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> 338cc5fcbbSMina Almasry #include <linux/migrate.h> 34d6606683SLinus Torvalds 3563551ae0SDavid Gibson #include <asm/page.h> 36ca15ca40SMike Rapoport #include <asm/pgalloc.h> 3724669e58SAneesh Kumar K.V #include <asm/tlb.h> 3863551ae0SDavid Gibson 3924669e58SAneesh Kumar K.V #include <linux/io.h> 4063551ae0SDavid Gibson #include <linux/hugetlb.h> 419dd540e2SAneesh Kumar K.V #include <linux/hugetlb_cgroup.h> 429a305230SLee Schermerhorn #include <linux/node.h> 43ab5ac90aSMichal Hocko #include <linux/page_owner.h> 447835e98bSNick Piggin #include "internal.h" 45f41f2ed4SMuchun Song #include "hugetlb_vmemmap.h" 461da177e4SLinus Torvalds 47c3f38a38SAneesh Kumar K.V int hugetlb_max_hstate __read_mostly; 48e5ff2159SAndi Kleen unsigned int default_hstate_idx; 49e5ff2159SAndi Kleen struct hstate hstates[HUGE_MAX_HSTATE]; 50cf11e85fSRoman Gushchin 51dbda8feaSBarry Song #ifdef CONFIG_CMA 52cf11e85fSRoman Gushchin static struct cma *hugetlb_cma[MAX_NUMNODES]; 53dbda8feaSBarry Song #endif 54dbda8feaSBarry Song static unsigned long hugetlb_cma_size __initdata; 55cf11e85fSRoman Gushchin 56641844f5SNaoya Horiguchi /* 57641844f5SNaoya Horiguchi * Minimum page order among possible hugepage sizes, set to a proper value 58641844f5SNaoya Horiguchi * at boot time. 59641844f5SNaoya Horiguchi */ 60641844f5SNaoya Horiguchi static unsigned int minimum_order __read_mostly = UINT_MAX; 61e5ff2159SAndi Kleen 6253ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 6353ba51d2SJon Tollefson 64e5ff2159SAndi Kleen /* for command line parsing */ 65e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 66e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 679fee021dSVaishali Thakkar static bool __initdata parsed_valid_hugepagesz = true; 68282f4214SMike Kravetz static bool __initdata parsed_default_hugepagesz; 69e5ff2159SAndi Kleen 703935baa9SDavid Gibson /* 7131caf665SNaoya Horiguchi * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, 7231caf665SNaoya Horiguchi * free_huge_pages, and surplus_huge_pages. 733935baa9SDavid Gibson */ 74c3f38a38SAneesh Kumar K.V DEFINE_SPINLOCK(hugetlb_lock); 750bd0f9fbSEric Paris 768382d914SDavidlohr Bueso /* 778382d914SDavidlohr Bueso * Serializes faults on the same logical page. This is used to 788382d914SDavidlohr Bueso * prevent spurious OOMs when the hugepage pool is fully utilized. 798382d914SDavidlohr Bueso */ 808382d914SDavidlohr Bueso static int num_fault_mutexes; 81c672c7f2SMike Kravetz struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; 828382d914SDavidlohr Bueso 837ca02d0aSMike Kravetz /* Forward declaration */ 847ca02d0aSMike Kravetz static int hugetlb_acct_memory(struct hstate *h, long delta); 857ca02d0aSMike Kravetz 861d88433bSMiaohe Lin static inline bool subpool_is_free(struct hugepage_subpool *spool) 871d88433bSMiaohe Lin { 881d88433bSMiaohe Lin if (spool->count) 891d88433bSMiaohe Lin return false; 901d88433bSMiaohe Lin if (spool->max_hpages != -1) 911d88433bSMiaohe Lin return spool->used_hpages == 0; 921d88433bSMiaohe Lin if (spool->min_hpages != -1) 931d88433bSMiaohe Lin return spool->rsv_hpages == spool->min_hpages; 941d88433bSMiaohe Lin 951d88433bSMiaohe Lin return true; 961d88433bSMiaohe Lin } 971d88433bSMiaohe Lin 98db71ef79SMike Kravetz static inline void unlock_or_release_subpool(struct hugepage_subpool *spool, 99db71ef79SMike Kravetz unsigned long irq_flags) 10090481622SDavid Gibson { 101db71ef79SMike Kravetz spin_unlock_irqrestore(&spool->lock, irq_flags); 10290481622SDavid Gibson 10390481622SDavid Gibson /* If no pages are used, and no other handles to the subpool 1047c8de358SEthon Paul * remain, give up any reservations based on minimum size and 1057ca02d0aSMike Kravetz * free the subpool */ 1061d88433bSMiaohe Lin if (subpool_is_free(spool)) { 1077ca02d0aSMike Kravetz if (spool->min_hpages != -1) 1087ca02d0aSMike Kravetz hugetlb_acct_memory(spool->hstate, 1097ca02d0aSMike Kravetz -spool->min_hpages); 11090481622SDavid Gibson kfree(spool); 11190481622SDavid Gibson } 1127ca02d0aSMike Kravetz } 11390481622SDavid Gibson 1147ca02d0aSMike Kravetz struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, 1157ca02d0aSMike Kravetz long min_hpages) 11690481622SDavid Gibson { 11790481622SDavid Gibson struct hugepage_subpool *spool; 11890481622SDavid Gibson 119c6a91820SMike Kravetz spool = kzalloc(sizeof(*spool), GFP_KERNEL); 12090481622SDavid Gibson if (!spool) 12190481622SDavid Gibson return NULL; 12290481622SDavid Gibson 12390481622SDavid Gibson spin_lock_init(&spool->lock); 12490481622SDavid Gibson spool->count = 1; 1257ca02d0aSMike Kravetz spool->max_hpages = max_hpages; 1267ca02d0aSMike Kravetz spool->hstate = h; 1277ca02d0aSMike Kravetz spool->min_hpages = min_hpages; 1287ca02d0aSMike Kravetz 1297ca02d0aSMike Kravetz if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { 1307ca02d0aSMike Kravetz kfree(spool); 1317ca02d0aSMike Kravetz return NULL; 1327ca02d0aSMike Kravetz } 1337ca02d0aSMike Kravetz spool->rsv_hpages = min_hpages; 13490481622SDavid Gibson 13590481622SDavid Gibson return spool; 13690481622SDavid Gibson } 13790481622SDavid Gibson 13890481622SDavid Gibson void hugepage_put_subpool(struct hugepage_subpool *spool) 13990481622SDavid Gibson { 140db71ef79SMike Kravetz unsigned long flags; 141db71ef79SMike Kravetz 142db71ef79SMike Kravetz spin_lock_irqsave(&spool->lock, flags); 14390481622SDavid Gibson BUG_ON(!spool->count); 14490481622SDavid Gibson spool->count--; 145db71ef79SMike Kravetz unlock_or_release_subpool(spool, flags); 14690481622SDavid Gibson } 14790481622SDavid Gibson 1481c5ecae3SMike Kravetz /* 1491c5ecae3SMike Kravetz * Subpool accounting for allocating and reserving pages. 1501c5ecae3SMike Kravetz * Return -ENOMEM if there are not enough resources to satisfy the 1519e7ee400SRandy Dunlap * request. Otherwise, return the number of pages by which the 1521c5ecae3SMike Kravetz * global pools must be adjusted (upward). The returned value may 1531c5ecae3SMike Kravetz * only be different than the passed value (delta) in the case where 1547c8de358SEthon Paul * a subpool minimum size must be maintained. 1551c5ecae3SMike Kravetz */ 1561c5ecae3SMike Kravetz static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, 15790481622SDavid Gibson long delta) 15890481622SDavid Gibson { 1591c5ecae3SMike Kravetz long ret = delta; 16090481622SDavid Gibson 16190481622SDavid Gibson if (!spool) 1621c5ecae3SMike Kravetz return ret; 16390481622SDavid Gibson 164db71ef79SMike Kravetz spin_lock_irq(&spool->lock); 16590481622SDavid Gibson 1661c5ecae3SMike Kravetz if (spool->max_hpages != -1) { /* maximum size accounting */ 1671c5ecae3SMike Kravetz if ((spool->used_hpages + delta) <= spool->max_hpages) 1681c5ecae3SMike Kravetz spool->used_hpages += delta; 1691c5ecae3SMike Kravetz else { 1701c5ecae3SMike Kravetz ret = -ENOMEM; 1711c5ecae3SMike Kravetz goto unlock_ret; 1721c5ecae3SMike Kravetz } 1731c5ecae3SMike Kravetz } 1741c5ecae3SMike Kravetz 17509a95e29SMike Kravetz /* minimum size accounting */ 17609a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->rsv_hpages) { 1771c5ecae3SMike Kravetz if (delta > spool->rsv_hpages) { 1781c5ecae3SMike Kravetz /* 1791c5ecae3SMike Kravetz * Asking for more reserves than those already taken on 1801c5ecae3SMike Kravetz * behalf of subpool. Return difference. 1811c5ecae3SMike Kravetz */ 1821c5ecae3SMike Kravetz ret = delta - spool->rsv_hpages; 1831c5ecae3SMike Kravetz spool->rsv_hpages = 0; 1841c5ecae3SMike Kravetz } else { 1851c5ecae3SMike Kravetz ret = 0; /* reserves already accounted for */ 1861c5ecae3SMike Kravetz spool->rsv_hpages -= delta; 1871c5ecae3SMike Kravetz } 1881c5ecae3SMike Kravetz } 1891c5ecae3SMike Kravetz 1901c5ecae3SMike Kravetz unlock_ret: 191db71ef79SMike Kravetz spin_unlock_irq(&spool->lock); 19290481622SDavid Gibson return ret; 19390481622SDavid Gibson } 19490481622SDavid Gibson 1951c5ecae3SMike Kravetz /* 1961c5ecae3SMike Kravetz * Subpool accounting for freeing and unreserving pages. 1971c5ecae3SMike Kravetz * Return the number of global page reservations that must be dropped. 1981c5ecae3SMike Kravetz * The return value may only be different than the passed value (delta) 1991c5ecae3SMike Kravetz * in the case where a subpool minimum size must be maintained. 2001c5ecae3SMike Kravetz */ 2011c5ecae3SMike Kravetz static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, 20290481622SDavid Gibson long delta) 20390481622SDavid Gibson { 2041c5ecae3SMike Kravetz long ret = delta; 205db71ef79SMike Kravetz unsigned long flags; 2061c5ecae3SMike Kravetz 20790481622SDavid Gibson if (!spool) 2081c5ecae3SMike Kravetz return delta; 20990481622SDavid Gibson 210db71ef79SMike Kravetz spin_lock_irqsave(&spool->lock, flags); 2111c5ecae3SMike Kravetz 2121c5ecae3SMike Kravetz if (spool->max_hpages != -1) /* maximum size accounting */ 21390481622SDavid Gibson spool->used_hpages -= delta; 2141c5ecae3SMike Kravetz 21509a95e29SMike Kravetz /* minimum size accounting */ 21609a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { 2171c5ecae3SMike Kravetz if (spool->rsv_hpages + delta <= spool->min_hpages) 2181c5ecae3SMike Kravetz ret = 0; 2191c5ecae3SMike Kravetz else 2201c5ecae3SMike Kravetz ret = spool->rsv_hpages + delta - spool->min_hpages; 2211c5ecae3SMike Kravetz 2221c5ecae3SMike Kravetz spool->rsv_hpages += delta; 2231c5ecae3SMike Kravetz if (spool->rsv_hpages > spool->min_hpages) 2241c5ecae3SMike Kravetz spool->rsv_hpages = spool->min_hpages; 2251c5ecae3SMike Kravetz } 2261c5ecae3SMike Kravetz 2271c5ecae3SMike Kravetz /* 2281c5ecae3SMike Kravetz * If hugetlbfs_put_super couldn't free spool due to an outstanding 2291c5ecae3SMike Kravetz * quota reference, free it now. 2301c5ecae3SMike Kravetz */ 231db71ef79SMike Kravetz unlock_or_release_subpool(spool, flags); 2321c5ecae3SMike Kravetz 2331c5ecae3SMike Kravetz return ret; 23490481622SDavid Gibson } 23590481622SDavid Gibson 23690481622SDavid Gibson static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 23790481622SDavid Gibson { 23890481622SDavid Gibson return HUGETLBFS_SB(inode->i_sb)->spool; 23990481622SDavid Gibson } 24090481622SDavid Gibson 24190481622SDavid Gibson static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 24290481622SDavid Gibson { 243496ad9aaSAl Viro return subpool_inode(file_inode(vma->vm_file)); 24490481622SDavid Gibson } 24590481622SDavid Gibson 2460db9d74eSMina Almasry /* Helper that removes a struct file_region from the resv_map cache and returns 2470db9d74eSMina Almasry * it for use. 2480db9d74eSMina Almasry */ 2490db9d74eSMina Almasry static struct file_region * 2500db9d74eSMina Almasry get_file_region_entry_from_cache(struct resv_map *resv, long from, long to) 2510db9d74eSMina Almasry { 2520db9d74eSMina Almasry struct file_region *nrg = NULL; 2530db9d74eSMina Almasry 2540db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count <= 0); 2550db9d74eSMina Almasry 2560db9d74eSMina Almasry resv->region_cache_count--; 2570db9d74eSMina Almasry nrg = list_first_entry(&resv->region_cache, struct file_region, link); 2580db9d74eSMina Almasry list_del(&nrg->link); 2590db9d74eSMina Almasry 2600db9d74eSMina Almasry nrg->from = from; 2610db9d74eSMina Almasry nrg->to = to; 2620db9d74eSMina Almasry 2630db9d74eSMina Almasry return nrg; 2640db9d74eSMina Almasry } 2650db9d74eSMina Almasry 266075a61d0SMina Almasry static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, 267075a61d0SMina Almasry struct file_region *rg) 268075a61d0SMina Almasry { 269075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 270075a61d0SMina Almasry nrg->reservation_counter = rg->reservation_counter; 271075a61d0SMina Almasry nrg->css = rg->css; 272075a61d0SMina Almasry if (rg->css) 273075a61d0SMina Almasry css_get(rg->css); 274075a61d0SMina Almasry #endif 275075a61d0SMina Almasry } 276075a61d0SMina Almasry 277075a61d0SMina Almasry /* Helper that records hugetlb_cgroup uncharge info. */ 278075a61d0SMina Almasry static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, 279075a61d0SMina Almasry struct hstate *h, 280075a61d0SMina Almasry struct resv_map *resv, 281075a61d0SMina Almasry struct file_region *nrg) 282075a61d0SMina Almasry { 283075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 284075a61d0SMina Almasry if (h_cg) { 285075a61d0SMina Almasry nrg->reservation_counter = 286075a61d0SMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 287075a61d0SMina Almasry nrg->css = &h_cg->css; 288d85aecf2SMiaohe Lin /* 289d85aecf2SMiaohe Lin * The caller will hold exactly one h_cg->css reference for the 290d85aecf2SMiaohe Lin * whole contiguous reservation region. But this area might be 291d85aecf2SMiaohe Lin * scattered when there are already some file_regions reside in 292d85aecf2SMiaohe Lin * it. As a result, many file_regions may share only one css 293d85aecf2SMiaohe Lin * reference. In order to ensure that one file_region must hold 294d85aecf2SMiaohe Lin * exactly one h_cg->css reference, we should do css_get for 295d85aecf2SMiaohe Lin * each file_region and leave the reference held by caller 296d85aecf2SMiaohe Lin * untouched. 297d85aecf2SMiaohe Lin */ 298d85aecf2SMiaohe Lin css_get(&h_cg->css); 299075a61d0SMina Almasry if (!resv->pages_per_hpage) 300075a61d0SMina Almasry resv->pages_per_hpage = pages_per_huge_page(h); 301075a61d0SMina Almasry /* pages_per_hpage should be the same for all entries in 302075a61d0SMina Almasry * a resv_map. 303075a61d0SMina Almasry */ 304075a61d0SMina Almasry VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); 305075a61d0SMina Almasry } else { 306075a61d0SMina Almasry nrg->reservation_counter = NULL; 307075a61d0SMina Almasry nrg->css = NULL; 308075a61d0SMina Almasry } 309075a61d0SMina Almasry #endif 310075a61d0SMina Almasry } 311075a61d0SMina Almasry 312d85aecf2SMiaohe Lin static void put_uncharge_info(struct file_region *rg) 313d85aecf2SMiaohe Lin { 314d85aecf2SMiaohe Lin #ifdef CONFIG_CGROUP_HUGETLB 315d85aecf2SMiaohe Lin if (rg->css) 316d85aecf2SMiaohe Lin css_put(rg->css); 317d85aecf2SMiaohe Lin #endif 318d85aecf2SMiaohe Lin } 319d85aecf2SMiaohe Lin 320a9b3f867SMina Almasry static bool has_same_uncharge_info(struct file_region *rg, 321a9b3f867SMina Almasry struct file_region *org) 322a9b3f867SMina Almasry { 323a9b3f867SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 324a9b3f867SMina Almasry return rg && org && 325a9b3f867SMina Almasry rg->reservation_counter == org->reservation_counter && 326a9b3f867SMina Almasry rg->css == org->css; 327a9b3f867SMina Almasry 328a9b3f867SMina Almasry #else 329a9b3f867SMina Almasry return true; 330a9b3f867SMina Almasry #endif 331a9b3f867SMina Almasry } 332a9b3f867SMina Almasry 333a9b3f867SMina Almasry static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) 334a9b3f867SMina Almasry { 335a9b3f867SMina Almasry struct file_region *nrg = NULL, *prg = NULL; 336a9b3f867SMina Almasry 337a9b3f867SMina Almasry prg = list_prev_entry(rg, link); 338a9b3f867SMina Almasry if (&prg->link != &resv->regions && prg->to == rg->from && 339a9b3f867SMina Almasry has_same_uncharge_info(prg, rg)) { 340a9b3f867SMina Almasry prg->to = rg->to; 341a9b3f867SMina Almasry 342a9b3f867SMina Almasry list_del(&rg->link); 343d85aecf2SMiaohe Lin put_uncharge_info(rg); 344a9b3f867SMina Almasry kfree(rg); 345a9b3f867SMina Almasry 3467db5e7b6SWei Yang rg = prg; 347a9b3f867SMina Almasry } 348a9b3f867SMina Almasry 349a9b3f867SMina Almasry nrg = list_next_entry(rg, link); 350a9b3f867SMina Almasry if (&nrg->link != &resv->regions && nrg->from == rg->to && 351a9b3f867SMina Almasry has_same_uncharge_info(nrg, rg)) { 352a9b3f867SMina Almasry nrg->from = rg->from; 353a9b3f867SMina Almasry 354a9b3f867SMina Almasry list_del(&rg->link); 355d85aecf2SMiaohe Lin put_uncharge_info(rg); 356a9b3f867SMina Almasry kfree(rg); 357a9b3f867SMina Almasry } 358a9b3f867SMina Almasry } 359a9b3f867SMina Almasry 3602103cf9cSPeter Xu static inline long 3612103cf9cSPeter Xu hugetlb_resv_map_add(struct resv_map *map, struct file_region *rg, long from, 3622103cf9cSPeter Xu long to, struct hstate *h, struct hugetlb_cgroup *cg, 3632103cf9cSPeter Xu long *regions_needed) 3642103cf9cSPeter Xu { 3652103cf9cSPeter Xu struct file_region *nrg; 3662103cf9cSPeter Xu 3672103cf9cSPeter Xu if (!regions_needed) { 3682103cf9cSPeter Xu nrg = get_file_region_entry_from_cache(map, from, to); 3692103cf9cSPeter Xu record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg); 3702103cf9cSPeter Xu list_add(&nrg->link, rg->link.prev); 3712103cf9cSPeter Xu coalesce_file_region(map, nrg); 3722103cf9cSPeter Xu } else 3732103cf9cSPeter Xu *regions_needed += 1; 3742103cf9cSPeter Xu 3752103cf9cSPeter Xu return to - from; 3762103cf9cSPeter Xu } 3772103cf9cSPeter Xu 378972a3da3SWei Yang /* 379972a3da3SWei Yang * Must be called with resv->lock held. 380972a3da3SWei Yang * 381972a3da3SWei Yang * Calling this with regions_needed != NULL will count the number of pages 382972a3da3SWei Yang * to be added but will not modify the linked list. And regions_needed will 383972a3da3SWei Yang * indicate the number of file_regions needed in the cache to carry out to add 384972a3da3SWei Yang * the regions for this range. 385d75c6af9SMina Almasry */ 386d75c6af9SMina Almasry static long add_reservation_in_range(struct resv_map *resv, long f, long t, 387075a61d0SMina Almasry struct hugetlb_cgroup *h_cg, 388972a3da3SWei Yang struct hstate *h, long *regions_needed) 389d75c6af9SMina Almasry { 3900db9d74eSMina Almasry long add = 0; 391d75c6af9SMina Almasry struct list_head *head = &resv->regions; 3920db9d74eSMina Almasry long last_accounted_offset = f; 3932103cf9cSPeter Xu struct file_region *rg = NULL, *trg = NULL; 394d75c6af9SMina Almasry 3950db9d74eSMina Almasry if (regions_needed) 3960db9d74eSMina Almasry *regions_needed = 0; 397d75c6af9SMina Almasry 3980db9d74eSMina Almasry /* In this loop, we essentially handle an entry for the range 3990db9d74eSMina Almasry * [last_accounted_offset, rg->from), at every iteration, with some 4000db9d74eSMina Almasry * bounds checking. 4010db9d74eSMina Almasry */ 4020db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, head, link) { 4030db9d74eSMina Almasry /* Skip irrelevant regions that start before our range. */ 4040db9d74eSMina Almasry if (rg->from < f) { 4050db9d74eSMina Almasry /* If this region ends after the last accounted offset, 4060db9d74eSMina Almasry * then we need to update last_accounted_offset. 4070db9d74eSMina Almasry */ 4080db9d74eSMina Almasry if (rg->to > last_accounted_offset) 4090db9d74eSMina Almasry last_accounted_offset = rg->to; 4100db9d74eSMina Almasry continue; 4110db9d74eSMina Almasry } 412d75c6af9SMina Almasry 4130db9d74eSMina Almasry /* When we find a region that starts beyond our range, we've 4140db9d74eSMina Almasry * finished. 4150db9d74eSMina Almasry */ 416ca7e0457SPeter Xu if (rg->from >= t) 417d75c6af9SMina Almasry break; 418d75c6af9SMina Almasry 4190db9d74eSMina Almasry /* Add an entry for last_accounted_offset -> rg->from, and 4200db9d74eSMina Almasry * update last_accounted_offset. 421d75c6af9SMina Almasry */ 4222103cf9cSPeter Xu if (rg->from > last_accounted_offset) 4232103cf9cSPeter Xu add += hugetlb_resv_map_add(resv, rg, 4242103cf9cSPeter Xu last_accounted_offset, 4252103cf9cSPeter Xu rg->from, h, h_cg, 4262103cf9cSPeter Xu regions_needed); 427d75c6af9SMina Almasry 4280db9d74eSMina Almasry last_accounted_offset = rg->to; 4290db9d74eSMina Almasry } 4300db9d74eSMina Almasry 4310db9d74eSMina Almasry /* Handle the case where our range extends beyond 4320db9d74eSMina Almasry * last_accounted_offset. 4330db9d74eSMina Almasry */ 4342103cf9cSPeter Xu if (last_accounted_offset < t) 4352103cf9cSPeter Xu add += hugetlb_resv_map_add(resv, rg, last_accounted_offset, 4362103cf9cSPeter Xu t, h, h_cg, regions_needed); 4370db9d74eSMina Almasry 4380db9d74eSMina Almasry VM_BUG_ON(add < 0); 4390db9d74eSMina Almasry return add; 4400db9d74eSMina Almasry } 4410db9d74eSMina Almasry 4420db9d74eSMina Almasry /* Must be called with resv->lock acquired. Will drop lock to allocate entries. 4430db9d74eSMina Almasry */ 4440db9d74eSMina Almasry static int allocate_file_region_entries(struct resv_map *resv, 4450db9d74eSMina Almasry int regions_needed) 4460db9d74eSMina Almasry __must_hold(&resv->lock) 4470db9d74eSMina Almasry { 4480db9d74eSMina Almasry struct list_head allocated_regions; 4490db9d74eSMina Almasry int to_allocate = 0, i = 0; 4500db9d74eSMina Almasry struct file_region *trg = NULL, *rg = NULL; 4510db9d74eSMina Almasry 4520db9d74eSMina Almasry VM_BUG_ON(regions_needed < 0); 4530db9d74eSMina Almasry 4540db9d74eSMina Almasry INIT_LIST_HEAD(&allocated_regions); 4550db9d74eSMina Almasry 4560db9d74eSMina Almasry /* 4570db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 4580db9d74eSMina Almasry * the number of in progress add operations plus regions_needed. 4590db9d74eSMina Almasry * 4600db9d74eSMina Almasry * This is a while loop because when we drop the lock, some other call 4610db9d74eSMina Almasry * to region_add or region_del may have consumed some region_entries, 4620db9d74eSMina Almasry * so we keep looping here until we finally have enough entries for 4630db9d74eSMina Almasry * (adds_in_progress + regions_needed). 4640db9d74eSMina Almasry */ 4650db9d74eSMina Almasry while (resv->region_cache_count < 4660db9d74eSMina Almasry (resv->adds_in_progress + regions_needed)) { 4670db9d74eSMina Almasry to_allocate = resv->adds_in_progress + regions_needed - 4680db9d74eSMina Almasry resv->region_cache_count; 4690db9d74eSMina Almasry 4700db9d74eSMina Almasry /* At this point, we should have enough entries in the cache 471f0953a1bSIngo Molnar * for all the existing adds_in_progress. We should only be 4720db9d74eSMina Almasry * needing to allocate for regions_needed. 4730db9d74eSMina Almasry */ 4740db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); 4750db9d74eSMina Almasry 4760db9d74eSMina Almasry spin_unlock(&resv->lock); 4770db9d74eSMina Almasry for (i = 0; i < to_allocate; i++) { 4780db9d74eSMina Almasry trg = kmalloc(sizeof(*trg), GFP_KERNEL); 4790db9d74eSMina Almasry if (!trg) 4800db9d74eSMina Almasry goto out_of_memory; 4810db9d74eSMina Almasry list_add(&trg->link, &allocated_regions); 4820db9d74eSMina Almasry } 4830db9d74eSMina Almasry 4840db9d74eSMina Almasry spin_lock(&resv->lock); 4850db9d74eSMina Almasry 486d3ec7b6eSWei Yang list_splice(&allocated_regions, &resv->region_cache); 487d3ec7b6eSWei Yang resv->region_cache_count += to_allocate; 4880db9d74eSMina Almasry } 4890db9d74eSMina Almasry 4900db9d74eSMina Almasry return 0; 4910db9d74eSMina Almasry 4920db9d74eSMina Almasry out_of_memory: 4930db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, &allocated_regions, link) { 494d75c6af9SMina Almasry list_del(&rg->link); 495d75c6af9SMina Almasry kfree(rg); 496d75c6af9SMina Almasry } 4970db9d74eSMina Almasry return -ENOMEM; 498d75c6af9SMina Almasry } 499d75c6af9SMina Almasry 5001dd308a7SMike Kravetz /* 5011dd308a7SMike Kravetz * Add the huge page range represented by [f, t) to the reserve 5020db9d74eSMina Almasry * map. Regions will be taken from the cache to fill in this range. 5030db9d74eSMina Almasry * Sufficient regions should exist in the cache due to the previous 5040db9d74eSMina Almasry * call to region_chg with the same range, but in some cases the cache will not 5050db9d74eSMina Almasry * have sufficient entries due to races with other code doing region_add or 5060db9d74eSMina Almasry * region_del. The extra needed entries will be allocated. 507cf3ad20bSMike Kravetz * 5080db9d74eSMina Almasry * regions_needed is the out value provided by a previous call to region_chg. 5090db9d74eSMina Almasry * 5100db9d74eSMina Almasry * Return the number of new huge pages added to the map. This number is greater 5110db9d74eSMina Almasry * than or equal to zero. If file_region entries needed to be allocated for 5127c8de358SEthon Paul * this operation and we were not able to allocate, it returns -ENOMEM. 5130db9d74eSMina Almasry * region_add of regions of length 1 never allocate file_regions and cannot 5140db9d74eSMina Almasry * fail; region_chg will always allocate at least 1 entry and a region_add for 5150db9d74eSMina Almasry * 1 page will only require at most 1 entry. 5161dd308a7SMike Kravetz */ 5170db9d74eSMina Almasry static long region_add(struct resv_map *resv, long f, long t, 518075a61d0SMina Almasry long in_regions_needed, struct hstate *h, 519075a61d0SMina Almasry struct hugetlb_cgroup *h_cg) 52096822904SAndy Whitcroft { 5210db9d74eSMina Almasry long add = 0, actual_regions_needed = 0; 52296822904SAndy Whitcroft 5237b24d861SDavidlohr Bueso spin_lock(&resv->lock); 5240db9d74eSMina Almasry retry: 5250db9d74eSMina Almasry 5260db9d74eSMina Almasry /* Count how many regions are actually needed to execute this add. */ 527972a3da3SWei Yang add_reservation_in_range(resv, f, t, NULL, NULL, 528972a3da3SWei Yang &actual_regions_needed); 52996822904SAndy Whitcroft 5305e911373SMike Kravetz /* 5310db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 5320db9d74eSMina Almasry * this add operation. Note that actual_regions_needed may be greater 5330db9d74eSMina Almasry * than in_regions_needed, as the resv_map may have been modified since 5340db9d74eSMina Almasry * the region_chg call. In this case, we need to make sure that we 5350db9d74eSMina Almasry * allocate extra entries, such that we have enough for all the 5360db9d74eSMina Almasry * existing adds_in_progress, plus the excess needed for this 5370db9d74eSMina Almasry * operation. 5385e911373SMike Kravetz */ 5390db9d74eSMina Almasry if (actual_regions_needed > in_regions_needed && 5400db9d74eSMina Almasry resv->region_cache_count < 5410db9d74eSMina Almasry resv->adds_in_progress + 5420db9d74eSMina Almasry (actual_regions_needed - in_regions_needed)) { 5430db9d74eSMina Almasry /* region_add operation of range 1 should never need to 5440db9d74eSMina Almasry * allocate file_region entries. 5450db9d74eSMina Almasry */ 5460db9d74eSMina Almasry VM_BUG_ON(t - f <= 1); 5475e911373SMike Kravetz 5480db9d74eSMina Almasry if (allocate_file_region_entries( 5490db9d74eSMina Almasry resv, actual_regions_needed - in_regions_needed)) { 5500db9d74eSMina Almasry return -ENOMEM; 5515e911373SMike Kravetz } 5525e911373SMike Kravetz 5530db9d74eSMina Almasry goto retry; 5540db9d74eSMina Almasry } 555cf3ad20bSMike Kravetz 556972a3da3SWei Yang add = add_reservation_in_range(resv, f, t, h_cg, h, NULL); 5570db9d74eSMina Almasry 5580db9d74eSMina Almasry resv->adds_in_progress -= in_regions_needed; 5590db9d74eSMina Almasry 5607b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 561cf3ad20bSMike Kravetz return add; 56296822904SAndy Whitcroft } 56396822904SAndy Whitcroft 5641dd308a7SMike Kravetz /* 5651dd308a7SMike Kravetz * Examine the existing reserve map and determine how many 5661dd308a7SMike Kravetz * huge pages in the specified range [f, t) are NOT currently 5671dd308a7SMike Kravetz * represented. This routine is called before a subsequent 5681dd308a7SMike Kravetz * call to region_add that will actually modify the reserve 5691dd308a7SMike Kravetz * map to add the specified range [f, t). region_chg does 5701dd308a7SMike Kravetz * not change the number of huge pages represented by the 5710db9d74eSMina Almasry * map. A number of new file_region structures is added to the cache as a 5720db9d74eSMina Almasry * placeholder, for the subsequent region_add call to use. At least 1 5730db9d74eSMina Almasry * file_region structure is added. 5740db9d74eSMina Almasry * 5750db9d74eSMina Almasry * out_regions_needed is the number of regions added to the 5760db9d74eSMina Almasry * resv->adds_in_progress. This value needs to be provided to a follow up call 5770db9d74eSMina Almasry * to region_add or region_abort for proper accounting. 5785e911373SMike Kravetz * 5795e911373SMike Kravetz * Returns the number of huge pages that need to be added to the existing 5805e911373SMike Kravetz * reservation map for the range [f, t). This number is greater or equal to 5815e911373SMike Kravetz * zero. -ENOMEM is returned if a new file_region structure or cache entry 5825e911373SMike Kravetz * is needed and can not be allocated. 5831dd308a7SMike Kravetz */ 5840db9d74eSMina Almasry static long region_chg(struct resv_map *resv, long f, long t, 5850db9d74eSMina Almasry long *out_regions_needed) 58696822904SAndy Whitcroft { 58796822904SAndy Whitcroft long chg = 0; 58896822904SAndy Whitcroft 5897b24d861SDavidlohr Bueso spin_lock(&resv->lock); 5905e911373SMike Kravetz 591972a3da3SWei Yang /* Count how many hugepages in this range are NOT represented. */ 592075a61d0SMina Almasry chg = add_reservation_in_range(resv, f, t, NULL, NULL, 593972a3da3SWei Yang out_regions_needed); 5945e911373SMike Kravetz 5950db9d74eSMina Almasry if (*out_regions_needed == 0) 5960db9d74eSMina Almasry *out_regions_needed = 1; 5975e911373SMike Kravetz 5980db9d74eSMina Almasry if (allocate_file_region_entries(resv, *out_regions_needed)) 5995e911373SMike Kravetz return -ENOMEM; 6005e911373SMike Kravetz 6010db9d74eSMina Almasry resv->adds_in_progress += *out_regions_needed; 60296822904SAndy Whitcroft 6037b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 60496822904SAndy Whitcroft return chg; 60596822904SAndy Whitcroft } 60696822904SAndy Whitcroft 6071dd308a7SMike Kravetz /* 6085e911373SMike Kravetz * Abort the in progress add operation. The adds_in_progress field 6095e911373SMike Kravetz * of the resv_map keeps track of the operations in progress between 6105e911373SMike Kravetz * calls to region_chg and region_add. Operations are sometimes 6115e911373SMike Kravetz * aborted after the call to region_chg. In such cases, region_abort 6120db9d74eSMina Almasry * is called to decrement the adds_in_progress counter. regions_needed 6130db9d74eSMina Almasry * is the value returned by the region_chg call, it is used to decrement 6140db9d74eSMina Almasry * the adds_in_progress counter. 6155e911373SMike Kravetz * 6165e911373SMike Kravetz * NOTE: The range arguments [f, t) are not needed or used in this 6175e911373SMike Kravetz * routine. They are kept to make reading the calling code easier as 6185e911373SMike Kravetz * arguments will match the associated region_chg call. 6195e911373SMike Kravetz */ 6200db9d74eSMina Almasry static void region_abort(struct resv_map *resv, long f, long t, 6210db9d74eSMina Almasry long regions_needed) 6225e911373SMike Kravetz { 6235e911373SMike Kravetz spin_lock(&resv->lock); 6245e911373SMike Kravetz VM_BUG_ON(!resv->region_cache_count); 6250db9d74eSMina Almasry resv->adds_in_progress -= regions_needed; 6265e911373SMike Kravetz spin_unlock(&resv->lock); 6275e911373SMike Kravetz } 6285e911373SMike Kravetz 6295e911373SMike Kravetz /* 630feba16e2SMike Kravetz * Delete the specified range [f, t) from the reserve map. If the 631feba16e2SMike Kravetz * t parameter is LONG_MAX, this indicates that ALL regions after f 632feba16e2SMike Kravetz * should be deleted. Locate the regions which intersect [f, t) 633feba16e2SMike Kravetz * and either trim, delete or split the existing regions. 634feba16e2SMike Kravetz * 635feba16e2SMike Kravetz * Returns the number of huge pages deleted from the reserve map. 636feba16e2SMike Kravetz * In the normal case, the return value is zero or more. In the 637feba16e2SMike Kravetz * case where a region must be split, a new region descriptor must 638feba16e2SMike Kravetz * be allocated. If the allocation fails, -ENOMEM will be returned. 639feba16e2SMike Kravetz * NOTE: If the parameter t == LONG_MAX, then we will never split 640feba16e2SMike Kravetz * a region and possibly return -ENOMEM. Callers specifying 641feba16e2SMike Kravetz * t == LONG_MAX do not need to check for -ENOMEM error. 6421dd308a7SMike Kravetz */ 643feba16e2SMike Kravetz static long region_del(struct resv_map *resv, long f, long t) 64496822904SAndy Whitcroft { 6451406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 64696822904SAndy Whitcroft struct file_region *rg, *trg; 647feba16e2SMike Kravetz struct file_region *nrg = NULL; 648feba16e2SMike Kravetz long del = 0; 64996822904SAndy Whitcroft 650feba16e2SMike Kravetz retry: 6517b24d861SDavidlohr Bueso spin_lock(&resv->lock); 652feba16e2SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 653dbe409e4SMike Kravetz /* 654dbe409e4SMike Kravetz * Skip regions before the range to be deleted. file_region 655dbe409e4SMike Kravetz * ranges are normally of the form [from, to). However, there 656dbe409e4SMike Kravetz * may be a "placeholder" entry in the map which is of the form 657dbe409e4SMike Kravetz * (from, to) with from == to. Check for placeholder entries 658dbe409e4SMike Kravetz * at the beginning of the range to be deleted. 659dbe409e4SMike Kravetz */ 660dbe409e4SMike Kravetz if (rg->to <= f && (rg->to != rg->from || rg->to != f)) 661feba16e2SMike Kravetz continue; 662dbe409e4SMike Kravetz 663feba16e2SMike Kravetz if (rg->from >= t) 66496822904SAndy Whitcroft break; 66596822904SAndy Whitcroft 666feba16e2SMike Kravetz if (f > rg->from && t < rg->to) { /* Must split region */ 667feba16e2SMike Kravetz /* 668feba16e2SMike Kravetz * Check for an entry in the cache before dropping 669feba16e2SMike Kravetz * lock and attempting allocation. 670feba16e2SMike Kravetz */ 671feba16e2SMike Kravetz if (!nrg && 672feba16e2SMike Kravetz resv->region_cache_count > resv->adds_in_progress) { 673feba16e2SMike Kravetz nrg = list_first_entry(&resv->region_cache, 674feba16e2SMike Kravetz struct file_region, 675feba16e2SMike Kravetz link); 676feba16e2SMike Kravetz list_del(&nrg->link); 677feba16e2SMike Kravetz resv->region_cache_count--; 67896822904SAndy Whitcroft } 67996822904SAndy Whitcroft 680feba16e2SMike Kravetz if (!nrg) { 681feba16e2SMike Kravetz spin_unlock(&resv->lock); 682feba16e2SMike Kravetz nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 683feba16e2SMike Kravetz if (!nrg) 684feba16e2SMike Kravetz return -ENOMEM; 685feba16e2SMike Kravetz goto retry; 686feba16e2SMike Kravetz } 687feba16e2SMike Kravetz 688feba16e2SMike Kravetz del += t - f; 68979aa925bSMike Kravetz hugetlb_cgroup_uncharge_file_region( 690d85aecf2SMiaohe Lin resv, rg, t - f, false); 691feba16e2SMike Kravetz 692feba16e2SMike Kravetz /* New entry for end of split region */ 693feba16e2SMike Kravetz nrg->from = t; 694feba16e2SMike Kravetz nrg->to = rg->to; 695075a61d0SMina Almasry 696075a61d0SMina Almasry copy_hugetlb_cgroup_uncharge_info(nrg, rg); 697075a61d0SMina Almasry 698feba16e2SMike Kravetz INIT_LIST_HEAD(&nrg->link); 699feba16e2SMike Kravetz 700feba16e2SMike Kravetz /* Original entry is trimmed */ 701feba16e2SMike Kravetz rg->to = f; 702feba16e2SMike Kravetz 703feba16e2SMike Kravetz list_add(&nrg->link, &rg->link); 704feba16e2SMike Kravetz nrg = NULL; 70596822904SAndy Whitcroft break; 706feba16e2SMike Kravetz } 707feba16e2SMike Kravetz 708feba16e2SMike Kravetz if (f <= rg->from && t >= rg->to) { /* Remove entire region */ 709feba16e2SMike Kravetz del += rg->to - rg->from; 710075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 711d85aecf2SMiaohe Lin rg->to - rg->from, true); 71296822904SAndy Whitcroft list_del(&rg->link); 71396822904SAndy Whitcroft kfree(rg); 714feba16e2SMike Kravetz continue; 71596822904SAndy Whitcroft } 7167b24d861SDavidlohr Bueso 717feba16e2SMike Kravetz if (f <= rg->from) { /* Trim beginning of region */ 718075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 719d85aecf2SMiaohe Lin t - rg->from, false); 720075a61d0SMina Almasry 72179aa925bSMike Kravetz del += t - rg->from; 72279aa925bSMike Kravetz rg->from = t; 72379aa925bSMike Kravetz } else { /* Trim end of region */ 724075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 725d85aecf2SMiaohe Lin rg->to - f, false); 72679aa925bSMike Kravetz 72779aa925bSMike Kravetz del += rg->to - f; 72879aa925bSMike Kravetz rg->to = f; 729feba16e2SMike Kravetz } 730feba16e2SMike Kravetz } 731feba16e2SMike Kravetz 7327b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 733feba16e2SMike Kravetz kfree(nrg); 734feba16e2SMike Kravetz return del; 73596822904SAndy Whitcroft } 73696822904SAndy Whitcroft 7371dd308a7SMike Kravetz /* 738b5cec28dSMike Kravetz * A rare out of memory error was encountered which prevented removal of 739b5cec28dSMike Kravetz * the reserve map region for a page. The huge page itself was free'ed 740b5cec28dSMike Kravetz * and removed from the page cache. This routine will adjust the subpool 741b5cec28dSMike Kravetz * usage count, and the global reserve count if needed. By incrementing 742b5cec28dSMike Kravetz * these counts, the reserve map entry which could not be deleted will 743b5cec28dSMike Kravetz * appear as a "reserved" entry instead of simply dangling with incorrect 744b5cec28dSMike Kravetz * counts. 745b5cec28dSMike Kravetz */ 74672e2936cSzhong jiang void hugetlb_fix_reserve_counts(struct inode *inode) 747b5cec28dSMike Kravetz { 748b5cec28dSMike Kravetz struct hugepage_subpool *spool = subpool_inode(inode); 749b5cec28dSMike Kravetz long rsv_adjust; 750da56388cSMiaohe Lin bool reserved = false; 751b5cec28dSMike Kravetz 752b5cec28dSMike Kravetz rsv_adjust = hugepage_subpool_get_pages(spool, 1); 753da56388cSMiaohe Lin if (rsv_adjust > 0) { 754b5cec28dSMike Kravetz struct hstate *h = hstate_inode(inode); 755b5cec28dSMike Kravetz 756da56388cSMiaohe Lin if (!hugetlb_acct_memory(h, 1)) 757da56388cSMiaohe Lin reserved = true; 758da56388cSMiaohe Lin } else if (!rsv_adjust) { 759da56388cSMiaohe Lin reserved = true; 760b5cec28dSMike Kravetz } 761da56388cSMiaohe Lin 762da56388cSMiaohe Lin if (!reserved) 763da56388cSMiaohe Lin pr_warn("hugetlb: Huge Page Reserved count may go negative.\n"); 764b5cec28dSMike Kravetz } 765b5cec28dSMike Kravetz 766b5cec28dSMike Kravetz /* 7671dd308a7SMike Kravetz * Count and return the number of huge pages in the reserve map 7681dd308a7SMike Kravetz * that intersect with the range [f, t). 7691dd308a7SMike Kravetz */ 7701406ec9bSJoonsoo Kim static long region_count(struct resv_map *resv, long f, long t) 77184afd99bSAndy Whitcroft { 7721406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 77384afd99bSAndy Whitcroft struct file_region *rg; 77484afd99bSAndy Whitcroft long chg = 0; 77584afd99bSAndy Whitcroft 7767b24d861SDavidlohr Bueso spin_lock(&resv->lock); 77784afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 77884afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 779f2135a4aSWang Sheng-Hui long seg_from; 780f2135a4aSWang Sheng-Hui long seg_to; 78184afd99bSAndy Whitcroft 78284afd99bSAndy Whitcroft if (rg->to <= f) 78384afd99bSAndy Whitcroft continue; 78484afd99bSAndy Whitcroft if (rg->from >= t) 78584afd99bSAndy Whitcroft break; 78684afd99bSAndy Whitcroft 78784afd99bSAndy Whitcroft seg_from = max(rg->from, f); 78884afd99bSAndy Whitcroft seg_to = min(rg->to, t); 78984afd99bSAndy Whitcroft 79084afd99bSAndy Whitcroft chg += seg_to - seg_from; 79184afd99bSAndy Whitcroft } 7927b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 79384afd99bSAndy Whitcroft 79484afd99bSAndy Whitcroft return chg; 79584afd99bSAndy Whitcroft } 79684afd99bSAndy Whitcroft 79796822904SAndy Whitcroft /* 798e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 799e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 800e7c4b0bfSAndy Whitcroft */ 801a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 802a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 803e7c4b0bfSAndy Whitcroft { 804a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 805a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 806e7c4b0bfSAndy Whitcroft } 807e7c4b0bfSAndy Whitcroft 8080fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 8090fe6e20bSNaoya Horiguchi unsigned long address) 8100fe6e20bSNaoya Horiguchi { 8110fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 8120fe6e20bSNaoya Horiguchi } 813dee41079SDan Williams EXPORT_SYMBOL_GPL(linear_hugepage_index); 8140fe6e20bSNaoya Horiguchi 81584afd99bSAndy Whitcroft /* 81608fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 81708fba699SMel Gorman * cases this will be same size as used by the page table entries. 81808fba699SMel Gorman */ 81908fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 82008fba699SMel Gorman { 82105ea8860SDan Williams if (vma->vm_ops && vma->vm_ops->pagesize) 82205ea8860SDan Williams return vma->vm_ops->pagesize(vma); 82308fba699SMel Gorman return PAGE_SIZE; 82408fba699SMel Gorman } 825f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 82608fba699SMel Gorman 82708fba699SMel Gorman /* 8283340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 8293340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 83009135cc5SDan Williams * architectures where it differs, an architecture-specific 'strong' 83109135cc5SDan Williams * version of this symbol is required. 8323340289dSMel Gorman */ 83309135cc5SDan Williams __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 8343340289dSMel Gorman { 8353340289dSMel Gorman return vma_kernel_pagesize(vma); 8363340289dSMel Gorman } 8373340289dSMel Gorman 8383340289dSMel Gorman /* 83984afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 84084afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 84184afd99bSAndy Whitcroft * alignment. 84284afd99bSAndy Whitcroft */ 84384afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 84484afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 84504f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 84684afd99bSAndy Whitcroft 847a1e78772SMel Gorman /* 848a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 849a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 850a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 851a1e78772SMel Gorman * 852a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 853a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 854a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 855a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 85684afd99bSAndy Whitcroft * 85784afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 85884afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 85984afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 86084afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 86184afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 86284afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 86384afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 86484afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 865a1e78772SMel Gorman */ 866e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 867e7c4b0bfSAndy Whitcroft { 868e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 869e7c4b0bfSAndy Whitcroft } 870e7c4b0bfSAndy Whitcroft 871e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 872e7c4b0bfSAndy Whitcroft unsigned long value) 873e7c4b0bfSAndy Whitcroft { 874e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 875e7c4b0bfSAndy Whitcroft } 876e7c4b0bfSAndy Whitcroft 877e9fe92aeSMina Almasry static void 878e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, 879e9fe92aeSMina Almasry struct hugetlb_cgroup *h_cg, 880e9fe92aeSMina Almasry struct hstate *h) 881e9fe92aeSMina Almasry { 882e9fe92aeSMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 883e9fe92aeSMina Almasry if (!h_cg || !h) { 884e9fe92aeSMina Almasry resv_map->reservation_counter = NULL; 885e9fe92aeSMina Almasry resv_map->pages_per_hpage = 0; 886e9fe92aeSMina Almasry resv_map->css = NULL; 887e9fe92aeSMina Almasry } else { 888e9fe92aeSMina Almasry resv_map->reservation_counter = 889e9fe92aeSMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 890e9fe92aeSMina Almasry resv_map->pages_per_hpage = pages_per_huge_page(h); 891e9fe92aeSMina Almasry resv_map->css = &h_cg->css; 892e9fe92aeSMina Almasry } 893e9fe92aeSMina Almasry #endif 894e9fe92aeSMina Almasry } 895e9fe92aeSMina Almasry 8969119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 89784afd99bSAndy Whitcroft { 89884afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 8995e911373SMike Kravetz struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 9005e911373SMike Kravetz 9015e911373SMike Kravetz if (!resv_map || !rg) { 9025e911373SMike Kravetz kfree(resv_map); 9035e911373SMike Kravetz kfree(rg); 90484afd99bSAndy Whitcroft return NULL; 9055e911373SMike Kravetz } 90684afd99bSAndy Whitcroft 90784afd99bSAndy Whitcroft kref_init(&resv_map->refs); 9087b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 90984afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 91084afd99bSAndy Whitcroft 9115e911373SMike Kravetz resv_map->adds_in_progress = 0; 912e9fe92aeSMina Almasry /* 913e9fe92aeSMina Almasry * Initialize these to 0. On shared mappings, 0's here indicate these 914e9fe92aeSMina Almasry * fields don't do cgroup accounting. On private mappings, these will be 915e9fe92aeSMina Almasry * re-initialized to the proper values, to indicate that hugetlb cgroup 916e9fe92aeSMina Almasry * reservations are to be un-charged from here. 917e9fe92aeSMina Almasry */ 918e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); 9195e911373SMike Kravetz 9205e911373SMike Kravetz INIT_LIST_HEAD(&resv_map->region_cache); 9215e911373SMike Kravetz list_add(&rg->link, &resv_map->region_cache); 9225e911373SMike Kravetz resv_map->region_cache_count = 1; 9235e911373SMike Kravetz 92484afd99bSAndy Whitcroft return resv_map; 92584afd99bSAndy Whitcroft } 92684afd99bSAndy Whitcroft 9279119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 92884afd99bSAndy Whitcroft { 92984afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 9305e911373SMike Kravetz struct list_head *head = &resv_map->region_cache; 9315e911373SMike Kravetz struct file_region *rg, *trg; 93284afd99bSAndy Whitcroft 93384afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 934feba16e2SMike Kravetz region_del(resv_map, 0, LONG_MAX); 9355e911373SMike Kravetz 9365e911373SMike Kravetz /* ... and any entries left in the cache */ 9375e911373SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 9385e911373SMike Kravetz list_del(&rg->link); 9395e911373SMike Kravetz kfree(rg); 9405e911373SMike Kravetz } 9415e911373SMike Kravetz 9425e911373SMike Kravetz VM_BUG_ON(resv_map->adds_in_progress); 9435e911373SMike Kravetz 94484afd99bSAndy Whitcroft kfree(resv_map); 94584afd99bSAndy Whitcroft } 94684afd99bSAndy Whitcroft 9474e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 9484e35f483SJoonsoo Kim { 949f27a5136SMike Kravetz /* 950f27a5136SMike Kravetz * At inode evict time, i_mapping may not point to the original 951f27a5136SMike Kravetz * address space within the inode. This original address space 952f27a5136SMike Kravetz * contains the pointer to the resv_map. So, always use the 953f27a5136SMike Kravetz * address space embedded within the inode. 954f27a5136SMike Kravetz * The VERY common case is inode->mapping == &inode->i_data but, 955f27a5136SMike Kravetz * this may not be true for device special inodes. 956f27a5136SMike Kravetz */ 957f27a5136SMike Kravetz return (struct resv_map *)(&inode->i_data)->private_data; 9584e35f483SJoonsoo Kim } 9594e35f483SJoonsoo Kim 96084afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 961a1e78772SMel Gorman { 96281d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 9634e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 9644e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 9654e35f483SJoonsoo Kim struct inode *inode = mapping->host; 9664e35f483SJoonsoo Kim 9674e35f483SJoonsoo Kim return inode_resv_map(inode); 9684e35f483SJoonsoo Kim 9694e35f483SJoonsoo Kim } else { 97084afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 97184afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 9724e35f483SJoonsoo Kim } 973a1e78772SMel Gorman } 974a1e78772SMel Gorman 97584afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 976a1e78772SMel Gorman { 97781d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 97881d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 979a1e78772SMel Gorman 98084afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 98184afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 98204f2cbe3SMel Gorman } 98304f2cbe3SMel Gorman 98404f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 98504f2cbe3SMel Gorman { 98681d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 98781d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 988e7c4b0bfSAndy Whitcroft 989e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 99004f2cbe3SMel Gorman } 99104f2cbe3SMel Gorman 99204f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 99304f2cbe3SMel Gorman { 99481d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 995e7c4b0bfSAndy Whitcroft 996e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 997a1e78772SMel Gorman } 998a1e78772SMel Gorman 99904f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 1000a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 1001a1e78772SMel Gorman { 100281d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 1003f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 1004a1e78772SMel Gorman vma->vm_private_data = (void *)0; 1005a1e78772SMel Gorman } 1006a1e78772SMel Gorman 1007a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 1008559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 1009a1e78772SMel Gorman { 1010af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 1011af0ed73eSJoonsoo Kim /* 1012af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 1013af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 1014af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 1015af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 1016af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 1017af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 1018af0ed73eSJoonsoo Kim * properly, so add work-around here. 1019af0ed73eSJoonsoo Kim */ 1020af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 1021559ec2f8SNicholas Krause return true; 1022af0ed73eSJoonsoo Kim else 1023559ec2f8SNicholas Krause return false; 1024af0ed73eSJoonsoo Kim } 1025a63884e9SJoonsoo Kim 1026a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 10271fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 10281fb1b0e9SMike Kravetz /* 10291fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 10301fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 10311fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 10327c8de358SEthon Paul * fallocate. In this case, there really are no reserves to 10331fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 10341fb1b0e9SMike Kravetz */ 10351fb1b0e9SMike Kravetz if (chg) 10361fb1b0e9SMike Kravetz return false; 10371fb1b0e9SMike Kravetz else 1038559ec2f8SNicholas Krause return true; 10391fb1b0e9SMike Kravetz } 1040a63884e9SJoonsoo Kim 1041a63884e9SJoonsoo Kim /* 1042a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 1043a63884e9SJoonsoo Kim * private mappings. 1044a63884e9SJoonsoo Kim */ 104567961f9dSMike Kravetz if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 104667961f9dSMike Kravetz /* 104767961f9dSMike Kravetz * Like the shared case above, a hole punch or truncate 104867961f9dSMike Kravetz * could have been performed on the private mapping. 104967961f9dSMike Kravetz * Examine the value of chg to determine if reserves 105067961f9dSMike Kravetz * actually exist or were previously consumed. 105167961f9dSMike Kravetz * Very Subtle - The value of chg comes from a previous 105267961f9dSMike Kravetz * call to vma_needs_reserves(). The reserve map for 105367961f9dSMike Kravetz * private mappings has different (opposite) semantics 105467961f9dSMike Kravetz * than that of shared mappings. vma_needs_reserves() 105567961f9dSMike Kravetz * has already taken this difference in semantics into 105667961f9dSMike Kravetz * account. Therefore, the meaning of chg is the same 105767961f9dSMike Kravetz * as in the shared case above. Code could easily be 105867961f9dSMike Kravetz * combined, but keeping it separate draws attention to 105967961f9dSMike Kravetz * subtle differences. 106067961f9dSMike Kravetz */ 106167961f9dSMike Kravetz if (chg) 106267961f9dSMike Kravetz return false; 106367961f9dSMike Kravetz else 1064559ec2f8SNicholas Krause return true; 106567961f9dSMike Kravetz } 1066a63884e9SJoonsoo Kim 1067559ec2f8SNicholas Krause return false; 1068a1e78772SMel Gorman } 1069a1e78772SMel Gorman 1070a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 10711da177e4SLinus Torvalds { 10721da177e4SLinus Torvalds int nid = page_to_nid(page); 10739487ca60SMike Kravetz 10749487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1075b65a4edaSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1076b65a4edaSMike Kravetz 10770edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 1078a5516438SAndi Kleen h->free_huge_pages++; 1079a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 10806c037149SMike Kravetz SetHPageFreed(page); 10811da177e4SLinus Torvalds } 10821da177e4SLinus Torvalds 108394310cbcSAnshuman Khandual static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) 1084bf50bab2SNaoya Horiguchi { 1085bf50bab2SNaoya Horiguchi struct page *page; 10861a08ae36SPavel Tatashin bool pin = !!(current->flags & PF_MEMALLOC_PIN); 1087bf50bab2SNaoya Horiguchi 10889487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1089bbe88753SJoonsoo Kim list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { 10908e3560d9SPavel Tatashin if (pin && !is_pinnable_page(page)) 1091bbe88753SJoonsoo Kim continue; 1092bbe88753SJoonsoo Kim 10936664bfc8SWei Yang if (PageHWPoison(page)) 10946664bfc8SWei Yang continue; 1095bbe88753SJoonsoo Kim 10960edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 1097a9869b83SNaoya Horiguchi set_page_refcounted(page); 10986c037149SMike Kravetz ClearHPageFreed(page); 1099bf50bab2SNaoya Horiguchi h->free_huge_pages--; 1100bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 1101bf50bab2SNaoya Horiguchi return page; 1102bf50bab2SNaoya Horiguchi } 1103bf50bab2SNaoya Horiguchi 11046664bfc8SWei Yang return NULL; 11056664bfc8SWei Yang } 11066664bfc8SWei Yang 11073e59fcb0SMichal Hocko static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, 11083e59fcb0SMichal Hocko nodemask_t *nmask) 110994310cbcSAnshuman Khandual { 11103e59fcb0SMichal Hocko unsigned int cpuset_mems_cookie; 11113e59fcb0SMichal Hocko struct zonelist *zonelist; 11123e59fcb0SMichal Hocko struct zone *zone; 11133e59fcb0SMichal Hocko struct zoneref *z; 111498fa15f3SAnshuman Khandual int node = NUMA_NO_NODE; 11153e59fcb0SMichal Hocko 11163e59fcb0SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 11173e59fcb0SMichal Hocko 11183e59fcb0SMichal Hocko retry_cpuset: 11193e59fcb0SMichal Hocko cpuset_mems_cookie = read_mems_allowed_begin(); 11203e59fcb0SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 112194310cbcSAnshuman Khandual struct page *page; 112294310cbcSAnshuman Khandual 11233e59fcb0SMichal Hocko if (!cpuset_zone_allowed(zone, gfp_mask)) 11243e59fcb0SMichal Hocko continue; 11253e59fcb0SMichal Hocko /* 11263e59fcb0SMichal Hocko * no need to ask again on the same node. Pool is node rather than 11273e59fcb0SMichal Hocko * zone aware 11283e59fcb0SMichal Hocko */ 11293e59fcb0SMichal Hocko if (zone_to_nid(zone) == node) 11303e59fcb0SMichal Hocko continue; 11313e59fcb0SMichal Hocko node = zone_to_nid(zone); 113294310cbcSAnshuman Khandual 113394310cbcSAnshuman Khandual page = dequeue_huge_page_node_exact(h, node); 113494310cbcSAnshuman Khandual if (page) 113594310cbcSAnshuman Khandual return page; 113694310cbcSAnshuman Khandual } 11373e59fcb0SMichal Hocko if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 11383e59fcb0SMichal Hocko goto retry_cpuset; 11393e59fcb0SMichal Hocko 114094310cbcSAnshuman Khandual return NULL; 114194310cbcSAnshuman Khandual } 114294310cbcSAnshuman Khandual 1143a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 1144a5516438SAndi Kleen struct vm_area_struct *vma, 1145af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 1146af0ed73eSJoonsoo Kim long chg) 11471da177e4SLinus Torvalds { 1148cfcaa66fSBen Widawsky struct page *page = NULL; 1149480eccf9SLee Schermerhorn struct mempolicy *mpol; 115004ec6264SVlastimil Babka gfp_t gfp_mask; 11513e59fcb0SMichal Hocko nodemask_t *nodemask; 115204ec6264SVlastimil Babka int nid; 11531da177e4SLinus Torvalds 1154a1e78772SMel Gorman /* 1155a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 1156a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 1157a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 1158a1e78772SMel Gorman */ 1159af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg) && 1160a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 1161c0ff7453SMiao Xie goto err; 1162a1e78772SMel Gorman 116304f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 1164a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 11656eab04a8SJustin P. Mattock goto err; 116604f2cbe3SMel Gorman 116704ec6264SVlastimil Babka gfp_mask = htlb_alloc_mask(h); 116804ec6264SVlastimil Babka nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 1169cfcaa66fSBen Widawsky 1170cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 11713e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 1172cfcaa66fSBen Widawsky 1173cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 1174cfcaa66fSBen Widawsky nodemask = NULL; 1175cfcaa66fSBen Widawsky } 1176cfcaa66fSBen Widawsky 1177cfcaa66fSBen Widawsky if (!page) 1178cfcaa66fSBen Widawsky page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 1179cfcaa66fSBen Widawsky 11803e59fcb0SMichal Hocko if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { 1181d6995da3SMike Kravetz SetHPageRestoreReserve(page); 1182a63884e9SJoonsoo Kim h->resv_huge_pages--; 1183bf50bab2SNaoya Horiguchi } 1184cc9a6c87SMel Gorman 1185cc9a6c87SMel Gorman mpol_cond_put(mpol); 1186cc9a6c87SMel Gorman return page; 1187cc9a6c87SMel Gorman 1188c0ff7453SMiao Xie err: 1189cc9a6c87SMel Gorman return NULL; 11901da177e4SLinus Torvalds } 11911da177e4SLinus Torvalds 11921cac6f2cSLuiz Capitulino /* 11931cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 11941cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 11951cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 11961cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 11971cac6f2cSLuiz Capitulino * node for alloc or free. 11981cac6f2cSLuiz Capitulino */ 11991cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 12001cac6f2cSLuiz Capitulino { 12010edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 12021cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 12031cac6f2cSLuiz Capitulino 12041cac6f2cSLuiz Capitulino return nid; 12051cac6f2cSLuiz Capitulino } 12061cac6f2cSLuiz Capitulino 12071cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 12081cac6f2cSLuiz Capitulino { 12091cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 12101cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 12111cac6f2cSLuiz Capitulino return nid; 12121cac6f2cSLuiz Capitulino } 12131cac6f2cSLuiz Capitulino 12141cac6f2cSLuiz Capitulino /* 12151cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 12161cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 12171cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 12181cac6f2cSLuiz Capitulino * mask. 12191cac6f2cSLuiz Capitulino */ 12201cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 12211cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 12221cac6f2cSLuiz Capitulino { 12231cac6f2cSLuiz Capitulino int nid; 12241cac6f2cSLuiz Capitulino 12251cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 12261cac6f2cSLuiz Capitulino 12271cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 12281cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 12291cac6f2cSLuiz Capitulino 12301cac6f2cSLuiz Capitulino return nid; 12311cac6f2cSLuiz Capitulino } 12321cac6f2cSLuiz Capitulino 12331cac6f2cSLuiz Capitulino /* 123410c6ec49SMike Kravetz * helper for remove_pool_huge_page() - return the previously saved 12351cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 12361cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 12371cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 12381cac6f2cSLuiz Capitulino */ 12391cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 12401cac6f2cSLuiz Capitulino { 12411cac6f2cSLuiz Capitulino int nid; 12421cac6f2cSLuiz Capitulino 12431cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 12441cac6f2cSLuiz Capitulino 12451cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 12461cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 12471cac6f2cSLuiz Capitulino 12481cac6f2cSLuiz Capitulino return nid; 12491cac6f2cSLuiz Capitulino } 12501cac6f2cSLuiz Capitulino 12511cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 12521cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 12531cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 12541cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 12551cac6f2cSLuiz Capitulino nr_nodes--) 12561cac6f2cSLuiz Capitulino 12571cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 12581cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 12591cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 12601cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 12611cac6f2cSLuiz Capitulino nr_nodes--) 12621cac6f2cSLuiz Capitulino 1263e1073d1eSAneesh Kumar K.V #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 1264944d9fecSLuiz Capitulino static void destroy_compound_gigantic_page(struct page *page, 1265d00181b9SKirill A. Shutemov unsigned int order) 1266944d9fecSLuiz Capitulino { 1267944d9fecSLuiz Capitulino int i; 1268944d9fecSLuiz Capitulino int nr_pages = 1 << order; 1269944d9fecSLuiz Capitulino struct page *p = page + 1; 1270944d9fecSLuiz Capitulino 1271c8cc708aSGerald Schaefer atomic_set(compound_mapcount_ptr(page), 0); 127247e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 127347e29d32SJohn Hubbard 1274944d9fecSLuiz Capitulino for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 12751d798ca3SKirill A. Shutemov clear_compound_head(p); 1276944d9fecSLuiz Capitulino set_page_refcounted(p); 1277944d9fecSLuiz Capitulino } 1278944d9fecSLuiz Capitulino 1279944d9fecSLuiz Capitulino set_compound_order(page, 0); 1280ba9c1201SGerald Schaefer page[1].compound_nr = 0; 1281944d9fecSLuiz Capitulino __ClearPageHead(page); 1282944d9fecSLuiz Capitulino } 1283944d9fecSLuiz Capitulino 1284d00181b9SKirill A. Shutemov static void free_gigantic_page(struct page *page, unsigned int order) 1285944d9fecSLuiz Capitulino { 1286cf11e85fSRoman Gushchin /* 1287cf11e85fSRoman Gushchin * If the page isn't allocated using the cma allocator, 1288cf11e85fSRoman Gushchin * cma_release() returns false. 1289cf11e85fSRoman Gushchin */ 1290dbda8feaSBarry Song #ifdef CONFIG_CMA 1291dbda8feaSBarry Song if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order)) 1292cf11e85fSRoman Gushchin return; 1293dbda8feaSBarry Song #endif 1294cf11e85fSRoman Gushchin 1295944d9fecSLuiz Capitulino free_contig_range(page_to_pfn(page), 1 << order); 1296944d9fecSLuiz Capitulino } 1297944d9fecSLuiz Capitulino 12984eb0716eSAlexandre Ghiti #ifdef CONFIG_CONTIG_ALLOC 1299d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 1300d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) 1301944d9fecSLuiz Capitulino { 130204adbc3fSMiaohe Lin unsigned long nr_pages = pages_per_huge_page(h); 1303953f064aSLi Xinhai if (nid == NUMA_NO_NODE) 1304953f064aSLi Xinhai nid = numa_mem_id(); 1305944d9fecSLuiz Capitulino 1306dbda8feaSBarry Song #ifdef CONFIG_CMA 1307dbda8feaSBarry Song { 1308cf11e85fSRoman Gushchin struct page *page; 1309cf11e85fSRoman Gushchin int node; 1310cf11e85fSRoman Gushchin 1311953f064aSLi Xinhai if (hugetlb_cma[nid]) { 1312953f064aSLi Xinhai page = cma_alloc(hugetlb_cma[nid], nr_pages, 1313953f064aSLi Xinhai huge_page_order(h), true); 1314953f064aSLi Xinhai if (page) 1315953f064aSLi Xinhai return page; 1316953f064aSLi Xinhai } 1317953f064aSLi Xinhai 1318953f064aSLi Xinhai if (!(gfp_mask & __GFP_THISNODE)) { 1319cf11e85fSRoman Gushchin for_each_node_mask(node, *nodemask) { 1320953f064aSLi Xinhai if (node == nid || !hugetlb_cma[node]) 1321cf11e85fSRoman Gushchin continue; 1322cf11e85fSRoman Gushchin 1323cf11e85fSRoman Gushchin page = cma_alloc(hugetlb_cma[node], nr_pages, 1324cf11e85fSRoman Gushchin huge_page_order(h), true); 1325cf11e85fSRoman Gushchin if (page) 1326cf11e85fSRoman Gushchin return page; 1327cf11e85fSRoman Gushchin } 1328cf11e85fSRoman Gushchin } 1329953f064aSLi Xinhai } 1330dbda8feaSBarry Song #endif 1331cf11e85fSRoman Gushchin 13325e27a2dfSAnshuman Khandual return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); 1333944d9fecSLuiz Capitulino } 1334944d9fecSLuiz Capitulino 13354eb0716eSAlexandre Ghiti #else /* !CONFIG_CONTIG_ALLOC */ 13364eb0716eSAlexandre Ghiti static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 13374eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 13384eb0716eSAlexandre Ghiti { 13394eb0716eSAlexandre Ghiti return NULL; 13404eb0716eSAlexandre Ghiti } 13414eb0716eSAlexandre Ghiti #endif /* CONFIG_CONTIG_ALLOC */ 1342944d9fecSLuiz Capitulino 1343e1073d1eSAneesh Kumar K.V #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 1344d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 13454eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 13464eb0716eSAlexandre Ghiti { 13474eb0716eSAlexandre Ghiti return NULL; 13484eb0716eSAlexandre Ghiti } 1349d00181b9SKirill A. Shutemov static inline void free_gigantic_page(struct page *page, unsigned int order) { } 1350944d9fecSLuiz Capitulino static inline void destroy_compound_gigantic_page(struct page *page, 1351d00181b9SKirill A. Shutemov unsigned int order) { } 1352944d9fecSLuiz Capitulino #endif 1353944d9fecSLuiz Capitulino 13546eb4e88aSMike Kravetz /* 13556eb4e88aSMike Kravetz * Remove hugetlb page from lists, and update dtor so that page appears 13566eb4e88aSMike Kravetz * as just a compound page. A reference is held on the page. 13576eb4e88aSMike Kravetz * 13586eb4e88aSMike Kravetz * Must be called with hugetlb lock held. 13596eb4e88aSMike Kravetz */ 13606eb4e88aSMike Kravetz static void remove_hugetlb_page(struct hstate *h, struct page *page, 13616eb4e88aSMike Kravetz bool adjust_surplus) 13626eb4e88aSMike Kravetz { 13636eb4e88aSMike Kravetz int nid = page_to_nid(page); 13646eb4e88aSMike Kravetz 13656eb4e88aSMike Kravetz VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 13666eb4e88aSMike Kravetz VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page); 13676eb4e88aSMike Kravetz 13689487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 13696eb4e88aSMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 13706eb4e88aSMike Kravetz return; 13716eb4e88aSMike Kravetz 13726eb4e88aSMike Kravetz list_del(&page->lru); 13736eb4e88aSMike Kravetz 13746eb4e88aSMike Kravetz if (HPageFreed(page)) { 13756eb4e88aSMike Kravetz h->free_huge_pages--; 13766eb4e88aSMike Kravetz h->free_huge_pages_node[nid]--; 13776eb4e88aSMike Kravetz } 13786eb4e88aSMike Kravetz if (adjust_surplus) { 13796eb4e88aSMike Kravetz h->surplus_huge_pages--; 13806eb4e88aSMike Kravetz h->surplus_huge_pages_node[nid]--; 13816eb4e88aSMike Kravetz } 13826eb4e88aSMike Kravetz 1383e32d20c0SMike Kravetz /* 1384e32d20c0SMike Kravetz * Very subtle 1385e32d20c0SMike Kravetz * 1386e32d20c0SMike Kravetz * For non-gigantic pages set the destructor to the normal compound 1387e32d20c0SMike Kravetz * page dtor. This is needed in case someone takes an additional 1388e32d20c0SMike Kravetz * temporary ref to the page, and freeing is delayed until they drop 1389e32d20c0SMike Kravetz * their reference. 1390e32d20c0SMike Kravetz * 1391e32d20c0SMike Kravetz * For gigantic pages set the destructor to the null dtor. This 1392e32d20c0SMike Kravetz * destructor will never be called. Before freeing the gigantic 1393e32d20c0SMike Kravetz * page destroy_compound_gigantic_page will turn the compound page 1394e32d20c0SMike Kravetz * into a simple group of pages. After this the destructor does not 1395e32d20c0SMike Kravetz * apply. 1396e32d20c0SMike Kravetz * 1397e32d20c0SMike Kravetz * This handles the case where more than one ref is held when and 1398e32d20c0SMike Kravetz * after update_and_free_page is called. 1399e32d20c0SMike Kravetz */ 14006eb4e88aSMike Kravetz set_page_refcounted(page); 1401e32d20c0SMike Kravetz if (hstate_is_gigantic(h)) 14026eb4e88aSMike Kravetz set_compound_page_dtor(page, NULL_COMPOUND_DTOR); 1403e32d20c0SMike Kravetz else 1404e32d20c0SMike Kravetz set_compound_page_dtor(page, COMPOUND_PAGE_DTOR); 14056eb4e88aSMike Kravetz 14066eb4e88aSMike Kravetz h->nr_huge_pages--; 14076eb4e88aSMike Kravetz h->nr_huge_pages_node[nid]--; 14086eb4e88aSMike Kravetz } 14096eb4e88aSMike Kravetz 1410ad2fa371SMuchun Song static void add_hugetlb_page(struct hstate *h, struct page *page, 1411ad2fa371SMuchun Song bool adjust_surplus) 1412ad2fa371SMuchun Song { 1413ad2fa371SMuchun Song int zeroed; 1414ad2fa371SMuchun Song int nid = page_to_nid(page); 1415ad2fa371SMuchun Song 1416ad2fa371SMuchun Song VM_BUG_ON_PAGE(!HPageVmemmapOptimized(page), page); 1417ad2fa371SMuchun Song 1418ad2fa371SMuchun Song lockdep_assert_held(&hugetlb_lock); 1419ad2fa371SMuchun Song 1420ad2fa371SMuchun Song INIT_LIST_HEAD(&page->lru); 1421ad2fa371SMuchun Song h->nr_huge_pages++; 1422ad2fa371SMuchun Song h->nr_huge_pages_node[nid]++; 1423ad2fa371SMuchun Song 1424ad2fa371SMuchun Song if (adjust_surplus) { 1425ad2fa371SMuchun Song h->surplus_huge_pages++; 1426ad2fa371SMuchun Song h->surplus_huge_pages_node[nid]++; 1427ad2fa371SMuchun Song } 1428ad2fa371SMuchun Song 1429ad2fa371SMuchun Song set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1430ad2fa371SMuchun Song set_page_private(page, 0); 1431ad2fa371SMuchun Song SetHPageVmemmapOptimized(page); 1432ad2fa371SMuchun Song 1433ad2fa371SMuchun Song /* 1434b65a4edaSMike Kravetz * This page is about to be managed by the hugetlb allocator and 1435b65a4edaSMike Kravetz * should have no users. Drop our reference, and check for others 1436b65a4edaSMike Kravetz * just in case. 1437ad2fa371SMuchun Song */ 1438ad2fa371SMuchun Song zeroed = put_page_testzero(page); 1439b65a4edaSMike Kravetz if (!zeroed) 1440b65a4edaSMike Kravetz /* 1441b65a4edaSMike Kravetz * It is VERY unlikely soneone else has taken a ref on 1442b65a4edaSMike Kravetz * the page. In this case, we simply return as the 1443b65a4edaSMike Kravetz * hugetlb destructor (free_huge_page) will be called 1444b65a4edaSMike Kravetz * when this other ref is dropped. 1445b65a4edaSMike Kravetz */ 1446b65a4edaSMike Kravetz return; 1447b65a4edaSMike Kravetz 1448ad2fa371SMuchun Song arch_clear_hugepage_flags(page); 1449ad2fa371SMuchun Song enqueue_huge_page(h, page); 1450ad2fa371SMuchun Song } 1451ad2fa371SMuchun Song 1452b65d4adbSMuchun Song static void __update_and_free_page(struct hstate *h, struct page *page) 14536af2acb6SAdam Litke { 14546af2acb6SAdam Litke int i; 1455dbfee5aeSMike Kravetz struct page *subpage = page; 1456a5516438SAndi Kleen 14574eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1458944d9fecSLuiz Capitulino return; 145918229df5SAndy Whitcroft 1460ad2fa371SMuchun Song if (alloc_huge_page_vmemmap(h, page)) { 1461ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 1462ad2fa371SMuchun Song /* 1463ad2fa371SMuchun Song * If we cannot allocate vmemmap pages, just refuse to free the 1464ad2fa371SMuchun Song * page and put the page back on the hugetlb free list and treat 1465ad2fa371SMuchun Song * as a surplus page. 1466ad2fa371SMuchun Song */ 1467ad2fa371SMuchun Song add_hugetlb_page(h, page, true); 1468ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 1469ad2fa371SMuchun Song return; 1470ad2fa371SMuchun Song } 1471ad2fa371SMuchun Song 1472dbfee5aeSMike Kravetz for (i = 0; i < pages_per_huge_page(h); 1473dbfee5aeSMike Kravetz i++, subpage = mem_map_next(subpage, page, i)) { 1474dbfee5aeSMike Kravetz subpage->flags &= ~(1 << PG_locked | 1 << PG_error | 147532f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1476a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1477a7407a27SLuiz Capitulino 1 << PG_writeback); 14786af2acb6SAdam Litke } 1479944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) { 1480944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 1481944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 1482944d9fecSLuiz Capitulino } else { 1483a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 14846af2acb6SAdam Litke } 1485944d9fecSLuiz Capitulino } 14866af2acb6SAdam Litke 1487b65d4adbSMuchun Song /* 1488b65d4adbSMuchun Song * As update_and_free_page() can be called under any context, so we cannot 1489b65d4adbSMuchun Song * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the 1490b65d4adbSMuchun Song * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate 1491b65d4adbSMuchun Song * the vmemmap pages. 1492b65d4adbSMuchun Song * 1493b65d4adbSMuchun Song * free_hpage_workfn() locklessly retrieves the linked list of pages to be 1494b65d4adbSMuchun Song * freed and frees them one-by-one. As the page->mapping pointer is going 1495b65d4adbSMuchun Song * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node 1496b65d4adbSMuchun Song * structure of a lockless linked list of huge pages to be freed. 1497b65d4adbSMuchun Song */ 1498b65d4adbSMuchun Song static LLIST_HEAD(hpage_freelist); 1499b65d4adbSMuchun Song 1500b65d4adbSMuchun Song static void free_hpage_workfn(struct work_struct *work) 1501b65d4adbSMuchun Song { 1502b65d4adbSMuchun Song struct llist_node *node; 1503b65d4adbSMuchun Song 1504b65d4adbSMuchun Song node = llist_del_all(&hpage_freelist); 1505b65d4adbSMuchun Song 1506b65d4adbSMuchun Song while (node) { 1507b65d4adbSMuchun Song struct page *page; 1508b65d4adbSMuchun Song struct hstate *h; 1509b65d4adbSMuchun Song 1510b65d4adbSMuchun Song page = container_of((struct address_space **)node, 1511b65d4adbSMuchun Song struct page, mapping); 1512b65d4adbSMuchun Song node = node->next; 1513b65d4adbSMuchun Song page->mapping = NULL; 1514b65d4adbSMuchun Song /* 1515b65d4adbSMuchun Song * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate() 1516b65d4adbSMuchun Song * is going to trigger because a previous call to 1517b65d4adbSMuchun Song * remove_hugetlb_page() will set_compound_page_dtor(page, 1518b65d4adbSMuchun Song * NULL_COMPOUND_DTOR), so do not use page_hstate() directly. 1519b65d4adbSMuchun Song */ 1520b65d4adbSMuchun Song h = size_to_hstate(page_size(page)); 1521b65d4adbSMuchun Song 1522b65d4adbSMuchun Song __update_and_free_page(h, page); 1523b65d4adbSMuchun Song 1524b65d4adbSMuchun Song cond_resched(); 1525b65d4adbSMuchun Song } 1526b65d4adbSMuchun Song } 1527b65d4adbSMuchun Song static DECLARE_WORK(free_hpage_work, free_hpage_workfn); 1528b65d4adbSMuchun Song 1529b65d4adbSMuchun Song static inline void flush_free_hpage_work(struct hstate *h) 1530b65d4adbSMuchun Song { 1531b65d4adbSMuchun Song if (free_vmemmap_pages_per_hpage(h)) 1532b65d4adbSMuchun Song flush_work(&free_hpage_work); 1533b65d4adbSMuchun Song } 1534b65d4adbSMuchun Song 1535b65d4adbSMuchun Song static void update_and_free_page(struct hstate *h, struct page *page, 1536b65d4adbSMuchun Song bool atomic) 1537b65d4adbSMuchun Song { 1538ad2fa371SMuchun Song if (!HPageVmemmapOptimized(page) || !atomic) { 1539b65d4adbSMuchun Song __update_and_free_page(h, page); 1540b65d4adbSMuchun Song return; 1541b65d4adbSMuchun Song } 1542b65d4adbSMuchun Song 1543b65d4adbSMuchun Song /* 1544b65d4adbSMuchun Song * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages. 1545b65d4adbSMuchun Song * 1546b65d4adbSMuchun Song * Only call schedule_work() if hpage_freelist is previously 1547b65d4adbSMuchun Song * empty. Otherwise, schedule_work() had been called but the workfn 1548b65d4adbSMuchun Song * hasn't retrieved the list yet. 1549b65d4adbSMuchun Song */ 1550b65d4adbSMuchun Song if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist)) 1551b65d4adbSMuchun Song schedule_work(&free_hpage_work); 1552b65d4adbSMuchun Song } 1553b65d4adbSMuchun Song 155410c6ec49SMike Kravetz static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list) 155510c6ec49SMike Kravetz { 155610c6ec49SMike Kravetz struct page *page, *t_page; 155710c6ec49SMike Kravetz 155810c6ec49SMike Kravetz list_for_each_entry_safe(page, t_page, list, lru) { 1559b65d4adbSMuchun Song update_and_free_page(h, page, false); 156010c6ec49SMike Kravetz cond_resched(); 156110c6ec49SMike Kravetz } 156210c6ec49SMike Kravetz } 156310c6ec49SMike Kravetz 1564e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1565e5ff2159SAndi Kleen { 1566e5ff2159SAndi Kleen struct hstate *h; 1567e5ff2159SAndi Kleen 1568e5ff2159SAndi Kleen for_each_hstate(h) { 1569e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1570e5ff2159SAndi Kleen return h; 1571e5ff2159SAndi Kleen } 1572e5ff2159SAndi Kleen return NULL; 1573e5ff2159SAndi Kleen } 1574e5ff2159SAndi Kleen 1575db71ef79SMike Kravetz void free_huge_page(struct page *page) 157627a85ef1SDavid Gibson { 1577a5516438SAndi Kleen /* 1578a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1579a5516438SAndi Kleen * compound page destructor. 1580a5516438SAndi Kleen */ 1581e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 15827893d1d5SAdam Litke int nid = page_to_nid(page); 1583d6995da3SMike Kravetz struct hugepage_subpool *spool = hugetlb_page_subpool(page); 158407443a85SJoonsoo Kim bool restore_reserve; 1585db71ef79SMike Kravetz unsigned long flags; 158627a85ef1SDavid Gibson 1587b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1588b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_mapcount(page), page); 15898ace22bcSYongkai Wu 1590d6995da3SMike Kravetz hugetlb_set_page_subpool(page, NULL); 15918ace22bcSYongkai Wu page->mapping = NULL; 1592d6995da3SMike Kravetz restore_reserve = HPageRestoreReserve(page); 1593d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 159427a85ef1SDavid Gibson 15951c5ecae3SMike Kravetz /* 1596d6995da3SMike Kravetz * If HPageRestoreReserve was set on page, page allocation consumed a 15970919e1b6SMike Kravetz * reservation. If the page was associated with a subpool, there 15980919e1b6SMike Kravetz * would have been a page reserved in the subpool before allocation 15990919e1b6SMike Kravetz * via hugepage_subpool_get_pages(). Since we are 'restoring' the 16006c26d310SMiaohe Lin * reservation, do not call hugepage_subpool_put_pages() as this will 16010919e1b6SMike Kravetz * remove the reserved page from the subpool. 16020919e1b6SMike Kravetz */ 16030919e1b6SMike Kravetz if (!restore_reserve) { 16040919e1b6SMike Kravetz /* 16050919e1b6SMike Kravetz * A return code of zero implies that the subpool will be 16060919e1b6SMike Kravetz * under its minimum size if the reservation is not restored 16070919e1b6SMike Kravetz * after page is free. Therefore, force restore_reserve 16080919e1b6SMike Kravetz * operation. 16091c5ecae3SMike Kravetz */ 16101c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 16111c5ecae3SMike Kravetz restore_reserve = true; 16120919e1b6SMike Kravetz } 16131c5ecae3SMike Kravetz 1614db71ef79SMike Kravetz spin_lock_irqsave(&hugetlb_lock, flags); 16158f251a3dSMike Kravetz ClearHPageMigratable(page); 16166d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 16176d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 161808cf9fafSMina Almasry hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 161908cf9fafSMina Almasry pages_per_huge_page(h), page); 162007443a85SJoonsoo Kim if (restore_reserve) 162107443a85SJoonsoo Kim h->resv_huge_pages++; 162207443a85SJoonsoo Kim 16239157c311SMike Kravetz if (HPageTemporary(page)) { 16246eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 1625db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1626b65d4adbSMuchun Song update_and_free_page(h, page, true); 1627ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 16280edaecfaSAneesh Kumar K.V /* remove the page from active list */ 16296eb4e88aSMike Kravetz remove_hugetlb_page(h, page, true); 1630db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1631b65d4adbSMuchun Song update_and_free_page(h, page, true); 16327893d1d5SAdam Litke } else { 16335d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1634a5516438SAndi Kleen enqueue_huge_page(h, page); 1635db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 163627a85ef1SDavid Gibson } 16371121828aSMike Kravetz } 163827a85ef1SDavid Gibson 1639d3d99fccSOscar Salvador /* 1640d3d99fccSOscar Salvador * Must be called with the hugetlb lock held 1641d3d99fccSOscar Salvador */ 1642d3d99fccSOscar Salvador static void __prep_account_new_huge_page(struct hstate *h, int nid) 1643d3d99fccSOscar Salvador { 1644d3d99fccSOscar Salvador lockdep_assert_held(&hugetlb_lock); 1645d3d99fccSOscar Salvador h->nr_huge_pages++; 1646d3d99fccSOscar Salvador h->nr_huge_pages_node[nid]++; 1647d3d99fccSOscar Salvador } 1648d3d99fccSOscar Salvador 1649f41f2ed4SMuchun Song static void __prep_new_huge_page(struct hstate *h, struct page *page) 1650b7ba30c6SAndi Kleen { 1651f41f2ed4SMuchun Song free_huge_page_vmemmap(h, page); 16520edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1653f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1654ff546117SMike Kravetz hugetlb_set_page_subpool(page, NULL); 16559dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 16561adc4d41SMina Almasry set_hugetlb_cgroup_rsvd(page, NULL); 1657d3d99fccSOscar Salvador } 1658d3d99fccSOscar Salvador 1659d3d99fccSOscar Salvador static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1660d3d99fccSOscar Salvador { 1661f41f2ed4SMuchun Song __prep_new_huge_page(h, page); 1662db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 1663d3d99fccSOscar Salvador __prep_account_new_huge_page(h, nid); 1664db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1665b7ba30c6SAndi Kleen } 1666b7ba30c6SAndi Kleen 16677118fc29SMike Kravetz static bool prep_compound_gigantic_page(struct page *page, unsigned int order) 166820a0307cSWu Fengguang { 16697118fc29SMike Kravetz int i, j; 167020a0307cSWu Fengguang int nr_pages = 1 << order; 167120a0307cSWu Fengguang struct page *p = page + 1; 167220a0307cSWu Fengguang 167320a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 167420a0307cSWu Fengguang set_compound_order(page, order); 1675ef5a22beSAndrea Arcangeli __ClearPageReserved(page); 1676de09d31dSKirill A. Shutemov __SetPageHead(page); 167720a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1678ef5a22beSAndrea Arcangeli /* 1679ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1680ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1681ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 16827c8de358SEthon Paul * too. Otherwise drivers using get_user_pages() to access tail 1683ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1684ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1685ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1686ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1687ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1688ef5a22beSAndrea Arcangeli * after get_user_pages(). 1689ef5a22beSAndrea Arcangeli */ 1690ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 16917118fc29SMike Kravetz /* 16927118fc29SMike Kravetz * Subtle and very unlikely 16937118fc29SMike Kravetz * 16947118fc29SMike Kravetz * Gigantic 'page allocators' such as memblock or cma will 16957118fc29SMike Kravetz * return a set of pages with each page ref counted. We need 16967118fc29SMike Kravetz * to turn this set of pages into a compound page with tail 16977118fc29SMike Kravetz * page ref counts set to zero. Code such as speculative page 16987118fc29SMike Kravetz * cache adding could take a ref on a 'to be' tail page. 16997118fc29SMike Kravetz * We need to respect any increased ref count, and only set 17007118fc29SMike Kravetz * the ref count to zero if count is currently 1. If count 1701416d85edSMike Kravetz * is not 1, we return an error. An error return indicates 1702416d85edSMike Kravetz * the set of pages can not be converted to a gigantic page. 1703416d85edSMike Kravetz * The caller who allocated the pages should then discard the 1704416d85edSMike Kravetz * pages using the appropriate free interface. 17057118fc29SMike Kravetz */ 17067118fc29SMike Kravetz if (!page_ref_freeze(p, 1)) { 1707416d85edSMike Kravetz pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n"); 17087118fc29SMike Kravetz goto out_error; 17097118fc29SMike Kravetz } 171058a84aa9SYouquan Song set_page_count(p, 0); 17111d798ca3SKirill A. Shutemov set_compound_head(p, page); 171220a0307cSWu Fengguang } 1713b4330afbSMike Kravetz atomic_set(compound_mapcount_ptr(page), -1); 171447e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 17157118fc29SMike Kravetz return true; 17167118fc29SMike Kravetz 17177118fc29SMike Kravetz out_error: 17187118fc29SMike Kravetz /* undo tail page modifications made above */ 17197118fc29SMike Kravetz p = page + 1; 17207118fc29SMike Kravetz for (j = 1; j < i; j++, p = mem_map_next(p, page, j)) { 17217118fc29SMike Kravetz clear_compound_head(p); 17227118fc29SMike Kravetz set_page_refcounted(p); 17237118fc29SMike Kravetz } 17247118fc29SMike Kravetz /* need to clear PG_reserved on remaining tail pages */ 17257118fc29SMike Kravetz for (; j < nr_pages; j++, p = mem_map_next(p, page, j)) 17267118fc29SMike Kravetz __ClearPageReserved(p); 17277118fc29SMike Kravetz set_compound_order(page, 0); 17287118fc29SMike Kravetz page[1].compound_nr = 0; 17297118fc29SMike Kravetz __ClearPageHead(page); 17307118fc29SMike Kravetz return false; 173120a0307cSWu Fengguang } 173220a0307cSWu Fengguang 17337795912cSAndrew Morton /* 17347795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 17357795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 17367795912cSAndrew Morton * details. 17377795912cSAndrew Morton */ 173820a0307cSWu Fengguang int PageHuge(struct page *page) 173920a0307cSWu Fengguang { 174020a0307cSWu Fengguang if (!PageCompound(page)) 174120a0307cSWu Fengguang return 0; 174220a0307cSWu Fengguang 174320a0307cSWu Fengguang page = compound_head(page); 1744f1e61557SKirill A. Shutemov return page[1].compound_dtor == HUGETLB_PAGE_DTOR; 174520a0307cSWu Fengguang } 174643131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 174743131e14SNaoya Horiguchi 174827c73ae7SAndrea Arcangeli /* 174927c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 175027c73ae7SAndrea Arcangeli * normal or transparent huge pages. 175127c73ae7SAndrea Arcangeli */ 175227c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 175327c73ae7SAndrea Arcangeli { 175427c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 175527c73ae7SAndrea Arcangeli return 0; 175627c73ae7SAndrea Arcangeli 1757d4af73e3SVlastimil Babka return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; 175827c73ae7SAndrea Arcangeli } 175927c73ae7SAndrea Arcangeli 1760c0d0381aSMike Kravetz /* 1761c0d0381aSMike Kravetz * Find and lock address space (mapping) in write mode. 1762c0d0381aSMike Kravetz * 1763336bf30eSMike Kravetz * Upon entry, the page is locked which means that page_mapping() is 1764336bf30eSMike Kravetz * stable. Due to locking order, we can only trylock_write. If we can 1765336bf30eSMike Kravetz * not get the lock, simply return NULL to caller. 1766c0d0381aSMike Kravetz */ 1767c0d0381aSMike Kravetz struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) 1768c0d0381aSMike Kravetz { 1769336bf30eSMike Kravetz struct address_space *mapping = page_mapping(hpage); 1770c0d0381aSMike Kravetz 1771c0d0381aSMike Kravetz if (!mapping) 1772c0d0381aSMike Kravetz return mapping; 1773c0d0381aSMike Kravetz 1774c0d0381aSMike Kravetz if (i_mmap_trylock_write(mapping)) 1775c0d0381aSMike Kravetz return mapping; 1776c0d0381aSMike Kravetz 1777c0d0381aSMike Kravetz return NULL; 1778c0d0381aSMike Kravetz } 1779c0d0381aSMike Kravetz 1780fe19bd3dSHugh Dickins pgoff_t hugetlb_basepage_index(struct page *page) 178113d60f4bSZhang Yi { 178213d60f4bSZhang Yi struct page *page_head = compound_head(page); 178313d60f4bSZhang Yi pgoff_t index = page_index(page_head); 178413d60f4bSZhang Yi unsigned long compound_idx; 178513d60f4bSZhang Yi 178613d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 178713d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 178813d60f4bSZhang Yi else 178913d60f4bSZhang Yi compound_idx = page - page_head; 179013d60f4bSZhang Yi 179113d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 179213d60f4bSZhang Yi } 179313d60f4bSZhang Yi 17940c397daeSMichal Hocko static struct page *alloc_buddy_huge_page(struct hstate *h, 1795f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1796f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 17971da177e4SLinus Torvalds { 1798af0fb9dfSMichal Hocko int order = huge_page_order(h); 17991da177e4SLinus Torvalds struct page *page; 1800f60858f9SMike Kravetz bool alloc_try_hard = true; 1801f96efd58SJoe Jin 1802f60858f9SMike Kravetz /* 1803f60858f9SMike Kravetz * By default we always try hard to allocate the page with 1804f60858f9SMike Kravetz * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in 1805f60858f9SMike Kravetz * a loop (to adjust global huge page counts) and previous allocation 1806f60858f9SMike Kravetz * failed, do not continue to try hard on the same node. Use the 1807f60858f9SMike Kravetz * node_alloc_noretry bitmap to manage this state information. 1808f60858f9SMike Kravetz */ 1809f60858f9SMike Kravetz if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) 1810f60858f9SMike Kravetz alloc_try_hard = false; 1811f60858f9SMike Kravetz gfp_mask |= __GFP_COMP|__GFP_NOWARN; 1812f60858f9SMike Kravetz if (alloc_try_hard) 1813f60858f9SMike Kravetz gfp_mask |= __GFP_RETRY_MAYFAIL; 1814af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 1815af0fb9dfSMichal Hocko nid = numa_mem_id(); 181684172f4bSMatthew Wilcox (Oracle) page = __alloc_pages(gfp_mask, order, nid, nmask); 1817af0fb9dfSMichal Hocko if (page) 1818af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC); 1819af0fb9dfSMichal Hocko else 1820af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 182163b4613cSNishanth Aravamudan 1822f60858f9SMike Kravetz /* 1823f60858f9SMike Kravetz * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this 1824f60858f9SMike Kravetz * indicates an overall state change. Clear bit so that we resume 1825f60858f9SMike Kravetz * normal 'try hard' allocations. 1826f60858f9SMike Kravetz */ 1827f60858f9SMike Kravetz if (node_alloc_noretry && page && !alloc_try_hard) 1828f60858f9SMike Kravetz node_clear(nid, *node_alloc_noretry); 1829f60858f9SMike Kravetz 1830f60858f9SMike Kravetz /* 1831f60858f9SMike Kravetz * If we tried hard to get a page but failed, set bit so that 1832f60858f9SMike Kravetz * subsequent attempts will not try as hard until there is an 1833f60858f9SMike Kravetz * overall state change. 1834f60858f9SMike Kravetz */ 1835f60858f9SMike Kravetz if (node_alloc_noretry && !page && alloc_try_hard) 1836f60858f9SMike Kravetz node_set(nid, *node_alloc_noretry); 1837f60858f9SMike Kravetz 183863b4613cSNishanth Aravamudan return page; 183963b4613cSNishanth Aravamudan } 184063b4613cSNishanth Aravamudan 1841af0fb9dfSMichal Hocko /* 18420c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 18430c397daeSMichal Hocko * should use this function to get new hugetlb pages 18440c397daeSMichal Hocko */ 18450c397daeSMichal Hocko static struct page *alloc_fresh_huge_page(struct hstate *h, 1846f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1847f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 18480c397daeSMichal Hocko { 18490c397daeSMichal Hocko struct page *page; 18507118fc29SMike Kravetz bool retry = false; 18510c397daeSMichal Hocko 18527118fc29SMike Kravetz retry: 18530c397daeSMichal Hocko if (hstate_is_gigantic(h)) 18540c397daeSMichal Hocko page = alloc_gigantic_page(h, gfp_mask, nid, nmask); 18550c397daeSMichal Hocko else 18560c397daeSMichal Hocko page = alloc_buddy_huge_page(h, gfp_mask, 1857f60858f9SMike Kravetz nid, nmask, node_alloc_noretry); 18580c397daeSMichal Hocko if (!page) 18590c397daeSMichal Hocko return NULL; 18600c397daeSMichal Hocko 18617118fc29SMike Kravetz if (hstate_is_gigantic(h)) { 18627118fc29SMike Kravetz if (!prep_compound_gigantic_page(page, huge_page_order(h))) { 18637118fc29SMike Kravetz /* 18647118fc29SMike Kravetz * Rare failure to convert pages to compound page. 18657118fc29SMike Kravetz * Free pages and try again - ONCE! 18667118fc29SMike Kravetz */ 18677118fc29SMike Kravetz free_gigantic_page(page, huge_page_order(h)); 18687118fc29SMike Kravetz if (!retry) { 18697118fc29SMike Kravetz retry = true; 18707118fc29SMike Kravetz goto retry; 18717118fc29SMike Kravetz } 18727118fc29SMike Kravetz return NULL; 18737118fc29SMike Kravetz } 18747118fc29SMike Kravetz } 18750c397daeSMichal Hocko prep_new_huge_page(h, page, page_to_nid(page)); 18760c397daeSMichal Hocko 18770c397daeSMichal Hocko return page; 18780c397daeSMichal Hocko } 18790c397daeSMichal Hocko 18800c397daeSMichal Hocko /* 1881af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 1882af0fb9dfSMichal Hocko * manner. 1883af0fb9dfSMichal Hocko */ 1884f60858f9SMike Kravetz static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 1885f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 1886b2261026SJoonsoo Kim { 1887b2261026SJoonsoo Kim struct page *page; 1888b2261026SJoonsoo Kim int nr_nodes, node; 1889af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 1890b2261026SJoonsoo Kim 1891b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1892f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, 1893f60858f9SMike Kravetz node_alloc_noretry); 1894af0fb9dfSMichal Hocko if (page) 1895b2261026SJoonsoo Kim break; 1896b2261026SJoonsoo Kim } 1897b2261026SJoonsoo Kim 1898af0fb9dfSMichal Hocko if (!page) 1899af0fb9dfSMichal Hocko return 0; 1900b2261026SJoonsoo Kim 1901af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 1902af0fb9dfSMichal Hocko 1903af0fb9dfSMichal Hocko return 1; 1904b2261026SJoonsoo Kim } 1905b2261026SJoonsoo Kim 1906e8c5c824SLee Schermerhorn /* 190710c6ec49SMike Kravetz * Remove huge page from pool from next node to free. Attempt to keep 190810c6ec49SMike Kravetz * persistent huge pages more or less balanced over allowed nodes. 190910c6ec49SMike Kravetz * This routine only 'removes' the hugetlb page. The caller must make 191010c6ec49SMike Kravetz * an additional call to free the page to low level allocators. 1911e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 1912e8c5c824SLee Schermerhorn */ 191310c6ec49SMike Kravetz static struct page *remove_pool_huge_page(struct hstate *h, 191410c6ec49SMike Kravetz nodemask_t *nodes_allowed, 19156ae11b27SLee Schermerhorn bool acct_surplus) 1916e8c5c824SLee Schermerhorn { 1917b2261026SJoonsoo Kim int nr_nodes, node; 191810c6ec49SMike Kravetz struct page *page = NULL; 1919e8c5c824SLee Schermerhorn 19209487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1921b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1922685f3457SLee Schermerhorn /* 1923685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 1924685f3457SLee Schermerhorn * nodes with surplus pages. 1925685f3457SLee Schermerhorn */ 1926b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 1927b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 192810c6ec49SMike Kravetz page = list_entry(h->hugepage_freelists[node].next, 1929e8c5c824SLee Schermerhorn struct page, lru); 19306eb4e88aSMike Kravetz remove_hugetlb_page(h, page, acct_surplus); 19319a76db09SLee Schermerhorn break; 1932e8c5c824SLee Schermerhorn } 1933b2261026SJoonsoo Kim } 1934e8c5c824SLee Schermerhorn 193510c6ec49SMike Kravetz return page; 1936e8c5c824SLee Schermerhorn } 1937e8c5c824SLee Schermerhorn 1938c8721bbbSNaoya Horiguchi /* 1939c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 1940faf53defSNaoya Horiguchi * nothing for in-use hugepages and non-hugepages. 1941faf53defSNaoya Horiguchi * This function returns values like below: 1942faf53defSNaoya Horiguchi * 1943ad2fa371SMuchun Song * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages 1944ad2fa371SMuchun Song * when the system is under memory pressure and the feature of 1945ad2fa371SMuchun Song * freeing unused vmemmap pages associated with each hugetlb page 1946ad2fa371SMuchun Song * is enabled. 1947faf53defSNaoya Horiguchi * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 1948faf53defSNaoya Horiguchi * (allocated or reserved.) 1949faf53defSNaoya Horiguchi * 0: successfully dissolved free hugepages or the page is not a 1950faf53defSNaoya Horiguchi * hugepage (considered as already dissolved) 1951c8721bbbSNaoya Horiguchi */ 1952c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 1953c8721bbbSNaoya Horiguchi { 19546bc9b564SNaoya Horiguchi int rc = -EBUSY; 1955082d5b6bSGerald Schaefer 19567ffddd49SMuchun Song retry: 1957faf53defSNaoya Horiguchi /* Not to disrupt normal path by vainly holding hugetlb_lock */ 1958faf53defSNaoya Horiguchi if (!PageHuge(page)) 1959faf53defSNaoya Horiguchi return 0; 1960faf53defSNaoya Horiguchi 1961db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 1962faf53defSNaoya Horiguchi if (!PageHuge(page)) { 1963faf53defSNaoya Horiguchi rc = 0; 1964faf53defSNaoya Horiguchi goto out; 1965faf53defSNaoya Horiguchi } 1966faf53defSNaoya Horiguchi 1967faf53defSNaoya Horiguchi if (!page_count(page)) { 19682247bb33SGerald Schaefer struct page *head = compound_head(page); 19692247bb33SGerald Schaefer struct hstate *h = page_hstate(head); 19706bc9b564SNaoya Horiguchi if (h->free_huge_pages - h->resv_huge_pages == 0) 1971082d5b6bSGerald Schaefer goto out; 19727ffddd49SMuchun Song 19737ffddd49SMuchun Song /* 19747ffddd49SMuchun Song * We should make sure that the page is already on the free list 19757ffddd49SMuchun Song * when it is dissolved. 19767ffddd49SMuchun Song */ 19776c037149SMike Kravetz if (unlikely(!HPageFreed(head))) { 1978db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 19797ffddd49SMuchun Song cond_resched(); 19807ffddd49SMuchun Song 19817ffddd49SMuchun Song /* 19827ffddd49SMuchun Song * Theoretically, we should return -EBUSY when we 19837ffddd49SMuchun Song * encounter this race. In fact, we have a chance 19847ffddd49SMuchun Song * to successfully dissolve the page if we do a 19857ffddd49SMuchun Song * retry. Because the race window is quite small. 19867ffddd49SMuchun Song * If we seize this opportunity, it is an optimization 19877ffddd49SMuchun Song * for increasing the success rate of dissolving page. 19887ffddd49SMuchun Song */ 19897ffddd49SMuchun Song goto retry; 19907ffddd49SMuchun Song } 19917ffddd49SMuchun Song 1992ad2fa371SMuchun Song remove_hugetlb_page(h, head, false); 1993ad2fa371SMuchun Song h->max_huge_pages--; 1994ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 1995ad2fa371SMuchun Song 1996c3114a84SAnshuman Khandual /* 1997ad2fa371SMuchun Song * Normally update_and_free_page will allocate required vmemmmap 1998ad2fa371SMuchun Song * before freeing the page. update_and_free_page will fail to 1999ad2fa371SMuchun Song * free the page if it can not allocate required vmemmap. We 2000ad2fa371SMuchun Song * need to adjust max_huge_pages if the page is not freed. 2001ad2fa371SMuchun Song * Attempt to allocate vmemmmap here so that we can take 2002ad2fa371SMuchun Song * appropriate action on failure. 2003ad2fa371SMuchun Song */ 2004ad2fa371SMuchun Song rc = alloc_huge_page_vmemmap(h, head); 2005ad2fa371SMuchun Song if (!rc) { 2006ad2fa371SMuchun Song /* 2007ad2fa371SMuchun Song * Move PageHWPoison flag from head page to the raw 2008ad2fa371SMuchun Song * error page, which makes any subpages rather than 2009ad2fa371SMuchun Song * the error page reusable. 2010c3114a84SAnshuman Khandual */ 2011c3114a84SAnshuman Khandual if (PageHWPoison(head) && page != head) { 2012c3114a84SAnshuman Khandual SetPageHWPoison(page); 2013c3114a84SAnshuman Khandual ClearPageHWPoison(head); 2014c3114a84SAnshuman Khandual } 2015b65d4adbSMuchun Song update_and_free_page(h, head, false); 2016ad2fa371SMuchun Song } else { 2017ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 2018ad2fa371SMuchun Song add_hugetlb_page(h, head, false); 2019ad2fa371SMuchun Song h->max_huge_pages++; 2020ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2021ad2fa371SMuchun Song } 2022ad2fa371SMuchun Song 2023ad2fa371SMuchun Song return rc; 2024c8721bbbSNaoya Horiguchi } 2025082d5b6bSGerald Schaefer out: 2026db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2027082d5b6bSGerald Schaefer return rc; 2028c8721bbbSNaoya Horiguchi } 2029c8721bbbSNaoya Horiguchi 2030c8721bbbSNaoya Horiguchi /* 2031c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 2032c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 20332247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 20342247bb33SGerald Schaefer * part of it lies within the given range. 2035082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 2036082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 2037c8721bbbSNaoya Horiguchi */ 2038082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 2039c8721bbbSNaoya Horiguchi { 2040c8721bbbSNaoya Horiguchi unsigned long pfn; 2041eb03aa00SGerald Schaefer struct page *page; 2042082d5b6bSGerald Schaefer int rc = 0; 2043c8721bbbSNaoya Horiguchi 2044d0177639SLi Zhong if (!hugepages_supported()) 2045082d5b6bSGerald Schaefer return rc; 2046d0177639SLi Zhong 2047eb03aa00SGerald Schaefer for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { 2048eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 2049eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 2050eb03aa00SGerald Schaefer if (rc) 2051082d5b6bSGerald Schaefer break; 2052eb03aa00SGerald Schaefer } 2053082d5b6bSGerald Schaefer 2054082d5b6bSGerald Schaefer return rc; 2055c8721bbbSNaoya Horiguchi } 2056c8721bbbSNaoya Horiguchi 2057ab5ac90aSMichal Hocko /* 2058ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 2059ab5ac90aSMichal Hocko */ 20600c397daeSMichal Hocko static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, 2061b65a4edaSMike Kravetz int nid, nodemask_t *nmask, bool zero_ref) 20627893d1d5SAdam Litke { 20639980d744SMichal Hocko struct page *page = NULL; 2064b65a4edaSMike Kravetz bool retry = false; 20657893d1d5SAdam Litke 2066bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2067aa888a74SAndi Kleen return NULL; 2068aa888a74SAndi Kleen 2069db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 20709980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 20719980d744SMichal Hocko goto out_unlock; 2072db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2073d1c3fb1fSNishanth Aravamudan 2074b65a4edaSMike Kravetz retry: 2075f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 20769980d744SMichal Hocko if (!page) 20770c397daeSMichal Hocko return NULL; 2078d1c3fb1fSNishanth Aravamudan 2079db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 20809980d744SMichal Hocko /* 20819980d744SMichal Hocko * We could have raced with the pool size change. 20829980d744SMichal Hocko * Double check that and simply deallocate the new page 20839980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 20849980d744SMichal Hocko * temporary page to workaround the nasty free_huge_page 20859980d744SMichal Hocko * codeflow 20869980d744SMichal Hocko */ 20879980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 20889157c311SMike Kravetz SetHPageTemporary(page); 2089db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 20909980d744SMichal Hocko put_page(page); 20912bf753e6SKai Shen return NULL; 2092b65a4edaSMike Kravetz } 2093b65a4edaSMike Kravetz 2094b65a4edaSMike Kravetz if (zero_ref) { 2095b65a4edaSMike Kravetz /* 2096b65a4edaSMike Kravetz * Caller requires a page with zero ref count. 2097b65a4edaSMike Kravetz * We will drop ref count here. If someone else is holding 2098b65a4edaSMike Kravetz * a ref, the page will be freed when they drop it. Abuse 2099b65a4edaSMike Kravetz * temporary page flag to accomplish this. 2100b65a4edaSMike Kravetz */ 2101b65a4edaSMike Kravetz SetHPageTemporary(page); 2102b65a4edaSMike Kravetz if (!put_page_testzero(page)) { 2103b65a4edaSMike Kravetz /* 2104b65a4edaSMike Kravetz * Unexpected inflated ref count on freshly allocated 2105b65a4edaSMike Kravetz * huge. Retry once. 2106b65a4edaSMike Kravetz */ 2107b65a4edaSMike Kravetz pr_info("HugeTLB unexpected inflated ref count on freshly allocated page\n"); 2108b65a4edaSMike Kravetz spin_unlock_irq(&hugetlb_lock); 2109b65a4edaSMike Kravetz if (retry) 2110b65a4edaSMike Kravetz return NULL; 2111b65a4edaSMike Kravetz 2112b65a4edaSMike Kravetz retry = true; 2113b65a4edaSMike Kravetz goto retry; 2114b65a4edaSMike Kravetz } 2115b65a4edaSMike Kravetz ClearHPageTemporary(page); 2116b65a4edaSMike Kravetz } 2117b65a4edaSMike Kravetz 21189980d744SMichal Hocko h->surplus_huge_pages++; 21194704dea3SMichal Hocko h->surplus_huge_pages_node[page_to_nid(page)]++; 21209980d744SMichal Hocko 21219980d744SMichal Hocko out_unlock: 2122db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 21237893d1d5SAdam Litke 21247893d1d5SAdam Litke return page; 21257893d1d5SAdam Litke } 21267893d1d5SAdam Litke 2127bbe88753SJoonsoo Kim static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, 2128ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 2129ab5ac90aSMichal Hocko { 2130ab5ac90aSMichal Hocko struct page *page; 2131ab5ac90aSMichal Hocko 2132ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 2133ab5ac90aSMichal Hocko return NULL; 2134ab5ac90aSMichal Hocko 2135f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 2136ab5ac90aSMichal Hocko if (!page) 2137ab5ac90aSMichal Hocko return NULL; 2138ab5ac90aSMichal Hocko 2139ab5ac90aSMichal Hocko /* 2140ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 2141ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 2142ab5ac90aSMichal Hocko */ 21439157c311SMike Kravetz SetHPageTemporary(page); 2144ab5ac90aSMichal Hocko 2145ab5ac90aSMichal Hocko return page; 2146ab5ac90aSMichal Hocko } 2147ab5ac90aSMichal Hocko 2148e4e574b7SAdam Litke /* 2149099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 2150099730d6SDave Hansen */ 2151e0ec90eeSDave Hansen static 21520c397daeSMichal Hocko struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, 2153099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 2154099730d6SDave Hansen { 2155cfcaa66fSBen Widawsky struct page *page = NULL; 2156aaf14e40SMichal Hocko struct mempolicy *mpol; 2157aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 2158aaf14e40SMichal Hocko int nid; 2159aaf14e40SMichal Hocko nodemask_t *nodemask; 2160aaf14e40SMichal Hocko 2161aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 2162cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 2163cfcaa66fSBen Widawsky gfp_t gfp = gfp_mask | __GFP_NOWARN; 2164cfcaa66fSBen Widawsky 2165cfcaa66fSBen Widawsky gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL); 2166cfcaa66fSBen Widawsky page = alloc_surplus_huge_page(h, gfp, nid, nodemask, false); 2167cfcaa66fSBen Widawsky 2168cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 2169cfcaa66fSBen Widawsky nodemask = NULL; 2170cfcaa66fSBen Widawsky } 2171cfcaa66fSBen Widawsky 2172cfcaa66fSBen Widawsky if (!page) 2173b65a4edaSMike Kravetz page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask, false); 2174aaf14e40SMichal Hocko mpol_cond_put(mpol); 2175aaf14e40SMichal Hocko return page; 2176099730d6SDave Hansen } 2177099730d6SDave Hansen 2178ab5ac90aSMichal Hocko /* page migration callback function */ 21793e59fcb0SMichal Hocko struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, 2180d92bbc27SJoonsoo Kim nodemask_t *nmask, gfp_t gfp_mask) 21814db9b2efSMichal Hocko { 2182db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 21834db9b2efSMichal Hocko if (h->free_huge_pages - h->resv_huge_pages > 0) { 21843e59fcb0SMichal Hocko struct page *page; 21853e59fcb0SMichal Hocko 21863e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); 21873e59fcb0SMichal Hocko if (page) { 2188db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 21893e59fcb0SMichal Hocko return page; 21904db9b2efSMichal Hocko } 21914db9b2efSMichal Hocko } 2192db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 21934db9b2efSMichal Hocko 21940c397daeSMichal Hocko return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); 21954db9b2efSMichal Hocko } 21964db9b2efSMichal Hocko 2197ebd63723SMichal Hocko /* mempolicy aware migration callback */ 2198389c8178SMichal Hocko struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, 2199389c8178SMichal Hocko unsigned long address) 2200ebd63723SMichal Hocko { 2201ebd63723SMichal Hocko struct mempolicy *mpol; 2202ebd63723SMichal Hocko nodemask_t *nodemask; 2203ebd63723SMichal Hocko struct page *page; 2204ebd63723SMichal Hocko gfp_t gfp_mask; 2205ebd63723SMichal Hocko int node; 2206ebd63723SMichal Hocko 2207ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 2208ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 2209d92bbc27SJoonsoo Kim page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); 2210ebd63723SMichal Hocko mpol_cond_put(mpol); 2211ebd63723SMichal Hocko 2212ebd63723SMichal Hocko return page; 2213ebd63723SMichal Hocko } 2214ebd63723SMichal Hocko 2215bf50bab2SNaoya Horiguchi /* 221625985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 2217e4e574b7SAdam Litke * of size 'delta'. 2218e4e574b7SAdam Litke */ 22190a4f3d1bSLiu Xiang static int gather_surplus_pages(struct hstate *h, long delta) 22201b2a1e7bSJules Irenge __must_hold(&hugetlb_lock) 2221e4e574b7SAdam Litke { 2222e4e574b7SAdam Litke struct list_head surplus_list; 2223e4e574b7SAdam Litke struct page *page, *tmp; 22240a4f3d1bSLiu Xiang int ret; 22250a4f3d1bSLiu Xiang long i; 22260a4f3d1bSLiu Xiang long needed, allocated; 222728073b02SHillf Danton bool alloc_ok = true; 2228e4e574b7SAdam Litke 22299487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2230a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 2231ac09b3a1SAdam Litke if (needed <= 0) { 2232a5516438SAndi Kleen h->resv_huge_pages += delta; 2233e4e574b7SAdam Litke return 0; 2234ac09b3a1SAdam Litke } 2235e4e574b7SAdam Litke 2236e4e574b7SAdam Litke allocated = 0; 2237e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 2238e4e574b7SAdam Litke 2239e4e574b7SAdam Litke ret = -ENOMEM; 2240e4e574b7SAdam Litke retry: 2241db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2242e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 22430c397daeSMichal Hocko page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), 2244b65a4edaSMike Kravetz NUMA_NO_NODE, NULL, true); 224528073b02SHillf Danton if (!page) { 224628073b02SHillf Danton alloc_ok = false; 224728073b02SHillf Danton break; 224828073b02SHillf Danton } 2249e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 225069ed779aSDavid Rientjes cond_resched(); 2251e4e574b7SAdam Litke } 225228073b02SHillf Danton allocated += i; 2253e4e574b7SAdam Litke 2254e4e574b7SAdam Litke /* 2255e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 2256e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 2257e4e574b7SAdam Litke */ 2258db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2259a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 2260a5516438SAndi Kleen (h->free_huge_pages + allocated); 226128073b02SHillf Danton if (needed > 0) { 226228073b02SHillf Danton if (alloc_ok) 2263e4e574b7SAdam Litke goto retry; 226428073b02SHillf Danton /* 226528073b02SHillf Danton * We were not able to allocate enough pages to 226628073b02SHillf Danton * satisfy the entire reservation so we free what 226728073b02SHillf Danton * we've allocated so far. 226828073b02SHillf Danton */ 226928073b02SHillf Danton goto free; 227028073b02SHillf Danton } 2271e4e574b7SAdam Litke /* 2272e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 227325985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 2274e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 2275ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 2276ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 2277ac09b3a1SAdam Litke * before they are reserved. 2278e4e574b7SAdam Litke */ 2279e4e574b7SAdam Litke needed += allocated; 2280a5516438SAndi Kleen h->resv_huge_pages += delta; 2281e4e574b7SAdam Litke ret = 0; 2282a9869b83SNaoya Horiguchi 228319fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 228419fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 228519fc3f0aSAdam Litke if ((--needed) < 0) 228619fc3f0aSAdam Litke break; 2287b65a4edaSMike Kravetz /* Add the page to the hugetlb allocator */ 2288a5516438SAndi Kleen enqueue_huge_page(h, page); 228919fc3f0aSAdam Litke } 229028073b02SHillf Danton free: 2291db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 229219fc3f0aSAdam Litke 2293b65a4edaSMike Kravetz /* 2294b65a4edaSMike Kravetz * Free unnecessary surplus pages to the buddy allocator. 2295b65a4edaSMike Kravetz * Pages have no ref count, call free_huge_page directly. 2296b65a4edaSMike Kravetz */ 2297c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 2298b65a4edaSMike Kravetz free_huge_page(page); 2299db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2300e4e574b7SAdam Litke 2301e4e574b7SAdam Litke return ret; 2302e4e574b7SAdam Litke } 2303e4e574b7SAdam Litke 2304e4e574b7SAdam Litke /* 2305e5bbc8a6SMike Kravetz * This routine has two main purposes: 2306e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 2307e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 2308e5bbc8a6SMike Kravetz * to the associated reservation map. 2309e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 2310e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 2311e4e574b7SAdam Litke */ 2312a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 2313a5516438SAndi Kleen unsigned long unused_resv_pages) 2314e4e574b7SAdam Litke { 2315e4e574b7SAdam Litke unsigned long nr_pages; 231610c6ec49SMike Kravetz struct page *page; 231710c6ec49SMike Kravetz LIST_HEAD(page_list); 231810c6ec49SMike Kravetz 23199487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 232010c6ec49SMike Kravetz /* Uncommit the reservation */ 232110c6ec49SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 2322e4e574b7SAdam Litke 2323aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 2324bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2325e5bbc8a6SMike Kravetz goto out; 2326aa888a74SAndi Kleen 2327e5bbc8a6SMike Kravetz /* 2328e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 2329e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 2330e5bbc8a6SMike Kravetz */ 2331a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 2332e4e574b7SAdam Litke 2333685f3457SLee Schermerhorn /* 2334685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 23359b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 23369b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 23379b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 233810c6ec49SMike Kravetz * remove_pool_huge_page() will balance the freed pages across the 23399b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 2340685f3457SLee Schermerhorn */ 2341685f3457SLee Schermerhorn while (nr_pages--) { 234210c6ec49SMike Kravetz page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1); 234310c6ec49SMike Kravetz if (!page) 2344e5bbc8a6SMike Kravetz goto out; 234510c6ec49SMike Kravetz 234610c6ec49SMike Kravetz list_add(&page->lru, &page_list); 2347e4e574b7SAdam Litke } 2348e5bbc8a6SMike Kravetz 2349e5bbc8a6SMike Kravetz out: 2350db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 235110c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 2352db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2353e4e574b7SAdam Litke } 2354e4e574b7SAdam Litke 23555e911373SMike Kravetz 2356c37f9fb1SAndy Whitcroft /* 2357feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 23585e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 2359cf3ad20bSMike Kravetz * 2360cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 2361cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 2362cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 2363cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 2364cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 2365feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 2366feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 2367feba16e2SMike Kravetz * is called in such cases. 2368cf3ad20bSMike Kravetz * 2369cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 2370cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 2371cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 2372cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 2373cf3ad20bSMike Kravetz * take appropriate action. 237496b96a96SMike Kravetz * 237596b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 237696b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 237796b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 237896b96a96SMike Kravetz * reservation exists. 2379846be085SMike Kravetz * 2380846be085SMike Kravetz * vma_del_reservation is used in error paths where an entry in the reserve 2381846be085SMike Kravetz * map was created during huge page allocation and must be removed. It is to 2382846be085SMike Kravetz * be called after calling vma_needs_reservation to determine if a reservation 2383846be085SMike Kravetz * exists. 2384c37f9fb1SAndy Whitcroft */ 23855e911373SMike Kravetz enum vma_resv_mode { 23865e911373SMike Kravetz VMA_NEEDS_RESV, 23875e911373SMike Kravetz VMA_COMMIT_RESV, 2388feba16e2SMike Kravetz VMA_END_RESV, 238996b96a96SMike Kravetz VMA_ADD_RESV, 2390846be085SMike Kravetz VMA_DEL_RESV, 23915e911373SMike Kravetz }; 2392cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 2393cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 23945e911373SMike Kravetz enum vma_resv_mode mode) 2395c37f9fb1SAndy Whitcroft { 23964e35f483SJoonsoo Kim struct resv_map *resv; 23974e35f483SJoonsoo Kim pgoff_t idx; 2398cf3ad20bSMike Kravetz long ret; 23990db9d74eSMina Almasry long dummy_out_regions_needed; 2400c37f9fb1SAndy Whitcroft 24014e35f483SJoonsoo Kim resv = vma_resv_map(vma); 24024e35f483SJoonsoo Kim if (!resv) 2403c37f9fb1SAndy Whitcroft return 1; 2404c37f9fb1SAndy Whitcroft 24054e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 24065e911373SMike Kravetz switch (mode) { 24075e911373SMike Kravetz case VMA_NEEDS_RESV: 24080db9d74eSMina Almasry ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); 24090db9d74eSMina Almasry /* We assume that vma_reservation_* routines always operate on 24100db9d74eSMina Almasry * 1 page, and that adding to resv map a 1 page entry can only 24110db9d74eSMina Almasry * ever require 1 region. 24120db9d74eSMina Almasry */ 24130db9d74eSMina Almasry VM_BUG_ON(dummy_out_regions_needed != 1); 24145e911373SMike Kravetz break; 24155e911373SMike Kravetz case VMA_COMMIT_RESV: 2416075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 24170db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 24180db9d74eSMina Almasry VM_BUG_ON(ret < 0); 24195e911373SMike Kravetz break; 2420feba16e2SMike Kravetz case VMA_END_RESV: 24210db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 24225e911373SMike Kravetz ret = 0; 24235e911373SMike Kravetz break; 242496b96a96SMike Kravetz case VMA_ADD_RESV: 24250db9d74eSMina Almasry if (vma->vm_flags & VM_MAYSHARE) { 2426075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 24270db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 24280db9d74eSMina Almasry VM_BUG_ON(ret < 0); 24290db9d74eSMina Almasry } else { 24300db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 243196b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 243296b96a96SMike Kravetz } 243396b96a96SMike Kravetz break; 2434846be085SMike Kravetz case VMA_DEL_RESV: 2435846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 2436846be085SMike Kravetz region_abort(resv, idx, idx + 1, 1); 2437846be085SMike Kravetz ret = region_del(resv, idx, idx + 1); 2438846be085SMike Kravetz } else { 2439846be085SMike Kravetz ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 2440846be085SMike Kravetz /* region_add calls of range 1 should never fail. */ 2441846be085SMike Kravetz VM_BUG_ON(ret < 0); 2442846be085SMike Kravetz } 2443846be085SMike Kravetz break; 24445e911373SMike Kravetz default: 24455e911373SMike Kravetz BUG(); 24465e911373SMike Kravetz } 244784afd99bSAndy Whitcroft 2448846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV) 2449cf3ad20bSMike Kravetz return ret; 245067961f9dSMike Kravetz /* 2451bf3d12b9SMiaohe Lin * We know private mapping must have HPAGE_RESV_OWNER set. 2452bf3d12b9SMiaohe Lin * 245367961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 245467961f9dSMike Kravetz * However, a file associated with mapping could have been 245567961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 245667961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 245767961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 245867961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 245967961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 246067961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 246167961f9dSMike Kravetz * reservation has already been consumed. As a result, the 246267961f9dSMike Kravetz * return value of this routine is the opposite of the 246367961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 246467961f9dSMike Kravetz */ 2465bf3d12b9SMiaohe Lin if (ret > 0) 246667961f9dSMike Kravetz return 0; 2467bf3d12b9SMiaohe Lin if (ret == 0) 246867961f9dSMike Kravetz return 1; 2469bf3d12b9SMiaohe Lin return ret; 247084afd99bSAndy Whitcroft } 2471cf3ad20bSMike Kravetz 2472cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 2473a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 2474c37f9fb1SAndy Whitcroft { 24755e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 2476cf3ad20bSMike Kravetz } 2477c37f9fb1SAndy Whitcroft 2478cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 2479cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 2480cf3ad20bSMike Kravetz { 24815e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 24825e911373SMike Kravetz } 24835e911373SMike Kravetz 2484feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 24855e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 24865e911373SMike Kravetz { 2487feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 2488c37f9fb1SAndy Whitcroft } 2489c37f9fb1SAndy Whitcroft 249096b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 249196b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 249296b96a96SMike Kravetz { 249396b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 249496b96a96SMike Kravetz } 249596b96a96SMike Kravetz 2496846be085SMike Kravetz static long vma_del_reservation(struct hstate *h, 2497846be085SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 249896b96a96SMike Kravetz { 2499846be085SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV); 2500846be085SMike Kravetz } 2501846be085SMike Kravetz 2502846be085SMike Kravetz /* 2503846be085SMike Kravetz * This routine is called to restore reservation information on error paths. 2504846be085SMike Kravetz * It should ONLY be called for pages allocated via alloc_huge_page(), and 2505846be085SMike Kravetz * the hugetlb mutex should remain held when calling this routine. 2506846be085SMike Kravetz * 2507846be085SMike Kravetz * It handles two specific cases: 2508846be085SMike Kravetz * 1) A reservation was in place and the page consumed the reservation. 2509846be085SMike Kravetz * HPageRestoreReserve is set in the page. 2510846be085SMike Kravetz * 2) No reservation was in place for the page, so HPageRestoreReserve is 2511846be085SMike Kravetz * not set. However, alloc_huge_page always updates the reserve map. 2512846be085SMike Kravetz * 2513846be085SMike Kravetz * In case 1, free_huge_page later in the error path will increment the 2514846be085SMike Kravetz * global reserve count. But, free_huge_page does not have enough context 2515846be085SMike Kravetz * to adjust the reservation map. This case deals primarily with private 2516846be085SMike Kravetz * mappings. Adjust the reserve map here to be consistent with global 2517846be085SMike Kravetz * reserve count adjustments to be made by free_huge_page. Make sure the 2518846be085SMike Kravetz * reserve map indicates there is a reservation present. 2519846be085SMike Kravetz * 2520846be085SMike Kravetz * In case 2, simply undo reserve map modifications done by alloc_huge_page. 2521846be085SMike Kravetz */ 2522846be085SMike Kravetz void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma, 2523846be085SMike Kravetz unsigned long address, struct page *page) 2524846be085SMike Kravetz { 252596b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 252696b96a96SMike Kravetz 2527846be085SMike Kravetz if (HPageRestoreReserve(page)) { 2528846be085SMike Kravetz if (unlikely(rc < 0)) 252996b96a96SMike Kravetz /* 253096b96a96SMike Kravetz * Rare out of memory condition in reserve map 2531d6995da3SMike Kravetz * manipulation. Clear HPageRestoreReserve so that 253296b96a96SMike Kravetz * global reserve count will not be incremented 253396b96a96SMike Kravetz * by free_huge_page. This will make it appear 253496b96a96SMike Kravetz * as though the reservation for this page was 253596b96a96SMike Kravetz * consumed. This may prevent the task from 253696b96a96SMike Kravetz * faulting in the page at a later time. This 253796b96a96SMike Kravetz * is better than inconsistent global huge page 253896b96a96SMike Kravetz * accounting of reserve counts. 253996b96a96SMike Kravetz */ 2540d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 2541846be085SMike Kravetz else if (rc) 2542846be085SMike Kravetz (void)vma_add_reservation(h, vma, address); 2543846be085SMike Kravetz else 2544846be085SMike Kravetz vma_end_reservation(h, vma, address); 2545846be085SMike Kravetz } else { 2546846be085SMike Kravetz if (!rc) { 254796b96a96SMike Kravetz /* 2548846be085SMike Kravetz * This indicates there is an entry in the reserve map 2549c7b1850dSMike Kravetz * not added by alloc_huge_page. We know it was added 2550846be085SMike Kravetz * before the alloc_huge_page call, otherwise 2551846be085SMike Kravetz * HPageRestoreReserve would be set on the page. 2552846be085SMike Kravetz * Remove the entry so that a subsequent allocation 2553846be085SMike Kravetz * does not consume a reservation. 255496b96a96SMike Kravetz */ 2555846be085SMike Kravetz rc = vma_del_reservation(h, vma, address); 2556846be085SMike Kravetz if (rc < 0) 2557846be085SMike Kravetz /* 2558846be085SMike Kravetz * VERY rare out of memory condition. Since 2559846be085SMike Kravetz * we can not delete the entry, set 2560846be085SMike Kravetz * HPageRestoreReserve so that the reserve 2561846be085SMike Kravetz * count will be incremented when the page 2562846be085SMike Kravetz * is freed. This reserve will be consumed 2563846be085SMike Kravetz * on a subsequent allocation. 2564846be085SMike Kravetz */ 2565846be085SMike Kravetz SetHPageRestoreReserve(page); 2566846be085SMike Kravetz } else if (rc < 0) { 2567846be085SMike Kravetz /* 2568846be085SMike Kravetz * Rare out of memory condition from 2569846be085SMike Kravetz * vma_needs_reservation call. Memory allocation is 2570846be085SMike Kravetz * only attempted if a new entry is needed. Therefore, 2571846be085SMike Kravetz * this implies there is not an entry in the 2572846be085SMike Kravetz * reserve map. 2573846be085SMike Kravetz * 2574846be085SMike Kravetz * For shared mappings, no entry in the map indicates 2575846be085SMike Kravetz * no reservation. We are done. 2576846be085SMike Kravetz */ 2577846be085SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 2578846be085SMike Kravetz /* 2579846be085SMike Kravetz * For private mappings, no entry indicates 2580846be085SMike Kravetz * a reservation is present. Since we can 2581846be085SMike Kravetz * not add an entry, set SetHPageRestoreReserve 2582846be085SMike Kravetz * on the page so reserve count will be 2583846be085SMike Kravetz * incremented when freed. This reserve will 2584846be085SMike Kravetz * be consumed on a subsequent allocation. 2585846be085SMike Kravetz */ 2586846be085SMike Kravetz SetHPageRestoreReserve(page); 258796b96a96SMike Kravetz } else 2588846be085SMike Kravetz /* 2589846be085SMike Kravetz * No reservation present, do nothing 2590846be085SMike Kravetz */ 259196b96a96SMike Kravetz vma_end_reservation(h, vma, address); 259296b96a96SMike Kravetz } 259396b96a96SMike Kravetz } 259496b96a96SMike Kravetz 2595369fa227SOscar Salvador /* 2596369fa227SOscar Salvador * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one 2597369fa227SOscar Salvador * @h: struct hstate old page belongs to 2598369fa227SOscar Salvador * @old_page: Old page to dissolve 2599ae37c7ffSOscar Salvador * @list: List to isolate the page in case we need to 2600369fa227SOscar Salvador * Returns 0 on success, otherwise negated error. 2601369fa227SOscar Salvador */ 2602ae37c7ffSOscar Salvador static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page, 2603ae37c7ffSOscar Salvador struct list_head *list) 2604369fa227SOscar Salvador { 2605369fa227SOscar Salvador gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2606369fa227SOscar Salvador int nid = page_to_nid(old_page); 2607b65a4edaSMike Kravetz bool alloc_retry = false; 2608369fa227SOscar Salvador struct page *new_page; 2609369fa227SOscar Salvador int ret = 0; 2610369fa227SOscar Salvador 2611369fa227SOscar Salvador /* 2612369fa227SOscar Salvador * Before dissolving the page, we need to allocate a new one for the 2613f41f2ed4SMuchun Song * pool to remain stable. Here, we allocate the page and 'prep' it 2614f41f2ed4SMuchun Song * by doing everything but actually updating counters and adding to 2615f41f2ed4SMuchun Song * the pool. This simplifies and let us do most of the processing 2616f41f2ed4SMuchun Song * under the lock. 2617369fa227SOscar Salvador */ 2618b65a4edaSMike Kravetz alloc_retry: 2619369fa227SOscar Salvador new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL); 2620369fa227SOscar Salvador if (!new_page) 2621369fa227SOscar Salvador return -ENOMEM; 2622b65a4edaSMike Kravetz /* 2623b65a4edaSMike Kravetz * If all goes well, this page will be directly added to the free 2624b65a4edaSMike Kravetz * list in the pool. For this the ref count needs to be zero. 2625b65a4edaSMike Kravetz * Attempt to drop now, and retry once if needed. It is VERY 2626b65a4edaSMike Kravetz * unlikely there is another ref on the page. 2627b65a4edaSMike Kravetz * 2628b65a4edaSMike Kravetz * If someone else has a reference to the page, it will be freed 2629b65a4edaSMike Kravetz * when they drop their ref. Abuse temporary page flag to accomplish 2630b65a4edaSMike Kravetz * this. Retry once if there is an inflated ref count. 2631b65a4edaSMike Kravetz */ 2632b65a4edaSMike Kravetz SetHPageTemporary(new_page); 2633b65a4edaSMike Kravetz if (!put_page_testzero(new_page)) { 2634b65a4edaSMike Kravetz if (alloc_retry) 2635b65a4edaSMike Kravetz return -EBUSY; 2636b65a4edaSMike Kravetz 2637b65a4edaSMike Kravetz alloc_retry = true; 2638b65a4edaSMike Kravetz goto alloc_retry; 2639b65a4edaSMike Kravetz } 2640b65a4edaSMike Kravetz ClearHPageTemporary(new_page); 2641b65a4edaSMike Kravetz 2642f41f2ed4SMuchun Song __prep_new_huge_page(h, new_page); 2643369fa227SOscar Salvador 2644369fa227SOscar Salvador retry: 2645369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2646369fa227SOscar Salvador if (!PageHuge(old_page)) { 2647369fa227SOscar Salvador /* 2648369fa227SOscar Salvador * Freed from under us. Drop new_page too. 2649369fa227SOscar Salvador */ 2650369fa227SOscar Salvador goto free_new; 2651369fa227SOscar Salvador } else if (page_count(old_page)) { 2652369fa227SOscar Salvador /* 2653ae37c7ffSOscar Salvador * Someone has grabbed the page, try to isolate it here. 2654ae37c7ffSOscar Salvador * Fail with -EBUSY if not possible. 2655369fa227SOscar Salvador */ 2656ae37c7ffSOscar Salvador spin_unlock_irq(&hugetlb_lock); 2657ae37c7ffSOscar Salvador if (!isolate_huge_page(old_page, list)) 2658369fa227SOscar Salvador ret = -EBUSY; 2659ae37c7ffSOscar Salvador spin_lock_irq(&hugetlb_lock); 2660369fa227SOscar Salvador goto free_new; 2661369fa227SOscar Salvador } else if (!HPageFreed(old_page)) { 2662369fa227SOscar Salvador /* 2663369fa227SOscar Salvador * Page's refcount is 0 but it has not been enqueued in the 2664369fa227SOscar Salvador * freelist yet. Race window is small, so we can succeed here if 2665369fa227SOscar Salvador * we retry. 2666369fa227SOscar Salvador */ 2667369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2668369fa227SOscar Salvador cond_resched(); 2669369fa227SOscar Salvador goto retry; 2670369fa227SOscar Salvador } else { 2671369fa227SOscar Salvador /* 2672369fa227SOscar Salvador * Ok, old_page is still a genuine free hugepage. Remove it from 2673369fa227SOscar Salvador * the freelist and decrease the counters. These will be 2674369fa227SOscar Salvador * incremented again when calling __prep_account_new_huge_page() 2675369fa227SOscar Salvador * and enqueue_huge_page() for new_page. The counters will remain 2676369fa227SOscar Salvador * stable since this happens under the lock. 2677369fa227SOscar Salvador */ 2678369fa227SOscar Salvador remove_hugetlb_page(h, old_page, false); 2679369fa227SOscar Salvador 2680369fa227SOscar Salvador /* 2681b65a4edaSMike Kravetz * Ref count on new page is already zero as it was dropped 2682b65a4edaSMike Kravetz * earlier. It can be directly added to the pool free list. 2683369fa227SOscar Salvador */ 2684369fa227SOscar Salvador __prep_account_new_huge_page(h, nid); 2685369fa227SOscar Salvador enqueue_huge_page(h, new_page); 2686369fa227SOscar Salvador 2687369fa227SOscar Salvador /* 2688369fa227SOscar Salvador * Pages have been replaced, we can safely free the old one. 2689369fa227SOscar Salvador */ 2690369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2691b65d4adbSMuchun Song update_and_free_page(h, old_page, false); 2692369fa227SOscar Salvador } 2693369fa227SOscar Salvador 2694369fa227SOscar Salvador return ret; 2695369fa227SOscar Salvador 2696369fa227SOscar Salvador free_new: 2697369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2698b65a4edaSMike Kravetz /* Page has a zero ref count, but needs a ref to be freed */ 2699b65a4edaSMike Kravetz set_page_refcounted(new_page); 2700b65d4adbSMuchun Song update_and_free_page(h, new_page, false); 2701369fa227SOscar Salvador 2702369fa227SOscar Salvador return ret; 2703369fa227SOscar Salvador } 2704369fa227SOscar Salvador 2705ae37c7ffSOscar Salvador int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list) 2706369fa227SOscar Salvador { 2707369fa227SOscar Salvador struct hstate *h; 2708369fa227SOscar Salvador struct page *head; 2709ae37c7ffSOscar Salvador int ret = -EBUSY; 2710369fa227SOscar Salvador 2711369fa227SOscar Salvador /* 2712369fa227SOscar Salvador * The page might have been dissolved from under our feet, so make sure 2713369fa227SOscar Salvador * to carefully check the state under the lock. 2714369fa227SOscar Salvador * Return success when racing as if we dissolved the page ourselves. 2715369fa227SOscar Salvador */ 2716369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2717369fa227SOscar Salvador if (PageHuge(page)) { 2718369fa227SOscar Salvador head = compound_head(page); 2719369fa227SOscar Salvador h = page_hstate(head); 2720369fa227SOscar Salvador } else { 2721369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2722369fa227SOscar Salvador return 0; 2723369fa227SOscar Salvador } 2724369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2725369fa227SOscar Salvador 2726369fa227SOscar Salvador /* 2727369fa227SOscar Salvador * Fence off gigantic pages as there is a cyclic dependency between 2728369fa227SOscar Salvador * alloc_contig_range and them. Return -ENOMEM as this has the effect 2729369fa227SOscar Salvador * of bailing out right away without further retrying. 2730369fa227SOscar Salvador */ 2731369fa227SOscar Salvador if (hstate_is_gigantic(h)) 2732369fa227SOscar Salvador return -ENOMEM; 2733369fa227SOscar Salvador 2734ae37c7ffSOscar Salvador if (page_count(head) && isolate_huge_page(head, list)) 2735ae37c7ffSOscar Salvador ret = 0; 2736ae37c7ffSOscar Salvador else if (!page_count(head)) 2737ae37c7ffSOscar Salvador ret = alloc_and_dissolve_huge_page(h, head, list); 2738ae37c7ffSOscar Salvador 2739ae37c7ffSOscar Salvador return ret; 2740369fa227SOscar Salvador } 2741369fa227SOscar Salvador 274270c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 274304f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 2744348ea204SAdam Litke { 274590481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 2746a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2747348ea204SAdam Litke struct page *page; 2748d85f69b0SMike Kravetz long map_chg, map_commit; 2749d85f69b0SMike Kravetz long gbl_chg; 27506d76dcf4SAneesh Kumar K.V int ret, idx; 27516d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 275208cf9fafSMina Almasry bool deferred_reserve; 27532fc39cecSAdam Litke 27546d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 2755a1e78772SMel Gorman /* 2756d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 2757d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 2758d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 2759a1e78772SMel Gorman */ 2760d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 2761d85f69b0SMike Kravetz if (map_chg < 0) 276276dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 2763d85f69b0SMike Kravetz 2764d85f69b0SMike Kravetz /* 2765d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 2766d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 2767d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 2768d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 2769d85f69b0SMike Kravetz * checked against any subpool limit. 2770d85f69b0SMike Kravetz */ 2771d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 2772d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 2773d85f69b0SMike Kravetz if (gbl_chg < 0) { 2774feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 277576dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 27765e911373SMike Kravetz } 277790d8b7e6SAdam Litke 2778d85f69b0SMike Kravetz /* 2779d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 2780d85f69b0SMike Kravetz * map, there could be reservations associated with the 2781d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 2782d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 2783d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 2784d85f69b0SMike Kravetz * the subpool reservations. 2785d85f69b0SMike Kravetz */ 2786d85f69b0SMike Kravetz if (avoid_reserve) 2787d85f69b0SMike Kravetz gbl_chg = 1; 2788d85f69b0SMike Kravetz } 2789d85f69b0SMike Kravetz 279008cf9fafSMina Almasry /* If this allocation is not consuming a reservation, charge it now. 279108cf9fafSMina Almasry */ 27926501fe5fSMiaohe Lin deferred_reserve = map_chg || avoid_reserve; 279308cf9fafSMina Almasry if (deferred_reserve) { 279408cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 279508cf9fafSMina Almasry idx, pages_per_huge_page(h), &h_cg); 27968f34af6fSJianyu Zhan if (ret) 27978f34af6fSJianyu Zhan goto out_subpool_put; 279808cf9fafSMina Almasry } 279908cf9fafSMina Almasry 280008cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 280108cf9fafSMina Almasry if (ret) 280208cf9fafSMina Almasry goto out_uncharge_cgroup_reservation; 28038f34af6fSJianyu Zhan 2804db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2805d85f69b0SMike Kravetz /* 2806d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 2807d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 2808d85f69b0SMike Kravetz * a reservation exists for the allocation. 2809d85f69b0SMike Kravetz */ 2810d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 281181a6fcaeSJoonsoo Kim if (!page) { 2812db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 28130c397daeSMichal Hocko page = alloc_buddy_huge_page_with_mpol(h, vma, addr); 28148f34af6fSJianyu Zhan if (!page) 28158f34af6fSJianyu Zhan goto out_uncharge_cgroup; 2816a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 2817d6995da3SMike Kravetz SetHPageRestoreReserve(page); 2818a88c7695SNaoya Horiguchi h->resv_huge_pages--; 2819a88c7695SNaoya Horiguchi } 2820db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 282115a8d68eSWei Yang list_add(&page->lru, &h->hugepage_activelist); 282281a6fcaeSJoonsoo Kim /* Fall through */ 2823a1e78772SMel Gorman } 282481a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 282508cf9fafSMina Almasry /* If allocation is not consuming a reservation, also store the 282608cf9fafSMina Almasry * hugetlb_cgroup pointer on the page. 282708cf9fafSMina Almasry */ 282808cf9fafSMina Almasry if (deferred_reserve) { 282908cf9fafSMina Almasry hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), 283008cf9fafSMina Almasry h_cg, page); 283108cf9fafSMina Almasry } 283208cf9fafSMina Almasry 2833db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2834a1e78772SMel Gorman 2835d6995da3SMike Kravetz hugetlb_set_page_subpool(page, spool); 2836a1e78772SMel Gorman 2837d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 2838d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 283933039678SMike Kravetz /* 284033039678SMike Kravetz * The page was added to the reservation map between 284133039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 284233039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 284333039678SMike Kravetz * Adjust for the subpool count incremented above AND 284433039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 284533039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 284633039678SMike Kravetz * no longer applies. 284733039678SMike Kravetz */ 284833039678SMike Kravetz long rsv_adjust; 284933039678SMike Kravetz 285033039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 285133039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 285279aa925bSMike Kravetz if (deferred_reserve) 285379aa925bSMike Kravetz hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 285479aa925bSMike Kravetz pages_per_huge_page(h), page); 285533039678SMike Kravetz } 28567893d1d5SAdam Litke return page; 28578f34af6fSJianyu Zhan 28588f34af6fSJianyu Zhan out_uncharge_cgroup: 28598f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 286008cf9fafSMina Almasry out_uncharge_cgroup_reservation: 286108cf9fafSMina Almasry if (deferred_reserve) 286208cf9fafSMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), 286308cf9fafSMina Almasry h_cg); 28648f34af6fSJianyu Zhan out_subpool_put: 2865d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 28668f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 2867feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 28688f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 2869b45b5bd6SDavid Gibson } 2870b45b5bd6SDavid Gibson 2871e24a1307SAneesh Kumar K.V int alloc_bootmem_huge_page(struct hstate *h) 2872e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 2873e24a1307SAneesh Kumar K.V int __alloc_bootmem_huge_page(struct hstate *h) 2874aa888a74SAndi Kleen { 2875aa888a74SAndi Kleen struct huge_bootmem_page *m; 2876b2261026SJoonsoo Kim int nr_nodes, node; 2877aa888a74SAndi Kleen 2878b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 2879aa888a74SAndi Kleen void *addr; 2880aa888a74SAndi Kleen 2881eb31d559SMike Rapoport addr = memblock_alloc_try_nid_raw( 28828b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 288397ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 2884aa888a74SAndi Kleen if (addr) { 2885aa888a74SAndi Kleen /* 2886aa888a74SAndi Kleen * Use the beginning of the huge page to store the 2887aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 2888aa888a74SAndi Kleen * puts them into the mem_map). 2889aa888a74SAndi Kleen */ 2890aa888a74SAndi Kleen m = addr; 2891aa888a74SAndi Kleen goto found; 2892aa888a74SAndi Kleen } 2893aa888a74SAndi Kleen } 2894aa888a74SAndi Kleen return 0; 2895aa888a74SAndi Kleen 2896aa888a74SAndi Kleen found: 2897df994eadSLuiz Capitulino BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); 2898aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 2899330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 2900aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 2901aa888a74SAndi Kleen m->hstate = h; 2902aa888a74SAndi Kleen return 1; 2903aa888a74SAndi Kleen } 2904aa888a74SAndi Kleen 290548b8d744SMike Kravetz /* 290648b8d744SMike Kravetz * Put bootmem huge pages into the standard lists after mem_map is up. 290748b8d744SMike Kravetz * Note: This only applies to gigantic (order > MAX_ORDER) pages. 290848b8d744SMike Kravetz */ 2909aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 2910aa888a74SAndi Kleen { 2911aa888a74SAndi Kleen struct huge_bootmem_page *m; 2912aa888a74SAndi Kleen 2913aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 291440d18ebfSMike Kravetz struct page *page = virt_to_page(m); 2915aa888a74SAndi Kleen struct hstate *h = m->hstate; 2916ee8f248dSBecky Bruce 291748b8d744SMike Kravetz VM_BUG_ON(!hstate_is_gigantic(h)); 2918aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 29197118fc29SMike Kravetz if (prep_compound_gigantic_page(page, huge_page_order(h))) { 2920ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 2921aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 29227118fc29SMike Kravetz put_page(page); /* add to the hugepage allocator */ 29237118fc29SMike Kravetz } else { 2924416d85edSMike Kravetz /* VERY unlikely inflated ref count on a tail page */ 29257118fc29SMike Kravetz free_gigantic_page(page, huge_page_order(h)); 29267118fc29SMike Kravetz } 2927af0fb9dfSMichal Hocko 2928b0320c7bSRafael Aquini /* 292948b8d744SMike Kravetz * We need to restore the 'stolen' pages to totalram_pages 293048b8d744SMike Kravetz * in order to fix confusing memory reports from free(1) and 293148b8d744SMike Kravetz * other side-effects, like CommitLimit going negative. 2932b0320c7bSRafael Aquini */ 2933c78a7f36SMiaohe Lin adjust_managed_page_count(page, pages_per_huge_page(h)); 2934520495feSCannon Matthews cond_resched(); 2935aa888a74SAndi Kleen } 2936aa888a74SAndi Kleen } 2937aa888a74SAndi Kleen 29388faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 29391da177e4SLinus Torvalds { 29401da177e4SLinus Torvalds unsigned long i; 2941f60858f9SMike Kravetz nodemask_t *node_alloc_noretry; 2942f60858f9SMike Kravetz 2943f60858f9SMike Kravetz if (!hstate_is_gigantic(h)) { 2944f60858f9SMike Kravetz /* 2945f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 2946f60858f9SMike Kravetz * Ignore errors as lower level routines can deal with 2947f60858f9SMike Kravetz * node_alloc_noretry == NULL. If this kmalloc fails at boot 2948f60858f9SMike Kravetz * time, we are likely in bigger trouble. 2949f60858f9SMike Kravetz */ 2950f60858f9SMike Kravetz node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), 2951f60858f9SMike Kravetz GFP_KERNEL); 2952f60858f9SMike Kravetz } else { 2953f60858f9SMike Kravetz /* allocations done at boot time */ 2954f60858f9SMike Kravetz node_alloc_noretry = NULL; 2955f60858f9SMike Kravetz } 2956f60858f9SMike Kravetz 2957f60858f9SMike Kravetz /* bit mask controlling how hard we retry per-node allocations */ 2958f60858f9SMike Kravetz if (node_alloc_noretry) 2959f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 29601da177e4SLinus Torvalds 2961e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 2962bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 2963dbda8feaSBarry Song if (hugetlb_cma_size) { 2964cf11e85fSRoman Gushchin pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); 29657ecc9565SChen Wandun goto free; 2966cf11e85fSRoman Gushchin } 2967aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 2968aa888a74SAndi Kleen break; 29690c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 2970f60858f9SMike Kravetz &node_states[N_MEMORY], 2971f60858f9SMike Kravetz node_alloc_noretry)) 29721da177e4SLinus Torvalds break; 297369ed779aSDavid Rientjes cond_resched(); 29741da177e4SLinus Torvalds } 2975d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 2976d715cf80SLiam R. Howlett char buf[32]; 2977d715cf80SLiam R. Howlett 2978c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2979d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 2980d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 29818faa8b07SAndi Kleen h->max_huge_pages = i; 2982e5ff2159SAndi Kleen } 29837ecc9565SChen Wandun free: 2984f60858f9SMike Kravetz kfree(node_alloc_noretry); 2985d715cf80SLiam R. Howlett } 2986e5ff2159SAndi Kleen 2987e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 2988e5ff2159SAndi Kleen { 2989*79dfc695SMike Kravetz struct hstate *h, *h2; 2990e5ff2159SAndi Kleen 2991e5ff2159SAndi Kleen for_each_hstate(h) { 2992641844f5SNaoya Horiguchi if (minimum_order > huge_page_order(h)) 2993641844f5SNaoya Horiguchi minimum_order = huge_page_order(h); 2994641844f5SNaoya Horiguchi 29958faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 2996bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 29978faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 2998*79dfc695SMike Kravetz 2999*79dfc695SMike Kravetz /* 3000*79dfc695SMike Kravetz * Set demote order for each hstate. Note that 3001*79dfc695SMike Kravetz * h->demote_order is initially 0. 3002*79dfc695SMike Kravetz * - We can not demote gigantic pages if runtime freeing 3003*79dfc695SMike Kravetz * is not supported, so skip this. 3004*79dfc695SMike Kravetz */ 3005*79dfc695SMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 3006*79dfc695SMike Kravetz continue; 3007*79dfc695SMike Kravetz for_each_hstate(h2) { 3008*79dfc695SMike Kravetz if (h2 == h) 3009*79dfc695SMike Kravetz continue; 3010*79dfc695SMike Kravetz if (h2->order < h->order && 3011*79dfc695SMike Kravetz h2->order > h->demote_order) 3012*79dfc695SMike Kravetz h->demote_order = h2->order; 3013*79dfc695SMike Kravetz } 3014e5ff2159SAndi Kleen } 3015641844f5SNaoya Horiguchi VM_BUG_ON(minimum_order == UINT_MAX); 3016e5ff2159SAndi Kleen } 3017e5ff2159SAndi Kleen 3018e5ff2159SAndi Kleen static void __init report_hugepages(void) 3019e5ff2159SAndi Kleen { 3020e5ff2159SAndi Kleen struct hstate *h; 3021e5ff2159SAndi Kleen 3022e5ff2159SAndi Kleen for_each_hstate(h) { 30234abd32dbSAndi Kleen char buf[32]; 3024c6247f72SMatthew Wilcox 3025c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3026ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 3027c6247f72SMatthew Wilcox buf, h->free_huge_pages); 3028e5ff2159SAndi Kleen } 3029e5ff2159SAndi Kleen } 3030e5ff2159SAndi Kleen 30311da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 30326ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 30336ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 30341da177e4SLinus Torvalds { 30354415cc8dSChristoph Lameter int i; 30361121828aSMike Kravetz LIST_HEAD(page_list); 30374415cc8dSChristoph Lameter 30389487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 3039bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3040aa888a74SAndi Kleen return; 3041aa888a74SAndi Kleen 30421121828aSMike Kravetz /* 30431121828aSMike Kravetz * Collect pages to be freed on a list, and free after dropping lock 30441121828aSMike Kravetz */ 30456ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 304610c6ec49SMike Kravetz struct page *page, *next; 3047a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 3048a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 3049a5516438SAndi Kleen if (count >= h->nr_huge_pages) 30501121828aSMike Kravetz goto out; 30511da177e4SLinus Torvalds if (PageHighMem(page)) 30521da177e4SLinus Torvalds continue; 30536eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 30541121828aSMike Kravetz list_add(&page->lru, &page_list); 30551121828aSMike Kravetz } 30561121828aSMike Kravetz } 30571121828aSMike Kravetz 30581121828aSMike Kravetz out: 3059db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 306010c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3061db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 30621da177e4SLinus Torvalds } 30631da177e4SLinus Torvalds #else 30646ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 30656ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 30661da177e4SLinus Torvalds { 30671da177e4SLinus Torvalds } 30681da177e4SLinus Torvalds #endif 30691da177e4SLinus Torvalds 307020a0307cSWu Fengguang /* 307120a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 307220a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 307320a0307cSWu Fengguang * Returns 1 if an adjustment was made. 307420a0307cSWu Fengguang */ 30756ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 30766ae11b27SLee Schermerhorn int delta) 307720a0307cSWu Fengguang { 3078b2261026SJoonsoo Kim int nr_nodes, node; 307920a0307cSWu Fengguang 30809487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 308120a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 308220a0307cSWu Fengguang 3083e8c5c824SLee Schermerhorn if (delta < 0) { 3084b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 3085b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 3086b2261026SJoonsoo Kim goto found; 3087b2261026SJoonsoo Kim } 3088b2261026SJoonsoo Kim } else { 3089b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 3090b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 3091b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 3092b2261026SJoonsoo Kim goto found; 3093e8c5c824SLee Schermerhorn } 30949a76db09SLee Schermerhorn } 3095b2261026SJoonsoo Kim return 0; 309620a0307cSWu Fengguang 3097b2261026SJoonsoo Kim found: 309820a0307cSWu Fengguang h->surplus_huge_pages += delta; 3099b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 3100b2261026SJoonsoo Kim return 1; 310120a0307cSWu Fengguang } 310220a0307cSWu Fengguang 3103a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 3104fd875dcaSMike Kravetz static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 31056ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 31061da177e4SLinus Torvalds { 31077893d1d5SAdam Litke unsigned long min_count, ret; 310810c6ec49SMike Kravetz struct page *page; 310910c6ec49SMike Kravetz LIST_HEAD(page_list); 3110f60858f9SMike Kravetz NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); 3111f60858f9SMike Kravetz 3112f60858f9SMike Kravetz /* 3113f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3114f60858f9SMike Kravetz * If we can not allocate the bit mask, do not attempt to allocate 3115f60858f9SMike Kravetz * the requested huge pages. 3116f60858f9SMike Kravetz */ 3117f60858f9SMike Kravetz if (node_alloc_noretry) 3118f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 3119f60858f9SMike Kravetz else 3120f60858f9SMike Kravetz return -ENOMEM; 31211da177e4SLinus Torvalds 312229383967SMike Kravetz /* 312329383967SMike Kravetz * resize_lock mutex prevents concurrent adjustments to number of 312429383967SMike Kravetz * pages in hstate via the proc/sysfs interfaces. 312529383967SMike Kravetz */ 312629383967SMike Kravetz mutex_lock(&h->resize_lock); 3127b65d4adbSMuchun Song flush_free_hpage_work(h); 3128db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 31294eb0716eSAlexandre Ghiti 31304eb0716eSAlexandre Ghiti /* 3131fd875dcaSMike Kravetz * Check for a node specific request. 3132fd875dcaSMike Kravetz * Changing node specific huge page count may require a corresponding 3133fd875dcaSMike Kravetz * change to the global count. In any case, the passed node mask 3134fd875dcaSMike Kravetz * (nodes_allowed) will restrict alloc/free to the specified node. 3135fd875dcaSMike Kravetz */ 3136fd875dcaSMike Kravetz if (nid != NUMA_NO_NODE) { 3137fd875dcaSMike Kravetz unsigned long old_count = count; 3138fd875dcaSMike Kravetz 3139fd875dcaSMike Kravetz count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 3140fd875dcaSMike Kravetz /* 3141fd875dcaSMike Kravetz * User may have specified a large count value which caused the 3142fd875dcaSMike Kravetz * above calculation to overflow. In this case, they wanted 3143fd875dcaSMike Kravetz * to allocate as many huge pages as possible. Set count to 3144fd875dcaSMike Kravetz * largest possible value to align with their intention. 3145fd875dcaSMike Kravetz */ 3146fd875dcaSMike Kravetz if (count < old_count) 3147fd875dcaSMike Kravetz count = ULONG_MAX; 3148fd875dcaSMike Kravetz } 3149fd875dcaSMike Kravetz 3150fd875dcaSMike Kravetz /* 31514eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 31524eb0716eSAlexandre Ghiti * page range allocation. 31534eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 31544eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 31554eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 31564eb0716eSAlexandre Ghiti */ 31574eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 31584eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 3159db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 316029383967SMike Kravetz mutex_unlock(&h->resize_lock); 3161f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 31624eb0716eSAlexandre Ghiti return -EINVAL; 31634eb0716eSAlexandre Ghiti } 31644eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 31654eb0716eSAlexandre Ghiti } 3166aa888a74SAndi Kleen 31677893d1d5SAdam Litke /* 31687893d1d5SAdam Litke * Increase the pool size 31697893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 31707893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 3171d1c3fb1fSNishanth Aravamudan * 31720c397daeSMichal Hocko * We might race with alloc_surplus_huge_page() here and be unable 3173d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 3174d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 3175d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 3176d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 31777893d1d5SAdam Litke */ 3178a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 31796ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 31807893d1d5SAdam Litke break; 31817893d1d5SAdam Litke } 31827893d1d5SAdam Litke 3183a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 31847893d1d5SAdam Litke /* 31857893d1d5SAdam Litke * If this allocation races such that we no longer need the 31867893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 31877893d1d5SAdam Litke * and reducing the surplus. 31887893d1d5SAdam Litke */ 3189db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3190649920c6SJia He 3191649920c6SJia He /* yield cpu to avoid soft lockup */ 3192649920c6SJia He cond_resched(); 3193649920c6SJia He 3194f60858f9SMike Kravetz ret = alloc_pool_huge_page(h, nodes_allowed, 3195f60858f9SMike Kravetz node_alloc_noretry); 3196db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 31977893d1d5SAdam Litke if (!ret) 31987893d1d5SAdam Litke goto out; 31997893d1d5SAdam Litke 3200536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 3201536240f2SMel Gorman if (signal_pending(current)) 3202536240f2SMel Gorman goto out; 32037893d1d5SAdam Litke } 32047893d1d5SAdam Litke 32057893d1d5SAdam Litke /* 32067893d1d5SAdam Litke * Decrease the pool size 32077893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 32087893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 32097893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 32107893d1d5SAdam Litke * to the desired size as pages become free. 3211d1c3fb1fSNishanth Aravamudan * 3212d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 3213d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 3214d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 32150c397daeSMichal Hocko * alloc_surplus_huge_page() is checking the global counter, 3216d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 3217d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 3218d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 32197893d1d5SAdam Litke */ 3220a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 32216b0c880dSAdam Litke min_count = max(count, min_count); 32226ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 322310c6ec49SMike Kravetz 322410c6ec49SMike Kravetz /* 322510c6ec49SMike Kravetz * Collect pages to be removed on list without dropping lock 322610c6ec49SMike Kravetz */ 3227a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 322810c6ec49SMike Kravetz page = remove_pool_huge_page(h, nodes_allowed, 0); 322910c6ec49SMike Kravetz if (!page) 32301da177e4SLinus Torvalds break; 323110c6ec49SMike Kravetz 323210c6ec49SMike Kravetz list_add(&page->lru, &page_list); 32331da177e4SLinus Torvalds } 323410c6ec49SMike Kravetz /* free the pages after dropping lock */ 3235db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 323610c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3237b65d4adbSMuchun Song flush_free_hpage_work(h); 3238db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 323910c6ec49SMike Kravetz 3240a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 32416ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 32427893d1d5SAdam Litke break; 32437893d1d5SAdam Litke } 32447893d1d5SAdam Litke out: 32454eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 3246db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 324729383967SMike Kravetz mutex_unlock(&h->resize_lock); 32484eb0716eSAlexandre Ghiti 3249f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 3250f60858f9SMike Kravetz 32514eb0716eSAlexandre Ghiti return 0; 32521da177e4SLinus Torvalds } 32531da177e4SLinus Torvalds 3254*79dfc695SMike Kravetz static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 3255*79dfc695SMike Kravetz __must_hold(&hugetlb_lock) 3256*79dfc695SMike Kravetz { 3257*79dfc695SMike Kravetz int rc = 0; 3258*79dfc695SMike Kravetz 3259*79dfc695SMike Kravetz lockdep_assert_held(&hugetlb_lock); 3260*79dfc695SMike Kravetz 3261*79dfc695SMike Kravetz /* We should never get here if no demote order */ 3262*79dfc695SMike Kravetz if (!h->demote_order) { 3263*79dfc695SMike Kravetz pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n"); 3264*79dfc695SMike Kravetz return -EINVAL; /* internal error */ 3265*79dfc695SMike Kravetz } 3266*79dfc695SMike Kravetz 3267*79dfc695SMike Kravetz /* 3268*79dfc695SMike Kravetz * TODO - demote fucntionality will be added in subsequent patch 3269*79dfc695SMike Kravetz */ 3270*79dfc695SMike Kravetz return rc; 3271*79dfc695SMike Kravetz } 3272*79dfc695SMike Kravetz 3273a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 3274a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 3275a3437870SNishanth Aravamudan 3276*79dfc695SMike Kravetz #define HSTATE_ATTR_WO(_name) \ 3277*79dfc695SMike Kravetz static struct kobj_attribute _name##_attr = __ATTR_WO(_name) 3278*79dfc695SMike Kravetz 3279a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 3280a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 3281a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 3282a3437870SNishanth Aravamudan 3283a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 3284a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 3285a3437870SNishanth Aravamudan 32869a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 32879a305230SLee Schermerhorn 32889a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 3289a3437870SNishanth Aravamudan { 3290a3437870SNishanth Aravamudan int i; 32919a305230SLee Schermerhorn 3292a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 32939a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 32949a305230SLee Schermerhorn if (nidp) 32959a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 3296a3437870SNishanth Aravamudan return &hstates[i]; 32979a305230SLee Schermerhorn } 32989a305230SLee Schermerhorn 32999a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 3300a3437870SNishanth Aravamudan } 3301a3437870SNishanth Aravamudan 330206808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 3303a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3304a3437870SNishanth Aravamudan { 33059a305230SLee Schermerhorn struct hstate *h; 33069a305230SLee Schermerhorn unsigned long nr_huge_pages; 33079a305230SLee Schermerhorn int nid; 33089a305230SLee Schermerhorn 33099a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 33109a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 33119a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 33129a305230SLee Schermerhorn else 33139a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 33149a305230SLee Schermerhorn 3315ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", nr_huge_pages); 3316a3437870SNishanth Aravamudan } 3317adbe8726SEric B Munson 3318238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 3319238d3c13SDavid Rientjes struct hstate *h, int nid, 3320238d3c13SDavid Rientjes unsigned long count, size_t len) 3321a3437870SNishanth Aravamudan { 3322a3437870SNishanth Aravamudan int err; 33232d0adf7eSOscar Salvador nodemask_t nodes_allowed, *n_mask; 3324a3437870SNishanth Aravamudan 33252d0adf7eSOscar Salvador if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 33262d0adf7eSOscar Salvador return -EINVAL; 3327adbe8726SEric B Munson 33289a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 33299a305230SLee Schermerhorn /* 33309a305230SLee Schermerhorn * global hstate attribute 33319a305230SLee Schermerhorn */ 33329a305230SLee Schermerhorn if (!(obey_mempolicy && 33332d0adf7eSOscar Salvador init_nodemask_of_mempolicy(&nodes_allowed))) 33342d0adf7eSOscar Salvador n_mask = &node_states[N_MEMORY]; 33352d0adf7eSOscar Salvador else 33362d0adf7eSOscar Salvador n_mask = &nodes_allowed; 33372d0adf7eSOscar Salvador } else { 33389a305230SLee Schermerhorn /* 3339fd875dcaSMike Kravetz * Node specific request. count adjustment happens in 3340fd875dcaSMike Kravetz * set_max_huge_pages() after acquiring hugetlb_lock. 33419a305230SLee Schermerhorn */ 33422d0adf7eSOscar Salvador init_nodemask_of_node(&nodes_allowed, nid); 33432d0adf7eSOscar Salvador n_mask = &nodes_allowed; 3344fd875dcaSMike Kravetz } 33459a305230SLee Schermerhorn 33462d0adf7eSOscar Salvador err = set_max_huge_pages(h, count, nid, n_mask); 334706808b08SLee Schermerhorn 33484eb0716eSAlexandre Ghiti return err ? err : len; 334906808b08SLee Schermerhorn } 335006808b08SLee Schermerhorn 3351238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 3352238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 3353238d3c13SDavid Rientjes size_t len) 3354238d3c13SDavid Rientjes { 3355238d3c13SDavid Rientjes struct hstate *h; 3356238d3c13SDavid Rientjes unsigned long count; 3357238d3c13SDavid Rientjes int nid; 3358238d3c13SDavid Rientjes int err; 3359238d3c13SDavid Rientjes 3360238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 3361238d3c13SDavid Rientjes if (err) 3362238d3c13SDavid Rientjes return err; 3363238d3c13SDavid Rientjes 3364238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 3365238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 3366238d3c13SDavid Rientjes } 3367238d3c13SDavid Rientjes 336806808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 336906808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 337006808b08SLee Schermerhorn { 337106808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 337206808b08SLee Schermerhorn } 337306808b08SLee Schermerhorn 337406808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 337506808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 337606808b08SLee Schermerhorn { 3377238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 3378a3437870SNishanth Aravamudan } 3379a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 3380a3437870SNishanth Aravamudan 338106808b08SLee Schermerhorn #ifdef CONFIG_NUMA 338206808b08SLee Schermerhorn 338306808b08SLee Schermerhorn /* 338406808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 338506808b08SLee Schermerhorn * huge page alloc/free. 338606808b08SLee Schermerhorn */ 338706808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 3388ae7a927dSJoe Perches struct kobj_attribute *attr, 3389ae7a927dSJoe Perches char *buf) 339006808b08SLee Schermerhorn { 339106808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 339206808b08SLee Schermerhorn } 339306808b08SLee Schermerhorn 339406808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 339506808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 339606808b08SLee Schermerhorn { 3397238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 339806808b08SLee Schermerhorn } 339906808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 340006808b08SLee Schermerhorn #endif 340106808b08SLee Schermerhorn 340206808b08SLee Schermerhorn 3403a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 3404a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3405a3437870SNishanth Aravamudan { 34069a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3407ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); 3408a3437870SNishanth Aravamudan } 3409adbe8726SEric B Munson 3410a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 3411a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 3412a3437870SNishanth Aravamudan { 3413a3437870SNishanth Aravamudan int err; 3414a3437870SNishanth Aravamudan unsigned long input; 34159a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3416a3437870SNishanth Aravamudan 3417bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3418adbe8726SEric B Munson return -EINVAL; 3419adbe8726SEric B Munson 34203dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 3421a3437870SNishanth Aravamudan if (err) 342273ae31e5SEric B Munson return err; 3423a3437870SNishanth Aravamudan 3424db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3425a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 3426db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3427a3437870SNishanth Aravamudan 3428a3437870SNishanth Aravamudan return count; 3429a3437870SNishanth Aravamudan } 3430a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 3431a3437870SNishanth Aravamudan 3432a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 3433a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3434a3437870SNishanth Aravamudan { 34359a305230SLee Schermerhorn struct hstate *h; 34369a305230SLee Schermerhorn unsigned long free_huge_pages; 34379a305230SLee Schermerhorn int nid; 34389a305230SLee Schermerhorn 34399a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 34409a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 34419a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 34429a305230SLee Schermerhorn else 34439a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 34449a305230SLee Schermerhorn 3445ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", free_huge_pages); 3446a3437870SNishanth Aravamudan } 3447a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 3448a3437870SNishanth Aravamudan 3449a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 3450a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3451a3437870SNishanth Aravamudan { 34529a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3453ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); 3454a3437870SNishanth Aravamudan } 3455a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 3456a3437870SNishanth Aravamudan 3457a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 3458a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3459a3437870SNishanth Aravamudan { 34609a305230SLee Schermerhorn struct hstate *h; 34619a305230SLee Schermerhorn unsigned long surplus_huge_pages; 34629a305230SLee Schermerhorn int nid; 34639a305230SLee Schermerhorn 34649a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 34659a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 34669a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 34679a305230SLee Schermerhorn else 34689a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 34699a305230SLee Schermerhorn 3470ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", surplus_huge_pages); 3471a3437870SNishanth Aravamudan } 3472a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 3473a3437870SNishanth Aravamudan 3474*79dfc695SMike Kravetz static ssize_t demote_store(struct kobject *kobj, 3475*79dfc695SMike Kravetz struct kobj_attribute *attr, const char *buf, size_t len) 3476*79dfc695SMike Kravetz { 3477*79dfc695SMike Kravetz unsigned long nr_demote; 3478*79dfc695SMike Kravetz unsigned long nr_available; 3479*79dfc695SMike Kravetz nodemask_t nodes_allowed, *n_mask; 3480*79dfc695SMike Kravetz struct hstate *h; 3481*79dfc695SMike Kravetz int err = 0; 3482*79dfc695SMike Kravetz int nid; 3483*79dfc695SMike Kravetz 3484*79dfc695SMike Kravetz err = kstrtoul(buf, 10, &nr_demote); 3485*79dfc695SMike Kravetz if (err) 3486*79dfc695SMike Kravetz return err; 3487*79dfc695SMike Kravetz h = kobj_to_hstate(kobj, &nid); 3488*79dfc695SMike Kravetz 3489*79dfc695SMike Kravetz if (nid != NUMA_NO_NODE) { 3490*79dfc695SMike Kravetz init_nodemask_of_node(&nodes_allowed, nid); 3491*79dfc695SMike Kravetz n_mask = &nodes_allowed; 3492*79dfc695SMike Kravetz } else { 3493*79dfc695SMike Kravetz n_mask = &node_states[N_MEMORY]; 3494*79dfc695SMike Kravetz } 3495*79dfc695SMike Kravetz 3496*79dfc695SMike Kravetz /* Synchronize with other sysfs operations modifying huge pages */ 3497*79dfc695SMike Kravetz mutex_lock(&h->resize_lock); 3498*79dfc695SMike Kravetz spin_lock_irq(&hugetlb_lock); 3499*79dfc695SMike Kravetz 3500*79dfc695SMike Kravetz while (nr_demote) { 3501*79dfc695SMike Kravetz /* 3502*79dfc695SMike Kravetz * Check for available pages to demote each time thorough the 3503*79dfc695SMike Kravetz * loop as demote_pool_huge_page will drop hugetlb_lock. 3504*79dfc695SMike Kravetz * 3505*79dfc695SMike Kravetz * NOTE: demote_pool_huge_page does not yet drop hugetlb_lock 3506*79dfc695SMike Kravetz * but will when full demote functionality is added in a later 3507*79dfc695SMike Kravetz * patch. 3508*79dfc695SMike Kravetz */ 3509*79dfc695SMike Kravetz if (nid != NUMA_NO_NODE) 3510*79dfc695SMike Kravetz nr_available = h->free_huge_pages_node[nid]; 3511*79dfc695SMike Kravetz else 3512*79dfc695SMike Kravetz nr_available = h->free_huge_pages; 3513*79dfc695SMike Kravetz nr_available -= h->resv_huge_pages; 3514*79dfc695SMike Kravetz if (!nr_available) 3515*79dfc695SMike Kravetz break; 3516*79dfc695SMike Kravetz 3517*79dfc695SMike Kravetz err = demote_pool_huge_page(h, n_mask); 3518*79dfc695SMike Kravetz if (err) 3519*79dfc695SMike Kravetz break; 3520*79dfc695SMike Kravetz 3521*79dfc695SMike Kravetz nr_demote--; 3522*79dfc695SMike Kravetz } 3523*79dfc695SMike Kravetz 3524*79dfc695SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3525*79dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 3526*79dfc695SMike Kravetz 3527*79dfc695SMike Kravetz if (err) 3528*79dfc695SMike Kravetz return err; 3529*79dfc695SMike Kravetz return len; 3530*79dfc695SMike Kravetz } 3531*79dfc695SMike Kravetz HSTATE_ATTR_WO(demote); 3532*79dfc695SMike Kravetz 3533*79dfc695SMike Kravetz static ssize_t demote_size_show(struct kobject *kobj, 3534*79dfc695SMike Kravetz struct kobj_attribute *attr, char *buf) 3535*79dfc695SMike Kravetz { 3536*79dfc695SMike Kravetz int nid; 3537*79dfc695SMike Kravetz struct hstate *h = kobj_to_hstate(kobj, &nid); 3538*79dfc695SMike Kravetz unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K; 3539*79dfc695SMike Kravetz 3540*79dfc695SMike Kravetz return sysfs_emit(buf, "%lukB\n", demote_size); 3541*79dfc695SMike Kravetz } 3542*79dfc695SMike Kravetz 3543*79dfc695SMike Kravetz static ssize_t demote_size_store(struct kobject *kobj, 3544*79dfc695SMike Kravetz struct kobj_attribute *attr, 3545*79dfc695SMike Kravetz const char *buf, size_t count) 3546*79dfc695SMike Kravetz { 3547*79dfc695SMike Kravetz struct hstate *h, *demote_hstate; 3548*79dfc695SMike Kravetz unsigned long demote_size; 3549*79dfc695SMike Kravetz unsigned int demote_order; 3550*79dfc695SMike Kravetz int nid; 3551*79dfc695SMike Kravetz 3552*79dfc695SMike Kravetz demote_size = (unsigned long)memparse(buf, NULL); 3553*79dfc695SMike Kravetz 3554*79dfc695SMike Kravetz demote_hstate = size_to_hstate(demote_size); 3555*79dfc695SMike Kravetz if (!demote_hstate) 3556*79dfc695SMike Kravetz return -EINVAL; 3557*79dfc695SMike Kravetz demote_order = demote_hstate->order; 3558*79dfc695SMike Kravetz 3559*79dfc695SMike Kravetz /* demote order must be smaller than hstate order */ 3560*79dfc695SMike Kravetz h = kobj_to_hstate(kobj, &nid); 3561*79dfc695SMike Kravetz if (demote_order >= h->order) 3562*79dfc695SMike Kravetz return -EINVAL; 3563*79dfc695SMike Kravetz 3564*79dfc695SMike Kravetz /* resize_lock synchronizes access to demote size and writes */ 3565*79dfc695SMike Kravetz mutex_lock(&h->resize_lock); 3566*79dfc695SMike Kravetz h->demote_order = demote_order; 3567*79dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 3568*79dfc695SMike Kravetz 3569*79dfc695SMike Kravetz return count; 3570*79dfc695SMike Kravetz } 3571*79dfc695SMike Kravetz HSTATE_ATTR(demote_size); 3572*79dfc695SMike Kravetz 3573a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 3574a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 3575a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 3576a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 3577a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 3578a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 357906808b08SLee Schermerhorn #ifdef CONFIG_NUMA 358006808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 358106808b08SLee Schermerhorn #endif 3582a3437870SNishanth Aravamudan NULL, 3583a3437870SNishanth Aravamudan }; 3584a3437870SNishanth Aravamudan 358567e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 3586a3437870SNishanth Aravamudan .attrs = hstate_attrs, 3587a3437870SNishanth Aravamudan }; 3588a3437870SNishanth Aravamudan 3589*79dfc695SMike Kravetz static struct attribute *hstate_demote_attrs[] = { 3590*79dfc695SMike Kravetz &demote_size_attr.attr, 3591*79dfc695SMike Kravetz &demote_attr.attr, 3592*79dfc695SMike Kravetz NULL, 3593*79dfc695SMike Kravetz }; 3594*79dfc695SMike Kravetz 3595*79dfc695SMike Kravetz static const struct attribute_group hstate_demote_attr_group = { 3596*79dfc695SMike Kravetz .attrs = hstate_demote_attrs, 3597*79dfc695SMike Kravetz }; 3598*79dfc695SMike Kravetz 3599094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 36009a305230SLee Schermerhorn struct kobject **hstate_kobjs, 360167e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 3602a3437870SNishanth Aravamudan { 3603a3437870SNishanth Aravamudan int retval; 3604972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 3605a3437870SNishanth Aravamudan 36069a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 36079a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 3608a3437870SNishanth Aravamudan return -ENOMEM; 3609a3437870SNishanth Aravamudan 36109a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 3611cc2205a6SMiaohe Lin if (retval) { 36129a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 3613cc2205a6SMiaohe Lin hstate_kobjs[hi] = NULL; 3614cc2205a6SMiaohe Lin } 3615a3437870SNishanth Aravamudan 3616*79dfc695SMike Kravetz if (h->demote_order) { 3617*79dfc695SMike Kravetz if (sysfs_create_group(hstate_kobjs[hi], 3618*79dfc695SMike Kravetz &hstate_demote_attr_group)) 3619*79dfc695SMike Kravetz pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name); 3620*79dfc695SMike Kravetz } 3621*79dfc695SMike Kravetz 3622a3437870SNishanth Aravamudan return retval; 3623a3437870SNishanth Aravamudan } 3624a3437870SNishanth Aravamudan 3625a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 3626a3437870SNishanth Aravamudan { 3627a3437870SNishanth Aravamudan struct hstate *h; 3628a3437870SNishanth Aravamudan int err; 3629a3437870SNishanth Aravamudan 3630a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 3631a3437870SNishanth Aravamudan if (!hugepages_kobj) 3632a3437870SNishanth Aravamudan return; 3633a3437870SNishanth Aravamudan 3634a3437870SNishanth Aravamudan for_each_hstate(h) { 36359a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 36369a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 3637a3437870SNishanth Aravamudan if (err) 3638282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s", h->name); 3639a3437870SNishanth Aravamudan } 3640a3437870SNishanth Aravamudan } 3641a3437870SNishanth Aravamudan 36429a305230SLee Schermerhorn #ifdef CONFIG_NUMA 36439a305230SLee Schermerhorn 36449a305230SLee Schermerhorn /* 36459a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 364610fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 364710fbcf4cSKay Sievers * index of a node device or _hstate == node id. 364810fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 36499a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 36509a305230SLee Schermerhorn */ 36519a305230SLee Schermerhorn struct node_hstate { 36529a305230SLee Schermerhorn struct kobject *hugepages_kobj; 36539a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 36549a305230SLee Schermerhorn }; 3655b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 36569a305230SLee Schermerhorn 36579a305230SLee Schermerhorn /* 365810fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 36599a305230SLee Schermerhorn */ 36609a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 36619a305230SLee Schermerhorn &nr_hugepages_attr.attr, 36629a305230SLee Schermerhorn &free_hugepages_attr.attr, 36639a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 36649a305230SLee Schermerhorn NULL, 36659a305230SLee Schermerhorn }; 36669a305230SLee Schermerhorn 366767e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 36689a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 36699a305230SLee Schermerhorn }; 36709a305230SLee Schermerhorn 36719a305230SLee Schermerhorn /* 367210fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 36739a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 36749a305230SLee Schermerhorn */ 36759a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 36769a305230SLee Schermerhorn { 36779a305230SLee Schermerhorn int nid; 36789a305230SLee Schermerhorn 36799a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 36809a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 36819a305230SLee Schermerhorn int i; 36829a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 36839a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 36849a305230SLee Schermerhorn if (nidp) 36859a305230SLee Schermerhorn *nidp = nid; 36869a305230SLee Schermerhorn return &hstates[i]; 36879a305230SLee Schermerhorn } 36889a305230SLee Schermerhorn } 36899a305230SLee Schermerhorn 36909a305230SLee Schermerhorn BUG(); 36919a305230SLee Schermerhorn return NULL; 36929a305230SLee Schermerhorn } 36939a305230SLee Schermerhorn 36949a305230SLee Schermerhorn /* 369510fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 36969a305230SLee Schermerhorn * No-op if no hstate attributes attached. 36979a305230SLee Schermerhorn */ 36983cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 36999a305230SLee Schermerhorn { 37009a305230SLee Schermerhorn struct hstate *h; 370110fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 37029a305230SLee Schermerhorn 37039a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 37049b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 37059a305230SLee Schermerhorn 3706972dc4deSAneesh Kumar K.V for_each_hstate(h) { 3707972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 3708972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 3709972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 3710972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 3711972dc4deSAneesh Kumar K.V } 37129a305230SLee Schermerhorn } 37139a305230SLee Schermerhorn 37149a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 37159a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 37169a305230SLee Schermerhorn } 37179a305230SLee Schermerhorn 37189a305230SLee Schermerhorn 37199a305230SLee Schermerhorn /* 372010fbcf4cSKay Sievers * Register hstate attributes for a single node device. 37219a305230SLee Schermerhorn * No-op if attributes already registered. 37229a305230SLee Schermerhorn */ 37233cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 37249a305230SLee Schermerhorn { 37259a305230SLee Schermerhorn struct hstate *h; 372610fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 37279a305230SLee Schermerhorn int err; 37289a305230SLee Schermerhorn 37299a305230SLee Schermerhorn if (nhs->hugepages_kobj) 37309a305230SLee Schermerhorn return; /* already allocated */ 37319a305230SLee Schermerhorn 37329a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 373310fbcf4cSKay Sievers &node->dev.kobj); 37349a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 37359a305230SLee Schermerhorn return; 37369a305230SLee Schermerhorn 37379a305230SLee Schermerhorn for_each_hstate(h) { 37389a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 37399a305230SLee Schermerhorn nhs->hstate_kobjs, 37409a305230SLee Schermerhorn &per_node_hstate_attr_group); 37419a305230SLee Schermerhorn if (err) { 3742282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s for node %d\n", 374310fbcf4cSKay Sievers h->name, node->dev.id); 37449a305230SLee Schermerhorn hugetlb_unregister_node(node); 37459a305230SLee Schermerhorn break; 37469a305230SLee Schermerhorn } 37479a305230SLee Schermerhorn } 37489a305230SLee Schermerhorn } 37499a305230SLee Schermerhorn 37509a305230SLee Schermerhorn /* 37519b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 375210fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 375310fbcf4cSKay Sievers * registered their associated device by this time. 37549a305230SLee Schermerhorn */ 37557d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 37569a305230SLee Schermerhorn { 37579a305230SLee Schermerhorn int nid; 37589a305230SLee Schermerhorn 37598cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 37608732794bSWen Congyang struct node *node = node_devices[nid]; 376110fbcf4cSKay Sievers if (node->dev.id == nid) 37629a305230SLee Schermerhorn hugetlb_register_node(node); 37639a305230SLee Schermerhorn } 37649a305230SLee Schermerhorn 37659a305230SLee Schermerhorn /* 376610fbcf4cSKay Sievers * Let the node device driver know we're here so it can 37679a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 37689a305230SLee Schermerhorn */ 37699a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 37709a305230SLee Schermerhorn hugetlb_unregister_node); 37719a305230SLee Schermerhorn } 37729a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 37739a305230SLee Schermerhorn 37749a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 37759a305230SLee Schermerhorn { 37769a305230SLee Schermerhorn BUG(); 37779a305230SLee Schermerhorn if (nidp) 37789a305230SLee Schermerhorn *nidp = -1; 37799a305230SLee Schermerhorn return NULL; 37809a305230SLee Schermerhorn } 37819a305230SLee Schermerhorn 37829a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 37839a305230SLee Schermerhorn 37849a305230SLee Schermerhorn #endif 37859a305230SLee Schermerhorn 3786a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 3787a3437870SNishanth Aravamudan { 37888382d914SDavidlohr Bueso int i; 37898382d914SDavidlohr Bueso 3790d6995da3SMike Kravetz BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < 3791d6995da3SMike Kravetz __NR_HPAGEFLAGS); 3792d6995da3SMike Kravetz 3793c2833a5bSMike Kravetz if (!hugepages_supported()) { 3794c2833a5bSMike Kravetz if (hugetlb_max_hstate || default_hstate_max_huge_pages) 3795c2833a5bSMike Kravetz pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); 37960ef89d25SBenjamin Herrenschmidt return 0; 3797d715cf80SLiam R. Howlett } 3798d715cf80SLiam R. Howlett 3799282f4214SMike Kravetz /* 3800282f4214SMike Kravetz * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some 3801282f4214SMike Kravetz * architectures depend on setup being done here. 3802282f4214SMike Kravetz */ 3803a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 3804282f4214SMike Kravetz if (!parsed_default_hugepagesz) { 3805282f4214SMike Kravetz /* 3806282f4214SMike Kravetz * If we did not parse a default huge page size, set 3807282f4214SMike Kravetz * default_hstate_idx to HPAGE_SIZE hstate. And, if the 3808282f4214SMike Kravetz * number of huge pages for this default size was implicitly 3809282f4214SMike Kravetz * specified, set that here as well. 3810282f4214SMike Kravetz * Note that the implicit setting will overwrite an explicit 3811282f4214SMike Kravetz * setting. A warning will be printed in this case. 3812282f4214SMike Kravetz */ 3813282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); 3814f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 3815282f4214SMike Kravetz if (default_hstate.max_huge_pages) { 3816282f4214SMike Kravetz char buf[32]; 3817282f4214SMike Kravetz 3818282f4214SMike Kravetz string_get_size(huge_page_size(&default_hstate), 3819282f4214SMike Kravetz 1, STRING_UNITS_2, buf, 32); 3820282f4214SMike Kravetz pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", 3821282f4214SMike Kravetz default_hstate.max_huge_pages, buf); 3822282f4214SMike Kravetz pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", 3823282f4214SMike Kravetz default_hstate_max_huge_pages); 3824282f4214SMike Kravetz } 3825282f4214SMike Kravetz default_hstate.max_huge_pages = 3826282f4214SMike Kravetz default_hstate_max_huge_pages; 3827282f4214SMike Kravetz } 3828f8b74815SVaishali Thakkar } 3829a3437870SNishanth Aravamudan 3830cf11e85fSRoman Gushchin hugetlb_cma_check(); 3831a3437870SNishanth Aravamudan hugetlb_init_hstates(); 3832aa888a74SAndi Kleen gather_bootmem_prealloc(); 3833a3437870SNishanth Aravamudan report_hugepages(); 3834a3437870SNishanth Aravamudan 3835a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 38369a305230SLee Schermerhorn hugetlb_register_all_nodes(); 38377179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 38389a305230SLee Schermerhorn 38398382d914SDavidlohr Bueso #ifdef CONFIG_SMP 38408382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 38418382d914SDavidlohr Bueso #else 38428382d914SDavidlohr Bueso num_fault_mutexes = 1; 38438382d914SDavidlohr Bueso #endif 3844c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 38456da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 38466da2ec56SKees Cook GFP_KERNEL); 3847c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 38488382d914SDavidlohr Bueso 38498382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 3850c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 3851a3437870SNishanth Aravamudan return 0; 3852a3437870SNishanth Aravamudan } 38533e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 3854a3437870SNishanth Aravamudan 3855ae94da89SMike Kravetz /* Overwritten by architectures with more huge page sizes */ 3856ae94da89SMike Kravetz bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) 38579fee021dSVaishali Thakkar { 3858ae94da89SMike Kravetz return size == HPAGE_SIZE; 38599fee021dSVaishali Thakkar } 38609fee021dSVaishali Thakkar 3861d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 3862a3437870SNishanth Aravamudan { 3863a3437870SNishanth Aravamudan struct hstate *h; 38648faa8b07SAndi Kleen unsigned long i; 38658faa8b07SAndi Kleen 3866a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 3867a3437870SNishanth Aravamudan return; 3868a3437870SNishanth Aravamudan } 386947d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 3870a3437870SNishanth Aravamudan BUG_ON(order == 0); 387147d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 387229383967SMike Kravetz mutex_init(&h->resize_lock); 3873a3437870SNishanth Aravamudan h->order = order; 3874aca78307SMiaohe Lin h->mask = ~(huge_page_size(h) - 1); 38758faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 38768faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 38770edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 387854f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 387954f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 3880a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 3881a3437870SNishanth Aravamudan huge_page_size(h)/1024); 388277490587SMuchun Song hugetlb_vmemmap_init(h); 38838faa8b07SAndi Kleen 3884a3437870SNishanth Aravamudan parsed_hstate = h; 3885a3437870SNishanth Aravamudan } 3886a3437870SNishanth Aravamudan 3887282f4214SMike Kravetz /* 3888282f4214SMike Kravetz * hugepages command line processing 3889282f4214SMike Kravetz * hugepages normally follows a valid hugepagsz or default_hugepagsz 3890282f4214SMike Kravetz * specification. If not, ignore the hugepages value. hugepages can also 3891282f4214SMike Kravetz * be the first huge page command line option in which case it implicitly 3892282f4214SMike Kravetz * specifies the number of huge pages for the default size. 3893282f4214SMike Kravetz */ 3894282f4214SMike Kravetz static int __init hugepages_setup(char *s) 3895a3437870SNishanth Aravamudan { 3896a3437870SNishanth Aravamudan unsigned long *mhp; 38978faa8b07SAndi Kleen static unsigned long *last_mhp; 3898a3437870SNishanth Aravamudan 38999fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 3900282f4214SMike Kravetz pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); 39019fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 3902282f4214SMike Kravetz return 0; 39039fee021dSVaishali Thakkar } 3904282f4214SMike Kravetz 3905a3437870SNishanth Aravamudan /* 3906282f4214SMike Kravetz * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter 3907282f4214SMike Kravetz * yet, so this hugepages= parameter goes to the "default hstate". 3908282f4214SMike Kravetz * Otherwise, it goes with the previously parsed hugepagesz or 3909282f4214SMike Kravetz * default_hugepagesz. 3910a3437870SNishanth Aravamudan */ 39119fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 3912a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 3913a3437870SNishanth Aravamudan else 3914a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 3915a3437870SNishanth Aravamudan 39168faa8b07SAndi Kleen if (mhp == last_mhp) { 3917282f4214SMike Kravetz pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); 3918282f4214SMike Kravetz return 0; 39198faa8b07SAndi Kleen } 39208faa8b07SAndi Kleen 3921a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 3922a3437870SNishanth Aravamudan *mhp = 0; 3923a3437870SNishanth Aravamudan 39248faa8b07SAndi Kleen /* 39258faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 392604adbc3fSMiaohe Lin * But we need to allocate gigantic hstates here early to still 39278faa8b07SAndi Kleen * use the bootmem allocator. 39288faa8b07SAndi Kleen */ 392904adbc3fSMiaohe Lin if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) 39308faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 39318faa8b07SAndi Kleen 39328faa8b07SAndi Kleen last_mhp = mhp; 39338faa8b07SAndi Kleen 3934a3437870SNishanth Aravamudan return 1; 3935a3437870SNishanth Aravamudan } 3936282f4214SMike Kravetz __setup("hugepages=", hugepages_setup); 3937e11bfbfcSNick Piggin 3938282f4214SMike Kravetz /* 3939282f4214SMike Kravetz * hugepagesz command line processing 3940282f4214SMike Kravetz * A specific huge page size can only be specified once with hugepagesz. 3941282f4214SMike Kravetz * hugepagesz is followed by hugepages on the command line. The global 3942282f4214SMike Kravetz * variable 'parsed_valid_hugepagesz' is used to determine if prior 3943282f4214SMike Kravetz * hugepagesz argument was valid. 3944282f4214SMike Kravetz */ 3945359f2544SMike Kravetz static int __init hugepagesz_setup(char *s) 3946e11bfbfcSNick Piggin { 3947359f2544SMike Kravetz unsigned long size; 3948282f4214SMike Kravetz struct hstate *h; 3949282f4214SMike Kravetz 3950282f4214SMike Kravetz parsed_valid_hugepagesz = false; 3951359f2544SMike Kravetz size = (unsigned long)memparse(s, NULL); 3952359f2544SMike Kravetz 3953359f2544SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 3954282f4214SMike Kravetz pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); 3955359f2544SMike Kravetz return 0; 3956359f2544SMike Kravetz } 3957359f2544SMike Kravetz 3958282f4214SMike Kravetz h = size_to_hstate(size); 3959282f4214SMike Kravetz if (h) { 3960282f4214SMike Kravetz /* 3961282f4214SMike Kravetz * hstate for this size already exists. This is normally 3962282f4214SMike Kravetz * an error, but is allowed if the existing hstate is the 3963282f4214SMike Kravetz * default hstate. More specifically, it is only allowed if 3964282f4214SMike Kravetz * the number of huge pages for the default hstate was not 3965282f4214SMike Kravetz * previously specified. 3966282f4214SMike Kravetz */ 3967282f4214SMike Kravetz if (!parsed_default_hugepagesz || h != &default_hstate || 3968282f4214SMike Kravetz default_hstate.max_huge_pages) { 3969282f4214SMike Kravetz pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); 397038237830SMike Kravetz return 0; 397138237830SMike Kravetz } 397238237830SMike Kravetz 3973282f4214SMike Kravetz /* 3974282f4214SMike Kravetz * No need to call hugetlb_add_hstate() as hstate already 3975282f4214SMike Kravetz * exists. But, do set parsed_hstate so that a following 3976282f4214SMike Kravetz * hugepages= parameter will be applied to this hstate. 3977282f4214SMike Kravetz */ 3978282f4214SMike Kravetz parsed_hstate = h; 3979282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3980e11bfbfcSNick Piggin return 1; 3981e11bfbfcSNick Piggin } 3982282f4214SMike Kravetz 3983359f2544SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 3984282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3985359f2544SMike Kravetz return 1; 3986359f2544SMike Kravetz } 3987359f2544SMike Kravetz __setup("hugepagesz=", hugepagesz_setup); 3988359f2544SMike Kravetz 3989282f4214SMike Kravetz /* 3990282f4214SMike Kravetz * default_hugepagesz command line input 3991282f4214SMike Kravetz * Only one instance of default_hugepagesz allowed on command line. 3992282f4214SMike Kravetz */ 3993ae94da89SMike Kravetz static int __init default_hugepagesz_setup(char *s) 3994e11bfbfcSNick Piggin { 3995ae94da89SMike Kravetz unsigned long size; 3996ae94da89SMike Kravetz 3997282f4214SMike Kravetz parsed_valid_hugepagesz = false; 3998282f4214SMike Kravetz if (parsed_default_hugepagesz) { 3999282f4214SMike Kravetz pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); 4000282f4214SMike Kravetz return 0; 4001282f4214SMike Kravetz } 4002282f4214SMike Kravetz 4003282f4214SMike Kravetz size = (unsigned long)memparse(s, NULL); 4004282f4214SMike Kravetz 4005282f4214SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4006282f4214SMike Kravetz pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); 4007282f4214SMike Kravetz return 0; 4008282f4214SMike Kravetz } 4009282f4214SMike Kravetz 4010282f4214SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4011282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4012282f4214SMike Kravetz parsed_default_hugepagesz = true; 4013282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(size)); 4014282f4214SMike Kravetz 4015282f4214SMike Kravetz /* 4016282f4214SMike Kravetz * The number of default huge pages (for this size) could have been 4017282f4214SMike Kravetz * specified as the first hugetlb parameter: hugepages=X. If so, 4018282f4214SMike Kravetz * then default_hstate_max_huge_pages is set. If the default huge 4019282f4214SMike Kravetz * page size is gigantic (>= MAX_ORDER), then the pages must be 4020282f4214SMike Kravetz * allocated here from bootmem allocator. 4021282f4214SMike Kravetz */ 4022282f4214SMike Kravetz if (default_hstate_max_huge_pages) { 4023282f4214SMike Kravetz default_hstate.max_huge_pages = default_hstate_max_huge_pages; 4024282f4214SMike Kravetz if (hstate_is_gigantic(&default_hstate)) 4025282f4214SMike Kravetz hugetlb_hstate_alloc_pages(&default_hstate); 4026282f4214SMike Kravetz default_hstate_max_huge_pages = 0; 4027282f4214SMike Kravetz } 4028282f4214SMike Kravetz 4029e11bfbfcSNick Piggin return 1; 4030e11bfbfcSNick Piggin } 4031ae94da89SMike Kravetz __setup("default_hugepagesz=", default_hugepagesz_setup); 4032a3437870SNishanth Aravamudan 40338ca39e68SMuchun Song static unsigned int allowed_mems_nr(struct hstate *h) 40348a213460SNishanth Aravamudan { 40358a213460SNishanth Aravamudan int node; 40368a213460SNishanth Aravamudan unsigned int nr = 0; 40378ca39e68SMuchun Song nodemask_t *mpol_allowed; 40388ca39e68SMuchun Song unsigned int *array = h->free_huge_pages_node; 40398ca39e68SMuchun Song gfp_t gfp_mask = htlb_alloc_mask(h); 40408a213460SNishanth Aravamudan 40418ca39e68SMuchun Song mpol_allowed = policy_nodemask_current(gfp_mask); 40428ca39e68SMuchun Song 40438ca39e68SMuchun Song for_each_node_mask(node, cpuset_current_mems_allowed) { 4044c93b0a99SJiapeng Zhong if (!mpol_allowed || node_isset(node, *mpol_allowed)) 40458a213460SNishanth Aravamudan nr += array[node]; 40468ca39e68SMuchun Song } 40478a213460SNishanth Aravamudan 40488a213460SNishanth Aravamudan return nr; 40498a213460SNishanth Aravamudan } 40508a213460SNishanth Aravamudan 40518a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 405217743798SMuchun Song static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, 405317743798SMuchun Song void *buffer, size_t *length, 405417743798SMuchun Song loff_t *ppos, unsigned long *out) 405517743798SMuchun Song { 405617743798SMuchun Song struct ctl_table dup_table; 405717743798SMuchun Song 405817743798SMuchun Song /* 405917743798SMuchun Song * In order to avoid races with __do_proc_doulongvec_minmax(), we 406017743798SMuchun Song * can duplicate the @table and alter the duplicate of it. 406117743798SMuchun Song */ 406217743798SMuchun Song dup_table = *table; 406317743798SMuchun Song dup_table.data = out; 406417743798SMuchun Song 406517743798SMuchun Song return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); 406617743798SMuchun Song } 406717743798SMuchun Song 406806808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 406906808b08SLee Schermerhorn struct ctl_table *table, int write, 407032927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 40711da177e4SLinus Torvalds { 4072e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 4073238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 407408d4a246SMichal Hocko int ret; 4075e5ff2159SAndi Kleen 4076457c1b27SNishanth Aravamudan if (!hugepages_supported()) 407786613628SJan Stancek return -EOPNOTSUPP; 4078457c1b27SNishanth Aravamudan 407917743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 408017743798SMuchun Song &tmp); 408108d4a246SMichal Hocko if (ret) 408208d4a246SMichal Hocko goto out; 4083e5ff2159SAndi Kleen 4084238d3c13SDavid Rientjes if (write) 4085238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 4086238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 408708d4a246SMichal Hocko out: 408808d4a246SMichal Hocko return ret; 40891da177e4SLinus Torvalds } 4090396faf03SMel Gorman 409106808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 409232927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 409306808b08SLee Schermerhorn { 409406808b08SLee Schermerhorn 409506808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 409606808b08SLee Schermerhorn buffer, length, ppos); 409706808b08SLee Schermerhorn } 409806808b08SLee Schermerhorn 409906808b08SLee Schermerhorn #ifdef CONFIG_NUMA 410006808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 410132927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 410206808b08SLee Schermerhorn { 410306808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 410406808b08SLee Schermerhorn buffer, length, ppos); 410506808b08SLee Schermerhorn } 410606808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 410706808b08SLee Schermerhorn 4108a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 410932927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 4110a3d0c6aaSNishanth Aravamudan { 4111a5516438SAndi Kleen struct hstate *h = &default_hstate; 4112e5ff2159SAndi Kleen unsigned long tmp; 411308d4a246SMichal Hocko int ret; 4114e5ff2159SAndi Kleen 4115457c1b27SNishanth Aravamudan if (!hugepages_supported()) 411686613628SJan Stancek return -EOPNOTSUPP; 4117457c1b27SNishanth Aravamudan 4118e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 4119e5ff2159SAndi Kleen 4120bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 4121adbe8726SEric B Munson return -EINVAL; 4122adbe8726SEric B Munson 412317743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 412417743798SMuchun Song &tmp); 412508d4a246SMichal Hocko if (ret) 412608d4a246SMichal Hocko goto out; 4127e5ff2159SAndi Kleen 4128e5ff2159SAndi Kleen if (write) { 4129db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4130e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 4131db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4132e5ff2159SAndi Kleen } 413308d4a246SMichal Hocko out: 413408d4a246SMichal Hocko return ret; 4135a3d0c6aaSNishanth Aravamudan } 4136a3d0c6aaSNishanth Aravamudan 41371da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 41381da177e4SLinus Torvalds 4139e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 41401da177e4SLinus Torvalds { 4141fcb2b0c5SRoman Gushchin struct hstate *h; 4142fcb2b0c5SRoman Gushchin unsigned long total = 0; 4143fcb2b0c5SRoman Gushchin 4144457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4145457c1b27SNishanth Aravamudan return; 4146fcb2b0c5SRoman Gushchin 4147fcb2b0c5SRoman Gushchin for_each_hstate(h) { 4148fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 4149fcb2b0c5SRoman Gushchin 4150aca78307SMiaohe Lin total += huge_page_size(h) * count; 4151fcb2b0c5SRoman Gushchin 4152fcb2b0c5SRoman Gushchin if (h == &default_hstate) 4153e1759c21SAlexey Dobriyan seq_printf(m, 41541da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 41551da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 4156b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 41577893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 41584f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 4159fcb2b0c5SRoman Gushchin count, 4160a5516438SAndi Kleen h->free_huge_pages, 4161a5516438SAndi Kleen h->resv_huge_pages, 4162a5516438SAndi Kleen h->surplus_huge_pages, 4163aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4164fcb2b0c5SRoman Gushchin } 4165fcb2b0c5SRoman Gushchin 4166aca78307SMiaohe Lin seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); 41671da177e4SLinus Torvalds } 41681da177e4SLinus Torvalds 41697981593bSJoe Perches int hugetlb_report_node_meminfo(char *buf, int len, int nid) 41701da177e4SLinus Torvalds { 4171a5516438SAndi Kleen struct hstate *h = &default_hstate; 41727981593bSJoe Perches 4173457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4174457c1b27SNishanth Aravamudan return 0; 41757981593bSJoe Perches 41767981593bSJoe Perches return sysfs_emit_at(buf, len, 41771da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 4178a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 4179a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 4180a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 4181a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 4182a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 41831da177e4SLinus Torvalds } 41841da177e4SLinus Torvalds 4185949f7ec5SDavid Rientjes void hugetlb_show_meminfo(void) 4186949f7ec5SDavid Rientjes { 4187949f7ec5SDavid Rientjes struct hstate *h; 4188949f7ec5SDavid Rientjes int nid; 4189949f7ec5SDavid Rientjes 4190457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4191457c1b27SNishanth Aravamudan return; 4192457c1b27SNishanth Aravamudan 4193949f7ec5SDavid Rientjes for_each_node_state(nid, N_MEMORY) 4194949f7ec5SDavid Rientjes for_each_hstate(h) 4195949f7ec5SDavid Rientjes pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 4196949f7ec5SDavid Rientjes nid, 4197949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 4198949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 4199949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 4200aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4201949f7ec5SDavid Rientjes } 4202949f7ec5SDavid Rientjes 42035d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 42045d317b2bSNaoya Horiguchi { 42055d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 42065d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 42075d317b2bSNaoya Horiguchi } 42085d317b2bSNaoya Horiguchi 42091da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 42101da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 42111da177e4SLinus Torvalds { 4212d0028588SWanpeng Li struct hstate *h; 4213d0028588SWanpeng Li unsigned long nr_total_pages = 0; 4214d0028588SWanpeng Li 4215d0028588SWanpeng Li for_each_hstate(h) 4216d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 4217d0028588SWanpeng Li return nr_total_pages; 42181da177e4SLinus Torvalds } 42191da177e4SLinus Torvalds 4220a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 4221fc1b8a73SMel Gorman { 4222fc1b8a73SMel Gorman int ret = -ENOMEM; 4223fc1b8a73SMel Gorman 42240aa7f354SMiaohe Lin if (!delta) 42250aa7f354SMiaohe Lin return 0; 42260aa7f354SMiaohe Lin 4227db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4228fc1b8a73SMel Gorman /* 4229fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 4230fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 4231fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 4232fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 4233fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 4234fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 4235fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 4236fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 4237fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 4238fc1b8a73SMel Gorman * 4239fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 4240fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 4241fc1b8a73SMel Gorman * we fall back to check against current free page availability as 4242fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 4243fc1b8a73SMel Gorman * semantics that cpuset has. 42448ca39e68SMuchun Song * 42458ca39e68SMuchun Song * Apart from cpuset, we also have memory policy mechanism that 42468ca39e68SMuchun Song * also determines from which node the kernel will allocate memory 42478ca39e68SMuchun Song * in a NUMA system. So similar to cpuset, we also should consider 42488ca39e68SMuchun Song * the memory policy of the current task. Similar to the description 42498ca39e68SMuchun Song * above. 4250fc1b8a73SMel Gorman */ 4251fc1b8a73SMel Gorman if (delta > 0) { 4252a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 4253fc1b8a73SMel Gorman goto out; 4254fc1b8a73SMel Gorman 42558ca39e68SMuchun Song if (delta > allowed_mems_nr(h)) { 4256a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 4257fc1b8a73SMel Gorman goto out; 4258fc1b8a73SMel Gorman } 4259fc1b8a73SMel Gorman } 4260fc1b8a73SMel Gorman 4261fc1b8a73SMel Gorman ret = 0; 4262fc1b8a73SMel Gorman if (delta < 0) 4263a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 4264fc1b8a73SMel Gorman 4265fc1b8a73SMel Gorman out: 4266db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4267fc1b8a73SMel Gorman return ret; 4268fc1b8a73SMel Gorman } 4269fc1b8a73SMel Gorman 427084afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 427184afd99bSAndy Whitcroft { 4272f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 427384afd99bSAndy Whitcroft 427484afd99bSAndy Whitcroft /* 427584afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 427684afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 427784afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 427825985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 427984afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 428084afd99bSAndy Whitcroft * new reference here without additional locking. 428184afd99bSAndy Whitcroft */ 428209a26e83SMike Kravetz if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 428309a26e83SMike Kravetz resv_map_dup_hugetlb_cgroup_uncharge_info(resv); 4284f522c3acSJoonsoo Kim kref_get(&resv->refs); 428584afd99bSAndy Whitcroft } 428609a26e83SMike Kravetz } 428784afd99bSAndy Whitcroft 4288a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 4289a1e78772SMel Gorman { 4290a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 4291f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 429290481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 42934e35f483SJoonsoo Kim unsigned long reserve, start, end; 42941c5ecae3SMike Kravetz long gbl_reserve; 429584afd99bSAndy Whitcroft 42964e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 42974e35f483SJoonsoo Kim return; 42984e35f483SJoonsoo Kim 4299a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 4300a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 430184afd99bSAndy Whitcroft 43024e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 4303e9fe92aeSMina Almasry hugetlb_cgroup_uncharge_counter(resv, start, end); 43047251ff78SAdam Litke if (reserve) { 43051c5ecae3SMike Kravetz /* 43061c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 43071c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 43081c5ecae3SMike Kravetz */ 43091c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 43101c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 43117251ff78SAdam Litke } 4312e9fe92aeSMina Almasry 4313e9fe92aeSMina Almasry kref_put(&resv->refs, resv_map_release); 4314a1e78772SMel Gorman } 4315a1e78772SMel Gorman 431631383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 431731383c68SDan Williams { 431831383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 431931383c68SDan Williams return -EINVAL; 432031383c68SDan Williams return 0; 432131383c68SDan Williams } 432231383c68SDan Williams 432305ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 432405ea8860SDan Williams { 4325aca78307SMiaohe Lin return huge_page_size(hstate_vma(vma)); 432605ea8860SDan Williams } 432705ea8860SDan Williams 43281da177e4SLinus Torvalds /* 43291da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 43301da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 43316c26d310SMiaohe Lin * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get 43321da177e4SLinus Torvalds * this far. 43331da177e4SLinus Torvalds */ 4334b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 43351da177e4SLinus Torvalds { 43361da177e4SLinus Torvalds BUG(); 4337d0217ac0SNick Piggin return 0; 43381da177e4SLinus Torvalds } 43391da177e4SLinus Torvalds 4340eec3636aSJane Chu /* 4341eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 4342eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 4343eec3636aSJane Chu * This is because under System V memory model, mappings created via 4344eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 4345eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 4346eec3636aSJane Chu */ 4347f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 4348d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 434984afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 4350a1e78772SMel Gorman .close = hugetlb_vm_op_close, 4351dd3b614fSDmitry Safonov .may_split = hugetlb_vm_op_split, 435205ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 43531da177e4SLinus Torvalds }; 43541da177e4SLinus Torvalds 43551e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 43561e8f889bSDavid Gibson int writable) 435763551ae0SDavid Gibson { 435863551ae0SDavid Gibson pte_t entry; 435979c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 436063551ae0SDavid Gibson 43611e8f889bSDavid Gibson if (writable) { 4362106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 4363106c992aSGerald Schaefer vma->vm_page_prot))); 436463551ae0SDavid Gibson } else { 4365106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 4366106c992aSGerald Schaefer vma->vm_page_prot)); 436763551ae0SDavid Gibson } 436863551ae0SDavid Gibson entry = pte_mkyoung(entry); 436963551ae0SDavid Gibson entry = pte_mkhuge(entry); 437079c1c594SChristophe Leroy entry = arch_make_huge_pte(entry, shift, vma->vm_flags); 437163551ae0SDavid Gibson 437263551ae0SDavid Gibson return entry; 437363551ae0SDavid Gibson } 437463551ae0SDavid Gibson 43751e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 43761e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 43771e8f889bSDavid Gibson { 43781e8f889bSDavid Gibson pte_t entry; 43791e8f889bSDavid Gibson 4380106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 438132f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 43824b3073e1SRussell King update_mmu_cache(vma, address, ptep); 43831e8f889bSDavid Gibson } 43841e8f889bSDavid Gibson 4385d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 43864a705fefSNaoya Horiguchi { 43874a705fefSNaoya Horiguchi swp_entry_t swp; 43884a705fefSNaoya Horiguchi 43894a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 4390d5ed7444SAneesh Kumar K.V return false; 43914a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4392d79d176aSBaoquan He if (is_migration_entry(swp)) 4393d5ed7444SAneesh Kumar K.V return true; 43944a705fefSNaoya Horiguchi else 4395d5ed7444SAneesh Kumar K.V return false; 43964a705fefSNaoya Horiguchi } 43974a705fefSNaoya Horiguchi 43983e5c3600SBaoquan He static bool is_hugetlb_entry_hwpoisoned(pte_t pte) 43994a705fefSNaoya Horiguchi { 44004a705fefSNaoya Horiguchi swp_entry_t swp; 44014a705fefSNaoya Horiguchi 44024a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 44033e5c3600SBaoquan He return false; 44044a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4405d79d176aSBaoquan He if (is_hwpoison_entry(swp)) 44063e5c3600SBaoquan He return true; 44074a705fefSNaoya Horiguchi else 44083e5c3600SBaoquan He return false; 44094a705fefSNaoya Horiguchi } 44101e8f889bSDavid Gibson 44114eae4efaSPeter Xu static void 44124eae4efaSPeter Xu hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, 44134eae4efaSPeter Xu struct page *new_page) 44144eae4efaSPeter Xu { 44154eae4efaSPeter Xu __SetPageUptodate(new_page); 44164eae4efaSPeter Xu set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1)); 44174eae4efaSPeter Xu hugepage_add_new_anon_rmap(new_page, vma, addr); 44184eae4efaSPeter Xu hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); 44194eae4efaSPeter Xu ClearHPageRestoreReserve(new_page); 44204eae4efaSPeter Xu SetHPageMigratable(new_page); 44214eae4efaSPeter Xu } 44224eae4efaSPeter Xu 442363551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 442463551ae0SDavid Gibson struct vm_area_struct *vma) 442563551ae0SDavid Gibson { 44265e41540cSMike Kravetz pte_t *src_pte, *dst_pte, entry, dst_entry; 442763551ae0SDavid Gibson struct page *ptepage; 44281c59827dSHugh Dickins unsigned long addr; 4429ca6eb14dSPeter Xu bool cow = is_cow_mapping(vma->vm_flags); 4430a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 4431a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 44324eae4efaSPeter Xu unsigned long npages = pages_per_huge_page(h); 4433c0d0381aSMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 4434ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 4435e8569dd2SAndreas Sandberg int ret = 0; 44361e8f889bSDavid Gibson 4437ac46d4f3SJérôme Glisse if (cow) { 44387269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src, 44396f4f13e8SJérôme Glisse vma->vm_start, 4440ac46d4f3SJérôme Glisse vma->vm_end); 4441ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4442c0d0381aSMike Kravetz } else { 4443c0d0381aSMike Kravetz /* 4444c0d0381aSMike Kravetz * For shared mappings i_mmap_rwsem must be held to call 4445c0d0381aSMike Kravetz * huge_pte_alloc, otherwise the returned ptep could go 4446c0d0381aSMike Kravetz * away if part of a shared pmd and another thread calls 4447c0d0381aSMike Kravetz * huge_pmd_unshare. 4448c0d0381aSMike Kravetz */ 4449c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 4450ac46d4f3SJérôme Glisse } 4451e8569dd2SAndreas Sandberg 4452a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 4453cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 44547868a208SPunit Agrawal src_pte = huge_pte_offset(src, addr, sz); 4455c74df32cSHugh Dickins if (!src_pte) 4456c74df32cSHugh Dickins continue; 4457aec44e0fSPeter Xu dst_pte = huge_pte_alloc(dst, vma, addr, sz); 4458e8569dd2SAndreas Sandberg if (!dst_pte) { 4459e8569dd2SAndreas Sandberg ret = -ENOMEM; 4460e8569dd2SAndreas Sandberg break; 4461e8569dd2SAndreas Sandberg } 4462c5c99429SLarry Woodman 44635e41540cSMike Kravetz /* 44645e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 44655e41540cSMike Kravetz * dst_pte == src_pte is the common case of src/dest sharing. 44665e41540cSMike Kravetz * 44675e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 44685e41540cSMike Kravetz * another vma. If dst_pte !none, this implies sharing. 44695e41540cSMike Kravetz * Check here before taking page table lock, and once again 44705e41540cSMike Kravetz * after taking the lock below. 44715e41540cSMike Kravetz */ 44725e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 44735e41540cSMike Kravetz if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) 4474c5c99429SLarry Woodman continue; 4475c5c99429SLarry Woodman 4476cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 4477cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 4478cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 44794a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 44805e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 44814eae4efaSPeter Xu again: 44825e41540cSMike Kravetz if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { 44835e41540cSMike Kravetz /* 44845e41540cSMike Kravetz * Skip if src entry none. Also, skip in the 44855e41540cSMike Kravetz * unlikely case dst entry !none as this implies 44865e41540cSMike Kravetz * sharing with another vma. 44875e41540cSMike Kravetz */ 44884a705fefSNaoya Horiguchi ; 44894a705fefSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry) || 44904a705fefSNaoya Horiguchi is_hugetlb_entry_hwpoisoned(entry))) { 44914a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 44924a705fefSNaoya Horiguchi 44934dd845b5SAlistair Popple if (is_writable_migration_entry(swp_entry) && cow) { 44944a705fefSNaoya Horiguchi /* 44954a705fefSNaoya Horiguchi * COW mappings require pages in both 44964a705fefSNaoya Horiguchi * parent and child to be set to read. 44974a705fefSNaoya Horiguchi */ 44984dd845b5SAlistair Popple swp_entry = make_readable_migration_entry( 44994dd845b5SAlistair Popple swp_offset(swp_entry)); 45004a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 4501e5251fd4SPunit Agrawal set_huge_swap_pte_at(src, addr, src_pte, 4502e5251fd4SPunit Agrawal entry, sz); 45034a705fefSNaoya Horiguchi } 4504e5251fd4SPunit Agrawal set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); 45054a705fefSNaoya Horiguchi } else { 45064eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 45074eae4efaSPeter Xu ptepage = pte_page(entry); 45084eae4efaSPeter Xu get_page(ptepage); 45094eae4efaSPeter Xu 45104eae4efaSPeter Xu /* 45114eae4efaSPeter Xu * This is a rare case where we see pinned hugetlb 45124eae4efaSPeter Xu * pages while they're prone to COW. We need to do the 45134eae4efaSPeter Xu * COW earlier during fork. 45144eae4efaSPeter Xu * 45154eae4efaSPeter Xu * When pre-allocating the page or copying data, we 45164eae4efaSPeter Xu * need to be without the pgtable locks since we could 45174eae4efaSPeter Xu * sleep during the process. 45184eae4efaSPeter Xu */ 45194eae4efaSPeter Xu if (unlikely(page_needs_cow_for_dma(vma, ptepage))) { 45204eae4efaSPeter Xu pte_t src_pte_old = entry; 45214eae4efaSPeter Xu struct page *new; 45224eae4efaSPeter Xu 45234eae4efaSPeter Xu spin_unlock(src_ptl); 45244eae4efaSPeter Xu spin_unlock(dst_ptl); 45254eae4efaSPeter Xu /* Do not use reserve as it's private owned */ 45264eae4efaSPeter Xu new = alloc_huge_page(vma, addr, 1); 45274eae4efaSPeter Xu if (IS_ERR(new)) { 45284eae4efaSPeter Xu put_page(ptepage); 45294eae4efaSPeter Xu ret = PTR_ERR(new); 45304eae4efaSPeter Xu break; 45314eae4efaSPeter Xu } 45324eae4efaSPeter Xu copy_user_huge_page(new, ptepage, addr, vma, 45334eae4efaSPeter Xu npages); 45344eae4efaSPeter Xu put_page(ptepage); 45354eae4efaSPeter Xu 45364eae4efaSPeter Xu /* Install the new huge page if src pte stable */ 45374eae4efaSPeter Xu dst_ptl = huge_pte_lock(h, dst, dst_pte); 45384eae4efaSPeter Xu src_ptl = huge_pte_lockptr(h, src, src_pte); 45394eae4efaSPeter Xu spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 45404eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 45414eae4efaSPeter Xu if (!pte_same(src_pte_old, entry)) { 4542846be085SMike Kravetz restore_reserve_on_error(h, vma, addr, 4543846be085SMike Kravetz new); 45444eae4efaSPeter Xu put_page(new); 45454eae4efaSPeter Xu /* dst_entry won't change as in child */ 45464eae4efaSPeter Xu goto again; 45474eae4efaSPeter Xu } 45484eae4efaSPeter Xu hugetlb_install_page(vma, dst_pte, addr, new); 45494eae4efaSPeter Xu spin_unlock(src_ptl); 45504eae4efaSPeter Xu spin_unlock(dst_ptl); 45514eae4efaSPeter Xu continue; 45524eae4efaSPeter Xu } 45534eae4efaSPeter Xu 455434ee645eSJoerg Roedel if (cow) { 45550f10851eSJérôme Glisse /* 45560f10851eSJérôme Glisse * No need to notify as we are downgrading page 45570f10851eSJérôme Glisse * table protection not changing it to point 45580f10851eSJérôme Glisse * to a new page. 45590f10851eSJérôme Glisse * 4560ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 45610f10851eSJérôme Glisse */ 45627f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 456384894e1cSPeter Xu entry = huge_pte_wrprotect(entry); 456434ee645eSJoerg Roedel } 45654eae4efaSPeter Xu 456653f9263bSKirill A. Shutemov page_dup_rmap(ptepage, true); 456763551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 45684eae4efaSPeter Xu hugetlb_count_add(npages, dst); 45691c59827dSHugh Dickins } 4570cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 4571cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 457263551ae0SDavid Gibson } 457363551ae0SDavid Gibson 4574e8569dd2SAndreas Sandberg if (cow) 4575ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 4576c0d0381aSMike Kravetz else 4577c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 4578e8569dd2SAndreas Sandberg 4579e8569dd2SAndreas Sandberg return ret; 458063551ae0SDavid Gibson } 458163551ae0SDavid Gibson 458273c54763SPeter Xu static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 458324669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 458424669e58SAneesh Kumar K.V struct page *ref_page) 458563551ae0SDavid Gibson { 458663551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 458763551ae0SDavid Gibson unsigned long address; 4588c7546f8fSDavid Gibson pte_t *ptep; 458963551ae0SDavid Gibson pte_t pte; 4590cb900f41SKirill A. Shutemov spinlock_t *ptl; 459163551ae0SDavid Gibson struct page *page; 4592a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 4593a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 4594ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 4595a5516438SAndi Kleen 459663551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 4597a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 4598a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 459963551ae0SDavid Gibson 460007e32661SAneesh Kumar K.V /* 460107e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 460207e32661SAneesh Kumar K.V * to huge page. 460307e32661SAneesh Kumar K.V */ 4604ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 460524669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 4606dff11abeSMike Kravetz 4607dff11abeSMike Kravetz /* 4608dff11abeSMike Kravetz * If sharing possible, alert mmu notifiers of worst case. 4609dff11abeSMike Kravetz */ 46106f4f13e8SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, 46116f4f13e8SJérôme Glisse end); 4612ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 4613ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4614569f48b8SHillf Danton address = start; 4615569f48b8SHillf Danton for (; address < end; address += sz) { 46167868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, sz); 4617c7546f8fSDavid Gibson if (!ptep) 4618c7546f8fSDavid Gibson continue; 4619c7546f8fSDavid Gibson 4620cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 462134ae204fSMike Kravetz if (huge_pmd_unshare(mm, vma, &address, ptep)) { 462231d49da5SAneesh Kumar K.V spin_unlock(ptl); 4623dff11abeSMike Kravetz /* 4624dff11abeSMike Kravetz * We just unmapped a page of PMDs by clearing a PUD. 4625dff11abeSMike Kravetz * The caller's TLB flush range should cover this area. 4626dff11abeSMike Kravetz */ 462731d49da5SAneesh Kumar K.V continue; 462831d49da5SAneesh Kumar K.V } 462939dde65cSChen, Kenneth W 46306629326bSHillf Danton pte = huge_ptep_get(ptep); 463131d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 463231d49da5SAneesh Kumar K.V spin_unlock(ptl); 463331d49da5SAneesh Kumar K.V continue; 463431d49da5SAneesh Kumar K.V } 46356629326bSHillf Danton 46366629326bSHillf Danton /* 46379fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 46389fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 46396629326bSHillf Danton */ 46409fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 46419386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 464231d49da5SAneesh Kumar K.V spin_unlock(ptl); 464331d49da5SAneesh Kumar K.V continue; 46448c4894c6SNaoya Horiguchi } 46456629326bSHillf Danton 46466629326bSHillf Danton page = pte_page(pte); 464704f2cbe3SMel Gorman /* 464804f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 464904f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 465004f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 465104f2cbe3SMel Gorman */ 465204f2cbe3SMel Gorman if (ref_page) { 465331d49da5SAneesh Kumar K.V if (page != ref_page) { 465431d49da5SAneesh Kumar K.V spin_unlock(ptl); 465531d49da5SAneesh Kumar K.V continue; 465631d49da5SAneesh Kumar K.V } 465704f2cbe3SMel Gorman /* 465804f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 465904f2cbe3SMel Gorman * future faults in this VMA will fail rather than 466004f2cbe3SMel Gorman * looking like data was lost 466104f2cbe3SMel Gorman */ 466204f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 466304f2cbe3SMel Gorman } 466404f2cbe3SMel Gorman 4665c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 4666b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 4667106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 46686649a386SKen Chen set_page_dirty(page); 46699e81130bSHillf Danton 46705d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 4671d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 467231d49da5SAneesh Kumar K.V 4673cb900f41SKirill A. Shutemov spin_unlock(ptl); 4674e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 467524669e58SAneesh Kumar K.V /* 467631d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 467724669e58SAneesh Kumar K.V */ 467831d49da5SAneesh Kumar K.V if (ref_page) 467931d49da5SAneesh Kumar K.V break; 4680fe1668aeSChen, Kenneth W } 4681ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 468224669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 46831da177e4SLinus Torvalds } 468463551ae0SDavid Gibson 4685d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 4686d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 4687d833352aSMel Gorman unsigned long end, struct page *ref_page) 4688d833352aSMel Gorman { 4689d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 4690d833352aSMel Gorman 4691d833352aSMel Gorman /* 4692d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 4693d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 4694d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 4695d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 4696d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 4697c8c06efaSDavidlohr Bueso * is to clear it before releasing the i_mmap_rwsem. This works 4698d833352aSMel Gorman * because in the context this is called, the VMA is about to be 4699c8c06efaSDavidlohr Bueso * destroyed and the i_mmap_rwsem is held. 4700d833352aSMel Gorman */ 4701d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 4702d833352aSMel Gorman } 4703d833352aSMel Gorman 4704502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 470504f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 4706502717f4SChen, Kenneth W { 470724669e58SAneesh Kumar K.V struct mmu_gather tlb; 4708dff11abeSMike Kravetz 4709a72afd87SWill Deacon tlb_gather_mmu(&tlb, vma->vm_mm); 471024669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 4711ae8eba8bSWill Deacon tlb_finish_mmu(&tlb); 4712502717f4SChen, Kenneth W } 4713502717f4SChen, Kenneth W 471404f2cbe3SMel Gorman /* 471504f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 4716578b7725SZhiyuan Dai * mapping it owns the reserve page for. The intention is to unmap the page 471704f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 471804f2cbe3SMel Gorman * same region. 471904f2cbe3SMel Gorman */ 47202f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 47212a4b3dedSHarvey Harrison struct page *page, unsigned long address) 472204f2cbe3SMel Gorman { 47237526674dSAdam Litke struct hstate *h = hstate_vma(vma); 472404f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 472504f2cbe3SMel Gorman struct address_space *mapping; 472604f2cbe3SMel Gorman pgoff_t pgoff; 472704f2cbe3SMel Gorman 472804f2cbe3SMel Gorman /* 472904f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 473004f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 473104f2cbe3SMel Gorman */ 47327526674dSAdam Litke address = address & huge_page_mask(h); 473336e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 473436e4f20aSMichal Hocko vma->vm_pgoff; 473593c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 473604f2cbe3SMel Gorman 47374eb2b1dcSMel Gorman /* 47384eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 47394eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 47404eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 47414eb2b1dcSMel Gorman */ 474283cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 47436b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 474404f2cbe3SMel Gorman /* Do not unmap the current VMA */ 474504f2cbe3SMel Gorman if (iter_vma == vma) 474604f2cbe3SMel Gorman continue; 474704f2cbe3SMel Gorman 474804f2cbe3SMel Gorman /* 47492f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 47502f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 47512f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 47522f84a899SMel Gorman */ 47532f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 47542f84a899SMel Gorman continue; 47552f84a899SMel Gorman 47562f84a899SMel Gorman /* 475704f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 475804f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 475904f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 476004f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 476104f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 476204f2cbe3SMel Gorman */ 476304f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 476424669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 476524669e58SAneesh Kumar K.V address + huge_page_size(h), page); 476604f2cbe3SMel Gorman } 476783cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 476804f2cbe3SMel Gorman } 476904f2cbe3SMel Gorman 47700fe6e20bSNaoya Horiguchi /* 47710fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 4772ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 4773ef009b25SMichal Hocko * cannot race with other handlers or page migration. 4774ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 47750fe6e20bSNaoya Horiguchi */ 47762b740303SSouptick Joarder static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 4777974e6d66SHuang Ying unsigned long address, pte_t *ptep, 4778cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 47791e8f889bSDavid Gibson { 47803999f52eSAneesh Kumar K.V pte_t pte; 4781a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 47821e8f889bSDavid Gibson struct page *old_page, *new_page; 47832b740303SSouptick Joarder int outside_reserve = 0; 47842b740303SSouptick Joarder vm_fault_t ret = 0; 4785974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 4786ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 47871e8f889bSDavid Gibson 47883999f52eSAneesh Kumar K.V pte = huge_ptep_get(ptep); 47891e8f889bSDavid Gibson old_page = pte_page(pte); 47901e8f889bSDavid Gibson 479104f2cbe3SMel Gorman retry_avoidcopy: 47921e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 47931e8f889bSDavid Gibson * and just make the page writable */ 479437a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 47955a49973dSHugh Dickins page_move_anon_rmap(old_page, vma); 47965b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 479783c54070SNick Piggin return 0; 47981e8f889bSDavid Gibson } 47991e8f889bSDavid Gibson 480004f2cbe3SMel Gorman /* 480104f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 480204f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 480304f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 480404f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 480504f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 480604f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 480704f2cbe3SMel Gorman * of the full address range. 480804f2cbe3SMel Gorman */ 48095944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 481004f2cbe3SMel Gorman old_page != pagecache_page) 481104f2cbe3SMel Gorman outside_reserve = 1; 481204f2cbe3SMel Gorman 481309cbfeafSKirill A. Shutemov get_page(old_page); 4814b76c8cfbSLarry Woodman 4815ad4404a2SDavidlohr Bueso /* 4816ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 4817ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 4818ad4404a2SDavidlohr Bueso */ 4819cb900f41SKirill A. Shutemov spin_unlock(ptl); 48205b7a1d40SHuang Ying new_page = alloc_huge_page(vma, haddr, outside_reserve); 48211e8f889bSDavid Gibson 48222fc39cecSAdam Litke if (IS_ERR(new_page)) { 482304f2cbe3SMel Gorman /* 482404f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 482504f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 482604f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 482704f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 482804f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 482904f2cbe3SMel Gorman */ 483004f2cbe3SMel Gorman if (outside_reserve) { 4831e7dd91c4SMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 4832e7dd91c4SMike Kravetz pgoff_t idx; 4833e7dd91c4SMike Kravetz u32 hash; 4834e7dd91c4SMike Kravetz 483509cbfeafSKirill A. Shutemov put_page(old_page); 483604f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 4837e7dd91c4SMike Kravetz /* 4838e7dd91c4SMike Kravetz * Drop hugetlb_fault_mutex and i_mmap_rwsem before 4839e7dd91c4SMike Kravetz * unmapping. unmapping needs to hold i_mmap_rwsem 4840e7dd91c4SMike Kravetz * in write mode. Dropping i_mmap_rwsem in read mode 4841e7dd91c4SMike Kravetz * here is OK as COW mappings do not interact with 4842e7dd91c4SMike Kravetz * PMD sharing. 4843e7dd91c4SMike Kravetz * 4844e7dd91c4SMike Kravetz * Reacquire both after unmap operation. 4845e7dd91c4SMike Kravetz */ 4846e7dd91c4SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 4847e7dd91c4SMike Kravetz hash = hugetlb_fault_mutex_hash(mapping, idx); 4848e7dd91c4SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 4849e7dd91c4SMike Kravetz i_mmap_unlock_read(mapping); 4850e7dd91c4SMike Kravetz 48515b7a1d40SHuang Ying unmap_ref_private(mm, vma, old_page, haddr); 4852e7dd91c4SMike Kravetz 4853e7dd91c4SMike Kravetz i_mmap_lock_read(mapping); 4854e7dd91c4SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 4855cb900f41SKirill A. Shutemov spin_lock(ptl); 48565b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4857a9af0c5dSNaoya Horiguchi if (likely(ptep && 4858a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 485904f2cbe3SMel Gorman goto retry_avoidcopy; 4860a734bcc8SHillf Danton /* 4861cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 4862cb900f41SKirill A. Shutemov * lock, and our job is done. 4863a734bcc8SHillf Danton */ 4864a734bcc8SHillf Danton return 0; 486504f2cbe3SMel Gorman } 486604f2cbe3SMel Gorman 48672b740303SSouptick Joarder ret = vmf_error(PTR_ERR(new_page)); 4868ad4404a2SDavidlohr Bueso goto out_release_old; 48691e8f889bSDavid Gibson } 48701e8f889bSDavid Gibson 48710fe6e20bSNaoya Horiguchi /* 48720fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 48730fe6e20bSNaoya Horiguchi * anon_vma prepared. 48740fe6e20bSNaoya Horiguchi */ 487544e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 4876ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 4877ad4404a2SDavidlohr Bueso goto out_release_all; 487844e2aa93SDean Nelson } 48790fe6e20bSNaoya Horiguchi 4880974e6d66SHuang Ying copy_user_huge_page(new_page, old_page, address, vma, 488147ad8475SAndrea Arcangeli pages_per_huge_page(h)); 48820ed361deSNick Piggin __SetPageUptodate(new_page); 48831e8f889bSDavid Gibson 48847269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, 48856f4f13e8SJérôme Glisse haddr + huge_page_size(h)); 4886ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4887ad4404a2SDavidlohr Bueso 4888b76c8cfbSLarry Woodman /* 4889cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 4890b76c8cfbSLarry Woodman * before the page tables are altered 4891b76c8cfbSLarry Woodman */ 4892cb900f41SKirill A. Shutemov spin_lock(ptl); 48935b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4894a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 4895d6995da3SMike Kravetz ClearHPageRestoreReserve(new_page); 489607443a85SJoonsoo Kim 48971e8f889bSDavid Gibson /* Break COW */ 48985b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 4899ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, range.end); 49005b7a1d40SHuang Ying set_huge_pte_at(mm, haddr, ptep, 49011e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 4902d281ee61SKirill A. Shutemov page_remove_rmap(old_page, true); 49035b7a1d40SHuang Ying hugepage_add_new_anon_rmap(new_page, vma, haddr); 49048f251a3dSMike Kravetz SetHPageMigratable(new_page); 49051e8f889bSDavid Gibson /* Make the old page be freed below */ 49061e8f889bSDavid Gibson new_page = old_page; 49071e8f889bSDavid Gibson } 4908cb900f41SKirill A. Shutemov spin_unlock(ptl); 4909ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 4910ad4404a2SDavidlohr Bueso out_release_all: 4911c7b1850dSMike Kravetz /* No restore in case of successful pagetable update (Break COW) */ 4912c7b1850dSMike Kravetz if (new_page != old_page) 49135b7a1d40SHuang Ying restore_reserve_on_error(h, vma, haddr, new_page); 491409cbfeafSKirill A. Shutemov put_page(new_page); 4915ad4404a2SDavidlohr Bueso out_release_old: 491609cbfeafSKirill A. Shutemov put_page(old_page); 49178312034fSJoonsoo Kim 4918ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 4919ad4404a2SDavidlohr Bueso return ret; 49201e8f889bSDavid Gibson } 49211e8f889bSDavid Gibson 492204f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 4923a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 4924a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 492504f2cbe3SMel Gorman { 492604f2cbe3SMel Gorman struct address_space *mapping; 4927e7c4b0bfSAndy Whitcroft pgoff_t idx; 492804f2cbe3SMel Gorman 492904f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 4930a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 493104f2cbe3SMel Gorman 493204f2cbe3SMel Gorman return find_lock_page(mapping, idx); 493304f2cbe3SMel Gorman } 493404f2cbe3SMel Gorman 49353ae77f43SHugh Dickins /* 49363ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 49373ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 49383ae77f43SHugh Dickins */ 49393ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 49402a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 49412a15efc9SHugh Dickins { 49422a15efc9SHugh Dickins struct address_space *mapping; 49432a15efc9SHugh Dickins pgoff_t idx; 49442a15efc9SHugh Dickins struct page *page; 49452a15efc9SHugh Dickins 49462a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 49472a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 49482a15efc9SHugh Dickins 49492a15efc9SHugh Dickins page = find_get_page(mapping, idx); 49502a15efc9SHugh Dickins if (page) 49512a15efc9SHugh Dickins put_page(page); 49522a15efc9SHugh Dickins return page != NULL; 49532a15efc9SHugh Dickins } 49542a15efc9SHugh Dickins 4955ab76ad54SMike Kravetz int huge_add_to_page_cache(struct page *page, struct address_space *mapping, 4956ab76ad54SMike Kravetz pgoff_t idx) 4957ab76ad54SMike Kravetz { 4958ab76ad54SMike Kravetz struct inode *inode = mapping->host; 4959ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 4960ab76ad54SMike Kravetz int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 4961ab76ad54SMike Kravetz 4962ab76ad54SMike Kravetz if (err) 4963ab76ad54SMike Kravetz return err; 4964d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 4965ab76ad54SMike Kravetz 496622146c3cSMike Kravetz /* 496722146c3cSMike Kravetz * set page dirty so that it will not be removed from cache/file 496822146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 496922146c3cSMike Kravetz */ 497022146c3cSMike Kravetz set_page_dirty(page); 497122146c3cSMike Kravetz 4972ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 4973ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 4974ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 4975ab76ad54SMike Kravetz return 0; 4976ab76ad54SMike Kravetz } 4977ab76ad54SMike Kravetz 49787677f7fdSAxel Rasmussen static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma, 49797677f7fdSAxel Rasmussen struct address_space *mapping, 49807677f7fdSAxel Rasmussen pgoff_t idx, 49817677f7fdSAxel Rasmussen unsigned int flags, 49827677f7fdSAxel Rasmussen unsigned long haddr, 49837677f7fdSAxel Rasmussen unsigned long reason) 49847677f7fdSAxel Rasmussen { 49857677f7fdSAxel Rasmussen vm_fault_t ret; 49867677f7fdSAxel Rasmussen u32 hash; 49877677f7fdSAxel Rasmussen struct vm_fault vmf = { 49887677f7fdSAxel Rasmussen .vma = vma, 49897677f7fdSAxel Rasmussen .address = haddr, 49907677f7fdSAxel Rasmussen .flags = flags, 49917677f7fdSAxel Rasmussen 49927677f7fdSAxel Rasmussen /* 49937677f7fdSAxel Rasmussen * Hard to debug if it ends up being 49947677f7fdSAxel Rasmussen * used by a callee that assumes 49957677f7fdSAxel Rasmussen * something about the other 49967677f7fdSAxel Rasmussen * uninitialized fields... same as in 49977677f7fdSAxel Rasmussen * memory.c 49987677f7fdSAxel Rasmussen */ 49997677f7fdSAxel Rasmussen }; 50007677f7fdSAxel Rasmussen 50017677f7fdSAxel Rasmussen /* 50027677f7fdSAxel Rasmussen * hugetlb_fault_mutex and i_mmap_rwsem must be 50037677f7fdSAxel Rasmussen * dropped before handling userfault. Reacquire 50047677f7fdSAxel Rasmussen * after handling fault to make calling code simpler. 50057677f7fdSAxel Rasmussen */ 50067677f7fdSAxel Rasmussen hash = hugetlb_fault_mutex_hash(mapping, idx); 50077677f7fdSAxel Rasmussen mutex_unlock(&hugetlb_fault_mutex_table[hash]); 50087677f7fdSAxel Rasmussen i_mmap_unlock_read(mapping); 50097677f7fdSAxel Rasmussen ret = handle_userfault(&vmf, reason); 50107677f7fdSAxel Rasmussen i_mmap_lock_read(mapping); 50117677f7fdSAxel Rasmussen mutex_lock(&hugetlb_fault_mutex_table[hash]); 50127677f7fdSAxel Rasmussen 50137677f7fdSAxel Rasmussen return ret; 50147677f7fdSAxel Rasmussen } 50157677f7fdSAxel Rasmussen 50162b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 50172b740303SSouptick Joarder struct vm_area_struct *vma, 50188382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 5019788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 5020ac9b9c66SHugh Dickins { 5021a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 50222b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 5023409eb8c2SHillf Danton int anon_rmap = 0; 50244c887265SAdam Litke unsigned long size; 50254c887265SAdam Litke struct page *page; 50261e8f889bSDavid Gibson pte_t new_pte; 5027cb900f41SKirill A. Shutemov spinlock_t *ptl; 5028285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 5029c7b1850dSMike Kravetz bool new_page, new_pagecache_page = false; 50304c887265SAdam Litke 503104f2cbe3SMel Gorman /* 503204f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 503304f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 503425985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 503504f2cbe3SMel Gorman */ 503604f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 5037910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 503804f2cbe3SMel Gorman current->pid); 503904f2cbe3SMel Gorman return ret; 504004f2cbe3SMel Gorman } 504104f2cbe3SMel Gorman 50424c887265SAdam Litke /* 504387bf91d3SMike Kravetz * We can not race with truncation due to holding i_mmap_rwsem. 504487bf91d3SMike Kravetz * i_size is modified when holding i_mmap_rwsem, so check here 504587bf91d3SMike Kravetz * once for faults beyond end of file. 50464c887265SAdam Litke */ 5047a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 5048ebed4bfcSHugh Dickins if (idx >= size) 5049ebed4bfcSHugh Dickins goto out; 50501a1aad8aSMike Kravetz 505187bf91d3SMike Kravetz retry: 5052c7b1850dSMike Kravetz new_page = false; 505387bf91d3SMike Kravetz page = find_lock_page(mapping, idx); 505487bf91d3SMike Kravetz if (!page) { 50557677f7fdSAxel Rasmussen /* Check for page in userfault range */ 50561a1aad8aSMike Kravetz if (userfaultfd_missing(vma)) { 50577677f7fdSAxel Rasmussen ret = hugetlb_handle_userfault(vma, mapping, idx, 50587677f7fdSAxel Rasmussen flags, haddr, 50597677f7fdSAxel Rasmussen VM_UFFD_MISSING); 50601a1aad8aSMike Kravetz goto out; 50611a1aad8aSMike Kravetz } 50621a1aad8aSMike Kravetz 5063285b8dcaSHuang Ying page = alloc_huge_page(vma, haddr, 0); 50642fc39cecSAdam Litke if (IS_ERR(page)) { 50654643d67eSMike Kravetz /* 50664643d67eSMike Kravetz * Returning error will result in faulting task being 50674643d67eSMike Kravetz * sent SIGBUS. The hugetlb fault mutex prevents two 50684643d67eSMike Kravetz * tasks from racing to fault in the same page which 50694643d67eSMike Kravetz * could result in false unable to allocate errors. 50704643d67eSMike Kravetz * Page migration does not take the fault mutex, but 50714643d67eSMike Kravetz * does a clear then write of pte's under page table 50724643d67eSMike Kravetz * lock. Page fault code could race with migration, 50734643d67eSMike Kravetz * notice the clear pte and try to allocate a page 50744643d67eSMike Kravetz * here. Before returning error, get ptl and make 50754643d67eSMike Kravetz * sure there really is no pte entry. 50764643d67eSMike Kravetz */ 50774643d67eSMike Kravetz ptl = huge_pte_lock(h, mm, ptep); 50784643d67eSMike Kravetz ret = 0; 5079d83e6c8aSMiaohe Lin if (huge_pte_none(huge_ptep_get(ptep))) 50802b740303SSouptick Joarder ret = vmf_error(PTR_ERR(page)); 5081d83e6c8aSMiaohe Lin spin_unlock(ptl); 50826bda666aSChristoph Lameter goto out; 50836bda666aSChristoph Lameter } 508447ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 50850ed361deSNick Piggin __SetPageUptodate(page); 5086cb6acd01SMike Kravetz new_page = true; 5087ac9b9c66SHugh Dickins 5088f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 5089ab76ad54SMike Kravetz int err = huge_add_to_page_cache(page, mapping, idx); 50906bda666aSChristoph Lameter if (err) { 50916bda666aSChristoph Lameter put_page(page); 50926bda666aSChristoph Lameter if (err == -EEXIST) 50936bda666aSChristoph Lameter goto retry; 50946bda666aSChristoph Lameter goto out; 50956bda666aSChristoph Lameter } 5096c7b1850dSMike Kravetz new_pagecache_page = true; 509723be7468SMel Gorman } else { 50986bda666aSChristoph Lameter lock_page(page); 50990fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 51000fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 51010fe6e20bSNaoya Horiguchi goto backout_unlocked; 510223be7468SMel Gorman } 5103409eb8c2SHillf Danton anon_rmap = 1; 51040fe6e20bSNaoya Horiguchi } 51050fe6e20bSNaoya Horiguchi } else { 510657303d80SAndy Whitcroft /* 5107998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 5108998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 5109998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 5110fd6a03edSNaoya Horiguchi */ 5111fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 51120eb98f15SMiaohe Lin ret = VM_FAULT_HWPOISON_LARGE | 5113972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 5114fd6a03edSNaoya Horiguchi goto backout_unlocked; 51156bda666aSChristoph Lameter } 51167677f7fdSAxel Rasmussen 51177677f7fdSAxel Rasmussen /* Check for page in userfault range. */ 51187677f7fdSAxel Rasmussen if (userfaultfd_minor(vma)) { 51197677f7fdSAxel Rasmussen unlock_page(page); 51207677f7fdSAxel Rasmussen put_page(page); 51217677f7fdSAxel Rasmussen ret = hugetlb_handle_userfault(vma, mapping, idx, 51227677f7fdSAxel Rasmussen flags, haddr, 51237677f7fdSAxel Rasmussen VM_UFFD_MINOR); 51247677f7fdSAxel Rasmussen goto out; 51257677f7fdSAxel Rasmussen } 5126998b4382SNaoya Horiguchi } 51271e8f889bSDavid Gibson 512857303d80SAndy Whitcroft /* 512957303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 513057303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 513157303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 513257303d80SAndy Whitcroft * the spinlock. 513357303d80SAndy Whitcroft */ 51345e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 5135285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 51362b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 51372b26736cSAndy Whitcroft goto backout_unlocked; 51382b26736cSAndy Whitcroft } 51395e911373SMike Kravetz /* Just decrements count, does not deallocate */ 5140285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 51415e911373SMike Kravetz } 514257303d80SAndy Whitcroft 51438bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 514483c54070SNick Piggin ret = 0; 51457f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 51464c887265SAdam Litke goto backout; 51474c887265SAdam Litke 514807443a85SJoonsoo Kim if (anon_rmap) { 5149d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 5150285b8dcaSHuang Ying hugepage_add_new_anon_rmap(page, vma, haddr); 5151ac714904SChoi Gi-yong } else 515253f9263bSKirill A. Shutemov page_dup_rmap(page, true); 51531e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 51541e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 5155285b8dcaSHuang Ying set_huge_pte_at(mm, haddr, ptep, new_pte); 51561e8f889bSDavid Gibson 51575d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 5158788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 51591e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 5160974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); 51611e8f889bSDavid Gibson } 51621e8f889bSDavid Gibson 5163cb900f41SKirill A. Shutemov spin_unlock(ptl); 5164cb6acd01SMike Kravetz 5165cb6acd01SMike Kravetz /* 51668f251a3dSMike Kravetz * Only set HPageMigratable in newly allocated pages. Existing pages 51678f251a3dSMike Kravetz * found in the pagecache may not have HPageMigratableset if they have 51688f251a3dSMike Kravetz * been isolated for migration. 5169cb6acd01SMike Kravetz */ 5170cb6acd01SMike Kravetz if (new_page) 51718f251a3dSMike Kravetz SetHPageMigratable(page); 5172cb6acd01SMike Kravetz 51734c887265SAdam Litke unlock_page(page); 51744c887265SAdam Litke out: 5175ac9b9c66SHugh Dickins return ret; 51764c887265SAdam Litke 51774c887265SAdam Litke backout: 5178cb900f41SKirill A. Shutemov spin_unlock(ptl); 51792b26736cSAndy Whitcroft backout_unlocked: 51804c887265SAdam Litke unlock_page(page); 5181c7b1850dSMike Kravetz /* restore reserve for newly allocated pages not in page cache */ 5182c7b1850dSMike Kravetz if (new_page && !new_pagecache_page) 5183285b8dcaSHuang Ying restore_reserve_on_error(h, vma, haddr, page); 51844c887265SAdam Litke put_page(page); 51854c887265SAdam Litke goto out; 5186ac9b9c66SHugh Dickins } 5187ac9b9c66SHugh Dickins 51888382d914SDavidlohr Bueso #ifdef CONFIG_SMP 5189188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 51908382d914SDavidlohr Bueso { 51918382d914SDavidlohr Bueso unsigned long key[2]; 51928382d914SDavidlohr Bueso u32 hash; 51938382d914SDavidlohr Bueso 51948382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 51958382d914SDavidlohr Bueso key[1] = idx; 51968382d914SDavidlohr Bueso 519755254636SMike Kravetz hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); 51988382d914SDavidlohr Bueso 51998382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 52008382d914SDavidlohr Bueso } 52018382d914SDavidlohr Bueso #else 52028382d914SDavidlohr Bueso /* 52036c26d310SMiaohe Lin * For uniprocessor systems we always use a single mutex, so just 52048382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 52058382d914SDavidlohr Bueso */ 5206188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 52078382d914SDavidlohr Bueso { 52088382d914SDavidlohr Bueso return 0; 52098382d914SDavidlohr Bueso } 52108382d914SDavidlohr Bueso #endif 52118382d914SDavidlohr Bueso 52122b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 5213788c7df4SHugh Dickins unsigned long address, unsigned int flags) 521486e5216fSAdam Litke { 52158382d914SDavidlohr Bueso pte_t *ptep, entry; 5216cb900f41SKirill A. Shutemov spinlock_t *ptl; 52172b740303SSouptick Joarder vm_fault_t ret; 52188382d914SDavidlohr Bueso u32 hash; 52198382d914SDavidlohr Bueso pgoff_t idx; 52200fe6e20bSNaoya Horiguchi struct page *page = NULL; 522157303d80SAndy Whitcroft struct page *pagecache_page = NULL; 5222a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 52238382d914SDavidlohr Bueso struct address_space *mapping; 52240f792cf9SNaoya Horiguchi int need_wait_lock = 0; 5225285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 522686e5216fSAdam Litke 5227285b8dcaSHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 5228fd6a03edSNaoya Horiguchi if (ptep) { 5229c0d0381aSMike Kravetz /* 5230c0d0381aSMike Kravetz * Since we hold no locks, ptep could be stale. That is 5231c0d0381aSMike Kravetz * OK as we are only making decisions based on content and 5232c0d0381aSMike Kravetz * not actually modifying content here. 5233c0d0381aSMike Kravetz */ 5234fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 5235290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 5236cb900f41SKirill A. Shutemov migration_entry_wait_huge(vma, mm, ptep); 5237290408d4SNaoya Horiguchi return 0; 5238290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 5239aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 5240972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 5241b43a9990SMike Kravetz } 52428382d914SDavidlohr Bueso 5243c0d0381aSMike Kravetz /* 5244c0d0381aSMike Kravetz * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold 524587bf91d3SMike Kravetz * until finished with ptep. This serves two purposes: 524687bf91d3SMike Kravetz * 1) It prevents huge_pmd_unshare from being called elsewhere 524787bf91d3SMike Kravetz * and making the ptep no longer valid. 524887bf91d3SMike Kravetz * 2) It synchronizes us with i_size modifications during truncation. 5249c0d0381aSMike Kravetz * 5250c0d0381aSMike Kravetz * ptep could have already be assigned via huge_pte_offset. That 5251c0d0381aSMike Kravetz * is OK, as huge_pte_alloc will return the same value unless 5252c0d0381aSMike Kravetz * something has changed. 5253c0d0381aSMike Kravetz */ 5254ddeaab32SMike Kravetz mapping = vma->vm_file->f_mapping; 5255c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 5256aec44e0fSPeter Xu ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); 5257c0d0381aSMike Kravetz if (!ptep) { 5258c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 5259c0d0381aSMike Kravetz return VM_FAULT_OOM; 5260c0d0381aSMike Kravetz } 5261ddeaab32SMike Kravetz 52623935baa9SDavid Gibson /* 52633935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 52643935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 52653935baa9SDavid Gibson * the same page in the page cache. 52663935baa9SDavid Gibson */ 5267c0d0381aSMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 5268188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 5269c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 52708382d914SDavidlohr Bueso 52717f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 52727f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 52738382d914SDavidlohr Bueso ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 5274b4d1d99fSDavid Gibson goto out_mutex; 52753935baa9SDavid Gibson } 527686e5216fSAdam Litke 527783c54070SNick Piggin ret = 0; 52781e8f889bSDavid Gibson 527957303d80SAndy Whitcroft /* 52800f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 52810f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 52827c8de358SEthon Paul * an active hugepage in pagecache. This goto expects the 2nd page 52837c8de358SEthon Paul * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will 52847c8de358SEthon Paul * properly handle it. 52850f792cf9SNaoya Horiguchi */ 52860f792cf9SNaoya Horiguchi if (!pte_present(entry)) 52870f792cf9SNaoya Horiguchi goto out_mutex; 52880f792cf9SNaoya Horiguchi 52890f792cf9SNaoya Horiguchi /* 529057303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 529157303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 529257303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 529357303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 529457303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 529557303d80SAndy Whitcroft * consumed. 529657303d80SAndy Whitcroft */ 5297106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 5298285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 52992b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 5300b4d1d99fSDavid Gibson goto out_mutex; 53012b26736cSAndy Whitcroft } 53025e911373SMike Kravetz /* Just decrements count, does not deallocate */ 5303285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 530457303d80SAndy Whitcroft 5305f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 530657303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 5307285b8dcaSHuang Ying vma, haddr); 530857303d80SAndy Whitcroft } 530957303d80SAndy Whitcroft 53100f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 53110fe6e20bSNaoya Horiguchi 53121e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 5313b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 5314cb900f41SKirill A. Shutemov goto out_ptl; 5315b4d1d99fSDavid Gibson 53160f792cf9SNaoya Horiguchi /* 53170f792cf9SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 53180f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 53190f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 53200f792cf9SNaoya Horiguchi */ 53210f792cf9SNaoya Horiguchi page = pte_page(entry); 53220f792cf9SNaoya Horiguchi if (page != pagecache_page) 53230f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 53240f792cf9SNaoya Horiguchi need_wait_lock = 1; 53250f792cf9SNaoya Horiguchi goto out_ptl; 53260f792cf9SNaoya Horiguchi } 53270f792cf9SNaoya Horiguchi 53280f792cf9SNaoya Horiguchi get_page(page); 5329b4d1d99fSDavid Gibson 5330788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 5331106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 5332974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, 5333cb900f41SKirill A. Shutemov pagecache_page, ptl); 53340f792cf9SNaoya Horiguchi goto out_put_page; 5335b4d1d99fSDavid Gibson } 5336106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 5337b4d1d99fSDavid Gibson } 5338b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 5339285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 5340788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 5341285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 53420f792cf9SNaoya Horiguchi out_put_page: 53430f792cf9SNaoya Horiguchi if (page != pagecache_page) 53440f792cf9SNaoya Horiguchi unlock_page(page); 53450f792cf9SNaoya Horiguchi put_page(page); 5346cb900f41SKirill A. Shutemov out_ptl: 5347cb900f41SKirill A. Shutemov spin_unlock(ptl); 534857303d80SAndy Whitcroft 534957303d80SAndy Whitcroft if (pagecache_page) { 535057303d80SAndy Whitcroft unlock_page(pagecache_page); 535157303d80SAndy Whitcroft put_page(pagecache_page); 535257303d80SAndy Whitcroft } 5353b4d1d99fSDavid Gibson out_mutex: 5354c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5355c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 53560f792cf9SNaoya Horiguchi /* 53570f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 53580f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 53590f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 53600f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 53610f792cf9SNaoya Horiguchi * here without taking refcount. 53620f792cf9SNaoya Horiguchi */ 53630f792cf9SNaoya Horiguchi if (need_wait_lock) 53640f792cf9SNaoya Horiguchi wait_on_page_locked(page); 53651e8f889bSDavid Gibson return ret; 536686e5216fSAdam Litke } 536786e5216fSAdam Litke 5368714c1891SAxel Rasmussen #ifdef CONFIG_USERFAULTFD 53698fb5debcSMike Kravetz /* 53708fb5debcSMike Kravetz * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with 53718fb5debcSMike Kravetz * modifications for huge pages. 53728fb5debcSMike Kravetz */ 53738fb5debcSMike Kravetz int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, 53748fb5debcSMike Kravetz pte_t *dst_pte, 53758fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 53768fb5debcSMike Kravetz unsigned long dst_addr, 53778fb5debcSMike Kravetz unsigned long src_addr, 5378f6191471SAxel Rasmussen enum mcopy_atomic_mode mode, 53798fb5debcSMike Kravetz struct page **pagep) 53808fb5debcSMike Kravetz { 5381f6191471SAxel Rasmussen bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE); 53828cc5fcbbSMina Almasry struct hstate *h = hstate_vma(dst_vma); 53838cc5fcbbSMina Almasry struct address_space *mapping = dst_vma->vm_file->f_mapping; 53848cc5fcbbSMina Almasry pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr); 53851e392147SAndrea Arcangeli unsigned long size; 53861c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 53878fb5debcSMike Kravetz pte_t _dst_pte; 53888fb5debcSMike Kravetz spinlock_t *ptl; 53898cc5fcbbSMina Almasry int ret = -ENOMEM; 53908fb5debcSMike Kravetz struct page *page; 5391f6191471SAxel Rasmussen int writable; 5392c7b1850dSMike Kravetz bool new_pagecache_page = false; 53938fb5debcSMike Kravetz 5394f6191471SAxel Rasmussen if (is_continue) { 5395f6191471SAxel Rasmussen ret = -EFAULT; 5396f6191471SAxel Rasmussen page = find_lock_page(mapping, idx); 5397f6191471SAxel Rasmussen if (!page) 5398f6191471SAxel Rasmussen goto out; 5399f6191471SAxel Rasmussen } else if (!*pagep) { 5400d84cf06eSMina Almasry /* If a page already exists, then it's UFFDIO_COPY for 5401d84cf06eSMina Almasry * a non-missing case. Return -EEXIST. 5402d84cf06eSMina Almasry */ 5403d84cf06eSMina Almasry if (vm_shared && 5404d84cf06eSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 5405d84cf06eSMina Almasry ret = -EEXIST; 54068fb5debcSMike Kravetz goto out; 5407d84cf06eSMina Almasry } 5408d84cf06eSMina Almasry 5409d84cf06eSMina Almasry page = alloc_huge_page(dst_vma, dst_addr, 0); 5410d84cf06eSMina Almasry if (IS_ERR(page)) { 5411d84cf06eSMina Almasry ret = -ENOMEM; 5412d84cf06eSMina Almasry goto out; 5413d84cf06eSMina Almasry } 54148fb5debcSMike Kravetz 54158fb5debcSMike Kravetz ret = copy_huge_page_from_user(page, 54168fb5debcSMike Kravetz (const void __user *) src_addr, 5417810a56b9SMike Kravetz pages_per_huge_page(h), false); 54188fb5debcSMike Kravetz 5419c1e8d7c6SMichel Lespinasse /* fallback to copy_from_user outside mmap_lock */ 54208fb5debcSMike Kravetz if (unlikely(ret)) { 54219e368259SAndrea Arcangeli ret = -ENOENT; 54228cc5fcbbSMina Almasry /* Free the allocated page which may have 54238cc5fcbbSMina Almasry * consumed a reservation. 54248cc5fcbbSMina Almasry */ 54258cc5fcbbSMina Almasry restore_reserve_on_error(h, dst_vma, dst_addr, page); 54268cc5fcbbSMina Almasry put_page(page); 54278cc5fcbbSMina Almasry 54288cc5fcbbSMina Almasry /* Allocate a temporary page to hold the copied 54298cc5fcbbSMina Almasry * contents. 54308cc5fcbbSMina Almasry */ 54318cc5fcbbSMina Almasry page = alloc_huge_page_vma(h, dst_vma, dst_addr); 54328cc5fcbbSMina Almasry if (!page) { 54338cc5fcbbSMina Almasry ret = -ENOMEM; 54348cc5fcbbSMina Almasry goto out; 54358cc5fcbbSMina Almasry } 54368fb5debcSMike Kravetz *pagep = page; 54378cc5fcbbSMina Almasry /* Set the outparam pagep and return to the caller to 54388cc5fcbbSMina Almasry * copy the contents outside the lock. Don't free the 54398cc5fcbbSMina Almasry * page. 54408cc5fcbbSMina Almasry */ 54418fb5debcSMike Kravetz goto out; 54428fb5debcSMike Kravetz } 54438fb5debcSMike Kravetz } else { 54448cc5fcbbSMina Almasry if (vm_shared && 54458cc5fcbbSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 54468cc5fcbbSMina Almasry put_page(*pagep); 54478cc5fcbbSMina Almasry ret = -EEXIST; 54488cc5fcbbSMina Almasry *pagep = NULL; 54498cc5fcbbSMina Almasry goto out; 54508cc5fcbbSMina Almasry } 54518cc5fcbbSMina Almasry 54528cc5fcbbSMina Almasry page = alloc_huge_page(dst_vma, dst_addr, 0); 54538cc5fcbbSMina Almasry if (IS_ERR(page)) { 54548cc5fcbbSMina Almasry ret = -ENOMEM; 54558cc5fcbbSMina Almasry *pagep = NULL; 54568cc5fcbbSMina Almasry goto out; 54578cc5fcbbSMina Almasry } 54588cc5fcbbSMina Almasry copy_huge_page(page, *pagep); 54598cc5fcbbSMina Almasry put_page(*pagep); 54608fb5debcSMike Kravetz *pagep = NULL; 54618fb5debcSMike Kravetz } 54628fb5debcSMike Kravetz 54638fb5debcSMike Kravetz /* 54648fb5debcSMike Kravetz * The memory barrier inside __SetPageUptodate makes sure that 54658fb5debcSMike Kravetz * preceding stores to the page contents become visible before 54668fb5debcSMike Kravetz * the set_pte_at() write. 54678fb5debcSMike Kravetz */ 54688fb5debcSMike Kravetz __SetPageUptodate(page); 54698fb5debcSMike Kravetz 5470f6191471SAxel Rasmussen /* Add shared, newly allocated pages to the page cache. */ 5471f6191471SAxel Rasmussen if (vm_shared && !is_continue) { 54721e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 54731e392147SAndrea Arcangeli ret = -EFAULT; 54741e392147SAndrea Arcangeli if (idx >= size) 54751e392147SAndrea Arcangeli goto out_release_nounlock; 54761c9e8defSMike Kravetz 54771e392147SAndrea Arcangeli /* 54781e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 54791e392147SAndrea Arcangeli * huge_add_to_page_cache() below happens through the 54801e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 54811e392147SAndrea Arcangeli * the caller. 54821e392147SAndrea Arcangeli */ 54831c9e8defSMike Kravetz ret = huge_add_to_page_cache(page, mapping, idx); 54841c9e8defSMike Kravetz if (ret) 54851c9e8defSMike Kravetz goto out_release_nounlock; 5486c7b1850dSMike Kravetz new_pagecache_page = true; 54871c9e8defSMike Kravetz } 54881c9e8defSMike Kravetz 54898fb5debcSMike Kravetz ptl = huge_pte_lockptr(h, dst_mm, dst_pte); 54908fb5debcSMike Kravetz spin_lock(ptl); 54918fb5debcSMike Kravetz 54921e392147SAndrea Arcangeli /* 54931e392147SAndrea Arcangeli * Recheck the i_size after holding PT lock to make sure not 54941e392147SAndrea Arcangeli * to leave any page mapped (as page_mapped()) beyond the end 54951e392147SAndrea Arcangeli * of the i_size (remove_inode_hugepages() is strict about 54961e392147SAndrea Arcangeli * enforcing that). If we bail out here, we'll also leave a 54971e392147SAndrea Arcangeli * page in the radix tree in the vm_shared case beyond the end 54981e392147SAndrea Arcangeli * of the i_size, but remove_inode_hugepages() will take care 54991e392147SAndrea Arcangeli * of it as soon as we drop the hugetlb_fault_mutex_table. 55001e392147SAndrea Arcangeli */ 55011e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 55021e392147SAndrea Arcangeli ret = -EFAULT; 55031e392147SAndrea Arcangeli if (idx >= size) 55041e392147SAndrea Arcangeli goto out_release_unlock; 55051e392147SAndrea Arcangeli 55068fb5debcSMike Kravetz ret = -EEXIST; 55078fb5debcSMike Kravetz if (!huge_pte_none(huge_ptep_get(dst_pte))) 55088fb5debcSMike Kravetz goto out_release_unlock; 55098fb5debcSMike Kravetz 55101c9e8defSMike Kravetz if (vm_shared) { 55111c9e8defSMike Kravetz page_dup_rmap(page, true); 55121c9e8defSMike Kravetz } else { 5513d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 55148fb5debcSMike Kravetz hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); 55151c9e8defSMike Kravetz } 55168fb5debcSMike Kravetz 5517f6191471SAxel Rasmussen /* For CONTINUE on a non-shared VMA, don't set VM_WRITE for CoW. */ 5518f6191471SAxel Rasmussen if (is_continue && !vm_shared) 5519f6191471SAxel Rasmussen writable = 0; 5520f6191471SAxel Rasmussen else 5521f6191471SAxel Rasmussen writable = dst_vma->vm_flags & VM_WRITE; 5522f6191471SAxel Rasmussen 5523f6191471SAxel Rasmussen _dst_pte = make_huge_pte(dst_vma, page, writable); 5524f6191471SAxel Rasmussen if (writable) 55258fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 55268fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 55278fb5debcSMike Kravetz 55288fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 55298fb5debcSMike Kravetz 55308fb5debcSMike Kravetz (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, 55318fb5debcSMike Kravetz dst_vma->vm_flags & VM_WRITE); 55328fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 55338fb5debcSMike Kravetz 55348fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 55358fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 55368fb5debcSMike Kravetz 55378fb5debcSMike Kravetz spin_unlock(ptl); 5538f6191471SAxel Rasmussen if (!is_continue) 55398f251a3dSMike Kravetz SetHPageMigratable(page); 5540f6191471SAxel Rasmussen if (vm_shared || is_continue) 55411c9e8defSMike Kravetz unlock_page(page); 55428fb5debcSMike Kravetz ret = 0; 55438fb5debcSMike Kravetz out: 55448fb5debcSMike Kravetz return ret; 55458fb5debcSMike Kravetz out_release_unlock: 55468fb5debcSMike Kravetz spin_unlock(ptl); 5547f6191471SAxel Rasmussen if (vm_shared || is_continue) 55481c9e8defSMike Kravetz unlock_page(page); 55495af10dfdSAndrea Arcangeli out_release_nounlock: 5550c7b1850dSMike Kravetz if (!new_pagecache_page) 5551846be085SMike Kravetz restore_reserve_on_error(h, dst_vma, dst_addr, page); 55528fb5debcSMike Kravetz put_page(page); 55538fb5debcSMike Kravetz goto out; 55548fb5debcSMike Kravetz } 5555714c1891SAxel Rasmussen #endif /* CONFIG_USERFAULTFD */ 55568fb5debcSMike Kravetz 555782e5d378SJoao Martins static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma, 555882e5d378SJoao Martins int refs, struct page **pages, 555982e5d378SJoao Martins struct vm_area_struct **vmas) 556082e5d378SJoao Martins { 556182e5d378SJoao Martins int nr; 556282e5d378SJoao Martins 556382e5d378SJoao Martins for (nr = 0; nr < refs; nr++) { 556482e5d378SJoao Martins if (likely(pages)) 556582e5d378SJoao Martins pages[nr] = mem_map_offset(page, nr); 556682e5d378SJoao Martins if (vmas) 556782e5d378SJoao Martins vmas[nr] = vma; 556882e5d378SJoao Martins } 556982e5d378SJoao Martins } 557082e5d378SJoao Martins 557128a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 557263551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 557328a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 55744f6da934SPeter Xu long i, unsigned int flags, int *locked) 557563551ae0SDavid Gibson { 5576d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 5577d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 557828a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 5579a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 55800fa5bc40SJoao Martins int err = -EFAULT, refs; 558163551ae0SDavid Gibson 558263551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 558363551ae0SDavid Gibson pte_t *pte; 5584cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 55852a15efc9SHugh Dickins int absent; 558663551ae0SDavid Gibson struct page *page; 558763551ae0SDavid Gibson 55884c887265SAdam Litke /* 558902057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 559002057967SDavid Rientjes * potentially allocating memory. 559102057967SDavid Rientjes */ 5592fa45f116SDavidlohr Bueso if (fatal_signal_pending(current)) { 559302057967SDavid Rientjes remainder = 0; 559402057967SDavid Rientjes break; 559502057967SDavid Rientjes } 559602057967SDavid Rientjes 559702057967SDavid Rientjes /* 55984c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 55992a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 56004c887265SAdam Litke * first, for the page indexing below to work. 5601cb900f41SKirill A. Shutemov * 5602cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 56034c887265SAdam Litke */ 56047868a208SPunit Agrawal pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), 56057868a208SPunit Agrawal huge_page_size(h)); 5606cb900f41SKirill A. Shutemov if (pte) 5607cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 56082a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 560963551ae0SDavid Gibson 56102a15efc9SHugh Dickins /* 56112a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 56123ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 56133ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 56143ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 56153ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 56162a15efc9SHugh Dickins */ 56173ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 56183ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 5619cb900f41SKirill A. Shutemov if (pte) 5620cb900f41SKirill A. Shutemov spin_unlock(ptl); 56212a15efc9SHugh Dickins remainder = 0; 56222a15efc9SHugh Dickins break; 56232a15efc9SHugh Dickins } 56242a15efc9SHugh Dickins 56259cc3a5bdSNaoya Horiguchi /* 56269cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 56279cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 56289cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 56299cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 56309cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 56319cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 56329cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 56339cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 56349cc3a5bdSNaoya Horiguchi */ 56359cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 5636106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 5637106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 56382b740303SSouptick Joarder vm_fault_t ret; 563987ffc118SAndrea Arcangeli unsigned int fault_flags = 0; 56404c887265SAdam Litke 5641cb900f41SKirill A. Shutemov if (pte) 5642cb900f41SKirill A. Shutemov spin_unlock(ptl); 564387ffc118SAndrea Arcangeli if (flags & FOLL_WRITE) 564487ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_WRITE; 56454f6da934SPeter Xu if (locked) 564671335f37SPeter Xu fault_flags |= FAULT_FLAG_ALLOW_RETRY | 564771335f37SPeter Xu FAULT_FLAG_KILLABLE; 564887ffc118SAndrea Arcangeli if (flags & FOLL_NOWAIT) 564987ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY | 565087ffc118SAndrea Arcangeli FAULT_FLAG_RETRY_NOWAIT; 565187ffc118SAndrea Arcangeli if (flags & FOLL_TRIED) { 56524426e945SPeter Xu /* 56534426e945SPeter Xu * Note: FAULT_FLAG_ALLOW_RETRY and 56544426e945SPeter Xu * FAULT_FLAG_TRIED can co-exist 56554426e945SPeter Xu */ 565687ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_TRIED; 565787ffc118SAndrea Arcangeli } 565887ffc118SAndrea Arcangeli ret = hugetlb_fault(mm, vma, vaddr, fault_flags); 565987ffc118SAndrea Arcangeli if (ret & VM_FAULT_ERROR) { 56602be7cfedSDaniel Jordan err = vm_fault_to_errno(ret, flags); 56611c59827dSHugh Dickins remainder = 0; 56621c59827dSHugh Dickins break; 56631c59827dSHugh Dickins } 566487ffc118SAndrea Arcangeli if (ret & VM_FAULT_RETRY) { 56654f6da934SPeter Xu if (locked && 56661ac25013SAndrea Arcangeli !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) 56674f6da934SPeter Xu *locked = 0; 566887ffc118SAndrea Arcangeli *nr_pages = 0; 566987ffc118SAndrea Arcangeli /* 567087ffc118SAndrea Arcangeli * VM_FAULT_RETRY must not return an 567187ffc118SAndrea Arcangeli * error, it will return zero 567287ffc118SAndrea Arcangeli * instead. 567387ffc118SAndrea Arcangeli * 567487ffc118SAndrea Arcangeli * No need to update "position" as the 567587ffc118SAndrea Arcangeli * caller will not check it after 567687ffc118SAndrea Arcangeli * *nr_pages is set to 0. 567787ffc118SAndrea Arcangeli */ 567887ffc118SAndrea Arcangeli return i; 567987ffc118SAndrea Arcangeli } 568087ffc118SAndrea Arcangeli continue; 568187ffc118SAndrea Arcangeli } 568263551ae0SDavid Gibson 5683a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 56847f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 56858fde12caSLinus Torvalds 56868fde12caSLinus Torvalds /* 5687acbfb087SZhigang Lu * If subpage information not requested, update counters 5688acbfb087SZhigang Lu * and skip the same_page loop below. 5689acbfb087SZhigang Lu */ 5690acbfb087SZhigang Lu if (!pages && !vmas && !pfn_offset && 5691acbfb087SZhigang Lu (vaddr + huge_page_size(h) < vma->vm_end) && 5692acbfb087SZhigang Lu (remainder >= pages_per_huge_page(h))) { 5693acbfb087SZhigang Lu vaddr += huge_page_size(h); 5694acbfb087SZhigang Lu remainder -= pages_per_huge_page(h); 5695acbfb087SZhigang Lu i += pages_per_huge_page(h); 5696acbfb087SZhigang Lu spin_unlock(ptl); 5697acbfb087SZhigang Lu continue; 5698acbfb087SZhigang Lu } 5699acbfb087SZhigang Lu 5700d08af0a5SJoao Martins /* vaddr may not be aligned to PAGE_SIZE */ 5701d08af0a5SJoao Martins refs = min3(pages_per_huge_page(h) - pfn_offset, remainder, 5702d08af0a5SJoao Martins (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT); 57030fa5bc40SJoao Martins 570482e5d378SJoao Martins if (pages || vmas) 570582e5d378SJoao Martins record_subpages_vmas(mem_map_offset(page, pfn_offset), 570682e5d378SJoao Martins vma, refs, 570782e5d378SJoao Martins likely(pages) ? pages + i : NULL, 570882e5d378SJoao Martins vmas ? vmas + i : NULL); 570963551ae0SDavid Gibson 571082e5d378SJoao Martins if (pages) { 57110fa5bc40SJoao Martins /* 57120fa5bc40SJoao Martins * try_grab_compound_head() should always succeed here, 57130fa5bc40SJoao Martins * because: a) we hold the ptl lock, and b) we've just 57140fa5bc40SJoao Martins * checked that the huge page is present in the page 57150fa5bc40SJoao Martins * tables. If the huge page is present, then the tail 57160fa5bc40SJoao Martins * pages must also be present. The ptl prevents the 57170fa5bc40SJoao Martins * head page and tail pages from being rearranged in 57180fa5bc40SJoao Martins * any way. So this page must be available at this 57190fa5bc40SJoao Martins * point, unless the page refcount overflowed: 57200fa5bc40SJoao Martins */ 572182e5d378SJoao Martins if (WARN_ON_ONCE(!try_grab_compound_head(pages[i], 57220fa5bc40SJoao Martins refs, 57230fa5bc40SJoao Martins flags))) { 57240fa5bc40SJoao Martins spin_unlock(ptl); 57250fa5bc40SJoao Martins remainder = 0; 57260fa5bc40SJoao Martins err = -ENOMEM; 57270fa5bc40SJoao Martins break; 57280fa5bc40SJoao Martins } 5729d5d4b0aaSChen, Kenneth W } 573082e5d378SJoao Martins 573182e5d378SJoao Martins vaddr += (refs << PAGE_SHIFT); 573282e5d378SJoao Martins remainder -= refs; 573382e5d378SJoao Martins i += refs; 573482e5d378SJoao Martins 5735cb900f41SKirill A. Shutemov spin_unlock(ptl); 573663551ae0SDavid Gibson } 573728a35716SMichel Lespinasse *nr_pages = remainder; 573887ffc118SAndrea Arcangeli /* 573987ffc118SAndrea Arcangeli * setting position is actually required only if remainder is 574087ffc118SAndrea Arcangeli * not zero but it's faster not to add a "if (remainder)" 574187ffc118SAndrea Arcangeli * branch. 574287ffc118SAndrea Arcangeli */ 574363551ae0SDavid Gibson *position = vaddr; 574463551ae0SDavid Gibson 57452be7cfedSDaniel Jordan return i ? i : err; 574663551ae0SDavid Gibson } 57478f860591SZhang, Yanmin 57487da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 57498f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 57508f860591SZhang, Yanmin { 57518f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 57528f860591SZhang, Yanmin unsigned long start = address; 57538f860591SZhang, Yanmin pte_t *ptep; 57548f860591SZhang, Yanmin pte_t pte; 5755a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 57567da4d641SPeter Zijlstra unsigned long pages = 0; 5757dff11abeSMike Kravetz bool shared_pmd = false; 5758ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 5759dff11abeSMike Kravetz 5760dff11abeSMike Kravetz /* 5761dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 5762ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 5763dff11abeSMike Kravetz * range if PMD sharing is possible. 5764dff11abeSMike Kravetz */ 57657269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 57667269f999SJérôme Glisse 0, vma, mm, start, end); 5767ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 57688f860591SZhang, Yanmin 57698f860591SZhang, Yanmin BUG_ON(address >= end); 5770ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 57718f860591SZhang, Yanmin 5772ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 577383cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 5774a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 5775cb900f41SKirill A. Shutemov spinlock_t *ptl; 57767868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, huge_page_size(h)); 57778f860591SZhang, Yanmin if (!ptep) 57788f860591SZhang, Yanmin continue; 5779cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 578034ae204fSMike Kravetz if (huge_pmd_unshare(mm, vma, &address, ptep)) { 57817da4d641SPeter Zijlstra pages++; 5782cb900f41SKirill A. Shutemov spin_unlock(ptl); 5783dff11abeSMike Kravetz shared_pmd = true; 578439dde65cSChen, Kenneth W continue; 57857da4d641SPeter Zijlstra } 5786a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 5787a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 5788a8bda28dSNaoya Horiguchi spin_unlock(ptl); 5789a8bda28dSNaoya Horiguchi continue; 5790a8bda28dSNaoya Horiguchi } 5791a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 5792a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 5793a8bda28dSNaoya Horiguchi 57944dd845b5SAlistair Popple if (is_writable_migration_entry(entry)) { 5795a8bda28dSNaoya Horiguchi pte_t newpte; 5796a8bda28dSNaoya Horiguchi 57974dd845b5SAlistair Popple entry = make_readable_migration_entry( 57984dd845b5SAlistair Popple swp_offset(entry)); 5799a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 5800e5251fd4SPunit Agrawal set_huge_swap_pte_at(mm, address, ptep, 5801e5251fd4SPunit Agrawal newpte, huge_page_size(h)); 5802a8bda28dSNaoya Horiguchi pages++; 5803a8bda28dSNaoya Horiguchi } 5804a8bda28dSNaoya Horiguchi spin_unlock(ptl); 5805a8bda28dSNaoya Horiguchi continue; 5806a8bda28dSNaoya Horiguchi } 5807a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 5808023bdd00SAneesh Kumar K.V pte_t old_pte; 580979c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 5810023bdd00SAneesh Kumar K.V 5811023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 5812023bdd00SAneesh Kumar K.V pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); 581379c1c594SChristophe Leroy pte = arch_make_huge_pte(pte, shift, vma->vm_flags); 5814023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 58157da4d641SPeter Zijlstra pages++; 58168f860591SZhang, Yanmin } 5817cb900f41SKirill A. Shutemov spin_unlock(ptl); 58188f860591SZhang, Yanmin } 5819d833352aSMel Gorman /* 5820c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 5821d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 5822c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 5823dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 5824dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 5825d833352aSMel Gorman */ 5826dff11abeSMike Kravetz if (shared_pmd) 5827ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 5828dff11abeSMike Kravetz else 58295491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 58300f10851eSJérôme Glisse /* 58310f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 58320f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 58330f10851eSJérôme Glisse * 5834ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 58350f10851eSJérôme Glisse */ 583683cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 5837ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 58387da4d641SPeter Zijlstra 58397da4d641SPeter Zijlstra return pages << h->order; 58408f860591SZhang, Yanmin } 58418f860591SZhang, Yanmin 584233b8f84aSMike Kravetz /* Return true if reservation was successful, false otherwise. */ 584333b8f84aSMike Kravetz bool hugetlb_reserve_pages(struct inode *inode, 5844a1e78772SMel Gorman long from, long to, 58455a6fe125SMel Gorman struct vm_area_struct *vma, 5846ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 5847e4e574b7SAdam Litke { 584833b8f84aSMike Kravetz long chg, add = -1; 5849a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 585090481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 58519119a41eSJoonsoo Kim struct resv_map *resv_map; 5852075a61d0SMina Almasry struct hugetlb_cgroup *h_cg = NULL; 58530db9d74eSMina Almasry long gbl_reserve, regions_needed = 0; 5854e4e574b7SAdam Litke 585563489f8eSMike Kravetz /* This should never happen */ 585663489f8eSMike Kravetz if (from > to) { 585763489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 585833b8f84aSMike Kravetz return false; 585963489f8eSMike Kravetz } 586063489f8eSMike Kravetz 5861a1e78772SMel Gorman /* 586217c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 586317c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 586490481622SDavid Gibson * without using reserves 586517c9d12eSMel Gorman */ 5866ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 586733b8f84aSMike Kravetz return true; 586817c9d12eSMel Gorman 586917c9d12eSMel Gorman /* 5870a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 5871a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 5872a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 5873a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 5874a1e78772SMel Gorman */ 58759119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 5876f27a5136SMike Kravetz /* 5877f27a5136SMike Kravetz * resv_map can not be NULL as hugetlb_reserve_pages is only 5878f27a5136SMike Kravetz * called for inodes for which resv_maps were created (see 5879f27a5136SMike Kravetz * hugetlbfs_get_inode). 5880f27a5136SMike Kravetz */ 58814e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 58829119a41eSJoonsoo Kim 58830db9d74eSMina Almasry chg = region_chg(resv_map, from, to, ®ions_needed); 58849119a41eSJoonsoo Kim 58859119a41eSJoonsoo Kim } else { 5886e9fe92aeSMina Almasry /* Private mapping. */ 58879119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 58885a6fe125SMel Gorman if (!resv_map) 588933b8f84aSMike Kravetz return false; 58905a6fe125SMel Gorman 589117c9d12eSMel Gorman chg = to - from; 589217c9d12eSMel Gorman 58935a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 58945a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 58955a6fe125SMel Gorman } 58965a6fe125SMel Gorman 589733b8f84aSMike Kravetz if (chg < 0) 5898c50ac050SDave Hansen goto out_err; 589917c9d12eSMel Gorman 590033b8f84aSMike Kravetz if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), 590133b8f84aSMike Kravetz chg * pages_per_huge_page(h), &h_cg) < 0) 5902075a61d0SMina Almasry goto out_err; 5903075a61d0SMina Almasry 5904075a61d0SMina Almasry if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { 5905075a61d0SMina Almasry /* For private mappings, the hugetlb_cgroup uncharge info hangs 5906075a61d0SMina Almasry * of the resv_map. 5907075a61d0SMina Almasry */ 5908075a61d0SMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); 5909075a61d0SMina Almasry } 5910075a61d0SMina Almasry 59111c5ecae3SMike Kravetz /* 59121c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 59131c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 59141c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 59151c5ecae3SMike Kravetz */ 59161c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 591733b8f84aSMike Kravetz if (gbl_reserve < 0) 5918075a61d0SMina Almasry goto out_uncharge_cgroup; 591917c9d12eSMel Gorman 592017c9d12eSMel Gorman /* 592117c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 592290481622SDavid Gibson * Hand the pages back to the subpool if there are not 592317c9d12eSMel Gorman */ 592433b8f84aSMike Kravetz if (hugetlb_acct_memory(h, gbl_reserve) < 0) 5925075a61d0SMina Almasry goto out_put_pages; 592617c9d12eSMel Gorman 592717c9d12eSMel Gorman /* 592817c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 592917c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 593017c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 593117c9d12eSMel Gorman * the reservation was and the page cache tells how much of 593217c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 593317c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 593417c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 593517c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 593617c9d12eSMel Gorman * else has to be done for private mappings here 593717c9d12eSMel Gorman */ 593833039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 5939075a61d0SMina Almasry add = region_add(resv_map, from, to, regions_needed, h, h_cg); 594033039678SMike Kravetz 59410db9d74eSMina Almasry if (unlikely(add < 0)) { 59420db9d74eSMina Almasry hugetlb_acct_memory(h, -gbl_reserve); 5943075a61d0SMina Almasry goto out_put_pages; 59440db9d74eSMina Almasry } else if (unlikely(chg > add)) { 594533039678SMike Kravetz /* 594633039678SMike Kravetz * pages in this range were added to the reserve 594733039678SMike Kravetz * map between region_chg and region_add. This 594833039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 594933039678SMike Kravetz * the subpool and reserve counts modified above 595033039678SMike Kravetz * based on the difference. 595133039678SMike Kravetz */ 595233039678SMike Kravetz long rsv_adjust; 595333039678SMike Kravetz 5954d85aecf2SMiaohe Lin /* 5955d85aecf2SMiaohe Lin * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the 5956d85aecf2SMiaohe Lin * reference to h_cg->css. See comment below for detail. 5957d85aecf2SMiaohe Lin */ 5958075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd( 5959075a61d0SMina Almasry hstate_index(h), 5960075a61d0SMina Almasry (chg - add) * pages_per_huge_page(h), h_cg); 5961075a61d0SMina Almasry 596233039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 596333039678SMike Kravetz chg - add); 596433039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 5965d85aecf2SMiaohe Lin } else if (h_cg) { 5966d85aecf2SMiaohe Lin /* 5967d85aecf2SMiaohe Lin * The file_regions will hold their own reference to 5968d85aecf2SMiaohe Lin * h_cg->css. So we should release the reference held 5969d85aecf2SMiaohe Lin * via hugetlb_cgroup_charge_cgroup_rsvd() when we are 5970d85aecf2SMiaohe Lin * done. 5971d85aecf2SMiaohe Lin */ 5972d85aecf2SMiaohe Lin hugetlb_cgroup_put_rsvd_cgroup(h_cg); 597333039678SMike Kravetz } 597433039678SMike Kravetz } 597533b8f84aSMike Kravetz return true; 597633b8f84aSMike Kravetz 5977075a61d0SMina Almasry out_put_pages: 5978075a61d0SMina Almasry /* put back original number of pages, chg */ 5979075a61d0SMina Almasry (void)hugepage_subpool_put_pages(spool, chg); 5980075a61d0SMina Almasry out_uncharge_cgroup: 5981075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), 5982075a61d0SMina Almasry chg * pages_per_huge_page(h), h_cg); 5983c50ac050SDave Hansen out_err: 59845e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 59850db9d74eSMina Almasry /* Only call region_abort if the region_chg succeeded but the 59860db9d74eSMina Almasry * region_add failed or didn't run. 59870db9d74eSMina Almasry */ 59880db9d74eSMina Almasry if (chg >= 0 && add < 0) 59890db9d74eSMina Almasry region_abort(resv_map, from, to, regions_needed); 5990f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 5991f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 599233b8f84aSMike Kravetz return false; 5993a43a8c39SChen, Kenneth W } 5994a43a8c39SChen, Kenneth W 5995b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 5996b5cec28dSMike Kravetz long freed) 5997a43a8c39SChen, Kenneth W { 5998a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 59994e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 60009119a41eSJoonsoo Kim long chg = 0; 600190481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 60021c5ecae3SMike Kravetz long gbl_reserve; 600345c682a6SKen Chen 6004f27a5136SMike Kravetz /* 6005f27a5136SMike Kravetz * Since this routine can be called in the evict inode path for all 6006f27a5136SMike Kravetz * hugetlbfs inodes, resv_map could be NULL. 6007f27a5136SMike Kravetz */ 6008b5cec28dSMike Kravetz if (resv_map) { 6009b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 6010b5cec28dSMike Kravetz /* 6011b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 6012b5cec28dSMike Kravetz * must be split and another region descriptor can not be 6013b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 6014b5cec28dSMike Kravetz */ 6015b5cec28dSMike Kravetz if (chg < 0) 6016b5cec28dSMike Kravetz return chg; 6017b5cec28dSMike Kravetz } 6018b5cec28dSMike Kravetz 601945c682a6SKen Chen spin_lock(&inode->i_lock); 6020e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 602145c682a6SKen Chen spin_unlock(&inode->i_lock); 602245c682a6SKen Chen 60231c5ecae3SMike Kravetz /* 60241c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 60251c5ecae3SMike Kravetz * reservations to be released may be adjusted. 6026dddf31a4SMiaohe Lin * 6027dddf31a4SMiaohe Lin * Note that !resv_map implies freed == 0. So (chg - freed) 6028dddf31a4SMiaohe Lin * won't go negative. 60291c5ecae3SMike Kravetz */ 60301c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 60311c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 6032b5cec28dSMike Kravetz 6033b5cec28dSMike Kravetz return 0; 6034a43a8c39SChen, Kenneth W } 603593f70f90SNaoya Horiguchi 60363212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 60373212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 60383212b535SSteve Capper struct vm_area_struct *vma, 60393212b535SSteve Capper unsigned long addr, pgoff_t idx) 60403212b535SSteve Capper { 60413212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 60423212b535SSteve Capper svma->vm_start; 60433212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 60443212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 60453212b535SSteve Capper 60463212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 6047de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 6048de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 60493212b535SSteve Capper 60503212b535SSteve Capper /* 60513212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 60523212b535SSteve Capper * page table page. 60533212b535SSteve Capper */ 60543212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 60553212b535SSteve Capper vm_flags != svm_flags || 605607e51edfSMiaohe Lin !range_in_vma(svma, sbase, s_end)) 60573212b535SSteve Capper return 0; 60583212b535SSteve Capper 60593212b535SSteve Capper return saddr; 60603212b535SSteve Capper } 60613212b535SSteve Capper 606231aafb45SNicholas Krause static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) 60633212b535SSteve Capper { 60643212b535SSteve Capper unsigned long base = addr & PUD_MASK; 60653212b535SSteve Capper unsigned long end = base + PUD_SIZE; 60663212b535SSteve Capper 60673212b535SSteve Capper /* 60683212b535SSteve Capper * check on proper vm_flags and page table alignment 60693212b535SSteve Capper */ 6070017b1660SMike Kravetz if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) 607131aafb45SNicholas Krause return true; 607231aafb45SNicholas Krause return false; 60733212b535SSteve Capper } 60743212b535SSteve Capper 6075c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 6076c1991e07SPeter Xu { 6077c1991e07SPeter Xu #ifdef CONFIG_USERFAULTFD 6078c1991e07SPeter Xu if (uffd_disable_huge_pmd_share(vma)) 6079c1991e07SPeter Xu return false; 6080c1991e07SPeter Xu #endif 6081c1991e07SPeter Xu return vma_shareable(vma, addr); 6082c1991e07SPeter Xu } 6083c1991e07SPeter Xu 60843212b535SSteve Capper /* 6085017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 6086017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 6087017b1660SMike Kravetz * shared pmd mappings. 6088017b1660SMike Kravetz */ 6089017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 6090017b1660SMike Kravetz unsigned long *start, unsigned long *end) 6091017b1660SMike Kravetz { 6092a1ba9da8SLi Xinhai unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), 6093a1ba9da8SLi Xinhai v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 6094017b1660SMike Kravetz 6095a1ba9da8SLi Xinhai /* 6096f0953a1bSIngo Molnar * vma needs to span at least one aligned PUD size, and the range 6097f0953a1bSIngo Molnar * must be at least partially within in. 6098a1ba9da8SLi Xinhai */ 6099a1ba9da8SLi Xinhai if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || 6100a1ba9da8SLi Xinhai (*end <= v_start) || (*start >= v_end)) 6101017b1660SMike Kravetz return; 6102017b1660SMike Kravetz 610375802ca6SPeter Xu /* Extend the range to be PUD aligned for a worst case scenario */ 6104a1ba9da8SLi Xinhai if (*start > v_start) 6105a1ba9da8SLi Xinhai *start = ALIGN_DOWN(*start, PUD_SIZE); 6106017b1660SMike Kravetz 6107a1ba9da8SLi Xinhai if (*end < v_end) 6108a1ba9da8SLi Xinhai *end = ALIGN(*end, PUD_SIZE); 6109017b1660SMike Kravetz } 6110017b1660SMike Kravetz 6111017b1660SMike Kravetz /* 61123212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 61133212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 61143212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 6115c0d0381aSMike Kravetz * code much cleaner. 6116c0d0381aSMike Kravetz * 61170bf7b64eSMike Kravetz * This routine must be called with i_mmap_rwsem held in at least read mode if 61180bf7b64eSMike Kravetz * sharing is possible. For hugetlbfs, this prevents removal of any page 61190bf7b64eSMike Kravetz * table entries associated with the address space. This is important as we 61200bf7b64eSMike Kravetz * are setting up sharing based on existing page table entries (mappings). 61210bf7b64eSMike Kravetz * 61220bf7b64eSMike Kravetz * NOTE: This routine is only called from huge_pte_alloc. Some callers of 61230bf7b64eSMike Kravetz * huge_pte_alloc know that sharing is not possible and do not take 61240bf7b64eSMike Kravetz * i_mmap_rwsem as a performance optimization. This is handled by the 61250bf7b64eSMike Kravetz * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is 61260bf7b64eSMike Kravetz * only required for subsequent processing. 61273212b535SSteve Capper */ 6128aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 6129aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 61303212b535SSteve Capper { 61313212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 61323212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 61333212b535SSteve Capper vma->vm_pgoff; 61343212b535SSteve Capper struct vm_area_struct *svma; 61353212b535SSteve Capper unsigned long saddr; 61363212b535SSteve Capper pte_t *spte = NULL; 61373212b535SSteve Capper pte_t *pte; 6138cb900f41SKirill A. Shutemov spinlock_t *ptl; 61393212b535SSteve Capper 61400bf7b64eSMike Kravetz i_mmap_assert_locked(mapping); 61413212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 61423212b535SSteve Capper if (svma == vma) 61433212b535SSteve Capper continue; 61443212b535SSteve Capper 61453212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 61463212b535SSteve Capper if (saddr) { 61477868a208SPunit Agrawal spte = huge_pte_offset(svma->vm_mm, saddr, 61487868a208SPunit Agrawal vma_mmu_pagesize(svma)); 61493212b535SSteve Capper if (spte) { 61503212b535SSteve Capper get_page(virt_to_page(spte)); 61513212b535SSteve Capper break; 61523212b535SSteve Capper } 61533212b535SSteve Capper } 61543212b535SSteve Capper } 61553212b535SSteve Capper 61563212b535SSteve Capper if (!spte) 61573212b535SSteve Capper goto out; 61583212b535SSteve Capper 61598bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 6160dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 61613212b535SSteve Capper pud_populate(mm, pud, 61623212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 6163c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 6164dc6c9a35SKirill A. Shutemov } else { 61653212b535SSteve Capper put_page(virt_to_page(spte)); 6166dc6c9a35SKirill A. Shutemov } 6167cb900f41SKirill A. Shutemov spin_unlock(ptl); 61683212b535SSteve Capper out: 61693212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 61703212b535SSteve Capper return pte; 61713212b535SSteve Capper } 61723212b535SSteve Capper 61733212b535SSteve Capper /* 61743212b535SSteve Capper * unmap huge page backed by shared pte. 61753212b535SSteve Capper * 61763212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 61773212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 61783212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 61793212b535SSteve Capper * 6180c0d0381aSMike Kravetz * Called with page table lock held and i_mmap_rwsem held in write mode. 61813212b535SSteve Capper * 61823212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 61833212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 61843212b535SSteve Capper */ 618534ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 618634ae204fSMike Kravetz unsigned long *addr, pte_t *ptep) 61873212b535SSteve Capper { 61883212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 6189c2febafcSKirill A. Shutemov p4d_t *p4d = p4d_offset(pgd, *addr); 6190c2febafcSKirill A. Shutemov pud_t *pud = pud_offset(p4d, *addr); 61913212b535SSteve Capper 619234ae204fSMike Kravetz i_mmap_assert_write_locked(vma->vm_file->f_mapping); 61933212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 61943212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 61953212b535SSteve Capper return 0; 61963212b535SSteve Capper 61973212b535SSteve Capper pud_clear(pud); 61983212b535SSteve Capper put_page(virt_to_page(ptep)); 6199dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 62003212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 62013212b535SSteve Capper return 1; 62023212b535SSteve Capper } 6203c1991e07SPeter Xu 62049e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 6205aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 6206aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 62079e5fc74cSSteve Capper { 62089e5fc74cSSteve Capper return NULL; 62099e5fc74cSSteve Capper } 6210e81f2d22SZhang Zhen 621134ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 621234ae204fSMike Kravetz unsigned long *addr, pte_t *ptep) 6213e81f2d22SZhang Zhen { 6214e81f2d22SZhang Zhen return 0; 6215e81f2d22SZhang Zhen } 6216017b1660SMike Kravetz 6217017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 6218017b1660SMike Kravetz unsigned long *start, unsigned long *end) 6219017b1660SMike Kravetz { 6220017b1660SMike Kravetz } 6221c1991e07SPeter Xu 6222c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 6223c1991e07SPeter Xu { 6224c1991e07SPeter Xu return false; 6225c1991e07SPeter Xu } 62263212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 62273212b535SSteve Capper 62289e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 6229aec44e0fSPeter Xu pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, 62309e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 62319e5fc74cSSteve Capper { 62329e5fc74cSSteve Capper pgd_t *pgd; 6233c2febafcSKirill A. Shutemov p4d_t *p4d; 62349e5fc74cSSteve Capper pud_t *pud; 62359e5fc74cSSteve Capper pte_t *pte = NULL; 62369e5fc74cSSteve Capper 62379e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 6238f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 6239f4f0a3d8SKirill A. Shutemov if (!p4d) 6240f4f0a3d8SKirill A. Shutemov return NULL; 6241c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 62429e5fc74cSSteve Capper if (pud) { 62439e5fc74cSSteve Capper if (sz == PUD_SIZE) { 62449e5fc74cSSteve Capper pte = (pte_t *)pud; 62459e5fc74cSSteve Capper } else { 62469e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 6247c1991e07SPeter Xu if (want_pmd_share(vma, addr) && pud_none(*pud)) 6248aec44e0fSPeter Xu pte = huge_pmd_share(mm, vma, addr, pud); 62499e5fc74cSSteve Capper else 62509e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 62519e5fc74cSSteve Capper } 62529e5fc74cSSteve Capper } 62534e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 62549e5fc74cSSteve Capper 62559e5fc74cSSteve Capper return pte; 62569e5fc74cSSteve Capper } 62579e5fc74cSSteve Capper 62589b19df29SPunit Agrawal /* 62599b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 62609b19df29SPunit Agrawal * entry at address @addr 62619b19df29SPunit Agrawal * 62628ac0b81aSLi Xinhai * Return: Pointer to page table entry (PUD or PMD) for 62638ac0b81aSLi Xinhai * address @addr, or NULL if a !p*d_present() entry is encountered and the 62649b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 62659b19df29SPunit Agrawal * table. 62669b19df29SPunit Agrawal */ 62677868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 62687868a208SPunit Agrawal unsigned long addr, unsigned long sz) 62699e5fc74cSSteve Capper { 62709e5fc74cSSteve Capper pgd_t *pgd; 6271c2febafcSKirill A. Shutemov p4d_t *p4d; 62728ac0b81aSLi Xinhai pud_t *pud; 62738ac0b81aSLi Xinhai pmd_t *pmd; 62749e5fc74cSSteve Capper 62759e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 6276c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 6277c2febafcSKirill A. Shutemov return NULL; 6278c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 6279c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 6280c2febafcSKirill A. Shutemov return NULL; 62819b19df29SPunit Agrawal 6282c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 62838ac0b81aSLi Xinhai if (sz == PUD_SIZE) 62848ac0b81aSLi Xinhai /* must be pud huge, non-present or none */ 62859e5fc74cSSteve Capper return (pte_t *)pud; 62868ac0b81aSLi Xinhai if (!pud_present(*pud)) 62878ac0b81aSLi Xinhai return NULL; 62888ac0b81aSLi Xinhai /* must have a valid entry and size to go further */ 62899b19df29SPunit Agrawal 62909e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 62918ac0b81aSLi Xinhai /* must be pmd huge, non-present or none */ 62929e5fc74cSSteve Capper return (pte_t *)pmd; 62939e5fc74cSSteve Capper } 62949e5fc74cSSteve Capper 629561f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 629661f77edaSNaoya Horiguchi 629761f77edaSNaoya Horiguchi /* 629861f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 629961f77edaSNaoya Horiguchi * behavior. 630061f77edaSNaoya Horiguchi */ 630161f77edaSNaoya Horiguchi struct page * __weak 630261f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 630361f77edaSNaoya Horiguchi int write) 630461f77edaSNaoya Horiguchi { 630561f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 630661f77edaSNaoya Horiguchi } 630761f77edaSNaoya Horiguchi 630861f77edaSNaoya Horiguchi struct page * __weak 63094dc71451SAneesh Kumar K.V follow_huge_pd(struct vm_area_struct *vma, 63104dc71451SAneesh Kumar K.V unsigned long address, hugepd_t hpd, int flags, int pdshift) 63114dc71451SAneesh Kumar K.V { 63124dc71451SAneesh Kumar K.V WARN(1, "hugepd follow called with no support for hugepage directory format\n"); 63134dc71451SAneesh Kumar K.V return NULL; 63144dc71451SAneesh Kumar K.V } 63154dc71451SAneesh Kumar K.V 63164dc71451SAneesh Kumar K.V struct page * __weak 63179e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 6318e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 63199e5fc74cSSteve Capper { 6320e66f17ffSNaoya Horiguchi struct page *page = NULL; 6321e66f17ffSNaoya Horiguchi spinlock_t *ptl; 6322c9d398faSNaoya Horiguchi pte_t pte; 63233faa52c0SJohn Hubbard 63243faa52c0SJohn Hubbard /* FOLL_GET and FOLL_PIN are mutually exclusive. */ 63253faa52c0SJohn Hubbard if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == 63263faa52c0SJohn Hubbard (FOLL_PIN | FOLL_GET))) 63273faa52c0SJohn Hubbard return NULL; 63283faa52c0SJohn Hubbard 6329e66f17ffSNaoya Horiguchi retry: 6330e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 6331e66f17ffSNaoya Horiguchi spin_lock(ptl); 6332e66f17ffSNaoya Horiguchi /* 6333e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 6334e66f17ffSNaoya Horiguchi * unmapped from other threads. 6335e66f17ffSNaoya Horiguchi */ 6336e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 6337e66f17ffSNaoya Horiguchi goto out; 6338c9d398faSNaoya Horiguchi pte = huge_ptep_get((pte_t *)pmd); 6339c9d398faSNaoya Horiguchi if (pte_present(pte)) { 634097534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 63413faa52c0SJohn Hubbard /* 63423faa52c0SJohn Hubbard * try_grab_page() should always succeed here, because: a) we 63433faa52c0SJohn Hubbard * hold the pmd (ptl) lock, and b) we've just checked that the 63443faa52c0SJohn Hubbard * huge pmd (head) page is present in the page tables. The ptl 63453faa52c0SJohn Hubbard * prevents the head page and tail pages from being rearranged 63463faa52c0SJohn Hubbard * in any way. So this page must be available at this point, 63473faa52c0SJohn Hubbard * unless the page refcount overflowed: 63483faa52c0SJohn Hubbard */ 63493faa52c0SJohn Hubbard if (WARN_ON_ONCE(!try_grab_page(page, flags))) { 63503faa52c0SJohn Hubbard page = NULL; 63513faa52c0SJohn Hubbard goto out; 63523faa52c0SJohn Hubbard } 6353e66f17ffSNaoya Horiguchi } else { 6354c9d398faSNaoya Horiguchi if (is_hugetlb_entry_migration(pte)) { 6355e66f17ffSNaoya Horiguchi spin_unlock(ptl); 6356e66f17ffSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pmd, ptl); 6357e66f17ffSNaoya Horiguchi goto retry; 6358e66f17ffSNaoya Horiguchi } 6359e66f17ffSNaoya Horiguchi /* 6360e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 6361e66f17ffSNaoya Horiguchi * follow_page_mask(). 6362e66f17ffSNaoya Horiguchi */ 6363e66f17ffSNaoya Horiguchi } 6364e66f17ffSNaoya Horiguchi out: 6365e66f17ffSNaoya Horiguchi spin_unlock(ptl); 63669e5fc74cSSteve Capper return page; 63679e5fc74cSSteve Capper } 63689e5fc74cSSteve Capper 636961f77edaSNaoya Horiguchi struct page * __weak 63709e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 6371e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 63729e5fc74cSSteve Capper { 63733faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 6374e66f17ffSNaoya Horiguchi return NULL; 63759e5fc74cSSteve Capper 6376e66f17ffSNaoya Horiguchi return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 63779e5fc74cSSteve Capper } 63789e5fc74cSSteve Capper 6379faaa5b62SAnshuman Khandual struct page * __weak 6380faaa5b62SAnshuman Khandual follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) 6381faaa5b62SAnshuman Khandual { 63823faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 6383faaa5b62SAnshuman Khandual return NULL; 6384faaa5b62SAnshuman Khandual 6385faaa5b62SAnshuman Khandual return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); 6386faaa5b62SAnshuman Khandual } 6387faaa5b62SAnshuman Khandual 638831caf665SNaoya Horiguchi bool isolate_huge_page(struct page *page, struct list_head *list) 638931caf665SNaoya Horiguchi { 6390bcc54222SNaoya Horiguchi bool ret = true; 6391bcc54222SNaoya Horiguchi 6392db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 63938f251a3dSMike Kravetz if (!PageHeadHuge(page) || 63948f251a3dSMike Kravetz !HPageMigratable(page) || 63950eb2df2bSMuchun Song !get_page_unless_zero(page)) { 6396bcc54222SNaoya Horiguchi ret = false; 6397bcc54222SNaoya Horiguchi goto unlock; 6398bcc54222SNaoya Horiguchi } 63998f251a3dSMike Kravetz ClearHPageMigratable(page); 640031caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 6401bcc54222SNaoya Horiguchi unlock: 6402db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 6403bcc54222SNaoya Horiguchi return ret; 640431caf665SNaoya Horiguchi } 640531caf665SNaoya Horiguchi 640625182f05SNaoya Horiguchi int get_hwpoison_huge_page(struct page *page, bool *hugetlb) 640725182f05SNaoya Horiguchi { 640825182f05SNaoya Horiguchi int ret = 0; 640925182f05SNaoya Horiguchi 641025182f05SNaoya Horiguchi *hugetlb = false; 641125182f05SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 641225182f05SNaoya Horiguchi if (PageHeadHuge(page)) { 641325182f05SNaoya Horiguchi *hugetlb = true; 641425182f05SNaoya Horiguchi if (HPageFreed(page) || HPageMigratable(page)) 641525182f05SNaoya Horiguchi ret = get_page_unless_zero(page); 64160ed950d1SNaoya Horiguchi else 64170ed950d1SNaoya Horiguchi ret = -EBUSY; 641825182f05SNaoya Horiguchi } 641925182f05SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 642025182f05SNaoya Horiguchi return ret; 642125182f05SNaoya Horiguchi } 642225182f05SNaoya Horiguchi 642331caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 642431caf665SNaoya Horiguchi { 6425db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 64268f251a3dSMike Kravetz SetHPageMigratable(page); 642731caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 6428db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 642931caf665SNaoya Horiguchi put_page(page); 643031caf665SNaoya Horiguchi } 6431ab5ac90aSMichal Hocko 6432ab5ac90aSMichal Hocko void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) 6433ab5ac90aSMichal Hocko { 6434ab5ac90aSMichal Hocko struct hstate *h = page_hstate(oldpage); 6435ab5ac90aSMichal Hocko 6436ab5ac90aSMichal Hocko hugetlb_cgroup_migrate(oldpage, newpage); 6437ab5ac90aSMichal Hocko set_page_owner_migrate_reason(newpage, reason); 6438ab5ac90aSMichal Hocko 6439ab5ac90aSMichal Hocko /* 6440ab5ac90aSMichal Hocko * transfer temporary state of the new huge page. This is 6441ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 6442ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 6443ab5ac90aSMichal Hocko * the temporary status. 6444ab5ac90aSMichal Hocko * 6445ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 6446ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 6447ab5ac90aSMichal Hocko * the per-node's. 6448ab5ac90aSMichal Hocko */ 64499157c311SMike Kravetz if (HPageTemporary(newpage)) { 6450ab5ac90aSMichal Hocko int old_nid = page_to_nid(oldpage); 6451ab5ac90aSMichal Hocko int new_nid = page_to_nid(newpage); 6452ab5ac90aSMichal Hocko 64539157c311SMike Kravetz SetHPageTemporary(oldpage); 64549157c311SMike Kravetz ClearHPageTemporary(newpage); 6455ab5ac90aSMichal Hocko 64565af1ab1dSMiaohe Lin /* 64575af1ab1dSMiaohe Lin * There is no need to transfer the per-node surplus state 64585af1ab1dSMiaohe Lin * when we do not cross the node. 64595af1ab1dSMiaohe Lin */ 64605af1ab1dSMiaohe Lin if (new_nid == old_nid) 64615af1ab1dSMiaohe Lin return; 6462db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 6463ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 6464ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 6465ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 6466ab5ac90aSMichal Hocko } 6467db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 6468ab5ac90aSMichal Hocko } 6469ab5ac90aSMichal Hocko } 6470cf11e85fSRoman Gushchin 64716dfeaff9SPeter Xu /* 64726dfeaff9SPeter Xu * This function will unconditionally remove all the shared pmd pgtable entries 64736dfeaff9SPeter Xu * within the specific vma for a hugetlbfs memory range. 64746dfeaff9SPeter Xu */ 64756dfeaff9SPeter Xu void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) 64766dfeaff9SPeter Xu { 64776dfeaff9SPeter Xu struct hstate *h = hstate_vma(vma); 64786dfeaff9SPeter Xu unsigned long sz = huge_page_size(h); 64796dfeaff9SPeter Xu struct mm_struct *mm = vma->vm_mm; 64806dfeaff9SPeter Xu struct mmu_notifier_range range; 64816dfeaff9SPeter Xu unsigned long address, start, end; 64826dfeaff9SPeter Xu spinlock_t *ptl; 64836dfeaff9SPeter Xu pte_t *ptep; 64846dfeaff9SPeter Xu 64856dfeaff9SPeter Xu if (!(vma->vm_flags & VM_MAYSHARE)) 64866dfeaff9SPeter Xu return; 64876dfeaff9SPeter Xu 64886dfeaff9SPeter Xu start = ALIGN(vma->vm_start, PUD_SIZE); 64896dfeaff9SPeter Xu end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 64906dfeaff9SPeter Xu 64916dfeaff9SPeter Xu if (start >= end) 64926dfeaff9SPeter Xu return; 64936dfeaff9SPeter Xu 64946dfeaff9SPeter Xu /* 64956dfeaff9SPeter Xu * No need to call adjust_range_if_pmd_sharing_possible(), because 64966dfeaff9SPeter Xu * we have already done the PUD_SIZE alignment. 64976dfeaff9SPeter Xu */ 64986dfeaff9SPeter Xu mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, 64996dfeaff9SPeter Xu start, end); 65006dfeaff9SPeter Xu mmu_notifier_invalidate_range_start(&range); 65016dfeaff9SPeter Xu i_mmap_lock_write(vma->vm_file->f_mapping); 65026dfeaff9SPeter Xu for (address = start; address < end; address += PUD_SIZE) { 65036dfeaff9SPeter Xu unsigned long tmp = address; 65046dfeaff9SPeter Xu 65056dfeaff9SPeter Xu ptep = huge_pte_offset(mm, address, sz); 65066dfeaff9SPeter Xu if (!ptep) 65076dfeaff9SPeter Xu continue; 65086dfeaff9SPeter Xu ptl = huge_pte_lock(h, mm, ptep); 65096dfeaff9SPeter Xu /* We don't want 'address' to be changed */ 65106dfeaff9SPeter Xu huge_pmd_unshare(mm, vma, &tmp, ptep); 65116dfeaff9SPeter Xu spin_unlock(ptl); 65126dfeaff9SPeter Xu } 65136dfeaff9SPeter Xu flush_hugetlb_tlb_range(vma, start, end); 65146dfeaff9SPeter Xu i_mmap_unlock_write(vma->vm_file->f_mapping); 65156dfeaff9SPeter Xu /* 65166dfeaff9SPeter Xu * No need to call mmu_notifier_invalidate_range(), see 65176dfeaff9SPeter Xu * Documentation/vm/mmu_notifier.rst. 65186dfeaff9SPeter Xu */ 65196dfeaff9SPeter Xu mmu_notifier_invalidate_range_end(&range); 65206dfeaff9SPeter Xu } 65216dfeaff9SPeter Xu 6522cf11e85fSRoman Gushchin #ifdef CONFIG_CMA 6523cf11e85fSRoman Gushchin static bool cma_reserve_called __initdata; 6524cf11e85fSRoman Gushchin 6525cf11e85fSRoman Gushchin static int __init cmdline_parse_hugetlb_cma(char *p) 6526cf11e85fSRoman Gushchin { 6527cf11e85fSRoman Gushchin hugetlb_cma_size = memparse(p, &p); 6528cf11e85fSRoman Gushchin return 0; 6529cf11e85fSRoman Gushchin } 6530cf11e85fSRoman Gushchin 6531cf11e85fSRoman Gushchin early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); 6532cf11e85fSRoman Gushchin 6533cf11e85fSRoman Gushchin void __init hugetlb_cma_reserve(int order) 6534cf11e85fSRoman Gushchin { 6535cf11e85fSRoman Gushchin unsigned long size, reserved, per_node; 6536cf11e85fSRoman Gushchin int nid; 6537cf11e85fSRoman Gushchin 6538cf11e85fSRoman Gushchin cma_reserve_called = true; 6539cf11e85fSRoman Gushchin 6540cf11e85fSRoman Gushchin if (!hugetlb_cma_size) 6541cf11e85fSRoman Gushchin return; 6542cf11e85fSRoman Gushchin 6543cf11e85fSRoman Gushchin if (hugetlb_cma_size < (PAGE_SIZE << order)) { 6544cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", 6545cf11e85fSRoman Gushchin (PAGE_SIZE << order) / SZ_1M); 6546cf11e85fSRoman Gushchin return; 6547cf11e85fSRoman Gushchin } 6548cf11e85fSRoman Gushchin 6549cf11e85fSRoman Gushchin /* 6550cf11e85fSRoman Gushchin * If 3 GB area is requested on a machine with 4 numa nodes, 6551cf11e85fSRoman Gushchin * let's allocate 1 GB on first three nodes and ignore the last one. 6552cf11e85fSRoman Gushchin */ 6553cf11e85fSRoman Gushchin per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); 6554cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", 6555cf11e85fSRoman Gushchin hugetlb_cma_size / SZ_1M, per_node / SZ_1M); 6556cf11e85fSRoman Gushchin 6557cf11e85fSRoman Gushchin reserved = 0; 6558cf11e85fSRoman Gushchin for_each_node_state(nid, N_ONLINE) { 6559cf11e85fSRoman Gushchin int res; 65602281f797SBarry Song char name[CMA_MAX_NAME]; 6561cf11e85fSRoman Gushchin 6562cf11e85fSRoman Gushchin size = min(per_node, hugetlb_cma_size - reserved); 6563cf11e85fSRoman Gushchin size = round_up(size, PAGE_SIZE << order); 6564cf11e85fSRoman Gushchin 65652281f797SBarry Song snprintf(name, sizeof(name), "hugetlb%d", nid); 6566cf11e85fSRoman Gushchin res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order, 656729d0f41dSBarry Song 0, false, name, 6568cf11e85fSRoman Gushchin &hugetlb_cma[nid], nid); 6569cf11e85fSRoman Gushchin if (res) { 6570cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: reservation failed: err %d, node %d", 6571cf11e85fSRoman Gushchin res, nid); 6572cf11e85fSRoman Gushchin continue; 6573cf11e85fSRoman Gushchin } 6574cf11e85fSRoman Gushchin 6575cf11e85fSRoman Gushchin reserved += size; 6576cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", 6577cf11e85fSRoman Gushchin size / SZ_1M, nid); 6578cf11e85fSRoman Gushchin 6579cf11e85fSRoman Gushchin if (reserved >= hugetlb_cma_size) 6580cf11e85fSRoman Gushchin break; 6581cf11e85fSRoman Gushchin } 6582cf11e85fSRoman Gushchin } 6583cf11e85fSRoman Gushchin 6584cf11e85fSRoman Gushchin void __init hugetlb_cma_check(void) 6585cf11e85fSRoman Gushchin { 6586cf11e85fSRoman Gushchin if (!hugetlb_cma_size || cma_reserve_called) 6587cf11e85fSRoman Gushchin return; 6588cf11e85fSRoman Gushchin 6589cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); 6590cf11e85fSRoman Gushchin } 6591cf11e85fSRoman Gushchin 6592cf11e85fSRoman Gushchin #endif /* CONFIG_CMA */ 6593