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/module.h> 81da177e4SLinus Torvalds #include <linux/mm.h> 9e1759c21SAlexey Dobriyan #include <linux/seq_file.h> 101da177e4SLinus Torvalds #include <linux/sysctl.h> 111da177e4SLinus Torvalds #include <linux/highmem.h> 12cddb8a5cSAndrea Arcangeli #include <linux/mmu_notifier.h> 131da177e4SLinus Torvalds #include <linux/nodemask.h> 1463551ae0SDavid Gibson #include <linux/pagemap.h> 155da7ca86SChristoph Lameter #include <linux/mempolicy.h> 163b32123dSGideon Israel Dsouza #include <linux/compiler.h> 17aea47ff3SChristoph Lameter #include <linux/cpuset.h> 183935baa9SDavid Gibson #include <linux/mutex.h> 19aa888a74SAndi Kleen #include <linux/bootmem.h> 20a3437870SNishanth Aravamudan #include <linux/sysfs.h> 215a0e3ad6STejun Heo #include <linux/slab.h> 220fe6e20bSNaoya Horiguchi #include <linux/rmap.h> 23fd6a03edSNaoya Horiguchi #include <linux/swap.h> 24fd6a03edSNaoya Horiguchi #include <linux/swapops.h> 25c8721bbbSNaoya Horiguchi #include <linux/page-isolation.h> 268382d914SDavidlohr Bueso #include <linux/jhash.h> 27d6606683SLinus Torvalds 2863551ae0SDavid Gibson #include <asm/page.h> 2963551ae0SDavid Gibson #include <asm/pgtable.h> 3024669e58SAneesh Kumar K.V #include <asm/tlb.h> 3163551ae0SDavid Gibson 3224669e58SAneesh Kumar K.V #include <linux/io.h> 3363551ae0SDavid Gibson #include <linux/hugetlb.h> 349dd540e2SAneesh Kumar K.V #include <linux/hugetlb_cgroup.h> 359a305230SLee Schermerhorn #include <linux/node.h> 367835e98bSNick Piggin #include "internal.h" 371da177e4SLinus Torvalds 38753162cdSAndrey Ryabinin int hugepages_treat_as_movable; 39a5516438SAndi Kleen 40c3f38a38SAneesh Kumar K.V int hugetlb_max_hstate __read_mostly; 41e5ff2159SAndi Kleen unsigned int default_hstate_idx; 42e5ff2159SAndi Kleen struct hstate hstates[HUGE_MAX_HSTATE]; 43641844f5SNaoya Horiguchi /* 44641844f5SNaoya Horiguchi * Minimum page order among possible hugepage sizes, set to a proper value 45641844f5SNaoya Horiguchi * at boot time. 46641844f5SNaoya Horiguchi */ 47641844f5SNaoya Horiguchi static unsigned int minimum_order __read_mostly = UINT_MAX; 48e5ff2159SAndi Kleen 4953ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 5053ba51d2SJon Tollefson 51e5ff2159SAndi Kleen /* for command line parsing */ 52e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 53e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 54e11bfbfcSNick Piggin static unsigned long __initdata default_hstate_size; 55e5ff2159SAndi Kleen 563935baa9SDavid Gibson /* 5731caf665SNaoya Horiguchi * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, 5831caf665SNaoya Horiguchi * free_huge_pages, and surplus_huge_pages. 593935baa9SDavid Gibson */ 60c3f38a38SAneesh Kumar K.V DEFINE_SPINLOCK(hugetlb_lock); 610bd0f9fbSEric Paris 628382d914SDavidlohr Bueso /* 638382d914SDavidlohr Bueso * Serializes faults on the same logical page. This is used to 648382d914SDavidlohr Bueso * prevent spurious OOMs when the hugepage pool is fully utilized. 658382d914SDavidlohr Bueso */ 668382d914SDavidlohr Bueso static int num_fault_mutexes; 67c672c7f2SMike Kravetz struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; 688382d914SDavidlohr Bueso 697ca02d0aSMike Kravetz /* Forward declaration */ 707ca02d0aSMike Kravetz static int hugetlb_acct_memory(struct hstate *h, long delta); 717ca02d0aSMike Kravetz 7290481622SDavid Gibson static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) 7390481622SDavid Gibson { 7490481622SDavid Gibson bool free = (spool->count == 0) && (spool->used_hpages == 0); 7590481622SDavid Gibson 7690481622SDavid Gibson spin_unlock(&spool->lock); 7790481622SDavid Gibson 7890481622SDavid Gibson /* If no pages are used, and no other handles to the subpool 797ca02d0aSMike Kravetz * remain, give up any reservations mased on minimum size and 807ca02d0aSMike Kravetz * free the subpool */ 817ca02d0aSMike Kravetz if (free) { 827ca02d0aSMike Kravetz if (spool->min_hpages != -1) 837ca02d0aSMike Kravetz hugetlb_acct_memory(spool->hstate, 847ca02d0aSMike Kravetz -spool->min_hpages); 8590481622SDavid Gibson kfree(spool); 8690481622SDavid Gibson } 877ca02d0aSMike Kravetz } 8890481622SDavid Gibson 897ca02d0aSMike Kravetz struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, 907ca02d0aSMike Kravetz long min_hpages) 9190481622SDavid Gibson { 9290481622SDavid Gibson struct hugepage_subpool *spool; 9390481622SDavid Gibson 94c6a91820SMike Kravetz spool = kzalloc(sizeof(*spool), GFP_KERNEL); 9590481622SDavid Gibson if (!spool) 9690481622SDavid Gibson return NULL; 9790481622SDavid Gibson 9890481622SDavid Gibson spin_lock_init(&spool->lock); 9990481622SDavid Gibson spool->count = 1; 1007ca02d0aSMike Kravetz spool->max_hpages = max_hpages; 1017ca02d0aSMike Kravetz spool->hstate = h; 1027ca02d0aSMike Kravetz spool->min_hpages = min_hpages; 1037ca02d0aSMike Kravetz 1047ca02d0aSMike Kravetz if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { 1057ca02d0aSMike Kravetz kfree(spool); 1067ca02d0aSMike Kravetz return NULL; 1077ca02d0aSMike Kravetz } 1087ca02d0aSMike Kravetz spool->rsv_hpages = min_hpages; 10990481622SDavid Gibson 11090481622SDavid Gibson return spool; 11190481622SDavid Gibson } 11290481622SDavid Gibson 11390481622SDavid Gibson void hugepage_put_subpool(struct hugepage_subpool *spool) 11490481622SDavid Gibson { 11590481622SDavid Gibson spin_lock(&spool->lock); 11690481622SDavid Gibson BUG_ON(!spool->count); 11790481622SDavid Gibson spool->count--; 11890481622SDavid Gibson unlock_or_release_subpool(spool); 11990481622SDavid Gibson } 12090481622SDavid Gibson 1211c5ecae3SMike Kravetz /* 1221c5ecae3SMike Kravetz * Subpool accounting for allocating and reserving pages. 1231c5ecae3SMike Kravetz * Return -ENOMEM if there are not enough resources to satisfy the 1241c5ecae3SMike Kravetz * the request. Otherwise, return the number of pages by which the 1251c5ecae3SMike Kravetz * global pools must be adjusted (upward). The returned value may 1261c5ecae3SMike Kravetz * only be different than the passed value (delta) in the case where 1271c5ecae3SMike Kravetz * a subpool minimum size must be manitained. 1281c5ecae3SMike Kravetz */ 1291c5ecae3SMike Kravetz static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, 13090481622SDavid Gibson long delta) 13190481622SDavid Gibson { 1321c5ecae3SMike Kravetz long ret = delta; 13390481622SDavid Gibson 13490481622SDavid Gibson if (!spool) 1351c5ecae3SMike Kravetz return ret; 13690481622SDavid Gibson 13790481622SDavid Gibson spin_lock(&spool->lock); 13890481622SDavid Gibson 1391c5ecae3SMike Kravetz if (spool->max_hpages != -1) { /* maximum size accounting */ 1401c5ecae3SMike Kravetz if ((spool->used_hpages + delta) <= spool->max_hpages) 1411c5ecae3SMike Kravetz spool->used_hpages += delta; 1421c5ecae3SMike Kravetz else { 1431c5ecae3SMike Kravetz ret = -ENOMEM; 1441c5ecae3SMike Kravetz goto unlock_ret; 1451c5ecae3SMike Kravetz } 1461c5ecae3SMike Kravetz } 1471c5ecae3SMike Kravetz 1481c5ecae3SMike Kravetz if (spool->min_hpages != -1) { /* minimum size accounting */ 1491c5ecae3SMike Kravetz if (delta > spool->rsv_hpages) { 1501c5ecae3SMike Kravetz /* 1511c5ecae3SMike Kravetz * Asking for more reserves than those already taken on 1521c5ecae3SMike Kravetz * behalf of subpool. Return difference. 1531c5ecae3SMike Kravetz */ 1541c5ecae3SMike Kravetz ret = delta - spool->rsv_hpages; 1551c5ecae3SMike Kravetz spool->rsv_hpages = 0; 1561c5ecae3SMike Kravetz } else { 1571c5ecae3SMike Kravetz ret = 0; /* reserves already accounted for */ 1581c5ecae3SMike Kravetz spool->rsv_hpages -= delta; 1591c5ecae3SMike Kravetz } 1601c5ecae3SMike Kravetz } 1611c5ecae3SMike Kravetz 1621c5ecae3SMike Kravetz unlock_ret: 1631c5ecae3SMike Kravetz spin_unlock(&spool->lock); 16490481622SDavid Gibson return ret; 16590481622SDavid Gibson } 16690481622SDavid Gibson 1671c5ecae3SMike Kravetz /* 1681c5ecae3SMike Kravetz * Subpool accounting for freeing and unreserving pages. 1691c5ecae3SMike Kravetz * Return the number of global page reservations that must be dropped. 1701c5ecae3SMike Kravetz * The return value may only be different than the passed value (delta) 1711c5ecae3SMike Kravetz * in the case where a subpool minimum size must be maintained. 1721c5ecae3SMike Kravetz */ 1731c5ecae3SMike Kravetz static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, 17490481622SDavid Gibson long delta) 17590481622SDavid Gibson { 1761c5ecae3SMike Kravetz long ret = delta; 1771c5ecae3SMike Kravetz 17890481622SDavid Gibson if (!spool) 1791c5ecae3SMike Kravetz return delta; 18090481622SDavid Gibson 18190481622SDavid Gibson spin_lock(&spool->lock); 1821c5ecae3SMike Kravetz 1831c5ecae3SMike Kravetz if (spool->max_hpages != -1) /* maximum size accounting */ 18490481622SDavid Gibson spool->used_hpages -= delta; 1851c5ecae3SMike Kravetz 1861c5ecae3SMike Kravetz if (spool->min_hpages != -1) { /* minimum size accounting */ 1871c5ecae3SMike Kravetz if (spool->rsv_hpages + delta <= spool->min_hpages) 1881c5ecae3SMike Kravetz ret = 0; 1891c5ecae3SMike Kravetz else 1901c5ecae3SMike Kravetz ret = spool->rsv_hpages + delta - spool->min_hpages; 1911c5ecae3SMike Kravetz 1921c5ecae3SMike Kravetz spool->rsv_hpages += delta; 1931c5ecae3SMike Kravetz if (spool->rsv_hpages > spool->min_hpages) 1941c5ecae3SMike Kravetz spool->rsv_hpages = spool->min_hpages; 1951c5ecae3SMike Kravetz } 1961c5ecae3SMike Kravetz 1971c5ecae3SMike Kravetz /* 1981c5ecae3SMike Kravetz * If hugetlbfs_put_super couldn't free spool due to an outstanding 1991c5ecae3SMike Kravetz * quota reference, free it now. 2001c5ecae3SMike Kravetz */ 20190481622SDavid Gibson unlock_or_release_subpool(spool); 2021c5ecae3SMike Kravetz 2031c5ecae3SMike Kravetz return ret; 20490481622SDavid Gibson } 20590481622SDavid Gibson 20690481622SDavid Gibson static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 20790481622SDavid Gibson { 20890481622SDavid Gibson return HUGETLBFS_SB(inode->i_sb)->spool; 20990481622SDavid Gibson } 21090481622SDavid Gibson 21190481622SDavid Gibson static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 21290481622SDavid Gibson { 213496ad9aaSAl Viro return subpool_inode(file_inode(vma->vm_file)); 21490481622SDavid Gibson } 21590481622SDavid Gibson 216e7c4b0bfSAndy Whitcroft /* 21796822904SAndy Whitcroft * Region tracking -- allows tracking of reservations and instantiated pages 21896822904SAndy Whitcroft * across the pages in a mapping. 21984afd99bSAndy Whitcroft * 2201dd308a7SMike Kravetz * The region data structures are embedded into a resv_map and protected 2211dd308a7SMike Kravetz * by a resv_map's lock. The set of regions within the resv_map represent 2221dd308a7SMike Kravetz * reservations for huge pages, or huge pages that have already been 2231dd308a7SMike Kravetz * instantiated within the map. The from and to elements are huge page 2241dd308a7SMike Kravetz * indicies into the associated mapping. from indicates the starting index 2251dd308a7SMike Kravetz * of the region. to represents the first index past the end of the region. 2261dd308a7SMike Kravetz * 2271dd308a7SMike Kravetz * For example, a file region structure with from == 0 and to == 4 represents 2281dd308a7SMike Kravetz * four huge pages in a mapping. It is important to note that the to element 2291dd308a7SMike Kravetz * represents the first element past the end of the region. This is used in 2301dd308a7SMike Kravetz * arithmetic as 4(to) - 0(from) = 4 huge pages in the region. 2311dd308a7SMike Kravetz * 2321dd308a7SMike Kravetz * Interval notation of the form [from, to) will be used to indicate that 2331dd308a7SMike Kravetz * the endpoint from is inclusive and to is exclusive. 23496822904SAndy Whitcroft */ 23596822904SAndy Whitcroft struct file_region { 23696822904SAndy Whitcroft struct list_head link; 23796822904SAndy Whitcroft long from; 23896822904SAndy Whitcroft long to; 23996822904SAndy Whitcroft }; 24096822904SAndy Whitcroft 2411dd308a7SMike Kravetz /* 2421dd308a7SMike Kravetz * Add the huge page range represented by [f, t) to the reserve 2435e911373SMike Kravetz * map. In the normal case, existing regions will be expanded 2445e911373SMike Kravetz * to accommodate the specified range. Sufficient regions should 2455e911373SMike Kravetz * exist for expansion due to the previous call to region_chg 2465e911373SMike Kravetz * with the same range. However, it is possible that region_del 2475e911373SMike Kravetz * could have been called after region_chg and modifed the map 2485e911373SMike Kravetz * in such a way that no region exists to be expanded. In this 2495e911373SMike Kravetz * case, pull a region descriptor from the cache associated with 2505e911373SMike Kravetz * the map and use that for the new range. 251cf3ad20bSMike Kravetz * 252cf3ad20bSMike Kravetz * Return the number of new huge pages added to the map. This 253cf3ad20bSMike Kravetz * number is greater than or equal to zero. 2541dd308a7SMike Kravetz */ 2551406ec9bSJoonsoo Kim static long region_add(struct resv_map *resv, long f, long t) 25696822904SAndy Whitcroft { 2571406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 25896822904SAndy Whitcroft struct file_region *rg, *nrg, *trg; 259cf3ad20bSMike Kravetz long add = 0; 26096822904SAndy Whitcroft 2617b24d861SDavidlohr Bueso spin_lock(&resv->lock); 26296822904SAndy Whitcroft /* Locate the region we are either in or before. */ 26396822904SAndy Whitcroft list_for_each_entry(rg, head, link) 26496822904SAndy Whitcroft if (f <= rg->to) 26596822904SAndy Whitcroft break; 26696822904SAndy Whitcroft 2675e911373SMike Kravetz /* 2685e911373SMike Kravetz * If no region exists which can be expanded to include the 2695e911373SMike Kravetz * specified range, the list must have been modified by an 2705e911373SMike Kravetz * interleving call to region_del(). Pull a region descriptor 2715e911373SMike Kravetz * from the cache and use it for this range. 2725e911373SMike Kravetz */ 2735e911373SMike Kravetz if (&rg->link == head || t < rg->from) { 2745e911373SMike Kravetz VM_BUG_ON(resv->region_cache_count <= 0); 2755e911373SMike Kravetz 2765e911373SMike Kravetz resv->region_cache_count--; 2775e911373SMike Kravetz nrg = list_first_entry(&resv->region_cache, struct file_region, 2785e911373SMike Kravetz link); 2795e911373SMike Kravetz list_del(&nrg->link); 2805e911373SMike Kravetz 2815e911373SMike Kravetz nrg->from = f; 2825e911373SMike Kravetz nrg->to = t; 2835e911373SMike Kravetz list_add(&nrg->link, rg->link.prev); 2845e911373SMike Kravetz 2855e911373SMike Kravetz add += t - f; 2865e911373SMike Kravetz goto out_locked; 2875e911373SMike Kravetz } 2885e911373SMike Kravetz 28996822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 29096822904SAndy Whitcroft if (f > rg->from) 29196822904SAndy Whitcroft f = rg->from; 29296822904SAndy Whitcroft 29396822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 29496822904SAndy Whitcroft nrg = rg; 29596822904SAndy Whitcroft list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 29696822904SAndy Whitcroft if (&rg->link == head) 29796822904SAndy Whitcroft break; 29896822904SAndy Whitcroft if (rg->from > t) 29996822904SAndy Whitcroft break; 30096822904SAndy Whitcroft 30196822904SAndy Whitcroft /* If this area reaches higher then extend our area to 30296822904SAndy Whitcroft * include it completely. If this is not the first area 30396822904SAndy Whitcroft * which we intend to reuse, free it. */ 30496822904SAndy Whitcroft if (rg->to > t) 30596822904SAndy Whitcroft t = rg->to; 30696822904SAndy Whitcroft if (rg != nrg) { 307cf3ad20bSMike Kravetz /* Decrement return value by the deleted range. 308cf3ad20bSMike Kravetz * Another range will span this area so that by 309cf3ad20bSMike Kravetz * end of routine add will be >= zero 310cf3ad20bSMike Kravetz */ 311cf3ad20bSMike Kravetz add -= (rg->to - rg->from); 31296822904SAndy Whitcroft list_del(&rg->link); 31396822904SAndy Whitcroft kfree(rg); 31496822904SAndy Whitcroft } 31596822904SAndy Whitcroft } 316cf3ad20bSMike Kravetz 317cf3ad20bSMike Kravetz add += (nrg->from - f); /* Added to beginning of region */ 31896822904SAndy Whitcroft nrg->from = f; 319cf3ad20bSMike Kravetz add += t - nrg->to; /* Added to end of region */ 32096822904SAndy Whitcroft nrg->to = t; 321cf3ad20bSMike Kravetz 3225e911373SMike Kravetz out_locked: 3235e911373SMike Kravetz resv->adds_in_progress--; 3247b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 325cf3ad20bSMike Kravetz VM_BUG_ON(add < 0); 326cf3ad20bSMike Kravetz return add; 32796822904SAndy Whitcroft } 32896822904SAndy Whitcroft 3291dd308a7SMike Kravetz /* 3301dd308a7SMike Kravetz * Examine the existing reserve map and determine how many 3311dd308a7SMike Kravetz * huge pages in the specified range [f, t) are NOT currently 3321dd308a7SMike Kravetz * represented. This routine is called before a subsequent 3331dd308a7SMike Kravetz * call to region_add that will actually modify the reserve 3341dd308a7SMike Kravetz * map to add the specified range [f, t). region_chg does 3351dd308a7SMike Kravetz * not change the number of huge pages represented by the 3361dd308a7SMike Kravetz * map. However, if the existing regions in the map can not 3371dd308a7SMike Kravetz * be expanded to represent the new range, a new file_region 3381dd308a7SMike Kravetz * structure is added to the map as a placeholder. This is 3391dd308a7SMike Kravetz * so that the subsequent region_add call will have all the 3401dd308a7SMike Kravetz * regions it needs and will not fail. 3411dd308a7SMike Kravetz * 3425e911373SMike Kravetz * Upon entry, region_chg will also examine the cache of region descriptors 3435e911373SMike Kravetz * associated with the map. If there are not enough descriptors cached, one 3445e911373SMike Kravetz * will be allocated for the in progress add operation. 3455e911373SMike Kravetz * 3465e911373SMike Kravetz * Returns the number of huge pages that need to be added to the existing 3475e911373SMike Kravetz * reservation map for the range [f, t). This number is greater or equal to 3485e911373SMike Kravetz * zero. -ENOMEM is returned if a new file_region structure or cache entry 3495e911373SMike Kravetz * is needed and can not be allocated. 3501dd308a7SMike Kravetz */ 3511406ec9bSJoonsoo Kim static long region_chg(struct resv_map *resv, long f, long t) 35296822904SAndy Whitcroft { 3531406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 3547b24d861SDavidlohr Bueso struct file_region *rg, *nrg = NULL; 35596822904SAndy Whitcroft long chg = 0; 35696822904SAndy Whitcroft 3577b24d861SDavidlohr Bueso retry: 3587b24d861SDavidlohr Bueso spin_lock(&resv->lock); 3595e911373SMike Kravetz retry_locked: 3605e911373SMike Kravetz resv->adds_in_progress++; 3615e911373SMike Kravetz 3625e911373SMike Kravetz /* 3635e911373SMike Kravetz * Check for sufficient descriptors in the cache to accommodate 3645e911373SMike Kravetz * the number of in progress add operations. 3655e911373SMike Kravetz */ 3665e911373SMike Kravetz if (resv->adds_in_progress > resv->region_cache_count) { 3675e911373SMike Kravetz struct file_region *trg; 3685e911373SMike Kravetz 3695e911373SMike Kravetz VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1); 3705e911373SMike Kravetz /* Must drop lock to allocate a new descriptor. */ 3715e911373SMike Kravetz resv->adds_in_progress--; 3725e911373SMike Kravetz spin_unlock(&resv->lock); 3735e911373SMike Kravetz 3745e911373SMike Kravetz trg = kmalloc(sizeof(*trg), GFP_KERNEL); 3755e911373SMike Kravetz if (!trg) 3765e911373SMike Kravetz return -ENOMEM; 3775e911373SMike Kravetz 3785e911373SMike Kravetz spin_lock(&resv->lock); 3795e911373SMike Kravetz list_add(&trg->link, &resv->region_cache); 3805e911373SMike Kravetz resv->region_cache_count++; 3815e911373SMike Kravetz goto retry_locked; 3825e911373SMike Kravetz } 3835e911373SMike Kravetz 38496822904SAndy Whitcroft /* Locate the region we are before or in. */ 38596822904SAndy Whitcroft list_for_each_entry(rg, head, link) 38696822904SAndy Whitcroft if (f <= rg->to) 38796822904SAndy Whitcroft break; 38896822904SAndy Whitcroft 38996822904SAndy Whitcroft /* If we are below the current region then a new region is required. 39096822904SAndy Whitcroft * Subtle, allocate a new region at the position but make it zero 39196822904SAndy Whitcroft * size such that we can guarantee to record the reservation. */ 39296822904SAndy Whitcroft if (&rg->link == head || t < rg->from) { 3937b24d861SDavidlohr Bueso if (!nrg) { 3945e911373SMike Kravetz resv->adds_in_progress--; 3957b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 39696822904SAndy Whitcroft nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 39796822904SAndy Whitcroft if (!nrg) 39896822904SAndy Whitcroft return -ENOMEM; 3997b24d861SDavidlohr Bueso 40096822904SAndy Whitcroft nrg->from = f; 40196822904SAndy Whitcroft nrg->to = f; 40296822904SAndy Whitcroft INIT_LIST_HEAD(&nrg->link); 4037b24d861SDavidlohr Bueso goto retry; 4047b24d861SDavidlohr Bueso } 40596822904SAndy Whitcroft 4067b24d861SDavidlohr Bueso list_add(&nrg->link, rg->link.prev); 4077b24d861SDavidlohr Bueso chg = t - f; 4087b24d861SDavidlohr Bueso goto out_nrg; 40996822904SAndy Whitcroft } 41096822904SAndy Whitcroft 41196822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 41296822904SAndy Whitcroft if (f > rg->from) 41396822904SAndy Whitcroft f = rg->from; 41496822904SAndy Whitcroft chg = t - f; 41596822904SAndy Whitcroft 41696822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 41796822904SAndy Whitcroft list_for_each_entry(rg, rg->link.prev, link) { 41896822904SAndy Whitcroft if (&rg->link == head) 41996822904SAndy Whitcroft break; 42096822904SAndy Whitcroft if (rg->from > t) 4217b24d861SDavidlohr Bueso goto out; 42296822904SAndy Whitcroft 42325985edcSLucas De Marchi /* We overlap with this area, if it extends further than 42496822904SAndy Whitcroft * us then we must extend ourselves. Account for its 42596822904SAndy Whitcroft * existing reservation. */ 42696822904SAndy Whitcroft if (rg->to > t) { 42796822904SAndy Whitcroft chg += rg->to - t; 42896822904SAndy Whitcroft t = rg->to; 42996822904SAndy Whitcroft } 43096822904SAndy Whitcroft chg -= rg->to - rg->from; 43196822904SAndy Whitcroft } 4327b24d861SDavidlohr Bueso 4337b24d861SDavidlohr Bueso out: 4347b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 4357b24d861SDavidlohr Bueso /* We already know we raced and no longer need the new region */ 4367b24d861SDavidlohr Bueso kfree(nrg); 4377b24d861SDavidlohr Bueso return chg; 4387b24d861SDavidlohr Bueso out_nrg: 4397b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 44096822904SAndy Whitcroft return chg; 44196822904SAndy Whitcroft } 44296822904SAndy Whitcroft 4431dd308a7SMike Kravetz /* 4445e911373SMike Kravetz * Abort the in progress add operation. The adds_in_progress field 4455e911373SMike Kravetz * of the resv_map keeps track of the operations in progress between 4465e911373SMike Kravetz * calls to region_chg and region_add. Operations are sometimes 4475e911373SMike Kravetz * aborted after the call to region_chg. In such cases, region_abort 4485e911373SMike Kravetz * is called to decrement the adds_in_progress counter. 4495e911373SMike Kravetz * 4505e911373SMike Kravetz * NOTE: The range arguments [f, t) are not needed or used in this 4515e911373SMike Kravetz * routine. They are kept to make reading the calling code easier as 4525e911373SMike Kravetz * arguments will match the associated region_chg call. 4535e911373SMike Kravetz */ 4545e911373SMike Kravetz static void region_abort(struct resv_map *resv, long f, long t) 4555e911373SMike Kravetz { 4565e911373SMike Kravetz spin_lock(&resv->lock); 4575e911373SMike Kravetz VM_BUG_ON(!resv->region_cache_count); 4585e911373SMike Kravetz resv->adds_in_progress--; 4595e911373SMike Kravetz spin_unlock(&resv->lock); 4605e911373SMike Kravetz } 4615e911373SMike Kravetz 4625e911373SMike Kravetz /* 463feba16e2SMike Kravetz * Delete the specified range [f, t) from the reserve map. If the 464feba16e2SMike Kravetz * t parameter is LONG_MAX, this indicates that ALL regions after f 465feba16e2SMike Kravetz * should be deleted. Locate the regions which intersect [f, t) 466feba16e2SMike Kravetz * and either trim, delete or split the existing regions. 467feba16e2SMike Kravetz * 468feba16e2SMike Kravetz * Returns the number of huge pages deleted from the reserve map. 469feba16e2SMike Kravetz * In the normal case, the return value is zero or more. In the 470feba16e2SMike Kravetz * case where a region must be split, a new region descriptor must 471feba16e2SMike Kravetz * be allocated. If the allocation fails, -ENOMEM will be returned. 472feba16e2SMike Kravetz * NOTE: If the parameter t == LONG_MAX, then we will never split 473feba16e2SMike Kravetz * a region and possibly return -ENOMEM. Callers specifying 474feba16e2SMike Kravetz * t == LONG_MAX do not need to check for -ENOMEM error. 4751dd308a7SMike Kravetz */ 476feba16e2SMike Kravetz static long region_del(struct resv_map *resv, long f, long t) 47796822904SAndy Whitcroft { 4781406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 47996822904SAndy Whitcroft struct file_region *rg, *trg; 480feba16e2SMike Kravetz struct file_region *nrg = NULL; 481feba16e2SMike Kravetz long del = 0; 48296822904SAndy Whitcroft 483feba16e2SMike Kravetz retry: 4847b24d861SDavidlohr Bueso spin_lock(&resv->lock); 485feba16e2SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 486feba16e2SMike Kravetz if (rg->to <= f) 487feba16e2SMike Kravetz continue; 488feba16e2SMike Kravetz if (rg->from >= t) 48996822904SAndy Whitcroft break; 49096822904SAndy Whitcroft 491feba16e2SMike Kravetz if (f > rg->from && t < rg->to) { /* Must split region */ 492feba16e2SMike Kravetz /* 493feba16e2SMike Kravetz * Check for an entry in the cache before dropping 494feba16e2SMike Kravetz * lock and attempting allocation. 495feba16e2SMike Kravetz */ 496feba16e2SMike Kravetz if (!nrg && 497feba16e2SMike Kravetz resv->region_cache_count > resv->adds_in_progress) { 498feba16e2SMike Kravetz nrg = list_first_entry(&resv->region_cache, 499feba16e2SMike Kravetz struct file_region, 500feba16e2SMike Kravetz link); 501feba16e2SMike Kravetz list_del(&nrg->link); 502feba16e2SMike Kravetz resv->region_cache_count--; 50396822904SAndy Whitcroft } 50496822904SAndy Whitcroft 505feba16e2SMike Kravetz if (!nrg) { 506feba16e2SMike Kravetz spin_unlock(&resv->lock); 507feba16e2SMike Kravetz nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 508feba16e2SMike Kravetz if (!nrg) 509feba16e2SMike Kravetz return -ENOMEM; 510feba16e2SMike Kravetz goto retry; 511feba16e2SMike Kravetz } 512feba16e2SMike Kravetz 513feba16e2SMike Kravetz del += t - f; 514feba16e2SMike Kravetz 515feba16e2SMike Kravetz /* New entry for end of split region */ 516feba16e2SMike Kravetz nrg->from = t; 517feba16e2SMike Kravetz nrg->to = rg->to; 518feba16e2SMike Kravetz INIT_LIST_HEAD(&nrg->link); 519feba16e2SMike Kravetz 520feba16e2SMike Kravetz /* Original entry is trimmed */ 521feba16e2SMike Kravetz rg->to = f; 522feba16e2SMike Kravetz 523feba16e2SMike Kravetz list_add(&nrg->link, &rg->link); 524feba16e2SMike Kravetz nrg = NULL; 52596822904SAndy Whitcroft break; 526feba16e2SMike Kravetz } 527feba16e2SMike Kravetz 528feba16e2SMike Kravetz if (f <= rg->from && t >= rg->to) { /* Remove entire region */ 529feba16e2SMike Kravetz del += rg->to - rg->from; 53096822904SAndy Whitcroft list_del(&rg->link); 53196822904SAndy Whitcroft kfree(rg); 532feba16e2SMike Kravetz continue; 53396822904SAndy Whitcroft } 5347b24d861SDavidlohr Bueso 535feba16e2SMike Kravetz if (f <= rg->from) { /* Trim beginning of region */ 536feba16e2SMike Kravetz del += t - rg->from; 537feba16e2SMike Kravetz rg->from = t; 538feba16e2SMike Kravetz } else { /* Trim end of region */ 539feba16e2SMike Kravetz del += rg->to - f; 540feba16e2SMike Kravetz rg->to = f; 541feba16e2SMike Kravetz } 542feba16e2SMike Kravetz } 543feba16e2SMike Kravetz 5447b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 545feba16e2SMike Kravetz kfree(nrg); 546feba16e2SMike Kravetz return del; 54796822904SAndy Whitcroft } 54896822904SAndy Whitcroft 5491dd308a7SMike Kravetz /* 550b5cec28dSMike Kravetz * A rare out of memory error was encountered which prevented removal of 551b5cec28dSMike Kravetz * the reserve map region for a page. The huge page itself was free'ed 552b5cec28dSMike Kravetz * and removed from the page cache. This routine will adjust the subpool 553b5cec28dSMike Kravetz * usage count, and the global reserve count if needed. By incrementing 554b5cec28dSMike Kravetz * these counts, the reserve map entry which could not be deleted will 555b5cec28dSMike Kravetz * appear as a "reserved" entry instead of simply dangling with incorrect 556b5cec28dSMike Kravetz * counts. 557b5cec28dSMike Kravetz */ 558b5cec28dSMike Kravetz void hugetlb_fix_reserve_counts(struct inode *inode, bool restore_reserve) 559b5cec28dSMike Kravetz { 560b5cec28dSMike Kravetz struct hugepage_subpool *spool = subpool_inode(inode); 561b5cec28dSMike Kravetz long rsv_adjust; 562b5cec28dSMike Kravetz 563b5cec28dSMike Kravetz rsv_adjust = hugepage_subpool_get_pages(spool, 1); 564b5cec28dSMike Kravetz if (restore_reserve && rsv_adjust) { 565b5cec28dSMike Kravetz struct hstate *h = hstate_inode(inode); 566b5cec28dSMike Kravetz 567b5cec28dSMike Kravetz hugetlb_acct_memory(h, 1); 568b5cec28dSMike Kravetz } 569b5cec28dSMike Kravetz } 570b5cec28dSMike Kravetz 571b5cec28dSMike Kravetz /* 5721dd308a7SMike Kravetz * Count and return the number of huge pages in the reserve map 5731dd308a7SMike Kravetz * that intersect with the range [f, t). 5741dd308a7SMike Kravetz */ 5751406ec9bSJoonsoo Kim static long region_count(struct resv_map *resv, long f, long t) 57684afd99bSAndy Whitcroft { 5771406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 57884afd99bSAndy Whitcroft struct file_region *rg; 57984afd99bSAndy Whitcroft long chg = 0; 58084afd99bSAndy Whitcroft 5817b24d861SDavidlohr Bueso spin_lock(&resv->lock); 58284afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 58384afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 584f2135a4aSWang Sheng-Hui long seg_from; 585f2135a4aSWang Sheng-Hui long seg_to; 58684afd99bSAndy Whitcroft 58784afd99bSAndy Whitcroft if (rg->to <= f) 58884afd99bSAndy Whitcroft continue; 58984afd99bSAndy Whitcroft if (rg->from >= t) 59084afd99bSAndy Whitcroft break; 59184afd99bSAndy Whitcroft 59284afd99bSAndy Whitcroft seg_from = max(rg->from, f); 59384afd99bSAndy Whitcroft seg_to = min(rg->to, t); 59484afd99bSAndy Whitcroft 59584afd99bSAndy Whitcroft chg += seg_to - seg_from; 59684afd99bSAndy Whitcroft } 5977b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 59884afd99bSAndy Whitcroft 59984afd99bSAndy Whitcroft return chg; 60084afd99bSAndy Whitcroft } 60184afd99bSAndy Whitcroft 60296822904SAndy Whitcroft /* 603e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 604e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 605e7c4b0bfSAndy Whitcroft */ 606a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 607a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 608e7c4b0bfSAndy Whitcroft { 609a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 610a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 611e7c4b0bfSAndy Whitcroft } 612e7c4b0bfSAndy Whitcroft 6130fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 6140fe6e20bSNaoya Horiguchi unsigned long address) 6150fe6e20bSNaoya Horiguchi { 6160fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 6170fe6e20bSNaoya Horiguchi } 6180fe6e20bSNaoya Horiguchi 61984afd99bSAndy Whitcroft /* 62008fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 62108fba699SMel Gorman * cases this will be same size as used by the page table entries. 62208fba699SMel Gorman */ 62308fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 62408fba699SMel Gorman { 62508fba699SMel Gorman struct hstate *hstate; 62608fba699SMel Gorman 62708fba699SMel Gorman if (!is_vm_hugetlb_page(vma)) 62808fba699SMel Gorman return PAGE_SIZE; 62908fba699SMel Gorman 63008fba699SMel Gorman hstate = hstate_vma(vma); 63108fba699SMel Gorman 6322415cf12SWanpeng Li return 1UL << huge_page_shift(hstate); 63308fba699SMel Gorman } 634f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 63508fba699SMel Gorman 63608fba699SMel Gorman /* 6373340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 6383340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 6393340289dSMel Gorman * architectures where it differs, an architecture-specific version of this 6403340289dSMel Gorman * function is required. 6413340289dSMel Gorman */ 6423340289dSMel Gorman #ifndef vma_mmu_pagesize 6433340289dSMel Gorman unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 6443340289dSMel Gorman { 6453340289dSMel Gorman return vma_kernel_pagesize(vma); 6463340289dSMel Gorman } 6473340289dSMel Gorman #endif 6483340289dSMel Gorman 6493340289dSMel Gorman /* 65084afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 65184afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 65284afd99bSAndy Whitcroft * alignment. 65384afd99bSAndy Whitcroft */ 65484afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 65584afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 65604f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 65784afd99bSAndy Whitcroft 658a1e78772SMel Gorman /* 659a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 660a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 661a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 662a1e78772SMel Gorman * 663a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 664a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 665a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 666a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 66784afd99bSAndy Whitcroft * 66884afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 66984afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 67084afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 67184afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 67284afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 67384afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 67484afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 67584afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 676a1e78772SMel Gorman */ 677e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 678e7c4b0bfSAndy Whitcroft { 679e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 680e7c4b0bfSAndy Whitcroft } 681e7c4b0bfSAndy Whitcroft 682e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 683e7c4b0bfSAndy Whitcroft unsigned long value) 684e7c4b0bfSAndy Whitcroft { 685e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 686e7c4b0bfSAndy Whitcroft } 687e7c4b0bfSAndy Whitcroft 6889119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 68984afd99bSAndy Whitcroft { 69084afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 6915e911373SMike Kravetz struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 6925e911373SMike Kravetz 6935e911373SMike Kravetz if (!resv_map || !rg) { 6945e911373SMike Kravetz kfree(resv_map); 6955e911373SMike Kravetz kfree(rg); 69684afd99bSAndy Whitcroft return NULL; 6975e911373SMike Kravetz } 69884afd99bSAndy Whitcroft 69984afd99bSAndy Whitcroft kref_init(&resv_map->refs); 7007b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 70184afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 70284afd99bSAndy Whitcroft 7035e911373SMike Kravetz resv_map->adds_in_progress = 0; 7045e911373SMike Kravetz 7055e911373SMike Kravetz INIT_LIST_HEAD(&resv_map->region_cache); 7065e911373SMike Kravetz list_add(&rg->link, &resv_map->region_cache); 7075e911373SMike Kravetz resv_map->region_cache_count = 1; 7085e911373SMike Kravetz 70984afd99bSAndy Whitcroft return resv_map; 71084afd99bSAndy Whitcroft } 71184afd99bSAndy Whitcroft 7129119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 71384afd99bSAndy Whitcroft { 71484afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 7155e911373SMike Kravetz struct list_head *head = &resv_map->region_cache; 7165e911373SMike Kravetz struct file_region *rg, *trg; 71784afd99bSAndy Whitcroft 71884afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 719feba16e2SMike Kravetz region_del(resv_map, 0, LONG_MAX); 7205e911373SMike Kravetz 7215e911373SMike Kravetz /* ... and any entries left in the cache */ 7225e911373SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 7235e911373SMike Kravetz list_del(&rg->link); 7245e911373SMike Kravetz kfree(rg); 7255e911373SMike Kravetz } 7265e911373SMike Kravetz 7275e911373SMike Kravetz VM_BUG_ON(resv_map->adds_in_progress); 7285e911373SMike Kravetz 72984afd99bSAndy Whitcroft kfree(resv_map); 73084afd99bSAndy Whitcroft } 73184afd99bSAndy Whitcroft 7324e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 7334e35f483SJoonsoo Kim { 7344e35f483SJoonsoo Kim return inode->i_mapping->private_data; 7354e35f483SJoonsoo Kim } 7364e35f483SJoonsoo Kim 73784afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 738a1e78772SMel Gorman { 73981d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 7404e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 7414e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 7424e35f483SJoonsoo Kim struct inode *inode = mapping->host; 7434e35f483SJoonsoo Kim 7444e35f483SJoonsoo Kim return inode_resv_map(inode); 7454e35f483SJoonsoo Kim 7464e35f483SJoonsoo Kim } else { 74784afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 74884afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 7494e35f483SJoonsoo Kim } 750a1e78772SMel Gorman } 751a1e78772SMel Gorman 75284afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 753a1e78772SMel Gorman { 75481d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 75581d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 756a1e78772SMel Gorman 75784afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 75884afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 75904f2cbe3SMel Gorman } 76004f2cbe3SMel Gorman 76104f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 76204f2cbe3SMel 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); 765e7c4b0bfSAndy Whitcroft 766e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 76704f2cbe3SMel Gorman } 76804f2cbe3SMel Gorman 76904f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 77004f2cbe3SMel Gorman { 77181d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 772e7c4b0bfSAndy Whitcroft 773e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 774a1e78772SMel Gorman } 775a1e78772SMel Gorman 77604f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 777a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 778a1e78772SMel Gorman { 77981d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 780f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 781a1e78772SMel Gorman vma->vm_private_data = (void *)0; 782a1e78772SMel Gorman } 783a1e78772SMel Gorman 784a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 785559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 786a1e78772SMel Gorman { 787af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 788af0ed73eSJoonsoo Kim /* 789af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 790af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 791af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 792af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 793af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 794af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 795af0ed73eSJoonsoo Kim * properly, so add work-around here. 796af0ed73eSJoonsoo Kim */ 797af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 798559ec2f8SNicholas Krause return true; 799af0ed73eSJoonsoo Kim else 800559ec2f8SNicholas Krause return false; 801af0ed73eSJoonsoo Kim } 802a63884e9SJoonsoo Kim 803a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 8041fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 8051fb1b0e9SMike Kravetz /* 8061fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 8071fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 8081fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 8091fb1b0e9SMike Kravetz * fallocate. In this case, there really are no reverves to 8101fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 8111fb1b0e9SMike Kravetz */ 8121fb1b0e9SMike Kravetz if (chg) 8131fb1b0e9SMike Kravetz return false; 8141fb1b0e9SMike Kravetz else 815559ec2f8SNicholas Krause return true; 8161fb1b0e9SMike Kravetz } 817a63884e9SJoonsoo Kim 818a63884e9SJoonsoo Kim /* 819a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 820a63884e9SJoonsoo Kim * private mappings. 821a63884e9SJoonsoo Kim */ 8227f09ca51SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 823559ec2f8SNicholas Krause return true; 824a63884e9SJoonsoo Kim 825559ec2f8SNicholas Krause return false; 826a1e78772SMel Gorman } 827a1e78772SMel Gorman 828a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 8291da177e4SLinus Torvalds { 8301da177e4SLinus Torvalds int nid = page_to_nid(page); 8310edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 832a5516438SAndi Kleen h->free_huge_pages++; 833a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 8341da177e4SLinus Torvalds } 8351da177e4SLinus Torvalds 836bf50bab2SNaoya Horiguchi static struct page *dequeue_huge_page_node(struct hstate *h, int nid) 837bf50bab2SNaoya Horiguchi { 838bf50bab2SNaoya Horiguchi struct page *page; 839bf50bab2SNaoya Horiguchi 840c8721bbbSNaoya Horiguchi list_for_each_entry(page, &h->hugepage_freelists[nid], lru) 841c8721bbbSNaoya Horiguchi if (!is_migrate_isolate_page(page)) 842c8721bbbSNaoya Horiguchi break; 843c8721bbbSNaoya Horiguchi /* 844c8721bbbSNaoya Horiguchi * if 'non-isolated free hugepage' not found on the list, 845c8721bbbSNaoya Horiguchi * the allocation fails. 846c8721bbbSNaoya Horiguchi */ 847c8721bbbSNaoya Horiguchi if (&h->hugepage_freelists[nid] == &page->lru) 848bf50bab2SNaoya Horiguchi return NULL; 8490edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 850a9869b83SNaoya Horiguchi set_page_refcounted(page); 851bf50bab2SNaoya Horiguchi h->free_huge_pages--; 852bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 853bf50bab2SNaoya Horiguchi return page; 854bf50bab2SNaoya Horiguchi } 855bf50bab2SNaoya Horiguchi 85686cdb465SNaoya Horiguchi /* Movability of hugepages depends on migration support. */ 85786cdb465SNaoya Horiguchi static inline gfp_t htlb_alloc_mask(struct hstate *h) 85886cdb465SNaoya Horiguchi { 859100873d7SNaoya Horiguchi if (hugepages_treat_as_movable || hugepage_migration_supported(h)) 86086cdb465SNaoya Horiguchi return GFP_HIGHUSER_MOVABLE; 86186cdb465SNaoya Horiguchi else 86286cdb465SNaoya Horiguchi return GFP_HIGHUSER; 86386cdb465SNaoya Horiguchi } 86486cdb465SNaoya Horiguchi 865a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 866a5516438SAndi Kleen struct vm_area_struct *vma, 867af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 868af0ed73eSJoonsoo Kim long chg) 8691da177e4SLinus Torvalds { 870b1c12cbcSKonstantin Khlebnikov struct page *page = NULL; 871480eccf9SLee Schermerhorn struct mempolicy *mpol; 87219770b32SMel Gorman nodemask_t *nodemask; 873c0ff7453SMiao Xie struct zonelist *zonelist; 874dd1a239fSMel Gorman struct zone *zone; 875dd1a239fSMel Gorman struct zoneref *z; 876cc9a6c87SMel Gorman unsigned int cpuset_mems_cookie; 8771da177e4SLinus Torvalds 878a1e78772SMel Gorman /* 879a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 880a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 881a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 882a1e78772SMel Gorman */ 883af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg) && 884a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 885c0ff7453SMiao Xie goto err; 886a1e78772SMel Gorman 88704f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 888a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 8896eab04a8SJustin P. Mattock goto err; 89004f2cbe3SMel Gorman 8919966c4bbSJoonsoo Kim retry_cpuset: 892d26914d1SMel Gorman cpuset_mems_cookie = read_mems_allowed_begin(); 8939966c4bbSJoonsoo Kim zonelist = huge_zonelist(vma, address, 89486cdb465SNaoya Horiguchi htlb_alloc_mask(h), &mpol, &nodemask); 8959966c4bbSJoonsoo Kim 89619770b32SMel Gorman for_each_zone_zonelist_nodemask(zone, z, zonelist, 89719770b32SMel Gorman MAX_NR_ZONES - 1, nodemask) { 898344736f2SVladimir Davydov if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) { 899bf50bab2SNaoya Horiguchi page = dequeue_huge_page_node(h, zone_to_nid(zone)); 900bf50bab2SNaoya Horiguchi if (page) { 901af0ed73eSJoonsoo Kim if (avoid_reserve) 902af0ed73eSJoonsoo Kim break; 903af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg)) 904af0ed73eSJoonsoo Kim break; 905af0ed73eSJoonsoo Kim 90607443a85SJoonsoo Kim SetPagePrivate(page); 907a63884e9SJoonsoo Kim h->resv_huge_pages--; 9085ab3ee7bSKen Chen break; 9091da177e4SLinus Torvalds } 9103abf7afdSAndrew Morton } 911bf50bab2SNaoya Horiguchi } 912cc9a6c87SMel Gorman 913cc9a6c87SMel Gorman mpol_cond_put(mpol); 914d26914d1SMel Gorman if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) 915cc9a6c87SMel Gorman goto retry_cpuset; 916cc9a6c87SMel Gorman return page; 917cc9a6c87SMel Gorman 918c0ff7453SMiao Xie err: 919cc9a6c87SMel Gorman return NULL; 9201da177e4SLinus Torvalds } 9211da177e4SLinus Torvalds 9221cac6f2cSLuiz Capitulino /* 9231cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 9241cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 9251cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 9261cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 9271cac6f2cSLuiz Capitulino * node for alloc or free. 9281cac6f2cSLuiz Capitulino */ 9291cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 9301cac6f2cSLuiz Capitulino { 9311cac6f2cSLuiz Capitulino nid = next_node(nid, *nodes_allowed); 9321cac6f2cSLuiz Capitulino if (nid == MAX_NUMNODES) 9331cac6f2cSLuiz Capitulino nid = first_node(*nodes_allowed); 9341cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 9351cac6f2cSLuiz Capitulino 9361cac6f2cSLuiz Capitulino return nid; 9371cac6f2cSLuiz Capitulino } 9381cac6f2cSLuiz Capitulino 9391cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 9401cac6f2cSLuiz Capitulino { 9411cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 9421cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 9431cac6f2cSLuiz Capitulino return nid; 9441cac6f2cSLuiz Capitulino } 9451cac6f2cSLuiz Capitulino 9461cac6f2cSLuiz Capitulino /* 9471cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 9481cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 9491cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 9501cac6f2cSLuiz Capitulino * mask. 9511cac6f2cSLuiz Capitulino */ 9521cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 9531cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 9541cac6f2cSLuiz Capitulino { 9551cac6f2cSLuiz Capitulino int nid; 9561cac6f2cSLuiz Capitulino 9571cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 9581cac6f2cSLuiz Capitulino 9591cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 9601cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 9611cac6f2cSLuiz Capitulino 9621cac6f2cSLuiz Capitulino return nid; 9631cac6f2cSLuiz Capitulino } 9641cac6f2cSLuiz Capitulino 9651cac6f2cSLuiz Capitulino /* 9661cac6f2cSLuiz Capitulino * helper for free_pool_huge_page() - return the previously saved 9671cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 9681cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 9691cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 9701cac6f2cSLuiz Capitulino */ 9711cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 9721cac6f2cSLuiz Capitulino { 9731cac6f2cSLuiz Capitulino int nid; 9741cac6f2cSLuiz Capitulino 9751cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 9761cac6f2cSLuiz Capitulino 9771cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 9781cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 9791cac6f2cSLuiz Capitulino 9801cac6f2cSLuiz Capitulino return nid; 9811cac6f2cSLuiz Capitulino } 9821cac6f2cSLuiz Capitulino 9831cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 9841cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 9851cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 9861cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 9871cac6f2cSLuiz Capitulino nr_nodes--) 9881cac6f2cSLuiz Capitulino 9891cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 9901cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 9911cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 9921cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 9931cac6f2cSLuiz Capitulino nr_nodes--) 9941cac6f2cSLuiz Capitulino 995944d9fecSLuiz Capitulino #if defined(CONFIG_CMA) && defined(CONFIG_X86_64) 996944d9fecSLuiz Capitulino static void destroy_compound_gigantic_page(struct page *page, 997d00181b9SKirill A. Shutemov unsigned int order) 998944d9fecSLuiz Capitulino { 999944d9fecSLuiz Capitulino int i; 1000944d9fecSLuiz Capitulino int nr_pages = 1 << order; 1001944d9fecSLuiz Capitulino struct page *p = page + 1; 1002944d9fecSLuiz Capitulino 1003944d9fecSLuiz Capitulino for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 10041d798ca3SKirill A. Shutemov clear_compound_head(p); 1005944d9fecSLuiz Capitulino set_page_refcounted(p); 1006944d9fecSLuiz Capitulino } 1007944d9fecSLuiz Capitulino 1008944d9fecSLuiz Capitulino set_compound_order(page, 0); 1009944d9fecSLuiz Capitulino __ClearPageHead(page); 1010944d9fecSLuiz Capitulino } 1011944d9fecSLuiz Capitulino 1012d00181b9SKirill A. Shutemov static void free_gigantic_page(struct page *page, unsigned int order) 1013944d9fecSLuiz Capitulino { 1014944d9fecSLuiz Capitulino free_contig_range(page_to_pfn(page), 1 << order); 1015944d9fecSLuiz Capitulino } 1016944d9fecSLuiz Capitulino 1017944d9fecSLuiz Capitulino static int __alloc_gigantic_page(unsigned long start_pfn, 1018944d9fecSLuiz Capitulino unsigned long nr_pages) 1019944d9fecSLuiz Capitulino { 1020944d9fecSLuiz Capitulino unsigned long end_pfn = start_pfn + nr_pages; 1021944d9fecSLuiz Capitulino return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE); 1022944d9fecSLuiz Capitulino } 1023944d9fecSLuiz Capitulino 1024944d9fecSLuiz Capitulino static bool pfn_range_valid_gigantic(unsigned long start_pfn, 1025944d9fecSLuiz Capitulino unsigned long nr_pages) 1026944d9fecSLuiz Capitulino { 1027944d9fecSLuiz Capitulino unsigned long i, end_pfn = start_pfn + nr_pages; 1028944d9fecSLuiz Capitulino struct page *page; 1029944d9fecSLuiz Capitulino 1030944d9fecSLuiz Capitulino for (i = start_pfn; i < end_pfn; i++) { 1031944d9fecSLuiz Capitulino if (!pfn_valid(i)) 1032944d9fecSLuiz Capitulino return false; 1033944d9fecSLuiz Capitulino 1034944d9fecSLuiz Capitulino page = pfn_to_page(i); 1035944d9fecSLuiz Capitulino 1036944d9fecSLuiz Capitulino if (PageReserved(page)) 1037944d9fecSLuiz Capitulino return false; 1038944d9fecSLuiz Capitulino 1039944d9fecSLuiz Capitulino if (page_count(page) > 0) 1040944d9fecSLuiz Capitulino return false; 1041944d9fecSLuiz Capitulino 1042944d9fecSLuiz Capitulino if (PageHuge(page)) 1043944d9fecSLuiz Capitulino return false; 1044944d9fecSLuiz Capitulino } 1045944d9fecSLuiz Capitulino 1046944d9fecSLuiz Capitulino return true; 1047944d9fecSLuiz Capitulino } 1048944d9fecSLuiz Capitulino 1049944d9fecSLuiz Capitulino static bool zone_spans_last_pfn(const struct zone *zone, 1050944d9fecSLuiz Capitulino unsigned long start_pfn, unsigned long nr_pages) 1051944d9fecSLuiz Capitulino { 1052944d9fecSLuiz Capitulino unsigned long last_pfn = start_pfn + nr_pages - 1; 1053944d9fecSLuiz Capitulino return zone_spans_pfn(zone, last_pfn); 1054944d9fecSLuiz Capitulino } 1055944d9fecSLuiz Capitulino 1056d00181b9SKirill A. Shutemov static struct page *alloc_gigantic_page(int nid, unsigned int order) 1057944d9fecSLuiz Capitulino { 1058944d9fecSLuiz Capitulino unsigned long nr_pages = 1 << order; 1059944d9fecSLuiz Capitulino unsigned long ret, pfn, flags; 1060944d9fecSLuiz Capitulino struct zone *z; 1061944d9fecSLuiz Capitulino 1062944d9fecSLuiz Capitulino z = NODE_DATA(nid)->node_zones; 1063944d9fecSLuiz Capitulino for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) { 1064944d9fecSLuiz Capitulino spin_lock_irqsave(&z->lock, flags); 1065944d9fecSLuiz Capitulino 1066944d9fecSLuiz Capitulino pfn = ALIGN(z->zone_start_pfn, nr_pages); 1067944d9fecSLuiz Capitulino while (zone_spans_last_pfn(z, pfn, nr_pages)) { 1068944d9fecSLuiz Capitulino if (pfn_range_valid_gigantic(pfn, nr_pages)) { 1069944d9fecSLuiz Capitulino /* 1070944d9fecSLuiz Capitulino * We release the zone lock here because 1071944d9fecSLuiz Capitulino * alloc_contig_range() will also lock the zone 1072944d9fecSLuiz Capitulino * at some point. If there's an allocation 1073944d9fecSLuiz Capitulino * spinning on this lock, it may win the race 1074944d9fecSLuiz Capitulino * and cause alloc_contig_range() to fail... 1075944d9fecSLuiz Capitulino */ 1076944d9fecSLuiz Capitulino spin_unlock_irqrestore(&z->lock, flags); 1077944d9fecSLuiz Capitulino ret = __alloc_gigantic_page(pfn, nr_pages); 1078944d9fecSLuiz Capitulino if (!ret) 1079944d9fecSLuiz Capitulino return pfn_to_page(pfn); 1080944d9fecSLuiz Capitulino spin_lock_irqsave(&z->lock, flags); 1081944d9fecSLuiz Capitulino } 1082944d9fecSLuiz Capitulino pfn += nr_pages; 1083944d9fecSLuiz Capitulino } 1084944d9fecSLuiz Capitulino 1085944d9fecSLuiz Capitulino spin_unlock_irqrestore(&z->lock, flags); 1086944d9fecSLuiz Capitulino } 1087944d9fecSLuiz Capitulino 1088944d9fecSLuiz Capitulino return NULL; 1089944d9fecSLuiz Capitulino } 1090944d9fecSLuiz Capitulino 1091944d9fecSLuiz Capitulino static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); 1092d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order); 1093944d9fecSLuiz Capitulino 1094944d9fecSLuiz Capitulino static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid) 1095944d9fecSLuiz Capitulino { 1096944d9fecSLuiz Capitulino struct page *page; 1097944d9fecSLuiz Capitulino 1098944d9fecSLuiz Capitulino page = alloc_gigantic_page(nid, huge_page_order(h)); 1099944d9fecSLuiz Capitulino if (page) { 1100944d9fecSLuiz Capitulino prep_compound_gigantic_page(page, huge_page_order(h)); 1101944d9fecSLuiz Capitulino prep_new_huge_page(h, page, nid); 1102944d9fecSLuiz Capitulino } 1103944d9fecSLuiz Capitulino 1104944d9fecSLuiz Capitulino return page; 1105944d9fecSLuiz Capitulino } 1106944d9fecSLuiz Capitulino 1107944d9fecSLuiz Capitulino static int alloc_fresh_gigantic_page(struct hstate *h, 1108944d9fecSLuiz Capitulino nodemask_t *nodes_allowed) 1109944d9fecSLuiz Capitulino { 1110944d9fecSLuiz Capitulino struct page *page = NULL; 1111944d9fecSLuiz Capitulino int nr_nodes, node; 1112944d9fecSLuiz Capitulino 1113944d9fecSLuiz Capitulino for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1114944d9fecSLuiz Capitulino page = alloc_fresh_gigantic_page_node(h, node); 1115944d9fecSLuiz Capitulino if (page) 1116944d9fecSLuiz Capitulino return 1; 1117944d9fecSLuiz Capitulino } 1118944d9fecSLuiz Capitulino 1119944d9fecSLuiz Capitulino return 0; 1120944d9fecSLuiz Capitulino } 1121944d9fecSLuiz Capitulino 1122944d9fecSLuiz Capitulino static inline bool gigantic_page_supported(void) { return true; } 1123944d9fecSLuiz Capitulino #else 1124944d9fecSLuiz Capitulino static inline bool gigantic_page_supported(void) { return false; } 1125d00181b9SKirill A. Shutemov static inline void free_gigantic_page(struct page *page, unsigned int order) { } 1126944d9fecSLuiz Capitulino static inline void destroy_compound_gigantic_page(struct page *page, 1127d00181b9SKirill A. Shutemov unsigned int order) { } 1128944d9fecSLuiz Capitulino static inline int alloc_fresh_gigantic_page(struct hstate *h, 1129944d9fecSLuiz Capitulino nodemask_t *nodes_allowed) { return 0; } 1130944d9fecSLuiz Capitulino #endif 1131944d9fecSLuiz Capitulino 1132a5516438SAndi Kleen static void update_and_free_page(struct hstate *h, struct page *page) 11336af2acb6SAdam Litke { 11346af2acb6SAdam Litke int i; 1135a5516438SAndi Kleen 1136944d9fecSLuiz Capitulino if (hstate_is_gigantic(h) && !gigantic_page_supported()) 1137944d9fecSLuiz Capitulino return; 113818229df5SAndy Whitcroft 1139a5516438SAndi Kleen h->nr_huge_pages--; 1140a5516438SAndi Kleen h->nr_huge_pages_node[page_to_nid(page)]--; 1141a5516438SAndi Kleen for (i = 0; i < pages_per_huge_page(h); i++) { 114232f84528SChris Forbes page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 114332f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1144a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1145a7407a27SLuiz Capitulino 1 << PG_writeback); 11466af2acb6SAdam Litke } 1147309381feSSasha Levin VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 1148f1e61557SKirill A. Shutemov set_compound_page_dtor(page, NULL_COMPOUND_DTOR); 11496af2acb6SAdam Litke set_page_refcounted(page); 1150944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) { 1151944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 1152944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 1153944d9fecSLuiz Capitulino } else { 1154a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 11556af2acb6SAdam Litke } 1156944d9fecSLuiz Capitulino } 11576af2acb6SAdam Litke 1158e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1159e5ff2159SAndi Kleen { 1160e5ff2159SAndi Kleen struct hstate *h; 1161e5ff2159SAndi Kleen 1162e5ff2159SAndi Kleen for_each_hstate(h) { 1163e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1164e5ff2159SAndi Kleen return h; 1165e5ff2159SAndi Kleen } 1166e5ff2159SAndi Kleen return NULL; 1167e5ff2159SAndi Kleen } 1168e5ff2159SAndi Kleen 1169bcc54222SNaoya Horiguchi /* 1170bcc54222SNaoya Horiguchi * Test to determine whether the hugepage is "active/in-use" (i.e. being linked 1171bcc54222SNaoya Horiguchi * to hstate->hugepage_activelist.) 1172bcc54222SNaoya Horiguchi * 1173bcc54222SNaoya Horiguchi * This function can be called for tail pages, but never returns true for them. 1174bcc54222SNaoya Horiguchi */ 1175bcc54222SNaoya Horiguchi bool page_huge_active(struct page *page) 1176bcc54222SNaoya Horiguchi { 1177bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHuge(page), page); 1178bcc54222SNaoya Horiguchi return PageHead(page) && PagePrivate(&page[1]); 1179bcc54222SNaoya Horiguchi } 1180bcc54222SNaoya Horiguchi 1181bcc54222SNaoya Horiguchi /* never called for tail page */ 1182bcc54222SNaoya Horiguchi static void set_page_huge_active(struct page *page) 1183bcc54222SNaoya Horiguchi { 1184bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 1185bcc54222SNaoya Horiguchi SetPagePrivate(&page[1]); 1186bcc54222SNaoya Horiguchi } 1187bcc54222SNaoya Horiguchi 1188bcc54222SNaoya Horiguchi static void clear_page_huge_active(struct page *page) 1189bcc54222SNaoya Horiguchi { 1190bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 1191bcc54222SNaoya Horiguchi ClearPagePrivate(&page[1]); 1192bcc54222SNaoya Horiguchi } 1193bcc54222SNaoya Horiguchi 11948f1d26d0SAtsushi Kumagai void free_huge_page(struct page *page) 119527a85ef1SDavid Gibson { 1196a5516438SAndi Kleen /* 1197a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1198a5516438SAndi Kleen * compound page destructor. 1199a5516438SAndi Kleen */ 1200e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 12017893d1d5SAdam Litke int nid = page_to_nid(page); 120290481622SDavid Gibson struct hugepage_subpool *spool = 120390481622SDavid Gibson (struct hugepage_subpool *)page_private(page); 120407443a85SJoonsoo Kim bool restore_reserve; 120527a85ef1SDavid Gibson 1206e5df70abSAndy Whitcroft set_page_private(page, 0); 120723be7468SMel Gorman page->mapping = NULL; 12087893d1d5SAdam Litke BUG_ON(page_count(page)); 12090fe6e20bSNaoya Horiguchi BUG_ON(page_mapcount(page)); 121007443a85SJoonsoo Kim restore_reserve = PagePrivate(page); 121116c794b4SJoonsoo Kim ClearPagePrivate(page); 121227a85ef1SDavid Gibson 12131c5ecae3SMike Kravetz /* 12141c5ecae3SMike Kravetz * A return code of zero implies that the subpool will be under its 12151c5ecae3SMike Kravetz * minimum size if the reservation is not restored after page is free. 12161c5ecae3SMike Kravetz * Therefore, force restore_reserve operation. 12171c5ecae3SMike Kravetz */ 12181c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 12191c5ecae3SMike Kravetz restore_reserve = true; 12201c5ecae3SMike Kravetz 122127a85ef1SDavid Gibson spin_lock(&hugetlb_lock); 1222bcc54222SNaoya Horiguchi clear_page_huge_active(page); 12236d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 12246d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 122507443a85SJoonsoo Kim if (restore_reserve) 122607443a85SJoonsoo Kim h->resv_huge_pages++; 122707443a85SJoonsoo Kim 1228944d9fecSLuiz Capitulino if (h->surplus_huge_pages_node[nid]) { 12290edaecfaSAneesh Kumar K.V /* remove the page from active list */ 12300edaecfaSAneesh Kumar K.V list_del(&page->lru); 1231a5516438SAndi Kleen update_and_free_page(h, page); 1232a5516438SAndi Kleen h->surplus_huge_pages--; 1233a5516438SAndi Kleen h->surplus_huge_pages_node[nid]--; 12347893d1d5SAdam Litke } else { 12355d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1236a5516438SAndi Kleen enqueue_huge_page(h, page); 12377893d1d5SAdam Litke } 123827a85ef1SDavid Gibson spin_unlock(&hugetlb_lock); 123927a85ef1SDavid Gibson } 124027a85ef1SDavid Gibson 1241a5516438SAndi Kleen static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1242b7ba30c6SAndi Kleen { 12430edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1244f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1245b7ba30c6SAndi Kleen spin_lock(&hugetlb_lock); 12469dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 1247a5516438SAndi Kleen h->nr_huge_pages++; 1248a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 1249b7ba30c6SAndi Kleen spin_unlock(&hugetlb_lock); 1250b7ba30c6SAndi Kleen put_page(page); /* free it into the hugepage allocator */ 1251b7ba30c6SAndi Kleen } 1252b7ba30c6SAndi Kleen 1253d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order) 125420a0307cSWu Fengguang { 125520a0307cSWu Fengguang int i; 125620a0307cSWu Fengguang int nr_pages = 1 << order; 125720a0307cSWu Fengguang struct page *p = page + 1; 125820a0307cSWu Fengguang 125920a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 126020a0307cSWu Fengguang set_compound_order(page, order); 126120a0307cSWu Fengguang __SetPageHead(page); 1262ef5a22beSAndrea Arcangeli __ClearPageReserved(page); 126320a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1264ef5a22beSAndrea Arcangeli /* 1265ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1266ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1267ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 1268ef5a22beSAndrea Arcangeli * too. Otherwse drivers using get_user_pages() to access tail 1269ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1270ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1271ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1272ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1273ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1274ef5a22beSAndrea Arcangeli * after get_user_pages(). 1275ef5a22beSAndrea Arcangeli */ 1276ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 127758a84aa9SYouquan Song set_page_count(p, 0); 12781d798ca3SKirill A. Shutemov set_compound_head(p, page); 127920a0307cSWu Fengguang } 128020a0307cSWu Fengguang } 128120a0307cSWu Fengguang 12827795912cSAndrew Morton /* 12837795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 12847795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 12857795912cSAndrew Morton * details. 12867795912cSAndrew Morton */ 128720a0307cSWu Fengguang int PageHuge(struct page *page) 128820a0307cSWu Fengguang { 128920a0307cSWu Fengguang if (!PageCompound(page)) 129020a0307cSWu Fengguang return 0; 129120a0307cSWu Fengguang 129220a0307cSWu Fengguang page = compound_head(page); 1293f1e61557SKirill A. Shutemov return page[1].compound_dtor == HUGETLB_PAGE_DTOR; 129420a0307cSWu Fengguang } 129543131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 129643131e14SNaoya Horiguchi 129727c73ae7SAndrea Arcangeli /* 129827c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 129927c73ae7SAndrea Arcangeli * normal or transparent huge pages. 130027c73ae7SAndrea Arcangeli */ 130127c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 130227c73ae7SAndrea Arcangeli { 130327c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 130427c73ae7SAndrea Arcangeli return 0; 130527c73ae7SAndrea Arcangeli 1306758f66a2SAndrew Morton return get_compound_page_dtor(page_head) == free_huge_page; 130727c73ae7SAndrea Arcangeli } 130827c73ae7SAndrea Arcangeli 130913d60f4bSZhang Yi pgoff_t __basepage_index(struct page *page) 131013d60f4bSZhang Yi { 131113d60f4bSZhang Yi struct page *page_head = compound_head(page); 131213d60f4bSZhang Yi pgoff_t index = page_index(page_head); 131313d60f4bSZhang Yi unsigned long compound_idx; 131413d60f4bSZhang Yi 131513d60f4bSZhang Yi if (!PageHuge(page_head)) 131613d60f4bSZhang Yi return page_index(page); 131713d60f4bSZhang Yi 131813d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 131913d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 132013d60f4bSZhang Yi else 132113d60f4bSZhang Yi compound_idx = page - page_head; 132213d60f4bSZhang Yi 132313d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 132413d60f4bSZhang Yi } 132513d60f4bSZhang Yi 1326a5516438SAndi Kleen static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) 13271da177e4SLinus Torvalds { 13281da177e4SLinus Torvalds struct page *page; 1329f96efd58SJoe Jin 133096db800fSVlastimil Babka page = __alloc_pages_node(nid, 133186cdb465SNaoya Horiguchi htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| 1332551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 1333a5516438SAndi Kleen huge_page_order(h)); 13341da177e4SLinus Torvalds if (page) { 1335a5516438SAndi Kleen prep_new_huge_page(h, page, nid); 13361da177e4SLinus Torvalds } 133763b4613cSNishanth Aravamudan 133863b4613cSNishanth Aravamudan return page; 133963b4613cSNishanth Aravamudan } 134063b4613cSNishanth Aravamudan 1341b2261026SJoonsoo Kim static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 1342b2261026SJoonsoo Kim { 1343b2261026SJoonsoo Kim struct page *page; 1344b2261026SJoonsoo Kim int nr_nodes, node; 1345b2261026SJoonsoo Kim int ret = 0; 1346b2261026SJoonsoo Kim 1347b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1348b2261026SJoonsoo Kim page = alloc_fresh_huge_page_node(h, node); 1349b2261026SJoonsoo Kim if (page) { 1350b2261026SJoonsoo Kim ret = 1; 1351b2261026SJoonsoo Kim break; 1352b2261026SJoonsoo Kim } 1353b2261026SJoonsoo Kim } 1354b2261026SJoonsoo Kim 1355b2261026SJoonsoo Kim if (ret) 1356b2261026SJoonsoo Kim count_vm_event(HTLB_BUDDY_PGALLOC); 1357b2261026SJoonsoo Kim else 1358b2261026SJoonsoo Kim count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 1359b2261026SJoonsoo Kim 1360b2261026SJoonsoo Kim return ret; 1361b2261026SJoonsoo Kim } 1362b2261026SJoonsoo Kim 1363e8c5c824SLee Schermerhorn /* 1364e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 1365e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 1366e8c5c824SLee Schermerhorn * balanced over allowed nodes. 1367e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 1368e8c5c824SLee Schermerhorn */ 13696ae11b27SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 13706ae11b27SLee Schermerhorn bool acct_surplus) 1371e8c5c824SLee Schermerhorn { 1372b2261026SJoonsoo Kim int nr_nodes, node; 1373e8c5c824SLee Schermerhorn int ret = 0; 1374e8c5c824SLee Schermerhorn 1375b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1376685f3457SLee Schermerhorn /* 1377685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 1378685f3457SLee Schermerhorn * nodes with surplus pages. 1379685f3457SLee Schermerhorn */ 1380b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 1381b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 1382e8c5c824SLee Schermerhorn struct page *page = 1383b2261026SJoonsoo Kim list_entry(h->hugepage_freelists[node].next, 1384e8c5c824SLee Schermerhorn struct page, lru); 1385e8c5c824SLee Schermerhorn list_del(&page->lru); 1386e8c5c824SLee Schermerhorn h->free_huge_pages--; 1387b2261026SJoonsoo Kim h->free_huge_pages_node[node]--; 1388685f3457SLee Schermerhorn if (acct_surplus) { 1389685f3457SLee Schermerhorn h->surplus_huge_pages--; 1390b2261026SJoonsoo Kim h->surplus_huge_pages_node[node]--; 1391685f3457SLee Schermerhorn } 1392e8c5c824SLee Schermerhorn update_and_free_page(h, page); 1393e8c5c824SLee Schermerhorn ret = 1; 13949a76db09SLee Schermerhorn break; 1395e8c5c824SLee Schermerhorn } 1396b2261026SJoonsoo Kim } 1397e8c5c824SLee Schermerhorn 1398e8c5c824SLee Schermerhorn return ret; 1399e8c5c824SLee Schermerhorn } 1400e8c5c824SLee Schermerhorn 1401c8721bbbSNaoya Horiguchi /* 1402c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 1403c8721bbbSNaoya Horiguchi * nothing for in-use (including surplus) hugepages. 1404c8721bbbSNaoya Horiguchi */ 1405c8721bbbSNaoya Horiguchi static void dissolve_free_huge_page(struct page *page) 1406c8721bbbSNaoya Horiguchi { 1407c8721bbbSNaoya Horiguchi spin_lock(&hugetlb_lock); 1408c8721bbbSNaoya Horiguchi if (PageHuge(page) && !page_count(page)) { 1409c8721bbbSNaoya Horiguchi struct hstate *h = page_hstate(page); 1410c8721bbbSNaoya Horiguchi int nid = page_to_nid(page); 1411c8721bbbSNaoya Horiguchi list_del(&page->lru); 1412c8721bbbSNaoya Horiguchi h->free_huge_pages--; 1413c8721bbbSNaoya Horiguchi h->free_huge_pages_node[nid]--; 1414c8721bbbSNaoya Horiguchi update_and_free_page(h, page); 1415c8721bbbSNaoya Horiguchi } 1416c8721bbbSNaoya Horiguchi spin_unlock(&hugetlb_lock); 1417c8721bbbSNaoya Horiguchi } 1418c8721bbbSNaoya Horiguchi 1419c8721bbbSNaoya Horiguchi /* 1420c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 1421c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 1422c8721bbbSNaoya Horiguchi * Note that start_pfn should aligned with (minimum) hugepage size. 1423c8721bbbSNaoya Horiguchi */ 1424c8721bbbSNaoya Horiguchi void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 1425c8721bbbSNaoya Horiguchi { 1426c8721bbbSNaoya Horiguchi unsigned long pfn; 1427c8721bbbSNaoya Horiguchi 1428d0177639SLi Zhong if (!hugepages_supported()) 1429d0177639SLi Zhong return; 1430d0177639SLi Zhong 1431641844f5SNaoya Horiguchi VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << minimum_order)); 1432641844f5SNaoya Horiguchi for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) 1433c8721bbbSNaoya Horiguchi dissolve_free_huge_page(pfn_to_page(pfn)); 1434c8721bbbSNaoya Horiguchi } 1435c8721bbbSNaoya Horiguchi 1436099730d6SDave Hansen /* 1437099730d6SDave Hansen * There are 3 ways this can get called: 1438099730d6SDave Hansen * 1. With vma+addr: we use the VMA's memory policy 1439099730d6SDave Hansen * 2. With !vma, but nid=NUMA_NO_NODE: We try to allocate a huge 1440099730d6SDave Hansen * page from any node, and let the buddy allocator itself figure 1441099730d6SDave Hansen * it out. 1442099730d6SDave Hansen * 3. With !vma, but nid!=NUMA_NO_NODE. We allocate a huge page 1443099730d6SDave Hansen * strictly from 'nid' 1444099730d6SDave Hansen */ 1445099730d6SDave Hansen static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h, 1446099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr, int nid) 1447099730d6SDave Hansen { 1448099730d6SDave Hansen int order = huge_page_order(h); 1449099730d6SDave Hansen gfp_t gfp = htlb_alloc_mask(h)|__GFP_COMP|__GFP_REPEAT|__GFP_NOWARN; 1450099730d6SDave Hansen unsigned int cpuset_mems_cookie; 1451099730d6SDave Hansen 1452099730d6SDave Hansen /* 1453099730d6SDave Hansen * We need a VMA to get a memory policy. If we do not 1454e0ec90eeSDave Hansen * have one, we use the 'nid' argument. 1455e0ec90eeSDave Hansen * 1456e0ec90eeSDave Hansen * The mempolicy stuff below has some non-inlined bits 1457e0ec90eeSDave Hansen * and calls ->vm_ops. That makes it hard to optimize at 1458e0ec90eeSDave Hansen * compile-time, even when NUMA is off and it does 1459e0ec90eeSDave Hansen * nothing. This helps the compiler optimize it out. 1460099730d6SDave Hansen */ 1461e0ec90eeSDave Hansen if (!IS_ENABLED(CONFIG_NUMA) || !vma) { 1462099730d6SDave Hansen /* 1463099730d6SDave Hansen * If a specific node is requested, make sure to 1464099730d6SDave Hansen * get memory from there, but only when a node 1465099730d6SDave Hansen * is explicitly specified. 1466099730d6SDave Hansen */ 1467099730d6SDave Hansen if (nid != NUMA_NO_NODE) 1468099730d6SDave Hansen gfp |= __GFP_THISNODE; 1469099730d6SDave Hansen /* 1470099730d6SDave Hansen * Make sure to call something that can handle 1471099730d6SDave Hansen * nid=NUMA_NO_NODE 1472099730d6SDave Hansen */ 1473099730d6SDave Hansen return alloc_pages_node(nid, gfp, order); 1474099730d6SDave Hansen } 1475099730d6SDave Hansen 1476099730d6SDave Hansen /* 1477099730d6SDave Hansen * OK, so we have a VMA. Fetch the mempolicy and try to 1478e0ec90eeSDave Hansen * allocate a huge page with it. We will only reach this 1479e0ec90eeSDave Hansen * when CONFIG_NUMA=y. 1480099730d6SDave Hansen */ 1481099730d6SDave Hansen do { 1482099730d6SDave Hansen struct page *page; 1483099730d6SDave Hansen struct mempolicy *mpol; 1484099730d6SDave Hansen struct zonelist *zl; 1485099730d6SDave Hansen nodemask_t *nodemask; 1486099730d6SDave Hansen 1487099730d6SDave Hansen cpuset_mems_cookie = read_mems_allowed_begin(); 1488099730d6SDave Hansen zl = huge_zonelist(vma, addr, gfp, &mpol, &nodemask); 1489099730d6SDave Hansen mpol_cond_put(mpol); 1490099730d6SDave Hansen page = __alloc_pages_nodemask(gfp, order, zl, nodemask); 1491099730d6SDave Hansen if (page) 1492099730d6SDave Hansen return page; 1493099730d6SDave Hansen } while (read_mems_allowed_retry(cpuset_mems_cookie)); 1494099730d6SDave Hansen 1495099730d6SDave Hansen return NULL; 1496099730d6SDave Hansen } 1497099730d6SDave Hansen 1498099730d6SDave Hansen /* 1499099730d6SDave Hansen * There are two ways to allocate a huge page: 1500099730d6SDave Hansen * 1. When you have a VMA and an address (like a fault) 1501099730d6SDave Hansen * 2. When you have no VMA (like when setting /proc/.../nr_hugepages) 1502099730d6SDave Hansen * 1503099730d6SDave Hansen * 'vma' and 'addr' are only for (1). 'nid' is always NUMA_NO_NODE in 1504099730d6SDave Hansen * this case which signifies that the allocation should be done with 1505099730d6SDave Hansen * respect for the VMA's memory policy. 1506099730d6SDave Hansen * 1507099730d6SDave Hansen * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This 1508099730d6SDave Hansen * implies that memory policies will not be taken in to account. 1509099730d6SDave Hansen */ 1510099730d6SDave Hansen static struct page *__alloc_buddy_huge_page(struct hstate *h, 1511099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr, int nid) 15127893d1d5SAdam Litke { 15137893d1d5SAdam Litke struct page *page; 1514bf50bab2SNaoya Horiguchi unsigned int r_nid; 15157893d1d5SAdam Litke 1516bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1517aa888a74SAndi Kleen return NULL; 1518aa888a74SAndi Kleen 1519d1c3fb1fSNishanth Aravamudan /* 1520099730d6SDave Hansen * Make sure that anyone specifying 'nid' is not also specifying a VMA. 1521099730d6SDave Hansen * This makes sure the caller is picking _one_ of the modes with which 1522099730d6SDave Hansen * we can call this function, not both. 1523099730d6SDave Hansen */ 1524099730d6SDave Hansen if (vma || (addr != -1)) { 1525e0ec90eeSDave Hansen VM_WARN_ON_ONCE(addr == -1); 1526e0ec90eeSDave Hansen VM_WARN_ON_ONCE(nid != NUMA_NO_NODE); 1527099730d6SDave Hansen } 1528099730d6SDave Hansen /* 1529d1c3fb1fSNishanth Aravamudan * Assume we will successfully allocate the surplus page to 1530d1c3fb1fSNishanth Aravamudan * prevent racing processes from causing the surplus to exceed 1531d1c3fb1fSNishanth Aravamudan * overcommit 1532d1c3fb1fSNishanth Aravamudan * 1533d1c3fb1fSNishanth Aravamudan * This however introduces a different race, where a process B 1534d1c3fb1fSNishanth Aravamudan * tries to grow the static hugepage pool while alloc_pages() is 1535d1c3fb1fSNishanth Aravamudan * called by process A. B will only examine the per-node 1536d1c3fb1fSNishanth Aravamudan * counters in determining if surplus huge pages can be 1537d1c3fb1fSNishanth Aravamudan * converted to normal huge pages in adjust_pool_surplus(). A 1538d1c3fb1fSNishanth Aravamudan * won't be able to increment the per-node counter, until the 1539d1c3fb1fSNishanth Aravamudan * lock is dropped by B, but B doesn't drop hugetlb_lock until 1540d1c3fb1fSNishanth Aravamudan * no more huge pages can be converted from surplus to normal 1541d1c3fb1fSNishanth Aravamudan * state (and doesn't try to convert again). Thus, we have a 1542d1c3fb1fSNishanth Aravamudan * case where a surplus huge page exists, the pool is grown, and 1543d1c3fb1fSNishanth Aravamudan * the surplus huge page still exists after, even though it 1544d1c3fb1fSNishanth Aravamudan * should just have been converted to a normal huge page. This 1545d1c3fb1fSNishanth Aravamudan * does not leak memory, though, as the hugepage will be freed 1546d1c3fb1fSNishanth Aravamudan * once it is out of use. It also does not allow the counters to 1547d1c3fb1fSNishanth Aravamudan * go out of whack in adjust_pool_surplus() as we don't modify 1548d1c3fb1fSNishanth Aravamudan * the node values until we've gotten the hugepage and only the 1549d1c3fb1fSNishanth Aravamudan * per-node value is checked there. 1550d1c3fb1fSNishanth Aravamudan */ 1551d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 1552a5516438SAndi Kleen if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 1553d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1554d1c3fb1fSNishanth Aravamudan return NULL; 1555d1c3fb1fSNishanth Aravamudan } else { 1556a5516438SAndi Kleen h->nr_huge_pages++; 1557a5516438SAndi Kleen h->surplus_huge_pages++; 1558d1c3fb1fSNishanth Aravamudan } 1559d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1560d1c3fb1fSNishanth Aravamudan 1561099730d6SDave Hansen page = __hugetlb_alloc_buddy_huge_page(h, vma, addr, nid); 1562d1c3fb1fSNishanth Aravamudan 15637893d1d5SAdam Litke spin_lock(&hugetlb_lock); 1564d1c3fb1fSNishanth Aravamudan if (page) { 15650edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1566bf50bab2SNaoya Horiguchi r_nid = page_to_nid(page); 1567f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 15689dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 1569d1c3fb1fSNishanth Aravamudan /* 1570d1c3fb1fSNishanth Aravamudan * We incremented the global counters already 1571d1c3fb1fSNishanth Aravamudan */ 1572bf50bab2SNaoya Horiguchi h->nr_huge_pages_node[r_nid]++; 1573bf50bab2SNaoya Horiguchi h->surplus_huge_pages_node[r_nid]++; 15743b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC); 1575d1c3fb1fSNishanth Aravamudan } else { 1576a5516438SAndi Kleen h->nr_huge_pages--; 1577a5516438SAndi Kleen h->surplus_huge_pages--; 15783b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 15797893d1d5SAdam Litke } 1580d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 15817893d1d5SAdam Litke 15827893d1d5SAdam Litke return page; 15837893d1d5SAdam Litke } 15847893d1d5SAdam Litke 1585e4e574b7SAdam Litke /* 1586099730d6SDave Hansen * Allocate a huge page from 'nid'. Note, 'nid' may be 1587099730d6SDave Hansen * NUMA_NO_NODE, which means that it may be allocated 1588099730d6SDave Hansen * anywhere. 1589099730d6SDave Hansen */ 1590e0ec90eeSDave Hansen static 1591099730d6SDave Hansen struct page *__alloc_buddy_huge_page_no_mpol(struct hstate *h, int nid) 1592099730d6SDave Hansen { 1593099730d6SDave Hansen unsigned long addr = -1; 1594099730d6SDave Hansen 1595099730d6SDave Hansen return __alloc_buddy_huge_page(h, NULL, addr, nid); 1596099730d6SDave Hansen } 1597099730d6SDave Hansen 1598099730d6SDave Hansen /* 1599099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 1600099730d6SDave Hansen */ 1601e0ec90eeSDave Hansen static 1602099730d6SDave Hansen struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h, 1603099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 1604099730d6SDave Hansen { 1605099730d6SDave Hansen return __alloc_buddy_huge_page(h, vma, addr, NUMA_NO_NODE); 1606099730d6SDave Hansen } 1607099730d6SDave Hansen 1608099730d6SDave Hansen /* 1609bf50bab2SNaoya Horiguchi * This allocation function is useful in the context where vma is irrelevant. 1610bf50bab2SNaoya Horiguchi * E.g. soft-offlining uses this function because it only cares physical 1611bf50bab2SNaoya Horiguchi * address of error page. 1612bf50bab2SNaoya Horiguchi */ 1613bf50bab2SNaoya Horiguchi struct page *alloc_huge_page_node(struct hstate *h, int nid) 1614bf50bab2SNaoya Horiguchi { 16154ef91848SJoonsoo Kim struct page *page = NULL; 1616bf50bab2SNaoya Horiguchi 1617bf50bab2SNaoya Horiguchi spin_lock(&hugetlb_lock); 16184ef91848SJoonsoo Kim if (h->free_huge_pages - h->resv_huge_pages > 0) 1619bf50bab2SNaoya Horiguchi page = dequeue_huge_page_node(h, nid); 1620bf50bab2SNaoya Horiguchi spin_unlock(&hugetlb_lock); 1621bf50bab2SNaoya Horiguchi 162294ae8ba7SAneesh Kumar K.V if (!page) 1623099730d6SDave Hansen page = __alloc_buddy_huge_page_no_mpol(h, nid); 1624bf50bab2SNaoya Horiguchi 1625bf50bab2SNaoya Horiguchi return page; 1626bf50bab2SNaoya Horiguchi } 1627bf50bab2SNaoya Horiguchi 1628bf50bab2SNaoya Horiguchi /* 162925985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 1630e4e574b7SAdam Litke * of size 'delta'. 1631e4e574b7SAdam Litke */ 1632a5516438SAndi Kleen static int gather_surplus_pages(struct hstate *h, int delta) 1633e4e574b7SAdam Litke { 1634e4e574b7SAdam Litke struct list_head surplus_list; 1635e4e574b7SAdam Litke struct page *page, *tmp; 1636e4e574b7SAdam Litke int ret, i; 1637e4e574b7SAdam Litke int needed, allocated; 163828073b02SHillf Danton bool alloc_ok = true; 1639e4e574b7SAdam Litke 1640a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 1641ac09b3a1SAdam Litke if (needed <= 0) { 1642a5516438SAndi Kleen h->resv_huge_pages += delta; 1643e4e574b7SAdam Litke return 0; 1644ac09b3a1SAdam Litke } 1645e4e574b7SAdam Litke 1646e4e574b7SAdam Litke allocated = 0; 1647e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 1648e4e574b7SAdam Litke 1649e4e574b7SAdam Litke ret = -ENOMEM; 1650e4e574b7SAdam Litke retry: 1651e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 1652e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 1653099730d6SDave Hansen page = __alloc_buddy_huge_page_no_mpol(h, NUMA_NO_NODE); 165428073b02SHillf Danton if (!page) { 165528073b02SHillf Danton alloc_ok = false; 165628073b02SHillf Danton break; 165728073b02SHillf Danton } 1658e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 1659e4e574b7SAdam Litke } 166028073b02SHillf Danton allocated += i; 1661e4e574b7SAdam Litke 1662e4e574b7SAdam Litke /* 1663e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 1664e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 1665e4e574b7SAdam Litke */ 1666e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 1667a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 1668a5516438SAndi Kleen (h->free_huge_pages + allocated); 166928073b02SHillf Danton if (needed > 0) { 167028073b02SHillf Danton if (alloc_ok) 1671e4e574b7SAdam Litke goto retry; 167228073b02SHillf Danton /* 167328073b02SHillf Danton * We were not able to allocate enough pages to 167428073b02SHillf Danton * satisfy the entire reservation so we free what 167528073b02SHillf Danton * we've allocated so far. 167628073b02SHillf Danton */ 167728073b02SHillf Danton goto free; 167828073b02SHillf Danton } 1679e4e574b7SAdam Litke /* 1680e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 168125985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 1682e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 1683ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 1684ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 1685ac09b3a1SAdam Litke * before they are reserved. 1686e4e574b7SAdam Litke */ 1687e4e574b7SAdam Litke needed += allocated; 1688a5516438SAndi Kleen h->resv_huge_pages += delta; 1689e4e574b7SAdam Litke ret = 0; 1690a9869b83SNaoya Horiguchi 169119fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 169219fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 169319fc3f0aSAdam Litke if ((--needed) < 0) 169419fc3f0aSAdam Litke break; 1695a9869b83SNaoya Horiguchi /* 1696a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 1697a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 1698a9869b83SNaoya Horiguchi */ 1699a9869b83SNaoya Horiguchi put_page_testzero(page); 1700309381feSSasha Levin VM_BUG_ON_PAGE(page_count(page), page); 1701a5516438SAndi Kleen enqueue_huge_page(h, page); 170219fc3f0aSAdam Litke } 170328073b02SHillf Danton free: 1704b0365c8dSHillf Danton spin_unlock(&hugetlb_lock); 170519fc3f0aSAdam Litke 170619fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 1707c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 1708a9869b83SNaoya Horiguchi put_page(page); 170919fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 1710e4e574b7SAdam Litke 1711e4e574b7SAdam Litke return ret; 1712e4e574b7SAdam Litke } 1713e4e574b7SAdam Litke 1714e4e574b7SAdam Litke /* 1715e4e574b7SAdam Litke * When releasing a hugetlb pool reservation, any surplus pages that were 1716e4e574b7SAdam Litke * allocated to satisfy the reservation must be explicitly freed if they were 1717e4e574b7SAdam Litke * never used. 1718685f3457SLee Schermerhorn * Called with hugetlb_lock held. 1719e4e574b7SAdam Litke */ 1720a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 1721a5516438SAndi Kleen unsigned long unused_resv_pages) 1722e4e574b7SAdam Litke { 1723e4e574b7SAdam Litke unsigned long nr_pages; 1724e4e574b7SAdam Litke 1725ac09b3a1SAdam Litke /* Uncommit the reservation */ 1726a5516438SAndi Kleen h->resv_huge_pages -= unused_resv_pages; 1727ac09b3a1SAdam Litke 1728aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 1729bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1730aa888a74SAndi Kleen return; 1731aa888a74SAndi Kleen 1732a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 1733e4e574b7SAdam Litke 1734685f3457SLee Schermerhorn /* 1735685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 17369b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 17379b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 17389b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 17399b5e5d0fSLee Schermerhorn * free_pool_huge_page() will balance the the freed pages across the 17409b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 1741685f3457SLee Schermerhorn */ 1742685f3457SLee Schermerhorn while (nr_pages--) { 17438cebfcd0SLai Jiangshan if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) 1744685f3457SLee Schermerhorn break; 17457848a4bfSMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 1746e4e574b7SAdam Litke } 1747e4e574b7SAdam Litke } 1748e4e574b7SAdam Litke 17495e911373SMike Kravetz 1750c37f9fb1SAndy Whitcroft /* 1751feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 17525e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 1753cf3ad20bSMike Kravetz * 1754cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 1755cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 1756cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 1757cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 1758cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 1759feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 1760feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 1761feba16e2SMike Kravetz * is called in such cases. 1762cf3ad20bSMike Kravetz * 1763cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 1764cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 1765cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 1766cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 1767cf3ad20bSMike Kravetz * take appropriate action. 1768c37f9fb1SAndy Whitcroft */ 17695e911373SMike Kravetz enum vma_resv_mode { 17705e911373SMike Kravetz VMA_NEEDS_RESV, 17715e911373SMike Kravetz VMA_COMMIT_RESV, 1772feba16e2SMike Kravetz VMA_END_RESV, 17735e911373SMike Kravetz }; 1774cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 1775cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 17765e911373SMike Kravetz enum vma_resv_mode mode) 1777c37f9fb1SAndy Whitcroft { 17784e35f483SJoonsoo Kim struct resv_map *resv; 17794e35f483SJoonsoo Kim pgoff_t idx; 1780cf3ad20bSMike Kravetz long ret; 1781c37f9fb1SAndy Whitcroft 17824e35f483SJoonsoo Kim resv = vma_resv_map(vma); 17834e35f483SJoonsoo Kim if (!resv) 1784c37f9fb1SAndy Whitcroft return 1; 1785c37f9fb1SAndy Whitcroft 17864e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 17875e911373SMike Kravetz switch (mode) { 17885e911373SMike Kravetz case VMA_NEEDS_RESV: 1789cf3ad20bSMike Kravetz ret = region_chg(resv, idx, idx + 1); 17905e911373SMike Kravetz break; 17915e911373SMike Kravetz case VMA_COMMIT_RESV: 17925e911373SMike Kravetz ret = region_add(resv, idx, idx + 1); 17935e911373SMike Kravetz break; 1794feba16e2SMike Kravetz case VMA_END_RESV: 17955e911373SMike Kravetz region_abort(resv, idx, idx + 1); 17965e911373SMike Kravetz ret = 0; 17975e911373SMike Kravetz break; 17985e911373SMike Kravetz default: 17995e911373SMike Kravetz BUG(); 18005e911373SMike Kravetz } 180184afd99bSAndy Whitcroft 18024e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) 1803cf3ad20bSMike Kravetz return ret; 18044e35f483SJoonsoo Kim else 1805cf3ad20bSMike Kravetz return ret < 0 ? ret : 0; 180684afd99bSAndy Whitcroft } 1807cf3ad20bSMike Kravetz 1808cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 1809a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 1810c37f9fb1SAndy Whitcroft { 18115e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 1812cf3ad20bSMike Kravetz } 1813c37f9fb1SAndy Whitcroft 1814cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 1815cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 1816cf3ad20bSMike Kravetz { 18175e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 18185e911373SMike Kravetz } 18195e911373SMike Kravetz 1820feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 18215e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 18225e911373SMike Kravetz { 1823feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 1824c37f9fb1SAndy Whitcroft } 1825c37f9fb1SAndy Whitcroft 182670c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 182704f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 1828348ea204SAdam Litke { 182990481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 1830a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 1831348ea204SAdam Litke struct page *page; 1832d85f69b0SMike Kravetz long map_chg, map_commit; 1833d85f69b0SMike Kravetz long gbl_chg; 18346d76dcf4SAneesh Kumar K.V int ret, idx; 18356d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 18362fc39cecSAdam Litke 18376d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 1838a1e78772SMel Gorman /* 1839d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 1840d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 1841d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 1842a1e78772SMel Gorman */ 1843d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 1844d85f69b0SMike Kravetz if (map_chg < 0) 184576dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 1846d85f69b0SMike Kravetz 1847d85f69b0SMike Kravetz /* 1848d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 1849d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 1850d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 1851d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 1852d85f69b0SMike Kravetz * checked against any subpool limit. 1853d85f69b0SMike Kravetz */ 1854d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 1855d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 1856d85f69b0SMike Kravetz if (gbl_chg < 0) { 1857feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 185876dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 18595e911373SMike Kravetz } 186090d8b7e6SAdam Litke 1861d85f69b0SMike Kravetz /* 1862d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 1863d85f69b0SMike Kravetz * map, there could be reservations associated with the 1864d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 1865d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 1866d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 1867d85f69b0SMike Kravetz * the subpool reservations. 1868d85f69b0SMike Kravetz */ 1869d85f69b0SMike Kravetz if (avoid_reserve) 1870d85f69b0SMike Kravetz gbl_chg = 1; 1871d85f69b0SMike Kravetz } 1872d85f69b0SMike Kravetz 18736d76dcf4SAneesh Kumar K.V ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 18748f34af6fSJianyu Zhan if (ret) 18758f34af6fSJianyu Zhan goto out_subpool_put; 18768f34af6fSJianyu Zhan 1877a1e78772SMel Gorman spin_lock(&hugetlb_lock); 1878d85f69b0SMike Kravetz /* 1879d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 1880d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 1881d85f69b0SMike Kravetz * a reservation exists for the allocation. 1882d85f69b0SMike Kravetz */ 1883d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 188481a6fcaeSJoonsoo Kim if (!page) { 188594ae8ba7SAneesh Kumar K.V spin_unlock(&hugetlb_lock); 1886099730d6SDave Hansen page = __alloc_buddy_huge_page_with_mpol(h, vma, addr); 18878f34af6fSJianyu Zhan if (!page) 18888f34af6fSJianyu Zhan goto out_uncharge_cgroup; 18898f34af6fSJianyu Zhan 189079dbb236SAneesh Kumar K.V spin_lock(&hugetlb_lock); 189179dbb236SAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 189281a6fcaeSJoonsoo Kim /* Fall through */ 1893a1e78772SMel Gorman } 189481a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 189581a6fcaeSJoonsoo Kim spin_unlock(&hugetlb_lock); 1896a1e78772SMel Gorman 189790481622SDavid Gibson set_page_private(page, (unsigned long)spool); 1898a1e78772SMel Gorman 1899d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 1900d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 190133039678SMike Kravetz /* 190233039678SMike Kravetz * The page was added to the reservation map between 190333039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 190433039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 190533039678SMike Kravetz * Adjust for the subpool count incremented above AND 190633039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 190733039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 190833039678SMike Kravetz * no longer applies. 190933039678SMike Kravetz */ 191033039678SMike Kravetz long rsv_adjust; 191133039678SMike Kravetz 191233039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 191333039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 191433039678SMike Kravetz } 19157893d1d5SAdam Litke return page; 19168f34af6fSJianyu Zhan 19178f34af6fSJianyu Zhan out_uncharge_cgroup: 19188f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 19198f34af6fSJianyu Zhan out_subpool_put: 1920d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 19218f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 1922feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 19238f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 1924b45b5bd6SDavid Gibson } 1925b45b5bd6SDavid Gibson 192674060e4dSNaoya Horiguchi /* 192774060e4dSNaoya Horiguchi * alloc_huge_page()'s wrapper which simply returns the page if allocation 192874060e4dSNaoya Horiguchi * succeeds, otherwise NULL. This function is called from new_vma_page(), 192974060e4dSNaoya Horiguchi * where no ERR_VALUE is expected to be returned. 193074060e4dSNaoya Horiguchi */ 193174060e4dSNaoya Horiguchi struct page *alloc_huge_page_noerr(struct vm_area_struct *vma, 193274060e4dSNaoya Horiguchi unsigned long addr, int avoid_reserve) 193374060e4dSNaoya Horiguchi { 193474060e4dSNaoya Horiguchi struct page *page = alloc_huge_page(vma, addr, avoid_reserve); 193574060e4dSNaoya Horiguchi if (IS_ERR(page)) 193674060e4dSNaoya Horiguchi page = NULL; 193774060e4dSNaoya Horiguchi return page; 193874060e4dSNaoya Horiguchi } 193974060e4dSNaoya Horiguchi 194091f47662SCyrill Gorcunov int __weak alloc_bootmem_huge_page(struct hstate *h) 1941aa888a74SAndi Kleen { 1942aa888a74SAndi Kleen struct huge_bootmem_page *m; 1943b2261026SJoonsoo Kim int nr_nodes, node; 1944aa888a74SAndi Kleen 1945b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 1946aa888a74SAndi Kleen void *addr; 1947aa888a74SAndi Kleen 19488b89a116SGrygorii Strashko addr = memblock_virt_alloc_try_nid_nopanic( 19498b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 19508b89a116SGrygorii Strashko 0, BOOTMEM_ALLOC_ACCESSIBLE, node); 1951aa888a74SAndi Kleen if (addr) { 1952aa888a74SAndi Kleen /* 1953aa888a74SAndi Kleen * Use the beginning of the huge page to store the 1954aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 1955aa888a74SAndi Kleen * puts them into the mem_map). 1956aa888a74SAndi Kleen */ 1957aa888a74SAndi Kleen m = addr; 1958aa888a74SAndi Kleen goto found; 1959aa888a74SAndi Kleen } 1960aa888a74SAndi Kleen } 1961aa888a74SAndi Kleen return 0; 1962aa888a74SAndi Kleen 1963aa888a74SAndi Kleen found: 1964df994eadSLuiz Capitulino BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); 1965aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 1966aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 1967aa888a74SAndi Kleen m->hstate = h; 1968aa888a74SAndi Kleen return 1; 1969aa888a74SAndi Kleen } 1970aa888a74SAndi Kleen 1971d00181b9SKirill A. Shutemov static void __init prep_compound_huge_page(struct page *page, 1972d00181b9SKirill A. Shutemov unsigned int order) 197318229df5SAndy Whitcroft { 197418229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 197518229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 197618229df5SAndy Whitcroft else 197718229df5SAndy Whitcroft prep_compound_page(page, order); 197818229df5SAndy Whitcroft } 197918229df5SAndy Whitcroft 1980aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 1981aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 1982aa888a74SAndi Kleen { 1983aa888a74SAndi Kleen struct huge_bootmem_page *m; 1984aa888a74SAndi Kleen 1985aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 1986aa888a74SAndi Kleen struct hstate *h = m->hstate; 1987ee8f248dSBecky Bruce struct page *page; 1988ee8f248dSBecky Bruce 1989ee8f248dSBecky Bruce #ifdef CONFIG_HIGHMEM 1990ee8f248dSBecky Bruce page = pfn_to_page(m->phys >> PAGE_SHIFT); 19918b89a116SGrygorii Strashko memblock_free_late(__pa(m), 1992ee8f248dSBecky Bruce sizeof(struct huge_bootmem_page)); 1993ee8f248dSBecky Bruce #else 1994ee8f248dSBecky Bruce page = virt_to_page(m); 1995ee8f248dSBecky Bruce #endif 1996aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 199718229df5SAndy Whitcroft prep_compound_huge_page(page, h->order); 1998ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 1999aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 2000b0320c7bSRafael Aquini /* 2001b0320c7bSRafael Aquini * If we had gigantic hugepages allocated at boot time, we need 2002b0320c7bSRafael Aquini * to restore the 'stolen' pages to totalram_pages in order to 2003b0320c7bSRafael Aquini * fix confusing memory reports from free(1) and another 2004b0320c7bSRafael Aquini * side-effects, like CommitLimit going negative. 2005b0320c7bSRafael Aquini */ 2006bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 20073dcc0571SJiang Liu adjust_managed_page_count(page, 1 << h->order); 2008aa888a74SAndi Kleen } 2009aa888a74SAndi Kleen } 2010aa888a74SAndi Kleen 20118faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 20121da177e4SLinus Torvalds { 20131da177e4SLinus Torvalds unsigned long i; 20141da177e4SLinus Torvalds 2015e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 2016bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 2017aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 2018aa888a74SAndi Kleen break; 20199b5e5d0fSLee Schermerhorn } else if (!alloc_fresh_huge_page(h, 20208cebfcd0SLai Jiangshan &node_states[N_MEMORY])) 20211da177e4SLinus Torvalds break; 20221da177e4SLinus Torvalds } 20238faa8b07SAndi Kleen h->max_huge_pages = i; 2024e5ff2159SAndi Kleen } 2025e5ff2159SAndi Kleen 2026e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 2027e5ff2159SAndi Kleen { 2028e5ff2159SAndi Kleen struct hstate *h; 2029e5ff2159SAndi Kleen 2030e5ff2159SAndi Kleen for_each_hstate(h) { 2031641844f5SNaoya Horiguchi if (minimum_order > huge_page_order(h)) 2032641844f5SNaoya Horiguchi minimum_order = huge_page_order(h); 2033641844f5SNaoya Horiguchi 20348faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 2035bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 20368faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 2037e5ff2159SAndi Kleen } 2038641844f5SNaoya Horiguchi VM_BUG_ON(minimum_order == UINT_MAX); 2039e5ff2159SAndi Kleen } 2040e5ff2159SAndi Kleen 20414abd32dbSAndi Kleen static char * __init memfmt(char *buf, unsigned long n) 20424abd32dbSAndi Kleen { 20434abd32dbSAndi Kleen if (n >= (1UL << 30)) 20444abd32dbSAndi Kleen sprintf(buf, "%lu GB", n >> 30); 20454abd32dbSAndi Kleen else if (n >= (1UL << 20)) 20464abd32dbSAndi Kleen sprintf(buf, "%lu MB", n >> 20); 20474abd32dbSAndi Kleen else 20484abd32dbSAndi Kleen sprintf(buf, "%lu KB", n >> 10); 20494abd32dbSAndi Kleen return buf; 20504abd32dbSAndi Kleen } 20514abd32dbSAndi Kleen 2052e5ff2159SAndi Kleen static void __init report_hugepages(void) 2053e5ff2159SAndi Kleen { 2054e5ff2159SAndi Kleen struct hstate *h; 2055e5ff2159SAndi Kleen 2056e5ff2159SAndi Kleen for_each_hstate(h) { 20574abd32dbSAndi Kleen char buf[32]; 2058ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 20594abd32dbSAndi Kleen memfmt(buf, huge_page_size(h)), 20604abd32dbSAndi Kleen h->free_huge_pages); 2061e5ff2159SAndi Kleen } 2062e5ff2159SAndi Kleen } 2063e5ff2159SAndi Kleen 20641da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 20656ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 20666ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 20671da177e4SLinus Torvalds { 20684415cc8dSChristoph Lameter int i; 20694415cc8dSChristoph Lameter 2070bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2071aa888a74SAndi Kleen return; 2072aa888a74SAndi Kleen 20736ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 20741da177e4SLinus Torvalds struct page *page, *next; 2075a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 2076a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 2077a5516438SAndi Kleen if (count >= h->nr_huge_pages) 20786b0c880dSAdam Litke return; 20791da177e4SLinus Torvalds if (PageHighMem(page)) 20801da177e4SLinus Torvalds continue; 20811da177e4SLinus Torvalds list_del(&page->lru); 2082e5ff2159SAndi Kleen update_and_free_page(h, page); 2083a5516438SAndi Kleen h->free_huge_pages--; 2084a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 20851da177e4SLinus Torvalds } 20861da177e4SLinus Torvalds } 20871da177e4SLinus Torvalds } 20881da177e4SLinus Torvalds #else 20896ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 20906ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 20911da177e4SLinus Torvalds { 20921da177e4SLinus Torvalds } 20931da177e4SLinus Torvalds #endif 20941da177e4SLinus Torvalds 209520a0307cSWu Fengguang /* 209620a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 209720a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 209820a0307cSWu Fengguang * Returns 1 if an adjustment was made. 209920a0307cSWu Fengguang */ 21006ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 21016ae11b27SLee Schermerhorn int delta) 210220a0307cSWu Fengguang { 2103b2261026SJoonsoo Kim int nr_nodes, node; 210420a0307cSWu Fengguang 210520a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 210620a0307cSWu Fengguang 2107e8c5c824SLee Schermerhorn if (delta < 0) { 2108b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 2109b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 2110b2261026SJoonsoo Kim goto found; 2111b2261026SJoonsoo Kim } 2112b2261026SJoonsoo Kim } else { 2113b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2114b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 2115b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 2116b2261026SJoonsoo Kim goto found; 2117e8c5c824SLee Schermerhorn } 21189a76db09SLee Schermerhorn } 2119b2261026SJoonsoo Kim return 0; 212020a0307cSWu Fengguang 2121b2261026SJoonsoo Kim found: 212220a0307cSWu Fengguang h->surplus_huge_pages += delta; 2123b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 2124b2261026SJoonsoo Kim return 1; 212520a0307cSWu Fengguang } 212620a0307cSWu Fengguang 2127a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 21286ae11b27SLee Schermerhorn static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, 21296ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 21301da177e4SLinus Torvalds { 21317893d1d5SAdam Litke unsigned long min_count, ret; 21321da177e4SLinus Torvalds 2133944d9fecSLuiz Capitulino if (hstate_is_gigantic(h) && !gigantic_page_supported()) 2134aa888a74SAndi Kleen return h->max_huge_pages; 2135aa888a74SAndi Kleen 21367893d1d5SAdam Litke /* 21377893d1d5SAdam Litke * Increase the pool size 21387893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 21397893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 2140d1c3fb1fSNishanth Aravamudan * 2141d1c3fb1fSNishanth Aravamudan * We might race with alloc_buddy_huge_page() here and be unable 2142d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 2143d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 2144d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 2145d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 21467893d1d5SAdam Litke */ 21471da177e4SLinus Torvalds spin_lock(&hugetlb_lock); 2148a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 21496ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 21507893d1d5SAdam Litke break; 21517893d1d5SAdam Litke } 21527893d1d5SAdam Litke 2153a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 21547893d1d5SAdam Litke /* 21557893d1d5SAdam Litke * If this allocation races such that we no longer need the 21567893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 21577893d1d5SAdam Litke * and reducing the surplus. 21587893d1d5SAdam Litke */ 21597893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 2160944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) 2161944d9fecSLuiz Capitulino ret = alloc_fresh_gigantic_page(h, nodes_allowed); 2162944d9fecSLuiz Capitulino else 21636ae11b27SLee Schermerhorn ret = alloc_fresh_huge_page(h, nodes_allowed); 21647893d1d5SAdam Litke spin_lock(&hugetlb_lock); 21657893d1d5SAdam Litke if (!ret) 21667893d1d5SAdam Litke goto out; 21677893d1d5SAdam Litke 2168536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 2169536240f2SMel Gorman if (signal_pending(current)) 2170536240f2SMel Gorman goto out; 21717893d1d5SAdam Litke } 21727893d1d5SAdam Litke 21737893d1d5SAdam Litke /* 21747893d1d5SAdam Litke * Decrease the pool size 21757893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 21767893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 21777893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 21787893d1d5SAdam Litke * to the desired size as pages become free. 2179d1c3fb1fSNishanth Aravamudan * 2180d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 2181d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 2182d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 2183d1c3fb1fSNishanth Aravamudan * alloc_buddy_huge_page() is checking the global counter, 2184d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 2185d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 2186d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 21877893d1d5SAdam Litke */ 2188a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 21896b0c880dSAdam Litke min_count = max(count, min_count); 21906ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 2191a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 21926ae11b27SLee Schermerhorn if (!free_pool_huge_page(h, nodes_allowed, 0)) 21931da177e4SLinus Torvalds break; 219455f67141SMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 21951da177e4SLinus Torvalds } 2196a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 21976ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 21987893d1d5SAdam Litke break; 21997893d1d5SAdam Litke } 22007893d1d5SAdam Litke out: 2201a5516438SAndi Kleen ret = persistent_huge_pages(h); 22021da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 22037893d1d5SAdam Litke return ret; 22041da177e4SLinus Torvalds } 22051da177e4SLinus Torvalds 2206a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 2207a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 2208a3437870SNishanth Aravamudan 2209a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 2210a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 2211a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 2212a3437870SNishanth Aravamudan 2213a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 2214a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 2215a3437870SNishanth Aravamudan 22169a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 22179a305230SLee Schermerhorn 22189a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 2219a3437870SNishanth Aravamudan { 2220a3437870SNishanth Aravamudan int i; 22219a305230SLee Schermerhorn 2222a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 22239a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 22249a305230SLee Schermerhorn if (nidp) 22259a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 2226a3437870SNishanth Aravamudan return &hstates[i]; 22279a305230SLee Schermerhorn } 22289a305230SLee Schermerhorn 22299a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 2230a3437870SNishanth Aravamudan } 2231a3437870SNishanth Aravamudan 223206808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 2233a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2234a3437870SNishanth Aravamudan { 22359a305230SLee Schermerhorn struct hstate *h; 22369a305230SLee Schermerhorn unsigned long nr_huge_pages; 22379a305230SLee Schermerhorn int nid; 22389a305230SLee Schermerhorn 22399a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 22409a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 22419a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 22429a305230SLee Schermerhorn else 22439a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 22449a305230SLee Schermerhorn 22459a305230SLee Schermerhorn return sprintf(buf, "%lu\n", nr_huge_pages); 2246a3437870SNishanth Aravamudan } 2247adbe8726SEric B Munson 2248238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 2249238d3c13SDavid Rientjes struct hstate *h, int nid, 2250238d3c13SDavid Rientjes unsigned long count, size_t len) 2251a3437870SNishanth Aravamudan { 2252a3437870SNishanth Aravamudan int err; 2253bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); 2254a3437870SNishanth Aravamudan 2255944d9fecSLuiz Capitulino if (hstate_is_gigantic(h) && !gigantic_page_supported()) { 2256adbe8726SEric B Munson err = -EINVAL; 2257adbe8726SEric B Munson goto out; 2258adbe8726SEric B Munson } 2259adbe8726SEric B Munson 22609a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 22619a305230SLee Schermerhorn /* 22629a305230SLee Schermerhorn * global hstate attribute 22639a305230SLee Schermerhorn */ 22649a305230SLee Schermerhorn if (!(obey_mempolicy && 22659a305230SLee Schermerhorn init_nodemask_of_mempolicy(nodes_allowed))) { 226606808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 22678cebfcd0SLai Jiangshan nodes_allowed = &node_states[N_MEMORY]; 226806808b08SLee Schermerhorn } 22699a305230SLee Schermerhorn } else if (nodes_allowed) { 22709a305230SLee Schermerhorn /* 22719a305230SLee Schermerhorn * per node hstate attribute: adjust count to global, 22729a305230SLee Schermerhorn * but restrict alloc/free to the specified node. 22739a305230SLee Schermerhorn */ 22749a305230SLee Schermerhorn count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 22759a305230SLee Schermerhorn init_nodemask_of_node(nodes_allowed, nid); 22769a305230SLee Schermerhorn } else 22778cebfcd0SLai Jiangshan nodes_allowed = &node_states[N_MEMORY]; 22789a305230SLee Schermerhorn 227906808b08SLee Schermerhorn h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); 2280a3437870SNishanth Aravamudan 22818cebfcd0SLai Jiangshan if (nodes_allowed != &node_states[N_MEMORY]) 228206808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 228306808b08SLee Schermerhorn 228406808b08SLee Schermerhorn return len; 2285adbe8726SEric B Munson out: 2286adbe8726SEric B Munson NODEMASK_FREE(nodes_allowed); 2287adbe8726SEric B Munson return err; 228806808b08SLee Schermerhorn } 228906808b08SLee Schermerhorn 2290238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 2291238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 2292238d3c13SDavid Rientjes size_t len) 2293238d3c13SDavid Rientjes { 2294238d3c13SDavid Rientjes struct hstate *h; 2295238d3c13SDavid Rientjes unsigned long count; 2296238d3c13SDavid Rientjes int nid; 2297238d3c13SDavid Rientjes int err; 2298238d3c13SDavid Rientjes 2299238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 2300238d3c13SDavid Rientjes if (err) 2301238d3c13SDavid Rientjes return err; 2302238d3c13SDavid Rientjes 2303238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 2304238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 2305238d3c13SDavid Rientjes } 2306238d3c13SDavid Rientjes 230706808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 230806808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 230906808b08SLee Schermerhorn { 231006808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 231106808b08SLee Schermerhorn } 231206808b08SLee Schermerhorn 231306808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 231406808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 231506808b08SLee Schermerhorn { 2316238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 2317a3437870SNishanth Aravamudan } 2318a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 2319a3437870SNishanth Aravamudan 232006808b08SLee Schermerhorn #ifdef CONFIG_NUMA 232106808b08SLee Schermerhorn 232206808b08SLee Schermerhorn /* 232306808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 232406808b08SLee Schermerhorn * huge page alloc/free. 232506808b08SLee Schermerhorn */ 232606808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 232706808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 232806808b08SLee Schermerhorn { 232906808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 233006808b08SLee Schermerhorn } 233106808b08SLee Schermerhorn 233206808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 233306808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 233406808b08SLee Schermerhorn { 2335238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 233606808b08SLee Schermerhorn } 233706808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 233806808b08SLee Schermerhorn #endif 233906808b08SLee Schermerhorn 234006808b08SLee Schermerhorn 2341a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 2342a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2343a3437870SNishanth Aravamudan { 23449a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2345a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 2346a3437870SNishanth Aravamudan } 2347adbe8726SEric B Munson 2348a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 2349a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 2350a3437870SNishanth Aravamudan { 2351a3437870SNishanth Aravamudan int err; 2352a3437870SNishanth Aravamudan unsigned long input; 23539a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2354a3437870SNishanth Aravamudan 2355bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2356adbe8726SEric B Munson return -EINVAL; 2357adbe8726SEric B Munson 23583dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 2359a3437870SNishanth Aravamudan if (err) 236073ae31e5SEric B Munson return err; 2361a3437870SNishanth Aravamudan 2362a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 2363a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 2364a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 2365a3437870SNishanth Aravamudan 2366a3437870SNishanth Aravamudan return count; 2367a3437870SNishanth Aravamudan } 2368a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 2369a3437870SNishanth Aravamudan 2370a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 2371a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2372a3437870SNishanth Aravamudan { 23739a305230SLee Schermerhorn struct hstate *h; 23749a305230SLee Schermerhorn unsigned long free_huge_pages; 23759a305230SLee Schermerhorn int nid; 23769a305230SLee Schermerhorn 23779a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 23789a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 23799a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 23809a305230SLee Schermerhorn else 23819a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 23829a305230SLee Schermerhorn 23839a305230SLee Schermerhorn return sprintf(buf, "%lu\n", free_huge_pages); 2384a3437870SNishanth Aravamudan } 2385a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 2386a3437870SNishanth Aravamudan 2387a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 2388a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2389a3437870SNishanth Aravamudan { 23909a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2391a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->resv_huge_pages); 2392a3437870SNishanth Aravamudan } 2393a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 2394a3437870SNishanth Aravamudan 2395a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 2396a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2397a3437870SNishanth Aravamudan { 23989a305230SLee Schermerhorn struct hstate *h; 23999a305230SLee Schermerhorn unsigned long surplus_huge_pages; 24009a305230SLee Schermerhorn int nid; 24019a305230SLee Schermerhorn 24029a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 24039a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 24049a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 24059a305230SLee Schermerhorn else 24069a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 24079a305230SLee Schermerhorn 24089a305230SLee Schermerhorn return sprintf(buf, "%lu\n", surplus_huge_pages); 2409a3437870SNishanth Aravamudan } 2410a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 2411a3437870SNishanth Aravamudan 2412a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 2413a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 2414a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 2415a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 2416a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 2417a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 241806808b08SLee Schermerhorn #ifdef CONFIG_NUMA 241906808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 242006808b08SLee Schermerhorn #endif 2421a3437870SNishanth Aravamudan NULL, 2422a3437870SNishanth Aravamudan }; 2423a3437870SNishanth Aravamudan 2424a3437870SNishanth Aravamudan static struct attribute_group hstate_attr_group = { 2425a3437870SNishanth Aravamudan .attrs = hstate_attrs, 2426a3437870SNishanth Aravamudan }; 2427a3437870SNishanth Aravamudan 2428094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 24299a305230SLee Schermerhorn struct kobject **hstate_kobjs, 24309a305230SLee Schermerhorn struct attribute_group *hstate_attr_group) 2431a3437870SNishanth Aravamudan { 2432a3437870SNishanth Aravamudan int retval; 2433972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 2434a3437870SNishanth Aravamudan 24359a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 24369a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 2437a3437870SNishanth Aravamudan return -ENOMEM; 2438a3437870SNishanth Aravamudan 24399a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 2440a3437870SNishanth Aravamudan if (retval) 24419a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 2442a3437870SNishanth Aravamudan 2443a3437870SNishanth Aravamudan return retval; 2444a3437870SNishanth Aravamudan } 2445a3437870SNishanth Aravamudan 2446a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 2447a3437870SNishanth Aravamudan { 2448a3437870SNishanth Aravamudan struct hstate *h; 2449a3437870SNishanth Aravamudan int err; 2450a3437870SNishanth Aravamudan 2451a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 2452a3437870SNishanth Aravamudan if (!hugepages_kobj) 2453a3437870SNishanth Aravamudan return; 2454a3437870SNishanth Aravamudan 2455a3437870SNishanth Aravamudan for_each_hstate(h) { 24569a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 24579a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 2458a3437870SNishanth Aravamudan if (err) 2459ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s", h->name); 2460a3437870SNishanth Aravamudan } 2461a3437870SNishanth Aravamudan } 2462a3437870SNishanth Aravamudan 24639a305230SLee Schermerhorn #ifdef CONFIG_NUMA 24649a305230SLee Schermerhorn 24659a305230SLee Schermerhorn /* 24669a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 246710fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 246810fbcf4cSKay Sievers * index of a node device or _hstate == node id. 246910fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 24709a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 24719a305230SLee Schermerhorn */ 24729a305230SLee Schermerhorn struct node_hstate { 24739a305230SLee Schermerhorn struct kobject *hugepages_kobj; 24749a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 24759a305230SLee Schermerhorn }; 2476b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 24779a305230SLee Schermerhorn 24789a305230SLee Schermerhorn /* 247910fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 24809a305230SLee Schermerhorn */ 24819a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 24829a305230SLee Schermerhorn &nr_hugepages_attr.attr, 24839a305230SLee Schermerhorn &free_hugepages_attr.attr, 24849a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 24859a305230SLee Schermerhorn NULL, 24869a305230SLee Schermerhorn }; 24879a305230SLee Schermerhorn 24889a305230SLee Schermerhorn static struct attribute_group per_node_hstate_attr_group = { 24899a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 24909a305230SLee Schermerhorn }; 24919a305230SLee Schermerhorn 24929a305230SLee Schermerhorn /* 249310fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 24949a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 24959a305230SLee Schermerhorn */ 24969a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 24979a305230SLee Schermerhorn { 24989a305230SLee Schermerhorn int nid; 24999a305230SLee Schermerhorn 25009a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 25019a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 25029a305230SLee Schermerhorn int i; 25039a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 25049a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 25059a305230SLee Schermerhorn if (nidp) 25069a305230SLee Schermerhorn *nidp = nid; 25079a305230SLee Schermerhorn return &hstates[i]; 25089a305230SLee Schermerhorn } 25099a305230SLee Schermerhorn } 25109a305230SLee Schermerhorn 25119a305230SLee Schermerhorn BUG(); 25129a305230SLee Schermerhorn return NULL; 25139a305230SLee Schermerhorn } 25149a305230SLee Schermerhorn 25159a305230SLee Schermerhorn /* 251610fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 25179a305230SLee Schermerhorn * No-op if no hstate attributes attached. 25189a305230SLee Schermerhorn */ 25193cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 25209a305230SLee Schermerhorn { 25219a305230SLee Schermerhorn struct hstate *h; 252210fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 25239a305230SLee Schermerhorn 25249a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 25259b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 25269a305230SLee Schermerhorn 2527972dc4deSAneesh Kumar K.V for_each_hstate(h) { 2528972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 2529972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 2530972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 2531972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 2532972dc4deSAneesh Kumar K.V } 25339a305230SLee Schermerhorn } 25349a305230SLee Schermerhorn 25359a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 25369a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 25379a305230SLee Schermerhorn } 25389a305230SLee Schermerhorn 25399a305230SLee Schermerhorn /* 254010fbcf4cSKay Sievers * hugetlb module exit: unregister hstate attributes from node devices 25419a305230SLee Schermerhorn * that have them. 25429a305230SLee Schermerhorn */ 25439a305230SLee Schermerhorn static void hugetlb_unregister_all_nodes(void) 25449a305230SLee Schermerhorn { 25459a305230SLee Schermerhorn int nid; 25469a305230SLee Schermerhorn 25479a305230SLee Schermerhorn /* 254810fbcf4cSKay Sievers * disable node device registrations. 25499a305230SLee Schermerhorn */ 25509a305230SLee Schermerhorn register_hugetlbfs_with_node(NULL, NULL); 25519a305230SLee Schermerhorn 25529a305230SLee Schermerhorn /* 25539a305230SLee Schermerhorn * remove hstate attributes from any nodes that have them. 25549a305230SLee Schermerhorn */ 25559a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) 25568732794bSWen Congyang hugetlb_unregister_node(node_devices[nid]); 25579a305230SLee Schermerhorn } 25589a305230SLee Schermerhorn 25599a305230SLee Schermerhorn /* 256010fbcf4cSKay Sievers * Register hstate attributes for a single node device. 25619a305230SLee Schermerhorn * No-op if attributes already registered. 25629a305230SLee Schermerhorn */ 25633cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 25649a305230SLee Schermerhorn { 25659a305230SLee Schermerhorn struct hstate *h; 256610fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 25679a305230SLee Schermerhorn int err; 25689a305230SLee Schermerhorn 25699a305230SLee Schermerhorn if (nhs->hugepages_kobj) 25709a305230SLee Schermerhorn return; /* already allocated */ 25719a305230SLee Schermerhorn 25729a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 257310fbcf4cSKay Sievers &node->dev.kobj); 25749a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 25759a305230SLee Schermerhorn return; 25769a305230SLee Schermerhorn 25779a305230SLee Schermerhorn for_each_hstate(h) { 25789a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 25799a305230SLee Schermerhorn nhs->hstate_kobjs, 25809a305230SLee Schermerhorn &per_node_hstate_attr_group); 25819a305230SLee Schermerhorn if (err) { 2582ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s for node %d\n", 258310fbcf4cSKay Sievers h->name, node->dev.id); 25849a305230SLee Schermerhorn hugetlb_unregister_node(node); 25859a305230SLee Schermerhorn break; 25869a305230SLee Schermerhorn } 25879a305230SLee Schermerhorn } 25889a305230SLee Schermerhorn } 25899a305230SLee Schermerhorn 25909a305230SLee Schermerhorn /* 25919b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 259210fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 259310fbcf4cSKay Sievers * registered their associated device by this time. 25949a305230SLee Schermerhorn */ 25957d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 25969a305230SLee Schermerhorn { 25979a305230SLee Schermerhorn int nid; 25989a305230SLee Schermerhorn 25998cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 26008732794bSWen Congyang struct node *node = node_devices[nid]; 260110fbcf4cSKay Sievers if (node->dev.id == nid) 26029a305230SLee Schermerhorn hugetlb_register_node(node); 26039a305230SLee Schermerhorn } 26049a305230SLee Schermerhorn 26059a305230SLee Schermerhorn /* 260610fbcf4cSKay Sievers * Let the node device driver know we're here so it can 26079a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 26089a305230SLee Schermerhorn */ 26099a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 26109a305230SLee Schermerhorn hugetlb_unregister_node); 26119a305230SLee Schermerhorn } 26129a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 26139a305230SLee Schermerhorn 26149a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 26159a305230SLee Schermerhorn { 26169a305230SLee Schermerhorn BUG(); 26179a305230SLee Schermerhorn if (nidp) 26189a305230SLee Schermerhorn *nidp = -1; 26199a305230SLee Schermerhorn return NULL; 26209a305230SLee Schermerhorn } 26219a305230SLee Schermerhorn 26229a305230SLee Schermerhorn static void hugetlb_unregister_all_nodes(void) { } 26239a305230SLee Schermerhorn 26249a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 26259a305230SLee Schermerhorn 26269a305230SLee Schermerhorn #endif 26279a305230SLee Schermerhorn 2628a3437870SNishanth Aravamudan static void __exit hugetlb_exit(void) 2629a3437870SNishanth Aravamudan { 2630a3437870SNishanth Aravamudan struct hstate *h; 2631a3437870SNishanth Aravamudan 26329a305230SLee Schermerhorn hugetlb_unregister_all_nodes(); 26339a305230SLee Schermerhorn 2634a3437870SNishanth Aravamudan for_each_hstate(h) { 2635972dc4deSAneesh Kumar K.V kobject_put(hstate_kobjs[hstate_index(h)]); 2636a3437870SNishanth Aravamudan } 2637a3437870SNishanth Aravamudan 2638a3437870SNishanth Aravamudan kobject_put(hugepages_kobj); 2639c672c7f2SMike Kravetz kfree(hugetlb_fault_mutex_table); 2640a3437870SNishanth Aravamudan } 2641a3437870SNishanth Aravamudan module_exit(hugetlb_exit); 2642a3437870SNishanth Aravamudan 2643a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 2644a3437870SNishanth Aravamudan { 26458382d914SDavidlohr Bueso int i; 26468382d914SDavidlohr Bueso 2647457c1b27SNishanth Aravamudan if (!hugepages_supported()) 26480ef89d25SBenjamin Herrenschmidt return 0; 2649a3437870SNishanth Aravamudan 2650e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) { 2651e11bfbfcSNick Piggin default_hstate_size = HPAGE_SIZE; 2652e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) 2653a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 2654a3437870SNishanth Aravamudan } 2655972dc4deSAneesh Kumar K.V default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); 2656e11bfbfcSNick Piggin if (default_hstate_max_huge_pages) 2657e11bfbfcSNick Piggin default_hstate.max_huge_pages = default_hstate_max_huge_pages; 2658a3437870SNishanth Aravamudan 2659a3437870SNishanth Aravamudan hugetlb_init_hstates(); 2660aa888a74SAndi Kleen gather_bootmem_prealloc(); 2661a3437870SNishanth Aravamudan report_hugepages(); 2662a3437870SNishanth Aravamudan 2663a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 26649a305230SLee Schermerhorn hugetlb_register_all_nodes(); 26657179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 26669a305230SLee Schermerhorn 26678382d914SDavidlohr Bueso #ifdef CONFIG_SMP 26688382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 26698382d914SDavidlohr Bueso #else 26708382d914SDavidlohr Bueso num_fault_mutexes = 1; 26718382d914SDavidlohr Bueso #endif 2672c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 26738382d914SDavidlohr Bueso kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL); 2674c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 26758382d914SDavidlohr Bueso 26768382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 2677c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 2678a3437870SNishanth Aravamudan return 0; 2679a3437870SNishanth Aravamudan } 2680a3437870SNishanth Aravamudan module_init(hugetlb_init); 2681a3437870SNishanth Aravamudan 2682a3437870SNishanth Aravamudan /* Should be called on processing a hugepagesz=... option */ 2683d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 2684a3437870SNishanth Aravamudan { 2685a3437870SNishanth Aravamudan struct hstate *h; 26868faa8b07SAndi Kleen unsigned long i; 26878faa8b07SAndi Kleen 2688a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 2689ffb22af5SAndrew Morton pr_warning("hugepagesz= specified twice, ignoring\n"); 2690a3437870SNishanth Aravamudan return; 2691a3437870SNishanth Aravamudan } 269247d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 2693a3437870SNishanth Aravamudan BUG_ON(order == 0); 269447d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 2695a3437870SNishanth Aravamudan h->order = order; 2696a3437870SNishanth Aravamudan h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 26978faa8b07SAndi Kleen h->nr_huge_pages = 0; 26988faa8b07SAndi Kleen h->free_huge_pages = 0; 26998faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 27008faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 27010edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 27028cebfcd0SLai Jiangshan h->next_nid_to_alloc = first_node(node_states[N_MEMORY]); 27038cebfcd0SLai Jiangshan h->next_nid_to_free = first_node(node_states[N_MEMORY]); 2704a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 2705a3437870SNishanth Aravamudan huge_page_size(h)/1024); 27068faa8b07SAndi Kleen 2707a3437870SNishanth Aravamudan parsed_hstate = h; 2708a3437870SNishanth Aravamudan } 2709a3437870SNishanth Aravamudan 2710e11bfbfcSNick Piggin static int __init hugetlb_nrpages_setup(char *s) 2711a3437870SNishanth Aravamudan { 2712a3437870SNishanth Aravamudan unsigned long *mhp; 27138faa8b07SAndi Kleen static unsigned long *last_mhp; 2714a3437870SNishanth Aravamudan 2715a3437870SNishanth Aravamudan /* 271647d38344SAneesh Kumar K.V * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, 2717a3437870SNishanth Aravamudan * so this hugepages= parameter goes to the "default hstate". 2718a3437870SNishanth Aravamudan */ 271947d38344SAneesh Kumar K.V if (!hugetlb_max_hstate) 2720a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 2721a3437870SNishanth Aravamudan else 2722a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 2723a3437870SNishanth Aravamudan 27248faa8b07SAndi Kleen if (mhp == last_mhp) { 2725ffb22af5SAndrew Morton pr_warning("hugepages= specified twice without " 27268faa8b07SAndi Kleen "interleaving hugepagesz=, ignoring\n"); 27278faa8b07SAndi Kleen return 1; 27288faa8b07SAndi Kleen } 27298faa8b07SAndi Kleen 2730a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 2731a3437870SNishanth Aravamudan *mhp = 0; 2732a3437870SNishanth Aravamudan 27338faa8b07SAndi Kleen /* 27348faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 27358faa8b07SAndi Kleen * But we need to allocate >= MAX_ORDER hstates here early to still 27368faa8b07SAndi Kleen * use the bootmem allocator. 27378faa8b07SAndi Kleen */ 273847d38344SAneesh Kumar K.V if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) 27398faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 27408faa8b07SAndi Kleen 27418faa8b07SAndi Kleen last_mhp = mhp; 27428faa8b07SAndi Kleen 2743a3437870SNishanth Aravamudan return 1; 2744a3437870SNishanth Aravamudan } 2745e11bfbfcSNick Piggin __setup("hugepages=", hugetlb_nrpages_setup); 2746e11bfbfcSNick Piggin 2747e11bfbfcSNick Piggin static int __init hugetlb_default_setup(char *s) 2748e11bfbfcSNick Piggin { 2749e11bfbfcSNick Piggin default_hstate_size = memparse(s, &s); 2750e11bfbfcSNick Piggin return 1; 2751e11bfbfcSNick Piggin } 2752e11bfbfcSNick Piggin __setup("default_hugepagesz=", hugetlb_default_setup); 2753a3437870SNishanth Aravamudan 27548a213460SNishanth Aravamudan static unsigned int cpuset_mems_nr(unsigned int *array) 27558a213460SNishanth Aravamudan { 27568a213460SNishanth Aravamudan int node; 27578a213460SNishanth Aravamudan unsigned int nr = 0; 27588a213460SNishanth Aravamudan 27598a213460SNishanth Aravamudan for_each_node_mask(node, cpuset_current_mems_allowed) 27608a213460SNishanth Aravamudan nr += array[node]; 27618a213460SNishanth Aravamudan 27628a213460SNishanth Aravamudan return nr; 27638a213460SNishanth Aravamudan } 27648a213460SNishanth Aravamudan 27658a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 276606808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 276706808b08SLee Schermerhorn struct ctl_table *table, int write, 276806808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 27691da177e4SLinus Torvalds { 2770e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 2771238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 277208d4a246SMichal Hocko int ret; 2773e5ff2159SAndi Kleen 2774457c1b27SNishanth Aravamudan if (!hugepages_supported()) 2775457c1b27SNishanth Aravamudan return -ENOTSUPP; 2776457c1b27SNishanth Aravamudan 2777e5ff2159SAndi Kleen table->data = &tmp; 2778e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 277908d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 278008d4a246SMichal Hocko if (ret) 278108d4a246SMichal Hocko goto out; 2782e5ff2159SAndi Kleen 2783238d3c13SDavid Rientjes if (write) 2784238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 2785238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 278608d4a246SMichal Hocko out: 278708d4a246SMichal Hocko return ret; 27881da177e4SLinus Torvalds } 2789396faf03SMel Gorman 279006808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 279106808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 279206808b08SLee Schermerhorn { 279306808b08SLee Schermerhorn 279406808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 279506808b08SLee Schermerhorn buffer, length, ppos); 279606808b08SLee Schermerhorn } 279706808b08SLee Schermerhorn 279806808b08SLee Schermerhorn #ifdef CONFIG_NUMA 279906808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 280006808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 280106808b08SLee Schermerhorn { 280206808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 280306808b08SLee Schermerhorn buffer, length, ppos); 280406808b08SLee Schermerhorn } 280506808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 280606808b08SLee Schermerhorn 2807a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 28088d65af78SAlexey Dobriyan void __user *buffer, 2809a3d0c6aaSNishanth Aravamudan size_t *length, loff_t *ppos) 2810a3d0c6aaSNishanth Aravamudan { 2811a5516438SAndi Kleen struct hstate *h = &default_hstate; 2812e5ff2159SAndi Kleen unsigned long tmp; 281308d4a246SMichal Hocko int ret; 2814e5ff2159SAndi Kleen 2815457c1b27SNishanth Aravamudan if (!hugepages_supported()) 2816457c1b27SNishanth Aravamudan return -ENOTSUPP; 2817457c1b27SNishanth Aravamudan 2818e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 2819e5ff2159SAndi Kleen 2820bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 2821adbe8726SEric B Munson return -EINVAL; 2822adbe8726SEric B Munson 2823e5ff2159SAndi Kleen table->data = &tmp; 2824e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 282508d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 282608d4a246SMichal Hocko if (ret) 282708d4a246SMichal Hocko goto out; 2828e5ff2159SAndi Kleen 2829e5ff2159SAndi Kleen if (write) { 2830064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 2831e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 2832a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 2833e5ff2159SAndi Kleen } 283408d4a246SMichal Hocko out: 283508d4a246SMichal Hocko return ret; 2836a3d0c6aaSNishanth Aravamudan } 2837a3d0c6aaSNishanth Aravamudan 28381da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 28391da177e4SLinus Torvalds 2840e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 28411da177e4SLinus Torvalds { 2842a5516438SAndi Kleen struct hstate *h = &default_hstate; 2843457c1b27SNishanth Aravamudan if (!hugepages_supported()) 2844457c1b27SNishanth Aravamudan return; 2845e1759c21SAlexey Dobriyan seq_printf(m, 28461da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 28471da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 2848b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 28497893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 28504f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 2851a5516438SAndi Kleen h->nr_huge_pages, 2852a5516438SAndi Kleen h->free_huge_pages, 2853a5516438SAndi Kleen h->resv_huge_pages, 2854a5516438SAndi Kleen h->surplus_huge_pages, 2855a5516438SAndi Kleen 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 28561da177e4SLinus Torvalds } 28571da177e4SLinus Torvalds 28581da177e4SLinus Torvalds int hugetlb_report_node_meminfo(int nid, char *buf) 28591da177e4SLinus Torvalds { 2860a5516438SAndi Kleen struct hstate *h = &default_hstate; 2861457c1b27SNishanth Aravamudan if (!hugepages_supported()) 2862457c1b27SNishanth Aravamudan return 0; 28631da177e4SLinus Torvalds return sprintf(buf, 28641da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 2865a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 2866a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 2867a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 2868a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 2869a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 28701da177e4SLinus Torvalds } 28711da177e4SLinus Torvalds 2872949f7ec5SDavid Rientjes void hugetlb_show_meminfo(void) 2873949f7ec5SDavid Rientjes { 2874949f7ec5SDavid Rientjes struct hstate *h; 2875949f7ec5SDavid Rientjes int nid; 2876949f7ec5SDavid Rientjes 2877457c1b27SNishanth Aravamudan if (!hugepages_supported()) 2878457c1b27SNishanth Aravamudan return; 2879457c1b27SNishanth Aravamudan 2880949f7ec5SDavid Rientjes for_each_node_state(nid, N_MEMORY) 2881949f7ec5SDavid Rientjes for_each_hstate(h) 2882949f7ec5SDavid Rientjes pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 2883949f7ec5SDavid Rientjes nid, 2884949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 2885949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 2886949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 2887949f7ec5SDavid Rientjes 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 2888949f7ec5SDavid Rientjes } 2889949f7ec5SDavid Rientjes 28905d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 28915d317b2bSNaoya Horiguchi { 28925d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 28935d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 28945d317b2bSNaoya Horiguchi } 28955d317b2bSNaoya Horiguchi 28961da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 28971da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 28981da177e4SLinus Torvalds { 2899d0028588SWanpeng Li struct hstate *h; 2900d0028588SWanpeng Li unsigned long nr_total_pages = 0; 2901d0028588SWanpeng Li 2902d0028588SWanpeng Li for_each_hstate(h) 2903d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 2904d0028588SWanpeng Li return nr_total_pages; 29051da177e4SLinus Torvalds } 29061da177e4SLinus Torvalds 2907a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 2908fc1b8a73SMel Gorman { 2909fc1b8a73SMel Gorman int ret = -ENOMEM; 2910fc1b8a73SMel Gorman 2911fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 2912fc1b8a73SMel Gorman /* 2913fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 2914fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 2915fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 2916fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 2917fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 2918fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 2919fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 2920fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 2921fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 2922fc1b8a73SMel Gorman * 2923fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 2924fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 2925fc1b8a73SMel Gorman * we fall back to check against current free page availability as 2926fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 2927fc1b8a73SMel Gorman * semantics that cpuset has. 2928fc1b8a73SMel Gorman */ 2929fc1b8a73SMel Gorman if (delta > 0) { 2930a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 2931fc1b8a73SMel Gorman goto out; 2932fc1b8a73SMel Gorman 2933a5516438SAndi Kleen if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 2934a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 2935fc1b8a73SMel Gorman goto out; 2936fc1b8a73SMel Gorman } 2937fc1b8a73SMel Gorman } 2938fc1b8a73SMel Gorman 2939fc1b8a73SMel Gorman ret = 0; 2940fc1b8a73SMel Gorman if (delta < 0) 2941a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 2942fc1b8a73SMel Gorman 2943fc1b8a73SMel Gorman out: 2944fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 2945fc1b8a73SMel Gorman return ret; 2946fc1b8a73SMel Gorman } 2947fc1b8a73SMel Gorman 294884afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 294984afd99bSAndy Whitcroft { 2950f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 295184afd99bSAndy Whitcroft 295284afd99bSAndy Whitcroft /* 295384afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 295484afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 295584afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 295625985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 295784afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 295884afd99bSAndy Whitcroft * new reference here without additional locking. 295984afd99bSAndy Whitcroft */ 29604e35f483SJoonsoo Kim if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 2961f522c3acSJoonsoo Kim kref_get(&resv->refs); 296284afd99bSAndy Whitcroft } 296384afd99bSAndy Whitcroft 2964a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 2965a1e78772SMel Gorman { 2966a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2967f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 296890481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 29694e35f483SJoonsoo Kim unsigned long reserve, start, end; 29701c5ecae3SMike Kravetz long gbl_reserve; 297184afd99bSAndy Whitcroft 29724e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 29734e35f483SJoonsoo Kim return; 29744e35f483SJoonsoo Kim 2975a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 2976a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 297784afd99bSAndy Whitcroft 29784e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 297984afd99bSAndy Whitcroft 2980f031dd27SJoonsoo Kim kref_put(&resv->refs, resv_map_release); 298184afd99bSAndy Whitcroft 29827251ff78SAdam Litke if (reserve) { 29831c5ecae3SMike Kravetz /* 29841c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 29851c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 29861c5ecae3SMike Kravetz */ 29871c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 29881c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 29897251ff78SAdam Litke } 2990a1e78772SMel Gorman } 2991a1e78772SMel Gorman 29921da177e4SLinus Torvalds /* 29931da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 29941da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 29951da177e4SLinus Torvalds * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 29961da177e4SLinus Torvalds * this far. 29971da177e4SLinus Torvalds */ 2998d0217ac0SNick Piggin static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) 29991da177e4SLinus Torvalds { 30001da177e4SLinus Torvalds BUG(); 3001d0217ac0SNick Piggin return 0; 30021da177e4SLinus Torvalds } 30031da177e4SLinus Torvalds 3004f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 3005d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 300684afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 3007a1e78772SMel Gorman .close = hugetlb_vm_op_close, 30081da177e4SLinus Torvalds }; 30091da177e4SLinus Torvalds 30101e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 30111e8f889bSDavid Gibson int writable) 301263551ae0SDavid Gibson { 301363551ae0SDavid Gibson pte_t entry; 301463551ae0SDavid Gibson 30151e8f889bSDavid Gibson if (writable) { 3016106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 3017106c992aSGerald Schaefer vma->vm_page_prot))); 301863551ae0SDavid Gibson } else { 3019106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 3020106c992aSGerald Schaefer vma->vm_page_prot)); 302163551ae0SDavid Gibson } 302263551ae0SDavid Gibson entry = pte_mkyoung(entry); 302363551ae0SDavid Gibson entry = pte_mkhuge(entry); 3024d9ed9faaSChris Metcalf entry = arch_make_huge_pte(entry, vma, page, writable); 302563551ae0SDavid Gibson 302663551ae0SDavid Gibson return entry; 302763551ae0SDavid Gibson } 302863551ae0SDavid Gibson 30291e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 30301e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 30311e8f889bSDavid Gibson { 30321e8f889bSDavid Gibson pte_t entry; 30331e8f889bSDavid Gibson 3034106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 303532f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 30364b3073e1SRussell King update_mmu_cache(vma, address, ptep); 30371e8f889bSDavid Gibson } 30381e8f889bSDavid Gibson 30394a705fefSNaoya Horiguchi static int is_hugetlb_entry_migration(pte_t pte) 30404a705fefSNaoya Horiguchi { 30414a705fefSNaoya Horiguchi swp_entry_t swp; 30424a705fefSNaoya Horiguchi 30434a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 30444a705fefSNaoya Horiguchi return 0; 30454a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 30464a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_migration_entry(swp)) 30474a705fefSNaoya Horiguchi return 1; 30484a705fefSNaoya Horiguchi else 30494a705fefSNaoya Horiguchi return 0; 30504a705fefSNaoya Horiguchi } 30514a705fefSNaoya Horiguchi 30524a705fefSNaoya Horiguchi static int is_hugetlb_entry_hwpoisoned(pte_t pte) 30534a705fefSNaoya Horiguchi { 30544a705fefSNaoya Horiguchi swp_entry_t swp; 30554a705fefSNaoya Horiguchi 30564a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 30574a705fefSNaoya Horiguchi return 0; 30584a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 30594a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_hwpoison_entry(swp)) 30604a705fefSNaoya Horiguchi return 1; 30614a705fefSNaoya Horiguchi else 30624a705fefSNaoya Horiguchi return 0; 30634a705fefSNaoya Horiguchi } 30641e8f889bSDavid Gibson 306563551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 306663551ae0SDavid Gibson struct vm_area_struct *vma) 306763551ae0SDavid Gibson { 306863551ae0SDavid Gibson pte_t *src_pte, *dst_pte, entry; 306963551ae0SDavid Gibson struct page *ptepage; 30701c59827dSHugh Dickins unsigned long addr; 30711e8f889bSDavid Gibson int cow; 3072a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3073a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 3074e8569dd2SAndreas Sandberg unsigned long mmun_start; /* For mmu_notifiers */ 3075e8569dd2SAndreas Sandberg unsigned long mmun_end; /* For mmu_notifiers */ 3076e8569dd2SAndreas Sandberg int ret = 0; 30771e8f889bSDavid Gibson 30781e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 307963551ae0SDavid Gibson 3080e8569dd2SAndreas Sandberg mmun_start = vma->vm_start; 3081e8569dd2SAndreas Sandberg mmun_end = vma->vm_end; 3082e8569dd2SAndreas Sandberg if (cow) 3083e8569dd2SAndreas Sandberg mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end); 3084e8569dd2SAndreas Sandberg 3085a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 3086cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 3087c74df32cSHugh Dickins src_pte = huge_pte_offset(src, addr); 3088c74df32cSHugh Dickins if (!src_pte) 3089c74df32cSHugh Dickins continue; 3090a5516438SAndi Kleen dst_pte = huge_pte_alloc(dst, addr, sz); 3091e8569dd2SAndreas Sandberg if (!dst_pte) { 3092e8569dd2SAndreas Sandberg ret = -ENOMEM; 3093e8569dd2SAndreas Sandberg break; 3094e8569dd2SAndreas Sandberg } 3095c5c99429SLarry Woodman 3096c5c99429SLarry Woodman /* If the pagetables are shared don't copy or take references */ 3097c5c99429SLarry Woodman if (dst_pte == src_pte) 3098c5c99429SLarry Woodman continue; 3099c5c99429SLarry Woodman 3100cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 3101cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 3102cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 31034a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 31044a705fefSNaoya Horiguchi if (huge_pte_none(entry)) { /* skip none entry */ 31054a705fefSNaoya Horiguchi ; 31064a705fefSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry) || 31074a705fefSNaoya Horiguchi is_hugetlb_entry_hwpoisoned(entry))) { 31084a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 31094a705fefSNaoya Horiguchi 31104a705fefSNaoya Horiguchi if (is_write_migration_entry(swp_entry) && cow) { 31114a705fefSNaoya Horiguchi /* 31124a705fefSNaoya Horiguchi * COW mappings require pages in both 31134a705fefSNaoya Horiguchi * parent and child to be set to read. 31144a705fefSNaoya Horiguchi */ 31154a705fefSNaoya Horiguchi make_migration_entry_read(&swp_entry); 31164a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 31174a705fefSNaoya Horiguchi set_huge_pte_at(src, addr, src_pte, entry); 31184a705fefSNaoya Horiguchi } 31194a705fefSNaoya Horiguchi set_huge_pte_at(dst, addr, dst_pte, entry); 31204a705fefSNaoya Horiguchi } else { 312134ee645eSJoerg Roedel if (cow) { 31227f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 312334ee645eSJoerg Roedel mmu_notifier_invalidate_range(src, mmun_start, 312434ee645eSJoerg Roedel mmun_end); 312534ee645eSJoerg Roedel } 31260253d634SNaoya Horiguchi entry = huge_ptep_get(src_pte); 312763551ae0SDavid Gibson ptepage = pte_page(entry); 312863551ae0SDavid Gibson get_page(ptepage); 31290fe6e20bSNaoya Horiguchi page_dup_rmap(ptepage); 313063551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 31315d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), dst); 31321c59827dSHugh Dickins } 3133cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 3134cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 313563551ae0SDavid Gibson } 313663551ae0SDavid Gibson 3137e8569dd2SAndreas Sandberg if (cow) 3138e8569dd2SAndreas Sandberg mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end); 3139e8569dd2SAndreas Sandberg 3140e8569dd2SAndreas Sandberg return ret; 314163551ae0SDavid Gibson } 314263551ae0SDavid Gibson 314324669e58SAneesh Kumar K.V void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 314424669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 314524669e58SAneesh Kumar K.V struct page *ref_page) 314663551ae0SDavid Gibson { 314724669e58SAneesh Kumar K.V int force_flush = 0; 314863551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 314963551ae0SDavid Gibson unsigned long address; 3150c7546f8fSDavid Gibson pte_t *ptep; 315163551ae0SDavid Gibson pte_t pte; 3152cb900f41SKirill A. Shutemov spinlock_t *ptl; 315363551ae0SDavid Gibson struct page *page; 3154a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3155a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 31562ec74c3eSSagi Grimberg const unsigned long mmun_start = start; /* For mmu_notifiers */ 31572ec74c3eSSagi Grimberg const unsigned long mmun_end = end; /* For mmu_notifiers */ 3158a5516438SAndi Kleen 315963551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 3160a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 3161a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 316263551ae0SDavid Gibson 316324669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 31642ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 3165569f48b8SHillf Danton address = start; 316624669e58SAneesh Kumar K.V again: 3167569f48b8SHillf Danton for (; address < end; address += sz) { 3168c7546f8fSDavid Gibson ptep = huge_pte_offset(mm, address); 3169c7546f8fSDavid Gibson if (!ptep) 3170c7546f8fSDavid Gibson continue; 3171c7546f8fSDavid Gibson 3172cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 317339dde65cSChen, Kenneth W if (huge_pmd_unshare(mm, &address, ptep)) 3174cb900f41SKirill A. Shutemov goto unlock; 317539dde65cSChen, Kenneth W 31766629326bSHillf Danton pte = huge_ptep_get(ptep); 31776629326bSHillf Danton if (huge_pte_none(pte)) 3178cb900f41SKirill A. Shutemov goto unlock; 31796629326bSHillf Danton 31806629326bSHillf Danton /* 31819fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 31829fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 31836629326bSHillf Danton */ 31849fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 3185106c992aSGerald Schaefer huge_pte_clear(mm, address, ptep); 3186cb900f41SKirill A. Shutemov goto unlock; 31878c4894c6SNaoya Horiguchi } 31886629326bSHillf Danton 31896629326bSHillf Danton page = pte_page(pte); 319004f2cbe3SMel Gorman /* 319104f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 319204f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 319304f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 319404f2cbe3SMel Gorman */ 319504f2cbe3SMel Gorman if (ref_page) { 319604f2cbe3SMel Gorman if (page != ref_page) 3197cb900f41SKirill A. Shutemov goto unlock; 319804f2cbe3SMel Gorman 319904f2cbe3SMel Gorman /* 320004f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 320104f2cbe3SMel Gorman * future faults in this VMA will fail rather than 320204f2cbe3SMel Gorman * looking like data was lost 320304f2cbe3SMel Gorman */ 320404f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 320504f2cbe3SMel Gorman } 320604f2cbe3SMel Gorman 3207c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 320824669e58SAneesh Kumar K.V tlb_remove_tlb_entry(tlb, ptep, address); 3209106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 32106649a386SKen Chen set_page_dirty(page); 32119e81130bSHillf Danton 32125d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 321324669e58SAneesh Kumar K.V page_remove_rmap(page); 321424669e58SAneesh Kumar K.V force_flush = !__tlb_remove_page(tlb, page); 3215cb900f41SKirill A. Shutemov if (force_flush) { 3216569f48b8SHillf Danton address += sz; 3217cb900f41SKirill A. Shutemov spin_unlock(ptl); 32189e81130bSHillf Danton break; 321963551ae0SDavid Gibson } 3220cb900f41SKirill A. Shutemov /* Bail out after unmapping reference page if supplied */ 3221cb900f41SKirill A. Shutemov if (ref_page) { 3222cb900f41SKirill A. Shutemov spin_unlock(ptl); 3223cb900f41SKirill A. Shutemov break; 3224cb900f41SKirill A. Shutemov } 3225cb900f41SKirill A. Shutemov unlock: 3226cb900f41SKirill A. Shutemov spin_unlock(ptl); 3227cb900f41SKirill A. Shutemov } 322824669e58SAneesh Kumar K.V /* 322924669e58SAneesh Kumar K.V * mmu_gather ran out of room to batch pages, we break out of 323024669e58SAneesh Kumar K.V * the PTE lock to avoid doing the potential expensive TLB invalidate 323124669e58SAneesh Kumar K.V * and page-free while holding it. 323224669e58SAneesh Kumar K.V */ 323324669e58SAneesh Kumar K.V if (force_flush) { 323424669e58SAneesh Kumar K.V force_flush = 0; 323524669e58SAneesh Kumar K.V tlb_flush_mmu(tlb); 323624669e58SAneesh Kumar K.V if (address < end && !ref_page) 323724669e58SAneesh Kumar K.V goto again; 3238fe1668aeSChen, Kenneth W } 32392ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 324024669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 32411da177e4SLinus Torvalds } 324263551ae0SDavid Gibson 3243d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 3244d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 3245d833352aSMel Gorman unsigned long end, struct page *ref_page) 3246d833352aSMel Gorman { 3247d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 3248d833352aSMel Gorman 3249d833352aSMel Gorman /* 3250d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 3251d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 3252d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 3253d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 3254d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 3255c8c06efaSDavidlohr Bueso * is to clear it before releasing the i_mmap_rwsem. This works 3256d833352aSMel Gorman * because in the context this is called, the VMA is about to be 3257c8c06efaSDavidlohr Bueso * destroyed and the i_mmap_rwsem is held. 3258d833352aSMel Gorman */ 3259d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 3260d833352aSMel Gorman } 3261d833352aSMel Gorman 3262502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 326304f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 3264502717f4SChen, Kenneth W { 326524669e58SAneesh Kumar K.V struct mm_struct *mm; 326624669e58SAneesh Kumar K.V struct mmu_gather tlb; 326724669e58SAneesh Kumar K.V 326824669e58SAneesh Kumar K.V mm = vma->vm_mm; 326924669e58SAneesh Kumar K.V 32702b047252SLinus Torvalds tlb_gather_mmu(&tlb, mm, start, end); 327124669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 327224669e58SAneesh Kumar K.V tlb_finish_mmu(&tlb, start, end); 3273502717f4SChen, Kenneth W } 3274502717f4SChen, Kenneth W 327504f2cbe3SMel Gorman /* 327604f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 327704f2cbe3SMel Gorman * mappping it owns the reserve page for. The intention is to unmap the page 327804f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 327904f2cbe3SMel Gorman * same region. 328004f2cbe3SMel Gorman */ 32812f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 32822a4b3dedSHarvey Harrison struct page *page, unsigned long address) 328304f2cbe3SMel Gorman { 32847526674dSAdam Litke struct hstate *h = hstate_vma(vma); 328504f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 328604f2cbe3SMel Gorman struct address_space *mapping; 328704f2cbe3SMel Gorman pgoff_t pgoff; 328804f2cbe3SMel Gorman 328904f2cbe3SMel Gorman /* 329004f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 329104f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 329204f2cbe3SMel Gorman */ 32937526674dSAdam Litke address = address & huge_page_mask(h); 329436e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 329536e4f20aSMichal Hocko vma->vm_pgoff; 3296496ad9aaSAl Viro mapping = file_inode(vma->vm_file)->i_mapping; 329704f2cbe3SMel Gorman 32984eb2b1dcSMel Gorman /* 32994eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 33004eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 33014eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 33024eb2b1dcSMel Gorman */ 330383cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 33046b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 330504f2cbe3SMel Gorman /* Do not unmap the current VMA */ 330604f2cbe3SMel Gorman if (iter_vma == vma) 330704f2cbe3SMel Gorman continue; 330804f2cbe3SMel Gorman 330904f2cbe3SMel Gorman /* 33102f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 33112f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 33122f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 33132f84a899SMel Gorman */ 33142f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 33152f84a899SMel Gorman continue; 33162f84a899SMel Gorman 33172f84a899SMel Gorman /* 331804f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 331904f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 332004f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 332104f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 332204f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 332304f2cbe3SMel Gorman */ 332404f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 332524669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 332624669e58SAneesh Kumar K.V address + huge_page_size(h), page); 332704f2cbe3SMel Gorman } 332883cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 332904f2cbe3SMel Gorman } 333004f2cbe3SMel Gorman 33310fe6e20bSNaoya Horiguchi /* 33320fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 3333ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 3334ef009b25SMichal Hocko * cannot race with other handlers or page migration. 3335ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 33360fe6e20bSNaoya Horiguchi */ 33371e8f889bSDavid Gibson static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 333804f2cbe3SMel Gorman unsigned long address, pte_t *ptep, pte_t pte, 3339cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 33401e8f889bSDavid Gibson { 3341a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 33421e8f889bSDavid Gibson struct page *old_page, *new_page; 3343ad4404a2SDavidlohr Bueso int ret = 0, outside_reserve = 0; 33442ec74c3eSSagi Grimberg unsigned long mmun_start; /* For mmu_notifiers */ 33452ec74c3eSSagi Grimberg unsigned long mmun_end; /* For mmu_notifiers */ 33461e8f889bSDavid Gibson 33471e8f889bSDavid Gibson old_page = pte_page(pte); 33481e8f889bSDavid Gibson 334904f2cbe3SMel Gorman retry_avoidcopy: 33501e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 33511e8f889bSDavid Gibson * and just make the page writable */ 335237a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 33530fe6e20bSNaoya Horiguchi page_move_anon_rmap(old_page, vma, address); 33541e8f889bSDavid Gibson set_huge_ptep_writable(vma, address, ptep); 335583c54070SNick Piggin return 0; 33561e8f889bSDavid Gibson } 33571e8f889bSDavid Gibson 335804f2cbe3SMel Gorman /* 335904f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 336004f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 336104f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 336204f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 336304f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 336404f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 336504f2cbe3SMel Gorman * of the full address range. 336604f2cbe3SMel Gorman */ 33675944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 336804f2cbe3SMel Gorman old_page != pagecache_page) 336904f2cbe3SMel Gorman outside_reserve = 1; 337004f2cbe3SMel Gorman 33711e8f889bSDavid Gibson page_cache_get(old_page); 3372b76c8cfbSLarry Woodman 3373ad4404a2SDavidlohr Bueso /* 3374ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 3375ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 3376ad4404a2SDavidlohr Bueso */ 3377cb900f41SKirill A. Shutemov spin_unlock(ptl); 337804f2cbe3SMel Gorman new_page = alloc_huge_page(vma, address, outside_reserve); 33791e8f889bSDavid Gibson 33802fc39cecSAdam Litke if (IS_ERR(new_page)) { 338104f2cbe3SMel Gorman /* 338204f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 338304f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 338404f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 338504f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 338604f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 338704f2cbe3SMel Gorman */ 338804f2cbe3SMel Gorman if (outside_reserve) { 3389ad4404a2SDavidlohr Bueso page_cache_release(old_page); 339004f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 33912f4612afSDavidlohr Bueso unmap_ref_private(mm, vma, old_page, address); 339204f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 3393cb900f41SKirill A. Shutemov spin_lock(ptl); 3394a734bcc8SHillf Danton ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 3395a9af0c5dSNaoya Horiguchi if (likely(ptep && 3396a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 339704f2cbe3SMel Gorman goto retry_avoidcopy; 3398a734bcc8SHillf Danton /* 3399cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 3400cb900f41SKirill A. Shutemov * lock, and our job is done. 3401a734bcc8SHillf Danton */ 3402a734bcc8SHillf Danton return 0; 340304f2cbe3SMel Gorman } 340404f2cbe3SMel Gorman 3405ad4404a2SDavidlohr Bueso ret = (PTR_ERR(new_page) == -ENOMEM) ? 3406ad4404a2SDavidlohr Bueso VM_FAULT_OOM : VM_FAULT_SIGBUS; 3407ad4404a2SDavidlohr Bueso goto out_release_old; 34081e8f889bSDavid Gibson } 34091e8f889bSDavid Gibson 34100fe6e20bSNaoya Horiguchi /* 34110fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 34120fe6e20bSNaoya Horiguchi * anon_vma prepared. 34130fe6e20bSNaoya Horiguchi */ 341444e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 3415ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 3416ad4404a2SDavidlohr Bueso goto out_release_all; 341744e2aa93SDean Nelson } 34180fe6e20bSNaoya Horiguchi 341947ad8475SAndrea Arcangeli copy_user_huge_page(new_page, old_page, address, vma, 342047ad8475SAndrea Arcangeli pages_per_huge_page(h)); 34210ed361deSNick Piggin __SetPageUptodate(new_page); 3422bcc54222SNaoya Horiguchi set_page_huge_active(new_page); 34231e8f889bSDavid Gibson 34242ec74c3eSSagi Grimberg mmun_start = address & huge_page_mask(h); 34252ec74c3eSSagi Grimberg mmun_end = mmun_start + huge_page_size(h); 34262ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 3427ad4404a2SDavidlohr Bueso 3428b76c8cfbSLarry Woodman /* 3429cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 3430b76c8cfbSLarry Woodman * before the page tables are altered 3431b76c8cfbSLarry Woodman */ 3432cb900f41SKirill A. Shutemov spin_lock(ptl); 3433a5516438SAndi Kleen ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 3434a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 343507443a85SJoonsoo Kim ClearPagePrivate(new_page); 343607443a85SJoonsoo Kim 34371e8f889bSDavid Gibson /* Break COW */ 34388fe627ecSGerald Schaefer huge_ptep_clear_flush(vma, address, ptep); 343934ee645eSJoerg Roedel mmu_notifier_invalidate_range(mm, mmun_start, mmun_end); 34401e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, 34411e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 34420fe6e20bSNaoya Horiguchi page_remove_rmap(old_page); 3443cd67f0d2SNaoya Horiguchi hugepage_add_new_anon_rmap(new_page, vma, address); 34441e8f889bSDavid Gibson /* Make the old page be freed below */ 34451e8f889bSDavid Gibson new_page = old_page; 34461e8f889bSDavid Gibson } 3447cb900f41SKirill A. Shutemov spin_unlock(ptl); 34482ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 3449ad4404a2SDavidlohr Bueso out_release_all: 34501e8f889bSDavid Gibson page_cache_release(new_page); 3451ad4404a2SDavidlohr Bueso out_release_old: 34521e8f889bSDavid Gibson page_cache_release(old_page); 34538312034fSJoonsoo Kim 3454ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 3455ad4404a2SDavidlohr Bueso return ret; 34561e8f889bSDavid Gibson } 34571e8f889bSDavid Gibson 345804f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 3459a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 3460a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 346104f2cbe3SMel Gorman { 346204f2cbe3SMel Gorman struct address_space *mapping; 3463e7c4b0bfSAndy Whitcroft pgoff_t idx; 346404f2cbe3SMel Gorman 346504f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 3466a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 346704f2cbe3SMel Gorman 346804f2cbe3SMel Gorman return find_lock_page(mapping, idx); 346904f2cbe3SMel Gorman } 347004f2cbe3SMel Gorman 34713ae77f43SHugh Dickins /* 34723ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 34733ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 34743ae77f43SHugh Dickins */ 34753ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 34762a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 34772a15efc9SHugh Dickins { 34782a15efc9SHugh Dickins struct address_space *mapping; 34792a15efc9SHugh Dickins pgoff_t idx; 34802a15efc9SHugh Dickins struct page *page; 34812a15efc9SHugh Dickins 34822a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 34832a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 34842a15efc9SHugh Dickins 34852a15efc9SHugh Dickins page = find_get_page(mapping, idx); 34862a15efc9SHugh Dickins if (page) 34872a15efc9SHugh Dickins put_page(page); 34882a15efc9SHugh Dickins return page != NULL; 34892a15efc9SHugh Dickins } 34902a15efc9SHugh Dickins 3491ab76ad54SMike Kravetz int huge_add_to_page_cache(struct page *page, struct address_space *mapping, 3492ab76ad54SMike Kravetz pgoff_t idx) 3493ab76ad54SMike Kravetz { 3494ab76ad54SMike Kravetz struct inode *inode = mapping->host; 3495ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 3496ab76ad54SMike Kravetz int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 3497ab76ad54SMike Kravetz 3498ab76ad54SMike Kravetz if (err) 3499ab76ad54SMike Kravetz return err; 3500ab76ad54SMike Kravetz ClearPagePrivate(page); 3501ab76ad54SMike Kravetz 3502ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 3503ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 3504ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 3505ab76ad54SMike Kravetz return 0; 3506ab76ad54SMike Kravetz } 3507ab76ad54SMike Kravetz 3508a1ed3ddaSRobert P. J. Day static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, 35098382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 3510788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 3511ac9b9c66SHugh Dickins { 3512a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3513ac9b9c66SHugh Dickins int ret = VM_FAULT_SIGBUS; 3514409eb8c2SHillf Danton int anon_rmap = 0; 35154c887265SAdam Litke unsigned long size; 35164c887265SAdam Litke struct page *page; 35171e8f889bSDavid Gibson pte_t new_pte; 3518cb900f41SKirill A. Shutemov spinlock_t *ptl; 35194c887265SAdam Litke 352004f2cbe3SMel Gorman /* 352104f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 352204f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 352325985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 352404f2cbe3SMel Gorman */ 352504f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 3526ffb22af5SAndrew Morton pr_warning("PID %d killed due to inadequate hugepage pool\n", 352704f2cbe3SMel Gorman current->pid); 352804f2cbe3SMel Gorman return ret; 352904f2cbe3SMel Gorman } 353004f2cbe3SMel Gorman 35314c887265SAdam Litke /* 35324c887265SAdam Litke * Use page lock to guard against racing truncation 35334c887265SAdam Litke * before we get page_table_lock. 35344c887265SAdam Litke */ 35356bda666aSChristoph Lameter retry: 35366bda666aSChristoph Lameter page = find_lock_page(mapping, idx); 35376bda666aSChristoph Lameter if (!page) { 3538a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 3539ebed4bfcSHugh Dickins if (idx >= size) 3540ebed4bfcSHugh Dickins goto out; 354104f2cbe3SMel Gorman page = alloc_huge_page(vma, address, 0); 35422fc39cecSAdam Litke if (IS_ERR(page)) { 354376dcee75SAneesh Kumar K.V ret = PTR_ERR(page); 354476dcee75SAneesh Kumar K.V if (ret == -ENOMEM) 354576dcee75SAneesh Kumar K.V ret = VM_FAULT_OOM; 354676dcee75SAneesh Kumar K.V else 354776dcee75SAneesh Kumar K.V ret = VM_FAULT_SIGBUS; 35486bda666aSChristoph Lameter goto out; 35496bda666aSChristoph Lameter } 355047ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 35510ed361deSNick Piggin __SetPageUptodate(page); 3552bcc54222SNaoya Horiguchi set_page_huge_active(page); 3553ac9b9c66SHugh Dickins 3554f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 3555ab76ad54SMike Kravetz int err = huge_add_to_page_cache(page, mapping, idx); 35566bda666aSChristoph Lameter if (err) { 35576bda666aSChristoph Lameter put_page(page); 35586bda666aSChristoph Lameter if (err == -EEXIST) 35596bda666aSChristoph Lameter goto retry; 35606bda666aSChristoph Lameter goto out; 35616bda666aSChristoph Lameter } 356223be7468SMel Gorman } else { 35636bda666aSChristoph Lameter lock_page(page); 35640fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 35650fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 35660fe6e20bSNaoya Horiguchi goto backout_unlocked; 356723be7468SMel Gorman } 3568409eb8c2SHillf Danton anon_rmap = 1; 35690fe6e20bSNaoya Horiguchi } 35700fe6e20bSNaoya Horiguchi } else { 357157303d80SAndy Whitcroft /* 3572998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 3573998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 3574998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 3575fd6a03edSNaoya Horiguchi */ 3576fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 3577aa50d3a7SAndi Kleen ret = VM_FAULT_HWPOISON | 3578972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 3579fd6a03edSNaoya Horiguchi goto backout_unlocked; 35806bda666aSChristoph Lameter } 3581998b4382SNaoya Horiguchi } 35821e8f889bSDavid Gibson 358357303d80SAndy Whitcroft /* 358457303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 358557303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 358657303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 358757303d80SAndy Whitcroft * the spinlock. 358857303d80SAndy Whitcroft */ 35895e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 35902b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 35912b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 35922b26736cSAndy Whitcroft goto backout_unlocked; 35932b26736cSAndy Whitcroft } 35945e911373SMike Kravetz /* Just decrements count, does not deallocate */ 3595feba16e2SMike Kravetz vma_end_reservation(h, vma, address); 35965e911373SMike Kravetz } 359757303d80SAndy Whitcroft 3598cb900f41SKirill A. Shutemov ptl = huge_pte_lockptr(h, mm, ptep); 3599cb900f41SKirill A. Shutemov spin_lock(ptl); 3600a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 36014c887265SAdam Litke if (idx >= size) 36024c887265SAdam Litke goto backout; 36034c887265SAdam Litke 360483c54070SNick Piggin ret = 0; 36057f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 36064c887265SAdam Litke goto backout; 36074c887265SAdam Litke 360807443a85SJoonsoo Kim if (anon_rmap) { 360907443a85SJoonsoo Kim ClearPagePrivate(page); 3610409eb8c2SHillf Danton hugepage_add_new_anon_rmap(page, vma, address); 3611ac714904SChoi Gi-yong } else 3612409eb8c2SHillf Danton page_dup_rmap(page); 36131e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 36141e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 36151e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, new_pte); 36161e8f889bSDavid Gibson 36175d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 3618788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 36191e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 3620cb900f41SKirill A. Shutemov ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl); 36211e8f889bSDavid Gibson } 36221e8f889bSDavid Gibson 3623cb900f41SKirill A. Shutemov spin_unlock(ptl); 36244c887265SAdam Litke unlock_page(page); 36254c887265SAdam Litke out: 3626ac9b9c66SHugh Dickins return ret; 36274c887265SAdam Litke 36284c887265SAdam Litke backout: 3629cb900f41SKirill A. Shutemov spin_unlock(ptl); 36302b26736cSAndy Whitcroft backout_unlocked: 36314c887265SAdam Litke unlock_page(page); 36324c887265SAdam Litke put_page(page); 36334c887265SAdam Litke goto out; 3634ac9b9c66SHugh Dickins } 3635ac9b9c66SHugh Dickins 36368382d914SDavidlohr Bueso #ifdef CONFIG_SMP 3637c672c7f2SMike Kravetz u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, 36388382d914SDavidlohr Bueso struct vm_area_struct *vma, 36398382d914SDavidlohr Bueso struct address_space *mapping, 36408382d914SDavidlohr Bueso pgoff_t idx, unsigned long address) 36418382d914SDavidlohr Bueso { 36428382d914SDavidlohr Bueso unsigned long key[2]; 36438382d914SDavidlohr Bueso u32 hash; 36448382d914SDavidlohr Bueso 36458382d914SDavidlohr Bueso if (vma->vm_flags & VM_SHARED) { 36468382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 36478382d914SDavidlohr Bueso key[1] = idx; 36488382d914SDavidlohr Bueso } else { 36498382d914SDavidlohr Bueso key[0] = (unsigned long) mm; 36508382d914SDavidlohr Bueso key[1] = address >> huge_page_shift(h); 36518382d914SDavidlohr Bueso } 36528382d914SDavidlohr Bueso 36538382d914SDavidlohr Bueso hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0); 36548382d914SDavidlohr Bueso 36558382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 36568382d914SDavidlohr Bueso } 36578382d914SDavidlohr Bueso #else 36588382d914SDavidlohr Bueso /* 36598382d914SDavidlohr Bueso * For uniprocesor systems we always use a single mutex, so just 36608382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 36618382d914SDavidlohr Bueso */ 3662c672c7f2SMike Kravetz u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, 36638382d914SDavidlohr Bueso struct vm_area_struct *vma, 36648382d914SDavidlohr Bueso struct address_space *mapping, 36658382d914SDavidlohr Bueso pgoff_t idx, unsigned long address) 36668382d914SDavidlohr Bueso { 36678382d914SDavidlohr Bueso return 0; 36688382d914SDavidlohr Bueso } 36698382d914SDavidlohr Bueso #endif 36708382d914SDavidlohr Bueso 367186e5216fSAdam Litke int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 3672788c7df4SHugh Dickins unsigned long address, unsigned int flags) 367386e5216fSAdam Litke { 36748382d914SDavidlohr Bueso pte_t *ptep, entry; 3675cb900f41SKirill A. Shutemov spinlock_t *ptl; 36761e8f889bSDavid Gibson int ret; 36778382d914SDavidlohr Bueso u32 hash; 36788382d914SDavidlohr Bueso pgoff_t idx; 36790fe6e20bSNaoya Horiguchi struct page *page = NULL; 368057303d80SAndy Whitcroft struct page *pagecache_page = NULL; 3681a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 36828382d914SDavidlohr Bueso struct address_space *mapping; 36830f792cf9SNaoya Horiguchi int need_wait_lock = 0; 368486e5216fSAdam Litke 36851e16a539SKAMEZAWA Hiroyuki address &= huge_page_mask(h); 36861e16a539SKAMEZAWA Hiroyuki 3687fd6a03edSNaoya Horiguchi ptep = huge_pte_offset(mm, address); 3688fd6a03edSNaoya Horiguchi if (ptep) { 3689fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 3690290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 3691cb900f41SKirill A. Shutemov migration_entry_wait_huge(vma, mm, ptep); 3692290408d4SNaoya Horiguchi return 0; 3693290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 3694aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 3695972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 3696fd6a03edSNaoya Horiguchi } 3697fd6a03edSNaoya Horiguchi 3698a5516438SAndi Kleen ptep = huge_pte_alloc(mm, address, huge_page_size(h)); 369986e5216fSAdam Litke if (!ptep) 370086e5216fSAdam Litke return VM_FAULT_OOM; 370186e5216fSAdam Litke 37028382d914SDavidlohr Bueso mapping = vma->vm_file->f_mapping; 37038382d914SDavidlohr Bueso idx = vma_hugecache_offset(h, vma, address); 37048382d914SDavidlohr Bueso 37053935baa9SDavid Gibson /* 37063935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 37073935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 37083935baa9SDavid Gibson * the same page in the page cache. 37093935baa9SDavid Gibson */ 3710c672c7f2SMike Kravetz hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address); 3711c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 37128382d914SDavidlohr Bueso 37137f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 37147f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 37158382d914SDavidlohr Bueso ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 3716b4d1d99fSDavid Gibson goto out_mutex; 37173935baa9SDavid Gibson } 371886e5216fSAdam Litke 371983c54070SNick Piggin ret = 0; 37201e8f889bSDavid Gibson 372157303d80SAndy Whitcroft /* 37220f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 37230f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 37240f792cf9SNaoya Horiguchi * a active hugepage in pagecache. This goto expects the 2nd page fault, 37250f792cf9SNaoya Horiguchi * and is_hugetlb_entry_(migration|hwpoisoned) check will properly 37260f792cf9SNaoya Horiguchi * handle it. 37270f792cf9SNaoya Horiguchi */ 37280f792cf9SNaoya Horiguchi if (!pte_present(entry)) 37290f792cf9SNaoya Horiguchi goto out_mutex; 37300f792cf9SNaoya Horiguchi 37310f792cf9SNaoya Horiguchi /* 373257303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 373357303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 373457303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 373557303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 373657303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 373757303d80SAndy Whitcroft * consumed. 373857303d80SAndy Whitcroft */ 3739106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 37402b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 37412b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 3742b4d1d99fSDavid Gibson goto out_mutex; 37432b26736cSAndy Whitcroft } 37445e911373SMike Kravetz /* Just decrements count, does not deallocate */ 3745feba16e2SMike Kravetz vma_end_reservation(h, vma, address); 374657303d80SAndy Whitcroft 3747f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 374857303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 374957303d80SAndy Whitcroft vma, address); 375057303d80SAndy Whitcroft } 375157303d80SAndy Whitcroft 37520f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 37530fe6e20bSNaoya Horiguchi 37541e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 3755b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 3756cb900f41SKirill A. Shutemov goto out_ptl; 3757b4d1d99fSDavid Gibson 37580f792cf9SNaoya Horiguchi /* 37590f792cf9SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 37600f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 37610f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 37620f792cf9SNaoya Horiguchi */ 37630f792cf9SNaoya Horiguchi page = pte_page(entry); 37640f792cf9SNaoya Horiguchi if (page != pagecache_page) 37650f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 37660f792cf9SNaoya Horiguchi need_wait_lock = 1; 37670f792cf9SNaoya Horiguchi goto out_ptl; 37680f792cf9SNaoya Horiguchi } 37690f792cf9SNaoya Horiguchi 37700f792cf9SNaoya Horiguchi get_page(page); 3771b4d1d99fSDavid Gibson 3772788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 3773106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 377457303d80SAndy Whitcroft ret = hugetlb_cow(mm, vma, address, ptep, entry, 3775cb900f41SKirill A. Shutemov pagecache_page, ptl); 37760f792cf9SNaoya Horiguchi goto out_put_page; 3777b4d1d99fSDavid Gibson } 3778106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 3779b4d1d99fSDavid Gibson } 3780b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 3781788c7df4SHugh Dickins if (huge_ptep_set_access_flags(vma, address, ptep, entry, 3782788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 37834b3073e1SRussell King update_mmu_cache(vma, address, ptep); 37840f792cf9SNaoya Horiguchi out_put_page: 37850f792cf9SNaoya Horiguchi if (page != pagecache_page) 37860f792cf9SNaoya Horiguchi unlock_page(page); 37870f792cf9SNaoya Horiguchi put_page(page); 3788cb900f41SKirill A. Shutemov out_ptl: 3789cb900f41SKirill A. Shutemov spin_unlock(ptl); 379057303d80SAndy Whitcroft 379157303d80SAndy Whitcroft if (pagecache_page) { 379257303d80SAndy Whitcroft unlock_page(pagecache_page); 379357303d80SAndy Whitcroft put_page(pagecache_page); 379457303d80SAndy Whitcroft } 3795b4d1d99fSDavid Gibson out_mutex: 3796c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 37970f792cf9SNaoya Horiguchi /* 37980f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 37990f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 38000f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 38010f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 38020f792cf9SNaoya Horiguchi * here without taking refcount. 38030f792cf9SNaoya Horiguchi */ 38040f792cf9SNaoya Horiguchi if (need_wait_lock) 38050f792cf9SNaoya Horiguchi wait_on_page_locked(page); 38061e8f889bSDavid Gibson return ret; 380786e5216fSAdam Litke } 380886e5216fSAdam Litke 380928a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 381063551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 381128a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 381228a35716SMichel Lespinasse long i, unsigned int flags) 381363551ae0SDavid Gibson { 3814d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 3815d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 381628a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 3817a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 381863551ae0SDavid Gibson 381963551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 382063551ae0SDavid Gibson pte_t *pte; 3821cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 38222a15efc9SHugh Dickins int absent; 382363551ae0SDavid Gibson struct page *page; 382463551ae0SDavid Gibson 38254c887265SAdam Litke /* 382602057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 382702057967SDavid Rientjes * potentially allocating memory. 382802057967SDavid Rientjes */ 382902057967SDavid Rientjes if (unlikely(fatal_signal_pending(current))) { 383002057967SDavid Rientjes remainder = 0; 383102057967SDavid Rientjes break; 383202057967SDavid Rientjes } 383302057967SDavid Rientjes 383402057967SDavid Rientjes /* 38354c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 38362a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 38374c887265SAdam Litke * first, for the page indexing below to work. 3838cb900f41SKirill A. Shutemov * 3839cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 38404c887265SAdam Litke */ 3841a5516438SAndi Kleen pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); 3842cb900f41SKirill A. Shutemov if (pte) 3843cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 38442a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 384563551ae0SDavid Gibson 38462a15efc9SHugh Dickins /* 38472a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 38483ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 38493ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 38503ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 38513ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 38522a15efc9SHugh Dickins */ 38533ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 38543ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 3855cb900f41SKirill A. Shutemov if (pte) 3856cb900f41SKirill A. Shutemov spin_unlock(ptl); 38572a15efc9SHugh Dickins remainder = 0; 38582a15efc9SHugh Dickins break; 38592a15efc9SHugh Dickins } 38602a15efc9SHugh Dickins 38619cc3a5bdSNaoya Horiguchi /* 38629cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 38639cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 38649cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 38659cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 38669cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 38679cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 38689cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 38699cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 38709cc3a5bdSNaoya Horiguchi */ 38719cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 3872106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 3873106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 38744c887265SAdam Litke int ret; 38754c887265SAdam Litke 3876cb900f41SKirill A. Shutemov if (pte) 3877cb900f41SKirill A. Shutemov spin_unlock(ptl); 38782a15efc9SHugh Dickins ret = hugetlb_fault(mm, vma, vaddr, 38792a15efc9SHugh Dickins (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); 3880a89182c7SAdam Litke if (!(ret & VM_FAULT_ERROR)) 38814c887265SAdam Litke continue; 38824c887265SAdam Litke 38831c59827dSHugh Dickins remainder = 0; 38841c59827dSHugh Dickins break; 38851c59827dSHugh Dickins } 388663551ae0SDavid Gibson 3887a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 38887f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 3889d5d4b0aaSChen, Kenneth W same_page: 3890d6692183SChen, Kenneth W if (pages) { 389169d177c2SAndy Whitcroft pages[i] = mem_map_offset(page, pfn_offset); 3892a0368d4eSAndrea Arcangeli get_page_foll(pages[i]); 3893d6692183SChen, Kenneth W } 389463551ae0SDavid Gibson 389563551ae0SDavid Gibson if (vmas) 389663551ae0SDavid Gibson vmas[i] = vma; 389763551ae0SDavid Gibson 389863551ae0SDavid Gibson vaddr += PAGE_SIZE; 3899d5d4b0aaSChen, Kenneth W ++pfn_offset; 390063551ae0SDavid Gibson --remainder; 390163551ae0SDavid Gibson ++i; 3902d5d4b0aaSChen, Kenneth W if (vaddr < vma->vm_end && remainder && 3903a5516438SAndi Kleen pfn_offset < pages_per_huge_page(h)) { 3904d5d4b0aaSChen, Kenneth W /* 3905d5d4b0aaSChen, Kenneth W * We use pfn_offset to avoid touching the pageframes 3906d5d4b0aaSChen, Kenneth W * of this compound page. 3907d5d4b0aaSChen, Kenneth W */ 3908d5d4b0aaSChen, Kenneth W goto same_page; 3909d5d4b0aaSChen, Kenneth W } 3910cb900f41SKirill A. Shutemov spin_unlock(ptl); 391163551ae0SDavid Gibson } 391228a35716SMichel Lespinasse *nr_pages = remainder; 391363551ae0SDavid Gibson *position = vaddr; 391463551ae0SDavid Gibson 39152a15efc9SHugh Dickins return i ? i : -EFAULT; 391663551ae0SDavid Gibson } 39178f860591SZhang, Yanmin 39187da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 39198f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 39208f860591SZhang, Yanmin { 39218f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 39228f860591SZhang, Yanmin unsigned long start = address; 39238f860591SZhang, Yanmin pte_t *ptep; 39248f860591SZhang, Yanmin pte_t pte; 3925a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 39267da4d641SPeter Zijlstra unsigned long pages = 0; 39278f860591SZhang, Yanmin 39288f860591SZhang, Yanmin BUG_ON(address >= end); 39298f860591SZhang, Yanmin flush_cache_range(vma, address, end); 39308f860591SZhang, Yanmin 3931a5338093SRik van Riel mmu_notifier_invalidate_range_start(mm, start, end); 393283cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 3933a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 3934cb900f41SKirill A. Shutemov spinlock_t *ptl; 39358f860591SZhang, Yanmin ptep = huge_pte_offset(mm, address); 39368f860591SZhang, Yanmin if (!ptep) 39378f860591SZhang, Yanmin continue; 3938cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 39397da4d641SPeter Zijlstra if (huge_pmd_unshare(mm, &address, ptep)) { 39407da4d641SPeter Zijlstra pages++; 3941cb900f41SKirill A. Shutemov spin_unlock(ptl); 394239dde65cSChen, Kenneth W continue; 39437da4d641SPeter Zijlstra } 3944a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 3945a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 3946a8bda28dSNaoya Horiguchi spin_unlock(ptl); 3947a8bda28dSNaoya Horiguchi continue; 3948a8bda28dSNaoya Horiguchi } 3949a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 3950a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 3951a8bda28dSNaoya Horiguchi 3952a8bda28dSNaoya Horiguchi if (is_write_migration_entry(entry)) { 3953a8bda28dSNaoya Horiguchi pte_t newpte; 3954a8bda28dSNaoya Horiguchi 3955a8bda28dSNaoya Horiguchi make_migration_entry_read(&entry); 3956a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 3957a8bda28dSNaoya Horiguchi set_huge_pte_at(mm, address, ptep, newpte); 3958a8bda28dSNaoya Horiguchi pages++; 3959a8bda28dSNaoya Horiguchi } 3960a8bda28dSNaoya Horiguchi spin_unlock(ptl); 3961a8bda28dSNaoya Horiguchi continue; 3962a8bda28dSNaoya Horiguchi } 3963a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 39648f860591SZhang, Yanmin pte = huge_ptep_get_and_clear(mm, address, ptep); 3965106c992aSGerald Schaefer pte = pte_mkhuge(huge_pte_modify(pte, newprot)); 3966be7517d6STony Lu pte = arch_make_huge_pte(pte, vma, NULL, 0); 39678f860591SZhang, Yanmin set_huge_pte_at(mm, address, ptep, pte); 39687da4d641SPeter Zijlstra pages++; 39698f860591SZhang, Yanmin } 3970cb900f41SKirill A. Shutemov spin_unlock(ptl); 39718f860591SZhang, Yanmin } 3972d833352aSMel Gorman /* 3973c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 3974d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 3975c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 3976d833352aSMel Gorman * and that page table be reused and filled with junk. 3977d833352aSMel Gorman */ 39788f860591SZhang, Yanmin flush_tlb_range(vma, start, end); 397934ee645eSJoerg Roedel mmu_notifier_invalidate_range(mm, start, end); 398083cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 3981a5338093SRik van Riel mmu_notifier_invalidate_range_end(mm, start, end); 39827da4d641SPeter Zijlstra 39837da4d641SPeter Zijlstra return pages << h->order; 39848f860591SZhang, Yanmin } 39858f860591SZhang, Yanmin 3986a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 3987a1e78772SMel Gorman long from, long to, 39885a6fe125SMel Gorman struct vm_area_struct *vma, 3989ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 3990e4e574b7SAdam Litke { 399117c9d12eSMel Gorman long ret, chg; 3992a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 399390481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 39949119a41eSJoonsoo Kim struct resv_map *resv_map; 39951c5ecae3SMike Kravetz long gbl_reserve; 3996e4e574b7SAdam Litke 3997a1e78772SMel Gorman /* 399817c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 399917c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 400090481622SDavid Gibson * without using reserves 400117c9d12eSMel Gorman */ 4002ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 400317c9d12eSMel Gorman return 0; 400417c9d12eSMel Gorman 400517c9d12eSMel Gorman /* 4006a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 4007a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 4008a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 4009a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 4010a1e78772SMel Gorman */ 40119119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 40124e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 40139119a41eSJoonsoo Kim 40141406ec9bSJoonsoo Kim chg = region_chg(resv_map, from, to); 40159119a41eSJoonsoo Kim 40169119a41eSJoonsoo Kim } else { 40179119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 40185a6fe125SMel Gorman if (!resv_map) 40195a6fe125SMel Gorman return -ENOMEM; 40205a6fe125SMel Gorman 402117c9d12eSMel Gorman chg = to - from; 402217c9d12eSMel Gorman 40235a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 40245a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 40255a6fe125SMel Gorman } 40265a6fe125SMel Gorman 4027c50ac050SDave Hansen if (chg < 0) { 4028c50ac050SDave Hansen ret = chg; 4029c50ac050SDave Hansen goto out_err; 4030c50ac050SDave Hansen } 403117c9d12eSMel Gorman 40321c5ecae3SMike Kravetz /* 40331c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 40341c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 40351c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 40361c5ecae3SMike Kravetz */ 40371c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 40381c5ecae3SMike Kravetz if (gbl_reserve < 0) { 4039c50ac050SDave Hansen ret = -ENOSPC; 4040c50ac050SDave Hansen goto out_err; 4041c50ac050SDave Hansen } 404217c9d12eSMel Gorman 404317c9d12eSMel Gorman /* 404417c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 404590481622SDavid Gibson * Hand the pages back to the subpool if there are not 404617c9d12eSMel Gorman */ 40471c5ecae3SMike Kravetz ret = hugetlb_acct_memory(h, gbl_reserve); 404817c9d12eSMel Gorman if (ret < 0) { 40491c5ecae3SMike Kravetz /* put back original number of pages, chg */ 40501c5ecae3SMike Kravetz (void)hugepage_subpool_put_pages(spool, chg); 4051c50ac050SDave Hansen goto out_err; 405217c9d12eSMel Gorman } 405317c9d12eSMel Gorman 405417c9d12eSMel Gorman /* 405517c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 405617c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 405717c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 405817c9d12eSMel Gorman * the reservation was and the page cache tells how much of 405917c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 406017c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 406117c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 406217c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 406317c9d12eSMel Gorman * else has to be done for private mappings here 406417c9d12eSMel Gorman */ 406533039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 406633039678SMike Kravetz long add = region_add(resv_map, from, to); 406733039678SMike Kravetz 406833039678SMike Kravetz if (unlikely(chg > add)) { 406933039678SMike Kravetz /* 407033039678SMike Kravetz * pages in this range were added to the reserve 407133039678SMike Kravetz * map between region_chg and region_add. This 407233039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 407333039678SMike Kravetz * the subpool and reserve counts modified above 407433039678SMike Kravetz * based on the difference. 407533039678SMike Kravetz */ 407633039678SMike Kravetz long rsv_adjust; 407733039678SMike Kravetz 407833039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 407933039678SMike Kravetz chg - add); 408033039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 408133039678SMike Kravetz } 408233039678SMike Kravetz } 4083a43a8c39SChen, Kenneth W return 0; 4084c50ac050SDave Hansen out_err: 40855e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 40865e911373SMike Kravetz region_abort(resv_map, from, to); 4087f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 4088f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 4089c50ac050SDave Hansen return ret; 4090a43a8c39SChen, Kenneth W } 4091a43a8c39SChen, Kenneth W 4092b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 4093b5cec28dSMike Kravetz long freed) 4094a43a8c39SChen, Kenneth W { 4095a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 40964e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 40979119a41eSJoonsoo Kim long chg = 0; 409890481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 40991c5ecae3SMike Kravetz long gbl_reserve; 410045c682a6SKen Chen 4101b5cec28dSMike Kravetz if (resv_map) { 4102b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 4103b5cec28dSMike Kravetz /* 4104b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 4105b5cec28dSMike Kravetz * must be split and another region descriptor can not be 4106b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 4107b5cec28dSMike Kravetz */ 4108b5cec28dSMike Kravetz if (chg < 0) 4109b5cec28dSMike Kravetz return chg; 4110b5cec28dSMike Kravetz } 4111b5cec28dSMike Kravetz 411245c682a6SKen Chen spin_lock(&inode->i_lock); 4113e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 411445c682a6SKen Chen spin_unlock(&inode->i_lock); 411545c682a6SKen Chen 41161c5ecae3SMike Kravetz /* 41171c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 41181c5ecae3SMike Kravetz * reservations to be released may be adjusted. 41191c5ecae3SMike Kravetz */ 41201c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 41211c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 4122b5cec28dSMike Kravetz 4123b5cec28dSMike Kravetz return 0; 4124a43a8c39SChen, Kenneth W } 412593f70f90SNaoya Horiguchi 41263212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 41273212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 41283212b535SSteve Capper struct vm_area_struct *vma, 41293212b535SSteve Capper unsigned long addr, pgoff_t idx) 41303212b535SSteve Capper { 41313212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 41323212b535SSteve Capper svma->vm_start; 41333212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 41343212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 41353212b535SSteve Capper 41363212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 4137de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 4138de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 41393212b535SSteve Capper 41403212b535SSteve Capper /* 41413212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 41423212b535SSteve Capper * page table page. 41433212b535SSteve Capper */ 41443212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 41453212b535SSteve Capper vm_flags != svm_flags || 41463212b535SSteve Capper sbase < svma->vm_start || svma->vm_end < s_end) 41473212b535SSteve Capper return 0; 41483212b535SSteve Capper 41493212b535SSteve Capper return saddr; 41503212b535SSteve Capper } 41513212b535SSteve Capper 415231aafb45SNicholas Krause static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) 41533212b535SSteve Capper { 41543212b535SSteve Capper unsigned long base = addr & PUD_MASK; 41553212b535SSteve Capper unsigned long end = base + PUD_SIZE; 41563212b535SSteve Capper 41573212b535SSteve Capper /* 41583212b535SSteve Capper * check on proper vm_flags and page table alignment 41593212b535SSteve Capper */ 41603212b535SSteve Capper if (vma->vm_flags & VM_MAYSHARE && 41613212b535SSteve Capper vma->vm_start <= base && end <= vma->vm_end) 416231aafb45SNicholas Krause return true; 416331aafb45SNicholas Krause return false; 41643212b535SSteve Capper } 41653212b535SSteve Capper 41663212b535SSteve Capper /* 41673212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 41683212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 41693212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 41703212b535SSteve Capper * code much cleaner. pmd allocation is essential for the shared case because 4171c8c06efaSDavidlohr Bueso * pud has to be populated inside the same i_mmap_rwsem section - otherwise 41723212b535SSteve Capper * racing tasks could either miss the sharing (see huge_pte_offset) or select a 41733212b535SSteve Capper * bad pmd for sharing. 41743212b535SSteve Capper */ 41753212b535SSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 41763212b535SSteve Capper { 41773212b535SSteve Capper struct vm_area_struct *vma = find_vma(mm, addr); 41783212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 41793212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 41803212b535SSteve Capper vma->vm_pgoff; 41813212b535SSteve Capper struct vm_area_struct *svma; 41823212b535SSteve Capper unsigned long saddr; 41833212b535SSteve Capper pte_t *spte = NULL; 41843212b535SSteve Capper pte_t *pte; 4185cb900f41SKirill A. Shutemov spinlock_t *ptl; 41863212b535SSteve Capper 41873212b535SSteve Capper if (!vma_shareable(vma, addr)) 41883212b535SSteve Capper return (pte_t *)pmd_alloc(mm, pud, addr); 41893212b535SSteve Capper 419083cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 41913212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 41923212b535SSteve Capper if (svma == vma) 41933212b535SSteve Capper continue; 41943212b535SSteve Capper 41953212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 41963212b535SSteve Capper if (saddr) { 41973212b535SSteve Capper spte = huge_pte_offset(svma->vm_mm, saddr); 41983212b535SSteve Capper if (spte) { 4199dc6c9a35SKirill A. Shutemov mm_inc_nr_pmds(mm); 42003212b535SSteve Capper get_page(virt_to_page(spte)); 42013212b535SSteve Capper break; 42023212b535SSteve Capper } 42033212b535SSteve Capper } 42043212b535SSteve Capper } 42053212b535SSteve Capper 42063212b535SSteve Capper if (!spte) 42073212b535SSteve Capper goto out; 42083212b535SSteve Capper 4209cb900f41SKirill A. Shutemov ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte); 4210cb900f41SKirill A. Shutemov spin_lock(ptl); 4211dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 42123212b535SSteve Capper pud_populate(mm, pud, 42133212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 4214dc6c9a35SKirill A. Shutemov } else { 42153212b535SSteve Capper put_page(virt_to_page(spte)); 4216dc6c9a35SKirill A. Shutemov mm_inc_nr_pmds(mm); 4217dc6c9a35SKirill A. Shutemov } 4218cb900f41SKirill A. Shutemov spin_unlock(ptl); 42193212b535SSteve Capper out: 42203212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 422183cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 42223212b535SSteve Capper return pte; 42233212b535SSteve Capper } 42243212b535SSteve Capper 42253212b535SSteve Capper /* 42263212b535SSteve Capper * unmap huge page backed by shared pte. 42273212b535SSteve Capper * 42283212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 42293212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 42303212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 42313212b535SSteve Capper * 4232cb900f41SKirill A. Shutemov * called with page table lock held. 42333212b535SSteve Capper * 42343212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 42353212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 42363212b535SSteve Capper */ 42373212b535SSteve Capper int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 42383212b535SSteve Capper { 42393212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 42403212b535SSteve Capper pud_t *pud = pud_offset(pgd, *addr); 42413212b535SSteve Capper 42423212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 42433212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 42443212b535SSteve Capper return 0; 42453212b535SSteve Capper 42463212b535SSteve Capper pud_clear(pud); 42473212b535SSteve Capper put_page(virt_to_page(ptep)); 4248dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 42493212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 42503212b535SSteve Capper return 1; 42513212b535SSteve Capper } 42529e5fc74cSSteve Capper #define want_pmd_share() (1) 42539e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 42549e5fc74cSSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 42559e5fc74cSSteve Capper { 42569e5fc74cSSteve Capper return NULL; 42579e5fc74cSSteve Capper } 4258e81f2d22SZhang Zhen 4259e81f2d22SZhang Zhen int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 4260e81f2d22SZhang Zhen { 4261e81f2d22SZhang Zhen return 0; 4262e81f2d22SZhang Zhen } 42639e5fc74cSSteve Capper #define want_pmd_share() (0) 42643212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 42653212b535SSteve Capper 42669e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 42679e5fc74cSSteve Capper pte_t *huge_pte_alloc(struct mm_struct *mm, 42689e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 42699e5fc74cSSteve Capper { 42709e5fc74cSSteve Capper pgd_t *pgd; 42719e5fc74cSSteve Capper pud_t *pud; 42729e5fc74cSSteve Capper pte_t *pte = NULL; 42739e5fc74cSSteve Capper 42749e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 42759e5fc74cSSteve Capper pud = pud_alloc(mm, pgd, addr); 42769e5fc74cSSteve Capper if (pud) { 42779e5fc74cSSteve Capper if (sz == PUD_SIZE) { 42789e5fc74cSSteve Capper pte = (pte_t *)pud; 42799e5fc74cSSteve Capper } else { 42809e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 42819e5fc74cSSteve Capper if (want_pmd_share() && pud_none(*pud)) 42829e5fc74cSSteve Capper pte = huge_pmd_share(mm, addr, pud); 42839e5fc74cSSteve Capper else 42849e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 42859e5fc74cSSteve Capper } 42869e5fc74cSSteve Capper } 42879e5fc74cSSteve Capper BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte)); 42889e5fc74cSSteve Capper 42899e5fc74cSSteve Capper return pte; 42909e5fc74cSSteve Capper } 42919e5fc74cSSteve Capper 42929e5fc74cSSteve Capper pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) 42939e5fc74cSSteve Capper { 42949e5fc74cSSteve Capper pgd_t *pgd; 42959e5fc74cSSteve Capper pud_t *pud; 42969e5fc74cSSteve Capper pmd_t *pmd = NULL; 42979e5fc74cSSteve Capper 42989e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 42999e5fc74cSSteve Capper if (pgd_present(*pgd)) { 43009e5fc74cSSteve Capper pud = pud_offset(pgd, addr); 43019e5fc74cSSteve Capper if (pud_present(*pud)) { 43029e5fc74cSSteve Capper if (pud_huge(*pud)) 43039e5fc74cSSteve Capper return (pte_t *)pud; 43049e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 43059e5fc74cSSteve Capper } 43069e5fc74cSSteve Capper } 43079e5fc74cSSteve Capper return (pte_t *) pmd; 43089e5fc74cSSteve Capper } 43099e5fc74cSSteve Capper 431061f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 431161f77edaSNaoya Horiguchi 431261f77edaSNaoya Horiguchi /* 431361f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 431461f77edaSNaoya Horiguchi * behavior. 431561f77edaSNaoya Horiguchi */ 431661f77edaSNaoya Horiguchi struct page * __weak 431761f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 431861f77edaSNaoya Horiguchi int write) 431961f77edaSNaoya Horiguchi { 432061f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 432161f77edaSNaoya Horiguchi } 432261f77edaSNaoya Horiguchi 432361f77edaSNaoya Horiguchi struct page * __weak 43249e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 4325e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 43269e5fc74cSSteve Capper { 4327e66f17ffSNaoya Horiguchi struct page *page = NULL; 4328e66f17ffSNaoya Horiguchi spinlock_t *ptl; 4329e66f17ffSNaoya Horiguchi retry: 4330e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 4331e66f17ffSNaoya Horiguchi spin_lock(ptl); 4332e66f17ffSNaoya Horiguchi /* 4333e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 4334e66f17ffSNaoya Horiguchi * unmapped from other threads. 4335e66f17ffSNaoya Horiguchi */ 4336e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 4337e66f17ffSNaoya Horiguchi goto out; 4338e66f17ffSNaoya Horiguchi if (pmd_present(*pmd)) { 433997534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 4340e66f17ffSNaoya Horiguchi if (flags & FOLL_GET) 4341e66f17ffSNaoya Horiguchi get_page(page); 4342e66f17ffSNaoya Horiguchi } else { 4343e66f17ffSNaoya Horiguchi if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) { 4344e66f17ffSNaoya Horiguchi spin_unlock(ptl); 4345e66f17ffSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pmd, ptl); 4346e66f17ffSNaoya Horiguchi goto retry; 4347e66f17ffSNaoya Horiguchi } 4348e66f17ffSNaoya Horiguchi /* 4349e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 4350e66f17ffSNaoya Horiguchi * follow_page_mask(). 4351e66f17ffSNaoya Horiguchi */ 4352e66f17ffSNaoya Horiguchi } 4353e66f17ffSNaoya Horiguchi out: 4354e66f17ffSNaoya Horiguchi spin_unlock(ptl); 43559e5fc74cSSteve Capper return page; 43569e5fc74cSSteve Capper } 43579e5fc74cSSteve Capper 435861f77edaSNaoya Horiguchi struct page * __weak 43599e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 4360e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 43619e5fc74cSSteve Capper { 4362e66f17ffSNaoya Horiguchi if (flags & FOLL_GET) 4363e66f17ffSNaoya Horiguchi return NULL; 43649e5fc74cSSteve Capper 4365e66f17ffSNaoya Horiguchi return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 43669e5fc74cSSteve Capper } 43679e5fc74cSSteve Capper 4368d5bd9106SAndi Kleen #ifdef CONFIG_MEMORY_FAILURE 4369d5bd9106SAndi Kleen 437093f70f90SNaoya Horiguchi /* 437193f70f90SNaoya Horiguchi * This function is called from memory failure code. 437293f70f90SNaoya Horiguchi * Assume the caller holds page lock of the head page. 437393f70f90SNaoya Horiguchi */ 43746de2b1aaSNaoya Horiguchi int dequeue_hwpoisoned_huge_page(struct page *hpage) 437593f70f90SNaoya Horiguchi { 437693f70f90SNaoya Horiguchi struct hstate *h = page_hstate(hpage); 437793f70f90SNaoya Horiguchi int nid = page_to_nid(hpage); 43786de2b1aaSNaoya Horiguchi int ret = -EBUSY; 437993f70f90SNaoya Horiguchi 438093f70f90SNaoya Horiguchi spin_lock(&hugetlb_lock); 43817e1f049eSNaoya Horiguchi /* 43827e1f049eSNaoya Horiguchi * Just checking !page_huge_active is not enough, because that could be 43837e1f049eSNaoya Horiguchi * an isolated/hwpoisoned hugepage (which have >0 refcount). 43847e1f049eSNaoya Horiguchi */ 43857e1f049eSNaoya Horiguchi if (!page_huge_active(hpage) && !page_count(hpage)) { 438656f2fb14SNaoya Horiguchi /* 438756f2fb14SNaoya Horiguchi * Hwpoisoned hugepage isn't linked to activelist or freelist, 438856f2fb14SNaoya Horiguchi * but dangling hpage->lru can trigger list-debug warnings 438956f2fb14SNaoya Horiguchi * (this happens when we call unpoison_memory() on it), 439056f2fb14SNaoya Horiguchi * so let it point to itself with list_del_init(). 439156f2fb14SNaoya Horiguchi */ 439256f2fb14SNaoya Horiguchi list_del_init(&hpage->lru); 43938c6c2ecbSNaoya Horiguchi set_page_refcounted(hpage); 439493f70f90SNaoya Horiguchi h->free_huge_pages--; 439593f70f90SNaoya Horiguchi h->free_huge_pages_node[nid]--; 43966de2b1aaSNaoya Horiguchi ret = 0; 439793f70f90SNaoya Horiguchi } 43986de2b1aaSNaoya Horiguchi spin_unlock(&hugetlb_lock); 43996de2b1aaSNaoya Horiguchi return ret; 44006de2b1aaSNaoya Horiguchi } 44016de2b1aaSNaoya Horiguchi #endif 440231caf665SNaoya Horiguchi 440331caf665SNaoya Horiguchi bool isolate_huge_page(struct page *page, struct list_head *list) 440431caf665SNaoya Horiguchi { 4405bcc54222SNaoya Horiguchi bool ret = true; 4406bcc54222SNaoya Horiguchi 4407309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 440831caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 4409bcc54222SNaoya Horiguchi if (!page_huge_active(page) || !get_page_unless_zero(page)) { 4410bcc54222SNaoya Horiguchi ret = false; 4411bcc54222SNaoya Horiguchi goto unlock; 4412bcc54222SNaoya Horiguchi } 4413bcc54222SNaoya Horiguchi clear_page_huge_active(page); 441431caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 4415bcc54222SNaoya Horiguchi unlock: 441631caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 4417bcc54222SNaoya Horiguchi return ret; 441831caf665SNaoya Horiguchi } 441931caf665SNaoya Horiguchi 442031caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 442131caf665SNaoya Horiguchi { 4422309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 442331caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 4424bcc54222SNaoya Horiguchi set_page_huge_active(page); 442531caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 442631caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 442731caf665SNaoya Horiguchi put_page(page); 442831caf665SNaoya Horiguchi } 4429