11da177e4SLinus Torvalds /* 21da177e4SLinus Torvalds * Generic hugetlb support. 36d49e352SNadia Yvette Chambers * (C) Nadia Yvette Chambers, April 2004 41da177e4SLinus Torvalds */ 51da177e4SLinus Torvalds #include <linux/list.h> 61da177e4SLinus Torvalds #include <linux/init.h> 71da177e4SLinus Torvalds #include <linux/mm.h> 8e1759c21SAlexey Dobriyan #include <linux/seq_file.h> 91da177e4SLinus Torvalds #include <linux/sysctl.h> 101da177e4SLinus Torvalds #include <linux/highmem.h> 11cddb8a5cSAndrea Arcangeli #include <linux/mmu_notifier.h> 121da177e4SLinus Torvalds #include <linux/nodemask.h> 1363551ae0SDavid Gibson #include <linux/pagemap.h> 145da7ca86SChristoph Lameter #include <linux/mempolicy.h> 153b32123dSGideon Israel Dsouza #include <linux/compiler.h> 16aea47ff3SChristoph Lameter #include <linux/cpuset.h> 173935baa9SDavid Gibson #include <linux/mutex.h> 18aa888a74SAndi Kleen #include <linux/bootmem.h> 19a3437870SNishanth Aravamudan #include <linux/sysfs.h> 205a0e3ad6STejun Heo #include <linux/slab.h> 21174cd4b1SIngo Molnar #include <linux/sched/signal.h> 220fe6e20bSNaoya Horiguchi #include <linux/rmap.h> 23c6247f72SMatthew Wilcox #include <linux/string_helpers.h> 24fd6a03edSNaoya Horiguchi #include <linux/swap.h> 25fd6a03edSNaoya Horiguchi #include <linux/swapops.h> 268382d914SDavidlohr Bueso #include <linux/jhash.h> 27d6606683SLinus Torvalds 2863551ae0SDavid Gibson #include <asm/page.h> 2963551ae0SDavid Gibson #include <asm/pgtable.h> 3024669e58SAneesh Kumar K.V #include <asm/tlb.h> 3163551ae0SDavid Gibson 3224669e58SAneesh Kumar K.V #include <linux/io.h> 3363551ae0SDavid Gibson #include <linux/hugetlb.h> 349dd540e2SAneesh Kumar K.V #include <linux/hugetlb_cgroup.h> 359a305230SLee Schermerhorn #include <linux/node.h> 361a1aad8aSMike Kravetz #include <linux/userfaultfd_k.h> 377835e98bSNick Piggin #include "internal.h" 381da177e4SLinus Torvalds 39753162cdSAndrey Ryabinin int hugepages_treat_as_movable; 40a5516438SAndi Kleen 41c3f38a38SAneesh Kumar K.V int hugetlb_max_hstate __read_mostly; 42e5ff2159SAndi Kleen unsigned int default_hstate_idx; 43e5ff2159SAndi Kleen struct hstate hstates[HUGE_MAX_HSTATE]; 44641844f5SNaoya Horiguchi /* 45641844f5SNaoya Horiguchi * Minimum page order among possible hugepage sizes, set to a proper value 46641844f5SNaoya Horiguchi * at boot time. 47641844f5SNaoya Horiguchi */ 48641844f5SNaoya Horiguchi static unsigned int minimum_order __read_mostly = UINT_MAX; 49e5ff2159SAndi Kleen 5053ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 5153ba51d2SJon Tollefson 52e5ff2159SAndi Kleen /* for command line parsing */ 53e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 54e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 55e11bfbfcSNick Piggin static unsigned long __initdata default_hstate_size; 569fee021dSVaishali Thakkar static bool __initdata parsed_valid_hugepagesz = true; 57e5ff2159SAndi Kleen 583935baa9SDavid Gibson /* 5931caf665SNaoya Horiguchi * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, 6031caf665SNaoya Horiguchi * free_huge_pages, and surplus_huge_pages. 613935baa9SDavid Gibson */ 62c3f38a38SAneesh Kumar K.V DEFINE_SPINLOCK(hugetlb_lock); 630bd0f9fbSEric Paris 648382d914SDavidlohr Bueso /* 658382d914SDavidlohr Bueso * Serializes faults on the same logical page. This is used to 668382d914SDavidlohr Bueso * prevent spurious OOMs when the hugepage pool is fully utilized. 678382d914SDavidlohr Bueso */ 688382d914SDavidlohr Bueso static int num_fault_mutexes; 69c672c7f2SMike Kravetz struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; 708382d914SDavidlohr Bueso 717ca02d0aSMike Kravetz /* Forward declaration */ 727ca02d0aSMike Kravetz static int hugetlb_acct_memory(struct hstate *h, long delta); 737ca02d0aSMike Kravetz 7490481622SDavid Gibson static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) 7590481622SDavid Gibson { 7690481622SDavid Gibson bool free = (spool->count == 0) && (spool->used_hpages == 0); 7790481622SDavid Gibson 7890481622SDavid Gibson spin_unlock(&spool->lock); 7990481622SDavid Gibson 8090481622SDavid Gibson /* If no pages are used, and no other handles to the subpool 817ca02d0aSMike Kravetz * remain, give up any reservations mased on minimum size and 827ca02d0aSMike Kravetz * free the subpool */ 837ca02d0aSMike Kravetz if (free) { 847ca02d0aSMike Kravetz if (spool->min_hpages != -1) 857ca02d0aSMike Kravetz hugetlb_acct_memory(spool->hstate, 867ca02d0aSMike Kravetz -spool->min_hpages); 8790481622SDavid Gibson kfree(spool); 8890481622SDavid Gibson } 897ca02d0aSMike Kravetz } 9090481622SDavid Gibson 917ca02d0aSMike Kravetz struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, 927ca02d0aSMike Kravetz long min_hpages) 9390481622SDavid Gibson { 9490481622SDavid Gibson struct hugepage_subpool *spool; 9590481622SDavid Gibson 96c6a91820SMike Kravetz spool = kzalloc(sizeof(*spool), GFP_KERNEL); 9790481622SDavid Gibson if (!spool) 9890481622SDavid Gibson return NULL; 9990481622SDavid Gibson 10090481622SDavid Gibson spin_lock_init(&spool->lock); 10190481622SDavid Gibson spool->count = 1; 1027ca02d0aSMike Kravetz spool->max_hpages = max_hpages; 1037ca02d0aSMike Kravetz spool->hstate = h; 1047ca02d0aSMike Kravetz spool->min_hpages = min_hpages; 1057ca02d0aSMike Kravetz 1067ca02d0aSMike Kravetz if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { 1077ca02d0aSMike Kravetz kfree(spool); 1087ca02d0aSMike Kravetz return NULL; 1097ca02d0aSMike Kravetz } 1107ca02d0aSMike Kravetz spool->rsv_hpages = min_hpages; 11190481622SDavid Gibson 11290481622SDavid Gibson return spool; 11390481622SDavid Gibson } 11490481622SDavid Gibson 11590481622SDavid Gibson void hugepage_put_subpool(struct hugepage_subpool *spool) 11690481622SDavid Gibson { 11790481622SDavid Gibson spin_lock(&spool->lock); 11890481622SDavid Gibson BUG_ON(!spool->count); 11990481622SDavid Gibson spool->count--; 12090481622SDavid Gibson unlock_or_release_subpool(spool); 12190481622SDavid Gibson } 12290481622SDavid Gibson 1231c5ecae3SMike Kravetz /* 1241c5ecae3SMike Kravetz * Subpool accounting for allocating and reserving pages. 1251c5ecae3SMike Kravetz * Return -ENOMEM if there are not enough resources to satisfy the 1261c5ecae3SMike Kravetz * the request. Otherwise, return the number of pages by which the 1271c5ecae3SMike Kravetz * global pools must be adjusted (upward). The returned value may 1281c5ecae3SMike Kravetz * only be different than the passed value (delta) in the case where 1291c5ecae3SMike Kravetz * a subpool minimum size must be manitained. 1301c5ecae3SMike Kravetz */ 1311c5ecae3SMike Kravetz static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, 13290481622SDavid Gibson long delta) 13390481622SDavid Gibson { 1341c5ecae3SMike Kravetz long ret = delta; 13590481622SDavid Gibson 13690481622SDavid Gibson if (!spool) 1371c5ecae3SMike Kravetz return ret; 13890481622SDavid Gibson 13990481622SDavid Gibson spin_lock(&spool->lock); 14090481622SDavid Gibson 1411c5ecae3SMike Kravetz if (spool->max_hpages != -1) { /* maximum size accounting */ 1421c5ecae3SMike Kravetz if ((spool->used_hpages + delta) <= spool->max_hpages) 1431c5ecae3SMike Kravetz spool->used_hpages += delta; 1441c5ecae3SMike Kravetz else { 1451c5ecae3SMike Kravetz ret = -ENOMEM; 1461c5ecae3SMike Kravetz goto unlock_ret; 1471c5ecae3SMike Kravetz } 1481c5ecae3SMike Kravetz } 1491c5ecae3SMike Kravetz 15009a95e29SMike Kravetz /* minimum size accounting */ 15109a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->rsv_hpages) { 1521c5ecae3SMike Kravetz if (delta > spool->rsv_hpages) { 1531c5ecae3SMike Kravetz /* 1541c5ecae3SMike Kravetz * Asking for more reserves than those already taken on 1551c5ecae3SMike Kravetz * behalf of subpool. Return difference. 1561c5ecae3SMike Kravetz */ 1571c5ecae3SMike Kravetz ret = delta - spool->rsv_hpages; 1581c5ecae3SMike Kravetz spool->rsv_hpages = 0; 1591c5ecae3SMike Kravetz } else { 1601c5ecae3SMike Kravetz ret = 0; /* reserves already accounted for */ 1611c5ecae3SMike Kravetz spool->rsv_hpages -= delta; 1621c5ecae3SMike Kravetz } 1631c5ecae3SMike Kravetz } 1641c5ecae3SMike Kravetz 1651c5ecae3SMike Kravetz unlock_ret: 1661c5ecae3SMike Kravetz spin_unlock(&spool->lock); 16790481622SDavid Gibson return ret; 16890481622SDavid Gibson } 16990481622SDavid Gibson 1701c5ecae3SMike Kravetz /* 1711c5ecae3SMike Kravetz * Subpool accounting for freeing and unreserving pages. 1721c5ecae3SMike Kravetz * Return the number of global page reservations that must be dropped. 1731c5ecae3SMike Kravetz * The return value may only be different than the passed value (delta) 1741c5ecae3SMike Kravetz * in the case where a subpool minimum size must be maintained. 1751c5ecae3SMike Kravetz */ 1761c5ecae3SMike Kravetz static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, 17790481622SDavid Gibson long delta) 17890481622SDavid Gibson { 1791c5ecae3SMike Kravetz long ret = delta; 1801c5ecae3SMike Kravetz 18190481622SDavid Gibson if (!spool) 1821c5ecae3SMike Kravetz return delta; 18390481622SDavid Gibson 18490481622SDavid Gibson spin_lock(&spool->lock); 1851c5ecae3SMike Kravetz 1861c5ecae3SMike Kravetz if (spool->max_hpages != -1) /* maximum size accounting */ 18790481622SDavid Gibson spool->used_hpages -= delta; 1881c5ecae3SMike Kravetz 18909a95e29SMike Kravetz /* minimum size accounting */ 19009a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { 1911c5ecae3SMike Kravetz if (spool->rsv_hpages + delta <= spool->min_hpages) 1921c5ecae3SMike Kravetz ret = 0; 1931c5ecae3SMike Kravetz else 1941c5ecae3SMike Kravetz ret = spool->rsv_hpages + delta - spool->min_hpages; 1951c5ecae3SMike Kravetz 1961c5ecae3SMike Kravetz spool->rsv_hpages += delta; 1971c5ecae3SMike Kravetz if (spool->rsv_hpages > spool->min_hpages) 1981c5ecae3SMike Kravetz spool->rsv_hpages = spool->min_hpages; 1991c5ecae3SMike Kravetz } 2001c5ecae3SMike Kravetz 2011c5ecae3SMike Kravetz /* 2021c5ecae3SMike Kravetz * If hugetlbfs_put_super couldn't free spool due to an outstanding 2031c5ecae3SMike Kravetz * quota reference, free it now. 2041c5ecae3SMike Kravetz */ 20590481622SDavid Gibson unlock_or_release_subpool(spool); 2061c5ecae3SMike Kravetz 2071c5ecae3SMike Kravetz return ret; 20890481622SDavid Gibson } 20990481622SDavid Gibson 21090481622SDavid Gibson static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 21190481622SDavid Gibson { 21290481622SDavid Gibson return HUGETLBFS_SB(inode->i_sb)->spool; 21390481622SDavid Gibson } 21490481622SDavid Gibson 21590481622SDavid Gibson static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 21690481622SDavid Gibson { 217496ad9aaSAl Viro return subpool_inode(file_inode(vma->vm_file)); 21890481622SDavid Gibson } 21990481622SDavid Gibson 220e7c4b0bfSAndy Whitcroft /* 22196822904SAndy Whitcroft * Region tracking -- allows tracking of reservations and instantiated pages 22296822904SAndy Whitcroft * across the pages in a mapping. 22384afd99bSAndy Whitcroft * 2241dd308a7SMike Kravetz * The region data structures are embedded into a resv_map and protected 2251dd308a7SMike Kravetz * by a resv_map's lock. The set of regions within the resv_map represent 2261dd308a7SMike Kravetz * reservations for huge pages, or huge pages that have already been 2271dd308a7SMike Kravetz * instantiated within the map. The from and to elements are huge page 2281dd308a7SMike Kravetz * indicies into the associated mapping. from indicates the starting index 2291dd308a7SMike Kravetz * of the region. to represents the first index past the end of the region. 2301dd308a7SMike Kravetz * 2311dd308a7SMike Kravetz * For example, a file region structure with from == 0 and to == 4 represents 2321dd308a7SMike Kravetz * four huge pages in a mapping. It is important to note that the to element 2331dd308a7SMike Kravetz * represents the first element past the end of the region. This is used in 2341dd308a7SMike Kravetz * arithmetic as 4(to) - 0(from) = 4 huge pages in the region. 2351dd308a7SMike Kravetz * 2361dd308a7SMike Kravetz * Interval notation of the form [from, to) will be used to indicate that 2371dd308a7SMike Kravetz * the endpoint from is inclusive and to is exclusive. 23896822904SAndy Whitcroft */ 23996822904SAndy Whitcroft struct file_region { 24096822904SAndy Whitcroft struct list_head link; 24196822904SAndy Whitcroft long from; 24296822904SAndy Whitcroft long to; 24396822904SAndy Whitcroft }; 24496822904SAndy Whitcroft 2451dd308a7SMike Kravetz /* 2461dd308a7SMike Kravetz * Add the huge page range represented by [f, t) to the reserve 2475e911373SMike Kravetz * map. In the normal case, existing regions will be expanded 2485e911373SMike Kravetz * to accommodate the specified range. Sufficient regions should 2495e911373SMike Kravetz * exist for expansion due to the previous call to region_chg 2505e911373SMike Kravetz * with the same range. However, it is possible that region_del 2515e911373SMike Kravetz * could have been called after region_chg and modifed the map 2525e911373SMike Kravetz * in such a way that no region exists to be expanded. In this 2535e911373SMike Kravetz * case, pull a region descriptor from the cache associated with 2545e911373SMike Kravetz * the map and use that for the new range. 255cf3ad20bSMike Kravetz * 256cf3ad20bSMike Kravetz * Return the number of new huge pages added to the map. This 257cf3ad20bSMike Kravetz * number is greater than or equal to zero. 2581dd308a7SMike Kravetz */ 2591406ec9bSJoonsoo Kim static long region_add(struct resv_map *resv, long f, long t) 26096822904SAndy Whitcroft { 2611406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 26296822904SAndy Whitcroft struct file_region *rg, *nrg, *trg; 263cf3ad20bSMike Kravetz long add = 0; 26496822904SAndy Whitcroft 2657b24d861SDavidlohr Bueso spin_lock(&resv->lock); 26696822904SAndy Whitcroft /* Locate the region we are either in or before. */ 26796822904SAndy Whitcroft list_for_each_entry(rg, head, link) 26896822904SAndy Whitcroft if (f <= rg->to) 26996822904SAndy Whitcroft break; 27096822904SAndy Whitcroft 2715e911373SMike Kravetz /* 2725e911373SMike Kravetz * If no region exists which can be expanded to include the 2735e911373SMike Kravetz * specified range, the list must have been modified by an 2745e911373SMike Kravetz * interleving call to region_del(). Pull a region descriptor 2755e911373SMike Kravetz * from the cache and use it for this range. 2765e911373SMike Kravetz */ 2775e911373SMike Kravetz if (&rg->link == head || t < rg->from) { 2785e911373SMike Kravetz VM_BUG_ON(resv->region_cache_count <= 0); 2795e911373SMike Kravetz 2805e911373SMike Kravetz resv->region_cache_count--; 2815e911373SMike Kravetz nrg = list_first_entry(&resv->region_cache, struct file_region, 2825e911373SMike Kravetz link); 2835e911373SMike Kravetz list_del(&nrg->link); 2845e911373SMike Kravetz 2855e911373SMike Kravetz nrg->from = f; 2865e911373SMike Kravetz nrg->to = t; 2875e911373SMike Kravetz list_add(&nrg->link, rg->link.prev); 2885e911373SMike Kravetz 2895e911373SMike Kravetz add += t - f; 2905e911373SMike Kravetz goto out_locked; 2915e911373SMike Kravetz } 2925e911373SMike Kravetz 29396822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 29496822904SAndy Whitcroft if (f > rg->from) 29596822904SAndy Whitcroft f = rg->from; 29696822904SAndy Whitcroft 29796822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 29896822904SAndy Whitcroft nrg = rg; 29996822904SAndy Whitcroft list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 30096822904SAndy Whitcroft if (&rg->link == head) 30196822904SAndy Whitcroft break; 30296822904SAndy Whitcroft if (rg->from > t) 30396822904SAndy Whitcroft break; 30496822904SAndy Whitcroft 30596822904SAndy Whitcroft /* If this area reaches higher then extend our area to 30696822904SAndy Whitcroft * include it completely. If this is not the first area 30796822904SAndy Whitcroft * which we intend to reuse, free it. */ 30896822904SAndy Whitcroft if (rg->to > t) 30996822904SAndy Whitcroft t = rg->to; 31096822904SAndy Whitcroft if (rg != nrg) { 311cf3ad20bSMike Kravetz /* Decrement return value by the deleted range. 312cf3ad20bSMike Kravetz * Another range will span this area so that by 313cf3ad20bSMike Kravetz * end of routine add will be >= zero 314cf3ad20bSMike Kravetz */ 315cf3ad20bSMike Kravetz add -= (rg->to - rg->from); 31696822904SAndy Whitcroft list_del(&rg->link); 31796822904SAndy Whitcroft kfree(rg); 31896822904SAndy Whitcroft } 31996822904SAndy Whitcroft } 320cf3ad20bSMike Kravetz 321cf3ad20bSMike Kravetz add += (nrg->from - f); /* Added to beginning of region */ 32296822904SAndy Whitcroft nrg->from = f; 323cf3ad20bSMike Kravetz add += t - nrg->to; /* Added to end of region */ 32496822904SAndy Whitcroft nrg->to = t; 325cf3ad20bSMike Kravetz 3265e911373SMike Kravetz out_locked: 3275e911373SMike Kravetz resv->adds_in_progress--; 3287b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 329cf3ad20bSMike Kravetz VM_BUG_ON(add < 0); 330cf3ad20bSMike Kravetz return add; 33196822904SAndy Whitcroft } 33296822904SAndy Whitcroft 3331dd308a7SMike Kravetz /* 3341dd308a7SMike Kravetz * Examine the existing reserve map and determine how many 3351dd308a7SMike Kravetz * huge pages in the specified range [f, t) are NOT currently 3361dd308a7SMike Kravetz * represented. This routine is called before a subsequent 3371dd308a7SMike Kravetz * call to region_add that will actually modify the reserve 3381dd308a7SMike Kravetz * map to add the specified range [f, t). region_chg does 3391dd308a7SMike Kravetz * not change the number of huge pages represented by the 3401dd308a7SMike Kravetz * map. However, if the existing regions in the map can not 3411dd308a7SMike Kravetz * be expanded to represent the new range, a new file_region 3421dd308a7SMike Kravetz * structure is added to the map as a placeholder. This is 3431dd308a7SMike Kravetz * so that the subsequent region_add call will have all the 3441dd308a7SMike Kravetz * regions it needs and will not fail. 3451dd308a7SMike Kravetz * 3465e911373SMike Kravetz * Upon entry, region_chg will also examine the cache of region descriptors 3475e911373SMike Kravetz * associated with the map. If there are not enough descriptors cached, one 3485e911373SMike Kravetz * will be allocated for the in progress add operation. 3495e911373SMike Kravetz * 3505e911373SMike Kravetz * Returns the number of huge pages that need to be added to the existing 3515e911373SMike Kravetz * reservation map for the range [f, t). This number is greater or equal to 3525e911373SMike Kravetz * zero. -ENOMEM is returned if a new file_region structure or cache entry 3535e911373SMike Kravetz * is needed and can not be allocated. 3541dd308a7SMike Kravetz */ 3551406ec9bSJoonsoo Kim static long region_chg(struct resv_map *resv, long f, long t) 35696822904SAndy Whitcroft { 3571406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 3587b24d861SDavidlohr Bueso struct file_region *rg, *nrg = NULL; 35996822904SAndy Whitcroft long chg = 0; 36096822904SAndy Whitcroft 3617b24d861SDavidlohr Bueso retry: 3627b24d861SDavidlohr Bueso spin_lock(&resv->lock); 3635e911373SMike Kravetz retry_locked: 3645e911373SMike Kravetz resv->adds_in_progress++; 3655e911373SMike Kravetz 3665e911373SMike Kravetz /* 3675e911373SMike Kravetz * Check for sufficient descriptors in the cache to accommodate 3685e911373SMike Kravetz * the number of in progress add operations. 3695e911373SMike Kravetz */ 3705e911373SMike Kravetz if (resv->adds_in_progress > resv->region_cache_count) { 3715e911373SMike Kravetz struct file_region *trg; 3725e911373SMike Kravetz 3735e911373SMike Kravetz VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1); 3745e911373SMike Kravetz /* Must drop lock to allocate a new descriptor. */ 3755e911373SMike Kravetz resv->adds_in_progress--; 3765e911373SMike Kravetz spin_unlock(&resv->lock); 3775e911373SMike Kravetz 3785e911373SMike Kravetz trg = kmalloc(sizeof(*trg), GFP_KERNEL); 379dbe409e4SMike Kravetz if (!trg) { 380dbe409e4SMike Kravetz kfree(nrg); 3815e911373SMike Kravetz return -ENOMEM; 382dbe409e4SMike Kravetz } 3835e911373SMike Kravetz 3845e911373SMike Kravetz spin_lock(&resv->lock); 3855e911373SMike Kravetz list_add(&trg->link, &resv->region_cache); 3865e911373SMike Kravetz resv->region_cache_count++; 3875e911373SMike Kravetz goto retry_locked; 3885e911373SMike Kravetz } 3895e911373SMike Kravetz 39096822904SAndy Whitcroft /* Locate the region we are before or in. */ 39196822904SAndy Whitcroft list_for_each_entry(rg, head, link) 39296822904SAndy Whitcroft if (f <= rg->to) 39396822904SAndy Whitcroft break; 39496822904SAndy Whitcroft 39596822904SAndy Whitcroft /* If we are below the current region then a new region is required. 39696822904SAndy Whitcroft * Subtle, allocate a new region at the position but make it zero 39796822904SAndy Whitcroft * size such that we can guarantee to record the reservation. */ 39896822904SAndy Whitcroft if (&rg->link == head || t < rg->from) { 3997b24d861SDavidlohr Bueso if (!nrg) { 4005e911373SMike Kravetz resv->adds_in_progress--; 4017b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 40296822904SAndy Whitcroft nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 40396822904SAndy Whitcroft if (!nrg) 40496822904SAndy Whitcroft return -ENOMEM; 4057b24d861SDavidlohr Bueso 40696822904SAndy Whitcroft nrg->from = f; 40796822904SAndy Whitcroft nrg->to = f; 40896822904SAndy Whitcroft INIT_LIST_HEAD(&nrg->link); 4097b24d861SDavidlohr Bueso goto retry; 4107b24d861SDavidlohr Bueso } 41196822904SAndy Whitcroft 4127b24d861SDavidlohr Bueso list_add(&nrg->link, rg->link.prev); 4137b24d861SDavidlohr Bueso chg = t - f; 4147b24d861SDavidlohr Bueso goto out_nrg; 41596822904SAndy Whitcroft } 41696822904SAndy Whitcroft 41796822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 41896822904SAndy Whitcroft if (f > rg->from) 41996822904SAndy Whitcroft f = rg->from; 42096822904SAndy Whitcroft chg = t - f; 42196822904SAndy Whitcroft 42296822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 42396822904SAndy Whitcroft list_for_each_entry(rg, rg->link.prev, link) { 42496822904SAndy Whitcroft if (&rg->link == head) 42596822904SAndy Whitcroft break; 42696822904SAndy Whitcroft if (rg->from > t) 4277b24d861SDavidlohr Bueso goto out; 42896822904SAndy Whitcroft 42925985edcSLucas De Marchi /* We overlap with this area, if it extends further than 43096822904SAndy Whitcroft * us then we must extend ourselves. Account for its 43196822904SAndy Whitcroft * existing reservation. */ 43296822904SAndy Whitcroft if (rg->to > t) { 43396822904SAndy Whitcroft chg += rg->to - t; 43496822904SAndy Whitcroft t = rg->to; 43596822904SAndy Whitcroft } 43696822904SAndy Whitcroft chg -= rg->to - rg->from; 43796822904SAndy Whitcroft } 4387b24d861SDavidlohr Bueso 4397b24d861SDavidlohr Bueso out: 4407b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 4417b24d861SDavidlohr Bueso /* We already know we raced and no longer need the new region */ 4427b24d861SDavidlohr Bueso kfree(nrg); 4437b24d861SDavidlohr Bueso return chg; 4447b24d861SDavidlohr Bueso out_nrg: 4457b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 44696822904SAndy Whitcroft return chg; 44796822904SAndy Whitcroft } 44896822904SAndy Whitcroft 4491dd308a7SMike Kravetz /* 4505e911373SMike Kravetz * Abort the in progress add operation. The adds_in_progress field 4515e911373SMike Kravetz * of the resv_map keeps track of the operations in progress between 4525e911373SMike Kravetz * calls to region_chg and region_add. Operations are sometimes 4535e911373SMike Kravetz * aborted after the call to region_chg. In such cases, region_abort 4545e911373SMike Kravetz * is called to decrement the adds_in_progress counter. 4555e911373SMike Kravetz * 4565e911373SMike Kravetz * NOTE: The range arguments [f, t) are not needed or used in this 4575e911373SMike Kravetz * routine. They are kept to make reading the calling code easier as 4585e911373SMike Kravetz * arguments will match the associated region_chg call. 4595e911373SMike Kravetz */ 4605e911373SMike Kravetz static void region_abort(struct resv_map *resv, long f, long t) 4615e911373SMike Kravetz { 4625e911373SMike Kravetz spin_lock(&resv->lock); 4635e911373SMike Kravetz VM_BUG_ON(!resv->region_cache_count); 4645e911373SMike Kravetz resv->adds_in_progress--; 4655e911373SMike Kravetz spin_unlock(&resv->lock); 4665e911373SMike Kravetz } 4675e911373SMike Kravetz 4685e911373SMike Kravetz /* 469feba16e2SMike Kravetz * Delete the specified range [f, t) from the reserve map. If the 470feba16e2SMike Kravetz * t parameter is LONG_MAX, this indicates that ALL regions after f 471feba16e2SMike Kravetz * should be deleted. Locate the regions which intersect [f, t) 472feba16e2SMike Kravetz * and either trim, delete or split the existing regions. 473feba16e2SMike Kravetz * 474feba16e2SMike Kravetz * Returns the number of huge pages deleted from the reserve map. 475feba16e2SMike Kravetz * In the normal case, the return value is zero or more. In the 476feba16e2SMike Kravetz * case where a region must be split, a new region descriptor must 477feba16e2SMike Kravetz * be allocated. If the allocation fails, -ENOMEM will be returned. 478feba16e2SMike Kravetz * NOTE: If the parameter t == LONG_MAX, then we will never split 479feba16e2SMike Kravetz * a region and possibly return -ENOMEM. Callers specifying 480feba16e2SMike Kravetz * t == LONG_MAX do not need to check for -ENOMEM error. 4811dd308a7SMike Kravetz */ 482feba16e2SMike Kravetz static long region_del(struct resv_map *resv, long f, long t) 48396822904SAndy Whitcroft { 4841406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 48596822904SAndy Whitcroft struct file_region *rg, *trg; 486feba16e2SMike Kravetz struct file_region *nrg = NULL; 487feba16e2SMike Kravetz long del = 0; 48896822904SAndy Whitcroft 489feba16e2SMike Kravetz retry: 4907b24d861SDavidlohr Bueso spin_lock(&resv->lock); 491feba16e2SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 492dbe409e4SMike Kravetz /* 493dbe409e4SMike Kravetz * Skip regions before the range to be deleted. file_region 494dbe409e4SMike Kravetz * ranges are normally of the form [from, to). However, there 495dbe409e4SMike Kravetz * may be a "placeholder" entry in the map which is of the form 496dbe409e4SMike Kravetz * (from, to) with from == to. Check for placeholder entries 497dbe409e4SMike Kravetz * at the beginning of the range to be deleted. 498dbe409e4SMike Kravetz */ 499dbe409e4SMike Kravetz if (rg->to <= f && (rg->to != rg->from || rg->to != f)) 500feba16e2SMike Kravetz continue; 501dbe409e4SMike Kravetz 502feba16e2SMike Kravetz if (rg->from >= t) 50396822904SAndy Whitcroft break; 50496822904SAndy Whitcroft 505feba16e2SMike Kravetz if (f > rg->from && t < rg->to) { /* Must split region */ 506feba16e2SMike Kravetz /* 507feba16e2SMike Kravetz * Check for an entry in the cache before dropping 508feba16e2SMike Kravetz * lock and attempting allocation. 509feba16e2SMike Kravetz */ 510feba16e2SMike Kravetz if (!nrg && 511feba16e2SMike Kravetz resv->region_cache_count > resv->adds_in_progress) { 512feba16e2SMike Kravetz nrg = list_first_entry(&resv->region_cache, 513feba16e2SMike Kravetz struct file_region, 514feba16e2SMike Kravetz link); 515feba16e2SMike Kravetz list_del(&nrg->link); 516feba16e2SMike Kravetz resv->region_cache_count--; 51796822904SAndy Whitcroft } 51896822904SAndy Whitcroft 519feba16e2SMike Kravetz if (!nrg) { 520feba16e2SMike Kravetz spin_unlock(&resv->lock); 521feba16e2SMike Kravetz nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 522feba16e2SMike Kravetz if (!nrg) 523feba16e2SMike Kravetz return -ENOMEM; 524feba16e2SMike Kravetz goto retry; 525feba16e2SMike Kravetz } 526feba16e2SMike Kravetz 527feba16e2SMike Kravetz del += t - f; 528feba16e2SMike Kravetz 529feba16e2SMike Kravetz /* New entry for end of split region */ 530feba16e2SMike Kravetz nrg->from = t; 531feba16e2SMike Kravetz nrg->to = rg->to; 532feba16e2SMike Kravetz INIT_LIST_HEAD(&nrg->link); 533feba16e2SMike Kravetz 534feba16e2SMike Kravetz /* Original entry is trimmed */ 535feba16e2SMike Kravetz rg->to = f; 536feba16e2SMike Kravetz 537feba16e2SMike Kravetz list_add(&nrg->link, &rg->link); 538feba16e2SMike Kravetz nrg = NULL; 53996822904SAndy Whitcroft break; 540feba16e2SMike Kravetz } 541feba16e2SMike Kravetz 542feba16e2SMike Kravetz if (f <= rg->from && t >= rg->to) { /* Remove entire region */ 543feba16e2SMike Kravetz del += rg->to - rg->from; 54496822904SAndy Whitcroft list_del(&rg->link); 54596822904SAndy Whitcroft kfree(rg); 546feba16e2SMike Kravetz continue; 54796822904SAndy Whitcroft } 5487b24d861SDavidlohr Bueso 549feba16e2SMike Kravetz if (f <= rg->from) { /* Trim beginning of region */ 550feba16e2SMike Kravetz del += t - rg->from; 551feba16e2SMike Kravetz rg->from = t; 552feba16e2SMike Kravetz } else { /* Trim end of region */ 553feba16e2SMike Kravetz del += rg->to - f; 554feba16e2SMike Kravetz rg->to = f; 555feba16e2SMike Kravetz } 556feba16e2SMike Kravetz } 557feba16e2SMike Kravetz 5587b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 559feba16e2SMike Kravetz kfree(nrg); 560feba16e2SMike Kravetz return del; 56196822904SAndy Whitcroft } 56296822904SAndy Whitcroft 5631dd308a7SMike Kravetz /* 564b5cec28dSMike Kravetz * A rare out of memory error was encountered which prevented removal of 565b5cec28dSMike Kravetz * the reserve map region for a page. The huge page itself was free'ed 566b5cec28dSMike Kravetz * and removed from the page cache. This routine will adjust the subpool 567b5cec28dSMike Kravetz * usage count, and the global reserve count if needed. By incrementing 568b5cec28dSMike Kravetz * these counts, the reserve map entry which could not be deleted will 569b5cec28dSMike Kravetz * appear as a "reserved" entry instead of simply dangling with incorrect 570b5cec28dSMike Kravetz * counts. 571b5cec28dSMike Kravetz */ 57272e2936cSzhong jiang void hugetlb_fix_reserve_counts(struct inode *inode) 573b5cec28dSMike Kravetz { 574b5cec28dSMike Kravetz struct hugepage_subpool *spool = subpool_inode(inode); 575b5cec28dSMike Kravetz long rsv_adjust; 576b5cec28dSMike Kravetz 577b5cec28dSMike Kravetz rsv_adjust = hugepage_subpool_get_pages(spool, 1); 57872e2936cSzhong jiang if (rsv_adjust) { 579b5cec28dSMike Kravetz struct hstate *h = hstate_inode(inode); 580b5cec28dSMike Kravetz 581b5cec28dSMike Kravetz hugetlb_acct_memory(h, 1); 582b5cec28dSMike Kravetz } 583b5cec28dSMike Kravetz } 584b5cec28dSMike Kravetz 585b5cec28dSMike Kravetz /* 5861dd308a7SMike Kravetz * Count and return the number of huge pages in the reserve map 5871dd308a7SMike Kravetz * that intersect with the range [f, t). 5881dd308a7SMike Kravetz */ 5891406ec9bSJoonsoo Kim static long region_count(struct resv_map *resv, long f, long t) 59084afd99bSAndy Whitcroft { 5911406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 59284afd99bSAndy Whitcroft struct file_region *rg; 59384afd99bSAndy Whitcroft long chg = 0; 59484afd99bSAndy Whitcroft 5957b24d861SDavidlohr Bueso spin_lock(&resv->lock); 59684afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 59784afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 598f2135a4aSWang Sheng-Hui long seg_from; 599f2135a4aSWang Sheng-Hui long seg_to; 60084afd99bSAndy Whitcroft 60184afd99bSAndy Whitcroft if (rg->to <= f) 60284afd99bSAndy Whitcroft continue; 60384afd99bSAndy Whitcroft if (rg->from >= t) 60484afd99bSAndy Whitcroft break; 60584afd99bSAndy Whitcroft 60684afd99bSAndy Whitcroft seg_from = max(rg->from, f); 60784afd99bSAndy Whitcroft seg_to = min(rg->to, t); 60884afd99bSAndy Whitcroft 60984afd99bSAndy Whitcroft chg += seg_to - seg_from; 61084afd99bSAndy Whitcroft } 6117b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 61284afd99bSAndy Whitcroft 61384afd99bSAndy Whitcroft return chg; 61484afd99bSAndy Whitcroft } 61584afd99bSAndy Whitcroft 61696822904SAndy Whitcroft /* 617e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 618e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 619e7c4b0bfSAndy Whitcroft */ 620a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 621a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 622e7c4b0bfSAndy Whitcroft { 623a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 624a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 625e7c4b0bfSAndy Whitcroft } 626e7c4b0bfSAndy Whitcroft 6270fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 6280fe6e20bSNaoya Horiguchi unsigned long address) 6290fe6e20bSNaoya Horiguchi { 6300fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 6310fe6e20bSNaoya Horiguchi } 632dee41079SDan Williams EXPORT_SYMBOL_GPL(linear_hugepage_index); 6330fe6e20bSNaoya Horiguchi 63484afd99bSAndy Whitcroft /* 63508fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 63608fba699SMel Gorman * cases this will be same size as used by the page table entries. 63708fba699SMel Gorman */ 63808fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 63908fba699SMel Gorman { 64008fba699SMel Gorman struct hstate *hstate; 64108fba699SMel Gorman 64208fba699SMel Gorman if (!is_vm_hugetlb_page(vma)) 64308fba699SMel Gorman return PAGE_SIZE; 64408fba699SMel Gorman 64508fba699SMel Gorman hstate = hstate_vma(vma); 64608fba699SMel Gorman 6472415cf12SWanpeng Li return 1UL << huge_page_shift(hstate); 64808fba699SMel Gorman } 649f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 65008fba699SMel Gorman 65108fba699SMel Gorman /* 6523340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 6533340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 6543340289dSMel Gorman * architectures where it differs, an architecture-specific version of this 6553340289dSMel Gorman * function is required. 6563340289dSMel Gorman */ 6573340289dSMel Gorman #ifndef vma_mmu_pagesize 6583340289dSMel Gorman unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 6593340289dSMel Gorman { 6603340289dSMel Gorman return vma_kernel_pagesize(vma); 6613340289dSMel Gorman } 6623340289dSMel Gorman #endif 6633340289dSMel Gorman 6643340289dSMel Gorman /* 66584afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 66684afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 66784afd99bSAndy Whitcroft * alignment. 66884afd99bSAndy Whitcroft */ 66984afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 67084afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 67104f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 67284afd99bSAndy Whitcroft 673a1e78772SMel Gorman /* 674a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 675a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 676a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 677a1e78772SMel Gorman * 678a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 679a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 680a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 681a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 68284afd99bSAndy Whitcroft * 68384afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 68484afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 68584afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 68684afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 68784afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 68884afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 68984afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 69084afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 691a1e78772SMel Gorman */ 692e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 693e7c4b0bfSAndy Whitcroft { 694e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 695e7c4b0bfSAndy Whitcroft } 696e7c4b0bfSAndy Whitcroft 697e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 698e7c4b0bfSAndy Whitcroft unsigned long value) 699e7c4b0bfSAndy Whitcroft { 700e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 701e7c4b0bfSAndy Whitcroft } 702e7c4b0bfSAndy Whitcroft 7039119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 70484afd99bSAndy Whitcroft { 70584afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 7065e911373SMike Kravetz struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 7075e911373SMike Kravetz 7085e911373SMike Kravetz if (!resv_map || !rg) { 7095e911373SMike Kravetz kfree(resv_map); 7105e911373SMike Kravetz kfree(rg); 71184afd99bSAndy Whitcroft return NULL; 7125e911373SMike Kravetz } 71384afd99bSAndy Whitcroft 71484afd99bSAndy Whitcroft kref_init(&resv_map->refs); 7157b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 71684afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 71784afd99bSAndy Whitcroft 7185e911373SMike Kravetz resv_map->adds_in_progress = 0; 7195e911373SMike Kravetz 7205e911373SMike Kravetz INIT_LIST_HEAD(&resv_map->region_cache); 7215e911373SMike Kravetz list_add(&rg->link, &resv_map->region_cache); 7225e911373SMike Kravetz resv_map->region_cache_count = 1; 7235e911373SMike Kravetz 72484afd99bSAndy Whitcroft return resv_map; 72584afd99bSAndy Whitcroft } 72684afd99bSAndy Whitcroft 7279119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 72884afd99bSAndy Whitcroft { 72984afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 7305e911373SMike Kravetz struct list_head *head = &resv_map->region_cache; 7315e911373SMike Kravetz struct file_region *rg, *trg; 73284afd99bSAndy Whitcroft 73384afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 734feba16e2SMike Kravetz region_del(resv_map, 0, LONG_MAX); 7355e911373SMike Kravetz 7365e911373SMike Kravetz /* ... and any entries left in the cache */ 7375e911373SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 7385e911373SMike Kravetz list_del(&rg->link); 7395e911373SMike Kravetz kfree(rg); 7405e911373SMike Kravetz } 7415e911373SMike Kravetz 7425e911373SMike Kravetz VM_BUG_ON(resv_map->adds_in_progress); 7435e911373SMike Kravetz 74484afd99bSAndy Whitcroft kfree(resv_map); 74584afd99bSAndy Whitcroft } 74684afd99bSAndy Whitcroft 7474e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 7484e35f483SJoonsoo Kim { 7494e35f483SJoonsoo Kim return inode->i_mapping->private_data; 7504e35f483SJoonsoo Kim } 7514e35f483SJoonsoo Kim 75284afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 753a1e78772SMel Gorman { 75481d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 7554e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 7564e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 7574e35f483SJoonsoo Kim struct inode *inode = mapping->host; 7584e35f483SJoonsoo Kim 7594e35f483SJoonsoo Kim return inode_resv_map(inode); 7604e35f483SJoonsoo Kim 7614e35f483SJoonsoo Kim } else { 76284afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 76384afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 7644e35f483SJoonsoo Kim } 765a1e78772SMel Gorman } 766a1e78772SMel Gorman 76784afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 768a1e78772SMel Gorman { 76981d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 77081d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 771a1e78772SMel Gorman 77284afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 77384afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 77404f2cbe3SMel Gorman } 77504f2cbe3SMel Gorman 77604f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 77704f2cbe3SMel Gorman { 77881d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 77981d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 780e7c4b0bfSAndy Whitcroft 781e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 78204f2cbe3SMel Gorman } 78304f2cbe3SMel Gorman 78404f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 78504f2cbe3SMel Gorman { 78681d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 787e7c4b0bfSAndy Whitcroft 788e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 789a1e78772SMel Gorman } 790a1e78772SMel Gorman 79104f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 792a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 793a1e78772SMel Gorman { 79481d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 795f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 796a1e78772SMel Gorman vma->vm_private_data = (void *)0; 797a1e78772SMel Gorman } 798a1e78772SMel Gorman 799a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 800559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 801a1e78772SMel Gorman { 802af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 803af0ed73eSJoonsoo Kim /* 804af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 805af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 806af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 807af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 808af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 809af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 810af0ed73eSJoonsoo Kim * properly, so add work-around here. 811af0ed73eSJoonsoo Kim */ 812af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 813559ec2f8SNicholas Krause return true; 814af0ed73eSJoonsoo Kim else 815559ec2f8SNicholas Krause return false; 816af0ed73eSJoonsoo Kim } 817a63884e9SJoonsoo Kim 818a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 8191fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 8201fb1b0e9SMike Kravetz /* 8211fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 8221fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 8231fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 8241fb1b0e9SMike Kravetz * fallocate. In this case, there really are no reverves to 8251fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 8261fb1b0e9SMike Kravetz */ 8271fb1b0e9SMike Kravetz if (chg) 8281fb1b0e9SMike Kravetz return false; 8291fb1b0e9SMike Kravetz else 830559ec2f8SNicholas Krause return true; 8311fb1b0e9SMike Kravetz } 832a63884e9SJoonsoo Kim 833a63884e9SJoonsoo Kim /* 834a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 835a63884e9SJoonsoo Kim * private mappings. 836a63884e9SJoonsoo Kim */ 83767961f9dSMike Kravetz if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 83867961f9dSMike Kravetz /* 83967961f9dSMike Kravetz * Like the shared case above, a hole punch or truncate 84067961f9dSMike Kravetz * could have been performed on the private mapping. 84167961f9dSMike Kravetz * Examine the value of chg to determine if reserves 84267961f9dSMike Kravetz * actually exist or were previously consumed. 84367961f9dSMike Kravetz * Very Subtle - The value of chg comes from a previous 84467961f9dSMike Kravetz * call to vma_needs_reserves(). The reserve map for 84567961f9dSMike Kravetz * private mappings has different (opposite) semantics 84667961f9dSMike Kravetz * than that of shared mappings. vma_needs_reserves() 84767961f9dSMike Kravetz * has already taken this difference in semantics into 84867961f9dSMike Kravetz * account. Therefore, the meaning of chg is the same 84967961f9dSMike Kravetz * as in the shared case above. Code could easily be 85067961f9dSMike Kravetz * combined, but keeping it separate draws attention to 85167961f9dSMike Kravetz * subtle differences. 85267961f9dSMike Kravetz */ 85367961f9dSMike Kravetz if (chg) 85467961f9dSMike Kravetz return false; 85567961f9dSMike Kravetz else 856559ec2f8SNicholas Krause return true; 85767961f9dSMike Kravetz } 858a63884e9SJoonsoo Kim 859559ec2f8SNicholas Krause return false; 860a1e78772SMel Gorman } 861a1e78772SMel Gorman 862a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 8631da177e4SLinus Torvalds { 8641da177e4SLinus Torvalds int nid = page_to_nid(page); 8650edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 866a5516438SAndi Kleen h->free_huge_pages++; 867a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 8681da177e4SLinus Torvalds } 8691da177e4SLinus Torvalds 87094310cbcSAnshuman Khandual static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) 871bf50bab2SNaoya Horiguchi { 872bf50bab2SNaoya Horiguchi struct page *page; 873bf50bab2SNaoya Horiguchi 874c8721bbbSNaoya Horiguchi list_for_each_entry(page, &h->hugepage_freelists[nid], lru) 875243abd5bSNaoya Horiguchi if (!PageHWPoison(page)) 876c8721bbbSNaoya Horiguchi break; 877c8721bbbSNaoya Horiguchi /* 878c8721bbbSNaoya Horiguchi * if 'non-isolated free hugepage' not found on the list, 879c8721bbbSNaoya Horiguchi * the allocation fails. 880c8721bbbSNaoya Horiguchi */ 881c8721bbbSNaoya Horiguchi if (&h->hugepage_freelists[nid] == &page->lru) 882bf50bab2SNaoya Horiguchi return NULL; 8830edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 884a9869b83SNaoya Horiguchi set_page_refcounted(page); 885bf50bab2SNaoya Horiguchi h->free_huge_pages--; 886bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 887bf50bab2SNaoya Horiguchi return page; 888bf50bab2SNaoya Horiguchi } 889bf50bab2SNaoya Horiguchi 8903e59fcb0SMichal Hocko static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, 8913e59fcb0SMichal Hocko nodemask_t *nmask) 89294310cbcSAnshuman Khandual { 8933e59fcb0SMichal Hocko unsigned int cpuset_mems_cookie; 8943e59fcb0SMichal Hocko struct zonelist *zonelist; 8953e59fcb0SMichal Hocko struct zone *zone; 8963e59fcb0SMichal Hocko struct zoneref *z; 8973e59fcb0SMichal Hocko int node = -1; 8983e59fcb0SMichal Hocko 8993e59fcb0SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 9003e59fcb0SMichal Hocko 9013e59fcb0SMichal Hocko retry_cpuset: 9023e59fcb0SMichal Hocko cpuset_mems_cookie = read_mems_allowed_begin(); 9033e59fcb0SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 90494310cbcSAnshuman Khandual struct page *page; 90594310cbcSAnshuman Khandual 9063e59fcb0SMichal Hocko if (!cpuset_zone_allowed(zone, gfp_mask)) 9073e59fcb0SMichal Hocko continue; 9083e59fcb0SMichal Hocko /* 9093e59fcb0SMichal Hocko * no need to ask again on the same node. Pool is node rather than 9103e59fcb0SMichal Hocko * zone aware 9113e59fcb0SMichal Hocko */ 9123e59fcb0SMichal Hocko if (zone_to_nid(zone) == node) 9133e59fcb0SMichal Hocko continue; 9143e59fcb0SMichal Hocko node = zone_to_nid(zone); 91594310cbcSAnshuman Khandual 91694310cbcSAnshuman Khandual page = dequeue_huge_page_node_exact(h, node); 91794310cbcSAnshuman Khandual if (page) 91894310cbcSAnshuman Khandual return page; 91994310cbcSAnshuman Khandual } 9203e59fcb0SMichal Hocko if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 9213e59fcb0SMichal Hocko goto retry_cpuset; 9223e59fcb0SMichal Hocko 92394310cbcSAnshuman Khandual return NULL; 92494310cbcSAnshuman Khandual } 92594310cbcSAnshuman Khandual 92686cdb465SNaoya Horiguchi /* Movability of hugepages depends on migration support. */ 92786cdb465SNaoya Horiguchi static inline gfp_t htlb_alloc_mask(struct hstate *h) 92886cdb465SNaoya Horiguchi { 929100873d7SNaoya Horiguchi if (hugepages_treat_as_movable || hugepage_migration_supported(h)) 93086cdb465SNaoya Horiguchi return GFP_HIGHUSER_MOVABLE; 93186cdb465SNaoya Horiguchi else 93286cdb465SNaoya Horiguchi return GFP_HIGHUSER; 93386cdb465SNaoya Horiguchi } 93486cdb465SNaoya Horiguchi 935a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 936a5516438SAndi Kleen struct vm_area_struct *vma, 937af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 938af0ed73eSJoonsoo Kim long chg) 9391da177e4SLinus Torvalds { 9403e59fcb0SMichal Hocko struct page *page; 941480eccf9SLee Schermerhorn struct mempolicy *mpol; 94204ec6264SVlastimil Babka gfp_t gfp_mask; 9433e59fcb0SMichal Hocko nodemask_t *nodemask; 94404ec6264SVlastimil Babka int nid; 9451da177e4SLinus Torvalds 946a1e78772SMel Gorman /* 947a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 948a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 949a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 950a1e78772SMel Gorman */ 951af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg) && 952a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 953c0ff7453SMiao Xie goto err; 954a1e78772SMel Gorman 95504f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 956a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 9576eab04a8SJustin P. Mattock goto err; 95804f2cbe3SMel Gorman 95904ec6264SVlastimil Babka gfp_mask = htlb_alloc_mask(h); 96004ec6264SVlastimil Babka nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 9613e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 9623e59fcb0SMichal Hocko if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { 96307443a85SJoonsoo Kim SetPagePrivate(page); 964a63884e9SJoonsoo Kim h->resv_huge_pages--; 965bf50bab2SNaoya Horiguchi } 966cc9a6c87SMel Gorman 967cc9a6c87SMel Gorman mpol_cond_put(mpol); 968cc9a6c87SMel Gorman return page; 969cc9a6c87SMel Gorman 970c0ff7453SMiao Xie err: 971cc9a6c87SMel Gorman return NULL; 9721da177e4SLinus Torvalds } 9731da177e4SLinus Torvalds 9741cac6f2cSLuiz Capitulino /* 9751cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 9761cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 9771cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 9781cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 9791cac6f2cSLuiz Capitulino * node for alloc or free. 9801cac6f2cSLuiz Capitulino */ 9811cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 9821cac6f2cSLuiz Capitulino { 9830edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 9841cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 9851cac6f2cSLuiz Capitulino 9861cac6f2cSLuiz Capitulino return nid; 9871cac6f2cSLuiz Capitulino } 9881cac6f2cSLuiz Capitulino 9891cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 9901cac6f2cSLuiz Capitulino { 9911cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 9921cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 9931cac6f2cSLuiz Capitulino return nid; 9941cac6f2cSLuiz Capitulino } 9951cac6f2cSLuiz Capitulino 9961cac6f2cSLuiz Capitulino /* 9971cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 9981cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 9991cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 10001cac6f2cSLuiz Capitulino * mask. 10011cac6f2cSLuiz Capitulino */ 10021cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 10031cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 10041cac6f2cSLuiz Capitulino { 10051cac6f2cSLuiz Capitulino int nid; 10061cac6f2cSLuiz Capitulino 10071cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 10081cac6f2cSLuiz Capitulino 10091cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 10101cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 10111cac6f2cSLuiz Capitulino 10121cac6f2cSLuiz Capitulino return nid; 10131cac6f2cSLuiz Capitulino } 10141cac6f2cSLuiz Capitulino 10151cac6f2cSLuiz Capitulino /* 10161cac6f2cSLuiz Capitulino * helper for free_pool_huge_page() - return the previously saved 10171cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 10181cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 10191cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 10201cac6f2cSLuiz Capitulino */ 10211cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 10221cac6f2cSLuiz Capitulino { 10231cac6f2cSLuiz Capitulino int nid; 10241cac6f2cSLuiz Capitulino 10251cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 10261cac6f2cSLuiz Capitulino 10271cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 10281cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 10291cac6f2cSLuiz Capitulino 10301cac6f2cSLuiz Capitulino return nid; 10311cac6f2cSLuiz Capitulino } 10321cac6f2cSLuiz Capitulino 10331cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 10341cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 10351cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 10361cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 10371cac6f2cSLuiz Capitulino nr_nodes--) 10381cac6f2cSLuiz Capitulino 10391cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 10401cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 10411cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 10421cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 10431cac6f2cSLuiz Capitulino nr_nodes--) 10441cac6f2cSLuiz Capitulino 1045e1073d1eSAneesh Kumar K.V #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 1046944d9fecSLuiz Capitulino static void destroy_compound_gigantic_page(struct page *page, 1047d00181b9SKirill A. Shutemov unsigned int order) 1048944d9fecSLuiz Capitulino { 1049944d9fecSLuiz Capitulino int i; 1050944d9fecSLuiz Capitulino int nr_pages = 1 << order; 1051944d9fecSLuiz Capitulino struct page *p = page + 1; 1052944d9fecSLuiz Capitulino 1053c8cc708aSGerald Schaefer atomic_set(compound_mapcount_ptr(page), 0); 1054944d9fecSLuiz Capitulino for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 10551d798ca3SKirill A. Shutemov clear_compound_head(p); 1056944d9fecSLuiz Capitulino set_page_refcounted(p); 1057944d9fecSLuiz Capitulino } 1058944d9fecSLuiz Capitulino 1059944d9fecSLuiz Capitulino set_compound_order(page, 0); 1060944d9fecSLuiz Capitulino __ClearPageHead(page); 1061944d9fecSLuiz Capitulino } 1062944d9fecSLuiz Capitulino 1063d00181b9SKirill A. Shutemov static void free_gigantic_page(struct page *page, unsigned int order) 1064944d9fecSLuiz Capitulino { 1065944d9fecSLuiz Capitulino free_contig_range(page_to_pfn(page), 1 << order); 1066944d9fecSLuiz Capitulino } 1067944d9fecSLuiz Capitulino 1068944d9fecSLuiz Capitulino static int __alloc_gigantic_page(unsigned long start_pfn, 106979b63f12SMichal Hocko unsigned long nr_pages, gfp_t gfp_mask) 1070944d9fecSLuiz Capitulino { 1071944d9fecSLuiz Capitulino unsigned long end_pfn = start_pfn + nr_pages; 1072ca96b625SLucas Stach return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE, 107379b63f12SMichal Hocko gfp_mask); 1074944d9fecSLuiz Capitulino } 1075944d9fecSLuiz Capitulino 1076f44b2ddaSJoonsoo Kim static bool pfn_range_valid_gigantic(struct zone *z, 1077f44b2ddaSJoonsoo Kim unsigned long start_pfn, unsigned long nr_pages) 1078944d9fecSLuiz Capitulino { 1079944d9fecSLuiz Capitulino unsigned long i, end_pfn = start_pfn + nr_pages; 1080944d9fecSLuiz Capitulino struct page *page; 1081944d9fecSLuiz Capitulino 1082944d9fecSLuiz Capitulino for (i = start_pfn; i < end_pfn; i++) { 1083944d9fecSLuiz Capitulino if (!pfn_valid(i)) 1084944d9fecSLuiz Capitulino return false; 1085944d9fecSLuiz Capitulino 1086944d9fecSLuiz Capitulino page = pfn_to_page(i); 1087944d9fecSLuiz Capitulino 1088f44b2ddaSJoonsoo Kim if (page_zone(page) != z) 1089f44b2ddaSJoonsoo Kim return false; 1090f44b2ddaSJoonsoo Kim 1091944d9fecSLuiz Capitulino if (PageReserved(page)) 1092944d9fecSLuiz Capitulino return false; 1093944d9fecSLuiz Capitulino 1094944d9fecSLuiz Capitulino if (page_count(page) > 0) 1095944d9fecSLuiz Capitulino return false; 1096944d9fecSLuiz Capitulino 1097944d9fecSLuiz Capitulino if (PageHuge(page)) 1098944d9fecSLuiz Capitulino return false; 1099944d9fecSLuiz Capitulino } 1100944d9fecSLuiz Capitulino 1101944d9fecSLuiz Capitulino return true; 1102944d9fecSLuiz Capitulino } 1103944d9fecSLuiz Capitulino 1104944d9fecSLuiz Capitulino static bool zone_spans_last_pfn(const struct zone *zone, 1105944d9fecSLuiz Capitulino unsigned long start_pfn, unsigned long nr_pages) 1106944d9fecSLuiz Capitulino { 1107944d9fecSLuiz Capitulino unsigned long last_pfn = start_pfn + nr_pages - 1; 1108944d9fecSLuiz Capitulino return zone_spans_pfn(zone, last_pfn); 1109944d9fecSLuiz Capitulino } 1110944d9fecSLuiz Capitulino 111179b63f12SMichal Hocko static struct page *alloc_gigantic_page(int nid, struct hstate *h) 1112944d9fecSLuiz Capitulino { 111379b63f12SMichal Hocko unsigned int order = huge_page_order(h); 1114944d9fecSLuiz Capitulino unsigned long nr_pages = 1 << order; 1115944d9fecSLuiz Capitulino unsigned long ret, pfn, flags; 111679b63f12SMichal Hocko struct zonelist *zonelist; 111779b63f12SMichal Hocko struct zone *zone; 111879b63f12SMichal Hocko struct zoneref *z; 111979b63f12SMichal Hocko gfp_t gfp_mask; 1120944d9fecSLuiz Capitulino 112179b63f12SMichal Hocko gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 112279b63f12SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 112379b63f12SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), NULL) { 112479b63f12SMichal Hocko spin_lock_irqsave(&zone->lock, flags); 1125944d9fecSLuiz Capitulino 112679b63f12SMichal Hocko pfn = ALIGN(zone->zone_start_pfn, nr_pages); 112779b63f12SMichal Hocko while (zone_spans_last_pfn(zone, pfn, nr_pages)) { 112879b63f12SMichal Hocko if (pfn_range_valid_gigantic(zone, pfn, nr_pages)) { 1129944d9fecSLuiz Capitulino /* 1130944d9fecSLuiz Capitulino * We release the zone lock here because 1131944d9fecSLuiz Capitulino * alloc_contig_range() will also lock the zone 1132944d9fecSLuiz Capitulino * at some point. If there's an allocation 1133944d9fecSLuiz Capitulino * spinning on this lock, it may win the race 1134944d9fecSLuiz Capitulino * and cause alloc_contig_range() to fail... 1135944d9fecSLuiz Capitulino */ 113679b63f12SMichal Hocko spin_unlock_irqrestore(&zone->lock, flags); 113779b63f12SMichal Hocko ret = __alloc_gigantic_page(pfn, nr_pages, gfp_mask); 1138944d9fecSLuiz Capitulino if (!ret) 1139944d9fecSLuiz Capitulino return pfn_to_page(pfn); 114079b63f12SMichal Hocko spin_lock_irqsave(&zone->lock, flags); 1141944d9fecSLuiz Capitulino } 1142944d9fecSLuiz Capitulino pfn += nr_pages; 1143944d9fecSLuiz Capitulino } 1144944d9fecSLuiz Capitulino 114579b63f12SMichal Hocko spin_unlock_irqrestore(&zone->lock, flags); 1146944d9fecSLuiz Capitulino } 1147944d9fecSLuiz Capitulino 1148944d9fecSLuiz Capitulino return NULL; 1149944d9fecSLuiz Capitulino } 1150944d9fecSLuiz Capitulino 1151944d9fecSLuiz Capitulino static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); 1152d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order); 1153944d9fecSLuiz Capitulino 1154944d9fecSLuiz Capitulino static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid) 1155944d9fecSLuiz Capitulino { 1156944d9fecSLuiz Capitulino struct page *page; 1157944d9fecSLuiz Capitulino 115879b63f12SMichal Hocko page = alloc_gigantic_page(nid, h); 1159944d9fecSLuiz Capitulino if (page) { 1160944d9fecSLuiz Capitulino prep_compound_gigantic_page(page, huge_page_order(h)); 1161944d9fecSLuiz Capitulino prep_new_huge_page(h, page, nid); 1162944d9fecSLuiz Capitulino } 1163944d9fecSLuiz Capitulino 1164944d9fecSLuiz Capitulino return page; 1165944d9fecSLuiz Capitulino } 1166944d9fecSLuiz Capitulino 1167944d9fecSLuiz Capitulino static int alloc_fresh_gigantic_page(struct hstate *h, 1168944d9fecSLuiz Capitulino nodemask_t *nodes_allowed) 1169944d9fecSLuiz Capitulino { 1170944d9fecSLuiz Capitulino struct page *page = NULL; 1171944d9fecSLuiz Capitulino int nr_nodes, node; 1172944d9fecSLuiz Capitulino 1173944d9fecSLuiz Capitulino for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1174944d9fecSLuiz Capitulino page = alloc_fresh_gigantic_page_node(h, node); 1175944d9fecSLuiz Capitulino if (page) 1176944d9fecSLuiz Capitulino return 1; 1177944d9fecSLuiz Capitulino } 1178944d9fecSLuiz Capitulino 1179944d9fecSLuiz Capitulino return 0; 1180944d9fecSLuiz Capitulino } 1181944d9fecSLuiz Capitulino 1182e1073d1eSAneesh Kumar K.V #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 1183944d9fecSLuiz Capitulino static inline bool gigantic_page_supported(void) { return false; } 1184d00181b9SKirill A. Shutemov static inline void free_gigantic_page(struct page *page, unsigned int order) { } 1185944d9fecSLuiz Capitulino static inline void destroy_compound_gigantic_page(struct page *page, 1186d00181b9SKirill A. Shutemov unsigned int order) { } 1187944d9fecSLuiz Capitulino static inline int alloc_fresh_gigantic_page(struct hstate *h, 1188944d9fecSLuiz Capitulino nodemask_t *nodes_allowed) { return 0; } 1189944d9fecSLuiz Capitulino #endif 1190944d9fecSLuiz Capitulino 1191a5516438SAndi Kleen static void update_and_free_page(struct hstate *h, struct page *page) 11926af2acb6SAdam Litke { 11936af2acb6SAdam Litke int i; 1194a5516438SAndi Kleen 1195944d9fecSLuiz Capitulino if (hstate_is_gigantic(h) && !gigantic_page_supported()) 1196944d9fecSLuiz Capitulino return; 119718229df5SAndy Whitcroft 1198a5516438SAndi Kleen h->nr_huge_pages--; 1199a5516438SAndi Kleen h->nr_huge_pages_node[page_to_nid(page)]--; 1200a5516438SAndi Kleen for (i = 0; i < pages_per_huge_page(h); i++) { 120132f84528SChris Forbes page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 120232f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1203a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1204a7407a27SLuiz Capitulino 1 << PG_writeback); 12056af2acb6SAdam Litke } 1206309381feSSasha Levin VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 1207f1e61557SKirill A. Shutemov set_compound_page_dtor(page, NULL_COMPOUND_DTOR); 12086af2acb6SAdam Litke set_page_refcounted(page); 1209944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) { 1210944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 1211944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 1212944d9fecSLuiz Capitulino } else { 1213a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 12146af2acb6SAdam Litke } 1215944d9fecSLuiz Capitulino } 12166af2acb6SAdam Litke 1217e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1218e5ff2159SAndi Kleen { 1219e5ff2159SAndi Kleen struct hstate *h; 1220e5ff2159SAndi Kleen 1221e5ff2159SAndi Kleen for_each_hstate(h) { 1222e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1223e5ff2159SAndi Kleen return h; 1224e5ff2159SAndi Kleen } 1225e5ff2159SAndi Kleen return NULL; 1226e5ff2159SAndi Kleen } 1227e5ff2159SAndi Kleen 1228bcc54222SNaoya Horiguchi /* 1229bcc54222SNaoya Horiguchi * Test to determine whether the hugepage is "active/in-use" (i.e. being linked 1230bcc54222SNaoya Horiguchi * to hstate->hugepage_activelist.) 1231bcc54222SNaoya Horiguchi * 1232bcc54222SNaoya Horiguchi * This function can be called for tail pages, but never returns true for them. 1233bcc54222SNaoya Horiguchi */ 1234bcc54222SNaoya Horiguchi bool page_huge_active(struct page *page) 1235bcc54222SNaoya Horiguchi { 1236bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHuge(page), page); 1237bcc54222SNaoya Horiguchi return PageHead(page) && PagePrivate(&page[1]); 1238bcc54222SNaoya Horiguchi } 1239bcc54222SNaoya Horiguchi 1240bcc54222SNaoya Horiguchi /* never called for tail page */ 1241bcc54222SNaoya Horiguchi static void set_page_huge_active(struct page *page) 1242bcc54222SNaoya Horiguchi { 1243bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 1244bcc54222SNaoya Horiguchi SetPagePrivate(&page[1]); 1245bcc54222SNaoya Horiguchi } 1246bcc54222SNaoya Horiguchi 1247bcc54222SNaoya Horiguchi static void clear_page_huge_active(struct page *page) 1248bcc54222SNaoya Horiguchi { 1249bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 1250bcc54222SNaoya Horiguchi ClearPagePrivate(&page[1]); 1251bcc54222SNaoya Horiguchi } 1252bcc54222SNaoya Horiguchi 12538f1d26d0SAtsushi Kumagai void free_huge_page(struct page *page) 125427a85ef1SDavid Gibson { 1255a5516438SAndi Kleen /* 1256a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1257a5516438SAndi Kleen * compound page destructor. 1258a5516438SAndi Kleen */ 1259e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 12607893d1d5SAdam Litke int nid = page_to_nid(page); 126190481622SDavid Gibson struct hugepage_subpool *spool = 126290481622SDavid Gibson (struct hugepage_subpool *)page_private(page); 126307443a85SJoonsoo Kim bool restore_reserve; 126427a85ef1SDavid Gibson 1265e5df70abSAndy Whitcroft set_page_private(page, 0); 126623be7468SMel Gorman page->mapping = NULL; 1267b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1268b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_mapcount(page), page); 126907443a85SJoonsoo Kim restore_reserve = PagePrivate(page); 127016c794b4SJoonsoo Kim ClearPagePrivate(page); 127127a85ef1SDavid Gibson 12721c5ecae3SMike Kravetz /* 12731c5ecae3SMike Kravetz * A return code of zero implies that the subpool will be under its 12741c5ecae3SMike Kravetz * minimum size if the reservation is not restored after page is free. 12751c5ecae3SMike Kravetz * Therefore, force restore_reserve operation. 12761c5ecae3SMike Kravetz */ 12771c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 12781c5ecae3SMike Kravetz restore_reserve = true; 12791c5ecae3SMike Kravetz 128027a85ef1SDavid Gibson spin_lock(&hugetlb_lock); 1281bcc54222SNaoya Horiguchi clear_page_huge_active(page); 12826d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 12836d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 128407443a85SJoonsoo Kim if (restore_reserve) 128507443a85SJoonsoo Kim h->resv_huge_pages++; 128607443a85SJoonsoo Kim 1287944d9fecSLuiz Capitulino if (h->surplus_huge_pages_node[nid]) { 12880edaecfaSAneesh Kumar K.V /* remove the page from active list */ 12890edaecfaSAneesh Kumar K.V list_del(&page->lru); 1290a5516438SAndi Kleen update_and_free_page(h, page); 1291a5516438SAndi Kleen h->surplus_huge_pages--; 1292a5516438SAndi Kleen h->surplus_huge_pages_node[nid]--; 12937893d1d5SAdam Litke } else { 12945d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1295a5516438SAndi Kleen enqueue_huge_page(h, page); 12967893d1d5SAdam Litke } 129727a85ef1SDavid Gibson spin_unlock(&hugetlb_lock); 129827a85ef1SDavid Gibson } 129927a85ef1SDavid Gibson 1300a5516438SAndi Kleen static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1301b7ba30c6SAndi Kleen { 13020edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1303f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1304b7ba30c6SAndi Kleen spin_lock(&hugetlb_lock); 13059dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 1306a5516438SAndi Kleen h->nr_huge_pages++; 1307a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 1308b7ba30c6SAndi Kleen spin_unlock(&hugetlb_lock); 1309b7ba30c6SAndi Kleen put_page(page); /* free it into the hugepage allocator */ 1310b7ba30c6SAndi Kleen } 1311b7ba30c6SAndi Kleen 1312d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order) 131320a0307cSWu Fengguang { 131420a0307cSWu Fengguang int i; 131520a0307cSWu Fengguang int nr_pages = 1 << order; 131620a0307cSWu Fengguang struct page *p = page + 1; 131720a0307cSWu Fengguang 131820a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 131920a0307cSWu Fengguang set_compound_order(page, order); 1320ef5a22beSAndrea Arcangeli __ClearPageReserved(page); 1321de09d31dSKirill A. Shutemov __SetPageHead(page); 132220a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1323ef5a22beSAndrea Arcangeli /* 1324ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1325ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1326ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 1327ef5a22beSAndrea Arcangeli * too. Otherwse drivers using get_user_pages() to access tail 1328ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1329ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1330ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1331ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1332ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1333ef5a22beSAndrea Arcangeli * after get_user_pages(). 1334ef5a22beSAndrea Arcangeli */ 1335ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 133658a84aa9SYouquan Song set_page_count(p, 0); 13371d798ca3SKirill A. Shutemov set_compound_head(p, page); 133820a0307cSWu Fengguang } 1339b4330afbSMike Kravetz atomic_set(compound_mapcount_ptr(page), -1); 134020a0307cSWu Fengguang } 134120a0307cSWu Fengguang 13427795912cSAndrew Morton /* 13437795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 13447795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 13457795912cSAndrew Morton * details. 13467795912cSAndrew Morton */ 134720a0307cSWu Fengguang int PageHuge(struct page *page) 134820a0307cSWu Fengguang { 134920a0307cSWu Fengguang if (!PageCompound(page)) 135020a0307cSWu Fengguang return 0; 135120a0307cSWu Fengguang 135220a0307cSWu Fengguang page = compound_head(page); 1353f1e61557SKirill A. Shutemov return page[1].compound_dtor == HUGETLB_PAGE_DTOR; 135420a0307cSWu Fengguang } 135543131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 135643131e14SNaoya Horiguchi 135727c73ae7SAndrea Arcangeli /* 135827c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 135927c73ae7SAndrea Arcangeli * normal or transparent huge pages. 136027c73ae7SAndrea Arcangeli */ 136127c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 136227c73ae7SAndrea Arcangeli { 136327c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 136427c73ae7SAndrea Arcangeli return 0; 136527c73ae7SAndrea Arcangeli 1366758f66a2SAndrew Morton return get_compound_page_dtor(page_head) == free_huge_page; 136727c73ae7SAndrea Arcangeli } 136827c73ae7SAndrea Arcangeli 136913d60f4bSZhang Yi pgoff_t __basepage_index(struct page *page) 137013d60f4bSZhang Yi { 137113d60f4bSZhang Yi struct page *page_head = compound_head(page); 137213d60f4bSZhang Yi pgoff_t index = page_index(page_head); 137313d60f4bSZhang Yi unsigned long compound_idx; 137413d60f4bSZhang Yi 137513d60f4bSZhang Yi if (!PageHuge(page_head)) 137613d60f4bSZhang Yi return page_index(page); 137713d60f4bSZhang Yi 137813d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 137913d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 138013d60f4bSZhang Yi else 138113d60f4bSZhang Yi compound_idx = page - page_head; 138213d60f4bSZhang Yi 138313d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 138413d60f4bSZhang Yi } 138513d60f4bSZhang Yi 1386a5516438SAndi Kleen static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) 13871da177e4SLinus Torvalds { 13881da177e4SLinus Torvalds struct page *page; 1389f96efd58SJoe Jin 139096db800fSVlastimil Babka page = __alloc_pages_node(nid, 139186cdb465SNaoya Horiguchi htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| 1392dcda9b04SMichal Hocko __GFP_RETRY_MAYFAIL|__GFP_NOWARN, 1393a5516438SAndi Kleen huge_page_order(h)); 13941da177e4SLinus Torvalds if (page) { 1395a5516438SAndi Kleen prep_new_huge_page(h, page, nid); 13961da177e4SLinus Torvalds } 139763b4613cSNishanth Aravamudan 139863b4613cSNishanth Aravamudan return page; 139963b4613cSNishanth Aravamudan } 140063b4613cSNishanth Aravamudan 1401b2261026SJoonsoo Kim static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 1402b2261026SJoonsoo Kim { 1403b2261026SJoonsoo Kim struct page *page; 1404b2261026SJoonsoo Kim int nr_nodes, node; 1405b2261026SJoonsoo Kim int ret = 0; 1406b2261026SJoonsoo Kim 1407b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1408b2261026SJoonsoo Kim page = alloc_fresh_huge_page_node(h, node); 1409b2261026SJoonsoo Kim if (page) { 1410b2261026SJoonsoo Kim ret = 1; 1411b2261026SJoonsoo Kim break; 1412b2261026SJoonsoo Kim } 1413b2261026SJoonsoo Kim } 1414b2261026SJoonsoo Kim 1415b2261026SJoonsoo Kim if (ret) 1416b2261026SJoonsoo Kim count_vm_event(HTLB_BUDDY_PGALLOC); 1417b2261026SJoonsoo Kim else 1418b2261026SJoonsoo Kim count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 1419b2261026SJoonsoo Kim 1420b2261026SJoonsoo Kim return ret; 1421b2261026SJoonsoo Kim } 1422b2261026SJoonsoo Kim 1423e8c5c824SLee Schermerhorn /* 1424e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 1425e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 1426e8c5c824SLee Schermerhorn * balanced over allowed nodes. 1427e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 1428e8c5c824SLee Schermerhorn */ 14296ae11b27SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 14306ae11b27SLee Schermerhorn bool acct_surplus) 1431e8c5c824SLee Schermerhorn { 1432b2261026SJoonsoo Kim int nr_nodes, node; 1433e8c5c824SLee Schermerhorn int ret = 0; 1434e8c5c824SLee Schermerhorn 1435b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1436685f3457SLee Schermerhorn /* 1437685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 1438685f3457SLee Schermerhorn * nodes with surplus pages. 1439685f3457SLee Schermerhorn */ 1440b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 1441b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 1442e8c5c824SLee Schermerhorn struct page *page = 1443b2261026SJoonsoo Kim list_entry(h->hugepage_freelists[node].next, 1444e8c5c824SLee Schermerhorn struct page, lru); 1445e8c5c824SLee Schermerhorn list_del(&page->lru); 1446e8c5c824SLee Schermerhorn h->free_huge_pages--; 1447b2261026SJoonsoo Kim h->free_huge_pages_node[node]--; 1448685f3457SLee Schermerhorn if (acct_surplus) { 1449685f3457SLee Schermerhorn h->surplus_huge_pages--; 1450b2261026SJoonsoo Kim h->surplus_huge_pages_node[node]--; 1451685f3457SLee Schermerhorn } 1452e8c5c824SLee Schermerhorn update_and_free_page(h, page); 1453e8c5c824SLee Schermerhorn ret = 1; 14549a76db09SLee Schermerhorn break; 1455e8c5c824SLee Schermerhorn } 1456b2261026SJoonsoo Kim } 1457e8c5c824SLee Schermerhorn 1458e8c5c824SLee Schermerhorn return ret; 1459e8c5c824SLee Schermerhorn } 1460e8c5c824SLee Schermerhorn 1461c8721bbbSNaoya Horiguchi /* 1462c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 1463082d5b6bSGerald Schaefer * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the 1464082d5b6bSGerald Schaefer * number of free hugepages would be reduced below the number of reserved 1465082d5b6bSGerald Schaefer * hugepages. 1466c8721bbbSNaoya Horiguchi */ 1467c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 1468c8721bbbSNaoya Horiguchi { 1469082d5b6bSGerald Schaefer int rc = 0; 1470082d5b6bSGerald Schaefer 1471c8721bbbSNaoya Horiguchi spin_lock(&hugetlb_lock); 1472c8721bbbSNaoya Horiguchi if (PageHuge(page) && !page_count(page)) { 14732247bb33SGerald Schaefer struct page *head = compound_head(page); 14742247bb33SGerald Schaefer struct hstate *h = page_hstate(head); 14752247bb33SGerald Schaefer int nid = page_to_nid(head); 1476082d5b6bSGerald Schaefer if (h->free_huge_pages - h->resv_huge_pages == 0) { 1477082d5b6bSGerald Schaefer rc = -EBUSY; 1478082d5b6bSGerald Schaefer goto out; 1479082d5b6bSGerald Schaefer } 1480c3114a84SAnshuman Khandual /* 1481c3114a84SAnshuman Khandual * Move PageHWPoison flag from head page to the raw error page, 1482c3114a84SAnshuman Khandual * which makes any subpages rather than the error page reusable. 1483c3114a84SAnshuman Khandual */ 1484c3114a84SAnshuman Khandual if (PageHWPoison(head) && page != head) { 1485c3114a84SAnshuman Khandual SetPageHWPoison(page); 1486c3114a84SAnshuman Khandual ClearPageHWPoison(head); 1487c3114a84SAnshuman Khandual } 14882247bb33SGerald Schaefer list_del(&head->lru); 1489c8721bbbSNaoya Horiguchi h->free_huge_pages--; 1490c8721bbbSNaoya Horiguchi h->free_huge_pages_node[nid]--; 1491c1470b33Szhong jiang h->max_huge_pages--; 14922247bb33SGerald Schaefer update_and_free_page(h, head); 1493c8721bbbSNaoya Horiguchi } 1494082d5b6bSGerald Schaefer out: 1495c8721bbbSNaoya Horiguchi spin_unlock(&hugetlb_lock); 1496082d5b6bSGerald Schaefer return rc; 1497c8721bbbSNaoya Horiguchi } 1498c8721bbbSNaoya Horiguchi 1499c8721bbbSNaoya Horiguchi /* 1500c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 1501c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 15022247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 15032247bb33SGerald Schaefer * part of it lies within the given range. 1504082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 1505082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 1506c8721bbbSNaoya Horiguchi */ 1507082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 1508c8721bbbSNaoya Horiguchi { 1509c8721bbbSNaoya Horiguchi unsigned long pfn; 1510eb03aa00SGerald Schaefer struct page *page; 1511082d5b6bSGerald Schaefer int rc = 0; 1512c8721bbbSNaoya Horiguchi 1513d0177639SLi Zhong if (!hugepages_supported()) 1514082d5b6bSGerald Schaefer return rc; 1515d0177639SLi Zhong 1516eb03aa00SGerald Schaefer for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { 1517eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 1518eb03aa00SGerald Schaefer if (PageHuge(page) && !page_count(page)) { 1519eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 1520eb03aa00SGerald Schaefer if (rc) 1521082d5b6bSGerald Schaefer break; 1522eb03aa00SGerald Schaefer } 1523eb03aa00SGerald Schaefer } 1524082d5b6bSGerald Schaefer 1525082d5b6bSGerald Schaefer return rc; 1526c8721bbbSNaoya Horiguchi } 1527c8721bbbSNaoya Horiguchi 1528099730d6SDave Hansen static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h, 1529aaf14e40SMichal Hocko gfp_t gfp_mask, int nid, nodemask_t *nmask) 1530099730d6SDave Hansen { 1531099730d6SDave Hansen int order = huge_page_order(h); 1532099730d6SDave Hansen 1533dcda9b04SMichal Hocko gfp_mask |= __GFP_COMP|__GFP_RETRY_MAYFAIL|__GFP_NOWARN; 1534aaf14e40SMichal Hocko if (nid == NUMA_NO_NODE) 1535aaf14e40SMichal Hocko nid = numa_mem_id(); 1536aaf14e40SMichal Hocko return __alloc_pages_nodemask(gfp_mask, order, nid, nmask); 1537099730d6SDave Hansen } 1538099730d6SDave Hansen 1539aaf14e40SMichal Hocko static struct page *__alloc_buddy_huge_page(struct hstate *h, gfp_t gfp_mask, 1540aaf14e40SMichal Hocko int nid, nodemask_t *nmask) 15417893d1d5SAdam Litke { 15427893d1d5SAdam Litke struct page *page; 1543bf50bab2SNaoya Horiguchi unsigned int r_nid; 15447893d1d5SAdam Litke 1545bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1546aa888a74SAndi Kleen return NULL; 1547aa888a74SAndi Kleen 1548d1c3fb1fSNishanth Aravamudan /* 1549d1c3fb1fSNishanth Aravamudan * Assume we will successfully allocate the surplus page to 1550d1c3fb1fSNishanth Aravamudan * prevent racing processes from causing the surplus to exceed 1551d1c3fb1fSNishanth Aravamudan * overcommit 1552d1c3fb1fSNishanth Aravamudan * 1553d1c3fb1fSNishanth Aravamudan * This however introduces a different race, where a process B 1554d1c3fb1fSNishanth Aravamudan * tries to grow the static hugepage pool while alloc_pages() is 1555d1c3fb1fSNishanth Aravamudan * called by process A. B will only examine the per-node 1556d1c3fb1fSNishanth Aravamudan * counters in determining if surplus huge pages can be 1557d1c3fb1fSNishanth Aravamudan * converted to normal huge pages in adjust_pool_surplus(). A 1558d1c3fb1fSNishanth Aravamudan * won't be able to increment the per-node counter, until the 1559d1c3fb1fSNishanth Aravamudan * lock is dropped by B, but B doesn't drop hugetlb_lock until 1560d1c3fb1fSNishanth Aravamudan * no more huge pages can be converted from surplus to normal 1561d1c3fb1fSNishanth Aravamudan * state (and doesn't try to convert again). Thus, we have a 1562d1c3fb1fSNishanth Aravamudan * case where a surplus huge page exists, the pool is grown, and 1563d1c3fb1fSNishanth Aravamudan * the surplus huge page still exists after, even though it 1564d1c3fb1fSNishanth Aravamudan * should just have been converted to a normal huge page. This 1565d1c3fb1fSNishanth Aravamudan * does not leak memory, though, as the hugepage will be freed 1566d1c3fb1fSNishanth Aravamudan * once it is out of use. It also does not allow the counters to 1567d1c3fb1fSNishanth Aravamudan * go out of whack in adjust_pool_surplus() as we don't modify 1568d1c3fb1fSNishanth Aravamudan * the node values until we've gotten the hugepage and only the 1569d1c3fb1fSNishanth Aravamudan * per-node value is checked there. 1570d1c3fb1fSNishanth Aravamudan */ 1571d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 1572a5516438SAndi Kleen if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 1573d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1574d1c3fb1fSNishanth Aravamudan return NULL; 1575d1c3fb1fSNishanth Aravamudan } else { 1576a5516438SAndi Kleen h->nr_huge_pages++; 1577a5516438SAndi Kleen h->surplus_huge_pages++; 1578d1c3fb1fSNishanth Aravamudan } 1579d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1580d1c3fb1fSNishanth Aravamudan 1581aaf14e40SMichal Hocko page = __hugetlb_alloc_buddy_huge_page(h, gfp_mask, nid, nmask); 1582d1c3fb1fSNishanth Aravamudan 15837893d1d5SAdam Litke spin_lock(&hugetlb_lock); 1584d1c3fb1fSNishanth Aravamudan if (page) { 15850edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1586bf50bab2SNaoya Horiguchi r_nid = page_to_nid(page); 1587f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 15889dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 1589d1c3fb1fSNishanth Aravamudan /* 1590d1c3fb1fSNishanth Aravamudan * We incremented the global counters already 1591d1c3fb1fSNishanth Aravamudan */ 1592bf50bab2SNaoya Horiguchi h->nr_huge_pages_node[r_nid]++; 1593bf50bab2SNaoya Horiguchi h->surplus_huge_pages_node[r_nid]++; 15943b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC); 1595d1c3fb1fSNishanth Aravamudan } else { 1596a5516438SAndi Kleen h->nr_huge_pages--; 1597a5516438SAndi Kleen h->surplus_huge_pages--; 15983b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 15997893d1d5SAdam Litke } 1600d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 16017893d1d5SAdam Litke 16027893d1d5SAdam Litke return page; 16037893d1d5SAdam Litke } 16047893d1d5SAdam Litke 1605e4e574b7SAdam Litke /* 1606099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 1607099730d6SDave Hansen */ 1608e0ec90eeSDave Hansen static 1609099730d6SDave Hansen struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h, 1610099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 1611099730d6SDave Hansen { 1612aaf14e40SMichal Hocko struct page *page; 1613aaf14e40SMichal Hocko struct mempolicy *mpol; 1614aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 1615aaf14e40SMichal Hocko int nid; 1616aaf14e40SMichal Hocko nodemask_t *nodemask; 1617aaf14e40SMichal Hocko 1618aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 1619aaf14e40SMichal Hocko page = __alloc_buddy_huge_page(h, gfp_mask, nid, nodemask); 1620aaf14e40SMichal Hocko mpol_cond_put(mpol); 1621aaf14e40SMichal Hocko 1622aaf14e40SMichal Hocko return page; 1623099730d6SDave Hansen } 1624099730d6SDave Hansen 1625099730d6SDave Hansen /* 1626bf50bab2SNaoya Horiguchi * This allocation function is useful in the context where vma is irrelevant. 1627bf50bab2SNaoya Horiguchi * E.g. soft-offlining uses this function because it only cares physical 1628bf50bab2SNaoya Horiguchi * address of error page. 1629bf50bab2SNaoya Horiguchi */ 1630bf50bab2SNaoya Horiguchi struct page *alloc_huge_page_node(struct hstate *h, int nid) 1631bf50bab2SNaoya Horiguchi { 1632aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 16334ef91848SJoonsoo Kim struct page *page = NULL; 1634bf50bab2SNaoya Horiguchi 1635aaf14e40SMichal Hocko if (nid != NUMA_NO_NODE) 1636aaf14e40SMichal Hocko gfp_mask |= __GFP_THISNODE; 1637aaf14e40SMichal Hocko 1638bf50bab2SNaoya Horiguchi spin_lock(&hugetlb_lock); 16394ef91848SJoonsoo Kim if (h->free_huge_pages - h->resv_huge_pages > 0) 16403e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL); 1641bf50bab2SNaoya Horiguchi spin_unlock(&hugetlb_lock); 1642bf50bab2SNaoya Horiguchi 164394ae8ba7SAneesh Kumar K.V if (!page) 1644aaf14e40SMichal Hocko page = __alloc_buddy_huge_page(h, gfp_mask, nid, NULL); 1645bf50bab2SNaoya Horiguchi 1646bf50bab2SNaoya Horiguchi return page; 1647bf50bab2SNaoya Horiguchi } 1648bf50bab2SNaoya Horiguchi 16493e59fcb0SMichal Hocko 16503e59fcb0SMichal Hocko struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, 16513e59fcb0SMichal Hocko nodemask_t *nmask) 16524db9b2efSMichal Hocko { 1653aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 16544db9b2efSMichal Hocko 16554db9b2efSMichal Hocko spin_lock(&hugetlb_lock); 16564db9b2efSMichal Hocko if (h->free_huge_pages - h->resv_huge_pages > 0) { 16573e59fcb0SMichal Hocko struct page *page; 16583e59fcb0SMichal Hocko 16593e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); 16603e59fcb0SMichal Hocko if (page) { 16613e59fcb0SMichal Hocko spin_unlock(&hugetlb_lock); 16623e59fcb0SMichal Hocko return page; 16634db9b2efSMichal Hocko } 16644db9b2efSMichal Hocko } 16654db9b2efSMichal Hocko spin_unlock(&hugetlb_lock); 16664db9b2efSMichal Hocko 16674db9b2efSMichal Hocko /* No reservations, try to overcommit */ 16683e59fcb0SMichal Hocko 16693e59fcb0SMichal Hocko return __alloc_buddy_huge_page(h, gfp_mask, preferred_nid, nmask); 16704db9b2efSMichal Hocko } 16714db9b2efSMichal Hocko 1672bf50bab2SNaoya Horiguchi /* 167325985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 1674e4e574b7SAdam Litke * of size 'delta'. 1675e4e574b7SAdam Litke */ 1676a5516438SAndi Kleen static int gather_surplus_pages(struct hstate *h, int delta) 1677e4e574b7SAdam Litke { 1678e4e574b7SAdam Litke struct list_head surplus_list; 1679e4e574b7SAdam Litke struct page *page, *tmp; 1680e4e574b7SAdam Litke int ret, i; 1681e4e574b7SAdam Litke int needed, allocated; 168228073b02SHillf Danton bool alloc_ok = true; 1683e4e574b7SAdam Litke 1684a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 1685ac09b3a1SAdam Litke if (needed <= 0) { 1686a5516438SAndi Kleen h->resv_huge_pages += delta; 1687e4e574b7SAdam Litke return 0; 1688ac09b3a1SAdam Litke } 1689e4e574b7SAdam Litke 1690e4e574b7SAdam Litke allocated = 0; 1691e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 1692e4e574b7SAdam Litke 1693e4e574b7SAdam Litke ret = -ENOMEM; 1694e4e574b7SAdam Litke retry: 1695e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 1696e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 1697aaf14e40SMichal Hocko page = __alloc_buddy_huge_page(h, htlb_alloc_mask(h), 1698aaf14e40SMichal Hocko NUMA_NO_NODE, NULL); 169928073b02SHillf Danton if (!page) { 170028073b02SHillf Danton alloc_ok = false; 170128073b02SHillf Danton break; 170228073b02SHillf Danton } 1703e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 170469ed779aSDavid Rientjes cond_resched(); 1705e4e574b7SAdam Litke } 170628073b02SHillf Danton allocated += i; 1707e4e574b7SAdam Litke 1708e4e574b7SAdam Litke /* 1709e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 1710e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 1711e4e574b7SAdam Litke */ 1712e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 1713a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 1714a5516438SAndi Kleen (h->free_huge_pages + allocated); 171528073b02SHillf Danton if (needed > 0) { 171628073b02SHillf Danton if (alloc_ok) 1717e4e574b7SAdam Litke goto retry; 171828073b02SHillf Danton /* 171928073b02SHillf Danton * We were not able to allocate enough pages to 172028073b02SHillf Danton * satisfy the entire reservation so we free what 172128073b02SHillf Danton * we've allocated so far. 172228073b02SHillf Danton */ 172328073b02SHillf Danton goto free; 172428073b02SHillf Danton } 1725e4e574b7SAdam Litke /* 1726e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 172725985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 1728e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 1729ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 1730ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 1731ac09b3a1SAdam Litke * before they are reserved. 1732e4e574b7SAdam Litke */ 1733e4e574b7SAdam Litke needed += allocated; 1734a5516438SAndi Kleen h->resv_huge_pages += delta; 1735e4e574b7SAdam Litke ret = 0; 1736a9869b83SNaoya Horiguchi 173719fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 173819fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 173919fc3f0aSAdam Litke if ((--needed) < 0) 174019fc3f0aSAdam Litke break; 1741a9869b83SNaoya Horiguchi /* 1742a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 1743a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 1744a9869b83SNaoya Horiguchi */ 1745a9869b83SNaoya Horiguchi put_page_testzero(page); 1746309381feSSasha Levin VM_BUG_ON_PAGE(page_count(page), page); 1747a5516438SAndi Kleen enqueue_huge_page(h, page); 174819fc3f0aSAdam Litke } 174928073b02SHillf Danton free: 1750b0365c8dSHillf Danton spin_unlock(&hugetlb_lock); 175119fc3f0aSAdam Litke 175219fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 1753c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 1754a9869b83SNaoya Horiguchi put_page(page); 175519fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 1756e4e574b7SAdam Litke 1757e4e574b7SAdam Litke return ret; 1758e4e574b7SAdam Litke } 1759e4e574b7SAdam Litke 1760e4e574b7SAdam Litke /* 1761e5bbc8a6SMike Kravetz * This routine has two main purposes: 1762e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 1763e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 1764e5bbc8a6SMike Kravetz * to the associated reservation map. 1765e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 1766e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 1767e5bbc8a6SMike Kravetz * 1768e5bbc8a6SMike Kravetz * Called with hugetlb_lock held. However, the lock could be dropped (and 1769e5bbc8a6SMike Kravetz * reacquired) during calls to cond_resched_lock. Whenever dropping the lock, 1770e5bbc8a6SMike Kravetz * we must make sure nobody else can claim pages we are in the process of 1771e5bbc8a6SMike Kravetz * freeing. Do this by ensuring resv_huge_page always is greater than the 1772e5bbc8a6SMike Kravetz * number of huge pages we plan to free when dropping the lock. 1773e4e574b7SAdam Litke */ 1774a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 1775a5516438SAndi Kleen unsigned long unused_resv_pages) 1776e4e574b7SAdam Litke { 1777e4e574b7SAdam Litke unsigned long nr_pages; 1778e4e574b7SAdam Litke 1779aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 1780bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1781e5bbc8a6SMike Kravetz goto out; 1782aa888a74SAndi Kleen 1783e5bbc8a6SMike Kravetz /* 1784e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 1785e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 1786e5bbc8a6SMike Kravetz */ 1787a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 1788e4e574b7SAdam Litke 1789685f3457SLee Schermerhorn /* 1790685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 17919b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 17929b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 17939b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 17949b5e5d0fSLee Schermerhorn * free_pool_huge_page() will balance the the freed pages across the 17959b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 1796e5bbc8a6SMike Kravetz * 1797e5bbc8a6SMike Kravetz * Note that we decrement resv_huge_pages as we free the pages. If 1798e5bbc8a6SMike Kravetz * we drop the lock, resv_huge_pages will still be sufficiently large 1799e5bbc8a6SMike Kravetz * to cover subsequent pages we may free. 1800685f3457SLee Schermerhorn */ 1801685f3457SLee Schermerhorn while (nr_pages--) { 1802e5bbc8a6SMike Kravetz h->resv_huge_pages--; 1803e5bbc8a6SMike Kravetz unused_resv_pages--; 18048cebfcd0SLai Jiangshan if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) 1805e5bbc8a6SMike Kravetz goto out; 18067848a4bfSMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 1807e4e574b7SAdam Litke } 1808e5bbc8a6SMike Kravetz 1809e5bbc8a6SMike Kravetz out: 1810e5bbc8a6SMike Kravetz /* Fully uncommit the reservation */ 1811e5bbc8a6SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 1812e4e574b7SAdam Litke } 1813e4e574b7SAdam Litke 18145e911373SMike Kravetz 1815c37f9fb1SAndy Whitcroft /* 1816feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 18175e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 1818cf3ad20bSMike Kravetz * 1819cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 1820cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 1821cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 1822cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 1823cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 1824feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 1825feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 1826feba16e2SMike Kravetz * is called in such cases. 1827cf3ad20bSMike Kravetz * 1828cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 1829cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 1830cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 1831cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 1832cf3ad20bSMike Kravetz * take appropriate action. 183396b96a96SMike Kravetz * 183496b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 183596b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 183696b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 183796b96a96SMike Kravetz * reservation exists. 1838c37f9fb1SAndy Whitcroft */ 18395e911373SMike Kravetz enum vma_resv_mode { 18405e911373SMike Kravetz VMA_NEEDS_RESV, 18415e911373SMike Kravetz VMA_COMMIT_RESV, 1842feba16e2SMike Kravetz VMA_END_RESV, 184396b96a96SMike Kravetz VMA_ADD_RESV, 18445e911373SMike Kravetz }; 1845cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 1846cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 18475e911373SMike Kravetz enum vma_resv_mode mode) 1848c37f9fb1SAndy Whitcroft { 18494e35f483SJoonsoo Kim struct resv_map *resv; 18504e35f483SJoonsoo Kim pgoff_t idx; 1851cf3ad20bSMike Kravetz long ret; 1852c37f9fb1SAndy Whitcroft 18534e35f483SJoonsoo Kim resv = vma_resv_map(vma); 18544e35f483SJoonsoo Kim if (!resv) 1855c37f9fb1SAndy Whitcroft return 1; 1856c37f9fb1SAndy Whitcroft 18574e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 18585e911373SMike Kravetz switch (mode) { 18595e911373SMike Kravetz case VMA_NEEDS_RESV: 1860cf3ad20bSMike Kravetz ret = region_chg(resv, idx, idx + 1); 18615e911373SMike Kravetz break; 18625e911373SMike Kravetz case VMA_COMMIT_RESV: 18635e911373SMike Kravetz ret = region_add(resv, idx, idx + 1); 18645e911373SMike Kravetz break; 1865feba16e2SMike Kravetz case VMA_END_RESV: 18665e911373SMike Kravetz region_abort(resv, idx, idx + 1); 18675e911373SMike Kravetz ret = 0; 18685e911373SMike Kravetz break; 186996b96a96SMike Kravetz case VMA_ADD_RESV: 187096b96a96SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) 187196b96a96SMike Kravetz ret = region_add(resv, idx, idx + 1); 187296b96a96SMike Kravetz else { 187396b96a96SMike Kravetz region_abort(resv, idx, idx + 1); 187496b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 187596b96a96SMike Kravetz } 187696b96a96SMike Kravetz break; 18775e911373SMike Kravetz default: 18785e911373SMike Kravetz BUG(); 18795e911373SMike Kravetz } 188084afd99bSAndy Whitcroft 18814e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) 1882cf3ad20bSMike Kravetz return ret; 188367961f9dSMike Kravetz else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { 188467961f9dSMike Kravetz /* 188567961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 188667961f9dSMike Kravetz * However, a file associated with mapping could have been 188767961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 188867961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 188967961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 189067961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 189167961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 189267961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 189367961f9dSMike Kravetz * reservation has already been consumed. As a result, the 189467961f9dSMike Kravetz * return value of this routine is the opposite of the 189567961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 189667961f9dSMike Kravetz */ 189767961f9dSMike Kravetz if (ret) 189867961f9dSMike Kravetz return 0; 189967961f9dSMike Kravetz else 190067961f9dSMike Kravetz return 1; 190167961f9dSMike Kravetz } 19024e35f483SJoonsoo Kim else 1903cf3ad20bSMike Kravetz return ret < 0 ? ret : 0; 190484afd99bSAndy Whitcroft } 1905cf3ad20bSMike Kravetz 1906cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 1907a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 1908c37f9fb1SAndy Whitcroft { 19095e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 1910cf3ad20bSMike Kravetz } 1911c37f9fb1SAndy Whitcroft 1912cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 1913cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 1914cf3ad20bSMike Kravetz { 19155e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 19165e911373SMike Kravetz } 19175e911373SMike Kravetz 1918feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 19195e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 19205e911373SMike Kravetz { 1921feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 1922c37f9fb1SAndy Whitcroft } 1923c37f9fb1SAndy Whitcroft 192496b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 192596b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 192696b96a96SMike Kravetz { 192796b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 192896b96a96SMike Kravetz } 192996b96a96SMike Kravetz 193096b96a96SMike Kravetz /* 193196b96a96SMike Kravetz * This routine is called to restore a reservation on error paths. In the 193296b96a96SMike Kravetz * specific error paths, a huge page was allocated (via alloc_huge_page) 193396b96a96SMike Kravetz * and is about to be freed. If a reservation for the page existed, 193496b96a96SMike Kravetz * alloc_huge_page would have consumed the reservation and set PagePrivate 193596b96a96SMike Kravetz * in the newly allocated page. When the page is freed via free_huge_page, 193696b96a96SMike Kravetz * the global reservation count will be incremented if PagePrivate is set. 193796b96a96SMike Kravetz * However, free_huge_page can not adjust the reserve map. Adjust the 193896b96a96SMike Kravetz * reserve map here to be consistent with global reserve count adjustments 193996b96a96SMike Kravetz * to be made by free_huge_page. 194096b96a96SMike Kravetz */ 194196b96a96SMike Kravetz static void restore_reserve_on_error(struct hstate *h, 194296b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long address, 194396b96a96SMike Kravetz struct page *page) 194496b96a96SMike Kravetz { 194596b96a96SMike Kravetz if (unlikely(PagePrivate(page))) { 194696b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 194796b96a96SMike Kravetz 194896b96a96SMike Kravetz if (unlikely(rc < 0)) { 194996b96a96SMike Kravetz /* 195096b96a96SMike Kravetz * Rare out of memory condition in reserve map 195196b96a96SMike Kravetz * manipulation. Clear PagePrivate so that 195296b96a96SMike Kravetz * global reserve count will not be incremented 195396b96a96SMike Kravetz * by free_huge_page. This will make it appear 195496b96a96SMike Kravetz * as though the reservation for this page was 195596b96a96SMike Kravetz * consumed. This may prevent the task from 195696b96a96SMike Kravetz * faulting in the page at a later time. This 195796b96a96SMike Kravetz * is better than inconsistent global huge page 195896b96a96SMike Kravetz * accounting of reserve counts. 195996b96a96SMike Kravetz */ 196096b96a96SMike Kravetz ClearPagePrivate(page); 196196b96a96SMike Kravetz } else if (rc) { 196296b96a96SMike Kravetz rc = vma_add_reservation(h, vma, address); 196396b96a96SMike Kravetz if (unlikely(rc < 0)) 196496b96a96SMike Kravetz /* 196596b96a96SMike Kravetz * See above comment about rare out of 196696b96a96SMike Kravetz * memory condition. 196796b96a96SMike Kravetz */ 196896b96a96SMike Kravetz ClearPagePrivate(page); 196996b96a96SMike Kravetz } else 197096b96a96SMike Kravetz vma_end_reservation(h, vma, address); 197196b96a96SMike Kravetz } 197296b96a96SMike Kravetz } 197396b96a96SMike Kravetz 197470c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 197504f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 1976348ea204SAdam Litke { 197790481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 1978a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 1979348ea204SAdam Litke struct page *page; 1980d85f69b0SMike Kravetz long map_chg, map_commit; 1981d85f69b0SMike Kravetz long gbl_chg; 19826d76dcf4SAneesh Kumar K.V int ret, idx; 19836d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 19842fc39cecSAdam Litke 19856d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 1986a1e78772SMel Gorman /* 1987d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 1988d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 1989d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 1990a1e78772SMel Gorman */ 1991d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 1992d85f69b0SMike Kravetz if (map_chg < 0) 199376dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 1994d85f69b0SMike Kravetz 1995d85f69b0SMike Kravetz /* 1996d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 1997d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 1998d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 1999d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 2000d85f69b0SMike Kravetz * checked against any subpool limit. 2001d85f69b0SMike Kravetz */ 2002d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 2003d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 2004d85f69b0SMike Kravetz if (gbl_chg < 0) { 2005feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 200676dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 20075e911373SMike Kravetz } 200890d8b7e6SAdam Litke 2009d85f69b0SMike Kravetz /* 2010d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 2011d85f69b0SMike Kravetz * map, there could be reservations associated with the 2012d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 2013d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 2014d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 2015d85f69b0SMike Kravetz * the subpool reservations. 2016d85f69b0SMike Kravetz */ 2017d85f69b0SMike Kravetz if (avoid_reserve) 2018d85f69b0SMike Kravetz gbl_chg = 1; 2019d85f69b0SMike Kravetz } 2020d85f69b0SMike Kravetz 20216d76dcf4SAneesh Kumar K.V ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 20228f34af6fSJianyu Zhan if (ret) 20238f34af6fSJianyu Zhan goto out_subpool_put; 20248f34af6fSJianyu Zhan 2025a1e78772SMel Gorman spin_lock(&hugetlb_lock); 2026d85f69b0SMike Kravetz /* 2027d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 2028d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 2029d85f69b0SMike Kravetz * a reservation exists for the allocation. 2030d85f69b0SMike Kravetz */ 2031d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 203281a6fcaeSJoonsoo Kim if (!page) { 203394ae8ba7SAneesh Kumar K.V spin_unlock(&hugetlb_lock); 2034099730d6SDave Hansen page = __alloc_buddy_huge_page_with_mpol(h, vma, addr); 20358f34af6fSJianyu Zhan if (!page) 20368f34af6fSJianyu Zhan goto out_uncharge_cgroup; 2037a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 2038a88c7695SNaoya Horiguchi SetPagePrivate(page); 2039a88c7695SNaoya Horiguchi h->resv_huge_pages--; 2040a88c7695SNaoya Horiguchi } 204179dbb236SAneesh Kumar K.V spin_lock(&hugetlb_lock); 204279dbb236SAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 204381a6fcaeSJoonsoo Kim /* Fall through */ 2044a1e78772SMel Gorman } 204581a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 204681a6fcaeSJoonsoo Kim spin_unlock(&hugetlb_lock); 2047a1e78772SMel Gorman 204890481622SDavid Gibson set_page_private(page, (unsigned long)spool); 2049a1e78772SMel Gorman 2050d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 2051d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 205233039678SMike Kravetz /* 205333039678SMike Kravetz * The page was added to the reservation map between 205433039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 205533039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 205633039678SMike Kravetz * Adjust for the subpool count incremented above AND 205733039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 205833039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 205933039678SMike Kravetz * no longer applies. 206033039678SMike Kravetz */ 206133039678SMike Kravetz long rsv_adjust; 206233039678SMike Kravetz 206333039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 206433039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 206533039678SMike Kravetz } 20667893d1d5SAdam Litke return page; 20678f34af6fSJianyu Zhan 20688f34af6fSJianyu Zhan out_uncharge_cgroup: 20698f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 20708f34af6fSJianyu Zhan out_subpool_put: 2071d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 20728f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 2073feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 20748f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 2075b45b5bd6SDavid Gibson } 2076b45b5bd6SDavid Gibson 207774060e4dSNaoya Horiguchi /* 207874060e4dSNaoya Horiguchi * alloc_huge_page()'s wrapper which simply returns the page if allocation 207974060e4dSNaoya Horiguchi * succeeds, otherwise NULL. This function is called from new_vma_page(), 208074060e4dSNaoya Horiguchi * where no ERR_VALUE is expected to be returned. 208174060e4dSNaoya Horiguchi */ 208274060e4dSNaoya Horiguchi struct page *alloc_huge_page_noerr(struct vm_area_struct *vma, 208374060e4dSNaoya Horiguchi unsigned long addr, int avoid_reserve) 208474060e4dSNaoya Horiguchi { 208574060e4dSNaoya Horiguchi struct page *page = alloc_huge_page(vma, addr, avoid_reserve); 208674060e4dSNaoya Horiguchi if (IS_ERR(page)) 208774060e4dSNaoya Horiguchi page = NULL; 208874060e4dSNaoya Horiguchi return page; 208974060e4dSNaoya Horiguchi } 209074060e4dSNaoya Horiguchi 2091e24a1307SAneesh Kumar K.V int alloc_bootmem_huge_page(struct hstate *h) 2092e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 2093e24a1307SAneesh Kumar K.V int __alloc_bootmem_huge_page(struct hstate *h) 2094aa888a74SAndi Kleen { 2095aa888a74SAndi Kleen struct huge_bootmem_page *m; 2096b2261026SJoonsoo Kim int nr_nodes, node; 2097aa888a74SAndi Kleen 2098b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 2099aa888a74SAndi Kleen void *addr; 2100aa888a74SAndi Kleen 21018b89a116SGrygorii Strashko addr = memblock_virt_alloc_try_nid_nopanic( 21028b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 21038b89a116SGrygorii Strashko 0, BOOTMEM_ALLOC_ACCESSIBLE, node); 2104aa888a74SAndi Kleen if (addr) { 2105aa888a74SAndi Kleen /* 2106aa888a74SAndi Kleen * Use the beginning of the huge page to store the 2107aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 2108aa888a74SAndi Kleen * puts them into the mem_map). 2109aa888a74SAndi Kleen */ 2110aa888a74SAndi Kleen m = addr; 2111aa888a74SAndi Kleen goto found; 2112aa888a74SAndi Kleen } 2113aa888a74SAndi Kleen } 2114aa888a74SAndi Kleen return 0; 2115aa888a74SAndi Kleen 2116aa888a74SAndi Kleen found: 2117df994eadSLuiz Capitulino BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); 2118aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 2119aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 2120aa888a74SAndi Kleen m->hstate = h; 2121aa888a74SAndi Kleen return 1; 2122aa888a74SAndi Kleen } 2123aa888a74SAndi Kleen 2124d00181b9SKirill A. Shutemov static void __init prep_compound_huge_page(struct page *page, 2125d00181b9SKirill A. Shutemov unsigned int order) 212618229df5SAndy Whitcroft { 212718229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 212818229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 212918229df5SAndy Whitcroft else 213018229df5SAndy Whitcroft prep_compound_page(page, order); 213118229df5SAndy Whitcroft } 213218229df5SAndy Whitcroft 2133aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 2134aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 2135aa888a74SAndi Kleen { 2136aa888a74SAndi Kleen struct huge_bootmem_page *m; 2137aa888a74SAndi Kleen 2138aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 2139aa888a74SAndi Kleen struct hstate *h = m->hstate; 2140ee8f248dSBecky Bruce struct page *page; 2141ee8f248dSBecky Bruce 2142ee8f248dSBecky Bruce #ifdef CONFIG_HIGHMEM 2143ee8f248dSBecky Bruce page = pfn_to_page(m->phys >> PAGE_SHIFT); 21448b89a116SGrygorii Strashko memblock_free_late(__pa(m), 2145ee8f248dSBecky Bruce sizeof(struct huge_bootmem_page)); 2146ee8f248dSBecky Bruce #else 2147ee8f248dSBecky Bruce page = virt_to_page(m); 2148ee8f248dSBecky Bruce #endif 2149aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 215018229df5SAndy Whitcroft prep_compound_huge_page(page, h->order); 2151ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 2152aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 2153b0320c7bSRafael Aquini /* 2154b0320c7bSRafael Aquini * If we had gigantic hugepages allocated at boot time, we need 2155b0320c7bSRafael Aquini * to restore the 'stolen' pages to totalram_pages in order to 2156b0320c7bSRafael Aquini * fix confusing memory reports from free(1) and another 2157b0320c7bSRafael Aquini * side-effects, like CommitLimit going negative. 2158b0320c7bSRafael Aquini */ 2159bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 21603dcc0571SJiang Liu adjust_managed_page_count(page, 1 << h->order); 2161aa888a74SAndi Kleen } 2162aa888a74SAndi Kleen } 2163aa888a74SAndi Kleen 21648faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 21651da177e4SLinus Torvalds { 21661da177e4SLinus Torvalds unsigned long i; 21671da177e4SLinus Torvalds 2168e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 2169bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 2170aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 2171aa888a74SAndi Kleen break; 21729b5e5d0fSLee Schermerhorn } else if (!alloc_fresh_huge_page(h, 21738cebfcd0SLai Jiangshan &node_states[N_MEMORY])) 21741da177e4SLinus Torvalds break; 217569ed779aSDavid Rientjes cond_resched(); 21761da177e4SLinus Torvalds } 2177d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 2178d715cf80SLiam R. Howlett char buf[32]; 2179d715cf80SLiam R. Howlett 2180c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2181d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 2182d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 21838faa8b07SAndi Kleen h->max_huge_pages = i; 2184e5ff2159SAndi Kleen } 2185d715cf80SLiam R. Howlett } 2186e5ff2159SAndi Kleen 2187e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 2188e5ff2159SAndi Kleen { 2189e5ff2159SAndi Kleen struct hstate *h; 2190e5ff2159SAndi Kleen 2191e5ff2159SAndi Kleen for_each_hstate(h) { 2192641844f5SNaoya Horiguchi if (minimum_order > huge_page_order(h)) 2193641844f5SNaoya Horiguchi minimum_order = huge_page_order(h); 2194641844f5SNaoya Horiguchi 21958faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 2196bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 21978faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 2198e5ff2159SAndi Kleen } 2199641844f5SNaoya Horiguchi VM_BUG_ON(minimum_order == UINT_MAX); 2200e5ff2159SAndi Kleen } 2201e5ff2159SAndi Kleen 2202e5ff2159SAndi Kleen static void __init report_hugepages(void) 2203e5ff2159SAndi Kleen { 2204e5ff2159SAndi Kleen struct hstate *h; 2205e5ff2159SAndi Kleen 2206e5ff2159SAndi Kleen for_each_hstate(h) { 22074abd32dbSAndi Kleen char buf[32]; 2208c6247f72SMatthew Wilcox 2209c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2210ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 2211c6247f72SMatthew Wilcox buf, h->free_huge_pages); 2212e5ff2159SAndi Kleen } 2213e5ff2159SAndi Kleen } 2214e5ff2159SAndi Kleen 22151da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 22166ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 22176ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 22181da177e4SLinus Torvalds { 22194415cc8dSChristoph Lameter int i; 22204415cc8dSChristoph Lameter 2221bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2222aa888a74SAndi Kleen return; 2223aa888a74SAndi Kleen 22246ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 22251da177e4SLinus Torvalds struct page *page, *next; 2226a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 2227a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 2228a5516438SAndi Kleen if (count >= h->nr_huge_pages) 22296b0c880dSAdam Litke return; 22301da177e4SLinus Torvalds if (PageHighMem(page)) 22311da177e4SLinus Torvalds continue; 22321da177e4SLinus Torvalds list_del(&page->lru); 2233e5ff2159SAndi Kleen update_and_free_page(h, page); 2234a5516438SAndi Kleen h->free_huge_pages--; 2235a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 22361da177e4SLinus Torvalds } 22371da177e4SLinus Torvalds } 22381da177e4SLinus Torvalds } 22391da177e4SLinus Torvalds #else 22406ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 22416ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 22421da177e4SLinus Torvalds { 22431da177e4SLinus Torvalds } 22441da177e4SLinus Torvalds #endif 22451da177e4SLinus Torvalds 224620a0307cSWu Fengguang /* 224720a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 224820a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 224920a0307cSWu Fengguang * Returns 1 if an adjustment was made. 225020a0307cSWu Fengguang */ 22516ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 22526ae11b27SLee Schermerhorn int delta) 225320a0307cSWu Fengguang { 2254b2261026SJoonsoo Kim int nr_nodes, node; 225520a0307cSWu Fengguang 225620a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 225720a0307cSWu Fengguang 2258e8c5c824SLee Schermerhorn if (delta < 0) { 2259b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 2260b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 2261b2261026SJoonsoo Kim goto found; 2262b2261026SJoonsoo Kim } 2263b2261026SJoonsoo Kim } else { 2264b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2265b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 2266b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 2267b2261026SJoonsoo Kim goto found; 2268e8c5c824SLee Schermerhorn } 22699a76db09SLee Schermerhorn } 2270b2261026SJoonsoo Kim return 0; 227120a0307cSWu Fengguang 2272b2261026SJoonsoo Kim found: 227320a0307cSWu Fengguang h->surplus_huge_pages += delta; 2274b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 2275b2261026SJoonsoo Kim return 1; 227620a0307cSWu Fengguang } 227720a0307cSWu Fengguang 2278a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 22796ae11b27SLee Schermerhorn static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, 22806ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 22811da177e4SLinus Torvalds { 22827893d1d5SAdam Litke unsigned long min_count, ret; 22831da177e4SLinus Torvalds 2284944d9fecSLuiz Capitulino if (hstate_is_gigantic(h) && !gigantic_page_supported()) 2285aa888a74SAndi Kleen return h->max_huge_pages; 2286aa888a74SAndi Kleen 22877893d1d5SAdam Litke /* 22887893d1d5SAdam Litke * Increase the pool size 22897893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 22907893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 2291d1c3fb1fSNishanth Aravamudan * 2292d15c7c09SNaoya Horiguchi * We might race with __alloc_buddy_huge_page() here and be unable 2293d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 2294d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 2295d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 2296d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 22977893d1d5SAdam Litke */ 22981da177e4SLinus Torvalds spin_lock(&hugetlb_lock); 2299a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 23006ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 23017893d1d5SAdam Litke break; 23027893d1d5SAdam Litke } 23037893d1d5SAdam Litke 2304a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 23057893d1d5SAdam Litke /* 23067893d1d5SAdam Litke * If this allocation races such that we no longer need the 23077893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 23087893d1d5SAdam Litke * and reducing the surplus. 23097893d1d5SAdam Litke */ 23107893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 2311649920c6SJia He 2312649920c6SJia He /* yield cpu to avoid soft lockup */ 2313649920c6SJia He cond_resched(); 2314649920c6SJia He 2315944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) 2316944d9fecSLuiz Capitulino ret = alloc_fresh_gigantic_page(h, nodes_allowed); 2317944d9fecSLuiz Capitulino else 23186ae11b27SLee Schermerhorn ret = alloc_fresh_huge_page(h, nodes_allowed); 23197893d1d5SAdam Litke spin_lock(&hugetlb_lock); 23207893d1d5SAdam Litke if (!ret) 23217893d1d5SAdam Litke goto out; 23227893d1d5SAdam Litke 2323536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 2324536240f2SMel Gorman if (signal_pending(current)) 2325536240f2SMel Gorman goto out; 23267893d1d5SAdam Litke } 23277893d1d5SAdam Litke 23287893d1d5SAdam Litke /* 23297893d1d5SAdam Litke * Decrease the pool size 23307893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 23317893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 23327893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 23337893d1d5SAdam Litke * to the desired size as pages become free. 2334d1c3fb1fSNishanth Aravamudan * 2335d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 2336d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 2337d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 2338d15c7c09SNaoya Horiguchi * __alloc_buddy_huge_page() is checking the global counter, 2339d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 2340d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 2341d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 23427893d1d5SAdam Litke */ 2343a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 23446b0c880dSAdam Litke min_count = max(count, min_count); 23456ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 2346a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 23476ae11b27SLee Schermerhorn if (!free_pool_huge_page(h, nodes_allowed, 0)) 23481da177e4SLinus Torvalds break; 234955f67141SMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 23501da177e4SLinus Torvalds } 2351a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 23526ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 23537893d1d5SAdam Litke break; 23547893d1d5SAdam Litke } 23557893d1d5SAdam Litke out: 2356a5516438SAndi Kleen ret = persistent_huge_pages(h); 23571da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 23587893d1d5SAdam Litke return ret; 23591da177e4SLinus Torvalds } 23601da177e4SLinus Torvalds 2361a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 2362a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 2363a3437870SNishanth Aravamudan 2364a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 2365a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 2366a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 2367a3437870SNishanth Aravamudan 2368a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 2369a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 2370a3437870SNishanth Aravamudan 23719a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 23729a305230SLee Schermerhorn 23739a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 2374a3437870SNishanth Aravamudan { 2375a3437870SNishanth Aravamudan int i; 23769a305230SLee Schermerhorn 2377a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 23789a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 23799a305230SLee Schermerhorn if (nidp) 23809a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 2381a3437870SNishanth Aravamudan return &hstates[i]; 23829a305230SLee Schermerhorn } 23839a305230SLee Schermerhorn 23849a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 2385a3437870SNishanth Aravamudan } 2386a3437870SNishanth Aravamudan 238706808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 2388a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2389a3437870SNishanth Aravamudan { 23909a305230SLee Schermerhorn struct hstate *h; 23919a305230SLee Schermerhorn unsigned long nr_huge_pages; 23929a305230SLee Schermerhorn int nid; 23939a305230SLee Schermerhorn 23949a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 23959a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 23969a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 23979a305230SLee Schermerhorn else 23989a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 23999a305230SLee Schermerhorn 24009a305230SLee Schermerhorn return sprintf(buf, "%lu\n", nr_huge_pages); 2401a3437870SNishanth Aravamudan } 2402adbe8726SEric B Munson 2403238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 2404238d3c13SDavid Rientjes struct hstate *h, int nid, 2405238d3c13SDavid Rientjes unsigned long count, size_t len) 2406a3437870SNishanth Aravamudan { 2407a3437870SNishanth Aravamudan int err; 2408bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); 2409a3437870SNishanth Aravamudan 2410944d9fecSLuiz Capitulino if (hstate_is_gigantic(h) && !gigantic_page_supported()) { 2411adbe8726SEric B Munson err = -EINVAL; 2412adbe8726SEric B Munson goto out; 2413adbe8726SEric B Munson } 2414adbe8726SEric B Munson 24159a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 24169a305230SLee Schermerhorn /* 24179a305230SLee Schermerhorn * global hstate attribute 24189a305230SLee Schermerhorn */ 24199a305230SLee Schermerhorn if (!(obey_mempolicy && 24209a305230SLee Schermerhorn init_nodemask_of_mempolicy(nodes_allowed))) { 242106808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 24228cebfcd0SLai Jiangshan nodes_allowed = &node_states[N_MEMORY]; 242306808b08SLee Schermerhorn } 24249a305230SLee Schermerhorn } else if (nodes_allowed) { 24259a305230SLee Schermerhorn /* 24269a305230SLee Schermerhorn * per node hstate attribute: adjust count to global, 24279a305230SLee Schermerhorn * but restrict alloc/free to the specified node. 24289a305230SLee Schermerhorn */ 24299a305230SLee Schermerhorn count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 24309a305230SLee Schermerhorn init_nodemask_of_node(nodes_allowed, nid); 24319a305230SLee Schermerhorn } else 24328cebfcd0SLai Jiangshan nodes_allowed = &node_states[N_MEMORY]; 24339a305230SLee Schermerhorn 243406808b08SLee Schermerhorn h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); 2435a3437870SNishanth Aravamudan 24368cebfcd0SLai Jiangshan if (nodes_allowed != &node_states[N_MEMORY]) 243706808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 243806808b08SLee Schermerhorn 243906808b08SLee Schermerhorn return len; 2440adbe8726SEric B Munson out: 2441adbe8726SEric B Munson NODEMASK_FREE(nodes_allowed); 2442adbe8726SEric B Munson return err; 244306808b08SLee Schermerhorn } 244406808b08SLee Schermerhorn 2445238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 2446238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 2447238d3c13SDavid Rientjes size_t len) 2448238d3c13SDavid Rientjes { 2449238d3c13SDavid Rientjes struct hstate *h; 2450238d3c13SDavid Rientjes unsigned long count; 2451238d3c13SDavid Rientjes int nid; 2452238d3c13SDavid Rientjes int err; 2453238d3c13SDavid Rientjes 2454238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 2455238d3c13SDavid Rientjes if (err) 2456238d3c13SDavid Rientjes return err; 2457238d3c13SDavid Rientjes 2458238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 2459238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 2460238d3c13SDavid Rientjes } 2461238d3c13SDavid Rientjes 246206808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 246306808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 246406808b08SLee Schermerhorn { 246506808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 246606808b08SLee Schermerhorn } 246706808b08SLee Schermerhorn 246806808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 246906808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 247006808b08SLee Schermerhorn { 2471238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 2472a3437870SNishanth Aravamudan } 2473a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 2474a3437870SNishanth Aravamudan 247506808b08SLee Schermerhorn #ifdef CONFIG_NUMA 247606808b08SLee Schermerhorn 247706808b08SLee Schermerhorn /* 247806808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 247906808b08SLee Schermerhorn * huge page alloc/free. 248006808b08SLee Schermerhorn */ 248106808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 248206808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 248306808b08SLee Schermerhorn { 248406808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 248506808b08SLee Schermerhorn } 248606808b08SLee Schermerhorn 248706808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 248806808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 248906808b08SLee Schermerhorn { 2490238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 249106808b08SLee Schermerhorn } 249206808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 249306808b08SLee Schermerhorn #endif 249406808b08SLee Schermerhorn 249506808b08SLee Schermerhorn 2496a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 2497a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2498a3437870SNishanth Aravamudan { 24999a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2500a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 2501a3437870SNishanth Aravamudan } 2502adbe8726SEric B Munson 2503a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 2504a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 2505a3437870SNishanth Aravamudan { 2506a3437870SNishanth Aravamudan int err; 2507a3437870SNishanth Aravamudan unsigned long input; 25089a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2509a3437870SNishanth Aravamudan 2510bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2511adbe8726SEric B Munson return -EINVAL; 2512adbe8726SEric B Munson 25133dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 2514a3437870SNishanth Aravamudan if (err) 251573ae31e5SEric B Munson return err; 2516a3437870SNishanth Aravamudan 2517a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 2518a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 2519a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 2520a3437870SNishanth Aravamudan 2521a3437870SNishanth Aravamudan return count; 2522a3437870SNishanth Aravamudan } 2523a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 2524a3437870SNishanth Aravamudan 2525a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 2526a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2527a3437870SNishanth Aravamudan { 25289a305230SLee Schermerhorn struct hstate *h; 25299a305230SLee Schermerhorn unsigned long free_huge_pages; 25309a305230SLee Schermerhorn int nid; 25319a305230SLee Schermerhorn 25329a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 25339a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 25349a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 25359a305230SLee Schermerhorn else 25369a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 25379a305230SLee Schermerhorn 25389a305230SLee Schermerhorn return sprintf(buf, "%lu\n", free_huge_pages); 2539a3437870SNishanth Aravamudan } 2540a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 2541a3437870SNishanth Aravamudan 2542a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 2543a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2544a3437870SNishanth Aravamudan { 25459a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2546a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->resv_huge_pages); 2547a3437870SNishanth Aravamudan } 2548a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 2549a3437870SNishanth Aravamudan 2550a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 2551a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2552a3437870SNishanth Aravamudan { 25539a305230SLee Schermerhorn struct hstate *h; 25549a305230SLee Schermerhorn unsigned long surplus_huge_pages; 25559a305230SLee Schermerhorn int nid; 25569a305230SLee Schermerhorn 25579a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 25589a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 25599a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 25609a305230SLee Schermerhorn else 25619a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 25629a305230SLee Schermerhorn 25639a305230SLee Schermerhorn return sprintf(buf, "%lu\n", surplus_huge_pages); 2564a3437870SNishanth Aravamudan } 2565a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 2566a3437870SNishanth Aravamudan 2567a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 2568a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 2569a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 2570a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 2571a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 2572a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 257306808b08SLee Schermerhorn #ifdef CONFIG_NUMA 257406808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 257506808b08SLee Schermerhorn #endif 2576a3437870SNishanth Aravamudan NULL, 2577a3437870SNishanth Aravamudan }; 2578a3437870SNishanth Aravamudan 257967e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 2580a3437870SNishanth Aravamudan .attrs = hstate_attrs, 2581a3437870SNishanth Aravamudan }; 2582a3437870SNishanth Aravamudan 2583094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 25849a305230SLee Schermerhorn struct kobject **hstate_kobjs, 258567e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 2586a3437870SNishanth Aravamudan { 2587a3437870SNishanth Aravamudan int retval; 2588972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 2589a3437870SNishanth Aravamudan 25909a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 25919a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 2592a3437870SNishanth Aravamudan return -ENOMEM; 2593a3437870SNishanth Aravamudan 25949a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 2595a3437870SNishanth Aravamudan if (retval) 25969a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 2597a3437870SNishanth Aravamudan 2598a3437870SNishanth Aravamudan return retval; 2599a3437870SNishanth Aravamudan } 2600a3437870SNishanth Aravamudan 2601a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 2602a3437870SNishanth Aravamudan { 2603a3437870SNishanth Aravamudan struct hstate *h; 2604a3437870SNishanth Aravamudan int err; 2605a3437870SNishanth Aravamudan 2606a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 2607a3437870SNishanth Aravamudan if (!hugepages_kobj) 2608a3437870SNishanth Aravamudan return; 2609a3437870SNishanth Aravamudan 2610a3437870SNishanth Aravamudan for_each_hstate(h) { 26119a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 26129a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 2613a3437870SNishanth Aravamudan if (err) 2614ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s", h->name); 2615a3437870SNishanth Aravamudan } 2616a3437870SNishanth Aravamudan } 2617a3437870SNishanth Aravamudan 26189a305230SLee Schermerhorn #ifdef CONFIG_NUMA 26199a305230SLee Schermerhorn 26209a305230SLee Schermerhorn /* 26219a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 262210fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 262310fbcf4cSKay Sievers * index of a node device or _hstate == node id. 262410fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 26259a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 26269a305230SLee Schermerhorn */ 26279a305230SLee Schermerhorn struct node_hstate { 26289a305230SLee Schermerhorn struct kobject *hugepages_kobj; 26299a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 26309a305230SLee Schermerhorn }; 2631b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 26329a305230SLee Schermerhorn 26339a305230SLee Schermerhorn /* 263410fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 26359a305230SLee Schermerhorn */ 26369a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 26379a305230SLee Schermerhorn &nr_hugepages_attr.attr, 26389a305230SLee Schermerhorn &free_hugepages_attr.attr, 26399a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 26409a305230SLee Schermerhorn NULL, 26419a305230SLee Schermerhorn }; 26429a305230SLee Schermerhorn 264367e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 26449a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 26459a305230SLee Schermerhorn }; 26469a305230SLee Schermerhorn 26479a305230SLee Schermerhorn /* 264810fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 26499a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 26509a305230SLee Schermerhorn */ 26519a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 26529a305230SLee Schermerhorn { 26539a305230SLee Schermerhorn int nid; 26549a305230SLee Schermerhorn 26559a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 26569a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 26579a305230SLee Schermerhorn int i; 26589a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 26599a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 26609a305230SLee Schermerhorn if (nidp) 26619a305230SLee Schermerhorn *nidp = nid; 26629a305230SLee Schermerhorn return &hstates[i]; 26639a305230SLee Schermerhorn } 26649a305230SLee Schermerhorn } 26659a305230SLee Schermerhorn 26669a305230SLee Schermerhorn BUG(); 26679a305230SLee Schermerhorn return NULL; 26689a305230SLee Schermerhorn } 26699a305230SLee Schermerhorn 26709a305230SLee Schermerhorn /* 267110fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 26729a305230SLee Schermerhorn * No-op if no hstate attributes attached. 26739a305230SLee Schermerhorn */ 26743cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 26759a305230SLee Schermerhorn { 26769a305230SLee Schermerhorn struct hstate *h; 267710fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 26789a305230SLee Schermerhorn 26799a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 26809b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 26819a305230SLee Schermerhorn 2682972dc4deSAneesh Kumar K.V for_each_hstate(h) { 2683972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 2684972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 2685972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 2686972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 2687972dc4deSAneesh Kumar K.V } 26889a305230SLee Schermerhorn } 26899a305230SLee Schermerhorn 26909a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 26919a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 26929a305230SLee Schermerhorn } 26939a305230SLee Schermerhorn 26949a305230SLee Schermerhorn 26959a305230SLee Schermerhorn /* 269610fbcf4cSKay Sievers * Register hstate attributes for a single node device. 26979a305230SLee Schermerhorn * No-op if attributes already registered. 26989a305230SLee Schermerhorn */ 26993cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 27009a305230SLee Schermerhorn { 27019a305230SLee Schermerhorn struct hstate *h; 270210fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 27039a305230SLee Schermerhorn int err; 27049a305230SLee Schermerhorn 27059a305230SLee Schermerhorn if (nhs->hugepages_kobj) 27069a305230SLee Schermerhorn return; /* already allocated */ 27079a305230SLee Schermerhorn 27089a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 270910fbcf4cSKay Sievers &node->dev.kobj); 27109a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 27119a305230SLee Schermerhorn return; 27129a305230SLee Schermerhorn 27139a305230SLee Schermerhorn for_each_hstate(h) { 27149a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 27159a305230SLee Schermerhorn nhs->hstate_kobjs, 27169a305230SLee Schermerhorn &per_node_hstate_attr_group); 27179a305230SLee Schermerhorn if (err) { 2718ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s for node %d\n", 271910fbcf4cSKay Sievers h->name, node->dev.id); 27209a305230SLee Schermerhorn hugetlb_unregister_node(node); 27219a305230SLee Schermerhorn break; 27229a305230SLee Schermerhorn } 27239a305230SLee Schermerhorn } 27249a305230SLee Schermerhorn } 27259a305230SLee Schermerhorn 27269a305230SLee Schermerhorn /* 27279b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 272810fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 272910fbcf4cSKay Sievers * registered their associated device by this time. 27309a305230SLee Schermerhorn */ 27317d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 27329a305230SLee Schermerhorn { 27339a305230SLee Schermerhorn int nid; 27349a305230SLee Schermerhorn 27358cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 27368732794bSWen Congyang struct node *node = node_devices[nid]; 273710fbcf4cSKay Sievers if (node->dev.id == nid) 27389a305230SLee Schermerhorn hugetlb_register_node(node); 27399a305230SLee Schermerhorn } 27409a305230SLee Schermerhorn 27419a305230SLee Schermerhorn /* 274210fbcf4cSKay Sievers * Let the node device driver know we're here so it can 27439a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 27449a305230SLee Schermerhorn */ 27459a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 27469a305230SLee Schermerhorn hugetlb_unregister_node); 27479a305230SLee Schermerhorn } 27489a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 27499a305230SLee Schermerhorn 27509a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 27519a305230SLee Schermerhorn { 27529a305230SLee Schermerhorn BUG(); 27539a305230SLee Schermerhorn if (nidp) 27549a305230SLee Schermerhorn *nidp = -1; 27559a305230SLee Schermerhorn return NULL; 27569a305230SLee Schermerhorn } 27579a305230SLee Schermerhorn 27589a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 27599a305230SLee Schermerhorn 27609a305230SLee Schermerhorn #endif 27619a305230SLee Schermerhorn 2762a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 2763a3437870SNishanth Aravamudan { 27648382d914SDavidlohr Bueso int i; 27658382d914SDavidlohr Bueso 2766457c1b27SNishanth Aravamudan if (!hugepages_supported()) 27670ef89d25SBenjamin Herrenschmidt return 0; 2768a3437870SNishanth Aravamudan 2769e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) { 2770d715cf80SLiam R. Howlett if (default_hstate_size != 0) { 2771d715cf80SLiam R. Howlett pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n", 2772d715cf80SLiam R. Howlett default_hstate_size, HPAGE_SIZE); 2773d715cf80SLiam R. Howlett } 2774d715cf80SLiam R. Howlett 2775e11bfbfcSNick Piggin default_hstate_size = HPAGE_SIZE; 2776e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) 2777a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 2778a3437870SNishanth Aravamudan } 2779972dc4deSAneesh Kumar K.V default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); 2780f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 2781f8b74815SVaishali Thakkar if (!default_hstate.max_huge_pages) 2782e11bfbfcSNick Piggin default_hstate.max_huge_pages = default_hstate_max_huge_pages; 2783f8b74815SVaishali Thakkar } 2784a3437870SNishanth Aravamudan 2785a3437870SNishanth Aravamudan hugetlb_init_hstates(); 2786aa888a74SAndi Kleen gather_bootmem_prealloc(); 2787a3437870SNishanth Aravamudan report_hugepages(); 2788a3437870SNishanth Aravamudan 2789a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 27909a305230SLee Schermerhorn hugetlb_register_all_nodes(); 27917179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 27929a305230SLee Schermerhorn 27938382d914SDavidlohr Bueso #ifdef CONFIG_SMP 27948382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 27958382d914SDavidlohr Bueso #else 27968382d914SDavidlohr Bueso num_fault_mutexes = 1; 27978382d914SDavidlohr Bueso #endif 2798c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 27998382d914SDavidlohr Bueso kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL); 2800c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 28018382d914SDavidlohr Bueso 28028382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 2803c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 2804a3437870SNishanth Aravamudan return 0; 2805a3437870SNishanth Aravamudan } 28063e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 2807a3437870SNishanth Aravamudan 2808a3437870SNishanth Aravamudan /* Should be called on processing a hugepagesz=... option */ 28099fee021dSVaishali Thakkar void __init hugetlb_bad_size(void) 28109fee021dSVaishali Thakkar { 28119fee021dSVaishali Thakkar parsed_valid_hugepagesz = false; 28129fee021dSVaishali Thakkar } 28139fee021dSVaishali Thakkar 2814d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 2815a3437870SNishanth Aravamudan { 2816a3437870SNishanth Aravamudan struct hstate *h; 28178faa8b07SAndi Kleen unsigned long i; 28188faa8b07SAndi Kleen 2819a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 2820598d8091SJoe Perches pr_warn("hugepagesz= specified twice, ignoring\n"); 2821a3437870SNishanth Aravamudan return; 2822a3437870SNishanth Aravamudan } 282347d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 2824a3437870SNishanth Aravamudan BUG_ON(order == 0); 282547d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 2826a3437870SNishanth Aravamudan h->order = order; 2827a3437870SNishanth Aravamudan h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 28288faa8b07SAndi Kleen h->nr_huge_pages = 0; 28298faa8b07SAndi Kleen h->free_huge_pages = 0; 28308faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 28318faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 28320edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 283354f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 283454f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 2835a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 2836a3437870SNishanth Aravamudan huge_page_size(h)/1024); 28378faa8b07SAndi Kleen 2838a3437870SNishanth Aravamudan parsed_hstate = h; 2839a3437870SNishanth Aravamudan } 2840a3437870SNishanth Aravamudan 2841e11bfbfcSNick Piggin static int __init hugetlb_nrpages_setup(char *s) 2842a3437870SNishanth Aravamudan { 2843a3437870SNishanth Aravamudan unsigned long *mhp; 28448faa8b07SAndi Kleen static unsigned long *last_mhp; 2845a3437870SNishanth Aravamudan 28469fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 28479fee021dSVaishali Thakkar pr_warn("hugepages = %s preceded by " 28489fee021dSVaishali Thakkar "an unsupported hugepagesz, ignoring\n", s); 28499fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 28509fee021dSVaishali Thakkar return 1; 28519fee021dSVaishali Thakkar } 2852a3437870SNishanth Aravamudan /* 285347d38344SAneesh Kumar K.V * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, 2854a3437870SNishanth Aravamudan * so this hugepages= parameter goes to the "default hstate". 2855a3437870SNishanth Aravamudan */ 28569fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 2857a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 2858a3437870SNishanth Aravamudan else 2859a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 2860a3437870SNishanth Aravamudan 28618faa8b07SAndi Kleen if (mhp == last_mhp) { 2862598d8091SJoe Perches pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n"); 28638faa8b07SAndi Kleen return 1; 28648faa8b07SAndi Kleen } 28658faa8b07SAndi Kleen 2866a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 2867a3437870SNishanth Aravamudan *mhp = 0; 2868a3437870SNishanth Aravamudan 28698faa8b07SAndi Kleen /* 28708faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 28718faa8b07SAndi Kleen * But we need to allocate >= MAX_ORDER hstates here early to still 28728faa8b07SAndi Kleen * use the bootmem allocator. 28738faa8b07SAndi Kleen */ 287447d38344SAneesh Kumar K.V if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) 28758faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 28768faa8b07SAndi Kleen 28778faa8b07SAndi Kleen last_mhp = mhp; 28788faa8b07SAndi Kleen 2879a3437870SNishanth Aravamudan return 1; 2880a3437870SNishanth Aravamudan } 2881e11bfbfcSNick Piggin __setup("hugepages=", hugetlb_nrpages_setup); 2882e11bfbfcSNick Piggin 2883e11bfbfcSNick Piggin static int __init hugetlb_default_setup(char *s) 2884e11bfbfcSNick Piggin { 2885e11bfbfcSNick Piggin default_hstate_size = memparse(s, &s); 2886e11bfbfcSNick Piggin return 1; 2887e11bfbfcSNick Piggin } 2888e11bfbfcSNick Piggin __setup("default_hugepagesz=", hugetlb_default_setup); 2889a3437870SNishanth Aravamudan 28908a213460SNishanth Aravamudan static unsigned int cpuset_mems_nr(unsigned int *array) 28918a213460SNishanth Aravamudan { 28928a213460SNishanth Aravamudan int node; 28938a213460SNishanth Aravamudan unsigned int nr = 0; 28948a213460SNishanth Aravamudan 28958a213460SNishanth Aravamudan for_each_node_mask(node, cpuset_current_mems_allowed) 28968a213460SNishanth Aravamudan nr += array[node]; 28978a213460SNishanth Aravamudan 28988a213460SNishanth Aravamudan return nr; 28998a213460SNishanth Aravamudan } 29008a213460SNishanth Aravamudan 29018a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 290206808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 290306808b08SLee Schermerhorn struct ctl_table *table, int write, 290406808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 29051da177e4SLinus Torvalds { 2906e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 2907238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 290808d4a246SMichal Hocko int ret; 2909e5ff2159SAndi Kleen 2910457c1b27SNishanth Aravamudan if (!hugepages_supported()) 291186613628SJan Stancek return -EOPNOTSUPP; 2912457c1b27SNishanth Aravamudan 2913e5ff2159SAndi Kleen table->data = &tmp; 2914e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 291508d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 291608d4a246SMichal Hocko if (ret) 291708d4a246SMichal Hocko goto out; 2918e5ff2159SAndi Kleen 2919238d3c13SDavid Rientjes if (write) 2920238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 2921238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 292208d4a246SMichal Hocko out: 292308d4a246SMichal Hocko return ret; 29241da177e4SLinus Torvalds } 2925396faf03SMel Gorman 292606808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 292706808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 292806808b08SLee Schermerhorn { 292906808b08SLee Schermerhorn 293006808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 293106808b08SLee Schermerhorn buffer, length, ppos); 293206808b08SLee Schermerhorn } 293306808b08SLee Schermerhorn 293406808b08SLee Schermerhorn #ifdef CONFIG_NUMA 293506808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 293606808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 293706808b08SLee Schermerhorn { 293806808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 293906808b08SLee Schermerhorn buffer, length, ppos); 294006808b08SLee Schermerhorn } 294106808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 294206808b08SLee Schermerhorn 2943a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 29448d65af78SAlexey Dobriyan void __user *buffer, 2945a3d0c6aaSNishanth Aravamudan size_t *length, loff_t *ppos) 2946a3d0c6aaSNishanth Aravamudan { 2947a5516438SAndi Kleen struct hstate *h = &default_hstate; 2948e5ff2159SAndi Kleen unsigned long tmp; 294908d4a246SMichal Hocko int ret; 2950e5ff2159SAndi Kleen 2951457c1b27SNishanth Aravamudan if (!hugepages_supported()) 295286613628SJan Stancek return -EOPNOTSUPP; 2953457c1b27SNishanth Aravamudan 2954e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 2955e5ff2159SAndi Kleen 2956bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 2957adbe8726SEric B Munson return -EINVAL; 2958adbe8726SEric B Munson 2959e5ff2159SAndi Kleen table->data = &tmp; 2960e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 296108d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 296208d4a246SMichal Hocko if (ret) 296308d4a246SMichal Hocko goto out; 2964e5ff2159SAndi Kleen 2965e5ff2159SAndi Kleen if (write) { 2966064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 2967e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 2968a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 2969e5ff2159SAndi Kleen } 297008d4a246SMichal Hocko out: 297108d4a246SMichal Hocko return ret; 2972a3d0c6aaSNishanth Aravamudan } 2973a3d0c6aaSNishanth Aravamudan 29741da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 29751da177e4SLinus Torvalds 2976e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 29771da177e4SLinus Torvalds { 2978fcb2b0c5SRoman Gushchin struct hstate *h; 2979fcb2b0c5SRoman Gushchin unsigned long total = 0; 2980fcb2b0c5SRoman Gushchin 2981457c1b27SNishanth Aravamudan if (!hugepages_supported()) 2982457c1b27SNishanth Aravamudan return; 2983fcb2b0c5SRoman Gushchin 2984fcb2b0c5SRoman Gushchin for_each_hstate(h) { 2985fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 2986fcb2b0c5SRoman Gushchin 2987fcb2b0c5SRoman Gushchin total += (PAGE_SIZE << huge_page_order(h)) * count; 2988fcb2b0c5SRoman Gushchin 2989fcb2b0c5SRoman Gushchin if (h == &default_hstate) 2990e1759c21SAlexey Dobriyan seq_printf(m, 29911da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 29921da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 2993b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 29947893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 29954f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 2996fcb2b0c5SRoman Gushchin count, 2997a5516438SAndi Kleen h->free_huge_pages, 2998a5516438SAndi Kleen h->resv_huge_pages, 2999a5516438SAndi Kleen h->surplus_huge_pages, 3000fcb2b0c5SRoman Gushchin (PAGE_SIZE << huge_page_order(h)) / 1024); 3001fcb2b0c5SRoman Gushchin } 3002fcb2b0c5SRoman Gushchin 3003fcb2b0c5SRoman Gushchin seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024); 30041da177e4SLinus Torvalds } 30051da177e4SLinus Torvalds 30061da177e4SLinus Torvalds int hugetlb_report_node_meminfo(int nid, char *buf) 30071da177e4SLinus Torvalds { 3008a5516438SAndi Kleen struct hstate *h = &default_hstate; 3009457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3010457c1b27SNishanth Aravamudan return 0; 30111da177e4SLinus Torvalds return sprintf(buf, 30121da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 3013a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 3014a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 3015a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 3016a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 3017a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 30181da177e4SLinus Torvalds } 30191da177e4SLinus Torvalds 3020949f7ec5SDavid Rientjes void hugetlb_show_meminfo(void) 3021949f7ec5SDavid Rientjes { 3022949f7ec5SDavid Rientjes struct hstate *h; 3023949f7ec5SDavid Rientjes int nid; 3024949f7ec5SDavid Rientjes 3025457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3026457c1b27SNishanth Aravamudan return; 3027457c1b27SNishanth Aravamudan 3028949f7ec5SDavid Rientjes for_each_node_state(nid, N_MEMORY) 3029949f7ec5SDavid Rientjes for_each_hstate(h) 3030949f7ec5SDavid Rientjes pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 3031949f7ec5SDavid Rientjes nid, 3032949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 3033949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 3034949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 3035949f7ec5SDavid Rientjes 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 3036949f7ec5SDavid Rientjes } 3037949f7ec5SDavid Rientjes 30385d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 30395d317b2bSNaoya Horiguchi { 30405d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 30415d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 30425d317b2bSNaoya Horiguchi } 30435d317b2bSNaoya Horiguchi 30441da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 30451da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 30461da177e4SLinus Torvalds { 3047d0028588SWanpeng Li struct hstate *h; 3048d0028588SWanpeng Li unsigned long nr_total_pages = 0; 3049d0028588SWanpeng Li 3050d0028588SWanpeng Li for_each_hstate(h) 3051d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 3052d0028588SWanpeng Li return nr_total_pages; 30531da177e4SLinus Torvalds } 30541da177e4SLinus Torvalds 3055a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 3056fc1b8a73SMel Gorman { 3057fc1b8a73SMel Gorman int ret = -ENOMEM; 3058fc1b8a73SMel Gorman 3059fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 3060fc1b8a73SMel Gorman /* 3061fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 3062fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 3063fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 3064fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 3065fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 3066fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 3067fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 3068fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 3069fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 3070fc1b8a73SMel Gorman * 3071fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 3072fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 3073fc1b8a73SMel Gorman * we fall back to check against current free page availability as 3074fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 3075fc1b8a73SMel Gorman * semantics that cpuset has. 3076fc1b8a73SMel Gorman */ 3077fc1b8a73SMel Gorman if (delta > 0) { 3078a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 3079fc1b8a73SMel Gorman goto out; 3080fc1b8a73SMel Gorman 3081a5516438SAndi Kleen if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 3082a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 3083fc1b8a73SMel Gorman goto out; 3084fc1b8a73SMel Gorman } 3085fc1b8a73SMel Gorman } 3086fc1b8a73SMel Gorman 3087fc1b8a73SMel Gorman ret = 0; 3088fc1b8a73SMel Gorman if (delta < 0) 3089a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 3090fc1b8a73SMel Gorman 3091fc1b8a73SMel Gorman out: 3092fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 3093fc1b8a73SMel Gorman return ret; 3094fc1b8a73SMel Gorman } 3095fc1b8a73SMel Gorman 309684afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 309784afd99bSAndy Whitcroft { 3098f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 309984afd99bSAndy Whitcroft 310084afd99bSAndy Whitcroft /* 310184afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 310284afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 310384afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 310425985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 310584afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 310684afd99bSAndy Whitcroft * new reference here without additional locking. 310784afd99bSAndy Whitcroft */ 31084e35f483SJoonsoo Kim if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 3109f522c3acSJoonsoo Kim kref_get(&resv->refs); 311084afd99bSAndy Whitcroft } 311184afd99bSAndy Whitcroft 3112a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 3113a1e78772SMel Gorman { 3114a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3115f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 311690481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 31174e35f483SJoonsoo Kim unsigned long reserve, start, end; 31181c5ecae3SMike Kravetz long gbl_reserve; 311984afd99bSAndy Whitcroft 31204e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 31214e35f483SJoonsoo Kim return; 31224e35f483SJoonsoo Kim 3123a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 3124a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 312584afd99bSAndy Whitcroft 31264e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 312784afd99bSAndy Whitcroft 3128f031dd27SJoonsoo Kim kref_put(&resv->refs, resv_map_release); 312984afd99bSAndy Whitcroft 31307251ff78SAdam Litke if (reserve) { 31311c5ecae3SMike Kravetz /* 31321c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 31331c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 31341c5ecae3SMike Kravetz */ 31351c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 31361c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 31377251ff78SAdam Litke } 3138a1e78772SMel Gorman } 3139a1e78772SMel Gorman 314031383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 314131383c68SDan Williams { 314231383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 314331383c68SDan Williams return -EINVAL; 314431383c68SDan Williams return 0; 314531383c68SDan Williams } 314631383c68SDan Williams 31471da177e4SLinus Torvalds /* 31481da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 31491da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 31501da177e4SLinus Torvalds * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 31511da177e4SLinus Torvalds * this far. 31521da177e4SLinus Torvalds */ 315311bac800SDave Jiang static int hugetlb_vm_op_fault(struct vm_fault *vmf) 31541da177e4SLinus Torvalds { 31551da177e4SLinus Torvalds BUG(); 3156d0217ac0SNick Piggin return 0; 31571da177e4SLinus Torvalds } 31581da177e4SLinus Torvalds 3159f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 3160d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 316184afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 3162a1e78772SMel Gorman .close = hugetlb_vm_op_close, 316331383c68SDan Williams .split = hugetlb_vm_op_split, 31641da177e4SLinus Torvalds }; 31651da177e4SLinus Torvalds 31661e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 31671e8f889bSDavid Gibson int writable) 316863551ae0SDavid Gibson { 316963551ae0SDavid Gibson pte_t entry; 317063551ae0SDavid Gibson 31711e8f889bSDavid Gibson if (writable) { 3172106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 3173106c992aSGerald Schaefer vma->vm_page_prot))); 317463551ae0SDavid Gibson } else { 3175106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 3176106c992aSGerald Schaefer vma->vm_page_prot)); 317763551ae0SDavid Gibson } 317863551ae0SDavid Gibson entry = pte_mkyoung(entry); 317963551ae0SDavid Gibson entry = pte_mkhuge(entry); 3180d9ed9faaSChris Metcalf entry = arch_make_huge_pte(entry, vma, page, writable); 318163551ae0SDavid Gibson 318263551ae0SDavid Gibson return entry; 318363551ae0SDavid Gibson } 318463551ae0SDavid Gibson 31851e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 31861e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 31871e8f889bSDavid Gibson { 31881e8f889bSDavid Gibson pte_t entry; 31891e8f889bSDavid Gibson 3190106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 319132f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 31924b3073e1SRussell King update_mmu_cache(vma, address, ptep); 31931e8f889bSDavid Gibson } 31941e8f889bSDavid Gibson 3195d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 31964a705fefSNaoya Horiguchi { 31974a705fefSNaoya Horiguchi swp_entry_t swp; 31984a705fefSNaoya Horiguchi 31994a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 3200d5ed7444SAneesh Kumar K.V return false; 32014a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 32024a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_migration_entry(swp)) 3203d5ed7444SAneesh Kumar K.V return true; 32044a705fefSNaoya Horiguchi else 3205d5ed7444SAneesh Kumar K.V return false; 32064a705fefSNaoya Horiguchi } 32074a705fefSNaoya Horiguchi 32084a705fefSNaoya Horiguchi static int is_hugetlb_entry_hwpoisoned(pte_t pte) 32094a705fefSNaoya Horiguchi { 32104a705fefSNaoya Horiguchi swp_entry_t swp; 32114a705fefSNaoya Horiguchi 32124a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 32134a705fefSNaoya Horiguchi return 0; 32144a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 32154a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_hwpoison_entry(swp)) 32164a705fefSNaoya Horiguchi return 1; 32174a705fefSNaoya Horiguchi else 32184a705fefSNaoya Horiguchi return 0; 32194a705fefSNaoya Horiguchi } 32201e8f889bSDavid Gibson 322163551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 322263551ae0SDavid Gibson struct vm_area_struct *vma) 322363551ae0SDavid Gibson { 322463551ae0SDavid Gibson pte_t *src_pte, *dst_pte, entry; 322563551ae0SDavid Gibson struct page *ptepage; 32261c59827dSHugh Dickins unsigned long addr; 32271e8f889bSDavid Gibson int cow; 3228a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3229a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 3230e8569dd2SAndreas Sandberg unsigned long mmun_start; /* For mmu_notifiers */ 3231e8569dd2SAndreas Sandberg unsigned long mmun_end; /* For mmu_notifiers */ 3232e8569dd2SAndreas Sandberg int ret = 0; 32331e8f889bSDavid Gibson 32341e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 323563551ae0SDavid Gibson 3236e8569dd2SAndreas Sandberg mmun_start = vma->vm_start; 3237e8569dd2SAndreas Sandberg mmun_end = vma->vm_end; 3238e8569dd2SAndreas Sandberg if (cow) 3239e8569dd2SAndreas Sandberg mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end); 3240e8569dd2SAndreas Sandberg 3241a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 3242cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 32437868a208SPunit Agrawal src_pte = huge_pte_offset(src, addr, sz); 3244c74df32cSHugh Dickins if (!src_pte) 3245c74df32cSHugh Dickins continue; 3246a5516438SAndi Kleen dst_pte = huge_pte_alloc(dst, addr, sz); 3247e8569dd2SAndreas Sandberg if (!dst_pte) { 3248e8569dd2SAndreas Sandberg ret = -ENOMEM; 3249e8569dd2SAndreas Sandberg break; 3250e8569dd2SAndreas Sandberg } 3251c5c99429SLarry Woodman 3252c5c99429SLarry Woodman /* If the pagetables are shared don't copy or take references */ 3253c5c99429SLarry Woodman if (dst_pte == src_pte) 3254c5c99429SLarry Woodman continue; 3255c5c99429SLarry Woodman 3256cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 3257cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 3258cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 32594a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 32604a705fefSNaoya Horiguchi if (huge_pte_none(entry)) { /* skip none entry */ 32614a705fefSNaoya Horiguchi ; 32624a705fefSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry) || 32634a705fefSNaoya Horiguchi is_hugetlb_entry_hwpoisoned(entry))) { 32644a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 32654a705fefSNaoya Horiguchi 32664a705fefSNaoya Horiguchi if (is_write_migration_entry(swp_entry) && cow) { 32674a705fefSNaoya Horiguchi /* 32684a705fefSNaoya Horiguchi * COW mappings require pages in both 32694a705fefSNaoya Horiguchi * parent and child to be set to read. 32704a705fefSNaoya Horiguchi */ 32714a705fefSNaoya Horiguchi make_migration_entry_read(&swp_entry); 32724a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 3273e5251fd4SPunit Agrawal set_huge_swap_pte_at(src, addr, src_pte, 3274e5251fd4SPunit Agrawal entry, sz); 32754a705fefSNaoya Horiguchi } 3276e5251fd4SPunit Agrawal set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); 32774a705fefSNaoya Horiguchi } else { 327834ee645eSJoerg Roedel if (cow) { 32790f10851eSJérôme Glisse /* 32800f10851eSJérôme Glisse * No need to notify as we are downgrading page 32810f10851eSJérôme Glisse * table protection not changing it to point 32820f10851eSJérôme Glisse * to a new page. 32830f10851eSJérôme Glisse * 32840f10851eSJérôme Glisse * See Documentation/vm/mmu_notifier.txt 32850f10851eSJérôme Glisse */ 32867f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 328734ee645eSJoerg Roedel } 32880253d634SNaoya Horiguchi entry = huge_ptep_get(src_pte); 328963551ae0SDavid Gibson ptepage = pte_page(entry); 329063551ae0SDavid Gibson get_page(ptepage); 329153f9263bSKirill A. Shutemov page_dup_rmap(ptepage, true); 329263551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 32935d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), dst); 32941c59827dSHugh Dickins } 3295cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 3296cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 329763551ae0SDavid Gibson } 329863551ae0SDavid Gibson 3299e8569dd2SAndreas Sandberg if (cow) 3300e8569dd2SAndreas Sandberg mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end); 3301e8569dd2SAndreas Sandberg 3302e8569dd2SAndreas Sandberg return ret; 330363551ae0SDavid Gibson } 330463551ae0SDavid Gibson 330524669e58SAneesh Kumar K.V void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 330624669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 330724669e58SAneesh Kumar K.V struct page *ref_page) 330863551ae0SDavid Gibson { 330963551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 331063551ae0SDavid Gibson unsigned long address; 3311c7546f8fSDavid Gibson pte_t *ptep; 331263551ae0SDavid Gibson pte_t pte; 3313cb900f41SKirill A. Shutemov spinlock_t *ptl; 331463551ae0SDavid Gibson struct page *page; 3315a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3316a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 33172ec74c3eSSagi Grimberg const unsigned long mmun_start = start; /* For mmu_notifiers */ 33182ec74c3eSSagi Grimberg const unsigned long mmun_end = end; /* For mmu_notifiers */ 3319a5516438SAndi Kleen 332063551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 3321a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 3322a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 332363551ae0SDavid Gibson 332407e32661SAneesh Kumar K.V /* 332507e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 332607e32661SAneesh Kumar K.V * to huge page. 332707e32661SAneesh Kumar K.V */ 332807e32661SAneesh Kumar K.V tlb_remove_check_page_size_change(tlb, sz); 332924669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 33302ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 3331569f48b8SHillf Danton address = start; 3332569f48b8SHillf Danton for (; address < end; address += sz) { 33337868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, sz); 3334c7546f8fSDavid Gibson if (!ptep) 3335c7546f8fSDavid Gibson continue; 3336c7546f8fSDavid Gibson 3337cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 333831d49da5SAneesh Kumar K.V if (huge_pmd_unshare(mm, &address, ptep)) { 333931d49da5SAneesh Kumar K.V spin_unlock(ptl); 334031d49da5SAneesh Kumar K.V continue; 334131d49da5SAneesh Kumar K.V } 334239dde65cSChen, Kenneth W 33436629326bSHillf Danton pte = huge_ptep_get(ptep); 334431d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 334531d49da5SAneesh Kumar K.V spin_unlock(ptl); 334631d49da5SAneesh Kumar K.V continue; 334731d49da5SAneesh Kumar K.V } 33486629326bSHillf Danton 33496629326bSHillf Danton /* 33509fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 33519fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 33526629326bSHillf Danton */ 33539fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 33549386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 335531d49da5SAneesh Kumar K.V spin_unlock(ptl); 335631d49da5SAneesh Kumar K.V continue; 33578c4894c6SNaoya Horiguchi } 33586629326bSHillf Danton 33596629326bSHillf Danton page = pte_page(pte); 336004f2cbe3SMel Gorman /* 336104f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 336204f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 336304f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 336404f2cbe3SMel Gorman */ 336504f2cbe3SMel Gorman if (ref_page) { 336631d49da5SAneesh Kumar K.V if (page != ref_page) { 336731d49da5SAneesh Kumar K.V spin_unlock(ptl); 336831d49da5SAneesh Kumar K.V continue; 336931d49da5SAneesh Kumar K.V } 337004f2cbe3SMel Gorman /* 337104f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 337204f2cbe3SMel Gorman * future faults in this VMA will fail rather than 337304f2cbe3SMel Gorman * looking like data was lost 337404f2cbe3SMel Gorman */ 337504f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 337604f2cbe3SMel Gorman } 337704f2cbe3SMel Gorman 3378c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 3379b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 3380106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 33816649a386SKen Chen set_page_dirty(page); 33829e81130bSHillf Danton 33835d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 3384d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 338531d49da5SAneesh Kumar K.V 3386cb900f41SKirill A. Shutemov spin_unlock(ptl); 3387e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 338824669e58SAneesh Kumar K.V /* 338931d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 339024669e58SAneesh Kumar K.V */ 339131d49da5SAneesh Kumar K.V if (ref_page) 339231d49da5SAneesh Kumar K.V break; 3393fe1668aeSChen, Kenneth W } 33942ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 339524669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 33961da177e4SLinus Torvalds } 339763551ae0SDavid Gibson 3398d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 3399d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 3400d833352aSMel Gorman unsigned long end, struct page *ref_page) 3401d833352aSMel Gorman { 3402d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 3403d833352aSMel Gorman 3404d833352aSMel Gorman /* 3405d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 3406d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 3407d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 3408d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 3409d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 3410c8c06efaSDavidlohr Bueso * is to clear it before releasing the i_mmap_rwsem. This works 3411d833352aSMel Gorman * because in the context this is called, the VMA is about to be 3412c8c06efaSDavidlohr Bueso * destroyed and the i_mmap_rwsem is held. 3413d833352aSMel Gorman */ 3414d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 3415d833352aSMel Gorman } 3416d833352aSMel Gorman 3417502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 341804f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 3419502717f4SChen, Kenneth W { 342024669e58SAneesh Kumar K.V struct mm_struct *mm; 342124669e58SAneesh Kumar K.V struct mmu_gather tlb; 342224669e58SAneesh Kumar K.V 342324669e58SAneesh Kumar K.V mm = vma->vm_mm; 342424669e58SAneesh Kumar K.V 34252b047252SLinus Torvalds tlb_gather_mmu(&tlb, mm, start, end); 342624669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 342724669e58SAneesh Kumar K.V tlb_finish_mmu(&tlb, start, end); 3428502717f4SChen, Kenneth W } 3429502717f4SChen, Kenneth W 343004f2cbe3SMel Gorman /* 343104f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 343204f2cbe3SMel Gorman * mappping it owns the reserve page for. The intention is to unmap the page 343304f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 343404f2cbe3SMel Gorman * same region. 343504f2cbe3SMel Gorman */ 34362f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 34372a4b3dedSHarvey Harrison struct page *page, unsigned long address) 343804f2cbe3SMel Gorman { 34397526674dSAdam Litke struct hstate *h = hstate_vma(vma); 344004f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 344104f2cbe3SMel Gorman struct address_space *mapping; 344204f2cbe3SMel Gorman pgoff_t pgoff; 344304f2cbe3SMel Gorman 344404f2cbe3SMel Gorman /* 344504f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 344604f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 344704f2cbe3SMel Gorman */ 34487526674dSAdam Litke address = address & huge_page_mask(h); 344936e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 345036e4f20aSMichal Hocko vma->vm_pgoff; 345193c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 345204f2cbe3SMel Gorman 34534eb2b1dcSMel Gorman /* 34544eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 34554eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 34564eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 34574eb2b1dcSMel Gorman */ 345883cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 34596b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 346004f2cbe3SMel Gorman /* Do not unmap the current VMA */ 346104f2cbe3SMel Gorman if (iter_vma == vma) 346204f2cbe3SMel Gorman continue; 346304f2cbe3SMel Gorman 346404f2cbe3SMel Gorman /* 34652f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 34662f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 34672f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 34682f84a899SMel Gorman */ 34692f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 34702f84a899SMel Gorman continue; 34712f84a899SMel Gorman 34722f84a899SMel Gorman /* 347304f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 347404f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 347504f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 347604f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 347704f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 347804f2cbe3SMel Gorman */ 347904f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 348024669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 348124669e58SAneesh Kumar K.V address + huge_page_size(h), page); 348204f2cbe3SMel Gorman } 348383cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 348404f2cbe3SMel Gorman } 348504f2cbe3SMel Gorman 34860fe6e20bSNaoya Horiguchi /* 34870fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 3488ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 3489ef009b25SMichal Hocko * cannot race with other handlers or page migration. 3490ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 34910fe6e20bSNaoya Horiguchi */ 34921e8f889bSDavid Gibson static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 34933999f52eSAneesh Kumar K.V unsigned long address, pte_t *ptep, 3494cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 34951e8f889bSDavid Gibson { 34963999f52eSAneesh Kumar K.V pte_t pte; 3497a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 34981e8f889bSDavid Gibson struct page *old_page, *new_page; 3499ad4404a2SDavidlohr Bueso int ret = 0, outside_reserve = 0; 35002ec74c3eSSagi Grimberg unsigned long mmun_start; /* For mmu_notifiers */ 35012ec74c3eSSagi Grimberg unsigned long mmun_end; /* For mmu_notifiers */ 35021e8f889bSDavid Gibson 35033999f52eSAneesh Kumar K.V pte = huge_ptep_get(ptep); 35041e8f889bSDavid Gibson old_page = pte_page(pte); 35051e8f889bSDavid Gibson 350604f2cbe3SMel Gorman retry_avoidcopy: 35071e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 35081e8f889bSDavid Gibson * and just make the page writable */ 350937a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 35105a49973dSHugh Dickins page_move_anon_rmap(old_page, vma); 35111e8f889bSDavid Gibson set_huge_ptep_writable(vma, address, ptep); 351283c54070SNick Piggin return 0; 35131e8f889bSDavid Gibson } 35141e8f889bSDavid Gibson 351504f2cbe3SMel Gorman /* 351604f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 351704f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 351804f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 351904f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 352004f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 352104f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 352204f2cbe3SMel Gorman * of the full address range. 352304f2cbe3SMel Gorman */ 35245944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 352504f2cbe3SMel Gorman old_page != pagecache_page) 352604f2cbe3SMel Gorman outside_reserve = 1; 352704f2cbe3SMel Gorman 352809cbfeafSKirill A. Shutemov get_page(old_page); 3529b76c8cfbSLarry Woodman 3530ad4404a2SDavidlohr Bueso /* 3531ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 3532ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 3533ad4404a2SDavidlohr Bueso */ 3534cb900f41SKirill A. Shutemov spin_unlock(ptl); 353504f2cbe3SMel Gorman new_page = alloc_huge_page(vma, address, outside_reserve); 35361e8f889bSDavid Gibson 35372fc39cecSAdam Litke if (IS_ERR(new_page)) { 353804f2cbe3SMel Gorman /* 353904f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 354004f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 354104f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 354204f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 354304f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 354404f2cbe3SMel Gorman */ 354504f2cbe3SMel Gorman if (outside_reserve) { 354609cbfeafSKirill A. Shutemov put_page(old_page); 354704f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 35482f4612afSDavidlohr Bueso unmap_ref_private(mm, vma, old_page, address); 354904f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 3550cb900f41SKirill A. Shutemov spin_lock(ptl); 35517868a208SPunit Agrawal ptep = huge_pte_offset(mm, address & huge_page_mask(h), 35527868a208SPunit Agrawal huge_page_size(h)); 3553a9af0c5dSNaoya Horiguchi if (likely(ptep && 3554a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 355504f2cbe3SMel Gorman goto retry_avoidcopy; 3556a734bcc8SHillf Danton /* 3557cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 3558cb900f41SKirill A. Shutemov * lock, and our job is done. 3559a734bcc8SHillf Danton */ 3560a734bcc8SHillf Danton return 0; 356104f2cbe3SMel Gorman } 356204f2cbe3SMel Gorman 3563ad4404a2SDavidlohr Bueso ret = (PTR_ERR(new_page) == -ENOMEM) ? 3564ad4404a2SDavidlohr Bueso VM_FAULT_OOM : VM_FAULT_SIGBUS; 3565ad4404a2SDavidlohr Bueso goto out_release_old; 35661e8f889bSDavid Gibson } 35671e8f889bSDavid Gibson 35680fe6e20bSNaoya Horiguchi /* 35690fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 35700fe6e20bSNaoya Horiguchi * anon_vma prepared. 35710fe6e20bSNaoya Horiguchi */ 357244e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 3573ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 3574ad4404a2SDavidlohr Bueso goto out_release_all; 357544e2aa93SDean Nelson } 35760fe6e20bSNaoya Horiguchi 357747ad8475SAndrea Arcangeli copy_user_huge_page(new_page, old_page, address, vma, 357847ad8475SAndrea Arcangeli pages_per_huge_page(h)); 35790ed361deSNick Piggin __SetPageUptodate(new_page); 3580bcc54222SNaoya Horiguchi set_page_huge_active(new_page); 35811e8f889bSDavid Gibson 35822ec74c3eSSagi Grimberg mmun_start = address & huge_page_mask(h); 35832ec74c3eSSagi Grimberg mmun_end = mmun_start + huge_page_size(h); 35842ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 3585ad4404a2SDavidlohr Bueso 3586b76c8cfbSLarry Woodman /* 3587cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 3588b76c8cfbSLarry Woodman * before the page tables are altered 3589b76c8cfbSLarry Woodman */ 3590cb900f41SKirill A. Shutemov spin_lock(ptl); 35917868a208SPunit Agrawal ptep = huge_pte_offset(mm, address & huge_page_mask(h), 35927868a208SPunit Agrawal huge_page_size(h)); 3593a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 359407443a85SJoonsoo Kim ClearPagePrivate(new_page); 359507443a85SJoonsoo Kim 35961e8f889bSDavid Gibson /* Break COW */ 35978fe627ecSGerald Schaefer huge_ptep_clear_flush(vma, address, ptep); 359834ee645eSJoerg Roedel mmu_notifier_invalidate_range(mm, mmun_start, mmun_end); 35991e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, 36001e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 3601d281ee61SKirill A. Shutemov page_remove_rmap(old_page, true); 3602cd67f0d2SNaoya Horiguchi hugepage_add_new_anon_rmap(new_page, vma, address); 36031e8f889bSDavid Gibson /* Make the old page be freed below */ 36041e8f889bSDavid Gibson new_page = old_page; 36051e8f889bSDavid Gibson } 3606cb900f41SKirill A. Shutemov spin_unlock(ptl); 36072ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 3608ad4404a2SDavidlohr Bueso out_release_all: 360996b96a96SMike Kravetz restore_reserve_on_error(h, vma, address, new_page); 361009cbfeafSKirill A. Shutemov put_page(new_page); 3611ad4404a2SDavidlohr Bueso out_release_old: 361209cbfeafSKirill A. Shutemov put_page(old_page); 36138312034fSJoonsoo Kim 3614ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 3615ad4404a2SDavidlohr Bueso return ret; 36161e8f889bSDavid Gibson } 36171e8f889bSDavid Gibson 361804f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 3619a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 3620a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 362104f2cbe3SMel Gorman { 362204f2cbe3SMel Gorman struct address_space *mapping; 3623e7c4b0bfSAndy Whitcroft pgoff_t idx; 362404f2cbe3SMel Gorman 362504f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 3626a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 362704f2cbe3SMel Gorman 362804f2cbe3SMel Gorman return find_lock_page(mapping, idx); 362904f2cbe3SMel Gorman } 363004f2cbe3SMel Gorman 36313ae77f43SHugh Dickins /* 36323ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 36333ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 36343ae77f43SHugh Dickins */ 36353ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 36362a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 36372a15efc9SHugh Dickins { 36382a15efc9SHugh Dickins struct address_space *mapping; 36392a15efc9SHugh Dickins pgoff_t idx; 36402a15efc9SHugh Dickins struct page *page; 36412a15efc9SHugh Dickins 36422a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 36432a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 36442a15efc9SHugh Dickins 36452a15efc9SHugh Dickins page = find_get_page(mapping, idx); 36462a15efc9SHugh Dickins if (page) 36472a15efc9SHugh Dickins put_page(page); 36482a15efc9SHugh Dickins return page != NULL; 36492a15efc9SHugh Dickins } 36502a15efc9SHugh Dickins 3651ab76ad54SMike Kravetz int huge_add_to_page_cache(struct page *page, struct address_space *mapping, 3652ab76ad54SMike Kravetz pgoff_t idx) 3653ab76ad54SMike Kravetz { 3654ab76ad54SMike Kravetz struct inode *inode = mapping->host; 3655ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 3656ab76ad54SMike Kravetz int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 3657ab76ad54SMike Kravetz 3658ab76ad54SMike Kravetz if (err) 3659ab76ad54SMike Kravetz return err; 3660ab76ad54SMike Kravetz ClearPagePrivate(page); 3661ab76ad54SMike Kravetz 3662ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 3663ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 3664ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 3665ab76ad54SMike Kravetz return 0; 3666ab76ad54SMike Kravetz } 3667ab76ad54SMike Kravetz 3668a1ed3ddaSRobert P. J. Day static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, 36698382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 3670788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 3671ac9b9c66SHugh Dickins { 3672a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3673ac9b9c66SHugh Dickins int ret = VM_FAULT_SIGBUS; 3674409eb8c2SHillf Danton int anon_rmap = 0; 36754c887265SAdam Litke unsigned long size; 36764c887265SAdam Litke struct page *page; 36771e8f889bSDavid Gibson pte_t new_pte; 3678cb900f41SKirill A. Shutemov spinlock_t *ptl; 36794c887265SAdam Litke 368004f2cbe3SMel Gorman /* 368104f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 368204f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 368325985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 368404f2cbe3SMel Gorman */ 368504f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 3686910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 368704f2cbe3SMel Gorman current->pid); 368804f2cbe3SMel Gorman return ret; 368904f2cbe3SMel Gorman } 369004f2cbe3SMel Gorman 36914c887265SAdam Litke /* 36924c887265SAdam Litke * Use page lock to guard against racing truncation 36934c887265SAdam Litke * before we get page_table_lock. 36944c887265SAdam Litke */ 36956bda666aSChristoph Lameter retry: 36966bda666aSChristoph Lameter page = find_lock_page(mapping, idx); 36976bda666aSChristoph Lameter if (!page) { 3698a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 3699ebed4bfcSHugh Dickins if (idx >= size) 3700ebed4bfcSHugh Dickins goto out; 37011a1aad8aSMike Kravetz 37021a1aad8aSMike Kravetz /* 37031a1aad8aSMike Kravetz * Check for page in userfault range 37041a1aad8aSMike Kravetz */ 37051a1aad8aSMike Kravetz if (userfaultfd_missing(vma)) { 37061a1aad8aSMike Kravetz u32 hash; 37071a1aad8aSMike Kravetz struct vm_fault vmf = { 37081a1aad8aSMike Kravetz .vma = vma, 37091a1aad8aSMike Kravetz .address = address, 37101a1aad8aSMike Kravetz .flags = flags, 37111a1aad8aSMike Kravetz /* 37121a1aad8aSMike Kravetz * Hard to debug if it ends up being 37131a1aad8aSMike Kravetz * used by a callee that assumes 37141a1aad8aSMike Kravetz * something about the other 37151a1aad8aSMike Kravetz * uninitialized fields... same as in 37161a1aad8aSMike Kravetz * memory.c 37171a1aad8aSMike Kravetz */ 37181a1aad8aSMike Kravetz }; 37191a1aad8aSMike Kravetz 37201a1aad8aSMike Kravetz /* 37211a1aad8aSMike Kravetz * hugetlb_fault_mutex must be dropped before 37221a1aad8aSMike Kravetz * handling userfault. Reacquire after handling 37231a1aad8aSMike Kravetz * fault to make calling code simpler. 37241a1aad8aSMike Kravetz */ 37251a1aad8aSMike Kravetz hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, 37261a1aad8aSMike Kravetz idx, address); 37271a1aad8aSMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 37281a1aad8aSMike Kravetz ret = handle_userfault(&vmf, VM_UFFD_MISSING); 37291a1aad8aSMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 37301a1aad8aSMike Kravetz goto out; 37311a1aad8aSMike Kravetz } 37321a1aad8aSMike Kravetz 373304f2cbe3SMel Gorman page = alloc_huge_page(vma, address, 0); 37342fc39cecSAdam Litke if (IS_ERR(page)) { 373576dcee75SAneesh Kumar K.V ret = PTR_ERR(page); 373676dcee75SAneesh Kumar K.V if (ret == -ENOMEM) 373776dcee75SAneesh Kumar K.V ret = VM_FAULT_OOM; 373876dcee75SAneesh Kumar K.V else 373976dcee75SAneesh Kumar K.V ret = VM_FAULT_SIGBUS; 37406bda666aSChristoph Lameter goto out; 37416bda666aSChristoph Lameter } 374247ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 37430ed361deSNick Piggin __SetPageUptodate(page); 3744bcc54222SNaoya Horiguchi set_page_huge_active(page); 3745ac9b9c66SHugh Dickins 3746f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 3747ab76ad54SMike Kravetz int err = huge_add_to_page_cache(page, mapping, idx); 37486bda666aSChristoph Lameter if (err) { 37496bda666aSChristoph Lameter put_page(page); 37506bda666aSChristoph Lameter if (err == -EEXIST) 37516bda666aSChristoph Lameter goto retry; 37526bda666aSChristoph Lameter goto out; 37536bda666aSChristoph Lameter } 375423be7468SMel Gorman } else { 37556bda666aSChristoph Lameter lock_page(page); 37560fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 37570fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 37580fe6e20bSNaoya Horiguchi goto backout_unlocked; 375923be7468SMel Gorman } 3760409eb8c2SHillf Danton anon_rmap = 1; 37610fe6e20bSNaoya Horiguchi } 37620fe6e20bSNaoya Horiguchi } else { 376357303d80SAndy Whitcroft /* 3764998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 3765998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 3766998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 3767fd6a03edSNaoya Horiguchi */ 3768fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 3769aa50d3a7SAndi Kleen ret = VM_FAULT_HWPOISON | 3770972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 3771fd6a03edSNaoya Horiguchi goto backout_unlocked; 37726bda666aSChristoph Lameter } 3773998b4382SNaoya Horiguchi } 37741e8f889bSDavid Gibson 377557303d80SAndy Whitcroft /* 377657303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 377757303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 377857303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 377957303d80SAndy Whitcroft * the spinlock. 378057303d80SAndy Whitcroft */ 37815e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 37822b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 37832b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 37842b26736cSAndy Whitcroft goto backout_unlocked; 37852b26736cSAndy Whitcroft } 37865e911373SMike Kravetz /* Just decrements count, does not deallocate */ 3787feba16e2SMike Kravetz vma_end_reservation(h, vma, address); 37885e911373SMike Kravetz } 378957303d80SAndy Whitcroft 37908bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 3791a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 37924c887265SAdam Litke if (idx >= size) 37934c887265SAdam Litke goto backout; 37944c887265SAdam Litke 379583c54070SNick Piggin ret = 0; 37967f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 37974c887265SAdam Litke goto backout; 37984c887265SAdam Litke 379907443a85SJoonsoo Kim if (anon_rmap) { 380007443a85SJoonsoo Kim ClearPagePrivate(page); 3801409eb8c2SHillf Danton hugepage_add_new_anon_rmap(page, vma, address); 3802ac714904SChoi Gi-yong } else 380353f9263bSKirill A. Shutemov page_dup_rmap(page, true); 38041e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 38051e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 38061e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, new_pte); 38071e8f889bSDavid Gibson 38085d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 3809788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 38101e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 38113999f52eSAneesh Kumar K.V ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); 38121e8f889bSDavid Gibson } 38131e8f889bSDavid Gibson 3814cb900f41SKirill A. Shutemov spin_unlock(ptl); 38154c887265SAdam Litke unlock_page(page); 38164c887265SAdam Litke out: 3817ac9b9c66SHugh Dickins return ret; 38184c887265SAdam Litke 38194c887265SAdam Litke backout: 3820cb900f41SKirill A. Shutemov spin_unlock(ptl); 38212b26736cSAndy Whitcroft backout_unlocked: 38224c887265SAdam Litke unlock_page(page); 382396b96a96SMike Kravetz restore_reserve_on_error(h, vma, address, page); 38244c887265SAdam Litke put_page(page); 38254c887265SAdam Litke goto out; 3826ac9b9c66SHugh Dickins } 3827ac9b9c66SHugh Dickins 38288382d914SDavidlohr Bueso #ifdef CONFIG_SMP 3829c672c7f2SMike Kravetz u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, 38308382d914SDavidlohr Bueso struct vm_area_struct *vma, 38318382d914SDavidlohr Bueso struct address_space *mapping, 38328382d914SDavidlohr Bueso pgoff_t idx, unsigned long address) 38338382d914SDavidlohr Bueso { 38348382d914SDavidlohr Bueso unsigned long key[2]; 38358382d914SDavidlohr Bueso u32 hash; 38368382d914SDavidlohr Bueso 38378382d914SDavidlohr Bueso if (vma->vm_flags & VM_SHARED) { 38388382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 38398382d914SDavidlohr Bueso key[1] = idx; 38408382d914SDavidlohr Bueso } else { 38418382d914SDavidlohr Bueso key[0] = (unsigned long) mm; 38428382d914SDavidlohr Bueso key[1] = address >> huge_page_shift(h); 38438382d914SDavidlohr Bueso } 38448382d914SDavidlohr Bueso 38458382d914SDavidlohr Bueso hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0); 38468382d914SDavidlohr Bueso 38478382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 38488382d914SDavidlohr Bueso } 38498382d914SDavidlohr Bueso #else 38508382d914SDavidlohr Bueso /* 38518382d914SDavidlohr Bueso * For uniprocesor systems we always use a single mutex, so just 38528382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 38538382d914SDavidlohr Bueso */ 3854c672c7f2SMike Kravetz u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, 38558382d914SDavidlohr Bueso struct vm_area_struct *vma, 38568382d914SDavidlohr Bueso struct address_space *mapping, 38578382d914SDavidlohr Bueso pgoff_t idx, unsigned long address) 38588382d914SDavidlohr Bueso { 38598382d914SDavidlohr Bueso return 0; 38608382d914SDavidlohr Bueso } 38618382d914SDavidlohr Bueso #endif 38628382d914SDavidlohr Bueso 386386e5216fSAdam Litke int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 3864788c7df4SHugh Dickins unsigned long address, unsigned int flags) 386586e5216fSAdam Litke { 38668382d914SDavidlohr Bueso pte_t *ptep, entry; 3867cb900f41SKirill A. Shutemov spinlock_t *ptl; 38681e8f889bSDavid Gibson int ret; 38698382d914SDavidlohr Bueso u32 hash; 38708382d914SDavidlohr Bueso pgoff_t idx; 38710fe6e20bSNaoya Horiguchi struct page *page = NULL; 387257303d80SAndy Whitcroft struct page *pagecache_page = NULL; 3873a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 38748382d914SDavidlohr Bueso struct address_space *mapping; 38750f792cf9SNaoya Horiguchi int need_wait_lock = 0; 387686e5216fSAdam Litke 38771e16a539SKAMEZAWA Hiroyuki address &= huge_page_mask(h); 38781e16a539SKAMEZAWA Hiroyuki 38797868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, huge_page_size(h)); 3880fd6a03edSNaoya Horiguchi if (ptep) { 3881fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 3882290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 3883cb900f41SKirill A. Shutemov migration_entry_wait_huge(vma, mm, ptep); 3884290408d4SNaoya Horiguchi return 0; 3885290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 3886aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 3887972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 38880d777df5SNaoya Horiguchi } else { 3889a5516438SAndi Kleen ptep = huge_pte_alloc(mm, address, huge_page_size(h)); 389086e5216fSAdam Litke if (!ptep) 389186e5216fSAdam Litke return VM_FAULT_OOM; 38920d777df5SNaoya Horiguchi } 389386e5216fSAdam Litke 38948382d914SDavidlohr Bueso mapping = vma->vm_file->f_mapping; 38958382d914SDavidlohr Bueso idx = vma_hugecache_offset(h, vma, address); 38968382d914SDavidlohr Bueso 38973935baa9SDavid Gibson /* 38983935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 38993935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 39003935baa9SDavid Gibson * the same page in the page cache. 39013935baa9SDavid Gibson */ 3902c672c7f2SMike Kravetz hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address); 3903c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 39048382d914SDavidlohr Bueso 39057f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 39067f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 39078382d914SDavidlohr Bueso ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 3908b4d1d99fSDavid Gibson goto out_mutex; 39093935baa9SDavid Gibson } 391086e5216fSAdam Litke 391183c54070SNick Piggin ret = 0; 39121e8f889bSDavid Gibson 391357303d80SAndy Whitcroft /* 39140f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 39150f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 39160f792cf9SNaoya Horiguchi * a active hugepage in pagecache. This goto expects the 2nd page fault, 39170f792cf9SNaoya Horiguchi * and is_hugetlb_entry_(migration|hwpoisoned) check will properly 39180f792cf9SNaoya Horiguchi * handle it. 39190f792cf9SNaoya Horiguchi */ 39200f792cf9SNaoya Horiguchi if (!pte_present(entry)) 39210f792cf9SNaoya Horiguchi goto out_mutex; 39220f792cf9SNaoya Horiguchi 39230f792cf9SNaoya Horiguchi /* 392457303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 392557303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 392657303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 392757303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 392857303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 392957303d80SAndy Whitcroft * consumed. 393057303d80SAndy Whitcroft */ 3931106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 39322b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 39332b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 3934b4d1d99fSDavid Gibson goto out_mutex; 39352b26736cSAndy Whitcroft } 39365e911373SMike Kravetz /* Just decrements count, does not deallocate */ 3937feba16e2SMike Kravetz vma_end_reservation(h, vma, address); 393857303d80SAndy Whitcroft 3939f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 394057303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 394157303d80SAndy Whitcroft vma, address); 394257303d80SAndy Whitcroft } 394357303d80SAndy Whitcroft 39440f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 39450fe6e20bSNaoya Horiguchi 39461e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 3947b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 3948cb900f41SKirill A. Shutemov goto out_ptl; 3949b4d1d99fSDavid Gibson 39500f792cf9SNaoya Horiguchi /* 39510f792cf9SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 39520f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 39530f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 39540f792cf9SNaoya Horiguchi */ 39550f792cf9SNaoya Horiguchi page = pte_page(entry); 39560f792cf9SNaoya Horiguchi if (page != pagecache_page) 39570f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 39580f792cf9SNaoya Horiguchi need_wait_lock = 1; 39590f792cf9SNaoya Horiguchi goto out_ptl; 39600f792cf9SNaoya Horiguchi } 39610f792cf9SNaoya Horiguchi 39620f792cf9SNaoya Horiguchi get_page(page); 3963b4d1d99fSDavid Gibson 3964788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 3965106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 39663999f52eSAneesh Kumar K.V ret = hugetlb_cow(mm, vma, address, ptep, 3967cb900f41SKirill A. Shutemov pagecache_page, ptl); 39680f792cf9SNaoya Horiguchi goto out_put_page; 3969b4d1d99fSDavid Gibson } 3970106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 3971b4d1d99fSDavid Gibson } 3972b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 3973788c7df4SHugh Dickins if (huge_ptep_set_access_flags(vma, address, ptep, entry, 3974788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 39754b3073e1SRussell King update_mmu_cache(vma, address, ptep); 39760f792cf9SNaoya Horiguchi out_put_page: 39770f792cf9SNaoya Horiguchi if (page != pagecache_page) 39780f792cf9SNaoya Horiguchi unlock_page(page); 39790f792cf9SNaoya Horiguchi put_page(page); 3980cb900f41SKirill A. Shutemov out_ptl: 3981cb900f41SKirill A. Shutemov spin_unlock(ptl); 398257303d80SAndy Whitcroft 398357303d80SAndy Whitcroft if (pagecache_page) { 398457303d80SAndy Whitcroft unlock_page(pagecache_page); 398557303d80SAndy Whitcroft put_page(pagecache_page); 398657303d80SAndy Whitcroft } 3987b4d1d99fSDavid Gibson out_mutex: 3988c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 39890f792cf9SNaoya Horiguchi /* 39900f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 39910f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 39920f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 39930f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 39940f792cf9SNaoya Horiguchi * here without taking refcount. 39950f792cf9SNaoya Horiguchi */ 39960f792cf9SNaoya Horiguchi if (need_wait_lock) 39970f792cf9SNaoya Horiguchi wait_on_page_locked(page); 39981e8f889bSDavid Gibson return ret; 399986e5216fSAdam Litke } 400086e5216fSAdam Litke 40018fb5debcSMike Kravetz /* 40028fb5debcSMike Kravetz * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with 40038fb5debcSMike Kravetz * modifications for huge pages. 40048fb5debcSMike Kravetz */ 40058fb5debcSMike Kravetz int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, 40068fb5debcSMike Kravetz pte_t *dst_pte, 40078fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 40088fb5debcSMike Kravetz unsigned long dst_addr, 40098fb5debcSMike Kravetz unsigned long src_addr, 40108fb5debcSMike Kravetz struct page **pagep) 40118fb5debcSMike Kravetz { 40121e392147SAndrea Arcangeli struct address_space *mapping; 40131e392147SAndrea Arcangeli pgoff_t idx; 40141e392147SAndrea Arcangeli unsigned long size; 40151c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 40168fb5debcSMike Kravetz struct hstate *h = hstate_vma(dst_vma); 40178fb5debcSMike Kravetz pte_t _dst_pte; 40188fb5debcSMike Kravetz spinlock_t *ptl; 40198fb5debcSMike Kravetz int ret; 40208fb5debcSMike Kravetz struct page *page; 40218fb5debcSMike Kravetz 40228fb5debcSMike Kravetz if (!*pagep) { 40238fb5debcSMike Kravetz ret = -ENOMEM; 40248fb5debcSMike Kravetz page = alloc_huge_page(dst_vma, dst_addr, 0); 40258fb5debcSMike Kravetz if (IS_ERR(page)) 40268fb5debcSMike Kravetz goto out; 40278fb5debcSMike Kravetz 40288fb5debcSMike Kravetz ret = copy_huge_page_from_user(page, 40298fb5debcSMike Kravetz (const void __user *) src_addr, 4030810a56b9SMike Kravetz pages_per_huge_page(h), false); 40318fb5debcSMike Kravetz 40328fb5debcSMike Kravetz /* fallback to copy_from_user outside mmap_sem */ 40338fb5debcSMike Kravetz if (unlikely(ret)) { 40348fb5debcSMike Kravetz ret = -EFAULT; 40358fb5debcSMike Kravetz *pagep = page; 40368fb5debcSMike Kravetz /* don't free the page */ 40378fb5debcSMike Kravetz goto out; 40388fb5debcSMike Kravetz } 40398fb5debcSMike Kravetz } else { 40408fb5debcSMike Kravetz page = *pagep; 40418fb5debcSMike Kravetz *pagep = NULL; 40428fb5debcSMike Kravetz } 40438fb5debcSMike Kravetz 40448fb5debcSMike Kravetz /* 40458fb5debcSMike Kravetz * The memory barrier inside __SetPageUptodate makes sure that 40468fb5debcSMike Kravetz * preceding stores to the page contents become visible before 40478fb5debcSMike Kravetz * the set_pte_at() write. 40488fb5debcSMike Kravetz */ 40498fb5debcSMike Kravetz __SetPageUptodate(page); 40508fb5debcSMike Kravetz set_page_huge_active(page); 40518fb5debcSMike Kravetz 40521e392147SAndrea Arcangeli mapping = dst_vma->vm_file->f_mapping; 40531e392147SAndrea Arcangeli idx = vma_hugecache_offset(h, dst_vma, dst_addr); 40541e392147SAndrea Arcangeli 40551c9e8defSMike Kravetz /* 40561c9e8defSMike Kravetz * If shared, add to page cache 40571c9e8defSMike Kravetz */ 40581c9e8defSMike Kravetz if (vm_shared) { 40591e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 40601e392147SAndrea Arcangeli ret = -EFAULT; 40611e392147SAndrea Arcangeli if (idx >= size) 40621e392147SAndrea Arcangeli goto out_release_nounlock; 40631c9e8defSMike Kravetz 40641e392147SAndrea Arcangeli /* 40651e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 40661e392147SAndrea Arcangeli * huge_add_to_page_cache() below happens through the 40671e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 40681e392147SAndrea Arcangeli * the caller. 40691e392147SAndrea Arcangeli */ 40701c9e8defSMike Kravetz ret = huge_add_to_page_cache(page, mapping, idx); 40711c9e8defSMike Kravetz if (ret) 40721c9e8defSMike Kravetz goto out_release_nounlock; 40731c9e8defSMike Kravetz } 40741c9e8defSMike Kravetz 40758fb5debcSMike Kravetz ptl = huge_pte_lockptr(h, dst_mm, dst_pte); 40768fb5debcSMike Kravetz spin_lock(ptl); 40778fb5debcSMike Kravetz 40781e392147SAndrea Arcangeli /* 40791e392147SAndrea Arcangeli * Recheck the i_size after holding PT lock to make sure not 40801e392147SAndrea Arcangeli * to leave any page mapped (as page_mapped()) beyond the end 40811e392147SAndrea Arcangeli * of the i_size (remove_inode_hugepages() is strict about 40821e392147SAndrea Arcangeli * enforcing that). If we bail out here, we'll also leave a 40831e392147SAndrea Arcangeli * page in the radix tree in the vm_shared case beyond the end 40841e392147SAndrea Arcangeli * of the i_size, but remove_inode_hugepages() will take care 40851e392147SAndrea Arcangeli * of it as soon as we drop the hugetlb_fault_mutex_table. 40861e392147SAndrea Arcangeli */ 40871e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 40881e392147SAndrea Arcangeli ret = -EFAULT; 40891e392147SAndrea Arcangeli if (idx >= size) 40901e392147SAndrea Arcangeli goto out_release_unlock; 40911e392147SAndrea Arcangeli 40928fb5debcSMike Kravetz ret = -EEXIST; 40938fb5debcSMike Kravetz if (!huge_pte_none(huge_ptep_get(dst_pte))) 40948fb5debcSMike Kravetz goto out_release_unlock; 40958fb5debcSMike Kravetz 40961c9e8defSMike Kravetz if (vm_shared) { 40971c9e8defSMike Kravetz page_dup_rmap(page, true); 40981c9e8defSMike Kravetz } else { 40998fb5debcSMike Kravetz ClearPagePrivate(page); 41008fb5debcSMike Kravetz hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); 41011c9e8defSMike Kravetz } 41028fb5debcSMike Kravetz 41038fb5debcSMike Kravetz _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); 41048fb5debcSMike Kravetz if (dst_vma->vm_flags & VM_WRITE) 41058fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 41068fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 41078fb5debcSMike Kravetz 41088fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 41098fb5debcSMike Kravetz 41108fb5debcSMike Kravetz (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, 41118fb5debcSMike Kravetz dst_vma->vm_flags & VM_WRITE); 41128fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 41138fb5debcSMike Kravetz 41148fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 41158fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 41168fb5debcSMike Kravetz 41178fb5debcSMike Kravetz spin_unlock(ptl); 41181c9e8defSMike Kravetz if (vm_shared) 41191c9e8defSMike Kravetz unlock_page(page); 41208fb5debcSMike Kravetz ret = 0; 41218fb5debcSMike Kravetz out: 41228fb5debcSMike Kravetz return ret; 41238fb5debcSMike Kravetz out_release_unlock: 41248fb5debcSMike Kravetz spin_unlock(ptl); 41251c9e8defSMike Kravetz if (vm_shared) 41261c9e8defSMike Kravetz unlock_page(page); 41275af10dfdSAndrea Arcangeli out_release_nounlock: 41288fb5debcSMike Kravetz put_page(page); 41298fb5debcSMike Kravetz goto out; 41308fb5debcSMike Kravetz } 41318fb5debcSMike Kravetz 413228a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 413363551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 413428a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 413587ffc118SAndrea Arcangeli long i, unsigned int flags, int *nonblocking) 413663551ae0SDavid Gibson { 4137d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 4138d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 413928a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 4140a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 41412be7cfedSDaniel Jordan int err = -EFAULT; 414263551ae0SDavid Gibson 414363551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 414463551ae0SDavid Gibson pte_t *pte; 4145cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 41462a15efc9SHugh Dickins int absent; 414763551ae0SDavid Gibson struct page *page; 414863551ae0SDavid Gibson 41494c887265SAdam Litke /* 415002057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 415102057967SDavid Rientjes * potentially allocating memory. 415202057967SDavid Rientjes */ 415302057967SDavid Rientjes if (unlikely(fatal_signal_pending(current))) { 415402057967SDavid Rientjes remainder = 0; 415502057967SDavid Rientjes break; 415602057967SDavid Rientjes } 415702057967SDavid Rientjes 415802057967SDavid Rientjes /* 41594c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 41602a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 41614c887265SAdam Litke * first, for the page indexing below to work. 4162cb900f41SKirill A. Shutemov * 4163cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 41644c887265SAdam Litke */ 41657868a208SPunit Agrawal pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), 41667868a208SPunit Agrawal huge_page_size(h)); 4167cb900f41SKirill A. Shutemov if (pte) 4168cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 41692a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 417063551ae0SDavid Gibson 41712a15efc9SHugh Dickins /* 41722a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 41733ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 41743ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 41753ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 41763ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 41772a15efc9SHugh Dickins */ 41783ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 41793ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 4180cb900f41SKirill A. Shutemov if (pte) 4181cb900f41SKirill A. Shutemov spin_unlock(ptl); 41822a15efc9SHugh Dickins remainder = 0; 41832a15efc9SHugh Dickins break; 41842a15efc9SHugh Dickins } 41852a15efc9SHugh Dickins 41869cc3a5bdSNaoya Horiguchi /* 41879cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 41889cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 41899cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 41909cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 41919cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 41929cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 41939cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 41949cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 41959cc3a5bdSNaoya Horiguchi */ 41969cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 4197106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 4198106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 41994c887265SAdam Litke int ret; 420087ffc118SAndrea Arcangeli unsigned int fault_flags = 0; 42014c887265SAdam Litke 4202cb900f41SKirill A. Shutemov if (pte) 4203cb900f41SKirill A. Shutemov spin_unlock(ptl); 420487ffc118SAndrea Arcangeli if (flags & FOLL_WRITE) 420587ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_WRITE; 420687ffc118SAndrea Arcangeli if (nonblocking) 420787ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY; 420887ffc118SAndrea Arcangeli if (flags & FOLL_NOWAIT) 420987ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY | 421087ffc118SAndrea Arcangeli FAULT_FLAG_RETRY_NOWAIT; 421187ffc118SAndrea Arcangeli if (flags & FOLL_TRIED) { 421287ffc118SAndrea Arcangeli VM_WARN_ON_ONCE(fault_flags & 421387ffc118SAndrea Arcangeli FAULT_FLAG_ALLOW_RETRY); 421487ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_TRIED; 421587ffc118SAndrea Arcangeli } 421687ffc118SAndrea Arcangeli ret = hugetlb_fault(mm, vma, vaddr, fault_flags); 421787ffc118SAndrea Arcangeli if (ret & VM_FAULT_ERROR) { 42182be7cfedSDaniel Jordan err = vm_fault_to_errno(ret, flags); 42191c59827dSHugh Dickins remainder = 0; 42201c59827dSHugh Dickins break; 42211c59827dSHugh Dickins } 422287ffc118SAndrea Arcangeli if (ret & VM_FAULT_RETRY) { 422387ffc118SAndrea Arcangeli if (nonblocking) 422487ffc118SAndrea Arcangeli *nonblocking = 0; 422587ffc118SAndrea Arcangeli *nr_pages = 0; 422687ffc118SAndrea Arcangeli /* 422787ffc118SAndrea Arcangeli * VM_FAULT_RETRY must not return an 422887ffc118SAndrea Arcangeli * error, it will return zero 422987ffc118SAndrea Arcangeli * instead. 423087ffc118SAndrea Arcangeli * 423187ffc118SAndrea Arcangeli * No need to update "position" as the 423287ffc118SAndrea Arcangeli * caller will not check it after 423387ffc118SAndrea Arcangeli * *nr_pages is set to 0. 423487ffc118SAndrea Arcangeli */ 423587ffc118SAndrea Arcangeli return i; 423687ffc118SAndrea Arcangeli } 423787ffc118SAndrea Arcangeli continue; 423887ffc118SAndrea Arcangeli } 423963551ae0SDavid Gibson 4240a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 42417f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 4242d5d4b0aaSChen, Kenneth W same_page: 4243d6692183SChen, Kenneth W if (pages) { 424469d177c2SAndy Whitcroft pages[i] = mem_map_offset(page, pfn_offset); 4245ddc58f27SKirill A. Shutemov get_page(pages[i]); 4246d6692183SChen, Kenneth W } 424763551ae0SDavid Gibson 424863551ae0SDavid Gibson if (vmas) 424963551ae0SDavid Gibson vmas[i] = vma; 425063551ae0SDavid Gibson 425163551ae0SDavid Gibson vaddr += PAGE_SIZE; 4252d5d4b0aaSChen, Kenneth W ++pfn_offset; 425363551ae0SDavid Gibson --remainder; 425463551ae0SDavid Gibson ++i; 4255d5d4b0aaSChen, Kenneth W if (vaddr < vma->vm_end && remainder && 4256a5516438SAndi Kleen pfn_offset < pages_per_huge_page(h)) { 4257d5d4b0aaSChen, Kenneth W /* 4258d5d4b0aaSChen, Kenneth W * We use pfn_offset to avoid touching the pageframes 4259d5d4b0aaSChen, Kenneth W * of this compound page. 4260d5d4b0aaSChen, Kenneth W */ 4261d5d4b0aaSChen, Kenneth W goto same_page; 4262d5d4b0aaSChen, Kenneth W } 4263cb900f41SKirill A. Shutemov spin_unlock(ptl); 426463551ae0SDavid Gibson } 426528a35716SMichel Lespinasse *nr_pages = remainder; 426687ffc118SAndrea Arcangeli /* 426787ffc118SAndrea Arcangeli * setting position is actually required only if remainder is 426887ffc118SAndrea Arcangeli * not zero but it's faster not to add a "if (remainder)" 426987ffc118SAndrea Arcangeli * branch. 427087ffc118SAndrea Arcangeli */ 427163551ae0SDavid Gibson *position = vaddr; 427263551ae0SDavid Gibson 42732be7cfedSDaniel Jordan return i ? i : err; 427463551ae0SDavid Gibson } 42758f860591SZhang, Yanmin 42765491ae7bSAneesh Kumar K.V #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE 42775491ae7bSAneesh Kumar K.V /* 42785491ae7bSAneesh Kumar K.V * ARCHes with special requirements for evicting HUGETLB backing TLB entries can 42795491ae7bSAneesh Kumar K.V * implement this. 42805491ae7bSAneesh Kumar K.V */ 42815491ae7bSAneesh Kumar K.V #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) 42825491ae7bSAneesh Kumar K.V #endif 42835491ae7bSAneesh Kumar K.V 42847da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 42858f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 42868f860591SZhang, Yanmin { 42878f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 42888f860591SZhang, Yanmin unsigned long start = address; 42898f860591SZhang, Yanmin pte_t *ptep; 42908f860591SZhang, Yanmin pte_t pte; 4291a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 42927da4d641SPeter Zijlstra unsigned long pages = 0; 42938f860591SZhang, Yanmin 42948f860591SZhang, Yanmin BUG_ON(address >= end); 42958f860591SZhang, Yanmin flush_cache_range(vma, address, end); 42968f860591SZhang, Yanmin 4297a5338093SRik van Riel mmu_notifier_invalidate_range_start(mm, start, end); 429883cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 4299a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 4300cb900f41SKirill A. Shutemov spinlock_t *ptl; 43017868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, huge_page_size(h)); 43028f860591SZhang, Yanmin if (!ptep) 43038f860591SZhang, Yanmin continue; 4304cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 43057da4d641SPeter Zijlstra if (huge_pmd_unshare(mm, &address, ptep)) { 43067da4d641SPeter Zijlstra pages++; 4307cb900f41SKirill A. Shutemov spin_unlock(ptl); 430839dde65cSChen, Kenneth W continue; 43097da4d641SPeter Zijlstra } 4310a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 4311a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 4312a8bda28dSNaoya Horiguchi spin_unlock(ptl); 4313a8bda28dSNaoya Horiguchi continue; 4314a8bda28dSNaoya Horiguchi } 4315a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 4316a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 4317a8bda28dSNaoya Horiguchi 4318a8bda28dSNaoya Horiguchi if (is_write_migration_entry(entry)) { 4319a8bda28dSNaoya Horiguchi pte_t newpte; 4320a8bda28dSNaoya Horiguchi 4321a8bda28dSNaoya Horiguchi make_migration_entry_read(&entry); 4322a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 4323e5251fd4SPunit Agrawal set_huge_swap_pte_at(mm, address, ptep, 4324e5251fd4SPunit Agrawal newpte, huge_page_size(h)); 4325a8bda28dSNaoya Horiguchi pages++; 4326a8bda28dSNaoya Horiguchi } 4327a8bda28dSNaoya Horiguchi spin_unlock(ptl); 4328a8bda28dSNaoya Horiguchi continue; 4329a8bda28dSNaoya Horiguchi } 4330a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 43318f860591SZhang, Yanmin pte = huge_ptep_get_and_clear(mm, address, ptep); 4332106c992aSGerald Schaefer pte = pte_mkhuge(huge_pte_modify(pte, newprot)); 4333be7517d6STony Lu pte = arch_make_huge_pte(pte, vma, NULL, 0); 43348f860591SZhang, Yanmin set_huge_pte_at(mm, address, ptep, pte); 43357da4d641SPeter Zijlstra pages++; 43368f860591SZhang, Yanmin } 4337cb900f41SKirill A. Shutemov spin_unlock(ptl); 43388f860591SZhang, Yanmin } 4339d833352aSMel Gorman /* 4340c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 4341d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 4342c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 4343d833352aSMel Gorman * and that page table be reused and filled with junk. 4344d833352aSMel Gorman */ 43455491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 43460f10851eSJérôme Glisse /* 43470f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 43480f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 43490f10851eSJérôme Glisse * 43500f10851eSJérôme Glisse * See Documentation/vm/mmu_notifier.txt 43510f10851eSJérôme Glisse */ 435283cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 4353a5338093SRik van Riel mmu_notifier_invalidate_range_end(mm, start, end); 43547da4d641SPeter Zijlstra 43557da4d641SPeter Zijlstra return pages << h->order; 43568f860591SZhang, Yanmin } 43578f860591SZhang, Yanmin 4358a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 4359a1e78772SMel Gorman long from, long to, 43605a6fe125SMel Gorman struct vm_area_struct *vma, 4361ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 4362e4e574b7SAdam Litke { 436317c9d12eSMel Gorman long ret, chg; 4364a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 436590481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 43669119a41eSJoonsoo Kim struct resv_map *resv_map; 43671c5ecae3SMike Kravetz long gbl_reserve; 4368e4e574b7SAdam Litke 4369a1e78772SMel Gorman /* 437017c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 437117c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 437290481622SDavid Gibson * without using reserves 437317c9d12eSMel Gorman */ 4374ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 437517c9d12eSMel Gorman return 0; 437617c9d12eSMel Gorman 437717c9d12eSMel Gorman /* 4378a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 4379a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 4380a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 4381a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 4382a1e78772SMel Gorman */ 43839119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 43844e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 43859119a41eSJoonsoo Kim 43861406ec9bSJoonsoo Kim chg = region_chg(resv_map, from, to); 43879119a41eSJoonsoo Kim 43889119a41eSJoonsoo Kim } else { 43899119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 43905a6fe125SMel Gorman if (!resv_map) 43915a6fe125SMel Gorman return -ENOMEM; 43925a6fe125SMel Gorman 439317c9d12eSMel Gorman chg = to - from; 439417c9d12eSMel Gorman 43955a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 43965a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 43975a6fe125SMel Gorman } 43985a6fe125SMel Gorman 4399c50ac050SDave Hansen if (chg < 0) { 4400c50ac050SDave Hansen ret = chg; 4401c50ac050SDave Hansen goto out_err; 4402c50ac050SDave Hansen } 440317c9d12eSMel Gorman 44041c5ecae3SMike Kravetz /* 44051c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 44061c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 44071c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 44081c5ecae3SMike Kravetz */ 44091c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 44101c5ecae3SMike Kravetz if (gbl_reserve < 0) { 4411c50ac050SDave Hansen ret = -ENOSPC; 4412c50ac050SDave Hansen goto out_err; 4413c50ac050SDave Hansen } 441417c9d12eSMel Gorman 441517c9d12eSMel Gorman /* 441617c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 441790481622SDavid Gibson * Hand the pages back to the subpool if there are not 441817c9d12eSMel Gorman */ 44191c5ecae3SMike Kravetz ret = hugetlb_acct_memory(h, gbl_reserve); 442017c9d12eSMel Gorman if (ret < 0) { 44211c5ecae3SMike Kravetz /* put back original number of pages, chg */ 44221c5ecae3SMike Kravetz (void)hugepage_subpool_put_pages(spool, chg); 4423c50ac050SDave Hansen goto out_err; 442417c9d12eSMel Gorman } 442517c9d12eSMel Gorman 442617c9d12eSMel Gorman /* 442717c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 442817c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 442917c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 443017c9d12eSMel Gorman * the reservation was and the page cache tells how much of 443117c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 443217c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 443317c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 443417c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 443517c9d12eSMel Gorman * else has to be done for private mappings here 443617c9d12eSMel Gorman */ 443733039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 443833039678SMike Kravetz long add = region_add(resv_map, from, to); 443933039678SMike Kravetz 444033039678SMike Kravetz if (unlikely(chg > add)) { 444133039678SMike Kravetz /* 444233039678SMike Kravetz * pages in this range were added to the reserve 444333039678SMike Kravetz * map between region_chg and region_add. This 444433039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 444533039678SMike Kravetz * the subpool and reserve counts modified above 444633039678SMike Kravetz * based on the difference. 444733039678SMike Kravetz */ 444833039678SMike Kravetz long rsv_adjust; 444933039678SMike Kravetz 445033039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 445133039678SMike Kravetz chg - add); 445233039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 445333039678SMike Kravetz } 445433039678SMike Kravetz } 4455a43a8c39SChen, Kenneth W return 0; 4456c50ac050SDave Hansen out_err: 44575e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 4458ff8c0c53SMike Kravetz /* Don't call region_abort if region_chg failed */ 4459ff8c0c53SMike Kravetz if (chg >= 0) 44605e911373SMike Kravetz region_abort(resv_map, from, to); 4461f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 4462f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 4463c50ac050SDave Hansen return ret; 4464a43a8c39SChen, Kenneth W } 4465a43a8c39SChen, Kenneth W 4466b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 4467b5cec28dSMike Kravetz long freed) 4468a43a8c39SChen, Kenneth W { 4469a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 44704e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 44719119a41eSJoonsoo Kim long chg = 0; 447290481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 44731c5ecae3SMike Kravetz long gbl_reserve; 447445c682a6SKen Chen 4475b5cec28dSMike Kravetz if (resv_map) { 4476b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 4477b5cec28dSMike Kravetz /* 4478b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 4479b5cec28dSMike Kravetz * must be split and another region descriptor can not be 4480b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 4481b5cec28dSMike Kravetz */ 4482b5cec28dSMike Kravetz if (chg < 0) 4483b5cec28dSMike Kravetz return chg; 4484b5cec28dSMike Kravetz } 4485b5cec28dSMike Kravetz 448645c682a6SKen Chen spin_lock(&inode->i_lock); 4487e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 448845c682a6SKen Chen spin_unlock(&inode->i_lock); 448945c682a6SKen Chen 44901c5ecae3SMike Kravetz /* 44911c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 44921c5ecae3SMike Kravetz * reservations to be released may be adjusted. 44931c5ecae3SMike Kravetz */ 44941c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 44951c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 4496b5cec28dSMike Kravetz 4497b5cec28dSMike Kravetz return 0; 4498a43a8c39SChen, Kenneth W } 449993f70f90SNaoya Horiguchi 45003212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 45013212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 45023212b535SSteve Capper struct vm_area_struct *vma, 45033212b535SSteve Capper unsigned long addr, pgoff_t idx) 45043212b535SSteve Capper { 45053212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 45063212b535SSteve Capper svma->vm_start; 45073212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 45083212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 45093212b535SSteve Capper 45103212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 4511de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 4512de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 45133212b535SSteve Capper 45143212b535SSteve Capper /* 45153212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 45163212b535SSteve Capper * page table page. 45173212b535SSteve Capper */ 45183212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 45193212b535SSteve Capper vm_flags != svm_flags || 45203212b535SSteve Capper sbase < svma->vm_start || svma->vm_end < s_end) 45213212b535SSteve Capper return 0; 45223212b535SSteve Capper 45233212b535SSteve Capper return saddr; 45243212b535SSteve Capper } 45253212b535SSteve Capper 452631aafb45SNicholas Krause static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) 45273212b535SSteve Capper { 45283212b535SSteve Capper unsigned long base = addr & PUD_MASK; 45293212b535SSteve Capper unsigned long end = base + PUD_SIZE; 45303212b535SSteve Capper 45313212b535SSteve Capper /* 45323212b535SSteve Capper * check on proper vm_flags and page table alignment 45333212b535SSteve Capper */ 45343212b535SSteve Capper if (vma->vm_flags & VM_MAYSHARE && 45353212b535SSteve Capper vma->vm_start <= base && end <= vma->vm_end) 453631aafb45SNicholas Krause return true; 453731aafb45SNicholas Krause return false; 45383212b535SSteve Capper } 45393212b535SSteve Capper 45403212b535SSteve Capper /* 45413212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 45423212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 45433212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 45443212b535SSteve Capper * code much cleaner. pmd allocation is essential for the shared case because 4545c8c06efaSDavidlohr Bueso * pud has to be populated inside the same i_mmap_rwsem section - otherwise 45463212b535SSteve Capper * racing tasks could either miss the sharing (see huge_pte_offset) or select a 45473212b535SSteve Capper * bad pmd for sharing. 45483212b535SSteve Capper */ 45493212b535SSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 45503212b535SSteve Capper { 45513212b535SSteve Capper struct vm_area_struct *vma = find_vma(mm, addr); 45523212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 45533212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 45543212b535SSteve Capper vma->vm_pgoff; 45553212b535SSteve Capper struct vm_area_struct *svma; 45563212b535SSteve Capper unsigned long saddr; 45573212b535SSteve Capper pte_t *spte = NULL; 45583212b535SSteve Capper pte_t *pte; 4559cb900f41SKirill A. Shutemov spinlock_t *ptl; 45603212b535SSteve Capper 45613212b535SSteve Capper if (!vma_shareable(vma, addr)) 45623212b535SSteve Capper return (pte_t *)pmd_alloc(mm, pud, addr); 45633212b535SSteve Capper 456483cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 45653212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 45663212b535SSteve Capper if (svma == vma) 45673212b535SSteve Capper continue; 45683212b535SSteve Capper 45693212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 45703212b535SSteve Capper if (saddr) { 45717868a208SPunit Agrawal spte = huge_pte_offset(svma->vm_mm, saddr, 45727868a208SPunit Agrawal vma_mmu_pagesize(svma)); 45733212b535SSteve Capper if (spte) { 45743212b535SSteve Capper get_page(virt_to_page(spte)); 45753212b535SSteve Capper break; 45763212b535SSteve Capper } 45773212b535SSteve Capper } 45783212b535SSteve Capper } 45793212b535SSteve Capper 45803212b535SSteve Capper if (!spte) 45813212b535SSteve Capper goto out; 45823212b535SSteve Capper 45838bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 4584dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 45853212b535SSteve Capper pud_populate(mm, pud, 45863212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 4587c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 4588dc6c9a35SKirill A. Shutemov } else { 45893212b535SSteve Capper put_page(virt_to_page(spte)); 4590dc6c9a35SKirill A. Shutemov } 4591cb900f41SKirill A. Shutemov spin_unlock(ptl); 45923212b535SSteve Capper out: 45933212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 459483cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 45953212b535SSteve Capper return pte; 45963212b535SSteve Capper } 45973212b535SSteve Capper 45983212b535SSteve Capper /* 45993212b535SSteve Capper * unmap huge page backed by shared pte. 46003212b535SSteve Capper * 46013212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 46023212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 46033212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 46043212b535SSteve Capper * 4605cb900f41SKirill A. Shutemov * called with page table lock held. 46063212b535SSteve Capper * 46073212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 46083212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 46093212b535SSteve Capper */ 46103212b535SSteve Capper int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 46113212b535SSteve Capper { 46123212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 4613c2febafcSKirill A. Shutemov p4d_t *p4d = p4d_offset(pgd, *addr); 4614c2febafcSKirill A. Shutemov pud_t *pud = pud_offset(p4d, *addr); 46153212b535SSteve Capper 46163212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 46173212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 46183212b535SSteve Capper return 0; 46193212b535SSteve Capper 46203212b535SSteve Capper pud_clear(pud); 46213212b535SSteve Capper put_page(virt_to_page(ptep)); 4622dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 46233212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 46243212b535SSteve Capper return 1; 46253212b535SSteve Capper } 46269e5fc74cSSteve Capper #define want_pmd_share() (1) 46279e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 46289e5fc74cSSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 46299e5fc74cSSteve Capper { 46309e5fc74cSSteve Capper return NULL; 46319e5fc74cSSteve Capper } 4632e81f2d22SZhang Zhen 4633e81f2d22SZhang Zhen int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 4634e81f2d22SZhang Zhen { 4635e81f2d22SZhang Zhen return 0; 4636e81f2d22SZhang Zhen } 46379e5fc74cSSteve Capper #define want_pmd_share() (0) 46383212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 46393212b535SSteve Capper 46409e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 46419e5fc74cSSteve Capper pte_t *huge_pte_alloc(struct mm_struct *mm, 46429e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 46439e5fc74cSSteve Capper { 46449e5fc74cSSteve Capper pgd_t *pgd; 4645c2febafcSKirill A. Shutemov p4d_t *p4d; 46469e5fc74cSSteve Capper pud_t *pud; 46479e5fc74cSSteve Capper pte_t *pte = NULL; 46489e5fc74cSSteve Capper 46499e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 4650f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 4651f4f0a3d8SKirill A. Shutemov if (!p4d) 4652f4f0a3d8SKirill A. Shutemov return NULL; 4653c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 46549e5fc74cSSteve Capper if (pud) { 46559e5fc74cSSteve Capper if (sz == PUD_SIZE) { 46569e5fc74cSSteve Capper pte = (pte_t *)pud; 46579e5fc74cSSteve Capper } else { 46589e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 46599e5fc74cSSteve Capper if (want_pmd_share() && pud_none(*pud)) 46609e5fc74cSSteve Capper pte = huge_pmd_share(mm, addr, pud); 46619e5fc74cSSteve Capper else 46629e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 46639e5fc74cSSteve Capper } 46649e5fc74cSSteve Capper } 46654e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 46669e5fc74cSSteve Capper 46679e5fc74cSSteve Capper return pte; 46689e5fc74cSSteve Capper } 46699e5fc74cSSteve Capper 46709b19df29SPunit Agrawal /* 46719b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 46729b19df29SPunit Agrawal * entry at address @addr 46739b19df29SPunit Agrawal * 46749b19df29SPunit Agrawal * Return: Pointer to page table or swap entry (PUD or PMD) for 46759b19df29SPunit Agrawal * address @addr, or NULL if a p*d_none() entry is encountered and the 46769b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 46779b19df29SPunit Agrawal * table. 46789b19df29SPunit Agrawal */ 46797868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 46807868a208SPunit Agrawal unsigned long addr, unsigned long sz) 46819e5fc74cSSteve Capper { 46829e5fc74cSSteve Capper pgd_t *pgd; 4683c2febafcSKirill A. Shutemov p4d_t *p4d; 46849e5fc74cSSteve Capper pud_t *pud; 4685c2febafcSKirill A. Shutemov pmd_t *pmd; 46869e5fc74cSSteve Capper 46879e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 4688c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 4689c2febafcSKirill A. Shutemov return NULL; 4690c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 4691c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 4692c2febafcSKirill A. Shutemov return NULL; 46939b19df29SPunit Agrawal 4694c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 46959b19df29SPunit Agrawal if (sz != PUD_SIZE && pud_none(*pud)) 4696c2febafcSKirill A. Shutemov return NULL; 46979b19df29SPunit Agrawal /* hugepage or swap? */ 46989b19df29SPunit Agrawal if (pud_huge(*pud) || !pud_present(*pud)) 46999e5fc74cSSteve Capper return (pte_t *)pud; 47009b19df29SPunit Agrawal 47019e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 47029b19df29SPunit Agrawal if (sz != PMD_SIZE && pmd_none(*pmd)) 47039b19df29SPunit Agrawal return NULL; 47049b19df29SPunit Agrawal /* hugepage or swap? */ 47059b19df29SPunit Agrawal if (pmd_huge(*pmd) || !pmd_present(*pmd)) 47069e5fc74cSSteve Capper return (pte_t *)pmd; 47079b19df29SPunit Agrawal 47089b19df29SPunit Agrawal return NULL; 47099e5fc74cSSteve Capper } 47109e5fc74cSSteve Capper 471161f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 471261f77edaSNaoya Horiguchi 471361f77edaSNaoya Horiguchi /* 471461f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 471561f77edaSNaoya Horiguchi * behavior. 471661f77edaSNaoya Horiguchi */ 471761f77edaSNaoya Horiguchi struct page * __weak 471861f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 471961f77edaSNaoya Horiguchi int write) 472061f77edaSNaoya Horiguchi { 472161f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 472261f77edaSNaoya Horiguchi } 472361f77edaSNaoya Horiguchi 472461f77edaSNaoya Horiguchi struct page * __weak 47254dc71451SAneesh Kumar K.V follow_huge_pd(struct vm_area_struct *vma, 47264dc71451SAneesh Kumar K.V unsigned long address, hugepd_t hpd, int flags, int pdshift) 47274dc71451SAneesh Kumar K.V { 47284dc71451SAneesh Kumar K.V WARN(1, "hugepd follow called with no support for hugepage directory format\n"); 47294dc71451SAneesh Kumar K.V return NULL; 47304dc71451SAneesh Kumar K.V } 47314dc71451SAneesh Kumar K.V 47324dc71451SAneesh Kumar K.V struct page * __weak 47339e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 4734e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 47359e5fc74cSSteve Capper { 4736e66f17ffSNaoya Horiguchi struct page *page = NULL; 4737e66f17ffSNaoya Horiguchi spinlock_t *ptl; 4738c9d398faSNaoya Horiguchi pte_t pte; 4739e66f17ffSNaoya Horiguchi retry: 4740e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 4741e66f17ffSNaoya Horiguchi spin_lock(ptl); 4742e66f17ffSNaoya Horiguchi /* 4743e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 4744e66f17ffSNaoya Horiguchi * unmapped from other threads. 4745e66f17ffSNaoya Horiguchi */ 4746e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 4747e66f17ffSNaoya Horiguchi goto out; 4748c9d398faSNaoya Horiguchi pte = huge_ptep_get((pte_t *)pmd); 4749c9d398faSNaoya Horiguchi if (pte_present(pte)) { 475097534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 4751e66f17ffSNaoya Horiguchi if (flags & FOLL_GET) 4752e66f17ffSNaoya Horiguchi get_page(page); 4753e66f17ffSNaoya Horiguchi } else { 4754c9d398faSNaoya Horiguchi if (is_hugetlb_entry_migration(pte)) { 4755e66f17ffSNaoya Horiguchi spin_unlock(ptl); 4756e66f17ffSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pmd, ptl); 4757e66f17ffSNaoya Horiguchi goto retry; 4758e66f17ffSNaoya Horiguchi } 4759e66f17ffSNaoya Horiguchi /* 4760e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 4761e66f17ffSNaoya Horiguchi * follow_page_mask(). 4762e66f17ffSNaoya Horiguchi */ 4763e66f17ffSNaoya Horiguchi } 4764e66f17ffSNaoya Horiguchi out: 4765e66f17ffSNaoya Horiguchi spin_unlock(ptl); 47669e5fc74cSSteve Capper return page; 47679e5fc74cSSteve Capper } 47689e5fc74cSSteve Capper 476961f77edaSNaoya Horiguchi struct page * __weak 47709e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 4771e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 47729e5fc74cSSteve Capper { 4773e66f17ffSNaoya Horiguchi if (flags & FOLL_GET) 4774e66f17ffSNaoya Horiguchi return NULL; 47759e5fc74cSSteve Capper 4776e66f17ffSNaoya Horiguchi return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 47779e5fc74cSSteve Capper } 47789e5fc74cSSteve Capper 4779faaa5b62SAnshuman Khandual struct page * __weak 4780faaa5b62SAnshuman Khandual follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) 4781faaa5b62SAnshuman Khandual { 4782faaa5b62SAnshuman Khandual if (flags & FOLL_GET) 4783faaa5b62SAnshuman Khandual return NULL; 4784faaa5b62SAnshuman Khandual 4785faaa5b62SAnshuman Khandual return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); 4786faaa5b62SAnshuman Khandual } 4787faaa5b62SAnshuman Khandual 478831caf665SNaoya Horiguchi bool isolate_huge_page(struct page *page, struct list_head *list) 478931caf665SNaoya Horiguchi { 4790bcc54222SNaoya Horiguchi bool ret = true; 4791bcc54222SNaoya Horiguchi 4792309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 479331caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 4794bcc54222SNaoya Horiguchi if (!page_huge_active(page) || !get_page_unless_zero(page)) { 4795bcc54222SNaoya Horiguchi ret = false; 4796bcc54222SNaoya Horiguchi goto unlock; 4797bcc54222SNaoya Horiguchi } 4798bcc54222SNaoya Horiguchi clear_page_huge_active(page); 479931caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 4800bcc54222SNaoya Horiguchi unlock: 480131caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 4802bcc54222SNaoya Horiguchi return ret; 480331caf665SNaoya Horiguchi } 480431caf665SNaoya Horiguchi 480531caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 480631caf665SNaoya Horiguchi { 4807309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 480831caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 4809bcc54222SNaoya Horiguchi set_page_huge_active(page); 481031caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 481131caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 481231caf665SNaoya Horiguchi put_page(page); 481331caf665SNaoya Horiguchi } 4814