11da177e4SLinus Torvalds /* 21da177e4SLinus Torvalds * Generic hugetlb support. 36d49e352SNadia Yvette Chambers * (C) Nadia Yvette Chambers, April 2004 41da177e4SLinus Torvalds */ 51da177e4SLinus Torvalds #include <linux/list.h> 61da177e4SLinus Torvalds #include <linux/init.h> 71da177e4SLinus Torvalds #include <linux/module.h> 81da177e4SLinus Torvalds #include <linux/mm.h> 9e1759c21SAlexey Dobriyan #include <linux/seq_file.h> 101da177e4SLinus Torvalds #include <linux/sysctl.h> 111da177e4SLinus Torvalds #include <linux/highmem.h> 12cddb8a5cSAndrea Arcangeli #include <linux/mmu_notifier.h> 131da177e4SLinus Torvalds #include <linux/nodemask.h> 1463551ae0SDavid Gibson #include <linux/pagemap.h> 155da7ca86SChristoph Lameter #include <linux/mempolicy.h> 163b32123dSGideon Israel Dsouza #include <linux/compiler.h> 17aea47ff3SChristoph Lameter #include <linux/cpuset.h> 183935baa9SDavid Gibson #include <linux/mutex.h> 19aa888a74SAndi Kleen #include <linux/bootmem.h> 20a3437870SNishanth Aravamudan #include <linux/sysfs.h> 215a0e3ad6STejun Heo #include <linux/slab.h> 220fe6e20bSNaoya Horiguchi #include <linux/rmap.h> 23fd6a03edSNaoya Horiguchi #include <linux/swap.h> 24fd6a03edSNaoya Horiguchi #include <linux/swapops.h> 25c8721bbbSNaoya Horiguchi #include <linux/page-isolation.h> 268382d914SDavidlohr Bueso #include <linux/jhash.h> 27d6606683SLinus Torvalds 2863551ae0SDavid Gibson #include <asm/page.h> 2963551ae0SDavid Gibson #include <asm/pgtable.h> 3024669e58SAneesh Kumar K.V #include <asm/tlb.h> 3163551ae0SDavid Gibson 3224669e58SAneesh Kumar K.V #include <linux/io.h> 3363551ae0SDavid Gibson #include <linux/hugetlb.h> 349dd540e2SAneesh Kumar K.V #include <linux/hugetlb_cgroup.h> 359a305230SLee Schermerhorn #include <linux/node.h> 367835e98bSNick Piggin #include "internal.h" 371da177e4SLinus Torvalds 38753162cdSAndrey Ryabinin int hugepages_treat_as_movable; 39a5516438SAndi Kleen 40c3f38a38SAneesh Kumar K.V int hugetlb_max_hstate __read_mostly; 41e5ff2159SAndi Kleen unsigned int default_hstate_idx; 42e5ff2159SAndi Kleen struct hstate hstates[HUGE_MAX_HSTATE]; 43641844f5SNaoya Horiguchi /* 44641844f5SNaoya Horiguchi * Minimum page order among possible hugepage sizes, set to a proper value 45641844f5SNaoya Horiguchi * at boot time. 46641844f5SNaoya Horiguchi */ 47641844f5SNaoya Horiguchi static unsigned int minimum_order __read_mostly = UINT_MAX; 48e5ff2159SAndi Kleen 4953ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 5053ba51d2SJon Tollefson 51e5ff2159SAndi Kleen /* for command line parsing */ 52e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 53e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 54e11bfbfcSNick Piggin static unsigned long __initdata default_hstate_size; 55e5ff2159SAndi Kleen 563935baa9SDavid Gibson /* 5731caf665SNaoya Horiguchi * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, 5831caf665SNaoya Horiguchi * free_huge_pages, and surplus_huge_pages. 593935baa9SDavid Gibson */ 60c3f38a38SAneesh Kumar K.V DEFINE_SPINLOCK(hugetlb_lock); 610bd0f9fbSEric Paris 628382d914SDavidlohr Bueso /* 638382d914SDavidlohr Bueso * Serializes faults on the same logical page. This is used to 648382d914SDavidlohr Bueso * prevent spurious OOMs when the hugepage pool is fully utilized. 658382d914SDavidlohr Bueso */ 668382d914SDavidlohr Bueso static int num_fault_mutexes; 678382d914SDavidlohr Bueso static struct mutex *htlb_fault_mutex_table ____cacheline_aligned_in_smp; 688382d914SDavidlohr Bueso 697ca02d0aSMike Kravetz /* Forward declaration */ 707ca02d0aSMike Kravetz static int hugetlb_acct_memory(struct hstate *h, long delta); 717ca02d0aSMike Kravetz 7290481622SDavid Gibson static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) 7390481622SDavid Gibson { 7490481622SDavid Gibson bool free = (spool->count == 0) && (spool->used_hpages == 0); 7590481622SDavid Gibson 7690481622SDavid Gibson spin_unlock(&spool->lock); 7790481622SDavid Gibson 7890481622SDavid Gibson /* If no pages are used, and no other handles to the subpool 797ca02d0aSMike Kravetz * remain, give up any reservations mased on minimum size and 807ca02d0aSMike Kravetz * free the subpool */ 817ca02d0aSMike Kravetz if (free) { 827ca02d0aSMike Kravetz if (spool->min_hpages != -1) 837ca02d0aSMike Kravetz hugetlb_acct_memory(spool->hstate, 847ca02d0aSMike Kravetz -spool->min_hpages); 8590481622SDavid Gibson kfree(spool); 8690481622SDavid Gibson } 877ca02d0aSMike Kravetz } 8890481622SDavid Gibson 897ca02d0aSMike Kravetz struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, 907ca02d0aSMike Kravetz long min_hpages) 9190481622SDavid Gibson { 9290481622SDavid Gibson struct hugepage_subpool *spool; 9390481622SDavid Gibson 94c6a91820SMike Kravetz spool = kzalloc(sizeof(*spool), GFP_KERNEL); 9590481622SDavid Gibson if (!spool) 9690481622SDavid Gibson return NULL; 9790481622SDavid Gibson 9890481622SDavid Gibson spin_lock_init(&spool->lock); 9990481622SDavid Gibson spool->count = 1; 1007ca02d0aSMike Kravetz spool->max_hpages = max_hpages; 1017ca02d0aSMike Kravetz spool->hstate = h; 1027ca02d0aSMike Kravetz spool->min_hpages = min_hpages; 1037ca02d0aSMike Kravetz 1047ca02d0aSMike Kravetz if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { 1057ca02d0aSMike Kravetz kfree(spool); 1067ca02d0aSMike Kravetz return NULL; 1077ca02d0aSMike Kravetz } 1087ca02d0aSMike Kravetz spool->rsv_hpages = min_hpages; 10990481622SDavid Gibson 11090481622SDavid Gibson return spool; 11190481622SDavid Gibson } 11290481622SDavid Gibson 11390481622SDavid Gibson void hugepage_put_subpool(struct hugepage_subpool *spool) 11490481622SDavid Gibson { 11590481622SDavid Gibson spin_lock(&spool->lock); 11690481622SDavid Gibson BUG_ON(!spool->count); 11790481622SDavid Gibson spool->count--; 11890481622SDavid Gibson unlock_or_release_subpool(spool); 11990481622SDavid Gibson } 12090481622SDavid Gibson 1211c5ecae3SMike Kravetz /* 1221c5ecae3SMike Kravetz * Subpool accounting for allocating and reserving pages. 1231c5ecae3SMike Kravetz * Return -ENOMEM if there are not enough resources to satisfy the 1241c5ecae3SMike Kravetz * the request. Otherwise, return the number of pages by which the 1251c5ecae3SMike Kravetz * global pools must be adjusted (upward). The returned value may 1261c5ecae3SMike Kravetz * only be different than the passed value (delta) in the case where 1271c5ecae3SMike Kravetz * a subpool minimum size must be manitained. 1281c5ecae3SMike Kravetz */ 1291c5ecae3SMike Kravetz static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, 13090481622SDavid Gibson long delta) 13190481622SDavid Gibson { 1321c5ecae3SMike Kravetz long ret = delta; 13390481622SDavid Gibson 13490481622SDavid Gibson if (!spool) 1351c5ecae3SMike Kravetz return ret; 13690481622SDavid Gibson 13790481622SDavid Gibson spin_lock(&spool->lock); 13890481622SDavid Gibson 1391c5ecae3SMike Kravetz if (spool->max_hpages != -1) { /* maximum size accounting */ 1401c5ecae3SMike Kravetz if ((spool->used_hpages + delta) <= spool->max_hpages) 1411c5ecae3SMike Kravetz spool->used_hpages += delta; 1421c5ecae3SMike Kravetz else { 1431c5ecae3SMike Kravetz ret = -ENOMEM; 1441c5ecae3SMike Kravetz goto unlock_ret; 1451c5ecae3SMike Kravetz } 1461c5ecae3SMike Kravetz } 1471c5ecae3SMike Kravetz 1481c5ecae3SMike Kravetz if (spool->min_hpages != -1) { /* minimum size accounting */ 1491c5ecae3SMike Kravetz if (delta > spool->rsv_hpages) { 1501c5ecae3SMike Kravetz /* 1511c5ecae3SMike Kravetz * Asking for more reserves than those already taken on 1521c5ecae3SMike Kravetz * behalf of subpool. Return difference. 1531c5ecae3SMike Kravetz */ 1541c5ecae3SMike Kravetz ret = delta - spool->rsv_hpages; 1551c5ecae3SMike Kravetz spool->rsv_hpages = 0; 1561c5ecae3SMike Kravetz } else { 1571c5ecae3SMike Kravetz ret = 0; /* reserves already accounted for */ 1581c5ecae3SMike Kravetz spool->rsv_hpages -= delta; 1591c5ecae3SMike Kravetz } 1601c5ecae3SMike Kravetz } 1611c5ecae3SMike Kravetz 1621c5ecae3SMike Kravetz unlock_ret: 1631c5ecae3SMike Kravetz spin_unlock(&spool->lock); 16490481622SDavid Gibson return ret; 16590481622SDavid Gibson } 16690481622SDavid Gibson 1671c5ecae3SMike Kravetz /* 1681c5ecae3SMike Kravetz * Subpool accounting for freeing and unreserving pages. 1691c5ecae3SMike Kravetz * Return the number of global page reservations that must be dropped. 1701c5ecae3SMike Kravetz * The return value may only be different than the passed value (delta) 1711c5ecae3SMike Kravetz * in the case where a subpool minimum size must be maintained. 1721c5ecae3SMike Kravetz */ 1731c5ecae3SMike Kravetz static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, 17490481622SDavid Gibson long delta) 17590481622SDavid Gibson { 1761c5ecae3SMike Kravetz long ret = delta; 1771c5ecae3SMike Kravetz 17890481622SDavid Gibson if (!spool) 1791c5ecae3SMike Kravetz return delta; 18090481622SDavid Gibson 18190481622SDavid Gibson spin_lock(&spool->lock); 1821c5ecae3SMike Kravetz 1831c5ecae3SMike Kravetz if (spool->max_hpages != -1) /* maximum size accounting */ 18490481622SDavid Gibson spool->used_hpages -= delta; 1851c5ecae3SMike Kravetz 1861c5ecae3SMike Kravetz if (spool->min_hpages != -1) { /* minimum size accounting */ 1871c5ecae3SMike Kravetz if (spool->rsv_hpages + delta <= spool->min_hpages) 1881c5ecae3SMike Kravetz ret = 0; 1891c5ecae3SMike Kravetz else 1901c5ecae3SMike Kravetz ret = spool->rsv_hpages + delta - spool->min_hpages; 1911c5ecae3SMike Kravetz 1921c5ecae3SMike Kravetz spool->rsv_hpages += delta; 1931c5ecae3SMike Kravetz if (spool->rsv_hpages > spool->min_hpages) 1941c5ecae3SMike Kravetz spool->rsv_hpages = spool->min_hpages; 1951c5ecae3SMike Kravetz } 1961c5ecae3SMike Kravetz 1971c5ecae3SMike Kravetz /* 1981c5ecae3SMike Kravetz * If hugetlbfs_put_super couldn't free spool due to an outstanding 1991c5ecae3SMike Kravetz * quota reference, free it now. 2001c5ecae3SMike Kravetz */ 20190481622SDavid Gibson unlock_or_release_subpool(spool); 2021c5ecae3SMike Kravetz 2031c5ecae3SMike Kravetz return ret; 20490481622SDavid Gibson } 20590481622SDavid Gibson 20690481622SDavid Gibson static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 20790481622SDavid Gibson { 20890481622SDavid Gibson return HUGETLBFS_SB(inode->i_sb)->spool; 20990481622SDavid Gibson } 21090481622SDavid Gibson 21190481622SDavid Gibson static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 21290481622SDavid Gibson { 213496ad9aaSAl Viro return subpool_inode(file_inode(vma->vm_file)); 21490481622SDavid Gibson } 21590481622SDavid Gibson 216e7c4b0bfSAndy Whitcroft /* 21796822904SAndy Whitcroft * Region tracking -- allows tracking of reservations and instantiated pages 21896822904SAndy Whitcroft * across the pages in a mapping. 21984afd99bSAndy Whitcroft * 2201dd308a7SMike Kravetz * The region data structures are embedded into a resv_map and protected 2211dd308a7SMike Kravetz * by a resv_map's lock. The set of regions within the resv_map represent 2221dd308a7SMike Kravetz * reservations for huge pages, or huge pages that have already been 2231dd308a7SMike Kravetz * instantiated within the map. The from and to elements are huge page 2241dd308a7SMike Kravetz * indicies into the associated mapping. from indicates the starting index 2251dd308a7SMike Kravetz * of the region. to represents the first index past the end of the region. 2261dd308a7SMike Kravetz * 2271dd308a7SMike Kravetz * For example, a file region structure with from == 0 and to == 4 represents 2281dd308a7SMike Kravetz * four huge pages in a mapping. It is important to note that the to element 2291dd308a7SMike Kravetz * represents the first element past the end of the region. This is used in 2301dd308a7SMike Kravetz * arithmetic as 4(to) - 0(from) = 4 huge pages in the region. 2311dd308a7SMike Kravetz * 2321dd308a7SMike Kravetz * Interval notation of the form [from, to) will be used to indicate that 2331dd308a7SMike Kravetz * the endpoint from is inclusive and to is exclusive. 23496822904SAndy Whitcroft */ 23596822904SAndy Whitcroft struct file_region { 23696822904SAndy Whitcroft struct list_head link; 23796822904SAndy Whitcroft long from; 23896822904SAndy Whitcroft long to; 23996822904SAndy Whitcroft }; 24096822904SAndy Whitcroft 2411dd308a7SMike Kravetz /* 2421dd308a7SMike Kravetz * Add the huge page range represented by [f, t) to the reserve 2431dd308a7SMike Kravetz * map. Existing regions will be expanded to accommodate the 2441dd308a7SMike Kravetz * specified range. We know only existing regions need to be 2451dd308a7SMike Kravetz * expanded, because region_add is only called after region_chg 2461dd308a7SMike Kravetz * with the same range. If a new file_region structure must 2471dd308a7SMike Kravetz * be allocated, it is done in region_chg. 2481dd308a7SMike Kravetz */ 2491406ec9bSJoonsoo Kim static long region_add(struct resv_map *resv, long f, long t) 25096822904SAndy Whitcroft { 2511406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 25296822904SAndy Whitcroft struct file_region *rg, *nrg, *trg; 25396822904SAndy Whitcroft 2547b24d861SDavidlohr Bueso spin_lock(&resv->lock); 25596822904SAndy Whitcroft /* Locate the region we are either in or before. */ 25696822904SAndy Whitcroft list_for_each_entry(rg, head, link) 25796822904SAndy Whitcroft if (f <= rg->to) 25896822904SAndy Whitcroft break; 25996822904SAndy Whitcroft 26096822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 26196822904SAndy Whitcroft if (f > rg->from) 26296822904SAndy Whitcroft f = rg->from; 26396822904SAndy Whitcroft 26496822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 26596822904SAndy Whitcroft nrg = rg; 26696822904SAndy Whitcroft list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 26796822904SAndy Whitcroft if (&rg->link == head) 26896822904SAndy Whitcroft break; 26996822904SAndy Whitcroft if (rg->from > t) 27096822904SAndy Whitcroft break; 27196822904SAndy Whitcroft 27296822904SAndy Whitcroft /* If this area reaches higher then extend our area to 27396822904SAndy Whitcroft * include it completely. If this is not the first area 27496822904SAndy Whitcroft * which we intend to reuse, free it. */ 27596822904SAndy Whitcroft if (rg->to > t) 27696822904SAndy Whitcroft t = rg->to; 27796822904SAndy Whitcroft if (rg != nrg) { 27896822904SAndy Whitcroft list_del(&rg->link); 27996822904SAndy Whitcroft kfree(rg); 28096822904SAndy Whitcroft } 28196822904SAndy Whitcroft } 28296822904SAndy Whitcroft nrg->from = f; 28396822904SAndy Whitcroft nrg->to = t; 2847b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 28596822904SAndy Whitcroft return 0; 28696822904SAndy Whitcroft } 28796822904SAndy Whitcroft 2881dd308a7SMike Kravetz /* 2891dd308a7SMike Kravetz * Examine the existing reserve map and determine how many 2901dd308a7SMike Kravetz * huge pages in the specified range [f, t) are NOT currently 2911dd308a7SMike Kravetz * represented. This routine is called before a subsequent 2921dd308a7SMike Kravetz * call to region_add that will actually modify the reserve 2931dd308a7SMike Kravetz * map to add the specified range [f, t). region_chg does 2941dd308a7SMike Kravetz * not change the number of huge pages represented by the 2951dd308a7SMike Kravetz * map. However, if the existing regions in the map can not 2961dd308a7SMike Kravetz * be expanded to represent the new range, a new file_region 2971dd308a7SMike Kravetz * structure is added to the map as a placeholder. This is 2981dd308a7SMike Kravetz * so that the subsequent region_add call will have all the 2991dd308a7SMike Kravetz * regions it needs and will not fail. 3001dd308a7SMike Kravetz * 3011dd308a7SMike Kravetz * Returns the number of huge pages that need to be added 3021dd308a7SMike Kravetz * to the existing reservation map for the range [f, t). 3031dd308a7SMike Kravetz * This number is greater or equal to zero. -ENOMEM is 3041dd308a7SMike Kravetz * returned if a new file_region structure is needed and can 3051dd308a7SMike Kravetz * not be allocated. 3061dd308a7SMike Kravetz */ 3071406ec9bSJoonsoo Kim static long region_chg(struct resv_map *resv, long f, long t) 30896822904SAndy Whitcroft { 3091406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 3107b24d861SDavidlohr Bueso struct file_region *rg, *nrg = NULL; 31196822904SAndy Whitcroft long chg = 0; 31296822904SAndy Whitcroft 3137b24d861SDavidlohr Bueso retry: 3147b24d861SDavidlohr Bueso spin_lock(&resv->lock); 31596822904SAndy Whitcroft /* Locate the region we are before or in. */ 31696822904SAndy Whitcroft list_for_each_entry(rg, head, link) 31796822904SAndy Whitcroft if (f <= rg->to) 31896822904SAndy Whitcroft break; 31996822904SAndy Whitcroft 32096822904SAndy Whitcroft /* If we are below the current region then a new region is required. 32196822904SAndy Whitcroft * Subtle, allocate a new region at the position but make it zero 32296822904SAndy Whitcroft * size such that we can guarantee to record the reservation. */ 32396822904SAndy Whitcroft if (&rg->link == head || t < rg->from) { 3247b24d861SDavidlohr Bueso if (!nrg) { 3257b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 32696822904SAndy Whitcroft nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 32796822904SAndy Whitcroft if (!nrg) 32896822904SAndy Whitcroft return -ENOMEM; 3297b24d861SDavidlohr Bueso 33096822904SAndy Whitcroft nrg->from = f; 33196822904SAndy Whitcroft nrg->to = f; 33296822904SAndy Whitcroft INIT_LIST_HEAD(&nrg->link); 3337b24d861SDavidlohr Bueso goto retry; 3347b24d861SDavidlohr Bueso } 33596822904SAndy Whitcroft 3367b24d861SDavidlohr Bueso list_add(&nrg->link, rg->link.prev); 3377b24d861SDavidlohr Bueso chg = t - f; 3387b24d861SDavidlohr Bueso goto out_nrg; 33996822904SAndy Whitcroft } 34096822904SAndy Whitcroft 34196822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 34296822904SAndy Whitcroft if (f > rg->from) 34396822904SAndy Whitcroft f = rg->from; 34496822904SAndy Whitcroft chg = t - f; 34596822904SAndy Whitcroft 34696822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 34796822904SAndy Whitcroft list_for_each_entry(rg, rg->link.prev, link) { 34896822904SAndy Whitcroft if (&rg->link == head) 34996822904SAndy Whitcroft break; 35096822904SAndy Whitcroft if (rg->from > t) 3517b24d861SDavidlohr Bueso goto out; 35296822904SAndy Whitcroft 35325985edcSLucas De Marchi /* We overlap with this area, if it extends further than 35496822904SAndy Whitcroft * us then we must extend ourselves. Account for its 35596822904SAndy Whitcroft * existing reservation. */ 35696822904SAndy Whitcroft if (rg->to > t) { 35796822904SAndy Whitcroft chg += rg->to - t; 35896822904SAndy Whitcroft t = rg->to; 35996822904SAndy Whitcroft } 36096822904SAndy Whitcroft chg -= rg->to - rg->from; 36196822904SAndy Whitcroft } 3627b24d861SDavidlohr Bueso 3637b24d861SDavidlohr Bueso out: 3647b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 3657b24d861SDavidlohr Bueso /* We already know we raced and no longer need the new region */ 3667b24d861SDavidlohr Bueso kfree(nrg); 3677b24d861SDavidlohr Bueso return chg; 3687b24d861SDavidlohr Bueso out_nrg: 3697b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 37096822904SAndy Whitcroft return chg; 37196822904SAndy Whitcroft } 37296822904SAndy Whitcroft 3731dd308a7SMike Kravetz /* 3741dd308a7SMike Kravetz * Truncate the reserve map at index 'end'. Modify/truncate any 3751dd308a7SMike Kravetz * region which contains end. Delete any regions past end. 3761dd308a7SMike Kravetz * Return the number of huge pages removed from the map. 3771dd308a7SMike Kravetz */ 3781406ec9bSJoonsoo Kim static long region_truncate(struct resv_map *resv, long end) 37996822904SAndy Whitcroft { 3801406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 38196822904SAndy Whitcroft struct file_region *rg, *trg; 38296822904SAndy Whitcroft long chg = 0; 38396822904SAndy Whitcroft 3847b24d861SDavidlohr Bueso spin_lock(&resv->lock); 38596822904SAndy Whitcroft /* Locate the region we are either in or before. */ 38696822904SAndy Whitcroft list_for_each_entry(rg, head, link) 38796822904SAndy Whitcroft if (end <= rg->to) 38896822904SAndy Whitcroft break; 38996822904SAndy Whitcroft if (&rg->link == head) 3907b24d861SDavidlohr Bueso goto out; 39196822904SAndy Whitcroft 39296822904SAndy Whitcroft /* If we are in the middle of a region then adjust it. */ 39396822904SAndy Whitcroft if (end > rg->from) { 39496822904SAndy Whitcroft chg = rg->to - end; 39596822904SAndy Whitcroft rg->to = end; 39696822904SAndy Whitcroft rg = list_entry(rg->link.next, typeof(*rg), link); 39796822904SAndy Whitcroft } 39896822904SAndy Whitcroft 39996822904SAndy Whitcroft /* Drop any remaining regions. */ 40096822904SAndy Whitcroft list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 40196822904SAndy Whitcroft if (&rg->link == head) 40296822904SAndy Whitcroft break; 40396822904SAndy Whitcroft chg += rg->to - rg->from; 40496822904SAndy Whitcroft list_del(&rg->link); 40596822904SAndy Whitcroft kfree(rg); 40696822904SAndy Whitcroft } 4077b24d861SDavidlohr Bueso 4087b24d861SDavidlohr Bueso out: 4097b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 41096822904SAndy Whitcroft return chg; 41196822904SAndy Whitcroft } 41296822904SAndy Whitcroft 4131dd308a7SMike Kravetz /* 4141dd308a7SMike Kravetz * Count and return the number of huge pages in the reserve map 4151dd308a7SMike Kravetz * that intersect with the range [f, t). 4161dd308a7SMike Kravetz */ 4171406ec9bSJoonsoo Kim static long region_count(struct resv_map *resv, long f, long t) 41884afd99bSAndy Whitcroft { 4191406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 42084afd99bSAndy Whitcroft struct file_region *rg; 42184afd99bSAndy Whitcroft long chg = 0; 42284afd99bSAndy Whitcroft 4237b24d861SDavidlohr Bueso spin_lock(&resv->lock); 42484afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 42584afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 426f2135a4aSWang Sheng-Hui long seg_from; 427f2135a4aSWang Sheng-Hui long seg_to; 42884afd99bSAndy Whitcroft 42984afd99bSAndy Whitcroft if (rg->to <= f) 43084afd99bSAndy Whitcroft continue; 43184afd99bSAndy Whitcroft if (rg->from >= t) 43284afd99bSAndy Whitcroft break; 43384afd99bSAndy Whitcroft 43484afd99bSAndy Whitcroft seg_from = max(rg->from, f); 43584afd99bSAndy Whitcroft seg_to = min(rg->to, t); 43684afd99bSAndy Whitcroft 43784afd99bSAndy Whitcroft chg += seg_to - seg_from; 43884afd99bSAndy Whitcroft } 4397b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 44084afd99bSAndy Whitcroft 44184afd99bSAndy Whitcroft return chg; 44284afd99bSAndy Whitcroft } 44384afd99bSAndy Whitcroft 44496822904SAndy Whitcroft /* 445e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 446e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 447e7c4b0bfSAndy Whitcroft */ 448a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 449a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 450e7c4b0bfSAndy Whitcroft { 451a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 452a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 453e7c4b0bfSAndy Whitcroft } 454e7c4b0bfSAndy Whitcroft 4550fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 4560fe6e20bSNaoya Horiguchi unsigned long address) 4570fe6e20bSNaoya Horiguchi { 4580fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 4590fe6e20bSNaoya Horiguchi } 4600fe6e20bSNaoya Horiguchi 46184afd99bSAndy Whitcroft /* 46208fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 46308fba699SMel Gorman * cases this will be same size as used by the page table entries. 46408fba699SMel Gorman */ 46508fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 46608fba699SMel Gorman { 46708fba699SMel Gorman struct hstate *hstate; 46808fba699SMel Gorman 46908fba699SMel Gorman if (!is_vm_hugetlb_page(vma)) 47008fba699SMel Gorman return PAGE_SIZE; 47108fba699SMel Gorman 47208fba699SMel Gorman hstate = hstate_vma(vma); 47308fba699SMel Gorman 4742415cf12SWanpeng Li return 1UL << huge_page_shift(hstate); 47508fba699SMel Gorman } 476f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 47708fba699SMel Gorman 47808fba699SMel Gorman /* 4793340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 4803340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 4813340289dSMel Gorman * architectures where it differs, an architecture-specific version of this 4823340289dSMel Gorman * function is required. 4833340289dSMel Gorman */ 4843340289dSMel Gorman #ifndef vma_mmu_pagesize 4853340289dSMel Gorman unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 4863340289dSMel Gorman { 4873340289dSMel Gorman return vma_kernel_pagesize(vma); 4883340289dSMel Gorman } 4893340289dSMel Gorman #endif 4903340289dSMel Gorman 4913340289dSMel Gorman /* 49284afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 49384afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 49484afd99bSAndy Whitcroft * alignment. 49584afd99bSAndy Whitcroft */ 49684afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 49784afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 49804f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 49984afd99bSAndy Whitcroft 500a1e78772SMel Gorman /* 501a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 502a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 503a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 504a1e78772SMel Gorman * 505a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 506a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 507a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 508a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 50984afd99bSAndy Whitcroft * 51084afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 51184afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 51284afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 51384afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 51484afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 51584afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 51684afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 51784afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 518a1e78772SMel Gorman */ 519e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 520e7c4b0bfSAndy Whitcroft { 521e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 522e7c4b0bfSAndy Whitcroft } 523e7c4b0bfSAndy Whitcroft 524e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 525e7c4b0bfSAndy Whitcroft unsigned long value) 526e7c4b0bfSAndy Whitcroft { 527e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 528e7c4b0bfSAndy Whitcroft } 529e7c4b0bfSAndy Whitcroft 5309119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 53184afd99bSAndy Whitcroft { 53284afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 53384afd99bSAndy Whitcroft if (!resv_map) 53484afd99bSAndy Whitcroft return NULL; 53584afd99bSAndy Whitcroft 53684afd99bSAndy Whitcroft kref_init(&resv_map->refs); 5377b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 53884afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 53984afd99bSAndy Whitcroft 54084afd99bSAndy Whitcroft return resv_map; 54184afd99bSAndy Whitcroft } 54284afd99bSAndy Whitcroft 5439119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 54484afd99bSAndy Whitcroft { 54584afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 54684afd99bSAndy Whitcroft 54784afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 5481406ec9bSJoonsoo Kim region_truncate(resv_map, 0); 54984afd99bSAndy Whitcroft kfree(resv_map); 55084afd99bSAndy Whitcroft } 55184afd99bSAndy Whitcroft 5524e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 5534e35f483SJoonsoo Kim { 5544e35f483SJoonsoo Kim return inode->i_mapping->private_data; 5554e35f483SJoonsoo Kim } 5564e35f483SJoonsoo Kim 55784afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 558a1e78772SMel Gorman { 55981d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 5604e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 5614e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 5624e35f483SJoonsoo Kim struct inode *inode = mapping->host; 5634e35f483SJoonsoo Kim 5644e35f483SJoonsoo Kim return inode_resv_map(inode); 5654e35f483SJoonsoo Kim 5664e35f483SJoonsoo Kim } else { 56784afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 56884afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 5694e35f483SJoonsoo Kim } 570a1e78772SMel Gorman } 571a1e78772SMel Gorman 57284afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 573a1e78772SMel Gorman { 57481d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 57581d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 576a1e78772SMel Gorman 57784afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 57884afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 57904f2cbe3SMel Gorman } 58004f2cbe3SMel Gorman 58104f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 58204f2cbe3SMel Gorman { 58381d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 58481d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 585e7c4b0bfSAndy Whitcroft 586e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 58704f2cbe3SMel Gorman } 58804f2cbe3SMel Gorman 58904f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 59004f2cbe3SMel Gorman { 59181d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 592e7c4b0bfSAndy Whitcroft 593e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 594a1e78772SMel Gorman } 595a1e78772SMel Gorman 59604f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 597a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 598a1e78772SMel Gorman { 59981d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 600f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 601a1e78772SMel Gorman vma->vm_private_data = (void *)0; 602a1e78772SMel Gorman } 603a1e78772SMel Gorman 604a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 605af0ed73eSJoonsoo Kim static int vma_has_reserves(struct vm_area_struct *vma, long chg) 606a1e78772SMel Gorman { 607af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 608af0ed73eSJoonsoo Kim /* 609af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 610af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 611af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 612af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 613af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 614af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 615af0ed73eSJoonsoo Kim * properly, so add work-around here. 616af0ed73eSJoonsoo Kim */ 617af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 618af0ed73eSJoonsoo Kim return 1; 619af0ed73eSJoonsoo Kim else 62072231b03SJoonsoo Kim return 0; 621af0ed73eSJoonsoo Kim } 622a63884e9SJoonsoo Kim 623a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 624f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) 625a1e78772SMel Gorman return 1; 626a63884e9SJoonsoo Kim 627a63884e9SJoonsoo Kim /* 628a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 629a63884e9SJoonsoo Kim * private mappings. 630a63884e9SJoonsoo Kim */ 6317f09ca51SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 6327f09ca51SMel Gorman return 1; 633a63884e9SJoonsoo Kim 6347f09ca51SMel Gorman return 0; 635a1e78772SMel Gorman } 636a1e78772SMel Gorman 637a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 6381da177e4SLinus Torvalds { 6391da177e4SLinus Torvalds int nid = page_to_nid(page); 6400edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 641a5516438SAndi Kleen h->free_huge_pages++; 642a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 6431da177e4SLinus Torvalds } 6441da177e4SLinus Torvalds 645bf50bab2SNaoya Horiguchi static struct page *dequeue_huge_page_node(struct hstate *h, int nid) 646bf50bab2SNaoya Horiguchi { 647bf50bab2SNaoya Horiguchi struct page *page; 648bf50bab2SNaoya Horiguchi 649c8721bbbSNaoya Horiguchi list_for_each_entry(page, &h->hugepage_freelists[nid], lru) 650c8721bbbSNaoya Horiguchi if (!is_migrate_isolate_page(page)) 651c8721bbbSNaoya Horiguchi break; 652c8721bbbSNaoya Horiguchi /* 653c8721bbbSNaoya Horiguchi * if 'non-isolated free hugepage' not found on the list, 654c8721bbbSNaoya Horiguchi * the allocation fails. 655c8721bbbSNaoya Horiguchi */ 656c8721bbbSNaoya Horiguchi if (&h->hugepage_freelists[nid] == &page->lru) 657bf50bab2SNaoya Horiguchi return NULL; 6580edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 659a9869b83SNaoya Horiguchi set_page_refcounted(page); 660bf50bab2SNaoya Horiguchi h->free_huge_pages--; 661bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 662bf50bab2SNaoya Horiguchi return page; 663bf50bab2SNaoya Horiguchi } 664bf50bab2SNaoya Horiguchi 66586cdb465SNaoya Horiguchi /* Movability of hugepages depends on migration support. */ 66686cdb465SNaoya Horiguchi static inline gfp_t htlb_alloc_mask(struct hstate *h) 66786cdb465SNaoya Horiguchi { 668100873d7SNaoya Horiguchi if (hugepages_treat_as_movable || hugepage_migration_supported(h)) 66986cdb465SNaoya Horiguchi return GFP_HIGHUSER_MOVABLE; 67086cdb465SNaoya Horiguchi else 67186cdb465SNaoya Horiguchi return GFP_HIGHUSER; 67286cdb465SNaoya Horiguchi } 67386cdb465SNaoya Horiguchi 674a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 675a5516438SAndi Kleen struct vm_area_struct *vma, 676af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 677af0ed73eSJoonsoo Kim long chg) 6781da177e4SLinus Torvalds { 679b1c12cbcSKonstantin Khlebnikov struct page *page = NULL; 680480eccf9SLee Schermerhorn struct mempolicy *mpol; 68119770b32SMel Gorman nodemask_t *nodemask; 682c0ff7453SMiao Xie struct zonelist *zonelist; 683dd1a239fSMel Gorman struct zone *zone; 684dd1a239fSMel Gorman struct zoneref *z; 685cc9a6c87SMel Gorman unsigned int cpuset_mems_cookie; 6861da177e4SLinus Torvalds 687a1e78772SMel Gorman /* 688a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 689a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 690a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 691a1e78772SMel Gorman */ 692af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg) && 693a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 694c0ff7453SMiao Xie goto err; 695a1e78772SMel Gorman 69604f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 697a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 6986eab04a8SJustin P. Mattock goto err; 69904f2cbe3SMel Gorman 7009966c4bbSJoonsoo Kim retry_cpuset: 701d26914d1SMel Gorman cpuset_mems_cookie = read_mems_allowed_begin(); 7029966c4bbSJoonsoo Kim zonelist = huge_zonelist(vma, address, 70386cdb465SNaoya Horiguchi htlb_alloc_mask(h), &mpol, &nodemask); 7049966c4bbSJoonsoo Kim 70519770b32SMel Gorman for_each_zone_zonelist_nodemask(zone, z, zonelist, 70619770b32SMel Gorman MAX_NR_ZONES - 1, nodemask) { 707344736f2SVladimir Davydov if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) { 708bf50bab2SNaoya Horiguchi page = dequeue_huge_page_node(h, zone_to_nid(zone)); 709bf50bab2SNaoya Horiguchi if (page) { 710af0ed73eSJoonsoo Kim if (avoid_reserve) 711af0ed73eSJoonsoo Kim break; 712af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg)) 713af0ed73eSJoonsoo Kim break; 714af0ed73eSJoonsoo Kim 71507443a85SJoonsoo Kim SetPagePrivate(page); 716a63884e9SJoonsoo Kim h->resv_huge_pages--; 7175ab3ee7bSKen Chen break; 7181da177e4SLinus Torvalds } 7193abf7afdSAndrew Morton } 720bf50bab2SNaoya Horiguchi } 721cc9a6c87SMel Gorman 722cc9a6c87SMel Gorman mpol_cond_put(mpol); 723d26914d1SMel Gorman if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) 724cc9a6c87SMel Gorman goto retry_cpuset; 725cc9a6c87SMel Gorman return page; 726cc9a6c87SMel Gorman 727c0ff7453SMiao Xie err: 728cc9a6c87SMel Gorman return NULL; 7291da177e4SLinus Torvalds } 7301da177e4SLinus Torvalds 7311cac6f2cSLuiz Capitulino /* 7321cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 7331cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 7341cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 7351cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 7361cac6f2cSLuiz Capitulino * node for alloc or free. 7371cac6f2cSLuiz Capitulino */ 7381cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 7391cac6f2cSLuiz Capitulino { 7401cac6f2cSLuiz Capitulino nid = next_node(nid, *nodes_allowed); 7411cac6f2cSLuiz Capitulino if (nid == MAX_NUMNODES) 7421cac6f2cSLuiz Capitulino nid = first_node(*nodes_allowed); 7431cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 7441cac6f2cSLuiz Capitulino 7451cac6f2cSLuiz Capitulino return nid; 7461cac6f2cSLuiz Capitulino } 7471cac6f2cSLuiz Capitulino 7481cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 7491cac6f2cSLuiz Capitulino { 7501cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 7511cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 7521cac6f2cSLuiz Capitulino return nid; 7531cac6f2cSLuiz Capitulino } 7541cac6f2cSLuiz Capitulino 7551cac6f2cSLuiz Capitulino /* 7561cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 7571cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 7581cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 7591cac6f2cSLuiz Capitulino * mask. 7601cac6f2cSLuiz Capitulino */ 7611cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 7621cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 7631cac6f2cSLuiz Capitulino { 7641cac6f2cSLuiz Capitulino int nid; 7651cac6f2cSLuiz Capitulino 7661cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 7671cac6f2cSLuiz Capitulino 7681cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 7691cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 7701cac6f2cSLuiz Capitulino 7711cac6f2cSLuiz Capitulino return nid; 7721cac6f2cSLuiz Capitulino } 7731cac6f2cSLuiz Capitulino 7741cac6f2cSLuiz Capitulino /* 7751cac6f2cSLuiz Capitulino * helper for free_pool_huge_page() - return the previously saved 7761cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 7771cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 7781cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 7791cac6f2cSLuiz Capitulino */ 7801cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 7811cac6f2cSLuiz Capitulino { 7821cac6f2cSLuiz Capitulino int nid; 7831cac6f2cSLuiz Capitulino 7841cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 7851cac6f2cSLuiz Capitulino 7861cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 7871cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 7881cac6f2cSLuiz Capitulino 7891cac6f2cSLuiz Capitulino return nid; 7901cac6f2cSLuiz Capitulino } 7911cac6f2cSLuiz Capitulino 7921cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 7931cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 7941cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 7951cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 7961cac6f2cSLuiz Capitulino nr_nodes--) 7971cac6f2cSLuiz Capitulino 7981cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 7991cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 8001cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 8011cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 8021cac6f2cSLuiz Capitulino nr_nodes--) 8031cac6f2cSLuiz Capitulino 804944d9fecSLuiz Capitulino #if defined(CONFIG_CMA) && defined(CONFIG_X86_64) 805944d9fecSLuiz Capitulino static void destroy_compound_gigantic_page(struct page *page, 806944d9fecSLuiz Capitulino unsigned long order) 807944d9fecSLuiz Capitulino { 808944d9fecSLuiz Capitulino int i; 809944d9fecSLuiz Capitulino int nr_pages = 1 << order; 810944d9fecSLuiz Capitulino struct page *p = page + 1; 811944d9fecSLuiz Capitulino 812944d9fecSLuiz Capitulino for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 813944d9fecSLuiz Capitulino __ClearPageTail(p); 814944d9fecSLuiz Capitulino set_page_refcounted(p); 815944d9fecSLuiz Capitulino p->first_page = NULL; 816944d9fecSLuiz Capitulino } 817944d9fecSLuiz Capitulino 818944d9fecSLuiz Capitulino set_compound_order(page, 0); 819944d9fecSLuiz Capitulino __ClearPageHead(page); 820944d9fecSLuiz Capitulino } 821944d9fecSLuiz Capitulino 822944d9fecSLuiz Capitulino static void free_gigantic_page(struct page *page, unsigned order) 823944d9fecSLuiz Capitulino { 824944d9fecSLuiz Capitulino free_contig_range(page_to_pfn(page), 1 << order); 825944d9fecSLuiz Capitulino } 826944d9fecSLuiz Capitulino 827944d9fecSLuiz Capitulino static int __alloc_gigantic_page(unsigned long start_pfn, 828944d9fecSLuiz Capitulino unsigned long nr_pages) 829944d9fecSLuiz Capitulino { 830944d9fecSLuiz Capitulino unsigned long end_pfn = start_pfn + nr_pages; 831944d9fecSLuiz Capitulino return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE); 832944d9fecSLuiz Capitulino } 833944d9fecSLuiz Capitulino 834944d9fecSLuiz Capitulino static bool pfn_range_valid_gigantic(unsigned long start_pfn, 835944d9fecSLuiz Capitulino unsigned long nr_pages) 836944d9fecSLuiz Capitulino { 837944d9fecSLuiz Capitulino unsigned long i, end_pfn = start_pfn + nr_pages; 838944d9fecSLuiz Capitulino struct page *page; 839944d9fecSLuiz Capitulino 840944d9fecSLuiz Capitulino for (i = start_pfn; i < end_pfn; i++) { 841944d9fecSLuiz Capitulino if (!pfn_valid(i)) 842944d9fecSLuiz Capitulino return false; 843944d9fecSLuiz Capitulino 844944d9fecSLuiz Capitulino page = pfn_to_page(i); 845944d9fecSLuiz Capitulino 846944d9fecSLuiz Capitulino if (PageReserved(page)) 847944d9fecSLuiz Capitulino return false; 848944d9fecSLuiz Capitulino 849944d9fecSLuiz Capitulino if (page_count(page) > 0) 850944d9fecSLuiz Capitulino return false; 851944d9fecSLuiz Capitulino 852944d9fecSLuiz Capitulino if (PageHuge(page)) 853944d9fecSLuiz Capitulino return false; 854944d9fecSLuiz Capitulino } 855944d9fecSLuiz Capitulino 856944d9fecSLuiz Capitulino return true; 857944d9fecSLuiz Capitulino } 858944d9fecSLuiz Capitulino 859944d9fecSLuiz Capitulino static bool zone_spans_last_pfn(const struct zone *zone, 860944d9fecSLuiz Capitulino unsigned long start_pfn, unsigned long nr_pages) 861944d9fecSLuiz Capitulino { 862944d9fecSLuiz Capitulino unsigned long last_pfn = start_pfn + nr_pages - 1; 863944d9fecSLuiz Capitulino return zone_spans_pfn(zone, last_pfn); 864944d9fecSLuiz Capitulino } 865944d9fecSLuiz Capitulino 866944d9fecSLuiz Capitulino static struct page *alloc_gigantic_page(int nid, unsigned order) 867944d9fecSLuiz Capitulino { 868944d9fecSLuiz Capitulino unsigned long nr_pages = 1 << order; 869944d9fecSLuiz Capitulino unsigned long ret, pfn, flags; 870944d9fecSLuiz Capitulino struct zone *z; 871944d9fecSLuiz Capitulino 872944d9fecSLuiz Capitulino z = NODE_DATA(nid)->node_zones; 873944d9fecSLuiz Capitulino for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) { 874944d9fecSLuiz Capitulino spin_lock_irqsave(&z->lock, flags); 875944d9fecSLuiz Capitulino 876944d9fecSLuiz Capitulino pfn = ALIGN(z->zone_start_pfn, nr_pages); 877944d9fecSLuiz Capitulino while (zone_spans_last_pfn(z, pfn, nr_pages)) { 878944d9fecSLuiz Capitulino if (pfn_range_valid_gigantic(pfn, nr_pages)) { 879944d9fecSLuiz Capitulino /* 880944d9fecSLuiz Capitulino * We release the zone lock here because 881944d9fecSLuiz Capitulino * alloc_contig_range() will also lock the zone 882944d9fecSLuiz Capitulino * at some point. If there's an allocation 883944d9fecSLuiz Capitulino * spinning on this lock, it may win the race 884944d9fecSLuiz Capitulino * and cause alloc_contig_range() to fail... 885944d9fecSLuiz Capitulino */ 886944d9fecSLuiz Capitulino spin_unlock_irqrestore(&z->lock, flags); 887944d9fecSLuiz Capitulino ret = __alloc_gigantic_page(pfn, nr_pages); 888944d9fecSLuiz Capitulino if (!ret) 889944d9fecSLuiz Capitulino return pfn_to_page(pfn); 890944d9fecSLuiz Capitulino spin_lock_irqsave(&z->lock, flags); 891944d9fecSLuiz Capitulino } 892944d9fecSLuiz Capitulino pfn += nr_pages; 893944d9fecSLuiz Capitulino } 894944d9fecSLuiz Capitulino 895944d9fecSLuiz Capitulino spin_unlock_irqrestore(&z->lock, flags); 896944d9fecSLuiz Capitulino } 897944d9fecSLuiz Capitulino 898944d9fecSLuiz Capitulino return NULL; 899944d9fecSLuiz Capitulino } 900944d9fecSLuiz Capitulino 901944d9fecSLuiz Capitulino static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); 902944d9fecSLuiz Capitulino static void prep_compound_gigantic_page(struct page *page, unsigned long order); 903944d9fecSLuiz Capitulino 904944d9fecSLuiz Capitulino static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid) 905944d9fecSLuiz Capitulino { 906944d9fecSLuiz Capitulino struct page *page; 907944d9fecSLuiz Capitulino 908944d9fecSLuiz Capitulino page = alloc_gigantic_page(nid, huge_page_order(h)); 909944d9fecSLuiz Capitulino if (page) { 910944d9fecSLuiz Capitulino prep_compound_gigantic_page(page, huge_page_order(h)); 911944d9fecSLuiz Capitulino prep_new_huge_page(h, page, nid); 912944d9fecSLuiz Capitulino } 913944d9fecSLuiz Capitulino 914944d9fecSLuiz Capitulino return page; 915944d9fecSLuiz Capitulino } 916944d9fecSLuiz Capitulino 917944d9fecSLuiz Capitulino static int alloc_fresh_gigantic_page(struct hstate *h, 918944d9fecSLuiz Capitulino nodemask_t *nodes_allowed) 919944d9fecSLuiz Capitulino { 920944d9fecSLuiz Capitulino struct page *page = NULL; 921944d9fecSLuiz Capitulino int nr_nodes, node; 922944d9fecSLuiz Capitulino 923944d9fecSLuiz Capitulino for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 924944d9fecSLuiz Capitulino page = alloc_fresh_gigantic_page_node(h, node); 925944d9fecSLuiz Capitulino if (page) 926944d9fecSLuiz Capitulino return 1; 927944d9fecSLuiz Capitulino } 928944d9fecSLuiz Capitulino 929944d9fecSLuiz Capitulino return 0; 930944d9fecSLuiz Capitulino } 931944d9fecSLuiz Capitulino 932944d9fecSLuiz Capitulino static inline bool gigantic_page_supported(void) { return true; } 933944d9fecSLuiz Capitulino #else 934944d9fecSLuiz Capitulino static inline bool gigantic_page_supported(void) { return false; } 935944d9fecSLuiz Capitulino static inline void free_gigantic_page(struct page *page, unsigned order) { } 936944d9fecSLuiz Capitulino static inline void destroy_compound_gigantic_page(struct page *page, 937944d9fecSLuiz Capitulino unsigned long order) { } 938944d9fecSLuiz Capitulino static inline int alloc_fresh_gigantic_page(struct hstate *h, 939944d9fecSLuiz Capitulino nodemask_t *nodes_allowed) { return 0; } 940944d9fecSLuiz Capitulino #endif 941944d9fecSLuiz Capitulino 942a5516438SAndi Kleen static void update_and_free_page(struct hstate *h, struct page *page) 9436af2acb6SAdam Litke { 9446af2acb6SAdam Litke int i; 945a5516438SAndi Kleen 946944d9fecSLuiz Capitulino if (hstate_is_gigantic(h) && !gigantic_page_supported()) 947944d9fecSLuiz Capitulino return; 94818229df5SAndy Whitcroft 949a5516438SAndi Kleen h->nr_huge_pages--; 950a5516438SAndi Kleen h->nr_huge_pages_node[page_to_nid(page)]--; 951a5516438SAndi Kleen for (i = 0; i < pages_per_huge_page(h); i++) { 95232f84528SChris Forbes page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 95332f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 954a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 955a7407a27SLuiz Capitulino 1 << PG_writeback); 9566af2acb6SAdam Litke } 957309381feSSasha Levin VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 9586af2acb6SAdam Litke set_compound_page_dtor(page, NULL); 9596af2acb6SAdam Litke set_page_refcounted(page); 960944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) { 961944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 962944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 963944d9fecSLuiz Capitulino } else { 9647f2e9525SGerald Schaefer arch_release_hugepage(page); 965a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 9666af2acb6SAdam Litke } 967944d9fecSLuiz Capitulino } 9686af2acb6SAdam Litke 969e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 970e5ff2159SAndi Kleen { 971e5ff2159SAndi Kleen struct hstate *h; 972e5ff2159SAndi Kleen 973e5ff2159SAndi Kleen for_each_hstate(h) { 974e5ff2159SAndi Kleen if (huge_page_size(h) == size) 975e5ff2159SAndi Kleen return h; 976e5ff2159SAndi Kleen } 977e5ff2159SAndi Kleen return NULL; 978e5ff2159SAndi Kleen } 979e5ff2159SAndi Kleen 980bcc54222SNaoya Horiguchi /* 981bcc54222SNaoya Horiguchi * Test to determine whether the hugepage is "active/in-use" (i.e. being linked 982bcc54222SNaoya Horiguchi * to hstate->hugepage_activelist.) 983bcc54222SNaoya Horiguchi * 984bcc54222SNaoya Horiguchi * This function can be called for tail pages, but never returns true for them. 985bcc54222SNaoya Horiguchi */ 986bcc54222SNaoya Horiguchi bool page_huge_active(struct page *page) 987bcc54222SNaoya Horiguchi { 988bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHuge(page), page); 989bcc54222SNaoya Horiguchi return PageHead(page) && PagePrivate(&page[1]); 990bcc54222SNaoya Horiguchi } 991bcc54222SNaoya Horiguchi 992bcc54222SNaoya Horiguchi /* never called for tail page */ 993bcc54222SNaoya Horiguchi static void set_page_huge_active(struct page *page) 994bcc54222SNaoya Horiguchi { 995bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 996bcc54222SNaoya Horiguchi SetPagePrivate(&page[1]); 997bcc54222SNaoya Horiguchi } 998bcc54222SNaoya Horiguchi 999bcc54222SNaoya Horiguchi static void clear_page_huge_active(struct page *page) 1000bcc54222SNaoya Horiguchi { 1001bcc54222SNaoya Horiguchi VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 1002bcc54222SNaoya Horiguchi ClearPagePrivate(&page[1]); 1003bcc54222SNaoya Horiguchi } 1004bcc54222SNaoya Horiguchi 10058f1d26d0SAtsushi Kumagai void free_huge_page(struct page *page) 100627a85ef1SDavid Gibson { 1007a5516438SAndi Kleen /* 1008a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1009a5516438SAndi Kleen * compound page destructor. 1010a5516438SAndi Kleen */ 1011e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 10127893d1d5SAdam Litke int nid = page_to_nid(page); 101390481622SDavid Gibson struct hugepage_subpool *spool = 101490481622SDavid Gibson (struct hugepage_subpool *)page_private(page); 101507443a85SJoonsoo Kim bool restore_reserve; 101627a85ef1SDavid Gibson 1017e5df70abSAndy Whitcroft set_page_private(page, 0); 101823be7468SMel Gorman page->mapping = NULL; 10197893d1d5SAdam Litke BUG_ON(page_count(page)); 10200fe6e20bSNaoya Horiguchi BUG_ON(page_mapcount(page)); 102107443a85SJoonsoo Kim restore_reserve = PagePrivate(page); 102216c794b4SJoonsoo Kim ClearPagePrivate(page); 102327a85ef1SDavid Gibson 10241c5ecae3SMike Kravetz /* 10251c5ecae3SMike Kravetz * A return code of zero implies that the subpool will be under its 10261c5ecae3SMike Kravetz * minimum size if the reservation is not restored after page is free. 10271c5ecae3SMike Kravetz * Therefore, force restore_reserve operation. 10281c5ecae3SMike Kravetz */ 10291c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 10301c5ecae3SMike Kravetz restore_reserve = true; 10311c5ecae3SMike Kravetz 103227a85ef1SDavid Gibson spin_lock(&hugetlb_lock); 1033bcc54222SNaoya Horiguchi clear_page_huge_active(page); 10346d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 10356d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 103607443a85SJoonsoo Kim if (restore_reserve) 103707443a85SJoonsoo Kim h->resv_huge_pages++; 103807443a85SJoonsoo Kim 1039944d9fecSLuiz Capitulino if (h->surplus_huge_pages_node[nid]) { 10400edaecfaSAneesh Kumar K.V /* remove the page from active list */ 10410edaecfaSAneesh Kumar K.V list_del(&page->lru); 1042a5516438SAndi Kleen update_and_free_page(h, page); 1043a5516438SAndi Kleen h->surplus_huge_pages--; 1044a5516438SAndi Kleen h->surplus_huge_pages_node[nid]--; 10457893d1d5SAdam Litke } else { 10465d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1047a5516438SAndi Kleen enqueue_huge_page(h, page); 10487893d1d5SAdam Litke } 104927a85ef1SDavid Gibson spin_unlock(&hugetlb_lock); 105027a85ef1SDavid Gibson } 105127a85ef1SDavid Gibson 1052a5516438SAndi Kleen static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1053b7ba30c6SAndi Kleen { 10540edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1055b7ba30c6SAndi Kleen set_compound_page_dtor(page, free_huge_page); 1056b7ba30c6SAndi Kleen spin_lock(&hugetlb_lock); 10579dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 1058a5516438SAndi Kleen h->nr_huge_pages++; 1059a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 1060b7ba30c6SAndi Kleen spin_unlock(&hugetlb_lock); 1061b7ba30c6SAndi Kleen put_page(page); /* free it into the hugepage allocator */ 1062b7ba30c6SAndi Kleen } 1063b7ba30c6SAndi Kleen 10642906dd52SLuiz Capitulino static void prep_compound_gigantic_page(struct page *page, unsigned long order) 106520a0307cSWu Fengguang { 106620a0307cSWu Fengguang int i; 106720a0307cSWu Fengguang int nr_pages = 1 << order; 106820a0307cSWu Fengguang struct page *p = page + 1; 106920a0307cSWu Fengguang 107020a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 107120a0307cSWu Fengguang set_compound_order(page, order); 107220a0307cSWu Fengguang __SetPageHead(page); 1073ef5a22beSAndrea Arcangeli __ClearPageReserved(page); 107420a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1075ef5a22beSAndrea Arcangeli /* 1076ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1077ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1078ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 1079ef5a22beSAndrea Arcangeli * too. Otherwse drivers using get_user_pages() to access tail 1080ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1081ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1082ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1083ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1084ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1085ef5a22beSAndrea Arcangeli * after get_user_pages(). 1086ef5a22beSAndrea Arcangeli */ 1087ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 108858a84aa9SYouquan Song set_page_count(p, 0); 108920a0307cSWu Fengguang p->first_page = page; 109044fc8057SDavid Rientjes /* Make sure p->first_page is always valid for PageTail() */ 109144fc8057SDavid Rientjes smp_wmb(); 109244fc8057SDavid Rientjes __SetPageTail(p); 109320a0307cSWu Fengguang } 109420a0307cSWu Fengguang } 109520a0307cSWu Fengguang 10967795912cSAndrew Morton /* 10977795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 10987795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 10997795912cSAndrew Morton * details. 11007795912cSAndrew Morton */ 110120a0307cSWu Fengguang int PageHuge(struct page *page) 110220a0307cSWu Fengguang { 110320a0307cSWu Fengguang if (!PageCompound(page)) 110420a0307cSWu Fengguang return 0; 110520a0307cSWu Fengguang 110620a0307cSWu Fengguang page = compound_head(page); 1107758f66a2SAndrew Morton return get_compound_page_dtor(page) == free_huge_page; 110820a0307cSWu Fengguang } 110943131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 111043131e14SNaoya Horiguchi 111127c73ae7SAndrea Arcangeli /* 111227c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 111327c73ae7SAndrea Arcangeli * normal or transparent huge pages. 111427c73ae7SAndrea Arcangeli */ 111527c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 111627c73ae7SAndrea Arcangeli { 111727c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 111827c73ae7SAndrea Arcangeli return 0; 111927c73ae7SAndrea Arcangeli 1120758f66a2SAndrew Morton return get_compound_page_dtor(page_head) == free_huge_page; 112127c73ae7SAndrea Arcangeli } 112227c73ae7SAndrea Arcangeli 112313d60f4bSZhang Yi pgoff_t __basepage_index(struct page *page) 112413d60f4bSZhang Yi { 112513d60f4bSZhang Yi struct page *page_head = compound_head(page); 112613d60f4bSZhang Yi pgoff_t index = page_index(page_head); 112713d60f4bSZhang Yi unsigned long compound_idx; 112813d60f4bSZhang Yi 112913d60f4bSZhang Yi if (!PageHuge(page_head)) 113013d60f4bSZhang Yi return page_index(page); 113113d60f4bSZhang Yi 113213d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 113313d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 113413d60f4bSZhang Yi else 113513d60f4bSZhang Yi compound_idx = page - page_head; 113613d60f4bSZhang Yi 113713d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 113813d60f4bSZhang Yi } 113913d60f4bSZhang Yi 1140a5516438SAndi Kleen static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) 11411da177e4SLinus Torvalds { 11421da177e4SLinus Torvalds struct page *page; 1143f96efd58SJoe Jin 11446484eb3eSMel Gorman page = alloc_pages_exact_node(nid, 114586cdb465SNaoya Horiguchi htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| 1146551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 1147a5516438SAndi Kleen huge_page_order(h)); 11481da177e4SLinus Torvalds if (page) { 11497f2e9525SGerald Schaefer if (arch_prepare_hugepage(page)) { 1150caff3a2cSGerald Schaefer __free_pages(page, huge_page_order(h)); 11517b8ee84dSHarvey Harrison return NULL; 11527f2e9525SGerald Schaefer } 1153a5516438SAndi Kleen prep_new_huge_page(h, page, nid); 11541da177e4SLinus Torvalds } 115563b4613cSNishanth Aravamudan 115663b4613cSNishanth Aravamudan return page; 115763b4613cSNishanth Aravamudan } 115863b4613cSNishanth Aravamudan 1159b2261026SJoonsoo Kim static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 1160b2261026SJoonsoo Kim { 1161b2261026SJoonsoo Kim struct page *page; 1162b2261026SJoonsoo Kim int nr_nodes, node; 1163b2261026SJoonsoo Kim int ret = 0; 1164b2261026SJoonsoo Kim 1165b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1166b2261026SJoonsoo Kim page = alloc_fresh_huge_page_node(h, node); 1167b2261026SJoonsoo Kim if (page) { 1168b2261026SJoonsoo Kim ret = 1; 1169b2261026SJoonsoo Kim break; 1170b2261026SJoonsoo Kim } 1171b2261026SJoonsoo Kim } 1172b2261026SJoonsoo Kim 1173b2261026SJoonsoo Kim if (ret) 1174b2261026SJoonsoo Kim count_vm_event(HTLB_BUDDY_PGALLOC); 1175b2261026SJoonsoo Kim else 1176b2261026SJoonsoo Kim count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 1177b2261026SJoonsoo Kim 1178b2261026SJoonsoo Kim return ret; 1179b2261026SJoonsoo Kim } 1180b2261026SJoonsoo Kim 1181e8c5c824SLee Schermerhorn /* 1182e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 1183e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 1184e8c5c824SLee Schermerhorn * balanced over allowed nodes. 1185e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 1186e8c5c824SLee Schermerhorn */ 11876ae11b27SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 11886ae11b27SLee Schermerhorn bool acct_surplus) 1189e8c5c824SLee Schermerhorn { 1190b2261026SJoonsoo Kim int nr_nodes, node; 1191e8c5c824SLee Schermerhorn int ret = 0; 1192e8c5c824SLee Schermerhorn 1193b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1194685f3457SLee Schermerhorn /* 1195685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 1196685f3457SLee Schermerhorn * nodes with surplus pages. 1197685f3457SLee Schermerhorn */ 1198b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 1199b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 1200e8c5c824SLee Schermerhorn struct page *page = 1201b2261026SJoonsoo Kim list_entry(h->hugepage_freelists[node].next, 1202e8c5c824SLee Schermerhorn struct page, lru); 1203e8c5c824SLee Schermerhorn list_del(&page->lru); 1204e8c5c824SLee Schermerhorn h->free_huge_pages--; 1205b2261026SJoonsoo Kim h->free_huge_pages_node[node]--; 1206685f3457SLee Schermerhorn if (acct_surplus) { 1207685f3457SLee Schermerhorn h->surplus_huge_pages--; 1208b2261026SJoonsoo Kim h->surplus_huge_pages_node[node]--; 1209685f3457SLee Schermerhorn } 1210e8c5c824SLee Schermerhorn update_and_free_page(h, page); 1211e8c5c824SLee Schermerhorn ret = 1; 12129a76db09SLee Schermerhorn break; 1213e8c5c824SLee Schermerhorn } 1214b2261026SJoonsoo Kim } 1215e8c5c824SLee Schermerhorn 1216e8c5c824SLee Schermerhorn return ret; 1217e8c5c824SLee Schermerhorn } 1218e8c5c824SLee Schermerhorn 1219c8721bbbSNaoya Horiguchi /* 1220c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 1221c8721bbbSNaoya Horiguchi * nothing for in-use (including surplus) hugepages. 1222c8721bbbSNaoya Horiguchi */ 1223c8721bbbSNaoya Horiguchi static void dissolve_free_huge_page(struct page *page) 1224c8721bbbSNaoya Horiguchi { 1225c8721bbbSNaoya Horiguchi spin_lock(&hugetlb_lock); 1226c8721bbbSNaoya Horiguchi if (PageHuge(page) && !page_count(page)) { 1227c8721bbbSNaoya Horiguchi struct hstate *h = page_hstate(page); 1228c8721bbbSNaoya Horiguchi int nid = page_to_nid(page); 1229c8721bbbSNaoya Horiguchi list_del(&page->lru); 1230c8721bbbSNaoya Horiguchi h->free_huge_pages--; 1231c8721bbbSNaoya Horiguchi h->free_huge_pages_node[nid]--; 1232c8721bbbSNaoya Horiguchi update_and_free_page(h, page); 1233c8721bbbSNaoya Horiguchi } 1234c8721bbbSNaoya Horiguchi spin_unlock(&hugetlb_lock); 1235c8721bbbSNaoya Horiguchi } 1236c8721bbbSNaoya Horiguchi 1237c8721bbbSNaoya Horiguchi /* 1238c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 1239c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 1240c8721bbbSNaoya Horiguchi * Note that start_pfn should aligned with (minimum) hugepage size. 1241c8721bbbSNaoya Horiguchi */ 1242c8721bbbSNaoya Horiguchi void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 1243c8721bbbSNaoya Horiguchi { 1244c8721bbbSNaoya Horiguchi unsigned long pfn; 1245c8721bbbSNaoya Horiguchi 1246d0177639SLi Zhong if (!hugepages_supported()) 1247d0177639SLi Zhong return; 1248d0177639SLi Zhong 1249641844f5SNaoya Horiguchi VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << minimum_order)); 1250641844f5SNaoya Horiguchi for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) 1251c8721bbbSNaoya Horiguchi dissolve_free_huge_page(pfn_to_page(pfn)); 1252c8721bbbSNaoya Horiguchi } 1253c8721bbbSNaoya Horiguchi 1254bf50bab2SNaoya Horiguchi static struct page *alloc_buddy_huge_page(struct hstate *h, int nid) 12557893d1d5SAdam Litke { 12567893d1d5SAdam Litke struct page *page; 1257bf50bab2SNaoya Horiguchi unsigned int r_nid; 12587893d1d5SAdam Litke 1259bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1260aa888a74SAndi Kleen return NULL; 1261aa888a74SAndi Kleen 1262d1c3fb1fSNishanth Aravamudan /* 1263d1c3fb1fSNishanth Aravamudan * Assume we will successfully allocate the surplus page to 1264d1c3fb1fSNishanth Aravamudan * prevent racing processes from causing the surplus to exceed 1265d1c3fb1fSNishanth Aravamudan * overcommit 1266d1c3fb1fSNishanth Aravamudan * 1267d1c3fb1fSNishanth Aravamudan * This however introduces a different race, where a process B 1268d1c3fb1fSNishanth Aravamudan * tries to grow the static hugepage pool while alloc_pages() is 1269d1c3fb1fSNishanth Aravamudan * called by process A. B will only examine the per-node 1270d1c3fb1fSNishanth Aravamudan * counters in determining if surplus huge pages can be 1271d1c3fb1fSNishanth Aravamudan * converted to normal huge pages in adjust_pool_surplus(). A 1272d1c3fb1fSNishanth Aravamudan * won't be able to increment the per-node counter, until the 1273d1c3fb1fSNishanth Aravamudan * lock is dropped by B, but B doesn't drop hugetlb_lock until 1274d1c3fb1fSNishanth Aravamudan * no more huge pages can be converted from surplus to normal 1275d1c3fb1fSNishanth Aravamudan * state (and doesn't try to convert again). Thus, we have a 1276d1c3fb1fSNishanth Aravamudan * case where a surplus huge page exists, the pool is grown, and 1277d1c3fb1fSNishanth Aravamudan * the surplus huge page still exists after, even though it 1278d1c3fb1fSNishanth Aravamudan * should just have been converted to a normal huge page. This 1279d1c3fb1fSNishanth Aravamudan * does not leak memory, though, as the hugepage will be freed 1280d1c3fb1fSNishanth Aravamudan * once it is out of use. It also does not allow the counters to 1281d1c3fb1fSNishanth Aravamudan * go out of whack in adjust_pool_surplus() as we don't modify 1282d1c3fb1fSNishanth Aravamudan * the node values until we've gotten the hugepage and only the 1283d1c3fb1fSNishanth Aravamudan * per-node value is checked there. 1284d1c3fb1fSNishanth Aravamudan */ 1285d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 1286a5516438SAndi Kleen if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 1287d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1288d1c3fb1fSNishanth Aravamudan return NULL; 1289d1c3fb1fSNishanth Aravamudan } else { 1290a5516438SAndi Kleen h->nr_huge_pages++; 1291a5516438SAndi Kleen h->surplus_huge_pages++; 1292d1c3fb1fSNishanth Aravamudan } 1293d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1294d1c3fb1fSNishanth Aravamudan 1295bf50bab2SNaoya Horiguchi if (nid == NUMA_NO_NODE) 129686cdb465SNaoya Horiguchi page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP| 1297551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 1298a5516438SAndi Kleen huge_page_order(h)); 1299bf50bab2SNaoya Horiguchi else 1300bf50bab2SNaoya Horiguchi page = alloc_pages_exact_node(nid, 130186cdb465SNaoya Horiguchi htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| 1302bf50bab2SNaoya Horiguchi __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h)); 1303d1c3fb1fSNishanth Aravamudan 1304caff3a2cSGerald Schaefer if (page && arch_prepare_hugepage(page)) { 1305caff3a2cSGerald Schaefer __free_pages(page, huge_page_order(h)); 1306ea5768c7SHillf Danton page = NULL; 1307caff3a2cSGerald Schaefer } 1308caff3a2cSGerald Schaefer 13097893d1d5SAdam Litke spin_lock(&hugetlb_lock); 1310d1c3fb1fSNishanth Aravamudan if (page) { 13110edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1312bf50bab2SNaoya Horiguchi r_nid = page_to_nid(page); 1313d1c3fb1fSNishanth Aravamudan set_compound_page_dtor(page, free_huge_page); 13149dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 1315d1c3fb1fSNishanth Aravamudan /* 1316d1c3fb1fSNishanth Aravamudan * We incremented the global counters already 1317d1c3fb1fSNishanth Aravamudan */ 1318bf50bab2SNaoya Horiguchi h->nr_huge_pages_node[r_nid]++; 1319bf50bab2SNaoya Horiguchi h->surplus_huge_pages_node[r_nid]++; 13203b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC); 1321d1c3fb1fSNishanth Aravamudan } else { 1322a5516438SAndi Kleen h->nr_huge_pages--; 1323a5516438SAndi Kleen h->surplus_huge_pages--; 13243b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 13257893d1d5SAdam Litke } 1326d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 13277893d1d5SAdam Litke 13287893d1d5SAdam Litke return page; 13297893d1d5SAdam Litke } 13307893d1d5SAdam Litke 1331e4e574b7SAdam Litke /* 1332bf50bab2SNaoya Horiguchi * This allocation function is useful in the context where vma is irrelevant. 1333bf50bab2SNaoya Horiguchi * E.g. soft-offlining uses this function because it only cares physical 1334bf50bab2SNaoya Horiguchi * address of error page. 1335bf50bab2SNaoya Horiguchi */ 1336bf50bab2SNaoya Horiguchi struct page *alloc_huge_page_node(struct hstate *h, int nid) 1337bf50bab2SNaoya Horiguchi { 13384ef91848SJoonsoo Kim struct page *page = NULL; 1339bf50bab2SNaoya Horiguchi 1340bf50bab2SNaoya Horiguchi spin_lock(&hugetlb_lock); 13414ef91848SJoonsoo Kim if (h->free_huge_pages - h->resv_huge_pages > 0) 1342bf50bab2SNaoya Horiguchi page = dequeue_huge_page_node(h, nid); 1343bf50bab2SNaoya Horiguchi spin_unlock(&hugetlb_lock); 1344bf50bab2SNaoya Horiguchi 134594ae8ba7SAneesh Kumar K.V if (!page) 1346bf50bab2SNaoya Horiguchi page = alloc_buddy_huge_page(h, nid); 1347bf50bab2SNaoya Horiguchi 1348bf50bab2SNaoya Horiguchi return page; 1349bf50bab2SNaoya Horiguchi } 1350bf50bab2SNaoya Horiguchi 1351bf50bab2SNaoya Horiguchi /* 135225985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 1353e4e574b7SAdam Litke * of size 'delta'. 1354e4e574b7SAdam Litke */ 1355a5516438SAndi Kleen static int gather_surplus_pages(struct hstate *h, int delta) 1356e4e574b7SAdam Litke { 1357e4e574b7SAdam Litke struct list_head surplus_list; 1358e4e574b7SAdam Litke struct page *page, *tmp; 1359e4e574b7SAdam Litke int ret, i; 1360e4e574b7SAdam Litke int needed, allocated; 136128073b02SHillf Danton bool alloc_ok = true; 1362e4e574b7SAdam Litke 1363a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 1364ac09b3a1SAdam Litke if (needed <= 0) { 1365a5516438SAndi Kleen h->resv_huge_pages += delta; 1366e4e574b7SAdam Litke return 0; 1367ac09b3a1SAdam Litke } 1368e4e574b7SAdam Litke 1369e4e574b7SAdam Litke allocated = 0; 1370e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 1371e4e574b7SAdam Litke 1372e4e574b7SAdam Litke ret = -ENOMEM; 1373e4e574b7SAdam Litke retry: 1374e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 1375e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 1376bf50bab2SNaoya Horiguchi page = alloc_buddy_huge_page(h, NUMA_NO_NODE); 137728073b02SHillf Danton if (!page) { 137828073b02SHillf Danton alloc_ok = false; 137928073b02SHillf Danton break; 138028073b02SHillf Danton } 1381e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 1382e4e574b7SAdam Litke } 138328073b02SHillf Danton allocated += i; 1384e4e574b7SAdam Litke 1385e4e574b7SAdam Litke /* 1386e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 1387e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 1388e4e574b7SAdam Litke */ 1389e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 1390a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 1391a5516438SAndi Kleen (h->free_huge_pages + allocated); 139228073b02SHillf Danton if (needed > 0) { 139328073b02SHillf Danton if (alloc_ok) 1394e4e574b7SAdam Litke goto retry; 139528073b02SHillf Danton /* 139628073b02SHillf Danton * We were not able to allocate enough pages to 139728073b02SHillf Danton * satisfy the entire reservation so we free what 139828073b02SHillf Danton * we've allocated so far. 139928073b02SHillf Danton */ 140028073b02SHillf Danton goto free; 140128073b02SHillf Danton } 1402e4e574b7SAdam Litke /* 1403e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 140425985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 1405e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 1406ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 1407ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 1408ac09b3a1SAdam Litke * before they are reserved. 1409e4e574b7SAdam Litke */ 1410e4e574b7SAdam Litke needed += allocated; 1411a5516438SAndi Kleen h->resv_huge_pages += delta; 1412e4e574b7SAdam Litke ret = 0; 1413a9869b83SNaoya Horiguchi 141419fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 141519fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 141619fc3f0aSAdam Litke if ((--needed) < 0) 141719fc3f0aSAdam Litke break; 1418a9869b83SNaoya Horiguchi /* 1419a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 1420a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 1421a9869b83SNaoya Horiguchi */ 1422a9869b83SNaoya Horiguchi put_page_testzero(page); 1423309381feSSasha Levin VM_BUG_ON_PAGE(page_count(page), page); 1424a5516438SAndi Kleen enqueue_huge_page(h, page); 142519fc3f0aSAdam Litke } 142628073b02SHillf Danton free: 1427b0365c8dSHillf Danton spin_unlock(&hugetlb_lock); 142819fc3f0aSAdam Litke 142919fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 1430c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 1431a9869b83SNaoya Horiguchi put_page(page); 143219fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 1433e4e574b7SAdam Litke 1434e4e574b7SAdam Litke return ret; 1435e4e574b7SAdam Litke } 1436e4e574b7SAdam Litke 1437e4e574b7SAdam Litke /* 1438e4e574b7SAdam Litke * When releasing a hugetlb pool reservation, any surplus pages that were 1439e4e574b7SAdam Litke * allocated to satisfy the reservation must be explicitly freed if they were 1440e4e574b7SAdam Litke * never used. 1441685f3457SLee Schermerhorn * Called with hugetlb_lock held. 1442e4e574b7SAdam Litke */ 1443a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 1444a5516438SAndi Kleen unsigned long unused_resv_pages) 1445e4e574b7SAdam Litke { 1446e4e574b7SAdam Litke unsigned long nr_pages; 1447e4e574b7SAdam Litke 1448ac09b3a1SAdam Litke /* Uncommit the reservation */ 1449a5516438SAndi Kleen h->resv_huge_pages -= unused_resv_pages; 1450ac09b3a1SAdam Litke 1451aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 1452bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1453aa888a74SAndi Kleen return; 1454aa888a74SAndi Kleen 1455a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 1456e4e574b7SAdam Litke 1457685f3457SLee Schermerhorn /* 1458685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 14599b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 14609b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 14619b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 14629b5e5d0fSLee Schermerhorn * free_pool_huge_page() will balance the the freed pages across the 14639b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 1464685f3457SLee Schermerhorn */ 1465685f3457SLee Schermerhorn while (nr_pages--) { 14668cebfcd0SLai Jiangshan if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) 1467685f3457SLee Schermerhorn break; 14687848a4bfSMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 1469e4e574b7SAdam Litke } 1470e4e574b7SAdam Litke } 1471e4e574b7SAdam Litke 1472c37f9fb1SAndy Whitcroft /* 1473c37f9fb1SAndy Whitcroft * Determine if the huge page at addr within the vma has an associated 1474c37f9fb1SAndy Whitcroft * reservation. Where it does not we will need to logically increase 147590481622SDavid Gibson * reservation and actually increase subpool usage before an allocation 147690481622SDavid Gibson * can occur. Where any new reservation would be required the 147790481622SDavid Gibson * reservation change is prepared, but not committed. Once the page 147890481622SDavid Gibson * has been allocated from the subpool and instantiated the change should 147990481622SDavid Gibson * be committed via vma_commit_reservation. No action is required on 148090481622SDavid Gibson * failure. 1481c37f9fb1SAndy Whitcroft */ 1482e2f17d94SRoel Kluin static long vma_needs_reservation(struct hstate *h, 1483a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 1484c37f9fb1SAndy Whitcroft { 14854e35f483SJoonsoo Kim struct resv_map *resv; 14864e35f483SJoonsoo Kim pgoff_t idx; 14874e35f483SJoonsoo Kim long chg; 1488c37f9fb1SAndy Whitcroft 14894e35f483SJoonsoo Kim resv = vma_resv_map(vma); 14904e35f483SJoonsoo Kim if (!resv) 1491c37f9fb1SAndy Whitcroft return 1; 1492c37f9fb1SAndy Whitcroft 14934e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 14944e35f483SJoonsoo Kim chg = region_chg(resv, idx, idx + 1); 149584afd99bSAndy Whitcroft 14964e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) 14974e35f483SJoonsoo Kim return chg; 14984e35f483SJoonsoo Kim else 14994e35f483SJoonsoo Kim return chg < 0 ? chg : 0; 150084afd99bSAndy Whitcroft } 1501a5516438SAndi Kleen static void vma_commit_reservation(struct hstate *h, 1502a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 1503c37f9fb1SAndy Whitcroft { 15044e35f483SJoonsoo Kim struct resv_map *resv; 15054e35f483SJoonsoo Kim pgoff_t idx; 1506c37f9fb1SAndy Whitcroft 15074e35f483SJoonsoo Kim resv = vma_resv_map(vma); 15084e35f483SJoonsoo Kim if (!resv) 15094e35f483SJoonsoo Kim return; 15109119a41eSJoonsoo Kim 15114e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 15121406ec9bSJoonsoo Kim region_add(resv, idx, idx + 1); 1513c37f9fb1SAndy Whitcroft } 1514c37f9fb1SAndy Whitcroft 1515348ea204SAdam Litke static struct page *alloc_huge_page(struct vm_area_struct *vma, 151604f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 1517348ea204SAdam Litke { 151890481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 1519a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 1520348ea204SAdam Litke struct page *page; 1521e2f17d94SRoel Kluin long chg; 15226d76dcf4SAneesh Kumar K.V int ret, idx; 15236d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 15242fc39cecSAdam Litke 15256d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 1526a1e78772SMel Gorman /* 152790481622SDavid Gibson * Processes that did not create the mapping will have no 152890481622SDavid Gibson * reserves and will not have accounted against subpool 152990481622SDavid Gibson * limit. Check that the subpool limit can be made before 153090481622SDavid Gibson * satisfying the allocation MAP_NORESERVE mappings may also 153190481622SDavid Gibson * need pages and subpool limit allocated allocated if no reserve 153290481622SDavid Gibson * mapping overlaps. 1533a1e78772SMel Gorman */ 1534a5516438SAndi Kleen chg = vma_needs_reservation(h, vma, addr); 1535c37f9fb1SAndy Whitcroft if (chg < 0) 153676dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 15378bb3f12eSJoonsoo Kim if (chg || avoid_reserve) 15381c5ecae3SMike Kravetz if (hugepage_subpool_get_pages(spool, 1) < 0) 153976dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 154090d8b7e6SAdam Litke 15416d76dcf4SAneesh Kumar K.V ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 15428f34af6fSJianyu Zhan if (ret) 15438f34af6fSJianyu Zhan goto out_subpool_put; 15448f34af6fSJianyu Zhan 1545a1e78772SMel Gorman spin_lock(&hugetlb_lock); 1546af0ed73eSJoonsoo Kim page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg); 154781a6fcaeSJoonsoo Kim if (!page) { 154894ae8ba7SAneesh Kumar K.V spin_unlock(&hugetlb_lock); 1549bf50bab2SNaoya Horiguchi page = alloc_buddy_huge_page(h, NUMA_NO_NODE); 15508f34af6fSJianyu Zhan if (!page) 15518f34af6fSJianyu Zhan goto out_uncharge_cgroup; 15528f34af6fSJianyu Zhan 155379dbb236SAneesh Kumar K.V spin_lock(&hugetlb_lock); 155479dbb236SAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 155581a6fcaeSJoonsoo Kim /* Fall through */ 1556a1e78772SMel Gorman } 155781a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 155881a6fcaeSJoonsoo Kim spin_unlock(&hugetlb_lock); 1559a1e78772SMel Gorman 156090481622SDavid Gibson set_page_private(page, (unsigned long)spool); 1561a1e78772SMel Gorman 1562a5516438SAndi Kleen vma_commit_reservation(h, vma, addr); 15637893d1d5SAdam Litke return page; 15648f34af6fSJianyu Zhan 15658f34af6fSJianyu Zhan out_uncharge_cgroup: 15668f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 15678f34af6fSJianyu Zhan out_subpool_put: 15688f34af6fSJianyu Zhan if (chg || avoid_reserve) 15698f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 15708f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 1571b45b5bd6SDavid Gibson } 1572b45b5bd6SDavid Gibson 157374060e4dSNaoya Horiguchi /* 157474060e4dSNaoya Horiguchi * alloc_huge_page()'s wrapper which simply returns the page if allocation 157574060e4dSNaoya Horiguchi * succeeds, otherwise NULL. This function is called from new_vma_page(), 157674060e4dSNaoya Horiguchi * where no ERR_VALUE is expected to be returned. 157774060e4dSNaoya Horiguchi */ 157874060e4dSNaoya Horiguchi struct page *alloc_huge_page_noerr(struct vm_area_struct *vma, 157974060e4dSNaoya Horiguchi unsigned long addr, int avoid_reserve) 158074060e4dSNaoya Horiguchi { 158174060e4dSNaoya Horiguchi struct page *page = alloc_huge_page(vma, addr, avoid_reserve); 158274060e4dSNaoya Horiguchi if (IS_ERR(page)) 158374060e4dSNaoya Horiguchi page = NULL; 158474060e4dSNaoya Horiguchi return page; 158574060e4dSNaoya Horiguchi } 158674060e4dSNaoya Horiguchi 158791f47662SCyrill Gorcunov int __weak alloc_bootmem_huge_page(struct hstate *h) 1588aa888a74SAndi Kleen { 1589aa888a74SAndi Kleen struct huge_bootmem_page *m; 1590b2261026SJoonsoo Kim int nr_nodes, node; 1591aa888a74SAndi Kleen 1592b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 1593aa888a74SAndi Kleen void *addr; 1594aa888a74SAndi Kleen 15958b89a116SGrygorii Strashko addr = memblock_virt_alloc_try_nid_nopanic( 15968b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 15978b89a116SGrygorii Strashko 0, BOOTMEM_ALLOC_ACCESSIBLE, node); 1598aa888a74SAndi Kleen if (addr) { 1599aa888a74SAndi Kleen /* 1600aa888a74SAndi Kleen * Use the beginning of the huge page to store the 1601aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 1602aa888a74SAndi Kleen * puts them into the mem_map). 1603aa888a74SAndi Kleen */ 1604aa888a74SAndi Kleen m = addr; 1605aa888a74SAndi Kleen goto found; 1606aa888a74SAndi Kleen } 1607aa888a74SAndi Kleen } 1608aa888a74SAndi Kleen return 0; 1609aa888a74SAndi Kleen 1610aa888a74SAndi Kleen found: 1611df994eadSLuiz Capitulino BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); 1612aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 1613aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 1614aa888a74SAndi Kleen m->hstate = h; 1615aa888a74SAndi Kleen return 1; 1616aa888a74SAndi Kleen } 1617aa888a74SAndi Kleen 1618f412c97aSDavid Rientjes static void __init prep_compound_huge_page(struct page *page, int order) 161918229df5SAndy Whitcroft { 162018229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 162118229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 162218229df5SAndy Whitcroft else 162318229df5SAndy Whitcroft prep_compound_page(page, order); 162418229df5SAndy Whitcroft } 162518229df5SAndy Whitcroft 1626aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 1627aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 1628aa888a74SAndi Kleen { 1629aa888a74SAndi Kleen struct huge_bootmem_page *m; 1630aa888a74SAndi Kleen 1631aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 1632aa888a74SAndi Kleen struct hstate *h = m->hstate; 1633ee8f248dSBecky Bruce struct page *page; 1634ee8f248dSBecky Bruce 1635ee8f248dSBecky Bruce #ifdef CONFIG_HIGHMEM 1636ee8f248dSBecky Bruce page = pfn_to_page(m->phys >> PAGE_SHIFT); 16378b89a116SGrygorii Strashko memblock_free_late(__pa(m), 1638ee8f248dSBecky Bruce sizeof(struct huge_bootmem_page)); 1639ee8f248dSBecky Bruce #else 1640ee8f248dSBecky Bruce page = virt_to_page(m); 1641ee8f248dSBecky Bruce #endif 1642aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 164318229df5SAndy Whitcroft prep_compound_huge_page(page, h->order); 1644ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 1645aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 1646b0320c7bSRafael Aquini /* 1647b0320c7bSRafael Aquini * If we had gigantic hugepages allocated at boot time, we need 1648b0320c7bSRafael Aquini * to restore the 'stolen' pages to totalram_pages in order to 1649b0320c7bSRafael Aquini * fix confusing memory reports from free(1) and another 1650b0320c7bSRafael Aquini * side-effects, like CommitLimit going negative. 1651b0320c7bSRafael Aquini */ 1652bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 16533dcc0571SJiang Liu adjust_managed_page_count(page, 1 << h->order); 1654aa888a74SAndi Kleen } 1655aa888a74SAndi Kleen } 1656aa888a74SAndi Kleen 16578faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 16581da177e4SLinus Torvalds { 16591da177e4SLinus Torvalds unsigned long i; 16601da177e4SLinus Torvalds 1661e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 1662bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 1663aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 1664aa888a74SAndi Kleen break; 16659b5e5d0fSLee Schermerhorn } else if (!alloc_fresh_huge_page(h, 16668cebfcd0SLai Jiangshan &node_states[N_MEMORY])) 16671da177e4SLinus Torvalds break; 16681da177e4SLinus Torvalds } 16698faa8b07SAndi Kleen h->max_huge_pages = i; 1670e5ff2159SAndi Kleen } 1671e5ff2159SAndi Kleen 1672e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 1673e5ff2159SAndi Kleen { 1674e5ff2159SAndi Kleen struct hstate *h; 1675e5ff2159SAndi Kleen 1676e5ff2159SAndi Kleen for_each_hstate(h) { 1677641844f5SNaoya Horiguchi if (minimum_order > huge_page_order(h)) 1678641844f5SNaoya Horiguchi minimum_order = huge_page_order(h); 1679641844f5SNaoya Horiguchi 16808faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 1681bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 16828faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 1683e5ff2159SAndi Kleen } 1684641844f5SNaoya Horiguchi VM_BUG_ON(minimum_order == UINT_MAX); 1685e5ff2159SAndi Kleen } 1686e5ff2159SAndi Kleen 16874abd32dbSAndi Kleen static char * __init memfmt(char *buf, unsigned long n) 16884abd32dbSAndi Kleen { 16894abd32dbSAndi Kleen if (n >= (1UL << 30)) 16904abd32dbSAndi Kleen sprintf(buf, "%lu GB", n >> 30); 16914abd32dbSAndi Kleen else if (n >= (1UL << 20)) 16924abd32dbSAndi Kleen sprintf(buf, "%lu MB", n >> 20); 16934abd32dbSAndi Kleen else 16944abd32dbSAndi Kleen sprintf(buf, "%lu KB", n >> 10); 16954abd32dbSAndi Kleen return buf; 16964abd32dbSAndi Kleen } 16974abd32dbSAndi Kleen 1698e5ff2159SAndi Kleen static void __init report_hugepages(void) 1699e5ff2159SAndi Kleen { 1700e5ff2159SAndi Kleen struct hstate *h; 1701e5ff2159SAndi Kleen 1702e5ff2159SAndi Kleen for_each_hstate(h) { 17034abd32dbSAndi Kleen char buf[32]; 1704ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 17054abd32dbSAndi Kleen memfmt(buf, huge_page_size(h)), 17064abd32dbSAndi Kleen h->free_huge_pages); 1707e5ff2159SAndi Kleen } 1708e5ff2159SAndi Kleen } 1709e5ff2159SAndi Kleen 17101da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 17116ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 17126ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 17131da177e4SLinus Torvalds { 17144415cc8dSChristoph Lameter int i; 17154415cc8dSChristoph Lameter 1716bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1717aa888a74SAndi Kleen return; 1718aa888a74SAndi Kleen 17196ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 17201da177e4SLinus Torvalds struct page *page, *next; 1721a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 1722a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 1723a5516438SAndi Kleen if (count >= h->nr_huge_pages) 17246b0c880dSAdam Litke return; 17251da177e4SLinus Torvalds if (PageHighMem(page)) 17261da177e4SLinus Torvalds continue; 17271da177e4SLinus Torvalds list_del(&page->lru); 1728e5ff2159SAndi Kleen update_and_free_page(h, page); 1729a5516438SAndi Kleen h->free_huge_pages--; 1730a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 17311da177e4SLinus Torvalds } 17321da177e4SLinus Torvalds } 17331da177e4SLinus Torvalds } 17341da177e4SLinus Torvalds #else 17356ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 17366ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 17371da177e4SLinus Torvalds { 17381da177e4SLinus Torvalds } 17391da177e4SLinus Torvalds #endif 17401da177e4SLinus Torvalds 174120a0307cSWu Fengguang /* 174220a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 174320a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 174420a0307cSWu Fengguang * Returns 1 if an adjustment was made. 174520a0307cSWu Fengguang */ 17466ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 17476ae11b27SLee Schermerhorn int delta) 174820a0307cSWu Fengguang { 1749b2261026SJoonsoo Kim int nr_nodes, node; 175020a0307cSWu Fengguang 175120a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 175220a0307cSWu Fengguang 1753e8c5c824SLee Schermerhorn if (delta < 0) { 1754b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1755b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 1756b2261026SJoonsoo Kim goto found; 1757b2261026SJoonsoo Kim } 1758b2261026SJoonsoo Kim } else { 1759b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1760b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 1761b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 1762b2261026SJoonsoo Kim goto found; 1763e8c5c824SLee Schermerhorn } 17649a76db09SLee Schermerhorn } 1765b2261026SJoonsoo Kim return 0; 176620a0307cSWu Fengguang 1767b2261026SJoonsoo Kim found: 176820a0307cSWu Fengguang h->surplus_huge_pages += delta; 1769b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 1770b2261026SJoonsoo Kim return 1; 177120a0307cSWu Fengguang } 177220a0307cSWu Fengguang 1773a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 17746ae11b27SLee Schermerhorn static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, 17756ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 17761da177e4SLinus Torvalds { 17777893d1d5SAdam Litke unsigned long min_count, ret; 17781da177e4SLinus Torvalds 1779944d9fecSLuiz Capitulino if (hstate_is_gigantic(h) && !gigantic_page_supported()) 1780aa888a74SAndi Kleen return h->max_huge_pages; 1781aa888a74SAndi Kleen 17827893d1d5SAdam Litke /* 17837893d1d5SAdam Litke * Increase the pool size 17847893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 17857893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 1786d1c3fb1fSNishanth Aravamudan * 1787d1c3fb1fSNishanth Aravamudan * We might race with alloc_buddy_huge_page() here and be unable 1788d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 1789d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 1790d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 1791d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 17927893d1d5SAdam Litke */ 17931da177e4SLinus Torvalds spin_lock(&hugetlb_lock); 1794a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 17956ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 17967893d1d5SAdam Litke break; 17977893d1d5SAdam Litke } 17987893d1d5SAdam Litke 1799a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 18007893d1d5SAdam Litke /* 18017893d1d5SAdam Litke * If this allocation races such that we no longer need the 18027893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 18037893d1d5SAdam Litke * and reducing the surplus. 18047893d1d5SAdam Litke */ 18057893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 1806944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) 1807944d9fecSLuiz Capitulino ret = alloc_fresh_gigantic_page(h, nodes_allowed); 1808944d9fecSLuiz Capitulino else 18096ae11b27SLee Schermerhorn ret = alloc_fresh_huge_page(h, nodes_allowed); 18107893d1d5SAdam Litke spin_lock(&hugetlb_lock); 18117893d1d5SAdam Litke if (!ret) 18127893d1d5SAdam Litke goto out; 18137893d1d5SAdam Litke 1814536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 1815536240f2SMel Gorman if (signal_pending(current)) 1816536240f2SMel Gorman goto out; 18177893d1d5SAdam Litke } 18187893d1d5SAdam Litke 18197893d1d5SAdam Litke /* 18207893d1d5SAdam Litke * Decrease the pool size 18217893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 18227893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 18237893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 18247893d1d5SAdam Litke * to the desired size as pages become free. 1825d1c3fb1fSNishanth Aravamudan * 1826d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 1827d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 1828d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 1829d1c3fb1fSNishanth Aravamudan * alloc_buddy_huge_page() is checking the global counter, 1830d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 1831d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 1832d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 18337893d1d5SAdam Litke */ 1834a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 18356b0c880dSAdam Litke min_count = max(count, min_count); 18366ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 1837a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 18386ae11b27SLee Schermerhorn if (!free_pool_huge_page(h, nodes_allowed, 0)) 18391da177e4SLinus Torvalds break; 184055f67141SMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 18411da177e4SLinus Torvalds } 1842a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 18436ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 18447893d1d5SAdam Litke break; 18457893d1d5SAdam Litke } 18467893d1d5SAdam Litke out: 1847a5516438SAndi Kleen ret = persistent_huge_pages(h); 18481da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 18497893d1d5SAdam Litke return ret; 18501da177e4SLinus Torvalds } 18511da177e4SLinus Torvalds 1852a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 1853a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 1854a3437870SNishanth Aravamudan 1855a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 1856a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 1857a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 1858a3437870SNishanth Aravamudan 1859a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 1860a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 1861a3437870SNishanth Aravamudan 18629a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 18639a305230SLee Schermerhorn 18649a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 1865a3437870SNishanth Aravamudan { 1866a3437870SNishanth Aravamudan int i; 18679a305230SLee Schermerhorn 1868a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 18699a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 18709a305230SLee Schermerhorn if (nidp) 18719a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 1872a3437870SNishanth Aravamudan return &hstates[i]; 18739a305230SLee Schermerhorn } 18749a305230SLee Schermerhorn 18759a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 1876a3437870SNishanth Aravamudan } 1877a3437870SNishanth Aravamudan 187806808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 1879a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1880a3437870SNishanth Aravamudan { 18819a305230SLee Schermerhorn struct hstate *h; 18829a305230SLee Schermerhorn unsigned long nr_huge_pages; 18839a305230SLee Schermerhorn int nid; 18849a305230SLee Schermerhorn 18859a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 18869a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 18879a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 18889a305230SLee Schermerhorn else 18899a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 18909a305230SLee Schermerhorn 18919a305230SLee Schermerhorn return sprintf(buf, "%lu\n", nr_huge_pages); 1892a3437870SNishanth Aravamudan } 1893adbe8726SEric B Munson 1894238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 1895238d3c13SDavid Rientjes struct hstate *h, int nid, 1896238d3c13SDavid Rientjes unsigned long count, size_t len) 1897a3437870SNishanth Aravamudan { 1898a3437870SNishanth Aravamudan int err; 1899bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); 1900a3437870SNishanth Aravamudan 1901944d9fecSLuiz Capitulino if (hstate_is_gigantic(h) && !gigantic_page_supported()) { 1902adbe8726SEric B Munson err = -EINVAL; 1903adbe8726SEric B Munson goto out; 1904adbe8726SEric B Munson } 1905adbe8726SEric B Munson 19069a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 19079a305230SLee Schermerhorn /* 19089a305230SLee Schermerhorn * global hstate attribute 19099a305230SLee Schermerhorn */ 19109a305230SLee Schermerhorn if (!(obey_mempolicy && 19119a305230SLee Schermerhorn init_nodemask_of_mempolicy(nodes_allowed))) { 191206808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 19138cebfcd0SLai Jiangshan nodes_allowed = &node_states[N_MEMORY]; 191406808b08SLee Schermerhorn } 19159a305230SLee Schermerhorn } else if (nodes_allowed) { 19169a305230SLee Schermerhorn /* 19179a305230SLee Schermerhorn * per node hstate attribute: adjust count to global, 19189a305230SLee Schermerhorn * but restrict alloc/free to the specified node. 19199a305230SLee Schermerhorn */ 19209a305230SLee Schermerhorn count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 19219a305230SLee Schermerhorn init_nodemask_of_node(nodes_allowed, nid); 19229a305230SLee Schermerhorn } else 19238cebfcd0SLai Jiangshan nodes_allowed = &node_states[N_MEMORY]; 19249a305230SLee Schermerhorn 192506808b08SLee Schermerhorn h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); 1926a3437870SNishanth Aravamudan 19278cebfcd0SLai Jiangshan if (nodes_allowed != &node_states[N_MEMORY]) 192806808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 192906808b08SLee Schermerhorn 193006808b08SLee Schermerhorn return len; 1931adbe8726SEric B Munson out: 1932adbe8726SEric B Munson NODEMASK_FREE(nodes_allowed); 1933adbe8726SEric B Munson return err; 193406808b08SLee Schermerhorn } 193506808b08SLee Schermerhorn 1936238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 1937238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 1938238d3c13SDavid Rientjes size_t len) 1939238d3c13SDavid Rientjes { 1940238d3c13SDavid Rientjes struct hstate *h; 1941238d3c13SDavid Rientjes unsigned long count; 1942238d3c13SDavid Rientjes int nid; 1943238d3c13SDavid Rientjes int err; 1944238d3c13SDavid Rientjes 1945238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 1946238d3c13SDavid Rientjes if (err) 1947238d3c13SDavid Rientjes return err; 1948238d3c13SDavid Rientjes 1949238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 1950238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 1951238d3c13SDavid Rientjes } 1952238d3c13SDavid Rientjes 195306808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 195406808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 195506808b08SLee Schermerhorn { 195606808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 195706808b08SLee Schermerhorn } 195806808b08SLee Schermerhorn 195906808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 196006808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 196106808b08SLee Schermerhorn { 1962238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 1963a3437870SNishanth Aravamudan } 1964a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 1965a3437870SNishanth Aravamudan 196606808b08SLee Schermerhorn #ifdef CONFIG_NUMA 196706808b08SLee Schermerhorn 196806808b08SLee Schermerhorn /* 196906808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 197006808b08SLee Schermerhorn * huge page alloc/free. 197106808b08SLee Schermerhorn */ 197206808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 197306808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 197406808b08SLee Schermerhorn { 197506808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 197606808b08SLee Schermerhorn } 197706808b08SLee Schermerhorn 197806808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 197906808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 198006808b08SLee Schermerhorn { 1981238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 198206808b08SLee Schermerhorn } 198306808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 198406808b08SLee Schermerhorn #endif 198506808b08SLee Schermerhorn 198606808b08SLee Schermerhorn 1987a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 1988a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1989a3437870SNishanth Aravamudan { 19909a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1991a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 1992a3437870SNishanth Aravamudan } 1993adbe8726SEric B Munson 1994a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 1995a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 1996a3437870SNishanth Aravamudan { 1997a3437870SNishanth Aravamudan int err; 1998a3437870SNishanth Aravamudan unsigned long input; 19999a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2000a3437870SNishanth Aravamudan 2001bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2002adbe8726SEric B Munson return -EINVAL; 2003adbe8726SEric B Munson 20043dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 2005a3437870SNishanth Aravamudan if (err) 200673ae31e5SEric B Munson return err; 2007a3437870SNishanth Aravamudan 2008a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 2009a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 2010a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 2011a3437870SNishanth Aravamudan 2012a3437870SNishanth Aravamudan return count; 2013a3437870SNishanth Aravamudan } 2014a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 2015a3437870SNishanth Aravamudan 2016a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 2017a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2018a3437870SNishanth Aravamudan { 20199a305230SLee Schermerhorn struct hstate *h; 20209a305230SLee Schermerhorn unsigned long free_huge_pages; 20219a305230SLee Schermerhorn int nid; 20229a305230SLee Schermerhorn 20239a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 20249a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 20259a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 20269a305230SLee Schermerhorn else 20279a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 20289a305230SLee Schermerhorn 20299a305230SLee Schermerhorn return sprintf(buf, "%lu\n", free_huge_pages); 2030a3437870SNishanth Aravamudan } 2031a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 2032a3437870SNishanth Aravamudan 2033a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 2034a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2035a3437870SNishanth Aravamudan { 20369a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2037a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->resv_huge_pages); 2038a3437870SNishanth Aravamudan } 2039a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 2040a3437870SNishanth Aravamudan 2041a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 2042a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2043a3437870SNishanth Aravamudan { 20449a305230SLee Schermerhorn struct hstate *h; 20459a305230SLee Schermerhorn unsigned long surplus_huge_pages; 20469a305230SLee Schermerhorn int nid; 20479a305230SLee Schermerhorn 20489a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 20499a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 20509a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 20519a305230SLee Schermerhorn else 20529a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 20539a305230SLee Schermerhorn 20549a305230SLee Schermerhorn return sprintf(buf, "%lu\n", surplus_huge_pages); 2055a3437870SNishanth Aravamudan } 2056a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 2057a3437870SNishanth Aravamudan 2058a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 2059a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 2060a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 2061a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 2062a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 2063a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 206406808b08SLee Schermerhorn #ifdef CONFIG_NUMA 206506808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 206606808b08SLee Schermerhorn #endif 2067a3437870SNishanth Aravamudan NULL, 2068a3437870SNishanth Aravamudan }; 2069a3437870SNishanth Aravamudan 2070a3437870SNishanth Aravamudan static struct attribute_group hstate_attr_group = { 2071a3437870SNishanth Aravamudan .attrs = hstate_attrs, 2072a3437870SNishanth Aravamudan }; 2073a3437870SNishanth Aravamudan 2074094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 20759a305230SLee Schermerhorn struct kobject **hstate_kobjs, 20769a305230SLee Schermerhorn struct attribute_group *hstate_attr_group) 2077a3437870SNishanth Aravamudan { 2078a3437870SNishanth Aravamudan int retval; 2079972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 2080a3437870SNishanth Aravamudan 20819a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 20829a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 2083a3437870SNishanth Aravamudan return -ENOMEM; 2084a3437870SNishanth Aravamudan 20859a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 2086a3437870SNishanth Aravamudan if (retval) 20879a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 2088a3437870SNishanth Aravamudan 2089a3437870SNishanth Aravamudan return retval; 2090a3437870SNishanth Aravamudan } 2091a3437870SNishanth Aravamudan 2092a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 2093a3437870SNishanth Aravamudan { 2094a3437870SNishanth Aravamudan struct hstate *h; 2095a3437870SNishanth Aravamudan int err; 2096a3437870SNishanth Aravamudan 2097a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 2098a3437870SNishanth Aravamudan if (!hugepages_kobj) 2099a3437870SNishanth Aravamudan return; 2100a3437870SNishanth Aravamudan 2101a3437870SNishanth Aravamudan for_each_hstate(h) { 21029a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 21039a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 2104a3437870SNishanth Aravamudan if (err) 2105ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s", h->name); 2106a3437870SNishanth Aravamudan } 2107a3437870SNishanth Aravamudan } 2108a3437870SNishanth Aravamudan 21099a305230SLee Schermerhorn #ifdef CONFIG_NUMA 21109a305230SLee Schermerhorn 21119a305230SLee Schermerhorn /* 21129a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 211310fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 211410fbcf4cSKay Sievers * index of a node device or _hstate == node id. 211510fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 21169a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 21179a305230SLee Schermerhorn */ 21189a305230SLee Schermerhorn struct node_hstate { 21199a305230SLee Schermerhorn struct kobject *hugepages_kobj; 21209a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 21219a305230SLee Schermerhorn }; 21229a305230SLee Schermerhorn struct node_hstate node_hstates[MAX_NUMNODES]; 21239a305230SLee Schermerhorn 21249a305230SLee Schermerhorn /* 212510fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 21269a305230SLee Schermerhorn */ 21279a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 21289a305230SLee Schermerhorn &nr_hugepages_attr.attr, 21299a305230SLee Schermerhorn &free_hugepages_attr.attr, 21309a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 21319a305230SLee Schermerhorn NULL, 21329a305230SLee Schermerhorn }; 21339a305230SLee Schermerhorn 21349a305230SLee Schermerhorn static struct attribute_group per_node_hstate_attr_group = { 21359a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 21369a305230SLee Schermerhorn }; 21379a305230SLee Schermerhorn 21389a305230SLee Schermerhorn /* 213910fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 21409a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 21419a305230SLee Schermerhorn */ 21429a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 21439a305230SLee Schermerhorn { 21449a305230SLee Schermerhorn int nid; 21459a305230SLee Schermerhorn 21469a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 21479a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 21489a305230SLee Schermerhorn int i; 21499a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 21509a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 21519a305230SLee Schermerhorn if (nidp) 21529a305230SLee Schermerhorn *nidp = nid; 21539a305230SLee Schermerhorn return &hstates[i]; 21549a305230SLee Schermerhorn } 21559a305230SLee Schermerhorn } 21569a305230SLee Schermerhorn 21579a305230SLee Schermerhorn BUG(); 21589a305230SLee Schermerhorn return NULL; 21599a305230SLee Schermerhorn } 21609a305230SLee Schermerhorn 21619a305230SLee Schermerhorn /* 216210fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 21639a305230SLee Schermerhorn * No-op if no hstate attributes attached. 21649a305230SLee Schermerhorn */ 21653cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 21669a305230SLee Schermerhorn { 21679a305230SLee Schermerhorn struct hstate *h; 216810fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 21699a305230SLee Schermerhorn 21709a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 21719b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 21729a305230SLee Schermerhorn 2173972dc4deSAneesh Kumar K.V for_each_hstate(h) { 2174972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 2175972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 2176972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 2177972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 2178972dc4deSAneesh Kumar K.V } 21799a305230SLee Schermerhorn } 21809a305230SLee Schermerhorn 21819a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 21829a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 21839a305230SLee Schermerhorn } 21849a305230SLee Schermerhorn 21859a305230SLee Schermerhorn /* 218610fbcf4cSKay Sievers * hugetlb module exit: unregister hstate attributes from node devices 21879a305230SLee Schermerhorn * that have them. 21889a305230SLee Schermerhorn */ 21899a305230SLee Schermerhorn static void hugetlb_unregister_all_nodes(void) 21909a305230SLee Schermerhorn { 21919a305230SLee Schermerhorn int nid; 21929a305230SLee Schermerhorn 21939a305230SLee Schermerhorn /* 219410fbcf4cSKay Sievers * disable node device registrations. 21959a305230SLee Schermerhorn */ 21969a305230SLee Schermerhorn register_hugetlbfs_with_node(NULL, NULL); 21979a305230SLee Schermerhorn 21989a305230SLee Schermerhorn /* 21999a305230SLee Schermerhorn * remove hstate attributes from any nodes that have them. 22009a305230SLee Schermerhorn */ 22019a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) 22028732794bSWen Congyang hugetlb_unregister_node(node_devices[nid]); 22039a305230SLee Schermerhorn } 22049a305230SLee Schermerhorn 22059a305230SLee Schermerhorn /* 220610fbcf4cSKay Sievers * Register hstate attributes for a single node device. 22079a305230SLee Schermerhorn * No-op if attributes already registered. 22089a305230SLee Schermerhorn */ 22093cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 22109a305230SLee Schermerhorn { 22119a305230SLee Schermerhorn struct hstate *h; 221210fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 22139a305230SLee Schermerhorn int err; 22149a305230SLee Schermerhorn 22159a305230SLee Schermerhorn if (nhs->hugepages_kobj) 22169a305230SLee Schermerhorn return; /* already allocated */ 22179a305230SLee Schermerhorn 22189a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 221910fbcf4cSKay Sievers &node->dev.kobj); 22209a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 22219a305230SLee Schermerhorn return; 22229a305230SLee Schermerhorn 22239a305230SLee Schermerhorn for_each_hstate(h) { 22249a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 22259a305230SLee Schermerhorn nhs->hstate_kobjs, 22269a305230SLee Schermerhorn &per_node_hstate_attr_group); 22279a305230SLee Schermerhorn if (err) { 2228ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s for node %d\n", 222910fbcf4cSKay Sievers h->name, node->dev.id); 22309a305230SLee Schermerhorn hugetlb_unregister_node(node); 22319a305230SLee Schermerhorn break; 22329a305230SLee Schermerhorn } 22339a305230SLee Schermerhorn } 22349a305230SLee Schermerhorn } 22359a305230SLee Schermerhorn 22369a305230SLee Schermerhorn /* 22379b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 223810fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 223910fbcf4cSKay Sievers * registered their associated device by this time. 22409a305230SLee Schermerhorn */ 22417d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 22429a305230SLee Schermerhorn { 22439a305230SLee Schermerhorn int nid; 22449a305230SLee Schermerhorn 22458cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 22468732794bSWen Congyang struct node *node = node_devices[nid]; 224710fbcf4cSKay Sievers if (node->dev.id == nid) 22489a305230SLee Schermerhorn hugetlb_register_node(node); 22499a305230SLee Schermerhorn } 22509a305230SLee Schermerhorn 22519a305230SLee Schermerhorn /* 225210fbcf4cSKay Sievers * Let the node device driver know we're here so it can 22539a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 22549a305230SLee Schermerhorn */ 22559a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 22569a305230SLee Schermerhorn hugetlb_unregister_node); 22579a305230SLee Schermerhorn } 22589a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 22599a305230SLee Schermerhorn 22609a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 22619a305230SLee Schermerhorn { 22629a305230SLee Schermerhorn BUG(); 22639a305230SLee Schermerhorn if (nidp) 22649a305230SLee Schermerhorn *nidp = -1; 22659a305230SLee Schermerhorn return NULL; 22669a305230SLee Schermerhorn } 22679a305230SLee Schermerhorn 22689a305230SLee Schermerhorn static void hugetlb_unregister_all_nodes(void) { } 22699a305230SLee Schermerhorn 22709a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 22719a305230SLee Schermerhorn 22729a305230SLee Schermerhorn #endif 22739a305230SLee Schermerhorn 2274a3437870SNishanth Aravamudan static void __exit hugetlb_exit(void) 2275a3437870SNishanth Aravamudan { 2276a3437870SNishanth Aravamudan struct hstate *h; 2277a3437870SNishanth Aravamudan 22789a305230SLee Schermerhorn hugetlb_unregister_all_nodes(); 22799a305230SLee Schermerhorn 2280a3437870SNishanth Aravamudan for_each_hstate(h) { 2281972dc4deSAneesh Kumar K.V kobject_put(hstate_kobjs[hstate_index(h)]); 2282a3437870SNishanth Aravamudan } 2283a3437870SNishanth Aravamudan 2284a3437870SNishanth Aravamudan kobject_put(hugepages_kobj); 22858382d914SDavidlohr Bueso kfree(htlb_fault_mutex_table); 2286a3437870SNishanth Aravamudan } 2287a3437870SNishanth Aravamudan module_exit(hugetlb_exit); 2288a3437870SNishanth Aravamudan 2289a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 2290a3437870SNishanth Aravamudan { 22918382d914SDavidlohr Bueso int i; 22928382d914SDavidlohr Bueso 2293457c1b27SNishanth Aravamudan if (!hugepages_supported()) 22940ef89d25SBenjamin Herrenschmidt return 0; 2295a3437870SNishanth Aravamudan 2296e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) { 2297e11bfbfcSNick Piggin default_hstate_size = HPAGE_SIZE; 2298e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) 2299a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 2300a3437870SNishanth Aravamudan } 2301972dc4deSAneesh Kumar K.V default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); 2302e11bfbfcSNick Piggin if (default_hstate_max_huge_pages) 2303e11bfbfcSNick Piggin default_hstate.max_huge_pages = default_hstate_max_huge_pages; 2304a3437870SNishanth Aravamudan 2305a3437870SNishanth Aravamudan hugetlb_init_hstates(); 2306aa888a74SAndi Kleen gather_bootmem_prealloc(); 2307a3437870SNishanth Aravamudan report_hugepages(); 2308a3437870SNishanth Aravamudan 2309a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 23109a305230SLee Schermerhorn hugetlb_register_all_nodes(); 23117179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 23129a305230SLee Schermerhorn 23138382d914SDavidlohr Bueso #ifdef CONFIG_SMP 23148382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 23158382d914SDavidlohr Bueso #else 23168382d914SDavidlohr Bueso num_fault_mutexes = 1; 23178382d914SDavidlohr Bueso #endif 23188382d914SDavidlohr Bueso htlb_fault_mutex_table = 23198382d914SDavidlohr Bueso kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL); 23208382d914SDavidlohr Bueso BUG_ON(!htlb_fault_mutex_table); 23218382d914SDavidlohr Bueso 23228382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 23238382d914SDavidlohr Bueso mutex_init(&htlb_fault_mutex_table[i]); 2324a3437870SNishanth Aravamudan return 0; 2325a3437870SNishanth Aravamudan } 2326a3437870SNishanth Aravamudan module_init(hugetlb_init); 2327a3437870SNishanth Aravamudan 2328a3437870SNishanth Aravamudan /* Should be called on processing a hugepagesz=... option */ 2329a3437870SNishanth Aravamudan void __init hugetlb_add_hstate(unsigned order) 2330a3437870SNishanth Aravamudan { 2331a3437870SNishanth Aravamudan struct hstate *h; 23328faa8b07SAndi Kleen unsigned long i; 23338faa8b07SAndi Kleen 2334a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 2335ffb22af5SAndrew Morton pr_warning("hugepagesz= specified twice, ignoring\n"); 2336a3437870SNishanth Aravamudan return; 2337a3437870SNishanth Aravamudan } 233847d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 2339a3437870SNishanth Aravamudan BUG_ON(order == 0); 234047d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 2341a3437870SNishanth Aravamudan h->order = order; 2342a3437870SNishanth Aravamudan h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 23438faa8b07SAndi Kleen h->nr_huge_pages = 0; 23448faa8b07SAndi Kleen h->free_huge_pages = 0; 23458faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 23468faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 23470edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 23488cebfcd0SLai Jiangshan h->next_nid_to_alloc = first_node(node_states[N_MEMORY]); 23498cebfcd0SLai Jiangshan h->next_nid_to_free = first_node(node_states[N_MEMORY]); 2350a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 2351a3437870SNishanth Aravamudan huge_page_size(h)/1024); 23528faa8b07SAndi Kleen 2353a3437870SNishanth Aravamudan parsed_hstate = h; 2354a3437870SNishanth Aravamudan } 2355a3437870SNishanth Aravamudan 2356e11bfbfcSNick Piggin static int __init hugetlb_nrpages_setup(char *s) 2357a3437870SNishanth Aravamudan { 2358a3437870SNishanth Aravamudan unsigned long *mhp; 23598faa8b07SAndi Kleen static unsigned long *last_mhp; 2360a3437870SNishanth Aravamudan 2361a3437870SNishanth Aravamudan /* 236247d38344SAneesh Kumar K.V * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, 2363a3437870SNishanth Aravamudan * so this hugepages= parameter goes to the "default hstate". 2364a3437870SNishanth Aravamudan */ 236547d38344SAneesh Kumar K.V if (!hugetlb_max_hstate) 2366a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 2367a3437870SNishanth Aravamudan else 2368a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 2369a3437870SNishanth Aravamudan 23708faa8b07SAndi Kleen if (mhp == last_mhp) { 2371ffb22af5SAndrew Morton pr_warning("hugepages= specified twice without " 23728faa8b07SAndi Kleen "interleaving hugepagesz=, ignoring\n"); 23738faa8b07SAndi Kleen return 1; 23748faa8b07SAndi Kleen } 23758faa8b07SAndi Kleen 2376a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 2377a3437870SNishanth Aravamudan *mhp = 0; 2378a3437870SNishanth Aravamudan 23798faa8b07SAndi Kleen /* 23808faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 23818faa8b07SAndi Kleen * But we need to allocate >= MAX_ORDER hstates here early to still 23828faa8b07SAndi Kleen * use the bootmem allocator. 23838faa8b07SAndi Kleen */ 238447d38344SAneesh Kumar K.V if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) 23858faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 23868faa8b07SAndi Kleen 23878faa8b07SAndi Kleen last_mhp = mhp; 23888faa8b07SAndi Kleen 2389a3437870SNishanth Aravamudan return 1; 2390a3437870SNishanth Aravamudan } 2391e11bfbfcSNick Piggin __setup("hugepages=", hugetlb_nrpages_setup); 2392e11bfbfcSNick Piggin 2393e11bfbfcSNick Piggin static int __init hugetlb_default_setup(char *s) 2394e11bfbfcSNick Piggin { 2395e11bfbfcSNick Piggin default_hstate_size = memparse(s, &s); 2396e11bfbfcSNick Piggin return 1; 2397e11bfbfcSNick Piggin } 2398e11bfbfcSNick Piggin __setup("default_hugepagesz=", hugetlb_default_setup); 2399a3437870SNishanth Aravamudan 24008a213460SNishanth Aravamudan static unsigned int cpuset_mems_nr(unsigned int *array) 24018a213460SNishanth Aravamudan { 24028a213460SNishanth Aravamudan int node; 24038a213460SNishanth Aravamudan unsigned int nr = 0; 24048a213460SNishanth Aravamudan 24058a213460SNishanth Aravamudan for_each_node_mask(node, cpuset_current_mems_allowed) 24068a213460SNishanth Aravamudan nr += array[node]; 24078a213460SNishanth Aravamudan 24088a213460SNishanth Aravamudan return nr; 24098a213460SNishanth Aravamudan } 24108a213460SNishanth Aravamudan 24118a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 241206808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 241306808b08SLee Schermerhorn struct ctl_table *table, int write, 241406808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 24151da177e4SLinus Torvalds { 2416e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 2417238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 241808d4a246SMichal Hocko int ret; 2419e5ff2159SAndi Kleen 2420457c1b27SNishanth Aravamudan if (!hugepages_supported()) 2421457c1b27SNishanth Aravamudan return -ENOTSUPP; 2422457c1b27SNishanth Aravamudan 2423e5ff2159SAndi Kleen table->data = &tmp; 2424e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 242508d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 242608d4a246SMichal Hocko if (ret) 242708d4a246SMichal Hocko goto out; 2428e5ff2159SAndi Kleen 2429238d3c13SDavid Rientjes if (write) 2430238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 2431238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 243208d4a246SMichal Hocko out: 243308d4a246SMichal Hocko return ret; 24341da177e4SLinus Torvalds } 2435396faf03SMel Gorman 243606808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 243706808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 243806808b08SLee Schermerhorn { 243906808b08SLee Schermerhorn 244006808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 244106808b08SLee Schermerhorn buffer, length, ppos); 244206808b08SLee Schermerhorn } 244306808b08SLee Schermerhorn 244406808b08SLee Schermerhorn #ifdef CONFIG_NUMA 244506808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 244606808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 244706808b08SLee Schermerhorn { 244806808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 244906808b08SLee Schermerhorn buffer, length, ppos); 245006808b08SLee Schermerhorn } 245106808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 245206808b08SLee Schermerhorn 2453a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 24548d65af78SAlexey Dobriyan void __user *buffer, 2455a3d0c6aaSNishanth Aravamudan size_t *length, loff_t *ppos) 2456a3d0c6aaSNishanth Aravamudan { 2457a5516438SAndi Kleen struct hstate *h = &default_hstate; 2458e5ff2159SAndi Kleen unsigned long tmp; 245908d4a246SMichal Hocko int ret; 2460e5ff2159SAndi Kleen 2461457c1b27SNishanth Aravamudan if (!hugepages_supported()) 2462457c1b27SNishanth Aravamudan return -ENOTSUPP; 2463457c1b27SNishanth Aravamudan 2464e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 2465e5ff2159SAndi Kleen 2466bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 2467adbe8726SEric B Munson return -EINVAL; 2468adbe8726SEric B Munson 2469e5ff2159SAndi Kleen table->data = &tmp; 2470e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 247108d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 247208d4a246SMichal Hocko if (ret) 247308d4a246SMichal Hocko goto out; 2474e5ff2159SAndi Kleen 2475e5ff2159SAndi Kleen if (write) { 2476064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 2477e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 2478a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 2479e5ff2159SAndi Kleen } 248008d4a246SMichal Hocko out: 248108d4a246SMichal Hocko return ret; 2482a3d0c6aaSNishanth Aravamudan } 2483a3d0c6aaSNishanth Aravamudan 24841da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 24851da177e4SLinus Torvalds 2486e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 24871da177e4SLinus Torvalds { 2488a5516438SAndi Kleen struct hstate *h = &default_hstate; 2489457c1b27SNishanth Aravamudan if (!hugepages_supported()) 2490457c1b27SNishanth Aravamudan return; 2491e1759c21SAlexey Dobriyan seq_printf(m, 24921da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 24931da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 2494b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 24957893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 24964f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 2497a5516438SAndi Kleen h->nr_huge_pages, 2498a5516438SAndi Kleen h->free_huge_pages, 2499a5516438SAndi Kleen h->resv_huge_pages, 2500a5516438SAndi Kleen h->surplus_huge_pages, 2501a5516438SAndi Kleen 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 25021da177e4SLinus Torvalds } 25031da177e4SLinus Torvalds 25041da177e4SLinus Torvalds int hugetlb_report_node_meminfo(int nid, char *buf) 25051da177e4SLinus Torvalds { 2506a5516438SAndi Kleen struct hstate *h = &default_hstate; 2507457c1b27SNishanth Aravamudan if (!hugepages_supported()) 2508457c1b27SNishanth Aravamudan return 0; 25091da177e4SLinus Torvalds return sprintf(buf, 25101da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 2511a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 2512a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 2513a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 2514a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 2515a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 25161da177e4SLinus Torvalds } 25171da177e4SLinus Torvalds 2518949f7ec5SDavid Rientjes void hugetlb_show_meminfo(void) 2519949f7ec5SDavid Rientjes { 2520949f7ec5SDavid Rientjes struct hstate *h; 2521949f7ec5SDavid Rientjes int nid; 2522949f7ec5SDavid Rientjes 2523457c1b27SNishanth Aravamudan if (!hugepages_supported()) 2524457c1b27SNishanth Aravamudan return; 2525457c1b27SNishanth Aravamudan 2526949f7ec5SDavid Rientjes for_each_node_state(nid, N_MEMORY) 2527949f7ec5SDavid Rientjes for_each_hstate(h) 2528949f7ec5SDavid Rientjes pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 2529949f7ec5SDavid Rientjes nid, 2530949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 2531949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 2532949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 2533949f7ec5SDavid Rientjes 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 2534949f7ec5SDavid Rientjes } 2535949f7ec5SDavid Rientjes 25361da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 25371da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 25381da177e4SLinus Torvalds { 2539d0028588SWanpeng Li struct hstate *h; 2540d0028588SWanpeng Li unsigned long nr_total_pages = 0; 2541d0028588SWanpeng Li 2542d0028588SWanpeng Li for_each_hstate(h) 2543d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 2544d0028588SWanpeng Li return nr_total_pages; 25451da177e4SLinus Torvalds } 25461da177e4SLinus Torvalds 2547a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 2548fc1b8a73SMel Gorman { 2549fc1b8a73SMel Gorman int ret = -ENOMEM; 2550fc1b8a73SMel Gorman 2551fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 2552fc1b8a73SMel Gorman /* 2553fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 2554fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 2555fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 2556fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 2557fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 2558fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 2559fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 2560fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 2561fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 2562fc1b8a73SMel Gorman * 2563fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 2564fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 2565fc1b8a73SMel Gorman * we fall back to check against current free page availability as 2566fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 2567fc1b8a73SMel Gorman * semantics that cpuset has. 2568fc1b8a73SMel Gorman */ 2569fc1b8a73SMel Gorman if (delta > 0) { 2570a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 2571fc1b8a73SMel Gorman goto out; 2572fc1b8a73SMel Gorman 2573a5516438SAndi Kleen if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 2574a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 2575fc1b8a73SMel Gorman goto out; 2576fc1b8a73SMel Gorman } 2577fc1b8a73SMel Gorman } 2578fc1b8a73SMel Gorman 2579fc1b8a73SMel Gorman ret = 0; 2580fc1b8a73SMel Gorman if (delta < 0) 2581a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 2582fc1b8a73SMel Gorman 2583fc1b8a73SMel Gorman out: 2584fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 2585fc1b8a73SMel Gorman return ret; 2586fc1b8a73SMel Gorman } 2587fc1b8a73SMel Gorman 258884afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 258984afd99bSAndy Whitcroft { 2590f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 259184afd99bSAndy Whitcroft 259284afd99bSAndy Whitcroft /* 259384afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 259484afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 259584afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 259625985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 259784afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 259884afd99bSAndy Whitcroft * new reference here without additional locking. 259984afd99bSAndy Whitcroft */ 26004e35f483SJoonsoo Kim if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 2601f522c3acSJoonsoo Kim kref_get(&resv->refs); 260284afd99bSAndy Whitcroft } 260384afd99bSAndy Whitcroft 2604a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 2605a1e78772SMel Gorman { 2606a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2607f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 260890481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 26094e35f483SJoonsoo Kim unsigned long reserve, start, end; 26101c5ecae3SMike Kravetz long gbl_reserve; 261184afd99bSAndy Whitcroft 26124e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 26134e35f483SJoonsoo Kim return; 26144e35f483SJoonsoo Kim 2615a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 2616a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 261784afd99bSAndy Whitcroft 26184e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 261984afd99bSAndy Whitcroft 2620f031dd27SJoonsoo Kim kref_put(&resv->refs, resv_map_release); 262184afd99bSAndy Whitcroft 26227251ff78SAdam Litke if (reserve) { 26231c5ecae3SMike Kravetz /* 26241c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 26251c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 26261c5ecae3SMike Kravetz */ 26271c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 26281c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 26297251ff78SAdam Litke } 2630a1e78772SMel Gorman } 2631a1e78772SMel Gorman 26321da177e4SLinus Torvalds /* 26331da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 26341da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 26351da177e4SLinus Torvalds * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 26361da177e4SLinus Torvalds * this far. 26371da177e4SLinus Torvalds */ 2638d0217ac0SNick Piggin static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) 26391da177e4SLinus Torvalds { 26401da177e4SLinus Torvalds BUG(); 2641d0217ac0SNick Piggin return 0; 26421da177e4SLinus Torvalds } 26431da177e4SLinus Torvalds 2644f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 2645d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 264684afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 2647a1e78772SMel Gorman .close = hugetlb_vm_op_close, 26481da177e4SLinus Torvalds }; 26491da177e4SLinus Torvalds 26501e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 26511e8f889bSDavid Gibson int writable) 265263551ae0SDavid Gibson { 265363551ae0SDavid Gibson pte_t entry; 265463551ae0SDavid Gibson 26551e8f889bSDavid Gibson if (writable) { 2656106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 2657106c992aSGerald Schaefer vma->vm_page_prot))); 265863551ae0SDavid Gibson } else { 2659106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 2660106c992aSGerald Schaefer vma->vm_page_prot)); 266163551ae0SDavid Gibson } 266263551ae0SDavid Gibson entry = pte_mkyoung(entry); 266363551ae0SDavid Gibson entry = pte_mkhuge(entry); 2664d9ed9faaSChris Metcalf entry = arch_make_huge_pte(entry, vma, page, writable); 266563551ae0SDavid Gibson 266663551ae0SDavid Gibson return entry; 266763551ae0SDavid Gibson } 266863551ae0SDavid Gibson 26691e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 26701e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 26711e8f889bSDavid Gibson { 26721e8f889bSDavid Gibson pte_t entry; 26731e8f889bSDavid Gibson 2674106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 267532f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 26764b3073e1SRussell King update_mmu_cache(vma, address, ptep); 26771e8f889bSDavid Gibson } 26781e8f889bSDavid Gibson 26794a705fefSNaoya Horiguchi static int is_hugetlb_entry_migration(pte_t pte) 26804a705fefSNaoya Horiguchi { 26814a705fefSNaoya Horiguchi swp_entry_t swp; 26824a705fefSNaoya Horiguchi 26834a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 26844a705fefSNaoya Horiguchi return 0; 26854a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 26864a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_migration_entry(swp)) 26874a705fefSNaoya Horiguchi return 1; 26884a705fefSNaoya Horiguchi else 26894a705fefSNaoya Horiguchi return 0; 26904a705fefSNaoya Horiguchi } 26914a705fefSNaoya Horiguchi 26924a705fefSNaoya Horiguchi static int is_hugetlb_entry_hwpoisoned(pte_t pte) 26934a705fefSNaoya Horiguchi { 26944a705fefSNaoya Horiguchi swp_entry_t swp; 26954a705fefSNaoya Horiguchi 26964a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 26974a705fefSNaoya Horiguchi return 0; 26984a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 26994a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_hwpoison_entry(swp)) 27004a705fefSNaoya Horiguchi return 1; 27014a705fefSNaoya Horiguchi else 27024a705fefSNaoya Horiguchi return 0; 27034a705fefSNaoya Horiguchi } 27041e8f889bSDavid Gibson 270563551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 270663551ae0SDavid Gibson struct vm_area_struct *vma) 270763551ae0SDavid Gibson { 270863551ae0SDavid Gibson pte_t *src_pte, *dst_pte, entry; 270963551ae0SDavid Gibson struct page *ptepage; 27101c59827dSHugh Dickins unsigned long addr; 27111e8f889bSDavid Gibson int cow; 2712a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2713a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 2714e8569dd2SAndreas Sandberg unsigned long mmun_start; /* For mmu_notifiers */ 2715e8569dd2SAndreas Sandberg unsigned long mmun_end; /* For mmu_notifiers */ 2716e8569dd2SAndreas Sandberg int ret = 0; 27171e8f889bSDavid Gibson 27181e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 271963551ae0SDavid Gibson 2720e8569dd2SAndreas Sandberg mmun_start = vma->vm_start; 2721e8569dd2SAndreas Sandberg mmun_end = vma->vm_end; 2722e8569dd2SAndreas Sandberg if (cow) 2723e8569dd2SAndreas Sandberg mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end); 2724e8569dd2SAndreas Sandberg 2725a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 2726cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 2727c74df32cSHugh Dickins src_pte = huge_pte_offset(src, addr); 2728c74df32cSHugh Dickins if (!src_pte) 2729c74df32cSHugh Dickins continue; 2730a5516438SAndi Kleen dst_pte = huge_pte_alloc(dst, addr, sz); 2731e8569dd2SAndreas Sandberg if (!dst_pte) { 2732e8569dd2SAndreas Sandberg ret = -ENOMEM; 2733e8569dd2SAndreas Sandberg break; 2734e8569dd2SAndreas Sandberg } 2735c5c99429SLarry Woodman 2736c5c99429SLarry Woodman /* If the pagetables are shared don't copy or take references */ 2737c5c99429SLarry Woodman if (dst_pte == src_pte) 2738c5c99429SLarry Woodman continue; 2739c5c99429SLarry Woodman 2740cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 2741cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 2742cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 27434a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 27444a705fefSNaoya Horiguchi if (huge_pte_none(entry)) { /* skip none entry */ 27454a705fefSNaoya Horiguchi ; 27464a705fefSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry) || 27474a705fefSNaoya Horiguchi is_hugetlb_entry_hwpoisoned(entry))) { 27484a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 27494a705fefSNaoya Horiguchi 27504a705fefSNaoya Horiguchi if (is_write_migration_entry(swp_entry) && cow) { 27514a705fefSNaoya Horiguchi /* 27524a705fefSNaoya Horiguchi * COW mappings require pages in both 27534a705fefSNaoya Horiguchi * parent and child to be set to read. 27544a705fefSNaoya Horiguchi */ 27554a705fefSNaoya Horiguchi make_migration_entry_read(&swp_entry); 27564a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 27574a705fefSNaoya Horiguchi set_huge_pte_at(src, addr, src_pte, entry); 27584a705fefSNaoya Horiguchi } 27594a705fefSNaoya Horiguchi set_huge_pte_at(dst, addr, dst_pte, entry); 27604a705fefSNaoya Horiguchi } else { 276134ee645eSJoerg Roedel if (cow) { 27627f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 276334ee645eSJoerg Roedel mmu_notifier_invalidate_range(src, mmun_start, 276434ee645eSJoerg Roedel mmun_end); 276534ee645eSJoerg Roedel } 27660253d634SNaoya Horiguchi entry = huge_ptep_get(src_pte); 276763551ae0SDavid Gibson ptepage = pte_page(entry); 276863551ae0SDavid Gibson get_page(ptepage); 27690fe6e20bSNaoya Horiguchi page_dup_rmap(ptepage); 277063551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 27711c59827dSHugh Dickins } 2772cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 2773cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 277463551ae0SDavid Gibson } 277563551ae0SDavid Gibson 2776e8569dd2SAndreas Sandberg if (cow) 2777e8569dd2SAndreas Sandberg mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end); 2778e8569dd2SAndreas Sandberg 2779e8569dd2SAndreas Sandberg return ret; 278063551ae0SDavid Gibson } 278163551ae0SDavid Gibson 278224669e58SAneesh Kumar K.V void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 278324669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 278424669e58SAneesh Kumar K.V struct page *ref_page) 278563551ae0SDavid Gibson { 278624669e58SAneesh Kumar K.V int force_flush = 0; 278763551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 278863551ae0SDavid Gibson unsigned long address; 2789c7546f8fSDavid Gibson pte_t *ptep; 279063551ae0SDavid Gibson pte_t pte; 2791cb900f41SKirill A. Shutemov spinlock_t *ptl; 279263551ae0SDavid Gibson struct page *page; 2793a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2794a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 27952ec74c3eSSagi Grimberg const unsigned long mmun_start = start; /* For mmu_notifiers */ 27962ec74c3eSSagi Grimberg const unsigned long mmun_end = end; /* For mmu_notifiers */ 2797a5516438SAndi Kleen 279863551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 2799a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 2800a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 280163551ae0SDavid Gibson 280224669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 28032ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 2804569f48b8SHillf Danton address = start; 280524669e58SAneesh Kumar K.V again: 2806569f48b8SHillf Danton for (; address < end; address += sz) { 2807c7546f8fSDavid Gibson ptep = huge_pte_offset(mm, address); 2808c7546f8fSDavid Gibson if (!ptep) 2809c7546f8fSDavid Gibson continue; 2810c7546f8fSDavid Gibson 2811cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 281239dde65cSChen, Kenneth W if (huge_pmd_unshare(mm, &address, ptep)) 2813cb900f41SKirill A. Shutemov goto unlock; 281439dde65cSChen, Kenneth W 28156629326bSHillf Danton pte = huge_ptep_get(ptep); 28166629326bSHillf Danton if (huge_pte_none(pte)) 2817cb900f41SKirill A. Shutemov goto unlock; 28186629326bSHillf Danton 28196629326bSHillf Danton /* 28209fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 28219fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 28226629326bSHillf Danton */ 28239fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 2824106c992aSGerald Schaefer huge_pte_clear(mm, address, ptep); 2825cb900f41SKirill A. Shutemov goto unlock; 28268c4894c6SNaoya Horiguchi } 28276629326bSHillf Danton 28286629326bSHillf Danton page = pte_page(pte); 282904f2cbe3SMel Gorman /* 283004f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 283104f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 283204f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 283304f2cbe3SMel Gorman */ 283404f2cbe3SMel Gorman if (ref_page) { 283504f2cbe3SMel Gorman if (page != ref_page) 2836cb900f41SKirill A. Shutemov goto unlock; 283704f2cbe3SMel Gorman 283804f2cbe3SMel Gorman /* 283904f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 284004f2cbe3SMel Gorman * future faults in this VMA will fail rather than 284104f2cbe3SMel Gorman * looking like data was lost 284204f2cbe3SMel Gorman */ 284304f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 284404f2cbe3SMel Gorman } 284504f2cbe3SMel Gorman 2846c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 284724669e58SAneesh Kumar K.V tlb_remove_tlb_entry(tlb, ptep, address); 2848106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 28496649a386SKen Chen set_page_dirty(page); 28509e81130bSHillf Danton 285124669e58SAneesh Kumar K.V page_remove_rmap(page); 285224669e58SAneesh Kumar K.V force_flush = !__tlb_remove_page(tlb, page); 2853cb900f41SKirill A. Shutemov if (force_flush) { 2854569f48b8SHillf Danton address += sz; 2855cb900f41SKirill A. Shutemov spin_unlock(ptl); 28569e81130bSHillf Danton break; 285763551ae0SDavid Gibson } 2858cb900f41SKirill A. Shutemov /* Bail out after unmapping reference page if supplied */ 2859cb900f41SKirill A. Shutemov if (ref_page) { 2860cb900f41SKirill A. Shutemov spin_unlock(ptl); 2861cb900f41SKirill A. Shutemov break; 2862cb900f41SKirill A. Shutemov } 2863cb900f41SKirill A. Shutemov unlock: 2864cb900f41SKirill A. Shutemov spin_unlock(ptl); 2865cb900f41SKirill A. Shutemov } 286624669e58SAneesh Kumar K.V /* 286724669e58SAneesh Kumar K.V * mmu_gather ran out of room to batch pages, we break out of 286824669e58SAneesh Kumar K.V * the PTE lock to avoid doing the potential expensive TLB invalidate 286924669e58SAneesh Kumar K.V * and page-free while holding it. 287024669e58SAneesh Kumar K.V */ 287124669e58SAneesh Kumar K.V if (force_flush) { 287224669e58SAneesh Kumar K.V force_flush = 0; 287324669e58SAneesh Kumar K.V tlb_flush_mmu(tlb); 287424669e58SAneesh Kumar K.V if (address < end && !ref_page) 287524669e58SAneesh Kumar K.V goto again; 2876fe1668aeSChen, Kenneth W } 28772ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 287824669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 28791da177e4SLinus Torvalds } 288063551ae0SDavid Gibson 2881d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 2882d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 2883d833352aSMel Gorman unsigned long end, struct page *ref_page) 2884d833352aSMel Gorman { 2885d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 2886d833352aSMel Gorman 2887d833352aSMel Gorman /* 2888d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 2889d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 2890d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 2891d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 2892d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 2893c8c06efaSDavidlohr Bueso * is to clear it before releasing the i_mmap_rwsem. This works 2894d833352aSMel Gorman * because in the context this is called, the VMA is about to be 2895c8c06efaSDavidlohr Bueso * destroyed and the i_mmap_rwsem is held. 2896d833352aSMel Gorman */ 2897d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 2898d833352aSMel Gorman } 2899d833352aSMel Gorman 2900502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 290104f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 2902502717f4SChen, Kenneth W { 290324669e58SAneesh Kumar K.V struct mm_struct *mm; 290424669e58SAneesh Kumar K.V struct mmu_gather tlb; 290524669e58SAneesh Kumar K.V 290624669e58SAneesh Kumar K.V mm = vma->vm_mm; 290724669e58SAneesh Kumar K.V 29082b047252SLinus Torvalds tlb_gather_mmu(&tlb, mm, start, end); 290924669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 291024669e58SAneesh Kumar K.V tlb_finish_mmu(&tlb, start, end); 2911502717f4SChen, Kenneth W } 2912502717f4SChen, Kenneth W 291304f2cbe3SMel Gorman /* 291404f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 291504f2cbe3SMel Gorman * mappping it owns the reserve page for. The intention is to unmap the page 291604f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 291704f2cbe3SMel Gorman * same region. 291804f2cbe3SMel Gorman */ 29192f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 29202a4b3dedSHarvey Harrison struct page *page, unsigned long address) 292104f2cbe3SMel Gorman { 29227526674dSAdam Litke struct hstate *h = hstate_vma(vma); 292304f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 292404f2cbe3SMel Gorman struct address_space *mapping; 292504f2cbe3SMel Gorman pgoff_t pgoff; 292604f2cbe3SMel Gorman 292704f2cbe3SMel Gorman /* 292804f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 292904f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 293004f2cbe3SMel Gorman */ 29317526674dSAdam Litke address = address & huge_page_mask(h); 293236e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 293336e4f20aSMichal Hocko vma->vm_pgoff; 2934496ad9aaSAl Viro mapping = file_inode(vma->vm_file)->i_mapping; 293504f2cbe3SMel Gorman 29364eb2b1dcSMel Gorman /* 29374eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 29384eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 29394eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 29404eb2b1dcSMel Gorman */ 294183cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 29426b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 294304f2cbe3SMel Gorman /* Do not unmap the current VMA */ 294404f2cbe3SMel Gorman if (iter_vma == vma) 294504f2cbe3SMel Gorman continue; 294604f2cbe3SMel Gorman 294704f2cbe3SMel Gorman /* 294804f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 294904f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 295004f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 295104f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 295204f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 295304f2cbe3SMel Gorman */ 295404f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 295524669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 295624669e58SAneesh Kumar K.V address + huge_page_size(h), page); 295704f2cbe3SMel Gorman } 295883cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 295904f2cbe3SMel Gorman } 296004f2cbe3SMel Gorman 29610fe6e20bSNaoya Horiguchi /* 29620fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 2963ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 2964ef009b25SMichal Hocko * cannot race with other handlers or page migration. 2965ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 29660fe6e20bSNaoya Horiguchi */ 29671e8f889bSDavid Gibson static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 296804f2cbe3SMel Gorman unsigned long address, pte_t *ptep, pte_t pte, 2969cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 29701e8f889bSDavid Gibson { 2971a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 29721e8f889bSDavid Gibson struct page *old_page, *new_page; 2973ad4404a2SDavidlohr Bueso int ret = 0, outside_reserve = 0; 29742ec74c3eSSagi Grimberg unsigned long mmun_start; /* For mmu_notifiers */ 29752ec74c3eSSagi Grimberg unsigned long mmun_end; /* For mmu_notifiers */ 29761e8f889bSDavid Gibson 29771e8f889bSDavid Gibson old_page = pte_page(pte); 29781e8f889bSDavid Gibson 297904f2cbe3SMel Gorman retry_avoidcopy: 29801e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 29811e8f889bSDavid Gibson * and just make the page writable */ 298237a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 29830fe6e20bSNaoya Horiguchi page_move_anon_rmap(old_page, vma, address); 29841e8f889bSDavid Gibson set_huge_ptep_writable(vma, address, ptep); 298583c54070SNick Piggin return 0; 29861e8f889bSDavid Gibson } 29871e8f889bSDavid Gibson 298804f2cbe3SMel Gorman /* 298904f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 299004f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 299104f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 299204f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 299304f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 299404f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 299504f2cbe3SMel Gorman * of the full address range. 299604f2cbe3SMel Gorman */ 29975944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 299804f2cbe3SMel Gorman old_page != pagecache_page) 299904f2cbe3SMel Gorman outside_reserve = 1; 300004f2cbe3SMel Gorman 30011e8f889bSDavid Gibson page_cache_get(old_page); 3002b76c8cfbSLarry Woodman 3003ad4404a2SDavidlohr Bueso /* 3004ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 3005ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 3006ad4404a2SDavidlohr Bueso */ 3007cb900f41SKirill A. Shutemov spin_unlock(ptl); 300804f2cbe3SMel Gorman new_page = alloc_huge_page(vma, address, outside_reserve); 30091e8f889bSDavid Gibson 30102fc39cecSAdam Litke if (IS_ERR(new_page)) { 301104f2cbe3SMel Gorman /* 301204f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 301304f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 301404f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 301504f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 301604f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 301704f2cbe3SMel Gorman */ 301804f2cbe3SMel Gorman if (outside_reserve) { 3019ad4404a2SDavidlohr Bueso page_cache_release(old_page); 302004f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 30212f4612afSDavidlohr Bueso unmap_ref_private(mm, vma, old_page, address); 302204f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 3023cb900f41SKirill A. Shutemov spin_lock(ptl); 3024a734bcc8SHillf Danton ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 3025a9af0c5dSNaoya Horiguchi if (likely(ptep && 3026a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 302704f2cbe3SMel Gorman goto retry_avoidcopy; 3028a734bcc8SHillf Danton /* 3029cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 3030cb900f41SKirill A. Shutemov * lock, and our job is done. 3031a734bcc8SHillf Danton */ 3032a734bcc8SHillf Danton return 0; 303304f2cbe3SMel Gorman } 303404f2cbe3SMel Gorman 3035ad4404a2SDavidlohr Bueso ret = (PTR_ERR(new_page) == -ENOMEM) ? 3036ad4404a2SDavidlohr Bueso VM_FAULT_OOM : VM_FAULT_SIGBUS; 3037ad4404a2SDavidlohr Bueso goto out_release_old; 30381e8f889bSDavid Gibson } 30391e8f889bSDavid Gibson 30400fe6e20bSNaoya Horiguchi /* 30410fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 30420fe6e20bSNaoya Horiguchi * anon_vma prepared. 30430fe6e20bSNaoya Horiguchi */ 304444e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 3045ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 3046ad4404a2SDavidlohr Bueso goto out_release_all; 304744e2aa93SDean Nelson } 30480fe6e20bSNaoya Horiguchi 304947ad8475SAndrea Arcangeli copy_user_huge_page(new_page, old_page, address, vma, 305047ad8475SAndrea Arcangeli pages_per_huge_page(h)); 30510ed361deSNick Piggin __SetPageUptodate(new_page); 3052bcc54222SNaoya Horiguchi set_page_huge_active(new_page); 30531e8f889bSDavid Gibson 30542ec74c3eSSagi Grimberg mmun_start = address & huge_page_mask(h); 30552ec74c3eSSagi Grimberg mmun_end = mmun_start + huge_page_size(h); 30562ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 3057ad4404a2SDavidlohr Bueso 3058b76c8cfbSLarry Woodman /* 3059cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 3060b76c8cfbSLarry Woodman * before the page tables are altered 3061b76c8cfbSLarry Woodman */ 3062cb900f41SKirill A. Shutemov spin_lock(ptl); 3063a5516438SAndi Kleen ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 3064a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 306507443a85SJoonsoo Kim ClearPagePrivate(new_page); 306607443a85SJoonsoo Kim 30671e8f889bSDavid Gibson /* Break COW */ 30688fe627ecSGerald Schaefer huge_ptep_clear_flush(vma, address, ptep); 306934ee645eSJoerg Roedel mmu_notifier_invalidate_range(mm, mmun_start, mmun_end); 30701e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, 30711e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 30720fe6e20bSNaoya Horiguchi page_remove_rmap(old_page); 3073cd67f0d2SNaoya Horiguchi hugepage_add_new_anon_rmap(new_page, vma, address); 30741e8f889bSDavid Gibson /* Make the old page be freed below */ 30751e8f889bSDavid Gibson new_page = old_page; 30761e8f889bSDavid Gibson } 3077cb900f41SKirill A. Shutemov spin_unlock(ptl); 30782ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 3079ad4404a2SDavidlohr Bueso out_release_all: 30801e8f889bSDavid Gibson page_cache_release(new_page); 3081ad4404a2SDavidlohr Bueso out_release_old: 30821e8f889bSDavid Gibson page_cache_release(old_page); 30838312034fSJoonsoo Kim 3084ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 3085ad4404a2SDavidlohr Bueso return ret; 30861e8f889bSDavid Gibson } 30871e8f889bSDavid Gibson 308804f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 3089a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 3090a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 309104f2cbe3SMel Gorman { 309204f2cbe3SMel Gorman struct address_space *mapping; 3093e7c4b0bfSAndy Whitcroft pgoff_t idx; 309404f2cbe3SMel Gorman 309504f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 3096a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 309704f2cbe3SMel Gorman 309804f2cbe3SMel Gorman return find_lock_page(mapping, idx); 309904f2cbe3SMel Gorman } 310004f2cbe3SMel Gorman 31013ae77f43SHugh Dickins /* 31023ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 31033ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 31043ae77f43SHugh Dickins */ 31053ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 31062a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 31072a15efc9SHugh Dickins { 31082a15efc9SHugh Dickins struct address_space *mapping; 31092a15efc9SHugh Dickins pgoff_t idx; 31102a15efc9SHugh Dickins struct page *page; 31112a15efc9SHugh Dickins 31122a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 31132a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 31142a15efc9SHugh Dickins 31152a15efc9SHugh Dickins page = find_get_page(mapping, idx); 31162a15efc9SHugh Dickins if (page) 31172a15efc9SHugh Dickins put_page(page); 31182a15efc9SHugh Dickins return page != NULL; 31192a15efc9SHugh Dickins } 31202a15efc9SHugh Dickins 3121a1ed3ddaSRobert P. J. Day static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, 31228382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 3123788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 3124ac9b9c66SHugh Dickins { 3125a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3126ac9b9c66SHugh Dickins int ret = VM_FAULT_SIGBUS; 3127409eb8c2SHillf Danton int anon_rmap = 0; 31284c887265SAdam Litke unsigned long size; 31294c887265SAdam Litke struct page *page; 31301e8f889bSDavid Gibson pte_t new_pte; 3131cb900f41SKirill A. Shutemov spinlock_t *ptl; 31324c887265SAdam Litke 313304f2cbe3SMel Gorman /* 313404f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 313504f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 313625985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 313704f2cbe3SMel Gorman */ 313804f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 3139ffb22af5SAndrew Morton pr_warning("PID %d killed due to inadequate hugepage pool\n", 314004f2cbe3SMel Gorman current->pid); 314104f2cbe3SMel Gorman return ret; 314204f2cbe3SMel Gorman } 314304f2cbe3SMel Gorman 31444c887265SAdam Litke /* 31454c887265SAdam Litke * Use page lock to guard against racing truncation 31464c887265SAdam Litke * before we get page_table_lock. 31474c887265SAdam Litke */ 31486bda666aSChristoph Lameter retry: 31496bda666aSChristoph Lameter page = find_lock_page(mapping, idx); 31506bda666aSChristoph Lameter if (!page) { 3151a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 3152ebed4bfcSHugh Dickins if (idx >= size) 3153ebed4bfcSHugh Dickins goto out; 315404f2cbe3SMel Gorman page = alloc_huge_page(vma, address, 0); 31552fc39cecSAdam Litke if (IS_ERR(page)) { 315676dcee75SAneesh Kumar K.V ret = PTR_ERR(page); 315776dcee75SAneesh Kumar K.V if (ret == -ENOMEM) 315876dcee75SAneesh Kumar K.V ret = VM_FAULT_OOM; 315976dcee75SAneesh Kumar K.V else 316076dcee75SAneesh Kumar K.V ret = VM_FAULT_SIGBUS; 31616bda666aSChristoph Lameter goto out; 31626bda666aSChristoph Lameter } 316347ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 31640ed361deSNick Piggin __SetPageUptodate(page); 3165bcc54222SNaoya Horiguchi set_page_huge_active(page); 3166ac9b9c66SHugh Dickins 3167f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 31686bda666aSChristoph Lameter int err; 316945c682a6SKen Chen struct inode *inode = mapping->host; 31706bda666aSChristoph Lameter 31716bda666aSChristoph Lameter err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 31726bda666aSChristoph Lameter if (err) { 31736bda666aSChristoph Lameter put_page(page); 31746bda666aSChristoph Lameter if (err == -EEXIST) 31756bda666aSChristoph Lameter goto retry; 31766bda666aSChristoph Lameter goto out; 31776bda666aSChristoph Lameter } 317807443a85SJoonsoo Kim ClearPagePrivate(page); 317945c682a6SKen Chen 318045c682a6SKen Chen spin_lock(&inode->i_lock); 3181a5516438SAndi Kleen inode->i_blocks += blocks_per_huge_page(h); 318245c682a6SKen Chen spin_unlock(&inode->i_lock); 318323be7468SMel Gorman } else { 31846bda666aSChristoph Lameter lock_page(page); 31850fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 31860fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 31870fe6e20bSNaoya Horiguchi goto backout_unlocked; 318823be7468SMel Gorman } 3189409eb8c2SHillf Danton anon_rmap = 1; 31900fe6e20bSNaoya Horiguchi } 31910fe6e20bSNaoya Horiguchi } else { 319257303d80SAndy Whitcroft /* 3193998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 3194998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 3195998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 3196fd6a03edSNaoya Horiguchi */ 3197fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 3198aa50d3a7SAndi Kleen ret = VM_FAULT_HWPOISON | 3199972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 3200fd6a03edSNaoya Horiguchi goto backout_unlocked; 32016bda666aSChristoph Lameter } 3202998b4382SNaoya Horiguchi } 32031e8f889bSDavid Gibson 320457303d80SAndy Whitcroft /* 320557303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 320657303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 320757303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 320857303d80SAndy Whitcroft * the spinlock. 320957303d80SAndy Whitcroft */ 3210788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) 32112b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 32122b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 32132b26736cSAndy Whitcroft goto backout_unlocked; 32142b26736cSAndy Whitcroft } 321557303d80SAndy Whitcroft 3216cb900f41SKirill A. Shutemov ptl = huge_pte_lockptr(h, mm, ptep); 3217cb900f41SKirill A. Shutemov spin_lock(ptl); 3218a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 32194c887265SAdam Litke if (idx >= size) 32204c887265SAdam Litke goto backout; 32214c887265SAdam Litke 322283c54070SNick Piggin ret = 0; 32237f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 32244c887265SAdam Litke goto backout; 32254c887265SAdam Litke 322607443a85SJoonsoo Kim if (anon_rmap) { 322707443a85SJoonsoo Kim ClearPagePrivate(page); 3228409eb8c2SHillf Danton hugepage_add_new_anon_rmap(page, vma, address); 3229ac714904SChoi Gi-yong } else 3230409eb8c2SHillf Danton page_dup_rmap(page); 32311e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 32321e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 32331e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, new_pte); 32341e8f889bSDavid Gibson 3235788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 32361e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 3237cb900f41SKirill A. Shutemov ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl); 32381e8f889bSDavid Gibson } 32391e8f889bSDavid Gibson 3240cb900f41SKirill A. Shutemov spin_unlock(ptl); 32414c887265SAdam Litke unlock_page(page); 32424c887265SAdam Litke out: 3243ac9b9c66SHugh Dickins return ret; 32444c887265SAdam Litke 32454c887265SAdam Litke backout: 3246cb900f41SKirill A. Shutemov spin_unlock(ptl); 32472b26736cSAndy Whitcroft backout_unlocked: 32484c887265SAdam Litke unlock_page(page); 32494c887265SAdam Litke put_page(page); 32504c887265SAdam Litke goto out; 3251ac9b9c66SHugh Dickins } 3252ac9b9c66SHugh Dickins 32538382d914SDavidlohr Bueso #ifdef CONFIG_SMP 32548382d914SDavidlohr Bueso static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm, 32558382d914SDavidlohr Bueso struct vm_area_struct *vma, 32568382d914SDavidlohr Bueso struct address_space *mapping, 32578382d914SDavidlohr Bueso pgoff_t idx, unsigned long address) 32588382d914SDavidlohr Bueso { 32598382d914SDavidlohr Bueso unsigned long key[2]; 32608382d914SDavidlohr Bueso u32 hash; 32618382d914SDavidlohr Bueso 32628382d914SDavidlohr Bueso if (vma->vm_flags & VM_SHARED) { 32638382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 32648382d914SDavidlohr Bueso key[1] = idx; 32658382d914SDavidlohr Bueso } else { 32668382d914SDavidlohr Bueso key[0] = (unsigned long) mm; 32678382d914SDavidlohr Bueso key[1] = address >> huge_page_shift(h); 32688382d914SDavidlohr Bueso } 32698382d914SDavidlohr Bueso 32708382d914SDavidlohr Bueso hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0); 32718382d914SDavidlohr Bueso 32728382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 32738382d914SDavidlohr Bueso } 32748382d914SDavidlohr Bueso #else 32758382d914SDavidlohr Bueso /* 32768382d914SDavidlohr Bueso * For uniprocesor systems we always use a single mutex, so just 32778382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 32788382d914SDavidlohr Bueso */ 32798382d914SDavidlohr Bueso static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm, 32808382d914SDavidlohr Bueso struct vm_area_struct *vma, 32818382d914SDavidlohr Bueso struct address_space *mapping, 32828382d914SDavidlohr Bueso pgoff_t idx, unsigned long address) 32838382d914SDavidlohr Bueso { 32848382d914SDavidlohr Bueso return 0; 32858382d914SDavidlohr Bueso } 32868382d914SDavidlohr Bueso #endif 32878382d914SDavidlohr Bueso 328886e5216fSAdam Litke int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 3289788c7df4SHugh Dickins unsigned long address, unsigned int flags) 329086e5216fSAdam Litke { 32918382d914SDavidlohr Bueso pte_t *ptep, entry; 3292cb900f41SKirill A. Shutemov spinlock_t *ptl; 32931e8f889bSDavid Gibson int ret; 32948382d914SDavidlohr Bueso u32 hash; 32958382d914SDavidlohr Bueso pgoff_t idx; 32960fe6e20bSNaoya Horiguchi struct page *page = NULL; 329757303d80SAndy Whitcroft struct page *pagecache_page = NULL; 3298a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 32998382d914SDavidlohr Bueso struct address_space *mapping; 33000f792cf9SNaoya Horiguchi int need_wait_lock = 0; 330186e5216fSAdam Litke 33021e16a539SKAMEZAWA Hiroyuki address &= huge_page_mask(h); 33031e16a539SKAMEZAWA Hiroyuki 3304fd6a03edSNaoya Horiguchi ptep = huge_pte_offset(mm, address); 3305fd6a03edSNaoya Horiguchi if (ptep) { 3306fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 3307290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 3308cb900f41SKirill A. Shutemov migration_entry_wait_huge(vma, mm, ptep); 3309290408d4SNaoya Horiguchi return 0; 3310290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 3311aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 3312972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 3313fd6a03edSNaoya Horiguchi } 3314fd6a03edSNaoya Horiguchi 3315a5516438SAndi Kleen ptep = huge_pte_alloc(mm, address, huge_page_size(h)); 331686e5216fSAdam Litke if (!ptep) 331786e5216fSAdam Litke return VM_FAULT_OOM; 331886e5216fSAdam Litke 33198382d914SDavidlohr Bueso mapping = vma->vm_file->f_mapping; 33208382d914SDavidlohr Bueso idx = vma_hugecache_offset(h, vma, address); 33218382d914SDavidlohr Bueso 33223935baa9SDavid Gibson /* 33233935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 33243935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 33253935baa9SDavid Gibson * the same page in the page cache. 33263935baa9SDavid Gibson */ 33278382d914SDavidlohr Bueso hash = fault_mutex_hash(h, mm, vma, mapping, idx, address); 33288382d914SDavidlohr Bueso mutex_lock(&htlb_fault_mutex_table[hash]); 33298382d914SDavidlohr Bueso 33307f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 33317f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 33328382d914SDavidlohr Bueso ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 3333b4d1d99fSDavid Gibson goto out_mutex; 33343935baa9SDavid Gibson } 333586e5216fSAdam Litke 333683c54070SNick Piggin ret = 0; 33371e8f889bSDavid Gibson 333857303d80SAndy Whitcroft /* 33390f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 33400f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 33410f792cf9SNaoya Horiguchi * a active hugepage in pagecache. This goto expects the 2nd page fault, 33420f792cf9SNaoya Horiguchi * and is_hugetlb_entry_(migration|hwpoisoned) check will properly 33430f792cf9SNaoya Horiguchi * handle it. 33440f792cf9SNaoya Horiguchi */ 33450f792cf9SNaoya Horiguchi if (!pte_present(entry)) 33460f792cf9SNaoya Horiguchi goto out_mutex; 33470f792cf9SNaoya Horiguchi 33480f792cf9SNaoya Horiguchi /* 334957303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 335057303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 335157303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 335257303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 335357303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 335457303d80SAndy Whitcroft * consumed. 335557303d80SAndy Whitcroft */ 3356106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 33572b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 33582b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 3359b4d1d99fSDavid Gibson goto out_mutex; 33602b26736cSAndy Whitcroft } 336157303d80SAndy Whitcroft 3362f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 336357303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 336457303d80SAndy Whitcroft vma, address); 336557303d80SAndy Whitcroft } 336657303d80SAndy Whitcroft 33670f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 33680fe6e20bSNaoya Horiguchi 33691e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 3370b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 3371cb900f41SKirill A. Shutemov goto out_ptl; 3372b4d1d99fSDavid Gibson 33730f792cf9SNaoya Horiguchi /* 33740f792cf9SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 33750f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 33760f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 33770f792cf9SNaoya Horiguchi */ 33780f792cf9SNaoya Horiguchi page = pte_page(entry); 33790f792cf9SNaoya Horiguchi if (page != pagecache_page) 33800f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 33810f792cf9SNaoya Horiguchi need_wait_lock = 1; 33820f792cf9SNaoya Horiguchi goto out_ptl; 33830f792cf9SNaoya Horiguchi } 33840f792cf9SNaoya Horiguchi 33850f792cf9SNaoya Horiguchi get_page(page); 3386b4d1d99fSDavid Gibson 3387788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 3388106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 338957303d80SAndy Whitcroft ret = hugetlb_cow(mm, vma, address, ptep, entry, 3390cb900f41SKirill A. Shutemov pagecache_page, ptl); 33910f792cf9SNaoya Horiguchi goto out_put_page; 3392b4d1d99fSDavid Gibson } 3393106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 3394b4d1d99fSDavid Gibson } 3395b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 3396788c7df4SHugh Dickins if (huge_ptep_set_access_flags(vma, address, ptep, entry, 3397788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 33984b3073e1SRussell King update_mmu_cache(vma, address, ptep); 33990f792cf9SNaoya Horiguchi out_put_page: 34000f792cf9SNaoya Horiguchi if (page != pagecache_page) 34010f792cf9SNaoya Horiguchi unlock_page(page); 34020f792cf9SNaoya Horiguchi put_page(page); 3403cb900f41SKirill A. Shutemov out_ptl: 3404cb900f41SKirill A. Shutemov spin_unlock(ptl); 340557303d80SAndy Whitcroft 340657303d80SAndy Whitcroft if (pagecache_page) { 340757303d80SAndy Whitcroft unlock_page(pagecache_page); 340857303d80SAndy Whitcroft put_page(pagecache_page); 340957303d80SAndy Whitcroft } 3410b4d1d99fSDavid Gibson out_mutex: 34118382d914SDavidlohr Bueso mutex_unlock(&htlb_fault_mutex_table[hash]); 34120f792cf9SNaoya Horiguchi /* 34130f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 34140f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 34150f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 34160f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 34170f792cf9SNaoya Horiguchi * here without taking refcount. 34180f792cf9SNaoya Horiguchi */ 34190f792cf9SNaoya Horiguchi if (need_wait_lock) 34200f792cf9SNaoya Horiguchi wait_on_page_locked(page); 34211e8f889bSDavid Gibson return ret; 342286e5216fSAdam Litke } 342386e5216fSAdam Litke 342428a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 342563551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 342628a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 342728a35716SMichel Lespinasse long i, unsigned int flags) 342863551ae0SDavid Gibson { 3429d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 3430d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 343128a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 3432a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 343363551ae0SDavid Gibson 343463551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 343563551ae0SDavid Gibson pte_t *pte; 3436cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 34372a15efc9SHugh Dickins int absent; 343863551ae0SDavid Gibson struct page *page; 343963551ae0SDavid Gibson 34404c887265SAdam Litke /* 344102057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 344202057967SDavid Rientjes * potentially allocating memory. 344302057967SDavid Rientjes */ 344402057967SDavid Rientjes if (unlikely(fatal_signal_pending(current))) { 344502057967SDavid Rientjes remainder = 0; 344602057967SDavid Rientjes break; 344702057967SDavid Rientjes } 344802057967SDavid Rientjes 344902057967SDavid Rientjes /* 34504c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 34512a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 34524c887265SAdam Litke * first, for the page indexing below to work. 3453cb900f41SKirill A. Shutemov * 3454cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 34554c887265SAdam Litke */ 3456a5516438SAndi Kleen pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); 3457cb900f41SKirill A. Shutemov if (pte) 3458cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 34592a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 346063551ae0SDavid Gibson 34612a15efc9SHugh Dickins /* 34622a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 34633ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 34643ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 34653ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 34663ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 34672a15efc9SHugh Dickins */ 34683ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 34693ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 3470cb900f41SKirill A. Shutemov if (pte) 3471cb900f41SKirill A. Shutemov spin_unlock(ptl); 34722a15efc9SHugh Dickins remainder = 0; 34732a15efc9SHugh Dickins break; 34742a15efc9SHugh Dickins } 34752a15efc9SHugh Dickins 34769cc3a5bdSNaoya Horiguchi /* 34779cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 34789cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 34799cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 34809cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 34819cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 34829cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 34839cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 34849cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 34859cc3a5bdSNaoya Horiguchi */ 34869cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 3487106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 3488106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 34894c887265SAdam Litke int ret; 34904c887265SAdam Litke 3491cb900f41SKirill A. Shutemov if (pte) 3492cb900f41SKirill A. Shutemov spin_unlock(ptl); 34932a15efc9SHugh Dickins ret = hugetlb_fault(mm, vma, vaddr, 34942a15efc9SHugh Dickins (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); 3495a89182c7SAdam Litke if (!(ret & VM_FAULT_ERROR)) 34964c887265SAdam Litke continue; 34974c887265SAdam Litke 34981c59827dSHugh Dickins remainder = 0; 34991c59827dSHugh Dickins break; 35001c59827dSHugh Dickins } 350163551ae0SDavid Gibson 3502a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 35037f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 3504d5d4b0aaSChen, Kenneth W same_page: 3505d6692183SChen, Kenneth W if (pages) { 350669d177c2SAndy Whitcroft pages[i] = mem_map_offset(page, pfn_offset); 3507a0368d4eSAndrea Arcangeli get_page_foll(pages[i]); 3508d6692183SChen, Kenneth W } 350963551ae0SDavid Gibson 351063551ae0SDavid Gibson if (vmas) 351163551ae0SDavid Gibson vmas[i] = vma; 351263551ae0SDavid Gibson 351363551ae0SDavid Gibson vaddr += PAGE_SIZE; 3514d5d4b0aaSChen, Kenneth W ++pfn_offset; 351563551ae0SDavid Gibson --remainder; 351663551ae0SDavid Gibson ++i; 3517d5d4b0aaSChen, Kenneth W if (vaddr < vma->vm_end && remainder && 3518a5516438SAndi Kleen pfn_offset < pages_per_huge_page(h)) { 3519d5d4b0aaSChen, Kenneth W /* 3520d5d4b0aaSChen, Kenneth W * We use pfn_offset to avoid touching the pageframes 3521d5d4b0aaSChen, Kenneth W * of this compound page. 3522d5d4b0aaSChen, Kenneth W */ 3523d5d4b0aaSChen, Kenneth W goto same_page; 3524d5d4b0aaSChen, Kenneth W } 3525cb900f41SKirill A. Shutemov spin_unlock(ptl); 352663551ae0SDavid Gibson } 352728a35716SMichel Lespinasse *nr_pages = remainder; 352863551ae0SDavid Gibson *position = vaddr; 352963551ae0SDavid Gibson 35302a15efc9SHugh Dickins return i ? i : -EFAULT; 353163551ae0SDavid Gibson } 35328f860591SZhang, Yanmin 35337da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 35348f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 35358f860591SZhang, Yanmin { 35368f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 35378f860591SZhang, Yanmin unsigned long start = address; 35388f860591SZhang, Yanmin pte_t *ptep; 35398f860591SZhang, Yanmin pte_t pte; 3540a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 35417da4d641SPeter Zijlstra unsigned long pages = 0; 35428f860591SZhang, Yanmin 35438f860591SZhang, Yanmin BUG_ON(address >= end); 35448f860591SZhang, Yanmin flush_cache_range(vma, address, end); 35458f860591SZhang, Yanmin 3546a5338093SRik van Riel mmu_notifier_invalidate_range_start(mm, start, end); 354783cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 3548a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 3549cb900f41SKirill A. Shutemov spinlock_t *ptl; 35508f860591SZhang, Yanmin ptep = huge_pte_offset(mm, address); 35518f860591SZhang, Yanmin if (!ptep) 35528f860591SZhang, Yanmin continue; 3553cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 35547da4d641SPeter Zijlstra if (huge_pmd_unshare(mm, &address, ptep)) { 35557da4d641SPeter Zijlstra pages++; 3556cb900f41SKirill A. Shutemov spin_unlock(ptl); 355739dde65cSChen, Kenneth W continue; 35587da4d641SPeter Zijlstra } 3559a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 3560a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 3561a8bda28dSNaoya Horiguchi spin_unlock(ptl); 3562a8bda28dSNaoya Horiguchi continue; 3563a8bda28dSNaoya Horiguchi } 3564a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 3565a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 3566a8bda28dSNaoya Horiguchi 3567a8bda28dSNaoya Horiguchi if (is_write_migration_entry(entry)) { 3568a8bda28dSNaoya Horiguchi pte_t newpte; 3569a8bda28dSNaoya Horiguchi 3570a8bda28dSNaoya Horiguchi make_migration_entry_read(&entry); 3571a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 3572a8bda28dSNaoya Horiguchi set_huge_pte_at(mm, address, ptep, newpte); 3573a8bda28dSNaoya Horiguchi pages++; 3574a8bda28dSNaoya Horiguchi } 3575a8bda28dSNaoya Horiguchi spin_unlock(ptl); 3576a8bda28dSNaoya Horiguchi continue; 3577a8bda28dSNaoya Horiguchi } 3578a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 35798f860591SZhang, Yanmin pte = huge_ptep_get_and_clear(mm, address, ptep); 3580106c992aSGerald Schaefer pte = pte_mkhuge(huge_pte_modify(pte, newprot)); 3581be7517d6STony Lu pte = arch_make_huge_pte(pte, vma, NULL, 0); 35828f860591SZhang, Yanmin set_huge_pte_at(mm, address, ptep, pte); 35837da4d641SPeter Zijlstra pages++; 35848f860591SZhang, Yanmin } 3585cb900f41SKirill A. Shutemov spin_unlock(ptl); 35868f860591SZhang, Yanmin } 3587d833352aSMel Gorman /* 3588c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 3589d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 3590c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 3591d833352aSMel Gorman * and that page table be reused and filled with junk. 3592d833352aSMel Gorman */ 35938f860591SZhang, Yanmin flush_tlb_range(vma, start, end); 359434ee645eSJoerg Roedel mmu_notifier_invalidate_range(mm, start, end); 359583cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 3596a5338093SRik van Riel mmu_notifier_invalidate_range_end(mm, start, end); 35977da4d641SPeter Zijlstra 35987da4d641SPeter Zijlstra return pages << h->order; 35998f860591SZhang, Yanmin } 36008f860591SZhang, Yanmin 3601a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 3602a1e78772SMel Gorman long from, long to, 36035a6fe125SMel Gorman struct vm_area_struct *vma, 3604ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 3605e4e574b7SAdam Litke { 360617c9d12eSMel Gorman long ret, chg; 3607a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 360890481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 36099119a41eSJoonsoo Kim struct resv_map *resv_map; 36101c5ecae3SMike Kravetz long gbl_reserve; 3611e4e574b7SAdam Litke 3612a1e78772SMel Gorman /* 361317c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 361417c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 361590481622SDavid Gibson * without using reserves 361617c9d12eSMel Gorman */ 3617ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 361817c9d12eSMel Gorman return 0; 361917c9d12eSMel Gorman 362017c9d12eSMel Gorman /* 3621a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 3622a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 3623a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 3624a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 3625a1e78772SMel Gorman */ 36269119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 36274e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 36289119a41eSJoonsoo Kim 36291406ec9bSJoonsoo Kim chg = region_chg(resv_map, from, to); 36309119a41eSJoonsoo Kim 36319119a41eSJoonsoo Kim } else { 36329119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 36335a6fe125SMel Gorman if (!resv_map) 36345a6fe125SMel Gorman return -ENOMEM; 36355a6fe125SMel Gorman 363617c9d12eSMel Gorman chg = to - from; 363717c9d12eSMel Gorman 36385a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 36395a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 36405a6fe125SMel Gorman } 36415a6fe125SMel Gorman 3642c50ac050SDave Hansen if (chg < 0) { 3643c50ac050SDave Hansen ret = chg; 3644c50ac050SDave Hansen goto out_err; 3645c50ac050SDave Hansen } 364617c9d12eSMel Gorman 36471c5ecae3SMike Kravetz /* 36481c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 36491c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 36501c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 36511c5ecae3SMike Kravetz */ 36521c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 36531c5ecae3SMike Kravetz if (gbl_reserve < 0) { 3654c50ac050SDave Hansen ret = -ENOSPC; 3655c50ac050SDave Hansen goto out_err; 3656c50ac050SDave Hansen } 365717c9d12eSMel Gorman 365817c9d12eSMel Gorman /* 365917c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 366090481622SDavid Gibson * Hand the pages back to the subpool if there are not 366117c9d12eSMel Gorman */ 36621c5ecae3SMike Kravetz ret = hugetlb_acct_memory(h, gbl_reserve); 366317c9d12eSMel Gorman if (ret < 0) { 36641c5ecae3SMike Kravetz /* put back original number of pages, chg */ 36651c5ecae3SMike Kravetz (void)hugepage_subpool_put_pages(spool, chg); 3666c50ac050SDave Hansen goto out_err; 366717c9d12eSMel Gorman } 366817c9d12eSMel Gorman 366917c9d12eSMel Gorman /* 367017c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 367117c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 367217c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 367317c9d12eSMel Gorman * the reservation was and the page cache tells how much of 367417c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 367517c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 367617c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 367717c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 367817c9d12eSMel Gorman * else has to be done for private mappings here 367917c9d12eSMel Gorman */ 3680f83a275dSMel Gorman if (!vma || vma->vm_flags & VM_MAYSHARE) 36811406ec9bSJoonsoo Kim region_add(resv_map, from, to); 3682a43a8c39SChen, Kenneth W return 0; 3683c50ac050SDave Hansen out_err: 3684f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 3685f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 3686c50ac050SDave Hansen return ret; 3687a43a8c39SChen, Kenneth W } 3688a43a8c39SChen, Kenneth W 3689a43a8c39SChen, Kenneth W void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) 3690a43a8c39SChen, Kenneth W { 3691a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 36924e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 36939119a41eSJoonsoo Kim long chg = 0; 369490481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 36951c5ecae3SMike Kravetz long gbl_reserve; 369645c682a6SKen Chen 36979119a41eSJoonsoo Kim if (resv_map) 36981406ec9bSJoonsoo Kim chg = region_truncate(resv_map, offset); 369945c682a6SKen Chen spin_lock(&inode->i_lock); 3700e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 370145c682a6SKen Chen spin_unlock(&inode->i_lock); 370245c682a6SKen Chen 37031c5ecae3SMike Kravetz /* 37041c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 37051c5ecae3SMike Kravetz * reservations to be released may be adjusted. 37061c5ecae3SMike Kravetz */ 37071c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 37081c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 3709a43a8c39SChen, Kenneth W } 371093f70f90SNaoya Horiguchi 37113212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 37123212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 37133212b535SSteve Capper struct vm_area_struct *vma, 37143212b535SSteve Capper unsigned long addr, pgoff_t idx) 37153212b535SSteve Capper { 37163212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 37173212b535SSteve Capper svma->vm_start; 37183212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 37193212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 37203212b535SSteve Capper 37213212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 37223212b535SSteve Capper unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED; 37233212b535SSteve Capper unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED; 37243212b535SSteve Capper 37253212b535SSteve Capper /* 37263212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 37273212b535SSteve Capper * page table page. 37283212b535SSteve Capper */ 37293212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 37303212b535SSteve Capper vm_flags != svm_flags || 37313212b535SSteve Capper sbase < svma->vm_start || svma->vm_end < s_end) 37323212b535SSteve Capper return 0; 37333212b535SSteve Capper 37343212b535SSteve Capper return saddr; 37353212b535SSteve Capper } 37363212b535SSteve Capper 37373212b535SSteve Capper static int vma_shareable(struct vm_area_struct *vma, unsigned long addr) 37383212b535SSteve Capper { 37393212b535SSteve Capper unsigned long base = addr & PUD_MASK; 37403212b535SSteve Capper unsigned long end = base + PUD_SIZE; 37413212b535SSteve Capper 37423212b535SSteve Capper /* 37433212b535SSteve Capper * check on proper vm_flags and page table alignment 37443212b535SSteve Capper */ 37453212b535SSteve Capper if (vma->vm_flags & VM_MAYSHARE && 37463212b535SSteve Capper vma->vm_start <= base && end <= vma->vm_end) 37473212b535SSteve Capper return 1; 37483212b535SSteve Capper return 0; 37493212b535SSteve Capper } 37503212b535SSteve Capper 37513212b535SSteve Capper /* 37523212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 37533212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 37543212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 37553212b535SSteve Capper * code much cleaner. pmd allocation is essential for the shared case because 3756c8c06efaSDavidlohr Bueso * pud has to be populated inside the same i_mmap_rwsem section - otherwise 37573212b535SSteve Capper * racing tasks could either miss the sharing (see huge_pte_offset) or select a 37583212b535SSteve Capper * bad pmd for sharing. 37593212b535SSteve Capper */ 37603212b535SSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 37613212b535SSteve Capper { 37623212b535SSteve Capper struct vm_area_struct *vma = find_vma(mm, addr); 37633212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 37643212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 37653212b535SSteve Capper vma->vm_pgoff; 37663212b535SSteve Capper struct vm_area_struct *svma; 37673212b535SSteve Capper unsigned long saddr; 37683212b535SSteve Capper pte_t *spte = NULL; 37693212b535SSteve Capper pte_t *pte; 3770cb900f41SKirill A. Shutemov spinlock_t *ptl; 37713212b535SSteve Capper 37723212b535SSteve Capper if (!vma_shareable(vma, addr)) 37733212b535SSteve Capper return (pte_t *)pmd_alloc(mm, pud, addr); 37743212b535SSteve Capper 377583cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 37763212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 37773212b535SSteve Capper if (svma == vma) 37783212b535SSteve Capper continue; 37793212b535SSteve Capper 37803212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 37813212b535SSteve Capper if (saddr) { 37823212b535SSteve Capper spte = huge_pte_offset(svma->vm_mm, saddr); 37833212b535SSteve Capper if (spte) { 3784dc6c9a35SKirill A. Shutemov mm_inc_nr_pmds(mm); 37853212b535SSteve Capper get_page(virt_to_page(spte)); 37863212b535SSteve Capper break; 37873212b535SSteve Capper } 37883212b535SSteve Capper } 37893212b535SSteve Capper } 37903212b535SSteve Capper 37913212b535SSteve Capper if (!spte) 37923212b535SSteve Capper goto out; 37933212b535SSteve Capper 3794cb900f41SKirill A. Shutemov ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte); 3795cb900f41SKirill A. Shutemov spin_lock(ptl); 3796dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 37973212b535SSteve Capper pud_populate(mm, pud, 37983212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 3799dc6c9a35SKirill A. Shutemov } else { 38003212b535SSteve Capper put_page(virt_to_page(spte)); 3801dc6c9a35SKirill A. Shutemov mm_inc_nr_pmds(mm); 3802dc6c9a35SKirill A. Shutemov } 3803cb900f41SKirill A. Shutemov spin_unlock(ptl); 38043212b535SSteve Capper out: 38053212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 380683cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 38073212b535SSteve Capper return pte; 38083212b535SSteve Capper } 38093212b535SSteve Capper 38103212b535SSteve Capper /* 38113212b535SSteve Capper * unmap huge page backed by shared pte. 38123212b535SSteve Capper * 38133212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 38143212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 38153212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 38163212b535SSteve Capper * 3817cb900f41SKirill A. Shutemov * called with page table lock held. 38183212b535SSteve Capper * 38193212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 38203212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 38213212b535SSteve Capper */ 38223212b535SSteve Capper int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 38233212b535SSteve Capper { 38243212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 38253212b535SSteve Capper pud_t *pud = pud_offset(pgd, *addr); 38263212b535SSteve Capper 38273212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 38283212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 38293212b535SSteve Capper return 0; 38303212b535SSteve Capper 38313212b535SSteve Capper pud_clear(pud); 38323212b535SSteve Capper put_page(virt_to_page(ptep)); 3833dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 38343212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 38353212b535SSteve Capper return 1; 38363212b535SSteve Capper } 38379e5fc74cSSteve Capper #define want_pmd_share() (1) 38389e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 38399e5fc74cSSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 38409e5fc74cSSteve Capper { 38419e5fc74cSSteve Capper return NULL; 38429e5fc74cSSteve Capper } 3843e81f2d22SZhang Zhen 3844e81f2d22SZhang Zhen int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 3845e81f2d22SZhang Zhen { 3846e81f2d22SZhang Zhen return 0; 3847e81f2d22SZhang Zhen } 38489e5fc74cSSteve Capper #define want_pmd_share() (0) 38493212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 38503212b535SSteve Capper 38519e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 38529e5fc74cSSteve Capper pte_t *huge_pte_alloc(struct mm_struct *mm, 38539e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 38549e5fc74cSSteve Capper { 38559e5fc74cSSteve Capper pgd_t *pgd; 38569e5fc74cSSteve Capper pud_t *pud; 38579e5fc74cSSteve Capper pte_t *pte = NULL; 38589e5fc74cSSteve Capper 38599e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 38609e5fc74cSSteve Capper pud = pud_alloc(mm, pgd, addr); 38619e5fc74cSSteve Capper if (pud) { 38629e5fc74cSSteve Capper if (sz == PUD_SIZE) { 38639e5fc74cSSteve Capper pte = (pte_t *)pud; 38649e5fc74cSSteve Capper } else { 38659e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 38669e5fc74cSSteve Capper if (want_pmd_share() && pud_none(*pud)) 38679e5fc74cSSteve Capper pte = huge_pmd_share(mm, addr, pud); 38689e5fc74cSSteve Capper else 38699e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 38709e5fc74cSSteve Capper } 38719e5fc74cSSteve Capper } 38729e5fc74cSSteve Capper BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte)); 38739e5fc74cSSteve Capper 38749e5fc74cSSteve Capper return pte; 38759e5fc74cSSteve Capper } 38769e5fc74cSSteve Capper 38779e5fc74cSSteve Capper pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) 38789e5fc74cSSteve Capper { 38799e5fc74cSSteve Capper pgd_t *pgd; 38809e5fc74cSSteve Capper pud_t *pud; 38819e5fc74cSSteve Capper pmd_t *pmd = NULL; 38829e5fc74cSSteve Capper 38839e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 38849e5fc74cSSteve Capper if (pgd_present(*pgd)) { 38859e5fc74cSSteve Capper pud = pud_offset(pgd, addr); 38869e5fc74cSSteve Capper if (pud_present(*pud)) { 38879e5fc74cSSteve Capper if (pud_huge(*pud)) 38889e5fc74cSSteve Capper return (pte_t *)pud; 38899e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 38909e5fc74cSSteve Capper } 38919e5fc74cSSteve Capper } 38929e5fc74cSSteve Capper return (pte_t *) pmd; 38939e5fc74cSSteve Capper } 38949e5fc74cSSteve Capper 389561f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 389661f77edaSNaoya Horiguchi 389761f77edaSNaoya Horiguchi /* 389861f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 389961f77edaSNaoya Horiguchi * behavior. 390061f77edaSNaoya Horiguchi */ 390161f77edaSNaoya Horiguchi struct page * __weak 390261f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 390361f77edaSNaoya Horiguchi int write) 390461f77edaSNaoya Horiguchi { 390561f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 390661f77edaSNaoya Horiguchi } 390761f77edaSNaoya Horiguchi 390861f77edaSNaoya Horiguchi struct page * __weak 39099e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 3910e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 39119e5fc74cSSteve Capper { 3912e66f17ffSNaoya Horiguchi struct page *page = NULL; 3913e66f17ffSNaoya Horiguchi spinlock_t *ptl; 3914e66f17ffSNaoya Horiguchi retry: 3915e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 3916e66f17ffSNaoya Horiguchi spin_lock(ptl); 3917e66f17ffSNaoya Horiguchi /* 3918e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 3919e66f17ffSNaoya Horiguchi * unmapped from other threads. 3920e66f17ffSNaoya Horiguchi */ 3921e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 3922e66f17ffSNaoya Horiguchi goto out; 3923e66f17ffSNaoya Horiguchi if (pmd_present(*pmd)) { 392497534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 3925e66f17ffSNaoya Horiguchi if (flags & FOLL_GET) 3926e66f17ffSNaoya Horiguchi get_page(page); 3927e66f17ffSNaoya Horiguchi } else { 3928e66f17ffSNaoya Horiguchi if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) { 3929e66f17ffSNaoya Horiguchi spin_unlock(ptl); 3930e66f17ffSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pmd, ptl); 3931e66f17ffSNaoya Horiguchi goto retry; 3932e66f17ffSNaoya Horiguchi } 3933e66f17ffSNaoya Horiguchi /* 3934e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 3935e66f17ffSNaoya Horiguchi * follow_page_mask(). 3936e66f17ffSNaoya Horiguchi */ 3937e66f17ffSNaoya Horiguchi } 3938e66f17ffSNaoya Horiguchi out: 3939e66f17ffSNaoya Horiguchi spin_unlock(ptl); 39409e5fc74cSSteve Capper return page; 39419e5fc74cSSteve Capper } 39429e5fc74cSSteve Capper 394361f77edaSNaoya Horiguchi struct page * __weak 39449e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 3945e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 39469e5fc74cSSteve Capper { 3947e66f17ffSNaoya Horiguchi if (flags & FOLL_GET) 3948e66f17ffSNaoya Horiguchi return NULL; 39499e5fc74cSSteve Capper 3950e66f17ffSNaoya Horiguchi return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 39519e5fc74cSSteve Capper } 39529e5fc74cSSteve Capper 3953d5bd9106SAndi Kleen #ifdef CONFIG_MEMORY_FAILURE 3954d5bd9106SAndi Kleen 395593f70f90SNaoya Horiguchi /* 395693f70f90SNaoya Horiguchi * This function is called from memory failure code. 395793f70f90SNaoya Horiguchi * Assume the caller holds page lock of the head page. 395893f70f90SNaoya Horiguchi */ 39596de2b1aaSNaoya Horiguchi int dequeue_hwpoisoned_huge_page(struct page *hpage) 396093f70f90SNaoya Horiguchi { 396193f70f90SNaoya Horiguchi struct hstate *h = page_hstate(hpage); 396293f70f90SNaoya Horiguchi int nid = page_to_nid(hpage); 39636de2b1aaSNaoya Horiguchi int ret = -EBUSY; 396493f70f90SNaoya Horiguchi 396593f70f90SNaoya Horiguchi spin_lock(&hugetlb_lock); 39667e1f049eSNaoya Horiguchi /* 39677e1f049eSNaoya Horiguchi * Just checking !page_huge_active is not enough, because that could be 39687e1f049eSNaoya Horiguchi * an isolated/hwpoisoned hugepage (which have >0 refcount). 39697e1f049eSNaoya Horiguchi */ 39707e1f049eSNaoya Horiguchi if (!page_huge_active(hpage) && !page_count(hpage)) { 397156f2fb14SNaoya Horiguchi /* 397256f2fb14SNaoya Horiguchi * Hwpoisoned hugepage isn't linked to activelist or freelist, 397356f2fb14SNaoya Horiguchi * but dangling hpage->lru can trigger list-debug warnings 397456f2fb14SNaoya Horiguchi * (this happens when we call unpoison_memory() on it), 397556f2fb14SNaoya Horiguchi * so let it point to itself with list_del_init(). 397656f2fb14SNaoya Horiguchi */ 397756f2fb14SNaoya Horiguchi list_del_init(&hpage->lru); 39788c6c2ecbSNaoya Horiguchi set_page_refcounted(hpage); 397993f70f90SNaoya Horiguchi h->free_huge_pages--; 398093f70f90SNaoya Horiguchi h->free_huge_pages_node[nid]--; 39816de2b1aaSNaoya Horiguchi ret = 0; 398293f70f90SNaoya Horiguchi } 39836de2b1aaSNaoya Horiguchi spin_unlock(&hugetlb_lock); 39846de2b1aaSNaoya Horiguchi return ret; 39856de2b1aaSNaoya Horiguchi } 39866de2b1aaSNaoya Horiguchi #endif 398731caf665SNaoya Horiguchi 398831caf665SNaoya Horiguchi bool isolate_huge_page(struct page *page, struct list_head *list) 398931caf665SNaoya Horiguchi { 3990bcc54222SNaoya Horiguchi bool ret = true; 3991bcc54222SNaoya Horiguchi 3992309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 399331caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 3994bcc54222SNaoya Horiguchi if (!page_huge_active(page) || !get_page_unless_zero(page)) { 3995bcc54222SNaoya Horiguchi ret = false; 3996bcc54222SNaoya Horiguchi goto unlock; 3997bcc54222SNaoya Horiguchi } 3998bcc54222SNaoya Horiguchi clear_page_huge_active(page); 399931caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 4000bcc54222SNaoya Horiguchi unlock: 400131caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 4002bcc54222SNaoya Horiguchi return ret; 400331caf665SNaoya Horiguchi } 400431caf665SNaoya Horiguchi 400531caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 400631caf665SNaoya Horiguchi { 4007309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 400831caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 4009bcc54222SNaoya Horiguchi set_page_huge_active(page); 401031caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 401131caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 401231caf665SNaoya Horiguchi put_page(page); 401331caf665SNaoya Horiguchi } 4014