be done for private mappings here 444017c9d12eSMel Gorman */ 444133039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 444233039678SMike Kravetz long add = region_add(resv_map, from, to); 444333039678SMike Kravetz 444433039678SMike Kravetz if (unlikely(chg > add)) { 444533039678SMike Kravetz /* 444633039678SMike Kravetz * pages in this range were added to the reserve 444733039678SMike Kravetz * map between region_chg and region_add. This 444833039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 444933039678SMike Kravetz * the subpool and reserve counts modified above 445033039678SMike Kravetz * based on the difference. 445133039678SMike Kravetz */ 445233039678SMike Kravetz long rsv_adjust; 445333039678SMike Kravetz 445433039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 445533039678SMike Kravetz chg - add); 445633039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 445733039678SMike Kravetz } 445833039678SMike Kravetz } 4459a43a8c39SChen, Kenneth W return 0; 4460c50ac050SDave Hansen out_err: 44615e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 4462ff8c0c53SMike Kravetz /* Don't call region_abort if region_chg failed */ 4463ff8c0c53SMike Kravetz if (chg >= 0) 44645e911373SMike Kravetz region_abort(resv_map, from, to); 4465f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 4466f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 4467c50ac050SDave Hansen return ret; 4468a43a8c39SChen, Kenneth W } 4469a43a8c39SChen, Kenneth W 4470b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 4471b5cec28dSMike Kravetz long freed) 4472a43a8c39SChen, Kenneth W { 4473a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 44744e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 44759119a41eSJoonsoo Kim long chg = 0; 447690481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 44771c5ecae3SMike Kravetz long gbl_reserve; 447845c682a6SKen Chen 4479b5cec28dSMike Kravetz if (resv_map) { 4480b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 4481b5cec28dSMike Kravetz /* 4482b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 4483b5cec28dSMike Kravetz * must be split and another region descriptor can not be 4484b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 4485b5cec28dSMike Kravetz */ 4486b5cec28dSMike Kravetz if (chg < 0) 4487b5cec28dSMike Kravetz return chg; 4488b5cec28dSMike Kravetz } 4489b5cec28dSMike Kravetz 449045c682a6SKen Chen spin_lock(&inode->i_lock); 4491e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 449245c682a6SKen Chen spin_unlock(&inode->i_lock); 449345c682a6SKen Chen 44941c5ecae3SMike Kravetz /* 44951c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 44961c5ecae3SMike Kravetz * reservations to be released may be adjusted. 44971c5ecae3SMike Kravetz */ 44981c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 44991c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 4500b5cec28dSMike Kravetz 4501b5cec28dSMike Kravetz return 0; 4502a43a8c39SChen, Kenneth W } 450393f70f90SNaoya Horiguchi 45043212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 45053212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 45063212b535SSteve Capper struct vm_area_struct *vma, 45073212b535SSteve Capper unsigned long addr, pgoff_t idx) 45083212b535SSteve Capper { 45093212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 45103212b535SSteve Capper svma->vm_start; 45113212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 45123212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 45133212b535SSteve Capper 45143212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 4515de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 4516de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 45173212b535SSteve Capper 45183212b535SSteve Capper /* 45193212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 45203212b535SSteve Capper * page table page. 45213212b535SSteve Capper */ 45223212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 45233212b535SSteve Capper vm_flags != svm_flags || 45243212b535SSteve Capper sbase < svma->vm_start || svma->vm_end < s_end) 45253212b535SSteve Capper return 0; 45263212b535SSteve Capper 45273212b535SSteve Capper return saddr; 45283212b535SSteve Capper } 45293212b535SSteve Capper 453031aafb45SNicholas Krause static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) 45313212b535SSteve Capper { 45323212b535SSteve Capper unsigned long base = addr & PUD_MASK; 45333212b535SSteve Capper unsigned long end = base + PUD_SIZE; 45343212b535SSteve Capper 45353212b535SSteve Capper /* 45363212b535SSteve Capper * check on proper vm_flags and page table alignment 45373212b535SSteve Capper */ 45383212b535SSteve Capper if (vma->vm_flags & VM_MAYSHARE && 45393212b535SSteve Capper vma->vm_start <= base && end <= vma->vm_end) 454031aafb45SNicholas Krause return true; 454131aafb45SNicholas Krause return false; 45423212b535SSteve Capper } 45433212b535SSteve Capper 45443212b535SSteve Capper /* 45453212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 45463212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 45473212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 45483212b535SSteve Capper * code much cleaner. pmd allocation is essential for the shared case because 4549c8c06efaSDavidlohr Bueso * pud has to be populated inside the same i_mmap_rwsem section - otherwise 45503212b535SSteve Capper * racing tasks could either miss the sharing (see huge_pte_offset) or select a 45513212b535SSteve Capper * bad pmd for sharing. 45523212b535SSteve Capper */ 45533212b535SSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 45543212b535SSteve Capper { 45553212b535SSteve Capper struct vm_area_struct *vma = find_vma(mm, addr); 45563212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 45573212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 45583212b535SSteve Capper vma->vm_pgoff; 45593212b535SSteve Capper struct vm_area_struct *svma; 45603212b535SSteve Capper unsigned long saddr; 45613212b535SSteve Capper pte_t *spte = NULL; 45623212b535SSteve Capper pte_t *pte; 4563cb900f41SKirill A. Shutemov spinlock_t *ptl; 45643212b535SSteve Capper 45653212b535SSteve Capper if (!vma_shareable(vma, addr)) 45663212b535SSteve Capper return (pte_t *)pmd_alloc(mm, pud, addr); 45673212b535SSteve Capper 456883cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 45693212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 45703212b535SSteve Capper if (svma == vma) 45713212b535SSteve Capper continue; 45723212b535SSteve Capper 45733212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 45743212b535SSteve Capper if (saddr) { 45757868a208SPunit Agrawal spte = huge_pte_offset(svma->vm_mm, saddr, 45767868a208SPunit Agrawal vma_mmu_pagesize(svma)); 45773212b535SSteve Capper if (spte) { 45783212b535SSteve Capper get_page(virt_to_page(spte)); 45793212b535SSteve Capper break; 45803212b535SSteve Capper } 45813212b535SSteve Capper } 45823212b535SSteve Capper } 45833212b535SSteve Capper 45843212b535SSteve Capper if (!spte) 45853212b535SSteve Capper goto out; 45863212b535SSteve Capper 45878bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 4588dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 45893212b535SSteve Capper pud_populate(mm, pud, 45903212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 4591c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 4592dc6c9a35SKirill A. Shutemov } else { 45933212b535SSteve Capper put_page(virt_to_page(spte)); 4594dc6c9a35SKirill A. Shutemov } 4595cb900f41SKirill A. Shutemov spin_unlock(ptl); 45963212b535SSteve Capper out: 45973212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 459883cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 45993212b535SSteve Capper return pte; 46003212b535SSteve Capper } 46013212b535SSteve Capper 46023212b535SSteve Capper /* 46033212b535SSteve Capper * unmap huge page backed by shared pte. 46043212b535SSteve Capper * 46053212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 46063212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 46073212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 46083212b535SSteve Capper * 4609cb900f41SKirill A. Shutemov * called with page table lock held. 46103212b535SSteve Capper * 46113212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 46123212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 46133212b535SSteve Capper */ 46143212b535SSteve Capper int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 46153212b535SSteve Capper { 46163212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 4617c2febafcSKirill A. Shutemov p4d_t *p4d = p4d_offset(pgd, *addr); 4618c2febafcSKirill A. Shutemov pud_t *pud = pud_offset(p4d, *addr); 46193212b535SSteve Capper 46203212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 46213212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 46223212b535SSteve Capper return 0; 46233212b535SSteve Capper 46243212b535SSteve Capper pud_clear(pud); 46253212b535SSteve Capper put_page(virt_to_page(ptep)); 4626dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 46273212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 46283212b535SSteve Capper return 1; 46293212b535SSteve Capper } 46309e5fc74cSSteve Capper #define want_pmd_share() (1) 46319e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 46329e5fc74cSSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 46339e5fc74cSSteve Capper { 46349e5fc74cSSteve Capper return NULL; 46359e5fc74cSSteve Capper } 4636e81f2d22SZhang Zhen 4637e81f2d22SZhang Zhen int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 4638e81f2d22SZhang Zhen { 4639e81f2d22SZhang Zhen return 0; 4640e81f2d22SZhang Zhen } 46419e5fc74cSSteve Capper #define want_pmd_share() (0) 46423212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 46433212b535SSteve Capper 46449e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 46459e5fc74cSSteve Capper pte_t *huge_pte_alloc(struct mm_struct *mm, 46469e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 46479e5fc74cSSteve Capper { 46489e5fc74cSSteve Capper pgd_t *pgd; 4649c2febafcSKirill A. Shutemov p4d_t *p4d; 46509e5fc74cSSteve Capper pud_t *pud; 46519e5fc74cSSteve Capper pte_t *pte = NULL; 46529e5fc74cSSteve Capper 46539e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 4654f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 4655f4f0a3d8SKirill A. Shutemov if (!p4d) 4656f4f0a3d8SKirill A. Shutemov return NULL; 4657c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 46589e5fc74cSSteve Capper if (pud) { 46599e5fc74cSSteve Capper if (sz == PUD_SIZE) { 46609e5fc74cSSteve Capper pte = (pte_t *)pud; 46619e5fc74cSSteve Capper } else { 46629e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 46639e5fc74cSSteve Capper if (want_pmd_share() && pud_none(*pud)) 46649e5fc74cSSteve Capper pte = huge_pmd_share(mm, addr, pud); 46659e5fc74cSSteve Capper else 46669e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 46679e5fc74cSSteve Capper } 46689e5fc74cSSteve Capper } 46694e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 46709e5fc74cSSteve Capper 46719e5fc74cSSteve Capper return pte; 46729e5fc74cSSteve Capper } 46739e5fc74cSSteve Capper 46749b19df29SPunit Agrawal /* 46759b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 46769b19df29SPunit Agrawal * entry at address @addr 46779b19df29SPunit Agrawal * 46789b19df29SPunit Agrawal * Return: Pointer to page table or swap entry (PUD or PMD) for 46799b19df29SPunit Agrawal * address @addr, or NULL if a p*d_none() entry is encountered and the 46809b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 46819b19df29SPunit Agrawal * table. 46829b19df29SPunit Agrawal */ 46837868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 46847868a208SPunit Agrawal unsigned long addr, unsigned long sz) 46859e5fc74cSSteve Capper { 46869e5fc74cSSteve Capper pgd_t *pgd; 4687c2febafcSKirill A. Shutemov p4d_t *p4d; 46889e5fc74cSSteve Capper pud_t *pud; 4689c2febafcSKirill A. Shutemov pmd_t *pmd; 46909e5fc74cSSteve Capper 46919e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 4692c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 4693c2febafcSKirill A. Shutemov return NULL; 4694c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 4695c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 4696c2febafcSKirill A. Shutemov return NULL; 46979b19df29SPunit Agrawal 4698c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 46999b19df29SPunit Agrawal if (sz != PUD_SIZE && pud_none(*pud)) 4700c2febafcSKirill A. Shutemov return NULL; 47019b19df29SPunit Agrawal /* hugepage or swap? */ 47029b19df29SPunit Agrawal if (pud_huge(*pud) || !pud_present(*pud)) 47039e5fc74cSSteve Capper return (pte_t *)pud; 47049b19df29SPunit Agrawal 47059e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 47069b19df29SPunit Agrawal if (sz != PMD_SIZE && pmd_none(*pmd)) 47079b19df29SPunit Agrawal return NULL; 47089b19df29SPunit Agrawal /* hugepage or swap? */ 47099b19df29SPunit Agrawal if (pmd_huge(*pmd) || !pmd_present(*pmd)) 47109e5fc74cSSteve Capper return (pte_t *)pmd; 47119b19df29SPunit Agrawal 47129b19df29SPunit Agrawal return NULL; 47139e5fc74cSSteve Capper } 47149e5fc74cSSteve Capper 471561f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 471661f77edaSNaoya Horiguchi 471761f77edaSNaoya Horiguchi /* 471861f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 471961f77edaSNaoya Horiguchi * behavior. 472061f77edaSNaoya Horiguchi */ 472161f77edaSNaoya Horiguchi struct page * __weak 472261f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 472361f77edaSNaoya Horiguchi int write) 472461f77edaSNaoya Horiguchi { 472561f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 472661f77edaSNaoya Horiguchi } 472761f77edaSNaoya Horiguchi 472861f77edaSNaoya Horiguchi struct page * __weak 47294dc71451SAneesh Kumar K.V follow_huge_pd(struct vm_area_struct *vma, 47304dc71451SAneesh Kumar K.V unsigned long address, hugepd_t hpd, int flags, int pdshift) 47314dc71451SAneesh Kumar K.V { 47324dc71451SAneesh Kumar K.V WARN(1, "hugepd follow called with no support for hugepage directory format\n"); 47334dc71451SAneesh Kumar K.V return NULL; 47344dc71451SAneesh Kumar K.V } 47354dc71451SAneesh Kumar K.V 47364dc71451SAneesh Kumar K.V struct page * __weak 47379e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 4738e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 47399e5fc74cSSteve Capper { 4740e66f17ffSNaoya Horiguchi struct page *page = NULL; 4741e66f17ffSNaoya Horiguchi spinlock_t *ptl; 4742c9d398faSNaoya Horiguchi pte_t pte; 4743e66f17ffSNaoya Horiguchi retry: 4744e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 4745e66f17ffSNaoya Horiguchi spin_lock(ptl); 4746e66f17ffSNaoya Horiguchi /* 4747e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 4748e66f17ffSNaoya Horiguchi * unmapped from other threads. 4749e66f17ffSNaoya Horiguchi */ 4750e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 4751e66f17ffSNaoya Horiguchi goto out; 4752c9d398faSNaoya Horiguchi pte = huge_ptep_get((pte_t *)pmd); 4753c9d398faSNaoya Horiguchi if (pte_present(pte)) { 475497534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 4755e66f17ffSNaoya Horiguchi if (flags & FOLL_GET) 4756e66f17ffSNaoya Horiguchi get_page(page); 4757e66f17ffSNaoya Horiguchi } else { 4758c9d398faSNaoya Horiguchi if (is_hugetlb_entry_migration(pte)) { 4759e66f17ffSNaoya Horiguchi spin_unlock(ptl); 4760e66f17ffSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pmd, ptl); 4761e66f17ffSNaoya Horiguchi goto retry; 4762e66f17ffSNaoya Horiguchi } 4763e66f17ffSNaoya Horiguchi /* 4764e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 4765e66f17ffSNaoya Horiguchi * follow_page_mask(). 4766e66f17ffSNaoya Horiguchi */ 4767e66f17ffSNaoya Horiguchi } 4768e66f17ffSNaoya Horiguchi out: 4769e66f17ffSNaoya Horiguchi spin_unlock(ptl); 47709e5fc74cSSteve Capper return page; 47719e5fc74cSSteve Capper } 47729e5fc74cSSteve Capper 477361f77edaSNaoya Horiguchi struct page * __weak 47749e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 4775e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 47769e5fc74cSSteve Capper { 4777e66f17ffSNaoya Horiguchi if (flags & FOLL_GET) 4778e66f17ffSNaoya Horiguchi return NULL; 47799e5fc74cSSteve Capper 4780e66f17ffSNaoya Horiguchi return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 47819e5fc74cSSteve Capper } 47829e5fc74cSSteve Capper 4783faaa5b62SAnshuman Khandual struct page * __weak 4784faaa5b62SAnshuman Khandual follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) 4785faaa5b62SAnshuman Khandual { 4786faaa5b62SAnshuman Khandual if (flags & FOLL_GET) 4787faaa5b62SAnshuman Khandual return NULL; 4788faaa5b62SAnshuman Khandual 4789faaa5b62SAnshuman Khandual return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); 4790faaa5b62SAnshuman Khandual } 4791faaa5b62SAnshuman Khandual 479231caf665SNaoya Horiguchi bool isolate_huge_page(struct page *page, struct list_head *list) 479331caf665SNaoya Horiguchi { 4794bcc54222SNaoya Horiguchi bool ret = true; 4795bcc54222SNaoya Horiguchi 4796309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 479731caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 4798bcc54222SNaoya Horiguchi if (!page_huge_active(page) || !get_page_unless_zero(page)) { 4799bcc54222SNaoya Horiguchi ret = false; 4800bcc54222SNaoya Horiguchi goto unlock; 4801bcc54222SNaoya Horiguchi } 4802bcc54222SNaoya Horiguchi clear_page_huge_active(page); 480331caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 4804bcc54222SNaoya Horiguchi unlock: 480531caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 4806bcc54222SNaoya Horiguchi return ret; 480731caf665SNaoya Horiguchi } 480831caf665SNaoya Horiguchi 480931caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 481031caf665SNaoya Horiguchi { 4811309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 481231caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 4813bcc54222SNaoya Horiguchi set_page_huge_active(page); 481431caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 481531caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 481631caf665SNaoya Horiguchi put_page(page); 481731caf665SNaoya Horiguchi } 4818ab5ac90aSMichal Hocko 4819ab5ac90aSMichal Hocko void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) 4820ab5ac90aSMichal Hocko { 4821ab5ac90aSMichal Hocko struct hstate *h = page_hstate(oldpage); 4822ab5ac90aSMichal Hocko 4823ab5ac90aSMichal Hocko hugetlb_cgroup_migrate(oldpage, newpage); 4824ab5ac90aSMichal Hocko set_page_owner_migrate_reason(newpage, reason); 4825ab5ac90aSMichal Hocko 4826ab5ac90aSMichal Hocko /* 4827ab5ac90aSMichal Hocko * transfer temporary state of the new huge page. This is 4828ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 4829ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 4830ab5ac90aSMichal Hocko * the temporary status. 4831ab5ac90aSMichal Hocko * 4832ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 4833ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 4834ab5ac90aSMichal Hocko * the per-node's. 4835ab5ac90aSMichal Hocko */ 4836ab5ac90aSMichal Hocko if (PageHugeTemporary(newpage)) { 4837ab5ac90aSMichal Hocko int old_nid = page_to_nid(oldpage); 4838ab5ac90aSMichal Hocko int new_nid = page_to_nid(newpage); 4839ab5ac90aSMichal Hocko 4840ab5ac90aSMichal Hocko SetPageHugeTemporary(oldpage); 4841ab5ac90aSMichal Hocko ClearPageHugeTemporary(newpage); 4842ab5ac90aSMichal Hocko 4843ab5ac90aSMichal Hocko spin_lock(&hugetlb_lock); 4844ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 4845ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 4846ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 4847ab5ac90aSMichal Hocko } 4848ab5ac90aSMichal Hocko spin_unlock(&hugetlb_lock); 4849ab5ac90aSMichal Hocko } 4850ab5ac90aSMichal Hocko } 4851