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 10624eb0716eSAlexandre Ghiti #ifdef CONFIG_CONTIG_ALLOC 1063944d9fecSLuiz Capitulino static int __alloc_gigantic_page(unsigned long start_pfn, 106479b63f12SMichal Hocko unsigned long nr_pages, gfp_t gfp_mask) 1065944d9fecSLuiz Capitulino { 1066944d9fecSLuiz Capitulino unsigned long end_pfn = start_pfn + nr_pages; 1067ca96b625SLucas Stach return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE, 106879b63f12SMichal Hocko gfp_mask); 1069944d9fecSLuiz Capitulino } 1070944d9fecSLuiz Capitulino 1071f44b2ddaSJoonsoo Kim static bool pfn_range_valid_gigantic(struct zone *z, 1072f44b2ddaSJoonsoo Kim unsigned long start_pfn, unsigned long nr_pages) 1073944d9fecSLuiz Capitulino { 1074944d9fecSLuiz Capitulino unsigned long i, end_pfn = start_pfn + nr_pages; 1075944d9fecSLuiz Capitulino struct page *page; 1076944d9fecSLuiz Capitulino 1077944d9fecSLuiz Capitulino for (i = start_pfn; i < end_pfn; i++) { 1078944d9fecSLuiz Capitulino if (!pfn_valid(i)) 1079944d9fecSLuiz Capitulino return false; 1080944d9fecSLuiz Capitulino 1081944d9fecSLuiz Capitulino page = pfn_to_page(i); 1082944d9fecSLuiz Capitulino 1083f44b2ddaSJoonsoo Kim if (page_zone(page) != z) 1084f44b2ddaSJoonsoo Kim return false; 1085f44b2ddaSJoonsoo Kim 1086944d9fecSLuiz Capitulino if (PageReserved(page)) 1087944d9fecSLuiz Capitulino return false; 1088944d9fecSLuiz Capitulino 1089944d9fecSLuiz Capitulino if (page_count(page) > 0) 1090944d9fecSLuiz Capitulino return false; 1091944d9fecSLuiz Capitulino 1092944d9fecSLuiz Capitulino if (PageHuge(page)) 1093944d9fecSLuiz Capitulino return false; 1094944d9fecSLuiz Capitulino } 1095944d9fecSLuiz Capitulino 1096944d9fecSLuiz Capitulino return true; 1097944d9fecSLuiz Capitulino } 1098944d9fecSLuiz Capitulino 1099944d9fecSLuiz Capitulino static bool zone_spans_last_pfn(const struct zone *zone, 1100944d9fecSLuiz Capitulino unsigned long start_pfn, unsigned long nr_pages) 1101944d9fecSLuiz Capitulino { 1102944d9fecSLuiz Capitulino unsigned long last_pfn = start_pfn + nr_pages - 1; 1103944d9fecSLuiz Capitulino return zone_spans_pfn(zone, last_pfn); 1104944d9fecSLuiz Capitulino } 1105944d9fecSLuiz Capitulino 1106d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 1107d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) 1108944d9fecSLuiz Capitulino { 110979b63f12SMichal Hocko unsigned int order = huge_page_order(h); 1110944d9fecSLuiz Capitulino unsigned long nr_pages = 1 << order; 1111944d9fecSLuiz Capitulino unsigned long ret, pfn, flags; 111279b63f12SMichal Hocko struct zonelist *zonelist; 111379b63f12SMichal Hocko struct zone *zone; 111479b63f12SMichal Hocko struct zoneref *z; 1115944d9fecSLuiz Capitulino 111679b63f12SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 1117d9cc948fSMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nodemask) { 111879b63f12SMichal Hocko spin_lock_irqsave(&zone->lock, flags); 1119944d9fecSLuiz Capitulino 112079b63f12SMichal Hocko pfn = ALIGN(zone->zone_start_pfn, nr_pages); 112179b63f12SMichal Hocko while (zone_spans_last_pfn(zone, pfn, nr_pages)) { 112279b63f12SMichal Hocko if (pfn_range_valid_gigantic(zone, pfn, nr_pages)) { 1123944d9fecSLuiz Capitulino /* 1124944d9fecSLuiz Capitulino * We release the zone lock here because 1125944d9fecSLuiz Capitulino * alloc_contig_range() will also lock the zone 1126944d9fecSLuiz Capitulino * at some point. If there's an allocation 1127944d9fecSLuiz Capitulino * spinning on this lock, it may win the race 1128944d9fecSLuiz Capitulino * and cause alloc_contig_range() to fail... 1129944d9fecSLuiz Capitulino */ 113079b63f12SMichal Hocko spin_unlock_irqrestore(&zone->lock, flags); 113179b63f12SMichal Hocko ret = __alloc_gigantic_page(pfn, nr_pages, gfp_mask); 1132944d9fecSLuiz Capitulino if (!ret) 1133944d9fecSLuiz Capitulino return pfn_to_page(pfn); 113479b63f12SMichal Hocko spin_lock_irqsave(&zone->lock, flags); 1135944d9fecSLuiz Capitulino } 1136944d9fecSLuiz Capitulino pfn += nr_pages; 1137944d9fecSLuiz Capitulino } 1138944d9fecSLuiz Capitulino 113979b63f12SMichal Hocko spin_unlock_irqrestore(&zone->lock, flags); 1140944d9fecSLuiz Capitulino } 1141944d9fecSLuiz Capitulino 1142944d9fecSLuiz Capitulino return NULL; 1143944d9fecSLuiz Capitulino } 1144944d9fecSLuiz Capitulino 1145944d9fecSLuiz Capitulino static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); 1146d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order); 11474eb0716eSAlexandre Ghiti #else /* !CONFIG_CONTIG_ALLOC */ 11484eb0716eSAlexandre Ghiti static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 11494eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 11504eb0716eSAlexandre Ghiti { 11514eb0716eSAlexandre Ghiti return NULL; 11524eb0716eSAlexandre Ghiti } 11534eb0716eSAlexandre Ghiti #endif /* CONFIG_CONTIG_ALLOC */ 1154944d9fecSLuiz Capitulino 1155e1073d1eSAneesh Kumar K.V #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 1156d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 11574eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 11584eb0716eSAlexandre Ghiti { 11594eb0716eSAlexandre Ghiti return NULL; 11604eb0716eSAlexandre Ghiti } 1161d00181b9SKirill A. Shutemov static inline void free_gigantic_page(struct page *page, unsigned int order) { } 1162944d9fecSLuiz Capitulino static inline void destroy_compound_gigantic_page(struct page *page, 1163d00181b9SKirill A. Shutemov unsigned int order) { } 1164944d9fecSLuiz Capitulino #endif 1165944d9fecSLuiz Capitulino 1166a5516438SAndi Kleen static void update_and_free_page(struct hstate *h, struct page *page) 11676af2acb6SAdam Litke { 11686af2acb6SAdam Litke int i; 1169a5516438SAndi Kleen 11704eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1171944d9fecSLuiz Capitulino return; 117218229df5SAndy Whitcroft 1173a5516438SAndi Kleen h->nr_huge_pages--; 1174a5516438SAndi Kleen h->nr_huge_pages_node[page_to_nid(page)]--; 1175a5516438SAndi Kleen for (i = 0; i < pages_per_huge_page(h); i++) { 117632f84528SChris Forbes page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 117732f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1178a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1179a7407a27SLuiz Capitulino 1 << PG_writeback); 11806af2acb6SAdam Litke } 1181309381feSSasha Levin VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 1182f1e61557SKirill A. Shutemov set_compound_page_dtor(page, NULL_COMPOUND_DTOR); 11836af2acb6SAdam Litke set_page_refcounted(page); 1184944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) { 1185944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 1186944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 1187944d9fecSLuiz Capitulino } else { 1188a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 11896af2acb6SAdam Litke } 1190944d9fecSLuiz Capitulino } 11916af2acb6SAdam Litke 1192e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1193e5ff2159SAndi Kleen { 1194e5ff2159SAndi Kleen struct hstate *h; 1195e5ff2159SAndi Kleen 1196e5ff2159SAndi Kleen for_each_hstate(h) { 1197e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1198e5ff2159SAndi Kleen return h; 1199e5ff2159SAndi Kleen } 1200e5ff2159SAndi Kleen return NULL; 1201e5ff2159SAndi Kleen } 1202e5ff2159SAndi Kleen 1203bcc54222SNaoya Horiguchi /* 1204bcc54222SNaoya Horiguchi * Test to determine whether the hugepage is "active/in-use" (i.e. being linked 1205bcc54222SNaoya Horiguchi * to hstate->hugepage_activelist.) 1206bcc54222SNaoya Horiguchi * 1207bcc54222SNaoya Horiguchi * This function can be called for tail pages, but never returns true for them. 1208bcc54222SNaoya Horiguchi */ 1209bcc54222SNaoya Horiguchi bool page_huge_active(struct page *page) 1210bcc54222SNaoya Horiguchi { 1211bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHuge(page), page); 1212bcc54222SNaoya Horiguchi return PageHead(page) && PagePrivate(&page[1]); 1213bcc54222SNaoya Horiguchi } 1214bcc54222SNaoya Horiguchi 1215bcc54222SNaoya Horiguchi /* never called for tail page */ 1216bcc54222SNaoya Horiguchi static void set_page_huge_active(struct page *page) 1217bcc54222SNaoya Horiguchi { 1218bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 1219bcc54222SNaoya Horiguchi SetPagePrivate(&page[1]); 1220bcc54222SNaoya Horiguchi } 1221bcc54222SNaoya Horiguchi 1222bcc54222SNaoya Horiguchi static void clear_page_huge_active(struct page *page) 1223bcc54222SNaoya Horiguchi { 1224bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 1225bcc54222SNaoya Horiguchi ClearPagePrivate(&page[1]); 1226bcc54222SNaoya Horiguchi } 1227bcc54222SNaoya Horiguchi 1228ab5ac90aSMichal Hocko /* 1229ab5ac90aSMichal Hocko * Internal hugetlb specific page flag. Do not use outside of the hugetlb 1230ab5ac90aSMichal Hocko * code 1231ab5ac90aSMichal Hocko */ 1232ab5ac90aSMichal Hocko static inline bool PageHugeTemporary(struct page *page) 1233ab5ac90aSMichal Hocko { 1234ab5ac90aSMichal Hocko if (!PageHuge(page)) 1235ab5ac90aSMichal Hocko return false; 1236ab5ac90aSMichal Hocko 1237ab5ac90aSMichal Hocko return (unsigned long)page[2].mapping == -1U; 1238ab5ac90aSMichal Hocko } 1239ab5ac90aSMichal Hocko 1240ab5ac90aSMichal Hocko static inline void SetPageHugeTemporary(struct page *page) 1241ab5ac90aSMichal Hocko { 1242ab5ac90aSMichal Hocko page[2].mapping = (void *)-1U; 1243ab5ac90aSMichal Hocko } 1244ab5ac90aSMichal Hocko 1245ab5ac90aSMichal Hocko static inline void ClearPageHugeTemporary(struct page *page) 1246ab5ac90aSMichal Hocko { 1247ab5ac90aSMichal Hocko page[2].mapping = NULL; 1248ab5ac90aSMichal Hocko } 1249ab5ac90aSMichal Hocko 12508f1d26d0SAtsushi Kumagai void free_huge_page(struct page *page) 125127a85ef1SDavid Gibson { 1252a5516438SAndi Kleen /* 1253a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1254a5516438SAndi Kleen * compound page destructor. 1255a5516438SAndi Kleen */ 1256e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 12577893d1d5SAdam Litke int nid = page_to_nid(page); 125890481622SDavid Gibson struct hugepage_subpool *spool = 125990481622SDavid Gibson (struct hugepage_subpool *)page_private(page); 126007443a85SJoonsoo Kim bool restore_reserve; 126127a85ef1SDavid Gibson 1262b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1263b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_mapcount(page), page); 12648ace22bcSYongkai Wu 12658ace22bcSYongkai Wu set_page_private(page, 0); 12668ace22bcSYongkai Wu page->mapping = NULL; 126707443a85SJoonsoo Kim restore_reserve = PagePrivate(page); 126816c794b4SJoonsoo Kim ClearPagePrivate(page); 126927a85ef1SDavid Gibson 12701c5ecae3SMike Kravetz /* 12711c5ecae3SMike Kravetz * A return code of zero implies that the subpool will be under its 12721c5ecae3SMike Kravetz * minimum size if the reservation is not restored after page is free. 12731c5ecae3SMike Kravetz * Therefore, force restore_reserve operation. 12741c5ecae3SMike Kravetz */ 12751c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 12761c5ecae3SMike Kravetz restore_reserve = true; 12771c5ecae3SMike Kravetz 127827a85ef1SDavid Gibson spin_lock(&hugetlb_lock); 1279bcc54222SNaoya Horiguchi clear_page_huge_active(page); 12806d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 12816d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 128207443a85SJoonsoo Kim if (restore_reserve) 128307443a85SJoonsoo Kim h->resv_huge_pages++; 128407443a85SJoonsoo Kim 1285ab5ac90aSMichal Hocko if (PageHugeTemporary(page)) { 1286ab5ac90aSMichal Hocko list_del(&page->lru); 1287ab5ac90aSMichal Hocko ClearPageHugeTemporary(page); 1288ab5ac90aSMichal Hocko update_and_free_page(h, page); 1289ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 12900edaecfaSAneesh Kumar K.V /* remove the page from active list */ 12910edaecfaSAneesh Kumar K.V list_del(&page->lru); 1292a5516438SAndi Kleen update_and_free_page(h, page); 1293a5516438SAndi Kleen h->surplus_huge_pages--; 1294a5516438SAndi Kleen h->surplus_huge_pages_node[nid]--; 12957893d1d5SAdam Litke } else { 12965d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1297a5516438SAndi Kleen enqueue_huge_page(h, page); 12987893d1d5SAdam Litke } 129927a85ef1SDavid Gibson spin_unlock(&hugetlb_lock); 130027a85ef1SDavid Gibson } 130127a85ef1SDavid Gibson 1302a5516438SAndi Kleen static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1303b7ba30c6SAndi Kleen { 13040edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1305f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1306b7ba30c6SAndi Kleen spin_lock(&hugetlb_lock); 13079dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 1308a5516438SAndi Kleen h->nr_huge_pages++; 1309a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 1310b7ba30c6SAndi Kleen spin_unlock(&hugetlb_lock); 1311b7ba30c6SAndi Kleen } 1312b7ba30c6SAndi Kleen 1313d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order) 131420a0307cSWu Fengguang { 131520a0307cSWu Fengguang int i; 131620a0307cSWu Fengguang int nr_pages = 1 << order; 131720a0307cSWu Fengguang struct page *p = page + 1; 131820a0307cSWu Fengguang 131920a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 132020a0307cSWu Fengguang set_compound_order(page, order); 1321ef5a22beSAndrea Arcangeli __ClearPageReserved(page); 1322de09d31dSKirill A. Shutemov __SetPageHead(page); 132320a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1324ef5a22beSAndrea Arcangeli /* 1325ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1326ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1327ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 1328ef5a22beSAndrea Arcangeli * too. Otherwse drivers using get_user_pages() to access tail 1329ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1330ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1331ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1332ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1333ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1334ef5a22beSAndrea Arcangeli * after get_user_pages(). 1335ef5a22beSAndrea Arcangeli */ 1336ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 133758a84aa9SYouquan Song set_page_count(p, 0); 13381d798ca3SKirill A. Shutemov set_compound_head(p, page); 133920a0307cSWu Fengguang } 1340b4330afbSMike Kravetz atomic_set(compound_mapcount_ptr(page), -1); 134120a0307cSWu Fengguang } 134220a0307cSWu Fengguang 13437795912cSAndrew Morton /* 13447795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 13457795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 13467795912cSAndrew Morton * details. 13477795912cSAndrew Morton */ 134820a0307cSWu Fengguang int PageHuge(struct page *page) 134920a0307cSWu Fengguang { 135020a0307cSWu Fengguang if (!PageCompound(page)) 135120a0307cSWu Fengguang return 0; 135220a0307cSWu Fengguang 135320a0307cSWu Fengguang page = compound_head(page); 1354f1e61557SKirill A. Shutemov return page[1].compound_dtor == HUGETLB_PAGE_DTOR; 135520a0307cSWu Fengguang } 135643131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 135743131e14SNaoya Horiguchi 135827c73ae7SAndrea Arcangeli /* 135927c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 136027c73ae7SAndrea Arcangeli * normal or transparent huge pages. 136127c73ae7SAndrea Arcangeli */ 136227c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 136327c73ae7SAndrea Arcangeli { 136427c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 136527c73ae7SAndrea Arcangeli return 0; 136627c73ae7SAndrea Arcangeli 1367758f66a2SAndrew Morton return get_compound_page_dtor(page_head) == free_huge_page; 136827c73ae7SAndrea Arcangeli } 136927c73ae7SAndrea Arcangeli 137013d60f4bSZhang Yi pgoff_t __basepage_index(struct page *page) 137113d60f4bSZhang Yi { 137213d60f4bSZhang Yi struct page *page_head = compound_head(page); 137313d60f4bSZhang Yi pgoff_t index = page_index(page_head); 137413d60f4bSZhang Yi unsigned long compound_idx; 137513d60f4bSZhang Yi 137613d60f4bSZhang Yi if (!PageHuge(page_head)) 137713d60f4bSZhang Yi return page_index(page); 137813d60f4bSZhang Yi 137913d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 138013d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 138113d60f4bSZhang Yi else 138213d60f4bSZhang Yi compound_idx = page - page_head; 138313d60f4bSZhang Yi 138413d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 138513d60f4bSZhang Yi } 138613d60f4bSZhang Yi 13870c397daeSMichal Hocko static struct page *alloc_buddy_huge_page(struct hstate *h, 1388af0fb9dfSMichal Hocko gfp_t gfp_mask, int nid, nodemask_t *nmask) 13891da177e4SLinus Torvalds { 1390af0fb9dfSMichal Hocko int order = huge_page_order(h); 13911da177e4SLinus Torvalds struct page *page; 1392f96efd58SJoe Jin 1393af0fb9dfSMichal Hocko gfp_mask |= __GFP_COMP|__GFP_RETRY_MAYFAIL|__GFP_NOWARN; 1394af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 1395af0fb9dfSMichal Hocko nid = numa_mem_id(); 1396af0fb9dfSMichal Hocko page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask); 1397af0fb9dfSMichal Hocko if (page) 1398af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC); 1399af0fb9dfSMichal Hocko else 1400af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 140163b4613cSNishanth Aravamudan 140263b4613cSNishanth Aravamudan return page; 140363b4613cSNishanth Aravamudan } 140463b4613cSNishanth Aravamudan 1405af0fb9dfSMichal Hocko /* 14060c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 14070c397daeSMichal Hocko * should use this function to get new hugetlb pages 14080c397daeSMichal Hocko */ 14090c397daeSMichal Hocko static struct page *alloc_fresh_huge_page(struct hstate *h, 14100c397daeSMichal Hocko gfp_t gfp_mask, int nid, nodemask_t *nmask) 14110c397daeSMichal Hocko { 14120c397daeSMichal Hocko struct page *page; 14130c397daeSMichal Hocko 14140c397daeSMichal Hocko if (hstate_is_gigantic(h)) 14150c397daeSMichal Hocko page = alloc_gigantic_page(h, gfp_mask, nid, nmask); 14160c397daeSMichal Hocko else 14170c397daeSMichal Hocko page = alloc_buddy_huge_page(h, gfp_mask, 14180c397daeSMichal Hocko nid, nmask); 14190c397daeSMichal Hocko if (!page) 14200c397daeSMichal Hocko return NULL; 14210c397daeSMichal Hocko 14220c397daeSMichal Hocko if (hstate_is_gigantic(h)) 14230c397daeSMichal Hocko prep_compound_gigantic_page(page, huge_page_order(h)); 14240c397daeSMichal Hocko prep_new_huge_page(h, page, page_to_nid(page)); 14250c397daeSMichal Hocko 14260c397daeSMichal Hocko return page; 14270c397daeSMichal Hocko } 14280c397daeSMichal Hocko 14290c397daeSMichal Hocko /* 1430af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 1431af0fb9dfSMichal Hocko * manner. 1432af0fb9dfSMichal Hocko */ 14330c397daeSMichal Hocko static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 1434b2261026SJoonsoo Kim { 1435b2261026SJoonsoo Kim struct page *page; 1436b2261026SJoonsoo Kim int nr_nodes, node; 1437af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 1438b2261026SJoonsoo Kim 1439b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 14400c397daeSMichal Hocko page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed); 1441af0fb9dfSMichal Hocko if (page) 1442b2261026SJoonsoo Kim break; 1443b2261026SJoonsoo Kim } 1444b2261026SJoonsoo Kim 1445af0fb9dfSMichal Hocko if (!page) 1446af0fb9dfSMichal Hocko return 0; 1447b2261026SJoonsoo Kim 1448af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 1449af0fb9dfSMichal Hocko 1450af0fb9dfSMichal Hocko return 1; 1451b2261026SJoonsoo Kim } 1452b2261026SJoonsoo Kim 1453e8c5c824SLee Schermerhorn /* 1454e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 1455e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 1456e8c5c824SLee Schermerhorn * balanced over allowed nodes. 1457e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 1458e8c5c824SLee Schermerhorn */ 14596ae11b27SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 14606ae11b27SLee Schermerhorn bool acct_surplus) 1461e8c5c824SLee Schermerhorn { 1462b2261026SJoonsoo Kim int nr_nodes, node; 1463e8c5c824SLee Schermerhorn int ret = 0; 1464e8c5c824SLee Schermerhorn 1465b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1466685f3457SLee Schermerhorn /* 1467685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 1468685f3457SLee Schermerhorn * nodes with surplus pages. 1469685f3457SLee Schermerhorn */ 1470b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 1471b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 1472e8c5c824SLee Schermerhorn struct page *page = 1473b2261026SJoonsoo Kim list_entry(h->hugepage_freelists[node].next, 1474e8c5c824SLee Schermerhorn struct page, lru); 1475e8c5c824SLee Schermerhorn list_del(&page->lru); 1476e8c5c824SLee Schermerhorn h->free_huge_pages--; 1477b2261026SJoonsoo Kim h->free_huge_pages_node[node]--; 1478685f3457SLee Schermerhorn if (acct_surplus) { 1479685f3457SLee Schermerhorn h->surplus_huge_pages--; 1480b2261026SJoonsoo Kim h->surplus_huge_pages_node[node]--; 1481685f3457SLee Schermerhorn } 1482e8c5c824SLee Schermerhorn update_and_free_page(h, page); 1483e8c5c824SLee Schermerhorn ret = 1; 14849a76db09SLee Schermerhorn break; 1485e8c5c824SLee Schermerhorn } 1486b2261026SJoonsoo Kim } 1487e8c5c824SLee Schermerhorn 1488e8c5c824SLee Schermerhorn return ret; 1489e8c5c824SLee Schermerhorn } 1490e8c5c824SLee Schermerhorn 1491c8721bbbSNaoya Horiguchi /* 1492c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 1493082d5b6bSGerald Schaefer * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the 14946bc9b564SNaoya Horiguchi * dissolution fails because a give page is not a free hugepage, or because 14956bc9b564SNaoya Horiguchi * free hugepages are fully reserved. 1496c8721bbbSNaoya Horiguchi */ 1497c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 1498c8721bbbSNaoya Horiguchi { 14996bc9b564SNaoya Horiguchi int rc = -EBUSY; 1500082d5b6bSGerald Schaefer 1501c8721bbbSNaoya Horiguchi spin_lock(&hugetlb_lock); 1502c8721bbbSNaoya Horiguchi if (PageHuge(page) && !page_count(page)) { 15032247bb33SGerald Schaefer struct page *head = compound_head(page); 15042247bb33SGerald Schaefer struct hstate *h = page_hstate(head); 15052247bb33SGerald Schaefer int nid = page_to_nid(head); 15066bc9b564SNaoya Horiguchi if (h->free_huge_pages - h->resv_huge_pages == 0) 1507082d5b6bSGerald Schaefer goto out; 1508c3114a84SAnshuman Khandual /* 1509c3114a84SAnshuman Khandual * Move PageHWPoison flag from head page to the raw error page, 1510c3114a84SAnshuman Khandual * which makes any subpages rather than the error page reusable. 1511c3114a84SAnshuman Khandual */ 1512c3114a84SAnshuman Khandual if (PageHWPoison(head) && page != head) { 1513c3114a84SAnshuman Khandual SetPageHWPoison(page); 1514c3114a84SAnshuman Khandual ClearPageHWPoison(head); 1515c3114a84SAnshuman Khandual } 15162247bb33SGerald Schaefer list_del(&head->lru); 1517c8721bbbSNaoya Horiguchi h->free_huge_pages--; 1518c8721bbbSNaoya Horiguchi h->free_huge_pages_node[nid]--; 1519c1470b33Szhong jiang h->max_huge_pages--; 15202247bb33SGerald Schaefer update_and_free_page(h, head); 15216bc9b564SNaoya Horiguchi rc = 0; 1522c8721bbbSNaoya Horiguchi } 1523082d5b6bSGerald Schaefer out: 1524c8721bbbSNaoya Horiguchi spin_unlock(&hugetlb_lock); 1525082d5b6bSGerald Schaefer return rc; 1526c8721bbbSNaoya Horiguchi } 1527c8721bbbSNaoya Horiguchi 1528c8721bbbSNaoya Horiguchi /* 1529c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 1530c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 15312247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 15322247bb33SGerald Schaefer * part of it lies within the given range. 1533082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 1534082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 1535c8721bbbSNaoya Horiguchi */ 1536082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 1537c8721bbbSNaoya Horiguchi { 1538c8721bbbSNaoya Horiguchi unsigned long pfn; 1539eb03aa00SGerald Schaefer struct page *page; 1540082d5b6bSGerald Schaefer int rc = 0; 1541c8721bbbSNaoya Horiguchi 1542d0177639SLi Zhong if (!hugepages_supported()) 1543082d5b6bSGerald Schaefer return rc; 1544d0177639SLi Zhong 1545eb03aa00SGerald Schaefer for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { 1546eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 1547eb03aa00SGerald Schaefer if (PageHuge(page) && !page_count(page)) { 1548eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 1549eb03aa00SGerald Schaefer if (rc) 1550082d5b6bSGerald Schaefer break; 1551eb03aa00SGerald Schaefer } 1552eb03aa00SGerald Schaefer } 1553082d5b6bSGerald Schaefer 1554082d5b6bSGerald Schaefer return rc; 1555c8721bbbSNaoya Horiguchi } 1556c8721bbbSNaoya Horiguchi 1557ab5ac90aSMichal Hocko /* 1558ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 1559ab5ac90aSMichal Hocko */ 15600c397daeSMichal Hocko static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, 1561aaf14e40SMichal Hocko int nid, nodemask_t *nmask) 15627893d1d5SAdam Litke { 15639980d744SMichal Hocko struct page *page = NULL; 15647893d1d5SAdam Litke 1565bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1566aa888a74SAndi Kleen return NULL; 1567aa888a74SAndi Kleen 1568d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 15699980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 15709980d744SMichal Hocko goto out_unlock; 1571d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1572d1c3fb1fSNishanth Aravamudan 15730c397daeSMichal Hocko page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask); 15749980d744SMichal Hocko if (!page) 15750c397daeSMichal Hocko return NULL; 1576d1c3fb1fSNishanth Aravamudan 15777893d1d5SAdam Litke spin_lock(&hugetlb_lock); 15789980d744SMichal Hocko /* 15799980d744SMichal Hocko * We could have raced with the pool size change. 15809980d744SMichal Hocko * Double check that and simply deallocate the new page 15819980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 15829980d744SMichal Hocko * temporary page to workaround the nasty free_huge_page 15839980d744SMichal Hocko * codeflow 15849980d744SMichal Hocko */ 15859980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 15869980d744SMichal Hocko SetPageHugeTemporary(page); 15872bf753e6SKai Shen spin_unlock(&hugetlb_lock); 15889980d744SMichal Hocko put_page(page); 15892bf753e6SKai Shen return NULL; 15909980d744SMichal Hocko } else { 15919980d744SMichal Hocko h->surplus_huge_pages++; 15924704dea3SMichal Hocko h->surplus_huge_pages_node[page_to_nid(page)]++; 15937893d1d5SAdam Litke } 15949980d744SMichal Hocko 15959980d744SMichal Hocko out_unlock: 1596d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 15977893d1d5SAdam Litke 15987893d1d5SAdam Litke return page; 15997893d1d5SAdam Litke } 16007893d1d5SAdam Litke 16019a4e9f3bSAneesh Kumar K.V struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, 1602ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 1603ab5ac90aSMichal Hocko { 1604ab5ac90aSMichal Hocko struct page *page; 1605ab5ac90aSMichal Hocko 1606ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 1607ab5ac90aSMichal Hocko return NULL; 1608ab5ac90aSMichal Hocko 16090c397daeSMichal Hocko page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask); 1610ab5ac90aSMichal Hocko if (!page) 1611ab5ac90aSMichal Hocko return NULL; 1612ab5ac90aSMichal Hocko 1613ab5ac90aSMichal Hocko /* 1614ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 1615ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 1616ab5ac90aSMichal Hocko */ 1617ab5ac90aSMichal Hocko SetPageHugeTemporary(page); 1618ab5ac90aSMichal Hocko 1619ab5ac90aSMichal Hocko return page; 1620ab5ac90aSMichal Hocko } 1621ab5ac90aSMichal Hocko 1622e4e574b7SAdam Litke /* 1623099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 1624099730d6SDave Hansen */ 1625e0ec90eeSDave Hansen static 16260c397daeSMichal Hocko struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, 1627099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 1628099730d6SDave Hansen { 1629aaf14e40SMichal Hocko struct page *page; 1630aaf14e40SMichal Hocko struct mempolicy *mpol; 1631aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 1632aaf14e40SMichal Hocko int nid; 1633aaf14e40SMichal Hocko nodemask_t *nodemask; 1634aaf14e40SMichal Hocko 1635aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 16360c397daeSMichal Hocko page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); 1637aaf14e40SMichal Hocko mpol_cond_put(mpol); 1638aaf14e40SMichal Hocko 1639aaf14e40SMichal Hocko return page; 1640099730d6SDave Hansen } 1641099730d6SDave Hansen 1642ab5ac90aSMichal Hocko /* page migration callback function */ 1643bf50bab2SNaoya Horiguchi struct page *alloc_huge_page_node(struct hstate *h, int nid) 1644bf50bab2SNaoya Horiguchi { 1645aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 16464ef91848SJoonsoo Kim struct page *page = NULL; 1647bf50bab2SNaoya Horiguchi 1648aaf14e40SMichal Hocko if (nid != NUMA_NO_NODE) 1649aaf14e40SMichal Hocko gfp_mask |= __GFP_THISNODE; 1650aaf14e40SMichal Hocko 1651bf50bab2SNaoya Horiguchi spin_lock(&hugetlb_lock); 16524ef91848SJoonsoo Kim if (h->free_huge_pages - h->resv_huge_pages > 0) 16533e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL); 1654bf50bab2SNaoya Horiguchi spin_unlock(&hugetlb_lock); 1655bf50bab2SNaoya Horiguchi 165694ae8ba7SAneesh Kumar K.V if (!page) 16570c397daeSMichal Hocko page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL); 1658bf50bab2SNaoya Horiguchi 1659bf50bab2SNaoya Horiguchi return page; 1660bf50bab2SNaoya Horiguchi } 1661bf50bab2SNaoya Horiguchi 1662ab5ac90aSMichal Hocko /* page migration callback function */ 16633e59fcb0SMichal Hocko struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, 16643e59fcb0SMichal Hocko nodemask_t *nmask) 16654db9b2efSMichal Hocko { 1666aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 16674db9b2efSMichal Hocko 16684db9b2efSMichal Hocko spin_lock(&hugetlb_lock); 16694db9b2efSMichal Hocko if (h->free_huge_pages - h->resv_huge_pages > 0) { 16703e59fcb0SMichal Hocko struct page *page; 16713e59fcb0SMichal Hocko 16723e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); 16733e59fcb0SMichal Hocko if (page) { 16743e59fcb0SMichal Hocko spin_unlock(&hugetlb_lock); 16753e59fcb0SMichal Hocko return page; 16764db9b2efSMichal Hocko } 16774db9b2efSMichal Hocko } 16784db9b2efSMichal Hocko spin_unlock(&hugetlb_lock); 16794db9b2efSMichal Hocko 16800c397daeSMichal Hocko return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); 16814db9b2efSMichal Hocko } 16824db9b2efSMichal Hocko 1683ebd63723SMichal Hocko /* mempolicy aware migration callback */ 1684389c8178SMichal Hocko struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, 1685389c8178SMichal Hocko unsigned long address) 1686ebd63723SMichal Hocko { 1687ebd63723SMichal Hocko struct mempolicy *mpol; 1688ebd63723SMichal Hocko nodemask_t *nodemask; 1689ebd63723SMichal Hocko struct page *page; 1690ebd63723SMichal Hocko gfp_t gfp_mask; 1691ebd63723SMichal Hocko int node; 1692ebd63723SMichal Hocko 1693ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 1694ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 1695ebd63723SMichal Hocko page = alloc_huge_page_nodemask(h, node, nodemask); 1696ebd63723SMichal Hocko mpol_cond_put(mpol); 1697ebd63723SMichal Hocko 1698ebd63723SMichal Hocko return page; 1699ebd63723SMichal Hocko } 1700ebd63723SMichal Hocko 1701bf50bab2SNaoya Horiguchi /* 170225985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 1703e4e574b7SAdam Litke * of size 'delta'. 1704e4e574b7SAdam Litke */ 1705a5516438SAndi Kleen static int gather_surplus_pages(struct hstate *h, int delta) 1706e4e574b7SAdam Litke { 1707e4e574b7SAdam Litke struct list_head surplus_list; 1708e4e574b7SAdam Litke struct page *page, *tmp; 1709e4e574b7SAdam Litke int ret, i; 1710e4e574b7SAdam Litke int needed, allocated; 171128073b02SHillf Danton bool alloc_ok = true; 1712e4e574b7SAdam Litke 1713a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 1714ac09b3a1SAdam Litke if (needed <= 0) { 1715a5516438SAndi Kleen h->resv_huge_pages += delta; 1716e4e574b7SAdam Litke return 0; 1717ac09b3a1SAdam Litke } 1718e4e574b7SAdam Litke 1719e4e574b7SAdam Litke allocated = 0; 1720e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 1721e4e574b7SAdam Litke 1722e4e574b7SAdam Litke ret = -ENOMEM; 1723e4e574b7SAdam Litke retry: 1724e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 1725e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 17260c397daeSMichal Hocko page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), 1727aaf14e40SMichal Hocko NUMA_NO_NODE, NULL); 172828073b02SHillf Danton if (!page) { 172928073b02SHillf Danton alloc_ok = false; 173028073b02SHillf Danton break; 173128073b02SHillf Danton } 1732e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 173369ed779aSDavid Rientjes cond_resched(); 1734e4e574b7SAdam Litke } 173528073b02SHillf Danton allocated += i; 1736e4e574b7SAdam Litke 1737e4e574b7SAdam Litke /* 1738e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 1739e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 1740e4e574b7SAdam Litke */ 1741e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 1742a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 1743a5516438SAndi Kleen (h->free_huge_pages + allocated); 174428073b02SHillf Danton if (needed > 0) { 174528073b02SHillf Danton if (alloc_ok) 1746e4e574b7SAdam Litke goto retry; 174728073b02SHillf Danton /* 174828073b02SHillf Danton * We were not able to allocate enough pages to 174928073b02SHillf Danton * satisfy the entire reservation so we free what 175028073b02SHillf Danton * we've allocated so far. 175128073b02SHillf Danton */ 175228073b02SHillf Danton goto free; 175328073b02SHillf Danton } 1754e4e574b7SAdam Litke /* 1755e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 175625985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 1757e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 1758ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 1759ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 1760ac09b3a1SAdam Litke * before they are reserved. 1761e4e574b7SAdam Litke */ 1762e4e574b7SAdam Litke needed += allocated; 1763a5516438SAndi Kleen h->resv_huge_pages += delta; 1764e4e574b7SAdam Litke ret = 0; 1765a9869b83SNaoya Horiguchi 176619fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 176719fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 176819fc3f0aSAdam Litke if ((--needed) < 0) 176919fc3f0aSAdam Litke break; 1770a9869b83SNaoya Horiguchi /* 1771a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 1772a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 1773a9869b83SNaoya Horiguchi */ 1774a9869b83SNaoya Horiguchi put_page_testzero(page); 1775309381feSSasha Levin VM_BUG_ON_PAGE(page_count(page), page); 1776a5516438SAndi Kleen enqueue_huge_page(h, page); 177719fc3f0aSAdam Litke } 177828073b02SHillf Danton free: 1779b0365c8dSHillf Danton spin_unlock(&hugetlb_lock); 178019fc3f0aSAdam Litke 178119fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 1782c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 1783a9869b83SNaoya Horiguchi put_page(page); 178419fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 1785e4e574b7SAdam Litke 1786e4e574b7SAdam Litke return ret; 1787e4e574b7SAdam Litke } 1788e4e574b7SAdam Litke 1789e4e574b7SAdam Litke /* 1790e5bbc8a6SMike Kravetz * This routine has two main purposes: 1791e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 1792e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 1793e5bbc8a6SMike Kravetz * to the associated reservation map. 1794e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 1795e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 1796e5bbc8a6SMike Kravetz * 1797e5bbc8a6SMike Kravetz * Called with hugetlb_lock held. However, the lock could be dropped (and 1798e5bbc8a6SMike Kravetz * reacquired) during calls to cond_resched_lock. Whenever dropping the lock, 1799e5bbc8a6SMike Kravetz * we must make sure nobody else can claim pages we are in the process of 1800e5bbc8a6SMike Kravetz * freeing. Do this by ensuring resv_huge_page always is greater than the 1801e5bbc8a6SMike Kravetz * number of huge pages we plan to free when dropping the lock. 1802e4e574b7SAdam Litke */ 1803a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 1804a5516438SAndi Kleen unsigned long unused_resv_pages) 1805e4e574b7SAdam Litke { 1806e4e574b7SAdam Litke unsigned long nr_pages; 1807e4e574b7SAdam Litke 1808aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 1809bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1810e5bbc8a6SMike Kravetz goto out; 1811aa888a74SAndi Kleen 1812e5bbc8a6SMike Kravetz /* 1813e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 1814e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 1815e5bbc8a6SMike Kravetz */ 1816a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 1817e4e574b7SAdam Litke 1818685f3457SLee Schermerhorn /* 1819685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 18209b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 18219b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 18229b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 18239b5e5d0fSLee Schermerhorn * free_pool_huge_page() will balance the the freed pages across the 18249b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 1825e5bbc8a6SMike Kravetz * 1826e5bbc8a6SMike Kravetz * Note that we decrement resv_huge_pages as we free the pages. If 1827e5bbc8a6SMike Kravetz * we drop the lock, resv_huge_pages will still be sufficiently large 1828e5bbc8a6SMike Kravetz * to cover subsequent pages we may free. 1829685f3457SLee Schermerhorn */ 1830685f3457SLee Schermerhorn while (nr_pages--) { 1831e5bbc8a6SMike Kravetz h->resv_huge_pages--; 1832e5bbc8a6SMike Kravetz unused_resv_pages--; 18338cebfcd0SLai Jiangshan if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) 1834e5bbc8a6SMike Kravetz goto out; 18357848a4bfSMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 1836e4e574b7SAdam Litke } 1837e5bbc8a6SMike Kravetz 1838e5bbc8a6SMike Kravetz out: 1839e5bbc8a6SMike Kravetz /* Fully uncommit the reservation */ 1840e5bbc8a6SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 1841e4e574b7SAdam Litke } 1842e4e574b7SAdam Litke 18435e911373SMike Kravetz 1844c37f9fb1SAndy Whitcroft /* 1845feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 18465e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 1847cf3ad20bSMike Kravetz * 1848cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 1849cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 1850cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 1851cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 1852cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 1853feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 1854feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 1855feba16e2SMike Kravetz * is called in such cases. 1856cf3ad20bSMike Kravetz * 1857cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 1858cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 1859cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 1860cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 1861cf3ad20bSMike Kravetz * take appropriate action. 186296b96a96SMike Kravetz * 186396b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 186496b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 186596b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 186696b96a96SMike Kravetz * reservation exists. 1867c37f9fb1SAndy Whitcroft */ 18685e911373SMike Kravetz enum vma_resv_mode { 18695e911373SMike Kravetz VMA_NEEDS_RESV, 18705e911373SMike Kravetz VMA_COMMIT_RESV, 1871feba16e2SMike Kravetz VMA_END_RESV, 187296b96a96SMike Kravetz VMA_ADD_RESV, 18735e911373SMike Kravetz }; 1874cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 1875cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 18765e911373SMike Kravetz enum vma_resv_mode mode) 1877c37f9fb1SAndy Whitcroft { 18784e35f483SJoonsoo Kim struct resv_map *resv; 18794e35f483SJoonsoo Kim pgoff_t idx; 1880cf3ad20bSMike Kravetz long ret; 1881c37f9fb1SAndy Whitcroft 18824e35f483SJoonsoo Kim resv = vma_resv_map(vma); 18834e35f483SJoonsoo Kim if (!resv) 1884c37f9fb1SAndy Whitcroft return 1; 1885c37f9fb1SAndy Whitcroft 18864e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 18875e911373SMike Kravetz switch (mode) { 18885e911373SMike Kravetz case VMA_NEEDS_RESV: 1889cf3ad20bSMike Kravetz ret = region_chg(resv, idx, idx + 1); 18905e911373SMike Kravetz break; 18915e911373SMike Kravetz case VMA_COMMIT_RESV: 18925e911373SMike Kravetz ret = region_add(resv, idx, idx + 1); 18935e911373SMike Kravetz break; 1894feba16e2SMike Kravetz case VMA_END_RESV: 18955e911373SMike Kravetz region_abort(resv, idx, idx + 1); 18965e911373SMike Kravetz ret = 0; 18975e911373SMike Kravetz break; 189896b96a96SMike Kravetz case VMA_ADD_RESV: 189996b96a96SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) 190096b96a96SMike Kravetz ret = region_add(resv, idx, idx + 1); 190196b96a96SMike Kravetz else { 190296b96a96SMike Kravetz region_abort(resv, idx, idx + 1); 190396b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 190496b96a96SMike Kravetz } 190596b96a96SMike Kravetz break; 19065e911373SMike Kravetz default: 19075e911373SMike Kravetz BUG(); 19085e911373SMike Kravetz } 190984afd99bSAndy Whitcroft 19104e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) 1911cf3ad20bSMike Kravetz return ret; 191267961f9dSMike Kravetz else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { 191367961f9dSMike Kravetz /* 191467961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 191567961f9dSMike Kravetz * However, a file associated with mapping could have been 191667961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 191767961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 191867961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 191967961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 192067961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 192167961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 192267961f9dSMike Kravetz * reservation has already been consumed. As a result, the 192367961f9dSMike Kravetz * return value of this routine is the opposite of the 192467961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 192567961f9dSMike Kravetz */ 192667961f9dSMike Kravetz if (ret) 192767961f9dSMike Kravetz return 0; 192867961f9dSMike Kravetz else 192967961f9dSMike Kravetz return 1; 193067961f9dSMike Kravetz } 19314e35f483SJoonsoo Kim else 1932cf3ad20bSMike Kravetz return ret < 0 ? ret : 0; 193384afd99bSAndy Whitcroft } 1934cf3ad20bSMike Kravetz 1935cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 1936a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 1937c37f9fb1SAndy Whitcroft { 19385e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 1939cf3ad20bSMike Kravetz } 1940c37f9fb1SAndy Whitcroft 1941cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 1942cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 1943cf3ad20bSMike Kravetz { 19445e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 19455e911373SMike Kravetz } 19465e911373SMike Kravetz 1947feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 19485e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 19495e911373SMike Kravetz { 1950feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 1951c37f9fb1SAndy Whitcroft } 1952c37f9fb1SAndy Whitcroft 195396b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 195496b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 195596b96a96SMike Kravetz { 195696b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 195796b96a96SMike Kravetz } 195896b96a96SMike Kravetz 195996b96a96SMike Kravetz /* 196096b96a96SMike Kravetz * This routine is called to restore a reservation on error paths. In the 196196b96a96SMike Kravetz * specific error paths, a huge page was allocated (via alloc_huge_page) 196296b96a96SMike Kravetz * and is about to be freed. If a reservation for the page existed, 196396b96a96SMike Kravetz * alloc_huge_page would have consumed the reservation and set PagePrivate 196496b96a96SMike Kravetz * in the newly allocated page. When the page is freed via free_huge_page, 196596b96a96SMike Kravetz * the global reservation count will be incremented if PagePrivate is set. 196696b96a96SMike Kravetz * However, free_huge_page can not adjust the reserve map. Adjust the 196796b96a96SMike Kravetz * reserve map here to be consistent with global reserve count adjustments 196896b96a96SMike Kravetz * to be made by free_huge_page. 196996b96a96SMike Kravetz */ 197096b96a96SMike Kravetz static void restore_reserve_on_error(struct hstate *h, 197196b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long address, 197296b96a96SMike Kravetz struct page *page) 197396b96a96SMike Kravetz { 197496b96a96SMike Kravetz if (unlikely(PagePrivate(page))) { 197596b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 197696b96a96SMike Kravetz 197796b96a96SMike Kravetz if (unlikely(rc < 0)) { 197896b96a96SMike Kravetz /* 197996b96a96SMike Kravetz * Rare out of memory condition in reserve map 198096b96a96SMike Kravetz * manipulation. Clear PagePrivate so that 198196b96a96SMike Kravetz * global reserve count will not be incremented 198296b96a96SMike Kravetz * by free_huge_page. This will make it appear 198396b96a96SMike Kravetz * as though the reservation for this page was 198496b96a96SMike Kravetz * consumed. This may prevent the task from 198596b96a96SMike Kravetz * faulting in the page at a later time. This 198696b96a96SMike Kravetz * is better than inconsistent global huge page 198796b96a96SMike Kravetz * accounting of reserve counts. 198896b96a96SMike Kravetz */ 198996b96a96SMike Kravetz ClearPagePrivate(page); 199096b96a96SMike Kravetz } else if (rc) { 199196b96a96SMike Kravetz rc = vma_add_reservation(h, vma, address); 199296b96a96SMike Kravetz if (unlikely(rc < 0)) 199396b96a96SMike Kravetz /* 199496b96a96SMike Kravetz * See above comment about rare out of 199596b96a96SMike Kravetz * memory condition. 199696b96a96SMike Kravetz */ 199796b96a96SMike Kravetz ClearPagePrivate(page); 199896b96a96SMike Kravetz } else 199996b96a96SMike Kravetz vma_end_reservation(h, vma, address); 200096b96a96SMike Kravetz } 200196b96a96SMike Kravetz } 200296b96a96SMike Kravetz 200370c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 200404f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 2005348ea204SAdam Litke { 200690481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 2007a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2008348ea204SAdam Litke struct page *page; 2009d85f69b0SMike Kravetz long map_chg, map_commit; 2010d85f69b0SMike Kravetz long gbl_chg; 20116d76dcf4SAneesh Kumar K.V int ret, idx; 20126d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 20132fc39cecSAdam Litke 20146d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 2015a1e78772SMel Gorman /* 2016d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 2017d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 2018d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 2019a1e78772SMel Gorman */ 2020d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 2021d85f69b0SMike Kravetz if (map_chg < 0) 202276dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 2023d85f69b0SMike Kravetz 2024d85f69b0SMike Kravetz /* 2025d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 2026d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 2027d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 2028d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 2029d85f69b0SMike Kravetz * checked against any subpool limit. 2030d85f69b0SMike Kravetz */ 2031d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 2032d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 2033d85f69b0SMike Kravetz if (gbl_chg < 0) { 2034feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 203576dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 20365e911373SMike Kravetz } 203790d8b7e6SAdam Litke 2038d85f69b0SMike Kravetz /* 2039d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 2040d85f69b0SMike Kravetz * map, there could be reservations associated with the 2041d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 2042d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 2043d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 2044d85f69b0SMike Kravetz * the subpool reservations. 2045d85f69b0SMike Kravetz */ 2046d85f69b0SMike Kravetz if (avoid_reserve) 2047d85f69b0SMike Kravetz gbl_chg = 1; 2048d85f69b0SMike Kravetz } 2049d85f69b0SMike Kravetz 20506d76dcf4SAneesh Kumar K.V ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 20518f34af6fSJianyu Zhan if (ret) 20528f34af6fSJianyu Zhan goto out_subpool_put; 20538f34af6fSJianyu Zhan 2054a1e78772SMel Gorman spin_lock(&hugetlb_lock); 2055d85f69b0SMike Kravetz /* 2056d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 2057d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 2058d85f69b0SMike Kravetz * a reservation exists for the allocation. 2059d85f69b0SMike Kravetz */ 2060d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 206181a6fcaeSJoonsoo Kim if (!page) { 206294ae8ba7SAneesh Kumar K.V spin_unlock(&hugetlb_lock); 20630c397daeSMichal Hocko page = alloc_buddy_huge_page_with_mpol(h, vma, addr); 20648f34af6fSJianyu Zhan if (!page) 20658f34af6fSJianyu Zhan goto out_uncharge_cgroup; 2066a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 2067a88c7695SNaoya Horiguchi SetPagePrivate(page); 2068a88c7695SNaoya Horiguchi h->resv_huge_pages--; 2069a88c7695SNaoya Horiguchi } 207079dbb236SAneesh Kumar K.V spin_lock(&hugetlb_lock); 207179dbb236SAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 207281a6fcaeSJoonsoo Kim /* Fall through */ 2073a1e78772SMel Gorman } 207481a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 207581a6fcaeSJoonsoo Kim spin_unlock(&hugetlb_lock); 2076a1e78772SMel Gorman 207790481622SDavid Gibson set_page_private(page, (unsigned long)spool); 2078a1e78772SMel Gorman 2079d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 2080d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 208133039678SMike Kravetz /* 208233039678SMike Kravetz * The page was added to the reservation map between 208333039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 208433039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 208533039678SMike Kravetz * Adjust for the subpool count incremented above AND 208633039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 208733039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 208833039678SMike Kravetz * no longer applies. 208933039678SMike Kravetz */ 209033039678SMike Kravetz long rsv_adjust; 209133039678SMike Kravetz 209233039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 209333039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 209433039678SMike Kravetz } 20957893d1d5SAdam Litke return page; 20968f34af6fSJianyu Zhan 20978f34af6fSJianyu Zhan out_uncharge_cgroup: 20988f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 20998f34af6fSJianyu Zhan out_subpool_put: 2100d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 21018f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 2102feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 21038f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 2104b45b5bd6SDavid Gibson } 2105b45b5bd6SDavid Gibson 2106e24a1307SAneesh Kumar K.V int alloc_bootmem_huge_page(struct hstate *h) 2107e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 2108e24a1307SAneesh Kumar K.V int __alloc_bootmem_huge_page(struct hstate *h) 2109aa888a74SAndi Kleen { 2110aa888a74SAndi Kleen struct huge_bootmem_page *m; 2111b2261026SJoonsoo Kim int nr_nodes, node; 2112aa888a74SAndi Kleen 2113b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 2114aa888a74SAndi Kleen void *addr; 2115aa888a74SAndi Kleen 2116eb31d559SMike Rapoport addr = memblock_alloc_try_nid_raw( 21178b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 211897ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 2119aa888a74SAndi Kleen if (addr) { 2120aa888a74SAndi Kleen /* 2121aa888a74SAndi Kleen * Use the beginning of the huge page to store the 2122aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 2123aa888a74SAndi Kleen * puts them into the mem_map). 2124aa888a74SAndi Kleen */ 2125aa888a74SAndi Kleen m = addr; 2126aa888a74SAndi Kleen goto found; 2127aa888a74SAndi Kleen } 2128aa888a74SAndi Kleen } 2129aa888a74SAndi Kleen return 0; 2130aa888a74SAndi Kleen 2131aa888a74SAndi Kleen found: 2132df994eadSLuiz Capitulino BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); 2133aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 2134330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 2135aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 2136aa888a74SAndi Kleen m->hstate = h; 2137aa888a74SAndi Kleen return 1; 2138aa888a74SAndi Kleen } 2139aa888a74SAndi Kleen 2140d00181b9SKirill A. Shutemov static void __init prep_compound_huge_page(struct page *page, 2141d00181b9SKirill A. Shutemov unsigned int order) 214218229df5SAndy Whitcroft { 214318229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 214418229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 214518229df5SAndy Whitcroft else 214618229df5SAndy Whitcroft prep_compound_page(page, order); 214718229df5SAndy Whitcroft } 214818229df5SAndy Whitcroft 2149aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 2150aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 2151aa888a74SAndi Kleen { 2152aa888a74SAndi Kleen struct huge_bootmem_page *m; 2153aa888a74SAndi Kleen 2154aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 215540d18ebfSMike Kravetz struct page *page = virt_to_page(m); 2156aa888a74SAndi Kleen struct hstate *h = m->hstate; 2157ee8f248dSBecky Bruce 2158aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 215918229df5SAndy Whitcroft prep_compound_huge_page(page, h->order); 2160ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 2161aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 2162af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 2163af0fb9dfSMichal Hocko 2164b0320c7bSRafael Aquini /* 2165b0320c7bSRafael Aquini * If we had gigantic hugepages allocated at boot time, we need 2166b0320c7bSRafael Aquini * to restore the 'stolen' pages to totalram_pages in order to 2167b0320c7bSRafael Aquini * fix confusing memory reports from free(1) and another 2168b0320c7bSRafael Aquini * side-effects, like CommitLimit going negative. 2169b0320c7bSRafael Aquini */ 2170bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 21713dcc0571SJiang Liu adjust_managed_page_count(page, 1 << h->order); 2172520495feSCannon Matthews cond_resched(); 2173aa888a74SAndi Kleen } 2174aa888a74SAndi Kleen } 2175aa888a74SAndi Kleen 21768faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 21771da177e4SLinus Torvalds { 21781da177e4SLinus Torvalds unsigned long i; 21791da177e4SLinus Torvalds 2180e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 2181bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 2182aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 2183aa888a74SAndi Kleen break; 21840c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 21858cebfcd0SLai Jiangshan &node_states[N_MEMORY])) 21861da177e4SLinus Torvalds break; 218769ed779aSDavid Rientjes cond_resched(); 21881da177e4SLinus Torvalds } 2189d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 2190d715cf80SLiam R. Howlett char buf[32]; 2191d715cf80SLiam R. Howlett 2192c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2193d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 2194d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 21958faa8b07SAndi Kleen h->max_huge_pages = i; 2196e5ff2159SAndi Kleen } 2197d715cf80SLiam R. Howlett } 2198e5ff2159SAndi Kleen 2199e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 2200e5ff2159SAndi Kleen { 2201e5ff2159SAndi Kleen struct hstate *h; 2202e5ff2159SAndi Kleen 2203e5ff2159SAndi Kleen for_each_hstate(h) { 2204641844f5SNaoya Horiguchi if (minimum_order > huge_page_order(h)) 2205641844f5SNaoya Horiguchi minimum_order = huge_page_order(h); 2206641844f5SNaoya Horiguchi 22078faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 2208bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 22098faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 2210e5ff2159SAndi Kleen } 2211641844f5SNaoya Horiguchi VM_BUG_ON(minimum_order == UINT_MAX); 2212e5ff2159SAndi Kleen } 2213e5ff2159SAndi Kleen 2214e5ff2159SAndi Kleen static void __init report_hugepages(void) 2215e5ff2159SAndi Kleen { 2216e5ff2159SAndi Kleen struct hstate *h; 2217e5ff2159SAndi Kleen 2218e5ff2159SAndi Kleen for_each_hstate(h) { 22194abd32dbSAndi Kleen char buf[32]; 2220c6247f72SMatthew Wilcox 2221c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2222ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 2223c6247f72SMatthew Wilcox buf, h->free_huge_pages); 2224e5ff2159SAndi Kleen } 2225e5ff2159SAndi Kleen } 2226e5ff2159SAndi Kleen 22271da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 22286ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 22296ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 22301da177e4SLinus Torvalds { 22314415cc8dSChristoph Lameter int i; 22324415cc8dSChristoph Lameter 2233bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2234aa888a74SAndi Kleen return; 2235aa888a74SAndi Kleen 22366ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 22371da177e4SLinus Torvalds struct page *page, *next; 2238a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 2239a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 2240a5516438SAndi Kleen if (count >= h->nr_huge_pages) 22416b0c880dSAdam Litke return; 22421da177e4SLinus Torvalds if (PageHighMem(page)) 22431da177e4SLinus Torvalds continue; 22441da177e4SLinus Torvalds list_del(&page->lru); 2245e5ff2159SAndi Kleen update_and_free_page(h, page); 2246a5516438SAndi Kleen h->free_huge_pages--; 2247a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 22481da177e4SLinus Torvalds } 22491da177e4SLinus Torvalds } 22501da177e4SLinus Torvalds } 22511da177e4SLinus Torvalds #else 22526ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 22536ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 22541da177e4SLinus Torvalds { 22551da177e4SLinus Torvalds } 22561da177e4SLinus Torvalds #endif 22571da177e4SLinus Torvalds 225820a0307cSWu Fengguang /* 225920a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 226020a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 226120a0307cSWu Fengguang * Returns 1 if an adjustment was made. 226220a0307cSWu Fengguang */ 22636ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 22646ae11b27SLee Schermerhorn int delta) 226520a0307cSWu Fengguang { 2266b2261026SJoonsoo Kim int nr_nodes, node; 226720a0307cSWu Fengguang 226820a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 226920a0307cSWu Fengguang 2270e8c5c824SLee Schermerhorn if (delta < 0) { 2271b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 2272b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 2273b2261026SJoonsoo Kim goto found; 2274b2261026SJoonsoo Kim } 2275b2261026SJoonsoo Kim } else { 2276b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2277b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 2278b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 2279b2261026SJoonsoo Kim goto found; 2280e8c5c824SLee Schermerhorn } 22819a76db09SLee Schermerhorn } 2282b2261026SJoonsoo Kim return 0; 228320a0307cSWu Fengguang 2284b2261026SJoonsoo Kim found: 228520a0307cSWu Fengguang h->surplus_huge_pages += delta; 2286b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 2287b2261026SJoonsoo Kim return 1; 228820a0307cSWu Fengguang } 228920a0307cSWu Fengguang 2290a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 22914eb0716eSAlexandre Ghiti static int set_max_huge_pages(struct hstate *h, unsigned long count, 22926ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 22931da177e4SLinus Torvalds { 22947893d1d5SAdam Litke unsigned long min_count, ret; 22951da177e4SLinus Torvalds 22964eb0716eSAlexandre Ghiti spin_lock(&hugetlb_lock); 22974eb0716eSAlexandre Ghiti 22984eb0716eSAlexandre Ghiti /* 22994eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 23004eb0716eSAlexandre Ghiti * page range allocation. 23014eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 23024eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 23034eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 23044eb0716eSAlexandre Ghiti */ 23054eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 23064eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 23074eb0716eSAlexandre Ghiti spin_unlock(&hugetlb_lock); 23084eb0716eSAlexandre Ghiti return -EINVAL; 23094eb0716eSAlexandre Ghiti } 23104eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 23114eb0716eSAlexandre Ghiti } 2312aa888a74SAndi Kleen 23137893d1d5SAdam Litke /* 23147893d1d5SAdam Litke * Increase the pool size 23157893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 23167893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 2317d1c3fb1fSNishanth Aravamudan * 23180c397daeSMichal Hocko * We might race with alloc_surplus_huge_page() here and be unable 2319d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 2320d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 2321d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 2322d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 23237893d1d5SAdam Litke */ 2324a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 23256ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 23267893d1d5SAdam Litke break; 23277893d1d5SAdam Litke } 23287893d1d5SAdam Litke 2329a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 23307893d1d5SAdam Litke /* 23317893d1d5SAdam Litke * If this allocation races such that we no longer need the 23327893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 23337893d1d5SAdam Litke * and reducing the surplus. 23347893d1d5SAdam Litke */ 23357893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 2336649920c6SJia He 2337649920c6SJia He /* yield cpu to avoid soft lockup */ 2338649920c6SJia He cond_resched(); 2339649920c6SJia He 23400c397daeSMichal Hocko ret = alloc_pool_huge_page(h, nodes_allowed); 23417893d1d5SAdam Litke spin_lock(&hugetlb_lock); 23427893d1d5SAdam Litke if (!ret) 23437893d1d5SAdam Litke goto out; 23447893d1d5SAdam Litke 2345536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 2346536240f2SMel Gorman if (signal_pending(current)) 2347536240f2SMel Gorman goto out; 23487893d1d5SAdam Litke } 23497893d1d5SAdam Litke 23507893d1d5SAdam Litke /* 23517893d1d5SAdam Litke * Decrease the pool size 23527893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 23537893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 23547893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 23557893d1d5SAdam Litke * to the desired size as pages become free. 2356d1c3fb1fSNishanth Aravamudan * 2357d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 2358d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 2359d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 23600c397daeSMichal Hocko * alloc_surplus_huge_page() is checking the global counter, 2361d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 2362d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 2363d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 23647893d1d5SAdam Litke */ 2365a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 23666b0c880dSAdam Litke min_count = max(count, min_count); 23676ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 2368a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 23696ae11b27SLee Schermerhorn if (!free_pool_huge_page(h, nodes_allowed, 0)) 23701da177e4SLinus Torvalds break; 237155f67141SMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 23721da177e4SLinus Torvalds } 2373a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 23746ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 23757893d1d5SAdam Litke break; 23767893d1d5SAdam Litke } 23777893d1d5SAdam Litke out: 23784eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 23791da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 23804eb0716eSAlexandre Ghiti 23814eb0716eSAlexandre Ghiti return 0; 23821da177e4SLinus Torvalds } 23831da177e4SLinus Torvalds 2384a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 2385a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 2386a3437870SNishanth Aravamudan 2387a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 2388a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 2389a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 2390a3437870SNishanth Aravamudan 2391a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 2392a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 2393a3437870SNishanth Aravamudan 23949a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 23959a305230SLee Schermerhorn 23969a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 2397a3437870SNishanth Aravamudan { 2398a3437870SNishanth Aravamudan int i; 23999a305230SLee Schermerhorn 2400a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 24019a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 24029a305230SLee Schermerhorn if (nidp) 24039a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 2404a3437870SNishanth Aravamudan return &hstates[i]; 24059a305230SLee Schermerhorn } 24069a305230SLee Schermerhorn 24079a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 2408a3437870SNishanth Aravamudan } 2409a3437870SNishanth Aravamudan 241006808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 2411a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2412a3437870SNishanth Aravamudan { 24139a305230SLee Schermerhorn struct hstate *h; 24149a305230SLee Schermerhorn unsigned long nr_huge_pages; 24159a305230SLee Schermerhorn int nid; 24169a305230SLee Schermerhorn 24179a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 24189a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 24199a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 24209a305230SLee Schermerhorn else 24219a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 24229a305230SLee Schermerhorn 24239a305230SLee Schermerhorn return sprintf(buf, "%lu\n", nr_huge_pages); 2424a3437870SNishanth Aravamudan } 2425adbe8726SEric B Munson 2426238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 2427238d3c13SDavid Rientjes struct hstate *h, int nid, 2428238d3c13SDavid Rientjes unsigned long count, size_t len) 2429a3437870SNishanth Aravamudan { 2430a3437870SNishanth Aravamudan int err; 2431bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); 2432a3437870SNishanth Aravamudan 24334eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) { 2434adbe8726SEric B Munson err = -EINVAL; 2435adbe8726SEric B Munson goto out; 2436adbe8726SEric B Munson } 2437adbe8726SEric B Munson 24389a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 24399a305230SLee Schermerhorn /* 24409a305230SLee Schermerhorn * global hstate attribute 24419a305230SLee Schermerhorn */ 24429a305230SLee Schermerhorn if (!(obey_mempolicy && 24439a305230SLee Schermerhorn init_nodemask_of_mempolicy(nodes_allowed))) { 244406808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 24458cebfcd0SLai Jiangshan nodes_allowed = &node_states[N_MEMORY]; 244606808b08SLee Schermerhorn } 24479a305230SLee Schermerhorn } else if (nodes_allowed) { 24489a305230SLee Schermerhorn /* 24499a305230SLee Schermerhorn * per node hstate attribute: adjust count to global, 24509a305230SLee Schermerhorn * but restrict alloc/free to the specified node. 24519a305230SLee Schermerhorn */ 24529a305230SLee Schermerhorn count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 24539a305230SLee Schermerhorn init_nodemask_of_node(nodes_allowed, nid); 24549a305230SLee Schermerhorn } else 24558cebfcd0SLai Jiangshan nodes_allowed = &node_states[N_MEMORY]; 24569a305230SLee Schermerhorn 24574eb0716eSAlexandre Ghiti err = set_max_huge_pages(h, count, nodes_allowed); 2458a3437870SNishanth Aravamudan 24594eb0716eSAlexandre Ghiti out: 24608cebfcd0SLai Jiangshan if (nodes_allowed != &node_states[N_MEMORY]) 246106808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 246206808b08SLee Schermerhorn 24634eb0716eSAlexandre Ghiti return err ? err : len; 246406808b08SLee Schermerhorn } 246506808b08SLee Schermerhorn 2466238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 2467238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 2468238d3c13SDavid Rientjes size_t len) 2469238d3c13SDavid Rientjes { 2470238d3c13SDavid Rientjes struct hstate *h; 2471238d3c13SDavid Rientjes unsigned long count; 2472238d3c13SDavid Rientjes int nid; 2473238d3c13SDavid Rientjes int err; 2474238d3c13SDavid Rientjes 2475238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 2476238d3c13SDavid Rientjes if (err) 2477238d3c13SDavid Rientjes return err; 2478238d3c13SDavid Rientjes 2479238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 2480238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 2481238d3c13SDavid Rientjes } 2482238d3c13SDavid Rientjes 248306808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 248406808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 248506808b08SLee Schermerhorn { 248606808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 248706808b08SLee Schermerhorn } 248806808b08SLee Schermerhorn 248906808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 249006808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 249106808b08SLee Schermerhorn { 2492238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 2493a3437870SNishanth Aravamudan } 2494a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 2495a3437870SNishanth Aravamudan 249606808b08SLee Schermerhorn #ifdef CONFIG_NUMA 249706808b08SLee Schermerhorn 249806808b08SLee Schermerhorn /* 249906808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 250006808b08SLee Schermerhorn * huge page alloc/free. 250106808b08SLee Schermerhorn */ 250206808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 250306808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 250406808b08SLee Schermerhorn { 250506808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 250606808b08SLee Schermerhorn } 250706808b08SLee Schermerhorn 250806808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 250906808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 251006808b08SLee Schermerhorn { 2511238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 251206808b08SLee Schermerhorn } 251306808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 251406808b08SLee Schermerhorn #endif 251506808b08SLee Schermerhorn 251606808b08SLee Schermerhorn 2517a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 2518a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2519a3437870SNishanth Aravamudan { 25209a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2521a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 2522a3437870SNishanth Aravamudan } 2523adbe8726SEric B Munson 2524a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 2525a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 2526a3437870SNishanth Aravamudan { 2527a3437870SNishanth Aravamudan int err; 2528a3437870SNishanth Aravamudan unsigned long input; 25299a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2530a3437870SNishanth Aravamudan 2531bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2532adbe8726SEric B Munson return -EINVAL; 2533adbe8726SEric B Munson 25343dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 2535a3437870SNishanth Aravamudan if (err) 253673ae31e5SEric B Munson return err; 2537a3437870SNishanth Aravamudan 2538a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 2539a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 2540a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 2541a3437870SNishanth Aravamudan 2542a3437870SNishanth Aravamudan return count; 2543a3437870SNishanth Aravamudan } 2544a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 2545a3437870SNishanth Aravamudan 2546a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 2547a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2548a3437870SNishanth Aravamudan { 25499a305230SLee Schermerhorn struct hstate *h; 25509a305230SLee Schermerhorn unsigned long free_huge_pages; 25519a305230SLee Schermerhorn int nid; 25529a305230SLee Schermerhorn 25539a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 25549a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 25559a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 25569a305230SLee Schermerhorn else 25579a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 25589a305230SLee Schermerhorn 25599a305230SLee Schermerhorn return sprintf(buf, "%lu\n", free_huge_pages); 2560a3437870SNishanth Aravamudan } 2561a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 2562a3437870SNishanth Aravamudan 2563a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 2564a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2565a3437870SNishanth Aravamudan { 25669a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2567a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->resv_huge_pages); 2568a3437870SNishanth Aravamudan } 2569a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 2570a3437870SNishanth Aravamudan 2571a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 2572a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2573a3437870SNishanth Aravamudan { 25749a305230SLee Schermerhorn struct hstate *h; 25759a305230SLee Schermerhorn unsigned long surplus_huge_pages; 25769a305230SLee Schermerhorn int nid; 25779a305230SLee Schermerhorn 25789a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 25799a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 25809a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 25819a305230SLee Schermerhorn else 25829a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 25839a305230SLee Schermerhorn 25849a305230SLee Schermerhorn return sprintf(buf, "%lu\n", surplus_huge_pages); 2585a3437870SNishanth Aravamudan } 2586a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 2587a3437870SNishanth Aravamudan 2588a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 2589a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 2590a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 2591a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 2592a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 2593a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 259406808b08SLee Schermerhorn #ifdef CONFIG_NUMA 259506808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 259606808b08SLee Schermerhorn #endif 2597a3437870SNishanth Aravamudan NULL, 2598a3437870SNishanth Aravamudan }; 2599a3437870SNishanth Aravamudan 260067e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 2601a3437870SNishanth Aravamudan .attrs = hstate_attrs, 2602a3437870SNishanth Aravamudan }; 2603a3437870SNishanth Aravamudan 2604094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 26059a305230SLee Schermerhorn struct kobject **hstate_kobjs, 260667e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 2607a3437870SNishanth Aravamudan { 2608a3437870SNishanth Aravamudan int retval; 2609972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 2610a3437870SNishanth Aravamudan 26119a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 26129a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 2613a3437870SNishanth Aravamudan return -ENOMEM; 2614a3437870SNishanth Aravamudan 26159a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 2616a3437870SNishanth Aravamudan if (retval) 26179a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 2618a3437870SNishanth Aravamudan 2619a3437870SNishanth Aravamudan return retval; 2620a3437870SNishanth Aravamudan } 2621a3437870SNishanth Aravamudan 2622a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 2623a3437870SNishanth Aravamudan { 2624a3437870SNishanth Aravamudan struct hstate *h; 2625a3437870SNishanth Aravamudan int err; 2626a3437870SNishanth Aravamudan 2627a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 2628a3437870SNishanth Aravamudan if (!hugepages_kobj) 2629a3437870SNishanth Aravamudan return; 2630a3437870SNishanth Aravamudan 2631a3437870SNishanth Aravamudan for_each_hstate(h) { 26329a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 26339a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 2634a3437870SNishanth Aravamudan if (err) 2635ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s", h->name); 2636a3437870SNishanth Aravamudan } 2637a3437870SNishanth Aravamudan } 2638a3437870SNishanth Aravamudan 26399a305230SLee Schermerhorn #ifdef CONFIG_NUMA 26409a305230SLee Schermerhorn 26419a305230SLee Schermerhorn /* 26429a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 264310fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 264410fbcf4cSKay Sievers * index of a node device or _hstate == node id. 264510fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 26469a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 26479a305230SLee Schermerhorn */ 26489a305230SLee Schermerhorn struct node_hstate { 26499a305230SLee Schermerhorn struct kobject *hugepages_kobj; 26509a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 26519a305230SLee Schermerhorn }; 2652b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 26539a305230SLee Schermerhorn 26549a305230SLee Schermerhorn /* 265510fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 26569a305230SLee Schermerhorn */ 26579a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 26589a305230SLee Schermerhorn &nr_hugepages_attr.attr, 26599a305230SLee Schermerhorn &free_hugepages_attr.attr, 26609a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 26619a305230SLee Schermerhorn NULL, 26629a305230SLee Schermerhorn }; 26639a305230SLee Schermerhorn 266467e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 26659a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 26669a305230SLee Schermerhorn }; 26679a305230SLee Schermerhorn 26689a305230SLee Schermerhorn /* 266910fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 26709a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 26719a305230SLee Schermerhorn */ 26729a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 26739a305230SLee Schermerhorn { 26749a305230SLee Schermerhorn int nid; 26759a305230SLee Schermerhorn 26769a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 26779a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 26789a305230SLee Schermerhorn int i; 26799a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 26809a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 26819a305230SLee Schermerhorn if (nidp) 26829a305230SLee Schermerhorn *nidp = nid; 26839a305230SLee Schermerhorn return &hstates[i]; 26849a305230SLee Schermerhorn } 26859a305230SLee Schermerhorn } 26869a305230SLee Schermerhorn 26879a305230SLee Schermerhorn BUG(); 26889a305230SLee Schermerhorn return NULL; 26899a305230SLee Schermerhorn } 26909a305230SLee Schermerhorn 26919a305230SLee Schermerhorn /* 269210fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 26939a305230SLee Schermerhorn * No-op if no hstate attributes attached. 26949a305230SLee Schermerhorn */ 26953cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 26969a305230SLee Schermerhorn { 26979a305230SLee Schermerhorn struct hstate *h; 269810fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 26999a305230SLee Schermerhorn 27009a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 27019b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 27029a305230SLee Schermerhorn 2703972dc4deSAneesh Kumar K.V for_each_hstate(h) { 2704972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 2705972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 2706972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 2707972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 2708972dc4deSAneesh Kumar K.V } 27099a305230SLee Schermerhorn } 27109a305230SLee Schermerhorn 27119a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 27129a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 27139a305230SLee Schermerhorn } 27149a305230SLee Schermerhorn 27159a305230SLee Schermerhorn 27169a305230SLee Schermerhorn /* 271710fbcf4cSKay Sievers * Register hstate attributes for a single node device. 27189a305230SLee Schermerhorn * No-op if attributes already registered. 27199a305230SLee Schermerhorn */ 27203cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 27219a305230SLee Schermerhorn { 27229a305230SLee Schermerhorn struct hstate *h; 272310fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 27249a305230SLee Schermerhorn int err; 27259a305230SLee Schermerhorn 27269a305230SLee Schermerhorn if (nhs->hugepages_kobj) 27279a305230SLee Schermerhorn return; /* already allocated */ 27289a305230SLee Schermerhorn 27299a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 273010fbcf4cSKay Sievers &node->dev.kobj); 27319a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 27329a305230SLee Schermerhorn return; 27339a305230SLee Schermerhorn 27349a305230SLee Schermerhorn for_each_hstate(h) { 27359a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 27369a305230SLee Schermerhorn nhs->hstate_kobjs, 27379a305230SLee Schermerhorn &per_node_hstate_attr_group); 27389a305230SLee Schermerhorn if (err) { 2739ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s for node %d\n", 274010fbcf4cSKay Sievers h->name, node->dev.id); 27419a305230SLee Schermerhorn hugetlb_unregister_node(node); 27429a305230SLee Schermerhorn break; 27439a305230SLee Schermerhorn } 27449a305230SLee Schermerhorn } 27459a305230SLee Schermerhorn } 27469a305230SLee Schermerhorn 27479a305230SLee Schermerhorn /* 27489b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 274910fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 275010fbcf4cSKay Sievers * registered their associated device by this time. 27519a305230SLee Schermerhorn */ 27527d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 27539a305230SLee Schermerhorn { 27549a305230SLee Schermerhorn int nid; 27559a305230SLee Schermerhorn 27568cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 27578732794bSWen Congyang struct node *node = node_devices[nid]; 275810fbcf4cSKay Sievers if (node->dev.id == nid) 27599a305230SLee Schermerhorn hugetlb_register_node(node); 27609a305230SLee Schermerhorn } 27619a305230SLee Schermerhorn 27629a305230SLee Schermerhorn /* 276310fbcf4cSKay Sievers * Let the node device driver know we're here so it can 27649a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 27659a305230SLee Schermerhorn */ 27669a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 27679a305230SLee Schermerhorn hugetlb_unregister_node); 27689a305230SLee Schermerhorn } 27699a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 27709a305230SLee Schermerhorn 27719a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 27729a305230SLee Schermerhorn { 27739a305230SLee Schermerhorn BUG(); 27749a305230SLee Schermerhorn if (nidp) 27759a305230SLee Schermerhorn *nidp = -1; 27769a305230SLee Schermerhorn return NULL; 27779a305230SLee Schermerhorn } 27789a305230SLee Schermerhorn 27799a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 27809a305230SLee Schermerhorn 27819a305230SLee Schermerhorn #endif 27829a305230SLee Schermerhorn 2783a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 2784a3437870SNishanth Aravamudan { 27858382d914SDavidlohr Bueso int i; 27868382d914SDavidlohr Bueso 2787457c1b27SNishanth Aravamudan if (!hugepages_supported()) 27880ef89d25SBenjamin Herrenschmidt return 0; 2789a3437870SNishanth Aravamudan 2790e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) { 2791d715cf80SLiam R. Howlett if (default_hstate_size != 0) { 2792d715cf80SLiam R. Howlett pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n", 2793d715cf80SLiam R. Howlett default_hstate_size, HPAGE_SIZE); 2794d715cf80SLiam R. Howlett } 2795d715cf80SLiam R. Howlett 2796e11bfbfcSNick Piggin default_hstate_size = HPAGE_SIZE; 2797e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) 2798a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 2799a3437870SNishanth Aravamudan } 2800972dc4deSAneesh Kumar K.V default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); 2801f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 2802f8b74815SVaishali Thakkar if (!default_hstate.max_huge_pages) 2803e11bfbfcSNick Piggin default_hstate.max_huge_pages = default_hstate_max_huge_pages; 2804f8b74815SVaishali Thakkar } 2805a3437870SNishanth Aravamudan 2806a3437870SNishanth Aravamudan hugetlb_init_hstates(); 2807aa888a74SAndi Kleen gather_bootmem_prealloc(); 2808a3437870SNishanth Aravamudan report_hugepages(); 2809a3437870SNishanth Aravamudan 2810a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 28119a305230SLee Schermerhorn hugetlb_register_all_nodes(); 28127179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 28139a305230SLee Schermerhorn 28148382d914SDavidlohr Bueso #ifdef CONFIG_SMP 28158382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 28168382d914SDavidlohr Bueso #else 28178382d914SDavidlohr Bueso num_fault_mutexes = 1; 28188382d914SDavidlohr Bueso #endif 2819c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 28206da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 28216da2ec56SKees Cook GFP_KERNEL); 2822c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 28238382d914SDavidlohr Bueso 28248382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 2825c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 2826a3437870SNishanth Aravamudan return 0; 2827a3437870SNishanth Aravamudan } 28283e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 2829a3437870SNishanth Aravamudan 2830a3437870SNishanth Aravamudan /* Should be called on processing a hugepagesz=... option */ 28319fee021dSVaishali Thakkar void __init hugetlb_bad_size(void) 28329fee021dSVaishali Thakkar { 28339fee021dSVaishali Thakkar parsed_valid_hugepagesz = false; 28349fee021dSVaishali Thakkar } 28359fee021dSVaishali Thakkar 2836d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 2837a3437870SNishanth Aravamudan { 2838a3437870SNishanth Aravamudan struct hstate *h; 28398faa8b07SAndi Kleen unsigned long i; 28408faa8b07SAndi Kleen 2841a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 2842598d8091SJoe Perches pr_warn("hugepagesz= specified twice, ignoring\n"); 2843a3437870SNishanth Aravamudan return; 2844a3437870SNishanth Aravamudan } 284547d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 2846a3437870SNishanth Aravamudan BUG_ON(order == 0); 284747d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 2848a3437870SNishanth Aravamudan h->order = order; 2849a3437870SNishanth Aravamudan h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 28508faa8b07SAndi Kleen h->nr_huge_pages = 0; 28518faa8b07SAndi Kleen h->free_huge_pages = 0; 28528faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 28538faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 28540edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 285554f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 285654f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 2857a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 2858a3437870SNishanth Aravamudan huge_page_size(h)/1024); 28598faa8b07SAndi Kleen 2860a3437870SNishanth Aravamudan parsed_hstate = h; 2861a3437870SNishanth Aravamudan } 2862a3437870SNishanth Aravamudan 2863e11bfbfcSNick Piggin static int __init hugetlb_nrpages_setup(char *s) 2864a3437870SNishanth Aravamudan { 2865a3437870SNishanth Aravamudan unsigned long *mhp; 28668faa8b07SAndi Kleen static unsigned long *last_mhp; 2867a3437870SNishanth Aravamudan 28689fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 28699fee021dSVaishali Thakkar pr_warn("hugepages = %s preceded by " 28709fee021dSVaishali Thakkar "an unsupported hugepagesz, ignoring\n", s); 28719fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 28729fee021dSVaishali Thakkar return 1; 28739fee021dSVaishali Thakkar } 2874a3437870SNishanth Aravamudan /* 287547d38344SAneesh Kumar K.V * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, 2876a3437870SNishanth Aravamudan * so this hugepages= parameter goes to the "default hstate". 2877a3437870SNishanth Aravamudan */ 28789fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 2879a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 2880a3437870SNishanth Aravamudan else 2881a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 2882a3437870SNishanth Aravamudan 28838faa8b07SAndi Kleen if (mhp == last_mhp) { 2884598d8091SJoe Perches pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n"); 28858faa8b07SAndi Kleen return 1; 28868faa8b07SAndi Kleen } 28878faa8b07SAndi Kleen 2888a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 2889a3437870SNishanth Aravamudan *mhp = 0; 2890a3437870SNishanth Aravamudan 28918faa8b07SAndi Kleen /* 28928faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 28938faa8b07SAndi Kleen * But we need to allocate >= MAX_ORDER hstates here early to still 28948faa8b07SAndi Kleen * use the bootmem allocator. 28958faa8b07SAndi Kleen */ 289647d38344SAneesh Kumar K.V if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) 28978faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 28988faa8b07SAndi Kleen 28998faa8b07SAndi Kleen last_mhp = mhp; 29008faa8b07SAndi Kleen 2901a3437870SNishanth Aravamudan return 1; 2902a3437870SNishanth Aravamudan } 2903e11bfbfcSNick Piggin __setup("hugepages=", hugetlb_nrpages_setup); 2904e11bfbfcSNick Piggin 2905e11bfbfcSNick Piggin static int __init hugetlb_default_setup(char *s) 2906e11bfbfcSNick Piggin { 2907e11bfbfcSNick Piggin default_hstate_size = memparse(s, &s); 2908e11bfbfcSNick Piggin return 1; 2909e11bfbfcSNick Piggin } 2910e11bfbfcSNick Piggin __setup("default_hugepagesz=", hugetlb_default_setup); 2911a3437870SNishanth Aravamudan 29128a213460SNishanth Aravamudan static unsigned int cpuset_mems_nr(unsigned int *array) 29138a213460SNishanth Aravamudan { 29148a213460SNishanth Aravamudan int node; 29158a213460SNishanth Aravamudan unsigned int nr = 0; 29168a213460SNishanth Aravamudan 29178a213460SNishanth Aravamudan for_each_node_mask(node, cpuset_current_mems_allowed) 29188a213460SNishanth Aravamudan nr += array[node]; 29198a213460SNishanth Aravamudan 29208a213460SNishanth Aravamudan return nr; 29218a213460SNishanth Aravamudan } 29228a213460SNishanth Aravamudan 29238a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 292406808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 292506808b08SLee Schermerhorn struct ctl_table *table, int write, 292606808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 29271da177e4SLinus Torvalds { 2928e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 2929238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 293008d4a246SMichal Hocko int ret; 2931e5ff2159SAndi Kleen 2932457c1b27SNishanth Aravamudan if (!hugepages_supported()) 293386613628SJan Stancek return -EOPNOTSUPP; 2934457c1b27SNishanth Aravamudan 2935e5ff2159SAndi Kleen table->data = &tmp; 2936e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 293708d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 293808d4a246SMichal Hocko if (ret) 293908d4a246SMichal Hocko goto out; 2940e5ff2159SAndi Kleen 2941238d3c13SDavid Rientjes if (write) 2942238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 2943238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 294408d4a246SMichal Hocko out: 294508d4a246SMichal Hocko return ret; 29461da177e4SLinus Torvalds } 2947396faf03SMel Gorman 294806808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 294906808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 295006808b08SLee Schermerhorn { 295106808b08SLee Schermerhorn 295206808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 295306808b08SLee Schermerhorn buffer, length, ppos); 295406808b08SLee Schermerhorn } 295506808b08SLee Schermerhorn 295606808b08SLee Schermerhorn #ifdef CONFIG_NUMA 295706808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 295806808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 295906808b08SLee Schermerhorn { 296006808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 296106808b08SLee Schermerhorn buffer, length, ppos); 296206808b08SLee Schermerhorn } 296306808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 296406808b08SLee Schermerhorn 2965a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 29668d65af78SAlexey Dobriyan void __user *buffer, 2967a3d0c6aaSNishanth Aravamudan size_t *length, loff_t *ppos) 2968a3d0c6aaSNishanth Aravamudan { 2969a5516438SAndi Kleen struct hstate *h = &default_hstate; 2970e5ff2159SAndi Kleen unsigned long tmp; 297108d4a246SMichal Hocko int ret; 2972e5ff2159SAndi Kleen 2973457c1b27SNishanth Aravamudan if (!hugepages_supported()) 297486613628SJan Stancek return -EOPNOTSUPP; 2975457c1b27SNishanth Aravamudan 2976e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 2977e5ff2159SAndi Kleen 2978bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 2979adbe8726SEric B Munson return -EINVAL; 2980adbe8726SEric B Munson 2981e5ff2159SAndi Kleen table->data = &tmp; 2982e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 298308d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 298408d4a246SMichal Hocko if (ret) 298508d4a246SMichal Hocko goto out; 2986e5ff2159SAndi Kleen 2987e5ff2159SAndi Kleen if (write) { 2988064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 2989e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 2990a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 2991e5ff2159SAndi Kleen } 299208d4a246SMichal Hocko out: 299308d4a246SMichal Hocko return ret; 2994a3d0c6aaSNishanth Aravamudan } 2995a3d0c6aaSNishanth Aravamudan 29961da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 29971da177e4SLinus Torvalds 2998e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 29991da177e4SLinus Torvalds { 3000fcb2b0c5SRoman Gushchin struct hstate *h; 3001fcb2b0c5SRoman Gushchin unsigned long total = 0; 3002fcb2b0c5SRoman Gushchin 3003457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3004457c1b27SNishanth Aravamudan return; 3005fcb2b0c5SRoman Gushchin 3006fcb2b0c5SRoman Gushchin for_each_hstate(h) { 3007fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 3008fcb2b0c5SRoman Gushchin 3009fcb2b0c5SRoman Gushchin total += (PAGE_SIZE << huge_page_order(h)) * count; 3010fcb2b0c5SRoman Gushchin 3011fcb2b0c5SRoman Gushchin if (h == &default_hstate) 3012e1759c21SAlexey Dobriyan seq_printf(m, 30131da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 30141da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 3015b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 30167893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 30174f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 3018fcb2b0c5SRoman Gushchin count, 3019a5516438SAndi Kleen h->free_huge_pages, 3020a5516438SAndi Kleen h->resv_huge_pages, 3021a5516438SAndi Kleen h->surplus_huge_pages, 3022fcb2b0c5SRoman Gushchin (PAGE_SIZE << huge_page_order(h)) / 1024); 3023fcb2b0c5SRoman Gushchin } 3024fcb2b0c5SRoman Gushchin 3025fcb2b0c5SRoman Gushchin seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024); 30261da177e4SLinus Torvalds } 30271da177e4SLinus Torvalds 30281da177e4SLinus Torvalds int hugetlb_report_node_meminfo(int nid, char *buf) 30291da177e4SLinus Torvalds { 3030a5516438SAndi Kleen struct hstate *h = &default_hstate; 3031457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3032457c1b27SNishanth Aravamudan return 0; 30331da177e4SLinus Torvalds return sprintf(buf, 30341da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 3035a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 3036a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 3037a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 3038a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 3039a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 30401da177e4SLinus Torvalds } 30411da177e4SLinus Torvalds 3042949f7ec5SDavid Rientjes void hugetlb_show_meminfo(void) 3043949f7ec5SDavid Rientjes { 3044949f7ec5SDavid Rientjes struct hstate *h; 3045949f7ec5SDavid Rientjes int nid; 3046949f7ec5SDavid Rientjes 3047457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3048457c1b27SNishanth Aravamudan return; 3049457c1b27SNishanth Aravamudan 3050949f7ec5SDavid Rientjes for_each_node_state(nid, N_MEMORY) 3051949f7ec5SDavid Rientjes for_each_hstate(h) 3052949f7ec5SDavid Rientjes pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 3053949f7ec5SDavid Rientjes nid, 3054949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 3055949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 3056949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 3057949f7ec5SDavid Rientjes 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 3058949f7ec5SDavid Rientjes } 3059949f7ec5SDavid Rientjes 30605d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 30615d317b2bSNaoya Horiguchi { 30625d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 30635d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 30645d317b2bSNaoya Horiguchi } 30655d317b2bSNaoya Horiguchi 30661da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 30671da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 30681da177e4SLinus Torvalds { 3069d0028588SWanpeng Li struct hstate *h; 3070d0028588SWanpeng Li unsigned long nr_total_pages = 0; 3071d0028588SWanpeng Li 3072d0028588SWanpeng Li for_each_hstate(h) 3073d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 3074d0028588SWanpeng Li return nr_total_pages; 30751da177e4SLinus Torvalds } 30761da177e4SLinus Torvalds 3077a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 3078fc1b8a73SMel Gorman { 3079fc1b8a73SMel Gorman int ret = -ENOMEM; 3080fc1b8a73SMel Gorman 3081fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 3082fc1b8a73SMel Gorman /* 3083fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 3084fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 3085fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 3086fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 3087fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 3088fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 3089fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 3090fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 3091fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 3092fc1b8a73SMel Gorman * 3093fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 3094fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 3095fc1b8a73SMel Gorman * we fall back to check against current free page availability as 3096fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 3097fc1b8a73SMel Gorman * semantics that cpuset has. 3098fc1b8a73SMel Gorman */ 3099fc1b8a73SMel Gorman if (delta > 0) { 3100a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 3101fc1b8a73SMel Gorman goto out; 3102fc1b8a73SMel Gorman 3103a5516438SAndi Kleen if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 3104a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 3105fc1b8a73SMel Gorman goto out; 3106fc1b8a73SMel Gorman } 3107fc1b8a73SMel Gorman } 3108fc1b8a73SMel Gorman 3109fc1b8a73SMel Gorman ret = 0; 3110fc1b8a73SMel Gorman if (delta < 0) 3111a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 3112fc1b8a73SMel Gorman 3113fc1b8a73SMel Gorman out: 3114fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 3115fc1b8a73SMel Gorman return ret; 3116fc1b8a73SMel Gorman } 3117fc1b8a73SMel Gorman 311884afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 311984afd99bSAndy Whitcroft { 3120f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 312184afd99bSAndy Whitcroft 312284afd99bSAndy Whitcroft /* 312384afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 312484afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 312584afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 312625985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 312784afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 312884afd99bSAndy Whitcroft * new reference here without additional locking. 312984afd99bSAndy Whitcroft */ 31304e35f483SJoonsoo Kim if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 3131f522c3acSJoonsoo Kim kref_get(&resv->refs); 313284afd99bSAndy Whitcroft } 313384afd99bSAndy Whitcroft 3134a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 3135a1e78772SMel Gorman { 3136a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3137f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 313890481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 31394e35f483SJoonsoo Kim unsigned long reserve, start, end; 31401c5ecae3SMike Kravetz long gbl_reserve; 314184afd99bSAndy Whitcroft 31424e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 31434e35f483SJoonsoo Kim return; 31444e35f483SJoonsoo Kim 3145a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 3146a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 314784afd99bSAndy Whitcroft 31484e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 314984afd99bSAndy Whitcroft 3150f031dd27SJoonsoo Kim kref_put(&resv->refs, resv_map_release); 315184afd99bSAndy Whitcroft 31527251ff78SAdam Litke if (reserve) { 31531c5ecae3SMike Kravetz /* 31541c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 31551c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 31561c5ecae3SMike Kravetz */ 31571c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 31581c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 31597251ff78SAdam Litke } 3160a1e78772SMel Gorman } 3161a1e78772SMel Gorman 316231383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 316331383c68SDan Williams { 316431383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 316531383c68SDan Williams return -EINVAL; 316631383c68SDan Williams return 0; 316731383c68SDan Williams } 316831383c68SDan Williams 316905ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 317005ea8860SDan Williams { 317105ea8860SDan Williams struct hstate *hstate = hstate_vma(vma); 317205ea8860SDan Williams 317305ea8860SDan Williams return 1UL << huge_page_shift(hstate); 317405ea8860SDan Williams } 317505ea8860SDan Williams 31761da177e4SLinus Torvalds /* 31771da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 31781da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 31791da177e4SLinus Torvalds * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 31801da177e4SLinus Torvalds * this far. 31811da177e4SLinus Torvalds */ 3182b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 31831da177e4SLinus Torvalds { 31841da177e4SLinus Torvalds BUG(); 3185d0217ac0SNick Piggin return 0; 31861da177e4SLinus Torvalds } 31871da177e4SLinus Torvalds 3188eec3636aSJane Chu /* 3189eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 3190eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 3191eec3636aSJane Chu * This is because under System V memory model, mappings created via 3192eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 3193eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 3194eec3636aSJane Chu */ 3195f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 3196d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 319784afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 3198a1e78772SMel Gorman .close = hugetlb_vm_op_close, 319931383c68SDan Williams .split = hugetlb_vm_op_split, 320005ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 32011da177e4SLinus Torvalds }; 32021da177e4SLinus Torvalds 32031e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 32041e8f889bSDavid Gibson int writable) 320563551ae0SDavid Gibson { 320663551ae0SDavid Gibson pte_t entry; 320763551ae0SDavid Gibson 32081e8f889bSDavid Gibson if (writable) { 3209106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 3210106c992aSGerald Schaefer vma->vm_page_prot))); 321163551ae0SDavid Gibson } else { 3212106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 3213106c992aSGerald Schaefer vma->vm_page_prot)); 321463551ae0SDavid Gibson } 321563551ae0SDavid Gibson entry = pte_mkyoung(entry); 321663551ae0SDavid Gibson entry = pte_mkhuge(entry); 3217d9ed9faaSChris Metcalf entry = arch_make_huge_pte(entry, vma, page, writable); 321863551ae0SDavid Gibson 321963551ae0SDavid Gibson return entry; 322063551ae0SDavid Gibson } 322163551ae0SDavid Gibson 32221e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 32231e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 32241e8f889bSDavid Gibson { 32251e8f889bSDavid Gibson pte_t entry; 32261e8f889bSDavid Gibson 3227106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 322832f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 32294b3073e1SRussell King update_mmu_cache(vma, address, ptep); 32301e8f889bSDavid Gibson } 32311e8f889bSDavid Gibson 3232d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 32334a705fefSNaoya Horiguchi { 32344a705fefSNaoya Horiguchi swp_entry_t swp; 32354a705fefSNaoya Horiguchi 32364a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 3237d5ed7444SAneesh Kumar K.V return false; 32384a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 32394a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_migration_entry(swp)) 3240d5ed7444SAneesh Kumar K.V return true; 32414a705fefSNaoya Horiguchi else 3242d5ed7444SAneesh Kumar K.V return false; 32434a705fefSNaoya Horiguchi } 32444a705fefSNaoya Horiguchi 32454a705fefSNaoya Horiguchi static int is_hugetlb_entry_hwpoisoned(pte_t pte) 32464a705fefSNaoya Horiguchi { 32474a705fefSNaoya Horiguchi swp_entry_t swp; 32484a705fefSNaoya Horiguchi 32494a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 32504a705fefSNaoya Horiguchi return 0; 32514a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 32524a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_hwpoison_entry(swp)) 32534a705fefSNaoya Horiguchi return 1; 32544a705fefSNaoya Horiguchi else 32554a705fefSNaoya Horiguchi return 0; 32564a705fefSNaoya Horiguchi } 32571e8f889bSDavid Gibson 325863551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 325963551ae0SDavid Gibson struct vm_area_struct *vma) 326063551ae0SDavid Gibson { 32615e41540cSMike Kravetz pte_t *src_pte, *dst_pte, entry, dst_entry; 326263551ae0SDavid Gibson struct page *ptepage; 32631c59827dSHugh Dickins unsigned long addr; 32641e8f889bSDavid Gibson int cow; 3265a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3266a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 3267ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 3268e8569dd2SAndreas Sandberg int ret = 0; 32691e8f889bSDavid Gibson 32701e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 327163551ae0SDavid Gibson 3272ac46d4f3SJérôme Glisse if (cow) { 3273ac46d4f3SJérôme Glisse mmu_notifier_range_init(&range, src, vma->vm_start, 3274ac46d4f3SJérôme Glisse vma->vm_end); 3275ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 3276ac46d4f3SJérôme Glisse } 3277e8569dd2SAndreas Sandberg 3278a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 3279cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 32807868a208SPunit Agrawal src_pte = huge_pte_offset(src, addr, sz); 3281c74df32cSHugh Dickins if (!src_pte) 3282c74df32cSHugh Dickins continue; 3283a5516438SAndi Kleen dst_pte = huge_pte_alloc(dst, addr, sz); 3284e8569dd2SAndreas Sandberg if (!dst_pte) { 3285e8569dd2SAndreas Sandberg ret = -ENOMEM; 3286e8569dd2SAndreas Sandberg break; 3287e8569dd2SAndreas Sandberg } 3288c5c99429SLarry Woodman 32895e41540cSMike Kravetz /* 32905e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 32915e41540cSMike Kravetz * dst_pte == src_pte is the common case of src/dest sharing. 32925e41540cSMike Kravetz * 32935e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 32945e41540cSMike Kravetz * another vma. If dst_pte !none, this implies sharing. 32955e41540cSMike Kravetz * Check here before taking page table lock, and once again 32965e41540cSMike Kravetz * after taking the lock below. 32975e41540cSMike Kravetz */ 32985e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 32995e41540cSMike Kravetz if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) 3300c5c99429SLarry Woodman continue; 3301c5c99429SLarry Woodman 3302cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 3303cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 3304cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 33054a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 33065e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 33075e41540cSMike Kravetz if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { 33085e41540cSMike Kravetz /* 33095e41540cSMike Kravetz * Skip if src entry none. Also, skip in the 33105e41540cSMike Kravetz * unlikely case dst entry !none as this implies 33115e41540cSMike Kravetz * sharing with another vma. 33125e41540cSMike Kravetz */ 33134a705fefSNaoya Horiguchi ; 33144a705fefSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry) || 33154a705fefSNaoya Horiguchi is_hugetlb_entry_hwpoisoned(entry))) { 33164a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 33174a705fefSNaoya Horiguchi 33184a705fefSNaoya Horiguchi if (is_write_migration_entry(swp_entry) && cow) { 33194a705fefSNaoya Horiguchi /* 33204a705fefSNaoya Horiguchi * COW mappings require pages in both 33214a705fefSNaoya Horiguchi * parent and child to be set to read. 33224a705fefSNaoya Horiguchi */ 33234a705fefSNaoya Horiguchi make_migration_entry_read(&swp_entry); 33244a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 3325e5251fd4SPunit Agrawal set_huge_swap_pte_at(src, addr, src_pte, 3326e5251fd4SPunit Agrawal entry, sz); 33274a705fefSNaoya Horiguchi } 3328e5251fd4SPunit Agrawal set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); 33294a705fefSNaoya Horiguchi } else { 333034ee645eSJoerg Roedel if (cow) { 33310f10851eSJérôme Glisse /* 33320f10851eSJérôme Glisse * No need to notify as we are downgrading page 33330f10851eSJérôme Glisse * table protection not changing it to point 33340f10851eSJérôme Glisse * to a new page. 33350f10851eSJérôme Glisse * 3336ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 33370f10851eSJérôme Glisse */ 33387f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 333934ee645eSJoerg Roedel } 33400253d634SNaoya Horiguchi entry = huge_ptep_get(src_pte); 334163551ae0SDavid Gibson ptepage = pte_page(entry); 334263551ae0SDavid Gibson get_page(ptepage); 334353f9263bSKirill A. Shutemov page_dup_rmap(ptepage, true); 334463551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 33455d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), dst); 33461c59827dSHugh Dickins } 3347cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 3348cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 334963551ae0SDavid Gibson } 335063551ae0SDavid Gibson 3351e8569dd2SAndreas Sandberg if (cow) 3352ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 3353e8569dd2SAndreas Sandberg 3354e8569dd2SAndreas Sandberg return ret; 335563551ae0SDavid Gibson } 335663551ae0SDavid Gibson 335724669e58SAneesh Kumar K.V void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 335824669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 335924669e58SAneesh Kumar K.V struct page *ref_page) 336063551ae0SDavid Gibson { 336163551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 336263551ae0SDavid Gibson unsigned long address; 3363c7546f8fSDavid Gibson pte_t *ptep; 336463551ae0SDavid Gibson pte_t pte; 3365cb900f41SKirill A. Shutemov spinlock_t *ptl; 336663551ae0SDavid Gibson struct page *page; 3367a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3368a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 3369ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 3370a5516438SAndi Kleen 337163551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 3372a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 3373a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 337463551ae0SDavid Gibson 337507e32661SAneesh Kumar K.V /* 337607e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 337707e32661SAneesh Kumar K.V * to huge page. 337807e32661SAneesh Kumar K.V */ 3379ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 338024669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 3381dff11abeSMike Kravetz 3382dff11abeSMike Kravetz /* 3383dff11abeSMike Kravetz * If sharing possible, alert mmu notifiers of worst case. 3384dff11abeSMike Kravetz */ 3385ac46d4f3SJérôme Glisse mmu_notifier_range_init(&range, mm, start, end); 3386ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 3387ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 3388569f48b8SHillf Danton address = start; 3389569f48b8SHillf Danton for (; address < end; address += sz) { 33907868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, sz); 3391c7546f8fSDavid Gibson if (!ptep) 3392c7546f8fSDavid Gibson continue; 3393c7546f8fSDavid Gibson 3394cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 339531d49da5SAneesh Kumar K.V if (huge_pmd_unshare(mm, &address, ptep)) { 339631d49da5SAneesh Kumar K.V spin_unlock(ptl); 3397dff11abeSMike Kravetz /* 3398dff11abeSMike Kravetz * We just unmapped a page of PMDs by clearing a PUD. 3399dff11abeSMike Kravetz * The caller's TLB flush range should cover this area. 3400dff11abeSMike Kravetz */ 340131d49da5SAneesh Kumar K.V continue; 340231d49da5SAneesh Kumar K.V } 340339dde65cSChen, Kenneth W 34046629326bSHillf Danton pte = huge_ptep_get(ptep); 340531d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 340631d49da5SAneesh Kumar K.V spin_unlock(ptl); 340731d49da5SAneesh Kumar K.V continue; 340831d49da5SAneesh Kumar K.V } 34096629326bSHillf Danton 34106629326bSHillf Danton /* 34119fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 34129fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 34136629326bSHillf Danton */ 34149fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 34159386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 341631d49da5SAneesh Kumar K.V spin_unlock(ptl); 341731d49da5SAneesh Kumar K.V continue; 34188c4894c6SNaoya Horiguchi } 34196629326bSHillf Danton 34206629326bSHillf Danton page = pte_page(pte); 342104f2cbe3SMel Gorman /* 342204f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 342304f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 342404f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 342504f2cbe3SMel Gorman */ 342604f2cbe3SMel Gorman if (ref_page) { 342731d49da5SAneesh Kumar K.V if (page != ref_page) { 342831d49da5SAneesh Kumar K.V spin_unlock(ptl); 342931d49da5SAneesh Kumar K.V continue; 343031d49da5SAneesh Kumar K.V } 343104f2cbe3SMel Gorman /* 343204f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 343304f2cbe3SMel Gorman * future faults in this VMA will fail rather than 343404f2cbe3SMel Gorman * looking like data was lost 343504f2cbe3SMel Gorman */ 343604f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 343704f2cbe3SMel Gorman } 343804f2cbe3SMel Gorman 3439c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 3440b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 3441106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 34426649a386SKen Chen set_page_dirty(page); 34439e81130bSHillf Danton 34445d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 3445d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 344631d49da5SAneesh Kumar K.V 3447cb900f41SKirill A. Shutemov spin_unlock(ptl); 3448e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 344924669e58SAneesh Kumar K.V /* 345031d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 345124669e58SAneesh Kumar K.V */ 345231d49da5SAneesh Kumar K.V if (ref_page) 345331d49da5SAneesh Kumar K.V break; 3454fe1668aeSChen, Kenneth W } 3455ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 345624669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 34571da177e4SLinus Torvalds } 345863551ae0SDavid Gibson 3459d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 3460d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 3461d833352aSMel Gorman unsigned long end, struct page *ref_page) 3462d833352aSMel Gorman { 3463d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 3464d833352aSMel Gorman 3465d833352aSMel Gorman /* 3466d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 3467d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 3468d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 3469d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 3470d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 3471c8c06efaSDavidlohr Bueso * is to clear it before releasing the i_mmap_rwsem. This works 3472d833352aSMel Gorman * because in the context this is called, the VMA is about to be 3473c8c06efaSDavidlohr Bueso * destroyed and the i_mmap_rwsem is held. 3474d833352aSMel Gorman */ 3475d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 3476d833352aSMel Gorman } 3477d833352aSMel Gorman 3478502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 347904f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 3480502717f4SChen, Kenneth W { 348124669e58SAneesh Kumar K.V struct mm_struct *mm; 348224669e58SAneesh Kumar K.V struct mmu_gather tlb; 3483dff11abeSMike Kravetz unsigned long tlb_start = start; 3484dff11abeSMike Kravetz unsigned long tlb_end = end; 3485dff11abeSMike Kravetz 3486dff11abeSMike Kravetz /* 3487dff11abeSMike Kravetz * If shared PMDs were possibly used within this vma range, adjust 3488dff11abeSMike Kravetz * start/end for worst case tlb flushing. 3489dff11abeSMike Kravetz * Note that we can not be sure if PMDs are shared until we try to 3490dff11abeSMike Kravetz * unmap pages. However, we want to make sure TLB flushing covers 3491dff11abeSMike Kravetz * the largest possible range. 3492dff11abeSMike Kravetz */ 3493dff11abeSMike Kravetz adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end); 349424669e58SAneesh Kumar K.V 349524669e58SAneesh Kumar K.V mm = vma->vm_mm; 349624669e58SAneesh Kumar K.V 3497dff11abeSMike Kravetz tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end); 349824669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 3499dff11abeSMike Kravetz tlb_finish_mmu(&tlb, tlb_start, tlb_end); 3500502717f4SChen, Kenneth W } 3501502717f4SChen, Kenneth W 350204f2cbe3SMel Gorman /* 350304f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 350404f2cbe3SMel Gorman * mappping it owns the reserve page for. The intention is to unmap the page 350504f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 350604f2cbe3SMel Gorman * same region. 350704f2cbe3SMel Gorman */ 35082f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 35092a4b3dedSHarvey Harrison struct page *page, unsigned long address) 351004f2cbe3SMel Gorman { 35117526674dSAdam Litke struct hstate *h = hstate_vma(vma); 351204f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 351304f2cbe3SMel Gorman struct address_space *mapping; 351404f2cbe3SMel Gorman pgoff_t pgoff; 351504f2cbe3SMel Gorman 351604f2cbe3SMel Gorman /* 351704f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 351804f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 351904f2cbe3SMel Gorman */ 35207526674dSAdam Litke address = address & huge_page_mask(h); 352136e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 352236e4f20aSMichal Hocko vma->vm_pgoff; 352393c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 352404f2cbe3SMel Gorman 35254eb2b1dcSMel Gorman /* 35264eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 35274eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 35284eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 35294eb2b1dcSMel Gorman */ 353083cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 35316b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 353204f2cbe3SMel Gorman /* Do not unmap the current VMA */ 353304f2cbe3SMel Gorman if (iter_vma == vma) 353404f2cbe3SMel Gorman continue; 353504f2cbe3SMel Gorman 353604f2cbe3SMel Gorman /* 35372f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 35382f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 35392f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 35402f84a899SMel Gorman */ 35412f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 35422f84a899SMel Gorman continue; 35432f84a899SMel Gorman 35442f84a899SMel Gorman /* 354504f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 354604f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 354704f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 354804f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 354904f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 355004f2cbe3SMel Gorman */ 355104f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 355224669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 355324669e58SAneesh Kumar K.V address + huge_page_size(h), page); 355404f2cbe3SMel Gorman } 355583cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 355604f2cbe3SMel Gorman } 355704f2cbe3SMel Gorman 35580fe6e20bSNaoya Horiguchi /* 35590fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 3560ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 3561ef009b25SMichal Hocko * cannot race with other handlers or page migration. 3562ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 35630fe6e20bSNaoya Horiguchi */ 35642b740303SSouptick Joarder static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 3565974e6d66SHuang Ying unsigned long address, pte_t *ptep, 3566cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 35671e8f889bSDavid Gibson { 35683999f52eSAneesh Kumar K.V pte_t pte; 3569a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 35701e8f889bSDavid Gibson struct page *old_page, *new_page; 35712b740303SSouptick Joarder int outside_reserve = 0; 35722b740303SSouptick Joarder vm_fault_t ret = 0; 3573974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 3574ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 35751e8f889bSDavid Gibson 35763999f52eSAneesh Kumar K.V pte = huge_ptep_get(ptep); 35771e8f889bSDavid Gibson old_page = pte_page(pte); 35781e8f889bSDavid Gibson 357904f2cbe3SMel Gorman retry_avoidcopy: 35801e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 35811e8f889bSDavid Gibson * and just make the page writable */ 358237a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 35835a49973dSHugh Dickins page_move_anon_rmap(old_page, vma); 35845b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 358583c54070SNick Piggin return 0; 35861e8f889bSDavid Gibson } 35871e8f889bSDavid Gibson 358804f2cbe3SMel Gorman /* 358904f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 359004f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 359104f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 359204f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 359304f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 359404f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 359504f2cbe3SMel Gorman * of the full address range. 359604f2cbe3SMel Gorman */ 35975944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 359804f2cbe3SMel Gorman old_page != pagecache_page) 359904f2cbe3SMel Gorman outside_reserve = 1; 360004f2cbe3SMel Gorman 360109cbfeafSKirill A. Shutemov get_page(old_page); 3602b76c8cfbSLarry Woodman 3603ad4404a2SDavidlohr Bueso /* 3604ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 3605ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 3606ad4404a2SDavidlohr Bueso */ 3607cb900f41SKirill A. Shutemov spin_unlock(ptl); 36085b7a1d40SHuang Ying new_page = alloc_huge_page(vma, haddr, outside_reserve); 36091e8f889bSDavid Gibson 36102fc39cecSAdam Litke if (IS_ERR(new_page)) { 361104f2cbe3SMel Gorman /* 361204f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 361304f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 361404f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 361504f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 361604f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 361704f2cbe3SMel Gorman */ 361804f2cbe3SMel Gorman if (outside_reserve) { 361909cbfeafSKirill A. Shutemov put_page(old_page); 362004f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 36215b7a1d40SHuang Ying unmap_ref_private(mm, vma, old_page, haddr); 362204f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 3623cb900f41SKirill A. Shutemov spin_lock(ptl); 36245b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 3625a9af0c5dSNaoya Horiguchi if (likely(ptep && 3626a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 362704f2cbe3SMel Gorman goto retry_avoidcopy; 3628a734bcc8SHillf Danton /* 3629cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 3630cb900f41SKirill A. Shutemov * lock, and our job is done. 3631a734bcc8SHillf Danton */ 3632a734bcc8SHillf Danton return 0; 363304f2cbe3SMel Gorman } 363404f2cbe3SMel Gorman 36352b740303SSouptick Joarder ret = vmf_error(PTR_ERR(new_page)); 3636ad4404a2SDavidlohr Bueso goto out_release_old; 36371e8f889bSDavid Gibson } 36381e8f889bSDavid Gibson 36390fe6e20bSNaoya Horiguchi /* 36400fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 36410fe6e20bSNaoya Horiguchi * anon_vma prepared. 36420fe6e20bSNaoya Horiguchi */ 364344e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 3644ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 3645ad4404a2SDavidlohr Bueso goto out_release_all; 364644e2aa93SDean Nelson } 36470fe6e20bSNaoya Horiguchi 3648974e6d66SHuang Ying copy_user_huge_page(new_page, old_page, address, vma, 364947ad8475SAndrea Arcangeli pages_per_huge_page(h)); 36500ed361deSNick Piggin __SetPageUptodate(new_page); 36511e8f889bSDavid Gibson 3652ac46d4f3SJérôme Glisse mmu_notifier_range_init(&range, mm, haddr, haddr + huge_page_size(h)); 3653ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 3654ad4404a2SDavidlohr Bueso 3655b76c8cfbSLarry Woodman /* 3656cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 3657b76c8cfbSLarry Woodman * before the page tables are altered 3658b76c8cfbSLarry Woodman */ 3659cb900f41SKirill A. Shutemov spin_lock(ptl); 36605b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 3661a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 366207443a85SJoonsoo Kim ClearPagePrivate(new_page); 366307443a85SJoonsoo Kim 36641e8f889bSDavid Gibson /* Break COW */ 36655b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 3666ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, range.end); 36675b7a1d40SHuang Ying set_huge_pte_at(mm, haddr, ptep, 36681e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 3669d281ee61SKirill A. Shutemov page_remove_rmap(old_page, true); 36705b7a1d40SHuang Ying hugepage_add_new_anon_rmap(new_page, vma, haddr); 3671cb6acd01SMike Kravetz set_page_huge_active(new_page); 36721e8f889bSDavid Gibson /* Make the old page be freed below */ 36731e8f889bSDavid Gibson new_page = old_page; 36741e8f889bSDavid Gibson } 3675cb900f41SKirill A. Shutemov spin_unlock(ptl); 3676ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 3677ad4404a2SDavidlohr Bueso out_release_all: 36785b7a1d40SHuang Ying restore_reserve_on_error(h, vma, haddr, new_page); 367909cbfeafSKirill A. Shutemov put_page(new_page); 3680ad4404a2SDavidlohr Bueso out_release_old: 368109cbfeafSKirill A. Shutemov put_page(old_page); 36828312034fSJoonsoo Kim 3683ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 3684ad4404a2SDavidlohr Bueso return ret; 36851e8f889bSDavid Gibson } 36861e8f889bSDavid Gibson 368704f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 3688a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 3689a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 369004f2cbe3SMel Gorman { 369104f2cbe3SMel Gorman struct address_space *mapping; 3692e7c4b0bfSAndy Whitcroft pgoff_t idx; 369304f2cbe3SMel Gorman 369404f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 3695a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 369604f2cbe3SMel Gorman 369704f2cbe3SMel Gorman return find_lock_page(mapping, idx); 369804f2cbe3SMel Gorman } 369904f2cbe3SMel Gorman 37003ae77f43SHugh Dickins /* 37013ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 37023ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 37033ae77f43SHugh Dickins */ 37043ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 37052a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 37062a15efc9SHugh Dickins { 37072a15efc9SHugh Dickins struct address_space *mapping; 37082a15efc9SHugh Dickins pgoff_t idx; 37092a15efc9SHugh Dickins struct page *page; 37102a15efc9SHugh Dickins 37112a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 37122a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 37132a15efc9SHugh Dickins 37142a15efc9SHugh Dickins page = find_get_page(mapping, idx); 37152a15efc9SHugh Dickins if (page) 37162a15efc9SHugh Dickins put_page(page); 37172a15efc9SHugh Dickins return page != NULL; 37182a15efc9SHugh Dickins } 37192a15efc9SHugh Dickins 3720ab76ad54SMike Kravetz int huge_add_to_page_cache(struct page *page, struct address_space *mapping, 3721ab76ad54SMike Kravetz pgoff_t idx) 3722ab76ad54SMike Kravetz { 3723ab76ad54SMike Kravetz struct inode *inode = mapping->host; 3724ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 3725ab76ad54SMike Kravetz int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 3726ab76ad54SMike Kravetz 3727ab76ad54SMike Kravetz if (err) 3728ab76ad54SMike Kravetz return err; 3729ab76ad54SMike Kravetz ClearPagePrivate(page); 3730ab76ad54SMike Kravetz 373122146c3cSMike Kravetz /* 373222146c3cSMike Kravetz * set page dirty so that it will not be removed from cache/file 373322146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 373422146c3cSMike Kravetz */ 373522146c3cSMike Kravetz set_page_dirty(page); 373622146c3cSMike Kravetz 3737ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 3738ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 3739ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 3740ab76ad54SMike Kravetz return 0; 3741ab76ad54SMike Kravetz } 3742ab76ad54SMike Kravetz 37432b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 37442b740303SSouptick Joarder struct vm_area_struct *vma, 37458382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 3746788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 3747ac9b9c66SHugh Dickins { 3748a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 37492b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 3750409eb8c2SHillf Danton int anon_rmap = 0; 37514c887265SAdam Litke unsigned long size; 37524c887265SAdam Litke struct page *page; 37531e8f889bSDavid Gibson pte_t new_pte; 3754cb900f41SKirill A. Shutemov spinlock_t *ptl; 3755285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 3756cb6acd01SMike Kravetz bool new_page = false; 37574c887265SAdam Litke 375804f2cbe3SMel Gorman /* 375904f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 376004f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 376125985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 376204f2cbe3SMel Gorman */ 376304f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 3764910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 376504f2cbe3SMel Gorman current->pid); 376604f2cbe3SMel Gorman return ret; 376704f2cbe3SMel Gorman } 376804f2cbe3SMel Gorman 37694c887265SAdam Litke /* 3770e7c58097SMike Kravetz * Use page lock to guard against racing truncation 3771e7c58097SMike Kravetz * before we get page_table_lock. 37724c887265SAdam Litke */ 3773e7c58097SMike Kravetz retry: 3774e7c58097SMike Kravetz page = find_lock_page(mapping, idx); 3775e7c58097SMike Kravetz if (!page) { 3776a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 3777ebed4bfcSHugh Dickins if (idx >= size) 3778ebed4bfcSHugh Dickins goto out; 37791a1aad8aSMike Kravetz 37801a1aad8aSMike Kravetz /* 37811a1aad8aSMike Kravetz * Check for page in userfault range 37821a1aad8aSMike Kravetz */ 37831a1aad8aSMike Kravetz if (userfaultfd_missing(vma)) { 37841a1aad8aSMike Kravetz u32 hash; 37851a1aad8aSMike Kravetz struct vm_fault vmf = { 37861a1aad8aSMike Kravetz .vma = vma, 3787285b8dcaSHuang Ying .address = haddr, 37881a1aad8aSMike Kravetz .flags = flags, 37891a1aad8aSMike Kravetz /* 37901a1aad8aSMike Kravetz * Hard to debug if it ends up being 37911a1aad8aSMike Kravetz * used by a callee that assumes 37921a1aad8aSMike Kravetz * something about the other 37931a1aad8aSMike Kravetz * uninitialized fields... same as in 37941a1aad8aSMike Kravetz * memory.c 37951a1aad8aSMike Kravetz */ 37961a1aad8aSMike Kravetz }; 37971a1aad8aSMike Kravetz 37981a1aad8aSMike Kravetz /* 3799ddeaab32SMike Kravetz * hugetlb_fault_mutex must be dropped before 3800ddeaab32SMike Kravetz * handling userfault. Reacquire after handling 3801ddeaab32SMike Kravetz * fault to make calling code simpler. 38021a1aad8aSMike Kravetz */ 38031a1aad8aSMike Kravetz hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, 3804285b8dcaSHuang Ying idx, haddr); 38051a1aad8aSMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 38061a1aad8aSMike Kravetz ret = handle_userfault(&vmf, VM_UFFD_MISSING); 38071a1aad8aSMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 38081a1aad8aSMike Kravetz goto out; 38091a1aad8aSMike Kravetz } 38101a1aad8aSMike Kravetz 3811285b8dcaSHuang Ying page = alloc_huge_page(vma, haddr, 0); 38122fc39cecSAdam Litke if (IS_ERR(page)) { 38132b740303SSouptick Joarder ret = vmf_error(PTR_ERR(page)); 38146bda666aSChristoph Lameter goto out; 38156bda666aSChristoph Lameter } 381647ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 38170ed361deSNick Piggin __SetPageUptodate(page); 3818cb6acd01SMike Kravetz new_page = true; 3819ac9b9c66SHugh Dickins 3820f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 3821ab76ad54SMike Kravetz int err = huge_add_to_page_cache(page, mapping, idx); 38226bda666aSChristoph Lameter if (err) { 38236bda666aSChristoph Lameter put_page(page); 38246bda666aSChristoph Lameter if (err == -EEXIST) 38256bda666aSChristoph Lameter goto retry; 38266bda666aSChristoph Lameter goto out; 38276bda666aSChristoph Lameter } 382823be7468SMel Gorman } else { 38296bda666aSChristoph Lameter lock_page(page); 38300fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 38310fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 38320fe6e20bSNaoya Horiguchi goto backout_unlocked; 383323be7468SMel Gorman } 3834409eb8c2SHillf Danton anon_rmap = 1; 38350fe6e20bSNaoya Horiguchi } 38360fe6e20bSNaoya Horiguchi } else { 383757303d80SAndy Whitcroft /* 3838998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 3839998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 3840998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 3841fd6a03edSNaoya Horiguchi */ 3842fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 3843aa50d3a7SAndi Kleen ret = VM_FAULT_HWPOISON | 3844972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 3845fd6a03edSNaoya Horiguchi goto backout_unlocked; 38466bda666aSChristoph Lameter } 3847998b4382SNaoya Horiguchi } 38481e8f889bSDavid Gibson 384957303d80SAndy Whitcroft /* 385057303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 385157303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 385257303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 385357303d80SAndy Whitcroft * the spinlock. 385457303d80SAndy Whitcroft */ 38555e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 3856285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 38572b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 38582b26736cSAndy Whitcroft goto backout_unlocked; 38592b26736cSAndy Whitcroft } 38605e911373SMike Kravetz /* Just decrements count, does not deallocate */ 3861285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 38625e911373SMike Kravetz } 386357303d80SAndy Whitcroft 38648bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 3865e7c58097SMike Kravetz size = i_size_read(mapping->host) >> huge_page_shift(h); 3866e7c58097SMike Kravetz if (idx >= size) 3867e7c58097SMike Kravetz goto backout; 38684c887265SAdam Litke 386983c54070SNick Piggin ret = 0; 38707f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 38714c887265SAdam Litke goto backout; 38724c887265SAdam Litke 387307443a85SJoonsoo Kim if (anon_rmap) { 387407443a85SJoonsoo Kim ClearPagePrivate(page); 3875285b8dcaSHuang Ying hugepage_add_new_anon_rmap(page, vma, haddr); 3876ac714904SChoi Gi-yong } else 387753f9263bSKirill A. Shutemov page_dup_rmap(page, true); 38781e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 38791e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 3880285b8dcaSHuang Ying set_huge_pte_at(mm, haddr, ptep, new_pte); 38811e8f889bSDavid Gibson 38825d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 3883788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 38841e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 3885974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); 38861e8f889bSDavid Gibson } 38871e8f889bSDavid Gibson 3888cb900f41SKirill A. Shutemov spin_unlock(ptl); 3889cb6acd01SMike Kravetz 3890cb6acd01SMike Kravetz /* 3891cb6acd01SMike Kravetz * Only make newly allocated pages active. Existing pages found 3892cb6acd01SMike Kravetz * in the pagecache could be !page_huge_active() if they have been 3893cb6acd01SMike Kravetz * isolated for migration. 3894cb6acd01SMike Kravetz */ 3895cb6acd01SMike Kravetz if (new_page) 3896cb6acd01SMike Kravetz set_page_huge_active(page); 3897cb6acd01SMike Kravetz 38984c887265SAdam Litke unlock_page(page); 38994c887265SAdam Litke out: 3900ac9b9c66SHugh Dickins return ret; 39014c887265SAdam Litke 39024c887265SAdam Litke backout: 3903cb900f41SKirill A. Shutemov spin_unlock(ptl); 39042b26736cSAndy Whitcroft backout_unlocked: 39054c887265SAdam Litke unlock_page(page); 3906285b8dcaSHuang Ying restore_reserve_on_error(h, vma, haddr, page); 39074c887265SAdam Litke put_page(page); 39084c887265SAdam Litke goto out; 3909ac9b9c66SHugh Dickins } 3910ac9b9c66SHugh Dickins 39118382d914SDavidlohr Bueso #ifdef CONFIG_SMP 3912c672c7f2SMike Kravetz u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, 39138382d914SDavidlohr Bueso struct vm_area_struct *vma, 39148382d914SDavidlohr Bueso struct address_space *mapping, 39158382d914SDavidlohr Bueso pgoff_t idx, unsigned long address) 39168382d914SDavidlohr Bueso { 39178382d914SDavidlohr Bueso unsigned long key[2]; 39188382d914SDavidlohr Bueso u32 hash; 39198382d914SDavidlohr Bueso 39208382d914SDavidlohr Bueso if (vma->vm_flags & VM_SHARED) { 39218382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 39228382d914SDavidlohr Bueso key[1] = idx; 39238382d914SDavidlohr Bueso } else { 39248382d914SDavidlohr Bueso key[0] = (unsigned long) mm; 39258382d914SDavidlohr Bueso key[1] = address >> huge_page_shift(h); 39268382d914SDavidlohr Bueso } 39278382d914SDavidlohr Bueso 39288382d914SDavidlohr Bueso hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0); 39298382d914SDavidlohr Bueso 39308382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 39318382d914SDavidlohr Bueso } 39328382d914SDavidlohr Bueso #else 39338382d914SDavidlohr Bueso /* 39348382d914SDavidlohr Bueso * For uniprocesor systems we always use a single mutex, so just 39358382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 39368382d914SDavidlohr Bueso */ 3937c672c7f2SMike Kravetz u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, 39388382d914SDavidlohr Bueso struct vm_area_struct *vma, 39398382d914SDavidlohr Bueso struct address_space *mapping, 39408382d914SDavidlohr Bueso pgoff_t idx, unsigned long address) 39418382d914SDavidlohr Bueso { 39428382d914SDavidlohr Bueso return 0; 39438382d914SDavidlohr Bueso } 39448382d914SDavidlohr Bueso #endif 39458382d914SDavidlohr Bueso 39462b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 3947788c7df4SHugh Dickins unsigned long address, unsigned int flags) 394886e5216fSAdam Litke { 39498382d914SDavidlohr Bueso pte_t *ptep, entry; 3950cb900f41SKirill A. Shutemov spinlock_t *ptl; 39512b740303SSouptick Joarder vm_fault_t ret; 39528382d914SDavidlohr Bueso u32 hash; 39538382d914SDavidlohr Bueso pgoff_t idx; 39540fe6e20bSNaoya Horiguchi struct page *page = NULL; 395557303d80SAndy Whitcroft struct page *pagecache_page = NULL; 3956a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 39578382d914SDavidlohr Bueso struct address_space *mapping; 39580f792cf9SNaoya Horiguchi int need_wait_lock = 0; 3959285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 396086e5216fSAdam Litke 3961285b8dcaSHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 3962fd6a03edSNaoya Horiguchi if (ptep) { 3963fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 3964290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 3965cb900f41SKirill A. Shutemov migration_entry_wait_huge(vma, mm, ptep); 3966290408d4SNaoya Horiguchi return 0; 3967290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 3968aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 3969972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 3970ddeaab32SMike Kravetz } else { 3971b43a9990SMike Kravetz ptep = huge_pte_alloc(mm, haddr, huge_page_size(h)); 3972ddeaab32SMike Kravetz if (!ptep) 3973b43a9990SMike Kravetz return VM_FAULT_OOM; 3974b43a9990SMike Kravetz } 39758382d914SDavidlohr Bueso 3976ddeaab32SMike Kravetz mapping = vma->vm_file->f_mapping; 3977ddeaab32SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 3978ddeaab32SMike Kravetz 39793935baa9SDavid Gibson /* 39803935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 39813935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 39823935baa9SDavid Gibson * the same page in the page cache. 39833935baa9SDavid Gibson */ 3984285b8dcaSHuang Ying hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, haddr); 3985c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 39868382d914SDavidlohr Bueso 39877f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 39887f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 39898382d914SDavidlohr Bueso ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 3990b4d1d99fSDavid Gibson goto out_mutex; 39913935baa9SDavid Gibson } 399286e5216fSAdam Litke 399383c54070SNick Piggin ret = 0; 39941e8f889bSDavid Gibson 399557303d80SAndy Whitcroft /* 39960f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 39970f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 39980f792cf9SNaoya Horiguchi * a active hugepage in pagecache. This goto expects the 2nd page fault, 39990f792cf9SNaoya Horiguchi * and is_hugetlb_entry_(migration|hwpoisoned) check will properly 40000f792cf9SNaoya Horiguchi * handle it. 40010f792cf9SNaoya Horiguchi */ 40020f792cf9SNaoya Horiguchi if (!pte_present(entry)) 40030f792cf9SNaoya Horiguchi goto out_mutex; 40040f792cf9SNaoya Horiguchi 40050f792cf9SNaoya Horiguchi /* 400657303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 400757303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 400857303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 400957303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 401057303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 401157303d80SAndy Whitcroft * consumed. 401257303d80SAndy Whitcroft */ 4013106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 4014285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 40152b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 4016b4d1d99fSDavid Gibson goto out_mutex; 40172b26736cSAndy Whitcroft } 40185e911373SMike Kravetz /* Just decrements count, does not deallocate */ 4019285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 402057303d80SAndy Whitcroft 4021f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 402257303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 4023285b8dcaSHuang Ying vma, haddr); 402457303d80SAndy Whitcroft } 402557303d80SAndy Whitcroft 40260f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 40270fe6e20bSNaoya Horiguchi 40281e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 4029b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 4030cb900f41SKirill A. Shutemov goto out_ptl; 4031b4d1d99fSDavid Gibson 40320f792cf9SNaoya Horiguchi /* 40330f792cf9SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 40340f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 40350f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 40360f792cf9SNaoya Horiguchi */ 40370f792cf9SNaoya Horiguchi page = pte_page(entry); 40380f792cf9SNaoya Horiguchi if (page != pagecache_page) 40390f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 40400f792cf9SNaoya Horiguchi need_wait_lock = 1; 40410f792cf9SNaoya Horiguchi goto out_ptl; 40420f792cf9SNaoya Horiguchi } 40430f792cf9SNaoya Horiguchi 40440f792cf9SNaoya Horiguchi get_page(page); 4045b4d1d99fSDavid Gibson 4046788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 4047106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 4048974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, 4049cb900f41SKirill A. Shutemov pagecache_page, ptl); 40500f792cf9SNaoya Horiguchi goto out_put_page; 4051b4d1d99fSDavid Gibson } 4052106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 4053b4d1d99fSDavid Gibson } 4054b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 4055285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 4056788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 4057285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 40580f792cf9SNaoya Horiguchi out_put_page: 40590f792cf9SNaoya Horiguchi if (page != pagecache_page) 40600f792cf9SNaoya Horiguchi unlock_page(page); 40610f792cf9SNaoya Horiguchi put_page(page); 4062cb900f41SKirill A. Shutemov out_ptl: 4063cb900f41SKirill A. Shutemov spin_unlock(ptl); 406457303d80SAndy Whitcroft 406557303d80SAndy Whitcroft if (pagecache_page) { 406657303d80SAndy Whitcroft unlock_page(pagecache_page); 406757303d80SAndy Whitcroft put_page(pagecache_page); 406857303d80SAndy Whitcroft } 4069b4d1d99fSDavid Gibson out_mutex: 4070c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 40710f792cf9SNaoya Horiguchi /* 40720f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 40730f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 40740f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 40750f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 40760f792cf9SNaoya Horiguchi * here without taking refcount. 40770f792cf9SNaoya Horiguchi */ 40780f792cf9SNaoya Horiguchi if (need_wait_lock) 40790f792cf9SNaoya Horiguchi wait_on_page_locked(page); 40801e8f889bSDavid Gibson return ret; 408186e5216fSAdam Litke } 408286e5216fSAdam Litke 40838fb5debcSMike Kravetz /* 40848fb5debcSMike Kravetz * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with 40858fb5debcSMike Kravetz * modifications for huge pages. 40868fb5debcSMike Kravetz */ 40878fb5debcSMike Kravetz int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, 40888fb5debcSMike Kravetz pte_t *dst_pte, 40898fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 40908fb5debcSMike Kravetz unsigned long dst_addr, 40918fb5debcSMike Kravetz unsigned long src_addr, 40928fb5debcSMike Kravetz struct page **pagep) 40938fb5debcSMike Kravetz { 40941e392147SAndrea Arcangeli struct address_space *mapping; 40951e392147SAndrea Arcangeli pgoff_t idx; 40961e392147SAndrea Arcangeli unsigned long size; 40971c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 40988fb5debcSMike Kravetz struct hstate *h = hstate_vma(dst_vma); 40998fb5debcSMike Kravetz pte_t _dst_pte; 41008fb5debcSMike Kravetz spinlock_t *ptl; 41018fb5debcSMike Kravetz int ret; 41028fb5debcSMike Kravetz struct page *page; 41038fb5debcSMike Kravetz 41048fb5debcSMike Kravetz if (!*pagep) { 41058fb5debcSMike Kravetz ret = -ENOMEM; 41068fb5debcSMike Kravetz page = alloc_huge_page(dst_vma, dst_addr, 0); 41078fb5debcSMike Kravetz if (IS_ERR(page)) 41088fb5debcSMike Kravetz goto out; 41098fb5debcSMike Kravetz 41108fb5debcSMike Kravetz ret = copy_huge_page_from_user(page, 41118fb5debcSMike Kravetz (const void __user *) src_addr, 4112810a56b9SMike Kravetz pages_per_huge_page(h), false); 41138fb5debcSMike Kravetz 41148fb5debcSMike Kravetz /* fallback to copy_from_user outside mmap_sem */ 41158fb5debcSMike Kravetz if (unlikely(ret)) { 41169e368259SAndrea Arcangeli ret = -ENOENT; 41178fb5debcSMike Kravetz *pagep = page; 41188fb5debcSMike Kravetz /* don't free the page */ 41198fb5debcSMike Kravetz goto out; 41208fb5debcSMike Kravetz } 41218fb5debcSMike Kravetz } else { 41228fb5debcSMike Kravetz page = *pagep; 41238fb5debcSMike Kravetz *pagep = NULL; 41248fb5debcSMike Kravetz } 41258fb5debcSMike Kravetz 41268fb5debcSMike Kravetz /* 41278fb5debcSMike Kravetz * The memory barrier inside __SetPageUptodate makes sure that 41288fb5debcSMike Kravetz * preceding stores to the page contents become visible before 41298fb5debcSMike Kravetz * the set_pte_at() write. 41308fb5debcSMike Kravetz */ 41318fb5debcSMike Kravetz __SetPageUptodate(page); 41328fb5debcSMike Kravetz 41331e392147SAndrea Arcangeli mapping = dst_vma->vm_file->f_mapping; 41341e392147SAndrea Arcangeli idx = vma_hugecache_offset(h, dst_vma, dst_addr); 41351e392147SAndrea Arcangeli 41361c9e8defSMike Kravetz /* 41371c9e8defSMike Kravetz * If shared, add to page cache 41381c9e8defSMike Kravetz */ 41391c9e8defSMike Kravetz if (vm_shared) { 41401e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 41411e392147SAndrea Arcangeli ret = -EFAULT; 41421e392147SAndrea Arcangeli if (idx >= size) 41431e392147SAndrea Arcangeli goto out_release_nounlock; 41441c9e8defSMike Kravetz 41451e392147SAndrea Arcangeli /* 41461e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 41471e392147SAndrea Arcangeli * huge_add_to_page_cache() below happens through the 41481e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 41491e392147SAndrea Arcangeli * the caller. 41501e392147SAndrea Arcangeli */ 41511c9e8defSMike Kravetz ret = huge_add_to_page_cache(page, mapping, idx); 41521c9e8defSMike Kravetz if (ret) 41531c9e8defSMike Kravetz goto out_release_nounlock; 41541c9e8defSMike Kravetz } 41551c9e8defSMike Kravetz 41568fb5debcSMike Kravetz ptl = huge_pte_lockptr(h, dst_mm, dst_pte); 41578fb5debcSMike Kravetz spin_lock(ptl); 41588fb5debcSMike Kravetz 41591e392147SAndrea Arcangeli /* 41601e392147SAndrea Arcangeli * Recheck the i_size after holding PT lock to make sure not 41611e392147SAndrea Arcangeli * to leave any page mapped (as page_mapped()) beyond the end 41621e392147SAndrea Arcangeli * of the i_size (remove_inode_hugepages() is strict about 41631e392147SAndrea Arcangeli * enforcing that). If we bail out here, we'll also leave a 41641e392147SAndrea Arcangeli * page in the radix tree in the vm_shared case beyond the end 41651e392147SAndrea Arcangeli * of the i_size, but remove_inode_hugepages() will take care 41661e392147SAndrea Arcangeli * of it as soon as we drop the hugetlb_fault_mutex_table. 41671e392147SAndrea Arcangeli */ 41681e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 41691e392147SAndrea Arcangeli ret = -EFAULT; 41701e392147SAndrea Arcangeli if (idx >= size) 41711e392147SAndrea Arcangeli goto out_release_unlock; 41721e392147SAndrea Arcangeli 41738fb5debcSMike Kravetz ret = -EEXIST; 41748fb5debcSMike Kravetz if (!huge_pte_none(huge_ptep_get(dst_pte))) 41758fb5debcSMike Kravetz goto out_release_unlock; 41768fb5debcSMike Kravetz 41771c9e8defSMike Kravetz if (vm_shared) { 41781c9e8defSMike Kravetz page_dup_rmap(page, true); 41791c9e8defSMike Kravetz } else { 41808fb5debcSMike Kravetz ClearPagePrivate(page); 41818fb5debcSMike Kravetz hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); 41821c9e8defSMike Kravetz } 41838fb5debcSMike Kravetz 41848fb5debcSMike Kravetz _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); 41858fb5debcSMike Kravetz if (dst_vma->vm_flags & VM_WRITE) 41868fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 41878fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 41888fb5debcSMike Kravetz 41898fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 41908fb5debcSMike Kravetz 41918fb5debcSMike Kravetz (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, 41928fb5debcSMike Kravetz dst_vma->vm_flags & VM_WRITE); 41938fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 41948fb5debcSMike Kravetz 41958fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 41968fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 41978fb5debcSMike Kravetz 41988fb5debcSMike Kravetz spin_unlock(ptl); 4199cb6acd01SMike Kravetz set_page_huge_active(page); 42001c9e8defSMike Kravetz if (vm_shared) 42011c9e8defSMike Kravetz unlock_page(page); 42028fb5debcSMike Kravetz ret = 0; 42038fb5debcSMike Kravetz out: 42048fb5debcSMike Kravetz return ret; 42058fb5debcSMike Kravetz out_release_unlock: 42068fb5debcSMike Kravetz spin_unlock(ptl); 42071c9e8defSMike Kravetz if (vm_shared) 42081c9e8defSMike Kravetz unlock_page(page); 42095af10dfdSAndrea Arcangeli out_release_nounlock: 42108fb5debcSMike Kravetz put_page(page); 42118fb5debcSMike Kravetz goto out; 42128fb5debcSMike Kravetz } 42138fb5debcSMike Kravetz 421428a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 421563551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 421628a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 421787ffc118SAndrea Arcangeli long i, unsigned int flags, int *nonblocking) 421863551ae0SDavid Gibson { 4219d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 4220d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 422128a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 4222a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 42232be7cfedSDaniel Jordan int err = -EFAULT; 422463551ae0SDavid Gibson 422563551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 422663551ae0SDavid Gibson pte_t *pte; 4227cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 42282a15efc9SHugh Dickins int absent; 422963551ae0SDavid Gibson struct page *page; 423063551ae0SDavid Gibson 42314c887265SAdam Litke /* 423202057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 423302057967SDavid Rientjes * potentially allocating memory. 423402057967SDavid Rientjes */ 4235fa45f116SDavidlohr Bueso if (fatal_signal_pending(current)) { 423602057967SDavid Rientjes remainder = 0; 423702057967SDavid Rientjes break; 423802057967SDavid Rientjes } 423902057967SDavid Rientjes 424002057967SDavid Rientjes /* 42414c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 42422a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 42434c887265SAdam Litke * first, for the page indexing below to work. 4244cb900f41SKirill A. Shutemov * 4245cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 42464c887265SAdam Litke */ 42477868a208SPunit Agrawal pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), 42487868a208SPunit Agrawal huge_page_size(h)); 4249cb900f41SKirill A. Shutemov if (pte) 4250cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 42512a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 425263551ae0SDavid Gibson 42532a15efc9SHugh Dickins /* 42542a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 42553ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 42563ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 42573ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 42583ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 42592a15efc9SHugh Dickins */ 42603ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 42613ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 4262cb900f41SKirill A. Shutemov if (pte) 4263cb900f41SKirill A. Shutemov spin_unlock(ptl); 42642a15efc9SHugh Dickins remainder = 0; 42652a15efc9SHugh Dickins break; 42662a15efc9SHugh Dickins } 42672a15efc9SHugh Dickins 42689cc3a5bdSNaoya Horiguchi /* 42699cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 42709cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 42719cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 42729cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 42739cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 42749cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 42759cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 42769cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 42779cc3a5bdSNaoya Horiguchi */ 42789cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 4279106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 4280106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 42812b740303SSouptick Joarder vm_fault_t ret; 428287ffc118SAndrea Arcangeli unsigned int fault_flags = 0; 42834c887265SAdam Litke 4284cb900f41SKirill A. Shutemov if (pte) 4285cb900f41SKirill A. Shutemov spin_unlock(ptl); 428687ffc118SAndrea Arcangeli if (flags & FOLL_WRITE) 428787ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_WRITE; 428887ffc118SAndrea Arcangeli if (nonblocking) 428987ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY; 429087ffc118SAndrea Arcangeli if (flags & FOLL_NOWAIT) 429187ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY | 429287ffc118SAndrea Arcangeli FAULT_FLAG_RETRY_NOWAIT; 429387ffc118SAndrea Arcangeli if (flags & FOLL_TRIED) { 429487ffc118SAndrea Arcangeli VM_WARN_ON_ONCE(fault_flags & 429587ffc118SAndrea Arcangeli FAULT_FLAG_ALLOW_RETRY); 429687ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_TRIED; 429787ffc118SAndrea Arcangeli } 429887ffc118SAndrea Arcangeli ret = hugetlb_fault(mm, vma, vaddr, fault_flags); 429987ffc118SAndrea Arcangeli if (ret & VM_FAULT_ERROR) { 43002be7cfedSDaniel Jordan err = vm_fault_to_errno(ret, flags); 43011c59827dSHugh Dickins remainder = 0; 43021c59827dSHugh Dickins break; 43031c59827dSHugh Dickins } 430487ffc118SAndrea Arcangeli if (ret & VM_FAULT_RETRY) { 43051ac25013SAndrea Arcangeli if (nonblocking && 43061ac25013SAndrea Arcangeli !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) 430787ffc118SAndrea Arcangeli *nonblocking = 0; 430887ffc118SAndrea Arcangeli *nr_pages = 0; 430987ffc118SAndrea Arcangeli /* 431087ffc118SAndrea Arcangeli * VM_FAULT_RETRY must not return an 431187ffc118SAndrea Arcangeli * error, it will return zero 431287ffc118SAndrea Arcangeli * instead. 431387ffc118SAndrea Arcangeli * 431487ffc118SAndrea Arcangeli * No need to update "position" as the 431587ffc118SAndrea Arcangeli * caller will not check it after 431687ffc118SAndrea Arcangeli * *nr_pages is set to 0. 431787ffc118SAndrea Arcangeli */ 431887ffc118SAndrea Arcangeli return i; 431987ffc118SAndrea Arcangeli } 432087ffc118SAndrea Arcangeli continue; 432187ffc118SAndrea Arcangeli } 432263551ae0SDavid Gibson 4323a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 43247f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 43258fde12caSLinus Torvalds 43268fde12caSLinus Torvalds /* 43278fde12caSLinus Torvalds * Instead of doing 'try_get_page()' below in the same_page 43288fde12caSLinus Torvalds * loop, just check the count once here. 43298fde12caSLinus Torvalds */ 43308fde12caSLinus Torvalds if (unlikely(page_count(page) <= 0)) { 43318fde12caSLinus Torvalds if (pages) { 43328fde12caSLinus Torvalds spin_unlock(ptl); 43338fde12caSLinus Torvalds remainder = 0; 43348fde12caSLinus Torvalds err = -ENOMEM; 43358fde12caSLinus Torvalds break; 43368fde12caSLinus Torvalds } 43378fde12caSLinus Torvalds } 4338d5d4b0aaSChen, Kenneth W same_page: 4339d6692183SChen, Kenneth W if (pages) { 434069d177c2SAndy Whitcroft pages[i] = mem_map_offset(page, pfn_offset); 4341ddc58f27SKirill A. Shutemov get_page(pages[i]); 4342d6692183SChen, Kenneth W } 434363551ae0SDavid Gibson 434463551ae0SDavid Gibson if (vmas) 434563551ae0SDavid Gibson vmas[i] = vma; 434663551ae0SDavid Gibson 434763551ae0SDavid Gibson vaddr += PAGE_SIZE; 4348d5d4b0aaSChen, Kenneth W ++pfn_offset; 434963551ae0SDavid Gibson --remainder; 435063551ae0SDavid Gibson ++i; 4351d5d4b0aaSChen, Kenneth W if (vaddr < vma->vm_end && remainder && 4352a5516438SAndi Kleen pfn_offset < pages_per_huge_page(h)) { 4353d5d4b0aaSChen, Kenneth W /* 4354d5d4b0aaSChen, Kenneth W * We use pfn_offset to avoid touching the pageframes 4355d5d4b0aaSChen, Kenneth W * of this compound page. 4356d5d4b0aaSChen, Kenneth W */ 4357d5d4b0aaSChen, Kenneth W goto same_page; 4358d5d4b0aaSChen, Kenneth W } 4359cb900f41SKirill A. Shutemov spin_unlock(ptl); 436063551ae0SDavid Gibson } 436128a35716SMichel Lespinasse *nr_pages = remainder; 436287ffc118SAndrea Arcangeli /* 436387ffc118SAndrea Arcangeli * setting position is actually required only if remainder is 436487ffc118SAndrea Arcangeli * not zero but it's faster not to add a "if (remainder)" 436587ffc118SAndrea Arcangeli * branch. 436687ffc118SAndrea Arcangeli */ 436763551ae0SDavid Gibson *position = vaddr; 436863551ae0SDavid Gibson 43692be7cfedSDaniel Jordan return i ? i : err; 437063551ae0SDavid Gibson } 43718f860591SZhang, Yanmin 43725491ae7bSAneesh Kumar K.V #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE 43735491ae7bSAneesh Kumar K.V /* 43745491ae7bSAneesh Kumar K.V * ARCHes with special requirements for evicting HUGETLB backing TLB entries can 43755491ae7bSAneesh Kumar K.V * implement this. 43765491ae7bSAneesh Kumar K.V */ 43775491ae7bSAneesh Kumar K.V #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) 43785491ae7bSAneesh Kumar K.V #endif 43795491ae7bSAneesh Kumar K.V 43807da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 43818f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 43828f860591SZhang, Yanmin { 43838f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 43848f860591SZhang, Yanmin unsigned long start = address; 43858f860591SZhang, Yanmin pte_t *ptep; 43868f860591SZhang, Yanmin pte_t pte; 4387a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 43887da4d641SPeter Zijlstra unsigned long pages = 0; 4389dff11abeSMike Kravetz bool shared_pmd = false; 4390ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 4391dff11abeSMike Kravetz 4392dff11abeSMike Kravetz /* 4393dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 4394ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 4395dff11abeSMike Kravetz * range if PMD sharing is possible. 4396dff11abeSMike Kravetz */ 4397ac46d4f3SJérôme Glisse mmu_notifier_range_init(&range, mm, start, end); 4398ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 43998f860591SZhang, Yanmin 44008f860591SZhang, Yanmin BUG_ON(address >= end); 4401ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 44028f860591SZhang, Yanmin 4403ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 440483cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 4405a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 4406cb900f41SKirill A. Shutemov spinlock_t *ptl; 44077868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, huge_page_size(h)); 44088f860591SZhang, Yanmin if (!ptep) 44098f860591SZhang, Yanmin continue; 4410cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 44117da4d641SPeter Zijlstra if (huge_pmd_unshare(mm, &address, ptep)) { 44127da4d641SPeter Zijlstra pages++; 4413cb900f41SKirill A. Shutemov spin_unlock(ptl); 4414dff11abeSMike Kravetz shared_pmd = true; 441539dde65cSChen, Kenneth W continue; 44167da4d641SPeter Zijlstra } 4417a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 4418a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 4419a8bda28dSNaoya Horiguchi spin_unlock(ptl); 4420a8bda28dSNaoya Horiguchi continue; 4421a8bda28dSNaoya Horiguchi } 4422a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 4423a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 4424a8bda28dSNaoya Horiguchi 4425a8bda28dSNaoya Horiguchi if (is_write_migration_entry(entry)) { 4426a8bda28dSNaoya Horiguchi pte_t newpte; 4427a8bda28dSNaoya Horiguchi 4428a8bda28dSNaoya Horiguchi make_migration_entry_read(&entry); 4429a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 4430e5251fd4SPunit Agrawal set_huge_swap_pte_at(mm, address, ptep, 4431e5251fd4SPunit Agrawal newpte, huge_page_size(h)); 4432a8bda28dSNaoya Horiguchi pages++; 4433a8bda28dSNaoya Horiguchi } 4434a8bda28dSNaoya Horiguchi spin_unlock(ptl); 4435a8bda28dSNaoya Horiguchi continue; 4436a8bda28dSNaoya Horiguchi } 4437a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 4438023bdd00SAneesh Kumar K.V pte_t old_pte; 4439023bdd00SAneesh Kumar K.V 4440023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 4441023bdd00SAneesh Kumar K.V pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); 4442be7517d6STony Lu pte = arch_make_huge_pte(pte, vma, NULL, 0); 4443023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 44447da4d641SPeter Zijlstra pages++; 44458f860591SZhang, Yanmin } 4446cb900f41SKirill A. Shutemov spin_unlock(ptl); 44478f860591SZhang, Yanmin } 4448d833352aSMel Gorman /* 4449c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 4450d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 4451c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 4452dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 4453dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 4454d833352aSMel Gorman */ 4455dff11abeSMike Kravetz if (shared_pmd) 4456ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 4457dff11abeSMike Kravetz else 44585491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 44590f10851eSJérôme Glisse /* 44600f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 44610f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 44620f10851eSJérôme Glisse * 4463ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 44640f10851eSJérôme Glisse */ 446583cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 4466ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 44677da4d641SPeter Zijlstra 44687da4d641SPeter Zijlstra return pages << h->order; 44698f860591SZhang, Yanmin } 44708f860591SZhang, Yanmin 4471a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 4472a1e78772SMel Gorman long from, long to, 44735a6fe125SMel Gorman struct vm_area_struct *vma, 4474ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 4475e4e574b7SAdam Litke { 447617c9d12eSMel Gorman long ret, chg; 4477a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 447890481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 44799119a41eSJoonsoo Kim struct resv_map *resv_map; 44801c5ecae3SMike Kravetz long gbl_reserve; 4481e4e574b7SAdam Litke 448263489f8eSMike Kravetz /* This should never happen */ 448363489f8eSMike Kravetz if (from > to) { 448463489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 448563489f8eSMike Kravetz return -EINVAL; 448663489f8eSMike Kravetz } 448763489f8eSMike Kravetz 4488a1e78772SMel Gorman /* 448917c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 449017c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 449190481622SDavid Gibson * without using reserves 449217c9d12eSMel Gorman */ 4493ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 449417c9d12eSMel Gorman return 0; 449517c9d12eSMel Gorman 449617c9d12eSMel Gorman /* 4497a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 4498a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 4499a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 4500a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 4501a1e78772SMel Gorman */ 45029119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 45034e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 45049119a41eSJoonsoo Kim 45051406ec9bSJoonsoo Kim chg = region_chg(resv_map, from, to); 45069119a41eSJoonsoo Kim 45079119a41eSJoonsoo Kim } else { 45089119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 45095a6fe125SMel Gorman if (!resv_map) 45105a6fe125SMel Gorman return -ENOMEM; 45115a6fe125SMel Gorman 451217c9d12eSMel Gorman chg = to - from; 451317c9d12eSMel Gorman 45145a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 45155a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 45165a6fe125SMel Gorman } 45175a6fe125SMel Gorman 4518c50ac050SDave Hansen if (chg < 0) { 4519c50ac050SDave Hansen ret = chg; 4520c50ac050SDave Hansen goto out_err; 4521c50ac050SDave Hansen } 452217c9d12eSMel Gorman 45231c5ecae3SMike Kravetz /* 45241c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 45251c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 45261c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 45271c5ecae3SMike Kravetz */ 45281c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 45291c5ecae3SMike Kravetz if (gbl_reserve < 0) { 4530c50ac050SDave Hansen ret = -ENOSPC; 4531c50ac050SDave Hansen goto out_err; 4532c50ac050SDave Hansen } 453317c9d12eSMel Gorman 453417c9d12eSMel Gorman /* 453517c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 453690481622SDavid Gibson * Hand the pages back to the subpool if there are not 453717c9d12eSMel Gorman */ 45381c5ecae3SMike Kravetz ret = hugetlb_acct_memory(h, gbl_reserve); 453917c9d12eSMel Gorman if (ret < 0) { 45401c5ecae3SMike Kravetz /* put back original number of pages, chg */ 45411c5ecae3SMike Kravetz (void)hugepage_subpool_put_pages(spool, chg); 4542c50ac050SDave Hansen goto out_err; 454317c9d12eSMel Gorman } 454417c9d12eSMel Gorman 454517c9d12eSMel Gorman /* 454617c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 454717c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 454817c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 454917c9d12eSMel Gorman * the reservation was and the page cache tells how much of 455017c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 455117c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 455217c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 455317c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 455417c9d12eSMel Gorman * else has to be done for private mappings here 455517c9d12eSMel Gorman */ 455633039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 455733039678SMike Kravetz long add = region_add(resv_map, from, to); 455833039678SMike Kravetz 455933039678SMike Kravetz if (unlikely(chg > add)) { 456033039678SMike Kravetz /* 456133039678SMike Kravetz * pages in this range were added to the reserve 456233039678SMike Kravetz * map between region_chg and region_add. This 456333039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 456433039678SMike Kravetz * the subpool and reserve counts modified above 456533039678SMike Kravetz * based on the difference. 456633039678SMike Kravetz */ 456733039678SMike Kravetz long rsv_adjust; 456833039678SMike Kravetz 456933039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 457033039678SMike Kravetz chg - add); 457133039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 457233039678SMike Kravetz } 457333039678SMike Kravetz } 4574a43a8c39SChen, Kenneth W return 0; 4575c50ac050SDave Hansen out_err: 45765e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 4577ff8c0c53SMike Kravetz /* Don't call region_abort if region_chg failed */ 4578ff8c0c53SMike Kravetz if (chg >= 0) 45795e911373SMike Kravetz region_abort(resv_map, from, to); 4580f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 4581f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 4582c50ac050SDave Hansen return ret; 4583a43a8c39SChen, Kenneth W } 4584a43a8c39SChen, Kenneth W 4585b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 4586b5cec28dSMike Kravetz long freed) 4587a43a8c39SChen, Kenneth W { 4588a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 45894e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 45909119a41eSJoonsoo Kim long chg = 0; 459190481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 45921c5ecae3SMike Kravetz long gbl_reserve; 459345c682a6SKen Chen 4594b5cec28dSMike Kravetz if (resv_map) { 4595b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 4596b5cec28dSMike Kravetz /* 4597b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 4598b5cec28dSMike Kravetz * must be split and another region descriptor can not be 4599b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 4600b5cec28dSMike Kravetz */ 4601b5cec28dSMike Kravetz if (chg < 0) 4602b5cec28dSMike Kravetz return chg; 4603b5cec28dSMike Kravetz } 4604b5cec28dSMike Kravetz 460545c682a6SKen Chen spin_lock(&inode->i_lock); 4606e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 460745c682a6SKen Chen spin_unlock(&inode->i_lock); 460845c682a6SKen Chen 46091c5ecae3SMike Kravetz /* 46101c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 46111c5ecae3SMike Kravetz * reservations to be released may be adjusted. 46121c5ecae3SMike Kravetz */ 46131c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 46141c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 4615b5cec28dSMike Kravetz 4616b5cec28dSMike Kravetz return 0; 4617a43a8c39SChen, Kenneth W } 461893f70f90SNaoya Horiguchi 46193212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 46203212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 46213212b535SSteve Capper struct vm_area_struct *vma, 46223212b535SSteve Capper unsigned long addr, pgoff_t idx) 46233212b535SSteve Capper { 46243212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 46253212b535SSteve Capper svma->vm_start; 46263212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 46273212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 46283212b535SSteve Capper 46293212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 4630de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 4631de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 46323212b535SSteve Capper 46333212b535SSteve Capper /* 46343212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 46353212b535SSteve Capper * page table page. 46363212b535SSteve Capper */ 46373212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 46383212b535SSteve Capper vm_flags != svm_flags || 46393212b535SSteve Capper sbase < svma->vm_start || svma->vm_end < s_end) 46403212b535SSteve Capper return 0; 46413212b535SSteve Capper 46423212b535SSteve Capper return saddr; 46433212b535SSteve Capper } 46443212b535SSteve Capper 464531aafb45SNicholas Krause static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) 46463212b535SSteve Capper { 46473212b535SSteve Capper unsigned long base = addr & PUD_MASK; 46483212b535SSteve Capper unsigned long end = base + PUD_SIZE; 46493212b535SSteve Capper 46503212b535SSteve Capper /* 46513212b535SSteve Capper * check on proper vm_flags and page table alignment 46523212b535SSteve Capper */ 4653017b1660SMike Kravetz if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) 465431aafb45SNicholas Krause return true; 465531aafb45SNicholas Krause return false; 46563212b535SSteve Capper } 46573212b535SSteve Capper 46583212b535SSteve Capper /* 4659017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 4660017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 4661017b1660SMike Kravetz * shared pmd mappings. 4662017b1660SMike Kravetz */ 4663017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 4664017b1660SMike Kravetz unsigned long *start, unsigned long *end) 4665017b1660SMike Kravetz { 4666017b1660SMike Kravetz unsigned long check_addr = *start; 4667017b1660SMike Kravetz 4668017b1660SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 4669017b1660SMike Kravetz return; 4670017b1660SMike Kravetz 4671017b1660SMike Kravetz for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) { 4672017b1660SMike Kravetz unsigned long a_start = check_addr & PUD_MASK; 4673017b1660SMike Kravetz unsigned long a_end = a_start + PUD_SIZE; 4674017b1660SMike Kravetz 4675017b1660SMike Kravetz /* 4676017b1660SMike Kravetz * If sharing is possible, adjust start/end if necessary. 4677017b1660SMike Kravetz */ 4678017b1660SMike Kravetz if (range_in_vma(vma, a_start, a_end)) { 4679017b1660SMike Kravetz if (a_start < *start) 4680017b1660SMike Kravetz *start = a_start; 4681017b1660SMike Kravetz if (a_end > *end) 4682017b1660SMike Kravetz *end = a_end; 4683017b1660SMike Kravetz } 4684017b1660SMike Kravetz } 4685017b1660SMike Kravetz } 4686017b1660SMike Kravetz 4687017b1660SMike Kravetz /* 46883212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 46893212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 46903212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 4691ddeaab32SMike Kravetz * code much cleaner. pmd allocation is essential for the shared case because 4692ddeaab32SMike Kravetz * pud has to be populated inside the same i_mmap_rwsem section - otherwise 4693ddeaab32SMike Kravetz * racing tasks could either miss the sharing (see huge_pte_offset) or select a 4694ddeaab32SMike Kravetz * bad pmd for sharing. 46953212b535SSteve Capper */ 46963212b535SSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 46973212b535SSteve Capper { 46983212b535SSteve Capper struct vm_area_struct *vma = find_vma(mm, addr); 46993212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 47003212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 47013212b535SSteve Capper vma->vm_pgoff; 47023212b535SSteve Capper struct vm_area_struct *svma; 47033212b535SSteve Capper unsigned long saddr; 47043212b535SSteve Capper pte_t *spte = NULL; 47053212b535SSteve Capper pte_t *pte; 4706cb900f41SKirill A. Shutemov spinlock_t *ptl; 47073212b535SSteve Capper 47083212b535SSteve Capper if (!vma_shareable(vma, addr)) 47093212b535SSteve Capper return (pte_t *)pmd_alloc(mm, pud, addr); 47103212b535SSteve Capper 4711ddeaab32SMike Kravetz i_mmap_lock_write(mapping); 47123212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 47133212b535SSteve Capper if (svma == vma) 47143212b535SSteve Capper continue; 47153212b535SSteve Capper 47163212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 47173212b535SSteve Capper if (saddr) { 47187868a208SPunit Agrawal spte = huge_pte_offset(svma->vm_mm, saddr, 47197868a208SPunit Agrawal vma_mmu_pagesize(svma)); 47203212b535SSteve Capper if (spte) { 47213212b535SSteve Capper get_page(virt_to_page(spte)); 47223212b535SSteve Capper break; 47233212b535SSteve Capper } 47243212b535SSteve Capper } 47253212b535SSteve Capper } 47263212b535SSteve Capper 47273212b535SSteve Capper if (!spte) 47283212b535SSteve Capper goto out; 47293212b535SSteve Capper 47308bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 4731dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 47323212b535SSteve Capper pud_populate(mm, pud, 47333212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 4734c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 4735dc6c9a35SKirill A. Shutemov } else { 47363212b535SSteve Capper put_page(virt_to_page(spte)); 4737dc6c9a35SKirill A. Shutemov } 4738cb900f41SKirill A. Shutemov spin_unlock(ptl); 47393212b535SSteve Capper out: 47403212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 4741ddeaab32SMike Kravetz i_mmap_unlock_write(mapping); 47423212b535SSteve Capper return pte; 47433212b535SSteve Capper } 47443212b535SSteve Capper 47453212b535SSteve Capper /* 47463212b535SSteve Capper * unmap huge page backed by shared pte. 47473212b535SSteve Capper * 47483212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 47493212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 47503212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 47513212b535SSteve Capper * 4752ddeaab32SMike Kravetz * called with page table lock held. 47533212b535SSteve Capper * 47543212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 47553212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 47563212b535SSteve Capper */ 47573212b535SSteve Capper int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 47583212b535SSteve Capper { 47593212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 4760c2febafcSKirill A. Shutemov p4d_t *p4d = p4d_offset(pgd, *addr); 4761c2febafcSKirill A. Shutemov pud_t *pud = pud_offset(p4d, *addr); 47623212b535SSteve Capper 47633212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 47643212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 47653212b535SSteve Capper return 0; 47663212b535SSteve Capper 47673212b535SSteve Capper pud_clear(pud); 47683212b535SSteve Capper put_page(virt_to_page(ptep)); 4769dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 47703212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 47713212b535SSteve Capper return 1; 47723212b535SSteve Capper } 47739e5fc74cSSteve Capper #define want_pmd_share() (1) 47749e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 47759e5fc74cSSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 47769e5fc74cSSteve Capper { 47779e5fc74cSSteve Capper return NULL; 47789e5fc74cSSteve Capper } 4779e81f2d22SZhang Zhen 4780e81f2d22SZhang Zhen int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 4781e81f2d22SZhang Zhen { 4782e81f2d22SZhang Zhen return 0; 4783e81f2d22SZhang Zhen } 4784017b1660SMike Kravetz 4785017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 4786017b1660SMike Kravetz unsigned long *start, unsigned long *end) 4787017b1660SMike Kravetz { 4788017b1660SMike Kravetz } 47899e5fc74cSSteve Capper #define want_pmd_share() (0) 47903212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 47913212b535SSteve Capper 47929e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 47939e5fc74cSSteve Capper pte_t *huge_pte_alloc(struct mm_struct *mm, 47949e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 47959e5fc74cSSteve Capper { 47969e5fc74cSSteve Capper pgd_t *pgd; 4797c2febafcSKirill A. Shutemov p4d_t *p4d; 47989e5fc74cSSteve Capper pud_t *pud; 47999e5fc74cSSteve Capper pte_t *pte = NULL; 48009e5fc74cSSteve Capper 48019e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 4802f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 4803f4f0a3d8SKirill A. Shutemov if (!p4d) 4804f4f0a3d8SKirill A. Shutemov return NULL; 4805c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 48069e5fc74cSSteve Capper if (pud) { 48079e5fc74cSSteve Capper if (sz == PUD_SIZE) { 48089e5fc74cSSteve Capper pte = (pte_t *)pud; 48099e5fc74cSSteve Capper } else { 48109e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 48119e5fc74cSSteve Capper if (want_pmd_share() && pud_none(*pud)) 48129e5fc74cSSteve Capper pte = huge_pmd_share(mm, addr, pud); 48139e5fc74cSSteve Capper else 48149e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 48159e5fc74cSSteve Capper } 48169e5fc74cSSteve Capper } 48174e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 48189e5fc74cSSteve Capper 48199e5fc74cSSteve Capper return pte; 48209e5fc74cSSteve Capper } 48219e5fc74cSSteve Capper 48229b19df29SPunit Agrawal /* 48239b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 48249b19df29SPunit Agrawal * entry at address @addr 48259b19df29SPunit Agrawal * 48269b19df29SPunit Agrawal * Return: Pointer to page table or swap entry (PUD or PMD) for 48279b19df29SPunit Agrawal * address @addr, or NULL if a p*d_none() entry is encountered and the 48289b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 48299b19df29SPunit Agrawal * table. 48309b19df29SPunit Agrawal */ 48317868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 48327868a208SPunit Agrawal unsigned long addr, unsigned long sz) 48339e5fc74cSSteve Capper { 48349e5fc74cSSteve Capper pgd_t *pgd; 4835c2febafcSKirill A. Shutemov p4d_t *p4d; 48369e5fc74cSSteve Capper pud_t *pud; 4837c2febafcSKirill A. Shutemov pmd_t *pmd; 48389e5fc74cSSteve Capper 48399e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 4840c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 4841c2febafcSKirill A. Shutemov return NULL; 4842c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 4843c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 4844c2febafcSKirill A. Shutemov return NULL; 48459b19df29SPunit Agrawal 4846c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 48479b19df29SPunit Agrawal if (sz != PUD_SIZE && pud_none(*pud)) 4848c2febafcSKirill A. Shutemov return NULL; 48499b19df29SPunit Agrawal /* hugepage or swap? */ 48509b19df29SPunit Agrawal if (pud_huge(*pud) || !pud_present(*pud)) 48519e5fc74cSSteve Capper return (pte_t *)pud; 48529b19df29SPunit Agrawal 48539e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 48549b19df29SPunit Agrawal if (sz != PMD_SIZE && pmd_none(*pmd)) 48559b19df29SPunit Agrawal return NULL; 48569b19df29SPunit Agrawal /* hugepage or swap? */ 48579b19df29SPunit Agrawal if (pmd_huge(*pmd) || !pmd_present(*pmd)) 48589e5fc74cSSteve Capper return (pte_t *)pmd; 48599b19df29SPunit Agrawal 48609b19df29SPunit Agrawal return NULL; 48619e5fc74cSSteve Capper } 48629e5fc74cSSteve Capper 486361f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 486461f77edaSNaoya Horiguchi 486561f77edaSNaoya Horiguchi /* 486661f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 486761f77edaSNaoya Horiguchi * behavior. 486861f77edaSNaoya Horiguchi */ 486961f77edaSNaoya Horiguchi struct page * __weak 487061f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 487161f77edaSNaoya Horiguchi int write) 487261f77edaSNaoya Horiguchi { 487361f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 487461f77edaSNaoya Horiguchi } 487561f77edaSNaoya Horiguchi 487661f77edaSNaoya Horiguchi struct page * __weak 48774dc71451SAneesh Kumar K.V follow_huge_pd(struct vm_area_struct *vma, 48784dc71451SAneesh Kumar K.V unsigned long address, hugepd_t hpd, int flags, int pdshift) 48794dc71451SAneesh Kumar K.V { 48804dc71451SAneesh Kumar K.V WARN(1, "hugepd follow called with no support for hugepage directory format\n"); 48814dc71451SAneesh Kumar K.V return NULL; 48824dc71451SAneesh Kumar K.V } 48834dc71451SAneesh Kumar K.V 48844dc71451SAneesh Kumar K.V struct page * __weak 48859e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 4886e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 48879e5fc74cSSteve Capper { 4888e66f17ffSNaoya Horiguchi struct page *page = NULL; 4889e66f17ffSNaoya Horiguchi spinlock_t *ptl; 4890c9d398faSNaoya Horiguchi pte_t pte; 4891e66f17ffSNaoya Horiguchi retry: 4892e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 4893e66f17ffSNaoya Horiguchi spin_lock(ptl); 4894e66f17ffSNaoya Horiguchi /* 4895e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 4896e66f17ffSNaoya Horiguchi * unmapped from other threads. 4897e66f17ffSNaoya Horiguchi */ 4898e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 4899e66f17ffSNaoya Horiguchi goto out; 4900c9d398faSNaoya Horiguchi pte = huge_ptep_get((pte_t *)pmd); 4901c9d398faSNaoya Horiguchi if (pte_present(pte)) { 490297534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 4903e66f17ffSNaoya Horiguchi if (flags & FOLL_GET) 4904e66f17ffSNaoya Horiguchi get_page(page); 4905e66f17ffSNaoya Horiguchi } else { 4906c9d398faSNaoya Horiguchi if (is_hugetlb_entry_migration(pte)) { 4907e66f17ffSNaoya Horiguchi spin_unlock(ptl); 4908e66f17ffSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pmd, ptl); 4909e66f17ffSNaoya Horiguchi goto retry; 4910e66f17ffSNaoya Horiguchi } 4911e66f17ffSNaoya Horiguchi /* 4912e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 4913e66f17ffSNaoya Horiguchi * follow_page_mask(). 4914e66f17ffSNaoya Horiguchi */ 4915e66f17ffSNaoya Horiguchi } 4916e66f17ffSNaoya Horiguchi out: 4917e66f17ffSNaoya Horiguchi spin_unlock(ptl); 49189e5fc74cSSteve Capper return page; 49199e5fc74cSSteve Capper } 49209e5fc74cSSteve Capper 492161f77edaSNaoya Horiguchi struct page * __weak 49229e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 4923e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 49249e5fc74cSSteve Capper { 4925e66f17ffSNaoya Horiguchi if (flags & FOLL_GET) 4926e66f17ffSNaoya Horiguchi return NULL; 49279e5fc74cSSteve Capper 4928e66f17ffSNaoya Horiguchi return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 49299e5fc74cSSteve Capper } 49309e5fc74cSSteve Capper 4931faaa5b62SAnshuman Khandual struct page * __weak 4932faaa5b62SAnshuman Khandual follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) 4933faaa5b62SAnshuman Khandual { 4934faaa5b62SAnshuman Khandual if (flags & FOLL_GET) 4935faaa5b62SAnshuman Khandual return NULL; 4936faaa5b62SAnshuman Khandual 4937faaa5b62SAnshuman Khandual return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); 4938faaa5b62SAnshuman Khandual } 4939faaa5b62SAnshuman Khandual 494031caf665SNaoya Horiguchi bool isolate_huge_page(struct page *page, struct list_head *list) 494131caf665SNaoya Horiguchi { 4942bcc54222SNaoya Horiguchi bool ret = true; 4943bcc54222SNaoya Horiguchi 4944309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 494531caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 4946bcc54222SNaoya Horiguchi if (!page_huge_active(page) || !get_page_unless_zero(page)) { 4947bcc54222SNaoya Horiguchi ret = false; 4948bcc54222SNaoya Horiguchi goto unlock; 4949bcc54222SNaoya Horiguchi } 4950bcc54222SNaoya Horiguchi clear_page_huge_active(page); 495131caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 4952bcc54222SNaoya Horiguchi unlock: 495331caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 4954bcc54222SNaoya Horiguchi return ret; 495531caf665SNaoya Horiguchi } 495631caf665SNaoya Horiguchi 495731caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 495831caf665SNaoya Horiguchi { 4959309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 496031caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 4961bcc54222SNaoya Horiguchi set_page_huge_active(page); 496231caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 496331caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 496431caf665SNaoya Horiguchi put_page(page); 496531caf665SNaoya Horiguchi } 4966ab5ac90aSMichal Hocko 4967ab5ac90aSMichal Hocko void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) 4968ab5ac90aSMichal Hocko { 4969ab5ac90aSMichal Hocko struct hstate *h = page_hstate(oldpage); 4970ab5ac90aSMichal Hocko 4971ab5ac90aSMichal Hocko hugetlb_cgroup_migrate(oldpage, newpage); 4972ab5ac90aSMichal Hocko set_page_owner_migrate_reason(newpage, reason); 4973ab5ac90aSMichal Hocko 4974ab5ac90aSMichal Hocko /* 4975ab5ac90aSMichal Hocko * transfer temporary state of the new huge page. This is 4976ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 4977ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 4978ab5ac90aSMichal Hocko * the temporary status. 4979ab5ac90aSMichal Hocko * 4980ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 4981ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 4982ab5ac90aSMichal Hocko * the per-node's. 4983ab5ac90aSMichal Hocko */ 4984ab5ac90aSMichal Hocko if (PageHugeTemporary(newpage)) { 4985ab5ac90aSMichal Hocko int old_nid = page_to_nid(oldpage); 4986ab5ac90aSMichal Hocko int new_nid = page_to_nid(newpage); 4987ab5ac90aSMichal Hocko 4988ab5ac90aSMichal Hocko SetPageHugeTemporary(oldpage); 4989ab5ac90aSMichal Hocko ClearPageHugeTemporary(newpage); 4990ab5ac90aSMichal Hocko 4991ab5ac90aSMichal Hocko spin_lock(&hugetlb_lock); 4992ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 4993ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 4994ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 4995ab5ac90aSMichal Hocko } 4996ab5ac90aSMichal Hocko spin_unlock(&hugetlb_lock); 4997ab5ac90aSMichal Hocko } 4998ab5ac90aSMichal Hocko } 4999