1457c8996SThomas Gleixner // SPDX-License-Identifier: GPL-2.0-only 21da177e4SLinus Torvalds /* 31da177e4SLinus Torvalds * Generic hugetlb support. 46d49e352SNadia Yvette Chambers * (C) Nadia Yvette Chambers, April 2004 51da177e4SLinus Torvalds */ 61da177e4SLinus Torvalds #include <linux/list.h> 71da177e4SLinus Torvalds #include <linux/init.h> 81da177e4SLinus Torvalds #include <linux/mm.h> 9e1759c21SAlexey Dobriyan #include <linux/seq_file.h> 101da177e4SLinus Torvalds #include <linux/sysctl.h> 111da177e4SLinus Torvalds #include <linux/highmem.h> 12cddb8a5cSAndrea Arcangeli #include <linux/mmu_notifier.h> 131da177e4SLinus Torvalds #include <linux/nodemask.h> 1463551ae0SDavid Gibson #include <linux/pagemap.h> 155da7ca86SChristoph Lameter #include <linux/mempolicy.h> 163b32123dSGideon Israel Dsouza #include <linux/compiler.h> 17aea47ff3SChristoph Lameter #include <linux/cpuset.h> 183935baa9SDavid Gibson #include <linux/mutex.h> 1997ad1087SMike Rapoport #include <linux/memblock.h> 20a3437870SNishanth Aravamudan #include <linux/sysfs.h> 215a0e3ad6STejun Heo #include <linux/slab.h> 22bbe88753SJoonsoo Kim #include <linux/sched/mm.h> 2363489f8eSMike Kravetz #include <linux/mmdebug.h> 24174cd4b1SIngo Molnar #include <linux/sched/signal.h> 250fe6e20bSNaoya Horiguchi #include <linux/rmap.h> 26c6247f72SMatthew Wilcox #include <linux/string_helpers.h> 27fd6a03edSNaoya Horiguchi #include <linux/swap.h> 28fd6a03edSNaoya Horiguchi #include <linux/swapops.h> 298382d914SDavidlohr Bueso #include <linux/jhash.h> 3098fa15f3SAnshuman Khandual #include <linux/numa.h> 31c77c0a8aSWaiman Long #include <linux/llist.h> 32cf11e85fSRoman Gushchin #include <linux/cma.h> 33d6606683SLinus Torvalds 3463551ae0SDavid Gibson #include <asm/page.h> 35ca15ca40SMike Rapoport #include <asm/pgalloc.h> 3624669e58SAneesh Kumar K.V #include <asm/tlb.h> 3763551ae0SDavid Gibson 3824669e58SAneesh Kumar K.V #include <linux/io.h> 3963551ae0SDavid Gibson #include <linux/hugetlb.h> 409dd540e2SAneesh Kumar K.V #include <linux/hugetlb_cgroup.h> 419a305230SLee Schermerhorn #include <linux/node.h> 421a1aad8aSMike Kravetz #include <linux/userfaultfd_k.h> 43ab5ac90aSMichal Hocko #include <linux/page_owner.h> 447835e98bSNick Piggin #include "internal.h" 451da177e4SLinus Torvalds 46c3f38a38SAneesh Kumar K.V int hugetlb_max_hstate __read_mostly; 47e5ff2159SAndi Kleen unsigned int default_hstate_idx; 48e5ff2159SAndi Kleen struct hstate hstates[HUGE_MAX_HSTATE]; 49cf11e85fSRoman Gushchin 50dbda8feaSBarry Song #ifdef CONFIG_CMA 51cf11e85fSRoman Gushchin static struct cma *hugetlb_cma[MAX_NUMNODES]; 52dbda8feaSBarry Song #endif 53dbda8feaSBarry Song static unsigned long hugetlb_cma_size __initdata; 54cf11e85fSRoman Gushchin 55641844f5SNaoya Horiguchi /* 56641844f5SNaoya Horiguchi * Minimum page order among possible hugepage sizes, set to a proper value 57641844f5SNaoya Horiguchi * at boot time. 58641844f5SNaoya Horiguchi */ 59641844f5SNaoya Horiguchi static unsigned int minimum_order __read_mostly = UINT_MAX; 60e5ff2159SAndi Kleen 6153ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 6253ba51d2SJon Tollefson 63e5ff2159SAndi Kleen /* for command line parsing */ 64e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 65e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 669fee021dSVaishali Thakkar static bool __initdata parsed_valid_hugepagesz = true; 67282f4214SMike Kravetz static bool __initdata parsed_default_hugepagesz; 68e5ff2159SAndi Kleen 693935baa9SDavid Gibson /* 7031caf665SNaoya Horiguchi * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, 7131caf665SNaoya Horiguchi * free_huge_pages, and surplus_huge_pages. 723935baa9SDavid Gibson */ 73c3f38a38SAneesh Kumar K.V DEFINE_SPINLOCK(hugetlb_lock); 740bd0f9fbSEric Paris 758382d914SDavidlohr Bueso /* 768382d914SDavidlohr Bueso * Serializes faults on the same logical page. This is used to 778382d914SDavidlohr Bueso * prevent spurious OOMs when the hugepage pool is fully utilized. 788382d914SDavidlohr Bueso */ 798382d914SDavidlohr Bueso static int num_fault_mutexes; 80c672c7f2SMike Kravetz struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; 818382d914SDavidlohr Bueso 827ca02d0aSMike Kravetz /* Forward declaration */ 837ca02d0aSMike Kravetz static int hugetlb_acct_memory(struct hstate *h, long delta); 847ca02d0aSMike Kravetz 851d88433bSMiaohe Lin static inline bool subpool_is_free(struct hugepage_subpool *spool) 861d88433bSMiaohe Lin { 871d88433bSMiaohe Lin if (spool->count) 881d88433bSMiaohe Lin return false; 891d88433bSMiaohe Lin if (spool->max_hpages != -1) 901d88433bSMiaohe Lin return spool->used_hpages == 0; 911d88433bSMiaohe Lin if (spool->min_hpages != -1) 921d88433bSMiaohe Lin return spool->rsv_hpages == spool->min_hpages; 931d88433bSMiaohe Lin 941d88433bSMiaohe Lin return true; 951d88433bSMiaohe Lin } 961d88433bSMiaohe Lin 97*db71ef79SMike Kravetz static inline void unlock_or_release_subpool(struct hugepage_subpool *spool, 98*db71ef79SMike Kravetz unsigned long irq_flags) 9990481622SDavid Gibson { 100*db71ef79SMike Kravetz spin_unlock_irqrestore(&spool->lock, irq_flags); 10190481622SDavid Gibson 10290481622SDavid Gibson /* If no pages are used, and no other handles to the subpool 1037c8de358SEthon Paul * remain, give up any reservations based on minimum size and 1047ca02d0aSMike Kravetz * free the subpool */ 1051d88433bSMiaohe Lin if (subpool_is_free(spool)) { 1067ca02d0aSMike Kravetz if (spool->min_hpages != -1) 1077ca02d0aSMike Kravetz hugetlb_acct_memory(spool->hstate, 1087ca02d0aSMike Kravetz -spool->min_hpages); 10990481622SDavid Gibson kfree(spool); 11090481622SDavid Gibson } 1117ca02d0aSMike Kravetz } 11290481622SDavid Gibson 1137ca02d0aSMike Kravetz struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, 1147ca02d0aSMike Kravetz long min_hpages) 11590481622SDavid Gibson { 11690481622SDavid Gibson struct hugepage_subpool *spool; 11790481622SDavid Gibson 118c6a91820SMike Kravetz spool = kzalloc(sizeof(*spool), GFP_KERNEL); 11990481622SDavid Gibson if (!spool) 12090481622SDavid Gibson return NULL; 12190481622SDavid Gibson 12290481622SDavid Gibson spin_lock_init(&spool->lock); 12390481622SDavid Gibson spool->count = 1; 1247ca02d0aSMike Kravetz spool->max_hpages = max_hpages; 1257ca02d0aSMike Kravetz spool->hstate = h; 1267ca02d0aSMike Kravetz spool->min_hpages = min_hpages; 1277ca02d0aSMike Kravetz 1287ca02d0aSMike Kravetz if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { 1297ca02d0aSMike Kravetz kfree(spool); 1307ca02d0aSMike Kravetz return NULL; 1317ca02d0aSMike Kravetz } 1327ca02d0aSMike Kravetz spool->rsv_hpages = min_hpages; 13390481622SDavid Gibson 13490481622SDavid Gibson return spool; 13590481622SDavid Gibson } 13690481622SDavid Gibson 13790481622SDavid Gibson void hugepage_put_subpool(struct hugepage_subpool *spool) 13890481622SDavid Gibson { 139*db71ef79SMike Kravetz unsigned long flags; 140*db71ef79SMike Kravetz 141*db71ef79SMike Kravetz spin_lock_irqsave(&spool->lock, flags); 14290481622SDavid Gibson BUG_ON(!spool->count); 14390481622SDavid Gibson spool->count--; 144*db71ef79SMike Kravetz unlock_or_release_subpool(spool, flags); 14590481622SDavid Gibson } 14690481622SDavid Gibson 1471c5ecae3SMike Kravetz /* 1481c5ecae3SMike Kravetz * Subpool accounting for allocating and reserving pages. 1491c5ecae3SMike Kravetz * Return -ENOMEM if there are not enough resources to satisfy the 1509e7ee400SRandy Dunlap * request. Otherwise, return the number of pages by which the 1511c5ecae3SMike Kravetz * global pools must be adjusted (upward). The returned value may 1521c5ecae3SMike Kravetz * only be different than the passed value (delta) in the case where 1537c8de358SEthon Paul * a subpool minimum size must be maintained. 1541c5ecae3SMike Kravetz */ 1551c5ecae3SMike Kravetz static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, 15690481622SDavid Gibson long delta) 15790481622SDavid Gibson { 1581c5ecae3SMike Kravetz long ret = delta; 15990481622SDavid Gibson 16090481622SDavid Gibson if (!spool) 1611c5ecae3SMike Kravetz return ret; 16290481622SDavid Gibson 163*db71ef79SMike Kravetz spin_lock_irq(&spool->lock); 16490481622SDavid Gibson 1651c5ecae3SMike Kravetz if (spool->max_hpages != -1) { /* maximum size accounting */ 1661c5ecae3SMike Kravetz if ((spool->used_hpages + delta) <= spool->max_hpages) 1671c5ecae3SMike Kravetz spool->used_hpages += delta; 1681c5ecae3SMike Kravetz else { 1691c5ecae3SMike Kravetz ret = -ENOMEM; 1701c5ecae3SMike Kravetz goto unlock_ret; 1711c5ecae3SMike Kravetz } 1721c5ecae3SMike Kravetz } 1731c5ecae3SMike Kravetz 17409a95e29SMike Kravetz /* minimum size accounting */ 17509a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->rsv_hpages) { 1761c5ecae3SMike Kravetz if (delta > spool->rsv_hpages) { 1771c5ecae3SMike Kravetz /* 1781c5ecae3SMike Kravetz * Asking for more reserves than those already taken on 1791c5ecae3SMike Kravetz * behalf of subpool. Return difference. 1801c5ecae3SMike Kravetz */ 1811c5ecae3SMike Kravetz ret = delta - spool->rsv_hpages; 1821c5ecae3SMike Kravetz spool->rsv_hpages = 0; 1831c5ecae3SMike Kravetz } else { 1841c5ecae3SMike Kravetz ret = 0; /* reserves already accounted for */ 1851c5ecae3SMike Kravetz spool->rsv_hpages -= delta; 1861c5ecae3SMike Kravetz } 1871c5ecae3SMike Kravetz } 1881c5ecae3SMike Kravetz 1891c5ecae3SMike Kravetz unlock_ret: 190*db71ef79SMike Kravetz spin_unlock_irq(&spool->lock); 19190481622SDavid Gibson return ret; 19290481622SDavid Gibson } 19390481622SDavid Gibson 1941c5ecae3SMike Kravetz /* 1951c5ecae3SMike Kravetz * Subpool accounting for freeing and unreserving pages. 1961c5ecae3SMike Kravetz * Return the number of global page reservations that must be dropped. 1971c5ecae3SMike Kravetz * The return value may only be different than the passed value (delta) 1981c5ecae3SMike Kravetz * in the case where a subpool minimum size must be maintained. 1991c5ecae3SMike Kravetz */ 2001c5ecae3SMike Kravetz static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, 20190481622SDavid Gibson long delta) 20290481622SDavid Gibson { 2031c5ecae3SMike Kravetz long ret = delta; 204*db71ef79SMike Kravetz unsigned long flags; 2051c5ecae3SMike Kravetz 20690481622SDavid Gibson if (!spool) 2071c5ecae3SMike Kravetz return delta; 20890481622SDavid Gibson 209*db71ef79SMike Kravetz spin_lock_irqsave(&spool->lock, flags); 2101c5ecae3SMike Kravetz 2111c5ecae3SMike Kravetz if (spool->max_hpages != -1) /* maximum size accounting */ 21290481622SDavid Gibson spool->used_hpages -= delta; 2131c5ecae3SMike Kravetz 21409a95e29SMike Kravetz /* minimum size accounting */ 21509a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { 2161c5ecae3SMike Kravetz if (spool->rsv_hpages + delta <= spool->min_hpages) 2171c5ecae3SMike Kravetz ret = 0; 2181c5ecae3SMike Kravetz else 2191c5ecae3SMike Kravetz ret = spool->rsv_hpages + delta - spool->min_hpages; 2201c5ecae3SMike Kravetz 2211c5ecae3SMike Kravetz spool->rsv_hpages += delta; 2221c5ecae3SMike Kravetz if (spool->rsv_hpages > spool->min_hpages) 2231c5ecae3SMike Kravetz spool->rsv_hpages = spool->min_hpages; 2241c5ecae3SMike Kravetz } 2251c5ecae3SMike Kravetz 2261c5ecae3SMike Kravetz /* 2271c5ecae3SMike Kravetz * If hugetlbfs_put_super couldn't free spool due to an outstanding 2281c5ecae3SMike Kravetz * quota reference, free it now. 2291c5ecae3SMike Kravetz */ 230*db71ef79SMike Kravetz unlock_or_release_subpool(spool, flags); 2311c5ecae3SMike Kravetz 2321c5ecae3SMike Kravetz return ret; 23390481622SDavid Gibson } 23490481622SDavid Gibson 23590481622SDavid Gibson static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 23690481622SDavid Gibson { 23790481622SDavid Gibson return HUGETLBFS_SB(inode->i_sb)->spool; 23890481622SDavid Gibson } 23990481622SDavid Gibson 24090481622SDavid Gibson static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 24190481622SDavid Gibson { 242496ad9aaSAl Viro return subpool_inode(file_inode(vma->vm_file)); 24390481622SDavid Gibson } 24490481622SDavid Gibson 2450db9d74eSMina Almasry /* Helper that removes a struct file_region from the resv_map cache and returns 2460db9d74eSMina Almasry * it for use. 2470db9d74eSMina Almasry */ 2480db9d74eSMina Almasry static struct file_region * 2490db9d74eSMina Almasry get_file_region_entry_from_cache(struct resv_map *resv, long from, long to) 2500db9d74eSMina Almasry { 2510db9d74eSMina Almasry struct file_region *nrg = NULL; 2520db9d74eSMina Almasry 2530db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count <= 0); 2540db9d74eSMina Almasry 2550db9d74eSMina Almasry resv->region_cache_count--; 2560db9d74eSMina Almasry nrg = list_first_entry(&resv->region_cache, struct file_region, link); 2570db9d74eSMina Almasry list_del(&nrg->link); 2580db9d74eSMina Almasry 2590db9d74eSMina Almasry nrg->from = from; 2600db9d74eSMina Almasry nrg->to = to; 2610db9d74eSMina Almasry 2620db9d74eSMina Almasry return nrg; 2630db9d74eSMina Almasry } 2640db9d74eSMina Almasry 265075a61d0SMina Almasry static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, 266075a61d0SMina Almasry struct file_region *rg) 267075a61d0SMina Almasry { 268075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 269075a61d0SMina Almasry nrg->reservation_counter = rg->reservation_counter; 270075a61d0SMina Almasry nrg->css = rg->css; 271075a61d0SMina Almasry if (rg->css) 272075a61d0SMina Almasry css_get(rg->css); 273075a61d0SMina Almasry #endif 274075a61d0SMina Almasry } 275075a61d0SMina Almasry 276075a61d0SMina Almasry /* Helper that records hugetlb_cgroup uncharge info. */ 277075a61d0SMina Almasry static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, 278075a61d0SMina Almasry struct hstate *h, 279075a61d0SMina Almasry struct resv_map *resv, 280075a61d0SMina Almasry struct file_region *nrg) 281075a61d0SMina Almasry { 282075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 283075a61d0SMina Almasry if (h_cg) { 284075a61d0SMina Almasry nrg->reservation_counter = 285075a61d0SMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 286075a61d0SMina Almasry nrg->css = &h_cg->css; 287d85aecf2SMiaohe Lin /* 288d85aecf2SMiaohe Lin * The caller will hold exactly one h_cg->css reference for the 289d85aecf2SMiaohe Lin * whole contiguous reservation region. But this area might be 290d85aecf2SMiaohe Lin * scattered when there are already some file_regions reside in 291d85aecf2SMiaohe Lin * it. As a result, many file_regions may share only one css 292d85aecf2SMiaohe Lin * reference. In order to ensure that one file_region must hold 293d85aecf2SMiaohe Lin * exactly one h_cg->css reference, we should do css_get for 294d85aecf2SMiaohe Lin * each file_region and leave the reference held by caller 295d85aecf2SMiaohe Lin * untouched. 296d85aecf2SMiaohe Lin */ 297d85aecf2SMiaohe Lin css_get(&h_cg->css); 298075a61d0SMina Almasry if (!resv->pages_per_hpage) 299075a61d0SMina Almasry resv->pages_per_hpage = pages_per_huge_page(h); 300075a61d0SMina Almasry /* pages_per_hpage should be the same for all entries in 301075a61d0SMina Almasry * a resv_map. 302075a61d0SMina Almasry */ 303075a61d0SMina Almasry VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); 304075a61d0SMina Almasry } else { 305075a61d0SMina Almasry nrg->reservation_counter = NULL; 306075a61d0SMina Almasry nrg->css = NULL; 307075a61d0SMina Almasry } 308075a61d0SMina Almasry #endif 309075a61d0SMina Almasry } 310075a61d0SMina Almasry 311d85aecf2SMiaohe Lin static void put_uncharge_info(struct file_region *rg) 312d85aecf2SMiaohe Lin { 313d85aecf2SMiaohe Lin #ifdef CONFIG_CGROUP_HUGETLB 314d85aecf2SMiaohe Lin if (rg->css) 315d85aecf2SMiaohe Lin css_put(rg->css); 316d85aecf2SMiaohe Lin #endif 317d85aecf2SMiaohe Lin } 318d85aecf2SMiaohe Lin 319a9b3f867SMina Almasry static bool has_same_uncharge_info(struct file_region *rg, 320a9b3f867SMina Almasry struct file_region *org) 321a9b3f867SMina Almasry { 322a9b3f867SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 323a9b3f867SMina Almasry return rg && org && 324a9b3f867SMina Almasry rg->reservation_counter == org->reservation_counter && 325a9b3f867SMina Almasry rg->css == org->css; 326a9b3f867SMina Almasry 327a9b3f867SMina Almasry #else 328a9b3f867SMina Almasry return true; 329a9b3f867SMina Almasry #endif 330a9b3f867SMina Almasry } 331a9b3f867SMina Almasry 332a9b3f867SMina Almasry static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) 333a9b3f867SMina Almasry { 334a9b3f867SMina Almasry struct file_region *nrg = NULL, *prg = NULL; 335a9b3f867SMina Almasry 336a9b3f867SMina Almasry prg = list_prev_entry(rg, link); 337a9b3f867SMina Almasry if (&prg->link != &resv->regions && prg->to == rg->from && 338a9b3f867SMina Almasry has_same_uncharge_info(prg, rg)) { 339a9b3f867SMina Almasry prg->to = rg->to; 340a9b3f867SMina Almasry 341a9b3f867SMina Almasry list_del(&rg->link); 342d85aecf2SMiaohe Lin put_uncharge_info(rg); 343a9b3f867SMina Almasry kfree(rg); 344a9b3f867SMina Almasry 3457db5e7b6SWei Yang rg = prg; 346a9b3f867SMina Almasry } 347a9b3f867SMina Almasry 348a9b3f867SMina Almasry nrg = list_next_entry(rg, link); 349a9b3f867SMina Almasry if (&nrg->link != &resv->regions && nrg->from == rg->to && 350a9b3f867SMina Almasry has_same_uncharge_info(nrg, rg)) { 351a9b3f867SMina Almasry nrg->from = rg->from; 352a9b3f867SMina Almasry 353a9b3f867SMina Almasry list_del(&rg->link); 354d85aecf2SMiaohe Lin put_uncharge_info(rg); 355a9b3f867SMina Almasry kfree(rg); 356a9b3f867SMina Almasry } 357a9b3f867SMina Almasry } 358a9b3f867SMina Almasry 3592103cf9cSPeter Xu static inline long 3602103cf9cSPeter Xu hugetlb_resv_map_add(struct resv_map *map, struct file_region *rg, long from, 3612103cf9cSPeter Xu long to, struct hstate *h, struct hugetlb_cgroup *cg, 3622103cf9cSPeter Xu long *regions_needed) 3632103cf9cSPeter Xu { 3642103cf9cSPeter Xu struct file_region *nrg; 3652103cf9cSPeter Xu 3662103cf9cSPeter Xu if (!regions_needed) { 3672103cf9cSPeter Xu nrg = get_file_region_entry_from_cache(map, from, to); 3682103cf9cSPeter Xu record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg); 3692103cf9cSPeter Xu list_add(&nrg->link, rg->link.prev); 3702103cf9cSPeter Xu coalesce_file_region(map, nrg); 3712103cf9cSPeter Xu } else 3722103cf9cSPeter Xu *regions_needed += 1; 3732103cf9cSPeter Xu 3742103cf9cSPeter Xu return to - from; 3752103cf9cSPeter Xu } 3762103cf9cSPeter Xu 377972a3da3SWei Yang /* 378972a3da3SWei Yang * Must be called with resv->lock held. 379972a3da3SWei Yang * 380972a3da3SWei Yang * Calling this with regions_needed != NULL will count the number of pages 381972a3da3SWei Yang * to be added but will not modify the linked list. And regions_needed will 382972a3da3SWei Yang * indicate the number of file_regions needed in the cache to carry out to add 383972a3da3SWei Yang * the regions for this range. 384d75c6af9SMina Almasry */ 385d75c6af9SMina Almasry static long add_reservation_in_range(struct resv_map *resv, long f, long t, 386075a61d0SMina Almasry struct hugetlb_cgroup *h_cg, 387972a3da3SWei Yang struct hstate *h, long *regions_needed) 388d75c6af9SMina Almasry { 3890db9d74eSMina Almasry long add = 0; 390d75c6af9SMina Almasry struct list_head *head = &resv->regions; 3910db9d74eSMina Almasry long last_accounted_offset = f; 3922103cf9cSPeter Xu struct file_region *rg = NULL, *trg = NULL; 393d75c6af9SMina Almasry 3940db9d74eSMina Almasry if (regions_needed) 3950db9d74eSMina Almasry *regions_needed = 0; 396d75c6af9SMina Almasry 3970db9d74eSMina Almasry /* In this loop, we essentially handle an entry for the range 3980db9d74eSMina Almasry * [last_accounted_offset, rg->from), at every iteration, with some 3990db9d74eSMina Almasry * bounds checking. 4000db9d74eSMina Almasry */ 4010db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, head, link) { 4020db9d74eSMina Almasry /* Skip irrelevant regions that start before our range. */ 4030db9d74eSMina Almasry if (rg->from < f) { 4040db9d74eSMina Almasry /* If this region ends after the last accounted offset, 4050db9d74eSMina Almasry * then we need to update last_accounted_offset. 4060db9d74eSMina Almasry */ 4070db9d74eSMina Almasry if (rg->to > last_accounted_offset) 4080db9d74eSMina Almasry last_accounted_offset = rg->to; 4090db9d74eSMina Almasry continue; 4100db9d74eSMina Almasry } 411d75c6af9SMina Almasry 4120db9d74eSMina Almasry /* When we find a region that starts beyond our range, we've 4130db9d74eSMina Almasry * finished. 4140db9d74eSMina Almasry */ 415ca7e0457SPeter Xu if (rg->from >= t) 416d75c6af9SMina Almasry break; 417d75c6af9SMina Almasry 4180db9d74eSMina Almasry /* Add an entry for last_accounted_offset -> rg->from, and 4190db9d74eSMina Almasry * update last_accounted_offset. 420d75c6af9SMina Almasry */ 4212103cf9cSPeter Xu if (rg->from > last_accounted_offset) 4222103cf9cSPeter Xu add += hugetlb_resv_map_add(resv, rg, 4232103cf9cSPeter Xu last_accounted_offset, 4242103cf9cSPeter Xu rg->from, h, h_cg, 4252103cf9cSPeter Xu regions_needed); 426d75c6af9SMina Almasry 4270db9d74eSMina Almasry last_accounted_offset = rg->to; 4280db9d74eSMina Almasry } 4290db9d74eSMina Almasry 4300db9d74eSMina Almasry /* Handle the case where our range extends beyond 4310db9d74eSMina Almasry * last_accounted_offset. 4320db9d74eSMina Almasry */ 4332103cf9cSPeter Xu if (last_accounted_offset < t) 4342103cf9cSPeter Xu add += hugetlb_resv_map_add(resv, rg, last_accounted_offset, 4352103cf9cSPeter Xu t, h, h_cg, regions_needed); 4360db9d74eSMina Almasry 4370db9d74eSMina Almasry VM_BUG_ON(add < 0); 4380db9d74eSMina Almasry return add; 4390db9d74eSMina Almasry } 4400db9d74eSMina Almasry 4410db9d74eSMina Almasry /* Must be called with resv->lock acquired. Will drop lock to allocate entries. 4420db9d74eSMina Almasry */ 4430db9d74eSMina Almasry static int allocate_file_region_entries(struct resv_map *resv, 4440db9d74eSMina Almasry int regions_needed) 4450db9d74eSMina Almasry __must_hold(&resv->lock) 4460db9d74eSMina Almasry { 4470db9d74eSMina Almasry struct list_head allocated_regions; 4480db9d74eSMina Almasry int to_allocate = 0, i = 0; 4490db9d74eSMina Almasry struct file_region *trg = NULL, *rg = NULL; 4500db9d74eSMina Almasry 4510db9d74eSMina Almasry VM_BUG_ON(regions_needed < 0); 4520db9d74eSMina Almasry 4530db9d74eSMina Almasry INIT_LIST_HEAD(&allocated_regions); 4540db9d74eSMina Almasry 4550db9d74eSMina Almasry /* 4560db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 4570db9d74eSMina Almasry * the number of in progress add operations plus regions_needed. 4580db9d74eSMina Almasry * 4590db9d74eSMina Almasry * This is a while loop because when we drop the lock, some other call 4600db9d74eSMina Almasry * to region_add or region_del may have consumed some region_entries, 4610db9d74eSMina Almasry * so we keep looping here until we finally have enough entries for 4620db9d74eSMina Almasry * (adds_in_progress + regions_needed). 4630db9d74eSMina Almasry */ 4640db9d74eSMina Almasry while (resv->region_cache_count < 4650db9d74eSMina Almasry (resv->adds_in_progress + regions_needed)) { 4660db9d74eSMina Almasry to_allocate = resv->adds_in_progress + regions_needed - 4670db9d74eSMina Almasry resv->region_cache_count; 4680db9d74eSMina Almasry 4690db9d74eSMina Almasry /* At this point, we should have enough entries in the cache 4700db9d74eSMina Almasry * for all the existings adds_in_progress. We should only be 4710db9d74eSMina Almasry * needing to allocate for regions_needed. 4720db9d74eSMina Almasry */ 4730db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); 4740db9d74eSMina Almasry 4750db9d74eSMina Almasry spin_unlock(&resv->lock); 4760db9d74eSMina Almasry for (i = 0; i < to_allocate; i++) { 4770db9d74eSMina Almasry trg = kmalloc(sizeof(*trg), GFP_KERNEL); 4780db9d74eSMina Almasry if (!trg) 4790db9d74eSMina Almasry goto out_of_memory; 4800db9d74eSMina Almasry list_add(&trg->link, &allocated_regions); 4810db9d74eSMina Almasry } 4820db9d74eSMina Almasry 4830db9d74eSMina Almasry spin_lock(&resv->lock); 4840db9d74eSMina Almasry 485d3ec7b6eSWei Yang list_splice(&allocated_regions, &resv->region_cache); 486d3ec7b6eSWei Yang resv->region_cache_count += to_allocate; 4870db9d74eSMina Almasry } 4880db9d74eSMina Almasry 4890db9d74eSMina Almasry return 0; 4900db9d74eSMina Almasry 4910db9d74eSMina Almasry out_of_memory: 4920db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, &allocated_regions, link) { 493d75c6af9SMina Almasry list_del(&rg->link); 494d75c6af9SMina Almasry kfree(rg); 495d75c6af9SMina Almasry } 4960db9d74eSMina Almasry return -ENOMEM; 497d75c6af9SMina Almasry } 498d75c6af9SMina Almasry 4991dd308a7SMike Kravetz /* 5001dd308a7SMike Kravetz * Add the huge page range represented by [f, t) to the reserve 5010db9d74eSMina Almasry * map. Regions will be taken from the cache to fill in this range. 5020db9d74eSMina Almasry * Sufficient regions should exist in the cache due to the previous 5030db9d74eSMina Almasry * call to region_chg with the same range, but in some cases the cache will not 5040db9d74eSMina Almasry * have sufficient entries due to races with other code doing region_add or 5050db9d74eSMina Almasry * region_del. The extra needed entries will be allocated. 506cf3ad20bSMike Kravetz * 5070db9d74eSMina Almasry * regions_needed is the out value provided by a previous call to region_chg. 5080db9d74eSMina Almasry * 5090db9d74eSMina Almasry * Return the number of new huge pages added to the map. This number is greater 5100db9d74eSMina Almasry * than or equal to zero. If file_region entries needed to be allocated for 5117c8de358SEthon Paul * this operation and we were not able to allocate, it returns -ENOMEM. 5120db9d74eSMina Almasry * region_add of regions of length 1 never allocate file_regions and cannot 5130db9d74eSMina Almasry * fail; region_chg will always allocate at least 1 entry and a region_add for 5140db9d74eSMina Almasry * 1 page will only require at most 1 entry. 5151dd308a7SMike Kravetz */ 5160db9d74eSMina Almasry static long region_add(struct resv_map *resv, long f, long t, 517075a61d0SMina Almasry long in_regions_needed, struct hstate *h, 518075a61d0SMina Almasry struct hugetlb_cgroup *h_cg) 51996822904SAndy Whitcroft { 5200db9d74eSMina Almasry long add = 0, actual_regions_needed = 0; 52196822904SAndy Whitcroft 5227b24d861SDavidlohr Bueso spin_lock(&resv->lock); 5230db9d74eSMina Almasry retry: 5240db9d74eSMina Almasry 5250db9d74eSMina Almasry /* Count how many regions are actually needed to execute this add. */ 526972a3da3SWei Yang add_reservation_in_range(resv, f, t, NULL, NULL, 527972a3da3SWei Yang &actual_regions_needed); 52896822904SAndy Whitcroft 5295e911373SMike Kravetz /* 5300db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 5310db9d74eSMina Almasry * this add operation. Note that actual_regions_needed may be greater 5320db9d74eSMina Almasry * than in_regions_needed, as the resv_map may have been modified since 5330db9d74eSMina Almasry * the region_chg call. In this case, we need to make sure that we 5340db9d74eSMina Almasry * allocate extra entries, such that we have enough for all the 5350db9d74eSMina Almasry * existing adds_in_progress, plus the excess needed for this 5360db9d74eSMina Almasry * operation. 5375e911373SMike Kravetz */ 5380db9d74eSMina Almasry if (actual_regions_needed > in_regions_needed && 5390db9d74eSMina Almasry resv->region_cache_count < 5400db9d74eSMina Almasry resv->adds_in_progress + 5410db9d74eSMina Almasry (actual_regions_needed - in_regions_needed)) { 5420db9d74eSMina Almasry /* region_add operation of range 1 should never need to 5430db9d74eSMina Almasry * allocate file_region entries. 5440db9d74eSMina Almasry */ 5450db9d74eSMina Almasry VM_BUG_ON(t - f <= 1); 5465e911373SMike Kravetz 5470db9d74eSMina Almasry if (allocate_file_region_entries( 5480db9d74eSMina Almasry resv, actual_regions_needed - in_regions_needed)) { 5490db9d74eSMina Almasry return -ENOMEM; 5505e911373SMike Kravetz } 5515e911373SMike Kravetz 5520db9d74eSMina Almasry goto retry; 5530db9d74eSMina Almasry } 554cf3ad20bSMike Kravetz 555972a3da3SWei Yang add = add_reservation_in_range(resv, f, t, h_cg, h, NULL); 5560db9d74eSMina Almasry 5570db9d74eSMina Almasry resv->adds_in_progress -= in_regions_needed; 5580db9d74eSMina Almasry 5597b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 560cf3ad20bSMike Kravetz return add; 56196822904SAndy Whitcroft } 56296822904SAndy Whitcroft 5631dd308a7SMike Kravetz /* 5641dd308a7SMike Kravetz * Examine the existing reserve map and determine how many 5651dd308a7SMike Kravetz * huge pages in the specified range [f, t) are NOT currently 5661dd308a7SMike Kravetz * represented. This routine is called before a subsequent 5671dd308a7SMike Kravetz * call to region_add that will actually modify the reserve 5681dd308a7SMike Kravetz * map to add the specified range [f, t). region_chg does 5691dd308a7SMike Kravetz * not change the number of huge pages represented by the 5700db9d74eSMina Almasry * map. A number of new file_region structures is added to the cache as a 5710db9d74eSMina Almasry * placeholder, for the subsequent region_add call to use. At least 1 5720db9d74eSMina Almasry * file_region structure is added. 5730db9d74eSMina Almasry * 5740db9d74eSMina Almasry * out_regions_needed is the number of regions added to the 5750db9d74eSMina Almasry * resv->adds_in_progress. This value needs to be provided to a follow up call 5760db9d74eSMina Almasry * to region_add or region_abort for proper accounting. 5775e911373SMike Kravetz * 5785e911373SMike Kravetz * Returns the number of huge pages that need to be added to the existing 5795e911373SMike Kravetz * reservation map for the range [f, t). This number is greater or equal to 5805e911373SMike Kravetz * zero. -ENOMEM is returned if a new file_region structure or cache entry 5815e911373SMike Kravetz * is needed and can not be allocated. 5821dd308a7SMike Kravetz */ 5830db9d74eSMina Almasry static long region_chg(struct resv_map *resv, long f, long t, 5840db9d74eSMina Almasry long *out_regions_needed) 58596822904SAndy Whitcroft { 58696822904SAndy Whitcroft long chg = 0; 58796822904SAndy Whitcroft 5887b24d861SDavidlohr Bueso spin_lock(&resv->lock); 5895e911373SMike Kravetz 590972a3da3SWei Yang /* Count how many hugepages in this range are NOT represented. */ 591075a61d0SMina Almasry chg = add_reservation_in_range(resv, f, t, NULL, NULL, 592972a3da3SWei Yang out_regions_needed); 5935e911373SMike Kravetz 5940db9d74eSMina Almasry if (*out_regions_needed == 0) 5950db9d74eSMina Almasry *out_regions_needed = 1; 5965e911373SMike Kravetz 5970db9d74eSMina Almasry if (allocate_file_region_entries(resv, *out_regions_needed)) 5985e911373SMike Kravetz return -ENOMEM; 5995e911373SMike Kravetz 6000db9d74eSMina Almasry resv->adds_in_progress += *out_regions_needed; 60196822904SAndy Whitcroft 6027b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 60396822904SAndy Whitcroft return chg; 60496822904SAndy Whitcroft } 60596822904SAndy Whitcroft 6061dd308a7SMike Kravetz /* 6075e911373SMike Kravetz * Abort the in progress add operation. The adds_in_progress field 6085e911373SMike Kravetz * of the resv_map keeps track of the operations in progress between 6095e911373SMike Kravetz * calls to region_chg and region_add. Operations are sometimes 6105e911373SMike Kravetz * aborted after the call to region_chg. In such cases, region_abort 6110db9d74eSMina Almasry * is called to decrement the adds_in_progress counter. regions_needed 6120db9d74eSMina Almasry * is the value returned by the region_chg call, it is used to decrement 6130db9d74eSMina Almasry * the adds_in_progress counter. 6145e911373SMike Kravetz * 6155e911373SMike Kravetz * NOTE: The range arguments [f, t) are not needed or used in this 6165e911373SMike Kravetz * routine. They are kept to make reading the calling code easier as 6175e911373SMike Kravetz * arguments will match the associated region_chg call. 6185e911373SMike Kravetz */ 6190db9d74eSMina Almasry static void region_abort(struct resv_map *resv, long f, long t, 6200db9d74eSMina Almasry long regions_needed) 6215e911373SMike Kravetz { 6225e911373SMike Kravetz spin_lock(&resv->lock); 6235e911373SMike Kravetz VM_BUG_ON(!resv->region_cache_count); 6240db9d74eSMina Almasry resv->adds_in_progress -= regions_needed; 6255e911373SMike Kravetz spin_unlock(&resv->lock); 6265e911373SMike Kravetz } 6275e911373SMike Kravetz 6285e911373SMike Kravetz /* 629feba16e2SMike Kravetz * Delete the specified range [f, t) from the reserve map. If the 630feba16e2SMike Kravetz * t parameter is LONG_MAX, this indicates that ALL regions after f 631feba16e2SMike Kravetz * should be deleted. Locate the regions which intersect [f, t) 632feba16e2SMike Kravetz * and either trim, delete or split the existing regions. 633feba16e2SMike Kravetz * 634feba16e2SMike Kravetz * Returns the number of huge pages deleted from the reserve map. 635feba16e2SMike Kravetz * In the normal case, the return value is zero or more. In the 636feba16e2SMike Kravetz * case where a region must be split, a new region descriptor must 637feba16e2SMike Kravetz * be allocated. If the allocation fails, -ENOMEM will be returned. 638feba16e2SMike Kravetz * NOTE: If the parameter t == LONG_MAX, then we will never split 639feba16e2SMike Kravetz * a region and possibly return -ENOMEM. Callers specifying 640feba16e2SMike Kravetz * t == LONG_MAX do not need to check for -ENOMEM error. 6411dd308a7SMike Kravetz */ 642feba16e2SMike Kravetz static long region_del(struct resv_map *resv, long f, long t) 64396822904SAndy Whitcroft { 6441406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 64596822904SAndy Whitcroft struct file_region *rg, *trg; 646feba16e2SMike Kravetz struct file_region *nrg = NULL; 647feba16e2SMike Kravetz long del = 0; 64896822904SAndy Whitcroft 649feba16e2SMike Kravetz retry: 6507b24d861SDavidlohr Bueso spin_lock(&resv->lock); 651feba16e2SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 652dbe409e4SMike Kravetz /* 653dbe409e4SMike Kravetz * Skip regions before the range to be deleted. file_region 654dbe409e4SMike Kravetz * ranges are normally of the form [from, to). However, there 655dbe409e4SMike Kravetz * may be a "placeholder" entry in the map which is of the form 656dbe409e4SMike Kravetz * (from, to) with from == to. Check for placeholder entries 657dbe409e4SMike Kravetz * at the beginning of the range to be deleted. 658dbe409e4SMike Kravetz */ 659dbe409e4SMike Kravetz if (rg->to <= f && (rg->to != rg->from || rg->to != f)) 660feba16e2SMike Kravetz continue; 661dbe409e4SMike Kravetz 662feba16e2SMike Kravetz if (rg->from >= t) 66396822904SAndy Whitcroft break; 66496822904SAndy Whitcroft 665feba16e2SMike Kravetz if (f > rg->from && t < rg->to) { /* Must split region */ 666feba16e2SMike Kravetz /* 667feba16e2SMike Kravetz * Check for an entry in the cache before dropping 668feba16e2SMike Kravetz * lock and attempting allocation. 669feba16e2SMike Kravetz */ 670feba16e2SMike Kravetz if (!nrg && 671feba16e2SMike Kravetz resv->region_cache_count > resv->adds_in_progress) { 672feba16e2SMike Kravetz nrg = list_first_entry(&resv->region_cache, 673feba16e2SMike Kravetz struct file_region, 674feba16e2SMike Kravetz link); 675feba16e2SMike Kravetz list_del(&nrg->link); 676feba16e2SMike Kravetz resv->region_cache_count--; 67796822904SAndy Whitcroft } 67896822904SAndy Whitcroft 679feba16e2SMike Kravetz if (!nrg) { 680feba16e2SMike Kravetz spin_unlock(&resv->lock); 681feba16e2SMike Kravetz nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 682feba16e2SMike Kravetz if (!nrg) 683feba16e2SMike Kravetz return -ENOMEM; 684feba16e2SMike Kravetz goto retry; 685feba16e2SMike Kravetz } 686feba16e2SMike Kravetz 687feba16e2SMike Kravetz del += t - f; 68879aa925bSMike Kravetz hugetlb_cgroup_uncharge_file_region( 689d85aecf2SMiaohe Lin resv, rg, t - f, false); 690feba16e2SMike Kravetz 691feba16e2SMike Kravetz /* New entry for end of split region */ 692feba16e2SMike Kravetz nrg->from = t; 693feba16e2SMike Kravetz nrg->to = rg->to; 694075a61d0SMina Almasry 695075a61d0SMina Almasry copy_hugetlb_cgroup_uncharge_info(nrg, rg); 696075a61d0SMina Almasry 697feba16e2SMike Kravetz INIT_LIST_HEAD(&nrg->link); 698feba16e2SMike Kravetz 699feba16e2SMike Kravetz /* Original entry is trimmed */ 700feba16e2SMike Kravetz rg->to = f; 701feba16e2SMike Kravetz 702feba16e2SMike Kravetz list_add(&nrg->link, &rg->link); 703feba16e2SMike Kravetz nrg = NULL; 70496822904SAndy Whitcroft break; 705feba16e2SMike Kravetz } 706feba16e2SMike Kravetz 707feba16e2SMike Kravetz if (f <= rg->from && t >= rg->to) { /* Remove entire region */ 708feba16e2SMike Kravetz del += rg->to - rg->from; 709075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 710d85aecf2SMiaohe Lin rg->to - rg->from, true); 71196822904SAndy Whitcroft list_del(&rg->link); 71296822904SAndy Whitcroft kfree(rg); 713feba16e2SMike Kravetz continue; 71496822904SAndy Whitcroft } 7157b24d861SDavidlohr Bueso 716feba16e2SMike Kravetz if (f <= rg->from) { /* Trim beginning of region */ 717075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 718d85aecf2SMiaohe Lin t - rg->from, false); 719075a61d0SMina Almasry 72079aa925bSMike Kravetz del += t - rg->from; 72179aa925bSMike Kravetz rg->from = t; 72279aa925bSMike Kravetz } else { /* Trim end of region */ 723075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 724d85aecf2SMiaohe Lin rg->to - f, false); 72579aa925bSMike Kravetz 72679aa925bSMike Kravetz del += rg->to - f; 72779aa925bSMike Kravetz rg->to = f; 728feba16e2SMike Kravetz } 729feba16e2SMike Kravetz } 730feba16e2SMike Kravetz 7317b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 732feba16e2SMike Kravetz kfree(nrg); 733feba16e2SMike Kravetz return del; 73496822904SAndy Whitcroft } 73596822904SAndy Whitcroft 7361dd308a7SMike Kravetz /* 737b5cec28dSMike Kravetz * A rare out of memory error was encountered which prevented removal of 738b5cec28dSMike Kravetz * the reserve map region for a page. The huge page itself was free'ed 739b5cec28dSMike Kravetz * and removed from the page cache. This routine will adjust the subpool 740b5cec28dSMike Kravetz * usage count, and the global reserve count if needed. By incrementing 741b5cec28dSMike Kravetz * these counts, the reserve map entry which could not be deleted will 742b5cec28dSMike Kravetz * appear as a "reserved" entry instead of simply dangling with incorrect 743b5cec28dSMike Kravetz * counts. 744b5cec28dSMike Kravetz */ 74572e2936cSzhong jiang void hugetlb_fix_reserve_counts(struct inode *inode) 746b5cec28dSMike Kravetz { 747b5cec28dSMike Kravetz struct hugepage_subpool *spool = subpool_inode(inode); 748b5cec28dSMike Kravetz long rsv_adjust; 749da56388cSMiaohe Lin bool reserved = false; 750b5cec28dSMike Kravetz 751b5cec28dSMike Kravetz rsv_adjust = hugepage_subpool_get_pages(spool, 1); 752da56388cSMiaohe Lin if (rsv_adjust > 0) { 753b5cec28dSMike Kravetz struct hstate *h = hstate_inode(inode); 754b5cec28dSMike Kravetz 755da56388cSMiaohe Lin if (!hugetlb_acct_memory(h, 1)) 756da56388cSMiaohe Lin reserved = true; 757da56388cSMiaohe Lin } else if (!rsv_adjust) { 758da56388cSMiaohe Lin reserved = true; 759b5cec28dSMike Kravetz } 760da56388cSMiaohe Lin 761da56388cSMiaohe Lin if (!reserved) 762da56388cSMiaohe Lin pr_warn("hugetlb: Huge Page Reserved count may go negative.\n"); 763b5cec28dSMike Kravetz } 764b5cec28dSMike Kravetz 765b5cec28dSMike Kravetz /* 7661dd308a7SMike Kravetz * Count and return the number of huge pages in the reserve map 7671dd308a7SMike Kravetz * that intersect with the range [f, t). 7681dd308a7SMike Kravetz */ 7691406ec9bSJoonsoo Kim static long region_count(struct resv_map *resv, long f, long t) 77084afd99bSAndy Whitcroft { 7711406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 77284afd99bSAndy Whitcroft struct file_region *rg; 77384afd99bSAndy Whitcroft long chg = 0; 77484afd99bSAndy Whitcroft 7757b24d861SDavidlohr Bueso spin_lock(&resv->lock); 77684afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 77784afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 778f2135a4aSWang Sheng-Hui long seg_from; 779f2135a4aSWang Sheng-Hui long seg_to; 78084afd99bSAndy Whitcroft 78184afd99bSAndy Whitcroft if (rg->to <= f) 78284afd99bSAndy Whitcroft continue; 78384afd99bSAndy Whitcroft if (rg->from >= t) 78484afd99bSAndy Whitcroft break; 78584afd99bSAndy Whitcroft 78684afd99bSAndy Whitcroft seg_from = max(rg->from, f); 78784afd99bSAndy Whitcroft seg_to = min(rg->to, t); 78884afd99bSAndy Whitcroft 78984afd99bSAndy Whitcroft chg += seg_to - seg_from; 79084afd99bSAndy Whitcroft } 7917b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 79284afd99bSAndy Whitcroft 79384afd99bSAndy Whitcroft return chg; 79484afd99bSAndy Whitcroft } 79584afd99bSAndy Whitcroft 79696822904SAndy Whitcroft /* 797e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 798e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 799e7c4b0bfSAndy Whitcroft */ 800a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 801a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 802e7c4b0bfSAndy Whitcroft { 803a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 804a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 805e7c4b0bfSAndy Whitcroft } 806e7c4b0bfSAndy Whitcroft 8070fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 8080fe6e20bSNaoya Horiguchi unsigned long address) 8090fe6e20bSNaoya Horiguchi { 8100fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 8110fe6e20bSNaoya Horiguchi } 812dee41079SDan Williams EXPORT_SYMBOL_GPL(linear_hugepage_index); 8130fe6e20bSNaoya Horiguchi 81484afd99bSAndy Whitcroft /* 81508fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 81608fba699SMel Gorman * cases this will be same size as used by the page table entries. 81708fba699SMel Gorman */ 81808fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 81908fba699SMel Gorman { 82005ea8860SDan Williams if (vma->vm_ops && vma->vm_ops->pagesize) 82105ea8860SDan Williams return vma->vm_ops->pagesize(vma); 82208fba699SMel Gorman return PAGE_SIZE; 82308fba699SMel Gorman } 824f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 82508fba699SMel Gorman 82608fba699SMel Gorman /* 8273340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 8283340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 82909135cc5SDan Williams * architectures where it differs, an architecture-specific 'strong' 83009135cc5SDan Williams * version of this symbol is required. 8313340289dSMel Gorman */ 83209135cc5SDan Williams __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 8333340289dSMel Gorman { 8343340289dSMel Gorman return vma_kernel_pagesize(vma); 8353340289dSMel Gorman } 8363340289dSMel Gorman 8373340289dSMel Gorman /* 83884afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 83984afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 84084afd99bSAndy Whitcroft * alignment. 84184afd99bSAndy Whitcroft */ 84284afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 84384afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 84404f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 84584afd99bSAndy Whitcroft 846a1e78772SMel Gorman /* 847a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 848a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 849a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 850a1e78772SMel Gorman * 851a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 852a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 853a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 854a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 85584afd99bSAndy Whitcroft * 85684afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 85784afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 85884afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 85984afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 86084afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 86184afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 86284afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 86384afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 864a1e78772SMel Gorman */ 865e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 866e7c4b0bfSAndy Whitcroft { 867e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 868e7c4b0bfSAndy Whitcroft } 869e7c4b0bfSAndy Whitcroft 870e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 871e7c4b0bfSAndy Whitcroft unsigned long value) 872e7c4b0bfSAndy Whitcroft { 873e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 874e7c4b0bfSAndy Whitcroft } 875e7c4b0bfSAndy Whitcroft 876e9fe92aeSMina Almasry static void 877e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, 878e9fe92aeSMina Almasry struct hugetlb_cgroup *h_cg, 879e9fe92aeSMina Almasry struct hstate *h) 880e9fe92aeSMina Almasry { 881e9fe92aeSMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 882e9fe92aeSMina Almasry if (!h_cg || !h) { 883e9fe92aeSMina Almasry resv_map->reservation_counter = NULL; 884e9fe92aeSMina Almasry resv_map->pages_per_hpage = 0; 885e9fe92aeSMina Almasry resv_map->css = NULL; 886e9fe92aeSMina Almasry } else { 887e9fe92aeSMina Almasry resv_map->reservation_counter = 888e9fe92aeSMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 889e9fe92aeSMina Almasry resv_map->pages_per_hpage = pages_per_huge_page(h); 890e9fe92aeSMina Almasry resv_map->css = &h_cg->css; 891e9fe92aeSMina Almasry } 892e9fe92aeSMina Almasry #endif 893e9fe92aeSMina Almasry } 894e9fe92aeSMina Almasry 8959119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 89684afd99bSAndy Whitcroft { 89784afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 8985e911373SMike Kravetz struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 8995e911373SMike Kravetz 9005e911373SMike Kravetz if (!resv_map || !rg) { 9015e911373SMike Kravetz kfree(resv_map); 9025e911373SMike Kravetz kfree(rg); 90384afd99bSAndy Whitcroft return NULL; 9045e911373SMike Kravetz } 90584afd99bSAndy Whitcroft 90684afd99bSAndy Whitcroft kref_init(&resv_map->refs); 9077b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 90884afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 90984afd99bSAndy Whitcroft 9105e911373SMike Kravetz resv_map->adds_in_progress = 0; 911e9fe92aeSMina Almasry /* 912e9fe92aeSMina Almasry * Initialize these to 0. On shared mappings, 0's here indicate these 913e9fe92aeSMina Almasry * fields don't do cgroup accounting. On private mappings, these will be 914e9fe92aeSMina Almasry * re-initialized to the proper values, to indicate that hugetlb cgroup 915e9fe92aeSMina Almasry * reservations are to be un-charged from here. 916e9fe92aeSMina Almasry */ 917e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); 9185e911373SMike Kravetz 9195e911373SMike Kravetz INIT_LIST_HEAD(&resv_map->region_cache); 9205e911373SMike Kravetz list_add(&rg->link, &resv_map->region_cache); 9215e911373SMike Kravetz resv_map->region_cache_count = 1; 9225e911373SMike Kravetz 92384afd99bSAndy Whitcroft return resv_map; 92484afd99bSAndy Whitcroft } 92584afd99bSAndy Whitcroft 9269119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 92784afd99bSAndy Whitcroft { 92884afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 9295e911373SMike Kravetz struct list_head *head = &resv_map->region_cache; 9305e911373SMike Kravetz struct file_region *rg, *trg; 93184afd99bSAndy Whitcroft 93284afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 933feba16e2SMike Kravetz region_del(resv_map, 0, LONG_MAX); 9345e911373SMike Kravetz 9355e911373SMike Kravetz /* ... and any entries left in the cache */ 9365e911373SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 9375e911373SMike Kravetz list_del(&rg->link); 9385e911373SMike Kravetz kfree(rg); 9395e911373SMike Kravetz } 9405e911373SMike Kravetz 9415e911373SMike Kravetz VM_BUG_ON(resv_map->adds_in_progress); 9425e911373SMike Kravetz 94384afd99bSAndy Whitcroft kfree(resv_map); 94484afd99bSAndy Whitcroft } 94584afd99bSAndy Whitcroft 9464e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 9474e35f483SJoonsoo Kim { 948f27a5136SMike Kravetz /* 949f27a5136SMike Kravetz * At inode evict time, i_mapping may not point to the original 950f27a5136SMike Kravetz * address space within the inode. This original address space 951f27a5136SMike Kravetz * contains the pointer to the resv_map. So, always use the 952f27a5136SMike Kravetz * address space embedded within the inode. 953f27a5136SMike Kravetz * The VERY common case is inode->mapping == &inode->i_data but, 954f27a5136SMike Kravetz * this may not be true for device special inodes. 955f27a5136SMike Kravetz */ 956f27a5136SMike Kravetz return (struct resv_map *)(&inode->i_data)->private_data; 9574e35f483SJoonsoo Kim } 9584e35f483SJoonsoo Kim 95984afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 960a1e78772SMel Gorman { 96181d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 9624e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 9634e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 9644e35f483SJoonsoo Kim struct inode *inode = mapping->host; 9654e35f483SJoonsoo Kim 9664e35f483SJoonsoo Kim return inode_resv_map(inode); 9674e35f483SJoonsoo Kim 9684e35f483SJoonsoo Kim } else { 96984afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 97084afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 9714e35f483SJoonsoo Kim } 972a1e78772SMel Gorman } 973a1e78772SMel Gorman 97484afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 975a1e78772SMel Gorman { 97681d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 97781d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 978a1e78772SMel Gorman 97984afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 98084afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 98104f2cbe3SMel Gorman } 98204f2cbe3SMel Gorman 98304f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 98404f2cbe3SMel Gorman { 98581d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 98681d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 987e7c4b0bfSAndy Whitcroft 988e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 98904f2cbe3SMel Gorman } 99004f2cbe3SMel Gorman 99104f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 99204f2cbe3SMel Gorman { 99381d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 994e7c4b0bfSAndy Whitcroft 995e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 996a1e78772SMel Gorman } 997a1e78772SMel Gorman 99804f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 999a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 1000a1e78772SMel Gorman { 100181d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 1002f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 1003a1e78772SMel Gorman vma->vm_private_data = (void *)0; 1004a1e78772SMel Gorman } 1005a1e78772SMel Gorman 1006a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 1007559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 1008a1e78772SMel Gorman { 1009af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 1010af0ed73eSJoonsoo Kim /* 1011af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 1012af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 1013af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 1014af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 1015af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 1016af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 1017af0ed73eSJoonsoo Kim * properly, so add work-around here. 1018af0ed73eSJoonsoo Kim */ 1019af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 1020559ec2f8SNicholas Krause return true; 1021af0ed73eSJoonsoo Kim else 1022559ec2f8SNicholas Krause return false; 1023af0ed73eSJoonsoo Kim } 1024a63884e9SJoonsoo Kim 1025a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 10261fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 10271fb1b0e9SMike Kravetz /* 10281fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 10291fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 10301fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 10317c8de358SEthon Paul * fallocate. In this case, there really are no reserves to 10321fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 10331fb1b0e9SMike Kravetz */ 10341fb1b0e9SMike Kravetz if (chg) 10351fb1b0e9SMike Kravetz return false; 10361fb1b0e9SMike Kravetz else 1037559ec2f8SNicholas Krause return true; 10381fb1b0e9SMike Kravetz } 1039a63884e9SJoonsoo Kim 1040a63884e9SJoonsoo Kim /* 1041a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 1042a63884e9SJoonsoo Kim * private mappings. 1043a63884e9SJoonsoo Kim */ 104467961f9dSMike Kravetz if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 104567961f9dSMike Kravetz /* 104667961f9dSMike Kravetz * Like the shared case above, a hole punch or truncate 104767961f9dSMike Kravetz * could have been performed on the private mapping. 104867961f9dSMike Kravetz * Examine the value of chg to determine if reserves 104967961f9dSMike Kravetz * actually exist or were previously consumed. 105067961f9dSMike Kravetz * Very Subtle - The value of chg comes from a previous 105167961f9dSMike Kravetz * call to vma_needs_reserves(). The reserve map for 105267961f9dSMike Kravetz * private mappings has different (opposite) semantics 105367961f9dSMike Kravetz * than that of shared mappings. vma_needs_reserves() 105467961f9dSMike Kravetz * has already taken this difference in semantics into 105567961f9dSMike Kravetz * account. Therefore, the meaning of chg is the same 105667961f9dSMike Kravetz * as in the shared case above. Code could easily be 105767961f9dSMike Kravetz * combined, but keeping it separate draws attention to 105867961f9dSMike Kravetz * subtle differences. 105967961f9dSMike Kravetz */ 106067961f9dSMike Kravetz if (chg) 106167961f9dSMike Kravetz return false; 106267961f9dSMike Kravetz else 1063559ec2f8SNicholas Krause return true; 106467961f9dSMike Kravetz } 1065a63884e9SJoonsoo Kim 1066559ec2f8SNicholas Krause return false; 1067a1e78772SMel Gorman } 1068a1e78772SMel Gorman 1069a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 10701da177e4SLinus Torvalds { 10711da177e4SLinus Torvalds int nid = page_to_nid(page); 10720edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 1073a5516438SAndi Kleen h->free_huge_pages++; 1074a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 10756c037149SMike Kravetz SetHPageFreed(page); 10761da177e4SLinus Torvalds } 10771da177e4SLinus Torvalds 107894310cbcSAnshuman Khandual static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) 1079bf50bab2SNaoya Horiguchi { 1080bf50bab2SNaoya Horiguchi struct page *page; 1081bbe88753SJoonsoo Kim bool nocma = !!(current->flags & PF_MEMALLOC_NOCMA); 1082bf50bab2SNaoya Horiguchi 1083bbe88753SJoonsoo Kim list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { 1084bbe88753SJoonsoo Kim if (nocma && is_migrate_cma_page(page)) 1085bbe88753SJoonsoo Kim continue; 1086bbe88753SJoonsoo Kim 10876664bfc8SWei Yang if (PageHWPoison(page)) 10886664bfc8SWei Yang continue; 1089bbe88753SJoonsoo Kim 10900edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 1091a9869b83SNaoya Horiguchi set_page_refcounted(page); 10926c037149SMike Kravetz ClearHPageFreed(page); 1093bf50bab2SNaoya Horiguchi h->free_huge_pages--; 1094bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 1095bf50bab2SNaoya Horiguchi return page; 1096bf50bab2SNaoya Horiguchi } 1097bf50bab2SNaoya Horiguchi 10986664bfc8SWei Yang return NULL; 10996664bfc8SWei Yang } 11006664bfc8SWei Yang 11013e59fcb0SMichal Hocko static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, 11023e59fcb0SMichal Hocko nodemask_t *nmask) 110394310cbcSAnshuman Khandual { 11043e59fcb0SMichal Hocko unsigned int cpuset_mems_cookie; 11053e59fcb0SMichal Hocko struct zonelist *zonelist; 11063e59fcb0SMichal Hocko struct zone *zone; 11073e59fcb0SMichal Hocko struct zoneref *z; 110898fa15f3SAnshuman Khandual int node = NUMA_NO_NODE; 11093e59fcb0SMichal Hocko 11103e59fcb0SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 11113e59fcb0SMichal Hocko 11123e59fcb0SMichal Hocko retry_cpuset: 11133e59fcb0SMichal Hocko cpuset_mems_cookie = read_mems_allowed_begin(); 11143e59fcb0SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 111594310cbcSAnshuman Khandual struct page *page; 111694310cbcSAnshuman Khandual 11173e59fcb0SMichal Hocko if (!cpuset_zone_allowed(zone, gfp_mask)) 11183e59fcb0SMichal Hocko continue; 11193e59fcb0SMichal Hocko /* 11203e59fcb0SMichal Hocko * no need to ask again on the same node. Pool is node rather than 11213e59fcb0SMichal Hocko * zone aware 11223e59fcb0SMichal Hocko */ 11233e59fcb0SMichal Hocko if (zone_to_nid(zone) == node) 11243e59fcb0SMichal Hocko continue; 11253e59fcb0SMichal Hocko node = zone_to_nid(zone); 112694310cbcSAnshuman Khandual 112794310cbcSAnshuman Khandual page = dequeue_huge_page_node_exact(h, node); 112894310cbcSAnshuman Khandual if (page) 112994310cbcSAnshuman Khandual return page; 113094310cbcSAnshuman Khandual } 11313e59fcb0SMichal Hocko if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 11323e59fcb0SMichal Hocko goto retry_cpuset; 11333e59fcb0SMichal Hocko 113494310cbcSAnshuman Khandual return NULL; 113594310cbcSAnshuman Khandual } 113694310cbcSAnshuman Khandual 1137a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 1138a5516438SAndi Kleen struct vm_area_struct *vma, 1139af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 1140af0ed73eSJoonsoo Kim long chg) 11411da177e4SLinus Torvalds { 11423e59fcb0SMichal Hocko struct page *page; 1143480eccf9SLee Schermerhorn struct mempolicy *mpol; 114404ec6264SVlastimil Babka gfp_t gfp_mask; 11453e59fcb0SMichal Hocko nodemask_t *nodemask; 114604ec6264SVlastimil Babka int nid; 11471da177e4SLinus Torvalds 1148a1e78772SMel Gorman /* 1149a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 1150a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 1151a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 1152a1e78772SMel Gorman */ 1153af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg) && 1154a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 1155c0ff7453SMiao Xie goto err; 1156a1e78772SMel Gorman 115704f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 1158a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 11596eab04a8SJustin P. Mattock goto err; 116004f2cbe3SMel Gorman 116104ec6264SVlastimil Babka gfp_mask = htlb_alloc_mask(h); 116204ec6264SVlastimil Babka nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 11633e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 11643e59fcb0SMichal Hocko if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { 1165d6995da3SMike Kravetz SetHPageRestoreReserve(page); 1166a63884e9SJoonsoo Kim h->resv_huge_pages--; 1167bf50bab2SNaoya Horiguchi } 1168cc9a6c87SMel Gorman 1169cc9a6c87SMel Gorman mpol_cond_put(mpol); 1170cc9a6c87SMel Gorman return page; 1171cc9a6c87SMel Gorman 1172c0ff7453SMiao Xie err: 1173cc9a6c87SMel Gorman return NULL; 11741da177e4SLinus Torvalds } 11751da177e4SLinus Torvalds 11761cac6f2cSLuiz Capitulino /* 11771cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 11781cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 11791cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 11801cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 11811cac6f2cSLuiz Capitulino * node for alloc or free. 11821cac6f2cSLuiz Capitulino */ 11831cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 11841cac6f2cSLuiz Capitulino { 11850edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 11861cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 11871cac6f2cSLuiz Capitulino 11881cac6f2cSLuiz Capitulino return nid; 11891cac6f2cSLuiz Capitulino } 11901cac6f2cSLuiz Capitulino 11911cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 11921cac6f2cSLuiz Capitulino { 11931cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 11941cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 11951cac6f2cSLuiz Capitulino return nid; 11961cac6f2cSLuiz Capitulino } 11971cac6f2cSLuiz Capitulino 11981cac6f2cSLuiz Capitulino /* 11991cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 12001cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 12011cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 12021cac6f2cSLuiz Capitulino * mask. 12031cac6f2cSLuiz Capitulino */ 12041cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 12051cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 12061cac6f2cSLuiz Capitulino { 12071cac6f2cSLuiz Capitulino int nid; 12081cac6f2cSLuiz Capitulino 12091cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 12101cac6f2cSLuiz Capitulino 12111cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 12121cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 12131cac6f2cSLuiz Capitulino 12141cac6f2cSLuiz Capitulino return nid; 12151cac6f2cSLuiz Capitulino } 12161cac6f2cSLuiz Capitulino 12171cac6f2cSLuiz Capitulino /* 121810c6ec49SMike Kravetz * helper for remove_pool_huge_page() - return the previously saved 12191cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 12201cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 12211cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 12221cac6f2cSLuiz Capitulino */ 12231cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 12241cac6f2cSLuiz Capitulino { 12251cac6f2cSLuiz Capitulino int nid; 12261cac6f2cSLuiz Capitulino 12271cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 12281cac6f2cSLuiz Capitulino 12291cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 12301cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 12311cac6f2cSLuiz Capitulino 12321cac6f2cSLuiz Capitulino return nid; 12331cac6f2cSLuiz Capitulino } 12341cac6f2cSLuiz Capitulino 12351cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 12361cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 12371cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 12381cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 12391cac6f2cSLuiz Capitulino nr_nodes--) 12401cac6f2cSLuiz Capitulino 12411cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 12421cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 12431cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 12441cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 12451cac6f2cSLuiz Capitulino nr_nodes--) 12461cac6f2cSLuiz Capitulino 1247e1073d1eSAneesh Kumar K.V #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 1248944d9fecSLuiz Capitulino static void destroy_compound_gigantic_page(struct page *page, 1249d00181b9SKirill A. Shutemov unsigned int order) 1250944d9fecSLuiz Capitulino { 1251944d9fecSLuiz Capitulino int i; 1252944d9fecSLuiz Capitulino int nr_pages = 1 << order; 1253944d9fecSLuiz Capitulino struct page *p = page + 1; 1254944d9fecSLuiz Capitulino 1255c8cc708aSGerald Schaefer atomic_set(compound_mapcount_ptr(page), 0); 125647e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 125747e29d32SJohn Hubbard 1258944d9fecSLuiz Capitulino for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 12591d798ca3SKirill A. Shutemov clear_compound_head(p); 1260944d9fecSLuiz Capitulino set_page_refcounted(p); 1261944d9fecSLuiz Capitulino } 1262944d9fecSLuiz Capitulino 1263944d9fecSLuiz Capitulino set_compound_order(page, 0); 1264ba9c1201SGerald Schaefer page[1].compound_nr = 0; 1265944d9fecSLuiz Capitulino __ClearPageHead(page); 1266944d9fecSLuiz Capitulino } 1267944d9fecSLuiz Capitulino 1268d00181b9SKirill A. Shutemov static void free_gigantic_page(struct page *page, unsigned int order) 1269944d9fecSLuiz Capitulino { 1270cf11e85fSRoman Gushchin /* 1271cf11e85fSRoman Gushchin * If the page isn't allocated using the cma allocator, 1272cf11e85fSRoman Gushchin * cma_release() returns false. 1273cf11e85fSRoman Gushchin */ 1274dbda8feaSBarry Song #ifdef CONFIG_CMA 1275dbda8feaSBarry Song if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order)) 1276cf11e85fSRoman Gushchin return; 1277dbda8feaSBarry Song #endif 1278cf11e85fSRoman Gushchin 1279944d9fecSLuiz Capitulino free_contig_range(page_to_pfn(page), 1 << order); 1280944d9fecSLuiz Capitulino } 1281944d9fecSLuiz Capitulino 12824eb0716eSAlexandre Ghiti #ifdef CONFIG_CONTIG_ALLOC 1283d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 1284d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) 1285944d9fecSLuiz Capitulino { 128604adbc3fSMiaohe Lin unsigned long nr_pages = pages_per_huge_page(h); 1287953f064aSLi Xinhai if (nid == NUMA_NO_NODE) 1288953f064aSLi Xinhai nid = numa_mem_id(); 1289944d9fecSLuiz Capitulino 1290dbda8feaSBarry Song #ifdef CONFIG_CMA 1291dbda8feaSBarry Song { 1292cf11e85fSRoman Gushchin struct page *page; 1293cf11e85fSRoman Gushchin int node; 1294cf11e85fSRoman Gushchin 1295953f064aSLi Xinhai if (hugetlb_cma[nid]) { 1296953f064aSLi Xinhai page = cma_alloc(hugetlb_cma[nid], nr_pages, 1297953f064aSLi Xinhai huge_page_order(h), true); 1298953f064aSLi Xinhai if (page) 1299953f064aSLi Xinhai return page; 1300953f064aSLi Xinhai } 1301953f064aSLi Xinhai 1302953f064aSLi Xinhai if (!(gfp_mask & __GFP_THISNODE)) { 1303cf11e85fSRoman Gushchin for_each_node_mask(node, *nodemask) { 1304953f064aSLi Xinhai if (node == nid || !hugetlb_cma[node]) 1305cf11e85fSRoman Gushchin continue; 1306cf11e85fSRoman Gushchin 1307cf11e85fSRoman Gushchin page = cma_alloc(hugetlb_cma[node], nr_pages, 1308cf11e85fSRoman Gushchin huge_page_order(h), true); 1309cf11e85fSRoman Gushchin if (page) 1310cf11e85fSRoman Gushchin return page; 1311cf11e85fSRoman Gushchin } 1312cf11e85fSRoman Gushchin } 1313953f064aSLi Xinhai } 1314dbda8feaSBarry Song #endif 1315cf11e85fSRoman Gushchin 13165e27a2dfSAnshuman Khandual return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); 1317944d9fecSLuiz Capitulino } 1318944d9fecSLuiz Capitulino 1319944d9fecSLuiz Capitulino static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); 1320d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order); 13214eb0716eSAlexandre Ghiti #else /* !CONFIG_CONTIG_ALLOC */ 13224eb0716eSAlexandre Ghiti static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 13234eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 13244eb0716eSAlexandre Ghiti { 13254eb0716eSAlexandre Ghiti return NULL; 13264eb0716eSAlexandre Ghiti } 13274eb0716eSAlexandre Ghiti #endif /* CONFIG_CONTIG_ALLOC */ 1328944d9fecSLuiz Capitulino 1329e1073d1eSAneesh Kumar K.V #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 1330d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 13314eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 13324eb0716eSAlexandre Ghiti { 13334eb0716eSAlexandre Ghiti return NULL; 13344eb0716eSAlexandre Ghiti } 1335d00181b9SKirill A. Shutemov static inline void free_gigantic_page(struct page *page, unsigned int order) { } 1336944d9fecSLuiz Capitulino static inline void destroy_compound_gigantic_page(struct page *page, 1337d00181b9SKirill A. Shutemov unsigned int order) { } 1338944d9fecSLuiz Capitulino #endif 1339944d9fecSLuiz Capitulino 13406eb4e88aSMike Kravetz /* 13416eb4e88aSMike Kravetz * Remove hugetlb page from lists, and update dtor so that page appears 13426eb4e88aSMike Kravetz * as just a compound page. A reference is held on the page. 13436eb4e88aSMike Kravetz * 13446eb4e88aSMike Kravetz * Must be called with hugetlb lock held. 13456eb4e88aSMike Kravetz */ 13466eb4e88aSMike Kravetz static void remove_hugetlb_page(struct hstate *h, struct page *page, 13476eb4e88aSMike Kravetz bool adjust_surplus) 13486eb4e88aSMike Kravetz { 13496eb4e88aSMike Kravetz int nid = page_to_nid(page); 13506eb4e88aSMike Kravetz 13516eb4e88aSMike Kravetz VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 13526eb4e88aSMike Kravetz VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page); 13536eb4e88aSMike Kravetz 13546eb4e88aSMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 13556eb4e88aSMike Kravetz return; 13566eb4e88aSMike Kravetz 13576eb4e88aSMike Kravetz list_del(&page->lru); 13586eb4e88aSMike Kravetz 13596eb4e88aSMike Kravetz if (HPageFreed(page)) { 13606eb4e88aSMike Kravetz h->free_huge_pages--; 13616eb4e88aSMike Kravetz h->free_huge_pages_node[nid]--; 13626eb4e88aSMike Kravetz } 13636eb4e88aSMike Kravetz if (adjust_surplus) { 13646eb4e88aSMike Kravetz h->surplus_huge_pages--; 13656eb4e88aSMike Kravetz h->surplus_huge_pages_node[nid]--; 13666eb4e88aSMike Kravetz } 13676eb4e88aSMike Kravetz 13686eb4e88aSMike Kravetz set_page_refcounted(page); 13696eb4e88aSMike Kravetz set_compound_page_dtor(page, NULL_COMPOUND_DTOR); 13706eb4e88aSMike Kravetz 13716eb4e88aSMike Kravetz h->nr_huge_pages--; 13726eb4e88aSMike Kravetz h->nr_huge_pages_node[nid]--; 13736eb4e88aSMike Kravetz } 13746eb4e88aSMike Kravetz 1375a5516438SAndi Kleen static void update_and_free_page(struct hstate *h, struct page *page) 13766af2acb6SAdam Litke { 13776af2acb6SAdam Litke int i; 1378dbfee5aeSMike Kravetz struct page *subpage = page; 1379a5516438SAndi Kleen 13804eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1381944d9fecSLuiz Capitulino return; 138218229df5SAndy Whitcroft 1383dbfee5aeSMike Kravetz for (i = 0; i < pages_per_huge_page(h); 1384dbfee5aeSMike Kravetz i++, subpage = mem_map_next(subpage, page, i)) { 1385dbfee5aeSMike Kravetz subpage->flags &= ~(1 << PG_locked | 1 << PG_error | 138632f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1387a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1388a7407a27SLuiz Capitulino 1 << PG_writeback); 13896af2acb6SAdam Litke } 1390944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) { 1391944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 1392944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 1393944d9fecSLuiz Capitulino } else { 1394a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 13956af2acb6SAdam Litke } 1396944d9fecSLuiz Capitulino } 13976af2acb6SAdam Litke 139810c6ec49SMike Kravetz static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list) 139910c6ec49SMike Kravetz { 140010c6ec49SMike Kravetz struct page *page, *t_page; 140110c6ec49SMike Kravetz 140210c6ec49SMike Kravetz list_for_each_entry_safe(page, t_page, list, lru) { 140310c6ec49SMike Kravetz update_and_free_page(h, page); 140410c6ec49SMike Kravetz cond_resched(); 140510c6ec49SMike Kravetz } 140610c6ec49SMike Kravetz } 140710c6ec49SMike Kravetz 1408e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1409e5ff2159SAndi Kleen { 1410e5ff2159SAndi Kleen struct hstate *h; 1411e5ff2159SAndi Kleen 1412e5ff2159SAndi Kleen for_each_hstate(h) { 1413e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1414e5ff2159SAndi Kleen return h; 1415e5ff2159SAndi Kleen } 1416e5ff2159SAndi Kleen return NULL; 1417e5ff2159SAndi Kleen } 1418e5ff2159SAndi Kleen 1419*db71ef79SMike Kravetz void free_huge_page(struct page *page) 142027a85ef1SDavid Gibson { 1421a5516438SAndi Kleen /* 1422a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1423a5516438SAndi Kleen * compound page destructor. 1424a5516438SAndi Kleen */ 1425e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 14267893d1d5SAdam Litke int nid = page_to_nid(page); 1427d6995da3SMike Kravetz struct hugepage_subpool *spool = hugetlb_page_subpool(page); 142807443a85SJoonsoo Kim bool restore_reserve; 1429*db71ef79SMike Kravetz unsigned long flags; 143027a85ef1SDavid Gibson 1431b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1432b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_mapcount(page), page); 14338ace22bcSYongkai Wu 1434d6995da3SMike Kravetz hugetlb_set_page_subpool(page, NULL); 14358ace22bcSYongkai Wu page->mapping = NULL; 1436d6995da3SMike Kravetz restore_reserve = HPageRestoreReserve(page); 1437d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 143827a85ef1SDavid Gibson 14391c5ecae3SMike Kravetz /* 1440d6995da3SMike Kravetz * If HPageRestoreReserve was set on page, page allocation consumed a 14410919e1b6SMike Kravetz * reservation. If the page was associated with a subpool, there 14420919e1b6SMike Kravetz * would have been a page reserved in the subpool before allocation 14430919e1b6SMike Kravetz * via hugepage_subpool_get_pages(). Since we are 'restoring' the 14446c26d310SMiaohe Lin * reservation, do not call hugepage_subpool_put_pages() as this will 14450919e1b6SMike Kravetz * remove the reserved page from the subpool. 14460919e1b6SMike Kravetz */ 14470919e1b6SMike Kravetz if (!restore_reserve) { 14480919e1b6SMike Kravetz /* 14490919e1b6SMike Kravetz * A return code of zero implies that the subpool will be 14500919e1b6SMike Kravetz * under its minimum size if the reservation is not restored 14510919e1b6SMike Kravetz * after page is free. Therefore, force restore_reserve 14520919e1b6SMike Kravetz * operation. 14531c5ecae3SMike Kravetz */ 14541c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 14551c5ecae3SMike Kravetz restore_reserve = true; 14560919e1b6SMike Kravetz } 14571c5ecae3SMike Kravetz 1458*db71ef79SMike Kravetz spin_lock_irqsave(&hugetlb_lock, flags); 14598f251a3dSMike Kravetz ClearHPageMigratable(page); 14606d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 14616d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 146208cf9fafSMina Almasry hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 146308cf9fafSMina Almasry pages_per_huge_page(h), page); 146407443a85SJoonsoo Kim if (restore_reserve) 146507443a85SJoonsoo Kim h->resv_huge_pages++; 146607443a85SJoonsoo Kim 14679157c311SMike Kravetz if (HPageTemporary(page)) { 14686eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 1469*db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1470ab5ac90aSMichal Hocko update_and_free_page(h, page); 1471ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 14720edaecfaSAneesh Kumar K.V /* remove the page from active list */ 14736eb4e88aSMike Kravetz remove_hugetlb_page(h, page, true); 1474*db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1475a5516438SAndi Kleen update_and_free_page(h, page); 14767893d1d5SAdam Litke } else { 14775d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1478a5516438SAndi Kleen enqueue_huge_page(h, page); 1479*db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 148027a85ef1SDavid Gibson } 14811121828aSMike Kravetz } 148227a85ef1SDavid Gibson 1483a5516438SAndi Kleen static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1484b7ba30c6SAndi Kleen { 14850edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1486f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1487ff546117SMike Kravetz hugetlb_set_page_subpool(page, NULL); 14889dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 14891adc4d41SMina Almasry set_hugetlb_cgroup_rsvd(page, NULL); 1490*db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 1491a5516438SAndi Kleen h->nr_huge_pages++; 1492a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 14936c037149SMike Kravetz ClearHPageFreed(page); 1494*db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1495b7ba30c6SAndi Kleen } 1496b7ba30c6SAndi Kleen 1497d00181b9SKirill A. Shutemov static void prep_compound_gigantic_page(struct page *page, unsigned int order) 149820a0307cSWu Fengguang { 149920a0307cSWu Fengguang int i; 150020a0307cSWu Fengguang int nr_pages = 1 << order; 150120a0307cSWu Fengguang struct page *p = page + 1; 150220a0307cSWu Fengguang 150320a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 150420a0307cSWu Fengguang set_compound_order(page, order); 1505ef5a22beSAndrea Arcangeli __ClearPageReserved(page); 1506de09d31dSKirill A. Shutemov __SetPageHead(page); 150720a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1508ef5a22beSAndrea Arcangeli /* 1509ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1510ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1511ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 15127c8de358SEthon Paul * too. Otherwise drivers using get_user_pages() to access tail 1513ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1514ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1515ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1516ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1517ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1518ef5a22beSAndrea Arcangeli * after get_user_pages(). 1519ef5a22beSAndrea Arcangeli */ 1520ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 152158a84aa9SYouquan Song set_page_count(p, 0); 15221d798ca3SKirill A. Shutemov set_compound_head(p, page); 152320a0307cSWu Fengguang } 1524b4330afbSMike Kravetz atomic_set(compound_mapcount_ptr(page), -1); 152547e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 152620a0307cSWu Fengguang } 152720a0307cSWu Fengguang 15287795912cSAndrew Morton /* 15297795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 15307795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 15317795912cSAndrew Morton * details. 15327795912cSAndrew Morton */ 153320a0307cSWu Fengguang int PageHuge(struct page *page) 153420a0307cSWu Fengguang { 153520a0307cSWu Fengguang if (!PageCompound(page)) 153620a0307cSWu Fengguang return 0; 153720a0307cSWu Fengguang 153820a0307cSWu Fengguang page = compound_head(page); 1539f1e61557SKirill A. Shutemov return page[1].compound_dtor == HUGETLB_PAGE_DTOR; 154020a0307cSWu Fengguang } 154143131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 154243131e14SNaoya Horiguchi 154327c73ae7SAndrea Arcangeli /* 154427c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 154527c73ae7SAndrea Arcangeli * normal or transparent huge pages. 154627c73ae7SAndrea Arcangeli */ 154727c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 154827c73ae7SAndrea Arcangeli { 154927c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 155027c73ae7SAndrea Arcangeli return 0; 155127c73ae7SAndrea Arcangeli 1552d4af73e3SVlastimil Babka return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; 155327c73ae7SAndrea Arcangeli } 155427c73ae7SAndrea Arcangeli 1555c0d0381aSMike Kravetz /* 1556c0d0381aSMike Kravetz * Find and lock address space (mapping) in write mode. 1557c0d0381aSMike Kravetz * 1558336bf30eSMike Kravetz * Upon entry, the page is locked which means that page_mapping() is 1559336bf30eSMike Kravetz * stable. Due to locking order, we can only trylock_write. If we can 1560336bf30eSMike Kravetz * not get the lock, simply return NULL to caller. 1561c0d0381aSMike Kravetz */ 1562c0d0381aSMike Kravetz struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) 1563c0d0381aSMike Kravetz { 1564336bf30eSMike Kravetz struct address_space *mapping = page_mapping(hpage); 1565c0d0381aSMike Kravetz 1566c0d0381aSMike Kravetz if (!mapping) 1567c0d0381aSMike Kravetz return mapping; 1568c0d0381aSMike Kravetz 1569c0d0381aSMike Kravetz if (i_mmap_trylock_write(mapping)) 1570c0d0381aSMike Kravetz return mapping; 1571c0d0381aSMike Kravetz 1572c0d0381aSMike Kravetz return NULL; 1573c0d0381aSMike Kravetz } 1574c0d0381aSMike Kravetz 157513d60f4bSZhang Yi pgoff_t __basepage_index(struct page *page) 157613d60f4bSZhang Yi { 157713d60f4bSZhang Yi struct page *page_head = compound_head(page); 157813d60f4bSZhang Yi pgoff_t index = page_index(page_head); 157913d60f4bSZhang Yi unsigned long compound_idx; 158013d60f4bSZhang Yi 158113d60f4bSZhang Yi if (!PageHuge(page_head)) 158213d60f4bSZhang Yi return page_index(page); 158313d60f4bSZhang Yi 158413d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 158513d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 158613d60f4bSZhang Yi else 158713d60f4bSZhang Yi compound_idx = page - page_head; 158813d60f4bSZhang Yi 158913d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 159013d60f4bSZhang Yi } 159113d60f4bSZhang Yi 15920c397daeSMichal Hocko static struct page *alloc_buddy_huge_page(struct hstate *h, 1593f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1594f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 15951da177e4SLinus Torvalds { 1596af0fb9dfSMichal Hocko int order = huge_page_order(h); 15971da177e4SLinus Torvalds struct page *page; 1598f60858f9SMike Kravetz bool alloc_try_hard = true; 1599f96efd58SJoe Jin 1600f60858f9SMike Kravetz /* 1601f60858f9SMike Kravetz * By default we always try hard to allocate the page with 1602f60858f9SMike Kravetz * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in 1603f60858f9SMike Kravetz * a loop (to adjust global huge page counts) and previous allocation 1604f60858f9SMike Kravetz * failed, do not continue to try hard on the same node. Use the 1605f60858f9SMike Kravetz * node_alloc_noretry bitmap to manage this state information. 1606f60858f9SMike Kravetz */ 1607f60858f9SMike Kravetz if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) 1608f60858f9SMike Kravetz alloc_try_hard = false; 1609f60858f9SMike Kravetz gfp_mask |= __GFP_COMP|__GFP_NOWARN; 1610f60858f9SMike Kravetz if (alloc_try_hard) 1611f60858f9SMike Kravetz gfp_mask |= __GFP_RETRY_MAYFAIL; 1612af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 1613af0fb9dfSMichal Hocko nid = numa_mem_id(); 161484172f4bSMatthew Wilcox (Oracle) page = __alloc_pages(gfp_mask, order, nid, nmask); 1615af0fb9dfSMichal Hocko if (page) 1616af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC); 1617af0fb9dfSMichal Hocko else 1618af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 161963b4613cSNishanth Aravamudan 1620f60858f9SMike Kravetz /* 1621f60858f9SMike Kravetz * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this 1622f60858f9SMike Kravetz * indicates an overall state change. Clear bit so that we resume 1623f60858f9SMike Kravetz * normal 'try hard' allocations. 1624f60858f9SMike Kravetz */ 1625f60858f9SMike Kravetz if (node_alloc_noretry && page && !alloc_try_hard) 1626f60858f9SMike Kravetz node_clear(nid, *node_alloc_noretry); 1627f60858f9SMike Kravetz 1628f60858f9SMike Kravetz /* 1629f60858f9SMike Kravetz * If we tried hard to get a page but failed, set bit so that 1630f60858f9SMike Kravetz * subsequent attempts will not try as hard until there is an 1631f60858f9SMike Kravetz * overall state change. 1632f60858f9SMike Kravetz */ 1633f60858f9SMike Kravetz if (node_alloc_noretry && !page && alloc_try_hard) 1634f60858f9SMike Kravetz node_set(nid, *node_alloc_noretry); 1635f60858f9SMike Kravetz 163663b4613cSNishanth Aravamudan return page; 163763b4613cSNishanth Aravamudan } 163863b4613cSNishanth Aravamudan 1639af0fb9dfSMichal Hocko /* 16400c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 16410c397daeSMichal Hocko * should use this function to get new hugetlb pages 16420c397daeSMichal Hocko */ 16430c397daeSMichal Hocko static struct page *alloc_fresh_huge_page(struct hstate *h, 1644f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1645f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 16460c397daeSMichal Hocko { 16470c397daeSMichal Hocko struct page *page; 16480c397daeSMichal Hocko 16490c397daeSMichal Hocko if (hstate_is_gigantic(h)) 16500c397daeSMichal Hocko page = alloc_gigantic_page(h, gfp_mask, nid, nmask); 16510c397daeSMichal Hocko else 16520c397daeSMichal Hocko page = alloc_buddy_huge_page(h, gfp_mask, 1653f60858f9SMike Kravetz nid, nmask, node_alloc_noretry); 16540c397daeSMichal Hocko if (!page) 16550c397daeSMichal Hocko return NULL; 16560c397daeSMichal Hocko 16570c397daeSMichal Hocko if (hstate_is_gigantic(h)) 16580c397daeSMichal Hocko prep_compound_gigantic_page(page, huge_page_order(h)); 16590c397daeSMichal Hocko prep_new_huge_page(h, page, page_to_nid(page)); 16600c397daeSMichal Hocko 16610c397daeSMichal Hocko return page; 16620c397daeSMichal Hocko } 16630c397daeSMichal Hocko 16640c397daeSMichal Hocko /* 1665af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 1666af0fb9dfSMichal Hocko * manner. 1667af0fb9dfSMichal Hocko */ 1668f60858f9SMike Kravetz static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 1669f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 1670b2261026SJoonsoo Kim { 1671b2261026SJoonsoo Kim struct page *page; 1672b2261026SJoonsoo Kim int nr_nodes, node; 1673af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 1674b2261026SJoonsoo Kim 1675b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1676f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, 1677f60858f9SMike Kravetz node_alloc_noretry); 1678af0fb9dfSMichal Hocko if (page) 1679b2261026SJoonsoo Kim break; 1680b2261026SJoonsoo Kim } 1681b2261026SJoonsoo Kim 1682af0fb9dfSMichal Hocko if (!page) 1683af0fb9dfSMichal Hocko return 0; 1684b2261026SJoonsoo Kim 1685af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 1686af0fb9dfSMichal Hocko 1687af0fb9dfSMichal Hocko return 1; 1688b2261026SJoonsoo Kim } 1689b2261026SJoonsoo Kim 1690e8c5c824SLee Schermerhorn /* 169110c6ec49SMike Kravetz * Remove huge page from pool from next node to free. Attempt to keep 169210c6ec49SMike Kravetz * persistent huge pages more or less balanced over allowed nodes. 169310c6ec49SMike Kravetz * This routine only 'removes' the hugetlb page. The caller must make 169410c6ec49SMike Kravetz * an additional call to free the page to low level allocators. 1695e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 1696e8c5c824SLee Schermerhorn */ 169710c6ec49SMike Kravetz static struct page *remove_pool_huge_page(struct hstate *h, 169810c6ec49SMike Kravetz nodemask_t *nodes_allowed, 16996ae11b27SLee Schermerhorn bool acct_surplus) 1700e8c5c824SLee Schermerhorn { 1701b2261026SJoonsoo Kim int nr_nodes, node; 170210c6ec49SMike Kravetz struct page *page = NULL; 1703e8c5c824SLee Schermerhorn 1704b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1705685f3457SLee Schermerhorn /* 1706685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 1707685f3457SLee Schermerhorn * nodes with surplus pages. 1708685f3457SLee Schermerhorn */ 1709b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 1710b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 171110c6ec49SMike Kravetz page = list_entry(h->hugepage_freelists[node].next, 1712e8c5c824SLee Schermerhorn struct page, lru); 17136eb4e88aSMike Kravetz remove_hugetlb_page(h, page, acct_surplus); 17149a76db09SLee Schermerhorn break; 1715e8c5c824SLee Schermerhorn } 1716b2261026SJoonsoo Kim } 1717e8c5c824SLee Schermerhorn 171810c6ec49SMike Kravetz return page; 1719e8c5c824SLee Schermerhorn } 1720e8c5c824SLee Schermerhorn 1721c8721bbbSNaoya Horiguchi /* 1722c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 1723faf53defSNaoya Horiguchi * nothing for in-use hugepages and non-hugepages. 1724faf53defSNaoya Horiguchi * This function returns values like below: 1725faf53defSNaoya Horiguchi * 1726faf53defSNaoya Horiguchi * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 1727faf53defSNaoya Horiguchi * (allocated or reserved.) 1728faf53defSNaoya Horiguchi * 0: successfully dissolved free hugepages or the page is not a 1729faf53defSNaoya Horiguchi * hugepage (considered as already dissolved) 1730c8721bbbSNaoya Horiguchi */ 1731c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 1732c8721bbbSNaoya Horiguchi { 17336bc9b564SNaoya Horiguchi int rc = -EBUSY; 1734082d5b6bSGerald Schaefer 17357ffddd49SMuchun Song retry: 1736faf53defSNaoya Horiguchi /* Not to disrupt normal path by vainly holding hugetlb_lock */ 1737faf53defSNaoya Horiguchi if (!PageHuge(page)) 1738faf53defSNaoya Horiguchi return 0; 1739faf53defSNaoya Horiguchi 1740*db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 1741faf53defSNaoya Horiguchi if (!PageHuge(page)) { 1742faf53defSNaoya Horiguchi rc = 0; 1743faf53defSNaoya Horiguchi goto out; 1744faf53defSNaoya Horiguchi } 1745faf53defSNaoya Horiguchi 1746faf53defSNaoya Horiguchi if (!page_count(page)) { 17472247bb33SGerald Schaefer struct page *head = compound_head(page); 17482247bb33SGerald Schaefer struct hstate *h = page_hstate(head); 17496bc9b564SNaoya Horiguchi if (h->free_huge_pages - h->resv_huge_pages == 0) 1750082d5b6bSGerald Schaefer goto out; 17517ffddd49SMuchun Song 17527ffddd49SMuchun Song /* 17537ffddd49SMuchun Song * We should make sure that the page is already on the free list 17547ffddd49SMuchun Song * when it is dissolved. 17557ffddd49SMuchun Song */ 17566c037149SMike Kravetz if (unlikely(!HPageFreed(head))) { 1757*db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 17587ffddd49SMuchun Song cond_resched(); 17597ffddd49SMuchun Song 17607ffddd49SMuchun Song /* 17617ffddd49SMuchun Song * Theoretically, we should return -EBUSY when we 17627ffddd49SMuchun Song * encounter this race. In fact, we have a chance 17637ffddd49SMuchun Song * to successfully dissolve the page if we do a 17647ffddd49SMuchun Song * retry. Because the race window is quite small. 17657ffddd49SMuchun Song * If we seize this opportunity, it is an optimization 17667ffddd49SMuchun Song * for increasing the success rate of dissolving page. 17677ffddd49SMuchun Song */ 17687ffddd49SMuchun Song goto retry; 17697ffddd49SMuchun Song } 17707ffddd49SMuchun Song 1771c3114a84SAnshuman Khandual /* 1772c3114a84SAnshuman Khandual * Move PageHWPoison flag from head page to the raw error page, 1773c3114a84SAnshuman Khandual * which makes any subpages rather than the error page reusable. 1774c3114a84SAnshuman Khandual */ 1775c3114a84SAnshuman Khandual if (PageHWPoison(head) && page != head) { 1776c3114a84SAnshuman Khandual SetPageHWPoison(page); 1777c3114a84SAnshuman Khandual ClearPageHWPoison(head); 1778c3114a84SAnshuman Khandual } 17796eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 1780c1470b33Szhong jiang h->max_huge_pages--; 1781*db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 17822247bb33SGerald Schaefer update_and_free_page(h, head); 17831121828aSMike Kravetz return 0; 1784c8721bbbSNaoya Horiguchi } 1785082d5b6bSGerald Schaefer out: 1786*db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1787082d5b6bSGerald Schaefer return rc; 1788c8721bbbSNaoya Horiguchi } 1789c8721bbbSNaoya Horiguchi 1790c8721bbbSNaoya Horiguchi /* 1791c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 1792c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 17932247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 17942247bb33SGerald Schaefer * part of it lies within the given range. 1795082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 1796082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 1797c8721bbbSNaoya Horiguchi */ 1798082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 1799c8721bbbSNaoya Horiguchi { 1800c8721bbbSNaoya Horiguchi unsigned long pfn; 1801eb03aa00SGerald Schaefer struct page *page; 1802082d5b6bSGerald Schaefer int rc = 0; 1803c8721bbbSNaoya Horiguchi 1804d0177639SLi Zhong if (!hugepages_supported()) 1805082d5b6bSGerald Schaefer return rc; 1806d0177639SLi Zhong 1807eb03aa00SGerald Schaefer for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { 1808eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 1809eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 1810eb03aa00SGerald Schaefer if (rc) 1811082d5b6bSGerald Schaefer break; 1812eb03aa00SGerald Schaefer } 1813082d5b6bSGerald Schaefer 1814082d5b6bSGerald Schaefer return rc; 1815c8721bbbSNaoya Horiguchi } 1816c8721bbbSNaoya Horiguchi 1817ab5ac90aSMichal Hocko /* 1818ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 1819ab5ac90aSMichal Hocko */ 18200c397daeSMichal Hocko static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, 1821aaf14e40SMichal Hocko int nid, nodemask_t *nmask) 18227893d1d5SAdam Litke { 18239980d744SMichal Hocko struct page *page = NULL; 18247893d1d5SAdam Litke 1825bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1826aa888a74SAndi Kleen return NULL; 1827aa888a74SAndi Kleen 1828*db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 18299980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 18309980d744SMichal Hocko goto out_unlock; 1831*db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1832d1c3fb1fSNishanth Aravamudan 1833f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 18349980d744SMichal Hocko if (!page) 18350c397daeSMichal Hocko return NULL; 1836d1c3fb1fSNishanth Aravamudan 1837*db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 18389980d744SMichal Hocko /* 18399980d744SMichal Hocko * We could have raced with the pool size change. 18409980d744SMichal Hocko * Double check that and simply deallocate the new page 18419980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 18429980d744SMichal Hocko * temporary page to workaround the nasty free_huge_page 18439980d744SMichal Hocko * codeflow 18449980d744SMichal Hocko */ 18459980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 18469157c311SMike Kravetz SetHPageTemporary(page); 1847*db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 18489980d744SMichal Hocko put_page(page); 18492bf753e6SKai Shen return NULL; 18509980d744SMichal Hocko } else { 18519980d744SMichal Hocko h->surplus_huge_pages++; 18524704dea3SMichal Hocko h->surplus_huge_pages_node[page_to_nid(page)]++; 18537893d1d5SAdam Litke } 18549980d744SMichal Hocko 18559980d744SMichal Hocko out_unlock: 1856*db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 18577893d1d5SAdam Litke 18587893d1d5SAdam Litke return page; 18597893d1d5SAdam Litke } 18607893d1d5SAdam Litke 1861bbe88753SJoonsoo Kim static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, 1862ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 1863ab5ac90aSMichal Hocko { 1864ab5ac90aSMichal Hocko struct page *page; 1865ab5ac90aSMichal Hocko 1866ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 1867ab5ac90aSMichal Hocko return NULL; 1868ab5ac90aSMichal Hocko 1869f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 1870ab5ac90aSMichal Hocko if (!page) 1871ab5ac90aSMichal Hocko return NULL; 1872ab5ac90aSMichal Hocko 1873ab5ac90aSMichal Hocko /* 1874ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 1875ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 1876ab5ac90aSMichal Hocko */ 18779157c311SMike Kravetz SetHPageTemporary(page); 1878ab5ac90aSMichal Hocko 1879ab5ac90aSMichal Hocko return page; 1880ab5ac90aSMichal Hocko } 1881ab5ac90aSMichal Hocko 1882e4e574b7SAdam Litke /* 1883099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 1884099730d6SDave Hansen */ 1885e0ec90eeSDave Hansen static 18860c397daeSMichal Hocko struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, 1887099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 1888099730d6SDave Hansen { 1889aaf14e40SMichal Hocko struct page *page; 1890aaf14e40SMichal Hocko struct mempolicy *mpol; 1891aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 1892aaf14e40SMichal Hocko int nid; 1893aaf14e40SMichal Hocko nodemask_t *nodemask; 1894aaf14e40SMichal Hocko 1895aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 18960c397daeSMichal Hocko page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); 1897aaf14e40SMichal Hocko mpol_cond_put(mpol); 1898aaf14e40SMichal Hocko 1899aaf14e40SMichal Hocko return page; 1900099730d6SDave Hansen } 1901099730d6SDave Hansen 1902ab5ac90aSMichal Hocko /* page migration callback function */ 19033e59fcb0SMichal Hocko struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, 1904d92bbc27SJoonsoo Kim nodemask_t *nmask, gfp_t gfp_mask) 19054db9b2efSMichal Hocko { 1906*db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 19074db9b2efSMichal Hocko if (h->free_huge_pages - h->resv_huge_pages > 0) { 19083e59fcb0SMichal Hocko struct page *page; 19093e59fcb0SMichal Hocko 19103e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); 19113e59fcb0SMichal Hocko if (page) { 1912*db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 19133e59fcb0SMichal Hocko return page; 19144db9b2efSMichal Hocko } 19154db9b2efSMichal Hocko } 1916*db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 19174db9b2efSMichal Hocko 19180c397daeSMichal Hocko return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); 19194db9b2efSMichal Hocko } 19204db9b2efSMichal Hocko 1921ebd63723SMichal Hocko /* mempolicy aware migration callback */ 1922389c8178SMichal Hocko struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, 1923389c8178SMichal Hocko unsigned long address) 1924ebd63723SMichal Hocko { 1925ebd63723SMichal Hocko struct mempolicy *mpol; 1926ebd63723SMichal Hocko nodemask_t *nodemask; 1927ebd63723SMichal Hocko struct page *page; 1928ebd63723SMichal Hocko gfp_t gfp_mask; 1929ebd63723SMichal Hocko int node; 1930ebd63723SMichal Hocko 1931ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 1932ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 1933d92bbc27SJoonsoo Kim page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); 1934ebd63723SMichal Hocko mpol_cond_put(mpol); 1935ebd63723SMichal Hocko 1936ebd63723SMichal Hocko return page; 1937ebd63723SMichal Hocko } 1938ebd63723SMichal Hocko 1939bf50bab2SNaoya Horiguchi /* 194025985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 1941e4e574b7SAdam Litke * of size 'delta'. 1942e4e574b7SAdam Litke */ 19430a4f3d1bSLiu Xiang static int gather_surplus_pages(struct hstate *h, long delta) 19441b2a1e7bSJules Irenge __must_hold(&hugetlb_lock) 1945e4e574b7SAdam Litke { 1946e4e574b7SAdam Litke struct list_head surplus_list; 1947e4e574b7SAdam Litke struct page *page, *tmp; 19480a4f3d1bSLiu Xiang int ret; 19490a4f3d1bSLiu Xiang long i; 19500a4f3d1bSLiu Xiang long needed, allocated; 195128073b02SHillf Danton bool alloc_ok = true; 1952e4e574b7SAdam Litke 1953a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 1954ac09b3a1SAdam Litke if (needed <= 0) { 1955a5516438SAndi Kleen h->resv_huge_pages += delta; 1956e4e574b7SAdam Litke return 0; 1957ac09b3a1SAdam Litke } 1958e4e574b7SAdam Litke 1959e4e574b7SAdam Litke allocated = 0; 1960e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 1961e4e574b7SAdam Litke 1962e4e574b7SAdam Litke ret = -ENOMEM; 1963e4e574b7SAdam Litke retry: 1964*db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1965e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 19660c397daeSMichal Hocko page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), 1967aaf14e40SMichal Hocko NUMA_NO_NODE, NULL); 196828073b02SHillf Danton if (!page) { 196928073b02SHillf Danton alloc_ok = false; 197028073b02SHillf Danton break; 197128073b02SHillf Danton } 1972e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 197369ed779aSDavid Rientjes cond_resched(); 1974e4e574b7SAdam Litke } 197528073b02SHillf Danton allocated += i; 1976e4e574b7SAdam Litke 1977e4e574b7SAdam Litke /* 1978e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 1979e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 1980e4e574b7SAdam Litke */ 1981*db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 1982a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 1983a5516438SAndi Kleen (h->free_huge_pages + allocated); 198428073b02SHillf Danton if (needed > 0) { 198528073b02SHillf Danton if (alloc_ok) 1986e4e574b7SAdam Litke goto retry; 198728073b02SHillf Danton /* 198828073b02SHillf Danton * We were not able to allocate enough pages to 198928073b02SHillf Danton * satisfy the entire reservation so we free what 199028073b02SHillf Danton * we've allocated so far. 199128073b02SHillf Danton */ 199228073b02SHillf Danton goto free; 199328073b02SHillf Danton } 1994e4e574b7SAdam Litke /* 1995e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 199625985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 1997e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 1998ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 1999ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 2000ac09b3a1SAdam Litke * before they are reserved. 2001e4e574b7SAdam Litke */ 2002e4e574b7SAdam Litke needed += allocated; 2003a5516438SAndi Kleen h->resv_huge_pages += delta; 2004e4e574b7SAdam Litke ret = 0; 2005a9869b83SNaoya Horiguchi 200619fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 200719fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 2008e558464bSMuchun Song int zeroed; 2009e558464bSMuchun Song 201019fc3f0aSAdam Litke if ((--needed) < 0) 201119fc3f0aSAdam Litke break; 2012a9869b83SNaoya Horiguchi /* 2013a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 2014a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 2015a9869b83SNaoya Horiguchi */ 2016e558464bSMuchun Song zeroed = put_page_testzero(page); 2017e558464bSMuchun Song VM_BUG_ON_PAGE(!zeroed, page); 2018a5516438SAndi Kleen enqueue_huge_page(h, page); 201919fc3f0aSAdam Litke } 202028073b02SHillf Danton free: 2021*db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 202219fc3f0aSAdam Litke 202319fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 2024c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 2025a9869b83SNaoya Horiguchi put_page(page); 2026*db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2027e4e574b7SAdam Litke 2028e4e574b7SAdam Litke return ret; 2029e4e574b7SAdam Litke } 2030e4e574b7SAdam Litke 2031e4e574b7SAdam Litke /* 2032e5bbc8a6SMike Kravetz * This routine has two main purposes: 2033e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 2034e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 2035e5bbc8a6SMike Kravetz * to the associated reservation map. 2036e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 2037e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 2038e4e574b7SAdam Litke */ 2039a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 2040a5516438SAndi Kleen unsigned long unused_resv_pages) 2041e4e574b7SAdam Litke { 2042e4e574b7SAdam Litke unsigned long nr_pages; 204310c6ec49SMike Kravetz struct page *page; 204410c6ec49SMike Kravetz LIST_HEAD(page_list); 204510c6ec49SMike Kravetz 204610c6ec49SMike Kravetz /* Uncommit the reservation */ 204710c6ec49SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 2048e4e574b7SAdam Litke 2049aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 2050bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2051e5bbc8a6SMike Kravetz goto out; 2052aa888a74SAndi Kleen 2053e5bbc8a6SMike Kravetz /* 2054e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 2055e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 2056e5bbc8a6SMike Kravetz */ 2057a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 2058e4e574b7SAdam Litke 2059685f3457SLee Schermerhorn /* 2060685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 20619b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 20629b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 20639b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 206410c6ec49SMike Kravetz * remove_pool_huge_page() will balance the freed pages across the 20659b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 2066685f3457SLee Schermerhorn */ 2067685f3457SLee Schermerhorn while (nr_pages--) { 206810c6ec49SMike Kravetz page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1); 206910c6ec49SMike Kravetz if (!page) 2070e5bbc8a6SMike Kravetz goto out; 207110c6ec49SMike Kravetz 207210c6ec49SMike Kravetz list_add(&page->lru, &page_list); 2073e4e574b7SAdam Litke } 2074e5bbc8a6SMike Kravetz 2075e5bbc8a6SMike Kravetz out: 2076*db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 207710c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 2078*db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2079e4e574b7SAdam Litke } 2080e4e574b7SAdam Litke 20815e911373SMike Kravetz 2082c37f9fb1SAndy Whitcroft /* 2083feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 20845e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 2085cf3ad20bSMike Kravetz * 2086cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 2087cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 2088cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 2089cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 2090cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 2091feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 2092feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 2093feba16e2SMike Kravetz * is called in such cases. 2094cf3ad20bSMike Kravetz * 2095cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 2096cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 2097cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 2098cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 2099cf3ad20bSMike Kravetz * take appropriate action. 210096b96a96SMike Kravetz * 210196b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 210296b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 210396b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 210496b96a96SMike Kravetz * reservation exists. 2105c37f9fb1SAndy Whitcroft */ 21065e911373SMike Kravetz enum vma_resv_mode { 21075e911373SMike Kravetz VMA_NEEDS_RESV, 21085e911373SMike Kravetz VMA_COMMIT_RESV, 2109feba16e2SMike Kravetz VMA_END_RESV, 211096b96a96SMike Kravetz VMA_ADD_RESV, 21115e911373SMike Kravetz }; 2112cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 2113cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 21145e911373SMike Kravetz enum vma_resv_mode mode) 2115c37f9fb1SAndy Whitcroft { 21164e35f483SJoonsoo Kim struct resv_map *resv; 21174e35f483SJoonsoo Kim pgoff_t idx; 2118cf3ad20bSMike Kravetz long ret; 21190db9d74eSMina Almasry long dummy_out_regions_needed; 2120c37f9fb1SAndy Whitcroft 21214e35f483SJoonsoo Kim resv = vma_resv_map(vma); 21224e35f483SJoonsoo Kim if (!resv) 2123c37f9fb1SAndy Whitcroft return 1; 2124c37f9fb1SAndy Whitcroft 21254e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 21265e911373SMike Kravetz switch (mode) { 21275e911373SMike Kravetz case VMA_NEEDS_RESV: 21280db9d74eSMina Almasry ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); 21290db9d74eSMina Almasry /* We assume that vma_reservation_* routines always operate on 21300db9d74eSMina Almasry * 1 page, and that adding to resv map a 1 page entry can only 21310db9d74eSMina Almasry * ever require 1 region. 21320db9d74eSMina Almasry */ 21330db9d74eSMina Almasry VM_BUG_ON(dummy_out_regions_needed != 1); 21345e911373SMike Kravetz break; 21355e911373SMike Kravetz case VMA_COMMIT_RESV: 2136075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 21370db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 21380db9d74eSMina Almasry VM_BUG_ON(ret < 0); 21395e911373SMike Kravetz break; 2140feba16e2SMike Kravetz case VMA_END_RESV: 21410db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 21425e911373SMike Kravetz ret = 0; 21435e911373SMike Kravetz break; 214496b96a96SMike Kravetz case VMA_ADD_RESV: 21450db9d74eSMina Almasry if (vma->vm_flags & VM_MAYSHARE) { 2146075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 21470db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 21480db9d74eSMina Almasry VM_BUG_ON(ret < 0); 21490db9d74eSMina Almasry } else { 21500db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 215196b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 215296b96a96SMike Kravetz } 215396b96a96SMike Kravetz break; 21545e911373SMike Kravetz default: 21555e911373SMike Kravetz BUG(); 21565e911373SMike Kravetz } 215784afd99bSAndy Whitcroft 21584e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) 2159cf3ad20bSMike Kravetz return ret; 216067961f9dSMike Kravetz /* 2161bf3d12b9SMiaohe Lin * We know private mapping must have HPAGE_RESV_OWNER set. 2162bf3d12b9SMiaohe Lin * 216367961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 216467961f9dSMike Kravetz * However, a file associated with mapping could have been 216567961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 216667961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 216767961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 216867961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 216967961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 217067961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 217167961f9dSMike Kravetz * reservation has already been consumed. As a result, the 217267961f9dSMike Kravetz * return value of this routine is the opposite of the 217367961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 217467961f9dSMike Kravetz */ 2175bf3d12b9SMiaohe Lin if (ret > 0) 217667961f9dSMike Kravetz return 0; 2177bf3d12b9SMiaohe Lin if (ret == 0) 217867961f9dSMike Kravetz return 1; 2179bf3d12b9SMiaohe Lin return ret; 218084afd99bSAndy Whitcroft } 2181cf3ad20bSMike Kravetz 2182cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 2183a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 2184c37f9fb1SAndy Whitcroft { 21855e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 2186cf3ad20bSMike Kravetz } 2187c37f9fb1SAndy Whitcroft 2188cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 2189cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 2190cf3ad20bSMike Kravetz { 21915e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 21925e911373SMike Kravetz } 21935e911373SMike Kravetz 2194feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 21955e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 21965e911373SMike Kravetz { 2197feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 2198c37f9fb1SAndy Whitcroft } 2199c37f9fb1SAndy Whitcroft 220096b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 220196b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 220296b96a96SMike Kravetz { 220396b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 220496b96a96SMike Kravetz } 220596b96a96SMike Kravetz 220696b96a96SMike Kravetz /* 220796b96a96SMike Kravetz * This routine is called to restore a reservation on error paths. In the 220896b96a96SMike Kravetz * specific error paths, a huge page was allocated (via alloc_huge_page) 220996b96a96SMike Kravetz * and is about to be freed. If a reservation for the page existed, 2210d6995da3SMike Kravetz * alloc_huge_page would have consumed the reservation and set 2211d6995da3SMike Kravetz * HPageRestoreReserve in the newly allocated page. When the page is freed 2212d6995da3SMike Kravetz * via free_huge_page, the global reservation count will be incremented if 2213d6995da3SMike Kravetz * HPageRestoreReserve is set. However, free_huge_page can not adjust the 2214d6995da3SMike Kravetz * reserve map. Adjust the reserve map here to be consistent with global 2215d6995da3SMike Kravetz * reserve count adjustments to be made by free_huge_page. 221696b96a96SMike Kravetz */ 221796b96a96SMike Kravetz static void restore_reserve_on_error(struct hstate *h, 221896b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long address, 221996b96a96SMike Kravetz struct page *page) 222096b96a96SMike Kravetz { 2221d6995da3SMike Kravetz if (unlikely(HPageRestoreReserve(page))) { 222296b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 222396b96a96SMike Kravetz 222496b96a96SMike Kravetz if (unlikely(rc < 0)) { 222596b96a96SMike Kravetz /* 222696b96a96SMike Kravetz * Rare out of memory condition in reserve map 2227d6995da3SMike Kravetz * manipulation. Clear HPageRestoreReserve so that 222896b96a96SMike Kravetz * global reserve count will not be incremented 222996b96a96SMike Kravetz * by free_huge_page. This will make it appear 223096b96a96SMike Kravetz * as though the reservation for this page was 223196b96a96SMike Kravetz * consumed. This may prevent the task from 223296b96a96SMike Kravetz * faulting in the page at a later time. This 223396b96a96SMike Kravetz * is better than inconsistent global huge page 223496b96a96SMike Kravetz * accounting of reserve counts. 223596b96a96SMike Kravetz */ 2236d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 223796b96a96SMike Kravetz } else if (rc) { 223896b96a96SMike Kravetz rc = vma_add_reservation(h, vma, address); 223996b96a96SMike Kravetz if (unlikely(rc < 0)) 224096b96a96SMike Kravetz /* 224196b96a96SMike Kravetz * See above comment about rare out of 224296b96a96SMike Kravetz * memory condition. 224396b96a96SMike Kravetz */ 2244d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 224596b96a96SMike Kravetz } else 224696b96a96SMike Kravetz vma_end_reservation(h, vma, address); 224796b96a96SMike Kravetz } 224896b96a96SMike Kravetz } 224996b96a96SMike Kravetz 225070c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 225104f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 2252348ea204SAdam Litke { 225390481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 2254a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2255348ea204SAdam Litke struct page *page; 2256d85f69b0SMike Kravetz long map_chg, map_commit; 2257d85f69b0SMike Kravetz long gbl_chg; 22586d76dcf4SAneesh Kumar K.V int ret, idx; 22596d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 226008cf9fafSMina Almasry bool deferred_reserve; 22612fc39cecSAdam Litke 22626d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 2263a1e78772SMel Gorman /* 2264d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 2265d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 2266d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 2267a1e78772SMel Gorman */ 2268d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 2269d85f69b0SMike Kravetz if (map_chg < 0) 227076dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 2271d85f69b0SMike Kravetz 2272d85f69b0SMike Kravetz /* 2273d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 2274d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 2275d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 2276d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 2277d85f69b0SMike Kravetz * checked against any subpool limit. 2278d85f69b0SMike Kravetz */ 2279d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 2280d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 2281d85f69b0SMike Kravetz if (gbl_chg < 0) { 2282feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 228376dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 22845e911373SMike Kravetz } 228590d8b7e6SAdam Litke 2286d85f69b0SMike Kravetz /* 2287d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 2288d85f69b0SMike Kravetz * map, there could be reservations associated with the 2289d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 2290d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 2291d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 2292d85f69b0SMike Kravetz * the subpool reservations. 2293d85f69b0SMike Kravetz */ 2294d85f69b0SMike Kravetz if (avoid_reserve) 2295d85f69b0SMike Kravetz gbl_chg = 1; 2296d85f69b0SMike Kravetz } 2297d85f69b0SMike Kravetz 229808cf9fafSMina Almasry /* If this allocation is not consuming a reservation, charge it now. 229908cf9fafSMina Almasry */ 23006501fe5fSMiaohe Lin deferred_reserve = map_chg || avoid_reserve; 230108cf9fafSMina Almasry if (deferred_reserve) { 230208cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 230308cf9fafSMina Almasry idx, pages_per_huge_page(h), &h_cg); 23048f34af6fSJianyu Zhan if (ret) 23058f34af6fSJianyu Zhan goto out_subpool_put; 230608cf9fafSMina Almasry } 230708cf9fafSMina Almasry 230808cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 230908cf9fafSMina Almasry if (ret) 231008cf9fafSMina Almasry goto out_uncharge_cgroup_reservation; 23118f34af6fSJianyu Zhan 2312*db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2313d85f69b0SMike Kravetz /* 2314d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 2315d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 2316d85f69b0SMike Kravetz * a reservation exists for the allocation. 2317d85f69b0SMike Kravetz */ 2318d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 231981a6fcaeSJoonsoo Kim if (!page) { 2320*db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 23210c397daeSMichal Hocko page = alloc_buddy_huge_page_with_mpol(h, vma, addr); 23228f34af6fSJianyu Zhan if (!page) 23238f34af6fSJianyu Zhan goto out_uncharge_cgroup; 2324a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 2325d6995da3SMike Kravetz SetHPageRestoreReserve(page); 2326a88c7695SNaoya Horiguchi h->resv_huge_pages--; 2327a88c7695SNaoya Horiguchi } 2328*db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 232915a8d68eSWei Yang list_add(&page->lru, &h->hugepage_activelist); 233081a6fcaeSJoonsoo Kim /* Fall through */ 2331a1e78772SMel Gorman } 233281a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 233308cf9fafSMina Almasry /* If allocation is not consuming a reservation, also store the 233408cf9fafSMina Almasry * hugetlb_cgroup pointer on the page. 233508cf9fafSMina Almasry */ 233608cf9fafSMina Almasry if (deferred_reserve) { 233708cf9fafSMina Almasry hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), 233808cf9fafSMina Almasry h_cg, page); 233908cf9fafSMina Almasry } 234008cf9fafSMina Almasry 2341*db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2342a1e78772SMel Gorman 2343d6995da3SMike Kravetz hugetlb_set_page_subpool(page, spool); 2344a1e78772SMel Gorman 2345d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 2346d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 234733039678SMike Kravetz /* 234833039678SMike Kravetz * The page was added to the reservation map between 234933039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 235033039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 235133039678SMike Kravetz * Adjust for the subpool count incremented above AND 235233039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 235333039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 235433039678SMike Kravetz * no longer applies. 235533039678SMike Kravetz */ 235633039678SMike Kravetz long rsv_adjust; 235733039678SMike Kravetz 235833039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 235933039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 236079aa925bSMike Kravetz if (deferred_reserve) 236179aa925bSMike Kravetz hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 236279aa925bSMike Kravetz pages_per_huge_page(h), page); 236333039678SMike Kravetz } 23647893d1d5SAdam Litke return page; 23658f34af6fSJianyu Zhan 23668f34af6fSJianyu Zhan out_uncharge_cgroup: 23678f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 236808cf9fafSMina Almasry out_uncharge_cgroup_reservation: 236908cf9fafSMina Almasry if (deferred_reserve) 237008cf9fafSMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), 237108cf9fafSMina Almasry h_cg); 23728f34af6fSJianyu Zhan out_subpool_put: 2373d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 23748f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 2375feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 23768f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 2377b45b5bd6SDavid Gibson } 2378b45b5bd6SDavid Gibson 2379e24a1307SAneesh Kumar K.V int alloc_bootmem_huge_page(struct hstate *h) 2380e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 2381e24a1307SAneesh Kumar K.V int __alloc_bootmem_huge_page(struct hstate *h) 2382aa888a74SAndi Kleen { 2383aa888a74SAndi Kleen struct huge_bootmem_page *m; 2384b2261026SJoonsoo Kim int nr_nodes, node; 2385aa888a74SAndi Kleen 2386b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 2387aa888a74SAndi Kleen void *addr; 2388aa888a74SAndi Kleen 2389eb31d559SMike Rapoport addr = memblock_alloc_try_nid_raw( 23908b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 239197ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 2392aa888a74SAndi Kleen if (addr) { 2393aa888a74SAndi Kleen /* 2394aa888a74SAndi Kleen * Use the beginning of the huge page to store the 2395aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 2396aa888a74SAndi Kleen * puts them into the mem_map). 2397aa888a74SAndi Kleen */ 2398aa888a74SAndi Kleen m = addr; 2399aa888a74SAndi Kleen goto found; 2400aa888a74SAndi Kleen } 2401aa888a74SAndi Kleen } 2402aa888a74SAndi Kleen return 0; 2403aa888a74SAndi Kleen 2404aa888a74SAndi Kleen found: 2405df994eadSLuiz Capitulino BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); 2406aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 2407330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 2408aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 2409aa888a74SAndi Kleen m->hstate = h; 2410aa888a74SAndi Kleen return 1; 2411aa888a74SAndi Kleen } 2412aa888a74SAndi Kleen 2413d00181b9SKirill A. Shutemov static void __init prep_compound_huge_page(struct page *page, 2414d00181b9SKirill A. Shutemov unsigned int order) 241518229df5SAndy Whitcroft { 241618229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 241718229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 241818229df5SAndy Whitcroft else 241918229df5SAndy Whitcroft prep_compound_page(page, order); 242018229df5SAndy Whitcroft } 242118229df5SAndy Whitcroft 2422aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 2423aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 2424aa888a74SAndi Kleen { 2425aa888a74SAndi Kleen struct huge_bootmem_page *m; 2426aa888a74SAndi Kleen 2427aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 242840d18ebfSMike Kravetz struct page *page = virt_to_page(m); 2429aa888a74SAndi Kleen struct hstate *h = m->hstate; 2430ee8f248dSBecky Bruce 2431aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 2432c78a7f36SMiaohe Lin prep_compound_huge_page(page, huge_page_order(h)); 2433ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 2434aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 2435af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 2436af0fb9dfSMichal Hocko 2437b0320c7bSRafael Aquini /* 2438b0320c7bSRafael Aquini * If we had gigantic hugepages allocated at boot time, we need 2439b0320c7bSRafael Aquini * to restore the 'stolen' pages to totalram_pages in order to 2440b0320c7bSRafael Aquini * fix confusing memory reports from free(1) and another 2441b0320c7bSRafael Aquini * side-effects, like CommitLimit going negative. 2442b0320c7bSRafael Aquini */ 2443bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2444c78a7f36SMiaohe Lin adjust_managed_page_count(page, pages_per_huge_page(h)); 2445520495feSCannon Matthews cond_resched(); 2446aa888a74SAndi Kleen } 2447aa888a74SAndi Kleen } 2448aa888a74SAndi Kleen 24498faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 24501da177e4SLinus Torvalds { 24511da177e4SLinus Torvalds unsigned long i; 2452f60858f9SMike Kravetz nodemask_t *node_alloc_noretry; 2453f60858f9SMike Kravetz 2454f60858f9SMike Kravetz if (!hstate_is_gigantic(h)) { 2455f60858f9SMike Kravetz /* 2456f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 2457f60858f9SMike Kravetz * Ignore errors as lower level routines can deal with 2458f60858f9SMike Kravetz * node_alloc_noretry == NULL. If this kmalloc fails at boot 2459f60858f9SMike Kravetz * time, we are likely in bigger trouble. 2460f60858f9SMike Kravetz */ 2461f60858f9SMike Kravetz node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), 2462f60858f9SMike Kravetz GFP_KERNEL); 2463f60858f9SMike Kravetz } else { 2464f60858f9SMike Kravetz /* allocations done at boot time */ 2465f60858f9SMike Kravetz node_alloc_noretry = NULL; 2466f60858f9SMike Kravetz } 2467f60858f9SMike Kravetz 2468f60858f9SMike Kravetz /* bit mask controlling how hard we retry per-node allocations */ 2469f60858f9SMike Kravetz if (node_alloc_noretry) 2470f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 24711da177e4SLinus Torvalds 2472e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 2473bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 2474dbda8feaSBarry Song if (hugetlb_cma_size) { 2475cf11e85fSRoman Gushchin pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); 24767ecc9565SChen Wandun goto free; 2477cf11e85fSRoman Gushchin } 2478aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 2479aa888a74SAndi Kleen break; 24800c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 2481f60858f9SMike Kravetz &node_states[N_MEMORY], 2482f60858f9SMike Kravetz node_alloc_noretry)) 24831da177e4SLinus Torvalds break; 248469ed779aSDavid Rientjes cond_resched(); 24851da177e4SLinus Torvalds } 2486d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 2487d715cf80SLiam R. Howlett char buf[32]; 2488d715cf80SLiam R. Howlett 2489c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2490d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 2491d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 24928faa8b07SAndi Kleen h->max_huge_pages = i; 2493e5ff2159SAndi Kleen } 24947ecc9565SChen Wandun free: 2495f60858f9SMike Kravetz kfree(node_alloc_noretry); 2496d715cf80SLiam R. Howlett } 2497e5ff2159SAndi Kleen 2498e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 2499e5ff2159SAndi Kleen { 2500e5ff2159SAndi Kleen struct hstate *h; 2501e5ff2159SAndi Kleen 2502e5ff2159SAndi Kleen for_each_hstate(h) { 2503641844f5SNaoya Horiguchi if (minimum_order > huge_page_order(h)) 2504641844f5SNaoya Horiguchi minimum_order = huge_page_order(h); 2505641844f5SNaoya Horiguchi 25068faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 2507bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 25088faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 2509e5ff2159SAndi Kleen } 2510641844f5SNaoya Horiguchi VM_BUG_ON(minimum_order == UINT_MAX); 2511e5ff2159SAndi Kleen } 2512e5ff2159SAndi Kleen 2513e5ff2159SAndi Kleen static void __init report_hugepages(void) 2514e5ff2159SAndi Kleen { 2515e5ff2159SAndi Kleen struct hstate *h; 2516e5ff2159SAndi Kleen 2517e5ff2159SAndi Kleen for_each_hstate(h) { 25184abd32dbSAndi Kleen char buf[32]; 2519c6247f72SMatthew Wilcox 2520c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 2521ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 2522c6247f72SMatthew Wilcox buf, h->free_huge_pages); 2523e5ff2159SAndi Kleen } 2524e5ff2159SAndi Kleen } 2525e5ff2159SAndi Kleen 25261da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 25276ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 25286ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 25291da177e4SLinus Torvalds { 25304415cc8dSChristoph Lameter int i; 25311121828aSMike Kravetz LIST_HEAD(page_list); 25324415cc8dSChristoph Lameter 2533bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2534aa888a74SAndi Kleen return; 2535aa888a74SAndi Kleen 25361121828aSMike Kravetz /* 25371121828aSMike Kravetz * Collect pages to be freed on a list, and free after dropping lock 25381121828aSMike Kravetz */ 25396ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 254010c6ec49SMike Kravetz struct page *page, *next; 2541a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 2542a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 2543a5516438SAndi Kleen if (count >= h->nr_huge_pages) 25441121828aSMike Kravetz goto out; 25451da177e4SLinus Torvalds if (PageHighMem(page)) 25461da177e4SLinus Torvalds continue; 25476eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 25481121828aSMike Kravetz list_add(&page->lru, &page_list); 25491121828aSMike Kravetz } 25501121828aSMike Kravetz } 25511121828aSMike Kravetz 25521121828aSMike Kravetz out: 2553*db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 255410c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 2555*db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 25561da177e4SLinus Torvalds } 25571da177e4SLinus Torvalds #else 25586ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 25596ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 25601da177e4SLinus Torvalds { 25611da177e4SLinus Torvalds } 25621da177e4SLinus Torvalds #endif 25631da177e4SLinus Torvalds 256420a0307cSWu Fengguang /* 256520a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 256620a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 256720a0307cSWu Fengguang * Returns 1 if an adjustment was made. 256820a0307cSWu Fengguang */ 25696ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 25706ae11b27SLee Schermerhorn int delta) 257120a0307cSWu Fengguang { 2572b2261026SJoonsoo Kim int nr_nodes, node; 257320a0307cSWu Fengguang 257420a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 257520a0307cSWu Fengguang 2576e8c5c824SLee Schermerhorn if (delta < 0) { 2577b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 2578b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 2579b2261026SJoonsoo Kim goto found; 2580b2261026SJoonsoo Kim } 2581b2261026SJoonsoo Kim } else { 2582b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2583b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 2584b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 2585b2261026SJoonsoo Kim goto found; 2586e8c5c824SLee Schermerhorn } 25879a76db09SLee Schermerhorn } 2588b2261026SJoonsoo Kim return 0; 258920a0307cSWu Fengguang 2590b2261026SJoonsoo Kim found: 259120a0307cSWu Fengguang h->surplus_huge_pages += delta; 2592b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 2593b2261026SJoonsoo Kim return 1; 259420a0307cSWu Fengguang } 259520a0307cSWu Fengguang 2596a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 2597fd875dcaSMike Kravetz static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 25986ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 25991da177e4SLinus Torvalds { 26007893d1d5SAdam Litke unsigned long min_count, ret; 260110c6ec49SMike Kravetz struct page *page; 260210c6ec49SMike Kravetz LIST_HEAD(page_list); 2603f60858f9SMike Kravetz NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); 2604f60858f9SMike Kravetz 2605f60858f9SMike Kravetz /* 2606f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 2607f60858f9SMike Kravetz * If we can not allocate the bit mask, do not attempt to allocate 2608f60858f9SMike Kravetz * the requested huge pages. 2609f60858f9SMike Kravetz */ 2610f60858f9SMike Kravetz if (node_alloc_noretry) 2611f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 2612f60858f9SMike Kravetz else 2613f60858f9SMike Kravetz return -ENOMEM; 26141da177e4SLinus Torvalds 261529383967SMike Kravetz /* 261629383967SMike Kravetz * resize_lock mutex prevents concurrent adjustments to number of 261729383967SMike Kravetz * pages in hstate via the proc/sysfs interfaces. 261829383967SMike Kravetz */ 261929383967SMike Kravetz mutex_lock(&h->resize_lock); 2620*db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 26214eb0716eSAlexandre Ghiti 26224eb0716eSAlexandre Ghiti /* 2623fd875dcaSMike Kravetz * Check for a node specific request. 2624fd875dcaSMike Kravetz * Changing node specific huge page count may require a corresponding 2625fd875dcaSMike Kravetz * change to the global count. In any case, the passed node mask 2626fd875dcaSMike Kravetz * (nodes_allowed) will restrict alloc/free to the specified node. 2627fd875dcaSMike Kravetz */ 2628fd875dcaSMike Kravetz if (nid != NUMA_NO_NODE) { 2629fd875dcaSMike Kravetz unsigned long old_count = count; 2630fd875dcaSMike Kravetz 2631fd875dcaSMike Kravetz count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 2632fd875dcaSMike Kravetz /* 2633fd875dcaSMike Kravetz * User may have specified a large count value which caused the 2634fd875dcaSMike Kravetz * above calculation to overflow. In this case, they wanted 2635fd875dcaSMike Kravetz * to allocate as many huge pages as possible. Set count to 2636fd875dcaSMike Kravetz * largest possible value to align with their intention. 2637fd875dcaSMike Kravetz */ 2638fd875dcaSMike Kravetz if (count < old_count) 2639fd875dcaSMike Kravetz count = ULONG_MAX; 2640fd875dcaSMike Kravetz } 2641fd875dcaSMike Kravetz 2642fd875dcaSMike Kravetz /* 26434eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 26444eb0716eSAlexandre Ghiti * page range allocation. 26454eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 26464eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 26474eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 26484eb0716eSAlexandre Ghiti */ 26494eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 26504eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 2651*db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 265229383967SMike Kravetz mutex_unlock(&h->resize_lock); 2653f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 26544eb0716eSAlexandre Ghiti return -EINVAL; 26554eb0716eSAlexandre Ghiti } 26564eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 26574eb0716eSAlexandre Ghiti } 2658aa888a74SAndi Kleen 26597893d1d5SAdam Litke /* 26607893d1d5SAdam Litke * Increase the pool size 26617893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 26627893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 2663d1c3fb1fSNishanth Aravamudan * 26640c397daeSMichal Hocko * We might race with alloc_surplus_huge_page() here and be unable 2665d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 2666d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 2667d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 2668d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 26697893d1d5SAdam Litke */ 2670a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 26716ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 26727893d1d5SAdam Litke break; 26737893d1d5SAdam Litke } 26747893d1d5SAdam Litke 2675a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 26767893d1d5SAdam Litke /* 26777893d1d5SAdam Litke * If this allocation races such that we no longer need the 26787893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 26797893d1d5SAdam Litke * and reducing the surplus. 26807893d1d5SAdam Litke */ 2681*db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2682649920c6SJia He 2683649920c6SJia He /* yield cpu to avoid soft lockup */ 2684649920c6SJia He cond_resched(); 2685649920c6SJia He 2686f60858f9SMike Kravetz ret = alloc_pool_huge_page(h, nodes_allowed, 2687f60858f9SMike Kravetz node_alloc_noretry); 2688*db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 26897893d1d5SAdam Litke if (!ret) 26907893d1d5SAdam Litke goto out; 26917893d1d5SAdam Litke 2692536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 2693536240f2SMel Gorman if (signal_pending(current)) 2694536240f2SMel Gorman goto out; 26957893d1d5SAdam Litke } 26967893d1d5SAdam Litke 26977893d1d5SAdam Litke /* 26987893d1d5SAdam Litke * Decrease the pool size 26997893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 27007893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 27017893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 27027893d1d5SAdam Litke * to the desired size as pages become free. 2703d1c3fb1fSNishanth Aravamudan * 2704d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 2705d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 2706d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 27070c397daeSMichal Hocko * alloc_surplus_huge_page() is checking the global counter, 2708d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 2709d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 2710d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 27117893d1d5SAdam Litke */ 2712a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 27136b0c880dSAdam Litke min_count = max(count, min_count); 27146ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 271510c6ec49SMike Kravetz 271610c6ec49SMike Kravetz /* 271710c6ec49SMike Kravetz * Collect pages to be removed on list without dropping lock 271810c6ec49SMike Kravetz */ 2719a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 272010c6ec49SMike Kravetz page = remove_pool_huge_page(h, nodes_allowed, 0); 272110c6ec49SMike Kravetz if (!page) 27221da177e4SLinus Torvalds break; 272310c6ec49SMike Kravetz 272410c6ec49SMike Kravetz list_add(&page->lru, &page_list); 27251da177e4SLinus Torvalds } 272610c6ec49SMike Kravetz /* free the pages after dropping lock */ 2727*db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 272810c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 2729*db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 273010c6ec49SMike Kravetz 2731a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 27326ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 27337893d1d5SAdam Litke break; 27347893d1d5SAdam Litke } 27357893d1d5SAdam Litke out: 27364eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 2737*db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 273829383967SMike Kravetz mutex_unlock(&h->resize_lock); 27394eb0716eSAlexandre Ghiti 2740f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 2741f60858f9SMike Kravetz 27424eb0716eSAlexandre Ghiti return 0; 27431da177e4SLinus Torvalds } 27441da177e4SLinus Torvalds 2745a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 2746a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 2747a3437870SNishanth Aravamudan 2748a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 2749a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 2750a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 2751a3437870SNishanth Aravamudan 2752a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 2753a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 2754a3437870SNishanth Aravamudan 27559a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 27569a305230SLee Schermerhorn 27579a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 2758a3437870SNishanth Aravamudan { 2759a3437870SNishanth Aravamudan int i; 27609a305230SLee Schermerhorn 2761a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 27629a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 27639a305230SLee Schermerhorn if (nidp) 27649a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 2765a3437870SNishanth Aravamudan return &hstates[i]; 27669a305230SLee Schermerhorn } 27679a305230SLee Schermerhorn 27689a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 2769a3437870SNishanth Aravamudan } 2770a3437870SNishanth Aravamudan 277106808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 2772a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2773a3437870SNishanth Aravamudan { 27749a305230SLee Schermerhorn struct hstate *h; 27759a305230SLee Schermerhorn unsigned long nr_huge_pages; 27769a305230SLee Schermerhorn int nid; 27779a305230SLee Schermerhorn 27789a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 27799a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 27809a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 27819a305230SLee Schermerhorn else 27829a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 27839a305230SLee Schermerhorn 2784ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", nr_huge_pages); 2785a3437870SNishanth Aravamudan } 2786adbe8726SEric B Munson 2787238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 2788238d3c13SDavid Rientjes struct hstate *h, int nid, 2789238d3c13SDavid Rientjes unsigned long count, size_t len) 2790a3437870SNishanth Aravamudan { 2791a3437870SNishanth Aravamudan int err; 27922d0adf7eSOscar Salvador nodemask_t nodes_allowed, *n_mask; 2793a3437870SNishanth Aravamudan 27942d0adf7eSOscar Salvador if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 27952d0adf7eSOscar Salvador return -EINVAL; 2796adbe8726SEric B Munson 27979a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 27989a305230SLee Schermerhorn /* 27999a305230SLee Schermerhorn * global hstate attribute 28009a305230SLee Schermerhorn */ 28019a305230SLee Schermerhorn if (!(obey_mempolicy && 28022d0adf7eSOscar Salvador init_nodemask_of_mempolicy(&nodes_allowed))) 28032d0adf7eSOscar Salvador n_mask = &node_states[N_MEMORY]; 28042d0adf7eSOscar Salvador else 28052d0adf7eSOscar Salvador n_mask = &nodes_allowed; 28062d0adf7eSOscar Salvador } else { 28079a305230SLee Schermerhorn /* 2808fd875dcaSMike Kravetz * Node specific request. count adjustment happens in 2809fd875dcaSMike Kravetz * set_max_huge_pages() after acquiring hugetlb_lock. 28109a305230SLee Schermerhorn */ 28112d0adf7eSOscar Salvador init_nodemask_of_node(&nodes_allowed, nid); 28122d0adf7eSOscar Salvador n_mask = &nodes_allowed; 2813fd875dcaSMike Kravetz } 28149a305230SLee Schermerhorn 28152d0adf7eSOscar Salvador err = set_max_huge_pages(h, count, nid, n_mask); 281606808b08SLee Schermerhorn 28174eb0716eSAlexandre Ghiti return err ? err : len; 281806808b08SLee Schermerhorn } 281906808b08SLee Schermerhorn 2820238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 2821238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 2822238d3c13SDavid Rientjes size_t len) 2823238d3c13SDavid Rientjes { 2824238d3c13SDavid Rientjes struct hstate *h; 2825238d3c13SDavid Rientjes unsigned long count; 2826238d3c13SDavid Rientjes int nid; 2827238d3c13SDavid Rientjes int err; 2828238d3c13SDavid Rientjes 2829238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 2830238d3c13SDavid Rientjes if (err) 2831238d3c13SDavid Rientjes return err; 2832238d3c13SDavid Rientjes 2833238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 2834238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 2835238d3c13SDavid Rientjes } 2836238d3c13SDavid Rientjes 283706808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 283806808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 283906808b08SLee Schermerhorn { 284006808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 284106808b08SLee Schermerhorn } 284206808b08SLee Schermerhorn 284306808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 284406808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 284506808b08SLee Schermerhorn { 2846238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 2847a3437870SNishanth Aravamudan } 2848a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 2849a3437870SNishanth Aravamudan 285006808b08SLee Schermerhorn #ifdef CONFIG_NUMA 285106808b08SLee Schermerhorn 285206808b08SLee Schermerhorn /* 285306808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 285406808b08SLee Schermerhorn * huge page alloc/free. 285506808b08SLee Schermerhorn */ 285606808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 2857ae7a927dSJoe Perches struct kobj_attribute *attr, 2858ae7a927dSJoe Perches char *buf) 285906808b08SLee Schermerhorn { 286006808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 286106808b08SLee Schermerhorn } 286206808b08SLee Schermerhorn 286306808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 286406808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 286506808b08SLee Schermerhorn { 2866238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 286706808b08SLee Schermerhorn } 286806808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 286906808b08SLee Schermerhorn #endif 287006808b08SLee Schermerhorn 287106808b08SLee Schermerhorn 2872a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 2873a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2874a3437870SNishanth Aravamudan { 28759a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2876ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); 2877a3437870SNishanth Aravamudan } 2878adbe8726SEric B Munson 2879a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 2880a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 2881a3437870SNishanth Aravamudan { 2882a3437870SNishanth Aravamudan int err; 2883a3437870SNishanth Aravamudan unsigned long input; 28849a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2885a3437870SNishanth Aravamudan 2886bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2887adbe8726SEric B Munson return -EINVAL; 2888adbe8726SEric B Munson 28893dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 2890a3437870SNishanth Aravamudan if (err) 289173ae31e5SEric B Munson return err; 2892a3437870SNishanth Aravamudan 2893*db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2894a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 2895*db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2896a3437870SNishanth Aravamudan 2897a3437870SNishanth Aravamudan return count; 2898a3437870SNishanth Aravamudan } 2899a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 2900a3437870SNishanth Aravamudan 2901a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 2902a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2903a3437870SNishanth Aravamudan { 29049a305230SLee Schermerhorn struct hstate *h; 29059a305230SLee Schermerhorn unsigned long free_huge_pages; 29069a305230SLee Schermerhorn int nid; 29079a305230SLee Schermerhorn 29089a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 29099a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 29109a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 29119a305230SLee Schermerhorn else 29129a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 29139a305230SLee Schermerhorn 2914ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", free_huge_pages); 2915a3437870SNishanth Aravamudan } 2916a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 2917a3437870SNishanth Aravamudan 2918a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 2919a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2920a3437870SNishanth Aravamudan { 29219a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 2922ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); 2923a3437870SNishanth Aravamudan } 2924a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 2925a3437870SNishanth Aravamudan 2926a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 2927a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 2928a3437870SNishanth Aravamudan { 29299a305230SLee Schermerhorn struct hstate *h; 29309a305230SLee Schermerhorn unsigned long surplus_huge_pages; 29319a305230SLee Schermerhorn int nid; 29329a305230SLee Schermerhorn 29339a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 29349a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 29359a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 29369a305230SLee Schermerhorn else 29379a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 29389a305230SLee Schermerhorn 2939ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", surplus_huge_pages); 2940a3437870SNishanth Aravamudan } 2941a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 2942a3437870SNishanth Aravamudan 2943a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 2944a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 2945a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 2946a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 2947a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 2948a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 294906808b08SLee Schermerhorn #ifdef CONFIG_NUMA 295006808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 295106808b08SLee Schermerhorn #endif 2952a3437870SNishanth Aravamudan NULL, 2953a3437870SNishanth Aravamudan }; 2954a3437870SNishanth Aravamudan 295567e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 2956a3437870SNishanth Aravamudan .attrs = hstate_attrs, 2957a3437870SNishanth Aravamudan }; 2958a3437870SNishanth Aravamudan 2959094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 29609a305230SLee Schermerhorn struct kobject **hstate_kobjs, 296167e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 2962a3437870SNishanth Aravamudan { 2963a3437870SNishanth Aravamudan int retval; 2964972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 2965a3437870SNishanth Aravamudan 29669a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 29679a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 2968a3437870SNishanth Aravamudan return -ENOMEM; 2969a3437870SNishanth Aravamudan 29709a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 2971cc2205a6SMiaohe Lin if (retval) { 29729a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 2973cc2205a6SMiaohe Lin hstate_kobjs[hi] = NULL; 2974cc2205a6SMiaohe Lin } 2975a3437870SNishanth Aravamudan 2976a3437870SNishanth Aravamudan return retval; 2977a3437870SNishanth Aravamudan } 2978a3437870SNishanth Aravamudan 2979a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 2980a3437870SNishanth Aravamudan { 2981a3437870SNishanth Aravamudan struct hstate *h; 2982a3437870SNishanth Aravamudan int err; 2983a3437870SNishanth Aravamudan 2984a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 2985a3437870SNishanth Aravamudan if (!hugepages_kobj) 2986a3437870SNishanth Aravamudan return; 2987a3437870SNishanth Aravamudan 2988a3437870SNishanth Aravamudan for_each_hstate(h) { 29899a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 29909a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 2991a3437870SNishanth Aravamudan if (err) 2992282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s", h->name); 2993a3437870SNishanth Aravamudan } 2994a3437870SNishanth Aravamudan } 2995a3437870SNishanth Aravamudan 29969a305230SLee Schermerhorn #ifdef CONFIG_NUMA 29979a305230SLee Schermerhorn 29989a305230SLee Schermerhorn /* 29999a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 300010fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 300110fbcf4cSKay Sievers * index of a node device or _hstate == node id. 300210fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 30039a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 30049a305230SLee Schermerhorn */ 30059a305230SLee Schermerhorn struct node_hstate { 30069a305230SLee Schermerhorn struct kobject *hugepages_kobj; 30079a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 30089a305230SLee Schermerhorn }; 3009b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 30109a305230SLee Schermerhorn 30119a305230SLee Schermerhorn /* 301210fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 30139a305230SLee Schermerhorn */ 30149a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 30159a305230SLee Schermerhorn &nr_hugepages_attr.attr, 30169a305230SLee Schermerhorn &free_hugepages_attr.attr, 30179a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 30189a305230SLee Schermerhorn NULL, 30199a305230SLee Schermerhorn }; 30209a305230SLee Schermerhorn 302167e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 30229a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 30239a305230SLee Schermerhorn }; 30249a305230SLee Schermerhorn 30259a305230SLee Schermerhorn /* 302610fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 30279a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 30289a305230SLee Schermerhorn */ 30299a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 30309a305230SLee Schermerhorn { 30319a305230SLee Schermerhorn int nid; 30329a305230SLee Schermerhorn 30339a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 30349a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 30359a305230SLee Schermerhorn int i; 30369a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 30379a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 30389a305230SLee Schermerhorn if (nidp) 30399a305230SLee Schermerhorn *nidp = nid; 30409a305230SLee Schermerhorn return &hstates[i]; 30419a305230SLee Schermerhorn } 30429a305230SLee Schermerhorn } 30439a305230SLee Schermerhorn 30449a305230SLee Schermerhorn BUG(); 30459a305230SLee Schermerhorn return NULL; 30469a305230SLee Schermerhorn } 30479a305230SLee Schermerhorn 30489a305230SLee Schermerhorn /* 304910fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 30509a305230SLee Schermerhorn * No-op if no hstate attributes attached. 30519a305230SLee Schermerhorn */ 30523cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 30539a305230SLee Schermerhorn { 30549a305230SLee Schermerhorn struct hstate *h; 305510fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 30569a305230SLee Schermerhorn 30579a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 30589b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 30599a305230SLee Schermerhorn 3060972dc4deSAneesh Kumar K.V for_each_hstate(h) { 3061972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 3062972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 3063972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 3064972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 3065972dc4deSAneesh Kumar K.V } 30669a305230SLee Schermerhorn } 30679a305230SLee Schermerhorn 30689a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 30699a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 30709a305230SLee Schermerhorn } 30719a305230SLee Schermerhorn 30729a305230SLee Schermerhorn 30739a305230SLee Schermerhorn /* 307410fbcf4cSKay Sievers * Register hstate attributes for a single node device. 30759a305230SLee Schermerhorn * No-op if attributes already registered. 30769a305230SLee Schermerhorn */ 30773cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 30789a305230SLee Schermerhorn { 30799a305230SLee Schermerhorn struct hstate *h; 308010fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 30819a305230SLee Schermerhorn int err; 30829a305230SLee Schermerhorn 30839a305230SLee Schermerhorn if (nhs->hugepages_kobj) 30849a305230SLee Schermerhorn return; /* already allocated */ 30859a305230SLee Schermerhorn 30869a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 308710fbcf4cSKay Sievers &node->dev.kobj); 30889a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 30899a305230SLee Schermerhorn return; 30909a305230SLee Schermerhorn 30919a305230SLee Schermerhorn for_each_hstate(h) { 30929a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 30939a305230SLee Schermerhorn nhs->hstate_kobjs, 30949a305230SLee Schermerhorn &per_node_hstate_attr_group); 30959a305230SLee Schermerhorn if (err) { 3096282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s for node %d\n", 309710fbcf4cSKay Sievers h->name, node->dev.id); 30989a305230SLee Schermerhorn hugetlb_unregister_node(node); 30999a305230SLee Schermerhorn break; 31009a305230SLee Schermerhorn } 31019a305230SLee Schermerhorn } 31029a305230SLee Schermerhorn } 31039a305230SLee Schermerhorn 31049a305230SLee Schermerhorn /* 31059b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 310610fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 310710fbcf4cSKay Sievers * registered their associated device by this time. 31089a305230SLee Schermerhorn */ 31097d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 31109a305230SLee Schermerhorn { 31119a305230SLee Schermerhorn int nid; 31129a305230SLee Schermerhorn 31138cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 31148732794bSWen Congyang struct node *node = node_devices[nid]; 311510fbcf4cSKay Sievers if (node->dev.id == nid) 31169a305230SLee Schermerhorn hugetlb_register_node(node); 31179a305230SLee Schermerhorn } 31189a305230SLee Schermerhorn 31199a305230SLee Schermerhorn /* 312010fbcf4cSKay Sievers * Let the node device driver know we're here so it can 31219a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 31229a305230SLee Schermerhorn */ 31239a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 31249a305230SLee Schermerhorn hugetlb_unregister_node); 31259a305230SLee Schermerhorn } 31269a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 31279a305230SLee Schermerhorn 31289a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 31299a305230SLee Schermerhorn { 31309a305230SLee Schermerhorn BUG(); 31319a305230SLee Schermerhorn if (nidp) 31329a305230SLee Schermerhorn *nidp = -1; 31339a305230SLee Schermerhorn return NULL; 31349a305230SLee Schermerhorn } 31359a305230SLee Schermerhorn 31369a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 31379a305230SLee Schermerhorn 31389a305230SLee Schermerhorn #endif 31399a305230SLee Schermerhorn 3140a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 3141a3437870SNishanth Aravamudan { 31428382d914SDavidlohr Bueso int i; 31438382d914SDavidlohr Bueso 3144d6995da3SMike Kravetz BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < 3145d6995da3SMike Kravetz __NR_HPAGEFLAGS); 3146d6995da3SMike Kravetz 3147c2833a5bSMike Kravetz if (!hugepages_supported()) { 3148c2833a5bSMike Kravetz if (hugetlb_max_hstate || default_hstate_max_huge_pages) 3149c2833a5bSMike Kravetz pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); 31500ef89d25SBenjamin Herrenschmidt return 0; 3151d715cf80SLiam R. Howlett } 3152d715cf80SLiam R. Howlett 3153282f4214SMike Kravetz /* 3154282f4214SMike Kravetz * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some 3155282f4214SMike Kravetz * architectures depend on setup being done here. 3156282f4214SMike Kravetz */ 3157a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 3158282f4214SMike Kravetz if (!parsed_default_hugepagesz) { 3159282f4214SMike Kravetz /* 3160282f4214SMike Kravetz * If we did not parse a default huge page size, set 3161282f4214SMike Kravetz * default_hstate_idx to HPAGE_SIZE hstate. And, if the 3162282f4214SMike Kravetz * number of huge pages for this default size was implicitly 3163282f4214SMike Kravetz * specified, set that here as well. 3164282f4214SMike Kravetz * Note that the implicit setting will overwrite an explicit 3165282f4214SMike Kravetz * setting. A warning will be printed in this case. 3166282f4214SMike Kravetz */ 3167282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); 3168f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 3169282f4214SMike Kravetz if (default_hstate.max_huge_pages) { 3170282f4214SMike Kravetz char buf[32]; 3171282f4214SMike Kravetz 3172282f4214SMike Kravetz string_get_size(huge_page_size(&default_hstate), 3173282f4214SMike Kravetz 1, STRING_UNITS_2, buf, 32); 3174282f4214SMike Kravetz pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", 3175282f4214SMike Kravetz default_hstate.max_huge_pages, buf); 3176282f4214SMike Kravetz pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", 3177282f4214SMike Kravetz default_hstate_max_huge_pages); 3178282f4214SMike Kravetz } 3179282f4214SMike Kravetz default_hstate.max_huge_pages = 3180282f4214SMike Kravetz default_hstate_max_huge_pages; 3181282f4214SMike Kravetz } 3182f8b74815SVaishali Thakkar } 3183a3437870SNishanth Aravamudan 3184cf11e85fSRoman Gushchin hugetlb_cma_check(); 3185a3437870SNishanth Aravamudan hugetlb_init_hstates(); 3186aa888a74SAndi Kleen gather_bootmem_prealloc(); 3187a3437870SNishanth Aravamudan report_hugepages(); 3188a3437870SNishanth Aravamudan 3189a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 31909a305230SLee Schermerhorn hugetlb_register_all_nodes(); 31917179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 31929a305230SLee Schermerhorn 31938382d914SDavidlohr Bueso #ifdef CONFIG_SMP 31948382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 31958382d914SDavidlohr Bueso #else 31968382d914SDavidlohr Bueso num_fault_mutexes = 1; 31978382d914SDavidlohr Bueso #endif 3198c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 31996da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 32006da2ec56SKees Cook GFP_KERNEL); 3201c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 32028382d914SDavidlohr Bueso 32038382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 3204c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 3205a3437870SNishanth Aravamudan return 0; 3206a3437870SNishanth Aravamudan } 32073e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 3208a3437870SNishanth Aravamudan 3209ae94da89SMike Kravetz /* Overwritten by architectures with more huge page sizes */ 3210ae94da89SMike Kravetz bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) 32119fee021dSVaishali Thakkar { 3212ae94da89SMike Kravetz return size == HPAGE_SIZE; 32139fee021dSVaishali Thakkar } 32149fee021dSVaishali Thakkar 3215d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 3216a3437870SNishanth Aravamudan { 3217a3437870SNishanth Aravamudan struct hstate *h; 32188faa8b07SAndi Kleen unsigned long i; 32198faa8b07SAndi Kleen 3220a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 3221a3437870SNishanth Aravamudan return; 3222a3437870SNishanth Aravamudan } 322347d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 3224a3437870SNishanth Aravamudan BUG_ON(order == 0); 322547d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 322629383967SMike Kravetz mutex_init(&h->resize_lock); 3227a3437870SNishanth Aravamudan h->order = order; 3228aca78307SMiaohe Lin h->mask = ~(huge_page_size(h) - 1); 32298faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 32308faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 32310edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 323254f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 323354f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 3234a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 3235a3437870SNishanth Aravamudan huge_page_size(h)/1024); 32368faa8b07SAndi Kleen 3237a3437870SNishanth Aravamudan parsed_hstate = h; 3238a3437870SNishanth Aravamudan } 3239a3437870SNishanth Aravamudan 3240282f4214SMike Kravetz /* 3241282f4214SMike Kravetz * hugepages command line processing 3242282f4214SMike Kravetz * hugepages normally follows a valid hugepagsz or default_hugepagsz 3243282f4214SMike Kravetz * specification. If not, ignore the hugepages value. hugepages can also 3244282f4214SMike Kravetz * be the first huge page command line option in which case it implicitly 3245282f4214SMike Kravetz * specifies the number of huge pages for the default size. 3246282f4214SMike Kravetz */ 3247282f4214SMike Kravetz static int __init hugepages_setup(char *s) 3248a3437870SNishanth Aravamudan { 3249a3437870SNishanth Aravamudan unsigned long *mhp; 32508faa8b07SAndi Kleen static unsigned long *last_mhp; 3251a3437870SNishanth Aravamudan 32529fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 3253282f4214SMike Kravetz pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); 32549fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 3255282f4214SMike Kravetz return 0; 32569fee021dSVaishali Thakkar } 3257282f4214SMike Kravetz 3258a3437870SNishanth Aravamudan /* 3259282f4214SMike Kravetz * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter 3260282f4214SMike Kravetz * yet, so this hugepages= parameter goes to the "default hstate". 3261282f4214SMike Kravetz * Otherwise, it goes with the previously parsed hugepagesz or 3262282f4214SMike Kravetz * default_hugepagesz. 3263a3437870SNishanth Aravamudan */ 32649fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 3265a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 3266a3437870SNishanth Aravamudan else 3267a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 3268a3437870SNishanth Aravamudan 32698faa8b07SAndi Kleen if (mhp == last_mhp) { 3270282f4214SMike Kravetz pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); 3271282f4214SMike Kravetz return 0; 32728faa8b07SAndi Kleen } 32738faa8b07SAndi Kleen 3274a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 3275a3437870SNishanth Aravamudan *mhp = 0; 3276a3437870SNishanth Aravamudan 32778faa8b07SAndi Kleen /* 32788faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 327904adbc3fSMiaohe Lin * But we need to allocate gigantic hstates here early to still 32808faa8b07SAndi Kleen * use the bootmem allocator. 32818faa8b07SAndi Kleen */ 328204adbc3fSMiaohe Lin if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) 32838faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 32848faa8b07SAndi Kleen 32858faa8b07SAndi Kleen last_mhp = mhp; 32868faa8b07SAndi Kleen 3287a3437870SNishanth Aravamudan return 1; 3288a3437870SNishanth Aravamudan } 3289282f4214SMike Kravetz __setup("hugepages=", hugepages_setup); 3290e11bfbfcSNick Piggin 3291282f4214SMike Kravetz /* 3292282f4214SMike Kravetz * hugepagesz command line processing 3293282f4214SMike Kravetz * A specific huge page size can only be specified once with hugepagesz. 3294282f4214SMike Kravetz * hugepagesz is followed by hugepages on the command line. The global 3295282f4214SMike Kravetz * variable 'parsed_valid_hugepagesz' is used to determine if prior 3296282f4214SMike Kravetz * hugepagesz argument was valid. 3297282f4214SMike Kravetz */ 3298359f2544SMike Kravetz static int __init hugepagesz_setup(char *s) 3299e11bfbfcSNick Piggin { 3300359f2544SMike Kravetz unsigned long size; 3301282f4214SMike Kravetz struct hstate *h; 3302282f4214SMike Kravetz 3303282f4214SMike Kravetz parsed_valid_hugepagesz = false; 3304359f2544SMike Kravetz size = (unsigned long)memparse(s, NULL); 3305359f2544SMike Kravetz 3306359f2544SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 3307282f4214SMike Kravetz pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); 3308359f2544SMike Kravetz return 0; 3309359f2544SMike Kravetz } 3310359f2544SMike Kravetz 3311282f4214SMike Kravetz h = size_to_hstate(size); 3312282f4214SMike Kravetz if (h) { 3313282f4214SMike Kravetz /* 3314282f4214SMike Kravetz * hstate for this size already exists. This is normally 3315282f4214SMike Kravetz * an error, but is allowed if the existing hstate is the 3316282f4214SMike Kravetz * default hstate. More specifically, it is only allowed if 3317282f4214SMike Kravetz * the number of huge pages for the default hstate was not 3318282f4214SMike Kravetz * previously specified. 3319282f4214SMike Kravetz */ 3320282f4214SMike Kravetz if (!parsed_default_hugepagesz || h != &default_hstate || 3321282f4214SMike Kravetz default_hstate.max_huge_pages) { 3322282f4214SMike Kravetz pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); 332338237830SMike Kravetz return 0; 332438237830SMike Kravetz } 332538237830SMike Kravetz 3326282f4214SMike Kravetz /* 3327282f4214SMike Kravetz * No need to call hugetlb_add_hstate() as hstate already 3328282f4214SMike Kravetz * exists. But, do set parsed_hstate so that a following 3329282f4214SMike Kravetz * hugepages= parameter will be applied to this hstate. 3330282f4214SMike Kravetz */ 3331282f4214SMike Kravetz parsed_hstate = h; 3332282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3333e11bfbfcSNick Piggin return 1; 3334e11bfbfcSNick Piggin } 3335282f4214SMike Kravetz 3336359f2544SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 3337282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3338359f2544SMike Kravetz return 1; 3339359f2544SMike Kravetz } 3340359f2544SMike Kravetz __setup("hugepagesz=", hugepagesz_setup); 3341359f2544SMike Kravetz 3342282f4214SMike Kravetz /* 3343282f4214SMike Kravetz * default_hugepagesz command line input 3344282f4214SMike Kravetz * Only one instance of default_hugepagesz allowed on command line. 3345282f4214SMike Kravetz */ 3346ae94da89SMike Kravetz static int __init default_hugepagesz_setup(char *s) 3347e11bfbfcSNick Piggin { 3348ae94da89SMike Kravetz unsigned long size; 3349ae94da89SMike Kravetz 3350282f4214SMike Kravetz parsed_valid_hugepagesz = false; 3351282f4214SMike Kravetz if (parsed_default_hugepagesz) { 3352282f4214SMike Kravetz pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); 3353282f4214SMike Kravetz return 0; 3354282f4214SMike Kravetz } 3355282f4214SMike Kravetz 3356282f4214SMike Kravetz size = (unsigned long)memparse(s, NULL); 3357282f4214SMike Kravetz 3358282f4214SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 3359282f4214SMike Kravetz pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); 3360282f4214SMike Kravetz return 0; 3361282f4214SMike Kravetz } 3362282f4214SMike Kravetz 3363282f4214SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 3364282f4214SMike Kravetz parsed_valid_hugepagesz = true; 3365282f4214SMike Kravetz parsed_default_hugepagesz = true; 3366282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(size)); 3367282f4214SMike Kravetz 3368282f4214SMike Kravetz /* 3369282f4214SMike Kravetz * The number of default huge pages (for this size) could have been 3370282f4214SMike Kravetz * specified as the first hugetlb parameter: hugepages=X. If so, 3371282f4214SMike Kravetz * then default_hstate_max_huge_pages is set. If the default huge 3372282f4214SMike Kravetz * page size is gigantic (>= MAX_ORDER), then the pages must be 3373282f4214SMike Kravetz * allocated here from bootmem allocator. 3374282f4214SMike Kravetz */ 3375282f4214SMike Kravetz if (default_hstate_max_huge_pages) { 3376282f4214SMike Kravetz default_hstate.max_huge_pages = default_hstate_max_huge_pages; 3377282f4214SMike Kravetz if (hstate_is_gigantic(&default_hstate)) 3378282f4214SMike Kravetz hugetlb_hstate_alloc_pages(&default_hstate); 3379282f4214SMike Kravetz default_hstate_max_huge_pages = 0; 3380282f4214SMike Kravetz } 3381282f4214SMike Kravetz 3382e11bfbfcSNick Piggin return 1; 3383e11bfbfcSNick Piggin } 3384ae94da89SMike Kravetz __setup("default_hugepagesz=", default_hugepagesz_setup); 3385a3437870SNishanth Aravamudan 33868ca39e68SMuchun Song static unsigned int allowed_mems_nr(struct hstate *h) 33878a213460SNishanth Aravamudan { 33888a213460SNishanth Aravamudan int node; 33898a213460SNishanth Aravamudan unsigned int nr = 0; 33908ca39e68SMuchun Song nodemask_t *mpol_allowed; 33918ca39e68SMuchun Song unsigned int *array = h->free_huge_pages_node; 33928ca39e68SMuchun Song gfp_t gfp_mask = htlb_alloc_mask(h); 33938a213460SNishanth Aravamudan 33948ca39e68SMuchun Song mpol_allowed = policy_nodemask_current(gfp_mask); 33958ca39e68SMuchun Song 33968ca39e68SMuchun Song for_each_node_mask(node, cpuset_current_mems_allowed) { 3397c93b0a99SJiapeng Zhong if (!mpol_allowed || node_isset(node, *mpol_allowed)) 33988a213460SNishanth Aravamudan nr += array[node]; 33998ca39e68SMuchun Song } 34008a213460SNishanth Aravamudan 34018a213460SNishanth Aravamudan return nr; 34028a213460SNishanth Aravamudan } 34038a213460SNishanth Aravamudan 34048a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 340517743798SMuchun Song static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, 340617743798SMuchun Song void *buffer, size_t *length, 340717743798SMuchun Song loff_t *ppos, unsigned long *out) 340817743798SMuchun Song { 340917743798SMuchun Song struct ctl_table dup_table; 341017743798SMuchun Song 341117743798SMuchun Song /* 341217743798SMuchun Song * In order to avoid races with __do_proc_doulongvec_minmax(), we 341317743798SMuchun Song * can duplicate the @table and alter the duplicate of it. 341417743798SMuchun Song */ 341517743798SMuchun Song dup_table = *table; 341617743798SMuchun Song dup_table.data = out; 341717743798SMuchun Song 341817743798SMuchun Song return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); 341917743798SMuchun Song } 342017743798SMuchun Song 342106808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 342206808b08SLee Schermerhorn struct ctl_table *table, int write, 342332927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 34241da177e4SLinus Torvalds { 3425e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 3426238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 342708d4a246SMichal Hocko int ret; 3428e5ff2159SAndi Kleen 3429457c1b27SNishanth Aravamudan if (!hugepages_supported()) 343086613628SJan Stancek return -EOPNOTSUPP; 3431457c1b27SNishanth Aravamudan 343217743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 343317743798SMuchun Song &tmp); 343408d4a246SMichal Hocko if (ret) 343508d4a246SMichal Hocko goto out; 3436e5ff2159SAndi Kleen 3437238d3c13SDavid Rientjes if (write) 3438238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 3439238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 344008d4a246SMichal Hocko out: 344108d4a246SMichal Hocko return ret; 34421da177e4SLinus Torvalds } 3443396faf03SMel Gorman 344406808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 344532927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 344606808b08SLee Schermerhorn { 344706808b08SLee Schermerhorn 344806808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 344906808b08SLee Schermerhorn buffer, length, ppos); 345006808b08SLee Schermerhorn } 345106808b08SLee Schermerhorn 345206808b08SLee Schermerhorn #ifdef CONFIG_NUMA 345306808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 345432927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 345506808b08SLee Schermerhorn { 345606808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 345706808b08SLee Schermerhorn buffer, length, ppos); 345806808b08SLee Schermerhorn } 345906808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 346006808b08SLee Schermerhorn 3461a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 346232927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 3463a3d0c6aaSNishanth Aravamudan { 3464a5516438SAndi Kleen struct hstate *h = &default_hstate; 3465e5ff2159SAndi Kleen unsigned long tmp; 346608d4a246SMichal Hocko int ret; 3467e5ff2159SAndi Kleen 3468457c1b27SNishanth Aravamudan if (!hugepages_supported()) 346986613628SJan Stancek return -EOPNOTSUPP; 3470457c1b27SNishanth Aravamudan 3471e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 3472e5ff2159SAndi Kleen 3473bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 3474adbe8726SEric B Munson return -EINVAL; 3475adbe8726SEric B Munson 347617743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 347717743798SMuchun Song &tmp); 347808d4a246SMichal Hocko if (ret) 347908d4a246SMichal Hocko goto out; 3480e5ff2159SAndi Kleen 3481e5ff2159SAndi Kleen if (write) { 3482*db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3483e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 3484*db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3485e5ff2159SAndi Kleen } 348608d4a246SMichal Hocko out: 348708d4a246SMichal Hocko return ret; 3488a3d0c6aaSNishanth Aravamudan } 3489a3d0c6aaSNishanth Aravamudan 34901da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 34911da177e4SLinus Torvalds 3492e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 34931da177e4SLinus Torvalds { 3494fcb2b0c5SRoman Gushchin struct hstate *h; 3495fcb2b0c5SRoman Gushchin unsigned long total = 0; 3496fcb2b0c5SRoman Gushchin 3497457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3498457c1b27SNishanth Aravamudan return; 3499fcb2b0c5SRoman Gushchin 3500fcb2b0c5SRoman Gushchin for_each_hstate(h) { 3501fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 3502fcb2b0c5SRoman Gushchin 3503aca78307SMiaohe Lin total += huge_page_size(h) * count; 3504fcb2b0c5SRoman Gushchin 3505fcb2b0c5SRoman Gushchin if (h == &default_hstate) 3506e1759c21SAlexey Dobriyan seq_printf(m, 35071da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 35081da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 3509b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 35107893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 35114f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 3512fcb2b0c5SRoman Gushchin count, 3513a5516438SAndi Kleen h->free_huge_pages, 3514a5516438SAndi Kleen h->resv_huge_pages, 3515a5516438SAndi Kleen h->surplus_huge_pages, 3516aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 3517fcb2b0c5SRoman Gushchin } 3518fcb2b0c5SRoman Gushchin 3519aca78307SMiaohe Lin seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); 35201da177e4SLinus Torvalds } 35211da177e4SLinus Torvalds 35227981593bSJoe Perches int hugetlb_report_node_meminfo(char *buf, int len, int nid) 35231da177e4SLinus Torvalds { 3524a5516438SAndi Kleen struct hstate *h = &default_hstate; 35257981593bSJoe Perches 3526457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3527457c1b27SNishanth Aravamudan return 0; 35287981593bSJoe Perches 35297981593bSJoe Perches return sysfs_emit_at(buf, len, 35301da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 3531a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 3532a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 3533a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 3534a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 3535a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 35361da177e4SLinus Torvalds } 35371da177e4SLinus Torvalds 3538949f7ec5SDavid Rientjes void hugetlb_show_meminfo(void) 3539949f7ec5SDavid Rientjes { 3540949f7ec5SDavid Rientjes struct hstate *h; 3541949f7ec5SDavid Rientjes int nid; 3542949f7ec5SDavid Rientjes 3543457c1b27SNishanth Aravamudan if (!hugepages_supported()) 3544457c1b27SNishanth Aravamudan return; 3545457c1b27SNishanth Aravamudan 3546949f7ec5SDavid Rientjes for_each_node_state(nid, N_MEMORY) 3547949f7ec5SDavid Rientjes for_each_hstate(h) 3548949f7ec5SDavid Rientjes pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 3549949f7ec5SDavid Rientjes nid, 3550949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 3551949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 3552949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 3553aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 3554949f7ec5SDavid Rientjes } 3555949f7ec5SDavid Rientjes 35565d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 35575d317b2bSNaoya Horiguchi { 35585d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 35595d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 35605d317b2bSNaoya Horiguchi } 35615d317b2bSNaoya Horiguchi 35621da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 35631da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 35641da177e4SLinus Torvalds { 3565d0028588SWanpeng Li struct hstate *h; 3566d0028588SWanpeng Li unsigned long nr_total_pages = 0; 3567d0028588SWanpeng Li 3568d0028588SWanpeng Li for_each_hstate(h) 3569d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 3570d0028588SWanpeng Li return nr_total_pages; 35711da177e4SLinus Torvalds } 35721da177e4SLinus Torvalds 3573a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 3574fc1b8a73SMel Gorman { 3575fc1b8a73SMel Gorman int ret = -ENOMEM; 3576fc1b8a73SMel Gorman 35770aa7f354SMiaohe Lin if (!delta) 35780aa7f354SMiaohe Lin return 0; 35790aa7f354SMiaohe Lin 3580*db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3581fc1b8a73SMel Gorman /* 3582fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 3583fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 3584fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 3585fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 3586fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 3587fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 3588fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 3589fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 3590fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 3591fc1b8a73SMel Gorman * 3592fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 3593fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 3594fc1b8a73SMel Gorman * we fall back to check against current free page availability as 3595fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 3596fc1b8a73SMel Gorman * semantics that cpuset has. 35978ca39e68SMuchun Song * 35988ca39e68SMuchun Song * Apart from cpuset, we also have memory policy mechanism that 35998ca39e68SMuchun Song * also determines from which node the kernel will allocate memory 36008ca39e68SMuchun Song * in a NUMA system. So similar to cpuset, we also should consider 36018ca39e68SMuchun Song * the memory policy of the current task. Similar to the description 36028ca39e68SMuchun Song * above. 3603fc1b8a73SMel Gorman */ 3604fc1b8a73SMel Gorman if (delta > 0) { 3605a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 3606fc1b8a73SMel Gorman goto out; 3607fc1b8a73SMel Gorman 36088ca39e68SMuchun Song if (delta > allowed_mems_nr(h)) { 3609a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 3610fc1b8a73SMel Gorman goto out; 3611fc1b8a73SMel Gorman } 3612fc1b8a73SMel Gorman } 3613fc1b8a73SMel Gorman 3614fc1b8a73SMel Gorman ret = 0; 3615fc1b8a73SMel Gorman if (delta < 0) 3616a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 3617fc1b8a73SMel Gorman 3618fc1b8a73SMel Gorman out: 3619*db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3620fc1b8a73SMel Gorman return ret; 3621fc1b8a73SMel Gorman } 3622fc1b8a73SMel Gorman 362384afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 362484afd99bSAndy Whitcroft { 3625f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 362684afd99bSAndy Whitcroft 362784afd99bSAndy Whitcroft /* 362884afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 362984afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 363084afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 363125985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 363284afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 363384afd99bSAndy Whitcroft * new reference here without additional locking. 363484afd99bSAndy Whitcroft */ 36354e35f483SJoonsoo Kim if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 3636f522c3acSJoonsoo Kim kref_get(&resv->refs); 363784afd99bSAndy Whitcroft } 363884afd99bSAndy Whitcroft 3639a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 3640a1e78772SMel Gorman { 3641a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3642f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 364390481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 36444e35f483SJoonsoo Kim unsigned long reserve, start, end; 36451c5ecae3SMike Kravetz long gbl_reserve; 364684afd99bSAndy Whitcroft 36474e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 36484e35f483SJoonsoo Kim return; 36494e35f483SJoonsoo Kim 3650a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 3651a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 365284afd99bSAndy Whitcroft 36534e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 3654e9fe92aeSMina Almasry hugetlb_cgroup_uncharge_counter(resv, start, end); 36557251ff78SAdam Litke if (reserve) { 36561c5ecae3SMike Kravetz /* 36571c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 36581c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 36591c5ecae3SMike Kravetz */ 36601c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 36611c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 36627251ff78SAdam Litke } 3663e9fe92aeSMina Almasry 3664e9fe92aeSMina Almasry kref_put(&resv->refs, resv_map_release); 3665a1e78772SMel Gorman } 3666a1e78772SMel Gorman 366731383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 366831383c68SDan Williams { 366931383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 367031383c68SDan Williams return -EINVAL; 367131383c68SDan Williams return 0; 367231383c68SDan Williams } 367331383c68SDan Williams 367405ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 367505ea8860SDan Williams { 3676aca78307SMiaohe Lin return huge_page_size(hstate_vma(vma)); 367705ea8860SDan Williams } 367805ea8860SDan Williams 36791da177e4SLinus Torvalds /* 36801da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 36811da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 36826c26d310SMiaohe Lin * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get 36831da177e4SLinus Torvalds * this far. 36841da177e4SLinus Torvalds */ 3685b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 36861da177e4SLinus Torvalds { 36871da177e4SLinus Torvalds BUG(); 3688d0217ac0SNick Piggin return 0; 36891da177e4SLinus Torvalds } 36901da177e4SLinus Torvalds 3691eec3636aSJane Chu /* 3692eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 3693eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 3694eec3636aSJane Chu * This is because under System V memory model, mappings created via 3695eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 3696eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 3697eec3636aSJane Chu */ 3698f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 3699d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 370084afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 3701a1e78772SMel Gorman .close = hugetlb_vm_op_close, 3702dd3b614fSDmitry Safonov .may_split = hugetlb_vm_op_split, 370305ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 37041da177e4SLinus Torvalds }; 37051da177e4SLinus Torvalds 37061e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 37071e8f889bSDavid Gibson int writable) 370863551ae0SDavid Gibson { 370963551ae0SDavid Gibson pte_t entry; 371063551ae0SDavid Gibson 37111e8f889bSDavid Gibson if (writable) { 3712106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 3713106c992aSGerald Schaefer vma->vm_page_prot))); 371463551ae0SDavid Gibson } else { 3715106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 3716106c992aSGerald Schaefer vma->vm_page_prot)); 371763551ae0SDavid Gibson } 371863551ae0SDavid Gibson entry = pte_mkyoung(entry); 371963551ae0SDavid Gibson entry = pte_mkhuge(entry); 3720d9ed9faaSChris Metcalf entry = arch_make_huge_pte(entry, vma, page, writable); 372163551ae0SDavid Gibson 372263551ae0SDavid Gibson return entry; 372363551ae0SDavid Gibson } 372463551ae0SDavid Gibson 37251e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 37261e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 37271e8f889bSDavid Gibson { 37281e8f889bSDavid Gibson pte_t entry; 37291e8f889bSDavid Gibson 3730106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 373132f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 37324b3073e1SRussell King update_mmu_cache(vma, address, ptep); 37331e8f889bSDavid Gibson } 37341e8f889bSDavid Gibson 3735d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 37364a705fefSNaoya Horiguchi { 37374a705fefSNaoya Horiguchi swp_entry_t swp; 37384a705fefSNaoya Horiguchi 37394a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 3740d5ed7444SAneesh Kumar K.V return false; 37414a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 3742d79d176aSBaoquan He if (is_migration_entry(swp)) 3743d5ed7444SAneesh Kumar K.V return true; 37444a705fefSNaoya Horiguchi else 3745d5ed7444SAneesh Kumar K.V return false; 37464a705fefSNaoya Horiguchi } 37474a705fefSNaoya Horiguchi 37483e5c3600SBaoquan He static bool is_hugetlb_entry_hwpoisoned(pte_t pte) 37494a705fefSNaoya Horiguchi { 37504a705fefSNaoya Horiguchi swp_entry_t swp; 37514a705fefSNaoya Horiguchi 37524a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 37533e5c3600SBaoquan He return false; 37544a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 3755d79d176aSBaoquan He if (is_hwpoison_entry(swp)) 37563e5c3600SBaoquan He return true; 37574a705fefSNaoya Horiguchi else 37583e5c3600SBaoquan He return false; 37594a705fefSNaoya Horiguchi } 37601e8f889bSDavid Gibson 37614eae4efaSPeter Xu static void 37624eae4efaSPeter Xu hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, 37634eae4efaSPeter Xu struct page *new_page) 37644eae4efaSPeter Xu { 37654eae4efaSPeter Xu __SetPageUptodate(new_page); 37664eae4efaSPeter Xu set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1)); 37674eae4efaSPeter Xu hugepage_add_new_anon_rmap(new_page, vma, addr); 37684eae4efaSPeter Xu hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); 37694eae4efaSPeter Xu ClearHPageRestoreReserve(new_page); 37704eae4efaSPeter Xu SetHPageMigratable(new_page); 37714eae4efaSPeter Xu } 37724eae4efaSPeter Xu 377363551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 377463551ae0SDavid Gibson struct vm_area_struct *vma) 377563551ae0SDavid Gibson { 37765e41540cSMike Kravetz pte_t *src_pte, *dst_pte, entry, dst_entry; 377763551ae0SDavid Gibson struct page *ptepage; 37781c59827dSHugh Dickins unsigned long addr; 3779ca6eb14dSPeter Xu bool cow = is_cow_mapping(vma->vm_flags); 3780a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3781a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 37824eae4efaSPeter Xu unsigned long npages = pages_per_huge_page(h); 3783c0d0381aSMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 3784ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 3785e8569dd2SAndreas Sandberg int ret = 0; 37861e8f889bSDavid Gibson 3787ac46d4f3SJérôme Glisse if (cow) { 37887269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src, 37896f4f13e8SJérôme Glisse vma->vm_start, 3790ac46d4f3SJérôme Glisse vma->vm_end); 3791ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 3792c0d0381aSMike Kravetz } else { 3793c0d0381aSMike Kravetz /* 3794c0d0381aSMike Kravetz * For shared mappings i_mmap_rwsem must be held to call 3795c0d0381aSMike Kravetz * huge_pte_alloc, otherwise the returned ptep could go 3796c0d0381aSMike Kravetz * away if part of a shared pmd and another thread calls 3797c0d0381aSMike Kravetz * huge_pmd_unshare. 3798c0d0381aSMike Kravetz */ 3799c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 3800ac46d4f3SJérôme Glisse } 3801e8569dd2SAndreas Sandberg 3802a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 3803cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 38047868a208SPunit Agrawal src_pte = huge_pte_offset(src, addr, sz); 3805c74df32cSHugh Dickins if (!src_pte) 3806c74df32cSHugh Dickins continue; 3807aec44e0fSPeter Xu dst_pte = huge_pte_alloc(dst, vma, addr, sz); 3808e8569dd2SAndreas Sandberg if (!dst_pte) { 3809e8569dd2SAndreas Sandberg ret = -ENOMEM; 3810e8569dd2SAndreas Sandberg break; 3811e8569dd2SAndreas Sandberg } 3812c5c99429SLarry Woodman 38135e41540cSMike Kravetz /* 38145e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 38155e41540cSMike Kravetz * dst_pte == src_pte is the common case of src/dest sharing. 38165e41540cSMike Kravetz * 38175e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 38185e41540cSMike Kravetz * another vma. If dst_pte !none, this implies sharing. 38195e41540cSMike Kravetz * Check here before taking page table lock, and once again 38205e41540cSMike Kravetz * after taking the lock below. 38215e41540cSMike Kravetz */ 38225e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 38235e41540cSMike Kravetz if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) 3824c5c99429SLarry Woodman continue; 3825c5c99429SLarry Woodman 3826cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 3827cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 3828cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 38294a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 38305e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 38314eae4efaSPeter Xu again: 38325e41540cSMike Kravetz if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { 38335e41540cSMike Kravetz /* 38345e41540cSMike Kravetz * Skip if src entry none. Also, skip in the 38355e41540cSMike Kravetz * unlikely case dst entry !none as this implies 38365e41540cSMike Kravetz * sharing with another vma. 38375e41540cSMike Kravetz */ 38384a705fefSNaoya Horiguchi ; 38394a705fefSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry) || 38404a705fefSNaoya Horiguchi is_hugetlb_entry_hwpoisoned(entry))) { 38414a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 38424a705fefSNaoya Horiguchi 38434a705fefSNaoya Horiguchi if (is_write_migration_entry(swp_entry) && cow) { 38444a705fefSNaoya Horiguchi /* 38454a705fefSNaoya Horiguchi * COW mappings require pages in both 38464a705fefSNaoya Horiguchi * parent and child to be set to read. 38474a705fefSNaoya Horiguchi */ 38484a705fefSNaoya Horiguchi make_migration_entry_read(&swp_entry); 38494a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 3850e5251fd4SPunit Agrawal set_huge_swap_pte_at(src, addr, src_pte, 3851e5251fd4SPunit Agrawal entry, sz); 38524a705fefSNaoya Horiguchi } 3853e5251fd4SPunit Agrawal set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); 38544a705fefSNaoya Horiguchi } else { 38554eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 38564eae4efaSPeter Xu ptepage = pte_page(entry); 38574eae4efaSPeter Xu get_page(ptepage); 38584eae4efaSPeter Xu 38594eae4efaSPeter Xu /* 38604eae4efaSPeter Xu * This is a rare case where we see pinned hugetlb 38614eae4efaSPeter Xu * pages while they're prone to COW. We need to do the 38624eae4efaSPeter Xu * COW earlier during fork. 38634eae4efaSPeter Xu * 38644eae4efaSPeter Xu * When pre-allocating the page or copying data, we 38654eae4efaSPeter Xu * need to be without the pgtable locks since we could 38664eae4efaSPeter Xu * sleep during the process. 38674eae4efaSPeter Xu */ 38684eae4efaSPeter Xu if (unlikely(page_needs_cow_for_dma(vma, ptepage))) { 38694eae4efaSPeter Xu pte_t src_pte_old = entry; 38704eae4efaSPeter Xu struct page *new; 38714eae4efaSPeter Xu 38724eae4efaSPeter Xu spin_unlock(src_ptl); 38734eae4efaSPeter Xu spin_unlock(dst_ptl); 38744eae4efaSPeter Xu /* Do not use reserve as it's private owned */ 38754eae4efaSPeter Xu new = alloc_huge_page(vma, addr, 1); 38764eae4efaSPeter Xu if (IS_ERR(new)) { 38774eae4efaSPeter Xu put_page(ptepage); 38784eae4efaSPeter Xu ret = PTR_ERR(new); 38794eae4efaSPeter Xu break; 38804eae4efaSPeter Xu } 38814eae4efaSPeter Xu copy_user_huge_page(new, ptepage, addr, vma, 38824eae4efaSPeter Xu npages); 38834eae4efaSPeter Xu put_page(ptepage); 38844eae4efaSPeter Xu 38854eae4efaSPeter Xu /* Install the new huge page if src pte stable */ 38864eae4efaSPeter Xu dst_ptl = huge_pte_lock(h, dst, dst_pte); 38874eae4efaSPeter Xu src_ptl = huge_pte_lockptr(h, src, src_pte); 38884eae4efaSPeter Xu spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 38894eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 38904eae4efaSPeter Xu if (!pte_same(src_pte_old, entry)) { 38914eae4efaSPeter Xu put_page(new); 38924eae4efaSPeter Xu /* dst_entry won't change as in child */ 38934eae4efaSPeter Xu goto again; 38944eae4efaSPeter Xu } 38954eae4efaSPeter Xu hugetlb_install_page(vma, dst_pte, addr, new); 38964eae4efaSPeter Xu spin_unlock(src_ptl); 38974eae4efaSPeter Xu spin_unlock(dst_ptl); 38984eae4efaSPeter Xu continue; 38994eae4efaSPeter Xu } 39004eae4efaSPeter Xu 390134ee645eSJoerg Roedel if (cow) { 39020f10851eSJérôme Glisse /* 39030f10851eSJérôme Glisse * No need to notify as we are downgrading page 39040f10851eSJérôme Glisse * table protection not changing it to point 39050f10851eSJérôme Glisse * to a new page. 39060f10851eSJérôme Glisse * 3907ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 39080f10851eSJérôme Glisse */ 39097f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 391034ee645eSJoerg Roedel } 39114eae4efaSPeter Xu 391253f9263bSKirill A. Shutemov page_dup_rmap(ptepage, true); 391363551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 39144eae4efaSPeter Xu hugetlb_count_add(npages, dst); 39151c59827dSHugh Dickins } 3916cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 3917cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 391863551ae0SDavid Gibson } 391963551ae0SDavid Gibson 3920e8569dd2SAndreas Sandberg if (cow) 3921ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 3922c0d0381aSMike Kravetz else 3923c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 3924e8569dd2SAndreas Sandberg 3925e8569dd2SAndreas Sandberg return ret; 392663551ae0SDavid Gibson } 392763551ae0SDavid Gibson 392824669e58SAneesh Kumar K.V void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 392924669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 393024669e58SAneesh Kumar K.V struct page *ref_page) 393163551ae0SDavid Gibson { 393263551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 393363551ae0SDavid Gibson unsigned long address; 3934c7546f8fSDavid Gibson pte_t *ptep; 393563551ae0SDavid Gibson pte_t pte; 3936cb900f41SKirill A. Shutemov spinlock_t *ptl; 393763551ae0SDavid Gibson struct page *page; 3938a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 3939a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 3940ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 3941a5516438SAndi Kleen 394263551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 3943a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 3944a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 394563551ae0SDavid Gibson 394607e32661SAneesh Kumar K.V /* 394707e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 394807e32661SAneesh Kumar K.V * to huge page. 394907e32661SAneesh Kumar K.V */ 3950ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 395124669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 3952dff11abeSMike Kravetz 3953dff11abeSMike Kravetz /* 3954dff11abeSMike Kravetz * If sharing possible, alert mmu notifiers of worst case. 3955dff11abeSMike Kravetz */ 39566f4f13e8SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, 39576f4f13e8SJérôme Glisse end); 3958ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 3959ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 3960569f48b8SHillf Danton address = start; 3961569f48b8SHillf Danton for (; address < end; address += sz) { 39627868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, sz); 3963c7546f8fSDavid Gibson if (!ptep) 3964c7546f8fSDavid Gibson continue; 3965c7546f8fSDavid Gibson 3966cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 396734ae204fSMike Kravetz if (huge_pmd_unshare(mm, vma, &address, ptep)) { 396831d49da5SAneesh Kumar K.V spin_unlock(ptl); 3969dff11abeSMike Kravetz /* 3970dff11abeSMike Kravetz * We just unmapped a page of PMDs by clearing a PUD. 3971dff11abeSMike Kravetz * The caller's TLB flush range should cover this area. 3972dff11abeSMike Kravetz */ 397331d49da5SAneesh Kumar K.V continue; 397431d49da5SAneesh Kumar K.V } 397539dde65cSChen, Kenneth W 39766629326bSHillf Danton pte = huge_ptep_get(ptep); 397731d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 397831d49da5SAneesh Kumar K.V spin_unlock(ptl); 397931d49da5SAneesh Kumar K.V continue; 398031d49da5SAneesh Kumar K.V } 39816629326bSHillf Danton 39826629326bSHillf Danton /* 39839fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 39849fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 39856629326bSHillf Danton */ 39869fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 39879386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 398831d49da5SAneesh Kumar K.V spin_unlock(ptl); 398931d49da5SAneesh Kumar K.V continue; 39908c4894c6SNaoya Horiguchi } 39916629326bSHillf Danton 39926629326bSHillf Danton page = pte_page(pte); 399304f2cbe3SMel Gorman /* 399404f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 399504f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 399604f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 399704f2cbe3SMel Gorman */ 399804f2cbe3SMel Gorman if (ref_page) { 399931d49da5SAneesh Kumar K.V if (page != ref_page) { 400031d49da5SAneesh Kumar K.V spin_unlock(ptl); 400131d49da5SAneesh Kumar K.V continue; 400231d49da5SAneesh Kumar K.V } 400304f2cbe3SMel Gorman /* 400404f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 400504f2cbe3SMel Gorman * future faults in this VMA will fail rather than 400604f2cbe3SMel Gorman * looking like data was lost 400704f2cbe3SMel Gorman */ 400804f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 400904f2cbe3SMel Gorman } 401004f2cbe3SMel Gorman 4011c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 4012b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 4013106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 40146649a386SKen Chen set_page_dirty(page); 40159e81130bSHillf Danton 40165d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 4017d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 401831d49da5SAneesh Kumar K.V 4019cb900f41SKirill A. Shutemov spin_unlock(ptl); 4020e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 402124669e58SAneesh Kumar K.V /* 402231d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 402324669e58SAneesh Kumar K.V */ 402431d49da5SAneesh Kumar K.V if (ref_page) 402531d49da5SAneesh Kumar K.V break; 4026fe1668aeSChen, Kenneth W } 4027ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 402824669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 40291da177e4SLinus Torvalds } 403063551ae0SDavid Gibson 4031d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 4032d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 4033d833352aSMel Gorman unsigned long end, struct page *ref_page) 4034d833352aSMel Gorman { 4035d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 4036d833352aSMel Gorman 4037d833352aSMel Gorman /* 4038d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 4039d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 4040d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 4041d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 4042d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 4043c8c06efaSDavidlohr Bueso * is to clear it before releasing the i_mmap_rwsem. This works 4044d833352aSMel Gorman * because in the context this is called, the VMA is about to be 4045c8c06efaSDavidlohr Bueso * destroyed and the i_mmap_rwsem is held. 4046d833352aSMel Gorman */ 4047d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 4048d833352aSMel Gorman } 4049d833352aSMel Gorman 4050502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 405104f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 4052502717f4SChen, Kenneth W { 405324669e58SAneesh Kumar K.V struct mmu_gather tlb; 4054dff11abeSMike Kravetz 4055a72afd87SWill Deacon tlb_gather_mmu(&tlb, vma->vm_mm); 405624669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 4057ae8eba8bSWill Deacon tlb_finish_mmu(&tlb); 4058502717f4SChen, Kenneth W } 4059502717f4SChen, Kenneth W 406004f2cbe3SMel Gorman /* 406104f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 4062578b7725SZhiyuan Dai * mapping it owns the reserve page for. The intention is to unmap the page 406304f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 406404f2cbe3SMel Gorman * same region. 406504f2cbe3SMel Gorman */ 40662f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 40672a4b3dedSHarvey Harrison struct page *page, unsigned long address) 406804f2cbe3SMel Gorman { 40697526674dSAdam Litke struct hstate *h = hstate_vma(vma); 407004f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 407104f2cbe3SMel Gorman struct address_space *mapping; 407204f2cbe3SMel Gorman pgoff_t pgoff; 407304f2cbe3SMel Gorman 407404f2cbe3SMel Gorman /* 407504f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 407604f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 407704f2cbe3SMel Gorman */ 40787526674dSAdam Litke address = address & huge_page_mask(h); 407936e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 408036e4f20aSMichal Hocko vma->vm_pgoff; 408193c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 408204f2cbe3SMel Gorman 40834eb2b1dcSMel Gorman /* 40844eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 40854eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 40864eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 40874eb2b1dcSMel Gorman */ 408883cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 40896b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 409004f2cbe3SMel Gorman /* Do not unmap the current VMA */ 409104f2cbe3SMel Gorman if (iter_vma == vma) 409204f2cbe3SMel Gorman continue; 409304f2cbe3SMel Gorman 409404f2cbe3SMel Gorman /* 40952f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 40962f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 40972f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 40982f84a899SMel Gorman */ 40992f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 41002f84a899SMel Gorman continue; 41012f84a899SMel Gorman 41022f84a899SMel Gorman /* 410304f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 410404f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 410504f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 410604f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 410704f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 410804f2cbe3SMel Gorman */ 410904f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 411024669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 411124669e58SAneesh Kumar K.V address + huge_page_size(h), page); 411204f2cbe3SMel Gorman } 411383cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 411404f2cbe3SMel Gorman } 411504f2cbe3SMel Gorman 41160fe6e20bSNaoya Horiguchi /* 41170fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 4118ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 4119ef009b25SMichal Hocko * cannot race with other handlers or page migration. 4120ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 41210fe6e20bSNaoya Horiguchi */ 41222b740303SSouptick Joarder static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 4123974e6d66SHuang Ying unsigned long address, pte_t *ptep, 4124cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 41251e8f889bSDavid Gibson { 41263999f52eSAneesh Kumar K.V pte_t pte; 4127a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 41281e8f889bSDavid Gibson struct page *old_page, *new_page; 41292b740303SSouptick Joarder int outside_reserve = 0; 41302b740303SSouptick Joarder vm_fault_t ret = 0; 4131974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 4132ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 41331e8f889bSDavid Gibson 41343999f52eSAneesh Kumar K.V pte = huge_ptep_get(ptep); 41351e8f889bSDavid Gibson old_page = pte_page(pte); 41361e8f889bSDavid Gibson 413704f2cbe3SMel Gorman retry_avoidcopy: 41381e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 41391e8f889bSDavid Gibson * and just make the page writable */ 414037a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 41415a49973dSHugh Dickins page_move_anon_rmap(old_page, vma); 41425b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 414383c54070SNick Piggin return 0; 41441e8f889bSDavid Gibson } 41451e8f889bSDavid Gibson 414604f2cbe3SMel Gorman /* 414704f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 414804f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 414904f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 415004f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 415104f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 415204f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 415304f2cbe3SMel Gorman * of the full address range. 415404f2cbe3SMel Gorman */ 41555944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 415604f2cbe3SMel Gorman old_page != pagecache_page) 415704f2cbe3SMel Gorman outside_reserve = 1; 415804f2cbe3SMel Gorman 415909cbfeafSKirill A. Shutemov get_page(old_page); 4160b76c8cfbSLarry Woodman 4161ad4404a2SDavidlohr Bueso /* 4162ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 4163ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 4164ad4404a2SDavidlohr Bueso */ 4165cb900f41SKirill A. Shutemov spin_unlock(ptl); 41665b7a1d40SHuang Ying new_page = alloc_huge_page(vma, haddr, outside_reserve); 41671e8f889bSDavid Gibson 41682fc39cecSAdam Litke if (IS_ERR(new_page)) { 416904f2cbe3SMel Gorman /* 417004f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 417104f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 417204f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 417304f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 417404f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 417504f2cbe3SMel Gorman */ 417604f2cbe3SMel Gorman if (outside_reserve) { 4177e7dd91c4SMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 4178e7dd91c4SMike Kravetz pgoff_t idx; 4179e7dd91c4SMike Kravetz u32 hash; 4180e7dd91c4SMike Kravetz 418109cbfeafSKirill A. Shutemov put_page(old_page); 418204f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 4183e7dd91c4SMike Kravetz /* 4184e7dd91c4SMike Kravetz * Drop hugetlb_fault_mutex and i_mmap_rwsem before 4185e7dd91c4SMike Kravetz * unmapping. unmapping needs to hold i_mmap_rwsem 4186e7dd91c4SMike Kravetz * in write mode. Dropping i_mmap_rwsem in read mode 4187e7dd91c4SMike Kravetz * here is OK as COW mappings do not interact with 4188e7dd91c4SMike Kravetz * PMD sharing. 4189e7dd91c4SMike Kravetz * 4190e7dd91c4SMike Kravetz * Reacquire both after unmap operation. 4191e7dd91c4SMike Kravetz */ 4192e7dd91c4SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 4193e7dd91c4SMike Kravetz hash = hugetlb_fault_mutex_hash(mapping, idx); 4194e7dd91c4SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 4195e7dd91c4SMike Kravetz i_mmap_unlock_read(mapping); 4196e7dd91c4SMike Kravetz 41975b7a1d40SHuang Ying unmap_ref_private(mm, vma, old_page, haddr); 4198e7dd91c4SMike Kravetz 4199e7dd91c4SMike Kravetz i_mmap_lock_read(mapping); 4200e7dd91c4SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 4201cb900f41SKirill A. Shutemov spin_lock(ptl); 42025b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4203a9af0c5dSNaoya Horiguchi if (likely(ptep && 4204a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 420504f2cbe3SMel Gorman goto retry_avoidcopy; 4206a734bcc8SHillf Danton /* 4207cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 4208cb900f41SKirill A. Shutemov * lock, and our job is done. 4209a734bcc8SHillf Danton */ 4210a734bcc8SHillf Danton return 0; 421104f2cbe3SMel Gorman } 421204f2cbe3SMel Gorman 42132b740303SSouptick Joarder ret = vmf_error(PTR_ERR(new_page)); 4214ad4404a2SDavidlohr Bueso goto out_release_old; 42151e8f889bSDavid Gibson } 42161e8f889bSDavid Gibson 42170fe6e20bSNaoya Horiguchi /* 42180fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 42190fe6e20bSNaoya Horiguchi * anon_vma prepared. 42200fe6e20bSNaoya Horiguchi */ 422144e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 4222ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 4223ad4404a2SDavidlohr Bueso goto out_release_all; 422444e2aa93SDean Nelson } 42250fe6e20bSNaoya Horiguchi 4226974e6d66SHuang Ying copy_user_huge_page(new_page, old_page, address, vma, 422747ad8475SAndrea Arcangeli pages_per_huge_page(h)); 42280ed361deSNick Piggin __SetPageUptodate(new_page); 42291e8f889bSDavid Gibson 42307269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, 42316f4f13e8SJérôme Glisse haddr + huge_page_size(h)); 4232ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4233ad4404a2SDavidlohr Bueso 4234b76c8cfbSLarry Woodman /* 4235cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 4236b76c8cfbSLarry Woodman * before the page tables are altered 4237b76c8cfbSLarry Woodman */ 4238cb900f41SKirill A. Shutemov spin_lock(ptl); 42395b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4240a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 4241d6995da3SMike Kravetz ClearHPageRestoreReserve(new_page); 424207443a85SJoonsoo Kim 42431e8f889bSDavid Gibson /* Break COW */ 42445b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 4245ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, range.end); 42465b7a1d40SHuang Ying set_huge_pte_at(mm, haddr, ptep, 42471e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 4248d281ee61SKirill A. Shutemov page_remove_rmap(old_page, true); 42495b7a1d40SHuang Ying hugepage_add_new_anon_rmap(new_page, vma, haddr); 42508f251a3dSMike Kravetz SetHPageMigratable(new_page); 42511e8f889bSDavid Gibson /* Make the old page be freed below */ 42521e8f889bSDavid Gibson new_page = old_page; 42531e8f889bSDavid Gibson } 4254cb900f41SKirill A. Shutemov spin_unlock(ptl); 4255ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 4256ad4404a2SDavidlohr Bueso out_release_all: 42575b7a1d40SHuang Ying restore_reserve_on_error(h, vma, haddr, new_page); 425809cbfeafSKirill A. Shutemov put_page(new_page); 4259ad4404a2SDavidlohr Bueso out_release_old: 426009cbfeafSKirill A. Shutemov put_page(old_page); 42618312034fSJoonsoo Kim 4262ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 4263ad4404a2SDavidlohr Bueso return ret; 42641e8f889bSDavid Gibson } 42651e8f889bSDavid Gibson 426604f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 4267a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 4268a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 426904f2cbe3SMel Gorman { 427004f2cbe3SMel Gorman struct address_space *mapping; 4271e7c4b0bfSAndy Whitcroft pgoff_t idx; 427204f2cbe3SMel Gorman 427304f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 4274a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 427504f2cbe3SMel Gorman 427604f2cbe3SMel Gorman return find_lock_page(mapping, idx); 427704f2cbe3SMel Gorman } 427804f2cbe3SMel Gorman 42793ae77f43SHugh Dickins /* 42803ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 42813ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 42823ae77f43SHugh Dickins */ 42833ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 42842a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 42852a15efc9SHugh Dickins { 42862a15efc9SHugh Dickins struct address_space *mapping; 42872a15efc9SHugh Dickins pgoff_t idx; 42882a15efc9SHugh Dickins struct page *page; 42892a15efc9SHugh Dickins 42902a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 42912a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 42922a15efc9SHugh Dickins 42932a15efc9SHugh Dickins page = find_get_page(mapping, idx); 42942a15efc9SHugh Dickins if (page) 42952a15efc9SHugh Dickins put_page(page); 42962a15efc9SHugh Dickins return page != NULL; 42972a15efc9SHugh Dickins } 42982a15efc9SHugh Dickins 4299ab76ad54SMike Kravetz int huge_add_to_page_cache(struct page *page, struct address_space *mapping, 4300ab76ad54SMike Kravetz pgoff_t idx) 4301ab76ad54SMike Kravetz { 4302ab76ad54SMike Kravetz struct inode *inode = mapping->host; 4303ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 4304ab76ad54SMike Kravetz int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 4305ab76ad54SMike Kravetz 4306ab76ad54SMike Kravetz if (err) 4307ab76ad54SMike Kravetz return err; 4308d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 4309ab76ad54SMike Kravetz 431022146c3cSMike Kravetz /* 431122146c3cSMike Kravetz * set page dirty so that it will not be removed from cache/file 431222146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 431322146c3cSMike Kravetz */ 431422146c3cSMike Kravetz set_page_dirty(page); 431522146c3cSMike Kravetz 4316ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 4317ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 4318ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 4319ab76ad54SMike Kravetz return 0; 4320ab76ad54SMike Kravetz } 4321ab76ad54SMike Kravetz 43222b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 43232b740303SSouptick Joarder struct vm_area_struct *vma, 43248382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 4325788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 4326ac9b9c66SHugh Dickins { 4327a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 43282b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 4329409eb8c2SHillf Danton int anon_rmap = 0; 43304c887265SAdam Litke unsigned long size; 43314c887265SAdam Litke struct page *page; 43321e8f889bSDavid Gibson pte_t new_pte; 4333cb900f41SKirill A. Shutemov spinlock_t *ptl; 4334285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 4335cb6acd01SMike Kravetz bool new_page = false; 43364c887265SAdam Litke 433704f2cbe3SMel Gorman /* 433804f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 433904f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 434025985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 434104f2cbe3SMel Gorman */ 434204f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 4343910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 434404f2cbe3SMel Gorman current->pid); 434504f2cbe3SMel Gorman return ret; 434604f2cbe3SMel Gorman } 434704f2cbe3SMel Gorman 43484c887265SAdam Litke /* 434987bf91d3SMike Kravetz * We can not race with truncation due to holding i_mmap_rwsem. 435087bf91d3SMike Kravetz * i_size is modified when holding i_mmap_rwsem, so check here 435187bf91d3SMike Kravetz * once for faults beyond end of file. 43524c887265SAdam Litke */ 4353a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 4354ebed4bfcSHugh Dickins if (idx >= size) 4355ebed4bfcSHugh Dickins goto out; 43561a1aad8aSMike Kravetz 435787bf91d3SMike Kravetz retry: 435887bf91d3SMike Kravetz page = find_lock_page(mapping, idx); 435987bf91d3SMike Kravetz if (!page) { 43601a1aad8aSMike Kravetz /* 43611a1aad8aSMike Kravetz * Check for page in userfault range 43621a1aad8aSMike Kravetz */ 43631a1aad8aSMike Kravetz if (userfaultfd_missing(vma)) { 43641a1aad8aSMike Kravetz u32 hash; 43651a1aad8aSMike Kravetz struct vm_fault vmf = { 43661a1aad8aSMike Kravetz .vma = vma, 4367285b8dcaSHuang Ying .address = haddr, 43681a1aad8aSMike Kravetz .flags = flags, 43691a1aad8aSMike Kravetz /* 43701a1aad8aSMike Kravetz * Hard to debug if it ends up being 43711a1aad8aSMike Kravetz * used by a callee that assumes 43721a1aad8aSMike Kravetz * something about the other 43731a1aad8aSMike Kravetz * uninitialized fields... same as in 43741a1aad8aSMike Kravetz * memory.c 43751a1aad8aSMike Kravetz */ 43761a1aad8aSMike Kravetz }; 43771a1aad8aSMike Kravetz 43781a1aad8aSMike Kravetz /* 4379c0d0381aSMike Kravetz * hugetlb_fault_mutex and i_mmap_rwsem must be 4380c0d0381aSMike Kravetz * dropped before handling userfault. Reacquire 4381c0d0381aSMike Kravetz * after handling fault to make calling code simpler. 43821a1aad8aSMike Kravetz */ 4383188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 43841a1aad8aSMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 4385c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 43861a1aad8aSMike Kravetz ret = handle_userfault(&vmf, VM_UFFD_MISSING); 4387c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 43881a1aad8aSMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 43891a1aad8aSMike Kravetz goto out; 43901a1aad8aSMike Kravetz } 43911a1aad8aSMike Kravetz 4392285b8dcaSHuang Ying page = alloc_huge_page(vma, haddr, 0); 43932fc39cecSAdam Litke if (IS_ERR(page)) { 43944643d67eSMike Kravetz /* 43954643d67eSMike Kravetz * Returning error will result in faulting task being 43964643d67eSMike Kravetz * sent SIGBUS. The hugetlb fault mutex prevents two 43974643d67eSMike Kravetz * tasks from racing to fault in the same page which 43984643d67eSMike Kravetz * could result in false unable to allocate errors. 43994643d67eSMike Kravetz * Page migration does not take the fault mutex, but 44004643d67eSMike Kravetz * does a clear then write of pte's under page table 44014643d67eSMike Kravetz * lock. Page fault code could race with migration, 44024643d67eSMike Kravetz * notice the clear pte and try to allocate a page 44034643d67eSMike Kravetz * here. Before returning error, get ptl and make 44044643d67eSMike Kravetz * sure there really is no pte entry. 44054643d67eSMike Kravetz */ 44064643d67eSMike Kravetz ptl = huge_pte_lock(h, mm, ptep); 44074643d67eSMike Kravetz ret = 0; 4408d83e6c8aSMiaohe Lin if (huge_pte_none(huge_ptep_get(ptep))) 44092b740303SSouptick Joarder ret = vmf_error(PTR_ERR(page)); 4410d83e6c8aSMiaohe Lin spin_unlock(ptl); 44116bda666aSChristoph Lameter goto out; 44126bda666aSChristoph Lameter } 441347ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 44140ed361deSNick Piggin __SetPageUptodate(page); 4415cb6acd01SMike Kravetz new_page = true; 4416ac9b9c66SHugh Dickins 4417f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 4418ab76ad54SMike Kravetz int err = huge_add_to_page_cache(page, mapping, idx); 44196bda666aSChristoph Lameter if (err) { 44206bda666aSChristoph Lameter put_page(page); 44216bda666aSChristoph Lameter if (err == -EEXIST) 44226bda666aSChristoph Lameter goto retry; 44236bda666aSChristoph Lameter goto out; 44246bda666aSChristoph Lameter } 442523be7468SMel Gorman } else { 44266bda666aSChristoph Lameter lock_page(page); 44270fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 44280fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 44290fe6e20bSNaoya Horiguchi goto backout_unlocked; 443023be7468SMel Gorman } 4431409eb8c2SHillf Danton anon_rmap = 1; 44320fe6e20bSNaoya Horiguchi } 44330fe6e20bSNaoya Horiguchi } else { 443457303d80SAndy Whitcroft /* 4435998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 4436998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 4437998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 4438fd6a03edSNaoya Horiguchi */ 4439fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 44400eb98f15SMiaohe Lin ret = VM_FAULT_HWPOISON_LARGE | 4441972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 4442fd6a03edSNaoya Horiguchi goto backout_unlocked; 44436bda666aSChristoph Lameter } 4444998b4382SNaoya Horiguchi } 44451e8f889bSDavid Gibson 444657303d80SAndy Whitcroft /* 444757303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 444857303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 444957303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 445057303d80SAndy Whitcroft * the spinlock. 445157303d80SAndy Whitcroft */ 44525e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 4453285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 44542b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 44552b26736cSAndy Whitcroft goto backout_unlocked; 44562b26736cSAndy Whitcroft } 44575e911373SMike Kravetz /* Just decrements count, does not deallocate */ 4458285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 44595e911373SMike Kravetz } 446057303d80SAndy Whitcroft 44618bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 446283c54070SNick Piggin ret = 0; 44637f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 44644c887265SAdam Litke goto backout; 44654c887265SAdam Litke 446607443a85SJoonsoo Kim if (anon_rmap) { 4467d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 4468285b8dcaSHuang Ying hugepage_add_new_anon_rmap(page, vma, haddr); 4469ac714904SChoi Gi-yong } else 447053f9263bSKirill A. Shutemov page_dup_rmap(page, true); 44711e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 44721e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 4473285b8dcaSHuang Ying set_huge_pte_at(mm, haddr, ptep, new_pte); 44741e8f889bSDavid Gibson 44755d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 4476788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 44771e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 4478974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); 44791e8f889bSDavid Gibson } 44801e8f889bSDavid Gibson 4481cb900f41SKirill A. Shutemov spin_unlock(ptl); 4482cb6acd01SMike Kravetz 4483cb6acd01SMike Kravetz /* 44848f251a3dSMike Kravetz * Only set HPageMigratable in newly allocated pages. Existing pages 44858f251a3dSMike Kravetz * found in the pagecache may not have HPageMigratableset if they have 44868f251a3dSMike Kravetz * been isolated for migration. 4487cb6acd01SMike Kravetz */ 4488cb6acd01SMike Kravetz if (new_page) 44898f251a3dSMike Kravetz SetHPageMigratable(page); 4490cb6acd01SMike Kravetz 44914c887265SAdam Litke unlock_page(page); 44924c887265SAdam Litke out: 4493ac9b9c66SHugh Dickins return ret; 44944c887265SAdam Litke 44954c887265SAdam Litke backout: 4496cb900f41SKirill A. Shutemov spin_unlock(ptl); 44972b26736cSAndy Whitcroft backout_unlocked: 44984c887265SAdam Litke unlock_page(page); 4499285b8dcaSHuang Ying restore_reserve_on_error(h, vma, haddr, page); 45004c887265SAdam Litke put_page(page); 45014c887265SAdam Litke goto out; 4502ac9b9c66SHugh Dickins } 4503ac9b9c66SHugh Dickins 45048382d914SDavidlohr Bueso #ifdef CONFIG_SMP 4505188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 45068382d914SDavidlohr Bueso { 45078382d914SDavidlohr Bueso unsigned long key[2]; 45088382d914SDavidlohr Bueso u32 hash; 45098382d914SDavidlohr Bueso 45108382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 45118382d914SDavidlohr Bueso key[1] = idx; 45128382d914SDavidlohr Bueso 451355254636SMike Kravetz hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); 45148382d914SDavidlohr Bueso 45158382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 45168382d914SDavidlohr Bueso } 45178382d914SDavidlohr Bueso #else 45188382d914SDavidlohr Bueso /* 45196c26d310SMiaohe Lin * For uniprocessor systems we always use a single mutex, so just 45208382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 45218382d914SDavidlohr Bueso */ 4522188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 45238382d914SDavidlohr Bueso { 45248382d914SDavidlohr Bueso return 0; 45258382d914SDavidlohr Bueso } 45268382d914SDavidlohr Bueso #endif 45278382d914SDavidlohr Bueso 45282b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 4529788c7df4SHugh Dickins unsigned long address, unsigned int flags) 453086e5216fSAdam Litke { 45318382d914SDavidlohr Bueso pte_t *ptep, entry; 4532cb900f41SKirill A. Shutemov spinlock_t *ptl; 45332b740303SSouptick Joarder vm_fault_t ret; 45348382d914SDavidlohr Bueso u32 hash; 45358382d914SDavidlohr Bueso pgoff_t idx; 45360fe6e20bSNaoya Horiguchi struct page *page = NULL; 453757303d80SAndy Whitcroft struct page *pagecache_page = NULL; 4538a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 45398382d914SDavidlohr Bueso struct address_space *mapping; 45400f792cf9SNaoya Horiguchi int need_wait_lock = 0; 4541285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 454286e5216fSAdam Litke 4543285b8dcaSHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 4544fd6a03edSNaoya Horiguchi if (ptep) { 4545c0d0381aSMike Kravetz /* 4546c0d0381aSMike Kravetz * Since we hold no locks, ptep could be stale. That is 4547c0d0381aSMike Kravetz * OK as we are only making decisions based on content and 4548c0d0381aSMike Kravetz * not actually modifying content here. 4549c0d0381aSMike Kravetz */ 4550fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 4551290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 4552cb900f41SKirill A. Shutemov migration_entry_wait_huge(vma, mm, ptep); 4553290408d4SNaoya Horiguchi return 0; 4554290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 4555aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 4556972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 4557b43a9990SMike Kravetz } 45588382d914SDavidlohr Bueso 4559c0d0381aSMike Kravetz /* 4560c0d0381aSMike Kravetz * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold 456187bf91d3SMike Kravetz * until finished with ptep. This serves two purposes: 456287bf91d3SMike Kravetz * 1) It prevents huge_pmd_unshare from being called elsewhere 456387bf91d3SMike Kravetz * and making the ptep no longer valid. 456487bf91d3SMike Kravetz * 2) It synchronizes us with i_size modifications during truncation. 4565c0d0381aSMike Kravetz * 4566c0d0381aSMike Kravetz * ptep could have already be assigned via huge_pte_offset. That 4567c0d0381aSMike Kravetz * is OK, as huge_pte_alloc will return the same value unless 4568c0d0381aSMike Kravetz * something has changed. 4569c0d0381aSMike Kravetz */ 4570ddeaab32SMike Kravetz mapping = vma->vm_file->f_mapping; 4571c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 4572aec44e0fSPeter Xu ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); 4573c0d0381aSMike Kravetz if (!ptep) { 4574c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 4575c0d0381aSMike Kravetz return VM_FAULT_OOM; 4576c0d0381aSMike Kravetz } 4577ddeaab32SMike Kravetz 45783935baa9SDavid Gibson /* 45793935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 45803935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 45813935baa9SDavid Gibson * the same page in the page cache. 45823935baa9SDavid Gibson */ 4583c0d0381aSMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 4584188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 4585c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 45868382d914SDavidlohr Bueso 45877f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 45887f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 45898382d914SDavidlohr Bueso ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 4590b4d1d99fSDavid Gibson goto out_mutex; 45913935baa9SDavid Gibson } 459286e5216fSAdam Litke 459383c54070SNick Piggin ret = 0; 45941e8f889bSDavid Gibson 459557303d80SAndy Whitcroft /* 45960f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 45970f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 45987c8de358SEthon Paul * an active hugepage in pagecache. This goto expects the 2nd page 45997c8de358SEthon Paul * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will 46007c8de358SEthon Paul * properly handle it. 46010f792cf9SNaoya Horiguchi */ 46020f792cf9SNaoya Horiguchi if (!pte_present(entry)) 46030f792cf9SNaoya Horiguchi goto out_mutex; 46040f792cf9SNaoya Horiguchi 46050f792cf9SNaoya Horiguchi /* 460657303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 460757303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 460857303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 460957303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 461057303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 461157303d80SAndy Whitcroft * consumed. 461257303d80SAndy Whitcroft */ 4613106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 4614285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 46152b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 4616b4d1d99fSDavid Gibson goto out_mutex; 46172b26736cSAndy Whitcroft } 46185e911373SMike Kravetz /* Just decrements count, does not deallocate */ 4619285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 462057303d80SAndy Whitcroft 4621f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 462257303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 4623285b8dcaSHuang Ying vma, haddr); 462457303d80SAndy Whitcroft } 462557303d80SAndy Whitcroft 46260f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 46270fe6e20bSNaoya Horiguchi 46281e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 4629b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 4630cb900f41SKirill A. Shutemov goto out_ptl; 4631b4d1d99fSDavid Gibson 46320f792cf9SNaoya Horiguchi /* 46330f792cf9SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 46340f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 46350f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 46360f792cf9SNaoya Horiguchi */ 46370f792cf9SNaoya Horiguchi page = pte_page(entry); 46380f792cf9SNaoya Horiguchi if (page != pagecache_page) 46390f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 46400f792cf9SNaoya Horiguchi need_wait_lock = 1; 46410f792cf9SNaoya Horiguchi goto out_ptl; 46420f792cf9SNaoya Horiguchi } 46430f792cf9SNaoya Horiguchi 46440f792cf9SNaoya Horiguchi get_page(page); 4645b4d1d99fSDavid Gibson 4646788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 4647106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 4648974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, 4649cb900f41SKirill A. Shutemov pagecache_page, ptl); 46500f792cf9SNaoya Horiguchi goto out_put_page; 4651b4d1d99fSDavid Gibson } 4652106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 4653b4d1d99fSDavid Gibson } 4654b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 4655285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 4656788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 4657285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 46580f792cf9SNaoya Horiguchi out_put_page: 46590f792cf9SNaoya Horiguchi if (page != pagecache_page) 46600f792cf9SNaoya Horiguchi unlock_page(page); 46610f792cf9SNaoya Horiguchi put_page(page); 4662cb900f41SKirill A. Shutemov out_ptl: 4663cb900f41SKirill A. Shutemov spin_unlock(ptl); 466457303d80SAndy Whitcroft 466557303d80SAndy Whitcroft if (pagecache_page) { 466657303d80SAndy Whitcroft unlock_page(pagecache_page); 466757303d80SAndy Whitcroft put_page(pagecache_page); 466857303d80SAndy Whitcroft } 4669b4d1d99fSDavid Gibson out_mutex: 4670c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 4671c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 46720f792cf9SNaoya Horiguchi /* 46730f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 46740f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 46750f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 46760f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 46770f792cf9SNaoya Horiguchi * here without taking refcount. 46780f792cf9SNaoya Horiguchi */ 46790f792cf9SNaoya Horiguchi if (need_wait_lock) 46800f792cf9SNaoya Horiguchi wait_on_page_locked(page); 46811e8f889bSDavid Gibson return ret; 468286e5216fSAdam Litke } 468386e5216fSAdam Litke 46848fb5debcSMike Kravetz /* 46858fb5debcSMike Kravetz * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with 46868fb5debcSMike Kravetz * modifications for huge pages. 46878fb5debcSMike Kravetz */ 46888fb5debcSMike Kravetz int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, 46898fb5debcSMike Kravetz pte_t *dst_pte, 46908fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 46918fb5debcSMike Kravetz unsigned long dst_addr, 46928fb5debcSMike Kravetz unsigned long src_addr, 46938fb5debcSMike Kravetz struct page **pagep) 46948fb5debcSMike Kravetz { 46951e392147SAndrea Arcangeli struct address_space *mapping; 46961e392147SAndrea Arcangeli pgoff_t idx; 46971e392147SAndrea Arcangeli unsigned long size; 46981c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 46998fb5debcSMike Kravetz struct hstate *h = hstate_vma(dst_vma); 47008fb5debcSMike Kravetz pte_t _dst_pte; 47018fb5debcSMike Kravetz spinlock_t *ptl; 47028fb5debcSMike Kravetz int ret; 47038fb5debcSMike Kravetz struct page *page; 47048fb5debcSMike Kravetz 47058fb5debcSMike Kravetz if (!*pagep) { 47068fb5debcSMike Kravetz ret = -ENOMEM; 47078fb5debcSMike Kravetz page = alloc_huge_page(dst_vma, dst_addr, 0); 47088fb5debcSMike Kravetz if (IS_ERR(page)) 47098fb5debcSMike Kravetz goto out; 47108fb5debcSMike Kravetz 47118fb5debcSMike Kravetz ret = copy_huge_page_from_user(page, 47128fb5debcSMike Kravetz (const void __user *) src_addr, 4713810a56b9SMike Kravetz pages_per_huge_page(h), false); 47148fb5debcSMike Kravetz 4715c1e8d7c6SMichel Lespinasse /* fallback to copy_from_user outside mmap_lock */ 47168fb5debcSMike Kravetz if (unlikely(ret)) { 47179e368259SAndrea Arcangeli ret = -ENOENT; 47188fb5debcSMike Kravetz *pagep = page; 47198fb5debcSMike Kravetz /* don't free the page */ 47208fb5debcSMike Kravetz goto out; 47218fb5debcSMike Kravetz } 47228fb5debcSMike Kravetz } else { 47238fb5debcSMike Kravetz page = *pagep; 47248fb5debcSMike Kravetz *pagep = NULL; 47258fb5debcSMike Kravetz } 47268fb5debcSMike Kravetz 47278fb5debcSMike Kravetz /* 47288fb5debcSMike Kravetz * The memory barrier inside __SetPageUptodate makes sure that 47298fb5debcSMike Kravetz * preceding stores to the page contents become visible before 47308fb5debcSMike Kravetz * the set_pte_at() write. 47318fb5debcSMike Kravetz */ 47328fb5debcSMike Kravetz __SetPageUptodate(page); 47338fb5debcSMike Kravetz 47341e392147SAndrea Arcangeli mapping = dst_vma->vm_file->f_mapping; 47351e392147SAndrea Arcangeli idx = vma_hugecache_offset(h, dst_vma, dst_addr); 47361e392147SAndrea Arcangeli 47371c9e8defSMike Kravetz /* 47381c9e8defSMike Kravetz * If shared, add to page cache 47391c9e8defSMike Kravetz */ 47401c9e8defSMike Kravetz if (vm_shared) { 47411e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 47421e392147SAndrea Arcangeli ret = -EFAULT; 47431e392147SAndrea Arcangeli if (idx >= size) 47441e392147SAndrea Arcangeli goto out_release_nounlock; 47451c9e8defSMike Kravetz 47461e392147SAndrea Arcangeli /* 47471e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 47481e392147SAndrea Arcangeli * huge_add_to_page_cache() below happens through the 47491e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 47501e392147SAndrea Arcangeli * the caller. 47511e392147SAndrea Arcangeli */ 47521c9e8defSMike Kravetz ret = huge_add_to_page_cache(page, mapping, idx); 47531c9e8defSMike Kravetz if (ret) 47541c9e8defSMike Kravetz goto out_release_nounlock; 47551c9e8defSMike Kravetz } 47561c9e8defSMike Kravetz 47578fb5debcSMike Kravetz ptl = huge_pte_lockptr(h, dst_mm, dst_pte); 47588fb5debcSMike Kravetz spin_lock(ptl); 47598fb5debcSMike Kravetz 47601e392147SAndrea Arcangeli /* 47611e392147SAndrea Arcangeli * Recheck the i_size after holding PT lock to make sure not 47621e392147SAndrea Arcangeli * to leave any page mapped (as page_mapped()) beyond the end 47631e392147SAndrea Arcangeli * of the i_size (remove_inode_hugepages() is strict about 47641e392147SAndrea Arcangeli * enforcing that). If we bail out here, we'll also leave a 47651e392147SAndrea Arcangeli * page in the radix tree in the vm_shared case beyond the end 47661e392147SAndrea Arcangeli * of the i_size, but remove_inode_hugepages() will take care 47671e392147SAndrea Arcangeli * of it as soon as we drop the hugetlb_fault_mutex_table. 47681e392147SAndrea Arcangeli */ 47691e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 47701e392147SAndrea Arcangeli ret = -EFAULT; 47711e392147SAndrea Arcangeli if (idx >= size) 47721e392147SAndrea Arcangeli goto out_release_unlock; 47731e392147SAndrea Arcangeli 47748fb5debcSMike Kravetz ret = -EEXIST; 47758fb5debcSMike Kravetz if (!huge_pte_none(huge_ptep_get(dst_pte))) 47768fb5debcSMike Kravetz goto out_release_unlock; 47778fb5debcSMike Kravetz 47781c9e8defSMike Kravetz if (vm_shared) { 47791c9e8defSMike Kravetz page_dup_rmap(page, true); 47801c9e8defSMike Kravetz } else { 4781d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 47828fb5debcSMike Kravetz hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); 47831c9e8defSMike Kravetz } 47848fb5debcSMike Kravetz 47858fb5debcSMike Kravetz _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); 47868fb5debcSMike Kravetz if (dst_vma->vm_flags & VM_WRITE) 47878fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 47888fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 47898fb5debcSMike Kravetz 47908fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 47918fb5debcSMike Kravetz 47928fb5debcSMike Kravetz (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, 47938fb5debcSMike Kravetz dst_vma->vm_flags & VM_WRITE); 47948fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 47958fb5debcSMike Kravetz 47968fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 47978fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 47988fb5debcSMike Kravetz 47998fb5debcSMike Kravetz spin_unlock(ptl); 48008f251a3dSMike Kravetz SetHPageMigratable(page); 48011c9e8defSMike Kravetz if (vm_shared) 48021c9e8defSMike Kravetz unlock_page(page); 48038fb5debcSMike Kravetz ret = 0; 48048fb5debcSMike Kravetz out: 48058fb5debcSMike Kravetz return ret; 48068fb5debcSMike Kravetz out_release_unlock: 48078fb5debcSMike Kravetz spin_unlock(ptl); 48081c9e8defSMike Kravetz if (vm_shared) 48091c9e8defSMike Kravetz unlock_page(page); 48105af10dfdSAndrea Arcangeli out_release_nounlock: 48118fb5debcSMike Kravetz put_page(page); 48128fb5debcSMike Kravetz goto out; 48138fb5debcSMike Kravetz } 48148fb5debcSMike Kravetz 481582e5d378SJoao Martins static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma, 481682e5d378SJoao Martins int refs, struct page **pages, 481782e5d378SJoao Martins struct vm_area_struct **vmas) 481882e5d378SJoao Martins { 481982e5d378SJoao Martins int nr; 482082e5d378SJoao Martins 482182e5d378SJoao Martins for (nr = 0; nr < refs; nr++) { 482282e5d378SJoao Martins if (likely(pages)) 482382e5d378SJoao Martins pages[nr] = mem_map_offset(page, nr); 482482e5d378SJoao Martins if (vmas) 482582e5d378SJoao Martins vmas[nr] = vma; 482682e5d378SJoao Martins } 482782e5d378SJoao Martins } 482882e5d378SJoao Martins 482928a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 483063551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 483128a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 48324f6da934SPeter Xu long i, unsigned int flags, int *locked) 483363551ae0SDavid Gibson { 4834d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 4835d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 483628a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 4837a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 48380fa5bc40SJoao Martins int err = -EFAULT, refs; 483963551ae0SDavid Gibson 484063551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 484163551ae0SDavid Gibson pte_t *pte; 4842cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 48432a15efc9SHugh Dickins int absent; 484463551ae0SDavid Gibson struct page *page; 484563551ae0SDavid Gibson 48464c887265SAdam Litke /* 484702057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 484802057967SDavid Rientjes * potentially allocating memory. 484902057967SDavid Rientjes */ 4850fa45f116SDavidlohr Bueso if (fatal_signal_pending(current)) { 485102057967SDavid Rientjes remainder = 0; 485202057967SDavid Rientjes break; 485302057967SDavid Rientjes } 485402057967SDavid Rientjes 485502057967SDavid Rientjes /* 48564c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 48572a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 48584c887265SAdam Litke * first, for the page indexing below to work. 4859cb900f41SKirill A. Shutemov * 4860cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 48614c887265SAdam Litke */ 48627868a208SPunit Agrawal pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), 48637868a208SPunit Agrawal huge_page_size(h)); 4864cb900f41SKirill A. Shutemov if (pte) 4865cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 48662a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 486763551ae0SDavid Gibson 48682a15efc9SHugh Dickins /* 48692a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 48703ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 48713ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 48723ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 48733ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 48742a15efc9SHugh Dickins */ 48753ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 48763ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 4877cb900f41SKirill A. Shutemov if (pte) 4878cb900f41SKirill A. Shutemov spin_unlock(ptl); 48792a15efc9SHugh Dickins remainder = 0; 48802a15efc9SHugh Dickins break; 48812a15efc9SHugh Dickins } 48822a15efc9SHugh Dickins 48839cc3a5bdSNaoya Horiguchi /* 48849cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 48859cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 48869cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 48879cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 48889cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 48899cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 48909cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 48919cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 48929cc3a5bdSNaoya Horiguchi */ 48939cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 4894106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 4895106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 48962b740303SSouptick Joarder vm_fault_t ret; 489787ffc118SAndrea Arcangeli unsigned int fault_flags = 0; 48984c887265SAdam Litke 4899cb900f41SKirill A. Shutemov if (pte) 4900cb900f41SKirill A. Shutemov spin_unlock(ptl); 490187ffc118SAndrea Arcangeli if (flags & FOLL_WRITE) 490287ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_WRITE; 49034f6da934SPeter Xu if (locked) 490471335f37SPeter Xu fault_flags |= FAULT_FLAG_ALLOW_RETRY | 490571335f37SPeter Xu FAULT_FLAG_KILLABLE; 490687ffc118SAndrea Arcangeli if (flags & FOLL_NOWAIT) 490787ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY | 490887ffc118SAndrea Arcangeli FAULT_FLAG_RETRY_NOWAIT; 490987ffc118SAndrea Arcangeli if (flags & FOLL_TRIED) { 49104426e945SPeter Xu /* 49114426e945SPeter Xu * Note: FAULT_FLAG_ALLOW_RETRY and 49124426e945SPeter Xu * FAULT_FLAG_TRIED can co-exist 49134426e945SPeter Xu */ 491487ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_TRIED; 491587ffc118SAndrea Arcangeli } 491687ffc118SAndrea Arcangeli ret = hugetlb_fault(mm, vma, vaddr, fault_flags); 491787ffc118SAndrea Arcangeli if (ret & VM_FAULT_ERROR) { 49182be7cfedSDaniel Jordan err = vm_fault_to_errno(ret, flags); 49191c59827dSHugh Dickins remainder = 0; 49201c59827dSHugh Dickins break; 49211c59827dSHugh Dickins } 492287ffc118SAndrea Arcangeli if (ret & VM_FAULT_RETRY) { 49234f6da934SPeter Xu if (locked && 49241ac25013SAndrea Arcangeli !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) 49254f6da934SPeter Xu *locked = 0; 492687ffc118SAndrea Arcangeli *nr_pages = 0; 492787ffc118SAndrea Arcangeli /* 492887ffc118SAndrea Arcangeli * VM_FAULT_RETRY must not return an 492987ffc118SAndrea Arcangeli * error, it will return zero 493087ffc118SAndrea Arcangeli * instead. 493187ffc118SAndrea Arcangeli * 493287ffc118SAndrea Arcangeli * No need to update "position" as the 493387ffc118SAndrea Arcangeli * caller will not check it after 493487ffc118SAndrea Arcangeli * *nr_pages is set to 0. 493587ffc118SAndrea Arcangeli */ 493687ffc118SAndrea Arcangeli return i; 493787ffc118SAndrea Arcangeli } 493887ffc118SAndrea Arcangeli continue; 493987ffc118SAndrea Arcangeli } 494063551ae0SDavid Gibson 4941a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 49427f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 49438fde12caSLinus Torvalds 49448fde12caSLinus Torvalds /* 4945acbfb087SZhigang Lu * If subpage information not requested, update counters 4946acbfb087SZhigang Lu * and skip the same_page loop below. 4947acbfb087SZhigang Lu */ 4948acbfb087SZhigang Lu if (!pages && !vmas && !pfn_offset && 4949acbfb087SZhigang Lu (vaddr + huge_page_size(h) < vma->vm_end) && 4950acbfb087SZhigang Lu (remainder >= pages_per_huge_page(h))) { 4951acbfb087SZhigang Lu vaddr += huge_page_size(h); 4952acbfb087SZhigang Lu remainder -= pages_per_huge_page(h); 4953acbfb087SZhigang Lu i += pages_per_huge_page(h); 4954acbfb087SZhigang Lu spin_unlock(ptl); 4955acbfb087SZhigang Lu continue; 4956acbfb087SZhigang Lu } 4957acbfb087SZhigang Lu 495882e5d378SJoao Martins refs = min3(pages_per_huge_page(h) - pfn_offset, 495982e5d378SJoao Martins (vma->vm_end - vaddr) >> PAGE_SHIFT, remainder); 49600fa5bc40SJoao Martins 496182e5d378SJoao Martins if (pages || vmas) 496282e5d378SJoao Martins record_subpages_vmas(mem_map_offset(page, pfn_offset), 496382e5d378SJoao Martins vma, refs, 496482e5d378SJoao Martins likely(pages) ? pages + i : NULL, 496582e5d378SJoao Martins vmas ? vmas + i : NULL); 496663551ae0SDavid Gibson 496782e5d378SJoao Martins if (pages) { 49680fa5bc40SJoao Martins /* 49690fa5bc40SJoao Martins * try_grab_compound_head() should always succeed here, 49700fa5bc40SJoao Martins * because: a) we hold the ptl lock, and b) we've just 49710fa5bc40SJoao Martins * checked that the huge page is present in the page 49720fa5bc40SJoao Martins * tables. If the huge page is present, then the tail 49730fa5bc40SJoao Martins * pages must also be present. The ptl prevents the 49740fa5bc40SJoao Martins * head page and tail pages from being rearranged in 49750fa5bc40SJoao Martins * any way. So this page must be available at this 49760fa5bc40SJoao Martins * point, unless the page refcount overflowed: 49770fa5bc40SJoao Martins */ 497882e5d378SJoao Martins if (WARN_ON_ONCE(!try_grab_compound_head(pages[i], 49790fa5bc40SJoao Martins refs, 49800fa5bc40SJoao Martins flags))) { 49810fa5bc40SJoao Martins spin_unlock(ptl); 49820fa5bc40SJoao Martins remainder = 0; 49830fa5bc40SJoao Martins err = -ENOMEM; 49840fa5bc40SJoao Martins break; 49850fa5bc40SJoao Martins } 4986d5d4b0aaSChen, Kenneth W } 498782e5d378SJoao Martins 498882e5d378SJoao Martins vaddr += (refs << PAGE_SHIFT); 498982e5d378SJoao Martins remainder -= refs; 499082e5d378SJoao Martins i += refs; 499182e5d378SJoao Martins 4992cb900f41SKirill A. Shutemov spin_unlock(ptl); 499363551ae0SDavid Gibson } 499428a35716SMichel Lespinasse *nr_pages = remainder; 499587ffc118SAndrea Arcangeli /* 499687ffc118SAndrea Arcangeli * setting position is actually required only if remainder is 499787ffc118SAndrea Arcangeli * not zero but it's faster not to add a "if (remainder)" 499887ffc118SAndrea Arcangeli * branch. 499987ffc118SAndrea Arcangeli */ 500063551ae0SDavid Gibson *position = vaddr; 500163551ae0SDavid Gibson 50022be7cfedSDaniel Jordan return i ? i : err; 500363551ae0SDavid Gibson } 50048f860591SZhang, Yanmin 50057da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 50068f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 50078f860591SZhang, Yanmin { 50088f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 50098f860591SZhang, Yanmin unsigned long start = address; 50108f860591SZhang, Yanmin pte_t *ptep; 50118f860591SZhang, Yanmin pte_t pte; 5012a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 50137da4d641SPeter Zijlstra unsigned long pages = 0; 5014dff11abeSMike Kravetz bool shared_pmd = false; 5015ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 5016dff11abeSMike Kravetz 5017dff11abeSMike Kravetz /* 5018dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 5019ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 5020dff11abeSMike Kravetz * range if PMD sharing is possible. 5021dff11abeSMike Kravetz */ 50227269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 50237269f999SJérôme Glisse 0, vma, mm, start, end); 5024ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 50258f860591SZhang, Yanmin 50268f860591SZhang, Yanmin BUG_ON(address >= end); 5027ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 50288f860591SZhang, Yanmin 5029ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 503083cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 5031a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 5032cb900f41SKirill A. Shutemov spinlock_t *ptl; 50337868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, huge_page_size(h)); 50348f860591SZhang, Yanmin if (!ptep) 50358f860591SZhang, Yanmin continue; 5036cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 503734ae204fSMike Kravetz if (huge_pmd_unshare(mm, vma, &address, ptep)) { 50387da4d641SPeter Zijlstra pages++; 5039cb900f41SKirill A. Shutemov spin_unlock(ptl); 5040dff11abeSMike Kravetz shared_pmd = true; 504139dde65cSChen, Kenneth W continue; 50427da4d641SPeter Zijlstra } 5043a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 5044a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 5045a8bda28dSNaoya Horiguchi spin_unlock(ptl); 5046a8bda28dSNaoya Horiguchi continue; 5047a8bda28dSNaoya Horiguchi } 5048a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 5049a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 5050a8bda28dSNaoya Horiguchi 5051a8bda28dSNaoya Horiguchi if (is_write_migration_entry(entry)) { 5052a8bda28dSNaoya Horiguchi pte_t newpte; 5053a8bda28dSNaoya Horiguchi 5054a8bda28dSNaoya Horiguchi make_migration_entry_read(&entry); 5055a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 5056e5251fd4SPunit Agrawal set_huge_swap_pte_at(mm, address, ptep, 5057e5251fd4SPunit Agrawal newpte, huge_page_size(h)); 5058a8bda28dSNaoya Horiguchi pages++; 5059a8bda28dSNaoya Horiguchi } 5060a8bda28dSNaoya Horiguchi spin_unlock(ptl); 5061a8bda28dSNaoya Horiguchi continue; 5062a8bda28dSNaoya Horiguchi } 5063a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 5064023bdd00SAneesh Kumar K.V pte_t old_pte; 5065023bdd00SAneesh Kumar K.V 5066023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 5067023bdd00SAneesh Kumar K.V pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); 5068be7517d6STony Lu pte = arch_make_huge_pte(pte, vma, NULL, 0); 5069023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 50707da4d641SPeter Zijlstra pages++; 50718f860591SZhang, Yanmin } 5072cb900f41SKirill A. Shutemov spin_unlock(ptl); 50738f860591SZhang, Yanmin } 5074d833352aSMel Gorman /* 5075c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 5076d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 5077c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 5078dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 5079dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 5080d833352aSMel Gorman */ 5081dff11abeSMike Kravetz if (shared_pmd) 5082ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 5083dff11abeSMike Kravetz else 50845491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 50850f10851eSJérôme Glisse /* 50860f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 50870f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 50880f10851eSJérôme Glisse * 5089ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 50900f10851eSJérôme Glisse */ 509183cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 5092ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 50937da4d641SPeter Zijlstra 50947da4d641SPeter Zijlstra return pages << h->order; 50958f860591SZhang, Yanmin } 50968f860591SZhang, Yanmin 509733b8f84aSMike Kravetz /* Return true if reservation was successful, false otherwise. */ 509833b8f84aSMike Kravetz bool hugetlb_reserve_pages(struct inode *inode, 5099a1e78772SMel Gorman long from, long to, 51005a6fe125SMel Gorman struct vm_area_struct *vma, 5101ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 5102e4e574b7SAdam Litke { 510333b8f84aSMike Kravetz long chg, add = -1; 5104a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 510590481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 51069119a41eSJoonsoo Kim struct resv_map *resv_map; 5107075a61d0SMina Almasry struct hugetlb_cgroup *h_cg = NULL; 51080db9d74eSMina Almasry long gbl_reserve, regions_needed = 0; 5109e4e574b7SAdam Litke 511063489f8eSMike Kravetz /* This should never happen */ 511163489f8eSMike Kravetz if (from > to) { 511263489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 511333b8f84aSMike Kravetz return false; 511463489f8eSMike Kravetz } 511563489f8eSMike Kravetz 5116a1e78772SMel Gorman /* 511717c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 511817c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 511990481622SDavid Gibson * without using reserves 512017c9d12eSMel Gorman */ 5121ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 512233b8f84aSMike Kravetz return true; 512317c9d12eSMel Gorman 512417c9d12eSMel Gorman /* 5125a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 5126a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 5127a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 5128a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 5129a1e78772SMel Gorman */ 51309119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 5131f27a5136SMike Kravetz /* 5132f27a5136SMike Kravetz * resv_map can not be NULL as hugetlb_reserve_pages is only 5133f27a5136SMike Kravetz * called for inodes for which resv_maps were created (see 5134f27a5136SMike Kravetz * hugetlbfs_get_inode). 5135f27a5136SMike Kravetz */ 51364e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 51379119a41eSJoonsoo Kim 51380db9d74eSMina Almasry chg = region_chg(resv_map, from, to, ®ions_needed); 51399119a41eSJoonsoo Kim 51409119a41eSJoonsoo Kim } else { 5141e9fe92aeSMina Almasry /* Private mapping. */ 51429119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 51435a6fe125SMel Gorman if (!resv_map) 514433b8f84aSMike Kravetz return false; 51455a6fe125SMel Gorman 514617c9d12eSMel Gorman chg = to - from; 514717c9d12eSMel Gorman 51485a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 51495a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 51505a6fe125SMel Gorman } 51515a6fe125SMel Gorman 515233b8f84aSMike Kravetz if (chg < 0) 5153c50ac050SDave Hansen goto out_err; 515417c9d12eSMel Gorman 515533b8f84aSMike Kravetz if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), 515633b8f84aSMike Kravetz chg * pages_per_huge_page(h), &h_cg) < 0) 5157075a61d0SMina Almasry goto out_err; 5158075a61d0SMina Almasry 5159075a61d0SMina Almasry if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { 5160075a61d0SMina Almasry /* For private mappings, the hugetlb_cgroup uncharge info hangs 5161075a61d0SMina Almasry * of the resv_map. 5162075a61d0SMina Almasry */ 5163075a61d0SMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); 5164075a61d0SMina Almasry } 5165075a61d0SMina Almasry 51661c5ecae3SMike Kravetz /* 51671c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 51681c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 51691c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 51701c5ecae3SMike Kravetz */ 51711c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 517233b8f84aSMike Kravetz if (gbl_reserve < 0) 5173075a61d0SMina Almasry goto out_uncharge_cgroup; 517417c9d12eSMel Gorman 517517c9d12eSMel Gorman /* 517617c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 517790481622SDavid Gibson * Hand the pages back to the subpool if there are not 517817c9d12eSMel Gorman */ 517933b8f84aSMike Kravetz if (hugetlb_acct_memory(h, gbl_reserve) < 0) 5180075a61d0SMina Almasry goto out_put_pages; 518117c9d12eSMel Gorman 518217c9d12eSMel Gorman /* 518317c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 518417c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 518517c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 518617c9d12eSMel Gorman * the reservation was and the page cache tells how much of 518717c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 518817c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 518917c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 519017c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 519117c9d12eSMel Gorman * else has to be done for private mappings here 519217c9d12eSMel Gorman */ 519333039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 5194075a61d0SMina Almasry add = region_add(resv_map, from, to, regions_needed, h, h_cg); 519533039678SMike Kravetz 51960db9d74eSMina Almasry if (unlikely(add < 0)) { 51970db9d74eSMina Almasry hugetlb_acct_memory(h, -gbl_reserve); 5198075a61d0SMina Almasry goto out_put_pages; 51990db9d74eSMina Almasry } else if (unlikely(chg > add)) { 520033039678SMike Kravetz /* 520133039678SMike Kravetz * pages in this range were added to the reserve 520233039678SMike Kravetz * map between region_chg and region_add. This 520333039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 520433039678SMike Kravetz * the subpool and reserve counts modified above 520533039678SMike Kravetz * based on the difference. 520633039678SMike Kravetz */ 520733039678SMike Kravetz long rsv_adjust; 520833039678SMike Kravetz 5209d85aecf2SMiaohe Lin /* 5210d85aecf2SMiaohe Lin * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the 5211d85aecf2SMiaohe Lin * reference to h_cg->css. See comment below for detail. 5212d85aecf2SMiaohe Lin */ 5213075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd( 5214075a61d0SMina Almasry hstate_index(h), 5215075a61d0SMina Almasry (chg - add) * pages_per_huge_page(h), h_cg); 5216075a61d0SMina Almasry 521733039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 521833039678SMike Kravetz chg - add); 521933039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 5220d85aecf2SMiaohe Lin } else if (h_cg) { 5221d85aecf2SMiaohe Lin /* 5222d85aecf2SMiaohe Lin * The file_regions will hold their own reference to 5223d85aecf2SMiaohe Lin * h_cg->css. So we should release the reference held 5224d85aecf2SMiaohe Lin * via hugetlb_cgroup_charge_cgroup_rsvd() when we are 5225d85aecf2SMiaohe Lin * done. 5226d85aecf2SMiaohe Lin */ 5227d85aecf2SMiaohe Lin hugetlb_cgroup_put_rsvd_cgroup(h_cg); 522833039678SMike Kravetz } 522933039678SMike Kravetz } 523033b8f84aSMike Kravetz return true; 523133b8f84aSMike Kravetz 5232075a61d0SMina Almasry out_put_pages: 5233075a61d0SMina Almasry /* put back original number of pages, chg */ 5234075a61d0SMina Almasry (void)hugepage_subpool_put_pages(spool, chg); 5235075a61d0SMina Almasry out_uncharge_cgroup: 5236075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), 5237075a61d0SMina Almasry chg * pages_per_huge_page(h), h_cg); 5238c50ac050SDave Hansen out_err: 52395e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 52400db9d74eSMina Almasry /* Only call region_abort if the region_chg succeeded but the 52410db9d74eSMina Almasry * region_add failed or didn't run. 52420db9d74eSMina Almasry */ 52430db9d74eSMina Almasry if (chg >= 0 && add < 0) 52440db9d74eSMina Almasry region_abort(resv_map, from, to, regions_needed); 5245f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 5246f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 524733b8f84aSMike Kravetz return false; 5248a43a8c39SChen, Kenneth W } 5249a43a8c39SChen, Kenneth W 5250b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 5251b5cec28dSMike Kravetz long freed) 5252a43a8c39SChen, Kenneth W { 5253a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 52544e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 52559119a41eSJoonsoo Kim long chg = 0; 525690481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 52571c5ecae3SMike Kravetz long gbl_reserve; 525845c682a6SKen Chen 5259f27a5136SMike Kravetz /* 5260f27a5136SMike Kravetz * Since this routine can be called in the evict inode path for all 5261f27a5136SMike Kravetz * hugetlbfs inodes, resv_map could be NULL. 5262f27a5136SMike Kravetz */ 5263b5cec28dSMike Kravetz if (resv_map) { 5264b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 5265b5cec28dSMike Kravetz /* 5266b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 5267b5cec28dSMike Kravetz * must be split and another region descriptor can not be 5268b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 5269b5cec28dSMike Kravetz */ 5270b5cec28dSMike Kravetz if (chg < 0) 5271b5cec28dSMike Kravetz return chg; 5272b5cec28dSMike Kravetz } 5273b5cec28dSMike Kravetz 527445c682a6SKen Chen spin_lock(&inode->i_lock); 5275e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 527645c682a6SKen Chen spin_unlock(&inode->i_lock); 527745c682a6SKen Chen 52781c5ecae3SMike Kravetz /* 52791c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 52801c5ecae3SMike Kravetz * reservations to be released may be adjusted. 5281dddf31a4SMiaohe Lin * 5282dddf31a4SMiaohe Lin * Note that !resv_map implies freed == 0. So (chg - freed) 5283dddf31a4SMiaohe Lin * won't go negative. 52841c5ecae3SMike Kravetz */ 52851c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 52861c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 5287b5cec28dSMike Kravetz 5288b5cec28dSMike Kravetz return 0; 5289a43a8c39SChen, Kenneth W } 529093f70f90SNaoya Horiguchi 52913212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 52923212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 52933212b535SSteve Capper struct vm_area_struct *vma, 52943212b535SSteve Capper unsigned long addr, pgoff_t idx) 52953212b535SSteve Capper { 52963212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 52973212b535SSteve Capper svma->vm_start; 52983212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 52993212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 53003212b535SSteve Capper 53013212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 5302de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 5303de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 53043212b535SSteve Capper 53053212b535SSteve Capper /* 53063212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 53073212b535SSteve Capper * page table page. 53083212b535SSteve Capper */ 53093212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 53103212b535SSteve Capper vm_flags != svm_flags || 531107e51edfSMiaohe Lin !range_in_vma(svma, sbase, s_end)) 53123212b535SSteve Capper return 0; 53133212b535SSteve Capper 53143212b535SSteve Capper return saddr; 53153212b535SSteve Capper } 53163212b535SSteve Capper 531731aafb45SNicholas Krause static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) 53183212b535SSteve Capper { 53193212b535SSteve Capper unsigned long base = addr & PUD_MASK; 53203212b535SSteve Capper unsigned long end = base + PUD_SIZE; 53213212b535SSteve Capper 53223212b535SSteve Capper /* 53233212b535SSteve Capper * check on proper vm_flags and page table alignment 53243212b535SSteve Capper */ 5325017b1660SMike Kravetz if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) 532631aafb45SNicholas Krause return true; 532731aafb45SNicholas Krause return false; 53283212b535SSteve Capper } 53293212b535SSteve Capper 5330c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 5331c1991e07SPeter Xu { 5332c1991e07SPeter Xu #ifdef CONFIG_USERFAULTFD 5333c1991e07SPeter Xu if (uffd_disable_huge_pmd_share(vma)) 5334c1991e07SPeter Xu return false; 5335c1991e07SPeter Xu #endif 5336c1991e07SPeter Xu return vma_shareable(vma, addr); 5337c1991e07SPeter Xu } 5338c1991e07SPeter Xu 53393212b535SSteve Capper /* 5340017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 5341017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 5342017b1660SMike Kravetz * shared pmd mappings. 5343017b1660SMike Kravetz */ 5344017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 5345017b1660SMike Kravetz unsigned long *start, unsigned long *end) 5346017b1660SMike Kravetz { 5347a1ba9da8SLi Xinhai unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), 5348a1ba9da8SLi Xinhai v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 5349017b1660SMike Kravetz 5350a1ba9da8SLi Xinhai /* 5351a1ba9da8SLi Xinhai * vma need span at least one aligned PUD size and the start,end range 5352a1ba9da8SLi Xinhai * must at least partialy within it. 5353a1ba9da8SLi Xinhai */ 5354a1ba9da8SLi Xinhai if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || 5355a1ba9da8SLi Xinhai (*end <= v_start) || (*start >= v_end)) 5356017b1660SMike Kravetz return; 5357017b1660SMike Kravetz 535875802ca6SPeter Xu /* Extend the range to be PUD aligned for a worst case scenario */ 5359a1ba9da8SLi Xinhai if (*start > v_start) 5360a1ba9da8SLi Xinhai *start = ALIGN_DOWN(*start, PUD_SIZE); 5361017b1660SMike Kravetz 5362a1ba9da8SLi Xinhai if (*end < v_end) 5363a1ba9da8SLi Xinhai *end = ALIGN(*end, PUD_SIZE); 5364017b1660SMike Kravetz } 5365017b1660SMike Kravetz 5366017b1660SMike Kravetz /* 53673212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 53683212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 53693212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 5370c0d0381aSMike Kravetz * code much cleaner. 5371c0d0381aSMike Kravetz * 53720bf7b64eSMike Kravetz * This routine must be called with i_mmap_rwsem held in at least read mode if 53730bf7b64eSMike Kravetz * sharing is possible. For hugetlbfs, this prevents removal of any page 53740bf7b64eSMike Kravetz * table entries associated with the address space. This is important as we 53750bf7b64eSMike Kravetz * are setting up sharing based on existing page table entries (mappings). 53760bf7b64eSMike Kravetz * 53770bf7b64eSMike Kravetz * NOTE: This routine is only called from huge_pte_alloc. Some callers of 53780bf7b64eSMike Kravetz * huge_pte_alloc know that sharing is not possible and do not take 53790bf7b64eSMike Kravetz * i_mmap_rwsem as a performance optimization. This is handled by the 53800bf7b64eSMike Kravetz * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is 53810bf7b64eSMike Kravetz * only required for subsequent processing. 53823212b535SSteve Capper */ 5383aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 5384aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 53853212b535SSteve Capper { 53863212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 53873212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 53883212b535SSteve Capper vma->vm_pgoff; 53893212b535SSteve Capper struct vm_area_struct *svma; 53903212b535SSteve Capper unsigned long saddr; 53913212b535SSteve Capper pte_t *spte = NULL; 53923212b535SSteve Capper pte_t *pte; 5393cb900f41SKirill A. Shutemov spinlock_t *ptl; 53943212b535SSteve Capper 53950bf7b64eSMike Kravetz i_mmap_assert_locked(mapping); 53963212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 53973212b535SSteve Capper if (svma == vma) 53983212b535SSteve Capper continue; 53993212b535SSteve Capper 54003212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 54013212b535SSteve Capper if (saddr) { 54027868a208SPunit Agrawal spte = huge_pte_offset(svma->vm_mm, saddr, 54037868a208SPunit Agrawal vma_mmu_pagesize(svma)); 54043212b535SSteve Capper if (spte) { 54053212b535SSteve Capper get_page(virt_to_page(spte)); 54063212b535SSteve Capper break; 54073212b535SSteve Capper } 54083212b535SSteve Capper } 54093212b535SSteve Capper } 54103212b535SSteve Capper 54113212b535SSteve Capper if (!spte) 54123212b535SSteve Capper goto out; 54133212b535SSteve Capper 54148bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 5415dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 54163212b535SSteve Capper pud_populate(mm, pud, 54173212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 5418c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 5419dc6c9a35SKirill A. Shutemov } else { 54203212b535SSteve Capper put_page(virt_to_page(spte)); 5421dc6c9a35SKirill A. Shutemov } 5422cb900f41SKirill A. Shutemov spin_unlock(ptl); 54233212b535SSteve Capper out: 54243212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 54253212b535SSteve Capper return pte; 54263212b535SSteve Capper } 54273212b535SSteve Capper 54283212b535SSteve Capper /* 54293212b535SSteve Capper * unmap huge page backed by shared pte. 54303212b535SSteve Capper * 54313212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 54323212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 54333212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 54343212b535SSteve Capper * 5435c0d0381aSMike Kravetz * Called with page table lock held and i_mmap_rwsem held in write mode. 54363212b535SSteve Capper * 54373212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 54383212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 54393212b535SSteve Capper */ 544034ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 544134ae204fSMike Kravetz unsigned long *addr, pte_t *ptep) 54423212b535SSteve Capper { 54433212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 5444c2febafcSKirill A. Shutemov p4d_t *p4d = p4d_offset(pgd, *addr); 5445c2febafcSKirill A. Shutemov pud_t *pud = pud_offset(p4d, *addr); 54463212b535SSteve Capper 544734ae204fSMike Kravetz i_mmap_assert_write_locked(vma->vm_file->f_mapping); 54483212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 54493212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 54503212b535SSteve Capper return 0; 54513212b535SSteve Capper 54523212b535SSteve Capper pud_clear(pud); 54533212b535SSteve Capper put_page(virt_to_page(ptep)); 5454dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 54553212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 54563212b535SSteve Capper return 1; 54573212b535SSteve Capper } 5458c1991e07SPeter Xu 54599e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 5460aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 5461aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 54629e5fc74cSSteve Capper { 54639e5fc74cSSteve Capper return NULL; 54649e5fc74cSSteve Capper } 5465e81f2d22SZhang Zhen 546634ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 546734ae204fSMike Kravetz unsigned long *addr, pte_t *ptep) 5468e81f2d22SZhang Zhen { 5469e81f2d22SZhang Zhen return 0; 5470e81f2d22SZhang Zhen } 5471017b1660SMike Kravetz 5472017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 5473017b1660SMike Kravetz unsigned long *start, unsigned long *end) 5474017b1660SMike Kravetz { 5475017b1660SMike Kravetz } 5476c1991e07SPeter Xu 5477c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 5478c1991e07SPeter Xu { 5479c1991e07SPeter Xu return false; 5480c1991e07SPeter Xu } 54813212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 54823212b535SSteve Capper 54839e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 5484aec44e0fSPeter Xu pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, 54859e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 54869e5fc74cSSteve Capper { 54879e5fc74cSSteve Capper pgd_t *pgd; 5488c2febafcSKirill A. Shutemov p4d_t *p4d; 54899e5fc74cSSteve Capper pud_t *pud; 54909e5fc74cSSteve Capper pte_t *pte = NULL; 54919e5fc74cSSteve Capper 54929e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 5493f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 5494f4f0a3d8SKirill A. Shutemov if (!p4d) 5495f4f0a3d8SKirill A. Shutemov return NULL; 5496c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 54979e5fc74cSSteve Capper if (pud) { 54989e5fc74cSSteve Capper if (sz == PUD_SIZE) { 54999e5fc74cSSteve Capper pte = (pte_t *)pud; 55009e5fc74cSSteve Capper } else { 55019e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 5502c1991e07SPeter Xu if (want_pmd_share(vma, addr) && pud_none(*pud)) 5503aec44e0fSPeter Xu pte = huge_pmd_share(mm, vma, addr, pud); 55049e5fc74cSSteve Capper else 55059e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 55069e5fc74cSSteve Capper } 55079e5fc74cSSteve Capper } 55084e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 55099e5fc74cSSteve Capper 55109e5fc74cSSteve Capper return pte; 55119e5fc74cSSteve Capper } 55129e5fc74cSSteve Capper 55139b19df29SPunit Agrawal /* 55149b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 55159b19df29SPunit Agrawal * entry at address @addr 55169b19df29SPunit Agrawal * 55178ac0b81aSLi Xinhai * Return: Pointer to page table entry (PUD or PMD) for 55188ac0b81aSLi Xinhai * address @addr, or NULL if a !p*d_present() entry is encountered and the 55199b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 55209b19df29SPunit Agrawal * table. 55219b19df29SPunit Agrawal */ 55227868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 55237868a208SPunit Agrawal unsigned long addr, unsigned long sz) 55249e5fc74cSSteve Capper { 55259e5fc74cSSteve Capper pgd_t *pgd; 5526c2febafcSKirill A. Shutemov p4d_t *p4d; 55278ac0b81aSLi Xinhai pud_t *pud; 55288ac0b81aSLi Xinhai pmd_t *pmd; 55299e5fc74cSSteve Capper 55309e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 5531c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 5532c2febafcSKirill A. Shutemov return NULL; 5533c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 5534c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 5535c2febafcSKirill A. Shutemov return NULL; 55369b19df29SPunit Agrawal 5537c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 55388ac0b81aSLi Xinhai if (sz == PUD_SIZE) 55398ac0b81aSLi Xinhai /* must be pud huge, non-present or none */ 55409e5fc74cSSteve Capper return (pte_t *)pud; 55418ac0b81aSLi Xinhai if (!pud_present(*pud)) 55428ac0b81aSLi Xinhai return NULL; 55438ac0b81aSLi Xinhai /* must have a valid entry and size to go further */ 55449b19df29SPunit Agrawal 55459e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 55468ac0b81aSLi Xinhai /* must be pmd huge, non-present or none */ 55479e5fc74cSSteve Capper return (pte_t *)pmd; 55489e5fc74cSSteve Capper } 55499e5fc74cSSteve Capper 555061f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 555161f77edaSNaoya Horiguchi 555261f77edaSNaoya Horiguchi /* 555361f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 555461f77edaSNaoya Horiguchi * behavior. 555561f77edaSNaoya Horiguchi */ 555661f77edaSNaoya Horiguchi struct page * __weak 555761f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 555861f77edaSNaoya Horiguchi int write) 555961f77edaSNaoya Horiguchi { 556061f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 556161f77edaSNaoya Horiguchi } 556261f77edaSNaoya Horiguchi 556361f77edaSNaoya Horiguchi struct page * __weak 55644dc71451SAneesh Kumar K.V follow_huge_pd(struct vm_area_struct *vma, 55654dc71451SAneesh Kumar K.V unsigned long address, hugepd_t hpd, int flags, int pdshift) 55664dc71451SAneesh Kumar K.V { 55674dc71451SAneesh Kumar K.V WARN(1, "hugepd follow called with no support for hugepage directory format\n"); 55684dc71451SAneesh Kumar K.V return NULL; 55694dc71451SAneesh Kumar K.V } 55704dc71451SAneesh Kumar K.V 55714dc71451SAneesh Kumar K.V struct page * __weak 55729e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 5573e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 55749e5fc74cSSteve Capper { 5575e66f17ffSNaoya Horiguchi struct page *page = NULL; 5576e66f17ffSNaoya Horiguchi spinlock_t *ptl; 5577c9d398faSNaoya Horiguchi pte_t pte; 55783faa52c0SJohn Hubbard 55793faa52c0SJohn Hubbard /* FOLL_GET and FOLL_PIN are mutually exclusive. */ 55803faa52c0SJohn Hubbard if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == 55813faa52c0SJohn Hubbard (FOLL_PIN | FOLL_GET))) 55823faa52c0SJohn Hubbard return NULL; 55833faa52c0SJohn Hubbard 5584e66f17ffSNaoya Horiguchi retry: 5585e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 5586e66f17ffSNaoya Horiguchi spin_lock(ptl); 5587e66f17ffSNaoya Horiguchi /* 5588e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 5589e66f17ffSNaoya Horiguchi * unmapped from other threads. 5590e66f17ffSNaoya Horiguchi */ 5591e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 5592e66f17ffSNaoya Horiguchi goto out; 5593c9d398faSNaoya Horiguchi pte = huge_ptep_get((pte_t *)pmd); 5594c9d398faSNaoya Horiguchi if (pte_present(pte)) { 559597534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 55963faa52c0SJohn Hubbard /* 55973faa52c0SJohn Hubbard * try_grab_page() should always succeed here, because: a) we 55983faa52c0SJohn Hubbard * hold the pmd (ptl) lock, and b) we've just checked that the 55993faa52c0SJohn Hubbard * huge pmd (head) page is present in the page tables. The ptl 56003faa52c0SJohn Hubbard * prevents the head page and tail pages from being rearranged 56013faa52c0SJohn Hubbard * in any way. So this page must be available at this point, 56023faa52c0SJohn Hubbard * unless the page refcount overflowed: 56033faa52c0SJohn Hubbard */ 56043faa52c0SJohn Hubbard if (WARN_ON_ONCE(!try_grab_page(page, flags))) { 56053faa52c0SJohn Hubbard page = NULL; 56063faa52c0SJohn Hubbard goto out; 56073faa52c0SJohn Hubbard } 5608e66f17ffSNaoya Horiguchi } else { 5609c9d398faSNaoya Horiguchi if (is_hugetlb_entry_migration(pte)) { 5610e66f17ffSNaoya Horiguchi spin_unlock(ptl); 5611e66f17ffSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pmd, ptl); 5612e66f17ffSNaoya Horiguchi goto retry; 5613e66f17ffSNaoya Horiguchi } 5614e66f17ffSNaoya Horiguchi /* 5615e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 5616e66f17ffSNaoya Horiguchi * follow_page_mask(). 5617e66f17ffSNaoya Horiguchi */ 5618e66f17ffSNaoya Horiguchi } 5619e66f17ffSNaoya Horiguchi out: 5620e66f17ffSNaoya Horiguchi spin_unlock(ptl); 56219e5fc74cSSteve Capper return page; 56229e5fc74cSSteve Capper } 56239e5fc74cSSteve Capper 562461f77edaSNaoya Horiguchi struct page * __weak 56259e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 5626e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 56279e5fc74cSSteve Capper { 56283faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 5629e66f17ffSNaoya Horiguchi return NULL; 56309e5fc74cSSteve Capper 5631e66f17ffSNaoya Horiguchi return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 56329e5fc74cSSteve Capper } 56339e5fc74cSSteve Capper 5634faaa5b62SAnshuman Khandual struct page * __weak 5635faaa5b62SAnshuman Khandual follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) 5636faaa5b62SAnshuman Khandual { 56373faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 5638faaa5b62SAnshuman Khandual return NULL; 5639faaa5b62SAnshuman Khandual 5640faaa5b62SAnshuman Khandual return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); 5641faaa5b62SAnshuman Khandual } 5642faaa5b62SAnshuman Khandual 564331caf665SNaoya Horiguchi bool isolate_huge_page(struct page *page, struct list_head *list) 564431caf665SNaoya Horiguchi { 5645bcc54222SNaoya Horiguchi bool ret = true; 5646bcc54222SNaoya Horiguchi 5647*db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 56488f251a3dSMike Kravetz if (!PageHeadHuge(page) || 56498f251a3dSMike Kravetz !HPageMigratable(page) || 56500eb2df2bSMuchun Song !get_page_unless_zero(page)) { 5651bcc54222SNaoya Horiguchi ret = false; 5652bcc54222SNaoya Horiguchi goto unlock; 5653bcc54222SNaoya Horiguchi } 56548f251a3dSMike Kravetz ClearHPageMigratable(page); 565531caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 5656bcc54222SNaoya Horiguchi unlock: 5657*db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 5658bcc54222SNaoya Horiguchi return ret; 565931caf665SNaoya Horiguchi } 566031caf665SNaoya Horiguchi 566131caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 566231caf665SNaoya Horiguchi { 5663*db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 56648f251a3dSMike Kravetz SetHPageMigratable(page); 566531caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 5666*db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 566731caf665SNaoya Horiguchi put_page(page); 566831caf665SNaoya Horiguchi } 5669ab5ac90aSMichal Hocko 5670ab5ac90aSMichal Hocko void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) 5671ab5ac90aSMichal Hocko { 5672ab5ac90aSMichal Hocko struct hstate *h = page_hstate(oldpage); 5673ab5ac90aSMichal Hocko 5674ab5ac90aSMichal Hocko hugetlb_cgroup_migrate(oldpage, newpage); 5675ab5ac90aSMichal Hocko set_page_owner_migrate_reason(newpage, reason); 5676ab5ac90aSMichal Hocko 5677ab5ac90aSMichal Hocko /* 5678ab5ac90aSMichal Hocko * transfer temporary state of the new huge page. This is 5679ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 5680ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 5681ab5ac90aSMichal Hocko * the temporary status. 5682ab5ac90aSMichal Hocko * 5683ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 5684ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 5685ab5ac90aSMichal Hocko * the per-node's. 5686ab5ac90aSMichal Hocko */ 56879157c311SMike Kravetz if (HPageTemporary(newpage)) { 5688ab5ac90aSMichal Hocko int old_nid = page_to_nid(oldpage); 5689ab5ac90aSMichal Hocko int new_nid = page_to_nid(newpage); 5690ab5ac90aSMichal Hocko 56919157c311SMike Kravetz SetHPageTemporary(oldpage); 56929157c311SMike Kravetz ClearHPageTemporary(newpage); 5693ab5ac90aSMichal Hocko 56945af1ab1dSMiaohe Lin /* 56955af1ab1dSMiaohe Lin * There is no need to transfer the per-node surplus state 56965af1ab1dSMiaohe Lin * when we do not cross the node. 56975af1ab1dSMiaohe Lin */ 56985af1ab1dSMiaohe Lin if (new_nid == old_nid) 56995af1ab1dSMiaohe Lin return; 5700*db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 5701ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 5702ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 5703ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 5704ab5ac90aSMichal Hocko } 5705*db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 5706ab5ac90aSMichal Hocko } 5707ab5ac90aSMichal Hocko } 5708cf11e85fSRoman Gushchin 57096dfeaff9SPeter Xu /* 57106dfeaff9SPeter Xu * This function will unconditionally remove all the shared pmd pgtable entries 57116dfeaff9SPeter Xu * within the specific vma for a hugetlbfs memory range. 57126dfeaff9SPeter Xu */ 57136dfeaff9SPeter Xu void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) 57146dfeaff9SPeter Xu { 57156dfeaff9SPeter Xu struct hstate *h = hstate_vma(vma); 57166dfeaff9SPeter Xu unsigned long sz = huge_page_size(h); 57176dfeaff9SPeter Xu struct mm_struct *mm = vma->vm_mm; 57186dfeaff9SPeter Xu struct mmu_notifier_range range; 57196dfeaff9SPeter Xu unsigned long address, start, end; 57206dfeaff9SPeter Xu spinlock_t *ptl; 57216dfeaff9SPeter Xu pte_t *ptep; 57226dfeaff9SPeter Xu 57236dfeaff9SPeter Xu if (!(vma->vm_flags & VM_MAYSHARE)) 57246dfeaff9SPeter Xu return; 57256dfeaff9SPeter Xu 57266dfeaff9SPeter Xu start = ALIGN(vma->vm_start, PUD_SIZE); 57276dfeaff9SPeter Xu end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 57286dfeaff9SPeter Xu 57296dfeaff9SPeter Xu if (start >= end) 57306dfeaff9SPeter Xu return; 57316dfeaff9SPeter Xu 57326dfeaff9SPeter Xu /* 57336dfeaff9SPeter Xu * No need to call adjust_range_if_pmd_sharing_possible(), because 57346dfeaff9SPeter Xu * we have already done the PUD_SIZE alignment. 57356dfeaff9SPeter Xu */ 57366dfeaff9SPeter Xu mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, 57376dfeaff9SPeter Xu start, end); 57386dfeaff9SPeter Xu mmu_notifier_invalidate_range_start(&range); 57396dfeaff9SPeter Xu i_mmap_lock_write(vma->vm_file->f_mapping); 57406dfeaff9SPeter Xu for (address = start; address < end; address += PUD_SIZE) { 57416dfeaff9SPeter Xu unsigned long tmp = address; 57426dfeaff9SPeter Xu 57436dfeaff9SPeter Xu ptep = huge_pte_offset(mm, address, sz); 57446dfeaff9SPeter Xu if (!ptep) 57456dfeaff9SPeter Xu continue; 57466dfeaff9SPeter Xu ptl = huge_pte_lock(h, mm, ptep); 57476dfeaff9SPeter Xu /* We don't want 'address' to be changed */ 57486dfeaff9SPeter Xu huge_pmd_unshare(mm, vma, &tmp, ptep); 57496dfeaff9SPeter Xu spin_unlock(ptl); 57506dfeaff9SPeter Xu } 57516dfeaff9SPeter Xu flush_hugetlb_tlb_range(vma, start, end); 57526dfeaff9SPeter Xu i_mmap_unlock_write(vma->vm_file->f_mapping); 57536dfeaff9SPeter Xu /* 57546dfeaff9SPeter Xu * No need to call mmu_notifier_invalidate_range(), see 57556dfeaff9SPeter Xu * Documentation/vm/mmu_notifier.rst. 57566dfeaff9SPeter Xu */ 57576dfeaff9SPeter Xu mmu_notifier_invalidate_range_end(&range); 57586dfeaff9SPeter Xu } 57596dfeaff9SPeter Xu 5760cf11e85fSRoman Gushchin #ifdef CONFIG_CMA 5761cf11e85fSRoman Gushchin static bool cma_reserve_called __initdata; 5762cf11e85fSRoman Gushchin 5763cf11e85fSRoman Gushchin static int __init cmdline_parse_hugetlb_cma(char *p) 5764cf11e85fSRoman Gushchin { 5765cf11e85fSRoman Gushchin hugetlb_cma_size = memparse(p, &p); 5766cf11e85fSRoman Gushchin return 0; 5767cf11e85fSRoman Gushchin } 5768cf11e85fSRoman Gushchin 5769cf11e85fSRoman Gushchin early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); 5770cf11e85fSRoman Gushchin 5771cf11e85fSRoman Gushchin void __init hugetlb_cma_reserve(int order) 5772cf11e85fSRoman Gushchin { 5773cf11e85fSRoman Gushchin unsigned long size, reserved, per_node; 5774cf11e85fSRoman Gushchin int nid; 5775cf11e85fSRoman Gushchin 5776cf11e85fSRoman Gushchin cma_reserve_called = true; 5777cf11e85fSRoman Gushchin 5778cf11e85fSRoman Gushchin if (!hugetlb_cma_size) 5779cf11e85fSRoman Gushchin return; 5780cf11e85fSRoman Gushchin 5781cf11e85fSRoman Gushchin if (hugetlb_cma_size < (PAGE_SIZE << order)) { 5782cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", 5783cf11e85fSRoman Gushchin (PAGE_SIZE << order) / SZ_1M); 5784cf11e85fSRoman Gushchin return; 5785cf11e85fSRoman Gushchin } 5786cf11e85fSRoman Gushchin 5787cf11e85fSRoman Gushchin /* 5788cf11e85fSRoman Gushchin * If 3 GB area is requested on a machine with 4 numa nodes, 5789cf11e85fSRoman Gushchin * let's allocate 1 GB on first three nodes and ignore the last one. 5790cf11e85fSRoman Gushchin */ 5791cf11e85fSRoman Gushchin per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); 5792cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", 5793cf11e85fSRoman Gushchin hugetlb_cma_size / SZ_1M, per_node / SZ_1M); 5794cf11e85fSRoman Gushchin 5795cf11e85fSRoman Gushchin reserved = 0; 5796cf11e85fSRoman Gushchin for_each_node_state(nid, N_ONLINE) { 5797cf11e85fSRoman Gushchin int res; 57982281f797SBarry Song char name[CMA_MAX_NAME]; 5799cf11e85fSRoman Gushchin 5800cf11e85fSRoman Gushchin size = min(per_node, hugetlb_cma_size - reserved); 5801cf11e85fSRoman Gushchin size = round_up(size, PAGE_SIZE << order); 5802cf11e85fSRoman Gushchin 58032281f797SBarry Song snprintf(name, sizeof(name), "hugetlb%d", nid); 5804cf11e85fSRoman Gushchin res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order, 580529d0f41dSBarry Song 0, false, name, 5806cf11e85fSRoman Gushchin &hugetlb_cma[nid], nid); 5807cf11e85fSRoman Gushchin if (res) { 5808cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: reservation failed: err %d, node %d", 5809cf11e85fSRoman Gushchin res, nid); 5810cf11e85fSRoman Gushchin continue; 5811cf11e85fSRoman Gushchin } 5812cf11e85fSRoman Gushchin 5813cf11e85fSRoman Gushchin reserved += size; 5814cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", 5815cf11e85fSRoman Gushchin size / SZ_1M, nid); 5816cf11e85fSRoman Gushchin 5817cf11e85fSRoman Gushchin if (reserved >= hugetlb_cma_size) 5818cf11e85fSRoman Gushchin break; 5819cf11e85fSRoman Gushchin } 5820cf11e85fSRoman Gushchin } 5821cf11e85fSRoman Gushchin 5822cf11e85fSRoman Gushchin void __init hugetlb_cma_check(void) 5823cf11e85fSRoman Gushchin { 5824cf11e85fSRoman Gushchin if (!hugetlb_cma_size || cma_reserve_called) 5825cf11e85fSRoman Gushchin return; 5826cf11e85fSRoman Gushchin 5827cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); 5828cf11e85fSRoman Gushchin } 5829cf11e85fSRoman Gushchin 5830cf11e85fSRoman Gushchin #endif /* CONFIG_CMA */ 5831