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); 10750edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 1076a5516438SAndi Kleen h->free_huge_pages++; 1077a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 10786c037149SMike Kravetz SetHPageFreed(page); 10791da177e4SLinus Torvalds } 10801da177e4SLinus Torvalds 108194310cbcSAnshuman Khandual static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) 1082bf50bab2SNaoya Horiguchi { 1083bf50bab2SNaoya Horiguchi struct page *page; 10841a08ae36SPavel Tatashin bool pin = !!(current->flags & PF_MEMALLOC_PIN); 1085bf50bab2SNaoya Horiguchi 10869487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1087bbe88753SJoonsoo Kim list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { 10888e3560d9SPavel Tatashin if (pin && !is_pinnable_page(page)) 1089bbe88753SJoonsoo Kim continue; 1090bbe88753SJoonsoo Kim 10916664bfc8SWei Yang if (PageHWPoison(page)) 10926664bfc8SWei Yang continue; 1093bbe88753SJoonsoo Kim 10940edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 1095a9869b83SNaoya Horiguchi set_page_refcounted(page); 10966c037149SMike Kravetz ClearHPageFreed(page); 1097bf50bab2SNaoya Horiguchi h->free_huge_pages--; 1098bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 1099bf50bab2SNaoya Horiguchi return page; 1100bf50bab2SNaoya Horiguchi } 1101bf50bab2SNaoya Horiguchi 11026664bfc8SWei Yang return NULL; 11036664bfc8SWei Yang } 11046664bfc8SWei Yang 11053e59fcb0SMichal Hocko static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, 11063e59fcb0SMichal Hocko nodemask_t *nmask) 110794310cbcSAnshuman Khandual { 11083e59fcb0SMichal Hocko unsigned int cpuset_mems_cookie; 11093e59fcb0SMichal Hocko struct zonelist *zonelist; 11103e59fcb0SMichal Hocko struct zone *zone; 11113e59fcb0SMichal Hocko struct zoneref *z; 111298fa15f3SAnshuman Khandual int node = NUMA_NO_NODE; 11133e59fcb0SMichal Hocko 11143e59fcb0SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 11153e59fcb0SMichal Hocko 11163e59fcb0SMichal Hocko retry_cpuset: 11173e59fcb0SMichal Hocko cpuset_mems_cookie = read_mems_allowed_begin(); 11183e59fcb0SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 111994310cbcSAnshuman Khandual struct page *page; 112094310cbcSAnshuman Khandual 11213e59fcb0SMichal Hocko if (!cpuset_zone_allowed(zone, gfp_mask)) 11223e59fcb0SMichal Hocko continue; 11233e59fcb0SMichal Hocko /* 11243e59fcb0SMichal Hocko * no need to ask again on the same node. Pool is node rather than 11253e59fcb0SMichal Hocko * zone aware 11263e59fcb0SMichal Hocko */ 11273e59fcb0SMichal Hocko if (zone_to_nid(zone) == node) 11283e59fcb0SMichal Hocko continue; 11293e59fcb0SMichal Hocko node = zone_to_nid(zone); 113094310cbcSAnshuman Khandual 113194310cbcSAnshuman Khandual page = dequeue_huge_page_node_exact(h, node); 113294310cbcSAnshuman Khandual if (page) 113394310cbcSAnshuman Khandual return page; 113494310cbcSAnshuman Khandual } 11353e59fcb0SMichal Hocko if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 11363e59fcb0SMichal Hocko goto retry_cpuset; 11373e59fcb0SMichal Hocko 113894310cbcSAnshuman Khandual return NULL; 113994310cbcSAnshuman Khandual } 114094310cbcSAnshuman Khandual 1141a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 1142a5516438SAndi Kleen struct vm_area_struct *vma, 1143af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 1144af0ed73eSJoonsoo Kim long chg) 11451da177e4SLinus Torvalds { 11463e59fcb0SMichal Hocko struct page *page; 1147480eccf9SLee Schermerhorn struct mempolicy *mpol; 114804ec6264SVlastimil Babka gfp_t gfp_mask; 11493e59fcb0SMichal Hocko nodemask_t *nodemask; 115004ec6264SVlastimil Babka int nid; 11511da177e4SLinus Torvalds 1152a1e78772SMel Gorman /* 1153a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 1154a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 1155a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 1156a1e78772SMel Gorman */ 1157af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg) && 1158a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 1159c0ff7453SMiao Xie goto err; 1160a1e78772SMel Gorman 116104f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 1162a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 11636eab04a8SJustin P. Mattock goto err; 116404f2cbe3SMel Gorman 116504ec6264SVlastimil Babka gfp_mask = htlb_alloc_mask(h); 116604ec6264SVlastimil Babka nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 11673e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 11683e59fcb0SMichal Hocko if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { 1169d6995da3SMike Kravetz SetHPageRestoreReserve(page); 1170a63884e9SJoonsoo Kim h->resv_huge_pages--; 1171bf50bab2SNaoya Horiguchi } 1172cc9a6c87SMel Gorman 1173cc9a6c87SMel Gorman mpol_cond_put(mpol); 1174cc9a6c87SMel Gorman return page; 1175cc9a6c87SMel Gorman 1176c0ff7453SMiao Xie err: 1177cc9a6c87SMel Gorman return NULL; 11781da177e4SLinus Torvalds } 11791da177e4SLinus Torvalds 11801cac6f2cSLuiz Capitulino /* 11811cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 11821cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 11831cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 11841cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 11851cac6f2cSLuiz Capitulino * node for alloc or free. 11861cac6f2cSLuiz Capitulino */ 11871cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 11881cac6f2cSLuiz Capitulino { 11890edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 11901cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 11911cac6f2cSLuiz Capitulino 11921cac6f2cSLuiz Capitulino return nid; 11931cac6f2cSLuiz Capitulino } 11941cac6f2cSLuiz Capitulino 11951cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 11961cac6f2cSLuiz Capitulino { 11971cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 11981cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 11991cac6f2cSLuiz Capitulino return nid; 12001cac6f2cSLuiz Capitulino } 12011cac6f2cSLuiz Capitulino 12021cac6f2cSLuiz Capitulino /* 12031cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 12041cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 12051cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 12061cac6f2cSLuiz Capitulino * mask. 12071cac6f2cSLuiz Capitulino */ 12081cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 12091cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 12101cac6f2cSLuiz Capitulino { 12111cac6f2cSLuiz Capitulino int nid; 12121cac6f2cSLuiz Capitulino 12131cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 12141cac6f2cSLuiz Capitulino 12151cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 12161cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 12171cac6f2cSLuiz Capitulino 12181cac6f2cSLuiz Capitulino return nid; 12191cac6f2cSLuiz Capitulino } 12201cac6f2cSLuiz Capitulino 12211cac6f2cSLuiz Capitulino /* 122210c6ec49SMike Kravetz * helper for remove_pool_huge_page() - return the previously saved 12231cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 12241cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 12251cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 12261cac6f2cSLuiz Capitulino */ 12271cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 12281cac6f2cSLuiz Capitulino { 12291cac6f2cSLuiz Capitulino int nid; 12301cac6f2cSLuiz Capitulino 12311cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 12321cac6f2cSLuiz Capitulino 12331cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 12341cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 12351cac6f2cSLuiz Capitulino 12361cac6f2cSLuiz Capitulino return nid; 12371cac6f2cSLuiz Capitulino } 12381cac6f2cSLuiz Capitulino 12391cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 12401cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 12411cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 12421cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 12431cac6f2cSLuiz Capitulino nr_nodes--) 12441cac6f2cSLuiz Capitulino 12451cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 12461cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 12471cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 12481cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 12491cac6f2cSLuiz Capitulino nr_nodes--) 12501cac6f2cSLuiz Capitulino 1251e1073d1eSAneesh Kumar K.V #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 1252944d9fecSLuiz Capitulino static void destroy_compound_gigantic_page(struct page *page, 1253d00181b9SKirill A. Shutemov unsigned int order) 1254944d9fecSLuiz Capitulino { 1255944d9fecSLuiz Capitulino int i; 1256944d9fecSLuiz Capitulino int nr_pages = 1 << order; 1257944d9fecSLuiz Capitulino struct page *p = page + 1; 1258944d9fecSLuiz Capitulino 1259c8cc708aSGerald Schaefer atomic_set(compound_mapcount_ptr(page), 0); 126047e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 126147e29d32SJohn Hubbard 1262944d9fecSLuiz Capitulino for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 12631d798ca3SKirill A. Shutemov clear_compound_head(p); 1264944d9fecSLuiz Capitulino set_page_refcounted(p); 1265944d9fecSLuiz Capitulino } 1266944d9fecSLuiz Capitulino 1267944d9fecSLuiz Capitulino set_compound_order(page, 0); 1268ba9c1201SGerald Schaefer page[1].compound_nr = 0; 1269944d9fecSLuiz Capitulino __ClearPageHead(page); 1270944d9fecSLuiz Capitulino } 1271944d9fecSLuiz Capitulino 1272d00181b9SKirill A. Shutemov static void free_gigantic_page(struct page *page, unsigned int order) 1273944d9fecSLuiz Capitulino { 1274cf11e85fSRoman Gushchin /* 1275cf11e85fSRoman Gushchin * If the page isn't allocated using the cma allocator, 1276cf11e85fSRoman Gushchin * cma_release() returns false. 1277cf11e85fSRoman Gushchin */ 1278dbda8feaSBarry Song #ifdef CONFIG_CMA 1279dbda8feaSBarry Song if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order)) 1280cf11e85fSRoman Gushchin return; 1281dbda8feaSBarry Song #endif 1282cf11e85fSRoman Gushchin 1283944d9fecSLuiz Capitulino free_contig_range(page_to_pfn(page), 1 << order); 1284944d9fecSLuiz Capitulino } 1285944d9fecSLuiz Capitulino 12864eb0716eSAlexandre Ghiti #ifdef CONFIG_CONTIG_ALLOC 1287d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 1288d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) 1289944d9fecSLuiz Capitulino { 129004adbc3fSMiaohe Lin unsigned long nr_pages = pages_per_huge_page(h); 1291953f064aSLi Xinhai if (nid == NUMA_NO_NODE) 1292953f064aSLi Xinhai nid = numa_mem_id(); 1293944d9fecSLuiz Capitulino 1294dbda8feaSBarry Song #ifdef CONFIG_CMA 1295dbda8feaSBarry Song { 1296cf11e85fSRoman Gushchin struct page *page; 1297cf11e85fSRoman Gushchin int node; 1298cf11e85fSRoman Gushchin 1299953f064aSLi Xinhai if (hugetlb_cma[nid]) { 1300953f064aSLi Xinhai page = cma_alloc(hugetlb_cma[nid], nr_pages, 1301953f064aSLi Xinhai huge_page_order(h), true); 1302953f064aSLi Xinhai if (page) 1303953f064aSLi Xinhai return page; 1304953f064aSLi Xinhai } 1305953f064aSLi Xinhai 1306953f064aSLi Xinhai if (!(gfp_mask & __GFP_THISNODE)) { 1307cf11e85fSRoman Gushchin for_each_node_mask(node, *nodemask) { 1308953f064aSLi Xinhai if (node == nid || !hugetlb_cma[node]) 1309cf11e85fSRoman Gushchin continue; 1310cf11e85fSRoman Gushchin 1311cf11e85fSRoman Gushchin page = cma_alloc(hugetlb_cma[node], nr_pages, 1312cf11e85fSRoman Gushchin huge_page_order(h), true); 1313cf11e85fSRoman Gushchin if (page) 1314cf11e85fSRoman Gushchin return page; 1315cf11e85fSRoman Gushchin } 1316cf11e85fSRoman Gushchin } 1317953f064aSLi Xinhai } 1318dbda8feaSBarry Song #endif 1319cf11e85fSRoman Gushchin 13205e27a2dfSAnshuman Khandual return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); 1321944d9fecSLuiz Capitulino } 1322944d9fecSLuiz Capitulino 13234eb0716eSAlexandre Ghiti #else /* !CONFIG_CONTIG_ALLOC */ 13244eb0716eSAlexandre Ghiti static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 13254eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 13264eb0716eSAlexandre Ghiti { 13274eb0716eSAlexandre Ghiti return NULL; 13284eb0716eSAlexandre Ghiti } 13294eb0716eSAlexandre Ghiti #endif /* CONFIG_CONTIG_ALLOC */ 1330944d9fecSLuiz Capitulino 1331e1073d1eSAneesh Kumar K.V #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 1332d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 13334eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 13344eb0716eSAlexandre Ghiti { 13354eb0716eSAlexandre Ghiti return NULL; 13364eb0716eSAlexandre Ghiti } 1337d00181b9SKirill A. Shutemov static inline void free_gigantic_page(struct page *page, unsigned int order) { } 1338944d9fecSLuiz Capitulino static inline void destroy_compound_gigantic_page(struct page *page, 1339d00181b9SKirill A. Shutemov unsigned int order) { } 1340944d9fecSLuiz Capitulino #endif 1341944d9fecSLuiz Capitulino 13426eb4e88aSMike Kravetz /* 13436eb4e88aSMike Kravetz * Remove hugetlb page from lists, and update dtor so that page appears 13446eb4e88aSMike Kravetz * as just a compound page. A reference is held on the page. 13456eb4e88aSMike Kravetz * 13466eb4e88aSMike Kravetz * Must be called with hugetlb lock held. 13476eb4e88aSMike Kravetz */ 13486eb4e88aSMike Kravetz static void remove_hugetlb_page(struct hstate *h, struct page *page, 13496eb4e88aSMike Kravetz bool adjust_surplus) 13506eb4e88aSMike Kravetz { 13516eb4e88aSMike Kravetz int nid = page_to_nid(page); 13526eb4e88aSMike Kravetz 13536eb4e88aSMike Kravetz VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 13546eb4e88aSMike Kravetz VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page); 13556eb4e88aSMike Kravetz 13569487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 13576eb4e88aSMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 13586eb4e88aSMike Kravetz return; 13596eb4e88aSMike Kravetz 13606eb4e88aSMike Kravetz list_del(&page->lru); 13616eb4e88aSMike Kravetz 13626eb4e88aSMike Kravetz if (HPageFreed(page)) { 13636eb4e88aSMike Kravetz h->free_huge_pages--; 13646eb4e88aSMike Kravetz h->free_huge_pages_node[nid]--; 13656eb4e88aSMike Kravetz } 13666eb4e88aSMike Kravetz if (adjust_surplus) { 13676eb4e88aSMike Kravetz h->surplus_huge_pages--; 13686eb4e88aSMike Kravetz h->surplus_huge_pages_node[nid]--; 13696eb4e88aSMike Kravetz } 13706eb4e88aSMike Kravetz 13716eb4e88aSMike Kravetz set_page_refcounted(page); 13726eb4e88aSMike Kravetz set_compound_page_dtor(page, NULL_COMPOUND_DTOR); 13736eb4e88aSMike Kravetz 13746eb4e88aSMike Kravetz h->nr_huge_pages--; 13756eb4e88aSMike Kravetz h->nr_huge_pages_node[nid]--; 13766eb4e88aSMike Kravetz } 13776eb4e88aSMike Kravetz 1378ad2fa371SMuchun Song static void add_hugetlb_page(struct hstate *h, struct page *page, 1379ad2fa371SMuchun Song bool adjust_surplus) 1380ad2fa371SMuchun Song { 1381ad2fa371SMuchun Song int zeroed; 1382ad2fa371SMuchun Song int nid = page_to_nid(page); 1383ad2fa371SMuchun Song 1384ad2fa371SMuchun Song VM_BUG_ON_PAGE(!HPageVmemmapOptimized(page), page); 1385ad2fa371SMuchun Song 1386ad2fa371SMuchun Song lockdep_assert_held(&hugetlb_lock); 1387ad2fa371SMuchun Song 1388ad2fa371SMuchun Song INIT_LIST_HEAD(&page->lru); 1389ad2fa371SMuchun Song h->nr_huge_pages++; 1390ad2fa371SMuchun Song h->nr_huge_pages_node[nid]++; 1391ad2fa371SMuchun Song 1392ad2fa371SMuchun Song if (adjust_surplus) { 1393ad2fa371SMuchun Song h->surplus_huge_pages++; 1394ad2fa371SMuchun Song h->surplus_huge_pages_node[nid]++; 1395ad2fa371SMuchun Song } 1396ad2fa371SMuchun Song 1397ad2fa371SMuchun Song set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1398ad2fa371SMuchun Song set_page_private(page, 0); 1399ad2fa371SMuchun Song SetHPageVmemmapOptimized(page); 1400ad2fa371SMuchun Song 1401ad2fa371SMuchun Song /* 1402ad2fa371SMuchun Song * This page is now managed by the hugetlb allocator and has 1403ad2fa371SMuchun Song * no users -- drop the last reference. 1404ad2fa371SMuchun Song */ 1405ad2fa371SMuchun Song zeroed = put_page_testzero(page); 1406ad2fa371SMuchun Song VM_BUG_ON_PAGE(!zeroed, page); 1407ad2fa371SMuchun Song arch_clear_hugepage_flags(page); 1408ad2fa371SMuchun Song enqueue_huge_page(h, page); 1409ad2fa371SMuchun Song } 1410ad2fa371SMuchun Song 1411b65d4adbSMuchun Song static void __update_and_free_page(struct hstate *h, struct page *page) 14126af2acb6SAdam Litke { 14136af2acb6SAdam Litke int i; 1414dbfee5aeSMike Kravetz struct page *subpage = page; 1415a5516438SAndi Kleen 14164eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1417944d9fecSLuiz Capitulino return; 141818229df5SAndy Whitcroft 1419ad2fa371SMuchun Song if (alloc_huge_page_vmemmap(h, page)) { 1420ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 1421ad2fa371SMuchun Song /* 1422ad2fa371SMuchun Song * If we cannot allocate vmemmap pages, just refuse to free the 1423ad2fa371SMuchun Song * page and put the page back on the hugetlb free list and treat 1424ad2fa371SMuchun Song * as a surplus page. 1425ad2fa371SMuchun Song */ 1426ad2fa371SMuchun Song add_hugetlb_page(h, page, true); 1427ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 1428ad2fa371SMuchun Song return; 1429ad2fa371SMuchun Song } 1430ad2fa371SMuchun Song 1431dbfee5aeSMike Kravetz for (i = 0; i < pages_per_huge_page(h); 1432dbfee5aeSMike Kravetz i++, subpage = mem_map_next(subpage, page, i)) { 1433dbfee5aeSMike Kravetz subpage->flags &= ~(1 << PG_locked | 1 << PG_error | 143432f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1435a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1436a7407a27SLuiz Capitulino 1 << PG_writeback); 14376af2acb6SAdam Litke } 1438944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) { 1439944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 1440944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 1441944d9fecSLuiz Capitulino } else { 1442a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 14436af2acb6SAdam Litke } 1444944d9fecSLuiz Capitulino } 14456af2acb6SAdam Litke 1446b65d4adbSMuchun Song /* 1447b65d4adbSMuchun Song * As update_and_free_page() can be called under any context, so we cannot 1448b65d4adbSMuchun Song * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the 1449b65d4adbSMuchun Song * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate 1450b65d4adbSMuchun Song * the vmemmap pages. 1451b65d4adbSMuchun Song * 1452b65d4adbSMuchun Song * free_hpage_workfn() locklessly retrieves the linked list of pages to be 1453b65d4adbSMuchun Song * freed and frees them one-by-one. As the page->mapping pointer is going 1454b65d4adbSMuchun Song * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node 1455b65d4adbSMuchun Song * structure of a lockless linked list of huge pages to be freed. 1456b65d4adbSMuchun Song */ 1457b65d4adbSMuchun Song static LLIST_HEAD(hpage_freelist); 1458b65d4adbSMuchun Song 1459b65d4adbSMuchun Song static void free_hpage_workfn(struct work_struct *work) 1460b65d4adbSMuchun Song { 1461b65d4adbSMuchun Song struct llist_node *node; 1462b65d4adbSMuchun Song 1463b65d4adbSMuchun Song node = llist_del_all(&hpage_freelist); 1464b65d4adbSMuchun Song 1465b65d4adbSMuchun Song while (node) { 1466b65d4adbSMuchun Song struct page *page; 1467b65d4adbSMuchun Song struct hstate *h; 1468b65d4adbSMuchun Song 1469b65d4adbSMuchun Song page = container_of((struct address_space **)node, 1470b65d4adbSMuchun Song struct page, mapping); 1471b65d4adbSMuchun Song node = node->next; 1472b65d4adbSMuchun Song page->mapping = NULL; 1473b65d4adbSMuchun Song /* 1474b65d4adbSMuchun Song * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate() 1475b65d4adbSMuchun Song * is going to trigger because a previous call to 1476b65d4adbSMuchun Song * remove_hugetlb_page() will set_compound_page_dtor(page, 1477b65d4adbSMuchun Song * NULL_COMPOUND_DTOR), so do not use page_hstate() directly. 1478b65d4adbSMuchun Song */ 1479b65d4adbSMuchun Song h = size_to_hstate(page_size(page)); 1480b65d4adbSMuchun Song 1481b65d4adbSMuchun Song __update_and_free_page(h, page); 1482b65d4adbSMuchun Song 1483b65d4adbSMuchun Song cond_resched(); 1484b65d4adbSMuchun Song } 1485b65d4adbSMuchun Song } 1486b65d4adbSMuchun Song static DECLARE_WORK(free_hpage_work, free_hpage_workfn); 1487b65d4adbSMuchun Song 1488b65d4adbSMuchun Song static inline void flush_free_hpage_work(struct hstate *h) 1489b65d4adbSMuchun Song { 1490b65d4adbSMuchun Song if (free_vmemmap_pages_per_hpage(h)) 1491b65d4adbSMuchun Song flush_work(&free_hpage_work); 1492b65d4adbSMuchun Song } 1493b65d4adbSMuchun Song 1494b65d4adbSMuchun Song static void update_and_free_page(struct hstate *h, struct page *page, 1495b65d4adbSMuchun Song bool atomic) 1496b65d4adbSMuchun Song { 1497ad2fa371SMuchun Song if (!HPageVmemmapOptimized(page) || !atomic) { 1498b65d4adbSMuchun Song __update_and_free_page(h, page); 1499b65d4adbSMuchun Song return; 1500b65d4adbSMuchun Song } 1501b65d4adbSMuchun Song 1502b65d4adbSMuchun Song /* 1503b65d4adbSMuchun Song * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages. 1504b65d4adbSMuchun Song * 1505b65d4adbSMuchun Song * Only call schedule_work() if hpage_freelist is previously 1506b65d4adbSMuchun Song * empty. Otherwise, schedule_work() had been called but the workfn 1507b65d4adbSMuchun Song * hasn't retrieved the list yet. 1508b65d4adbSMuchun Song */ 1509b65d4adbSMuchun Song if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist)) 1510b65d4adbSMuchun Song schedule_work(&free_hpage_work); 1511b65d4adbSMuchun Song } 1512b65d4adbSMuchun Song 151310c6ec49SMike Kravetz static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list) 151410c6ec49SMike Kravetz { 151510c6ec49SMike Kravetz struct page *page, *t_page; 151610c6ec49SMike Kravetz 151710c6ec49SMike Kravetz list_for_each_entry_safe(page, t_page, list, lru) { 1518b65d4adbSMuchun Song update_and_free_page(h, page, false); 151910c6ec49SMike Kravetz cond_resched(); 152010c6ec49SMike Kravetz } 152110c6ec49SMike Kravetz } 152210c6ec49SMike Kravetz 1523e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1524e5ff2159SAndi Kleen { 1525e5ff2159SAndi Kleen struct hstate *h; 1526e5ff2159SAndi Kleen 1527e5ff2159SAndi Kleen for_each_hstate(h) { 1528e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1529e5ff2159SAndi Kleen return h; 1530e5ff2159SAndi Kleen } 1531e5ff2159SAndi Kleen return NULL; 1532e5ff2159SAndi Kleen } 1533e5ff2159SAndi Kleen 1534db71ef79SMike Kravetz void free_huge_page(struct page *page) 153527a85ef1SDavid Gibson { 1536a5516438SAndi Kleen /* 1537a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1538a5516438SAndi Kleen * compound page destructor. 1539a5516438SAndi Kleen */ 1540e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 15417893d1d5SAdam Litke int nid = page_to_nid(page); 1542d6995da3SMike Kravetz struct hugepage_subpool *spool = hugetlb_page_subpool(page); 154307443a85SJoonsoo Kim bool restore_reserve; 1544db71ef79SMike Kravetz unsigned long flags; 154527a85ef1SDavid Gibson 1546b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1547b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_mapcount(page), page); 15488ace22bcSYongkai Wu 1549d6995da3SMike Kravetz hugetlb_set_page_subpool(page, NULL); 15508ace22bcSYongkai Wu page->mapping = NULL; 1551d6995da3SMike Kravetz restore_reserve = HPageRestoreReserve(page); 1552d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 155327a85ef1SDavid Gibson 15541c5ecae3SMike Kravetz /* 1555d6995da3SMike Kravetz * If HPageRestoreReserve was set on page, page allocation consumed a 15560919e1b6SMike Kravetz * reservation. If the page was associated with a subpool, there 15570919e1b6SMike Kravetz * would have been a page reserved in the subpool before allocation 15580919e1b6SMike Kravetz * via hugepage_subpool_get_pages(). Since we are 'restoring' the 15596c26d310SMiaohe Lin * reservation, do not call hugepage_subpool_put_pages() as this will 15600919e1b6SMike Kravetz * remove the reserved page from the subpool. 15610919e1b6SMike Kravetz */ 15620919e1b6SMike Kravetz if (!restore_reserve) { 15630919e1b6SMike Kravetz /* 15640919e1b6SMike Kravetz * A return code of zero implies that the subpool will be 15650919e1b6SMike Kravetz * under its minimum size if the reservation is not restored 15660919e1b6SMike Kravetz * after page is free. Therefore, force restore_reserve 15670919e1b6SMike Kravetz * operation. 15681c5ecae3SMike Kravetz */ 15691c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 15701c5ecae3SMike Kravetz restore_reserve = true; 15710919e1b6SMike Kravetz } 15721c5ecae3SMike Kravetz 1573db71ef79SMike Kravetz spin_lock_irqsave(&hugetlb_lock, flags); 15748f251a3dSMike Kravetz ClearHPageMigratable(page); 15756d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 15766d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 157708cf9fafSMina Almasry hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 157808cf9fafSMina Almasry pages_per_huge_page(h), page); 157907443a85SJoonsoo Kim if (restore_reserve) 158007443a85SJoonsoo Kim h->resv_huge_pages++; 158107443a85SJoonsoo Kim 15829157c311SMike Kravetz if (HPageTemporary(page)) { 15836eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 1584db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1585b65d4adbSMuchun Song update_and_free_page(h, page, true); 1586ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 15870edaecfaSAneesh Kumar K.V /* remove the page from active list */ 15886eb4e88aSMike Kravetz remove_hugetlb_page(h, page, true); 1589db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1590b65d4adbSMuchun Song update_and_free_page(h, page, true); 15917893d1d5SAdam Litke } else { 15925d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1593a5516438SAndi Kleen enqueue_huge_page(h, page); 1594db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 159527a85ef1SDavid Gibson } 15961121828aSMike Kravetz } 159727a85ef1SDavid Gibson 1598d3d99fccSOscar Salvador /* 1599d3d99fccSOscar Salvador * Must be called with the hugetlb lock held 1600d3d99fccSOscar Salvador */ 1601d3d99fccSOscar Salvador static void __prep_account_new_huge_page(struct hstate *h, int nid) 1602d3d99fccSOscar Salvador { 1603d3d99fccSOscar Salvador lockdep_assert_held(&hugetlb_lock); 1604d3d99fccSOscar Salvador h->nr_huge_pages++; 1605d3d99fccSOscar Salvador h->nr_huge_pages_node[nid]++; 1606d3d99fccSOscar Salvador } 1607d3d99fccSOscar Salvador 1608f41f2ed4SMuchun Song static void __prep_new_huge_page(struct hstate *h, struct page *page) 1609b7ba30c6SAndi Kleen { 1610f41f2ed4SMuchun Song free_huge_page_vmemmap(h, page); 16110edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1612f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1613ff546117SMike Kravetz hugetlb_set_page_subpool(page, NULL); 16149dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 16151adc4d41SMina Almasry set_hugetlb_cgroup_rsvd(page, NULL); 1616d3d99fccSOscar Salvador } 1617d3d99fccSOscar Salvador 1618d3d99fccSOscar Salvador static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1619d3d99fccSOscar Salvador { 1620f41f2ed4SMuchun Song __prep_new_huge_page(h, page); 1621db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 1622d3d99fccSOscar Salvador __prep_account_new_huge_page(h, nid); 1623db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1624b7ba30c6SAndi Kleen } 1625b7ba30c6SAndi Kleen 1626*7118fc29SMike Kravetz static bool prep_compound_gigantic_page(struct page *page, unsigned int order) 162720a0307cSWu Fengguang { 1628*7118fc29SMike Kravetz int i, j; 162920a0307cSWu Fengguang int nr_pages = 1 << order; 163020a0307cSWu Fengguang struct page *p = page + 1; 163120a0307cSWu Fengguang 163220a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 163320a0307cSWu Fengguang set_compound_order(page, order); 1634ef5a22beSAndrea Arcangeli __ClearPageReserved(page); 1635de09d31dSKirill A. Shutemov __SetPageHead(page); 163620a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1637ef5a22beSAndrea Arcangeli /* 1638ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1639ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1640ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 16417c8de358SEthon Paul * too. Otherwise drivers using get_user_pages() to access tail 1642ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1643ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1644ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1645ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1646ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1647ef5a22beSAndrea Arcangeli * after get_user_pages(). 1648ef5a22beSAndrea Arcangeli */ 1649ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 1650*7118fc29SMike Kravetz /* 1651*7118fc29SMike Kravetz * Subtle and very unlikely 1652*7118fc29SMike Kravetz * 1653*7118fc29SMike Kravetz * Gigantic 'page allocators' such as memblock or cma will 1654*7118fc29SMike Kravetz * return a set of pages with each page ref counted. We need 1655*7118fc29SMike Kravetz * to turn this set of pages into a compound page with tail 1656*7118fc29SMike Kravetz * page ref counts set to zero. Code such as speculative page 1657*7118fc29SMike Kravetz * cache adding could take a ref on a 'to be' tail page. 1658*7118fc29SMike Kravetz * We need to respect any increased ref count, and only set 1659*7118fc29SMike Kravetz * the ref count to zero if count is currently 1. If count 1660*7118fc29SMike Kravetz * is not 1, we call synchronize_rcu in the hope that a rcu 1661*7118fc29SMike Kravetz * grace period will cause ref count to drop and then retry. 1662*7118fc29SMike Kravetz * If count is still inflated on retry we return an error and 1663*7118fc29SMike Kravetz * must discard the pages. 1664*7118fc29SMike Kravetz */ 1665*7118fc29SMike Kravetz if (!page_ref_freeze(p, 1)) { 1666*7118fc29SMike Kravetz pr_info("HugeTLB unexpected inflated ref count on freshly allocated page\n"); 1667*7118fc29SMike Kravetz synchronize_rcu(); 1668*7118fc29SMike Kravetz if (!page_ref_freeze(p, 1)) 1669*7118fc29SMike Kravetz goto out_error; 1670*7118fc29SMike Kravetz } 167158a84aa9SYouquan Song set_page_count(p, 0); 16721d798ca3SKirill A. Shutemov set_compound_head(p, page); 167320a0307cSWu Fengguang } 1674b4330afbSMike Kravetz atomic_set(compound_mapcount_ptr(page), -1); 167547e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 1676*7118fc29SMike Kravetz return true; 1677*7118fc29SMike Kravetz 1678*7118fc29SMike Kravetz out_error: 1679*7118fc29SMike Kravetz /* undo tail page modifications made above */ 1680*7118fc29SMike Kravetz p = page + 1; 1681*7118fc29SMike Kravetz for (j = 1; j < i; j++, p = mem_map_next(p, page, j)) { 1682*7118fc29SMike Kravetz clear_compound_head(p); 1683*7118fc29SMike Kravetz set_page_refcounted(p); 1684*7118fc29SMike Kravetz } 1685*7118fc29SMike Kravetz /* need to clear PG_reserved on remaining tail pages */ 1686*7118fc29SMike Kravetz for (; j < nr_pages; j++, p = mem_map_next(p, page, j)) 1687*7118fc29SMike Kravetz __ClearPageReserved(p); 1688*7118fc29SMike Kravetz set_compound_order(page, 0); 1689*7118fc29SMike Kravetz page[1].compound_nr = 0; 1690*7118fc29SMike Kravetz __ClearPageHead(page); 1691*7118fc29SMike Kravetz return false; 169220a0307cSWu Fengguang } 169320a0307cSWu Fengguang 16947795912cSAndrew Morton /* 16957795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 16967795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 16977795912cSAndrew Morton * details. 16987795912cSAndrew Morton */ 169920a0307cSWu Fengguang int PageHuge(struct page *page) 170020a0307cSWu Fengguang { 170120a0307cSWu Fengguang if (!PageCompound(page)) 170220a0307cSWu Fengguang return 0; 170320a0307cSWu Fengguang 170420a0307cSWu Fengguang page = compound_head(page); 1705f1e61557SKirill A. Shutemov return page[1].compound_dtor == HUGETLB_PAGE_DTOR; 170620a0307cSWu Fengguang } 170743131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 170843131e14SNaoya Horiguchi 170927c73ae7SAndrea Arcangeli /* 171027c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 171127c73ae7SAndrea Arcangeli * normal or transparent huge pages. 171227c73ae7SAndrea Arcangeli */ 171327c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 171427c73ae7SAndrea Arcangeli { 171527c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 171627c73ae7SAndrea Arcangeli return 0; 171727c73ae7SAndrea Arcangeli 1718d4af73e3SVlastimil Babka return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; 171927c73ae7SAndrea Arcangeli } 172027c73ae7SAndrea Arcangeli 1721c0d0381aSMike Kravetz /* 1722c0d0381aSMike Kravetz * Find and lock address space (mapping) in write mode. 1723c0d0381aSMike Kravetz * 1724336bf30eSMike Kravetz * Upon entry, the page is locked which means that page_mapping() is 1725336bf30eSMike Kravetz * stable. Due to locking order, we can only trylock_write. If we can 1726336bf30eSMike Kravetz * not get the lock, simply return NULL to caller. 1727c0d0381aSMike Kravetz */ 1728c0d0381aSMike Kravetz struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) 1729c0d0381aSMike Kravetz { 1730336bf30eSMike Kravetz struct address_space *mapping = page_mapping(hpage); 1731c0d0381aSMike Kravetz 1732c0d0381aSMike Kravetz if (!mapping) 1733c0d0381aSMike Kravetz return mapping; 1734c0d0381aSMike Kravetz 1735c0d0381aSMike Kravetz if (i_mmap_trylock_write(mapping)) 1736c0d0381aSMike Kravetz return mapping; 1737c0d0381aSMike Kravetz 1738c0d0381aSMike Kravetz return NULL; 1739c0d0381aSMike Kravetz } 1740c0d0381aSMike Kravetz 1741fe19bd3dSHugh Dickins pgoff_t hugetlb_basepage_index(struct page *page) 174213d60f4bSZhang Yi { 174313d60f4bSZhang Yi struct page *page_head = compound_head(page); 174413d60f4bSZhang Yi pgoff_t index = page_index(page_head); 174513d60f4bSZhang Yi unsigned long compound_idx; 174613d60f4bSZhang Yi 174713d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 174813d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 174913d60f4bSZhang Yi else 175013d60f4bSZhang Yi compound_idx = page - page_head; 175113d60f4bSZhang Yi 175213d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 175313d60f4bSZhang Yi } 175413d60f4bSZhang Yi 17550c397daeSMichal Hocko static struct page *alloc_buddy_huge_page(struct hstate *h, 1756f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1757f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 17581da177e4SLinus Torvalds { 1759af0fb9dfSMichal Hocko int order = huge_page_order(h); 17601da177e4SLinus Torvalds struct page *page; 1761f60858f9SMike Kravetz bool alloc_try_hard = true; 1762f96efd58SJoe Jin 1763f60858f9SMike Kravetz /* 1764f60858f9SMike Kravetz * By default we always try hard to allocate the page with 1765f60858f9SMike Kravetz * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in 1766f60858f9SMike Kravetz * a loop (to adjust global huge page counts) and previous allocation 1767f60858f9SMike Kravetz * failed, do not continue to try hard on the same node. Use the 1768f60858f9SMike Kravetz * node_alloc_noretry bitmap to manage this state information. 1769f60858f9SMike Kravetz */ 1770f60858f9SMike Kravetz if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) 1771f60858f9SMike Kravetz alloc_try_hard = false; 1772f60858f9SMike Kravetz gfp_mask |= __GFP_COMP|__GFP_NOWARN; 1773f60858f9SMike Kravetz if (alloc_try_hard) 1774f60858f9SMike Kravetz gfp_mask |= __GFP_RETRY_MAYFAIL; 1775af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 1776af0fb9dfSMichal Hocko nid = numa_mem_id(); 177784172f4bSMatthew Wilcox (Oracle) page = __alloc_pages(gfp_mask, order, nid, nmask); 1778af0fb9dfSMichal Hocko if (page) 1779af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC); 1780af0fb9dfSMichal Hocko else 1781af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 178263b4613cSNishanth Aravamudan 1783f60858f9SMike Kravetz /* 1784f60858f9SMike Kravetz * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this 1785f60858f9SMike Kravetz * indicates an overall state change. Clear bit so that we resume 1786f60858f9SMike Kravetz * normal 'try hard' allocations. 1787f60858f9SMike Kravetz */ 1788f60858f9SMike Kravetz if (node_alloc_noretry && page && !alloc_try_hard) 1789f60858f9SMike Kravetz node_clear(nid, *node_alloc_noretry); 1790f60858f9SMike Kravetz 1791f60858f9SMike Kravetz /* 1792f60858f9SMike Kravetz * If we tried hard to get a page but failed, set bit so that 1793f60858f9SMike Kravetz * subsequent attempts will not try as hard until there is an 1794f60858f9SMike Kravetz * overall state change. 1795f60858f9SMike Kravetz */ 1796f60858f9SMike Kravetz if (node_alloc_noretry && !page && alloc_try_hard) 1797f60858f9SMike Kravetz node_set(nid, *node_alloc_noretry); 1798f60858f9SMike Kravetz 179963b4613cSNishanth Aravamudan return page; 180063b4613cSNishanth Aravamudan } 180163b4613cSNishanth Aravamudan 1802af0fb9dfSMichal Hocko /* 18030c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 18040c397daeSMichal Hocko * should use this function to get new hugetlb pages 18050c397daeSMichal Hocko */ 18060c397daeSMichal Hocko static struct page *alloc_fresh_huge_page(struct hstate *h, 1807f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1808f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 18090c397daeSMichal Hocko { 18100c397daeSMichal Hocko struct page *page; 1811*7118fc29SMike Kravetz bool retry = false; 18120c397daeSMichal Hocko 1813*7118fc29SMike Kravetz retry: 18140c397daeSMichal Hocko if (hstate_is_gigantic(h)) 18150c397daeSMichal Hocko page = alloc_gigantic_page(h, gfp_mask, nid, nmask); 18160c397daeSMichal Hocko else 18170c397daeSMichal Hocko page = alloc_buddy_huge_page(h, gfp_mask, 1818f60858f9SMike Kravetz nid, nmask, node_alloc_noretry); 18190c397daeSMichal Hocko if (!page) 18200c397daeSMichal Hocko return NULL; 18210c397daeSMichal Hocko 1822*7118fc29SMike Kravetz if (hstate_is_gigantic(h)) { 1823*7118fc29SMike Kravetz if (!prep_compound_gigantic_page(page, huge_page_order(h))) { 1824*7118fc29SMike Kravetz /* 1825*7118fc29SMike Kravetz * Rare failure to convert pages to compound page. 1826*7118fc29SMike Kravetz * Free pages and try again - ONCE! 1827*7118fc29SMike Kravetz */ 1828*7118fc29SMike Kravetz free_gigantic_page(page, huge_page_order(h)); 1829*7118fc29SMike Kravetz if (!retry) { 1830*7118fc29SMike Kravetz retry = true; 1831*7118fc29SMike Kravetz goto retry; 1832*7118fc29SMike Kravetz } 1833*7118fc29SMike Kravetz pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n"); 1834*7118fc29SMike Kravetz return NULL; 1835*7118fc29SMike Kravetz } 1836*7118fc29SMike Kravetz } 18370c397daeSMichal Hocko prep_new_huge_page(h, page, page_to_nid(page)); 18380c397daeSMichal Hocko 18390c397daeSMichal Hocko return page; 18400c397daeSMichal Hocko } 18410c397daeSMichal Hocko 18420c397daeSMichal Hocko /* 1843af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 1844af0fb9dfSMichal Hocko * manner. 1845af0fb9dfSMichal Hocko */ 1846f60858f9SMike Kravetz static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 1847f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 1848b2261026SJoonsoo Kim { 1849b2261026SJoonsoo Kim struct page *page; 1850b2261026SJoonsoo Kim int nr_nodes, node; 1851af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 1852b2261026SJoonsoo Kim 1853b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1854f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, 1855f60858f9SMike Kravetz node_alloc_noretry); 1856af0fb9dfSMichal Hocko if (page) 1857b2261026SJoonsoo Kim break; 1858b2261026SJoonsoo Kim } 1859b2261026SJoonsoo Kim 1860af0fb9dfSMichal Hocko if (!page) 1861af0fb9dfSMichal Hocko return 0; 1862b2261026SJoonsoo Kim 1863af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 1864af0fb9dfSMichal Hocko 1865af0fb9dfSMichal Hocko return 1; 1866b2261026SJoonsoo Kim } 1867b2261026SJoonsoo Kim 1868e8c5c824SLee Schermerhorn /* 186910c6ec49SMike Kravetz * Remove huge page from pool from next node to free. Attempt to keep 187010c6ec49SMike Kravetz * persistent huge pages more or less balanced over allowed nodes. 187110c6ec49SMike Kravetz * This routine only 'removes' the hugetlb page. The caller must make 187210c6ec49SMike Kravetz * an additional call to free the page to low level allocators. 1873e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 1874e8c5c824SLee Schermerhorn */ 187510c6ec49SMike Kravetz static struct page *remove_pool_huge_page(struct hstate *h, 187610c6ec49SMike Kravetz nodemask_t *nodes_allowed, 18776ae11b27SLee Schermerhorn bool acct_surplus) 1878e8c5c824SLee Schermerhorn { 1879b2261026SJoonsoo Kim int nr_nodes, node; 188010c6ec49SMike Kravetz struct page *page = NULL; 1881e8c5c824SLee Schermerhorn 18829487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1883b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1884685f3457SLee Schermerhorn /* 1885685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 1886685f3457SLee Schermerhorn * nodes with surplus pages. 1887685f3457SLee Schermerhorn */ 1888b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 1889b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 189010c6ec49SMike Kravetz page = list_entry(h->hugepage_freelists[node].next, 1891e8c5c824SLee Schermerhorn struct page, lru); 18926eb4e88aSMike Kravetz remove_hugetlb_page(h, page, acct_surplus); 18939a76db09SLee Schermerhorn break; 1894e8c5c824SLee Schermerhorn } 1895b2261026SJoonsoo Kim } 1896e8c5c824SLee Schermerhorn 189710c6ec49SMike Kravetz return page; 1898e8c5c824SLee Schermerhorn } 1899e8c5c824SLee Schermerhorn 1900c8721bbbSNaoya Horiguchi /* 1901c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 1902faf53defSNaoya Horiguchi * nothing for in-use hugepages and non-hugepages. 1903faf53defSNaoya Horiguchi * This function returns values like below: 1904faf53defSNaoya Horiguchi * 1905ad2fa371SMuchun Song * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages 1906ad2fa371SMuchun Song * when the system is under memory pressure and the feature of 1907ad2fa371SMuchun Song * freeing unused vmemmap pages associated with each hugetlb page 1908ad2fa371SMuchun Song * is enabled. 1909faf53defSNaoya Horiguchi * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 1910faf53defSNaoya Horiguchi * (allocated or reserved.) 1911faf53defSNaoya Horiguchi * 0: successfully dissolved free hugepages or the page is not a 1912faf53defSNaoya Horiguchi * hugepage (considered as already dissolved) 1913c8721bbbSNaoya Horiguchi */ 1914c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 1915c8721bbbSNaoya Horiguchi { 19166bc9b564SNaoya Horiguchi int rc = -EBUSY; 1917082d5b6bSGerald Schaefer 19187ffddd49SMuchun Song retry: 1919faf53defSNaoya Horiguchi /* Not to disrupt normal path by vainly holding hugetlb_lock */ 1920faf53defSNaoya Horiguchi if (!PageHuge(page)) 1921faf53defSNaoya Horiguchi return 0; 1922faf53defSNaoya Horiguchi 1923db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 1924faf53defSNaoya Horiguchi if (!PageHuge(page)) { 1925faf53defSNaoya Horiguchi rc = 0; 1926faf53defSNaoya Horiguchi goto out; 1927faf53defSNaoya Horiguchi } 1928faf53defSNaoya Horiguchi 1929faf53defSNaoya Horiguchi if (!page_count(page)) { 19302247bb33SGerald Schaefer struct page *head = compound_head(page); 19312247bb33SGerald Schaefer struct hstate *h = page_hstate(head); 19326bc9b564SNaoya Horiguchi if (h->free_huge_pages - h->resv_huge_pages == 0) 1933082d5b6bSGerald Schaefer goto out; 19347ffddd49SMuchun Song 19357ffddd49SMuchun Song /* 19367ffddd49SMuchun Song * We should make sure that the page is already on the free list 19377ffddd49SMuchun Song * when it is dissolved. 19387ffddd49SMuchun Song */ 19396c037149SMike Kravetz if (unlikely(!HPageFreed(head))) { 1940db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 19417ffddd49SMuchun Song cond_resched(); 19427ffddd49SMuchun Song 19437ffddd49SMuchun Song /* 19447ffddd49SMuchun Song * Theoretically, we should return -EBUSY when we 19457ffddd49SMuchun Song * encounter this race. In fact, we have a chance 19467ffddd49SMuchun Song * to successfully dissolve the page if we do a 19477ffddd49SMuchun Song * retry. Because the race window is quite small. 19487ffddd49SMuchun Song * If we seize this opportunity, it is an optimization 19497ffddd49SMuchun Song * for increasing the success rate of dissolving page. 19507ffddd49SMuchun Song */ 19517ffddd49SMuchun Song goto retry; 19527ffddd49SMuchun Song } 19537ffddd49SMuchun Song 1954ad2fa371SMuchun Song remove_hugetlb_page(h, head, false); 1955ad2fa371SMuchun Song h->max_huge_pages--; 1956ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 1957ad2fa371SMuchun Song 1958c3114a84SAnshuman Khandual /* 1959ad2fa371SMuchun Song * Normally update_and_free_page will allocate required vmemmmap 1960ad2fa371SMuchun Song * before freeing the page. update_and_free_page will fail to 1961ad2fa371SMuchun Song * free the page if it can not allocate required vmemmap. We 1962ad2fa371SMuchun Song * need to adjust max_huge_pages if the page is not freed. 1963ad2fa371SMuchun Song * Attempt to allocate vmemmmap here so that we can take 1964ad2fa371SMuchun Song * appropriate action on failure. 1965ad2fa371SMuchun Song */ 1966ad2fa371SMuchun Song rc = alloc_huge_page_vmemmap(h, head); 1967ad2fa371SMuchun Song if (!rc) { 1968ad2fa371SMuchun Song /* 1969ad2fa371SMuchun Song * Move PageHWPoison flag from head page to the raw 1970ad2fa371SMuchun Song * error page, which makes any subpages rather than 1971ad2fa371SMuchun Song * the error page reusable. 1972c3114a84SAnshuman Khandual */ 1973c3114a84SAnshuman Khandual if (PageHWPoison(head) && page != head) { 1974c3114a84SAnshuman Khandual SetPageHWPoison(page); 1975c3114a84SAnshuman Khandual ClearPageHWPoison(head); 1976c3114a84SAnshuman Khandual } 1977b65d4adbSMuchun Song update_and_free_page(h, head, false); 1978ad2fa371SMuchun Song } else { 1979ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 1980ad2fa371SMuchun Song add_hugetlb_page(h, head, false); 1981ad2fa371SMuchun Song h->max_huge_pages++; 1982ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 1983ad2fa371SMuchun Song } 1984ad2fa371SMuchun Song 1985ad2fa371SMuchun Song return rc; 1986c8721bbbSNaoya Horiguchi } 1987082d5b6bSGerald Schaefer out: 1988db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1989082d5b6bSGerald Schaefer return rc; 1990c8721bbbSNaoya Horiguchi } 1991c8721bbbSNaoya Horiguchi 1992c8721bbbSNaoya Horiguchi /* 1993c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 1994c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 19952247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 19962247bb33SGerald Schaefer * part of it lies within the given range. 1997082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 1998082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 1999c8721bbbSNaoya Horiguchi */ 2000082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 2001c8721bbbSNaoya Horiguchi { 2002c8721bbbSNaoya Horiguchi unsigned long pfn; 2003eb03aa00SGerald Schaefer struct page *page; 2004082d5b6bSGerald Schaefer int rc = 0; 2005c8721bbbSNaoya Horiguchi 2006d0177639SLi Zhong if (!hugepages_supported()) 2007082d5b6bSGerald Schaefer return rc; 2008d0177639SLi Zhong 2009eb03aa00SGerald Schaefer for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { 2010eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 2011eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 2012eb03aa00SGerald Schaefer if (rc) 2013082d5b6bSGerald Schaefer break; 2014eb03aa00SGerald Schaefer } 2015082d5b6bSGerald Schaefer 2016082d5b6bSGerald Schaefer return rc; 2017c8721bbbSNaoya Horiguchi } 2018c8721bbbSNaoya Horiguchi 2019ab5ac90aSMichal Hocko /* 2020ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 2021ab5ac90aSMichal Hocko */ 20220c397daeSMichal Hocko static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, 2023aaf14e40SMichal Hocko int nid, nodemask_t *nmask) 20247893d1d5SAdam Litke { 20259980d744SMichal Hocko struct page *page = NULL; 20267893d1d5SAdam Litke 2027bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2028aa888a74SAndi Kleen return NULL; 2029aa888a74SAndi Kleen 2030db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 20319980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 20329980d744SMichal Hocko goto out_unlock; 2033db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2034d1c3fb1fSNishanth Aravamudan 2035f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 20369980d744SMichal Hocko if (!page) 20370c397daeSMichal Hocko return NULL; 2038d1c3fb1fSNishanth Aravamudan 2039db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 20409980d744SMichal Hocko /* 20419980d744SMichal Hocko * We could have raced with the pool size change. 20429980d744SMichal Hocko * Double check that and simply deallocate the new page 20439980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 20449980d744SMichal Hocko * temporary page to workaround the nasty free_huge_page 20459980d744SMichal Hocko * codeflow 20469980d744SMichal Hocko */ 20479980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 20489157c311SMike Kravetz SetHPageTemporary(page); 2049db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 20509980d744SMichal Hocko put_page(page); 20512bf753e6SKai Shen return NULL; 20529980d744SMichal Hocko } else { 20539980d744SMichal Hocko h->surplus_huge_pages++; 20544704dea3SMichal Hocko h->surplus_huge_pages_node[page_to_nid(page)]++; 20557893d1d5SAdam Litke } 20569980d744SMichal Hocko 20579980d744SMichal Hocko out_unlock: 2058db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 20597893d1d5SAdam Litke 20607893d1d5SAdam Litke return page; 20617893d1d5SAdam Litke } 20627893d1d5SAdam Litke 2063bbe88753SJoonsoo Kim static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, 2064ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 2065ab5ac90aSMichal Hocko { 2066ab5ac90aSMichal Hocko struct page *page; 2067ab5ac90aSMichal Hocko 2068ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 2069ab5ac90aSMichal Hocko return NULL; 2070ab5ac90aSMichal Hocko 2071f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 2072ab5ac90aSMichal Hocko if (!page) 2073ab5ac90aSMichal Hocko return NULL; 2074ab5ac90aSMichal Hocko 2075ab5ac90aSMichal Hocko /* 2076ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 2077ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 2078ab5ac90aSMichal Hocko */ 20799157c311SMike Kravetz SetHPageTemporary(page); 2080ab5ac90aSMichal Hocko 2081ab5ac90aSMichal Hocko return page; 2082ab5ac90aSMichal Hocko } 2083ab5ac90aSMichal Hocko 2084e4e574b7SAdam Litke /* 2085099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 2086099730d6SDave Hansen */ 2087e0ec90eeSDave Hansen static 20880c397daeSMichal Hocko struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, 2089099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 2090099730d6SDave Hansen { 2091aaf14e40SMichal Hocko struct page *page; 2092aaf14e40SMichal Hocko struct mempolicy *mpol; 2093aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 2094aaf14e40SMichal Hocko int nid; 2095aaf14e40SMichal Hocko nodemask_t *nodemask; 2096aaf14e40SMichal Hocko 2097aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 20980c397daeSMichal Hocko page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); 2099aaf14e40SMichal Hocko mpol_cond_put(mpol); 2100aaf14e40SMichal Hocko 2101aaf14e40SMichal Hocko return page; 2102099730d6SDave Hansen } 2103099730d6SDave Hansen 2104ab5ac90aSMichal Hocko /* page migration callback function */ 21053e59fcb0SMichal Hocko struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, 2106d92bbc27SJoonsoo Kim nodemask_t *nmask, gfp_t gfp_mask) 21074db9b2efSMichal Hocko { 2108db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 21094db9b2efSMichal Hocko if (h->free_huge_pages - h->resv_huge_pages > 0) { 21103e59fcb0SMichal Hocko struct page *page; 21113e59fcb0SMichal Hocko 21123e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); 21133e59fcb0SMichal Hocko if (page) { 2114db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 21153e59fcb0SMichal Hocko return page; 21164db9b2efSMichal Hocko } 21174db9b2efSMichal Hocko } 2118db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 21194db9b2efSMichal Hocko 21200c397daeSMichal Hocko return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); 21214db9b2efSMichal Hocko } 21224db9b2efSMichal Hocko 2123ebd63723SMichal Hocko /* mempolicy aware migration callback */ 2124389c8178SMichal Hocko struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, 2125389c8178SMichal Hocko unsigned long address) 2126ebd63723SMichal Hocko { 2127ebd63723SMichal Hocko struct mempolicy *mpol; 2128ebd63723SMichal Hocko nodemask_t *nodemask; 2129ebd63723SMichal Hocko struct page *page; 2130ebd63723SMichal Hocko gfp_t gfp_mask; 2131ebd63723SMichal Hocko int node; 2132ebd63723SMichal Hocko 2133ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 2134ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 2135d92bbc27SJoonsoo Kim page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); 2136ebd63723SMichal Hocko mpol_cond_put(mpol); 2137ebd63723SMichal Hocko 2138ebd63723SMichal Hocko return page; 2139ebd63723SMichal Hocko } 2140ebd63723SMichal Hocko 2141bf50bab2SNaoya Horiguchi /* 214225985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 2143e4e574b7SAdam Litke * of size 'delta'. 2144e4e574b7SAdam Litke */ 21450a4f3d1bSLiu Xiang static int gather_surplus_pages(struct hstate *h, long delta) 21461b2a1e7bSJules Irenge __must_hold(&hugetlb_lock) 2147e4e574b7SAdam Litke { 2148e4e574b7SAdam Litke struct list_head surplus_list; 2149e4e574b7SAdam Litke struct page *page, *tmp; 21500a4f3d1bSLiu Xiang int ret; 21510a4f3d1bSLiu Xiang long i; 21520a4f3d1bSLiu Xiang long needed, allocated; 215328073b02SHillf Danton bool alloc_ok = true; 2154e4e574b7SAdam Litke 21559487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2156a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 2157ac09b3a1SAdam Litke if (needed <= 0) { 2158a5516438SAndi Kleen h->resv_huge_pages += delta; 2159e4e574b7SAdam Litke return 0; 2160ac09b3a1SAdam Litke } 2161e4e574b7SAdam Litke 2162e4e574b7SAdam Litke allocated = 0; 2163e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 2164e4e574b7SAdam Litke 2165e4e574b7SAdam Litke ret = -ENOMEM; 2166e4e574b7SAdam Litke retry: 2167db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2168e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 21690c397daeSMichal Hocko page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), 2170aaf14e40SMichal Hocko NUMA_NO_NODE, NULL); 217128073b02SHillf Danton if (!page) { 217228073b02SHillf Danton alloc_ok = false; 217328073b02SHillf Danton break; 217428073b02SHillf Danton } 2175e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 217669ed779aSDavid Rientjes cond_resched(); 2177e4e574b7SAdam Litke } 217828073b02SHillf Danton allocated += i; 2179e4e574b7SAdam Litke 2180e4e574b7SAdam Litke /* 2181e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 2182e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 2183e4e574b7SAdam Litke */ 2184db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2185a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 2186a5516438SAndi Kleen (h->free_huge_pages + allocated); 218728073b02SHillf Danton if (needed > 0) { 218828073b02SHillf Danton if (alloc_ok) 2189e4e574b7SAdam Litke goto retry; 219028073b02SHillf Danton /* 219128073b02SHillf Danton * We were not able to allocate enough pages to 219228073b02SHillf Danton * satisfy the entire reservation so we free what 219328073b02SHillf Danton * we've allocated so far. 219428073b02SHillf Danton */ 219528073b02SHillf Danton goto free; 219628073b02SHillf Danton } 2197e4e574b7SAdam Litke /* 2198e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 219925985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 2200e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 2201ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 2202ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 2203ac09b3a1SAdam Litke * before they are reserved. 2204e4e574b7SAdam Litke */ 2205e4e574b7SAdam Litke needed += allocated; 2206a5516438SAndi Kleen h->resv_huge_pages += delta; 2207e4e574b7SAdam Litke ret = 0; 2208a9869b83SNaoya Horiguchi 220919fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 221019fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 2211e558464bSMuchun Song int zeroed; 2212e558464bSMuchun Song 221319fc3f0aSAdam Litke if ((--needed) < 0) 221419fc3f0aSAdam Litke break; 2215a9869b83SNaoya Horiguchi /* 2216a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 2217a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 2218a9869b83SNaoya Horiguchi */ 2219e558464bSMuchun Song zeroed = put_page_testzero(page); 2220e558464bSMuchun Song VM_BUG_ON_PAGE(!zeroed, page); 2221a5516438SAndi Kleen enqueue_huge_page(h, page); 222219fc3f0aSAdam Litke } 222328073b02SHillf Danton free: 2224db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 222519fc3f0aSAdam Litke 222619fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 2227c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 2228a9869b83SNaoya Horiguchi put_page(page); 2229db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2230e4e574b7SAdam Litke 2231e4e574b7SAdam Litke return ret; 2232e4e574b7SAdam Litke } 2233e4e574b7SAdam Litke 2234e4e574b7SAdam Litke /* 2235e5bbc8a6SMike Kravetz * This routine has two main purposes: 2236e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 2237e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 2238e5bbc8a6SMike Kravetz * to the associated reservation map. 2239e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 2240e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 2241e4e574b7SAdam Litke */ 2242a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 2243a5516438SAndi Kleen unsigned long unused_resv_pages) 2244e4e574b7SAdam Litke { 2245e4e574b7SAdam Litke unsigned long nr_pages; 224610c6ec49SMike Kravetz struct page *page; 224710c6ec49SMike Kravetz LIST_HEAD(page_list); 224810c6ec49SMike Kravetz 22499487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 225010c6ec49SMike Kravetz /* Uncommit the reservation */ 225110c6ec49SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 2252e4e574b7SAdam Litke 2253aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 2254bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2255e5bbc8a6SMike Kravetz goto out; 2256aa888a74SAndi Kleen 2257e5bbc8a6SMike Kravetz /* 2258e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 2259e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 2260e5bbc8a6SMike Kravetz */ 2261a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 2262e4e574b7SAdam Litke 2263685f3457SLee Schermerhorn /* 2264685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 22659b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 22669b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 22679b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 226810c6ec49SMike Kravetz * remove_pool_huge_page() will balance the freed pages across the 22699b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 2270685f3457SLee Schermerhorn */ 2271685f3457SLee Schermerhorn while (nr_pages--) { 227210c6ec49SMike Kravetz page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1); 227310c6ec49SMike Kravetz if (!page) 2274e5bbc8a6SMike Kravetz goto out; 227510c6ec49SMike Kravetz 227610c6ec49SMike Kravetz list_add(&page->lru, &page_list); 2277e4e574b7SAdam Litke } 2278e5bbc8a6SMike Kravetz 2279e5bbc8a6SMike Kravetz out: 2280db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 228110c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 2282db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2283e4e574b7SAdam Litke } 2284e4e574b7SAdam Litke 22855e911373SMike Kravetz 2286c37f9fb1SAndy Whitcroft /* 2287feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 22885e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 2289cf3ad20bSMike Kravetz * 2290cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 2291cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 2292cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 2293cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 2294cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 2295feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 2296feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 2297feba16e2SMike Kravetz * is called in such cases. 2298cf3ad20bSMike Kravetz * 2299cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 2300cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 2301cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 2302cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 2303cf3ad20bSMike Kravetz * take appropriate action. 230496b96a96SMike Kravetz * 230596b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 230696b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 230796b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 230896b96a96SMike Kravetz * reservation exists. 2309846be085SMike Kravetz * 2310846be085SMike Kravetz * vma_del_reservation is used in error paths where an entry in the reserve 2311846be085SMike Kravetz * map was created during huge page allocation and must be removed. It is to 2312846be085SMike Kravetz * be called after calling vma_needs_reservation to determine if a reservation 2313846be085SMike Kravetz * exists. 2314c37f9fb1SAndy Whitcroft */ 23155e911373SMike Kravetz enum vma_resv_mode { 23165e911373SMike Kravetz VMA_NEEDS_RESV, 23175e911373SMike Kravetz VMA_COMMIT_RESV, 2318feba16e2SMike Kravetz VMA_END_RESV, 231996b96a96SMike Kravetz VMA_ADD_RESV, 2320846be085SMike Kravetz VMA_DEL_RESV, 23215e911373SMike Kravetz }; 2322cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 2323cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 23245e911373SMike Kravetz enum vma_resv_mode mode) 2325c37f9fb1SAndy Whitcroft { 23264e35f483SJoonsoo Kim struct resv_map *resv; 23274e35f483SJoonsoo Kim pgoff_t idx; 2328cf3ad20bSMike Kravetz long ret; 23290db9d74eSMina Almasry long dummy_out_regions_needed; 2330c37f9fb1SAndy Whitcroft 23314e35f483SJoonsoo Kim resv = vma_resv_map(vma); 23324e35f483SJoonsoo Kim if (!resv) 2333c37f9fb1SAndy Whitcroft return 1; 2334c37f9fb1SAndy Whitcroft 23354e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 23365e911373SMike Kravetz switch (mode) { 23375e911373SMike Kravetz case VMA_NEEDS_RESV: 23380db9d74eSMina Almasry ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); 23390db9d74eSMina Almasry /* We assume that vma_reservation_* routines always operate on 23400db9d74eSMina Almasry * 1 page, and that adding to resv map a 1 page entry can only 23410db9d74eSMina Almasry * ever require 1 region. 23420db9d74eSMina Almasry */ 23430db9d74eSMina Almasry VM_BUG_ON(dummy_out_regions_needed != 1); 23445e911373SMike Kravetz break; 23455e911373SMike Kravetz case VMA_COMMIT_RESV: 2346075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 23470db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 23480db9d74eSMina Almasry VM_BUG_ON(ret < 0); 23495e911373SMike Kravetz break; 2350feba16e2SMike Kravetz case VMA_END_RESV: 23510db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 23525e911373SMike Kravetz ret = 0; 23535e911373SMike Kravetz break; 235496b96a96SMike Kravetz case VMA_ADD_RESV: 23550db9d74eSMina Almasry if (vma->vm_flags & VM_MAYSHARE) { 2356075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 23570db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 23580db9d74eSMina Almasry VM_BUG_ON(ret < 0); 23590db9d74eSMina Almasry } else { 23600db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 236196b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 236296b96a96SMike Kravetz } 236396b96a96SMike Kravetz break; 2364846be085SMike Kravetz case VMA_DEL_RESV: 2365846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 2366846be085SMike Kravetz region_abort(resv, idx, idx + 1, 1); 2367846be085SMike Kravetz ret = region_del(resv, idx, idx + 1); 2368846be085SMike Kravetz } else { 2369846be085SMike Kravetz ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 2370846be085SMike Kravetz /* region_add calls of range 1 should never fail. */ 2371846be085SMike Kravetz VM_BUG_ON(ret < 0); 2372846be085SMike Kravetz } 2373846be085SMike Kravetz break; 23745e911373SMike Kravetz default: 23755e911373SMike Kravetz BUG(); 23765e911373SMike Kravetz } 237784afd99bSAndy Whitcroft 2378846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV) 2379cf3ad20bSMike Kravetz return ret; 238067961f9dSMike Kravetz /* 2381bf3d12b9SMiaohe Lin * We know private mapping must have HPAGE_RESV_OWNER set. 2382bf3d12b9SMiaohe Lin * 238367961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 238467961f9dSMike Kravetz * However, a file associated with mapping could have been 238567961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 238667961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 238767961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 238867961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 238967961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 239067961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 239167961f9dSMike Kravetz * reservation has already been consumed. As a result, the 239267961f9dSMike Kravetz * return value of this routine is the opposite of the 239367961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 239467961f9dSMike Kravetz */ 2395bf3d12b9SMiaohe Lin if (ret > 0) 239667961f9dSMike Kravetz return 0; 2397bf3d12b9SMiaohe Lin if (ret == 0) 239867961f9dSMike Kravetz return 1; 2399bf3d12b9SMiaohe Lin return ret; 240084afd99bSAndy Whitcroft } 2401cf3ad20bSMike Kravetz 2402cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 2403a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 2404c37f9fb1SAndy Whitcroft { 24055e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 2406cf3ad20bSMike Kravetz } 2407c37f9fb1SAndy Whitcroft 2408cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 2409cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 2410cf3ad20bSMike Kravetz { 24115e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 24125e911373SMike Kravetz } 24135e911373SMike Kravetz 2414feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 24155e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 24165e911373SMike Kravetz { 2417feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 2418c37f9fb1SAndy Whitcroft } 2419c37f9fb1SAndy Whitcroft 242096b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 242196b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 242296b96a96SMike Kravetz { 242396b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 242496b96a96SMike Kravetz } 242596b96a96SMike Kravetz 2426846be085SMike Kravetz static long vma_del_reservation(struct hstate *h, 2427846be085SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 242896b96a96SMike Kravetz { 2429846be085SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV); 2430846be085SMike Kravetz } 2431846be085SMike Kravetz 2432846be085SMike Kravetz /* 2433846be085SMike Kravetz * This routine is called to restore reservation information on error paths. 2434846be085SMike Kravetz * It should ONLY be called for pages allocated via alloc_huge_page(), and 2435846be085SMike Kravetz * the hugetlb mutex should remain held when calling this routine. 2436846be085SMike Kravetz * 2437846be085SMike Kravetz * It handles two specific cases: 2438846be085SMike Kravetz * 1) A reservation was in place and the page consumed the reservation. 2439846be085SMike Kravetz * HPageRestoreReserve is set in the page. 2440846be085SMike Kravetz * 2) No reservation was in place for the page, so HPageRestoreReserve is 2441846be085SMike Kravetz * not set. However, alloc_huge_page always updates the reserve map. 2442846be085SMike Kravetz * 2443846be085SMike Kravetz * In case 1, free_huge_page later in the error path will increment the 2444846be085SMike Kravetz * global reserve count. But, free_huge_page does not have enough context 2445846be085SMike Kravetz * to adjust the reservation map. This case deals primarily with private 2446846be085SMike Kravetz * mappings. Adjust the reserve map here to be consistent with global 2447846be085SMike Kravetz * reserve count adjustments to be made by free_huge_page. Make sure the 2448846be085SMike Kravetz * reserve map indicates there is a reservation present. 2449846be085SMike Kravetz * 2450846be085SMike Kravetz * In case 2, simply undo reserve map modifications done by alloc_huge_page. 2451846be085SMike Kravetz */ 2452846be085SMike Kravetz void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma, 2453846be085SMike Kravetz unsigned long address, struct page *page) 2454846be085SMike Kravetz { 245596b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 245696b96a96SMike Kravetz 2457846be085SMike Kravetz if (HPageRestoreReserve(page)) { 2458846be085SMike Kravetz if (unlikely(rc < 0)) 245996b96a96SMike Kravetz /* 246096b96a96SMike Kravetz * Rare out of memory condition in reserve map 2461d6995da3SMike Kravetz * manipulation. Clear HPageRestoreReserve so that 246296b96a96SMike Kravetz * global reserve count will not be incremented 246396b96a96SMike Kravetz * by free_huge_page. This will make it appear 246496b96a96SMike Kravetz * as though the reservation for this page was 246596b96a96SMike Kravetz * consumed. This may prevent the task from 246696b96a96SMike Kravetz * faulting in the page at a later time. This 246796b96a96SMike Kravetz * is better than inconsistent global huge page 246896b96a96SMike Kravetz * accounting of reserve counts. 246996b96a96SMike Kravetz */ 2470d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 2471846be085SMike Kravetz else if (rc) 2472846be085SMike Kravetz (void)vma_add_reservation(h, vma, address); 2473846be085SMike Kravetz else 2474846be085SMike Kravetz vma_end_reservation(h, vma, address); 2475846be085SMike Kravetz } else { 2476846be085SMike Kravetz if (!rc) { 247796b96a96SMike Kravetz /* 2478846be085SMike Kravetz * This indicates there is an entry in the reserve map 2479846be085SMike Kravetz * added by alloc_huge_page. We know it was added 2480846be085SMike Kravetz * before the alloc_huge_page call, otherwise 2481846be085SMike Kravetz * HPageRestoreReserve would be set on the page. 2482846be085SMike Kravetz * Remove the entry so that a subsequent allocation 2483846be085SMike Kravetz * does not consume a reservation. 248496b96a96SMike Kravetz */ 2485846be085SMike Kravetz rc = vma_del_reservation(h, vma, address); 2486846be085SMike Kravetz if (rc < 0) 2487846be085SMike Kravetz /* 2488846be085SMike Kravetz * VERY rare out of memory condition. Since 2489846be085SMike Kravetz * we can not delete the entry, set 2490846be085SMike Kravetz * HPageRestoreReserve so that the reserve 2491846be085SMike Kravetz * count will be incremented when the page 2492846be085SMike Kravetz * is freed. This reserve will be consumed 2493846be085SMike Kravetz * on a subsequent allocation. 2494846be085SMike Kravetz */ 2495846be085SMike Kravetz SetHPageRestoreReserve(page); 2496846be085SMike Kravetz } else if (rc < 0) { 2497846be085SMike Kravetz /* 2498846be085SMike Kravetz * Rare out of memory condition from 2499846be085SMike Kravetz * vma_needs_reservation call. Memory allocation is 2500846be085SMike Kravetz * only attempted if a new entry is needed. Therefore, 2501846be085SMike Kravetz * this implies there is not an entry in the 2502846be085SMike Kravetz * reserve map. 2503846be085SMike Kravetz * 2504846be085SMike Kravetz * For shared mappings, no entry in the map indicates 2505846be085SMike Kravetz * no reservation. We are done. 2506846be085SMike Kravetz */ 2507846be085SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 2508846be085SMike Kravetz /* 2509846be085SMike Kravetz * For private mappings, no entry indicates 2510846be085SMike Kravetz * a reservation is present. Since we can 2511846be085SMike Kravetz * not add an entry, set SetHPageRestoreReserve 2512846be085SMike Kravetz * on the page so reserve count will be 2513846be085SMike Kravetz * incremented when freed. This reserve will 2514846be085SMike Kravetz * be consumed on a subsequent allocation. 2515846be085SMike Kravetz */ 2516846be085SMike Kravetz SetHPageRestoreReserve(page); 251796b96a96SMike Kravetz } else 2518846be085SMike Kravetz /* 2519846be085SMike Kravetz * No reservation present, do nothing 2520846be085SMike Kravetz */ 252196b96a96SMike Kravetz vma_end_reservation(h, vma, address); 252296b96a96SMike Kravetz } 252396b96a96SMike Kravetz } 252496b96a96SMike Kravetz 2525369fa227SOscar Salvador /* 2526369fa227SOscar Salvador * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one 2527369fa227SOscar Salvador * @h: struct hstate old page belongs to 2528369fa227SOscar Salvador * @old_page: Old page to dissolve 2529ae37c7ffSOscar Salvador * @list: List to isolate the page in case we need to 2530369fa227SOscar Salvador * Returns 0 on success, otherwise negated error. 2531369fa227SOscar Salvador */ 2532ae37c7ffSOscar Salvador static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page, 2533ae37c7ffSOscar Salvador struct list_head *list) 2534369fa227SOscar Salvador { 2535369fa227SOscar Salvador gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2536369fa227SOscar Salvador int nid = page_to_nid(old_page); 2537369fa227SOscar Salvador struct page *new_page; 2538369fa227SOscar Salvador int ret = 0; 2539369fa227SOscar Salvador 2540369fa227SOscar Salvador /* 2541369fa227SOscar Salvador * Before dissolving the page, we need to allocate a new one for the 2542f41f2ed4SMuchun Song * pool to remain stable. Here, we allocate the page and 'prep' it 2543f41f2ed4SMuchun Song * by doing everything but actually updating counters and adding to 2544f41f2ed4SMuchun Song * the pool. This simplifies and let us do most of the processing 2545f41f2ed4SMuchun Song * under the lock. 2546369fa227SOscar Salvador */ 2547369fa227SOscar Salvador new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL); 2548369fa227SOscar Salvador if (!new_page) 2549369fa227SOscar Salvador return -ENOMEM; 2550f41f2ed4SMuchun Song __prep_new_huge_page(h, new_page); 2551369fa227SOscar Salvador 2552369fa227SOscar Salvador retry: 2553369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2554369fa227SOscar Salvador if (!PageHuge(old_page)) { 2555369fa227SOscar Salvador /* 2556369fa227SOscar Salvador * Freed from under us. Drop new_page too. 2557369fa227SOscar Salvador */ 2558369fa227SOscar Salvador goto free_new; 2559369fa227SOscar Salvador } else if (page_count(old_page)) { 2560369fa227SOscar Salvador /* 2561ae37c7ffSOscar Salvador * Someone has grabbed the page, try to isolate it here. 2562ae37c7ffSOscar Salvador * Fail with -EBUSY if not possible. 2563369fa227SOscar Salvador */ 2564ae37c7ffSOscar Salvador spin_unlock_irq(&hugetlb_lock); 2565ae37c7ffSOscar Salvador if (!isolate_huge_page(old_page, list)) 2566369fa227SOscar Salvador ret = -EBUSY; 2567ae37c7ffSOscar Salvador spin_lock_irq(&hugetlb_lock); 2568369fa227SOscar Salvador goto free_new; 2569369fa227SOscar Salvador } else if (!HPageFreed(old_page)) { 2570369fa227SOscar Salvador /* 2571369fa227SOscar Salvador * Page's refcount is 0 but it has not been enqueued in the 2572369fa227SOscar Salvador * freelist yet. Race window is small, so we can succeed here if 2573369fa227SOscar Salvador * we retry. 2574369fa227SOscar Salvador */ 2575369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2576369fa227SOscar Salvador cond_resched(); 2577369fa227SOscar Salvador goto retry; 2578369fa227SOscar Salvador } else { 2579369fa227SOscar Salvador /* 2580369fa227SOscar Salvador * Ok, old_page is still a genuine free hugepage. Remove it from 2581369fa227SOscar Salvador * the freelist and decrease the counters. These will be 2582369fa227SOscar Salvador * incremented again when calling __prep_account_new_huge_page() 2583369fa227SOscar Salvador * and enqueue_huge_page() for new_page. The counters will remain 2584369fa227SOscar Salvador * stable since this happens under the lock. 2585369fa227SOscar Salvador */ 2586369fa227SOscar Salvador remove_hugetlb_page(h, old_page, false); 2587369fa227SOscar Salvador 2588369fa227SOscar Salvador /* 2589369fa227SOscar Salvador * Reference count trick is needed because allocator gives us 2590369fa227SOscar Salvador * referenced page but the pool requires pages with 0 refcount. 2591369fa227SOscar Salvador */ 2592369fa227SOscar Salvador __prep_account_new_huge_page(h, nid); 2593369fa227SOscar Salvador page_ref_dec(new_page); 2594369fa227SOscar Salvador enqueue_huge_page(h, new_page); 2595369fa227SOscar Salvador 2596369fa227SOscar Salvador /* 2597369fa227SOscar Salvador * Pages have been replaced, we can safely free the old one. 2598369fa227SOscar Salvador */ 2599369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2600b65d4adbSMuchun Song update_and_free_page(h, old_page, false); 2601369fa227SOscar Salvador } 2602369fa227SOscar Salvador 2603369fa227SOscar Salvador return ret; 2604369fa227SOscar Salvador 2605369fa227SOscar Salvador free_new: 2606369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2607b65d4adbSMuchun Song update_and_free_page(h, new_page, false); 2608369fa227SOscar Salvador 2609369fa227SOscar Salvador return ret; 2610369fa227SOscar Salvador } 2611369fa227SOscar Salvador 2612ae37c7ffSOscar Salvador int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list) 2613369fa227SOscar Salvador { 2614369fa227SOscar Salvador struct hstate *h; 2615369fa227SOscar Salvador struct page *head; 2616ae37c7ffSOscar Salvador int ret = -EBUSY; 2617369fa227SOscar Salvador 2618369fa227SOscar Salvador /* 2619369fa227SOscar Salvador * The page might have been dissolved from under our feet, so make sure 2620369fa227SOscar Salvador * to carefully check the state under the lock. 2621369fa227SOscar Salvador * Return success when racing as if we dissolved the page ourselves. 2622369fa227SOscar Salvador */ 2623369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2624369fa227SOscar Salvador if (PageHuge(page)) { 2625369fa227SOscar Salvador head = compound_head(page); 2626369fa227SOscar Salvador h = page_hstate(head); 2627369fa227SOscar Salvador } else { 2628369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2629369fa227SOscar Salvador return 0; 2630369fa227SOscar Salvador } 2631369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2632369fa227SOscar Salvador 2633369fa227SOscar Salvador /* 2634369fa227SOscar Salvador * Fence off gigantic pages as there is a cyclic dependency between 2635369fa227SOscar Salvador * alloc_contig_range and them. Return -ENOMEM as this has the effect 2636369fa227SOscar Salvador * of bailing out right away without further retrying. 2637369fa227SOscar Salvador */ 2638369fa227SOscar Salvador if (hstate_is_gigantic(h)) 2639369fa227SOscar Salvador return -ENOMEM; 2640369fa227SOscar Salvador 2641ae37c7ffSOscar Salvador if (page_count(head) && isolate_huge_page(head, list)) 2642ae37c7ffSOscar Salvador ret = 0; 2643ae37c7ffSOscar Salvador else if (!page_count(head)) 2644ae37c7ffSOscar Salvador ret = alloc_and_dissolve_huge_page(h, head, list); 2645ae37c7ffSOscar Salvador 2646ae37c7ffSOscar Salvador return ret; 2647369fa227SOscar Salvador } 2648369fa227SOscar Salvador 264970c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 265004f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 2651348ea204SAdam Litke { 265290481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 2653a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2654348ea204SAdam Litke struct page *page; 2655d85f69b0SMike Kravetz long map_chg, map_commit; 2656d85f69b0SMike Kravetz long gbl_chg; 26576d76dcf4SAneesh Kumar K.V int ret, idx; 26586d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 265908cf9fafSMina Almasry bool deferred_reserve; 26602fc39cecSAdam Litke 26616d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 2662a1e78772SMel Gorman /* 2663d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 2664d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 2665d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 2666a1e78772SMel Gorman */ 2667d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 2668d85f69b0SMike Kravetz if (map_chg < 0) 266976dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 2670d85f69b0SMike Kravetz 2671d85f69b0SMike Kravetz /* 2672d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 2673d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 2674d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 2675d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 2676d85f69b0SMike Kravetz * checked against any subpool limit. 2677d85f69b0SMike Kravetz */ 2678d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 2679d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 2680d85f69b0SMike Kravetz if (gbl_chg < 0) { 2681feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 268276dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 26835e911373SMike Kravetz } 268490d8b7e6SAdam Litke 2685d85f69b0SMike Kravetz /* 2686d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 2687d85f69b0SMike Kravetz * map, there could be reservations associated with the 2688d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 2689d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 2690d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 2691d85f69b0SMike Kravetz * the subpool reservations. 2692d85f69b0SMike Kravetz */ 2693d85f69b0SMike Kravetz if (avoid_reserve) 2694d85f69b0SMike Kravetz gbl_chg = 1; 2695d85f69b0SMike Kravetz } 2696d85f69b0SMike Kravetz 269708cf9fafSMina Almasry /* If this allocation is not consuming a reservation, charge it now. 269808cf9fafSMina Almasry */ 26996501fe5fSMiaohe Lin deferred_reserve = map_chg || avoid_reserve; 270008cf9fafSMina Almasry if (deferred_reserve) { 270108cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 270208cf9fafSMina Almasry idx, pages_per_huge_page(h), &h_cg); 27038f34af6fSJianyu Zhan if (ret) 27048f34af6fSJianyu Zhan goto out_subpool_put; 270508cf9fafSMina Almasry } 270608cf9fafSMina Almasry 270708cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 270808cf9fafSMina Almasry if (ret) 270908cf9fafSMina Almasry goto out_uncharge_cgroup_reservation; 27108f34af6fSJianyu Zhan 2711db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2712d85f69b0SMike Kravetz /* 2713d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 2714d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 2715d85f69b0SMike Kravetz * a reservation exists for the allocation. 2716d85f69b0SMike Kravetz */ 2717d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 271881a6fcaeSJoonsoo Kim if (!page) { 2719db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 27200c397daeSMichal Hocko page = alloc_buddy_huge_page_with_mpol(h, vma, addr); 27218f34af6fSJianyu Zhan if (!page) 27228f34af6fSJianyu Zhan goto out_uncharge_cgroup; 2723a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 2724d6995da3SMike Kravetz SetHPageRestoreReserve(page); 2725a88c7695SNaoya Horiguchi h->resv_huge_pages--; 2726a88c7695SNaoya Horiguchi } 2727db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 272815a8d68eSWei Yang list_add(&page->lru, &h->hugepage_activelist); 272981a6fcaeSJoonsoo Kim /* Fall through */ 2730a1e78772SMel Gorman } 273181a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 273208cf9fafSMina Almasry /* If allocation is not consuming a reservation, also store the 273308cf9fafSMina Almasry * hugetlb_cgroup pointer on the page. 273408cf9fafSMina Almasry */ 273508cf9fafSMina Almasry if (deferred_reserve) { 273608cf9fafSMina Almasry hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), 273708cf9fafSMina Almasry h_cg, page); 273808cf9fafSMina Almasry } 273908cf9fafSMina Almasry 2740db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2741a1e78772SMel Gorman 2742d6995da3SMike Kravetz hugetlb_set_page_subpool(page, spool); 2743a1e78772SMel Gorman 2744d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 2745d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 274633039678SMike Kravetz /* 274733039678SMike Kravetz * The page was added to the reservation map between 274833039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 274933039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 275033039678SMike Kravetz * Adjust for the subpool count incremented above AND 275133039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 275233039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 275333039678SMike Kravetz * no longer applies. 275433039678SMike Kravetz */ 275533039678SMike Kravetz long rsv_adjust; 275633039678SMike Kravetz 275733039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 275833039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 275979aa925bSMike Kravetz if (deferred_reserve) 276079aa925bSMike Kravetz hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 276179aa925bSMike Kravetz pages_per_huge_page(h), page); 276233039678SMike Kravetz } 27637893d1d5SAdam Litke return page; 27648f34af6fSJianyu Zhan 27658f34af6fSJianyu Zhan out_uncharge_cgroup: 27668f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 276708cf9fafSMina Almasry out_uncharge_cgroup_reservation: 276808cf9fafSMina Almasry if (deferred_reserve) 276908cf9fafSMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), 277008cf9fafSMina Almasry h_cg); 27718f34af6fSJianyu Zhan out_subpool_put: 2772d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 27738f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 2774feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 27758f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 2776b45b5bd6SDavid Gibson } 2777b45b5bd6SDavid Gibson 2778e24a1307SAneesh Kumar K.V int alloc_bootmem_huge_page(struct hstate *h) 2779e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 2780e24a1307SAneesh Kumar K.V int __alloc_bootmem_huge_page(struct hstate *h) 2781aa888a74SAndi Kleen { 2782aa888a74SAndi Kleen struct huge_bootmem_page *m; 2783b2261026SJoonsoo Kim int nr_nodes, node; 2784aa888a74SAndi Kleen 2785b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 2786aa888a74SAndi Kleen void *addr; 2787aa888a74SAndi Kleen 2788eb31d559SMike Rapoport addr = memblock_alloc_try_nid_raw( 27898b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 279097ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 2791aa888a74SAndi Kleen if (addr) { 2792aa888a74SAndi Kleen /* 2793aa888a74SAndi Kleen * Use the beginning of the huge page to store the 2794aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 2795aa888a74SAndi Kleen * puts them into the mem_map). 2796aa888a74SAndi Kleen */ 2797aa888a74SAndi Kleen m = addr; 2798aa888a74SAndi Kleen goto found; 2799aa888a74SAndi Kleen } 2800aa888a74SAndi Kleen } 2801aa888a74SAndi Kleen return 0; 2802aa888a74SAndi Kleen 2803aa888a74SAndi Kleen found: 2804df994eadSLuiz Capitulino BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); 2805aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 2806330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 2807aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 2808aa888a74SAndi Kleen m->hstate = h; 2809aa888a74SAndi Kleen return 1; 2810aa888a74SAndi Kleen } 2811aa888a74SAndi Kleen 281248b8d744SMike Kravetz /* 281348b8d744SMike Kravetz * Put bootmem huge pages into the standard lists after mem_map is up. 281448b8d744SMike Kravetz * Note: This only applies to gigantic (order > MAX_ORDER) pages. 281548b8d744SMike Kravetz */ 2816aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 2817aa888a74SAndi Kleen { 2818aa888a74SAndi Kleen struct huge_bootmem_page *m; 2819aa888a74SAndi Kleen 2820aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 282140d18ebfSMike Kravetz struct page *page = virt_to_page(m); 2822aa888a74SAndi Kleen struct hstate *h = m->hstate; 2823ee8f248dSBecky Bruce 282448b8d744SMike Kravetz VM_BUG_ON(!hstate_is_gigantic(h)); 2825aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 2826*7118fc29SMike Kravetz if (prep_compound_gigantic_page(page, huge_page_order(h))) { 2827ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 2828aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 2829*7118fc29SMike Kravetz put_page(page); /* add to the hugepage allocator */ 2830*7118fc29SMike Kravetz } else { 2831*7118fc29SMike Kravetz free_gigantic_page(page, huge_page_order(h)); 2832*7118fc29SMike Kravetz pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n"); 2833*7118fc29SMike Kravetz } 2834af0fb9dfSMichal Hocko 2835b0320c7bSRafael Aquini /* 283648b8d744SMike Kravetz * We need to restore the 'stolen' pages to totalram_pages 283748b8d744SMike Kravetz * in order to fix confusing memory reports from free(1) and 283848b8d744SMike Kravetz * other side-effects, like CommitLimit going negative. 2839b0320c7bSRafael Aquini */ 2840c78a7f36SMiaohe Lin adjust_managed_page_count(page, pages_per_huge_page(h)); 2841520495feSCannon Matthews cond_resched(); 2842aa888a74SAndi Kleen } 2843aa888a74SAndi Kleen } 2844aa888a74SAndi Kleen 28458faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 28461da177e4SLinus Torvalds { 28471da177e4SLinus Torvalds unsigned long i; 2848f60858f9SMike Kravetz nodemask_t *node_alloc_noretry; 2849f60858f9SMike Kravetz 2850f60858f9SMike Kravetz if (!hstate_is_gigantic(h)) { 2851f60858f9SMike Kravetz /* 2852f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 2853f60858f9SMike Kravetz * Ignore errors as lower level routines can deal with 2854f60858f9SMike Kravetz * node_alloc_noretry == NULL. If this kmalloc fails at boot 2855f60858f9SMike Kravetz * time, we are likely in bigger trouble. 2856f60858f9SMike Kravetz */ 2857f60858f9SMike Kravetz node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), 2858f60858f9SMike Kravetz GFP_KERNEL); 2859f60858f9SMike Kravetz } else { 2860f60858f9SMike Kravetz /* allocations done at boot time */ 2861f60858f9SMike Kravetz node_alloc_noretry = NULL; 2862f60858f9SMike Kravetz } 2863f60858f9SMike Kravetz 2864f60858f9SMike Kravetz /* bit mask controlling how hard we retry per-node allocations */ 2865f60858f9SMike Kravetz if (node_alloc_noretry) 2866f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 28671da177e4SLinus Torvalds 2868e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 2869bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 2870dbda8feaSBarry Song if (hugetlb_cma_size) { 2871cf11e85fSRoman Gushchin pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); 28727ecc9565SChen Wandun goto free; 2873cf11e85fSRoman Gushchin } 2874aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 2875aa888a74SAndi Kleen break; 28760c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 2877f60858f9SMike Kravetz &node_states[N_MEMORY], 2878f60858f9SMike Kravetz node_alloc_noretry)) 28791da177e4SLinus Torvalds break; 288069ed779aSDavid Rientjes cond_resched(); 28811da177e4SLinus Torvalds } 2882d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 2883d715cf80SLiam R. Howlett char buf[32]; 2884d715cf80SLiam R. Howlett 2885c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2886d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 2887d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 28888faa8b07SAndi Kleen h->max_huge_pages = i; 2889e5ff2159SAndi Kleen } 28907ecc9565SChen Wandun free: 2891f60858f9SMike Kravetz kfree(node_alloc_noretry); 2892d715cf80SLiam R. Howlett } 2893e5ff2159SAndi Kleen 2894e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 2895e5ff2159SAndi Kleen { 2896e5ff2159SAndi Kleen struct hstate *h; 2897e5ff2159SAndi Kleen 2898e5ff2159SAndi Kleen for_each_hstate(h) { 2899641844f5SNaoya Horiguchi if (minimum_order > huge_page_order(h)) 2900641844f5SNaoya Horiguchi minimum_order = huge_page_order(h); 2901641844f5SNaoya Horiguchi 29028faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 2903bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 29048faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 2905e5ff2159SAndi Kleen } 2906641844f5SNaoya Horiguchi VM_BUG_ON(minimum_order == UINT_MAX); 2907e5ff2159SAndi Kleen } 2908e5ff2159SAndi Kleen 2909e5ff2159SAndi Kleen static void __init report_hugepages(void) 2910e5ff2159SAndi Kleen { 2911e5ff2159SAndi Kleen struct hstate *h; 2912e5ff2159SAndi Kleen 2913e5ff2159SAndi Kleen for_each_hstate(h) { 29144abd32dbSAndi Kleen char buf[32]; 2915c6247f72SMatthew Wilcox 2916c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2917ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 2918c6247f72SMatthew Wilcox buf, h->free_huge_pages); 2919e5ff2159SAndi Kleen } 2920e5ff2159SAndi Kleen } 2921e5ff2159SAndi Kleen 29221da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 29236ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 29246ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 29251da177e4SLinus Torvalds { 29264415cc8dSChristoph Lameter int i; 29271121828aSMike Kravetz LIST_HEAD(page_list); 29284415cc8dSChristoph Lameter 29299487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2930bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2931aa888a74SAndi Kleen return; 2932aa888a74SAndi Kleen 29331121828aSMike Kravetz /* 29341121828aSMike Kravetz * Collect pages to be freed on a list, and free after dropping lock 29351121828aSMike Kravetz */ 29366ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 293710c6ec49SMike Kravetz struct page *page, *next; 2938a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 2939a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 2940a5516438SAndi Kleen if (count >= h->nr_huge_pages) 29411121828aSMike Kravetz goto out; 29421da177e4SLinus Torvalds if (PageHighMem(page)) 29431da177e4SLinus Torvalds continue; 29446eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 29451121828aSMike Kravetz list_add(&page->lru, &page_list); 29461121828aSMike Kravetz } 29471121828aSMike Kravetz } 29481121828aSMike Kravetz 29491121828aSMike Kravetz out: 2950db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 295110c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 2952db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 29531da177e4SLinus Torvalds } 29541da177e4SLinus Torvalds #else 29556ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 29566ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 29571da177e4SLinus Torvalds { 29581da177e4SLinus Torvalds } 29591da177e4SLinus Torvalds #endif 29601da177e4SLinus Torvalds 296120a0307cSWu Fengguang /* 296220a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 296320a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 296420a0307cSWu Fengguang * Returns 1 if an adjustment was made. 296520a0307cSWu Fengguang */ 29666ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 29676ae11b27SLee Schermerhorn int delta) 296820a0307cSWu Fengguang { 2969b2261026SJoonsoo Kim int nr_nodes, node; 297020a0307cSWu Fengguang 29719487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 297220a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 297320a0307cSWu Fengguang 2974e8c5c824SLee Schermerhorn if (delta < 0) { 2975b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 2976b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 2977b2261026SJoonsoo Kim goto found; 2978b2261026SJoonsoo Kim } 2979b2261026SJoonsoo Kim } else { 2980b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2981b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 2982b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 2983b2261026SJoonsoo Kim goto found; 2984e8c5c824SLee Schermerhorn } 29859a76db09SLee Schermerhorn } 2986b2261026SJoonsoo Kim return 0; 298720a0307cSWu Fengguang 2988b2261026SJoonsoo Kim found: 298920a0307cSWu Fengguang h->surplus_huge_pages += delta; 2990b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 2991b2261026SJoonsoo Kim return 1; 299220a0307cSWu Fengguang } 299320a0307cSWu Fengguang 2994a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 2995fd875dcaSMike Kravetz static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 29966ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 29971da177e4SLinus Torvalds { 29987893d1d5SAdam Litke unsigned long min_count, ret; 299910c6ec49SMike Kravetz struct page *page; 300010c6ec49SMike Kravetz LIST_HEAD(page_list); 3001f60858f9SMike Kravetz NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); 3002f60858f9SMike Kravetz 3003f60858f9SMike Kravetz /* 3004f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3005f60858f9SMike Kravetz * If we can not allocate the bit mask, do not attempt to allocate 3006f60858f9SMike Kravetz * the requested huge pages. 3007f60858f9SMike Kravetz */ 3008f60858f9SMike Kravetz if (node_alloc_noretry) 3009f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 3010f60858f9SMike Kravetz else 3011f60858f9SMike Kravetz return -ENOMEM; 30121da177e4SLinus Torvalds 301329383967SMike Kravetz /* 301429383967SMike Kravetz * resize_lock mutex prevents concurrent adjustments to number of 301529383967SMike Kravetz * pages in hstate via the proc/sysfs interfaces. 301629383967SMike Kravetz */ 301729383967SMike Kravetz mutex_lock(&h->resize_lock); 3018b65d4adbSMuchun Song flush_free_hpage_work(h); 3019db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 30204eb0716eSAlexandre Ghiti 30214eb0716eSAlexandre Ghiti /* 3022fd875dcaSMike Kravetz * Check for a node specific request. 3023fd875dcaSMike Kravetz * Changing node specific huge page count may require a corresponding 3024fd875dcaSMike Kravetz * change to the global count. In any case, the passed node mask 3025fd875dcaSMike Kravetz * (nodes_allowed) will restrict alloc/free to the specified node. 3026fd875dcaSMike Kravetz */ 3027fd875dcaSMike Kravetz if (nid != NUMA_NO_NODE) { 3028fd875dcaSMike Kravetz unsigned long old_count = count; 3029fd875dcaSMike Kravetz 3030fd875dcaSMike Kravetz count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 3031fd875dcaSMike Kravetz /* 3032fd875dcaSMike Kravetz * User may have specified a large count value which caused the 3033fd875dcaSMike Kravetz * above calculation to overflow. In this case, they wanted 3034fd875dcaSMike Kravetz * to allocate as many huge pages as possible. Set count to 3035fd875dcaSMike Kravetz * largest possible value to align with their intention. 3036fd875dcaSMike Kravetz */ 3037fd875dcaSMike Kravetz if (count < old_count) 3038fd875dcaSMike Kravetz count = ULONG_MAX; 3039fd875dcaSMike Kravetz } 3040fd875dcaSMike Kravetz 3041fd875dcaSMike Kravetz /* 30424eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 30434eb0716eSAlexandre Ghiti * page range allocation. 30444eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 30454eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 30464eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 30474eb0716eSAlexandre Ghiti */ 30484eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 30494eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 3050db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 305129383967SMike Kravetz mutex_unlock(&h->resize_lock); 3052f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 30534eb0716eSAlexandre Ghiti return -EINVAL; 30544eb0716eSAlexandre Ghiti } 30554eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 30564eb0716eSAlexandre Ghiti } 3057aa888a74SAndi Kleen 30587893d1d5SAdam Litke /* 30597893d1d5SAdam Litke * Increase the pool size 30607893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 30617893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 3062d1c3fb1fSNishanth Aravamudan * 30630c397daeSMichal Hocko * We might race with alloc_surplus_huge_page() here and be unable 3064d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 3065d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 3066d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 3067d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 30687893d1d5SAdam Litke */ 3069a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 30706ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 30717893d1d5SAdam Litke break; 30727893d1d5SAdam Litke } 30737893d1d5SAdam Litke 3074a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 30757893d1d5SAdam Litke /* 30767893d1d5SAdam Litke * If this allocation races such that we no longer need the 30777893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 30787893d1d5SAdam Litke * and reducing the surplus. 30797893d1d5SAdam Litke */ 3080db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3081649920c6SJia He 3082649920c6SJia He /* yield cpu to avoid soft lockup */ 3083649920c6SJia He cond_resched(); 3084649920c6SJia He 3085f60858f9SMike Kravetz ret = alloc_pool_huge_page(h, nodes_allowed, 3086f60858f9SMike Kravetz node_alloc_noretry); 3087db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 30887893d1d5SAdam Litke if (!ret) 30897893d1d5SAdam Litke goto out; 30907893d1d5SAdam Litke 3091536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 3092536240f2SMel Gorman if (signal_pending(current)) 3093536240f2SMel Gorman goto out; 30947893d1d5SAdam Litke } 30957893d1d5SAdam Litke 30967893d1d5SAdam Litke /* 30977893d1d5SAdam Litke * Decrease the pool size 30987893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 30997893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 31007893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 31017893d1d5SAdam Litke * to the desired size as pages become free. 3102d1c3fb1fSNishanth Aravamudan * 3103d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 3104d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 3105d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 31060c397daeSMichal Hocko * alloc_surplus_huge_page() is checking the global counter, 3107d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 3108d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 3109d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 31107893d1d5SAdam Litke */ 3111a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 31126b0c880dSAdam Litke min_count = max(count, min_count); 31136ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 311410c6ec49SMike Kravetz 311510c6ec49SMike Kravetz /* 311610c6ec49SMike Kravetz * Collect pages to be removed on list without dropping lock 311710c6ec49SMike Kravetz */ 3118a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 311910c6ec49SMike Kravetz page = remove_pool_huge_page(h, nodes_allowed, 0); 312010c6ec49SMike Kravetz if (!page) 31211da177e4SLinus Torvalds break; 312210c6ec49SMike Kravetz 312310c6ec49SMike Kravetz list_add(&page->lru, &page_list); 31241da177e4SLinus Torvalds } 312510c6ec49SMike Kravetz /* free the pages after dropping lock */ 3126db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 312710c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3128b65d4adbSMuchun Song flush_free_hpage_work(h); 3129db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 313010c6ec49SMike Kravetz 3131a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 31326ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 31337893d1d5SAdam Litke break; 31347893d1d5SAdam Litke } 31357893d1d5SAdam Litke out: 31364eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 3137db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 313829383967SMike Kravetz mutex_unlock(&h->resize_lock); 31394eb0716eSAlexandre Ghiti 3140f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 3141f60858f9SMike Kravetz 31424eb0716eSAlexandre Ghiti return 0; 31431da177e4SLinus Torvalds } 31441da177e4SLinus Torvalds 3145a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 3146a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 3147a3437870SNishanth Aravamudan 3148a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 3149a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 3150a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 3151a3437870SNishanth Aravamudan 3152a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 3153a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 3154a3437870SNishanth Aravamudan 31559a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 31569a305230SLee Schermerhorn 31579a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 3158a3437870SNishanth Aravamudan { 3159a3437870SNishanth Aravamudan int i; 31609a305230SLee Schermerhorn 3161a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 31629a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 31639a305230SLee Schermerhorn if (nidp) 31649a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 3165a3437870SNishanth Aravamudan return &hstates[i]; 31669a305230SLee Schermerhorn } 31679a305230SLee Schermerhorn 31689a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 3169a3437870SNishanth Aravamudan } 3170a3437870SNishanth Aravamudan 317106808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 3172a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3173a3437870SNishanth Aravamudan { 31749a305230SLee Schermerhorn struct hstate *h; 31759a305230SLee Schermerhorn unsigned long nr_huge_pages; 31769a305230SLee Schermerhorn int nid; 31779a305230SLee Schermerhorn 31789a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 31799a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 31809a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 31819a305230SLee Schermerhorn else 31829a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 31839a305230SLee Schermerhorn 3184ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", nr_huge_pages); 3185a3437870SNishanth Aravamudan } 3186adbe8726SEric B Munson 3187238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 3188238d3c13SDavid Rientjes struct hstate *h, int nid, 3189238d3c13SDavid Rientjes unsigned long count, size_t len) 3190a3437870SNishanth Aravamudan { 3191a3437870SNishanth Aravamudan int err; 31922d0adf7eSOscar Salvador nodemask_t nodes_allowed, *n_mask; 3193a3437870SNishanth Aravamudan 31942d0adf7eSOscar Salvador if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 31952d0adf7eSOscar Salvador return -EINVAL; 3196adbe8726SEric B Munson 31979a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 31989a305230SLee Schermerhorn /* 31999a305230SLee Schermerhorn * global hstate attribute 32009a305230SLee Schermerhorn */ 32019a305230SLee Schermerhorn if (!(obey_mempolicy && 32022d0adf7eSOscar Salvador init_nodemask_of_mempolicy(&nodes_allowed))) 32032d0adf7eSOscar Salvador n_mask = &node_states[N_MEMORY]; 32042d0adf7eSOscar Salvador else 32052d0adf7eSOscar Salvador n_mask = &nodes_allowed; 32062d0adf7eSOscar Salvador } else { 32079a305230SLee Schermerhorn /* 3208fd875dcaSMike Kravetz * Node specific request. count adjustment happens in 3209fd875dcaSMike Kravetz * set_max_huge_pages() after acquiring hugetlb_lock. 32109a305230SLee Schermerhorn */ 32112d0adf7eSOscar Salvador init_nodemask_of_node(&nodes_allowed, nid); 32122d0adf7eSOscar Salvador n_mask = &nodes_allowed; 3213fd875dcaSMike Kravetz } 32149a305230SLee Schermerhorn 32152d0adf7eSOscar Salvador err = set_max_huge_pages(h, count, nid, n_mask); 321606808b08SLee Schermerhorn 32174eb0716eSAlexandre Ghiti return err ? err : len; 321806808b08SLee Schermerhorn } 321906808b08SLee Schermerhorn 3220238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 3221238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 3222238d3c13SDavid Rientjes size_t len) 3223238d3c13SDavid Rientjes { 3224238d3c13SDavid Rientjes struct hstate *h; 3225238d3c13SDavid Rientjes unsigned long count; 3226238d3c13SDavid Rientjes int nid; 3227238d3c13SDavid Rientjes int err; 3228238d3c13SDavid Rientjes 3229238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 3230238d3c13SDavid Rientjes if (err) 3231238d3c13SDavid Rientjes return err; 3232238d3c13SDavid Rientjes 3233238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 3234238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 3235238d3c13SDavid Rientjes } 3236238d3c13SDavid Rientjes 323706808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 323806808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 323906808b08SLee Schermerhorn { 324006808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 324106808b08SLee Schermerhorn } 324206808b08SLee Schermerhorn 324306808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 324406808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 324506808b08SLee Schermerhorn { 3246238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 3247a3437870SNishanth Aravamudan } 3248a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 3249a3437870SNishanth Aravamudan 325006808b08SLee Schermerhorn #ifdef CONFIG_NUMA 325106808b08SLee Schermerhorn 325206808b08SLee Schermerhorn /* 325306808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 325406808b08SLee Schermerhorn * huge page alloc/free. 325506808b08SLee Schermerhorn */ 325606808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 3257ae7a927dSJoe Perches struct kobj_attribute *attr, 3258ae7a927dSJoe Perches char *buf) 325906808b08SLee Schermerhorn { 326006808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 326106808b08SLee Schermerhorn } 326206808b08SLee Schermerhorn 326306808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 326406808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 326506808b08SLee Schermerhorn { 3266238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 326706808b08SLee Schermerhorn } 326806808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 326906808b08SLee Schermerhorn #endif 327006808b08SLee Schermerhorn 327106808b08SLee Schermerhorn 3272a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 3273a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3274a3437870SNishanth Aravamudan { 32759a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3276ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); 3277a3437870SNishanth Aravamudan } 3278adbe8726SEric B Munson 3279a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 3280a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 3281a3437870SNishanth Aravamudan { 3282a3437870SNishanth Aravamudan int err; 3283a3437870SNishanth Aravamudan unsigned long input; 32849a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3285a3437870SNishanth Aravamudan 3286bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3287adbe8726SEric B Munson return -EINVAL; 3288adbe8726SEric B Munson 32893dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 3290a3437870SNishanth Aravamudan if (err) 329173ae31e5SEric B Munson return err; 3292a3437870SNishanth Aravamudan 3293db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3294a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 3295db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3296a3437870SNishanth Aravamudan 3297a3437870SNishanth Aravamudan return count; 3298a3437870SNishanth Aravamudan } 3299a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 3300a3437870SNishanth Aravamudan 3301a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 3302a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3303a3437870SNishanth Aravamudan { 33049a305230SLee Schermerhorn struct hstate *h; 33059a305230SLee Schermerhorn unsigned long free_huge_pages; 33069a305230SLee Schermerhorn int nid; 33079a305230SLee Schermerhorn 33089a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 33099a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 33109a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 33119a305230SLee Schermerhorn else 33129a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 33139a305230SLee Schermerhorn 3314ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", free_huge_pages); 3315a3437870SNishanth Aravamudan } 3316a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 3317a3437870SNishanth Aravamudan 3318a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 3319a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3320a3437870SNishanth Aravamudan { 33219a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3322ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); 3323a3437870SNishanth Aravamudan } 3324a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 3325a3437870SNishanth Aravamudan 3326a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 3327a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3328a3437870SNishanth Aravamudan { 33299a305230SLee Schermerhorn struct hstate *h; 33309a305230SLee Schermerhorn unsigned long surplus_huge_pages; 33319a305230SLee Schermerhorn int nid; 33329a305230SLee Schermerhorn 33339a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 33349a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 33359a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 33369a305230SLee Schermerhorn else 33379a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 33389a305230SLee Schermerhorn 3339ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", surplus_huge_pages); 3340a3437870SNishanth Aravamudan } 3341a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 3342a3437870SNishanth Aravamudan 3343a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 3344a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 3345a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 3346a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 3347a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 3348a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 334906808b08SLee Schermerhorn #ifdef CONFIG_NUMA 335006808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 335106808b08SLee Schermerhorn #endif 3352a3437870SNishanth Aravamudan NULL, 3353a3437870SNishanth Aravamudan }; 3354a3437870SNishanth Aravamudan 335567e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 3356a3437870SNishanth Aravamudan .attrs = hstate_attrs, 3357a3437870SNishanth Aravamudan }; 3358a3437870SNishanth Aravamudan 3359094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 33609a305230SLee Schermerhorn struct kobject **hstate_kobjs, 336167e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 3362a3437870SNishanth Aravamudan { 3363a3437870SNishanth Aravamudan int retval; 3364972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 3365a3437870SNishanth Aravamudan 33669a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 33679a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 3368a3437870SNishanth Aravamudan return -ENOMEM; 3369a3437870SNishanth Aravamudan 33709a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 3371cc2205a6SMiaohe Lin if (retval) { 33729a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 3373cc2205a6SMiaohe Lin hstate_kobjs[hi] = NULL; 3374cc2205a6SMiaohe Lin } 3375a3437870SNishanth Aravamudan 3376a3437870SNishanth Aravamudan return retval; 3377a3437870SNishanth Aravamudan } 3378a3437870SNishanth Aravamudan 3379a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 3380a3437870SNishanth Aravamudan { 3381a3437870SNishanth Aravamudan struct hstate *h; 3382a3437870SNishanth Aravamudan int err; 3383a3437870SNishanth Aravamudan 3384a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 3385a3437870SNishanth Aravamudan if (!hugepages_kobj) 3386a3437870SNishanth Aravamudan return; 3387a3437870SNishanth Aravamudan 3388a3437870SNishanth Aravamudan for_each_hstate(h) { 33899a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 33909a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 3391a3437870SNishanth Aravamudan if (err) 3392282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s", h->name); 3393a3437870SNishanth Aravamudan } 3394a3437870SNishanth Aravamudan } 3395a3437870SNishanth Aravamudan 33969a305230SLee Schermerhorn #ifdef CONFIG_NUMA 33979a305230SLee Schermerhorn 33989a305230SLee Schermerhorn /* 33999a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 340010fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 340110fbcf4cSKay Sievers * index of a node device or _hstate == node id. 340210fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 34039a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 34049a305230SLee Schermerhorn */ 34059a305230SLee Schermerhorn struct node_hstate { 34069a305230SLee Schermerhorn struct kobject *hugepages_kobj; 34079a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 34089a305230SLee Schermerhorn }; 3409b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 34109a305230SLee Schermerhorn 34119a305230SLee Schermerhorn /* 341210fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 34139a305230SLee Schermerhorn */ 34149a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 34159a305230SLee Schermerhorn &nr_hugepages_attr.attr, 34169a305230SLee Schermerhorn &free_hugepages_attr.attr, 34179a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 34189a305230SLee Schermerhorn NULL, 34199a305230SLee Schermerhorn }; 34209a305230SLee Schermerhorn 342167e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 34229a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 34239a305230SLee Schermerhorn }; 34249a305230SLee Schermerhorn 34259a305230SLee Schermerhorn /* 342610fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 34279a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 34289a305230SLee Schermerhorn */ 34299a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 34309a305230SLee Schermerhorn { 34319a305230SLee Schermerhorn int nid; 34329a305230SLee Schermerhorn 34339a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 34349a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 34359a305230SLee Schermerhorn int i; 34369a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 34379a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 34389a305230SLee Schermerhorn if (nidp) 34399a305230SLee Schermerhorn *nidp = nid; 34409a305230SLee Schermerhorn return &hstates[i]; 34419a305230SLee Schermerhorn } 34429a305230SLee Schermerhorn } 34439a305230SLee Schermerhorn 34449a305230SLee Schermerhorn BUG(); 34459a305230SLee Schermerhorn return NULL; 34469a305230SLee Schermerhorn } 34479a305230SLee Schermerhorn 34489a305230SLee Schermerhorn /* 344910fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 34509a305230SLee Schermerhorn * No-op if no hstate attributes attached. 34519a305230SLee Schermerhorn */ 34523cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 34539a305230SLee Schermerhorn { 34549a305230SLee Schermerhorn struct hstate *h; 345510fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 34569a305230SLee Schermerhorn 34579a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 34589b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 34599a305230SLee Schermerhorn 3460972dc4deSAneesh Kumar K.V for_each_hstate(h) { 3461972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 3462972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 3463972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 3464972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 3465972dc4deSAneesh Kumar K.V } 34669a305230SLee Schermerhorn } 34679a305230SLee Schermerhorn 34689a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 34699a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 34709a305230SLee Schermerhorn } 34719a305230SLee Schermerhorn 34729a305230SLee Schermerhorn 34739a305230SLee Schermerhorn /* 347410fbcf4cSKay Sievers * Register hstate attributes for a single node device. 34759a305230SLee Schermerhorn * No-op if attributes already registered. 34769a305230SLee Schermerhorn */ 34773cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 34789a305230SLee Schermerhorn { 34799a305230SLee Schermerhorn struct hstate *h; 348010fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 34819a305230SLee Schermerhorn int err; 34829a305230SLee Schermerhorn 34839a305230SLee Schermerhorn if (nhs->hugepages_kobj) 34849a305230SLee Schermerhorn return; /* already allocated */ 34859a305230SLee Schermerhorn 34869a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 348710fbcf4cSKay Sievers &node->dev.kobj); 34889a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 34899a305230SLee Schermerhorn return; 34909a305230SLee Schermerhorn 34919a305230SLee Schermerhorn for_each_hstate(h) { 34929a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 34939a305230SLee Schermerhorn nhs->hstate_kobjs, 34949a305230SLee Schermerhorn &per_node_hstate_attr_group); 34959a305230SLee Schermerhorn if (err) { 3496282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s for node %d\n", 349710fbcf4cSKay Sievers h->name, node->dev.id); 34989a305230SLee Schermerhorn hugetlb_unregister_node(node); 34999a305230SLee Schermerhorn break; 35009a305230SLee Schermerhorn } 35019a305230SLee Schermerhorn } 35029a305230SLee Schermerhorn } 35039a305230SLee Schermerhorn 35049a305230SLee Schermerhorn /* 35059b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 350610fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 350710fbcf4cSKay Sievers * registered their associated device by this time. 35089a305230SLee Schermerhorn */ 35097d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 35109a305230SLee Schermerhorn { 35119a305230SLee Schermerhorn int nid; 35129a305230SLee Schermerhorn 35138cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 35148732794bSWen Congyang struct node *node = node_devices[nid]; 351510fbcf4cSKay Sievers if (node->dev.id == nid) 35169a305230SLee Schermerhorn hugetlb_register_node(node); 35179a305230SLee Schermerhorn } 35189a305230SLee Schermerhorn 35199a305230SLee Schermerhorn /* 352010fbcf4cSKay Sievers * Let the node device driver know we're here so it can 35219a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 35229a305230SLee Schermerhorn */ 35239a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 35249a305230SLee Schermerhorn hugetlb_unregister_node); 35259a305230SLee Schermerhorn } 35269a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 35279a305230SLee Schermerhorn 35289a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 35299a305230SLee Schermerhorn { 35309a305230SLee Schermerhorn BUG(); 35319a305230SLee Schermerhorn if (nidp) 35329a305230SLee Schermerhorn *nidp = -1; 35339a305230SLee Schermerhorn return NULL; 35349a305230SLee Schermerhorn } 35359a305230SLee Schermerhorn 35369a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 35379a305230SLee Schermerhorn 35389a305230SLee Schermerhorn #endif 35399a305230SLee Schermerhorn 3540a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 3541a3437870SNishanth Aravamudan { 35428382d914SDavidlohr Bueso int i; 35438382d914SDavidlohr Bueso 3544d6995da3SMike Kravetz BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < 3545d6995da3SMike Kravetz __NR_HPAGEFLAGS); 3546d6995da3SMike Kravetz 3547c2833a5bSMike Kravetz if (!hugepages_supported()) { 3548c2833a5bSMike Kravetz if (hugetlb_max_hstate || default_hstate_max_huge_pages) 3549c2833a5bSMike Kravetz pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); 35500ef89d25SBenjamin Herrenschmidt return 0; 3551d715cf80SLiam R. Howlett } 3552d715cf80SLiam R. Howlett 3553282f4214SMike Kravetz /* 3554282f4214SMike Kravetz * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some 3555282f4214SMike Kravetz * architectures depend on setup being done here. 3556282f4214SMike Kravetz */ 3557a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 3558282f4214SMike Kravetz if (!parsed_default_hugepagesz) { 3559282f4214SMike Kravetz /* 3560282f4214SMike Kravetz * If we did not parse a default huge page size, set 3561282f4214SMike Kravetz * default_hstate_idx to HPAGE_SIZE hstate. And, if the 3562282f4214SMike Kravetz * number of huge pages for this default size was implicitly 3563282f4214SMike Kravetz * specified, set that here as well. 3564282f4214SMike Kravetz * Note that the implicit setting will overwrite an explicit 3565282f4214SMike Kravetz * setting. A warning will be printed in this case. 3566282f4214SMike Kravetz */ 3567282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); 3568f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 3569282f4214SMike Kravetz if (default_hstate.max_huge_pages) { 3570282f4214SMike Kravetz char buf[32]; 3571282f4214SMike Kravetz 3572282f4214SMike Kravetz string_get_size(huge_page_size(&default_hstate), 3573282f4214SMike Kravetz 1, STRING_UNITS_2, buf, 32); 3574282f4214SMike Kravetz pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", 3575282f4214SMike Kravetz default_hstate.max_huge_pages, buf); 3576282f4214SMike Kravetz pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", 3577282f4214SMike Kravetz default_hstate_max_huge_pages); 3578282f4214SMike Kravetz } 3579282f4214SMike Kravetz default_hstate.max_huge_pages = 3580282f4214SMike Kravetz default_hstate_max_huge_pages; 3581282f4214SMike Kravetz } 3582f8b74815SVaishali Thakkar } 3583a3437870SNishanth Aravamudan 3584cf11e85fSRoman Gushchin hugetlb_cma_check(); 3585a3437870SNishanth Aravamudan hugetlb_init_hstates(); 3586aa888a74SAndi Kleen gather_bootmem_prealloc(); 3587a3437870SNishanth Aravamudan report_hugepages(); 3588a3437870SNishanth Aravamudan 3589a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 35909a305230SLee Schermerhorn hugetlb_register_all_nodes(); 35917179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 35929a305230SLee Schermerhorn 35938382d914SDavidlohr Bueso #ifdef CONFIG_SMP 35948382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 35958382d914SDavidlohr Bueso #else 35968382d914SDavidlohr Bueso num_fault_mutexes = 1; 35978382d914SDavidlohr Bueso #endif 3598c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 35996da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 36006da2ec56SKees Cook GFP_KERNEL); 3601c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 36028382d914SDavidlohr Bueso 36038382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 3604c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 3605a3437870SNishanth Aravamudan return 0; 3606a3437870SNishanth Aravamudan } 36073e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 3608a3437870SNishanth Aravamudan 3609ae94da89SMike Kravetz /* Overwritten by architectures with more huge page sizes */ 3610ae94da89SMike Kravetz bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) 36119fee021dSVaishali Thakkar { 3612ae94da89SMike Kravetz return size == HPAGE_SIZE; 36139fee021dSVaishali Thakkar } 36149fee021dSVaishali Thakkar 3615d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 3616a3437870SNishanth Aravamudan { 3617a3437870SNishanth Aravamudan struct hstate *h; 36188faa8b07SAndi Kleen unsigned long i; 36198faa8b07SAndi Kleen 3620a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 3621a3437870SNishanth Aravamudan return; 3622a3437870SNishanth Aravamudan } 362347d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 3624a3437870SNishanth Aravamudan BUG_ON(order == 0); 362547d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 362629383967SMike Kravetz mutex_init(&h->resize_lock); 3627a3437870SNishanth Aravamudan h->order = order; 3628aca78307SMiaohe Lin h->mask = ~(huge_page_size(h) - 1); 36298faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 36308faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 36310edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 363254f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 363354f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 3634a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 3635a3437870SNishanth Aravamudan huge_page_size(h)/1024); 363677490587SMuchun Song hugetlb_vmemmap_init(h); 36378faa8b07SAndi Kleen 3638a3437870SNishanth Aravamudan parsed_hstate = h; 3639a3437870SNishanth Aravamudan } 3640a3437870SNishanth Aravamudan 3641282f4214SMike Kravetz /* 3642282f4214SMike Kravetz * hugepages command line processing 3643282f4214SMike Kravetz * hugepages normally follows a valid hugepagsz or default_hugepagsz 3644282f4214SMike Kravetz * specification. If not, ignore the hugepages value. hugepages can also 3645282f4214SMike Kravetz * be the first huge page command line option in which case it implicitly 3646282f4214SMike Kravetz * specifies the number of huge pages for the default size. 3647282f4214SMike Kravetz */ 3648282f4214SMike Kravetz static int __init hugepages_setup(char *s) 3649a3437870SNishanth Aravamudan { 3650a3437870SNishanth Aravamudan unsigned long *mhp; 36518faa8b07SAndi Kleen static unsigned long *last_mhp; 3652a3437870SNishanth Aravamudan 36539fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 3654282f4214SMike Kravetz pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); 36559fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 3656282f4214SMike Kravetz return 0; 36579fee021dSVaishali Thakkar } 3658282f4214SMike Kravetz 3659a3437870SNishanth Aravamudan /* 3660282f4214SMike Kravetz * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter 3661282f4214SMike Kravetz * yet, so this hugepages= parameter goes to the "default hstate". 3662282f4214SMike Kravetz * Otherwise, it goes with the previously parsed hugepagesz or 3663282f4214SMike Kravetz * default_hugepagesz. 3664a3437870SNishanth Aravamudan */ 36659fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 3666a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 3667a3437870SNishanth Aravamudan else 3668a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 3669a3437870SNishanth Aravamudan 36708faa8b07SAndi Kleen if (mhp == last_mhp) { 3671282f4214SMike Kravetz pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); 3672282f4214SMike Kravetz return 0; 36738faa8b07SAndi Kleen } 36748faa8b07SAndi Kleen 3675a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 3676a3437870SNishanth Aravamudan *mhp = 0; 3677a3437870SNishanth Aravamudan 36788faa8b07SAndi Kleen /* 36798faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 368004adbc3fSMiaohe Lin * But we need to allocate gigantic hstates here early to still 36818faa8b07SAndi Kleen * use the bootmem allocator. 36828faa8b07SAndi Kleen */ 368304adbc3fSMiaohe Lin if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) 36848faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 36858faa8b07SAndi Kleen 36868faa8b07SAndi Kleen last_mhp = mhp; 36878faa8b07SAndi Kleen 3688a3437870SNishanth Aravamudan return 1; 3689a3437870SNishanth Aravamudan } 3690282f4214SMike Kravetz __setup("hugepages=", hugepages_setup); 3691e11bfbfcSNick Piggin 3692282f4214SMike Kravetz /* 3693282f4214SMike Kravetz * hugepagesz command line processing 3694282f4214SMike Kravetz * A specific huge page size can only be specified once with hugepagesz. 3695282f4214SMike Kravetz * hugepagesz is followed by hugepages on the command line. The global 3696282f4214SMike Kravetz * variable 'parsed_valid_hugepagesz' is used to determine if prior 3697282f4214SMike Kravetz * hugepagesz argument was valid. 3698282f4214SMike Kravetz */ 3699359f2544SMike Kravetz static int __init hugepagesz_setup(char *s) 3700e11bfbfcSNick Piggin { 3701359f2544SMike Kravetz unsigned long size; 3702282f4214SMike Kravetz struct hstate *h; 3703282f4214SMike Kravetz 3704282f4214SMike Kravetz parsed_valid_hugepagesz = false; 3705359f2544SMike Kravetz size = (unsigned long)memparse(s, NULL); 3706359f2544SMike Kravetz 3707359f2544SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 3708282f4214SMike Kravetz pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); 3709359f2544SMike Kravetz return 0; 3710359f2544SMike Kravetz } 3711359f2544SMike Kravetz 3712282f4214SMike Kravetz h = size_to_hstate(size); 3713282f4214SMike Kravetz if (h) { 3714282f4214SMike Kravetz /* 3715282f4214SMike Kravetz * hstate for this size already exists. This is normally 3716282f4214SMike Kravetz * an error, but is allowed if the existing hstate is the 3717282f4214SMike Kravetz * default hstate. More specifically, it is only allowed if 3718282f4214SMike Kravetz * the number of huge pages for the default hstate was not 3719282f4214SMike Kravetz * previously specified. 3720282f4214SMike Kravetz */ 3721282f4214SMike Kravetz if (!parsed_default_hugepagesz || h != &default_hstate || 3722282f4214SMike Kravetz default_hstate.max_huge_pages) { 3723282f4214SMike Kravetz pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); 372438237830SMike Kravetz return 0; 372538237830SMike Kravetz } 372638237830SMike Kravetz 3727282f4214SMike Kravetz /* 3728282f4214SMike Kravetz * No need to call hugetlb_add_hstate() as hstate already 3729282f4214SMike Kravetz * exists. But, do set parsed_hstate so that a following 3730282f4214SMike Kravetz * hugepages= parameter will be applied to this hstate. 3731282f4214SMike Kravetz */ 3732282f4214SMike Kravetz parsed_hstate = h; 3733282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3734e11bfbfcSNick Piggin return 1; 3735e11bfbfcSNick Piggin } 3736282f4214SMike Kravetz 3737359f2544SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 3738282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3739359f2544SMike Kravetz return 1; 3740359f2544SMike Kravetz } 3741359f2544SMike Kravetz __setup("hugepagesz=", hugepagesz_setup); 3742359f2544SMike Kravetz 3743282f4214SMike Kravetz /* 3744282f4214SMike Kravetz * default_hugepagesz command line input 3745282f4214SMike Kravetz * Only one instance of default_hugepagesz allowed on command line. 3746282f4214SMike Kravetz */ 3747ae94da89SMike Kravetz static int __init default_hugepagesz_setup(char *s) 3748e11bfbfcSNick Piggin { 3749ae94da89SMike Kravetz unsigned long size; 3750ae94da89SMike Kravetz 3751282f4214SMike Kravetz parsed_valid_hugepagesz = false; 3752282f4214SMike Kravetz if (parsed_default_hugepagesz) { 3753282f4214SMike Kravetz pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); 3754282f4214SMike Kravetz return 0; 3755282f4214SMike Kravetz } 3756282f4214SMike Kravetz 3757282f4214SMike Kravetz size = (unsigned long)memparse(s, NULL); 3758282f4214SMike Kravetz 3759282f4214SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 3760282f4214SMike Kravetz pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); 3761282f4214SMike Kravetz return 0; 3762282f4214SMike Kravetz } 3763282f4214SMike Kravetz 3764282f4214SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 3765282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3766282f4214SMike Kravetz parsed_default_hugepagesz = true; 3767282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(size)); 3768282f4214SMike Kravetz 3769282f4214SMike Kravetz /* 3770282f4214SMike Kravetz * The number of default huge pages (for this size) could have been 3771282f4214SMike Kravetz * specified as the first hugetlb parameter: hugepages=X. If so, 3772282f4214SMike Kravetz * then default_hstate_max_huge_pages is set. If the default huge 3773282f4214SMike Kravetz * page size is gigantic (>= MAX_ORDER), then the pages must be 3774282f4214SMike Kravetz * allocated here from bootmem allocator. 3775282f4214SMike Kravetz */ 3776282f4214SMike Kravetz if (default_hstate_max_huge_pages) { 3777282f4214SMike Kravetz default_hstate.max_huge_pages = default_hstate_max_huge_pages; 3778282f4214SMike Kravetz if (hstate_is_gigantic(&default_hstate)) 3779282f4214SMike Kravetz hugetlb_hstate_alloc_pages(&default_hstate); 3780282f4214SMike Kravetz default_hstate_max_huge_pages = 0; 3781282f4214SMike Kravetz } 3782282f4214SMike Kravetz 3783e11bfbfcSNick Piggin return 1; 3784e11bfbfcSNick Piggin } 3785ae94da89SMike Kravetz __setup("default_hugepagesz=", default_hugepagesz_setup); 3786a3437870SNishanth Aravamudan 37878ca39e68SMuchun Song static unsigned int allowed_mems_nr(struct hstate *h) 37888a213460SNishanth Aravamudan { 37898a213460SNishanth Aravamudan int node; 37908a213460SNishanth Aravamudan unsigned int nr = 0; 37918ca39e68SMuchun Song nodemask_t *mpol_allowed; 37928ca39e68SMuchun Song unsigned int *array = h->free_huge_pages_node; 37938ca39e68SMuchun Song gfp_t gfp_mask = htlb_alloc_mask(h); 37948a213460SNishanth Aravamudan 37958ca39e68SMuchun Song mpol_allowed = policy_nodemask_current(gfp_mask); 37968ca39e68SMuchun Song 37978ca39e68SMuchun Song for_each_node_mask(node, cpuset_current_mems_allowed) { 3798c93b0a99SJiapeng Zhong if (!mpol_allowed || node_isset(node, *mpol_allowed)) 37998a213460SNishanth Aravamudan nr += array[node]; 38008ca39e68SMuchun Song } 38018a213460SNishanth Aravamudan 38028a213460SNishanth Aravamudan return nr; 38038a213460SNishanth Aravamudan } 38048a213460SNishanth Aravamudan 38058a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 380617743798SMuchun Song static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, 380717743798SMuchun Song void *buffer, size_t *length, 380817743798SMuchun Song loff_t *ppos, unsigned long *out) 380917743798SMuchun Song { 381017743798SMuchun Song struct ctl_table dup_table; 381117743798SMuchun Song 381217743798SMuchun Song /* 381317743798SMuchun Song * In order to avoid races with __do_proc_doulongvec_minmax(), we 381417743798SMuchun Song * can duplicate the @table and alter the duplicate of it. 381517743798SMuchun Song */ 381617743798SMuchun Song dup_table = *table; 381717743798SMuchun Song dup_table.data = out; 381817743798SMuchun Song 381917743798SMuchun Song return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); 382017743798SMuchun Song } 382117743798SMuchun Song 382206808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 382306808b08SLee Schermerhorn struct ctl_table *table, int write, 382432927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 38251da177e4SLinus Torvalds { 3826e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 3827238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 382808d4a246SMichal Hocko int ret; 3829e5ff2159SAndi Kleen 3830457c1b27SNishanth Aravamudan if (!hugepages_supported()) 383186613628SJan Stancek return -EOPNOTSUPP; 3832457c1b27SNishanth Aravamudan 383317743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 383417743798SMuchun Song &tmp); 383508d4a246SMichal Hocko if (ret) 383608d4a246SMichal Hocko goto out; 3837e5ff2159SAndi Kleen 3838238d3c13SDavid Rientjes if (write) 3839238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 3840238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 384108d4a246SMichal Hocko out: 384208d4a246SMichal Hocko return ret; 38431da177e4SLinus Torvalds } 3844396faf03SMel Gorman 384506808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 384632927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 384706808b08SLee Schermerhorn { 384806808b08SLee Schermerhorn 384906808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 385006808b08SLee Schermerhorn buffer, length, ppos); 385106808b08SLee Schermerhorn } 385206808b08SLee Schermerhorn 385306808b08SLee Schermerhorn #ifdef CONFIG_NUMA 385406808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 385532927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 385606808b08SLee Schermerhorn { 385706808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 385806808b08SLee Schermerhorn buffer, length, ppos); 385906808b08SLee Schermerhorn } 386006808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 386106808b08SLee Schermerhorn 3862a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 386332927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 3864a3d0c6aaSNishanth Aravamudan { 3865a5516438SAndi Kleen struct hstate *h = &default_hstate; 3866e5ff2159SAndi Kleen unsigned long tmp; 386708d4a246SMichal Hocko int ret; 3868e5ff2159SAndi Kleen 3869457c1b27SNishanth Aravamudan if (!hugepages_supported()) 387086613628SJan Stancek return -EOPNOTSUPP; 3871457c1b27SNishanth Aravamudan 3872e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 3873e5ff2159SAndi Kleen 3874bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 3875adbe8726SEric B Munson return -EINVAL; 3876adbe8726SEric B Munson 387717743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 387817743798SMuchun Song &tmp); 387908d4a246SMichal Hocko if (ret) 388008d4a246SMichal Hocko goto out; 3881e5ff2159SAndi Kleen 3882e5ff2159SAndi Kleen if (write) { 3883db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3884e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 3885db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3886e5ff2159SAndi Kleen } 388708d4a246SMichal Hocko out: 388808d4a246SMichal Hocko return ret; 3889a3d0c6aaSNishanth Aravamudan } 3890a3d0c6aaSNishanth Aravamudan 38911da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 38921da177e4SLinus Torvalds 3893e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 38941da177e4SLinus Torvalds { 3895fcb2b0c5SRoman Gushchin struct hstate *h; 3896fcb2b0c5SRoman Gushchin unsigned long total = 0; 3897fcb2b0c5SRoman Gushchin 3898457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3899457c1b27SNishanth Aravamudan return; 3900fcb2b0c5SRoman Gushchin 3901fcb2b0c5SRoman Gushchin for_each_hstate(h) { 3902fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 3903fcb2b0c5SRoman Gushchin 3904aca78307SMiaohe Lin total += huge_page_size(h) * count; 3905fcb2b0c5SRoman Gushchin 3906fcb2b0c5SRoman Gushchin if (h == &default_hstate) 3907e1759c21SAlexey Dobriyan seq_printf(m, 39081da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 39091da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 3910b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 39117893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 39124f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 3913fcb2b0c5SRoman Gushchin count, 3914a5516438SAndi Kleen h->free_huge_pages, 3915a5516438SAndi Kleen h->resv_huge_pages, 3916a5516438SAndi Kleen h->surplus_huge_pages, 3917aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 3918fcb2b0c5SRoman Gushchin } 3919fcb2b0c5SRoman Gushchin 3920aca78307SMiaohe Lin seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); 39211da177e4SLinus Torvalds } 39221da177e4SLinus Torvalds 39237981593bSJoe Perches int hugetlb_report_node_meminfo(char *buf, int len, int nid) 39241da177e4SLinus Torvalds { 3925a5516438SAndi Kleen struct hstate *h = &default_hstate; 39267981593bSJoe Perches 3927457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3928457c1b27SNishanth Aravamudan return 0; 39297981593bSJoe Perches 39307981593bSJoe Perches return sysfs_emit_at(buf, len, 39311da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 3932a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 3933a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 3934a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 3935a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 3936a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 39371da177e4SLinus Torvalds } 39381da177e4SLinus Torvalds 3939949f7ec5SDavid Rientjes void hugetlb_show_meminfo(void) 3940949f7ec5SDavid Rientjes { 3941949f7ec5SDavid Rientjes struct hstate *h; 3942949f7ec5SDavid Rientjes int nid; 3943949f7ec5SDavid Rientjes 3944457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3945457c1b27SNishanth Aravamudan return; 3946457c1b27SNishanth Aravamudan 3947949f7ec5SDavid Rientjes for_each_node_state(nid, N_MEMORY) 3948949f7ec5SDavid Rientjes for_each_hstate(h) 3949949f7ec5SDavid Rientjes pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 3950949f7ec5SDavid Rientjes nid, 3951949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 3952949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 3953949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 3954aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 3955949f7ec5SDavid Rientjes } 3956949f7ec5SDavid Rientjes 39575d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 39585d317b2bSNaoya Horiguchi { 39595d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 39605d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 39615d317b2bSNaoya Horiguchi } 39625d317b2bSNaoya Horiguchi 39631da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 39641da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 39651da177e4SLinus Torvalds { 3966d0028588SWanpeng Li struct hstate *h; 3967d0028588SWanpeng Li unsigned long nr_total_pages = 0; 3968d0028588SWanpeng Li 3969d0028588SWanpeng Li for_each_hstate(h) 3970d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 3971d0028588SWanpeng Li return nr_total_pages; 39721da177e4SLinus Torvalds } 39731da177e4SLinus Torvalds 3974a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 3975fc1b8a73SMel Gorman { 3976fc1b8a73SMel Gorman int ret = -ENOMEM; 3977fc1b8a73SMel Gorman 39780aa7f354SMiaohe Lin if (!delta) 39790aa7f354SMiaohe Lin return 0; 39800aa7f354SMiaohe Lin 3981db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3982fc1b8a73SMel Gorman /* 3983fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 3984fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 3985fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 3986fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 3987fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 3988fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 3989fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 3990fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 3991fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 3992fc1b8a73SMel Gorman * 3993fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 3994fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 3995fc1b8a73SMel Gorman * we fall back to check against current free page availability as 3996fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 3997fc1b8a73SMel Gorman * semantics that cpuset has. 39988ca39e68SMuchun Song * 39998ca39e68SMuchun Song * Apart from cpuset, we also have memory policy mechanism that 40008ca39e68SMuchun Song * also determines from which node the kernel will allocate memory 40018ca39e68SMuchun Song * in a NUMA system. So similar to cpuset, we also should consider 40028ca39e68SMuchun Song * the memory policy of the current task. Similar to the description 40038ca39e68SMuchun Song * above. 4004fc1b8a73SMel Gorman */ 4005fc1b8a73SMel Gorman if (delta > 0) { 4006a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 4007fc1b8a73SMel Gorman goto out; 4008fc1b8a73SMel Gorman 40098ca39e68SMuchun Song if (delta > allowed_mems_nr(h)) { 4010a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 4011fc1b8a73SMel Gorman goto out; 4012fc1b8a73SMel Gorman } 4013fc1b8a73SMel Gorman } 4014fc1b8a73SMel Gorman 4015fc1b8a73SMel Gorman ret = 0; 4016fc1b8a73SMel Gorman if (delta < 0) 4017a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 4018fc1b8a73SMel Gorman 4019fc1b8a73SMel Gorman out: 4020db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4021fc1b8a73SMel Gorman return ret; 4022fc1b8a73SMel Gorman } 4023fc1b8a73SMel Gorman 402484afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 402584afd99bSAndy Whitcroft { 4026f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 402784afd99bSAndy Whitcroft 402884afd99bSAndy Whitcroft /* 402984afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 403084afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 403184afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 403225985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 403384afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 403484afd99bSAndy Whitcroft * new reference here without additional locking. 403584afd99bSAndy Whitcroft */ 40364e35f483SJoonsoo Kim if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 4037f522c3acSJoonsoo Kim kref_get(&resv->refs); 403884afd99bSAndy Whitcroft } 403984afd99bSAndy Whitcroft 4040a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 4041a1e78772SMel Gorman { 4042a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 4043f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 404490481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 40454e35f483SJoonsoo Kim unsigned long reserve, start, end; 40461c5ecae3SMike Kravetz long gbl_reserve; 404784afd99bSAndy Whitcroft 40484e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 40494e35f483SJoonsoo Kim return; 40504e35f483SJoonsoo Kim 4051a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 4052a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 405384afd99bSAndy Whitcroft 40544e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 4055e9fe92aeSMina Almasry hugetlb_cgroup_uncharge_counter(resv, start, end); 40567251ff78SAdam Litke if (reserve) { 40571c5ecae3SMike Kravetz /* 40581c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 40591c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 40601c5ecae3SMike Kravetz */ 40611c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 40621c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 40637251ff78SAdam Litke } 4064e9fe92aeSMina Almasry 4065e9fe92aeSMina Almasry kref_put(&resv->refs, resv_map_release); 4066a1e78772SMel Gorman } 4067a1e78772SMel Gorman 406831383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 406931383c68SDan Williams { 407031383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 407131383c68SDan Williams return -EINVAL; 407231383c68SDan Williams return 0; 407331383c68SDan Williams } 407431383c68SDan Williams 407505ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 407605ea8860SDan Williams { 4077aca78307SMiaohe Lin return huge_page_size(hstate_vma(vma)); 407805ea8860SDan Williams } 407905ea8860SDan Williams 40801da177e4SLinus Torvalds /* 40811da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 40821da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 40836c26d310SMiaohe Lin * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get 40841da177e4SLinus Torvalds * this far. 40851da177e4SLinus Torvalds */ 4086b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 40871da177e4SLinus Torvalds { 40881da177e4SLinus Torvalds BUG(); 4089d0217ac0SNick Piggin return 0; 40901da177e4SLinus Torvalds } 40911da177e4SLinus Torvalds 4092eec3636aSJane Chu /* 4093eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 4094eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 4095eec3636aSJane Chu * This is because under System V memory model, mappings created via 4096eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 4097eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 4098eec3636aSJane Chu */ 4099f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 4100d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 410184afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 4102a1e78772SMel Gorman .close = hugetlb_vm_op_close, 4103dd3b614fSDmitry Safonov .may_split = hugetlb_vm_op_split, 410405ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 41051da177e4SLinus Torvalds }; 41061da177e4SLinus Torvalds 41071e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 41081e8f889bSDavid Gibson int writable) 410963551ae0SDavid Gibson { 411063551ae0SDavid Gibson pte_t entry; 411179c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 411263551ae0SDavid Gibson 41131e8f889bSDavid Gibson if (writable) { 4114106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 4115106c992aSGerald Schaefer vma->vm_page_prot))); 411663551ae0SDavid Gibson } else { 4117106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 4118106c992aSGerald Schaefer vma->vm_page_prot)); 411963551ae0SDavid Gibson } 412063551ae0SDavid Gibson entry = pte_mkyoung(entry); 412163551ae0SDavid Gibson entry = pte_mkhuge(entry); 412279c1c594SChristophe Leroy entry = arch_make_huge_pte(entry, shift, vma->vm_flags); 412363551ae0SDavid Gibson 412463551ae0SDavid Gibson return entry; 412563551ae0SDavid Gibson } 412663551ae0SDavid Gibson 41271e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 41281e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 41291e8f889bSDavid Gibson { 41301e8f889bSDavid Gibson pte_t entry; 41311e8f889bSDavid Gibson 4132106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 413332f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 41344b3073e1SRussell King update_mmu_cache(vma, address, ptep); 41351e8f889bSDavid Gibson } 41361e8f889bSDavid Gibson 4137d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 41384a705fefSNaoya Horiguchi { 41394a705fefSNaoya Horiguchi swp_entry_t swp; 41404a705fefSNaoya Horiguchi 41414a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 4142d5ed7444SAneesh Kumar K.V return false; 41434a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4144d79d176aSBaoquan He if (is_migration_entry(swp)) 4145d5ed7444SAneesh Kumar K.V return true; 41464a705fefSNaoya Horiguchi else 4147d5ed7444SAneesh Kumar K.V return false; 41484a705fefSNaoya Horiguchi } 41494a705fefSNaoya Horiguchi 41503e5c3600SBaoquan He static bool is_hugetlb_entry_hwpoisoned(pte_t pte) 41514a705fefSNaoya Horiguchi { 41524a705fefSNaoya Horiguchi swp_entry_t swp; 41534a705fefSNaoya Horiguchi 41544a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 41553e5c3600SBaoquan He return false; 41564a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4157d79d176aSBaoquan He if (is_hwpoison_entry(swp)) 41583e5c3600SBaoquan He return true; 41594a705fefSNaoya Horiguchi else 41603e5c3600SBaoquan He return false; 41614a705fefSNaoya Horiguchi } 41621e8f889bSDavid Gibson 41634eae4efaSPeter Xu static void 41644eae4efaSPeter Xu hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, 41654eae4efaSPeter Xu struct page *new_page) 41664eae4efaSPeter Xu { 41674eae4efaSPeter Xu __SetPageUptodate(new_page); 41684eae4efaSPeter Xu set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1)); 41694eae4efaSPeter Xu hugepage_add_new_anon_rmap(new_page, vma, addr); 41704eae4efaSPeter Xu hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); 41714eae4efaSPeter Xu ClearHPageRestoreReserve(new_page); 41724eae4efaSPeter Xu SetHPageMigratable(new_page); 41734eae4efaSPeter Xu } 41744eae4efaSPeter Xu 417563551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 417663551ae0SDavid Gibson struct vm_area_struct *vma) 417763551ae0SDavid Gibson { 41785e41540cSMike Kravetz pte_t *src_pte, *dst_pte, entry, dst_entry; 417963551ae0SDavid Gibson struct page *ptepage; 41801c59827dSHugh Dickins unsigned long addr; 4181ca6eb14dSPeter Xu bool cow = is_cow_mapping(vma->vm_flags); 4182a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 4183a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 41844eae4efaSPeter Xu unsigned long npages = pages_per_huge_page(h); 4185c0d0381aSMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 4186ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 4187e8569dd2SAndreas Sandberg int ret = 0; 41881e8f889bSDavid Gibson 4189ac46d4f3SJérôme Glisse if (cow) { 41907269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src, 41916f4f13e8SJérôme Glisse vma->vm_start, 4192ac46d4f3SJérôme Glisse vma->vm_end); 4193ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4194c0d0381aSMike Kravetz } else { 4195c0d0381aSMike Kravetz /* 4196c0d0381aSMike Kravetz * For shared mappings i_mmap_rwsem must be held to call 4197c0d0381aSMike Kravetz * huge_pte_alloc, otherwise the returned ptep could go 4198c0d0381aSMike Kravetz * away if part of a shared pmd and another thread calls 4199c0d0381aSMike Kravetz * huge_pmd_unshare. 4200c0d0381aSMike Kravetz */ 4201c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 4202ac46d4f3SJérôme Glisse } 4203e8569dd2SAndreas Sandberg 4204a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 4205cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 42067868a208SPunit Agrawal src_pte = huge_pte_offset(src, addr, sz); 4207c74df32cSHugh Dickins if (!src_pte) 4208c74df32cSHugh Dickins continue; 4209aec44e0fSPeter Xu dst_pte = huge_pte_alloc(dst, vma, addr, sz); 4210e8569dd2SAndreas Sandberg if (!dst_pte) { 4211e8569dd2SAndreas Sandberg ret = -ENOMEM; 4212e8569dd2SAndreas Sandberg break; 4213e8569dd2SAndreas Sandberg } 4214c5c99429SLarry Woodman 42155e41540cSMike Kravetz /* 42165e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 42175e41540cSMike Kravetz * dst_pte == src_pte is the common case of src/dest sharing. 42185e41540cSMike Kravetz * 42195e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 42205e41540cSMike Kravetz * another vma. If dst_pte !none, this implies sharing. 42215e41540cSMike Kravetz * Check here before taking page table lock, and once again 42225e41540cSMike Kravetz * after taking the lock below. 42235e41540cSMike Kravetz */ 42245e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 42255e41540cSMike Kravetz if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) 4226c5c99429SLarry Woodman continue; 4227c5c99429SLarry Woodman 4228cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 4229cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 4230cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 42314a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 42325e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 42334eae4efaSPeter Xu again: 42345e41540cSMike Kravetz if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { 42355e41540cSMike Kravetz /* 42365e41540cSMike Kravetz * Skip if src entry none. Also, skip in the 42375e41540cSMike Kravetz * unlikely case dst entry !none as this implies 42385e41540cSMike Kravetz * sharing with another vma. 42395e41540cSMike Kravetz */ 42404a705fefSNaoya Horiguchi ; 42414a705fefSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry) || 42424a705fefSNaoya Horiguchi is_hugetlb_entry_hwpoisoned(entry))) { 42434a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 42444a705fefSNaoya Horiguchi 42454a705fefSNaoya Horiguchi if (is_write_migration_entry(swp_entry) && cow) { 42464a705fefSNaoya Horiguchi /* 42474a705fefSNaoya Horiguchi * COW mappings require pages in both 42484a705fefSNaoya Horiguchi * parent and child to be set to read. 42494a705fefSNaoya Horiguchi */ 42504a705fefSNaoya Horiguchi make_migration_entry_read(&swp_entry); 42514a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 4252e5251fd4SPunit Agrawal set_huge_swap_pte_at(src, addr, src_pte, 4253e5251fd4SPunit Agrawal entry, sz); 42544a705fefSNaoya Horiguchi } 4255e5251fd4SPunit Agrawal set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); 42564a705fefSNaoya Horiguchi } else { 42574eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 42584eae4efaSPeter Xu ptepage = pte_page(entry); 42594eae4efaSPeter Xu get_page(ptepage); 42604eae4efaSPeter Xu 42614eae4efaSPeter Xu /* 42624eae4efaSPeter Xu * This is a rare case where we see pinned hugetlb 42634eae4efaSPeter Xu * pages while they're prone to COW. We need to do the 42644eae4efaSPeter Xu * COW earlier during fork. 42654eae4efaSPeter Xu * 42664eae4efaSPeter Xu * When pre-allocating the page or copying data, we 42674eae4efaSPeter Xu * need to be without the pgtable locks since we could 42684eae4efaSPeter Xu * sleep during the process. 42694eae4efaSPeter Xu */ 42704eae4efaSPeter Xu if (unlikely(page_needs_cow_for_dma(vma, ptepage))) { 42714eae4efaSPeter Xu pte_t src_pte_old = entry; 42724eae4efaSPeter Xu struct page *new; 42734eae4efaSPeter Xu 42744eae4efaSPeter Xu spin_unlock(src_ptl); 42754eae4efaSPeter Xu spin_unlock(dst_ptl); 42764eae4efaSPeter Xu /* Do not use reserve as it's private owned */ 42774eae4efaSPeter Xu new = alloc_huge_page(vma, addr, 1); 42784eae4efaSPeter Xu if (IS_ERR(new)) { 42794eae4efaSPeter Xu put_page(ptepage); 42804eae4efaSPeter Xu ret = PTR_ERR(new); 42814eae4efaSPeter Xu break; 42824eae4efaSPeter Xu } 42834eae4efaSPeter Xu copy_user_huge_page(new, ptepage, addr, vma, 42844eae4efaSPeter Xu npages); 42854eae4efaSPeter Xu put_page(ptepage); 42864eae4efaSPeter Xu 42874eae4efaSPeter Xu /* Install the new huge page if src pte stable */ 42884eae4efaSPeter Xu dst_ptl = huge_pte_lock(h, dst, dst_pte); 42894eae4efaSPeter Xu src_ptl = huge_pte_lockptr(h, src, src_pte); 42904eae4efaSPeter Xu spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 42914eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 42924eae4efaSPeter Xu if (!pte_same(src_pte_old, entry)) { 4293846be085SMike Kravetz restore_reserve_on_error(h, vma, addr, 4294846be085SMike Kravetz new); 42954eae4efaSPeter Xu put_page(new); 42964eae4efaSPeter Xu /* dst_entry won't change as in child */ 42974eae4efaSPeter Xu goto again; 42984eae4efaSPeter Xu } 42994eae4efaSPeter Xu hugetlb_install_page(vma, dst_pte, addr, new); 43004eae4efaSPeter Xu spin_unlock(src_ptl); 43014eae4efaSPeter Xu spin_unlock(dst_ptl); 43024eae4efaSPeter Xu continue; 43034eae4efaSPeter Xu } 43044eae4efaSPeter Xu 430534ee645eSJoerg Roedel if (cow) { 43060f10851eSJérôme Glisse /* 43070f10851eSJérôme Glisse * No need to notify as we are downgrading page 43080f10851eSJérôme Glisse * table protection not changing it to point 43090f10851eSJérôme Glisse * to a new page. 43100f10851eSJérôme Glisse * 4311ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 43120f10851eSJérôme Glisse */ 43137f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 431484894e1cSPeter Xu entry = huge_pte_wrprotect(entry); 431534ee645eSJoerg Roedel } 43164eae4efaSPeter Xu 431753f9263bSKirill A. Shutemov page_dup_rmap(ptepage, true); 431863551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 43194eae4efaSPeter Xu hugetlb_count_add(npages, dst); 43201c59827dSHugh Dickins } 4321cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 4322cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 432363551ae0SDavid Gibson } 432463551ae0SDavid Gibson 4325e8569dd2SAndreas Sandberg if (cow) 4326ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 4327c0d0381aSMike Kravetz else 4328c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 4329e8569dd2SAndreas Sandberg 4330e8569dd2SAndreas Sandberg return ret; 433163551ae0SDavid Gibson } 433263551ae0SDavid Gibson 433324669e58SAneesh Kumar K.V void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 433424669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 433524669e58SAneesh Kumar K.V struct page *ref_page) 433663551ae0SDavid Gibson { 433763551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 433863551ae0SDavid Gibson unsigned long address; 4339c7546f8fSDavid Gibson pte_t *ptep; 434063551ae0SDavid Gibson pte_t pte; 4341cb900f41SKirill A. Shutemov spinlock_t *ptl; 434263551ae0SDavid Gibson struct page *page; 4343a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 4344a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 4345ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 4346a5516438SAndi Kleen 434763551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 4348a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 4349a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 435063551ae0SDavid Gibson 435107e32661SAneesh Kumar K.V /* 435207e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 435307e32661SAneesh Kumar K.V * to huge page. 435407e32661SAneesh Kumar K.V */ 4355ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 435624669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 4357dff11abeSMike Kravetz 4358dff11abeSMike Kravetz /* 4359dff11abeSMike Kravetz * If sharing possible, alert mmu notifiers of worst case. 4360dff11abeSMike Kravetz */ 43616f4f13e8SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, 43626f4f13e8SJérôme Glisse end); 4363ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 4364ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4365569f48b8SHillf Danton address = start; 4366569f48b8SHillf Danton for (; address < end; address += sz) { 43677868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, sz); 4368c7546f8fSDavid Gibson if (!ptep) 4369c7546f8fSDavid Gibson continue; 4370c7546f8fSDavid Gibson 4371cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 437234ae204fSMike Kravetz if (huge_pmd_unshare(mm, vma, &address, ptep)) { 437331d49da5SAneesh Kumar K.V spin_unlock(ptl); 4374dff11abeSMike Kravetz /* 4375dff11abeSMike Kravetz * We just unmapped a page of PMDs by clearing a PUD. 4376dff11abeSMike Kravetz * The caller's TLB flush range should cover this area. 4377dff11abeSMike Kravetz */ 437831d49da5SAneesh Kumar K.V continue; 437931d49da5SAneesh Kumar K.V } 438039dde65cSChen, Kenneth W 43816629326bSHillf Danton pte = huge_ptep_get(ptep); 438231d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 438331d49da5SAneesh Kumar K.V spin_unlock(ptl); 438431d49da5SAneesh Kumar K.V continue; 438531d49da5SAneesh Kumar K.V } 43866629326bSHillf Danton 43876629326bSHillf Danton /* 43889fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 43899fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 43906629326bSHillf Danton */ 43919fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 43929386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 439331d49da5SAneesh Kumar K.V spin_unlock(ptl); 439431d49da5SAneesh Kumar K.V continue; 43958c4894c6SNaoya Horiguchi } 43966629326bSHillf Danton 43976629326bSHillf Danton page = pte_page(pte); 439804f2cbe3SMel Gorman /* 439904f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 440004f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 440104f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 440204f2cbe3SMel Gorman */ 440304f2cbe3SMel Gorman if (ref_page) { 440431d49da5SAneesh Kumar K.V if (page != ref_page) { 440531d49da5SAneesh Kumar K.V spin_unlock(ptl); 440631d49da5SAneesh Kumar K.V continue; 440731d49da5SAneesh Kumar K.V } 440804f2cbe3SMel Gorman /* 440904f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 441004f2cbe3SMel Gorman * future faults in this VMA will fail rather than 441104f2cbe3SMel Gorman * looking like data was lost 441204f2cbe3SMel Gorman */ 441304f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 441404f2cbe3SMel Gorman } 441504f2cbe3SMel Gorman 4416c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 4417b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 4418106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 44196649a386SKen Chen set_page_dirty(page); 44209e81130bSHillf Danton 44215d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 4422d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 442331d49da5SAneesh Kumar K.V 4424cb900f41SKirill A. Shutemov spin_unlock(ptl); 4425e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 442624669e58SAneesh Kumar K.V /* 442731d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 442824669e58SAneesh Kumar K.V */ 442931d49da5SAneesh Kumar K.V if (ref_page) 443031d49da5SAneesh Kumar K.V break; 4431fe1668aeSChen, Kenneth W } 4432ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 443324669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 44341da177e4SLinus Torvalds } 443563551ae0SDavid Gibson 4436d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 4437d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 4438d833352aSMel Gorman unsigned long end, struct page *ref_page) 4439d833352aSMel Gorman { 4440d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 4441d833352aSMel Gorman 4442d833352aSMel Gorman /* 4443d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 4444d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 4445d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 4446d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 4447d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 4448c8c06efaSDavidlohr Bueso * is to clear it before releasing the i_mmap_rwsem. This works 4449d833352aSMel Gorman * because in the context this is called, the VMA is about to be 4450c8c06efaSDavidlohr Bueso * destroyed and the i_mmap_rwsem is held. 4451d833352aSMel Gorman */ 4452d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 4453d833352aSMel Gorman } 4454d833352aSMel Gorman 4455502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 445604f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 4457502717f4SChen, Kenneth W { 445824669e58SAneesh Kumar K.V struct mmu_gather tlb; 4459dff11abeSMike Kravetz 4460a72afd87SWill Deacon tlb_gather_mmu(&tlb, vma->vm_mm); 446124669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 4462ae8eba8bSWill Deacon tlb_finish_mmu(&tlb); 4463502717f4SChen, Kenneth W } 4464502717f4SChen, Kenneth W 446504f2cbe3SMel Gorman /* 446604f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 4467578b7725SZhiyuan Dai * mapping it owns the reserve page for. The intention is to unmap the page 446804f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 446904f2cbe3SMel Gorman * same region. 447004f2cbe3SMel Gorman */ 44712f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 44722a4b3dedSHarvey Harrison struct page *page, unsigned long address) 447304f2cbe3SMel Gorman { 44747526674dSAdam Litke struct hstate *h = hstate_vma(vma); 447504f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 447604f2cbe3SMel Gorman struct address_space *mapping; 447704f2cbe3SMel Gorman pgoff_t pgoff; 447804f2cbe3SMel Gorman 447904f2cbe3SMel Gorman /* 448004f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 448104f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 448204f2cbe3SMel Gorman */ 44837526674dSAdam Litke address = address & huge_page_mask(h); 448436e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 448536e4f20aSMichal Hocko vma->vm_pgoff; 448693c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 448704f2cbe3SMel Gorman 44884eb2b1dcSMel Gorman /* 44894eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 44904eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 44914eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 44924eb2b1dcSMel Gorman */ 449383cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 44946b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 449504f2cbe3SMel Gorman /* Do not unmap the current VMA */ 449604f2cbe3SMel Gorman if (iter_vma == vma) 449704f2cbe3SMel Gorman continue; 449804f2cbe3SMel Gorman 449904f2cbe3SMel Gorman /* 45002f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 45012f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 45022f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 45032f84a899SMel Gorman */ 45042f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 45052f84a899SMel Gorman continue; 45062f84a899SMel Gorman 45072f84a899SMel Gorman /* 450804f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 450904f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 451004f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 451104f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 451204f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 451304f2cbe3SMel Gorman */ 451404f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 451524669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 451624669e58SAneesh Kumar K.V address + huge_page_size(h), page); 451704f2cbe3SMel Gorman } 451883cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 451904f2cbe3SMel Gorman } 452004f2cbe3SMel Gorman 45210fe6e20bSNaoya Horiguchi /* 45220fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 4523ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 4524ef009b25SMichal Hocko * cannot race with other handlers or page migration. 4525ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 45260fe6e20bSNaoya Horiguchi */ 45272b740303SSouptick Joarder static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 4528974e6d66SHuang Ying unsigned long address, pte_t *ptep, 4529cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 45301e8f889bSDavid Gibson { 45313999f52eSAneesh Kumar K.V pte_t pte; 4532a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 45331e8f889bSDavid Gibson struct page *old_page, *new_page; 45342b740303SSouptick Joarder int outside_reserve = 0; 45352b740303SSouptick Joarder vm_fault_t ret = 0; 4536974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 4537ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 45381e8f889bSDavid Gibson 45393999f52eSAneesh Kumar K.V pte = huge_ptep_get(ptep); 45401e8f889bSDavid Gibson old_page = pte_page(pte); 45411e8f889bSDavid Gibson 454204f2cbe3SMel Gorman retry_avoidcopy: 45431e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 45441e8f889bSDavid Gibson * and just make the page writable */ 454537a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 45465a49973dSHugh Dickins page_move_anon_rmap(old_page, vma); 45475b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 454883c54070SNick Piggin return 0; 45491e8f889bSDavid Gibson } 45501e8f889bSDavid Gibson 455104f2cbe3SMel Gorman /* 455204f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 455304f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 455404f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 455504f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 455604f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 455704f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 455804f2cbe3SMel Gorman * of the full address range. 455904f2cbe3SMel Gorman */ 45605944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 456104f2cbe3SMel Gorman old_page != pagecache_page) 456204f2cbe3SMel Gorman outside_reserve = 1; 456304f2cbe3SMel Gorman 456409cbfeafSKirill A. Shutemov get_page(old_page); 4565b76c8cfbSLarry Woodman 4566ad4404a2SDavidlohr Bueso /* 4567ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 4568ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 4569ad4404a2SDavidlohr Bueso */ 4570cb900f41SKirill A. Shutemov spin_unlock(ptl); 45715b7a1d40SHuang Ying new_page = alloc_huge_page(vma, haddr, outside_reserve); 45721e8f889bSDavid Gibson 45732fc39cecSAdam Litke if (IS_ERR(new_page)) { 457404f2cbe3SMel Gorman /* 457504f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 457604f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 457704f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 457804f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 457904f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 458004f2cbe3SMel Gorman */ 458104f2cbe3SMel Gorman if (outside_reserve) { 4582e7dd91c4SMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 4583e7dd91c4SMike Kravetz pgoff_t idx; 4584e7dd91c4SMike Kravetz u32 hash; 4585e7dd91c4SMike Kravetz 458609cbfeafSKirill A. Shutemov put_page(old_page); 458704f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 4588e7dd91c4SMike Kravetz /* 4589e7dd91c4SMike Kravetz * Drop hugetlb_fault_mutex and i_mmap_rwsem before 4590e7dd91c4SMike Kravetz * unmapping. unmapping needs to hold i_mmap_rwsem 4591e7dd91c4SMike Kravetz * in write mode. Dropping i_mmap_rwsem in read mode 4592e7dd91c4SMike Kravetz * here is OK as COW mappings do not interact with 4593e7dd91c4SMike Kravetz * PMD sharing. 4594e7dd91c4SMike Kravetz * 4595e7dd91c4SMike Kravetz * Reacquire both after unmap operation. 4596e7dd91c4SMike Kravetz */ 4597e7dd91c4SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 4598e7dd91c4SMike Kravetz hash = hugetlb_fault_mutex_hash(mapping, idx); 4599e7dd91c4SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 4600e7dd91c4SMike Kravetz i_mmap_unlock_read(mapping); 4601e7dd91c4SMike Kravetz 46025b7a1d40SHuang Ying unmap_ref_private(mm, vma, old_page, haddr); 4603e7dd91c4SMike Kravetz 4604e7dd91c4SMike Kravetz i_mmap_lock_read(mapping); 4605e7dd91c4SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 4606cb900f41SKirill A. Shutemov spin_lock(ptl); 46075b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4608a9af0c5dSNaoya Horiguchi if (likely(ptep && 4609a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 461004f2cbe3SMel Gorman goto retry_avoidcopy; 4611a734bcc8SHillf Danton /* 4612cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 4613cb900f41SKirill A. Shutemov * lock, and our job is done. 4614a734bcc8SHillf Danton */ 4615a734bcc8SHillf Danton return 0; 461604f2cbe3SMel Gorman } 461704f2cbe3SMel Gorman 46182b740303SSouptick Joarder ret = vmf_error(PTR_ERR(new_page)); 4619ad4404a2SDavidlohr Bueso goto out_release_old; 46201e8f889bSDavid Gibson } 46211e8f889bSDavid Gibson 46220fe6e20bSNaoya Horiguchi /* 46230fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 46240fe6e20bSNaoya Horiguchi * anon_vma prepared. 46250fe6e20bSNaoya Horiguchi */ 462644e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 4627ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 4628ad4404a2SDavidlohr Bueso goto out_release_all; 462944e2aa93SDean Nelson } 46300fe6e20bSNaoya Horiguchi 4631974e6d66SHuang Ying copy_user_huge_page(new_page, old_page, address, vma, 463247ad8475SAndrea Arcangeli pages_per_huge_page(h)); 46330ed361deSNick Piggin __SetPageUptodate(new_page); 46341e8f889bSDavid Gibson 46357269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, 46366f4f13e8SJérôme Glisse haddr + huge_page_size(h)); 4637ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4638ad4404a2SDavidlohr Bueso 4639b76c8cfbSLarry Woodman /* 4640cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 4641b76c8cfbSLarry Woodman * before the page tables are altered 4642b76c8cfbSLarry Woodman */ 4643cb900f41SKirill A. Shutemov spin_lock(ptl); 46445b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4645a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 4646d6995da3SMike Kravetz ClearHPageRestoreReserve(new_page); 464707443a85SJoonsoo Kim 46481e8f889bSDavid Gibson /* Break COW */ 46495b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 4650ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, range.end); 46515b7a1d40SHuang Ying set_huge_pte_at(mm, haddr, ptep, 46521e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 4653d281ee61SKirill A. Shutemov page_remove_rmap(old_page, true); 46545b7a1d40SHuang Ying hugepage_add_new_anon_rmap(new_page, vma, haddr); 46558f251a3dSMike Kravetz SetHPageMigratable(new_page); 46561e8f889bSDavid Gibson /* Make the old page be freed below */ 46571e8f889bSDavid Gibson new_page = old_page; 46581e8f889bSDavid Gibson } 4659cb900f41SKirill A. Shutemov spin_unlock(ptl); 4660ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 4661ad4404a2SDavidlohr Bueso out_release_all: 46625b7a1d40SHuang Ying restore_reserve_on_error(h, vma, haddr, new_page); 466309cbfeafSKirill A. Shutemov put_page(new_page); 4664ad4404a2SDavidlohr Bueso out_release_old: 466509cbfeafSKirill A. Shutemov put_page(old_page); 46668312034fSJoonsoo Kim 4667ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 4668ad4404a2SDavidlohr Bueso return ret; 46691e8f889bSDavid Gibson } 46701e8f889bSDavid Gibson 467104f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 4672a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 4673a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 467404f2cbe3SMel Gorman { 467504f2cbe3SMel Gorman struct address_space *mapping; 4676e7c4b0bfSAndy Whitcroft pgoff_t idx; 467704f2cbe3SMel Gorman 467804f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 4679a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 468004f2cbe3SMel Gorman 468104f2cbe3SMel Gorman return find_lock_page(mapping, idx); 468204f2cbe3SMel Gorman } 468304f2cbe3SMel Gorman 46843ae77f43SHugh Dickins /* 46853ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 46863ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 46873ae77f43SHugh Dickins */ 46883ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 46892a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 46902a15efc9SHugh Dickins { 46912a15efc9SHugh Dickins struct address_space *mapping; 46922a15efc9SHugh Dickins pgoff_t idx; 46932a15efc9SHugh Dickins struct page *page; 46942a15efc9SHugh Dickins 46952a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 46962a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 46972a15efc9SHugh Dickins 46982a15efc9SHugh Dickins page = find_get_page(mapping, idx); 46992a15efc9SHugh Dickins if (page) 47002a15efc9SHugh Dickins put_page(page); 47012a15efc9SHugh Dickins return page != NULL; 47022a15efc9SHugh Dickins } 47032a15efc9SHugh Dickins 4704ab76ad54SMike Kravetz int huge_add_to_page_cache(struct page *page, struct address_space *mapping, 4705ab76ad54SMike Kravetz pgoff_t idx) 4706ab76ad54SMike Kravetz { 4707ab76ad54SMike Kravetz struct inode *inode = mapping->host; 4708ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 4709ab76ad54SMike Kravetz int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 4710ab76ad54SMike Kravetz 4711ab76ad54SMike Kravetz if (err) 4712ab76ad54SMike Kravetz return err; 4713d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 4714ab76ad54SMike Kravetz 471522146c3cSMike Kravetz /* 471622146c3cSMike Kravetz * set page dirty so that it will not be removed from cache/file 471722146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 471822146c3cSMike Kravetz */ 471922146c3cSMike Kravetz set_page_dirty(page); 472022146c3cSMike Kravetz 4721ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 4722ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 4723ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 4724ab76ad54SMike Kravetz return 0; 4725ab76ad54SMike Kravetz } 4726ab76ad54SMike Kravetz 47277677f7fdSAxel Rasmussen static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma, 47287677f7fdSAxel Rasmussen struct address_space *mapping, 47297677f7fdSAxel Rasmussen pgoff_t idx, 47307677f7fdSAxel Rasmussen unsigned int flags, 47317677f7fdSAxel Rasmussen unsigned long haddr, 47327677f7fdSAxel Rasmussen unsigned long reason) 47337677f7fdSAxel Rasmussen { 47347677f7fdSAxel Rasmussen vm_fault_t ret; 47357677f7fdSAxel Rasmussen u32 hash; 47367677f7fdSAxel Rasmussen struct vm_fault vmf = { 47377677f7fdSAxel Rasmussen .vma = vma, 47387677f7fdSAxel Rasmussen .address = haddr, 47397677f7fdSAxel Rasmussen .flags = flags, 47407677f7fdSAxel Rasmussen 47417677f7fdSAxel Rasmussen /* 47427677f7fdSAxel Rasmussen * Hard to debug if it ends up being 47437677f7fdSAxel Rasmussen * used by a callee that assumes 47447677f7fdSAxel Rasmussen * something about the other 47457677f7fdSAxel Rasmussen * uninitialized fields... same as in 47467677f7fdSAxel Rasmussen * memory.c 47477677f7fdSAxel Rasmussen */ 47487677f7fdSAxel Rasmussen }; 47497677f7fdSAxel Rasmussen 47507677f7fdSAxel Rasmussen /* 47517677f7fdSAxel Rasmussen * hugetlb_fault_mutex and i_mmap_rwsem must be 47527677f7fdSAxel Rasmussen * dropped before handling userfault. Reacquire 47537677f7fdSAxel Rasmussen * after handling fault to make calling code simpler. 47547677f7fdSAxel Rasmussen */ 47557677f7fdSAxel Rasmussen hash = hugetlb_fault_mutex_hash(mapping, idx); 47567677f7fdSAxel Rasmussen mutex_unlock(&hugetlb_fault_mutex_table[hash]); 47577677f7fdSAxel Rasmussen i_mmap_unlock_read(mapping); 47587677f7fdSAxel Rasmussen ret = handle_userfault(&vmf, reason); 47597677f7fdSAxel Rasmussen i_mmap_lock_read(mapping); 47607677f7fdSAxel Rasmussen mutex_lock(&hugetlb_fault_mutex_table[hash]); 47617677f7fdSAxel Rasmussen 47627677f7fdSAxel Rasmussen return ret; 47637677f7fdSAxel Rasmussen } 47647677f7fdSAxel Rasmussen 47652b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 47662b740303SSouptick Joarder struct vm_area_struct *vma, 47678382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 4768788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 4769ac9b9c66SHugh Dickins { 4770a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 47712b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 4772409eb8c2SHillf Danton int anon_rmap = 0; 47734c887265SAdam Litke unsigned long size; 47744c887265SAdam Litke struct page *page; 47751e8f889bSDavid Gibson pte_t new_pte; 4776cb900f41SKirill A. Shutemov spinlock_t *ptl; 4777285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 4778cb6acd01SMike Kravetz bool new_page = false; 47794c887265SAdam Litke 478004f2cbe3SMel Gorman /* 478104f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 478204f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 478325985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 478404f2cbe3SMel Gorman */ 478504f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 4786910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 478704f2cbe3SMel Gorman current->pid); 478804f2cbe3SMel Gorman return ret; 478904f2cbe3SMel Gorman } 479004f2cbe3SMel Gorman 47914c887265SAdam Litke /* 479287bf91d3SMike Kravetz * We can not race with truncation due to holding i_mmap_rwsem. 479387bf91d3SMike Kravetz * i_size is modified when holding i_mmap_rwsem, so check here 479487bf91d3SMike Kravetz * once for faults beyond end of file. 47954c887265SAdam Litke */ 4796a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 4797ebed4bfcSHugh Dickins if (idx >= size) 4798ebed4bfcSHugh Dickins goto out; 47991a1aad8aSMike Kravetz 480087bf91d3SMike Kravetz retry: 480187bf91d3SMike Kravetz page = find_lock_page(mapping, idx); 480287bf91d3SMike Kravetz if (!page) { 48037677f7fdSAxel Rasmussen /* Check for page in userfault range */ 48041a1aad8aSMike Kravetz if (userfaultfd_missing(vma)) { 48057677f7fdSAxel Rasmussen ret = hugetlb_handle_userfault(vma, mapping, idx, 48067677f7fdSAxel Rasmussen flags, haddr, 48077677f7fdSAxel Rasmussen VM_UFFD_MISSING); 48081a1aad8aSMike Kravetz goto out; 48091a1aad8aSMike Kravetz } 48101a1aad8aSMike Kravetz 4811285b8dcaSHuang Ying page = alloc_huge_page(vma, haddr, 0); 48122fc39cecSAdam Litke if (IS_ERR(page)) { 48134643d67eSMike Kravetz /* 48144643d67eSMike Kravetz * Returning error will result in faulting task being 48154643d67eSMike Kravetz * sent SIGBUS. The hugetlb fault mutex prevents two 48164643d67eSMike Kravetz * tasks from racing to fault in the same page which 48174643d67eSMike Kravetz * could result in false unable to allocate errors. 48184643d67eSMike Kravetz * Page migration does not take the fault mutex, but 48194643d67eSMike Kravetz * does a clear then write of pte's under page table 48204643d67eSMike Kravetz * lock. Page fault code could race with migration, 48214643d67eSMike Kravetz * notice the clear pte and try to allocate a page 48224643d67eSMike Kravetz * here. Before returning error, get ptl and make 48234643d67eSMike Kravetz * sure there really is no pte entry. 48244643d67eSMike Kravetz */ 48254643d67eSMike Kravetz ptl = huge_pte_lock(h, mm, ptep); 48264643d67eSMike Kravetz ret = 0; 4827d83e6c8aSMiaohe Lin if (huge_pte_none(huge_ptep_get(ptep))) 48282b740303SSouptick Joarder ret = vmf_error(PTR_ERR(page)); 4829d83e6c8aSMiaohe Lin spin_unlock(ptl); 48306bda666aSChristoph Lameter goto out; 48316bda666aSChristoph Lameter } 483247ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 48330ed361deSNick Piggin __SetPageUptodate(page); 4834cb6acd01SMike Kravetz new_page = true; 4835ac9b9c66SHugh Dickins 4836f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 4837ab76ad54SMike Kravetz int err = huge_add_to_page_cache(page, mapping, idx); 48386bda666aSChristoph Lameter if (err) { 48396bda666aSChristoph Lameter put_page(page); 48406bda666aSChristoph Lameter if (err == -EEXIST) 48416bda666aSChristoph Lameter goto retry; 48426bda666aSChristoph Lameter goto out; 48436bda666aSChristoph Lameter } 484423be7468SMel Gorman } else { 48456bda666aSChristoph Lameter lock_page(page); 48460fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 48470fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 48480fe6e20bSNaoya Horiguchi goto backout_unlocked; 484923be7468SMel Gorman } 4850409eb8c2SHillf Danton anon_rmap = 1; 48510fe6e20bSNaoya Horiguchi } 48520fe6e20bSNaoya Horiguchi } else { 485357303d80SAndy Whitcroft /* 4854998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 4855998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 4856998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 4857fd6a03edSNaoya Horiguchi */ 4858fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 48590eb98f15SMiaohe Lin ret = VM_FAULT_HWPOISON_LARGE | 4860972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 4861fd6a03edSNaoya Horiguchi goto backout_unlocked; 48626bda666aSChristoph Lameter } 48637677f7fdSAxel Rasmussen 48647677f7fdSAxel Rasmussen /* Check for page in userfault range. */ 48657677f7fdSAxel Rasmussen if (userfaultfd_minor(vma)) { 48667677f7fdSAxel Rasmussen unlock_page(page); 48677677f7fdSAxel Rasmussen put_page(page); 48687677f7fdSAxel Rasmussen ret = hugetlb_handle_userfault(vma, mapping, idx, 48697677f7fdSAxel Rasmussen flags, haddr, 48707677f7fdSAxel Rasmussen VM_UFFD_MINOR); 48717677f7fdSAxel Rasmussen goto out; 48727677f7fdSAxel Rasmussen } 4873998b4382SNaoya Horiguchi } 48741e8f889bSDavid Gibson 487557303d80SAndy Whitcroft /* 487657303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 487757303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 487857303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 487957303d80SAndy Whitcroft * the spinlock. 488057303d80SAndy Whitcroft */ 48815e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 4882285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 48832b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 48842b26736cSAndy Whitcroft goto backout_unlocked; 48852b26736cSAndy Whitcroft } 48865e911373SMike Kravetz /* Just decrements count, does not deallocate */ 4887285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 48885e911373SMike Kravetz } 488957303d80SAndy Whitcroft 48908bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 489183c54070SNick Piggin ret = 0; 48927f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 48934c887265SAdam Litke goto backout; 48944c887265SAdam Litke 489507443a85SJoonsoo Kim if (anon_rmap) { 4896d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 4897285b8dcaSHuang Ying hugepage_add_new_anon_rmap(page, vma, haddr); 4898ac714904SChoi Gi-yong } else 489953f9263bSKirill A. Shutemov page_dup_rmap(page, true); 49001e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 49011e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 4902285b8dcaSHuang Ying set_huge_pte_at(mm, haddr, ptep, new_pte); 49031e8f889bSDavid Gibson 49045d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 4905788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 49061e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 4907974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); 49081e8f889bSDavid Gibson } 49091e8f889bSDavid Gibson 4910cb900f41SKirill A. Shutemov spin_unlock(ptl); 4911cb6acd01SMike Kravetz 4912cb6acd01SMike Kravetz /* 49138f251a3dSMike Kravetz * Only set HPageMigratable in newly allocated pages. Existing pages 49148f251a3dSMike Kravetz * found in the pagecache may not have HPageMigratableset if they have 49158f251a3dSMike Kravetz * been isolated for migration. 4916cb6acd01SMike Kravetz */ 4917cb6acd01SMike Kravetz if (new_page) 49188f251a3dSMike Kravetz SetHPageMigratable(page); 4919cb6acd01SMike Kravetz 49204c887265SAdam Litke unlock_page(page); 49214c887265SAdam Litke out: 4922ac9b9c66SHugh Dickins return ret; 49234c887265SAdam Litke 49244c887265SAdam Litke backout: 4925cb900f41SKirill A. Shutemov spin_unlock(ptl); 49262b26736cSAndy Whitcroft backout_unlocked: 49274c887265SAdam Litke unlock_page(page); 4928285b8dcaSHuang Ying restore_reserve_on_error(h, vma, haddr, page); 49294c887265SAdam Litke put_page(page); 49304c887265SAdam Litke goto out; 4931ac9b9c66SHugh Dickins } 4932ac9b9c66SHugh Dickins 49338382d914SDavidlohr Bueso #ifdef CONFIG_SMP 4934188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 49358382d914SDavidlohr Bueso { 49368382d914SDavidlohr Bueso unsigned long key[2]; 49378382d914SDavidlohr Bueso u32 hash; 49388382d914SDavidlohr Bueso 49398382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 49408382d914SDavidlohr Bueso key[1] = idx; 49418382d914SDavidlohr Bueso 494255254636SMike Kravetz hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); 49438382d914SDavidlohr Bueso 49448382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 49458382d914SDavidlohr Bueso } 49468382d914SDavidlohr Bueso #else 49478382d914SDavidlohr Bueso /* 49486c26d310SMiaohe Lin * For uniprocessor systems we always use a single mutex, so just 49498382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 49508382d914SDavidlohr Bueso */ 4951188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 49528382d914SDavidlohr Bueso { 49538382d914SDavidlohr Bueso return 0; 49548382d914SDavidlohr Bueso } 49558382d914SDavidlohr Bueso #endif 49568382d914SDavidlohr Bueso 49572b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 4958788c7df4SHugh Dickins unsigned long address, unsigned int flags) 495986e5216fSAdam Litke { 49608382d914SDavidlohr Bueso pte_t *ptep, entry; 4961cb900f41SKirill A. Shutemov spinlock_t *ptl; 49622b740303SSouptick Joarder vm_fault_t ret; 49638382d914SDavidlohr Bueso u32 hash; 49648382d914SDavidlohr Bueso pgoff_t idx; 49650fe6e20bSNaoya Horiguchi struct page *page = NULL; 496657303d80SAndy Whitcroft struct page *pagecache_page = NULL; 4967a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 49688382d914SDavidlohr Bueso struct address_space *mapping; 49690f792cf9SNaoya Horiguchi int need_wait_lock = 0; 4970285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 497186e5216fSAdam Litke 4972285b8dcaSHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4973fd6a03edSNaoya Horiguchi if (ptep) { 4974c0d0381aSMike Kravetz /* 4975c0d0381aSMike Kravetz * Since we hold no locks, ptep could be stale. That is 4976c0d0381aSMike Kravetz * OK as we are only making decisions based on content and 4977c0d0381aSMike Kravetz * not actually modifying content here. 4978c0d0381aSMike Kravetz */ 4979fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 4980290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 4981cb900f41SKirill A. Shutemov migration_entry_wait_huge(vma, mm, ptep); 4982290408d4SNaoya Horiguchi return 0; 4983290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 4984aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 4985972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 4986b43a9990SMike Kravetz } 49878382d914SDavidlohr Bueso 4988c0d0381aSMike Kravetz /* 4989c0d0381aSMike Kravetz * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold 499087bf91d3SMike Kravetz * until finished with ptep. This serves two purposes: 499187bf91d3SMike Kravetz * 1) It prevents huge_pmd_unshare from being called elsewhere 499287bf91d3SMike Kravetz * and making the ptep no longer valid. 499387bf91d3SMike Kravetz * 2) It synchronizes us with i_size modifications during truncation. 4994c0d0381aSMike Kravetz * 4995c0d0381aSMike Kravetz * ptep could have already be assigned via huge_pte_offset. That 4996c0d0381aSMike Kravetz * is OK, as huge_pte_alloc will return the same value unless 4997c0d0381aSMike Kravetz * something has changed. 4998c0d0381aSMike Kravetz */ 4999ddeaab32SMike Kravetz mapping = vma->vm_file->f_mapping; 5000c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 5001aec44e0fSPeter Xu ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); 5002c0d0381aSMike Kravetz if (!ptep) { 5003c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 5004c0d0381aSMike Kravetz return VM_FAULT_OOM; 5005c0d0381aSMike Kravetz } 5006ddeaab32SMike Kravetz 50073935baa9SDavid Gibson /* 50083935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 50093935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 50103935baa9SDavid Gibson * the same page in the page cache. 50113935baa9SDavid Gibson */ 5012c0d0381aSMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 5013188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 5014c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 50158382d914SDavidlohr Bueso 50167f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 50177f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 50188382d914SDavidlohr Bueso ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 5019b4d1d99fSDavid Gibson goto out_mutex; 50203935baa9SDavid Gibson } 502186e5216fSAdam Litke 502283c54070SNick Piggin ret = 0; 50231e8f889bSDavid Gibson 502457303d80SAndy Whitcroft /* 50250f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 50260f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 50277c8de358SEthon Paul * an active hugepage in pagecache. This goto expects the 2nd page 50287c8de358SEthon Paul * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will 50297c8de358SEthon Paul * properly handle it. 50300f792cf9SNaoya Horiguchi */ 50310f792cf9SNaoya Horiguchi if (!pte_present(entry)) 50320f792cf9SNaoya Horiguchi goto out_mutex; 50330f792cf9SNaoya Horiguchi 50340f792cf9SNaoya Horiguchi /* 503557303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 503657303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 503757303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 503857303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 503957303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 504057303d80SAndy Whitcroft * consumed. 504157303d80SAndy Whitcroft */ 5042106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 5043285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 50442b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 5045b4d1d99fSDavid Gibson goto out_mutex; 50462b26736cSAndy Whitcroft } 50475e911373SMike Kravetz /* Just decrements count, does not deallocate */ 5048285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 504957303d80SAndy Whitcroft 5050f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 505157303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 5052285b8dcaSHuang Ying vma, haddr); 505357303d80SAndy Whitcroft } 505457303d80SAndy Whitcroft 50550f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 50560fe6e20bSNaoya Horiguchi 50571e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 5058b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 5059cb900f41SKirill A. Shutemov goto out_ptl; 5060b4d1d99fSDavid Gibson 50610f792cf9SNaoya Horiguchi /* 50620f792cf9SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 50630f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 50640f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 50650f792cf9SNaoya Horiguchi */ 50660f792cf9SNaoya Horiguchi page = pte_page(entry); 50670f792cf9SNaoya Horiguchi if (page != pagecache_page) 50680f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 50690f792cf9SNaoya Horiguchi need_wait_lock = 1; 50700f792cf9SNaoya Horiguchi goto out_ptl; 50710f792cf9SNaoya Horiguchi } 50720f792cf9SNaoya Horiguchi 50730f792cf9SNaoya Horiguchi get_page(page); 5074b4d1d99fSDavid Gibson 5075788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 5076106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 5077974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, 5078cb900f41SKirill A. Shutemov pagecache_page, ptl); 50790f792cf9SNaoya Horiguchi goto out_put_page; 5080b4d1d99fSDavid Gibson } 5081106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 5082b4d1d99fSDavid Gibson } 5083b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 5084285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 5085788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 5086285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 50870f792cf9SNaoya Horiguchi out_put_page: 50880f792cf9SNaoya Horiguchi if (page != pagecache_page) 50890f792cf9SNaoya Horiguchi unlock_page(page); 50900f792cf9SNaoya Horiguchi put_page(page); 5091cb900f41SKirill A. Shutemov out_ptl: 5092cb900f41SKirill A. Shutemov spin_unlock(ptl); 509357303d80SAndy Whitcroft 509457303d80SAndy Whitcroft if (pagecache_page) { 509557303d80SAndy Whitcroft unlock_page(pagecache_page); 509657303d80SAndy Whitcroft put_page(pagecache_page); 509757303d80SAndy Whitcroft } 5098b4d1d99fSDavid Gibson out_mutex: 5099c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5100c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 51010f792cf9SNaoya Horiguchi /* 51020f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 51030f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 51040f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 51050f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 51060f792cf9SNaoya Horiguchi * here without taking refcount. 51070f792cf9SNaoya Horiguchi */ 51080f792cf9SNaoya Horiguchi if (need_wait_lock) 51090f792cf9SNaoya Horiguchi wait_on_page_locked(page); 51101e8f889bSDavid Gibson return ret; 511186e5216fSAdam Litke } 511286e5216fSAdam Litke 5113714c1891SAxel Rasmussen #ifdef CONFIG_USERFAULTFD 51148fb5debcSMike Kravetz /* 51158fb5debcSMike Kravetz * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with 51168fb5debcSMike Kravetz * modifications for huge pages. 51178fb5debcSMike Kravetz */ 51188fb5debcSMike Kravetz int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, 51198fb5debcSMike Kravetz pte_t *dst_pte, 51208fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 51218fb5debcSMike Kravetz unsigned long dst_addr, 51228fb5debcSMike Kravetz unsigned long src_addr, 5123f6191471SAxel Rasmussen enum mcopy_atomic_mode mode, 51248fb5debcSMike Kravetz struct page **pagep) 51258fb5debcSMike Kravetz { 5126f6191471SAxel Rasmussen bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE); 51278cc5fcbbSMina Almasry struct hstate *h = hstate_vma(dst_vma); 51288cc5fcbbSMina Almasry struct address_space *mapping = dst_vma->vm_file->f_mapping; 51298cc5fcbbSMina Almasry pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr); 51301e392147SAndrea Arcangeli unsigned long size; 51311c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 51328fb5debcSMike Kravetz pte_t _dst_pte; 51338fb5debcSMike Kravetz spinlock_t *ptl; 51348cc5fcbbSMina Almasry int ret = -ENOMEM; 51358fb5debcSMike Kravetz struct page *page; 5136f6191471SAxel Rasmussen int writable; 51378fb5debcSMike Kravetz 5138f6191471SAxel Rasmussen if (is_continue) { 5139f6191471SAxel Rasmussen ret = -EFAULT; 5140f6191471SAxel Rasmussen page = find_lock_page(mapping, idx); 5141f6191471SAxel Rasmussen if (!page) 5142f6191471SAxel Rasmussen goto out; 5143f6191471SAxel Rasmussen } else if (!*pagep) { 5144d84cf06eSMina Almasry /* If a page already exists, then it's UFFDIO_COPY for 5145d84cf06eSMina Almasry * a non-missing case. Return -EEXIST. 5146d84cf06eSMina Almasry */ 5147d84cf06eSMina Almasry if (vm_shared && 5148d84cf06eSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 5149d84cf06eSMina Almasry ret = -EEXIST; 51508fb5debcSMike Kravetz goto out; 5151d84cf06eSMina Almasry } 5152d84cf06eSMina Almasry 5153d84cf06eSMina Almasry page = alloc_huge_page(dst_vma, dst_addr, 0); 5154d84cf06eSMina Almasry if (IS_ERR(page)) { 5155d84cf06eSMina Almasry ret = -ENOMEM; 5156d84cf06eSMina Almasry goto out; 5157d84cf06eSMina Almasry } 51588fb5debcSMike Kravetz 51598fb5debcSMike Kravetz ret = copy_huge_page_from_user(page, 51608fb5debcSMike Kravetz (const void __user *) src_addr, 5161810a56b9SMike Kravetz pages_per_huge_page(h), false); 51628fb5debcSMike Kravetz 5163c1e8d7c6SMichel Lespinasse /* fallback to copy_from_user outside mmap_lock */ 51648fb5debcSMike Kravetz if (unlikely(ret)) { 51659e368259SAndrea Arcangeli ret = -ENOENT; 51668cc5fcbbSMina Almasry /* Free the allocated page which may have 51678cc5fcbbSMina Almasry * consumed a reservation. 51688cc5fcbbSMina Almasry */ 51698cc5fcbbSMina Almasry restore_reserve_on_error(h, dst_vma, dst_addr, page); 51708cc5fcbbSMina Almasry put_page(page); 51718cc5fcbbSMina Almasry 51728cc5fcbbSMina Almasry /* Allocate a temporary page to hold the copied 51738cc5fcbbSMina Almasry * contents. 51748cc5fcbbSMina Almasry */ 51758cc5fcbbSMina Almasry page = alloc_huge_page_vma(h, dst_vma, dst_addr); 51768cc5fcbbSMina Almasry if (!page) { 51778cc5fcbbSMina Almasry ret = -ENOMEM; 51788cc5fcbbSMina Almasry goto out; 51798cc5fcbbSMina Almasry } 51808fb5debcSMike Kravetz *pagep = page; 51818cc5fcbbSMina Almasry /* Set the outparam pagep and return to the caller to 51828cc5fcbbSMina Almasry * copy the contents outside the lock. Don't free the 51838cc5fcbbSMina Almasry * page. 51848cc5fcbbSMina Almasry */ 51858fb5debcSMike Kravetz goto out; 51868fb5debcSMike Kravetz } 51878fb5debcSMike Kravetz } else { 51888cc5fcbbSMina Almasry if (vm_shared && 51898cc5fcbbSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 51908cc5fcbbSMina Almasry put_page(*pagep); 51918cc5fcbbSMina Almasry ret = -EEXIST; 51928cc5fcbbSMina Almasry *pagep = NULL; 51938cc5fcbbSMina Almasry goto out; 51948cc5fcbbSMina Almasry } 51958cc5fcbbSMina Almasry 51968cc5fcbbSMina Almasry page = alloc_huge_page(dst_vma, dst_addr, 0); 51978cc5fcbbSMina Almasry if (IS_ERR(page)) { 51988cc5fcbbSMina Almasry ret = -ENOMEM; 51998cc5fcbbSMina Almasry *pagep = NULL; 52008cc5fcbbSMina Almasry goto out; 52018cc5fcbbSMina Almasry } 52028cc5fcbbSMina Almasry copy_huge_page(page, *pagep); 52038cc5fcbbSMina Almasry put_page(*pagep); 52048fb5debcSMike Kravetz *pagep = NULL; 52058fb5debcSMike Kravetz } 52068fb5debcSMike Kravetz 52078fb5debcSMike Kravetz /* 52088fb5debcSMike Kravetz * The memory barrier inside __SetPageUptodate makes sure that 52098fb5debcSMike Kravetz * preceding stores to the page contents become visible before 52108fb5debcSMike Kravetz * the set_pte_at() write. 52118fb5debcSMike Kravetz */ 52128fb5debcSMike Kravetz __SetPageUptodate(page); 52138fb5debcSMike Kravetz 5214f6191471SAxel Rasmussen /* Add shared, newly allocated pages to the page cache. */ 5215f6191471SAxel Rasmussen if (vm_shared && !is_continue) { 52161e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 52171e392147SAndrea Arcangeli ret = -EFAULT; 52181e392147SAndrea Arcangeli if (idx >= size) 52191e392147SAndrea Arcangeli goto out_release_nounlock; 52201c9e8defSMike Kravetz 52211e392147SAndrea Arcangeli /* 52221e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 52231e392147SAndrea Arcangeli * huge_add_to_page_cache() below happens through the 52241e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 52251e392147SAndrea Arcangeli * the caller. 52261e392147SAndrea Arcangeli */ 52271c9e8defSMike Kravetz ret = huge_add_to_page_cache(page, mapping, idx); 52281c9e8defSMike Kravetz if (ret) 52291c9e8defSMike Kravetz goto out_release_nounlock; 52301c9e8defSMike Kravetz } 52311c9e8defSMike Kravetz 52328fb5debcSMike Kravetz ptl = huge_pte_lockptr(h, dst_mm, dst_pte); 52338fb5debcSMike Kravetz spin_lock(ptl); 52348fb5debcSMike Kravetz 52351e392147SAndrea Arcangeli /* 52361e392147SAndrea Arcangeli * Recheck the i_size after holding PT lock to make sure not 52371e392147SAndrea Arcangeli * to leave any page mapped (as page_mapped()) beyond the end 52381e392147SAndrea Arcangeli * of the i_size (remove_inode_hugepages() is strict about 52391e392147SAndrea Arcangeli * enforcing that). If we bail out here, we'll also leave a 52401e392147SAndrea Arcangeli * page in the radix tree in the vm_shared case beyond the end 52411e392147SAndrea Arcangeli * of the i_size, but remove_inode_hugepages() will take care 52421e392147SAndrea Arcangeli * of it as soon as we drop the hugetlb_fault_mutex_table. 52431e392147SAndrea Arcangeli */ 52441e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 52451e392147SAndrea Arcangeli ret = -EFAULT; 52461e392147SAndrea Arcangeli if (idx >= size) 52471e392147SAndrea Arcangeli goto out_release_unlock; 52481e392147SAndrea Arcangeli 52498fb5debcSMike Kravetz ret = -EEXIST; 52508fb5debcSMike Kravetz if (!huge_pte_none(huge_ptep_get(dst_pte))) 52518fb5debcSMike Kravetz goto out_release_unlock; 52528fb5debcSMike Kravetz 52531c9e8defSMike Kravetz if (vm_shared) { 52541c9e8defSMike Kravetz page_dup_rmap(page, true); 52551c9e8defSMike Kravetz } else { 5256d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 52578fb5debcSMike Kravetz hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); 52581c9e8defSMike Kravetz } 52598fb5debcSMike Kravetz 5260f6191471SAxel Rasmussen /* For CONTINUE on a non-shared VMA, don't set VM_WRITE for CoW. */ 5261f6191471SAxel Rasmussen if (is_continue && !vm_shared) 5262f6191471SAxel Rasmussen writable = 0; 5263f6191471SAxel Rasmussen else 5264f6191471SAxel Rasmussen writable = dst_vma->vm_flags & VM_WRITE; 5265f6191471SAxel Rasmussen 5266f6191471SAxel Rasmussen _dst_pte = make_huge_pte(dst_vma, page, writable); 5267f6191471SAxel Rasmussen if (writable) 52688fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 52698fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 52708fb5debcSMike Kravetz 52718fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 52728fb5debcSMike Kravetz 52738fb5debcSMike Kravetz (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, 52748fb5debcSMike Kravetz dst_vma->vm_flags & VM_WRITE); 52758fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 52768fb5debcSMike Kravetz 52778fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 52788fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 52798fb5debcSMike Kravetz 52808fb5debcSMike Kravetz spin_unlock(ptl); 5281f6191471SAxel Rasmussen if (!is_continue) 52828f251a3dSMike Kravetz SetHPageMigratable(page); 5283f6191471SAxel Rasmussen if (vm_shared || is_continue) 52841c9e8defSMike Kravetz unlock_page(page); 52858fb5debcSMike Kravetz ret = 0; 52868fb5debcSMike Kravetz out: 52878fb5debcSMike Kravetz return ret; 52888fb5debcSMike Kravetz out_release_unlock: 52898fb5debcSMike Kravetz spin_unlock(ptl); 5290f6191471SAxel Rasmussen if (vm_shared || is_continue) 52911c9e8defSMike Kravetz unlock_page(page); 52925af10dfdSAndrea Arcangeli out_release_nounlock: 5293846be085SMike Kravetz restore_reserve_on_error(h, dst_vma, dst_addr, page); 52948fb5debcSMike Kravetz put_page(page); 52958fb5debcSMike Kravetz goto out; 52968fb5debcSMike Kravetz } 5297714c1891SAxel Rasmussen #endif /* CONFIG_USERFAULTFD */ 52988fb5debcSMike Kravetz 529982e5d378SJoao Martins static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma, 530082e5d378SJoao Martins int refs, struct page **pages, 530182e5d378SJoao Martins struct vm_area_struct **vmas) 530282e5d378SJoao Martins { 530382e5d378SJoao Martins int nr; 530482e5d378SJoao Martins 530582e5d378SJoao Martins for (nr = 0; nr < refs; nr++) { 530682e5d378SJoao Martins if (likely(pages)) 530782e5d378SJoao Martins pages[nr] = mem_map_offset(page, nr); 530882e5d378SJoao Martins if (vmas) 530982e5d378SJoao Martins vmas[nr] = vma; 531082e5d378SJoao Martins } 531182e5d378SJoao Martins } 531282e5d378SJoao Martins 531328a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 531463551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 531528a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 53164f6da934SPeter Xu long i, unsigned int flags, int *locked) 531763551ae0SDavid Gibson { 5318d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 5319d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 532028a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 5321a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 53220fa5bc40SJoao Martins int err = -EFAULT, refs; 532363551ae0SDavid Gibson 532463551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 532563551ae0SDavid Gibson pte_t *pte; 5326cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 53272a15efc9SHugh Dickins int absent; 532863551ae0SDavid Gibson struct page *page; 532963551ae0SDavid Gibson 53304c887265SAdam Litke /* 533102057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 533202057967SDavid Rientjes * potentially allocating memory. 533302057967SDavid Rientjes */ 5334fa45f116SDavidlohr Bueso if (fatal_signal_pending(current)) { 533502057967SDavid Rientjes remainder = 0; 533602057967SDavid Rientjes break; 533702057967SDavid Rientjes } 533802057967SDavid Rientjes 533902057967SDavid Rientjes /* 53404c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 53412a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 53424c887265SAdam Litke * first, for the page indexing below to work. 5343cb900f41SKirill A. Shutemov * 5344cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 53454c887265SAdam Litke */ 53467868a208SPunit Agrawal pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), 53477868a208SPunit Agrawal huge_page_size(h)); 5348cb900f41SKirill A. Shutemov if (pte) 5349cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 53502a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 535163551ae0SDavid Gibson 53522a15efc9SHugh Dickins /* 53532a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 53543ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 53553ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 53563ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 53573ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 53582a15efc9SHugh Dickins */ 53593ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 53603ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 5361cb900f41SKirill A. Shutemov if (pte) 5362cb900f41SKirill A. Shutemov spin_unlock(ptl); 53632a15efc9SHugh Dickins remainder = 0; 53642a15efc9SHugh Dickins break; 53652a15efc9SHugh Dickins } 53662a15efc9SHugh Dickins 53679cc3a5bdSNaoya Horiguchi /* 53689cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 53699cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 53709cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 53719cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 53729cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 53739cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 53749cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 53759cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 53769cc3a5bdSNaoya Horiguchi */ 53779cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 5378106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 5379106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 53802b740303SSouptick Joarder vm_fault_t ret; 538187ffc118SAndrea Arcangeli unsigned int fault_flags = 0; 53824c887265SAdam Litke 5383cb900f41SKirill A. Shutemov if (pte) 5384cb900f41SKirill A. Shutemov spin_unlock(ptl); 538587ffc118SAndrea Arcangeli if (flags & FOLL_WRITE) 538687ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_WRITE; 53874f6da934SPeter Xu if (locked) 538871335f37SPeter Xu fault_flags |= FAULT_FLAG_ALLOW_RETRY | 538971335f37SPeter Xu FAULT_FLAG_KILLABLE; 539087ffc118SAndrea Arcangeli if (flags & FOLL_NOWAIT) 539187ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY | 539287ffc118SAndrea Arcangeli FAULT_FLAG_RETRY_NOWAIT; 539387ffc118SAndrea Arcangeli if (flags & FOLL_TRIED) { 53944426e945SPeter Xu /* 53954426e945SPeter Xu * Note: FAULT_FLAG_ALLOW_RETRY and 53964426e945SPeter Xu * FAULT_FLAG_TRIED can co-exist 53974426e945SPeter Xu */ 539887ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_TRIED; 539987ffc118SAndrea Arcangeli } 540087ffc118SAndrea Arcangeli ret = hugetlb_fault(mm, vma, vaddr, fault_flags); 540187ffc118SAndrea Arcangeli if (ret & VM_FAULT_ERROR) { 54022be7cfedSDaniel Jordan err = vm_fault_to_errno(ret, flags); 54031c59827dSHugh Dickins remainder = 0; 54041c59827dSHugh Dickins break; 54051c59827dSHugh Dickins } 540687ffc118SAndrea Arcangeli if (ret & VM_FAULT_RETRY) { 54074f6da934SPeter Xu if (locked && 54081ac25013SAndrea Arcangeli !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) 54094f6da934SPeter Xu *locked = 0; 541087ffc118SAndrea Arcangeli *nr_pages = 0; 541187ffc118SAndrea Arcangeli /* 541287ffc118SAndrea Arcangeli * VM_FAULT_RETRY must not return an 541387ffc118SAndrea Arcangeli * error, it will return zero 541487ffc118SAndrea Arcangeli * instead. 541587ffc118SAndrea Arcangeli * 541687ffc118SAndrea Arcangeli * No need to update "position" as the 541787ffc118SAndrea Arcangeli * caller will not check it after 541887ffc118SAndrea Arcangeli * *nr_pages is set to 0. 541987ffc118SAndrea Arcangeli */ 542087ffc118SAndrea Arcangeli return i; 542187ffc118SAndrea Arcangeli } 542287ffc118SAndrea Arcangeli continue; 542387ffc118SAndrea Arcangeli } 542463551ae0SDavid Gibson 5425a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 54267f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 54278fde12caSLinus Torvalds 54288fde12caSLinus Torvalds /* 5429acbfb087SZhigang Lu * If subpage information not requested, update counters 5430acbfb087SZhigang Lu * and skip the same_page loop below. 5431acbfb087SZhigang Lu */ 5432acbfb087SZhigang Lu if (!pages && !vmas && !pfn_offset && 5433acbfb087SZhigang Lu (vaddr + huge_page_size(h) < vma->vm_end) && 5434acbfb087SZhigang Lu (remainder >= pages_per_huge_page(h))) { 5435acbfb087SZhigang Lu vaddr += huge_page_size(h); 5436acbfb087SZhigang Lu remainder -= pages_per_huge_page(h); 5437acbfb087SZhigang Lu i += pages_per_huge_page(h); 5438acbfb087SZhigang Lu spin_unlock(ptl); 5439acbfb087SZhigang Lu continue; 5440acbfb087SZhigang Lu } 5441acbfb087SZhigang Lu 544282e5d378SJoao Martins refs = min3(pages_per_huge_page(h) - pfn_offset, 544382e5d378SJoao Martins (vma->vm_end - vaddr) >> PAGE_SHIFT, remainder); 54440fa5bc40SJoao Martins 544582e5d378SJoao Martins if (pages || vmas) 544682e5d378SJoao Martins record_subpages_vmas(mem_map_offset(page, pfn_offset), 544782e5d378SJoao Martins vma, refs, 544882e5d378SJoao Martins likely(pages) ? pages + i : NULL, 544982e5d378SJoao Martins vmas ? vmas + i : NULL); 545063551ae0SDavid Gibson 545182e5d378SJoao Martins if (pages) { 54520fa5bc40SJoao Martins /* 54530fa5bc40SJoao Martins * try_grab_compound_head() should always succeed here, 54540fa5bc40SJoao Martins * because: a) we hold the ptl lock, and b) we've just 54550fa5bc40SJoao Martins * checked that the huge page is present in the page 54560fa5bc40SJoao Martins * tables. If the huge page is present, then the tail 54570fa5bc40SJoao Martins * pages must also be present. The ptl prevents the 54580fa5bc40SJoao Martins * head page and tail pages from being rearranged in 54590fa5bc40SJoao Martins * any way. So this page must be available at this 54600fa5bc40SJoao Martins * point, unless the page refcount overflowed: 54610fa5bc40SJoao Martins */ 546282e5d378SJoao Martins if (WARN_ON_ONCE(!try_grab_compound_head(pages[i], 54630fa5bc40SJoao Martins refs, 54640fa5bc40SJoao Martins flags))) { 54650fa5bc40SJoao Martins spin_unlock(ptl); 54660fa5bc40SJoao Martins remainder = 0; 54670fa5bc40SJoao Martins err = -ENOMEM; 54680fa5bc40SJoao Martins break; 54690fa5bc40SJoao Martins } 5470d5d4b0aaSChen, Kenneth W } 547182e5d378SJoao Martins 547282e5d378SJoao Martins vaddr += (refs << PAGE_SHIFT); 547382e5d378SJoao Martins remainder -= refs; 547482e5d378SJoao Martins i += refs; 547582e5d378SJoao Martins 5476cb900f41SKirill A. Shutemov spin_unlock(ptl); 547763551ae0SDavid Gibson } 547828a35716SMichel Lespinasse *nr_pages = remainder; 547987ffc118SAndrea Arcangeli /* 548087ffc118SAndrea Arcangeli * setting position is actually required only if remainder is 548187ffc118SAndrea Arcangeli * not zero but it's faster not to add a "if (remainder)" 548287ffc118SAndrea Arcangeli * branch. 548387ffc118SAndrea Arcangeli */ 548463551ae0SDavid Gibson *position = vaddr; 548563551ae0SDavid Gibson 54862be7cfedSDaniel Jordan return i ? i : err; 548763551ae0SDavid Gibson } 54888f860591SZhang, Yanmin 54897da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 54908f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 54918f860591SZhang, Yanmin { 54928f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 54938f860591SZhang, Yanmin unsigned long start = address; 54948f860591SZhang, Yanmin pte_t *ptep; 54958f860591SZhang, Yanmin pte_t pte; 5496a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 54977da4d641SPeter Zijlstra unsigned long pages = 0; 5498dff11abeSMike Kravetz bool shared_pmd = false; 5499ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 5500dff11abeSMike Kravetz 5501dff11abeSMike Kravetz /* 5502dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 5503ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 5504dff11abeSMike Kravetz * range if PMD sharing is possible. 5505dff11abeSMike Kravetz */ 55067269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 55077269f999SJérôme Glisse 0, vma, mm, start, end); 5508ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 55098f860591SZhang, Yanmin 55108f860591SZhang, Yanmin BUG_ON(address >= end); 5511ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 55128f860591SZhang, Yanmin 5513ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 551483cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 5515a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 5516cb900f41SKirill A. Shutemov spinlock_t *ptl; 55177868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, huge_page_size(h)); 55188f860591SZhang, Yanmin if (!ptep) 55198f860591SZhang, Yanmin continue; 5520cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 552134ae204fSMike Kravetz if (huge_pmd_unshare(mm, vma, &address, ptep)) { 55227da4d641SPeter Zijlstra pages++; 5523cb900f41SKirill A. Shutemov spin_unlock(ptl); 5524dff11abeSMike Kravetz shared_pmd = true; 552539dde65cSChen, Kenneth W continue; 55267da4d641SPeter Zijlstra } 5527a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 5528a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 5529a8bda28dSNaoya Horiguchi spin_unlock(ptl); 5530a8bda28dSNaoya Horiguchi continue; 5531a8bda28dSNaoya Horiguchi } 5532a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 5533a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 5534a8bda28dSNaoya Horiguchi 5535a8bda28dSNaoya Horiguchi if (is_write_migration_entry(entry)) { 5536a8bda28dSNaoya Horiguchi pte_t newpte; 5537a8bda28dSNaoya Horiguchi 5538a8bda28dSNaoya Horiguchi make_migration_entry_read(&entry); 5539a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 5540e5251fd4SPunit Agrawal set_huge_swap_pte_at(mm, address, ptep, 5541e5251fd4SPunit Agrawal newpte, huge_page_size(h)); 5542a8bda28dSNaoya Horiguchi pages++; 5543a8bda28dSNaoya Horiguchi } 5544a8bda28dSNaoya Horiguchi spin_unlock(ptl); 5545a8bda28dSNaoya Horiguchi continue; 5546a8bda28dSNaoya Horiguchi } 5547a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 5548023bdd00SAneesh Kumar K.V pte_t old_pte; 554979c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 5550023bdd00SAneesh Kumar K.V 5551023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 5552023bdd00SAneesh Kumar K.V pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); 555379c1c594SChristophe Leroy pte = arch_make_huge_pte(pte, shift, vma->vm_flags); 5554023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 55557da4d641SPeter Zijlstra pages++; 55568f860591SZhang, Yanmin } 5557cb900f41SKirill A. Shutemov spin_unlock(ptl); 55588f860591SZhang, Yanmin } 5559d833352aSMel Gorman /* 5560c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 5561d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 5562c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 5563dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 5564dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 5565d833352aSMel Gorman */ 5566dff11abeSMike Kravetz if (shared_pmd) 5567ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 5568dff11abeSMike Kravetz else 55695491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 55700f10851eSJérôme Glisse /* 55710f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 55720f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 55730f10851eSJérôme Glisse * 5574ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 55750f10851eSJérôme Glisse */ 557683cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 5577ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 55787da4d641SPeter Zijlstra 55797da4d641SPeter Zijlstra return pages << h->order; 55808f860591SZhang, Yanmin } 55818f860591SZhang, Yanmin 558233b8f84aSMike Kravetz /* Return true if reservation was successful, false otherwise. */ 558333b8f84aSMike Kravetz bool hugetlb_reserve_pages(struct inode *inode, 5584a1e78772SMel Gorman long from, long to, 55855a6fe125SMel Gorman struct vm_area_struct *vma, 5586ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 5587e4e574b7SAdam Litke { 558833b8f84aSMike Kravetz long chg, add = -1; 5589a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 559090481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 55919119a41eSJoonsoo Kim struct resv_map *resv_map; 5592075a61d0SMina Almasry struct hugetlb_cgroup *h_cg = NULL; 55930db9d74eSMina Almasry long gbl_reserve, regions_needed = 0; 5594e4e574b7SAdam Litke 559563489f8eSMike Kravetz /* This should never happen */ 559663489f8eSMike Kravetz if (from > to) { 559763489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 559833b8f84aSMike Kravetz return false; 559963489f8eSMike Kravetz } 560063489f8eSMike Kravetz 5601a1e78772SMel Gorman /* 560217c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 560317c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 560490481622SDavid Gibson * without using reserves 560517c9d12eSMel Gorman */ 5606ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 560733b8f84aSMike Kravetz return true; 560817c9d12eSMel Gorman 560917c9d12eSMel Gorman /* 5610a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 5611a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 5612a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 5613a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 5614a1e78772SMel Gorman */ 56159119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 5616f27a5136SMike Kravetz /* 5617f27a5136SMike Kravetz * resv_map can not be NULL as hugetlb_reserve_pages is only 5618f27a5136SMike Kravetz * called for inodes for which resv_maps were created (see 5619f27a5136SMike Kravetz * hugetlbfs_get_inode). 5620f27a5136SMike Kravetz */ 56214e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 56229119a41eSJoonsoo Kim 56230db9d74eSMina Almasry chg = region_chg(resv_map, from, to, ®ions_needed); 56249119a41eSJoonsoo Kim 56259119a41eSJoonsoo Kim } else { 5626e9fe92aeSMina Almasry /* Private mapping. */ 56279119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 56285a6fe125SMel Gorman if (!resv_map) 562933b8f84aSMike Kravetz return false; 56305a6fe125SMel Gorman 563117c9d12eSMel Gorman chg = to - from; 563217c9d12eSMel Gorman 56335a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 56345a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 56355a6fe125SMel Gorman } 56365a6fe125SMel Gorman 563733b8f84aSMike Kravetz if (chg < 0) 5638c50ac050SDave Hansen goto out_err; 563917c9d12eSMel Gorman 564033b8f84aSMike Kravetz if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), 564133b8f84aSMike Kravetz chg * pages_per_huge_page(h), &h_cg) < 0) 5642075a61d0SMina Almasry goto out_err; 5643075a61d0SMina Almasry 5644075a61d0SMina Almasry if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { 5645075a61d0SMina Almasry /* For private mappings, the hugetlb_cgroup uncharge info hangs 5646075a61d0SMina Almasry * of the resv_map. 5647075a61d0SMina Almasry */ 5648075a61d0SMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); 5649075a61d0SMina Almasry } 5650075a61d0SMina Almasry 56511c5ecae3SMike Kravetz /* 56521c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 56531c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 56541c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 56551c5ecae3SMike Kravetz */ 56561c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 565733b8f84aSMike Kravetz if (gbl_reserve < 0) 5658075a61d0SMina Almasry goto out_uncharge_cgroup; 565917c9d12eSMel Gorman 566017c9d12eSMel Gorman /* 566117c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 566290481622SDavid Gibson * Hand the pages back to the subpool if there are not 566317c9d12eSMel Gorman */ 566433b8f84aSMike Kravetz if (hugetlb_acct_memory(h, gbl_reserve) < 0) 5665075a61d0SMina Almasry goto out_put_pages; 566617c9d12eSMel Gorman 566717c9d12eSMel Gorman /* 566817c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 566917c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 567017c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 567117c9d12eSMel Gorman * the reservation was and the page cache tells how much of 567217c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 567317c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 567417c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 567517c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 567617c9d12eSMel Gorman * else has to be done for private mappings here 567717c9d12eSMel Gorman */ 567833039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 5679075a61d0SMina Almasry add = region_add(resv_map, from, to, regions_needed, h, h_cg); 568033039678SMike Kravetz 56810db9d74eSMina Almasry if (unlikely(add < 0)) { 56820db9d74eSMina Almasry hugetlb_acct_memory(h, -gbl_reserve); 5683075a61d0SMina Almasry goto out_put_pages; 56840db9d74eSMina Almasry } else if (unlikely(chg > add)) { 568533039678SMike Kravetz /* 568633039678SMike Kravetz * pages in this range were added to the reserve 568733039678SMike Kravetz * map between region_chg and region_add. This 568833039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 568933039678SMike Kravetz * the subpool and reserve counts modified above 569033039678SMike Kravetz * based on the difference. 569133039678SMike Kravetz */ 569233039678SMike Kravetz long rsv_adjust; 569333039678SMike Kravetz 5694d85aecf2SMiaohe Lin /* 5695d85aecf2SMiaohe Lin * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the 5696d85aecf2SMiaohe Lin * reference to h_cg->css. See comment below for detail. 5697d85aecf2SMiaohe Lin */ 5698075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd( 5699075a61d0SMina Almasry hstate_index(h), 5700075a61d0SMina Almasry (chg - add) * pages_per_huge_page(h), h_cg); 5701075a61d0SMina Almasry 570233039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 570333039678SMike Kravetz chg - add); 570433039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 5705d85aecf2SMiaohe Lin } else if (h_cg) { 5706d85aecf2SMiaohe Lin /* 5707d85aecf2SMiaohe Lin * The file_regions will hold their own reference to 5708d85aecf2SMiaohe Lin * h_cg->css. So we should release the reference held 5709d85aecf2SMiaohe Lin * via hugetlb_cgroup_charge_cgroup_rsvd() when we are 5710d85aecf2SMiaohe Lin * done. 5711d85aecf2SMiaohe Lin */ 5712d85aecf2SMiaohe Lin hugetlb_cgroup_put_rsvd_cgroup(h_cg); 571333039678SMike Kravetz } 571433039678SMike Kravetz } 571533b8f84aSMike Kravetz return true; 571633b8f84aSMike Kravetz 5717075a61d0SMina Almasry out_put_pages: 5718075a61d0SMina Almasry /* put back original number of pages, chg */ 5719075a61d0SMina Almasry (void)hugepage_subpool_put_pages(spool, chg); 5720075a61d0SMina Almasry out_uncharge_cgroup: 5721075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), 5722075a61d0SMina Almasry chg * pages_per_huge_page(h), h_cg); 5723c50ac050SDave Hansen out_err: 57245e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 57250db9d74eSMina Almasry /* Only call region_abort if the region_chg succeeded but the 57260db9d74eSMina Almasry * region_add failed or didn't run. 57270db9d74eSMina Almasry */ 57280db9d74eSMina Almasry if (chg >= 0 && add < 0) 57290db9d74eSMina Almasry region_abort(resv_map, from, to, regions_needed); 5730f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 5731f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 573233b8f84aSMike Kravetz return false; 5733a43a8c39SChen, Kenneth W } 5734a43a8c39SChen, Kenneth W 5735b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 5736b5cec28dSMike Kravetz long freed) 5737a43a8c39SChen, Kenneth W { 5738a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 57394e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 57409119a41eSJoonsoo Kim long chg = 0; 574190481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 57421c5ecae3SMike Kravetz long gbl_reserve; 574345c682a6SKen Chen 5744f27a5136SMike Kravetz /* 5745f27a5136SMike Kravetz * Since this routine can be called in the evict inode path for all 5746f27a5136SMike Kravetz * hugetlbfs inodes, resv_map could be NULL. 5747f27a5136SMike Kravetz */ 5748b5cec28dSMike Kravetz if (resv_map) { 5749b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 5750b5cec28dSMike Kravetz /* 5751b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 5752b5cec28dSMike Kravetz * must be split and another region descriptor can not be 5753b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 5754b5cec28dSMike Kravetz */ 5755b5cec28dSMike Kravetz if (chg < 0) 5756b5cec28dSMike Kravetz return chg; 5757b5cec28dSMike Kravetz } 5758b5cec28dSMike Kravetz 575945c682a6SKen Chen spin_lock(&inode->i_lock); 5760e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 576145c682a6SKen Chen spin_unlock(&inode->i_lock); 576245c682a6SKen Chen 57631c5ecae3SMike Kravetz /* 57641c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 57651c5ecae3SMike Kravetz * reservations to be released may be adjusted. 5766dddf31a4SMiaohe Lin * 5767dddf31a4SMiaohe Lin * Note that !resv_map implies freed == 0. So (chg - freed) 5768dddf31a4SMiaohe Lin * won't go negative. 57691c5ecae3SMike Kravetz */ 57701c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 57711c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 5772b5cec28dSMike Kravetz 5773b5cec28dSMike Kravetz return 0; 5774a43a8c39SChen, Kenneth W } 577593f70f90SNaoya Horiguchi 57763212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 57773212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 57783212b535SSteve Capper struct vm_area_struct *vma, 57793212b535SSteve Capper unsigned long addr, pgoff_t idx) 57803212b535SSteve Capper { 57813212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 57823212b535SSteve Capper svma->vm_start; 57833212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 57843212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 57853212b535SSteve Capper 57863212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 5787de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 5788de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 57893212b535SSteve Capper 57903212b535SSteve Capper /* 57913212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 57923212b535SSteve Capper * page table page. 57933212b535SSteve Capper */ 57943212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 57953212b535SSteve Capper vm_flags != svm_flags || 579607e51edfSMiaohe Lin !range_in_vma(svma, sbase, s_end)) 57973212b535SSteve Capper return 0; 57983212b535SSteve Capper 57993212b535SSteve Capper return saddr; 58003212b535SSteve Capper } 58013212b535SSteve Capper 580231aafb45SNicholas Krause static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) 58033212b535SSteve Capper { 58043212b535SSteve Capper unsigned long base = addr & PUD_MASK; 58053212b535SSteve Capper unsigned long end = base + PUD_SIZE; 58063212b535SSteve Capper 58073212b535SSteve Capper /* 58083212b535SSteve Capper * check on proper vm_flags and page table alignment 58093212b535SSteve Capper */ 5810017b1660SMike Kravetz if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) 581131aafb45SNicholas Krause return true; 581231aafb45SNicholas Krause return false; 58133212b535SSteve Capper } 58143212b535SSteve Capper 5815c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 5816c1991e07SPeter Xu { 5817c1991e07SPeter Xu #ifdef CONFIG_USERFAULTFD 5818c1991e07SPeter Xu if (uffd_disable_huge_pmd_share(vma)) 5819c1991e07SPeter Xu return false; 5820c1991e07SPeter Xu #endif 5821c1991e07SPeter Xu return vma_shareable(vma, addr); 5822c1991e07SPeter Xu } 5823c1991e07SPeter Xu 58243212b535SSteve Capper /* 5825017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 5826017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 5827017b1660SMike Kravetz * shared pmd mappings. 5828017b1660SMike Kravetz */ 5829017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 5830017b1660SMike Kravetz unsigned long *start, unsigned long *end) 5831017b1660SMike Kravetz { 5832a1ba9da8SLi Xinhai unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), 5833a1ba9da8SLi Xinhai v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 5834017b1660SMike Kravetz 5835a1ba9da8SLi Xinhai /* 5836f0953a1bSIngo Molnar * vma needs to span at least one aligned PUD size, and the range 5837f0953a1bSIngo Molnar * must be at least partially within in. 5838a1ba9da8SLi Xinhai */ 5839a1ba9da8SLi Xinhai if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || 5840a1ba9da8SLi Xinhai (*end <= v_start) || (*start >= v_end)) 5841017b1660SMike Kravetz return; 5842017b1660SMike Kravetz 584375802ca6SPeter Xu /* Extend the range to be PUD aligned for a worst case scenario */ 5844a1ba9da8SLi Xinhai if (*start > v_start) 5845a1ba9da8SLi Xinhai *start = ALIGN_DOWN(*start, PUD_SIZE); 5846017b1660SMike Kravetz 5847a1ba9da8SLi Xinhai if (*end < v_end) 5848a1ba9da8SLi Xinhai *end = ALIGN(*end, PUD_SIZE); 5849017b1660SMike Kravetz } 5850017b1660SMike Kravetz 5851017b1660SMike Kravetz /* 58523212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 58533212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 58543212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 5855c0d0381aSMike Kravetz * code much cleaner. 5856c0d0381aSMike Kravetz * 58570bf7b64eSMike Kravetz * This routine must be called with i_mmap_rwsem held in at least read mode if 58580bf7b64eSMike Kravetz * sharing is possible. For hugetlbfs, this prevents removal of any page 58590bf7b64eSMike Kravetz * table entries associated with the address space. This is important as we 58600bf7b64eSMike Kravetz * are setting up sharing based on existing page table entries (mappings). 58610bf7b64eSMike Kravetz * 58620bf7b64eSMike Kravetz * NOTE: This routine is only called from huge_pte_alloc. Some callers of 58630bf7b64eSMike Kravetz * huge_pte_alloc know that sharing is not possible and do not take 58640bf7b64eSMike Kravetz * i_mmap_rwsem as a performance optimization. This is handled by the 58650bf7b64eSMike Kravetz * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is 58660bf7b64eSMike Kravetz * only required for subsequent processing. 58673212b535SSteve Capper */ 5868aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 5869aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 58703212b535SSteve Capper { 58713212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 58723212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 58733212b535SSteve Capper vma->vm_pgoff; 58743212b535SSteve Capper struct vm_area_struct *svma; 58753212b535SSteve Capper unsigned long saddr; 58763212b535SSteve Capper pte_t *spte = NULL; 58773212b535SSteve Capper pte_t *pte; 5878cb900f41SKirill A. Shutemov spinlock_t *ptl; 58793212b535SSteve Capper 58800bf7b64eSMike Kravetz i_mmap_assert_locked(mapping); 58813212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 58823212b535SSteve Capper if (svma == vma) 58833212b535SSteve Capper continue; 58843212b535SSteve Capper 58853212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 58863212b535SSteve Capper if (saddr) { 58877868a208SPunit Agrawal spte = huge_pte_offset(svma->vm_mm, saddr, 58887868a208SPunit Agrawal vma_mmu_pagesize(svma)); 58893212b535SSteve Capper if (spte) { 58903212b535SSteve Capper get_page(virt_to_page(spte)); 58913212b535SSteve Capper break; 58923212b535SSteve Capper } 58933212b535SSteve Capper } 58943212b535SSteve Capper } 58953212b535SSteve Capper 58963212b535SSteve Capper if (!spte) 58973212b535SSteve Capper goto out; 58983212b535SSteve Capper 58998bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 5900dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 59013212b535SSteve Capper pud_populate(mm, pud, 59023212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 5903c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 5904dc6c9a35SKirill A. Shutemov } else { 59053212b535SSteve Capper put_page(virt_to_page(spte)); 5906dc6c9a35SKirill A. Shutemov } 5907cb900f41SKirill A. Shutemov spin_unlock(ptl); 59083212b535SSteve Capper out: 59093212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 59103212b535SSteve Capper return pte; 59113212b535SSteve Capper } 59123212b535SSteve Capper 59133212b535SSteve Capper /* 59143212b535SSteve Capper * unmap huge page backed by shared pte. 59153212b535SSteve Capper * 59163212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 59173212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 59183212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 59193212b535SSteve Capper * 5920c0d0381aSMike Kravetz * Called with page table lock held and i_mmap_rwsem held in write mode. 59213212b535SSteve Capper * 59223212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 59233212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 59243212b535SSteve Capper */ 592534ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 592634ae204fSMike Kravetz unsigned long *addr, pte_t *ptep) 59273212b535SSteve Capper { 59283212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 5929c2febafcSKirill A. Shutemov p4d_t *p4d = p4d_offset(pgd, *addr); 5930c2febafcSKirill A. Shutemov pud_t *pud = pud_offset(p4d, *addr); 59313212b535SSteve Capper 593234ae204fSMike Kravetz i_mmap_assert_write_locked(vma->vm_file->f_mapping); 59333212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 59343212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 59353212b535SSteve Capper return 0; 59363212b535SSteve Capper 59373212b535SSteve Capper pud_clear(pud); 59383212b535SSteve Capper put_page(virt_to_page(ptep)); 5939dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 59403212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 59413212b535SSteve Capper return 1; 59423212b535SSteve Capper } 5943c1991e07SPeter Xu 59449e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 5945aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 5946aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 59479e5fc74cSSteve Capper { 59489e5fc74cSSteve Capper return NULL; 59499e5fc74cSSteve Capper } 5950e81f2d22SZhang Zhen 595134ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 595234ae204fSMike Kravetz unsigned long *addr, pte_t *ptep) 5953e81f2d22SZhang Zhen { 5954e81f2d22SZhang Zhen return 0; 5955e81f2d22SZhang Zhen } 5956017b1660SMike Kravetz 5957017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 5958017b1660SMike Kravetz unsigned long *start, unsigned long *end) 5959017b1660SMike Kravetz { 5960017b1660SMike Kravetz } 5961c1991e07SPeter Xu 5962c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 5963c1991e07SPeter Xu { 5964c1991e07SPeter Xu return false; 5965c1991e07SPeter Xu } 59663212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 59673212b535SSteve Capper 59689e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 5969aec44e0fSPeter Xu pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, 59709e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 59719e5fc74cSSteve Capper { 59729e5fc74cSSteve Capper pgd_t *pgd; 5973c2febafcSKirill A. Shutemov p4d_t *p4d; 59749e5fc74cSSteve Capper pud_t *pud; 59759e5fc74cSSteve Capper pte_t *pte = NULL; 59769e5fc74cSSteve Capper 59779e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 5978f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 5979f4f0a3d8SKirill A. Shutemov if (!p4d) 5980f4f0a3d8SKirill A. Shutemov return NULL; 5981c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 59829e5fc74cSSteve Capper if (pud) { 59839e5fc74cSSteve Capper if (sz == PUD_SIZE) { 59849e5fc74cSSteve Capper pte = (pte_t *)pud; 59859e5fc74cSSteve Capper } else { 59869e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 5987c1991e07SPeter Xu if (want_pmd_share(vma, addr) && pud_none(*pud)) 5988aec44e0fSPeter Xu pte = huge_pmd_share(mm, vma, addr, pud); 59899e5fc74cSSteve Capper else 59909e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 59919e5fc74cSSteve Capper } 59929e5fc74cSSteve Capper } 59934e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 59949e5fc74cSSteve Capper 59959e5fc74cSSteve Capper return pte; 59969e5fc74cSSteve Capper } 59979e5fc74cSSteve Capper 59989b19df29SPunit Agrawal /* 59999b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 60009b19df29SPunit Agrawal * entry at address @addr 60019b19df29SPunit Agrawal * 60028ac0b81aSLi Xinhai * Return: Pointer to page table entry (PUD or PMD) for 60038ac0b81aSLi Xinhai * address @addr, or NULL if a !p*d_present() entry is encountered and the 60049b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 60059b19df29SPunit Agrawal * table. 60069b19df29SPunit Agrawal */ 60077868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 60087868a208SPunit Agrawal unsigned long addr, unsigned long sz) 60099e5fc74cSSteve Capper { 60109e5fc74cSSteve Capper pgd_t *pgd; 6011c2febafcSKirill A. Shutemov p4d_t *p4d; 60128ac0b81aSLi Xinhai pud_t *pud; 60138ac0b81aSLi Xinhai pmd_t *pmd; 60149e5fc74cSSteve Capper 60159e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 6016c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 6017c2febafcSKirill A. Shutemov return NULL; 6018c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 6019c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 6020c2febafcSKirill A. Shutemov return NULL; 60219b19df29SPunit Agrawal 6022c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 60238ac0b81aSLi Xinhai if (sz == PUD_SIZE) 60248ac0b81aSLi Xinhai /* must be pud huge, non-present or none */ 60259e5fc74cSSteve Capper return (pte_t *)pud; 60268ac0b81aSLi Xinhai if (!pud_present(*pud)) 60278ac0b81aSLi Xinhai return NULL; 60288ac0b81aSLi Xinhai /* must have a valid entry and size to go further */ 60299b19df29SPunit Agrawal 60309e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 60318ac0b81aSLi Xinhai /* must be pmd huge, non-present or none */ 60329e5fc74cSSteve Capper return (pte_t *)pmd; 60339e5fc74cSSteve Capper } 60349e5fc74cSSteve Capper 603561f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 603661f77edaSNaoya Horiguchi 603761f77edaSNaoya Horiguchi /* 603861f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 603961f77edaSNaoya Horiguchi * behavior. 604061f77edaSNaoya Horiguchi */ 604161f77edaSNaoya Horiguchi struct page * __weak 604261f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 604361f77edaSNaoya Horiguchi int write) 604461f77edaSNaoya Horiguchi { 604561f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 604661f77edaSNaoya Horiguchi } 604761f77edaSNaoya Horiguchi 604861f77edaSNaoya Horiguchi struct page * __weak 60494dc71451SAneesh Kumar K.V follow_huge_pd(struct vm_area_struct *vma, 60504dc71451SAneesh Kumar K.V unsigned long address, hugepd_t hpd, int flags, int pdshift) 60514dc71451SAneesh Kumar K.V { 60524dc71451SAneesh Kumar K.V WARN(1, "hugepd follow called with no support for hugepage directory format\n"); 60534dc71451SAneesh Kumar K.V return NULL; 60544dc71451SAneesh Kumar K.V } 60554dc71451SAneesh Kumar K.V 60564dc71451SAneesh Kumar K.V struct page * __weak 60579e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 6058e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 60599e5fc74cSSteve Capper { 6060e66f17ffSNaoya Horiguchi struct page *page = NULL; 6061e66f17ffSNaoya Horiguchi spinlock_t *ptl; 6062c9d398faSNaoya Horiguchi pte_t pte; 60633faa52c0SJohn Hubbard 60643faa52c0SJohn Hubbard /* FOLL_GET and FOLL_PIN are mutually exclusive. */ 60653faa52c0SJohn Hubbard if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == 60663faa52c0SJohn Hubbard (FOLL_PIN | FOLL_GET))) 60673faa52c0SJohn Hubbard return NULL; 60683faa52c0SJohn Hubbard 6069e66f17ffSNaoya Horiguchi retry: 6070e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 6071e66f17ffSNaoya Horiguchi spin_lock(ptl); 6072e66f17ffSNaoya Horiguchi /* 6073e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 6074e66f17ffSNaoya Horiguchi * unmapped from other threads. 6075e66f17ffSNaoya Horiguchi */ 6076e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 6077e66f17ffSNaoya Horiguchi goto out; 6078c9d398faSNaoya Horiguchi pte = huge_ptep_get((pte_t *)pmd); 6079c9d398faSNaoya Horiguchi if (pte_present(pte)) { 608097534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 60813faa52c0SJohn Hubbard /* 60823faa52c0SJohn Hubbard * try_grab_page() should always succeed here, because: a) we 60833faa52c0SJohn Hubbard * hold the pmd (ptl) lock, and b) we've just checked that the 60843faa52c0SJohn Hubbard * huge pmd (head) page is present in the page tables. The ptl 60853faa52c0SJohn Hubbard * prevents the head page and tail pages from being rearranged 60863faa52c0SJohn Hubbard * in any way. So this page must be available at this point, 60873faa52c0SJohn Hubbard * unless the page refcount overflowed: 60883faa52c0SJohn Hubbard */ 60893faa52c0SJohn Hubbard if (WARN_ON_ONCE(!try_grab_page(page, flags))) { 60903faa52c0SJohn Hubbard page = NULL; 60913faa52c0SJohn Hubbard goto out; 60923faa52c0SJohn Hubbard } 6093e66f17ffSNaoya Horiguchi } else { 6094c9d398faSNaoya Horiguchi if (is_hugetlb_entry_migration(pte)) { 6095e66f17ffSNaoya Horiguchi spin_unlock(ptl); 6096e66f17ffSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pmd, ptl); 6097e66f17ffSNaoya Horiguchi goto retry; 6098e66f17ffSNaoya Horiguchi } 6099e66f17ffSNaoya Horiguchi /* 6100e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 6101e66f17ffSNaoya Horiguchi * follow_page_mask(). 6102e66f17ffSNaoya Horiguchi */ 6103e66f17ffSNaoya Horiguchi } 6104e66f17ffSNaoya Horiguchi out: 6105e66f17ffSNaoya Horiguchi spin_unlock(ptl); 61069e5fc74cSSteve Capper return page; 61079e5fc74cSSteve Capper } 61089e5fc74cSSteve Capper 610961f77edaSNaoya Horiguchi struct page * __weak 61109e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 6111e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 61129e5fc74cSSteve Capper { 61133faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 6114e66f17ffSNaoya Horiguchi return NULL; 61159e5fc74cSSteve Capper 6116e66f17ffSNaoya Horiguchi return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 61179e5fc74cSSteve Capper } 61189e5fc74cSSteve Capper 6119faaa5b62SAnshuman Khandual struct page * __weak 6120faaa5b62SAnshuman Khandual follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) 6121faaa5b62SAnshuman Khandual { 61223faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 6123faaa5b62SAnshuman Khandual return NULL; 6124faaa5b62SAnshuman Khandual 6125faaa5b62SAnshuman Khandual return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); 6126faaa5b62SAnshuman Khandual } 6127faaa5b62SAnshuman Khandual 612831caf665SNaoya Horiguchi bool isolate_huge_page(struct page *page, struct list_head *list) 612931caf665SNaoya Horiguchi { 6130bcc54222SNaoya Horiguchi bool ret = true; 6131bcc54222SNaoya Horiguchi 6132db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 61338f251a3dSMike Kravetz if (!PageHeadHuge(page) || 61348f251a3dSMike Kravetz !HPageMigratable(page) || 61350eb2df2bSMuchun Song !get_page_unless_zero(page)) { 6136bcc54222SNaoya Horiguchi ret = false; 6137bcc54222SNaoya Horiguchi goto unlock; 6138bcc54222SNaoya Horiguchi } 61398f251a3dSMike Kravetz ClearHPageMigratable(page); 614031caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 6141bcc54222SNaoya Horiguchi unlock: 6142db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 6143bcc54222SNaoya Horiguchi return ret; 614431caf665SNaoya Horiguchi } 614531caf665SNaoya Horiguchi 614625182f05SNaoya Horiguchi int get_hwpoison_huge_page(struct page *page, bool *hugetlb) 614725182f05SNaoya Horiguchi { 614825182f05SNaoya Horiguchi int ret = 0; 614925182f05SNaoya Horiguchi 615025182f05SNaoya Horiguchi *hugetlb = false; 615125182f05SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 615225182f05SNaoya Horiguchi if (PageHeadHuge(page)) { 615325182f05SNaoya Horiguchi *hugetlb = true; 615425182f05SNaoya Horiguchi if (HPageFreed(page) || HPageMigratable(page)) 615525182f05SNaoya Horiguchi ret = get_page_unless_zero(page); 61560ed950d1SNaoya Horiguchi else 61570ed950d1SNaoya Horiguchi ret = -EBUSY; 615825182f05SNaoya Horiguchi } 615925182f05SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 616025182f05SNaoya Horiguchi return ret; 616125182f05SNaoya Horiguchi } 616225182f05SNaoya Horiguchi 616331caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 616431caf665SNaoya Horiguchi { 6165db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 61668f251a3dSMike Kravetz SetHPageMigratable(page); 616731caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 6168db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 616931caf665SNaoya Horiguchi put_page(page); 617031caf665SNaoya Horiguchi } 6171ab5ac90aSMichal Hocko 6172ab5ac90aSMichal Hocko void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) 6173ab5ac90aSMichal Hocko { 6174ab5ac90aSMichal Hocko struct hstate *h = page_hstate(oldpage); 6175ab5ac90aSMichal Hocko 6176ab5ac90aSMichal Hocko hugetlb_cgroup_migrate(oldpage, newpage); 6177ab5ac90aSMichal Hocko set_page_owner_migrate_reason(newpage, reason); 6178ab5ac90aSMichal Hocko 6179ab5ac90aSMichal Hocko /* 6180ab5ac90aSMichal Hocko * transfer temporary state of the new huge page. This is 6181ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 6182ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 6183ab5ac90aSMichal Hocko * the temporary status. 6184ab5ac90aSMichal Hocko * 6185ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 6186ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 6187ab5ac90aSMichal Hocko * the per-node's. 6188ab5ac90aSMichal Hocko */ 61899157c311SMike Kravetz if (HPageTemporary(newpage)) { 6190ab5ac90aSMichal Hocko int old_nid = page_to_nid(oldpage); 6191ab5ac90aSMichal Hocko int new_nid = page_to_nid(newpage); 6192ab5ac90aSMichal Hocko 61939157c311SMike Kravetz SetHPageTemporary(oldpage); 61949157c311SMike Kravetz ClearHPageTemporary(newpage); 6195ab5ac90aSMichal Hocko 61965af1ab1dSMiaohe Lin /* 61975af1ab1dSMiaohe Lin * There is no need to transfer the per-node surplus state 61985af1ab1dSMiaohe Lin * when we do not cross the node. 61995af1ab1dSMiaohe Lin */ 62005af1ab1dSMiaohe Lin if (new_nid == old_nid) 62015af1ab1dSMiaohe Lin return; 6202db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 6203ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 6204ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 6205ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 6206ab5ac90aSMichal Hocko } 6207db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 6208ab5ac90aSMichal Hocko } 6209ab5ac90aSMichal Hocko } 6210cf11e85fSRoman Gushchin 62116dfeaff9SPeter Xu /* 62126dfeaff9SPeter Xu * This function will unconditionally remove all the shared pmd pgtable entries 62136dfeaff9SPeter Xu * within the specific vma for a hugetlbfs memory range. 62146dfeaff9SPeter Xu */ 62156dfeaff9SPeter Xu void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) 62166dfeaff9SPeter Xu { 62176dfeaff9SPeter Xu struct hstate *h = hstate_vma(vma); 62186dfeaff9SPeter Xu unsigned long sz = huge_page_size(h); 62196dfeaff9SPeter Xu struct mm_struct *mm = vma->vm_mm; 62206dfeaff9SPeter Xu struct mmu_notifier_range range; 62216dfeaff9SPeter Xu unsigned long address, start, end; 62226dfeaff9SPeter Xu spinlock_t *ptl; 62236dfeaff9SPeter Xu pte_t *ptep; 62246dfeaff9SPeter Xu 62256dfeaff9SPeter Xu if (!(vma->vm_flags & VM_MAYSHARE)) 62266dfeaff9SPeter Xu return; 62276dfeaff9SPeter Xu 62286dfeaff9SPeter Xu start = ALIGN(vma->vm_start, PUD_SIZE); 62296dfeaff9SPeter Xu end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 62306dfeaff9SPeter Xu 62316dfeaff9SPeter Xu if (start >= end) 62326dfeaff9SPeter Xu return; 62336dfeaff9SPeter Xu 62346dfeaff9SPeter Xu /* 62356dfeaff9SPeter Xu * No need to call adjust_range_if_pmd_sharing_possible(), because 62366dfeaff9SPeter Xu * we have already done the PUD_SIZE alignment. 62376dfeaff9SPeter Xu */ 62386dfeaff9SPeter Xu mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, 62396dfeaff9SPeter Xu start, end); 62406dfeaff9SPeter Xu mmu_notifier_invalidate_range_start(&range); 62416dfeaff9SPeter Xu i_mmap_lock_write(vma->vm_file->f_mapping); 62426dfeaff9SPeter Xu for (address = start; address < end; address += PUD_SIZE) { 62436dfeaff9SPeter Xu unsigned long tmp = address; 62446dfeaff9SPeter Xu 62456dfeaff9SPeter Xu ptep = huge_pte_offset(mm, address, sz); 62466dfeaff9SPeter Xu if (!ptep) 62476dfeaff9SPeter Xu continue; 62486dfeaff9SPeter Xu ptl = huge_pte_lock(h, mm, ptep); 62496dfeaff9SPeter Xu /* We don't want 'address' to be changed */ 62506dfeaff9SPeter Xu huge_pmd_unshare(mm, vma, &tmp, ptep); 62516dfeaff9SPeter Xu spin_unlock(ptl); 62526dfeaff9SPeter Xu } 62536dfeaff9SPeter Xu flush_hugetlb_tlb_range(vma, start, end); 62546dfeaff9SPeter Xu i_mmap_unlock_write(vma->vm_file->f_mapping); 62556dfeaff9SPeter Xu /* 62566dfeaff9SPeter Xu * No need to call mmu_notifier_invalidate_range(), see 62576dfeaff9SPeter Xu * Documentation/vm/mmu_notifier.rst. 62586dfeaff9SPeter Xu */ 62596dfeaff9SPeter Xu mmu_notifier_invalidate_range_end(&range); 62606dfeaff9SPeter Xu } 62616dfeaff9SPeter Xu 6262cf11e85fSRoman Gushchin #ifdef CONFIG_CMA 6263cf11e85fSRoman Gushchin static bool cma_reserve_called __initdata; 6264cf11e85fSRoman Gushchin 6265cf11e85fSRoman Gushchin static int __init cmdline_parse_hugetlb_cma(char *p) 6266cf11e85fSRoman Gushchin { 6267cf11e85fSRoman Gushchin hugetlb_cma_size = memparse(p, &p); 6268cf11e85fSRoman Gushchin return 0; 6269cf11e85fSRoman Gushchin } 6270cf11e85fSRoman Gushchin 6271cf11e85fSRoman Gushchin early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); 6272cf11e85fSRoman Gushchin 6273cf11e85fSRoman Gushchin void __init hugetlb_cma_reserve(int order) 6274cf11e85fSRoman Gushchin { 6275cf11e85fSRoman Gushchin unsigned long size, reserved, per_node; 6276cf11e85fSRoman Gushchin int nid; 6277cf11e85fSRoman Gushchin 6278cf11e85fSRoman Gushchin cma_reserve_called = true; 6279cf11e85fSRoman Gushchin 6280cf11e85fSRoman Gushchin if (!hugetlb_cma_size) 6281cf11e85fSRoman Gushchin return; 6282cf11e85fSRoman Gushchin 6283cf11e85fSRoman Gushchin if (hugetlb_cma_size < (PAGE_SIZE << order)) { 6284cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", 6285cf11e85fSRoman Gushchin (PAGE_SIZE << order) / SZ_1M); 6286cf11e85fSRoman Gushchin return; 6287cf11e85fSRoman Gushchin } 6288cf11e85fSRoman Gushchin 6289cf11e85fSRoman Gushchin /* 6290cf11e85fSRoman Gushchin * If 3 GB area is requested on a machine with 4 numa nodes, 6291cf11e85fSRoman Gushchin * let's allocate 1 GB on first three nodes and ignore the last one. 6292cf11e85fSRoman Gushchin */ 6293cf11e85fSRoman Gushchin per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); 6294cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", 6295cf11e85fSRoman Gushchin hugetlb_cma_size / SZ_1M, per_node / SZ_1M); 6296cf11e85fSRoman Gushchin 6297cf11e85fSRoman Gushchin reserved = 0; 6298cf11e85fSRoman Gushchin for_each_node_state(nid, N_ONLINE) { 6299cf11e85fSRoman Gushchin int res; 63002281f797SBarry Song char name[CMA_MAX_NAME]; 6301cf11e85fSRoman Gushchin 6302cf11e85fSRoman Gushchin size = min(per_node, hugetlb_cma_size - reserved); 6303cf11e85fSRoman Gushchin size = round_up(size, PAGE_SIZE << order); 6304cf11e85fSRoman Gushchin 63052281f797SBarry Song snprintf(name, sizeof(name), "hugetlb%d", nid); 6306cf11e85fSRoman Gushchin res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order, 630729d0f41dSBarry Song 0, false, name, 6308cf11e85fSRoman Gushchin &hugetlb_cma[nid], nid); 6309cf11e85fSRoman Gushchin if (res) { 6310cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: reservation failed: err %d, node %d", 6311cf11e85fSRoman Gushchin res, nid); 6312cf11e85fSRoman Gushchin continue; 6313cf11e85fSRoman Gushchin } 6314cf11e85fSRoman Gushchin 6315cf11e85fSRoman Gushchin reserved += size; 6316cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", 6317cf11e85fSRoman Gushchin size / SZ_1M, nid); 6318cf11e85fSRoman Gushchin 6319cf11e85fSRoman Gushchin if (reserved >= hugetlb_cma_size) 6320cf11e85fSRoman Gushchin break; 6321cf11e85fSRoman Gushchin } 6322cf11e85fSRoman Gushchin } 6323cf11e85fSRoman Gushchin 6324cf11e85fSRoman Gushchin void __init hugetlb_cma_check(void) 6325cf11e85fSRoman Gushchin { 6326cf11e85fSRoman Gushchin if (!hugetlb_cma_size || cma_reserve_called) 6327cf11e85fSRoman Gushchin return; 6328cf11e85fSRoman Gushchin 6329cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); 6330cf11e85fSRoman Gushchin } 6331cf11e85fSRoman Gushchin 6332cf11e85fSRoman Gushchin #endif /* CONFIG_CMA */ 6333