11da177e4SLinus Torvalds /* 21da177e4SLinus Torvalds * Generic hugetlb support. 36d49e352SNadia Yvette Chambers * (C) Nadia Yvette Chambers, April 2004 41da177e4SLinus Torvalds */ 51da177e4SLinus Torvalds #include <linux/list.h> 61da177e4SLinus Torvalds #include <linux/init.h> 71da177e4SLinus Torvalds #include <linux/mm.h> 8e1759c21SAlexey Dobriyan #include <linux/seq_file.h> 91da177e4SLinus Torvalds #include <linux/sysctl.h> 101da177e4SLinus Torvalds #include <linux/highmem.h> 11cddb8a5cSAndrea Arcangeli #include <linux/mmu_notifier.h> 121da177e4SLinus Torvalds #include <linux/nodemask.h> 1363551ae0SDavid Gibson #include <linux/pagemap.h> 145da7ca86SChristoph Lameter #include <linux/mempolicy.h> 153b32123dSGideon Israel Dsouza #include <linux/compiler.h> 16aea47ff3SChristoph Lameter #include <linux/cpuset.h> 173935baa9SDavid Gibson #include <linux/mutex.h> 18aa888a74SAndi Kleen #include <linux/bootmem.h> 19a3437870SNishanth Aravamudan #include <linux/sysfs.h> 205a0e3ad6STejun Heo #include <linux/slab.h> 210fe6e20bSNaoya Horiguchi #include <linux/rmap.h> 22fd6a03edSNaoya Horiguchi #include <linux/swap.h> 23fd6a03edSNaoya Horiguchi #include <linux/swapops.h> 24c8721bbbSNaoya Horiguchi #include <linux/page-isolation.h> 258382d914SDavidlohr Bueso #include <linux/jhash.h> 26d6606683SLinus Torvalds 2763551ae0SDavid Gibson #include <asm/page.h> 2863551ae0SDavid Gibson #include <asm/pgtable.h> 2924669e58SAneesh Kumar K.V #include <asm/tlb.h> 3063551ae0SDavid Gibson 3124669e58SAneesh Kumar K.V #include <linux/io.h> 3263551ae0SDavid Gibson #include <linux/hugetlb.h> 339dd540e2SAneesh Kumar K.V #include <linux/hugetlb_cgroup.h> 349a305230SLee Schermerhorn #include <linux/node.h> 357835e98bSNick Piggin #include "internal.h" 361da177e4SLinus Torvalds 37753162cdSAndrey Ryabinin int hugepages_treat_as_movable; 38a5516438SAndi Kleen 39c3f38a38SAneesh Kumar K.V int hugetlb_max_hstate __read_mostly; 40e5ff2159SAndi Kleen unsigned int default_hstate_idx; 41e5ff2159SAndi Kleen struct hstate hstates[HUGE_MAX_HSTATE]; 42641844f5SNaoya Horiguchi /* 43641844f5SNaoya Horiguchi * Minimum page order among possible hugepage sizes, set to a proper value 44641844f5SNaoya Horiguchi * at boot time. 45641844f5SNaoya Horiguchi */ 46641844f5SNaoya Horiguchi static unsigned int minimum_order __read_mostly = UINT_MAX; 47e5ff2159SAndi Kleen 4853ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 4953ba51d2SJon Tollefson 50e5ff2159SAndi Kleen /* for command line parsing */ 51e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 52e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 53e11bfbfcSNick Piggin static unsigned long __initdata default_hstate_size; 549fee021dSVaishali Thakkar static bool __initdata parsed_valid_hugepagesz = true; 55e5ff2159SAndi Kleen 563935baa9SDavid Gibson /* 5731caf665SNaoya Horiguchi * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, 5831caf665SNaoya Horiguchi * free_huge_pages, and surplus_huge_pages. 593935baa9SDavid Gibson */ 60c3f38a38SAneesh Kumar K.V DEFINE_SPINLOCK(hugetlb_lock); 610bd0f9fbSEric Paris 628382d914SDavidlohr Bueso /* 638382d914SDavidlohr Bueso * Serializes faults on the same logical page. This is used to 648382d914SDavidlohr Bueso * prevent spurious OOMs when the hugepage pool is fully utilized. 658382d914SDavidlohr Bueso */ 668382d914SDavidlohr Bueso static int num_fault_mutexes; 67c672c7f2SMike Kravetz struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; 688382d914SDavidlohr Bueso 697ca02d0aSMike Kravetz /* Forward declaration */ 707ca02d0aSMike Kravetz static int hugetlb_acct_memory(struct hstate *h, long delta); 717ca02d0aSMike Kravetz 7290481622SDavid Gibson static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) 7390481622SDavid Gibson { 7490481622SDavid Gibson bool free = (spool->count == 0) && (spool->used_hpages == 0); 7590481622SDavid Gibson 7690481622SDavid Gibson spin_unlock(&spool->lock); 7790481622SDavid Gibson 7890481622SDavid Gibson /* If no pages are used, and no other handles to the subpool 797ca02d0aSMike Kravetz * remain, give up any reservations mased on minimum size and 807ca02d0aSMike Kravetz * free the subpool */ 817ca02d0aSMike Kravetz if (free) { 827ca02d0aSMike Kravetz if (spool->min_hpages != -1) 837ca02d0aSMike Kravetz hugetlb_acct_memory(spool->hstate, 847ca02d0aSMike Kravetz -spool->min_hpages); 8590481622SDavid Gibson kfree(spool); 8690481622SDavid Gibson } 877ca02d0aSMike Kravetz } 8890481622SDavid Gibson 897ca02d0aSMike Kravetz struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, 907ca02d0aSMike Kravetz long min_hpages) 9190481622SDavid Gibson { 9290481622SDavid Gibson struct hugepage_subpool *spool; 9390481622SDavid Gibson 94c6a91820SMike Kravetz spool = kzalloc(sizeof(*spool), GFP_KERNEL); 9590481622SDavid Gibson if (!spool) 9690481622SDavid Gibson return NULL; 9790481622SDavid Gibson 9890481622SDavid Gibson spin_lock_init(&spool->lock); 9990481622SDavid Gibson spool->count = 1; 1007ca02d0aSMike Kravetz spool->max_hpages = max_hpages; 1017ca02d0aSMike Kravetz spool->hstate = h; 1027ca02d0aSMike Kravetz spool->min_hpages = min_hpages; 1037ca02d0aSMike Kravetz 1047ca02d0aSMike Kravetz if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { 1057ca02d0aSMike Kravetz kfree(spool); 1067ca02d0aSMike Kravetz return NULL; 1077ca02d0aSMike Kravetz } 1087ca02d0aSMike Kravetz spool->rsv_hpages = min_hpages; 10990481622SDavid Gibson 11090481622SDavid Gibson return spool; 11190481622SDavid Gibson } 11290481622SDavid Gibson 11390481622SDavid Gibson void hugepage_put_subpool(struct hugepage_subpool *spool) 11490481622SDavid Gibson { 11590481622SDavid Gibson spin_lock(&spool->lock); 11690481622SDavid Gibson BUG_ON(!spool->count); 11790481622SDavid Gibson spool->count--; 11890481622SDavid Gibson unlock_or_release_subpool(spool); 11990481622SDavid Gibson } 12090481622SDavid Gibson 1211c5ecae3SMike Kravetz /* 1221c5ecae3SMike Kravetz * Subpool accounting for allocating and reserving pages. 1231c5ecae3SMike Kravetz * Return -ENOMEM if there are not enough resources to satisfy the 1241c5ecae3SMike Kravetz * the request. Otherwise, return the number of pages by which the 1251c5ecae3SMike Kravetz * global pools must be adjusted (upward). The returned value may 1261c5ecae3SMike Kravetz * only be different than the passed value (delta) in the case where 1271c5ecae3SMike Kravetz * a subpool minimum size must be manitained. 1281c5ecae3SMike Kravetz */ 1291c5ecae3SMike Kravetz static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, 13090481622SDavid Gibson long delta) 13190481622SDavid Gibson { 1321c5ecae3SMike Kravetz long ret = delta; 13390481622SDavid Gibson 13490481622SDavid Gibson if (!spool) 1351c5ecae3SMike Kravetz return ret; 13690481622SDavid Gibson 13790481622SDavid Gibson spin_lock(&spool->lock); 13890481622SDavid Gibson 1391c5ecae3SMike Kravetz if (spool->max_hpages != -1) { /* maximum size accounting */ 1401c5ecae3SMike Kravetz if ((spool->used_hpages + delta) <= spool->max_hpages) 1411c5ecae3SMike Kravetz spool->used_hpages += delta; 1421c5ecae3SMike Kravetz else { 1431c5ecae3SMike Kravetz ret = -ENOMEM; 1441c5ecae3SMike Kravetz goto unlock_ret; 1451c5ecae3SMike Kravetz } 1461c5ecae3SMike Kravetz } 1471c5ecae3SMike Kravetz 14809a95e29SMike Kravetz /* minimum size accounting */ 14909a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->rsv_hpages) { 1501c5ecae3SMike Kravetz if (delta > spool->rsv_hpages) { 1511c5ecae3SMike Kravetz /* 1521c5ecae3SMike Kravetz * Asking for more reserves than those already taken on 1531c5ecae3SMike Kravetz * behalf of subpool. Return difference. 1541c5ecae3SMike Kravetz */ 1551c5ecae3SMike Kravetz ret = delta - spool->rsv_hpages; 1561c5ecae3SMike Kravetz spool->rsv_hpages = 0; 1571c5ecae3SMike Kravetz } else { 1581c5ecae3SMike Kravetz ret = 0; /* reserves already accounted for */ 1591c5ecae3SMike Kravetz spool->rsv_hpages -= delta; 1601c5ecae3SMike Kravetz } 1611c5ecae3SMike Kravetz } 1621c5ecae3SMike Kravetz 1631c5ecae3SMike Kravetz unlock_ret: 1641c5ecae3SMike Kravetz spin_unlock(&spool->lock); 16590481622SDavid Gibson return ret; 16690481622SDavid Gibson } 16790481622SDavid Gibson 1681c5ecae3SMike Kravetz /* 1691c5ecae3SMike Kravetz * Subpool accounting for freeing and unreserving pages. 1701c5ecae3SMike Kravetz * Return the number of global page reservations that must be dropped. 1711c5ecae3SMike Kravetz * The return value may only be different than the passed value (delta) 1721c5ecae3SMike Kravetz * in the case where a subpool minimum size must be maintained. 1731c5ecae3SMike Kravetz */ 1741c5ecae3SMike Kravetz static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, 17590481622SDavid Gibson long delta) 17690481622SDavid Gibson { 1771c5ecae3SMike Kravetz long ret = delta; 1781c5ecae3SMike Kravetz 17990481622SDavid Gibson if (!spool) 1801c5ecae3SMike Kravetz return delta; 18190481622SDavid Gibson 18290481622SDavid Gibson spin_lock(&spool->lock); 1831c5ecae3SMike Kravetz 1841c5ecae3SMike Kravetz if (spool->max_hpages != -1) /* maximum size accounting */ 18590481622SDavid Gibson spool->used_hpages -= delta; 1861c5ecae3SMike Kravetz 18709a95e29SMike Kravetz /* minimum size accounting */ 18809a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { 1891c5ecae3SMike Kravetz if (spool->rsv_hpages + delta <= spool->min_hpages) 1901c5ecae3SMike Kravetz ret = 0; 1911c5ecae3SMike Kravetz else 1921c5ecae3SMike Kravetz ret = spool->rsv_hpages + delta - spool->min_hpages; 1931c5ecae3SMike Kravetz 1941c5ecae3SMike Kravetz spool->rsv_hpages += delta; 1951c5ecae3SMike Kravetz if (spool->rsv_hpages > spool->min_hpages) 1961c5ecae3SMike Kravetz spool->rsv_hpages = spool->min_hpages; 1971c5ecae3SMike Kravetz } 1981c5ecae3SMike Kravetz 1991c5ecae3SMike Kravetz /* 2001c5ecae3SMike Kravetz * If hugetlbfs_put_super couldn't free spool due to an outstanding 2011c5ecae3SMike Kravetz * quota reference, free it now. 2021c5ecae3SMike Kravetz */ 20390481622SDavid Gibson unlock_or_release_subpool(spool); 2041c5ecae3SMike Kravetz 2051c5ecae3SMike Kravetz return ret; 20690481622SDavid Gibson } 20790481622SDavid Gibson 20890481622SDavid Gibson static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 20990481622SDavid Gibson { 21090481622SDavid Gibson return HUGETLBFS_SB(inode->i_sb)->spool; 21190481622SDavid Gibson } 21290481622SDavid Gibson 21390481622SDavid Gibson static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 21490481622SDavid Gibson { 215496ad9aaSAl Viro return subpool_inode(file_inode(vma->vm_file)); 21690481622SDavid Gibson } 21790481622SDavid Gibson 218e7c4b0bfSAndy Whitcroft /* 21996822904SAndy Whitcroft * Region tracking -- allows tracking of reservations and instantiated pages 22096822904SAndy Whitcroft * across the pages in a mapping. 22184afd99bSAndy Whitcroft * 2221dd308a7SMike Kravetz * The region data structures are embedded into a resv_map and protected 2231dd308a7SMike Kravetz * by a resv_map's lock. The set of regions within the resv_map represent 2241dd308a7SMike Kravetz * reservations for huge pages, or huge pages that have already been 2251dd308a7SMike Kravetz * instantiated within the map. The from and to elements are huge page 2261dd308a7SMike Kravetz * indicies into the associated mapping. from indicates the starting index 2271dd308a7SMike Kravetz * of the region. to represents the first index past the end of the region. 2281dd308a7SMike Kravetz * 2291dd308a7SMike Kravetz * For example, a file region structure with from == 0 and to == 4 represents 2301dd308a7SMike Kravetz * four huge pages in a mapping. It is important to note that the to element 2311dd308a7SMike Kravetz * represents the first element past the end of the region. This is used in 2321dd308a7SMike Kravetz * arithmetic as 4(to) - 0(from) = 4 huge pages in the region. 2331dd308a7SMike Kravetz * 2341dd308a7SMike Kravetz * Interval notation of the form [from, to) will be used to indicate that 2351dd308a7SMike Kravetz * the endpoint from is inclusive and to is exclusive. 23696822904SAndy Whitcroft */ 23796822904SAndy Whitcroft struct file_region { 23896822904SAndy Whitcroft struct list_head link; 23996822904SAndy Whitcroft long from; 24096822904SAndy Whitcroft long to; 24196822904SAndy Whitcroft }; 24296822904SAndy Whitcroft 2431dd308a7SMike Kravetz /* 2441dd308a7SMike Kravetz * Add the huge page range represented by [f, t) to the reserve 2455e911373SMike Kravetz * map. In the normal case, existing regions will be expanded 2465e911373SMike Kravetz * to accommodate the specified range. Sufficient regions should 2475e911373SMike Kravetz * exist for expansion due to the previous call to region_chg 2485e911373SMike Kravetz * with the same range. However, it is possible that region_del 2495e911373SMike Kravetz * could have been called after region_chg and modifed the map 2505e911373SMike Kravetz * in such a way that no region exists to be expanded. In this 2515e911373SMike Kravetz * case, pull a region descriptor from the cache associated with 2525e911373SMike Kravetz * the map and use that for the new range. 253cf3ad20bSMike Kravetz * 254cf3ad20bSMike Kravetz * Return the number of new huge pages added to the map. This 255cf3ad20bSMike Kravetz * number is greater than or equal to zero. 2561dd308a7SMike Kravetz */ 2571406ec9bSJoonsoo Kim static long region_add(struct resv_map *resv, long f, long t) 25896822904SAndy Whitcroft { 2591406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 26096822904SAndy Whitcroft struct file_region *rg, *nrg, *trg; 261cf3ad20bSMike Kravetz long add = 0; 26296822904SAndy Whitcroft 2637b24d861SDavidlohr Bueso spin_lock(&resv->lock); 26496822904SAndy Whitcroft /* Locate the region we are either in or before. */ 26596822904SAndy Whitcroft list_for_each_entry(rg, head, link) 26696822904SAndy Whitcroft if (f <= rg->to) 26796822904SAndy Whitcroft break; 26896822904SAndy Whitcroft 2695e911373SMike Kravetz /* 2705e911373SMike Kravetz * If no region exists which can be expanded to include the 2715e911373SMike Kravetz * specified range, the list must have been modified by an 2725e911373SMike Kravetz * interleving call to region_del(). Pull a region descriptor 2735e911373SMike Kravetz * from the cache and use it for this range. 2745e911373SMike Kravetz */ 2755e911373SMike Kravetz if (&rg->link == head || t < rg->from) { 2765e911373SMike Kravetz VM_BUG_ON(resv->region_cache_count <= 0); 2775e911373SMike Kravetz 2785e911373SMike Kravetz resv->region_cache_count--; 2795e911373SMike Kravetz nrg = list_first_entry(&resv->region_cache, struct file_region, 2805e911373SMike Kravetz link); 2815e911373SMike Kravetz list_del(&nrg->link); 2825e911373SMike Kravetz 2835e911373SMike Kravetz nrg->from = f; 2845e911373SMike Kravetz nrg->to = t; 2855e911373SMike Kravetz list_add(&nrg->link, rg->link.prev); 2865e911373SMike Kravetz 2875e911373SMike Kravetz add += t - f; 2885e911373SMike Kravetz goto out_locked; 2895e911373SMike Kravetz } 2905e911373SMike Kravetz 29196822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 29296822904SAndy Whitcroft if (f > rg->from) 29396822904SAndy Whitcroft f = rg->from; 29496822904SAndy Whitcroft 29596822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 29696822904SAndy Whitcroft nrg = rg; 29796822904SAndy Whitcroft list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 29896822904SAndy Whitcroft if (&rg->link == head) 29996822904SAndy Whitcroft break; 30096822904SAndy Whitcroft if (rg->from > t) 30196822904SAndy Whitcroft break; 30296822904SAndy Whitcroft 30396822904SAndy Whitcroft /* If this area reaches higher then extend our area to 30496822904SAndy Whitcroft * include it completely. If this is not the first area 30596822904SAndy Whitcroft * which we intend to reuse, free it. */ 30696822904SAndy Whitcroft if (rg->to > t) 30796822904SAndy Whitcroft t = rg->to; 30896822904SAndy Whitcroft if (rg != nrg) { 309cf3ad20bSMike Kravetz /* Decrement return value by the deleted range. 310cf3ad20bSMike Kravetz * Another range will span this area so that by 311cf3ad20bSMike Kravetz * end of routine add will be >= zero 312cf3ad20bSMike Kravetz */ 313cf3ad20bSMike Kravetz add -= (rg->to - rg->from); 31496822904SAndy Whitcroft list_del(&rg->link); 31596822904SAndy Whitcroft kfree(rg); 31696822904SAndy Whitcroft } 31796822904SAndy Whitcroft } 318cf3ad20bSMike Kravetz 319cf3ad20bSMike Kravetz add += (nrg->from - f); /* Added to beginning of region */ 32096822904SAndy Whitcroft nrg->from = f; 321cf3ad20bSMike Kravetz add += t - nrg->to; /* Added to end of region */ 32296822904SAndy Whitcroft nrg->to = t; 323cf3ad20bSMike Kravetz 3245e911373SMike Kravetz out_locked: 3255e911373SMike Kravetz resv->adds_in_progress--; 3267b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 327cf3ad20bSMike Kravetz VM_BUG_ON(add < 0); 328cf3ad20bSMike Kravetz return add; 32996822904SAndy Whitcroft } 33096822904SAndy Whitcroft 3311dd308a7SMike Kravetz /* 3321dd308a7SMike Kravetz * Examine the existing reserve map and determine how many 3331dd308a7SMike Kravetz * huge pages in the specified range [f, t) are NOT currently 3341dd308a7SMike Kravetz * represented. This routine is called before a subsequent 3351dd308a7SMike Kravetz * call to region_add that will actually modify the reserve 3361dd308a7SMike Kravetz * map to add the specified range [f, t). region_chg does 3371dd308a7SMike Kravetz * not change the number of huge pages represented by the 3381dd308a7SMike Kravetz * map. However, if the existing regions in the map can not 3391dd308a7SMike Kravetz * be expanded to represent the new range, a new file_region 3401dd308a7SMike Kravetz * structure is added to the map as a placeholder. This is 3411dd308a7SMike Kravetz * so that the subsequent region_add call will have all the 3421dd308a7SMike Kravetz * regions it needs and will not fail. 3431dd308a7SMike Kravetz * 3445e911373SMike Kravetz * Upon entry, region_chg will also examine the cache of region descriptors 3455e911373SMike Kravetz * associated with the map. If there are not enough descriptors cached, one 3465e911373SMike Kravetz * will be allocated for the in progress add operation. 3475e911373SMike Kravetz * 3485e911373SMike Kravetz * Returns the number of huge pages that need to be added to the existing 3495e911373SMike Kravetz * reservation map for the range [f, t). This number is greater or equal to 3505e911373SMike Kravetz * zero. -ENOMEM is returned if a new file_region structure or cache entry 3515e911373SMike Kravetz * is needed and can not be allocated. 3521dd308a7SMike Kravetz */ 3531406ec9bSJoonsoo Kim static long region_chg(struct resv_map *resv, long f, long t) 35496822904SAndy Whitcroft { 3551406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 3567b24d861SDavidlohr Bueso struct file_region *rg, *nrg = NULL; 35796822904SAndy Whitcroft long chg = 0; 35896822904SAndy Whitcroft 3597b24d861SDavidlohr Bueso retry: 3607b24d861SDavidlohr Bueso spin_lock(&resv->lock); 3615e911373SMike Kravetz retry_locked: 3625e911373SMike Kravetz resv->adds_in_progress++; 3635e911373SMike Kravetz 3645e911373SMike Kravetz /* 3655e911373SMike Kravetz * Check for sufficient descriptors in the cache to accommodate 3665e911373SMike Kravetz * the number of in progress add operations. 3675e911373SMike Kravetz */ 3685e911373SMike Kravetz if (resv->adds_in_progress > resv->region_cache_count) { 3695e911373SMike Kravetz struct file_region *trg; 3705e911373SMike Kravetz 3715e911373SMike Kravetz VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1); 3725e911373SMike Kravetz /* Must drop lock to allocate a new descriptor. */ 3735e911373SMike Kravetz resv->adds_in_progress--; 3745e911373SMike Kravetz spin_unlock(&resv->lock); 3755e911373SMike Kravetz 3765e911373SMike Kravetz trg = kmalloc(sizeof(*trg), GFP_KERNEL); 377dbe409e4SMike Kravetz if (!trg) { 378dbe409e4SMike Kravetz kfree(nrg); 3795e911373SMike Kravetz return -ENOMEM; 380dbe409e4SMike Kravetz } 3815e911373SMike Kravetz 3825e911373SMike Kravetz spin_lock(&resv->lock); 3835e911373SMike Kravetz list_add(&trg->link, &resv->region_cache); 3845e911373SMike Kravetz resv->region_cache_count++; 3855e911373SMike Kravetz goto retry_locked; 3865e911373SMike Kravetz } 3875e911373SMike Kravetz 38896822904SAndy Whitcroft /* Locate the region we are before or in. */ 38996822904SAndy Whitcroft list_for_each_entry(rg, head, link) 39096822904SAndy Whitcroft if (f <= rg->to) 39196822904SAndy Whitcroft break; 39296822904SAndy Whitcroft 39396822904SAndy Whitcroft /* If we are below the current region then a new region is required. 39496822904SAndy Whitcroft * Subtle, allocate a new region at the position but make it zero 39596822904SAndy Whitcroft * size such that we can guarantee to record the reservation. */ 39696822904SAndy Whitcroft if (&rg->link == head || t < rg->from) { 3977b24d861SDavidlohr Bueso if (!nrg) { 3985e911373SMike Kravetz resv->adds_in_progress--; 3997b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 40096822904SAndy Whitcroft nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 40196822904SAndy Whitcroft if (!nrg) 40296822904SAndy Whitcroft return -ENOMEM; 4037b24d861SDavidlohr Bueso 40496822904SAndy Whitcroft nrg->from = f; 40596822904SAndy Whitcroft nrg->to = f; 40696822904SAndy Whitcroft INIT_LIST_HEAD(&nrg->link); 4077b24d861SDavidlohr Bueso goto retry; 4087b24d861SDavidlohr Bueso } 40996822904SAndy Whitcroft 4107b24d861SDavidlohr Bueso list_add(&nrg->link, rg->link.prev); 4117b24d861SDavidlohr Bueso chg = t - f; 4127b24d861SDavidlohr Bueso goto out_nrg; 41396822904SAndy Whitcroft } 41496822904SAndy Whitcroft 41596822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 41696822904SAndy Whitcroft if (f > rg->from) 41796822904SAndy Whitcroft f = rg->from; 41896822904SAndy Whitcroft chg = t - f; 41996822904SAndy Whitcroft 42096822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 42196822904SAndy Whitcroft list_for_each_entry(rg, rg->link.prev, link) { 42296822904SAndy Whitcroft if (&rg->link == head) 42396822904SAndy Whitcroft break; 42496822904SAndy Whitcroft if (rg->from > t) 4257b24d861SDavidlohr Bueso goto out; 42696822904SAndy Whitcroft 42725985edcSLucas De Marchi /* We overlap with this area, if it extends further than 42896822904SAndy Whitcroft * us then we must extend ourselves. Account for its 42996822904SAndy Whitcroft * existing reservation. */ 43096822904SAndy Whitcroft if (rg->to > t) { 43196822904SAndy Whitcroft chg += rg->to - t; 43296822904SAndy Whitcroft t = rg->to; 43396822904SAndy Whitcroft } 43496822904SAndy Whitcroft chg -= rg->to - rg->from; 43596822904SAndy Whitcroft } 4367b24d861SDavidlohr Bueso 4377b24d861SDavidlohr Bueso out: 4387b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 4397b24d861SDavidlohr Bueso /* We already know we raced and no longer need the new region */ 4407b24d861SDavidlohr Bueso kfree(nrg); 4417b24d861SDavidlohr Bueso return chg; 4427b24d861SDavidlohr Bueso out_nrg: 4437b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 44496822904SAndy Whitcroft return chg; 44596822904SAndy Whitcroft } 44696822904SAndy Whitcroft 4471dd308a7SMike Kravetz /* 4485e911373SMike Kravetz * Abort the in progress add operation. The adds_in_progress field 4495e911373SMike Kravetz * of the resv_map keeps track of the operations in progress between 4505e911373SMike Kravetz * calls to region_chg and region_add. Operations are sometimes 4515e911373SMike Kravetz * aborted after the call to region_chg. In such cases, region_abort 4525e911373SMike Kravetz * is called to decrement the adds_in_progress counter. 4535e911373SMike Kravetz * 4545e911373SMike Kravetz * NOTE: The range arguments [f, t) are not needed or used in this 4555e911373SMike Kravetz * routine. They are kept to make reading the calling code easier as 4565e911373SMike Kravetz * arguments will match the associated region_chg call. 4575e911373SMike Kravetz */ 4585e911373SMike Kravetz static void region_abort(struct resv_map *resv, long f, long t) 4595e911373SMike Kravetz { 4605e911373SMike Kravetz spin_lock(&resv->lock); 4615e911373SMike Kravetz VM_BUG_ON(!resv->region_cache_count); 4625e911373SMike Kravetz resv->adds_in_progress--; 4635e911373SMike Kravetz spin_unlock(&resv->lock); 4645e911373SMike Kravetz } 4655e911373SMike Kravetz 4665e911373SMike Kravetz /* 467feba16e2SMike Kravetz * Delete the specified range [f, t) from the reserve map. If the 468feba16e2SMike Kravetz * t parameter is LONG_MAX, this indicates that ALL regions after f 469feba16e2SMike Kravetz * should be deleted. Locate the regions which intersect [f, t) 470feba16e2SMike Kravetz * and either trim, delete or split the existing regions. 471feba16e2SMike Kravetz * 472feba16e2SMike Kravetz * Returns the number of huge pages deleted from the reserve map. 473feba16e2SMike Kravetz * In the normal case, the return value is zero or more. In the 474feba16e2SMike Kravetz * case where a region must be split, a new region descriptor must 475feba16e2SMike Kravetz * be allocated. If the allocation fails, -ENOMEM will be returned. 476feba16e2SMike Kravetz * NOTE: If the parameter t == LONG_MAX, then we will never split 477feba16e2SMike Kravetz * a region and possibly return -ENOMEM. Callers specifying 478feba16e2SMike Kravetz * t == LONG_MAX do not need to check for -ENOMEM error. 4791dd308a7SMike Kravetz */ 480feba16e2SMike Kravetz static long region_del(struct resv_map *resv, long f, long t) 48196822904SAndy Whitcroft { 4821406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 48396822904SAndy Whitcroft struct file_region *rg, *trg; 484feba16e2SMike Kravetz struct file_region *nrg = NULL; 485feba16e2SMike Kravetz long del = 0; 48696822904SAndy Whitcroft 487feba16e2SMike Kravetz retry: 4887b24d861SDavidlohr Bueso spin_lock(&resv->lock); 489feba16e2SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 490dbe409e4SMike Kravetz /* 491dbe409e4SMike Kravetz * Skip regions before the range to be deleted. file_region 492dbe409e4SMike Kravetz * ranges are normally of the form [from, to). However, there 493dbe409e4SMike Kravetz * may be a "placeholder" entry in the map which is of the form 494dbe409e4SMike Kravetz * (from, to) with from == to. Check for placeholder entries 495dbe409e4SMike Kravetz * at the beginning of the range to be deleted. 496dbe409e4SMike Kravetz */ 497dbe409e4SMike Kravetz if (rg->to <= f && (rg->to != rg->from || rg->to != f)) 498feba16e2SMike Kravetz continue; 499dbe409e4SMike Kravetz 500feba16e2SMike Kravetz if (rg->from >= t) 50196822904SAndy Whitcroft break; 50296822904SAndy Whitcroft 503feba16e2SMike Kravetz if (f > rg->from && t < rg->to) { /* Must split region */ 504feba16e2SMike Kravetz /* 505feba16e2SMike Kravetz * Check for an entry in the cache before dropping 506feba16e2SMike Kravetz * lock and attempting allocation. 507feba16e2SMike Kravetz */ 508feba16e2SMike Kravetz if (!nrg && 509feba16e2SMike Kravetz resv->region_cache_count > resv->adds_in_progress) { 510feba16e2SMike Kravetz nrg = list_first_entry(&resv->region_cache, 511feba16e2SMike Kravetz struct file_region, 512feba16e2SMike Kravetz link); 513feba16e2SMike Kravetz list_del(&nrg->link); 514feba16e2SMike Kravetz resv->region_cache_count--; 51596822904SAndy Whitcroft } 51696822904SAndy Whitcroft 517feba16e2SMike Kravetz if (!nrg) { 518feba16e2SMike Kravetz spin_unlock(&resv->lock); 519feba16e2SMike Kravetz nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 520feba16e2SMike Kravetz if (!nrg) 521feba16e2SMike Kravetz return -ENOMEM; 522feba16e2SMike Kravetz goto retry; 523feba16e2SMike Kravetz } 524feba16e2SMike Kravetz 525feba16e2SMike Kravetz del += t - f; 526feba16e2SMike Kravetz 527feba16e2SMike Kravetz /* New entry for end of split region */ 528feba16e2SMike Kravetz nrg->from = t; 529feba16e2SMike Kravetz nrg->to = rg->to; 530feba16e2SMike Kravetz INIT_LIST_HEAD(&nrg->link); 531feba16e2SMike Kravetz 532feba16e2SMike Kravetz /* Original entry is trimmed */ 533feba16e2SMike Kravetz rg->to = f; 534feba16e2SMike Kravetz 535feba16e2SMike Kravetz list_add(&nrg->link, &rg->link); 536feba16e2SMike Kravetz nrg = NULL; 53796822904SAndy Whitcroft break; 538feba16e2SMike Kravetz } 539feba16e2SMike Kravetz 540feba16e2SMike Kravetz if (f <= rg->from && t >= rg->to) { /* Remove entire region */ 541feba16e2SMike Kravetz del += rg->to - rg->from; 54296822904SAndy Whitcroft list_del(&rg->link); 54396822904SAndy Whitcroft kfree(rg); 544feba16e2SMike Kravetz continue; 54596822904SAndy Whitcroft } 5467b24d861SDavidlohr Bueso 547feba16e2SMike Kravetz if (f <= rg->from) { /* Trim beginning of region */ 548feba16e2SMike Kravetz del += t - rg->from; 549feba16e2SMike Kravetz rg->from = t; 550feba16e2SMike Kravetz } else { /* Trim end of region */ 551feba16e2SMike Kravetz del += rg->to - f; 552feba16e2SMike Kravetz rg->to = f; 553feba16e2SMike Kravetz } 554feba16e2SMike Kravetz } 555feba16e2SMike Kravetz 5567b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 557feba16e2SMike Kravetz kfree(nrg); 558feba16e2SMike Kravetz return del; 55996822904SAndy Whitcroft } 56096822904SAndy Whitcroft 5611dd308a7SMike Kravetz /* 562b5cec28dSMike Kravetz * A rare out of memory error was encountered which prevented removal of 563b5cec28dSMike Kravetz * the reserve map region for a page. The huge page itself was free'ed 564b5cec28dSMike Kravetz * and removed from the page cache. This routine will adjust the subpool 565b5cec28dSMike Kravetz * usage count, and the global reserve count if needed. By incrementing 566b5cec28dSMike Kravetz * these counts, the reserve map entry which could not be deleted will 567b5cec28dSMike Kravetz * appear as a "reserved" entry instead of simply dangling with incorrect 568b5cec28dSMike Kravetz * counts. 569b5cec28dSMike Kravetz */ 57072e2936cSzhong jiang void hugetlb_fix_reserve_counts(struct inode *inode) 571b5cec28dSMike Kravetz { 572b5cec28dSMike Kravetz struct hugepage_subpool *spool = subpool_inode(inode); 573b5cec28dSMike Kravetz long rsv_adjust; 574b5cec28dSMike Kravetz 575b5cec28dSMike Kravetz rsv_adjust = hugepage_subpool_get_pages(spool, 1); 57672e2936cSzhong jiang if (rsv_adjust) { 577b5cec28dSMike Kravetz struct hstate *h = hstate_inode(inode); 578b5cec28dSMike Kravetz 579b5cec28dSMike Kravetz hugetlb_acct_memory(h, 1); 580b5cec28dSMike Kravetz } 581b5cec28dSMike Kravetz } 582b5cec28dSMike Kravetz 583b5cec28dSMike Kravetz /* 5841dd308a7SMike Kravetz * Count and return the number of huge pages in the reserve map 5851dd308a7SMike Kravetz * that intersect with the range [f, t). 5861dd308a7SMike Kravetz */ 5871406ec9bSJoonsoo Kim static long region_count(struct resv_map *resv, long f, long t) 58884afd99bSAndy Whitcroft { 5891406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 59084afd99bSAndy Whitcroft struct file_region *rg; 59184afd99bSAndy Whitcroft long chg = 0; 59284afd99bSAndy Whitcroft 5937b24d861SDavidlohr Bueso spin_lock(&resv->lock); 59484afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 59584afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 596f2135a4aSWang Sheng-Hui long seg_from; 597f2135a4aSWang Sheng-Hui long seg_to; 59884afd99bSAndy Whitcroft 59984afd99bSAndy Whitcroft if (rg->to <= f) 60084afd99bSAndy Whitcroft continue; 60184afd99bSAndy Whitcroft if (rg->from >= t) 60284afd99bSAndy Whitcroft break; 60384afd99bSAndy Whitcroft 60484afd99bSAndy Whitcroft seg_from = max(rg->from, f); 60584afd99bSAndy Whitcroft seg_to = min(rg->to, t); 60684afd99bSAndy Whitcroft 60784afd99bSAndy Whitcroft chg += seg_to - seg_from; 60884afd99bSAndy Whitcroft } 6097b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 61084afd99bSAndy Whitcroft 61184afd99bSAndy Whitcroft return chg; 61284afd99bSAndy Whitcroft } 61384afd99bSAndy Whitcroft 61496822904SAndy Whitcroft /* 615e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 616e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 617e7c4b0bfSAndy Whitcroft */ 618a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 619a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 620e7c4b0bfSAndy Whitcroft { 621a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 622a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 623e7c4b0bfSAndy Whitcroft } 624e7c4b0bfSAndy Whitcroft 6250fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 6260fe6e20bSNaoya Horiguchi unsigned long address) 6270fe6e20bSNaoya Horiguchi { 6280fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 6290fe6e20bSNaoya Horiguchi } 630dee41079SDan Williams EXPORT_SYMBOL_GPL(linear_hugepage_index); 6310fe6e20bSNaoya Horiguchi 63284afd99bSAndy Whitcroft /* 63308fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 63408fba699SMel Gorman * cases this will be same size as used by the page table entries. 63508fba699SMel Gorman */ 63608fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 63708fba699SMel Gorman { 63808fba699SMel Gorman struct hstate *hstate; 63908fba699SMel Gorman 64008fba699SMel Gorman if (!is_vm_hugetlb_page(vma)) 64108fba699SMel Gorman return PAGE_SIZE; 64208fba699SMel Gorman 64308fba699SMel Gorman hstate = hstate_vma(vma); 64408fba699SMel Gorman 6452415cf12SWanpeng Li return 1UL << huge_page_shift(hstate); 64608fba699SMel Gorman } 647f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 64808fba699SMel Gorman 64908fba699SMel Gorman /* 6503340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 6513340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 6523340289dSMel Gorman * architectures where it differs, an architecture-specific version of this 6533340289dSMel Gorman * function is required. 6543340289dSMel Gorman */ 6553340289dSMel Gorman #ifndef vma_mmu_pagesize 6563340289dSMel Gorman unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 6573340289dSMel Gorman { 6583340289dSMel Gorman return vma_kernel_pagesize(vma); 6593340289dSMel Gorman } 6603340289dSMel Gorman #endif 6613340289dSMel Gorman 6623340289dSMel Gorman /* 66384afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 66484afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 66584afd99bSAndy Whitcroft * alignment. 66684afd99bSAndy Whitcroft */ 66784afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 66884afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 66904f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 67084afd99bSAndy Whitcroft 671a1e78772SMel Gorman /* 672a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 673a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 674a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 675a1e78772SMel Gorman * 676a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 677a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 678a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 679a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 68084afd99bSAndy Whitcroft * 68184afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 68284afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 68384afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 68484afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 68584afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 68684afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 68784afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 68884afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 689a1e78772SMel Gorman */ 690e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 691e7c4b0bfSAndy Whitcroft { 692e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 693e7c4b0bfSAndy Whitcroft } 694e7c4b0bfSAndy Whitcroft 695e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 696e7c4b0bfSAndy Whitcroft unsigned long value) 697e7c4b0bfSAndy Whitcroft { 698e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 699e7c4b0bfSAndy Whitcroft } 700e7c4b0bfSAndy Whitcroft 7019119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 70284afd99bSAndy Whitcroft { 70384afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 7045e911373SMike Kravetz struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 7055e911373SMike Kravetz 7065e911373SMike Kravetz if (!resv_map || !rg) { 7075e911373SMike Kravetz kfree(resv_map); 7085e911373SMike Kravetz kfree(rg); 70984afd99bSAndy Whitcroft return NULL; 7105e911373SMike Kravetz } 71184afd99bSAndy Whitcroft 71284afd99bSAndy Whitcroft kref_init(&resv_map->refs); 7137b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 71484afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 71584afd99bSAndy Whitcroft 7165e911373SMike Kravetz resv_map->adds_in_progress = 0; 7175e911373SMike Kravetz 7185e911373SMike Kravetz INIT_LIST_HEAD(&resv_map->region_cache); 7195e911373SMike Kravetz list_add(&rg->link, &resv_map->region_cache); 7205e911373SMike Kravetz resv_map->region_cache_count = 1; 7215e911373SMike Kravetz 72284afd99bSAndy Whitcroft return resv_map; 72384afd99bSAndy Whitcroft } 72484afd99bSAndy Whitcroft 7259119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 72684afd99bSAndy Whitcroft { 72784afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 7285e911373SMike Kravetz struct list_head *head = &resv_map->region_cache; 7295e911373SMike Kravetz struct file_region *rg, *trg; 73084afd99bSAndy Whitcroft 73184afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 732feba16e2SMike Kravetz region_del(resv_map, 0, LONG_MAX); 7335e911373SMike Kravetz 7345e911373SMike Kravetz /* ... and any entries left in the cache */ 7355e911373SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 7365e911373SMike Kravetz list_del(&rg->link); 7375e911373SMike Kravetz kfree(rg); 7385e911373SMike Kravetz } 7395e911373SMike Kravetz 7405e911373SMike Kravetz VM_BUG_ON(resv_map->adds_in_progress); 7415e911373SMike Kravetz 74284afd99bSAndy Whitcroft kfree(resv_map); 74384afd99bSAndy Whitcroft } 74484afd99bSAndy Whitcroft 7454e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 7464e35f483SJoonsoo Kim { 7474e35f483SJoonsoo Kim return inode->i_mapping->private_data; 7484e35f483SJoonsoo Kim } 7494e35f483SJoonsoo Kim 75084afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 751a1e78772SMel Gorman { 75281d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 7534e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 7544e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 7554e35f483SJoonsoo Kim struct inode *inode = mapping->host; 7564e35f483SJoonsoo Kim 7574e35f483SJoonsoo Kim return inode_resv_map(inode); 7584e35f483SJoonsoo Kim 7594e35f483SJoonsoo Kim } else { 76084afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 76184afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 7624e35f483SJoonsoo Kim } 763a1e78772SMel Gorman } 764a1e78772SMel Gorman 76584afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 766a1e78772SMel Gorman { 76781d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 76881d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 769a1e78772SMel Gorman 77084afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 77184afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 77204f2cbe3SMel Gorman } 77304f2cbe3SMel Gorman 77404f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 77504f2cbe3SMel Gorman { 77681d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 77781d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 778e7c4b0bfSAndy Whitcroft 779e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 78004f2cbe3SMel Gorman } 78104f2cbe3SMel Gorman 78204f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 78304f2cbe3SMel Gorman { 78481d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 785e7c4b0bfSAndy Whitcroft 786e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 787a1e78772SMel Gorman } 788a1e78772SMel Gorman 78904f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 790a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 791a1e78772SMel Gorman { 79281d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 793f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 794a1e78772SMel Gorman vma->vm_private_data = (void *)0; 795a1e78772SMel Gorman } 796a1e78772SMel Gorman 797a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 798559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 799a1e78772SMel Gorman { 800af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 801af0ed73eSJoonsoo Kim /* 802af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 803af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 804af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 805af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 806af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 807af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 808af0ed73eSJoonsoo Kim * properly, so add work-around here. 809af0ed73eSJoonsoo Kim */ 810af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 811559ec2f8SNicholas Krause return true; 812af0ed73eSJoonsoo Kim else 813559ec2f8SNicholas Krause return false; 814af0ed73eSJoonsoo Kim } 815a63884e9SJoonsoo Kim 816a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 8171fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 8181fb1b0e9SMike Kravetz /* 8191fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 8201fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 8211fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 8221fb1b0e9SMike Kravetz * fallocate. In this case, there really are no reverves to 8231fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 8241fb1b0e9SMike Kravetz */ 8251fb1b0e9SMike Kravetz if (chg) 8261fb1b0e9SMike Kravetz return false; 8271fb1b0e9SMike Kravetz else 828559ec2f8SNicholas Krause return true; 8291fb1b0e9SMike Kravetz } 830a63884e9SJoonsoo Kim 831a63884e9SJoonsoo Kim /* 832a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 833a63884e9SJoonsoo Kim * private mappings. 834a63884e9SJoonsoo Kim */ 83567961f9dSMike Kravetz if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 83667961f9dSMike Kravetz /* 83767961f9dSMike Kravetz * Like the shared case above, a hole punch or truncate 83867961f9dSMike Kravetz * could have been performed on the private mapping. 83967961f9dSMike Kravetz * Examine the value of chg to determine if reserves 84067961f9dSMike Kravetz * actually exist or were previously consumed. 84167961f9dSMike Kravetz * Very Subtle - The value of chg comes from a previous 84267961f9dSMike Kravetz * call to vma_needs_reserves(). The reserve map for 84367961f9dSMike Kravetz * private mappings has different (opposite) semantics 84467961f9dSMike Kravetz * than that of shared mappings. vma_needs_reserves() 84567961f9dSMike Kravetz * has already taken this difference in semantics into 84667961f9dSMike Kravetz * account. Therefore, the meaning of chg is the same 84767961f9dSMike Kravetz * as in the shared case above. Code could easily be 84867961f9dSMike Kravetz * combined, but keeping it separate draws attention to 84967961f9dSMike Kravetz * subtle differences. 85067961f9dSMike Kravetz */ 85167961f9dSMike Kravetz if (chg) 85267961f9dSMike Kravetz return false; 85367961f9dSMike Kravetz else 854559ec2f8SNicholas Krause return true; 85567961f9dSMike Kravetz } 856a63884e9SJoonsoo Kim 857559ec2f8SNicholas Krause return false; 858a1e78772SMel Gorman } 859a1e78772SMel Gorman 860a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 8611da177e4SLinus Torvalds { 8621da177e4SLinus Torvalds int nid = page_to_nid(page); 8630edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 864a5516438SAndi Kleen h->free_huge_pages++; 865a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 8661da177e4SLinus Torvalds } 8671da177e4SLinus Torvalds 868bf50bab2SNaoya Horiguchi static struct page *dequeue_huge_page_node(struct hstate *h, int nid) 869bf50bab2SNaoya Horiguchi { 870bf50bab2SNaoya Horiguchi struct page *page; 871bf50bab2SNaoya Horiguchi 872c8721bbbSNaoya Horiguchi list_for_each_entry(page, &h->hugepage_freelists[nid], lru) 873c8721bbbSNaoya Horiguchi if (!is_migrate_isolate_page(page)) 874c8721bbbSNaoya Horiguchi break; 875c8721bbbSNaoya Horiguchi /* 876c8721bbbSNaoya Horiguchi * if 'non-isolated free hugepage' not found on the list, 877c8721bbbSNaoya Horiguchi * the allocation fails. 878c8721bbbSNaoya Horiguchi */ 879c8721bbbSNaoya Horiguchi if (&h->hugepage_freelists[nid] == &page->lru) 880bf50bab2SNaoya Horiguchi return NULL; 8810edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 882a9869b83SNaoya Horiguchi set_page_refcounted(page); 883bf50bab2SNaoya Horiguchi h->free_huge_pages--; 884bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 885bf50bab2SNaoya Horiguchi return page; 886bf50bab2SNaoya Horiguchi } 887bf50bab2SNaoya Horiguchi 88886cdb465SNaoya Horiguchi /* Movability of hugepages depends on migration support. */ 88986cdb465SNaoya Horiguchi static inline gfp_t htlb_alloc_mask(struct hstate *h) 89086cdb465SNaoya Horiguchi { 891100873d7SNaoya Horiguchi if (hugepages_treat_as_movable || hugepage_migration_supported(h)) 89286cdb465SNaoya Horiguchi return GFP_HIGHUSER_MOVABLE; 89386cdb465SNaoya Horiguchi else 89486cdb465SNaoya Horiguchi return GFP_HIGHUSER; 89586cdb465SNaoya Horiguchi } 89686cdb465SNaoya Horiguchi 897a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 898a5516438SAndi Kleen struct vm_area_struct *vma, 899af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 900af0ed73eSJoonsoo Kim long chg) 9011da177e4SLinus Torvalds { 902b1c12cbcSKonstantin Khlebnikov struct page *page = NULL; 903480eccf9SLee Schermerhorn struct mempolicy *mpol; 90419770b32SMel Gorman nodemask_t *nodemask; 905c0ff7453SMiao Xie struct zonelist *zonelist; 906dd1a239fSMel Gorman struct zone *zone; 907dd1a239fSMel Gorman struct zoneref *z; 908cc9a6c87SMel Gorman unsigned int cpuset_mems_cookie; 9091da177e4SLinus Torvalds 910a1e78772SMel Gorman /* 911a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 912a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 913a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 914a1e78772SMel Gorman */ 915af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg) && 916a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 917c0ff7453SMiao Xie goto err; 918a1e78772SMel Gorman 91904f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 920a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 9216eab04a8SJustin P. Mattock goto err; 92204f2cbe3SMel Gorman 9239966c4bbSJoonsoo Kim retry_cpuset: 924d26914d1SMel Gorman cpuset_mems_cookie = read_mems_allowed_begin(); 9259966c4bbSJoonsoo Kim zonelist = huge_zonelist(vma, address, 92686cdb465SNaoya Horiguchi htlb_alloc_mask(h), &mpol, &nodemask); 9279966c4bbSJoonsoo Kim 92819770b32SMel Gorman for_each_zone_zonelist_nodemask(zone, z, zonelist, 92919770b32SMel Gorman MAX_NR_ZONES - 1, nodemask) { 930344736f2SVladimir Davydov if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) { 931bf50bab2SNaoya Horiguchi page = dequeue_huge_page_node(h, zone_to_nid(zone)); 932bf50bab2SNaoya Horiguchi if (page) { 933af0ed73eSJoonsoo Kim if (avoid_reserve) 934af0ed73eSJoonsoo Kim break; 935af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg)) 936af0ed73eSJoonsoo Kim break; 937af0ed73eSJoonsoo Kim 93807443a85SJoonsoo Kim SetPagePrivate(page); 939a63884e9SJoonsoo Kim h->resv_huge_pages--; 9405ab3ee7bSKen Chen break; 9411da177e4SLinus Torvalds } 9423abf7afdSAndrew Morton } 943bf50bab2SNaoya Horiguchi } 944cc9a6c87SMel Gorman 945cc9a6c87SMel Gorman mpol_cond_put(mpol); 946d26914d1SMel Gorman if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) 947cc9a6c87SMel Gorman goto retry_cpuset; 948cc9a6c87SMel Gorman return page; 949cc9a6c87SMel Gorman 950c0ff7453SMiao Xie err: 951cc9a6c87SMel Gorman return NULL; 9521da177e4SLinus Torvalds } 9531da177e4SLinus Torvalds 9541cac6f2cSLuiz Capitulino /* 9551cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 9561cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 9571cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 9581cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 9591cac6f2cSLuiz Capitulino * node for alloc or free. 9601cac6f2cSLuiz Capitulino */ 9611cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 9621cac6f2cSLuiz Capitulino { 9630edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 9641cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 9651cac6f2cSLuiz Capitulino 9661cac6f2cSLuiz Capitulino return nid; 9671cac6f2cSLuiz Capitulino } 9681cac6f2cSLuiz Capitulino 9691cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 9701cac6f2cSLuiz Capitulino { 9711cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 9721cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 9731cac6f2cSLuiz Capitulino return nid; 9741cac6f2cSLuiz Capitulino } 9751cac6f2cSLuiz Capitulino 9761cac6f2cSLuiz Capitulino /* 9771cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 9781cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 9791cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 9801cac6f2cSLuiz Capitulino * mask. 9811cac6f2cSLuiz Capitulino */ 9821cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 9831cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 9841cac6f2cSLuiz Capitulino { 9851cac6f2cSLuiz Capitulino int nid; 9861cac6f2cSLuiz Capitulino 9871cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 9881cac6f2cSLuiz Capitulino 9891cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 9901cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 9911cac6f2cSLuiz Capitulino 9921cac6f2cSLuiz Capitulino return nid; 9931cac6f2cSLuiz Capitulino } 9941cac6f2cSLuiz Capitulino 9951cac6f2cSLuiz Capitulino /* 9961cac6f2cSLuiz Capitulino * helper for free_pool_huge_page() - return the previously saved 9971cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 9981cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 9991cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 10001cac6f2cSLuiz Capitulino */ 10011cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 10021cac6f2cSLuiz Capitulino { 10031cac6f2cSLuiz Capitulino int nid; 10041cac6f2cSLuiz Capitulino 10051cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 10061cac6f2cSLuiz Capitulino 10071cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 10081cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 10091cac6f2cSLuiz Capitulino 10101cac6f2cSLuiz Capitulino return nid; 10111cac6f2cSLuiz Capitulino } 10121cac6f2cSLuiz Capitulino 10131cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 10141cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 10151cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 10161cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 10171cac6f2cSLuiz Capitulino nr_nodes--) 10181cac6f2cSLuiz Capitulino 10191cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 10201cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 10211cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 10221cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 10231cac6f2cSLuiz Capitulino nr_nodes--) 10241cac6f2cSLuiz Capitulino 1025461a7184SYisheng Xie #if defined(CONFIG_ARCH_HAS_GIGANTIC_PAGE) && \ 1026d08de8e2SGerald Schaefer ((defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || \ 1027d08de8e2SGerald Schaefer defined(CONFIG_CMA)) 1028944d9fecSLuiz Capitulino static void destroy_compound_gigantic_page(struct page *page, 1029d00181b9SKirill A. Shutemov unsigned int order) 1030944d9fecSLuiz Capitulino { 1031944d9fecSLuiz Capitulino int i; 1032944d9fecSLuiz Capitulino int nr_pages = 1 << order; 1033944d9fecSLuiz Capitulino struct page *p = page + 1; 1034944d9fecSLuiz Capitulino 1035c8cc708aSGerald Schaefer atomic_set(compound_mapcount_ptr(page), 0); 1036944d9fecSLuiz Capitulino for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 10371d798ca3SKirill A. Shutemov clear_compound_head(p); 1038944d9fecSLuiz Capitulino set_page_refcounted(p); 1039944d9fecSLuiz Capitulino } 1040944d9fecSLuiz Capitulino 1041944d9fecSLuiz Capitulino set_compound_order(page, 0); 1042944d9fecSLuiz Capitulino __ClearPageHead(page); 1043944d9fecSLuiz Capitulino } 1044944d9fecSLuiz Capitulino 1045d00181b9SKirill A. Shutemov static void free_gigantic_page(struct page *page, unsigned int order) 1046944d9fecSLuiz Capitulino { 1047944d9fecSLuiz Capitulino free_contig_range(page_to_pfn(page), 1 << order); 1048944d9fecSLuiz Capitulino } 1049944d9fecSLuiz Capitulino 1050944d9fecSLuiz Capitulino static int __alloc_gigantic_page(unsigned long start_pfn, 1051944d9fecSLuiz Capitulino unsigned long nr_pages) 1052944d9fecSLuiz Capitulino { 1053944d9fecSLuiz Capitulino unsigned long end_pfn = start_pfn + nr_pages; 1054944d9fecSLuiz Capitulino return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE); 1055944d9fecSLuiz Capitulino } 1056944d9fecSLuiz Capitulino 1057f44b2ddaSJoonsoo Kim static bool pfn_range_valid_gigantic(struct zone *z, 1058f44b2ddaSJoonsoo Kim unsigned long start_pfn, unsigned long nr_pages) 1059944d9fecSLuiz Capitulino { 1060944d9fecSLuiz Capitulino unsigned long i, end_pfn = start_pfn + nr_pages; 1061944d9fecSLuiz Capitulino struct page *page; 1062944d9fecSLuiz Capitulino 1063944d9fecSLuiz Capitulino for (i = start_pfn; i < end_pfn; i++) { 1064944d9fecSLuiz Capitulino if (!pfn_valid(i)) 1065944d9fecSLuiz Capitulino return false; 1066944d9fecSLuiz Capitulino 1067944d9fecSLuiz Capitulino page = pfn_to_page(i); 1068944d9fecSLuiz Capitulino 1069f44b2ddaSJoonsoo Kim if (page_zone(page) != z) 1070f44b2ddaSJoonsoo Kim return false; 1071f44b2ddaSJoonsoo Kim 1072944d9fecSLuiz Capitulino if (PageReserved(page)) 1073944d9fecSLuiz Capitulino return false; 1074944d9fecSLuiz Capitulino 1075944d9fecSLuiz Capitulino if (page_count(page) > 0) 1076944d9fecSLuiz Capitulino return false; 1077944d9fecSLuiz Capitulino 1078944d9fecSLuiz Capitulino if (PageHuge(page)) 1079944d9fecSLuiz Capitulino return false; 1080944d9fecSLuiz Capitulino } 1081944d9fecSLuiz Capitulino 1082944d9fecSLuiz Capitulino return true; 1083944d9fecSLuiz Capitulino } 1084944d9fecSLuiz Capitulino 1085944d9fecSLuiz Capitulino static bool zone_spans_last_pfn(const struct zone *zone, 1086944d9fecSLuiz Capitulino unsigned long start_pfn, unsigned long nr_pages) 1087944d9fecSLuiz Capitulino { 1088944d9fecSLuiz Capitulino unsigned long last_pfn = start_pfn + nr_pages - 1; 1089944d9fecSLuiz Capitulino return zone_spans_pfn(zone, last_pfn); 1090944d9fecSLuiz Capitulino } 1091944d9fecSLuiz Capitulino 1092d00181b9SKirill A. Shutemov static struct page *alloc_gigantic_page(int nid, unsigned int order) 1093944d9fecSLuiz Capitulino { 1094944d9fecSLuiz Capitulino unsigned long nr_pages = 1 << order; 1095944d9fecSLuiz Capitulino unsigned long ret, pfn, flags; 1096944d9fecSLuiz Capitulino struct zone *z; 1097944d9fecSLuiz Capitulino 1098944d9fecSLuiz Capitulino z = NODE_DATA(nid)->node_zones; 1099944d9fecSLuiz Capitulino for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) { 1100944d9fecSLuiz Capitulino spin_lock_irqsave(&z->lock, flags); 1101944d9fecSLuiz Capitulino 1102944d9fecSLuiz Capitulino pfn = ALIGN(z->zone_start_pfn, nr_pages); 1103944d9fecSLuiz Capitulino while (zone_spans_last_pfn(z, pfn, nr_pages)) { 1104f44b2ddaSJoonsoo Kim if (pfn_range_valid_gigantic(z, pfn, nr_pages)) { 1105944d9fecSLuiz Capitulino /* 1106944d9fecSLuiz Capitulino * We release the zone lock here because 1107944d9fecSLuiz Capitulino * alloc_contig_range() will also lock the zone 1108944d9fecSLuiz Capitulino * at some point. If there's an allocation 1109944d9fecSLuiz Capitulino * spinning on this lock, it may win the race 1110944d9fecSLuiz Capitulino * and cause alloc_contig_range() to fail... 1111944d9fecSLuiz Capitulino */ 1112944d9fecSLuiz Capitulino spin_unlock_irqrestore(&z->lock, flags); 1113944d9fecSLuiz Capitulino ret = __alloc_gigantic_page(pfn, nr_pages); 1114944d9fecSLuiz Capitulino if (!ret) 1115944d9fecSLuiz Capitulino return pfn_to_page(pfn); 1116944d9fecSLuiz Capitulino spin_lock_irqsave(&z->lock, flags); 1117944d9fecSLuiz Capitulino } 1118944d9fecSLuiz Capitulino pfn += nr_pages; 1119944d9fecSLuiz Capitulino } 1120944d9fecSLuiz Capitulino 1121944d9fecSLuiz Capitulino spin_unlock_irqrestore(&z->lock, flags); 1122944d9fecSLuiz Capitulino } 1123944d9fecSLuiz Capitulino 1124944d9fecSLuiz Capitulino return NULL; 1125944d9fecSLuiz Capitulino } 1126944d9fecSLuiz Capitulino 1127944d9fecSLuiz Capitulino static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); 1128d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order); 1129944d9fecSLuiz Capitulino 1130944d9fecSLuiz Capitulino static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid) 1131944d9fecSLuiz Capitulino { 1132944d9fecSLuiz Capitulino struct page *page; 1133944d9fecSLuiz Capitulino 1134944d9fecSLuiz Capitulino page = alloc_gigantic_page(nid, huge_page_order(h)); 1135944d9fecSLuiz Capitulino if (page) { 1136944d9fecSLuiz Capitulino prep_compound_gigantic_page(page, huge_page_order(h)); 1137944d9fecSLuiz Capitulino prep_new_huge_page(h, page, nid); 1138944d9fecSLuiz Capitulino } 1139944d9fecSLuiz Capitulino 1140944d9fecSLuiz Capitulino return page; 1141944d9fecSLuiz Capitulino } 1142944d9fecSLuiz Capitulino 1143944d9fecSLuiz Capitulino static int alloc_fresh_gigantic_page(struct hstate *h, 1144944d9fecSLuiz Capitulino nodemask_t *nodes_allowed) 1145944d9fecSLuiz Capitulino { 1146944d9fecSLuiz Capitulino struct page *page = NULL; 1147944d9fecSLuiz Capitulino int nr_nodes, node; 1148944d9fecSLuiz Capitulino 1149944d9fecSLuiz Capitulino for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1150944d9fecSLuiz Capitulino page = alloc_fresh_gigantic_page_node(h, node); 1151944d9fecSLuiz Capitulino if (page) 1152944d9fecSLuiz Capitulino return 1; 1153944d9fecSLuiz Capitulino } 1154944d9fecSLuiz Capitulino 1155944d9fecSLuiz Capitulino return 0; 1156944d9fecSLuiz Capitulino } 1157944d9fecSLuiz Capitulino 1158944d9fecSLuiz Capitulino static inline bool gigantic_page_supported(void) { return true; } 1159944d9fecSLuiz Capitulino #else 1160944d9fecSLuiz Capitulino static inline bool gigantic_page_supported(void) { return false; } 1161d00181b9SKirill A. Shutemov static inline void free_gigantic_page(struct page *page, unsigned int order) { } 1162944d9fecSLuiz Capitulino static inline void destroy_compound_gigantic_page(struct page *page, 1163d00181b9SKirill A. Shutemov unsigned int order) { } 1164944d9fecSLuiz Capitulino static inline int alloc_fresh_gigantic_page(struct hstate *h, 1165944d9fecSLuiz Capitulino nodemask_t *nodes_allowed) { return 0; } 1166944d9fecSLuiz Capitulino #endif 1167944d9fecSLuiz Capitulino 1168a5516438SAndi Kleen static void update_and_free_page(struct hstate *h, struct page *page) 11696af2acb6SAdam Litke { 11706af2acb6SAdam Litke int i; 1171a5516438SAndi Kleen 1172944d9fecSLuiz Capitulino if (hstate_is_gigantic(h) && !gigantic_page_supported()) 1173944d9fecSLuiz Capitulino return; 117418229df5SAndy Whitcroft 1175a5516438SAndi Kleen h->nr_huge_pages--; 1176a5516438SAndi Kleen h->nr_huge_pages_node[page_to_nid(page)]--; 1177a5516438SAndi Kleen for (i = 0; i < pages_per_huge_page(h); i++) { 117832f84528SChris Forbes page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 117932f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1180a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1181a7407a27SLuiz Capitulino 1 << PG_writeback); 11826af2acb6SAdam Litke } 1183309381feSSasha Levin VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 1184f1e61557SKirill A. Shutemov set_compound_page_dtor(page, NULL_COMPOUND_DTOR); 11856af2acb6SAdam Litke set_page_refcounted(page); 1186944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) { 1187944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 1188944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 1189944d9fecSLuiz Capitulino } else { 1190a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 11916af2acb6SAdam Litke } 1192944d9fecSLuiz Capitulino } 11936af2acb6SAdam Litke 1194e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1195e5ff2159SAndi Kleen { 1196e5ff2159SAndi Kleen struct hstate *h; 1197e5ff2159SAndi Kleen 1198e5ff2159SAndi Kleen for_each_hstate(h) { 1199e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1200e5ff2159SAndi Kleen return h; 1201e5ff2159SAndi Kleen } 1202e5ff2159SAndi Kleen return NULL; 1203e5ff2159SAndi Kleen } 1204e5ff2159SAndi Kleen 1205bcc54222SNaoya Horiguchi /* 1206bcc54222SNaoya Horiguchi * Test to determine whether the hugepage is "active/in-use" (i.e. being linked 1207bcc54222SNaoya Horiguchi * to hstate->hugepage_activelist.) 1208bcc54222SNaoya Horiguchi * 1209bcc54222SNaoya Horiguchi * This function can be called for tail pages, but never returns true for them. 1210bcc54222SNaoya Horiguchi */ 1211bcc54222SNaoya Horiguchi bool page_huge_active(struct page *page) 1212bcc54222SNaoya Horiguchi { 1213bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHuge(page), page); 1214bcc54222SNaoya Horiguchi return PageHead(page) && PagePrivate(&page[1]); 1215bcc54222SNaoya Horiguchi } 1216bcc54222SNaoya Horiguchi 1217bcc54222SNaoya Horiguchi /* never called for tail page */ 1218bcc54222SNaoya Horiguchi static void set_page_huge_active(struct page *page) 1219bcc54222SNaoya Horiguchi { 1220bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 1221bcc54222SNaoya Horiguchi SetPagePrivate(&page[1]); 1222bcc54222SNaoya Horiguchi } 1223bcc54222SNaoya Horiguchi 1224bcc54222SNaoya Horiguchi static void clear_page_huge_active(struct page *page) 1225bcc54222SNaoya Horiguchi { 1226bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 1227bcc54222SNaoya Horiguchi ClearPagePrivate(&page[1]); 1228bcc54222SNaoya Horiguchi } 1229bcc54222SNaoya Horiguchi 12308f1d26d0SAtsushi Kumagai void free_huge_page(struct page *page) 123127a85ef1SDavid Gibson { 1232a5516438SAndi Kleen /* 1233a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1234a5516438SAndi Kleen * compound page destructor. 1235a5516438SAndi Kleen */ 1236e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 12377893d1d5SAdam Litke int nid = page_to_nid(page); 123890481622SDavid Gibson struct hugepage_subpool *spool = 123990481622SDavid Gibson (struct hugepage_subpool *)page_private(page); 124007443a85SJoonsoo Kim bool restore_reserve; 124127a85ef1SDavid Gibson 1242e5df70abSAndy Whitcroft set_page_private(page, 0); 124323be7468SMel Gorman page->mapping = NULL; 1244b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1245b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_mapcount(page), page); 124607443a85SJoonsoo Kim restore_reserve = PagePrivate(page); 124716c794b4SJoonsoo Kim ClearPagePrivate(page); 124827a85ef1SDavid Gibson 12491c5ecae3SMike Kravetz /* 12501c5ecae3SMike Kravetz * A return code of zero implies that the subpool will be under its 12511c5ecae3SMike Kravetz * minimum size if the reservation is not restored after page is free. 12521c5ecae3SMike Kravetz * Therefore, force restore_reserve operation. 12531c5ecae3SMike Kravetz */ 12541c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 12551c5ecae3SMike Kravetz restore_reserve = true; 12561c5ecae3SMike Kravetz 125727a85ef1SDavid Gibson spin_lock(&hugetlb_lock); 1258bcc54222SNaoya Horiguchi clear_page_huge_active(page); 12596d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 12606d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 126107443a85SJoonsoo Kim if (restore_reserve) 126207443a85SJoonsoo Kim h->resv_huge_pages++; 126307443a85SJoonsoo Kim 1264944d9fecSLuiz Capitulino if (h->surplus_huge_pages_node[nid]) { 12650edaecfaSAneesh Kumar K.V /* remove the page from active list */ 12660edaecfaSAneesh Kumar K.V list_del(&page->lru); 1267a5516438SAndi Kleen update_and_free_page(h, page); 1268a5516438SAndi Kleen h->surplus_huge_pages--; 1269a5516438SAndi Kleen h->surplus_huge_pages_node[nid]--; 12707893d1d5SAdam Litke } else { 12715d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1272a5516438SAndi Kleen enqueue_huge_page(h, page); 12737893d1d5SAdam Litke } 127427a85ef1SDavid Gibson spin_unlock(&hugetlb_lock); 127527a85ef1SDavid Gibson } 127627a85ef1SDavid Gibson 1277a5516438SAndi Kleen static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1278b7ba30c6SAndi Kleen { 12790edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1280f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1281b7ba30c6SAndi Kleen spin_lock(&hugetlb_lock); 12829dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 1283a5516438SAndi Kleen h->nr_huge_pages++; 1284a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 1285b7ba30c6SAndi Kleen spin_unlock(&hugetlb_lock); 1286b7ba30c6SAndi Kleen put_page(page); /* free it into the hugepage allocator */ 1287b7ba30c6SAndi Kleen } 1288b7ba30c6SAndi Kleen 1289d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order) 129020a0307cSWu Fengguang { 129120a0307cSWu Fengguang int i; 129220a0307cSWu Fengguang int nr_pages = 1 << order; 129320a0307cSWu Fengguang struct page *p = page + 1; 129420a0307cSWu Fengguang 129520a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 129620a0307cSWu Fengguang set_compound_order(page, order); 1297ef5a22beSAndrea Arcangeli __ClearPageReserved(page); 1298de09d31dSKirill A. Shutemov __SetPageHead(page); 129920a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1300ef5a22beSAndrea Arcangeli /* 1301ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1302ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1303ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 1304ef5a22beSAndrea Arcangeli * too. Otherwse drivers using get_user_pages() to access tail 1305ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1306ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1307ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1308ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1309ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1310ef5a22beSAndrea Arcangeli * after get_user_pages(). 1311ef5a22beSAndrea Arcangeli */ 1312ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 131358a84aa9SYouquan Song set_page_count(p, 0); 13141d798ca3SKirill A. Shutemov set_compound_head(p, page); 131520a0307cSWu Fengguang } 1316b4330afbSMike Kravetz atomic_set(compound_mapcount_ptr(page), -1); 131720a0307cSWu Fengguang } 131820a0307cSWu Fengguang 13197795912cSAndrew Morton /* 13207795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 13217795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 13227795912cSAndrew Morton * details. 13237795912cSAndrew Morton */ 132420a0307cSWu Fengguang int PageHuge(struct page *page) 132520a0307cSWu Fengguang { 132620a0307cSWu Fengguang if (!PageCompound(page)) 132720a0307cSWu Fengguang return 0; 132820a0307cSWu Fengguang 132920a0307cSWu Fengguang page = compound_head(page); 1330f1e61557SKirill A. Shutemov return page[1].compound_dtor == HUGETLB_PAGE_DTOR; 133120a0307cSWu Fengguang } 133243131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 133343131e14SNaoya Horiguchi 133427c73ae7SAndrea Arcangeli /* 133527c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 133627c73ae7SAndrea Arcangeli * normal or transparent huge pages. 133727c73ae7SAndrea Arcangeli */ 133827c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 133927c73ae7SAndrea Arcangeli { 134027c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 134127c73ae7SAndrea Arcangeli return 0; 134227c73ae7SAndrea Arcangeli 1343758f66a2SAndrew Morton return get_compound_page_dtor(page_head) == free_huge_page; 134427c73ae7SAndrea Arcangeli } 134527c73ae7SAndrea Arcangeli 134613d60f4bSZhang Yi pgoff_t __basepage_index(struct page *page) 134713d60f4bSZhang Yi { 134813d60f4bSZhang Yi struct page *page_head = compound_head(page); 134913d60f4bSZhang Yi pgoff_t index = page_index(page_head); 135013d60f4bSZhang Yi unsigned long compound_idx; 135113d60f4bSZhang Yi 135213d60f4bSZhang Yi if (!PageHuge(page_head)) 135313d60f4bSZhang Yi return page_index(page); 135413d60f4bSZhang Yi 135513d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 135613d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 135713d60f4bSZhang Yi else 135813d60f4bSZhang Yi compound_idx = page - page_head; 135913d60f4bSZhang Yi 136013d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 136113d60f4bSZhang Yi } 136213d60f4bSZhang Yi 1363a5516438SAndi Kleen static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) 13641da177e4SLinus Torvalds { 13651da177e4SLinus Torvalds struct page *page; 1366f96efd58SJoe Jin 136796db800fSVlastimil Babka page = __alloc_pages_node(nid, 136886cdb465SNaoya Horiguchi htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| 1369551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 1370a5516438SAndi Kleen huge_page_order(h)); 13711da177e4SLinus Torvalds if (page) { 1372a5516438SAndi Kleen prep_new_huge_page(h, page, nid); 13731da177e4SLinus Torvalds } 137463b4613cSNishanth Aravamudan 137563b4613cSNishanth Aravamudan return page; 137663b4613cSNishanth Aravamudan } 137763b4613cSNishanth Aravamudan 1378b2261026SJoonsoo Kim static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 1379b2261026SJoonsoo Kim { 1380b2261026SJoonsoo Kim struct page *page; 1381b2261026SJoonsoo Kim int nr_nodes, node; 1382b2261026SJoonsoo Kim int ret = 0; 1383b2261026SJoonsoo Kim 1384b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1385b2261026SJoonsoo Kim page = alloc_fresh_huge_page_node(h, node); 1386b2261026SJoonsoo Kim if (page) { 1387b2261026SJoonsoo Kim ret = 1; 1388b2261026SJoonsoo Kim break; 1389b2261026SJoonsoo Kim } 1390b2261026SJoonsoo Kim } 1391b2261026SJoonsoo Kim 1392b2261026SJoonsoo Kim if (ret) 1393b2261026SJoonsoo Kim count_vm_event(HTLB_BUDDY_PGALLOC); 1394b2261026SJoonsoo Kim else 1395b2261026SJoonsoo Kim count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 1396b2261026SJoonsoo Kim 1397b2261026SJoonsoo Kim return ret; 1398b2261026SJoonsoo Kim } 1399b2261026SJoonsoo Kim 1400e8c5c824SLee Schermerhorn /* 1401e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 1402e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 1403e8c5c824SLee Schermerhorn * balanced over allowed nodes. 1404e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 1405e8c5c824SLee Schermerhorn */ 14066ae11b27SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 14076ae11b27SLee Schermerhorn bool acct_surplus) 1408e8c5c824SLee Schermerhorn { 1409b2261026SJoonsoo Kim int nr_nodes, node; 1410e8c5c824SLee Schermerhorn int ret = 0; 1411e8c5c824SLee Schermerhorn 1412b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1413685f3457SLee Schermerhorn /* 1414685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 1415685f3457SLee Schermerhorn * nodes with surplus pages. 1416685f3457SLee Schermerhorn */ 1417b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 1418b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 1419e8c5c824SLee Schermerhorn struct page *page = 1420b2261026SJoonsoo Kim list_entry(h->hugepage_freelists[node].next, 1421e8c5c824SLee Schermerhorn struct page, lru); 1422e8c5c824SLee Schermerhorn list_del(&page->lru); 1423e8c5c824SLee Schermerhorn h->free_huge_pages--; 1424b2261026SJoonsoo Kim h->free_huge_pages_node[node]--; 1425685f3457SLee Schermerhorn if (acct_surplus) { 1426685f3457SLee Schermerhorn h->surplus_huge_pages--; 1427b2261026SJoonsoo Kim h->surplus_huge_pages_node[node]--; 1428685f3457SLee Schermerhorn } 1429e8c5c824SLee Schermerhorn update_and_free_page(h, page); 1430e8c5c824SLee Schermerhorn ret = 1; 14319a76db09SLee Schermerhorn break; 1432e8c5c824SLee Schermerhorn } 1433b2261026SJoonsoo Kim } 1434e8c5c824SLee Schermerhorn 1435e8c5c824SLee Schermerhorn return ret; 1436e8c5c824SLee Schermerhorn } 1437e8c5c824SLee Schermerhorn 1438c8721bbbSNaoya Horiguchi /* 1439c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 1440082d5b6bSGerald Schaefer * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the 1441082d5b6bSGerald Schaefer * number of free hugepages would be reduced below the number of reserved 1442082d5b6bSGerald Schaefer * hugepages. 1443c8721bbbSNaoya Horiguchi */ 1444082d5b6bSGerald Schaefer static int dissolve_free_huge_page(struct page *page) 1445c8721bbbSNaoya Horiguchi { 1446082d5b6bSGerald Schaefer int rc = 0; 1447082d5b6bSGerald Schaefer 1448c8721bbbSNaoya Horiguchi spin_lock(&hugetlb_lock); 1449c8721bbbSNaoya Horiguchi if (PageHuge(page) && !page_count(page)) { 14502247bb33SGerald Schaefer struct page *head = compound_head(page); 14512247bb33SGerald Schaefer struct hstate *h = page_hstate(head); 14522247bb33SGerald Schaefer int nid = page_to_nid(head); 1453082d5b6bSGerald Schaefer if (h->free_huge_pages - h->resv_huge_pages == 0) { 1454082d5b6bSGerald Schaefer rc = -EBUSY; 1455082d5b6bSGerald Schaefer goto out; 1456082d5b6bSGerald Schaefer } 14572247bb33SGerald Schaefer list_del(&head->lru); 1458c8721bbbSNaoya Horiguchi h->free_huge_pages--; 1459c8721bbbSNaoya Horiguchi h->free_huge_pages_node[nid]--; 1460c1470b33Szhong jiang h->max_huge_pages--; 14612247bb33SGerald Schaefer update_and_free_page(h, head); 1462c8721bbbSNaoya Horiguchi } 1463082d5b6bSGerald Schaefer out: 1464c8721bbbSNaoya Horiguchi spin_unlock(&hugetlb_lock); 1465082d5b6bSGerald Schaefer return rc; 1466c8721bbbSNaoya Horiguchi } 1467c8721bbbSNaoya Horiguchi 1468c8721bbbSNaoya Horiguchi /* 1469c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 1470c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 14712247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 14722247bb33SGerald Schaefer * part of it lies within the given range. 1473082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 1474082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 1475c8721bbbSNaoya Horiguchi */ 1476082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 1477c8721bbbSNaoya Horiguchi { 1478c8721bbbSNaoya Horiguchi unsigned long pfn; 1479eb03aa00SGerald Schaefer struct page *page; 1480082d5b6bSGerald Schaefer int rc = 0; 1481c8721bbbSNaoya Horiguchi 1482d0177639SLi Zhong if (!hugepages_supported()) 1483082d5b6bSGerald Schaefer return rc; 1484d0177639SLi Zhong 1485eb03aa00SGerald Schaefer for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { 1486eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 1487eb03aa00SGerald Schaefer if (PageHuge(page) && !page_count(page)) { 1488eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 1489eb03aa00SGerald Schaefer if (rc) 1490082d5b6bSGerald Schaefer break; 1491eb03aa00SGerald Schaefer } 1492eb03aa00SGerald Schaefer } 1493082d5b6bSGerald Schaefer 1494082d5b6bSGerald Schaefer return rc; 1495c8721bbbSNaoya Horiguchi } 1496c8721bbbSNaoya Horiguchi 1497099730d6SDave Hansen /* 1498099730d6SDave Hansen * There are 3 ways this can get called: 1499099730d6SDave Hansen * 1. With vma+addr: we use the VMA's memory policy 1500099730d6SDave Hansen * 2. With !vma, but nid=NUMA_NO_NODE: We try to allocate a huge 1501099730d6SDave Hansen * page from any node, and let the buddy allocator itself figure 1502099730d6SDave Hansen * it out. 1503099730d6SDave Hansen * 3. With !vma, but nid!=NUMA_NO_NODE. We allocate a huge page 1504099730d6SDave Hansen * strictly from 'nid' 1505099730d6SDave Hansen */ 1506099730d6SDave Hansen static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h, 1507099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr, int nid) 1508099730d6SDave Hansen { 1509099730d6SDave Hansen int order = huge_page_order(h); 1510099730d6SDave Hansen gfp_t gfp = htlb_alloc_mask(h)|__GFP_COMP|__GFP_REPEAT|__GFP_NOWARN; 1511099730d6SDave Hansen unsigned int cpuset_mems_cookie; 1512099730d6SDave Hansen 1513099730d6SDave Hansen /* 1514099730d6SDave Hansen * We need a VMA to get a memory policy. If we do not 1515e0ec90eeSDave Hansen * have one, we use the 'nid' argument. 1516e0ec90eeSDave Hansen * 1517e0ec90eeSDave Hansen * The mempolicy stuff below has some non-inlined bits 1518e0ec90eeSDave Hansen * and calls ->vm_ops. That makes it hard to optimize at 1519e0ec90eeSDave Hansen * compile-time, even when NUMA is off and it does 1520e0ec90eeSDave Hansen * nothing. This helps the compiler optimize it out. 1521099730d6SDave Hansen */ 1522e0ec90eeSDave Hansen if (!IS_ENABLED(CONFIG_NUMA) || !vma) { 1523099730d6SDave Hansen /* 1524099730d6SDave Hansen * If a specific node is requested, make sure to 1525099730d6SDave Hansen * get memory from there, but only when a node 1526099730d6SDave Hansen * is explicitly specified. 1527099730d6SDave Hansen */ 1528099730d6SDave Hansen if (nid != NUMA_NO_NODE) 1529099730d6SDave Hansen gfp |= __GFP_THISNODE; 1530099730d6SDave Hansen /* 1531099730d6SDave Hansen * Make sure to call something that can handle 1532099730d6SDave Hansen * nid=NUMA_NO_NODE 1533099730d6SDave Hansen */ 1534099730d6SDave Hansen return alloc_pages_node(nid, gfp, order); 1535099730d6SDave Hansen } 1536099730d6SDave Hansen 1537099730d6SDave Hansen /* 1538099730d6SDave Hansen * OK, so we have a VMA. Fetch the mempolicy and try to 1539e0ec90eeSDave Hansen * allocate a huge page with it. We will only reach this 1540e0ec90eeSDave Hansen * when CONFIG_NUMA=y. 1541099730d6SDave Hansen */ 1542099730d6SDave Hansen do { 1543099730d6SDave Hansen struct page *page; 1544099730d6SDave Hansen struct mempolicy *mpol; 1545099730d6SDave Hansen struct zonelist *zl; 1546099730d6SDave Hansen nodemask_t *nodemask; 1547099730d6SDave Hansen 1548099730d6SDave Hansen cpuset_mems_cookie = read_mems_allowed_begin(); 1549099730d6SDave Hansen zl = huge_zonelist(vma, addr, gfp, &mpol, &nodemask); 1550099730d6SDave Hansen mpol_cond_put(mpol); 1551099730d6SDave Hansen page = __alloc_pages_nodemask(gfp, order, zl, nodemask); 1552099730d6SDave Hansen if (page) 1553099730d6SDave Hansen return page; 1554099730d6SDave Hansen } while (read_mems_allowed_retry(cpuset_mems_cookie)); 1555099730d6SDave Hansen 1556099730d6SDave Hansen return NULL; 1557099730d6SDave Hansen } 1558099730d6SDave Hansen 1559099730d6SDave Hansen /* 1560099730d6SDave Hansen * There are two ways to allocate a huge page: 1561099730d6SDave Hansen * 1. When you have a VMA and an address (like a fault) 1562099730d6SDave Hansen * 2. When you have no VMA (like when setting /proc/.../nr_hugepages) 1563099730d6SDave Hansen * 1564099730d6SDave Hansen * 'vma' and 'addr' are only for (1). 'nid' is always NUMA_NO_NODE in 1565099730d6SDave Hansen * this case which signifies that the allocation should be done with 1566099730d6SDave Hansen * respect for the VMA's memory policy. 1567099730d6SDave Hansen * 1568099730d6SDave Hansen * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This 1569099730d6SDave Hansen * implies that memory policies will not be taken in to account. 1570099730d6SDave Hansen */ 1571099730d6SDave Hansen static struct page *__alloc_buddy_huge_page(struct hstate *h, 1572099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr, int nid) 15737893d1d5SAdam Litke { 15747893d1d5SAdam Litke struct page *page; 1575bf50bab2SNaoya Horiguchi unsigned int r_nid; 15767893d1d5SAdam Litke 1577bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1578aa888a74SAndi Kleen return NULL; 1579aa888a74SAndi Kleen 1580d1c3fb1fSNishanth Aravamudan /* 1581099730d6SDave Hansen * Make sure that anyone specifying 'nid' is not also specifying a VMA. 1582099730d6SDave Hansen * This makes sure the caller is picking _one_ of the modes with which 1583099730d6SDave Hansen * we can call this function, not both. 1584099730d6SDave Hansen */ 1585099730d6SDave Hansen if (vma || (addr != -1)) { 1586e0ec90eeSDave Hansen VM_WARN_ON_ONCE(addr == -1); 1587e0ec90eeSDave Hansen VM_WARN_ON_ONCE(nid != NUMA_NO_NODE); 1588099730d6SDave Hansen } 1589099730d6SDave Hansen /* 1590d1c3fb1fSNishanth Aravamudan * Assume we will successfully allocate the surplus page to 1591d1c3fb1fSNishanth Aravamudan * prevent racing processes from causing the surplus to exceed 1592d1c3fb1fSNishanth Aravamudan * overcommit 1593d1c3fb1fSNishanth Aravamudan * 1594d1c3fb1fSNishanth Aravamudan * This however introduces a different race, where a process B 1595d1c3fb1fSNishanth Aravamudan * tries to grow the static hugepage pool while alloc_pages() is 1596d1c3fb1fSNishanth Aravamudan * called by process A. B will only examine the per-node 1597d1c3fb1fSNishanth Aravamudan * counters in determining if surplus huge pages can be 1598d1c3fb1fSNishanth Aravamudan * converted to normal huge pages in adjust_pool_surplus(). A 1599d1c3fb1fSNishanth Aravamudan * won't be able to increment the per-node counter, until the 1600d1c3fb1fSNishanth Aravamudan * lock is dropped by B, but B doesn't drop hugetlb_lock until 1601d1c3fb1fSNishanth Aravamudan * no more huge pages can be converted from surplus to normal 1602d1c3fb1fSNishanth Aravamudan * state (and doesn't try to convert again). Thus, we have a 1603d1c3fb1fSNishanth Aravamudan * case where a surplus huge page exists, the pool is grown, and 1604d1c3fb1fSNishanth Aravamudan * the surplus huge page still exists after, even though it 1605d1c3fb1fSNishanth Aravamudan * should just have been converted to a normal huge page. This 1606d1c3fb1fSNishanth Aravamudan * does not leak memory, though, as the hugepage will be freed 1607d1c3fb1fSNishanth Aravamudan * once it is out of use. It also does not allow the counters to 1608d1c3fb1fSNishanth Aravamudan * go out of whack in adjust_pool_surplus() as we don't modify 1609d1c3fb1fSNishanth Aravamudan * the node values until we've gotten the hugepage and only the 1610d1c3fb1fSNishanth Aravamudan * per-node value is checked there. 1611d1c3fb1fSNishanth Aravamudan */ 1612d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 1613a5516438SAndi Kleen if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 1614d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1615d1c3fb1fSNishanth Aravamudan return NULL; 1616d1c3fb1fSNishanth Aravamudan } else { 1617a5516438SAndi Kleen h->nr_huge_pages++; 1618a5516438SAndi Kleen h->surplus_huge_pages++; 1619d1c3fb1fSNishanth Aravamudan } 1620d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1621d1c3fb1fSNishanth Aravamudan 1622099730d6SDave Hansen page = __hugetlb_alloc_buddy_huge_page(h, vma, addr, nid); 1623d1c3fb1fSNishanth Aravamudan 16247893d1d5SAdam Litke spin_lock(&hugetlb_lock); 1625d1c3fb1fSNishanth Aravamudan if (page) { 16260edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1627bf50bab2SNaoya Horiguchi r_nid = page_to_nid(page); 1628f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 16299dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 1630d1c3fb1fSNishanth Aravamudan /* 1631d1c3fb1fSNishanth Aravamudan * We incremented the global counters already 1632d1c3fb1fSNishanth Aravamudan */ 1633bf50bab2SNaoya Horiguchi h->nr_huge_pages_node[r_nid]++; 1634bf50bab2SNaoya Horiguchi h->surplus_huge_pages_node[r_nid]++; 16353b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC); 1636d1c3fb1fSNishanth Aravamudan } else { 1637a5516438SAndi Kleen h->nr_huge_pages--; 1638a5516438SAndi Kleen h->surplus_huge_pages--; 16393b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 16407893d1d5SAdam Litke } 1641d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 16427893d1d5SAdam Litke 16437893d1d5SAdam Litke return page; 16447893d1d5SAdam Litke } 16457893d1d5SAdam Litke 1646e4e574b7SAdam Litke /* 1647099730d6SDave Hansen * Allocate a huge page from 'nid'. Note, 'nid' may be 1648099730d6SDave Hansen * NUMA_NO_NODE, which means that it may be allocated 1649099730d6SDave Hansen * anywhere. 1650099730d6SDave Hansen */ 1651e0ec90eeSDave Hansen static 1652099730d6SDave Hansen struct page *__alloc_buddy_huge_page_no_mpol(struct hstate *h, int nid) 1653099730d6SDave Hansen { 1654099730d6SDave Hansen unsigned long addr = -1; 1655099730d6SDave Hansen 1656099730d6SDave Hansen return __alloc_buddy_huge_page(h, NULL, addr, nid); 1657099730d6SDave Hansen } 1658099730d6SDave Hansen 1659099730d6SDave Hansen /* 1660099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 1661099730d6SDave Hansen */ 1662e0ec90eeSDave Hansen static 1663099730d6SDave Hansen struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h, 1664099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 1665099730d6SDave Hansen { 1666099730d6SDave Hansen return __alloc_buddy_huge_page(h, vma, addr, NUMA_NO_NODE); 1667099730d6SDave Hansen } 1668099730d6SDave Hansen 1669099730d6SDave Hansen /* 1670bf50bab2SNaoya Horiguchi * This allocation function is useful in the context where vma is irrelevant. 1671bf50bab2SNaoya Horiguchi * E.g. soft-offlining uses this function because it only cares physical 1672bf50bab2SNaoya Horiguchi * address of error page. 1673bf50bab2SNaoya Horiguchi */ 1674bf50bab2SNaoya Horiguchi struct page *alloc_huge_page_node(struct hstate *h, int nid) 1675bf50bab2SNaoya Horiguchi { 16764ef91848SJoonsoo Kim struct page *page = NULL; 1677bf50bab2SNaoya Horiguchi 1678bf50bab2SNaoya Horiguchi spin_lock(&hugetlb_lock); 16794ef91848SJoonsoo Kim if (h->free_huge_pages - h->resv_huge_pages > 0) 1680bf50bab2SNaoya Horiguchi page = dequeue_huge_page_node(h, nid); 1681bf50bab2SNaoya Horiguchi spin_unlock(&hugetlb_lock); 1682bf50bab2SNaoya Horiguchi 168394ae8ba7SAneesh Kumar K.V if (!page) 1684099730d6SDave Hansen page = __alloc_buddy_huge_page_no_mpol(h, nid); 1685bf50bab2SNaoya Horiguchi 1686bf50bab2SNaoya Horiguchi return page; 1687bf50bab2SNaoya Horiguchi } 1688bf50bab2SNaoya Horiguchi 1689bf50bab2SNaoya Horiguchi /* 169025985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 1691e4e574b7SAdam Litke * of size 'delta'. 1692e4e574b7SAdam Litke */ 1693a5516438SAndi Kleen static int gather_surplus_pages(struct hstate *h, int delta) 1694e4e574b7SAdam Litke { 1695e4e574b7SAdam Litke struct list_head surplus_list; 1696e4e574b7SAdam Litke struct page *page, *tmp; 1697e4e574b7SAdam Litke int ret, i; 1698e4e574b7SAdam Litke int needed, allocated; 169928073b02SHillf Danton bool alloc_ok = true; 1700e4e574b7SAdam Litke 1701a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 1702ac09b3a1SAdam Litke if (needed <= 0) { 1703a5516438SAndi Kleen h->resv_huge_pages += delta; 1704e4e574b7SAdam Litke return 0; 1705ac09b3a1SAdam Litke } 1706e4e574b7SAdam Litke 1707e4e574b7SAdam Litke allocated = 0; 1708e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 1709e4e574b7SAdam Litke 1710e4e574b7SAdam Litke ret = -ENOMEM; 1711e4e574b7SAdam Litke retry: 1712e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 1713e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 1714099730d6SDave Hansen page = __alloc_buddy_huge_page_no_mpol(h, NUMA_NO_NODE); 171528073b02SHillf Danton if (!page) { 171628073b02SHillf Danton alloc_ok = false; 171728073b02SHillf Danton break; 171828073b02SHillf Danton } 1719e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 1720e4e574b7SAdam Litke } 172128073b02SHillf Danton allocated += i; 1722e4e574b7SAdam Litke 1723e4e574b7SAdam Litke /* 1724e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 1725e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 1726e4e574b7SAdam Litke */ 1727e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 1728a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 1729a5516438SAndi Kleen (h->free_huge_pages + allocated); 173028073b02SHillf Danton if (needed > 0) { 173128073b02SHillf Danton if (alloc_ok) 1732e4e574b7SAdam Litke goto retry; 173328073b02SHillf Danton /* 173428073b02SHillf Danton * We were not able to allocate enough pages to 173528073b02SHillf Danton * satisfy the entire reservation so we free what 173628073b02SHillf Danton * we've allocated so far. 173728073b02SHillf Danton */ 173828073b02SHillf Danton goto free; 173928073b02SHillf Danton } 1740e4e574b7SAdam Litke /* 1741e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 174225985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 1743e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 1744ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 1745ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 1746ac09b3a1SAdam Litke * before they are reserved. 1747e4e574b7SAdam Litke */ 1748e4e574b7SAdam Litke needed += allocated; 1749a5516438SAndi Kleen h->resv_huge_pages += delta; 1750e4e574b7SAdam Litke ret = 0; 1751a9869b83SNaoya Horiguchi 175219fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 175319fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 175419fc3f0aSAdam Litke if ((--needed) < 0) 175519fc3f0aSAdam Litke break; 1756a9869b83SNaoya Horiguchi /* 1757a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 1758a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 1759a9869b83SNaoya Horiguchi */ 1760a9869b83SNaoya Horiguchi put_page_testzero(page); 1761309381feSSasha Levin VM_BUG_ON_PAGE(page_count(page), page); 1762a5516438SAndi Kleen enqueue_huge_page(h, page); 176319fc3f0aSAdam Litke } 176428073b02SHillf Danton free: 1765b0365c8dSHillf Danton spin_unlock(&hugetlb_lock); 176619fc3f0aSAdam Litke 176719fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 1768c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 1769a9869b83SNaoya Horiguchi put_page(page); 177019fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 1771e4e574b7SAdam Litke 1772e4e574b7SAdam Litke return ret; 1773e4e574b7SAdam Litke } 1774e4e574b7SAdam Litke 1775e4e574b7SAdam Litke /* 1776e5bbc8a6SMike Kravetz * This routine has two main purposes: 1777e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 1778e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 1779e5bbc8a6SMike Kravetz * to the associated reservation map. 1780e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 1781e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 1782e5bbc8a6SMike Kravetz * 1783e5bbc8a6SMike Kravetz * Called with hugetlb_lock held. However, the lock could be dropped (and 1784e5bbc8a6SMike Kravetz * reacquired) during calls to cond_resched_lock. Whenever dropping the lock, 1785e5bbc8a6SMike Kravetz * we must make sure nobody else can claim pages we are in the process of 1786e5bbc8a6SMike Kravetz * freeing. Do this by ensuring resv_huge_page always is greater than the 1787e5bbc8a6SMike Kravetz * number of huge pages we plan to free when dropping the lock. 1788e4e574b7SAdam Litke */ 1789a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 1790a5516438SAndi Kleen unsigned long unused_resv_pages) 1791e4e574b7SAdam Litke { 1792e4e574b7SAdam Litke unsigned long nr_pages; 1793e4e574b7SAdam Litke 1794aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 1795bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1796e5bbc8a6SMike Kravetz goto out; 1797aa888a74SAndi Kleen 1798e5bbc8a6SMike Kravetz /* 1799e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 1800e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 1801e5bbc8a6SMike Kravetz */ 1802a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 1803e4e574b7SAdam Litke 1804685f3457SLee Schermerhorn /* 1805685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 18069b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 18079b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 18089b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 18099b5e5d0fSLee Schermerhorn * free_pool_huge_page() will balance the the freed pages across the 18109b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 1811e5bbc8a6SMike Kravetz * 1812e5bbc8a6SMike Kravetz * Note that we decrement resv_huge_pages as we free the pages. If 1813e5bbc8a6SMike Kravetz * we drop the lock, resv_huge_pages will still be sufficiently large 1814e5bbc8a6SMike Kravetz * to cover subsequent pages we may free. 1815685f3457SLee Schermerhorn */ 1816685f3457SLee Schermerhorn while (nr_pages--) { 1817e5bbc8a6SMike Kravetz h->resv_huge_pages--; 1818e5bbc8a6SMike Kravetz unused_resv_pages--; 18198cebfcd0SLai Jiangshan if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) 1820e5bbc8a6SMike Kravetz goto out; 18217848a4bfSMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 1822e4e574b7SAdam Litke } 1823e5bbc8a6SMike Kravetz 1824e5bbc8a6SMike Kravetz out: 1825e5bbc8a6SMike Kravetz /* Fully uncommit the reservation */ 1826e5bbc8a6SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 1827e4e574b7SAdam Litke } 1828e4e574b7SAdam Litke 18295e911373SMike Kravetz 1830c37f9fb1SAndy Whitcroft /* 1831feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 18325e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 1833cf3ad20bSMike Kravetz * 1834cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 1835cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 1836cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 1837cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 1838cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 1839feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 1840feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 1841feba16e2SMike Kravetz * is called in such cases. 1842cf3ad20bSMike Kravetz * 1843cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 1844cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 1845cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 1846cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 1847cf3ad20bSMike Kravetz * take appropriate action. 184896b96a96SMike Kravetz * 184996b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 185096b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 185196b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 185296b96a96SMike Kravetz * reservation exists. 1853c37f9fb1SAndy Whitcroft */ 18545e911373SMike Kravetz enum vma_resv_mode { 18555e911373SMike Kravetz VMA_NEEDS_RESV, 18565e911373SMike Kravetz VMA_COMMIT_RESV, 1857feba16e2SMike Kravetz VMA_END_RESV, 185896b96a96SMike Kravetz VMA_ADD_RESV, 18595e911373SMike Kravetz }; 1860cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 1861cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 18625e911373SMike Kravetz enum vma_resv_mode mode) 1863c37f9fb1SAndy Whitcroft { 18644e35f483SJoonsoo Kim struct resv_map *resv; 18654e35f483SJoonsoo Kim pgoff_t idx; 1866cf3ad20bSMike Kravetz long ret; 1867c37f9fb1SAndy Whitcroft 18684e35f483SJoonsoo Kim resv = vma_resv_map(vma); 18694e35f483SJoonsoo Kim if (!resv) 1870c37f9fb1SAndy Whitcroft return 1; 1871c37f9fb1SAndy Whitcroft 18724e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 18735e911373SMike Kravetz switch (mode) { 18745e911373SMike Kravetz case VMA_NEEDS_RESV: 1875cf3ad20bSMike Kravetz ret = region_chg(resv, idx, idx + 1); 18765e911373SMike Kravetz break; 18775e911373SMike Kravetz case VMA_COMMIT_RESV: 18785e911373SMike Kravetz ret = region_add(resv, idx, idx + 1); 18795e911373SMike Kravetz break; 1880feba16e2SMike Kravetz case VMA_END_RESV: 18815e911373SMike Kravetz region_abort(resv, idx, idx + 1); 18825e911373SMike Kravetz ret = 0; 18835e911373SMike Kravetz break; 188496b96a96SMike Kravetz case VMA_ADD_RESV: 188596b96a96SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) 188696b96a96SMike Kravetz ret = region_add(resv, idx, idx + 1); 188796b96a96SMike Kravetz else { 188896b96a96SMike Kravetz region_abort(resv, idx, idx + 1); 188996b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 189096b96a96SMike Kravetz } 189196b96a96SMike Kravetz break; 18925e911373SMike Kravetz default: 18935e911373SMike Kravetz BUG(); 18945e911373SMike Kravetz } 189584afd99bSAndy Whitcroft 18964e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) 1897cf3ad20bSMike Kravetz return ret; 189867961f9dSMike Kravetz else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { 189967961f9dSMike Kravetz /* 190067961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 190167961f9dSMike Kravetz * However, a file associated with mapping could have been 190267961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 190367961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 190467961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 190567961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 190667961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 190767961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 190867961f9dSMike Kravetz * reservation has already been consumed. As a result, the 190967961f9dSMike Kravetz * return value of this routine is the opposite of the 191067961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 191167961f9dSMike Kravetz */ 191267961f9dSMike Kravetz if (ret) 191367961f9dSMike Kravetz return 0; 191467961f9dSMike Kravetz else 191567961f9dSMike Kravetz return 1; 191667961f9dSMike Kravetz } 19174e35f483SJoonsoo Kim else 1918cf3ad20bSMike Kravetz return ret < 0 ? ret : 0; 191984afd99bSAndy Whitcroft } 1920cf3ad20bSMike Kravetz 1921cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 1922a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 1923c37f9fb1SAndy Whitcroft { 19245e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 1925cf3ad20bSMike Kravetz } 1926c37f9fb1SAndy Whitcroft 1927cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 1928cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 1929cf3ad20bSMike Kravetz { 19305e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 19315e911373SMike Kravetz } 19325e911373SMike Kravetz 1933feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 19345e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 19355e911373SMike Kravetz { 1936feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 1937c37f9fb1SAndy Whitcroft } 1938c37f9fb1SAndy Whitcroft 193996b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 194096b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 194196b96a96SMike Kravetz { 194296b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 194396b96a96SMike Kravetz } 194496b96a96SMike Kravetz 194596b96a96SMike Kravetz /* 194696b96a96SMike Kravetz * This routine is called to restore a reservation on error paths. In the 194796b96a96SMike Kravetz * specific error paths, a huge page was allocated (via alloc_huge_page) 194896b96a96SMike Kravetz * and is about to be freed. If a reservation for the page existed, 194996b96a96SMike Kravetz * alloc_huge_page would have consumed the reservation and set PagePrivate 195096b96a96SMike Kravetz * in the newly allocated page. When the page is freed via free_huge_page, 195196b96a96SMike Kravetz * the global reservation count will be incremented if PagePrivate is set. 195296b96a96SMike Kravetz * However, free_huge_page can not adjust the reserve map. Adjust the 195396b96a96SMike Kravetz * reserve map here to be consistent with global reserve count adjustments 195496b96a96SMike Kravetz * to be made by free_huge_page. 195596b96a96SMike Kravetz */ 195696b96a96SMike Kravetz static void restore_reserve_on_error(struct hstate *h, 195796b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long address, 195896b96a96SMike Kravetz struct page *page) 195996b96a96SMike Kravetz { 196096b96a96SMike Kravetz if (unlikely(PagePrivate(page))) { 196196b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 196296b96a96SMike Kravetz 196396b96a96SMike Kravetz if (unlikely(rc < 0)) { 196496b96a96SMike Kravetz /* 196596b96a96SMike Kravetz * Rare out of memory condition in reserve map 196696b96a96SMike Kravetz * manipulation. Clear PagePrivate so that 196796b96a96SMike Kravetz * global reserve count will not be incremented 196896b96a96SMike Kravetz * by free_huge_page. This will make it appear 196996b96a96SMike Kravetz * as though the reservation for this page was 197096b96a96SMike Kravetz * consumed. This may prevent the task from 197196b96a96SMike Kravetz * faulting in the page at a later time. This 197296b96a96SMike Kravetz * is better than inconsistent global huge page 197396b96a96SMike Kravetz * accounting of reserve counts. 197496b96a96SMike Kravetz */ 197596b96a96SMike Kravetz ClearPagePrivate(page); 197696b96a96SMike Kravetz } else if (rc) { 197796b96a96SMike Kravetz rc = vma_add_reservation(h, vma, address); 197896b96a96SMike Kravetz if (unlikely(rc < 0)) 197996b96a96SMike Kravetz /* 198096b96a96SMike Kravetz * See above comment about rare out of 198196b96a96SMike Kravetz * memory condition. 198296b96a96SMike Kravetz */ 198396b96a96SMike Kravetz ClearPagePrivate(page); 198496b96a96SMike Kravetz } else 198596b96a96SMike Kravetz vma_end_reservation(h, vma, address); 198696b96a96SMike Kravetz } 198796b96a96SMike Kravetz } 198896b96a96SMike Kravetz 198970c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 199004f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 1991348ea204SAdam Litke { 199290481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 1993a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 1994348ea204SAdam Litke struct page *page; 1995d85f69b0SMike Kravetz long map_chg, map_commit; 1996d85f69b0SMike Kravetz long gbl_chg; 19976d76dcf4SAneesh Kumar K.V int ret, idx; 19986d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 19992fc39cecSAdam Litke 20006d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 2001a1e78772SMel Gorman /* 2002d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 2003d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 2004d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 2005a1e78772SMel Gorman */ 2006d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 2007d85f69b0SMike Kravetz if (map_chg < 0) 200876dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 2009d85f69b0SMike Kravetz 2010d85f69b0SMike Kravetz /* 2011d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 2012d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 2013d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 2014d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 2015d85f69b0SMike Kravetz * checked against any subpool limit. 2016d85f69b0SMike Kravetz */ 2017d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 2018d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 2019d85f69b0SMike Kravetz if (gbl_chg < 0) { 2020feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 202176dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 20225e911373SMike Kravetz } 202390d8b7e6SAdam Litke 2024d85f69b0SMike Kravetz /* 2025d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 2026d85f69b0SMike Kravetz * map, there could be reservations associated with the 2027d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 2028d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 2029d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 2030d85f69b0SMike Kravetz * the subpool reservations. 2031d85f69b0SMike Kravetz */ 2032d85f69b0SMike Kravetz if (avoid_reserve) 2033d85f69b0SMike Kravetz gbl_chg = 1; 2034d85f69b0SMike Kravetz } 2035d85f69b0SMike Kravetz 20366d76dcf4SAneesh Kumar K.V ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 20378f34af6fSJianyu Zhan if (ret) 20388f34af6fSJianyu Zhan goto out_subpool_put; 20398f34af6fSJianyu Zhan 2040a1e78772SMel Gorman spin_lock(&hugetlb_lock); 2041d85f69b0SMike Kravetz /* 2042d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 2043d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 2044d85f69b0SMike Kravetz * a reservation exists for the allocation. 2045d85f69b0SMike Kravetz */ 2046d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 204781a6fcaeSJoonsoo Kim if (!page) { 204894ae8ba7SAneesh Kumar K.V spin_unlock(&hugetlb_lock); 2049099730d6SDave Hansen page = __alloc_buddy_huge_page_with_mpol(h, vma, addr); 20508f34af6fSJianyu Zhan if (!page) 20518f34af6fSJianyu Zhan goto out_uncharge_cgroup; 2052a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 2053a88c7695SNaoya Horiguchi SetPagePrivate(page); 2054a88c7695SNaoya Horiguchi h->resv_huge_pages--; 2055a88c7695SNaoya Horiguchi } 205679dbb236SAneesh Kumar K.V spin_lock(&hugetlb_lock); 205779dbb236SAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 205881a6fcaeSJoonsoo Kim /* Fall through */ 2059a1e78772SMel Gorman } 206081a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 206181a6fcaeSJoonsoo Kim spin_unlock(&hugetlb_lock); 2062a1e78772SMel Gorman 206390481622SDavid Gibson set_page_private(page, (unsigned long)spool); 2064a1e78772SMel Gorman 2065d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 2066d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 206733039678SMike Kravetz /* 206833039678SMike Kravetz * The page was added to the reservation map between 206933039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 207033039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 207133039678SMike Kravetz * Adjust for the subpool count incremented above AND 207233039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 207333039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 207433039678SMike Kravetz * no longer applies. 207533039678SMike Kravetz */ 207633039678SMike Kravetz long rsv_adjust; 207733039678SMike Kravetz 207833039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 207933039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 208033039678SMike Kravetz } 20817893d1d5SAdam Litke return page; 20828f34af6fSJianyu Zhan 20838f34af6fSJianyu Zhan out_uncharge_cgroup: 20848f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 20858f34af6fSJianyu Zhan out_subpool_put: 2086d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 20878f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 2088feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 20898f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 2090b45b5bd6SDavid Gibson } 2091b45b5bd6SDavid Gibson 209274060e4dSNaoya Horiguchi /* 209374060e4dSNaoya Horiguchi * alloc_huge_page()'s wrapper which simply returns the page if allocation 209474060e4dSNaoya Horiguchi * succeeds, otherwise NULL. This function is called from new_vma_page(), 209574060e4dSNaoya Horiguchi * where no ERR_VALUE is expected to be returned. 209674060e4dSNaoya Horiguchi */ 209774060e4dSNaoya Horiguchi struct page *alloc_huge_page_noerr(struct vm_area_struct *vma, 209874060e4dSNaoya Horiguchi unsigned long addr, int avoid_reserve) 209974060e4dSNaoya Horiguchi { 210074060e4dSNaoya Horiguchi struct page *page = alloc_huge_page(vma, addr, avoid_reserve); 210174060e4dSNaoya Horiguchi if (IS_ERR(page)) 210274060e4dSNaoya Horiguchi page = NULL; 210374060e4dSNaoya Horiguchi return page; 210474060e4dSNaoya Horiguchi } 210574060e4dSNaoya Horiguchi 210691f47662SCyrill Gorcunov int __weak alloc_bootmem_huge_page(struct hstate *h) 2107aa888a74SAndi Kleen { 2108aa888a74SAndi Kleen struct huge_bootmem_page *m; 2109b2261026SJoonsoo Kim int nr_nodes, node; 2110aa888a74SAndi Kleen 2111b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 2112aa888a74SAndi Kleen void *addr; 2113aa888a74SAndi Kleen 21148b89a116SGrygorii Strashko addr = memblock_virt_alloc_try_nid_nopanic( 21158b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 21168b89a116SGrygorii Strashko 0, BOOTMEM_ALLOC_ACCESSIBLE, node); 2117aa888a74SAndi Kleen if (addr) { 2118aa888a74SAndi Kleen /* 2119aa888a74SAndi Kleen * Use the beginning of the huge page to store the 2120aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 2121aa888a74SAndi Kleen * puts them into the mem_map). 2122aa888a74SAndi Kleen */ 2123aa888a74SAndi Kleen m = addr; 2124aa888a74SAndi Kleen goto found; 2125aa888a74SAndi Kleen } 2126aa888a74SAndi Kleen } 2127aa888a74SAndi Kleen return 0; 2128aa888a74SAndi Kleen 2129aa888a74SAndi Kleen found: 2130df994eadSLuiz Capitulino BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); 2131aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 2132aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 2133aa888a74SAndi Kleen m->hstate = h; 2134aa888a74SAndi Kleen return 1; 2135aa888a74SAndi Kleen } 2136aa888a74SAndi Kleen 2137d00181b9SKirill A. Shutemov static void __init prep_compound_huge_page(struct page *page, 2138d00181b9SKirill A. Shutemov unsigned int order) 213918229df5SAndy Whitcroft { 214018229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 214118229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 214218229df5SAndy Whitcroft else 214318229df5SAndy Whitcroft prep_compound_page(page, order); 214418229df5SAndy Whitcroft } 214518229df5SAndy Whitcroft 2146aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 2147aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 2148aa888a74SAndi Kleen { 2149aa888a74SAndi Kleen struct huge_bootmem_page *m; 2150aa888a74SAndi Kleen 2151aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 2152aa888a74SAndi Kleen struct hstate *h = m->hstate; 2153ee8f248dSBecky Bruce struct page *page; 2154ee8f248dSBecky Bruce 2155ee8f248dSBecky Bruce #ifdef CONFIG_HIGHMEM 2156ee8f248dSBecky Bruce page = pfn_to_page(m->phys >> PAGE_SHIFT); 21578b89a116SGrygorii Strashko memblock_free_late(__pa(m), 2158ee8f248dSBecky Bruce sizeof(struct huge_bootmem_page)); 2159ee8f248dSBecky Bruce #else 2160ee8f248dSBecky Bruce page = virt_to_page(m); 2161ee8f248dSBecky Bruce #endif 2162aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 216318229df5SAndy Whitcroft prep_compound_huge_page(page, h->order); 2164ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 2165aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 2166b0320c7bSRafael Aquini /* 2167b0320c7bSRafael Aquini * If we had gigantic hugepages allocated at boot time, we need 2168b0320c7bSRafael Aquini * to restore the 'stolen' pages to totalram_pages in order to 2169b0320c7bSRafael Aquini * fix confusing memory reports from free(1) and another 2170b0320c7bSRafael Aquini * side-effects, like CommitLimit going negative. 2171b0320c7bSRafael Aquini */ 2172bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 21733dcc0571SJiang Liu adjust_managed_page_count(page, 1 << h->order); 2174aa888a74SAndi Kleen } 2175aa888a74SAndi Kleen } 2176aa888a74SAndi Kleen 21778faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 21781da177e4SLinus Torvalds { 21791da177e4SLinus Torvalds unsigned long i; 21801da177e4SLinus Torvalds 2181e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 2182bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 2183aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 2184aa888a74SAndi Kleen break; 21859b5e5d0fSLee Schermerhorn } else if (!alloc_fresh_huge_page(h, 21868cebfcd0SLai Jiangshan &node_states[N_MEMORY])) 21871da177e4SLinus Torvalds break; 21881da177e4SLinus Torvalds } 21898faa8b07SAndi Kleen h->max_huge_pages = i; 2190e5ff2159SAndi Kleen } 2191e5ff2159SAndi Kleen 2192e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 2193e5ff2159SAndi Kleen { 2194e5ff2159SAndi Kleen struct hstate *h; 2195e5ff2159SAndi Kleen 2196e5ff2159SAndi Kleen for_each_hstate(h) { 2197641844f5SNaoya Horiguchi if (minimum_order > huge_page_order(h)) 2198641844f5SNaoya Horiguchi minimum_order = huge_page_order(h); 2199641844f5SNaoya Horiguchi 22008faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 2201bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 22028faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 2203e5ff2159SAndi Kleen } 2204641844f5SNaoya Horiguchi VM_BUG_ON(minimum_order == UINT_MAX); 2205e5ff2159SAndi Kleen } 2206e5ff2159SAndi Kleen 22074abd32dbSAndi Kleen static char * __init memfmt(char *buf, unsigned long n) 22084abd32dbSAndi Kleen { 22094abd32dbSAndi Kleen if (n >= (1UL << 30)) 22104abd32dbSAndi Kleen sprintf(buf, "%lu GB", n >> 30); 22114abd32dbSAndi Kleen else if (n >= (1UL << 20)) 22124abd32dbSAndi Kleen sprintf(buf, "%lu MB", n >> 20); 22134abd32dbSAndi Kleen else 22144abd32dbSAndi Kleen sprintf(buf, "%lu KB", n >> 10); 22154abd32dbSAndi Kleen return buf; 22164abd32dbSAndi Kleen } 22174abd32dbSAndi Kleen 2218e5ff2159SAndi Kleen static void __init report_hugepages(void) 2219e5ff2159SAndi Kleen { 2220e5ff2159SAndi Kleen struct hstate *h; 2221e5ff2159SAndi Kleen 2222e5ff2159SAndi Kleen for_each_hstate(h) { 22234abd32dbSAndi Kleen char buf[32]; 2224ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 22254abd32dbSAndi Kleen memfmt(buf, huge_page_size(h)), 22264abd32dbSAndi Kleen h->free_huge_pages); 2227e5ff2159SAndi Kleen } 2228e5ff2159SAndi Kleen } 2229e5ff2159SAndi Kleen 22301da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 22316ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 22326ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 22331da177e4SLinus Torvalds { 22344415cc8dSChristoph Lameter int i; 22354415cc8dSChristoph Lameter 2236bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2237aa888a74SAndi Kleen return; 2238aa888a74SAndi Kleen 22396ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 22401da177e4SLinus Torvalds struct page *page, *next; 2241a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 2242a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 2243a5516438SAndi Kleen if (count >= h->nr_huge_pages) 22446b0c880dSAdam Litke return; 22451da177e4SLinus Torvalds if (PageHighMem(page)) 22461da177e4SLinus Torvalds continue; 22471da177e4SLinus Torvalds list_del(&page->lru); 2248e5ff2159SAndi Kleen update_and_free_page(h, page); 2249a5516438SAndi Kleen h->free_huge_pages--; 2250a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 22511da177e4SLinus Torvalds } 22521da177e4SLinus Torvalds } 22531da177e4SLinus Torvalds } 22541da177e4SLinus Torvalds #else 22556ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 22566ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 22571da177e4SLinus Torvalds { 22581da177e4SLinus Torvalds } 22591da177e4SLinus Torvalds #endif 22601da177e4SLinus Torvalds 226120a0307cSWu Fengguang /* 226220a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 226320a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 226420a0307cSWu Fengguang * Returns 1 if an adjustment was made. 226520a0307cSWu Fengguang */ 22666ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 22676ae11b27SLee Schermerhorn int delta) 226820a0307cSWu Fengguang { 2269b2261026SJoonsoo Kim int nr_nodes, node; 227020a0307cSWu Fengguang 227120a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 227220a0307cSWu Fengguang 2273e8c5c824SLee Schermerhorn if (delta < 0) { 2274b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 2275b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 2276b2261026SJoonsoo Kim goto found; 2277b2261026SJoonsoo Kim } 2278b2261026SJoonsoo Kim } else { 2279b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2280b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 2281b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 2282b2261026SJoonsoo Kim goto found; 2283e8c5c824SLee Schermerhorn } 22849a76db09SLee Schermerhorn } 2285b2261026SJoonsoo Kim return 0; 228620a0307cSWu Fengguang 2287b2261026SJoonsoo Kim found: 228820a0307cSWu Fengguang h->surplus_huge_pages += delta; 2289b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 2290b2261026SJoonsoo Kim return 1; 229120a0307cSWu Fengguang } 229220a0307cSWu Fengguang 2293a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 22946ae11b27SLee Schermerhorn static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, 22956ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 22961da177e4SLinus Torvalds { 22977893d1d5SAdam Litke unsigned long min_count, ret; 22981da177e4SLinus Torvalds 2299944d9fecSLuiz Capitulino if (hstate_is_gigantic(h) && !gigantic_page_supported()) 2300aa888a74SAndi Kleen return h->max_huge_pages; 2301aa888a74SAndi Kleen 23027893d1d5SAdam Litke /* 23037893d1d5SAdam Litke * Increase the pool size 23047893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 23057893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 2306d1c3fb1fSNishanth Aravamudan * 2307d15c7c09SNaoya Horiguchi * We might race with __alloc_buddy_huge_page() here and be unable 2308d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 2309d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 2310d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 2311d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 23127893d1d5SAdam Litke */ 23131da177e4SLinus Torvalds spin_lock(&hugetlb_lock); 2314a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 23156ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 23167893d1d5SAdam Litke break; 23177893d1d5SAdam Litke } 23187893d1d5SAdam Litke 2319a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 23207893d1d5SAdam Litke /* 23217893d1d5SAdam Litke * If this allocation races such that we no longer need the 23227893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 23237893d1d5SAdam Litke * and reducing the surplus. 23247893d1d5SAdam Litke */ 23257893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 2326649920c6SJia He 2327649920c6SJia He /* yield cpu to avoid soft lockup */ 2328649920c6SJia He cond_resched(); 2329649920c6SJia He 2330944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) 2331944d9fecSLuiz Capitulino ret = alloc_fresh_gigantic_page(h, nodes_allowed); 2332944d9fecSLuiz Capitulino else 23336ae11b27SLee Schermerhorn ret = alloc_fresh_huge_page(h, nodes_allowed); 23347893d1d5SAdam Litke spin_lock(&hugetlb_lock); 23357893d1d5SAdam Litke if (!ret) 23367893d1d5SAdam Litke goto out; 23377893d1d5SAdam Litke 2338536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 2339536240f2SMel Gorman if (signal_pending(current)) 2340536240f2SMel Gorman goto out; 23417893d1d5SAdam Litke } 23427893d1d5SAdam Litke 23437893d1d5SAdam Litke /* 23447893d1d5SAdam Litke * Decrease the pool size 23457893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 23467893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 23477893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 23487893d1d5SAdam Litke * to the desired size as pages become free. 2349d1c3fb1fSNishanth Aravamudan * 2350d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 2351d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 2352d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 2353d15c7c09SNaoya Horiguchi * __alloc_buddy_huge_page() is checking the global counter, 2354d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 2355d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 2356d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 23577893d1d5SAdam Litke */ 2358a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 23596b0c880dSAdam Litke min_count = max(count, min_count); 23606ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 2361a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 23626ae11b27SLee Schermerhorn if (!free_pool_huge_page(h, nodes_allowed, 0)) 23631da177e4SLinus Torvalds break; 236455f67141SMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 23651da177e4SLinus Torvalds } 2366a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 23676ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 23687893d1d5SAdam Litke break; 23697893d1d5SAdam Litke } 23707893d1d5SAdam Litke out: 2371a5516438SAndi Kleen ret = persistent_huge_pages(h); 23721da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 23737893d1d5SAdam Litke return ret; 23741da177e4SLinus Torvalds } 23751da177e4SLinus Torvalds 2376a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 2377a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 2378a3437870SNishanth Aravamudan 2379a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 2380a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 2381a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 2382a3437870SNishanth Aravamudan 2383a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 2384a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 2385a3437870SNishanth Aravamudan 23869a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 23879a305230SLee Schermerhorn 23889a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 2389a3437870SNishanth Aravamudan { 2390a3437870SNishanth Aravamudan int i; 23919a305230SLee Schermerhorn 2392a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 23939a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 23949a305230SLee Schermerhorn if (nidp) 23959a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 2396a3437870SNishanth Aravamudan return &hstates[i]; 23979a305230SLee Schermerhorn } 23989a305230SLee Schermerhorn 23999a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 2400a3437870SNishanth Aravamudan } 2401a3437870SNishanth Aravamudan 240206808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 2403a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2404a3437870SNishanth Aravamudan { 24059a305230SLee Schermerhorn struct hstate *h; 24069a305230SLee Schermerhorn unsigned long nr_huge_pages; 24079a305230SLee Schermerhorn int nid; 24089a305230SLee Schermerhorn 24099a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 24109a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 24119a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 24129a305230SLee Schermerhorn else 24139a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 24149a305230SLee Schermerhorn 24159a305230SLee Schermerhorn return sprintf(buf, "%lu\n", nr_huge_pages); 2416a3437870SNishanth Aravamudan } 2417adbe8726SEric B Munson 2418238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 2419238d3c13SDavid Rientjes struct hstate *h, int nid, 2420238d3c13SDavid Rientjes unsigned long count, size_t len) 2421a3437870SNishanth Aravamudan { 2422a3437870SNishanth Aravamudan int err; 2423bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); 2424a3437870SNishanth Aravamudan 2425944d9fecSLuiz Capitulino if (hstate_is_gigantic(h) && !gigantic_page_supported()) { 2426adbe8726SEric B Munson err = -EINVAL; 2427adbe8726SEric B Munson goto out; 2428adbe8726SEric B Munson } 2429adbe8726SEric B Munson 24309a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 24319a305230SLee Schermerhorn /* 24329a305230SLee Schermerhorn * global hstate attribute 24339a305230SLee Schermerhorn */ 24349a305230SLee Schermerhorn if (!(obey_mempolicy && 24359a305230SLee Schermerhorn init_nodemask_of_mempolicy(nodes_allowed))) { 243606808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 24378cebfcd0SLai Jiangshan nodes_allowed = &node_states[N_MEMORY]; 243806808b08SLee Schermerhorn } 24399a305230SLee Schermerhorn } else if (nodes_allowed) { 24409a305230SLee Schermerhorn /* 24419a305230SLee Schermerhorn * per node hstate attribute: adjust count to global, 24429a305230SLee Schermerhorn * but restrict alloc/free to the specified node. 24439a305230SLee Schermerhorn */ 24449a305230SLee Schermerhorn count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 24459a305230SLee Schermerhorn init_nodemask_of_node(nodes_allowed, nid); 24469a305230SLee Schermerhorn } else 24478cebfcd0SLai Jiangshan nodes_allowed = &node_states[N_MEMORY]; 24489a305230SLee Schermerhorn 244906808b08SLee Schermerhorn h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); 2450a3437870SNishanth Aravamudan 24518cebfcd0SLai Jiangshan if (nodes_allowed != &node_states[N_MEMORY]) 245206808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 245306808b08SLee Schermerhorn 245406808b08SLee Schermerhorn return len; 2455adbe8726SEric B Munson out: 2456adbe8726SEric B Munson NODEMASK_FREE(nodes_allowed); 2457adbe8726SEric B Munson return err; 245806808b08SLee Schermerhorn } 245906808b08SLee Schermerhorn 2460238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 2461238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 2462238d3c13SDavid Rientjes size_t len) 2463238d3c13SDavid Rientjes { 2464238d3c13SDavid Rientjes struct hstate *h; 2465238d3c13SDavid Rientjes unsigned long count; 2466238d3c13SDavid Rientjes int nid; 2467238d3c13SDavid Rientjes int err; 2468238d3c13SDavid Rientjes 2469238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 2470238d3c13SDavid Rientjes if (err) 2471238d3c13SDavid Rientjes return err; 2472238d3c13SDavid Rientjes 2473238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 2474238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 2475238d3c13SDavid Rientjes } 2476238d3c13SDavid Rientjes 247706808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 247806808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 247906808b08SLee Schermerhorn { 248006808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 248106808b08SLee Schermerhorn } 248206808b08SLee Schermerhorn 248306808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 248406808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 248506808b08SLee Schermerhorn { 2486238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 2487a3437870SNishanth Aravamudan } 2488a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 2489a3437870SNishanth Aravamudan 249006808b08SLee Schermerhorn #ifdef CONFIG_NUMA 249106808b08SLee Schermerhorn 249206808b08SLee Schermerhorn /* 249306808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 249406808b08SLee Schermerhorn * huge page alloc/free. 249506808b08SLee Schermerhorn */ 249606808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 249706808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 249806808b08SLee Schermerhorn { 249906808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 250006808b08SLee Schermerhorn } 250106808b08SLee Schermerhorn 250206808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 250306808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 250406808b08SLee Schermerhorn { 2505238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 250606808b08SLee Schermerhorn } 250706808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 250806808b08SLee Schermerhorn #endif 250906808b08SLee Schermerhorn 251006808b08SLee Schermerhorn 2511a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 2512a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2513a3437870SNishanth Aravamudan { 25149a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2515a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 2516a3437870SNishanth Aravamudan } 2517adbe8726SEric B Munson 2518a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 2519a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 2520a3437870SNishanth Aravamudan { 2521a3437870SNishanth Aravamudan int err; 2522a3437870SNishanth Aravamudan unsigned long input; 25239a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2524a3437870SNishanth Aravamudan 2525bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2526adbe8726SEric B Munson return -EINVAL; 2527adbe8726SEric B Munson 25283dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 2529a3437870SNishanth Aravamudan if (err) 253073ae31e5SEric B Munson return err; 2531a3437870SNishanth Aravamudan 2532a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 2533a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 2534a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 2535a3437870SNishanth Aravamudan 2536a3437870SNishanth Aravamudan return count; 2537a3437870SNishanth Aravamudan } 2538a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 2539a3437870SNishanth Aravamudan 2540a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 2541a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2542a3437870SNishanth Aravamudan { 25439a305230SLee Schermerhorn struct hstate *h; 25449a305230SLee Schermerhorn unsigned long free_huge_pages; 25459a305230SLee Schermerhorn int nid; 25469a305230SLee Schermerhorn 25479a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 25489a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 25499a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 25509a305230SLee Schermerhorn else 25519a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 25529a305230SLee Schermerhorn 25539a305230SLee Schermerhorn return sprintf(buf, "%lu\n", free_huge_pages); 2554a3437870SNishanth Aravamudan } 2555a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 2556a3437870SNishanth Aravamudan 2557a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 2558a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2559a3437870SNishanth Aravamudan { 25609a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2561a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->resv_huge_pages); 2562a3437870SNishanth Aravamudan } 2563a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 2564a3437870SNishanth Aravamudan 2565a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 2566a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2567a3437870SNishanth Aravamudan { 25689a305230SLee Schermerhorn struct hstate *h; 25699a305230SLee Schermerhorn unsigned long surplus_huge_pages; 25709a305230SLee Schermerhorn int nid; 25719a305230SLee Schermerhorn 25729a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 25739a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 25749a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 25759a305230SLee Schermerhorn else 25769a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 25779a305230SLee Schermerhorn 25789a305230SLee Schermerhorn return sprintf(buf, "%lu\n", surplus_huge_pages); 2579a3437870SNishanth Aravamudan } 2580a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 2581a3437870SNishanth Aravamudan 2582a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 2583a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 2584a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 2585a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 2586a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 2587a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 258806808b08SLee Schermerhorn #ifdef CONFIG_NUMA 258906808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 259006808b08SLee Schermerhorn #endif 2591a3437870SNishanth Aravamudan NULL, 2592a3437870SNishanth Aravamudan }; 2593a3437870SNishanth Aravamudan 2594a3437870SNishanth Aravamudan static struct attribute_group hstate_attr_group = { 2595a3437870SNishanth Aravamudan .attrs = hstate_attrs, 2596a3437870SNishanth Aravamudan }; 2597a3437870SNishanth Aravamudan 2598094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 25999a305230SLee Schermerhorn struct kobject **hstate_kobjs, 26009a305230SLee Schermerhorn struct attribute_group *hstate_attr_group) 2601a3437870SNishanth Aravamudan { 2602a3437870SNishanth Aravamudan int retval; 2603972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 2604a3437870SNishanth Aravamudan 26059a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 26069a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 2607a3437870SNishanth Aravamudan return -ENOMEM; 2608a3437870SNishanth Aravamudan 26099a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 2610a3437870SNishanth Aravamudan if (retval) 26119a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 2612a3437870SNishanth Aravamudan 2613a3437870SNishanth Aravamudan return retval; 2614a3437870SNishanth Aravamudan } 2615a3437870SNishanth Aravamudan 2616a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 2617a3437870SNishanth Aravamudan { 2618a3437870SNishanth Aravamudan struct hstate *h; 2619a3437870SNishanth Aravamudan int err; 2620a3437870SNishanth Aravamudan 2621a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 2622a3437870SNishanth Aravamudan if (!hugepages_kobj) 2623a3437870SNishanth Aravamudan return; 2624a3437870SNishanth Aravamudan 2625a3437870SNishanth Aravamudan for_each_hstate(h) { 26269a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 26279a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 2628a3437870SNishanth Aravamudan if (err) 2629ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s", h->name); 2630a3437870SNishanth Aravamudan } 2631a3437870SNishanth Aravamudan } 2632a3437870SNishanth Aravamudan 26339a305230SLee Schermerhorn #ifdef CONFIG_NUMA 26349a305230SLee Schermerhorn 26359a305230SLee Schermerhorn /* 26369a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 263710fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 263810fbcf4cSKay Sievers * index of a node device or _hstate == node id. 263910fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 26409a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 26419a305230SLee Schermerhorn */ 26429a305230SLee Schermerhorn struct node_hstate { 26439a305230SLee Schermerhorn struct kobject *hugepages_kobj; 26449a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 26459a305230SLee Schermerhorn }; 2646b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 26479a305230SLee Schermerhorn 26489a305230SLee Schermerhorn /* 264910fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 26509a305230SLee Schermerhorn */ 26519a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 26529a305230SLee Schermerhorn &nr_hugepages_attr.attr, 26539a305230SLee Schermerhorn &free_hugepages_attr.attr, 26549a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 26559a305230SLee Schermerhorn NULL, 26569a305230SLee Schermerhorn }; 26579a305230SLee Schermerhorn 26589a305230SLee Schermerhorn static struct attribute_group per_node_hstate_attr_group = { 26599a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 26609a305230SLee Schermerhorn }; 26619a305230SLee Schermerhorn 26629a305230SLee Schermerhorn /* 266310fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 26649a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 26659a305230SLee Schermerhorn */ 26669a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 26679a305230SLee Schermerhorn { 26689a305230SLee Schermerhorn int nid; 26699a305230SLee Schermerhorn 26709a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 26719a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 26729a305230SLee Schermerhorn int i; 26739a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 26749a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 26759a305230SLee Schermerhorn if (nidp) 26769a305230SLee Schermerhorn *nidp = nid; 26779a305230SLee Schermerhorn return &hstates[i]; 26789a305230SLee Schermerhorn } 26799a305230SLee Schermerhorn } 26809a305230SLee Schermerhorn 26819a305230SLee Schermerhorn BUG(); 26829a305230SLee Schermerhorn return NULL; 26839a305230SLee Schermerhorn } 26849a305230SLee Schermerhorn 26859a305230SLee Schermerhorn /* 268610fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 26879a305230SLee Schermerhorn * No-op if no hstate attributes attached. 26889a305230SLee Schermerhorn */ 26893cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 26909a305230SLee Schermerhorn { 26919a305230SLee Schermerhorn struct hstate *h; 269210fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 26939a305230SLee Schermerhorn 26949a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 26959b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 26969a305230SLee Schermerhorn 2697972dc4deSAneesh Kumar K.V for_each_hstate(h) { 2698972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 2699972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 2700972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 2701972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 2702972dc4deSAneesh Kumar K.V } 27039a305230SLee Schermerhorn } 27049a305230SLee Schermerhorn 27059a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 27069a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 27079a305230SLee Schermerhorn } 27089a305230SLee Schermerhorn 27099a305230SLee Schermerhorn 27109a305230SLee Schermerhorn /* 271110fbcf4cSKay Sievers * Register hstate attributes for a single node device. 27129a305230SLee Schermerhorn * No-op if attributes already registered. 27139a305230SLee Schermerhorn */ 27143cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 27159a305230SLee Schermerhorn { 27169a305230SLee Schermerhorn struct hstate *h; 271710fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 27189a305230SLee Schermerhorn int err; 27199a305230SLee Schermerhorn 27209a305230SLee Schermerhorn if (nhs->hugepages_kobj) 27219a305230SLee Schermerhorn return; /* already allocated */ 27229a305230SLee Schermerhorn 27239a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 272410fbcf4cSKay Sievers &node->dev.kobj); 27259a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 27269a305230SLee Schermerhorn return; 27279a305230SLee Schermerhorn 27289a305230SLee Schermerhorn for_each_hstate(h) { 27299a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 27309a305230SLee Schermerhorn nhs->hstate_kobjs, 27319a305230SLee Schermerhorn &per_node_hstate_attr_group); 27329a305230SLee Schermerhorn if (err) { 2733ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s for node %d\n", 273410fbcf4cSKay Sievers h->name, node->dev.id); 27359a305230SLee Schermerhorn hugetlb_unregister_node(node); 27369a305230SLee Schermerhorn break; 27379a305230SLee Schermerhorn } 27389a305230SLee Schermerhorn } 27399a305230SLee Schermerhorn } 27409a305230SLee Schermerhorn 27419a305230SLee Schermerhorn /* 27429b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 274310fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 274410fbcf4cSKay Sievers * registered their associated device by this time. 27459a305230SLee Schermerhorn */ 27467d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 27479a305230SLee Schermerhorn { 27489a305230SLee Schermerhorn int nid; 27499a305230SLee Schermerhorn 27508cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 27518732794bSWen Congyang struct node *node = node_devices[nid]; 275210fbcf4cSKay Sievers if (node->dev.id == nid) 27539a305230SLee Schermerhorn hugetlb_register_node(node); 27549a305230SLee Schermerhorn } 27559a305230SLee Schermerhorn 27569a305230SLee Schermerhorn /* 275710fbcf4cSKay Sievers * Let the node device driver know we're here so it can 27589a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 27599a305230SLee Schermerhorn */ 27609a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 27619a305230SLee Schermerhorn hugetlb_unregister_node); 27629a305230SLee Schermerhorn } 27639a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 27649a305230SLee Schermerhorn 27659a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 27669a305230SLee Schermerhorn { 27679a305230SLee Schermerhorn BUG(); 27689a305230SLee Schermerhorn if (nidp) 27699a305230SLee Schermerhorn *nidp = -1; 27709a305230SLee Schermerhorn return NULL; 27719a305230SLee Schermerhorn } 27729a305230SLee Schermerhorn 27739a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 27749a305230SLee Schermerhorn 27759a305230SLee Schermerhorn #endif 27769a305230SLee Schermerhorn 2777a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 2778a3437870SNishanth Aravamudan { 27798382d914SDavidlohr Bueso int i; 27808382d914SDavidlohr Bueso 2781457c1b27SNishanth Aravamudan if (!hugepages_supported()) 27820ef89d25SBenjamin Herrenschmidt return 0; 2783a3437870SNishanth Aravamudan 2784e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) { 2785e11bfbfcSNick Piggin default_hstate_size = HPAGE_SIZE; 2786e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) 2787a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 2788a3437870SNishanth Aravamudan } 2789972dc4deSAneesh Kumar K.V default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); 2790f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 2791f8b74815SVaishali Thakkar if (!default_hstate.max_huge_pages) 2792e11bfbfcSNick Piggin default_hstate.max_huge_pages = default_hstate_max_huge_pages; 2793f8b74815SVaishali Thakkar } 2794a3437870SNishanth Aravamudan 2795a3437870SNishanth Aravamudan hugetlb_init_hstates(); 2796aa888a74SAndi Kleen gather_bootmem_prealloc(); 2797a3437870SNishanth Aravamudan report_hugepages(); 2798a3437870SNishanth Aravamudan 2799a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 28009a305230SLee Schermerhorn hugetlb_register_all_nodes(); 28017179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 28029a305230SLee Schermerhorn 28038382d914SDavidlohr Bueso #ifdef CONFIG_SMP 28048382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 28058382d914SDavidlohr Bueso #else 28068382d914SDavidlohr Bueso num_fault_mutexes = 1; 28078382d914SDavidlohr Bueso #endif 2808c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 28098382d914SDavidlohr Bueso kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL); 2810c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 28118382d914SDavidlohr Bueso 28128382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 2813c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 2814a3437870SNishanth Aravamudan return 0; 2815a3437870SNishanth Aravamudan } 28163e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 2817a3437870SNishanth Aravamudan 2818a3437870SNishanth Aravamudan /* Should be called on processing a hugepagesz=... option */ 28199fee021dSVaishali Thakkar void __init hugetlb_bad_size(void) 28209fee021dSVaishali Thakkar { 28219fee021dSVaishali Thakkar parsed_valid_hugepagesz = false; 28229fee021dSVaishali Thakkar } 28239fee021dSVaishali Thakkar 2824d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 2825a3437870SNishanth Aravamudan { 2826a3437870SNishanth Aravamudan struct hstate *h; 28278faa8b07SAndi Kleen unsigned long i; 28288faa8b07SAndi Kleen 2829a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 2830598d8091SJoe Perches pr_warn("hugepagesz= specified twice, ignoring\n"); 2831a3437870SNishanth Aravamudan return; 2832a3437870SNishanth Aravamudan } 283347d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 2834a3437870SNishanth Aravamudan BUG_ON(order == 0); 283547d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 2836a3437870SNishanth Aravamudan h->order = order; 2837a3437870SNishanth Aravamudan h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 28388faa8b07SAndi Kleen h->nr_huge_pages = 0; 28398faa8b07SAndi Kleen h->free_huge_pages = 0; 28408faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 28418faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 28420edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 284354f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 284454f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 2845a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 2846a3437870SNishanth Aravamudan huge_page_size(h)/1024); 28478faa8b07SAndi Kleen 2848a3437870SNishanth Aravamudan parsed_hstate = h; 2849a3437870SNishanth Aravamudan } 2850a3437870SNishanth Aravamudan 2851e11bfbfcSNick Piggin static int __init hugetlb_nrpages_setup(char *s) 2852a3437870SNishanth Aravamudan { 2853a3437870SNishanth Aravamudan unsigned long *mhp; 28548faa8b07SAndi Kleen static unsigned long *last_mhp; 2855a3437870SNishanth Aravamudan 28569fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 28579fee021dSVaishali Thakkar pr_warn("hugepages = %s preceded by " 28589fee021dSVaishali Thakkar "an unsupported hugepagesz, ignoring\n", s); 28599fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 28609fee021dSVaishali Thakkar return 1; 28619fee021dSVaishali Thakkar } 2862a3437870SNishanth Aravamudan /* 286347d38344SAneesh Kumar K.V * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, 2864a3437870SNishanth Aravamudan * so this hugepages= parameter goes to the "default hstate". 2865a3437870SNishanth Aravamudan */ 28669fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 2867a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 2868a3437870SNishanth Aravamudan else 2869a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 2870a3437870SNishanth Aravamudan 28718faa8b07SAndi Kleen if (mhp == last_mhp) { 2872598d8091SJoe Perches pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n"); 28738faa8b07SAndi Kleen return 1; 28748faa8b07SAndi Kleen } 28758faa8b07SAndi Kleen 2876a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 2877a3437870SNishanth Aravamudan *mhp = 0; 2878a3437870SNishanth Aravamudan 28798faa8b07SAndi Kleen /* 28808faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 28818faa8b07SAndi Kleen * But we need to allocate >= MAX_ORDER hstates here early to still 28828faa8b07SAndi Kleen * use the bootmem allocator. 28838faa8b07SAndi Kleen */ 288447d38344SAneesh Kumar K.V if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) 28858faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 28868faa8b07SAndi Kleen 28878faa8b07SAndi Kleen last_mhp = mhp; 28888faa8b07SAndi Kleen 2889a3437870SNishanth Aravamudan return 1; 2890a3437870SNishanth Aravamudan } 2891e11bfbfcSNick Piggin __setup("hugepages=", hugetlb_nrpages_setup); 2892e11bfbfcSNick Piggin 2893e11bfbfcSNick Piggin static int __init hugetlb_default_setup(char *s) 2894e11bfbfcSNick Piggin { 2895e11bfbfcSNick Piggin default_hstate_size = memparse(s, &s); 2896e11bfbfcSNick Piggin return 1; 2897e11bfbfcSNick Piggin } 2898e11bfbfcSNick Piggin __setup("default_hugepagesz=", hugetlb_default_setup); 2899a3437870SNishanth Aravamudan 29008a213460SNishanth Aravamudan static unsigned int cpuset_mems_nr(unsigned int *array) 29018a213460SNishanth Aravamudan { 29028a213460SNishanth Aravamudan int node; 29038a213460SNishanth Aravamudan unsigned int nr = 0; 29048a213460SNishanth Aravamudan 29058a213460SNishanth Aravamudan for_each_node_mask(node, cpuset_current_mems_allowed) 29068a213460SNishanth Aravamudan nr += array[node]; 29078a213460SNishanth Aravamudan 29088a213460SNishanth Aravamudan return nr; 29098a213460SNishanth Aravamudan } 29108a213460SNishanth Aravamudan 29118a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 291206808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 291306808b08SLee Schermerhorn struct ctl_table *table, int write, 291406808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 29151da177e4SLinus Torvalds { 2916e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 2917238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 291808d4a246SMichal Hocko int ret; 2919e5ff2159SAndi Kleen 2920457c1b27SNishanth Aravamudan if (!hugepages_supported()) 292186613628SJan Stancek return -EOPNOTSUPP; 2922457c1b27SNishanth Aravamudan 2923e5ff2159SAndi Kleen table->data = &tmp; 2924e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 292508d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 292608d4a246SMichal Hocko if (ret) 292708d4a246SMichal Hocko goto out; 2928e5ff2159SAndi Kleen 2929238d3c13SDavid Rientjes if (write) 2930238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 2931238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 293208d4a246SMichal Hocko out: 293308d4a246SMichal Hocko return ret; 29341da177e4SLinus Torvalds } 2935396faf03SMel Gorman 293606808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 293706808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 293806808b08SLee Schermerhorn { 293906808b08SLee Schermerhorn 294006808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 294106808b08SLee Schermerhorn buffer, length, ppos); 294206808b08SLee Schermerhorn } 294306808b08SLee Schermerhorn 294406808b08SLee Schermerhorn #ifdef CONFIG_NUMA 294506808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 294606808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 294706808b08SLee Schermerhorn { 294806808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 294906808b08SLee Schermerhorn buffer, length, ppos); 295006808b08SLee Schermerhorn } 295106808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 295206808b08SLee Schermerhorn 2953a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 29548d65af78SAlexey Dobriyan void __user *buffer, 2955a3d0c6aaSNishanth Aravamudan size_t *length, loff_t *ppos) 2956a3d0c6aaSNishanth Aravamudan { 2957a5516438SAndi Kleen struct hstate *h = &default_hstate; 2958e5ff2159SAndi Kleen unsigned long tmp; 295908d4a246SMichal Hocko int ret; 2960e5ff2159SAndi Kleen 2961457c1b27SNishanth Aravamudan if (!hugepages_supported()) 296286613628SJan Stancek return -EOPNOTSUPP; 2963457c1b27SNishanth Aravamudan 2964e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 2965e5ff2159SAndi Kleen 2966bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 2967adbe8726SEric B Munson return -EINVAL; 2968adbe8726SEric B Munson 2969e5ff2159SAndi Kleen table->data = &tmp; 2970e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 297108d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 297208d4a246SMichal Hocko if (ret) 297308d4a246SMichal Hocko goto out; 2974e5ff2159SAndi Kleen 2975e5ff2159SAndi Kleen if (write) { 2976064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 2977e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 2978a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 2979e5ff2159SAndi Kleen } 298008d4a246SMichal Hocko out: 298108d4a246SMichal Hocko return ret; 2982a3d0c6aaSNishanth Aravamudan } 2983a3d0c6aaSNishanth Aravamudan 29841da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 29851da177e4SLinus Torvalds 2986e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 29871da177e4SLinus Torvalds { 2988a5516438SAndi Kleen struct hstate *h = &default_hstate; 2989457c1b27SNishanth Aravamudan if (!hugepages_supported()) 2990457c1b27SNishanth Aravamudan return; 2991e1759c21SAlexey Dobriyan seq_printf(m, 29921da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 29931da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 2994b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 29957893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 29964f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 2997a5516438SAndi Kleen h->nr_huge_pages, 2998a5516438SAndi Kleen h->free_huge_pages, 2999a5516438SAndi Kleen h->resv_huge_pages, 3000a5516438SAndi Kleen h->surplus_huge_pages, 3001a5516438SAndi Kleen 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 30021da177e4SLinus Torvalds } 30031da177e4SLinus Torvalds 30041da177e4SLinus Torvalds int hugetlb_report_node_meminfo(int nid, char *buf) 30051da177e4SLinus Torvalds { 3006a5516438SAndi Kleen struct hstate *h = &default_hstate; 3007457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3008457c1b27SNishanth Aravamudan return 0; 30091da177e4SLinus Torvalds return sprintf(buf, 30101da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 3011a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 3012a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 3013a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 3014a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 3015a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 30161da177e4SLinus Torvalds } 30171da177e4SLinus Torvalds 3018949f7ec5SDavid Rientjes void hugetlb_show_meminfo(void) 3019949f7ec5SDavid Rientjes { 3020949f7ec5SDavid Rientjes struct hstate *h; 3021949f7ec5SDavid Rientjes int nid; 3022949f7ec5SDavid Rientjes 3023457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3024457c1b27SNishanth Aravamudan return; 3025457c1b27SNishanth Aravamudan 3026949f7ec5SDavid Rientjes for_each_node_state(nid, N_MEMORY) 3027949f7ec5SDavid Rientjes for_each_hstate(h) 3028949f7ec5SDavid Rientjes pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 3029949f7ec5SDavid Rientjes nid, 3030949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 3031949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 3032949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 3033949f7ec5SDavid Rientjes 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 3034949f7ec5SDavid Rientjes } 3035949f7ec5SDavid Rientjes 30365d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 30375d317b2bSNaoya Horiguchi { 30385d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 30395d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 30405d317b2bSNaoya Horiguchi } 30415d317b2bSNaoya Horiguchi 30421da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 30431da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 30441da177e4SLinus Torvalds { 3045d0028588SWanpeng Li struct hstate *h; 3046d0028588SWanpeng Li unsigned long nr_total_pages = 0; 3047d0028588SWanpeng Li 3048d0028588SWanpeng Li for_each_hstate(h) 3049d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 3050d0028588SWanpeng Li return nr_total_pages; 30511da177e4SLinus Torvalds } 30521da177e4SLinus Torvalds 3053a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 3054fc1b8a73SMel Gorman { 3055fc1b8a73SMel Gorman int ret = -ENOMEM; 3056fc1b8a73SMel Gorman 3057fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 3058fc1b8a73SMel Gorman /* 3059fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 3060fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 3061fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 3062fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 3063fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 3064fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 3065fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 3066fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 3067fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 3068fc1b8a73SMel Gorman * 3069fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 3070fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 3071fc1b8a73SMel Gorman * we fall back to check against current free page availability as 3072fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 3073fc1b8a73SMel Gorman * semantics that cpuset has. 3074fc1b8a73SMel Gorman */ 3075fc1b8a73SMel Gorman if (delta > 0) { 3076a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 3077fc1b8a73SMel Gorman goto out; 3078fc1b8a73SMel Gorman 3079a5516438SAndi Kleen if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 3080a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 3081fc1b8a73SMel Gorman goto out; 3082fc1b8a73SMel Gorman } 3083fc1b8a73SMel Gorman } 3084fc1b8a73SMel Gorman 3085fc1b8a73SMel Gorman ret = 0; 3086fc1b8a73SMel Gorman if (delta < 0) 3087a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 3088fc1b8a73SMel Gorman 3089fc1b8a73SMel Gorman out: 3090fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 3091fc1b8a73SMel Gorman return ret; 3092fc1b8a73SMel Gorman } 3093fc1b8a73SMel Gorman 309484afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 309584afd99bSAndy Whitcroft { 3096f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 309784afd99bSAndy Whitcroft 309884afd99bSAndy Whitcroft /* 309984afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 310084afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 310184afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 310225985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 310384afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 310484afd99bSAndy Whitcroft * new reference here without additional locking. 310584afd99bSAndy Whitcroft */ 31064e35f483SJoonsoo Kim if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 3107f522c3acSJoonsoo Kim kref_get(&resv->refs); 310884afd99bSAndy Whitcroft } 310984afd99bSAndy Whitcroft 3110a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 3111a1e78772SMel Gorman { 3112a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3113f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 311490481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 31154e35f483SJoonsoo Kim unsigned long reserve, start, end; 31161c5ecae3SMike Kravetz long gbl_reserve; 311784afd99bSAndy Whitcroft 31184e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 31194e35f483SJoonsoo Kim return; 31204e35f483SJoonsoo Kim 3121a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 3122a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 312384afd99bSAndy Whitcroft 31244e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 312584afd99bSAndy Whitcroft 3126f031dd27SJoonsoo Kim kref_put(&resv->refs, resv_map_release); 312784afd99bSAndy Whitcroft 31287251ff78SAdam Litke if (reserve) { 31291c5ecae3SMike Kravetz /* 31301c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 31311c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 31321c5ecae3SMike Kravetz */ 31331c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 31341c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 31357251ff78SAdam Litke } 3136a1e78772SMel Gorman } 3137a1e78772SMel Gorman 31381da177e4SLinus Torvalds /* 31391da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 31401da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 31411da177e4SLinus Torvalds * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 31421da177e4SLinus Torvalds * this far. 31431da177e4SLinus Torvalds */ 3144d0217ac0SNick Piggin static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) 31451da177e4SLinus Torvalds { 31461da177e4SLinus Torvalds BUG(); 3147d0217ac0SNick Piggin return 0; 31481da177e4SLinus Torvalds } 31491da177e4SLinus Torvalds 3150f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 3151d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 315284afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 3153a1e78772SMel Gorman .close = hugetlb_vm_op_close, 31541da177e4SLinus Torvalds }; 31551da177e4SLinus Torvalds 31561e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 31571e8f889bSDavid Gibson int writable) 315863551ae0SDavid Gibson { 315963551ae0SDavid Gibson pte_t entry; 316063551ae0SDavid Gibson 31611e8f889bSDavid Gibson if (writable) { 3162106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 3163106c992aSGerald Schaefer vma->vm_page_prot))); 316463551ae0SDavid Gibson } else { 3165106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 3166106c992aSGerald Schaefer vma->vm_page_prot)); 316763551ae0SDavid Gibson } 316863551ae0SDavid Gibson entry = pte_mkyoung(entry); 316963551ae0SDavid Gibson entry = pte_mkhuge(entry); 3170d9ed9faaSChris Metcalf entry = arch_make_huge_pte(entry, vma, page, writable); 317163551ae0SDavid Gibson 317263551ae0SDavid Gibson return entry; 317363551ae0SDavid Gibson } 317463551ae0SDavid Gibson 31751e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 31761e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 31771e8f889bSDavid Gibson { 31781e8f889bSDavid Gibson pte_t entry; 31791e8f889bSDavid Gibson 3180106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 318132f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 31824b3073e1SRussell King update_mmu_cache(vma, address, ptep); 31831e8f889bSDavid Gibson } 31841e8f889bSDavid Gibson 31854a705fefSNaoya Horiguchi static int is_hugetlb_entry_migration(pte_t pte) 31864a705fefSNaoya Horiguchi { 31874a705fefSNaoya Horiguchi swp_entry_t swp; 31884a705fefSNaoya Horiguchi 31894a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 31904a705fefSNaoya Horiguchi return 0; 31914a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 31924a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_migration_entry(swp)) 31934a705fefSNaoya Horiguchi return 1; 31944a705fefSNaoya Horiguchi else 31954a705fefSNaoya Horiguchi return 0; 31964a705fefSNaoya Horiguchi } 31974a705fefSNaoya Horiguchi 31984a705fefSNaoya Horiguchi static int is_hugetlb_entry_hwpoisoned(pte_t pte) 31994a705fefSNaoya Horiguchi { 32004a705fefSNaoya Horiguchi swp_entry_t swp; 32014a705fefSNaoya Horiguchi 32024a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 32034a705fefSNaoya Horiguchi return 0; 32044a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 32054a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_hwpoison_entry(swp)) 32064a705fefSNaoya Horiguchi return 1; 32074a705fefSNaoya Horiguchi else 32084a705fefSNaoya Horiguchi return 0; 32094a705fefSNaoya Horiguchi } 32101e8f889bSDavid Gibson 321163551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 321263551ae0SDavid Gibson struct vm_area_struct *vma) 321363551ae0SDavid Gibson { 321463551ae0SDavid Gibson pte_t *src_pte, *dst_pte, entry; 321563551ae0SDavid Gibson struct page *ptepage; 32161c59827dSHugh Dickins unsigned long addr; 32171e8f889bSDavid Gibson int cow; 3218a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3219a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 3220e8569dd2SAndreas Sandberg unsigned long mmun_start; /* For mmu_notifiers */ 3221e8569dd2SAndreas Sandberg unsigned long mmun_end; /* For mmu_notifiers */ 3222e8569dd2SAndreas Sandberg int ret = 0; 32231e8f889bSDavid Gibson 32241e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 322563551ae0SDavid Gibson 3226e8569dd2SAndreas Sandberg mmun_start = vma->vm_start; 3227e8569dd2SAndreas Sandberg mmun_end = vma->vm_end; 3228e8569dd2SAndreas Sandberg if (cow) 3229e8569dd2SAndreas Sandberg mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end); 3230e8569dd2SAndreas Sandberg 3231a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 3232cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 3233c74df32cSHugh Dickins src_pte = huge_pte_offset(src, addr); 3234c74df32cSHugh Dickins if (!src_pte) 3235c74df32cSHugh Dickins continue; 3236a5516438SAndi Kleen dst_pte = huge_pte_alloc(dst, addr, sz); 3237e8569dd2SAndreas Sandberg if (!dst_pte) { 3238e8569dd2SAndreas Sandberg ret = -ENOMEM; 3239e8569dd2SAndreas Sandberg break; 3240e8569dd2SAndreas Sandberg } 3241c5c99429SLarry Woodman 3242c5c99429SLarry Woodman /* If the pagetables are shared don't copy or take references */ 3243c5c99429SLarry Woodman if (dst_pte == src_pte) 3244c5c99429SLarry Woodman continue; 3245c5c99429SLarry Woodman 3246cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 3247cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 3248cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 32494a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 32504a705fefSNaoya Horiguchi if (huge_pte_none(entry)) { /* skip none entry */ 32514a705fefSNaoya Horiguchi ; 32524a705fefSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry) || 32534a705fefSNaoya Horiguchi is_hugetlb_entry_hwpoisoned(entry))) { 32544a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 32554a705fefSNaoya Horiguchi 32564a705fefSNaoya Horiguchi if (is_write_migration_entry(swp_entry) && cow) { 32574a705fefSNaoya Horiguchi /* 32584a705fefSNaoya Horiguchi * COW mappings require pages in both 32594a705fefSNaoya Horiguchi * parent and child to be set to read. 32604a705fefSNaoya Horiguchi */ 32614a705fefSNaoya Horiguchi make_migration_entry_read(&swp_entry); 32624a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 32634a705fefSNaoya Horiguchi set_huge_pte_at(src, addr, src_pte, entry); 32644a705fefSNaoya Horiguchi } 32654a705fefSNaoya Horiguchi set_huge_pte_at(dst, addr, dst_pte, entry); 32664a705fefSNaoya Horiguchi } else { 326734ee645eSJoerg Roedel if (cow) { 32687f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 326934ee645eSJoerg Roedel mmu_notifier_invalidate_range(src, mmun_start, 327034ee645eSJoerg Roedel mmun_end); 327134ee645eSJoerg Roedel } 32720253d634SNaoya Horiguchi entry = huge_ptep_get(src_pte); 327363551ae0SDavid Gibson ptepage = pte_page(entry); 327463551ae0SDavid Gibson get_page(ptepage); 327553f9263bSKirill A. Shutemov page_dup_rmap(ptepage, true); 327663551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 32775d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), dst); 32781c59827dSHugh Dickins } 3279cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 3280cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 328163551ae0SDavid Gibson } 328263551ae0SDavid Gibson 3283e8569dd2SAndreas Sandberg if (cow) 3284e8569dd2SAndreas Sandberg mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end); 3285e8569dd2SAndreas Sandberg 3286e8569dd2SAndreas Sandberg return ret; 328763551ae0SDavid Gibson } 328863551ae0SDavid Gibson 328924669e58SAneesh Kumar K.V void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 329024669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 329124669e58SAneesh Kumar K.V struct page *ref_page) 329263551ae0SDavid Gibson { 329363551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 329463551ae0SDavid Gibson unsigned long address; 3295c7546f8fSDavid Gibson pte_t *ptep; 329663551ae0SDavid Gibson pte_t pte; 3297cb900f41SKirill A. Shutemov spinlock_t *ptl; 329863551ae0SDavid Gibson struct page *page; 3299a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3300a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 33012ec74c3eSSagi Grimberg const unsigned long mmun_start = start; /* For mmu_notifiers */ 33022ec74c3eSSagi Grimberg const unsigned long mmun_end = end; /* For mmu_notifiers */ 3303a5516438SAndi Kleen 330463551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 3305a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 3306a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 330763551ae0SDavid Gibson 330807e32661SAneesh Kumar K.V /* 330907e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 331007e32661SAneesh Kumar K.V * to huge page. 331107e32661SAneesh Kumar K.V */ 331207e32661SAneesh Kumar K.V tlb_remove_check_page_size_change(tlb, sz); 331324669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 33142ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 3315569f48b8SHillf Danton address = start; 3316569f48b8SHillf Danton for (; address < end; address += sz) { 3317c7546f8fSDavid Gibson ptep = huge_pte_offset(mm, address); 3318c7546f8fSDavid Gibson if (!ptep) 3319c7546f8fSDavid Gibson continue; 3320c7546f8fSDavid Gibson 3321cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 332231d49da5SAneesh Kumar K.V if (huge_pmd_unshare(mm, &address, ptep)) { 332331d49da5SAneesh Kumar K.V spin_unlock(ptl); 332431d49da5SAneesh Kumar K.V continue; 332531d49da5SAneesh Kumar K.V } 332639dde65cSChen, Kenneth W 33276629326bSHillf Danton pte = huge_ptep_get(ptep); 332831d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 332931d49da5SAneesh Kumar K.V spin_unlock(ptl); 333031d49da5SAneesh Kumar K.V continue; 333131d49da5SAneesh Kumar K.V } 33326629326bSHillf Danton 33336629326bSHillf Danton /* 33349fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 33359fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 33366629326bSHillf Danton */ 33379fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 3338106c992aSGerald Schaefer huge_pte_clear(mm, address, ptep); 333931d49da5SAneesh Kumar K.V spin_unlock(ptl); 334031d49da5SAneesh Kumar K.V continue; 33418c4894c6SNaoya Horiguchi } 33426629326bSHillf Danton 33436629326bSHillf Danton page = pte_page(pte); 334404f2cbe3SMel Gorman /* 334504f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 334604f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 334704f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 334804f2cbe3SMel Gorman */ 334904f2cbe3SMel Gorman if (ref_page) { 335031d49da5SAneesh Kumar K.V if (page != ref_page) { 335131d49da5SAneesh Kumar K.V spin_unlock(ptl); 335231d49da5SAneesh Kumar K.V continue; 335331d49da5SAneesh Kumar K.V } 335404f2cbe3SMel Gorman /* 335504f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 335604f2cbe3SMel Gorman * future faults in this VMA will fail rather than 335704f2cbe3SMel Gorman * looking like data was lost 335804f2cbe3SMel Gorman */ 335904f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 336004f2cbe3SMel Gorman } 336104f2cbe3SMel Gorman 3362c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 3363b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 3364106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 33656649a386SKen Chen set_page_dirty(page); 33669e81130bSHillf Danton 33675d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 3368d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 336931d49da5SAneesh Kumar K.V 3370cb900f41SKirill A. Shutemov spin_unlock(ptl); 3371e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 337224669e58SAneesh Kumar K.V /* 337331d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 337424669e58SAneesh Kumar K.V */ 337531d49da5SAneesh Kumar K.V if (ref_page) 337631d49da5SAneesh Kumar K.V break; 3377fe1668aeSChen, Kenneth W } 33782ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 337924669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 33801da177e4SLinus Torvalds } 338163551ae0SDavid Gibson 3382d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 3383d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 3384d833352aSMel Gorman unsigned long end, struct page *ref_page) 3385d833352aSMel Gorman { 3386d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 3387d833352aSMel Gorman 3388d833352aSMel Gorman /* 3389d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 3390d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 3391d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 3392d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 3393d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 3394c8c06efaSDavidlohr Bueso * is to clear it before releasing the i_mmap_rwsem. This works 3395d833352aSMel Gorman * because in the context this is called, the VMA is about to be 3396c8c06efaSDavidlohr Bueso * destroyed and the i_mmap_rwsem is held. 3397d833352aSMel Gorman */ 3398d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 3399d833352aSMel Gorman } 3400d833352aSMel Gorman 3401502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 340204f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 3403502717f4SChen, Kenneth W { 340424669e58SAneesh Kumar K.V struct mm_struct *mm; 340524669e58SAneesh Kumar K.V struct mmu_gather tlb; 340624669e58SAneesh Kumar K.V 340724669e58SAneesh Kumar K.V mm = vma->vm_mm; 340824669e58SAneesh Kumar K.V 34092b047252SLinus Torvalds tlb_gather_mmu(&tlb, mm, start, end); 341024669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 341124669e58SAneesh Kumar K.V tlb_finish_mmu(&tlb, start, end); 3412502717f4SChen, Kenneth W } 3413502717f4SChen, Kenneth W 341404f2cbe3SMel Gorman /* 341504f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 341604f2cbe3SMel Gorman * mappping it owns the reserve page for. The intention is to unmap the page 341704f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 341804f2cbe3SMel Gorman * same region. 341904f2cbe3SMel Gorman */ 34202f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 34212a4b3dedSHarvey Harrison struct page *page, unsigned long address) 342204f2cbe3SMel Gorman { 34237526674dSAdam Litke struct hstate *h = hstate_vma(vma); 342404f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 342504f2cbe3SMel Gorman struct address_space *mapping; 342604f2cbe3SMel Gorman pgoff_t pgoff; 342704f2cbe3SMel Gorman 342804f2cbe3SMel Gorman /* 342904f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 343004f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 343104f2cbe3SMel Gorman */ 34327526674dSAdam Litke address = address & huge_page_mask(h); 343336e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 343436e4f20aSMichal Hocko vma->vm_pgoff; 343593c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 343604f2cbe3SMel Gorman 34374eb2b1dcSMel Gorman /* 34384eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 34394eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 34404eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 34414eb2b1dcSMel Gorman */ 344283cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 34436b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 344404f2cbe3SMel Gorman /* Do not unmap the current VMA */ 344504f2cbe3SMel Gorman if (iter_vma == vma) 344604f2cbe3SMel Gorman continue; 344704f2cbe3SMel Gorman 344804f2cbe3SMel Gorman /* 34492f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 34502f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 34512f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 34522f84a899SMel Gorman */ 34532f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 34542f84a899SMel Gorman continue; 34552f84a899SMel Gorman 34562f84a899SMel Gorman /* 345704f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 345804f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 345904f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 346004f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 346104f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 346204f2cbe3SMel Gorman */ 346304f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 346424669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 346524669e58SAneesh Kumar K.V address + huge_page_size(h), page); 346604f2cbe3SMel Gorman } 346783cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 346804f2cbe3SMel Gorman } 346904f2cbe3SMel Gorman 34700fe6e20bSNaoya Horiguchi /* 34710fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 3472ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 3473ef009b25SMichal Hocko * cannot race with other handlers or page migration. 3474ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 34750fe6e20bSNaoya Horiguchi */ 34761e8f889bSDavid Gibson static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 34773999f52eSAneesh Kumar K.V unsigned long address, pte_t *ptep, 3478cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 34791e8f889bSDavid Gibson { 34803999f52eSAneesh Kumar K.V pte_t pte; 3481a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 34821e8f889bSDavid Gibson struct page *old_page, *new_page; 3483ad4404a2SDavidlohr Bueso int ret = 0, outside_reserve = 0; 34842ec74c3eSSagi Grimberg unsigned long mmun_start; /* For mmu_notifiers */ 34852ec74c3eSSagi Grimberg unsigned long mmun_end; /* For mmu_notifiers */ 34861e8f889bSDavid Gibson 34873999f52eSAneesh Kumar K.V pte = huge_ptep_get(ptep); 34881e8f889bSDavid Gibson old_page = pte_page(pte); 34891e8f889bSDavid Gibson 349004f2cbe3SMel Gorman retry_avoidcopy: 34911e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 34921e8f889bSDavid Gibson * and just make the page writable */ 349337a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 34945a49973dSHugh Dickins page_move_anon_rmap(old_page, vma); 34951e8f889bSDavid Gibson set_huge_ptep_writable(vma, address, ptep); 349683c54070SNick Piggin return 0; 34971e8f889bSDavid Gibson } 34981e8f889bSDavid Gibson 349904f2cbe3SMel Gorman /* 350004f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 350104f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 350204f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 350304f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 350404f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 350504f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 350604f2cbe3SMel Gorman * of the full address range. 350704f2cbe3SMel Gorman */ 35085944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 350904f2cbe3SMel Gorman old_page != pagecache_page) 351004f2cbe3SMel Gorman outside_reserve = 1; 351104f2cbe3SMel Gorman 351209cbfeafSKirill A. Shutemov get_page(old_page); 3513b76c8cfbSLarry Woodman 3514ad4404a2SDavidlohr Bueso /* 3515ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 3516ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 3517ad4404a2SDavidlohr Bueso */ 3518cb900f41SKirill A. Shutemov spin_unlock(ptl); 351904f2cbe3SMel Gorman new_page = alloc_huge_page(vma, address, outside_reserve); 35201e8f889bSDavid Gibson 35212fc39cecSAdam Litke if (IS_ERR(new_page)) { 352204f2cbe3SMel Gorman /* 352304f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 352404f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 352504f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 352604f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 352704f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 352804f2cbe3SMel Gorman */ 352904f2cbe3SMel Gorman if (outside_reserve) { 353009cbfeafSKirill A. Shutemov put_page(old_page); 353104f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 35322f4612afSDavidlohr Bueso unmap_ref_private(mm, vma, old_page, address); 353304f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 3534cb900f41SKirill A. Shutemov spin_lock(ptl); 3535a734bcc8SHillf Danton ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 3536a9af0c5dSNaoya Horiguchi if (likely(ptep && 3537a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 353804f2cbe3SMel Gorman goto retry_avoidcopy; 3539a734bcc8SHillf Danton /* 3540cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 3541cb900f41SKirill A. Shutemov * lock, and our job is done. 3542a734bcc8SHillf Danton */ 3543a734bcc8SHillf Danton return 0; 354404f2cbe3SMel Gorman } 354504f2cbe3SMel Gorman 3546ad4404a2SDavidlohr Bueso ret = (PTR_ERR(new_page) == -ENOMEM) ? 3547ad4404a2SDavidlohr Bueso VM_FAULT_OOM : VM_FAULT_SIGBUS; 3548ad4404a2SDavidlohr Bueso goto out_release_old; 35491e8f889bSDavid Gibson } 35501e8f889bSDavid Gibson 35510fe6e20bSNaoya Horiguchi /* 35520fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 35530fe6e20bSNaoya Horiguchi * anon_vma prepared. 35540fe6e20bSNaoya Horiguchi */ 355544e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 3556ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 3557ad4404a2SDavidlohr Bueso goto out_release_all; 355844e2aa93SDean Nelson } 35590fe6e20bSNaoya Horiguchi 356047ad8475SAndrea Arcangeli copy_user_huge_page(new_page, old_page, address, vma, 356147ad8475SAndrea Arcangeli pages_per_huge_page(h)); 35620ed361deSNick Piggin __SetPageUptodate(new_page); 3563bcc54222SNaoya Horiguchi set_page_huge_active(new_page); 35641e8f889bSDavid Gibson 35652ec74c3eSSagi Grimberg mmun_start = address & huge_page_mask(h); 35662ec74c3eSSagi Grimberg mmun_end = mmun_start + huge_page_size(h); 35672ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 3568ad4404a2SDavidlohr Bueso 3569b76c8cfbSLarry Woodman /* 3570cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 3571b76c8cfbSLarry Woodman * before the page tables are altered 3572b76c8cfbSLarry Woodman */ 3573cb900f41SKirill A. Shutemov spin_lock(ptl); 3574a5516438SAndi Kleen ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 3575a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 357607443a85SJoonsoo Kim ClearPagePrivate(new_page); 357707443a85SJoonsoo Kim 35781e8f889bSDavid Gibson /* Break COW */ 35798fe627ecSGerald Schaefer huge_ptep_clear_flush(vma, address, ptep); 358034ee645eSJoerg Roedel mmu_notifier_invalidate_range(mm, mmun_start, mmun_end); 35811e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, 35821e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 3583d281ee61SKirill A. Shutemov page_remove_rmap(old_page, true); 3584cd67f0d2SNaoya Horiguchi hugepage_add_new_anon_rmap(new_page, vma, address); 35851e8f889bSDavid Gibson /* Make the old page be freed below */ 35861e8f889bSDavid Gibson new_page = old_page; 35871e8f889bSDavid Gibson } 3588cb900f41SKirill A. Shutemov spin_unlock(ptl); 35892ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 3590ad4404a2SDavidlohr Bueso out_release_all: 359196b96a96SMike Kravetz restore_reserve_on_error(h, vma, address, new_page); 359209cbfeafSKirill A. Shutemov put_page(new_page); 3593ad4404a2SDavidlohr Bueso out_release_old: 359409cbfeafSKirill A. Shutemov put_page(old_page); 35958312034fSJoonsoo Kim 3596ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 3597ad4404a2SDavidlohr Bueso return ret; 35981e8f889bSDavid Gibson } 35991e8f889bSDavid Gibson 360004f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 3601a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 3602a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 360304f2cbe3SMel Gorman { 360404f2cbe3SMel Gorman struct address_space *mapping; 3605e7c4b0bfSAndy Whitcroft pgoff_t idx; 360604f2cbe3SMel Gorman 360704f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 3608a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 360904f2cbe3SMel Gorman 361004f2cbe3SMel Gorman return find_lock_page(mapping, idx); 361104f2cbe3SMel Gorman } 361204f2cbe3SMel Gorman 36133ae77f43SHugh Dickins /* 36143ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 36153ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 36163ae77f43SHugh Dickins */ 36173ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 36182a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 36192a15efc9SHugh Dickins { 36202a15efc9SHugh Dickins struct address_space *mapping; 36212a15efc9SHugh Dickins pgoff_t idx; 36222a15efc9SHugh Dickins struct page *page; 36232a15efc9SHugh Dickins 36242a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 36252a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 36262a15efc9SHugh Dickins 36272a15efc9SHugh Dickins page = find_get_page(mapping, idx); 36282a15efc9SHugh Dickins if (page) 36292a15efc9SHugh Dickins put_page(page); 36302a15efc9SHugh Dickins return page != NULL; 36312a15efc9SHugh Dickins } 36322a15efc9SHugh Dickins 3633ab76ad54SMike Kravetz int huge_add_to_page_cache(struct page *page, struct address_space *mapping, 3634ab76ad54SMike Kravetz pgoff_t idx) 3635ab76ad54SMike Kravetz { 3636ab76ad54SMike Kravetz struct inode *inode = mapping->host; 3637ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 3638ab76ad54SMike Kravetz int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 3639ab76ad54SMike Kravetz 3640ab76ad54SMike Kravetz if (err) 3641ab76ad54SMike Kravetz return err; 3642ab76ad54SMike Kravetz ClearPagePrivate(page); 3643ab76ad54SMike Kravetz 3644ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 3645ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 3646ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 3647ab76ad54SMike Kravetz return 0; 3648ab76ad54SMike Kravetz } 3649ab76ad54SMike Kravetz 3650a1ed3ddaSRobert P. J. Day static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, 36518382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 3652788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 3653ac9b9c66SHugh Dickins { 3654a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3655ac9b9c66SHugh Dickins int ret = VM_FAULT_SIGBUS; 3656409eb8c2SHillf Danton int anon_rmap = 0; 36574c887265SAdam Litke unsigned long size; 36584c887265SAdam Litke struct page *page; 36591e8f889bSDavid Gibson pte_t new_pte; 3660cb900f41SKirill A. Shutemov spinlock_t *ptl; 36614c887265SAdam Litke 366204f2cbe3SMel Gorman /* 366304f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 366404f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 366525985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 366604f2cbe3SMel Gorman */ 366704f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 3668910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 366904f2cbe3SMel Gorman current->pid); 367004f2cbe3SMel Gorman return ret; 367104f2cbe3SMel Gorman } 367204f2cbe3SMel Gorman 36734c887265SAdam Litke /* 36744c887265SAdam Litke * Use page lock to guard against racing truncation 36754c887265SAdam Litke * before we get page_table_lock. 36764c887265SAdam Litke */ 36776bda666aSChristoph Lameter retry: 36786bda666aSChristoph Lameter page = find_lock_page(mapping, idx); 36796bda666aSChristoph Lameter if (!page) { 3680a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 3681ebed4bfcSHugh Dickins if (idx >= size) 3682ebed4bfcSHugh Dickins goto out; 368304f2cbe3SMel Gorman page = alloc_huge_page(vma, address, 0); 36842fc39cecSAdam Litke if (IS_ERR(page)) { 368576dcee75SAneesh Kumar K.V ret = PTR_ERR(page); 368676dcee75SAneesh Kumar K.V if (ret == -ENOMEM) 368776dcee75SAneesh Kumar K.V ret = VM_FAULT_OOM; 368876dcee75SAneesh Kumar K.V else 368976dcee75SAneesh Kumar K.V ret = VM_FAULT_SIGBUS; 36906bda666aSChristoph Lameter goto out; 36916bda666aSChristoph Lameter } 369247ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 36930ed361deSNick Piggin __SetPageUptodate(page); 3694bcc54222SNaoya Horiguchi set_page_huge_active(page); 3695ac9b9c66SHugh Dickins 3696f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 3697ab76ad54SMike Kravetz int err = huge_add_to_page_cache(page, mapping, idx); 36986bda666aSChristoph Lameter if (err) { 36996bda666aSChristoph Lameter put_page(page); 37006bda666aSChristoph Lameter if (err == -EEXIST) 37016bda666aSChristoph Lameter goto retry; 37026bda666aSChristoph Lameter goto out; 37036bda666aSChristoph Lameter } 370423be7468SMel Gorman } else { 37056bda666aSChristoph Lameter lock_page(page); 37060fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 37070fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 37080fe6e20bSNaoya Horiguchi goto backout_unlocked; 370923be7468SMel Gorman } 3710409eb8c2SHillf Danton anon_rmap = 1; 37110fe6e20bSNaoya Horiguchi } 37120fe6e20bSNaoya Horiguchi } else { 371357303d80SAndy Whitcroft /* 3714998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 3715998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 3716998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 3717fd6a03edSNaoya Horiguchi */ 3718fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 3719aa50d3a7SAndi Kleen ret = VM_FAULT_HWPOISON | 3720972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 3721fd6a03edSNaoya Horiguchi goto backout_unlocked; 37226bda666aSChristoph Lameter } 3723998b4382SNaoya Horiguchi } 37241e8f889bSDavid Gibson 372557303d80SAndy Whitcroft /* 372657303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 372757303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 372857303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 372957303d80SAndy Whitcroft * the spinlock. 373057303d80SAndy Whitcroft */ 37315e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 37322b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 37332b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 37342b26736cSAndy Whitcroft goto backout_unlocked; 37352b26736cSAndy Whitcroft } 37365e911373SMike Kravetz /* Just decrements count, does not deallocate */ 3737feba16e2SMike Kravetz vma_end_reservation(h, vma, address); 37385e911373SMike Kravetz } 373957303d80SAndy Whitcroft 37408bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 3741a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 37424c887265SAdam Litke if (idx >= size) 37434c887265SAdam Litke goto backout; 37444c887265SAdam Litke 374583c54070SNick Piggin ret = 0; 37467f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 37474c887265SAdam Litke goto backout; 37484c887265SAdam Litke 374907443a85SJoonsoo Kim if (anon_rmap) { 375007443a85SJoonsoo Kim ClearPagePrivate(page); 3751409eb8c2SHillf Danton hugepage_add_new_anon_rmap(page, vma, address); 3752ac714904SChoi Gi-yong } else 375353f9263bSKirill A. Shutemov page_dup_rmap(page, true); 37541e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 37551e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 37561e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, new_pte); 37571e8f889bSDavid Gibson 37585d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 3759788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 37601e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 37613999f52eSAneesh Kumar K.V ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); 37621e8f889bSDavid Gibson } 37631e8f889bSDavid Gibson 3764cb900f41SKirill A. Shutemov spin_unlock(ptl); 37654c887265SAdam Litke unlock_page(page); 37664c887265SAdam Litke out: 3767ac9b9c66SHugh Dickins return ret; 37684c887265SAdam Litke 37694c887265SAdam Litke backout: 3770cb900f41SKirill A. Shutemov spin_unlock(ptl); 37712b26736cSAndy Whitcroft backout_unlocked: 37724c887265SAdam Litke unlock_page(page); 377396b96a96SMike Kravetz restore_reserve_on_error(h, vma, address, page); 37744c887265SAdam Litke put_page(page); 37754c887265SAdam Litke goto out; 3776ac9b9c66SHugh Dickins } 3777ac9b9c66SHugh Dickins 37788382d914SDavidlohr Bueso #ifdef CONFIG_SMP 3779c672c7f2SMike Kravetz u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, 37808382d914SDavidlohr Bueso struct vm_area_struct *vma, 37818382d914SDavidlohr Bueso struct address_space *mapping, 37828382d914SDavidlohr Bueso pgoff_t idx, unsigned long address) 37838382d914SDavidlohr Bueso { 37848382d914SDavidlohr Bueso unsigned long key[2]; 37858382d914SDavidlohr Bueso u32 hash; 37868382d914SDavidlohr Bueso 37878382d914SDavidlohr Bueso if (vma->vm_flags & VM_SHARED) { 37888382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 37898382d914SDavidlohr Bueso key[1] = idx; 37908382d914SDavidlohr Bueso } else { 37918382d914SDavidlohr Bueso key[0] = (unsigned long) mm; 37928382d914SDavidlohr Bueso key[1] = address >> huge_page_shift(h); 37938382d914SDavidlohr Bueso } 37948382d914SDavidlohr Bueso 37958382d914SDavidlohr Bueso hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0); 37968382d914SDavidlohr Bueso 37978382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 37988382d914SDavidlohr Bueso } 37998382d914SDavidlohr Bueso #else 38008382d914SDavidlohr Bueso /* 38018382d914SDavidlohr Bueso * For uniprocesor systems we always use a single mutex, so just 38028382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 38038382d914SDavidlohr Bueso */ 3804c672c7f2SMike Kravetz u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, 38058382d914SDavidlohr Bueso struct vm_area_struct *vma, 38068382d914SDavidlohr Bueso struct address_space *mapping, 38078382d914SDavidlohr Bueso pgoff_t idx, unsigned long address) 38088382d914SDavidlohr Bueso { 38098382d914SDavidlohr Bueso return 0; 38108382d914SDavidlohr Bueso } 38118382d914SDavidlohr Bueso #endif 38128382d914SDavidlohr Bueso 381386e5216fSAdam Litke int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 3814788c7df4SHugh Dickins unsigned long address, unsigned int flags) 381586e5216fSAdam Litke { 38168382d914SDavidlohr Bueso pte_t *ptep, entry; 3817cb900f41SKirill A. Shutemov spinlock_t *ptl; 38181e8f889bSDavid Gibson int ret; 38198382d914SDavidlohr Bueso u32 hash; 38208382d914SDavidlohr Bueso pgoff_t idx; 38210fe6e20bSNaoya Horiguchi struct page *page = NULL; 382257303d80SAndy Whitcroft struct page *pagecache_page = NULL; 3823a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 38248382d914SDavidlohr Bueso struct address_space *mapping; 38250f792cf9SNaoya Horiguchi int need_wait_lock = 0; 382686e5216fSAdam Litke 38271e16a539SKAMEZAWA Hiroyuki address &= huge_page_mask(h); 38281e16a539SKAMEZAWA Hiroyuki 3829fd6a03edSNaoya Horiguchi ptep = huge_pte_offset(mm, address); 3830fd6a03edSNaoya Horiguchi if (ptep) { 3831fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 3832290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 3833cb900f41SKirill A. Shutemov migration_entry_wait_huge(vma, mm, ptep); 3834290408d4SNaoya Horiguchi return 0; 3835290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 3836aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 3837972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 38380d777df5SNaoya Horiguchi } else { 3839a5516438SAndi Kleen ptep = huge_pte_alloc(mm, address, huge_page_size(h)); 384086e5216fSAdam Litke if (!ptep) 384186e5216fSAdam Litke return VM_FAULT_OOM; 38420d777df5SNaoya Horiguchi } 384386e5216fSAdam Litke 38448382d914SDavidlohr Bueso mapping = vma->vm_file->f_mapping; 38458382d914SDavidlohr Bueso idx = vma_hugecache_offset(h, vma, address); 38468382d914SDavidlohr Bueso 38473935baa9SDavid Gibson /* 38483935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 38493935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 38503935baa9SDavid Gibson * the same page in the page cache. 38513935baa9SDavid Gibson */ 3852c672c7f2SMike Kravetz hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address); 3853c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 38548382d914SDavidlohr Bueso 38557f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 38567f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 38578382d914SDavidlohr Bueso ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 3858b4d1d99fSDavid Gibson goto out_mutex; 38593935baa9SDavid Gibson } 386086e5216fSAdam Litke 386183c54070SNick Piggin ret = 0; 38621e8f889bSDavid Gibson 386357303d80SAndy Whitcroft /* 38640f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 38650f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 38660f792cf9SNaoya Horiguchi * a active hugepage in pagecache. This goto expects the 2nd page fault, 38670f792cf9SNaoya Horiguchi * and is_hugetlb_entry_(migration|hwpoisoned) check will properly 38680f792cf9SNaoya Horiguchi * handle it. 38690f792cf9SNaoya Horiguchi */ 38700f792cf9SNaoya Horiguchi if (!pte_present(entry)) 38710f792cf9SNaoya Horiguchi goto out_mutex; 38720f792cf9SNaoya Horiguchi 38730f792cf9SNaoya Horiguchi /* 387457303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 387557303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 387657303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 387757303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 387857303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 387957303d80SAndy Whitcroft * consumed. 388057303d80SAndy Whitcroft */ 3881106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 38822b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 38832b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 3884b4d1d99fSDavid Gibson goto out_mutex; 38852b26736cSAndy Whitcroft } 38865e911373SMike Kravetz /* Just decrements count, does not deallocate */ 3887feba16e2SMike Kravetz vma_end_reservation(h, vma, address); 388857303d80SAndy Whitcroft 3889f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 389057303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 389157303d80SAndy Whitcroft vma, address); 389257303d80SAndy Whitcroft } 389357303d80SAndy Whitcroft 38940f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 38950fe6e20bSNaoya Horiguchi 38961e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 3897b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 3898cb900f41SKirill A. Shutemov goto out_ptl; 3899b4d1d99fSDavid Gibson 39000f792cf9SNaoya Horiguchi /* 39010f792cf9SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 39020f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 39030f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 39040f792cf9SNaoya Horiguchi */ 39050f792cf9SNaoya Horiguchi page = pte_page(entry); 39060f792cf9SNaoya Horiguchi if (page != pagecache_page) 39070f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 39080f792cf9SNaoya Horiguchi need_wait_lock = 1; 39090f792cf9SNaoya Horiguchi goto out_ptl; 39100f792cf9SNaoya Horiguchi } 39110f792cf9SNaoya Horiguchi 39120f792cf9SNaoya Horiguchi get_page(page); 3913b4d1d99fSDavid Gibson 3914788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 3915106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 39163999f52eSAneesh Kumar K.V ret = hugetlb_cow(mm, vma, address, ptep, 3917cb900f41SKirill A. Shutemov pagecache_page, ptl); 39180f792cf9SNaoya Horiguchi goto out_put_page; 3919b4d1d99fSDavid Gibson } 3920106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 3921b4d1d99fSDavid Gibson } 3922b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 3923788c7df4SHugh Dickins if (huge_ptep_set_access_flags(vma, address, ptep, entry, 3924788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 39254b3073e1SRussell King update_mmu_cache(vma, address, ptep); 39260f792cf9SNaoya Horiguchi out_put_page: 39270f792cf9SNaoya Horiguchi if (page != pagecache_page) 39280f792cf9SNaoya Horiguchi unlock_page(page); 39290f792cf9SNaoya Horiguchi put_page(page); 3930cb900f41SKirill A. Shutemov out_ptl: 3931cb900f41SKirill A. Shutemov spin_unlock(ptl); 393257303d80SAndy Whitcroft 393357303d80SAndy Whitcroft if (pagecache_page) { 393457303d80SAndy Whitcroft unlock_page(pagecache_page); 393557303d80SAndy Whitcroft put_page(pagecache_page); 393657303d80SAndy Whitcroft } 3937b4d1d99fSDavid Gibson out_mutex: 3938c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 39390f792cf9SNaoya Horiguchi /* 39400f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 39410f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 39420f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 39430f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 39440f792cf9SNaoya Horiguchi * here without taking refcount. 39450f792cf9SNaoya Horiguchi */ 39460f792cf9SNaoya Horiguchi if (need_wait_lock) 39470f792cf9SNaoya Horiguchi wait_on_page_locked(page); 39481e8f889bSDavid Gibson return ret; 394986e5216fSAdam Litke } 395086e5216fSAdam Litke 395128a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 395263551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 395328a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 395428a35716SMichel Lespinasse long i, unsigned int flags) 395563551ae0SDavid Gibson { 3956d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 3957d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 395828a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 3959a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 396063551ae0SDavid Gibson 396163551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 396263551ae0SDavid Gibson pte_t *pte; 3963cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 39642a15efc9SHugh Dickins int absent; 396563551ae0SDavid Gibson struct page *page; 396663551ae0SDavid Gibson 39674c887265SAdam Litke /* 396802057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 396902057967SDavid Rientjes * potentially allocating memory. 397002057967SDavid Rientjes */ 397102057967SDavid Rientjes if (unlikely(fatal_signal_pending(current))) { 397202057967SDavid Rientjes remainder = 0; 397302057967SDavid Rientjes break; 397402057967SDavid Rientjes } 397502057967SDavid Rientjes 397602057967SDavid Rientjes /* 39774c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 39782a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 39794c887265SAdam Litke * first, for the page indexing below to work. 3980cb900f41SKirill A. Shutemov * 3981cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 39824c887265SAdam Litke */ 3983a5516438SAndi Kleen pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); 3984cb900f41SKirill A. Shutemov if (pte) 3985cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 39862a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 398763551ae0SDavid Gibson 39882a15efc9SHugh Dickins /* 39892a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 39903ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 39913ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 39923ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 39933ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 39942a15efc9SHugh Dickins */ 39953ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 39963ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 3997cb900f41SKirill A. Shutemov if (pte) 3998cb900f41SKirill A. Shutemov spin_unlock(ptl); 39992a15efc9SHugh Dickins remainder = 0; 40002a15efc9SHugh Dickins break; 40012a15efc9SHugh Dickins } 40022a15efc9SHugh Dickins 40039cc3a5bdSNaoya Horiguchi /* 40049cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 40059cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 40069cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 40079cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 40089cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 40099cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 40109cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 40119cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 40129cc3a5bdSNaoya Horiguchi */ 40139cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 4014106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 4015106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 40164c887265SAdam Litke int ret; 40174c887265SAdam Litke 4018cb900f41SKirill A. Shutemov if (pte) 4019cb900f41SKirill A. Shutemov spin_unlock(ptl); 40202a15efc9SHugh Dickins ret = hugetlb_fault(mm, vma, vaddr, 40212a15efc9SHugh Dickins (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); 4022a89182c7SAdam Litke if (!(ret & VM_FAULT_ERROR)) 40234c887265SAdam Litke continue; 40244c887265SAdam Litke 40251c59827dSHugh Dickins remainder = 0; 40261c59827dSHugh Dickins break; 40271c59827dSHugh Dickins } 402863551ae0SDavid Gibson 4029a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 40307f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 4031d5d4b0aaSChen, Kenneth W same_page: 4032d6692183SChen, Kenneth W if (pages) { 403369d177c2SAndy Whitcroft pages[i] = mem_map_offset(page, pfn_offset); 4034ddc58f27SKirill A. Shutemov get_page(pages[i]); 4035d6692183SChen, Kenneth W } 403663551ae0SDavid Gibson 403763551ae0SDavid Gibson if (vmas) 403863551ae0SDavid Gibson vmas[i] = vma; 403963551ae0SDavid Gibson 404063551ae0SDavid Gibson vaddr += PAGE_SIZE; 4041d5d4b0aaSChen, Kenneth W ++pfn_offset; 404263551ae0SDavid Gibson --remainder; 404363551ae0SDavid Gibson ++i; 4044d5d4b0aaSChen, Kenneth W if (vaddr < vma->vm_end && remainder && 4045a5516438SAndi Kleen pfn_offset < pages_per_huge_page(h)) { 4046d5d4b0aaSChen, Kenneth W /* 4047d5d4b0aaSChen, Kenneth W * We use pfn_offset to avoid touching the pageframes 4048d5d4b0aaSChen, Kenneth W * of this compound page. 4049d5d4b0aaSChen, Kenneth W */ 4050d5d4b0aaSChen, Kenneth W goto same_page; 4051d5d4b0aaSChen, Kenneth W } 4052cb900f41SKirill A. Shutemov spin_unlock(ptl); 405363551ae0SDavid Gibson } 405428a35716SMichel Lespinasse *nr_pages = remainder; 405563551ae0SDavid Gibson *position = vaddr; 405663551ae0SDavid Gibson 40572a15efc9SHugh Dickins return i ? i : -EFAULT; 405863551ae0SDavid Gibson } 40598f860591SZhang, Yanmin 40605491ae7bSAneesh Kumar K.V #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE 40615491ae7bSAneesh Kumar K.V /* 40625491ae7bSAneesh Kumar K.V * ARCHes with special requirements for evicting HUGETLB backing TLB entries can 40635491ae7bSAneesh Kumar K.V * implement this. 40645491ae7bSAneesh Kumar K.V */ 40655491ae7bSAneesh Kumar K.V #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) 40665491ae7bSAneesh Kumar K.V #endif 40675491ae7bSAneesh Kumar K.V 40687da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 40698f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 40708f860591SZhang, Yanmin { 40718f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 40728f860591SZhang, Yanmin unsigned long start = address; 40738f860591SZhang, Yanmin pte_t *ptep; 40748f860591SZhang, Yanmin pte_t pte; 4075a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 40767da4d641SPeter Zijlstra unsigned long pages = 0; 40778f860591SZhang, Yanmin 40788f860591SZhang, Yanmin BUG_ON(address >= end); 40798f860591SZhang, Yanmin flush_cache_range(vma, address, end); 40808f860591SZhang, Yanmin 4081a5338093SRik van Riel mmu_notifier_invalidate_range_start(mm, start, end); 408283cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 4083a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 4084cb900f41SKirill A. Shutemov spinlock_t *ptl; 40858f860591SZhang, Yanmin ptep = huge_pte_offset(mm, address); 40868f860591SZhang, Yanmin if (!ptep) 40878f860591SZhang, Yanmin continue; 4088cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 40897da4d641SPeter Zijlstra if (huge_pmd_unshare(mm, &address, ptep)) { 40907da4d641SPeter Zijlstra pages++; 4091cb900f41SKirill A. Shutemov spin_unlock(ptl); 409239dde65cSChen, Kenneth W continue; 40937da4d641SPeter Zijlstra } 4094a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 4095a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 4096a8bda28dSNaoya Horiguchi spin_unlock(ptl); 4097a8bda28dSNaoya Horiguchi continue; 4098a8bda28dSNaoya Horiguchi } 4099a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 4100a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 4101a8bda28dSNaoya Horiguchi 4102a8bda28dSNaoya Horiguchi if (is_write_migration_entry(entry)) { 4103a8bda28dSNaoya Horiguchi pte_t newpte; 4104a8bda28dSNaoya Horiguchi 4105a8bda28dSNaoya Horiguchi make_migration_entry_read(&entry); 4106a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 4107a8bda28dSNaoya Horiguchi set_huge_pte_at(mm, address, ptep, newpte); 4108a8bda28dSNaoya Horiguchi pages++; 4109a8bda28dSNaoya Horiguchi } 4110a8bda28dSNaoya Horiguchi spin_unlock(ptl); 4111a8bda28dSNaoya Horiguchi continue; 4112a8bda28dSNaoya Horiguchi } 4113a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 41148f860591SZhang, Yanmin pte = huge_ptep_get_and_clear(mm, address, ptep); 4115106c992aSGerald Schaefer pte = pte_mkhuge(huge_pte_modify(pte, newprot)); 4116be7517d6STony Lu pte = arch_make_huge_pte(pte, vma, NULL, 0); 41178f860591SZhang, Yanmin set_huge_pte_at(mm, address, ptep, pte); 41187da4d641SPeter Zijlstra pages++; 41198f860591SZhang, Yanmin } 4120cb900f41SKirill A. Shutemov spin_unlock(ptl); 41218f860591SZhang, Yanmin } 4122d833352aSMel Gorman /* 4123c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 4124d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 4125c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 4126d833352aSMel Gorman * and that page table be reused and filled with junk. 4127d833352aSMel Gorman */ 41285491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 412934ee645eSJoerg Roedel mmu_notifier_invalidate_range(mm, start, end); 413083cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 4131a5338093SRik van Riel mmu_notifier_invalidate_range_end(mm, start, end); 41327da4d641SPeter Zijlstra 41337da4d641SPeter Zijlstra return pages << h->order; 41348f860591SZhang, Yanmin } 41358f860591SZhang, Yanmin 4136a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 4137a1e78772SMel Gorman long from, long to, 41385a6fe125SMel Gorman struct vm_area_struct *vma, 4139ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 4140e4e574b7SAdam Litke { 414117c9d12eSMel Gorman long ret, chg; 4142a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 414390481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 41449119a41eSJoonsoo Kim struct resv_map *resv_map; 41451c5ecae3SMike Kravetz long gbl_reserve; 4146e4e574b7SAdam Litke 4147a1e78772SMel Gorman /* 414817c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 414917c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 415090481622SDavid Gibson * without using reserves 415117c9d12eSMel Gorman */ 4152ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 415317c9d12eSMel Gorman return 0; 415417c9d12eSMel Gorman 415517c9d12eSMel Gorman /* 4156a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 4157a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 4158a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 4159a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 4160a1e78772SMel Gorman */ 41619119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 41624e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 41639119a41eSJoonsoo Kim 41641406ec9bSJoonsoo Kim chg = region_chg(resv_map, from, to); 41659119a41eSJoonsoo Kim 41669119a41eSJoonsoo Kim } else { 41679119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 41685a6fe125SMel Gorman if (!resv_map) 41695a6fe125SMel Gorman return -ENOMEM; 41705a6fe125SMel Gorman 417117c9d12eSMel Gorman chg = to - from; 417217c9d12eSMel Gorman 41735a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 41745a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 41755a6fe125SMel Gorman } 41765a6fe125SMel Gorman 4177c50ac050SDave Hansen if (chg < 0) { 4178c50ac050SDave Hansen ret = chg; 4179c50ac050SDave Hansen goto out_err; 4180c50ac050SDave Hansen } 418117c9d12eSMel Gorman 41821c5ecae3SMike Kravetz /* 41831c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 41841c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 41851c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 41861c5ecae3SMike Kravetz */ 41871c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 41881c5ecae3SMike Kravetz if (gbl_reserve < 0) { 4189c50ac050SDave Hansen ret = -ENOSPC; 4190c50ac050SDave Hansen goto out_err; 4191c50ac050SDave Hansen } 419217c9d12eSMel Gorman 419317c9d12eSMel Gorman /* 419417c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 419590481622SDavid Gibson * Hand the pages back to the subpool if there are not 419617c9d12eSMel Gorman */ 41971c5ecae3SMike Kravetz ret = hugetlb_acct_memory(h, gbl_reserve); 419817c9d12eSMel Gorman if (ret < 0) { 41991c5ecae3SMike Kravetz /* put back original number of pages, chg */ 42001c5ecae3SMike Kravetz (void)hugepage_subpool_put_pages(spool, chg); 4201c50ac050SDave Hansen goto out_err; 420217c9d12eSMel Gorman } 420317c9d12eSMel Gorman 420417c9d12eSMel Gorman /* 420517c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 420617c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 420717c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 420817c9d12eSMel Gorman * the reservation was and the page cache tells how much of 420917c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 421017c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 421117c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 421217c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 421317c9d12eSMel Gorman * else has to be done for private mappings here 421417c9d12eSMel Gorman */ 421533039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 421633039678SMike Kravetz long add = region_add(resv_map, from, to); 421733039678SMike Kravetz 421833039678SMike Kravetz if (unlikely(chg > add)) { 421933039678SMike Kravetz /* 422033039678SMike Kravetz * pages in this range were added to the reserve 422133039678SMike Kravetz * map between region_chg and region_add. This 422233039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 422333039678SMike Kravetz * the subpool and reserve counts modified above 422433039678SMike Kravetz * based on the difference. 422533039678SMike Kravetz */ 422633039678SMike Kravetz long rsv_adjust; 422733039678SMike Kravetz 422833039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 422933039678SMike Kravetz chg - add); 423033039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 423133039678SMike Kravetz } 423233039678SMike Kravetz } 4233a43a8c39SChen, Kenneth W return 0; 4234c50ac050SDave Hansen out_err: 42355e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 42365e911373SMike Kravetz region_abort(resv_map, from, to); 4237f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 4238f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 4239c50ac050SDave Hansen return ret; 4240a43a8c39SChen, Kenneth W } 4241a43a8c39SChen, Kenneth W 4242b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 4243b5cec28dSMike Kravetz long freed) 4244a43a8c39SChen, Kenneth W { 4245a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 42464e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 42479119a41eSJoonsoo Kim long chg = 0; 424890481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 42491c5ecae3SMike Kravetz long gbl_reserve; 425045c682a6SKen Chen 4251b5cec28dSMike Kravetz if (resv_map) { 4252b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 4253b5cec28dSMike Kravetz /* 4254b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 4255b5cec28dSMike Kravetz * must be split and another region descriptor can not be 4256b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 4257b5cec28dSMike Kravetz */ 4258b5cec28dSMike Kravetz if (chg < 0) 4259b5cec28dSMike Kravetz return chg; 4260b5cec28dSMike Kravetz } 4261b5cec28dSMike Kravetz 426245c682a6SKen Chen spin_lock(&inode->i_lock); 4263e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 426445c682a6SKen Chen spin_unlock(&inode->i_lock); 426545c682a6SKen Chen 42661c5ecae3SMike Kravetz /* 42671c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 42681c5ecae3SMike Kravetz * reservations to be released may be adjusted. 42691c5ecae3SMike Kravetz */ 42701c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 42711c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 4272b5cec28dSMike Kravetz 4273b5cec28dSMike Kravetz return 0; 4274a43a8c39SChen, Kenneth W } 427593f70f90SNaoya Horiguchi 42763212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 42773212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 42783212b535SSteve Capper struct vm_area_struct *vma, 42793212b535SSteve Capper unsigned long addr, pgoff_t idx) 42803212b535SSteve Capper { 42813212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 42823212b535SSteve Capper svma->vm_start; 42833212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 42843212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 42853212b535SSteve Capper 42863212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 4287de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 4288de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 42893212b535SSteve Capper 42903212b535SSteve Capper /* 42913212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 42923212b535SSteve Capper * page table page. 42933212b535SSteve Capper */ 42943212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 42953212b535SSteve Capper vm_flags != svm_flags || 42963212b535SSteve Capper sbase < svma->vm_start || svma->vm_end < s_end) 42973212b535SSteve Capper return 0; 42983212b535SSteve Capper 42993212b535SSteve Capper return saddr; 43003212b535SSteve Capper } 43013212b535SSteve Capper 430231aafb45SNicholas Krause static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) 43033212b535SSteve Capper { 43043212b535SSteve Capper unsigned long base = addr & PUD_MASK; 43053212b535SSteve Capper unsigned long end = base + PUD_SIZE; 43063212b535SSteve Capper 43073212b535SSteve Capper /* 43083212b535SSteve Capper * check on proper vm_flags and page table alignment 43093212b535SSteve Capper */ 43103212b535SSteve Capper if (vma->vm_flags & VM_MAYSHARE && 43113212b535SSteve Capper vma->vm_start <= base && end <= vma->vm_end) 431231aafb45SNicholas Krause return true; 431331aafb45SNicholas Krause return false; 43143212b535SSteve Capper } 43153212b535SSteve Capper 43163212b535SSteve Capper /* 43173212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 43183212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 43193212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 43203212b535SSteve Capper * code much cleaner. pmd allocation is essential for the shared case because 4321c8c06efaSDavidlohr Bueso * pud has to be populated inside the same i_mmap_rwsem section - otherwise 43223212b535SSteve Capper * racing tasks could either miss the sharing (see huge_pte_offset) or select a 43233212b535SSteve Capper * bad pmd for sharing. 43243212b535SSteve Capper */ 43253212b535SSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 43263212b535SSteve Capper { 43273212b535SSteve Capper struct vm_area_struct *vma = find_vma(mm, addr); 43283212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 43293212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 43303212b535SSteve Capper vma->vm_pgoff; 43313212b535SSteve Capper struct vm_area_struct *svma; 43323212b535SSteve Capper unsigned long saddr; 43333212b535SSteve Capper pte_t *spte = NULL; 43343212b535SSteve Capper pte_t *pte; 4335cb900f41SKirill A. Shutemov spinlock_t *ptl; 43363212b535SSteve Capper 43373212b535SSteve Capper if (!vma_shareable(vma, addr)) 43383212b535SSteve Capper return (pte_t *)pmd_alloc(mm, pud, addr); 43393212b535SSteve Capper 434083cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 43413212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 43423212b535SSteve Capper if (svma == vma) 43433212b535SSteve Capper continue; 43443212b535SSteve Capper 43453212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 43463212b535SSteve Capper if (saddr) { 43473212b535SSteve Capper spte = huge_pte_offset(svma->vm_mm, saddr); 43483212b535SSteve Capper if (spte) { 43493212b535SSteve Capper get_page(virt_to_page(spte)); 43503212b535SSteve Capper break; 43513212b535SSteve Capper } 43523212b535SSteve Capper } 43533212b535SSteve Capper } 43543212b535SSteve Capper 43553212b535SSteve Capper if (!spte) 43563212b535SSteve Capper goto out; 43573212b535SSteve Capper 43588bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 4359dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 43603212b535SSteve Capper pud_populate(mm, pud, 43613212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 4362c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 4363dc6c9a35SKirill A. Shutemov } else { 43643212b535SSteve Capper put_page(virt_to_page(spte)); 4365dc6c9a35SKirill A. Shutemov } 4366cb900f41SKirill A. Shutemov spin_unlock(ptl); 43673212b535SSteve Capper out: 43683212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 436983cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 43703212b535SSteve Capper return pte; 43713212b535SSteve Capper } 43723212b535SSteve Capper 43733212b535SSteve Capper /* 43743212b535SSteve Capper * unmap huge page backed by shared pte. 43753212b535SSteve Capper * 43763212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 43773212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 43783212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 43793212b535SSteve Capper * 4380cb900f41SKirill A. Shutemov * called with page table lock held. 43813212b535SSteve Capper * 43823212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 43833212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 43843212b535SSteve Capper */ 43853212b535SSteve Capper int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 43863212b535SSteve Capper { 43873212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 43883212b535SSteve Capper pud_t *pud = pud_offset(pgd, *addr); 43893212b535SSteve Capper 43903212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 43913212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 43923212b535SSteve Capper return 0; 43933212b535SSteve Capper 43943212b535SSteve Capper pud_clear(pud); 43953212b535SSteve Capper put_page(virt_to_page(ptep)); 4396dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 43973212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 43983212b535SSteve Capper return 1; 43993212b535SSteve Capper } 44009e5fc74cSSteve Capper #define want_pmd_share() (1) 44019e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 44029e5fc74cSSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 44039e5fc74cSSteve Capper { 44049e5fc74cSSteve Capper return NULL; 44059e5fc74cSSteve Capper } 4406e81f2d22SZhang Zhen 4407e81f2d22SZhang Zhen int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 4408e81f2d22SZhang Zhen { 4409e81f2d22SZhang Zhen return 0; 4410e81f2d22SZhang Zhen } 44119e5fc74cSSteve Capper #define want_pmd_share() (0) 44123212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 44133212b535SSteve Capper 44149e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 44159e5fc74cSSteve Capper pte_t *huge_pte_alloc(struct mm_struct *mm, 44169e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 44179e5fc74cSSteve Capper { 44189e5fc74cSSteve Capper pgd_t *pgd; 44199e5fc74cSSteve Capper pud_t *pud; 44209e5fc74cSSteve Capper pte_t *pte = NULL; 44219e5fc74cSSteve Capper 44229e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 44239e5fc74cSSteve Capper pud = pud_alloc(mm, pgd, addr); 44249e5fc74cSSteve Capper if (pud) { 44259e5fc74cSSteve Capper if (sz == PUD_SIZE) { 44269e5fc74cSSteve Capper pte = (pte_t *)pud; 44279e5fc74cSSteve Capper } else { 44289e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 44299e5fc74cSSteve Capper if (want_pmd_share() && pud_none(*pud)) 44309e5fc74cSSteve Capper pte = huge_pmd_share(mm, addr, pud); 44319e5fc74cSSteve Capper else 44329e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 44339e5fc74cSSteve Capper } 44349e5fc74cSSteve Capper } 44354e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 44369e5fc74cSSteve Capper 44379e5fc74cSSteve Capper return pte; 44389e5fc74cSSteve Capper } 44399e5fc74cSSteve Capper 44409e5fc74cSSteve Capper pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) 44419e5fc74cSSteve Capper { 44429e5fc74cSSteve Capper pgd_t *pgd; 44439e5fc74cSSteve Capper pud_t *pud; 44449e5fc74cSSteve Capper pmd_t *pmd = NULL; 44459e5fc74cSSteve Capper 44469e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 44479e5fc74cSSteve Capper if (pgd_present(*pgd)) { 44489e5fc74cSSteve Capper pud = pud_offset(pgd, addr); 44499e5fc74cSSteve Capper if (pud_present(*pud)) { 44509e5fc74cSSteve Capper if (pud_huge(*pud)) 44519e5fc74cSSteve Capper return (pte_t *)pud; 44529e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 44539e5fc74cSSteve Capper } 44549e5fc74cSSteve Capper } 44559e5fc74cSSteve Capper return (pte_t *) pmd; 44569e5fc74cSSteve Capper } 44579e5fc74cSSteve Capper 445861f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 445961f77edaSNaoya Horiguchi 446061f77edaSNaoya Horiguchi /* 446161f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 446261f77edaSNaoya Horiguchi * behavior. 446361f77edaSNaoya Horiguchi */ 446461f77edaSNaoya Horiguchi struct page * __weak 446561f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 446661f77edaSNaoya Horiguchi int write) 446761f77edaSNaoya Horiguchi { 446861f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 446961f77edaSNaoya Horiguchi } 447061f77edaSNaoya Horiguchi 447161f77edaSNaoya Horiguchi struct page * __weak 44729e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 4473e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 44749e5fc74cSSteve Capper { 4475e66f17ffSNaoya Horiguchi struct page *page = NULL; 4476e66f17ffSNaoya Horiguchi spinlock_t *ptl; 4477e66f17ffSNaoya Horiguchi retry: 4478e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 4479e66f17ffSNaoya Horiguchi spin_lock(ptl); 4480e66f17ffSNaoya Horiguchi /* 4481e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 4482e66f17ffSNaoya Horiguchi * unmapped from other threads. 4483e66f17ffSNaoya Horiguchi */ 4484e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 4485e66f17ffSNaoya Horiguchi goto out; 4486e66f17ffSNaoya Horiguchi if (pmd_present(*pmd)) { 448797534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 4488e66f17ffSNaoya Horiguchi if (flags & FOLL_GET) 4489e66f17ffSNaoya Horiguchi get_page(page); 4490e66f17ffSNaoya Horiguchi } else { 4491e66f17ffSNaoya Horiguchi if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) { 4492e66f17ffSNaoya Horiguchi spin_unlock(ptl); 4493e66f17ffSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pmd, ptl); 4494e66f17ffSNaoya Horiguchi goto retry; 4495e66f17ffSNaoya Horiguchi } 4496e66f17ffSNaoya Horiguchi /* 4497e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 4498e66f17ffSNaoya Horiguchi * follow_page_mask(). 4499e66f17ffSNaoya Horiguchi */ 4500e66f17ffSNaoya Horiguchi } 4501e66f17ffSNaoya Horiguchi out: 4502e66f17ffSNaoya Horiguchi spin_unlock(ptl); 45039e5fc74cSSteve Capper return page; 45049e5fc74cSSteve Capper } 45059e5fc74cSSteve Capper 450661f77edaSNaoya Horiguchi struct page * __weak 45079e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 4508e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 45099e5fc74cSSteve Capper { 4510e66f17ffSNaoya Horiguchi if (flags & FOLL_GET) 4511e66f17ffSNaoya Horiguchi return NULL; 45129e5fc74cSSteve Capper 4513e66f17ffSNaoya Horiguchi return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 45149e5fc74cSSteve Capper } 45159e5fc74cSSteve Capper 4516d5bd9106SAndi Kleen #ifdef CONFIG_MEMORY_FAILURE 4517d5bd9106SAndi Kleen 451893f70f90SNaoya Horiguchi /* 451993f70f90SNaoya Horiguchi * This function is called from memory failure code. 452093f70f90SNaoya Horiguchi */ 45216de2b1aaSNaoya Horiguchi int dequeue_hwpoisoned_huge_page(struct page *hpage) 452293f70f90SNaoya Horiguchi { 452393f70f90SNaoya Horiguchi struct hstate *h = page_hstate(hpage); 452493f70f90SNaoya Horiguchi int nid = page_to_nid(hpage); 45256de2b1aaSNaoya Horiguchi int ret = -EBUSY; 452693f70f90SNaoya Horiguchi 452793f70f90SNaoya Horiguchi spin_lock(&hugetlb_lock); 45287e1f049eSNaoya Horiguchi /* 45297e1f049eSNaoya Horiguchi * Just checking !page_huge_active is not enough, because that could be 45307e1f049eSNaoya Horiguchi * an isolated/hwpoisoned hugepage (which have >0 refcount). 45317e1f049eSNaoya Horiguchi */ 45327e1f049eSNaoya Horiguchi if (!page_huge_active(hpage) && !page_count(hpage)) { 453356f2fb14SNaoya Horiguchi /* 453456f2fb14SNaoya Horiguchi * Hwpoisoned hugepage isn't linked to activelist or freelist, 453556f2fb14SNaoya Horiguchi * but dangling hpage->lru can trigger list-debug warnings 453656f2fb14SNaoya Horiguchi * (this happens when we call unpoison_memory() on it), 453756f2fb14SNaoya Horiguchi * so let it point to itself with list_del_init(). 453856f2fb14SNaoya Horiguchi */ 453956f2fb14SNaoya Horiguchi list_del_init(&hpage->lru); 45408c6c2ecbSNaoya Horiguchi set_page_refcounted(hpage); 454193f70f90SNaoya Horiguchi h->free_huge_pages--; 454293f70f90SNaoya Horiguchi h->free_huge_pages_node[nid]--; 45436de2b1aaSNaoya Horiguchi ret = 0; 454493f70f90SNaoya Horiguchi } 45456de2b1aaSNaoya Horiguchi spin_unlock(&hugetlb_lock); 45466de2b1aaSNaoya Horiguchi return ret; 45476de2b1aaSNaoya Horiguchi } 45486de2b1aaSNaoya Horiguchi #endif 454931caf665SNaoya Horiguchi 455031caf665SNaoya Horiguchi bool isolate_huge_page(struct page *page, struct list_head *list) 455131caf665SNaoya Horiguchi { 4552bcc54222SNaoya Horiguchi bool ret = true; 4553bcc54222SNaoya Horiguchi 4554309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 455531caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 4556bcc54222SNaoya Horiguchi if (!page_huge_active(page) || !get_page_unless_zero(page)) { 4557bcc54222SNaoya Horiguchi ret = false; 4558bcc54222SNaoya Horiguchi goto unlock; 4559bcc54222SNaoya Horiguchi } 4560bcc54222SNaoya Horiguchi clear_page_huge_active(page); 456131caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 4562bcc54222SNaoya Horiguchi unlock: 456331caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 4564bcc54222SNaoya Horiguchi return ret; 456531caf665SNaoya Horiguchi } 456631caf665SNaoya Horiguchi 456731caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 456831caf665SNaoya Horiguchi { 4569309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 457031caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 4571bcc54222SNaoya Horiguchi set_page_huge_active(page); 457231caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 457331caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 457431caf665SNaoya Horiguchi put_page(page); 457531caf665SNaoya Horiguchi } 4576