11da177e4SLinus Torvalds /* 21da177e4SLinus Torvalds * Generic hugetlb support. 36d49e352SNadia Yvette Chambers * (C) Nadia Yvette Chambers, April 2004 41da177e4SLinus Torvalds */ 51da177e4SLinus Torvalds #include <linux/list.h> 61da177e4SLinus Torvalds #include <linux/init.h> 71da177e4SLinus Torvalds #include <linux/mm.h> 8e1759c21SAlexey Dobriyan #include <linux/seq_file.h> 91da177e4SLinus Torvalds #include <linux/sysctl.h> 101da177e4SLinus Torvalds #include <linux/highmem.h> 11cddb8a5cSAndrea Arcangeli #include <linux/mmu_notifier.h> 121da177e4SLinus Torvalds #include <linux/nodemask.h> 1363551ae0SDavid Gibson #include <linux/pagemap.h> 145da7ca86SChristoph Lameter #include <linux/mempolicy.h> 153b32123dSGideon Israel Dsouza #include <linux/compiler.h> 16aea47ff3SChristoph Lameter #include <linux/cpuset.h> 173935baa9SDavid Gibson #include <linux/mutex.h> 18aa888a74SAndi Kleen #include <linux/bootmem.h> 19a3437870SNishanth Aravamudan #include <linux/sysfs.h> 205a0e3ad6STejun Heo #include <linux/slab.h> 210fe6e20bSNaoya Horiguchi #include <linux/rmap.h> 22fd6a03edSNaoya Horiguchi #include <linux/swap.h> 23fd6a03edSNaoya Horiguchi #include <linux/swapops.h> 24c8721bbbSNaoya Horiguchi #include <linux/page-isolation.h> 258382d914SDavidlohr Bueso #include <linux/jhash.h> 26d6606683SLinus Torvalds 2763551ae0SDavid Gibson #include <asm/page.h> 2863551ae0SDavid Gibson #include <asm/pgtable.h> 2924669e58SAneesh Kumar K.V #include <asm/tlb.h> 3063551ae0SDavid Gibson 3124669e58SAneesh Kumar K.V #include <linux/io.h> 3263551ae0SDavid Gibson #include <linux/hugetlb.h> 339dd540e2SAneesh Kumar K.V #include <linux/hugetlb_cgroup.h> 349a305230SLee Schermerhorn #include <linux/node.h> 357835e98bSNick Piggin #include "internal.h" 361da177e4SLinus Torvalds 37753162cdSAndrey Ryabinin int hugepages_treat_as_movable; 38a5516438SAndi Kleen 39c3f38a38SAneesh Kumar K.V int hugetlb_max_hstate __read_mostly; 40e5ff2159SAndi Kleen unsigned int default_hstate_idx; 41e5ff2159SAndi Kleen struct hstate hstates[HUGE_MAX_HSTATE]; 42641844f5SNaoya Horiguchi /* 43641844f5SNaoya Horiguchi * Minimum page order among possible hugepage sizes, set to a proper value 44641844f5SNaoya Horiguchi * at boot time. 45641844f5SNaoya Horiguchi */ 46641844f5SNaoya Horiguchi static unsigned int minimum_order __read_mostly = UINT_MAX; 47e5ff2159SAndi Kleen 4853ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 4953ba51d2SJon Tollefson 50e5ff2159SAndi Kleen /* for command line parsing */ 51e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 52e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 53e11bfbfcSNick Piggin static unsigned long __initdata default_hstate_size; 54e5ff2159SAndi Kleen 553935baa9SDavid Gibson /* 5631caf665SNaoya Horiguchi * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, 5731caf665SNaoya Horiguchi * free_huge_pages, and surplus_huge_pages. 583935baa9SDavid Gibson */ 59c3f38a38SAneesh Kumar K.V DEFINE_SPINLOCK(hugetlb_lock); 600bd0f9fbSEric Paris 618382d914SDavidlohr Bueso /* 628382d914SDavidlohr Bueso * Serializes faults on the same logical page. This is used to 638382d914SDavidlohr Bueso * prevent spurious OOMs when the hugepage pool is fully utilized. 648382d914SDavidlohr Bueso */ 658382d914SDavidlohr Bueso static int num_fault_mutexes; 66c672c7f2SMike Kravetz struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; 678382d914SDavidlohr Bueso 687ca02d0aSMike Kravetz /* Forward declaration */ 697ca02d0aSMike Kravetz static int hugetlb_acct_memory(struct hstate *h, long delta); 707ca02d0aSMike Kravetz 7190481622SDavid Gibson static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) 7290481622SDavid Gibson { 7390481622SDavid Gibson bool free = (spool->count == 0) && (spool->used_hpages == 0); 7490481622SDavid Gibson 7590481622SDavid Gibson spin_unlock(&spool->lock); 7690481622SDavid Gibson 7790481622SDavid Gibson /* If no pages are used, and no other handles to the subpool 787ca02d0aSMike Kravetz * remain, give up any reservations mased on minimum size and 797ca02d0aSMike Kravetz * free the subpool */ 807ca02d0aSMike Kravetz if (free) { 817ca02d0aSMike Kravetz if (spool->min_hpages != -1) 827ca02d0aSMike Kravetz hugetlb_acct_memory(spool->hstate, 837ca02d0aSMike Kravetz -spool->min_hpages); 8490481622SDavid Gibson kfree(spool); 8590481622SDavid Gibson } 867ca02d0aSMike Kravetz } 8790481622SDavid Gibson 887ca02d0aSMike Kravetz struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, 897ca02d0aSMike Kravetz long min_hpages) 9090481622SDavid Gibson { 9190481622SDavid Gibson struct hugepage_subpool *spool; 9290481622SDavid Gibson 93c6a91820SMike Kravetz spool = kzalloc(sizeof(*spool), GFP_KERNEL); 9490481622SDavid Gibson if (!spool) 9590481622SDavid Gibson return NULL; 9690481622SDavid Gibson 9790481622SDavid Gibson spin_lock_init(&spool->lock); 9890481622SDavid Gibson spool->count = 1; 997ca02d0aSMike Kravetz spool->max_hpages = max_hpages; 1007ca02d0aSMike Kravetz spool->hstate = h; 1017ca02d0aSMike Kravetz spool->min_hpages = min_hpages; 1027ca02d0aSMike Kravetz 1037ca02d0aSMike Kravetz if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { 1047ca02d0aSMike Kravetz kfree(spool); 1057ca02d0aSMike Kravetz return NULL; 1067ca02d0aSMike Kravetz } 1077ca02d0aSMike Kravetz spool->rsv_hpages = min_hpages; 10890481622SDavid Gibson 10990481622SDavid Gibson return spool; 11090481622SDavid Gibson } 11190481622SDavid Gibson 11290481622SDavid Gibson void hugepage_put_subpool(struct hugepage_subpool *spool) 11390481622SDavid Gibson { 11490481622SDavid Gibson spin_lock(&spool->lock); 11590481622SDavid Gibson BUG_ON(!spool->count); 11690481622SDavid Gibson spool->count--; 11790481622SDavid Gibson unlock_or_release_subpool(spool); 11890481622SDavid Gibson } 11990481622SDavid Gibson 1201c5ecae3SMike Kravetz /* 1211c5ecae3SMike Kravetz * Subpool accounting for allocating and reserving pages. 1221c5ecae3SMike Kravetz * Return -ENOMEM if there are not enough resources to satisfy the 1231c5ecae3SMike Kravetz * the request. Otherwise, return the number of pages by which the 1241c5ecae3SMike Kravetz * global pools must be adjusted (upward). The returned value may 1251c5ecae3SMike Kravetz * only be different than the passed value (delta) in the case where 1261c5ecae3SMike Kravetz * a subpool minimum size must be manitained. 1271c5ecae3SMike Kravetz */ 1281c5ecae3SMike Kravetz static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, 12990481622SDavid Gibson long delta) 13090481622SDavid Gibson { 1311c5ecae3SMike Kravetz long ret = delta; 13290481622SDavid Gibson 13390481622SDavid Gibson if (!spool) 1341c5ecae3SMike Kravetz return ret; 13590481622SDavid Gibson 13690481622SDavid Gibson spin_lock(&spool->lock); 13790481622SDavid Gibson 1381c5ecae3SMike Kravetz if (spool->max_hpages != -1) { /* maximum size accounting */ 1391c5ecae3SMike Kravetz if ((spool->used_hpages + delta) <= spool->max_hpages) 1401c5ecae3SMike Kravetz spool->used_hpages += delta; 1411c5ecae3SMike Kravetz else { 1421c5ecae3SMike Kravetz ret = -ENOMEM; 1431c5ecae3SMike Kravetz goto unlock_ret; 1441c5ecae3SMike Kravetz } 1451c5ecae3SMike Kravetz } 1461c5ecae3SMike Kravetz 1471c5ecae3SMike Kravetz if (spool->min_hpages != -1) { /* minimum size accounting */ 1481c5ecae3SMike Kravetz if (delta > spool->rsv_hpages) { 1491c5ecae3SMike Kravetz /* 1501c5ecae3SMike Kravetz * Asking for more reserves than those already taken on 1511c5ecae3SMike Kravetz * behalf of subpool. Return difference. 1521c5ecae3SMike Kravetz */ 1531c5ecae3SMike Kravetz ret = delta - spool->rsv_hpages; 1541c5ecae3SMike Kravetz spool->rsv_hpages = 0; 1551c5ecae3SMike Kravetz } else { 1561c5ecae3SMike Kravetz ret = 0; /* reserves already accounted for */ 1571c5ecae3SMike Kravetz spool->rsv_hpages -= delta; 1581c5ecae3SMike Kravetz } 1591c5ecae3SMike Kravetz } 1601c5ecae3SMike Kravetz 1611c5ecae3SMike Kravetz unlock_ret: 1621c5ecae3SMike Kravetz spin_unlock(&spool->lock); 16390481622SDavid Gibson return ret; 16490481622SDavid Gibson } 16590481622SDavid Gibson 1661c5ecae3SMike Kravetz /* 1671c5ecae3SMike Kravetz * Subpool accounting for freeing and unreserving pages. 1681c5ecae3SMike Kravetz * Return the number of global page reservations that must be dropped. 1691c5ecae3SMike Kravetz * The return value may only be different than the passed value (delta) 1701c5ecae3SMike Kravetz * in the case where a subpool minimum size must be maintained. 1711c5ecae3SMike Kravetz */ 1721c5ecae3SMike Kravetz static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, 17390481622SDavid Gibson long delta) 17490481622SDavid Gibson { 1751c5ecae3SMike Kravetz long ret = delta; 1761c5ecae3SMike Kravetz 17790481622SDavid Gibson if (!spool) 1781c5ecae3SMike Kravetz return delta; 17990481622SDavid Gibson 18090481622SDavid Gibson spin_lock(&spool->lock); 1811c5ecae3SMike Kravetz 1821c5ecae3SMike Kravetz if (spool->max_hpages != -1) /* maximum size accounting */ 18390481622SDavid Gibson spool->used_hpages -= delta; 1841c5ecae3SMike Kravetz 1851c5ecae3SMike Kravetz if (spool->min_hpages != -1) { /* minimum size accounting */ 1861c5ecae3SMike Kravetz if (spool->rsv_hpages + delta <= spool->min_hpages) 1871c5ecae3SMike Kravetz ret = 0; 1881c5ecae3SMike Kravetz else 1891c5ecae3SMike Kravetz ret = spool->rsv_hpages + delta - spool->min_hpages; 1901c5ecae3SMike Kravetz 1911c5ecae3SMike Kravetz spool->rsv_hpages += delta; 1921c5ecae3SMike Kravetz if (spool->rsv_hpages > spool->min_hpages) 1931c5ecae3SMike Kravetz spool->rsv_hpages = spool->min_hpages; 1941c5ecae3SMike Kravetz } 1951c5ecae3SMike Kravetz 1961c5ecae3SMike Kravetz /* 1971c5ecae3SMike Kravetz * If hugetlbfs_put_super couldn't free spool due to an outstanding 1981c5ecae3SMike Kravetz * quota reference, free it now. 1991c5ecae3SMike Kravetz */ 20090481622SDavid Gibson unlock_or_release_subpool(spool); 2011c5ecae3SMike Kravetz 2021c5ecae3SMike Kravetz return ret; 20390481622SDavid Gibson } 20490481622SDavid Gibson 20590481622SDavid Gibson static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 20690481622SDavid Gibson { 20790481622SDavid Gibson return HUGETLBFS_SB(inode->i_sb)->spool; 20890481622SDavid Gibson } 20990481622SDavid Gibson 21090481622SDavid Gibson static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 21190481622SDavid Gibson { 212496ad9aaSAl Viro return subpool_inode(file_inode(vma->vm_file)); 21390481622SDavid Gibson } 21490481622SDavid Gibson 215e7c4b0bfSAndy Whitcroft /* 21696822904SAndy Whitcroft * Region tracking -- allows tracking of reservations and instantiated pages 21796822904SAndy Whitcroft * across the pages in a mapping. 21884afd99bSAndy Whitcroft * 2191dd308a7SMike Kravetz * The region data structures are embedded into a resv_map and protected 2201dd308a7SMike Kravetz * by a resv_map's lock. The set of regions within the resv_map represent 2211dd308a7SMike Kravetz * reservations for huge pages, or huge pages that have already been 2221dd308a7SMike Kravetz * instantiated within the map. The from and to elements are huge page 2231dd308a7SMike Kravetz * indicies into the associated mapping. from indicates the starting index 2241dd308a7SMike Kravetz * of the region. to represents the first index past the end of the region. 2251dd308a7SMike Kravetz * 2261dd308a7SMike Kravetz * For example, a file region structure with from == 0 and to == 4 represents 2271dd308a7SMike Kravetz * four huge pages in a mapping. It is important to note that the to element 2281dd308a7SMike Kravetz * represents the first element past the end of the region. This is used in 2291dd308a7SMike Kravetz * arithmetic as 4(to) - 0(from) = 4 huge pages in the region. 2301dd308a7SMike Kravetz * 2311dd308a7SMike Kravetz * Interval notation of the form [from, to) will be used to indicate that 2321dd308a7SMike Kravetz * the endpoint from is inclusive and to is exclusive. 23396822904SAndy Whitcroft */ 23496822904SAndy Whitcroft struct file_region { 23596822904SAndy Whitcroft struct list_head link; 23696822904SAndy Whitcroft long from; 23796822904SAndy Whitcroft long to; 23896822904SAndy Whitcroft }; 23996822904SAndy Whitcroft 2401dd308a7SMike Kravetz /* 2411dd308a7SMike Kravetz * Add the huge page range represented by [f, t) to the reserve 2425e911373SMike Kravetz * map. In the normal case, existing regions will be expanded 2435e911373SMike Kravetz * to accommodate the specified range. Sufficient regions should 2445e911373SMike Kravetz * exist for expansion due to the previous call to region_chg 2455e911373SMike Kravetz * with the same range. However, it is possible that region_del 2465e911373SMike Kravetz * could have been called after region_chg and modifed the map 2475e911373SMike Kravetz * in such a way that no region exists to be expanded. In this 2485e911373SMike Kravetz * case, pull a region descriptor from the cache associated with 2495e911373SMike Kravetz * the map and use that for the new range. 250cf3ad20bSMike Kravetz * 251cf3ad20bSMike Kravetz * Return the number of new huge pages added to the map. This 252cf3ad20bSMike Kravetz * number is greater than or equal to zero. 2531dd308a7SMike Kravetz */ 2541406ec9bSJoonsoo Kim static long region_add(struct resv_map *resv, long f, long t) 25596822904SAndy Whitcroft { 2561406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 25796822904SAndy Whitcroft struct file_region *rg, *nrg, *trg; 258cf3ad20bSMike Kravetz long add = 0; 25996822904SAndy Whitcroft 2607b24d861SDavidlohr Bueso spin_lock(&resv->lock); 26196822904SAndy Whitcroft /* Locate the region we are either in or before. */ 26296822904SAndy Whitcroft list_for_each_entry(rg, head, link) 26396822904SAndy Whitcroft if (f <= rg->to) 26496822904SAndy Whitcroft break; 26596822904SAndy Whitcroft 2665e911373SMike Kravetz /* 2675e911373SMike Kravetz * If no region exists which can be expanded to include the 2685e911373SMike Kravetz * specified range, the list must have been modified by an 2695e911373SMike Kravetz * interleving call to region_del(). Pull a region descriptor 2705e911373SMike Kravetz * from the cache and use it for this range. 2715e911373SMike Kravetz */ 2725e911373SMike Kravetz if (&rg->link == head || t < rg->from) { 2735e911373SMike Kravetz VM_BUG_ON(resv->region_cache_count <= 0); 2745e911373SMike Kravetz 2755e911373SMike Kravetz resv->region_cache_count--; 2765e911373SMike Kravetz nrg = list_first_entry(&resv->region_cache, struct file_region, 2775e911373SMike Kravetz link); 2785e911373SMike Kravetz list_del(&nrg->link); 2795e911373SMike Kravetz 2805e911373SMike Kravetz nrg->from = f; 2815e911373SMike Kravetz nrg->to = t; 2825e911373SMike Kravetz list_add(&nrg->link, rg->link.prev); 2835e911373SMike Kravetz 2845e911373SMike Kravetz add += t - f; 2855e911373SMike Kravetz goto out_locked; 2865e911373SMike Kravetz } 2875e911373SMike Kravetz 28896822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 28996822904SAndy Whitcroft if (f > rg->from) 29096822904SAndy Whitcroft f = rg->from; 29196822904SAndy Whitcroft 29296822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 29396822904SAndy Whitcroft nrg = rg; 29496822904SAndy Whitcroft list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 29596822904SAndy Whitcroft if (&rg->link == head) 29696822904SAndy Whitcroft break; 29796822904SAndy Whitcroft if (rg->from > t) 29896822904SAndy Whitcroft break; 29996822904SAndy Whitcroft 30096822904SAndy Whitcroft /* If this area reaches higher then extend our area to 30196822904SAndy Whitcroft * include it completely. If this is not the first area 30296822904SAndy Whitcroft * which we intend to reuse, free it. */ 30396822904SAndy Whitcroft if (rg->to > t) 30496822904SAndy Whitcroft t = rg->to; 30596822904SAndy Whitcroft if (rg != nrg) { 306cf3ad20bSMike Kravetz /* Decrement return value by the deleted range. 307cf3ad20bSMike Kravetz * Another range will span this area so that by 308cf3ad20bSMike Kravetz * end of routine add will be >= zero 309cf3ad20bSMike Kravetz */ 310cf3ad20bSMike Kravetz add -= (rg->to - rg->from); 31196822904SAndy Whitcroft list_del(&rg->link); 31296822904SAndy Whitcroft kfree(rg); 31396822904SAndy Whitcroft } 31496822904SAndy Whitcroft } 315cf3ad20bSMike Kravetz 316cf3ad20bSMike Kravetz add += (nrg->from - f); /* Added to beginning of region */ 31796822904SAndy Whitcroft nrg->from = f; 318cf3ad20bSMike Kravetz add += t - nrg->to; /* Added to end of region */ 31996822904SAndy Whitcroft nrg->to = t; 320cf3ad20bSMike Kravetz 3215e911373SMike Kravetz out_locked: 3225e911373SMike Kravetz resv->adds_in_progress--; 3237b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 324cf3ad20bSMike Kravetz VM_BUG_ON(add < 0); 325cf3ad20bSMike Kravetz return add; 32696822904SAndy Whitcroft } 32796822904SAndy Whitcroft 3281dd308a7SMike Kravetz /* 3291dd308a7SMike Kravetz * Examine the existing reserve map and determine how many 3301dd308a7SMike Kravetz * huge pages in the specified range [f, t) are NOT currently 3311dd308a7SMike Kravetz * represented. This routine is called before a subsequent 3321dd308a7SMike Kravetz * call to region_add that will actually modify the reserve 3331dd308a7SMike Kravetz * map to add the specified range [f, t). region_chg does 3341dd308a7SMike Kravetz * not change the number of huge pages represented by the 3351dd308a7SMike Kravetz * map. However, if the existing regions in the map can not 3361dd308a7SMike Kravetz * be expanded to represent the new range, a new file_region 3371dd308a7SMike Kravetz * structure is added to the map as a placeholder. This is 3381dd308a7SMike Kravetz * so that the subsequent region_add call will have all the 3391dd308a7SMike Kravetz * regions it needs and will not fail. 3401dd308a7SMike Kravetz * 3415e911373SMike Kravetz * Upon entry, region_chg will also examine the cache of region descriptors 3425e911373SMike Kravetz * associated with the map. If there are not enough descriptors cached, one 3435e911373SMike Kravetz * will be allocated for the in progress add operation. 3445e911373SMike Kravetz * 3455e911373SMike Kravetz * Returns the number of huge pages that need to be added to the existing 3465e911373SMike Kravetz * reservation map for the range [f, t). This number is greater or equal to 3475e911373SMike Kravetz * zero. -ENOMEM is returned if a new file_region structure or cache entry 3485e911373SMike Kravetz * is needed and can not be allocated. 3491dd308a7SMike Kravetz */ 3501406ec9bSJoonsoo Kim static long region_chg(struct resv_map *resv, long f, long t) 35196822904SAndy Whitcroft { 3521406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 3537b24d861SDavidlohr Bueso struct file_region *rg, *nrg = NULL; 35496822904SAndy Whitcroft long chg = 0; 35596822904SAndy Whitcroft 3567b24d861SDavidlohr Bueso retry: 3577b24d861SDavidlohr Bueso spin_lock(&resv->lock); 3585e911373SMike Kravetz retry_locked: 3595e911373SMike Kravetz resv->adds_in_progress++; 3605e911373SMike Kravetz 3615e911373SMike Kravetz /* 3625e911373SMike Kravetz * Check for sufficient descriptors in the cache to accommodate 3635e911373SMike Kravetz * the number of in progress add operations. 3645e911373SMike Kravetz */ 3655e911373SMike Kravetz if (resv->adds_in_progress > resv->region_cache_count) { 3665e911373SMike Kravetz struct file_region *trg; 3675e911373SMike Kravetz 3685e911373SMike Kravetz VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1); 3695e911373SMike Kravetz /* Must drop lock to allocate a new descriptor. */ 3705e911373SMike Kravetz resv->adds_in_progress--; 3715e911373SMike Kravetz spin_unlock(&resv->lock); 3725e911373SMike Kravetz 3735e911373SMike Kravetz trg = kmalloc(sizeof(*trg), GFP_KERNEL); 374dbe409e4SMike Kravetz if (!trg) { 375dbe409e4SMike Kravetz kfree(nrg); 3765e911373SMike Kravetz return -ENOMEM; 377dbe409e4SMike Kravetz } 3785e911373SMike Kravetz 3795e911373SMike Kravetz spin_lock(&resv->lock); 3805e911373SMike Kravetz list_add(&trg->link, &resv->region_cache); 3815e911373SMike Kravetz resv->region_cache_count++; 3825e911373SMike Kravetz goto retry_locked; 3835e911373SMike Kravetz } 3845e911373SMike Kravetz 38596822904SAndy Whitcroft /* Locate the region we are before or in. */ 38696822904SAndy Whitcroft list_for_each_entry(rg, head, link) 38796822904SAndy Whitcroft if (f <= rg->to) 38896822904SAndy Whitcroft break; 38996822904SAndy Whitcroft 39096822904SAndy Whitcroft /* If we are below the current region then a new region is required. 39196822904SAndy Whitcroft * Subtle, allocate a new region at the position but make it zero 39296822904SAndy Whitcroft * size such that we can guarantee to record the reservation. */ 39396822904SAndy Whitcroft if (&rg->link == head || t < rg->from) { 3947b24d861SDavidlohr Bueso if (!nrg) { 3955e911373SMike Kravetz resv->adds_in_progress--; 3967b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 39796822904SAndy Whitcroft nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 39896822904SAndy Whitcroft if (!nrg) 39996822904SAndy Whitcroft return -ENOMEM; 4007b24d861SDavidlohr Bueso 40196822904SAndy Whitcroft nrg->from = f; 40296822904SAndy Whitcroft nrg->to = f; 40396822904SAndy Whitcroft INIT_LIST_HEAD(&nrg->link); 4047b24d861SDavidlohr Bueso goto retry; 4057b24d861SDavidlohr Bueso } 40696822904SAndy Whitcroft 4077b24d861SDavidlohr Bueso list_add(&nrg->link, rg->link.prev); 4087b24d861SDavidlohr Bueso chg = t - f; 4097b24d861SDavidlohr Bueso goto out_nrg; 41096822904SAndy Whitcroft } 41196822904SAndy Whitcroft 41296822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 41396822904SAndy Whitcroft if (f > rg->from) 41496822904SAndy Whitcroft f = rg->from; 41596822904SAndy Whitcroft chg = t - f; 41696822904SAndy Whitcroft 41796822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 41896822904SAndy Whitcroft list_for_each_entry(rg, rg->link.prev, link) { 41996822904SAndy Whitcroft if (&rg->link == head) 42096822904SAndy Whitcroft break; 42196822904SAndy Whitcroft if (rg->from > t) 4227b24d861SDavidlohr Bueso goto out; 42396822904SAndy Whitcroft 42425985edcSLucas De Marchi /* We overlap with this area, if it extends further than 42596822904SAndy Whitcroft * us then we must extend ourselves. Account for its 42696822904SAndy Whitcroft * existing reservation. */ 42796822904SAndy Whitcroft if (rg->to > t) { 42896822904SAndy Whitcroft chg += rg->to - t; 42996822904SAndy Whitcroft t = rg->to; 43096822904SAndy Whitcroft } 43196822904SAndy Whitcroft chg -= rg->to - rg->from; 43296822904SAndy Whitcroft } 4337b24d861SDavidlohr Bueso 4347b24d861SDavidlohr Bueso out: 4357b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 4367b24d861SDavidlohr Bueso /* We already know we raced and no longer need the new region */ 4377b24d861SDavidlohr Bueso kfree(nrg); 4387b24d861SDavidlohr Bueso return chg; 4397b24d861SDavidlohr Bueso out_nrg: 4407b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 44196822904SAndy Whitcroft return chg; 44296822904SAndy Whitcroft } 44396822904SAndy Whitcroft 4441dd308a7SMike Kravetz /* 4455e911373SMike Kravetz * Abort the in progress add operation. The adds_in_progress field 4465e911373SMike Kravetz * of the resv_map keeps track of the operations in progress between 4475e911373SMike Kravetz * calls to region_chg and region_add. Operations are sometimes 4485e911373SMike Kravetz * aborted after the call to region_chg. In such cases, region_abort 4495e911373SMike Kravetz * is called to decrement the adds_in_progress counter. 4505e911373SMike Kravetz * 4515e911373SMike Kravetz * NOTE: The range arguments [f, t) are not needed or used in this 4525e911373SMike Kravetz * routine. They are kept to make reading the calling code easier as 4535e911373SMike Kravetz * arguments will match the associated region_chg call. 4545e911373SMike Kravetz */ 4555e911373SMike Kravetz static void region_abort(struct resv_map *resv, long f, long t) 4565e911373SMike Kravetz { 4575e911373SMike Kravetz spin_lock(&resv->lock); 4585e911373SMike Kravetz VM_BUG_ON(!resv->region_cache_count); 4595e911373SMike Kravetz resv->adds_in_progress--; 4605e911373SMike Kravetz spin_unlock(&resv->lock); 4615e911373SMike Kravetz } 4625e911373SMike Kravetz 4635e911373SMike Kravetz /* 464feba16e2SMike Kravetz * Delete the specified range [f, t) from the reserve map. If the 465feba16e2SMike Kravetz * t parameter is LONG_MAX, this indicates that ALL regions after f 466feba16e2SMike Kravetz * should be deleted. Locate the regions which intersect [f, t) 467feba16e2SMike Kravetz * and either trim, delete or split the existing regions. 468feba16e2SMike Kravetz * 469feba16e2SMike Kravetz * Returns the number of huge pages deleted from the reserve map. 470feba16e2SMike Kravetz * In the normal case, the return value is zero or more. In the 471feba16e2SMike Kravetz * case where a region must be split, a new region descriptor must 472feba16e2SMike Kravetz * be allocated. If the allocation fails, -ENOMEM will be returned. 473feba16e2SMike Kravetz * NOTE: If the parameter t == LONG_MAX, then we will never split 474feba16e2SMike Kravetz * a region and possibly return -ENOMEM. Callers specifying 475feba16e2SMike Kravetz * t == LONG_MAX do not need to check for -ENOMEM error. 4761dd308a7SMike Kravetz */ 477feba16e2SMike Kravetz static long region_del(struct resv_map *resv, long f, long t) 47896822904SAndy Whitcroft { 4791406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 48096822904SAndy Whitcroft struct file_region *rg, *trg; 481feba16e2SMike Kravetz struct file_region *nrg = NULL; 482feba16e2SMike Kravetz long del = 0; 48396822904SAndy Whitcroft 484feba16e2SMike Kravetz retry: 4857b24d861SDavidlohr Bueso spin_lock(&resv->lock); 486feba16e2SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 487dbe409e4SMike Kravetz /* 488dbe409e4SMike Kravetz * Skip regions before the range to be deleted. file_region 489dbe409e4SMike Kravetz * ranges are normally of the form [from, to). However, there 490dbe409e4SMike Kravetz * may be a "placeholder" entry in the map which is of the form 491dbe409e4SMike Kravetz * (from, to) with from == to. Check for placeholder entries 492dbe409e4SMike Kravetz * at the beginning of the range to be deleted. 493dbe409e4SMike Kravetz */ 494dbe409e4SMike Kravetz if (rg->to <= f && (rg->to != rg->from || rg->to != f)) 495feba16e2SMike Kravetz continue; 496dbe409e4SMike Kravetz 497feba16e2SMike Kravetz if (rg->from >= t) 49896822904SAndy Whitcroft break; 49996822904SAndy Whitcroft 500feba16e2SMike Kravetz if (f > rg->from && t < rg->to) { /* Must split region */ 501feba16e2SMike Kravetz /* 502feba16e2SMike Kravetz * Check for an entry in the cache before dropping 503feba16e2SMike Kravetz * lock and attempting allocation. 504feba16e2SMike Kravetz */ 505feba16e2SMike Kravetz if (!nrg && 506feba16e2SMike Kravetz resv->region_cache_count > resv->adds_in_progress) { 507feba16e2SMike Kravetz nrg = list_first_entry(&resv->region_cache, 508feba16e2SMike Kravetz struct file_region, 509feba16e2SMike Kravetz link); 510feba16e2SMike Kravetz list_del(&nrg->link); 511feba16e2SMike Kravetz resv->region_cache_count--; 51296822904SAndy Whitcroft } 51396822904SAndy Whitcroft 514feba16e2SMike Kravetz if (!nrg) { 515feba16e2SMike Kravetz spin_unlock(&resv->lock); 516feba16e2SMike Kravetz nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 517feba16e2SMike Kravetz if (!nrg) 518feba16e2SMike Kravetz return -ENOMEM; 519feba16e2SMike Kravetz goto retry; 520feba16e2SMike Kravetz } 521feba16e2SMike Kravetz 522feba16e2SMike Kravetz del += t - f; 523feba16e2SMike Kravetz 524feba16e2SMike Kravetz /* New entry for end of split region */ 525feba16e2SMike Kravetz nrg->from = t; 526feba16e2SMike Kravetz nrg->to = rg->to; 527feba16e2SMike Kravetz INIT_LIST_HEAD(&nrg->link); 528feba16e2SMike Kravetz 529feba16e2SMike Kravetz /* Original entry is trimmed */ 530feba16e2SMike Kravetz rg->to = f; 531feba16e2SMike Kravetz 532feba16e2SMike Kravetz list_add(&nrg->link, &rg->link); 533feba16e2SMike Kravetz nrg = NULL; 53496822904SAndy Whitcroft break; 535feba16e2SMike Kravetz } 536feba16e2SMike Kravetz 537feba16e2SMike Kravetz if (f <= rg->from && t >= rg->to) { /* Remove entire region */ 538feba16e2SMike Kravetz del += rg->to - rg->from; 53996822904SAndy Whitcroft list_del(&rg->link); 54096822904SAndy Whitcroft kfree(rg); 541feba16e2SMike Kravetz continue; 54296822904SAndy Whitcroft } 5437b24d861SDavidlohr Bueso 544feba16e2SMike Kravetz if (f <= rg->from) { /* Trim beginning of region */ 545feba16e2SMike Kravetz del += t - rg->from; 546feba16e2SMike Kravetz rg->from = t; 547feba16e2SMike Kravetz } else { /* Trim end of region */ 548feba16e2SMike Kravetz del += rg->to - f; 549feba16e2SMike Kravetz rg->to = f; 550feba16e2SMike Kravetz } 551feba16e2SMike Kravetz } 552feba16e2SMike Kravetz 5537b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 554feba16e2SMike Kravetz kfree(nrg); 555feba16e2SMike Kravetz return del; 55696822904SAndy Whitcroft } 55796822904SAndy Whitcroft 5581dd308a7SMike Kravetz /* 559b5cec28dSMike Kravetz * A rare out of memory error was encountered which prevented removal of 560b5cec28dSMike Kravetz * the reserve map region for a page. The huge page itself was free'ed 561b5cec28dSMike Kravetz * and removed from the page cache. This routine will adjust the subpool 562b5cec28dSMike Kravetz * usage count, and the global reserve count if needed. By incrementing 563b5cec28dSMike Kravetz * these counts, the reserve map entry which could not be deleted will 564b5cec28dSMike Kravetz * appear as a "reserved" entry instead of simply dangling with incorrect 565b5cec28dSMike Kravetz * counts. 566b5cec28dSMike Kravetz */ 567b5cec28dSMike Kravetz void hugetlb_fix_reserve_counts(struct inode *inode, bool restore_reserve) 568b5cec28dSMike Kravetz { 569b5cec28dSMike Kravetz struct hugepage_subpool *spool = subpool_inode(inode); 570b5cec28dSMike Kravetz long rsv_adjust; 571b5cec28dSMike Kravetz 572b5cec28dSMike Kravetz rsv_adjust = hugepage_subpool_get_pages(spool, 1); 573b5cec28dSMike Kravetz if (restore_reserve && rsv_adjust) { 574b5cec28dSMike Kravetz struct hstate *h = hstate_inode(inode); 575b5cec28dSMike Kravetz 576b5cec28dSMike Kravetz hugetlb_acct_memory(h, 1); 577b5cec28dSMike Kravetz } 578b5cec28dSMike Kravetz } 579b5cec28dSMike Kravetz 580b5cec28dSMike Kravetz /* 5811dd308a7SMike Kravetz * Count and return the number of huge pages in the reserve map 5821dd308a7SMike Kravetz * that intersect with the range [f, t). 5831dd308a7SMike Kravetz */ 5841406ec9bSJoonsoo Kim static long region_count(struct resv_map *resv, long f, long t) 58584afd99bSAndy Whitcroft { 5861406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 58784afd99bSAndy Whitcroft struct file_region *rg; 58884afd99bSAndy Whitcroft long chg = 0; 58984afd99bSAndy Whitcroft 5907b24d861SDavidlohr Bueso spin_lock(&resv->lock); 59184afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 59284afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 593f2135a4aSWang Sheng-Hui long seg_from; 594f2135a4aSWang Sheng-Hui long seg_to; 59584afd99bSAndy Whitcroft 59684afd99bSAndy Whitcroft if (rg->to <= f) 59784afd99bSAndy Whitcroft continue; 59884afd99bSAndy Whitcroft if (rg->from >= t) 59984afd99bSAndy Whitcroft break; 60084afd99bSAndy Whitcroft 60184afd99bSAndy Whitcroft seg_from = max(rg->from, f); 60284afd99bSAndy Whitcroft seg_to = min(rg->to, t); 60384afd99bSAndy Whitcroft 60484afd99bSAndy Whitcroft chg += seg_to - seg_from; 60584afd99bSAndy Whitcroft } 6067b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 60784afd99bSAndy Whitcroft 60884afd99bSAndy Whitcroft return chg; 60984afd99bSAndy Whitcroft } 61084afd99bSAndy Whitcroft 61196822904SAndy Whitcroft /* 612e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 613e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 614e7c4b0bfSAndy Whitcroft */ 615a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 616a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 617e7c4b0bfSAndy Whitcroft { 618a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 619a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 620e7c4b0bfSAndy Whitcroft } 621e7c4b0bfSAndy Whitcroft 6220fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 6230fe6e20bSNaoya Horiguchi unsigned long address) 6240fe6e20bSNaoya Horiguchi { 6250fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 6260fe6e20bSNaoya Horiguchi } 6270fe6e20bSNaoya Horiguchi 62884afd99bSAndy Whitcroft /* 62908fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 63008fba699SMel Gorman * cases this will be same size as used by the page table entries. 63108fba699SMel Gorman */ 63208fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 63308fba699SMel Gorman { 63408fba699SMel Gorman struct hstate *hstate; 63508fba699SMel Gorman 63608fba699SMel Gorman if (!is_vm_hugetlb_page(vma)) 63708fba699SMel Gorman return PAGE_SIZE; 63808fba699SMel Gorman 63908fba699SMel Gorman hstate = hstate_vma(vma); 64008fba699SMel Gorman 6412415cf12SWanpeng Li return 1UL << huge_page_shift(hstate); 64208fba699SMel Gorman } 643f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 64408fba699SMel Gorman 64508fba699SMel Gorman /* 6463340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 6473340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 6483340289dSMel Gorman * architectures where it differs, an architecture-specific version of this 6493340289dSMel Gorman * function is required. 6503340289dSMel Gorman */ 6513340289dSMel Gorman #ifndef vma_mmu_pagesize 6523340289dSMel Gorman unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 6533340289dSMel Gorman { 6543340289dSMel Gorman return vma_kernel_pagesize(vma); 6553340289dSMel Gorman } 6563340289dSMel Gorman #endif 6573340289dSMel Gorman 6583340289dSMel Gorman /* 65984afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 66084afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 66184afd99bSAndy Whitcroft * alignment. 66284afd99bSAndy Whitcroft */ 66384afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 66484afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 66504f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 66684afd99bSAndy Whitcroft 667a1e78772SMel Gorman /* 668a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 669a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 670a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 671a1e78772SMel Gorman * 672a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 673a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 674a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 675a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 67684afd99bSAndy Whitcroft * 67784afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 67884afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 67984afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 68084afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 68184afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 68284afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 68384afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 68484afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 685a1e78772SMel Gorman */ 686e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 687e7c4b0bfSAndy Whitcroft { 688e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 689e7c4b0bfSAndy Whitcroft } 690e7c4b0bfSAndy Whitcroft 691e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 692e7c4b0bfSAndy Whitcroft unsigned long value) 693e7c4b0bfSAndy Whitcroft { 694e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 695e7c4b0bfSAndy Whitcroft } 696e7c4b0bfSAndy Whitcroft 6979119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 69884afd99bSAndy Whitcroft { 69984afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 7005e911373SMike Kravetz struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 7015e911373SMike Kravetz 7025e911373SMike Kravetz if (!resv_map || !rg) { 7035e911373SMike Kravetz kfree(resv_map); 7045e911373SMike Kravetz kfree(rg); 70584afd99bSAndy Whitcroft return NULL; 7065e911373SMike Kravetz } 70784afd99bSAndy Whitcroft 70884afd99bSAndy Whitcroft kref_init(&resv_map->refs); 7097b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 71084afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 71184afd99bSAndy Whitcroft 7125e911373SMike Kravetz resv_map->adds_in_progress = 0; 7135e911373SMike Kravetz 7145e911373SMike Kravetz INIT_LIST_HEAD(&resv_map->region_cache); 7155e911373SMike Kravetz list_add(&rg->link, &resv_map->region_cache); 7165e911373SMike Kravetz resv_map->region_cache_count = 1; 7175e911373SMike Kravetz 71884afd99bSAndy Whitcroft return resv_map; 71984afd99bSAndy Whitcroft } 72084afd99bSAndy Whitcroft 7219119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 72284afd99bSAndy Whitcroft { 72384afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 7245e911373SMike Kravetz struct list_head *head = &resv_map->region_cache; 7255e911373SMike Kravetz struct file_region *rg, *trg; 72684afd99bSAndy Whitcroft 72784afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 728feba16e2SMike Kravetz region_del(resv_map, 0, LONG_MAX); 7295e911373SMike Kravetz 7305e911373SMike Kravetz /* ... and any entries left in the cache */ 7315e911373SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 7325e911373SMike Kravetz list_del(&rg->link); 7335e911373SMike Kravetz kfree(rg); 7345e911373SMike Kravetz } 7355e911373SMike Kravetz 7365e911373SMike Kravetz VM_BUG_ON(resv_map->adds_in_progress); 7375e911373SMike Kravetz 73884afd99bSAndy Whitcroft kfree(resv_map); 73984afd99bSAndy Whitcroft } 74084afd99bSAndy Whitcroft 7414e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 7424e35f483SJoonsoo Kim { 7434e35f483SJoonsoo Kim return inode->i_mapping->private_data; 7444e35f483SJoonsoo Kim } 7454e35f483SJoonsoo Kim 74684afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 747a1e78772SMel Gorman { 74881d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 7494e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 7504e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 7514e35f483SJoonsoo Kim struct inode *inode = mapping->host; 7524e35f483SJoonsoo Kim 7534e35f483SJoonsoo Kim return inode_resv_map(inode); 7544e35f483SJoonsoo Kim 7554e35f483SJoonsoo Kim } else { 75684afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 75784afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 7584e35f483SJoonsoo Kim } 759a1e78772SMel Gorman } 760a1e78772SMel Gorman 76184afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 762a1e78772SMel Gorman { 76381d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 76481d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 765a1e78772SMel Gorman 76684afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 76784afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 76804f2cbe3SMel Gorman } 76904f2cbe3SMel Gorman 77004f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 77104f2cbe3SMel Gorman { 77281d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 77381d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 774e7c4b0bfSAndy Whitcroft 775e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 77604f2cbe3SMel Gorman } 77704f2cbe3SMel Gorman 77804f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 77904f2cbe3SMel Gorman { 78081d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 781e7c4b0bfSAndy Whitcroft 782e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 783a1e78772SMel Gorman } 784a1e78772SMel Gorman 78504f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 786a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 787a1e78772SMel Gorman { 78881d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 789f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 790a1e78772SMel Gorman vma->vm_private_data = (void *)0; 791a1e78772SMel Gorman } 792a1e78772SMel Gorman 793a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 794559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 795a1e78772SMel Gorman { 796af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 797af0ed73eSJoonsoo Kim /* 798af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 799af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 800af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 801af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 802af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 803af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 804af0ed73eSJoonsoo Kim * properly, so add work-around here. 805af0ed73eSJoonsoo Kim */ 806af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 807559ec2f8SNicholas Krause return true; 808af0ed73eSJoonsoo Kim else 809559ec2f8SNicholas Krause return false; 810af0ed73eSJoonsoo Kim } 811a63884e9SJoonsoo Kim 812a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 8131fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 8141fb1b0e9SMike Kravetz /* 8151fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 8161fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 8171fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 8181fb1b0e9SMike Kravetz * fallocate. In this case, there really are no reverves to 8191fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 8201fb1b0e9SMike Kravetz */ 8211fb1b0e9SMike Kravetz if (chg) 8221fb1b0e9SMike Kravetz return false; 8231fb1b0e9SMike Kravetz else 824559ec2f8SNicholas Krause return true; 8251fb1b0e9SMike Kravetz } 826a63884e9SJoonsoo Kim 827a63884e9SJoonsoo Kim /* 828a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 829a63884e9SJoonsoo Kim * private mappings. 830a63884e9SJoonsoo Kim */ 8317f09ca51SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 832559ec2f8SNicholas Krause return true; 833a63884e9SJoonsoo Kim 834559ec2f8SNicholas Krause return false; 835a1e78772SMel Gorman } 836a1e78772SMel Gorman 837a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 8381da177e4SLinus Torvalds { 8391da177e4SLinus Torvalds int nid = page_to_nid(page); 8400edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 841a5516438SAndi Kleen h->free_huge_pages++; 842a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 8431da177e4SLinus Torvalds } 8441da177e4SLinus Torvalds 845bf50bab2SNaoya Horiguchi static struct page *dequeue_huge_page_node(struct hstate *h, int nid) 846bf50bab2SNaoya Horiguchi { 847bf50bab2SNaoya Horiguchi struct page *page; 848bf50bab2SNaoya Horiguchi 849c8721bbbSNaoya Horiguchi list_for_each_entry(page, &h->hugepage_freelists[nid], lru) 850c8721bbbSNaoya Horiguchi if (!is_migrate_isolate_page(page)) 851c8721bbbSNaoya Horiguchi break; 852c8721bbbSNaoya Horiguchi /* 853c8721bbbSNaoya Horiguchi * if 'non-isolated free hugepage' not found on the list, 854c8721bbbSNaoya Horiguchi * the allocation fails. 855c8721bbbSNaoya Horiguchi */ 856c8721bbbSNaoya Horiguchi if (&h->hugepage_freelists[nid] == &page->lru) 857bf50bab2SNaoya Horiguchi return NULL; 8580edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 859a9869b83SNaoya Horiguchi set_page_refcounted(page); 860bf50bab2SNaoya Horiguchi h->free_huge_pages--; 861bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 862bf50bab2SNaoya Horiguchi return page; 863bf50bab2SNaoya Horiguchi } 864bf50bab2SNaoya Horiguchi 86586cdb465SNaoya Horiguchi /* Movability of hugepages depends on migration support. */ 86686cdb465SNaoya Horiguchi static inline gfp_t htlb_alloc_mask(struct hstate *h) 86786cdb465SNaoya Horiguchi { 868100873d7SNaoya Horiguchi if (hugepages_treat_as_movable || hugepage_migration_supported(h)) 86986cdb465SNaoya Horiguchi return GFP_HIGHUSER_MOVABLE; 87086cdb465SNaoya Horiguchi else 87186cdb465SNaoya Horiguchi return GFP_HIGHUSER; 87286cdb465SNaoya Horiguchi } 87386cdb465SNaoya Horiguchi 874a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 875a5516438SAndi Kleen struct vm_area_struct *vma, 876af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 877af0ed73eSJoonsoo Kim long chg) 8781da177e4SLinus Torvalds { 879b1c12cbcSKonstantin Khlebnikov struct page *page = NULL; 880480eccf9SLee Schermerhorn struct mempolicy *mpol; 88119770b32SMel Gorman nodemask_t *nodemask; 882c0ff7453SMiao Xie struct zonelist *zonelist; 883dd1a239fSMel Gorman struct zone *zone; 884dd1a239fSMel Gorman struct zoneref *z; 885cc9a6c87SMel Gorman unsigned int cpuset_mems_cookie; 8861da177e4SLinus Torvalds 887a1e78772SMel Gorman /* 888a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 889a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 890a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 891a1e78772SMel Gorman */ 892af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg) && 893a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 894c0ff7453SMiao Xie goto err; 895a1e78772SMel Gorman 89604f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 897a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 8986eab04a8SJustin P. Mattock goto err; 89904f2cbe3SMel Gorman 9009966c4bbSJoonsoo Kim retry_cpuset: 901d26914d1SMel Gorman cpuset_mems_cookie = read_mems_allowed_begin(); 9029966c4bbSJoonsoo Kim zonelist = huge_zonelist(vma, address, 90386cdb465SNaoya Horiguchi htlb_alloc_mask(h), &mpol, &nodemask); 9049966c4bbSJoonsoo Kim 90519770b32SMel Gorman for_each_zone_zonelist_nodemask(zone, z, zonelist, 90619770b32SMel Gorman MAX_NR_ZONES - 1, nodemask) { 907344736f2SVladimir Davydov if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) { 908bf50bab2SNaoya Horiguchi page = dequeue_huge_page_node(h, zone_to_nid(zone)); 909bf50bab2SNaoya Horiguchi if (page) { 910af0ed73eSJoonsoo Kim if (avoid_reserve) 911af0ed73eSJoonsoo Kim break; 912af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg)) 913af0ed73eSJoonsoo Kim break; 914af0ed73eSJoonsoo Kim 91507443a85SJoonsoo Kim SetPagePrivate(page); 916a63884e9SJoonsoo Kim h->resv_huge_pages--; 9175ab3ee7bSKen Chen break; 9181da177e4SLinus Torvalds } 9193abf7afdSAndrew Morton } 920bf50bab2SNaoya Horiguchi } 921cc9a6c87SMel Gorman 922cc9a6c87SMel Gorman mpol_cond_put(mpol); 923d26914d1SMel Gorman if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) 924cc9a6c87SMel Gorman goto retry_cpuset; 925cc9a6c87SMel Gorman return page; 926cc9a6c87SMel Gorman 927c0ff7453SMiao Xie err: 928cc9a6c87SMel Gorman return NULL; 9291da177e4SLinus Torvalds } 9301da177e4SLinus Torvalds 9311cac6f2cSLuiz Capitulino /* 9321cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 9331cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 9341cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 9351cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 9361cac6f2cSLuiz Capitulino * node for alloc or free. 9371cac6f2cSLuiz Capitulino */ 9381cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 9391cac6f2cSLuiz Capitulino { 9400edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 9411cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 9421cac6f2cSLuiz Capitulino 9431cac6f2cSLuiz Capitulino return nid; 9441cac6f2cSLuiz Capitulino } 9451cac6f2cSLuiz Capitulino 9461cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 9471cac6f2cSLuiz Capitulino { 9481cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 9491cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 9501cac6f2cSLuiz Capitulino return nid; 9511cac6f2cSLuiz Capitulino } 9521cac6f2cSLuiz Capitulino 9531cac6f2cSLuiz Capitulino /* 9541cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 9551cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 9561cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 9571cac6f2cSLuiz Capitulino * mask. 9581cac6f2cSLuiz Capitulino */ 9591cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 9601cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 9611cac6f2cSLuiz Capitulino { 9621cac6f2cSLuiz Capitulino int nid; 9631cac6f2cSLuiz Capitulino 9641cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 9651cac6f2cSLuiz Capitulino 9661cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 9671cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 9681cac6f2cSLuiz Capitulino 9691cac6f2cSLuiz Capitulino return nid; 9701cac6f2cSLuiz Capitulino } 9711cac6f2cSLuiz Capitulino 9721cac6f2cSLuiz Capitulino /* 9731cac6f2cSLuiz Capitulino * helper for free_pool_huge_page() - return the previously saved 9741cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 9751cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 9761cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 9771cac6f2cSLuiz Capitulino */ 9781cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 9791cac6f2cSLuiz Capitulino { 9801cac6f2cSLuiz Capitulino int nid; 9811cac6f2cSLuiz Capitulino 9821cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 9831cac6f2cSLuiz Capitulino 9841cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 9851cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 9861cac6f2cSLuiz Capitulino 9871cac6f2cSLuiz Capitulino return nid; 9881cac6f2cSLuiz Capitulino } 9891cac6f2cSLuiz Capitulino 9901cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 9911cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 9921cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 9931cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 9941cac6f2cSLuiz Capitulino nr_nodes--) 9951cac6f2cSLuiz Capitulino 9961cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 9971cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 9981cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 9991cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 10001cac6f2cSLuiz Capitulino nr_nodes--) 10011cac6f2cSLuiz Capitulino 1002080fe206SVlastimil Babka #if defined(CONFIG_X86_64) && ((defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)) 1003944d9fecSLuiz Capitulino static void destroy_compound_gigantic_page(struct page *page, 1004d00181b9SKirill A. Shutemov unsigned int order) 1005944d9fecSLuiz Capitulino { 1006944d9fecSLuiz Capitulino int i; 1007944d9fecSLuiz Capitulino int nr_pages = 1 << order; 1008944d9fecSLuiz Capitulino struct page *p = page + 1; 1009944d9fecSLuiz Capitulino 1010944d9fecSLuiz Capitulino for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 10111d798ca3SKirill A. Shutemov clear_compound_head(p); 1012944d9fecSLuiz Capitulino set_page_refcounted(p); 1013944d9fecSLuiz Capitulino } 1014944d9fecSLuiz Capitulino 1015944d9fecSLuiz Capitulino set_compound_order(page, 0); 1016944d9fecSLuiz Capitulino __ClearPageHead(page); 1017944d9fecSLuiz Capitulino } 1018944d9fecSLuiz Capitulino 1019d00181b9SKirill A. Shutemov static void free_gigantic_page(struct page *page, unsigned int order) 1020944d9fecSLuiz Capitulino { 1021944d9fecSLuiz Capitulino free_contig_range(page_to_pfn(page), 1 << order); 1022944d9fecSLuiz Capitulino } 1023944d9fecSLuiz Capitulino 1024944d9fecSLuiz Capitulino static int __alloc_gigantic_page(unsigned long start_pfn, 1025944d9fecSLuiz Capitulino unsigned long nr_pages) 1026944d9fecSLuiz Capitulino { 1027944d9fecSLuiz Capitulino unsigned long end_pfn = start_pfn + nr_pages; 1028944d9fecSLuiz Capitulino return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE); 1029944d9fecSLuiz Capitulino } 1030944d9fecSLuiz Capitulino 1031944d9fecSLuiz Capitulino static bool pfn_range_valid_gigantic(unsigned long start_pfn, 1032944d9fecSLuiz Capitulino unsigned long nr_pages) 1033944d9fecSLuiz Capitulino { 1034944d9fecSLuiz Capitulino unsigned long i, end_pfn = start_pfn + nr_pages; 1035944d9fecSLuiz Capitulino struct page *page; 1036944d9fecSLuiz Capitulino 1037944d9fecSLuiz Capitulino for (i = start_pfn; i < end_pfn; i++) { 1038944d9fecSLuiz Capitulino if (!pfn_valid(i)) 1039944d9fecSLuiz Capitulino return false; 1040944d9fecSLuiz Capitulino 1041944d9fecSLuiz Capitulino page = pfn_to_page(i); 1042944d9fecSLuiz Capitulino 1043944d9fecSLuiz Capitulino if (PageReserved(page)) 1044944d9fecSLuiz Capitulino return false; 1045944d9fecSLuiz Capitulino 1046944d9fecSLuiz Capitulino if (page_count(page) > 0) 1047944d9fecSLuiz Capitulino return false; 1048944d9fecSLuiz Capitulino 1049944d9fecSLuiz Capitulino if (PageHuge(page)) 1050944d9fecSLuiz Capitulino return false; 1051944d9fecSLuiz Capitulino } 1052944d9fecSLuiz Capitulino 1053944d9fecSLuiz Capitulino return true; 1054944d9fecSLuiz Capitulino } 1055944d9fecSLuiz Capitulino 1056944d9fecSLuiz Capitulino static bool zone_spans_last_pfn(const struct zone *zone, 1057944d9fecSLuiz Capitulino unsigned long start_pfn, unsigned long nr_pages) 1058944d9fecSLuiz Capitulino { 1059944d9fecSLuiz Capitulino unsigned long last_pfn = start_pfn + nr_pages - 1; 1060944d9fecSLuiz Capitulino return zone_spans_pfn(zone, last_pfn); 1061944d9fecSLuiz Capitulino } 1062944d9fecSLuiz Capitulino 1063d00181b9SKirill A. Shutemov static struct page *alloc_gigantic_page(int nid, unsigned int order) 1064944d9fecSLuiz Capitulino { 1065944d9fecSLuiz Capitulino unsigned long nr_pages = 1 << order; 1066944d9fecSLuiz Capitulino unsigned long ret, pfn, flags; 1067944d9fecSLuiz Capitulino struct zone *z; 1068944d9fecSLuiz Capitulino 1069944d9fecSLuiz Capitulino z = NODE_DATA(nid)->node_zones; 1070944d9fecSLuiz Capitulino for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) { 1071944d9fecSLuiz Capitulino spin_lock_irqsave(&z->lock, flags); 1072944d9fecSLuiz Capitulino 1073944d9fecSLuiz Capitulino pfn = ALIGN(z->zone_start_pfn, nr_pages); 1074944d9fecSLuiz Capitulino while (zone_spans_last_pfn(z, pfn, nr_pages)) { 1075944d9fecSLuiz Capitulino if (pfn_range_valid_gigantic(pfn, nr_pages)) { 1076944d9fecSLuiz Capitulino /* 1077944d9fecSLuiz Capitulino * We release the zone lock here because 1078944d9fecSLuiz Capitulino * alloc_contig_range() will also lock the zone 1079944d9fecSLuiz Capitulino * at some point. If there's an allocation 1080944d9fecSLuiz Capitulino * spinning on this lock, it may win the race 1081944d9fecSLuiz Capitulino * and cause alloc_contig_range() to fail... 1082944d9fecSLuiz Capitulino */ 1083944d9fecSLuiz Capitulino spin_unlock_irqrestore(&z->lock, flags); 1084944d9fecSLuiz Capitulino ret = __alloc_gigantic_page(pfn, nr_pages); 1085944d9fecSLuiz Capitulino if (!ret) 1086944d9fecSLuiz Capitulino return pfn_to_page(pfn); 1087944d9fecSLuiz Capitulino spin_lock_irqsave(&z->lock, flags); 1088944d9fecSLuiz Capitulino } 1089944d9fecSLuiz Capitulino pfn += nr_pages; 1090944d9fecSLuiz Capitulino } 1091944d9fecSLuiz Capitulino 1092944d9fecSLuiz Capitulino spin_unlock_irqrestore(&z->lock, flags); 1093944d9fecSLuiz Capitulino } 1094944d9fecSLuiz Capitulino 1095944d9fecSLuiz Capitulino return NULL; 1096944d9fecSLuiz Capitulino } 1097944d9fecSLuiz Capitulino 1098944d9fecSLuiz Capitulino static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); 1099d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order); 1100944d9fecSLuiz Capitulino 1101944d9fecSLuiz Capitulino static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid) 1102944d9fecSLuiz Capitulino { 1103944d9fecSLuiz Capitulino struct page *page; 1104944d9fecSLuiz Capitulino 1105944d9fecSLuiz Capitulino page = alloc_gigantic_page(nid, huge_page_order(h)); 1106944d9fecSLuiz Capitulino if (page) { 1107944d9fecSLuiz Capitulino prep_compound_gigantic_page(page, huge_page_order(h)); 1108944d9fecSLuiz Capitulino prep_new_huge_page(h, page, nid); 1109944d9fecSLuiz Capitulino } 1110944d9fecSLuiz Capitulino 1111944d9fecSLuiz Capitulino return page; 1112944d9fecSLuiz Capitulino } 1113944d9fecSLuiz Capitulino 1114944d9fecSLuiz Capitulino static int alloc_fresh_gigantic_page(struct hstate *h, 1115944d9fecSLuiz Capitulino nodemask_t *nodes_allowed) 1116944d9fecSLuiz Capitulino { 1117944d9fecSLuiz Capitulino struct page *page = NULL; 1118944d9fecSLuiz Capitulino int nr_nodes, node; 1119944d9fecSLuiz Capitulino 1120944d9fecSLuiz Capitulino for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1121944d9fecSLuiz Capitulino page = alloc_fresh_gigantic_page_node(h, node); 1122944d9fecSLuiz Capitulino if (page) 1123944d9fecSLuiz Capitulino return 1; 1124944d9fecSLuiz Capitulino } 1125944d9fecSLuiz Capitulino 1126944d9fecSLuiz Capitulino return 0; 1127944d9fecSLuiz Capitulino } 1128944d9fecSLuiz Capitulino 1129944d9fecSLuiz Capitulino static inline bool gigantic_page_supported(void) { return true; } 1130944d9fecSLuiz Capitulino #else 1131944d9fecSLuiz Capitulino static inline bool gigantic_page_supported(void) { return false; } 1132d00181b9SKirill A. Shutemov static inline void free_gigantic_page(struct page *page, unsigned int order) { } 1133944d9fecSLuiz Capitulino static inline void destroy_compound_gigantic_page(struct page *page, 1134d00181b9SKirill A. Shutemov unsigned int order) { } 1135944d9fecSLuiz Capitulino static inline int alloc_fresh_gigantic_page(struct hstate *h, 1136944d9fecSLuiz Capitulino nodemask_t *nodes_allowed) { return 0; } 1137944d9fecSLuiz Capitulino #endif 1138944d9fecSLuiz Capitulino 1139a5516438SAndi Kleen static void update_and_free_page(struct hstate *h, struct page *page) 11406af2acb6SAdam Litke { 11416af2acb6SAdam Litke int i; 1142a5516438SAndi Kleen 1143944d9fecSLuiz Capitulino if (hstate_is_gigantic(h) && !gigantic_page_supported()) 1144944d9fecSLuiz Capitulino return; 114518229df5SAndy Whitcroft 1146a5516438SAndi Kleen h->nr_huge_pages--; 1147a5516438SAndi Kleen h->nr_huge_pages_node[page_to_nid(page)]--; 1148a5516438SAndi Kleen for (i = 0; i < pages_per_huge_page(h); i++) { 114932f84528SChris Forbes page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 115032f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1151a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1152a7407a27SLuiz Capitulino 1 << PG_writeback); 11536af2acb6SAdam Litke } 1154309381feSSasha Levin VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 1155f1e61557SKirill A. Shutemov set_compound_page_dtor(page, NULL_COMPOUND_DTOR); 11566af2acb6SAdam Litke set_page_refcounted(page); 1157944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) { 1158944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 1159944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 1160944d9fecSLuiz Capitulino } else { 1161a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 11626af2acb6SAdam Litke } 1163944d9fecSLuiz Capitulino } 11646af2acb6SAdam Litke 1165e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1166e5ff2159SAndi Kleen { 1167e5ff2159SAndi Kleen struct hstate *h; 1168e5ff2159SAndi Kleen 1169e5ff2159SAndi Kleen for_each_hstate(h) { 1170e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1171e5ff2159SAndi Kleen return h; 1172e5ff2159SAndi Kleen } 1173e5ff2159SAndi Kleen return NULL; 1174e5ff2159SAndi Kleen } 1175e5ff2159SAndi Kleen 1176bcc54222SNaoya Horiguchi /* 1177bcc54222SNaoya Horiguchi * Test to determine whether the hugepage is "active/in-use" (i.e. being linked 1178bcc54222SNaoya Horiguchi * to hstate->hugepage_activelist.) 1179bcc54222SNaoya Horiguchi * 1180bcc54222SNaoya Horiguchi * This function can be called for tail pages, but never returns true for them. 1181bcc54222SNaoya Horiguchi */ 1182bcc54222SNaoya Horiguchi bool page_huge_active(struct page *page) 1183bcc54222SNaoya Horiguchi { 1184bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHuge(page), page); 1185bcc54222SNaoya Horiguchi return PageHead(page) && PagePrivate(&page[1]); 1186bcc54222SNaoya Horiguchi } 1187bcc54222SNaoya Horiguchi 1188bcc54222SNaoya Horiguchi /* never called for tail page */ 1189bcc54222SNaoya Horiguchi static void set_page_huge_active(struct page *page) 1190bcc54222SNaoya Horiguchi { 1191bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 1192bcc54222SNaoya Horiguchi SetPagePrivate(&page[1]); 1193bcc54222SNaoya Horiguchi } 1194bcc54222SNaoya Horiguchi 1195bcc54222SNaoya Horiguchi static void clear_page_huge_active(struct page *page) 1196bcc54222SNaoya Horiguchi { 1197bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 1198bcc54222SNaoya Horiguchi ClearPagePrivate(&page[1]); 1199bcc54222SNaoya Horiguchi } 1200bcc54222SNaoya Horiguchi 12018f1d26d0SAtsushi Kumagai void free_huge_page(struct page *page) 120227a85ef1SDavid Gibson { 1203a5516438SAndi Kleen /* 1204a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1205a5516438SAndi Kleen * compound page destructor. 1206a5516438SAndi Kleen */ 1207e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 12087893d1d5SAdam Litke int nid = page_to_nid(page); 120990481622SDavid Gibson struct hugepage_subpool *spool = 121090481622SDavid Gibson (struct hugepage_subpool *)page_private(page); 121107443a85SJoonsoo Kim bool restore_reserve; 121227a85ef1SDavid Gibson 1213e5df70abSAndy Whitcroft set_page_private(page, 0); 121423be7468SMel Gorman page->mapping = NULL; 1215b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1216b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_mapcount(page), page); 121707443a85SJoonsoo Kim restore_reserve = PagePrivate(page); 121816c794b4SJoonsoo Kim ClearPagePrivate(page); 121927a85ef1SDavid Gibson 12201c5ecae3SMike Kravetz /* 12211c5ecae3SMike Kravetz * A return code of zero implies that the subpool will be under its 12221c5ecae3SMike Kravetz * minimum size if the reservation is not restored after page is free. 12231c5ecae3SMike Kravetz * Therefore, force restore_reserve operation. 12241c5ecae3SMike Kravetz */ 12251c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 12261c5ecae3SMike Kravetz restore_reserve = true; 12271c5ecae3SMike Kravetz 122827a85ef1SDavid Gibson spin_lock(&hugetlb_lock); 1229bcc54222SNaoya Horiguchi clear_page_huge_active(page); 12306d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 12316d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 123207443a85SJoonsoo Kim if (restore_reserve) 123307443a85SJoonsoo Kim h->resv_huge_pages++; 123407443a85SJoonsoo Kim 1235944d9fecSLuiz Capitulino if (h->surplus_huge_pages_node[nid]) { 12360edaecfaSAneesh Kumar K.V /* remove the page from active list */ 12370edaecfaSAneesh Kumar K.V list_del(&page->lru); 1238a5516438SAndi Kleen update_and_free_page(h, page); 1239a5516438SAndi Kleen h->surplus_huge_pages--; 1240a5516438SAndi Kleen h->surplus_huge_pages_node[nid]--; 12417893d1d5SAdam Litke } else { 12425d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1243a5516438SAndi Kleen enqueue_huge_page(h, page); 12447893d1d5SAdam Litke } 124527a85ef1SDavid Gibson spin_unlock(&hugetlb_lock); 124627a85ef1SDavid Gibson } 124727a85ef1SDavid Gibson 1248a5516438SAndi Kleen static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1249b7ba30c6SAndi Kleen { 12500edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1251f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1252b7ba30c6SAndi Kleen spin_lock(&hugetlb_lock); 12539dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 1254a5516438SAndi Kleen h->nr_huge_pages++; 1255a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 1256b7ba30c6SAndi Kleen spin_unlock(&hugetlb_lock); 1257b7ba30c6SAndi Kleen put_page(page); /* free it into the hugepage allocator */ 1258b7ba30c6SAndi Kleen } 1259b7ba30c6SAndi Kleen 1260d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order) 126120a0307cSWu Fengguang { 126220a0307cSWu Fengguang int i; 126320a0307cSWu Fengguang int nr_pages = 1 << order; 126420a0307cSWu Fengguang struct page *p = page + 1; 126520a0307cSWu Fengguang 126620a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 126720a0307cSWu Fengguang set_compound_order(page, order); 1268ef5a22beSAndrea Arcangeli __ClearPageReserved(page); 1269de09d31dSKirill A. Shutemov __SetPageHead(page); 127020a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1271ef5a22beSAndrea Arcangeli /* 1272ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1273ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1274ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 1275ef5a22beSAndrea Arcangeli * too. Otherwse drivers using get_user_pages() to access tail 1276ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1277ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1278ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1279ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1280ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1281ef5a22beSAndrea Arcangeli * after get_user_pages(). 1282ef5a22beSAndrea Arcangeli */ 1283ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 128458a84aa9SYouquan Song set_page_count(p, 0); 12851d798ca3SKirill A. Shutemov set_compound_head(p, page); 128620a0307cSWu Fengguang } 1287b4330afbSMike Kravetz atomic_set(compound_mapcount_ptr(page), -1); 128820a0307cSWu Fengguang } 128920a0307cSWu Fengguang 12907795912cSAndrew Morton /* 12917795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 12927795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 12937795912cSAndrew Morton * details. 12947795912cSAndrew Morton */ 129520a0307cSWu Fengguang int PageHuge(struct page *page) 129620a0307cSWu Fengguang { 129720a0307cSWu Fengguang if (!PageCompound(page)) 129820a0307cSWu Fengguang return 0; 129920a0307cSWu Fengguang 130020a0307cSWu Fengguang page = compound_head(page); 1301f1e61557SKirill A. Shutemov return page[1].compound_dtor == HUGETLB_PAGE_DTOR; 130220a0307cSWu Fengguang } 130343131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 130443131e14SNaoya Horiguchi 130527c73ae7SAndrea Arcangeli /* 130627c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 130727c73ae7SAndrea Arcangeli * normal or transparent huge pages. 130827c73ae7SAndrea Arcangeli */ 130927c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 131027c73ae7SAndrea Arcangeli { 131127c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 131227c73ae7SAndrea Arcangeli return 0; 131327c73ae7SAndrea Arcangeli 1314758f66a2SAndrew Morton return get_compound_page_dtor(page_head) == free_huge_page; 131527c73ae7SAndrea Arcangeli } 131627c73ae7SAndrea Arcangeli 131713d60f4bSZhang Yi pgoff_t __basepage_index(struct page *page) 131813d60f4bSZhang Yi { 131913d60f4bSZhang Yi struct page *page_head = compound_head(page); 132013d60f4bSZhang Yi pgoff_t index = page_index(page_head); 132113d60f4bSZhang Yi unsigned long compound_idx; 132213d60f4bSZhang Yi 132313d60f4bSZhang Yi if (!PageHuge(page_head)) 132413d60f4bSZhang Yi return page_index(page); 132513d60f4bSZhang Yi 132613d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 132713d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 132813d60f4bSZhang Yi else 132913d60f4bSZhang Yi compound_idx = page - page_head; 133013d60f4bSZhang Yi 133113d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 133213d60f4bSZhang Yi } 133313d60f4bSZhang Yi 1334a5516438SAndi Kleen static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) 13351da177e4SLinus Torvalds { 13361da177e4SLinus Torvalds struct page *page; 1337f96efd58SJoe Jin 133896db800fSVlastimil Babka page = __alloc_pages_node(nid, 133986cdb465SNaoya Horiguchi htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| 1340551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 1341a5516438SAndi Kleen huge_page_order(h)); 13421da177e4SLinus Torvalds if (page) { 1343a5516438SAndi Kleen prep_new_huge_page(h, page, nid); 13441da177e4SLinus Torvalds } 134563b4613cSNishanth Aravamudan 134663b4613cSNishanth Aravamudan return page; 134763b4613cSNishanth Aravamudan } 134863b4613cSNishanth Aravamudan 1349b2261026SJoonsoo Kim static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 1350b2261026SJoonsoo Kim { 1351b2261026SJoonsoo Kim struct page *page; 1352b2261026SJoonsoo Kim int nr_nodes, node; 1353b2261026SJoonsoo Kim int ret = 0; 1354b2261026SJoonsoo Kim 1355b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1356b2261026SJoonsoo Kim page = alloc_fresh_huge_page_node(h, node); 1357b2261026SJoonsoo Kim if (page) { 1358b2261026SJoonsoo Kim ret = 1; 1359b2261026SJoonsoo Kim break; 1360b2261026SJoonsoo Kim } 1361b2261026SJoonsoo Kim } 1362b2261026SJoonsoo Kim 1363b2261026SJoonsoo Kim if (ret) 1364b2261026SJoonsoo Kim count_vm_event(HTLB_BUDDY_PGALLOC); 1365b2261026SJoonsoo Kim else 1366b2261026SJoonsoo Kim count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 1367b2261026SJoonsoo Kim 1368b2261026SJoonsoo Kim return ret; 1369b2261026SJoonsoo Kim } 1370b2261026SJoonsoo Kim 1371e8c5c824SLee Schermerhorn /* 1372e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 1373e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 1374e8c5c824SLee Schermerhorn * balanced over allowed nodes. 1375e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 1376e8c5c824SLee Schermerhorn */ 13776ae11b27SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 13786ae11b27SLee Schermerhorn bool acct_surplus) 1379e8c5c824SLee Schermerhorn { 1380b2261026SJoonsoo Kim int nr_nodes, node; 1381e8c5c824SLee Schermerhorn int ret = 0; 1382e8c5c824SLee Schermerhorn 1383b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1384685f3457SLee Schermerhorn /* 1385685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 1386685f3457SLee Schermerhorn * nodes with surplus pages. 1387685f3457SLee Schermerhorn */ 1388b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 1389b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 1390e8c5c824SLee Schermerhorn struct page *page = 1391b2261026SJoonsoo Kim list_entry(h->hugepage_freelists[node].next, 1392e8c5c824SLee Schermerhorn struct page, lru); 1393e8c5c824SLee Schermerhorn list_del(&page->lru); 1394e8c5c824SLee Schermerhorn h->free_huge_pages--; 1395b2261026SJoonsoo Kim h->free_huge_pages_node[node]--; 1396685f3457SLee Schermerhorn if (acct_surplus) { 1397685f3457SLee Schermerhorn h->surplus_huge_pages--; 1398b2261026SJoonsoo Kim h->surplus_huge_pages_node[node]--; 1399685f3457SLee Schermerhorn } 1400e8c5c824SLee Schermerhorn update_and_free_page(h, page); 1401e8c5c824SLee Schermerhorn ret = 1; 14029a76db09SLee Schermerhorn break; 1403e8c5c824SLee Schermerhorn } 1404b2261026SJoonsoo Kim } 1405e8c5c824SLee Schermerhorn 1406e8c5c824SLee Schermerhorn return ret; 1407e8c5c824SLee Schermerhorn } 1408e8c5c824SLee Schermerhorn 1409c8721bbbSNaoya Horiguchi /* 1410c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 1411c8721bbbSNaoya Horiguchi * nothing for in-use (including surplus) hugepages. 1412c8721bbbSNaoya Horiguchi */ 1413c8721bbbSNaoya Horiguchi static void dissolve_free_huge_page(struct page *page) 1414c8721bbbSNaoya Horiguchi { 1415c8721bbbSNaoya Horiguchi spin_lock(&hugetlb_lock); 1416c8721bbbSNaoya Horiguchi if (PageHuge(page) && !page_count(page)) { 1417c8721bbbSNaoya Horiguchi struct hstate *h = page_hstate(page); 1418c8721bbbSNaoya Horiguchi int nid = page_to_nid(page); 1419c8721bbbSNaoya Horiguchi list_del(&page->lru); 1420c8721bbbSNaoya Horiguchi h->free_huge_pages--; 1421c8721bbbSNaoya Horiguchi h->free_huge_pages_node[nid]--; 1422c8721bbbSNaoya Horiguchi update_and_free_page(h, page); 1423c8721bbbSNaoya Horiguchi } 1424c8721bbbSNaoya Horiguchi spin_unlock(&hugetlb_lock); 1425c8721bbbSNaoya Horiguchi } 1426c8721bbbSNaoya Horiguchi 1427c8721bbbSNaoya Horiguchi /* 1428c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 1429c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 1430c8721bbbSNaoya Horiguchi * Note that start_pfn should aligned with (minimum) hugepage size. 1431c8721bbbSNaoya Horiguchi */ 1432c8721bbbSNaoya Horiguchi void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 1433c8721bbbSNaoya Horiguchi { 1434c8721bbbSNaoya Horiguchi unsigned long pfn; 1435c8721bbbSNaoya Horiguchi 1436d0177639SLi Zhong if (!hugepages_supported()) 1437d0177639SLi Zhong return; 1438d0177639SLi Zhong 1439641844f5SNaoya Horiguchi VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << minimum_order)); 1440641844f5SNaoya Horiguchi for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) 1441c8721bbbSNaoya Horiguchi dissolve_free_huge_page(pfn_to_page(pfn)); 1442c8721bbbSNaoya Horiguchi } 1443c8721bbbSNaoya Horiguchi 1444099730d6SDave Hansen /* 1445099730d6SDave Hansen * There are 3 ways this can get called: 1446099730d6SDave Hansen * 1. With vma+addr: we use the VMA's memory policy 1447099730d6SDave Hansen * 2. With !vma, but nid=NUMA_NO_NODE: We try to allocate a huge 1448099730d6SDave Hansen * page from any node, and let the buddy allocator itself figure 1449099730d6SDave Hansen * it out. 1450099730d6SDave Hansen * 3. With !vma, but nid!=NUMA_NO_NODE. We allocate a huge page 1451099730d6SDave Hansen * strictly from 'nid' 1452099730d6SDave Hansen */ 1453099730d6SDave Hansen static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h, 1454099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr, int nid) 1455099730d6SDave Hansen { 1456099730d6SDave Hansen int order = huge_page_order(h); 1457099730d6SDave Hansen gfp_t gfp = htlb_alloc_mask(h)|__GFP_COMP|__GFP_REPEAT|__GFP_NOWARN; 1458099730d6SDave Hansen unsigned int cpuset_mems_cookie; 1459099730d6SDave Hansen 1460099730d6SDave Hansen /* 1461099730d6SDave Hansen * We need a VMA to get a memory policy. If we do not 1462e0ec90eeSDave Hansen * have one, we use the 'nid' argument. 1463e0ec90eeSDave Hansen * 1464e0ec90eeSDave Hansen * The mempolicy stuff below has some non-inlined bits 1465e0ec90eeSDave Hansen * and calls ->vm_ops. That makes it hard to optimize at 1466e0ec90eeSDave Hansen * compile-time, even when NUMA is off and it does 1467e0ec90eeSDave Hansen * nothing. This helps the compiler optimize it out. 1468099730d6SDave Hansen */ 1469e0ec90eeSDave Hansen if (!IS_ENABLED(CONFIG_NUMA) || !vma) { 1470099730d6SDave Hansen /* 1471099730d6SDave Hansen * If a specific node is requested, make sure to 1472099730d6SDave Hansen * get memory from there, but only when a node 1473099730d6SDave Hansen * is explicitly specified. 1474099730d6SDave Hansen */ 1475099730d6SDave Hansen if (nid != NUMA_NO_NODE) 1476099730d6SDave Hansen gfp |= __GFP_THISNODE; 1477099730d6SDave Hansen /* 1478099730d6SDave Hansen * Make sure to call something that can handle 1479099730d6SDave Hansen * nid=NUMA_NO_NODE 1480099730d6SDave Hansen */ 1481099730d6SDave Hansen return alloc_pages_node(nid, gfp, order); 1482099730d6SDave Hansen } 1483099730d6SDave Hansen 1484099730d6SDave Hansen /* 1485099730d6SDave Hansen * OK, so we have a VMA. Fetch the mempolicy and try to 1486e0ec90eeSDave Hansen * allocate a huge page with it. We will only reach this 1487e0ec90eeSDave Hansen * when CONFIG_NUMA=y. 1488099730d6SDave Hansen */ 1489099730d6SDave Hansen do { 1490099730d6SDave Hansen struct page *page; 1491099730d6SDave Hansen struct mempolicy *mpol; 1492099730d6SDave Hansen struct zonelist *zl; 1493099730d6SDave Hansen nodemask_t *nodemask; 1494099730d6SDave Hansen 1495099730d6SDave Hansen cpuset_mems_cookie = read_mems_allowed_begin(); 1496099730d6SDave Hansen zl = huge_zonelist(vma, addr, gfp, &mpol, &nodemask); 1497099730d6SDave Hansen mpol_cond_put(mpol); 1498099730d6SDave Hansen page = __alloc_pages_nodemask(gfp, order, zl, nodemask); 1499099730d6SDave Hansen if (page) 1500099730d6SDave Hansen return page; 1501099730d6SDave Hansen } while (read_mems_allowed_retry(cpuset_mems_cookie)); 1502099730d6SDave Hansen 1503099730d6SDave Hansen return NULL; 1504099730d6SDave Hansen } 1505099730d6SDave Hansen 1506099730d6SDave Hansen /* 1507099730d6SDave Hansen * There are two ways to allocate a huge page: 1508099730d6SDave Hansen * 1. When you have a VMA and an address (like a fault) 1509099730d6SDave Hansen * 2. When you have no VMA (like when setting /proc/.../nr_hugepages) 1510099730d6SDave Hansen * 1511099730d6SDave Hansen * 'vma' and 'addr' are only for (1). 'nid' is always NUMA_NO_NODE in 1512099730d6SDave Hansen * this case which signifies that the allocation should be done with 1513099730d6SDave Hansen * respect for the VMA's memory policy. 1514099730d6SDave Hansen * 1515099730d6SDave Hansen * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This 1516099730d6SDave Hansen * implies that memory policies will not be taken in to account. 1517099730d6SDave Hansen */ 1518099730d6SDave Hansen static struct page *__alloc_buddy_huge_page(struct hstate *h, 1519099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr, int nid) 15207893d1d5SAdam Litke { 15217893d1d5SAdam Litke struct page *page; 1522bf50bab2SNaoya Horiguchi unsigned int r_nid; 15237893d1d5SAdam Litke 1524bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1525aa888a74SAndi Kleen return NULL; 1526aa888a74SAndi Kleen 1527d1c3fb1fSNishanth Aravamudan /* 1528099730d6SDave Hansen * Make sure that anyone specifying 'nid' is not also specifying a VMA. 1529099730d6SDave Hansen * This makes sure the caller is picking _one_ of the modes with which 1530099730d6SDave Hansen * we can call this function, not both. 1531099730d6SDave Hansen */ 1532099730d6SDave Hansen if (vma || (addr != -1)) { 1533e0ec90eeSDave Hansen VM_WARN_ON_ONCE(addr == -1); 1534e0ec90eeSDave Hansen VM_WARN_ON_ONCE(nid != NUMA_NO_NODE); 1535099730d6SDave Hansen } 1536099730d6SDave Hansen /* 1537d1c3fb1fSNishanth Aravamudan * Assume we will successfully allocate the surplus page to 1538d1c3fb1fSNishanth Aravamudan * prevent racing processes from causing the surplus to exceed 1539d1c3fb1fSNishanth Aravamudan * overcommit 1540d1c3fb1fSNishanth Aravamudan * 1541d1c3fb1fSNishanth Aravamudan * This however introduces a different race, where a process B 1542d1c3fb1fSNishanth Aravamudan * tries to grow the static hugepage pool while alloc_pages() is 1543d1c3fb1fSNishanth Aravamudan * called by process A. B will only examine the per-node 1544d1c3fb1fSNishanth Aravamudan * counters in determining if surplus huge pages can be 1545d1c3fb1fSNishanth Aravamudan * converted to normal huge pages in adjust_pool_surplus(). A 1546d1c3fb1fSNishanth Aravamudan * won't be able to increment the per-node counter, until the 1547d1c3fb1fSNishanth Aravamudan * lock is dropped by B, but B doesn't drop hugetlb_lock until 1548d1c3fb1fSNishanth Aravamudan * no more huge pages can be converted from surplus to normal 1549d1c3fb1fSNishanth Aravamudan * state (and doesn't try to convert again). Thus, we have a 1550d1c3fb1fSNishanth Aravamudan * case where a surplus huge page exists, the pool is grown, and 1551d1c3fb1fSNishanth Aravamudan * the surplus huge page still exists after, even though it 1552d1c3fb1fSNishanth Aravamudan * should just have been converted to a normal huge page. This 1553d1c3fb1fSNishanth Aravamudan * does not leak memory, though, as the hugepage will be freed 1554d1c3fb1fSNishanth Aravamudan * once it is out of use. It also does not allow the counters to 1555d1c3fb1fSNishanth Aravamudan * go out of whack in adjust_pool_surplus() as we don't modify 1556d1c3fb1fSNishanth Aravamudan * the node values until we've gotten the hugepage and only the 1557d1c3fb1fSNishanth Aravamudan * per-node value is checked there. 1558d1c3fb1fSNishanth Aravamudan */ 1559d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 1560a5516438SAndi Kleen if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 1561d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1562d1c3fb1fSNishanth Aravamudan return NULL; 1563d1c3fb1fSNishanth Aravamudan } else { 1564a5516438SAndi Kleen h->nr_huge_pages++; 1565a5516438SAndi Kleen h->surplus_huge_pages++; 1566d1c3fb1fSNishanth Aravamudan } 1567d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1568d1c3fb1fSNishanth Aravamudan 1569099730d6SDave Hansen page = __hugetlb_alloc_buddy_huge_page(h, vma, addr, nid); 1570d1c3fb1fSNishanth Aravamudan 15717893d1d5SAdam Litke spin_lock(&hugetlb_lock); 1572d1c3fb1fSNishanth Aravamudan if (page) { 15730edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1574bf50bab2SNaoya Horiguchi r_nid = page_to_nid(page); 1575f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 15769dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 1577d1c3fb1fSNishanth Aravamudan /* 1578d1c3fb1fSNishanth Aravamudan * We incremented the global counters already 1579d1c3fb1fSNishanth Aravamudan */ 1580bf50bab2SNaoya Horiguchi h->nr_huge_pages_node[r_nid]++; 1581bf50bab2SNaoya Horiguchi h->surplus_huge_pages_node[r_nid]++; 15823b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC); 1583d1c3fb1fSNishanth Aravamudan } else { 1584a5516438SAndi Kleen h->nr_huge_pages--; 1585a5516438SAndi Kleen h->surplus_huge_pages--; 15863b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 15877893d1d5SAdam Litke } 1588d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 15897893d1d5SAdam Litke 15907893d1d5SAdam Litke return page; 15917893d1d5SAdam Litke } 15927893d1d5SAdam Litke 1593e4e574b7SAdam Litke /* 1594099730d6SDave Hansen * Allocate a huge page from 'nid'. Note, 'nid' may be 1595099730d6SDave Hansen * NUMA_NO_NODE, which means that it may be allocated 1596099730d6SDave Hansen * anywhere. 1597099730d6SDave Hansen */ 1598e0ec90eeSDave Hansen static 1599099730d6SDave Hansen struct page *__alloc_buddy_huge_page_no_mpol(struct hstate *h, int nid) 1600099730d6SDave Hansen { 1601099730d6SDave Hansen unsigned long addr = -1; 1602099730d6SDave Hansen 1603099730d6SDave Hansen return __alloc_buddy_huge_page(h, NULL, addr, nid); 1604099730d6SDave Hansen } 1605099730d6SDave Hansen 1606099730d6SDave Hansen /* 1607099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 1608099730d6SDave Hansen */ 1609e0ec90eeSDave Hansen static 1610099730d6SDave Hansen struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h, 1611099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 1612099730d6SDave Hansen { 1613099730d6SDave Hansen return __alloc_buddy_huge_page(h, vma, addr, NUMA_NO_NODE); 1614099730d6SDave Hansen } 1615099730d6SDave Hansen 1616099730d6SDave Hansen /* 1617bf50bab2SNaoya Horiguchi * This allocation function is useful in the context where vma is irrelevant. 1618bf50bab2SNaoya Horiguchi * E.g. soft-offlining uses this function because it only cares physical 1619bf50bab2SNaoya Horiguchi * address of error page. 1620bf50bab2SNaoya Horiguchi */ 1621bf50bab2SNaoya Horiguchi struct page *alloc_huge_page_node(struct hstate *h, int nid) 1622bf50bab2SNaoya Horiguchi { 16234ef91848SJoonsoo Kim struct page *page = NULL; 1624bf50bab2SNaoya Horiguchi 1625bf50bab2SNaoya Horiguchi spin_lock(&hugetlb_lock); 16264ef91848SJoonsoo Kim if (h->free_huge_pages - h->resv_huge_pages > 0) 1627bf50bab2SNaoya Horiguchi page = dequeue_huge_page_node(h, nid); 1628bf50bab2SNaoya Horiguchi spin_unlock(&hugetlb_lock); 1629bf50bab2SNaoya Horiguchi 163094ae8ba7SAneesh Kumar K.V if (!page) 1631099730d6SDave Hansen page = __alloc_buddy_huge_page_no_mpol(h, nid); 1632bf50bab2SNaoya Horiguchi 1633bf50bab2SNaoya Horiguchi return page; 1634bf50bab2SNaoya Horiguchi } 1635bf50bab2SNaoya Horiguchi 1636bf50bab2SNaoya Horiguchi /* 163725985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 1638e4e574b7SAdam Litke * of size 'delta'. 1639e4e574b7SAdam Litke */ 1640a5516438SAndi Kleen static int gather_surplus_pages(struct hstate *h, int delta) 1641e4e574b7SAdam Litke { 1642e4e574b7SAdam Litke struct list_head surplus_list; 1643e4e574b7SAdam Litke struct page *page, *tmp; 1644e4e574b7SAdam Litke int ret, i; 1645e4e574b7SAdam Litke int needed, allocated; 164628073b02SHillf Danton bool alloc_ok = true; 1647e4e574b7SAdam Litke 1648a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 1649ac09b3a1SAdam Litke if (needed <= 0) { 1650a5516438SAndi Kleen h->resv_huge_pages += delta; 1651e4e574b7SAdam Litke return 0; 1652ac09b3a1SAdam Litke } 1653e4e574b7SAdam Litke 1654e4e574b7SAdam Litke allocated = 0; 1655e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 1656e4e574b7SAdam Litke 1657e4e574b7SAdam Litke ret = -ENOMEM; 1658e4e574b7SAdam Litke retry: 1659e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 1660e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 1661099730d6SDave Hansen page = __alloc_buddy_huge_page_no_mpol(h, NUMA_NO_NODE); 166228073b02SHillf Danton if (!page) { 166328073b02SHillf Danton alloc_ok = false; 166428073b02SHillf Danton break; 166528073b02SHillf Danton } 1666e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 1667e4e574b7SAdam Litke } 166828073b02SHillf Danton allocated += i; 1669e4e574b7SAdam Litke 1670e4e574b7SAdam Litke /* 1671e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 1672e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 1673e4e574b7SAdam Litke */ 1674e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 1675a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 1676a5516438SAndi Kleen (h->free_huge_pages + allocated); 167728073b02SHillf Danton if (needed > 0) { 167828073b02SHillf Danton if (alloc_ok) 1679e4e574b7SAdam Litke goto retry; 168028073b02SHillf Danton /* 168128073b02SHillf Danton * We were not able to allocate enough pages to 168228073b02SHillf Danton * satisfy the entire reservation so we free what 168328073b02SHillf Danton * we've allocated so far. 168428073b02SHillf Danton */ 168528073b02SHillf Danton goto free; 168628073b02SHillf Danton } 1687e4e574b7SAdam Litke /* 1688e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 168925985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 1690e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 1691ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 1692ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 1693ac09b3a1SAdam Litke * before they are reserved. 1694e4e574b7SAdam Litke */ 1695e4e574b7SAdam Litke needed += allocated; 1696a5516438SAndi Kleen h->resv_huge_pages += delta; 1697e4e574b7SAdam Litke ret = 0; 1698a9869b83SNaoya Horiguchi 169919fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 170019fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 170119fc3f0aSAdam Litke if ((--needed) < 0) 170219fc3f0aSAdam Litke break; 1703a9869b83SNaoya Horiguchi /* 1704a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 1705a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 1706a9869b83SNaoya Horiguchi */ 1707a9869b83SNaoya Horiguchi put_page_testzero(page); 1708309381feSSasha Levin VM_BUG_ON_PAGE(page_count(page), page); 1709a5516438SAndi Kleen enqueue_huge_page(h, page); 171019fc3f0aSAdam Litke } 171128073b02SHillf Danton free: 1712b0365c8dSHillf Danton spin_unlock(&hugetlb_lock); 171319fc3f0aSAdam Litke 171419fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 1715c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 1716a9869b83SNaoya Horiguchi put_page(page); 171719fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 1718e4e574b7SAdam Litke 1719e4e574b7SAdam Litke return ret; 1720e4e574b7SAdam Litke } 1721e4e574b7SAdam Litke 1722e4e574b7SAdam Litke /* 1723e4e574b7SAdam Litke * When releasing a hugetlb pool reservation, any surplus pages that were 1724e4e574b7SAdam Litke * allocated to satisfy the reservation must be explicitly freed if they were 1725e4e574b7SAdam Litke * never used. 1726685f3457SLee Schermerhorn * Called with hugetlb_lock held. 1727e4e574b7SAdam Litke */ 1728a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 1729a5516438SAndi Kleen unsigned long unused_resv_pages) 1730e4e574b7SAdam Litke { 1731e4e574b7SAdam Litke unsigned long nr_pages; 1732e4e574b7SAdam Litke 1733ac09b3a1SAdam Litke /* Uncommit the reservation */ 1734a5516438SAndi Kleen h->resv_huge_pages -= unused_resv_pages; 1735ac09b3a1SAdam Litke 1736aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 1737bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1738aa888a74SAndi Kleen return; 1739aa888a74SAndi Kleen 1740a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 1741e4e574b7SAdam Litke 1742685f3457SLee Schermerhorn /* 1743685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 17449b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 17459b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 17469b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 17479b5e5d0fSLee Schermerhorn * free_pool_huge_page() will balance the the freed pages across the 17489b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 1749685f3457SLee Schermerhorn */ 1750685f3457SLee Schermerhorn while (nr_pages--) { 17518cebfcd0SLai Jiangshan if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) 1752685f3457SLee Schermerhorn break; 17537848a4bfSMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 1754e4e574b7SAdam Litke } 1755e4e574b7SAdam Litke } 1756e4e574b7SAdam Litke 17575e911373SMike Kravetz 1758c37f9fb1SAndy Whitcroft /* 1759feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 17605e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 1761cf3ad20bSMike Kravetz * 1762cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 1763cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 1764cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 1765cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 1766cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 1767feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 1768feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 1769feba16e2SMike Kravetz * is called in such cases. 1770cf3ad20bSMike Kravetz * 1771cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 1772cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 1773cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 1774cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 1775cf3ad20bSMike Kravetz * take appropriate action. 1776c37f9fb1SAndy Whitcroft */ 17775e911373SMike Kravetz enum vma_resv_mode { 17785e911373SMike Kravetz VMA_NEEDS_RESV, 17795e911373SMike Kravetz VMA_COMMIT_RESV, 1780feba16e2SMike Kravetz VMA_END_RESV, 17815e911373SMike Kravetz }; 1782cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 1783cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 17845e911373SMike Kravetz enum vma_resv_mode mode) 1785c37f9fb1SAndy Whitcroft { 17864e35f483SJoonsoo Kim struct resv_map *resv; 17874e35f483SJoonsoo Kim pgoff_t idx; 1788cf3ad20bSMike Kravetz long ret; 1789c37f9fb1SAndy Whitcroft 17904e35f483SJoonsoo Kim resv = vma_resv_map(vma); 17914e35f483SJoonsoo Kim if (!resv) 1792c37f9fb1SAndy Whitcroft return 1; 1793c37f9fb1SAndy Whitcroft 17944e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 17955e911373SMike Kravetz switch (mode) { 17965e911373SMike Kravetz case VMA_NEEDS_RESV: 1797cf3ad20bSMike Kravetz ret = region_chg(resv, idx, idx + 1); 17985e911373SMike Kravetz break; 17995e911373SMike Kravetz case VMA_COMMIT_RESV: 18005e911373SMike Kravetz ret = region_add(resv, idx, idx + 1); 18015e911373SMike Kravetz break; 1802feba16e2SMike Kravetz case VMA_END_RESV: 18035e911373SMike Kravetz region_abort(resv, idx, idx + 1); 18045e911373SMike Kravetz ret = 0; 18055e911373SMike Kravetz break; 18065e911373SMike Kravetz default: 18075e911373SMike Kravetz BUG(); 18085e911373SMike Kravetz } 180984afd99bSAndy Whitcroft 18104e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) 1811cf3ad20bSMike Kravetz return ret; 18124e35f483SJoonsoo Kim else 1813cf3ad20bSMike Kravetz return ret < 0 ? ret : 0; 181484afd99bSAndy Whitcroft } 1815cf3ad20bSMike Kravetz 1816cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 1817a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 1818c37f9fb1SAndy Whitcroft { 18195e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 1820cf3ad20bSMike Kravetz } 1821c37f9fb1SAndy Whitcroft 1822cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 1823cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 1824cf3ad20bSMike Kravetz { 18255e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 18265e911373SMike Kravetz } 18275e911373SMike Kravetz 1828feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 18295e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 18305e911373SMike Kravetz { 1831feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 1832c37f9fb1SAndy Whitcroft } 1833c37f9fb1SAndy Whitcroft 183470c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 183504f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 1836348ea204SAdam Litke { 183790481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 1838a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 1839348ea204SAdam Litke struct page *page; 1840d85f69b0SMike Kravetz long map_chg, map_commit; 1841d85f69b0SMike Kravetz long gbl_chg; 18426d76dcf4SAneesh Kumar K.V int ret, idx; 18436d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 18442fc39cecSAdam Litke 18456d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 1846a1e78772SMel Gorman /* 1847d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 1848d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 1849d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 1850a1e78772SMel Gorman */ 1851d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 1852d85f69b0SMike Kravetz if (map_chg < 0) 185376dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 1854d85f69b0SMike Kravetz 1855d85f69b0SMike Kravetz /* 1856d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 1857d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 1858d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 1859d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 1860d85f69b0SMike Kravetz * checked against any subpool limit. 1861d85f69b0SMike Kravetz */ 1862d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 1863d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 1864d85f69b0SMike Kravetz if (gbl_chg < 0) { 1865feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 186676dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 18675e911373SMike Kravetz } 186890d8b7e6SAdam Litke 1869d85f69b0SMike Kravetz /* 1870d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 1871d85f69b0SMike Kravetz * map, there could be reservations associated with the 1872d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 1873d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 1874d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 1875d85f69b0SMike Kravetz * the subpool reservations. 1876d85f69b0SMike Kravetz */ 1877d85f69b0SMike Kravetz if (avoid_reserve) 1878d85f69b0SMike Kravetz gbl_chg = 1; 1879d85f69b0SMike Kravetz } 1880d85f69b0SMike Kravetz 18816d76dcf4SAneesh Kumar K.V ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 18828f34af6fSJianyu Zhan if (ret) 18838f34af6fSJianyu Zhan goto out_subpool_put; 18848f34af6fSJianyu Zhan 1885a1e78772SMel Gorman spin_lock(&hugetlb_lock); 1886d85f69b0SMike Kravetz /* 1887d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 1888d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 1889d85f69b0SMike Kravetz * a reservation exists for the allocation. 1890d85f69b0SMike Kravetz */ 1891d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 189281a6fcaeSJoonsoo Kim if (!page) { 189394ae8ba7SAneesh Kumar K.V spin_unlock(&hugetlb_lock); 1894099730d6SDave Hansen page = __alloc_buddy_huge_page_with_mpol(h, vma, addr); 18958f34af6fSJianyu Zhan if (!page) 18968f34af6fSJianyu Zhan goto out_uncharge_cgroup; 1897a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 1898a88c7695SNaoya Horiguchi SetPagePrivate(page); 1899a88c7695SNaoya Horiguchi h->resv_huge_pages--; 1900a88c7695SNaoya Horiguchi } 190179dbb236SAneesh Kumar K.V spin_lock(&hugetlb_lock); 190279dbb236SAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 190381a6fcaeSJoonsoo Kim /* Fall through */ 1904a1e78772SMel Gorman } 190581a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 190681a6fcaeSJoonsoo Kim spin_unlock(&hugetlb_lock); 1907a1e78772SMel Gorman 190890481622SDavid Gibson set_page_private(page, (unsigned long)spool); 1909a1e78772SMel Gorman 1910d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 1911d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 191233039678SMike Kravetz /* 191333039678SMike Kravetz * The page was added to the reservation map between 191433039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 191533039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 191633039678SMike Kravetz * Adjust for the subpool count incremented above AND 191733039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 191833039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 191933039678SMike Kravetz * no longer applies. 192033039678SMike Kravetz */ 192133039678SMike Kravetz long rsv_adjust; 192233039678SMike Kravetz 192333039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 192433039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 192533039678SMike Kravetz } 19267893d1d5SAdam Litke return page; 19278f34af6fSJianyu Zhan 19288f34af6fSJianyu Zhan out_uncharge_cgroup: 19298f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 19308f34af6fSJianyu Zhan out_subpool_put: 1931d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 19328f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 1933feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 19348f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 1935b45b5bd6SDavid Gibson } 1936b45b5bd6SDavid Gibson 193774060e4dSNaoya Horiguchi /* 193874060e4dSNaoya Horiguchi * alloc_huge_page()'s wrapper which simply returns the page if allocation 193974060e4dSNaoya Horiguchi * succeeds, otherwise NULL. This function is called from new_vma_page(), 194074060e4dSNaoya Horiguchi * where no ERR_VALUE is expected to be returned. 194174060e4dSNaoya Horiguchi */ 194274060e4dSNaoya Horiguchi struct page *alloc_huge_page_noerr(struct vm_area_struct *vma, 194374060e4dSNaoya Horiguchi unsigned long addr, int avoid_reserve) 194474060e4dSNaoya Horiguchi { 194574060e4dSNaoya Horiguchi struct page *page = alloc_huge_page(vma, addr, avoid_reserve); 194674060e4dSNaoya Horiguchi if (IS_ERR(page)) 194774060e4dSNaoya Horiguchi page = NULL; 194874060e4dSNaoya Horiguchi return page; 194974060e4dSNaoya Horiguchi } 195074060e4dSNaoya Horiguchi 195191f47662SCyrill Gorcunov int __weak alloc_bootmem_huge_page(struct hstate *h) 1952aa888a74SAndi Kleen { 1953aa888a74SAndi Kleen struct huge_bootmem_page *m; 1954b2261026SJoonsoo Kim int nr_nodes, node; 1955aa888a74SAndi Kleen 1956b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 1957aa888a74SAndi Kleen void *addr; 1958aa888a74SAndi Kleen 19598b89a116SGrygorii Strashko addr = memblock_virt_alloc_try_nid_nopanic( 19608b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 19618b89a116SGrygorii Strashko 0, BOOTMEM_ALLOC_ACCESSIBLE, node); 1962aa888a74SAndi Kleen if (addr) { 1963aa888a74SAndi Kleen /* 1964aa888a74SAndi Kleen * Use the beginning of the huge page to store the 1965aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 1966aa888a74SAndi Kleen * puts them into the mem_map). 1967aa888a74SAndi Kleen */ 1968aa888a74SAndi Kleen m = addr; 1969aa888a74SAndi Kleen goto found; 1970aa888a74SAndi Kleen } 1971aa888a74SAndi Kleen } 1972aa888a74SAndi Kleen return 0; 1973aa888a74SAndi Kleen 1974aa888a74SAndi Kleen found: 1975df994eadSLuiz Capitulino BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); 1976aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 1977aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 1978aa888a74SAndi Kleen m->hstate = h; 1979aa888a74SAndi Kleen return 1; 1980aa888a74SAndi Kleen } 1981aa888a74SAndi Kleen 1982d00181b9SKirill A. Shutemov static void __init prep_compound_huge_page(struct page *page, 1983d00181b9SKirill A. Shutemov unsigned int order) 198418229df5SAndy Whitcroft { 198518229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 198618229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 198718229df5SAndy Whitcroft else 198818229df5SAndy Whitcroft prep_compound_page(page, order); 198918229df5SAndy Whitcroft } 199018229df5SAndy Whitcroft 1991aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 1992aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 1993aa888a74SAndi Kleen { 1994aa888a74SAndi Kleen struct huge_bootmem_page *m; 1995aa888a74SAndi Kleen 1996aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 1997aa888a74SAndi Kleen struct hstate *h = m->hstate; 1998ee8f248dSBecky Bruce struct page *page; 1999ee8f248dSBecky Bruce 2000ee8f248dSBecky Bruce #ifdef CONFIG_HIGHMEM 2001ee8f248dSBecky Bruce page = pfn_to_page(m->phys >> PAGE_SHIFT); 20028b89a116SGrygorii Strashko memblock_free_late(__pa(m), 2003ee8f248dSBecky Bruce sizeof(struct huge_bootmem_page)); 2004ee8f248dSBecky Bruce #else 2005ee8f248dSBecky Bruce page = virt_to_page(m); 2006ee8f248dSBecky Bruce #endif 2007aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 200818229df5SAndy Whitcroft prep_compound_huge_page(page, h->order); 2009ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 2010aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 2011b0320c7bSRafael Aquini /* 2012b0320c7bSRafael Aquini * If we had gigantic hugepages allocated at boot time, we need 2013b0320c7bSRafael Aquini * to restore the 'stolen' pages to totalram_pages in order to 2014b0320c7bSRafael Aquini * fix confusing memory reports from free(1) and another 2015b0320c7bSRafael Aquini * side-effects, like CommitLimit going negative. 2016b0320c7bSRafael Aquini */ 2017bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 20183dcc0571SJiang Liu adjust_managed_page_count(page, 1 << h->order); 2019aa888a74SAndi Kleen } 2020aa888a74SAndi Kleen } 2021aa888a74SAndi Kleen 20228faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 20231da177e4SLinus Torvalds { 20241da177e4SLinus Torvalds unsigned long i; 20251da177e4SLinus Torvalds 2026e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 2027bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 2028aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 2029aa888a74SAndi Kleen break; 20309b5e5d0fSLee Schermerhorn } else if (!alloc_fresh_huge_page(h, 20318cebfcd0SLai Jiangshan &node_states[N_MEMORY])) 20321da177e4SLinus Torvalds break; 20331da177e4SLinus Torvalds } 20348faa8b07SAndi Kleen h->max_huge_pages = i; 2035e5ff2159SAndi Kleen } 2036e5ff2159SAndi Kleen 2037e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 2038e5ff2159SAndi Kleen { 2039e5ff2159SAndi Kleen struct hstate *h; 2040e5ff2159SAndi Kleen 2041e5ff2159SAndi Kleen for_each_hstate(h) { 2042641844f5SNaoya Horiguchi if (minimum_order > huge_page_order(h)) 2043641844f5SNaoya Horiguchi minimum_order = huge_page_order(h); 2044641844f5SNaoya Horiguchi 20458faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 2046bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 20478faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 2048e5ff2159SAndi Kleen } 2049641844f5SNaoya Horiguchi VM_BUG_ON(minimum_order == UINT_MAX); 2050e5ff2159SAndi Kleen } 2051e5ff2159SAndi Kleen 20524abd32dbSAndi Kleen static char * __init memfmt(char *buf, unsigned long n) 20534abd32dbSAndi Kleen { 20544abd32dbSAndi Kleen if (n >= (1UL << 30)) 20554abd32dbSAndi Kleen sprintf(buf, "%lu GB", n >> 30); 20564abd32dbSAndi Kleen else if (n >= (1UL << 20)) 20574abd32dbSAndi Kleen sprintf(buf, "%lu MB", n >> 20); 20584abd32dbSAndi Kleen else 20594abd32dbSAndi Kleen sprintf(buf, "%lu KB", n >> 10); 20604abd32dbSAndi Kleen return buf; 20614abd32dbSAndi Kleen } 20624abd32dbSAndi Kleen 2063e5ff2159SAndi Kleen static void __init report_hugepages(void) 2064e5ff2159SAndi Kleen { 2065e5ff2159SAndi Kleen struct hstate *h; 2066e5ff2159SAndi Kleen 2067e5ff2159SAndi Kleen for_each_hstate(h) { 20684abd32dbSAndi Kleen char buf[32]; 2069ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 20704abd32dbSAndi Kleen memfmt(buf, huge_page_size(h)), 20714abd32dbSAndi Kleen h->free_huge_pages); 2072e5ff2159SAndi Kleen } 2073e5ff2159SAndi Kleen } 2074e5ff2159SAndi Kleen 20751da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 20766ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 20776ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 20781da177e4SLinus Torvalds { 20794415cc8dSChristoph Lameter int i; 20804415cc8dSChristoph Lameter 2081bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2082aa888a74SAndi Kleen return; 2083aa888a74SAndi Kleen 20846ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 20851da177e4SLinus Torvalds struct page *page, *next; 2086a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 2087a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 2088a5516438SAndi Kleen if (count >= h->nr_huge_pages) 20896b0c880dSAdam Litke return; 20901da177e4SLinus Torvalds if (PageHighMem(page)) 20911da177e4SLinus Torvalds continue; 20921da177e4SLinus Torvalds list_del(&page->lru); 2093e5ff2159SAndi Kleen update_and_free_page(h, page); 2094a5516438SAndi Kleen h->free_huge_pages--; 2095a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 20961da177e4SLinus Torvalds } 20971da177e4SLinus Torvalds } 20981da177e4SLinus Torvalds } 20991da177e4SLinus Torvalds #else 21006ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 21016ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 21021da177e4SLinus Torvalds { 21031da177e4SLinus Torvalds } 21041da177e4SLinus Torvalds #endif 21051da177e4SLinus Torvalds 210620a0307cSWu Fengguang /* 210720a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 210820a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 210920a0307cSWu Fengguang * Returns 1 if an adjustment was made. 211020a0307cSWu Fengguang */ 21116ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 21126ae11b27SLee Schermerhorn int delta) 211320a0307cSWu Fengguang { 2114b2261026SJoonsoo Kim int nr_nodes, node; 211520a0307cSWu Fengguang 211620a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 211720a0307cSWu Fengguang 2118e8c5c824SLee Schermerhorn if (delta < 0) { 2119b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 2120b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 2121b2261026SJoonsoo Kim goto found; 2122b2261026SJoonsoo Kim } 2123b2261026SJoonsoo Kim } else { 2124b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2125b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 2126b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 2127b2261026SJoonsoo Kim goto found; 2128e8c5c824SLee Schermerhorn } 21299a76db09SLee Schermerhorn } 2130b2261026SJoonsoo Kim return 0; 213120a0307cSWu Fengguang 2132b2261026SJoonsoo Kim found: 213320a0307cSWu Fengguang h->surplus_huge_pages += delta; 2134b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 2135b2261026SJoonsoo Kim return 1; 213620a0307cSWu Fengguang } 213720a0307cSWu Fengguang 2138a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 21396ae11b27SLee Schermerhorn static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, 21406ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 21411da177e4SLinus Torvalds { 21427893d1d5SAdam Litke unsigned long min_count, ret; 21431da177e4SLinus Torvalds 2144944d9fecSLuiz Capitulino if (hstate_is_gigantic(h) && !gigantic_page_supported()) 2145aa888a74SAndi Kleen return h->max_huge_pages; 2146aa888a74SAndi Kleen 21477893d1d5SAdam Litke /* 21487893d1d5SAdam Litke * Increase the pool size 21497893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 21507893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 2151d1c3fb1fSNishanth Aravamudan * 2152d15c7c09SNaoya Horiguchi * We might race with __alloc_buddy_huge_page() here and be unable 2153d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 2154d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 2155d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 2156d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 21577893d1d5SAdam Litke */ 21581da177e4SLinus Torvalds spin_lock(&hugetlb_lock); 2159a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 21606ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 21617893d1d5SAdam Litke break; 21627893d1d5SAdam Litke } 21637893d1d5SAdam Litke 2164a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 21657893d1d5SAdam Litke /* 21667893d1d5SAdam Litke * If this allocation races such that we no longer need the 21677893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 21687893d1d5SAdam Litke * and reducing the surplus. 21697893d1d5SAdam Litke */ 21707893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 2171944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) 2172944d9fecSLuiz Capitulino ret = alloc_fresh_gigantic_page(h, nodes_allowed); 2173944d9fecSLuiz Capitulino else 21746ae11b27SLee Schermerhorn ret = alloc_fresh_huge_page(h, nodes_allowed); 21757893d1d5SAdam Litke spin_lock(&hugetlb_lock); 21767893d1d5SAdam Litke if (!ret) 21777893d1d5SAdam Litke goto out; 21787893d1d5SAdam Litke 2179536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 2180536240f2SMel Gorman if (signal_pending(current)) 2181536240f2SMel Gorman goto out; 21827893d1d5SAdam Litke } 21837893d1d5SAdam Litke 21847893d1d5SAdam Litke /* 21857893d1d5SAdam Litke * Decrease the pool size 21867893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 21877893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 21887893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 21897893d1d5SAdam Litke * to the desired size as pages become free. 2190d1c3fb1fSNishanth Aravamudan * 2191d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 2192d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 2193d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 2194d15c7c09SNaoya Horiguchi * __alloc_buddy_huge_page() is checking the global counter, 2195d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 2196d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 2197d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 21987893d1d5SAdam Litke */ 2199a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 22006b0c880dSAdam Litke min_count = max(count, min_count); 22016ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 2202a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 22036ae11b27SLee Schermerhorn if (!free_pool_huge_page(h, nodes_allowed, 0)) 22041da177e4SLinus Torvalds break; 220555f67141SMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 22061da177e4SLinus Torvalds } 2207a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 22086ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 22097893d1d5SAdam Litke break; 22107893d1d5SAdam Litke } 22117893d1d5SAdam Litke out: 2212a5516438SAndi Kleen ret = persistent_huge_pages(h); 22131da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 22147893d1d5SAdam Litke return ret; 22151da177e4SLinus Torvalds } 22161da177e4SLinus Torvalds 2217a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 2218a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 2219a3437870SNishanth Aravamudan 2220a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 2221a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 2222a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 2223a3437870SNishanth Aravamudan 2224a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 2225a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 2226a3437870SNishanth Aravamudan 22279a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 22289a305230SLee Schermerhorn 22299a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 2230a3437870SNishanth Aravamudan { 2231a3437870SNishanth Aravamudan int i; 22329a305230SLee Schermerhorn 2233a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 22349a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 22359a305230SLee Schermerhorn if (nidp) 22369a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 2237a3437870SNishanth Aravamudan return &hstates[i]; 22389a305230SLee Schermerhorn } 22399a305230SLee Schermerhorn 22409a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 2241a3437870SNishanth Aravamudan } 2242a3437870SNishanth Aravamudan 224306808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 2244a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2245a3437870SNishanth Aravamudan { 22469a305230SLee Schermerhorn struct hstate *h; 22479a305230SLee Schermerhorn unsigned long nr_huge_pages; 22489a305230SLee Schermerhorn int nid; 22499a305230SLee Schermerhorn 22509a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 22519a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 22529a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 22539a305230SLee Schermerhorn else 22549a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 22559a305230SLee Schermerhorn 22569a305230SLee Schermerhorn return sprintf(buf, "%lu\n", nr_huge_pages); 2257a3437870SNishanth Aravamudan } 2258adbe8726SEric B Munson 2259238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 2260238d3c13SDavid Rientjes struct hstate *h, int nid, 2261238d3c13SDavid Rientjes unsigned long count, size_t len) 2262a3437870SNishanth Aravamudan { 2263a3437870SNishanth Aravamudan int err; 2264bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); 2265a3437870SNishanth Aravamudan 2266944d9fecSLuiz Capitulino if (hstate_is_gigantic(h) && !gigantic_page_supported()) { 2267adbe8726SEric B Munson err = -EINVAL; 2268adbe8726SEric B Munson goto out; 2269adbe8726SEric B Munson } 2270adbe8726SEric B Munson 22719a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 22729a305230SLee Schermerhorn /* 22739a305230SLee Schermerhorn * global hstate attribute 22749a305230SLee Schermerhorn */ 22759a305230SLee Schermerhorn if (!(obey_mempolicy && 22769a305230SLee Schermerhorn init_nodemask_of_mempolicy(nodes_allowed))) { 227706808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 22788cebfcd0SLai Jiangshan nodes_allowed = &node_states[N_MEMORY]; 227906808b08SLee Schermerhorn } 22809a305230SLee Schermerhorn } else if (nodes_allowed) { 22819a305230SLee Schermerhorn /* 22829a305230SLee Schermerhorn * per node hstate attribute: adjust count to global, 22839a305230SLee Schermerhorn * but restrict alloc/free to the specified node. 22849a305230SLee Schermerhorn */ 22859a305230SLee Schermerhorn count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 22869a305230SLee Schermerhorn init_nodemask_of_node(nodes_allowed, nid); 22879a305230SLee Schermerhorn } else 22888cebfcd0SLai Jiangshan nodes_allowed = &node_states[N_MEMORY]; 22899a305230SLee Schermerhorn 229006808b08SLee Schermerhorn h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); 2291a3437870SNishanth Aravamudan 22928cebfcd0SLai Jiangshan if (nodes_allowed != &node_states[N_MEMORY]) 229306808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 229406808b08SLee Schermerhorn 229506808b08SLee Schermerhorn return len; 2296adbe8726SEric B Munson out: 2297adbe8726SEric B Munson NODEMASK_FREE(nodes_allowed); 2298adbe8726SEric B Munson return err; 229906808b08SLee Schermerhorn } 230006808b08SLee Schermerhorn 2301238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 2302238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 2303238d3c13SDavid Rientjes size_t len) 2304238d3c13SDavid Rientjes { 2305238d3c13SDavid Rientjes struct hstate *h; 2306238d3c13SDavid Rientjes unsigned long count; 2307238d3c13SDavid Rientjes int nid; 2308238d3c13SDavid Rientjes int err; 2309238d3c13SDavid Rientjes 2310238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 2311238d3c13SDavid Rientjes if (err) 2312238d3c13SDavid Rientjes return err; 2313238d3c13SDavid Rientjes 2314238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 2315238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 2316238d3c13SDavid Rientjes } 2317238d3c13SDavid Rientjes 231806808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 231906808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 232006808b08SLee Schermerhorn { 232106808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 232206808b08SLee Schermerhorn } 232306808b08SLee Schermerhorn 232406808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 232506808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 232606808b08SLee Schermerhorn { 2327238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 2328a3437870SNishanth Aravamudan } 2329a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 2330a3437870SNishanth Aravamudan 233106808b08SLee Schermerhorn #ifdef CONFIG_NUMA 233206808b08SLee Schermerhorn 233306808b08SLee Schermerhorn /* 233406808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 233506808b08SLee Schermerhorn * huge page alloc/free. 233606808b08SLee Schermerhorn */ 233706808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 233806808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 233906808b08SLee Schermerhorn { 234006808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 234106808b08SLee Schermerhorn } 234206808b08SLee Schermerhorn 234306808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 234406808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 234506808b08SLee Schermerhorn { 2346238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 234706808b08SLee Schermerhorn } 234806808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 234906808b08SLee Schermerhorn #endif 235006808b08SLee Schermerhorn 235106808b08SLee Schermerhorn 2352a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 2353a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2354a3437870SNishanth Aravamudan { 23559a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2356a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 2357a3437870SNishanth Aravamudan } 2358adbe8726SEric B Munson 2359a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 2360a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 2361a3437870SNishanth Aravamudan { 2362a3437870SNishanth Aravamudan int err; 2363a3437870SNishanth Aravamudan unsigned long input; 23649a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2365a3437870SNishanth Aravamudan 2366bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2367adbe8726SEric B Munson return -EINVAL; 2368adbe8726SEric B Munson 23693dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 2370a3437870SNishanth Aravamudan if (err) 237173ae31e5SEric B Munson return err; 2372a3437870SNishanth Aravamudan 2373a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 2374a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 2375a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 2376a3437870SNishanth Aravamudan 2377a3437870SNishanth Aravamudan return count; 2378a3437870SNishanth Aravamudan } 2379a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 2380a3437870SNishanth Aravamudan 2381a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 2382a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2383a3437870SNishanth Aravamudan { 23849a305230SLee Schermerhorn struct hstate *h; 23859a305230SLee Schermerhorn unsigned long free_huge_pages; 23869a305230SLee Schermerhorn int nid; 23879a305230SLee Schermerhorn 23889a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 23899a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 23909a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 23919a305230SLee Schermerhorn else 23929a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 23939a305230SLee Schermerhorn 23949a305230SLee Schermerhorn return sprintf(buf, "%lu\n", free_huge_pages); 2395a3437870SNishanth Aravamudan } 2396a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 2397a3437870SNishanth Aravamudan 2398a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 2399a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2400a3437870SNishanth Aravamudan { 24019a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2402a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->resv_huge_pages); 2403a3437870SNishanth Aravamudan } 2404a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 2405a3437870SNishanth Aravamudan 2406a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 2407a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2408a3437870SNishanth Aravamudan { 24099a305230SLee Schermerhorn struct hstate *h; 24109a305230SLee Schermerhorn unsigned long surplus_huge_pages; 24119a305230SLee Schermerhorn int nid; 24129a305230SLee Schermerhorn 24139a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 24149a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 24159a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 24169a305230SLee Schermerhorn else 24179a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 24189a305230SLee Schermerhorn 24199a305230SLee Schermerhorn return sprintf(buf, "%lu\n", surplus_huge_pages); 2420a3437870SNishanth Aravamudan } 2421a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 2422a3437870SNishanth Aravamudan 2423a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 2424a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 2425a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 2426a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 2427a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 2428a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 242906808b08SLee Schermerhorn #ifdef CONFIG_NUMA 243006808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 243106808b08SLee Schermerhorn #endif 2432a3437870SNishanth Aravamudan NULL, 2433a3437870SNishanth Aravamudan }; 2434a3437870SNishanth Aravamudan 2435a3437870SNishanth Aravamudan static struct attribute_group hstate_attr_group = { 2436a3437870SNishanth Aravamudan .attrs = hstate_attrs, 2437a3437870SNishanth Aravamudan }; 2438a3437870SNishanth Aravamudan 2439094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 24409a305230SLee Schermerhorn struct kobject **hstate_kobjs, 24419a305230SLee Schermerhorn struct attribute_group *hstate_attr_group) 2442a3437870SNishanth Aravamudan { 2443a3437870SNishanth Aravamudan int retval; 2444972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 2445a3437870SNishanth Aravamudan 24469a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 24479a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 2448a3437870SNishanth Aravamudan return -ENOMEM; 2449a3437870SNishanth Aravamudan 24509a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 2451a3437870SNishanth Aravamudan if (retval) 24529a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 2453a3437870SNishanth Aravamudan 2454a3437870SNishanth Aravamudan return retval; 2455a3437870SNishanth Aravamudan } 2456a3437870SNishanth Aravamudan 2457a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 2458a3437870SNishanth Aravamudan { 2459a3437870SNishanth Aravamudan struct hstate *h; 2460a3437870SNishanth Aravamudan int err; 2461a3437870SNishanth Aravamudan 2462a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 2463a3437870SNishanth Aravamudan if (!hugepages_kobj) 2464a3437870SNishanth Aravamudan return; 2465a3437870SNishanth Aravamudan 2466a3437870SNishanth Aravamudan for_each_hstate(h) { 24679a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 24689a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 2469a3437870SNishanth Aravamudan if (err) 2470ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s", h->name); 2471a3437870SNishanth Aravamudan } 2472a3437870SNishanth Aravamudan } 2473a3437870SNishanth Aravamudan 24749a305230SLee Schermerhorn #ifdef CONFIG_NUMA 24759a305230SLee Schermerhorn 24769a305230SLee Schermerhorn /* 24779a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 247810fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 247910fbcf4cSKay Sievers * index of a node device or _hstate == node id. 248010fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 24819a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 24829a305230SLee Schermerhorn */ 24839a305230SLee Schermerhorn struct node_hstate { 24849a305230SLee Schermerhorn struct kobject *hugepages_kobj; 24859a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 24869a305230SLee Schermerhorn }; 2487b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 24889a305230SLee Schermerhorn 24899a305230SLee Schermerhorn /* 249010fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 24919a305230SLee Schermerhorn */ 24929a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 24939a305230SLee Schermerhorn &nr_hugepages_attr.attr, 24949a305230SLee Schermerhorn &free_hugepages_attr.attr, 24959a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 24969a305230SLee Schermerhorn NULL, 24979a305230SLee Schermerhorn }; 24989a305230SLee Schermerhorn 24999a305230SLee Schermerhorn static struct attribute_group per_node_hstate_attr_group = { 25009a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 25019a305230SLee Schermerhorn }; 25029a305230SLee Schermerhorn 25039a305230SLee Schermerhorn /* 250410fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 25059a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 25069a305230SLee Schermerhorn */ 25079a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 25089a305230SLee Schermerhorn { 25099a305230SLee Schermerhorn int nid; 25109a305230SLee Schermerhorn 25119a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 25129a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 25139a305230SLee Schermerhorn int i; 25149a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 25159a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 25169a305230SLee Schermerhorn if (nidp) 25179a305230SLee Schermerhorn *nidp = nid; 25189a305230SLee Schermerhorn return &hstates[i]; 25199a305230SLee Schermerhorn } 25209a305230SLee Schermerhorn } 25219a305230SLee Schermerhorn 25229a305230SLee Schermerhorn BUG(); 25239a305230SLee Schermerhorn return NULL; 25249a305230SLee Schermerhorn } 25259a305230SLee Schermerhorn 25269a305230SLee Schermerhorn /* 252710fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 25289a305230SLee Schermerhorn * No-op if no hstate attributes attached. 25299a305230SLee Schermerhorn */ 25303cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 25319a305230SLee Schermerhorn { 25329a305230SLee Schermerhorn struct hstate *h; 253310fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 25349a305230SLee Schermerhorn 25359a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 25369b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 25379a305230SLee Schermerhorn 2538972dc4deSAneesh Kumar K.V for_each_hstate(h) { 2539972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 2540972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 2541972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 2542972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 2543972dc4deSAneesh Kumar K.V } 25449a305230SLee Schermerhorn } 25459a305230SLee Schermerhorn 25469a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 25479a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 25489a305230SLee Schermerhorn } 25499a305230SLee Schermerhorn 25509a305230SLee Schermerhorn 25519a305230SLee Schermerhorn /* 255210fbcf4cSKay Sievers * Register hstate attributes for a single node device. 25539a305230SLee Schermerhorn * No-op if attributes already registered. 25549a305230SLee Schermerhorn */ 25553cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 25569a305230SLee Schermerhorn { 25579a305230SLee Schermerhorn struct hstate *h; 255810fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 25599a305230SLee Schermerhorn int err; 25609a305230SLee Schermerhorn 25619a305230SLee Schermerhorn if (nhs->hugepages_kobj) 25629a305230SLee Schermerhorn return; /* already allocated */ 25639a305230SLee Schermerhorn 25649a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 256510fbcf4cSKay Sievers &node->dev.kobj); 25669a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 25679a305230SLee Schermerhorn return; 25689a305230SLee Schermerhorn 25699a305230SLee Schermerhorn for_each_hstate(h) { 25709a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 25719a305230SLee Schermerhorn nhs->hstate_kobjs, 25729a305230SLee Schermerhorn &per_node_hstate_attr_group); 25739a305230SLee Schermerhorn if (err) { 2574ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s for node %d\n", 257510fbcf4cSKay Sievers h->name, node->dev.id); 25769a305230SLee Schermerhorn hugetlb_unregister_node(node); 25779a305230SLee Schermerhorn break; 25789a305230SLee Schermerhorn } 25799a305230SLee Schermerhorn } 25809a305230SLee Schermerhorn } 25819a305230SLee Schermerhorn 25829a305230SLee Schermerhorn /* 25839b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 258410fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 258510fbcf4cSKay Sievers * registered their associated device by this time. 25869a305230SLee Schermerhorn */ 25877d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 25889a305230SLee Schermerhorn { 25899a305230SLee Schermerhorn int nid; 25909a305230SLee Schermerhorn 25918cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 25928732794bSWen Congyang struct node *node = node_devices[nid]; 259310fbcf4cSKay Sievers if (node->dev.id == nid) 25949a305230SLee Schermerhorn hugetlb_register_node(node); 25959a305230SLee Schermerhorn } 25969a305230SLee Schermerhorn 25979a305230SLee Schermerhorn /* 259810fbcf4cSKay Sievers * Let the node device driver know we're here so it can 25999a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 26009a305230SLee Schermerhorn */ 26019a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 26029a305230SLee Schermerhorn hugetlb_unregister_node); 26039a305230SLee Schermerhorn } 26049a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 26059a305230SLee Schermerhorn 26069a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 26079a305230SLee Schermerhorn { 26089a305230SLee Schermerhorn BUG(); 26099a305230SLee Schermerhorn if (nidp) 26109a305230SLee Schermerhorn *nidp = -1; 26119a305230SLee Schermerhorn return NULL; 26129a305230SLee Schermerhorn } 26139a305230SLee Schermerhorn 26149a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 26159a305230SLee Schermerhorn 26169a305230SLee Schermerhorn #endif 26179a305230SLee Schermerhorn 2618a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 2619a3437870SNishanth Aravamudan { 26208382d914SDavidlohr Bueso int i; 26218382d914SDavidlohr Bueso 2622457c1b27SNishanth Aravamudan if (!hugepages_supported()) 26230ef89d25SBenjamin Herrenschmidt return 0; 2624a3437870SNishanth Aravamudan 2625e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) { 2626e11bfbfcSNick Piggin default_hstate_size = HPAGE_SIZE; 2627e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) 2628a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 2629a3437870SNishanth Aravamudan } 2630972dc4deSAneesh Kumar K.V default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); 2631f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 2632f8b74815SVaishali Thakkar if (!default_hstate.max_huge_pages) 2633e11bfbfcSNick Piggin default_hstate.max_huge_pages = default_hstate_max_huge_pages; 2634f8b74815SVaishali Thakkar } 2635a3437870SNishanth Aravamudan 2636a3437870SNishanth Aravamudan hugetlb_init_hstates(); 2637aa888a74SAndi Kleen gather_bootmem_prealloc(); 2638a3437870SNishanth Aravamudan report_hugepages(); 2639a3437870SNishanth Aravamudan 2640a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 26419a305230SLee Schermerhorn hugetlb_register_all_nodes(); 26427179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 26439a305230SLee Schermerhorn 26448382d914SDavidlohr Bueso #ifdef CONFIG_SMP 26458382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 26468382d914SDavidlohr Bueso #else 26478382d914SDavidlohr Bueso num_fault_mutexes = 1; 26488382d914SDavidlohr Bueso #endif 2649c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 26508382d914SDavidlohr Bueso kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL); 2651c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 26528382d914SDavidlohr Bueso 26538382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 2654c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 2655a3437870SNishanth Aravamudan return 0; 2656a3437870SNishanth Aravamudan } 26573e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 2658a3437870SNishanth Aravamudan 2659a3437870SNishanth Aravamudan /* Should be called on processing a hugepagesz=... option */ 2660d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 2661a3437870SNishanth Aravamudan { 2662a3437870SNishanth Aravamudan struct hstate *h; 26638faa8b07SAndi Kleen unsigned long i; 26648faa8b07SAndi Kleen 2665a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 2666598d8091SJoe Perches pr_warn("hugepagesz= specified twice, ignoring\n"); 2667a3437870SNishanth Aravamudan return; 2668a3437870SNishanth Aravamudan } 266947d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 2670a3437870SNishanth Aravamudan BUG_ON(order == 0); 267147d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 2672a3437870SNishanth Aravamudan h->order = order; 2673a3437870SNishanth Aravamudan h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 26748faa8b07SAndi Kleen h->nr_huge_pages = 0; 26758faa8b07SAndi Kleen h->free_huge_pages = 0; 26768faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 26778faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 26780edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 26798cebfcd0SLai Jiangshan h->next_nid_to_alloc = first_node(node_states[N_MEMORY]); 26808cebfcd0SLai Jiangshan h->next_nid_to_free = first_node(node_states[N_MEMORY]); 2681a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 2682a3437870SNishanth Aravamudan huge_page_size(h)/1024); 26838faa8b07SAndi Kleen 2684a3437870SNishanth Aravamudan parsed_hstate = h; 2685a3437870SNishanth Aravamudan } 2686a3437870SNishanth Aravamudan 2687e11bfbfcSNick Piggin static int __init hugetlb_nrpages_setup(char *s) 2688a3437870SNishanth Aravamudan { 2689a3437870SNishanth Aravamudan unsigned long *mhp; 26908faa8b07SAndi Kleen static unsigned long *last_mhp; 2691a3437870SNishanth Aravamudan 2692a3437870SNishanth Aravamudan /* 269347d38344SAneesh Kumar K.V * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, 2694a3437870SNishanth Aravamudan * so this hugepages= parameter goes to the "default hstate". 2695a3437870SNishanth Aravamudan */ 269647d38344SAneesh Kumar K.V if (!hugetlb_max_hstate) 2697a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 2698a3437870SNishanth Aravamudan else 2699a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 2700a3437870SNishanth Aravamudan 27018faa8b07SAndi Kleen if (mhp == last_mhp) { 2702598d8091SJoe Perches pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n"); 27038faa8b07SAndi Kleen return 1; 27048faa8b07SAndi Kleen } 27058faa8b07SAndi Kleen 2706a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 2707a3437870SNishanth Aravamudan *mhp = 0; 2708a3437870SNishanth Aravamudan 27098faa8b07SAndi Kleen /* 27108faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 27118faa8b07SAndi Kleen * But we need to allocate >= MAX_ORDER hstates here early to still 27128faa8b07SAndi Kleen * use the bootmem allocator. 27138faa8b07SAndi Kleen */ 271447d38344SAneesh Kumar K.V if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) 27158faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 27168faa8b07SAndi Kleen 27178faa8b07SAndi Kleen last_mhp = mhp; 27188faa8b07SAndi Kleen 2719a3437870SNishanth Aravamudan return 1; 2720a3437870SNishanth Aravamudan } 2721e11bfbfcSNick Piggin __setup("hugepages=", hugetlb_nrpages_setup); 2722e11bfbfcSNick Piggin 2723e11bfbfcSNick Piggin static int __init hugetlb_default_setup(char *s) 2724e11bfbfcSNick Piggin { 2725e11bfbfcSNick Piggin default_hstate_size = memparse(s, &s); 2726e11bfbfcSNick Piggin return 1; 2727e11bfbfcSNick Piggin } 2728e11bfbfcSNick Piggin __setup("default_hugepagesz=", hugetlb_default_setup); 2729a3437870SNishanth Aravamudan 27308a213460SNishanth Aravamudan static unsigned int cpuset_mems_nr(unsigned int *array) 27318a213460SNishanth Aravamudan { 27328a213460SNishanth Aravamudan int node; 27338a213460SNishanth Aravamudan unsigned int nr = 0; 27348a213460SNishanth Aravamudan 27358a213460SNishanth Aravamudan for_each_node_mask(node, cpuset_current_mems_allowed) 27368a213460SNishanth Aravamudan nr += array[node]; 27378a213460SNishanth Aravamudan 27388a213460SNishanth Aravamudan return nr; 27398a213460SNishanth Aravamudan } 27408a213460SNishanth Aravamudan 27418a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 274206808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 274306808b08SLee Schermerhorn struct ctl_table *table, int write, 274406808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 27451da177e4SLinus Torvalds { 2746e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 2747238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 274808d4a246SMichal Hocko int ret; 2749e5ff2159SAndi Kleen 2750457c1b27SNishanth Aravamudan if (!hugepages_supported()) 275186613628SJan Stancek return -EOPNOTSUPP; 2752457c1b27SNishanth Aravamudan 2753e5ff2159SAndi Kleen table->data = &tmp; 2754e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 275508d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 275608d4a246SMichal Hocko if (ret) 275708d4a246SMichal Hocko goto out; 2758e5ff2159SAndi Kleen 2759238d3c13SDavid Rientjes if (write) 2760238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 2761238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 276208d4a246SMichal Hocko out: 276308d4a246SMichal Hocko return ret; 27641da177e4SLinus Torvalds } 2765396faf03SMel Gorman 276606808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 276706808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 276806808b08SLee Schermerhorn { 276906808b08SLee Schermerhorn 277006808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 277106808b08SLee Schermerhorn buffer, length, ppos); 277206808b08SLee Schermerhorn } 277306808b08SLee Schermerhorn 277406808b08SLee Schermerhorn #ifdef CONFIG_NUMA 277506808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 277606808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 277706808b08SLee Schermerhorn { 277806808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 277906808b08SLee Schermerhorn buffer, length, ppos); 278006808b08SLee Schermerhorn } 278106808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 278206808b08SLee Schermerhorn 2783a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 27848d65af78SAlexey Dobriyan void __user *buffer, 2785a3d0c6aaSNishanth Aravamudan size_t *length, loff_t *ppos) 2786a3d0c6aaSNishanth Aravamudan { 2787a5516438SAndi Kleen struct hstate *h = &default_hstate; 2788e5ff2159SAndi Kleen unsigned long tmp; 278908d4a246SMichal Hocko int ret; 2790e5ff2159SAndi Kleen 2791457c1b27SNishanth Aravamudan if (!hugepages_supported()) 279286613628SJan Stancek return -EOPNOTSUPP; 2793457c1b27SNishanth Aravamudan 2794e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 2795e5ff2159SAndi Kleen 2796bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 2797adbe8726SEric B Munson return -EINVAL; 2798adbe8726SEric B Munson 2799e5ff2159SAndi Kleen table->data = &tmp; 2800e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 280108d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 280208d4a246SMichal Hocko if (ret) 280308d4a246SMichal Hocko goto out; 2804e5ff2159SAndi Kleen 2805e5ff2159SAndi Kleen if (write) { 2806064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 2807e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 2808a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 2809e5ff2159SAndi Kleen } 281008d4a246SMichal Hocko out: 281108d4a246SMichal Hocko return ret; 2812a3d0c6aaSNishanth Aravamudan } 2813a3d0c6aaSNishanth Aravamudan 28141da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 28151da177e4SLinus Torvalds 2816e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 28171da177e4SLinus Torvalds { 2818a5516438SAndi Kleen struct hstate *h = &default_hstate; 2819457c1b27SNishanth Aravamudan if (!hugepages_supported()) 2820457c1b27SNishanth Aravamudan return; 2821e1759c21SAlexey Dobriyan seq_printf(m, 28221da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 28231da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 2824b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 28257893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 28264f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 2827a5516438SAndi Kleen h->nr_huge_pages, 2828a5516438SAndi Kleen h->free_huge_pages, 2829a5516438SAndi Kleen h->resv_huge_pages, 2830a5516438SAndi Kleen h->surplus_huge_pages, 2831a5516438SAndi Kleen 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 28321da177e4SLinus Torvalds } 28331da177e4SLinus Torvalds 28341da177e4SLinus Torvalds int hugetlb_report_node_meminfo(int nid, char *buf) 28351da177e4SLinus Torvalds { 2836a5516438SAndi Kleen struct hstate *h = &default_hstate; 2837457c1b27SNishanth Aravamudan if (!hugepages_supported()) 2838457c1b27SNishanth Aravamudan return 0; 28391da177e4SLinus Torvalds return sprintf(buf, 28401da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 2841a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 2842a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 2843a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 2844a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 2845a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 28461da177e4SLinus Torvalds } 28471da177e4SLinus Torvalds 2848949f7ec5SDavid Rientjes void hugetlb_show_meminfo(void) 2849949f7ec5SDavid Rientjes { 2850949f7ec5SDavid Rientjes struct hstate *h; 2851949f7ec5SDavid Rientjes int nid; 2852949f7ec5SDavid Rientjes 2853457c1b27SNishanth Aravamudan if (!hugepages_supported()) 2854457c1b27SNishanth Aravamudan return; 2855457c1b27SNishanth Aravamudan 2856949f7ec5SDavid Rientjes for_each_node_state(nid, N_MEMORY) 2857949f7ec5SDavid Rientjes for_each_hstate(h) 2858949f7ec5SDavid Rientjes pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 2859949f7ec5SDavid Rientjes nid, 2860949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 2861949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 2862949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 2863949f7ec5SDavid Rientjes 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 2864949f7ec5SDavid Rientjes } 2865949f7ec5SDavid Rientjes 28665d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 28675d317b2bSNaoya Horiguchi { 28685d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 28695d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 28705d317b2bSNaoya Horiguchi } 28715d317b2bSNaoya Horiguchi 28721da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 28731da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 28741da177e4SLinus Torvalds { 2875d0028588SWanpeng Li struct hstate *h; 2876d0028588SWanpeng Li unsigned long nr_total_pages = 0; 2877d0028588SWanpeng Li 2878d0028588SWanpeng Li for_each_hstate(h) 2879d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 2880d0028588SWanpeng Li return nr_total_pages; 28811da177e4SLinus Torvalds } 28821da177e4SLinus Torvalds 2883a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 2884fc1b8a73SMel Gorman { 2885fc1b8a73SMel Gorman int ret = -ENOMEM; 2886fc1b8a73SMel Gorman 2887fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 2888fc1b8a73SMel Gorman /* 2889fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 2890fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 2891fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 2892fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 2893fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 2894fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 2895fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 2896fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 2897fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 2898fc1b8a73SMel Gorman * 2899fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 2900fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 2901fc1b8a73SMel Gorman * we fall back to check against current free page availability as 2902fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 2903fc1b8a73SMel Gorman * semantics that cpuset has. 2904fc1b8a73SMel Gorman */ 2905fc1b8a73SMel Gorman if (delta > 0) { 2906a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 2907fc1b8a73SMel Gorman goto out; 2908fc1b8a73SMel Gorman 2909a5516438SAndi Kleen if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 2910a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 2911fc1b8a73SMel Gorman goto out; 2912fc1b8a73SMel Gorman } 2913fc1b8a73SMel Gorman } 2914fc1b8a73SMel Gorman 2915fc1b8a73SMel Gorman ret = 0; 2916fc1b8a73SMel Gorman if (delta < 0) 2917a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 2918fc1b8a73SMel Gorman 2919fc1b8a73SMel Gorman out: 2920fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 2921fc1b8a73SMel Gorman return ret; 2922fc1b8a73SMel Gorman } 2923fc1b8a73SMel Gorman 292484afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 292584afd99bSAndy Whitcroft { 2926f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 292784afd99bSAndy Whitcroft 292884afd99bSAndy Whitcroft /* 292984afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 293084afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 293184afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 293225985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 293384afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 293484afd99bSAndy Whitcroft * new reference here without additional locking. 293584afd99bSAndy Whitcroft */ 29364e35f483SJoonsoo Kim if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 2937f522c3acSJoonsoo Kim kref_get(&resv->refs); 293884afd99bSAndy Whitcroft } 293984afd99bSAndy Whitcroft 2940a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 2941a1e78772SMel Gorman { 2942a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2943f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 294490481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 29454e35f483SJoonsoo Kim unsigned long reserve, start, end; 29461c5ecae3SMike Kravetz long gbl_reserve; 294784afd99bSAndy Whitcroft 29484e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 29494e35f483SJoonsoo Kim return; 29504e35f483SJoonsoo Kim 2951a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 2952a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 295384afd99bSAndy Whitcroft 29544e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 295584afd99bSAndy Whitcroft 2956f031dd27SJoonsoo Kim kref_put(&resv->refs, resv_map_release); 295784afd99bSAndy Whitcroft 29587251ff78SAdam Litke if (reserve) { 29591c5ecae3SMike Kravetz /* 29601c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 29611c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 29621c5ecae3SMike Kravetz */ 29631c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 29641c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 29657251ff78SAdam Litke } 2966a1e78772SMel Gorman } 2967a1e78772SMel Gorman 29681da177e4SLinus Torvalds /* 29691da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 29701da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 29711da177e4SLinus Torvalds * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 29721da177e4SLinus Torvalds * this far. 29731da177e4SLinus Torvalds */ 2974d0217ac0SNick Piggin static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) 29751da177e4SLinus Torvalds { 29761da177e4SLinus Torvalds BUG(); 2977d0217ac0SNick Piggin return 0; 29781da177e4SLinus Torvalds } 29791da177e4SLinus Torvalds 2980f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 2981d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 298284afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 2983a1e78772SMel Gorman .close = hugetlb_vm_op_close, 29841da177e4SLinus Torvalds }; 29851da177e4SLinus Torvalds 29861e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 29871e8f889bSDavid Gibson int writable) 298863551ae0SDavid Gibson { 298963551ae0SDavid Gibson pte_t entry; 299063551ae0SDavid Gibson 29911e8f889bSDavid Gibson if (writable) { 2992106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 2993106c992aSGerald Schaefer vma->vm_page_prot))); 299463551ae0SDavid Gibson } else { 2995106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 2996106c992aSGerald Schaefer vma->vm_page_prot)); 299763551ae0SDavid Gibson } 299863551ae0SDavid Gibson entry = pte_mkyoung(entry); 299963551ae0SDavid Gibson entry = pte_mkhuge(entry); 3000d9ed9faaSChris Metcalf entry = arch_make_huge_pte(entry, vma, page, writable); 300163551ae0SDavid Gibson 300263551ae0SDavid Gibson return entry; 300363551ae0SDavid Gibson } 300463551ae0SDavid Gibson 30051e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 30061e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 30071e8f889bSDavid Gibson { 30081e8f889bSDavid Gibson pte_t entry; 30091e8f889bSDavid Gibson 3010106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 301132f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 30124b3073e1SRussell King update_mmu_cache(vma, address, ptep); 30131e8f889bSDavid Gibson } 30141e8f889bSDavid Gibson 30154a705fefSNaoya Horiguchi static int is_hugetlb_entry_migration(pte_t pte) 30164a705fefSNaoya Horiguchi { 30174a705fefSNaoya Horiguchi swp_entry_t swp; 30184a705fefSNaoya Horiguchi 30194a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 30204a705fefSNaoya Horiguchi return 0; 30214a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 30224a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_migration_entry(swp)) 30234a705fefSNaoya Horiguchi return 1; 30244a705fefSNaoya Horiguchi else 30254a705fefSNaoya Horiguchi return 0; 30264a705fefSNaoya Horiguchi } 30274a705fefSNaoya Horiguchi 30284a705fefSNaoya Horiguchi static int is_hugetlb_entry_hwpoisoned(pte_t pte) 30294a705fefSNaoya Horiguchi { 30304a705fefSNaoya Horiguchi swp_entry_t swp; 30314a705fefSNaoya Horiguchi 30324a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 30334a705fefSNaoya Horiguchi return 0; 30344a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 30354a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_hwpoison_entry(swp)) 30364a705fefSNaoya Horiguchi return 1; 30374a705fefSNaoya Horiguchi else 30384a705fefSNaoya Horiguchi return 0; 30394a705fefSNaoya Horiguchi } 30401e8f889bSDavid Gibson 304163551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 304263551ae0SDavid Gibson struct vm_area_struct *vma) 304363551ae0SDavid Gibson { 304463551ae0SDavid Gibson pte_t *src_pte, *dst_pte, entry; 304563551ae0SDavid Gibson struct page *ptepage; 30461c59827dSHugh Dickins unsigned long addr; 30471e8f889bSDavid Gibson int cow; 3048a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3049a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 3050e8569dd2SAndreas Sandberg unsigned long mmun_start; /* For mmu_notifiers */ 3051e8569dd2SAndreas Sandberg unsigned long mmun_end; /* For mmu_notifiers */ 3052e8569dd2SAndreas Sandberg int ret = 0; 30531e8f889bSDavid Gibson 30541e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 305563551ae0SDavid Gibson 3056e8569dd2SAndreas Sandberg mmun_start = vma->vm_start; 3057e8569dd2SAndreas Sandberg mmun_end = vma->vm_end; 3058e8569dd2SAndreas Sandberg if (cow) 3059e8569dd2SAndreas Sandberg mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end); 3060e8569dd2SAndreas Sandberg 3061a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 3062cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 3063c74df32cSHugh Dickins src_pte = huge_pte_offset(src, addr); 3064c74df32cSHugh Dickins if (!src_pte) 3065c74df32cSHugh Dickins continue; 3066a5516438SAndi Kleen dst_pte = huge_pte_alloc(dst, addr, sz); 3067e8569dd2SAndreas Sandberg if (!dst_pte) { 3068e8569dd2SAndreas Sandberg ret = -ENOMEM; 3069e8569dd2SAndreas Sandberg break; 3070e8569dd2SAndreas Sandberg } 3071c5c99429SLarry Woodman 3072c5c99429SLarry Woodman /* If the pagetables are shared don't copy or take references */ 3073c5c99429SLarry Woodman if (dst_pte == src_pte) 3074c5c99429SLarry Woodman continue; 3075c5c99429SLarry Woodman 3076cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 3077cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 3078cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 30794a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 30804a705fefSNaoya Horiguchi if (huge_pte_none(entry)) { /* skip none entry */ 30814a705fefSNaoya Horiguchi ; 30824a705fefSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry) || 30834a705fefSNaoya Horiguchi is_hugetlb_entry_hwpoisoned(entry))) { 30844a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 30854a705fefSNaoya Horiguchi 30864a705fefSNaoya Horiguchi if (is_write_migration_entry(swp_entry) && cow) { 30874a705fefSNaoya Horiguchi /* 30884a705fefSNaoya Horiguchi * COW mappings require pages in both 30894a705fefSNaoya Horiguchi * parent and child to be set to read. 30904a705fefSNaoya Horiguchi */ 30914a705fefSNaoya Horiguchi make_migration_entry_read(&swp_entry); 30924a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 30934a705fefSNaoya Horiguchi set_huge_pte_at(src, addr, src_pte, entry); 30944a705fefSNaoya Horiguchi } 30954a705fefSNaoya Horiguchi set_huge_pte_at(dst, addr, dst_pte, entry); 30964a705fefSNaoya Horiguchi } else { 309734ee645eSJoerg Roedel if (cow) { 30987f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 309934ee645eSJoerg Roedel mmu_notifier_invalidate_range(src, mmun_start, 310034ee645eSJoerg Roedel mmun_end); 310134ee645eSJoerg Roedel } 31020253d634SNaoya Horiguchi entry = huge_ptep_get(src_pte); 310363551ae0SDavid Gibson ptepage = pte_page(entry); 310463551ae0SDavid Gibson get_page(ptepage); 310553f9263bSKirill A. Shutemov page_dup_rmap(ptepage, true); 310663551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 31075d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), dst); 31081c59827dSHugh Dickins } 3109cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 3110cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 311163551ae0SDavid Gibson } 311263551ae0SDavid Gibson 3113e8569dd2SAndreas Sandberg if (cow) 3114e8569dd2SAndreas Sandberg mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end); 3115e8569dd2SAndreas Sandberg 3116e8569dd2SAndreas Sandberg return ret; 311763551ae0SDavid Gibson } 311863551ae0SDavid Gibson 311924669e58SAneesh Kumar K.V void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 312024669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 312124669e58SAneesh Kumar K.V struct page *ref_page) 312263551ae0SDavid Gibson { 312324669e58SAneesh Kumar K.V int force_flush = 0; 312463551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 312563551ae0SDavid Gibson unsigned long address; 3126c7546f8fSDavid Gibson pte_t *ptep; 312763551ae0SDavid Gibson pte_t pte; 3128cb900f41SKirill A. Shutemov spinlock_t *ptl; 312963551ae0SDavid Gibson struct page *page; 3130a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3131a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 31322ec74c3eSSagi Grimberg const unsigned long mmun_start = start; /* For mmu_notifiers */ 31332ec74c3eSSagi Grimberg const unsigned long mmun_end = end; /* For mmu_notifiers */ 3134a5516438SAndi Kleen 313563551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 3136a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 3137a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 313863551ae0SDavid Gibson 313924669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 31402ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 3141569f48b8SHillf Danton address = start; 314224669e58SAneesh Kumar K.V again: 3143569f48b8SHillf Danton for (; address < end; address += sz) { 3144c7546f8fSDavid Gibson ptep = huge_pte_offset(mm, address); 3145c7546f8fSDavid Gibson if (!ptep) 3146c7546f8fSDavid Gibson continue; 3147c7546f8fSDavid Gibson 3148cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 314939dde65cSChen, Kenneth W if (huge_pmd_unshare(mm, &address, ptep)) 3150cb900f41SKirill A. Shutemov goto unlock; 315139dde65cSChen, Kenneth W 31526629326bSHillf Danton pte = huge_ptep_get(ptep); 31536629326bSHillf Danton if (huge_pte_none(pte)) 3154cb900f41SKirill A. Shutemov goto unlock; 31556629326bSHillf Danton 31566629326bSHillf Danton /* 31579fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 31589fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 31596629326bSHillf Danton */ 31609fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 3161106c992aSGerald Schaefer huge_pte_clear(mm, address, ptep); 3162cb900f41SKirill A. Shutemov goto unlock; 31638c4894c6SNaoya Horiguchi } 31646629326bSHillf Danton 31656629326bSHillf Danton page = pte_page(pte); 316604f2cbe3SMel Gorman /* 316704f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 316804f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 316904f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 317004f2cbe3SMel Gorman */ 317104f2cbe3SMel Gorman if (ref_page) { 317204f2cbe3SMel Gorman if (page != ref_page) 3173cb900f41SKirill A. Shutemov goto unlock; 317404f2cbe3SMel Gorman 317504f2cbe3SMel Gorman /* 317604f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 317704f2cbe3SMel Gorman * future faults in this VMA will fail rather than 317804f2cbe3SMel Gorman * looking like data was lost 317904f2cbe3SMel Gorman */ 318004f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 318104f2cbe3SMel Gorman } 318204f2cbe3SMel Gorman 3183c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 318424669e58SAneesh Kumar K.V tlb_remove_tlb_entry(tlb, ptep, address); 3185106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 31866649a386SKen Chen set_page_dirty(page); 31879e81130bSHillf Danton 31885d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 3189d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 319024669e58SAneesh Kumar K.V force_flush = !__tlb_remove_page(tlb, page); 3191cb900f41SKirill A. Shutemov if (force_flush) { 3192569f48b8SHillf Danton address += sz; 3193cb900f41SKirill A. Shutemov spin_unlock(ptl); 31949e81130bSHillf Danton break; 319563551ae0SDavid Gibson } 3196cb900f41SKirill A. Shutemov /* Bail out after unmapping reference page if supplied */ 3197cb900f41SKirill A. Shutemov if (ref_page) { 3198cb900f41SKirill A. Shutemov spin_unlock(ptl); 3199cb900f41SKirill A. Shutemov break; 3200cb900f41SKirill A. Shutemov } 3201cb900f41SKirill A. Shutemov unlock: 3202cb900f41SKirill A. Shutemov spin_unlock(ptl); 3203cb900f41SKirill A. Shutemov } 320424669e58SAneesh Kumar K.V /* 320524669e58SAneesh Kumar K.V * mmu_gather ran out of room to batch pages, we break out of 320624669e58SAneesh Kumar K.V * the PTE lock to avoid doing the potential expensive TLB invalidate 320724669e58SAneesh Kumar K.V * and page-free while holding it. 320824669e58SAneesh Kumar K.V */ 320924669e58SAneesh Kumar K.V if (force_flush) { 321024669e58SAneesh Kumar K.V force_flush = 0; 321124669e58SAneesh Kumar K.V tlb_flush_mmu(tlb); 321224669e58SAneesh Kumar K.V if (address < end && !ref_page) 321324669e58SAneesh Kumar K.V goto again; 3214fe1668aeSChen, Kenneth W } 32152ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 321624669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 32171da177e4SLinus Torvalds } 321863551ae0SDavid Gibson 3219d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 3220d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 3221d833352aSMel Gorman unsigned long end, struct page *ref_page) 3222d833352aSMel Gorman { 3223d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 3224d833352aSMel Gorman 3225d833352aSMel Gorman /* 3226d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 3227d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 3228d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 3229d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 3230d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 3231c8c06efaSDavidlohr Bueso * is to clear it before releasing the i_mmap_rwsem. This works 3232d833352aSMel Gorman * because in the context this is called, the VMA is about to be 3233c8c06efaSDavidlohr Bueso * destroyed and the i_mmap_rwsem is held. 3234d833352aSMel Gorman */ 3235d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 3236d833352aSMel Gorman } 3237d833352aSMel Gorman 3238502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 323904f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 3240502717f4SChen, Kenneth W { 324124669e58SAneesh Kumar K.V struct mm_struct *mm; 324224669e58SAneesh Kumar K.V struct mmu_gather tlb; 324324669e58SAneesh Kumar K.V 324424669e58SAneesh Kumar K.V mm = vma->vm_mm; 324524669e58SAneesh Kumar K.V 32462b047252SLinus Torvalds tlb_gather_mmu(&tlb, mm, start, end); 324724669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 324824669e58SAneesh Kumar K.V tlb_finish_mmu(&tlb, start, end); 3249502717f4SChen, Kenneth W } 3250502717f4SChen, Kenneth W 325104f2cbe3SMel Gorman /* 325204f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 325304f2cbe3SMel Gorman * mappping it owns the reserve page for. The intention is to unmap the page 325404f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 325504f2cbe3SMel Gorman * same region. 325604f2cbe3SMel Gorman */ 32572f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 32582a4b3dedSHarvey Harrison struct page *page, unsigned long address) 325904f2cbe3SMel Gorman { 32607526674dSAdam Litke struct hstate *h = hstate_vma(vma); 326104f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 326204f2cbe3SMel Gorman struct address_space *mapping; 326304f2cbe3SMel Gorman pgoff_t pgoff; 326404f2cbe3SMel Gorman 326504f2cbe3SMel Gorman /* 326604f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 326704f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 326804f2cbe3SMel Gorman */ 32697526674dSAdam Litke address = address & huge_page_mask(h); 327036e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 327136e4f20aSMichal Hocko vma->vm_pgoff; 3272496ad9aaSAl Viro mapping = file_inode(vma->vm_file)->i_mapping; 327304f2cbe3SMel Gorman 32744eb2b1dcSMel Gorman /* 32754eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 32764eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 32774eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 32784eb2b1dcSMel Gorman */ 327983cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 32806b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 328104f2cbe3SMel Gorman /* Do not unmap the current VMA */ 328204f2cbe3SMel Gorman if (iter_vma == vma) 328304f2cbe3SMel Gorman continue; 328404f2cbe3SMel Gorman 328504f2cbe3SMel Gorman /* 32862f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 32872f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 32882f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 32892f84a899SMel Gorman */ 32902f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 32912f84a899SMel Gorman continue; 32922f84a899SMel Gorman 32932f84a899SMel Gorman /* 329404f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 329504f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 329604f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 329704f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 329804f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 329904f2cbe3SMel Gorman */ 330004f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 330124669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 330224669e58SAneesh Kumar K.V address + huge_page_size(h), page); 330304f2cbe3SMel Gorman } 330483cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 330504f2cbe3SMel Gorman } 330604f2cbe3SMel Gorman 33070fe6e20bSNaoya Horiguchi /* 33080fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 3309ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 3310ef009b25SMichal Hocko * cannot race with other handlers or page migration. 3311ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 33120fe6e20bSNaoya Horiguchi */ 33131e8f889bSDavid Gibson static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 331404f2cbe3SMel Gorman unsigned long address, pte_t *ptep, pte_t pte, 3315cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 33161e8f889bSDavid Gibson { 3317a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 33181e8f889bSDavid Gibson struct page *old_page, *new_page; 3319ad4404a2SDavidlohr Bueso int ret = 0, outside_reserve = 0; 33202ec74c3eSSagi Grimberg unsigned long mmun_start; /* For mmu_notifiers */ 33212ec74c3eSSagi Grimberg unsigned long mmun_end; /* For mmu_notifiers */ 33221e8f889bSDavid Gibson 33231e8f889bSDavid Gibson old_page = pte_page(pte); 33241e8f889bSDavid Gibson 332504f2cbe3SMel Gorman retry_avoidcopy: 33261e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 33271e8f889bSDavid Gibson * and just make the page writable */ 332837a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 33290fe6e20bSNaoya Horiguchi page_move_anon_rmap(old_page, vma, address); 33301e8f889bSDavid Gibson set_huge_ptep_writable(vma, address, ptep); 333183c54070SNick Piggin return 0; 33321e8f889bSDavid Gibson } 33331e8f889bSDavid Gibson 333404f2cbe3SMel Gorman /* 333504f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 333604f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 333704f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 333804f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 333904f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 334004f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 334104f2cbe3SMel Gorman * of the full address range. 334204f2cbe3SMel Gorman */ 33435944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 334404f2cbe3SMel Gorman old_page != pagecache_page) 334504f2cbe3SMel Gorman outside_reserve = 1; 334604f2cbe3SMel Gorman 334709cbfeafSKirill A. Shutemov get_page(old_page); 3348b76c8cfbSLarry Woodman 3349ad4404a2SDavidlohr Bueso /* 3350ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 3351ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 3352ad4404a2SDavidlohr Bueso */ 3353cb900f41SKirill A. Shutemov spin_unlock(ptl); 335404f2cbe3SMel Gorman new_page = alloc_huge_page(vma, address, outside_reserve); 33551e8f889bSDavid Gibson 33562fc39cecSAdam Litke if (IS_ERR(new_page)) { 335704f2cbe3SMel Gorman /* 335804f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 335904f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 336004f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 336104f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 336204f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 336304f2cbe3SMel Gorman */ 336404f2cbe3SMel Gorman if (outside_reserve) { 336509cbfeafSKirill A. Shutemov put_page(old_page); 336604f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 33672f4612afSDavidlohr Bueso unmap_ref_private(mm, vma, old_page, address); 336804f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 3369cb900f41SKirill A. Shutemov spin_lock(ptl); 3370a734bcc8SHillf Danton ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 3371a9af0c5dSNaoya Horiguchi if (likely(ptep && 3372a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 337304f2cbe3SMel Gorman goto retry_avoidcopy; 3374a734bcc8SHillf Danton /* 3375cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 3376cb900f41SKirill A. Shutemov * lock, and our job is done. 3377a734bcc8SHillf Danton */ 3378a734bcc8SHillf Danton return 0; 337904f2cbe3SMel Gorman } 338004f2cbe3SMel Gorman 3381ad4404a2SDavidlohr Bueso ret = (PTR_ERR(new_page) == -ENOMEM) ? 3382ad4404a2SDavidlohr Bueso VM_FAULT_OOM : VM_FAULT_SIGBUS; 3383ad4404a2SDavidlohr Bueso goto out_release_old; 33841e8f889bSDavid Gibson } 33851e8f889bSDavid Gibson 33860fe6e20bSNaoya Horiguchi /* 33870fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 33880fe6e20bSNaoya Horiguchi * anon_vma prepared. 33890fe6e20bSNaoya Horiguchi */ 339044e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 3391ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 3392ad4404a2SDavidlohr Bueso goto out_release_all; 339344e2aa93SDean Nelson } 33940fe6e20bSNaoya Horiguchi 339547ad8475SAndrea Arcangeli copy_user_huge_page(new_page, old_page, address, vma, 339647ad8475SAndrea Arcangeli pages_per_huge_page(h)); 33970ed361deSNick Piggin __SetPageUptodate(new_page); 3398bcc54222SNaoya Horiguchi set_page_huge_active(new_page); 33991e8f889bSDavid Gibson 34002ec74c3eSSagi Grimberg mmun_start = address & huge_page_mask(h); 34012ec74c3eSSagi Grimberg mmun_end = mmun_start + huge_page_size(h); 34022ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 3403ad4404a2SDavidlohr Bueso 3404b76c8cfbSLarry Woodman /* 3405cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 3406b76c8cfbSLarry Woodman * before the page tables are altered 3407b76c8cfbSLarry Woodman */ 3408cb900f41SKirill A. Shutemov spin_lock(ptl); 3409a5516438SAndi Kleen ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 3410a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 341107443a85SJoonsoo Kim ClearPagePrivate(new_page); 341207443a85SJoonsoo Kim 34131e8f889bSDavid Gibson /* Break COW */ 34148fe627ecSGerald Schaefer huge_ptep_clear_flush(vma, address, ptep); 341534ee645eSJoerg Roedel mmu_notifier_invalidate_range(mm, mmun_start, mmun_end); 34161e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, 34171e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 3418d281ee61SKirill A. Shutemov page_remove_rmap(old_page, true); 3419cd67f0d2SNaoya Horiguchi hugepage_add_new_anon_rmap(new_page, vma, address); 34201e8f889bSDavid Gibson /* Make the old page be freed below */ 34211e8f889bSDavid Gibson new_page = old_page; 34221e8f889bSDavid Gibson } 3423cb900f41SKirill A. Shutemov spin_unlock(ptl); 34242ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 3425ad4404a2SDavidlohr Bueso out_release_all: 342609cbfeafSKirill A. Shutemov put_page(new_page); 3427ad4404a2SDavidlohr Bueso out_release_old: 342809cbfeafSKirill A. Shutemov put_page(old_page); 34298312034fSJoonsoo Kim 3430ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 3431ad4404a2SDavidlohr Bueso return ret; 34321e8f889bSDavid Gibson } 34331e8f889bSDavid Gibson 343404f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 3435a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 3436a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 343704f2cbe3SMel Gorman { 343804f2cbe3SMel Gorman struct address_space *mapping; 3439e7c4b0bfSAndy Whitcroft pgoff_t idx; 344004f2cbe3SMel Gorman 344104f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 3442a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 344304f2cbe3SMel Gorman 344404f2cbe3SMel Gorman return find_lock_page(mapping, idx); 344504f2cbe3SMel Gorman } 344604f2cbe3SMel Gorman 34473ae77f43SHugh Dickins /* 34483ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 34493ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 34503ae77f43SHugh Dickins */ 34513ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 34522a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 34532a15efc9SHugh Dickins { 34542a15efc9SHugh Dickins struct address_space *mapping; 34552a15efc9SHugh Dickins pgoff_t idx; 34562a15efc9SHugh Dickins struct page *page; 34572a15efc9SHugh Dickins 34582a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 34592a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 34602a15efc9SHugh Dickins 34612a15efc9SHugh Dickins page = find_get_page(mapping, idx); 34622a15efc9SHugh Dickins if (page) 34632a15efc9SHugh Dickins put_page(page); 34642a15efc9SHugh Dickins return page != NULL; 34652a15efc9SHugh Dickins } 34662a15efc9SHugh Dickins 3467ab76ad54SMike Kravetz int huge_add_to_page_cache(struct page *page, struct address_space *mapping, 3468ab76ad54SMike Kravetz pgoff_t idx) 3469ab76ad54SMike Kravetz { 3470ab76ad54SMike Kravetz struct inode *inode = mapping->host; 3471ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 3472ab76ad54SMike Kravetz int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 3473ab76ad54SMike Kravetz 3474ab76ad54SMike Kravetz if (err) 3475ab76ad54SMike Kravetz return err; 3476ab76ad54SMike Kravetz ClearPagePrivate(page); 3477ab76ad54SMike Kravetz 3478ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 3479ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 3480ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 3481ab76ad54SMike Kravetz return 0; 3482ab76ad54SMike Kravetz } 3483ab76ad54SMike Kravetz 3484a1ed3ddaSRobert P. J. Day static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, 34858382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 3486788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 3487ac9b9c66SHugh Dickins { 3488a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3489ac9b9c66SHugh Dickins int ret = VM_FAULT_SIGBUS; 3490409eb8c2SHillf Danton int anon_rmap = 0; 34914c887265SAdam Litke unsigned long size; 34924c887265SAdam Litke struct page *page; 34931e8f889bSDavid Gibson pte_t new_pte; 3494cb900f41SKirill A. Shutemov spinlock_t *ptl; 34954c887265SAdam Litke 349604f2cbe3SMel Gorman /* 349704f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 349804f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 349925985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 350004f2cbe3SMel Gorman */ 350104f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 3502910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 350304f2cbe3SMel Gorman current->pid); 350404f2cbe3SMel Gorman return ret; 350504f2cbe3SMel Gorman } 350604f2cbe3SMel Gorman 35074c887265SAdam Litke /* 35084c887265SAdam Litke * Use page lock to guard against racing truncation 35094c887265SAdam Litke * before we get page_table_lock. 35104c887265SAdam Litke */ 35116bda666aSChristoph Lameter retry: 35126bda666aSChristoph Lameter page = find_lock_page(mapping, idx); 35136bda666aSChristoph Lameter if (!page) { 3514a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 3515ebed4bfcSHugh Dickins if (idx >= size) 3516ebed4bfcSHugh Dickins goto out; 351704f2cbe3SMel Gorman page = alloc_huge_page(vma, address, 0); 35182fc39cecSAdam Litke if (IS_ERR(page)) { 351976dcee75SAneesh Kumar K.V ret = PTR_ERR(page); 352076dcee75SAneesh Kumar K.V if (ret == -ENOMEM) 352176dcee75SAneesh Kumar K.V ret = VM_FAULT_OOM; 352276dcee75SAneesh Kumar K.V else 352376dcee75SAneesh Kumar K.V ret = VM_FAULT_SIGBUS; 35246bda666aSChristoph Lameter goto out; 35256bda666aSChristoph Lameter } 352647ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 35270ed361deSNick Piggin __SetPageUptodate(page); 3528bcc54222SNaoya Horiguchi set_page_huge_active(page); 3529ac9b9c66SHugh Dickins 3530f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 3531ab76ad54SMike Kravetz int err = huge_add_to_page_cache(page, mapping, idx); 35326bda666aSChristoph Lameter if (err) { 35336bda666aSChristoph Lameter put_page(page); 35346bda666aSChristoph Lameter if (err == -EEXIST) 35356bda666aSChristoph Lameter goto retry; 35366bda666aSChristoph Lameter goto out; 35376bda666aSChristoph Lameter } 353823be7468SMel Gorman } else { 35396bda666aSChristoph Lameter lock_page(page); 35400fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 35410fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 35420fe6e20bSNaoya Horiguchi goto backout_unlocked; 354323be7468SMel Gorman } 3544409eb8c2SHillf Danton anon_rmap = 1; 35450fe6e20bSNaoya Horiguchi } 35460fe6e20bSNaoya Horiguchi } else { 354757303d80SAndy Whitcroft /* 3548998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 3549998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 3550998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 3551fd6a03edSNaoya Horiguchi */ 3552fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 3553aa50d3a7SAndi Kleen ret = VM_FAULT_HWPOISON | 3554972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 3555fd6a03edSNaoya Horiguchi goto backout_unlocked; 35566bda666aSChristoph Lameter } 3557998b4382SNaoya Horiguchi } 35581e8f889bSDavid Gibson 355957303d80SAndy Whitcroft /* 356057303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 356157303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 356257303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 356357303d80SAndy Whitcroft * the spinlock. 356457303d80SAndy Whitcroft */ 35655e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 35662b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 35672b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 35682b26736cSAndy Whitcroft goto backout_unlocked; 35692b26736cSAndy Whitcroft } 35705e911373SMike Kravetz /* Just decrements count, does not deallocate */ 3571feba16e2SMike Kravetz vma_end_reservation(h, vma, address); 35725e911373SMike Kravetz } 357357303d80SAndy Whitcroft 3574cb900f41SKirill A. Shutemov ptl = huge_pte_lockptr(h, mm, ptep); 3575cb900f41SKirill A. Shutemov spin_lock(ptl); 3576a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 35774c887265SAdam Litke if (idx >= size) 35784c887265SAdam Litke goto backout; 35794c887265SAdam Litke 358083c54070SNick Piggin ret = 0; 35817f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 35824c887265SAdam Litke goto backout; 35834c887265SAdam Litke 358407443a85SJoonsoo Kim if (anon_rmap) { 358507443a85SJoonsoo Kim ClearPagePrivate(page); 3586409eb8c2SHillf Danton hugepage_add_new_anon_rmap(page, vma, address); 3587ac714904SChoi Gi-yong } else 358853f9263bSKirill A. Shutemov page_dup_rmap(page, true); 35891e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 35901e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 35911e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, new_pte); 35921e8f889bSDavid Gibson 35935d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 3594788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 35951e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 3596cb900f41SKirill A. Shutemov ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl); 35971e8f889bSDavid Gibson } 35981e8f889bSDavid Gibson 3599cb900f41SKirill A. Shutemov spin_unlock(ptl); 36004c887265SAdam Litke unlock_page(page); 36014c887265SAdam Litke out: 3602ac9b9c66SHugh Dickins return ret; 36034c887265SAdam Litke 36044c887265SAdam Litke backout: 3605cb900f41SKirill A. Shutemov spin_unlock(ptl); 36062b26736cSAndy Whitcroft backout_unlocked: 36074c887265SAdam Litke unlock_page(page); 36084c887265SAdam Litke put_page(page); 36094c887265SAdam Litke goto out; 3610ac9b9c66SHugh Dickins } 3611ac9b9c66SHugh Dickins 36128382d914SDavidlohr Bueso #ifdef CONFIG_SMP 3613c672c7f2SMike Kravetz u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, 36148382d914SDavidlohr Bueso struct vm_area_struct *vma, 36158382d914SDavidlohr Bueso struct address_space *mapping, 36168382d914SDavidlohr Bueso pgoff_t idx, unsigned long address) 36178382d914SDavidlohr Bueso { 36188382d914SDavidlohr Bueso unsigned long key[2]; 36198382d914SDavidlohr Bueso u32 hash; 36208382d914SDavidlohr Bueso 36218382d914SDavidlohr Bueso if (vma->vm_flags & VM_SHARED) { 36228382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 36238382d914SDavidlohr Bueso key[1] = idx; 36248382d914SDavidlohr Bueso } else { 36258382d914SDavidlohr Bueso key[0] = (unsigned long) mm; 36268382d914SDavidlohr Bueso key[1] = address >> huge_page_shift(h); 36278382d914SDavidlohr Bueso } 36288382d914SDavidlohr Bueso 36298382d914SDavidlohr Bueso hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0); 36308382d914SDavidlohr Bueso 36318382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 36328382d914SDavidlohr Bueso } 36338382d914SDavidlohr Bueso #else 36348382d914SDavidlohr Bueso /* 36358382d914SDavidlohr Bueso * For uniprocesor systems we always use a single mutex, so just 36368382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 36378382d914SDavidlohr Bueso */ 3638c672c7f2SMike Kravetz u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, 36398382d914SDavidlohr Bueso struct vm_area_struct *vma, 36408382d914SDavidlohr Bueso struct address_space *mapping, 36418382d914SDavidlohr Bueso pgoff_t idx, unsigned long address) 36428382d914SDavidlohr Bueso { 36438382d914SDavidlohr Bueso return 0; 36448382d914SDavidlohr Bueso } 36458382d914SDavidlohr Bueso #endif 36468382d914SDavidlohr Bueso 364786e5216fSAdam Litke int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 3648788c7df4SHugh Dickins unsigned long address, unsigned int flags) 364986e5216fSAdam Litke { 36508382d914SDavidlohr Bueso pte_t *ptep, entry; 3651cb900f41SKirill A. Shutemov spinlock_t *ptl; 36521e8f889bSDavid Gibson int ret; 36538382d914SDavidlohr Bueso u32 hash; 36548382d914SDavidlohr Bueso pgoff_t idx; 36550fe6e20bSNaoya Horiguchi struct page *page = NULL; 365657303d80SAndy Whitcroft struct page *pagecache_page = NULL; 3657a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 36588382d914SDavidlohr Bueso struct address_space *mapping; 36590f792cf9SNaoya Horiguchi int need_wait_lock = 0; 366086e5216fSAdam Litke 36611e16a539SKAMEZAWA Hiroyuki address &= huge_page_mask(h); 36621e16a539SKAMEZAWA Hiroyuki 3663fd6a03edSNaoya Horiguchi ptep = huge_pte_offset(mm, address); 3664fd6a03edSNaoya Horiguchi if (ptep) { 3665fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 3666290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 3667cb900f41SKirill A. Shutemov migration_entry_wait_huge(vma, mm, ptep); 3668290408d4SNaoya Horiguchi return 0; 3669290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 3670aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 3671972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 36720d777df5SNaoya Horiguchi } else { 3673a5516438SAndi Kleen ptep = huge_pte_alloc(mm, address, huge_page_size(h)); 367486e5216fSAdam Litke if (!ptep) 367586e5216fSAdam Litke return VM_FAULT_OOM; 36760d777df5SNaoya Horiguchi } 367786e5216fSAdam Litke 36788382d914SDavidlohr Bueso mapping = vma->vm_file->f_mapping; 36798382d914SDavidlohr Bueso idx = vma_hugecache_offset(h, vma, address); 36808382d914SDavidlohr Bueso 36813935baa9SDavid Gibson /* 36823935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 36833935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 36843935baa9SDavid Gibson * the same page in the page cache. 36853935baa9SDavid Gibson */ 3686c672c7f2SMike Kravetz hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address); 3687c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 36888382d914SDavidlohr Bueso 36897f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 36907f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 36918382d914SDavidlohr Bueso ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 3692b4d1d99fSDavid Gibson goto out_mutex; 36933935baa9SDavid Gibson } 369486e5216fSAdam Litke 369583c54070SNick Piggin ret = 0; 36961e8f889bSDavid Gibson 369757303d80SAndy Whitcroft /* 36980f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 36990f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 37000f792cf9SNaoya Horiguchi * a active hugepage in pagecache. This goto expects the 2nd page fault, 37010f792cf9SNaoya Horiguchi * and is_hugetlb_entry_(migration|hwpoisoned) check will properly 37020f792cf9SNaoya Horiguchi * handle it. 37030f792cf9SNaoya Horiguchi */ 37040f792cf9SNaoya Horiguchi if (!pte_present(entry)) 37050f792cf9SNaoya Horiguchi goto out_mutex; 37060f792cf9SNaoya Horiguchi 37070f792cf9SNaoya Horiguchi /* 370857303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 370957303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 371057303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 371157303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 371257303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 371357303d80SAndy Whitcroft * consumed. 371457303d80SAndy Whitcroft */ 3715106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 37162b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 37172b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 3718b4d1d99fSDavid Gibson goto out_mutex; 37192b26736cSAndy Whitcroft } 37205e911373SMike Kravetz /* Just decrements count, does not deallocate */ 3721feba16e2SMike Kravetz vma_end_reservation(h, vma, address); 372257303d80SAndy Whitcroft 3723f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 372457303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 372557303d80SAndy Whitcroft vma, address); 372657303d80SAndy Whitcroft } 372757303d80SAndy Whitcroft 37280f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 37290fe6e20bSNaoya Horiguchi 37301e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 3731b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 3732cb900f41SKirill A. Shutemov goto out_ptl; 3733b4d1d99fSDavid Gibson 37340f792cf9SNaoya Horiguchi /* 37350f792cf9SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 37360f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 37370f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 37380f792cf9SNaoya Horiguchi */ 37390f792cf9SNaoya Horiguchi page = pte_page(entry); 37400f792cf9SNaoya Horiguchi if (page != pagecache_page) 37410f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 37420f792cf9SNaoya Horiguchi need_wait_lock = 1; 37430f792cf9SNaoya Horiguchi goto out_ptl; 37440f792cf9SNaoya Horiguchi } 37450f792cf9SNaoya Horiguchi 37460f792cf9SNaoya Horiguchi get_page(page); 3747b4d1d99fSDavid Gibson 3748788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 3749106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 375057303d80SAndy Whitcroft ret = hugetlb_cow(mm, vma, address, ptep, entry, 3751cb900f41SKirill A. Shutemov pagecache_page, ptl); 37520f792cf9SNaoya Horiguchi goto out_put_page; 3753b4d1d99fSDavid Gibson } 3754106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 3755b4d1d99fSDavid Gibson } 3756b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 3757788c7df4SHugh Dickins if (huge_ptep_set_access_flags(vma, address, ptep, entry, 3758788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 37594b3073e1SRussell King update_mmu_cache(vma, address, ptep); 37600f792cf9SNaoya Horiguchi out_put_page: 37610f792cf9SNaoya Horiguchi if (page != pagecache_page) 37620f792cf9SNaoya Horiguchi unlock_page(page); 37630f792cf9SNaoya Horiguchi put_page(page); 3764cb900f41SKirill A. Shutemov out_ptl: 3765cb900f41SKirill A. Shutemov spin_unlock(ptl); 376657303d80SAndy Whitcroft 376757303d80SAndy Whitcroft if (pagecache_page) { 376857303d80SAndy Whitcroft unlock_page(pagecache_page); 376957303d80SAndy Whitcroft put_page(pagecache_page); 377057303d80SAndy Whitcroft } 3771b4d1d99fSDavid Gibson out_mutex: 3772c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 37730f792cf9SNaoya Horiguchi /* 37740f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 37750f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 37760f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 37770f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 37780f792cf9SNaoya Horiguchi * here without taking refcount. 37790f792cf9SNaoya Horiguchi */ 37800f792cf9SNaoya Horiguchi if (need_wait_lock) 37810f792cf9SNaoya Horiguchi wait_on_page_locked(page); 37821e8f889bSDavid Gibson return ret; 378386e5216fSAdam Litke } 378486e5216fSAdam Litke 378528a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 378663551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 378728a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 378828a35716SMichel Lespinasse long i, unsigned int flags) 378963551ae0SDavid Gibson { 3790d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 3791d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 379228a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 3793a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 379463551ae0SDavid Gibson 379563551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 379663551ae0SDavid Gibson pte_t *pte; 3797cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 37982a15efc9SHugh Dickins int absent; 379963551ae0SDavid Gibson struct page *page; 380063551ae0SDavid Gibson 38014c887265SAdam Litke /* 380202057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 380302057967SDavid Rientjes * potentially allocating memory. 380402057967SDavid Rientjes */ 380502057967SDavid Rientjes if (unlikely(fatal_signal_pending(current))) { 380602057967SDavid Rientjes remainder = 0; 380702057967SDavid Rientjes break; 380802057967SDavid Rientjes } 380902057967SDavid Rientjes 381002057967SDavid Rientjes /* 38114c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 38122a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 38134c887265SAdam Litke * first, for the page indexing below to work. 3814cb900f41SKirill A. Shutemov * 3815cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 38164c887265SAdam Litke */ 3817a5516438SAndi Kleen pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); 3818cb900f41SKirill A. Shutemov if (pte) 3819cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 38202a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 382163551ae0SDavid Gibson 38222a15efc9SHugh Dickins /* 38232a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 38243ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 38253ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 38263ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 38273ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 38282a15efc9SHugh Dickins */ 38293ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 38303ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 3831cb900f41SKirill A. Shutemov if (pte) 3832cb900f41SKirill A. Shutemov spin_unlock(ptl); 38332a15efc9SHugh Dickins remainder = 0; 38342a15efc9SHugh Dickins break; 38352a15efc9SHugh Dickins } 38362a15efc9SHugh Dickins 38379cc3a5bdSNaoya Horiguchi /* 38389cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 38399cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 38409cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 38419cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 38429cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 38439cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 38449cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 38459cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 38469cc3a5bdSNaoya Horiguchi */ 38479cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 3848106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 3849106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 38504c887265SAdam Litke int ret; 38514c887265SAdam Litke 3852cb900f41SKirill A. Shutemov if (pte) 3853cb900f41SKirill A. Shutemov spin_unlock(ptl); 38542a15efc9SHugh Dickins ret = hugetlb_fault(mm, vma, vaddr, 38552a15efc9SHugh Dickins (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); 3856a89182c7SAdam Litke if (!(ret & VM_FAULT_ERROR)) 38574c887265SAdam Litke continue; 38584c887265SAdam Litke 38591c59827dSHugh Dickins remainder = 0; 38601c59827dSHugh Dickins break; 38611c59827dSHugh Dickins } 386263551ae0SDavid Gibson 3863a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 38647f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 3865d5d4b0aaSChen, Kenneth W same_page: 3866d6692183SChen, Kenneth W if (pages) { 386769d177c2SAndy Whitcroft pages[i] = mem_map_offset(page, pfn_offset); 3868ddc58f27SKirill A. Shutemov get_page(pages[i]); 3869d6692183SChen, Kenneth W } 387063551ae0SDavid Gibson 387163551ae0SDavid Gibson if (vmas) 387263551ae0SDavid Gibson vmas[i] = vma; 387363551ae0SDavid Gibson 387463551ae0SDavid Gibson vaddr += PAGE_SIZE; 3875d5d4b0aaSChen, Kenneth W ++pfn_offset; 387663551ae0SDavid Gibson --remainder; 387763551ae0SDavid Gibson ++i; 3878d5d4b0aaSChen, Kenneth W if (vaddr < vma->vm_end && remainder && 3879a5516438SAndi Kleen pfn_offset < pages_per_huge_page(h)) { 3880d5d4b0aaSChen, Kenneth W /* 3881d5d4b0aaSChen, Kenneth W * We use pfn_offset to avoid touching the pageframes 3882d5d4b0aaSChen, Kenneth W * of this compound page. 3883d5d4b0aaSChen, Kenneth W */ 3884d5d4b0aaSChen, Kenneth W goto same_page; 3885d5d4b0aaSChen, Kenneth W } 3886cb900f41SKirill A. Shutemov spin_unlock(ptl); 388763551ae0SDavid Gibson } 388828a35716SMichel Lespinasse *nr_pages = remainder; 388963551ae0SDavid Gibson *position = vaddr; 389063551ae0SDavid Gibson 38912a15efc9SHugh Dickins return i ? i : -EFAULT; 389263551ae0SDavid Gibson } 38938f860591SZhang, Yanmin 38947da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 38958f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 38968f860591SZhang, Yanmin { 38978f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 38988f860591SZhang, Yanmin unsigned long start = address; 38998f860591SZhang, Yanmin pte_t *ptep; 39008f860591SZhang, Yanmin pte_t pte; 3901a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 39027da4d641SPeter Zijlstra unsigned long pages = 0; 39038f860591SZhang, Yanmin 39048f860591SZhang, Yanmin BUG_ON(address >= end); 39058f860591SZhang, Yanmin flush_cache_range(vma, address, end); 39068f860591SZhang, Yanmin 3907a5338093SRik van Riel mmu_notifier_invalidate_range_start(mm, start, end); 390883cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 3909a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 3910cb900f41SKirill A. Shutemov spinlock_t *ptl; 39118f860591SZhang, Yanmin ptep = huge_pte_offset(mm, address); 39128f860591SZhang, Yanmin if (!ptep) 39138f860591SZhang, Yanmin continue; 3914cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 39157da4d641SPeter Zijlstra if (huge_pmd_unshare(mm, &address, ptep)) { 39167da4d641SPeter Zijlstra pages++; 3917cb900f41SKirill A. Shutemov spin_unlock(ptl); 391839dde65cSChen, Kenneth W continue; 39197da4d641SPeter Zijlstra } 3920a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 3921a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 3922a8bda28dSNaoya Horiguchi spin_unlock(ptl); 3923a8bda28dSNaoya Horiguchi continue; 3924a8bda28dSNaoya Horiguchi } 3925a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 3926a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 3927a8bda28dSNaoya Horiguchi 3928a8bda28dSNaoya Horiguchi if (is_write_migration_entry(entry)) { 3929a8bda28dSNaoya Horiguchi pte_t newpte; 3930a8bda28dSNaoya Horiguchi 3931a8bda28dSNaoya Horiguchi make_migration_entry_read(&entry); 3932a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 3933a8bda28dSNaoya Horiguchi set_huge_pte_at(mm, address, ptep, newpte); 3934a8bda28dSNaoya Horiguchi pages++; 3935a8bda28dSNaoya Horiguchi } 3936a8bda28dSNaoya Horiguchi spin_unlock(ptl); 3937a8bda28dSNaoya Horiguchi continue; 3938a8bda28dSNaoya Horiguchi } 3939a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 39408f860591SZhang, Yanmin pte = huge_ptep_get_and_clear(mm, address, ptep); 3941106c992aSGerald Schaefer pte = pte_mkhuge(huge_pte_modify(pte, newprot)); 3942be7517d6STony Lu pte = arch_make_huge_pte(pte, vma, NULL, 0); 39438f860591SZhang, Yanmin set_huge_pte_at(mm, address, ptep, pte); 39447da4d641SPeter Zijlstra pages++; 39458f860591SZhang, Yanmin } 3946cb900f41SKirill A. Shutemov spin_unlock(ptl); 39478f860591SZhang, Yanmin } 3948d833352aSMel Gorman /* 3949c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 3950d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 3951c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 3952d833352aSMel Gorman * and that page table be reused and filled with junk. 3953d833352aSMel Gorman */ 39548f860591SZhang, Yanmin flush_tlb_range(vma, start, end); 395534ee645eSJoerg Roedel mmu_notifier_invalidate_range(mm, start, end); 395683cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 3957a5338093SRik van Riel mmu_notifier_invalidate_range_end(mm, start, end); 39587da4d641SPeter Zijlstra 39597da4d641SPeter Zijlstra return pages << h->order; 39608f860591SZhang, Yanmin } 39618f860591SZhang, Yanmin 3962a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 3963a1e78772SMel Gorman long from, long to, 39645a6fe125SMel Gorman struct vm_area_struct *vma, 3965ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 3966e4e574b7SAdam Litke { 396717c9d12eSMel Gorman long ret, chg; 3968a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 396990481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 39709119a41eSJoonsoo Kim struct resv_map *resv_map; 39711c5ecae3SMike Kravetz long gbl_reserve; 3972e4e574b7SAdam Litke 3973a1e78772SMel Gorman /* 397417c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 397517c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 397690481622SDavid Gibson * without using reserves 397717c9d12eSMel Gorman */ 3978ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 397917c9d12eSMel Gorman return 0; 398017c9d12eSMel Gorman 398117c9d12eSMel Gorman /* 3982a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 3983a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 3984a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 3985a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 3986a1e78772SMel Gorman */ 39879119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 39884e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 39899119a41eSJoonsoo Kim 39901406ec9bSJoonsoo Kim chg = region_chg(resv_map, from, to); 39919119a41eSJoonsoo Kim 39929119a41eSJoonsoo Kim } else { 39939119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 39945a6fe125SMel Gorman if (!resv_map) 39955a6fe125SMel Gorman return -ENOMEM; 39965a6fe125SMel Gorman 399717c9d12eSMel Gorman chg = to - from; 399817c9d12eSMel Gorman 39995a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 40005a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 40015a6fe125SMel Gorman } 40025a6fe125SMel Gorman 4003c50ac050SDave Hansen if (chg < 0) { 4004c50ac050SDave Hansen ret = chg; 4005c50ac050SDave Hansen goto out_err; 4006c50ac050SDave Hansen } 400717c9d12eSMel Gorman 40081c5ecae3SMike Kravetz /* 40091c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 40101c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 40111c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 40121c5ecae3SMike Kravetz */ 40131c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 40141c5ecae3SMike Kravetz if (gbl_reserve < 0) { 4015c50ac050SDave Hansen ret = -ENOSPC; 4016c50ac050SDave Hansen goto out_err; 4017c50ac050SDave Hansen } 401817c9d12eSMel Gorman 401917c9d12eSMel Gorman /* 402017c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 402190481622SDavid Gibson * Hand the pages back to the subpool if there are not 402217c9d12eSMel Gorman */ 40231c5ecae3SMike Kravetz ret = hugetlb_acct_memory(h, gbl_reserve); 402417c9d12eSMel Gorman if (ret < 0) { 40251c5ecae3SMike Kravetz /* put back original number of pages, chg */ 40261c5ecae3SMike Kravetz (void)hugepage_subpool_put_pages(spool, chg); 4027c50ac050SDave Hansen goto out_err; 402817c9d12eSMel Gorman } 402917c9d12eSMel Gorman 403017c9d12eSMel Gorman /* 403117c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 403217c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 403317c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 403417c9d12eSMel Gorman * the reservation was and the page cache tells how much of 403517c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 403617c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 403717c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 403817c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 403917c9d12eSMel Gorman * else has to be done for private mappings here 404017c9d12eSMel Gorman */ 404133039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 404233039678SMike Kravetz long add = region_add(resv_map, from, to); 404333039678SMike Kravetz 404433039678SMike Kravetz if (unlikely(chg > add)) { 404533039678SMike Kravetz /* 404633039678SMike Kravetz * pages in this range were added to the reserve 404733039678SMike Kravetz * map between region_chg and region_add. This 404833039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 404933039678SMike Kravetz * the subpool and reserve counts modified above 405033039678SMike Kravetz * based on the difference. 405133039678SMike Kravetz */ 405233039678SMike Kravetz long rsv_adjust; 405333039678SMike Kravetz 405433039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 405533039678SMike Kravetz chg - add); 405633039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 405733039678SMike Kravetz } 405833039678SMike Kravetz } 4059a43a8c39SChen, Kenneth W return 0; 4060c50ac050SDave Hansen out_err: 40615e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 40625e911373SMike Kravetz region_abort(resv_map, from, to); 4063f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 4064f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 4065c50ac050SDave Hansen return ret; 4066a43a8c39SChen, Kenneth W } 4067a43a8c39SChen, Kenneth W 4068b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 4069b5cec28dSMike Kravetz long freed) 4070a43a8c39SChen, Kenneth W { 4071a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 40724e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 40739119a41eSJoonsoo Kim long chg = 0; 407490481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 40751c5ecae3SMike Kravetz long gbl_reserve; 407645c682a6SKen Chen 4077b5cec28dSMike Kravetz if (resv_map) { 4078b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 4079b5cec28dSMike Kravetz /* 4080b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 4081b5cec28dSMike Kravetz * must be split and another region descriptor can not be 4082b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 4083b5cec28dSMike Kravetz */ 4084b5cec28dSMike Kravetz if (chg < 0) 4085b5cec28dSMike Kravetz return chg; 4086b5cec28dSMike Kravetz } 4087b5cec28dSMike Kravetz 408845c682a6SKen Chen spin_lock(&inode->i_lock); 4089e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 409045c682a6SKen Chen spin_unlock(&inode->i_lock); 409145c682a6SKen Chen 40921c5ecae3SMike Kravetz /* 40931c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 40941c5ecae3SMike Kravetz * reservations to be released may be adjusted. 40951c5ecae3SMike Kravetz */ 40961c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 40971c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 4098b5cec28dSMike Kravetz 4099b5cec28dSMike Kravetz return 0; 4100a43a8c39SChen, Kenneth W } 410193f70f90SNaoya Horiguchi 41023212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 41033212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 41043212b535SSteve Capper struct vm_area_struct *vma, 41053212b535SSteve Capper unsigned long addr, pgoff_t idx) 41063212b535SSteve Capper { 41073212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 41083212b535SSteve Capper svma->vm_start; 41093212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 41103212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 41113212b535SSteve Capper 41123212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 4113de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 4114de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 41153212b535SSteve Capper 41163212b535SSteve Capper /* 41173212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 41183212b535SSteve Capper * page table page. 41193212b535SSteve Capper */ 41203212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 41213212b535SSteve Capper vm_flags != svm_flags || 41223212b535SSteve Capper sbase < svma->vm_start || svma->vm_end < s_end) 41233212b535SSteve Capper return 0; 41243212b535SSteve Capper 41253212b535SSteve Capper return saddr; 41263212b535SSteve Capper } 41273212b535SSteve Capper 412831aafb45SNicholas Krause static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) 41293212b535SSteve Capper { 41303212b535SSteve Capper unsigned long base = addr & PUD_MASK; 41313212b535SSteve Capper unsigned long end = base + PUD_SIZE; 41323212b535SSteve Capper 41333212b535SSteve Capper /* 41343212b535SSteve Capper * check on proper vm_flags and page table alignment 41353212b535SSteve Capper */ 41363212b535SSteve Capper if (vma->vm_flags & VM_MAYSHARE && 41373212b535SSteve Capper vma->vm_start <= base && end <= vma->vm_end) 413831aafb45SNicholas Krause return true; 413931aafb45SNicholas Krause return false; 41403212b535SSteve Capper } 41413212b535SSteve Capper 41423212b535SSteve Capper /* 41433212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 41443212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 41453212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 41463212b535SSteve Capper * code much cleaner. pmd allocation is essential for the shared case because 4147c8c06efaSDavidlohr Bueso * pud has to be populated inside the same i_mmap_rwsem section - otherwise 41483212b535SSteve Capper * racing tasks could either miss the sharing (see huge_pte_offset) or select a 41493212b535SSteve Capper * bad pmd for sharing. 41503212b535SSteve Capper */ 41513212b535SSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 41523212b535SSteve Capper { 41533212b535SSteve Capper struct vm_area_struct *vma = find_vma(mm, addr); 41543212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 41553212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 41563212b535SSteve Capper vma->vm_pgoff; 41573212b535SSteve Capper struct vm_area_struct *svma; 41583212b535SSteve Capper unsigned long saddr; 41593212b535SSteve Capper pte_t *spte = NULL; 41603212b535SSteve Capper pte_t *pte; 4161cb900f41SKirill A. Shutemov spinlock_t *ptl; 41623212b535SSteve Capper 41633212b535SSteve Capper if (!vma_shareable(vma, addr)) 41643212b535SSteve Capper return (pte_t *)pmd_alloc(mm, pud, addr); 41653212b535SSteve Capper 416683cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 41673212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 41683212b535SSteve Capper if (svma == vma) 41693212b535SSteve Capper continue; 41703212b535SSteve Capper 41713212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 41723212b535SSteve Capper if (saddr) { 41733212b535SSteve Capper spte = huge_pte_offset(svma->vm_mm, saddr); 41743212b535SSteve Capper if (spte) { 4175dc6c9a35SKirill A. Shutemov mm_inc_nr_pmds(mm); 41763212b535SSteve Capper get_page(virt_to_page(spte)); 41773212b535SSteve Capper break; 41783212b535SSteve Capper } 41793212b535SSteve Capper } 41803212b535SSteve Capper } 41813212b535SSteve Capper 41823212b535SSteve Capper if (!spte) 41833212b535SSteve Capper goto out; 41843212b535SSteve Capper 4185cb900f41SKirill A. Shutemov ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte); 4186cb900f41SKirill A. Shutemov spin_lock(ptl); 4187dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 41883212b535SSteve Capper pud_populate(mm, pud, 41893212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 4190dc6c9a35SKirill A. Shutemov } else { 41913212b535SSteve Capper put_page(virt_to_page(spte)); 4192dc6c9a35SKirill A. Shutemov mm_inc_nr_pmds(mm); 4193dc6c9a35SKirill A. Shutemov } 4194cb900f41SKirill A. Shutemov spin_unlock(ptl); 41953212b535SSteve Capper out: 41963212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 419783cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 41983212b535SSteve Capper return pte; 41993212b535SSteve Capper } 42003212b535SSteve Capper 42013212b535SSteve Capper /* 42023212b535SSteve Capper * unmap huge page backed by shared pte. 42033212b535SSteve Capper * 42043212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 42053212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 42063212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 42073212b535SSteve Capper * 4208cb900f41SKirill A. Shutemov * called with page table lock held. 42093212b535SSteve Capper * 42103212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 42113212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 42123212b535SSteve Capper */ 42133212b535SSteve Capper int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 42143212b535SSteve Capper { 42153212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 42163212b535SSteve Capper pud_t *pud = pud_offset(pgd, *addr); 42173212b535SSteve Capper 42183212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 42193212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 42203212b535SSteve Capper return 0; 42213212b535SSteve Capper 42223212b535SSteve Capper pud_clear(pud); 42233212b535SSteve Capper put_page(virt_to_page(ptep)); 4224dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 42253212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 42263212b535SSteve Capper return 1; 42273212b535SSteve Capper } 42289e5fc74cSSteve Capper #define want_pmd_share() (1) 42299e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 42309e5fc74cSSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 42319e5fc74cSSteve Capper { 42329e5fc74cSSteve Capper return NULL; 42339e5fc74cSSteve Capper } 4234e81f2d22SZhang Zhen 4235e81f2d22SZhang Zhen int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 4236e81f2d22SZhang Zhen { 4237e81f2d22SZhang Zhen return 0; 4238e81f2d22SZhang Zhen } 42399e5fc74cSSteve Capper #define want_pmd_share() (0) 42403212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 42413212b535SSteve Capper 42429e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 42439e5fc74cSSteve Capper pte_t *huge_pte_alloc(struct mm_struct *mm, 42449e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 42459e5fc74cSSteve Capper { 42469e5fc74cSSteve Capper pgd_t *pgd; 42479e5fc74cSSteve Capper pud_t *pud; 42489e5fc74cSSteve Capper pte_t *pte = NULL; 42499e5fc74cSSteve Capper 42509e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 42519e5fc74cSSteve Capper pud = pud_alloc(mm, pgd, addr); 42529e5fc74cSSteve Capper if (pud) { 42539e5fc74cSSteve Capper if (sz == PUD_SIZE) { 42549e5fc74cSSteve Capper pte = (pte_t *)pud; 42559e5fc74cSSteve Capper } else { 42569e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 42579e5fc74cSSteve Capper if (want_pmd_share() && pud_none(*pud)) 42589e5fc74cSSteve Capper pte = huge_pmd_share(mm, addr, pud); 42599e5fc74cSSteve Capper else 42609e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 42619e5fc74cSSteve Capper } 42629e5fc74cSSteve Capper } 42639e5fc74cSSteve Capper BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte)); 42649e5fc74cSSteve Capper 42659e5fc74cSSteve Capper return pte; 42669e5fc74cSSteve Capper } 42679e5fc74cSSteve Capper 42689e5fc74cSSteve Capper pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) 42699e5fc74cSSteve Capper { 42709e5fc74cSSteve Capper pgd_t *pgd; 42719e5fc74cSSteve Capper pud_t *pud; 42729e5fc74cSSteve Capper pmd_t *pmd = NULL; 42739e5fc74cSSteve Capper 42749e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 42759e5fc74cSSteve Capper if (pgd_present(*pgd)) { 42769e5fc74cSSteve Capper pud = pud_offset(pgd, addr); 42779e5fc74cSSteve Capper if (pud_present(*pud)) { 42789e5fc74cSSteve Capper if (pud_huge(*pud)) 42799e5fc74cSSteve Capper return (pte_t *)pud; 42809e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 42819e5fc74cSSteve Capper } 42829e5fc74cSSteve Capper } 42839e5fc74cSSteve Capper return (pte_t *) pmd; 42849e5fc74cSSteve Capper } 42859e5fc74cSSteve Capper 428661f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 428761f77edaSNaoya Horiguchi 428861f77edaSNaoya Horiguchi /* 428961f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 429061f77edaSNaoya Horiguchi * behavior. 429161f77edaSNaoya Horiguchi */ 429261f77edaSNaoya Horiguchi struct page * __weak 429361f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 429461f77edaSNaoya Horiguchi int write) 429561f77edaSNaoya Horiguchi { 429661f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 429761f77edaSNaoya Horiguchi } 429861f77edaSNaoya Horiguchi 429961f77edaSNaoya Horiguchi struct page * __weak 43009e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 4301e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 43029e5fc74cSSteve Capper { 4303e66f17ffSNaoya Horiguchi struct page *page = NULL; 4304e66f17ffSNaoya Horiguchi spinlock_t *ptl; 4305e66f17ffSNaoya Horiguchi retry: 4306e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 4307e66f17ffSNaoya Horiguchi spin_lock(ptl); 4308e66f17ffSNaoya Horiguchi /* 4309e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 4310e66f17ffSNaoya Horiguchi * unmapped from other threads. 4311e66f17ffSNaoya Horiguchi */ 4312e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 4313e66f17ffSNaoya Horiguchi goto out; 4314e66f17ffSNaoya Horiguchi if (pmd_present(*pmd)) { 431597534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 4316e66f17ffSNaoya Horiguchi if (flags & FOLL_GET) 4317e66f17ffSNaoya Horiguchi get_page(page); 4318e66f17ffSNaoya Horiguchi } else { 4319e66f17ffSNaoya Horiguchi if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) { 4320e66f17ffSNaoya Horiguchi spin_unlock(ptl); 4321e66f17ffSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pmd, ptl); 4322e66f17ffSNaoya Horiguchi goto retry; 4323e66f17ffSNaoya Horiguchi } 4324e66f17ffSNaoya Horiguchi /* 4325e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 4326e66f17ffSNaoya Horiguchi * follow_page_mask(). 4327e66f17ffSNaoya Horiguchi */ 4328e66f17ffSNaoya Horiguchi } 4329e66f17ffSNaoya Horiguchi out: 4330e66f17ffSNaoya Horiguchi spin_unlock(ptl); 43319e5fc74cSSteve Capper return page; 43329e5fc74cSSteve Capper } 43339e5fc74cSSteve Capper 433461f77edaSNaoya Horiguchi struct page * __weak 43359e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 4336e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 43379e5fc74cSSteve Capper { 4338e66f17ffSNaoya Horiguchi if (flags & FOLL_GET) 4339e66f17ffSNaoya Horiguchi return NULL; 43409e5fc74cSSteve Capper 4341e66f17ffSNaoya Horiguchi return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 43429e5fc74cSSteve Capper } 43439e5fc74cSSteve Capper 4344d5bd9106SAndi Kleen #ifdef CONFIG_MEMORY_FAILURE 4345d5bd9106SAndi Kleen 434693f70f90SNaoya Horiguchi /* 434793f70f90SNaoya Horiguchi * This function is called from memory failure code. 434893f70f90SNaoya Horiguchi * Assume the caller holds page lock of the head page. 434993f70f90SNaoya Horiguchi */ 43506de2b1aaSNaoya Horiguchi int dequeue_hwpoisoned_huge_page(struct page *hpage) 435193f70f90SNaoya Horiguchi { 435293f70f90SNaoya Horiguchi struct hstate *h = page_hstate(hpage); 435393f70f90SNaoya Horiguchi int nid = page_to_nid(hpage); 43546de2b1aaSNaoya Horiguchi int ret = -EBUSY; 435593f70f90SNaoya Horiguchi 435693f70f90SNaoya Horiguchi spin_lock(&hugetlb_lock); 43577e1f049eSNaoya Horiguchi /* 43587e1f049eSNaoya Horiguchi * Just checking !page_huge_active is not enough, because that could be 43597e1f049eSNaoya Horiguchi * an isolated/hwpoisoned hugepage (which have >0 refcount). 43607e1f049eSNaoya Horiguchi */ 43617e1f049eSNaoya Horiguchi if (!page_huge_active(hpage) && !page_count(hpage)) { 436256f2fb14SNaoya Horiguchi /* 436356f2fb14SNaoya Horiguchi * Hwpoisoned hugepage isn't linked to activelist or freelist, 436456f2fb14SNaoya Horiguchi * but dangling hpage->lru can trigger list-debug warnings 436556f2fb14SNaoya Horiguchi * (this happens when we call unpoison_memory() on it), 436656f2fb14SNaoya Horiguchi * so let it point to itself with list_del_init(). 436756f2fb14SNaoya Horiguchi */ 436856f2fb14SNaoya Horiguchi list_del_init(&hpage->lru); 43698c6c2ecbSNaoya Horiguchi set_page_refcounted(hpage); 437093f70f90SNaoya Horiguchi h->free_huge_pages--; 437193f70f90SNaoya Horiguchi h->free_huge_pages_node[nid]--; 43726de2b1aaSNaoya Horiguchi ret = 0; 437393f70f90SNaoya Horiguchi } 43746de2b1aaSNaoya Horiguchi spin_unlock(&hugetlb_lock); 43756de2b1aaSNaoya Horiguchi return ret; 43766de2b1aaSNaoya Horiguchi } 43776de2b1aaSNaoya Horiguchi #endif 437831caf665SNaoya Horiguchi 437931caf665SNaoya Horiguchi bool isolate_huge_page(struct page *page, struct list_head *list) 438031caf665SNaoya Horiguchi { 4381bcc54222SNaoya Horiguchi bool ret = true; 4382bcc54222SNaoya Horiguchi 4383309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 438431caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 4385bcc54222SNaoya Horiguchi if (!page_huge_active(page) || !get_page_unless_zero(page)) { 4386bcc54222SNaoya Horiguchi ret = false; 4387bcc54222SNaoya Horiguchi goto unlock; 4388bcc54222SNaoya Horiguchi } 4389bcc54222SNaoya Horiguchi clear_page_huge_active(page); 439031caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 4391bcc54222SNaoya Horiguchi unlock: 439231caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 4393bcc54222SNaoya Horiguchi return ret; 439431caf665SNaoya Horiguchi } 439531caf665SNaoya Horiguchi 439631caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 439731caf665SNaoya Horiguchi { 4398309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 439931caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 4400bcc54222SNaoya Horiguchi set_page_huge_active(page); 440131caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 440231caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 440331caf665SNaoya Horiguchi put_page(page); 440431caf665SNaoya Horiguchi } 4405