1457c8996SThomas Gleixner // SPDX-License-Identifier: GPL-2.0-only 21da177e4SLinus Torvalds /* 31da177e4SLinus Torvalds * Generic hugetlb support. 46d49e352SNadia Yvette Chambers * (C) Nadia Yvette Chambers, April 2004 51da177e4SLinus Torvalds */ 61da177e4SLinus Torvalds #include <linux/list.h> 71da177e4SLinus Torvalds #include <linux/init.h> 81da177e4SLinus Torvalds #include <linux/mm.h> 9e1759c21SAlexey Dobriyan #include <linux/seq_file.h> 101da177e4SLinus Torvalds #include <linux/sysctl.h> 111da177e4SLinus Torvalds #include <linux/highmem.h> 12cddb8a5cSAndrea Arcangeli #include <linux/mmu_notifier.h> 131da177e4SLinus Torvalds #include <linux/nodemask.h> 1463551ae0SDavid Gibson #include <linux/pagemap.h> 155da7ca86SChristoph Lameter #include <linux/mempolicy.h> 163b32123dSGideon Israel Dsouza #include <linux/compiler.h> 17aea47ff3SChristoph Lameter #include <linux/cpuset.h> 183935baa9SDavid Gibson #include <linux/mutex.h> 1997ad1087SMike Rapoport #include <linux/memblock.h> 20a3437870SNishanth Aravamudan #include <linux/sysfs.h> 215a0e3ad6STejun Heo #include <linux/slab.h> 22bbe88753SJoonsoo Kim #include <linux/sched/mm.h> 2363489f8eSMike Kravetz #include <linux/mmdebug.h> 24174cd4b1SIngo Molnar #include <linux/sched/signal.h> 250fe6e20bSNaoya Horiguchi #include <linux/rmap.h> 26c6247f72SMatthew Wilcox #include <linux/string_helpers.h> 27fd6a03edSNaoya Horiguchi #include <linux/swap.h> 28fd6a03edSNaoya Horiguchi #include <linux/swapops.h> 298382d914SDavidlohr Bueso #include <linux/jhash.h> 3098fa15f3SAnshuman Khandual #include <linux/numa.h> 31c77c0a8aSWaiman Long #include <linux/llist.h> 32cf11e85fSRoman Gushchin #include <linux/cma.h> 338cc5fcbbSMina Almasry #include <linux/migrate.h> 34f9317f77SMike Kravetz #include <linux/nospec.h> 35d6606683SLinus Torvalds 3663551ae0SDavid Gibson #include <asm/page.h> 37ca15ca40SMike Rapoport #include <asm/pgalloc.h> 3824669e58SAneesh Kumar K.V #include <asm/tlb.h> 3963551ae0SDavid Gibson 4024669e58SAneesh Kumar K.V #include <linux/io.h> 4163551ae0SDavid Gibson #include <linux/hugetlb.h> 429dd540e2SAneesh Kumar K.V #include <linux/hugetlb_cgroup.h> 439a305230SLee Schermerhorn #include <linux/node.h> 44ab5ac90aSMichal Hocko #include <linux/page_owner.h> 457835e98bSNick Piggin #include "internal.h" 46f41f2ed4SMuchun Song #include "hugetlb_vmemmap.h" 471da177e4SLinus Torvalds 48c3f38a38SAneesh Kumar K.V int hugetlb_max_hstate __read_mostly; 49e5ff2159SAndi Kleen unsigned int default_hstate_idx; 50e5ff2159SAndi Kleen struct hstate hstates[HUGE_MAX_HSTATE]; 51cf11e85fSRoman Gushchin 52dbda8feaSBarry Song #ifdef CONFIG_CMA 53cf11e85fSRoman Gushchin static struct cma *hugetlb_cma[MAX_NUMNODES]; 5438e719abSBaolin Wang static unsigned long hugetlb_cma_size_in_node[MAX_NUMNODES] __initdata; 55a01f4390SMike Kravetz static bool hugetlb_cma_page(struct page *page, unsigned int order) 56a01f4390SMike Kravetz { 57a01f4390SMike Kravetz return cma_pages_valid(hugetlb_cma[page_to_nid(page)], page, 58a01f4390SMike Kravetz 1 << order); 59a01f4390SMike Kravetz } 60a01f4390SMike Kravetz #else 61a01f4390SMike Kravetz static bool hugetlb_cma_page(struct page *page, unsigned int order) 62a01f4390SMike Kravetz { 63a01f4390SMike Kravetz return false; 64a01f4390SMike Kravetz } 65dbda8feaSBarry Song #endif 66dbda8feaSBarry Song static unsigned long hugetlb_cma_size __initdata; 67cf11e85fSRoman Gushchin 68641844f5SNaoya Horiguchi /* 69641844f5SNaoya Horiguchi * Minimum page order among possible hugepage sizes, set to a proper value 70641844f5SNaoya Horiguchi * at boot time. 71641844f5SNaoya Horiguchi */ 72641844f5SNaoya Horiguchi static unsigned int minimum_order __read_mostly = UINT_MAX; 73e5ff2159SAndi Kleen 7453ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 7553ba51d2SJon Tollefson 76e5ff2159SAndi Kleen /* for command line parsing */ 77e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 78e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 799fee021dSVaishali Thakkar static bool __initdata parsed_valid_hugepagesz = true; 80282f4214SMike Kravetz static bool __initdata parsed_default_hugepagesz; 81b5389086SZhenguo Yao static unsigned int default_hugepages_in_node[MAX_NUMNODES] __initdata; 82e5ff2159SAndi Kleen 833935baa9SDavid Gibson /* 8431caf665SNaoya Horiguchi * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, 8531caf665SNaoya Horiguchi * free_huge_pages, and surplus_huge_pages. 863935baa9SDavid Gibson */ 87c3f38a38SAneesh Kumar K.V DEFINE_SPINLOCK(hugetlb_lock); 880bd0f9fbSEric Paris 898382d914SDavidlohr Bueso /* 908382d914SDavidlohr Bueso * Serializes faults on the same logical page. This is used to 918382d914SDavidlohr Bueso * prevent spurious OOMs when the hugepage pool is fully utilized. 928382d914SDavidlohr Bueso */ 938382d914SDavidlohr Bueso static int num_fault_mutexes; 94c672c7f2SMike Kravetz struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; 958382d914SDavidlohr Bueso 967ca02d0aSMike Kravetz /* Forward declaration */ 977ca02d0aSMike Kravetz static int hugetlb_acct_memory(struct hstate *h, long delta); 987ca02d0aSMike Kravetz 991d88433bSMiaohe Lin static inline bool subpool_is_free(struct hugepage_subpool *spool) 1001d88433bSMiaohe Lin { 1011d88433bSMiaohe Lin if (spool->count) 1021d88433bSMiaohe Lin return false; 1031d88433bSMiaohe Lin if (spool->max_hpages != -1) 1041d88433bSMiaohe Lin return spool->used_hpages == 0; 1051d88433bSMiaohe Lin if (spool->min_hpages != -1) 1061d88433bSMiaohe Lin return spool->rsv_hpages == spool->min_hpages; 1071d88433bSMiaohe Lin 1081d88433bSMiaohe Lin return true; 1091d88433bSMiaohe Lin } 1101d88433bSMiaohe Lin 111db71ef79SMike Kravetz static inline void unlock_or_release_subpool(struct hugepage_subpool *spool, 112db71ef79SMike Kravetz unsigned long irq_flags) 11390481622SDavid Gibson { 114db71ef79SMike Kravetz spin_unlock_irqrestore(&spool->lock, irq_flags); 11590481622SDavid Gibson 11690481622SDavid Gibson /* If no pages are used, and no other handles to the subpool 1177c8de358SEthon Paul * remain, give up any reservations based on minimum size and 1187ca02d0aSMike Kravetz * free the subpool */ 1191d88433bSMiaohe Lin if (subpool_is_free(spool)) { 1207ca02d0aSMike Kravetz if (spool->min_hpages != -1) 1217ca02d0aSMike Kravetz hugetlb_acct_memory(spool->hstate, 1227ca02d0aSMike Kravetz -spool->min_hpages); 12390481622SDavid Gibson kfree(spool); 12490481622SDavid Gibson } 1257ca02d0aSMike Kravetz } 12690481622SDavid Gibson 1277ca02d0aSMike Kravetz struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, 1287ca02d0aSMike Kravetz long min_hpages) 12990481622SDavid Gibson { 13090481622SDavid Gibson struct hugepage_subpool *spool; 13190481622SDavid Gibson 132c6a91820SMike Kravetz spool = kzalloc(sizeof(*spool), GFP_KERNEL); 13390481622SDavid Gibson if (!spool) 13490481622SDavid Gibson return NULL; 13590481622SDavid Gibson 13690481622SDavid Gibson spin_lock_init(&spool->lock); 13790481622SDavid Gibson spool->count = 1; 1387ca02d0aSMike Kravetz spool->max_hpages = max_hpages; 1397ca02d0aSMike Kravetz spool->hstate = h; 1407ca02d0aSMike Kravetz spool->min_hpages = min_hpages; 1417ca02d0aSMike Kravetz 1427ca02d0aSMike Kravetz if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { 1437ca02d0aSMike Kravetz kfree(spool); 1447ca02d0aSMike Kravetz return NULL; 1457ca02d0aSMike Kravetz } 1467ca02d0aSMike Kravetz spool->rsv_hpages = min_hpages; 14790481622SDavid Gibson 14890481622SDavid Gibson return spool; 14990481622SDavid Gibson } 15090481622SDavid Gibson 15190481622SDavid Gibson void hugepage_put_subpool(struct hugepage_subpool *spool) 15290481622SDavid Gibson { 153db71ef79SMike Kravetz unsigned long flags; 154db71ef79SMike Kravetz 155db71ef79SMike Kravetz spin_lock_irqsave(&spool->lock, flags); 15690481622SDavid Gibson BUG_ON(!spool->count); 15790481622SDavid Gibson spool->count--; 158db71ef79SMike Kravetz unlock_or_release_subpool(spool, flags); 15990481622SDavid Gibson } 16090481622SDavid Gibson 1611c5ecae3SMike Kravetz /* 1621c5ecae3SMike Kravetz * Subpool accounting for allocating and reserving pages. 1631c5ecae3SMike Kravetz * Return -ENOMEM if there are not enough resources to satisfy the 1649e7ee400SRandy Dunlap * request. Otherwise, return the number of pages by which the 1651c5ecae3SMike Kravetz * global pools must be adjusted (upward). The returned value may 1661c5ecae3SMike Kravetz * only be different than the passed value (delta) in the case where 1677c8de358SEthon Paul * a subpool minimum size must be maintained. 1681c5ecae3SMike Kravetz */ 1691c5ecae3SMike Kravetz static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, 17090481622SDavid Gibson long delta) 17190481622SDavid Gibson { 1721c5ecae3SMike Kravetz long ret = delta; 17390481622SDavid Gibson 17490481622SDavid Gibson if (!spool) 1751c5ecae3SMike Kravetz return ret; 17690481622SDavid Gibson 177db71ef79SMike Kravetz spin_lock_irq(&spool->lock); 17890481622SDavid Gibson 1791c5ecae3SMike Kravetz if (spool->max_hpages != -1) { /* maximum size accounting */ 1801c5ecae3SMike Kravetz if ((spool->used_hpages + delta) <= spool->max_hpages) 1811c5ecae3SMike Kravetz spool->used_hpages += delta; 1821c5ecae3SMike Kravetz else { 1831c5ecae3SMike Kravetz ret = -ENOMEM; 1841c5ecae3SMike Kravetz goto unlock_ret; 1851c5ecae3SMike Kravetz } 1861c5ecae3SMike Kravetz } 1871c5ecae3SMike Kravetz 18809a95e29SMike Kravetz /* minimum size accounting */ 18909a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->rsv_hpages) { 1901c5ecae3SMike Kravetz if (delta > spool->rsv_hpages) { 1911c5ecae3SMike Kravetz /* 1921c5ecae3SMike Kravetz * Asking for more reserves than those already taken on 1931c5ecae3SMike Kravetz * behalf of subpool. Return difference. 1941c5ecae3SMike Kravetz */ 1951c5ecae3SMike Kravetz ret = delta - spool->rsv_hpages; 1961c5ecae3SMike Kravetz spool->rsv_hpages = 0; 1971c5ecae3SMike Kravetz } else { 1981c5ecae3SMike Kravetz ret = 0; /* reserves already accounted for */ 1991c5ecae3SMike Kravetz spool->rsv_hpages -= delta; 2001c5ecae3SMike Kravetz } 2011c5ecae3SMike Kravetz } 2021c5ecae3SMike Kravetz 2031c5ecae3SMike Kravetz unlock_ret: 204db71ef79SMike Kravetz spin_unlock_irq(&spool->lock); 20590481622SDavid Gibson return ret; 20690481622SDavid Gibson } 20790481622SDavid Gibson 2081c5ecae3SMike Kravetz /* 2091c5ecae3SMike Kravetz * Subpool accounting for freeing and unreserving pages. 2101c5ecae3SMike Kravetz * Return the number of global page reservations that must be dropped. 2111c5ecae3SMike Kravetz * The return value may only be different than the passed value (delta) 2121c5ecae3SMike Kravetz * in the case where a subpool minimum size must be maintained. 2131c5ecae3SMike Kravetz */ 2141c5ecae3SMike Kravetz static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, 21590481622SDavid Gibson long delta) 21690481622SDavid Gibson { 2171c5ecae3SMike Kravetz long ret = delta; 218db71ef79SMike Kravetz unsigned long flags; 2191c5ecae3SMike Kravetz 22090481622SDavid Gibson if (!spool) 2211c5ecae3SMike Kravetz return delta; 22290481622SDavid Gibson 223db71ef79SMike Kravetz spin_lock_irqsave(&spool->lock, flags); 2241c5ecae3SMike Kravetz 2251c5ecae3SMike Kravetz if (spool->max_hpages != -1) /* maximum size accounting */ 22690481622SDavid Gibson spool->used_hpages -= delta; 2271c5ecae3SMike Kravetz 22809a95e29SMike Kravetz /* minimum size accounting */ 22909a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { 2301c5ecae3SMike Kravetz if (spool->rsv_hpages + delta <= spool->min_hpages) 2311c5ecae3SMike Kravetz ret = 0; 2321c5ecae3SMike Kravetz else 2331c5ecae3SMike Kravetz ret = spool->rsv_hpages + delta - spool->min_hpages; 2341c5ecae3SMike Kravetz 2351c5ecae3SMike Kravetz spool->rsv_hpages += delta; 2361c5ecae3SMike Kravetz if (spool->rsv_hpages > spool->min_hpages) 2371c5ecae3SMike Kravetz spool->rsv_hpages = spool->min_hpages; 2381c5ecae3SMike Kravetz } 2391c5ecae3SMike Kravetz 2401c5ecae3SMike Kravetz /* 2411c5ecae3SMike Kravetz * If hugetlbfs_put_super couldn't free spool due to an outstanding 2421c5ecae3SMike Kravetz * quota reference, free it now. 2431c5ecae3SMike Kravetz */ 244db71ef79SMike Kravetz unlock_or_release_subpool(spool, flags); 2451c5ecae3SMike Kravetz 2461c5ecae3SMike Kravetz return ret; 24790481622SDavid Gibson } 24890481622SDavid Gibson 24990481622SDavid Gibson static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 25090481622SDavid Gibson { 25190481622SDavid Gibson return HUGETLBFS_SB(inode->i_sb)->spool; 25290481622SDavid Gibson } 25390481622SDavid Gibson 25490481622SDavid Gibson static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 25590481622SDavid Gibson { 256496ad9aaSAl Viro return subpool_inode(file_inode(vma->vm_file)); 25790481622SDavid Gibson } 25890481622SDavid Gibson 2590db9d74eSMina Almasry /* Helper that removes a struct file_region from the resv_map cache and returns 2600db9d74eSMina Almasry * it for use. 2610db9d74eSMina Almasry */ 2620db9d74eSMina Almasry static struct file_region * 2630db9d74eSMina Almasry get_file_region_entry_from_cache(struct resv_map *resv, long from, long to) 2640db9d74eSMina Almasry { 2650db9d74eSMina Almasry struct file_region *nrg = NULL; 2660db9d74eSMina Almasry 2670db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count <= 0); 2680db9d74eSMina Almasry 2690db9d74eSMina Almasry resv->region_cache_count--; 2700db9d74eSMina Almasry nrg = list_first_entry(&resv->region_cache, struct file_region, link); 2710db9d74eSMina Almasry list_del(&nrg->link); 2720db9d74eSMina Almasry 2730db9d74eSMina Almasry nrg->from = from; 2740db9d74eSMina Almasry nrg->to = to; 2750db9d74eSMina Almasry 2760db9d74eSMina Almasry return nrg; 2770db9d74eSMina Almasry } 2780db9d74eSMina Almasry 279075a61d0SMina Almasry static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, 280075a61d0SMina Almasry struct file_region *rg) 281075a61d0SMina Almasry { 282075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 283075a61d0SMina Almasry nrg->reservation_counter = rg->reservation_counter; 284075a61d0SMina Almasry nrg->css = rg->css; 285075a61d0SMina Almasry if (rg->css) 286075a61d0SMina Almasry css_get(rg->css); 287075a61d0SMina Almasry #endif 288075a61d0SMina Almasry } 289075a61d0SMina Almasry 290075a61d0SMina Almasry /* Helper that records hugetlb_cgroup uncharge info. */ 291075a61d0SMina Almasry static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, 292075a61d0SMina Almasry struct hstate *h, 293075a61d0SMina Almasry struct resv_map *resv, 294075a61d0SMina Almasry struct file_region *nrg) 295075a61d0SMina Almasry { 296075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 297075a61d0SMina Almasry if (h_cg) { 298075a61d0SMina Almasry nrg->reservation_counter = 299075a61d0SMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 300075a61d0SMina Almasry nrg->css = &h_cg->css; 301d85aecf2SMiaohe Lin /* 302d85aecf2SMiaohe Lin * The caller will hold exactly one h_cg->css reference for the 303d85aecf2SMiaohe Lin * whole contiguous reservation region. But this area might be 304d85aecf2SMiaohe Lin * scattered when there are already some file_regions reside in 305d85aecf2SMiaohe Lin * it. As a result, many file_regions may share only one css 306d85aecf2SMiaohe Lin * reference. In order to ensure that one file_region must hold 307d85aecf2SMiaohe Lin * exactly one h_cg->css reference, we should do css_get for 308d85aecf2SMiaohe Lin * each file_region and leave the reference held by caller 309d85aecf2SMiaohe Lin * untouched. 310d85aecf2SMiaohe Lin */ 311d85aecf2SMiaohe Lin css_get(&h_cg->css); 312075a61d0SMina Almasry if (!resv->pages_per_hpage) 313075a61d0SMina Almasry resv->pages_per_hpage = pages_per_huge_page(h); 314075a61d0SMina Almasry /* pages_per_hpage should be the same for all entries in 315075a61d0SMina Almasry * a resv_map. 316075a61d0SMina Almasry */ 317075a61d0SMina Almasry VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); 318075a61d0SMina Almasry } else { 319075a61d0SMina Almasry nrg->reservation_counter = NULL; 320075a61d0SMina Almasry nrg->css = NULL; 321075a61d0SMina Almasry } 322075a61d0SMina Almasry #endif 323075a61d0SMina Almasry } 324075a61d0SMina Almasry 325d85aecf2SMiaohe Lin static void put_uncharge_info(struct file_region *rg) 326d85aecf2SMiaohe Lin { 327d85aecf2SMiaohe Lin #ifdef CONFIG_CGROUP_HUGETLB 328d85aecf2SMiaohe Lin if (rg->css) 329d85aecf2SMiaohe Lin css_put(rg->css); 330d85aecf2SMiaohe Lin #endif 331d85aecf2SMiaohe Lin } 332d85aecf2SMiaohe Lin 333a9b3f867SMina Almasry static bool has_same_uncharge_info(struct file_region *rg, 334a9b3f867SMina Almasry struct file_region *org) 335a9b3f867SMina Almasry { 336a9b3f867SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 3370739eb43SBaolin Wang return rg->reservation_counter == org->reservation_counter && 338a9b3f867SMina Almasry rg->css == org->css; 339a9b3f867SMina Almasry 340a9b3f867SMina Almasry #else 341a9b3f867SMina Almasry return true; 342a9b3f867SMina Almasry #endif 343a9b3f867SMina Almasry } 344a9b3f867SMina Almasry 345a9b3f867SMina Almasry static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) 346a9b3f867SMina Almasry { 347a9b3f867SMina Almasry struct file_region *nrg = NULL, *prg = NULL; 348a9b3f867SMina Almasry 349a9b3f867SMina Almasry prg = list_prev_entry(rg, link); 350a9b3f867SMina Almasry if (&prg->link != &resv->regions && prg->to == rg->from && 351a9b3f867SMina Almasry has_same_uncharge_info(prg, rg)) { 352a9b3f867SMina Almasry prg->to = rg->to; 353a9b3f867SMina Almasry 354a9b3f867SMina Almasry list_del(&rg->link); 355d85aecf2SMiaohe Lin put_uncharge_info(rg); 356a9b3f867SMina Almasry kfree(rg); 357a9b3f867SMina Almasry 3587db5e7b6SWei Yang rg = prg; 359a9b3f867SMina Almasry } 360a9b3f867SMina Almasry 361a9b3f867SMina Almasry nrg = list_next_entry(rg, link); 362a9b3f867SMina Almasry if (&nrg->link != &resv->regions && nrg->from == rg->to && 363a9b3f867SMina Almasry has_same_uncharge_info(nrg, rg)) { 364a9b3f867SMina Almasry nrg->from = rg->from; 365a9b3f867SMina Almasry 366a9b3f867SMina Almasry list_del(&rg->link); 367d85aecf2SMiaohe Lin put_uncharge_info(rg); 368a9b3f867SMina Almasry kfree(rg); 369a9b3f867SMina Almasry } 370a9b3f867SMina Almasry } 371a9b3f867SMina Almasry 3722103cf9cSPeter Xu static inline long 3732103cf9cSPeter Xu hugetlb_resv_map_add(struct resv_map *map, struct file_region *rg, long from, 3742103cf9cSPeter Xu long to, struct hstate *h, struct hugetlb_cgroup *cg, 3752103cf9cSPeter Xu long *regions_needed) 3762103cf9cSPeter Xu { 3772103cf9cSPeter Xu struct file_region *nrg; 3782103cf9cSPeter Xu 3792103cf9cSPeter Xu if (!regions_needed) { 3802103cf9cSPeter Xu nrg = get_file_region_entry_from_cache(map, from, to); 3812103cf9cSPeter Xu record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg); 3822103cf9cSPeter Xu list_add(&nrg->link, rg->link.prev); 3832103cf9cSPeter Xu coalesce_file_region(map, nrg); 3842103cf9cSPeter Xu } else 3852103cf9cSPeter Xu *regions_needed += 1; 3862103cf9cSPeter Xu 3872103cf9cSPeter Xu return to - from; 3882103cf9cSPeter Xu } 3892103cf9cSPeter Xu 390972a3da3SWei Yang /* 391972a3da3SWei Yang * Must be called with resv->lock held. 392972a3da3SWei Yang * 393972a3da3SWei Yang * Calling this with regions_needed != NULL will count the number of pages 394972a3da3SWei Yang * to be added but will not modify the linked list. And regions_needed will 395972a3da3SWei Yang * indicate the number of file_regions needed in the cache to carry out to add 396972a3da3SWei Yang * the regions for this range. 397d75c6af9SMina Almasry */ 398d75c6af9SMina Almasry static long add_reservation_in_range(struct resv_map *resv, long f, long t, 399075a61d0SMina Almasry struct hugetlb_cgroup *h_cg, 400972a3da3SWei Yang struct hstate *h, long *regions_needed) 401d75c6af9SMina Almasry { 4020db9d74eSMina Almasry long add = 0; 403d75c6af9SMina Almasry struct list_head *head = &resv->regions; 4040db9d74eSMina Almasry long last_accounted_offset = f; 4052103cf9cSPeter Xu struct file_region *rg = NULL, *trg = NULL; 406d75c6af9SMina Almasry 4070db9d74eSMina Almasry if (regions_needed) 4080db9d74eSMina Almasry *regions_needed = 0; 409d75c6af9SMina Almasry 4100db9d74eSMina Almasry /* In this loop, we essentially handle an entry for the range 4110db9d74eSMina Almasry * [last_accounted_offset, rg->from), at every iteration, with some 4120db9d74eSMina Almasry * bounds checking. 4130db9d74eSMina Almasry */ 4140db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, head, link) { 4150db9d74eSMina Almasry /* Skip irrelevant regions that start before our range. */ 4160db9d74eSMina Almasry if (rg->from < f) { 4170db9d74eSMina Almasry /* If this region ends after the last accounted offset, 4180db9d74eSMina Almasry * then we need to update last_accounted_offset. 4190db9d74eSMina Almasry */ 4200db9d74eSMina Almasry if (rg->to > last_accounted_offset) 4210db9d74eSMina Almasry last_accounted_offset = rg->to; 4220db9d74eSMina Almasry continue; 4230db9d74eSMina Almasry } 424d75c6af9SMina Almasry 4250db9d74eSMina Almasry /* When we find a region that starts beyond our range, we've 4260db9d74eSMina Almasry * finished. 4270db9d74eSMina Almasry */ 428ca7e0457SPeter Xu if (rg->from >= t) 429d75c6af9SMina Almasry break; 430d75c6af9SMina Almasry 4310db9d74eSMina Almasry /* Add an entry for last_accounted_offset -> rg->from, and 4320db9d74eSMina Almasry * update last_accounted_offset. 433d75c6af9SMina Almasry */ 4342103cf9cSPeter Xu if (rg->from > last_accounted_offset) 4352103cf9cSPeter Xu add += hugetlb_resv_map_add(resv, rg, 4362103cf9cSPeter Xu last_accounted_offset, 4372103cf9cSPeter Xu rg->from, h, h_cg, 4382103cf9cSPeter Xu regions_needed); 439d75c6af9SMina Almasry 4400db9d74eSMina Almasry last_accounted_offset = rg->to; 4410db9d74eSMina Almasry } 4420db9d74eSMina Almasry 4430db9d74eSMina Almasry /* Handle the case where our range extends beyond 4440db9d74eSMina Almasry * last_accounted_offset. 4450db9d74eSMina Almasry */ 4462103cf9cSPeter Xu if (last_accounted_offset < t) 4472103cf9cSPeter Xu add += hugetlb_resv_map_add(resv, rg, last_accounted_offset, 4482103cf9cSPeter Xu t, h, h_cg, regions_needed); 4490db9d74eSMina Almasry 4500db9d74eSMina Almasry return add; 4510db9d74eSMina Almasry } 4520db9d74eSMina Almasry 4530db9d74eSMina Almasry /* Must be called with resv->lock acquired. Will drop lock to allocate entries. 4540db9d74eSMina Almasry */ 4550db9d74eSMina Almasry static int allocate_file_region_entries(struct resv_map *resv, 4560db9d74eSMina Almasry int regions_needed) 4570db9d74eSMina Almasry __must_hold(&resv->lock) 4580db9d74eSMina Almasry { 4590db9d74eSMina Almasry struct list_head allocated_regions; 4600db9d74eSMina Almasry int to_allocate = 0, i = 0; 4610db9d74eSMina Almasry struct file_region *trg = NULL, *rg = NULL; 4620db9d74eSMina Almasry 4630db9d74eSMina Almasry VM_BUG_ON(regions_needed < 0); 4640db9d74eSMina Almasry 4650db9d74eSMina Almasry INIT_LIST_HEAD(&allocated_regions); 4660db9d74eSMina Almasry 4670db9d74eSMina Almasry /* 4680db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 4690db9d74eSMina Almasry * the number of in progress add operations plus regions_needed. 4700db9d74eSMina Almasry * 4710db9d74eSMina Almasry * This is a while loop because when we drop the lock, some other call 4720db9d74eSMina Almasry * to region_add or region_del may have consumed some region_entries, 4730db9d74eSMina Almasry * so we keep looping here until we finally have enough entries for 4740db9d74eSMina Almasry * (adds_in_progress + regions_needed). 4750db9d74eSMina Almasry */ 4760db9d74eSMina Almasry while (resv->region_cache_count < 4770db9d74eSMina Almasry (resv->adds_in_progress + regions_needed)) { 4780db9d74eSMina Almasry to_allocate = resv->adds_in_progress + regions_needed - 4790db9d74eSMina Almasry resv->region_cache_count; 4800db9d74eSMina Almasry 4810db9d74eSMina Almasry /* At this point, we should have enough entries in the cache 482f0953a1bSIngo Molnar * for all the existing adds_in_progress. We should only be 4830db9d74eSMina Almasry * needing to allocate for regions_needed. 4840db9d74eSMina Almasry */ 4850db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); 4860db9d74eSMina Almasry 4870db9d74eSMina Almasry spin_unlock(&resv->lock); 4880db9d74eSMina Almasry for (i = 0; i < to_allocate; i++) { 4890db9d74eSMina Almasry trg = kmalloc(sizeof(*trg), GFP_KERNEL); 4900db9d74eSMina Almasry if (!trg) 4910db9d74eSMina Almasry goto out_of_memory; 4920db9d74eSMina Almasry list_add(&trg->link, &allocated_regions); 4930db9d74eSMina Almasry } 4940db9d74eSMina Almasry 4950db9d74eSMina Almasry spin_lock(&resv->lock); 4960db9d74eSMina Almasry 497d3ec7b6eSWei Yang list_splice(&allocated_regions, &resv->region_cache); 498d3ec7b6eSWei Yang resv->region_cache_count += to_allocate; 4990db9d74eSMina Almasry } 5000db9d74eSMina Almasry 5010db9d74eSMina Almasry return 0; 5020db9d74eSMina Almasry 5030db9d74eSMina Almasry out_of_memory: 5040db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, &allocated_regions, link) { 505d75c6af9SMina Almasry list_del(&rg->link); 506d75c6af9SMina Almasry kfree(rg); 507d75c6af9SMina Almasry } 5080db9d74eSMina Almasry return -ENOMEM; 509d75c6af9SMina Almasry } 510d75c6af9SMina Almasry 5111dd308a7SMike Kravetz /* 5121dd308a7SMike Kravetz * Add the huge page range represented by [f, t) to the reserve 5130db9d74eSMina Almasry * map. Regions will be taken from the cache to fill in this range. 5140db9d74eSMina Almasry * Sufficient regions should exist in the cache due to the previous 5150db9d74eSMina Almasry * call to region_chg with the same range, but in some cases the cache will not 5160db9d74eSMina Almasry * have sufficient entries due to races with other code doing region_add or 5170db9d74eSMina Almasry * region_del. The extra needed entries will be allocated. 518cf3ad20bSMike Kravetz * 5190db9d74eSMina Almasry * regions_needed is the out value provided by a previous call to region_chg. 5200db9d74eSMina Almasry * 5210db9d74eSMina Almasry * Return the number of new huge pages added to the map. This number is greater 5220db9d74eSMina Almasry * than or equal to zero. If file_region entries needed to be allocated for 5237c8de358SEthon Paul * this operation and we were not able to allocate, it returns -ENOMEM. 5240db9d74eSMina Almasry * region_add of regions of length 1 never allocate file_regions and cannot 5250db9d74eSMina Almasry * fail; region_chg will always allocate at least 1 entry and a region_add for 5260db9d74eSMina Almasry * 1 page will only require at most 1 entry. 5271dd308a7SMike Kravetz */ 5280db9d74eSMina Almasry static long region_add(struct resv_map *resv, long f, long t, 529075a61d0SMina Almasry long in_regions_needed, struct hstate *h, 530075a61d0SMina Almasry struct hugetlb_cgroup *h_cg) 53196822904SAndy Whitcroft { 5320db9d74eSMina Almasry long add = 0, actual_regions_needed = 0; 53396822904SAndy Whitcroft 5347b24d861SDavidlohr Bueso spin_lock(&resv->lock); 5350db9d74eSMina Almasry retry: 5360db9d74eSMina Almasry 5370db9d74eSMina Almasry /* Count how many regions are actually needed to execute this add. */ 538972a3da3SWei Yang add_reservation_in_range(resv, f, t, NULL, NULL, 539972a3da3SWei Yang &actual_regions_needed); 54096822904SAndy Whitcroft 5415e911373SMike Kravetz /* 5420db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 5430db9d74eSMina Almasry * this add operation. Note that actual_regions_needed may be greater 5440db9d74eSMina Almasry * than in_regions_needed, as the resv_map may have been modified since 5450db9d74eSMina Almasry * the region_chg call. In this case, we need to make sure that we 5460db9d74eSMina Almasry * allocate extra entries, such that we have enough for all the 5470db9d74eSMina Almasry * existing adds_in_progress, plus the excess needed for this 5480db9d74eSMina Almasry * operation. 5495e911373SMike Kravetz */ 5500db9d74eSMina Almasry if (actual_regions_needed > in_regions_needed && 5510db9d74eSMina Almasry resv->region_cache_count < 5520db9d74eSMina Almasry resv->adds_in_progress + 5530db9d74eSMina Almasry (actual_regions_needed - in_regions_needed)) { 5540db9d74eSMina Almasry /* region_add operation of range 1 should never need to 5550db9d74eSMina Almasry * allocate file_region entries. 5560db9d74eSMina Almasry */ 5570db9d74eSMina Almasry VM_BUG_ON(t - f <= 1); 5585e911373SMike Kravetz 5590db9d74eSMina Almasry if (allocate_file_region_entries( 5600db9d74eSMina Almasry resv, actual_regions_needed - in_regions_needed)) { 5610db9d74eSMina Almasry return -ENOMEM; 5625e911373SMike Kravetz } 5635e911373SMike Kravetz 5640db9d74eSMina Almasry goto retry; 5650db9d74eSMina Almasry } 566cf3ad20bSMike Kravetz 567972a3da3SWei Yang add = add_reservation_in_range(resv, f, t, h_cg, h, NULL); 5680db9d74eSMina Almasry 5690db9d74eSMina Almasry resv->adds_in_progress -= in_regions_needed; 5700db9d74eSMina Almasry 5717b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 572cf3ad20bSMike Kravetz return add; 57396822904SAndy Whitcroft } 57496822904SAndy Whitcroft 5751dd308a7SMike Kravetz /* 5761dd308a7SMike Kravetz * Examine the existing reserve map and determine how many 5771dd308a7SMike Kravetz * huge pages in the specified range [f, t) are NOT currently 5781dd308a7SMike Kravetz * represented. This routine is called before a subsequent 5791dd308a7SMike Kravetz * call to region_add that will actually modify the reserve 5801dd308a7SMike Kravetz * map to add the specified range [f, t). region_chg does 5811dd308a7SMike Kravetz * not change the number of huge pages represented by the 5820db9d74eSMina Almasry * map. A number of new file_region structures is added to the cache as a 5830db9d74eSMina Almasry * placeholder, for the subsequent region_add call to use. At least 1 5840db9d74eSMina Almasry * file_region structure is added. 5850db9d74eSMina Almasry * 5860db9d74eSMina Almasry * out_regions_needed is the number of regions added to the 5870db9d74eSMina Almasry * resv->adds_in_progress. This value needs to be provided to a follow up call 5880db9d74eSMina Almasry * to region_add or region_abort for proper accounting. 5895e911373SMike Kravetz * 5905e911373SMike Kravetz * Returns the number of huge pages that need to be added to the existing 5915e911373SMike Kravetz * reservation map for the range [f, t). This number is greater or equal to 5925e911373SMike Kravetz * zero. -ENOMEM is returned if a new file_region structure or cache entry 5935e911373SMike Kravetz * is needed and can not be allocated. 5941dd308a7SMike Kravetz */ 5950db9d74eSMina Almasry static long region_chg(struct resv_map *resv, long f, long t, 5960db9d74eSMina Almasry long *out_regions_needed) 59796822904SAndy Whitcroft { 59896822904SAndy Whitcroft long chg = 0; 59996822904SAndy Whitcroft 6007b24d861SDavidlohr Bueso spin_lock(&resv->lock); 6015e911373SMike Kravetz 602972a3da3SWei Yang /* Count how many hugepages in this range are NOT represented. */ 603075a61d0SMina Almasry chg = add_reservation_in_range(resv, f, t, NULL, NULL, 604972a3da3SWei Yang out_regions_needed); 6055e911373SMike Kravetz 6060db9d74eSMina Almasry if (*out_regions_needed == 0) 6070db9d74eSMina Almasry *out_regions_needed = 1; 6085e911373SMike Kravetz 6090db9d74eSMina Almasry if (allocate_file_region_entries(resv, *out_regions_needed)) 6105e911373SMike Kravetz return -ENOMEM; 6115e911373SMike Kravetz 6120db9d74eSMina Almasry resv->adds_in_progress += *out_regions_needed; 61396822904SAndy Whitcroft 6147b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 61596822904SAndy Whitcroft return chg; 61696822904SAndy Whitcroft } 61796822904SAndy Whitcroft 6181dd308a7SMike Kravetz /* 6195e911373SMike Kravetz * Abort the in progress add operation. The adds_in_progress field 6205e911373SMike Kravetz * of the resv_map keeps track of the operations in progress between 6215e911373SMike Kravetz * calls to region_chg and region_add. Operations are sometimes 6225e911373SMike Kravetz * aborted after the call to region_chg. In such cases, region_abort 6230db9d74eSMina Almasry * is called to decrement the adds_in_progress counter. regions_needed 6240db9d74eSMina Almasry * is the value returned by the region_chg call, it is used to decrement 6250db9d74eSMina Almasry * the adds_in_progress counter. 6265e911373SMike Kravetz * 6275e911373SMike Kravetz * NOTE: The range arguments [f, t) are not needed or used in this 6285e911373SMike Kravetz * routine. They are kept to make reading the calling code easier as 6295e911373SMike Kravetz * arguments will match the associated region_chg call. 6305e911373SMike Kravetz */ 6310db9d74eSMina Almasry static void region_abort(struct resv_map *resv, long f, long t, 6320db9d74eSMina Almasry long regions_needed) 6335e911373SMike Kravetz { 6345e911373SMike Kravetz spin_lock(&resv->lock); 6355e911373SMike Kravetz VM_BUG_ON(!resv->region_cache_count); 6360db9d74eSMina Almasry resv->adds_in_progress -= regions_needed; 6375e911373SMike Kravetz spin_unlock(&resv->lock); 6385e911373SMike Kravetz } 6395e911373SMike Kravetz 6405e911373SMike Kravetz /* 641feba16e2SMike Kravetz * Delete the specified range [f, t) from the reserve map. If the 642feba16e2SMike Kravetz * t parameter is LONG_MAX, this indicates that ALL regions after f 643feba16e2SMike Kravetz * should be deleted. Locate the regions which intersect [f, t) 644feba16e2SMike Kravetz * and either trim, delete or split the existing regions. 645feba16e2SMike Kravetz * 646feba16e2SMike Kravetz * Returns the number of huge pages deleted from the reserve map. 647feba16e2SMike Kravetz * In the normal case, the return value is zero or more. In the 648feba16e2SMike Kravetz * case where a region must be split, a new region descriptor must 649feba16e2SMike Kravetz * be allocated. If the allocation fails, -ENOMEM will be returned. 650feba16e2SMike Kravetz * NOTE: If the parameter t == LONG_MAX, then we will never split 651feba16e2SMike Kravetz * a region and possibly return -ENOMEM. Callers specifying 652feba16e2SMike Kravetz * t == LONG_MAX do not need to check for -ENOMEM error. 6531dd308a7SMike Kravetz */ 654feba16e2SMike Kravetz static long region_del(struct resv_map *resv, long f, long t) 65596822904SAndy Whitcroft { 6561406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 65796822904SAndy Whitcroft struct file_region *rg, *trg; 658feba16e2SMike Kravetz struct file_region *nrg = NULL; 659feba16e2SMike Kravetz long del = 0; 66096822904SAndy Whitcroft 661feba16e2SMike Kravetz retry: 6627b24d861SDavidlohr Bueso spin_lock(&resv->lock); 663feba16e2SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 664dbe409e4SMike Kravetz /* 665dbe409e4SMike Kravetz * Skip regions before the range to be deleted. file_region 666dbe409e4SMike Kravetz * ranges are normally of the form [from, to). However, there 667dbe409e4SMike Kravetz * may be a "placeholder" entry in the map which is of the form 668dbe409e4SMike Kravetz * (from, to) with from == to. Check for placeholder entries 669dbe409e4SMike Kravetz * at the beginning of the range to be deleted. 670dbe409e4SMike Kravetz */ 671dbe409e4SMike Kravetz if (rg->to <= f && (rg->to != rg->from || rg->to != f)) 672feba16e2SMike Kravetz continue; 673dbe409e4SMike Kravetz 674feba16e2SMike Kravetz if (rg->from >= t) 67596822904SAndy Whitcroft break; 67696822904SAndy Whitcroft 677feba16e2SMike Kravetz if (f > rg->from && t < rg->to) { /* Must split region */ 678feba16e2SMike Kravetz /* 679feba16e2SMike Kravetz * Check for an entry in the cache before dropping 680feba16e2SMike Kravetz * lock and attempting allocation. 681feba16e2SMike Kravetz */ 682feba16e2SMike Kravetz if (!nrg && 683feba16e2SMike Kravetz resv->region_cache_count > resv->adds_in_progress) { 684feba16e2SMike Kravetz nrg = list_first_entry(&resv->region_cache, 685feba16e2SMike Kravetz struct file_region, 686feba16e2SMike Kravetz link); 687feba16e2SMike Kravetz list_del(&nrg->link); 688feba16e2SMike Kravetz resv->region_cache_count--; 68996822904SAndy Whitcroft } 69096822904SAndy Whitcroft 691feba16e2SMike Kravetz if (!nrg) { 692feba16e2SMike Kravetz spin_unlock(&resv->lock); 693feba16e2SMike Kravetz nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 694feba16e2SMike Kravetz if (!nrg) 695feba16e2SMike Kravetz return -ENOMEM; 696feba16e2SMike Kravetz goto retry; 697feba16e2SMike Kravetz } 698feba16e2SMike Kravetz 699feba16e2SMike Kravetz del += t - f; 70079aa925bSMike Kravetz hugetlb_cgroup_uncharge_file_region( 701d85aecf2SMiaohe Lin resv, rg, t - f, false); 702feba16e2SMike Kravetz 703feba16e2SMike Kravetz /* New entry for end of split region */ 704feba16e2SMike Kravetz nrg->from = t; 705feba16e2SMike Kravetz nrg->to = rg->to; 706075a61d0SMina Almasry 707075a61d0SMina Almasry copy_hugetlb_cgroup_uncharge_info(nrg, rg); 708075a61d0SMina Almasry 709feba16e2SMike Kravetz INIT_LIST_HEAD(&nrg->link); 710feba16e2SMike Kravetz 711feba16e2SMike Kravetz /* Original entry is trimmed */ 712feba16e2SMike Kravetz rg->to = f; 713feba16e2SMike Kravetz 714feba16e2SMike Kravetz list_add(&nrg->link, &rg->link); 715feba16e2SMike Kravetz nrg = NULL; 71696822904SAndy Whitcroft break; 717feba16e2SMike Kravetz } 718feba16e2SMike Kravetz 719feba16e2SMike Kravetz if (f <= rg->from && t >= rg->to) { /* Remove entire region */ 720feba16e2SMike Kravetz del += rg->to - rg->from; 721075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 722d85aecf2SMiaohe Lin rg->to - rg->from, true); 72396822904SAndy Whitcroft list_del(&rg->link); 72496822904SAndy Whitcroft kfree(rg); 725feba16e2SMike Kravetz continue; 72696822904SAndy Whitcroft } 7277b24d861SDavidlohr Bueso 728feba16e2SMike Kravetz if (f <= rg->from) { /* Trim beginning of region */ 729075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 730d85aecf2SMiaohe Lin t - rg->from, false); 731075a61d0SMina Almasry 73279aa925bSMike Kravetz del += t - rg->from; 73379aa925bSMike Kravetz rg->from = t; 73479aa925bSMike Kravetz } else { /* Trim end of region */ 735075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 736d85aecf2SMiaohe Lin rg->to - f, false); 73779aa925bSMike Kravetz 73879aa925bSMike Kravetz del += rg->to - f; 73979aa925bSMike Kravetz rg->to = f; 740feba16e2SMike Kravetz } 741feba16e2SMike Kravetz } 742feba16e2SMike Kravetz 7437b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 744feba16e2SMike Kravetz kfree(nrg); 745feba16e2SMike Kravetz return del; 74696822904SAndy Whitcroft } 74796822904SAndy Whitcroft 7481dd308a7SMike Kravetz /* 749b5cec28dSMike Kravetz * A rare out of memory error was encountered which prevented removal of 750b5cec28dSMike Kravetz * the reserve map region for a page. The huge page itself was free'ed 751b5cec28dSMike Kravetz * and removed from the page cache. This routine will adjust the subpool 752b5cec28dSMike Kravetz * usage count, and the global reserve count if needed. By incrementing 753b5cec28dSMike Kravetz * these counts, the reserve map entry which could not be deleted will 754b5cec28dSMike Kravetz * appear as a "reserved" entry instead of simply dangling with incorrect 755b5cec28dSMike Kravetz * counts. 756b5cec28dSMike Kravetz */ 75772e2936cSzhong jiang void hugetlb_fix_reserve_counts(struct inode *inode) 758b5cec28dSMike Kravetz { 759b5cec28dSMike Kravetz struct hugepage_subpool *spool = subpool_inode(inode); 760b5cec28dSMike Kravetz long rsv_adjust; 761da56388cSMiaohe Lin bool reserved = false; 762b5cec28dSMike Kravetz 763b5cec28dSMike Kravetz rsv_adjust = hugepage_subpool_get_pages(spool, 1); 764da56388cSMiaohe Lin if (rsv_adjust > 0) { 765b5cec28dSMike Kravetz struct hstate *h = hstate_inode(inode); 766b5cec28dSMike Kravetz 767da56388cSMiaohe Lin if (!hugetlb_acct_memory(h, 1)) 768da56388cSMiaohe Lin reserved = true; 769da56388cSMiaohe Lin } else if (!rsv_adjust) { 770da56388cSMiaohe Lin reserved = true; 771b5cec28dSMike Kravetz } 772da56388cSMiaohe Lin 773da56388cSMiaohe Lin if (!reserved) 774da56388cSMiaohe Lin pr_warn("hugetlb: Huge Page Reserved count may go negative.\n"); 775b5cec28dSMike Kravetz } 776b5cec28dSMike Kravetz 777b5cec28dSMike Kravetz /* 7781dd308a7SMike Kravetz * Count and return the number of huge pages in the reserve map 7791dd308a7SMike Kravetz * that intersect with the range [f, t). 7801dd308a7SMike Kravetz */ 7811406ec9bSJoonsoo Kim static long region_count(struct resv_map *resv, long f, long t) 78284afd99bSAndy Whitcroft { 7831406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 78484afd99bSAndy Whitcroft struct file_region *rg; 78584afd99bSAndy Whitcroft long chg = 0; 78684afd99bSAndy Whitcroft 7877b24d861SDavidlohr Bueso spin_lock(&resv->lock); 78884afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 78984afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 790f2135a4aSWang Sheng-Hui long seg_from; 791f2135a4aSWang Sheng-Hui long seg_to; 79284afd99bSAndy Whitcroft 79384afd99bSAndy Whitcroft if (rg->to <= f) 79484afd99bSAndy Whitcroft continue; 79584afd99bSAndy Whitcroft if (rg->from >= t) 79684afd99bSAndy Whitcroft break; 79784afd99bSAndy Whitcroft 79884afd99bSAndy Whitcroft seg_from = max(rg->from, f); 79984afd99bSAndy Whitcroft seg_to = min(rg->to, t); 80084afd99bSAndy Whitcroft 80184afd99bSAndy Whitcroft chg += seg_to - seg_from; 80284afd99bSAndy Whitcroft } 8037b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 80484afd99bSAndy Whitcroft 80584afd99bSAndy Whitcroft return chg; 80684afd99bSAndy Whitcroft } 80784afd99bSAndy Whitcroft 80896822904SAndy Whitcroft /* 809e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 810e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 811e7c4b0bfSAndy Whitcroft */ 812a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 813a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 814e7c4b0bfSAndy Whitcroft { 815a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 816a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 817e7c4b0bfSAndy Whitcroft } 818e7c4b0bfSAndy Whitcroft 8190fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 8200fe6e20bSNaoya Horiguchi unsigned long address) 8210fe6e20bSNaoya Horiguchi { 8220fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 8230fe6e20bSNaoya Horiguchi } 824dee41079SDan Williams EXPORT_SYMBOL_GPL(linear_hugepage_index); 8250fe6e20bSNaoya Horiguchi 82684afd99bSAndy Whitcroft /* 82708fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 82808fba699SMel Gorman * cases this will be same size as used by the page table entries. 82908fba699SMel Gorman */ 83008fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 83108fba699SMel Gorman { 83205ea8860SDan Williams if (vma->vm_ops && vma->vm_ops->pagesize) 83305ea8860SDan Williams return vma->vm_ops->pagesize(vma); 83408fba699SMel Gorman return PAGE_SIZE; 83508fba699SMel Gorman } 836f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 83708fba699SMel Gorman 83808fba699SMel Gorman /* 8393340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 8403340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 84109135cc5SDan Williams * architectures where it differs, an architecture-specific 'strong' 84209135cc5SDan Williams * version of this symbol is required. 8433340289dSMel Gorman */ 84409135cc5SDan Williams __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 8453340289dSMel Gorman { 8463340289dSMel Gorman return vma_kernel_pagesize(vma); 8473340289dSMel Gorman } 8483340289dSMel Gorman 8493340289dSMel Gorman /* 85084afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 85184afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 85284afd99bSAndy Whitcroft * alignment. 85384afd99bSAndy Whitcroft */ 85484afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 85584afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 85604f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 85784afd99bSAndy Whitcroft 858a1e78772SMel Gorman /* 859a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 860a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 861a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 862a1e78772SMel Gorman * 863a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 864a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 865a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 866a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 86784afd99bSAndy Whitcroft * 86884afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 86984afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 87084afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 87184afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 87284afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 87384afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 87484afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 87584afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 876a1e78772SMel Gorman */ 877e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 878e7c4b0bfSAndy Whitcroft { 879e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 880e7c4b0bfSAndy Whitcroft } 881e7c4b0bfSAndy Whitcroft 882e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 883e7c4b0bfSAndy Whitcroft unsigned long value) 884e7c4b0bfSAndy Whitcroft { 885e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 886e7c4b0bfSAndy Whitcroft } 887e7c4b0bfSAndy Whitcroft 888e9fe92aeSMina Almasry static void 889e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, 890e9fe92aeSMina Almasry struct hugetlb_cgroup *h_cg, 891e9fe92aeSMina Almasry struct hstate *h) 892e9fe92aeSMina Almasry { 893e9fe92aeSMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 894e9fe92aeSMina Almasry if (!h_cg || !h) { 895e9fe92aeSMina Almasry resv_map->reservation_counter = NULL; 896e9fe92aeSMina Almasry resv_map->pages_per_hpage = 0; 897e9fe92aeSMina Almasry resv_map->css = NULL; 898e9fe92aeSMina Almasry } else { 899e9fe92aeSMina Almasry resv_map->reservation_counter = 900e9fe92aeSMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 901e9fe92aeSMina Almasry resv_map->pages_per_hpage = pages_per_huge_page(h); 902e9fe92aeSMina Almasry resv_map->css = &h_cg->css; 903e9fe92aeSMina Almasry } 904e9fe92aeSMina Almasry #endif 905e9fe92aeSMina Almasry } 906e9fe92aeSMina Almasry 9079119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 90884afd99bSAndy Whitcroft { 90984afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 9105e911373SMike Kravetz struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 9115e911373SMike Kravetz 9125e911373SMike Kravetz if (!resv_map || !rg) { 9135e911373SMike Kravetz kfree(resv_map); 9145e911373SMike Kravetz kfree(rg); 91584afd99bSAndy Whitcroft return NULL; 9165e911373SMike Kravetz } 91784afd99bSAndy Whitcroft 91884afd99bSAndy Whitcroft kref_init(&resv_map->refs); 9197b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 92084afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 92184afd99bSAndy Whitcroft 9225e911373SMike Kravetz resv_map->adds_in_progress = 0; 923e9fe92aeSMina Almasry /* 924e9fe92aeSMina Almasry * Initialize these to 0. On shared mappings, 0's here indicate these 925e9fe92aeSMina Almasry * fields don't do cgroup accounting. On private mappings, these will be 926e9fe92aeSMina Almasry * re-initialized to the proper values, to indicate that hugetlb cgroup 927e9fe92aeSMina Almasry * reservations are to be un-charged from here. 928e9fe92aeSMina Almasry */ 929e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); 9305e911373SMike Kravetz 9315e911373SMike Kravetz INIT_LIST_HEAD(&resv_map->region_cache); 9325e911373SMike Kravetz list_add(&rg->link, &resv_map->region_cache); 9335e911373SMike Kravetz resv_map->region_cache_count = 1; 9345e911373SMike Kravetz 93584afd99bSAndy Whitcroft return resv_map; 93684afd99bSAndy Whitcroft } 93784afd99bSAndy Whitcroft 9389119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 93984afd99bSAndy Whitcroft { 94084afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 9415e911373SMike Kravetz struct list_head *head = &resv_map->region_cache; 9425e911373SMike Kravetz struct file_region *rg, *trg; 94384afd99bSAndy Whitcroft 94484afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 945feba16e2SMike Kravetz region_del(resv_map, 0, LONG_MAX); 9465e911373SMike Kravetz 9475e911373SMike Kravetz /* ... and any entries left in the cache */ 9485e911373SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 9495e911373SMike Kravetz list_del(&rg->link); 9505e911373SMike Kravetz kfree(rg); 9515e911373SMike Kravetz } 9525e911373SMike Kravetz 9535e911373SMike Kravetz VM_BUG_ON(resv_map->adds_in_progress); 9545e911373SMike Kravetz 95584afd99bSAndy Whitcroft kfree(resv_map); 95684afd99bSAndy Whitcroft } 95784afd99bSAndy Whitcroft 9584e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 9594e35f483SJoonsoo Kim { 960f27a5136SMike Kravetz /* 961f27a5136SMike Kravetz * At inode evict time, i_mapping may not point to the original 962f27a5136SMike Kravetz * address space within the inode. This original address space 963f27a5136SMike Kravetz * contains the pointer to the resv_map. So, always use the 964f27a5136SMike Kravetz * address space embedded within the inode. 965f27a5136SMike Kravetz * The VERY common case is inode->mapping == &inode->i_data but, 966f27a5136SMike Kravetz * this may not be true for device special inodes. 967f27a5136SMike Kravetz */ 968f27a5136SMike Kravetz return (struct resv_map *)(&inode->i_data)->private_data; 9694e35f483SJoonsoo Kim } 9704e35f483SJoonsoo Kim 97184afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 972a1e78772SMel Gorman { 97381d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 9744e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 9754e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 9764e35f483SJoonsoo Kim struct inode *inode = mapping->host; 9774e35f483SJoonsoo Kim 9784e35f483SJoonsoo Kim return inode_resv_map(inode); 9794e35f483SJoonsoo Kim 9804e35f483SJoonsoo Kim } else { 98184afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 98284afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 9834e35f483SJoonsoo Kim } 984a1e78772SMel Gorman } 985a1e78772SMel Gorman 98684afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 987a1e78772SMel Gorman { 98881d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 98981d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 990a1e78772SMel Gorman 99184afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 99284afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 99304f2cbe3SMel Gorman } 99404f2cbe3SMel Gorman 99504f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 99604f2cbe3SMel Gorman { 99781d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 99881d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 999e7c4b0bfSAndy Whitcroft 1000e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 100104f2cbe3SMel Gorman } 100204f2cbe3SMel Gorman 100304f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 100404f2cbe3SMel Gorman { 100581d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 1006e7c4b0bfSAndy Whitcroft 1007e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 1008a1e78772SMel Gorman } 1009a1e78772SMel Gorman 101004f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 1011a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 1012a1e78772SMel Gorman { 101381d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 1014f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 1015a1e78772SMel Gorman vma->vm_private_data = (void *)0; 1016a1e78772SMel Gorman } 1017a1e78772SMel Gorman 1018550a7d60SMina Almasry /* 1019550a7d60SMina Almasry * Reset and decrement one ref on hugepage private reservation. 1020550a7d60SMina Almasry * Called with mm->mmap_sem writer semaphore held. 1021550a7d60SMina Almasry * This function should be only used by move_vma() and operate on 1022550a7d60SMina Almasry * same sized vma. It should never come here with last ref on the 1023550a7d60SMina Almasry * reservation. 1024550a7d60SMina Almasry */ 1025550a7d60SMina Almasry void clear_vma_resv_huge_pages(struct vm_area_struct *vma) 1026550a7d60SMina Almasry { 1027550a7d60SMina Almasry /* 1028550a7d60SMina Almasry * Clear the old hugetlb private page reservation. 1029550a7d60SMina Almasry * It has already been transferred to new_vma. 1030550a7d60SMina Almasry * 1031550a7d60SMina Almasry * During a mremap() operation of a hugetlb vma we call move_vma() 1032550a7d60SMina Almasry * which copies vma into new_vma and unmaps vma. After the copy 1033550a7d60SMina Almasry * operation both new_vma and vma share a reference to the resv_map 1034550a7d60SMina Almasry * struct, and at that point vma is about to be unmapped. We don't 1035550a7d60SMina Almasry * want to return the reservation to the pool at unmap of vma because 1036550a7d60SMina Almasry * the reservation still lives on in new_vma, so simply decrement the 1037550a7d60SMina Almasry * ref here and remove the resv_map reference from this vma. 1038550a7d60SMina Almasry */ 1039550a7d60SMina Almasry struct resv_map *reservations = vma_resv_map(vma); 1040550a7d60SMina Almasry 1041afe041c2SBui Quang Minh if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 1042afe041c2SBui Quang Minh resv_map_put_hugetlb_cgroup_uncharge_info(reservations); 1043550a7d60SMina Almasry kref_put(&reservations->refs, resv_map_release); 1044afe041c2SBui Quang Minh } 1045550a7d60SMina Almasry 1046550a7d60SMina Almasry reset_vma_resv_huge_pages(vma); 1047550a7d60SMina Almasry } 1048550a7d60SMina Almasry 1049a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 1050559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 1051a1e78772SMel Gorman { 1052af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 1053af0ed73eSJoonsoo Kim /* 1054af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 1055af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 1056af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 1057af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 1058af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 1059af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 1060af0ed73eSJoonsoo Kim * properly, so add work-around here. 1061af0ed73eSJoonsoo Kim */ 1062af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 1063559ec2f8SNicholas Krause return true; 1064af0ed73eSJoonsoo Kim else 1065559ec2f8SNicholas Krause return false; 1066af0ed73eSJoonsoo Kim } 1067a63884e9SJoonsoo Kim 1068a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 10691fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 10701fb1b0e9SMike Kravetz /* 10711fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 10721fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 10731fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 10747c8de358SEthon Paul * fallocate. In this case, there really are no reserves to 10751fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 10761fb1b0e9SMike Kravetz */ 10771fb1b0e9SMike Kravetz if (chg) 10781fb1b0e9SMike Kravetz return false; 10791fb1b0e9SMike Kravetz else 1080559ec2f8SNicholas Krause return true; 10811fb1b0e9SMike Kravetz } 1082a63884e9SJoonsoo Kim 1083a63884e9SJoonsoo Kim /* 1084a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 1085a63884e9SJoonsoo Kim * private mappings. 1086a63884e9SJoonsoo Kim */ 108767961f9dSMike Kravetz if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 108867961f9dSMike Kravetz /* 108967961f9dSMike Kravetz * Like the shared case above, a hole punch or truncate 109067961f9dSMike Kravetz * could have been performed on the private mapping. 109167961f9dSMike Kravetz * Examine the value of chg to determine if reserves 109267961f9dSMike Kravetz * actually exist or were previously consumed. 109367961f9dSMike Kravetz * Very Subtle - The value of chg comes from a previous 109467961f9dSMike Kravetz * call to vma_needs_reserves(). The reserve map for 109567961f9dSMike Kravetz * private mappings has different (opposite) semantics 109667961f9dSMike Kravetz * than that of shared mappings. vma_needs_reserves() 109767961f9dSMike Kravetz * has already taken this difference in semantics into 109867961f9dSMike Kravetz * account. Therefore, the meaning of chg is the same 109967961f9dSMike Kravetz * as in the shared case above. Code could easily be 110067961f9dSMike Kravetz * combined, but keeping it separate draws attention to 110167961f9dSMike Kravetz * subtle differences. 110267961f9dSMike Kravetz */ 110367961f9dSMike Kravetz if (chg) 110467961f9dSMike Kravetz return false; 110567961f9dSMike Kravetz else 1106559ec2f8SNicholas Krause return true; 110767961f9dSMike Kravetz } 1108a63884e9SJoonsoo Kim 1109559ec2f8SNicholas Krause return false; 1110a1e78772SMel Gorman } 1111a1e78772SMel Gorman 1112a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 11131da177e4SLinus Torvalds { 11141da177e4SLinus Torvalds int nid = page_to_nid(page); 11159487ca60SMike Kravetz 11169487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1117b65a4edaSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1118b65a4edaSMike Kravetz 11190edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 1120a5516438SAndi Kleen h->free_huge_pages++; 1121a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 11226c037149SMike Kravetz SetHPageFreed(page); 11231da177e4SLinus Torvalds } 11241da177e4SLinus Torvalds 112594310cbcSAnshuman Khandual static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) 1126bf50bab2SNaoya Horiguchi { 1127bf50bab2SNaoya Horiguchi struct page *page; 11281a08ae36SPavel Tatashin bool pin = !!(current->flags & PF_MEMALLOC_PIN); 1129bf50bab2SNaoya Horiguchi 11309487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1131bbe88753SJoonsoo Kim list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { 11328e3560d9SPavel Tatashin if (pin && !is_pinnable_page(page)) 1133bbe88753SJoonsoo Kim continue; 1134bbe88753SJoonsoo Kim 11356664bfc8SWei Yang if (PageHWPoison(page)) 11366664bfc8SWei Yang continue; 1137bbe88753SJoonsoo Kim 11380edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 1139a9869b83SNaoya Horiguchi set_page_refcounted(page); 11406c037149SMike Kravetz ClearHPageFreed(page); 1141bf50bab2SNaoya Horiguchi h->free_huge_pages--; 1142bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 1143bf50bab2SNaoya Horiguchi return page; 1144bf50bab2SNaoya Horiguchi } 1145bf50bab2SNaoya Horiguchi 11466664bfc8SWei Yang return NULL; 11476664bfc8SWei Yang } 11486664bfc8SWei Yang 11493e59fcb0SMichal Hocko static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, 11503e59fcb0SMichal Hocko nodemask_t *nmask) 115194310cbcSAnshuman Khandual { 11523e59fcb0SMichal Hocko unsigned int cpuset_mems_cookie; 11533e59fcb0SMichal Hocko struct zonelist *zonelist; 11543e59fcb0SMichal Hocko struct zone *zone; 11553e59fcb0SMichal Hocko struct zoneref *z; 115698fa15f3SAnshuman Khandual int node = NUMA_NO_NODE; 11573e59fcb0SMichal Hocko 11583e59fcb0SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 11593e59fcb0SMichal Hocko 11603e59fcb0SMichal Hocko retry_cpuset: 11613e59fcb0SMichal Hocko cpuset_mems_cookie = read_mems_allowed_begin(); 11623e59fcb0SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 116394310cbcSAnshuman Khandual struct page *page; 116494310cbcSAnshuman Khandual 11653e59fcb0SMichal Hocko if (!cpuset_zone_allowed(zone, gfp_mask)) 11663e59fcb0SMichal Hocko continue; 11673e59fcb0SMichal Hocko /* 11683e59fcb0SMichal Hocko * no need to ask again on the same node. Pool is node rather than 11693e59fcb0SMichal Hocko * zone aware 11703e59fcb0SMichal Hocko */ 11713e59fcb0SMichal Hocko if (zone_to_nid(zone) == node) 11723e59fcb0SMichal Hocko continue; 11733e59fcb0SMichal Hocko node = zone_to_nid(zone); 117494310cbcSAnshuman Khandual 117594310cbcSAnshuman Khandual page = dequeue_huge_page_node_exact(h, node); 117694310cbcSAnshuman Khandual if (page) 117794310cbcSAnshuman Khandual return page; 117894310cbcSAnshuman Khandual } 11793e59fcb0SMichal Hocko if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 11803e59fcb0SMichal Hocko goto retry_cpuset; 11813e59fcb0SMichal Hocko 118294310cbcSAnshuman Khandual return NULL; 118394310cbcSAnshuman Khandual } 118494310cbcSAnshuman Khandual 1185a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 1186a5516438SAndi Kleen struct vm_area_struct *vma, 1187af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 1188af0ed73eSJoonsoo Kim long chg) 11891da177e4SLinus Torvalds { 1190cfcaa66fSBen Widawsky struct page *page = NULL; 1191480eccf9SLee Schermerhorn struct mempolicy *mpol; 119204ec6264SVlastimil Babka gfp_t gfp_mask; 11933e59fcb0SMichal Hocko nodemask_t *nodemask; 119404ec6264SVlastimil Babka int nid; 11951da177e4SLinus Torvalds 1196a1e78772SMel Gorman /* 1197a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 1198a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 1199a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 1200a1e78772SMel Gorman */ 1201af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg) && 1202a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 1203c0ff7453SMiao Xie goto err; 1204a1e78772SMel Gorman 120504f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 1206a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 12076eab04a8SJustin P. Mattock goto err; 120804f2cbe3SMel Gorman 120904ec6264SVlastimil Babka gfp_mask = htlb_alloc_mask(h); 121004ec6264SVlastimil Babka nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 1211cfcaa66fSBen Widawsky 1212cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 12133e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 1214cfcaa66fSBen Widawsky 1215cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 1216cfcaa66fSBen Widawsky nodemask = NULL; 1217cfcaa66fSBen Widawsky } 1218cfcaa66fSBen Widawsky 1219cfcaa66fSBen Widawsky if (!page) 1220cfcaa66fSBen Widawsky page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 1221cfcaa66fSBen Widawsky 12223e59fcb0SMichal Hocko if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { 1223d6995da3SMike Kravetz SetHPageRestoreReserve(page); 1224a63884e9SJoonsoo Kim h->resv_huge_pages--; 1225bf50bab2SNaoya Horiguchi } 1226cc9a6c87SMel Gorman 1227cc9a6c87SMel Gorman mpol_cond_put(mpol); 1228cc9a6c87SMel Gorman return page; 1229cc9a6c87SMel Gorman 1230c0ff7453SMiao Xie err: 1231cc9a6c87SMel Gorman return NULL; 12321da177e4SLinus Torvalds } 12331da177e4SLinus Torvalds 12341cac6f2cSLuiz Capitulino /* 12351cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 12361cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 12371cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 12381cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 12391cac6f2cSLuiz Capitulino * node for alloc or free. 12401cac6f2cSLuiz Capitulino */ 12411cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 12421cac6f2cSLuiz Capitulino { 12430edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 12441cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 12451cac6f2cSLuiz Capitulino 12461cac6f2cSLuiz Capitulino return nid; 12471cac6f2cSLuiz Capitulino } 12481cac6f2cSLuiz Capitulino 12491cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 12501cac6f2cSLuiz Capitulino { 12511cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 12521cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 12531cac6f2cSLuiz Capitulino return nid; 12541cac6f2cSLuiz Capitulino } 12551cac6f2cSLuiz Capitulino 12561cac6f2cSLuiz Capitulino /* 12571cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 12581cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 12591cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 12601cac6f2cSLuiz Capitulino * mask. 12611cac6f2cSLuiz Capitulino */ 12621cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 12631cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 12641cac6f2cSLuiz Capitulino { 12651cac6f2cSLuiz Capitulino int nid; 12661cac6f2cSLuiz Capitulino 12671cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 12681cac6f2cSLuiz Capitulino 12691cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 12701cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 12711cac6f2cSLuiz Capitulino 12721cac6f2cSLuiz Capitulino return nid; 12731cac6f2cSLuiz Capitulino } 12741cac6f2cSLuiz Capitulino 12751cac6f2cSLuiz Capitulino /* 127610c6ec49SMike Kravetz * helper for remove_pool_huge_page() - return the previously saved 12771cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 12781cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 12791cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 12801cac6f2cSLuiz Capitulino */ 12811cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 12821cac6f2cSLuiz Capitulino { 12831cac6f2cSLuiz Capitulino int nid; 12841cac6f2cSLuiz Capitulino 12851cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 12861cac6f2cSLuiz Capitulino 12871cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 12881cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 12891cac6f2cSLuiz Capitulino 12901cac6f2cSLuiz Capitulino return nid; 12911cac6f2cSLuiz Capitulino } 12921cac6f2cSLuiz Capitulino 12931cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 12941cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 12951cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 12961cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 12971cac6f2cSLuiz Capitulino nr_nodes--) 12981cac6f2cSLuiz Capitulino 12991cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 13001cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 13011cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 13021cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 13031cac6f2cSLuiz Capitulino nr_nodes--) 13041cac6f2cSLuiz Capitulino 13058531fc6fSMike Kravetz /* used to demote non-gigantic_huge pages as well */ 130634d9e35bSMike Kravetz static void __destroy_compound_gigantic_page(struct page *page, 130734d9e35bSMike Kravetz unsigned int order, bool demote) 1308944d9fecSLuiz Capitulino { 1309944d9fecSLuiz Capitulino int i; 1310944d9fecSLuiz Capitulino int nr_pages = 1 << order; 1311944d9fecSLuiz Capitulino struct page *p = page + 1; 1312944d9fecSLuiz Capitulino 1313c8cc708aSGerald Schaefer atomic_set(compound_mapcount_ptr(page), 0); 131447e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 131547e29d32SJohn Hubbard 1316944d9fecSLuiz Capitulino for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1317a01f4390SMike Kravetz p->mapping = NULL; 13181d798ca3SKirill A. Shutemov clear_compound_head(p); 131934d9e35bSMike Kravetz if (!demote) 1320944d9fecSLuiz Capitulino set_page_refcounted(p); 1321944d9fecSLuiz Capitulino } 1322944d9fecSLuiz Capitulino 1323944d9fecSLuiz Capitulino set_compound_order(page, 0); 13245232c63fSMatthew Wilcox (Oracle) #ifdef CONFIG_64BIT 1325ba9c1201SGerald Schaefer page[1].compound_nr = 0; 13265232c63fSMatthew Wilcox (Oracle) #endif 1327944d9fecSLuiz Capitulino __ClearPageHead(page); 1328944d9fecSLuiz Capitulino } 1329944d9fecSLuiz Capitulino 13308531fc6fSMike Kravetz static void destroy_compound_hugetlb_page_for_demote(struct page *page, 13318531fc6fSMike Kravetz unsigned int order) 13328531fc6fSMike Kravetz { 13338531fc6fSMike Kravetz __destroy_compound_gigantic_page(page, order, true); 13348531fc6fSMike Kravetz } 13358531fc6fSMike Kravetz 13368531fc6fSMike Kravetz #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 133734d9e35bSMike Kravetz static void destroy_compound_gigantic_page(struct page *page, 133834d9e35bSMike Kravetz unsigned int order) 133934d9e35bSMike Kravetz { 134034d9e35bSMike Kravetz __destroy_compound_gigantic_page(page, order, false); 134134d9e35bSMike Kravetz } 134234d9e35bSMike Kravetz 1343d00181b9SKirill A. Shutemov static void free_gigantic_page(struct page *page, unsigned int order) 1344944d9fecSLuiz Capitulino { 1345cf11e85fSRoman Gushchin /* 1346cf11e85fSRoman Gushchin * If the page isn't allocated using the cma allocator, 1347cf11e85fSRoman Gushchin * cma_release() returns false. 1348cf11e85fSRoman Gushchin */ 1349dbda8feaSBarry Song #ifdef CONFIG_CMA 1350dbda8feaSBarry Song if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order)) 1351cf11e85fSRoman Gushchin return; 1352dbda8feaSBarry Song #endif 1353cf11e85fSRoman Gushchin 1354944d9fecSLuiz Capitulino free_contig_range(page_to_pfn(page), 1 << order); 1355944d9fecSLuiz Capitulino } 1356944d9fecSLuiz Capitulino 13574eb0716eSAlexandre Ghiti #ifdef CONFIG_CONTIG_ALLOC 1358d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 1359d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) 1360944d9fecSLuiz Capitulino { 136104adbc3fSMiaohe Lin unsigned long nr_pages = pages_per_huge_page(h); 1362953f064aSLi Xinhai if (nid == NUMA_NO_NODE) 1363953f064aSLi Xinhai nid = numa_mem_id(); 1364944d9fecSLuiz Capitulino 1365dbda8feaSBarry Song #ifdef CONFIG_CMA 1366dbda8feaSBarry Song { 1367cf11e85fSRoman Gushchin struct page *page; 1368cf11e85fSRoman Gushchin int node; 1369cf11e85fSRoman Gushchin 1370953f064aSLi Xinhai if (hugetlb_cma[nid]) { 1371953f064aSLi Xinhai page = cma_alloc(hugetlb_cma[nid], nr_pages, 1372953f064aSLi Xinhai huge_page_order(h), true); 1373953f064aSLi Xinhai if (page) 1374953f064aSLi Xinhai return page; 1375953f064aSLi Xinhai } 1376953f064aSLi Xinhai 1377953f064aSLi Xinhai if (!(gfp_mask & __GFP_THISNODE)) { 1378cf11e85fSRoman Gushchin for_each_node_mask(node, *nodemask) { 1379953f064aSLi Xinhai if (node == nid || !hugetlb_cma[node]) 1380cf11e85fSRoman Gushchin continue; 1381cf11e85fSRoman Gushchin 1382cf11e85fSRoman Gushchin page = cma_alloc(hugetlb_cma[node], nr_pages, 1383cf11e85fSRoman Gushchin huge_page_order(h), true); 1384cf11e85fSRoman Gushchin if (page) 1385cf11e85fSRoman Gushchin return page; 1386cf11e85fSRoman Gushchin } 1387cf11e85fSRoman Gushchin } 1388953f064aSLi Xinhai } 1389dbda8feaSBarry Song #endif 1390cf11e85fSRoman Gushchin 13915e27a2dfSAnshuman Khandual return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); 1392944d9fecSLuiz Capitulino } 1393944d9fecSLuiz Capitulino 13944eb0716eSAlexandre Ghiti #else /* !CONFIG_CONTIG_ALLOC */ 13954eb0716eSAlexandre Ghiti static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 13964eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 13974eb0716eSAlexandre Ghiti { 13984eb0716eSAlexandre Ghiti return NULL; 13994eb0716eSAlexandre Ghiti } 14004eb0716eSAlexandre Ghiti #endif /* CONFIG_CONTIG_ALLOC */ 1401944d9fecSLuiz Capitulino 1402e1073d1eSAneesh Kumar K.V #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 1403d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 14044eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 14054eb0716eSAlexandre Ghiti { 14064eb0716eSAlexandre Ghiti return NULL; 14074eb0716eSAlexandre Ghiti } 1408d00181b9SKirill A. Shutemov static inline void free_gigantic_page(struct page *page, unsigned int order) { } 1409944d9fecSLuiz Capitulino static inline void destroy_compound_gigantic_page(struct page *page, 1410d00181b9SKirill A. Shutemov unsigned int order) { } 1411944d9fecSLuiz Capitulino #endif 1412944d9fecSLuiz Capitulino 14136eb4e88aSMike Kravetz /* 14146eb4e88aSMike Kravetz * Remove hugetlb page from lists, and update dtor so that page appears 141534d9e35bSMike Kravetz * as just a compound page. 141634d9e35bSMike Kravetz * 141734d9e35bSMike Kravetz * A reference is held on the page, except in the case of demote. 14186eb4e88aSMike Kravetz * 14196eb4e88aSMike Kravetz * Must be called with hugetlb lock held. 14206eb4e88aSMike Kravetz */ 142134d9e35bSMike Kravetz static void __remove_hugetlb_page(struct hstate *h, struct page *page, 142234d9e35bSMike Kravetz bool adjust_surplus, 142334d9e35bSMike Kravetz bool demote) 14246eb4e88aSMike Kravetz { 14256eb4e88aSMike Kravetz int nid = page_to_nid(page); 14266eb4e88aSMike Kravetz 14276eb4e88aSMike Kravetz VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 14286eb4e88aSMike Kravetz VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page); 14296eb4e88aSMike Kravetz 14309487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 14316eb4e88aSMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 14326eb4e88aSMike Kravetz return; 14336eb4e88aSMike Kravetz 14346eb4e88aSMike Kravetz list_del(&page->lru); 14356eb4e88aSMike Kravetz 14366eb4e88aSMike Kravetz if (HPageFreed(page)) { 14376eb4e88aSMike Kravetz h->free_huge_pages--; 14386eb4e88aSMike Kravetz h->free_huge_pages_node[nid]--; 14396eb4e88aSMike Kravetz } 14406eb4e88aSMike Kravetz if (adjust_surplus) { 14416eb4e88aSMike Kravetz h->surplus_huge_pages--; 14426eb4e88aSMike Kravetz h->surplus_huge_pages_node[nid]--; 14436eb4e88aSMike Kravetz } 14446eb4e88aSMike Kravetz 1445e32d20c0SMike Kravetz /* 1446e32d20c0SMike Kravetz * Very subtle 1447e32d20c0SMike Kravetz * 1448e32d20c0SMike Kravetz * For non-gigantic pages set the destructor to the normal compound 1449e32d20c0SMike Kravetz * page dtor. This is needed in case someone takes an additional 1450e32d20c0SMike Kravetz * temporary ref to the page, and freeing is delayed until they drop 1451e32d20c0SMike Kravetz * their reference. 1452e32d20c0SMike Kravetz * 1453e32d20c0SMike Kravetz * For gigantic pages set the destructor to the null dtor. This 1454e32d20c0SMike Kravetz * destructor will never be called. Before freeing the gigantic 1455e32d20c0SMike Kravetz * page destroy_compound_gigantic_page will turn the compound page 1456e32d20c0SMike Kravetz * into a simple group of pages. After this the destructor does not 1457e32d20c0SMike Kravetz * apply. 1458e32d20c0SMike Kravetz * 1459e32d20c0SMike Kravetz * This handles the case where more than one ref is held when and 1460e32d20c0SMike Kravetz * after update_and_free_page is called. 146134d9e35bSMike Kravetz * 146234d9e35bSMike Kravetz * In the case of demote we do not ref count the page as it will soon 146334d9e35bSMike Kravetz * be turned into a page of smaller size. 1464e32d20c0SMike Kravetz */ 146534d9e35bSMike Kravetz if (!demote) 14666eb4e88aSMike Kravetz set_page_refcounted(page); 1467e32d20c0SMike Kravetz if (hstate_is_gigantic(h)) 14686eb4e88aSMike Kravetz set_compound_page_dtor(page, NULL_COMPOUND_DTOR); 1469e32d20c0SMike Kravetz else 1470e32d20c0SMike Kravetz set_compound_page_dtor(page, COMPOUND_PAGE_DTOR); 14716eb4e88aSMike Kravetz 14726eb4e88aSMike Kravetz h->nr_huge_pages--; 14736eb4e88aSMike Kravetz h->nr_huge_pages_node[nid]--; 14746eb4e88aSMike Kravetz } 14756eb4e88aSMike Kravetz 147634d9e35bSMike Kravetz static void remove_hugetlb_page(struct hstate *h, struct page *page, 147734d9e35bSMike Kravetz bool adjust_surplus) 147834d9e35bSMike Kravetz { 147934d9e35bSMike Kravetz __remove_hugetlb_page(h, page, adjust_surplus, false); 148034d9e35bSMike Kravetz } 148134d9e35bSMike Kravetz 14828531fc6fSMike Kravetz static void remove_hugetlb_page_for_demote(struct hstate *h, struct page *page, 14838531fc6fSMike Kravetz bool adjust_surplus) 14848531fc6fSMike Kravetz { 14858531fc6fSMike Kravetz __remove_hugetlb_page(h, page, adjust_surplus, true); 14868531fc6fSMike Kravetz } 14878531fc6fSMike Kravetz 1488ad2fa371SMuchun Song static void add_hugetlb_page(struct hstate *h, struct page *page, 1489ad2fa371SMuchun Song bool adjust_surplus) 1490ad2fa371SMuchun Song { 1491ad2fa371SMuchun Song int zeroed; 1492ad2fa371SMuchun Song int nid = page_to_nid(page); 1493ad2fa371SMuchun Song 1494ad2fa371SMuchun Song VM_BUG_ON_PAGE(!HPageVmemmapOptimized(page), page); 1495ad2fa371SMuchun Song 1496ad2fa371SMuchun Song lockdep_assert_held(&hugetlb_lock); 1497ad2fa371SMuchun Song 1498ad2fa371SMuchun Song INIT_LIST_HEAD(&page->lru); 1499ad2fa371SMuchun Song h->nr_huge_pages++; 1500ad2fa371SMuchun Song h->nr_huge_pages_node[nid]++; 1501ad2fa371SMuchun Song 1502ad2fa371SMuchun Song if (adjust_surplus) { 1503ad2fa371SMuchun Song h->surplus_huge_pages++; 1504ad2fa371SMuchun Song h->surplus_huge_pages_node[nid]++; 1505ad2fa371SMuchun Song } 1506ad2fa371SMuchun Song 1507ad2fa371SMuchun Song set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1508ad2fa371SMuchun Song set_page_private(page, 0); 1509ad2fa371SMuchun Song SetHPageVmemmapOptimized(page); 1510ad2fa371SMuchun Song 1511ad2fa371SMuchun Song /* 1512b65a4edaSMike Kravetz * This page is about to be managed by the hugetlb allocator and 1513b65a4edaSMike Kravetz * should have no users. Drop our reference, and check for others 1514b65a4edaSMike Kravetz * just in case. 1515ad2fa371SMuchun Song */ 1516ad2fa371SMuchun Song zeroed = put_page_testzero(page); 1517b65a4edaSMike Kravetz if (!zeroed) 1518b65a4edaSMike Kravetz /* 1519b65a4edaSMike Kravetz * It is VERY unlikely soneone else has taken a ref on 1520b65a4edaSMike Kravetz * the page. In this case, we simply return as the 1521b65a4edaSMike Kravetz * hugetlb destructor (free_huge_page) will be called 1522b65a4edaSMike Kravetz * when this other ref is dropped. 1523b65a4edaSMike Kravetz */ 1524b65a4edaSMike Kravetz return; 1525b65a4edaSMike Kravetz 1526ad2fa371SMuchun Song arch_clear_hugepage_flags(page); 1527ad2fa371SMuchun Song enqueue_huge_page(h, page); 1528ad2fa371SMuchun Song } 1529ad2fa371SMuchun Song 1530b65d4adbSMuchun Song static void __update_and_free_page(struct hstate *h, struct page *page) 15316af2acb6SAdam Litke { 15326af2acb6SAdam Litke int i; 1533dbfee5aeSMike Kravetz struct page *subpage = page; 1534a5516438SAndi Kleen 15354eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1536944d9fecSLuiz Capitulino return; 153718229df5SAndy Whitcroft 1538ad2fa371SMuchun Song if (alloc_huge_page_vmemmap(h, page)) { 1539ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 1540ad2fa371SMuchun Song /* 1541ad2fa371SMuchun Song * If we cannot allocate vmemmap pages, just refuse to free the 1542ad2fa371SMuchun Song * page and put the page back on the hugetlb free list and treat 1543ad2fa371SMuchun Song * as a surplus page. 1544ad2fa371SMuchun Song */ 1545ad2fa371SMuchun Song add_hugetlb_page(h, page, true); 1546ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 1547ad2fa371SMuchun Song return; 1548ad2fa371SMuchun Song } 1549ad2fa371SMuchun Song 1550dbfee5aeSMike Kravetz for (i = 0; i < pages_per_huge_page(h); 1551dbfee5aeSMike Kravetz i++, subpage = mem_map_next(subpage, page, i)) { 1552dbfee5aeSMike Kravetz subpage->flags &= ~(1 << PG_locked | 1 << PG_error | 155332f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1554a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1555a7407a27SLuiz Capitulino 1 << PG_writeback); 15566af2acb6SAdam Litke } 1557a01f4390SMike Kravetz 1558a01f4390SMike Kravetz /* 1559a01f4390SMike Kravetz * Non-gigantic pages demoted from CMA allocated gigantic pages 1560a01f4390SMike Kravetz * need to be given back to CMA in free_gigantic_page. 1561a01f4390SMike Kravetz */ 1562a01f4390SMike Kravetz if (hstate_is_gigantic(h) || 1563a01f4390SMike Kravetz hugetlb_cma_page(page, huge_page_order(h))) { 1564944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 1565944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 1566944d9fecSLuiz Capitulino } else { 1567a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 15686af2acb6SAdam Litke } 1569944d9fecSLuiz Capitulino } 15706af2acb6SAdam Litke 1571b65d4adbSMuchun Song /* 1572b65d4adbSMuchun Song * As update_and_free_page() can be called under any context, so we cannot 1573b65d4adbSMuchun Song * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the 1574b65d4adbSMuchun Song * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate 1575b65d4adbSMuchun Song * the vmemmap pages. 1576b65d4adbSMuchun Song * 1577b65d4adbSMuchun Song * free_hpage_workfn() locklessly retrieves the linked list of pages to be 1578b65d4adbSMuchun Song * freed and frees them one-by-one. As the page->mapping pointer is going 1579b65d4adbSMuchun Song * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node 1580b65d4adbSMuchun Song * structure of a lockless linked list of huge pages to be freed. 1581b65d4adbSMuchun Song */ 1582b65d4adbSMuchun Song static LLIST_HEAD(hpage_freelist); 1583b65d4adbSMuchun Song 1584b65d4adbSMuchun Song static void free_hpage_workfn(struct work_struct *work) 1585b65d4adbSMuchun Song { 1586b65d4adbSMuchun Song struct llist_node *node; 1587b65d4adbSMuchun Song 1588b65d4adbSMuchun Song node = llist_del_all(&hpage_freelist); 1589b65d4adbSMuchun Song 1590b65d4adbSMuchun Song while (node) { 1591b65d4adbSMuchun Song struct page *page; 1592b65d4adbSMuchun Song struct hstate *h; 1593b65d4adbSMuchun Song 1594b65d4adbSMuchun Song page = container_of((struct address_space **)node, 1595b65d4adbSMuchun Song struct page, mapping); 1596b65d4adbSMuchun Song node = node->next; 1597b65d4adbSMuchun Song page->mapping = NULL; 1598b65d4adbSMuchun Song /* 1599b65d4adbSMuchun Song * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate() 1600b65d4adbSMuchun Song * is going to trigger because a previous call to 1601b65d4adbSMuchun Song * remove_hugetlb_page() will set_compound_page_dtor(page, 1602b65d4adbSMuchun Song * NULL_COMPOUND_DTOR), so do not use page_hstate() directly. 1603b65d4adbSMuchun Song */ 1604b65d4adbSMuchun Song h = size_to_hstate(page_size(page)); 1605b65d4adbSMuchun Song 1606b65d4adbSMuchun Song __update_and_free_page(h, page); 1607b65d4adbSMuchun Song 1608b65d4adbSMuchun Song cond_resched(); 1609b65d4adbSMuchun Song } 1610b65d4adbSMuchun Song } 1611b65d4adbSMuchun Song static DECLARE_WORK(free_hpage_work, free_hpage_workfn); 1612b65d4adbSMuchun Song 1613b65d4adbSMuchun Song static inline void flush_free_hpage_work(struct hstate *h) 1614b65d4adbSMuchun Song { 1615b65d4adbSMuchun Song if (free_vmemmap_pages_per_hpage(h)) 1616b65d4adbSMuchun Song flush_work(&free_hpage_work); 1617b65d4adbSMuchun Song } 1618b65d4adbSMuchun Song 1619b65d4adbSMuchun Song static void update_and_free_page(struct hstate *h, struct page *page, 1620b65d4adbSMuchun Song bool atomic) 1621b65d4adbSMuchun Song { 1622ad2fa371SMuchun Song if (!HPageVmemmapOptimized(page) || !atomic) { 1623b65d4adbSMuchun Song __update_and_free_page(h, page); 1624b65d4adbSMuchun Song return; 1625b65d4adbSMuchun Song } 1626b65d4adbSMuchun Song 1627b65d4adbSMuchun Song /* 1628b65d4adbSMuchun Song * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages. 1629b65d4adbSMuchun Song * 1630b65d4adbSMuchun Song * Only call schedule_work() if hpage_freelist is previously 1631b65d4adbSMuchun Song * empty. Otherwise, schedule_work() had been called but the workfn 1632b65d4adbSMuchun Song * hasn't retrieved the list yet. 1633b65d4adbSMuchun Song */ 1634b65d4adbSMuchun Song if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist)) 1635b65d4adbSMuchun Song schedule_work(&free_hpage_work); 1636b65d4adbSMuchun Song } 1637b65d4adbSMuchun Song 163810c6ec49SMike Kravetz static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list) 163910c6ec49SMike Kravetz { 164010c6ec49SMike Kravetz struct page *page, *t_page; 164110c6ec49SMike Kravetz 164210c6ec49SMike Kravetz list_for_each_entry_safe(page, t_page, list, lru) { 1643b65d4adbSMuchun Song update_and_free_page(h, page, false); 164410c6ec49SMike Kravetz cond_resched(); 164510c6ec49SMike Kravetz } 164610c6ec49SMike Kravetz } 164710c6ec49SMike Kravetz 1648e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1649e5ff2159SAndi Kleen { 1650e5ff2159SAndi Kleen struct hstate *h; 1651e5ff2159SAndi Kleen 1652e5ff2159SAndi Kleen for_each_hstate(h) { 1653e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1654e5ff2159SAndi Kleen return h; 1655e5ff2159SAndi Kleen } 1656e5ff2159SAndi Kleen return NULL; 1657e5ff2159SAndi Kleen } 1658e5ff2159SAndi Kleen 1659db71ef79SMike Kravetz void free_huge_page(struct page *page) 166027a85ef1SDavid Gibson { 1661a5516438SAndi Kleen /* 1662a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1663a5516438SAndi Kleen * compound page destructor. 1664a5516438SAndi Kleen */ 1665e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 16667893d1d5SAdam Litke int nid = page_to_nid(page); 1667d6995da3SMike Kravetz struct hugepage_subpool *spool = hugetlb_page_subpool(page); 166807443a85SJoonsoo Kim bool restore_reserve; 1669db71ef79SMike Kravetz unsigned long flags; 167027a85ef1SDavid Gibson 1671b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1672b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_mapcount(page), page); 16738ace22bcSYongkai Wu 1674d6995da3SMike Kravetz hugetlb_set_page_subpool(page, NULL); 16758ace22bcSYongkai Wu page->mapping = NULL; 1676d6995da3SMike Kravetz restore_reserve = HPageRestoreReserve(page); 1677d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 167827a85ef1SDavid Gibson 16791c5ecae3SMike Kravetz /* 1680d6995da3SMike Kravetz * If HPageRestoreReserve was set on page, page allocation consumed a 16810919e1b6SMike Kravetz * reservation. If the page was associated with a subpool, there 16820919e1b6SMike Kravetz * would have been a page reserved in the subpool before allocation 16830919e1b6SMike Kravetz * via hugepage_subpool_get_pages(). Since we are 'restoring' the 16846c26d310SMiaohe Lin * reservation, do not call hugepage_subpool_put_pages() as this will 16850919e1b6SMike Kravetz * remove the reserved page from the subpool. 16860919e1b6SMike Kravetz */ 16870919e1b6SMike Kravetz if (!restore_reserve) { 16880919e1b6SMike Kravetz /* 16890919e1b6SMike Kravetz * A return code of zero implies that the subpool will be 16900919e1b6SMike Kravetz * under its minimum size if the reservation is not restored 16910919e1b6SMike Kravetz * after page is free. Therefore, force restore_reserve 16920919e1b6SMike Kravetz * operation. 16931c5ecae3SMike Kravetz */ 16941c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 16951c5ecae3SMike Kravetz restore_reserve = true; 16960919e1b6SMike Kravetz } 16971c5ecae3SMike Kravetz 1698db71ef79SMike Kravetz spin_lock_irqsave(&hugetlb_lock, flags); 16998f251a3dSMike Kravetz ClearHPageMigratable(page); 17006d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 17016d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 170208cf9fafSMina Almasry hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 170308cf9fafSMina Almasry pages_per_huge_page(h), page); 170407443a85SJoonsoo Kim if (restore_reserve) 170507443a85SJoonsoo Kim h->resv_huge_pages++; 170607443a85SJoonsoo Kim 17079157c311SMike Kravetz if (HPageTemporary(page)) { 17086eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 1709db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1710b65d4adbSMuchun Song update_and_free_page(h, page, true); 1711ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 17120edaecfaSAneesh Kumar K.V /* remove the page from active list */ 17136eb4e88aSMike Kravetz remove_hugetlb_page(h, page, true); 1714db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1715b65d4adbSMuchun Song update_and_free_page(h, page, true); 17167893d1d5SAdam Litke } else { 17175d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1718a5516438SAndi Kleen enqueue_huge_page(h, page); 1719db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 172027a85ef1SDavid Gibson } 17211121828aSMike Kravetz } 172227a85ef1SDavid Gibson 1723d3d99fccSOscar Salvador /* 1724d3d99fccSOscar Salvador * Must be called with the hugetlb lock held 1725d3d99fccSOscar Salvador */ 1726d3d99fccSOscar Salvador static void __prep_account_new_huge_page(struct hstate *h, int nid) 1727d3d99fccSOscar Salvador { 1728d3d99fccSOscar Salvador lockdep_assert_held(&hugetlb_lock); 1729d3d99fccSOscar Salvador h->nr_huge_pages++; 1730d3d99fccSOscar Salvador h->nr_huge_pages_node[nid]++; 1731d3d99fccSOscar Salvador } 1732d3d99fccSOscar Salvador 1733f41f2ed4SMuchun Song static void __prep_new_huge_page(struct hstate *h, struct page *page) 1734b7ba30c6SAndi Kleen { 1735f41f2ed4SMuchun Song free_huge_page_vmemmap(h, page); 17360edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1737f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1738ff546117SMike Kravetz hugetlb_set_page_subpool(page, NULL); 17399dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 17401adc4d41SMina Almasry set_hugetlb_cgroup_rsvd(page, NULL); 1741d3d99fccSOscar Salvador } 1742d3d99fccSOscar Salvador 1743d3d99fccSOscar Salvador static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1744d3d99fccSOscar Salvador { 1745f41f2ed4SMuchun Song __prep_new_huge_page(h, page); 1746db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 1747d3d99fccSOscar Salvador __prep_account_new_huge_page(h, nid); 1748db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1749b7ba30c6SAndi Kleen } 1750b7ba30c6SAndi Kleen 175134d9e35bSMike Kravetz static bool __prep_compound_gigantic_page(struct page *page, unsigned int order, 175234d9e35bSMike Kravetz bool demote) 175320a0307cSWu Fengguang { 17547118fc29SMike Kravetz int i, j; 175520a0307cSWu Fengguang int nr_pages = 1 << order; 175620a0307cSWu Fengguang struct page *p = page + 1; 175720a0307cSWu Fengguang 175820a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 175920a0307cSWu Fengguang set_compound_order(page, order); 1760ef5a22beSAndrea Arcangeli __ClearPageReserved(page); 1761de09d31dSKirill A. Shutemov __SetPageHead(page); 176220a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1763ef5a22beSAndrea Arcangeli /* 1764ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1765ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1766ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 17677c8de358SEthon Paul * too. Otherwise drivers using get_user_pages() to access tail 1768ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1769ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1770ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1771ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1772ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1773ef5a22beSAndrea Arcangeli * after get_user_pages(). 1774ef5a22beSAndrea Arcangeli */ 1775ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 17767118fc29SMike Kravetz /* 17777118fc29SMike Kravetz * Subtle and very unlikely 17787118fc29SMike Kravetz * 17797118fc29SMike Kravetz * Gigantic 'page allocators' such as memblock or cma will 17807118fc29SMike Kravetz * return a set of pages with each page ref counted. We need 17817118fc29SMike Kravetz * to turn this set of pages into a compound page with tail 17827118fc29SMike Kravetz * page ref counts set to zero. Code such as speculative page 17837118fc29SMike Kravetz * cache adding could take a ref on a 'to be' tail page. 17847118fc29SMike Kravetz * We need to respect any increased ref count, and only set 17857118fc29SMike Kravetz * the ref count to zero if count is currently 1. If count 1786416d85edSMike Kravetz * is not 1, we return an error. An error return indicates 1787416d85edSMike Kravetz * the set of pages can not be converted to a gigantic page. 1788416d85edSMike Kravetz * The caller who allocated the pages should then discard the 1789416d85edSMike Kravetz * pages using the appropriate free interface. 179034d9e35bSMike Kravetz * 179134d9e35bSMike Kravetz * In the case of demote, the ref count will be zero. 17927118fc29SMike Kravetz */ 179334d9e35bSMike Kravetz if (!demote) { 17947118fc29SMike Kravetz if (!page_ref_freeze(p, 1)) { 1795416d85edSMike Kravetz pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n"); 17967118fc29SMike Kravetz goto out_error; 17977118fc29SMike Kravetz } 179834d9e35bSMike Kravetz } else { 179934d9e35bSMike Kravetz VM_BUG_ON_PAGE(page_count(p), p); 180034d9e35bSMike Kravetz } 18011d798ca3SKirill A. Shutemov set_compound_head(p, page); 180220a0307cSWu Fengguang } 1803b4330afbSMike Kravetz atomic_set(compound_mapcount_ptr(page), -1); 180447e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 18057118fc29SMike Kravetz return true; 18067118fc29SMike Kravetz 18077118fc29SMike Kravetz out_error: 18087118fc29SMike Kravetz /* undo tail page modifications made above */ 18097118fc29SMike Kravetz p = page + 1; 18107118fc29SMike Kravetz for (j = 1; j < i; j++, p = mem_map_next(p, page, j)) { 18117118fc29SMike Kravetz clear_compound_head(p); 18127118fc29SMike Kravetz set_page_refcounted(p); 18137118fc29SMike Kravetz } 18147118fc29SMike Kravetz /* need to clear PG_reserved on remaining tail pages */ 18157118fc29SMike Kravetz for (; j < nr_pages; j++, p = mem_map_next(p, page, j)) 18167118fc29SMike Kravetz __ClearPageReserved(p); 18177118fc29SMike Kravetz set_compound_order(page, 0); 18185232c63fSMatthew Wilcox (Oracle) #ifdef CONFIG_64BIT 18197118fc29SMike Kravetz page[1].compound_nr = 0; 18205232c63fSMatthew Wilcox (Oracle) #endif 18217118fc29SMike Kravetz __ClearPageHead(page); 18227118fc29SMike Kravetz return false; 182320a0307cSWu Fengguang } 182420a0307cSWu Fengguang 182534d9e35bSMike Kravetz static bool prep_compound_gigantic_page(struct page *page, unsigned int order) 182634d9e35bSMike Kravetz { 182734d9e35bSMike Kravetz return __prep_compound_gigantic_page(page, order, false); 182834d9e35bSMike Kravetz } 182934d9e35bSMike Kravetz 18308531fc6fSMike Kravetz static bool prep_compound_gigantic_page_for_demote(struct page *page, 18318531fc6fSMike Kravetz unsigned int order) 18328531fc6fSMike Kravetz { 18338531fc6fSMike Kravetz return __prep_compound_gigantic_page(page, order, true); 18348531fc6fSMike Kravetz } 18358531fc6fSMike Kravetz 18367795912cSAndrew Morton /* 18377795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 18387795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 18397795912cSAndrew Morton * details. 18407795912cSAndrew Morton */ 184120a0307cSWu Fengguang int PageHuge(struct page *page) 184220a0307cSWu Fengguang { 184320a0307cSWu Fengguang if (!PageCompound(page)) 184420a0307cSWu Fengguang return 0; 184520a0307cSWu Fengguang 184620a0307cSWu Fengguang page = compound_head(page); 1847f1e61557SKirill A. Shutemov return page[1].compound_dtor == HUGETLB_PAGE_DTOR; 184820a0307cSWu Fengguang } 184943131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 185043131e14SNaoya Horiguchi 185127c73ae7SAndrea Arcangeli /* 185227c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 185327c73ae7SAndrea Arcangeli * normal or transparent huge pages. 185427c73ae7SAndrea Arcangeli */ 185527c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 185627c73ae7SAndrea Arcangeli { 185727c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 185827c73ae7SAndrea Arcangeli return 0; 185927c73ae7SAndrea Arcangeli 1860d4af73e3SVlastimil Babka return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; 186127c73ae7SAndrea Arcangeli } 18624e936eccSDavid Howells EXPORT_SYMBOL_GPL(PageHeadHuge); 186327c73ae7SAndrea Arcangeli 1864c0d0381aSMike Kravetz /* 1865c0d0381aSMike Kravetz * Find and lock address space (mapping) in write mode. 1866c0d0381aSMike Kravetz * 1867336bf30eSMike Kravetz * Upon entry, the page is locked which means that page_mapping() is 1868336bf30eSMike Kravetz * stable. Due to locking order, we can only trylock_write. If we can 1869336bf30eSMike Kravetz * not get the lock, simply return NULL to caller. 1870c0d0381aSMike Kravetz */ 1871c0d0381aSMike Kravetz struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) 1872c0d0381aSMike Kravetz { 1873336bf30eSMike Kravetz struct address_space *mapping = page_mapping(hpage); 1874c0d0381aSMike Kravetz 1875c0d0381aSMike Kravetz if (!mapping) 1876c0d0381aSMike Kravetz return mapping; 1877c0d0381aSMike Kravetz 1878c0d0381aSMike Kravetz if (i_mmap_trylock_write(mapping)) 1879c0d0381aSMike Kravetz return mapping; 1880c0d0381aSMike Kravetz 1881c0d0381aSMike Kravetz return NULL; 1882c0d0381aSMike Kravetz } 1883c0d0381aSMike Kravetz 1884fe19bd3dSHugh Dickins pgoff_t hugetlb_basepage_index(struct page *page) 188513d60f4bSZhang Yi { 188613d60f4bSZhang Yi struct page *page_head = compound_head(page); 188713d60f4bSZhang Yi pgoff_t index = page_index(page_head); 188813d60f4bSZhang Yi unsigned long compound_idx; 188913d60f4bSZhang Yi 189013d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 189113d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 189213d60f4bSZhang Yi else 189313d60f4bSZhang Yi compound_idx = page - page_head; 189413d60f4bSZhang Yi 189513d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 189613d60f4bSZhang Yi } 189713d60f4bSZhang Yi 18980c397daeSMichal Hocko static struct page *alloc_buddy_huge_page(struct hstate *h, 1899f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1900f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 19011da177e4SLinus Torvalds { 1902af0fb9dfSMichal Hocko int order = huge_page_order(h); 19031da177e4SLinus Torvalds struct page *page; 1904f60858f9SMike Kravetz bool alloc_try_hard = true; 1905f96efd58SJoe Jin 1906f60858f9SMike Kravetz /* 1907f60858f9SMike Kravetz * By default we always try hard to allocate the page with 1908f60858f9SMike Kravetz * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in 1909f60858f9SMike Kravetz * a loop (to adjust global huge page counts) and previous allocation 1910f60858f9SMike Kravetz * failed, do not continue to try hard on the same node. Use the 1911f60858f9SMike Kravetz * node_alloc_noretry bitmap to manage this state information. 1912f60858f9SMike Kravetz */ 1913f60858f9SMike Kravetz if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) 1914f60858f9SMike Kravetz alloc_try_hard = false; 1915f60858f9SMike Kravetz gfp_mask |= __GFP_COMP|__GFP_NOWARN; 1916f60858f9SMike Kravetz if (alloc_try_hard) 1917f60858f9SMike Kravetz gfp_mask |= __GFP_RETRY_MAYFAIL; 1918af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 1919af0fb9dfSMichal Hocko nid = numa_mem_id(); 192084172f4bSMatthew Wilcox (Oracle) page = __alloc_pages(gfp_mask, order, nid, nmask); 1921af0fb9dfSMichal Hocko if (page) 1922af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC); 1923af0fb9dfSMichal Hocko else 1924af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 192563b4613cSNishanth Aravamudan 1926f60858f9SMike Kravetz /* 1927f60858f9SMike Kravetz * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this 1928f60858f9SMike Kravetz * indicates an overall state change. Clear bit so that we resume 1929f60858f9SMike Kravetz * normal 'try hard' allocations. 1930f60858f9SMike Kravetz */ 1931f60858f9SMike Kravetz if (node_alloc_noretry && page && !alloc_try_hard) 1932f60858f9SMike Kravetz node_clear(nid, *node_alloc_noretry); 1933f60858f9SMike Kravetz 1934f60858f9SMike Kravetz /* 1935f60858f9SMike Kravetz * If we tried hard to get a page but failed, set bit so that 1936f60858f9SMike Kravetz * subsequent attempts will not try as hard until there is an 1937f60858f9SMike Kravetz * overall state change. 1938f60858f9SMike Kravetz */ 1939f60858f9SMike Kravetz if (node_alloc_noretry && !page && alloc_try_hard) 1940f60858f9SMike Kravetz node_set(nid, *node_alloc_noretry); 1941f60858f9SMike Kravetz 194263b4613cSNishanth Aravamudan return page; 194363b4613cSNishanth Aravamudan } 194463b4613cSNishanth Aravamudan 1945af0fb9dfSMichal Hocko /* 19460c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 19470c397daeSMichal Hocko * should use this function to get new hugetlb pages 19480c397daeSMichal Hocko */ 19490c397daeSMichal Hocko static struct page *alloc_fresh_huge_page(struct hstate *h, 1950f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1951f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 19520c397daeSMichal Hocko { 19530c397daeSMichal Hocko struct page *page; 19547118fc29SMike Kravetz bool retry = false; 19550c397daeSMichal Hocko 19567118fc29SMike Kravetz retry: 19570c397daeSMichal Hocko if (hstate_is_gigantic(h)) 19580c397daeSMichal Hocko page = alloc_gigantic_page(h, gfp_mask, nid, nmask); 19590c397daeSMichal Hocko else 19600c397daeSMichal Hocko page = alloc_buddy_huge_page(h, gfp_mask, 1961f60858f9SMike Kravetz nid, nmask, node_alloc_noretry); 19620c397daeSMichal Hocko if (!page) 19630c397daeSMichal Hocko return NULL; 19640c397daeSMichal Hocko 19657118fc29SMike Kravetz if (hstate_is_gigantic(h)) { 19667118fc29SMike Kravetz if (!prep_compound_gigantic_page(page, huge_page_order(h))) { 19677118fc29SMike Kravetz /* 19687118fc29SMike Kravetz * Rare failure to convert pages to compound page. 19697118fc29SMike Kravetz * Free pages and try again - ONCE! 19707118fc29SMike Kravetz */ 19717118fc29SMike Kravetz free_gigantic_page(page, huge_page_order(h)); 19727118fc29SMike Kravetz if (!retry) { 19737118fc29SMike Kravetz retry = true; 19747118fc29SMike Kravetz goto retry; 19757118fc29SMike Kravetz } 19767118fc29SMike Kravetz return NULL; 19777118fc29SMike Kravetz } 19787118fc29SMike Kravetz } 19790c397daeSMichal Hocko prep_new_huge_page(h, page, page_to_nid(page)); 19800c397daeSMichal Hocko 19810c397daeSMichal Hocko return page; 19820c397daeSMichal Hocko } 19830c397daeSMichal Hocko 19840c397daeSMichal Hocko /* 1985af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 1986af0fb9dfSMichal Hocko * manner. 1987af0fb9dfSMichal Hocko */ 1988f60858f9SMike Kravetz static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 1989f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 1990b2261026SJoonsoo Kim { 1991b2261026SJoonsoo Kim struct page *page; 1992b2261026SJoonsoo Kim int nr_nodes, node; 1993af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 1994b2261026SJoonsoo Kim 1995b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1996f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, 1997f60858f9SMike Kravetz node_alloc_noretry); 1998af0fb9dfSMichal Hocko if (page) 1999b2261026SJoonsoo Kim break; 2000b2261026SJoonsoo Kim } 2001b2261026SJoonsoo Kim 2002af0fb9dfSMichal Hocko if (!page) 2003af0fb9dfSMichal Hocko return 0; 2004b2261026SJoonsoo Kim 2005af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 2006af0fb9dfSMichal Hocko 2007af0fb9dfSMichal Hocko return 1; 2008b2261026SJoonsoo Kim } 2009b2261026SJoonsoo Kim 2010e8c5c824SLee Schermerhorn /* 201110c6ec49SMike Kravetz * Remove huge page from pool from next node to free. Attempt to keep 201210c6ec49SMike Kravetz * persistent huge pages more or less balanced over allowed nodes. 201310c6ec49SMike Kravetz * This routine only 'removes' the hugetlb page. The caller must make 201410c6ec49SMike Kravetz * an additional call to free the page to low level allocators. 2015e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 2016e8c5c824SLee Schermerhorn */ 201710c6ec49SMike Kravetz static struct page *remove_pool_huge_page(struct hstate *h, 201810c6ec49SMike Kravetz nodemask_t *nodes_allowed, 20196ae11b27SLee Schermerhorn bool acct_surplus) 2020e8c5c824SLee Schermerhorn { 2021b2261026SJoonsoo Kim int nr_nodes, node; 202210c6ec49SMike Kravetz struct page *page = NULL; 2023e8c5c824SLee Schermerhorn 20249487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2025b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2026685f3457SLee Schermerhorn /* 2027685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 2028685f3457SLee Schermerhorn * nodes with surplus pages. 2029685f3457SLee Schermerhorn */ 2030b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 2031b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 203210c6ec49SMike Kravetz page = list_entry(h->hugepage_freelists[node].next, 2033e8c5c824SLee Schermerhorn struct page, lru); 20346eb4e88aSMike Kravetz remove_hugetlb_page(h, page, acct_surplus); 20359a76db09SLee Schermerhorn break; 2036e8c5c824SLee Schermerhorn } 2037b2261026SJoonsoo Kim } 2038e8c5c824SLee Schermerhorn 203910c6ec49SMike Kravetz return page; 2040e8c5c824SLee Schermerhorn } 2041e8c5c824SLee Schermerhorn 2042c8721bbbSNaoya Horiguchi /* 2043c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 2044faf53defSNaoya Horiguchi * nothing for in-use hugepages and non-hugepages. 2045faf53defSNaoya Horiguchi * This function returns values like below: 2046faf53defSNaoya Horiguchi * 2047ad2fa371SMuchun Song * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages 2048ad2fa371SMuchun Song * when the system is under memory pressure and the feature of 2049ad2fa371SMuchun Song * freeing unused vmemmap pages associated with each hugetlb page 2050ad2fa371SMuchun Song * is enabled. 2051faf53defSNaoya Horiguchi * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 2052faf53defSNaoya Horiguchi * (allocated or reserved.) 2053faf53defSNaoya Horiguchi * 0: successfully dissolved free hugepages or the page is not a 2054faf53defSNaoya Horiguchi * hugepage (considered as already dissolved) 2055c8721bbbSNaoya Horiguchi */ 2056c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 2057c8721bbbSNaoya Horiguchi { 20586bc9b564SNaoya Horiguchi int rc = -EBUSY; 2059082d5b6bSGerald Schaefer 20607ffddd49SMuchun Song retry: 2061faf53defSNaoya Horiguchi /* Not to disrupt normal path by vainly holding hugetlb_lock */ 2062faf53defSNaoya Horiguchi if (!PageHuge(page)) 2063faf53defSNaoya Horiguchi return 0; 2064faf53defSNaoya Horiguchi 2065db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2066faf53defSNaoya Horiguchi if (!PageHuge(page)) { 2067faf53defSNaoya Horiguchi rc = 0; 2068faf53defSNaoya Horiguchi goto out; 2069faf53defSNaoya Horiguchi } 2070faf53defSNaoya Horiguchi 2071faf53defSNaoya Horiguchi if (!page_count(page)) { 20722247bb33SGerald Schaefer struct page *head = compound_head(page); 20732247bb33SGerald Schaefer struct hstate *h = page_hstate(head); 20746bc9b564SNaoya Horiguchi if (h->free_huge_pages - h->resv_huge_pages == 0) 2075082d5b6bSGerald Schaefer goto out; 20767ffddd49SMuchun Song 20777ffddd49SMuchun Song /* 20787ffddd49SMuchun Song * We should make sure that the page is already on the free list 20797ffddd49SMuchun Song * when it is dissolved. 20807ffddd49SMuchun Song */ 20816c037149SMike Kravetz if (unlikely(!HPageFreed(head))) { 2082db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 20837ffddd49SMuchun Song cond_resched(); 20847ffddd49SMuchun Song 20857ffddd49SMuchun Song /* 20867ffddd49SMuchun Song * Theoretically, we should return -EBUSY when we 20877ffddd49SMuchun Song * encounter this race. In fact, we have a chance 20887ffddd49SMuchun Song * to successfully dissolve the page if we do a 20897ffddd49SMuchun Song * retry. Because the race window is quite small. 20907ffddd49SMuchun Song * If we seize this opportunity, it is an optimization 20917ffddd49SMuchun Song * for increasing the success rate of dissolving page. 20927ffddd49SMuchun Song */ 20937ffddd49SMuchun Song goto retry; 20947ffddd49SMuchun Song } 20957ffddd49SMuchun Song 2096ad2fa371SMuchun Song remove_hugetlb_page(h, head, false); 2097ad2fa371SMuchun Song h->max_huge_pages--; 2098ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2099ad2fa371SMuchun Song 2100c3114a84SAnshuman Khandual /* 2101ad2fa371SMuchun Song * Normally update_and_free_page will allocate required vmemmmap 2102ad2fa371SMuchun Song * before freeing the page. update_and_free_page will fail to 2103ad2fa371SMuchun Song * free the page if it can not allocate required vmemmap. We 2104ad2fa371SMuchun Song * need to adjust max_huge_pages if the page is not freed. 2105ad2fa371SMuchun Song * Attempt to allocate vmemmmap here so that we can take 2106ad2fa371SMuchun Song * appropriate action on failure. 2107ad2fa371SMuchun Song */ 2108ad2fa371SMuchun Song rc = alloc_huge_page_vmemmap(h, head); 2109ad2fa371SMuchun Song if (!rc) { 2110ad2fa371SMuchun Song /* 2111ad2fa371SMuchun Song * Move PageHWPoison flag from head page to the raw 2112ad2fa371SMuchun Song * error page, which makes any subpages rather than 2113ad2fa371SMuchun Song * the error page reusable. 2114c3114a84SAnshuman Khandual */ 2115c3114a84SAnshuman Khandual if (PageHWPoison(head) && page != head) { 2116c3114a84SAnshuman Khandual SetPageHWPoison(page); 2117c3114a84SAnshuman Khandual ClearPageHWPoison(head); 2118c3114a84SAnshuman Khandual } 2119b65d4adbSMuchun Song update_and_free_page(h, head, false); 2120ad2fa371SMuchun Song } else { 2121ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 2122ad2fa371SMuchun Song add_hugetlb_page(h, head, false); 2123ad2fa371SMuchun Song h->max_huge_pages++; 2124ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2125ad2fa371SMuchun Song } 2126ad2fa371SMuchun Song 2127ad2fa371SMuchun Song return rc; 2128c8721bbbSNaoya Horiguchi } 2129082d5b6bSGerald Schaefer out: 2130db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2131082d5b6bSGerald Schaefer return rc; 2132c8721bbbSNaoya Horiguchi } 2133c8721bbbSNaoya Horiguchi 2134c8721bbbSNaoya Horiguchi /* 2135c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 2136c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 21372247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 21382247bb33SGerald Schaefer * part of it lies within the given range. 2139082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 2140082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 2141c8721bbbSNaoya Horiguchi */ 2142082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 2143c8721bbbSNaoya Horiguchi { 2144c8721bbbSNaoya Horiguchi unsigned long pfn; 2145eb03aa00SGerald Schaefer struct page *page; 2146082d5b6bSGerald Schaefer int rc = 0; 2147c8721bbbSNaoya Horiguchi 2148d0177639SLi Zhong if (!hugepages_supported()) 2149082d5b6bSGerald Schaefer return rc; 2150d0177639SLi Zhong 2151eb03aa00SGerald Schaefer for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { 2152eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 2153eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 2154eb03aa00SGerald Schaefer if (rc) 2155082d5b6bSGerald Schaefer break; 2156eb03aa00SGerald Schaefer } 2157082d5b6bSGerald Schaefer 2158082d5b6bSGerald Schaefer return rc; 2159c8721bbbSNaoya Horiguchi } 2160c8721bbbSNaoya Horiguchi 2161ab5ac90aSMichal Hocko /* 2162ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 2163ab5ac90aSMichal Hocko */ 21640c397daeSMichal Hocko static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, 2165b65a4edaSMike Kravetz int nid, nodemask_t *nmask, bool zero_ref) 21667893d1d5SAdam Litke { 21679980d744SMichal Hocko struct page *page = NULL; 2168b65a4edaSMike Kravetz bool retry = false; 21697893d1d5SAdam Litke 2170bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2171aa888a74SAndi Kleen return NULL; 2172aa888a74SAndi Kleen 2173db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 21749980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 21759980d744SMichal Hocko goto out_unlock; 2176db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2177d1c3fb1fSNishanth Aravamudan 2178b65a4edaSMike Kravetz retry: 2179f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 21809980d744SMichal Hocko if (!page) 21810c397daeSMichal Hocko return NULL; 2182d1c3fb1fSNishanth Aravamudan 2183db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 21849980d744SMichal Hocko /* 21859980d744SMichal Hocko * We could have raced with the pool size change. 21869980d744SMichal Hocko * Double check that and simply deallocate the new page 21879980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 21889980d744SMichal Hocko * temporary page to workaround the nasty free_huge_page 21899980d744SMichal Hocko * codeflow 21909980d744SMichal Hocko */ 21919980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 21929157c311SMike Kravetz SetHPageTemporary(page); 2193db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 21949980d744SMichal Hocko put_page(page); 21952bf753e6SKai Shen return NULL; 2196b65a4edaSMike Kravetz } 2197b65a4edaSMike Kravetz 2198b65a4edaSMike Kravetz if (zero_ref) { 2199b65a4edaSMike Kravetz /* 2200b65a4edaSMike Kravetz * Caller requires a page with zero ref count. 2201b65a4edaSMike Kravetz * We will drop ref count here. If someone else is holding 2202b65a4edaSMike Kravetz * a ref, the page will be freed when they drop it. Abuse 2203b65a4edaSMike Kravetz * temporary page flag to accomplish this. 2204b65a4edaSMike Kravetz */ 2205b65a4edaSMike Kravetz SetHPageTemporary(page); 2206b65a4edaSMike Kravetz if (!put_page_testzero(page)) { 2207b65a4edaSMike Kravetz /* 2208b65a4edaSMike Kravetz * Unexpected inflated ref count on freshly allocated 2209b65a4edaSMike Kravetz * huge. Retry once. 2210b65a4edaSMike Kravetz */ 2211b65a4edaSMike Kravetz pr_info("HugeTLB unexpected inflated ref count on freshly allocated page\n"); 2212b65a4edaSMike Kravetz spin_unlock_irq(&hugetlb_lock); 2213b65a4edaSMike Kravetz if (retry) 2214b65a4edaSMike Kravetz return NULL; 2215b65a4edaSMike Kravetz 2216b65a4edaSMike Kravetz retry = true; 2217b65a4edaSMike Kravetz goto retry; 2218b65a4edaSMike Kravetz } 2219b65a4edaSMike Kravetz ClearHPageTemporary(page); 2220b65a4edaSMike Kravetz } 2221b65a4edaSMike Kravetz 22229980d744SMichal Hocko h->surplus_huge_pages++; 22234704dea3SMichal Hocko h->surplus_huge_pages_node[page_to_nid(page)]++; 22249980d744SMichal Hocko 22259980d744SMichal Hocko out_unlock: 2226db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 22277893d1d5SAdam Litke 22287893d1d5SAdam Litke return page; 22297893d1d5SAdam Litke } 22307893d1d5SAdam Litke 2231bbe88753SJoonsoo Kim static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, 2232ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 2233ab5ac90aSMichal Hocko { 2234ab5ac90aSMichal Hocko struct page *page; 2235ab5ac90aSMichal Hocko 2236ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 2237ab5ac90aSMichal Hocko return NULL; 2238ab5ac90aSMichal Hocko 2239f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 2240ab5ac90aSMichal Hocko if (!page) 2241ab5ac90aSMichal Hocko return NULL; 2242ab5ac90aSMichal Hocko 2243ab5ac90aSMichal Hocko /* 2244ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 2245ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 2246ab5ac90aSMichal Hocko */ 22479157c311SMike Kravetz SetHPageTemporary(page); 2248ab5ac90aSMichal Hocko 2249ab5ac90aSMichal Hocko return page; 2250ab5ac90aSMichal Hocko } 2251ab5ac90aSMichal Hocko 2252e4e574b7SAdam Litke /* 2253099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 2254099730d6SDave Hansen */ 2255e0ec90eeSDave Hansen static 22560c397daeSMichal Hocko struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, 2257099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 2258099730d6SDave Hansen { 2259cfcaa66fSBen Widawsky struct page *page = NULL; 2260aaf14e40SMichal Hocko struct mempolicy *mpol; 2261aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 2262aaf14e40SMichal Hocko int nid; 2263aaf14e40SMichal Hocko nodemask_t *nodemask; 2264aaf14e40SMichal Hocko 2265aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 2266cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 2267cfcaa66fSBen Widawsky gfp_t gfp = gfp_mask | __GFP_NOWARN; 2268cfcaa66fSBen Widawsky 2269cfcaa66fSBen Widawsky gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL); 2270cfcaa66fSBen Widawsky page = alloc_surplus_huge_page(h, gfp, nid, nodemask, false); 2271cfcaa66fSBen Widawsky 2272cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 2273cfcaa66fSBen Widawsky nodemask = NULL; 2274cfcaa66fSBen Widawsky } 2275cfcaa66fSBen Widawsky 2276cfcaa66fSBen Widawsky if (!page) 2277b65a4edaSMike Kravetz page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask, false); 2278aaf14e40SMichal Hocko mpol_cond_put(mpol); 2279aaf14e40SMichal Hocko return page; 2280099730d6SDave Hansen } 2281099730d6SDave Hansen 2282ab5ac90aSMichal Hocko /* page migration callback function */ 22833e59fcb0SMichal Hocko struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, 2284d92bbc27SJoonsoo Kim nodemask_t *nmask, gfp_t gfp_mask) 22854db9b2efSMichal Hocko { 2286db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 22874db9b2efSMichal Hocko if (h->free_huge_pages - h->resv_huge_pages > 0) { 22883e59fcb0SMichal Hocko struct page *page; 22893e59fcb0SMichal Hocko 22903e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); 22913e59fcb0SMichal Hocko if (page) { 2292db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 22933e59fcb0SMichal Hocko return page; 22944db9b2efSMichal Hocko } 22954db9b2efSMichal Hocko } 2296db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 22974db9b2efSMichal Hocko 22980c397daeSMichal Hocko return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); 22994db9b2efSMichal Hocko } 23004db9b2efSMichal Hocko 2301ebd63723SMichal Hocko /* mempolicy aware migration callback */ 2302389c8178SMichal Hocko struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, 2303389c8178SMichal Hocko unsigned long address) 2304ebd63723SMichal Hocko { 2305ebd63723SMichal Hocko struct mempolicy *mpol; 2306ebd63723SMichal Hocko nodemask_t *nodemask; 2307ebd63723SMichal Hocko struct page *page; 2308ebd63723SMichal Hocko gfp_t gfp_mask; 2309ebd63723SMichal Hocko int node; 2310ebd63723SMichal Hocko 2311ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 2312ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 2313d92bbc27SJoonsoo Kim page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); 2314ebd63723SMichal Hocko mpol_cond_put(mpol); 2315ebd63723SMichal Hocko 2316ebd63723SMichal Hocko return page; 2317ebd63723SMichal Hocko } 2318ebd63723SMichal Hocko 2319bf50bab2SNaoya Horiguchi /* 232025985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 2321e4e574b7SAdam Litke * of size 'delta'. 2322e4e574b7SAdam Litke */ 23230a4f3d1bSLiu Xiang static int gather_surplus_pages(struct hstate *h, long delta) 23241b2a1e7bSJules Irenge __must_hold(&hugetlb_lock) 2325e4e574b7SAdam Litke { 2326e4e574b7SAdam Litke struct list_head surplus_list; 2327e4e574b7SAdam Litke struct page *page, *tmp; 23280a4f3d1bSLiu Xiang int ret; 23290a4f3d1bSLiu Xiang long i; 23300a4f3d1bSLiu Xiang long needed, allocated; 233128073b02SHillf Danton bool alloc_ok = true; 2332e4e574b7SAdam Litke 23339487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2334a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 2335ac09b3a1SAdam Litke if (needed <= 0) { 2336a5516438SAndi Kleen h->resv_huge_pages += delta; 2337e4e574b7SAdam Litke return 0; 2338ac09b3a1SAdam Litke } 2339e4e574b7SAdam Litke 2340e4e574b7SAdam Litke allocated = 0; 2341e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 2342e4e574b7SAdam Litke 2343e4e574b7SAdam Litke ret = -ENOMEM; 2344e4e574b7SAdam Litke retry: 2345db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2346e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 23470c397daeSMichal Hocko page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), 2348b65a4edaSMike Kravetz NUMA_NO_NODE, NULL, true); 234928073b02SHillf Danton if (!page) { 235028073b02SHillf Danton alloc_ok = false; 235128073b02SHillf Danton break; 235228073b02SHillf Danton } 2353e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 235469ed779aSDavid Rientjes cond_resched(); 2355e4e574b7SAdam Litke } 235628073b02SHillf Danton allocated += i; 2357e4e574b7SAdam Litke 2358e4e574b7SAdam Litke /* 2359e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 2360e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 2361e4e574b7SAdam Litke */ 2362db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2363a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 2364a5516438SAndi Kleen (h->free_huge_pages + allocated); 236528073b02SHillf Danton if (needed > 0) { 236628073b02SHillf Danton if (alloc_ok) 2367e4e574b7SAdam Litke goto retry; 236828073b02SHillf Danton /* 236928073b02SHillf Danton * We were not able to allocate enough pages to 237028073b02SHillf Danton * satisfy the entire reservation so we free what 237128073b02SHillf Danton * we've allocated so far. 237228073b02SHillf Danton */ 237328073b02SHillf Danton goto free; 237428073b02SHillf Danton } 2375e4e574b7SAdam Litke /* 2376e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 237725985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 2378e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 2379ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 2380ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 2381ac09b3a1SAdam Litke * before they are reserved. 2382e4e574b7SAdam Litke */ 2383e4e574b7SAdam Litke needed += allocated; 2384a5516438SAndi Kleen h->resv_huge_pages += delta; 2385e4e574b7SAdam Litke ret = 0; 2386a9869b83SNaoya Horiguchi 238719fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 238819fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 238919fc3f0aSAdam Litke if ((--needed) < 0) 239019fc3f0aSAdam Litke break; 2391b65a4edaSMike Kravetz /* Add the page to the hugetlb allocator */ 2392a5516438SAndi Kleen enqueue_huge_page(h, page); 239319fc3f0aSAdam Litke } 239428073b02SHillf Danton free: 2395db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 239619fc3f0aSAdam Litke 2397b65a4edaSMike Kravetz /* 2398b65a4edaSMike Kravetz * Free unnecessary surplus pages to the buddy allocator. 2399b65a4edaSMike Kravetz * Pages have no ref count, call free_huge_page directly. 2400b65a4edaSMike Kravetz */ 2401c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 2402b65a4edaSMike Kravetz free_huge_page(page); 2403db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2404e4e574b7SAdam Litke 2405e4e574b7SAdam Litke return ret; 2406e4e574b7SAdam Litke } 2407e4e574b7SAdam Litke 2408e4e574b7SAdam Litke /* 2409e5bbc8a6SMike Kravetz * This routine has two main purposes: 2410e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 2411e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 2412e5bbc8a6SMike Kravetz * to the associated reservation map. 2413e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 2414e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 2415e4e574b7SAdam Litke */ 2416a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 2417a5516438SAndi Kleen unsigned long unused_resv_pages) 2418e4e574b7SAdam Litke { 2419e4e574b7SAdam Litke unsigned long nr_pages; 242010c6ec49SMike Kravetz struct page *page; 242110c6ec49SMike Kravetz LIST_HEAD(page_list); 242210c6ec49SMike Kravetz 24239487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 242410c6ec49SMike Kravetz /* Uncommit the reservation */ 242510c6ec49SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 2426e4e574b7SAdam Litke 2427aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 2428bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2429e5bbc8a6SMike Kravetz goto out; 2430aa888a74SAndi Kleen 2431e5bbc8a6SMike Kravetz /* 2432e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 2433e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 2434e5bbc8a6SMike Kravetz */ 2435a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 2436e4e574b7SAdam Litke 2437685f3457SLee Schermerhorn /* 2438685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 24399b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 24409b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 24419b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 244210c6ec49SMike Kravetz * remove_pool_huge_page() will balance the freed pages across the 24439b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 2444685f3457SLee Schermerhorn */ 2445685f3457SLee Schermerhorn while (nr_pages--) { 244610c6ec49SMike Kravetz page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1); 244710c6ec49SMike Kravetz if (!page) 2448e5bbc8a6SMike Kravetz goto out; 244910c6ec49SMike Kravetz 245010c6ec49SMike Kravetz list_add(&page->lru, &page_list); 2451e4e574b7SAdam Litke } 2452e5bbc8a6SMike Kravetz 2453e5bbc8a6SMike Kravetz out: 2454db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 245510c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 2456db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2457e4e574b7SAdam Litke } 2458e4e574b7SAdam Litke 24595e911373SMike Kravetz 2460c37f9fb1SAndy Whitcroft /* 2461feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 24625e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 2463cf3ad20bSMike Kravetz * 2464cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 2465cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 2466cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 2467cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 2468cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 2469feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 2470feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 2471feba16e2SMike Kravetz * is called in such cases. 2472cf3ad20bSMike Kravetz * 2473cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 2474cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 2475cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 2476cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 2477cf3ad20bSMike Kravetz * take appropriate action. 247896b96a96SMike Kravetz * 247996b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 248096b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 248196b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 248296b96a96SMike Kravetz * reservation exists. 2483846be085SMike Kravetz * 2484846be085SMike Kravetz * vma_del_reservation is used in error paths where an entry in the reserve 2485846be085SMike Kravetz * map was created during huge page allocation and must be removed. It is to 2486846be085SMike Kravetz * be called after calling vma_needs_reservation to determine if a reservation 2487846be085SMike Kravetz * exists. 2488c37f9fb1SAndy Whitcroft */ 24895e911373SMike Kravetz enum vma_resv_mode { 24905e911373SMike Kravetz VMA_NEEDS_RESV, 24915e911373SMike Kravetz VMA_COMMIT_RESV, 2492feba16e2SMike Kravetz VMA_END_RESV, 249396b96a96SMike Kravetz VMA_ADD_RESV, 2494846be085SMike Kravetz VMA_DEL_RESV, 24955e911373SMike Kravetz }; 2496cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 2497cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 24985e911373SMike Kravetz enum vma_resv_mode mode) 2499c37f9fb1SAndy Whitcroft { 25004e35f483SJoonsoo Kim struct resv_map *resv; 25014e35f483SJoonsoo Kim pgoff_t idx; 2502cf3ad20bSMike Kravetz long ret; 25030db9d74eSMina Almasry long dummy_out_regions_needed; 2504c37f9fb1SAndy Whitcroft 25054e35f483SJoonsoo Kim resv = vma_resv_map(vma); 25064e35f483SJoonsoo Kim if (!resv) 2507c37f9fb1SAndy Whitcroft return 1; 2508c37f9fb1SAndy Whitcroft 25094e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 25105e911373SMike Kravetz switch (mode) { 25115e911373SMike Kravetz case VMA_NEEDS_RESV: 25120db9d74eSMina Almasry ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); 25130db9d74eSMina Almasry /* We assume that vma_reservation_* routines always operate on 25140db9d74eSMina Almasry * 1 page, and that adding to resv map a 1 page entry can only 25150db9d74eSMina Almasry * ever require 1 region. 25160db9d74eSMina Almasry */ 25170db9d74eSMina Almasry VM_BUG_ON(dummy_out_regions_needed != 1); 25185e911373SMike Kravetz break; 25195e911373SMike Kravetz case VMA_COMMIT_RESV: 2520075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 25210db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 25220db9d74eSMina Almasry VM_BUG_ON(ret < 0); 25235e911373SMike Kravetz break; 2524feba16e2SMike Kravetz case VMA_END_RESV: 25250db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 25265e911373SMike Kravetz ret = 0; 25275e911373SMike Kravetz break; 252896b96a96SMike Kravetz case VMA_ADD_RESV: 25290db9d74eSMina Almasry if (vma->vm_flags & VM_MAYSHARE) { 2530075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 25310db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 25320db9d74eSMina Almasry VM_BUG_ON(ret < 0); 25330db9d74eSMina Almasry } else { 25340db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 253596b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 253696b96a96SMike Kravetz } 253796b96a96SMike Kravetz break; 2538846be085SMike Kravetz case VMA_DEL_RESV: 2539846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 2540846be085SMike Kravetz region_abort(resv, idx, idx + 1, 1); 2541846be085SMike Kravetz ret = region_del(resv, idx, idx + 1); 2542846be085SMike Kravetz } else { 2543846be085SMike Kravetz ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 2544846be085SMike Kravetz /* region_add calls of range 1 should never fail. */ 2545846be085SMike Kravetz VM_BUG_ON(ret < 0); 2546846be085SMike Kravetz } 2547846be085SMike Kravetz break; 25485e911373SMike Kravetz default: 25495e911373SMike Kravetz BUG(); 25505e911373SMike Kravetz } 255184afd99bSAndy Whitcroft 2552846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV) 2553cf3ad20bSMike Kravetz return ret; 255467961f9dSMike Kravetz /* 2555bf3d12b9SMiaohe Lin * We know private mapping must have HPAGE_RESV_OWNER set. 2556bf3d12b9SMiaohe Lin * 255767961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 255867961f9dSMike Kravetz * However, a file associated with mapping could have been 255967961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 256067961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 256167961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 256267961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 256367961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 256467961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 256567961f9dSMike Kravetz * reservation has already been consumed. As a result, the 256667961f9dSMike Kravetz * return value of this routine is the opposite of the 256767961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 256867961f9dSMike Kravetz */ 2569bf3d12b9SMiaohe Lin if (ret > 0) 257067961f9dSMike Kravetz return 0; 2571bf3d12b9SMiaohe Lin if (ret == 0) 257267961f9dSMike Kravetz return 1; 2573bf3d12b9SMiaohe Lin return ret; 257484afd99bSAndy Whitcroft } 2575cf3ad20bSMike Kravetz 2576cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 2577a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 2578c37f9fb1SAndy Whitcroft { 25795e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 2580cf3ad20bSMike Kravetz } 2581c37f9fb1SAndy Whitcroft 2582cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 2583cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 2584cf3ad20bSMike Kravetz { 25855e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 25865e911373SMike Kravetz } 25875e911373SMike Kravetz 2588feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 25895e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 25905e911373SMike Kravetz { 2591feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 2592c37f9fb1SAndy Whitcroft } 2593c37f9fb1SAndy Whitcroft 259496b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 259596b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 259696b96a96SMike Kravetz { 259796b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 259896b96a96SMike Kravetz } 259996b96a96SMike Kravetz 2600846be085SMike Kravetz static long vma_del_reservation(struct hstate *h, 2601846be085SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 260296b96a96SMike Kravetz { 2603846be085SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV); 2604846be085SMike Kravetz } 2605846be085SMike Kravetz 2606846be085SMike Kravetz /* 2607846be085SMike Kravetz * This routine is called to restore reservation information on error paths. 2608846be085SMike Kravetz * It should ONLY be called for pages allocated via alloc_huge_page(), and 2609846be085SMike Kravetz * the hugetlb mutex should remain held when calling this routine. 2610846be085SMike Kravetz * 2611846be085SMike Kravetz * It handles two specific cases: 2612846be085SMike Kravetz * 1) A reservation was in place and the page consumed the reservation. 2613846be085SMike Kravetz * HPageRestoreReserve is set in the page. 2614846be085SMike Kravetz * 2) No reservation was in place for the page, so HPageRestoreReserve is 2615846be085SMike Kravetz * not set. However, alloc_huge_page always updates the reserve map. 2616846be085SMike Kravetz * 2617846be085SMike Kravetz * In case 1, free_huge_page later in the error path will increment the 2618846be085SMike Kravetz * global reserve count. But, free_huge_page does not have enough context 2619846be085SMike Kravetz * to adjust the reservation map. This case deals primarily with private 2620846be085SMike Kravetz * mappings. Adjust the reserve map here to be consistent with global 2621846be085SMike Kravetz * reserve count adjustments to be made by free_huge_page. Make sure the 2622846be085SMike Kravetz * reserve map indicates there is a reservation present. 2623846be085SMike Kravetz * 2624846be085SMike Kravetz * In case 2, simply undo reserve map modifications done by alloc_huge_page. 2625846be085SMike Kravetz */ 2626846be085SMike Kravetz void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma, 2627846be085SMike Kravetz unsigned long address, struct page *page) 2628846be085SMike Kravetz { 262996b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 263096b96a96SMike Kravetz 2631846be085SMike Kravetz if (HPageRestoreReserve(page)) { 2632846be085SMike Kravetz if (unlikely(rc < 0)) 263396b96a96SMike Kravetz /* 263496b96a96SMike Kravetz * Rare out of memory condition in reserve map 2635d6995da3SMike Kravetz * manipulation. Clear HPageRestoreReserve so that 263696b96a96SMike Kravetz * global reserve count will not be incremented 263796b96a96SMike Kravetz * by free_huge_page. This will make it appear 263896b96a96SMike Kravetz * as though the reservation for this page was 263996b96a96SMike Kravetz * consumed. This may prevent the task from 264096b96a96SMike Kravetz * faulting in the page at a later time. This 264196b96a96SMike Kravetz * is better than inconsistent global huge page 264296b96a96SMike Kravetz * accounting of reserve counts. 264396b96a96SMike Kravetz */ 2644d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 2645846be085SMike Kravetz else if (rc) 2646846be085SMike Kravetz (void)vma_add_reservation(h, vma, address); 2647846be085SMike Kravetz else 2648846be085SMike Kravetz vma_end_reservation(h, vma, address); 2649846be085SMike Kravetz } else { 2650846be085SMike Kravetz if (!rc) { 265196b96a96SMike Kravetz /* 2652846be085SMike Kravetz * This indicates there is an entry in the reserve map 2653c7b1850dSMike Kravetz * not added by alloc_huge_page. We know it was added 2654846be085SMike Kravetz * before the alloc_huge_page call, otherwise 2655846be085SMike Kravetz * HPageRestoreReserve would be set on the page. 2656846be085SMike Kravetz * Remove the entry so that a subsequent allocation 2657846be085SMike Kravetz * does not consume a reservation. 265896b96a96SMike Kravetz */ 2659846be085SMike Kravetz rc = vma_del_reservation(h, vma, address); 2660846be085SMike Kravetz if (rc < 0) 2661846be085SMike Kravetz /* 2662846be085SMike Kravetz * VERY rare out of memory condition. Since 2663846be085SMike Kravetz * we can not delete the entry, set 2664846be085SMike Kravetz * HPageRestoreReserve so that the reserve 2665846be085SMike Kravetz * count will be incremented when the page 2666846be085SMike Kravetz * is freed. This reserve will be consumed 2667846be085SMike Kravetz * on a subsequent allocation. 2668846be085SMike Kravetz */ 2669846be085SMike Kravetz SetHPageRestoreReserve(page); 2670846be085SMike Kravetz } else if (rc < 0) { 2671846be085SMike Kravetz /* 2672846be085SMike Kravetz * Rare out of memory condition from 2673846be085SMike Kravetz * vma_needs_reservation call. Memory allocation is 2674846be085SMike Kravetz * only attempted if a new entry is needed. Therefore, 2675846be085SMike Kravetz * this implies there is not an entry in the 2676846be085SMike Kravetz * reserve map. 2677846be085SMike Kravetz * 2678846be085SMike Kravetz * For shared mappings, no entry in the map indicates 2679846be085SMike Kravetz * no reservation. We are done. 2680846be085SMike Kravetz */ 2681846be085SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 2682846be085SMike Kravetz /* 2683846be085SMike Kravetz * For private mappings, no entry indicates 2684846be085SMike Kravetz * a reservation is present. Since we can 2685846be085SMike Kravetz * not add an entry, set SetHPageRestoreReserve 2686846be085SMike Kravetz * on the page so reserve count will be 2687846be085SMike Kravetz * incremented when freed. This reserve will 2688846be085SMike Kravetz * be consumed on a subsequent allocation. 2689846be085SMike Kravetz */ 2690846be085SMike Kravetz SetHPageRestoreReserve(page); 269196b96a96SMike Kravetz } else 2692846be085SMike Kravetz /* 2693846be085SMike Kravetz * No reservation present, do nothing 2694846be085SMike Kravetz */ 269596b96a96SMike Kravetz vma_end_reservation(h, vma, address); 269696b96a96SMike Kravetz } 269796b96a96SMike Kravetz } 269896b96a96SMike Kravetz 2699369fa227SOscar Salvador /* 2700369fa227SOscar Salvador * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one 2701369fa227SOscar Salvador * @h: struct hstate old page belongs to 2702369fa227SOscar Salvador * @old_page: Old page to dissolve 2703ae37c7ffSOscar Salvador * @list: List to isolate the page in case we need to 2704369fa227SOscar Salvador * Returns 0 on success, otherwise negated error. 2705369fa227SOscar Salvador */ 2706ae37c7ffSOscar Salvador static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page, 2707ae37c7ffSOscar Salvador struct list_head *list) 2708369fa227SOscar Salvador { 2709369fa227SOscar Salvador gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2710369fa227SOscar Salvador int nid = page_to_nid(old_page); 2711b65a4edaSMike Kravetz bool alloc_retry = false; 2712369fa227SOscar Salvador struct page *new_page; 2713369fa227SOscar Salvador int ret = 0; 2714369fa227SOscar Salvador 2715369fa227SOscar Salvador /* 2716369fa227SOscar Salvador * Before dissolving the page, we need to allocate a new one for the 2717f41f2ed4SMuchun Song * pool to remain stable. Here, we allocate the page and 'prep' it 2718f41f2ed4SMuchun Song * by doing everything but actually updating counters and adding to 2719f41f2ed4SMuchun Song * the pool. This simplifies and let us do most of the processing 2720f41f2ed4SMuchun Song * under the lock. 2721369fa227SOscar Salvador */ 2722b65a4edaSMike Kravetz alloc_retry: 2723369fa227SOscar Salvador new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL); 2724369fa227SOscar Salvador if (!new_page) 2725369fa227SOscar Salvador return -ENOMEM; 2726b65a4edaSMike Kravetz /* 2727b65a4edaSMike Kravetz * If all goes well, this page will be directly added to the free 2728b65a4edaSMike Kravetz * list in the pool. For this the ref count needs to be zero. 2729b65a4edaSMike Kravetz * Attempt to drop now, and retry once if needed. It is VERY 2730b65a4edaSMike Kravetz * unlikely there is another ref on the page. 2731b65a4edaSMike Kravetz * 2732b65a4edaSMike Kravetz * If someone else has a reference to the page, it will be freed 2733b65a4edaSMike Kravetz * when they drop their ref. Abuse temporary page flag to accomplish 2734b65a4edaSMike Kravetz * this. Retry once if there is an inflated ref count. 2735b65a4edaSMike Kravetz */ 2736b65a4edaSMike Kravetz SetHPageTemporary(new_page); 2737b65a4edaSMike Kravetz if (!put_page_testzero(new_page)) { 2738b65a4edaSMike Kravetz if (alloc_retry) 2739b65a4edaSMike Kravetz return -EBUSY; 2740b65a4edaSMike Kravetz 2741b65a4edaSMike Kravetz alloc_retry = true; 2742b65a4edaSMike Kravetz goto alloc_retry; 2743b65a4edaSMike Kravetz } 2744b65a4edaSMike Kravetz ClearHPageTemporary(new_page); 2745b65a4edaSMike Kravetz 2746f41f2ed4SMuchun Song __prep_new_huge_page(h, new_page); 2747369fa227SOscar Salvador 2748369fa227SOscar Salvador retry: 2749369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2750369fa227SOscar Salvador if (!PageHuge(old_page)) { 2751369fa227SOscar Salvador /* 2752369fa227SOscar Salvador * Freed from under us. Drop new_page too. 2753369fa227SOscar Salvador */ 2754369fa227SOscar Salvador goto free_new; 2755369fa227SOscar Salvador } else if (page_count(old_page)) { 2756369fa227SOscar Salvador /* 2757ae37c7ffSOscar Salvador * Someone has grabbed the page, try to isolate it here. 2758ae37c7ffSOscar Salvador * Fail with -EBUSY if not possible. 2759369fa227SOscar Salvador */ 2760ae37c7ffSOscar Salvador spin_unlock_irq(&hugetlb_lock); 2761ae37c7ffSOscar Salvador if (!isolate_huge_page(old_page, list)) 2762369fa227SOscar Salvador ret = -EBUSY; 2763ae37c7ffSOscar Salvador spin_lock_irq(&hugetlb_lock); 2764369fa227SOscar Salvador goto free_new; 2765369fa227SOscar Salvador } else if (!HPageFreed(old_page)) { 2766369fa227SOscar Salvador /* 2767369fa227SOscar Salvador * Page's refcount is 0 but it has not been enqueued in the 2768369fa227SOscar Salvador * freelist yet. Race window is small, so we can succeed here if 2769369fa227SOscar Salvador * we retry. 2770369fa227SOscar Salvador */ 2771369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2772369fa227SOscar Salvador cond_resched(); 2773369fa227SOscar Salvador goto retry; 2774369fa227SOscar Salvador } else { 2775369fa227SOscar Salvador /* 2776369fa227SOscar Salvador * Ok, old_page is still a genuine free hugepage. Remove it from 2777369fa227SOscar Salvador * the freelist and decrease the counters. These will be 2778369fa227SOscar Salvador * incremented again when calling __prep_account_new_huge_page() 2779369fa227SOscar Salvador * and enqueue_huge_page() for new_page. The counters will remain 2780369fa227SOscar Salvador * stable since this happens under the lock. 2781369fa227SOscar Salvador */ 2782369fa227SOscar Salvador remove_hugetlb_page(h, old_page, false); 2783369fa227SOscar Salvador 2784369fa227SOscar Salvador /* 2785b65a4edaSMike Kravetz * Ref count on new page is already zero as it was dropped 2786b65a4edaSMike Kravetz * earlier. It can be directly added to the pool free list. 2787369fa227SOscar Salvador */ 2788369fa227SOscar Salvador __prep_account_new_huge_page(h, nid); 2789369fa227SOscar Salvador enqueue_huge_page(h, new_page); 2790369fa227SOscar Salvador 2791369fa227SOscar Salvador /* 2792369fa227SOscar Salvador * Pages have been replaced, we can safely free the old one. 2793369fa227SOscar Salvador */ 2794369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2795b65d4adbSMuchun Song update_and_free_page(h, old_page, false); 2796369fa227SOscar Salvador } 2797369fa227SOscar Salvador 2798369fa227SOscar Salvador return ret; 2799369fa227SOscar Salvador 2800369fa227SOscar Salvador free_new: 2801369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2802b65a4edaSMike Kravetz /* Page has a zero ref count, but needs a ref to be freed */ 2803b65a4edaSMike Kravetz set_page_refcounted(new_page); 2804b65d4adbSMuchun Song update_and_free_page(h, new_page, false); 2805369fa227SOscar Salvador 2806369fa227SOscar Salvador return ret; 2807369fa227SOscar Salvador } 2808369fa227SOscar Salvador 2809ae37c7ffSOscar Salvador int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list) 2810369fa227SOscar Salvador { 2811369fa227SOscar Salvador struct hstate *h; 2812369fa227SOscar Salvador struct page *head; 2813ae37c7ffSOscar Salvador int ret = -EBUSY; 2814369fa227SOscar Salvador 2815369fa227SOscar Salvador /* 2816369fa227SOscar Salvador * The page might have been dissolved from under our feet, so make sure 2817369fa227SOscar Salvador * to carefully check the state under the lock. 2818369fa227SOscar Salvador * Return success when racing as if we dissolved the page ourselves. 2819369fa227SOscar Salvador */ 2820369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2821369fa227SOscar Salvador if (PageHuge(page)) { 2822369fa227SOscar Salvador head = compound_head(page); 2823369fa227SOscar Salvador h = page_hstate(head); 2824369fa227SOscar Salvador } else { 2825369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2826369fa227SOscar Salvador return 0; 2827369fa227SOscar Salvador } 2828369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2829369fa227SOscar Salvador 2830369fa227SOscar Salvador /* 2831369fa227SOscar Salvador * Fence off gigantic pages as there is a cyclic dependency between 2832369fa227SOscar Salvador * alloc_contig_range and them. Return -ENOMEM as this has the effect 2833369fa227SOscar Salvador * of bailing out right away without further retrying. 2834369fa227SOscar Salvador */ 2835369fa227SOscar Salvador if (hstate_is_gigantic(h)) 2836369fa227SOscar Salvador return -ENOMEM; 2837369fa227SOscar Salvador 2838ae37c7ffSOscar Salvador if (page_count(head) && isolate_huge_page(head, list)) 2839ae37c7ffSOscar Salvador ret = 0; 2840ae37c7ffSOscar Salvador else if (!page_count(head)) 2841ae37c7ffSOscar Salvador ret = alloc_and_dissolve_huge_page(h, head, list); 2842ae37c7ffSOscar Salvador 2843ae37c7ffSOscar Salvador return ret; 2844369fa227SOscar Salvador } 2845369fa227SOscar Salvador 284670c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 284704f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 2848348ea204SAdam Litke { 284990481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 2850a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2851348ea204SAdam Litke struct page *page; 2852d85f69b0SMike Kravetz long map_chg, map_commit; 2853d85f69b0SMike Kravetz long gbl_chg; 28546d76dcf4SAneesh Kumar K.V int ret, idx; 28556d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 285608cf9fafSMina Almasry bool deferred_reserve; 28572fc39cecSAdam Litke 28586d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 2859a1e78772SMel Gorman /* 2860d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 2861d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 2862d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 2863a1e78772SMel Gorman */ 2864d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 2865d85f69b0SMike Kravetz if (map_chg < 0) 286676dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 2867d85f69b0SMike Kravetz 2868d85f69b0SMike Kravetz /* 2869d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 2870d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 2871d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 2872d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 2873d85f69b0SMike Kravetz * checked against any subpool limit. 2874d85f69b0SMike Kravetz */ 2875d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 2876d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 2877d85f69b0SMike Kravetz if (gbl_chg < 0) { 2878feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 287976dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 28805e911373SMike Kravetz } 288190d8b7e6SAdam Litke 2882d85f69b0SMike Kravetz /* 2883d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 2884d85f69b0SMike Kravetz * map, there could be reservations associated with the 2885d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 2886d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 2887d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 2888d85f69b0SMike Kravetz * the subpool reservations. 2889d85f69b0SMike Kravetz */ 2890d85f69b0SMike Kravetz if (avoid_reserve) 2891d85f69b0SMike Kravetz gbl_chg = 1; 2892d85f69b0SMike Kravetz } 2893d85f69b0SMike Kravetz 289408cf9fafSMina Almasry /* If this allocation is not consuming a reservation, charge it now. 289508cf9fafSMina Almasry */ 28966501fe5fSMiaohe Lin deferred_reserve = map_chg || avoid_reserve; 289708cf9fafSMina Almasry if (deferred_reserve) { 289808cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 289908cf9fafSMina Almasry idx, pages_per_huge_page(h), &h_cg); 29008f34af6fSJianyu Zhan if (ret) 29018f34af6fSJianyu Zhan goto out_subpool_put; 290208cf9fafSMina Almasry } 290308cf9fafSMina Almasry 290408cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 290508cf9fafSMina Almasry if (ret) 290608cf9fafSMina Almasry goto out_uncharge_cgroup_reservation; 29078f34af6fSJianyu Zhan 2908db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2909d85f69b0SMike Kravetz /* 2910d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 2911d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 2912d85f69b0SMike Kravetz * a reservation exists for the allocation. 2913d85f69b0SMike Kravetz */ 2914d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 291581a6fcaeSJoonsoo Kim if (!page) { 2916db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 29170c397daeSMichal Hocko page = alloc_buddy_huge_page_with_mpol(h, vma, addr); 29188f34af6fSJianyu Zhan if (!page) 29198f34af6fSJianyu Zhan goto out_uncharge_cgroup; 2920a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 2921d6995da3SMike Kravetz SetHPageRestoreReserve(page); 2922a88c7695SNaoya Horiguchi h->resv_huge_pages--; 2923a88c7695SNaoya Horiguchi } 2924db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 292515a8d68eSWei Yang list_add(&page->lru, &h->hugepage_activelist); 292681a6fcaeSJoonsoo Kim /* Fall through */ 2927a1e78772SMel Gorman } 292881a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 292908cf9fafSMina Almasry /* If allocation is not consuming a reservation, also store the 293008cf9fafSMina Almasry * hugetlb_cgroup pointer on the page. 293108cf9fafSMina Almasry */ 293208cf9fafSMina Almasry if (deferred_reserve) { 293308cf9fafSMina Almasry hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), 293408cf9fafSMina Almasry h_cg, page); 293508cf9fafSMina Almasry } 293608cf9fafSMina Almasry 2937db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2938a1e78772SMel Gorman 2939d6995da3SMike Kravetz hugetlb_set_page_subpool(page, spool); 2940a1e78772SMel Gorman 2941d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 2942d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 294333039678SMike Kravetz /* 294433039678SMike Kravetz * The page was added to the reservation map between 294533039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 294633039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 294733039678SMike Kravetz * Adjust for the subpool count incremented above AND 294833039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 294933039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 295033039678SMike Kravetz * no longer applies. 295133039678SMike Kravetz */ 295233039678SMike Kravetz long rsv_adjust; 295333039678SMike Kravetz 295433039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 295533039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 295679aa925bSMike Kravetz if (deferred_reserve) 295779aa925bSMike Kravetz hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 295879aa925bSMike Kravetz pages_per_huge_page(h), page); 295933039678SMike Kravetz } 29607893d1d5SAdam Litke return page; 29618f34af6fSJianyu Zhan 29628f34af6fSJianyu Zhan out_uncharge_cgroup: 29638f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 296408cf9fafSMina Almasry out_uncharge_cgroup_reservation: 296508cf9fafSMina Almasry if (deferred_reserve) 296608cf9fafSMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), 296708cf9fafSMina Almasry h_cg); 29688f34af6fSJianyu Zhan out_subpool_put: 2969d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 29708f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 2971feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 29728f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 2973b45b5bd6SDavid Gibson } 2974b45b5bd6SDavid Gibson 2975b5389086SZhenguo Yao int alloc_bootmem_huge_page(struct hstate *h, int nid) 2976e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 2977b5389086SZhenguo Yao int __alloc_bootmem_huge_page(struct hstate *h, int nid) 2978aa888a74SAndi Kleen { 2979b5389086SZhenguo Yao struct huge_bootmem_page *m = NULL; /* initialize for clang */ 2980b2261026SJoonsoo Kim int nr_nodes, node; 2981aa888a74SAndi Kleen 29824178158eSZhenguo Yao if (nid != NUMA_NO_NODE && nid >= nr_online_nodes) 2983b5389086SZhenguo Yao return 0; 2984b5389086SZhenguo Yao /* do node specific alloc */ 2985b5389086SZhenguo Yao if (nid != NUMA_NO_NODE) { 2986b5389086SZhenguo Yao m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h), 2987b5389086SZhenguo Yao 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid); 2988b5389086SZhenguo Yao if (!m) 2989b5389086SZhenguo Yao return 0; 2990b5389086SZhenguo Yao goto found; 2991b5389086SZhenguo Yao } 2992b5389086SZhenguo Yao /* allocate from next node when distributing huge pages */ 2993b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 2994b5389086SZhenguo Yao m = memblock_alloc_try_nid_raw( 29958b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 299697ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 2997aa888a74SAndi Kleen /* 2998aa888a74SAndi Kleen * Use the beginning of the huge page to store the 2999aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 3000aa888a74SAndi Kleen * puts them into the mem_map). 3001aa888a74SAndi Kleen */ 3002b5389086SZhenguo Yao if (!m) 3003b5389086SZhenguo Yao return 0; 3004aa888a74SAndi Kleen goto found; 3005aa888a74SAndi Kleen } 3006aa888a74SAndi Kleen 3007aa888a74SAndi Kleen found: 3008aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 3009330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 3010aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 3011aa888a74SAndi Kleen m->hstate = h; 3012aa888a74SAndi Kleen return 1; 3013aa888a74SAndi Kleen } 3014aa888a74SAndi Kleen 301548b8d744SMike Kravetz /* 301648b8d744SMike Kravetz * Put bootmem huge pages into the standard lists after mem_map is up. 301748b8d744SMike Kravetz * Note: This only applies to gigantic (order > MAX_ORDER) pages. 301848b8d744SMike Kravetz */ 3019aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 3020aa888a74SAndi Kleen { 3021aa888a74SAndi Kleen struct huge_bootmem_page *m; 3022aa888a74SAndi Kleen 3023aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 302440d18ebfSMike Kravetz struct page *page = virt_to_page(m); 3025aa888a74SAndi Kleen struct hstate *h = m->hstate; 3026ee8f248dSBecky Bruce 302748b8d744SMike Kravetz VM_BUG_ON(!hstate_is_gigantic(h)); 3028aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 30297118fc29SMike Kravetz if (prep_compound_gigantic_page(page, huge_page_order(h))) { 3030ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 3031aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 30327118fc29SMike Kravetz put_page(page); /* add to the hugepage allocator */ 30337118fc29SMike Kravetz } else { 3034416d85edSMike Kravetz /* VERY unlikely inflated ref count on a tail page */ 30357118fc29SMike Kravetz free_gigantic_page(page, huge_page_order(h)); 30367118fc29SMike Kravetz } 3037af0fb9dfSMichal Hocko 3038b0320c7bSRafael Aquini /* 303948b8d744SMike Kravetz * We need to restore the 'stolen' pages to totalram_pages 304048b8d744SMike Kravetz * in order to fix confusing memory reports from free(1) and 304148b8d744SMike Kravetz * other side-effects, like CommitLimit going negative. 3042b0320c7bSRafael Aquini */ 3043c78a7f36SMiaohe Lin adjust_managed_page_count(page, pages_per_huge_page(h)); 3044520495feSCannon Matthews cond_resched(); 3045aa888a74SAndi Kleen } 3046aa888a74SAndi Kleen } 3047b5389086SZhenguo Yao static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid) 3048b5389086SZhenguo Yao { 3049b5389086SZhenguo Yao unsigned long i; 3050b5389086SZhenguo Yao char buf[32]; 3051b5389086SZhenguo Yao 3052b5389086SZhenguo Yao for (i = 0; i < h->max_huge_pages_node[nid]; ++i) { 3053b5389086SZhenguo Yao if (hstate_is_gigantic(h)) { 3054b5389086SZhenguo Yao if (!alloc_bootmem_huge_page(h, nid)) 3055b5389086SZhenguo Yao break; 3056b5389086SZhenguo Yao } else { 3057b5389086SZhenguo Yao struct page *page; 3058b5389086SZhenguo Yao gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 3059b5389086SZhenguo Yao 3060b5389086SZhenguo Yao page = alloc_fresh_huge_page(h, gfp_mask, nid, 3061b5389086SZhenguo Yao &node_states[N_MEMORY], NULL); 3062b5389086SZhenguo Yao if (!page) 3063b5389086SZhenguo Yao break; 3064b5389086SZhenguo Yao put_page(page); /* free it into the hugepage allocator */ 3065b5389086SZhenguo Yao } 3066b5389086SZhenguo Yao cond_resched(); 3067b5389086SZhenguo Yao } 3068b5389086SZhenguo Yao if (i == h->max_huge_pages_node[nid]) 3069b5389086SZhenguo Yao return; 3070b5389086SZhenguo Yao 3071b5389086SZhenguo Yao string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3072b5389086SZhenguo Yao pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n", 3073b5389086SZhenguo Yao h->max_huge_pages_node[nid], buf, nid, i); 3074b5389086SZhenguo Yao h->max_huge_pages -= (h->max_huge_pages_node[nid] - i); 3075b5389086SZhenguo Yao h->max_huge_pages_node[nid] = i; 3076b5389086SZhenguo Yao } 3077aa888a74SAndi Kleen 30788faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 30791da177e4SLinus Torvalds { 30801da177e4SLinus Torvalds unsigned long i; 3081f60858f9SMike Kravetz nodemask_t *node_alloc_noretry; 3082b5389086SZhenguo Yao bool node_specific_alloc = false; 3083f60858f9SMike Kravetz 3084b5389086SZhenguo Yao /* skip gigantic hugepages allocation if hugetlb_cma enabled */ 3085b5389086SZhenguo Yao if (hstate_is_gigantic(h) && hugetlb_cma_size) { 3086b5389086SZhenguo Yao pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); 3087b5389086SZhenguo Yao return; 3088b5389086SZhenguo Yao } 3089b5389086SZhenguo Yao 3090b5389086SZhenguo Yao /* do node specific alloc */ 3091b5389086SZhenguo Yao for (i = 0; i < nr_online_nodes; i++) { 3092b5389086SZhenguo Yao if (h->max_huge_pages_node[i] > 0) { 3093b5389086SZhenguo Yao hugetlb_hstate_alloc_pages_onenode(h, i); 3094b5389086SZhenguo Yao node_specific_alloc = true; 3095b5389086SZhenguo Yao } 3096b5389086SZhenguo Yao } 3097b5389086SZhenguo Yao 3098b5389086SZhenguo Yao if (node_specific_alloc) 3099b5389086SZhenguo Yao return; 3100b5389086SZhenguo Yao 3101b5389086SZhenguo Yao /* below will do all node balanced alloc */ 3102f60858f9SMike Kravetz if (!hstate_is_gigantic(h)) { 3103f60858f9SMike Kravetz /* 3104f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3105f60858f9SMike Kravetz * Ignore errors as lower level routines can deal with 3106f60858f9SMike Kravetz * node_alloc_noretry == NULL. If this kmalloc fails at boot 3107f60858f9SMike Kravetz * time, we are likely in bigger trouble. 3108f60858f9SMike Kravetz */ 3109f60858f9SMike Kravetz node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), 3110f60858f9SMike Kravetz GFP_KERNEL); 3111f60858f9SMike Kravetz } else { 3112f60858f9SMike Kravetz /* allocations done at boot time */ 3113f60858f9SMike Kravetz node_alloc_noretry = NULL; 3114f60858f9SMike Kravetz } 3115f60858f9SMike Kravetz 3116f60858f9SMike Kravetz /* bit mask controlling how hard we retry per-node allocations */ 3117f60858f9SMike Kravetz if (node_alloc_noretry) 3118f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 31191da177e4SLinus Torvalds 3120e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 3121bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 3122b5389086SZhenguo Yao if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE)) 3123aa888a74SAndi Kleen break; 31240c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 3125f60858f9SMike Kravetz &node_states[N_MEMORY], 3126f60858f9SMike Kravetz node_alloc_noretry)) 31271da177e4SLinus Torvalds break; 312869ed779aSDavid Rientjes cond_resched(); 31291da177e4SLinus Torvalds } 3130d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 3131d715cf80SLiam R. Howlett char buf[32]; 3132d715cf80SLiam R. Howlett 3133c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3134d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 3135d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 31368faa8b07SAndi Kleen h->max_huge_pages = i; 3137e5ff2159SAndi Kleen } 3138f60858f9SMike Kravetz kfree(node_alloc_noretry); 3139d715cf80SLiam R. Howlett } 3140e5ff2159SAndi Kleen 3141e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 3142e5ff2159SAndi Kleen { 314379dfc695SMike Kravetz struct hstate *h, *h2; 3144e5ff2159SAndi Kleen 3145e5ff2159SAndi Kleen for_each_hstate(h) { 3146641844f5SNaoya Horiguchi if (minimum_order > huge_page_order(h)) 3147641844f5SNaoya Horiguchi minimum_order = huge_page_order(h); 3148641844f5SNaoya Horiguchi 31498faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 3150bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 31518faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 315279dfc695SMike Kravetz 315379dfc695SMike Kravetz /* 315479dfc695SMike Kravetz * Set demote order for each hstate. Note that 315579dfc695SMike Kravetz * h->demote_order is initially 0. 315679dfc695SMike Kravetz * - We can not demote gigantic pages if runtime freeing 315779dfc695SMike Kravetz * is not supported, so skip this. 3158a01f4390SMike Kravetz * - If CMA allocation is possible, we can not demote 3159a01f4390SMike Kravetz * HUGETLB_PAGE_ORDER or smaller size pages. 316079dfc695SMike Kravetz */ 316179dfc695SMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 316279dfc695SMike Kravetz continue; 3163a01f4390SMike Kravetz if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER) 3164a01f4390SMike Kravetz continue; 316579dfc695SMike Kravetz for_each_hstate(h2) { 316679dfc695SMike Kravetz if (h2 == h) 316779dfc695SMike Kravetz continue; 316879dfc695SMike Kravetz if (h2->order < h->order && 316979dfc695SMike Kravetz h2->order > h->demote_order) 317079dfc695SMike Kravetz h->demote_order = h2->order; 317179dfc695SMike Kravetz } 3172e5ff2159SAndi Kleen } 3173641844f5SNaoya Horiguchi VM_BUG_ON(minimum_order == UINT_MAX); 3174e5ff2159SAndi Kleen } 3175e5ff2159SAndi Kleen 3176e5ff2159SAndi Kleen static void __init report_hugepages(void) 3177e5ff2159SAndi Kleen { 3178e5ff2159SAndi Kleen struct hstate *h; 3179e5ff2159SAndi Kleen 3180e5ff2159SAndi Kleen for_each_hstate(h) { 31814abd32dbSAndi Kleen char buf[32]; 3182c6247f72SMatthew Wilcox 3183c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3184ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 3185c6247f72SMatthew Wilcox buf, h->free_huge_pages); 3186e5ff2159SAndi Kleen } 3187e5ff2159SAndi Kleen } 3188e5ff2159SAndi Kleen 31891da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 31906ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 31916ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 31921da177e4SLinus Torvalds { 31934415cc8dSChristoph Lameter int i; 31941121828aSMike Kravetz LIST_HEAD(page_list); 31954415cc8dSChristoph Lameter 31969487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 3197bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3198aa888a74SAndi Kleen return; 3199aa888a74SAndi Kleen 32001121828aSMike Kravetz /* 32011121828aSMike Kravetz * Collect pages to be freed on a list, and free after dropping lock 32021121828aSMike Kravetz */ 32036ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 320410c6ec49SMike Kravetz struct page *page, *next; 3205a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 3206a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 3207a5516438SAndi Kleen if (count >= h->nr_huge_pages) 32081121828aSMike Kravetz goto out; 32091da177e4SLinus Torvalds if (PageHighMem(page)) 32101da177e4SLinus Torvalds continue; 32116eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 32121121828aSMike Kravetz list_add(&page->lru, &page_list); 32131121828aSMike Kravetz } 32141121828aSMike Kravetz } 32151121828aSMike Kravetz 32161121828aSMike Kravetz out: 3217db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 321810c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3219db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 32201da177e4SLinus Torvalds } 32211da177e4SLinus Torvalds #else 32226ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 32236ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 32241da177e4SLinus Torvalds { 32251da177e4SLinus Torvalds } 32261da177e4SLinus Torvalds #endif 32271da177e4SLinus Torvalds 322820a0307cSWu Fengguang /* 322920a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 323020a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 323120a0307cSWu Fengguang * Returns 1 if an adjustment was made. 323220a0307cSWu Fengguang */ 32336ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 32346ae11b27SLee Schermerhorn int delta) 323520a0307cSWu Fengguang { 3236b2261026SJoonsoo Kim int nr_nodes, node; 323720a0307cSWu Fengguang 32389487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 323920a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 324020a0307cSWu Fengguang 3241e8c5c824SLee Schermerhorn if (delta < 0) { 3242b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 3243b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 3244b2261026SJoonsoo Kim goto found; 3245b2261026SJoonsoo Kim } 3246b2261026SJoonsoo Kim } else { 3247b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 3248b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 3249b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 3250b2261026SJoonsoo Kim goto found; 3251e8c5c824SLee Schermerhorn } 32529a76db09SLee Schermerhorn } 3253b2261026SJoonsoo Kim return 0; 325420a0307cSWu Fengguang 3255b2261026SJoonsoo Kim found: 325620a0307cSWu Fengguang h->surplus_huge_pages += delta; 3257b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 3258b2261026SJoonsoo Kim return 1; 325920a0307cSWu Fengguang } 326020a0307cSWu Fengguang 3261a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 3262fd875dcaSMike Kravetz static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 32636ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 32641da177e4SLinus Torvalds { 32657893d1d5SAdam Litke unsigned long min_count, ret; 326610c6ec49SMike Kravetz struct page *page; 326710c6ec49SMike Kravetz LIST_HEAD(page_list); 3268f60858f9SMike Kravetz NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); 3269f60858f9SMike Kravetz 3270f60858f9SMike Kravetz /* 3271f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3272f60858f9SMike Kravetz * If we can not allocate the bit mask, do not attempt to allocate 3273f60858f9SMike Kravetz * the requested huge pages. 3274f60858f9SMike Kravetz */ 3275f60858f9SMike Kravetz if (node_alloc_noretry) 3276f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 3277f60858f9SMike Kravetz else 3278f60858f9SMike Kravetz return -ENOMEM; 32791da177e4SLinus Torvalds 328029383967SMike Kravetz /* 328129383967SMike Kravetz * resize_lock mutex prevents concurrent adjustments to number of 328229383967SMike Kravetz * pages in hstate via the proc/sysfs interfaces. 328329383967SMike Kravetz */ 328429383967SMike Kravetz mutex_lock(&h->resize_lock); 3285b65d4adbSMuchun Song flush_free_hpage_work(h); 3286db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 32874eb0716eSAlexandre Ghiti 32884eb0716eSAlexandre Ghiti /* 3289fd875dcaSMike Kravetz * Check for a node specific request. 3290fd875dcaSMike Kravetz * Changing node specific huge page count may require a corresponding 3291fd875dcaSMike Kravetz * change to the global count. In any case, the passed node mask 3292fd875dcaSMike Kravetz * (nodes_allowed) will restrict alloc/free to the specified node. 3293fd875dcaSMike Kravetz */ 3294fd875dcaSMike Kravetz if (nid != NUMA_NO_NODE) { 3295fd875dcaSMike Kravetz unsigned long old_count = count; 3296fd875dcaSMike Kravetz 3297fd875dcaSMike Kravetz count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 3298fd875dcaSMike Kravetz /* 3299fd875dcaSMike Kravetz * User may have specified a large count value which caused the 3300fd875dcaSMike Kravetz * above calculation to overflow. In this case, they wanted 3301fd875dcaSMike Kravetz * to allocate as many huge pages as possible. Set count to 3302fd875dcaSMike Kravetz * largest possible value to align with their intention. 3303fd875dcaSMike Kravetz */ 3304fd875dcaSMike Kravetz if (count < old_count) 3305fd875dcaSMike Kravetz count = ULONG_MAX; 3306fd875dcaSMike Kravetz } 3307fd875dcaSMike Kravetz 3308fd875dcaSMike Kravetz /* 33094eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 33104eb0716eSAlexandre Ghiti * page range allocation. 33114eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 33124eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 33134eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 33144eb0716eSAlexandre Ghiti */ 33154eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 33164eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 3317db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 331829383967SMike Kravetz mutex_unlock(&h->resize_lock); 3319f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 33204eb0716eSAlexandre Ghiti return -EINVAL; 33214eb0716eSAlexandre Ghiti } 33224eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 33234eb0716eSAlexandre Ghiti } 3324aa888a74SAndi Kleen 33257893d1d5SAdam Litke /* 33267893d1d5SAdam Litke * Increase the pool size 33277893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 33287893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 3329d1c3fb1fSNishanth Aravamudan * 33300c397daeSMichal Hocko * We might race with alloc_surplus_huge_page() here and be unable 3331d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 3332d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 3333d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 3334d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 33357893d1d5SAdam Litke */ 3336a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 33376ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 33387893d1d5SAdam Litke break; 33397893d1d5SAdam Litke } 33407893d1d5SAdam Litke 3341a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 33427893d1d5SAdam Litke /* 33437893d1d5SAdam Litke * If this allocation races such that we no longer need the 33447893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 33457893d1d5SAdam Litke * and reducing the surplus. 33467893d1d5SAdam Litke */ 3347db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3348649920c6SJia He 3349649920c6SJia He /* yield cpu to avoid soft lockup */ 3350649920c6SJia He cond_resched(); 3351649920c6SJia He 3352f60858f9SMike Kravetz ret = alloc_pool_huge_page(h, nodes_allowed, 3353f60858f9SMike Kravetz node_alloc_noretry); 3354db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 33557893d1d5SAdam Litke if (!ret) 33567893d1d5SAdam Litke goto out; 33577893d1d5SAdam Litke 3358536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 3359536240f2SMel Gorman if (signal_pending(current)) 3360536240f2SMel Gorman goto out; 33617893d1d5SAdam Litke } 33627893d1d5SAdam Litke 33637893d1d5SAdam Litke /* 33647893d1d5SAdam Litke * Decrease the pool size 33657893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 33667893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 33677893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 33687893d1d5SAdam Litke * to the desired size as pages become free. 3369d1c3fb1fSNishanth Aravamudan * 3370d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 3371d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 3372d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 33730c397daeSMichal Hocko * alloc_surplus_huge_page() is checking the global counter, 3374d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 3375d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 3376d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 33777893d1d5SAdam Litke */ 3378a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 33796b0c880dSAdam Litke min_count = max(count, min_count); 33806ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 338110c6ec49SMike Kravetz 338210c6ec49SMike Kravetz /* 338310c6ec49SMike Kravetz * Collect pages to be removed on list without dropping lock 338410c6ec49SMike Kravetz */ 3385a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 338610c6ec49SMike Kravetz page = remove_pool_huge_page(h, nodes_allowed, 0); 338710c6ec49SMike Kravetz if (!page) 33881da177e4SLinus Torvalds break; 338910c6ec49SMike Kravetz 339010c6ec49SMike Kravetz list_add(&page->lru, &page_list); 33911da177e4SLinus Torvalds } 339210c6ec49SMike Kravetz /* free the pages after dropping lock */ 3393db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 339410c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3395b65d4adbSMuchun Song flush_free_hpage_work(h); 3396db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 339710c6ec49SMike Kravetz 3398a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 33996ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 34007893d1d5SAdam Litke break; 34017893d1d5SAdam Litke } 34027893d1d5SAdam Litke out: 34034eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 3404db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 340529383967SMike Kravetz mutex_unlock(&h->resize_lock); 34064eb0716eSAlexandre Ghiti 3407f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 3408f60858f9SMike Kravetz 34094eb0716eSAlexandre Ghiti return 0; 34101da177e4SLinus Torvalds } 34111da177e4SLinus Torvalds 34128531fc6fSMike Kravetz static int demote_free_huge_page(struct hstate *h, struct page *page) 34138531fc6fSMike Kravetz { 34148531fc6fSMike Kravetz int i, nid = page_to_nid(page); 34158531fc6fSMike Kravetz struct hstate *target_hstate; 34168531fc6fSMike Kravetz int rc = 0; 34178531fc6fSMike Kravetz 34188531fc6fSMike Kravetz target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order); 34198531fc6fSMike Kravetz 34208531fc6fSMike Kravetz remove_hugetlb_page_for_demote(h, page, false); 34218531fc6fSMike Kravetz spin_unlock_irq(&hugetlb_lock); 34228531fc6fSMike Kravetz 34238531fc6fSMike Kravetz rc = alloc_huge_page_vmemmap(h, page); 34248531fc6fSMike Kravetz if (rc) { 34258531fc6fSMike Kravetz /* Allocation of vmemmmap failed, we can not demote page */ 34268531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 34278531fc6fSMike Kravetz set_page_refcounted(page); 34288531fc6fSMike Kravetz add_hugetlb_page(h, page, false); 34298531fc6fSMike Kravetz return rc; 34308531fc6fSMike Kravetz } 34318531fc6fSMike Kravetz 34328531fc6fSMike Kravetz /* 34338531fc6fSMike Kravetz * Use destroy_compound_hugetlb_page_for_demote for all huge page 34348531fc6fSMike Kravetz * sizes as it will not ref count pages. 34358531fc6fSMike Kravetz */ 34368531fc6fSMike Kravetz destroy_compound_hugetlb_page_for_demote(page, huge_page_order(h)); 34378531fc6fSMike Kravetz 34388531fc6fSMike Kravetz /* 34398531fc6fSMike Kravetz * Taking target hstate mutex synchronizes with set_max_huge_pages. 34408531fc6fSMike Kravetz * Without the mutex, pages added to target hstate could be marked 34418531fc6fSMike Kravetz * as surplus. 34428531fc6fSMike Kravetz * 34438531fc6fSMike Kravetz * Note that we already hold h->resize_lock. To prevent deadlock, 34448531fc6fSMike Kravetz * use the convention of always taking larger size hstate mutex first. 34458531fc6fSMike Kravetz */ 34468531fc6fSMike Kravetz mutex_lock(&target_hstate->resize_lock); 34478531fc6fSMike Kravetz for (i = 0; i < pages_per_huge_page(h); 34488531fc6fSMike Kravetz i += pages_per_huge_page(target_hstate)) { 34498531fc6fSMike Kravetz if (hstate_is_gigantic(target_hstate)) 34508531fc6fSMike Kravetz prep_compound_gigantic_page_for_demote(page + i, 34518531fc6fSMike Kravetz target_hstate->order); 34528531fc6fSMike Kravetz else 34538531fc6fSMike Kravetz prep_compound_page(page + i, target_hstate->order); 34548531fc6fSMike Kravetz set_page_private(page + i, 0); 34558531fc6fSMike Kravetz set_page_refcounted(page + i); 34568531fc6fSMike Kravetz prep_new_huge_page(target_hstate, page + i, nid); 34578531fc6fSMike Kravetz put_page(page + i); 34588531fc6fSMike Kravetz } 34598531fc6fSMike Kravetz mutex_unlock(&target_hstate->resize_lock); 34608531fc6fSMike Kravetz 34618531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 34628531fc6fSMike Kravetz 34638531fc6fSMike Kravetz /* 34648531fc6fSMike Kravetz * Not absolutely necessary, but for consistency update max_huge_pages 34658531fc6fSMike Kravetz * based on pool changes for the demoted page. 34668531fc6fSMike Kravetz */ 34678531fc6fSMike Kravetz h->max_huge_pages--; 34688531fc6fSMike Kravetz target_hstate->max_huge_pages += pages_per_huge_page(h); 34698531fc6fSMike Kravetz 34708531fc6fSMike Kravetz return rc; 34718531fc6fSMike Kravetz } 34728531fc6fSMike Kravetz 347379dfc695SMike Kravetz static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 347479dfc695SMike Kravetz __must_hold(&hugetlb_lock) 347579dfc695SMike Kravetz { 34768531fc6fSMike Kravetz int nr_nodes, node; 34778531fc6fSMike Kravetz struct page *page; 347879dfc695SMike Kravetz 347979dfc695SMike Kravetz lockdep_assert_held(&hugetlb_lock); 348079dfc695SMike Kravetz 348179dfc695SMike Kravetz /* We should never get here if no demote order */ 348279dfc695SMike Kravetz if (!h->demote_order) { 348379dfc695SMike Kravetz pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n"); 348479dfc695SMike Kravetz return -EINVAL; /* internal error */ 348579dfc695SMike Kravetz } 348679dfc695SMike Kravetz 34878531fc6fSMike Kravetz for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 3488*5a317412SMike Kravetz list_for_each_entry(page, &h->hugepage_freelists[node], lru) { 3489*5a317412SMike Kravetz if (PageHWPoison(page)) 3490*5a317412SMike Kravetz continue; 3491*5a317412SMike Kravetz 3492*5a317412SMike Kravetz return demote_free_huge_page(h, page); 34938531fc6fSMike Kravetz } 34948531fc6fSMike Kravetz } 34958531fc6fSMike Kravetz 3496*5a317412SMike Kravetz /* 3497*5a317412SMike Kravetz * Only way to get here is if all pages on free lists are poisoned. 3498*5a317412SMike Kravetz * Return -EBUSY so that caller will not retry. 3499*5a317412SMike Kravetz */ 3500*5a317412SMike Kravetz return -EBUSY; 350179dfc695SMike Kravetz } 350279dfc695SMike Kravetz 3503a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 3504a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 3505a3437870SNishanth Aravamudan 350679dfc695SMike Kravetz #define HSTATE_ATTR_WO(_name) \ 350779dfc695SMike Kravetz static struct kobj_attribute _name##_attr = __ATTR_WO(_name) 350879dfc695SMike Kravetz 3509a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 351098bc26acSMiaohe Lin static struct kobj_attribute _name##_attr = __ATTR_RW(_name) 3511a3437870SNishanth Aravamudan 3512a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 3513a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 3514a3437870SNishanth Aravamudan 35159a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 35169a305230SLee Schermerhorn 35179a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 3518a3437870SNishanth Aravamudan { 3519a3437870SNishanth Aravamudan int i; 35209a305230SLee Schermerhorn 3521a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 35229a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 35239a305230SLee Schermerhorn if (nidp) 35249a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 3525a3437870SNishanth Aravamudan return &hstates[i]; 35269a305230SLee Schermerhorn } 35279a305230SLee Schermerhorn 35289a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 3529a3437870SNishanth Aravamudan } 3530a3437870SNishanth Aravamudan 353106808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 3532a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3533a3437870SNishanth Aravamudan { 35349a305230SLee Schermerhorn struct hstate *h; 35359a305230SLee Schermerhorn unsigned long nr_huge_pages; 35369a305230SLee Schermerhorn int nid; 35379a305230SLee Schermerhorn 35389a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 35399a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 35409a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 35419a305230SLee Schermerhorn else 35429a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 35439a305230SLee Schermerhorn 3544ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", nr_huge_pages); 3545a3437870SNishanth Aravamudan } 3546adbe8726SEric B Munson 3547238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 3548238d3c13SDavid Rientjes struct hstate *h, int nid, 3549238d3c13SDavid Rientjes unsigned long count, size_t len) 3550a3437870SNishanth Aravamudan { 3551a3437870SNishanth Aravamudan int err; 35522d0adf7eSOscar Salvador nodemask_t nodes_allowed, *n_mask; 3553a3437870SNishanth Aravamudan 35542d0adf7eSOscar Salvador if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 35552d0adf7eSOscar Salvador return -EINVAL; 3556adbe8726SEric B Munson 35579a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 35589a305230SLee Schermerhorn /* 35599a305230SLee Schermerhorn * global hstate attribute 35609a305230SLee Schermerhorn */ 35619a305230SLee Schermerhorn if (!(obey_mempolicy && 35622d0adf7eSOscar Salvador init_nodemask_of_mempolicy(&nodes_allowed))) 35632d0adf7eSOscar Salvador n_mask = &node_states[N_MEMORY]; 35642d0adf7eSOscar Salvador else 35652d0adf7eSOscar Salvador n_mask = &nodes_allowed; 35662d0adf7eSOscar Salvador } else { 35679a305230SLee Schermerhorn /* 3568fd875dcaSMike Kravetz * Node specific request. count adjustment happens in 3569fd875dcaSMike Kravetz * set_max_huge_pages() after acquiring hugetlb_lock. 35709a305230SLee Schermerhorn */ 35712d0adf7eSOscar Salvador init_nodemask_of_node(&nodes_allowed, nid); 35722d0adf7eSOscar Salvador n_mask = &nodes_allowed; 3573fd875dcaSMike Kravetz } 35749a305230SLee Schermerhorn 35752d0adf7eSOscar Salvador err = set_max_huge_pages(h, count, nid, n_mask); 357606808b08SLee Schermerhorn 35774eb0716eSAlexandre Ghiti return err ? err : len; 357806808b08SLee Schermerhorn } 357906808b08SLee Schermerhorn 3580238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 3581238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 3582238d3c13SDavid Rientjes size_t len) 3583238d3c13SDavid Rientjes { 3584238d3c13SDavid Rientjes struct hstate *h; 3585238d3c13SDavid Rientjes unsigned long count; 3586238d3c13SDavid Rientjes int nid; 3587238d3c13SDavid Rientjes int err; 3588238d3c13SDavid Rientjes 3589238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 3590238d3c13SDavid Rientjes if (err) 3591238d3c13SDavid Rientjes return err; 3592238d3c13SDavid Rientjes 3593238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 3594238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 3595238d3c13SDavid Rientjes } 3596238d3c13SDavid Rientjes 359706808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 359806808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 359906808b08SLee Schermerhorn { 360006808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 360106808b08SLee Schermerhorn } 360206808b08SLee Schermerhorn 360306808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 360406808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 360506808b08SLee Schermerhorn { 3606238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 3607a3437870SNishanth Aravamudan } 3608a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 3609a3437870SNishanth Aravamudan 361006808b08SLee Schermerhorn #ifdef CONFIG_NUMA 361106808b08SLee Schermerhorn 361206808b08SLee Schermerhorn /* 361306808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 361406808b08SLee Schermerhorn * huge page alloc/free. 361506808b08SLee Schermerhorn */ 361606808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 3617ae7a927dSJoe Perches struct kobj_attribute *attr, 3618ae7a927dSJoe Perches char *buf) 361906808b08SLee Schermerhorn { 362006808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 362106808b08SLee Schermerhorn } 362206808b08SLee Schermerhorn 362306808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 362406808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 362506808b08SLee Schermerhorn { 3626238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 362706808b08SLee Schermerhorn } 362806808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 362906808b08SLee Schermerhorn #endif 363006808b08SLee Schermerhorn 363106808b08SLee Schermerhorn 3632a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 3633a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3634a3437870SNishanth Aravamudan { 36359a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3636ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); 3637a3437870SNishanth Aravamudan } 3638adbe8726SEric B Munson 3639a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 3640a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 3641a3437870SNishanth Aravamudan { 3642a3437870SNishanth Aravamudan int err; 3643a3437870SNishanth Aravamudan unsigned long input; 36449a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3645a3437870SNishanth Aravamudan 3646bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3647adbe8726SEric B Munson return -EINVAL; 3648adbe8726SEric B Munson 36493dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 3650a3437870SNishanth Aravamudan if (err) 365173ae31e5SEric B Munson return err; 3652a3437870SNishanth Aravamudan 3653db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3654a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 3655db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3656a3437870SNishanth Aravamudan 3657a3437870SNishanth Aravamudan return count; 3658a3437870SNishanth Aravamudan } 3659a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 3660a3437870SNishanth Aravamudan 3661a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 3662a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3663a3437870SNishanth Aravamudan { 36649a305230SLee Schermerhorn struct hstate *h; 36659a305230SLee Schermerhorn unsigned long free_huge_pages; 36669a305230SLee Schermerhorn int nid; 36679a305230SLee Schermerhorn 36689a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 36699a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 36709a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 36719a305230SLee Schermerhorn else 36729a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 36739a305230SLee Schermerhorn 3674ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", free_huge_pages); 3675a3437870SNishanth Aravamudan } 3676a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 3677a3437870SNishanth Aravamudan 3678a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 3679a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3680a3437870SNishanth Aravamudan { 36819a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3682ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); 3683a3437870SNishanth Aravamudan } 3684a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 3685a3437870SNishanth Aravamudan 3686a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 3687a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3688a3437870SNishanth Aravamudan { 36899a305230SLee Schermerhorn struct hstate *h; 36909a305230SLee Schermerhorn unsigned long surplus_huge_pages; 36919a305230SLee Schermerhorn int nid; 36929a305230SLee Schermerhorn 36939a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 36949a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 36959a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 36969a305230SLee Schermerhorn else 36979a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 36989a305230SLee Schermerhorn 3699ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", surplus_huge_pages); 3700a3437870SNishanth Aravamudan } 3701a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 3702a3437870SNishanth Aravamudan 370379dfc695SMike Kravetz static ssize_t demote_store(struct kobject *kobj, 370479dfc695SMike Kravetz struct kobj_attribute *attr, const char *buf, size_t len) 370579dfc695SMike Kravetz { 370679dfc695SMike Kravetz unsigned long nr_demote; 370779dfc695SMike Kravetz unsigned long nr_available; 370879dfc695SMike Kravetz nodemask_t nodes_allowed, *n_mask; 370979dfc695SMike Kravetz struct hstate *h; 371079dfc695SMike Kravetz int err = 0; 371179dfc695SMike Kravetz int nid; 371279dfc695SMike Kravetz 371379dfc695SMike Kravetz err = kstrtoul(buf, 10, &nr_demote); 371479dfc695SMike Kravetz if (err) 371579dfc695SMike Kravetz return err; 371679dfc695SMike Kravetz h = kobj_to_hstate(kobj, &nid); 371779dfc695SMike Kravetz 371879dfc695SMike Kravetz if (nid != NUMA_NO_NODE) { 371979dfc695SMike Kravetz init_nodemask_of_node(&nodes_allowed, nid); 372079dfc695SMike Kravetz n_mask = &nodes_allowed; 372179dfc695SMike Kravetz } else { 372279dfc695SMike Kravetz n_mask = &node_states[N_MEMORY]; 372379dfc695SMike Kravetz } 372479dfc695SMike Kravetz 372579dfc695SMike Kravetz /* Synchronize with other sysfs operations modifying huge pages */ 372679dfc695SMike Kravetz mutex_lock(&h->resize_lock); 372779dfc695SMike Kravetz spin_lock_irq(&hugetlb_lock); 372879dfc695SMike Kravetz 372979dfc695SMike Kravetz while (nr_demote) { 373079dfc695SMike Kravetz /* 373179dfc695SMike Kravetz * Check for available pages to demote each time thorough the 373279dfc695SMike Kravetz * loop as demote_pool_huge_page will drop hugetlb_lock. 373379dfc695SMike Kravetz */ 373479dfc695SMike Kravetz if (nid != NUMA_NO_NODE) 373579dfc695SMike Kravetz nr_available = h->free_huge_pages_node[nid]; 373679dfc695SMike Kravetz else 373779dfc695SMike Kravetz nr_available = h->free_huge_pages; 373879dfc695SMike Kravetz nr_available -= h->resv_huge_pages; 373979dfc695SMike Kravetz if (!nr_available) 374079dfc695SMike Kravetz break; 374179dfc695SMike Kravetz 374279dfc695SMike Kravetz err = demote_pool_huge_page(h, n_mask); 374379dfc695SMike Kravetz if (err) 374479dfc695SMike Kravetz break; 374579dfc695SMike Kravetz 374679dfc695SMike Kravetz nr_demote--; 374779dfc695SMike Kravetz } 374879dfc695SMike Kravetz 374979dfc695SMike Kravetz spin_unlock_irq(&hugetlb_lock); 375079dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 375179dfc695SMike Kravetz 375279dfc695SMike Kravetz if (err) 375379dfc695SMike Kravetz return err; 375479dfc695SMike Kravetz return len; 375579dfc695SMike Kravetz } 375679dfc695SMike Kravetz HSTATE_ATTR_WO(demote); 375779dfc695SMike Kravetz 375879dfc695SMike Kravetz static ssize_t demote_size_show(struct kobject *kobj, 375979dfc695SMike Kravetz struct kobj_attribute *attr, char *buf) 376079dfc695SMike Kravetz { 376179dfc695SMike Kravetz int nid; 376279dfc695SMike Kravetz struct hstate *h = kobj_to_hstate(kobj, &nid); 376379dfc695SMike Kravetz unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K; 376479dfc695SMike Kravetz 376579dfc695SMike Kravetz return sysfs_emit(buf, "%lukB\n", demote_size); 376679dfc695SMike Kravetz } 376779dfc695SMike Kravetz 376879dfc695SMike Kravetz static ssize_t demote_size_store(struct kobject *kobj, 376979dfc695SMike Kravetz struct kobj_attribute *attr, 377079dfc695SMike Kravetz const char *buf, size_t count) 377179dfc695SMike Kravetz { 377279dfc695SMike Kravetz struct hstate *h, *demote_hstate; 377379dfc695SMike Kravetz unsigned long demote_size; 377479dfc695SMike Kravetz unsigned int demote_order; 377579dfc695SMike Kravetz int nid; 377679dfc695SMike Kravetz 377779dfc695SMike Kravetz demote_size = (unsigned long)memparse(buf, NULL); 377879dfc695SMike Kravetz 377979dfc695SMike Kravetz demote_hstate = size_to_hstate(demote_size); 378079dfc695SMike Kravetz if (!demote_hstate) 378179dfc695SMike Kravetz return -EINVAL; 378279dfc695SMike Kravetz demote_order = demote_hstate->order; 3783a01f4390SMike Kravetz if (demote_order < HUGETLB_PAGE_ORDER) 3784a01f4390SMike Kravetz return -EINVAL; 378579dfc695SMike Kravetz 378679dfc695SMike Kravetz /* demote order must be smaller than hstate order */ 378779dfc695SMike Kravetz h = kobj_to_hstate(kobj, &nid); 378879dfc695SMike Kravetz if (demote_order >= h->order) 378979dfc695SMike Kravetz return -EINVAL; 379079dfc695SMike Kravetz 379179dfc695SMike Kravetz /* resize_lock synchronizes access to demote size and writes */ 379279dfc695SMike Kravetz mutex_lock(&h->resize_lock); 379379dfc695SMike Kravetz h->demote_order = demote_order; 379479dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 379579dfc695SMike Kravetz 379679dfc695SMike Kravetz return count; 379779dfc695SMike Kravetz } 379879dfc695SMike Kravetz HSTATE_ATTR(demote_size); 379979dfc695SMike Kravetz 3800a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 3801a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 3802a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 3803a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 3804a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 3805a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 380606808b08SLee Schermerhorn #ifdef CONFIG_NUMA 380706808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 380806808b08SLee Schermerhorn #endif 3809a3437870SNishanth Aravamudan NULL, 3810a3437870SNishanth Aravamudan }; 3811a3437870SNishanth Aravamudan 381267e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 3813a3437870SNishanth Aravamudan .attrs = hstate_attrs, 3814a3437870SNishanth Aravamudan }; 3815a3437870SNishanth Aravamudan 381679dfc695SMike Kravetz static struct attribute *hstate_demote_attrs[] = { 381779dfc695SMike Kravetz &demote_size_attr.attr, 381879dfc695SMike Kravetz &demote_attr.attr, 381979dfc695SMike Kravetz NULL, 382079dfc695SMike Kravetz }; 382179dfc695SMike Kravetz 382279dfc695SMike Kravetz static const struct attribute_group hstate_demote_attr_group = { 382379dfc695SMike Kravetz .attrs = hstate_demote_attrs, 382479dfc695SMike Kravetz }; 382579dfc695SMike Kravetz 3826094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 38279a305230SLee Schermerhorn struct kobject **hstate_kobjs, 382867e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 3829a3437870SNishanth Aravamudan { 3830a3437870SNishanth Aravamudan int retval; 3831972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 3832a3437870SNishanth Aravamudan 38339a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 38349a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 3835a3437870SNishanth Aravamudan return -ENOMEM; 3836a3437870SNishanth Aravamudan 38379a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 3838cc2205a6SMiaohe Lin if (retval) { 38399a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 3840cc2205a6SMiaohe Lin hstate_kobjs[hi] = NULL; 3841cc2205a6SMiaohe Lin } 3842a3437870SNishanth Aravamudan 384379dfc695SMike Kravetz if (h->demote_order) { 384479dfc695SMike Kravetz if (sysfs_create_group(hstate_kobjs[hi], 384579dfc695SMike Kravetz &hstate_demote_attr_group)) 384679dfc695SMike Kravetz pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name); 384779dfc695SMike Kravetz } 384879dfc695SMike Kravetz 3849a3437870SNishanth Aravamudan return retval; 3850a3437870SNishanth Aravamudan } 3851a3437870SNishanth Aravamudan 3852a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 3853a3437870SNishanth Aravamudan { 3854a3437870SNishanth Aravamudan struct hstate *h; 3855a3437870SNishanth Aravamudan int err; 3856a3437870SNishanth Aravamudan 3857a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 3858a3437870SNishanth Aravamudan if (!hugepages_kobj) 3859a3437870SNishanth Aravamudan return; 3860a3437870SNishanth Aravamudan 3861a3437870SNishanth Aravamudan for_each_hstate(h) { 38629a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 38639a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 3864a3437870SNishanth Aravamudan if (err) 3865282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s", h->name); 3866a3437870SNishanth Aravamudan } 3867a3437870SNishanth Aravamudan } 3868a3437870SNishanth Aravamudan 38699a305230SLee Schermerhorn #ifdef CONFIG_NUMA 38709a305230SLee Schermerhorn 38719a305230SLee Schermerhorn /* 38729a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 387310fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 387410fbcf4cSKay Sievers * index of a node device or _hstate == node id. 387510fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 38769a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 38779a305230SLee Schermerhorn */ 38789a305230SLee Schermerhorn struct node_hstate { 38799a305230SLee Schermerhorn struct kobject *hugepages_kobj; 38809a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 38819a305230SLee Schermerhorn }; 3882b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 38839a305230SLee Schermerhorn 38849a305230SLee Schermerhorn /* 388510fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 38869a305230SLee Schermerhorn */ 38879a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 38889a305230SLee Schermerhorn &nr_hugepages_attr.attr, 38899a305230SLee Schermerhorn &free_hugepages_attr.attr, 38909a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 38919a305230SLee Schermerhorn NULL, 38929a305230SLee Schermerhorn }; 38939a305230SLee Schermerhorn 389467e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 38959a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 38969a305230SLee Schermerhorn }; 38979a305230SLee Schermerhorn 38989a305230SLee Schermerhorn /* 389910fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 39009a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 39019a305230SLee Schermerhorn */ 39029a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 39039a305230SLee Schermerhorn { 39049a305230SLee Schermerhorn int nid; 39059a305230SLee Schermerhorn 39069a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 39079a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 39089a305230SLee Schermerhorn int i; 39099a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 39109a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 39119a305230SLee Schermerhorn if (nidp) 39129a305230SLee Schermerhorn *nidp = nid; 39139a305230SLee Schermerhorn return &hstates[i]; 39149a305230SLee Schermerhorn } 39159a305230SLee Schermerhorn } 39169a305230SLee Schermerhorn 39179a305230SLee Schermerhorn BUG(); 39189a305230SLee Schermerhorn return NULL; 39199a305230SLee Schermerhorn } 39209a305230SLee Schermerhorn 39219a305230SLee Schermerhorn /* 392210fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 39239a305230SLee Schermerhorn * No-op if no hstate attributes attached. 39249a305230SLee Schermerhorn */ 39253cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 39269a305230SLee Schermerhorn { 39279a305230SLee Schermerhorn struct hstate *h; 392810fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 39299a305230SLee Schermerhorn 39309a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 39319b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 39329a305230SLee Schermerhorn 3933972dc4deSAneesh Kumar K.V for_each_hstate(h) { 3934972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 3935972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 3936972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 3937972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 3938972dc4deSAneesh Kumar K.V } 39399a305230SLee Schermerhorn } 39409a305230SLee Schermerhorn 39419a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 39429a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 39439a305230SLee Schermerhorn } 39449a305230SLee Schermerhorn 39459a305230SLee Schermerhorn 39469a305230SLee Schermerhorn /* 394710fbcf4cSKay Sievers * Register hstate attributes for a single node device. 39489a305230SLee Schermerhorn * No-op if attributes already registered. 39499a305230SLee Schermerhorn */ 39503cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 39519a305230SLee Schermerhorn { 39529a305230SLee Schermerhorn struct hstate *h; 395310fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 39549a305230SLee Schermerhorn int err; 39559a305230SLee Schermerhorn 39569a305230SLee Schermerhorn if (nhs->hugepages_kobj) 39579a305230SLee Schermerhorn return; /* already allocated */ 39589a305230SLee Schermerhorn 39599a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 396010fbcf4cSKay Sievers &node->dev.kobj); 39619a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 39629a305230SLee Schermerhorn return; 39639a305230SLee Schermerhorn 39649a305230SLee Schermerhorn for_each_hstate(h) { 39659a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 39669a305230SLee Schermerhorn nhs->hstate_kobjs, 39679a305230SLee Schermerhorn &per_node_hstate_attr_group); 39689a305230SLee Schermerhorn if (err) { 3969282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s for node %d\n", 397010fbcf4cSKay Sievers h->name, node->dev.id); 39719a305230SLee Schermerhorn hugetlb_unregister_node(node); 39729a305230SLee Schermerhorn break; 39739a305230SLee Schermerhorn } 39749a305230SLee Schermerhorn } 39759a305230SLee Schermerhorn } 39769a305230SLee Schermerhorn 39779a305230SLee Schermerhorn /* 39789b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 397910fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 398010fbcf4cSKay Sievers * registered their associated device by this time. 39819a305230SLee Schermerhorn */ 39827d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 39839a305230SLee Schermerhorn { 39849a305230SLee Schermerhorn int nid; 39859a305230SLee Schermerhorn 39868cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 39878732794bSWen Congyang struct node *node = node_devices[nid]; 398810fbcf4cSKay Sievers if (node->dev.id == nid) 39899a305230SLee Schermerhorn hugetlb_register_node(node); 39909a305230SLee Schermerhorn } 39919a305230SLee Schermerhorn 39929a305230SLee Schermerhorn /* 399310fbcf4cSKay Sievers * Let the node device driver know we're here so it can 39949a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 39959a305230SLee Schermerhorn */ 39969a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 39979a305230SLee Schermerhorn hugetlb_unregister_node); 39989a305230SLee Schermerhorn } 39999a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 40009a305230SLee Schermerhorn 40019a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 40029a305230SLee Schermerhorn { 40039a305230SLee Schermerhorn BUG(); 40049a305230SLee Schermerhorn if (nidp) 40059a305230SLee Schermerhorn *nidp = -1; 40069a305230SLee Schermerhorn return NULL; 40079a305230SLee Schermerhorn } 40089a305230SLee Schermerhorn 40099a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 40109a305230SLee Schermerhorn 40119a305230SLee Schermerhorn #endif 40129a305230SLee Schermerhorn 4013a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 4014a3437870SNishanth Aravamudan { 40158382d914SDavidlohr Bueso int i; 40168382d914SDavidlohr Bueso 4017d6995da3SMike Kravetz BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < 4018d6995da3SMike Kravetz __NR_HPAGEFLAGS); 4019d6995da3SMike Kravetz 4020c2833a5bSMike Kravetz if (!hugepages_supported()) { 4021c2833a5bSMike Kravetz if (hugetlb_max_hstate || default_hstate_max_huge_pages) 4022c2833a5bSMike Kravetz pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); 40230ef89d25SBenjamin Herrenschmidt return 0; 4024d715cf80SLiam R. Howlett } 4025d715cf80SLiam R. Howlett 4026282f4214SMike Kravetz /* 4027282f4214SMike Kravetz * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some 4028282f4214SMike Kravetz * architectures depend on setup being done here. 4029282f4214SMike Kravetz */ 4030a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 4031282f4214SMike Kravetz if (!parsed_default_hugepagesz) { 4032282f4214SMike Kravetz /* 4033282f4214SMike Kravetz * If we did not parse a default huge page size, set 4034282f4214SMike Kravetz * default_hstate_idx to HPAGE_SIZE hstate. And, if the 4035282f4214SMike Kravetz * number of huge pages for this default size was implicitly 4036282f4214SMike Kravetz * specified, set that here as well. 4037282f4214SMike Kravetz * Note that the implicit setting will overwrite an explicit 4038282f4214SMike Kravetz * setting. A warning will be printed in this case. 4039282f4214SMike Kravetz */ 4040282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); 4041f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 4042282f4214SMike Kravetz if (default_hstate.max_huge_pages) { 4043282f4214SMike Kravetz char buf[32]; 4044282f4214SMike Kravetz 4045282f4214SMike Kravetz string_get_size(huge_page_size(&default_hstate), 4046282f4214SMike Kravetz 1, STRING_UNITS_2, buf, 32); 4047282f4214SMike Kravetz pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", 4048282f4214SMike Kravetz default_hstate.max_huge_pages, buf); 4049282f4214SMike Kravetz pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", 4050282f4214SMike Kravetz default_hstate_max_huge_pages); 4051282f4214SMike Kravetz } 4052282f4214SMike Kravetz default_hstate.max_huge_pages = 4053282f4214SMike Kravetz default_hstate_max_huge_pages; 4054b5389086SZhenguo Yao 4055b5389086SZhenguo Yao for (i = 0; i < nr_online_nodes; i++) 4056b5389086SZhenguo Yao default_hstate.max_huge_pages_node[i] = 4057b5389086SZhenguo Yao default_hugepages_in_node[i]; 4058282f4214SMike Kravetz } 4059f8b74815SVaishali Thakkar } 4060a3437870SNishanth Aravamudan 4061cf11e85fSRoman Gushchin hugetlb_cma_check(); 4062a3437870SNishanth Aravamudan hugetlb_init_hstates(); 4063aa888a74SAndi Kleen gather_bootmem_prealloc(); 4064a3437870SNishanth Aravamudan report_hugepages(); 4065a3437870SNishanth Aravamudan 4066a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 40679a305230SLee Schermerhorn hugetlb_register_all_nodes(); 40687179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 40699a305230SLee Schermerhorn 40708382d914SDavidlohr Bueso #ifdef CONFIG_SMP 40718382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 40728382d914SDavidlohr Bueso #else 40738382d914SDavidlohr Bueso num_fault_mutexes = 1; 40748382d914SDavidlohr Bueso #endif 4075c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 40766da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 40776da2ec56SKees Cook GFP_KERNEL); 4078c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 40798382d914SDavidlohr Bueso 40808382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 4081c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 4082a3437870SNishanth Aravamudan return 0; 4083a3437870SNishanth Aravamudan } 40843e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 4085a3437870SNishanth Aravamudan 4086ae94da89SMike Kravetz /* Overwritten by architectures with more huge page sizes */ 4087ae94da89SMike Kravetz bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) 40889fee021dSVaishali Thakkar { 4089ae94da89SMike Kravetz return size == HPAGE_SIZE; 40909fee021dSVaishali Thakkar } 40919fee021dSVaishali Thakkar 4092d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 4093a3437870SNishanth Aravamudan { 4094a3437870SNishanth Aravamudan struct hstate *h; 40958faa8b07SAndi Kleen unsigned long i; 40968faa8b07SAndi Kleen 4097a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 4098a3437870SNishanth Aravamudan return; 4099a3437870SNishanth Aravamudan } 410047d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 4101a3437870SNishanth Aravamudan BUG_ON(order == 0); 410247d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 410329383967SMike Kravetz mutex_init(&h->resize_lock); 4104a3437870SNishanth Aravamudan h->order = order; 4105aca78307SMiaohe Lin h->mask = ~(huge_page_size(h) - 1); 41068faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 41078faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 41080edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 410954f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 411054f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 4111a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 4112a3437870SNishanth Aravamudan huge_page_size(h)/1024); 411377490587SMuchun Song hugetlb_vmemmap_init(h); 41148faa8b07SAndi Kleen 4115a3437870SNishanth Aravamudan parsed_hstate = h; 4116a3437870SNishanth Aravamudan } 4117a3437870SNishanth Aravamudan 4118b5389086SZhenguo Yao bool __init __weak hugetlb_node_alloc_supported(void) 4119b5389086SZhenguo Yao { 4120b5389086SZhenguo Yao return true; 4121b5389086SZhenguo Yao } 4122282f4214SMike Kravetz /* 4123282f4214SMike Kravetz * hugepages command line processing 4124282f4214SMike Kravetz * hugepages normally follows a valid hugepagsz or default_hugepagsz 4125282f4214SMike Kravetz * specification. If not, ignore the hugepages value. hugepages can also 4126282f4214SMike Kravetz * be the first huge page command line option in which case it implicitly 4127282f4214SMike Kravetz * specifies the number of huge pages for the default size. 4128282f4214SMike Kravetz */ 4129282f4214SMike Kravetz static int __init hugepages_setup(char *s) 4130a3437870SNishanth Aravamudan { 4131a3437870SNishanth Aravamudan unsigned long *mhp; 41328faa8b07SAndi Kleen static unsigned long *last_mhp; 4133b5389086SZhenguo Yao int node = NUMA_NO_NODE; 4134b5389086SZhenguo Yao int count; 4135b5389086SZhenguo Yao unsigned long tmp; 4136b5389086SZhenguo Yao char *p = s; 4137a3437870SNishanth Aravamudan 41389fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 4139282f4214SMike Kravetz pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); 41409fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 4141282f4214SMike Kravetz return 0; 41429fee021dSVaishali Thakkar } 4143282f4214SMike Kravetz 4144a3437870SNishanth Aravamudan /* 4145282f4214SMike Kravetz * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter 4146282f4214SMike Kravetz * yet, so this hugepages= parameter goes to the "default hstate". 4147282f4214SMike Kravetz * Otherwise, it goes with the previously parsed hugepagesz or 4148282f4214SMike Kravetz * default_hugepagesz. 4149a3437870SNishanth Aravamudan */ 41509fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 4151a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 4152a3437870SNishanth Aravamudan else 4153a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 4154a3437870SNishanth Aravamudan 41558faa8b07SAndi Kleen if (mhp == last_mhp) { 4156282f4214SMike Kravetz pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); 4157282f4214SMike Kravetz return 0; 41588faa8b07SAndi Kleen } 41598faa8b07SAndi Kleen 4160b5389086SZhenguo Yao while (*p) { 4161b5389086SZhenguo Yao count = 0; 4162b5389086SZhenguo Yao if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4163b5389086SZhenguo Yao goto invalid; 4164b5389086SZhenguo Yao /* Parameter is node format */ 4165b5389086SZhenguo Yao if (p[count] == ':') { 4166b5389086SZhenguo Yao if (!hugetlb_node_alloc_supported()) { 4167b5389086SZhenguo Yao pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n"); 4168b5389086SZhenguo Yao return 0; 4169b5389086SZhenguo Yao } 4170e79ce983SLiu Yuntao if (tmp >= nr_online_nodes) 4171e79ce983SLiu Yuntao goto invalid; 4172f9317f77SMike Kravetz node = array_index_nospec(tmp, nr_online_nodes); 4173b5389086SZhenguo Yao p += count + 1; 4174b5389086SZhenguo Yao /* Parse hugepages */ 4175b5389086SZhenguo Yao if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4176b5389086SZhenguo Yao goto invalid; 4177b5389086SZhenguo Yao if (!hugetlb_max_hstate) 4178b5389086SZhenguo Yao default_hugepages_in_node[node] = tmp; 4179b5389086SZhenguo Yao else 4180b5389086SZhenguo Yao parsed_hstate->max_huge_pages_node[node] = tmp; 4181b5389086SZhenguo Yao *mhp += tmp; 4182b5389086SZhenguo Yao /* Go to parse next node*/ 4183b5389086SZhenguo Yao if (p[count] == ',') 4184b5389086SZhenguo Yao p += count + 1; 4185b5389086SZhenguo Yao else 4186b5389086SZhenguo Yao break; 4187b5389086SZhenguo Yao } else { 4188b5389086SZhenguo Yao if (p != s) 4189b5389086SZhenguo Yao goto invalid; 4190b5389086SZhenguo Yao *mhp = tmp; 4191b5389086SZhenguo Yao break; 4192b5389086SZhenguo Yao } 4193b5389086SZhenguo Yao } 4194a3437870SNishanth Aravamudan 41958faa8b07SAndi Kleen /* 41968faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 419704adbc3fSMiaohe Lin * But we need to allocate gigantic hstates here early to still 41988faa8b07SAndi Kleen * use the bootmem allocator. 41998faa8b07SAndi Kleen */ 420004adbc3fSMiaohe Lin if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) 42018faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 42028faa8b07SAndi Kleen 42038faa8b07SAndi Kleen last_mhp = mhp; 42048faa8b07SAndi Kleen 4205a3437870SNishanth Aravamudan return 1; 4206b5389086SZhenguo Yao 4207b5389086SZhenguo Yao invalid: 4208b5389086SZhenguo Yao pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p); 4209b5389086SZhenguo Yao return 0; 4210a3437870SNishanth Aravamudan } 4211282f4214SMike Kravetz __setup("hugepages=", hugepages_setup); 4212e11bfbfcSNick Piggin 4213282f4214SMike Kravetz /* 4214282f4214SMike Kravetz * hugepagesz command line processing 4215282f4214SMike Kravetz * A specific huge page size can only be specified once with hugepagesz. 4216282f4214SMike Kravetz * hugepagesz is followed by hugepages on the command line. The global 4217282f4214SMike Kravetz * variable 'parsed_valid_hugepagesz' is used to determine if prior 4218282f4214SMike Kravetz * hugepagesz argument was valid. 4219282f4214SMike Kravetz */ 4220359f2544SMike Kravetz static int __init hugepagesz_setup(char *s) 4221e11bfbfcSNick Piggin { 4222359f2544SMike Kravetz unsigned long size; 4223282f4214SMike Kravetz struct hstate *h; 4224282f4214SMike Kravetz 4225282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4226359f2544SMike Kravetz size = (unsigned long)memparse(s, NULL); 4227359f2544SMike Kravetz 4228359f2544SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4229282f4214SMike Kravetz pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); 4230359f2544SMike Kravetz return 0; 4231359f2544SMike Kravetz } 4232359f2544SMike Kravetz 4233282f4214SMike Kravetz h = size_to_hstate(size); 4234282f4214SMike Kravetz if (h) { 4235282f4214SMike Kravetz /* 4236282f4214SMike Kravetz * hstate for this size already exists. This is normally 4237282f4214SMike Kravetz * an error, but is allowed if the existing hstate is the 4238282f4214SMike Kravetz * default hstate. More specifically, it is only allowed if 4239282f4214SMike Kravetz * the number of huge pages for the default hstate was not 4240282f4214SMike Kravetz * previously specified. 4241282f4214SMike Kravetz */ 4242282f4214SMike Kravetz if (!parsed_default_hugepagesz || h != &default_hstate || 4243282f4214SMike Kravetz default_hstate.max_huge_pages) { 4244282f4214SMike Kravetz pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); 424538237830SMike Kravetz return 0; 424638237830SMike Kravetz } 424738237830SMike Kravetz 4248282f4214SMike Kravetz /* 4249282f4214SMike Kravetz * No need to call hugetlb_add_hstate() as hstate already 4250282f4214SMike Kravetz * exists. But, do set parsed_hstate so that a following 4251282f4214SMike Kravetz * hugepages= parameter will be applied to this hstate. 4252282f4214SMike Kravetz */ 4253282f4214SMike Kravetz parsed_hstate = h; 4254282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4255e11bfbfcSNick Piggin return 1; 4256e11bfbfcSNick Piggin } 4257282f4214SMike Kravetz 4258359f2544SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4259282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4260359f2544SMike Kravetz return 1; 4261359f2544SMike Kravetz } 4262359f2544SMike Kravetz __setup("hugepagesz=", hugepagesz_setup); 4263359f2544SMike Kravetz 4264282f4214SMike Kravetz /* 4265282f4214SMike Kravetz * default_hugepagesz command line input 4266282f4214SMike Kravetz * Only one instance of default_hugepagesz allowed on command line. 4267282f4214SMike Kravetz */ 4268ae94da89SMike Kravetz static int __init default_hugepagesz_setup(char *s) 4269e11bfbfcSNick Piggin { 4270ae94da89SMike Kravetz unsigned long size; 4271b5389086SZhenguo Yao int i; 4272ae94da89SMike Kravetz 4273282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4274282f4214SMike Kravetz if (parsed_default_hugepagesz) { 4275282f4214SMike Kravetz pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); 4276282f4214SMike Kravetz return 0; 4277282f4214SMike Kravetz } 4278282f4214SMike Kravetz 4279282f4214SMike Kravetz size = (unsigned long)memparse(s, NULL); 4280282f4214SMike Kravetz 4281282f4214SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4282282f4214SMike Kravetz pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); 4283282f4214SMike Kravetz return 0; 4284282f4214SMike Kravetz } 4285282f4214SMike Kravetz 4286282f4214SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4287282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4288282f4214SMike Kravetz parsed_default_hugepagesz = true; 4289282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(size)); 4290282f4214SMike Kravetz 4291282f4214SMike Kravetz /* 4292282f4214SMike Kravetz * The number of default huge pages (for this size) could have been 4293282f4214SMike Kravetz * specified as the first hugetlb parameter: hugepages=X. If so, 4294282f4214SMike Kravetz * then default_hstate_max_huge_pages is set. If the default huge 4295282f4214SMike Kravetz * page size is gigantic (>= MAX_ORDER), then the pages must be 4296282f4214SMike Kravetz * allocated here from bootmem allocator. 4297282f4214SMike Kravetz */ 4298282f4214SMike Kravetz if (default_hstate_max_huge_pages) { 4299282f4214SMike Kravetz default_hstate.max_huge_pages = default_hstate_max_huge_pages; 4300b5389086SZhenguo Yao for (i = 0; i < nr_online_nodes; i++) 4301b5389086SZhenguo Yao default_hstate.max_huge_pages_node[i] = 4302b5389086SZhenguo Yao default_hugepages_in_node[i]; 4303282f4214SMike Kravetz if (hstate_is_gigantic(&default_hstate)) 4304282f4214SMike Kravetz hugetlb_hstate_alloc_pages(&default_hstate); 4305282f4214SMike Kravetz default_hstate_max_huge_pages = 0; 4306282f4214SMike Kravetz } 4307282f4214SMike Kravetz 4308e11bfbfcSNick Piggin return 1; 4309e11bfbfcSNick Piggin } 4310ae94da89SMike Kravetz __setup("default_hugepagesz=", default_hugepagesz_setup); 4311a3437870SNishanth Aravamudan 43128ca39e68SMuchun Song static unsigned int allowed_mems_nr(struct hstate *h) 43138a213460SNishanth Aravamudan { 43148a213460SNishanth Aravamudan int node; 43158a213460SNishanth Aravamudan unsigned int nr = 0; 43168ca39e68SMuchun Song nodemask_t *mpol_allowed; 43178ca39e68SMuchun Song unsigned int *array = h->free_huge_pages_node; 43188ca39e68SMuchun Song gfp_t gfp_mask = htlb_alloc_mask(h); 43198a213460SNishanth Aravamudan 43208ca39e68SMuchun Song mpol_allowed = policy_nodemask_current(gfp_mask); 43218ca39e68SMuchun Song 43228ca39e68SMuchun Song for_each_node_mask(node, cpuset_current_mems_allowed) { 4323c93b0a99SJiapeng Zhong if (!mpol_allowed || node_isset(node, *mpol_allowed)) 43248a213460SNishanth Aravamudan nr += array[node]; 43258ca39e68SMuchun Song } 43268a213460SNishanth Aravamudan 43278a213460SNishanth Aravamudan return nr; 43288a213460SNishanth Aravamudan } 43298a213460SNishanth Aravamudan 43308a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 433117743798SMuchun Song static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, 433217743798SMuchun Song void *buffer, size_t *length, 433317743798SMuchun Song loff_t *ppos, unsigned long *out) 433417743798SMuchun Song { 433517743798SMuchun Song struct ctl_table dup_table; 433617743798SMuchun Song 433717743798SMuchun Song /* 433817743798SMuchun Song * In order to avoid races with __do_proc_doulongvec_minmax(), we 433917743798SMuchun Song * can duplicate the @table and alter the duplicate of it. 434017743798SMuchun Song */ 434117743798SMuchun Song dup_table = *table; 434217743798SMuchun Song dup_table.data = out; 434317743798SMuchun Song 434417743798SMuchun Song return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); 434517743798SMuchun Song } 434617743798SMuchun Song 434706808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 434806808b08SLee Schermerhorn struct ctl_table *table, int write, 434932927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 43501da177e4SLinus Torvalds { 4351e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 4352238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 435308d4a246SMichal Hocko int ret; 4354e5ff2159SAndi Kleen 4355457c1b27SNishanth Aravamudan if (!hugepages_supported()) 435686613628SJan Stancek return -EOPNOTSUPP; 4357457c1b27SNishanth Aravamudan 435817743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 435917743798SMuchun Song &tmp); 436008d4a246SMichal Hocko if (ret) 436108d4a246SMichal Hocko goto out; 4362e5ff2159SAndi Kleen 4363238d3c13SDavid Rientjes if (write) 4364238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 4365238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 436608d4a246SMichal Hocko out: 436708d4a246SMichal Hocko return ret; 43681da177e4SLinus Torvalds } 4369396faf03SMel Gorman 437006808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 437132927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 437206808b08SLee Schermerhorn { 437306808b08SLee Schermerhorn 437406808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 437506808b08SLee Schermerhorn buffer, length, ppos); 437606808b08SLee Schermerhorn } 437706808b08SLee Schermerhorn 437806808b08SLee Schermerhorn #ifdef CONFIG_NUMA 437906808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 438032927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 438106808b08SLee Schermerhorn { 438206808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 438306808b08SLee Schermerhorn buffer, length, ppos); 438406808b08SLee Schermerhorn } 438506808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 438606808b08SLee Schermerhorn 4387a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 438832927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 4389a3d0c6aaSNishanth Aravamudan { 4390a5516438SAndi Kleen struct hstate *h = &default_hstate; 4391e5ff2159SAndi Kleen unsigned long tmp; 439208d4a246SMichal Hocko int ret; 4393e5ff2159SAndi Kleen 4394457c1b27SNishanth Aravamudan if (!hugepages_supported()) 439586613628SJan Stancek return -EOPNOTSUPP; 4396457c1b27SNishanth Aravamudan 4397e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 4398e5ff2159SAndi Kleen 4399bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 4400adbe8726SEric B Munson return -EINVAL; 4401adbe8726SEric B Munson 440217743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 440317743798SMuchun Song &tmp); 440408d4a246SMichal Hocko if (ret) 440508d4a246SMichal Hocko goto out; 4406e5ff2159SAndi Kleen 4407e5ff2159SAndi Kleen if (write) { 4408db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4409e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 4410db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4411e5ff2159SAndi Kleen } 441208d4a246SMichal Hocko out: 441308d4a246SMichal Hocko return ret; 4414a3d0c6aaSNishanth Aravamudan } 4415a3d0c6aaSNishanth Aravamudan 44161da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 44171da177e4SLinus Torvalds 4418e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 44191da177e4SLinus Torvalds { 4420fcb2b0c5SRoman Gushchin struct hstate *h; 4421fcb2b0c5SRoman Gushchin unsigned long total = 0; 4422fcb2b0c5SRoman Gushchin 4423457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4424457c1b27SNishanth Aravamudan return; 4425fcb2b0c5SRoman Gushchin 4426fcb2b0c5SRoman Gushchin for_each_hstate(h) { 4427fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 4428fcb2b0c5SRoman Gushchin 4429aca78307SMiaohe Lin total += huge_page_size(h) * count; 4430fcb2b0c5SRoman Gushchin 4431fcb2b0c5SRoman Gushchin if (h == &default_hstate) 4432e1759c21SAlexey Dobriyan seq_printf(m, 44331da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 44341da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 4435b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 44367893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 44374f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 4438fcb2b0c5SRoman Gushchin count, 4439a5516438SAndi Kleen h->free_huge_pages, 4440a5516438SAndi Kleen h->resv_huge_pages, 4441a5516438SAndi Kleen h->surplus_huge_pages, 4442aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4443fcb2b0c5SRoman Gushchin } 4444fcb2b0c5SRoman Gushchin 4445aca78307SMiaohe Lin seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); 44461da177e4SLinus Torvalds } 44471da177e4SLinus Torvalds 44487981593bSJoe Perches int hugetlb_report_node_meminfo(char *buf, int len, int nid) 44491da177e4SLinus Torvalds { 4450a5516438SAndi Kleen struct hstate *h = &default_hstate; 44517981593bSJoe Perches 4452457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4453457c1b27SNishanth Aravamudan return 0; 44547981593bSJoe Perches 44557981593bSJoe Perches return sysfs_emit_at(buf, len, 44561da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 4457a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 4458a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 4459a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 4460a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 4461a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 44621da177e4SLinus Torvalds } 44631da177e4SLinus Torvalds 4464949f7ec5SDavid Rientjes void hugetlb_show_meminfo(void) 4465949f7ec5SDavid Rientjes { 4466949f7ec5SDavid Rientjes struct hstate *h; 4467949f7ec5SDavid Rientjes int nid; 4468949f7ec5SDavid Rientjes 4469457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4470457c1b27SNishanth Aravamudan return; 4471457c1b27SNishanth Aravamudan 4472949f7ec5SDavid Rientjes for_each_node_state(nid, N_MEMORY) 4473949f7ec5SDavid Rientjes for_each_hstate(h) 4474949f7ec5SDavid Rientjes pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 4475949f7ec5SDavid Rientjes nid, 4476949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 4477949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 4478949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 4479aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4480949f7ec5SDavid Rientjes } 4481949f7ec5SDavid Rientjes 44825d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 44835d317b2bSNaoya Horiguchi { 44845d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 44855d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 44865d317b2bSNaoya Horiguchi } 44875d317b2bSNaoya Horiguchi 44881da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 44891da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 44901da177e4SLinus Torvalds { 4491d0028588SWanpeng Li struct hstate *h; 4492d0028588SWanpeng Li unsigned long nr_total_pages = 0; 4493d0028588SWanpeng Li 4494d0028588SWanpeng Li for_each_hstate(h) 4495d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 4496d0028588SWanpeng Li return nr_total_pages; 44971da177e4SLinus Torvalds } 44981da177e4SLinus Torvalds 4499a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 4500fc1b8a73SMel Gorman { 4501fc1b8a73SMel Gorman int ret = -ENOMEM; 4502fc1b8a73SMel Gorman 45030aa7f354SMiaohe Lin if (!delta) 45040aa7f354SMiaohe Lin return 0; 45050aa7f354SMiaohe Lin 4506db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4507fc1b8a73SMel Gorman /* 4508fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 4509fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 4510fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 4511fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 4512fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 4513fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 4514fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 4515fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 4516fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 4517fc1b8a73SMel Gorman * 4518fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 4519fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 4520fc1b8a73SMel Gorman * we fall back to check against current free page availability as 4521fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 4522fc1b8a73SMel Gorman * semantics that cpuset has. 45238ca39e68SMuchun Song * 45248ca39e68SMuchun Song * Apart from cpuset, we also have memory policy mechanism that 45258ca39e68SMuchun Song * also determines from which node the kernel will allocate memory 45268ca39e68SMuchun Song * in a NUMA system. So similar to cpuset, we also should consider 45278ca39e68SMuchun Song * the memory policy of the current task. Similar to the description 45288ca39e68SMuchun Song * above. 4529fc1b8a73SMel Gorman */ 4530fc1b8a73SMel Gorman if (delta > 0) { 4531a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 4532fc1b8a73SMel Gorman goto out; 4533fc1b8a73SMel Gorman 45348ca39e68SMuchun Song if (delta > allowed_mems_nr(h)) { 4535a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 4536fc1b8a73SMel Gorman goto out; 4537fc1b8a73SMel Gorman } 4538fc1b8a73SMel Gorman } 4539fc1b8a73SMel Gorman 4540fc1b8a73SMel Gorman ret = 0; 4541fc1b8a73SMel Gorman if (delta < 0) 4542a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 4543fc1b8a73SMel Gorman 4544fc1b8a73SMel Gorman out: 4545db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4546fc1b8a73SMel Gorman return ret; 4547fc1b8a73SMel Gorman } 4548fc1b8a73SMel Gorman 454984afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 455084afd99bSAndy Whitcroft { 4551f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 455284afd99bSAndy Whitcroft 455384afd99bSAndy Whitcroft /* 455484afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 455584afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 455684afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 455725985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 455884afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 455984afd99bSAndy Whitcroft * new reference here without additional locking. 456084afd99bSAndy Whitcroft */ 456109a26e83SMike Kravetz if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 456209a26e83SMike Kravetz resv_map_dup_hugetlb_cgroup_uncharge_info(resv); 4563f522c3acSJoonsoo Kim kref_get(&resv->refs); 456484afd99bSAndy Whitcroft } 456509a26e83SMike Kravetz } 456684afd99bSAndy Whitcroft 4567a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 4568a1e78772SMel Gorman { 4569a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 4570f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 457190481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 45724e35f483SJoonsoo Kim unsigned long reserve, start, end; 45731c5ecae3SMike Kravetz long gbl_reserve; 457484afd99bSAndy Whitcroft 45754e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 45764e35f483SJoonsoo Kim return; 45774e35f483SJoonsoo Kim 4578a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 4579a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 458084afd99bSAndy Whitcroft 45814e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 4582e9fe92aeSMina Almasry hugetlb_cgroup_uncharge_counter(resv, start, end); 45837251ff78SAdam Litke if (reserve) { 45841c5ecae3SMike Kravetz /* 45851c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 45861c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 45871c5ecae3SMike Kravetz */ 45881c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 45891c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 45907251ff78SAdam Litke } 4591e9fe92aeSMina Almasry 4592e9fe92aeSMina Almasry kref_put(&resv->refs, resv_map_release); 4593a1e78772SMel Gorman } 4594a1e78772SMel Gorman 459531383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 459631383c68SDan Williams { 459731383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 459831383c68SDan Williams return -EINVAL; 459931383c68SDan Williams return 0; 460031383c68SDan Williams } 460131383c68SDan Williams 460205ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 460305ea8860SDan Williams { 4604aca78307SMiaohe Lin return huge_page_size(hstate_vma(vma)); 460505ea8860SDan Williams } 460605ea8860SDan Williams 46071da177e4SLinus Torvalds /* 46081da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 46091da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 46106c26d310SMiaohe Lin * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get 46111da177e4SLinus Torvalds * this far. 46121da177e4SLinus Torvalds */ 4613b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 46141da177e4SLinus Torvalds { 46151da177e4SLinus Torvalds BUG(); 4616d0217ac0SNick Piggin return 0; 46171da177e4SLinus Torvalds } 46181da177e4SLinus Torvalds 4619eec3636aSJane Chu /* 4620eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 4621eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 4622eec3636aSJane Chu * This is because under System V memory model, mappings created via 4623eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 4624eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 4625eec3636aSJane Chu */ 4626f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 4627d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 462884afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 4629a1e78772SMel Gorman .close = hugetlb_vm_op_close, 4630dd3b614fSDmitry Safonov .may_split = hugetlb_vm_op_split, 463105ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 46321da177e4SLinus Torvalds }; 46331da177e4SLinus Torvalds 46341e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 46351e8f889bSDavid Gibson int writable) 463663551ae0SDavid Gibson { 463763551ae0SDavid Gibson pte_t entry; 463879c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 463963551ae0SDavid Gibson 46401e8f889bSDavid Gibson if (writable) { 4641106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 4642106c992aSGerald Schaefer vma->vm_page_prot))); 464363551ae0SDavid Gibson } else { 4644106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 4645106c992aSGerald Schaefer vma->vm_page_prot)); 464663551ae0SDavid Gibson } 464763551ae0SDavid Gibson entry = pte_mkyoung(entry); 464879c1c594SChristophe Leroy entry = arch_make_huge_pte(entry, shift, vma->vm_flags); 464963551ae0SDavid Gibson 465063551ae0SDavid Gibson return entry; 465163551ae0SDavid Gibson } 465263551ae0SDavid Gibson 46531e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 46541e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 46551e8f889bSDavid Gibson { 46561e8f889bSDavid Gibson pte_t entry; 46571e8f889bSDavid Gibson 4658106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 465932f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 46604b3073e1SRussell King update_mmu_cache(vma, address, ptep); 46611e8f889bSDavid Gibson } 46621e8f889bSDavid Gibson 4663d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 46644a705fefSNaoya Horiguchi { 46654a705fefSNaoya Horiguchi swp_entry_t swp; 46664a705fefSNaoya Horiguchi 46674a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 4668d5ed7444SAneesh Kumar K.V return false; 46694a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4670d79d176aSBaoquan He if (is_migration_entry(swp)) 4671d5ed7444SAneesh Kumar K.V return true; 46724a705fefSNaoya Horiguchi else 4673d5ed7444SAneesh Kumar K.V return false; 46744a705fefSNaoya Horiguchi } 46754a705fefSNaoya Horiguchi 46763e5c3600SBaoquan He static bool is_hugetlb_entry_hwpoisoned(pte_t pte) 46774a705fefSNaoya Horiguchi { 46784a705fefSNaoya Horiguchi swp_entry_t swp; 46794a705fefSNaoya Horiguchi 46804a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 46813e5c3600SBaoquan He return false; 46824a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4683d79d176aSBaoquan He if (is_hwpoison_entry(swp)) 46843e5c3600SBaoquan He return true; 46854a705fefSNaoya Horiguchi else 46863e5c3600SBaoquan He return false; 46874a705fefSNaoya Horiguchi } 46881e8f889bSDavid Gibson 46894eae4efaSPeter Xu static void 46904eae4efaSPeter Xu hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, 46914eae4efaSPeter Xu struct page *new_page) 46924eae4efaSPeter Xu { 46934eae4efaSPeter Xu __SetPageUptodate(new_page); 46944eae4efaSPeter Xu hugepage_add_new_anon_rmap(new_page, vma, addr); 46951eba86c0SPasha Tatashin set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1)); 46964eae4efaSPeter Xu hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); 46974eae4efaSPeter Xu ClearHPageRestoreReserve(new_page); 46984eae4efaSPeter Xu SetHPageMigratable(new_page); 46994eae4efaSPeter Xu } 47004eae4efaSPeter Xu 470163551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 470263551ae0SDavid Gibson struct vm_area_struct *vma) 470363551ae0SDavid Gibson { 47045e41540cSMike Kravetz pte_t *src_pte, *dst_pte, entry, dst_entry; 470563551ae0SDavid Gibson struct page *ptepage; 47061c59827dSHugh Dickins unsigned long addr; 4707ca6eb14dSPeter Xu bool cow = is_cow_mapping(vma->vm_flags); 4708a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 4709a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 47104eae4efaSPeter Xu unsigned long npages = pages_per_huge_page(h); 4711c0d0381aSMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 4712ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 4713e8569dd2SAndreas Sandberg int ret = 0; 47141e8f889bSDavid Gibson 4715ac46d4f3SJérôme Glisse if (cow) { 47167269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src, 47176f4f13e8SJérôme Glisse vma->vm_start, 4718ac46d4f3SJérôme Glisse vma->vm_end); 4719ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4720c0d0381aSMike Kravetz } else { 4721c0d0381aSMike Kravetz /* 4722c0d0381aSMike Kravetz * For shared mappings i_mmap_rwsem must be held to call 4723c0d0381aSMike Kravetz * huge_pte_alloc, otherwise the returned ptep could go 4724c0d0381aSMike Kravetz * away if part of a shared pmd and another thread calls 4725c0d0381aSMike Kravetz * huge_pmd_unshare. 4726c0d0381aSMike Kravetz */ 4727c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 4728ac46d4f3SJérôme Glisse } 4729e8569dd2SAndreas Sandberg 4730a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 4731cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 47327868a208SPunit Agrawal src_pte = huge_pte_offset(src, addr, sz); 4733c74df32cSHugh Dickins if (!src_pte) 4734c74df32cSHugh Dickins continue; 4735aec44e0fSPeter Xu dst_pte = huge_pte_alloc(dst, vma, addr, sz); 4736e8569dd2SAndreas Sandberg if (!dst_pte) { 4737e8569dd2SAndreas Sandberg ret = -ENOMEM; 4738e8569dd2SAndreas Sandberg break; 4739e8569dd2SAndreas Sandberg } 4740c5c99429SLarry Woodman 47415e41540cSMike Kravetz /* 47425e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 47435e41540cSMike Kravetz * dst_pte == src_pte is the common case of src/dest sharing. 47445e41540cSMike Kravetz * 47455e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 47465e41540cSMike Kravetz * another vma. If dst_pte !none, this implies sharing. 47475e41540cSMike Kravetz * Check here before taking page table lock, and once again 47485e41540cSMike Kravetz * after taking the lock below. 47495e41540cSMike Kravetz */ 47505e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 47515e41540cSMike Kravetz if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) 4752c5c99429SLarry Woodman continue; 4753c5c99429SLarry Woodman 4754cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 4755cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 4756cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 47574a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 47585e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 47594eae4efaSPeter Xu again: 47605e41540cSMike Kravetz if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { 47615e41540cSMike Kravetz /* 47625e41540cSMike Kravetz * Skip if src entry none. Also, skip in the 47635e41540cSMike Kravetz * unlikely case dst entry !none as this implies 47645e41540cSMike Kravetz * sharing with another vma. 47655e41540cSMike Kravetz */ 47664a705fefSNaoya Horiguchi ; 47674a705fefSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry) || 47684a705fefSNaoya Horiguchi is_hugetlb_entry_hwpoisoned(entry))) { 47694a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 47704a705fefSNaoya Horiguchi 47714dd845b5SAlistair Popple if (is_writable_migration_entry(swp_entry) && cow) { 47724a705fefSNaoya Horiguchi /* 47734a705fefSNaoya Horiguchi * COW mappings require pages in both 47744a705fefSNaoya Horiguchi * parent and child to be set to read. 47754a705fefSNaoya Horiguchi */ 47764dd845b5SAlistair Popple swp_entry = make_readable_migration_entry( 47774dd845b5SAlistair Popple swp_offset(swp_entry)); 47784a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 4779e5251fd4SPunit Agrawal set_huge_swap_pte_at(src, addr, src_pte, 4780e5251fd4SPunit Agrawal entry, sz); 47814a705fefSNaoya Horiguchi } 4782e5251fd4SPunit Agrawal set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); 47834a705fefSNaoya Horiguchi } else { 47844eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 47854eae4efaSPeter Xu ptepage = pte_page(entry); 47864eae4efaSPeter Xu get_page(ptepage); 47874eae4efaSPeter Xu 47884eae4efaSPeter Xu /* 47894eae4efaSPeter Xu * This is a rare case where we see pinned hugetlb 47904eae4efaSPeter Xu * pages while they're prone to COW. We need to do the 47914eae4efaSPeter Xu * COW earlier during fork. 47924eae4efaSPeter Xu * 47934eae4efaSPeter Xu * When pre-allocating the page or copying data, we 47944eae4efaSPeter Xu * need to be without the pgtable locks since we could 47954eae4efaSPeter Xu * sleep during the process. 47964eae4efaSPeter Xu */ 47974eae4efaSPeter Xu if (unlikely(page_needs_cow_for_dma(vma, ptepage))) { 47984eae4efaSPeter Xu pte_t src_pte_old = entry; 47994eae4efaSPeter Xu struct page *new; 48004eae4efaSPeter Xu 48014eae4efaSPeter Xu spin_unlock(src_ptl); 48024eae4efaSPeter Xu spin_unlock(dst_ptl); 48034eae4efaSPeter Xu /* Do not use reserve as it's private owned */ 48044eae4efaSPeter Xu new = alloc_huge_page(vma, addr, 1); 48054eae4efaSPeter Xu if (IS_ERR(new)) { 48064eae4efaSPeter Xu put_page(ptepage); 48074eae4efaSPeter Xu ret = PTR_ERR(new); 48084eae4efaSPeter Xu break; 48094eae4efaSPeter Xu } 48104eae4efaSPeter Xu copy_user_huge_page(new, ptepage, addr, vma, 48114eae4efaSPeter Xu npages); 48124eae4efaSPeter Xu put_page(ptepage); 48134eae4efaSPeter Xu 48144eae4efaSPeter Xu /* Install the new huge page if src pte stable */ 48154eae4efaSPeter Xu dst_ptl = huge_pte_lock(h, dst, dst_pte); 48164eae4efaSPeter Xu src_ptl = huge_pte_lockptr(h, src, src_pte); 48174eae4efaSPeter Xu spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 48184eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 48194eae4efaSPeter Xu if (!pte_same(src_pte_old, entry)) { 4820846be085SMike Kravetz restore_reserve_on_error(h, vma, addr, 4821846be085SMike Kravetz new); 48224eae4efaSPeter Xu put_page(new); 48234eae4efaSPeter Xu /* dst_entry won't change as in child */ 48244eae4efaSPeter Xu goto again; 48254eae4efaSPeter Xu } 48264eae4efaSPeter Xu hugetlb_install_page(vma, dst_pte, addr, new); 48274eae4efaSPeter Xu spin_unlock(src_ptl); 48284eae4efaSPeter Xu spin_unlock(dst_ptl); 48294eae4efaSPeter Xu continue; 48304eae4efaSPeter Xu } 48314eae4efaSPeter Xu 483234ee645eSJoerg Roedel if (cow) { 48330f10851eSJérôme Glisse /* 48340f10851eSJérôme Glisse * No need to notify as we are downgrading page 48350f10851eSJérôme Glisse * table protection not changing it to point 48360f10851eSJérôme Glisse * to a new page. 48370f10851eSJérôme Glisse * 4838ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 48390f10851eSJérôme Glisse */ 48407f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 484184894e1cSPeter Xu entry = huge_pte_wrprotect(entry); 484234ee645eSJoerg Roedel } 48434eae4efaSPeter Xu 484453f9263bSKirill A. Shutemov page_dup_rmap(ptepage, true); 484563551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 48464eae4efaSPeter Xu hugetlb_count_add(npages, dst); 48471c59827dSHugh Dickins } 4848cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 4849cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 485063551ae0SDavid Gibson } 485163551ae0SDavid Gibson 4852e8569dd2SAndreas Sandberg if (cow) 4853ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 4854c0d0381aSMike Kravetz else 4855c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 4856e8569dd2SAndreas Sandberg 4857e8569dd2SAndreas Sandberg return ret; 485863551ae0SDavid Gibson } 485963551ae0SDavid Gibson 4860550a7d60SMina Almasry static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr, 4861db110a99SAneesh Kumar K.V unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte) 4862550a7d60SMina Almasry { 4863550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 4864550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 4865550a7d60SMina Almasry spinlock_t *src_ptl, *dst_ptl; 4866db110a99SAneesh Kumar K.V pte_t pte; 4867550a7d60SMina Almasry 4868550a7d60SMina Almasry dst_ptl = huge_pte_lock(h, mm, dst_pte); 4869550a7d60SMina Almasry src_ptl = huge_pte_lockptr(h, mm, src_pte); 4870550a7d60SMina Almasry 4871550a7d60SMina Almasry /* 4872550a7d60SMina Almasry * We don't have to worry about the ordering of src and dst ptlocks 4873550a7d60SMina Almasry * because exclusive mmap_sem (or the i_mmap_lock) prevents deadlock. 4874550a7d60SMina Almasry */ 4875550a7d60SMina Almasry if (src_ptl != dst_ptl) 4876550a7d60SMina Almasry spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 4877550a7d60SMina Almasry 4878550a7d60SMina Almasry pte = huge_ptep_get_and_clear(mm, old_addr, src_pte); 4879550a7d60SMina Almasry set_huge_pte_at(mm, new_addr, dst_pte, pte); 4880550a7d60SMina Almasry 4881550a7d60SMina Almasry if (src_ptl != dst_ptl) 4882550a7d60SMina Almasry spin_unlock(src_ptl); 4883550a7d60SMina Almasry spin_unlock(dst_ptl); 4884550a7d60SMina Almasry } 4885550a7d60SMina Almasry 4886550a7d60SMina Almasry int move_hugetlb_page_tables(struct vm_area_struct *vma, 4887550a7d60SMina Almasry struct vm_area_struct *new_vma, 4888550a7d60SMina Almasry unsigned long old_addr, unsigned long new_addr, 4889550a7d60SMina Almasry unsigned long len) 4890550a7d60SMina Almasry { 4891550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 4892550a7d60SMina Almasry struct address_space *mapping = vma->vm_file->f_mapping; 4893550a7d60SMina Almasry unsigned long sz = huge_page_size(h); 4894550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 4895550a7d60SMina Almasry unsigned long old_end = old_addr + len; 4896550a7d60SMina Almasry unsigned long old_addr_copy; 4897550a7d60SMina Almasry pte_t *src_pte, *dst_pte; 4898550a7d60SMina Almasry struct mmu_notifier_range range; 4899550a7d60SMina Almasry 4900550a7d60SMina Almasry mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, old_addr, 4901550a7d60SMina Almasry old_end); 4902550a7d60SMina Almasry adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 4903550a7d60SMina Almasry mmu_notifier_invalidate_range_start(&range); 4904550a7d60SMina Almasry /* Prevent race with file truncation */ 4905550a7d60SMina Almasry i_mmap_lock_write(mapping); 4906550a7d60SMina Almasry for (; old_addr < old_end; old_addr += sz, new_addr += sz) { 4907550a7d60SMina Almasry src_pte = huge_pte_offset(mm, old_addr, sz); 4908550a7d60SMina Almasry if (!src_pte) 4909550a7d60SMina Almasry continue; 4910550a7d60SMina Almasry if (huge_pte_none(huge_ptep_get(src_pte))) 4911550a7d60SMina Almasry continue; 4912550a7d60SMina Almasry 4913550a7d60SMina Almasry /* old_addr arg to huge_pmd_unshare() is a pointer and so the 4914550a7d60SMina Almasry * arg may be modified. Pass a copy instead to preserve the 4915550a7d60SMina Almasry * value in old_addr. 4916550a7d60SMina Almasry */ 4917550a7d60SMina Almasry old_addr_copy = old_addr; 4918550a7d60SMina Almasry 4919550a7d60SMina Almasry if (huge_pmd_unshare(mm, vma, &old_addr_copy, src_pte)) 4920550a7d60SMina Almasry continue; 4921550a7d60SMina Almasry 4922550a7d60SMina Almasry dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz); 4923550a7d60SMina Almasry if (!dst_pte) 4924550a7d60SMina Almasry break; 4925550a7d60SMina Almasry 4926db110a99SAneesh Kumar K.V move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte); 4927550a7d60SMina Almasry } 4928550a7d60SMina Almasry flush_tlb_range(vma, old_end - len, old_end); 4929550a7d60SMina Almasry mmu_notifier_invalidate_range_end(&range); 493013e4ad2cSNadav Amit i_mmap_unlock_write(mapping); 4931550a7d60SMina Almasry 4932550a7d60SMina Almasry return len + old_addr - old_end; 4933550a7d60SMina Almasry } 4934550a7d60SMina Almasry 493573c54763SPeter Xu static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 493624669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 493724669e58SAneesh Kumar K.V struct page *ref_page) 493863551ae0SDavid Gibson { 493963551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 494063551ae0SDavid Gibson unsigned long address; 4941c7546f8fSDavid Gibson pte_t *ptep; 494263551ae0SDavid Gibson pte_t pte; 4943cb900f41SKirill A. Shutemov spinlock_t *ptl; 494463551ae0SDavid Gibson struct page *page; 4945a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 4946a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 4947ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 4948a4a118f2SNadav Amit bool force_flush = false; 4949a5516438SAndi Kleen 495063551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 4951a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 4952a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 495363551ae0SDavid Gibson 495407e32661SAneesh Kumar K.V /* 495507e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 495607e32661SAneesh Kumar K.V * to huge page. 495707e32661SAneesh Kumar K.V */ 4958ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 495924669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 4960dff11abeSMike Kravetz 4961dff11abeSMike Kravetz /* 4962dff11abeSMike Kravetz * If sharing possible, alert mmu notifiers of worst case. 4963dff11abeSMike Kravetz */ 49646f4f13e8SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, 49656f4f13e8SJérôme Glisse end); 4966ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 4967ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4968569f48b8SHillf Danton address = start; 4969569f48b8SHillf Danton for (; address < end; address += sz) { 49707868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, sz); 4971c7546f8fSDavid Gibson if (!ptep) 4972c7546f8fSDavid Gibson continue; 4973c7546f8fSDavid Gibson 4974cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 497534ae204fSMike Kravetz if (huge_pmd_unshare(mm, vma, &address, ptep)) { 497631d49da5SAneesh Kumar K.V spin_unlock(ptl); 4977a4a118f2SNadav Amit tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE); 4978a4a118f2SNadav Amit force_flush = true; 497931d49da5SAneesh Kumar K.V continue; 498031d49da5SAneesh Kumar K.V } 498139dde65cSChen, Kenneth W 49826629326bSHillf Danton pte = huge_ptep_get(ptep); 498331d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 498431d49da5SAneesh Kumar K.V spin_unlock(ptl); 498531d49da5SAneesh Kumar K.V continue; 498631d49da5SAneesh Kumar K.V } 49876629326bSHillf Danton 49886629326bSHillf Danton /* 49899fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 49909fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 49916629326bSHillf Danton */ 49929fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 49939386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 499431d49da5SAneesh Kumar K.V spin_unlock(ptl); 499531d49da5SAneesh Kumar K.V continue; 49968c4894c6SNaoya Horiguchi } 49976629326bSHillf Danton 49986629326bSHillf Danton page = pte_page(pte); 499904f2cbe3SMel Gorman /* 500004f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 500104f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 500204f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 500304f2cbe3SMel Gorman */ 500404f2cbe3SMel Gorman if (ref_page) { 500531d49da5SAneesh Kumar K.V if (page != ref_page) { 500631d49da5SAneesh Kumar K.V spin_unlock(ptl); 500731d49da5SAneesh Kumar K.V continue; 500831d49da5SAneesh Kumar K.V } 500904f2cbe3SMel Gorman /* 501004f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 501104f2cbe3SMel Gorman * future faults in this VMA will fail rather than 501204f2cbe3SMel Gorman * looking like data was lost 501304f2cbe3SMel Gorman */ 501404f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 501504f2cbe3SMel Gorman } 501604f2cbe3SMel Gorman 5017c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 5018b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 5019106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 50206649a386SKen Chen set_page_dirty(page); 50219e81130bSHillf Danton 50225d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 5023cea86fe2SHugh Dickins page_remove_rmap(page, vma, true); 502431d49da5SAneesh Kumar K.V 5025cb900f41SKirill A. Shutemov spin_unlock(ptl); 5026e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 502724669e58SAneesh Kumar K.V /* 502831d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 502924669e58SAneesh Kumar K.V */ 503031d49da5SAneesh Kumar K.V if (ref_page) 503131d49da5SAneesh Kumar K.V break; 5032fe1668aeSChen, Kenneth W } 5033ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 503424669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 5035a4a118f2SNadav Amit 5036a4a118f2SNadav Amit /* 5037a4a118f2SNadav Amit * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We 5038a4a118f2SNadav Amit * could defer the flush until now, since by holding i_mmap_rwsem we 5039a4a118f2SNadav Amit * guaranteed that the last refernece would not be dropped. But we must 5040a4a118f2SNadav Amit * do the flushing before we return, as otherwise i_mmap_rwsem will be 5041a4a118f2SNadav Amit * dropped and the last reference to the shared PMDs page might be 5042a4a118f2SNadav Amit * dropped as well. 5043a4a118f2SNadav Amit * 5044a4a118f2SNadav Amit * In theory we could defer the freeing of the PMD pages as well, but 5045a4a118f2SNadav Amit * huge_pmd_unshare() relies on the exact page_count for the PMD page to 5046a4a118f2SNadav Amit * detect sharing, so we cannot defer the release of the page either. 5047a4a118f2SNadav Amit * Instead, do flush now. 5048a4a118f2SNadav Amit */ 5049a4a118f2SNadav Amit if (force_flush) 5050a4a118f2SNadav Amit tlb_flush_mmu_tlbonly(tlb); 50511da177e4SLinus Torvalds } 505263551ae0SDavid Gibson 5053d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 5054d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 5055d833352aSMel Gorman unsigned long end, struct page *ref_page) 5056d833352aSMel Gorman { 5057d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 5058d833352aSMel Gorman 5059d833352aSMel Gorman /* 5060d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 5061d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 5062d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 5063d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 5064d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 5065c8c06efaSDavidlohr Bueso * is to clear it before releasing the i_mmap_rwsem. This works 5066d833352aSMel Gorman * because in the context this is called, the VMA is about to be 5067c8c06efaSDavidlohr Bueso * destroyed and the i_mmap_rwsem is held. 5068d833352aSMel Gorman */ 5069d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 5070d833352aSMel Gorman } 5071d833352aSMel Gorman 5072502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 507304f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 5074502717f4SChen, Kenneth W { 507524669e58SAneesh Kumar K.V struct mmu_gather tlb; 5076dff11abeSMike Kravetz 5077a72afd87SWill Deacon tlb_gather_mmu(&tlb, vma->vm_mm); 507824669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 5079ae8eba8bSWill Deacon tlb_finish_mmu(&tlb); 5080502717f4SChen, Kenneth W } 5081502717f4SChen, Kenneth W 508204f2cbe3SMel Gorman /* 508304f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 5084578b7725SZhiyuan Dai * mapping it owns the reserve page for. The intention is to unmap the page 508504f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 508604f2cbe3SMel Gorman * same region. 508704f2cbe3SMel Gorman */ 50882f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 50892a4b3dedSHarvey Harrison struct page *page, unsigned long address) 509004f2cbe3SMel Gorman { 50917526674dSAdam Litke struct hstate *h = hstate_vma(vma); 509204f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 509304f2cbe3SMel Gorman struct address_space *mapping; 509404f2cbe3SMel Gorman pgoff_t pgoff; 509504f2cbe3SMel Gorman 509604f2cbe3SMel Gorman /* 509704f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 509804f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 509904f2cbe3SMel Gorman */ 51007526674dSAdam Litke address = address & huge_page_mask(h); 510136e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 510236e4f20aSMichal Hocko vma->vm_pgoff; 510393c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 510404f2cbe3SMel Gorman 51054eb2b1dcSMel Gorman /* 51064eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 51074eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 51084eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 51094eb2b1dcSMel Gorman */ 511083cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 51116b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 511204f2cbe3SMel Gorman /* Do not unmap the current VMA */ 511304f2cbe3SMel Gorman if (iter_vma == vma) 511404f2cbe3SMel Gorman continue; 511504f2cbe3SMel Gorman 511604f2cbe3SMel Gorman /* 51172f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 51182f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 51192f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 51202f84a899SMel Gorman */ 51212f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 51222f84a899SMel Gorman continue; 51232f84a899SMel Gorman 51242f84a899SMel Gorman /* 512504f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 512604f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 512704f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 512804f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 512904f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 513004f2cbe3SMel Gorman */ 513104f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 513224669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 513324669e58SAneesh Kumar K.V address + huge_page_size(h), page); 513404f2cbe3SMel Gorman } 513583cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 513604f2cbe3SMel Gorman } 513704f2cbe3SMel Gorman 51380fe6e20bSNaoya Horiguchi /* 51390fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 5140aa6d2e8cSBaolin Wang * Called with hugetlb_fault_mutex_table held and pte_page locked so we 5141ef009b25SMichal Hocko * cannot race with other handlers or page migration. 5142ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 51430fe6e20bSNaoya Horiguchi */ 51442b740303SSouptick Joarder static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 5145974e6d66SHuang Ying unsigned long address, pte_t *ptep, 5146cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 51471e8f889bSDavid Gibson { 51483999f52eSAneesh Kumar K.V pte_t pte; 5149a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 51501e8f889bSDavid Gibson struct page *old_page, *new_page; 51512b740303SSouptick Joarder int outside_reserve = 0; 51522b740303SSouptick Joarder vm_fault_t ret = 0; 5153974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 5154ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 51551e8f889bSDavid Gibson 51563999f52eSAneesh Kumar K.V pte = huge_ptep_get(ptep); 51571e8f889bSDavid Gibson old_page = pte_page(pte); 51581e8f889bSDavid Gibson 515904f2cbe3SMel Gorman retry_avoidcopy: 51601e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 51611e8f889bSDavid Gibson * and just make the page writable */ 516237a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 51635a49973dSHugh Dickins page_move_anon_rmap(old_page, vma); 51645b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 516583c54070SNick Piggin return 0; 51661e8f889bSDavid Gibson } 51671e8f889bSDavid Gibson 516804f2cbe3SMel Gorman /* 516904f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 517004f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 517104f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 517204f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 517304f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 517404f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 517504f2cbe3SMel Gorman * of the full address range. 517604f2cbe3SMel Gorman */ 51775944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 517804f2cbe3SMel Gorman old_page != pagecache_page) 517904f2cbe3SMel Gorman outside_reserve = 1; 518004f2cbe3SMel Gorman 518109cbfeafSKirill A. Shutemov get_page(old_page); 5182b76c8cfbSLarry Woodman 5183ad4404a2SDavidlohr Bueso /* 5184ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 5185ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 5186ad4404a2SDavidlohr Bueso */ 5187cb900f41SKirill A. Shutemov spin_unlock(ptl); 51885b7a1d40SHuang Ying new_page = alloc_huge_page(vma, haddr, outside_reserve); 51891e8f889bSDavid Gibson 51902fc39cecSAdam Litke if (IS_ERR(new_page)) { 519104f2cbe3SMel Gorman /* 519204f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 519304f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 519404f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 519504f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 519604f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 519704f2cbe3SMel Gorman */ 519804f2cbe3SMel Gorman if (outside_reserve) { 5199e7dd91c4SMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 5200e7dd91c4SMike Kravetz pgoff_t idx; 5201e7dd91c4SMike Kravetz u32 hash; 5202e7dd91c4SMike Kravetz 520309cbfeafSKirill A. Shutemov put_page(old_page); 520404f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 5205e7dd91c4SMike Kravetz /* 5206e7dd91c4SMike Kravetz * Drop hugetlb_fault_mutex and i_mmap_rwsem before 5207e7dd91c4SMike Kravetz * unmapping. unmapping needs to hold i_mmap_rwsem 5208e7dd91c4SMike Kravetz * in write mode. Dropping i_mmap_rwsem in read mode 5209e7dd91c4SMike Kravetz * here is OK as COW mappings do not interact with 5210e7dd91c4SMike Kravetz * PMD sharing. 5211e7dd91c4SMike Kravetz * 5212e7dd91c4SMike Kravetz * Reacquire both after unmap operation. 5213e7dd91c4SMike Kravetz */ 5214e7dd91c4SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 5215e7dd91c4SMike Kravetz hash = hugetlb_fault_mutex_hash(mapping, idx); 5216e7dd91c4SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5217e7dd91c4SMike Kravetz i_mmap_unlock_read(mapping); 5218e7dd91c4SMike Kravetz 52195b7a1d40SHuang Ying unmap_ref_private(mm, vma, old_page, haddr); 5220e7dd91c4SMike Kravetz 5221e7dd91c4SMike Kravetz i_mmap_lock_read(mapping); 5222e7dd91c4SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 5223cb900f41SKirill A. Shutemov spin_lock(ptl); 52245b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 5225a9af0c5dSNaoya Horiguchi if (likely(ptep && 5226a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 522704f2cbe3SMel Gorman goto retry_avoidcopy; 5228a734bcc8SHillf Danton /* 5229cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 5230cb900f41SKirill A. Shutemov * lock, and our job is done. 5231a734bcc8SHillf Danton */ 5232a734bcc8SHillf Danton return 0; 523304f2cbe3SMel Gorman } 523404f2cbe3SMel Gorman 52352b740303SSouptick Joarder ret = vmf_error(PTR_ERR(new_page)); 5236ad4404a2SDavidlohr Bueso goto out_release_old; 52371e8f889bSDavid Gibson } 52381e8f889bSDavid Gibson 52390fe6e20bSNaoya Horiguchi /* 52400fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 52410fe6e20bSNaoya Horiguchi * anon_vma prepared. 52420fe6e20bSNaoya Horiguchi */ 524344e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 5244ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 5245ad4404a2SDavidlohr Bueso goto out_release_all; 524644e2aa93SDean Nelson } 52470fe6e20bSNaoya Horiguchi 5248974e6d66SHuang Ying copy_user_huge_page(new_page, old_page, address, vma, 524947ad8475SAndrea Arcangeli pages_per_huge_page(h)); 52500ed361deSNick Piggin __SetPageUptodate(new_page); 52511e8f889bSDavid Gibson 52527269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, 52536f4f13e8SJérôme Glisse haddr + huge_page_size(h)); 5254ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 5255ad4404a2SDavidlohr Bueso 5256b76c8cfbSLarry Woodman /* 5257cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 5258b76c8cfbSLarry Woodman * before the page tables are altered 5259b76c8cfbSLarry Woodman */ 5260cb900f41SKirill A. Shutemov spin_lock(ptl); 52615b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 5262a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 5263d6995da3SMike Kravetz ClearHPageRestoreReserve(new_page); 526407443a85SJoonsoo Kim 52651e8f889bSDavid Gibson /* Break COW */ 52665b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 5267ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, range.end); 5268cea86fe2SHugh Dickins page_remove_rmap(old_page, vma, true); 52695b7a1d40SHuang Ying hugepage_add_new_anon_rmap(new_page, vma, haddr); 52701eba86c0SPasha Tatashin set_huge_pte_at(mm, haddr, ptep, 52711eba86c0SPasha Tatashin make_huge_pte(vma, new_page, 1)); 52728f251a3dSMike Kravetz SetHPageMigratable(new_page); 52731e8f889bSDavid Gibson /* Make the old page be freed below */ 52741e8f889bSDavid Gibson new_page = old_page; 52751e8f889bSDavid Gibson } 5276cb900f41SKirill A. Shutemov spin_unlock(ptl); 5277ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 5278ad4404a2SDavidlohr Bueso out_release_all: 5279c7b1850dSMike Kravetz /* No restore in case of successful pagetable update (Break COW) */ 5280c7b1850dSMike Kravetz if (new_page != old_page) 52815b7a1d40SHuang Ying restore_reserve_on_error(h, vma, haddr, new_page); 528209cbfeafSKirill A. Shutemov put_page(new_page); 5283ad4404a2SDavidlohr Bueso out_release_old: 528409cbfeafSKirill A. Shutemov put_page(old_page); 52858312034fSJoonsoo Kim 5286ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 5287ad4404a2SDavidlohr Bueso return ret; 52881e8f889bSDavid Gibson } 52891e8f889bSDavid Gibson 529004f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 5291a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 5292a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 529304f2cbe3SMel Gorman { 529404f2cbe3SMel Gorman struct address_space *mapping; 5295e7c4b0bfSAndy Whitcroft pgoff_t idx; 529604f2cbe3SMel Gorman 529704f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 5298a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 529904f2cbe3SMel Gorman 530004f2cbe3SMel Gorman return find_lock_page(mapping, idx); 530104f2cbe3SMel Gorman } 530204f2cbe3SMel Gorman 53033ae77f43SHugh Dickins /* 53043ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 53053ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 53063ae77f43SHugh Dickins */ 53073ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 53082a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 53092a15efc9SHugh Dickins { 53102a15efc9SHugh Dickins struct address_space *mapping; 53112a15efc9SHugh Dickins pgoff_t idx; 53122a15efc9SHugh Dickins struct page *page; 53132a15efc9SHugh Dickins 53142a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 53152a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 53162a15efc9SHugh Dickins 53172a15efc9SHugh Dickins page = find_get_page(mapping, idx); 53182a15efc9SHugh Dickins if (page) 53192a15efc9SHugh Dickins put_page(page); 53202a15efc9SHugh Dickins return page != NULL; 53212a15efc9SHugh Dickins } 53222a15efc9SHugh Dickins 5323ab76ad54SMike Kravetz int huge_add_to_page_cache(struct page *page, struct address_space *mapping, 5324ab76ad54SMike Kravetz pgoff_t idx) 5325ab76ad54SMike Kravetz { 5326ab76ad54SMike Kravetz struct inode *inode = mapping->host; 5327ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 5328ab76ad54SMike Kravetz int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 5329ab76ad54SMike Kravetz 5330ab76ad54SMike Kravetz if (err) 5331ab76ad54SMike Kravetz return err; 5332d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 5333ab76ad54SMike Kravetz 533422146c3cSMike Kravetz /* 533522146c3cSMike Kravetz * set page dirty so that it will not be removed from cache/file 533622146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 533722146c3cSMike Kravetz */ 533822146c3cSMike Kravetz set_page_dirty(page); 533922146c3cSMike Kravetz 5340ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 5341ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 5342ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 5343ab76ad54SMike Kravetz return 0; 5344ab76ad54SMike Kravetz } 5345ab76ad54SMike Kravetz 53467677f7fdSAxel Rasmussen static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma, 53477677f7fdSAxel Rasmussen struct address_space *mapping, 53487677f7fdSAxel Rasmussen pgoff_t idx, 53497677f7fdSAxel Rasmussen unsigned int flags, 53507677f7fdSAxel Rasmussen unsigned long haddr, 5351824ddc60SNadav Amit unsigned long addr, 53527677f7fdSAxel Rasmussen unsigned long reason) 53537677f7fdSAxel Rasmussen { 53547677f7fdSAxel Rasmussen vm_fault_t ret; 53557677f7fdSAxel Rasmussen u32 hash; 53567677f7fdSAxel Rasmussen struct vm_fault vmf = { 53577677f7fdSAxel Rasmussen .vma = vma, 53587677f7fdSAxel Rasmussen .address = haddr, 5359824ddc60SNadav Amit .real_address = addr, 53607677f7fdSAxel Rasmussen .flags = flags, 53617677f7fdSAxel Rasmussen 53627677f7fdSAxel Rasmussen /* 53637677f7fdSAxel Rasmussen * Hard to debug if it ends up being 53647677f7fdSAxel Rasmussen * used by a callee that assumes 53657677f7fdSAxel Rasmussen * something about the other 53667677f7fdSAxel Rasmussen * uninitialized fields... same as in 53677677f7fdSAxel Rasmussen * memory.c 53687677f7fdSAxel Rasmussen */ 53697677f7fdSAxel Rasmussen }; 53707677f7fdSAxel Rasmussen 53717677f7fdSAxel Rasmussen /* 53727677f7fdSAxel Rasmussen * hugetlb_fault_mutex and i_mmap_rwsem must be 53737677f7fdSAxel Rasmussen * dropped before handling userfault. Reacquire 53747677f7fdSAxel Rasmussen * after handling fault to make calling code simpler. 53757677f7fdSAxel Rasmussen */ 53767677f7fdSAxel Rasmussen hash = hugetlb_fault_mutex_hash(mapping, idx); 53777677f7fdSAxel Rasmussen mutex_unlock(&hugetlb_fault_mutex_table[hash]); 53787677f7fdSAxel Rasmussen i_mmap_unlock_read(mapping); 53797677f7fdSAxel Rasmussen ret = handle_userfault(&vmf, reason); 53807677f7fdSAxel Rasmussen i_mmap_lock_read(mapping); 53817677f7fdSAxel Rasmussen mutex_lock(&hugetlb_fault_mutex_table[hash]); 53827677f7fdSAxel Rasmussen 53837677f7fdSAxel Rasmussen return ret; 53847677f7fdSAxel Rasmussen } 53857677f7fdSAxel Rasmussen 53862b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 53872b740303SSouptick Joarder struct vm_area_struct *vma, 53888382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 5389788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 5390ac9b9c66SHugh Dickins { 5391a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 53922b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 5393409eb8c2SHillf Danton int anon_rmap = 0; 53944c887265SAdam Litke unsigned long size; 53954c887265SAdam Litke struct page *page; 53961e8f889bSDavid Gibson pte_t new_pte; 5397cb900f41SKirill A. Shutemov spinlock_t *ptl; 5398285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 5399c7b1850dSMike Kravetz bool new_page, new_pagecache_page = false; 54004c887265SAdam Litke 540104f2cbe3SMel Gorman /* 540204f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 540304f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 540425985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 540504f2cbe3SMel Gorman */ 540604f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 5407910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 540804f2cbe3SMel Gorman current->pid); 540904f2cbe3SMel Gorman return ret; 541004f2cbe3SMel Gorman } 541104f2cbe3SMel Gorman 54124c887265SAdam Litke /* 541387bf91d3SMike Kravetz * We can not race with truncation due to holding i_mmap_rwsem. 541487bf91d3SMike Kravetz * i_size is modified when holding i_mmap_rwsem, so check here 541587bf91d3SMike Kravetz * once for faults beyond end of file. 54164c887265SAdam Litke */ 5417a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 5418ebed4bfcSHugh Dickins if (idx >= size) 5419ebed4bfcSHugh Dickins goto out; 54201a1aad8aSMike Kravetz 542187bf91d3SMike Kravetz retry: 5422c7b1850dSMike Kravetz new_page = false; 542387bf91d3SMike Kravetz page = find_lock_page(mapping, idx); 542487bf91d3SMike Kravetz if (!page) { 54257677f7fdSAxel Rasmussen /* Check for page in userfault range */ 54261a1aad8aSMike Kravetz if (userfaultfd_missing(vma)) { 54277677f7fdSAxel Rasmussen ret = hugetlb_handle_userfault(vma, mapping, idx, 5428824ddc60SNadav Amit flags, haddr, address, 54297677f7fdSAxel Rasmussen VM_UFFD_MISSING); 54301a1aad8aSMike Kravetz goto out; 54311a1aad8aSMike Kravetz } 54321a1aad8aSMike Kravetz 5433285b8dcaSHuang Ying page = alloc_huge_page(vma, haddr, 0); 54342fc39cecSAdam Litke if (IS_ERR(page)) { 54354643d67eSMike Kravetz /* 54364643d67eSMike Kravetz * Returning error will result in faulting task being 54374643d67eSMike Kravetz * sent SIGBUS. The hugetlb fault mutex prevents two 54384643d67eSMike Kravetz * tasks from racing to fault in the same page which 54394643d67eSMike Kravetz * could result in false unable to allocate errors. 54404643d67eSMike Kravetz * Page migration does not take the fault mutex, but 54414643d67eSMike Kravetz * does a clear then write of pte's under page table 54424643d67eSMike Kravetz * lock. Page fault code could race with migration, 54434643d67eSMike Kravetz * notice the clear pte and try to allocate a page 54444643d67eSMike Kravetz * here. Before returning error, get ptl and make 54454643d67eSMike Kravetz * sure there really is no pte entry. 54464643d67eSMike Kravetz */ 54474643d67eSMike Kravetz ptl = huge_pte_lock(h, mm, ptep); 54484643d67eSMike Kravetz ret = 0; 5449d83e6c8aSMiaohe Lin if (huge_pte_none(huge_ptep_get(ptep))) 54502b740303SSouptick Joarder ret = vmf_error(PTR_ERR(page)); 5451d83e6c8aSMiaohe Lin spin_unlock(ptl); 54526bda666aSChristoph Lameter goto out; 54536bda666aSChristoph Lameter } 545447ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 54550ed361deSNick Piggin __SetPageUptodate(page); 5456cb6acd01SMike Kravetz new_page = true; 5457ac9b9c66SHugh Dickins 5458f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 5459ab76ad54SMike Kravetz int err = huge_add_to_page_cache(page, mapping, idx); 54606bda666aSChristoph Lameter if (err) { 54616bda666aSChristoph Lameter put_page(page); 54626bda666aSChristoph Lameter if (err == -EEXIST) 54636bda666aSChristoph Lameter goto retry; 54646bda666aSChristoph Lameter goto out; 54656bda666aSChristoph Lameter } 5466c7b1850dSMike Kravetz new_pagecache_page = true; 546723be7468SMel Gorman } else { 54686bda666aSChristoph Lameter lock_page(page); 54690fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 54700fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 54710fe6e20bSNaoya Horiguchi goto backout_unlocked; 547223be7468SMel Gorman } 5473409eb8c2SHillf Danton anon_rmap = 1; 54740fe6e20bSNaoya Horiguchi } 54750fe6e20bSNaoya Horiguchi } else { 547657303d80SAndy Whitcroft /* 5477998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 5478998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 5479998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 5480fd6a03edSNaoya Horiguchi */ 5481fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 54820eb98f15SMiaohe Lin ret = VM_FAULT_HWPOISON_LARGE | 5483972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 5484fd6a03edSNaoya Horiguchi goto backout_unlocked; 54856bda666aSChristoph Lameter } 54867677f7fdSAxel Rasmussen 54877677f7fdSAxel Rasmussen /* Check for page in userfault range. */ 54887677f7fdSAxel Rasmussen if (userfaultfd_minor(vma)) { 54897677f7fdSAxel Rasmussen unlock_page(page); 54907677f7fdSAxel Rasmussen put_page(page); 54917677f7fdSAxel Rasmussen ret = hugetlb_handle_userfault(vma, mapping, idx, 5492824ddc60SNadav Amit flags, haddr, address, 54937677f7fdSAxel Rasmussen VM_UFFD_MINOR); 54947677f7fdSAxel Rasmussen goto out; 54957677f7fdSAxel Rasmussen } 5496998b4382SNaoya Horiguchi } 54971e8f889bSDavid Gibson 549857303d80SAndy Whitcroft /* 549957303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 550057303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 550157303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 550257303d80SAndy Whitcroft * the spinlock. 550357303d80SAndy Whitcroft */ 55045e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 5505285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 55062b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 55072b26736cSAndy Whitcroft goto backout_unlocked; 55082b26736cSAndy Whitcroft } 55095e911373SMike Kravetz /* Just decrements count, does not deallocate */ 5510285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 55115e911373SMike Kravetz } 551257303d80SAndy Whitcroft 55138bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 551483c54070SNick Piggin ret = 0; 55157f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 55164c887265SAdam Litke goto backout; 55174c887265SAdam Litke 551807443a85SJoonsoo Kim if (anon_rmap) { 5519d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 5520285b8dcaSHuang Ying hugepage_add_new_anon_rmap(page, vma, haddr); 5521ac714904SChoi Gi-yong } else 552253f9263bSKirill A. Shutemov page_dup_rmap(page, true); 55231e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 55241e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 5525285b8dcaSHuang Ying set_huge_pte_at(mm, haddr, ptep, new_pte); 55261e8f889bSDavid Gibson 55275d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 5528788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 55291e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 5530974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); 55311e8f889bSDavid Gibson } 55321e8f889bSDavid Gibson 5533cb900f41SKirill A. Shutemov spin_unlock(ptl); 5534cb6acd01SMike Kravetz 5535cb6acd01SMike Kravetz /* 55368f251a3dSMike Kravetz * Only set HPageMigratable in newly allocated pages. Existing pages 55378f251a3dSMike Kravetz * found in the pagecache may not have HPageMigratableset if they have 55388f251a3dSMike Kravetz * been isolated for migration. 5539cb6acd01SMike Kravetz */ 5540cb6acd01SMike Kravetz if (new_page) 55418f251a3dSMike Kravetz SetHPageMigratable(page); 5542cb6acd01SMike Kravetz 55434c887265SAdam Litke unlock_page(page); 55444c887265SAdam Litke out: 5545ac9b9c66SHugh Dickins return ret; 55464c887265SAdam Litke 55474c887265SAdam Litke backout: 5548cb900f41SKirill A. Shutemov spin_unlock(ptl); 55492b26736cSAndy Whitcroft backout_unlocked: 55504c887265SAdam Litke unlock_page(page); 5551c7b1850dSMike Kravetz /* restore reserve for newly allocated pages not in page cache */ 5552c7b1850dSMike Kravetz if (new_page && !new_pagecache_page) 5553285b8dcaSHuang Ying restore_reserve_on_error(h, vma, haddr, page); 55544c887265SAdam Litke put_page(page); 55554c887265SAdam Litke goto out; 5556ac9b9c66SHugh Dickins } 5557ac9b9c66SHugh Dickins 55588382d914SDavidlohr Bueso #ifdef CONFIG_SMP 5559188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 55608382d914SDavidlohr Bueso { 55618382d914SDavidlohr Bueso unsigned long key[2]; 55628382d914SDavidlohr Bueso u32 hash; 55638382d914SDavidlohr Bueso 55648382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 55658382d914SDavidlohr Bueso key[1] = idx; 55668382d914SDavidlohr Bueso 556755254636SMike Kravetz hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); 55688382d914SDavidlohr Bueso 55698382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 55708382d914SDavidlohr Bueso } 55718382d914SDavidlohr Bueso #else 55728382d914SDavidlohr Bueso /* 55736c26d310SMiaohe Lin * For uniprocessor systems we always use a single mutex, so just 55748382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 55758382d914SDavidlohr Bueso */ 5576188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 55778382d914SDavidlohr Bueso { 55788382d914SDavidlohr Bueso return 0; 55798382d914SDavidlohr Bueso } 55808382d914SDavidlohr Bueso #endif 55818382d914SDavidlohr Bueso 55822b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 5583788c7df4SHugh Dickins unsigned long address, unsigned int flags) 558486e5216fSAdam Litke { 55858382d914SDavidlohr Bueso pte_t *ptep, entry; 5586cb900f41SKirill A. Shutemov spinlock_t *ptl; 55872b740303SSouptick Joarder vm_fault_t ret; 55888382d914SDavidlohr Bueso u32 hash; 55898382d914SDavidlohr Bueso pgoff_t idx; 55900fe6e20bSNaoya Horiguchi struct page *page = NULL; 559157303d80SAndy Whitcroft struct page *pagecache_page = NULL; 5592a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 55938382d914SDavidlohr Bueso struct address_space *mapping; 55940f792cf9SNaoya Horiguchi int need_wait_lock = 0; 5595285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 559686e5216fSAdam Litke 5597285b8dcaSHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 5598fd6a03edSNaoya Horiguchi if (ptep) { 5599c0d0381aSMike Kravetz /* 5600c0d0381aSMike Kravetz * Since we hold no locks, ptep could be stale. That is 5601c0d0381aSMike Kravetz * OK as we are only making decisions based on content and 5602c0d0381aSMike Kravetz * not actually modifying content here. 5603c0d0381aSMike Kravetz */ 5604fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 5605290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 5606cb900f41SKirill A. Shutemov migration_entry_wait_huge(vma, mm, ptep); 5607290408d4SNaoya Horiguchi return 0; 5608290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 5609aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 5610972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 5611b43a9990SMike Kravetz } 56128382d914SDavidlohr Bueso 5613c0d0381aSMike Kravetz /* 5614c0d0381aSMike Kravetz * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold 561587bf91d3SMike Kravetz * until finished with ptep. This serves two purposes: 561687bf91d3SMike Kravetz * 1) It prevents huge_pmd_unshare from being called elsewhere 561787bf91d3SMike Kravetz * and making the ptep no longer valid. 561887bf91d3SMike Kravetz * 2) It synchronizes us with i_size modifications during truncation. 5619c0d0381aSMike Kravetz * 5620c0d0381aSMike Kravetz * ptep could have already be assigned via huge_pte_offset. That 5621c0d0381aSMike Kravetz * is OK, as huge_pte_alloc will return the same value unless 5622c0d0381aSMike Kravetz * something has changed. 5623c0d0381aSMike Kravetz */ 5624ddeaab32SMike Kravetz mapping = vma->vm_file->f_mapping; 5625c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 5626aec44e0fSPeter Xu ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); 5627c0d0381aSMike Kravetz if (!ptep) { 5628c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 5629c0d0381aSMike Kravetz return VM_FAULT_OOM; 5630c0d0381aSMike Kravetz } 5631ddeaab32SMike Kravetz 56323935baa9SDavid Gibson /* 56333935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 56343935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 56353935baa9SDavid Gibson * the same page in the page cache. 56363935baa9SDavid Gibson */ 5637c0d0381aSMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 5638188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 5639c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 56408382d914SDavidlohr Bueso 56417f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 56427f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 56438382d914SDavidlohr Bueso ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 5644b4d1d99fSDavid Gibson goto out_mutex; 56453935baa9SDavid Gibson } 564686e5216fSAdam Litke 564783c54070SNick Piggin ret = 0; 56481e8f889bSDavid Gibson 564957303d80SAndy Whitcroft /* 56500f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 56510f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 56527c8de358SEthon Paul * an active hugepage in pagecache. This goto expects the 2nd page 56537c8de358SEthon Paul * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will 56547c8de358SEthon Paul * properly handle it. 56550f792cf9SNaoya Horiguchi */ 56560f792cf9SNaoya Horiguchi if (!pte_present(entry)) 56570f792cf9SNaoya Horiguchi goto out_mutex; 56580f792cf9SNaoya Horiguchi 56590f792cf9SNaoya Horiguchi /* 566057303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 566157303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 566257303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 566357303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 566457303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 566557303d80SAndy Whitcroft * consumed. 566657303d80SAndy Whitcroft */ 5667106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 5668285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 56692b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 5670b4d1d99fSDavid Gibson goto out_mutex; 56712b26736cSAndy Whitcroft } 56725e911373SMike Kravetz /* Just decrements count, does not deallocate */ 5673285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 567457303d80SAndy Whitcroft 5675f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 567657303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 5677285b8dcaSHuang Ying vma, haddr); 567857303d80SAndy Whitcroft } 567957303d80SAndy Whitcroft 56800f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 56810fe6e20bSNaoya Horiguchi 56821e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 5683b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 5684cb900f41SKirill A. Shutemov goto out_ptl; 5685b4d1d99fSDavid Gibson 56860f792cf9SNaoya Horiguchi /* 56870f792cf9SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 56880f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 56890f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 56900f792cf9SNaoya Horiguchi */ 56910f792cf9SNaoya Horiguchi page = pte_page(entry); 56920f792cf9SNaoya Horiguchi if (page != pagecache_page) 56930f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 56940f792cf9SNaoya Horiguchi need_wait_lock = 1; 56950f792cf9SNaoya Horiguchi goto out_ptl; 56960f792cf9SNaoya Horiguchi } 56970f792cf9SNaoya Horiguchi 56980f792cf9SNaoya Horiguchi get_page(page); 5699b4d1d99fSDavid Gibson 5700788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 5701106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 5702974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, 5703cb900f41SKirill A. Shutemov pagecache_page, ptl); 57040f792cf9SNaoya Horiguchi goto out_put_page; 5705b4d1d99fSDavid Gibson } 5706106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 5707b4d1d99fSDavid Gibson } 5708b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 5709285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 5710788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 5711285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 57120f792cf9SNaoya Horiguchi out_put_page: 57130f792cf9SNaoya Horiguchi if (page != pagecache_page) 57140f792cf9SNaoya Horiguchi unlock_page(page); 57150f792cf9SNaoya Horiguchi put_page(page); 5716cb900f41SKirill A. Shutemov out_ptl: 5717cb900f41SKirill A. Shutemov spin_unlock(ptl); 571857303d80SAndy Whitcroft 571957303d80SAndy Whitcroft if (pagecache_page) { 572057303d80SAndy Whitcroft unlock_page(pagecache_page); 572157303d80SAndy Whitcroft put_page(pagecache_page); 572257303d80SAndy Whitcroft } 5723b4d1d99fSDavid Gibson out_mutex: 5724c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5725c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 57260f792cf9SNaoya Horiguchi /* 57270f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 57280f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 57290f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 57300f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 57310f792cf9SNaoya Horiguchi * here without taking refcount. 57320f792cf9SNaoya Horiguchi */ 57330f792cf9SNaoya Horiguchi if (need_wait_lock) 57340f792cf9SNaoya Horiguchi wait_on_page_locked(page); 57351e8f889bSDavid Gibson return ret; 573686e5216fSAdam Litke } 573786e5216fSAdam Litke 5738714c1891SAxel Rasmussen #ifdef CONFIG_USERFAULTFD 57398fb5debcSMike Kravetz /* 57408fb5debcSMike Kravetz * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with 57418fb5debcSMike Kravetz * modifications for huge pages. 57428fb5debcSMike Kravetz */ 57438fb5debcSMike Kravetz int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, 57448fb5debcSMike Kravetz pte_t *dst_pte, 57458fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 57468fb5debcSMike Kravetz unsigned long dst_addr, 57478fb5debcSMike Kravetz unsigned long src_addr, 5748f6191471SAxel Rasmussen enum mcopy_atomic_mode mode, 57498fb5debcSMike Kravetz struct page **pagep) 57508fb5debcSMike Kravetz { 5751f6191471SAxel Rasmussen bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE); 57528cc5fcbbSMina Almasry struct hstate *h = hstate_vma(dst_vma); 57538cc5fcbbSMina Almasry struct address_space *mapping = dst_vma->vm_file->f_mapping; 57548cc5fcbbSMina Almasry pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr); 57551e392147SAndrea Arcangeli unsigned long size; 57561c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 57578fb5debcSMike Kravetz pte_t _dst_pte; 57588fb5debcSMike Kravetz spinlock_t *ptl; 57598cc5fcbbSMina Almasry int ret = -ENOMEM; 57608fb5debcSMike Kravetz struct page *page; 5761f6191471SAxel Rasmussen int writable; 5762cc30042dSMina Almasry bool page_in_pagecache = false; 57638fb5debcSMike Kravetz 5764f6191471SAxel Rasmussen if (is_continue) { 5765f6191471SAxel Rasmussen ret = -EFAULT; 5766f6191471SAxel Rasmussen page = find_lock_page(mapping, idx); 5767f6191471SAxel Rasmussen if (!page) 5768f6191471SAxel Rasmussen goto out; 5769cc30042dSMina Almasry page_in_pagecache = true; 5770f6191471SAxel Rasmussen } else if (!*pagep) { 5771d84cf06eSMina Almasry /* If a page already exists, then it's UFFDIO_COPY for 5772d84cf06eSMina Almasry * a non-missing case. Return -EEXIST. 5773d84cf06eSMina Almasry */ 5774d84cf06eSMina Almasry if (vm_shared && 5775d84cf06eSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 5776d84cf06eSMina Almasry ret = -EEXIST; 57778fb5debcSMike Kravetz goto out; 5778d84cf06eSMina Almasry } 5779d84cf06eSMina Almasry 5780d84cf06eSMina Almasry page = alloc_huge_page(dst_vma, dst_addr, 0); 5781d84cf06eSMina Almasry if (IS_ERR(page)) { 5782d84cf06eSMina Almasry ret = -ENOMEM; 5783d84cf06eSMina Almasry goto out; 5784d84cf06eSMina Almasry } 57858fb5debcSMike Kravetz 57868fb5debcSMike Kravetz ret = copy_huge_page_from_user(page, 57878fb5debcSMike Kravetz (const void __user *) src_addr, 5788810a56b9SMike Kravetz pages_per_huge_page(h), false); 57898fb5debcSMike Kravetz 5790c1e8d7c6SMichel Lespinasse /* fallback to copy_from_user outside mmap_lock */ 57918fb5debcSMike Kravetz if (unlikely(ret)) { 57929e368259SAndrea Arcangeli ret = -ENOENT; 57938cc5fcbbSMina Almasry /* Free the allocated page which may have 57948cc5fcbbSMina Almasry * consumed a reservation. 57958cc5fcbbSMina Almasry */ 57968cc5fcbbSMina Almasry restore_reserve_on_error(h, dst_vma, dst_addr, page); 57978cc5fcbbSMina Almasry put_page(page); 57988cc5fcbbSMina Almasry 57998cc5fcbbSMina Almasry /* Allocate a temporary page to hold the copied 58008cc5fcbbSMina Almasry * contents. 58018cc5fcbbSMina Almasry */ 58028cc5fcbbSMina Almasry page = alloc_huge_page_vma(h, dst_vma, dst_addr); 58038cc5fcbbSMina Almasry if (!page) { 58048cc5fcbbSMina Almasry ret = -ENOMEM; 58058cc5fcbbSMina Almasry goto out; 58068cc5fcbbSMina Almasry } 58078fb5debcSMike Kravetz *pagep = page; 58088cc5fcbbSMina Almasry /* Set the outparam pagep and return to the caller to 58098cc5fcbbSMina Almasry * copy the contents outside the lock. Don't free the 58108cc5fcbbSMina Almasry * page. 58118cc5fcbbSMina Almasry */ 58128fb5debcSMike Kravetz goto out; 58138fb5debcSMike Kravetz } 58148fb5debcSMike Kravetz } else { 58158cc5fcbbSMina Almasry if (vm_shared && 58168cc5fcbbSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 58178cc5fcbbSMina Almasry put_page(*pagep); 58188cc5fcbbSMina Almasry ret = -EEXIST; 58198cc5fcbbSMina Almasry *pagep = NULL; 58208cc5fcbbSMina Almasry goto out; 58218cc5fcbbSMina Almasry } 58228cc5fcbbSMina Almasry 58238cc5fcbbSMina Almasry page = alloc_huge_page(dst_vma, dst_addr, 0); 58248cc5fcbbSMina Almasry if (IS_ERR(page)) { 58258cc5fcbbSMina Almasry ret = -ENOMEM; 58268cc5fcbbSMina Almasry *pagep = NULL; 58278cc5fcbbSMina Almasry goto out; 58288cc5fcbbSMina Almasry } 582934892366SMuchun Song copy_user_huge_page(page, *pagep, dst_addr, dst_vma, 583034892366SMuchun Song pages_per_huge_page(h)); 58318cc5fcbbSMina Almasry put_page(*pagep); 58328fb5debcSMike Kravetz *pagep = NULL; 58338fb5debcSMike Kravetz } 58348fb5debcSMike Kravetz 58358fb5debcSMike Kravetz /* 58368fb5debcSMike Kravetz * The memory barrier inside __SetPageUptodate makes sure that 58378fb5debcSMike Kravetz * preceding stores to the page contents become visible before 58388fb5debcSMike Kravetz * the set_pte_at() write. 58398fb5debcSMike Kravetz */ 58408fb5debcSMike Kravetz __SetPageUptodate(page); 58418fb5debcSMike Kravetz 5842f6191471SAxel Rasmussen /* Add shared, newly allocated pages to the page cache. */ 5843f6191471SAxel Rasmussen if (vm_shared && !is_continue) { 58441e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 58451e392147SAndrea Arcangeli ret = -EFAULT; 58461e392147SAndrea Arcangeli if (idx >= size) 58471e392147SAndrea Arcangeli goto out_release_nounlock; 58481c9e8defSMike Kravetz 58491e392147SAndrea Arcangeli /* 58501e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 58511e392147SAndrea Arcangeli * huge_add_to_page_cache() below happens through the 58521e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 58531e392147SAndrea Arcangeli * the caller. 58541e392147SAndrea Arcangeli */ 58551c9e8defSMike Kravetz ret = huge_add_to_page_cache(page, mapping, idx); 58561c9e8defSMike Kravetz if (ret) 58571c9e8defSMike Kravetz goto out_release_nounlock; 5858cc30042dSMina Almasry page_in_pagecache = true; 58591c9e8defSMike Kravetz } 58601c9e8defSMike Kravetz 58618fb5debcSMike Kravetz ptl = huge_pte_lockptr(h, dst_mm, dst_pte); 58628fb5debcSMike Kravetz spin_lock(ptl); 58638fb5debcSMike Kravetz 58641e392147SAndrea Arcangeli /* 58651e392147SAndrea Arcangeli * Recheck the i_size after holding PT lock to make sure not 58661e392147SAndrea Arcangeli * to leave any page mapped (as page_mapped()) beyond the end 58671e392147SAndrea Arcangeli * of the i_size (remove_inode_hugepages() is strict about 58681e392147SAndrea Arcangeli * enforcing that). If we bail out here, we'll also leave a 58691e392147SAndrea Arcangeli * page in the radix tree in the vm_shared case beyond the end 58701e392147SAndrea Arcangeli * of the i_size, but remove_inode_hugepages() will take care 58711e392147SAndrea Arcangeli * of it as soon as we drop the hugetlb_fault_mutex_table. 58721e392147SAndrea Arcangeli */ 58731e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 58741e392147SAndrea Arcangeli ret = -EFAULT; 58751e392147SAndrea Arcangeli if (idx >= size) 58761e392147SAndrea Arcangeli goto out_release_unlock; 58771e392147SAndrea Arcangeli 58788fb5debcSMike Kravetz ret = -EEXIST; 58798fb5debcSMike Kravetz if (!huge_pte_none(huge_ptep_get(dst_pte))) 58808fb5debcSMike Kravetz goto out_release_unlock; 58818fb5debcSMike Kravetz 58821c9e8defSMike Kravetz if (vm_shared) { 58831c9e8defSMike Kravetz page_dup_rmap(page, true); 58841c9e8defSMike Kravetz } else { 5885d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 58868fb5debcSMike Kravetz hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); 58871c9e8defSMike Kravetz } 58888fb5debcSMike Kravetz 5889f6191471SAxel Rasmussen /* For CONTINUE on a non-shared VMA, don't set VM_WRITE for CoW. */ 5890f6191471SAxel Rasmussen if (is_continue && !vm_shared) 5891f6191471SAxel Rasmussen writable = 0; 5892f6191471SAxel Rasmussen else 5893f6191471SAxel Rasmussen writable = dst_vma->vm_flags & VM_WRITE; 5894f6191471SAxel Rasmussen 5895f6191471SAxel Rasmussen _dst_pte = make_huge_pte(dst_vma, page, writable); 5896f6191471SAxel Rasmussen if (writable) 58978fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 58988fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 58998fb5debcSMike Kravetz 59008fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 59018fb5debcSMike Kravetz 59028fb5debcSMike Kravetz (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, 59038fb5debcSMike Kravetz dst_vma->vm_flags & VM_WRITE); 59048fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 59058fb5debcSMike Kravetz 59068fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 59078fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 59088fb5debcSMike Kravetz 59098fb5debcSMike Kravetz spin_unlock(ptl); 5910f6191471SAxel Rasmussen if (!is_continue) 59118f251a3dSMike Kravetz SetHPageMigratable(page); 5912f6191471SAxel Rasmussen if (vm_shared || is_continue) 59131c9e8defSMike Kravetz unlock_page(page); 59148fb5debcSMike Kravetz ret = 0; 59158fb5debcSMike Kravetz out: 59168fb5debcSMike Kravetz return ret; 59178fb5debcSMike Kravetz out_release_unlock: 59188fb5debcSMike Kravetz spin_unlock(ptl); 5919f6191471SAxel Rasmussen if (vm_shared || is_continue) 59201c9e8defSMike Kravetz unlock_page(page); 59215af10dfdSAndrea Arcangeli out_release_nounlock: 5922cc30042dSMina Almasry if (!page_in_pagecache) 5923846be085SMike Kravetz restore_reserve_on_error(h, dst_vma, dst_addr, page); 59248fb5debcSMike Kravetz put_page(page); 59258fb5debcSMike Kravetz goto out; 59268fb5debcSMike Kravetz } 5927714c1891SAxel Rasmussen #endif /* CONFIG_USERFAULTFD */ 59288fb5debcSMike Kravetz 592982e5d378SJoao Martins static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma, 593082e5d378SJoao Martins int refs, struct page **pages, 593182e5d378SJoao Martins struct vm_area_struct **vmas) 593282e5d378SJoao Martins { 593382e5d378SJoao Martins int nr; 593482e5d378SJoao Martins 593582e5d378SJoao Martins for (nr = 0; nr < refs; nr++) { 593682e5d378SJoao Martins if (likely(pages)) 593782e5d378SJoao Martins pages[nr] = mem_map_offset(page, nr); 593882e5d378SJoao Martins if (vmas) 593982e5d378SJoao Martins vmas[nr] = vma; 594082e5d378SJoao Martins } 594182e5d378SJoao Martins } 594282e5d378SJoao Martins 594328a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 594463551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 594528a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 59464f6da934SPeter Xu long i, unsigned int flags, int *locked) 594763551ae0SDavid Gibson { 5948d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 5949d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 595028a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 5951a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 59520fa5bc40SJoao Martins int err = -EFAULT, refs; 595363551ae0SDavid Gibson 595463551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 595563551ae0SDavid Gibson pte_t *pte; 5956cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 59572a15efc9SHugh Dickins int absent; 595863551ae0SDavid Gibson struct page *page; 595963551ae0SDavid Gibson 59604c887265SAdam Litke /* 596102057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 596202057967SDavid Rientjes * potentially allocating memory. 596302057967SDavid Rientjes */ 5964fa45f116SDavidlohr Bueso if (fatal_signal_pending(current)) { 596502057967SDavid Rientjes remainder = 0; 596602057967SDavid Rientjes break; 596702057967SDavid Rientjes } 596802057967SDavid Rientjes 596902057967SDavid Rientjes /* 59704c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 59712a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 59724c887265SAdam Litke * first, for the page indexing below to work. 5973cb900f41SKirill A. Shutemov * 5974cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 59754c887265SAdam Litke */ 59767868a208SPunit Agrawal pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), 59777868a208SPunit Agrawal huge_page_size(h)); 5978cb900f41SKirill A. Shutemov if (pte) 5979cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 59802a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 598163551ae0SDavid Gibson 59822a15efc9SHugh Dickins /* 59832a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 59843ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 59853ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 59863ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 59873ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 59882a15efc9SHugh Dickins */ 59893ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 59903ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 5991cb900f41SKirill A. Shutemov if (pte) 5992cb900f41SKirill A. Shutemov spin_unlock(ptl); 59932a15efc9SHugh Dickins remainder = 0; 59942a15efc9SHugh Dickins break; 59952a15efc9SHugh Dickins } 59962a15efc9SHugh Dickins 59979cc3a5bdSNaoya Horiguchi /* 59989cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 59999cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 60009cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 60019cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 60029cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 60039cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 60049cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 60059cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 60069cc3a5bdSNaoya Horiguchi */ 60079cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 6008106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 6009106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 60102b740303SSouptick Joarder vm_fault_t ret; 601187ffc118SAndrea Arcangeli unsigned int fault_flags = 0; 60124c887265SAdam Litke 6013cb900f41SKirill A. Shutemov if (pte) 6014cb900f41SKirill A. Shutemov spin_unlock(ptl); 601587ffc118SAndrea Arcangeli if (flags & FOLL_WRITE) 601687ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_WRITE; 60174f6da934SPeter Xu if (locked) 601871335f37SPeter Xu fault_flags |= FAULT_FLAG_ALLOW_RETRY | 601971335f37SPeter Xu FAULT_FLAG_KILLABLE; 602087ffc118SAndrea Arcangeli if (flags & FOLL_NOWAIT) 602187ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY | 602287ffc118SAndrea Arcangeli FAULT_FLAG_RETRY_NOWAIT; 602387ffc118SAndrea Arcangeli if (flags & FOLL_TRIED) { 60244426e945SPeter Xu /* 60254426e945SPeter Xu * Note: FAULT_FLAG_ALLOW_RETRY and 60264426e945SPeter Xu * FAULT_FLAG_TRIED can co-exist 60274426e945SPeter Xu */ 602887ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_TRIED; 602987ffc118SAndrea Arcangeli } 603087ffc118SAndrea Arcangeli ret = hugetlb_fault(mm, vma, vaddr, fault_flags); 603187ffc118SAndrea Arcangeli if (ret & VM_FAULT_ERROR) { 60322be7cfedSDaniel Jordan err = vm_fault_to_errno(ret, flags); 60331c59827dSHugh Dickins remainder = 0; 60341c59827dSHugh Dickins break; 60351c59827dSHugh Dickins } 603687ffc118SAndrea Arcangeli if (ret & VM_FAULT_RETRY) { 60374f6da934SPeter Xu if (locked && 60381ac25013SAndrea Arcangeli !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) 60394f6da934SPeter Xu *locked = 0; 604087ffc118SAndrea Arcangeli *nr_pages = 0; 604187ffc118SAndrea Arcangeli /* 604287ffc118SAndrea Arcangeli * VM_FAULT_RETRY must not return an 604387ffc118SAndrea Arcangeli * error, it will return zero 604487ffc118SAndrea Arcangeli * instead. 604587ffc118SAndrea Arcangeli * 604687ffc118SAndrea Arcangeli * No need to update "position" as the 604787ffc118SAndrea Arcangeli * caller will not check it after 604887ffc118SAndrea Arcangeli * *nr_pages is set to 0. 604987ffc118SAndrea Arcangeli */ 605087ffc118SAndrea Arcangeli return i; 605187ffc118SAndrea Arcangeli } 605287ffc118SAndrea Arcangeli continue; 605387ffc118SAndrea Arcangeli } 605463551ae0SDavid Gibson 6055a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 60567f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 60578fde12caSLinus Torvalds 60588fde12caSLinus Torvalds /* 6059acbfb087SZhigang Lu * If subpage information not requested, update counters 6060acbfb087SZhigang Lu * and skip the same_page loop below. 6061acbfb087SZhigang Lu */ 6062acbfb087SZhigang Lu if (!pages && !vmas && !pfn_offset && 6063acbfb087SZhigang Lu (vaddr + huge_page_size(h) < vma->vm_end) && 6064acbfb087SZhigang Lu (remainder >= pages_per_huge_page(h))) { 6065acbfb087SZhigang Lu vaddr += huge_page_size(h); 6066acbfb087SZhigang Lu remainder -= pages_per_huge_page(h); 6067acbfb087SZhigang Lu i += pages_per_huge_page(h); 6068acbfb087SZhigang Lu spin_unlock(ptl); 6069acbfb087SZhigang Lu continue; 6070acbfb087SZhigang Lu } 6071acbfb087SZhigang Lu 6072d08af0a5SJoao Martins /* vaddr may not be aligned to PAGE_SIZE */ 6073d08af0a5SJoao Martins refs = min3(pages_per_huge_page(h) - pfn_offset, remainder, 6074d08af0a5SJoao Martins (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT); 60750fa5bc40SJoao Martins 607682e5d378SJoao Martins if (pages || vmas) 607782e5d378SJoao Martins record_subpages_vmas(mem_map_offset(page, pfn_offset), 607882e5d378SJoao Martins vma, refs, 607982e5d378SJoao Martins likely(pages) ? pages + i : NULL, 608082e5d378SJoao Martins vmas ? vmas + i : NULL); 608163551ae0SDavid Gibson 608282e5d378SJoao Martins if (pages) { 60830fa5bc40SJoao Martins /* 6084822951d8SMatthew Wilcox (Oracle) * try_grab_folio() should always succeed here, 60850fa5bc40SJoao Martins * because: a) we hold the ptl lock, and b) we've just 60860fa5bc40SJoao Martins * checked that the huge page is present in the page 60870fa5bc40SJoao Martins * tables. If the huge page is present, then the tail 60880fa5bc40SJoao Martins * pages must also be present. The ptl prevents the 60890fa5bc40SJoao Martins * head page and tail pages from being rearranged in 60900fa5bc40SJoao Martins * any way. So this page must be available at this 60910fa5bc40SJoao Martins * point, unless the page refcount overflowed: 60920fa5bc40SJoao Martins */ 6093822951d8SMatthew Wilcox (Oracle) if (WARN_ON_ONCE(!try_grab_folio(pages[i], refs, 60940fa5bc40SJoao Martins flags))) { 60950fa5bc40SJoao Martins spin_unlock(ptl); 60960fa5bc40SJoao Martins remainder = 0; 60970fa5bc40SJoao Martins err = -ENOMEM; 60980fa5bc40SJoao Martins break; 60990fa5bc40SJoao Martins } 6100d5d4b0aaSChen, Kenneth W } 610182e5d378SJoao Martins 610282e5d378SJoao Martins vaddr += (refs << PAGE_SHIFT); 610382e5d378SJoao Martins remainder -= refs; 610482e5d378SJoao Martins i += refs; 610582e5d378SJoao Martins 6106cb900f41SKirill A. Shutemov spin_unlock(ptl); 610763551ae0SDavid Gibson } 610828a35716SMichel Lespinasse *nr_pages = remainder; 610987ffc118SAndrea Arcangeli /* 611087ffc118SAndrea Arcangeli * setting position is actually required only if remainder is 611187ffc118SAndrea Arcangeli * not zero but it's faster not to add a "if (remainder)" 611287ffc118SAndrea Arcangeli * branch. 611387ffc118SAndrea Arcangeli */ 611463551ae0SDavid Gibson *position = vaddr; 611563551ae0SDavid Gibson 61162be7cfedSDaniel Jordan return i ? i : err; 611763551ae0SDavid Gibson } 61188f860591SZhang, Yanmin 61197da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 61208f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 61218f860591SZhang, Yanmin { 61228f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 61238f860591SZhang, Yanmin unsigned long start = address; 61248f860591SZhang, Yanmin pte_t *ptep; 61258f860591SZhang, Yanmin pte_t pte; 6126a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 61277da4d641SPeter Zijlstra unsigned long pages = 0; 6128dff11abeSMike Kravetz bool shared_pmd = false; 6129ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 6130dff11abeSMike Kravetz 6131dff11abeSMike Kravetz /* 6132dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 6133ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 6134dff11abeSMike Kravetz * range if PMD sharing is possible. 6135dff11abeSMike Kravetz */ 61367269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 61377269f999SJérôme Glisse 0, vma, mm, start, end); 6138ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 61398f860591SZhang, Yanmin 61408f860591SZhang, Yanmin BUG_ON(address >= end); 6141ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 61428f860591SZhang, Yanmin 6143ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 614483cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 6145a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 6146cb900f41SKirill A. Shutemov spinlock_t *ptl; 61477868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, huge_page_size(h)); 61488f860591SZhang, Yanmin if (!ptep) 61498f860591SZhang, Yanmin continue; 6150cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 615134ae204fSMike Kravetz if (huge_pmd_unshare(mm, vma, &address, ptep)) { 61527da4d641SPeter Zijlstra pages++; 6153cb900f41SKirill A. Shutemov spin_unlock(ptl); 6154dff11abeSMike Kravetz shared_pmd = true; 615539dde65cSChen, Kenneth W continue; 61567da4d641SPeter Zijlstra } 6157a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 6158a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 6159a8bda28dSNaoya Horiguchi spin_unlock(ptl); 6160a8bda28dSNaoya Horiguchi continue; 6161a8bda28dSNaoya Horiguchi } 6162a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 6163a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 6164a8bda28dSNaoya Horiguchi 61654dd845b5SAlistair Popple if (is_writable_migration_entry(entry)) { 6166a8bda28dSNaoya Horiguchi pte_t newpte; 6167a8bda28dSNaoya Horiguchi 61684dd845b5SAlistair Popple entry = make_readable_migration_entry( 61694dd845b5SAlistair Popple swp_offset(entry)); 6170a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 6171e5251fd4SPunit Agrawal set_huge_swap_pte_at(mm, address, ptep, 6172e5251fd4SPunit Agrawal newpte, huge_page_size(h)); 6173a8bda28dSNaoya Horiguchi pages++; 6174a8bda28dSNaoya Horiguchi } 6175a8bda28dSNaoya Horiguchi spin_unlock(ptl); 6176a8bda28dSNaoya Horiguchi continue; 6177a8bda28dSNaoya Horiguchi } 6178a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 6179023bdd00SAneesh Kumar K.V pte_t old_pte; 618079c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 6181023bdd00SAneesh Kumar K.V 6182023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 618316785bd7SAnshuman Khandual pte = huge_pte_modify(old_pte, newprot); 618479c1c594SChristophe Leroy pte = arch_make_huge_pte(pte, shift, vma->vm_flags); 6185023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 61867da4d641SPeter Zijlstra pages++; 61878f860591SZhang, Yanmin } 6188cb900f41SKirill A. Shutemov spin_unlock(ptl); 61898f860591SZhang, Yanmin } 6190d833352aSMel Gorman /* 6191c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 6192d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 6193c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 6194dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 6195dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 6196d833352aSMel Gorman */ 6197dff11abeSMike Kravetz if (shared_pmd) 6198ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 6199dff11abeSMike Kravetz else 62005491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 62010f10851eSJérôme Glisse /* 62020f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 62030f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 62040f10851eSJérôme Glisse * 6205ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 62060f10851eSJérôme Glisse */ 620783cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 6208ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 62097da4d641SPeter Zijlstra 62107da4d641SPeter Zijlstra return pages << h->order; 62118f860591SZhang, Yanmin } 62128f860591SZhang, Yanmin 621333b8f84aSMike Kravetz /* Return true if reservation was successful, false otherwise. */ 621433b8f84aSMike Kravetz bool hugetlb_reserve_pages(struct inode *inode, 6215a1e78772SMel Gorman long from, long to, 62165a6fe125SMel Gorman struct vm_area_struct *vma, 6217ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 6218e4e574b7SAdam Litke { 621933b8f84aSMike Kravetz long chg, add = -1; 6220a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 622190481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 62229119a41eSJoonsoo Kim struct resv_map *resv_map; 6223075a61d0SMina Almasry struct hugetlb_cgroup *h_cg = NULL; 62240db9d74eSMina Almasry long gbl_reserve, regions_needed = 0; 6225e4e574b7SAdam Litke 622663489f8eSMike Kravetz /* This should never happen */ 622763489f8eSMike Kravetz if (from > to) { 622863489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 622933b8f84aSMike Kravetz return false; 623063489f8eSMike Kravetz } 623163489f8eSMike Kravetz 6232a1e78772SMel Gorman /* 623317c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 623417c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 623590481622SDavid Gibson * without using reserves 623617c9d12eSMel Gorman */ 6237ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 623833b8f84aSMike Kravetz return true; 623917c9d12eSMel Gorman 624017c9d12eSMel Gorman /* 6241a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 6242a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 6243a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 6244a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 6245a1e78772SMel Gorman */ 62469119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 6247f27a5136SMike Kravetz /* 6248f27a5136SMike Kravetz * resv_map can not be NULL as hugetlb_reserve_pages is only 6249f27a5136SMike Kravetz * called for inodes for which resv_maps were created (see 6250f27a5136SMike Kravetz * hugetlbfs_get_inode). 6251f27a5136SMike Kravetz */ 62524e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 62539119a41eSJoonsoo Kim 62540db9d74eSMina Almasry chg = region_chg(resv_map, from, to, ®ions_needed); 62559119a41eSJoonsoo Kim 62569119a41eSJoonsoo Kim } else { 6257e9fe92aeSMina Almasry /* Private mapping. */ 62589119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 62595a6fe125SMel Gorman if (!resv_map) 626033b8f84aSMike Kravetz return false; 62615a6fe125SMel Gorman 626217c9d12eSMel Gorman chg = to - from; 626317c9d12eSMel Gorman 62645a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 62655a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 62665a6fe125SMel Gorman } 62675a6fe125SMel Gorman 626833b8f84aSMike Kravetz if (chg < 0) 6269c50ac050SDave Hansen goto out_err; 627017c9d12eSMel Gorman 627133b8f84aSMike Kravetz if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), 627233b8f84aSMike Kravetz chg * pages_per_huge_page(h), &h_cg) < 0) 6273075a61d0SMina Almasry goto out_err; 6274075a61d0SMina Almasry 6275075a61d0SMina Almasry if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { 6276075a61d0SMina Almasry /* For private mappings, the hugetlb_cgroup uncharge info hangs 6277075a61d0SMina Almasry * of the resv_map. 6278075a61d0SMina Almasry */ 6279075a61d0SMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); 6280075a61d0SMina Almasry } 6281075a61d0SMina Almasry 62821c5ecae3SMike Kravetz /* 62831c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 62841c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 62851c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 62861c5ecae3SMike Kravetz */ 62871c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 628833b8f84aSMike Kravetz if (gbl_reserve < 0) 6289075a61d0SMina Almasry goto out_uncharge_cgroup; 629017c9d12eSMel Gorman 629117c9d12eSMel Gorman /* 629217c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 629390481622SDavid Gibson * Hand the pages back to the subpool if there are not 629417c9d12eSMel Gorman */ 629533b8f84aSMike Kravetz if (hugetlb_acct_memory(h, gbl_reserve) < 0) 6296075a61d0SMina Almasry goto out_put_pages; 629717c9d12eSMel Gorman 629817c9d12eSMel Gorman /* 629917c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 630017c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 630117c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 630217c9d12eSMel Gorman * the reservation was and the page cache tells how much of 630317c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 630417c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 630517c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 630617c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 630717c9d12eSMel Gorman * else has to be done for private mappings here 630817c9d12eSMel Gorman */ 630933039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 6310075a61d0SMina Almasry add = region_add(resv_map, from, to, regions_needed, h, h_cg); 631133039678SMike Kravetz 63120db9d74eSMina Almasry if (unlikely(add < 0)) { 63130db9d74eSMina Almasry hugetlb_acct_memory(h, -gbl_reserve); 6314075a61d0SMina Almasry goto out_put_pages; 63150db9d74eSMina Almasry } else if (unlikely(chg > add)) { 631633039678SMike Kravetz /* 631733039678SMike Kravetz * pages in this range were added to the reserve 631833039678SMike Kravetz * map between region_chg and region_add. This 631933039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 632033039678SMike Kravetz * the subpool and reserve counts modified above 632133039678SMike Kravetz * based on the difference. 632233039678SMike Kravetz */ 632333039678SMike Kravetz long rsv_adjust; 632433039678SMike Kravetz 6325d85aecf2SMiaohe Lin /* 6326d85aecf2SMiaohe Lin * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the 6327d85aecf2SMiaohe Lin * reference to h_cg->css. See comment below for detail. 6328d85aecf2SMiaohe Lin */ 6329075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd( 6330075a61d0SMina Almasry hstate_index(h), 6331075a61d0SMina Almasry (chg - add) * pages_per_huge_page(h), h_cg); 6332075a61d0SMina Almasry 633333039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 633433039678SMike Kravetz chg - add); 633533039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 6336d85aecf2SMiaohe Lin } else if (h_cg) { 6337d85aecf2SMiaohe Lin /* 6338d85aecf2SMiaohe Lin * The file_regions will hold their own reference to 6339d85aecf2SMiaohe Lin * h_cg->css. So we should release the reference held 6340d85aecf2SMiaohe Lin * via hugetlb_cgroup_charge_cgroup_rsvd() when we are 6341d85aecf2SMiaohe Lin * done. 6342d85aecf2SMiaohe Lin */ 6343d85aecf2SMiaohe Lin hugetlb_cgroup_put_rsvd_cgroup(h_cg); 634433039678SMike Kravetz } 634533039678SMike Kravetz } 634633b8f84aSMike Kravetz return true; 634733b8f84aSMike Kravetz 6348075a61d0SMina Almasry out_put_pages: 6349075a61d0SMina Almasry /* put back original number of pages, chg */ 6350075a61d0SMina Almasry (void)hugepage_subpool_put_pages(spool, chg); 6351075a61d0SMina Almasry out_uncharge_cgroup: 6352075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), 6353075a61d0SMina Almasry chg * pages_per_huge_page(h), h_cg); 6354c50ac050SDave Hansen out_err: 63555e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 63560db9d74eSMina Almasry /* Only call region_abort if the region_chg succeeded but the 63570db9d74eSMina Almasry * region_add failed or didn't run. 63580db9d74eSMina Almasry */ 63590db9d74eSMina Almasry if (chg >= 0 && add < 0) 63600db9d74eSMina Almasry region_abort(resv_map, from, to, regions_needed); 6361f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 6362f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 636333b8f84aSMike Kravetz return false; 6364a43a8c39SChen, Kenneth W } 6365a43a8c39SChen, Kenneth W 6366b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 6367b5cec28dSMike Kravetz long freed) 6368a43a8c39SChen, Kenneth W { 6369a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 63704e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 63719119a41eSJoonsoo Kim long chg = 0; 637290481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 63731c5ecae3SMike Kravetz long gbl_reserve; 637445c682a6SKen Chen 6375f27a5136SMike Kravetz /* 6376f27a5136SMike Kravetz * Since this routine can be called in the evict inode path for all 6377f27a5136SMike Kravetz * hugetlbfs inodes, resv_map could be NULL. 6378f27a5136SMike Kravetz */ 6379b5cec28dSMike Kravetz if (resv_map) { 6380b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 6381b5cec28dSMike Kravetz /* 6382b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 6383b5cec28dSMike Kravetz * must be split and another region descriptor can not be 6384b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 6385b5cec28dSMike Kravetz */ 6386b5cec28dSMike Kravetz if (chg < 0) 6387b5cec28dSMike Kravetz return chg; 6388b5cec28dSMike Kravetz } 6389b5cec28dSMike Kravetz 639045c682a6SKen Chen spin_lock(&inode->i_lock); 6391e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 639245c682a6SKen Chen spin_unlock(&inode->i_lock); 639345c682a6SKen Chen 63941c5ecae3SMike Kravetz /* 63951c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 63961c5ecae3SMike Kravetz * reservations to be released may be adjusted. 6397dddf31a4SMiaohe Lin * 6398dddf31a4SMiaohe Lin * Note that !resv_map implies freed == 0. So (chg - freed) 6399dddf31a4SMiaohe Lin * won't go negative. 64001c5ecae3SMike Kravetz */ 64011c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 64021c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 6403b5cec28dSMike Kravetz 6404b5cec28dSMike Kravetz return 0; 6405a43a8c39SChen, Kenneth W } 640693f70f90SNaoya Horiguchi 64073212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 64083212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 64093212b535SSteve Capper struct vm_area_struct *vma, 64103212b535SSteve Capper unsigned long addr, pgoff_t idx) 64113212b535SSteve Capper { 64123212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 64133212b535SSteve Capper svma->vm_start; 64143212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 64153212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 64163212b535SSteve Capper 64173212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 6418de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 6419de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 64203212b535SSteve Capper 64213212b535SSteve Capper /* 64223212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 64233212b535SSteve Capper * page table page. 64243212b535SSteve Capper */ 64253212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 64263212b535SSteve Capper vm_flags != svm_flags || 642707e51edfSMiaohe Lin !range_in_vma(svma, sbase, s_end)) 64283212b535SSteve Capper return 0; 64293212b535SSteve Capper 64303212b535SSteve Capper return saddr; 64313212b535SSteve Capper } 64323212b535SSteve Capper 643331aafb45SNicholas Krause static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) 64343212b535SSteve Capper { 64353212b535SSteve Capper unsigned long base = addr & PUD_MASK; 64363212b535SSteve Capper unsigned long end = base + PUD_SIZE; 64373212b535SSteve Capper 64383212b535SSteve Capper /* 64393212b535SSteve Capper * check on proper vm_flags and page table alignment 64403212b535SSteve Capper */ 6441017b1660SMike Kravetz if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) 644231aafb45SNicholas Krause return true; 644331aafb45SNicholas Krause return false; 64443212b535SSteve Capper } 64453212b535SSteve Capper 6446c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 6447c1991e07SPeter Xu { 6448c1991e07SPeter Xu #ifdef CONFIG_USERFAULTFD 6449c1991e07SPeter Xu if (uffd_disable_huge_pmd_share(vma)) 6450c1991e07SPeter Xu return false; 6451c1991e07SPeter Xu #endif 6452c1991e07SPeter Xu return vma_shareable(vma, addr); 6453c1991e07SPeter Xu } 6454c1991e07SPeter Xu 64553212b535SSteve Capper /* 6456017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 6457017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 6458017b1660SMike Kravetz * shared pmd mappings. 6459017b1660SMike Kravetz */ 6460017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 6461017b1660SMike Kravetz unsigned long *start, unsigned long *end) 6462017b1660SMike Kravetz { 6463a1ba9da8SLi Xinhai unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), 6464a1ba9da8SLi Xinhai v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 6465017b1660SMike Kravetz 6466a1ba9da8SLi Xinhai /* 6467f0953a1bSIngo Molnar * vma needs to span at least one aligned PUD size, and the range 6468f0953a1bSIngo Molnar * must be at least partially within in. 6469a1ba9da8SLi Xinhai */ 6470a1ba9da8SLi Xinhai if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || 6471a1ba9da8SLi Xinhai (*end <= v_start) || (*start >= v_end)) 6472017b1660SMike Kravetz return; 6473017b1660SMike Kravetz 647475802ca6SPeter Xu /* Extend the range to be PUD aligned for a worst case scenario */ 6475a1ba9da8SLi Xinhai if (*start > v_start) 6476a1ba9da8SLi Xinhai *start = ALIGN_DOWN(*start, PUD_SIZE); 6477017b1660SMike Kravetz 6478a1ba9da8SLi Xinhai if (*end < v_end) 6479a1ba9da8SLi Xinhai *end = ALIGN(*end, PUD_SIZE); 6480017b1660SMike Kravetz } 6481017b1660SMike Kravetz 6482017b1660SMike Kravetz /* 64833212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 64843212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 64853212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 6486c0d0381aSMike Kravetz * code much cleaner. 6487c0d0381aSMike Kravetz * 64880bf7b64eSMike Kravetz * This routine must be called with i_mmap_rwsem held in at least read mode if 64890bf7b64eSMike Kravetz * sharing is possible. For hugetlbfs, this prevents removal of any page 64900bf7b64eSMike Kravetz * table entries associated with the address space. This is important as we 64910bf7b64eSMike Kravetz * are setting up sharing based on existing page table entries (mappings). 64923212b535SSteve Capper */ 6493aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 6494aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 64953212b535SSteve Capper { 64963212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 64973212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 64983212b535SSteve Capper vma->vm_pgoff; 64993212b535SSteve Capper struct vm_area_struct *svma; 65003212b535SSteve Capper unsigned long saddr; 65013212b535SSteve Capper pte_t *spte = NULL; 65023212b535SSteve Capper pte_t *pte; 6503cb900f41SKirill A. Shutemov spinlock_t *ptl; 65043212b535SSteve Capper 65050bf7b64eSMike Kravetz i_mmap_assert_locked(mapping); 65063212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 65073212b535SSteve Capper if (svma == vma) 65083212b535SSteve Capper continue; 65093212b535SSteve Capper 65103212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 65113212b535SSteve Capper if (saddr) { 65127868a208SPunit Agrawal spte = huge_pte_offset(svma->vm_mm, saddr, 65137868a208SPunit Agrawal vma_mmu_pagesize(svma)); 65143212b535SSteve Capper if (spte) { 65153212b535SSteve Capper get_page(virt_to_page(spte)); 65163212b535SSteve Capper break; 65173212b535SSteve Capper } 65183212b535SSteve Capper } 65193212b535SSteve Capper } 65203212b535SSteve Capper 65213212b535SSteve Capper if (!spte) 65223212b535SSteve Capper goto out; 65233212b535SSteve Capper 65248bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 6525dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 65263212b535SSteve Capper pud_populate(mm, pud, 65273212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 6528c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 6529dc6c9a35SKirill A. Shutemov } else { 65303212b535SSteve Capper put_page(virt_to_page(spte)); 6531dc6c9a35SKirill A. Shutemov } 6532cb900f41SKirill A. Shutemov spin_unlock(ptl); 65333212b535SSteve Capper out: 65343212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 65353212b535SSteve Capper return pte; 65363212b535SSteve Capper } 65373212b535SSteve Capper 65383212b535SSteve Capper /* 65393212b535SSteve Capper * unmap huge page backed by shared pte. 65403212b535SSteve Capper * 65413212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 65423212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 65433212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 65443212b535SSteve Capper * 6545c0d0381aSMike Kravetz * Called with page table lock held and i_mmap_rwsem held in write mode. 65463212b535SSteve Capper * 65473212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 65483212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 65493212b535SSteve Capper */ 655034ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 655134ae204fSMike Kravetz unsigned long *addr, pte_t *ptep) 65523212b535SSteve Capper { 65533212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 6554c2febafcSKirill A. Shutemov p4d_t *p4d = p4d_offset(pgd, *addr); 6555c2febafcSKirill A. Shutemov pud_t *pud = pud_offset(p4d, *addr); 65563212b535SSteve Capper 655734ae204fSMike Kravetz i_mmap_assert_write_locked(vma->vm_file->f_mapping); 65583212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 65593212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 65603212b535SSteve Capper return 0; 65613212b535SSteve Capper 65623212b535SSteve Capper pud_clear(pud); 65633212b535SSteve Capper put_page(virt_to_page(ptep)); 6564dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 65653212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 65663212b535SSteve Capper return 1; 65673212b535SSteve Capper } 6568c1991e07SPeter Xu 65699e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 6570aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 6571aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 65729e5fc74cSSteve Capper { 65739e5fc74cSSteve Capper return NULL; 65749e5fc74cSSteve Capper } 6575e81f2d22SZhang Zhen 657634ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 657734ae204fSMike Kravetz unsigned long *addr, pte_t *ptep) 6578e81f2d22SZhang Zhen { 6579e81f2d22SZhang Zhen return 0; 6580e81f2d22SZhang Zhen } 6581017b1660SMike Kravetz 6582017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 6583017b1660SMike Kravetz unsigned long *start, unsigned long *end) 6584017b1660SMike Kravetz { 6585017b1660SMike Kravetz } 6586c1991e07SPeter Xu 6587c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 6588c1991e07SPeter Xu { 6589c1991e07SPeter Xu return false; 6590c1991e07SPeter Xu } 65913212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 65923212b535SSteve Capper 65939e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 6594aec44e0fSPeter Xu pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, 65959e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 65969e5fc74cSSteve Capper { 65979e5fc74cSSteve Capper pgd_t *pgd; 6598c2febafcSKirill A. Shutemov p4d_t *p4d; 65999e5fc74cSSteve Capper pud_t *pud; 66009e5fc74cSSteve Capper pte_t *pte = NULL; 66019e5fc74cSSteve Capper 66029e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 6603f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 6604f4f0a3d8SKirill A. Shutemov if (!p4d) 6605f4f0a3d8SKirill A. Shutemov return NULL; 6606c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 66079e5fc74cSSteve Capper if (pud) { 66089e5fc74cSSteve Capper if (sz == PUD_SIZE) { 66099e5fc74cSSteve Capper pte = (pte_t *)pud; 66109e5fc74cSSteve Capper } else { 66119e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 6612c1991e07SPeter Xu if (want_pmd_share(vma, addr) && pud_none(*pud)) 6613aec44e0fSPeter Xu pte = huge_pmd_share(mm, vma, addr, pud); 66149e5fc74cSSteve Capper else 66159e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 66169e5fc74cSSteve Capper } 66179e5fc74cSSteve Capper } 66184e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 66199e5fc74cSSteve Capper 66209e5fc74cSSteve Capper return pte; 66219e5fc74cSSteve Capper } 66229e5fc74cSSteve Capper 66239b19df29SPunit Agrawal /* 66249b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 66259b19df29SPunit Agrawal * entry at address @addr 66269b19df29SPunit Agrawal * 66278ac0b81aSLi Xinhai * Return: Pointer to page table entry (PUD or PMD) for 66288ac0b81aSLi Xinhai * address @addr, or NULL if a !p*d_present() entry is encountered and the 66299b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 66309b19df29SPunit Agrawal * table. 66319b19df29SPunit Agrawal */ 66327868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 66337868a208SPunit Agrawal unsigned long addr, unsigned long sz) 66349e5fc74cSSteve Capper { 66359e5fc74cSSteve Capper pgd_t *pgd; 6636c2febafcSKirill A. Shutemov p4d_t *p4d; 66378ac0b81aSLi Xinhai pud_t *pud; 66388ac0b81aSLi Xinhai pmd_t *pmd; 66399e5fc74cSSteve Capper 66409e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 6641c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 6642c2febafcSKirill A. Shutemov return NULL; 6643c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 6644c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 6645c2febafcSKirill A. Shutemov return NULL; 66469b19df29SPunit Agrawal 6647c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 66488ac0b81aSLi Xinhai if (sz == PUD_SIZE) 66498ac0b81aSLi Xinhai /* must be pud huge, non-present or none */ 66509e5fc74cSSteve Capper return (pte_t *)pud; 66518ac0b81aSLi Xinhai if (!pud_present(*pud)) 66528ac0b81aSLi Xinhai return NULL; 66538ac0b81aSLi Xinhai /* must have a valid entry and size to go further */ 66549b19df29SPunit Agrawal 66559e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 66568ac0b81aSLi Xinhai /* must be pmd huge, non-present or none */ 66579e5fc74cSSteve Capper return (pte_t *)pmd; 66589e5fc74cSSteve Capper } 66599e5fc74cSSteve Capper 666061f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 666161f77edaSNaoya Horiguchi 666261f77edaSNaoya Horiguchi /* 666361f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 666461f77edaSNaoya Horiguchi * behavior. 666561f77edaSNaoya Horiguchi */ 666661f77edaSNaoya Horiguchi struct page * __weak 666761f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 666861f77edaSNaoya Horiguchi int write) 666961f77edaSNaoya Horiguchi { 667061f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 667161f77edaSNaoya Horiguchi } 667261f77edaSNaoya Horiguchi 667361f77edaSNaoya Horiguchi struct page * __weak 66744dc71451SAneesh Kumar K.V follow_huge_pd(struct vm_area_struct *vma, 66754dc71451SAneesh Kumar K.V unsigned long address, hugepd_t hpd, int flags, int pdshift) 66764dc71451SAneesh Kumar K.V { 66774dc71451SAneesh Kumar K.V WARN(1, "hugepd follow called with no support for hugepage directory format\n"); 66784dc71451SAneesh Kumar K.V return NULL; 66794dc71451SAneesh Kumar K.V } 66804dc71451SAneesh Kumar K.V 66814dc71451SAneesh Kumar K.V struct page * __weak 66829e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 6683e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 66849e5fc74cSSteve Capper { 6685e66f17ffSNaoya Horiguchi struct page *page = NULL; 6686e66f17ffSNaoya Horiguchi spinlock_t *ptl; 6687c9d398faSNaoya Horiguchi pte_t pte; 66883faa52c0SJohn Hubbard 66893faa52c0SJohn Hubbard /* FOLL_GET and FOLL_PIN are mutually exclusive. */ 66903faa52c0SJohn Hubbard if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == 66913faa52c0SJohn Hubbard (FOLL_PIN | FOLL_GET))) 66923faa52c0SJohn Hubbard return NULL; 66933faa52c0SJohn Hubbard 6694e66f17ffSNaoya Horiguchi retry: 6695e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 6696e66f17ffSNaoya Horiguchi spin_lock(ptl); 6697e66f17ffSNaoya Horiguchi /* 6698e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 6699e66f17ffSNaoya Horiguchi * unmapped from other threads. 6700e66f17ffSNaoya Horiguchi */ 6701e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 6702e66f17ffSNaoya Horiguchi goto out; 6703c9d398faSNaoya Horiguchi pte = huge_ptep_get((pte_t *)pmd); 6704c9d398faSNaoya Horiguchi if (pte_present(pte)) { 670597534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 67063faa52c0SJohn Hubbard /* 67073faa52c0SJohn Hubbard * try_grab_page() should always succeed here, because: a) we 67083faa52c0SJohn Hubbard * hold the pmd (ptl) lock, and b) we've just checked that the 67093faa52c0SJohn Hubbard * huge pmd (head) page is present in the page tables. The ptl 67103faa52c0SJohn Hubbard * prevents the head page and tail pages from being rearranged 67113faa52c0SJohn Hubbard * in any way. So this page must be available at this point, 67123faa52c0SJohn Hubbard * unless the page refcount overflowed: 67133faa52c0SJohn Hubbard */ 67143faa52c0SJohn Hubbard if (WARN_ON_ONCE(!try_grab_page(page, flags))) { 67153faa52c0SJohn Hubbard page = NULL; 67163faa52c0SJohn Hubbard goto out; 67173faa52c0SJohn Hubbard } 6718e66f17ffSNaoya Horiguchi } else { 6719c9d398faSNaoya Horiguchi if (is_hugetlb_entry_migration(pte)) { 6720e66f17ffSNaoya Horiguchi spin_unlock(ptl); 6721e66f17ffSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pmd, ptl); 6722e66f17ffSNaoya Horiguchi goto retry; 6723e66f17ffSNaoya Horiguchi } 6724e66f17ffSNaoya Horiguchi /* 6725e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 6726e66f17ffSNaoya Horiguchi * follow_page_mask(). 6727e66f17ffSNaoya Horiguchi */ 6728e66f17ffSNaoya Horiguchi } 6729e66f17ffSNaoya Horiguchi out: 6730e66f17ffSNaoya Horiguchi spin_unlock(ptl); 67319e5fc74cSSteve Capper return page; 67329e5fc74cSSteve Capper } 67339e5fc74cSSteve Capper 673461f77edaSNaoya Horiguchi struct page * __weak 67359e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 6736e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 67379e5fc74cSSteve Capper { 67383faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 6739e66f17ffSNaoya Horiguchi return NULL; 67409e5fc74cSSteve Capper 6741e66f17ffSNaoya Horiguchi return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 67429e5fc74cSSteve Capper } 67439e5fc74cSSteve Capper 6744faaa5b62SAnshuman Khandual struct page * __weak 6745faaa5b62SAnshuman Khandual follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) 6746faaa5b62SAnshuman Khandual { 67473faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 6748faaa5b62SAnshuman Khandual return NULL; 6749faaa5b62SAnshuman Khandual 6750faaa5b62SAnshuman Khandual return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); 6751faaa5b62SAnshuman Khandual } 6752faaa5b62SAnshuman Khandual 675331caf665SNaoya Horiguchi bool isolate_huge_page(struct page *page, struct list_head *list) 675431caf665SNaoya Horiguchi { 6755bcc54222SNaoya Horiguchi bool ret = true; 6756bcc54222SNaoya Horiguchi 6757db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 67588f251a3dSMike Kravetz if (!PageHeadHuge(page) || 67598f251a3dSMike Kravetz !HPageMigratable(page) || 67600eb2df2bSMuchun Song !get_page_unless_zero(page)) { 6761bcc54222SNaoya Horiguchi ret = false; 6762bcc54222SNaoya Horiguchi goto unlock; 6763bcc54222SNaoya Horiguchi } 67648f251a3dSMike Kravetz ClearHPageMigratable(page); 676531caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 6766bcc54222SNaoya Horiguchi unlock: 6767db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 6768bcc54222SNaoya Horiguchi return ret; 676931caf665SNaoya Horiguchi } 677031caf665SNaoya Horiguchi 677125182f05SNaoya Horiguchi int get_hwpoison_huge_page(struct page *page, bool *hugetlb) 677225182f05SNaoya Horiguchi { 677325182f05SNaoya Horiguchi int ret = 0; 677425182f05SNaoya Horiguchi 677525182f05SNaoya Horiguchi *hugetlb = false; 677625182f05SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 677725182f05SNaoya Horiguchi if (PageHeadHuge(page)) { 677825182f05SNaoya Horiguchi *hugetlb = true; 677925182f05SNaoya Horiguchi if (HPageFreed(page) || HPageMigratable(page)) 678025182f05SNaoya Horiguchi ret = get_page_unless_zero(page); 67810ed950d1SNaoya Horiguchi else 67820ed950d1SNaoya Horiguchi ret = -EBUSY; 678325182f05SNaoya Horiguchi } 678425182f05SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 678525182f05SNaoya Horiguchi return ret; 678625182f05SNaoya Horiguchi } 678725182f05SNaoya Horiguchi 678831caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 678931caf665SNaoya Horiguchi { 6790db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 67918f251a3dSMike Kravetz SetHPageMigratable(page); 679231caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 6793db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 679431caf665SNaoya Horiguchi put_page(page); 679531caf665SNaoya Horiguchi } 6796ab5ac90aSMichal Hocko 6797ab5ac90aSMichal Hocko void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) 6798ab5ac90aSMichal Hocko { 6799ab5ac90aSMichal Hocko struct hstate *h = page_hstate(oldpage); 6800ab5ac90aSMichal Hocko 6801ab5ac90aSMichal Hocko hugetlb_cgroup_migrate(oldpage, newpage); 6802ab5ac90aSMichal Hocko set_page_owner_migrate_reason(newpage, reason); 6803ab5ac90aSMichal Hocko 6804ab5ac90aSMichal Hocko /* 6805ab5ac90aSMichal Hocko * transfer temporary state of the new huge page. This is 6806ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 6807ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 6808ab5ac90aSMichal Hocko * the temporary status. 6809ab5ac90aSMichal Hocko * 6810ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 6811ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 6812ab5ac90aSMichal Hocko * the per-node's. 6813ab5ac90aSMichal Hocko */ 68149157c311SMike Kravetz if (HPageTemporary(newpage)) { 6815ab5ac90aSMichal Hocko int old_nid = page_to_nid(oldpage); 6816ab5ac90aSMichal Hocko int new_nid = page_to_nid(newpage); 6817ab5ac90aSMichal Hocko 68189157c311SMike Kravetz SetHPageTemporary(oldpage); 68199157c311SMike Kravetz ClearHPageTemporary(newpage); 6820ab5ac90aSMichal Hocko 68215af1ab1dSMiaohe Lin /* 68225af1ab1dSMiaohe Lin * There is no need to transfer the per-node surplus state 68235af1ab1dSMiaohe Lin * when we do not cross the node. 68245af1ab1dSMiaohe Lin */ 68255af1ab1dSMiaohe Lin if (new_nid == old_nid) 68265af1ab1dSMiaohe Lin return; 6827db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 6828ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 6829ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 6830ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 6831ab5ac90aSMichal Hocko } 6832db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 6833ab5ac90aSMichal Hocko } 6834ab5ac90aSMichal Hocko } 6835cf11e85fSRoman Gushchin 68366dfeaff9SPeter Xu /* 68376dfeaff9SPeter Xu * This function will unconditionally remove all the shared pmd pgtable entries 68386dfeaff9SPeter Xu * within the specific vma for a hugetlbfs memory range. 68396dfeaff9SPeter Xu */ 68406dfeaff9SPeter Xu void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) 68416dfeaff9SPeter Xu { 68426dfeaff9SPeter Xu struct hstate *h = hstate_vma(vma); 68436dfeaff9SPeter Xu unsigned long sz = huge_page_size(h); 68446dfeaff9SPeter Xu struct mm_struct *mm = vma->vm_mm; 68456dfeaff9SPeter Xu struct mmu_notifier_range range; 68466dfeaff9SPeter Xu unsigned long address, start, end; 68476dfeaff9SPeter Xu spinlock_t *ptl; 68486dfeaff9SPeter Xu pte_t *ptep; 68496dfeaff9SPeter Xu 68506dfeaff9SPeter Xu if (!(vma->vm_flags & VM_MAYSHARE)) 68516dfeaff9SPeter Xu return; 68526dfeaff9SPeter Xu 68536dfeaff9SPeter Xu start = ALIGN(vma->vm_start, PUD_SIZE); 68546dfeaff9SPeter Xu end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 68556dfeaff9SPeter Xu 68566dfeaff9SPeter Xu if (start >= end) 68576dfeaff9SPeter Xu return; 68586dfeaff9SPeter Xu 68596dfeaff9SPeter Xu /* 68606dfeaff9SPeter Xu * No need to call adjust_range_if_pmd_sharing_possible(), because 68616dfeaff9SPeter Xu * we have already done the PUD_SIZE alignment. 68626dfeaff9SPeter Xu */ 68636dfeaff9SPeter Xu mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, 68646dfeaff9SPeter Xu start, end); 68656dfeaff9SPeter Xu mmu_notifier_invalidate_range_start(&range); 68666dfeaff9SPeter Xu i_mmap_lock_write(vma->vm_file->f_mapping); 68676dfeaff9SPeter Xu for (address = start; address < end; address += PUD_SIZE) { 68686dfeaff9SPeter Xu unsigned long tmp = address; 68696dfeaff9SPeter Xu 68706dfeaff9SPeter Xu ptep = huge_pte_offset(mm, address, sz); 68716dfeaff9SPeter Xu if (!ptep) 68726dfeaff9SPeter Xu continue; 68736dfeaff9SPeter Xu ptl = huge_pte_lock(h, mm, ptep); 68746dfeaff9SPeter Xu /* We don't want 'address' to be changed */ 68756dfeaff9SPeter Xu huge_pmd_unshare(mm, vma, &tmp, ptep); 68766dfeaff9SPeter Xu spin_unlock(ptl); 68776dfeaff9SPeter Xu } 68786dfeaff9SPeter Xu flush_hugetlb_tlb_range(vma, start, end); 68796dfeaff9SPeter Xu i_mmap_unlock_write(vma->vm_file->f_mapping); 68806dfeaff9SPeter Xu /* 68816dfeaff9SPeter Xu * No need to call mmu_notifier_invalidate_range(), see 68826dfeaff9SPeter Xu * Documentation/vm/mmu_notifier.rst. 68836dfeaff9SPeter Xu */ 68846dfeaff9SPeter Xu mmu_notifier_invalidate_range_end(&range); 68856dfeaff9SPeter Xu } 68866dfeaff9SPeter Xu 6887cf11e85fSRoman Gushchin #ifdef CONFIG_CMA 6888cf11e85fSRoman Gushchin static bool cma_reserve_called __initdata; 6889cf11e85fSRoman Gushchin 6890cf11e85fSRoman Gushchin static int __init cmdline_parse_hugetlb_cma(char *p) 6891cf11e85fSRoman Gushchin { 689238e719abSBaolin Wang int nid, count = 0; 689338e719abSBaolin Wang unsigned long tmp; 689438e719abSBaolin Wang char *s = p; 689538e719abSBaolin Wang 689638e719abSBaolin Wang while (*s) { 689738e719abSBaolin Wang if (sscanf(s, "%lu%n", &tmp, &count) != 1) 689838e719abSBaolin Wang break; 689938e719abSBaolin Wang 690038e719abSBaolin Wang if (s[count] == ':') { 6901f9317f77SMike Kravetz if (tmp >= MAX_NUMNODES) 690238e719abSBaolin Wang break; 6903f9317f77SMike Kravetz nid = array_index_nospec(tmp, MAX_NUMNODES); 690438e719abSBaolin Wang 690538e719abSBaolin Wang s += count + 1; 690638e719abSBaolin Wang tmp = memparse(s, &s); 690738e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = tmp; 690838e719abSBaolin Wang hugetlb_cma_size += tmp; 690938e719abSBaolin Wang 691038e719abSBaolin Wang /* 691138e719abSBaolin Wang * Skip the separator if have one, otherwise 691238e719abSBaolin Wang * break the parsing. 691338e719abSBaolin Wang */ 691438e719abSBaolin Wang if (*s == ',') 691538e719abSBaolin Wang s++; 691638e719abSBaolin Wang else 691738e719abSBaolin Wang break; 691838e719abSBaolin Wang } else { 6919cf11e85fSRoman Gushchin hugetlb_cma_size = memparse(p, &p); 692038e719abSBaolin Wang break; 692138e719abSBaolin Wang } 692238e719abSBaolin Wang } 692338e719abSBaolin Wang 6924cf11e85fSRoman Gushchin return 0; 6925cf11e85fSRoman Gushchin } 6926cf11e85fSRoman Gushchin 6927cf11e85fSRoman Gushchin early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); 6928cf11e85fSRoman Gushchin 6929cf11e85fSRoman Gushchin void __init hugetlb_cma_reserve(int order) 6930cf11e85fSRoman Gushchin { 6931cf11e85fSRoman Gushchin unsigned long size, reserved, per_node; 693238e719abSBaolin Wang bool node_specific_cma_alloc = false; 6933cf11e85fSRoman Gushchin int nid; 6934cf11e85fSRoman Gushchin 6935cf11e85fSRoman Gushchin cma_reserve_called = true; 6936cf11e85fSRoman Gushchin 6937cf11e85fSRoman Gushchin if (!hugetlb_cma_size) 6938cf11e85fSRoman Gushchin return; 6939cf11e85fSRoman Gushchin 694038e719abSBaolin Wang for (nid = 0; nid < MAX_NUMNODES; nid++) { 694138e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 694238e719abSBaolin Wang continue; 694338e719abSBaolin Wang 694438e719abSBaolin Wang if (!node_state(nid, N_ONLINE)) { 694538e719abSBaolin Wang pr_warn("hugetlb_cma: invalid node %d specified\n", nid); 694638e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 694738e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 694838e719abSBaolin Wang continue; 694938e719abSBaolin Wang } 695038e719abSBaolin Wang 695138e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) { 695238e719abSBaolin Wang pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n", 695338e719abSBaolin Wang nid, (PAGE_SIZE << order) / SZ_1M); 695438e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 695538e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 695638e719abSBaolin Wang } else { 695738e719abSBaolin Wang node_specific_cma_alloc = true; 695838e719abSBaolin Wang } 695938e719abSBaolin Wang } 696038e719abSBaolin Wang 696138e719abSBaolin Wang /* Validate the CMA size again in case some invalid nodes specified. */ 696238e719abSBaolin Wang if (!hugetlb_cma_size) 696338e719abSBaolin Wang return; 696438e719abSBaolin Wang 6965cf11e85fSRoman Gushchin if (hugetlb_cma_size < (PAGE_SIZE << order)) { 6966cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", 6967cf11e85fSRoman Gushchin (PAGE_SIZE << order) / SZ_1M); 6968a01f4390SMike Kravetz hugetlb_cma_size = 0; 6969cf11e85fSRoman Gushchin return; 6970cf11e85fSRoman Gushchin } 6971cf11e85fSRoman Gushchin 697238e719abSBaolin Wang if (!node_specific_cma_alloc) { 6973cf11e85fSRoman Gushchin /* 6974cf11e85fSRoman Gushchin * If 3 GB area is requested on a machine with 4 numa nodes, 6975cf11e85fSRoman Gushchin * let's allocate 1 GB on first three nodes and ignore the last one. 6976cf11e85fSRoman Gushchin */ 6977cf11e85fSRoman Gushchin per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); 6978cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", 6979cf11e85fSRoman Gushchin hugetlb_cma_size / SZ_1M, per_node / SZ_1M); 698038e719abSBaolin Wang } 6981cf11e85fSRoman Gushchin 6982cf11e85fSRoman Gushchin reserved = 0; 6983cf11e85fSRoman Gushchin for_each_node_state(nid, N_ONLINE) { 6984cf11e85fSRoman Gushchin int res; 69852281f797SBarry Song char name[CMA_MAX_NAME]; 6986cf11e85fSRoman Gushchin 698738e719abSBaolin Wang if (node_specific_cma_alloc) { 698838e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 698938e719abSBaolin Wang continue; 699038e719abSBaolin Wang 699138e719abSBaolin Wang size = hugetlb_cma_size_in_node[nid]; 699238e719abSBaolin Wang } else { 6993cf11e85fSRoman Gushchin size = min(per_node, hugetlb_cma_size - reserved); 699438e719abSBaolin Wang } 699538e719abSBaolin Wang 6996cf11e85fSRoman Gushchin size = round_up(size, PAGE_SIZE << order); 6997cf11e85fSRoman Gushchin 69982281f797SBarry Song snprintf(name, sizeof(name), "hugetlb%d", nid); 6999a01f4390SMike Kravetz /* 7000a01f4390SMike Kravetz * Note that 'order per bit' is based on smallest size that 7001a01f4390SMike Kravetz * may be returned to CMA allocator in the case of 7002a01f4390SMike Kravetz * huge page demotion. 7003a01f4390SMike Kravetz */ 7004a01f4390SMike Kravetz res = cma_declare_contiguous_nid(0, size, 0, 7005a01f4390SMike Kravetz PAGE_SIZE << HUGETLB_PAGE_ORDER, 700629d0f41dSBarry Song 0, false, name, 7007cf11e85fSRoman Gushchin &hugetlb_cma[nid], nid); 7008cf11e85fSRoman Gushchin if (res) { 7009cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: reservation failed: err %d, node %d", 7010cf11e85fSRoman Gushchin res, nid); 7011cf11e85fSRoman Gushchin continue; 7012cf11e85fSRoman Gushchin } 7013cf11e85fSRoman Gushchin 7014cf11e85fSRoman Gushchin reserved += size; 7015cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", 7016cf11e85fSRoman Gushchin size / SZ_1M, nid); 7017cf11e85fSRoman Gushchin 7018cf11e85fSRoman Gushchin if (reserved >= hugetlb_cma_size) 7019cf11e85fSRoman Gushchin break; 7020cf11e85fSRoman Gushchin } 7021a01f4390SMike Kravetz 7022a01f4390SMike Kravetz if (!reserved) 7023a01f4390SMike Kravetz /* 7024a01f4390SMike Kravetz * hugetlb_cma_size is used to determine if allocations from 7025a01f4390SMike Kravetz * cma are possible. Set to zero if no cma regions are set up. 7026a01f4390SMike Kravetz */ 7027a01f4390SMike Kravetz hugetlb_cma_size = 0; 7028cf11e85fSRoman Gushchin } 7029cf11e85fSRoman Gushchin 7030cf11e85fSRoman Gushchin void __init hugetlb_cma_check(void) 7031cf11e85fSRoman Gushchin { 7032cf11e85fSRoman Gushchin if (!hugetlb_cma_size || cma_reserve_called) 7033cf11e85fSRoman Gushchin return; 7034cf11e85fSRoman Gushchin 7035cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); 7036cf11e85fSRoman Gushchin } 7037cf11e85fSRoman Gushchin 7038cf11e85fSRoman Gushchin #endif /* CONFIG_CMA */ 7039