1457c8996SThomas Gleixner // SPDX-License-Identifier: GPL-2.0-only 21da177e4SLinus Torvalds /* 31da177e4SLinus Torvalds * Generic hugetlb support. 46d49e352SNadia Yvette Chambers * (C) Nadia Yvette Chambers, April 2004 51da177e4SLinus Torvalds */ 61da177e4SLinus Torvalds #include <linux/list.h> 71da177e4SLinus Torvalds #include <linux/init.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> 1997ad1087SMike Rapoport #include <linux/memblock.h> 20a3437870SNishanth Aravamudan #include <linux/sysfs.h> 215a0e3ad6STejun Heo #include <linux/slab.h> 22bbe88753SJoonsoo Kim #include <linux/sched/mm.h> 2363489f8eSMike Kravetz #include <linux/mmdebug.h> 24174cd4b1SIngo Molnar #include <linux/sched/signal.h> 250fe6e20bSNaoya Horiguchi #include <linux/rmap.h> 26c6247f72SMatthew Wilcox #include <linux/string_helpers.h> 27fd6a03edSNaoya Horiguchi #include <linux/swap.h> 28fd6a03edSNaoya Horiguchi #include <linux/swapops.h> 298382d914SDavidlohr Bueso #include <linux/jhash.h> 3098fa15f3SAnshuman Khandual #include <linux/numa.h> 31c77c0a8aSWaiman Long #include <linux/llist.h> 32cf11e85fSRoman Gushchin #include <linux/cma.h> 33d6606683SLinus Torvalds 3463551ae0SDavid Gibson #include <asm/page.h> 35ca15ca40SMike Rapoport #include <asm/pgalloc.h> 3624669e58SAneesh Kumar K.V #include <asm/tlb.h> 3763551ae0SDavid Gibson 3824669e58SAneesh Kumar K.V #include <linux/io.h> 3963551ae0SDavid Gibson #include <linux/hugetlb.h> 409dd540e2SAneesh Kumar K.V #include <linux/hugetlb_cgroup.h> 419a305230SLee Schermerhorn #include <linux/node.h> 421a1aad8aSMike Kravetz #include <linux/userfaultfd_k.h> 43ab5ac90aSMichal Hocko #include <linux/page_owner.h> 447835e98bSNick Piggin #include "internal.h" 451da177e4SLinus Torvalds 46c3f38a38SAneesh Kumar K.V int hugetlb_max_hstate __read_mostly; 47e5ff2159SAndi Kleen unsigned int default_hstate_idx; 48e5ff2159SAndi Kleen struct hstate hstates[HUGE_MAX_HSTATE]; 49cf11e85fSRoman Gushchin 50dbda8feaSBarry Song #ifdef CONFIG_CMA 51cf11e85fSRoman Gushchin static struct cma *hugetlb_cma[MAX_NUMNODES]; 52dbda8feaSBarry Song #endif 53dbda8feaSBarry Song static unsigned long hugetlb_cma_size __initdata; 54cf11e85fSRoman Gushchin 55641844f5SNaoya Horiguchi /* 56641844f5SNaoya Horiguchi * Minimum page order among possible hugepage sizes, set to a proper value 57641844f5SNaoya Horiguchi * at boot time. 58641844f5SNaoya Horiguchi */ 59641844f5SNaoya Horiguchi static unsigned int minimum_order __read_mostly = UINT_MAX; 60e5ff2159SAndi Kleen 6153ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 6253ba51d2SJon Tollefson 63e5ff2159SAndi Kleen /* for command line parsing */ 64e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 65e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 669fee021dSVaishali Thakkar static bool __initdata parsed_valid_hugepagesz = true; 67282f4214SMike Kravetz static bool __initdata parsed_default_hugepagesz; 68e5ff2159SAndi Kleen 693935baa9SDavid Gibson /* 7031caf665SNaoya Horiguchi * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, 7131caf665SNaoya Horiguchi * free_huge_pages, and surplus_huge_pages. 723935baa9SDavid Gibson */ 73c3f38a38SAneesh Kumar K.V DEFINE_SPINLOCK(hugetlb_lock); 740bd0f9fbSEric Paris 758382d914SDavidlohr Bueso /* 768382d914SDavidlohr Bueso * Serializes faults on the same logical page. This is used to 778382d914SDavidlohr Bueso * prevent spurious OOMs when the hugepage pool is fully utilized. 788382d914SDavidlohr Bueso */ 798382d914SDavidlohr Bueso static int num_fault_mutexes; 80c672c7f2SMike Kravetz struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; 818382d914SDavidlohr Bueso 827ffddd49SMuchun Song static inline bool PageHugeFreed(struct page *head) 837ffddd49SMuchun Song { 847ffddd49SMuchun Song return page_private(head + 4) == -1UL; 857ffddd49SMuchun Song } 867ffddd49SMuchun Song 877ffddd49SMuchun Song static inline void SetPageHugeFreed(struct page *head) 887ffddd49SMuchun Song { 897ffddd49SMuchun Song set_page_private(head + 4, -1UL); 907ffddd49SMuchun Song } 917ffddd49SMuchun Song 927ffddd49SMuchun Song static inline void ClearPageHugeFreed(struct page *head) 937ffddd49SMuchun Song { 947ffddd49SMuchun Song set_page_private(head + 4, 0); 957ffddd49SMuchun Song } 967ffddd49SMuchun Song 977ca02d0aSMike Kravetz /* Forward declaration */ 987ca02d0aSMike Kravetz static int hugetlb_acct_memory(struct hstate *h, long delta); 997ca02d0aSMike Kravetz 1001d88433bSMiaohe Lin static inline bool subpool_is_free(struct hugepage_subpool *spool) 1011d88433bSMiaohe Lin { 1021d88433bSMiaohe Lin if (spool->count) 1031d88433bSMiaohe Lin return false; 1041d88433bSMiaohe Lin if (spool->max_hpages != -1) 1051d88433bSMiaohe Lin return spool->used_hpages == 0; 1061d88433bSMiaohe Lin if (spool->min_hpages != -1) 1071d88433bSMiaohe Lin return spool->rsv_hpages == spool->min_hpages; 1081d88433bSMiaohe Lin 1091d88433bSMiaohe Lin return true; 1101d88433bSMiaohe Lin } 1111d88433bSMiaohe Lin 11290481622SDavid Gibson static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) 11390481622SDavid Gibson { 11490481622SDavid Gibson spin_unlock(&spool->lock); 11590481622SDavid Gibson 11690481622SDavid Gibson /* If no pages are used, and no other handles to the subpool 1177c8de358SEthon Paul * remain, give up any reservations based on minimum size and 1187ca02d0aSMike Kravetz * free the subpool */ 1191d88433bSMiaohe Lin if (subpool_is_free(spool)) { 1207ca02d0aSMike Kravetz if (spool->min_hpages != -1) 1217ca02d0aSMike Kravetz hugetlb_acct_memory(spool->hstate, 1227ca02d0aSMike Kravetz -spool->min_hpages); 12390481622SDavid Gibson kfree(spool); 12490481622SDavid Gibson } 1257ca02d0aSMike Kravetz } 12690481622SDavid Gibson 1277ca02d0aSMike Kravetz struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, 1287ca02d0aSMike Kravetz long min_hpages) 12990481622SDavid Gibson { 13090481622SDavid Gibson struct hugepage_subpool *spool; 13190481622SDavid Gibson 132c6a91820SMike Kravetz spool = kzalloc(sizeof(*spool), GFP_KERNEL); 13390481622SDavid Gibson if (!spool) 13490481622SDavid Gibson return NULL; 13590481622SDavid Gibson 13690481622SDavid Gibson spin_lock_init(&spool->lock); 13790481622SDavid Gibson spool->count = 1; 1387ca02d0aSMike Kravetz spool->max_hpages = max_hpages; 1397ca02d0aSMike Kravetz spool->hstate = h; 1407ca02d0aSMike Kravetz spool->min_hpages = min_hpages; 1417ca02d0aSMike Kravetz 1427ca02d0aSMike Kravetz if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { 1437ca02d0aSMike Kravetz kfree(spool); 1447ca02d0aSMike Kravetz return NULL; 1457ca02d0aSMike Kravetz } 1467ca02d0aSMike Kravetz spool->rsv_hpages = min_hpages; 14790481622SDavid Gibson 14890481622SDavid Gibson return spool; 14990481622SDavid Gibson } 15090481622SDavid Gibson 15190481622SDavid Gibson void hugepage_put_subpool(struct hugepage_subpool *spool) 15290481622SDavid Gibson { 15390481622SDavid Gibson spin_lock(&spool->lock); 15490481622SDavid Gibson BUG_ON(!spool->count); 15590481622SDavid Gibson spool->count--; 15690481622SDavid Gibson unlock_or_release_subpool(spool); 15790481622SDavid Gibson } 15890481622SDavid Gibson 1591c5ecae3SMike Kravetz /* 1601c5ecae3SMike Kravetz * Subpool accounting for allocating and reserving pages. 1611c5ecae3SMike Kravetz * Return -ENOMEM if there are not enough resources to satisfy the 1629e7ee400SRandy Dunlap * request. Otherwise, return the number of pages by which the 1631c5ecae3SMike Kravetz * global pools must be adjusted (upward). The returned value may 1641c5ecae3SMike Kravetz * only be different than the passed value (delta) in the case where 1657c8de358SEthon Paul * a subpool minimum size must be maintained. 1661c5ecae3SMike Kravetz */ 1671c5ecae3SMike Kravetz static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, 16890481622SDavid Gibson long delta) 16990481622SDavid Gibson { 1701c5ecae3SMike Kravetz long ret = delta; 17190481622SDavid Gibson 17290481622SDavid Gibson if (!spool) 1731c5ecae3SMike Kravetz return ret; 17490481622SDavid Gibson 17590481622SDavid Gibson spin_lock(&spool->lock); 17690481622SDavid Gibson 1771c5ecae3SMike Kravetz if (spool->max_hpages != -1) { /* maximum size accounting */ 1781c5ecae3SMike Kravetz if ((spool->used_hpages + delta) <= spool->max_hpages) 1791c5ecae3SMike Kravetz spool->used_hpages += delta; 1801c5ecae3SMike Kravetz else { 1811c5ecae3SMike Kravetz ret = -ENOMEM; 1821c5ecae3SMike Kravetz goto unlock_ret; 1831c5ecae3SMike Kravetz } 1841c5ecae3SMike Kravetz } 1851c5ecae3SMike Kravetz 18609a95e29SMike Kravetz /* minimum size accounting */ 18709a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->rsv_hpages) { 1881c5ecae3SMike Kravetz if (delta > spool->rsv_hpages) { 1891c5ecae3SMike Kravetz /* 1901c5ecae3SMike Kravetz * Asking for more reserves than those already taken on 1911c5ecae3SMike Kravetz * behalf of subpool. Return difference. 1921c5ecae3SMike Kravetz */ 1931c5ecae3SMike Kravetz ret = delta - spool->rsv_hpages; 1941c5ecae3SMike Kravetz spool->rsv_hpages = 0; 1951c5ecae3SMike Kravetz } else { 1961c5ecae3SMike Kravetz ret = 0; /* reserves already accounted for */ 1971c5ecae3SMike Kravetz spool->rsv_hpages -= delta; 1981c5ecae3SMike Kravetz } 1991c5ecae3SMike Kravetz } 2001c5ecae3SMike Kravetz 2011c5ecae3SMike Kravetz unlock_ret: 2021c5ecae3SMike Kravetz spin_unlock(&spool->lock); 20390481622SDavid Gibson return ret; 20490481622SDavid Gibson } 20590481622SDavid Gibson 2061c5ecae3SMike Kravetz /* 2071c5ecae3SMike Kravetz * Subpool accounting for freeing and unreserving pages. 2081c5ecae3SMike Kravetz * Return the number of global page reservations that must be dropped. 2091c5ecae3SMike Kravetz * The return value may only be different than the passed value (delta) 2101c5ecae3SMike Kravetz * in the case where a subpool minimum size must be maintained. 2111c5ecae3SMike Kravetz */ 2121c5ecae3SMike Kravetz static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, 21390481622SDavid Gibson long delta) 21490481622SDavid Gibson { 2151c5ecae3SMike Kravetz long ret = delta; 2161c5ecae3SMike Kravetz 21790481622SDavid Gibson if (!spool) 2181c5ecae3SMike Kravetz return delta; 21990481622SDavid Gibson 22090481622SDavid Gibson spin_lock(&spool->lock); 2211c5ecae3SMike Kravetz 2221c5ecae3SMike Kravetz if (spool->max_hpages != -1) /* maximum size accounting */ 22390481622SDavid Gibson spool->used_hpages -= delta; 2241c5ecae3SMike Kravetz 22509a95e29SMike Kravetz /* minimum size accounting */ 22609a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { 2271c5ecae3SMike Kravetz if (spool->rsv_hpages + delta <= spool->min_hpages) 2281c5ecae3SMike Kravetz ret = 0; 2291c5ecae3SMike Kravetz else 2301c5ecae3SMike Kravetz ret = spool->rsv_hpages + delta - spool->min_hpages; 2311c5ecae3SMike Kravetz 2321c5ecae3SMike Kravetz spool->rsv_hpages += delta; 2331c5ecae3SMike Kravetz if (spool->rsv_hpages > spool->min_hpages) 2341c5ecae3SMike Kravetz spool->rsv_hpages = spool->min_hpages; 2351c5ecae3SMike Kravetz } 2361c5ecae3SMike Kravetz 2371c5ecae3SMike Kravetz /* 2381c5ecae3SMike Kravetz * If hugetlbfs_put_super couldn't free spool due to an outstanding 2391c5ecae3SMike Kravetz * quota reference, free it now. 2401c5ecae3SMike Kravetz */ 24190481622SDavid Gibson unlock_or_release_subpool(spool); 2421c5ecae3SMike Kravetz 2431c5ecae3SMike Kravetz return ret; 24490481622SDavid Gibson } 24590481622SDavid Gibson 24690481622SDavid Gibson static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 24790481622SDavid Gibson { 24890481622SDavid Gibson return HUGETLBFS_SB(inode->i_sb)->spool; 24990481622SDavid Gibson } 25090481622SDavid Gibson 25190481622SDavid Gibson static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 25290481622SDavid Gibson { 253496ad9aaSAl Viro return subpool_inode(file_inode(vma->vm_file)); 25490481622SDavid Gibson } 25590481622SDavid Gibson 2560db9d74eSMina Almasry /* Helper that removes a struct file_region from the resv_map cache and returns 2570db9d74eSMina Almasry * it for use. 2580db9d74eSMina Almasry */ 2590db9d74eSMina Almasry static struct file_region * 2600db9d74eSMina Almasry get_file_region_entry_from_cache(struct resv_map *resv, long from, long to) 2610db9d74eSMina Almasry { 2620db9d74eSMina Almasry struct file_region *nrg = NULL; 2630db9d74eSMina Almasry 2640db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count <= 0); 2650db9d74eSMina Almasry 2660db9d74eSMina Almasry resv->region_cache_count--; 2670db9d74eSMina Almasry nrg = list_first_entry(&resv->region_cache, struct file_region, link); 2680db9d74eSMina Almasry list_del(&nrg->link); 2690db9d74eSMina Almasry 2700db9d74eSMina Almasry nrg->from = from; 2710db9d74eSMina Almasry nrg->to = to; 2720db9d74eSMina Almasry 2730db9d74eSMina Almasry return nrg; 2740db9d74eSMina Almasry } 2750db9d74eSMina Almasry 276075a61d0SMina Almasry static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, 277075a61d0SMina Almasry struct file_region *rg) 278075a61d0SMina Almasry { 279075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 280075a61d0SMina Almasry nrg->reservation_counter = rg->reservation_counter; 281075a61d0SMina Almasry nrg->css = rg->css; 282075a61d0SMina Almasry if (rg->css) 283075a61d0SMina Almasry css_get(rg->css); 284075a61d0SMina Almasry #endif 285075a61d0SMina Almasry } 286075a61d0SMina Almasry 287075a61d0SMina Almasry /* Helper that records hugetlb_cgroup uncharge info. */ 288075a61d0SMina Almasry static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, 289075a61d0SMina Almasry struct hstate *h, 290075a61d0SMina Almasry struct resv_map *resv, 291075a61d0SMina Almasry struct file_region *nrg) 292075a61d0SMina Almasry { 293075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 294075a61d0SMina Almasry if (h_cg) { 295075a61d0SMina Almasry nrg->reservation_counter = 296075a61d0SMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 297075a61d0SMina Almasry nrg->css = &h_cg->css; 298075a61d0SMina Almasry if (!resv->pages_per_hpage) 299075a61d0SMina Almasry resv->pages_per_hpage = pages_per_huge_page(h); 300075a61d0SMina Almasry /* pages_per_hpage should be the same for all entries in 301075a61d0SMina Almasry * a resv_map. 302075a61d0SMina Almasry */ 303075a61d0SMina Almasry VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); 304075a61d0SMina Almasry } else { 305075a61d0SMina Almasry nrg->reservation_counter = NULL; 306075a61d0SMina Almasry nrg->css = NULL; 307075a61d0SMina Almasry } 308075a61d0SMina Almasry #endif 309075a61d0SMina Almasry } 310075a61d0SMina Almasry 311a9b3f867SMina Almasry static bool has_same_uncharge_info(struct file_region *rg, 312a9b3f867SMina Almasry struct file_region *org) 313a9b3f867SMina Almasry { 314a9b3f867SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 315a9b3f867SMina Almasry return rg && org && 316a9b3f867SMina Almasry rg->reservation_counter == org->reservation_counter && 317a9b3f867SMina Almasry rg->css == org->css; 318a9b3f867SMina Almasry 319a9b3f867SMina Almasry #else 320a9b3f867SMina Almasry return true; 321a9b3f867SMina Almasry #endif 322a9b3f867SMina Almasry } 323a9b3f867SMina Almasry 324a9b3f867SMina Almasry static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) 325a9b3f867SMina Almasry { 326a9b3f867SMina Almasry struct file_region *nrg = NULL, *prg = NULL; 327a9b3f867SMina Almasry 328a9b3f867SMina Almasry prg = list_prev_entry(rg, link); 329a9b3f867SMina Almasry if (&prg->link != &resv->regions && prg->to == rg->from && 330a9b3f867SMina Almasry has_same_uncharge_info(prg, rg)) { 331a9b3f867SMina Almasry prg->to = rg->to; 332a9b3f867SMina Almasry 333a9b3f867SMina Almasry list_del(&rg->link); 334a9b3f867SMina Almasry kfree(rg); 335a9b3f867SMina Almasry 3367db5e7b6SWei Yang rg = prg; 337a9b3f867SMina Almasry } 338a9b3f867SMina Almasry 339a9b3f867SMina Almasry nrg = list_next_entry(rg, link); 340a9b3f867SMina Almasry if (&nrg->link != &resv->regions && nrg->from == rg->to && 341a9b3f867SMina Almasry has_same_uncharge_info(nrg, rg)) { 342a9b3f867SMina Almasry nrg->from = rg->from; 343a9b3f867SMina Almasry 344a9b3f867SMina Almasry list_del(&rg->link); 345a9b3f867SMina Almasry kfree(rg); 346a9b3f867SMina Almasry } 347a9b3f867SMina Almasry } 348a9b3f867SMina Almasry 349972a3da3SWei Yang /* 350972a3da3SWei Yang * Must be called with resv->lock held. 351972a3da3SWei Yang * 352972a3da3SWei Yang * Calling this with regions_needed != NULL will count the number of pages 353972a3da3SWei Yang * to be added but will not modify the linked list. And regions_needed will 354972a3da3SWei Yang * indicate the number of file_regions needed in the cache to carry out to add 355972a3da3SWei Yang * the regions for this range. 356d75c6af9SMina Almasry */ 357d75c6af9SMina Almasry static long add_reservation_in_range(struct resv_map *resv, long f, long t, 358075a61d0SMina Almasry struct hugetlb_cgroup *h_cg, 359972a3da3SWei Yang struct hstate *h, long *regions_needed) 360d75c6af9SMina Almasry { 3610db9d74eSMina Almasry long add = 0; 362d75c6af9SMina Almasry struct list_head *head = &resv->regions; 3630db9d74eSMina Almasry long last_accounted_offset = f; 364d75c6af9SMina Almasry struct file_region *rg = NULL, *trg = NULL, *nrg = NULL; 365d75c6af9SMina Almasry 3660db9d74eSMina Almasry if (regions_needed) 3670db9d74eSMina Almasry *regions_needed = 0; 368d75c6af9SMina Almasry 3690db9d74eSMina Almasry /* In this loop, we essentially handle an entry for the range 3700db9d74eSMina Almasry * [last_accounted_offset, rg->from), at every iteration, with some 3710db9d74eSMina Almasry * bounds checking. 3720db9d74eSMina Almasry */ 3730db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, head, link) { 3740db9d74eSMina Almasry /* Skip irrelevant regions that start before our range. */ 3750db9d74eSMina Almasry if (rg->from < f) { 3760db9d74eSMina Almasry /* If this region ends after the last accounted offset, 3770db9d74eSMina Almasry * then we need to update last_accounted_offset. 3780db9d74eSMina Almasry */ 3790db9d74eSMina Almasry if (rg->to > last_accounted_offset) 3800db9d74eSMina Almasry last_accounted_offset = rg->to; 3810db9d74eSMina Almasry continue; 3820db9d74eSMina Almasry } 383d75c6af9SMina Almasry 3840db9d74eSMina Almasry /* When we find a region that starts beyond our range, we've 3850db9d74eSMina Almasry * finished. 3860db9d74eSMina Almasry */ 387d75c6af9SMina Almasry if (rg->from > t) 388d75c6af9SMina Almasry break; 389d75c6af9SMina Almasry 3900db9d74eSMina Almasry /* Add an entry for last_accounted_offset -> rg->from, and 3910db9d74eSMina Almasry * update last_accounted_offset. 392d75c6af9SMina Almasry */ 3930db9d74eSMina Almasry if (rg->from > last_accounted_offset) { 3940db9d74eSMina Almasry add += rg->from - last_accounted_offset; 395972a3da3SWei Yang if (!regions_needed) { 3960db9d74eSMina Almasry nrg = get_file_region_entry_from_cache( 3970db9d74eSMina Almasry resv, last_accounted_offset, rg->from); 398075a61d0SMina Almasry record_hugetlb_cgroup_uncharge_info(h_cg, h, 399075a61d0SMina Almasry resv, nrg); 4000db9d74eSMina Almasry list_add(&nrg->link, rg->link.prev); 401a9b3f867SMina Almasry coalesce_file_region(resv, nrg); 402972a3da3SWei Yang } else 4030db9d74eSMina Almasry *regions_needed += 1; 404d75c6af9SMina Almasry } 405d75c6af9SMina Almasry 4060db9d74eSMina Almasry last_accounted_offset = rg->to; 4070db9d74eSMina Almasry } 4080db9d74eSMina Almasry 4090db9d74eSMina Almasry /* Handle the case where our range extends beyond 4100db9d74eSMina Almasry * last_accounted_offset. 4110db9d74eSMina Almasry */ 4120db9d74eSMina Almasry if (last_accounted_offset < t) { 4130db9d74eSMina Almasry add += t - last_accounted_offset; 414972a3da3SWei Yang if (!regions_needed) { 4150db9d74eSMina Almasry nrg = get_file_region_entry_from_cache( 4160db9d74eSMina Almasry resv, last_accounted_offset, t); 417075a61d0SMina Almasry record_hugetlb_cgroup_uncharge_info(h_cg, h, resv, nrg); 4180db9d74eSMina Almasry list_add(&nrg->link, rg->link.prev); 419a9b3f867SMina Almasry coalesce_file_region(resv, nrg); 420972a3da3SWei Yang } else 4210db9d74eSMina Almasry *regions_needed += 1; 4220db9d74eSMina Almasry } 4230db9d74eSMina Almasry 4240db9d74eSMina Almasry VM_BUG_ON(add < 0); 4250db9d74eSMina Almasry return add; 4260db9d74eSMina Almasry } 4270db9d74eSMina Almasry 4280db9d74eSMina Almasry /* Must be called with resv->lock acquired. Will drop lock to allocate entries. 4290db9d74eSMina Almasry */ 4300db9d74eSMina Almasry static int allocate_file_region_entries(struct resv_map *resv, 4310db9d74eSMina Almasry int regions_needed) 4320db9d74eSMina Almasry __must_hold(&resv->lock) 4330db9d74eSMina Almasry { 4340db9d74eSMina Almasry struct list_head allocated_regions; 4350db9d74eSMina Almasry int to_allocate = 0, i = 0; 4360db9d74eSMina Almasry struct file_region *trg = NULL, *rg = NULL; 4370db9d74eSMina Almasry 4380db9d74eSMina Almasry VM_BUG_ON(regions_needed < 0); 4390db9d74eSMina Almasry 4400db9d74eSMina Almasry INIT_LIST_HEAD(&allocated_regions); 4410db9d74eSMina Almasry 4420db9d74eSMina Almasry /* 4430db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 4440db9d74eSMina Almasry * the number of in progress add operations plus regions_needed. 4450db9d74eSMina Almasry * 4460db9d74eSMina Almasry * This is a while loop because when we drop the lock, some other call 4470db9d74eSMina Almasry * to region_add or region_del may have consumed some region_entries, 4480db9d74eSMina Almasry * so we keep looping here until we finally have enough entries for 4490db9d74eSMina Almasry * (adds_in_progress + regions_needed). 4500db9d74eSMina Almasry */ 4510db9d74eSMina Almasry while (resv->region_cache_count < 4520db9d74eSMina Almasry (resv->adds_in_progress + regions_needed)) { 4530db9d74eSMina Almasry to_allocate = resv->adds_in_progress + regions_needed - 4540db9d74eSMina Almasry resv->region_cache_count; 4550db9d74eSMina Almasry 4560db9d74eSMina Almasry /* At this point, we should have enough entries in the cache 4570db9d74eSMina Almasry * for all the existings adds_in_progress. We should only be 4580db9d74eSMina Almasry * needing to allocate for regions_needed. 4590db9d74eSMina Almasry */ 4600db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); 4610db9d74eSMina Almasry 4620db9d74eSMina Almasry spin_unlock(&resv->lock); 4630db9d74eSMina Almasry for (i = 0; i < to_allocate; i++) { 4640db9d74eSMina Almasry trg = kmalloc(sizeof(*trg), GFP_KERNEL); 4650db9d74eSMina Almasry if (!trg) 4660db9d74eSMina Almasry goto out_of_memory; 4670db9d74eSMina Almasry list_add(&trg->link, &allocated_regions); 4680db9d74eSMina Almasry } 4690db9d74eSMina Almasry 4700db9d74eSMina Almasry spin_lock(&resv->lock); 4710db9d74eSMina Almasry 472d3ec7b6eSWei Yang list_splice(&allocated_regions, &resv->region_cache); 473d3ec7b6eSWei Yang resv->region_cache_count += to_allocate; 4740db9d74eSMina Almasry } 4750db9d74eSMina Almasry 4760db9d74eSMina Almasry return 0; 4770db9d74eSMina Almasry 4780db9d74eSMina Almasry out_of_memory: 4790db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, &allocated_regions, link) { 480d75c6af9SMina Almasry list_del(&rg->link); 481d75c6af9SMina Almasry kfree(rg); 482d75c6af9SMina Almasry } 4830db9d74eSMina Almasry return -ENOMEM; 484d75c6af9SMina Almasry } 485d75c6af9SMina Almasry 4861dd308a7SMike Kravetz /* 4871dd308a7SMike Kravetz * Add the huge page range represented by [f, t) to the reserve 4880db9d74eSMina Almasry * map. Regions will be taken from the cache to fill in this range. 4890db9d74eSMina Almasry * Sufficient regions should exist in the cache due to the previous 4900db9d74eSMina Almasry * call to region_chg with the same range, but in some cases the cache will not 4910db9d74eSMina Almasry * have sufficient entries due to races with other code doing region_add or 4920db9d74eSMina Almasry * region_del. The extra needed entries will be allocated. 493cf3ad20bSMike Kravetz * 4940db9d74eSMina Almasry * regions_needed is the out value provided by a previous call to region_chg. 4950db9d74eSMina Almasry * 4960db9d74eSMina Almasry * Return the number of new huge pages added to the map. This number is greater 4970db9d74eSMina Almasry * than or equal to zero. If file_region entries needed to be allocated for 4987c8de358SEthon Paul * this operation and we were not able to allocate, it returns -ENOMEM. 4990db9d74eSMina Almasry * region_add of regions of length 1 never allocate file_regions and cannot 5000db9d74eSMina Almasry * fail; region_chg will always allocate at least 1 entry and a region_add for 5010db9d74eSMina Almasry * 1 page will only require at most 1 entry. 5021dd308a7SMike Kravetz */ 5030db9d74eSMina Almasry static long region_add(struct resv_map *resv, long f, long t, 504075a61d0SMina Almasry long in_regions_needed, struct hstate *h, 505075a61d0SMina Almasry struct hugetlb_cgroup *h_cg) 50696822904SAndy Whitcroft { 5070db9d74eSMina Almasry long add = 0, actual_regions_needed = 0; 50896822904SAndy Whitcroft 5097b24d861SDavidlohr Bueso spin_lock(&resv->lock); 5100db9d74eSMina Almasry retry: 5110db9d74eSMina Almasry 5120db9d74eSMina Almasry /* Count how many regions are actually needed to execute this add. */ 513972a3da3SWei Yang add_reservation_in_range(resv, f, t, NULL, NULL, 514972a3da3SWei Yang &actual_regions_needed); 51596822904SAndy Whitcroft 5165e911373SMike Kravetz /* 5170db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 5180db9d74eSMina Almasry * this add operation. Note that actual_regions_needed may be greater 5190db9d74eSMina Almasry * than in_regions_needed, as the resv_map may have been modified since 5200db9d74eSMina Almasry * the region_chg call. In this case, we need to make sure that we 5210db9d74eSMina Almasry * allocate extra entries, such that we have enough for all the 5220db9d74eSMina Almasry * existing adds_in_progress, plus the excess needed for this 5230db9d74eSMina Almasry * operation. 5245e911373SMike Kravetz */ 5250db9d74eSMina Almasry if (actual_regions_needed > in_regions_needed && 5260db9d74eSMina Almasry resv->region_cache_count < 5270db9d74eSMina Almasry resv->adds_in_progress + 5280db9d74eSMina Almasry (actual_regions_needed - in_regions_needed)) { 5290db9d74eSMina Almasry /* region_add operation of range 1 should never need to 5300db9d74eSMina Almasry * allocate file_region entries. 5310db9d74eSMina Almasry */ 5320db9d74eSMina Almasry VM_BUG_ON(t - f <= 1); 5335e911373SMike Kravetz 5340db9d74eSMina Almasry if (allocate_file_region_entries( 5350db9d74eSMina Almasry resv, actual_regions_needed - in_regions_needed)) { 5360db9d74eSMina Almasry return -ENOMEM; 5375e911373SMike Kravetz } 5385e911373SMike Kravetz 5390db9d74eSMina Almasry goto retry; 5400db9d74eSMina Almasry } 541cf3ad20bSMike Kravetz 542972a3da3SWei Yang add = add_reservation_in_range(resv, f, t, h_cg, h, NULL); 5430db9d74eSMina Almasry 5440db9d74eSMina Almasry resv->adds_in_progress -= in_regions_needed; 5450db9d74eSMina Almasry 5467b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 547cf3ad20bSMike Kravetz VM_BUG_ON(add < 0); 548cf3ad20bSMike Kravetz return add; 54996822904SAndy Whitcroft } 55096822904SAndy Whitcroft 5511dd308a7SMike Kravetz /* 5521dd308a7SMike Kravetz * Examine the existing reserve map and determine how many 5531dd308a7SMike Kravetz * huge pages in the specified range [f, t) are NOT currently 5541dd308a7SMike Kravetz * represented. This routine is called before a subsequent 5551dd308a7SMike Kravetz * call to region_add that will actually modify the reserve 5561dd308a7SMike Kravetz * map to add the specified range [f, t). region_chg does 5571dd308a7SMike Kravetz * not change the number of huge pages represented by the 5580db9d74eSMina Almasry * map. A number of new file_region structures is added to the cache as a 5590db9d74eSMina Almasry * placeholder, for the subsequent region_add call to use. At least 1 5600db9d74eSMina Almasry * file_region structure is added. 5610db9d74eSMina Almasry * 5620db9d74eSMina Almasry * out_regions_needed is the number of regions added to the 5630db9d74eSMina Almasry * resv->adds_in_progress. This value needs to be provided to a follow up call 5640db9d74eSMina Almasry * to region_add or region_abort for proper accounting. 5655e911373SMike Kravetz * 5665e911373SMike Kravetz * Returns the number of huge pages that need to be added to the existing 5675e911373SMike Kravetz * reservation map for the range [f, t). This number is greater or equal to 5685e911373SMike Kravetz * zero. -ENOMEM is returned if a new file_region structure or cache entry 5695e911373SMike Kravetz * is needed and can not be allocated. 5701dd308a7SMike Kravetz */ 5710db9d74eSMina Almasry static long region_chg(struct resv_map *resv, long f, long t, 5720db9d74eSMina Almasry long *out_regions_needed) 57396822904SAndy Whitcroft { 57496822904SAndy Whitcroft long chg = 0; 57596822904SAndy Whitcroft 5767b24d861SDavidlohr Bueso spin_lock(&resv->lock); 5775e911373SMike Kravetz 578972a3da3SWei Yang /* Count how many hugepages in this range are NOT represented. */ 579075a61d0SMina Almasry chg = add_reservation_in_range(resv, f, t, NULL, NULL, 580972a3da3SWei Yang out_regions_needed); 5815e911373SMike Kravetz 5820db9d74eSMina Almasry if (*out_regions_needed == 0) 5830db9d74eSMina Almasry *out_regions_needed = 1; 5845e911373SMike Kravetz 5850db9d74eSMina Almasry if (allocate_file_region_entries(resv, *out_regions_needed)) 5865e911373SMike Kravetz return -ENOMEM; 5875e911373SMike Kravetz 5880db9d74eSMina Almasry resv->adds_in_progress += *out_regions_needed; 58996822904SAndy Whitcroft 5907b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 59196822904SAndy Whitcroft return chg; 59296822904SAndy Whitcroft } 59396822904SAndy Whitcroft 5941dd308a7SMike Kravetz /* 5955e911373SMike Kravetz * Abort the in progress add operation. The adds_in_progress field 5965e911373SMike Kravetz * of the resv_map keeps track of the operations in progress between 5975e911373SMike Kravetz * calls to region_chg and region_add. Operations are sometimes 5985e911373SMike Kravetz * aborted after the call to region_chg. In such cases, region_abort 5990db9d74eSMina Almasry * is called to decrement the adds_in_progress counter. regions_needed 6000db9d74eSMina Almasry * is the value returned by the region_chg call, it is used to decrement 6010db9d74eSMina Almasry * the adds_in_progress counter. 6025e911373SMike Kravetz * 6035e911373SMike Kravetz * NOTE: The range arguments [f, t) are not needed or used in this 6045e911373SMike Kravetz * routine. They are kept to make reading the calling code easier as 6055e911373SMike Kravetz * arguments will match the associated region_chg call. 6065e911373SMike Kravetz */ 6070db9d74eSMina Almasry static void region_abort(struct resv_map *resv, long f, long t, 6080db9d74eSMina Almasry long regions_needed) 6095e911373SMike Kravetz { 6105e911373SMike Kravetz spin_lock(&resv->lock); 6115e911373SMike Kravetz VM_BUG_ON(!resv->region_cache_count); 6120db9d74eSMina Almasry resv->adds_in_progress -= regions_needed; 6135e911373SMike Kravetz spin_unlock(&resv->lock); 6145e911373SMike Kravetz } 6155e911373SMike Kravetz 6165e911373SMike Kravetz /* 617feba16e2SMike Kravetz * Delete the specified range [f, t) from the reserve map. If the 618feba16e2SMike Kravetz * t parameter is LONG_MAX, this indicates that ALL regions after f 619feba16e2SMike Kravetz * should be deleted. Locate the regions which intersect [f, t) 620feba16e2SMike Kravetz * and either trim, delete or split the existing regions. 621feba16e2SMike Kravetz * 622feba16e2SMike Kravetz * Returns the number of huge pages deleted from the reserve map. 623feba16e2SMike Kravetz * In the normal case, the return value is zero or more. In the 624feba16e2SMike Kravetz * case where a region must be split, a new region descriptor must 625feba16e2SMike Kravetz * be allocated. If the allocation fails, -ENOMEM will be returned. 626feba16e2SMike Kravetz * NOTE: If the parameter t == LONG_MAX, then we will never split 627feba16e2SMike Kravetz * a region and possibly return -ENOMEM. Callers specifying 628feba16e2SMike Kravetz * t == LONG_MAX do not need to check for -ENOMEM error. 6291dd308a7SMike Kravetz */ 630feba16e2SMike Kravetz static long region_del(struct resv_map *resv, long f, long t) 63196822904SAndy Whitcroft { 6321406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 63396822904SAndy Whitcroft struct file_region *rg, *trg; 634feba16e2SMike Kravetz struct file_region *nrg = NULL; 635feba16e2SMike Kravetz long del = 0; 63696822904SAndy Whitcroft 637feba16e2SMike Kravetz retry: 6387b24d861SDavidlohr Bueso spin_lock(&resv->lock); 639feba16e2SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 640dbe409e4SMike Kravetz /* 641dbe409e4SMike Kravetz * Skip regions before the range to be deleted. file_region 642dbe409e4SMike Kravetz * ranges are normally of the form [from, to). However, there 643dbe409e4SMike Kravetz * may be a "placeholder" entry in the map which is of the form 644dbe409e4SMike Kravetz * (from, to) with from == to. Check for placeholder entries 645dbe409e4SMike Kravetz * at the beginning of the range to be deleted. 646dbe409e4SMike Kravetz */ 647dbe409e4SMike Kravetz if (rg->to <= f && (rg->to != rg->from || rg->to != f)) 648feba16e2SMike Kravetz continue; 649dbe409e4SMike Kravetz 650feba16e2SMike Kravetz if (rg->from >= t) 65196822904SAndy Whitcroft break; 65296822904SAndy Whitcroft 653feba16e2SMike Kravetz if (f > rg->from && t < rg->to) { /* Must split region */ 654feba16e2SMike Kravetz /* 655feba16e2SMike Kravetz * Check for an entry in the cache before dropping 656feba16e2SMike Kravetz * lock and attempting allocation. 657feba16e2SMike Kravetz */ 658feba16e2SMike Kravetz if (!nrg && 659feba16e2SMike Kravetz resv->region_cache_count > resv->adds_in_progress) { 660feba16e2SMike Kravetz nrg = list_first_entry(&resv->region_cache, 661feba16e2SMike Kravetz struct file_region, 662feba16e2SMike Kravetz link); 663feba16e2SMike Kravetz list_del(&nrg->link); 664feba16e2SMike Kravetz resv->region_cache_count--; 66596822904SAndy Whitcroft } 66696822904SAndy Whitcroft 667feba16e2SMike Kravetz if (!nrg) { 668feba16e2SMike Kravetz spin_unlock(&resv->lock); 669feba16e2SMike Kravetz nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 670feba16e2SMike Kravetz if (!nrg) 671feba16e2SMike Kravetz return -ENOMEM; 672feba16e2SMike Kravetz goto retry; 673feba16e2SMike Kravetz } 674feba16e2SMike Kravetz 675feba16e2SMike Kravetz del += t - f; 67679aa925bSMike Kravetz hugetlb_cgroup_uncharge_file_region( 67779aa925bSMike Kravetz resv, rg, t - f); 678feba16e2SMike Kravetz 679feba16e2SMike Kravetz /* New entry for end of split region */ 680feba16e2SMike Kravetz nrg->from = t; 681feba16e2SMike Kravetz nrg->to = rg->to; 682075a61d0SMina Almasry 683075a61d0SMina Almasry copy_hugetlb_cgroup_uncharge_info(nrg, rg); 684075a61d0SMina Almasry 685feba16e2SMike Kravetz INIT_LIST_HEAD(&nrg->link); 686feba16e2SMike Kravetz 687feba16e2SMike Kravetz /* Original entry is trimmed */ 688feba16e2SMike Kravetz rg->to = f; 689feba16e2SMike Kravetz 690feba16e2SMike Kravetz list_add(&nrg->link, &rg->link); 691feba16e2SMike Kravetz nrg = NULL; 69296822904SAndy Whitcroft break; 693feba16e2SMike Kravetz } 694feba16e2SMike Kravetz 695feba16e2SMike Kravetz if (f <= rg->from && t >= rg->to) { /* Remove entire region */ 696feba16e2SMike Kravetz del += rg->to - rg->from; 697075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 698075a61d0SMina Almasry rg->to - rg->from); 69996822904SAndy Whitcroft list_del(&rg->link); 70096822904SAndy Whitcroft kfree(rg); 701feba16e2SMike Kravetz continue; 70296822904SAndy Whitcroft } 7037b24d861SDavidlohr Bueso 704feba16e2SMike Kravetz if (f <= rg->from) { /* Trim beginning of region */ 705075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 706075a61d0SMina Almasry t - rg->from); 707075a61d0SMina Almasry 70879aa925bSMike Kravetz del += t - rg->from; 70979aa925bSMike Kravetz rg->from = t; 71079aa925bSMike Kravetz } else { /* Trim end of region */ 711075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 712075a61d0SMina Almasry rg->to - f); 71379aa925bSMike Kravetz 71479aa925bSMike Kravetz del += rg->to - f; 71579aa925bSMike Kravetz rg->to = f; 716feba16e2SMike Kravetz } 717feba16e2SMike Kravetz } 718feba16e2SMike Kravetz 7197b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 720feba16e2SMike Kravetz kfree(nrg); 721feba16e2SMike Kravetz return del; 72296822904SAndy Whitcroft } 72396822904SAndy Whitcroft 7241dd308a7SMike Kravetz /* 725b5cec28dSMike Kravetz * A rare out of memory error was encountered which prevented removal of 726b5cec28dSMike Kravetz * the reserve map region for a page. The huge page itself was free'ed 727b5cec28dSMike Kravetz * and removed from the page cache. This routine will adjust the subpool 728b5cec28dSMike Kravetz * usage count, and the global reserve count if needed. By incrementing 729b5cec28dSMike Kravetz * these counts, the reserve map entry which could not be deleted will 730b5cec28dSMike Kravetz * appear as a "reserved" entry instead of simply dangling with incorrect 731b5cec28dSMike Kravetz * counts. 732b5cec28dSMike Kravetz */ 73372e2936cSzhong jiang void hugetlb_fix_reserve_counts(struct inode *inode) 734b5cec28dSMike Kravetz { 735b5cec28dSMike Kravetz struct hugepage_subpool *spool = subpool_inode(inode); 736b5cec28dSMike Kravetz long rsv_adjust; 737b5cec28dSMike Kravetz 738b5cec28dSMike Kravetz rsv_adjust = hugepage_subpool_get_pages(spool, 1); 73972e2936cSzhong jiang if (rsv_adjust) { 740b5cec28dSMike Kravetz struct hstate *h = hstate_inode(inode); 741b5cec28dSMike Kravetz 742b5cec28dSMike Kravetz hugetlb_acct_memory(h, 1); 743b5cec28dSMike Kravetz } 744b5cec28dSMike Kravetz } 745b5cec28dSMike Kravetz 746b5cec28dSMike Kravetz /* 7471dd308a7SMike Kravetz * Count and return the number of huge pages in the reserve map 7481dd308a7SMike Kravetz * that intersect with the range [f, t). 7491dd308a7SMike Kravetz */ 7501406ec9bSJoonsoo Kim static long region_count(struct resv_map *resv, long f, long t) 75184afd99bSAndy Whitcroft { 7521406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 75384afd99bSAndy Whitcroft struct file_region *rg; 75484afd99bSAndy Whitcroft long chg = 0; 75584afd99bSAndy Whitcroft 7567b24d861SDavidlohr Bueso spin_lock(&resv->lock); 75784afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 75884afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 759f2135a4aSWang Sheng-Hui long seg_from; 760f2135a4aSWang Sheng-Hui long seg_to; 76184afd99bSAndy Whitcroft 76284afd99bSAndy Whitcroft if (rg->to <= f) 76384afd99bSAndy Whitcroft continue; 76484afd99bSAndy Whitcroft if (rg->from >= t) 76584afd99bSAndy Whitcroft break; 76684afd99bSAndy Whitcroft 76784afd99bSAndy Whitcroft seg_from = max(rg->from, f); 76884afd99bSAndy Whitcroft seg_to = min(rg->to, t); 76984afd99bSAndy Whitcroft 77084afd99bSAndy Whitcroft chg += seg_to - seg_from; 77184afd99bSAndy Whitcroft } 7727b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 77384afd99bSAndy Whitcroft 77484afd99bSAndy Whitcroft return chg; 77584afd99bSAndy Whitcroft } 77684afd99bSAndy Whitcroft 77796822904SAndy Whitcroft /* 778e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 779e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 780e7c4b0bfSAndy Whitcroft */ 781a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 782a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 783e7c4b0bfSAndy Whitcroft { 784a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 785a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 786e7c4b0bfSAndy Whitcroft } 787e7c4b0bfSAndy Whitcroft 7880fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 7890fe6e20bSNaoya Horiguchi unsigned long address) 7900fe6e20bSNaoya Horiguchi { 7910fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 7920fe6e20bSNaoya Horiguchi } 793dee41079SDan Williams EXPORT_SYMBOL_GPL(linear_hugepage_index); 7940fe6e20bSNaoya Horiguchi 79584afd99bSAndy Whitcroft /* 79608fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 79708fba699SMel Gorman * cases this will be same size as used by the page table entries. 79808fba699SMel Gorman */ 79908fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 80008fba699SMel Gorman { 80105ea8860SDan Williams if (vma->vm_ops && vma->vm_ops->pagesize) 80205ea8860SDan Williams return vma->vm_ops->pagesize(vma); 80308fba699SMel Gorman return PAGE_SIZE; 80408fba699SMel Gorman } 805f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 80608fba699SMel Gorman 80708fba699SMel Gorman /* 8083340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 8093340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 81009135cc5SDan Williams * architectures where it differs, an architecture-specific 'strong' 81109135cc5SDan Williams * version of this symbol is required. 8123340289dSMel Gorman */ 81309135cc5SDan Williams __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 8143340289dSMel Gorman { 8153340289dSMel Gorman return vma_kernel_pagesize(vma); 8163340289dSMel Gorman } 8173340289dSMel Gorman 8183340289dSMel Gorman /* 81984afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 82084afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 82184afd99bSAndy Whitcroft * alignment. 82284afd99bSAndy Whitcroft */ 82384afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 82484afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 82504f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 82684afd99bSAndy Whitcroft 827a1e78772SMel Gorman /* 828a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 829a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 830a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 831a1e78772SMel Gorman * 832a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 833a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 834a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 835a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 83684afd99bSAndy Whitcroft * 83784afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 83884afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 83984afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 84084afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 84184afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 84284afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 84384afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 84484afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 845a1e78772SMel Gorman */ 846e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 847e7c4b0bfSAndy Whitcroft { 848e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 849e7c4b0bfSAndy Whitcroft } 850e7c4b0bfSAndy Whitcroft 851e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 852e7c4b0bfSAndy Whitcroft unsigned long value) 853e7c4b0bfSAndy Whitcroft { 854e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 855e7c4b0bfSAndy Whitcroft } 856e7c4b0bfSAndy Whitcroft 857e9fe92aeSMina Almasry static void 858e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, 859e9fe92aeSMina Almasry struct hugetlb_cgroup *h_cg, 860e9fe92aeSMina Almasry struct hstate *h) 861e9fe92aeSMina Almasry { 862e9fe92aeSMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 863e9fe92aeSMina Almasry if (!h_cg || !h) { 864e9fe92aeSMina Almasry resv_map->reservation_counter = NULL; 865e9fe92aeSMina Almasry resv_map->pages_per_hpage = 0; 866e9fe92aeSMina Almasry resv_map->css = NULL; 867e9fe92aeSMina Almasry } else { 868e9fe92aeSMina Almasry resv_map->reservation_counter = 869e9fe92aeSMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 870e9fe92aeSMina Almasry resv_map->pages_per_hpage = pages_per_huge_page(h); 871e9fe92aeSMina Almasry resv_map->css = &h_cg->css; 872e9fe92aeSMina Almasry } 873e9fe92aeSMina Almasry #endif 874e9fe92aeSMina Almasry } 875e9fe92aeSMina Almasry 8769119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 87784afd99bSAndy Whitcroft { 87884afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 8795e911373SMike Kravetz struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 8805e911373SMike Kravetz 8815e911373SMike Kravetz if (!resv_map || !rg) { 8825e911373SMike Kravetz kfree(resv_map); 8835e911373SMike Kravetz kfree(rg); 88484afd99bSAndy Whitcroft return NULL; 8855e911373SMike Kravetz } 88684afd99bSAndy Whitcroft 88784afd99bSAndy Whitcroft kref_init(&resv_map->refs); 8887b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 88984afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 89084afd99bSAndy Whitcroft 8915e911373SMike Kravetz resv_map->adds_in_progress = 0; 892e9fe92aeSMina Almasry /* 893e9fe92aeSMina Almasry * Initialize these to 0. On shared mappings, 0's here indicate these 894e9fe92aeSMina Almasry * fields don't do cgroup accounting. On private mappings, these will be 895e9fe92aeSMina Almasry * re-initialized to the proper values, to indicate that hugetlb cgroup 896e9fe92aeSMina Almasry * reservations are to be un-charged from here. 897e9fe92aeSMina Almasry */ 898e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); 8995e911373SMike Kravetz 9005e911373SMike Kravetz INIT_LIST_HEAD(&resv_map->region_cache); 9015e911373SMike Kravetz list_add(&rg->link, &resv_map->region_cache); 9025e911373SMike Kravetz resv_map->region_cache_count = 1; 9035e911373SMike Kravetz 90484afd99bSAndy Whitcroft return resv_map; 90584afd99bSAndy Whitcroft } 90684afd99bSAndy Whitcroft 9079119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 90884afd99bSAndy Whitcroft { 90984afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 9105e911373SMike Kravetz struct list_head *head = &resv_map->region_cache; 9115e911373SMike Kravetz struct file_region *rg, *trg; 91284afd99bSAndy Whitcroft 91384afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 914feba16e2SMike Kravetz region_del(resv_map, 0, LONG_MAX); 9155e911373SMike Kravetz 9165e911373SMike Kravetz /* ... and any entries left in the cache */ 9175e911373SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 9185e911373SMike Kravetz list_del(&rg->link); 9195e911373SMike Kravetz kfree(rg); 9205e911373SMike Kravetz } 9215e911373SMike Kravetz 9225e911373SMike Kravetz VM_BUG_ON(resv_map->adds_in_progress); 9235e911373SMike Kravetz 92484afd99bSAndy Whitcroft kfree(resv_map); 92584afd99bSAndy Whitcroft } 92684afd99bSAndy Whitcroft 9274e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 9284e35f483SJoonsoo Kim { 929f27a5136SMike Kravetz /* 930f27a5136SMike Kravetz * At inode evict time, i_mapping may not point to the original 931f27a5136SMike Kravetz * address space within the inode. This original address space 932f27a5136SMike Kravetz * contains the pointer to the resv_map. So, always use the 933f27a5136SMike Kravetz * address space embedded within the inode. 934f27a5136SMike Kravetz * The VERY common case is inode->mapping == &inode->i_data but, 935f27a5136SMike Kravetz * this may not be true for device special inodes. 936f27a5136SMike Kravetz */ 937f27a5136SMike Kravetz return (struct resv_map *)(&inode->i_data)->private_data; 9384e35f483SJoonsoo Kim } 9394e35f483SJoonsoo Kim 94084afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 941a1e78772SMel Gorman { 94281d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 9434e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 9444e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 9454e35f483SJoonsoo Kim struct inode *inode = mapping->host; 9464e35f483SJoonsoo Kim 9474e35f483SJoonsoo Kim return inode_resv_map(inode); 9484e35f483SJoonsoo Kim 9494e35f483SJoonsoo Kim } else { 95084afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 95184afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 9524e35f483SJoonsoo Kim } 953a1e78772SMel Gorman } 954a1e78772SMel Gorman 95584afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 956a1e78772SMel Gorman { 95781d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 95881d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 959a1e78772SMel Gorman 96084afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 96184afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 96204f2cbe3SMel Gorman } 96304f2cbe3SMel Gorman 96404f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 96504f2cbe3SMel Gorman { 96681d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 96781d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 968e7c4b0bfSAndy Whitcroft 969e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 97004f2cbe3SMel Gorman } 97104f2cbe3SMel Gorman 97204f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 97304f2cbe3SMel Gorman { 97481d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 975e7c4b0bfSAndy Whitcroft 976e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 977a1e78772SMel Gorman } 978a1e78772SMel Gorman 97904f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 980a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 981a1e78772SMel Gorman { 98281d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 983f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 984a1e78772SMel Gorman vma->vm_private_data = (void *)0; 985a1e78772SMel Gorman } 986a1e78772SMel Gorman 987a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 988559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 989a1e78772SMel Gorman { 990af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 991af0ed73eSJoonsoo Kim /* 992af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 993af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 994af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 995af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 996af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 997af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 998af0ed73eSJoonsoo Kim * properly, so add work-around here. 999af0ed73eSJoonsoo Kim */ 1000af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 1001559ec2f8SNicholas Krause return true; 1002af0ed73eSJoonsoo Kim else 1003559ec2f8SNicholas Krause return false; 1004af0ed73eSJoonsoo Kim } 1005a63884e9SJoonsoo Kim 1006a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 10071fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 10081fb1b0e9SMike Kravetz /* 10091fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 10101fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 10111fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 10127c8de358SEthon Paul * fallocate. In this case, there really are no reserves to 10131fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 10141fb1b0e9SMike Kravetz */ 10151fb1b0e9SMike Kravetz if (chg) 10161fb1b0e9SMike Kravetz return false; 10171fb1b0e9SMike Kravetz else 1018559ec2f8SNicholas Krause return true; 10191fb1b0e9SMike Kravetz } 1020a63884e9SJoonsoo Kim 1021a63884e9SJoonsoo Kim /* 1022a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 1023a63884e9SJoonsoo Kim * private mappings. 1024a63884e9SJoonsoo Kim */ 102567961f9dSMike Kravetz if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 102667961f9dSMike Kravetz /* 102767961f9dSMike Kravetz * Like the shared case above, a hole punch or truncate 102867961f9dSMike Kravetz * could have been performed on the private mapping. 102967961f9dSMike Kravetz * Examine the value of chg to determine if reserves 103067961f9dSMike Kravetz * actually exist or were previously consumed. 103167961f9dSMike Kravetz * Very Subtle - The value of chg comes from a previous 103267961f9dSMike Kravetz * call to vma_needs_reserves(). The reserve map for 103367961f9dSMike Kravetz * private mappings has different (opposite) semantics 103467961f9dSMike Kravetz * than that of shared mappings. vma_needs_reserves() 103567961f9dSMike Kravetz * has already taken this difference in semantics into 103667961f9dSMike Kravetz * account. Therefore, the meaning of chg is the same 103767961f9dSMike Kravetz * as in the shared case above. Code could easily be 103867961f9dSMike Kravetz * combined, but keeping it separate draws attention to 103967961f9dSMike Kravetz * subtle differences. 104067961f9dSMike Kravetz */ 104167961f9dSMike Kravetz if (chg) 104267961f9dSMike Kravetz return false; 104367961f9dSMike Kravetz else 1044559ec2f8SNicholas Krause return true; 104567961f9dSMike Kravetz } 1046a63884e9SJoonsoo Kim 1047559ec2f8SNicholas Krause return false; 1048a1e78772SMel Gorman } 1049a1e78772SMel Gorman 1050a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 10511da177e4SLinus Torvalds { 10521da177e4SLinus Torvalds int nid = page_to_nid(page); 10530edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 1054a5516438SAndi Kleen h->free_huge_pages++; 1055a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 10567ffddd49SMuchun Song SetPageHugeFreed(page); 10571da177e4SLinus Torvalds } 10581da177e4SLinus Torvalds 105994310cbcSAnshuman Khandual static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) 1060bf50bab2SNaoya Horiguchi { 1061bf50bab2SNaoya Horiguchi struct page *page; 1062bbe88753SJoonsoo Kim bool nocma = !!(current->flags & PF_MEMALLOC_NOCMA); 1063bf50bab2SNaoya Horiguchi 1064bbe88753SJoonsoo Kim list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { 1065bbe88753SJoonsoo Kim if (nocma && is_migrate_cma_page(page)) 1066bbe88753SJoonsoo Kim continue; 1067bbe88753SJoonsoo Kim 10686664bfc8SWei Yang if (PageHWPoison(page)) 10696664bfc8SWei Yang continue; 1070bbe88753SJoonsoo Kim 10710edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 1072a9869b83SNaoya Horiguchi set_page_refcounted(page); 10737ffddd49SMuchun Song ClearPageHugeFreed(page); 1074bf50bab2SNaoya Horiguchi h->free_huge_pages--; 1075bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 1076bf50bab2SNaoya Horiguchi return page; 1077bf50bab2SNaoya Horiguchi } 1078bf50bab2SNaoya Horiguchi 10796664bfc8SWei Yang return NULL; 10806664bfc8SWei Yang } 10816664bfc8SWei Yang 10823e59fcb0SMichal Hocko static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, 10833e59fcb0SMichal Hocko nodemask_t *nmask) 108494310cbcSAnshuman Khandual { 10853e59fcb0SMichal Hocko unsigned int cpuset_mems_cookie; 10863e59fcb0SMichal Hocko struct zonelist *zonelist; 10873e59fcb0SMichal Hocko struct zone *zone; 10883e59fcb0SMichal Hocko struct zoneref *z; 108998fa15f3SAnshuman Khandual int node = NUMA_NO_NODE; 10903e59fcb0SMichal Hocko 10913e59fcb0SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 10923e59fcb0SMichal Hocko 10933e59fcb0SMichal Hocko retry_cpuset: 10943e59fcb0SMichal Hocko cpuset_mems_cookie = read_mems_allowed_begin(); 10953e59fcb0SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 109694310cbcSAnshuman Khandual struct page *page; 109794310cbcSAnshuman Khandual 10983e59fcb0SMichal Hocko if (!cpuset_zone_allowed(zone, gfp_mask)) 10993e59fcb0SMichal Hocko continue; 11003e59fcb0SMichal Hocko /* 11013e59fcb0SMichal Hocko * no need to ask again on the same node. Pool is node rather than 11023e59fcb0SMichal Hocko * zone aware 11033e59fcb0SMichal Hocko */ 11043e59fcb0SMichal Hocko if (zone_to_nid(zone) == node) 11053e59fcb0SMichal Hocko continue; 11063e59fcb0SMichal Hocko node = zone_to_nid(zone); 110794310cbcSAnshuman Khandual 110894310cbcSAnshuman Khandual page = dequeue_huge_page_node_exact(h, node); 110994310cbcSAnshuman Khandual if (page) 111094310cbcSAnshuman Khandual return page; 111194310cbcSAnshuman Khandual } 11123e59fcb0SMichal Hocko if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 11133e59fcb0SMichal Hocko goto retry_cpuset; 11143e59fcb0SMichal Hocko 111594310cbcSAnshuman Khandual return NULL; 111694310cbcSAnshuman Khandual } 111794310cbcSAnshuman Khandual 1118a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 1119a5516438SAndi Kleen struct vm_area_struct *vma, 1120af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 1121af0ed73eSJoonsoo Kim long chg) 11221da177e4SLinus Torvalds { 11233e59fcb0SMichal Hocko struct page *page; 1124480eccf9SLee Schermerhorn struct mempolicy *mpol; 112504ec6264SVlastimil Babka gfp_t gfp_mask; 11263e59fcb0SMichal Hocko nodemask_t *nodemask; 112704ec6264SVlastimil Babka int nid; 11281da177e4SLinus Torvalds 1129a1e78772SMel Gorman /* 1130a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 1131a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 1132a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 1133a1e78772SMel Gorman */ 1134af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg) && 1135a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 1136c0ff7453SMiao Xie goto err; 1137a1e78772SMel Gorman 113804f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 1139a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 11406eab04a8SJustin P. Mattock goto err; 114104f2cbe3SMel Gorman 114204ec6264SVlastimil Babka gfp_mask = htlb_alloc_mask(h); 114304ec6264SVlastimil Babka nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 11443e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 11453e59fcb0SMichal Hocko if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { 1146d6995da3SMike Kravetz SetHPageRestoreReserve(page); 1147a63884e9SJoonsoo Kim h->resv_huge_pages--; 1148bf50bab2SNaoya Horiguchi } 1149cc9a6c87SMel Gorman 1150cc9a6c87SMel Gorman mpol_cond_put(mpol); 1151cc9a6c87SMel Gorman return page; 1152cc9a6c87SMel Gorman 1153c0ff7453SMiao Xie err: 1154cc9a6c87SMel Gorman return NULL; 11551da177e4SLinus Torvalds } 11561da177e4SLinus Torvalds 11571cac6f2cSLuiz Capitulino /* 11581cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 11591cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 11601cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 11611cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 11621cac6f2cSLuiz Capitulino * node for alloc or free. 11631cac6f2cSLuiz Capitulino */ 11641cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 11651cac6f2cSLuiz Capitulino { 11660edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 11671cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 11681cac6f2cSLuiz Capitulino 11691cac6f2cSLuiz Capitulino return nid; 11701cac6f2cSLuiz Capitulino } 11711cac6f2cSLuiz Capitulino 11721cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 11731cac6f2cSLuiz Capitulino { 11741cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 11751cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 11761cac6f2cSLuiz Capitulino return nid; 11771cac6f2cSLuiz Capitulino } 11781cac6f2cSLuiz Capitulino 11791cac6f2cSLuiz Capitulino /* 11801cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 11811cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 11821cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 11831cac6f2cSLuiz Capitulino * mask. 11841cac6f2cSLuiz Capitulino */ 11851cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 11861cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 11871cac6f2cSLuiz Capitulino { 11881cac6f2cSLuiz Capitulino int nid; 11891cac6f2cSLuiz Capitulino 11901cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 11911cac6f2cSLuiz Capitulino 11921cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 11931cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 11941cac6f2cSLuiz Capitulino 11951cac6f2cSLuiz Capitulino return nid; 11961cac6f2cSLuiz Capitulino } 11971cac6f2cSLuiz Capitulino 11981cac6f2cSLuiz Capitulino /* 11991cac6f2cSLuiz Capitulino * helper for free_pool_huge_page() - return the previously saved 12001cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 12011cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 12021cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 12031cac6f2cSLuiz Capitulino */ 12041cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 12051cac6f2cSLuiz Capitulino { 12061cac6f2cSLuiz Capitulino int nid; 12071cac6f2cSLuiz Capitulino 12081cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 12091cac6f2cSLuiz Capitulino 12101cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 12111cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 12121cac6f2cSLuiz Capitulino 12131cac6f2cSLuiz Capitulino return nid; 12141cac6f2cSLuiz Capitulino } 12151cac6f2cSLuiz Capitulino 12161cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 12171cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 12181cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 12191cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 12201cac6f2cSLuiz Capitulino nr_nodes--) 12211cac6f2cSLuiz Capitulino 12221cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 12231cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 12241cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 12251cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 12261cac6f2cSLuiz Capitulino nr_nodes--) 12271cac6f2cSLuiz Capitulino 1228e1073d1eSAneesh Kumar K.V #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 1229944d9fecSLuiz Capitulino static void destroy_compound_gigantic_page(struct page *page, 1230d00181b9SKirill A. Shutemov unsigned int order) 1231944d9fecSLuiz Capitulino { 1232944d9fecSLuiz Capitulino int i; 1233944d9fecSLuiz Capitulino int nr_pages = 1 << order; 1234944d9fecSLuiz Capitulino struct page *p = page + 1; 1235944d9fecSLuiz Capitulino 1236c8cc708aSGerald Schaefer atomic_set(compound_mapcount_ptr(page), 0); 123747e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 123847e29d32SJohn Hubbard 1239944d9fecSLuiz Capitulino for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 12401d798ca3SKirill A. Shutemov clear_compound_head(p); 1241944d9fecSLuiz Capitulino set_page_refcounted(p); 1242944d9fecSLuiz Capitulino } 1243944d9fecSLuiz Capitulino 1244944d9fecSLuiz Capitulino set_compound_order(page, 0); 1245ba9c1201SGerald Schaefer page[1].compound_nr = 0; 1246944d9fecSLuiz Capitulino __ClearPageHead(page); 1247944d9fecSLuiz Capitulino } 1248944d9fecSLuiz Capitulino 1249d00181b9SKirill A. Shutemov static void free_gigantic_page(struct page *page, unsigned int order) 1250944d9fecSLuiz Capitulino { 1251cf11e85fSRoman Gushchin /* 1252cf11e85fSRoman Gushchin * If the page isn't allocated using the cma allocator, 1253cf11e85fSRoman Gushchin * cma_release() returns false. 1254cf11e85fSRoman Gushchin */ 1255dbda8feaSBarry Song #ifdef CONFIG_CMA 1256dbda8feaSBarry Song if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order)) 1257cf11e85fSRoman Gushchin return; 1258dbda8feaSBarry Song #endif 1259cf11e85fSRoman Gushchin 1260944d9fecSLuiz Capitulino free_contig_range(page_to_pfn(page), 1 << order); 1261944d9fecSLuiz Capitulino } 1262944d9fecSLuiz Capitulino 12634eb0716eSAlexandre Ghiti #ifdef CONFIG_CONTIG_ALLOC 1264d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 1265d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) 1266944d9fecSLuiz Capitulino { 12675e27a2dfSAnshuman Khandual unsigned long nr_pages = 1UL << huge_page_order(h); 1268953f064aSLi Xinhai if (nid == NUMA_NO_NODE) 1269953f064aSLi Xinhai nid = numa_mem_id(); 1270944d9fecSLuiz Capitulino 1271dbda8feaSBarry Song #ifdef CONFIG_CMA 1272dbda8feaSBarry Song { 1273cf11e85fSRoman Gushchin struct page *page; 1274cf11e85fSRoman Gushchin int node; 1275cf11e85fSRoman Gushchin 1276953f064aSLi Xinhai if (hugetlb_cma[nid]) { 1277953f064aSLi Xinhai page = cma_alloc(hugetlb_cma[nid], nr_pages, 1278953f064aSLi Xinhai huge_page_order(h), true); 1279953f064aSLi Xinhai if (page) 1280953f064aSLi Xinhai return page; 1281953f064aSLi Xinhai } 1282953f064aSLi Xinhai 1283953f064aSLi Xinhai if (!(gfp_mask & __GFP_THISNODE)) { 1284cf11e85fSRoman Gushchin for_each_node_mask(node, *nodemask) { 1285953f064aSLi Xinhai if (node == nid || !hugetlb_cma[node]) 1286cf11e85fSRoman Gushchin continue; 1287cf11e85fSRoman Gushchin 1288cf11e85fSRoman Gushchin page = cma_alloc(hugetlb_cma[node], nr_pages, 1289cf11e85fSRoman Gushchin huge_page_order(h), true); 1290cf11e85fSRoman Gushchin if (page) 1291cf11e85fSRoman Gushchin return page; 1292cf11e85fSRoman Gushchin } 1293cf11e85fSRoman Gushchin } 1294953f064aSLi Xinhai } 1295dbda8feaSBarry Song #endif 1296cf11e85fSRoman Gushchin 12975e27a2dfSAnshuman Khandual return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); 1298944d9fecSLuiz Capitulino } 1299944d9fecSLuiz Capitulino 1300944d9fecSLuiz Capitulino static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); 1301d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order); 13024eb0716eSAlexandre Ghiti #else /* !CONFIG_CONTIG_ALLOC */ 13034eb0716eSAlexandre Ghiti static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 13044eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 13054eb0716eSAlexandre Ghiti { 13064eb0716eSAlexandre Ghiti return NULL; 13074eb0716eSAlexandre Ghiti } 13084eb0716eSAlexandre Ghiti #endif /* CONFIG_CONTIG_ALLOC */ 1309944d9fecSLuiz Capitulino 1310e1073d1eSAneesh Kumar K.V #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 1311d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 13124eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 13134eb0716eSAlexandre Ghiti { 13144eb0716eSAlexandre Ghiti return NULL; 13154eb0716eSAlexandre Ghiti } 1316d00181b9SKirill A. Shutemov static inline void free_gigantic_page(struct page *page, unsigned int order) { } 1317944d9fecSLuiz Capitulino static inline void destroy_compound_gigantic_page(struct page *page, 1318d00181b9SKirill A. Shutemov unsigned int order) { } 1319944d9fecSLuiz Capitulino #endif 1320944d9fecSLuiz Capitulino 1321a5516438SAndi Kleen static void update_and_free_page(struct hstate *h, struct page *page) 13226af2acb6SAdam Litke { 13236af2acb6SAdam Litke int i; 1324dbfee5aeSMike Kravetz struct page *subpage = page; 1325a5516438SAndi Kleen 13264eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1327944d9fecSLuiz Capitulino return; 132818229df5SAndy Whitcroft 1329a5516438SAndi Kleen h->nr_huge_pages--; 1330a5516438SAndi Kleen h->nr_huge_pages_node[page_to_nid(page)]--; 1331dbfee5aeSMike Kravetz for (i = 0; i < pages_per_huge_page(h); 1332dbfee5aeSMike Kravetz i++, subpage = mem_map_next(subpage, page, i)) { 1333dbfee5aeSMike Kravetz subpage->flags &= ~(1 << PG_locked | 1 << PG_error | 133432f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1335a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1336a7407a27SLuiz Capitulino 1 << PG_writeback); 13376af2acb6SAdam Litke } 1338309381feSSasha Levin VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 13391adc4d41SMina Almasry VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page); 1340f1e61557SKirill A. Shutemov set_compound_page_dtor(page, NULL_COMPOUND_DTOR); 13416af2acb6SAdam Litke set_page_refcounted(page); 1342944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) { 1343cf11e85fSRoman Gushchin /* 1344cf11e85fSRoman Gushchin * Temporarily drop the hugetlb_lock, because 1345cf11e85fSRoman Gushchin * we might block in free_gigantic_page(). 1346cf11e85fSRoman Gushchin */ 1347cf11e85fSRoman Gushchin spin_unlock(&hugetlb_lock); 1348944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 1349944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 1350cf11e85fSRoman Gushchin spin_lock(&hugetlb_lock); 1351944d9fecSLuiz Capitulino } else { 1352a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 13536af2acb6SAdam Litke } 1354944d9fecSLuiz Capitulino } 13556af2acb6SAdam Litke 1356e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1357e5ff2159SAndi Kleen { 1358e5ff2159SAndi Kleen struct hstate *h; 1359e5ff2159SAndi Kleen 1360e5ff2159SAndi Kleen for_each_hstate(h) { 1361e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1362e5ff2159SAndi Kleen return h; 1363e5ff2159SAndi Kleen } 1364e5ff2159SAndi Kleen return NULL; 1365e5ff2159SAndi Kleen } 1366e5ff2159SAndi Kleen 1367bcc54222SNaoya Horiguchi /* 1368ab5ac90aSMichal Hocko * Internal hugetlb specific page flag. Do not use outside of the hugetlb 1369ab5ac90aSMichal Hocko * code 1370ab5ac90aSMichal Hocko */ 1371ab5ac90aSMichal Hocko static inline bool PageHugeTemporary(struct page *page) 1372ab5ac90aSMichal Hocko { 1373ab5ac90aSMichal Hocko if (!PageHuge(page)) 1374ab5ac90aSMichal Hocko return false; 1375ab5ac90aSMichal Hocko 1376ab5ac90aSMichal Hocko return (unsigned long)page[2].mapping == -1U; 1377ab5ac90aSMichal Hocko } 1378ab5ac90aSMichal Hocko 1379ab5ac90aSMichal Hocko static inline void SetPageHugeTemporary(struct page *page) 1380ab5ac90aSMichal Hocko { 1381ab5ac90aSMichal Hocko page[2].mapping = (void *)-1U; 1382ab5ac90aSMichal Hocko } 1383ab5ac90aSMichal Hocko 1384ab5ac90aSMichal Hocko static inline void ClearPageHugeTemporary(struct page *page) 1385ab5ac90aSMichal Hocko { 1386ab5ac90aSMichal Hocko page[2].mapping = NULL; 1387ab5ac90aSMichal Hocko } 1388ab5ac90aSMichal Hocko 1389c77c0a8aSWaiman Long static void __free_huge_page(struct page *page) 139027a85ef1SDavid Gibson { 1391a5516438SAndi Kleen /* 1392a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1393a5516438SAndi Kleen * compound page destructor. 1394a5516438SAndi Kleen */ 1395e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 13967893d1d5SAdam Litke int nid = page_to_nid(page); 1397d6995da3SMike Kravetz struct hugepage_subpool *spool = hugetlb_page_subpool(page); 139807443a85SJoonsoo Kim bool restore_reserve; 139927a85ef1SDavid Gibson 1400b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1401b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_mapcount(page), page); 14028ace22bcSYongkai Wu 1403d6995da3SMike Kravetz hugetlb_set_page_subpool(page, NULL); 14048ace22bcSYongkai Wu page->mapping = NULL; 1405d6995da3SMike Kravetz restore_reserve = HPageRestoreReserve(page); 1406d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 140727a85ef1SDavid Gibson 14081c5ecae3SMike Kravetz /* 1409d6995da3SMike Kravetz * If HPageRestoreReserve was set on page, page allocation consumed a 14100919e1b6SMike Kravetz * reservation. If the page was associated with a subpool, there 14110919e1b6SMike Kravetz * would have been a page reserved in the subpool before allocation 14120919e1b6SMike Kravetz * via hugepage_subpool_get_pages(). Since we are 'restoring' the 14136c26d310SMiaohe Lin * reservation, do not call hugepage_subpool_put_pages() as this will 14140919e1b6SMike Kravetz * remove the reserved page from the subpool. 14150919e1b6SMike Kravetz */ 14160919e1b6SMike Kravetz if (!restore_reserve) { 14170919e1b6SMike Kravetz /* 14180919e1b6SMike Kravetz * A return code of zero implies that the subpool will be 14190919e1b6SMike Kravetz * under its minimum size if the reservation is not restored 14200919e1b6SMike Kravetz * after page is free. Therefore, force restore_reserve 14210919e1b6SMike Kravetz * operation. 14221c5ecae3SMike Kravetz */ 14231c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 14241c5ecae3SMike Kravetz restore_reserve = true; 14250919e1b6SMike Kravetz } 14261c5ecae3SMike Kravetz 142727a85ef1SDavid Gibson spin_lock(&hugetlb_lock); 1428*8f251a3dSMike Kravetz ClearHPageMigratable(page); 14296d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 14306d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 143108cf9fafSMina Almasry hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 143208cf9fafSMina Almasry pages_per_huge_page(h), page); 143307443a85SJoonsoo Kim if (restore_reserve) 143407443a85SJoonsoo Kim h->resv_huge_pages++; 143507443a85SJoonsoo Kim 1436ab5ac90aSMichal Hocko if (PageHugeTemporary(page)) { 1437ab5ac90aSMichal Hocko list_del(&page->lru); 1438ab5ac90aSMichal Hocko ClearPageHugeTemporary(page); 1439ab5ac90aSMichal Hocko update_and_free_page(h, page); 1440ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 14410edaecfaSAneesh Kumar K.V /* remove the page from active list */ 14420edaecfaSAneesh Kumar K.V list_del(&page->lru); 1443a5516438SAndi Kleen update_and_free_page(h, page); 1444a5516438SAndi Kleen h->surplus_huge_pages--; 1445a5516438SAndi Kleen h->surplus_huge_pages_node[nid]--; 14467893d1d5SAdam Litke } else { 14475d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1448a5516438SAndi Kleen enqueue_huge_page(h, page); 14497893d1d5SAdam Litke } 145027a85ef1SDavid Gibson spin_unlock(&hugetlb_lock); 145127a85ef1SDavid Gibson } 145227a85ef1SDavid Gibson 1453c77c0a8aSWaiman Long /* 1454c77c0a8aSWaiman Long * As free_huge_page() can be called from a non-task context, we have 1455c77c0a8aSWaiman Long * to defer the actual freeing in a workqueue to prevent potential 1456c77c0a8aSWaiman Long * hugetlb_lock deadlock. 1457c77c0a8aSWaiman Long * 1458c77c0a8aSWaiman Long * free_hpage_workfn() locklessly retrieves the linked list of pages to 1459c77c0a8aSWaiman Long * be freed and frees them one-by-one. As the page->mapping pointer is 1460c77c0a8aSWaiman Long * going to be cleared in __free_huge_page() anyway, it is reused as the 1461c77c0a8aSWaiman Long * llist_node structure of a lockless linked list of huge pages to be freed. 1462c77c0a8aSWaiman Long */ 1463c77c0a8aSWaiman Long static LLIST_HEAD(hpage_freelist); 1464c77c0a8aSWaiman Long 1465c77c0a8aSWaiman Long static void free_hpage_workfn(struct work_struct *work) 1466c77c0a8aSWaiman Long { 1467c77c0a8aSWaiman Long struct llist_node *node; 1468c77c0a8aSWaiman Long struct page *page; 1469c77c0a8aSWaiman Long 1470c77c0a8aSWaiman Long node = llist_del_all(&hpage_freelist); 1471c77c0a8aSWaiman Long 1472c77c0a8aSWaiman Long while (node) { 1473c77c0a8aSWaiman Long page = container_of((struct address_space **)node, 1474c77c0a8aSWaiman Long struct page, mapping); 1475c77c0a8aSWaiman Long node = node->next; 1476c77c0a8aSWaiman Long __free_huge_page(page); 1477c77c0a8aSWaiman Long } 1478c77c0a8aSWaiman Long } 1479c77c0a8aSWaiman Long static DECLARE_WORK(free_hpage_work, free_hpage_workfn); 1480c77c0a8aSWaiman Long 1481c77c0a8aSWaiman Long void free_huge_page(struct page *page) 1482c77c0a8aSWaiman Long { 1483c77c0a8aSWaiman Long /* 1484c77c0a8aSWaiman Long * Defer freeing if in non-task context to avoid hugetlb_lock deadlock. 1485c77c0a8aSWaiman Long */ 1486c77c0a8aSWaiman Long if (!in_task()) { 1487c77c0a8aSWaiman Long /* 1488c77c0a8aSWaiman Long * Only call schedule_work() if hpage_freelist is previously 1489c77c0a8aSWaiman Long * empty. Otherwise, schedule_work() had been called but the 1490c77c0a8aSWaiman Long * workfn hasn't retrieved the list yet. 1491c77c0a8aSWaiman Long */ 1492c77c0a8aSWaiman Long if (llist_add((struct llist_node *)&page->mapping, 1493c77c0a8aSWaiman Long &hpage_freelist)) 1494c77c0a8aSWaiman Long schedule_work(&free_hpage_work); 1495c77c0a8aSWaiman Long return; 1496c77c0a8aSWaiman Long } 1497c77c0a8aSWaiman Long 1498c77c0a8aSWaiman Long __free_huge_page(page); 1499c77c0a8aSWaiman Long } 1500c77c0a8aSWaiman Long 1501a5516438SAndi Kleen static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1502b7ba30c6SAndi Kleen { 15030edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1504f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 15059dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 15061adc4d41SMina Almasry set_hugetlb_cgroup_rsvd(page, NULL); 15072f37511cSWei Yang spin_lock(&hugetlb_lock); 1508a5516438SAndi Kleen h->nr_huge_pages++; 1509a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 15107ffddd49SMuchun Song ClearPageHugeFreed(page); 1511b7ba30c6SAndi Kleen spin_unlock(&hugetlb_lock); 1512b7ba30c6SAndi Kleen } 1513b7ba30c6SAndi Kleen 1514d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order) 151520a0307cSWu Fengguang { 151620a0307cSWu Fengguang int i; 151720a0307cSWu Fengguang int nr_pages = 1 << order; 151820a0307cSWu Fengguang struct page *p = page + 1; 151920a0307cSWu Fengguang 152020a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 152120a0307cSWu Fengguang set_compound_order(page, order); 1522ef5a22beSAndrea Arcangeli __ClearPageReserved(page); 1523de09d31dSKirill A. Shutemov __SetPageHead(page); 152420a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1525ef5a22beSAndrea Arcangeli /* 1526ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1527ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1528ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 15297c8de358SEthon Paul * too. Otherwise drivers using get_user_pages() to access tail 1530ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1531ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1532ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1533ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1534ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1535ef5a22beSAndrea Arcangeli * after get_user_pages(). 1536ef5a22beSAndrea Arcangeli */ 1537ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 153858a84aa9SYouquan Song set_page_count(p, 0); 15391d798ca3SKirill A. Shutemov set_compound_head(p, page); 154020a0307cSWu Fengguang } 1541b4330afbSMike Kravetz atomic_set(compound_mapcount_ptr(page), -1); 154247e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 154320a0307cSWu Fengguang } 154420a0307cSWu Fengguang 15457795912cSAndrew Morton /* 15467795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 15477795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 15487795912cSAndrew Morton * details. 15497795912cSAndrew Morton */ 155020a0307cSWu Fengguang int PageHuge(struct page *page) 155120a0307cSWu Fengguang { 155220a0307cSWu Fengguang if (!PageCompound(page)) 155320a0307cSWu Fengguang return 0; 155420a0307cSWu Fengguang 155520a0307cSWu Fengguang page = compound_head(page); 1556f1e61557SKirill A. Shutemov return page[1].compound_dtor == HUGETLB_PAGE_DTOR; 155720a0307cSWu Fengguang } 155843131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 155943131e14SNaoya Horiguchi 156027c73ae7SAndrea Arcangeli /* 156127c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 156227c73ae7SAndrea Arcangeli * normal or transparent huge pages. 156327c73ae7SAndrea Arcangeli */ 156427c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 156527c73ae7SAndrea Arcangeli { 156627c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 156727c73ae7SAndrea Arcangeli return 0; 156827c73ae7SAndrea Arcangeli 1569d4af73e3SVlastimil Babka return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; 157027c73ae7SAndrea Arcangeli } 157127c73ae7SAndrea Arcangeli 1572c0d0381aSMike Kravetz /* 1573c0d0381aSMike Kravetz * Find and lock address space (mapping) in write mode. 1574c0d0381aSMike Kravetz * 1575336bf30eSMike Kravetz * Upon entry, the page is locked which means that page_mapping() is 1576336bf30eSMike Kravetz * stable. Due to locking order, we can only trylock_write. If we can 1577336bf30eSMike Kravetz * not get the lock, simply return NULL to caller. 1578c0d0381aSMike Kravetz */ 1579c0d0381aSMike Kravetz struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) 1580c0d0381aSMike Kravetz { 1581336bf30eSMike Kravetz struct address_space *mapping = page_mapping(hpage); 1582c0d0381aSMike Kravetz 1583c0d0381aSMike Kravetz if (!mapping) 1584c0d0381aSMike Kravetz return mapping; 1585c0d0381aSMike Kravetz 1586c0d0381aSMike Kravetz if (i_mmap_trylock_write(mapping)) 1587c0d0381aSMike Kravetz return mapping; 1588c0d0381aSMike Kravetz 1589c0d0381aSMike Kravetz return NULL; 1590c0d0381aSMike Kravetz } 1591c0d0381aSMike Kravetz 159213d60f4bSZhang Yi pgoff_t __basepage_index(struct page *page) 159313d60f4bSZhang Yi { 159413d60f4bSZhang Yi struct page *page_head = compound_head(page); 159513d60f4bSZhang Yi pgoff_t index = page_index(page_head); 159613d60f4bSZhang Yi unsigned long compound_idx; 159713d60f4bSZhang Yi 159813d60f4bSZhang Yi if (!PageHuge(page_head)) 159913d60f4bSZhang Yi return page_index(page); 160013d60f4bSZhang Yi 160113d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 160213d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 160313d60f4bSZhang Yi else 160413d60f4bSZhang Yi compound_idx = page - page_head; 160513d60f4bSZhang Yi 160613d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 160713d60f4bSZhang Yi } 160813d60f4bSZhang Yi 16090c397daeSMichal Hocko static struct page *alloc_buddy_huge_page(struct hstate *h, 1610f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1611f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 16121da177e4SLinus Torvalds { 1613af0fb9dfSMichal Hocko int order = huge_page_order(h); 16141da177e4SLinus Torvalds struct page *page; 1615f60858f9SMike Kravetz bool alloc_try_hard = true; 1616f96efd58SJoe Jin 1617f60858f9SMike Kravetz /* 1618f60858f9SMike Kravetz * By default we always try hard to allocate the page with 1619f60858f9SMike Kravetz * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in 1620f60858f9SMike Kravetz * a loop (to adjust global huge page counts) and previous allocation 1621f60858f9SMike Kravetz * failed, do not continue to try hard on the same node. Use the 1622f60858f9SMike Kravetz * node_alloc_noretry bitmap to manage this state information. 1623f60858f9SMike Kravetz */ 1624f60858f9SMike Kravetz if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) 1625f60858f9SMike Kravetz alloc_try_hard = false; 1626f60858f9SMike Kravetz gfp_mask |= __GFP_COMP|__GFP_NOWARN; 1627f60858f9SMike Kravetz if (alloc_try_hard) 1628f60858f9SMike Kravetz gfp_mask |= __GFP_RETRY_MAYFAIL; 1629af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 1630af0fb9dfSMichal Hocko nid = numa_mem_id(); 1631af0fb9dfSMichal Hocko page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask); 1632af0fb9dfSMichal Hocko if (page) 1633af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC); 1634af0fb9dfSMichal Hocko else 1635af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 163663b4613cSNishanth Aravamudan 1637f60858f9SMike Kravetz /* 1638f60858f9SMike Kravetz * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this 1639f60858f9SMike Kravetz * indicates an overall state change. Clear bit so that we resume 1640f60858f9SMike Kravetz * normal 'try hard' allocations. 1641f60858f9SMike Kravetz */ 1642f60858f9SMike Kravetz if (node_alloc_noretry && page && !alloc_try_hard) 1643f60858f9SMike Kravetz node_clear(nid, *node_alloc_noretry); 1644f60858f9SMike Kravetz 1645f60858f9SMike Kravetz /* 1646f60858f9SMike Kravetz * If we tried hard to get a page but failed, set bit so that 1647f60858f9SMike Kravetz * subsequent attempts will not try as hard until there is an 1648f60858f9SMike Kravetz * overall state change. 1649f60858f9SMike Kravetz */ 1650f60858f9SMike Kravetz if (node_alloc_noretry && !page && alloc_try_hard) 1651f60858f9SMike Kravetz node_set(nid, *node_alloc_noretry); 1652f60858f9SMike Kravetz 165363b4613cSNishanth Aravamudan return page; 165463b4613cSNishanth Aravamudan } 165563b4613cSNishanth Aravamudan 1656af0fb9dfSMichal Hocko /* 16570c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 16580c397daeSMichal Hocko * should use this function to get new hugetlb pages 16590c397daeSMichal Hocko */ 16600c397daeSMichal Hocko static struct page *alloc_fresh_huge_page(struct hstate *h, 1661f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1662f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 16630c397daeSMichal Hocko { 16640c397daeSMichal Hocko struct page *page; 16650c397daeSMichal Hocko 16660c397daeSMichal Hocko if (hstate_is_gigantic(h)) 16670c397daeSMichal Hocko page = alloc_gigantic_page(h, gfp_mask, nid, nmask); 16680c397daeSMichal Hocko else 16690c397daeSMichal Hocko page = alloc_buddy_huge_page(h, gfp_mask, 1670f60858f9SMike Kravetz nid, nmask, node_alloc_noretry); 16710c397daeSMichal Hocko if (!page) 16720c397daeSMichal Hocko return NULL; 16730c397daeSMichal Hocko 16740c397daeSMichal Hocko if (hstate_is_gigantic(h)) 16750c397daeSMichal Hocko prep_compound_gigantic_page(page, huge_page_order(h)); 16760c397daeSMichal Hocko prep_new_huge_page(h, page, page_to_nid(page)); 16770c397daeSMichal Hocko 16780c397daeSMichal Hocko return page; 16790c397daeSMichal Hocko } 16800c397daeSMichal Hocko 16810c397daeSMichal Hocko /* 1682af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 1683af0fb9dfSMichal Hocko * manner. 1684af0fb9dfSMichal Hocko */ 1685f60858f9SMike Kravetz static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 1686f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 1687b2261026SJoonsoo Kim { 1688b2261026SJoonsoo Kim struct page *page; 1689b2261026SJoonsoo Kim int nr_nodes, node; 1690af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 1691b2261026SJoonsoo Kim 1692b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1693f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, 1694f60858f9SMike Kravetz node_alloc_noretry); 1695af0fb9dfSMichal Hocko if (page) 1696b2261026SJoonsoo Kim break; 1697b2261026SJoonsoo Kim } 1698b2261026SJoonsoo Kim 1699af0fb9dfSMichal Hocko if (!page) 1700af0fb9dfSMichal Hocko return 0; 1701b2261026SJoonsoo Kim 1702af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 1703af0fb9dfSMichal Hocko 1704af0fb9dfSMichal Hocko return 1; 1705b2261026SJoonsoo Kim } 1706b2261026SJoonsoo Kim 1707e8c5c824SLee Schermerhorn /* 1708e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 1709e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 1710e8c5c824SLee Schermerhorn * balanced over allowed nodes. 1711e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 1712e8c5c824SLee Schermerhorn */ 17136ae11b27SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 17146ae11b27SLee Schermerhorn bool acct_surplus) 1715e8c5c824SLee Schermerhorn { 1716b2261026SJoonsoo Kim int nr_nodes, node; 1717e8c5c824SLee Schermerhorn int ret = 0; 1718e8c5c824SLee Schermerhorn 1719b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1720685f3457SLee Schermerhorn /* 1721685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 1722685f3457SLee Schermerhorn * nodes with surplus pages. 1723685f3457SLee Schermerhorn */ 1724b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 1725b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 1726e8c5c824SLee Schermerhorn struct page *page = 1727b2261026SJoonsoo Kim list_entry(h->hugepage_freelists[node].next, 1728e8c5c824SLee Schermerhorn struct page, lru); 1729e8c5c824SLee Schermerhorn list_del(&page->lru); 1730e8c5c824SLee Schermerhorn h->free_huge_pages--; 1731b2261026SJoonsoo Kim h->free_huge_pages_node[node]--; 1732685f3457SLee Schermerhorn if (acct_surplus) { 1733685f3457SLee Schermerhorn h->surplus_huge_pages--; 1734b2261026SJoonsoo Kim h->surplus_huge_pages_node[node]--; 1735685f3457SLee Schermerhorn } 1736e8c5c824SLee Schermerhorn update_and_free_page(h, page); 1737e8c5c824SLee Schermerhorn ret = 1; 17389a76db09SLee Schermerhorn break; 1739e8c5c824SLee Schermerhorn } 1740b2261026SJoonsoo Kim } 1741e8c5c824SLee Schermerhorn 1742e8c5c824SLee Schermerhorn return ret; 1743e8c5c824SLee Schermerhorn } 1744e8c5c824SLee Schermerhorn 1745c8721bbbSNaoya Horiguchi /* 1746c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 1747faf53defSNaoya Horiguchi * nothing for in-use hugepages and non-hugepages. 1748faf53defSNaoya Horiguchi * This function returns values like below: 1749faf53defSNaoya Horiguchi * 1750faf53defSNaoya Horiguchi * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 1751faf53defSNaoya Horiguchi * (allocated or reserved.) 1752faf53defSNaoya Horiguchi * 0: successfully dissolved free hugepages or the page is not a 1753faf53defSNaoya Horiguchi * hugepage (considered as already dissolved) 1754c8721bbbSNaoya Horiguchi */ 1755c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 1756c8721bbbSNaoya Horiguchi { 17576bc9b564SNaoya Horiguchi int rc = -EBUSY; 1758082d5b6bSGerald Schaefer 17597ffddd49SMuchun Song retry: 1760faf53defSNaoya Horiguchi /* Not to disrupt normal path by vainly holding hugetlb_lock */ 1761faf53defSNaoya Horiguchi if (!PageHuge(page)) 1762faf53defSNaoya Horiguchi return 0; 1763faf53defSNaoya Horiguchi 1764c8721bbbSNaoya Horiguchi spin_lock(&hugetlb_lock); 1765faf53defSNaoya Horiguchi if (!PageHuge(page)) { 1766faf53defSNaoya Horiguchi rc = 0; 1767faf53defSNaoya Horiguchi goto out; 1768faf53defSNaoya Horiguchi } 1769faf53defSNaoya Horiguchi 1770faf53defSNaoya Horiguchi if (!page_count(page)) { 17712247bb33SGerald Schaefer struct page *head = compound_head(page); 17722247bb33SGerald Schaefer struct hstate *h = page_hstate(head); 17732247bb33SGerald Schaefer int nid = page_to_nid(head); 17746bc9b564SNaoya Horiguchi if (h->free_huge_pages - h->resv_huge_pages == 0) 1775082d5b6bSGerald Schaefer goto out; 17767ffddd49SMuchun Song 17777ffddd49SMuchun Song /* 17787ffddd49SMuchun Song * We should make sure that the page is already on the free list 17797ffddd49SMuchun Song * when it is dissolved. 17807ffddd49SMuchun Song */ 17817ffddd49SMuchun Song if (unlikely(!PageHugeFreed(head))) { 17827ffddd49SMuchun Song spin_unlock(&hugetlb_lock); 17837ffddd49SMuchun Song cond_resched(); 17847ffddd49SMuchun Song 17857ffddd49SMuchun Song /* 17867ffddd49SMuchun Song * Theoretically, we should return -EBUSY when we 17877ffddd49SMuchun Song * encounter this race. In fact, we have a chance 17887ffddd49SMuchun Song * to successfully dissolve the page if we do a 17897ffddd49SMuchun Song * retry. Because the race window is quite small. 17907ffddd49SMuchun Song * If we seize this opportunity, it is an optimization 17917ffddd49SMuchun Song * for increasing the success rate of dissolving page. 17927ffddd49SMuchun Song */ 17937ffddd49SMuchun Song goto retry; 17947ffddd49SMuchun Song } 17957ffddd49SMuchun Song 1796c3114a84SAnshuman Khandual /* 1797c3114a84SAnshuman Khandual * Move PageHWPoison flag from head page to the raw error page, 1798c3114a84SAnshuman Khandual * which makes any subpages rather than the error page reusable. 1799c3114a84SAnshuman Khandual */ 1800c3114a84SAnshuman Khandual if (PageHWPoison(head) && page != head) { 1801c3114a84SAnshuman Khandual SetPageHWPoison(page); 1802c3114a84SAnshuman Khandual ClearPageHWPoison(head); 1803c3114a84SAnshuman Khandual } 18042247bb33SGerald Schaefer list_del(&head->lru); 1805c8721bbbSNaoya Horiguchi h->free_huge_pages--; 1806c8721bbbSNaoya Horiguchi h->free_huge_pages_node[nid]--; 1807c1470b33Szhong jiang h->max_huge_pages--; 18082247bb33SGerald Schaefer update_and_free_page(h, head); 18096bc9b564SNaoya Horiguchi rc = 0; 1810c8721bbbSNaoya Horiguchi } 1811082d5b6bSGerald Schaefer out: 1812c8721bbbSNaoya Horiguchi spin_unlock(&hugetlb_lock); 1813082d5b6bSGerald Schaefer return rc; 1814c8721bbbSNaoya Horiguchi } 1815c8721bbbSNaoya Horiguchi 1816c8721bbbSNaoya Horiguchi /* 1817c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 1818c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 18192247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 18202247bb33SGerald Schaefer * part of it lies within the given range. 1821082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 1822082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 1823c8721bbbSNaoya Horiguchi */ 1824082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 1825c8721bbbSNaoya Horiguchi { 1826c8721bbbSNaoya Horiguchi unsigned long pfn; 1827eb03aa00SGerald Schaefer struct page *page; 1828082d5b6bSGerald Schaefer int rc = 0; 1829c8721bbbSNaoya Horiguchi 1830d0177639SLi Zhong if (!hugepages_supported()) 1831082d5b6bSGerald Schaefer return rc; 1832d0177639SLi Zhong 1833eb03aa00SGerald Schaefer for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { 1834eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 1835eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 1836eb03aa00SGerald Schaefer if (rc) 1837082d5b6bSGerald Schaefer break; 1838eb03aa00SGerald Schaefer } 1839082d5b6bSGerald Schaefer 1840082d5b6bSGerald Schaefer return rc; 1841c8721bbbSNaoya Horiguchi } 1842c8721bbbSNaoya Horiguchi 1843ab5ac90aSMichal Hocko /* 1844ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 1845ab5ac90aSMichal Hocko */ 18460c397daeSMichal Hocko static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, 1847aaf14e40SMichal Hocko int nid, nodemask_t *nmask) 18487893d1d5SAdam Litke { 18499980d744SMichal Hocko struct page *page = NULL; 18507893d1d5SAdam Litke 1851bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1852aa888a74SAndi Kleen return NULL; 1853aa888a74SAndi Kleen 1854d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 18559980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 18569980d744SMichal Hocko goto out_unlock; 1857d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1858d1c3fb1fSNishanth Aravamudan 1859f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 18609980d744SMichal Hocko if (!page) 18610c397daeSMichal Hocko return NULL; 1862d1c3fb1fSNishanth Aravamudan 18637893d1d5SAdam Litke spin_lock(&hugetlb_lock); 18649980d744SMichal Hocko /* 18659980d744SMichal Hocko * We could have raced with the pool size change. 18669980d744SMichal Hocko * Double check that and simply deallocate the new page 18679980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 18689980d744SMichal Hocko * temporary page to workaround the nasty free_huge_page 18699980d744SMichal Hocko * codeflow 18709980d744SMichal Hocko */ 18719980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 18729980d744SMichal Hocko SetPageHugeTemporary(page); 18732bf753e6SKai Shen spin_unlock(&hugetlb_lock); 18749980d744SMichal Hocko put_page(page); 18752bf753e6SKai Shen return NULL; 18769980d744SMichal Hocko } else { 18779980d744SMichal Hocko h->surplus_huge_pages++; 18784704dea3SMichal Hocko h->surplus_huge_pages_node[page_to_nid(page)]++; 18797893d1d5SAdam Litke } 18809980d744SMichal Hocko 18819980d744SMichal Hocko out_unlock: 1882d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 18837893d1d5SAdam Litke 18847893d1d5SAdam Litke return page; 18857893d1d5SAdam Litke } 18867893d1d5SAdam Litke 1887bbe88753SJoonsoo Kim static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, 1888ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 1889ab5ac90aSMichal Hocko { 1890ab5ac90aSMichal Hocko struct page *page; 1891ab5ac90aSMichal Hocko 1892ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 1893ab5ac90aSMichal Hocko return NULL; 1894ab5ac90aSMichal Hocko 1895f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 1896ab5ac90aSMichal Hocko if (!page) 1897ab5ac90aSMichal Hocko return NULL; 1898ab5ac90aSMichal Hocko 1899ab5ac90aSMichal Hocko /* 1900ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 1901ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 1902ab5ac90aSMichal Hocko */ 1903ab5ac90aSMichal Hocko SetPageHugeTemporary(page); 1904ab5ac90aSMichal Hocko 1905ab5ac90aSMichal Hocko return page; 1906ab5ac90aSMichal Hocko } 1907ab5ac90aSMichal Hocko 1908e4e574b7SAdam Litke /* 1909099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 1910099730d6SDave Hansen */ 1911e0ec90eeSDave Hansen static 19120c397daeSMichal Hocko struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, 1913099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 1914099730d6SDave Hansen { 1915aaf14e40SMichal Hocko struct page *page; 1916aaf14e40SMichal Hocko struct mempolicy *mpol; 1917aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 1918aaf14e40SMichal Hocko int nid; 1919aaf14e40SMichal Hocko nodemask_t *nodemask; 1920aaf14e40SMichal Hocko 1921aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 19220c397daeSMichal Hocko page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); 1923aaf14e40SMichal Hocko mpol_cond_put(mpol); 1924aaf14e40SMichal Hocko 1925aaf14e40SMichal Hocko return page; 1926099730d6SDave Hansen } 1927099730d6SDave Hansen 1928ab5ac90aSMichal Hocko /* page migration callback function */ 19293e59fcb0SMichal Hocko struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, 1930d92bbc27SJoonsoo Kim nodemask_t *nmask, gfp_t gfp_mask) 19314db9b2efSMichal Hocko { 19324db9b2efSMichal Hocko spin_lock(&hugetlb_lock); 19334db9b2efSMichal Hocko if (h->free_huge_pages - h->resv_huge_pages > 0) { 19343e59fcb0SMichal Hocko struct page *page; 19353e59fcb0SMichal Hocko 19363e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); 19373e59fcb0SMichal Hocko if (page) { 19383e59fcb0SMichal Hocko spin_unlock(&hugetlb_lock); 19393e59fcb0SMichal Hocko return page; 19404db9b2efSMichal Hocko } 19414db9b2efSMichal Hocko } 19424db9b2efSMichal Hocko spin_unlock(&hugetlb_lock); 19434db9b2efSMichal Hocko 19440c397daeSMichal Hocko return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); 19454db9b2efSMichal Hocko } 19464db9b2efSMichal Hocko 1947ebd63723SMichal Hocko /* mempolicy aware migration callback */ 1948389c8178SMichal Hocko struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, 1949389c8178SMichal Hocko unsigned long address) 1950ebd63723SMichal Hocko { 1951ebd63723SMichal Hocko struct mempolicy *mpol; 1952ebd63723SMichal Hocko nodemask_t *nodemask; 1953ebd63723SMichal Hocko struct page *page; 1954ebd63723SMichal Hocko gfp_t gfp_mask; 1955ebd63723SMichal Hocko int node; 1956ebd63723SMichal Hocko 1957ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 1958ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 1959d92bbc27SJoonsoo Kim page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); 1960ebd63723SMichal Hocko mpol_cond_put(mpol); 1961ebd63723SMichal Hocko 1962ebd63723SMichal Hocko return page; 1963ebd63723SMichal Hocko } 1964ebd63723SMichal Hocko 1965bf50bab2SNaoya Horiguchi /* 196625985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 1967e4e574b7SAdam Litke * of size 'delta'. 1968e4e574b7SAdam Litke */ 19690a4f3d1bSLiu Xiang static int gather_surplus_pages(struct hstate *h, long delta) 19701b2a1e7bSJules Irenge __must_hold(&hugetlb_lock) 1971e4e574b7SAdam Litke { 1972e4e574b7SAdam Litke struct list_head surplus_list; 1973e4e574b7SAdam Litke struct page *page, *tmp; 19740a4f3d1bSLiu Xiang int ret; 19750a4f3d1bSLiu Xiang long i; 19760a4f3d1bSLiu Xiang long needed, allocated; 197728073b02SHillf Danton bool alloc_ok = true; 1978e4e574b7SAdam Litke 1979a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 1980ac09b3a1SAdam Litke if (needed <= 0) { 1981a5516438SAndi Kleen h->resv_huge_pages += delta; 1982e4e574b7SAdam Litke return 0; 1983ac09b3a1SAdam Litke } 1984e4e574b7SAdam Litke 1985e4e574b7SAdam Litke allocated = 0; 1986e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 1987e4e574b7SAdam Litke 1988e4e574b7SAdam Litke ret = -ENOMEM; 1989e4e574b7SAdam Litke retry: 1990e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 1991e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 19920c397daeSMichal Hocko page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), 1993aaf14e40SMichal Hocko NUMA_NO_NODE, NULL); 199428073b02SHillf Danton if (!page) { 199528073b02SHillf Danton alloc_ok = false; 199628073b02SHillf Danton break; 199728073b02SHillf Danton } 1998e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 199969ed779aSDavid Rientjes cond_resched(); 2000e4e574b7SAdam Litke } 200128073b02SHillf Danton allocated += i; 2002e4e574b7SAdam Litke 2003e4e574b7SAdam Litke /* 2004e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 2005e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 2006e4e574b7SAdam Litke */ 2007e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 2008a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 2009a5516438SAndi Kleen (h->free_huge_pages + allocated); 201028073b02SHillf Danton if (needed > 0) { 201128073b02SHillf Danton if (alloc_ok) 2012e4e574b7SAdam Litke goto retry; 201328073b02SHillf Danton /* 201428073b02SHillf Danton * We were not able to allocate enough pages to 201528073b02SHillf Danton * satisfy the entire reservation so we free what 201628073b02SHillf Danton * we've allocated so far. 201728073b02SHillf Danton */ 201828073b02SHillf Danton goto free; 201928073b02SHillf Danton } 2020e4e574b7SAdam Litke /* 2021e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 202225985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 2023e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 2024ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 2025ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 2026ac09b3a1SAdam Litke * before they are reserved. 2027e4e574b7SAdam Litke */ 2028e4e574b7SAdam Litke needed += allocated; 2029a5516438SAndi Kleen h->resv_huge_pages += delta; 2030e4e574b7SAdam Litke ret = 0; 2031a9869b83SNaoya Horiguchi 203219fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 203319fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 2034e558464bSMuchun Song int zeroed; 2035e558464bSMuchun Song 203619fc3f0aSAdam Litke if ((--needed) < 0) 203719fc3f0aSAdam Litke break; 2038a9869b83SNaoya Horiguchi /* 2039a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 2040a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 2041a9869b83SNaoya Horiguchi */ 2042e558464bSMuchun Song zeroed = put_page_testzero(page); 2043e558464bSMuchun Song VM_BUG_ON_PAGE(!zeroed, page); 2044a5516438SAndi Kleen enqueue_huge_page(h, page); 204519fc3f0aSAdam Litke } 204628073b02SHillf Danton free: 2047b0365c8dSHillf Danton spin_unlock(&hugetlb_lock); 204819fc3f0aSAdam Litke 204919fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 2050c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 2051a9869b83SNaoya Horiguchi put_page(page); 205219fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 2053e4e574b7SAdam Litke 2054e4e574b7SAdam Litke return ret; 2055e4e574b7SAdam Litke } 2056e4e574b7SAdam Litke 2057e4e574b7SAdam Litke /* 2058e5bbc8a6SMike Kravetz * This routine has two main purposes: 2059e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 2060e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 2061e5bbc8a6SMike Kravetz * to the associated reservation map. 2062e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 2063e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 2064e5bbc8a6SMike Kravetz * 2065e5bbc8a6SMike Kravetz * Called with hugetlb_lock held. However, the lock could be dropped (and 2066e5bbc8a6SMike Kravetz * reacquired) during calls to cond_resched_lock. Whenever dropping the lock, 2067e5bbc8a6SMike Kravetz * we must make sure nobody else can claim pages we are in the process of 2068e5bbc8a6SMike Kravetz * freeing. Do this by ensuring resv_huge_page always is greater than the 2069e5bbc8a6SMike Kravetz * number of huge pages we plan to free when dropping the lock. 2070e4e574b7SAdam Litke */ 2071a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 2072a5516438SAndi Kleen unsigned long unused_resv_pages) 2073e4e574b7SAdam Litke { 2074e4e574b7SAdam Litke unsigned long nr_pages; 2075e4e574b7SAdam Litke 2076aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 2077bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2078e5bbc8a6SMike Kravetz goto out; 2079aa888a74SAndi Kleen 2080e5bbc8a6SMike Kravetz /* 2081e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 2082e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 2083e5bbc8a6SMike Kravetz */ 2084a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 2085e4e574b7SAdam Litke 2086685f3457SLee Schermerhorn /* 2087685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 20889b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 20899b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 20909b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 20919e7ee400SRandy Dunlap * free_pool_huge_page() will balance the freed pages across the 20929b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 2093e5bbc8a6SMike Kravetz * 2094e5bbc8a6SMike Kravetz * Note that we decrement resv_huge_pages as we free the pages. If 2095e5bbc8a6SMike Kravetz * we drop the lock, resv_huge_pages will still be sufficiently large 2096e5bbc8a6SMike Kravetz * to cover subsequent pages we may free. 2097685f3457SLee Schermerhorn */ 2098685f3457SLee Schermerhorn while (nr_pages--) { 2099e5bbc8a6SMike Kravetz h->resv_huge_pages--; 2100e5bbc8a6SMike Kravetz unused_resv_pages--; 21018cebfcd0SLai Jiangshan if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) 2102e5bbc8a6SMike Kravetz goto out; 21037848a4bfSMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 2104e4e574b7SAdam Litke } 2105e5bbc8a6SMike Kravetz 2106e5bbc8a6SMike Kravetz out: 2107e5bbc8a6SMike Kravetz /* Fully uncommit the reservation */ 2108e5bbc8a6SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 2109e4e574b7SAdam Litke } 2110e4e574b7SAdam Litke 21115e911373SMike Kravetz 2112c37f9fb1SAndy Whitcroft /* 2113feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 21145e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 2115cf3ad20bSMike Kravetz * 2116cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 2117cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 2118cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 2119cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 2120cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 2121feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 2122feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 2123feba16e2SMike Kravetz * is called in such cases. 2124cf3ad20bSMike Kravetz * 2125cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 2126cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 2127cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 2128cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 2129cf3ad20bSMike Kravetz * take appropriate action. 213096b96a96SMike Kravetz * 213196b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 213296b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 213396b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 213496b96a96SMike Kravetz * reservation exists. 2135c37f9fb1SAndy Whitcroft */ 21365e911373SMike Kravetz enum vma_resv_mode { 21375e911373SMike Kravetz VMA_NEEDS_RESV, 21385e911373SMike Kravetz VMA_COMMIT_RESV, 2139feba16e2SMike Kravetz VMA_END_RESV, 214096b96a96SMike Kravetz VMA_ADD_RESV, 21415e911373SMike Kravetz }; 2142cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 2143cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 21445e911373SMike Kravetz enum vma_resv_mode mode) 2145c37f9fb1SAndy Whitcroft { 21464e35f483SJoonsoo Kim struct resv_map *resv; 21474e35f483SJoonsoo Kim pgoff_t idx; 2148cf3ad20bSMike Kravetz long ret; 21490db9d74eSMina Almasry long dummy_out_regions_needed; 2150c37f9fb1SAndy Whitcroft 21514e35f483SJoonsoo Kim resv = vma_resv_map(vma); 21524e35f483SJoonsoo Kim if (!resv) 2153c37f9fb1SAndy Whitcroft return 1; 2154c37f9fb1SAndy Whitcroft 21554e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 21565e911373SMike Kravetz switch (mode) { 21575e911373SMike Kravetz case VMA_NEEDS_RESV: 21580db9d74eSMina Almasry ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); 21590db9d74eSMina Almasry /* We assume that vma_reservation_* routines always operate on 21600db9d74eSMina Almasry * 1 page, and that adding to resv map a 1 page entry can only 21610db9d74eSMina Almasry * ever require 1 region. 21620db9d74eSMina Almasry */ 21630db9d74eSMina Almasry VM_BUG_ON(dummy_out_regions_needed != 1); 21645e911373SMike Kravetz break; 21655e911373SMike Kravetz case VMA_COMMIT_RESV: 2166075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 21670db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 21680db9d74eSMina Almasry VM_BUG_ON(ret < 0); 21695e911373SMike Kravetz break; 2170feba16e2SMike Kravetz case VMA_END_RESV: 21710db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 21725e911373SMike Kravetz ret = 0; 21735e911373SMike Kravetz break; 217496b96a96SMike Kravetz case VMA_ADD_RESV: 21750db9d74eSMina Almasry if (vma->vm_flags & VM_MAYSHARE) { 2176075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 21770db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 21780db9d74eSMina Almasry VM_BUG_ON(ret < 0); 21790db9d74eSMina Almasry } else { 21800db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 218196b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 218296b96a96SMike Kravetz } 218396b96a96SMike Kravetz break; 21845e911373SMike Kravetz default: 21855e911373SMike Kravetz BUG(); 21865e911373SMike Kravetz } 218784afd99bSAndy Whitcroft 21884e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) 2189cf3ad20bSMike Kravetz return ret; 219067961f9dSMike Kravetz else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { 219167961f9dSMike Kravetz /* 219267961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 219367961f9dSMike Kravetz * However, a file associated with mapping could have been 219467961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 219567961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 219667961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 219767961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 219867961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 219967961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 220067961f9dSMike Kravetz * reservation has already been consumed. As a result, the 220167961f9dSMike Kravetz * return value of this routine is the opposite of the 220267961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 220367961f9dSMike Kravetz */ 220467961f9dSMike Kravetz if (ret) 220567961f9dSMike Kravetz return 0; 220667961f9dSMike Kravetz else 220767961f9dSMike Kravetz return 1; 220867961f9dSMike Kravetz } 22094e35f483SJoonsoo Kim else 2210cf3ad20bSMike Kravetz return ret < 0 ? ret : 0; 221184afd99bSAndy Whitcroft } 2212cf3ad20bSMike Kravetz 2213cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 2214a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 2215c37f9fb1SAndy Whitcroft { 22165e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 2217cf3ad20bSMike Kravetz } 2218c37f9fb1SAndy Whitcroft 2219cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 2220cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 2221cf3ad20bSMike Kravetz { 22225e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 22235e911373SMike Kravetz } 22245e911373SMike Kravetz 2225feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 22265e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 22275e911373SMike Kravetz { 2228feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 2229c37f9fb1SAndy Whitcroft } 2230c37f9fb1SAndy Whitcroft 223196b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 223296b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 223396b96a96SMike Kravetz { 223496b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 223596b96a96SMike Kravetz } 223696b96a96SMike Kravetz 223796b96a96SMike Kravetz /* 223896b96a96SMike Kravetz * This routine is called to restore a reservation on error paths. In the 223996b96a96SMike Kravetz * specific error paths, a huge page was allocated (via alloc_huge_page) 224096b96a96SMike Kravetz * and is about to be freed. If a reservation for the page existed, 2241d6995da3SMike Kravetz * alloc_huge_page would have consumed the reservation and set 2242d6995da3SMike Kravetz * HPageRestoreReserve in the newly allocated page. When the page is freed 2243d6995da3SMike Kravetz * via free_huge_page, the global reservation count will be incremented if 2244d6995da3SMike Kravetz * HPageRestoreReserve is set. However, free_huge_page can not adjust the 2245d6995da3SMike Kravetz * reserve map. Adjust the reserve map here to be consistent with global 2246d6995da3SMike Kravetz * reserve count adjustments to be made by free_huge_page. 224796b96a96SMike Kravetz */ 224896b96a96SMike Kravetz static void restore_reserve_on_error(struct hstate *h, 224996b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long address, 225096b96a96SMike Kravetz struct page *page) 225196b96a96SMike Kravetz { 2252d6995da3SMike Kravetz if (unlikely(HPageRestoreReserve(page))) { 225396b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 225496b96a96SMike Kravetz 225596b96a96SMike Kravetz if (unlikely(rc < 0)) { 225696b96a96SMike Kravetz /* 225796b96a96SMike Kravetz * Rare out of memory condition in reserve map 2258d6995da3SMike Kravetz * manipulation. Clear HPageRestoreReserve so that 225996b96a96SMike Kravetz * global reserve count will not be incremented 226096b96a96SMike Kravetz * by free_huge_page. This will make it appear 226196b96a96SMike Kravetz * as though the reservation for this page was 226296b96a96SMike Kravetz * consumed. This may prevent the task from 226396b96a96SMike Kravetz * faulting in the page at a later time. This 226496b96a96SMike Kravetz * is better than inconsistent global huge page 226596b96a96SMike Kravetz * accounting of reserve counts. 226696b96a96SMike Kravetz */ 2267d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 226896b96a96SMike Kravetz } else if (rc) { 226996b96a96SMike Kravetz rc = vma_add_reservation(h, vma, address); 227096b96a96SMike Kravetz if (unlikely(rc < 0)) 227196b96a96SMike Kravetz /* 227296b96a96SMike Kravetz * See above comment about rare out of 227396b96a96SMike Kravetz * memory condition. 227496b96a96SMike Kravetz */ 2275d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 227696b96a96SMike Kravetz } else 227796b96a96SMike Kravetz vma_end_reservation(h, vma, address); 227896b96a96SMike Kravetz } 227996b96a96SMike Kravetz } 228096b96a96SMike Kravetz 228170c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 228204f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 2283348ea204SAdam Litke { 228490481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 2285a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2286348ea204SAdam Litke struct page *page; 2287d85f69b0SMike Kravetz long map_chg, map_commit; 2288d85f69b0SMike Kravetz long gbl_chg; 22896d76dcf4SAneesh Kumar K.V int ret, idx; 22906d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 229108cf9fafSMina Almasry bool deferred_reserve; 22922fc39cecSAdam Litke 22936d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 2294a1e78772SMel Gorman /* 2295d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 2296d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 2297d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 2298a1e78772SMel Gorman */ 2299d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 2300d85f69b0SMike Kravetz if (map_chg < 0) 230176dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 2302d85f69b0SMike Kravetz 2303d85f69b0SMike Kravetz /* 2304d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 2305d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 2306d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 2307d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 2308d85f69b0SMike Kravetz * checked against any subpool limit. 2309d85f69b0SMike Kravetz */ 2310d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 2311d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 2312d85f69b0SMike Kravetz if (gbl_chg < 0) { 2313feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 231476dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 23155e911373SMike Kravetz } 231690d8b7e6SAdam Litke 2317d85f69b0SMike Kravetz /* 2318d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 2319d85f69b0SMike Kravetz * map, there could be reservations associated with the 2320d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 2321d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 2322d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 2323d85f69b0SMike Kravetz * the subpool reservations. 2324d85f69b0SMike Kravetz */ 2325d85f69b0SMike Kravetz if (avoid_reserve) 2326d85f69b0SMike Kravetz gbl_chg = 1; 2327d85f69b0SMike Kravetz } 2328d85f69b0SMike Kravetz 232908cf9fafSMina Almasry /* If this allocation is not consuming a reservation, charge it now. 233008cf9fafSMina Almasry */ 233108cf9fafSMina Almasry deferred_reserve = map_chg || avoid_reserve || !vma_resv_map(vma); 233208cf9fafSMina Almasry if (deferred_reserve) { 233308cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 233408cf9fafSMina Almasry idx, pages_per_huge_page(h), &h_cg); 23358f34af6fSJianyu Zhan if (ret) 23368f34af6fSJianyu Zhan goto out_subpool_put; 233708cf9fafSMina Almasry } 233808cf9fafSMina Almasry 233908cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 234008cf9fafSMina Almasry if (ret) 234108cf9fafSMina Almasry goto out_uncharge_cgroup_reservation; 23428f34af6fSJianyu Zhan 2343a1e78772SMel Gorman spin_lock(&hugetlb_lock); 2344d85f69b0SMike Kravetz /* 2345d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 2346d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 2347d85f69b0SMike Kravetz * a reservation exists for the allocation. 2348d85f69b0SMike Kravetz */ 2349d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 235081a6fcaeSJoonsoo Kim if (!page) { 235194ae8ba7SAneesh Kumar K.V spin_unlock(&hugetlb_lock); 23520c397daeSMichal Hocko page = alloc_buddy_huge_page_with_mpol(h, vma, addr); 23538f34af6fSJianyu Zhan if (!page) 23548f34af6fSJianyu Zhan goto out_uncharge_cgroup; 2355a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 2356d6995da3SMike Kravetz SetHPageRestoreReserve(page); 2357a88c7695SNaoya Horiguchi h->resv_huge_pages--; 2358a88c7695SNaoya Horiguchi } 235979dbb236SAneesh Kumar K.V spin_lock(&hugetlb_lock); 236015a8d68eSWei Yang list_add(&page->lru, &h->hugepage_activelist); 236181a6fcaeSJoonsoo Kim /* Fall through */ 2362a1e78772SMel Gorman } 236381a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 236408cf9fafSMina Almasry /* If allocation is not consuming a reservation, also store the 236508cf9fafSMina Almasry * hugetlb_cgroup pointer on the page. 236608cf9fafSMina Almasry */ 236708cf9fafSMina Almasry if (deferred_reserve) { 236808cf9fafSMina Almasry hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), 236908cf9fafSMina Almasry h_cg, page); 237008cf9fafSMina Almasry } 237108cf9fafSMina Almasry 237281a6fcaeSJoonsoo Kim spin_unlock(&hugetlb_lock); 2373a1e78772SMel Gorman 2374d6995da3SMike Kravetz hugetlb_set_page_subpool(page, spool); 2375a1e78772SMel Gorman 2376d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 2377d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 237833039678SMike Kravetz /* 237933039678SMike Kravetz * The page was added to the reservation map between 238033039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 238133039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 238233039678SMike Kravetz * Adjust for the subpool count incremented above AND 238333039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 238433039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 238533039678SMike Kravetz * no longer applies. 238633039678SMike Kravetz */ 238733039678SMike Kravetz long rsv_adjust; 238833039678SMike Kravetz 238933039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 239033039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 239179aa925bSMike Kravetz if (deferred_reserve) 239279aa925bSMike Kravetz hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 239379aa925bSMike Kravetz pages_per_huge_page(h), page); 239433039678SMike Kravetz } 23957893d1d5SAdam Litke return page; 23968f34af6fSJianyu Zhan 23978f34af6fSJianyu Zhan out_uncharge_cgroup: 23988f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 239908cf9fafSMina Almasry out_uncharge_cgroup_reservation: 240008cf9fafSMina Almasry if (deferred_reserve) 240108cf9fafSMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), 240208cf9fafSMina Almasry h_cg); 24038f34af6fSJianyu Zhan out_subpool_put: 2404d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 24058f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 2406feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 24078f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 2408b45b5bd6SDavid Gibson } 2409b45b5bd6SDavid Gibson 2410e24a1307SAneesh Kumar K.V int alloc_bootmem_huge_page(struct hstate *h) 2411e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 2412e24a1307SAneesh Kumar K.V int __alloc_bootmem_huge_page(struct hstate *h) 2413aa888a74SAndi Kleen { 2414aa888a74SAndi Kleen struct huge_bootmem_page *m; 2415b2261026SJoonsoo Kim int nr_nodes, node; 2416aa888a74SAndi Kleen 2417b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 2418aa888a74SAndi Kleen void *addr; 2419aa888a74SAndi Kleen 2420eb31d559SMike Rapoport addr = memblock_alloc_try_nid_raw( 24218b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 242297ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 2423aa888a74SAndi Kleen if (addr) { 2424aa888a74SAndi Kleen /* 2425aa888a74SAndi Kleen * Use the beginning of the huge page to store the 2426aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 2427aa888a74SAndi Kleen * puts them into the mem_map). 2428aa888a74SAndi Kleen */ 2429aa888a74SAndi Kleen m = addr; 2430aa888a74SAndi Kleen goto found; 2431aa888a74SAndi Kleen } 2432aa888a74SAndi Kleen } 2433aa888a74SAndi Kleen return 0; 2434aa888a74SAndi Kleen 2435aa888a74SAndi Kleen found: 2436df994eadSLuiz Capitulino BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); 2437aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 2438330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 2439aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 2440aa888a74SAndi Kleen m->hstate = h; 2441aa888a74SAndi Kleen return 1; 2442aa888a74SAndi Kleen } 2443aa888a74SAndi Kleen 2444d00181b9SKirill A. Shutemov static void __init prep_compound_huge_page(struct page *page, 2445d00181b9SKirill A. Shutemov unsigned int order) 244618229df5SAndy Whitcroft { 244718229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 244818229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 244918229df5SAndy Whitcroft else 245018229df5SAndy Whitcroft prep_compound_page(page, order); 245118229df5SAndy Whitcroft } 245218229df5SAndy Whitcroft 2453aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 2454aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 2455aa888a74SAndi Kleen { 2456aa888a74SAndi Kleen struct huge_bootmem_page *m; 2457aa888a74SAndi Kleen 2458aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 245940d18ebfSMike Kravetz struct page *page = virt_to_page(m); 2460aa888a74SAndi Kleen struct hstate *h = m->hstate; 2461ee8f248dSBecky Bruce 2462aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 2463c78a7f36SMiaohe Lin prep_compound_huge_page(page, huge_page_order(h)); 2464ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 2465aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 2466af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 2467af0fb9dfSMichal Hocko 2468b0320c7bSRafael Aquini /* 2469b0320c7bSRafael Aquini * If we had gigantic hugepages allocated at boot time, we need 2470b0320c7bSRafael Aquini * to restore the 'stolen' pages to totalram_pages in order to 2471b0320c7bSRafael Aquini * fix confusing memory reports from free(1) and another 2472b0320c7bSRafael Aquini * side-effects, like CommitLimit going negative. 2473b0320c7bSRafael Aquini */ 2474bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2475c78a7f36SMiaohe Lin adjust_managed_page_count(page, pages_per_huge_page(h)); 2476520495feSCannon Matthews cond_resched(); 2477aa888a74SAndi Kleen } 2478aa888a74SAndi Kleen } 2479aa888a74SAndi Kleen 24808faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 24811da177e4SLinus Torvalds { 24821da177e4SLinus Torvalds unsigned long i; 2483f60858f9SMike Kravetz nodemask_t *node_alloc_noretry; 2484f60858f9SMike Kravetz 2485f60858f9SMike Kravetz if (!hstate_is_gigantic(h)) { 2486f60858f9SMike Kravetz /* 2487f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 2488f60858f9SMike Kravetz * Ignore errors as lower level routines can deal with 2489f60858f9SMike Kravetz * node_alloc_noretry == NULL. If this kmalloc fails at boot 2490f60858f9SMike Kravetz * time, we are likely in bigger trouble. 2491f60858f9SMike Kravetz */ 2492f60858f9SMike Kravetz node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), 2493f60858f9SMike Kravetz GFP_KERNEL); 2494f60858f9SMike Kravetz } else { 2495f60858f9SMike Kravetz /* allocations done at boot time */ 2496f60858f9SMike Kravetz node_alloc_noretry = NULL; 2497f60858f9SMike Kravetz } 2498f60858f9SMike Kravetz 2499f60858f9SMike Kravetz /* bit mask controlling how hard we retry per-node allocations */ 2500f60858f9SMike Kravetz if (node_alloc_noretry) 2501f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 25021da177e4SLinus Torvalds 2503e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 2504bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 2505dbda8feaSBarry Song if (hugetlb_cma_size) { 2506cf11e85fSRoman Gushchin pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); 25077ecc9565SChen Wandun goto free; 2508cf11e85fSRoman Gushchin } 2509aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 2510aa888a74SAndi Kleen break; 25110c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 2512f60858f9SMike Kravetz &node_states[N_MEMORY], 2513f60858f9SMike Kravetz node_alloc_noretry)) 25141da177e4SLinus Torvalds break; 251569ed779aSDavid Rientjes cond_resched(); 25161da177e4SLinus Torvalds } 2517d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 2518d715cf80SLiam R. Howlett char buf[32]; 2519d715cf80SLiam R. Howlett 2520c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2521d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 2522d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 25238faa8b07SAndi Kleen h->max_huge_pages = i; 2524e5ff2159SAndi Kleen } 25257ecc9565SChen Wandun free: 2526f60858f9SMike Kravetz kfree(node_alloc_noretry); 2527d715cf80SLiam R. Howlett } 2528e5ff2159SAndi Kleen 2529e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 2530e5ff2159SAndi Kleen { 2531e5ff2159SAndi Kleen struct hstate *h; 2532e5ff2159SAndi Kleen 2533e5ff2159SAndi Kleen for_each_hstate(h) { 2534641844f5SNaoya Horiguchi if (minimum_order > huge_page_order(h)) 2535641844f5SNaoya Horiguchi minimum_order = huge_page_order(h); 2536641844f5SNaoya Horiguchi 25378faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 2538bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 25398faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 2540e5ff2159SAndi Kleen } 2541641844f5SNaoya Horiguchi VM_BUG_ON(minimum_order == UINT_MAX); 2542e5ff2159SAndi Kleen } 2543e5ff2159SAndi Kleen 2544e5ff2159SAndi Kleen static void __init report_hugepages(void) 2545e5ff2159SAndi Kleen { 2546e5ff2159SAndi Kleen struct hstate *h; 2547e5ff2159SAndi Kleen 2548e5ff2159SAndi Kleen for_each_hstate(h) { 25494abd32dbSAndi Kleen char buf[32]; 2550c6247f72SMatthew Wilcox 2551c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2552ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 2553c6247f72SMatthew Wilcox buf, h->free_huge_pages); 2554e5ff2159SAndi Kleen } 2555e5ff2159SAndi Kleen } 2556e5ff2159SAndi Kleen 25571da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 25586ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 25596ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 25601da177e4SLinus Torvalds { 25614415cc8dSChristoph Lameter int i; 25624415cc8dSChristoph Lameter 2563bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2564aa888a74SAndi Kleen return; 2565aa888a74SAndi Kleen 25666ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 25671da177e4SLinus Torvalds struct page *page, *next; 2568a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 2569a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 2570a5516438SAndi Kleen if (count >= h->nr_huge_pages) 25716b0c880dSAdam Litke return; 25721da177e4SLinus Torvalds if (PageHighMem(page)) 25731da177e4SLinus Torvalds continue; 25741da177e4SLinus Torvalds list_del(&page->lru); 2575e5ff2159SAndi Kleen update_and_free_page(h, page); 2576a5516438SAndi Kleen h->free_huge_pages--; 2577a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 25781da177e4SLinus Torvalds } 25791da177e4SLinus Torvalds } 25801da177e4SLinus Torvalds } 25811da177e4SLinus Torvalds #else 25826ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 25836ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 25841da177e4SLinus Torvalds { 25851da177e4SLinus Torvalds } 25861da177e4SLinus Torvalds #endif 25871da177e4SLinus Torvalds 258820a0307cSWu Fengguang /* 258920a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 259020a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 259120a0307cSWu Fengguang * Returns 1 if an adjustment was made. 259220a0307cSWu Fengguang */ 25936ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 25946ae11b27SLee Schermerhorn int delta) 259520a0307cSWu Fengguang { 2596b2261026SJoonsoo Kim int nr_nodes, node; 259720a0307cSWu Fengguang 259820a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 259920a0307cSWu Fengguang 2600e8c5c824SLee Schermerhorn if (delta < 0) { 2601b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 2602b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 2603b2261026SJoonsoo Kim goto found; 2604b2261026SJoonsoo Kim } 2605b2261026SJoonsoo Kim } else { 2606b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2607b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 2608b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 2609b2261026SJoonsoo Kim goto found; 2610e8c5c824SLee Schermerhorn } 26119a76db09SLee Schermerhorn } 2612b2261026SJoonsoo Kim return 0; 261320a0307cSWu Fengguang 2614b2261026SJoonsoo Kim found: 261520a0307cSWu Fengguang h->surplus_huge_pages += delta; 2616b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 2617b2261026SJoonsoo Kim return 1; 261820a0307cSWu Fengguang } 261920a0307cSWu Fengguang 2620a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 2621fd875dcaSMike Kravetz static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 26226ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 26231da177e4SLinus Torvalds { 26247893d1d5SAdam Litke unsigned long min_count, ret; 2625f60858f9SMike Kravetz NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); 2626f60858f9SMike Kravetz 2627f60858f9SMike Kravetz /* 2628f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 2629f60858f9SMike Kravetz * If we can not allocate the bit mask, do not attempt to allocate 2630f60858f9SMike Kravetz * the requested huge pages. 2631f60858f9SMike Kravetz */ 2632f60858f9SMike Kravetz if (node_alloc_noretry) 2633f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 2634f60858f9SMike Kravetz else 2635f60858f9SMike Kravetz return -ENOMEM; 26361da177e4SLinus Torvalds 26374eb0716eSAlexandre Ghiti spin_lock(&hugetlb_lock); 26384eb0716eSAlexandre Ghiti 26394eb0716eSAlexandre Ghiti /* 2640fd875dcaSMike Kravetz * Check for a node specific request. 2641fd875dcaSMike Kravetz * Changing node specific huge page count may require a corresponding 2642fd875dcaSMike Kravetz * change to the global count. In any case, the passed node mask 2643fd875dcaSMike Kravetz * (nodes_allowed) will restrict alloc/free to the specified node. 2644fd875dcaSMike Kravetz */ 2645fd875dcaSMike Kravetz if (nid != NUMA_NO_NODE) { 2646fd875dcaSMike Kravetz unsigned long old_count = count; 2647fd875dcaSMike Kravetz 2648fd875dcaSMike Kravetz count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 2649fd875dcaSMike Kravetz /* 2650fd875dcaSMike Kravetz * User may have specified a large count value which caused the 2651fd875dcaSMike Kravetz * above calculation to overflow. In this case, they wanted 2652fd875dcaSMike Kravetz * to allocate as many huge pages as possible. Set count to 2653fd875dcaSMike Kravetz * largest possible value to align with their intention. 2654fd875dcaSMike Kravetz */ 2655fd875dcaSMike Kravetz if (count < old_count) 2656fd875dcaSMike Kravetz count = ULONG_MAX; 2657fd875dcaSMike Kravetz } 2658fd875dcaSMike Kravetz 2659fd875dcaSMike Kravetz /* 26604eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 26614eb0716eSAlexandre Ghiti * page range allocation. 26624eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 26634eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 26644eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 26654eb0716eSAlexandre Ghiti */ 26664eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 26674eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 26684eb0716eSAlexandre Ghiti spin_unlock(&hugetlb_lock); 2669f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 26704eb0716eSAlexandre Ghiti return -EINVAL; 26714eb0716eSAlexandre Ghiti } 26724eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 26734eb0716eSAlexandre Ghiti } 2674aa888a74SAndi Kleen 26757893d1d5SAdam Litke /* 26767893d1d5SAdam Litke * Increase the pool size 26777893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 26787893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 2679d1c3fb1fSNishanth Aravamudan * 26800c397daeSMichal Hocko * We might race with alloc_surplus_huge_page() here and be unable 2681d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 2682d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 2683d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 2684d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 26857893d1d5SAdam Litke */ 2686a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 26876ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 26887893d1d5SAdam Litke break; 26897893d1d5SAdam Litke } 26907893d1d5SAdam Litke 2691a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 26927893d1d5SAdam Litke /* 26937893d1d5SAdam Litke * If this allocation races such that we no longer need the 26947893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 26957893d1d5SAdam Litke * and reducing the surplus. 26967893d1d5SAdam Litke */ 26977893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 2698649920c6SJia He 2699649920c6SJia He /* yield cpu to avoid soft lockup */ 2700649920c6SJia He cond_resched(); 2701649920c6SJia He 2702f60858f9SMike Kravetz ret = alloc_pool_huge_page(h, nodes_allowed, 2703f60858f9SMike Kravetz node_alloc_noretry); 27047893d1d5SAdam Litke spin_lock(&hugetlb_lock); 27057893d1d5SAdam Litke if (!ret) 27067893d1d5SAdam Litke goto out; 27077893d1d5SAdam Litke 2708536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 2709536240f2SMel Gorman if (signal_pending(current)) 2710536240f2SMel Gorman goto out; 27117893d1d5SAdam Litke } 27127893d1d5SAdam Litke 27137893d1d5SAdam Litke /* 27147893d1d5SAdam Litke * Decrease the pool size 27157893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 27167893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 27177893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 27187893d1d5SAdam Litke * to the desired size as pages become free. 2719d1c3fb1fSNishanth Aravamudan * 2720d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 2721d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 2722d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 27230c397daeSMichal Hocko * alloc_surplus_huge_page() is checking the global counter, 2724d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 2725d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 2726d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 27277893d1d5SAdam Litke */ 2728a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 27296b0c880dSAdam Litke min_count = max(count, min_count); 27306ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 2731a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 27326ae11b27SLee Schermerhorn if (!free_pool_huge_page(h, nodes_allowed, 0)) 27331da177e4SLinus Torvalds break; 273455f67141SMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 27351da177e4SLinus Torvalds } 2736a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 27376ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 27387893d1d5SAdam Litke break; 27397893d1d5SAdam Litke } 27407893d1d5SAdam Litke out: 27414eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 27421da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 27434eb0716eSAlexandre Ghiti 2744f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 2745f60858f9SMike Kravetz 27464eb0716eSAlexandre Ghiti return 0; 27471da177e4SLinus Torvalds } 27481da177e4SLinus Torvalds 2749a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 2750a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 2751a3437870SNishanth Aravamudan 2752a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 2753a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 2754a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 2755a3437870SNishanth Aravamudan 2756a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 2757a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 2758a3437870SNishanth Aravamudan 27599a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 27609a305230SLee Schermerhorn 27619a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 2762a3437870SNishanth Aravamudan { 2763a3437870SNishanth Aravamudan int i; 27649a305230SLee Schermerhorn 2765a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 27669a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 27679a305230SLee Schermerhorn if (nidp) 27689a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 2769a3437870SNishanth Aravamudan return &hstates[i]; 27709a305230SLee Schermerhorn } 27719a305230SLee Schermerhorn 27729a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 2773a3437870SNishanth Aravamudan } 2774a3437870SNishanth Aravamudan 277506808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 2776a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2777a3437870SNishanth Aravamudan { 27789a305230SLee Schermerhorn struct hstate *h; 27799a305230SLee Schermerhorn unsigned long nr_huge_pages; 27809a305230SLee Schermerhorn int nid; 27819a305230SLee Schermerhorn 27829a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 27839a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 27849a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 27859a305230SLee Schermerhorn else 27869a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 27879a305230SLee Schermerhorn 2788ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", nr_huge_pages); 2789a3437870SNishanth Aravamudan } 2790adbe8726SEric B Munson 2791238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 2792238d3c13SDavid Rientjes struct hstate *h, int nid, 2793238d3c13SDavid Rientjes unsigned long count, size_t len) 2794a3437870SNishanth Aravamudan { 2795a3437870SNishanth Aravamudan int err; 27962d0adf7eSOscar Salvador nodemask_t nodes_allowed, *n_mask; 2797a3437870SNishanth Aravamudan 27982d0adf7eSOscar Salvador if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 27992d0adf7eSOscar Salvador return -EINVAL; 2800adbe8726SEric B Munson 28019a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 28029a305230SLee Schermerhorn /* 28039a305230SLee Schermerhorn * global hstate attribute 28049a305230SLee Schermerhorn */ 28059a305230SLee Schermerhorn if (!(obey_mempolicy && 28062d0adf7eSOscar Salvador init_nodemask_of_mempolicy(&nodes_allowed))) 28072d0adf7eSOscar Salvador n_mask = &node_states[N_MEMORY]; 28082d0adf7eSOscar Salvador else 28092d0adf7eSOscar Salvador n_mask = &nodes_allowed; 28102d0adf7eSOscar Salvador } else { 28119a305230SLee Schermerhorn /* 2812fd875dcaSMike Kravetz * Node specific request. count adjustment happens in 2813fd875dcaSMike Kravetz * set_max_huge_pages() after acquiring hugetlb_lock. 28149a305230SLee Schermerhorn */ 28152d0adf7eSOscar Salvador init_nodemask_of_node(&nodes_allowed, nid); 28162d0adf7eSOscar Salvador n_mask = &nodes_allowed; 2817fd875dcaSMike Kravetz } 28189a305230SLee Schermerhorn 28192d0adf7eSOscar Salvador err = set_max_huge_pages(h, count, nid, n_mask); 282006808b08SLee Schermerhorn 28214eb0716eSAlexandre Ghiti return err ? err : len; 282206808b08SLee Schermerhorn } 282306808b08SLee Schermerhorn 2824238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 2825238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 2826238d3c13SDavid Rientjes size_t len) 2827238d3c13SDavid Rientjes { 2828238d3c13SDavid Rientjes struct hstate *h; 2829238d3c13SDavid Rientjes unsigned long count; 2830238d3c13SDavid Rientjes int nid; 2831238d3c13SDavid Rientjes int err; 2832238d3c13SDavid Rientjes 2833238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 2834238d3c13SDavid Rientjes if (err) 2835238d3c13SDavid Rientjes return err; 2836238d3c13SDavid Rientjes 2837238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 2838238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 2839238d3c13SDavid Rientjes } 2840238d3c13SDavid Rientjes 284106808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 284206808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 284306808b08SLee Schermerhorn { 284406808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 284506808b08SLee Schermerhorn } 284606808b08SLee Schermerhorn 284706808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 284806808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 284906808b08SLee Schermerhorn { 2850238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 2851a3437870SNishanth Aravamudan } 2852a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 2853a3437870SNishanth Aravamudan 285406808b08SLee Schermerhorn #ifdef CONFIG_NUMA 285506808b08SLee Schermerhorn 285606808b08SLee Schermerhorn /* 285706808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 285806808b08SLee Schermerhorn * huge page alloc/free. 285906808b08SLee Schermerhorn */ 286006808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 2861ae7a927dSJoe Perches struct kobj_attribute *attr, 2862ae7a927dSJoe Perches char *buf) 286306808b08SLee Schermerhorn { 286406808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 286506808b08SLee Schermerhorn } 286606808b08SLee Schermerhorn 286706808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 286806808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 286906808b08SLee Schermerhorn { 2870238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 287106808b08SLee Schermerhorn } 287206808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 287306808b08SLee Schermerhorn #endif 287406808b08SLee Schermerhorn 287506808b08SLee Schermerhorn 2876a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 2877a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2878a3437870SNishanth Aravamudan { 28799a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2880ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); 2881a3437870SNishanth Aravamudan } 2882adbe8726SEric B Munson 2883a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 2884a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 2885a3437870SNishanth Aravamudan { 2886a3437870SNishanth Aravamudan int err; 2887a3437870SNishanth Aravamudan unsigned long input; 28889a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2889a3437870SNishanth Aravamudan 2890bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2891adbe8726SEric B Munson return -EINVAL; 2892adbe8726SEric B Munson 28933dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 2894a3437870SNishanth Aravamudan if (err) 289573ae31e5SEric B Munson return err; 2896a3437870SNishanth Aravamudan 2897a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 2898a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 2899a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 2900a3437870SNishanth Aravamudan 2901a3437870SNishanth Aravamudan return count; 2902a3437870SNishanth Aravamudan } 2903a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 2904a3437870SNishanth Aravamudan 2905a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 2906a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2907a3437870SNishanth Aravamudan { 29089a305230SLee Schermerhorn struct hstate *h; 29099a305230SLee Schermerhorn unsigned long free_huge_pages; 29109a305230SLee Schermerhorn int nid; 29119a305230SLee Schermerhorn 29129a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 29139a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 29149a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 29159a305230SLee Schermerhorn else 29169a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 29179a305230SLee Schermerhorn 2918ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", free_huge_pages); 2919a3437870SNishanth Aravamudan } 2920a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 2921a3437870SNishanth Aravamudan 2922a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 2923a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2924a3437870SNishanth Aravamudan { 29259a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2926ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); 2927a3437870SNishanth Aravamudan } 2928a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 2929a3437870SNishanth Aravamudan 2930a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 2931a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2932a3437870SNishanth Aravamudan { 29339a305230SLee Schermerhorn struct hstate *h; 29349a305230SLee Schermerhorn unsigned long surplus_huge_pages; 29359a305230SLee Schermerhorn int nid; 29369a305230SLee Schermerhorn 29379a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 29389a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 29399a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 29409a305230SLee Schermerhorn else 29419a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 29429a305230SLee Schermerhorn 2943ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", surplus_huge_pages); 2944a3437870SNishanth Aravamudan } 2945a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 2946a3437870SNishanth Aravamudan 2947a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 2948a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 2949a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 2950a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 2951a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 2952a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 295306808b08SLee Schermerhorn #ifdef CONFIG_NUMA 295406808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 295506808b08SLee Schermerhorn #endif 2956a3437870SNishanth Aravamudan NULL, 2957a3437870SNishanth Aravamudan }; 2958a3437870SNishanth Aravamudan 295967e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 2960a3437870SNishanth Aravamudan .attrs = hstate_attrs, 2961a3437870SNishanth Aravamudan }; 2962a3437870SNishanth Aravamudan 2963094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 29649a305230SLee Schermerhorn struct kobject **hstate_kobjs, 296567e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 2966a3437870SNishanth Aravamudan { 2967a3437870SNishanth Aravamudan int retval; 2968972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 2969a3437870SNishanth Aravamudan 29709a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 29719a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 2972a3437870SNishanth Aravamudan return -ENOMEM; 2973a3437870SNishanth Aravamudan 29749a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 2975cc2205a6SMiaohe Lin if (retval) { 29769a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 2977cc2205a6SMiaohe Lin hstate_kobjs[hi] = NULL; 2978cc2205a6SMiaohe Lin } 2979a3437870SNishanth Aravamudan 2980a3437870SNishanth Aravamudan return retval; 2981a3437870SNishanth Aravamudan } 2982a3437870SNishanth Aravamudan 2983a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 2984a3437870SNishanth Aravamudan { 2985a3437870SNishanth Aravamudan struct hstate *h; 2986a3437870SNishanth Aravamudan int err; 2987a3437870SNishanth Aravamudan 2988a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 2989a3437870SNishanth Aravamudan if (!hugepages_kobj) 2990a3437870SNishanth Aravamudan return; 2991a3437870SNishanth Aravamudan 2992a3437870SNishanth Aravamudan for_each_hstate(h) { 29939a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 29949a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 2995a3437870SNishanth Aravamudan if (err) 2996282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s", h->name); 2997a3437870SNishanth Aravamudan } 2998a3437870SNishanth Aravamudan } 2999a3437870SNishanth Aravamudan 30009a305230SLee Schermerhorn #ifdef CONFIG_NUMA 30019a305230SLee Schermerhorn 30029a305230SLee Schermerhorn /* 30039a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 300410fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 300510fbcf4cSKay Sievers * index of a node device or _hstate == node id. 300610fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 30079a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 30089a305230SLee Schermerhorn */ 30099a305230SLee Schermerhorn struct node_hstate { 30109a305230SLee Schermerhorn struct kobject *hugepages_kobj; 30119a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 30129a305230SLee Schermerhorn }; 3013b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 30149a305230SLee Schermerhorn 30159a305230SLee Schermerhorn /* 301610fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 30179a305230SLee Schermerhorn */ 30189a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 30199a305230SLee Schermerhorn &nr_hugepages_attr.attr, 30209a305230SLee Schermerhorn &free_hugepages_attr.attr, 30219a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 30229a305230SLee Schermerhorn NULL, 30239a305230SLee Schermerhorn }; 30249a305230SLee Schermerhorn 302567e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 30269a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 30279a305230SLee Schermerhorn }; 30289a305230SLee Schermerhorn 30299a305230SLee Schermerhorn /* 303010fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 30319a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 30329a305230SLee Schermerhorn */ 30339a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 30349a305230SLee Schermerhorn { 30359a305230SLee Schermerhorn int nid; 30369a305230SLee Schermerhorn 30379a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 30389a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 30399a305230SLee Schermerhorn int i; 30409a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 30419a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 30429a305230SLee Schermerhorn if (nidp) 30439a305230SLee Schermerhorn *nidp = nid; 30449a305230SLee Schermerhorn return &hstates[i]; 30459a305230SLee Schermerhorn } 30469a305230SLee Schermerhorn } 30479a305230SLee Schermerhorn 30489a305230SLee Schermerhorn BUG(); 30499a305230SLee Schermerhorn return NULL; 30509a305230SLee Schermerhorn } 30519a305230SLee Schermerhorn 30529a305230SLee Schermerhorn /* 305310fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 30549a305230SLee Schermerhorn * No-op if no hstate attributes attached. 30559a305230SLee Schermerhorn */ 30563cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 30579a305230SLee Schermerhorn { 30589a305230SLee Schermerhorn struct hstate *h; 305910fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 30609a305230SLee Schermerhorn 30619a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 30629b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 30639a305230SLee Schermerhorn 3064972dc4deSAneesh Kumar K.V for_each_hstate(h) { 3065972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 3066972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 3067972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 3068972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 3069972dc4deSAneesh Kumar K.V } 30709a305230SLee Schermerhorn } 30719a305230SLee Schermerhorn 30729a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 30739a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 30749a305230SLee Schermerhorn } 30759a305230SLee Schermerhorn 30769a305230SLee Schermerhorn 30779a305230SLee Schermerhorn /* 307810fbcf4cSKay Sievers * Register hstate attributes for a single node device. 30799a305230SLee Schermerhorn * No-op if attributes already registered. 30809a305230SLee Schermerhorn */ 30813cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 30829a305230SLee Schermerhorn { 30839a305230SLee Schermerhorn struct hstate *h; 308410fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 30859a305230SLee Schermerhorn int err; 30869a305230SLee Schermerhorn 30879a305230SLee Schermerhorn if (nhs->hugepages_kobj) 30889a305230SLee Schermerhorn return; /* already allocated */ 30899a305230SLee Schermerhorn 30909a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 309110fbcf4cSKay Sievers &node->dev.kobj); 30929a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 30939a305230SLee Schermerhorn return; 30949a305230SLee Schermerhorn 30959a305230SLee Schermerhorn for_each_hstate(h) { 30969a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 30979a305230SLee Schermerhorn nhs->hstate_kobjs, 30989a305230SLee Schermerhorn &per_node_hstate_attr_group); 30999a305230SLee Schermerhorn if (err) { 3100282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s for node %d\n", 310110fbcf4cSKay Sievers h->name, node->dev.id); 31029a305230SLee Schermerhorn hugetlb_unregister_node(node); 31039a305230SLee Schermerhorn break; 31049a305230SLee Schermerhorn } 31059a305230SLee Schermerhorn } 31069a305230SLee Schermerhorn } 31079a305230SLee Schermerhorn 31089a305230SLee Schermerhorn /* 31099b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 311010fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 311110fbcf4cSKay Sievers * registered their associated device by this time. 31129a305230SLee Schermerhorn */ 31137d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 31149a305230SLee Schermerhorn { 31159a305230SLee Schermerhorn int nid; 31169a305230SLee Schermerhorn 31178cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 31188732794bSWen Congyang struct node *node = node_devices[nid]; 311910fbcf4cSKay Sievers if (node->dev.id == nid) 31209a305230SLee Schermerhorn hugetlb_register_node(node); 31219a305230SLee Schermerhorn } 31229a305230SLee Schermerhorn 31239a305230SLee Schermerhorn /* 312410fbcf4cSKay Sievers * Let the node device driver know we're here so it can 31259a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 31269a305230SLee Schermerhorn */ 31279a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 31289a305230SLee Schermerhorn hugetlb_unregister_node); 31299a305230SLee Schermerhorn } 31309a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 31319a305230SLee Schermerhorn 31329a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 31339a305230SLee Schermerhorn { 31349a305230SLee Schermerhorn BUG(); 31359a305230SLee Schermerhorn if (nidp) 31369a305230SLee Schermerhorn *nidp = -1; 31379a305230SLee Schermerhorn return NULL; 31389a305230SLee Schermerhorn } 31399a305230SLee Schermerhorn 31409a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 31419a305230SLee Schermerhorn 31429a305230SLee Schermerhorn #endif 31439a305230SLee Schermerhorn 3144a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 3145a3437870SNishanth Aravamudan { 31468382d914SDavidlohr Bueso int i; 31478382d914SDavidlohr Bueso 3148d6995da3SMike Kravetz BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < 3149d6995da3SMike Kravetz __NR_HPAGEFLAGS); 3150d6995da3SMike Kravetz 3151c2833a5bSMike Kravetz if (!hugepages_supported()) { 3152c2833a5bSMike Kravetz if (hugetlb_max_hstate || default_hstate_max_huge_pages) 3153c2833a5bSMike Kravetz pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); 31540ef89d25SBenjamin Herrenschmidt return 0; 3155d715cf80SLiam R. Howlett } 3156d715cf80SLiam R. Howlett 3157282f4214SMike Kravetz /* 3158282f4214SMike Kravetz * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some 3159282f4214SMike Kravetz * architectures depend on setup being done here. 3160282f4214SMike Kravetz */ 3161a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 3162282f4214SMike Kravetz if (!parsed_default_hugepagesz) { 3163282f4214SMike Kravetz /* 3164282f4214SMike Kravetz * If we did not parse a default huge page size, set 3165282f4214SMike Kravetz * default_hstate_idx to HPAGE_SIZE hstate. And, if the 3166282f4214SMike Kravetz * number of huge pages for this default size was implicitly 3167282f4214SMike Kravetz * specified, set that here as well. 3168282f4214SMike Kravetz * Note that the implicit setting will overwrite an explicit 3169282f4214SMike Kravetz * setting. A warning will be printed in this case. 3170282f4214SMike Kravetz */ 3171282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); 3172f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 3173282f4214SMike Kravetz if (default_hstate.max_huge_pages) { 3174282f4214SMike Kravetz char buf[32]; 3175282f4214SMike Kravetz 3176282f4214SMike Kravetz string_get_size(huge_page_size(&default_hstate), 3177282f4214SMike Kravetz 1, STRING_UNITS_2, buf, 32); 3178282f4214SMike Kravetz pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", 3179282f4214SMike Kravetz default_hstate.max_huge_pages, buf); 3180282f4214SMike Kravetz pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", 3181282f4214SMike Kravetz default_hstate_max_huge_pages); 3182282f4214SMike Kravetz } 3183282f4214SMike Kravetz default_hstate.max_huge_pages = 3184282f4214SMike Kravetz default_hstate_max_huge_pages; 3185282f4214SMike Kravetz } 3186f8b74815SVaishali Thakkar } 3187a3437870SNishanth Aravamudan 3188cf11e85fSRoman Gushchin hugetlb_cma_check(); 3189a3437870SNishanth Aravamudan hugetlb_init_hstates(); 3190aa888a74SAndi Kleen gather_bootmem_prealloc(); 3191a3437870SNishanth Aravamudan report_hugepages(); 3192a3437870SNishanth Aravamudan 3193a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 31949a305230SLee Schermerhorn hugetlb_register_all_nodes(); 31957179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 31969a305230SLee Schermerhorn 31978382d914SDavidlohr Bueso #ifdef CONFIG_SMP 31988382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 31998382d914SDavidlohr Bueso #else 32008382d914SDavidlohr Bueso num_fault_mutexes = 1; 32018382d914SDavidlohr Bueso #endif 3202c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 32036da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 32046da2ec56SKees Cook GFP_KERNEL); 3205c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 32068382d914SDavidlohr Bueso 32078382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 3208c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 3209a3437870SNishanth Aravamudan return 0; 3210a3437870SNishanth Aravamudan } 32113e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 3212a3437870SNishanth Aravamudan 3213ae94da89SMike Kravetz /* Overwritten by architectures with more huge page sizes */ 3214ae94da89SMike Kravetz bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) 32159fee021dSVaishali Thakkar { 3216ae94da89SMike Kravetz return size == HPAGE_SIZE; 32179fee021dSVaishali Thakkar } 32189fee021dSVaishali Thakkar 3219d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 3220a3437870SNishanth Aravamudan { 3221a3437870SNishanth Aravamudan struct hstate *h; 32228faa8b07SAndi Kleen unsigned long i; 32238faa8b07SAndi Kleen 3224a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 3225a3437870SNishanth Aravamudan return; 3226a3437870SNishanth Aravamudan } 322747d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 3228a3437870SNishanth Aravamudan BUG_ON(order == 0); 322947d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 3230a3437870SNishanth Aravamudan h->order = order; 3231aca78307SMiaohe Lin h->mask = ~(huge_page_size(h) - 1); 32328faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 32338faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 32340edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 323554f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 323654f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 3237a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 3238a3437870SNishanth Aravamudan huge_page_size(h)/1024); 32398faa8b07SAndi Kleen 3240a3437870SNishanth Aravamudan parsed_hstate = h; 3241a3437870SNishanth Aravamudan } 3242a3437870SNishanth Aravamudan 3243282f4214SMike Kravetz /* 3244282f4214SMike Kravetz * hugepages command line processing 3245282f4214SMike Kravetz * hugepages normally follows a valid hugepagsz or default_hugepagsz 3246282f4214SMike Kravetz * specification. If not, ignore the hugepages value. hugepages can also 3247282f4214SMike Kravetz * be the first huge page command line option in which case it implicitly 3248282f4214SMike Kravetz * specifies the number of huge pages for the default size. 3249282f4214SMike Kravetz */ 3250282f4214SMike Kravetz static int __init hugepages_setup(char *s) 3251a3437870SNishanth Aravamudan { 3252a3437870SNishanth Aravamudan unsigned long *mhp; 32538faa8b07SAndi Kleen static unsigned long *last_mhp; 3254a3437870SNishanth Aravamudan 32559fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 3256282f4214SMike Kravetz pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); 32579fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 3258282f4214SMike Kravetz return 0; 32599fee021dSVaishali Thakkar } 3260282f4214SMike Kravetz 3261a3437870SNishanth Aravamudan /* 3262282f4214SMike Kravetz * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter 3263282f4214SMike Kravetz * yet, so this hugepages= parameter goes to the "default hstate". 3264282f4214SMike Kravetz * Otherwise, it goes with the previously parsed hugepagesz or 3265282f4214SMike Kravetz * default_hugepagesz. 3266a3437870SNishanth Aravamudan */ 32679fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 3268a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 3269a3437870SNishanth Aravamudan else 3270a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 3271a3437870SNishanth Aravamudan 32728faa8b07SAndi Kleen if (mhp == last_mhp) { 3273282f4214SMike Kravetz pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); 3274282f4214SMike Kravetz return 0; 32758faa8b07SAndi Kleen } 32768faa8b07SAndi Kleen 3277a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 3278a3437870SNishanth Aravamudan *mhp = 0; 3279a3437870SNishanth Aravamudan 32808faa8b07SAndi Kleen /* 32818faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 32828faa8b07SAndi Kleen * But we need to allocate >= MAX_ORDER hstates here early to still 32838faa8b07SAndi Kleen * use the bootmem allocator. 32848faa8b07SAndi Kleen */ 328547d38344SAneesh Kumar K.V if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) 32868faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 32878faa8b07SAndi Kleen 32888faa8b07SAndi Kleen last_mhp = mhp; 32898faa8b07SAndi Kleen 3290a3437870SNishanth Aravamudan return 1; 3291a3437870SNishanth Aravamudan } 3292282f4214SMike Kravetz __setup("hugepages=", hugepages_setup); 3293e11bfbfcSNick Piggin 3294282f4214SMike Kravetz /* 3295282f4214SMike Kravetz * hugepagesz command line processing 3296282f4214SMike Kravetz * A specific huge page size can only be specified once with hugepagesz. 3297282f4214SMike Kravetz * hugepagesz is followed by hugepages on the command line. The global 3298282f4214SMike Kravetz * variable 'parsed_valid_hugepagesz' is used to determine if prior 3299282f4214SMike Kravetz * hugepagesz argument was valid. 3300282f4214SMike Kravetz */ 3301359f2544SMike Kravetz static int __init hugepagesz_setup(char *s) 3302e11bfbfcSNick Piggin { 3303359f2544SMike Kravetz unsigned long size; 3304282f4214SMike Kravetz struct hstate *h; 3305282f4214SMike Kravetz 3306282f4214SMike Kravetz parsed_valid_hugepagesz = false; 3307359f2544SMike Kravetz size = (unsigned long)memparse(s, NULL); 3308359f2544SMike Kravetz 3309359f2544SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 3310282f4214SMike Kravetz pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); 3311359f2544SMike Kravetz return 0; 3312359f2544SMike Kravetz } 3313359f2544SMike Kravetz 3314282f4214SMike Kravetz h = size_to_hstate(size); 3315282f4214SMike Kravetz if (h) { 3316282f4214SMike Kravetz /* 3317282f4214SMike Kravetz * hstate for this size already exists. This is normally 3318282f4214SMike Kravetz * an error, but is allowed if the existing hstate is the 3319282f4214SMike Kravetz * default hstate. More specifically, it is only allowed if 3320282f4214SMike Kravetz * the number of huge pages for the default hstate was not 3321282f4214SMike Kravetz * previously specified. 3322282f4214SMike Kravetz */ 3323282f4214SMike Kravetz if (!parsed_default_hugepagesz || h != &default_hstate || 3324282f4214SMike Kravetz default_hstate.max_huge_pages) { 3325282f4214SMike Kravetz pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); 332638237830SMike Kravetz return 0; 332738237830SMike Kravetz } 332838237830SMike Kravetz 3329282f4214SMike Kravetz /* 3330282f4214SMike Kravetz * No need to call hugetlb_add_hstate() as hstate already 3331282f4214SMike Kravetz * exists. But, do set parsed_hstate so that a following 3332282f4214SMike Kravetz * hugepages= parameter will be applied to this hstate. 3333282f4214SMike Kravetz */ 3334282f4214SMike Kravetz parsed_hstate = h; 3335282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3336e11bfbfcSNick Piggin return 1; 3337e11bfbfcSNick Piggin } 3338282f4214SMike Kravetz 3339359f2544SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 3340282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3341359f2544SMike Kravetz return 1; 3342359f2544SMike Kravetz } 3343359f2544SMike Kravetz __setup("hugepagesz=", hugepagesz_setup); 3344359f2544SMike Kravetz 3345282f4214SMike Kravetz /* 3346282f4214SMike Kravetz * default_hugepagesz command line input 3347282f4214SMike Kravetz * Only one instance of default_hugepagesz allowed on command line. 3348282f4214SMike Kravetz */ 3349ae94da89SMike Kravetz static int __init default_hugepagesz_setup(char *s) 3350e11bfbfcSNick Piggin { 3351ae94da89SMike Kravetz unsigned long size; 3352ae94da89SMike Kravetz 3353282f4214SMike Kravetz parsed_valid_hugepagesz = false; 3354282f4214SMike Kravetz if (parsed_default_hugepagesz) { 3355282f4214SMike Kravetz pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); 3356282f4214SMike Kravetz return 0; 3357282f4214SMike Kravetz } 3358282f4214SMike Kravetz 3359282f4214SMike Kravetz size = (unsigned long)memparse(s, NULL); 3360282f4214SMike Kravetz 3361282f4214SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 3362282f4214SMike Kravetz pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); 3363282f4214SMike Kravetz return 0; 3364282f4214SMike Kravetz } 3365282f4214SMike Kravetz 3366282f4214SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 3367282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3368282f4214SMike Kravetz parsed_default_hugepagesz = true; 3369282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(size)); 3370282f4214SMike Kravetz 3371282f4214SMike Kravetz /* 3372282f4214SMike Kravetz * The number of default huge pages (for this size) could have been 3373282f4214SMike Kravetz * specified as the first hugetlb parameter: hugepages=X. If so, 3374282f4214SMike Kravetz * then default_hstate_max_huge_pages is set. If the default huge 3375282f4214SMike Kravetz * page size is gigantic (>= MAX_ORDER), then the pages must be 3376282f4214SMike Kravetz * allocated here from bootmem allocator. 3377282f4214SMike Kravetz */ 3378282f4214SMike Kravetz if (default_hstate_max_huge_pages) { 3379282f4214SMike Kravetz default_hstate.max_huge_pages = default_hstate_max_huge_pages; 3380282f4214SMike Kravetz if (hstate_is_gigantic(&default_hstate)) 3381282f4214SMike Kravetz hugetlb_hstate_alloc_pages(&default_hstate); 3382282f4214SMike Kravetz default_hstate_max_huge_pages = 0; 3383282f4214SMike Kravetz } 3384282f4214SMike Kravetz 3385e11bfbfcSNick Piggin return 1; 3386e11bfbfcSNick Piggin } 3387ae94da89SMike Kravetz __setup("default_hugepagesz=", default_hugepagesz_setup); 3388a3437870SNishanth Aravamudan 33898ca39e68SMuchun Song static unsigned int allowed_mems_nr(struct hstate *h) 33908a213460SNishanth Aravamudan { 33918a213460SNishanth Aravamudan int node; 33928a213460SNishanth Aravamudan unsigned int nr = 0; 33938ca39e68SMuchun Song nodemask_t *mpol_allowed; 33948ca39e68SMuchun Song unsigned int *array = h->free_huge_pages_node; 33958ca39e68SMuchun Song gfp_t gfp_mask = htlb_alloc_mask(h); 33968a213460SNishanth Aravamudan 33978ca39e68SMuchun Song mpol_allowed = policy_nodemask_current(gfp_mask); 33988ca39e68SMuchun Song 33998ca39e68SMuchun Song for_each_node_mask(node, cpuset_current_mems_allowed) { 3400c93b0a99SJiapeng Zhong if (!mpol_allowed || node_isset(node, *mpol_allowed)) 34018a213460SNishanth Aravamudan nr += array[node]; 34028ca39e68SMuchun Song } 34038a213460SNishanth Aravamudan 34048a213460SNishanth Aravamudan return nr; 34058a213460SNishanth Aravamudan } 34068a213460SNishanth Aravamudan 34078a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 340817743798SMuchun Song static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, 340917743798SMuchun Song void *buffer, size_t *length, 341017743798SMuchun Song loff_t *ppos, unsigned long *out) 341117743798SMuchun Song { 341217743798SMuchun Song struct ctl_table dup_table; 341317743798SMuchun Song 341417743798SMuchun Song /* 341517743798SMuchun Song * In order to avoid races with __do_proc_doulongvec_minmax(), we 341617743798SMuchun Song * can duplicate the @table and alter the duplicate of it. 341717743798SMuchun Song */ 341817743798SMuchun Song dup_table = *table; 341917743798SMuchun Song dup_table.data = out; 342017743798SMuchun Song 342117743798SMuchun Song return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); 342217743798SMuchun Song } 342317743798SMuchun Song 342406808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 342506808b08SLee Schermerhorn struct ctl_table *table, int write, 342632927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 34271da177e4SLinus Torvalds { 3428e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 3429238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 343008d4a246SMichal Hocko int ret; 3431e5ff2159SAndi Kleen 3432457c1b27SNishanth Aravamudan if (!hugepages_supported()) 343386613628SJan Stancek return -EOPNOTSUPP; 3434457c1b27SNishanth Aravamudan 343517743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 343617743798SMuchun Song &tmp); 343708d4a246SMichal Hocko if (ret) 343808d4a246SMichal Hocko goto out; 3439e5ff2159SAndi Kleen 3440238d3c13SDavid Rientjes if (write) 3441238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 3442238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 344308d4a246SMichal Hocko out: 344408d4a246SMichal Hocko return ret; 34451da177e4SLinus Torvalds } 3446396faf03SMel Gorman 344706808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 344832927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 344906808b08SLee Schermerhorn { 345006808b08SLee Schermerhorn 345106808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 345206808b08SLee Schermerhorn buffer, length, ppos); 345306808b08SLee Schermerhorn } 345406808b08SLee Schermerhorn 345506808b08SLee Schermerhorn #ifdef CONFIG_NUMA 345606808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 345732927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 345806808b08SLee Schermerhorn { 345906808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 346006808b08SLee Schermerhorn buffer, length, ppos); 346106808b08SLee Schermerhorn } 346206808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 346306808b08SLee Schermerhorn 3464a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 346532927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 3466a3d0c6aaSNishanth Aravamudan { 3467a5516438SAndi Kleen struct hstate *h = &default_hstate; 3468e5ff2159SAndi Kleen unsigned long tmp; 346908d4a246SMichal Hocko int ret; 3470e5ff2159SAndi Kleen 3471457c1b27SNishanth Aravamudan if (!hugepages_supported()) 347286613628SJan Stancek return -EOPNOTSUPP; 3473457c1b27SNishanth Aravamudan 3474e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 3475e5ff2159SAndi Kleen 3476bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 3477adbe8726SEric B Munson return -EINVAL; 3478adbe8726SEric B Munson 347917743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 348017743798SMuchun Song &tmp); 348108d4a246SMichal Hocko if (ret) 348208d4a246SMichal Hocko goto out; 3483e5ff2159SAndi Kleen 3484e5ff2159SAndi Kleen if (write) { 3485064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 3486e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 3487a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 3488e5ff2159SAndi Kleen } 348908d4a246SMichal Hocko out: 349008d4a246SMichal Hocko return ret; 3491a3d0c6aaSNishanth Aravamudan } 3492a3d0c6aaSNishanth Aravamudan 34931da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 34941da177e4SLinus Torvalds 3495e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 34961da177e4SLinus Torvalds { 3497fcb2b0c5SRoman Gushchin struct hstate *h; 3498fcb2b0c5SRoman Gushchin unsigned long total = 0; 3499fcb2b0c5SRoman Gushchin 3500457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3501457c1b27SNishanth Aravamudan return; 3502fcb2b0c5SRoman Gushchin 3503fcb2b0c5SRoman Gushchin for_each_hstate(h) { 3504fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 3505fcb2b0c5SRoman Gushchin 3506aca78307SMiaohe Lin total += huge_page_size(h) * count; 3507fcb2b0c5SRoman Gushchin 3508fcb2b0c5SRoman Gushchin if (h == &default_hstate) 3509e1759c21SAlexey Dobriyan seq_printf(m, 35101da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 35111da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 3512b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 35137893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 35144f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 3515fcb2b0c5SRoman Gushchin count, 3516a5516438SAndi Kleen h->free_huge_pages, 3517a5516438SAndi Kleen h->resv_huge_pages, 3518a5516438SAndi Kleen h->surplus_huge_pages, 3519aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 3520fcb2b0c5SRoman Gushchin } 3521fcb2b0c5SRoman Gushchin 3522aca78307SMiaohe Lin seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); 35231da177e4SLinus Torvalds } 35241da177e4SLinus Torvalds 35257981593bSJoe Perches int hugetlb_report_node_meminfo(char *buf, int len, int nid) 35261da177e4SLinus Torvalds { 3527a5516438SAndi Kleen struct hstate *h = &default_hstate; 35287981593bSJoe Perches 3529457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3530457c1b27SNishanth Aravamudan return 0; 35317981593bSJoe Perches 35327981593bSJoe Perches return sysfs_emit_at(buf, len, 35331da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 3534a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 3535a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 3536a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 3537a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 3538a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 35391da177e4SLinus Torvalds } 35401da177e4SLinus Torvalds 3541949f7ec5SDavid Rientjes void hugetlb_show_meminfo(void) 3542949f7ec5SDavid Rientjes { 3543949f7ec5SDavid Rientjes struct hstate *h; 3544949f7ec5SDavid Rientjes int nid; 3545949f7ec5SDavid Rientjes 3546457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3547457c1b27SNishanth Aravamudan return; 3548457c1b27SNishanth Aravamudan 3549949f7ec5SDavid Rientjes for_each_node_state(nid, N_MEMORY) 3550949f7ec5SDavid Rientjes for_each_hstate(h) 3551949f7ec5SDavid Rientjes pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 3552949f7ec5SDavid Rientjes nid, 3553949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 3554949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 3555949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 3556aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 3557949f7ec5SDavid Rientjes } 3558949f7ec5SDavid Rientjes 35595d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 35605d317b2bSNaoya Horiguchi { 35615d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 35625d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 35635d317b2bSNaoya Horiguchi } 35645d317b2bSNaoya Horiguchi 35651da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 35661da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 35671da177e4SLinus Torvalds { 3568d0028588SWanpeng Li struct hstate *h; 3569d0028588SWanpeng Li unsigned long nr_total_pages = 0; 3570d0028588SWanpeng Li 3571d0028588SWanpeng Li for_each_hstate(h) 3572d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 3573d0028588SWanpeng Li return nr_total_pages; 35741da177e4SLinus Torvalds } 35751da177e4SLinus Torvalds 3576a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 3577fc1b8a73SMel Gorman { 3578fc1b8a73SMel Gorman int ret = -ENOMEM; 3579fc1b8a73SMel Gorman 35800aa7f354SMiaohe Lin if (!delta) 35810aa7f354SMiaohe Lin return 0; 35820aa7f354SMiaohe Lin 3583fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 3584fc1b8a73SMel Gorman /* 3585fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 3586fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 3587fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 3588fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 3589fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 3590fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 3591fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 3592fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 3593fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 3594fc1b8a73SMel Gorman * 3595fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 3596fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 3597fc1b8a73SMel Gorman * we fall back to check against current free page availability as 3598fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 3599fc1b8a73SMel Gorman * semantics that cpuset has. 36008ca39e68SMuchun Song * 36018ca39e68SMuchun Song * Apart from cpuset, we also have memory policy mechanism that 36028ca39e68SMuchun Song * also determines from which node the kernel will allocate memory 36038ca39e68SMuchun Song * in a NUMA system. So similar to cpuset, we also should consider 36048ca39e68SMuchun Song * the memory policy of the current task. Similar to the description 36058ca39e68SMuchun Song * above. 3606fc1b8a73SMel Gorman */ 3607fc1b8a73SMel Gorman if (delta > 0) { 3608a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 3609fc1b8a73SMel Gorman goto out; 3610fc1b8a73SMel Gorman 36118ca39e68SMuchun Song if (delta > allowed_mems_nr(h)) { 3612a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 3613fc1b8a73SMel Gorman goto out; 3614fc1b8a73SMel Gorman } 3615fc1b8a73SMel Gorman } 3616fc1b8a73SMel Gorman 3617fc1b8a73SMel Gorman ret = 0; 3618fc1b8a73SMel Gorman if (delta < 0) 3619a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 3620fc1b8a73SMel Gorman 3621fc1b8a73SMel Gorman out: 3622fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 3623fc1b8a73SMel Gorman return ret; 3624fc1b8a73SMel Gorman } 3625fc1b8a73SMel Gorman 362684afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 362784afd99bSAndy Whitcroft { 3628f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 362984afd99bSAndy Whitcroft 363084afd99bSAndy Whitcroft /* 363184afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 363284afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 363384afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 363425985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 363584afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 363684afd99bSAndy Whitcroft * new reference here without additional locking. 363784afd99bSAndy Whitcroft */ 36384e35f483SJoonsoo Kim if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 3639f522c3acSJoonsoo Kim kref_get(&resv->refs); 364084afd99bSAndy Whitcroft } 364184afd99bSAndy Whitcroft 3642a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 3643a1e78772SMel Gorman { 3644a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3645f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 364690481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 36474e35f483SJoonsoo Kim unsigned long reserve, start, end; 36481c5ecae3SMike Kravetz long gbl_reserve; 364984afd99bSAndy Whitcroft 36504e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 36514e35f483SJoonsoo Kim return; 36524e35f483SJoonsoo Kim 3653a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 3654a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 365584afd99bSAndy Whitcroft 36564e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 3657e9fe92aeSMina Almasry hugetlb_cgroup_uncharge_counter(resv, start, end); 36587251ff78SAdam Litke if (reserve) { 36591c5ecae3SMike Kravetz /* 36601c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 36611c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 36621c5ecae3SMike Kravetz */ 36631c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 36641c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 36657251ff78SAdam Litke } 3666e9fe92aeSMina Almasry 3667e9fe92aeSMina Almasry kref_put(&resv->refs, resv_map_release); 3668a1e78772SMel Gorman } 3669a1e78772SMel Gorman 367031383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 367131383c68SDan Williams { 367231383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 367331383c68SDan Williams return -EINVAL; 367431383c68SDan Williams return 0; 367531383c68SDan Williams } 367631383c68SDan Williams 367705ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 367805ea8860SDan Williams { 3679aca78307SMiaohe Lin return huge_page_size(hstate_vma(vma)); 368005ea8860SDan Williams } 368105ea8860SDan Williams 36821da177e4SLinus Torvalds /* 36831da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 36841da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 36856c26d310SMiaohe Lin * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get 36861da177e4SLinus Torvalds * this far. 36871da177e4SLinus Torvalds */ 3688b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 36891da177e4SLinus Torvalds { 36901da177e4SLinus Torvalds BUG(); 3691d0217ac0SNick Piggin return 0; 36921da177e4SLinus Torvalds } 36931da177e4SLinus Torvalds 3694eec3636aSJane Chu /* 3695eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 3696eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 3697eec3636aSJane Chu * This is because under System V memory model, mappings created via 3698eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 3699eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 3700eec3636aSJane Chu */ 3701f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 3702d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 370384afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 3704a1e78772SMel Gorman .close = hugetlb_vm_op_close, 3705dd3b614fSDmitry Safonov .may_split = hugetlb_vm_op_split, 370605ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 37071da177e4SLinus Torvalds }; 37081da177e4SLinus Torvalds 37091e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 37101e8f889bSDavid Gibson int writable) 371163551ae0SDavid Gibson { 371263551ae0SDavid Gibson pte_t entry; 371363551ae0SDavid Gibson 37141e8f889bSDavid Gibson if (writable) { 3715106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 3716106c992aSGerald Schaefer vma->vm_page_prot))); 371763551ae0SDavid Gibson } else { 3718106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 3719106c992aSGerald Schaefer vma->vm_page_prot)); 372063551ae0SDavid Gibson } 372163551ae0SDavid Gibson entry = pte_mkyoung(entry); 372263551ae0SDavid Gibson entry = pte_mkhuge(entry); 3723d9ed9faaSChris Metcalf entry = arch_make_huge_pte(entry, vma, page, writable); 372463551ae0SDavid Gibson 372563551ae0SDavid Gibson return entry; 372663551ae0SDavid Gibson } 372763551ae0SDavid Gibson 37281e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 37291e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 37301e8f889bSDavid Gibson { 37311e8f889bSDavid Gibson pte_t entry; 37321e8f889bSDavid Gibson 3733106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 373432f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 37354b3073e1SRussell King update_mmu_cache(vma, address, ptep); 37361e8f889bSDavid Gibson } 37371e8f889bSDavid Gibson 3738d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 37394a705fefSNaoya Horiguchi { 37404a705fefSNaoya Horiguchi swp_entry_t swp; 37414a705fefSNaoya Horiguchi 37424a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 3743d5ed7444SAneesh Kumar K.V return false; 37444a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 3745d79d176aSBaoquan He if (is_migration_entry(swp)) 3746d5ed7444SAneesh Kumar K.V return true; 37474a705fefSNaoya Horiguchi else 3748d5ed7444SAneesh Kumar K.V return false; 37494a705fefSNaoya Horiguchi } 37504a705fefSNaoya Horiguchi 37513e5c3600SBaoquan He static bool is_hugetlb_entry_hwpoisoned(pte_t pte) 37524a705fefSNaoya Horiguchi { 37534a705fefSNaoya Horiguchi swp_entry_t swp; 37544a705fefSNaoya Horiguchi 37554a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 37563e5c3600SBaoquan He return false; 37574a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 3758d79d176aSBaoquan He if (is_hwpoison_entry(swp)) 37593e5c3600SBaoquan He return true; 37604a705fefSNaoya Horiguchi else 37613e5c3600SBaoquan He return false; 37624a705fefSNaoya Horiguchi } 37631e8f889bSDavid Gibson 376463551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 376563551ae0SDavid Gibson struct vm_area_struct *vma) 376663551ae0SDavid Gibson { 37675e41540cSMike Kravetz pte_t *src_pte, *dst_pte, entry, dst_entry; 376863551ae0SDavid Gibson struct page *ptepage; 37691c59827dSHugh Dickins unsigned long addr; 37701e8f889bSDavid Gibson int cow; 3771a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3772a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 3773c0d0381aSMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 3774ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 3775e8569dd2SAndreas Sandberg int ret = 0; 37761e8f889bSDavid Gibson 37771e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 377863551ae0SDavid Gibson 3779ac46d4f3SJérôme Glisse if (cow) { 37807269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src, 37816f4f13e8SJérôme Glisse vma->vm_start, 3782ac46d4f3SJérôme Glisse vma->vm_end); 3783ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 3784c0d0381aSMike Kravetz } else { 3785c0d0381aSMike Kravetz /* 3786c0d0381aSMike Kravetz * For shared mappings i_mmap_rwsem must be held to call 3787c0d0381aSMike Kravetz * huge_pte_alloc, otherwise the returned ptep could go 3788c0d0381aSMike Kravetz * away if part of a shared pmd and another thread calls 3789c0d0381aSMike Kravetz * huge_pmd_unshare. 3790c0d0381aSMike Kravetz */ 3791c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 3792ac46d4f3SJérôme Glisse } 3793e8569dd2SAndreas Sandberg 3794a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 3795cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 37967868a208SPunit Agrawal src_pte = huge_pte_offset(src, addr, sz); 3797c74df32cSHugh Dickins if (!src_pte) 3798c74df32cSHugh Dickins continue; 3799a5516438SAndi Kleen dst_pte = huge_pte_alloc(dst, addr, sz); 3800e8569dd2SAndreas Sandberg if (!dst_pte) { 3801e8569dd2SAndreas Sandberg ret = -ENOMEM; 3802e8569dd2SAndreas Sandberg break; 3803e8569dd2SAndreas Sandberg } 3804c5c99429SLarry Woodman 38055e41540cSMike Kravetz /* 38065e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 38075e41540cSMike Kravetz * dst_pte == src_pte is the common case of src/dest sharing. 38085e41540cSMike Kravetz * 38095e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 38105e41540cSMike Kravetz * another vma. If dst_pte !none, this implies sharing. 38115e41540cSMike Kravetz * Check here before taking page table lock, and once again 38125e41540cSMike Kravetz * after taking the lock below. 38135e41540cSMike Kravetz */ 38145e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 38155e41540cSMike Kravetz if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) 3816c5c99429SLarry Woodman continue; 3817c5c99429SLarry Woodman 3818cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 3819cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 3820cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 38214a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 38225e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 38235e41540cSMike Kravetz if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { 38245e41540cSMike Kravetz /* 38255e41540cSMike Kravetz * Skip if src entry none. Also, skip in the 38265e41540cSMike Kravetz * unlikely case dst entry !none as this implies 38275e41540cSMike Kravetz * sharing with another vma. 38285e41540cSMike Kravetz */ 38294a705fefSNaoya Horiguchi ; 38304a705fefSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry) || 38314a705fefSNaoya Horiguchi is_hugetlb_entry_hwpoisoned(entry))) { 38324a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 38334a705fefSNaoya Horiguchi 38344a705fefSNaoya Horiguchi if (is_write_migration_entry(swp_entry) && cow) { 38354a705fefSNaoya Horiguchi /* 38364a705fefSNaoya Horiguchi * COW mappings require pages in both 38374a705fefSNaoya Horiguchi * parent and child to be set to read. 38384a705fefSNaoya Horiguchi */ 38394a705fefSNaoya Horiguchi make_migration_entry_read(&swp_entry); 38404a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 3841e5251fd4SPunit Agrawal set_huge_swap_pte_at(src, addr, src_pte, 3842e5251fd4SPunit Agrawal entry, sz); 38434a705fefSNaoya Horiguchi } 3844e5251fd4SPunit Agrawal set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); 38454a705fefSNaoya Horiguchi } else { 384634ee645eSJoerg Roedel if (cow) { 38470f10851eSJérôme Glisse /* 38480f10851eSJérôme Glisse * No need to notify as we are downgrading page 38490f10851eSJérôme Glisse * table protection not changing it to point 38500f10851eSJérôme Glisse * to a new page. 38510f10851eSJérôme Glisse * 3852ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 38530f10851eSJérôme Glisse */ 38547f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 385534ee645eSJoerg Roedel } 38560253d634SNaoya Horiguchi entry = huge_ptep_get(src_pte); 385763551ae0SDavid Gibson ptepage = pte_page(entry); 385863551ae0SDavid Gibson get_page(ptepage); 385953f9263bSKirill A. Shutemov page_dup_rmap(ptepage, true); 386063551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 38615d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), dst); 38621c59827dSHugh Dickins } 3863cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 3864cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 386563551ae0SDavid Gibson } 386663551ae0SDavid Gibson 3867e8569dd2SAndreas Sandberg if (cow) 3868ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 3869c0d0381aSMike Kravetz else 3870c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 3871e8569dd2SAndreas Sandberg 3872e8569dd2SAndreas Sandberg return ret; 387363551ae0SDavid Gibson } 387463551ae0SDavid Gibson 387524669e58SAneesh Kumar K.V void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 387624669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 387724669e58SAneesh Kumar K.V struct page *ref_page) 387863551ae0SDavid Gibson { 387963551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 388063551ae0SDavid Gibson unsigned long address; 3881c7546f8fSDavid Gibson pte_t *ptep; 388263551ae0SDavid Gibson pte_t pte; 3883cb900f41SKirill A. Shutemov spinlock_t *ptl; 388463551ae0SDavid Gibson struct page *page; 3885a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3886a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 3887ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 3888a5516438SAndi Kleen 388963551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 3890a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 3891a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 389263551ae0SDavid Gibson 389307e32661SAneesh Kumar K.V /* 389407e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 389507e32661SAneesh Kumar K.V * to huge page. 389607e32661SAneesh Kumar K.V */ 3897ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 389824669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 3899dff11abeSMike Kravetz 3900dff11abeSMike Kravetz /* 3901dff11abeSMike Kravetz * If sharing possible, alert mmu notifiers of worst case. 3902dff11abeSMike Kravetz */ 39036f4f13e8SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, 39046f4f13e8SJérôme Glisse end); 3905ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 3906ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 3907569f48b8SHillf Danton address = start; 3908569f48b8SHillf Danton for (; address < end; address += sz) { 39097868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, sz); 3910c7546f8fSDavid Gibson if (!ptep) 3911c7546f8fSDavid Gibson continue; 3912c7546f8fSDavid Gibson 3913cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 391434ae204fSMike Kravetz if (huge_pmd_unshare(mm, vma, &address, ptep)) { 391531d49da5SAneesh Kumar K.V spin_unlock(ptl); 3916dff11abeSMike Kravetz /* 3917dff11abeSMike Kravetz * We just unmapped a page of PMDs by clearing a PUD. 3918dff11abeSMike Kravetz * The caller's TLB flush range should cover this area. 3919dff11abeSMike Kravetz */ 392031d49da5SAneesh Kumar K.V continue; 392131d49da5SAneesh Kumar K.V } 392239dde65cSChen, Kenneth W 39236629326bSHillf Danton pte = huge_ptep_get(ptep); 392431d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 392531d49da5SAneesh Kumar K.V spin_unlock(ptl); 392631d49da5SAneesh Kumar K.V continue; 392731d49da5SAneesh Kumar K.V } 39286629326bSHillf Danton 39296629326bSHillf Danton /* 39309fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 39319fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 39326629326bSHillf Danton */ 39339fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 39349386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 393531d49da5SAneesh Kumar K.V spin_unlock(ptl); 393631d49da5SAneesh Kumar K.V continue; 39378c4894c6SNaoya Horiguchi } 39386629326bSHillf Danton 39396629326bSHillf Danton page = pte_page(pte); 394004f2cbe3SMel Gorman /* 394104f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 394204f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 394304f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 394404f2cbe3SMel Gorman */ 394504f2cbe3SMel Gorman if (ref_page) { 394631d49da5SAneesh Kumar K.V if (page != ref_page) { 394731d49da5SAneesh Kumar K.V spin_unlock(ptl); 394831d49da5SAneesh Kumar K.V continue; 394931d49da5SAneesh Kumar K.V } 395004f2cbe3SMel Gorman /* 395104f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 395204f2cbe3SMel Gorman * future faults in this VMA will fail rather than 395304f2cbe3SMel Gorman * looking like data was lost 395404f2cbe3SMel Gorman */ 395504f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 395604f2cbe3SMel Gorman } 395704f2cbe3SMel Gorman 3958c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 3959b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 3960106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 39616649a386SKen Chen set_page_dirty(page); 39629e81130bSHillf Danton 39635d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 3964d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 396531d49da5SAneesh Kumar K.V 3966cb900f41SKirill A. Shutemov spin_unlock(ptl); 3967e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 396824669e58SAneesh Kumar K.V /* 396931d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 397024669e58SAneesh Kumar K.V */ 397131d49da5SAneesh Kumar K.V if (ref_page) 397231d49da5SAneesh Kumar K.V break; 3973fe1668aeSChen, Kenneth W } 3974ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 397524669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 39761da177e4SLinus Torvalds } 397763551ae0SDavid Gibson 3978d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 3979d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 3980d833352aSMel Gorman unsigned long end, struct page *ref_page) 3981d833352aSMel Gorman { 3982d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 3983d833352aSMel Gorman 3984d833352aSMel Gorman /* 3985d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 3986d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 3987d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 3988d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 3989d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 3990c8c06efaSDavidlohr Bueso * is to clear it before releasing the i_mmap_rwsem. This works 3991d833352aSMel Gorman * because in the context this is called, the VMA is about to be 3992c8c06efaSDavidlohr Bueso * destroyed and the i_mmap_rwsem is held. 3993d833352aSMel Gorman */ 3994d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 3995d833352aSMel Gorman } 3996d833352aSMel Gorman 3997502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 399804f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 3999502717f4SChen, Kenneth W { 400024669e58SAneesh Kumar K.V struct mmu_gather tlb; 4001dff11abeSMike Kravetz 4002a72afd87SWill Deacon tlb_gather_mmu(&tlb, vma->vm_mm); 400324669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 4004ae8eba8bSWill Deacon tlb_finish_mmu(&tlb); 4005502717f4SChen, Kenneth W } 4006502717f4SChen, Kenneth W 400704f2cbe3SMel Gorman /* 400804f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 4009578b7725SZhiyuan Dai * mapping it owns the reserve page for. The intention is to unmap the page 401004f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 401104f2cbe3SMel Gorman * same region. 401204f2cbe3SMel Gorman */ 40132f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 40142a4b3dedSHarvey Harrison struct page *page, unsigned long address) 401504f2cbe3SMel Gorman { 40167526674dSAdam Litke struct hstate *h = hstate_vma(vma); 401704f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 401804f2cbe3SMel Gorman struct address_space *mapping; 401904f2cbe3SMel Gorman pgoff_t pgoff; 402004f2cbe3SMel Gorman 402104f2cbe3SMel Gorman /* 402204f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 402304f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 402404f2cbe3SMel Gorman */ 40257526674dSAdam Litke address = address & huge_page_mask(h); 402636e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 402736e4f20aSMichal Hocko vma->vm_pgoff; 402893c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 402904f2cbe3SMel Gorman 40304eb2b1dcSMel Gorman /* 40314eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 40324eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 40334eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 40344eb2b1dcSMel Gorman */ 403583cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 40366b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 403704f2cbe3SMel Gorman /* Do not unmap the current VMA */ 403804f2cbe3SMel Gorman if (iter_vma == vma) 403904f2cbe3SMel Gorman continue; 404004f2cbe3SMel Gorman 404104f2cbe3SMel Gorman /* 40422f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 40432f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 40442f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 40452f84a899SMel Gorman */ 40462f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 40472f84a899SMel Gorman continue; 40482f84a899SMel Gorman 40492f84a899SMel Gorman /* 405004f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 405104f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 405204f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 405304f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 405404f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 405504f2cbe3SMel Gorman */ 405604f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 405724669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 405824669e58SAneesh Kumar K.V address + huge_page_size(h), page); 405904f2cbe3SMel Gorman } 406083cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 406104f2cbe3SMel Gorman } 406204f2cbe3SMel Gorman 40630fe6e20bSNaoya Horiguchi /* 40640fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 4065ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 4066ef009b25SMichal Hocko * cannot race with other handlers or page migration. 4067ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 40680fe6e20bSNaoya Horiguchi */ 40692b740303SSouptick Joarder static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 4070974e6d66SHuang Ying unsigned long address, pte_t *ptep, 4071cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 40721e8f889bSDavid Gibson { 40733999f52eSAneesh Kumar K.V pte_t pte; 4074a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 40751e8f889bSDavid Gibson struct page *old_page, *new_page; 40762b740303SSouptick Joarder int outside_reserve = 0; 40772b740303SSouptick Joarder vm_fault_t ret = 0; 4078974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 4079ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 40801e8f889bSDavid Gibson 40813999f52eSAneesh Kumar K.V pte = huge_ptep_get(ptep); 40821e8f889bSDavid Gibson old_page = pte_page(pte); 40831e8f889bSDavid Gibson 408404f2cbe3SMel Gorman retry_avoidcopy: 40851e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 40861e8f889bSDavid Gibson * and just make the page writable */ 408737a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 40885a49973dSHugh Dickins page_move_anon_rmap(old_page, vma); 40895b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 409083c54070SNick Piggin return 0; 40911e8f889bSDavid Gibson } 40921e8f889bSDavid Gibson 409304f2cbe3SMel Gorman /* 409404f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 409504f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 409604f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 409704f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 409804f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 409904f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 410004f2cbe3SMel Gorman * of the full address range. 410104f2cbe3SMel Gorman */ 41025944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 410304f2cbe3SMel Gorman old_page != pagecache_page) 410404f2cbe3SMel Gorman outside_reserve = 1; 410504f2cbe3SMel Gorman 410609cbfeafSKirill A. Shutemov get_page(old_page); 4107b76c8cfbSLarry Woodman 4108ad4404a2SDavidlohr Bueso /* 4109ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 4110ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 4111ad4404a2SDavidlohr Bueso */ 4112cb900f41SKirill A. Shutemov spin_unlock(ptl); 41135b7a1d40SHuang Ying new_page = alloc_huge_page(vma, haddr, outside_reserve); 41141e8f889bSDavid Gibson 41152fc39cecSAdam Litke if (IS_ERR(new_page)) { 411604f2cbe3SMel Gorman /* 411704f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 411804f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 411904f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 412004f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 412104f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 412204f2cbe3SMel Gorman */ 412304f2cbe3SMel Gorman if (outside_reserve) { 4124e7dd91c4SMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 4125e7dd91c4SMike Kravetz pgoff_t idx; 4126e7dd91c4SMike Kravetz u32 hash; 4127e7dd91c4SMike Kravetz 412809cbfeafSKirill A. Shutemov put_page(old_page); 412904f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 4130e7dd91c4SMike Kravetz /* 4131e7dd91c4SMike Kravetz * Drop hugetlb_fault_mutex and i_mmap_rwsem before 4132e7dd91c4SMike Kravetz * unmapping. unmapping needs to hold i_mmap_rwsem 4133e7dd91c4SMike Kravetz * in write mode. Dropping i_mmap_rwsem in read mode 4134e7dd91c4SMike Kravetz * here is OK as COW mappings do not interact with 4135e7dd91c4SMike Kravetz * PMD sharing. 4136e7dd91c4SMike Kravetz * 4137e7dd91c4SMike Kravetz * Reacquire both after unmap operation. 4138e7dd91c4SMike Kravetz */ 4139e7dd91c4SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 4140e7dd91c4SMike Kravetz hash = hugetlb_fault_mutex_hash(mapping, idx); 4141e7dd91c4SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 4142e7dd91c4SMike Kravetz i_mmap_unlock_read(mapping); 4143e7dd91c4SMike Kravetz 41445b7a1d40SHuang Ying unmap_ref_private(mm, vma, old_page, haddr); 4145e7dd91c4SMike Kravetz 4146e7dd91c4SMike Kravetz i_mmap_lock_read(mapping); 4147e7dd91c4SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 4148cb900f41SKirill A. Shutemov spin_lock(ptl); 41495b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4150a9af0c5dSNaoya Horiguchi if (likely(ptep && 4151a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 415204f2cbe3SMel Gorman goto retry_avoidcopy; 4153a734bcc8SHillf Danton /* 4154cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 4155cb900f41SKirill A. Shutemov * lock, and our job is done. 4156a734bcc8SHillf Danton */ 4157a734bcc8SHillf Danton return 0; 415804f2cbe3SMel Gorman } 415904f2cbe3SMel Gorman 41602b740303SSouptick Joarder ret = vmf_error(PTR_ERR(new_page)); 4161ad4404a2SDavidlohr Bueso goto out_release_old; 41621e8f889bSDavid Gibson } 41631e8f889bSDavid Gibson 41640fe6e20bSNaoya Horiguchi /* 41650fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 41660fe6e20bSNaoya Horiguchi * anon_vma prepared. 41670fe6e20bSNaoya Horiguchi */ 416844e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 4169ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 4170ad4404a2SDavidlohr Bueso goto out_release_all; 417144e2aa93SDean Nelson } 41720fe6e20bSNaoya Horiguchi 4173974e6d66SHuang Ying copy_user_huge_page(new_page, old_page, address, vma, 417447ad8475SAndrea Arcangeli pages_per_huge_page(h)); 41750ed361deSNick Piggin __SetPageUptodate(new_page); 41761e8f889bSDavid Gibson 41777269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, 41786f4f13e8SJérôme Glisse haddr + huge_page_size(h)); 4179ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4180ad4404a2SDavidlohr Bueso 4181b76c8cfbSLarry Woodman /* 4182cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 4183b76c8cfbSLarry Woodman * before the page tables are altered 4184b76c8cfbSLarry Woodman */ 4185cb900f41SKirill A. Shutemov spin_lock(ptl); 41865b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4187a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 4188d6995da3SMike Kravetz ClearHPageRestoreReserve(new_page); 418907443a85SJoonsoo Kim 41901e8f889bSDavid Gibson /* Break COW */ 41915b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 4192ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, range.end); 41935b7a1d40SHuang Ying set_huge_pte_at(mm, haddr, ptep, 41941e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 4195d281ee61SKirill A. Shutemov page_remove_rmap(old_page, true); 41965b7a1d40SHuang Ying hugepage_add_new_anon_rmap(new_page, vma, haddr); 4197*8f251a3dSMike Kravetz SetHPageMigratable(new_page); 41981e8f889bSDavid Gibson /* Make the old page be freed below */ 41991e8f889bSDavid Gibson new_page = old_page; 42001e8f889bSDavid Gibson } 4201cb900f41SKirill A. Shutemov spin_unlock(ptl); 4202ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 4203ad4404a2SDavidlohr Bueso out_release_all: 42045b7a1d40SHuang Ying restore_reserve_on_error(h, vma, haddr, new_page); 420509cbfeafSKirill A. Shutemov put_page(new_page); 4206ad4404a2SDavidlohr Bueso out_release_old: 420709cbfeafSKirill A. Shutemov put_page(old_page); 42088312034fSJoonsoo Kim 4209ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 4210ad4404a2SDavidlohr Bueso return ret; 42111e8f889bSDavid Gibson } 42121e8f889bSDavid Gibson 421304f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 4214a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 4215a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 421604f2cbe3SMel Gorman { 421704f2cbe3SMel Gorman struct address_space *mapping; 4218e7c4b0bfSAndy Whitcroft pgoff_t idx; 421904f2cbe3SMel Gorman 422004f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 4221a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 422204f2cbe3SMel Gorman 422304f2cbe3SMel Gorman return find_lock_page(mapping, idx); 422404f2cbe3SMel Gorman } 422504f2cbe3SMel Gorman 42263ae77f43SHugh Dickins /* 42273ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 42283ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 42293ae77f43SHugh Dickins */ 42303ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 42312a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 42322a15efc9SHugh Dickins { 42332a15efc9SHugh Dickins struct address_space *mapping; 42342a15efc9SHugh Dickins pgoff_t idx; 42352a15efc9SHugh Dickins struct page *page; 42362a15efc9SHugh Dickins 42372a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 42382a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 42392a15efc9SHugh Dickins 42402a15efc9SHugh Dickins page = find_get_page(mapping, idx); 42412a15efc9SHugh Dickins if (page) 42422a15efc9SHugh Dickins put_page(page); 42432a15efc9SHugh Dickins return page != NULL; 42442a15efc9SHugh Dickins } 42452a15efc9SHugh Dickins 4246ab76ad54SMike Kravetz int huge_add_to_page_cache(struct page *page, struct address_space *mapping, 4247ab76ad54SMike Kravetz pgoff_t idx) 4248ab76ad54SMike Kravetz { 4249ab76ad54SMike Kravetz struct inode *inode = mapping->host; 4250ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 4251ab76ad54SMike Kravetz int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 4252ab76ad54SMike Kravetz 4253ab76ad54SMike Kravetz if (err) 4254ab76ad54SMike Kravetz return err; 4255d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 4256ab76ad54SMike Kravetz 425722146c3cSMike Kravetz /* 425822146c3cSMike Kravetz * set page dirty so that it will not be removed from cache/file 425922146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 426022146c3cSMike Kravetz */ 426122146c3cSMike Kravetz set_page_dirty(page); 426222146c3cSMike Kravetz 4263ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 4264ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 4265ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 4266ab76ad54SMike Kravetz return 0; 4267ab76ad54SMike Kravetz } 4268ab76ad54SMike Kravetz 42692b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 42702b740303SSouptick Joarder struct vm_area_struct *vma, 42718382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 4272788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 4273ac9b9c66SHugh Dickins { 4274a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 42752b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 4276409eb8c2SHillf Danton int anon_rmap = 0; 42774c887265SAdam Litke unsigned long size; 42784c887265SAdam Litke struct page *page; 42791e8f889bSDavid Gibson pte_t new_pte; 4280cb900f41SKirill A. Shutemov spinlock_t *ptl; 4281285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 4282cb6acd01SMike Kravetz bool new_page = false; 42834c887265SAdam Litke 428404f2cbe3SMel Gorman /* 428504f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 428604f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 428725985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 428804f2cbe3SMel Gorman */ 428904f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 4290910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 429104f2cbe3SMel Gorman current->pid); 429204f2cbe3SMel Gorman return ret; 429304f2cbe3SMel Gorman } 429404f2cbe3SMel Gorman 42954c887265SAdam Litke /* 429687bf91d3SMike Kravetz * We can not race with truncation due to holding i_mmap_rwsem. 429787bf91d3SMike Kravetz * i_size is modified when holding i_mmap_rwsem, so check here 429887bf91d3SMike Kravetz * once for faults beyond end of file. 42994c887265SAdam Litke */ 4300a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 4301ebed4bfcSHugh Dickins if (idx >= size) 4302ebed4bfcSHugh Dickins goto out; 43031a1aad8aSMike Kravetz 430487bf91d3SMike Kravetz retry: 430587bf91d3SMike Kravetz page = find_lock_page(mapping, idx); 430687bf91d3SMike Kravetz if (!page) { 43071a1aad8aSMike Kravetz /* 43081a1aad8aSMike Kravetz * Check for page in userfault range 43091a1aad8aSMike Kravetz */ 43101a1aad8aSMike Kravetz if (userfaultfd_missing(vma)) { 43111a1aad8aSMike Kravetz u32 hash; 43121a1aad8aSMike Kravetz struct vm_fault vmf = { 43131a1aad8aSMike Kravetz .vma = vma, 4314285b8dcaSHuang Ying .address = haddr, 43151a1aad8aSMike Kravetz .flags = flags, 43161a1aad8aSMike Kravetz /* 43171a1aad8aSMike Kravetz * Hard to debug if it ends up being 43181a1aad8aSMike Kravetz * used by a callee that assumes 43191a1aad8aSMike Kravetz * something about the other 43201a1aad8aSMike Kravetz * uninitialized fields... same as in 43211a1aad8aSMike Kravetz * memory.c 43221a1aad8aSMike Kravetz */ 43231a1aad8aSMike Kravetz }; 43241a1aad8aSMike Kravetz 43251a1aad8aSMike Kravetz /* 4326c0d0381aSMike Kravetz * hugetlb_fault_mutex and i_mmap_rwsem must be 4327c0d0381aSMike Kravetz * dropped before handling userfault. Reacquire 4328c0d0381aSMike Kravetz * after handling fault to make calling code simpler. 43291a1aad8aSMike Kravetz */ 4330188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 43311a1aad8aSMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 4332c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 43331a1aad8aSMike Kravetz ret = handle_userfault(&vmf, VM_UFFD_MISSING); 4334c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 43351a1aad8aSMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 43361a1aad8aSMike Kravetz goto out; 43371a1aad8aSMike Kravetz } 43381a1aad8aSMike Kravetz 4339285b8dcaSHuang Ying page = alloc_huge_page(vma, haddr, 0); 43402fc39cecSAdam Litke if (IS_ERR(page)) { 43414643d67eSMike Kravetz /* 43424643d67eSMike Kravetz * Returning error will result in faulting task being 43434643d67eSMike Kravetz * sent SIGBUS. The hugetlb fault mutex prevents two 43444643d67eSMike Kravetz * tasks from racing to fault in the same page which 43454643d67eSMike Kravetz * could result in false unable to allocate errors. 43464643d67eSMike Kravetz * Page migration does not take the fault mutex, but 43474643d67eSMike Kravetz * does a clear then write of pte's under page table 43484643d67eSMike Kravetz * lock. Page fault code could race with migration, 43494643d67eSMike Kravetz * notice the clear pte and try to allocate a page 43504643d67eSMike Kravetz * here. Before returning error, get ptl and make 43514643d67eSMike Kravetz * sure there really is no pte entry. 43524643d67eSMike Kravetz */ 43534643d67eSMike Kravetz ptl = huge_pte_lock(h, mm, ptep); 43544643d67eSMike Kravetz if (!huge_pte_none(huge_ptep_get(ptep))) { 43554643d67eSMike Kravetz ret = 0; 43564643d67eSMike Kravetz spin_unlock(ptl); 43574643d67eSMike Kravetz goto out; 43584643d67eSMike Kravetz } 43594643d67eSMike Kravetz spin_unlock(ptl); 43602b740303SSouptick Joarder ret = vmf_error(PTR_ERR(page)); 43616bda666aSChristoph Lameter goto out; 43626bda666aSChristoph Lameter } 436347ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 43640ed361deSNick Piggin __SetPageUptodate(page); 4365cb6acd01SMike Kravetz new_page = true; 4366ac9b9c66SHugh Dickins 4367f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 4368ab76ad54SMike Kravetz int err = huge_add_to_page_cache(page, mapping, idx); 43696bda666aSChristoph Lameter if (err) { 43706bda666aSChristoph Lameter put_page(page); 43716bda666aSChristoph Lameter if (err == -EEXIST) 43726bda666aSChristoph Lameter goto retry; 43736bda666aSChristoph Lameter goto out; 43746bda666aSChristoph Lameter } 437523be7468SMel Gorman } else { 43766bda666aSChristoph Lameter lock_page(page); 43770fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 43780fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 43790fe6e20bSNaoya Horiguchi goto backout_unlocked; 438023be7468SMel Gorman } 4381409eb8c2SHillf Danton anon_rmap = 1; 43820fe6e20bSNaoya Horiguchi } 43830fe6e20bSNaoya Horiguchi } else { 438457303d80SAndy Whitcroft /* 4385998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 4386998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 4387998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 4388fd6a03edSNaoya Horiguchi */ 4389fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 43900eb98f15SMiaohe Lin ret = VM_FAULT_HWPOISON_LARGE | 4391972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 4392fd6a03edSNaoya Horiguchi goto backout_unlocked; 43936bda666aSChristoph Lameter } 4394998b4382SNaoya Horiguchi } 43951e8f889bSDavid Gibson 439657303d80SAndy Whitcroft /* 439757303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 439857303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 439957303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 440057303d80SAndy Whitcroft * the spinlock. 440157303d80SAndy Whitcroft */ 44025e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 4403285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 44042b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 44052b26736cSAndy Whitcroft goto backout_unlocked; 44062b26736cSAndy Whitcroft } 44075e911373SMike Kravetz /* Just decrements count, does not deallocate */ 4408285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 44095e911373SMike Kravetz } 441057303d80SAndy Whitcroft 44118bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 441283c54070SNick Piggin ret = 0; 44137f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 44144c887265SAdam Litke goto backout; 44154c887265SAdam Litke 441607443a85SJoonsoo Kim if (anon_rmap) { 4417d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 4418285b8dcaSHuang Ying hugepage_add_new_anon_rmap(page, vma, haddr); 4419ac714904SChoi Gi-yong } else 442053f9263bSKirill A. Shutemov page_dup_rmap(page, true); 44211e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 44221e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 4423285b8dcaSHuang Ying set_huge_pte_at(mm, haddr, ptep, new_pte); 44241e8f889bSDavid Gibson 44255d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 4426788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 44271e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 4428974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); 44291e8f889bSDavid Gibson } 44301e8f889bSDavid Gibson 4431cb900f41SKirill A. Shutemov spin_unlock(ptl); 4432cb6acd01SMike Kravetz 4433cb6acd01SMike Kravetz /* 4434*8f251a3dSMike Kravetz * Only set HPageMigratable in newly allocated pages. Existing pages 4435*8f251a3dSMike Kravetz * found in the pagecache may not have HPageMigratableset if they have 4436*8f251a3dSMike Kravetz * been isolated for migration. 4437cb6acd01SMike Kravetz */ 4438cb6acd01SMike Kravetz if (new_page) 4439*8f251a3dSMike Kravetz SetHPageMigratable(page); 4440cb6acd01SMike Kravetz 44414c887265SAdam Litke unlock_page(page); 44424c887265SAdam Litke out: 4443ac9b9c66SHugh Dickins return ret; 44444c887265SAdam Litke 44454c887265SAdam Litke backout: 4446cb900f41SKirill A. Shutemov spin_unlock(ptl); 44472b26736cSAndy Whitcroft backout_unlocked: 44484c887265SAdam Litke unlock_page(page); 4449285b8dcaSHuang Ying restore_reserve_on_error(h, vma, haddr, page); 44504c887265SAdam Litke put_page(page); 44514c887265SAdam Litke goto out; 4452ac9b9c66SHugh Dickins } 4453ac9b9c66SHugh Dickins 44548382d914SDavidlohr Bueso #ifdef CONFIG_SMP 4455188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 44568382d914SDavidlohr Bueso { 44578382d914SDavidlohr Bueso unsigned long key[2]; 44588382d914SDavidlohr Bueso u32 hash; 44598382d914SDavidlohr Bueso 44608382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 44618382d914SDavidlohr Bueso key[1] = idx; 44628382d914SDavidlohr Bueso 446355254636SMike Kravetz hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); 44648382d914SDavidlohr Bueso 44658382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 44668382d914SDavidlohr Bueso } 44678382d914SDavidlohr Bueso #else 44688382d914SDavidlohr Bueso /* 44696c26d310SMiaohe Lin * For uniprocessor systems we always use a single mutex, so just 44708382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 44718382d914SDavidlohr Bueso */ 4472188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 44738382d914SDavidlohr Bueso { 44748382d914SDavidlohr Bueso return 0; 44758382d914SDavidlohr Bueso } 44768382d914SDavidlohr Bueso #endif 44778382d914SDavidlohr Bueso 44782b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 4479788c7df4SHugh Dickins unsigned long address, unsigned int flags) 448086e5216fSAdam Litke { 44818382d914SDavidlohr Bueso pte_t *ptep, entry; 4482cb900f41SKirill A. Shutemov spinlock_t *ptl; 44832b740303SSouptick Joarder vm_fault_t ret; 44848382d914SDavidlohr Bueso u32 hash; 44858382d914SDavidlohr Bueso pgoff_t idx; 44860fe6e20bSNaoya Horiguchi struct page *page = NULL; 448757303d80SAndy Whitcroft struct page *pagecache_page = NULL; 4488a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 44898382d914SDavidlohr Bueso struct address_space *mapping; 44900f792cf9SNaoya Horiguchi int need_wait_lock = 0; 4491285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 449286e5216fSAdam Litke 4493285b8dcaSHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4494fd6a03edSNaoya Horiguchi if (ptep) { 4495c0d0381aSMike Kravetz /* 4496c0d0381aSMike Kravetz * Since we hold no locks, ptep could be stale. That is 4497c0d0381aSMike Kravetz * OK as we are only making decisions based on content and 4498c0d0381aSMike Kravetz * not actually modifying content here. 4499c0d0381aSMike Kravetz */ 4500fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 4501290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 4502cb900f41SKirill A. Shutemov migration_entry_wait_huge(vma, mm, ptep); 4503290408d4SNaoya Horiguchi return 0; 4504290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 4505aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 4506972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 4507b43a9990SMike Kravetz } 45088382d914SDavidlohr Bueso 4509c0d0381aSMike Kravetz /* 4510c0d0381aSMike Kravetz * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold 451187bf91d3SMike Kravetz * until finished with ptep. This serves two purposes: 451287bf91d3SMike Kravetz * 1) It prevents huge_pmd_unshare from being called elsewhere 451387bf91d3SMike Kravetz * and making the ptep no longer valid. 451487bf91d3SMike Kravetz * 2) It synchronizes us with i_size modifications during truncation. 4515c0d0381aSMike Kravetz * 4516c0d0381aSMike Kravetz * ptep could have already be assigned via huge_pte_offset. That 4517c0d0381aSMike Kravetz * is OK, as huge_pte_alloc will return the same value unless 4518c0d0381aSMike Kravetz * something has changed. 4519c0d0381aSMike Kravetz */ 4520ddeaab32SMike Kravetz mapping = vma->vm_file->f_mapping; 4521c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 4522c0d0381aSMike Kravetz ptep = huge_pte_alloc(mm, haddr, huge_page_size(h)); 4523c0d0381aSMike Kravetz if (!ptep) { 4524c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 4525c0d0381aSMike Kravetz return VM_FAULT_OOM; 4526c0d0381aSMike Kravetz } 4527ddeaab32SMike Kravetz 45283935baa9SDavid Gibson /* 45293935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 45303935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 45313935baa9SDavid Gibson * the same page in the page cache. 45323935baa9SDavid Gibson */ 4533c0d0381aSMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 4534188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 4535c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 45368382d914SDavidlohr Bueso 45377f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 45387f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 45398382d914SDavidlohr Bueso ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 4540b4d1d99fSDavid Gibson goto out_mutex; 45413935baa9SDavid Gibson } 454286e5216fSAdam Litke 454383c54070SNick Piggin ret = 0; 45441e8f889bSDavid Gibson 454557303d80SAndy Whitcroft /* 45460f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 45470f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 45487c8de358SEthon Paul * an active hugepage in pagecache. This goto expects the 2nd page 45497c8de358SEthon Paul * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will 45507c8de358SEthon Paul * properly handle it. 45510f792cf9SNaoya Horiguchi */ 45520f792cf9SNaoya Horiguchi if (!pte_present(entry)) 45530f792cf9SNaoya Horiguchi goto out_mutex; 45540f792cf9SNaoya Horiguchi 45550f792cf9SNaoya Horiguchi /* 455657303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 455757303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 455857303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 455957303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 456057303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 456157303d80SAndy Whitcroft * consumed. 456257303d80SAndy Whitcroft */ 4563106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 4564285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 45652b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 4566b4d1d99fSDavid Gibson goto out_mutex; 45672b26736cSAndy Whitcroft } 45685e911373SMike Kravetz /* Just decrements count, does not deallocate */ 4569285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 457057303d80SAndy Whitcroft 4571f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 457257303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 4573285b8dcaSHuang Ying vma, haddr); 457457303d80SAndy Whitcroft } 457557303d80SAndy Whitcroft 45760f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 45770fe6e20bSNaoya Horiguchi 45781e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 4579b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 4580cb900f41SKirill A. Shutemov goto out_ptl; 4581b4d1d99fSDavid Gibson 45820f792cf9SNaoya Horiguchi /* 45830f792cf9SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 45840f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 45850f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 45860f792cf9SNaoya Horiguchi */ 45870f792cf9SNaoya Horiguchi page = pte_page(entry); 45880f792cf9SNaoya Horiguchi if (page != pagecache_page) 45890f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 45900f792cf9SNaoya Horiguchi need_wait_lock = 1; 45910f792cf9SNaoya Horiguchi goto out_ptl; 45920f792cf9SNaoya Horiguchi } 45930f792cf9SNaoya Horiguchi 45940f792cf9SNaoya Horiguchi get_page(page); 4595b4d1d99fSDavid Gibson 4596788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 4597106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 4598974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, 4599cb900f41SKirill A. Shutemov pagecache_page, ptl); 46000f792cf9SNaoya Horiguchi goto out_put_page; 4601b4d1d99fSDavid Gibson } 4602106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 4603b4d1d99fSDavid Gibson } 4604b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 4605285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 4606788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 4607285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 46080f792cf9SNaoya Horiguchi out_put_page: 46090f792cf9SNaoya Horiguchi if (page != pagecache_page) 46100f792cf9SNaoya Horiguchi unlock_page(page); 46110f792cf9SNaoya Horiguchi put_page(page); 4612cb900f41SKirill A. Shutemov out_ptl: 4613cb900f41SKirill A. Shutemov spin_unlock(ptl); 461457303d80SAndy Whitcroft 461557303d80SAndy Whitcroft if (pagecache_page) { 461657303d80SAndy Whitcroft unlock_page(pagecache_page); 461757303d80SAndy Whitcroft put_page(pagecache_page); 461857303d80SAndy Whitcroft } 4619b4d1d99fSDavid Gibson out_mutex: 4620c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 4621c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 46220f792cf9SNaoya Horiguchi /* 46230f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 46240f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 46250f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 46260f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 46270f792cf9SNaoya Horiguchi * here without taking refcount. 46280f792cf9SNaoya Horiguchi */ 46290f792cf9SNaoya Horiguchi if (need_wait_lock) 46300f792cf9SNaoya Horiguchi wait_on_page_locked(page); 46311e8f889bSDavid Gibson return ret; 463286e5216fSAdam Litke } 463386e5216fSAdam Litke 46348fb5debcSMike Kravetz /* 46358fb5debcSMike Kravetz * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with 46368fb5debcSMike Kravetz * modifications for huge pages. 46378fb5debcSMike Kravetz */ 46388fb5debcSMike Kravetz int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, 46398fb5debcSMike Kravetz pte_t *dst_pte, 46408fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 46418fb5debcSMike Kravetz unsigned long dst_addr, 46428fb5debcSMike Kravetz unsigned long src_addr, 46438fb5debcSMike Kravetz struct page **pagep) 46448fb5debcSMike Kravetz { 46451e392147SAndrea Arcangeli struct address_space *mapping; 46461e392147SAndrea Arcangeli pgoff_t idx; 46471e392147SAndrea Arcangeli unsigned long size; 46481c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 46498fb5debcSMike Kravetz struct hstate *h = hstate_vma(dst_vma); 46508fb5debcSMike Kravetz pte_t _dst_pte; 46518fb5debcSMike Kravetz spinlock_t *ptl; 46528fb5debcSMike Kravetz int ret; 46538fb5debcSMike Kravetz struct page *page; 46548fb5debcSMike Kravetz 46558fb5debcSMike Kravetz if (!*pagep) { 46568fb5debcSMike Kravetz ret = -ENOMEM; 46578fb5debcSMike Kravetz page = alloc_huge_page(dst_vma, dst_addr, 0); 46588fb5debcSMike Kravetz if (IS_ERR(page)) 46598fb5debcSMike Kravetz goto out; 46608fb5debcSMike Kravetz 46618fb5debcSMike Kravetz ret = copy_huge_page_from_user(page, 46628fb5debcSMike Kravetz (const void __user *) src_addr, 4663810a56b9SMike Kravetz pages_per_huge_page(h), false); 46648fb5debcSMike Kravetz 4665c1e8d7c6SMichel Lespinasse /* fallback to copy_from_user outside mmap_lock */ 46668fb5debcSMike Kravetz if (unlikely(ret)) { 46679e368259SAndrea Arcangeli ret = -ENOENT; 46688fb5debcSMike Kravetz *pagep = page; 46698fb5debcSMike Kravetz /* don't free the page */ 46708fb5debcSMike Kravetz goto out; 46718fb5debcSMike Kravetz } 46728fb5debcSMike Kravetz } else { 46738fb5debcSMike Kravetz page = *pagep; 46748fb5debcSMike Kravetz *pagep = NULL; 46758fb5debcSMike Kravetz } 46768fb5debcSMike Kravetz 46778fb5debcSMike Kravetz /* 46788fb5debcSMike Kravetz * The memory barrier inside __SetPageUptodate makes sure that 46798fb5debcSMike Kravetz * preceding stores to the page contents become visible before 46808fb5debcSMike Kravetz * the set_pte_at() write. 46818fb5debcSMike Kravetz */ 46828fb5debcSMike Kravetz __SetPageUptodate(page); 46838fb5debcSMike Kravetz 46841e392147SAndrea Arcangeli mapping = dst_vma->vm_file->f_mapping; 46851e392147SAndrea Arcangeli idx = vma_hugecache_offset(h, dst_vma, dst_addr); 46861e392147SAndrea Arcangeli 46871c9e8defSMike Kravetz /* 46881c9e8defSMike Kravetz * If shared, add to page cache 46891c9e8defSMike Kravetz */ 46901c9e8defSMike Kravetz if (vm_shared) { 46911e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 46921e392147SAndrea Arcangeli ret = -EFAULT; 46931e392147SAndrea Arcangeli if (idx >= size) 46941e392147SAndrea Arcangeli goto out_release_nounlock; 46951c9e8defSMike Kravetz 46961e392147SAndrea Arcangeli /* 46971e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 46981e392147SAndrea Arcangeli * huge_add_to_page_cache() below happens through the 46991e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 47001e392147SAndrea Arcangeli * the caller. 47011e392147SAndrea Arcangeli */ 47021c9e8defSMike Kravetz ret = huge_add_to_page_cache(page, mapping, idx); 47031c9e8defSMike Kravetz if (ret) 47041c9e8defSMike Kravetz goto out_release_nounlock; 47051c9e8defSMike Kravetz } 47061c9e8defSMike Kravetz 47078fb5debcSMike Kravetz ptl = huge_pte_lockptr(h, dst_mm, dst_pte); 47088fb5debcSMike Kravetz spin_lock(ptl); 47098fb5debcSMike Kravetz 47101e392147SAndrea Arcangeli /* 47111e392147SAndrea Arcangeli * Recheck the i_size after holding PT lock to make sure not 47121e392147SAndrea Arcangeli * to leave any page mapped (as page_mapped()) beyond the end 47131e392147SAndrea Arcangeli * of the i_size (remove_inode_hugepages() is strict about 47141e392147SAndrea Arcangeli * enforcing that). If we bail out here, we'll also leave a 47151e392147SAndrea Arcangeli * page in the radix tree in the vm_shared case beyond the end 47161e392147SAndrea Arcangeli * of the i_size, but remove_inode_hugepages() will take care 47171e392147SAndrea Arcangeli * of it as soon as we drop the hugetlb_fault_mutex_table. 47181e392147SAndrea Arcangeli */ 47191e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 47201e392147SAndrea Arcangeli ret = -EFAULT; 47211e392147SAndrea Arcangeli if (idx >= size) 47221e392147SAndrea Arcangeli goto out_release_unlock; 47231e392147SAndrea Arcangeli 47248fb5debcSMike Kravetz ret = -EEXIST; 47258fb5debcSMike Kravetz if (!huge_pte_none(huge_ptep_get(dst_pte))) 47268fb5debcSMike Kravetz goto out_release_unlock; 47278fb5debcSMike Kravetz 47281c9e8defSMike Kravetz if (vm_shared) { 47291c9e8defSMike Kravetz page_dup_rmap(page, true); 47301c9e8defSMike Kravetz } else { 4731d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 47328fb5debcSMike Kravetz hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); 47331c9e8defSMike Kravetz } 47348fb5debcSMike Kravetz 47358fb5debcSMike Kravetz _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); 47368fb5debcSMike Kravetz if (dst_vma->vm_flags & VM_WRITE) 47378fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 47388fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 47398fb5debcSMike Kravetz 47408fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 47418fb5debcSMike Kravetz 47428fb5debcSMike Kravetz (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, 47438fb5debcSMike Kravetz dst_vma->vm_flags & VM_WRITE); 47448fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 47458fb5debcSMike Kravetz 47468fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 47478fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 47488fb5debcSMike Kravetz 47498fb5debcSMike Kravetz spin_unlock(ptl); 4750*8f251a3dSMike Kravetz SetHPageMigratable(page); 47511c9e8defSMike Kravetz if (vm_shared) 47521c9e8defSMike Kravetz unlock_page(page); 47538fb5debcSMike Kravetz ret = 0; 47548fb5debcSMike Kravetz out: 47558fb5debcSMike Kravetz return ret; 47568fb5debcSMike Kravetz out_release_unlock: 47578fb5debcSMike Kravetz spin_unlock(ptl); 47581c9e8defSMike Kravetz if (vm_shared) 47591c9e8defSMike Kravetz unlock_page(page); 47605af10dfdSAndrea Arcangeli out_release_nounlock: 47618fb5debcSMike Kravetz put_page(page); 47628fb5debcSMike Kravetz goto out; 47638fb5debcSMike Kravetz } 47648fb5debcSMike Kravetz 476582e5d378SJoao Martins static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma, 476682e5d378SJoao Martins int refs, struct page **pages, 476782e5d378SJoao Martins struct vm_area_struct **vmas) 476882e5d378SJoao Martins { 476982e5d378SJoao Martins int nr; 477082e5d378SJoao Martins 477182e5d378SJoao Martins for (nr = 0; nr < refs; nr++) { 477282e5d378SJoao Martins if (likely(pages)) 477382e5d378SJoao Martins pages[nr] = mem_map_offset(page, nr); 477482e5d378SJoao Martins if (vmas) 477582e5d378SJoao Martins vmas[nr] = vma; 477682e5d378SJoao Martins } 477782e5d378SJoao Martins } 477882e5d378SJoao Martins 477928a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 478063551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 478128a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 47824f6da934SPeter Xu long i, unsigned int flags, int *locked) 478363551ae0SDavid Gibson { 4784d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 4785d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 478628a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 4787a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 47880fa5bc40SJoao Martins int err = -EFAULT, refs; 478963551ae0SDavid Gibson 479063551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 479163551ae0SDavid Gibson pte_t *pte; 4792cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 47932a15efc9SHugh Dickins int absent; 479463551ae0SDavid Gibson struct page *page; 479563551ae0SDavid Gibson 47964c887265SAdam Litke /* 479702057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 479802057967SDavid Rientjes * potentially allocating memory. 479902057967SDavid Rientjes */ 4800fa45f116SDavidlohr Bueso if (fatal_signal_pending(current)) { 480102057967SDavid Rientjes remainder = 0; 480202057967SDavid Rientjes break; 480302057967SDavid Rientjes } 480402057967SDavid Rientjes 480502057967SDavid Rientjes /* 48064c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 48072a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 48084c887265SAdam Litke * first, for the page indexing below to work. 4809cb900f41SKirill A. Shutemov * 4810cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 48114c887265SAdam Litke */ 48127868a208SPunit Agrawal pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), 48137868a208SPunit Agrawal huge_page_size(h)); 4814cb900f41SKirill A. Shutemov if (pte) 4815cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 48162a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 481763551ae0SDavid Gibson 48182a15efc9SHugh Dickins /* 48192a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 48203ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 48213ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 48223ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 48233ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 48242a15efc9SHugh Dickins */ 48253ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 48263ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 4827cb900f41SKirill A. Shutemov if (pte) 4828cb900f41SKirill A. Shutemov spin_unlock(ptl); 48292a15efc9SHugh Dickins remainder = 0; 48302a15efc9SHugh Dickins break; 48312a15efc9SHugh Dickins } 48322a15efc9SHugh Dickins 48339cc3a5bdSNaoya Horiguchi /* 48349cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 48359cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 48369cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 48379cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 48389cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 48399cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 48409cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 48419cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 48429cc3a5bdSNaoya Horiguchi */ 48439cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 4844106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 4845106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 48462b740303SSouptick Joarder vm_fault_t ret; 484787ffc118SAndrea Arcangeli unsigned int fault_flags = 0; 48484c887265SAdam Litke 4849cb900f41SKirill A. Shutemov if (pte) 4850cb900f41SKirill A. Shutemov spin_unlock(ptl); 485187ffc118SAndrea Arcangeli if (flags & FOLL_WRITE) 485287ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_WRITE; 48534f6da934SPeter Xu if (locked) 485471335f37SPeter Xu fault_flags |= FAULT_FLAG_ALLOW_RETRY | 485571335f37SPeter Xu FAULT_FLAG_KILLABLE; 485687ffc118SAndrea Arcangeli if (flags & FOLL_NOWAIT) 485787ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY | 485887ffc118SAndrea Arcangeli FAULT_FLAG_RETRY_NOWAIT; 485987ffc118SAndrea Arcangeli if (flags & FOLL_TRIED) { 48604426e945SPeter Xu /* 48614426e945SPeter Xu * Note: FAULT_FLAG_ALLOW_RETRY and 48624426e945SPeter Xu * FAULT_FLAG_TRIED can co-exist 48634426e945SPeter Xu */ 486487ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_TRIED; 486587ffc118SAndrea Arcangeli } 486687ffc118SAndrea Arcangeli ret = hugetlb_fault(mm, vma, vaddr, fault_flags); 486787ffc118SAndrea Arcangeli if (ret & VM_FAULT_ERROR) { 48682be7cfedSDaniel Jordan err = vm_fault_to_errno(ret, flags); 48691c59827dSHugh Dickins remainder = 0; 48701c59827dSHugh Dickins break; 48711c59827dSHugh Dickins } 487287ffc118SAndrea Arcangeli if (ret & VM_FAULT_RETRY) { 48734f6da934SPeter Xu if (locked && 48741ac25013SAndrea Arcangeli !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) 48754f6da934SPeter Xu *locked = 0; 487687ffc118SAndrea Arcangeli *nr_pages = 0; 487787ffc118SAndrea Arcangeli /* 487887ffc118SAndrea Arcangeli * VM_FAULT_RETRY must not return an 487987ffc118SAndrea Arcangeli * error, it will return zero 488087ffc118SAndrea Arcangeli * instead. 488187ffc118SAndrea Arcangeli * 488287ffc118SAndrea Arcangeli * No need to update "position" as the 488387ffc118SAndrea Arcangeli * caller will not check it after 488487ffc118SAndrea Arcangeli * *nr_pages is set to 0. 488587ffc118SAndrea Arcangeli */ 488687ffc118SAndrea Arcangeli return i; 488787ffc118SAndrea Arcangeli } 488887ffc118SAndrea Arcangeli continue; 488987ffc118SAndrea Arcangeli } 489063551ae0SDavid Gibson 4891a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 48927f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 48938fde12caSLinus Torvalds 48948fde12caSLinus Torvalds /* 4895acbfb087SZhigang Lu * If subpage information not requested, update counters 4896acbfb087SZhigang Lu * and skip the same_page loop below. 4897acbfb087SZhigang Lu */ 4898acbfb087SZhigang Lu if (!pages && !vmas && !pfn_offset && 4899acbfb087SZhigang Lu (vaddr + huge_page_size(h) < vma->vm_end) && 4900acbfb087SZhigang Lu (remainder >= pages_per_huge_page(h))) { 4901acbfb087SZhigang Lu vaddr += huge_page_size(h); 4902acbfb087SZhigang Lu remainder -= pages_per_huge_page(h); 4903acbfb087SZhigang Lu i += pages_per_huge_page(h); 4904acbfb087SZhigang Lu spin_unlock(ptl); 4905acbfb087SZhigang Lu continue; 4906acbfb087SZhigang Lu } 4907acbfb087SZhigang Lu 490882e5d378SJoao Martins refs = min3(pages_per_huge_page(h) - pfn_offset, 490982e5d378SJoao Martins (vma->vm_end - vaddr) >> PAGE_SHIFT, remainder); 49100fa5bc40SJoao Martins 491182e5d378SJoao Martins if (pages || vmas) 491282e5d378SJoao Martins record_subpages_vmas(mem_map_offset(page, pfn_offset), 491382e5d378SJoao Martins vma, refs, 491482e5d378SJoao Martins likely(pages) ? pages + i : NULL, 491582e5d378SJoao Martins vmas ? vmas + i : NULL); 491663551ae0SDavid Gibson 491782e5d378SJoao Martins if (pages) { 49180fa5bc40SJoao Martins /* 49190fa5bc40SJoao Martins * try_grab_compound_head() should always succeed here, 49200fa5bc40SJoao Martins * because: a) we hold the ptl lock, and b) we've just 49210fa5bc40SJoao Martins * checked that the huge page is present in the page 49220fa5bc40SJoao Martins * tables. If the huge page is present, then the tail 49230fa5bc40SJoao Martins * pages must also be present. The ptl prevents the 49240fa5bc40SJoao Martins * head page and tail pages from being rearranged in 49250fa5bc40SJoao Martins * any way. So this page must be available at this 49260fa5bc40SJoao Martins * point, unless the page refcount overflowed: 49270fa5bc40SJoao Martins */ 492882e5d378SJoao Martins if (WARN_ON_ONCE(!try_grab_compound_head(pages[i], 49290fa5bc40SJoao Martins refs, 49300fa5bc40SJoao Martins flags))) { 49310fa5bc40SJoao Martins spin_unlock(ptl); 49320fa5bc40SJoao Martins remainder = 0; 49330fa5bc40SJoao Martins err = -ENOMEM; 49340fa5bc40SJoao Martins break; 49350fa5bc40SJoao Martins } 4936d5d4b0aaSChen, Kenneth W } 493782e5d378SJoao Martins 493882e5d378SJoao Martins vaddr += (refs << PAGE_SHIFT); 493982e5d378SJoao Martins remainder -= refs; 494082e5d378SJoao Martins i += refs; 494182e5d378SJoao Martins 4942cb900f41SKirill A. Shutemov spin_unlock(ptl); 494363551ae0SDavid Gibson } 494428a35716SMichel Lespinasse *nr_pages = remainder; 494587ffc118SAndrea Arcangeli /* 494687ffc118SAndrea Arcangeli * setting position is actually required only if remainder is 494787ffc118SAndrea Arcangeli * not zero but it's faster not to add a "if (remainder)" 494887ffc118SAndrea Arcangeli * branch. 494987ffc118SAndrea Arcangeli */ 495063551ae0SDavid Gibson *position = vaddr; 495163551ae0SDavid Gibson 49522be7cfedSDaniel Jordan return i ? i : err; 495363551ae0SDavid Gibson } 49548f860591SZhang, Yanmin 49555491ae7bSAneesh Kumar K.V #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE 49565491ae7bSAneesh Kumar K.V /* 49575491ae7bSAneesh Kumar K.V * ARCHes with special requirements for evicting HUGETLB backing TLB entries can 49585491ae7bSAneesh Kumar K.V * implement this. 49595491ae7bSAneesh Kumar K.V */ 49605491ae7bSAneesh Kumar K.V #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) 49615491ae7bSAneesh Kumar K.V #endif 49625491ae7bSAneesh Kumar K.V 49637da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 49648f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 49658f860591SZhang, Yanmin { 49668f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 49678f860591SZhang, Yanmin unsigned long start = address; 49688f860591SZhang, Yanmin pte_t *ptep; 49698f860591SZhang, Yanmin pte_t pte; 4970a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 49717da4d641SPeter Zijlstra unsigned long pages = 0; 4972dff11abeSMike Kravetz bool shared_pmd = false; 4973ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 4974dff11abeSMike Kravetz 4975dff11abeSMike Kravetz /* 4976dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 4977ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 4978dff11abeSMike Kravetz * range if PMD sharing is possible. 4979dff11abeSMike Kravetz */ 49807269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 49817269f999SJérôme Glisse 0, vma, mm, start, end); 4982ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 49838f860591SZhang, Yanmin 49848f860591SZhang, Yanmin BUG_ON(address >= end); 4985ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 49868f860591SZhang, Yanmin 4987ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 498883cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 4989a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 4990cb900f41SKirill A. Shutemov spinlock_t *ptl; 49917868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, huge_page_size(h)); 49928f860591SZhang, Yanmin if (!ptep) 49938f860591SZhang, Yanmin continue; 4994cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 499534ae204fSMike Kravetz if (huge_pmd_unshare(mm, vma, &address, ptep)) { 49967da4d641SPeter Zijlstra pages++; 4997cb900f41SKirill A. Shutemov spin_unlock(ptl); 4998dff11abeSMike Kravetz shared_pmd = true; 499939dde65cSChen, Kenneth W continue; 50007da4d641SPeter Zijlstra } 5001a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 5002a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 5003a8bda28dSNaoya Horiguchi spin_unlock(ptl); 5004a8bda28dSNaoya Horiguchi continue; 5005a8bda28dSNaoya Horiguchi } 5006a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 5007a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 5008a8bda28dSNaoya Horiguchi 5009a8bda28dSNaoya Horiguchi if (is_write_migration_entry(entry)) { 5010a8bda28dSNaoya Horiguchi pte_t newpte; 5011a8bda28dSNaoya Horiguchi 5012a8bda28dSNaoya Horiguchi make_migration_entry_read(&entry); 5013a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 5014e5251fd4SPunit Agrawal set_huge_swap_pte_at(mm, address, ptep, 5015e5251fd4SPunit Agrawal newpte, huge_page_size(h)); 5016a8bda28dSNaoya Horiguchi pages++; 5017a8bda28dSNaoya Horiguchi } 5018a8bda28dSNaoya Horiguchi spin_unlock(ptl); 5019a8bda28dSNaoya Horiguchi continue; 5020a8bda28dSNaoya Horiguchi } 5021a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 5022023bdd00SAneesh Kumar K.V pte_t old_pte; 5023023bdd00SAneesh Kumar K.V 5024023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 5025023bdd00SAneesh Kumar K.V pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); 5026be7517d6STony Lu pte = arch_make_huge_pte(pte, vma, NULL, 0); 5027023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 50287da4d641SPeter Zijlstra pages++; 50298f860591SZhang, Yanmin } 5030cb900f41SKirill A. Shutemov spin_unlock(ptl); 50318f860591SZhang, Yanmin } 5032d833352aSMel Gorman /* 5033c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 5034d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 5035c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 5036dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 5037dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 5038d833352aSMel Gorman */ 5039dff11abeSMike Kravetz if (shared_pmd) 5040ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 5041dff11abeSMike Kravetz else 50425491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 50430f10851eSJérôme Glisse /* 50440f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 50450f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 50460f10851eSJérôme Glisse * 5047ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 50480f10851eSJérôme Glisse */ 504983cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 5050ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 50517da4d641SPeter Zijlstra 50527da4d641SPeter Zijlstra return pages << h->order; 50538f860591SZhang, Yanmin } 50548f860591SZhang, Yanmin 5055a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 5056a1e78772SMel Gorman long from, long to, 50575a6fe125SMel Gorman struct vm_area_struct *vma, 5058ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 5059e4e574b7SAdam Litke { 50600db9d74eSMina Almasry long ret, chg, add = -1; 5061a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 506290481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 50639119a41eSJoonsoo Kim struct resv_map *resv_map; 5064075a61d0SMina Almasry struct hugetlb_cgroup *h_cg = NULL; 50650db9d74eSMina Almasry long gbl_reserve, regions_needed = 0; 5066e4e574b7SAdam Litke 506763489f8eSMike Kravetz /* This should never happen */ 506863489f8eSMike Kravetz if (from > to) { 506963489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 507063489f8eSMike Kravetz return -EINVAL; 507163489f8eSMike Kravetz } 507263489f8eSMike Kravetz 5073a1e78772SMel Gorman /* 507417c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 507517c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 507690481622SDavid Gibson * without using reserves 507717c9d12eSMel Gorman */ 5078ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 507917c9d12eSMel Gorman return 0; 508017c9d12eSMel Gorman 508117c9d12eSMel Gorman /* 5082a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 5083a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 5084a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 5085a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 5086a1e78772SMel Gorman */ 50879119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 5088f27a5136SMike Kravetz /* 5089f27a5136SMike Kravetz * resv_map can not be NULL as hugetlb_reserve_pages is only 5090f27a5136SMike Kravetz * called for inodes for which resv_maps were created (see 5091f27a5136SMike Kravetz * hugetlbfs_get_inode). 5092f27a5136SMike Kravetz */ 50934e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 50949119a41eSJoonsoo Kim 50950db9d74eSMina Almasry chg = region_chg(resv_map, from, to, ®ions_needed); 50969119a41eSJoonsoo Kim 50979119a41eSJoonsoo Kim } else { 5098e9fe92aeSMina Almasry /* Private mapping. */ 50999119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 51005a6fe125SMel Gorman if (!resv_map) 51015a6fe125SMel Gorman return -ENOMEM; 51025a6fe125SMel Gorman 510317c9d12eSMel Gorman chg = to - from; 510417c9d12eSMel Gorman 51055a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 51065a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 51075a6fe125SMel Gorman } 51085a6fe125SMel Gorman 5109c50ac050SDave Hansen if (chg < 0) { 5110c50ac050SDave Hansen ret = chg; 5111c50ac050SDave Hansen goto out_err; 5112c50ac050SDave Hansen } 511317c9d12eSMel Gorman 5114075a61d0SMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 5115075a61d0SMina Almasry hstate_index(h), chg * pages_per_huge_page(h), &h_cg); 5116075a61d0SMina Almasry 5117075a61d0SMina Almasry if (ret < 0) { 5118075a61d0SMina Almasry ret = -ENOMEM; 5119075a61d0SMina Almasry goto out_err; 5120075a61d0SMina Almasry } 5121075a61d0SMina Almasry 5122075a61d0SMina Almasry if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { 5123075a61d0SMina Almasry /* For private mappings, the hugetlb_cgroup uncharge info hangs 5124075a61d0SMina Almasry * of the resv_map. 5125075a61d0SMina Almasry */ 5126075a61d0SMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); 5127075a61d0SMina Almasry } 5128075a61d0SMina Almasry 51291c5ecae3SMike Kravetz /* 51301c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 51311c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 51321c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 51331c5ecae3SMike Kravetz */ 51341c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 51351c5ecae3SMike Kravetz if (gbl_reserve < 0) { 5136c50ac050SDave Hansen ret = -ENOSPC; 5137075a61d0SMina Almasry goto out_uncharge_cgroup; 5138c50ac050SDave Hansen } 513917c9d12eSMel Gorman 514017c9d12eSMel Gorman /* 514117c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 514290481622SDavid Gibson * Hand the pages back to the subpool if there are not 514317c9d12eSMel Gorman */ 51441c5ecae3SMike Kravetz ret = hugetlb_acct_memory(h, gbl_reserve); 514517c9d12eSMel Gorman if (ret < 0) { 5146075a61d0SMina Almasry goto out_put_pages; 514717c9d12eSMel Gorman } 514817c9d12eSMel Gorman 514917c9d12eSMel Gorman /* 515017c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 515117c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 515217c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 515317c9d12eSMel Gorman * the reservation was and the page cache tells how much of 515417c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 515517c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 515617c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 515717c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 515817c9d12eSMel Gorman * else has to be done for private mappings here 515917c9d12eSMel Gorman */ 516033039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 5161075a61d0SMina Almasry add = region_add(resv_map, from, to, regions_needed, h, h_cg); 516233039678SMike Kravetz 51630db9d74eSMina Almasry if (unlikely(add < 0)) { 51640db9d74eSMina Almasry hugetlb_acct_memory(h, -gbl_reserve); 51657fc2513aSDan Carpenter ret = add; 5166075a61d0SMina Almasry goto out_put_pages; 51670db9d74eSMina Almasry } else if (unlikely(chg > add)) { 516833039678SMike Kravetz /* 516933039678SMike Kravetz * pages in this range were added to the reserve 517033039678SMike Kravetz * map between region_chg and region_add. This 517133039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 517233039678SMike Kravetz * the subpool and reserve counts modified above 517333039678SMike Kravetz * based on the difference. 517433039678SMike Kravetz */ 517533039678SMike Kravetz long rsv_adjust; 517633039678SMike Kravetz 5177075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd( 5178075a61d0SMina Almasry hstate_index(h), 5179075a61d0SMina Almasry (chg - add) * pages_per_huge_page(h), h_cg); 5180075a61d0SMina Almasry 518133039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 518233039678SMike Kravetz chg - add); 518333039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 518433039678SMike Kravetz } 518533039678SMike Kravetz } 5186a43a8c39SChen, Kenneth W return 0; 5187075a61d0SMina Almasry out_put_pages: 5188075a61d0SMina Almasry /* put back original number of pages, chg */ 5189075a61d0SMina Almasry (void)hugepage_subpool_put_pages(spool, chg); 5190075a61d0SMina Almasry out_uncharge_cgroup: 5191075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), 5192075a61d0SMina Almasry chg * pages_per_huge_page(h), h_cg); 5193c50ac050SDave Hansen out_err: 51945e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 51950db9d74eSMina Almasry /* Only call region_abort if the region_chg succeeded but the 51960db9d74eSMina Almasry * region_add failed or didn't run. 51970db9d74eSMina Almasry */ 51980db9d74eSMina Almasry if (chg >= 0 && add < 0) 51990db9d74eSMina Almasry region_abort(resv_map, from, to, regions_needed); 5200f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 5201f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 5202c50ac050SDave Hansen return ret; 5203a43a8c39SChen, Kenneth W } 5204a43a8c39SChen, Kenneth W 5205b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 5206b5cec28dSMike Kravetz long freed) 5207a43a8c39SChen, Kenneth W { 5208a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 52094e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 52109119a41eSJoonsoo Kim long chg = 0; 521190481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 52121c5ecae3SMike Kravetz long gbl_reserve; 521345c682a6SKen Chen 5214f27a5136SMike Kravetz /* 5215f27a5136SMike Kravetz * Since this routine can be called in the evict inode path for all 5216f27a5136SMike Kravetz * hugetlbfs inodes, resv_map could be NULL. 5217f27a5136SMike Kravetz */ 5218b5cec28dSMike Kravetz if (resv_map) { 5219b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 5220b5cec28dSMike Kravetz /* 5221b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 5222b5cec28dSMike Kravetz * must be split and another region descriptor can not be 5223b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 5224b5cec28dSMike Kravetz */ 5225b5cec28dSMike Kravetz if (chg < 0) 5226b5cec28dSMike Kravetz return chg; 5227b5cec28dSMike Kravetz } 5228b5cec28dSMike Kravetz 522945c682a6SKen Chen spin_lock(&inode->i_lock); 5230e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 523145c682a6SKen Chen spin_unlock(&inode->i_lock); 523245c682a6SKen Chen 52331c5ecae3SMike Kravetz /* 52341c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 52351c5ecae3SMike Kravetz * reservations to be released may be adjusted. 52361c5ecae3SMike Kravetz */ 52371c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 52381c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 5239b5cec28dSMike Kravetz 5240b5cec28dSMike Kravetz return 0; 5241a43a8c39SChen, Kenneth W } 524293f70f90SNaoya Horiguchi 52433212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 52443212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 52453212b535SSteve Capper struct vm_area_struct *vma, 52463212b535SSteve Capper unsigned long addr, pgoff_t idx) 52473212b535SSteve Capper { 52483212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 52493212b535SSteve Capper svma->vm_start; 52503212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 52513212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 52523212b535SSteve Capper 52533212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 5254de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 5255de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 52563212b535SSteve Capper 52573212b535SSteve Capper /* 52583212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 52593212b535SSteve Capper * page table page. 52603212b535SSteve Capper */ 52613212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 52623212b535SSteve Capper vm_flags != svm_flags || 526307e51edfSMiaohe Lin !range_in_vma(svma, sbase, s_end)) 52643212b535SSteve Capper return 0; 52653212b535SSteve Capper 52663212b535SSteve Capper return saddr; 52673212b535SSteve Capper } 52683212b535SSteve Capper 526931aafb45SNicholas Krause static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) 52703212b535SSteve Capper { 52713212b535SSteve Capper unsigned long base = addr & PUD_MASK; 52723212b535SSteve Capper unsigned long end = base + PUD_SIZE; 52733212b535SSteve Capper 52743212b535SSteve Capper /* 52753212b535SSteve Capper * check on proper vm_flags and page table alignment 52763212b535SSteve Capper */ 5277017b1660SMike Kravetz if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) 527831aafb45SNicholas Krause return true; 527931aafb45SNicholas Krause return false; 52803212b535SSteve Capper } 52813212b535SSteve Capper 52823212b535SSteve Capper /* 5283017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 5284017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 5285017b1660SMike Kravetz * shared pmd mappings. 5286017b1660SMike Kravetz */ 5287017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 5288017b1660SMike Kravetz unsigned long *start, unsigned long *end) 5289017b1660SMike Kravetz { 5290a1ba9da8SLi Xinhai unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), 5291a1ba9da8SLi Xinhai v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 5292017b1660SMike Kravetz 5293a1ba9da8SLi Xinhai /* 5294a1ba9da8SLi Xinhai * vma need span at least one aligned PUD size and the start,end range 5295a1ba9da8SLi Xinhai * must at least partialy within it. 5296a1ba9da8SLi Xinhai */ 5297a1ba9da8SLi Xinhai if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || 5298a1ba9da8SLi Xinhai (*end <= v_start) || (*start >= v_end)) 5299017b1660SMike Kravetz return; 5300017b1660SMike Kravetz 530175802ca6SPeter Xu /* Extend the range to be PUD aligned for a worst case scenario */ 5302a1ba9da8SLi Xinhai if (*start > v_start) 5303a1ba9da8SLi Xinhai *start = ALIGN_DOWN(*start, PUD_SIZE); 5304017b1660SMike Kravetz 5305a1ba9da8SLi Xinhai if (*end < v_end) 5306a1ba9da8SLi Xinhai *end = ALIGN(*end, PUD_SIZE); 5307017b1660SMike Kravetz } 5308017b1660SMike Kravetz 5309017b1660SMike Kravetz /* 53103212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 53113212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 53123212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 5313c0d0381aSMike Kravetz * code much cleaner. 5314c0d0381aSMike Kravetz * 53150bf7b64eSMike Kravetz * This routine must be called with i_mmap_rwsem held in at least read mode if 53160bf7b64eSMike Kravetz * sharing is possible. For hugetlbfs, this prevents removal of any page 53170bf7b64eSMike Kravetz * table entries associated with the address space. This is important as we 53180bf7b64eSMike Kravetz * are setting up sharing based on existing page table entries (mappings). 53190bf7b64eSMike Kravetz * 53200bf7b64eSMike Kravetz * NOTE: This routine is only called from huge_pte_alloc. Some callers of 53210bf7b64eSMike Kravetz * huge_pte_alloc know that sharing is not possible and do not take 53220bf7b64eSMike Kravetz * i_mmap_rwsem as a performance optimization. This is handled by the 53230bf7b64eSMike Kravetz * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is 53240bf7b64eSMike Kravetz * only required for subsequent processing. 53253212b535SSteve Capper */ 53263212b535SSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 53273212b535SSteve Capper { 53283212b535SSteve Capper struct vm_area_struct *vma = find_vma(mm, addr); 53293212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 53303212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 53313212b535SSteve Capper vma->vm_pgoff; 53323212b535SSteve Capper struct vm_area_struct *svma; 53333212b535SSteve Capper unsigned long saddr; 53343212b535SSteve Capper pte_t *spte = NULL; 53353212b535SSteve Capper pte_t *pte; 5336cb900f41SKirill A. Shutemov spinlock_t *ptl; 53373212b535SSteve Capper 53383212b535SSteve Capper if (!vma_shareable(vma, addr)) 53393212b535SSteve Capper return (pte_t *)pmd_alloc(mm, pud, addr); 53403212b535SSteve Capper 53410bf7b64eSMike Kravetz i_mmap_assert_locked(mapping); 53423212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 53433212b535SSteve Capper if (svma == vma) 53443212b535SSteve Capper continue; 53453212b535SSteve Capper 53463212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 53473212b535SSteve Capper if (saddr) { 53487868a208SPunit Agrawal spte = huge_pte_offset(svma->vm_mm, saddr, 53497868a208SPunit Agrawal vma_mmu_pagesize(svma)); 53503212b535SSteve Capper if (spte) { 53513212b535SSteve Capper get_page(virt_to_page(spte)); 53523212b535SSteve Capper break; 53533212b535SSteve Capper } 53543212b535SSteve Capper } 53553212b535SSteve Capper } 53563212b535SSteve Capper 53573212b535SSteve Capper if (!spte) 53583212b535SSteve Capper goto out; 53593212b535SSteve Capper 53608bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 5361dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 53623212b535SSteve Capper pud_populate(mm, pud, 53633212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 5364c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 5365dc6c9a35SKirill A. Shutemov } else { 53663212b535SSteve Capper put_page(virt_to_page(spte)); 5367dc6c9a35SKirill A. Shutemov } 5368cb900f41SKirill A. Shutemov spin_unlock(ptl); 53693212b535SSteve Capper out: 53703212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 53713212b535SSteve Capper return pte; 53723212b535SSteve Capper } 53733212b535SSteve Capper 53743212b535SSteve Capper /* 53753212b535SSteve Capper * unmap huge page backed by shared pte. 53763212b535SSteve Capper * 53773212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 53783212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 53793212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 53803212b535SSteve Capper * 5381c0d0381aSMike Kravetz * Called with page table lock held and i_mmap_rwsem held in write mode. 53823212b535SSteve Capper * 53833212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 53843212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 53853212b535SSteve Capper */ 538634ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 538734ae204fSMike Kravetz unsigned long *addr, pte_t *ptep) 53883212b535SSteve Capper { 53893212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 5390c2febafcSKirill A. Shutemov p4d_t *p4d = p4d_offset(pgd, *addr); 5391c2febafcSKirill A. Shutemov pud_t *pud = pud_offset(p4d, *addr); 53923212b535SSteve Capper 539334ae204fSMike Kravetz i_mmap_assert_write_locked(vma->vm_file->f_mapping); 53943212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 53953212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 53963212b535SSteve Capper return 0; 53973212b535SSteve Capper 53983212b535SSteve Capper pud_clear(pud); 53993212b535SSteve Capper put_page(virt_to_page(ptep)); 5400dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 54013212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 54023212b535SSteve Capper return 1; 54033212b535SSteve Capper } 54049e5fc74cSSteve Capper #define want_pmd_share() (1) 54059e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 54069e5fc74cSSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 54079e5fc74cSSteve Capper { 54089e5fc74cSSteve Capper return NULL; 54099e5fc74cSSteve Capper } 5410e81f2d22SZhang Zhen 541134ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 541234ae204fSMike Kravetz unsigned long *addr, pte_t *ptep) 5413e81f2d22SZhang Zhen { 5414e81f2d22SZhang Zhen return 0; 5415e81f2d22SZhang Zhen } 5416017b1660SMike Kravetz 5417017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 5418017b1660SMike Kravetz unsigned long *start, unsigned long *end) 5419017b1660SMike Kravetz { 5420017b1660SMike Kravetz } 54219e5fc74cSSteve Capper #define want_pmd_share() (0) 54223212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 54233212b535SSteve Capper 54249e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 54259e5fc74cSSteve Capper pte_t *huge_pte_alloc(struct mm_struct *mm, 54269e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 54279e5fc74cSSteve Capper { 54289e5fc74cSSteve Capper pgd_t *pgd; 5429c2febafcSKirill A. Shutemov p4d_t *p4d; 54309e5fc74cSSteve Capper pud_t *pud; 54319e5fc74cSSteve Capper pte_t *pte = NULL; 54329e5fc74cSSteve Capper 54339e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 5434f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 5435f4f0a3d8SKirill A. Shutemov if (!p4d) 5436f4f0a3d8SKirill A. Shutemov return NULL; 5437c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 54389e5fc74cSSteve Capper if (pud) { 54399e5fc74cSSteve Capper if (sz == PUD_SIZE) { 54409e5fc74cSSteve Capper pte = (pte_t *)pud; 54419e5fc74cSSteve Capper } else { 54429e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 54439e5fc74cSSteve Capper if (want_pmd_share() && pud_none(*pud)) 54449e5fc74cSSteve Capper pte = huge_pmd_share(mm, addr, pud); 54459e5fc74cSSteve Capper else 54469e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 54479e5fc74cSSteve Capper } 54489e5fc74cSSteve Capper } 54494e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 54509e5fc74cSSteve Capper 54519e5fc74cSSteve Capper return pte; 54529e5fc74cSSteve Capper } 54539e5fc74cSSteve Capper 54549b19df29SPunit Agrawal /* 54559b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 54569b19df29SPunit Agrawal * entry at address @addr 54579b19df29SPunit Agrawal * 54588ac0b81aSLi Xinhai * Return: Pointer to page table entry (PUD or PMD) for 54598ac0b81aSLi Xinhai * address @addr, or NULL if a !p*d_present() entry is encountered and the 54609b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 54619b19df29SPunit Agrawal * table. 54629b19df29SPunit Agrawal */ 54637868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 54647868a208SPunit Agrawal unsigned long addr, unsigned long sz) 54659e5fc74cSSteve Capper { 54669e5fc74cSSteve Capper pgd_t *pgd; 5467c2febafcSKirill A. Shutemov p4d_t *p4d; 54688ac0b81aSLi Xinhai pud_t *pud; 54698ac0b81aSLi Xinhai pmd_t *pmd; 54709e5fc74cSSteve Capper 54719e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 5472c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 5473c2febafcSKirill A. Shutemov return NULL; 5474c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 5475c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 5476c2febafcSKirill A. Shutemov return NULL; 54779b19df29SPunit Agrawal 5478c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 54798ac0b81aSLi Xinhai if (sz == PUD_SIZE) 54808ac0b81aSLi Xinhai /* must be pud huge, non-present or none */ 54819e5fc74cSSteve Capper return (pte_t *)pud; 54828ac0b81aSLi Xinhai if (!pud_present(*pud)) 54838ac0b81aSLi Xinhai return NULL; 54848ac0b81aSLi Xinhai /* must have a valid entry and size to go further */ 54859b19df29SPunit Agrawal 54869e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 54878ac0b81aSLi Xinhai /* must be pmd huge, non-present or none */ 54889e5fc74cSSteve Capper return (pte_t *)pmd; 54899e5fc74cSSteve Capper } 54909e5fc74cSSteve Capper 549161f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 549261f77edaSNaoya Horiguchi 549361f77edaSNaoya Horiguchi /* 549461f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 549561f77edaSNaoya Horiguchi * behavior. 549661f77edaSNaoya Horiguchi */ 549761f77edaSNaoya Horiguchi struct page * __weak 549861f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 549961f77edaSNaoya Horiguchi int write) 550061f77edaSNaoya Horiguchi { 550161f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 550261f77edaSNaoya Horiguchi } 550361f77edaSNaoya Horiguchi 550461f77edaSNaoya Horiguchi struct page * __weak 55054dc71451SAneesh Kumar K.V follow_huge_pd(struct vm_area_struct *vma, 55064dc71451SAneesh Kumar K.V unsigned long address, hugepd_t hpd, int flags, int pdshift) 55074dc71451SAneesh Kumar K.V { 55084dc71451SAneesh Kumar K.V WARN(1, "hugepd follow called with no support for hugepage directory format\n"); 55094dc71451SAneesh Kumar K.V return NULL; 55104dc71451SAneesh Kumar K.V } 55114dc71451SAneesh Kumar K.V 55124dc71451SAneesh Kumar K.V struct page * __weak 55139e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 5514e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 55159e5fc74cSSteve Capper { 5516e66f17ffSNaoya Horiguchi struct page *page = NULL; 5517e66f17ffSNaoya Horiguchi spinlock_t *ptl; 5518c9d398faSNaoya Horiguchi pte_t pte; 55193faa52c0SJohn Hubbard 55203faa52c0SJohn Hubbard /* FOLL_GET and FOLL_PIN are mutually exclusive. */ 55213faa52c0SJohn Hubbard if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == 55223faa52c0SJohn Hubbard (FOLL_PIN | FOLL_GET))) 55233faa52c0SJohn Hubbard return NULL; 55243faa52c0SJohn Hubbard 5525e66f17ffSNaoya Horiguchi retry: 5526e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 5527e66f17ffSNaoya Horiguchi spin_lock(ptl); 5528e66f17ffSNaoya Horiguchi /* 5529e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 5530e66f17ffSNaoya Horiguchi * unmapped from other threads. 5531e66f17ffSNaoya Horiguchi */ 5532e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 5533e66f17ffSNaoya Horiguchi goto out; 5534c9d398faSNaoya Horiguchi pte = huge_ptep_get((pte_t *)pmd); 5535c9d398faSNaoya Horiguchi if (pte_present(pte)) { 553697534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 55373faa52c0SJohn Hubbard /* 55383faa52c0SJohn Hubbard * try_grab_page() should always succeed here, because: a) we 55393faa52c0SJohn Hubbard * hold the pmd (ptl) lock, and b) we've just checked that the 55403faa52c0SJohn Hubbard * huge pmd (head) page is present in the page tables. The ptl 55413faa52c0SJohn Hubbard * prevents the head page and tail pages from being rearranged 55423faa52c0SJohn Hubbard * in any way. So this page must be available at this point, 55433faa52c0SJohn Hubbard * unless the page refcount overflowed: 55443faa52c0SJohn Hubbard */ 55453faa52c0SJohn Hubbard if (WARN_ON_ONCE(!try_grab_page(page, flags))) { 55463faa52c0SJohn Hubbard page = NULL; 55473faa52c0SJohn Hubbard goto out; 55483faa52c0SJohn Hubbard } 5549e66f17ffSNaoya Horiguchi } else { 5550c9d398faSNaoya Horiguchi if (is_hugetlb_entry_migration(pte)) { 5551e66f17ffSNaoya Horiguchi spin_unlock(ptl); 5552e66f17ffSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pmd, ptl); 5553e66f17ffSNaoya Horiguchi goto retry; 5554e66f17ffSNaoya Horiguchi } 5555e66f17ffSNaoya Horiguchi /* 5556e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 5557e66f17ffSNaoya Horiguchi * follow_page_mask(). 5558e66f17ffSNaoya Horiguchi */ 5559e66f17ffSNaoya Horiguchi } 5560e66f17ffSNaoya Horiguchi out: 5561e66f17ffSNaoya Horiguchi spin_unlock(ptl); 55629e5fc74cSSteve Capper return page; 55639e5fc74cSSteve Capper } 55649e5fc74cSSteve Capper 556561f77edaSNaoya Horiguchi struct page * __weak 55669e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 5567e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 55689e5fc74cSSteve Capper { 55693faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 5570e66f17ffSNaoya Horiguchi return NULL; 55719e5fc74cSSteve Capper 5572e66f17ffSNaoya Horiguchi return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 55739e5fc74cSSteve Capper } 55749e5fc74cSSteve Capper 5575faaa5b62SAnshuman Khandual struct page * __weak 5576faaa5b62SAnshuman Khandual follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) 5577faaa5b62SAnshuman Khandual { 55783faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 5579faaa5b62SAnshuman Khandual return NULL; 5580faaa5b62SAnshuman Khandual 5581faaa5b62SAnshuman Khandual return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); 5582faaa5b62SAnshuman Khandual } 5583faaa5b62SAnshuman Khandual 558431caf665SNaoya Horiguchi bool isolate_huge_page(struct page *page, struct list_head *list) 558531caf665SNaoya Horiguchi { 5586bcc54222SNaoya Horiguchi bool ret = true; 5587bcc54222SNaoya Horiguchi 558831caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 5589*8f251a3dSMike Kravetz if (!PageHeadHuge(page) || 5590*8f251a3dSMike Kravetz !HPageMigratable(page) || 55910eb2df2bSMuchun Song !get_page_unless_zero(page)) { 5592bcc54222SNaoya Horiguchi ret = false; 5593bcc54222SNaoya Horiguchi goto unlock; 5594bcc54222SNaoya Horiguchi } 5595*8f251a3dSMike Kravetz ClearHPageMigratable(page); 559631caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 5597bcc54222SNaoya Horiguchi unlock: 559831caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 5599bcc54222SNaoya Horiguchi return ret; 560031caf665SNaoya Horiguchi } 560131caf665SNaoya Horiguchi 560231caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 560331caf665SNaoya Horiguchi { 560431caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 5605*8f251a3dSMike Kravetz SetHPageMigratable(page); 560631caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 560731caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 560831caf665SNaoya Horiguchi put_page(page); 560931caf665SNaoya Horiguchi } 5610ab5ac90aSMichal Hocko 5611ab5ac90aSMichal Hocko void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) 5612ab5ac90aSMichal Hocko { 5613ab5ac90aSMichal Hocko struct hstate *h = page_hstate(oldpage); 5614ab5ac90aSMichal Hocko 5615ab5ac90aSMichal Hocko hugetlb_cgroup_migrate(oldpage, newpage); 5616ab5ac90aSMichal Hocko set_page_owner_migrate_reason(newpage, reason); 5617ab5ac90aSMichal Hocko 5618ab5ac90aSMichal Hocko /* 5619ab5ac90aSMichal Hocko * transfer temporary state of the new huge page. This is 5620ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 5621ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 5622ab5ac90aSMichal Hocko * the temporary status. 5623ab5ac90aSMichal Hocko * 5624ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 5625ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 5626ab5ac90aSMichal Hocko * the per-node's. 5627ab5ac90aSMichal Hocko */ 5628ab5ac90aSMichal Hocko if (PageHugeTemporary(newpage)) { 5629ab5ac90aSMichal Hocko int old_nid = page_to_nid(oldpage); 5630ab5ac90aSMichal Hocko int new_nid = page_to_nid(newpage); 5631ab5ac90aSMichal Hocko 5632ab5ac90aSMichal Hocko SetPageHugeTemporary(oldpage); 5633ab5ac90aSMichal Hocko ClearPageHugeTemporary(newpage); 5634ab5ac90aSMichal Hocko 5635ab5ac90aSMichal Hocko spin_lock(&hugetlb_lock); 5636ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 5637ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 5638ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 5639ab5ac90aSMichal Hocko } 5640ab5ac90aSMichal Hocko spin_unlock(&hugetlb_lock); 5641ab5ac90aSMichal Hocko } 5642ab5ac90aSMichal Hocko } 5643cf11e85fSRoman Gushchin 5644cf11e85fSRoman Gushchin #ifdef CONFIG_CMA 5645cf11e85fSRoman Gushchin static bool cma_reserve_called __initdata; 5646cf11e85fSRoman Gushchin 5647cf11e85fSRoman Gushchin static int __init cmdline_parse_hugetlb_cma(char *p) 5648cf11e85fSRoman Gushchin { 5649cf11e85fSRoman Gushchin hugetlb_cma_size = memparse(p, &p); 5650cf11e85fSRoman Gushchin return 0; 5651cf11e85fSRoman Gushchin } 5652cf11e85fSRoman Gushchin 5653cf11e85fSRoman Gushchin early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); 5654cf11e85fSRoman Gushchin 5655cf11e85fSRoman Gushchin void __init hugetlb_cma_reserve(int order) 5656cf11e85fSRoman Gushchin { 5657cf11e85fSRoman Gushchin unsigned long size, reserved, per_node; 5658cf11e85fSRoman Gushchin int nid; 5659cf11e85fSRoman Gushchin 5660cf11e85fSRoman Gushchin cma_reserve_called = true; 5661cf11e85fSRoman Gushchin 5662cf11e85fSRoman Gushchin if (!hugetlb_cma_size) 5663cf11e85fSRoman Gushchin return; 5664cf11e85fSRoman Gushchin 5665cf11e85fSRoman Gushchin if (hugetlb_cma_size < (PAGE_SIZE << order)) { 5666cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", 5667cf11e85fSRoman Gushchin (PAGE_SIZE << order) / SZ_1M); 5668cf11e85fSRoman Gushchin return; 5669cf11e85fSRoman Gushchin } 5670cf11e85fSRoman Gushchin 5671cf11e85fSRoman Gushchin /* 5672cf11e85fSRoman Gushchin * If 3 GB area is requested on a machine with 4 numa nodes, 5673cf11e85fSRoman Gushchin * let's allocate 1 GB on first three nodes and ignore the last one. 5674cf11e85fSRoman Gushchin */ 5675cf11e85fSRoman Gushchin per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); 5676cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", 5677cf11e85fSRoman Gushchin hugetlb_cma_size / SZ_1M, per_node / SZ_1M); 5678cf11e85fSRoman Gushchin 5679cf11e85fSRoman Gushchin reserved = 0; 5680cf11e85fSRoman Gushchin for_each_node_state(nid, N_ONLINE) { 5681cf11e85fSRoman Gushchin int res; 56822281f797SBarry Song char name[CMA_MAX_NAME]; 5683cf11e85fSRoman Gushchin 5684cf11e85fSRoman Gushchin size = min(per_node, hugetlb_cma_size - reserved); 5685cf11e85fSRoman Gushchin size = round_up(size, PAGE_SIZE << order); 5686cf11e85fSRoman Gushchin 56872281f797SBarry Song snprintf(name, sizeof(name), "hugetlb%d", nid); 5688cf11e85fSRoman Gushchin res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order, 568929d0f41dSBarry Song 0, false, name, 5690cf11e85fSRoman Gushchin &hugetlb_cma[nid], nid); 5691cf11e85fSRoman Gushchin if (res) { 5692cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: reservation failed: err %d, node %d", 5693cf11e85fSRoman Gushchin res, nid); 5694cf11e85fSRoman Gushchin continue; 5695cf11e85fSRoman Gushchin } 5696cf11e85fSRoman Gushchin 5697cf11e85fSRoman Gushchin reserved += size; 5698cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", 5699cf11e85fSRoman Gushchin size / SZ_1M, nid); 5700cf11e85fSRoman Gushchin 5701cf11e85fSRoman Gushchin if (reserved >= hugetlb_cma_size) 5702cf11e85fSRoman Gushchin break; 5703cf11e85fSRoman Gushchin } 5704cf11e85fSRoman Gushchin } 5705cf11e85fSRoman Gushchin 5706cf11e85fSRoman Gushchin void __init hugetlb_cma_check(void) 5707cf11e85fSRoman Gushchin { 5708cf11e85fSRoman Gushchin if (!hugetlb_cma_size || cma_reserve_called) 5709cf11e85fSRoman Gushchin return; 5710cf11e85fSRoman Gushchin 5711cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); 5712cf11e85fSRoman Gushchin } 5713cf11e85fSRoman Gushchin 5714cf11e85fSRoman Gushchin #endif /* CONFIG_CMA */ 5715