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> 1897ad1087SMike Rapoport #include <linux/memblock.h> 19a3437870SNishanth Aravamudan #include <linux/sysfs.h> 205a0e3ad6STejun Heo #include <linux/slab.h> 2163489f8eSMike Kravetz #include <linux/mmdebug.h> 22174cd4b1SIngo Molnar #include <linux/sched/signal.h> 230fe6e20bSNaoya Horiguchi #include <linux/rmap.h> 24c6247f72SMatthew Wilcox #include <linux/string_helpers.h> 25fd6a03edSNaoya Horiguchi #include <linux/swap.h> 26fd6a03edSNaoya Horiguchi #include <linux/swapops.h> 278382d914SDavidlohr Bueso #include <linux/jhash.h> 2898fa15f3SAnshuman Khandual #include <linux/numa.h> 29d6606683SLinus Torvalds 3063551ae0SDavid Gibson #include <asm/page.h> 3163551ae0SDavid Gibson #include <asm/pgtable.h> 3224669e58SAneesh Kumar K.V #include <asm/tlb.h> 3363551ae0SDavid Gibson 3424669e58SAneesh Kumar K.V #include <linux/io.h> 3563551ae0SDavid Gibson #include <linux/hugetlb.h> 369dd540e2SAneesh Kumar K.V #include <linux/hugetlb_cgroup.h> 379a305230SLee Schermerhorn #include <linux/node.h> 381a1aad8aSMike Kravetz #include <linux/userfaultfd_k.h> 39ab5ac90aSMichal Hocko #include <linux/page_owner.h> 407835e98bSNick Piggin #include "internal.h" 411da177e4SLinus Torvalds 42c3f38a38SAneesh Kumar K.V int hugetlb_max_hstate __read_mostly; 43e5ff2159SAndi Kleen unsigned int default_hstate_idx; 44e5ff2159SAndi Kleen struct hstate hstates[HUGE_MAX_HSTATE]; 45641844f5SNaoya Horiguchi /* 46641844f5SNaoya Horiguchi * Minimum page order among possible hugepage sizes, set to a proper value 47641844f5SNaoya Horiguchi * at boot time. 48641844f5SNaoya Horiguchi */ 49641844f5SNaoya Horiguchi static unsigned int minimum_order __read_mostly = UINT_MAX; 50e5ff2159SAndi Kleen 5153ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 5253ba51d2SJon Tollefson 53e5ff2159SAndi Kleen /* for command line parsing */ 54e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 55e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 56e11bfbfcSNick Piggin static unsigned long __initdata default_hstate_size; 579fee021dSVaishali Thakkar static bool __initdata parsed_valid_hugepagesz = true; 58e5ff2159SAndi Kleen 593935baa9SDavid Gibson /* 6031caf665SNaoya Horiguchi * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, 6131caf665SNaoya Horiguchi * free_huge_pages, and surplus_huge_pages. 623935baa9SDavid Gibson */ 63c3f38a38SAneesh Kumar K.V DEFINE_SPINLOCK(hugetlb_lock); 640bd0f9fbSEric Paris 658382d914SDavidlohr Bueso /* 668382d914SDavidlohr Bueso * Serializes faults on the same logical page. This is used to 678382d914SDavidlohr Bueso * prevent spurious OOMs when the hugepage pool is fully utilized. 688382d914SDavidlohr Bueso */ 698382d914SDavidlohr Bueso static int num_fault_mutexes; 70c672c7f2SMike Kravetz struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; 718382d914SDavidlohr Bueso 727ca02d0aSMike Kravetz /* Forward declaration */ 737ca02d0aSMike Kravetz static int hugetlb_acct_memory(struct hstate *h, long delta); 747ca02d0aSMike Kravetz 7590481622SDavid Gibson static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) 7690481622SDavid Gibson { 7790481622SDavid Gibson bool free = (spool->count == 0) && (spool->used_hpages == 0); 7890481622SDavid Gibson 7990481622SDavid Gibson spin_unlock(&spool->lock); 8090481622SDavid Gibson 8190481622SDavid Gibson /* If no pages are used, and no other handles to the subpool 827ca02d0aSMike Kravetz * remain, give up any reservations mased on minimum size and 837ca02d0aSMike Kravetz * free the subpool */ 847ca02d0aSMike Kravetz if (free) { 857ca02d0aSMike Kravetz if (spool->min_hpages != -1) 867ca02d0aSMike Kravetz hugetlb_acct_memory(spool->hstate, 877ca02d0aSMike Kravetz -spool->min_hpages); 8890481622SDavid Gibson kfree(spool); 8990481622SDavid Gibson } 907ca02d0aSMike Kravetz } 9190481622SDavid Gibson 927ca02d0aSMike Kravetz struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, 937ca02d0aSMike Kravetz long min_hpages) 9490481622SDavid Gibson { 9590481622SDavid Gibson struct hugepage_subpool *spool; 9690481622SDavid Gibson 97c6a91820SMike Kravetz spool = kzalloc(sizeof(*spool), GFP_KERNEL); 9890481622SDavid Gibson if (!spool) 9990481622SDavid Gibson return NULL; 10090481622SDavid Gibson 10190481622SDavid Gibson spin_lock_init(&spool->lock); 10290481622SDavid Gibson spool->count = 1; 1037ca02d0aSMike Kravetz spool->max_hpages = max_hpages; 1047ca02d0aSMike Kravetz spool->hstate = h; 1057ca02d0aSMike Kravetz spool->min_hpages = min_hpages; 1067ca02d0aSMike Kravetz 1077ca02d0aSMike Kravetz if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { 1087ca02d0aSMike Kravetz kfree(spool); 1097ca02d0aSMike Kravetz return NULL; 1107ca02d0aSMike Kravetz } 1117ca02d0aSMike Kravetz spool->rsv_hpages = min_hpages; 11290481622SDavid Gibson 11390481622SDavid Gibson return spool; 11490481622SDavid Gibson } 11590481622SDavid Gibson 11690481622SDavid Gibson void hugepage_put_subpool(struct hugepage_subpool *spool) 11790481622SDavid Gibson { 11890481622SDavid Gibson spin_lock(&spool->lock); 11990481622SDavid Gibson BUG_ON(!spool->count); 12090481622SDavid Gibson spool->count--; 12190481622SDavid Gibson unlock_or_release_subpool(spool); 12290481622SDavid Gibson } 12390481622SDavid Gibson 1241c5ecae3SMike Kravetz /* 1251c5ecae3SMike Kravetz * Subpool accounting for allocating and reserving pages. 1261c5ecae3SMike Kravetz * Return -ENOMEM if there are not enough resources to satisfy the 1271c5ecae3SMike Kravetz * the request. Otherwise, return the number of pages by which the 1281c5ecae3SMike Kravetz * global pools must be adjusted (upward). The returned value may 1291c5ecae3SMike Kravetz * only be different than the passed value (delta) in the case where 1301c5ecae3SMike Kravetz * a subpool minimum size must be manitained. 1311c5ecae3SMike Kravetz */ 1321c5ecae3SMike Kravetz static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, 13390481622SDavid Gibson long delta) 13490481622SDavid Gibson { 1351c5ecae3SMike Kravetz long ret = delta; 13690481622SDavid Gibson 13790481622SDavid Gibson if (!spool) 1381c5ecae3SMike Kravetz return ret; 13990481622SDavid Gibson 14090481622SDavid Gibson spin_lock(&spool->lock); 14190481622SDavid Gibson 1421c5ecae3SMike Kravetz if (spool->max_hpages != -1) { /* maximum size accounting */ 1431c5ecae3SMike Kravetz if ((spool->used_hpages + delta) <= spool->max_hpages) 1441c5ecae3SMike Kravetz spool->used_hpages += delta; 1451c5ecae3SMike Kravetz else { 1461c5ecae3SMike Kravetz ret = -ENOMEM; 1471c5ecae3SMike Kravetz goto unlock_ret; 1481c5ecae3SMike Kravetz } 1491c5ecae3SMike Kravetz } 1501c5ecae3SMike Kravetz 15109a95e29SMike Kravetz /* minimum size accounting */ 15209a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->rsv_hpages) { 1531c5ecae3SMike Kravetz if (delta > spool->rsv_hpages) { 1541c5ecae3SMike Kravetz /* 1551c5ecae3SMike Kravetz * Asking for more reserves than those already taken on 1561c5ecae3SMike Kravetz * behalf of subpool. Return difference. 1571c5ecae3SMike Kravetz */ 1581c5ecae3SMike Kravetz ret = delta - spool->rsv_hpages; 1591c5ecae3SMike Kravetz spool->rsv_hpages = 0; 1601c5ecae3SMike Kravetz } else { 1611c5ecae3SMike Kravetz ret = 0; /* reserves already accounted for */ 1621c5ecae3SMike Kravetz spool->rsv_hpages -= delta; 1631c5ecae3SMike Kravetz } 1641c5ecae3SMike Kravetz } 1651c5ecae3SMike Kravetz 1661c5ecae3SMike Kravetz unlock_ret: 1671c5ecae3SMike Kravetz spin_unlock(&spool->lock); 16890481622SDavid Gibson return ret; 16990481622SDavid Gibson } 17090481622SDavid Gibson 1711c5ecae3SMike Kravetz /* 1721c5ecae3SMike Kravetz * Subpool accounting for freeing and unreserving pages. 1731c5ecae3SMike Kravetz * Return the number of global page reservations that must be dropped. 1741c5ecae3SMike Kravetz * The return value may only be different than the passed value (delta) 1751c5ecae3SMike Kravetz * in the case where a subpool minimum size must be maintained. 1761c5ecae3SMike Kravetz */ 1771c5ecae3SMike Kravetz static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, 17890481622SDavid Gibson long delta) 17990481622SDavid Gibson { 1801c5ecae3SMike Kravetz long ret = delta; 1811c5ecae3SMike Kravetz 18290481622SDavid Gibson if (!spool) 1831c5ecae3SMike Kravetz return delta; 18490481622SDavid Gibson 18590481622SDavid Gibson spin_lock(&spool->lock); 1861c5ecae3SMike Kravetz 1871c5ecae3SMike Kravetz if (spool->max_hpages != -1) /* maximum size accounting */ 18890481622SDavid Gibson spool->used_hpages -= delta; 1891c5ecae3SMike Kravetz 19009a95e29SMike Kravetz /* minimum size accounting */ 19109a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { 1921c5ecae3SMike Kravetz if (spool->rsv_hpages + delta <= spool->min_hpages) 1931c5ecae3SMike Kravetz ret = 0; 1941c5ecae3SMike Kravetz else 1951c5ecae3SMike Kravetz ret = spool->rsv_hpages + delta - spool->min_hpages; 1961c5ecae3SMike Kravetz 1971c5ecae3SMike Kravetz spool->rsv_hpages += delta; 1981c5ecae3SMike Kravetz if (spool->rsv_hpages > spool->min_hpages) 1991c5ecae3SMike Kravetz spool->rsv_hpages = spool->min_hpages; 2001c5ecae3SMike Kravetz } 2011c5ecae3SMike Kravetz 2021c5ecae3SMike Kravetz /* 2031c5ecae3SMike Kravetz * If hugetlbfs_put_super couldn't free spool due to an outstanding 2041c5ecae3SMike Kravetz * quota reference, free it now. 2051c5ecae3SMike Kravetz */ 20690481622SDavid Gibson unlock_or_release_subpool(spool); 2071c5ecae3SMike Kravetz 2081c5ecae3SMike Kravetz return ret; 20990481622SDavid Gibson } 21090481622SDavid Gibson 21190481622SDavid Gibson static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 21290481622SDavid Gibson { 21390481622SDavid Gibson return HUGETLBFS_SB(inode->i_sb)->spool; 21490481622SDavid Gibson } 21590481622SDavid Gibson 21690481622SDavid Gibson static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 21790481622SDavid Gibson { 218496ad9aaSAl Viro return subpool_inode(file_inode(vma->vm_file)); 21990481622SDavid Gibson } 22090481622SDavid Gibson 221e7c4b0bfSAndy Whitcroft /* 22296822904SAndy Whitcroft * Region tracking -- allows tracking of reservations and instantiated pages 22396822904SAndy Whitcroft * across the pages in a mapping. 22484afd99bSAndy Whitcroft * 2251dd308a7SMike Kravetz * The region data structures are embedded into a resv_map and protected 2261dd308a7SMike Kravetz * by a resv_map's lock. The set of regions within the resv_map represent 2271dd308a7SMike Kravetz * reservations for huge pages, or huge pages that have already been 2281dd308a7SMike Kravetz * instantiated within the map. The from and to elements are huge page 2291dd308a7SMike Kravetz * indicies into the associated mapping. from indicates the starting index 2301dd308a7SMike Kravetz * of the region. to represents the first index past the end of the region. 2311dd308a7SMike Kravetz * 2321dd308a7SMike Kravetz * For example, a file region structure with from == 0 and to == 4 represents 2331dd308a7SMike Kravetz * four huge pages in a mapping. It is important to note that the to element 2341dd308a7SMike Kravetz * represents the first element past the end of the region. This is used in 2351dd308a7SMike Kravetz * arithmetic as 4(to) - 0(from) = 4 huge pages in the region. 2361dd308a7SMike Kravetz * 2371dd308a7SMike Kravetz * Interval notation of the form [from, to) will be used to indicate that 2381dd308a7SMike Kravetz * the endpoint from is inclusive and to is exclusive. 23996822904SAndy Whitcroft */ 24096822904SAndy Whitcroft struct file_region { 24196822904SAndy Whitcroft struct list_head link; 24296822904SAndy Whitcroft long from; 24396822904SAndy Whitcroft long to; 24496822904SAndy Whitcroft }; 24596822904SAndy Whitcroft 2461dd308a7SMike Kravetz /* 2471dd308a7SMike Kravetz * Add the huge page range represented by [f, t) to the reserve 2485e911373SMike Kravetz * map. In the normal case, existing regions will be expanded 2495e911373SMike Kravetz * to accommodate the specified range. Sufficient regions should 2505e911373SMike Kravetz * exist for expansion due to the previous call to region_chg 2515e911373SMike Kravetz * with the same range. However, it is possible that region_del 2525e911373SMike Kravetz * could have been called after region_chg and modifed the map 2535e911373SMike Kravetz * in such a way that no region exists to be expanded. In this 2545e911373SMike Kravetz * case, pull a region descriptor from the cache associated with 2555e911373SMike Kravetz * the map and use that for the new range. 256cf3ad20bSMike Kravetz * 257cf3ad20bSMike Kravetz * Return the number of new huge pages added to the map. This 258cf3ad20bSMike Kravetz * number is greater than or equal to zero. 2591dd308a7SMike Kravetz */ 2601406ec9bSJoonsoo Kim static long region_add(struct resv_map *resv, long f, long t) 26196822904SAndy Whitcroft { 2621406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 26396822904SAndy Whitcroft struct file_region *rg, *nrg, *trg; 264cf3ad20bSMike Kravetz long add = 0; 26596822904SAndy Whitcroft 2667b24d861SDavidlohr Bueso spin_lock(&resv->lock); 26796822904SAndy Whitcroft /* Locate the region we are either in or before. */ 26896822904SAndy Whitcroft list_for_each_entry(rg, head, link) 26996822904SAndy Whitcroft if (f <= rg->to) 27096822904SAndy Whitcroft break; 27196822904SAndy Whitcroft 2725e911373SMike Kravetz /* 2735e911373SMike Kravetz * If no region exists which can be expanded to include the 2745e911373SMike Kravetz * specified range, the list must have been modified by an 2755e911373SMike Kravetz * interleving call to region_del(). Pull a region descriptor 2765e911373SMike Kravetz * from the cache and use it for this range. 2775e911373SMike Kravetz */ 2785e911373SMike Kravetz if (&rg->link == head || t < rg->from) { 2795e911373SMike Kravetz VM_BUG_ON(resv->region_cache_count <= 0); 2805e911373SMike Kravetz 2815e911373SMike Kravetz resv->region_cache_count--; 2825e911373SMike Kravetz nrg = list_first_entry(&resv->region_cache, struct file_region, 2835e911373SMike Kravetz link); 2845e911373SMike Kravetz list_del(&nrg->link); 2855e911373SMike Kravetz 2865e911373SMike Kravetz nrg->from = f; 2875e911373SMike Kravetz nrg->to = t; 2885e911373SMike Kravetz list_add(&nrg->link, rg->link.prev); 2895e911373SMike Kravetz 2905e911373SMike Kravetz add += t - f; 2915e911373SMike Kravetz goto out_locked; 2925e911373SMike Kravetz } 2935e911373SMike Kravetz 29496822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 29596822904SAndy Whitcroft if (f > rg->from) 29696822904SAndy Whitcroft f = rg->from; 29796822904SAndy Whitcroft 29896822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 29996822904SAndy Whitcroft nrg = rg; 30096822904SAndy Whitcroft list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 30196822904SAndy Whitcroft if (&rg->link == head) 30296822904SAndy Whitcroft break; 30396822904SAndy Whitcroft if (rg->from > t) 30496822904SAndy Whitcroft break; 30596822904SAndy Whitcroft 30696822904SAndy Whitcroft /* If this area reaches higher then extend our area to 30796822904SAndy Whitcroft * include it completely. If this is not the first area 30896822904SAndy Whitcroft * which we intend to reuse, free it. */ 30996822904SAndy Whitcroft if (rg->to > t) 31096822904SAndy Whitcroft t = rg->to; 31196822904SAndy Whitcroft if (rg != nrg) { 312cf3ad20bSMike Kravetz /* Decrement return value by the deleted range. 313cf3ad20bSMike Kravetz * Another range will span this area so that by 314cf3ad20bSMike Kravetz * end of routine add will be >= zero 315cf3ad20bSMike Kravetz */ 316cf3ad20bSMike Kravetz add -= (rg->to - rg->from); 31796822904SAndy Whitcroft list_del(&rg->link); 31896822904SAndy Whitcroft kfree(rg); 31996822904SAndy Whitcroft } 32096822904SAndy Whitcroft } 321cf3ad20bSMike Kravetz 322cf3ad20bSMike Kravetz add += (nrg->from - f); /* Added to beginning of region */ 32396822904SAndy Whitcroft nrg->from = f; 324cf3ad20bSMike Kravetz add += t - nrg->to; /* Added to end of region */ 32596822904SAndy Whitcroft nrg->to = t; 326cf3ad20bSMike Kravetz 3275e911373SMike Kravetz out_locked: 3285e911373SMike Kravetz resv->adds_in_progress--; 3297b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 330cf3ad20bSMike Kravetz VM_BUG_ON(add < 0); 331cf3ad20bSMike Kravetz return add; 33296822904SAndy Whitcroft } 33396822904SAndy Whitcroft 3341dd308a7SMike Kravetz /* 3351dd308a7SMike Kravetz * Examine the existing reserve map and determine how many 3361dd308a7SMike Kravetz * huge pages in the specified range [f, t) are NOT currently 3371dd308a7SMike Kravetz * represented. This routine is called before a subsequent 3381dd308a7SMike Kravetz * call to region_add that will actually modify the reserve 3391dd308a7SMike Kravetz * map to add the specified range [f, t). region_chg does 3401dd308a7SMike Kravetz * not change the number of huge pages represented by the 3411dd308a7SMike Kravetz * map. However, if the existing regions in the map can not 3421dd308a7SMike Kravetz * be expanded to represent the new range, a new file_region 3431dd308a7SMike Kravetz * structure is added to the map as a placeholder. This is 3441dd308a7SMike Kravetz * so that the subsequent region_add call will have all the 3451dd308a7SMike Kravetz * regions it needs and will not fail. 3461dd308a7SMike Kravetz * 3475e911373SMike Kravetz * Upon entry, region_chg will also examine the cache of region descriptors 3485e911373SMike Kravetz * associated with the map. If there are not enough descriptors cached, one 3495e911373SMike Kravetz * will be allocated for the in progress add operation. 3505e911373SMike Kravetz * 3515e911373SMike Kravetz * Returns the number of huge pages that need to be added to the existing 3525e911373SMike Kravetz * reservation map for the range [f, t). This number is greater or equal to 3535e911373SMike Kravetz * zero. -ENOMEM is returned if a new file_region structure or cache entry 3545e911373SMike Kravetz * is needed and can not be allocated. 3551dd308a7SMike Kravetz */ 3561406ec9bSJoonsoo Kim static long region_chg(struct resv_map *resv, long f, long t) 35796822904SAndy Whitcroft { 3581406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 3597b24d861SDavidlohr Bueso struct file_region *rg, *nrg = NULL; 36096822904SAndy Whitcroft long chg = 0; 36196822904SAndy Whitcroft 3627b24d861SDavidlohr Bueso retry: 3637b24d861SDavidlohr Bueso spin_lock(&resv->lock); 3645e911373SMike Kravetz retry_locked: 3655e911373SMike Kravetz resv->adds_in_progress++; 3665e911373SMike Kravetz 3675e911373SMike Kravetz /* 3685e911373SMike Kravetz * Check for sufficient descriptors in the cache to accommodate 3695e911373SMike Kravetz * the number of in progress add operations. 3705e911373SMike Kravetz */ 3715e911373SMike Kravetz if (resv->adds_in_progress > resv->region_cache_count) { 3725e911373SMike Kravetz struct file_region *trg; 3735e911373SMike Kravetz 3745e911373SMike Kravetz VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1); 3755e911373SMike Kravetz /* Must drop lock to allocate a new descriptor. */ 3765e911373SMike Kravetz resv->adds_in_progress--; 3775e911373SMike Kravetz spin_unlock(&resv->lock); 3785e911373SMike Kravetz 3795e911373SMike Kravetz trg = kmalloc(sizeof(*trg), GFP_KERNEL); 380dbe409e4SMike Kravetz if (!trg) { 381dbe409e4SMike Kravetz kfree(nrg); 3825e911373SMike Kravetz return -ENOMEM; 383dbe409e4SMike Kravetz } 3845e911373SMike Kravetz 3855e911373SMike Kravetz spin_lock(&resv->lock); 3865e911373SMike Kravetz list_add(&trg->link, &resv->region_cache); 3875e911373SMike Kravetz resv->region_cache_count++; 3885e911373SMike Kravetz goto retry_locked; 3895e911373SMike Kravetz } 3905e911373SMike Kravetz 39196822904SAndy Whitcroft /* Locate the region we are before or in. */ 39296822904SAndy Whitcroft list_for_each_entry(rg, head, link) 39396822904SAndy Whitcroft if (f <= rg->to) 39496822904SAndy Whitcroft break; 39596822904SAndy Whitcroft 39696822904SAndy Whitcroft /* If we are below the current region then a new region is required. 39796822904SAndy Whitcroft * Subtle, allocate a new region at the position but make it zero 39896822904SAndy Whitcroft * size such that we can guarantee to record the reservation. */ 39996822904SAndy Whitcroft if (&rg->link == head || t < rg->from) { 4007b24d861SDavidlohr Bueso if (!nrg) { 4015e911373SMike Kravetz resv->adds_in_progress--; 4027b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 40396822904SAndy Whitcroft nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 40496822904SAndy Whitcroft if (!nrg) 40596822904SAndy Whitcroft return -ENOMEM; 4067b24d861SDavidlohr Bueso 40796822904SAndy Whitcroft nrg->from = f; 40896822904SAndy Whitcroft nrg->to = f; 40996822904SAndy Whitcroft INIT_LIST_HEAD(&nrg->link); 4107b24d861SDavidlohr Bueso goto retry; 4117b24d861SDavidlohr Bueso } 41296822904SAndy Whitcroft 4137b24d861SDavidlohr Bueso list_add(&nrg->link, rg->link.prev); 4147b24d861SDavidlohr Bueso chg = t - f; 4157b24d861SDavidlohr Bueso goto out_nrg; 41696822904SAndy Whitcroft } 41796822904SAndy Whitcroft 41896822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 41996822904SAndy Whitcroft if (f > rg->from) 42096822904SAndy Whitcroft f = rg->from; 42196822904SAndy Whitcroft chg = t - f; 42296822904SAndy Whitcroft 42396822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 42496822904SAndy Whitcroft list_for_each_entry(rg, rg->link.prev, link) { 42596822904SAndy Whitcroft if (&rg->link == head) 42696822904SAndy Whitcroft break; 42796822904SAndy Whitcroft if (rg->from > t) 4287b24d861SDavidlohr Bueso goto out; 42996822904SAndy Whitcroft 43025985edcSLucas De Marchi /* We overlap with this area, if it extends further than 43196822904SAndy Whitcroft * us then we must extend ourselves. Account for its 43296822904SAndy Whitcroft * existing reservation. */ 43396822904SAndy Whitcroft if (rg->to > t) { 43496822904SAndy Whitcroft chg += rg->to - t; 43596822904SAndy Whitcroft t = rg->to; 43696822904SAndy Whitcroft } 43796822904SAndy Whitcroft chg -= rg->to - rg->from; 43896822904SAndy Whitcroft } 4397b24d861SDavidlohr Bueso 4407b24d861SDavidlohr Bueso out: 4417b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 4427b24d861SDavidlohr Bueso /* We already know we raced and no longer need the new region */ 4437b24d861SDavidlohr Bueso kfree(nrg); 4447b24d861SDavidlohr Bueso return chg; 4457b24d861SDavidlohr Bueso out_nrg: 4467b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 44796822904SAndy Whitcroft return chg; 44896822904SAndy Whitcroft } 44996822904SAndy Whitcroft 4501dd308a7SMike Kravetz /* 4515e911373SMike Kravetz * Abort the in progress add operation. The adds_in_progress field 4525e911373SMike Kravetz * of the resv_map keeps track of the operations in progress between 4535e911373SMike Kravetz * calls to region_chg and region_add. Operations are sometimes 4545e911373SMike Kravetz * aborted after the call to region_chg. In such cases, region_abort 4555e911373SMike Kravetz * is called to decrement the adds_in_progress counter. 4565e911373SMike Kravetz * 4575e911373SMike Kravetz * NOTE: The range arguments [f, t) are not needed or used in this 4585e911373SMike Kravetz * routine. They are kept to make reading the calling code easier as 4595e911373SMike Kravetz * arguments will match the associated region_chg call. 4605e911373SMike Kravetz */ 4615e911373SMike Kravetz static void region_abort(struct resv_map *resv, long f, long t) 4625e911373SMike Kravetz { 4635e911373SMike Kravetz spin_lock(&resv->lock); 4645e911373SMike Kravetz VM_BUG_ON(!resv->region_cache_count); 4655e911373SMike Kravetz resv->adds_in_progress--; 4665e911373SMike Kravetz spin_unlock(&resv->lock); 4675e911373SMike Kravetz } 4685e911373SMike Kravetz 4695e911373SMike Kravetz /* 470feba16e2SMike Kravetz * Delete the specified range [f, t) from the reserve map. If the 471feba16e2SMike Kravetz * t parameter is LONG_MAX, this indicates that ALL regions after f 472feba16e2SMike Kravetz * should be deleted. Locate the regions which intersect [f, t) 473feba16e2SMike Kravetz * and either trim, delete or split the existing regions. 474feba16e2SMike Kravetz * 475feba16e2SMike Kravetz * Returns the number of huge pages deleted from the reserve map. 476feba16e2SMike Kravetz * In the normal case, the return value is zero or more. In the 477feba16e2SMike Kravetz * case where a region must be split, a new region descriptor must 478feba16e2SMike Kravetz * be allocated. If the allocation fails, -ENOMEM will be returned. 479feba16e2SMike Kravetz * NOTE: If the parameter t == LONG_MAX, then we will never split 480feba16e2SMike Kravetz * a region and possibly return -ENOMEM. Callers specifying 481feba16e2SMike Kravetz * t == LONG_MAX do not need to check for -ENOMEM error. 4821dd308a7SMike Kravetz */ 483feba16e2SMike Kravetz static long region_del(struct resv_map *resv, long f, long t) 48496822904SAndy Whitcroft { 4851406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 48696822904SAndy Whitcroft struct file_region *rg, *trg; 487feba16e2SMike Kravetz struct file_region *nrg = NULL; 488feba16e2SMike Kravetz long del = 0; 48996822904SAndy Whitcroft 490feba16e2SMike Kravetz retry: 4917b24d861SDavidlohr Bueso spin_lock(&resv->lock); 492feba16e2SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 493dbe409e4SMike Kravetz /* 494dbe409e4SMike Kravetz * Skip regions before the range to be deleted. file_region 495dbe409e4SMike Kravetz * ranges are normally of the form [from, to). However, there 496dbe409e4SMike Kravetz * may be a "placeholder" entry in the map which is of the form 497dbe409e4SMike Kravetz * (from, to) with from == to. Check for placeholder entries 498dbe409e4SMike Kravetz * at the beginning of the range to be deleted. 499dbe409e4SMike Kravetz */ 500dbe409e4SMike Kravetz if (rg->to <= f && (rg->to != rg->from || rg->to != f)) 501feba16e2SMike Kravetz continue; 502dbe409e4SMike Kravetz 503feba16e2SMike Kravetz if (rg->from >= t) 50496822904SAndy Whitcroft break; 50596822904SAndy Whitcroft 506feba16e2SMike Kravetz if (f > rg->from && t < rg->to) { /* Must split region */ 507feba16e2SMike Kravetz /* 508feba16e2SMike Kravetz * Check for an entry in the cache before dropping 509feba16e2SMike Kravetz * lock and attempting allocation. 510feba16e2SMike Kravetz */ 511feba16e2SMike Kravetz if (!nrg && 512feba16e2SMike Kravetz resv->region_cache_count > resv->adds_in_progress) { 513feba16e2SMike Kravetz nrg = list_first_entry(&resv->region_cache, 514feba16e2SMike Kravetz struct file_region, 515feba16e2SMike Kravetz link); 516feba16e2SMike Kravetz list_del(&nrg->link); 517feba16e2SMike Kravetz resv->region_cache_count--; 51896822904SAndy Whitcroft } 51996822904SAndy Whitcroft 520feba16e2SMike Kravetz if (!nrg) { 521feba16e2SMike Kravetz spin_unlock(&resv->lock); 522feba16e2SMike Kravetz nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 523feba16e2SMike Kravetz if (!nrg) 524feba16e2SMike Kravetz return -ENOMEM; 525feba16e2SMike Kravetz goto retry; 526feba16e2SMike Kravetz } 527feba16e2SMike Kravetz 528feba16e2SMike Kravetz del += t - f; 529feba16e2SMike Kravetz 530feba16e2SMike Kravetz /* New entry for end of split region */ 531feba16e2SMike Kravetz nrg->from = t; 532feba16e2SMike Kravetz nrg->to = rg->to; 533feba16e2SMike Kravetz INIT_LIST_HEAD(&nrg->link); 534feba16e2SMike Kravetz 535feba16e2SMike Kravetz /* Original entry is trimmed */ 536feba16e2SMike Kravetz rg->to = f; 537feba16e2SMike Kravetz 538feba16e2SMike Kravetz list_add(&nrg->link, &rg->link); 539feba16e2SMike Kravetz nrg = NULL; 54096822904SAndy Whitcroft break; 541feba16e2SMike Kravetz } 542feba16e2SMike Kravetz 543feba16e2SMike Kravetz if (f <= rg->from && t >= rg->to) { /* Remove entire region */ 544feba16e2SMike Kravetz del += rg->to - rg->from; 54596822904SAndy Whitcroft list_del(&rg->link); 54696822904SAndy Whitcroft kfree(rg); 547feba16e2SMike Kravetz continue; 54896822904SAndy Whitcroft } 5497b24d861SDavidlohr Bueso 550feba16e2SMike Kravetz if (f <= rg->from) { /* Trim beginning of region */ 551feba16e2SMike Kravetz del += t - rg->from; 552feba16e2SMike Kravetz rg->from = t; 553feba16e2SMike Kravetz } else { /* Trim end of region */ 554feba16e2SMike Kravetz del += rg->to - f; 555feba16e2SMike Kravetz rg->to = f; 556feba16e2SMike Kravetz } 557feba16e2SMike Kravetz } 558feba16e2SMike Kravetz 5597b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 560feba16e2SMike Kravetz kfree(nrg); 561feba16e2SMike Kravetz return del; 56296822904SAndy Whitcroft } 56396822904SAndy Whitcroft 5641dd308a7SMike Kravetz /* 565b5cec28dSMike Kravetz * A rare out of memory error was encountered which prevented removal of 566b5cec28dSMike Kravetz * the reserve map region for a page. The huge page itself was free'ed 567b5cec28dSMike Kravetz * and removed from the page cache. This routine will adjust the subpool 568b5cec28dSMike Kravetz * usage count, and the global reserve count if needed. By incrementing 569b5cec28dSMike Kravetz * these counts, the reserve map entry which could not be deleted will 570b5cec28dSMike Kravetz * appear as a "reserved" entry instead of simply dangling with incorrect 571b5cec28dSMike Kravetz * counts. 572b5cec28dSMike Kravetz */ 57372e2936cSzhong jiang void hugetlb_fix_reserve_counts(struct inode *inode) 574b5cec28dSMike Kravetz { 575b5cec28dSMike Kravetz struct hugepage_subpool *spool = subpool_inode(inode); 576b5cec28dSMike Kravetz long rsv_adjust; 577b5cec28dSMike Kravetz 578b5cec28dSMike Kravetz rsv_adjust = hugepage_subpool_get_pages(spool, 1); 57972e2936cSzhong jiang if (rsv_adjust) { 580b5cec28dSMike Kravetz struct hstate *h = hstate_inode(inode); 581b5cec28dSMike Kravetz 582b5cec28dSMike Kravetz hugetlb_acct_memory(h, 1); 583b5cec28dSMike Kravetz } 584b5cec28dSMike Kravetz } 585b5cec28dSMike Kravetz 586b5cec28dSMike Kravetz /* 5871dd308a7SMike Kravetz * Count and return the number of huge pages in the reserve map 5881dd308a7SMike Kravetz * that intersect with the range [f, t). 5891dd308a7SMike Kravetz */ 5901406ec9bSJoonsoo Kim static long region_count(struct resv_map *resv, long f, long t) 59184afd99bSAndy Whitcroft { 5921406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 59384afd99bSAndy Whitcroft struct file_region *rg; 59484afd99bSAndy Whitcroft long chg = 0; 59584afd99bSAndy Whitcroft 5967b24d861SDavidlohr Bueso spin_lock(&resv->lock); 59784afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 59884afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 599f2135a4aSWang Sheng-Hui long seg_from; 600f2135a4aSWang Sheng-Hui long seg_to; 60184afd99bSAndy Whitcroft 60284afd99bSAndy Whitcroft if (rg->to <= f) 60384afd99bSAndy Whitcroft continue; 60484afd99bSAndy Whitcroft if (rg->from >= t) 60584afd99bSAndy Whitcroft break; 60684afd99bSAndy Whitcroft 60784afd99bSAndy Whitcroft seg_from = max(rg->from, f); 60884afd99bSAndy Whitcroft seg_to = min(rg->to, t); 60984afd99bSAndy Whitcroft 61084afd99bSAndy Whitcroft chg += seg_to - seg_from; 61184afd99bSAndy Whitcroft } 6127b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 61384afd99bSAndy Whitcroft 61484afd99bSAndy Whitcroft return chg; 61584afd99bSAndy Whitcroft } 61684afd99bSAndy Whitcroft 61796822904SAndy Whitcroft /* 618e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 619e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 620e7c4b0bfSAndy Whitcroft */ 621a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 622a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 623e7c4b0bfSAndy Whitcroft { 624a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 625a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 626e7c4b0bfSAndy Whitcroft } 627e7c4b0bfSAndy Whitcroft 6280fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 6290fe6e20bSNaoya Horiguchi unsigned long address) 6300fe6e20bSNaoya Horiguchi { 6310fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 6320fe6e20bSNaoya Horiguchi } 633dee41079SDan Williams EXPORT_SYMBOL_GPL(linear_hugepage_index); 6340fe6e20bSNaoya Horiguchi 63584afd99bSAndy Whitcroft /* 63608fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 63708fba699SMel Gorman * cases this will be same size as used by the page table entries. 63808fba699SMel Gorman */ 63908fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 64008fba699SMel Gorman { 64105ea8860SDan Williams if (vma->vm_ops && vma->vm_ops->pagesize) 64205ea8860SDan Williams return vma->vm_ops->pagesize(vma); 64308fba699SMel Gorman return PAGE_SIZE; 64408fba699SMel Gorman } 645f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 64608fba699SMel Gorman 64708fba699SMel Gorman /* 6483340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 6493340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 65009135cc5SDan Williams * architectures where it differs, an architecture-specific 'strong' 65109135cc5SDan Williams * version of this symbol is required. 6523340289dSMel Gorman */ 65309135cc5SDan Williams __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 6543340289dSMel Gorman { 6553340289dSMel Gorman return vma_kernel_pagesize(vma); 6563340289dSMel Gorman } 6573340289dSMel Gorman 6583340289dSMel Gorman /* 65984afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 66084afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 66184afd99bSAndy Whitcroft * alignment. 66284afd99bSAndy Whitcroft */ 66384afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 66484afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 66504f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 66684afd99bSAndy Whitcroft 667a1e78772SMel Gorman /* 668a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 669a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 670a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 671a1e78772SMel Gorman * 672a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 673a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 674a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 675a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 67684afd99bSAndy Whitcroft * 67784afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 67884afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 67984afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 68084afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 68184afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 68284afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 68384afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 68484afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 685a1e78772SMel Gorman */ 686e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 687e7c4b0bfSAndy Whitcroft { 688e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 689e7c4b0bfSAndy Whitcroft } 690e7c4b0bfSAndy Whitcroft 691e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 692e7c4b0bfSAndy Whitcroft unsigned long value) 693e7c4b0bfSAndy Whitcroft { 694e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 695e7c4b0bfSAndy Whitcroft } 696e7c4b0bfSAndy Whitcroft 6979119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 69884afd99bSAndy Whitcroft { 69984afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 7005e911373SMike Kravetz struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 7015e911373SMike Kravetz 7025e911373SMike Kravetz if (!resv_map || !rg) { 7035e911373SMike Kravetz kfree(resv_map); 7045e911373SMike Kravetz kfree(rg); 70584afd99bSAndy Whitcroft return NULL; 7065e911373SMike Kravetz } 70784afd99bSAndy Whitcroft 70884afd99bSAndy Whitcroft kref_init(&resv_map->refs); 7097b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 71084afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 71184afd99bSAndy Whitcroft 7125e911373SMike Kravetz resv_map->adds_in_progress = 0; 7135e911373SMike Kravetz 7145e911373SMike Kravetz INIT_LIST_HEAD(&resv_map->region_cache); 7155e911373SMike Kravetz list_add(&rg->link, &resv_map->region_cache); 7165e911373SMike Kravetz resv_map->region_cache_count = 1; 7175e911373SMike Kravetz 71884afd99bSAndy Whitcroft return resv_map; 71984afd99bSAndy Whitcroft } 72084afd99bSAndy Whitcroft 7219119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 72284afd99bSAndy Whitcroft { 72384afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 7245e911373SMike Kravetz struct list_head *head = &resv_map->region_cache; 7255e911373SMike Kravetz struct file_region *rg, *trg; 72684afd99bSAndy Whitcroft 72784afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 728feba16e2SMike Kravetz region_del(resv_map, 0, LONG_MAX); 7295e911373SMike Kravetz 7305e911373SMike Kravetz /* ... and any entries left in the cache */ 7315e911373SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 7325e911373SMike Kravetz list_del(&rg->link); 7335e911373SMike Kravetz kfree(rg); 7345e911373SMike Kravetz } 7355e911373SMike Kravetz 7365e911373SMike Kravetz VM_BUG_ON(resv_map->adds_in_progress); 7375e911373SMike Kravetz 73884afd99bSAndy Whitcroft kfree(resv_map); 73984afd99bSAndy Whitcroft } 74084afd99bSAndy Whitcroft 7414e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 7424e35f483SJoonsoo Kim { 7434e35f483SJoonsoo Kim return inode->i_mapping->private_data; 7444e35f483SJoonsoo Kim } 7454e35f483SJoonsoo Kim 74684afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 747a1e78772SMel Gorman { 74881d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 7494e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 7504e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 7514e35f483SJoonsoo Kim struct inode *inode = mapping->host; 7524e35f483SJoonsoo Kim 7534e35f483SJoonsoo Kim return inode_resv_map(inode); 7544e35f483SJoonsoo Kim 7554e35f483SJoonsoo Kim } else { 75684afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 75784afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 7584e35f483SJoonsoo Kim } 759a1e78772SMel Gorman } 760a1e78772SMel Gorman 76184afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 762a1e78772SMel Gorman { 76381d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 76481d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 765a1e78772SMel Gorman 76684afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 76784afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 76804f2cbe3SMel Gorman } 76904f2cbe3SMel Gorman 77004f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 77104f2cbe3SMel Gorman { 77281d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 77381d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 774e7c4b0bfSAndy Whitcroft 775e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 77604f2cbe3SMel Gorman } 77704f2cbe3SMel Gorman 77804f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 77904f2cbe3SMel Gorman { 78081d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 781e7c4b0bfSAndy Whitcroft 782e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 783a1e78772SMel Gorman } 784a1e78772SMel Gorman 78504f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 786a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 787a1e78772SMel Gorman { 78881d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 789f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 790a1e78772SMel Gorman vma->vm_private_data = (void *)0; 791a1e78772SMel Gorman } 792a1e78772SMel Gorman 793a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 794559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 795a1e78772SMel Gorman { 796af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 797af0ed73eSJoonsoo Kim /* 798af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 799af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 800af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 801af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 802af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 803af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 804af0ed73eSJoonsoo Kim * properly, so add work-around here. 805af0ed73eSJoonsoo Kim */ 806af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 807559ec2f8SNicholas Krause return true; 808af0ed73eSJoonsoo Kim else 809559ec2f8SNicholas Krause return false; 810af0ed73eSJoonsoo Kim } 811a63884e9SJoonsoo Kim 812a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 8131fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 8141fb1b0e9SMike Kravetz /* 8151fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 8161fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 8171fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 8181fb1b0e9SMike Kravetz * fallocate. In this case, there really are no reverves to 8191fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 8201fb1b0e9SMike Kravetz */ 8211fb1b0e9SMike Kravetz if (chg) 8221fb1b0e9SMike Kravetz return false; 8231fb1b0e9SMike Kravetz else 824559ec2f8SNicholas Krause return true; 8251fb1b0e9SMike Kravetz } 826a63884e9SJoonsoo Kim 827a63884e9SJoonsoo Kim /* 828a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 829a63884e9SJoonsoo Kim * private mappings. 830a63884e9SJoonsoo Kim */ 83167961f9dSMike Kravetz if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 83267961f9dSMike Kravetz /* 83367961f9dSMike Kravetz * Like the shared case above, a hole punch or truncate 83467961f9dSMike Kravetz * could have been performed on the private mapping. 83567961f9dSMike Kravetz * Examine the value of chg to determine if reserves 83667961f9dSMike Kravetz * actually exist or were previously consumed. 83767961f9dSMike Kravetz * Very Subtle - The value of chg comes from a previous 83867961f9dSMike Kravetz * call to vma_needs_reserves(). The reserve map for 83967961f9dSMike Kravetz * private mappings has different (opposite) semantics 84067961f9dSMike Kravetz * than that of shared mappings. vma_needs_reserves() 84167961f9dSMike Kravetz * has already taken this difference in semantics into 84267961f9dSMike Kravetz * account. Therefore, the meaning of chg is the same 84367961f9dSMike Kravetz * as in the shared case above. Code could easily be 84467961f9dSMike Kravetz * combined, but keeping it separate draws attention to 84567961f9dSMike Kravetz * subtle differences. 84667961f9dSMike Kravetz */ 84767961f9dSMike Kravetz if (chg) 84867961f9dSMike Kravetz return false; 84967961f9dSMike Kravetz else 850559ec2f8SNicholas Krause return true; 85167961f9dSMike Kravetz } 852a63884e9SJoonsoo Kim 853559ec2f8SNicholas Krause return false; 854a1e78772SMel Gorman } 855a1e78772SMel Gorman 856a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 8571da177e4SLinus Torvalds { 8581da177e4SLinus Torvalds int nid = page_to_nid(page); 8590edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 860a5516438SAndi Kleen h->free_huge_pages++; 861a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 8621da177e4SLinus Torvalds } 8631da177e4SLinus Torvalds 86494310cbcSAnshuman Khandual static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) 865bf50bab2SNaoya Horiguchi { 866bf50bab2SNaoya Horiguchi struct page *page; 867bf50bab2SNaoya Horiguchi 868c8721bbbSNaoya Horiguchi list_for_each_entry(page, &h->hugepage_freelists[nid], lru) 869243abd5bSNaoya Horiguchi if (!PageHWPoison(page)) 870c8721bbbSNaoya Horiguchi break; 871c8721bbbSNaoya Horiguchi /* 872c8721bbbSNaoya Horiguchi * if 'non-isolated free hugepage' not found on the list, 873c8721bbbSNaoya Horiguchi * the allocation fails. 874c8721bbbSNaoya Horiguchi */ 875c8721bbbSNaoya Horiguchi if (&h->hugepage_freelists[nid] == &page->lru) 876bf50bab2SNaoya Horiguchi return NULL; 8770edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 878a9869b83SNaoya Horiguchi set_page_refcounted(page); 879bf50bab2SNaoya Horiguchi h->free_huge_pages--; 880bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 881bf50bab2SNaoya Horiguchi return page; 882bf50bab2SNaoya Horiguchi } 883bf50bab2SNaoya Horiguchi 8843e59fcb0SMichal Hocko static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, 8853e59fcb0SMichal Hocko nodemask_t *nmask) 88694310cbcSAnshuman Khandual { 8873e59fcb0SMichal Hocko unsigned int cpuset_mems_cookie; 8883e59fcb0SMichal Hocko struct zonelist *zonelist; 8893e59fcb0SMichal Hocko struct zone *zone; 8903e59fcb0SMichal Hocko struct zoneref *z; 89198fa15f3SAnshuman Khandual int node = NUMA_NO_NODE; 8923e59fcb0SMichal Hocko 8933e59fcb0SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 8943e59fcb0SMichal Hocko 8953e59fcb0SMichal Hocko retry_cpuset: 8963e59fcb0SMichal Hocko cpuset_mems_cookie = read_mems_allowed_begin(); 8973e59fcb0SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 89894310cbcSAnshuman Khandual struct page *page; 89994310cbcSAnshuman Khandual 9003e59fcb0SMichal Hocko if (!cpuset_zone_allowed(zone, gfp_mask)) 9013e59fcb0SMichal Hocko continue; 9023e59fcb0SMichal Hocko /* 9033e59fcb0SMichal Hocko * no need to ask again on the same node. Pool is node rather than 9043e59fcb0SMichal Hocko * zone aware 9053e59fcb0SMichal Hocko */ 9063e59fcb0SMichal Hocko if (zone_to_nid(zone) == node) 9073e59fcb0SMichal Hocko continue; 9083e59fcb0SMichal Hocko node = zone_to_nid(zone); 90994310cbcSAnshuman Khandual 91094310cbcSAnshuman Khandual page = dequeue_huge_page_node_exact(h, node); 91194310cbcSAnshuman Khandual if (page) 91294310cbcSAnshuman Khandual return page; 91394310cbcSAnshuman Khandual } 9143e59fcb0SMichal Hocko if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 9153e59fcb0SMichal Hocko goto retry_cpuset; 9163e59fcb0SMichal Hocko 91794310cbcSAnshuman Khandual return NULL; 91894310cbcSAnshuman Khandual } 91994310cbcSAnshuman Khandual 92086cdb465SNaoya Horiguchi /* Movability of hugepages depends on migration support. */ 92186cdb465SNaoya Horiguchi static inline gfp_t htlb_alloc_mask(struct hstate *h) 92286cdb465SNaoya Horiguchi { 9237ed2c31dSAnshuman Khandual if (hugepage_movable_supported(h)) 92486cdb465SNaoya Horiguchi return GFP_HIGHUSER_MOVABLE; 92586cdb465SNaoya Horiguchi else 92686cdb465SNaoya Horiguchi return GFP_HIGHUSER; 92786cdb465SNaoya Horiguchi } 92886cdb465SNaoya Horiguchi 929a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 930a5516438SAndi Kleen struct vm_area_struct *vma, 931af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 932af0ed73eSJoonsoo Kim long chg) 9331da177e4SLinus Torvalds { 9343e59fcb0SMichal Hocko struct page *page; 935480eccf9SLee Schermerhorn struct mempolicy *mpol; 93604ec6264SVlastimil Babka gfp_t gfp_mask; 9373e59fcb0SMichal Hocko nodemask_t *nodemask; 93804ec6264SVlastimil Babka int nid; 9391da177e4SLinus Torvalds 940a1e78772SMel Gorman /* 941a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 942a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 943a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 944a1e78772SMel Gorman */ 945af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg) && 946a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 947c0ff7453SMiao Xie goto err; 948a1e78772SMel Gorman 94904f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 950a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 9516eab04a8SJustin P. Mattock goto err; 95204f2cbe3SMel Gorman 95304ec6264SVlastimil Babka gfp_mask = htlb_alloc_mask(h); 95404ec6264SVlastimil Babka nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 9553e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 9563e59fcb0SMichal Hocko if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { 95707443a85SJoonsoo Kim SetPagePrivate(page); 958a63884e9SJoonsoo Kim h->resv_huge_pages--; 959bf50bab2SNaoya Horiguchi } 960cc9a6c87SMel Gorman 961cc9a6c87SMel Gorman mpol_cond_put(mpol); 962cc9a6c87SMel Gorman return page; 963cc9a6c87SMel Gorman 964c0ff7453SMiao Xie err: 965cc9a6c87SMel Gorman return NULL; 9661da177e4SLinus Torvalds } 9671da177e4SLinus Torvalds 9681cac6f2cSLuiz Capitulino /* 9691cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 9701cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 9711cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 9721cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 9731cac6f2cSLuiz Capitulino * node for alloc or free. 9741cac6f2cSLuiz Capitulino */ 9751cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 9761cac6f2cSLuiz Capitulino { 9770edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 9781cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 9791cac6f2cSLuiz Capitulino 9801cac6f2cSLuiz Capitulino return nid; 9811cac6f2cSLuiz Capitulino } 9821cac6f2cSLuiz Capitulino 9831cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 9841cac6f2cSLuiz Capitulino { 9851cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 9861cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 9871cac6f2cSLuiz Capitulino return nid; 9881cac6f2cSLuiz Capitulino } 9891cac6f2cSLuiz Capitulino 9901cac6f2cSLuiz Capitulino /* 9911cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 9921cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 9931cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 9941cac6f2cSLuiz Capitulino * mask. 9951cac6f2cSLuiz Capitulino */ 9961cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 9971cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 9981cac6f2cSLuiz Capitulino { 9991cac6f2cSLuiz Capitulino int nid; 10001cac6f2cSLuiz Capitulino 10011cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 10021cac6f2cSLuiz Capitulino 10031cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 10041cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 10051cac6f2cSLuiz Capitulino 10061cac6f2cSLuiz Capitulino return nid; 10071cac6f2cSLuiz Capitulino } 10081cac6f2cSLuiz Capitulino 10091cac6f2cSLuiz Capitulino /* 10101cac6f2cSLuiz Capitulino * helper for free_pool_huge_page() - return the previously saved 10111cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 10121cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 10131cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 10141cac6f2cSLuiz Capitulino */ 10151cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 10161cac6f2cSLuiz Capitulino { 10171cac6f2cSLuiz Capitulino int nid; 10181cac6f2cSLuiz Capitulino 10191cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 10201cac6f2cSLuiz Capitulino 10211cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 10221cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 10231cac6f2cSLuiz Capitulino 10241cac6f2cSLuiz Capitulino return nid; 10251cac6f2cSLuiz Capitulino } 10261cac6f2cSLuiz Capitulino 10271cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 10281cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 10291cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 10301cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 10311cac6f2cSLuiz Capitulino nr_nodes--) 10321cac6f2cSLuiz Capitulino 10331cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(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_free(hs, mask)) || 1); \ 10371cac6f2cSLuiz Capitulino nr_nodes--) 10381cac6f2cSLuiz Capitulino 1039e1073d1eSAneesh Kumar K.V #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 1040944d9fecSLuiz Capitulino static void destroy_compound_gigantic_page(struct page *page, 1041d00181b9SKirill A. Shutemov unsigned int order) 1042944d9fecSLuiz Capitulino { 1043944d9fecSLuiz Capitulino int i; 1044944d9fecSLuiz Capitulino int nr_pages = 1 << order; 1045944d9fecSLuiz Capitulino struct page *p = page + 1; 1046944d9fecSLuiz Capitulino 1047c8cc708aSGerald Schaefer atomic_set(compound_mapcount_ptr(page), 0); 1048944d9fecSLuiz Capitulino for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 10491d798ca3SKirill A. Shutemov clear_compound_head(p); 1050944d9fecSLuiz Capitulino set_page_refcounted(p); 1051944d9fecSLuiz Capitulino } 1052944d9fecSLuiz Capitulino 1053944d9fecSLuiz Capitulino set_compound_order(page, 0); 1054944d9fecSLuiz Capitulino __ClearPageHead(page); 1055944d9fecSLuiz Capitulino } 1056944d9fecSLuiz Capitulino 1057d00181b9SKirill A. Shutemov static void free_gigantic_page(struct page *page, unsigned int order) 1058944d9fecSLuiz Capitulino { 1059944d9fecSLuiz Capitulino free_contig_range(page_to_pfn(page), 1 << order); 1060944d9fecSLuiz Capitulino } 1061944d9fecSLuiz Capitulino 1062944d9fecSLuiz Capitulino static int __alloc_gigantic_page(unsigned long start_pfn, 106379b63f12SMichal Hocko unsigned long nr_pages, gfp_t gfp_mask) 1064944d9fecSLuiz Capitulino { 1065944d9fecSLuiz Capitulino unsigned long end_pfn = start_pfn + nr_pages; 1066ca96b625SLucas Stach return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE, 106779b63f12SMichal Hocko gfp_mask); 1068944d9fecSLuiz Capitulino } 1069944d9fecSLuiz Capitulino 1070f44b2ddaSJoonsoo Kim static bool pfn_range_valid_gigantic(struct zone *z, 1071f44b2ddaSJoonsoo Kim unsigned long start_pfn, unsigned long nr_pages) 1072944d9fecSLuiz Capitulino { 1073944d9fecSLuiz Capitulino unsigned long i, end_pfn = start_pfn + nr_pages; 1074944d9fecSLuiz Capitulino struct page *page; 1075944d9fecSLuiz Capitulino 1076944d9fecSLuiz Capitulino for (i = start_pfn; i < end_pfn; i++) { 1077944d9fecSLuiz Capitulino if (!pfn_valid(i)) 1078944d9fecSLuiz Capitulino return false; 1079944d9fecSLuiz Capitulino 1080944d9fecSLuiz Capitulino page = pfn_to_page(i); 1081944d9fecSLuiz Capitulino 1082f44b2ddaSJoonsoo Kim if (page_zone(page) != z) 1083f44b2ddaSJoonsoo Kim return false; 1084f44b2ddaSJoonsoo Kim 1085944d9fecSLuiz Capitulino if (PageReserved(page)) 1086944d9fecSLuiz Capitulino return false; 1087944d9fecSLuiz Capitulino 1088944d9fecSLuiz Capitulino if (page_count(page) > 0) 1089944d9fecSLuiz Capitulino return false; 1090944d9fecSLuiz Capitulino 1091944d9fecSLuiz Capitulino if (PageHuge(page)) 1092944d9fecSLuiz Capitulino return false; 1093944d9fecSLuiz Capitulino } 1094944d9fecSLuiz Capitulino 1095944d9fecSLuiz Capitulino return true; 1096944d9fecSLuiz Capitulino } 1097944d9fecSLuiz Capitulino 1098944d9fecSLuiz Capitulino static bool zone_spans_last_pfn(const struct zone *zone, 1099944d9fecSLuiz Capitulino unsigned long start_pfn, unsigned long nr_pages) 1100944d9fecSLuiz Capitulino { 1101944d9fecSLuiz Capitulino unsigned long last_pfn = start_pfn + nr_pages - 1; 1102944d9fecSLuiz Capitulino return zone_spans_pfn(zone, last_pfn); 1103944d9fecSLuiz Capitulino } 1104944d9fecSLuiz Capitulino 1105d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 1106d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) 1107944d9fecSLuiz Capitulino { 110879b63f12SMichal Hocko unsigned int order = huge_page_order(h); 1109944d9fecSLuiz Capitulino unsigned long nr_pages = 1 << order; 1110944d9fecSLuiz Capitulino unsigned long ret, pfn, flags; 111179b63f12SMichal Hocko struct zonelist *zonelist; 111279b63f12SMichal Hocko struct zone *zone; 111379b63f12SMichal Hocko struct zoneref *z; 1114944d9fecSLuiz Capitulino 111579b63f12SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 1116d9cc948fSMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nodemask) { 111779b63f12SMichal Hocko spin_lock_irqsave(&zone->lock, flags); 1118944d9fecSLuiz Capitulino 111979b63f12SMichal Hocko pfn = ALIGN(zone->zone_start_pfn, nr_pages); 112079b63f12SMichal Hocko while (zone_spans_last_pfn(zone, pfn, nr_pages)) { 112179b63f12SMichal Hocko if (pfn_range_valid_gigantic(zone, pfn, nr_pages)) { 1122944d9fecSLuiz Capitulino /* 1123944d9fecSLuiz Capitulino * We release the zone lock here because 1124944d9fecSLuiz Capitulino * alloc_contig_range() will also lock the zone 1125944d9fecSLuiz Capitulino * at some point. If there's an allocation 1126944d9fecSLuiz Capitulino * spinning on this lock, it may win the race 1127944d9fecSLuiz Capitulino * and cause alloc_contig_range() to fail... 1128944d9fecSLuiz Capitulino */ 112979b63f12SMichal Hocko spin_unlock_irqrestore(&zone->lock, flags); 113079b63f12SMichal Hocko ret = __alloc_gigantic_page(pfn, nr_pages, gfp_mask); 1131944d9fecSLuiz Capitulino if (!ret) 1132944d9fecSLuiz Capitulino return pfn_to_page(pfn); 113379b63f12SMichal Hocko spin_lock_irqsave(&zone->lock, flags); 1134944d9fecSLuiz Capitulino } 1135944d9fecSLuiz Capitulino pfn += nr_pages; 1136944d9fecSLuiz Capitulino } 1137944d9fecSLuiz Capitulino 113879b63f12SMichal Hocko spin_unlock_irqrestore(&zone->lock, flags); 1139944d9fecSLuiz Capitulino } 1140944d9fecSLuiz Capitulino 1141944d9fecSLuiz Capitulino return NULL; 1142944d9fecSLuiz Capitulino } 1143944d9fecSLuiz Capitulino 1144944d9fecSLuiz Capitulino static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); 1145d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order); 1146944d9fecSLuiz Capitulino 1147e1073d1eSAneesh Kumar K.V #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 1148944d9fecSLuiz Capitulino static inline bool gigantic_page_supported(void) { return false; } 1149d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 1150d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) { return NULL; } 1151d00181b9SKirill A. Shutemov static inline void free_gigantic_page(struct page *page, unsigned int order) { } 1152944d9fecSLuiz Capitulino static inline void destroy_compound_gigantic_page(struct page *page, 1153d00181b9SKirill A. Shutemov unsigned int order) { } 1154944d9fecSLuiz Capitulino #endif 1155944d9fecSLuiz Capitulino 1156a5516438SAndi Kleen static void update_and_free_page(struct hstate *h, struct page *page) 11576af2acb6SAdam Litke { 11586af2acb6SAdam Litke int i; 1159a5516438SAndi Kleen 1160944d9fecSLuiz Capitulino if (hstate_is_gigantic(h) && !gigantic_page_supported()) 1161944d9fecSLuiz Capitulino return; 116218229df5SAndy Whitcroft 1163a5516438SAndi Kleen h->nr_huge_pages--; 1164a5516438SAndi Kleen h->nr_huge_pages_node[page_to_nid(page)]--; 1165a5516438SAndi Kleen for (i = 0; i < pages_per_huge_page(h); i++) { 116632f84528SChris Forbes page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 116732f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1168a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1169a7407a27SLuiz Capitulino 1 << PG_writeback); 11706af2acb6SAdam Litke } 1171309381feSSasha Levin VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 1172f1e61557SKirill A. Shutemov set_compound_page_dtor(page, NULL_COMPOUND_DTOR); 11736af2acb6SAdam Litke set_page_refcounted(page); 1174944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) { 1175944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 1176944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 1177944d9fecSLuiz Capitulino } else { 1178a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 11796af2acb6SAdam Litke } 1180944d9fecSLuiz Capitulino } 11816af2acb6SAdam Litke 1182e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1183e5ff2159SAndi Kleen { 1184e5ff2159SAndi Kleen struct hstate *h; 1185e5ff2159SAndi Kleen 1186e5ff2159SAndi Kleen for_each_hstate(h) { 1187e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1188e5ff2159SAndi Kleen return h; 1189e5ff2159SAndi Kleen } 1190e5ff2159SAndi Kleen return NULL; 1191e5ff2159SAndi Kleen } 1192e5ff2159SAndi Kleen 1193bcc54222SNaoya Horiguchi /* 1194bcc54222SNaoya Horiguchi * Test to determine whether the hugepage is "active/in-use" (i.e. being linked 1195bcc54222SNaoya Horiguchi * to hstate->hugepage_activelist.) 1196bcc54222SNaoya Horiguchi * 1197bcc54222SNaoya Horiguchi * This function can be called for tail pages, but never returns true for them. 1198bcc54222SNaoya Horiguchi */ 1199bcc54222SNaoya Horiguchi bool page_huge_active(struct page *page) 1200bcc54222SNaoya Horiguchi { 1201bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHuge(page), page); 1202bcc54222SNaoya Horiguchi return PageHead(page) && PagePrivate(&page[1]); 1203bcc54222SNaoya Horiguchi } 1204bcc54222SNaoya Horiguchi 1205bcc54222SNaoya Horiguchi /* never called for tail page */ 1206bcc54222SNaoya Horiguchi static void set_page_huge_active(struct page *page) 1207bcc54222SNaoya Horiguchi { 1208bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 1209bcc54222SNaoya Horiguchi SetPagePrivate(&page[1]); 1210bcc54222SNaoya Horiguchi } 1211bcc54222SNaoya Horiguchi 1212bcc54222SNaoya Horiguchi static void clear_page_huge_active(struct page *page) 1213bcc54222SNaoya Horiguchi { 1214bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 1215bcc54222SNaoya Horiguchi ClearPagePrivate(&page[1]); 1216bcc54222SNaoya Horiguchi } 1217bcc54222SNaoya Horiguchi 1218ab5ac90aSMichal Hocko /* 1219ab5ac90aSMichal Hocko * Internal hugetlb specific page flag. Do not use outside of the hugetlb 1220ab5ac90aSMichal Hocko * code 1221ab5ac90aSMichal Hocko */ 1222ab5ac90aSMichal Hocko static inline bool PageHugeTemporary(struct page *page) 1223ab5ac90aSMichal Hocko { 1224ab5ac90aSMichal Hocko if (!PageHuge(page)) 1225ab5ac90aSMichal Hocko return false; 1226ab5ac90aSMichal Hocko 1227ab5ac90aSMichal Hocko return (unsigned long)page[2].mapping == -1U; 1228ab5ac90aSMichal Hocko } 1229ab5ac90aSMichal Hocko 1230ab5ac90aSMichal Hocko static inline void SetPageHugeTemporary(struct page *page) 1231ab5ac90aSMichal Hocko { 1232ab5ac90aSMichal Hocko page[2].mapping = (void *)-1U; 1233ab5ac90aSMichal Hocko } 1234ab5ac90aSMichal Hocko 1235ab5ac90aSMichal Hocko static inline void ClearPageHugeTemporary(struct page *page) 1236ab5ac90aSMichal Hocko { 1237ab5ac90aSMichal Hocko page[2].mapping = NULL; 1238ab5ac90aSMichal Hocko } 1239ab5ac90aSMichal Hocko 12408f1d26d0SAtsushi Kumagai void free_huge_page(struct page *page) 124127a85ef1SDavid Gibson { 1242a5516438SAndi Kleen /* 1243a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1244a5516438SAndi Kleen * compound page destructor. 1245a5516438SAndi Kleen */ 1246e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 12477893d1d5SAdam Litke int nid = page_to_nid(page); 124890481622SDavid Gibson struct hugepage_subpool *spool = 124990481622SDavid Gibson (struct hugepage_subpool *)page_private(page); 125007443a85SJoonsoo Kim bool restore_reserve; 125127a85ef1SDavid Gibson 1252b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1253b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_mapcount(page), page); 12548ace22bcSYongkai Wu 12558ace22bcSYongkai Wu set_page_private(page, 0); 12568ace22bcSYongkai Wu page->mapping = NULL; 125707443a85SJoonsoo Kim restore_reserve = PagePrivate(page); 125816c794b4SJoonsoo Kim ClearPagePrivate(page); 125927a85ef1SDavid Gibson 12601c5ecae3SMike Kravetz /* 12611c5ecae3SMike Kravetz * A return code of zero implies that the subpool will be under its 12621c5ecae3SMike Kravetz * minimum size if the reservation is not restored after page is free. 12631c5ecae3SMike Kravetz * Therefore, force restore_reserve operation. 12641c5ecae3SMike Kravetz */ 12651c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 12661c5ecae3SMike Kravetz restore_reserve = true; 12671c5ecae3SMike Kravetz 126827a85ef1SDavid Gibson spin_lock(&hugetlb_lock); 1269bcc54222SNaoya Horiguchi clear_page_huge_active(page); 12706d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 12716d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 127207443a85SJoonsoo Kim if (restore_reserve) 127307443a85SJoonsoo Kim h->resv_huge_pages++; 127407443a85SJoonsoo Kim 1275ab5ac90aSMichal Hocko if (PageHugeTemporary(page)) { 1276ab5ac90aSMichal Hocko list_del(&page->lru); 1277ab5ac90aSMichal Hocko ClearPageHugeTemporary(page); 1278ab5ac90aSMichal Hocko update_and_free_page(h, page); 1279ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 12800edaecfaSAneesh Kumar K.V /* remove the page from active list */ 12810edaecfaSAneesh Kumar K.V list_del(&page->lru); 1282a5516438SAndi Kleen update_and_free_page(h, page); 1283a5516438SAndi Kleen h->surplus_huge_pages--; 1284a5516438SAndi Kleen h->surplus_huge_pages_node[nid]--; 12857893d1d5SAdam Litke } else { 12865d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1287a5516438SAndi Kleen enqueue_huge_page(h, page); 12887893d1d5SAdam Litke } 128927a85ef1SDavid Gibson spin_unlock(&hugetlb_lock); 129027a85ef1SDavid Gibson } 129127a85ef1SDavid Gibson 1292a5516438SAndi Kleen static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1293b7ba30c6SAndi Kleen { 12940edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1295f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1296b7ba30c6SAndi Kleen spin_lock(&hugetlb_lock); 12979dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 1298a5516438SAndi Kleen h->nr_huge_pages++; 1299a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 1300b7ba30c6SAndi Kleen spin_unlock(&hugetlb_lock); 1301b7ba30c6SAndi Kleen } 1302b7ba30c6SAndi Kleen 1303d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order) 130420a0307cSWu Fengguang { 130520a0307cSWu Fengguang int i; 130620a0307cSWu Fengguang int nr_pages = 1 << order; 130720a0307cSWu Fengguang struct page *p = page + 1; 130820a0307cSWu Fengguang 130920a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 131020a0307cSWu Fengguang set_compound_order(page, order); 1311ef5a22beSAndrea Arcangeli __ClearPageReserved(page); 1312de09d31dSKirill A. Shutemov __SetPageHead(page); 131320a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1314ef5a22beSAndrea Arcangeli /* 1315ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1316ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1317ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 1318ef5a22beSAndrea Arcangeli * too. Otherwse drivers using get_user_pages() to access tail 1319ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1320ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1321ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1322ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1323ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1324ef5a22beSAndrea Arcangeli * after get_user_pages(). 1325ef5a22beSAndrea Arcangeli */ 1326ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 132758a84aa9SYouquan Song set_page_count(p, 0); 13281d798ca3SKirill A. Shutemov set_compound_head(p, page); 132920a0307cSWu Fengguang } 1330b4330afbSMike Kravetz atomic_set(compound_mapcount_ptr(page), -1); 133120a0307cSWu Fengguang } 133220a0307cSWu Fengguang 13337795912cSAndrew Morton /* 13347795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 13357795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 13367795912cSAndrew Morton * details. 13377795912cSAndrew Morton */ 133820a0307cSWu Fengguang int PageHuge(struct page *page) 133920a0307cSWu Fengguang { 134020a0307cSWu Fengguang if (!PageCompound(page)) 134120a0307cSWu Fengguang return 0; 134220a0307cSWu Fengguang 134320a0307cSWu Fengguang page = compound_head(page); 1344f1e61557SKirill A. Shutemov return page[1].compound_dtor == HUGETLB_PAGE_DTOR; 134520a0307cSWu Fengguang } 134643131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 134743131e14SNaoya Horiguchi 134827c73ae7SAndrea Arcangeli /* 134927c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 135027c73ae7SAndrea Arcangeli * normal or transparent huge pages. 135127c73ae7SAndrea Arcangeli */ 135227c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 135327c73ae7SAndrea Arcangeli { 135427c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 135527c73ae7SAndrea Arcangeli return 0; 135627c73ae7SAndrea Arcangeli 1357758f66a2SAndrew Morton return get_compound_page_dtor(page_head) == free_huge_page; 135827c73ae7SAndrea Arcangeli } 135927c73ae7SAndrea Arcangeli 136013d60f4bSZhang Yi pgoff_t __basepage_index(struct page *page) 136113d60f4bSZhang Yi { 136213d60f4bSZhang Yi struct page *page_head = compound_head(page); 136313d60f4bSZhang Yi pgoff_t index = page_index(page_head); 136413d60f4bSZhang Yi unsigned long compound_idx; 136513d60f4bSZhang Yi 136613d60f4bSZhang Yi if (!PageHuge(page_head)) 136713d60f4bSZhang Yi return page_index(page); 136813d60f4bSZhang Yi 136913d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 137013d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 137113d60f4bSZhang Yi else 137213d60f4bSZhang Yi compound_idx = page - page_head; 137313d60f4bSZhang Yi 137413d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 137513d60f4bSZhang Yi } 137613d60f4bSZhang Yi 13770c397daeSMichal Hocko static struct page *alloc_buddy_huge_page(struct hstate *h, 1378af0fb9dfSMichal Hocko gfp_t gfp_mask, int nid, nodemask_t *nmask) 13791da177e4SLinus Torvalds { 1380af0fb9dfSMichal Hocko int order = huge_page_order(h); 13811da177e4SLinus Torvalds struct page *page; 1382f96efd58SJoe Jin 1383af0fb9dfSMichal Hocko gfp_mask |= __GFP_COMP|__GFP_RETRY_MAYFAIL|__GFP_NOWARN; 1384af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 1385af0fb9dfSMichal Hocko nid = numa_mem_id(); 1386af0fb9dfSMichal Hocko page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask); 1387af0fb9dfSMichal Hocko if (page) 1388af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC); 1389af0fb9dfSMichal Hocko else 1390af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 139163b4613cSNishanth Aravamudan 139263b4613cSNishanth Aravamudan return page; 139363b4613cSNishanth Aravamudan } 139463b4613cSNishanth Aravamudan 1395af0fb9dfSMichal Hocko /* 13960c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 13970c397daeSMichal Hocko * should use this function to get new hugetlb pages 13980c397daeSMichal Hocko */ 13990c397daeSMichal Hocko static struct page *alloc_fresh_huge_page(struct hstate *h, 14000c397daeSMichal Hocko gfp_t gfp_mask, int nid, nodemask_t *nmask) 14010c397daeSMichal Hocko { 14020c397daeSMichal Hocko struct page *page; 14030c397daeSMichal Hocko 14040c397daeSMichal Hocko if (hstate_is_gigantic(h)) 14050c397daeSMichal Hocko page = alloc_gigantic_page(h, gfp_mask, nid, nmask); 14060c397daeSMichal Hocko else 14070c397daeSMichal Hocko page = alloc_buddy_huge_page(h, gfp_mask, 14080c397daeSMichal Hocko nid, nmask); 14090c397daeSMichal Hocko if (!page) 14100c397daeSMichal Hocko return NULL; 14110c397daeSMichal Hocko 14120c397daeSMichal Hocko if (hstate_is_gigantic(h)) 14130c397daeSMichal Hocko prep_compound_gigantic_page(page, huge_page_order(h)); 14140c397daeSMichal Hocko prep_new_huge_page(h, page, page_to_nid(page)); 14150c397daeSMichal Hocko 14160c397daeSMichal Hocko return page; 14170c397daeSMichal Hocko } 14180c397daeSMichal Hocko 14190c397daeSMichal Hocko /* 1420af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 1421af0fb9dfSMichal Hocko * manner. 1422af0fb9dfSMichal Hocko */ 14230c397daeSMichal Hocko static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 1424b2261026SJoonsoo Kim { 1425b2261026SJoonsoo Kim struct page *page; 1426b2261026SJoonsoo Kim int nr_nodes, node; 1427af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 1428b2261026SJoonsoo Kim 1429b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 14300c397daeSMichal Hocko page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed); 1431af0fb9dfSMichal Hocko if (page) 1432b2261026SJoonsoo Kim break; 1433b2261026SJoonsoo Kim } 1434b2261026SJoonsoo Kim 1435af0fb9dfSMichal Hocko if (!page) 1436af0fb9dfSMichal Hocko return 0; 1437b2261026SJoonsoo Kim 1438af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 1439af0fb9dfSMichal Hocko 1440af0fb9dfSMichal Hocko return 1; 1441b2261026SJoonsoo Kim } 1442b2261026SJoonsoo Kim 1443e8c5c824SLee Schermerhorn /* 1444e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 1445e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 1446e8c5c824SLee Schermerhorn * balanced over allowed nodes. 1447e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 1448e8c5c824SLee Schermerhorn */ 14496ae11b27SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 14506ae11b27SLee Schermerhorn bool acct_surplus) 1451e8c5c824SLee Schermerhorn { 1452b2261026SJoonsoo Kim int nr_nodes, node; 1453e8c5c824SLee Schermerhorn int ret = 0; 1454e8c5c824SLee Schermerhorn 1455b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1456685f3457SLee Schermerhorn /* 1457685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 1458685f3457SLee Schermerhorn * nodes with surplus pages. 1459685f3457SLee Schermerhorn */ 1460b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 1461b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 1462e8c5c824SLee Schermerhorn struct page *page = 1463b2261026SJoonsoo Kim list_entry(h->hugepage_freelists[node].next, 1464e8c5c824SLee Schermerhorn struct page, lru); 1465e8c5c824SLee Schermerhorn list_del(&page->lru); 1466e8c5c824SLee Schermerhorn h->free_huge_pages--; 1467b2261026SJoonsoo Kim h->free_huge_pages_node[node]--; 1468685f3457SLee Schermerhorn if (acct_surplus) { 1469685f3457SLee Schermerhorn h->surplus_huge_pages--; 1470b2261026SJoonsoo Kim h->surplus_huge_pages_node[node]--; 1471685f3457SLee Schermerhorn } 1472e8c5c824SLee Schermerhorn update_and_free_page(h, page); 1473e8c5c824SLee Schermerhorn ret = 1; 14749a76db09SLee Schermerhorn break; 1475e8c5c824SLee Schermerhorn } 1476b2261026SJoonsoo Kim } 1477e8c5c824SLee Schermerhorn 1478e8c5c824SLee Schermerhorn return ret; 1479e8c5c824SLee Schermerhorn } 1480e8c5c824SLee Schermerhorn 1481c8721bbbSNaoya Horiguchi /* 1482c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 1483082d5b6bSGerald Schaefer * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the 14846bc9b564SNaoya Horiguchi * dissolution fails because a give page is not a free hugepage, or because 14856bc9b564SNaoya Horiguchi * free hugepages are fully reserved. 1486c8721bbbSNaoya Horiguchi */ 1487c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 1488c8721bbbSNaoya Horiguchi { 14896bc9b564SNaoya Horiguchi int rc = -EBUSY; 1490082d5b6bSGerald Schaefer 1491c8721bbbSNaoya Horiguchi spin_lock(&hugetlb_lock); 1492c8721bbbSNaoya Horiguchi if (PageHuge(page) && !page_count(page)) { 14932247bb33SGerald Schaefer struct page *head = compound_head(page); 14942247bb33SGerald Schaefer struct hstate *h = page_hstate(head); 14952247bb33SGerald Schaefer int nid = page_to_nid(head); 14966bc9b564SNaoya Horiguchi if (h->free_huge_pages - h->resv_huge_pages == 0) 1497082d5b6bSGerald Schaefer goto out; 1498c3114a84SAnshuman Khandual /* 1499c3114a84SAnshuman Khandual * Move PageHWPoison flag from head page to the raw error page, 1500c3114a84SAnshuman Khandual * which makes any subpages rather than the error page reusable. 1501c3114a84SAnshuman Khandual */ 1502c3114a84SAnshuman Khandual if (PageHWPoison(head) && page != head) { 1503c3114a84SAnshuman Khandual SetPageHWPoison(page); 1504c3114a84SAnshuman Khandual ClearPageHWPoison(head); 1505c3114a84SAnshuman Khandual } 15062247bb33SGerald Schaefer list_del(&head->lru); 1507c8721bbbSNaoya Horiguchi h->free_huge_pages--; 1508c8721bbbSNaoya Horiguchi h->free_huge_pages_node[nid]--; 1509c1470b33Szhong jiang h->max_huge_pages--; 15102247bb33SGerald Schaefer update_and_free_page(h, head); 15116bc9b564SNaoya Horiguchi rc = 0; 1512c8721bbbSNaoya Horiguchi } 1513082d5b6bSGerald Schaefer out: 1514c8721bbbSNaoya Horiguchi spin_unlock(&hugetlb_lock); 1515082d5b6bSGerald Schaefer return rc; 1516c8721bbbSNaoya Horiguchi } 1517c8721bbbSNaoya Horiguchi 1518c8721bbbSNaoya Horiguchi /* 1519c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 1520c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 15212247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 15222247bb33SGerald Schaefer * part of it lies within the given range. 1523082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 1524082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 1525c8721bbbSNaoya Horiguchi */ 1526082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 1527c8721bbbSNaoya Horiguchi { 1528c8721bbbSNaoya Horiguchi unsigned long pfn; 1529eb03aa00SGerald Schaefer struct page *page; 1530082d5b6bSGerald Schaefer int rc = 0; 1531c8721bbbSNaoya Horiguchi 1532d0177639SLi Zhong if (!hugepages_supported()) 1533082d5b6bSGerald Schaefer return rc; 1534d0177639SLi Zhong 1535eb03aa00SGerald Schaefer for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { 1536eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 1537eb03aa00SGerald Schaefer if (PageHuge(page) && !page_count(page)) { 1538eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 1539eb03aa00SGerald Schaefer if (rc) 1540082d5b6bSGerald Schaefer break; 1541eb03aa00SGerald Schaefer } 1542eb03aa00SGerald Schaefer } 1543082d5b6bSGerald Schaefer 1544082d5b6bSGerald Schaefer return rc; 1545c8721bbbSNaoya Horiguchi } 1546c8721bbbSNaoya Horiguchi 1547ab5ac90aSMichal Hocko /* 1548ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 1549ab5ac90aSMichal Hocko */ 15500c397daeSMichal Hocko static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, 1551aaf14e40SMichal Hocko int nid, nodemask_t *nmask) 15527893d1d5SAdam Litke { 15539980d744SMichal Hocko struct page *page = NULL; 15547893d1d5SAdam Litke 1555bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1556aa888a74SAndi Kleen return NULL; 1557aa888a74SAndi Kleen 1558d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 15599980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 15609980d744SMichal Hocko goto out_unlock; 1561d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1562d1c3fb1fSNishanth Aravamudan 15630c397daeSMichal Hocko page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask); 15649980d744SMichal Hocko if (!page) 15650c397daeSMichal Hocko return NULL; 1566d1c3fb1fSNishanth Aravamudan 15677893d1d5SAdam Litke spin_lock(&hugetlb_lock); 15689980d744SMichal Hocko /* 15699980d744SMichal Hocko * We could have raced with the pool size change. 15709980d744SMichal Hocko * Double check that and simply deallocate the new page 15719980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 15729980d744SMichal Hocko * temporary page to workaround the nasty free_huge_page 15739980d744SMichal Hocko * codeflow 15749980d744SMichal Hocko */ 15759980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 15769980d744SMichal Hocko SetPageHugeTemporary(page); 15772bf753e6SKai Shen spin_unlock(&hugetlb_lock); 15789980d744SMichal Hocko put_page(page); 15792bf753e6SKai Shen return NULL; 15809980d744SMichal Hocko } else { 15819980d744SMichal Hocko h->surplus_huge_pages++; 15824704dea3SMichal Hocko h->surplus_huge_pages_node[page_to_nid(page)]++; 15837893d1d5SAdam Litke } 15849980d744SMichal Hocko 15859980d744SMichal Hocko out_unlock: 1586d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 15877893d1d5SAdam Litke 15887893d1d5SAdam Litke return page; 15897893d1d5SAdam Litke } 15907893d1d5SAdam Litke 15919a4e9f3bSAneesh Kumar K.V struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, 1592ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 1593ab5ac90aSMichal Hocko { 1594ab5ac90aSMichal Hocko struct page *page; 1595ab5ac90aSMichal Hocko 1596ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 1597ab5ac90aSMichal Hocko return NULL; 1598ab5ac90aSMichal Hocko 15990c397daeSMichal Hocko page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask); 1600ab5ac90aSMichal Hocko if (!page) 1601ab5ac90aSMichal Hocko return NULL; 1602ab5ac90aSMichal Hocko 1603ab5ac90aSMichal Hocko /* 1604ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 1605ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 1606ab5ac90aSMichal Hocko */ 1607ab5ac90aSMichal Hocko SetPageHugeTemporary(page); 1608ab5ac90aSMichal Hocko 1609ab5ac90aSMichal Hocko return page; 1610ab5ac90aSMichal Hocko } 1611ab5ac90aSMichal Hocko 1612e4e574b7SAdam Litke /* 1613099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 1614099730d6SDave Hansen */ 1615e0ec90eeSDave Hansen static 16160c397daeSMichal Hocko struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, 1617099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 1618099730d6SDave Hansen { 1619aaf14e40SMichal Hocko struct page *page; 1620aaf14e40SMichal Hocko struct mempolicy *mpol; 1621aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 1622aaf14e40SMichal Hocko int nid; 1623aaf14e40SMichal Hocko nodemask_t *nodemask; 1624aaf14e40SMichal Hocko 1625aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 16260c397daeSMichal Hocko page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); 1627aaf14e40SMichal Hocko mpol_cond_put(mpol); 1628aaf14e40SMichal Hocko 1629aaf14e40SMichal Hocko return page; 1630099730d6SDave Hansen } 1631099730d6SDave Hansen 1632ab5ac90aSMichal Hocko /* page migration callback function */ 1633bf50bab2SNaoya Horiguchi struct page *alloc_huge_page_node(struct hstate *h, int nid) 1634bf50bab2SNaoya Horiguchi { 1635aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 16364ef91848SJoonsoo Kim struct page *page = NULL; 1637bf50bab2SNaoya Horiguchi 1638aaf14e40SMichal Hocko if (nid != NUMA_NO_NODE) 1639aaf14e40SMichal Hocko gfp_mask |= __GFP_THISNODE; 1640aaf14e40SMichal Hocko 1641bf50bab2SNaoya Horiguchi spin_lock(&hugetlb_lock); 16424ef91848SJoonsoo Kim if (h->free_huge_pages - h->resv_huge_pages > 0) 16433e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL); 1644bf50bab2SNaoya Horiguchi spin_unlock(&hugetlb_lock); 1645bf50bab2SNaoya Horiguchi 164694ae8ba7SAneesh Kumar K.V if (!page) 16470c397daeSMichal Hocko page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL); 1648bf50bab2SNaoya Horiguchi 1649bf50bab2SNaoya Horiguchi return page; 1650bf50bab2SNaoya Horiguchi } 1651bf50bab2SNaoya Horiguchi 1652ab5ac90aSMichal Hocko /* page migration callback function */ 16533e59fcb0SMichal Hocko struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, 16543e59fcb0SMichal Hocko nodemask_t *nmask) 16554db9b2efSMichal Hocko { 1656aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 16574db9b2efSMichal Hocko 16584db9b2efSMichal Hocko spin_lock(&hugetlb_lock); 16594db9b2efSMichal Hocko if (h->free_huge_pages - h->resv_huge_pages > 0) { 16603e59fcb0SMichal Hocko struct page *page; 16613e59fcb0SMichal Hocko 16623e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); 16633e59fcb0SMichal Hocko if (page) { 16643e59fcb0SMichal Hocko spin_unlock(&hugetlb_lock); 16653e59fcb0SMichal Hocko return page; 16664db9b2efSMichal Hocko } 16674db9b2efSMichal Hocko } 16684db9b2efSMichal Hocko spin_unlock(&hugetlb_lock); 16694db9b2efSMichal Hocko 16700c397daeSMichal Hocko return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); 16714db9b2efSMichal Hocko } 16724db9b2efSMichal Hocko 1673ebd63723SMichal Hocko /* mempolicy aware migration callback */ 1674389c8178SMichal Hocko struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, 1675389c8178SMichal Hocko unsigned long address) 1676ebd63723SMichal Hocko { 1677ebd63723SMichal Hocko struct mempolicy *mpol; 1678ebd63723SMichal Hocko nodemask_t *nodemask; 1679ebd63723SMichal Hocko struct page *page; 1680ebd63723SMichal Hocko gfp_t gfp_mask; 1681ebd63723SMichal Hocko int node; 1682ebd63723SMichal Hocko 1683ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 1684ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 1685ebd63723SMichal Hocko page = alloc_huge_page_nodemask(h, node, nodemask); 1686ebd63723SMichal Hocko mpol_cond_put(mpol); 1687ebd63723SMichal Hocko 1688ebd63723SMichal Hocko return page; 1689ebd63723SMichal Hocko } 1690ebd63723SMichal Hocko 1691bf50bab2SNaoya Horiguchi /* 169225985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 1693e4e574b7SAdam Litke * of size 'delta'. 1694e4e574b7SAdam Litke */ 1695a5516438SAndi Kleen static int gather_surplus_pages(struct hstate *h, int delta) 1696e4e574b7SAdam Litke { 1697e4e574b7SAdam Litke struct list_head surplus_list; 1698e4e574b7SAdam Litke struct page *page, *tmp; 1699e4e574b7SAdam Litke int ret, i; 1700e4e574b7SAdam Litke int needed, allocated; 170128073b02SHillf Danton bool alloc_ok = true; 1702e4e574b7SAdam Litke 1703a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 1704ac09b3a1SAdam Litke if (needed <= 0) { 1705a5516438SAndi Kleen h->resv_huge_pages += delta; 1706e4e574b7SAdam Litke return 0; 1707ac09b3a1SAdam Litke } 1708e4e574b7SAdam Litke 1709e4e574b7SAdam Litke allocated = 0; 1710e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 1711e4e574b7SAdam Litke 1712e4e574b7SAdam Litke ret = -ENOMEM; 1713e4e574b7SAdam Litke retry: 1714e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 1715e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 17160c397daeSMichal Hocko page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), 1717aaf14e40SMichal Hocko NUMA_NO_NODE, NULL); 171828073b02SHillf Danton if (!page) { 171928073b02SHillf Danton alloc_ok = false; 172028073b02SHillf Danton break; 172128073b02SHillf Danton } 1722e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 172369ed779aSDavid Rientjes cond_resched(); 1724e4e574b7SAdam Litke } 172528073b02SHillf Danton allocated += i; 1726e4e574b7SAdam Litke 1727e4e574b7SAdam Litke /* 1728e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 1729e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 1730e4e574b7SAdam Litke */ 1731e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 1732a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 1733a5516438SAndi Kleen (h->free_huge_pages + allocated); 173428073b02SHillf Danton if (needed > 0) { 173528073b02SHillf Danton if (alloc_ok) 1736e4e574b7SAdam Litke goto retry; 173728073b02SHillf Danton /* 173828073b02SHillf Danton * We were not able to allocate enough pages to 173928073b02SHillf Danton * satisfy the entire reservation so we free what 174028073b02SHillf Danton * we've allocated so far. 174128073b02SHillf Danton */ 174228073b02SHillf Danton goto free; 174328073b02SHillf Danton } 1744e4e574b7SAdam Litke /* 1745e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 174625985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 1747e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 1748ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 1749ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 1750ac09b3a1SAdam Litke * before they are reserved. 1751e4e574b7SAdam Litke */ 1752e4e574b7SAdam Litke needed += allocated; 1753a5516438SAndi Kleen h->resv_huge_pages += delta; 1754e4e574b7SAdam Litke ret = 0; 1755a9869b83SNaoya Horiguchi 175619fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 175719fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 175819fc3f0aSAdam Litke if ((--needed) < 0) 175919fc3f0aSAdam Litke break; 1760a9869b83SNaoya Horiguchi /* 1761a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 1762a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 1763a9869b83SNaoya Horiguchi */ 1764a9869b83SNaoya Horiguchi put_page_testzero(page); 1765309381feSSasha Levin VM_BUG_ON_PAGE(page_count(page), page); 1766a5516438SAndi Kleen enqueue_huge_page(h, page); 176719fc3f0aSAdam Litke } 176828073b02SHillf Danton free: 1769b0365c8dSHillf Danton spin_unlock(&hugetlb_lock); 177019fc3f0aSAdam Litke 177119fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 1772c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 1773a9869b83SNaoya Horiguchi put_page(page); 177419fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 1775e4e574b7SAdam Litke 1776e4e574b7SAdam Litke return ret; 1777e4e574b7SAdam Litke } 1778e4e574b7SAdam Litke 1779e4e574b7SAdam Litke /* 1780e5bbc8a6SMike Kravetz * This routine has two main purposes: 1781e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 1782e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 1783e5bbc8a6SMike Kravetz * to the associated reservation map. 1784e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 1785e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 1786e5bbc8a6SMike Kravetz * 1787e5bbc8a6SMike Kravetz * Called with hugetlb_lock held. However, the lock could be dropped (and 1788e5bbc8a6SMike Kravetz * reacquired) during calls to cond_resched_lock. Whenever dropping the lock, 1789e5bbc8a6SMike Kravetz * we must make sure nobody else can claim pages we are in the process of 1790e5bbc8a6SMike Kravetz * freeing. Do this by ensuring resv_huge_page always is greater than the 1791e5bbc8a6SMike Kravetz * number of huge pages we plan to free when dropping the lock. 1792e4e574b7SAdam Litke */ 1793a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 1794a5516438SAndi Kleen unsigned long unused_resv_pages) 1795e4e574b7SAdam Litke { 1796e4e574b7SAdam Litke unsigned long nr_pages; 1797e4e574b7SAdam Litke 1798aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 1799bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1800e5bbc8a6SMike Kravetz goto out; 1801aa888a74SAndi Kleen 1802e5bbc8a6SMike Kravetz /* 1803e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 1804e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 1805e5bbc8a6SMike Kravetz */ 1806a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 1807e4e574b7SAdam Litke 1808685f3457SLee Schermerhorn /* 1809685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 18109b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 18119b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 18129b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 18139b5e5d0fSLee Schermerhorn * free_pool_huge_page() will balance the the freed pages across the 18149b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 1815e5bbc8a6SMike Kravetz * 1816e5bbc8a6SMike Kravetz * Note that we decrement resv_huge_pages as we free the pages. If 1817e5bbc8a6SMike Kravetz * we drop the lock, resv_huge_pages will still be sufficiently large 1818e5bbc8a6SMike Kravetz * to cover subsequent pages we may free. 1819685f3457SLee Schermerhorn */ 1820685f3457SLee Schermerhorn while (nr_pages--) { 1821e5bbc8a6SMike Kravetz h->resv_huge_pages--; 1822e5bbc8a6SMike Kravetz unused_resv_pages--; 18238cebfcd0SLai Jiangshan if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) 1824e5bbc8a6SMike Kravetz goto out; 18257848a4bfSMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 1826e4e574b7SAdam Litke } 1827e5bbc8a6SMike Kravetz 1828e5bbc8a6SMike Kravetz out: 1829e5bbc8a6SMike Kravetz /* Fully uncommit the reservation */ 1830e5bbc8a6SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 1831e4e574b7SAdam Litke } 1832e4e574b7SAdam Litke 18335e911373SMike Kravetz 1834c37f9fb1SAndy Whitcroft /* 1835feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 18365e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 1837cf3ad20bSMike Kravetz * 1838cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 1839cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 1840cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 1841cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 1842cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 1843feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 1844feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 1845feba16e2SMike Kravetz * is called in such cases. 1846cf3ad20bSMike Kravetz * 1847cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 1848cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 1849cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 1850cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 1851cf3ad20bSMike Kravetz * take appropriate action. 185296b96a96SMike Kravetz * 185396b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 185496b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 185596b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 185696b96a96SMike Kravetz * reservation exists. 1857c37f9fb1SAndy Whitcroft */ 18585e911373SMike Kravetz enum vma_resv_mode { 18595e911373SMike Kravetz VMA_NEEDS_RESV, 18605e911373SMike Kravetz VMA_COMMIT_RESV, 1861feba16e2SMike Kravetz VMA_END_RESV, 186296b96a96SMike Kravetz VMA_ADD_RESV, 18635e911373SMike Kravetz }; 1864cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 1865cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 18665e911373SMike Kravetz enum vma_resv_mode mode) 1867c37f9fb1SAndy Whitcroft { 18684e35f483SJoonsoo Kim struct resv_map *resv; 18694e35f483SJoonsoo Kim pgoff_t idx; 1870cf3ad20bSMike Kravetz long ret; 1871c37f9fb1SAndy Whitcroft 18724e35f483SJoonsoo Kim resv = vma_resv_map(vma); 18734e35f483SJoonsoo Kim if (!resv) 1874c37f9fb1SAndy Whitcroft return 1; 1875c37f9fb1SAndy Whitcroft 18764e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 18775e911373SMike Kravetz switch (mode) { 18785e911373SMike Kravetz case VMA_NEEDS_RESV: 1879cf3ad20bSMike Kravetz ret = region_chg(resv, idx, idx + 1); 18805e911373SMike Kravetz break; 18815e911373SMike Kravetz case VMA_COMMIT_RESV: 18825e911373SMike Kravetz ret = region_add(resv, idx, idx + 1); 18835e911373SMike Kravetz break; 1884feba16e2SMike Kravetz case VMA_END_RESV: 18855e911373SMike Kravetz region_abort(resv, idx, idx + 1); 18865e911373SMike Kravetz ret = 0; 18875e911373SMike Kravetz break; 188896b96a96SMike Kravetz case VMA_ADD_RESV: 188996b96a96SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) 189096b96a96SMike Kravetz ret = region_add(resv, idx, idx + 1); 189196b96a96SMike Kravetz else { 189296b96a96SMike Kravetz region_abort(resv, idx, idx + 1); 189396b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 189496b96a96SMike Kravetz } 189596b96a96SMike Kravetz break; 18965e911373SMike Kravetz default: 18975e911373SMike Kravetz BUG(); 18985e911373SMike Kravetz } 189984afd99bSAndy Whitcroft 19004e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) 1901cf3ad20bSMike Kravetz return ret; 190267961f9dSMike Kravetz else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { 190367961f9dSMike Kravetz /* 190467961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 190567961f9dSMike Kravetz * However, a file associated with mapping could have been 190667961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 190767961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 190867961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 190967961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 191067961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 191167961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 191267961f9dSMike Kravetz * reservation has already been consumed. As a result, the 191367961f9dSMike Kravetz * return value of this routine is the opposite of the 191467961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 191567961f9dSMike Kravetz */ 191667961f9dSMike Kravetz if (ret) 191767961f9dSMike Kravetz return 0; 191867961f9dSMike Kravetz else 191967961f9dSMike Kravetz return 1; 192067961f9dSMike Kravetz } 19214e35f483SJoonsoo Kim else 1922cf3ad20bSMike Kravetz return ret < 0 ? ret : 0; 192384afd99bSAndy Whitcroft } 1924cf3ad20bSMike Kravetz 1925cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 1926a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 1927c37f9fb1SAndy Whitcroft { 19285e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 1929cf3ad20bSMike Kravetz } 1930c37f9fb1SAndy Whitcroft 1931cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 1932cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 1933cf3ad20bSMike Kravetz { 19345e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 19355e911373SMike Kravetz } 19365e911373SMike Kravetz 1937feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 19385e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 19395e911373SMike Kravetz { 1940feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 1941c37f9fb1SAndy Whitcroft } 1942c37f9fb1SAndy Whitcroft 194396b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 194496b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 194596b96a96SMike Kravetz { 194696b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 194796b96a96SMike Kravetz } 194896b96a96SMike Kravetz 194996b96a96SMike Kravetz /* 195096b96a96SMike Kravetz * This routine is called to restore a reservation on error paths. In the 195196b96a96SMike Kravetz * specific error paths, a huge page was allocated (via alloc_huge_page) 195296b96a96SMike Kravetz * and is about to be freed. If a reservation for the page existed, 195396b96a96SMike Kravetz * alloc_huge_page would have consumed the reservation and set PagePrivate 195496b96a96SMike Kravetz * in the newly allocated page. When the page is freed via free_huge_page, 195596b96a96SMike Kravetz * the global reservation count will be incremented if PagePrivate is set. 195696b96a96SMike Kravetz * However, free_huge_page can not adjust the reserve map. Adjust the 195796b96a96SMike Kravetz * reserve map here to be consistent with global reserve count adjustments 195896b96a96SMike Kravetz * to be made by free_huge_page. 195996b96a96SMike Kravetz */ 196096b96a96SMike Kravetz static void restore_reserve_on_error(struct hstate *h, 196196b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long address, 196296b96a96SMike Kravetz struct page *page) 196396b96a96SMike Kravetz { 196496b96a96SMike Kravetz if (unlikely(PagePrivate(page))) { 196596b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 196696b96a96SMike Kravetz 196796b96a96SMike Kravetz if (unlikely(rc < 0)) { 196896b96a96SMike Kravetz /* 196996b96a96SMike Kravetz * Rare out of memory condition in reserve map 197096b96a96SMike Kravetz * manipulation. Clear PagePrivate so that 197196b96a96SMike Kravetz * global reserve count will not be incremented 197296b96a96SMike Kravetz * by free_huge_page. This will make it appear 197396b96a96SMike Kravetz * as though the reservation for this page was 197496b96a96SMike Kravetz * consumed. This may prevent the task from 197596b96a96SMike Kravetz * faulting in the page at a later time. This 197696b96a96SMike Kravetz * is better than inconsistent global huge page 197796b96a96SMike Kravetz * accounting of reserve counts. 197896b96a96SMike Kravetz */ 197996b96a96SMike Kravetz ClearPagePrivate(page); 198096b96a96SMike Kravetz } else if (rc) { 198196b96a96SMike Kravetz rc = vma_add_reservation(h, vma, address); 198296b96a96SMike Kravetz if (unlikely(rc < 0)) 198396b96a96SMike Kravetz /* 198496b96a96SMike Kravetz * See above comment about rare out of 198596b96a96SMike Kravetz * memory condition. 198696b96a96SMike Kravetz */ 198796b96a96SMike Kravetz ClearPagePrivate(page); 198896b96a96SMike Kravetz } else 198996b96a96SMike Kravetz vma_end_reservation(h, vma, address); 199096b96a96SMike Kravetz } 199196b96a96SMike Kravetz } 199296b96a96SMike Kravetz 199370c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 199404f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 1995348ea204SAdam Litke { 199690481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 1997a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 1998348ea204SAdam Litke struct page *page; 1999d85f69b0SMike Kravetz long map_chg, map_commit; 2000d85f69b0SMike Kravetz long gbl_chg; 20016d76dcf4SAneesh Kumar K.V int ret, idx; 20026d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 20032fc39cecSAdam Litke 20046d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 2005a1e78772SMel Gorman /* 2006d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 2007d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 2008d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 2009a1e78772SMel Gorman */ 2010d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 2011d85f69b0SMike Kravetz if (map_chg < 0) 201276dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 2013d85f69b0SMike Kravetz 2014d85f69b0SMike Kravetz /* 2015d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 2016d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 2017d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 2018d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 2019d85f69b0SMike Kravetz * checked against any subpool limit. 2020d85f69b0SMike Kravetz */ 2021d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 2022d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 2023d85f69b0SMike Kravetz if (gbl_chg < 0) { 2024feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 202576dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 20265e911373SMike Kravetz } 202790d8b7e6SAdam Litke 2028d85f69b0SMike Kravetz /* 2029d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 2030d85f69b0SMike Kravetz * map, there could be reservations associated with the 2031d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 2032d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 2033d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 2034d85f69b0SMike Kravetz * the subpool reservations. 2035d85f69b0SMike Kravetz */ 2036d85f69b0SMike Kravetz if (avoid_reserve) 2037d85f69b0SMike Kravetz gbl_chg = 1; 2038d85f69b0SMike Kravetz } 2039d85f69b0SMike Kravetz 20406d76dcf4SAneesh Kumar K.V ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 20418f34af6fSJianyu Zhan if (ret) 20428f34af6fSJianyu Zhan goto out_subpool_put; 20438f34af6fSJianyu Zhan 2044a1e78772SMel Gorman spin_lock(&hugetlb_lock); 2045d85f69b0SMike Kravetz /* 2046d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 2047d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 2048d85f69b0SMike Kravetz * a reservation exists for the allocation. 2049d85f69b0SMike Kravetz */ 2050d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 205181a6fcaeSJoonsoo Kim if (!page) { 205294ae8ba7SAneesh Kumar K.V spin_unlock(&hugetlb_lock); 20530c397daeSMichal Hocko page = alloc_buddy_huge_page_with_mpol(h, vma, addr); 20548f34af6fSJianyu Zhan if (!page) 20558f34af6fSJianyu Zhan goto out_uncharge_cgroup; 2056a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 2057a88c7695SNaoya Horiguchi SetPagePrivate(page); 2058a88c7695SNaoya Horiguchi h->resv_huge_pages--; 2059a88c7695SNaoya Horiguchi } 206079dbb236SAneesh Kumar K.V spin_lock(&hugetlb_lock); 206179dbb236SAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 206281a6fcaeSJoonsoo Kim /* Fall through */ 2063a1e78772SMel Gorman } 206481a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 206581a6fcaeSJoonsoo Kim spin_unlock(&hugetlb_lock); 2066a1e78772SMel Gorman 206790481622SDavid Gibson set_page_private(page, (unsigned long)spool); 2068a1e78772SMel Gorman 2069d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 2070d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 207133039678SMike Kravetz /* 207233039678SMike Kravetz * The page was added to the reservation map between 207333039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 207433039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 207533039678SMike Kravetz * Adjust for the subpool count incremented above AND 207633039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 207733039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 207833039678SMike Kravetz * no longer applies. 207933039678SMike Kravetz */ 208033039678SMike Kravetz long rsv_adjust; 208133039678SMike Kravetz 208233039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 208333039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 208433039678SMike Kravetz } 20857893d1d5SAdam Litke return page; 20868f34af6fSJianyu Zhan 20878f34af6fSJianyu Zhan out_uncharge_cgroup: 20888f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 20898f34af6fSJianyu Zhan out_subpool_put: 2090d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 20918f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 2092feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 20938f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 2094b45b5bd6SDavid Gibson } 2095b45b5bd6SDavid Gibson 2096e24a1307SAneesh Kumar K.V int alloc_bootmem_huge_page(struct hstate *h) 2097e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 2098e24a1307SAneesh Kumar K.V int __alloc_bootmem_huge_page(struct hstate *h) 2099aa888a74SAndi Kleen { 2100aa888a74SAndi Kleen struct huge_bootmem_page *m; 2101b2261026SJoonsoo Kim int nr_nodes, node; 2102aa888a74SAndi Kleen 2103b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 2104aa888a74SAndi Kleen void *addr; 2105aa888a74SAndi Kleen 2106eb31d559SMike Rapoport addr = memblock_alloc_try_nid_raw( 21078b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 210897ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 2109aa888a74SAndi Kleen if (addr) { 2110aa888a74SAndi Kleen /* 2111aa888a74SAndi Kleen * Use the beginning of the huge page to store the 2112aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 2113aa888a74SAndi Kleen * puts them into the mem_map). 2114aa888a74SAndi Kleen */ 2115aa888a74SAndi Kleen m = addr; 2116aa888a74SAndi Kleen goto found; 2117aa888a74SAndi Kleen } 2118aa888a74SAndi Kleen } 2119aa888a74SAndi Kleen return 0; 2120aa888a74SAndi Kleen 2121aa888a74SAndi Kleen found: 2122df994eadSLuiz Capitulino BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); 2123aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 2124330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 2125aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 2126aa888a74SAndi Kleen m->hstate = h; 2127aa888a74SAndi Kleen return 1; 2128aa888a74SAndi Kleen } 2129aa888a74SAndi Kleen 2130d00181b9SKirill A. Shutemov static void __init prep_compound_huge_page(struct page *page, 2131d00181b9SKirill A. Shutemov unsigned int order) 213218229df5SAndy Whitcroft { 213318229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 213418229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 213518229df5SAndy Whitcroft else 213618229df5SAndy Whitcroft prep_compound_page(page, order); 213718229df5SAndy Whitcroft } 213818229df5SAndy Whitcroft 2139aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 2140aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 2141aa888a74SAndi Kleen { 2142aa888a74SAndi Kleen struct huge_bootmem_page *m; 2143aa888a74SAndi Kleen 2144aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 214540d18ebfSMike Kravetz struct page *page = virt_to_page(m); 2146aa888a74SAndi Kleen struct hstate *h = m->hstate; 2147ee8f248dSBecky Bruce 2148aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 214918229df5SAndy Whitcroft prep_compound_huge_page(page, h->order); 2150ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 2151aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 2152af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 2153af0fb9dfSMichal Hocko 2154b0320c7bSRafael Aquini /* 2155b0320c7bSRafael Aquini * If we had gigantic hugepages allocated at boot time, we need 2156b0320c7bSRafael Aquini * to restore the 'stolen' pages to totalram_pages in order to 2157b0320c7bSRafael Aquini * fix confusing memory reports from free(1) and another 2158b0320c7bSRafael Aquini * side-effects, like CommitLimit going negative. 2159b0320c7bSRafael Aquini */ 2160bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 21613dcc0571SJiang Liu adjust_managed_page_count(page, 1 << h->order); 2162520495feSCannon Matthews cond_resched(); 2163aa888a74SAndi Kleen } 2164aa888a74SAndi Kleen } 2165aa888a74SAndi Kleen 21668faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 21671da177e4SLinus Torvalds { 21681da177e4SLinus Torvalds unsigned long i; 21691da177e4SLinus Torvalds 2170e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 2171bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 2172aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 2173aa888a74SAndi Kleen break; 21740c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 21758cebfcd0SLai Jiangshan &node_states[N_MEMORY])) 21761da177e4SLinus Torvalds break; 217769ed779aSDavid Rientjes cond_resched(); 21781da177e4SLinus Torvalds } 2179d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 2180d715cf80SLiam R. Howlett char buf[32]; 2181d715cf80SLiam R. Howlett 2182c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2183d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 2184d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 21858faa8b07SAndi Kleen h->max_huge_pages = i; 2186e5ff2159SAndi Kleen } 2187d715cf80SLiam R. Howlett } 2188e5ff2159SAndi Kleen 2189e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 2190e5ff2159SAndi Kleen { 2191e5ff2159SAndi Kleen struct hstate *h; 2192e5ff2159SAndi Kleen 2193e5ff2159SAndi Kleen for_each_hstate(h) { 2194641844f5SNaoya Horiguchi if (minimum_order > huge_page_order(h)) 2195641844f5SNaoya Horiguchi minimum_order = huge_page_order(h); 2196641844f5SNaoya Horiguchi 21978faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 2198bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 21998faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 2200e5ff2159SAndi Kleen } 2201641844f5SNaoya Horiguchi VM_BUG_ON(minimum_order == UINT_MAX); 2202e5ff2159SAndi Kleen } 2203e5ff2159SAndi Kleen 2204e5ff2159SAndi Kleen static void __init report_hugepages(void) 2205e5ff2159SAndi Kleen { 2206e5ff2159SAndi Kleen struct hstate *h; 2207e5ff2159SAndi Kleen 2208e5ff2159SAndi Kleen for_each_hstate(h) { 22094abd32dbSAndi Kleen char buf[32]; 2210c6247f72SMatthew Wilcox 2211c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2212ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 2213c6247f72SMatthew Wilcox buf, h->free_huge_pages); 2214e5ff2159SAndi Kleen } 2215e5ff2159SAndi Kleen } 2216e5ff2159SAndi Kleen 22171da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 22186ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 22196ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 22201da177e4SLinus Torvalds { 22214415cc8dSChristoph Lameter int i; 22224415cc8dSChristoph Lameter 2223bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2224aa888a74SAndi Kleen return; 2225aa888a74SAndi Kleen 22266ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 22271da177e4SLinus Torvalds struct page *page, *next; 2228a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 2229a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 2230a5516438SAndi Kleen if (count >= h->nr_huge_pages) 22316b0c880dSAdam Litke return; 22321da177e4SLinus Torvalds if (PageHighMem(page)) 22331da177e4SLinus Torvalds continue; 22341da177e4SLinus Torvalds list_del(&page->lru); 2235e5ff2159SAndi Kleen update_and_free_page(h, page); 2236a5516438SAndi Kleen h->free_huge_pages--; 2237a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 22381da177e4SLinus Torvalds } 22391da177e4SLinus Torvalds } 22401da177e4SLinus Torvalds } 22411da177e4SLinus Torvalds #else 22426ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 22436ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 22441da177e4SLinus Torvalds { 22451da177e4SLinus Torvalds } 22461da177e4SLinus Torvalds #endif 22471da177e4SLinus Torvalds 224820a0307cSWu Fengguang /* 224920a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 225020a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 225120a0307cSWu Fengguang * Returns 1 if an adjustment was made. 225220a0307cSWu Fengguang */ 22536ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 22546ae11b27SLee Schermerhorn int delta) 225520a0307cSWu Fengguang { 2256b2261026SJoonsoo Kim int nr_nodes, node; 225720a0307cSWu Fengguang 225820a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 225920a0307cSWu Fengguang 2260e8c5c824SLee Schermerhorn if (delta < 0) { 2261b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 2262b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 2263b2261026SJoonsoo Kim goto found; 2264b2261026SJoonsoo Kim } 2265b2261026SJoonsoo Kim } else { 2266b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2267b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 2268b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 2269b2261026SJoonsoo Kim goto found; 2270e8c5c824SLee Schermerhorn } 22719a76db09SLee Schermerhorn } 2272b2261026SJoonsoo Kim return 0; 227320a0307cSWu Fengguang 2274b2261026SJoonsoo Kim found: 227520a0307cSWu Fengguang h->surplus_huge_pages += delta; 2276b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 2277b2261026SJoonsoo Kim return 1; 227820a0307cSWu Fengguang } 227920a0307cSWu Fengguang 2280a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 22816ae11b27SLee Schermerhorn static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, 22826ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 22831da177e4SLinus Torvalds { 22847893d1d5SAdam Litke unsigned long min_count, ret; 22851da177e4SLinus Torvalds 2286944d9fecSLuiz Capitulino if (hstate_is_gigantic(h) && !gigantic_page_supported()) 2287aa888a74SAndi Kleen return h->max_huge_pages; 2288aa888a74SAndi Kleen 22897893d1d5SAdam Litke /* 22907893d1d5SAdam Litke * Increase the pool size 22917893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 22927893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 2293d1c3fb1fSNishanth Aravamudan * 22940c397daeSMichal Hocko * We might race with alloc_surplus_huge_page() here and be unable 2295d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 2296d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 2297d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 2298d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 22997893d1d5SAdam Litke */ 23001da177e4SLinus Torvalds spin_lock(&hugetlb_lock); 2301a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 23026ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 23037893d1d5SAdam Litke break; 23047893d1d5SAdam Litke } 23057893d1d5SAdam Litke 2306a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 23077893d1d5SAdam Litke /* 23087893d1d5SAdam Litke * If this allocation races such that we no longer need the 23097893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 23107893d1d5SAdam Litke * and reducing the surplus. 23117893d1d5SAdam Litke */ 23127893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 2313649920c6SJia He 2314649920c6SJia He /* yield cpu to avoid soft lockup */ 2315649920c6SJia He cond_resched(); 2316649920c6SJia He 23170c397daeSMichal Hocko ret = alloc_pool_huge_page(h, nodes_allowed); 23187893d1d5SAdam Litke spin_lock(&hugetlb_lock); 23197893d1d5SAdam Litke if (!ret) 23207893d1d5SAdam Litke goto out; 23217893d1d5SAdam Litke 2322536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 2323536240f2SMel Gorman if (signal_pending(current)) 2324536240f2SMel Gorman goto out; 23257893d1d5SAdam Litke } 23267893d1d5SAdam Litke 23277893d1d5SAdam Litke /* 23287893d1d5SAdam Litke * Decrease the pool size 23297893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 23307893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 23317893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 23327893d1d5SAdam Litke * to the desired size as pages become free. 2333d1c3fb1fSNishanth Aravamudan * 2334d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 2335d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 2336d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 23370c397daeSMichal Hocko * alloc_surplus_huge_page() is checking the global counter, 2338d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 2339d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 2340d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 23417893d1d5SAdam Litke */ 2342a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 23436b0c880dSAdam Litke min_count = max(count, min_count); 23446ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 2345a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 23466ae11b27SLee Schermerhorn if (!free_pool_huge_page(h, nodes_allowed, 0)) 23471da177e4SLinus Torvalds break; 234855f67141SMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 23491da177e4SLinus Torvalds } 2350a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 23516ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 23527893d1d5SAdam Litke break; 23537893d1d5SAdam Litke } 23547893d1d5SAdam Litke out: 2355a5516438SAndi Kleen ret = persistent_huge_pages(h); 23561da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 23577893d1d5SAdam Litke return ret; 23581da177e4SLinus Torvalds } 23591da177e4SLinus Torvalds 2360a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 2361a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 2362a3437870SNishanth Aravamudan 2363a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 2364a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 2365a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 2366a3437870SNishanth Aravamudan 2367a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 2368a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 2369a3437870SNishanth Aravamudan 23709a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 23719a305230SLee Schermerhorn 23729a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 2373a3437870SNishanth Aravamudan { 2374a3437870SNishanth Aravamudan int i; 23759a305230SLee Schermerhorn 2376a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 23779a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 23789a305230SLee Schermerhorn if (nidp) 23799a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 2380a3437870SNishanth Aravamudan return &hstates[i]; 23819a305230SLee Schermerhorn } 23829a305230SLee Schermerhorn 23839a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 2384a3437870SNishanth Aravamudan } 2385a3437870SNishanth Aravamudan 238606808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 2387a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2388a3437870SNishanth Aravamudan { 23899a305230SLee Schermerhorn struct hstate *h; 23909a305230SLee Schermerhorn unsigned long nr_huge_pages; 23919a305230SLee Schermerhorn int nid; 23929a305230SLee Schermerhorn 23939a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 23949a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 23959a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 23969a305230SLee Schermerhorn else 23979a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 23989a305230SLee Schermerhorn 23999a305230SLee Schermerhorn return sprintf(buf, "%lu\n", nr_huge_pages); 2400a3437870SNishanth Aravamudan } 2401adbe8726SEric B Munson 2402238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 2403238d3c13SDavid Rientjes struct hstate *h, int nid, 2404238d3c13SDavid Rientjes unsigned long count, size_t len) 2405a3437870SNishanth Aravamudan { 2406a3437870SNishanth Aravamudan int err; 2407bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); 2408a3437870SNishanth Aravamudan 2409944d9fecSLuiz Capitulino if (hstate_is_gigantic(h) && !gigantic_page_supported()) { 2410adbe8726SEric B Munson err = -EINVAL; 2411adbe8726SEric B Munson goto out; 2412adbe8726SEric B Munson } 2413adbe8726SEric B Munson 24149a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 24159a305230SLee Schermerhorn /* 24169a305230SLee Schermerhorn * global hstate attribute 24179a305230SLee Schermerhorn */ 24189a305230SLee Schermerhorn if (!(obey_mempolicy && 24199a305230SLee Schermerhorn init_nodemask_of_mempolicy(nodes_allowed))) { 242006808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 24218cebfcd0SLai Jiangshan nodes_allowed = &node_states[N_MEMORY]; 242206808b08SLee Schermerhorn } 24239a305230SLee Schermerhorn } else if (nodes_allowed) { 24249a305230SLee Schermerhorn /* 24259a305230SLee Schermerhorn * per node hstate attribute: adjust count to global, 24269a305230SLee Schermerhorn * but restrict alloc/free to the specified node. 24279a305230SLee Schermerhorn */ 24289a305230SLee Schermerhorn count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 24299a305230SLee Schermerhorn init_nodemask_of_node(nodes_allowed, nid); 24309a305230SLee Schermerhorn } else 24318cebfcd0SLai Jiangshan nodes_allowed = &node_states[N_MEMORY]; 24329a305230SLee Schermerhorn 243306808b08SLee Schermerhorn h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); 2434a3437870SNishanth Aravamudan 24358cebfcd0SLai Jiangshan if (nodes_allowed != &node_states[N_MEMORY]) 243606808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 243706808b08SLee Schermerhorn 243806808b08SLee Schermerhorn return len; 2439adbe8726SEric B Munson out: 2440adbe8726SEric B Munson NODEMASK_FREE(nodes_allowed); 2441adbe8726SEric B Munson return err; 244206808b08SLee Schermerhorn } 244306808b08SLee Schermerhorn 2444238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 2445238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 2446238d3c13SDavid Rientjes size_t len) 2447238d3c13SDavid Rientjes { 2448238d3c13SDavid Rientjes struct hstate *h; 2449238d3c13SDavid Rientjes unsigned long count; 2450238d3c13SDavid Rientjes int nid; 2451238d3c13SDavid Rientjes int err; 2452238d3c13SDavid Rientjes 2453238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 2454238d3c13SDavid Rientjes if (err) 2455238d3c13SDavid Rientjes return err; 2456238d3c13SDavid Rientjes 2457238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 2458238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 2459238d3c13SDavid Rientjes } 2460238d3c13SDavid Rientjes 246106808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 246206808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 246306808b08SLee Schermerhorn { 246406808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 246506808b08SLee Schermerhorn } 246606808b08SLee Schermerhorn 246706808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 246806808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 246906808b08SLee Schermerhorn { 2470238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 2471a3437870SNishanth Aravamudan } 2472a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 2473a3437870SNishanth Aravamudan 247406808b08SLee Schermerhorn #ifdef CONFIG_NUMA 247506808b08SLee Schermerhorn 247606808b08SLee Schermerhorn /* 247706808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 247806808b08SLee Schermerhorn * huge page alloc/free. 247906808b08SLee Schermerhorn */ 248006808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 248106808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 248206808b08SLee Schermerhorn { 248306808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 248406808b08SLee Schermerhorn } 248506808b08SLee Schermerhorn 248606808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 248706808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 248806808b08SLee Schermerhorn { 2489238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 249006808b08SLee Schermerhorn } 249106808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 249206808b08SLee Schermerhorn #endif 249306808b08SLee Schermerhorn 249406808b08SLee Schermerhorn 2495a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 2496a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2497a3437870SNishanth Aravamudan { 24989a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2499a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 2500a3437870SNishanth Aravamudan } 2501adbe8726SEric B Munson 2502a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 2503a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 2504a3437870SNishanth Aravamudan { 2505a3437870SNishanth Aravamudan int err; 2506a3437870SNishanth Aravamudan unsigned long input; 25079a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2508a3437870SNishanth Aravamudan 2509bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2510adbe8726SEric B Munson return -EINVAL; 2511adbe8726SEric B Munson 25123dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 2513a3437870SNishanth Aravamudan if (err) 251473ae31e5SEric B Munson return err; 2515a3437870SNishanth Aravamudan 2516a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 2517a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 2518a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 2519a3437870SNishanth Aravamudan 2520a3437870SNishanth Aravamudan return count; 2521a3437870SNishanth Aravamudan } 2522a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 2523a3437870SNishanth Aravamudan 2524a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 2525a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2526a3437870SNishanth Aravamudan { 25279a305230SLee Schermerhorn struct hstate *h; 25289a305230SLee Schermerhorn unsigned long free_huge_pages; 25299a305230SLee Schermerhorn int nid; 25309a305230SLee Schermerhorn 25319a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 25329a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 25339a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 25349a305230SLee Schermerhorn else 25359a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 25369a305230SLee Schermerhorn 25379a305230SLee Schermerhorn return sprintf(buf, "%lu\n", free_huge_pages); 2538a3437870SNishanth Aravamudan } 2539a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 2540a3437870SNishanth Aravamudan 2541a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 2542a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2543a3437870SNishanth Aravamudan { 25449a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2545a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->resv_huge_pages); 2546a3437870SNishanth Aravamudan } 2547a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 2548a3437870SNishanth Aravamudan 2549a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 2550a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2551a3437870SNishanth Aravamudan { 25529a305230SLee Schermerhorn struct hstate *h; 25539a305230SLee Schermerhorn unsigned long surplus_huge_pages; 25549a305230SLee Schermerhorn int nid; 25559a305230SLee Schermerhorn 25569a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 25579a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 25589a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 25599a305230SLee Schermerhorn else 25609a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 25619a305230SLee Schermerhorn 25629a305230SLee Schermerhorn return sprintf(buf, "%lu\n", surplus_huge_pages); 2563a3437870SNishanth Aravamudan } 2564a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 2565a3437870SNishanth Aravamudan 2566a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 2567a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 2568a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 2569a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 2570a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 2571a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 257206808b08SLee Schermerhorn #ifdef CONFIG_NUMA 257306808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 257406808b08SLee Schermerhorn #endif 2575a3437870SNishanth Aravamudan NULL, 2576a3437870SNishanth Aravamudan }; 2577a3437870SNishanth Aravamudan 257867e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 2579a3437870SNishanth Aravamudan .attrs = hstate_attrs, 2580a3437870SNishanth Aravamudan }; 2581a3437870SNishanth Aravamudan 2582094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 25839a305230SLee Schermerhorn struct kobject **hstate_kobjs, 258467e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 2585a3437870SNishanth Aravamudan { 2586a3437870SNishanth Aravamudan int retval; 2587972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 2588a3437870SNishanth Aravamudan 25899a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 25909a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 2591a3437870SNishanth Aravamudan return -ENOMEM; 2592a3437870SNishanth Aravamudan 25939a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 2594a3437870SNishanth Aravamudan if (retval) 25959a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 2596a3437870SNishanth Aravamudan 2597a3437870SNishanth Aravamudan return retval; 2598a3437870SNishanth Aravamudan } 2599a3437870SNishanth Aravamudan 2600a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 2601a3437870SNishanth Aravamudan { 2602a3437870SNishanth Aravamudan struct hstate *h; 2603a3437870SNishanth Aravamudan int err; 2604a3437870SNishanth Aravamudan 2605a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 2606a3437870SNishanth Aravamudan if (!hugepages_kobj) 2607a3437870SNishanth Aravamudan return; 2608a3437870SNishanth Aravamudan 2609a3437870SNishanth Aravamudan for_each_hstate(h) { 26109a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 26119a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 2612a3437870SNishanth Aravamudan if (err) 2613ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s", h->name); 2614a3437870SNishanth Aravamudan } 2615a3437870SNishanth Aravamudan } 2616a3437870SNishanth Aravamudan 26179a305230SLee Schermerhorn #ifdef CONFIG_NUMA 26189a305230SLee Schermerhorn 26199a305230SLee Schermerhorn /* 26209a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 262110fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 262210fbcf4cSKay Sievers * index of a node device or _hstate == node id. 262310fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 26249a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 26259a305230SLee Schermerhorn */ 26269a305230SLee Schermerhorn struct node_hstate { 26279a305230SLee Schermerhorn struct kobject *hugepages_kobj; 26289a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 26299a305230SLee Schermerhorn }; 2630b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 26319a305230SLee Schermerhorn 26329a305230SLee Schermerhorn /* 263310fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 26349a305230SLee Schermerhorn */ 26359a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 26369a305230SLee Schermerhorn &nr_hugepages_attr.attr, 26379a305230SLee Schermerhorn &free_hugepages_attr.attr, 26389a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 26399a305230SLee Schermerhorn NULL, 26409a305230SLee Schermerhorn }; 26419a305230SLee Schermerhorn 264267e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 26439a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 26449a305230SLee Schermerhorn }; 26459a305230SLee Schermerhorn 26469a305230SLee Schermerhorn /* 264710fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 26489a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 26499a305230SLee Schermerhorn */ 26509a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 26519a305230SLee Schermerhorn { 26529a305230SLee Schermerhorn int nid; 26539a305230SLee Schermerhorn 26549a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 26559a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 26569a305230SLee Schermerhorn int i; 26579a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 26589a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 26599a305230SLee Schermerhorn if (nidp) 26609a305230SLee Schermerhorn *nidp = nid; 26619a305230SLee Schermerhorn return &hstates[i]; 26629a305230SLee Schermerhorn } 26639a305230SLee Schermerhorn } 26649a305230SLee Schermerhorn 26659a305230SLee Schermerhorn BUG(); 26669a305230SLee Schermerhorn return NULL; 26679a305230SLee Schermerhorn } 26689a305230SLee Schermerhorn 26699a305230SLee Schermerhorn /* 267010fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 26719a305230SLee Schermerhorn * No-op if no hstate attributes attached. 26729a305230SLee Schermerhorn */ 26733cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 26749a305230SLee Schermerhorn { 26759a305230SLee Schermerhorn struct hstate *h; 267610fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 26779a305230SLee Schermerhorn 26789a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 26799b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 26809a305230SLee Schermerhorn 2681972dc4deSAneesh Kumar K.V for_each_hstate(h) { 2682972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 2683972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 2684972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 2685972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 2686972dc4deSAneesh Kumar K.V } 26879a305230SLee Schermerhorn } 26889a305230SLee Schermerhorn 26899a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 26909a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 26919a305230SLee Schermerhorn } 26929a305230SLee Schermerhorn 26939a305230SLee Schermerhorn 26949a305230SLee Schermerhorn /* 269510fbcf4cSKay Sievers * Register hstate attributes for a single node device. 26969a305230SLee Schermerhorn * No-op if attributes already registered. 26979a305230SLee Schermerhorn */ 26983cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 26999a305230SLee Schermerhorn { 27009a305230SLee Schermerhorn struct hstate *h; 270110fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 27029a305230SLee Schermerhorn int err; 27039a305230SLee Schermerhorn 27049a305230SLee Schermerhorn if (nhs->hugepages_kobj) 27059a305230SLee Schermerhorn return; /* already allocated */ 27069a305230SLee Schermerhorn 27079a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 270810fbcf4cSKay Sievers &node->dev.kobj); 27099a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 27109a305230SLee Schermerhorn return; 27119a305230SLee Schermerhorn 27129a305230SLee Schermerhorn for_each_hstate(h) { 27139a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 27149a305230SLee Schermerhorn nhs->hstate_kobjs, 27159a305230SLee Schermerhorn &per_node_hstate_attr_group); 27169a305230SLee Schermerhorn if (err) { 2717ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s for node %d\n", 271810fbcf4cSKay Sievers h->name, node->dev.id); 27199a305230SLee Schermerhorn hugetlb_unregister_node(node); 27209a305230SLee Schermerhorn break; 27219a305230SLee Schermerhorn } 27229a305230SLee Schermerhorn } 27239a305230SLee Schermerhorn } 27249a305230SLee Schermerhorn 27259a305230SLee Schermerhorn /* 27269b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 272710fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 272810fbcf4cSKay Sievers * registered their associated device by this time. 27299a305230SLee Schermerhorn */ 27307d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 27319a305230SLee Schermerhorn { 27329a305230SLee Schermerhorn int nid; 27339a305230SLee Schermerhorn 27348cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 27358732794bSWen Congyang struct node *node = node_devices[nid]; 273610fbcf4cSKay Sievers if (node->dev.id == nid) 27379a305230SLee Schermerhorn hugetlb_register_node(node); 27389a305230SLee Schermerhorn } 27399a305230SLee Schermerhorn 27409a305230SLee Schermerhorn /* 274110fbcf4cSKay Sievers * Let the node device driver know we're here so it can 27429a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 27439a305230SLee Schermerhorn */ 27449a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 27459a305230SLee Schermerhorn hugetlb_unregister_node); 27469a305230SLee Schermerhorn } 27479a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 27489a305230SLee Schermerhorn 27499a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 27509a305230SLee Schermerhorn { 27519a305230SLee Schermerhorn BUG(); 27529a305230SLee Schermerhorn if (nidp) 27539a305230SLee Schermerhorn *nidp = -1; 27549a305230SLee Schermerhorn return NULL; 27559a305230SLee Schermerhorn } 27569a305230SLee Schermerhorn 27579a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 27589a305230SLee Schermerhorn 27599a305230SLee Schermerhorn #endif 27609a305230SLee Schermerhorn 2761a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 2762a3437870SNishanth Aravamudan { 27638382d914SDavidlohr Bueso int i; 27648382d914SDavidlohr Bueso 2765457c1b27SNishanth Aravamudan if (!hugepages_supported()) 27660ef89d25SBenjamin Herrenschmidt return 0; 2767a3437870SNishanth Aravamudan 2768e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) { 2769d715cf80SLiam R. Howlett if (default_hstate_size != 0) { 2770d715cf80SLiam R. Howlett pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n", 2771d715cf80SLiam R. Howlett default_hstate_size, HPAGE_SIZE); 2772d715cf80SLiam R. Howlett } 2773d715cf80SLiam R. Howlett 2774e11bfbfcSNick Piggin default_hstate_size = HPAGE_SIZE; 2775e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) 2776a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 2777a3437870SNishanth Aravamudan } 2778972dc4deSAneesh Kumar K.V default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); 2779f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 2780f8b74815SVaishali Thakkar if (!default_hstate.max_huge_pages) 2781e11bfbfcSNick Piggin default_hstate.max_huge_pages = default_hstate_max_huge_pages; 2782f8b74815SVaishali Thakkar } 2783a3437870SNishanth Aravamudan 2784a3437870SNishanth Aravamudan hugetlb_init_hstates(); 2785aa888a74SAndi Kleen gather_bootmem_prealloc(); 2786a3437870SNishanth Aravamudan report_hugepages(); 2787a3437870SNishanth Aravamudan 2788a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 27899a305230SLee Schermerhorn hugetlb_register_all_nodes(); 27907179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 27919a305230SLee Schermerhorn 27928382d914SDavidlohr Bueso #ifdef CONFIG_SMP 27938382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 27948382d914SDavidlohr Bueso #else 27958382d914SDavidlohr Bueso num_fault_mutexes = 1; 27968382d914SDavidlohr Bueso #endif 2797c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 27986da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 27996da2ec56SKees Cook 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 314705ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 314805ea8860SDan Williams { 314905ea8860SDan Williams struct hstate *hstate = hstate_vma(vma); 315005ea8860SDan Williams 315105ea8860SDan Williams return 1UL << huge_page_shift(hstate); 315205ea8860SDan Williams } 315305ea8860SDan Williams 31541da177e4SLinus Torvalds /* 31551da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 31561da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 31571da177e4SLinus Torvalds * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 31581da177e4SLinus Torvalds * this far. 31591da177e4SLinus Torvalds */ 3160b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 31611da177e4SLinus Torvalds { 31621da177e4SLinus Torvalds BUG(); 3163d0217ac0SNick Piggin return 0; 31641da177e4SLinus Torvalds } 31651da177e4SLinus Torvalds 3166eec3636aSJane Chu /* 3167eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 3168eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 3169eec3636aSJane Chu * This is because under System V memory model, mappings created via 3170eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 3171eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 3172eec3636aSJane Chu */ 3173f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 3174d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 317584afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 3176a1e78772SMel Gorman .close = hugetlb_vm_op_close, 317731383c68SDan Williams .split = hugetlb_vm_op_split, 317805ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 31791da177e4SLinus Torvalds }; 31801da177e4SLinus Torvalds 31811e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 31821e8f889bSDavid Gibson int writable) 318363551ae0SDavid Gibson { 318463551ae0SDavid Gibson pte_t entry; 318563551ae0SDavid Gibson 31861e8f889bSDavid Gibson if (writable) { 3187106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 3188106c992aSGerald Schaefer vma->vm_page_prot))); 318963551ae0SDavid Gibson } else { 3190106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 3191106c992aSGerald Schaefer vma->vm_page_prot)); 319263551ae0SDavid Gibson } 319363551ae0SDavid Gibson entry = pte_mkyoung(entry); 319463551ae0SDavid Gibson entry = pte_mkhuge(entry); 3195d9ed9faaSChris Metcalf entry = arch_make_huge_pte(entry, vma, page, writable); 319663551ae0SDavid Gibson 319763551ae0SDavid Gibson return entry; 319863551ae0SDavid Gibson } 319963551ae0SDavid Gibson 32001e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 32011e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 32021e8f889bSDavid Gibson { 32031e8f889bSDavid Gibson pte_t entry; 32041e8f889bSDavid Gibson 3205106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 320632f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 32074b3073e1SRussell King update_mmu_cache(vma, address, ptep); 32081e8f889bSDavid Gibson } 32091e8f889bSDavid Gibson 3210d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 32114a705fefSNaoya Horiguchi { 32124a705fefSNaoya Horiguchi swp_entry_t swp; 32134a705fefSNaoya Horiguchi 32144a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 3215d5ed7444SAneesh Kumar K.V return false; 32164a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 32174a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_migration_entry(swp)) 3218d5ed7444SAneesh Kumar K.V return true; 32194a705fefSNaoya Horiguchi else 3220d5ed7444SAneesh Kumar K.V return false; 32214a705fefSNaoya Horiguchi } 32224a705fefSNaoya Horiguchi 32234a705fefSNaoya Horiguchi static int is_hugetlb_entry_hwpoisoned(pte_t pte) 32244a705fefSNaoya Horiguchi { 32254a705fefSNaoya Horiguchi swp_entry_t swp; 32264a705fefSNaoya Horiguchi 32274a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 32284a705fefSNaoya Horiguchi return 0; 32294a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 32304a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_hwpoison_entry(swp)) 32314a705fefSNaoya Horiguchi return 1; 32324a705fefSNaoya Horiguchi else 32334a705fefSNaoya Horiguchi return 0; 32344a705fefSNaoya Horiguchi } 32351e8f889bSDavid Gibson 323663551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 323763551ae0SDavid Gibson struct vm_area_struct *vma) 323863551ae0SDavid Gibson { 32395e41540cSMike Kravetz pte_t *src_pte, *dst_pte, entry, dst_entry; 324063551ae0SDavid Gibson struct page *ptepage; 32411c59827dSHugh Dickins unsigned long addr; 32421e8f889bSDavid Gibson int cow; 3243a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3244a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 3245ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 3246e8569dd2SAndreas Sandberg int ret = 0; 32471e8f889bSDavid Gibson 32481e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 324963551ae0SDavid Gibson 3250ac46d4f3SJérôme Glisse if (cow) { 3251ac46d4f3SJérôme Glisse mmu_notifier_range_init(&range, src, vma->vm_start, 3252ac46d4f3SJérôme Glisse vma->vm_end); 3253ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 3254ac46d4f3SJérôme Glisse } 3255e8569dd2SAndreas Sandberg 3256a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 3257cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 32587868a208SPunit Agrawal src_pte = huge_pte_offset(src, addr, sz); 3259c74df32cSHugh Dickins if (!src_pte) 3260c74df32cSHugh Dickins continue; 3261a5516438SAndi Kleen dst_pte = huge_pte_alloc(dst, addr, sz); 3262e8569dd2SAndreas Sandberg if (!dst_pte) { 3263e8569dd2SAndreas Sandberg ret = -ENOMEM; 3264e8569dd2SAndreas Sandberg break; 3265e8569dd2SAndreas Sandberg } 3266c5c99429SLarry Woodman 32675e41540cSMike Kravetz /* 32685e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 32695e41540cSMike Kravetz * dst_pte == src_pte is the common case of src/dest sharing. 32705e41540cSMike Kravetz * 32715e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 32725e41540cSMike Kravetz * another vma. If dst_pte !none, this implies sharing. 32735e41540cSMike Kravetz * Check here before taking page table lock, and once again 32745e41540cSMike Kravetz * after taking the lock below. 32755e41540cSMike Kravetz */ 32765e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 32775e41540cSMike Kravetz if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) 3278c5c99429SLarry Woodman continue; 3279c5c99429SLarry Woodman 3280cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 3281cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 3282cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 32834a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 32845e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 32855e41540cSMike Kravetz if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { 32865e41540cSMike Kravetz /* 32875e41540cSMike Kravetz * Skip if src entry none. Also, skip in the 32885e41540cSMike Kravetz * unlikely case dst entry !none as this implies 32895e41540cSMike Kravetz * sharing with another vma. 32905e41540cSMike Kravetz */ 32914a705fefSNaoya Horiguchi ; 32924a705fefSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry) || 32934a705fefSNaoya Horiguchi is_hugetlb_entry_hwpoisoned(entry))) { 32944a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 32954a705fefSNaoya Horiguchi 32964a705fefSNaoya Horiguchi if (is_write_migration_entry(swp_entry) && cow) { 32974a705fefSNaoya Horiguchi /* 32984a705fefSNaoya Horiguchi * COW mappings require pages in both 32994a705fefSNaoya Horiguchi * parent and child to be set to read. 33004a705fefSNaoya Horiguchi */ 33014a705fefSNaoya Horiguchi make_migration_entry_read(&swp_entry); 33024a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 3303e5251fd4SPunit Agrawal set_huge_swap_pte_at(src, addr, src_pte, 3304e5251fd4SPunit Agrawal entry, sz); 33054a705fefSNaoya Horiguchi } 3306e5251fd4SPunit Agrawal set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); 33074a705fefSNaoya Horiguchi } else { 330834ee645eSJoerg Roedel if (cow) { 33090f10851eSJérôme Glisse /* 33100f10851eSJérôme Glisse * No need to notify as we are downgrading page 33110f10851eSJérôme Glisse * table protection not changing it to point 33120f10851eSJérôme Glisse * to a new page. 33130f10851eSJérôme Glisse * 3314ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 33150f10851eSJérôme Glisse */ 33167f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 331734ee645eSJoerg Roedel } 33180253d634SNaoya Horiguchi entry = huge_ptep_get(src_pte); 331963551ae0SDavid Gibson ptepage = pte_page(entry); 332063551ae0SDavid Gibson get_page(ptepage); 332153f9263bSKirill A. Shutemov page_dup_rmap(ptepage, true); 332263551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 33235d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), dst); 33241c59827dSHugh Dickins } 3325cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 3326cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 332763551ae0SDavid Gibson } 332863551ae0SDavid Gibson 3329e8569dd2SAndreas Sandberg if (cow) 3330ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 3331e8569dd2SAndreas Sandberg 3332e8569dd2SAndreas Sandberg return ret; 333363551ae0SDavid Gibson } 333463551ae0SDavid Gibson 333524669e58SAneesh Kumar K.V void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 333624669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 333724669e58SAneesh Kumar K.V struct page *ref_page) 333863551ae0SDavid Gibson { 333963551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 334063551ae0SDavid Gibson unsigned long address; 3341c7546f8fSDavid Gibson pte_t *ptep; 334263551ae0SDavid Gibson pte_t pte; 3343cb900f41SKirill A. Shutemov spinlock_t *ptl; 334463551ae0SDavid Gibson struct page *page; 3345a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3346a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 3347ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 3348a5516438SAndi Kleen 334963551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 3350a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 3351a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 335263551ae0SDavid Gibson 335307e32661SAneesh Kumar K.V /* 335407e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 335507e32661SAneesh Kumar K.V * to huge page. 335607e32661SAneesh Kumar K.V */ 3357ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 335824669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 3359dff11abeSMike Kravetz 3360dff11abeSMike Kravetz /* 3361dff11abeSMike Kravetz * If sharing possible, alert mmu notifiers of worst case. 3362dff11abeSMike Kravetz */ 3363ac46d4f3SJérôme Glisse mmu_notifier_range_init(&range, mm, start, end); 3364ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 3365ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 3366569f48b8SHillf Danton address = start; 3367569f48b8SHillf Danton for (; address < end; address += sz) { 33687868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, sz); 3369c7546f8fSDavid Gibson if (!ptep) 3370c7546f8fSDavid Gibson continue; 3371c7546f8fSDavid Gibson 3372cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 337331d49da5SAneesh Kumar K.V if (huge_pmd_unshare(mm, &address, ptep)) { 337431d49da5SAneesh Kumar K.V spin_unlock(ptl); 3375dff11abeSMike Kravetz /* 3376dff11abeSMike Kravetz * We just unmapped a page of PMDs by clearing a PUD. 3377dff11abeSMike Kravetz * The caller's TLB flush range should cover this area. 3378dff11abeSMike Kravetz */ 337931d49da5SAneesh Kumar K.V continue; 338031d49da5SAneesh Kumar K.V } 338139dde65cSChen, Kenneth W 33826629326bSHillf Danton pte = huge_ptep_get(ptep); 338331d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 338431d49da5SAneesh Kumar K.V spin_unlock(ptl); 338531d49da5SAneesh Kumar K.V continue; 338631d49da5SAneesh Kumar K.V } 33876629326bSHillf Danton 33886629326bSHillf Danton /* 33899fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 33909fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 33916629326bSHillf Danton */ 33929fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 33939386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 339431d49da5SAneesh Kumar K.V spin_unlock(ptl); 339531d49da5SAneesh Kumar K.V continue; 33968c4894c6SNaoya Horiguchi } 33976629326bSHillf Danton 33986629326bSHillf Danton page = pte_page(pte); 339904f2cbe3SMel Gorman /* 340004f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 340104f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 340204f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 340304f2cbe3SMel Gorman */ 340404f2cbe3SMel Gorman if (ref_page) { 340531d49da5SAneesh Kumar K.V if (page != ref_page) { 340631d49da5SAneesh Kumar K.V spin_unlock(ptl); 340731d49da5SAneesh Kumar K.V continue; 340831d49da5SAneesh Kumar K.V } 340904f2cbe3SMel Gorman /* 341004f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 341104f2cbe3SMel Gorman * future faults in this VMA will fail rather than 341204f2cbe3SMel Gorman * looking like data was lost 341304f2cbe3SMel Gorman */ 341404f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 341504f2cbe3SMel Gorman } 341604f2cbe3SMel Gorman 3417c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 3418b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 3419106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 34206649a386SKen Chen set_page_dirty(page); 34219e81130bSHillf Danton 34225d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 3423d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 342431d49da5SAneesh Kumar K.V 3425cb900f41SKirill A. Shutemov spin_unlock(ptl); 3426e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 342724669e58SAneesh Kumar K.V /* 342831d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 342924669e58SAneesh Kumar K.V */ 343031d49da5SAneesh Kumar K.V if (ref_page) 343131d49da5SAneesh Kumar K.V break; 3432fe1668aeSChen, Kenneth W } 3433ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 343424669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 34351da177e4SLinus Torvalds } 343663551ae0SDavid Gibson 3437d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 3438d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 3439d833352aSMel Gorman unsigned long end, struct page *ref_page) 3440d833352aSMel Gorman { 3441d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 3442d833352aSMel Gorman 3443d833352aSMel Gorman /* 3444d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 3445d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 3446d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 3447d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 3448d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 3449c8c06efaSDavidlohr Bueso * is to clear it before releasing the i_mmap_rwsem. This works 3450d833352aSMel Gorman * because in the context this is called, the VMA is about to be 3451c8c06efaSDavidlohr Bueso * destroyed and the i_mmap_rwsem is held. 3452d833352aSMel Gorman */ 3453d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 3454d833352aSMel Gorman } 3455d833352aSMel Gorman 3456502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 345704f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 3458502717f4SChen, Kenneth W { 345924669e58SAneesh Kumar K.V struct mm_struct *mm; 346024669e58SAneesh Kumar K.V struct mmu_gather tlb; 3461dff11abeSMike Kravetz unsigned long tlb_start = start; 3462dff11abeSMike Kravetz unsigned long tlb_end = end; 3463dff11abeSMike Kravetz 3464dff11abeSMike Kravetz /* 3465dff11abeSMike Kravetz * If shared PMDs were possibly used within this vma range, adjust 3466dff11abeSMike Kravetz * start/end for worst case tlb flushing. 3467dff11abeSMike Kravetz * Note that we can not be sure if PMDs are shared until we try to 3468dff11abeSMike Kravetz * unmap pages. However, we want to make sure TLB flushing covers 3469dff11abeSMike Kravetz * the largest possible range. 3470dff11abeSMike Kravetz */ 3471dff11abeSMike Kravetz adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end); 347224669e58SAneesh Kumar K.V 347324669e58SAneesh Kumar K.V mm = vma->vm_mm; 347424669e58SAneesh Kumar K.V 3475dff11abeSMike Kravetz tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end); 347624669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 3477dff11abeSMike Kravetz tlb_finish_mmu(&tlb, tlb_start, tlb_end); 3478502717f4SChen, Kenneth W } 3479502717f4SChen, Kenneth W 348004f2cbe3SMel Gorman /* 348104f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 348204f2cbe3SMel Gorman * mappping it owns the reserve page for. The intention is to unmap the page 348304f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 348404f2cbe3SMel Gorman * same region. 348504f2cbe3SMel Gorman */ 34862f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 34872a4b3dedSHarvey Harrison struct page *page, unsigned long address) 348804f2cbe3SMel Gorman { 34897526674dSAdam Litke struct hstate *h = hstate_vma(vma); 349004f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 349104f2cbe3SMel Gorman struct address_space *mapping; 349204f2cbe3SMel Gorman pgoff_t pgoff; 349304f2cbe3SMel Gorman 349404f2cbe3SMel Gorman /* 349504f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 349604f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 349704f2cbe3SMel Gorman */ 34987526674dSAdam Litke address = address & huge_page_mask(h); 349936e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 350036e4f20aSMichal Hocko vma->vm_pgoff; 350193c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 350204f2cbe3SMel Gorman 35034eb2b1dcSMel Gorman /* 35044eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 35054eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 35064eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 35074eb2b1dcSMel Gorman */ 350883cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 35096b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 351004f2cbe3SMel Gorman /* Do not unmap the current VMA */ 351104f2cbe3SMel Gorman if (iter_vma == vma) 351204f2cbe3SMel Gorman continue; 351304f2cbe3SMel Gorman 351404f2cbe3SMel Gorman /* 35152f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 35162f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 35172f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 35182f84a899SMel Gorman */ 35192f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 35202f84a899SMel Gorman continue; 35212f84a899SMel Gorman 35222f84a899SMel Gorman /* 352304f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 352404f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 352504f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 352604f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 352704f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 352804f2cbe3SMel Gorman */ 352904f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 353024669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 353124669e58SAneesh Kumar K.V address + huge_page_size(h), page); 353204f2cbe3SMel Gorman } 353383cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 353404f2cbe3SMel Gorman } 353504f2cbe3SMel Gorman 35360fe6e20bSNaoya Horiguchi /* 35370fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 3538ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 3539ef009b25SMichal Hocko * cannot race with other handlers or page migration. 3540ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 35410fe6e20bSNaoya Horiguchi */ 35422b740303SSouptick Joarder static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 3543974e6d66SHuang Ying unsigned long address, pte_t *ptep, 3544cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 35451e8f889bSDavid Gibson { 35463999f52eSAneesh Kumar K.V pte_t pte; 3547a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 35481e8f889bSDavid Gibson struct page *old_page, *new_page; 35492b740303SSouptick Joarder int outside_reserve = 0; 35502b740303SSouptick Joarder vm_fault_t ret = 0; 3551974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 3552ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 35531e8f889bSDavid Gibson 35543999f52eSAneesh Kumar K.V pte = huge_ptep_get(ptep); 35551e8f889bSDavid Gibson old_page = pte_page(pte); 35561e8f889bSDavid Gibson 355704f2cbe3SMel Gorman retry_avoidcopy: 35581e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 35591e8f889bSDavid Gibson * and just make the page writable */ 356037a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 35615a49973dSHugh Dickins page_move_anon_rmap(old_page, vma); 35625b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 356383c54070SNick Piggin return 0; 35641e8f889bSDavid Gibson } 35651e8f889bSDavid Gibson 356604f2cbe3SMel Gorman /* 356704f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 356804f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 356904f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 357004f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 357104f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 357204f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 357304f2cbe3SMel Gorman * of the full address range. 357404f2cbe3SMel Gorman */ 35755944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 357604f2cbe3SMel Gorman old_page != pagecache_page) 357704f2cbe3SMel Gorman outside_reserve = 1; 357804f2cbe3SMel Gorman 357909cbfeafSKirill A. Shutemov get_page(old_page); 3580b76c8cfbSLarry Woodman 3581ad4404a2SDavidlohr Bueso /* 3582ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 3583ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 3584ad4404a2SDavidlohr Bueso */ 3585cb900f41SKirill A. Shutemov spin_unlock(ptl); 35865b7a1d40SHuang Ying new_page = alloc_huge_page(vma, haddr, outside_reserve); 35871e8f889bSDavid Gibson 35882fc39cecSAdam Litke if (IS_ERR(new_page)) { 358904f2cbe3SMel Gorman /* 359004f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 359104f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 359204f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 359304f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 359404f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 359504f2cbe3SMel Gorman */ 359604f2cbe3SMel Gorman if (outside_reserve) { 359709cbfeafSKirill A. Shutemov put_page(old_page); 359804f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 35995b7a1d40SHuang Ying unmap_ref_private(mm, vma, old_page, haddr); 360004f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 3601cb900f41SKirill A. Shutemov spin_lock(ptl); 36025b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 3603a9af0c5dSNaoya Horiguchi if (likely(ptep && 3604a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 360504f2cbe3SMel Gorman goto retry_avoidcopy; 3606a734bcc8SHillf Danton /* 3607cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 3608cb900f41SKirill A. Shutemov * lock, and our job is done. 3609a734bcc8SHillf Danton */ 3610a734bcc8SHillf Danton return 0; 361104f2cbe3SMel Gorman } 361204f2cbe3SMel Gorman 36132b740303SSouptick Joarder ret = vmf_error(PTR_ERR(new_page)); 3614ad4404a2SDavidlohr Bueso goto out_release_old; 36151e8f889bSDavid Gibson } 36161e8f889bSDavid Gibson 36170fe6e20bSNaoya Horiguchi /* 36180fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 36190fe6e20bSNaoya Horiguchi * anon_vma prepared. 36200fe6e20bSNaoya Horiguchi */ 362144e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 3622ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 3623ad4404a2SDavidlohr Bueso goto out_release_all; 362444e2aa93SDean Nelson } 36250fe6e20bSNaoya Horiguchi 3626974e6d66SHuang Ying copy_user_huge_page(new_page, old_page, address, vma, 362747ad8475SAndrea Arcangeli pages_per_huge_page(h)); 36280ed361deSNick Piggin __SetPageUptodate(new_page); 36291e8f889bSDavid Gibson 3630ac46d4f3SJérôme Glisse mmu_notifier_range_init(&range, mm, haddr, haddr + huge_page_size(h)); 3631ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 3632ad4404a2SDavidlohr Bueso 3633b76c8cfbSLarry Woodman /* 3634cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 3635b76c8cfbSLarry Woodman * before the page tables are altered 3636b76c8cfbSLarry Woodman */ 3637cb900f41SKirill A. Shutemov spin_lock(ptl); 36385b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 3639a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 364007443a85SJoonsoo Kim ClearPagePrivate(new_page); 364107443a85SJoonsoo Kim 36421e8f889bSDavid Gibson /* Break COW */ 36435b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 3644ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, range.end); 36455b7a1d40SHuang Ying set_huge_pte_at(mm, haddr, ptep, 36461e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 3647d281ee61SKirill A. Shutemov page_remove_rmap(old_page, true); 36485b7a1d40SHuang Ying hugepage_add_new_anon_rmap(new_page, vma, haddr); 3649cb6acd01SMike Kravetz set_page_huge_active(new_page); 36501e8f889bSDavid Gibson /* Make the old page be freed below */ 36511e8f889bSDavid Gibson new_page = old_page; 36521e8f889bSDavid Gibson } 3653cb900f41SKirill A. Shutemov spin_unlock(ptl); 3654ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 3655ad4404a2SDavidlohr Bueso out_release_all: 36565b7a1d40SHuang Ying restore_reserve_on_error(h, vma, haddr, new_page); 365709cbfeafSKirill A. Shutemov put_page(new_page); 3658ad4404a2SDavidlohr Bueso out_release_old: 365909cbfeafSKirill A. Shutemov put_page(old_page); 36608312034fSJoonsoo Kim 3661ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 3662ad4404a2SDavidlohr Bueso return ret; 36631e8f889bSDavid Gibson } 36641e8f889bSDavid Gibson 366504f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 3666a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 3667a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 366804f2cbe3SMel Gorman { 366904f2cbe3SMel Gorman struct address_space *mapping; 3670e7c4b0bfSAndy Whitcroft pgoff_t idx; 367104f2cbe3SMel Gorman 367204f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 3673a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 367404f2cbe3SMel Gorman 367504f2cbe3SMel Gorman return find_lock_page(mapping, idx); 367604f2cbe3SMel Gorman } 367704f2cbe3SMel Gorman 36783ae77f43SHugh Dickins /* 36793ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 36803ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 36813ae77f43SHugh Dickins */ 36823ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 36832a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 36842a15efc9SHugh Dickins { 36852a15efc9SHugh Dickins struct address_space *mapping; 36862a15efc9SHugh Dickins pgoff_t idx; 36872a15efc9SHugh Dickins struct page *page; 36882a15efc9SHugh Dickins 36892a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 36902a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 36912a15efc9SHugh Dickins 36922a15efc9SHugh Dickins page = find_get_page(mapping, idx); 36932a15efc9SHugh Dickins if (page) 36942a15efc9SHugh Dickins put_page(page); 36952a15efc9SHugh Dickins return page != NULL; 36962a15efc9SHugh Dickins } 36972a15efc9SHugh Dickins 3698ab76ad54SMike Kravetz int huge_add_to_page_cache(struct page *page, struct address_space *mapping, 3699ab76ad54SMike Kravetz pgoff_t idx) 3700ab76ad54SMike Kravetz { 3701ab76ad54SMike Kravetz struct inode *inode = mapping->host; 3702ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 3703ab76ad54SMike Kravetz int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 3704ab76ad54SMike Kravetz 3705ab76ad54SMike Kravetz if (err) 3706ab76ad54SMike Kravetz return err; 3707ab76ad54SMike Kravetz ClearPagePrivate(page); 3708ab76ad54SMike Kravetz 370922146c3cSMike Kravetz /* 371022146c3cSMike Kravetz * set page dirty so that it will not be removed from cache/file 371122146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 371222146c3cSMike Kravetz */ 371322146c3cSMike Kravetz set_page_dirty(page); 371422146c3cSMike Kravetz 3715ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 3716ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 3717ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 3718ab76ad54SMike Kravetz return 0; 3719ab76ad54SMike Kravetz } 3720ab76ad54SMike Kravetz 37212b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 37222b740303SSouptick Joarder struct vm_area_struct *vma, 37238382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 3724788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 3725ac9b9c66SHugh Dickins { 3726a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 37272b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 3728409eb8c2SHillf Danton int anon_rmap = 0; 37294c887265SAdam Litke unsigned long size; 37304c887265SAdam Litke struct page *page; 37311e8f889bSDavid Gibson pte_t new_pte; 3732cb900f41SKirill A. Shutemov spinlock_t *ptl; 3733285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 3734cb6acd01SMike Kravetz bool new_page = false; 37354c887265SAdam Litke 373604f2cbe3SMel Gorman /* 373704f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 373804f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 373925985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 374004f2cbe3SMel Gorman */ 374104f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 3742910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 374304f2cbe3SMel Gorman current->pid); 374404f2cbe3SMel Gorman return ret; 374504f2cbe3SMel Gorman } 374604f2cbe3SMel Gorman 37474c887265SAdam Litke /* 3748e7c58097SMike Kravetz * Use page lock to guard against racing truncation 3749e7c58097SMike Kravetz * before we get page_table_lock. 37504c887265SAdam Litke */ 3751e7c58097SMike Kravetz retry: 3752e7c58097SMike Kravetz page = find_lock_page(mapping, idx); 3753e7c58097SMike Kravetz if (!page) { 3754a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 3755ebed4bfcSHugh Dickins if (idx >= size) 3756ebed4bfcSHugh Dickins goto out; 37571a1aad8aSMike Kravetz 37581a1aad8aSMike Kravetz /* 37591a1aad8aSMike Kravetz * Check for page in userfault range 37601a1aad8aSMike Kravetz */ 37611a1aad8aSMike Kravetz if (userfaultfd_missing(vma)) { 37621a1aad8aSMike Kravetz u32 hash; 37631a1aad8aSMike Kravetz struct vm_fault vmf = { 37641a1aad8aSMike Kravetz .vma = vma, 3765285b8dcaSHuang Ying .address = haddr, 37661a1aad8aSMike Kravetz .flags = flags, 37671a1aad8aSMike Kravetz /* 37681a1aad8aSMike Kravetz * Hard to debug if it ends up being 37691a1aad8aSMike Kravetz * used by a callee that assumes 37701a1aad8aSMike Kravetz * something about the other 37711a1aad8aSMike Kravetz * uninitialized fields... same as in 37721a1aad8aSMike Kravetz * memory.c 37731a1aad8aSMike Kravetz */ 37741a1aad8aSMike Kravetz }; 37751a1aad8aSMike Kravetz 37761a1aad8aSMike Kravetz /* 3777ddeaab32SMike Kravetz * hugetlb_fault_mutex must be dropped before 3778ddeaab32SMike Kravetz * handling userfault. Reacquire after handling 3779ddeaab32SMike Kravetz * fault to make calling code simpler. 37801a1aad8aSMike Kravetz */ 37811a1aad8aSMike Kravetz hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, 3782285b8dcaSHuang Ying idx, haddr); 37831a1aad8aSMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 37841a1aad8aSMike Kravetz ret = handle_userfault(&vmf, VM_UFFD_MISSING); 37851a1aad8aSMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 37861a1aad8aSMike Kravetz goto out; 37871a1aad8aSMike Kravetz } 37881a1aad8aSMike Kravetz 3789285b8dcaSHuang Ying page = alloc_huge_page(vma, haddr, 0); 37902fc39cecSAdam Litke if (IS_ERR(page)) { 37912b740303SSouptick Joarder ret = vmf_error(PTR_ERR(page)); 37926bda666aSChristoph Lameter goto out; 37936bda666aSChristoph Lameter } 379447ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 37950ed361deSNick Piggin __SetPageUptodate(page); 3796cb6acd01SMike Kravetz new_page = true; 3797ac9b9c66SHugh Dickins 3798f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 3799ab76ad54SMike Kravetz int err = huge_add_to_page_cache(page, mapping, idx); 38006bda666aSChristoph Lameter if (err) { 38016bda666aSChristoph Lameter put_page(page); 38026bda666aSChristoph Lameter if (err == -EEXIST) 38036bda666aSChristoph Lameter goto retry; 38046bda666aSChristoph Lameter goto out; 38056bda666aSChristoph Lameter } 380623be7468SMel Gorman } else { 38076bda666aSChristoph Lameter lock_page(page); 38080fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 38090fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 38100fe6e20bSNaoya Horiguchi goto backout_unlocked; 381123be7468SMel Gorman } 3812409eb8c2SHillf Danton anon_rmap = 1; 38130fe6e20bSNaoya Horiguchi } 38140fe6e20bSNaoya Horiguchi } else { 381557303d80SAndy Whitcroft /* 3816998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 3817998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 3818998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 3819fd6a03edSNaoya Horiguchi */ 3820fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 3821aa50d3a7SAndi Kleen ret = VM_FAULT_HWPOISON | 3822972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 3823fd6a03edSNaoya Horiguchi goto backout_unlocked; 38246bda666aSChristoph Lameter } 3825998b4382SNaoya Horiguchi } 38261e8f889bSDavid Gibson 382757303d80SAndy Whitcroft /* 382857303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 382957303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 383057303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 383157303d80SAndy Whitcroft * the spinlock. 383257303d80SAndy Whitcroft */ 38335e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 3834285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 38352b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 38362b26736cSAndy Whitcroft goto backout_unlocked; 38372b26736cSAndy Whitcroft } 38385e911373SMike Kravetz /* Just decrements count, does not deallocate */ 3839285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 38405e911373SMike Kravetz } 384157303d80SAndy Whitcroft 38428bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 3843e7c58097SMike Kravetz size = i_size_read(mapping->host) >> huge_page_shift(h); 3844e7c58097SMike Kravetz if (idx >= size) 3845e7c58097SMike Kravetz goto backout; 38464c887265SAdam Litke 384783c54070SNick Piggin ret = 0; 38487f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 38494c887265SAdam Litke goto backout; 38504c887265SAdam Litke 385107443a85SJoonsoo Kim if (anon_rmap) { 385207443a85SJoonsoo Kim ClearPagePrivate(page); 3853285b8dcaSHuang Ying hugepage_add_new_anon_rmap(page, vma, haddr); 3854ac714904SChoi Gi-yong } else 385553f9263bSKirill A. Shutemov page_dup_rmap(page, true); 38561e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 38571e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 3858285b8dcaSHuang Ying set_huge_pte_at(mm, haddr, ptep, new_pte); 38591e8f889bSDavid Gibson 38605d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 3861788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 38621e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 3863974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); 38641e8f889bSDavid Gibson } 38651e8f889bSDavid Gibson 3866cb900f41SKirill A. Shutemov spin_unlock(ptl); 3867cb6acd01SMike Kravetz 3868cb6acd01SMike Kravetz /* 3869cb6acd01SMike Kravetz * Only make newly allocated pages active. Existing pages found 3870cb6acd01SMike Kravetz * in the pagecache could be !page_huge_active() if they have been 3871cb6acd01SMike Kravetz * isolated for migration. 3872cb6acd01SMike Kravetz */ 3873cb6acd01SMike Kravetz if (new_page) 3874cb6acd01SMike Kravetz set_page_huge_active(page); 3875cb6acd01SMike Kravetz 38764c887265SAdam Litke unlock_page(page); 38774c887265SAdam Litke out: 3878ac9b9c66SHugh Dickins return ret; 38794c887265SAdam Litke 38804c887265SAdam Litke backout: 3881cb900f41SKirill A. Shutemov spin_unlock(ptl); 38822b26736cSAndy Whitcroft backout_unlocked: 38834c887265SAdam Litke unlock_page(page); 3884285b8dcaSHuang Ying restore_reserve_on_error(h, vma, haddr, page); 38854c887265SAdam Litke put_page(page); 38864c887265SAdam Litke goto out; 3887ac9b9c66SHugh Dickins } 3888ac9b9c66SHugh Dickins 38898382d914SDavidlohr Bueso #ifdef CONFIG_SMP 3890c672c7f2SMike Kravetz u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, 38918382d914SDavidlohr Bueso struct vm_area_struct *vma, 38928382d914SDavidlohr Bueso struct address_space *mapping, 38938382d914SDavidlohr Bueso pgoff_t idx, unsigned long address) 38948382d914SDavidlohr Bueso { 38958382d914SDavidlohr Bueso unsigned long key[2]; 38968382d914SDavidlohr Bueso u32 hash; 38978382d914SDavidlohr Bueso 38988382d914SDavidlohr Bueso if (vma->vm_flags & VM_SHARED) { 38998382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 39008382d914SDavidlohr Bueso key[1] = idx; 39018382d914SDavidlohr Bueso } else { 39028382d914SDavidlohr Bueso key[0] = (unsigned long) mm; 39038382d914SDavidlohr Bueso key[1] = address >> huge_page_shift(h); 39048382d914SDavidlohr Bueso } 39058382d914SDavidlohr Bueso 39068382d914SDavidlohr Bueso hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0); 39078382d914SDavidlohr Bueso 39088382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 39098382d914SDavidlohr Bueso } 39108382d914SDavidlohr Bueso #else 39118382d914SDavidlohr Bueso /* 39128382d914SDavidlohr Bueso * For uniprocesor systems we always use a single mutex, so just 39138382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 39148382d914SDavidlohr Bueso */ 3915c672c7f2SMike Kravetz u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, 39168382d914SDavidlohr Bueso struct vm_area_struct *vma, 39178382d914SDavidlohr Bueso struct address_space *mapping, 39188382d914SDavidlohr Bueso pgoff_t idx, unsigned long address) 39198382d914SDavidlohr Bueso { 39208382d914SDavidlohr Bueso return 0; 39218382d914SDavidlohr Bueso } 39228382d914SDavidlohr Bueso #endif 39238382d914SDavidlohr Bueso 39242b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 3925788c7df4SHugh Dickins unsigned long address, unsigned int flags) 392686e5216fSAdam Litke { 39278382d914SDavidlohr Bueso pte_t *ptep, entry; 3928cb900f41SKirill A. Shutemov spinlock_t *ptl; 39292b740303SSouptick Joarder vm_fault_t ret; 39308382d914SDavidlohr Bueso u32 hash; 39318382d914SDavidlohr Bueso pgoff_t idx; 39320fe6e20bSNaoya Horiguchi struct page *page = NULL; 393357303d80SAndy Whitcroft struct page *pagecache_page = NULL; 3934a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 39358382d914SDavidlohr Bueso struct address_space *mapping; 39360f792cf9SNaoya Horiguchi int need_wait_lock = 0; 3937285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 393886e5216fSAdam Litke 3939285b8dcaSHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 3940fd6a03edSNaoya Horiguchi if (ptep) { 3941fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 3942290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 3943cb900f41SKirill A. Shutemov migration_entry_wait_huge(vma, mm, ptep); 3944290408d4SNaoya Horiguchi return 0; 3945290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 3946aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 3947972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 3948ddeaab32SMike Kravetz } else { 3949b43a9990SMike Kravetz ptep = huge_pte_alloc(mm, haddr, huge_page_size(h)); 3950ddeaab32SMike Kravetz if (!ptep) 3951b43a9990SMike Kravetz return VM_FAULT_OOM; 3952b43a9990SMike Kravetz } 39538382d914SDavidlohr Bueso 3954ddeaab32SMike Kravetz mapping = vma->vm_file->f_mapping; 3955ddeaab32SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 3956ddeaab32SMike Kravetz 39573935baa9SDavid Gibson /* 39583935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 39593935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 39603935baa9SDavid Gibson * the same page in the page cache. 39613935baa9SDavid Gibson */ 3962285b8dcaSHuang Ying hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, haddr); 3963c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 39648382d914SDavidlohr Bueso 39657f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 39667f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 39678382d914SDavidlohr Bueso ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 3968b4d1d99fSDavid Gibson goto out_mutex; 39693935baa9SDavid Gibson } 397086e5216fSAdam Litke 397183c54070SNick Piggin ret = 0; 39721e8f889bSDavid Gibson 397357303d80SAndy Whitcroft /* 39740f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 39750f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 39760f792cf9SNaoya Horiguchi * a active hugepage in pagecache. This goto expects the 2nd page fault, 39770f792cf9SNaoya Horiguchi * and is_hugetlb_entry_(migration|hwpoisoned) check will properly 39780f792cf9SNaoya Horiguchi * handle it. 39790f792cf9SNaoya Horiguchi */ 39800f792cf9SNaoya Horiguchi if (!pte_present(entry)) 39810f792cf9SNaoya Horiguchi goto out_mutex; 39820f792cf9SNaoya Horiguchi 39830f792cf9SNaoya Horiguchi /* 398457303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 398557303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 398657303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 398757303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 398857303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 398957303d80SAndy Whitcroft * consumed. 399057303d80SAndy Whitcroft */ 3991106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 3992285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 39932b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 3994b4d1d99fSDavid Gibson goto out_mutex; 39952b26736cSAndy Whitcroft } 39965e911373SMike Kravetz /* Just decrements count, does not deallocate */ 3997285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 399857303d80SAndy Whitcroft 3999f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 400057303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 4001285b8dcaSHuang Ying vma, haddr); 400257303d80SAndy Whitcroft } 400357303d80SAndy Whitcroft 40040f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 40050fe6e20bSNaoya Horiguchi 40061e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 4007b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 4008cb900f41SKirill A. Shutemov goto out_ptl; 4009b4d1d99fSDavid Gibson 40100f792cf9SNaoya Horiguchi /* 40110f792cf9SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 40120f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 40130f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 40140f792cf9SNaoya Horiguchi */ 40150f792cf9SNaoya Horiguchi page = pte_page(entry); 40160f792cf9SNaoya Horiguchi if (page != pagecache_page) 40170f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 40180f792cf9SNaoya Horiguchi need_wait_lock = 1; 40190f792cf9SNaoya Horiguchi goto out_ptl; 40200f792cf9SNaoya Horiguchi } 40210f792cf9SNaoya Horiguchi 40220f792cf9SNaoya Horiguchi get_page(page); 4023b4d1d99fSDavid Gibson 4024788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 4025106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 4026974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, 4027cb900f41SKirill A. Shutemov pagecache_page, ptl); 40280f792cf9SNaoya Horiguchi goto out_put_page; 4029b4d1d99fSDavid Gibson } 4030106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 4031b4d1d99fSDavid Gibson } 4032b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 4033285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 4034788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 4035285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 40360f792cf9SNaoya Horiguchi out_put_page: 40370f792cf9SNaoya Horiguchi if (page != pagecache_page) 40380f792cf9SNaoya Horiguchi unlock_page(page); 40390f792cf9SNaoya Horiguchi put_page(page); 4040cb900f41SKirill A. Shutemov out_ptl: 4041cb900f41SKirill A. Shutemov spin_unlock(ptl); 404257303d80SAndy Whitcroft 404357303d80SAndy Whitcroft if (pagecache_page) { 404457303d80SAndy Whitcroft unlock_page(pagecache_page); 404557303d80SAndy Whitcroft put_page(pagecache_page); 404657303d80SAndy Whitcroft } 4047b4d1d99fSDavid Gibson out_mutex: 4048c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 40490f792cf9SNaoya Horiguchi /* 40500f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 40510f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 40520f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 40530f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 40540f792cf9SNaoya Horiguchi * here without taking refcount. 40550f792cf9SNaoya Horiguchi */ 40560f792cf9SNaoya Horiguchi if (need_wait_lock) 40570f792cf9SNaoya Horiguchi wait_on_page_locked(page); 40581e8f889bSDavid Gibson return ret; 405986e5216fSAdam Litke } 406086e5216fSAdam Litke 40618fb5debcSMike Kravetz /* 40628fb5debcSMike Kravetz * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with 40638fb5debcSMike Kravetz * modifications for huge pages. 40648fb5debcSMike Kravetz */ 40658fb5debcSMike Kravetz int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, 40668fb5debcSMike Kravetz pte_t *dst_pte, 40678fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 40688fb5debcSMike Kravetz unsigned long dst_addr, 40698fb5debcSMike Kravetz unsigned long src_addr, 40708fb5debcSMike Kravetz struct page **pagep) 40718fb5debcSMike Kravetz { 40721e392147SAndrea Arcangeli struct address_space *mapping; 40731e392147SAndrea Arcangeli pgoff_t idx; 40741e392147SAndrea Arcangeli unsigned long size; 40751c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 40768fb5debcSMike Kravetz struct hstate *h = hstate_vma(dst_vma); 40778fb5debcSMike Kravetz pte_t _dst_pte; 40788fb5debcSMike Kravetz spinlock_t *ptl; 40798fb5debcSMike Kravetz int ret; 40808fb5debcSMike Kravetz struct page *page; 40818fb5debcSMike Kravetz 40828fb5debcSMike Kravetz if (!*pagep) { 40838fb5debcSMike Kravetz ret = -ENOMEM; 40848fb5debcSMike Kravetz page = alloc_huge_page(dst_vma, dst_addr, 0); 40858fb5debcSMike Kravetz if (IS_ERR(page)) 40868fb5debcSMike Kravetz goto out; 40878fb5debcSMike Kravetz 40888fb5debcSMike Kravetz ret = copy_huge_page_from_user(page, 40898fb5debcSMike Kravetz (const void __user *) src_addr, 4090810a56b9SMike Kravetz pages_per_huge_page(h), false); 40918fb5debcSMike Kravetz 40928fb5debcSMike Kravetz /* fallback to copy_from_user outside mmap_sem */ 40938fb5debcSMike Kravetz if (unlikely(ret)) { 40949e368259SAndrea Arcangeli ret = -ENOENT; 40958fb5debcSMike Kravetz *pagep = page; 40968fb5debcSMike Kravetz /* don't free the page */ 40978fb5debcSMike Kravetz goto out; 40988fb5debcSMike Kravetz } 40998fb5debcSMike Kravetz } else { 41008fb5debcSMike Kravetz page = *pagep; 41018fb5debcSMike Kravetz *pagep = NULL; 41028fb5debcSMike Kravetz } 41038fb5debcSMike Kravetz 41048fb5debcSMike Kravetz /* 41058fb5debcSMike Kravetz * The memory barrier inside __SetPageUptodate makes sure that 41068fb5debcSMike Kravetz * preceding stores to the page contents become visible before 41078fb5debcSMike Kravetz * the set_pte_at() write. 41088fb5debcSMike Kravetz */ 41098fb5debcSMike Kravetz __SetPageUptodate(page); 41108fb5debcSMike Kravetz 41111e392147SAndrea Arcangeli mapping = dst_vma->vm_file->f_mapping; 41121e392147SAndrea Arcangeli idx = vma_hugecache_offset(h, dst_vma, dst_addr); 41131e392147SAndrea Arcangeli 41141c9e8defSMike Kravetz /* 41151c9e8defSMike Kravetz * If shared, add to page cache 41161c9e8defSMike Kravetz */ 41171c9e8defSMike Kravetz if (vm_shared) { 41181e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 41191e392147SAndrea Arcangeli ret = -EFAULT; 41201e392147SAndrea Arcangeli if (idx >= size) 41211e392147SAndrea Arcangeli goto out_release_nounlock; 41221c9e8defSMike Kravetz 41231e392147SAndrea Arcangeli /* 41241e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 41251e392147SAndrea Arcangeli * huge_add_to_page_cache() below happens through the 41261e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 41271e392147SAndrea Arcangeli * the caller. 41281e392147SAndrea Arcangeli */ 41291c9e8defSMike Kravetz ret = huge_add_to_page_cache(page, mapping, idx); 41301c9e8defSMike Kravetz if (ret) 41311c9e8defSMike Kravetz goto out_release_nounlock; 41321c9e8defSMike Kravetz } 41331c9e8defSMike Kravetz 41348fb5debcSMike Kravetz ptl = huge_pte_lockptr(h, dst_mm, dst_pte); 41358fb5debcSMike Kravetz spin_lock(ptl); 41368fb5debcSMike Kravetz 41371e392147SAndrea Arcangeli /* 41381e392147SAndrea Arcangeli * Recheck the i_size after holding PT lock to make sure not 41391e392147SAndrea Arcangeli * to leave any page mapped (as page_mapped()) beyond the end 41401e392147SAndrea Arcangeli * of the i_size (remove_inode_hugepages() is strict about 41411e392147SAndrea Arcangeli * enforcing that). If we bail out here, we'll also leave a 41421e392147SAndrea Arcangeli * page in the radix tree in the vm_shared case beyond the end 41431e392147SAndrea Arcangeli * of the i_size, but remove_inode_hugepages() will take care 41441e392147SAndrea Arcangeli * of it as soon as we drop the hugetlb_fault_mutex_table. 41451e392147SAndrea Arcangeli */ 41461e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 41471e392147SAndrea Arcangeli ret = -EFAULT; 41481e392147SAndrea Arcangeli if (idx >= size) 41491e392147SAndrea Arcangeli goto out_release_unlock; 41501e392147SAndrea Arcangeli 41518fb5debcSMike Kravetz ret = -EEXIST; 41528fb5debcSMike Kravetz if (!huge_pte_none(huge_ptep_get(dst_pte))) 41538fb5debcSMike Kravetz goto out_release_unlock; 41548fb5debcSMike Kravetz 41551c9e8defSMike Kravetz if (vm_shared) { 41561c9e8defSMike Kravetz page_dup_rmap(page, true); 41571c9e8defSMike Kravetz } else { 41588fb5debcSMike Kravetz ClearPagePrivate(page); 41598fb5debcSMike Kravetz hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); 41601c9e8defSMike Kravetz } 41618fb5debcSMike Kravetz 41628fb5debcSMike Kravetz _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); 41638fb5debcSMike Kravetz if (dst_vma->vm_flags & VM_WRITE) 41648fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 41658fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 41668fb5debcSMike Kravetz 41678fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 41688fb5debcSMike Kravetz 41698fb5debcSMike Kravetz (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, 41708fb5debcSMike Kravetz dst_vma->vm_flags & VM_WRITE); 41718fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 41728fb5debcSMike Kravetz 41738fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 41748fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 41758fb5debcSMike Kravetz 41768fb5debcSMike Kravetz spin_unlock(ptl); 4177cb6acd01SMike Kravetz set_page_huge_active(page); 41781c9e8defSMike Kravetz if (vm_shared) 41791c9e8defSMike Kravetz unlock_page(page); 41808fb5debcSMike Kravetz ret = 0; 41818fb5debcSMike Kravetz out: 41828fb5debcSMike Kravetz return ret; 41838fb5debcSMike Kravetz out_release_unlock: 41848fb5debcSMike Kravetz spin_unlock(ptl); 41851c9e8defSMike Kravetz if (vm_shared) 41861c9e8defSMike Kravetz unlock_page(page); 41875af10dfdSAndrea Arcangeli out_release_nounlock: 41888fb5debcSMike Kravetz put_page(page); 41898fb5debcSMike Kravetz goto out; 41908fb5debcSMike Kravetz } 41918fb5debcSMike Kravetz 419228a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 419363551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 419428a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 419587ffc118SAndrea Arcangeli long i, unsigned int flags, int *nonblocking) 419663551ae0SDavid Gibson { 4197d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 4198d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 419928a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 4200a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 42012be7cfedSDaniel Jordan int err = -EFAULT; 420263551ae0SDavid Gibson 420363551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 420463551ae0SDavid Gibson pte_t *pte; 4205cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 42062a15efc9SHugh Dickins int absent; 420763551ae0SDavid Gibson struct page *page; 420863551ae0SDavid Gibson 42094c887265SAdam Litke /* 421002057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 421102057967SDavid Rientjes * potentially allocating memory. 421202057967SDavid Rientjes */ 4213fa45f116SDavidlohr Bueso if (fatal_signal_pending(current)) { 421402057967SDavid Rientjes remainder = 0; 421502057967SDavid Rientjes break; 421602057967SDavid Rientjes } 421702057967SDavid Rientjes 421802057967SDavid Rientjes /* 42194c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 42202a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 42214c887265SAdam Litke * first, for the page indexing below to work. 4222cb900f41SKirill A. Shutemov * 4223cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 42244c887265SAdam Litke */ 42257868a208SPunit Agrawal pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), 42267868a208SPunit Agrawal huge_page_size(h)); 4227cb900f41SKirill A. Shutemov if (pte) 4228cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 42292a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 423063551ae0SDavid Gibson 42312a15efc9SHugh Dickins /* 42322a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 42333ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 42343ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 42353ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 42363ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 42372a15efc9SHugh Dickins */ 42383ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 42393ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 4240cb900f41SKirill A. Shutemov if (pte) 4241cb900f41SKirill A. Shutemov spin_unlock(ptl); 42422a15efc9SHugh Dickins remainder = 0; 42432a15efc9SHugh Dickins break; 42442a15efc9SHugh Dickins } 42452a15efc9SHugh Dickins 42469cc3a5bdSNaoya Horiguchi /* 42479cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 42489cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 42499cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 42509cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 42519cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 42529cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 42539cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 42549cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 42559cc3a5bdSNaoya Horiguchi */ 42569cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 4257106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 4258106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 42592b740303SSouptick Joarder vm_fault_t ret; 426087ffc118SAndrea Arcangeli unsigned int fault_flags = 0; 42614c887265SAdam Litke 4262cb900f41SKirill A. Shutemov if (pte) 4263cb900f41SKirill A. Shutemov spin_unlock(ptl); 426487ffc118SAndrea Arcangeli if (flags & FOLL_WRITE) 426587ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_WRITE; 426687ffc118SAndrea Arcangeli if (nonblocking) 426787ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY; 426887ffc118SAndrea Arcangeli if (flags & FOLL_NOWAIT) 426987ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY | 427087ffc118SAndrea Arcangeli FAULT_FLAG_RETRY_NOWAIT; 427187ffc118SAndrea Arcangeli if (flags & FOLL_TRIED) { 427287ffc118SAndrea Arcangeli VM_WARN_ON_ONCE(fault_flags & 427387ffc118SAndrea Arcangeli FAULT_FLAG_ALLOW_RETRY); 427487ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_TRIED; 427587ffc118SAndrea Arcangeli } 427687ffc118SAndrea Arcangeli ret = hugetlb_fault(mm, vma, vaddr, fault_flags); 427787ffc118SAndrea Arcangeli if (ret & VM_FAULT_ERROR) { 42782be7cfedSDaniel Jordan err = vm_fault_to_errno(ret, flags); 42791c59827dSHugh Dickins remainder = 0; 42801c59827dSHugh Dickins break; 42811c59827dSHugh Dickins } 428287ffc118SAndrea Arcangeli if (ret & VM_FAULT_RETRY) { 42831ac25013SAndrea Arcangeli if (nonblocking && 42841ac25013SAndrea Arcangeli !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) 428587ffc118SAndrea Arcangeli *nonblocking = 0; 428687ffc118SAndrea Arcangeli *nr_pages = 0; 428787ffc118SAndrea Arcangeli /* 428887ffc118SAndrea Arcangeli * VM_FAULT_RETRY must not return an 428987ffc118SAndrea Arcangeli * error, it will return zero 429087ffc118SAndrea Arcangeli * instead. 429187ffc118SAndrea Arcangeli * 429287ffc118SAndrea Arcangeli * No need to update "position" as the 429387ffc118SAndrea Arcangeli * caller will not check it after 429487ffc118SAndrea Arcangeli * *nr_pages is set to 0. 429587ffc118SAndrea Arcangeli */ 429687ffc118SAndrea Arcangeli return i; 429787ffc118SAndrea Arcangeli } 429887ffc118SAndrea Arcangeli continue; 429987ffc118SAndrea Arcangeli } 430063551ae0SDavid Gibson 4301a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 43027f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 43038fde12caSLinus Torvalds 43048fde12caSLinus Torvalds /* 43058fde12caSLinus Torvalds * Instead of doing 'try_get_page()' below in the same_page 43068fde12caSLinus Torvalds * loop, just check the count once here. 43078fde12caSLinus Torvalds */ 43088fde12caSLinus Torvalds if (unlikely(page_count(page) <= 0)) { 43098fde12caSLinus Torvalds if (pages) { 43108fde12caSLinus Torvalds spin_unlock(ptl); 43118fde12caSLinus Torvalds remainder = 0; 43128fde12caSLinus Torvalds err = -ENOMEM; 43138fde12caSLinus Torvalds break; 43148fde12caSLinus Torvalds } 43158fde12caSLinus Torvalds } 4316d5d4b0aaSChen, Kenneth W same_page: 4317d6692183SChen, Kenneth W if (pages) { 431869d177c2SAndy Whitcroft pages[i] = mem_map_offset(page, pfn_offset); 4319ddc58f27SKirill A. Shutemov get_page(pages[i]); 4320d6692183SChen, Kenneth W } 432163551ae0SDavid Gibson 432263551ae0SDavid Gibson if (vmas) 432363551ae0SDavid Gibson vmas[i] = vma; 432463551ae0SDavid Gibson 432563551ae0SDavid Gibson vaddr += PAGE_SIZE; 4326d5d4b0aaSChen, Kenneth W ++pfn_offset; 432763551ae0SDavid Gibson --remainder; 432863551ae0SDavid Gibson ++i; 4329d5d4b0aaSChen, Kenneth W if (vaddr < vma->vm_end && remainder && 4330a5516438SAndi Kleen pfn_offset < pages_per_huge_page(h)) { 4331d5d4b0aaSChen, Kenneth W /* 4332d5d4b0aaSChen, Kenneth W * We use pfn_offset to avoid touching the pageframes 4333d5d4b0aaSChen, Kenneth W * of this compound page. 4334d5d4b0aaSChen, Kenneth W */ 4335d5d4b0aaSChen, Kenneth W goto same_page; 4336d5d4b0aaSChen, Kenneth W } 4337cb900f41SKirill A. Shutemov spin_unlock(ptl); 433863551ae0SDavid Gibson } 433928a35716SMichel Lespinasse *nr_pages = remainder; 434087ffc118SAndrea Arcangeli /* 434187ffc118SAndrea Arcangeli * setting position is actually required only if remainder is 434287ffc118SAndrea Arcangeli * not zero but it's faster not to add a "if (remainder)" 434387ffc118SAndrea Arcangeli * branch. 434487ffc118SAndrea Arcangeli */ 434563551ae0SDavid Gibson *position = vaddr; 434663551ae0SDavid Gibson 43472be7cfedSDaniel Jordan return i ? i : err; 434863551ae0SDavid Gibson } 43498f860591SZhang, Yanmin 43505491ae7bSAneesh Kumar K.V #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE 43515491ae7bSAneesh Kumar K.V /* 43525491ae7bSAneesh Kumar K.V * ARCHes with special requirements for evicting HUGETLB backing TLB entries can 43535491ae7bSAneesh Kumar K.V * implement this. 43545491ae7bSAneesh Kumar K.V */ 43555491ae7bSAneesh Kumar K.V #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) 43565491ae7bSAneesh Kumar K.V #endif 43575491ae7bSAneesh Kumar K.V 43587da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 43598f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 43608f860591SZhang, Yanmin { 43618f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 43628f860591SZhang, Yanmin unsigned long start = address; 43638f860591SZhang, Yanmin pte_t *ptep; 43648f860591SZhang, Yanmin pte_t pte; 4365a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 43667da4d641SPeter Zijlstra unsigned long pages = 0; 4367dff11abeSMike Kravetz bool shared_pmd = false; 4368ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 4369dff11abeSMike Kravetz 4370dff11abeSMike Kravetz /* 4371dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 4372ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 4373dff11abeSMike Kravetz * range if PMD sharing is possible. 4374dff11abeSMike Kravetz */ 4375ac46d4f3SJérôme Glisse mmu_notifier_range_init(&range, mm, start, end); 4376ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 43778f860591SZhang, Yanmin 43788f860591SZhang, Yanmin BUG_ON(address >= end); 4379ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 43808f860591SZhang, Yanmin 4381ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 438283cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 4383a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 4384cb900f41SKirill A. Shutemov spinlock_t *ptl; 43857868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, huge_page_size(h)); 43868f860591SZhang, Yanmin if (!ptep) 43878f860591SZhang, Yanmin continue; 4388cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 43897da4d641SPeter Zijlstra if (huge_pmd_unshare(mm, &address, ptep)) { 43907da4d641SPeter Zijlstra pages++; 4391cb900f41SKirill A. Shutemov spin_unlock(ptl); 4392dff11abeSMike Kravetz shared_pmd = true; 439339dde65cSChen, Kenneth W continue; 43947da4d641SPeter Zijlstra } 4395a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 4396a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 4397a8bda28dSNaoya Horiguchi spin_unlock(ptl); 4398a8bda28dSNaoya Horiguchi continue; 4399a8bda28dSNaoya Horiguchi } 4400a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 4401a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 4402a8bda28dSNaoya Horiguchi 4403a8bda28dSNaoya Horiguchi if (is_write_migration_entry(entry)) { 4404a8bda28dSNaoya Horiguchi pte_t newpte; 4405a8bda28dSNaoya Horiguchi 4406a8bda28dSNaoya Horiguchi make_migration_entry_read(&entry); 4407a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 4408e5251fd4SPunit Agrawal set_huge_swap_pte_at(mm, address, ptep, 4409e5251fd4SPunit Agrawal newpte, huge_page_size(h)); 4410a8bda28dSNaoya Horiguchi pages++; 4411a8bda28dSNaoya Horiguchi } 4412a8bda28dSNaoya Horiguchi spin_unlock(ptl); 4413a8bda28dSNaoya Horiguchi continue; 4414a8bda28dSNaoya Horiguchi } 4415a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 4416023bdd00SAneesh Kumar K.V pte_t old_pte; 4417023bdd00SAneesh Kumar K.V 4418023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 4419023bdd00SAneesh Kumar K.V pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); 4420be7517d6STony Lu pte = arch_make_huge_pte(pte, vma, NULL, 0); 4421023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 44227da4d641SPeter Zijlstra pages++; 44238f860591SZhang, Yanmin } 4424cb900f41SKirill A. Shutemov spin_unlock(ptl); 44258f860591SZhang, Yanmin } 4426d833352aSMel Gorman /* 4427c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 4428d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 4429c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 4430dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 4431dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 4432d833352aSMel Gorman */ 4433dff11abeSMike Kravetz if (shared_pmd) 4434ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 4435dff11abeSMike Kravetz else 44365491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 44370f10851eSJérôme Glisse /* 44380f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 44390f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 44400f10851eSJérôme Glisse * 4441ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 44420f10851eSJérôme Glisse */ 444383cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 4444ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 44457da4d641SPeter Zijlstra 44467da4d641SPeter Zijlstra return pages << h->order; 44478f860591SZhang, Yanmin } 44488f860591SZhang, Yanmin 4449a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 4450a1e78772SMel Gorman long from, long to, 44515a6fe125SMel Gorman struct vm_area_struct *vma, 4452ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 4453e4e574b7SAdam Litke { 445417c9d12eSMel Gorman long ret, chg; 4455a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 445690481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 44579119a41eSJoonsoo Kim struct resv_map *resv_map; 44581c5ecae3SMike Kravetz long gbl_reserve; 4459e4e574b7SAdam Litke 446063489f8eSMike Kravetz /* This should never happen */ 446163489f8eSMike Kravetz if (from > to) { 446263489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 446363489f8eSMike Kravetz return -EINVAL; 446463489f8eSMike Kravetz } 446563489f8eSMike Kravetz 4466a1e78772SMel Gorman /* 446717c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 446817c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 446990481622SDavid Gibson * without using reserves 447017c9d12eSMel Gorman */ 4471ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 447217c9d12eSMel Gorman return 0; 447317c9d12eSMel Gorman 447417c9d12eSMel Gorman /* 4475a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 4476a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 4477a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 4478a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 4479a1e78772SMel Gorman */ 44809119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 44814e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 44829119a41eSJoonsoo Kim 44831406ec9bSJoonsoo Kim chg = region_chg(resv_map, from, to); 44849119a41eSJoonsoo Kim 44859119a41eSJoonsoo Kim } else { 44869119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 44875a6fe125SMel Gorman if (!resv_map) 44885a6fe125SMel Gorman return -ENOMEM; 44895a6fe125SMel Gorman 449017c9d12eSMel Gorman chg = to - from; 449117c9d12eSMel Gorman 44925a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 44935a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 44945a6fe125SMel Gorman } 44955a6fe125SMel Gorman 4496c50ac050SDave Hansen if (chg < 0) { 4497c50ac050SDave Hansen ret = chg; 4498c50ac050SDave Hansen goto out_err; 4499c50ac050SDave Hansen } 450017c9d12eSMel Gorman 45011c5ecae3SMike Kravetz /* 45021c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 45031c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 45041c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 45051c5ecae3SMike Kravetz */ 45061c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 45071c5ecae3SMike Kravetz if (gbl_reserve < 0) { 4508c50ac050SDave Hansen ret = -ENOSPC; 4509c50ac050SDave Hansen goto out_err; 4510c50ac050SDave Hansen } 451117c9d12eSMel Gorman 451217c9d12eSMel Gorman /* 451317c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 451490481622SDavid Gibson * Hand the pages back to the subpool if there are not 451517c9d12eSMel Gorman */ 45161c5ecae3SMike Kravetz ret = hugetlb_acct_memory(h, gbl_reserve); 451717c9d12eSMel Gorman if (ret < 0) { 45181c5ecae3SMike Kravetz /* put back original number of pages, chg */ 45191c5ecae3SMike Kravetz (void)hugepage_subpool_put_pages(spool, chg); 4520c50ac050SDave Hansen goto out_err; 452117c9d12eSMel Gorman } 452217c9d12eSMel Gorman 452317c9d12eSMel Gorman /* 452417c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 452517c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 452617c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 452717c9d12eSMel Gorman * the reservation was and the page cache tells how much of 452817c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 452917c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 453017c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 453117c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 453217c9d12eSMel Gorman * else has to be done for private mappings here 453317c9d12eSMel Gorman */ 453433039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 453533039678SMike Kravetz long add = region_add(resv_map, from, to); 453633039678SMike Kravetz 453733039678SMike Kravetz if (unlikely(chg > add)) { 453833039678SMike Kravetz /* 453933039678SMike Kravetz * pages in this range were added to the reserve 454033039678SMike Kravetz * map between region_chg and region_add. This 454133039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 454233039678SMike Kravetz * the subpool and reserve counts modified above 454333039678SMike Kravetz * based on the difference. 454433039678SMike Kravetz */ 454533039678SMike Kravetz long rsv_adjust; 454633039678SMike Kravetz 454733039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 454833039678SMike Kravetz chg - add); 454933039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 455033039678SMike Kravetz } 455133039678SMike Kravetz } 4552a43a8c39SChen, Kenneth W return 0; 4553c50ac050SDave Hansen out_err: 45545e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 4555ff8c0c53SMike Kravetz /* Don't call region_abort if region_chg failed */ 4556ff8c0c53SMike Kravetz if (chg >= 0) 45575e911373SMike Kravetz region_abort(resv_map, from, to); 4558f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 4559f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 4560c50ac050SDave Hansen return ret; 4561a43a8c39SChen, Kenneth W } 4562a43a8c39SChen, Kenneth W 4563b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 4564b5cec28dSMike Kravetz long freed) 4565a43a8c39SChen, Kenneth W { 4566a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 45674e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 45689119a41eSJoonsoo Kim long chg = 0; 456990481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 45701c5ecae3SMike Kravetz long gbl_reserve; 457145c682a6SKen Chen 4572b5cec28dSMike Kravetz if (resv_map) { 4573b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 4574b5cec28dSMike Kravetz /* 4575b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 4576b5cec28dSMike Kravetz * must be split and another region descriptor can not be 4577b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 4578b5cec28dSMike Kravetz */ 4579b5cec28dSMike Kravetz if (chg < 0) 4580b5cec28dSMike Kravetz return chg; 4581b5cec28dSMike Kravetz } 4582b5cec28dSMike Kravetz 458345c682a6SKen Chen spin_lock(&inode->i_lock); 4584e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 458545c682a6SKen Chen spin_unlock(&inode->i_lock); 458645c682a6SKen Chen 45871c5ecae3SMike Kravetz /* 45881c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 45891c5ecae3SMike Kravetz * reservations to be released may be adjusted. 45901c5ecae3SMike Kravetz */ 45911c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 45921c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 4593b5cec28dSMike Kravetz 4594b5cec28dSMike Kravetz return 0; 4595a43a8c39SChen, Kenneth W } 459693f70f90SNaoya Horiguchi 45973212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 45983212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 45993212b535SSteve Capper struct vm_area_struct *vma, 46003212b535SSteve Capper unsigned long addr, pgoff_t idx) 46013212b535SSteve Capper { 46023212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 46033212b535SSteve Capper svma->vm_start; 46043212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 46053212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 46063212b535SSteve Capper 46073212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 4608de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 4609de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 46103212b535SSteve Capper 46113212b535SSteve Capper /* 46123212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 46133212b535SSteve Capper * page table page. 46143212b535SSteve Capper */ 46153212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 46163212b535SSteve Capper vm_flags != svm_flags || 46173212b535SSteve Capper sbase < svma->vm_start || svma->vm_end < s_end) 46183212b535SSteve Capper return 0; 46193212b535SSteve Capper 46203212b535SSteve Capper return saddr; 46213212b535SSteve Capper } 46223212b535SSteve Capper 462331aafb45SNicholas Krause static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) 46243212b535SSteve Capper { 46253212b535SSteve Capper unsigned long base = addr & PUD_MASK; 46263212b535SSteve Capper unsigned long end = base + PUD_SIZE; 46273212b535SSteve Capper 46283212b535SSteve Capper /* 46293212b535SSteve Capper * check on proper vm_flags and page table alignment 46303212b535SSteve Capper */ 4631017b1660SMike Kravetz if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) 463231aafb45SNicholas Krause return true; 463331aafb45SNicholas Krause return false; 46343212b535SSteve Capper } 46353212b535SSteve Capper 46363212b535SSteve Capper /* 4637017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 4638017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 4639017b1660SMike Kravetz * shared pmd mappings. 4640017b1660SMike Kravetz */ 4641017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 4642017b1660SMike Kravetz unsigned long *start, unsigned long *end) 4643017b1660SMike Kravetz { 4644017b1660SMike Kravetz unsigned long check_addr = *start; 4645017b1660SMike Kravetz 4646017b1660SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 4647017b1660SMike Kravetz return; 4648017b1660SMike Kravetz 4649017b1660SMike Kravetz for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) { 4650017b1660SMike Kravetz unsigned long a_start = check_addr & PUD_MASK; 4651017b1660SMike Kravetz unsigned long a_end = a_start + PUD_SIZE; 4652017b1660SMike Kravetz 4653017b1660SMike Kravetz /* 4654017b1660SMike Kravetz * If sharing is possible, adjust start/end if necessary. 4655017b1660SMike Kravetz */ 4656017b1660SMike Kravetz if (range_in_vma(vma, a_start, a_end)) { 4657017b1660SMike Kravetz if (a_start < *start) 4658017b1660SMike Kravetz *start = a_start; 4659017b1660SMike Kravetz if (a_end > *end) 4660017b1660SMike Kravetz *end = a_end; 4661017b1660SMike Kravetz } 4662017b1660SMike Kravetz } 4663017b1660SMike Kravetz } 4664017b1660SMike Kravetz 4665017b1660SMike Kravetz /* 46663212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 46673212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 46683212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 4669ddeaab32SMike Kravetz * code much cleaner. pmd allocation is essential for the shared case because 4670ddeaab32SMike Kravetz * pud has to be populated inside the same i_mmap_rwsem section - otherwise 4671ddeaab32SMike Kravetz * racing tasks could either miss the sharing (see huge_pte_offset) or select a 4672ddeaab32SMike Kravetz * bad pmd for sharing. 46733212b535SSteve Capper */ 46743212b535SSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 46753212b535SSteve Capper { 46763212b535SSteve Capper struct vm_area_struct *vma = find_vma(mm, addr); 46773212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 46783212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 46793212b535SSteve Capper vma->vm_pgoff; 46803212b535SSteve Capper struct vm_area_struct *svma; 46813212b535SSteve Capper unsigned long saddr; 46823212b535SSteve Capper pte_t *spte = NULL; 46833212b535SSteve Capper pte_t *pte; 4684cb900f41SKirill A. Shutemov spinlock_t *ptl; 46853212b535SSteve Capper 46863212b535SSteve Capper if (!vma_shareable(vma, addr)) 46873212b535SSteve Capper return (pte_t *)pmd_alloc(mm, pud, addr); 46883212b535SSteve Capper 4689ddeaab32SMike Kravetz i_mmap_lock_write(mapping); 46903212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 46913212b535SSteve Capper if (svma == vma) 46923212b535SSteve Capper continue; 46933212b535SSteve Capper 46943212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 46953212b535SSteve Capper if (saddr) { 46967868a208SPunit Agrawal spte = huge_pte_offset(svma->vm_mm, saddr, 46977868a208SPunit Agrawal vma_mmu_pagesize(svma)); 46983212b535SSteve Capper if (spte) { 46993212b535SSteve Capper get_page(virt_to_page(spte)); 47003212b535SSteve Capper break; 47013212b535SSteve Capper } 47023212b535SSteve Capper } 47033212b535SSteve Capper } 47043212b535SSteve Capper 47053212b535SSteve Capper if (!spte) 47063212b535SSteve Capper goto out; 47073212b535SSteve Capper 47088bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 4709dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 47103212b535SSteve Capper pud_populate(mm, pud, 47113212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 4712c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 4713dc6c9a35SKirill A. Shutemov } else { 47143212b535SSteve Capper put_page(virt_to_page(spte)); 4715dc6c9a35SKirill A. Shutemov } 4716cb900f41SKirill A. Shutemov spin_unlock(ptl); 47173212b535SSteve Capper out: 47183212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 4719ddeaab32SMike Kravetz i_mmap_unlock_write(mapping); 47203212b535SSteve Capper return pte; 47213212b535SSteve Capper } 47223212b535SSteve Capper 47233212b535SSteve Capper /* 47243212b535SSteve Capper * unmap huge page backed by shared pte. 47253212b535SSteve Capper * 47263212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 47273212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 47283212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 47293212b535SSteve Capper * 4730ddeaab32SMike Kravetz * called with page table lock held. 47313212b535SSteve Capper * 47323212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 47333212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 47343212b535SSteve Capper */ 47353212b535SSteve Capper int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 47363212b535SSteve Capper { 47373212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 4738c2febafcSKirill A. Shutemov p4d_t *p4d = p4d_offset(pgd, *addr); 4739c2febafcSKirill A. Shutemov pud_t *pud = pud_offset(p4d, *addr); 47403212b535SSteve Capper 47413212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 47423212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 47433212b535SSteve Capper return 0; 47443212b535SSteve Capper 47453212b535SSteve Capper pud_clear(pud); 47463212b535SSteve Capper put_page(virt_to_page(ptep)); 4747dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 47483212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 47493212b535SSteve Capper return 1; 47503212b535SSteve Capper } 47519e5fc74cSSteve Capper #define want_pmd_share() (1) 47529e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 47539e5fc74cSSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 47549e5fc74cSSteve Capper { 47559e5fc74cSSteve Capper return NULL; 47569e5fc74cSSteve Capper } 4757e81f2d22SZhang Zhen 4758e81f2d22SZhang Zhen int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 4759e81f2d22SZhang Zhen { 4760e81f2d22SZhang Zhen return 0; 4761e81f2d22SZhang Zhen } 4762017b1660SMike Kravetz 4763017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 4764017b1660SMike Kravetz unsigned long *start, unsigned long *end) 4765017b1660SMike Kravetz { 4766017b1660SMike Kravetz } 47679e5fc74cSSteve Capper #define want_pmd_share() (0) 47683212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 47693212b535SSteve Capper 47709e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 47719e5fc74cSSteve Capper pte_t *huge_pte_alloc(struct mm_struct *mm, 47729e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 47739e5fc74cSSteve Capper { 47749e5fc74cSSteve Capper pgd_t *pgd; 4775c2febafcSKirill A. Shutemov p4d_t *p4d; 47769e5fc74cSSteve Capper pud_t *pud; 47779e5fc74cSSteve Capper pte_t *pte = NULL; 47789e5fc74cSSteve Capper 47799e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 4780f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 4781f4f0a3d8SKirill A. Shutemov if (!p4d) 4782f4f0a3d8SKirill A. Shutemov return NULL; 4783c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 47849e5fc74cSSteve Capper if (pud) { 47859e5fc74cSSteve Capper if (sz == PUD_SIZE) { 47869e5fc74cSSteve Capper pte = (pte_t *)pud; 47879e5fc74cSSteve Capper } else { 47889e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 47899e5fc74cSSteve Capper if (want_pmd_share() && pud_none(*pud)) 47909e5fc74cSSteve Capper pte = huge_pmd_share(mm, addr, pud); 47919e5fc74cSSteve Capper else 47929e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 47939e5fc74cSSteve Capper } 47949e5fc74cSSteve Capper } 47954e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 47969e5fc74cSSteve Capper 47979e5fc74cSSteve Capper return pte; 47989e5fc74cSSteve Capper } 47999e5fc74cSSteve Capper 48009b19df29SPunit Agrawal /* 48019b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 48029b19df29SPunit Agrawal * entry at address @addr 48039b19df29SPunit Agrawal * 48049b19df29SPunit Agrawal * Return: Pointer to page table or swap entry (PUD or PMD) for 48059b19df29SPunit Agrawal * address @addr, or NULL if a p*d_none() entry is encountered and the 48069b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 48079b19df29SPunit Agrawal * table. 48089b19df29SPunit Agrawal */ 48097868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 48107868a208SPunit Agrawal unsigned long addr, unsigned long sz) 48119e5fc74cSSteve Capper { 48129e5fc74cSSteve Capper pgd_t *pgd; 4813c2febafcSKirill A. Shutemov p4d_t *p4d; 48149e5fc74cSSteve Capper pud_t *pud; 4815c2febafcSKirill A. Shutemov pmd_t *pmd; 48169e5fc74cSSteve Capper 48179e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 4818c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 4819c2febafcSKirill A. Shutemov return NULL; 4820c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 4821c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 4822c2febafcSKirill A. Shutemov return NULL; 48239b19df29SPunit Agrawal 4824c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 48259b19df29SPunit Agrawal if (sz != PUD_SIZE && pud_none(*pud)) 4826c2febafcSKirill A. Shutemov return NULL; 48279b19df29SPunit Agrawal /* hugepage or swap? */ 48289b19df29SPunit Agrawal if (pud_huge(*pud) || !pud_present(*pud)) 48299e5fc74cSSteve Capper return (pte_t *)pud; 48309b19df29SPunit Agrawal 48319e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 48329b19df29SPunit Agrawal if (sz != PMD_SIZE && pmd_none(*pmd)) 48339b19df29SPunit Agrawal return NULL; 48349b19df29SPunit Agrawal /* hugepage or swap? */ 48359b19df29SPunit Agrawal if (pmd_huge(*pmd) || !pmd_present(*pmd)) 48369e5fc74cSSteve Capper return (pte_t *)pmd; 48379b19df29SPunit Agrawal 48389b19df29SPunit Agrawal return NULL; 48399e5fc74cSSteve Capper } 48409e5fc74cSSteve Capper 484161f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 484261f77edaSNaoya Horiguchi 484361f77edaSNaoya Horiguchi /* 484461f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 484561f77edaSNaoya Horiguchi * behavior. 484661f77edaSNaoya Horiguchi */ 484761f77edaSNaoya Horiguchi struct page * __weak 484861f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 484961f77edaSNaoya Horiguchi int write) 485061f77edaSNaoya Horiguchi { 485161f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 485261f77edaSNaoya Horiguchi } 485361f77edaSNaoya Horiguchi 485461f77edaSNaoya Horiguchi struct page * __weak 48554dc71451SAneesh Kumar K.V follow_huge_pd(struct vm_area_struct *vma, 48564dc71451SAneesh Kumar K.V unsigned long address, hugepd_t hpd, int flags, int pdshift) 48574dc71451SAneesh Kumar K.V { 48584dc71451SAneesh Kumar K.V WARN(1, "hugepd follow called with no support for hugepage directory format\n"); 48594dc71451SAneesh Kumar K.V return NULL; 48604dc71451SAneesh Kumar K.V } 48614dc71451SAneesh Kumar K.V 48624dc71451SAneesh Kumar K.V struct page * __weak 48639e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 4864e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 48659e5fc74cSSteve Capper { 4866e66f17ffSNaoya Horiguchi struct page *page = NULL; 4867e66f17ffSNaoya Horiguchi spinlock_t *ptl; 4868c9d398faSNaoya Horiguchi pte_t pte; 4869e66f17ffSNaoya Horiguchi retry: 4870e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 4871e66f17ffSNaoya Horiguchi spin_lock(ptl); 4872e66f17ffSNaoya Horiguchi /* 4873e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 4874e66f17ffSNaoya Horiguchi * unmapped from other threads. 4875e66f17ffSNaoya Horiguchi */ 4876e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 4877e66f17ffSNaoya Horiguchi goto out; 4878c9d398faSNaoya Horiguchi pte = huge_ptep_get((pte_t *)pmd); 4879c9d398faSNaoya Horiguchi if (pte_present(pte)) { 488097534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 4881e66f17ffSNaoya Horiguchi if (flags & FOLL_GET) 4882e66f17ffSNaoya Horiguchi get_page(page); 4883e66f17ffSNaoya Horiguchi } else { 4884c9d398faSNaoya Horiguchi if (is_hugetlb_entry_migration(pte)) { 4885e66f17ffSNaoya Horiguchi spin_unlock(ptl); 4886e66f17ffSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pmd, ptl); 4887e66f17ffSNaoya Horiguchi goto retry; 4888e66f17ffSNaoya Horiguchi } 4889e66f17ffSNaoya Horiguchi /* 4890e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 4891e66f17ffSNaoya Horiguchi * follow_page_mask(). 4892e66f17ffSNaoya Horiguchi */ 4893e66f17ffSNaoya Horiguchi } 4894e66f17ffSNaoya Horiguchi out: 4895e66f17ffSNaoya Horiguchi spin_unlock(ptl); 48969e5fc74cSSteve Capper return page; 48979e5fc74cSSteve Capper } 48989e5fc74cSSteve Capper 489961f77edaSNaoya Horiguchi struct page * __weak 49009e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 4901e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 49029e5fc74cSSteve Capper { 4903e66f17ffSNaoya Horiguchi if (flags & FOLL_GET) 4904e66f17ffSNaoya Horiguchi return NULL; 49059e5fc74cSSteve Capper 4906e66f17ffSNaoya Horiguchi return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 49079e5fc74cSSteve Capper } 49089e5fc74cSSteve Capper 4909faaa5b62SAnshuman Khandual struct page * __weak 4910faaa5b62SAnshuman Khandual follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) 4911faaa5b62SAnshuman Khandual { 4912faaa5b62SAnshuman Khandual if (flags & FOLL_GET) 4913faaa5b62SAnshuman Khandual return NULL; 4914faaa5b62SAnshuman Khandual 4915faaa5b62SAnshuman Khandual return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); 4916faaa5b62SAnshuman Khandual } 4917faaa5b62SAnshuman Khandual 491831caf665SNaoya Horiguchi bool isolate_huge_page(struct page *page, struct list_head *list) 491931caf665SNaoya Horiguchi { 4920bcc54222SNaoya Horiguchi bool ret = true; 4921bcc54222SNaoya Horiguchi 4922309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 492331caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 4924bcc54222SNaoya Horiguchi if (!page_huge_active(page) || !get_page_unless_zero(page)) { 4925bcc54222SNaoya Horiguchi ret = false; 4926bcc54222SNaoya Horiguchi goto unlock; 4927bcc54222SNaoya Horiguchi } 4928bcc54222SNaoya Horiguchi clear_page_huge_active(page); 492931caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 4930bcc54222SNaoya Horiguchi unlock: 493131caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 4932bcc54222SNaoya Horiguchi return ret; 493331caf665SNaoya Horiguchi } 493431caf665SNaoya Horiguchi 493531caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 493631caf665SNaoya Horiguchi { 4937309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 493831caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 4939bcc54222SNaoya Horiguchi set_page_huge_active(page); 494031caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 494131caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 494231caf665SNaoya Horiguchi put_page(page); 494331caf665SNaoya Horiguchi } 4944ab5ac90aSMichal Hocko 4945ab5ac90aSMichal Hocko void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) 4946ab5ac90aSMichal Hocko { 4947ab5ac90aSMichal Hocko struct hstate *h = page_hstate(oldpage); 4948ab5ac90aSMichal Hocko 4949ab5ac90aSMichal Hocko hugetlb_cgroup_migrate(oldpage, newpage); 4950ab5ac90aSMichal Hocko set_page_owner_migrate_reason(newpage, reason); 4951ab5ac90aSMichal Hocko 4952ab5ac90aSMichal Hocko /* 4953ab5ac90aSMichal Hocko * transfer temporary state of the new huge page. This is 4954ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 4955ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 4956ab5ac90aSMichal Hocko * the temporary status. 4957ab5ac90aSMichal Hocko * 4958ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 4959ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 4960ab5ac90aSMichal Hocko * the per-node's. 4961ab5ac90aSMichal Hocko */ 4962ab5ac90aSMichal Hocko if (PageHugeTemporary(newpage)) { 4963ab5ac90aSMichal Hocko int old_nid = page_to_nid(oldpage); 4964ab5ac90aSMichal Hocko int new_nid = page_to_nid(newpage); 4965ab5ac90aSMichal Hocko 4966ab5ac90aSMichal Hocko SetPageHugeTemporary(oldpage); 4967ab5ac90aSMichal Hocko ClearPageHugeTemporary(newpage); 4968ab5ac90aSMichal Hocko 4969ab5ac90aSMichal Hocko spin_lock(&hugetlb_lock); 4970ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 4971ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 4972ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 4973ab5ac90aSMichal Hocko } 4974ab5ac90aSMichal Hocko spin_unlock(&hugetlb_lock); 4975ab5ac90aSMichal Hocko } 4976ab5ac90aSMichal Hocko } 4977