1457c8996SThomas Gleixner // SPDX-License-Identifier: GPL-2.0-only 21da177e4SLinus Torvalds /* 31da177e4SLinus Torvalds * Generic hugetlb support. 46d49e352SNadia Yvette Chambers * (C) Nadia Yvette Chambers, April 2004 51da177e4SLinus Torvalds */ 61da177e4SLinus Torvalds #include <linux/list.h> 71da177e4SLinus Torvalds #include <linux/init.h> 81da177e4SLinus Torvalds #include <linux/mm.h> 9e1759c21SAlexey Dobriyan #include <linux/seq_file.h> 101da177e4SLinus Torvalds #include <linux/sysctl.h> 111da177e4SLinus Torvalds #include <linux/highmem.h> 12cddb8a5cSAndrea Arcangeli #include <linux/mmu_notifier.h> 131da177e4SLinus Torvalds #include <linux/nodemask.h> 1463551ae0SDavid Gibson #include <linux/pagemap.h> 155da7ca86SChristoph Lameter #include <linux/mempolicy.h> 163b32123dSGideon Israel Dsouza #include <linux/compiler.h> 17aea47ff3SChristoph Lameter #include <linux/cpuset.h> 183935baa9SDavid Gibson #include <linux/mutex.h> 1997ad1087SMike Rapoport #include <linux/memblock.h> 20a3437870SNishanth Aravamudan #include <linux/sysfs.h> 215a0e3ad6STejun Heo #include <linux/slab.h> 22bbe88753SJoonsoo Kim #include <linux/sched/mm.h> 2363489f8eSMike Kravetz #include <linux/mmdebug.h> 24174cd4b1SIngo Molnar #include <linux/sched/signal.h> 250fe6e20bSNaoya Horiguchi #include <linux/rmap.h> 26c6247f72SMatthew Wilcox #include <linux/string_helpers.h> 27fd6a03edSNaoya Horiguchi #include <linux/swap.h> 28fd6a03edSNaoya Horiguchi #include <linux/swapops.h> 298382d914SDavidlohr Bueso #include <linux/jhash.h> 3098fa15f3SAnshuman Khandual #include <linux/numa.h> 31c77c0a8aSWaiman Long #include <linux/llist.h> 32cf11e85fSRoman Gushchin #include <linux/cma.h> 338cc5fcbbSMina Almasry #include <linux/migrate.h> 34d6606683SLinus Torvalds 3563551ae0SDavid Gibson #include <asm/page.h> 36ca15ca40SMike Rapoport #include <asm/pgalloc.h> 3724669e58SAneesh Kumar K.V #include <asm/tlb.h> 3863551ae0SDavid Gibson 3924669e58SAneesh Kumar K.V #include <linux/io.h> 4063551ae0SDavid Gibson #include <linux/hugetlb.h> 419dd540e2SAneesh Kumar K.V #include <linux/hugetlb_cgroup.h> 429a305230SLee Schermerhorn #include <linux/node.h> 43ab5ac90aSMichal Hocko #include <linux/page_owner.h> 447835e98bSNick Piggin #include "internal.h" 45f41f2ed4SMuchun Song #include "hugetlb_vmemmap.h" 461da177e4SLinus Torvalds 47c3f38a38SAneesh Kumar K.V int hugetlb_max_hstate __read_mostly; 48e5ff2159SAndi Kleen unsigned int default_hstate_idx; 49e5ff2159SAndi Kleen struct hstate hstates[HUGE_MAX_HSTATE]; 50cf11e85fSRoman Gushchin 51dbda8feaSBarry Song #ifdef CONFIG_CMA 52cf11e85fSRoman Gushchin static struct cma *hugetlb_cma[MAX_NUMNODES]; 5338e719abSBaolin Wang static unsigned long hugetlb_cma_size_in_node[MAX_NUMNODES] __initdata; 54a01f4390SMike Kravetz static bool hugetlb_cma_page(struct page *page, unsigned int order) 55a01f4390SMike Kravetz { 56a01f4390SMike Kravetz return cma_pages_valid(hugetlb_cma[page_to_nid(page)], page, 57a01f4390SMike Kravetz 1 << order); 58a01f4390SMike Kravetz } 59a01f4390SMike Kravetz #else 60a01f4390SMike Kravetz static bool hugetlb_cma_page(struct page *page, unsigned int order) 61a01f4390SMike Kravetz { 62a01f4390SMike Kravetz return false; 63a01f4390SMike Kravetz } 64dbda8feaSBarry Song #endif 65dbda8feaSBarry Song static unsigned long hugetlb_cma_size __initdata; 66cf11e85fSRoman Gushchin 67641844f5SNaoya Horiguchi /* 68641844f5SNaoya Horiguchi * Minimum page order among possible hugepage sizes, set to a proper value 69641844f5SNaoya Horiguchi * at boot time. 70641844f5SNaoya Horiguchi */ 71641844f5SNaoya Horiguchi static unsigned int minimum_order __read_mostly = UINT_MAX; 72e5ff2159SAndi Kleen 7353ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 7453ba51d2SJon Tollefson 75e5ff2159SAndi Kleen /* for command line parsing */ 76e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 77e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 789fee021dSVaishali Thakkar static bool __initdata parsed_valid_hugepagesz = true; 79282f4214SMike Kravetz static bool __initdata parsed_default_hugepagesz; 80e5ff2159SAndi Kleen 813935baa9SDavid Gibson /* 8231caf665SNaoya Horiguchi * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, 8331caf665SNaoya Horiguchi * free_huge_pages, and surplus_huge_pages. 843935baa9SDavid Gibson */ 85c3f38a38SAneesh Kumar K.V DEFINE_SPINLOCK(hugetlb_lock); 860bd0f9fbSEric Paris 878382d914SDavidlohr Bueso /* 888382d914SDavidlohr Bueso * Serializes faults on the same logical page. This is used to 898382d914SDavidlohr Bueso * prevent spurious OOMs when the hugepage pool is fully utilized. 908382d914SDavidlohr Bueso */ 918382d914SDavidlohr Bueso static int num_fault_mutexes; 92c672c7f2SMike Kravetz struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; 938382d914SDavidlohr Bueso 947ca02d0aSMike Kravetz /* Forward declaration */ 957ca02d0aSMike Kravetz static int hugetlb_acct_memory(struct hstate *h, long delta); 967ca02d0aSMike Kravetz 971d88433bSMiaohe Lin static inline bool subpool_is_free(struct hugepage_subpool *spool) 981d88433bSMiaohe Lin { 991d88433bSMiaohe Lin if (spool->count) 1001d88433bSMiaohe Lin return false; 1011d88433bSMiaohe Lin if (spool->max_hpages != -1) 1021d88433bSMiaohe Lin return spool->used_hpages == 0; 1031d88433bSMiaohe Lin if (spool->min_hpages != -1) 1041d88433bSMiaohe Lin return spool->rsv_hpages == spool->min_hpages; 1051d88433bSMiaohe Lin 1061d88433bSMiaohe Lin return true; 1071d88433bSMiaohe Lin } 1081d88433bSMiaohe Lin 109db71ef79SMike Kravetz static inline void unlock_or_release_subpool(struct hugepage_subpool *spool, 110db71ef79SMike Kravetz unsigned long irq_flags) 11190481622SDavid Gibson { 112db71ef79SMike Kravetz spin_unlock_irqrestore(&spool->lock, irq_flags); 11390481622SDavid Gibson 11490481622SDavid Gibson /* If no pages are used, and no other handles to the subpool 1157c8de358SEthon Paul * remain, give up any reservations based on minimum size and 1167ca02d0aSMike Kravetz * free the subpool */ 1171d88433bSMiaohe Lin if (subpool_is_free(spool)) { 1187ca02d0aSMike Kravetz if (spool->min_hpages != -1) 1197ca02d0aSMike Kravetz hugetlb_acct_memory(spool->hstate, 1207ca02d0aSMike Kravetz -spool->min_hpages); 12190481622SDavid Gibson kfree(spool); 12290481622SDavid Gibson } 1237ca02d0aSMike Kravetz } 12490481622SDavid Gibson 1257ca02d0aSMike Kravetz struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, 1267ca02d0aSMike Kravetz long min_hpages) 12790481622SDavid Gibson { 12890481622SDavid Gibson struct hugepage_subpool *spool; 12990481622SDavid Gibson 130c6a91820SMike Kravetz spool = kzalloc(sizeof(*spool), GFP_KERNEL); 13190481622SDavid Gibson if (!spool) 13290481622SDavid Gibson return NULL; 13390481622SDavid Gibson 13490481622SDavid Gibson spin_lock_init(&spool->lock); 13590481622SDavid Gibson spool->count = 1; 1367ca02d0aSMike Kravetz spool->max_hpages = max_hpages; 1377ca02d0aSMike Kravetz spool->hstate = h; 1387ca02d0aSMike Kravetz spool->min_hpages = min_hpages; 1397ca02d0aSMike Kravetz 1407ca02d0aSMike Kravetz if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { 1417ca02d0aSMike Kravetz kfree(spool); 1427ca02d0aSMike Kravetz return NULL; 1437ca02d0aSMike Kravetz } 1447ca02d0aSMike Kravetz spool->rsv_hpages = min_hpages; 14590481622SDavid Gibson 14690481622SDavid Gibson return spool; 14790481622SDavid Gibson } 14890481622SDavid Gibson 14990481622SDavid Gibson void hugepage_put_subpool(struct hugepage_subpool *spool) 15090481622SDavid Gibson { 151db71ef79SMike Kravetz unsigned long flags; 152db71ef79SMike Kravetz 153db71ef79SMike Kravetz spin_lock_irqsave(&spool->lock, flags); 15490481622SDavid Gibson BUG_ON(!spool->count); 15590481622SDavid Gibson spool->count--; 156db71ef79SMike Kravetz unlock_or_release_subpool(spool, flags); 15790481622SDavid Gibson } 15890481622SDavid Gibson 1591c5ecae3SMike Kravetz /* 1601c5ecae3SMike Kravetz * Subpool accounting for allocating and reserving pages. 1611c5ecae3SMike Kravetz * Return -ENOMEM if there are not enough resources to satisfy the 1629e7ee400SRandy Dunlap * request. Otherwise, return the number of pages by which the 1631c5ecae3SMike Kravetz * global pools must be adjusted (upward). The returned value may 1641c5ecae3SMike Kravetz * only be different than the passed value (delta) in the case where 1657c8de358SEthon Paul * a subpool minimum size must be maintained. 1661c5ecae3SMike Kravetz */ 1671c5ecae3SMike Kravetz static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, 16890481622SDavid Gibson long delta) 16990481622SDavid Gibson { 1701c5ecae3SMike Kravetz long ret = delta; 17190481622SDavid Gibson 17290481622SDavid Gibson if (!spool) 1731c5ecae3SMike Kravetz return ret; 17490481622SDavid Gibson 175db71ef79SMike Kravetz spin_lock_irq(&spool->lock); 17690481622SDavid Gibson 1771c5ecae3SMike Kravetz if (spool->max_hpages != -1) { /* maximum size accounting */ 1781c5ecae3SMike Kravetz if ((spool->used_hpages + delta) <= spool->max_hpages) 1791c5ecae3SMike Kravetz spool->used_hpages += delta; 1801c5ecae3SMike Kravetz else { 1811c5ecae3SMike Kravetz ret = -ENOMEM; 1821c5ecae3SMike Kravetz goto unlock_ret; 1831c5ecae3SMike Kravetz } 1841c5ecae3SMike Kravetz } 1851c5ecae3SMike Kravetz 18609a95e29SMike Kravetz /* minimum size accounting */ 18709a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->rsv_hpages) { 1881c5ecae3SMike Kravetz if (delta > spool->rsv_hpages) { 1891c5ecae3SMike Kravetz /* 1901c5ecae3SMike Kravetz * Asking for more reserves than those already taken on 1911c5ecae3SMike Kravetz * behalf of subpool. Return difference. 1921c5ecae3SMike Kravetz */ 1931c5ecae3SMike Kravetz ret = delta - spool->rsv_hpages; 1941c5ecae3SMike Kravetz spool->rsv_hpages = 0; 1951c5ecae3SMike Kravetz } else { 1961c5ecae3SMike Kravetz ret = 0; /* reserves already accounted for */ 1971c5ecae3SMike Kravetz spool->rsv_hpages -= delta; 1981c5ecae3SMike Kravetz } 1991c5ecae3SMike Kravetz } 2001c5ecae3SMike Kravetz 2011c5ecae3SMike Kravetz unlock_ret: 202db71ef79SMike Kravetz spin_unlock_irq(&spool->lock); 20390481622SDavid Gibson return ret; 20490481622SDavid Gibson } 20590481622SDavid Gibson 2061c5ecae3SMike Kravetz /* 2071c5ecae3SMike Kravetz * Subpool accounting for freeing and unreserving pages. 2081c5ecae3SMike Kravetz * Return the number of global page reservations that must be dropped. 2091c5ecae3SMike Kravetz * The return value may only be different than the passed value (delta) 2101c5ecae3SMike Kravetz * in the case where a subpool minimum size must be maintained. 2111c5ecae3SMike Kravetz */ 2121c5ecae3SMike Kravetz static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, 21390481622SDavid Gibson long delta) 21490481622SDavid Gibson { 2151c5ecae3SMike Kravetz long ret = delta; 216db71ef79SMike Kravetz unsigned long flags; 2171c5ecae3SMike Kravetz 21890481622SDavid Gibson if (!spool) 2191c5ecae3SMike Kravetz return delta; 22090481622SDavid Gibson 221db71ef79SMike Kravetz spin_lock_irqsave(&spool->lock, flags); 2221c5ecae3SMike Kravetz 2231c5ecae3SMike Kravetz if (spool->max_hpages != -1) /* maximum size accounting */ 22490481622SDavid Gibson spool->used_hpages -= delta; 2251c5ecae3SMike Kravetz 22609a95e29SMike Kravetz /* minimum size accounting */ 22709a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { 2281c5ecae3SMike Kravetz if (spool->rsv_hpages + delta <= spool->min_hpages) 2291c5ecae3SMike Kravetz ret = 0; 2301c5ecae3SMike Kravetz else 2311c5ecae3SMike Kravetz ret = spool->rsv_hpages + delta - spool->min_hpages; 2321c5ecae3SMike Kravetz 2331c5ecae3SMike Kravetz spool->rsv_hpages += delta; 2341c5ecae3SMike Kravetz if (spool->rsv_hpages > spool->min_hpages) 2351c5ecae3SMike Kravetz spool->rsv_hpages = spool->min_hpages; 2361c5ecae3SMike Kravetz } 2371c5ecae3SMike Kravetz 2381c5ecae3SMike Kravetz /* 2391c5ecae3SMike Kravetz * If hugetlbfs_put_super couldn't free spool due to an outstanding 2401c5ecae3SMike Kravetz * quota reference, free it now. 2411c5ecae3SMike Kravetz */ 242db71ef79SMike Kravetz unlock_or_release_subpool(spool, flags); 2431c5ecae3SMike Kravetz 2441c5ecae3SMike Kravetz return ret; 24590481622SDavid Gibson } 24690481622SDavid Gibson 24790481622SDavid Gibson static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 24890481622SDavid Gibson { 24990481622SDavid Gibson return HUGETLBFS_SB(inode->i_sb)->spool; 25090481622SDavid Gibson } 25190481622SDavid Gibson 25290481622SDavid Gibson static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 25390481622SDavid Gibson { 254496ad9aaSAl Viro return subpool_inode(file_inode(vma->vm_file)); 25590481622SDavid Gibson } 25690481622SDavid Gibson 2570db9d74eSMina Almasry /* Helper that removes a struct file_region from the resv_map cache and returns 2580db9d74eSMina Almasry * it for use. 2590db9d74eSMina Almasry */ 2600db9d74eSMina Almasry static struct file_region * 2610db9d74eSMina Almasry get_file_region_entry_from_cache(struct resv_map *resv, long from, long to) 2620db9d74eSMina Almasry { 2630db9d74eSMina Almasry struct file_region *nrg = NULL; 2640db9d74eSMina Almasry 2650db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count <= 0); 2660db9d74eSMina Almasry 2670db9d74eSMina Almasry resv->region_cache_count--; 2680db9d74eSMina Almasry nrg = list_first_entry(&resv->region_cache, struct file_region, link); 2690db9d74eSMina Almasry list_del(&nrg->link); 2700db9d74eSMina Almasry 2710db9d74eSMina Almasry nrg->from = from; 2720db9d74eSMina Almasry nrg->to = to; 2730db9d74eSMina Almasry 2740db9d74eSMina Almasry return nrg; 2750db9d74eSMina Almasry } 2760db9d74eSMina Almasry 277075a61d0SMina Almasry static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, 278075a61d0SMina Almasry struct file_region *rg) 279075a61d0SMina Almasry { 280075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 281075a61d0SMina Almasry nrg->reservation_counter = rg->reservation_counter; 282075a61d0SMina Almasry nrg->css = rg->css; 283075a61d0SMina Almasry if (rg->css) 284075a61d0SMina Almasry css_get(rg->css); 285075a61d0SMina Almasry #endif 286075a61d0SMina Almasry } 287075a61d0SMina Almasry 288075a61d0SMina Almasry /* Helper that records hugetlb_cgroup uncharge info. */ 289075a61d0SMina Almasry static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, 290075a61d0SMina Almasry struct hstate *h, 291075a61d0SMina Almasry struct resv_map *resv, 292075a61d0SMina Almasry struct file_region *nrg) 293075a61d0SMina Almasry { 294075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 295075a61d0SMina Almasry if (h_cg) { 296075a61d0SMina Almasry nrg->reservation_counter = 297075a61d0SMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 298075a61d0SMina Almasry nrg->css = &h_cg->css; 299d85aecf2SMiaohe Lin /* 300d85aecf2SMiaohe Lin * The caller will hold exactly one h_cg->css reference for the 301d85aecf2SMiaohe Lin * whole contiguous reservation region. But this area might be 302d85aecf2SMiaohe Lin * scattered when there are already some file_regions reside in 303d85aecf2SMiaohe Lin * it. As a result, many file_regions may share only one css 304d85aecf2SMiaohe Lin * reference. In order to ensure that one file_region must hold 305d85aecf2SMiaohe Lin * exactly one h_cg->css reference, we should do css_get for 306d85aecf2SMiaohe Lin * each file_region and leave the reference held by caller 307d85aecf2SMiaohe Lin * untouched. 308d85aecf2SMiaohe Lin */ 309d85aecf2SMiaohe Lin css_get(&h_cg->css); 310075a61d0SMina Almasry if (!resv->pages_per_hpage) 311075a61d0SMina Almasry resv->pages_per_hpage = pages_per_huge_page(h); 312075a61d0SMina Almasry /* pages_per_hpage should be the same for all entries in 313075a61d0SMina Almasry * a resv_map. 314075a61d0SMina Almasry */ 315075a61d0SMina Almasry VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); 316075a61d0SMina Almasry } else { 317075a61d0SMina Almasry nrg->reservation_counter = NULL; 318075a61d0SMina Almasry nrg->css = NULL; 319075a61d0SMina Almasry } 320075a61d0SMina Almasry #endif 321075a61d0SMina Almasry } 322075a61d0SMina Almasry 323d85aecf2SMiaohe Lin static void put_uncharge_info(struct file_region *rg) 324d85aecf2SMiaohe Lin { 325d85aecf2SMiaohe Lin #ifdef CONFIG_CGROUP_HUGETLB 326d85aecf2SMiaohe Lin if (rg->css) 327d85aecf2SMiaohe Lin css_put(rg->css); 328d85aecf2SMiaohe Lin #endif 329d85aecf2SMiaohe Lin } 330d85aecf2SMiaohe Lin 331a9b3f867SMina Almasry static bool has_same_uncharge_info(struct file_region *rg, 332a9b3f867SMina Almasry struct file_region *org) 333a9b3f867SMina Almasry { 334a9b3f867SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 335*0739eb43SBaolin Wang return rg->reservation_counter == org->reservation_counter && 336a9b3f867SMina Almasry rg->css == org->css; 337a9b3f867SMina Almasry 338a9b3f867SMina Almasry #else 339a9b3f867SMina Almasry return true; 340a9b3f867SMina Almasry #endif 341a9b3f867SMina Almasry } 342a9b3f867SMina Almasry 343a9b3f867SMina Almasry static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) 344a9b3f867SMina Almasry { 345a9b3f867SMina Almasry struct file_region *nrg = NULL, *prg = NULL; 346a9b3f867SMina Almasry 347a9b3f867SMina Almasry prg = list_prev_entry(rg, link); 348a9b3f867SMina Almasry if (&prg->link != &resv->regions && prg->to == rg->from && 349a9b3f867SMina Almasry has_same_uncharge_info(prg, rg)) { 350a9b3f867SMina Almasry prg->to = rg->to; 351a9b3f867SMina Almasry 352a9b3f867SMina Almasry list_del(&rg->link); 353d85aecf2SMiaohe Lin put_uncharge_info(rg); 354a9b3f867SMina Almasry kfree(rg); 355a9b3f867SMina Almasry 3567db5e7b6SWei Yang rg = prg; 357a9b3f867SMina Almasry } 358a9b3f867SMina Almasry 359a9b3f867SMina Almasry nrg = list_next_entry(rg, link); 360a9b3f867SMina Almasry if (&nrg->link != &resv->regions && nrg->from == rg->to && 361a9b3f867SMina Almasry has_same_uncharge_info(nrg, rg)) { 362a9b3f867SMina Almasry nrg->from = rg->from; 363a9b3f867SMina Almasry 364a9b3f867SMina Almasry list_del(&rg->link); 365d85aecf2SMiaohe Lin put_uncharge_info(rg); 366a9b3f867SMina Almasry kfree(rg); 367a9b3f867SMina Almasry } 368a9b3f867SMina Almasry } 369a9b3f867SMina Almasry 3702103cf9cSPeter Xu static inline long 3712103cf9cSPeter Xu hugetlb_resv_map_add(struct resv_map *map, struct file_region *rg, long from, 3722103cf9cSPeter Xu long to, struct hstate *h, struct hugetlb_cgroup *cg, 3732103cf9cSPeter Xu long *regions_needed) 3742103cf9cSPeter Xu { 3752103cf9cSPeter Xu struct file_region *nrg; 3762103cf9cSPeter Xu 3772103cf9cSPeter Xu if (!regions_needed) { 3782103cf9cSPeter Xu nrg = get_file_region_entry_from_cache(map, from, to); 3792103cf9cSPeter Xu record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg); 3802103cf9cSPeter Xu list_add(&nrg->link, rg->link.prev); 3812103cf9cSPeter Xu coalesce_file_region(map, nrg); 3822103cf9cSPeter Xu } else 3832103cf9cSPeter Xu *regions_needed += 1; 3842103cf9cSPeter Xu 3852103cf9cSPeter Xu return to - from; 3862103cf9cSPeter Xu } 3872103cf9cSPeter Xu 388972a3da3SWei Yang /* 389972a3da3SWei Yang * Must be called with resv->lock held. 390972a3da3SWei Yang * 391972a3da3SWei Yang * Calling this with regions_needed != NULL will count the number of pages 392972a3da3SWei Yang * to be added but will not modify the linked list. And regions_needed will 393972a3da3SWei Yang * indicate the number of file_regions needed in the cache to carry out to add 394972a3da3SWei Yang * the regions for this range. 395d75c6af9SMina Almasry */ 396d75c6af9SMina Almasry static long add_reservation_in_range(struct resv_map *resv, long f, long t, 397075a61d0SMina Almasry struct hugetlb_cgroup *h_cg, 398972a3da3SWei Yang struct hstate *h, long *regions_needed) 399d75c6af9SMina Almasry { 4000db9d74eSMina Almasry long add = 0; 401d75c6af9SMina Almasry struct list_head *head = &resv->regions; 4020db9d74eSMina Almasry long last_accounted_offset = f; 4032103cf9cSPeter Xu struct file_region *rg = NULL, *trg = NULL; 404d75c6af9SMina Almasry 4050db9d74eSMina Almasry if (regions_needed) 4060db9d74eSMina Almasry *regions_needed = 0; 407d75c6af9SMina Almasry 4080db9d74eSMina Almasry /* In this loop, we essentially handle an entry for the range 4090db9d74eSMina Almasry * [last_accounted_offset, rg->from), at every iteration, with some 4100db9d74eSMina Almasry * bounds checking. 4110db9d74eSMina Almasry */ 4120db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, head, link) { 4130db9d74eSMina Almasry /* Skip irrelevant regions that start before our range. */ 4140db9d74eSMina Almasry if (rg->from < f) { 4150db9d74eSMina Almasry /* If this region ends after the last accounted offset, 4160db9d74eSMina Almasry * then we need to update last_accounted_offset. 4170db9d74eSMina Almasry */ 4180db9d74eSMina Almasry if (rg->to > last_accounted_offset) 4190db9d74eSMina Almasry last_accounted_offset = rg->to; 4200db9d74eSMina Almasry continue; 4210db9d74eSMina Almasry } 422d75c6af9SMina Almasry 4230db9d74eSMina Almasry /* When we find a region that starts beyond our range, we've 4240db9d74eSMina Almasry * finished. 4250db9d74eSMina Almasry */ 426ca7e0457SPeter Xu if (rg->from >= t) 427d75c6af9SMina Almasry break; 428d75c6af9SMina Almasry 4290db9d74eSMina Almasry /* Add an entry for last_accounted_offset -> rg->from, and 4300db9d74eSMina Almasry * update last_accounted_offset. 431d75c6af9SMina Almasry */ 4322103cf9cSPeter Xu if (rg->from > last_accounted_offset) 4332103cf9cSPeter Xu add += hugetlb_resv_map_add(resv, rg, 4342103cf9cSPeter Xu last_accounted_offset, 4352103cf9cSPeter Xu rg->from, h, h_cg, 4362103cf9cSPeter Xu regions_needed); 437d75c6af9SMina Almasry 4380db9d74eSMina Almasry last_accounted_offset = rg->to; 4390db9d74eSMina Almasry } 4400db9d74eSMina Almasry 4410db9d74eSMina Almasry /* Handle the case where our range extends beyond 4420db9d74eSMina Almasry * last_accounted_offset. 4430db9d74eSMina Almasry */ 4442103cf9cSPeter Xu if (last_accounted_offset < t) 4452103cf9cSPeter Xu add += hugetlb_resv_map_add(resv, rg, last_accounted_offset, 4462103cf9cSPeter Xu t, h, h_cg, regions_needed); 4470db9d74eSMina Almasry 4480db9d74eSMina Almasry VM_BUG_ON(add < 0); 4490db9d74eSMina Almasry return add; 4500db9d74eSMina Almasry } 4510db9d74eSMina Almasry 4520db9d74eSMina Almasry /* Must be called with resv->lock acquired. Will drop lock to allocate entries. 4530db9d74eSMina Almasry */ 4540db9d74eSMina Almasry static int allocate_file_region_entries(struct resv_map *resv, 4550db9d74eSMina Almasry int regions_needed) 4560db9d74eSMina Almasry __must_hold(&resv->lock) 4570db9d74eSMina Almasry { 4580db9d74eSMina Almasry struct list_head allocated_regions; 4590db9d74eSMina Almasry int to_allocate = 0, i = 0; 4600db9d74eSMina Almasry struct file_region *trg = NULL, *rg = NULL; 4610db9d74eSMina Almasry 4620db9d74eSMina Almasry VM_BUG_ON(regions_needed < 0); 4630db9d74eSMina Almasry 4640db9d74eSMina Almasry INIT_LIST_HEAD(&allocated_regions); 4650db9d74eSMina Almasry 4660db9d74eSMina Almasry /* 4670db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 4680db9d74eSMina Almasry * the number of in progress add operations plus regions_needed. 4690db9d74eSMina Almasry * 4700db9d74eSMina Almasry * This is a while loop because when we drop the lock, some other call 4710db9d74eSMina Almasry * to region_add or region_del may have consumed some region_entries, 4720db9d74eSMina Almasry * so we keep looping here until we finally have enough entries for 4730db9d74eSMina Almasry * (adds_in_progress + regions_needed). 4740db9d74eSMina Almasry */ 4750db9d74eSMina Almasry while (resv->region_cache_count < 4760db9d74eSMina Almasry (resv->adds_in_progress + regions_needed)) { 4770db9d74eSMina Almasry to_allocate = resv->adds_in_progress + regions_needed - 4780db9d74eSMina Almasry resv->region_cache_count; 4790db9d74eSMina Almasry 4800db9d74eSMina Almasry /* At this point, we should have enough entries in the cache 481f0953a1bSIngo Molnar * for all the existing adds_in_progress. We should only be 4820db9d74eSMina Almasry * needing to allocate for regions_needed. 4830db9d74eSMina Almasry */ 4840db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); 4850db9d74eSMina Almasry 4860db9d74eSMina Almasry spin_unlock(&resv->lock); 4870db9d74eSMina Almasry for (i = 0; i < to_allocate; i++) { 4880db9d74eSMina Almasry trg = kmalloc(sizeof(*trg), GFP_KERNEL); 4890db9d74eSMina Almasry if (!trg) 4900db9d74eSMina Almasry goto out_of_memory; 4910db9d74eSMina Almasry list_add(&trg->link, &allocated_regions); 4920db9d74eSMina Almasry } 4930db9d74eSMina Almasry 4940db9d74eSMina Almasry spin_lock(&resv->lock); 4950db9d74eSMina Almasry 496d3ec7b6eSWei Yang list_splice(&allocated_regions, &resv->region_cache); 497d3ec7b6eSWei Yang resv->region_cache_count += to_allocate; 4980db9d74eSMina Almasry } 4990db9d74eSMina Almasry 5000db9d74eSMina Almasry return 0; 5010db9d74eSMina Almasry 5020db9d74eSMina Almasry out_of_memory: 5030db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, &allocated_regions, link) { 504d75c6af9SMina Almasry list_del(&rg->link); 505d75c6af9SMina Almasry kfree(rg); 506d75c6af9SMina Almasry } 5070db9d74eSMina Almasry return -ENOMEM; 508d75c6af9SMina Almasry } 509d75c6af9SMina Almasry 5101dd308a7SMike Kravetz /* 5111dd308a7SMike Kravetz * Add the huge page range represented by [f, t) to the reserve 5120db9d74eSMina Almasry * map. Regions will be taken from the cache to fill in this range. 5130db9d74eSMina Almasry * Sufficient regions should exist in the cache due to the previous 5140db9d74eSMina Almasry * call to region_chg with the same range, but in some cases the cache will not 5150db9d74eSMina Almasry * have sufficient entries due to races with other code doing region_add or 5160db9d74eSMina Almasry * region_del. The extra needed entries will be allocated. 517cf3ad20bSMike Kravetz * 5180db9d74eSMina Almasry * regions_needed is the out value provided by a previous call to region_chg. 5190db9d74eSMina Almasry * 5200db9d74eSMina Almasry * Return the number of new huge pages added to the map. This number is greater 5210db9d74eSMina Almasry * than or equal to zero. If file_region entries needed to be allocated for 5227c8de358SEthon Paul * this operation and we were not able to allocate, it returns -ENOMEM. 5230db9d74eSMina Almasry * region_add of regions of length 1 never allocate file_regions and cannot 5240db9d74eSMina Almasry * fail; region_chg will always allocate at least 1 entry and a region_add for 5250db9d74eSMina Almasry * 1 page will only require at most 1 entry. 5261dd308a7SMike Kravetz */ 5270db9d74eSMina Almasry static long region_add(struct resv_map *resv, long f, long t, 528075a61d0SMina Almasry long in_regions_needed, struct hstate *h, 529075a61d0SMina Almasry struct hugetlb_cgroup *h_cg) 53096822904SAndy Whitcroft { 5310db9d74eSMina Almasry long add = 0, actual_regions_needed = 0; 53296822904SAndy Whitcroft 5337b24d861SDavidlohr Bueso spin_lock(&resv->lock); 5340db9d74eSMina Almasry retry: 5350db9d74eSMina Almasry 5360db9d74eSMina Almasry /* Count how many regions are actually needed to execute this add. */ 537972a3da3SWei Yang add_reservation_in_range(resv, f, t, NULL, NULL, 538972a3da3SWei Yang &actual_regions_needed); 53996822904SAndy Whitcroft 5405e911373SMike Kravetz /* 5410db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 5420db9d74eSMina Almasry * this add operation. Note that actual_regions_needed may be greater 5430db9d74eSMina Almasry * than in_regions_needed, as the resv_map may have been modified since 5440db9d74eSMina Almasry * the region_chg call. In this case, we need to make sure that we 5450db9d74eSMina Almasry * allocate extra entries, such that we have enough for all the 5460db9d74eSMina Almasry * existing adds_in_progress, plus the excess needed for this 5470db9d74eSMina Almasry * operation. 5485e911373SMike Kravetz */ 5490db9d74eSMina Almasry if (actual_regions_needed > in_regions_needed && 5500db9d74eSMina Almasry resv->region_cache_count < 5510db9d74eSMina Almasry resv->adds_in_progress + 5520db9d74eSMina Almasry (actual_regions_needed - in_regions_needed)) { 5530db9d74eSMina Almasry /* region_add operation of range 1 should never need to 5540db9d74eSMina Almasry * allocate file_region entries. 5550db9d74eSMina Almasry */ 5560db9d74eSMina Almasry VM_BUG_ON(t - f <= 1); 5575e911373SMike Kravetz 5580db9d74eSMina Almasry if (allocate_file_region_entries( 5590db9d74eSMina Almasry resv, actual_regions_needed - in_regions_needed)) { 5600db9d74eSMina Almasry return -ENOMEM; 5615e911373SMike Kravetz } 5625e911373SMike Kravetz 5630db9d74eSMina Almasry goto retry; 5640db9d74eSMina Almasry } 565cf3ad20bSMike Kravetz 566972a3da3SWei Yang add = add_reservation_in_range(resv, f, t, h_cg, h, NULL); 5670db9d74eSMina Almasry 5680db9d74eSMina Almasry resv->adds_in_progress -= in_regions_needed; 5690db9d74eSMina Almasry 5707b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 571cf3ad20bSMike Kravetz return add; 57296822904SAndy Whitcroft } 57396822904SAndy Whitcroft 5741dd308a7SMike Kravetz /* 5751dd308a7SMike Kravetz * Examine the existing reserve map and determine how many 5761dd308a7SMike Kravetz * huge pages in the specified range [f, t) are NOT currently 5771dd308a7SMike Kravetz * represented. This routine is called before a subsequent 5781dd308a7SMike Kravetz * call to region_add that will actually modify the reserve 5791dd308a7SMike Kravetz * map to add the specified range [f, t). region_chg does 5801dd308a7SMike Kravetz * not change the number of huge pages represented by the 5810db9d74eSMina Almasry * map. A number of new file_region structures is added to the cache as a 5820db9d74eSMina Almasry * placeholder, for the subsequent region_add call to use. At least 1 5830db9d74eSMina Almasry * file_region structure is added. 5840db9d74eSMina Almasry * 5850db9d74eSMina Almasry * out_regions_needed is the number of regions added to the 5860db9d74eSMina Almasry * resv->adds_in_progress. This value needs to be provided to a follow up call 5870db9d74eSMina Almasry * to region_add or region_abort for proper accounting. 5885e911373SMike Kravetz * 5895e911373SMike Kravetz * Returns the number of huge pages that need to be added to the existing 5905e911373SMike Kravetz * reservation map for the range [f, t). This number is greater or equal to 5915e911373SMike Kravetz * zero. -ENOMEM is returned if a new file_region structure or cache entry 5925e911373SMike Kravetz * is needed and can not be allocated. 5931dd308a7SMike Kravetz */ 5940db9d74eSMina Almasry static long region_chg(struct resv_map *resv, long f, long t, 5950db9d74eSMina Almasry long *out_regions_needed) 59696822904SAndy Whitcroft { 59796822904SAndy Whitcroft long chg = 0; 59896822904SAndy Whitcroft 5997b24d861SDavidlohr Bueso spin_lock(&resv->lock); 6005e911373SMike Kravetz 601972a3da3SWei Yang /* Count how many hugepages in this range are NOT represented. */ 602075a61d0SMina Almasry chg = add_reservation_in_range(resv, f, t, NULL, NULL, 603972a3da3SWei Yang out_regions_needed); 6045e911373SMike Kravetz 6050db9d74eSMina Almasry if (*out_regions_needed == 0) 6060db9d74eSMina Almasry *out_regions_needed = 1; 6075e911373SMike Kravetz 6080db9d74eSMina Almasry if (allocate_file_region_entries(resv, *out_regions_needed)) 6095e911373SMike Kravetz return -ENOMEM; 6105e911373SMike Kravetz 6110db9d74eSMina Almasry resv->adds_in_progress += *out_regions_needed; 61296822904SAndy Whitcroft 6137b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 61496822904SAndy Whitcroft return chg; 61596822904SAndy Whitcroft } 61696822904SAndy Whitcroft 6171dd308a7SMike Kravetz /* 6185e911373SMike Kravetz * Abort the in progress add operation. The adds_in_progress field 6195e911373SMike Kravetz * of the resv_map keeps track of the operations in progress between 6205e911373SMike Kravetz * calls to region_chg and region_add. Operations are sometimes 6215e911373SMike Kravetz * aborted after the call to region_chg. In such cases, region_abort 6220db9d74eSMina Almasry * is called to decrement the adds_in_progress counter. regions_needed 6230db9d74eSMina Almasry * is the value returned by the region_chg call, it is used to decrement 6240db9d74eSMina Almasry * the adds_in_progress counter. 6255e911373SMike Kravetz * 6265e911373SMike Kravetz * NOTE: The range arguments [f, t) are not needed or used in this 6275e911373SMike Kravetz * routine. They are kept to make reading the calling code easier as 6285e911373SMike Kravetz * arguments will match the associated region_chg call. 6295e911373SMike Kravetz */ 6300db9d74eSMina Almasry static void region_abort(struct resv_map *resv, long f, long t, 6310db9d74eSMina Almasry long regions_needed) 6325e911373SMike Kravetz { 6335e911373SMike Kravetz spin_lock(&resv->lock); 6345e911373SMike Kravetz VM_BUG_ON(!resv->region_cache_count); 6350db9d74eSMina Almasry resv->adds_in_progress -= regions_needed; 6365e911373SMike Kravetz spin_unlock(&resv->lock); 6375e911373SMike Kravetz } 6385e911373SMike Kravetz 6395e911373SMike Kravetz /* 640feba16e2SMike Kravetz * Delete the specified range [f, t) from the reserve map. If the 641feba16e2SMike Kravetz * t parameter is LONG_MAX, this indicates that ALL regions after f 642feba16e2SMike Kravetz * should be deleted. Locate the regions which intersect [f, t) 643feba16e2SMike Kravetz * and either trim, delete or split the existing regions. 644feba16e2SMike Kravetz * 645feba16e2SMike Kravetz * Returns the number of huge pages deleted from the reserve map. 646feba16e2SMike Kravetz * In the normal case, the return value is zero or more. In the 647feba16e2SMike Kravetz * case where a region must be split, a new region descriptor must 648feba16e2SMike Kravetz * be allocated. If the allocation fails, -ENOMEM will be returned. 649feba16e2SMike Kravetz * NOTE: If the parameter t == LONG_MAX, then we will never split 650feba16e2SMike Kravetz * a region and possibly return -ENOMEM. Callers specifying 651feba16e2SMike Kravetz * t == LONG_MAX do not need to check for -ENOMEM error. 6521dd308a7SMike Kravetz */ 653feba16e2SMike Kravetz static long region_del(struct resv_map *resv, long f, long t) 65496822904SAndy Whitcroft { 6551406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 65696822904SAndy Whitcroft struct file_region *rg, *trg; 657feba16e2SMike Kravetz struct file_region *nrg = NULL; 658feba16e2SMike Kravetz long del = 0; 65996822904SAndy Whitcroft 660feba16e2SMike Kravetz retry: 6617b24d861SDavidlohr Bueso spin_lock(&resv->lock); 662feba16e2SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 663dbe409e4SMike Kravetz /* 664dbe409e4SMike Kravetz * Skip regions before the range to be deleted. file_region 665dbe409e4SMike Kravetz * ranges are normally of the form [from, to). However, there 666dbe409e4SMike Kravetz * may be a "placeholder" entry in the map which is of the form 667dbe409e4SMike Kravetz * (from, to) with from == to. Check for placeholder entries 668dbe409e4SMike Kravetz * at the beginning of the range to be deleted. 669dbe409e4SMike Kravetz */ 670dbe409e4SMike Kravetz if (rg->to <= f && (rg->to != rg->from || rg->to != f)) 671feba16e2SMike Kravetz continue; 672dbe409e4SMike Kravetz 673feba16e2SMike Kravetz if (rg->from >= t) 67496822904SAndy Whitcroft break; 67596822904SAndy Whitcroft 676feba16e2SMike Kravetz if (f > rg->from && t < rg->to) { /* Must split region */ 677feba16e2SMike Kravetz /* 678feba16e2SMike Kravetz * Check for an entry in the cache before dropping 679feba16e2SMike Kravetz * lock and attempting allocation. 680feba16e2SMike Kravetz */ 681feba16e2SMike Kravetz if (!nrg && 682feba16e2SMike Kravetz resv->region_cache_count > resv->adds_in_progress) { 683feba16e2SMike Kravetz nrg = list_first_entry(&resv->region_cache, 684feba16e2SMike Kravetz struct file_region, 685feba16e2SMike Kravetz link); 686feba16e2SMike Kravetz list_del(&nrg->link); 687feba16e2SMike Kravetz resv->region_cache_count--; 68896822904SAndy Whitcroft } 68996822904SAndy Whitcroft 690feba16e2SMike Kravetz if (!nrg) { 691feba16e2SMike Kravetz spin_unlock(&resv->lock); 692feba16e2SMike Kravetz nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 693feba16e2SMike Kravetz if (!nrg) 694feba16e2SMike Kravetz return -ENOMEM; 695feba16e2SMike Kravetz goto retry; 696feba16e2SMike Kravetz } 697feba16e2SMike Kravetz 698feba16e2SMike Kravetz del += t - f; 69979aa925bSMike Kravetz hugetlb_cgroup_uncharge_file_region( 700d85aecf2SMiaohe Lin resv, rg, t - f, false); 701feba16e2SMike Kravetz 702feba16e2SMike Kravetz /* New entry for end of split region */ 703feba16e2SMike Kravetz nrg->from = t; 704feba16e2SMike Kravetz nrg->to = rg->to; 705075a61d0SMina Almasry 706075a61d0SMina Almasry copy_hugetlb_cgroup_uncharge_info(nrg, rg); 707075a61d0SMina Almasry 708feba16e2SMike Kravetz INIT_LIST_HEAD(&nrg->link); 709feba16e2SMike Kravetz 710feba16e2SMike Kravetz /* Original entry is trimmed */ 711feba16e2SMike Kravetz rg->to = f; 712feba16e2SMike Kravetz 713feba16e2SMike Kravetz list_add(&nrg->link, &rg->link); 714feba16e2SMike Kravetz nrg = NULL; 71596822904SAndy Whitcroft break; 716feba16e2SMike Kravetz } 717feba16e2SMike Kravetz 718feba16e2SMike Kravetz if (f <= rg->from && t >= rg->to) { /* Remove entire region */ 719feba16e2SMike Kravetz del += rg->to - rg->from; 720075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 721d85aecf2SMiaohe Lin rg->to - rg->from, true); 72296822904SAndy Whitcroft list_del(&rg->link); 72396822904SAndy Whitcroft kfree(rg); 724feba16e2SMike Kravetz continue; 72596822904SAndy Whitcroft } 7267b24d861SDavidlohr Bueso 727feba16e2SMike Kravetz if (f <= rg->from) { /* Trim beginning of region */ 728075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 729d85aecf2SMiaohe Lin t - rg->from, false); 730075a61d0SMina Almasry 73179aa925bSMike Kravetz del += t - rg->from; 73279aa925bSMike Kravetz rg->from = t; 73379aa925bSMike Kravetz } else { /* Trim end of region */ 734075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 735d85aecf2SMiaohe Lin rg->to - f, false); 73679aa925bSMike Kravetz 73779aa925bSMike Kravetz del += rg->to - f; 73879aa925bSMike Kravetz rg->to = f; 739feba16e2SMike Kravetz } 740feba16e2SMike Kravetz } 741feba16e2SMike Kravetz 7427b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 743feba16e2SMike Kravetz kfree(nrg); 744feba16e2SMike Kravetz return del; 74596822904SAndy Whitcroft } 74696822904SAndy Whitcroft 7471dd308a7SMike Kravetz /* 748b5cec28dSMike Kravetz * A rare out of memory error was encountered which prevented removal of 749b5cec28dSMike Kravetz * the reserve map region for a page. The huge page itself was free'ed 750b5cec28dSMike Kravetz * and removed from the page cache. This routine will adjust the subpool 751b5cec28dSMike Kravetz * usage count, and the global reserve count if needed. By incrementing 752b5cec28dSMike Kravetz * these counts, the reserve map entry which could not be deleted will 753b5cec28dSMike Kravetz * appear as a "reserved" entry instead of simply dangling with incorrect 754b5cec28dSMike Kravetz * counts. 755b5cec28dSMike Kravetz */ 75672e2936cSzhong jiang void hugetlb_fix_reserve_counts(struct inode *inode) 757b5cec28dSMike Kravetz { 758b5cec28dSMike Kravetz struct hugepage_subpool *spool = subpool_inode(inode); 759b5cec28dSMike Kravetz long rsv_adjust; 760da56388cSMiaohe Lin bool reserved = false; 761b5cec28dSMike Kravetz 762b5cec28dSMike Kravetz rsv_adjust = hugepage_subpool_get_pages(spool, 1); 763da56388cSMiaohe Lin if (rsv_adjust > 0) { 764b5cec28dSMike Kravetz struct hstate *h = hstate_inode(inode); 765b5cec28dSMike Kravetz 766da56388cSMiaohe Lin if (!hugetlb_acct_memory(h, 1)) 767da56388cSMiaohe Lin reserved = true; 768da56388cSMiaohe Lin } else if (!rsv_adjust) { 769da56388cSMiaohe Lin reserved = true; 770b5cec28dSMike Kravetz } 771da56388cSMiaohe Lin 772da56388cSMiaohe Lin if (!reserved) 773da56388cSMiaohe Lin pr_warn("hugetlb: Huge Page Reserved count may go negative.\n"); 774b5cec28dSMike Kravetz } 775b5cec28dSMike Kravetz 776b5cec28dSMike Kravetz /* 7771dd308a7SMike Kravetz * Count and return the number of huge pages in the reserve map 7781dd308a7SMike Kravetz * that intersect with the range [f, t). 7791dd308a7SMike Kravetz */ 7801406ec9bSJoonsoo Kim static long region_count(struct resv_map *resv, long f, long t) 78184afd99bSAndy Whitcroft { 7821406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 78384afd99bSAndy Whitcroft struct file_region *rg; 78484afd99bSAndy Whitcroft long chg = 0; 78584afd99bSAndy Whitcroft 7867b24d861SDavidlohr Bueso spin_lock(&resv->lock); 78784afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 78884afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 789f2135a4aSWang Sheng-Hui long seg_from; 790f2135a4aSWang Sheng-Hui long seg_to; 79184afd99bSAndy Whitcroft 79284afd99bSAndy Whitcroft if (rg->to <= f) 79384afd99bSAndy Whitcroft continue; 79484afd99bSAndy Whitcroft if (rg->from >= t) 79584afd99bSAndy Whitcroft break; 79684afd99bSAndy Whitcroft 79784afd99bSAndy Whitcroft seg_from = max(rg->from, f); 79884afd99bSAndy Whitcroft seg_to = min(rg->to, t); 79984afd99bSAndy Whitcroft 80084afd99bSAndy Whitcroft chg += seg_to - seg_from; 80184afd99bSAndy Whitcroft } 8027b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 80384afd99bSAndy Whitcroft 80484afd99bSAndy Whitcroft return chg; 80584afd99bSAndy Whitcroft } 80684afd99bSAndy Whitcroft 80796822904SAndy Whitcroft /* 808e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 809e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 810e7c4b0bfSAndy Whitcroft */ 811a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 812a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 813e7c4b0bfSAndy Whitcroft { 814a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 815a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 816e7c4b0bfSAndy Whitcroft } 817e7c4b0bfSAndy Whitcroft 8180fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 8190fe6e20bSNaoya Horiguchi unsigned long address) 8200fe6e20bSNaoya Horiguchi { 8210fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 8220fe6e20bSNaoya Horiguchi } 823dee41079SDan Williams EXPORT_SYMBOL_GPL(linear_hugepage_index); 8240fe6e20bSNaoya Horiguchi 82584afd99bSAndy Whitcroft /* 82608fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 82708fba699SMel Gorman * cases this will be same size as used by the page table entries. 82808fba699SMel Gorman */ 82908fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 83008fba699SMel Gorman { 83105ea8860SDan Williams if (vma->vm_ops && vma->vm_ops->pagesize) 83205ea8860SDan Williams return vma->vm_ops->pagesize(vma); 83308fba699SMel Gorman return PAGE_SIZE; 83408fba699SMel Gorman } 835f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 83608fba699SMel Gorman 83708fba699SMel Gorman /* 8383340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 8393340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 84009135cc5SDan Williams * architectures where it differs, an architecture-specific 'strong' 84109135cc5SDan Williams * version of this symbol is required. 8423340289dSMel Gorman */ 84309135cc5SDan Williams __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 8443340289dSMel Gorman { 8453340289dSMel Gorman return vma_kernel_pagesize(vma); 8463340289dSMel Gorman } 8473340289dSMel Gorman 8483340289dSMel Gorman /* 84984afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 85084afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 85184afd99bSAndy Whitcroft * alignment. 85284afd99bSAndy Whitcroft */ 85384afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 85484afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 85504f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 85684afd99bSAndy Whitcroft 857a1e78772SMel Gorman /* 858a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 859a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 860a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 861a1e78772SMel Gorman * 862a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 863a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 864a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 865a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 86684afd99bSAndy Whitcroft * 86784afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 86884afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 86984afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 87084afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 87184afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 87284afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 87384afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 87484afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 875a1e78772SMel Gorman */ 876e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 877e7c4b0bfSAndy Whitcroft { 878e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 879e7c4b0bfSAndy Whitcroft } 880e7c4b0bfSAndy Whitcroft 881e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 882e7c4b0bfSAndy Whitcroft unsigned long value) 883e7c4b0bfSAndy Whitcroft { 884e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 885e7c4b0bfSAndy Whitcroft } 886e7c4b0bfSAndy Whitcroft 887e9fe92aeSMina Almasry static void 888e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, 889e9fe92aeSMina Almasry struct hugetlb_cgroup *h_cg, 890e9fe92aeSMina Almasry struct hstate *h) 891e9fe92aeSMina Almasry { 892e9fe92aeSMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 893e9fe92aeSMina Almasry if (!h_cg || !h) { 894e9fe92aeSMina Almasry resv_map->reservation_counter = NULL; 895e9fe92aeSMina Almasry resv_map->pages_per_hpage = 0; 896e9fe92aeSMina Almasry resv_map->css = NULL; 897e9fe92aeSMina Almasry } else { 898e9fe92aeSMina Almasry resv_map->reservation_counter = 899e9fe92aeSMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 900e9fe92aeSMina Almasry resv_map->pages_per_hpage = pages_per_huge_page(h); 901e9fe92aeSMina Almasry resv_map->css = &h_cg->css; 902e9fe92aeSMina Almasry } 903e9fe92aeSMina Almasry #endif 904e9fe92aeSMina Almasry } 905e9fe92aeSMina Almasry 9069119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 90784afd99bSAndy Whitcroft { 90884afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 9095e911373SMike Kravetz struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 9105e911373SMike Kravetz 9115e911373SMike Kravetz if (!resv_map || !rg) { 9125e911373SMike Kravetz kfree(resv_map); 9135e911373SMike Kravetz kfree(rg); 91484afd99bSAndy Whitcroft return NULL; 9155e911373SMike Kravetz } 91684afd99bSAndy Whitcroft 91784afd99bSAndy Whitcroft kref_init(&resv_map->refs); 9187b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 91984afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 92084afd99bSAndy Whitcroft 9215e911373SMike Kravetz resv_map->adds_in_progress = 0; 922e9fe92aeSMina Almasry /* 923e9fe92aeSMina Almasry * Initialize these to 0. On shared mappings, 0's here indicate these 924e9fe92aeSMina Almasry * fields don't do cgroup accounting. On private mappings, these will be 925e9fe92aeSMina Almasry * re-initialized to the proper values, to indicate that hugetlb cgroup 926e9fe92aeSMina Almasry * reservations are to be un-charged from here. 927e9fe92aeSMina Almasry */ 928e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); 9295e911373SMike Kravetz 9305e911373SMike Kravetz INIT_LIST_HEAD(&resv_map->region_cache); 9315e911373SMike Kravetz list_add(&rg->link, &resv_map->region_cache); 9325e911373SMike Kravetz resv_map->region_cache_count = 1; 9335e911373SMike Kravetz 93484afd99bSAndy Whitcroft return resv_map; 93584afd99bSAndy Whitcroft } 93684afd99bSAndy Whitcroft 9379119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 93884afd99bSAndy Whitcroft { 93984afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 9405e911373SMike Kravetz struct list_head *head = &resv_map->region_cache; 9415e911373SMike Kravetz struct file_region *rg, *trg; 94284afd99bSAndy Whitcroft 94384afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 944feba16e2SMike Kravetz region_del(resv_map, 0, LONG_MAX); 9455e911373SMike Kravetz 9465e911373SMike Kravetz /* ... and any entries left in the cache */ 9475e911373SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 9485e911373SMike Kravetz list_del(&rg->link); 9495e911373SMike Kravetz kfree(rg); 9505e911373SMike Kravetz } 9515e911373SMike Kravetz 9525e911373SMike Kravetz VM_BUG_ON(resv_map->adds_in_progress); 9535e911373SMike Kravetz 95484afd99bSAndy Whitcroft kfree(resv_map); 95584afd99bSAndy Whitcroft } 95684afd99bSAndy Whitcroft 9574e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 9584e35f483SJoonsoo Kim { 959f27a5136SMike Kravetz /* 960f27a5136SMike Kravetz * At inode evict time, i_mapping may not point to the original 961f27a5136SMike Kravetz * address space within the inode. This original address space 962f27a5136SMike Kravetz * contains the pointer to the resv_map. So, always use the 963f27a5136SMike Kravetz * address space embedded within the inode. 964f27a5136SMike Kravetz * The VERY common case is inode->mapping == &inode->i_data but, 965f27a5136SMike Kravetz * this may not be true for device special inodes. 966f27a5136SMike Kravetz */ 967f27a5136SMike Kravetz return (struct resv_map *)(&inode->i_data)->private_data; 9684e35f483SJoonsoo Kim } 9694e35f483SJoonsoo Kim 97084afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 971a1e78772SMel Gorman { 97281d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 9734e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 9744e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 9754e35f483SJoonsoo Kim struct inode *inode = mapping->host; 9764e35f483SJoonsoo Kim 9774e35f483SJoonsoo Kim return inode_resv_map(inode); 9784e35f483SJoonsoo Kim 9794e35f483SJoonsoo Kim } else { 98084afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 98184afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 9824e35f483SJoonsoo Kim } 983a1e78772SMel Gorman } 984a1e78772SMel Gorman 98584afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 986a1e78772SMel Gorman { 98781d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 98881d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 989a1e78772SMel Gorman 99084afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 99184afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 99204f2cbe3SMel Gorman } 99304f2cbe3SMel Gorman 99404f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 99504f2cbe3SMel Gorman { 99681d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 99781d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 998e7c4b0bfSAndy Whitcroft 999e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 100004f2cbe3SMel Gorman } 100104f2cbe3SMel Gorman 100204f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 100304f2cbe3SMel Gorman { 100481d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 1005e7c4b0bfSAndy Whitcroft 1006e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 1007a1e78772SMel Gorman } 1008a1e78772SMel Gorman 100904f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 1010a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 1011a1e78772SMel Gorman { 101281d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 1013f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 1014a1e78772SMel Gorman vma->vm_private_data = (void *)0; 1015a1e78772SMel Gorman } 1016a1e78772SMel Gorman 1017550a7d60SMina Almasry /* 1018550a7d60SMina Almasry * Reset and decrement one ref on hugepage private reservation. 1019550a7d60SMina Almasry * Called with mm->mmap_sem writer semaphore held. 1020550a7d60SMina Almasry * This function should be only used by move_vma() and operate on 1021550a7d60SMina Almasry * same sized vma. It should never come here with last ref on the 1022550a7d60SMina Almasry * reservation. 1023550a7d60SMina Almasry */ 1024550a7d60SMina Almasry void clear_vma_resv_huge_pages(struct vm_area_struct *vma) 1025550a7d60SMina Almasry { 1026550a7d60SMina Almasry /* 1027550a7d60SMina Almasry * Clear the old hugetlb private page reservation. 1028550a7d60SMina Almasry * It has already been transferred to new_vma. 1029550a7d60SMina Almasry * 1030550a7d60SMina Almasry * During a mremap() operation of a hugetlb vma we call move_vma() 1031550a7d60SMina Almasry * which copies vma into new_vma and unmaps vma. After the copy 1032550a7d60SMina Almasry * operation both new_vma and vma share a reference to the resv_map 1033550a7d60SMina Almasry * struct, and at that point vma is about to be unmapped. We don't 1034550a7d60SMina Almasry * want to return the reservation to the pool at unmap of vma because 1035550a7d60SMina Almasry * the reservation still lives on in new_vma, so simply decrement the 1036550a7d60SMina Almasry * ref here and remove the resv_map reference from this vma. 1037550a7d60SMina Almasry */ 1038550a7d60SMina Almasry struct resv_map *reservations = vma_resv_map(vma); 1039550a7d60SMina Almasry 1040550a7d60SMina Almasry if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 1041550a7d60SMina Almasry kref_put(&reservations->refs, resv_map_release); 1042550a7d60SMina Almasry 1043550a7d60SMina Almasry reset_vma_resv_huge_pages(vma); 1044550a7d60SMina Almasry } 1045550a7d60SMina Almasry 1046a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 1047559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 1048a1e78772SMel Gorman { 1049af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 1050af0ed73eSJoonsoo Kim /* 1051af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 1052af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 1053af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 1054af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 1055af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 1056af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 1057af0ed73eSJoonsoo Kim * properly, so add work-around here. 1058af0ed73eSJoonsoo Kim */ 1059af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 1060559ec2f8SNicholas Krause return true; 1061af0ed73eSJoonsoo Kim else 1062559ec2f8SNicholas Krause return false; 1063af0ed73eSJoonsoo Kim } 1064a63884e9SJoonsoo Kim 1065a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 10661fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 10671fb1b0e9SMike Kravetz /* 10681fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 10691fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 10701fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 10717c8de358SEthon Paul * fallocate. In this case, there really are no reserves to 10721fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 10731fb1b0e9SMike Kravetz */ 10741fb1b0e9SMike Kravetz if (chg) 10751fb1b0e9SMike Kravetz return false; 10761fb1b0e9SMike Kravetz else 1077559ec2f8SNicholas Krause return true; 10781fb1b0e9SMike Kravetz } 1079a63884e9SJoonsoo Kim 1080a63884e9SJoonsoo Kim /* 1081a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 1082a63884e9SJoonsoo Kim * private mappings. 1083a63884e9SJoonsoo Kim */ 108467961f9dSMike Kravetz if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 108567961f9dSMike Kravetz /* 108667961f9dSMike Kravetz * Like the shared case above, a hole punch or truncate 108767961f9dSMike Kravetz * could have been performed on the private mapping. 108867961f9dSMike Kravetz * Examine the value of chg to determine if reserves 108967961f9dSMike Kravetz * actually exist or were previously consumed. 109067961f9dSMike Kravetz * Very Subtle - The value of chg comes from a previous 109167961f9dSMike Kravetz * call to vma_needs_reserves(). The reserve map for 109267961f9dSMike Kravetz * private mappings has different (opposite) semantics 109367961f9dSMike Kravetz * than that of shared mappings. vma_needs_reserves() 109467961f9dSMike Kravetz * has already taken this difference in semantics into 109567961f9dSMike Kravetz * account. Therefore, the meaning of chg is the same 109667961f9dSMike Kravetz * as in the shared case above. Code could easily be 109767961f9dSMike Kravetz * combined, but keeping it separate draws attention to 109867961f9dSMike Kravetz * subtle differences. 109967961f9dSMike Kravetz */ 110067961f9dSMike Kravetz if (chg) 110167961f9dSMike Kravetz return false; 110267961f9dSMike Kravetz else 1103559ec2f8SNicholas Krause return true; 110467961f9dSMike Kravetz } 1105a63884e9SJoonsoo Kim 1106559ec2f8SNicholas Krause return false; 1107a1e78772SMel Gorman } 1108a1e78772SMel Gorman 1109a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 11101da177e4SLinus Torvalds { 11111da177e4SLinus Torvalds int nid = page_to_nid(page); 11129487ca60SMike Kravetz 11139487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1114b65a4edaSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1115b65a4edaSMike Kravetz 11160edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 1117a5516438SAndi Kleen h->free_huge_pages++; 1118a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 11196c037149SMike Kravetz SetHPageFreed(page); 11201da177e4SLinus Torvalds } 11211da177e4SLinus Torvalds 112294310cbcSAnshuman Khandual static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) 1123bf50bab2SNaoya Horiguchi { 1124bf50bab2SNaoya Horiguchi struct page *page; 11251a08ae36SPavel Tatashin bool pin = !!(current->flags & PF_MEMALLOC_PIN); 1126bf50bab2SNaoya Horiguchi 11279487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1128bbe88753SJoonsoo Kim list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { 11298e3560d9SPavel Tatashin if (pin && !is_pinnable_page(page)) 1130bbe88753SJoonsoo Kim continue; 1131bbe88753SJoonsoo Kim 11326664bfc8SWei Yang if (PageHWPoison(page)) 11336664bfc8SWei Yang continue; 1134bbe88753SJoonsoo Kim 11350edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 1136a9869b83SNaoya Horiguchi set_page_refcounted(page); 11376c037149SMike Kravetz ClearHPageFreed(page); 1138bf50bab2SNaoya Horiguchi h->free_huge_pages--; 1139bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 1140bf50bab2SNaoya Horiguchi return page; 1141bf50bab2SNaoya Horiguchi } 1142bf50bab2SNaoya Horiguchi 11436664bfc8SWei Yang return NULL; 11446664bfc8SWei Yang } 11456664bfc8SWei Yang 11463e59fcb0SMichal Hocko static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, 11473e59fcb0SMichal Hocko nodemask_t *nmask) 114894310cbcSAnshuman Khandual { 11493e59fcb0SMichal Hocko unsigned int cpuset_mems_cookie; 11503e59fcb0SMichal Hocko struct zonelist *zonelist; 11513e59fcb0SMichal Hocko struct zone *zone; 11523e59fcb0SMichal Hocko struct zoneref *z; 115398fa15f3SAnshuman Khandual int node = NUMA_NO_NODE; 11543e59fcb0SMichal Hocko 11553e59fcb0SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 11563e59fcb0SMichal Hocko 11573e59fcb0SMichal Hocko retry_cpuset: 11583e59fcb0SMichal Hocko cpuset_mems_cookie = read_mems_allowed_begin(); 11593e59fcb0SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 116094310cbcSAnshuman Khandual struct page *page; 116194310cbcSAnshuman Khandual 11623e59fcb0SMichal Hocko if (!cpuset_zone_allowed(zone, gfp_mask)) 11633e59fcb0SMichal Hocko continue; 11643e59fcb0SMichal Hocko /* 11653e59fcb0SMichal Hocko * no need to ask again on the same node. Pool is node rather than 11663e59fcb0SMichal Hocko * zone aware 11673e59fcb0SMichal Hocko */ 11683e59fcb0SMichal Hocko if (zone_to_nid(zone) == node) 11693e59fcb0SMichal Hocko continue; 11703e59fcb0SMichal Hocko node = zone_to_nid(zone); 117194310cbcSAnshuman Khandual 117294310cbcSAnshuman Khandual page = dequeue_huge_page_node_exact(h, node); 117394310cbcSAnshuman Khandual if (page) 117494310cbcSAnshuman Khandual return page; 117594310cbcSAnshuman Khandual } 11763e59fcb0SMichal Hocko if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 11773e59fcb0SMichal Hocko goto retry_cpuset; 11783e59fcb0SMichal Hocko 117994310cbcSAnshuman Khandual return NULL; 118094310cbcSAnshuman Khandual } 118194310cbcSAnshuman Khandual 1182a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 1183a5516438SAndi Kleen struct vm_area_struct *vma, 1184af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 1185af0ed73eSJoonsoo Kim long chg) 11861da177e4SLinus Torvalds { 1187cfcaa66fSBen Widawsky struct page *page = NULL; 1188480eccf9SLee Schermerhorn struct mempolicy *mpol; 118904ec6264SVlastimil Babka gfp_t gfp_mask; 11903e59fcb0SMichal Hocko nodemask_t *nodemask; 119104ec6264SVlastimil Babka int nid; 11921da177e4SLinus Torvalds 1193a1e78772SMel Gorman /* 1194a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 1195a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 1196a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 1197a1e78772SMel Gorman */ 1198af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg) && 1199a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 1200c0ff7453SMiao Xie goto err; 1201a1e78772SMel Gorman 120204f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 1203a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 12046eab04a8SJustin P. Mattock goto err; 120504f2cbe3SMel Gorman 120604ec6264SVlastimil Babka gfp_mask = htlb_alloc_mask(h); 120704ec6264SVlastimil Babka nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 1208cfcaa66fSBen Widawsky 1209cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 12103e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 1211cfcaa66fSBen Widawsky 1212cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 1213cfcaa66fSBen Widawsky nodemask = NULL; 1214cfcaa66fSBen Widawsky } 1215cfcaa66fSBen Widawsky 1216cfcaa66fSBen Widawsky if (!page) 1217cfcaa66fSBen Widawsky page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 1218cfcaa66fSBen Widawsky 12193e59fcb0SMichal Hocko if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { 1220d6995da3SMike Kravetz SetHPageRestoreReserve(page); 1221a63884e9SJoonsoo Kim h->resv_huge_pages--; 1222bf50bab2SNaoya Horiguchi } 1223cc9a6c87SMel Gorman 1224cc9a6c87SMel Gorman mpol_cond_put(mpol); 1225cc9a6c87SMel Gorman return page; 1226cc9a6c87SMel Gorman 1227c0ff7453SMiao Xie err: 1228cc9a6c87SMel Gorman return NULL; 12291da177e4SLinus Torvalds } 12301da177e4SLinus Torvalds 12311cac6f2cSLuiz Capitulino /* 12321cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 12331cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 12341cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 12351cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 12361cac6f2cSLuiz Capitulino * node for alloc or free. 12371cac6f2cSLuiz Capitulino */ 12381cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 12391cac6f2cSLuiz Capitulino { 12400edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 12411cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 12421cac6f2cSLuiz Capitulino 12431cac6f2cSLuiz Capitulino return nid; 12441cac6f2cSLuiz Capitulino } 12451cac6f2cSLuiz Capitulino 12461cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 12471cac6f2cSLuiz Capitulino { 12481cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 12491cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 12501cac6f2cSLuiz Capitulino return nid; 12511cac6f2cSLuiz Capitulino } 12521cac6f2cSLuiz Capitulino 12531cac6f2cSLuiz Capitulino /* 12541cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 12551cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 12561cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 12571cac6f2cSLuiz Capitulino * mask. 12581cac6f2cSLuiz Capitulino */ 12591cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 12601cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 12611cac6f2cSLuiz Capitulino { 12621cac6f2cSLuiz Capitulino int nid; 12631cac6f2cSLuiz Capitulino 12641cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 12651cac6f2cSLuiz Capitulino 12661cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 12671cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 12681cac6f2cSLuiz Capitulino 12691cac6f2cSLuiz Capitulino return nid; 12701cac6f2cSLuiz Capitulino } 12711cac6f2cSLuiz Capitulino 12721cac6f2cSLuiz Capitulino /* 127310c6ec49SMike Kravetz * helper for remove_pool_huge_page() - return the previously saved 12741cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 12751cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 12761cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 12771cac6f2cSLuiz Capitulino */ 12781cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 12791cac6f2cSLuiz Capitulino { 12801cac6f2cSLuiz Capitulino int nid; 12811cac6f2cSLuiz Capitulino 12821cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 12831cac6f2cSLuiz Capitulino 12841cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 12851cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 12861cac6f2cSLuiz Capitulino 12871cac6f2cSLuiz Capitulino return nid; 12881cac6f2cSLuiz Capitulino } 12891cac6f2cSLuiz Capitulino 12901cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 12911cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 12921cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 12931cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 12941cac6f2cSLuiz Capitulino nr_nodes--) 12951cac6f2cSLuiz Capitulino 12961cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 12971cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 12981cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 12991cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 13001cac6f2cSLuiz Capitulino nr_nodes--) 13011cac6f2cSLuiz Capitulino 13028531fc6fSMike Kravetz /* used to demote non-gigantic_huge pages as well */ 130334d9e35bSMike Kravetz static void __destroy_compound_gigantic_page(struct page *page, 130434d9e35bSMike Kravetz unsigned int order, bool demote) 1305944d9fecSLuiz Capitulino { 1306944d9fecSLuiz Capitulino int i; 1307944d9fecSLuiz Capitulino int nr_pages = 1 << order; 1308944d9fecSLuiz Capitulino struct page *p = page + 1; 1309944d9fecSLuiz Capitulino 1310c8cc708aSGerald Schaefer atomic_set(compound_mapcount_ptr(page), 0); 131147e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 131247e29d32SJohn Hubbard 1313944d9fecSLuiz Capitulino for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1314a01f4390SMike Kravetz p->mapping = NULL; 13151d798ca3SKirill A. Shutemov clear_compound_head(p); 131634d9e35bSMike Kravetz if (!demote) 1317944d9fecSLuiz Capitulino set_page_refcounted(p); 1318944d9fecSLuiz Capitulino } 1319944d9fecSLuiz Capitulino 1320944d9fecSLuiz Capitulino set_compound_order(page, 0); 1321ba9c1201SGerald Schaefer page[1].compound_nr = 0; 1322944d9fecSLuiz Capitulino __ClearPageHead(page); 1323944d9fecSLuiz Capitulino } 1324944d9fecSLuiz Capitulino 13258531fc6fSMike Kravetz static void destroy_compound_hugetlb_page_for_demote(struct page *page, 13268531fc6fSMike Kravetz unsigned int order) 13278531fc6fSMike Kravetz { 13288531fc6fSMike Kravetz __destroy_compound_gigantic_page(page, order, true); 13298531fc6fSMike Kravetz } 13308531fc6fSMike Kravetz 13318531fc6fSMike Kravetz #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 133234d9e35bSMike Kravetz static void destroy_compound_gigantic_page(struct page *page, 133334d9e35bSMike Kravetz unsigned int order) 133434d9e35bSMike Kravetz { 133534d9e35bSMike Kravetz __destroy_compound_gigantic_page(page, order, false); 133634d9e35bSMike Kravetz } 133734d9e35bSMike Kravetz 1338d00181b9SKirill A. Shutemov static void free_gigantic_page(struct page *page, unsigned int order) 1339944d9fecSLuiz Capitulino { 1340cf11e85fSRoman Gushchin /* 1341cf11e85fSRoman Gushchin * If the page isn't allocated using the cma allocator, 1342cf11e85fSRoman Gushchin * cma_release() returns false. 1343cf11e85fSRoman Gushchin */ 1344dbda8feaSBarry Song #ifdef CONFIG_CMA 1345dbda8feaSBarry Song if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order)) 1346cf11e85fSRoman Gushchin return; 1347dbda8feaSBarry Song #endif 1348cf11e85fSRoman Gushchin 1349944d9fecSLuiz Capitulino free_contig_range(page_to_pfn(page), 1 << order); 1350944d9fecSLuiz Capitulino } 1351944d9fecSLuiz Capitulino 13524eb0716eSAlexandre Ghiti #ifdef CONFIG_CONTIG_ALLOC 1353d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 1354d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) 1355944d9fecSLuiz Capitulino { 135604adbc3fSMiaohe Lin unsigned long nr_pages = pages_per_huge_page(h); 1357953f064aSLi Xinhai if (nid == NUMA_NO_NODE) 1358953f064aSLi Xinhai nid = numa_mem_id(); 1359944d9fecSLuiz Capitulino 1360dbda8feaSBarry Song #ifdef CONFIG_CMA 1361dbda8feaSBarry Song { 1362cf11e85fSRoman Gushchin struct page *page; 1363cf11e85fSRoman Gushchin int node; 1364cf11e85fSRoman Gushchin 1365953f064aSLi Xinhai if (hugetlb_cma[nid]) { 1366953f064aSLi Xinhai page = cma_alloc(hugetlb_cma[nid], nr_pages, 1367953f064aSLi Xinhai huge_page_order(h), true); 1368953f064aSLi Xinhai if (page) 1369953f064aSLi Xinhai return page; 1370953f064aSLi Xinhai } 1371953f064aSLi Xinhai 1372953f064aSLi Xinhai if (!(gfp_mask & __GFP_THISNODE)) { 1373cf11e85fSRoman Gushchin for_each_node_mask(node, *nodemask) { 1374953f064aSLi Xinhai if (node == nid || !hugetlb_cma[node]) 1375cf11e85fSRoman Gushchin continue; 1376cf11e85fSRoman Gushchin 1377cf11e85fSRoman Gushchin page = cma_alloc(hugetlb_cma[node], nr_pages, 1378cf11e85fSRoman Gushchin huge_page_order(h), true); 1379cf11e85fSRoman Gushchin if (page) 1380cf11e85fSRoman Gushchin return page; 1381cf11e85fSRoman Gushchin } 1382cf11e85fSRoman Gushchin } 1383953f064aSLi Xinhai } 1384dbda8feaSBarry Song #endif 1385cf11e85fSRoman Gushchin 13865e27a2dfSAnshuman Khandual return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); 1387944d9fecSLuiz Capitulino } 1388944d9fecSLuiz Capitulino 13894eb0716eSAlexandre Ghiti #else /* !CONFIG_CONTIG_ALLOC */ 13904eb0716eSAlexandre Ghiti static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 13914eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 13924eb0716eSAlexandre Ghiti { 13934eb0716eSAlexandre Ghiti return NULL; 13944eb0716eSAlexandre Ghiti } 13954eb0716eSAlexandre Ghiti #endif /* CONFIG_CONTIG_ALLOC */ 1396944d9fecSLuiz Capitulino 1397e1073d1eSAneesh Kumar K.V #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 1398d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 13994eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 14004eb0716eSAlexandre Ghiti { 14014eb0716eSAlexandre Ghiti return NULL; 14024eb0716eSAlexandre Ghiti } 1403d00181b9SKirill A. Shutemov static inline void free_gigantic_page(struct page *page, unsigned int order) { } 1404944d9fecSLuiz Capitulino static inline void destroy_compound_gigantic_page(struct page *page, 1405d00181b9SKirill A. Shutemov unsigned int order) { } 1406944d9fecSLuiz Capitulino #endif 1407944d9fecSLuiz Capitulino 14086eb4e88aSMike Kravetz /* 14096eb4e88aSMike Kravetz * Remove hugetlb page from lists, and update dtor so that page appears 141034d9e35bSMike Kravetz * as just a compound page. 141134d9e35bSMike Kravetz * 141234d9e35bSMike Kravetz * A reference is held on the page, except in the case of demote. 14136eb4e88aSMike Kravetz * 14146eb4e88aSMike Kravetz * Must be called with hugetlb lock held. 14156eb4e88aSMike Kravetz */ 141634d9e35bSMike Kravetz static void __remove_hugetlb_page(struct hstate *h, struct page *page, 141734d9e35bSMike Kravetz bool adjust_surplus, 141834d9e35bSMike Kravetz bool demote) 14196eb4e88aSMike Kravetz { 14206eb4e88aSMike Kravetz int nid = page_to_nid(page); 14216eb4e88aSMike Kravetz 14226eb4e88aSMike Kravetz VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 14236eb4e88aSMike Kravetz VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page); 14246eb4e88aSMike Kravetz 14259487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 14266eb4e88aSMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 14276eb4e88aSMike Kravetz return; 14286eb4e88aSMike Kravetz 14296eb4e88aSMike Kravetz list_del(&page->lru); 14306eb4e88aSMike Kravetz 14316eb4e88aSMike Kravetz if (HPageFreed(page)) { 14326eb4e88aSMike Kravetz h->free_huge_pages--; 14336eb4e88aSMike Kravetz h->free_huge_pages_node[nid]--; 14346eb4e88aSMike Kravetz } 14356eb4e88aSMike Kravetz if (adjust_surplus) { 14366eb4e88aSMike Kravetz h->surplus_huge_pages--; 14376eb4e88aSMike Kravetz h->surplus_huge_pages_node[nid]--; 14386eb4e88aSMike Kravetz } 14396eb4e88aSMike Kravetz 1440e32d20c0SMike Kravetz /* 1441e32d20c0SMike Kravetz * Very subtle 1442e32d20c0SMike Kravetz * 1443e32d20c0SMike Kravetz * For non-gigantic pages set the destructor to the normal compound 1444e32d20c0SMike Kravetz * page dtor. This is needed in case someone takes an additional 1445e32d20c0SMike Kravetz * temporary ref to the page, and freeing is delayed until they drop 1446e32d20c0SMike Kravetz * their reference. 1447e32d20c0SMike Kravetz * 1448e32d20c0SMike Kravetz * For gigantic pages set the destructor to the null dtor. This 1449e32d20c0SMike Kravetz * destructor will never be called. Before freeing the gigantic 1450e32d20c0SMike Kravetz * page destroy_compound_gigantic_page will turn the compound page 1451e32d20c0SMike Kravetz * into a simple group of pages. After this the destructor does not 1452e32d20c0SMike Kravetz * apply. 1453e32d20c0SMike Kravetz * 1454e32d20c0SMike Kravetz * This handles the case where more than one ref is held when and 1455e32d20c0SMike Kravetz * after update_and_free_page is called. 145634d9e35bSMike Kravetz * 145734d9e35bSMike Kravetz * In the case of demote we do not ref count the page as it will soon 145834d9e35bSMike Kravetz * be turned into a page of smaller size. 1459e32d20c0SMike Kravetz */ 146034d9e35bSMike Kravetz if (!demote) 14616eb4e88aSMike Kravetz set_page_refcounted(page); 1462e32d20c0SMike Kravetz if (hstate_is_gigantic(h)) 14636eb4e88aSMike Kravetz set_compound_page_dtor(page, NULL_COMPOUND_DTOR); 1464e32d20c0SMike Kravetz else 1465e32d20c0SMike Kravetz set_compound_page_dtor(page, COMPOUND_PAGE_DTOR); 14666eb4e88aSMike Kravetz 14676eb4e88aSMike Kravetz h->nr_huge_pages--; 14686eb4e88aSMike Kravetz h->nr_huge_pages_node[nid]--; 14696eb4e88aSMike Kravetz } 14706eb4e88aSMike Kravetz 147134d9e35bSMike Kravetz static void remove_hugetlb_page(struct hstate *h, struct page *page, 147234d9e35bSMike Kravetz bool adjust_surplus) 147334d9e35bSMike Kravetz { 147434d9e35bSMike Kravetz __remove_hugetlb_page(h, page, adjust_surplus, false); 147534d9e35bSMike Kravetz } 147634d9e35bSMike Kravetz 14778531fc6fSMike Kravetz static void remove_hugetlb_page_for_demote(struct hstate *h, struct page *page, 14788531fc6fSMike Kravetz bool adjust_surplus) 14798531fc6fSMike Kravetz { 14808531fc6fSMike Kravetz __remove_hugetlb_page(h, page, adjust_surplus, true); 14818531fc6fSMike Kravetz } 14828531fc6fSMike Kravetz 1483ad2fa371SMuchun Song static void add_hugetlb_page(struct hstate *h, struct page *page, 1484ad2fa371SMuchun Song bool adjust_surplus) 1485ad2fa371SMuchun Song { 1486ad2fa371SMuchun Song int zeroed; 1487ad2fa371SMuchun Song int nid = page_to_nid(page); 1488ad2fa371SMuchun Song 1489ad2fa371SMuchun Song VM_BUG_ON_PAGE(!HPageVmemmapOptimized(page), page); 1490ad2fa371SMuchun Song 1491ad2fa371SMuchun Song lockdep_assert_held(&hugetlb_lock); 1492ad2fa371SMuchun Song 1493ad2fa371SMuchun Song INIT_LIST_HEAD(&page->lru); 1494ad2fa371SMuchun Song h->nr_huge_pages++; 1495ad2fa371SMuchun Song h->nr_huge_pages_node[nid]++; 1496ad2fa371SMuchun Song 1497ad2fa371SMuchun Song if (adjust_surplus) { 1498ad2fa371SMuchun Song h->surplus_huge_pages++; 1499ad2fa371SMuchun Song h->surplus_huge_pages_node[nid]++; 1500ad2fa371SMuchun Song } 1501ad2fa371SMuchun Song 1502ad2fa371SMuchun Song set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1503ad2fa371SMuchun Song set_page_private(page, 0); 1504ad2fa371SMuchun Song SetHPageVmemmapOptimized(page); 1505ad2fa371SMuchun Song 1506ad2fa371SMuchun Song /* 1507b65a4edaSMike Kravetz * This page is about to be managed by the hugetlb allocator and 1508b65a4edaSMike Kravetz * should have no users. Drop our reference, and check for others 1509b65a4edaSMike Kravetz * just in case. 1510ad2fa371SMuchun Song */ 1511ad2fa371SMuchun Song zeroed = put_page_testzero(page); 1512b65a4edaSMike Kravetz if (!zeroed) 1513b65a4edaSMike Kravetz /* 1514b65a4edaSMike Kravetz * It is VERY unlikely soneone else has taken a ref on 1515b65a4edaSMike Kravetz * the page. In this case, we simply return as the 1516b65a4edaSMike Kravetz * hugetlb destructor (free_huge_page) will be called 1517b65a4edaSMike Kravetz * when this other ref is dropped. 1518b65a4edaSMike Kravetz */ 1519b65a4edaSMike Kravetz return; 1520b65a4edaSMike Kravetz 1521ad2fa371SMuchun Song arch_clear_hugepage_flags(page); 1522ad2fa371SMuchun Song enqueue_huge_page(h, page); 1523ad2fa371SMuchun Song } 1524ad2fa371SMuchun Song 1525b65d4adbSMuchun Song static void __update_and_free_page(struct hstate *h, struct page *page) 15266af2acb6SAdam Litke { 15276af2acb6SAdam Litke int i; 1528dbfee5aeSMike Kravetz struct page *subpage = page; 1529a5516438SAndi Kleen 15304eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1531944d9fecSLuiz Capitulino return; 153218229df5SAndy Whitcroft 1533ad2fa371SMuchun Song if (alloc_huge_page_vmemmap(h, page)) { 1534ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 1535ad2fa371SMuchun Song /* 1536ad2fa371SMuchun Song * If we cannot allocate vmemmap pages, just refuse to free the 1537ad2fa371SMuchun Song * page and put the page back on the hugetlb free list and treat 1538ad2fa371SMuchun Song * as a surplus page. 1539ad2fa371SMuchun Song */ 1540ad2fa371SMuchun Song add_hugetlb_page(h, page, true); 1541ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 1542ad2fa371SMuchun Song return; 1543ad2fa371SMuchun Song } 1544ad2fa371SMuchun Song 1545dbfee5aeSMike Kravetz for (i = 0; i < pages_per_huge_page(h); 1546dbfee5aeSMike Kravetz i++, subpage = mem_map_next(subpage, page, i)) { 1547dbfee5aeSMike Kravetz subpage->flags &= ~(1 << PG_locked | 1 << PG_error | 154832f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1549a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1550a7407a27SLuiz Capitulino 1 << PG_writeback); 15516af2acb6SAdam Litke } 1552a01f4390SMike Kravetz 1553a01f4390SMike Kravetz /* 1554a01f4390SMike Kravetz * Non-gigantic pages demoted from CMA allocated gigantic pages 1555a01f4390SMike Kravetz * need to be given back to CMA in free_gigantic_page. 1556a01f4390SMike Kravetz */ 1557a01f4390SMike Kravetz if (hstate_is_gigantic(h) || 1558a01f4390SMike Kravetz hugetlb_cma_page(page, huge_page_order(h))) { 1559944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 1560944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 1561944d9fecSLuiz Capitulino } else { 1562a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 15636af2acb6SAdam Litke } 1564944d9fecSLuiz Capitulino } 15656af2acb6SAdam Litke 1566b65d4adbSMuchun Song /* 1567b65d4adbSMuchun Song * As update_and_free_page() can be called under any context, so we cannot 1568b65d4adbSMuchun Song * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the 1569b65d4adbSMuchun Song * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate 1570b65d4adbSMuchun Song * the vmemmap pages. 1571b65d4adbSMuchun Song * 1572b65d4adbSMuchun Song * free_hpage_workfn() locklessly retrieves the linked list of pages to be 1573b65d4adbSMuchun Song * freed and frees them one-by-one. As the page->mapping pointer is going 1574b65d4adbSMuchun Song * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node 1575b65d4adbSMuchun Song * structure of a lockless linked list of huge pages to be freed. 1576b65d4adbSMuchun Song */ 1577b65d4adbSMuchun Song static LLIST_HEAD(hpage_freelist); 1578b65d4adbSMuchun Song 1579b65d4adbSMuchun Song static void free_hpage_workfn(struct work_struct *work) 1580b65d4adbSMuchun Song { 1581b65d4adbSMuchun Song struct llist_node *node; 1582b65d4adbSMuchun Song 1583b65d4adbSMuchun Song node = llist_del_all(&hpage_freelist); 1584b65d4adbSMuchun Song 1585b65d4adbSMuchun Song while (node) { 1586b65d4adbSMuchun Song struct page *page; 1587b65d4adbSMuchun Song struct hstate *h; 1588b65d4adbSMuchun Song 1589b65d4adbSMuchun Song page = container_of((struct address_space **)node, 1590b65d4adbSMuchun Song struct page, mapping); 1591b65d4adbSMuchun Song node = node->next; 1592b65d4adbSMuchun Song page->mapping = NULL; 1593b65d4adbSMuchun Song /* 1594b65d4adbSMuchun Song * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate() 1595b65d4adbSMuchun Song * is going to trigger because a previous call to 1596b65d4adbSMuchun Song * remove_hugetlb_page() will set_compound_page_dtor(page, 1597b65d4adbSMuchun Song * NULL_COMPOUND_DTOR), so do not use page_hstate() directly. 1598b65d4adbSMuchun Song */ 1599b65d4adbSMuchun Song h = size_to_hstate(page_size(page)); 1600b65d4adbSMuchun Song 1601b65d4adbSMuchun Song __update_and_free_page(h, page); 1602b65d4adbSMuchun Song 1603b65d4adbSMuchun Song cond_resched(); 1604b65d4adbSMuchun Song } 1605b65d4adbSMuchun Song } 1606b65d4adbSMuchun Song static DECLARE_WORK(free_hpage_work, free_hpage_workfn); 1607b65d4adbSMuchun Song 1608b65d4adbSMuchun Song static inline void flush_free_hpage_work(struct hstate *h) 1609b65d4adbSMuchun Song { 1610b65d4adbSMuchun Song if (free_vmemmap_pages_per_hpage(h)) 1611b65d4adbSMuchun Song flush_work(&free_hpage_work); 1612b65d4adbSMuchun Song } 1613b65d4adbSMuchun Song 1614b65d4adbSMuchun Song static void update_and_free_page(struct hstate *h, struct page *page, 1615b65d4adbSMuchun Song bool atomic) 1616b65d4adbSMuchun Song { 1617ad2fa371SMuchun Song if (!HPageVmemmapOptimized(page) || !atomic) { 1618b65d4adbSMuchun Song __update_and_free_page(h, page); 1619b65d4adbSMuchun Song return; 1620b65d4adbSMuchun Song } 1621b65d4adbSMuchun Song 1622b65d4adbSMuchun Song /* 1623b65d4adbSMuchun Song * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages. 1624b65d4adbSMuchun Song * 1625b65d4adbSMuchun Song * Only call schedule_work() if hpage_freelist is previously 1626b65d4adbSMuchun Song * empty. Otherwise, schedule_work() had been called but the workfn 1627b65d4adbSMuchun Song * hasn't retrieved the list yet. 1628b65d4adbSMuchun Song */ 1629b65d4adbSMuchun Song if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist)) 1630b65d4adbSMuchun Song schedule_work(&free_hpage_work); 1631b65d4adbSMuchun Song } 1632b65d4adbSMuchun Song 163310c6ec49SMike Kravetz static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list) 163410c6ec49SMike Kravetz { 163510c6ec49SMike Kravetz struct page *page, *t_page; 163610c6ec49SMike Kravetz 163710c6ec49SMike Kravetz list_for_each_entry_safe(page, t_page, list, lru) { 1638b65d4adbSMuchun Song update_and_free_page(h, page, false); 163910c6ec49SMike Kravetz cond_resched(); 164010c6ec49SMike Kravetz } 164110c6ec49SMike Kravetz } 164210c6ec49SMike Kravetz 1643e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1644e5ff2159SAndi Kleen { 1645e5ff2159SAndi Kleen struct hstate *h; 1646e5ff2159SAndi Kleen 1647e5ff2159SAndi Kleen for_each_hstate(h) { 1648e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1649e5ff2159SAndi Kleen return h; 1650e5ff2159SAndi Kleen } 1651e5ff2159SAndi Kleen return NULL; 1652e5ff2159SAndi Kleen } 1653e5ff2159SAndi Kleen 1654db71ef79SMike Kravetz void free_huge_page(struct page *page) 165527a85ef1SDavid Gibson { 1656a5516438SAndi Kleen /* 1657a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1658a5516438SAndi Kleen * compound page destructor. 1659a5516438SAndi Kleen */ 1660e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 16617893d1d5SAdam Litke int nid = page_to_nid(page); 1662d6995da3SMike Kravetz struct hugepage_subpool *spool = hugetlb_page_subpool(page); 166307443a85SJoonsoo Kim bool restore_reserve; 1664db71ef79SMike Kravetz unsigned long flags; 166527a85ef1SDavid Gibson 1666b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1667b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_mapcount(page), page); 16688ace22bcSYongkai Wu 1669d6995da3SMike Kravetz hugetlb_set_page_subpool(page, NULL); 16708ace22bcSYongkai Wu page->mapping = NULL; 1671d6995da3SMike Kravetz restore_reserve = HPageRestoreReserve(page); 1672d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 167327a85ef1SDavid Gibson 16741c5ecae3SMike Kravetz /* 1675d6995da3SMike Kravetz * If HPageRestoreReserve was set on page, page allocation consumed a 16760919e1b6SMike Kravetz * reservation. If the page was associated with a subpool, there 16770919e1b6SMike Kravetz * would have been a page reserved in the subpool before allocation 16780919e1b6SMike Kravetz * via hugepage_subpool_get_pages(). Since we are 'restoring' the 16796c26d310SMiaohe Lin * reservation, do not call hugepage_subpool_put_pages() as this will 16800919e1b6SMike Kravetz * remove the reserved page from the subpool. 16810919e1b6SMike Kravetz */ 16820919e1b6SMike Kravetz if (!restore_reserve) { 16830919e1b6SMike Kravetz /* 16840919e1b6SMike Kravetz * A return code of zero implies that the subpool will be 16850919e1b6SMike Kravetz * under its minimum size if the reservation is not restored 16860919e1b6SMike Kravetz * after page is free. Therefore, force restore_reserve 16870919e1b6SMike Kravetz * operation. 16881c5ecae3SMike Kravetz */ 16891c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 16901c5ecae3SMike Kravetz restore_reserve = true; 16910919e1b6SMike Kravetz } 16921c5ecae3SMike Kravetz 1693db71ef79SMike Kravetz spin_lock_irqsave(&hugetlb_lock, flags); 16948f251a3dSMike Kravetz ClearHPageMigratable(page); 16956d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 16966d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 169708cf9fafSMina Almasry hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 169808cf9fafSMina Almasry pages_per_huge_page(h), page); 169907443a85SJoonsoo Kim if (restore_reserve) 170007443a85SJoonsoo Kim h->resv_huge_pages++; 170107443a85SJoonsoo Kim 17029157c311SMike Kravetz if (HPageTemporary(page)) { 17036eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 1704db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1705b65d4adbSMuchun Song update_and_free_page(h, page, true); 1706ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 17070edaecfaSAneesh Kumar K.V /* remove the page from active list */ 17086eb4e88aSMike Kravetz remove_hugetlb_page(h, page, true); 1709db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1710b65d4adbSMuchun Song update_and_free_page(h, page, true); 17117893d1d5SAdam Litke } else { 17125d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1713a5516438SAndi Kleen enqueue_huge_page(h, page); 1714db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 171527a85ef1SDavid Gibson } 17161121828aSMike Kravetz } 171727a85ef1SDavid Gibson 1718d3d99fccSOscar Salvador /* 1719d3d99fccSOscar Salvador * Must be called with the hugetlb lock held 1720d3d99fccSOscar Salvador */ 1721d3d99fccSOscar Salvador static void __prep_account_new_huge_page(struct hstate *h, int nid) 1722d3d99fccSOscar Salvador { 1723d3d99fccSOscar Salvador lockdep_assert_held(&hugetlb_lock); 1724d3d99fccSOscar Salvador h->nr_huge_pages++; 1725d3d99fccSOscar Salvador h->nr_huge_pages_node[nid]++; 1726d3d99fccSOscar Salvador } 1727d3d99fccSOscar Salvador 1728f41f2ed4SMuchun Song static void __prep_new_huge_page(struct hstate *h, struct page *page) 1729b7ba30c6SAndi Kleen { 1730f41f2ed4SMuchun Song free_huge_page_vmemmap(h, page); 17310edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1732f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1733ff546117SMike Kravetz hugetlb_set_page_subpool(page, NULL); 17349dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 17351adc4d41SMina Almasry set_hugetlb_cgroup_rsvd(page, NULL); 1736d3d99fccSOscar Salvador } 1737d3d99fccSOscar Salvador 1738d3d99fccSOscar Salvador static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1739d3d99fccSOscar Salvador { 1740f41f2ed4SMuchun Song __prep_new_huge_page(h, page); 1741db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 1742d3d99fccSOscar Salvador __prep_account_new_huge_page(h, nid); 1743db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1744b7ba30c6SAndi Kleen } 1745b7ba30c6SAndi Kleen 174634d9e35bSMike Kravetz static bool __prep_compound_gigantic_page(struct page *page, unsigned int order, 174734d9e35bSMike Kravetz bool demote) 174820a0307cSWu Fengguang { 17497118fc29SMike Kravetz int i, j; 175020a0307cSWu Fengguang int nr_pages = 1 << order; 175120a0307cSWu Fengguang struct page *p = page + 1; 175220a0307cSWu Fengguang 175320a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 175420a0307cSWu Fengguang set_compound_order(page, order); 1755ef5a22beSAndrea Arcangeli __ClearPageReserved(page); 1756de09d31dSKirill A. Shutemov __SetPageHead(page); 175720a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1758ef5a22beSAndrea Arcangeli /* 1759ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1760ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1761ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 17627c8de358SEthon Paul * too. Otherwise drivers using get_user_pages() to access tail 1763ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1764ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1765ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1766ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1767ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1768ef5a22beSAndrea Arcangeli * after get_user_pages(). 1769ef5a22beSAndrea Arcangeli */ 1770ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 17717118fc29SMike Kravetz /* 17727118fc29SMike Kravetz * Subtle and very unlikely 17737118fc29SMike Kravetz * 17747118fc29SMike Kravetz * Gigantic 'page allocators' such as memblock or cma will 17757118fc29SMike Kravetz * return a set of pages with each page ref counted. We need 17767118fc29SMike Kravetz * to turn this set of pages into a compound page with tail 17777118fc29SMike Kravetz * page ref counts set to zero. Code such as speculative page 17787118fc29SMike Kravetz * cache adding could take a ref on a 'to be' tail page. 17797118fc29SMike Kravetz * We need to respect any increased ref count, and only set 17807118fc29SMike Kravetz * the ref count to zero if count is currently 1. If count 1781416d85edSMike Kravetz * is not 1, we return an error. An error return indicates 1782416d85edSMike Kravetz * the set of pages can not be converted to a gigantic page. 1783416d85edSMike Kravetz * The caller who allocated the pages should then discard the 1784416d85edSMike Kravetz * pages using the appropriate free interface. 178534d9e35bSMike Kravetz * 178634d9e35bSMike Kravetz * In the case of demote, the ref count will be zero. 17877118fc29SMike Kravetz */ 178834d9e35bSMike Kravetz if (!demote) { 17897118fc29SMike Kravetz if (!page_ref_freeze(p, 1)) { 1790416d85edSMike Kravetz pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n"); 17917118fc29SMike Kravetz goto out_error; 17927118fc29SMike Kravetz } 179334d9e35bSMike Kravetz } else { 179434d9e35bSMike Kravetz VM_BUG_ON_PAGE(page_count(p), p); 179534d9e35bSMike Kravetz } 179658a84aa9SYouquan Song set_page_count(p, 0); 17971d798ca3SKirill A. Shutemov set_compound_head(p, page); 179820a0307cSWu Fengguang } 1799b4330afbSMike Kravetz atomic_set(compound_mapcount_ptr(page), -1); 180047e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 18017118fc29SMike Kravetz return true; 18027118fc29SMike Kravetz 18037118fc29SMike Kravetz out_error: 18047118fc29SMike Kravetz /* undo tail page modifications made above */ 18057118fc29SMike Kravetz p = page + 1; 18067118fc29SMike Kravetz for (j = 1; j < i; j++, p = mem_map_next(p, page, j)) { 18077118fc29SMike Kravetz clear_compound_head(p); 18087118fc29SMike Kravetz set_page_refcounted(p); 18097118fc29SMike Kravetz } 18107118fc29SMike Kravetz /* need to clear PG_reserved on remaining tail pages */ 18117118fc29SMike Kravetz for (; j < nr_pages; j++, p = mem_map_next(p, page, j)) 18127118fc29SMike Kravetz __ClearPageReserved(p); 18137118fc29SMike Kravetz set_compound_order(page, 0); 18147118fc29SMike Kravetz page[1].compound_nr = 0; 18157118fc29SMike Kravetz __ClearPageHead(page); 18167118fc29SMike Kravetz return false; 181720a0307cSWu Fengguang } 181820a0307cSWu Fengguang 181934d9e35bSMike Kravetz static bool prep_compound_gigantic_page(struct page *page, unsigned int order) 182034d9e35bSMike Kravetz { 182134d9e35bSMike Kravetz return __prep_compound_gigantic_page(page, order, false); 182234d9e35bSMike Kravetz } 182334d9e35bSMike Kravetz 18248531fc6fSMike Kravetz static bool prep_compound_gigantic_page_for_demote(struct page *page, 18258531fc6fSMike Kravetz unsigned int order) 18268531fc6fSMike Kravetz { 18278531fc6fSMike Kravetz return __prep_compound_gigantic_page(page, order, true); 18288531fc6fSMike Kravetz } 18298531fc6fSMike Kravetz 18307795912cSAndrew Morton /* 18317795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 18327795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 18337795912cSAndrew Morton * details. 18347795912cSAndrew Morton */ 183520a0307cSWu Fengguang int PageHuge(struct page *page) 183620a0307cSWu Fengguang { 183720a0307cSWu Fengguang if (!PageCompound(page)) 183820a0307cSWu Fengguang return 0; 183920a0307cSWu Fengguang 184020a0307cSWu Fengguang page = compound_head(page); 1841f1e61557SKirill A. Shutemov return page[1].compound_dtor == HUGETLB_PAGE_DTOR; 184220a0307cSWu Fengguang } 184343131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 184443131e14SNaoya Horiguchi 184527c73ae7SAndrea Arcangeli /* 184627c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 184727c73ae7SAndrea Arcangeli * normal or transparent huge pages. 184827c73ae7SAndrea Arcangeli */ 184927c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 185027c73ae7SAndrea Arcangeli { 185127c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 185227c73ae7SAndrea Arcangeli return 0; 185327c73ae7SAndrea Arcangeli 1854d4af73e3SVlastimil Babka return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; 185527c73ae7SAndrea Arcangeli } 185627c73ae7SAndrea Arcangeli 1857c0d0381aSMike Kravetz /* 1858c0d0381aSMike Kravetz * Find and lock address space (mapping) in write mode. 1859c0d0381aSMike Kravetz * 1860336bf30eSMike Kravetz * Upon entry, the page is locked which means that page_mapping() is 1861336bf30eSMike Kravetz * stable. Due to locking order, we can only trylock_write. If we can 1862336bf30eSMike Kravetz * not get the lock, simply return NULL to caller. 1863c0d0381aSMike Kravetz */ 1864c0d0381aSMike Kravetz struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) 1865c0d0381aSMike Kravetz { 1866336bf30eSMike Kravetz struct address_space *mapping = page_mapping(hpage); 1867c0d0381aSMike Kravetz 1868c0d0381aSMike Kravetz if (!mapping) 1869c0d0381aSMike Kravetz return mapping; 1870c0d0381aSMike Kravetz 1871c0d0381aSMike Kravetz if (i_mmap_trylock_write(mapping)) 1872c0d0381aSMike Kravetz return mapping; 1873c0d0381aSMike Kravetz 1874c0d0381aSMike Kravetz return NULL; 1875c0d0381aSMike Kravetz } 1876c0d0381aSMike Kravetz 1877fe19bd3dSHugh Dickins pgoff_t hugetlb_basepage_index(struct page *page) 187813d60f4bSZhang Yi { 187913d60f4bSZhang Yi struct page *page_head = compound_head(page); 188013d60f4bSZhang Yi pgoff_t index = page_index(page_head); 188113d60f4bSZhang Yi unsigned long compound_idx; 188213d60f4bSZhang Yi 188313d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 188413d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 188513d60f4bSZhang Yi else 188613d60f4bSZhang Yi compound_idx = page - page_head; 188713d60f4bSZhang Yi 188813d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 188913d60f4bSZhang Yi } 189013d60f4bSZhang Yi 18910c397daeSMichal Hocko static struct page *alloc_buddy_huge_page(struct hstate *h, 1892f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1893f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 18941da177e4SLinus Torvalds { 1895af0fb9dfSMichal Hocko int order = huge_page_order(h); 18961da177e4SLinus Torvalds struct page *page; 1897f60858f9SMike Kravetz bool alloc_try_hard = true; 1898f96efd58SJoe Jin 1899f60858f9SMike Kravetz /* 1900f60858f9SMike Kravetz * By default we always try hard to allocate the page with 1901f60858f9SMike Kravetz * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in 1902f60858f9SMike Kravetz * a loop (to adjust global huge page counts) and previous allocation 1903f60858f9SMike Kravetz * failed, do not continue to try hard on the same node. Use the 1904f60858f9SMike Kravetz * node_alloc_noretry bitmap to manage this state information. 1905f60858f9SMike Kravetz */ 1906f60858f9SMike Kravetz if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) 1907f60858f9SMike Kravetz alloc_try_hard = false; 1908f60858f9SMike Kravetz gfp_mask |= __GFP_COMP|__GFP_NOWARN; 1909f60858f9SMike Kravetz if (alloc_try_hard) 1910f60858f9SMike Kravetz gfp_mask |= __GFP_RETRY_MAYFAIL; 1911af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 1912af0fb9dfSMichal Hocko nid = numa_mem_id(); 191384172f4bSMatthew Wilcox (Oracle) page = __alloc_pages(gfp_mask, order, nid, nmask); 1914af0fb9dfSMichal Hocko if (page) 1915af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC); 1916af0fb9dfSMichal Hocko else 1917af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 191863b4613cSNishanth Aravamudan 1919f60858f9SMike Kravetz /* 1920f60858f9SMike Kravetz * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this 1921f60858f9SMike Kravetz * indicates an overall state change. Clear bit so that we resume 1922f60858f9SMike Kravetz * normal 'try hard' allocations. 1923f60858f9SMike Kravetz */ 1924f60858f9SMike Kravetz if (node_alloc_noretry && page && !alloc_try_hard) 1925f60858f9SMike Kravetz node_clear(nid, *node_alloc_noretry); 1926f60858f9SMike Kravetz 1927f60858f9SMike Kravetz /* 1928f60858f9SMike Kravetz * If we tried hard to get a page but failed, set bit so that 1929f60858f9SMike Kravetz * subsequent attempts will not try as hard until there is an 1930f60858f9SMike Kravetz * overall state change. 1931f60858f9SMike Kravetz */ 1932f60858f9SMike Kravetz if (node_alloc_noretry && !page && alloc_try_hard) 1933f60858f9SMike Kravetz node_set(nid, *node_alloc_noretry); 1934f60858f9SMike Kravetz 193563b4613cSNishanth Aravamudan return page; 193663b4613cSNishanth Aravamudan } 193763b4613cSNishanth Aravamudan 1938af0fb9dfSMichal Hocko /* 19390c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 19400c397daeSMichal Hocko * should use this function to get new hugetlb pages 19410c397daeSMichal Hocko */ 19420c397daeSMichal Hocko static struct page *alloc_fresh_huge_page(struct hstate *h, 1943f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1944f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 19450c397daeSMichal Hocko { 19460c397daeSMichal Hocko struct page *page; 19477118fc29SMike Kravetz bool retry = false; 19480c397daeSMichal Hocko 19497118fc29SMike Kravetz retry: 19500c397daeSMichal Hocko if (hstate_is_gigantic(h)) 19510c397daeSMichal Hocko page = alloc_gigantic_page(h, gfp_mask, nid, nmask); 19520c397daeSMichal Hocko else 19530c397daeSMichal Hocko page = alloc_buddy_huge_page(h, gfp_mask, 1954f60858f9SMike Kravetz nid, nmask, node_alloc_noretry); 19550c397daeSMichal Hocko if (!page) 19560c397daeSMichal Hocko return NULL; 19570c397daeSMichal Hocko 19587118fc29SMike Kravetz if (hstate_is_gigantic(h)) { 19597118fc29SMike Kravetz if (!prep_compound_gigantic_page(page, huge_page_order(h))) { 19607118fc29SMike Kravetz /* 19617118fc29SMike Kravetz * Rare failure to convert pages to compound page. 19627118fc29SMike Kravetz * Free pages and try again - ONCE! 19637118fc29SMike Kravetz */ 19647118fc29SMike Kravetz free_gigantic_page(page, huge_page_order(h)); 19657118fc29SMike Kravetz if (!retry) { 19667118fc29SMike Kravetz retry = true; 19677118fc29SMike Kravetz goto retry; 19687118fc29SMike Kravetz } 19697118fc29SMike Kravetz return NULL; 19707118fc29SMike Kravetz } 19717118fc29SMike Kravetz } 19720c397daeSMichal Hocko prep_new_huge_page(h, page, page_to_nid(page)); 19730c397daeSMichal Hocko 19740c397daeSMichal Hocko return page; 19750c397daeSMichal Hocko } 19760c397daeSMichal Hocko 19770c397daeSMichal Hocko /* 1978af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 1979af0fb9dfSMichal Hocko * manner. 1980af0fb9dfSMichal Hocko */ 1981f60858f9SMike Kravetz static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 1982f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 1983b2261026SJoonsoo Kim { 1984b2261026SJoonsoo Kim struct page *page; 1985b2261026SJoonsoo Kim int nr_nodes, node; 1986af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 1987b2261026SJoonsoo Kim 1988b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1989f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, 1990f60858f9SMike Kravetz node_alloc_noretry); 1991af0fb9dfSMichal Hocko if (page) 1992b2261026SJoonsoo Kim break; 1993b2261026SJoonsoo Kim } 1994b2261026SJoonsoo Kim 1995af0fb9dfSMichal Hocko if (!page) 1996af0fb9dfSMichal Hocko return 0; 1997b2261026SJoonsoo Kim 1998af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 1999af0fb9dfSMichal Hocko 2000af0fb9dfSMichal Hocko return 1; 2001b2261026SJoonsoo Kim } 2002b2261026SJoonsoo Kim 2003e8c5c824SLee Schermerhorn /* 200410c6ec49SMike Kravetz * Remove huge page from pool from next node to free. Attempt to keep 200510c6ec49SMike Kravetz * persistent huge pages more or less balanced over allowed nodes. 200610c6ec49SMike Kravetz * This routine only 'removes' the hugetlb page. The caller must make 200710c6ec49SMike Kravetz * an additional call to free the page to low level allocators. 2008e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 2009e8c5c824SLee Schermerhorn */ 201010c6ec49SMike Kravetz static struct page *remove_pool_huge_page(struct hstate *h, 201110c6ec49SMike Kravetz nodemask_t *nodes_allowed, 20126ae11b27SLee Schermerhorn bool acct_surplus) 2013e8c5c824SLee Schermerhorn { 2014b2261026SJoonsoo Kim int nr_nodes, node; 201510c6ec49SMike Kravetz struct page *page = NULL; 2016e8c5c824SLee Schermerhorn 20179487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2018b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2019685f3457SLee Schermerhorn /* 2020685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 2021685f3457SLee Schermerhorn * nodes with surplus pages. 2022685f3457SLee Schermerhorn */ 2023b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 2024b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 202510c6ec49SMike Kravetz page = list_entry(h->hugepage_freelists[node].next, 2026e8c5c824SLee Schermerhorn struct page, lru); 20276eb4e88aSMike Kravetz remove_hugetlb_page(h, page, acct_surplus); 20289a76db09SLee Schermerhorn break; 2029e8c5c824SLee Schermerhorn } 2030b2261026SJoonsoo Kim } 2031e8c5c824SLee Schermerhorn 203210c6ec49SMike Kravetz return page; 2033e8c5c824SLee Schermerhorn } 2034e8c5c824SLee Schermerhorn 2035c8721bbbSNaoya Horiguchi /* 2036c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 2037faf53defSNaoya Horiguchi * nothing for in-use hugepages and non-hugepages. 2038faf53defSNaoya Horiguchi * This function returns values like below: 2039faf53defSNaoya Horiguchi * 2040ad2fa371SMuchun Song * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages 2041ad2fa371SMuchun Song * when the system is under memory pressure and the feature of 2042ad2fa371SMuchun Song * freeing unused vmemmap pages associated with each hugetlb page 2043ad2fa371SMuchun Song * is enabled. 2044faf53defSNaoya Horiguchi * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 2045faf53defSNaoya Horiguchi * (allocated or reserved.) 2046faf53defSNaoya Horiguchi * 0: successfully dissolved free hugepages or the page is not a 2047faf53defSNaoya Horiguchi * hugepage (considered as already dissolved) 2048c8721bbbSNaoya Horiguchi */ 2049c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 2050c8721bbbSNaoya Horiguchi { 20516bc9b564SNaoya Horiguchi int rc = -EBUSY; 2052082d5b6bSGerald Schaefer 20537ffddd49SMuchun Song retry: 2054faf53defSNaoya Horiguchi /* Not to disrupt normal path by vainly holding hugetlb_lock */ 2055faf53defSNaoya Horiguchi if (!PageHuge(page)) 2056faf53defSNaoya Horiguchi return 0; 2057faf53defSNaoya Horiguchi 2058db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2059faf53defSNaoya Horiguchi if (!PageHuge(page)) { 2060faf53defSNaoya Horiguchi rc = 0; 2061faf53defSNaoya Horiguchi goto out; 2062faf53defSNaoya Horiguchi } 2063faf53defSNaoya Horiguchi 2064faf53defSNaoya Horiguchi if (!page_count(page)) { 20652247bb33SGerald Schaefer struct page *head = compound_head(page); 20662247bb33SGerald Schaefer struct hstate *h = page_hstate(head); 20676bc9b564SNaoya Horiguchi if (h->free_huge_pages - h->resv_huge_pages == 0) 2068082d5b6bSGerald Schaefer goto out; 20697ffddd49SMuchun Song 20707ffddd49SMuchun Song /* 20717ffddd49SMuchun Song * We should make sure that the page is already on the free list 20727ffddd49SMuchun Song * when it is dissolved. 20737ffddd49SMuchun Song */ 20746c037149SMike Kravetz if (unlikely(!HPageFreed(head))) { 2075db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 20767ffddd49SMuchun Song cond_resched(); 20777ffddd49SMuchun Song 20787ffddd49SMuchun Song /* 20797ffddd49SMuchun Song * Theoretically, we should return -EBUSY when we 20807ffddd49SMuchun Song * encounter this race. In fact, we have a chance 20817ffddd49SMuchun Song * to successfully dissolve the page if we do a 20827ffddd49SMuchun Song * retry. Because the race window is quite small. 20837ffddd49SMuchun Song * If we seize this opportunity, it is an optimization 20847ffddd49SMuchun Song * for increasing the success rate of dissolving page. 20857ffddd49SMuchun Song */ 20867ffddd49SMuchun Song goto retry; 20877ffddd49SMuchun Song } 20887ffddd49SMuchun Song 2089ad2fa371SMuchun Song remove_hugetlb_page(h, head, false); 2090ad2fa371SMuchun Song h->max_huge_pages--; 2091ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2092ad2fa371SMuchun Song 2093c3114a84SAnshuman Khandual /* 2094ad2fa371SMuchun Song * Normally update_and_free_page will allocate required vmemmmap 2095ad2fa371SMuchun Song * before freeing the page. update_and_free_page will fail to 2096ad2fa371SMuchun Song * free the page if it can not allocate required vmemmap. We 2097ad2fa371SMuchun Song * need to adjust max_huge_pages if the page is not freed. 2098ad2fa371SMuchun Song * Attempt to allocate vmemmmap here so that we can take 2099ad2fa371SMuchun Song * appropriate action on failure. 2100ad2fa371SMuchun Song */ 2101ad2fa371SMuchun Song rc = alloc_huge_page_vmemmap(h, head); 2102ad2fa371SMuchun Song if (!rc) { 2103ad2fa371SMuchun Song /* 2104ad2fa371SMuchun Song * Move PageHWPoison flag from head page to the raw 2105ad2fa371SMuchun Song * error page, which makes any subpages rather than 2106ad2fa371SMuchun Song * the error page reusable. 2107c3114a84SAnshuman Khandual */ 2108c3114a84SAnshuman Khandual if (PageHWPoison(head) && page != head) { 2109c3114a84SAnshuman Khandual SetPageHWPoison(page); 2110c3114a84SAnshuman Khandual ClearPageHWPoison(head); 2111c3114a84SAnshuman Khandual } 2112b65d4adbSMuchun Song update_and_free_page(h, head, false); 2113ad2fa371SMuchun Song } else { 2114ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 2115ad2fa371SMuchun Song add_hugetlb_page(h, head, false); 2116ad2fa371SMuchun Song h->max_huge_pages++; 2117ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2118ad2fa371SMuchun Song } 2119ad2fa371SMuchun Song 2120ad2fa371SMuchun Song return rc; 2121c8721bbbSNaoya Horiguchi } 2122082d5b6bSGerald Schaefer out: 2123db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2124082d5b6bSGerald Schaefer return rc; 2125c8721bbbSNaoya Horiguchi } 2126c8721bbbSNaoya Horiguchi 2127c8721bbbSNaoya Horiguchi /* 2128c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 2129c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 21302247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 21312247bb33SGerald Schaefer * part of it lies within the given range. 2132082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 2133082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 2134c8721bbbSNaoya Horiguchi */ 2135082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 2136c8721bbbSNaoya Horiguchi { 2137c8721bbbSNaoya Horiguchi unsigned long pfn; 2138eb03aa00SGerald Schaefer struct page *page; 2139082d5b6bSGerald Schaefer int rc = 0; 2140c8721bbbSNaoya Horiguchi 2141d0177639SLi Zhong if (!hugepages_supported()) 2142082d5b6bSGerald Schaefer return rc; 2143d0177639SLi Zhong 2144eb03aa00SGerald Schaefer for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { 2145eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 2146eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 2147eb03aa00SGerald Schaefer if (rc) 2148082d5b6bSGerald Schaefer break; 2149eb03aa00SGerald Schaefer } 2150082d5b6bSGerald Schaefer 2151082d5b6bSGerald Schaefer return rc; 2152c8721bbbSNaoya Horiguchi } 2153c8721bbbSNaoya Horiguchi 2154ab5ac90aSMichal Hocko /* 2155ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 2156ab5ac90aSMichal Hocko */ 21570c397daeSMichal Hocko static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, 2158b65a4edaSMike Kravetz int nid, nodemask_t *nmask, bool zero_ref) 21597893d1d5SAdam Litke { 21609980d744SMichal Hocko struct page *page = NULL; 2161b65a4edaSMike Kravetz bool retry = false; 21627893d1d5SAdam Litke 2163bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2164aa888a74SAndi Kleen return NULL; 2165aa888a74SAndi Kleen 2166db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 21679980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 21689980d744SMichal Hocko goto out_unlock; 2169db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2170d1c3fb1fSNishanth Aravamudan 2171b65a4edaSMike Kravetz retry: 2172f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 21739980d744SMichal Hocko if (!page) 21740c397daeSMichal Hocko return NULL; 2175d1c3fb1fSNishanth Aravamudan 2176db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 21779980d744SMichal Hocko /* 21789980d744SMichal Hocko * We could have raced with the pool size change. 21799980d744SMichal Hocko * Double check that and simply deallocate the new page 21809980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 21819980d744SMichal Hocko * temporary page to workaround the nasty free_huge_page 21829980d744SMichal Hocko * codeflow 21839980d744SMichal Hocko */ 21849980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 21859157c311SMike Kravetz SetHPageTemporary(page); 2186db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 21879980d744SMichal Hocko put_page(page); 21882bf753e6SKai Shen return NULL; 2189b65a4edaSMike Kravetz } 2190b65a4edaSMike Kravetz 2191b65a4edaSMike Kravetz if (zero_ref) { 2192b65a4edaSMike Kravetz /* 2193b65a4edaSMike Kravetz * Caller requires a page with zero ref count. 2194b65a4edaSMike Kravetz * We will drop ref count here. If someone else is holding 2195b65a4edaSMike Kravetz * a ref, the page will be freed when they drop it. Abuse 2196b65a4edaSMike Kravetz * temporary page flag to accomplish this. 2197b65a4edaSMike Kravetz */ 2198b65a4edaSMike Kravetz SetHPageTemporary(page); 2199b65a4edaSMike Kravetz if (!put_page_testzero(page)) { 2200b65a4edaSMike Kravetz /* 2201b65a4edaSMike Kravetz * Unexpected inflated ref count on freshly allocated 2202b65a4edaSMike Kravetz * huge. Retry once. 2203b65a4edaSMike Kravetz */ 2204b65a4edaSMike Kravetz pr_info("HugeTLB unexpected inflated ref count on freshly allocated page\n"); 2205b65a4edaSMike Kravetz spin_unlock_irq(&hugetlb_lock); 2206b65a4edaSMike Kravetz if (retry) 2207b65a4edaSMike Kravetz return NULL; 2208b65a4edaSMike Kravetz 2209b65a4edaSMike Kravetz retry = true; 2210b65a4edaSMike Kravetz goto retry; 2211b65a4edaSMike Kravetz } 2212b65a4edaSMike Kravetz ClearHPageTemporary(page); 2213b65a4edaSMike Kravetz } 2214b65a4edaSMike Kravetz 22159980d744SMichal Hocko h->surplus_huge_pages++; 22164704dea3SMichal Hocko h->surplus_huge_pages_node[page_to_nid(page)]++; 22179980d744SMichal Hocko 22189980d744SMichal Hocko out_unlock: 2219db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 22207893d1d5SAdam Litke 22217893d1d5SAdam Litke return page; 22227893d1d5SAdam Litke } 22237893d1d5SAdam Litke 2224bbe88753SJoonsoo Kim static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, 2225ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 2226ab5ac90aSMichal Hocko { 2227ab5ac90aSMichal Hocko struct page *page; 2228ab5ac90aSMichal Hocko 2229ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 2230ab5ac90aSMichal Hocko return NULL; 2231ab5ac90aSMichal Hocko 2232f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 2233ab5ac90aSMichal Hocko if (!page) 2234ab5ac90aSMichal Hocko return NULL; 2235ab5ac90aSMichal Hocko 2236ab5ac90aSMichal Hocko /* 2237ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 2238ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 2239ab5ac90aSMichal Hocko */ 22409157c311SMike Kravetz SetHPageTemporary(page); 2241ab5ac90aSMichal Hocko 2242ab5ac90aSMichal Hocko return page; 2243ab5ac90aSMichal Hocko } 2244ab5ac90aSMichal Hocko 2245e4e574b7SAdam Litke /* 2246099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 2247099730d6SDave Hansen */ 2248e0ec90eeSDave Hansen static 22490c397daeSMichal Hocko struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, 2250099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 2251099730d6SDave Hansen { 2252cfcaa66fSBen Widawsky struct page *page = NULL; 2253aaf14e40SMichal Hocko struct mempolicy *mpol; 2254aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 2255aaf14e40SMichal Hocko int nid; 2256aaf14e40SMichal Hocko nodemask_t *nodemask; 2257aaf14e40SMichal Hocko 2258aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 2259cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 2260cfcaa66fSBen Widawsky gfp_t gfp = gfp_mask | __GFP_NOWARN; 2261cfcaa66fSBen Widawsky 2262cfcaa66fSBen Widawsky gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL); 2263cfcaa66fSBen Widawsky page = alloc_surplus_huge_page(h, gfp, nid, nodemask, false); 2264cfcaa66fSBen Widawsky 2265cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 2266cfcaa66fSBen Widawsky nodemask = NULL; 2267cfcaa66fSBen Widawsky } 2268cfcaa66fSBen Widawsky 2269cfcaa66fSBen Widawsky if (!page) 2270b65a4edaSMike Kravetz page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask, false); 2271aaf14e40SMichal Hocko mpol_cond_put(mpol); 2272aaf14e40SMichal Hocko return page; 2273099730d6SDave Hansen } 2274099730d6SDave Hansen 2275ab5ac90aSMichal Hocko /* page migration callback function */ 22763e59fcb0SMichal Hocko struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, 2277d92bbc27SJoonsoo Kim nodemask_t *nmask, gfp_t gfp_mask) 22784db9b2efSMichal Hocko { 2279db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 22804db9b2efSMichal Hocko if (h->free_huge_pages - h->resv_huge_pages > 0) { 22813e59fcb0SMichal Hocko struct page *page; 22823e59fcb0SMichal Hocko 22833e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); 22843e59fcb0SMichal Hocko if (page) { 2285db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 22863e59fcb0SMichal Hocko return page; 22874db9b2efSMichal Hocko } 22884db9b2efSMichal Hocko } 2289db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 22904db9b2efSMichal Hocko 22910c397daeSMichal Hocko return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); 22924db9b2efSMichal Hocko } 22934db9b2efSMichal Hocko 2294ebd63723SMichal Hocko /* mempolicy aware migration callback */ 2295389c8178SMichal Hocko struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, 2296389c8178SMichal Hocko unsigned long address) 2297ebd63723SMichal Hocko { 2298ebd63723SMichal Hocko struct mempolicy *mpol; 2299ebd63723SMichal Hocko nodemask_t *nodemask; 2300ebd63723SMichal Hocko struct page *page; 2301ebd63723SMichal Hocko gfp_t gfp_mask; 2302ebd63723SMichal Hocko int node; 2303ebd63723SMichal Hocko 2304ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 2305ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 2306d92bbc27SJoonsoo Kim page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); 2307ebd63723SMichal Hocko mpol_cond_put(mpol); 2308ebd63723SMichal Hocko 2309ebd63723SMichal Hocko return page; 2310ebd63723SMichal Hocko } 2311ebd63723SMichal Hocko 2312bf50bab2SNaoya Horiguchi /* 231325985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 2314e4e574b7SAdam Litke * of size 'delta'. 2315e4e574b7SAdam Litke */ 23160a4f3d1bSLiu Xiang static int gather_surplus_pages(struct hstate *h, long delta) 23171b2a1e7bSJules Irenge __must_hold(&hugetlb_lock) 2318e4e574b7SAdam Litke { 2319e4e574b7SAdam Litke struct list_head surplus_list; 2320e4e574b7SAdam Litke struct page *page, *tmp; 23210a4f3d1bSLiu Xiang int ret; 23220a4f3d1bSLiu Xiang long i; 23230a4f3d1bSLiu Xiang long needed, allocated; 232428073b02SHillf Danton bool alloc_ok = true; 2325e4e574b7SAdam Litke 23269487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2327a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 2328ac09b3a1SAdam Litke if (needed <= 0) { 2329a5516438SAndi Kleen h->resv_huge_pages += delta; 2330e4e574b7SAdam Litke return 0; 2331ac09b3a1SAdam Litke } 2332e4e574b7SAdam Litke 2333e4e574b7SAdam Litke allocated = 0; 2334e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 2335e4e574b7SAdam Litke 2336e4e574b7SAdam Litke ret = -ENOMEM; 2337e4e574b7SAdam Litke retry: 2338db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2339e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 23400c397daeSMichal Hocko page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), 2341b65a4edaSMike Kravetz NUMA_NO_NODE, NULL, true); 234228073b02SHillf Danton if (!page) { 234328073b02SHillf Danton alloc_ok = false; 234428073b02SHillf Danton break; 234528073b02SHillf Danton } 2346e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 234769ed779aSDavid Rientjes cond_resched(); 2348e4e574b7SAdam Litke } 234928073b02SHillf Danton allocated += i; 2350e4e574b7SAdam Litke 2351e4e574b7SAdam Litke /* 2352e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 2353e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 2354e4e574b7SAdam Litke */ 2355db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2356a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 2357a5516438SAndi Kleen (h->free_huge_pages + allocated); 235828073b02SHillf Danton if (needed > 0) { 235928073b02SHillf Danton if (alloc_ok) 2360e4e574b7SAdam Litke goto retry; 236128073b02SHillf Danton /* 236228073b02SHillf Danton * We were not able to allocate enough pages to 236328073b02SHillf Danton * satisfy the entire reservation so we free what 236428073b02SHillf Danton * we've allocated so far. 236528073b02SHillf Danton */ 236628073b02SHillf Danton goto free; 236728073b02SHillf Danton } 2368e4e574b7SAdam Litke /* 2369e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 237025985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 2371e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 2372ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 2373ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 2374ac09b3a1SAdam Litke * before they are reserved. 2375e4e574b7SAdam Litke */ 2376e4e574b7SAdam Litke needed += allocated; 2377a5516438SAndi Kleen h->resv_huge_pages += delta; 2378e4e574b7SAdam Litke ret = 0; 2379a9869b83SNaoya Horiguchi 238019fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 238119fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 238219fc3f0aSAdam Litke if ((--needed) < 0) 238319fc3f0aSAdam Litke break; 2384b65a4edaSMike Kravetz /* Add the page to the hugetlb allocator */ 2385a5516438SAndi Kleen enqueue_huge_page(h, page); 238619fc3f0aSAdam Litke } 238728073b02SHillf Danton free: 2388db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 238919fc3f0aSAdam Litke 2390b65a4edaSMike Kravetz /* 2391b65a4edaSMike Kravetz * Free unnecessary surplus pages to the buddy allocator. 2392b65a4edaSMike Kravetz * Pages have no ref count, call free_huge_page directly. 2393b65a4edaSMike Kravetz */ 2394c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 2395b65a4edaSMike Kravetz free_huge_page(page); 2396db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2397e4e574b7SAdam Litke 2398e4e574b7SAdam Litke return ret; 2399e4e574b7SAdam Litke } 2400e4e574b7SAdam Litke 2401e4e574b7SAdam Litke /* 2402e5bbc8a6SMike Kravetz * This routine has two main purposes: 2403e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 2404e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 2405e5bbc8a6SMike Kravetz * to the associated reservation map. 2406e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 2407e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 2408e4e574b7SAdam Litke */ 2409a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 2410a5516438SAndi Kleen unsigned long unused_resv_pages) 2411e4e574b7SAdam Litke { 2412e4e574b7SAdam Litke unsigned long nr_pages; 241310c6ec49SMike Kravetz struct page *page; 241410c6ec49SMike Kravetz LIST_HEAD(page_list); 241510c6ec49SMike Kravetz 24169487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 241710c6ec49SMike Kravetz /* Uncommit the reservation */ 241810c6ec49SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 2419e4e574b7SAdam Litke 2420aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 2421bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2422e5bbc8a6SMike Kravetz goto out; 2423aa888a74SAndi Kleen 2424e5bbc8a6SMike Kravetz /* 2425e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 2426e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 2427e5bbc8a6SMike Kravetz */ 2428a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 2429e4e574b7SAdam Litke 2430685f3457SLee Schermerhorn /* 2431685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 24329b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 24339b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 24349b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 243510c6ec49SMike Kravetz * remove_pool_huge_page() will balance the freed pages across the 24369b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 2437685f3457SLee Schermerhorn */ 2438685f3457SLee Schermerhorn while (nr_pages--) { 243910c6ec49SMike Kravetz page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1); 244010c6ec49SMike Kravetz if (!page) 2441e5bbc8a6SMike Kravetz goto out; 244210c6ec49SMike Kravetz 244310c6ec49SMike Kravetz list_add(&page->lru, &page_list); 2444e4e574b7SAdam Litke } 2445e5bbc8a6SMike Kravetz 2446e5bbc8a6SMike Kravetz out: 2447db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 244810c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 2449db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2450e4e574b7SAdam Litke } 2451e4e574b7SAdam Litke 24525e911373SMike Kravetz 2453c37f9fb1SAndy Whitcroft /* 2454feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 24555e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 2456cf3ad20bSMike Kravetz * 2457cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 2458cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 2459cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 2460cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 2461cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 2462feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 2463feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 2464feba16e2SMike Kravetz * is called in such cases. 2465cf3ad20bSMike Kravetz * 2466cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 2467cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 2468cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 2469cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 2470cf3ad20bSMike Kravetz * take appropriate action. 247196b96a96SMike Kravetz * 247296b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 247396b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 247496b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 247596b96a96SMike Kravetz * reservation exists. 2476846be085SMike Kravetz * 2477846be085SMike Kravetz * vma_del_reservation is used in error paths where an entry in the reserve 2478846be085SMike Kravetz * map was created during huge page allocation and must be removed. It is to 2479846be085SMike Kravetz * be called after calling vma_needs_reservation to determine if a reservation 2480846be085SMike Kravetz * exists. 2481c37f9fb1SAndy Whitcroft */ 24825e911373SMike Kravetz enum vma_resv_mode { 24835e911373SMike Kravetz VMA_NEEDS_RESV, 24845e911373SMike Kravetz VMA_COMMIT_RESV, 2485feba16e2SMike Kravetz VMA_END_RESV, 248696b96a96SMike Kravetz VMA_ADD_RESV, 2487846be085SMike Kravetz VMA_DEL_RESV, 24885e911373SMike Kravetz }; 2489cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 2490cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 24915e911373SMike Kravetz enum vma_resv_mode mode) 2492c37f9fb1SAndy Whitcroft { 24934e35f483SJoonsoo Kim struct resv_map *resv; 24944e35f483SJoonsoo Kim pgoff_t idx; 2495cf3ad20bSMike Kravetz long ret; 24960db9d74eSMina Almasry long dummy_out_regions_needed; 2497c37f9fb1SAndy Whitcroft 24984e35f483SJoonsoo Kim resv = vma_resv_map(vma); 24994e35f483SJoonsoo Kim if (!resv) 2500c37f9fb1SAndy Whitcroft return 1; 2501c37f9fb1SAndy Whitcroft 25024e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 25035e911373SMike Kravetz switch (mode) { 25045e911373SMike Kravetz case VMA_NEEDS_RESV: 25050db9d74eSMina Almasry ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); 25060db9d74eSMina Almasry /* We assume that vma_reservation_* routines always operate on 25070db9d74eSMina Almasry * 1 page, and that adding to resv map a 1 page entry can only 25080db9d74eSMina Almasry * ever require 1 region. 25090db9d74eSMina Almasry */ 25100db9d74eSMina Almasry VM_BUG_ON(dummy_out_regions_needed != 1); 25115e911373SMike Kravetz break; 25125e911373SMike Kravetz case VMA_COMMIT_RESV: 2513075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 25140db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 25150db9d74eSMina Almasry VM_BUG_ON(ret < 0); 25165e911373SMike Kravetz break; 2517feba16e2SMike Kravetz case VMA_END_RESV: 25180db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 25195e911373SMike Kravetz ret = 0; 25205e911373SMike Kravetz break; 252196b96a96SMike Kravetz case VMA_ADD_RESV: 25220db9d74eSMina Almasry if (vma->vm_flags & VM_MAYSHARE) { 2523075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 25240db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 25250db9d74eSMina Almasry VM_BUG_ON(ret < 0); 25260db9d74eSMina Almasry } else { 25270db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 252896b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 252996b96a96SMike Kravetz } 253096b96a96SMike Kravetz break; 2531846be085SMike Kravetz case VMA_DEL_RESV: 2532846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 2533846be085SMike Kravetz region_abort(resv, idx, idx + 1, 1); 2534846be085SMike Kravetz ret = region_del(resv, idx, idx + 1); 2535846be085SMike Kravetz } else { 2536846be085SMike Kravetz ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 2537846be085SMike Kravetz /* region_add calls of range 1 should never fail. */ 2538846be085SMike Kravetz VM_BUG_ON(ret < 0); 2539846be085SMike Kravetz } 2540846be085SMike Kravetz break; 25415e911373SMike Kravetz default: 25425e911373SMike Kravetz BUG(); 25435e911373SMike Kravetz } 254484afd99bSAndy Whitcroft 2545846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV) 2546cf3ad20bSMike Kravetz return ret; 254767961f9dSMike Kravetz /* 2548bf3d12b9SMiaohe Lin * We know private mapping must have HPAGE_RESV_OWNER set. 2549bf3d12b9SMiaohe Lin * 255067961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 255167961f9dSMike Kravetz * However, a file associated with mapping could have been 255267961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 255367961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 255467961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 255567961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 255667961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 255767961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 255867961f9dSMike Kravetz * reservation has already been consumed. As a result, the 255967961f9dSMike Kravetz * return value of this routine is the opposite of the 256067961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 256167961f9dSMike Kravetz */ 2562bf3d12b9SMiaohe Lin if (ret > 0) 256367961f9dSMike Kravetz return 0; 2564bf3d12b9SMiaohe Lin if (ret == 0) 256567961f9dSMike Kravetz return 1; 2566bf3d12b9SMiaohe Lin return ret; 256784afd99bSAndy Whitcroft } 2568cf3ad20bSMike Kravetz 2569cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 2570a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 2571c37f9fb1SAndy Whitcroft { 25725e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 2573cf3ad20bSMike Kravetz } 2574c37f9fb1SAndy Whitcroft 2575cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 2576cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 2577cf3ad20bSMike Kravetz { 25785e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 25795e911373SMike Kravetz } 25805e911373SMike Kravetz 2581feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 25825e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 25835e911373SMike Kravetz { 2584feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 2585c37f9fb1SAndy Whitcroft } 2586c37f9fb1SAndy Whitcroft 258796b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 258896b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 258996b96a96SMike Kravetz { 259096b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 259196b96a96SMike Kravetz } 259296b96a96SMike Kravetz 2593846be085SMike Kravetz static long vma_del_reservation(struct hstate *h, 2594846be085SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 259596b96a96SMike Kravetz { 2596846be085SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV); 2597846be085SMike Kravetz } 2598846be085SMike Kravetz 2599846be085SMike Kravetz /* 2600846be085SMike Kravetz * This routine is called to restore reservation information on error paths. 2601846be085SMike Kravetz * It should ONLY be called for pages allocated via alloc_huge_page(), and 2602846be085SMike Kravetz * the hugetlb mutex should remain held when calling this routine. 2603846be085SMike Kravetz * 2604846be085SMike Kravetz * It handles two specific cases: 2605846be085SMike Kravetz * 1) A reservation was in place and the page consumed the reservation. 2606846be085SMike Kravetz * HPageRestoreReserve is set in the page. 2607846be085SMike Kravetz * 2) No reservation was in place for the page, so HPageRestoreReserve is 2608846be085SMike Kravetz * not set. However, alloc_huge_page always updates the reserve map. 2609846be085SMike Kravetz * 2610846be085SMike Kravetz * In case 1, free_huge_page later in the error path will increment the 2611846be085SMike Kravetz * global reserve count. But, free_huge_page does not have enough context 2612846be085SMike Kravetz * to adjust the reservation map. This case deals primarily with private 2613846be085SMike Kravetz * mappings. Adjust the reserve map here to be consistent with global 2614846be085SMike Kravetz * reserve count adjustments to be made by free_huge_page. Make sure the 2615846be085SMike Kravetz * reserve map indicates there is a reservation present. 2616846be085SMike Kravetz * 2617846be085SMike Kravetz * In case 2, simply undo reserve map modifications done by alloc_huge_page. 2618846be085SMike Kravetz */ 2619846be085SMike Kravetz void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma, 2620846be085SMike Kravetz unsigned long address, struct page *page) 2621846be085SMike Kravetz { 262296b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 262396b96a96SMike Kravetz 2624846be085SMike Kravetz if (HPageRestoreReserve(page)) { 2625846be085SMike Kravetz if (unlikely(rc < 0)) 262696b96a96SMike Kravetz /* 262796b96a96SMike Kravetz * Rare out of memory condition in reserve map 2628d6995da3SMike Kravetz * manipulation. Clear HPageRestoreReserve so that 262996b96a96SMike Kravetz * global reserve count will not be incremented 263096b96a96SMike Kravetz * by free_huge_page. This will make it appear 263196b96a96SMike Kravetz * as though the reservation for this page was 263296b96a96SMike Kravetz * consumed. This may prevent the task from 263396b96a96SMike Kravetz * faulting in the page at a later time. This 263496b96a96SMike Kravetz * is better than inconsistent global huge page 263596b96a96SMike Kravetz * accounting of reserve counts. 263696b96a96SMike Kravetz */ 2637d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 2638846be085SMike Kravetz else if (rc) 2639846be085SMike Kravetz (void)vma_add_reservation(h, vma, address); 2640846be085SMike Kravetz else 2641846be085SMike Kravetz vma_end_reservation(h, vma, address); 2642846be085SMike Kravetz } else { 2643846be085SMike Kravetz if (!rc) { 264496b96a96SMike Kravetz /* 2645846be085SMike Kravetz * This indicates there is an entry in the reserve map 2646c7b1850dSMike Kravetz * not added by alloc_huge_page. We know it was added 2647846be085SMike Kravetz * before the alloc_huge_page call, otherwise 2648846be085SMike Kravetz * HPageRestoreReserve would be set on the page. 2649846be085SMike Kravetz * Remove the entry so that a subsequent allocation 2650846be085SMike Kravetz * does not consume a reservation. 265196b96a96SMike Kravetz */ 2652846be085SMike Kravetz rc = vma_del_reservation(h, vma, address); 2653846be085SMike Kravetz if (rc < 0) 2654846be085SMike Kravetz /* 2655846be085SMike Kravetz * VERY rare out of memory condition. Since 2656846be085SMike Kravetz * we can not delete the entry, set 2657846be085SMike Kravetz * HPageRestoreReserve so that the reserve 2658846be085SMike Kravetz * count will be incremented when the page 2659846be085SMike Kravetz * is freed. This reserve will be consumed 2660846be085SMike Kravetz * on a subsequent allocation. 2661846be085SMike Kravetz */ 2662846be085SMike Kravetz SetHPageRestoreReserve(page); 2663846be085SMike Kravetz } else if (rc < 0) { 2664846be085SMike Kravetz /* 2665846be085SMike Kravetz * Rare out of memory condition from 2666846be085SMike Kravetz * vma_needs_reservation call. Memory allocation is 2667846be085SMike Kravetz * only attempted if a new entry is needed. Therefore, 2668846be085SMike Kravetz * this implies there is not an entry in the 2669846be085SMike Kravetz * reserve map. 2670846be085SMike Kravetz * 2671846be085SMike Kravetz * For shared mappings, no entry in the map indicates 2672846be085SMike Kravetz * no reservation. We are done. 2673846be085SMike Kravetz */ 2674846be085SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 2675846be085SMike Kravetz /* 2676846be085SMike Kravetz * For private mappings, no entry indicates 2677846be085SMike Kravetz * a reservation is present. Since we can 2678846be085SMike Kravetz * not add an entry, set SetHPageRestoreReserve 2679846be085SMike Kravetz * on the page so reserve count will be 2680846be085SMike Kravetz * incremented when freed. This reserve will 2681846be085SMike Kravetz * be consumed on a subsequent allocation. 2682846be085SMike Kravetz */ 2683846be085SMike Kravetz SetHPageRestoreReserve(page); 268496b96a96SMike Kravetz } else 2685846be085SMike Kravetz /* 2686846be085SMike Kravetz * No reservation present, do nothing 2687846be085SMike Kravetz */ 268896b96a96SMike Kravetz vma_end_reservation(h, vma, address); 268996b96a96SMike Kravetz } 269096b96a96SMike Kravetz } 269196b96a96SMike Kravetz 2692369fa227SOscar Salvador /* 2693369fa227SOscar Salvador * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one 2694369fa227SOscar Salvador * @h: struct hstate old page belongs to 2695369fa227SOscar Salvador * @old_page: Old page to dissolve 2696ae37c7ffSOscar Salvador * @list: List to isolate the page in case we need to 2697369fa227SOscar Salvador * Returns 0 on success, otherwise negated error. 2698369fa227SOscar Salvador */ 2699ae37c7ffSOscar Salvador static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page, 2700ae37c7ffSOscar Salvador struct list_head *list) 2701369fa227SOscar Salvador { 2702369fa227SOscar Salvador gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2703369fa227SOscar Salvador int nid = page_to_nid(old_page); 2704b65a4edaSMike Kravetz bool alloc_retry = false; 2705369fa227SOscar Salvador struct page *new_page; 2706369fa227SOscar Salvador int ret = 0; 2707369fa227SOscar Salvador 2708369fa227SOscar Salvador /* 2709369fa227SOscar Salvador * Before dissolving the page, we need to allocate a new one for the 2710f41f2ed4SMuchun Song * pool to remain stable. Here, we allocate the page and 'prep' it 2711f41f2ed4SMuchun Song * by doing everything but actually updating counters and adding to 2712f41f2ed4SMuchun Song * the pool. This simplifies and let us do most of the processing 2713f41f2ed4SMuchun Song * under the lock. 2714369fa227SOscar Salvador */ 2715b65a4edaSMike Kravetz alloc_retry: 2716369fa227SOscar Salvador new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL); 2717369fa227SOscar Salvador if (!new_page) 2718369fa227SOscar Salvador return -ENOMEM; 2719b65a4edaSMike Kravetz /* 2720b65a4edaSMike Kravetz * If all goes well, this page will be directly added to the free 2721b65a4edaSMike Kravetz * list in the pool. For this the ref count needs to be zero. 2722b65a4edaSMike Kravetz * Attempt to drop now, and retry once if needed. It is VERY 2723b65a4edaSMike Kravetz * unlikely there is another ref on the page. 2724b65a4edaSMike Kravetz * 2725b65a4edaSMike Kravetz * If someone else has a reference to the page, it will be freed 2726b65a4edaSMike Kravetz * when they drop their ref. Abuse temporary page flag to accomplish 2727b65a4edaSMike Kravetz * this. Retry once if there is an inflated ref count. 2728b65a4edaSMike Kravetz */ 2729b65a4edaSMike Kravetz SetHPageTemporary(new_page); 2730b65a4edaSMike Kravetz if (!put_page_testzero(new_page)) { 2731b65a4edaSMike Kravetz if (alloc_retry) 2732b65a4edaSMike Kravetz return -EBUSY; 2733b65a4edaSMike Kravetz 2734b65a4edaSMike Kravetz alloc_retry = true; 2735b65a4edaSMike Kravetz goto alloc_retry; 2736b65a4edaSMike Kravetz } 2737b65a4edaSMike Kravetz ClearHPageTemporary(new_page); 2738b65a4edaSMike Kravetz 2739f41f2ed4SMuchun Song __prep_new_huge_page(h, new_page); 2740369fa227SOscar Salvador 2741369fa227SOscar Salvador retry: 2742369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2743369fa227SOscar Salvador if (!PageHuge(old_page)) { 2744369fa227SOscar Salvador /* 2745369fa227SOscar Salvador * Freed from under us. Drop new_page too. 2746369fa227SOscar Salvador */ 2747369fa227SOscar Salvador goto free_new; 2748369fa227SOscar Salvador } else if (page_count(old_page)) { 2749369fa227SOscar Salvador /* 2750ae37c7ffSOscar Salvador * Someone has grabbed the page, try to isolate it here. 2751ae37c7ffSOscar Salvador * Fail with -EBUSY if not possible. 2752369fa227SOscar Salvador */ 2753ae37c7ffSOscar Salvador spin_unlock_irq(&hugetlb_lock); 2754ae37c7ffSOscar Salvador if (!isolate_huge_page(old_page, list)) 2755369fa227SOscar Salvador ret = -EBUSY; 2756ae37c7ffSOscar Salvador spin_lock_irq(&hugetlb_lock); 2757369fa227SOscar Salvador goto free_new; 2758369fa227SOscar Salvador } else if (!HPageFreed(old_page)) { 2759369fa227SOscar Salvador /* 2760369fa227SOscar Salvador * Page's refcount is 0 but it has not been enqueued in the 2761369fa227SOscar Salvador * freelist yet. Race window is small, so we can succeed here if 2762369fa227SOscar Salvador * we retry. 2763369fa227SOscar Salvador */ 2764369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2765369fa227SOscar Salvador cond_resched(); 2766369fa227SOscar Salvador goto retry; 2767369fa227SOscar Salvador } else { 2768369fa227SOscar Salvador /* 2769369fa227SOscar Salvador * Ok, old_page is still a genuine free hugepage. Remove it from 2770369fa227SOscar Salvador * the freelist and decrease the counters. These will be 2771369fa227SOscar Salvador * incremented again when calling __prep_account_new_huge_page() 2772369fa227SOscar Salvador * and enqueue_huge_page() for new_page. The counters will remain 2773369fa227SOscar Salvador * stable since this happens under the lock. 2774369fa227SOscar Salvador */ 2775369fa227SOscar Salvador remove_hugetlb_page(h, old_page, false); 2776369fa227SOscar Salvador 2777369fa227SOscar Salvador /* 2778b65a4edaSMike Kravetz * Ref count on new page is already zero as it was dropped 2779b65a4edaSMike Kravetz * earlier. It can be directly added to the pool free list. 2780369fa227SOscar Salvador */ 2781369fa227SOscar Salvador __prep_account_new_huge_page(h, nid); 2782369fa227SOscar Salvador enqueue_huge_page(h, new_page); 2783369fa227SOscar Salvador 2784369fa227SOscar Salvador /* 2785369fa227SOscar Salvador * Pages have been replaced, we can safely free the old one. 2786369fa227SOscar Salvador */ 2787369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2788b65d4adbSMuchun Song update_and_free_page(h, old_page, false); 2789369fa227SOscar Salvador } 2790369fa227SOscar Salvador 2791369fa227SOscar Salvador return ret; 2792369fa227SOscar Salvador 2793369fa227SOscar Salvador free_new: 2794369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2795b65a4edaSMike Kravetz /* Page has a zero ref count, but needs a ref to be freed */ 2796b65a4edaSMike Kravetz set_page_refcounted(new_page); 2797b65d4adbSMuchun Song update_and_free_page(h, new_page, false); 2798369fa227SOscar Salvador 2799369fa227SOscar Salvador return ret; 2800369fa227SOscar Salvador } 2801369fa227SOscar Salvador 2802ae37c7ffSOscar Salvador int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list) 2803369fa227SOscar Salvador { 2804369fa227SOscar Salvador struct hstate *h; 2805369fa227SOscar Salvador struct page *head; 2806ae37c7ffSOscar Salvador int ret = -EBUSY; 2807369fa227SOscar Salvador 2808369fa227SOscar Salvador /* 2809369fa227SOscar Salvador * The page might have been dissolved from under our feet, so make sure 2810369fa227SOscar Salvador * to carefully check the state under the lock. 2811369fa227SOscar Salvador * Return success when racing as if we dissolved the page ourselves. 2812369fa227SOscar Salvador */ 2813369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2814369fa227SOscar Salvador if (PageHuge(page)) { 2815369fa227SOscar Salvador head = compound_head(page); 2816369fa227SOscar Salvador h = page_hstate(head); 2817369fa227SOscar Salvador } else { 2818369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2819369fa227SOscar Salvador return 0; 2820369fa227SOscar Salvador } 2821369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2822369fa227SOscar Salvador 2823369fa227SOscar Salvador /* 2824369fa227SOscar Salvador * Fence off gigantic pages as there is a cyclic dependency between 2825369fa227SOscar Salvador * alloc_contig_range and them. Return -ENOMEM as this has the effect 2826369fa227SOscar Salvador * of bailing out right away without further retrying. 2827369fa227SOscar Salvador */ 2828369fa227SOscar Salvador if (hstate_is_gigantic(h)) 2829369fa227SOscar Salvador return -ENOMEM; 2830369fa227SOscar Salvador 2831ae37c7ffSOscar Salvador if (page_count(head) && isolate_huge_page(head, list)) 2832ae37c7ffSOscar Salvador ret = 0; 2833ae37c7ffSOscar Salvador else if (!page_count(head)) 2834ae37c7ffSOscar Salvador ret = alloc_and_dissolve_huge_page(h, head, list); 2835ae37c7ffSOscar Salvador 2836ae37c7ffSOscar Salvador return ret; 2837369fa227SOscar Salvador } 2838369fa227SOscar Salvador 283970c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 284004f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 2841348ea204SAdam Litke { 284290481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 2843a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2844348ea204SAdam Litke struct page *page; 2845d85f69b0SMike Kravetz long map_chg, map_commit; 2846d85f69b0SMike Kravetz long gbl_chg; 28476d76dcf4SAneesh Kumar K.V int ret, idx; 28486d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 284908cf9fafSMina Almasry bool deferred_reserve; 28502fc39cecSAdam Litke 28516d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 2852a1e78772SMel Gorman /* 2853d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 2854d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 2855d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 2856a1e78772SMel Gorman */ 2857d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 2858d85f69b0SMike Kravetz if (map_chg < 0) 285976dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 2860d85f69b0SMike Kravetz 2861d85f69b0SMike Kravetz /* 2862d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 2863d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 2864d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 2865d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 2866d85f69b0SMike Kravetz * checked against any subpool limit. 2867d85f69b0SMike Kravetz */ 2868d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 2869d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 2870d85f69b0SMike Kravetz if (gbl_chg < 0) { 2871feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 287276dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 28735e911373SMike Kravetz } 287490d8b7e6SAdam Litke 2875d85f69b0SMike Kravetz /* 2876d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 2877d85f69b0SMike Kravetz * map, there could be reservations associated with the 2878d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 2879d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 2880d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 2881d85f69b0SMike Kravetz * the subpool reservations. 2882d85f69b0SMike Kravetz */ 2883d85f69b0SMike Kravetz if (avoid_reserve) 2884d85f69b0SMike Kravetz gbl_chg = 1; 2885d85f69b0SMike Kravetz } 2886d85f69b0SMike Kravetz 288708cf9fafSMina Almasry /* If this allocation is not consuming a reservation, charge it now. 288808cf9fafSMina Almasry */ 28896501fe5fSMiaohe Lin deferred_reserve = map_chg || avoid_reserve; 289008cf9fafSMina Almasry if (deferred_reserve) { 289108cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 289208cf9fafSMina Almasry idx, pages_per_huge_page(h), &h_cg); 28938f34af6fSJianyu Zhan if (ret) 28948f34af6fSJianyu Zhan goto out_subpool_put; 289508cf9fafSMina Almasry } 289608cf9fafSMina Almasry 289708cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 289808cf9fafSMina Almasry if (ret) 289908cf9fafSMina Almasry goto out_uncharge_cgroup_reservation; 29008f34af6fSJianyu Zhan 2901db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2902d85f69b0SMike Kravetz /* 2903d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 2904d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 2905d85f69b0SMike Kravetz * a reservation exists for the allocation. 2906d85f69b0SMike Kravetz */ 2907d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 290881a6fcaeSJoonsoo Kim if (!page) { 2909db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 29100c397daeSMichal Hocko page = alloc_buddy_huge_page_with_mpol(h, vma, addr); 29118f34af6fSJianyu Zhan if (!page) 29128f34af6fSJianyu Zhan goto out_uncharge_cgroup; 2913a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 2914d6995da3SMike Kravetz SetHPageRestoreReserve(page); 2915a88c7695SNaoya Horiguchi h->resv_huge_pages--; 2916a88c7695SNaoya Horiguchi } 2917db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 291815a8d68eSWei Yang list_add(&page->lru, &h->hugepage_activelist); 291981a6fcaeSJoonsoo Kim /* Fall through */ 2920a1e78772SMel Gorman } 292181a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 292208cf9fafSMina Almasry /* If allocation is not consuming a reservation, also store the 292308cf9fafSMina Almasry * hugetlb_cgroup pointer on the page. 292408cf9fafSMina Almasry */ 292508cf9fafSMina Almasry if (deferred_reserve) { 292608cf9fafSMina Almasry hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), 292708cf9fafSMina Almasry h_cg, page); 292808cf9fafSMina Almasry } 292908cf9fafSMina Almasry 2930db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2931a1e78772SMel Gorman 2932d6995da3SMike Kravetz hugetlb_set_page_subpool(page, spool); 2933a1e78772SMel Gorman 2934d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 2935d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 293633039678SMike Kravetz /* 293733039678SMike Kravetz * The page was added to the reservation map between 293833039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 293933039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 294033039678SMike Kravetz * Adjust for the subpool count incremented above AND 294133039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 294233039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 294333039678SMike Kravetz * no longer applies. 294433039678SMike Kravetz */ 294533039678SMike Kravetz long rsv_adjust; 294633039678SMike Kravetz 294733039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 294833039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 294979aa925bSMike Kravetz if (deferred_reserve) 295079aa925bSMike Kravetz hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 295179aa925bSMike Kravetz pages_per_huge_page(h), page); 295233039678SMike Kravetz } 29537893d1d5SAdam Litke return page; 29548f34af6fSJianyu Zhan 29558f34af6fSJianyu Zhan out_uncharge_cgroup: 29568f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 295708cf9fafSMina Almasry out_uncharge_cgroup_reservation: 295808cf9fafSMina Almasry if (deferred_reserve) 295908cf9fafSMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), 296008cf9fafSMina Almasry h_cg); 29618f34af6fSJianyu Zhan out_subpool_put: 2962d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 29638f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 2964feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 29658f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 2966b45b5bd6SDavid Gibson } 2967b45b5bd6SDavid Gibson 2968e24a1307SAneesh Kumar K.V int alloc_bootmem_huge_page(struct hstate *h) 2969e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 2970e24a1307SAneesh Kumar K.V int __alloc_bootmem_huge_page(struct hstate *h) 2971aa888a74SAndi Kleen { 2972aa888a74SAndi Kleen struct huge_bootmem_page *m; 2973b2261026SJoonsoo Kim int nr_nodes, node; 2974aa888a74SAndi Kleen 2975b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 2976aa888a74SAndi Kleen void *addr; 2977aa888a74SAndi Kleen 2978eb31d559SMike Rapoport addr = memblock_alloc_try_nid_raw( 29798b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 298097ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 2981aa888a74SAndi Kleen if (addr) { 2982aa888a74SAndi Kleen /* 2983aa888a74SAndi Kleen * Use the beginning of the huge page to store the 2984aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 2985aa888a74SAndi Kleen * puts them into the mem_map). 2986aa888a74SAndi Kleen */ 2987aa888a74SAndi Kleen m = addr; 2988aa888a74SAndi Kleen goto found; 2989aa888a74SAndi Kleen } 2990aa888a74SAndi Kleen } 2991aa888a74SAndi Kleen return 0; 2992aa888a74SAndi Kleen 2993aa888a74SAndi Kleen found: 2994df994eadSLuiz Capitulino BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); 2995aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 2996330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 2997aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 2998aa888a74SAndi Kleen m->hstate = h; 2999aa888a74SAndi Kleen return 1; 3000aa888a74SAndi Kleen } 3001aa888a74SAndi Kleen 300248b8d744SMike Kravetz /* 300348b8d744SMike Kravetz * Put bootmem huge pages into the standard lists after mem_map is up. 300448b8d744SMike Kravetz * Note: This only applies to gigantic (order > MAX_ORDER) pages. 300548b8d744SMike Kravetz */ 3006aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 3007aa888a74SAndi Kleen { 3008aa888a74SAndi Kleen struct huge_bootmem_page *m; 3009aa888a74SAndi Kleen 3010aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 301140d18ebfSMike Kravetz struct page *page = virt_to_page(m); 3012aa888a74SAndi Kleen struct hstate *h = m->hstate; 3013ee8f248dSBecky Bruce 301448b8d744SMike Kravetz VM_BUG_ON(!hstate_is_gigantic(h)); 3015aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 30167118fc29SMike Kravetz if (prep_compound_gigantic_page(page, huge_page_order(h))) { 3017ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 3018aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 30197118fc29SMike Kravetz put_page(page); /* add to the hugepage allocator */ 30207118fc29SMike Kravetz } else { 3021416d85edSMike Kravetz /* VERY unlikely inflated ref count on a tail page */ 30227118fc29SMike Kravetz free_gigantic_page(page, huge_page_order(h)); 30237118fc29SMike Kravetz } 3024af0fb9dfSMichal Hocko 3025b0320c7bSRafael Aquini /* 302648b8d744SMike Kravetz * We need to restore the 'stolen' pages to totalram_pages 302748b8d744SMike Kravetz * in order to fix confusing memory reports from free(1) and 302848b8d744SMike Kravetz * other side-effects, like CommitLimit going negative. 3029b0320c7bSRafael Aquini */ 3030c78a7f36SMiaohe Lin adjust_managed_page_count(page, pages_per_huge_page(h)); 3031520495feSCannon Matthews cond_resched(); 3032aa888a74SAndi Kleen } 3033aa888a74SAndi Kleen } 3034aa888a74SAndi Kleen 30358faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 30361da177e4SLinus Torvalds { 30371da177e4SLinus Torvalds unsigned long i; 3038f60858f9SMike Kravetz nodemask_t *node_alloc_noretry; 3039f60858f9SMike Kravetz 3040f60858f9SMike Kravetz if (!hstate_is_gigantic(h)) { 3041f60858f9SMike Kravetz /* 3042f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3043f60858f9SMike Kravetz * Ignore errors as lower level routines can deal with 3044f60858f9SMike Kravetz * node_alloc_noretry == NULL. If this kmalloc fails at boot 3045f60858f9SMike Kravetz * time, we are likely in bigger trouble. 3046f60858f9SMike Kravetz */ 3047f60858f9SMike Kravetz node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), 3048f60858f9SMike Kravetz GFP_KERNEL); 3049f60858f9SMike Kravetz } else { 3050f60858f9SMike Kravetz /* allocations done at boot time */ 3051f60858f9SMike Kravetz node_alloc_noretry = NULL; 3052f60858f9SMike Kravetz } 3053f60858f9SMike Kravetz 3054f60858f9SMike Kravetz /* bit mask controlling how hard we retry per-node allocations */ 3055f60858f9SMike Kravetz if (node_alloc_noretry) 3056f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 30571da177e4SLinus Torvalds 3058e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 3059bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 3060dbda8feaSBarry Song if (hugetlb_cma_size) { 3061cf11e85fSRoman Gushchin pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); 30627ecc9565SChen Wandun goto free; 3063cf11e85fSRoman Gushchin } 3064aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 3065aa888a74SAndi Kleen break; 30660c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 3067f60858f9SMike Kravetz &node_states[N_MEMORY], 3068f60858f9SMike Kravetz node_alloc_noretry)) 30691da177e4SLinus Torvalds break; 307069ed779aSDavid Rientjes cond_resched(); 30711da177e4SLinus Torvalds } 3072d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 3073d715cf80SLiam R. Howlett char buf[32]; 3074d715cf80SLiam R. Howlett 3075c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3076d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 3077d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 30788faa8b07SAndi Kleen h->max_huge_pages = i; 3079e5ff2159SAndi Kleen } 30807ecc9565SChen Wandun free: 3081f60858f9SMike Kravetz kfree(node_alloc_noretry); 3082d715cf80SLiam R. Howlett } 3083e5ff2159SAndi Kleen 3084e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 3085e5ff2159SAndi Kleen { 308679dfc695SMike Kravetz struct hstate *h, *h2; 3087e5ff2159SAndi Kleen 3088e5ff2159SAndi Kleen for_each_hstate(h) { 3089641844f5SNaoya Horiguchi if (minimum_order > huge_page_order(h)) 3090641844f5SNaoya Horiguchi minimum_order = huge_page_order(h); 3091641844f5SNaoya Horiguchi 30928faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 3093bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 30948faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 309579dfc695SMike Kravetz 309679dfc695SMike Kravetz /* 309779dfc695SMike Kravetz * Set demote order for each hstate. Note that 309879dfc695SMike Kravetz * h->demote_order is initially 0. 309979dfc695SMike Kravetz * - We can not demote gigantic pages if runtime freeing 310079dfc695SMike Kravetz * is not supported, so skip this. 3101a01f4390SMike Kravetz * - If CMA allocation is possible, we can not demote 3102a01f4390SMike Kravetz * HUGETLB_PAGE_ORDER or smaller size pages. 310379dfc695SMike Kravetz */ 310479dfc695SMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 310579dfc695SMike Kravetz continue; 3106a01f4390SMike Kravetz if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER) 3107a01f4390SMike Kravetz continue; 310879dfc695SMike Kravetz for_each_hstate(h2) { 310979dfc695SMike Kravetz if (h2 == h) 311079dfc695SMike Kravetz continue; 311179dfc695SMike Kravetz if (h2->order < h->order && 311279dfc695SMike Kravetz h2->order > h->demote_order) 311379dfc695SMike Kravetz h->demote_order = h2->order; 311479dfc695SMike Kravetz } 3115e5ff2159SAndi Kleen } 3116641844f5SNaoya Horiguchi VM_BUG_ON(minimum_order == UINT_MAX); 3117e5ff2159SAndi Kleen } 3118e5ff2159SAndi Kleen 3119e5ff2159SAndi Kleen static void __init report_hugepages(void) 3120e5ff2159SAndi Kleen { 3121e5ff2159SAndi Kleen struct hstate *h; 3122e5ff2159SAndi Kleen 3123e5ff2159SAndi Kleen for_each_hstate(h) { 31244abd32dbSAndi Kleen char buf[32]; 3125c6247f72SMatthew Wilcox 3126c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3127ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 3128c6247f72SMatthew Wilcox buf, h->free_huge_pages); 3129e5ff2159SAndi Kleen } 3130e5ff2159SAndi Kleen } 3131e5ff2159SAndi Kleen 31321da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 31336ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 31346ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 31351da177e4SLinus Torvalds { 31364415cc8dSChristoph Lameter int i; 31371121828aSMike Kravetz LIST_HEAD(page_list); 31384415cc8dSChristoph Lameter 31399487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 3140bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3141aa888a74SAndi Kleen return; 3142aa888a74SAndi Kleen 31431121828aSMike Kravetz /* 31441121828aSMike Kravetz * Collect pages to be freed on a list, and free after dropping lock 31451121828aSMike Kravetz */ 31466ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 314710c6ec49SMike Kravetz struct page *page, *next; 3148a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 3149a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 3150a5516438SAndi Kleen if (count >= h->nr_huge_pages) 31511121828aSMike Kravetz goto out; 31521da177e4SLinus Torvalds if (PageHighMem(page)) 31531da177e4SLinus Torvalds continue; 31546eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 31551121828aSMike Kravetz list_add(&page->lru, &page_list); 31561121828aSMike Kravetz } 31571121828aSMike Kravetz } 31581121828aSMike Kravetz 31591121828aSMike Kravetz out: 3160db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 316110c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3162db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 31631da177e4SLinus Torvalds } 31641da177e4SLinus Torvalds #else 31656ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 31666ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 31671da177e4SLinus Torvalds { 31681da177e4SLinus Torvalds } 31691da177e4SLinus Torvalds #endif 31701da177e4SLinus Torvalds 317120a0307cSWu Fengguang /* 317220a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 317320a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 317420a0307cSWu Fengguang * Returns 1 if an adjustment was made. 317520a0307cSWu Fengguang */ 31766ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 31776ae11b27SLee Schermerhorn int delta) 317820a0307cSWu Fengguang { 3179b2261026SJoonsoo Kim int nr_nodes, node; 318020a0307cSWu Fengguang 31819487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 318220a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 318320a0307cSWu Fengguang 3184e8c5c824SLee Schermerhorn if (delta < 0) { 3185b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 3186b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 3187b2261026SJoonsoo Kim goto found; 3188b2261026SJoonsoo Kim } 3189b2261026SJoonsoo Kim } else { 3190b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 3191b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 3192b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 3193b2261026SJoonsoo Kim goto found; 3194e8c5c824SLee Schermerhorn } 31959a76db09SLee Schermerhorn } 3196b2261026SJoonsoo Kim return 0; 319720a0307cSWu Fengguang 3198b2261026SJoonsoo Kim found: 319920a0307cSWu Fengguang h->surplus_huge_pages += delta; 3200b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 3201b2261026SJoonsoo Kim return 1; 320220a0307cSWu Fengguang } 320320a0307cSWu Fengguang 3204a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 3205fd875dcaSMike Kravetz static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 32066ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 32071da177e4SLinus Torvalds { 32087893d1d5SAdam Litke unsigned long min_count, ret; 320910c6ec49SMike Kravetz struct page *page; 321010c6ec49SMike Kravetz LIST_HEAD(page_list); 3211f60858f9SMike Kravetz NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); 3212f60858f9SMike Kravetz 3213f60858f9SMike Kravetz /* 3214f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3215f60858f9SMike Kravetz * If we can not allocate the bit mask, do not attempt to allocate 3216f60858f9SMike Kravetz * the requested huge pages. 3217f60858f9SMike Kravetz */ 3218f60858f9SMike Kravetz if (node_alloc_noretry) 3219f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 3220f60858f9SMike Kravetz else 3221f60858f9SMike Kravetz return -ENOMEM; 32221da177e4SLinus Torvalds 322329383967SMike Kravetz /* 322429383967SMike Kravetz * resize_lock mutex prevents concurrent adjustments to number of 322529383967SMike Kravetz * pages in hstate via the proc/sysfs interfaces. 322629383967SMike Kravetz */ 322729383967SMike Kravetz mutex_lock(&h->resize_lock); 3228b65d4adbSMuchun Song flush_free_hpage_work(h); 3229db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 32304eb0716eSAlexandre Ghiti 32314eb0716eSAlexandre Ghiti /* 3232fd875dcaSMike Kravetz * Check for a node specific request. 3233fd875dcaSMike Kravetz * Changing node specific huge page count may require a corresponding 3234fd875dcaSMike Kravetz * change to the global count. In any case, the passed node mask 3235fd875dcaSMike Kravetz * (nodes_allowed) will restrict alloc/free to the specified node. 3236fd875dcaSMike Kravetz */ 3237fd875dcaSMike Kravetz if (nid != NUMA_NO_NODE) { 3238fd875dcaSMike Kravetz unsigned long old_count = count; 3239fd875dcaSMike Kravetz 3240fd875dcaSMike Kravetz count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 3241fd875dcaSMike Kravetz /* 3242fd875dcaSMike Kravetz * User may have specified a large count value which caused the 3243fd875dcaSMike Kravetz * above calculation to overflow. In this case, they wanted 3244fd875dcaSMike Kravetz * to allocate as many huge pages as possible. Set count to 3245fd875dcaSMike Kravetz * largest possible value to align with their intention. 3246fd875dcaSMike Kravetz */ 3247fd875dcaSMike Kravetz if (count < old_count) 3248fd875dcaSMike Kravetz count = ULONG_MAX; 3249fd875dcaSMike Kravetz } 3250fd875dcaSMike Kravetz 3251fd875dcaSMike Kravetz /* 32524eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 32534eb0716eSAlexandre Ghiti * page range allocation. 32544eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 32554eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 32564eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 32574eb0716eSAlexandre Ghiti */ 32584eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 32594eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 3260db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 326129383967SMike Kravetz mutex_unlock(&h->resize_lock); 3262f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 32634eb0716eSAlexandre Ghiti return -EINVAL; 32644eb0716eSAlexandre Ghiti } 32654eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 32664eb0716eSAlexandre Ghiti } 3267aa888a74SAndi Kleen 32687893d1d5SAdam Litke /* 32697893d1d5SAdam Litke * Increase the pool size 32707893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 32717893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 3272d1c3fb1fSNishanth Aravamudan * 32730c397daeSMichal Hocko * We might race with alloc_surplus_huge_page() here and be unable 3274d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 3275d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 3276d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 3277d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 32787893d1d5SAdam Litke */ 3279a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 32806ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 32817893d1d5SAdam Litke break; 32827893d1d5SAdam Litke } 32837893d1d5SAdam Litke 3284a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 32857893d1d5SAdam Litke /* 32867893d1d5SAdam Litke * If this allocation races such that we no longer need the 32877893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 32887893d1d5SAdam Litke * and reducing the surplus. 32897893d1d5SAdam Litke */ 3290db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3291649920c6SJia He 3292649920c6SJia He /* yield cpu to avoid soft lockup */ 3293649920c6SJia He cond_resched(); 3294649920c6SJia He 3295f60858f9SMike Kravetz ret = alloc_pool_huge_page(h, nodes_allowed, 3296f60858f9SMike Kravetz node_alloc_noretry); 3297db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 32987893d1d5SAdam Litke if (!ret) 32997893d1d5SAdam Litke goto out; 33007893d1d5SAdam Litke 3301536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 3302536240f2SMel Gorman if (signal_pending(current)) 3303536240f2SMel Gorman goto out; 33047893d1d5SAdam Litke } 33057893d1d5SAdam Litke 33067893d1d5SAdam Litke /* 33077893d1d5SAdam Litke * Decrease the pool size 33087893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 33097893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 33107893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 33117893d1d5SAdam Litke * to the desired size as pages become free. 3312d1c3fb1fSNishanth Aravamudan * 3313d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 3314d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 3315d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 33160c397daeSMichal Hocko * alloc_surplus_huge_page() is checking the global counter, 3317d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 3318d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 3319d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 33207893d1d5SAdam Litke */ 3321a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 33226b0c880dSAdam Litke min_count = max(count, min_count); 33236ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 332410c6ec49SMike Kravetz 332510c6ec49SMike Kravetz /* 332610c6ec49SMike Kravetz * Collect pages to be removed on list without dropping lock 332710c6ec49SMike Kravetz */ 3328a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 332910c6ec49SMike Kravetz page = remove_pool_huge_page(h, nodes_allowed, 0); 333010c6ec49SMike Kravetz if (!page) 33311da177e4SLinus Torvalds break; 333210c6ec49SMike Kravetz 333310c6ec49SMike Kravetz list_add(&page->lru, &page_list); 33341da177e4SLinus Torvalds } 333510c6ec49SMike Kravetz /* free the pages after dropping lock */ 3336db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 333710c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3338b65d4adbSMuchun Song flush_free_hpage_work(h); 3339db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 334010c6ec49SMike Kravetz 3341a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 33426ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 33437893d1d5SAdam Litke break; 33447893d1d5SAdam Litke } 33457893d1d5SAdam Litke out: 33464eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 3347db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 334829383967SMike Kravetz mutex_unlock(&h->resize_lock); 33494eb0716eSAlexandre Ghiti 3350f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 3351f60858f9SMike Kravetz 33524eb0716eSAlexandre Ghiti return 0; 33531da177e4SLinus Torvalds } 33541da177e4SLinus Torvalds 33558531fc6fSMike Kravetz static int demote_free_huge_page(struct hstate *h, struct page *page) 33568531fc6fSMike Kravetz { 33578531fc6fSMike Kravetz int i, nid = page_to_nid(page); 33588531fc6fSMike Kravetz struct hstate *target_hstate; 33598531fc6fSMike Kravetz int rc = 0; 33608531fc6fSMike Kravetz 33618531fc6fSMike Kravetz target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order); 33628531fc6fSMike Kravetz 33638531fc6fSMike Kravetz remove_hugetlb_page_for_demote(h, page, false); 33648531fc6fSMike Kravetz spin_unlock_irq(&hugetlb_lock); 33658531fc6fSMike Kravetz 33668531fc6fSMike Kravetz rc = alloc_huge_page_vmemmap(h, page); 33678531fc6fSMike Kravetz if (rc) { 33688531fc6fSMike Kravetz /* Allocation of vmemmmap failed, we can not demote page */ 33698531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 33708531fc6fSMike Kravetz set_page_refcounted(page); 33718531fc6fSMike Kravetz add_hugetlb_page(h, page, false); 33728531fc6fSMike Kravetz return rc; 33738531fc6fSMike Kravetz } 33748531fc6fSMike Kravetz 33758531fc6fSMike Kravetz /* 33768531fc6fSMike Kravetz * Use destroy_compound_hugetlb_page_for_demote for all huge page 33778531fc6fSMike Kravetz * sizes as it will not ref count pages. 33788531fc6fSMike Kravetz */ 33798531fc6fSMike Kravetz destroy_compound_hugetlb_page_for_demote(page, huge_page_order(h)); 33808531fc6fSMike Kravetz 33818531fc6fSMike Kravetz /* 33828531fc6fSMike Kravetz * Taking target hstate mutex synchronizes with set_max_huge_pages. 33838531fc6fSMike Kravetz * Without the mutex, pages added to target hstate could be marked 33848531fc6fSMike Kravetz * as surplus. 33858531fc6fSMike Kravetz * 33868531fc6fSMike Kravetz * Note that we already hold h->resize_lock. To prevent deadlock, 33878531fc6fSMike Kravetz * use the convention of always taking larger size hstate mutex first. 33888531fc6fSMike Kravetz */ 33898531fc6fSMike Kravetz mutex_lock(&target_hstate->resize_lock); 33908531fc6fSMike Kravetz for (i = 0; i < pages_per_huge_page(h); 33918531fc6fSMike Kravetz i += pages_per_huge_page(target_hstate)) { 33928531fc6fSMike Kravetz if (hstate_is_gigantic(target_hstate)) 33938531fc6fSMike Kravetz prep_compound_gigantic_page_for_demote(page + i, 33948531fc6fSMike Kravetz target_hstate->order); 33958531fc6fSMike Kravetz else 33968531fc6fSMike Kravetz prep_compound_page(page + i, target_hstate->order); 33978531fc6fSMike Kravetz set_page_private(page + i, 0); 33988531fc6fSMike Kravetz set_page_refcounted(page + i); 33998531fc6fSMike Kravetz prep_new_huge_page(target_hstate, page + i, nid); 34008531fc6fSMike Kravetz put_page(page + i); 34018531fc6fSMike Kravetz } 34028531fc6fSMike Kravetz mutex_unlock(&target_hstate->resize_lock); 34038531fc6fSMike Kravetz 34048531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 34058531fc6fSMike Kravetz 34068531fc6fSMike Kravetz /* 34078531fc6fSMike Kravetz * Not absolutely necessary, but for consistency update max_huge_pages 34088531fc6fSMike Kravetz * based on pool changes for the demoted page. 34098531fc6fSMike Kravetz */ 34108531fc6fSMike Kravetz h->max_huge_pages--; 34118531fc6fSMike Kravetz target_hstate->max_huge_pages += pages_per_huge_page(h); 34128531fc6fSMike Kravetz 34138531fc6fSMike Kravetz return rc; 34148531fc6fSMike Kravetz } 34158531fc6fSMike Kravetz 341679dfc695SMike Kravetz static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 341779dfc695SMike Kravetz __must_hold(&hugetlb_lock) 341879dfc695SMike Kravetz { 34198531fc6fSMike Kravetz int nr_nodes, node; 34208531fc6fSMike Kravetz struct page *page; 342179dfc695SMike Kravetz int rc = 0; 342279dfc695SMike Kravetz 342379dfc695SMike Kravetz lockdep_assert_held(&hugetlb_lock); 342479dfc695SMike Kravetz 342579dfc695SMike Kravetz /* We should never get here if no demote order */ 342679dfc695SMike Kravetz if (!h->demote_order) { 342779dfc695SMike Kravetz pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n"); 342879dfc695SMike Kravetz return -EINVAL; /* internal error */ 342979dfc695SMike Kravetz } 343079dfc695SMike Kravetz 34318531fc6fSMike Kravetz for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 34328531fc6fSMike Kravetz if (!list_empty(&h->hugepage_freelists[node])) { 34338531fc6fSMike Kravetz page = list_entry(h->hugepage_freelists[node].next, 34348531fc6fSMike Kravetz struct page, lru); 34358531fc6fSMike Kravetz rc = demote_free_huge_page(h, page); 34368531fc6fSMike Kravetz break; 34378531fc6fSMike Kravetz } 34388531fc6fSMike Kravetz } 34398531fc6fSMike Kravetz 344079dfc695SMike Kravetz return rc; 344179dfc695SMike Kravetz } 344279dfc695SMike Kravetz 3443a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 3444a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 3445a3437870SNishanth Aravamudan 344679dfc695SMike Kravetz #define HSTATE_ATTR_WO(_name) \ 344779dfc695SMike Kravetz static struct kobj_attribute _name##_attr = __ATTR_WO(_name) 344879dfc695SMike Kravetz 3449a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 3450a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 3451a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 3452a3437870SNishanth Aravamudan 3453a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 3454a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 3455a3437870SNishanth Aravamudan 34569a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 34579a305230SLee Schermerhorn 34589a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 3459a3437870SNishanth Aravamudan { 3460a3437870SNishanth Aravamudan int i; 34619a305230SLee Schermerhorn 3462a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 34639a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 34649a305230SLee Schermerhorn if (nidp) 34659a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 3466a3437870SNishanth Aravamudan return &hstates[i]; 34679a305230SLee Schermerhorn } 34689a305230SLee Schermerhorn 34699a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 3470a3437870SNishanth Aravamudan } 3471a3437870SNishanth Aravamudan 347206808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 3473a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3474a3437870SNishanth Aravamudan { 34759a305230SLee Schermerhorn struct hstate *h; 34769a305230SLee Schermerhorn unsigned long nr_huge_pages; 34779a305230SLee Schermerhorn int nid; 34789a305230SLee Schermerhorn 34799a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 34809a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 34819a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 34829a305230SLee Schermerhorn else 34839a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 34849a305230SLee Schermerhorn 3485ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", nr_huge_pages); 3486a3437870SNishanth Aravamudan } 3487adbe8726SEric B Munson 3488238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 3489238d3c13SDavid Rientjes struct hstate *h, int nid, 3490238d3c13SDavid Rientjes unsigned long count, size_t len) 3491a3437870SNishanth Aravamudan { 3492a3437870SNishanth Aravamudan int err; 34932d0adf7eSOscar Salvador nodemask_t nodes_allowed, *n_mask; 3494a3437870SNishanth Aravamudan 34952d0adf7eSOscar Salvador if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 34962d0adf7eSOscar Salvador return -EINVAL; 3497adbe8726SEric B Munson 34989a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 34999a305230SLee Schermerhorn /* 35009a305230SLee Schermerhorn * global hstate attribute 35019a305230SLee Schermerhorn */ 35029a305230SLee Schermerhorn if (!(obey_mempolicy && 35032d0adf7eSOscar Salvador init_nodemask_of_mempolicy(&nodes_allowed))) 35042d0adf7eSOscar Salvador n_mask = &node_states[N_MEMORY]; 35052d0adf7eSOscar Salvador else 35062d0adf7eSOscar Salvador n_mask = &nodes_allowed; 35072d0adf7eSOscar Salvador } else { 35089a305230SLee Schermerhorn /* 3509fd875dcaSMike Kravetz * Node specific request. count adjustment happens in 3510fd875dcaSMike Kravetz * set_max_huge_pages() after acquiring hugetlb_lock. 35119a305230SLee Schermerhorn */ 35122d0adf7eSOscar Salvador init_nodemask_of_node(&nodes_allowed, nid); 35132d0adf7eSOscar Salvador n_mask = &nodes_allowed; 3514fd875dcaSMike Kravetz } 35159a305230SLee Schermerhorn 35162d0adf7eSOscar Salvador err = set_max_huge_pages(h, count, nid, n_mask); 351706808b08SLee Schermerhorn 35184eb0716eSAlexandre Ghiti return err ? err : len; 351906808b08SLee Schermerhorn } 352006808b08SLee Schermerhorn 3521238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 3522238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 3523238d3c13SDavid Rientjes size_t len) 3524238d3c13SDavid Rientjes { 3525238d3c13SDavid Rientjes struct hstate *h; 3526238d3c13SDavid Rientjes unsigned long count; 3527238d3c13SDavid Rientjes int nid; 3528238d3c13SDavid Rientjes int err; 3529238d3c13SDavid Rientjes 3530238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 3531238d3c13SDavid Rientjes if (err) 3532238d3c13SDavid Rientjes return err; 3533238d3c13SDavid Rientjes 3534238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 3535238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 3536238d3c13SDavid Rientjes } 3537238d3c13SDavid Rientjes 353806808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 353906808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 354006808b08SLee Schermerhorn { 354106808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 354206808b08SLee Schermerhorn } 354306808b08SLee Schermerhorn 354406808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 354506808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 354606808b08SLee Schermerhorn { 3547238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 3548a3437870SNishanth Aravamudan } 3549a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 3550a3437870SNishanth Aravamudan 355106808b08SLee Schermerhorn #ifdef CONFIG_NUMA 355206808b08SLee Schermerhorn 355306808b08SLee Schermerhorn /* 355406808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 355506808b08SLee Schermerhorn * huge page alloc/free. 355606808b08SLee Schermerhorn */ 355706808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 3558ae7a927dSJoe Perches struct kobj_attribute *attr, 3559ae7a927dSJoe Perches char *buf) 356006808b08SLee Schermerhorn { 356106808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 356206808b08SLee Schermerhorn } 356306808b08SLee Schermerhorn 356406808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 356506808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 356606808b08SLee Schermerhorn { 3567238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 356806808b08SLee Schermerhorn } 356906808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 357006808b08SLee Schermerhorn #endif 357106808b08SLee Schermerhorn 357206808b08SLee Schermerhorn 3573a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 3574a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3575a3437870SNishanth Aravamudan { 35769a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3577ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); 3578a3437870SNishanth Aravamudan } 3579adbe8726SEric B Munson 3580a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 3581a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 3582a3437870SNishanth Aravamudan { 3583a3437870SNishanth Aravamudan int err; 3584a3437870SNishanth Aravamudan unsigned long input; 35859a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3586a3437870SNishanth Aravamudan 3587bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3588adbe8726SEric B Munson return -EINVAL; 3589adbe8726SEric B Munson 35903dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 3591a3437870SNishanth Aravamudan if (err) 359273ae31e5SEric B Munson return err; 3593a3437870SNishanth Aravamudan 3594db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3595a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 3596db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3597a3437870SNishanth Aravamudan 3598a3437870SNishanth Aravamudan return count; 3599a3437870SNishanth Aravamudan } 3600a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 3601a3437870SNishanth Aravamudan 3602a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 3603a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3604a3437870SNishanth Aravamudan { 36059a305230SLee Schermerhorn struct hstate *h; 36069a305230SLee Schermerhorn unsigned long free_huge_pages; 36079a305230SLee Schermerhorn int nid; 36089a305230SLee Schermerhorn 36099a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 36109a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 36119a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 36129a305230SLee Schermerhorn else 36139a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 36149a305230SLee Schermerhorn 3615ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", free_huge_pages); 3616a3437870SNishanth Aravamudan } 3617a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 3618a3437870SNishanth Aravamudan 3619a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 3620a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3621a3437870SNishanth Aravamudan { 36229a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3623ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); 3624a3437870SNishanth Aravamudan } 3625a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 3626a3437870SNishanth Aravamudan 3627a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 3628a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3629a3437870SNishanth Aravamudan { 36309a305230SLee Schermerhorn struct hstate *h; 36319a305230SLee Schermerhorn unsigned long surplus_huge_pages; 36329a305230SLee Schermerhorn int nid; 36339a305230SLee Schermerhorn 36349a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 36359a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 36369a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 36379a305230SLee Schermerhorn else 36389a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 36399a305230SLee Schermerhorn 3640ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", surplus_huge_pages); 3641a3437870SNishanth Aravamudan } 3642a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 3643a3437870SNishanth Aravamudan 364479dfc695SMike Kravetz static ssize_t demote_store(struct kobject *kobj, 364579dfc695SMike Kravetz struct kobj_attribute *attr, const char *buf, size_t len) 364679dfc695SMike Kravetz { 364779dfc695SMike Kravetz unsigned long nr_demote; 364879dfc695SMike Kravetz unsigned long nr_available; 364979dfc695SMike Kravetz nodemask_t nodes_allowed, *n_mask; 365079dfc695SMike Kravetz struct hstate *h; 365179dfc695SMike Kravetz int err = 0; 365279dfc695SMike Kravetz int nid; 365379dfc695SMike Kravetz 365479dfc695SMike Kravetz err = kstrtoul(buf, 10, &nr_demote); 365579dfc695SMike Kravetz if (err) 365679dfc695SMike Kravetz return err; 365779dfc695SMike Kravetz h = kobj_to_hstate(kobj, &nid); 365879dfc695SMike Kravetz 365979dfc695SMike Kravetz if (nid != NUMA_NO_NODE) { 366079dfc695SMike Kravetz init_nodemask_of_node(&nodes_allowed, nid); 366179dfc695SMike Kravetz n_mask = &nodes_allowed; 366279dfc695SMike Kravetz } else { 366379dfc695SMike Kravetz n_mask = &node_states[N_MEMORY]; 366479dfc695SMike Kravetz } 366579dfc695SMike Kravetz 366679dfc695SMike Kravetz /* Synchronize with other sysfs operations modifying huge pages */ 366779dfc695SMike Kravetz mutex_lock(&h->resize_lock); 366879dfc695SMike Kravetz spin_lock_irq(&hugetlb_lock); 366979dfc695SMike Kravetz 367079dfc695SMike Kravetz while (nr_demote) { 367179dfc695SMike Kravetz /* 367279dfc695SMike Kravetz * Check for available pages to demote each time thorough the 367379dfc695SMike Kravetz * loop as demote_pool_huge_page will drop hugetlb_lock. 367479dfc695SMike Kravetz */ 367579dfc695SMike Kravetz if (nid != NUMA_NO_NODE) 367679dfc695SMike Kravetz nr_available = h->free_huge_pages_node[nid]; 367779dfc695SMike Kravetz else 367879dfc695SMike Kravetz nr_available = h->free_huge_pages; 367979dfc695SMike Kravetz nr_available -= h->resv_huge_pages; 368079dfc695SMike Kravetz if (!nr_available) 368179dfc695SMike Kravetz break; 368279dfc695SMike Kravetz 368379dfc695SMike Kravetz err = demote_pool_huge_page(h, n_mask); 368479dfc695SMike Kravetz if (err) 368579dfc695SMike Kravetz break; 368679dfc695SMike Kravetz 368779dfc695SMike Kravetz nr_demote--; 368879dfc695SMike Kravetz } 368979dfc695SMike Kravetz 369079dfc695SMike Kravetz spin_unlock_irq(&hugetlb_lock); 369179dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 369279dfc695SMike Kravetz 369379dfc695SMike Kravetz if (err) 369479dfc695SMike Kravetz return err; 369579dfc695SMike Kravetz return len; 369679dfc695SMike Kravetz } 369779dfc695SMike Kravetz HSTATE_ATTR_WO(demote); 369879dfc695SMike Kravetz 369979dfc695SMike Kravetz static ssize_t demote_size_show(struct kobject *kobj, 370079dfc695SMike Kravetz struct kobj_attribute *attr, char *buf) 370179dfc695SMike Kravetz { 370279dfc695SMike Kravetz int nid; 370379dfc695SMike Kravetz struct hstate *h = kobj_to_hstate(kobj, &nid); 370479dfc695SMike Kravetz unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K; 370579dfc695SMike Kravetz 370679dfc695SMike Kravetz return sysfs_emit(buf, "%lukB\n", demote_size); 370779dfc695SMike Kravetz } 370879dfc695SMike Kravetz 370979dfc695SMike Kravetz static ssize_t demote_size_store(struct kobject *kobj, 371079dfc695SMike Kravetz struct kobj_attribute *attr, 371179dfc695SMike Kravetz const char *buf, size_t count) 371279dfc695SMike Kravetz { 371379dfc695SMike Kravetz struct hstate *h, *demote_hstate; 371479dfc695SMike Kravetz unsigned long demote_size; 371579dfc695SMike Kravetz unsigned int demote_order; 371679dfc695SMike Kravetz int nid; 371779dfc695SMike Kravetz 371879dfc695SMike Kravetz demote_size = (unsigned long)memparse(buf, NULL); 371979dfc695SMike Kravetz 372079dfc695SMike Kravetz demote_hstate = size_to_hstate(demote_size); 372179dfc695SMike Kravetz if (!demote_hstate) 372279dfc695SMike Kravetz return -EINVAL; 372379dfc695SMike Kravetz demote_order = demote_hstate->order; 3724a01f4390SMike Kravetz if (demote_order < HUGETLB_PAGE_ORDER) 3725a01f4390SMike Kravetz return -EINVAL; 372679dfc695SMike Kravetz 372779dfc695SMike Kravetz /* demote order must be smaller than hstate order */ 372879dfc695SMike Kravetz h = kobj_to_hstate(kobj, &nid); 372979dfc695SMike Kravetz if (demote_order >= h->order) 373079dfc695SMike Kravetz return -EINVAL; 373179dfc695SMike Kravetz 373279dfc695SMike Kravetz /* resize_lock synchronizes access to demote size and writes */ 373379dfc695SMike Kravetz mutex_lock(&h->resize_lock); 373479dfc695SMike Kravetz h->demote_order = demote_order; 373579dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 373679dfc695SMike Kravetz 373779dfc695SMike Kravetz return count; 373879dfc695SMike Kravetz } 373979dfc695SMike Kravetz HSTATE_ATTR(demote_size); 374079dfc695SMike Kravetz 3741a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 3742a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 3743a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 3744a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 3745a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 3746a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 374706808b08SLee Schermerhorn #ifdef CONFIG_NUMA 374806808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 374906808b08SLee Schermerhorn #endif 3750a3437870SNishanth Aravamudan NULL, 3751a3437870SNishanth Aravamudan }; 3752a3437870SNishanth Aravamudan 375367e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 3754a3437870SNishanth Aravamudan .attrs = hstate_attrs, 3755a3437870SNishanth Aravamudan }; 3756a3437870SNishanth Aravamudan 375779dfc695SMike Kravetz static struct attribute *hstate_demote_attrs[] = { 375879dfc695SMike Kravetz &demote_size_attr.attr, 375979dfc695SMike Kravetz &demote_attr.attr, 376079dfc695SMike Kravetz NULL, 376179dfc695SMike Kravetz }; 376279dfc695SMike Kravetz 376379dfc695SMike Kravetz static const struct attribute_group hstate_demote_attr_group = { 376479dfc695SMike Kravetz .attrs = hstate_demote_attrs, 376579dfc695SMike Kravetz }; 376679dfc695SMike Kravetz 3767094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 37689a305230SLee Schermerhorn struct kobject **hstate_kobjs, 376967e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 3770a3437870SNishanth Aravamudan { 3771a3437870SNishanth Aravamudan int retval; 3772972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 3773a3437870SNishanth Aravamudan 37749a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 37759a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 3776a3437870SNishanth Aravamudan return -ENOMEM; 3777a3437870SNishanth Aravamudan 37789a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 3779cc2205a6SMiaohe Lin if (retval) { 37809a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 3781cc2205a6SMiaohe Lin hstate_kobjs[hi] = NULL; 3782cc2205a6SMiaohe Lin } 3783a3437870SNishanth Aravamudan 378479dfc695SMike Kravetz if (h->demote_order) { 378579dfc695SMike Kravetz if (sysfs_create_group(hstate_kobjs[hi], 378679dfc695SMike Kravetz &hstate_demote_attr_group)) 378779dfc695SMike Kravetz pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name); 378879dfc695SMike Kravetz } 378979dfc695SMike Kravetz 3790a3437870SNishanth Aravamudan return retval; 3791a3437870SNishanth Aravamudan } 3792a3437870SNishanth Aravamudan 3793a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 3794a3437870SNishanth Aravamudan { 3795a3437870SNishanth Aravamudan struct hstate *h; 3796a3437870SNishanth Aravamudan int err; 3797a3437870SNishanth Aravamudan 3798a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 3799a3437870SNishanth Aravamudan if (!hugepages_kobj) 3800a3437870SNishanth Aravamudan return; 3801a3437870SNishanth Aravamudan 3802a3437870SNishanth Aravamudan for_each_hstate(h) { 38039a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 38049a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 3805a3437870SNishanth Aravamudan if (err) 3806282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s", h->name); 3807a3437870SNishanth Aravamudan } 3808a3437870SNishanth Aravamudan } 3809a3437870SNishanth Aravamudan 38109a305230SLee Schermerhorn #ifdef CONFIG_NUMA 38119a305230SLee Schermerhorn 38129a305230SLee Schermerhorn /* 38139a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 381410fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 381510fbcf4cSKay Sievers * index of a node device or _hstate == node id. 381610fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 38179a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 38189a305230SLee Schermerhorn */ 38199a305230SLee Schermerhorn struct node_hstate { 38209a305230SLee Schermerhorn struct kobject *hugepages_kobj; 38219a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 38229a305230SLee Schermerhorn }; 3823b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 38249a305230SLee Schermerhorn 38259a305230SLee Schermerhorn /* 382610fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 38279a305230SLee Schermerhorn */ 38289a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 38299a305230SLee Schermerhorn &nr_hugepages_attr.attr, 38309a305230SLee Schermerhorn &free_hugepages_attr.attr, 38319a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 38329a305230SLee Schermerhorn NULL, 38339a305230SLee Schermerhorn }; 38349a305230SLee Schermerhorn 383567e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 38369a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 38379a305230SLee Schermerhorn }; 38389a305230SLee Schermerhorn 38399a305230SLee Schermerhorn /* 384010fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 38419a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 38429a305230SLee Schermerhorn */ 38439a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 38449a305230SLee Schermerhorn { 38459a305230SLee Schermerhorn int nid; 38469a305230SLee Schermerhorn 38479a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 38489a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 38499a305230SLee Schermerhorn int i; 38509a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 38519a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 38529a305230SLee Schermerhorn if (nidp) 38539a305230SLee Schermerhorn *nidp = nid; 38549a305230SLee Schermerhorn return &hstates[i]; 38559a305230SLee Schermerhorn } 38569a305230SLee Schermerhorn } 38579a305230SLee Schermerhorn 38589a305230SLee Schermerhorn BUG(); 38599a305230SLee Schermerhorn return NULL; 38609a305230SLee Schermerhorn } 38619a305230SLee Schermerhorn 38629a305230SLee Schermerhorn /* 386310fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 38649a305230SLee Schermerhorn * No-op if no hstate attributes attached. 38659a305230SLee Schermerhorn */ 38663cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 38679a305230SLee Schermerhorn { 38689a305230SLee Schermerhorn struct hstate *h; 386910fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 38709a305230SLee Schermerhorn 38719a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 38729b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 38739a305230SLee Schermerhorn 3874972dc4deSAneesh Kumar K.V for_each_hstate(h) { 3875972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 3876972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 3877972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 3878972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 3879972dc4deSAneesh Kumar K.V } 38809a305230SLee Schermerhorn } 38819a305230SLee Schermerhorn 38829a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 38839a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 38849a305230SLee Schermerhorn } 38859a305230SLee Schermerhorn 38869a305230SLee Schermerhorn 38879a305230SLee Schermerhorn /* 388810fbcf4cSKay Sievers * Register hstate attributes for a single node device. 38899a305230SLee Schermerhorn * No-op if attributes already registered. 38909a305230SLee Schermerhorn */ 38913cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 38929a305230SLee Schermerhorn { 38939a305230SLee Schermerhorn struct hstate *h; 389410fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 38959a305230SLee Schermerhorn int err; 38969a305230SLee Schermerhorn 38979a305230SLee Schermerhorn if (nhs->hugepages_kobj) 38989a305230SLee Schermerhorn return; /* already allocated */ 38999a305230SLee Schermerhorn 39009a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 390110fbcf4cSKay Sievers &node->dev.kobj); 39029a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 39039a305230SLee Schermerhorn return; 39049a305230SLee Schermerhorn 39059a305230SLee Schermerhorn for_each_hstate(h) { 39069a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 39079a305230SLee Schermerhorn nhs->hstate_kobjs, 39089a305230SLee Schermerhorn &per_node_hstate_attr_group); 39099a305230SLee Schermerhorn if (err) { 3910282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s for node %d\n", 391110fbcf4cSKay Sievers h->name, node->dev.id); 39129a305230SLee Schermerhorn hugetlb_unregister_node(node); 39139a305230SLee Schermerhorn break; 39149a305230SLee Schermerhorn } 39159a305230SLee Schermerhorn } 39169a305230SLee Schermerhorn } 39179a305230SLee Schermerhorn 39189a305230SLee Schermerhorn /* 39199b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 392010fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 392110fbcf4cSKay Sievers * registered their associated device by this time. 39229a305230SLee Schermerhorn */ 39237d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 39249a305230SLee Schermerhorn { 39259a305230SLee Schermerhorn int nid; 39269a305230SLee Schermerhorn 39278cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 39288732794bSWen Congyang struct node *node = node_devices[nid]; 392910fbcf4cSKay Sievers if (node->dev.id == nid) 39309a305230SLee Schermerhorn hugetlb_register_node(node); 39319a305230SLee Schermerhorn } 39329a305230SLee Schermerhorn 39339a305230SLee Schermerhorn /* 393410fbcf4cSKay Sievers * Let the node device driver know we're here so it can 39359a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 39369a305230SLee Schermerhorn */ 39379a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 39389a305230SLee Schermerhorn hugetlb_unregister_node); 39399a305230SLee Schermerhorn } 39409a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 39419a305230SLee Schermerhorn 39429a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 39439a305230SLee Schermerhorn { 39449a305230SLee Schermerhorn BUG(); 39459a305230SLee Schermerhorn if (nidp) 39469a305230SLee Schermerhorn *nidp = -1; 39479a305230SLee Schermerhorn return NULL; 39489a305230SLee Schermerhorn } 39499a305230SLee Schermerhorn 39509a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 39519a305230SLee Schermerhorn 39529a305230SLee Schermerhorn #endif 39539a305230SLee Schermerhorn 3954a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 3955a3437870SNishanth Aravamudan { 39568382d914SDavidlohr Bueso int i; 39578382d914SDavidlohr Bueso 3958d6995da3SMike Kravetz BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < 3959d6995da3SMike Kravetz __NR_HPAGEFLAGS); 3960d6995da3SMike Kravetz 3961c2833a5bSMike Kravetz if (!hugepages_supported()) { 3962c2833a5bSMike Kravetz if (hugetlb_max_hstate || default_hstate_max_huge_pages) 3963c2833a5bSMike Kravetz pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); 39640ef89d25SBenjamin Herrenschmidt return 0; 3965d715cf80SLiam R. Howlett } 3966d715cf80SLiam R. Howlett 3967282f4214SMike Kravetz /* 3968282f4214SMike Kravetz * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some 3969282f4214SMike Kravetz * architectures depend on setup being done here. 3970282f4214SMike Kravetz */ 3971a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 3972282f4214SMike Kravetz if (!parsed_default_hugepagesz) { 3973282f4214SMike Kravetz /* 3974282f4214SMike Kravetz * If we did not parse a default huge page size, set 3975282f4214SMike Kravetz * default_hstate_idx to HPAGE_SIZE hstate. And, if the 3976282f4214SMike Kravetz * number of huge pages for this default size was implicitly 3977282f4214SMike Kravetz * specified, set that here as well. 3978282f4214SMike Kravetz * Note that the implicit setting will overwrite an explicit 3979282f4214SMike Kravetz * setting. A warning will be printed in this case. 3980282f4214SMike Kravetz */ 3981282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); 3982f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 3983282f4214SMike Kravetz if (default_hstate.max_huge_pages) { 3984282f4214SMike Kravetz char buf[32]; 3985282f4214SMike Kravetz 3986282f4214SMike Kravetz string_get_size(huge_page_size(&default_hstate), 3987282f4214SMike Kravetz 1, STRING_UNITS_2, buf, 32); 3988282f4214SMike Kravetz pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", 3989282f4214SMike Kravetz default_hstate.max_huge_pages, buf); 3990282f4214SMike Kravetz pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", 3991282f4214SMike Kravetz default_hstate_max_huge_pages); 3992282f4214SMike Kravetz } 3993282f4214SMike Kravetz default_hstate.max_huge_pages = 3994282f4214SMike Kravetz default_hstate_max_huge_pages; 3995282f4214SMike Kravetz } 3996f8b74815SVaishali Thakkar } 3997a3437870SNishanth Aravamudan 3998cf11e85fSRoman Gushchin hugetlb_cma_check(); 3999a3437870SNishanth Aravamudan hugetlb_init_hstates(); 4000aa888a74SAndi Kleen gather_bootmem_prealloc(); 4001a3437870SNishanth Aravamudan report_hugepages(); 4002a3437870SNishanth Aravamudan 4003a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 40049a305230SLee Schermerhorn hugetlb_register_all_nodes(); 40057179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 40069a305230SLee Schermerhorn 40078382d914SDavidlohr Bueso #ifdef CONFIG_SMP 40088382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 40098382d914SDavidlohr Bueso #else 40108382d914SDavidlohr Bueso num_fault_mutexes = 1; 40118382d914SDavidlohr Bueso #endif 4012c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 40136da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 40146da2ec56SKees Cook GFP_KERNEL); 4015c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 40168382d914SDavidlohr Bueso 40178382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 4018c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 4019a3437870SNishanth Aravamudan return 0; 4020a3437870SNishanth Aravamudan } 40213e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 4022a3437870SNishanth Aravamudan 4023ae94da89SMike Kravetz /* Overwritten by architectures with more huge page sizes */ 4024ae94da89SMike Kravetz bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) 40259fee021dSVaishali Thakkar { 4026ae94da89SMike Kravetz return size == HPAGE_SIZE; 40279fee021dSVaishali Thakkar } 40289fee021dSVaishali Thakkar 4029d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 4030a3437870SNishanth Aravamudan { 4031a3437870SNishanth Aravamudan struct hstate *h; 40328faa8b07SAndi Kleen unsigned long i; 40338faa8b07SAndi Kleen 4034a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 4035a3437870SNishanth Aravamudan return; 4036a3437870SNishanth Aravamudan } 403747d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 4038a3437870SNishanth Aravamudan BUG_ON(order == 0); 403947d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 404029383967SMike Kravetz mutex_init(&h->resize_lock); 4041a3437870SNishanth Aravamudan h->order = order; 4042aca78307SMiaohe Lin h->mask = ~(huge_page_size(h) - 1); 40438faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 40448faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 40450edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 404654f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 404754f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 4048a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 4049a3437870SNishanth Aravamudan huge_page_size(h)/1024); 405077490587SMuchun Song hugetlb_vmemmap_init(h); 40518faa8b07SAndi Kleen 4052a3437870SNishanth Aravamudan parsed_hstate = h; 4053a3437870SNishanth Aravamudan } 4054a3437870SNishanth Aravamudan 4055282f4214SMike Kravetz /* 4056282f4214SMike Kravetz * hugepages command line processing 4057282f4214SMike Kravetz * hugepages normally follows a valid hugepagsz or default_hugepagsz 4058282f4214SMike Kravetz * specification. If not, ignore the hugepages value. hugepages can also 4059282f4214SMike Kravetz * be the first huge page command line option in which case it implicitly 4060282f4214SMike Kravetz * specifies the number of huge pages for the default size. 4061282f4214SMike Kravetz */ 4062282f4214SMike Kravetz static int __init hugepages_setup(char *s) 4063a3437870SNishanth Aravamudan { 4064a3437870SNishanth Aravamudan unsigned long *mhp; 40658faa8b07SAndi Kleen static unsigned long *last_mhp; 4066a3437870SNishanth Aravamudan 40679fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 4068282f4214SMike Kravetz pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); 40699fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 4070282f4214SMike Kravetz return 0; 40719fee021dSVaishali Thakkar } 4072282f4214SMike Kravetz 4073a3437870SNishanth Aravamudan /* 4074282f4214SMike Kravetz * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter 4075282f4214SMike Kravetz * yet, so this hugepages= parameter goes to the "default hstate". 4076282f4214SMike Kravetz * Otherwise, it goes with the previously parsed hugepagesz or 4077282f4214SMike Kravetz * default_hugepagesz. 4078a3437870SNishanth Aravamudan */ 40799fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 4080a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 4081a3437870SNishanth Aravamudan else 4082a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 4083a3437870SNishanth Aravamudan 40848faa8b07SAndi Kleen if (mhp == last_mhp) { 4085282f4214SMike Kravetz pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); 4086282f4214SMike Kravetz return 0; 40878faa8b07SAndi Kleen } 40888faa8b07SAndi Kleen 4089a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 4090a3437870SNishanth Aravamudan *mhp = 0; 4091a3437870SNishanth Aravamudan 40928faa8b07SAndi Kleen /* 40938faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 409404adbc3fSMiaohe Lin * But we need to allocate gigantic hstates here early to still 40958faa8b07SAndi Kleen * use the bootmem allocator. 40968faa8b07SAndi Kleen */ 409704adbc3fSMiaohe Lin if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) 40988faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 40998faa8b07SAndi Kleen 41008faa8b07SAndi Kleen last_mhp = mhp; 41018faa8b07SAndi Kleen 4102a3437870SNishanth Aravamudan return 1; 4103a3437870SNishanth Aravamudan } 4104282f4214SMike Kravetz __setup("hugepages=", hugepages_setup); 4105e11bfbfcSNick Piggin 4106282f4214SMike Kravetz /* 4107282f4214SMike Kravetz * hugepagesz command line processing 4108282f4214SMike Kravetz * A specific huge page size can only be specified once with hugepagesz. 4109282f4214SMike Kravetz * hugepagesz is followed by hugepages on the command line. The global 4110282f4214SMike Kravetz * variable 'parsed_valid_hugepagesz' is used to determine if prior 4111282f4214SMike Kravetz * hugepagesz argument was valid. 4112282f4214SMike Kravetz */ 4113359f2544SMike Kravetz static int __init hugepagesz_setup(char *s) 4114e11bfbfcSNick Piggin { 4115359f2544SMike Kravetz unsigned long size; 4116282f4214SMike Kravetz struct hstate *h; 4117282f4214SMike Kravetz 4118282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4119359f2544SMike Kravetz size = (unsigned long)memparse(s, NULL); 4120359f2544SMike Kravetz 4121359f2544SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4122282f4214SMike Kravetz pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); 4123359f2544SMike Kravetz return 0; 4124359f2544SMike Kravetz } 4125359f2544SMike Kravetz 4126282f4214SMike Kravetz h = size_to_hstate(size); 4127282f4214SMike Kravetz if (h) { 4128282f4214SMike Kravetz /* 4129282f4214SMike Kravetz * hstate for this size already exists. This is normally 4130282f4214SMike Kravetz * an error, but is allowed if the existing hstate is the 4131282f4214SMike Kravetz * default hstate. More specifically, it is only allowed if 4132282f4214SMike Kravetz * the number of huge pages for the default hstate was not 4133282f4214SMike Kravetz * previously specified. 4134282f4214SMike Kravetz */ 4135282f4214SMike Kravetz if (!parsed_default_hugepagesz || h != &default_hstate || 4136282f4214SMike Kravetz default_hstate.max_huge_pages) { 4137282f4214SMike Kravetz pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); 413838237830SMike Kravetz return 0; 413938237830SMike Kravetz } 414038237830SMike Kravetz 4141282f4214SMike Kravetz /* 4142282f4214SMike Kravetz * No need to call hugetlb_add_hstate() as hstate already 4143282f4214SMike Kravetz * exists. But, do set parsed_hstate so that a following 4144282f4214SMike Kravetz * hugepages= parameter will be applied to this hstate. 4145282f4214SMike Kravetz */ 4146282f4214SMike Kravetz parsed_hstate = h; 4147282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4148e11bfbfcSNick Piggin return 1; 4149e11bfbfcSNick Piggin } 4150282f4214SMike Kravetz 4151359f2544SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4152282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4153359f2544SMike Kravetz return 1; 4154359f2544SMike Kravetz } 4155359f2544SMike Kravetz __setup("hugepagesz=", hugepagesz_setup); 4156359f2544SMike Kravetz 4157282f4214SMike Kravetz /* 4158282f4214SMike Kravetz * default_hugepagesz command line input 4159282f4214SMike Kravetz * Only one instance of default_hugepagesz allowed on command line. 4160282f4214SMike Kravetz */ 4161ae94da89SMike Kravetz static int __init default_hugepagesz_setup(char *s) 4162e11bfbfcSNick Piggin { 4163ae94da89SMike Kravetz unsigned long size; 4164ae94da89SMike Kravetz 4165282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4166282f4214SMike Kravetz if (parsed_default_hugepagesz) { 4167282f4214SMike Kravetz pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); 4168282f4214SMike Kravetz return 0; 4169282f4214SMike Kravetz } 4170282f4214SMike Kravetz 4171282f4214SMike Kravetz size = (unsigned long)memparse(s, NULL); 4172282f4214SMike Kravetz 4173282f4214SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4174282f4214SMike Kravetz pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); 4175282f4214SMike Kravetz return 0; 4176282f4214SMike Kravetz } 4177282f4214SMike Kravetz 4178282f4214SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4179282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4180282f4214SMike Kravetz parsed_default_hugepagesz = true; 4181282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(size)); 4182282f4214SMike Kravetz 4183282f4214SMike Kravetz /* 4184282f4214SMike Kravetz * The number of default huge pages (for this size) could have been 4185282f4214SMike Kravetz * specified as the first hugetlb parameter: hugepages=X. If so, 4186282f4214SMike Kravetz * then default_hstate_max_huge_pages is set. If the default huge 4187282f4214SMike Kravetz * page size is gigantic (>= MAX_ORDER), then the pages must be 4188282f4214SMike Kravetz * allocated here from bootmem allocator. 4189282f4214SMike Kravetz */ 4190282f4214SMike Kravetz if (default_hstate_max_huge_pages) { 4191282f4214SMike Kravetz default_hstate.max_huge_pages = default_hstate_max_huge_pages; 4192282f4214SMike Kravetz if (hstate_is_gigantic(&default_hstate)) 4193282f4214SMike Kravetz hugetlb_hstate_alloc_pages(&default_hstate); 4194282f4214SMike Kravetz default_hstate_max_huge_pages = 0; 4195282f4214SMike Kravetz } 4196282f4214SMike Kravetz 4197e11bfbfcSNick Piggin return 1; 4198e11bfbfcSNick Piggin } 4199ae94da89SMike Kravetz __setup("default_hugepagesz=", default_hugepagesz_setup); 4200a3437870SNishanth Aravamudan 42018ca39e68SMuchun Song static unsigned int allowed_mems_nr(struct hstate *h) 42028a213460SNishanth Aravamudan { 42038a213460SNishanth Aravamudan int node; 42048a213460SNishanth Aravamudan unsigned int nr = 0; 42058ca39e68SMuchun Song nodemask_t *mpol_allowed; 42068ca39e68SMuchun Song unsigned int *array = h->free_huge_pages_node; 42078ca39e68SMuchun Song gfp_t gfp_mask = htlb_alloc_mask(h); 42088a213460SNishanth Aravamudan 42098ca39e68SMuchun Song mpol_allowed = policy_nodemask_current(gfp_mask); 42108ca39e68SMuchun Song 42118ca39e68SMuchun Song for_each_node_mask(node, cpuset_current_mems_allowed) { 4212c93b0a99SJiapeng Zhong if (!mpol_allowed || node_isset(node, *mpol_allowed)) 42138a213460SNishanth Aravamudan nr += array[node]; 42148ca39e68SMuchun Song } 42158a213460SNishanth Aravamudan 42168a213460SNishanth Aravamudan return nr; 42178a213460SNishanth Aravamudan } 42188a213460SNishanth Aravamudan 42198a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 422017743798SMuchun Song static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, 422117743798SMuchun Song void *buffer, size_t *length, 422217743798SMuchun Song loff_t *ppos, unsigned long *out) 422317743798SMuchun Song { 422417743798SMuchun Song struct ctl_table dup_table; 422517743798SMuchun Song 422617743798SMuchun Song /* 422717743798SMuchun Song * In order to avoid races with __do_proc_doulongvec_minmax(), we 422817743798SMuchun Song * can duplicate the @table and alter the duplicate of it. 422917743798SMuchun Song */ 423017743798SMuchun Song dup_table = *table; 423117743798SMuchun Song dup_table.data = out; 423217743798SMuchun Song 423317743798SMuchun Song return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); 423417743798SMuchun Song } 423517743798SMuchun Song 423606808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 423706808b08SLee Schermerhorn struct ctl_table *table, int write, 423832927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 42391da177e4SLinus Torvalds { 4240e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 4241238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 424208d4a246SMichal Hocko int ret; 4243e5ff2159SAndi Kleen 4244457c1b27SNishanth Aravamudan if (!hugepages_supported()) 424586613628SJan Stancek return -EOPNOTSUPP; 4246457c1b27SNishanth Aravamudan 424717743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 424817743798SMuchun Song &tmp); 424908d4a246SMichal Hocko if (ret) 425008d4a246SMichal Hocko goto out; 4251e5ff2159SAndi Kleen 4252238d3c13SDavid Rientjes if (write) 4253238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 4254238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 425508d4a246SMichal Hocko out: 425608d4a246SMichal Hocko return ret; 42571da177e4SLinus Torvalds } 4258396faf03SMel Gorman 425906808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 426032927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 426106808b08SLee Schermerhorn { 426206808b08SLee Schermerhorn 426306808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 426406808b08SLee Schermerhorn buffer, length, ppos); 426506808b08SLee Schermerhorn } 426606808b08SLee Schermerhorn 426706808b08SLee Schermerhorn #ifdef CONFIG_NUMA 426806808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 426932927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 427006808b08SLee Schermerhorn { 427106808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 427206808b08SLee Schermerhorn buffer, length, ppos); 427306808b08SLee Schermerhorn } 427406808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 427506808b08SLee Schermerhorn 4276a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 427732927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 4278a3d0c6aaSNishanth Aravamudan { 4279a5516438SAndi Kleen struct hstate *h = &default_hstate; 4280e5ff2159SAndi Kleen unsigned long tmp; 428108d4a246SMichal Hocko int ret; 4282e5ff2159SAndi Kleen 4283457c1b27SNishanth Aravamudan if (!hugepages_supported()) 428486613628SJan Stancek return -EOPNOTSUPP; 4285457c1b27SNishanth Aravamudan 4286e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 4287e5ff2159SAndi Kleen 4288bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 4289adbe8726SEric B Munson return -EINVAL; 4290adbe8726SEric B Munson 429117743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 429217743798SMuchun Song &tmp); 429308d4a246SMichal Hocko if (ret) 429408d4a246SMichal Hocko goto out; 4295e5ff2159SAndi Kleen 4296e5ff2159SAndi Kleen if (write) { 4297db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4298e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 4299db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4300e5ff2159SAndi Kleen } 430108d4a246SMichal Hocko out: 430208d4a246SMichal Hocko return ret; 4303a3d0c6aaSNishanth Aravamudan } 4304a3d0c6aaSNishanth Aravamudan 43051da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 43061da177e4SLinus Torvalds 4307e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 43081da177e4SLinus Torvalds { 4309fcb2b0c5SRoman Gushchin struct hstate *h; 4310fcb2b0c5SRoman Gushchin unsigned long total = 0; 4311fcb2b0c5SRoman Gushchin 4312457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4313457c1b27SNishanth Aravamudan return; 4314fcb2b0c5SRoman Gushchin 4315fcb2b0c5SRoman Gushchin for_each_hstate(h) { 4316fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 4317fcb2b0c5SRoman Gushchin 4318aca78307SMiaohe Lin total += huge_page_size(h) * count; 4319fcb2b0c5SRoman Gushchin 4320fcb2b0c5SRoman Gushchin if (h == &default_hstate) 4321e1759c21SAlexey Dobriyan seq_printf(m, 43221da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 43231da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 4324b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 43257893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 43264f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 4327fcb2b0c5SRoman Gushchin count, 4328a5516438SAndi Kleen h->free_huge_pages, 4329a5516438SAndi Kleen h->resv_huge_pages, 4330a5516438SAndi Kleen h->surplus_huge_pages, 4331aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4332fcb2b0c5SRoman Gushchin } 4333fcb2b0c5SRoman Gushchin 4334aca78307SMiaohe Lin seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); 43351da177e4SLinus Torvalds } 43361da177e4SLinus Torvalds 43377981593bSJoe Perches int hugetlb_report_node_meminfo(char *buf, int len, int nid) 43381da177e4SLinus Torvalds { 4339a5516438SAndi Kleen struct hstate *h = &default_hstate; 43407981593bSJoe Perches 4341457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4342457c1b27SNishanth Aravamudan return 0; 43437981593bSJoe Perches 43447981593bSJoe Perches return sysfs_emit_at(buf, len, 43451da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 4346a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 4347a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 4348a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 4349a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 4350a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 43511da177e4SLinus Torvalds } 43521da177e4SLinus Torvalds 4353949f7ec5SDavid Rientjes void hugetlb_show_meminfo(void) 4354949f7ec5SDavid Rientjes { 4355949f7ec5SDavid Rientjes struct hstate *h; 4356949f7ec5SDavid Rientjes int nid; 4357949f7ec5SDavid Rientjes 4358457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4359457c1b27SNishanth Aravamudan return; 4360457c1b27SNishanth Aravamudan 4361949f7ec5SDavid Rientjes for_each_node_state(nid, N_MEMORY) 4362949f7ec5SDavid Rientjes for_each_hstate(h) 4363949f7ec5SDavid Rientjes pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 4364949f7ec5SDavid Rientjes nid, 4365949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 4366949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 4367949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 4368aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4369949f7ec5SDavid Rientjes } 4370949f7ec5SDavid Rientjes 43715d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 43725d317b2bSNaoya Horiguchi { 43735d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 43745d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 43755d317b2bSNaoya Horiguchi } 43765d317b2bSNaoya Horiguchi 43771da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 43781da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 43791da177e4SLinus Torvalds { 4380d0028588SWanpeng Li struct hstate *h; 4381d0028588SWanpeng Li unsigned long nr_total_pages = 0; 4382d0028588SWanpeng Li 4383d0028588SWanpeng Li for_each_hstate(h) 4384d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 4385d0028588SWanpeng Li return nr_total_pages; 43861da177e4SLinus Torvalds } 43871da177e4SLinus Torvalds 4388a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 4389fc1b8a73SMel Gorman { 4390fc1b8a73SMel Gorman int ret = -ENOMEM; 4391fc1b8a73SMel Gorman 43920aa7f354SMiaohe Lin if (!delta) 43930aa7f354SMiaohe Lin return 0; 43940aa7f354SMiaohe Lin 4395db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4396fc1b8a73SMel Gorman /* 4397fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 4398fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 4399fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 4400fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 4401fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 4402fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 4403fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 4404fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 4405fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 4406fc1b8a73SMel Gorman * 4407fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 4408fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 4409fc1b8a73SMel Gorman * we fall back to check against current free page availability as 4410fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 4411fc1b8a73SMel Gorman * semantics that cpuset has. 44128ca39e68SMuchun Song * 44138ca39e68SMuchun Song * Apart from cpuset, we also have memory policy mechanism that 44148ca39e68SMuchun Song * also determines from which node the kernel will allocate memory 44158ca39e68SMuchun Song * in a NUMA system. So similar to cpuset, we also should consider 44168ca39e68SMuchun Song * the memory policy of the current task. Similar to the description 44178ca39e68SMuchun Song * above. 4418fc1b8a73SMel Gorman */ 4419fc1b8a73SMel Gorman if (delta > 0) { 4420a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 4421fc1b8a73SMel Gorman goto out; 4422fc1b8a73SMel Gorman 44238ca39e68SMuchun Song if (delta > allowed_mems_nr(h)) { 4424a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 4425fc1b8a73SMel Gorman goto out; 4426fc1b8a73SMel Gorman } 4427fc1b8a73SMel Gorman } 4428fc1b8a73SMel Gorman 4429fc1b8a73SMel Gorman ret = 0; 4430fc1b8a73SMel Gorman if (delta < 0) 4431a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 4432fc1b8a73SMel Gorman 4433fc1b8a73SMel Gorman out: 4434db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4435fc1b8a73SMel Gorman return ret; 4436fc1b8a73SMel Gorman } 4437fc1b8a73SMel Gorman 443884afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 443984afd99bSAndy Whitcroft { 4440f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 444184afd99bSAndy Whitcroft 444284afd99bSAndy Whitcroft /* 444384afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 444484afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 444584afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 444625985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 444784afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 444884afd99bSAndy Whitcroft * new reference here without additional locking. 444984afd99bSAndy Whitcroft */ 445009a26e83SMike Kravetz if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 445109a26e83SMike Kravetz resv_map_dup_hugetlb_cgroup_uncharge_info(resv); 4452f522c3acSJoonsoo Kim kref_get(&resv->refs); 445384afd99bSAndy Whitcroft } 445409a26e83SMike Kravetz } 445584afd99bSAndy Whitcroft 4456a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 4457a1e78772SMel Gorman { 4458a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 4459f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 446090481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 44614e35f483SJoonsoo Kim unsigned long reserve, start, end; 44621c5ecae3SMike Kravetz long gbl_reserve; 446384afd99bSAndy Whitcroft 44644e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 44654e35f483SJoonsoo Kim return; 44664e35f483SJoonsoo Kim 4467a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 4468a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 446984afd99bSAndy Whitcroft 44704e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 4471e9fe92aeSMina Almasry hugetlb_cgroup_uncharge_counter(resv, start, end); 44727251ff78SAdam Litke if (reserve) { 44731c5ecae3SMike Kravetz /* 44741c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 44751c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 44761c5ecae3SMike Kravetz */ 44771c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 44781c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 44797251ff78SAdam Litke } 4480e9fe92aeSMina Almasry 4481e9fe92aeSMina Almasry kref_put(&resv->refs, resv_map_release); 4482a1e78772SMel Gorman } 4483a1e78772SMel Gorman 448431383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 448531383c68SDan Williams { 448631383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 448731383c68SDan Williams return -EINVAL; 448831383c68SDan Williams return 0; 448931383c68SDan Williams } 449031383c68SDan Williams 449105ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 449205ea8860SDan Williams { 4493aca78307SMiaohe Lin return huge_page_size(hstate_vma(vma)); 449405ea8860SDan Williams } 449505ea8860SDan Williams 44961da177e4SLinus Torvalds /* 44971da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 44981da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 44996c26d310SMiaohe Lin * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get 45001da177e4SLinus Torvalds * this far. 45011da177e4SLinus Torvalds */ 4502b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 45031da177e4SLinus Torvalds { 45041da177e4SLinus Torvalds BUG(); 4505d0217ac0SNick Piggin return 0; 45061da177e4SLinus Torvalds } 45071da177e4SLinus Torvalds 4508eec3636aSJane Chu /* 4509eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 4510eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 4511eec3636aSJane Chu * This is because under System V memory model, mappings created via 4512eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 4513eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 4514eec3636aSJane Chu */ 4515f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 4516d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 451784afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 4518a1e78772SMel Gorman .close = hugetlb_vm_op_close, 4519dd3b614fSDmitry Safonov .may_split = hugetlb_vm_op_split, 452005ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 45211da177e4SLinus Torvalds }; 45221da177e4SLinus Torvalds 45231e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 45241e8f889bSDavid Gibson int writable) 452563551ae0SDavid Gibson { 452663551ae0SDavid Gibson pte_t entry; 452779c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 452863551ae0SDavid Gibson 45291e8f889bSDavid Gibson if (writable) { 4530106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 4531106c992aSGerald Schaefer vma->vm_page_prot))); 453263551ae0SDavid Gibson } else { 4533106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 4534106c992aSGerald Schaefer vma->vm_page_prot)); 453563551ae0SDavid Gibson } 453663551ae0SDavid Gibson entry = pte_mkyoung(entry); 453763551ae0SDavid Gibson entry = pte_mkhuge(entry); 453879c1c594SChristophe Leroy entry = arch_make_huge_pte(entry, shift, vma->vm_flags); 453963551ae0SDavid Gibson 454063551ae0SDavid Gibson return entry; 454163551ae0SDavid Gibson } 454263551ae0SDavid Gibson 45431e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 45441e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 45451e8f889bSDavid Gibson { 45461e8f889bSDavid Gibson pte_t entry; 45471e8f889bSDavid Gibson 4548106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 454932f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 45504b3073e1SRussell King update_mmu_cache(vma, address, ptep); 45511e8f889bSDavid Gibson } 45521e8f889bSDavid Gibson 4553d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 45544a705fefSNaoya Horiguchi { 45554a705fefSNaoya Horiguchi swp_entry_t swp; 45564a705fefSNaoya Horiguchi 45574a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 4558d5ed7444SAneesh Kumar K.V return false; 45594a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4560d79d176aSBaoquan He if (is_migration_entry(swp)) 4561d5ed7444SAneesh Kumar K.V return true; 45624a705fefSNaoya Horiguchi else 4563d5ed7444SAneesh Kumar K.V return false; 45644a705fefSNaoya Horiguchi } 45654a705fefSNaoya Horiguchi 45663e5c3600SBaoquan He static bool is_hugetlb_entry_hwpoisoned(pte_t pte) 45674a705fefSNaoya Horiguchi { 45684a705fefSNaoya Horiguchi swp_entry_t swp; 45694a705fefSNaoya Horiguchi 45704a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 45713e5c3600SBaoquan He return false; 45724a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4573d79d176aSBaoquan He if (is_hwpoison_entry(swp)) 45743e5c3600SBaoquan He return true; 45754a705fefSNaoya Horiguchi else 45763e5c3600SBaoquan He return false; 45774a705fefSNaoya Horiguchi } 45781e8f889bSDavid Gibson 45794eae4efaSPeter Xu static void 45804eae4efaSPeter Xu hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, 45814eae4efaSPeter Xu struct page *new_page) 45824eae4efaSPeter Xu { 45834eae4efaSPeter Xu __SetPageUptodate(new_page); 45844eae4efaSPeter Xu set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1)); 45854eae4efaSPeter Xu hugepage_add_new_anon_rmap(new_page, vma, addr); 45864eae4efaSPeter Xu hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); 45874eae4efaSPeter Xu ClearHPageRestoreReserve(new_page); 45884eae4efaSPeter Xu SetHPageMigratable(new_page); 45894eae4efaSPeter Xu } 45904eae4efaSPeter Xu 459163551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 459263551ae0SDavid Gibson struct vm_area_struct *vma) 459363551ae0SDavid Gibson { 45945e41540cSMike Kravetz pte_t *src_pte, *dst_pte, entry, dst_entry; 459563551ae0SDavid Gibson struct page *ptepage; 45961c59827dSHugh Dickins unsigned long addr; 4597ca6eb14dSPeter Xu bool cow = is_cow_mapping(vma->vm_flags); 4598a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 4599a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 46004eae4efaSPeter Xu unsigned long npages = pages_per_huge_page(h); 4601c0d0381aSMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 4602ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 4603e8569dd2SAndreas Sandberg int ret = 0; 46041e8f889bSDavid Gibson 4605ac46d4f3SJérôme Glisse if (cow) { 46067269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src, 46076f4f13e8SJérôme Glisse vma->vm_start, 4608ac46d4f3SJérôme Glisse vma->vm_end); 4609ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4610c0d0381aSMike Kravetz } else { 4611c0d0381aSMike Kravetz /* 4612c0d0381aSMike Kravetz * For shared mappings i_mmap_rwsem must be held to call 4613c0d0381aSMike Kravetz * huge_pte_alloc, otherwise the returned ptep could go 4614c0d0381aSMike Kravetz * away if part of a shared pmd and another thread calls 4615c0d0381aSMike Kravetz * huge_pmd_unshare. 4616c0d0381aSMike Kravetz */ 4617c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 4618ac46d4f3SJérôme Glisse } 4619e8569dd2SAndreas Sandberg 4620a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 4621cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 46227868a208SPunit Agrawal src_pte = huge_pte_offset(src, addr, sz); 4623c74df32cSHugh Dickins if (!src_pte) 4624c74df32cSHugh Dickins continue; 4625aec44e0fSPeter Xu dst_pte = huge_pte_alloc(dst, vma, addr, sz); 4626e8569dd2SAndreas Sandberg if (!dst_pte) { 4627e8569dd2SAndreas Sandberg ret = -ENOMEM; 4628e8569dd2SAndreas Sandberg break; 4629e8569dd2SAndreas Sandberg } 4630c5c99429SLarry Woodman 46315e41540cSMike Kravetz /* 46325e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 46335e41540cSMike Kravetz * dst_pte == src_pte is the common case of src/dest sharing. 46345e41540cSMike Kravetz * 46355e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 46365e41540cSMike Kravetz * another vma. If dst_pte !none, this implies sharing. 46375e41540cSMike Kravetz * Check here before taking page table lock, and once again 46385e41540cSMike Kravetz * after taking the lock below. 46395e41540cSMike Kravetz */ 46405e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 46415e41540cSMike Kravetz if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) 4642c5c99429SLarry Woodman continue; 4643c5c99429SLarry Woodman 4644cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 4645cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 4646cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 46474a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 46485e41540cSMike Kravetz dst_entry = huge_ptep_get(dst_pte); 46494eae4efaSPeter Xu again: 46505e41540cSMike Kravetz if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { 46515e41540cSMike Kravetz /* 46525e41540cSMike Kravetz * Skip if src entry none. Also, skip in the 46535e41540cSMike Kravetz * unlikely case dst entry !none as this implies 46545e41540cSMike Kravetz * sharing with another vma. 46555e41540cSMike Kravetz */ 46564a705fefSNaoya Horiguchi ; 46574a705fefSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry) || 46584a705fefSNaoya Horiguchi is_hugetlb_entry_hwpoisoned(entry))) { 46594a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 46604a705fefSNaoya Horiguchi 46614dd845b5SAlistair Popple if (is_writable_migration_entry(swp_entry) && cow) { 46624a705fefSNaoya Horiguchi /* 46634a705fefSNaoya Horiguchi * COW mappings require pages in both 46644a705fefSNaoya Horiguchi * parent and child to be set to read. 46654a705fefSNaoya Horiguchi */ 46664dd845b5SAlistair Popple swp_entry = make_readable_migration_entry( 46674dd845b5SAlistair Popple swp_offset(swp_entry)); 46684a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 4669e5251fd4SPunit Agrawal set_huge_swap_pte_at(src, addr, src_pte, 4670e5251fd4SPunit Agrawal entry, sz); 46714a705fefSNaoya Horiguchi } 4672e5251fd4SPunit Agrawal set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); 46734a705fefSNaoya Horiguchi } else { 46744eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 46754eae4efaSPeter Xu ptepage = pte_page(entry); 46764eae4efaSPeter Xu get_page(ptepage); 46774eae4efaSPeter Xu 46784eae4efaSPeter Xu /* 46794eae4efaSPeter Xu * This is a rare case where we see pinned hugetlb 46804eae4efaSPeter Xu * pages while they're prone to COW. We need to do the 46814eae4efaSPeter Xu * COW earlier during fork. 46824eae4efaSPeter Xu * 46834eae4efaSPeter Xu * When pre-allocating the page or copying data, we 46844eae4efaSPeter Xu * need to be without the pgtable locks since we could 46854eae4efaSPeter Xu * sleep during the process. 46864eae4efaSPeter Xu */ 46874eae4efaSPeter Xu if (unlikely(page_needs_cow_for_dma(vma, ptepage))) { 46884eae4efaSPeter Xu pte_t src_pte_old = entry; 46894eae4efaSPeter Xu struct page *new; 46904eae4efaSPeter Xu 46914eae4efaSPeter Xu spin_unlock(src_ptl); 46924eae4efaSPeter Xu spin_unlock(dst_ptl); 46934eae4efaSPeter Xu /* Do not use reserve as it's private owned */ 46944eae4efaSPeter Xu new = alloc_huge_page(vma, addr, 1); 46954eae4efaSPeter Xu if (IS_ERR(new)) { 46964eae4efaSPeter Xu put_page(ptepage); 46974eae4efaSPeter Xu ret = PTR_ERR(new); 46984eae4efaSPeter Xu break; 46994eae4efaSPeter Xu } 47004eae4efaSPeter Xu copy_user_huge_page(new, ptepage, addr, vma, 47014eae4efaSPeter Xu npages); 47024eae4efaSPeter Xu put_page(ptepage); 47034eae4efaSPeter Xu 47044eae4efaSPeter Xu /* Install the new huge page if src pte stable */ 47054eae4efaSPeter Xu dst_ptl = huge_pte_lock(h, dst, dst_pte); 47064eae4efaSPeter Xu src_ptl = huge_pte_lockptr(h, src, src_pte); 47074eae4efaSPeter Xu spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 47084eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 47094eae4efaSPeter Xu if (!pte_same(src_pte_old, entry)) { 4710846be085SMike Kravetz restore_reserve_on_error(h, vma, addr, 4711846be085SMike Kravetz new); 47124eae4efaSPeter Xu put_page(new); 47134eae4efaSPeter Xu /* dst_entry won't change as in child */ 47144eae4efaSPeter Xu goto again; 47154eae4efaSPeter Xu } 47164eae4efaSPeter Xu hugetlb_install_page(vma, dst_pte, addr, new); 47174eae4efaSPeter Xu spin_unlock(src_ptl); 47184eae4efaSPeter Xu spin_unlock(dst_ptl); 47194eae4efaSPeter Xu continue; 47204eae4efaSPeter Xu } 47214eae4efaSPeter Xu 472234ee645eSJoerg Roedel if (cow) { 47230f10851eSJérôme Glisse /* 47240f10851eSJérôme Glisse * No need to notify as we are downgrading page 47250f10851eSJérôme Glisse * table protection not changing it to point 47260f10851eSJérôme Glisse * to a new page. 47270f10851eSJérôme Glisse * 4728ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 47290f10851eSJérôme Glisse */ 47307f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 473184894e1cSPeter Xu entry = huge_pte_wrprotect(entry); 473234ee645eSJoerg Roedel } 47334eae4efaSPeter Xu 473453f9263bSKirill A. Shutemov page_dup_rmap(ptepage, true); 473563551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 47364eae4efaSPeter Xu hugetlb_count_add(npages, dst); 47371c59827dSHugh Dickins } 4738cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 4739cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 474063551ae0SDavid Gibson } 474163551ae0SDavid Gibson 4742e8569dd2SAndreas Sandberg if (cow) 4743ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 4744c0d0381aSMike Kravetz else 4745c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 4746e8569dd2SAndreas Sandberg 4747e8569dd2SAndreas Sandberg return ret; 474863551ae0SDavid Gibson } 474963551ae0SDavid Gibson 4750550a7d60SMina Almasry static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr, 4751550a7d60SMina Almasry unsigned long new_addr, pte_t *src_pte) 4752550a7d60SMina Almasry { 4753550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 4754550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 4755550a7d60SMina Almasry pte_t *dst_pte, pte; 4756550a7d60SMina Almasry spinlock_t *src_ptl, *dst_ptl; 4757550a7d60SMina Almasry 4758550a7d60SMina Almasry dst_pte = huge_pte_offset(mm, new_addr, huge_page_size(h)); 4759550a7d60SMina Almasry dst_ptl = huge_pte_lock(h, mm, dst_pte); 4760550a7d60SMina Almasry src_ptl = huge_pte_lockptr(h, mm, src_pte); 4761550a7d60SMina Almasry 4762550a7d60SMina Almasry /* 4763550a7d60SMina Almasry * We don't have to worry about the ordering of src and dst ptlocks 4764550a7d60SMina Almasry * because exclusive mmap_sem (or the i_mmap_lock) prevents deadlock. 4765550a7d60SMina Almasry */ 4766550a7d60SMina Almasry if (src_ptl != dst_ptl) 4767550a7d60SMina Almasry spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 4768550a7d60SMina Almasry 4769550a7d60SMina Almasry pte = huge_ptep_get_and_clear(mm, old_addr, src_pte); 4770550a7d60SMina Almasry set_huge_pte_at(mm, new_addr, dst_pte, pte); 4771550a7d60SMina Almasry 4772550a7d60SMina Almasry if (src_ptl != dst_ptl) 4773550a7d60SMina Almasry spin_unlock(src_ptl); 4774550a7d60SMina Almasry spin_unlock(dst_ptl); 4775550a7d60SMina Almasry } 4776550a7d60SMina Almasry 4777550a7d60SMina Almasry int move_hugetlb_page_tables(struct vm_area_struct *vma, 4778550a7d60SMina Almasry struct vm_area_struct *new_vma, 4779550a7d60SMina Almasry unsigned long old_addr, unsigned long new_addr, 4780550a7d60SMina Almasry unsigned long len) 4781550a7d60SMina Almasry { 4782550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 4783550a7d60SMina Almasry struct address_space *mapping = vma->vm_file->f_mapping; 4784550a7d60SMina Almasry unsigned long sz = huge_page_size(h); 4785550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 4786550a7d60SMina Almasry unsigned long old_end = old_addr + len; 4787550a7d60SMina Almasry unsigned long old_addr_copy; 4788550a7d60SMina Almasry pte_t *src_pte, *dst_pte; 4789550a7d60SMina Almasry struct mmu_notifier_range range; 4790550a7d60SMina Almasry 4791550a7d60SMina Almasry mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, old_addr, 4792550a7d60SMina Almasry old_end); 4793550a7d60SMina Almasry adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 4794550a7d60SMina Almasry mmu_notifier_invalidate_range_start(&range); 4795550a7d60SMina Almasry /* Prevent race with file truncation */ 4796550a7d60SMina Almasry i_mmap_lock_write(mapping); 4797550a7d60SMina Almasry for (; old_addr < old_end; old_addr += sz, new_addr += sz) { 4798550a7d60SMina Almasry src_pte = huge_pte_offset(mm, old_addr, sz); 4799550a7d60SMina Almasry if (!src_pte) 4800550a7d60SMina Almasry continue; 4801550a7d60SMina Almasry if (huge_pte_none(huge_ptep_get(src_pte))) 4802550a7d60SMina Almasry continue; 4803550a7d60SMina Almasry 4804550a7d60SMina Almasry /* old_addr arg to huge_pmd_unshare() is a pointer and so the 4805550a7d60SMina Almasry * arg may be modified. Pass a copy instead to preserve the 4806550a7d60SMina Almasry * value in old_addr. 4807550a7d60SMina Almasry */ 4808550a7d60SMina Almasry old_addr_copy = old_addr; 4809550a7d60SMina Almasry 4810550a7d60SMina Almasry if (huge_pmd_unshare(mm, vma, &old_addr_copy, src_pte)) 4811550a7d60SMina Almasry continue; 4812550a7d60SMina Almasry 4813550a7d60SMina Almasry dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz); 4814550a7d60SMina Almasry if (!dst_pte) 4815550a7d60SMina Almasry break; 4816550a7d60SMina Almasry 4817550a7d60SMina Almasry move_huge_pte(vma, old_addr, new_addr, src_pte); 4818550a7d60SMina Almasry } 4819550a7d60SMina Almasry i_mmap_unlock_write(mapping); 4820550a7d60SMina Almasry flush_tlb_range(vma, old_end - len, old_end); 4821550a7d60SMina Almasry mmu_notifier_invalidate_range_end(&range); 4822550a7d60SMina Almasry 4823550a7d60SMina Almasry return len + old_addr - old_end; 4824550a7d60SMina Almasry } 4825550a7d60SMina Almasry 482673c54763SPeter Xu static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 482724669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 482824669e58SAneesh Kumar K.V struct page *ref_page) 482963551ae0SDavid Gibson { 483063551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 483163551ae0SDavid Gibson unsigned long address; 4832c7546f8fSDavid Gibson pte_t *ptep; 483363551ae0SDavid Gibson pte_t pte; 4834cb900f41SKirill A. Shutemov spinlock_t *ptl; 483563551ae0SDavid Gibson struct page *page; 4836a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 4837a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 4838ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 4839a5516438SAndi Kleen 484063551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 4841a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 4842a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 484363551ae0SDavid Gibson 484407e32661SAneesh Kumar K.V /* 484507e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 484607e32661SAneesh Kumar K.V * to huge page. 484707e32661SAneesh Kumar K.V */ 4848ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 484924669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 4850dff11abeSMike Kravetz 4851dff11abeSMike Kravetz /* 4852dff11abeSMike Kravetz * If sharing possible, alert mmu notifiers of worst case. 4853dff11abeSMike Kravetz */ 48546f4f13e8SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, 48556f4f13e8SJérôme Glisse end); 4856ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 4857ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4858569f48b8SHillf Danton address = start; 4859569f48b8SHillf Danton for (; address < end; address += sz) { 48607868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, sz); 4861c7546f8fSDavid Gibson if (!ptep) 4862c7546f8fSDavid Gibson continue; 4863c7546f8fSDavid Gibson 4864cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 486534ae204fSMike Kravetz if (huge_pmd_unshare(mm, vma, &address, ptep)) { 486631d49da5SAneesh Kumar K.V spin_unlock(ptl); 4867dff11abeSMike Kravetz /* 4868dff11abeSMike Kravetz * We just unmapped a page of PMDs by clearing a PUD. 4869dff11abeSMike Kravetz * The caller's TLB flush range should cover this area. 4870dff11abeSMike Kravetz */ 487131d49da5SAneesh Kumar K.V continue; 487231d49da5SAneesh Kumar K.V } 487339dde65cSChen, Kenneth W 48746629326bSHillf Danton pte = huge_ptep_get(ptep); 487531d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 487631d49da5SAneesh Kumar K.V spin_unlock(ptl); 487731d49da5SAneesh Kumar K.V continue; 487831d49da5SAneesh Kumar K.V } 48796629326bSHillf Danton 48806629326bSHillf Danton /* 48819fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 48829fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 48836629326bSHillf Danton */ 48849fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 48859386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 488631d49da5SAneesh Kumar K.V spin_unlock(ptl); 488731d49da5SAneesh Kumar K.V continue; 48888c4894c6SNaoya Horiguchi } 48896629326bSHillf Danton 48906629326bSHillf Danton page = pte_page(pte); 489104f2cbe3SMel Gorman /* 489204f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 489304f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 489404f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 489504f2cbe3SMel Gorman */ 489604f2cbe3SMel Gorman if (ref_page) { 489731d49da5SAneesh Kumar K.V if (page != ref_page) { 489831d49da5SAneesh Kumar K.V spin_unlock(ptl); 489931d49da5SAneesh Kumar K.V continue; 490031d49da5SAneesh Kumar K.V } 490104f2cbe3SMel Gorman /* 490204f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 490304f2cbe3SMel Gorman * future faults in this VMA will fail rather than 490404f2cbe3SMel Gorman * looking like data was lost 490504f2cbe3SMel Gorman */ 490604f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 490704f2cbe3SMel Gorman } 490804f2cbe3SMel Gorman 4909c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 4910b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 4911106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 49126649a386SKen Chen set_page_dirty(page); 49139e81130bSHillf Danton 49145d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 4915d281ee61SKirill A. Shutemov page_remove_rmap(page, true); 491631d49da5SAneesh Kumar K.V 4917cb900f41SKirill A. Shutemov spin_unlock(ptl); 4918e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 491924669e58SAneesh Kumar K.V /* 492031d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 492124669e58SAneesh Kumar K.V */ 492231d49da5SAneesh Kumar K.V if (ref_page) 492331d49da5SAneesh Kumar K.V break; 4924fe1668aeSChen, Kenneth W } 4925ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 492624669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 49271da177e4SLinus Torvalds } 492863551ae0SDavid Gibson 4929d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 4930d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 4931d833352aSMel Gorman unsigned long end, struct page *ref_page) 4932d833352aSMel Gorman { 4933d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 4934d833352aSMel Gorman 4935d833352aSMel Gorman /* 4936d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 4937d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 4938d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 4939d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 4940d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 4941c8c06efaSDavidlohr Bueso * is to clear it before releasing the i_mmap_rwsem. This works 4942d833352aSMel Gorman * because in the context this is called, the VMA is about to be 4943c8c06efaSDavidlohr Bueso * destroyed and the i_mmap_rwsem is held. 4944d833352aSMel Gorman */ 4945d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 4946d833352aSMel Gorman } 4947d833352aSMel Gorman 4948502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 494904f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 4950502717f4SChen, Kenneth W { 495124669e58SAneesh Kumar K.V struct mmu_gather tlb; 4952dff11abeSMike Kravetz 4953a72afd87SWill Deacon tlb_gather_mmu(&tlb, vma->vm_mm); 495424669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 4955ae8eba8bSWill Deacon tlb_finish_mmu(&tlb); 4956502717f4SChen, Kenneth W } 4957502717f4SChen, Kenneth W 495804f2cbe3SMel Gorman /* 495904f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 4960578b7725SZhiyuan Dai * mapping it owns the reserve page for. The intention is to unmap the page 496104f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 496204f2cbe3SMel Gorman * same region. 496304f2cbe3SMel Gorman */ 49642f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 49652a4b3dedSHarvey Harrison struct page *page, unsigned long address) 496604f2cbe3SMel Gorman { 49677526674dSAdam Litke struct hstate *h = hstate_vma(vma); 496804f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 496904f2cbe3SMel Gorman struct address_space *mapping; 497004f2cbe3SMel Gorman pgoff_t pgoff; 497104f2cbe3SMel Gorman 497204f2cbe3SMel Gorman /* 497304f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 497404f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 497504f2cbe3SMel Gorman */ 49767526674dSAdam Litke address = address & huge_page_mask(h); 497736e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 497836e4f20aSMichal Hocko vma->vm_pgoff; 497993c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 498004f2cbe3SMel Gorman 49814eb2b1dcSMel Gorman /* 49824eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 49834eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 49844eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 49854eb2b1dcSMel Gorman */ 498683cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 49876b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 498804f2cbe3SMel Gorman /* Do not unmap the current VMA */ 498904f2cbe3SMel Gorman if (iter_vma == vma) 499004f2cbe3SMel Gorman continue; 499104f2cbe3SMel Gorman 499204f2cbe3SMel Gorman /* 49932f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 49942f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 49952f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 49962f84a899SMel Gorman */ 49972f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 49982f84a899SMel Gorman continue; 49992f84a899SMel Gorman 50002f84a899SMel Gorman /* 500104f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 500204f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 500304f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 500404f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 500504f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 500604f2cbe3SMel Gorman */ 500704f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 500824669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 500924669e58SAneesh Kumar K.V address + huge_page_size(h), page); 501004f2cbe3SMel Gorman } 501183cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 501204f2cbe3SMel Gorman } 501304f2cbe3SMel Gorman 50140fe6e20bSNaoya Horiguchi /* 50150fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 5016aa6d2e8cSBaolin Wang * Called with hugetlb_fault_mutex_table held and pte_page locked so we 5017ef009b25SMichal Hocko * cannot race with other handlers or page migration. 5018ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 50190fe6e20bSNaoya Horiguchi */ 50202b740303SSouptick Joarder static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 5021974e6d66SHuang Ying unsigned long address, pte_t *ptep, 5022cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 50231e8f889bSDavid Gibson { 50243999f52eSAneesh Kumar K.V pte_t pte; 5025a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 50261e8f889bSDavid Gibson struct page *old_page, *new_page; 50272b740303SSouptick Joarder int outside_reserve = 0; 50282b740303SSouptick Joarder vm_fault_t ret = 0; 5029974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 5030ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 50311e8f889bSDavid Gibson 50323999f52eSAneesh Kumar K.V pte = huge_ptep_get(ptep); 50331e8f889bSDavid Gibson old_page = pte_page(pte); 50341e8f889bSDavid Gibson 503504f2cbe3SMel Gorman retry_avoidcopy: 50361e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 50371e8f889bSDavid Gibson * and just make the page writable */ 503837a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 50395a49973dSHugh Dickins page_move_anon_rmap(old_page, vma); 50405b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 504183c54070SNick Piggin return 0; 50421e8f889bSDavid Gibson } 50431e8f889bSDavid Gibson 504404f2cbe3SMel Gorman /* 504504f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 504604f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 504704f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 504804f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 504904f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 505004f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 505104f2cbe3SMel Gorman * of the full address range. 505204f2cbe3SMel Gorman */ 50535944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 505404f2cbe3SMel Gorman old_page != pagecache_page) 505504f2cbe3SMel Gorman outside_reserve = 1; 505604f2cbe3SMel Gorman 505709cbfeafSKirill A. Shutemov get_page(old_page); 5058b76c8cfbSLarry Woodman 5059ad4404a2SDavidlohr Bueso /* 5060ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 5061ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 5062ad4404a2SDavidlohr Bueso */ 5063cb900f41SKirill A. Shutemov spin_unlock(ptl); 50645b7a1d40SHuang Ying new_page = alloc_huge_page(vma, haddr, outside_reserve); 50651e8f889bSDavid Gibson 50662fc39cecSAdam Litke if (IS_ERR(new_page)) { 506704f2cbe3SMel Gorman /* 506804f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 506904f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 507004f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 507104f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 507204f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 507304f2cbe3SMel Gorman */ 507404f2cbe3SMel Gorman if (outside_reserve) { 5075e7dd91c4SMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 5076e7dd91c4SMike Kravetz pgoff_t idx; 5077e7dd91c4SMike Kravetz u32 hash; 5078e7dd91c4SMike Kravetz 507909cbfeafSKirill A. Shutemov put_page(old_page); 508004f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 5081e7dd91c4SMike Kravetz /* 5082e7dd91c4SMike Kravetz * Drop hugetlb_fault_mutex and i_mmap_rwsem before 5083e7dd91c4SMike Kravetz * unmapping. unmapping needs to hold i_mmap_rwsem 5084e7dd91c4SMike Kravetz * in write mode. Dropping i_mmap_rwsem in read mode 5085e7dd91c4SMike Kravetz * here is OK as COW mappings do not interact with 5086e7dd91c4SMike Kravetz * PMD sharing. 5087e7dd91c4SMike Kravetz * 5088e7dd91c4SMike Kravetz * Reacquire both after unmap operation. 5089e7dd91c4SMike Kravetz */ 5090e7dd91c4SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 5091e7dd91c4SMike Kravetz hash = hugetlb_fault_mutex_hash(mapping, idx); 5092e7dd91c4SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5093e7dd91c4SMike Kravetz i_mmap_unlock_read(mapping); 5094e7dd91c4SMike Kravetz 50955b7a1d40SHuang Ying unmap_ref_private(mm, vma, old_page, haddr); 5096e7dd91c4SMike Kravetz 5097e7dd91c4SMike Kravetz i_mmap_lock_read(mapping); 5098e7dd91c4SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 5099cb900f41SKirill A. Shutemov spin_lock(ptl); 51005b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 5101a9af0c5dSNaoya Horiguchi if (likely(ptep && 5102a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 510304f2cbe3SMel Gorman goto retry_avoidcopy; 5104a734bcc8SHillf Danton /* 5105cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 5106cb900f41SKirill A. Shutemov * lock, and our job is done. 5107a734bcc8SHillf Danton */ 5108a734bcc8SHillf Danton return 0; 510904f2cbe3SMel Gorman } 511004f2cbe3SMel Gorman 51112b740303SSouptick Joarder ret = vmf_error(PTR_ERR(new_page)); 5112ad4404a2SDavidlohr Bueso goto out_release_old; 51131e8f889bSDavid Gibson } 51141e8f889bSDavid Gibson 51150fe6e20bSNaoya Horiguchi /* 51160fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 51170fe6e20bSNaoya Horiguchi * anon_vma prepared. 51180fe6e20bSNaoya Horiguchi */ 511944e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 5120ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 5121ad4404a2SDavidlohr Bueso goto out_release_all; 512244e2aa93SDean Nelson } 51230fe6e20bSNaoya Horiguchi 5124974e6d66SHuang Ying copy_user_huge_page(new_page, old_page, address, vma, 512547ad8475SAndrea Arcangeli pages_per_huge_page(h)); 51260ed361deSNick Piggin __SetPageUptodate(new_page); 51271e8f889bSDavid Gibson 51287269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, 51296f4f13e8SJérôme Glisse haddr + huge_page_size(h)); 5130ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 5131ad4404a2SDavidlohr Bueso 5132b76c8cfbSLarry Woodman /* 5133cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 5134b76c8cfbSLarry Woodman * before the page tables are altered 5135b76c8cfbSLarry Woodman */ 5136cb900f41SKirill A. Shutemov spin_lock(ptl); 51375b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 5138a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 5139d6995da3SMike Kravetz ClearHPageRestoreReserve(new_page); 514007443a85SJoonsoo Kim 51411e8f889bSDavid Gibson /* Break COW */ 51425b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 5143ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, range.end); 51445b7a1d40SHuang Ying set_huge_pte_at(mm, haddr, ptep, 51451e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 5146d281ee61SKirill A. Shutemov page_remove_rmap(old_page, true); 51475b7a1d40SHuang Ying hugepage_add_new_anon_rmap(new_page, vma, haddr); 51488f251a3dSMike Kravetz SetHPageMigratable(new_page); 51491e8f889bSDavid Gibson /* Make the old page be freed below */ 51501e8f889bSDavid Gibson new_page = old_page; 51511e8f889bSDavid Gibson } 5152cb900f41SKirill A. Shutemov spin_unlock(ptl); 5153ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 5154ad4404a2SDavidlohr Bueso out_release_all: 5155c7b1850dSMike Kravetz /* No restore in case of successful pagetable update (Break COW) */ 5156c7b1850dSMike Kravetz if (new_page != old_page) 51575b7a1d40SHuang Ying restore_reserve_on_error(h, vma, haddr, new_page); 515809cbfeafSKirill A. Shutemov put_page(new_page); 5159ad4404a2SDavidlohr Bueso out_release_old: 516009cbfeafSKirill A. Shutemov put_page(old_page); 51618312034fSJoonsoo Kim 5162ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 5163ad4404a2SDavidlohr Bueso return ret; 51641e8f889bSDavid Gibson } 51651e8f889bSDavid Gibson 516604f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 5167a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 5168a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 516904f2cbe3SMel Gorman { 517004f2cbe3SMel Gorman struct address_space *mapping; 5171e7c4b0bfSAndy Whitcroft pgoff_t idx; 517204f2cbe3SMel Gorman 517304f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 5174a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 517504f2cbe3SMel Gorman 517604f2cbe3SMel Gorman return find_lock_page(mapping, idx); 517704f2cbe3SMel Gorman } 517804f2cbe3SMel Gorman 51793ae77f43SHugh Dickins /* 51803ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 51813ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 51823ae77f43SHugh Dickins */ 51833ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 51842a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 51852a15efc9SHugh Dickins { 51862a15efc9SHugh Dickins struct address_space *mapping; 51872a15efc9SHugh Dickins pgoff_t idx; 51882a15efc9SHugh Dickins struct page *page; 51892a15efc9SHugh Dickins 51902a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 51912a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 51922a15efc9SHugh Dickins 51932a15efc9SHugh Dickins page = find_get_page(mapping, idx); 51942a15efc9SHugh Dickins if (page) 51952a15efc9SHugh Dickins put_page(page); 51962a15efc9SHugh Dickins return page != NULL; 51972a15efc9SHugh Dickins } 51982a15efc9SHugh Dickins 5199ab76ad54SMike Kravetz int huge_add_to_page_cache(struct page *page, struct address_space *mapping, 5200ab76ad54SMike Kravetz pgoff_t idx) 5201ab76ad54SMike Kravetz { 5202ab76ad54SMike Kravetz struct inode *inode = mapping->host; 5203ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 5204ab76ad54SMike Kravetz int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 5205ab76ad54SMike Kravetz 5206ab76ad54SMike Kravetz if (err) 5207ab76ad54SMike Kravetz return err; 5208d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 5209ab76ad54SMike Kravetz 521022146c3cSMike Kravetz /* 521122146c3cSMike Kravetz * set page dirty so that it will not be removed from cache/file 521222146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 521322146c3cSMike Kravetz */ 521422146c3cSMike Kravetz set_page_dirty(page); 521522146c3cSMike Kravetz 5216ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 5217ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 5218ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 5219ab76ad54SMike Kravetz return 0; 5220ab76ad54SMike Kravetz } 5221ab76ad54SMike Kravetz 52227677f7fdSAxel Rasmussen static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma, 52237677f7fdSAxel Rasmussen struct address_space *mapping, 52247677f7fdSAxel Rasmussen pgoff_t idx, 52257677f7fdSAxel Rasmussen unsigned int flags, 52267677f7fdSAxel Rasmussen unsigned long haddr, 52277677f7fdSAxel Rasmussen unsigned long reason) 52287677f7fdSAxel Rasmussen { 52297677f7fdSAxel Rasmussen vm_fault_t ret; 52307677f7fdSAxel Rasmussen u32 hash; 52317677f7fdSAxel Rasmussen struct vm_fault vmf = { 52327677f7fdSAxel Rasmussen .vma = vma, 52337677f7fdSAxel Rasmussen .address = haddr, 52347677f7fdSAxel Rasmussen .flags = flags, 52357677f7fdSAxel Rasmussen 52367677f7fdSAxel Rasmussen /* 52377677f7fdSAxel Rasmussen * Hard to debug if it ends up being 52387677f7fdSAxel Rasmussen * used by a callee that assumes 52397677f7fdSAxel Rasmussen * something about the other 52407677f7fdSAxel Rasmussen * uninitialized fields... same as in 52417677f7fdSAxel Rasmussen * memory.c 52427677f7fdSAxel Rasmussen */ 52437677f7fdSAxel Rasmussen }; 52447677f7fdSAxel Rasmussen 52457677f7fdSAxel Rasmussen /* 52467677f7fdSAxel Rasmussen * hugetlb_fault_mutex and i_mmap_rwsem must be 52477677f7fdSAxel Rasmussen * dropped before handling userfault. Reacquire 52487677f7fdSAxel Rasmussen * after handling fault to make calling code simpler. 52497677f7fdSAxel Rasmussen */ 52507677f7fdSAxel Rasmussen hash = hugetlb_fault_mutex_hash(mapping, idx); 52517677f7fdSAxel Rasmussen mutex_unlock(&hugetlb_fault_mutex_table[hash]); 52527677f7fdSAxel Rasmussen i_mmap_unlock_read(mapping); 52537677f7fdSAxel Rasmussen ret = handle_userfault(&vmf, reason); 52547677f7fdSAxel Rasmussen i_mmap_lock_read(mapping); 52557677f7fdSAxel Rasmussen mutex_lock(&hugetlb_fault_mutex_table[hash]); 52567677f7fdSAxel Rasmussen 52577677f7fdSAxel Rasmussen return ret; 52587677f7fdSAxel Rasmussen } 52597677f7fdSAxel Rasmussen 52602b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 52612b740303SSouptick Joarder struct vm_area_struct *vma, 52628382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 5263788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 5264ac9b9c66SHugh Dickins { 5265a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 52662b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 5267409eb8c2SHillf Danton int anon_rmap = 0; 52684c887265SAdam Litke unsigned long size; 52694c887265SAdam Litke struct page *page; 52701e8f889bSDavid Gibson pte_t new_pte; 5271cb900f41SKirill A. Shutemov spinlock_t *ptl; 5272285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 5273c7b1850dSMike Kravetz bool new_page, new_pagecache_page = false; 52744c887265SAdam Litke 527504f2cbe3SMel Gorman /* 527604f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 527704f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 527825985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 527904f2cbe3SMel Gorman */ 528004f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 5281910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 528204f2cbe3SMel Gorman current->pid); 528304f2cbe3SMel Gorman return ret; 528404f2cbe3SMel Gorman } 528504f2cbe3SMel Gorman 52864c887265SAdam Litke /* 528787bf91d3SMike Kravetz * We can not race with truncation due to holding i_mmap_rwsem. 528887bf91d3SMike Kravetz * i_size is modified when holding i_mmap_rwsem, so check here 528987bf91d3SMike Kravetz * once for faults beyond end of file. 52904c887265SAdam Litke */ 5291a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 5292ebed4bfcSHugh Dickins if (idx >= size) 5293ebed4bfcSHugh Dickins goto out; 52941a1aad8aSMike Kravetz 529587bf91d3SMike Kravetz retry: 5296c7b1850dSMike Kravetz new_page = false; 529787bf91d3SMike Kravetz page = find_lock_page(mapping, idx); 529887bf91d3SMike Kravetz if (!page) { 52997677f7fdSAxel Rasmussen /* Check for page in userfault range */ 53001a1aad8aSMike Kravetz if (userfaultfd_missing(vma)) { 53017677f7fdSAxel Rasmussen ret = hugetlb_handle_userfault(vma, mapping, idx, 53027677f7fdSAxel Rasmussen flags, haddr, 53037677f7fdSAxel Rasmussen VM_UFFD_MISSING); 53041a1aad8aSMike Kravetz goto out; 53051a1aad8aSMike Kravetz } 53061a1aad8aSMike Kravetz 5307285b8dcaSHuang Ying page = alloc_huge_page(vma, haddr, 0); 53082fc39cecSAdam Litke if (IS_ERR(page)) { 53094643d67eSMike Kravetz /* 53104643d67eSMike Kravetz * Returning error will result in faulting task being 53114643d67eSMike Kravetz * sent SIGBUS. The hugetlb fault mutex prevents two 53124643d67eSMike Kravetz * tasks from racing to fault in the same page which 53134643d67eSMike Kravetz * could result in false unable to allocate errors. 53144643d67eSMike Kravetz * Page migration does not take the fault mutex, but 53154643d67eSMike Kravetz * does a clear then write of pte's under page table 53164643d67eSMike Kravetz * lock. Page fault code could race with migration, 53174643d67eSMike Kravetz * notice the clear pte and try to allocate a page 53184643d67eSMike Kravetz * here. Before returning error, get ptl and make 53194643d67eSMike Kravetz * sure there really is no pte entry. 53204643d67eSMike Kravetz */ 53214643d67eSMike Kravetz ptl = huge_pte_lock(h, mm, ptep); 53224643d67eSMike Kravetz ret = 0; 5323d83e6c8aSMiaohe Lin if (huge_pte_none(huge_ptep_get(ptep))) 53242b740303SSouptick Joarder ret = vmf_error(PTR_ERR(page)); 5325d83e6c8aSMiaohe Lin spin_unlock(ptl); 53266bda666aSChristoph Lameter goto out; 53276bda666aSChristoph Lameter } 532847ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 53290ed361deSNick Piggin __SetPageUptodate(page); 5330cb6acd01SMike Kravetz new_page = true; 5331ac9b9c66SHugh Dickins 5332f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 5333ab76ad54SMike Kravetz int err = huge_add_to_page_cache(page, mapping, idx); 53346bda666aSChristoph Lameter if (err) { 53356bda666aSChristoph Lameter put_page(page); 53366bda666aSChristoph Lameter if (err == -EEXIST) 53376bda666aSChristoph Lameter goto retry; 53386bda666aSChristoph Lameter goto out; 53396bda666aSChristoph Lameter } 5340c7b1850dSMike Kravetz new_pagecache_page = true; 534123be7468SMel Gorman } else { 53426bda666aSChristoph Lameter lock_page(page); 53430fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 53440fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 53450fe6e20bSNaoya Horiguchi goto backout_unlocked; 534623be7468SMel Gorman } 5347409eb8c2SHillf Danton anon_rmap = 1; 53480fe6e20bSNaoya Horiguchi } 53490fe6e20bSNaoya Horiguchi } else { 535057303d80SAndy Whitcroft /* 5351998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 5352998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 5353998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 5354fd6a03edSNaoya Horiguchi */ 5355fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 53560eb98f15SMiaohe Lin ret = VM_FAULT_HWPOISON_LARGE | 5357972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 5358fd6a03edSNaoya Horiguchi goto backout_unlocked; 53596bda666aSChristoph Lameter } 53607677f7fdSAxel Rasmussen 53617677f7fdSAxel Rasmussen /* Check for page in userfault range. */ 53627677f7fdSAxel Rasmussen if (userfaultfd_minor(vma)) { 53637677f7fdSAxel Rasmussen unlock_page(page); 53647677f7fdSAxel Rasmussen put_page(page); 53657677f7fdSAxel Rasmussen ret = hugetlb_handle_userfault(vma, mapping, idx, 53667677f7fdSAxel Rasmussen flags, haddr, 53677677f7fdSAxel Rasmussen VM_UFFD_MINOR); 53687677f7fdSAxel Rasmussen goto out; 53697677f7fdSAxel Rasmussen } 5370998b4382SNaoya Horiguchi } 53711e8f889bSDavid Gibson 537257303d80SAndy Whitcroft /* 537357303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 537457303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 537557303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 537657303d80SAndy Whitcroft * the spinlock. 537757303d80SAndy Whitcroft */ 53785e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 5379285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 53802b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 53812b26736cSAndy Whitcroft goto backout_unlocked; 53822b26736cSAndy Whitcroft } 53835e911373SMike Kravetz /* Just decrements count, does not deallocate */ 5384285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 53855e911373SMike Kravetz } 538657303d80SAndy Whitcroft 53878bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 538883c54070SNick Piggin ret = 0; 53897f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 53904c887265SAdam Litke goto backout; 53914c887265SAdam Litke 539207443a85SJoonsoo Kim if (anon_rmap) { 5393d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 5394285b8dcaSHuang Ying hugepage_add_new_anon_rmap(page, vma, haddr); 5395ac714904SChoi Gi-yong } else 539653f9263bSKirill A. Shutemov page_dup_rmap(page, true); 53971e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 53981e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 5399285b8dcaSHuang Ying set_huge_pte_at(mm, haddr, ptep, new_pte); 54001e8f889bSDavid Gibson 54015d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 5402788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 54031e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 5404974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); 54051e8f889bSDavid Gibson } 54061e8f889bSDavid Gibson 5407cb900f41SKirill A. Shutemov spin_unlock(ptl); 5408cb6acd01SMike Kravetz 5409cb6acd01SMike Kravetz /* 54108f251a3dSMike Kravetz * Only set HPageMigratable in newly allocated pages. Existing pages 54118f251a3dSMike Kravetz * found in the pagecache may not have HPageMigratableset if they have 54128f251a3dSMike Kravetz * been isolated for migration. 5413cb6acd01SMike Kravetz */ 5414cb6acd01SMike Kravetz if (new_page) 54158f251a3dSMike Kravetz SetHPageMigratable(page); 5416cb6acd01SMike Kravetz 54174c887265SAdam Litke unlock_page(page); 54184c887265SAdam Litke out: 5419ac9b9c66SHugh Dickins return ret; 54204c887265SAdam Litke 54214c887265SAdam Litke backout: 5422cb900f41SKirill A. Shutemov spin_unlock(ptl); 54232b26736cSAndy Whitcroft backout_unlocked: 54244c887265SAdam Litke unlock_page(page); 5425c7b1850dSMike Kravetz /* restore reserve for newly allocated pages not in page cache */ 5426c7b1850dSMike Kravetz if (new_page && !new_pagecache_page) 5427285b8dcaSHuang Ying restore_reserve_on_error(h, vma, haddr, page); 54284c887265SAdam Litke put_page(page); 54294c887265SAdam Litke goto out; 5430ac9b9c66SHugh Dickins } 5431ac9b9c66SHugh Dickins 54328382d914SDavidlohr Bueso #ifdef CONFIG_SMP 5433188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 54348382d914SDavidlohr Bueso { 54358382d914SDavidlohr Bueso unsigned long key[2]; 54368382d914SDavidlohr Bueso u32 hash; 54378382d914SDavidlohr Bueso 54388382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 54398382d914SDavidlohr Bueso key[1] = idx; 54408382d914SDavidlohr Bueso 544155254636SMike Kravetz hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); 54428382d914SDavidlohr Bueso 54438382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 54448382d914SDavidlohr Bueso } 54458382d914SDavidlohr Bueso #else 54468382d914SDavidlohr Bueso /* 54476c26d310SMiaohe Lin * For uniprocessor systems we always use a single mutex, so just 54488382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 54498382d914SDavidlohr Bueso */ 5450188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 54518382d914SDavidlohr Bueso { 54528382d914SDavidlohr Bueso return 0; 54538382d914SDavidlohr Bueso } 54548382d914SDavidlohr Bueso #endif 54558382d914SDavidlohr Bueso 54562b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 5457788c7df4SHugh Dickins unsigned long address, unsigned int flags) 545886e5216fSAdam Litke { 54598382d914SDavidlohr Bueso pte_t *ptep, entry; 5460cb900f41SKirill A. Shutemov spinlock_t *ptl; 54612b740303SSouptick Joarder vm_fault_t ret; 54628382d914SDavidlohr Bueso u32 hash; 54638382d914SDavidlohr Bueso pgoff_t idx; 54640fe6e20bSNaoya Horiguchi struct page *page = NULL; 546557303d80SAndy Whitcroft struct page *pagecache_page = NULL; 5466a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 54678382d914SDavidlohr Bueso struct address_space *mapping; 54680f792cf9SNaoya Horiguchi int need_wait_lock = 0; 5469285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 547086e5216fSAdam Litke 5471285b8dcaSHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 5472fd6a03edSNaoya Horiguchi if (ptep) { 5473c0d0381aSMike Kravetz /* 5474c0d0381aSMike Kravetz * Since we hold no locks, ptep could be stale. That is 5475c0d0381aSMike Kravetz * OK as we are only making decisions based on content and 5476c0d0381aSMike Kravetz * not actually modifying content here. 5477c0d0381aSMike Kravetz */ 5478fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 5479290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 5480cb900f41SKirill A. Shutemov migration_entry_wait_huge(vma, mm, ptep); 5481290408d4SNaoya Horiguchi return 0; 5482290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 5483aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 5484972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 5485b43a9990SMike Kravetz } 54868382d914SDavidlohr Bueso 5487c0d0381aSMike Kravetz /* 5488c0d0381aSMike Kravetz * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold 548987bf91d3SMike Kravetz * until finished with ptep. This serves two purposes: 549087bf91d3SMike Kravetz * 1) It prevents huge_pmd_unshare from being called elsewhere 549187bf91d3SMike Kravetz * and making the ptep no longer valid. 549287bf91d3SMike Kravetz * 2) It synchronizes us with i_size modifications during truncation. 5493c0d0381aSMike Kravetz * 5494c0d0381aSMike Kravetz * ptep could have already be assigned via huge_pte_offset. That 5495c0d0381aSMike Kravetz * is OK, as huge_pte_alloc will return the same value unless 5496c0d0381aSMike Kravetz * something has changed. 5497c0d0381aSMike Kravetz */ 5498ddeaab32SMike Kravetz mapping = vma->vm_file->f_mapping; 5499c0d0381aSMike Kravetz i_mmap_lock_read(mapping); 5500aec44e0fSPeter Xu ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); 5501c0d0381aSMike Kravetz if (!ptep) { 5502c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 5503c0d0381aSMike Kravetz return VM_FAULT_OOM; 5504c0d0381aSMike Kravetz } 5505ddeaab32SMike Kravetz 55063935baa9SDavid Gibson /* 55073935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 55083935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 55093935baa9SDavid Gibson * the same page in the page cache. 55103935baa9SDavid Gibson */ 5511c0d0381aSMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 5512188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 5513c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 55148382d914SDavidlohr Bueso 55157f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 55167f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 55178382d914SDavidlohr Bueso ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 5518b4d1d99fSDavid Gibson goto out_mutex; 55193935baa9SDavid Gibson } 552086e5216fSAdam Litke 552183c54070SNick Piggin ret = 0; 55221e8f889bSDavid Gibson 552357303d80SAndy Whitcroft /* 55240f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 55250f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 55267c8de358SEthon Paul * an active hugepage in pagecache. This goto expects the 2nd page 55277c8de358SEthon Paul * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will 55287c8de358SEthon Paul * properly handle it. 55290f792cf9SNaoya Horiguchi */ 55300f792cf9SNaoya Horiguchi if (!pte_present(entry)) 55310f792cf9SNaoya Horiguchi goto out_mutex; 55320f792cf9SNaoya Horiguchi 55330f792cf9SNaoya Horiguchi /* 553457303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 553557303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 553657303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 553757303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 553857303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 553957303d80SAndy Whitcroft * consumed. 554057303d80SAndy Whitcroft */ 5541106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 5542285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 55432b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 5544b4d1d99fSDavid Gibson goto out_mutex; 55452b26736cSAndy Whitcroft } 55465e911373SMike Kravetz /* Just decrements count, does not deallocate */ 5547285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 554857303d80SAndy Whitcroft 5549f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 555057303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 5551285b8dcaSHuang Ying vma, haddr); 555257303d80SAndy Whitcroft } 555357303d80SAndy Whitcroft 55540f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 55550fe6e20bSNaoya Horiguchi 55561e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 5557b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 5558cb900f41SKirill A. Shutemov goto out_ptl; 5559b4d1d99fSDavid Gibson 55600f792cf9SNaoya Horiguchi /* 55610f792cf9SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 55620f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 55630f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 55640f792cf9SNaoya Horiguchi */ 55650f792cf9SNaoya Horiguchi page = pte_page(entry); 55660f792cf9SNaoya Horiguchi if (page != pagecache_page) 55670f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 55680f792cf9SNaoya Horiguchi need_wait_lock = 1; 55690f792cf9SNaoya Horiguchi goto out_ptl; 55700f792cf9SNaoya Horiguchi } 55710f792cf9SNaoya Horiguchi 55720f792cf9SNaoya Horiguchi get_page(page); 5573b4d1d99fSDavid Gibson 5574788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 5575106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 5576974e6d66SHuang Ying ret = hugetlb_cow(mm, vma, address, ptep, 5577cb900f41SKirill A. Shutemov pagecache_page, ptl); 55780f792cf9SNaoya Horiguchi goto out_put_page; 5579b4d1d99fSDavid Gibson } 5580106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 5581b4d1d99fSDavid Gibson } 5582b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 5583285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 5584788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 5585285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 55860f792cf9SNaoya Horiguchi out_put_page: 55870f792cf9SNaoya Horiguchi if (page != pagecache_page) 55880f792cf9SNaoya Horiguchi unlock_page(page); 55890f792cf9SNaoya Horiguchi put_page(page); 5590cb900f41SKirill A. Shutemov out_ptl: 5591cb900f41SKirill A. Shutemov spin_unlock(ptl); 559257303d80SAndy Whitcroft 559357303d80SAndy Whitcroft if (pagecache_page) { 559457303d80SAndy Whitcroft unlock_page(pagecache_page); 559557303d80SAndy Whitcroft put_page(pagecache_page); 559657303d80SAndy Whitcroft } 5597b4d1d99fSDavid Gibson out_mutex: 5598c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5599c0d0381aSMike Kravetz i_mmap_unlock_read(mapping); 56000f792cf9SNaoya Horiguchi /* 56010f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 56020f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 56030f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 56040f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 56050f792cf9SNaoya Horiguchi * here without taking refcount. 56060f792cf9SNaoya Horiguchi */ 56070f792cf9SNaoya Horiguchi if (need_wait_lock) 56080f792cf9SNaoya Horiguchi wait_on_page_locked(page); 56091e8f889bSDavid Gibson return ret; 561086e5216fSAdam Litke } 561186e5216fSAdam Litke 5612714c1891SAxel Rasmussen #ifdef CONFIG_USERFAULTFD 56138fb5debcSMike Kravetz /* 56148fb5debcSMike Kravetz * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with 56158fb5debcSMike Kravetz * modifications for huge pages. 56168fb5debcSMike Kravetz */ 56178fb5debcSMike Kravetz int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, 56188fb5debcSMike Kravetz pte_t *dst_pte, 56198fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 56208fb5debcSMike Kravetz unsigned long dst_addr, 56218fb5debcSMike Kravetz unsigned long src_addr, 5622f6191471SAxel Rasmussen enum mcopy_atomic_mode mode, 56238fb5debcSMike Kravetz struct page **pagep) 56248fb5debcSMike Kravetz { 5625f6191471SAxel Rasmussen bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE); 56268cc5fcbbSMina Almasry struct hstate *h = hstate_vma(dst_vma); 56278cc5fcbbSMina Almasry struct address_space *mapping = dst_vma->vm_file->f_mapping; 56288cc5fcbbSMina Almasry pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr); 56291e392147SAndrea Arcangeli unsigned long size; 56301c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 56318fb5debcSMike Kravetz pte_t _dst_pte; 56328fb5debcSMike Kravetz spinlock_t *ptl; 56338cc5fcbbSMina Almasry int ret = -ENOMEM; 56348fb5debcSMike Kravetz struct page *page; 5635f6191471SAxel Rasmussen int writable; 5636c7b1850dSMike Kravetz bool new_pagecache_page = false; 56378fb5debcSMike Kravetz 5638f6191471SAxel Rasmussen if (is_continue) { 5639f6191471SAxel Rasmussen ret = -EFAULT; 5640f6191471SAxel Rasmussen page = find_lock_page(mapping, idx); 5641f6191471SAxel Rasmussen if (!page) 5642f6191471SAxel Rasmussen goto out; 5643f6191471SAxel Rasmussen } else if (!*pagep) { 5644d84cf06eSMina Almasry /* If a page already exists, then it's UFFDIO_COPY for 5645d84cf06eSMina Almasry * a non-missing case. Return -EEXIST. 5646d84cf06eSMina Almasry */ 5647d84cf06eSMina Almasry if (vm_shared && 5648d84cf06eSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 5649d84cf06eSMina Almasry ret = -EEXIST; 56508fb5debcSMike Kravetz goto out; 5651d84cf06eSMina Almasry } 5652d84cf06eSMina Almasry 5653d84cf06eSMina Almasry page = alloc_huge_page(dst_vma, dst_addr, 0); 5654d84cf06eSMina Almasry if (IS_ERR(page)) { 5655d84cf06eSMina Almasry ret = -ENOMEM; 5656d84cf06eSMina Almasry goto out; 5657d84cf06eSMina Almasry } 56588fb5debcSMike Kravetz 56598fb5debcSMike Kravetz ret = copy_huge_page_from_user(page, 56608fb5debcSMike Kravetz (const void __user *) src_addr, 5661810a56b9SMike Kravetz pages_per_huge_page(h), false); 56628fb5debcSMike Kravetz 5663c1e8d7c6SMichel Lespinasse /* fallback to copy_from_user outside mmap_lock */ 56648fb5debcSMike Kravetz if (unlikely(ret)) { 56659e368259SAndrea Arcangeli ret = -ENOENT; 56668cc5fcbbSMina Almasry /* Free the allocated page which may have 56678cc5fcbbSMina Almasry * consumed a reservation. 56688cc5fcbbSMina Almasry */ 56698cc5fcbbSMina Almasry restore_reserve_on_error(h, dst_vma, dst_addr, page); 56708cc5fcbbSMina Almasry put_page(page); 56718cc5fcbbSMina Almasry 56728cc5fcbbSMina Almasry /* Allocate a temporary page to hold the copied 56738cc5fcbbSMina Almasry * contents. 56748cc5fcbbSMina Almasry */ 56758cc5fcbbSMina Almasry page = alloc_huge_page_vma(h, dst_vma, dst_addr); 56768cc5fcbbSMina Almasry if (!page) { 56778cc5fcbbSMina Almasry ret = -ENOMEM; 56788cc5fcbbSMina Almasry goto out; 56798cc5fcbbSMina Almasry } 56808fb5debcSMike Kravetz *pagep = page; 56818cc5fcbbSMina Almasry /* Set the outparam pagep and return to the caller to 56828cc5fcbbSMina Almasry * copy the contents outside the lock. Don't free the 56838cc5fcbbSMina Almasry * page. 56848cc5fcbbSMina Almasry */ 56858fb5debcSMike Kravetz goto out; 56868fb5debcSMike Kravetz } 56878fb5debcSMike Kravetz } else { 56888cc5fcbbSMina Almasry if (vm_shared && 56898cc5fcbbSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 56908cc5fcbbSMina Almasry put_page(*pagep); 56918cc5fcbbSMina Almasry ret = -EEXIST; 56928cc5fcbbSMina Almasry *pagep = NULL; 56938cc5fcbbSMina Almasry goto out; 56948cc5fcbbSMina Almasry } 56958cc5fcbbSMina Almasry 56968cc5fcbbSMina Almasry page = alloc_huge_page(dst_vma, dst_addr, 0); 56978cc5fcbbSMina Almasry if (IS_ERR(page)) { 56988cc5fcbbSMina Almasry ret = -ENOMEM; 56998cc5fcbbSMina Almasry *pagep = NULL; 57008cc5fcbbSMina Almasry goto out; 57018cc5fcbbSMina Almasry } 57028cc5fcbbSMina Almasry copy_huge_page(page, *pagep); 57038cc5fcbbSMina Almasry put_page(*pagep); 57048fb5debcSMike Kravetz *pagep = NULL; 57058fb5debcSMike Kravetz } 57068fb5debcSMike Kravetz 57078fb5debcSMike Kravetz /* 57088fb5debcSMike Kravetz * The memory barrier inside __SetPageUptodate makes sure that 57098fb5debcSMike Kravetz * preceding stores to the page contents become visible before 57108fb5debcSMike Kravetz * the set_pte_at() write. 57118fb5debcSMike Kravetz */ 57128fb5debcSMike Kravetz __SetPageUptodate(page); 57138fb5debcSMike Kravetz 5714f6191471SAxel Rasmussen /* Add shared, newly allocated pages to the page cache. */ 5715f6191471SAxel Rasmussen if (vm_shared && !is_continue) { 57161e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 57171e392147SAndrea Arcangeli ret = -EFAULT; 57181e392147SAndrea Arcangeli if (idx >= size) 57191e392147SAndrea Arcangeli goto out_release_nounlock; 57201c9e8defSMike Kravetz 57211e392147SAndrea Arcangeli /* 57221e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 57231e392147SAndrea Arcangeli * huge_add_to_page_cache() below happens through the 57241e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 57251e392147SAndrea Arcangeli * the caller. 57261e392147SAndrea Arcangeli */ 57271c9e8defSMike Kravetz ret = huge_add_to_page_cache(page, mapping, idx); 57281c9e8defSMike Kravetz if (ret) 57291c9e8defSMike Kravetz goto out_release_nounlock; 5730c7b1850dSMike Kravetz new_pagecache_page = true; 57311c9e8defSMike Kravetz } 57321c9e8defSMike Kravetz 57338fb5debcSMike Kravetz ptl = huge_pte_lockptr(h, dst_mm, dst_pte); 57348fb5debcSMike Kravetz spin_lock(ptl); 57358fb5debcSMike Kravetz 57361e392147SAndrea Arcangeli /* 57371e392147SAndrea Arcangeli * Recheck the i_size after holding PT lock to make sure not 57381e392147SAndrea Arcangeli * to leave any page mapped (as page_mapped()) beyond the end 57391e392147SAndrea Arcangeli * of the i_size (remove_inode_hugepages() is strict about 57401e392147SAndrea Arcangeli * enforcing that). If we bail out here, we'll also leave a 57411e392147SAndrea Arcangeli * page in the radix tree in the vm_shared case beyond the end 57421e392147SAndrea Arcangeli * of the i_size, but remove_inode_hugepages() will take care 57431e392147SAndrea Arcangeli * of it as soon as we drop the hugetlb_fault_mutex_table. 57441e392147SAndrea Arcangeli */ 57451e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 57461e392147SAndrea Arcangeli ret = -EFAULT; 57471e392147SAndrea Arcangeli if (idx >= size) 57481e392147SAndrea Arcangeli goto out_release_unlock; 57491e392147SAndrea Arcangeli 57508fb5debcSMike Kravetz ret = -EEXIST; 57518fb5debcSMike Kravetz if (!huge_pte_none(huge_ptep_get(dst_pte))) 57528fb5debcSMike Kravetz goto out_release_unlock; 57538fb5debcSMike Kravetz 57541c9e8defSMike Kravetz if (vm_shared) { 57551c9e8defSMike Kravetz page_dup_rmap(page, true); 57561c9e8defSMike Kravetz } else { 5757d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 57588fb5debcSMike Kravetz hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); 57591c9e8defSMike Kravetz } 57608fb5debcSMike Kravetz 5761f6191471SAxel Rasmussen /* For CONTINUE on a non-shared VMA, don't set VM_WRITE for CoW. */ 5762f6191471SAxel Rasmussen if (is_continue && !vm_shared) 5763f6191471SAxel Rasmussen writable = 0; 5764f6191471SAxel Rasmussen else 5765f6191471SAxel Rasmussen writable = dst_vma->vm_flags & VM_WRITE; 5766f6191471SAxel Rasmussen 5767f6191471SAxel Rasmussen _dst_pte = make_huge_pte(dst_vma, page, writable); 5768f6191471SAxel Rasmussen if (writable) 57698fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 57708fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 57718fb5debcSMike Kravetz 57728fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 57738fb5debcSMike Kravetz 57748fb5debcSMike Kravetz (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, 57758fb5debcSMike Kravetz dst_vma->vm_flags & VM_WRITE); 57768fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 57778fb5debcSMike Kravetz 57788fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 57798fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 57808fb5debcSMike Kravetz 57818fb5debcSMike Kravetz spin_unlock(ptl); 5782f6191471SAxel Rasmussen if (!is_continue) 57838f251a3dSMike Kravetz SetHPageMigratable(page); 5784f6191471SAxel Rasmussen if (vm_shared || is_continue) 57851c9e8defSMike Kravetz unlock_page(page); 57868fb5debcSMike Kravetz ret = 0; 57878fb5debcSMike Kravetz out: 57888fb5debcSMike Kravetz return ret; 57898fb5debcSMike Kravetz out_release_unlock: 57908fb5debcSMike Kravetz spin_unlock(ptl); 5791f6191471SAxel Rasmussen if (vm_shared || is_continue) 57921c9e8defSMike Kravetz unlock_page(page); 57935af10dfdSAndrea Arcangeli out_release_nounlock: 5794c7b1850dSMike Kravetz if (!new_pagecache_page) 5795846be085SMike Kravetz restore_reserve_on_error(h, dst_vma, dst_addr, page); 57968fb5debcSMike Kravetz put_page(page); 57978fb5debcSMike Kravetz goto out; 57988fb5debcSMike Kravetz } 5799714c1891SAxel Rasmussen #endif /* CONFIG_USERFAULTFD */ 58008fb5debcSMike Kravetz 580182e5d378SJoao Martins static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma, 580282e5d378SJoao Martins int refs, struct page **pages, 580382e5d378SJoao Martins struct vm_area_struct **vmas) 580482e5d378SJoao Martins { 580582e5d378SJoao Martins int nr; 580682e5d378SJoao Martins 580782e5d378SJoao Martins for (nr = 0; nr < refs; nr++) { 580882e5d378SJoao Martins if (likely(pages)) 580982e5d378SJoao Martins pages[nr] = mem_map_offset(page, nr); 581082e5d378SJoao Martins if (vmas) 581182e5d378SJoao Martins vmas[nr] = vma; 581282e5d378SJoao Martins } 581382e5d378SJoao Martins } 581482e5d378SJoao Martins 581528a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 581663551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 581728a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 58184f6da934SPeter Xu long i, unsigned int flags, int *locked) 581963551ae0SDavid Gibson { 5820d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 5821d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 582228a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 5823a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 58240fa5bc40SJoao Martins int err = -EFAULT, refs; 582563551ae0SDavid Gibson 582663551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 582763551ae0SDavid Gibson pte_t *pte; 5828cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 58292a15efc9SHugh Dickins int absent; 583063551ae0SDavid Gibson struct page *page; 583163551ae0SDavid Gibson 58324c887265SAdam Litke /* 583302057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 583402057967SDavid Rientjes * potentially allocating memory. 583502057967SDavid Rientjes */ 5836fa45f116SDavidlohr Bueso if (fatal_signal_pending(current)) { 583702057967SDavid Rientjes remainder = 0; 583802057967SDavid Rientjes break; 583902057967SDavid Rientjes } 584002057967SDavid Rientjes 584102057967SDavid Rientjes /* 58424c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 58432a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 58444c887265SAdam Litke * first, for the page indexing below to work. 5845cb900f41SKirill A. Shutemov * 5846cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 58474c887265SAdam Litke */ 58487868a208SPunit Agrawal pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), 58497868a208SPunit Agrawal huge_page_size(h)); 5850cb900f41SKirill A. Shutemov if (pte) 5851cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 58522a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 585363551ae0SDavid Gibson 58542a15efc9SHugh Dickins /* 58552a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 58563ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 58573ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 58583ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 58593ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 58602a15efc9SHugh Dickins */ 58613ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 58623ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 5863cb900f41SKirill A. Shutemov if (pte) 5864cb900f41SKirill A. Shutemov spin_unlock(ptl); 58652a15efc9SHugh Dickins remainder = 0; 58662a15efc9SHugh Dickins break; 58672a15efc9SHugh Dickins } 58682a15efc9SHugh Dickins 58699cc3a5bdSNaoya Horiguchi /* 58709cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 58719cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 58729cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 58739cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 58749cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 58759cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 58769cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 58779cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 58789cc3a5bdSNaoya Horiguchi */ 58799cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 5880106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 5881106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 58822b740303SSouptick Joarder vm_fault_t ret; 588387ffc118SAndrea Arcangeli unsigned int fault_flags = 0; 58844c887265SAdam Litke 5885cb900f41SKirill A. Shutemov if (pte) 5886cb900f41SKirill A. Shutemov spin_unlock(ptl); 588787ffc118SAndrea Arcangeli if (flags & FOLL_WRITE) 588887ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_WRITE; 58894f6da934SPeter Xu if (locked) 589071335f37SPeter Xu fault_flags |= FAULT_FLAG_ALLOW_RETRY | 589171335f37SPeter Xu FAULT_FLAG_KILLABLE; 589287ffc118SAndrea Arcangeli if (flags & FOLL_NOWAIT) 589387ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY | 589487ffc118SAndrea Arcangeli FAULT_FLAG_RETRY_NOWAIT; 589587ffc118SAndrea Arcangeli if (flags & FOLL_TRIED) { 58964426e945SPeter Xu /* 58974426e945SPeter Xu * Note: FAULT_FLAG_ALLOW_RETRY and 58984426e945SPeter Xu * FAULT_FLAG_TRIED can co-exist 58994426e945SPeter Xu */ 590087ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_TRIED; 590187ffc118SAndrea Arcangeli } 590287ffc118SAndrea Arcangeli ret = hugetlb_fault(mm, vma, vaddr, fault_flags); 590387ffc118SAndrea Arcangeli if (ret & VM_FAULT_ERROR) { 59042be7cfedSDaniel Jordan err = vm_fault_to_errno(ret, flags); 59051c59827dSHugh Dickins remainder = 0; 59061c59827dSHugh Dickins break; 59071c59827dSHugh Dickins } 590887ffc118SAndrea Arcangeli if (ret & VM_FAULT_RETRY) { 59094f6da934SPeter Xu if (locked && 59101ac25013SAndrea Arcangeli !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) 59114f6da934SPeter Xu *locked = 0; 591287ffc118SAndrea Arcangeli *nr_pages = 0; 591387ffc118SAndrea Arcangeli /* 591487ffc118SAndrea Arcangeli * VM_FAULT_RETRY must not return an 591587ffc118SAndrea Arcangeli * error, it will return zero 591687ffc118SAndrea Arcangeli * instead. 591787ffc118SAndrea Arcangeli * 591887ffc118SAndrea Arcangeli * No need to update "position" as the 591987ffc118SAndrea Arcangeli * caller will not check it after 592087ffc118SAndrea Arcangeli * *nr_pages is set to 0. 592187ffc118SAndrea Arcangeli */ 592287ffc118SAndrea Arcangeli return i; 592387ffc118SAndrea Arcangeli } 592487ffc118SAndrea Arcangeli continue; 592587ffc118SAndrea Arcangeli } 592663551ae0SDavid Gibson 5927a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 59287f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 59298fde12caSLinus Torvalds 59308fde12caSLinus Torvalds /* 5931acbfb087SZhigang Lu * If subpage information not requested, update counters 5932acbfb087SZhigang Lu * and skip the same_page loop below. 5933acbfb087SZhigang Lu */ 5934acbfb087SZhigang Lu if (!pages && !vmas && !pfn_offset && 5935acbfb087SZhigang Lu (vaddr + huge_page_size(h) < vma->vm_end) && 5936acbfb087SZhigang Lu (remainder >= pages_per_huge_page(h))) { 5937acbfb087SZhigang Lu vaddr += huge_page_size(h); 5938acbfb087SZhigang Lu remainder -= pages_per_huge_page(h); 5939acbfb087SZhigang Lu i += pages_per_huge_page(h); 5940acbfb087SZhigang Lu spin_unlock(ptl); 5941acbfb087SZhigang Lu continue; 5942acbfb087SZhigang Lu } 5943acbfb087SZhigang Lu 5944d08af0a5SJoao Martins /* vaddr may not be aligned to PAGE_SIZE */ 5945d08af0a5SJoao Martins refs = min3(pages_per_huge_page(h) - pfn_offset, remainder, 5946d08af0a5SJoao Martins (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT); 59470fa5bc40SJoao Martins 594882e5d378SJoao Martins if (pages || vmas) 594982e5d378SJoao Martins record_subpages_vmas(mem_map_offset(page, pfn_offset), 595082e5d378SJoao Martins vma, refs, 595182e5d378SJoao Martins likely(pages) ? pages + i : NULL, 595282e5d378SJoao Martins vmas ? vmas + i : NULL); 595363551ae0SDavid Gibson 595482e5d378SJoao Martins if (pages) { 59550fa5bc40SJoao Martins /* 59560fa5bc40SJoao Martins * try_grab_compound_head() should always succeed here, 59570fa5bc40SJoao Martins * because: a) we hold the ptl lock, and b) we've just 59580fa5bc40SJoao Martins * checked that the huge page is present in the page 59590fa5bc40SJoao Martins * tables. If the huge page is present, then the tail 59600fa5bc40SJoao Martins * pages must also be present. The ptl prevents the 59610fa5bc40SJoao Martins * head page and tail pages from being rearranged in 59620fa5bc40SJoao Martins * any way. So this page must be available at this 59630fa5bc40SJoao Martins * point, unless the page refcount overflowed: 59640fa5bc40SJoao Martins */ 596582e5d378SJoao Martins if (WARN_ON_ONCE(!try_grab_compound_head(pages[i], 59660fa5bc40SJoao Martins refs, 59670fa5bc40SJoao Martins flags))) { 59680fa5bc40SJoao Martins spin_unlock(ptl); 59690fa5bc40SJoao Martins remainder = 0; 59700fa5bc40SJoao Martins err = -ENOMEM; 59710fa5bc40SJoao Martins break; 59720fa5bc40SJoao Martins } 5973d5d4b0aaSChen, Kenneth W } 597482e5d378SJoao Martins 597582e5d378SJoao Martins vaddr += (refs << PAGE_SHIFT); 597682e5d378SJoao Martins remainder -= refs; 597782e5d378SJoao Martins i += refs; 597882e5d378SJoao Martins 5979cb900f41SKirill A. Shutemov spin_unlock(ptl); 598063551ae0SDavid Gibson } 598128a35716SMichel Lespinasse *nr_pages = remainder; 598287ffc118SAndrea Arcangeli /* 598387ffc118SAndrea Arcangeli * setting position is actually required only if remainder is 598487ffc118SAndrea Arcangeli * not zero but it's faster not to add a "if (remainder)" 598587ffc118SAndrea Arcangeli * branch. 598687ffc118SAndrea Arcangeli */ 598763551ae0SDavid Gibson *position = vaddr; 598863551ae0SDavid Gibson 59892be7cfedSDaniel Jordan return i ? i : err; 599063551ae0SDavid Gibson } 59918f860591SZhang, Yanmin 59927da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 59938f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 59948f860591SZhang, Yanmin { 59958f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 59968f860591SZhang, Yanmin unsigned long start = address; 59978f860591SZhang, Yanmin pte_t *ptep; 59988f860591SZhang, Yanmin pte_t pte; 5999a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 60007da4d641SPeter Zijlstra unsigned long pages = 0; 6001dff11abeSMike Kravetz bool shared_pmd = false; 6002ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 6003dff11abeSMike Kravetz 6004dff11abeSMike Kravetz /* 6005dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 6006ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 6007dff11abeSMike Kravetz * range if PMD sharing is possible. 6008dff11abeSMike Kravetz */ 60097269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 60107269f999SJérôme Glisse 0, vma, mm, start, end); 6011ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 60128f860591SZhang, Yanmin 60138f860591SZhang, Yanmin BUG_ON(address >= end); 6014ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 60158f860591SZhang, Yanmin 6016ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 601783cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 6018a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 6019cb900f41SKirill A. Shutemov spinlock_t *ptl; 60207868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, huge_page_size(h)); 60218f860591SZhang, Yanmin if (!ptep) 60228f860591SZhang, Yanmin continue; 6023cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 602434ae204fSMike Kravetz if (huge_pmd_unshare(mm, vma, &address, ptep)) { 60257da4d641SPeter Zijlstra pages++; 6026cb900f41SKirill A. Shutemov spin_unlock(ptl); 6027dff11abeSMike Kravetz shared_pmd = true; 602839dde65cSChen, Kenneth W continue; 60297da4d641SPeter Zijlstra } 6030a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 6031a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 6032a8bda28dSNaoya Horiguchi spin_unlock(ptl); 6033a8bda28dSNaoya Horiguchi continue; 6034a8bda28dSNaoya Horiguchi } 6035a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 6036a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 6037a8bda28dSNaoya Horiguchi 60384dd845b5SAlistair Popple if (is_writable_migration_entry(entry)) { 6039a8bda28dSNaoya Horiguchi pte_t newpte; 6040a8bda28dSNaoya Horiguchi 60414dd845b5SAlistair Popple entry = make_readable_migration_entry( 60424dd845b5SAlistair Popple swp_offset(entry)); 6043a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 6044e5251fd4SPunit Agrawal set_huge_swap_pte_at(mm, address, ptep, 6045e5251fd4SPunit Agrawal newpte, huge_page_size(h)); 6046a8bda28dSNaoya Horiguchi pages++; 6047a8bda28dSNaoya Horiguchi } 6048a8bda28dSNaoya Horiguchi spin_unlock(ptl); 6049a8bda28dSNaoya Horiguchi continue; 6050a8bda28dSNaoya Horiguchi } 6051a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 6052023bdd00SAneesh Kumar K.V pte_t old_pte; 605379c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 6054023bdd00SAneesh Kumar K.V 6055023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 6056023bdd00SAneesh Kumar K.V pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); 605779c1c594SChristophe Leroy pte = arch_make_huge_pte(pte, shift, vma->vm_flags); 6058023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 60597da4d641SPeter Zijlstra pages++; 60608f860591SZhang, Yanmin } 6061cb900f41SKirill A. Shutemov spin_unlock(ptl); 60628f860591SZhang, Yanmin } 6063d833352aSMel Gorman /* 6064c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 6065d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 6066c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 6067dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 6068dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 6069d833352aSMel Gorman */ 6070dff11abeSMike Kravetz if (shared_pmd) 6071ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 6072dff11abeSMike Kravetz else 60735491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 60740f10851eSJérôme Glisse /* 60750f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 60760f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 60770f10851eSJérôme Glisse * 6078ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 60790f10851eSJérôme Glisse */ 608083cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 6081ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 60827da4d641SPeter Zijlstra 60837da4d641SPeter Zijlstra return pages << h->order; 60848f860591SZhang, Yanmin } 60858f860591SZhang, Yanmin 608633b8f84aSMike Kravetz /* Return true if reservation was successful, false otherwise. */ 608733b8f84aSMike Kravetz bool hugetlb_reserve_pages(struct inode *inode, 6088a1e78772SMel Gorman long from, long to, 60895a6fe125SMel Gorman struct vm_area_struct *vma, 6090ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 6091e4e574b7SAdam Litke { 609233b8f84aSMike Kravetz long chg, add = -1; 6093a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 609490481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 60959119a41eSJoonsoo Kim struct resv_map *resv_map; 6096075a61d0SMina Almasry struct hugetlb_cgroup *h_cg = NULL; 60970db9d74eSMina Almasry long gbl_reserve, regions_needed = 0; 6098e4e574b7SAdam Litke 609963489f8eSMike Kravetz /* This should never happen */ 610063489f8eSMike Kravetz if (from > to) { 610163489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 610233b8f84aSMike Kravetz return false; 610363489f8eSMike Kravetz } 610463489f8eSMike Kravetz 6105a1e78772SMel Gorman /* 610617c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 610717c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 610890481622SDavid Gibson * without using reserves 610917c9d12eSMel Gorman */ 6110ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 611133b8f84aSMike Kravetz return true; 611217c9d12eSMel Gorman 611317c9d12eSMel Gorman /* 6114a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 6115a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 6116a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 6117a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 6118a1e78772SMel Gorman */ 61199119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 6120f27a5136SMike Kravetz /* 6121f27a5136SMike Kravetz * resv_map can not be NULL as hugetlb_reserve_pages is only 6122f27a5136SMike Kravetz * called for inodes for which resv_maps were created (see 6123f27a5136SMike Kravetz * hugetlbfs_get_inode). 6124f27a5136SMike Kravetz */ 61254e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 61269119a41eSJoonsoo Kim 61270db9d74eSMina Almasry chg = region_chg(resv_map, from, to, ®ions_needed); 61289119a41eSJoonsoo Kim 61299119a41eSJoonsoo Kim } else { 6130e9fe92aeSMina Almasry /* Private mapping. */ 61319119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 61325a6fe125SMel Gorman if (!resv_map) 613333b8f84aSMike Kravetz return false; 61345a6fe125SMel Gorman 613517c9d12eSMel Gorman chg = to - from; 613617c9d12eSMel Gorman 61375a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 61385a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 61395a6fe125SMel Gorman } 61405a6fe125SMel Gorman 614133b8f84aSMike Kravetz if (chg < 0) 6142c50ac050SDave Hansen goto out_err; 614317c9d12eSMel Gorman 614433b8f84aSMike Kravetz if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), 614533b8f84aSMike Kravetz chg * pages_per_huge_page(h), &h_cg) < 0) 6146075a61d0SMina Almasry goto out_err; 6147075a61d0SMina Almasry 6148075a61d0SMina Almasry if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { 6149075a61d0SMina Almasry /* For private mappings, the hugetlb_cgroup uncharge info hangs 6150075a61d0SMina Almasry * of the resv_map. 6151075a61d0SMina Almasry */ 6152075a61d0SMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); 6153075a61d0SMina Almasry } 6154075a61d0SMina Almasry 61551c5ecae3SMike Kravetz /* 61561c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 61571c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 61581c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 61591c5ecae3SMike Kravetz */ 61601c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 616133b8f84aSMike Kravetz if (gbl_reserve < 0) 6162075a61d0SMina Almasry goto out_uncharge_cgroup; 616317c9d12eSMel Gorman 616417c9d12eSMel Gorman /* 616517c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 616690481622SDavid Gibson * Hand the pages back to the subpool if there are not 616717c9d12eSMel Gorman */ 616833b8f84aSMike Kravetz if (hugetlb_acct_memory(h, gbl_reserve) < 0) 6169075a61d0SMina Almasry goto out_put_pages; 617017c9d12eSMel Gorman 617117c9d12eSMel Gorman /* 617217c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 617317c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 617417c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 617517c9d12eSMel Gorman * the reservation was and the page cache tells how much of 617617c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 617717c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 617817c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 617917c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 618017c9d12eSMel Gorman * else has to be done for private mappings here 618117c9d12eSMel Gorman */ 618233039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 6183075a61d0SMina Almasry add = region_add(resv_map, from, to, regions_needed, h, h_cg); 618433039678SMike Kravetz 61850db9d74eSMina Almasry if (unlikely(add < 0)) { 61860db9d74eSMina Almasry hugetlb_acct_memory(h, -gbl_reserve); 6187075a61d0SMina Almasry goto out_put_pages; 61880db9d74eSMina Almasry } else if (unlikely(chg > add)) { 618933039678SMike Kravetz /* 619033039678SMike Kravetz * pages in this range were added to the reserve 619133039678SMike Kravetz * map between region_chg and region_add. This 619233039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 619333039678SMike Kravetz * the subpool and reserve counts modified above 619433039678SMike Kravetz * based on the difference. 619533039678SMike Kravetz */ 619633039678SMike Kravetz long rsv_adjust; 619733039678SMike Kravetz 6198d85aecf2SMiaohe Lin /* 6199d85aecf2SMiaohe Lin * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the 6200d85aecf2SMiaohe Lin * reference to h_cg->css. See comment below for detail. 6201d85aecf2SMiaohe Lin */ 6202075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd( 6203075a61d0SMina Almasry hstate_index(h), 6204075a61d0SMina Almasry (chg - add) * pages_per_huge_page(h), h_cg); 6205075a61d0SMina Almasry 620633039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 620733039678SMike Kravetz chg - add); 620833039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 6209d85aecf2SMiaohe Lin } else if (h_cg) { 6210d85aecf2SMiaohe Lin /* 6211d85aecf2SMiaohe Lin * The file_regions will hold their own reference to 6212d85aecf2SMiaohe Lin * h_cg->css. So we should release the reference held 6213d85aecf2SMiaohe Lin * via hugetlb_cgroup_charge_cgroup_rsvd() when we are 6214d85aecf2SMiaohe Lin * done. 6215d85aecf2SMiaohe Lin */ 6216d85aecf2SMiaohe Lin hugetlb_cgroup_put_rsvd_cgroup(h_cg); 621733039678SMike Kravetz } 621833039678SMike Kravetz } 621933b8f84aSMike Kravetz return true; 622033b8f84aSMike Kravetz 6221075a61d0SMina Almasry out_put_pages: 6222075a61d0SMina Almasry /* put back original number of pages, chg */ 6223075a61d0SMina Almasry (void)hugepage_subpool_put_pages(spool, chg); 6224075a61d0SMina Almasry out_uncharge_cgroup: 6225075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), 6226075a61d0SMina Almasry chg * pages_per_huge_page(h), h_cg); 6227c50ac050SDave Hansen out_err: 62285e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 62290db9d74eSMina Almasry /* Only call region_abort if the region_chg succeeded but the 62300db9d74eSMina Almasry * region_add failed or didn't run. 62310db9d74eSMina Almasry */ 62320db9d74eSMina Almasry if (chg >= 0 && add < 0) 62330db9d74eSMina Almasry region_abort(resv_map, from, to, regions_needed); 6234f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 6235f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 623633b8f84aSMike Kravetz return false; 6237a43a8c39SChen, Kenneth W } 6238a43a8c39SChen, Kenneth W 6239b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 6240b5cec28dSMike Kravetz long freed) 6241a43a8c39SChen, Kenneth W { 6242a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 62434e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 62449119a41eSJoonsoo Kim long chg = 0; 624590481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 62461c5ecae3SMike Kravetz long gbl_reserve; 624745c682a6SKen Chen 6248f27a5136SMike Kravetz /* 6249f27a5136SMike Kravetz * Since this routine can be called in the evict inode path for all 6250f27a5136SMike Kravetz * hugetlbfs inodes, resv_map could be NULL. 6251f27a5136SMike Kravetz */ 6252b5cec28dSMike Kravetz if (resv_map) { 6253b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 6254b5cec28dSMike Kravetz /* 6255b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 6256b5cec28dSMike Kravetz * must be split and another region descriptor can not be 6257b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 6258b5cec28dSMike Kravetz */ 6259b5cec28dSMike Kravetz if (chg < 0) 6260b5cec28dSMike Kravetz return chg; 6261b5cec28dSMike Kravetz } 6262b5cec28dSMike Kravetz 626345c682a6SKen Chen spin_lock(&inode->i_lock); 6264e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 626545c682a6SKen Chen spin_unlock(&inode->i_lock); 626645c682a6SKen Chen 62671c5ecae3SMike Kravetz /* 62681c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 62691c5ecae3SMike Kravetz * reservations to be released may be adjusted. 6270dddf31a4SMiaohe Lin * 6271dddf31a4SMiaohe Lin * Note that !resv_map implies freed == 0. So (chg - freed) 6272dddf31a4SMiaohe Lin * won't go negative. 62731c5ecae3SMike Kravetz */ 62741c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 62751c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 6276b5cec28dSMike Kravetz 6277b5cec28dSMike Kravetz return 0; 6278a43a8c39SChen, Kenneth W } 627993f70f90SNaoya Horiguchi 62803212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 62813212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 62823212b535SSteve Capper struct vm_area_struct *vma, 62833212b535SSteve Capper unsigned long addr, pgoff_t idx) 62843212b535SSteve Capper { 62853212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 62863212b535SSteve Capper svma->vm_start; 62873212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 62883212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 62893212b535SSteve Capper 62903212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 6291de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 6292de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 62933212b535SSteve Capper 62943212b535SSteve Capper /* 62953212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 62963212b535SSteve Capper * page table page. 62973212b535SSteve Capper */ 62983212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 62993212b535SSteve Capper vm_flags != svm_flags || 630007e51edfSMiaohe Lin !range_in_vma(svma, sbase, s_end)) 63013212b535SSteve Capper return 0; 63023212b535SSteve Capper 63033212b535SSteve Capper return saddr; 63043212b535SSteve Capper } 63053212b535SSteve Capper 630631aafb45SNicholas Krause static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) 63073212b535SSteve Capper { 63083212b535SSteve Capper unsigned long base = addr & PUD_MASK; 63093212b535SSteve Capper unsigned long end = base + PUD_SIZE; 63103212b535SSteve Capper 63113212b535SSteve Capper /* 63123212b535SSteve Capper * check on proper vm_flags and page table alignment 63133212b535SSteve Capper */ 6314017b1660SMike Kravetz if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) 631531aafb45SNicholas Krause return true; 631631aafb45SNicholas Krause return false; 63173212b535SSteve Capper } 63183212b535SSteve Capper 6319c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 6320c1991e07SPeter Xu { 6321c1991e07SPeter Xu #ifdef CONFIG_USERFAULTFD 6322c1991e07SPeter Xu if (uffd_disable_huge_pmd_share(vma)) 6323c1991e07SPeter Xu return false; 6324c1991e07SPeter Xu #endif 6325c1991e07SPeter Xu return vma_shareable(vma, addr); 6326c1991e07SPeter Xu } 6327c1991e07SPeter Xu 63283212b535SSteve Capper /* 6329017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 6330017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 6331017b1660SMike Kravetz * shared pmd mappings. 6332017b1660SMike Kravetz */ 6333017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 6334017b1660SMike Kravetz unsigned long *start, unsigned long *end) 6335017b1660SMike Kravetz { 6336a1ba9da8SLi Xinhai unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), 6337a1ba9da8SLi Xinhai v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 6338017b1660SMike Kravetz 6339a1ba9da8SLi Xinhai /* 6340f0953a1bSIngo Molnar * vma needs to span at least one aligned PUD size, and the range 6341f0953a1bSIngo Molnar * must be at least partially within in. 6342a1ba9da8SLi Xinhai */ 6343a1ba9da8SLi Xinhai if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || 6344a1ba9da8SLi Xinhai (*end <= v_start) || (*start >= v_end)) 6345017b1660SMike Kravetz return; 6346017b1660SMike Kravetz 634775802ca6SPeter Xu /* Extend the range to be PUD aligned for a worst case scenario */ 6348a1ba9da8SLi Xinhai if (*start > v_start) 6349a1ba9da8SLi Xinhai *start = ALIGN_DOWN(*start, PUD_SIZE); 6350017b1660SMike Kravetz 6351a1ba9da8SLi Xinhai if (*end < v_end) 6352a1ba9da8SLi Xinhai *end = ALIGN(*end, PUD_SIZE); 6353017b1660SMike Kravetz } 6354017b1660SMike Kravetz 6355017b1660SMike Kravetz /* 63563212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 63573212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 63583212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 6359c0d0381aSMike Kravetz * code much cleaner. 6360c0d0381aSMike Kravetz * 63610bf7b64eSMike Kravetz * This routine must be called with i_mmap_rwsem held in at least read mode if 63620bf7b64eSMike Kravetz * sharing is possible. For hugetlbfs, this prevents removal of any page 63630bf7b64eSMike Kravetz * table entries associated with the address space. This is important as we 63640bf7b64eSMike Kravetz * are setting up sharing based on existing page table entries (mappings). 63653212b535SSteve Capper */ 6366aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 6367aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 63683212b535SSteve Capper { 63693212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 63703212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 63713212b535SSteve Capper vma->vm_pgoff; 63723212b535SSteve Capper struct vm_area_struct *svma; 63733212b535SSteve Capper unsigned long saddr; 63743212b535SSteve Capper pte_t *spte = NULL; 63753212b535SSteve Capper pte_t *pte; 6376cb900f41SKirill A. Shutemov spinlock_t *ptl; 63773212b535SSteve Capper 63780bf7b64eSMike Kravetz i_mmap_assert_locked(mapping); 63793212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 63803212b535SSteve Capper if (svma == vma) 63813212b535SSteve Capper continue; 63823212b535SSteve Capper 63833212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 63843212b535SSteve Capper if (saddr) { 63857868a208SPunit Agrawal spte = huge_pte_offset(svma->vm_mm, saddr, 63867868a208SPunit Agrawal vma_mmu_pagesize(svma)); 63873212b535SSteve Capper if (spte) { 63883212b535SSteve Capper get_page(virt_to_page(spte)); 63893212b535SSteve Capper break; 63903212b535SSteve Capper } 63913212b535SSteve Capper } 63923212b535SSteve Capper } 63933212b535SSteve Capper 63943212b535SSteve Capper if (!spte) 63953212b535SSteve Capper goto out; 63963212b535SSteve Capper 63978bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 6398dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 63993212b535SSteve Capper pud_populate(mm, pud, 64003212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 6401c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 6402dc6c9a35SKirill A. Shutemov } else { 64033212b535SSteve Capper put_page(virt_to_page(spte)); 6404dc6c9a35SKirill A. Shutemov } 6405cb900f41SKirill A. Shutemov spin_unlock(ptl); 64063212b535SSteve Capper out: 64073212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 64083212b535SSteve Capper return pte; 64093212b535SSteve Capper } 64103212b535SSteve Capper 64113212b535SSteve Capper /* 64123212b535SSteve Capper * unmap huge page backed by shared pte. 64133212b535SSteve Capper * 64143212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 64153212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 64163212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 64173212b535SSteve Capper * 6418c0d0381aSMike Kravetz * Called with page table lock held and i_mmap_rwsem held in write mode. 64193212b535SSteve Capper * 64203212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 64213212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 64223212b535SSteve Capper */ 642334ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 642434ae204fSMike Kravetz unsigned long *addr, pte_t *ptep) 64253212b535SSteve Capper { 64263212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 6427c2febafcSKirill A. Shutemov p4d_t *p4d = p4d_offset(pgd, *addr); 6428c2febafcSKirill A. Shutemov pud_t *pud = pud_offset(p4d, *addr); 64293212b535SSteve Capper 643034ae204fSMike Kravetz i_mmap_assert_write_locked(vma->vm_file->f_mapping); 64313212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 64323212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 64333212b535SSteve Capper return 0; 64343212b535SSteve Capper 64353212b535SSteve Capper pud_clear(pud); 64363212b535SSteve Capper put_page(virt_to_page(ptep)); 6437dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 64383212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 64393212b535SSteve Capper return 1; 64403212b535SSteve Capper } 6441c1991e07SPeter Xu 64429e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 6443aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 6444aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 64459e5fc74cSSteve Capper { 64469e5fc74cSSteve Capper return NULL; 64479e5fc74cSSteve Capper } 6448e81f2d22SZhang Zhen 644934ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 645034ae204fSMike Kravetz unsigned long *addr, pte_t *ptep) 6451e81f2d22SZhang Zhen { 6452e81f2d22SZhang Zhen return 0; 6453e81f2d22SZhang Zhen } 6454017b1660SMike Kravetz 6455017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 6456017b1660SMike Kravetz unsigned long *start, unsigned long *end) 6457017b1660SMike Kravetz { 6458017b1660SMike Kravetz } 6459c1991e07SPeter Xu 6460c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 6461c1991e07SPeter Xu { 6462c1991e07SPeter Xu return false; 6463c1991e07SPeter Xu } 64643212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 64653212b535SSteve Capper 64669e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 6467aec44e0fSPeter Xu pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, 64689e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 64699e5fc74cSSteve Capper { 64709e5fc74cSSteve Capper pgd_t *pgd; 6471c2febafcSKirill A. Shutemov p4d_t *p4d; 64729e5fc74cSSteve Capper pud_t *pud; 64739e5fc74cSSteve Capper pte_t *pte = NULL; 64749e5fc74cSSteve Capper 64759e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 6476f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 6477f4f0a3d8SKirill A. Shutemov if (!p4d) 6478f4f0a3d8SKirill A. Shutemov return NULL; 6479c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 64809e5fc74cSSteve Capper if (pud) { 64819e5fc74cSSteve Capper if (sz == PUD_SIZE) { 64829e5fc74cSSteve Capper pte = (pte_t *)pud; 64839e5fc74cSSteve Capper } else { 64849e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 6485c1991e07SPeter Xu if (want_pmd_share(vma, addr) && pud_none(*pud)) 6486aec44e0fSPeter Xu pte = huge_pmd_share(mm, vma, addr, pud); 64879e5fc74cSSteve Capper else 64889e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 64899e5fc74cSSteve Capper } 64909e5fc74cSSteve Capper } 64914e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 64929e5fc74cSSteve Capper 64939e5fc74cSSteve Capper return pte; 64949e5fc74cSSteve Capper } 64959e5fc74cSSteve Capper 64969b19df29SPunit Agrawal /* 64979b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 64989b19df29SPunit Agrawal * entry at address @addr 64999b19df29SPunit Agrawal * 65008ac0b81aSLi Xinhai * Return: Pointer to page table entry (PUD or PMD) for 65018ac0b81aSLi Xinhai * address @addr, or NULL if a !p*d_present() entry is encountered and the 65029b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 65039b19df29SPunit Agrawal * table. 65049b19df29SPunit Agrawal */ 65057868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 65067868a208SPunit Agrawal unsigned long addr, unsigned long sz) 65079e5fc74cSSteve Capper { 65089e5fc74cSSteve Capper pgd_t *pgd; 6509c2febafcSKirill A. Shutemov p4d_t *p4d; 65108ac0b81aSLi Xinhai pud_t *pud; 65118ac0b81aSLi Xinhai pmd_t *pmd; 65129e5fc74cSSteve Capper 65139e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 6514c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 6515c2febafcSKirill A. Shutemov return NULL; 6516c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 6517c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 6518c2febafcSKirill A. Shutemov return NULL; 65199b19df29SPunit Agrawal 6520c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 65218ac0b81aSLi Xinhai if (sz == PUD_SIZE) 65228ac0b81aSLi Xinhai /* must be pud huge, non-present or none */ 65239e5fc74cSSteve Capper return (pte_t *)pud; 65248ac0b81aSLi Xinhai if (!pud_present(*pud)) 65258ac0b81aSLi Xinhai return NULL; 65268ac0b81aSLi Xinhai /* must have a valid entry and size to go further */ 65279b19df29SPunit Agrawal 65289e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 65298ac0b81aSLi Xinhai /* must be pmd huge, non-present or none */ 65309e5fc74cSSteve Capper return (pte_t *)pmd; 65319e5fc74cSSteve Capper } 65329e5fc74cSSteve Capper 653361f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 653461f77edaSNaoya Horiguchi 653561f77edaSNaoya Horiguchi /* 653661f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 653761f77edaSNaoya Horiguchi * behavior. 653861f77edaSNaoya Horiguchi */ 653961f77edaSNaoya Horiguchi struct page * __weak 654061f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 654161f77edaSNaoya Horiguchi int write) 654261f77edaSNaoya Horiguchi { 654361f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 654461f77edaSNaoya Horiguchi } 654561f77edaSNaoya Horiguchi 654661f77edaSNaoya Horiguchi struct page * __weak 65474dc71451SAneesh Kumar K.V follow_huge_pd(struct vm_area_struct *vma, 65484dc71451SAneesh Kumar K.V unsigned long address, hugepd_t hpd, int flags, int pdshift) 65494dc71451SAneesh Kumar K.V { 65504dc71451SAneesh Kumar K.V WARN(1, "hugepd follow called with no support for hugepage directory format\n"); 65514dc71451SAneesh Kumar K.V return NULL; 65524dc71451SAneesh Kumar K.V } 65534dc71451SAneesh Kumar K.V 65544dc71451SAneesh Kumar K.V struct page * __weak 65559e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 6556e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 65579e5fc74cSSteve Capper { 6558e66f17ffSNaoya Horiguchi struct page *page = NULL; 6559e66f17ffSNaoya Horiguchi spinlock_t *ptl; 6560c9d398faSNaoya Horiguchi pte_t pte; 65613faa52c0SJohn Hubbard 65623faa52c0SJohn Hubbard /* FOLL_GET and FOLL_PIN are mutually exclusive. */ 65633faa52c0SJohn Hubbard if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == 65643faa52c0SJohn Hubbard (FOLL_PIN | FOLL_GET))) 65653faa52c0SJohn Hubbard return NULL; 65663faa52c0SJohn Hubbard 6567e66f17ffSNaoya Horiguchi retry: 6568e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 6569e66f17ffSNaoya Horiguchi spin_lock(ptl); 6570e66f17ffSNaoya Horiguchi /* 6571e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 6572e66f17ffSNaoya Horiguchi * unmapped from other threads. 6573e66f17ffSNaoya Horiguchi */ 6574e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 6575e66f17ffSNaoya Horiguchi goto out; 6576c9d398faSNaoya Horiguchi pte = huge_ptep_get((pte_t *)pmd); 6577c9d398faSNaoya Horiguchi if (pte_present(pte)) { 657897534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 65793faa52c0SJohn Hubbard /* 65803faa52c0SJohn Hubbard * try_grab_page() should always succeed here, because: a) we 65813faa52c0SJohn Hubbard * hold the pmd (ptl) lock, and b) we've just checked that the 65823faa52c0SJohn Hubbard * huge pmd (head) page is present in the page tables. The ptl 65833faa52c0SJohn Hubbard * prevents the head page and tail pages from being rearranged 65843faa52c0SJohn Hubbard * in any way. So this page must be available at this point, 65853faa52c0SJohn Hubbard * unless the page refcount overflowed: 65863faa52c0SJohn Hubbard */ 65873faa52c0SJohn Hubbard if (WARN_ON_ONCE(!try_grab_page(page, flags))) { 65883faa52c0SJohn Hubbard page = NULL; 65893faa52c0SJohn Hubbard goto out; 65903faa52c0SJohn Hubbard } 6591e66f17ffSNaoya Horiguchi } else { 6592c9d398faSNaoya Horiguchi if (is_hugetlb_entry_migration(pte)) { 6593e66f17ffSNaoya Horiguchi spin_unlock(ptl); 6594e66f17ffSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pmd, ptl); 6595e66f17ffSNaoya Horiguchi goto retry; 6596e66f17ffSNaoya Horiguchi } 6597e66f17ffSNaoya Horiguchi /* 6598e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 6599e66f17ffSNaoya Horiguchi * follow_page_mask(). 6600e66f17ffSNaoya Horiguchi */ 6601e66f17ffSNaoya Horiguchi } 6602e66f17ffSNaoya Horiguchi out: 6603e66f17ffSNaoya Horiguchi spin_unlock(ptl); 66049e5fc74cSSteve Capper return page; 66059e5fc74cSSteve Capper } 66069e5fc74cSSteve Capper 660761f77edaSNaoya Horiguchi struct page * __weak 66089e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 6609e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 66109e5fc74cSSteve Capper { 66113faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 6612e66f17ffSNaoya Horiguchi return NULL; 66139e5fc74cSSteve Capper 6614e66f17ffSNaoya Horiguchi return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 66159e5fc74cSSteve Capper } 66169e5fc74cSSteve Capper 6617faaa5b62SAnshuman Khandual struct page * __weak 6618faaa5b62SAnshuman Khandual follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) 6619faaa5b62SAnshuman Khandual { 66203faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 6621faaa5b62SAnshuman Khandual return NULL; 6622faaa5b62SAnshuman Khandual 6623faaa5b62SAnshuman Khandual return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); 6624faaa5b62SAnshuman Khandual } 6625faaa5b62SAnshuman Khandual 662631caf665SNaoya Horiguchi bool isolate_huge_page(struct page *page, struct list_head *list) 662731caf665SNaoya Horiguchi { 6628bcc54222SNaoya Horiguchi bool ret = true; 6629bcc54222SNaoya Horiguchi 6630db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 66318f251a3dSMike Kravetz if (!PageHeadHuge(page) || 66328f251a3dSMike Kravetz !HPageMigratable(page) || 66330eb2df2bSMuchun Song !get_page_unless_zero(page)) { 6634bcc54222SNaoya Horiguchi ret = false; 6635bcc54222SNaoya Horiguchi goto unlock; 6636bcc54222SNaoya Horiguchi } 66378f251a3dSMike Kravetz ClearHPageMigratable(page); 663831caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 6639bcc54222SNaoya Horiguchi unlock: 6640db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 6641bcc54222SNaoya Horiguchi return ret; 664231caf665SNaoya Horiguchi } 664331caf665SNaoya Horiguchi 664425182f05SNaoya Horiguchi int get_hwpoison_huge_page(struct page *page, bool *hugetlb) 664525182f05SNaoya Horiguchi { 664625182f05SNaoya Horiguchi int ret = 0; 664725182f05SNaoya Horiguchi 664825182f05SNaoya Horiguchi *hugetlb = false; 664925182f05SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 665025182f05SNaoya Horiguchi if (PageHeadHuge(page)) { 665125182f05SNaoya Horiguchi *hugetlb = true; 665225182f05SNaoya Horiguchi if (HPageFreed(page) || HPageMigratable(page)) 665325182f05SNaoya Horiguchi ret = get_page_unless_zero(page); 66540ed950d1SNaoya Horiguchi else 66550ed950d1SNaoya Horiguchi ret = -EBUSY; 665625182f05SNaoya Horiguchi } 665725182f05SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 665825182f05SNaoya Horiguchi return ret; 665925182f05SNaoya Horiguchi } 666025182f05SNaoya Horiguchi 666131caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 666231caf665SNaoya Horiguchi { 6663db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 66648f251a3dSMike Kravetz SetHPageMigratable(page); 666531caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 6666db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 666731caf665SNaoya Horiguchi put_page(page); 666831caf665SNaoya Horiguchi } 6669ab5ac90aSMichal Hocko 6670ab5ac90aSMichal Hocko void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) 6671ab5ac90aSMichal Hocko { 6672ab5ac90aSMichal Hocko struct hstate *h = page_hstate(oldpage); 6673ab5ac90aSMichal Hocko 6674ab5ac90aSMichal Hocko hugetlb_cgroup_migrate(oldpage, newpage); 6675ab5ac90aSMichal Hocko set_page_owner_migrate_reason(newpage, reason); 6676ab5ac90aSMichal Hocko 6677ab5ac90aSMichal Hocko /* 6678ab5ac90aSMichal Hocko * transfer temporary state of the new huge page. This is 6679ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 6680ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 6681ab5ac90aSMichal Hocko * the temporary status. 6682ab5ac90aSMichal Hocko * 6683ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 6684ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 6685ab5ac90aSMichal Hocko * the per-node's. 6686ab5ac90aSMichal Hocko */ 66879157c311SMike Kravetz if (HPageTemporary(newpage)) { 6688ab5ac90aSMichal Hocko int old_nid = page_to_nid(oldpage); 6689ab5ac90aSMichal Hocko int new_nid = page_to_nid(newpage); 6690ab5ac90aSMichal Hocko 66919157c311SMike Kravetz SetHPageTemporary(oldpage); 66929157c311SMike Kravetz ClearHPageTemporary(newpage); 6693ab5ac90aSMichal Hocko 66945af1ab1dSMiaohe Lin /* 66955af1ab1dSMiaohe Lin * There is no need to transfer the per-node surplus state 66965af1ab1dSMiaohe Lin * when we do not cross the node. 66975af1ab1dSMiaohe Lin */ 66985af1ab1dSMiaohe Lin if (new_nid == old_nid) 66995af1ab1dSMiaohe Lin return; 6700db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 6701ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 6702ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 6703ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 6704ab5ac90aSMichal Hocko } 6705db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 6706ab5ac90aSMichal Hocko } 6707ab5ac90aSMichal Hocko } 6708cf11e85fSRoman Gushchin 67096dfeaff9SPeter Xu /* 67106dfeaff9SPeter Xu * This function will unconditionally remove all the shared pmd pgtable entries 67116dfeaff9SPeter Xu * within the specific vma for a hugetlbfs memory range. 67126dfeaff9SPeter Xu */ 67136dfeaff9SPeter Xu void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) 67146dfeaff9SPeter Xu { 67156dfeaff9SPeter Xu struct hstate *h = hstate_vma(vma); 67166dfeaff9SPeter Xu unsigned long sz = huge_page_size(h); 67176dfeaff9SPeter Xu struct mm_struct *mm = vma->vm_mm; 67186dfeaff9SPeter Xu struct mmu_notifier_range range; 67196dfeaff9SPeter Xu unsigned long address, start, end; 67206dfeaff9SPeter Xu spinlock_t *ptl; 67216dfeaff9SPeter Xu pte_t *ptep; 67226dfeaff9SPeter Xu 67236dfeaff9SPeter Xu if (!(vma->vm_flags & VM_MAYSHARE)) 67246dfeaff9SPeter Xu return; 67256dfeaff9SPeter Xu 67266dfeaff9SPeter Xu start = ALIGN(vma->vm_start, PUD_SIZE); 67276dfeaff9SPeter Xu end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 67286dfeaff9SPeter Xu 67296dfeaff9SPeter Xu if (start >= end) 67306dfeaff9SPeter Xu return; 67316dfeaff9SPeter Xu 67326dfeaff9SPeter Xu /* 67336dfeaff9SPeter Xu * No need to call adjust_range_if_pmd_sharing_possible(), because 67346dfeaff9SPeter Xu * we have already done the PUD_SIZE alignment. 67356dfeaff9SPeter Xu */ 67366dfeaff9SPeter Xu mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, 67376dfeaff9SPeter Xu start, end); 67386dfeaff9SPeter Xu mmu_notifier_invalidate_range_start(&range); 67396dfeaff9SPeter Xu i_mmap_lock_write(vma->vm_file->f_mapping); 67406dfeaff9SPeter Xu for (address = start; address < end; address += PUD_SIZE) { 67416dfeaff9SPeter Xu unsigned long tmp = address; 67426dfeaff9SPeter Xu 67436dfeaff9SPeter Xu ptep = huge_pte_offset(mm, address, sz); 67446dfeaff9SPeter Xu if (!ptep) 67456dfeaff9SPeter Xu continue; 67466dfeaff9SPeter Xu ptl = huge_pte_lock(h, mm, ptep); 67476dfeaff9SPeter Xu /* We don't want 'address' to be changed */ 67486dfeaff9SPeter Xu huge_pmd_unshare(mm, vma, &tmp, ptep); 67496dfeaff9SPeter Xu spin_unlock(ptl); 67506dfeaff9SPeter Xu } 67516dfeaff9SPeter Xu flush_hugetlb_tlb_range(vma, start, end); 67526dfeaff9SPeter Xu i_mmap_unlock_write(vma->vm_file->f_mapping); 67536dfeaff9SPeter Xu /* 67546dfeaff9SPeter Xu * No need to call mmu_notifier_invalidate_range(), see 67556dfeaff9SPeter Xu * Documentation/vm/mmu_notifier.rst. 67566dfeaff9SPeter Xu */ 67576dfeaff9SPeter Xu mmu_notifier_invalidate_range_end(&range); 67586dfeaff9SPeter Xu } 67596dfeaff9SPeter Xu 6760cf11e85fSRoman Gushchin #ifdef CONFIG_CMA 6761cf11e85fSRoman Gushchin static bool cma_reserve_called __initdata; 6762cf11e85fSRoman Gushchin 6763cf11e85fSRoman Gushchin static int __init cmdline_parse_hugetlb_cma(char *p) 6764cf11e85fSRoman Gushchin { 676538e719abSBaolin Wang int nid, count = 0; 676638e719abSBaolin Wang unsigned long tmp; 676738e719abSBaolin Wang char *s = p; 676838e719abSBaolin Wang 676938e719abSBaolin Wang while (*s) { 677038e719abSBaolin Wang if (sscanf(s, "%lu%n", &tmp, &count) != 1) 677138e719abSBaolin Wang break; 677238e719abSBaolin Wang 677338e719abSBaolin Wang if (s[count] == ':') { 677438e719abSBaolin Wang nid = tmp; 677538e719abSBaolin Wang if (nid < 0 || nid >= MAX_NUMNODES) 677638e719abSBaolin Wang break; 677738e719abSBaolin Wang 677838e719abSBaolin Wang s += count + 1; 677938e719abSBaolin Wang tmp = memparse(s, &s); 678038e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = tmp; 678138e719abSBaolin Wang hugetlb_cma_size += tmp; 678238e719abSBaolin Wang 678338e719abSBaolin Wang /* 678438e719abSBaolin Wang * Skip the separator if have one, otherwise 678538e719abSBaolin Wang * break the parsing. 678638e719abSBaolin Wang */ 678738e719abSBaolin Wang if (*s == ',') 678838e719abSBaolin Wang s++; 678938e719abSBaolin Wang else 679038e719abSBaolin Wang break; 679138e719abSBaolin Wang } else { 6792cf11e85fSRoman Gushchin hugetlb_cma_size = memparse(p, &p); 679338e719abSBaolin Wang break; 679438e719abSBaolin Wang } 679538e719abSBaolin Wang } 679638e719abSBaolin Wang 6797cf11e85fSRoman Gushchin return 0; 6798cf11e85fSRoman Gushchin } 6799cf11e85fSRoman Gushchin 6800cf11e85fSRoman Gushchin early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); 6801cf11e85fSRoman Gushchin 6802cf11e85fSRoman Gushchin void __init hugetlb_cma_reserve(int order) 6803cf11e85fSRoman Gushchin { 6804cf11e85fSRoman Gushchin unsigned long size, reserved, per_node; 680538e719abSBaolin Wang bool node_specific_cma_alloc = false; 6806cf11e85fSRoman Gushchin int nid; 6807cf11e85fSRoman Gushchin 6808cf11e85fSRoman Gushchin cma_reserve_called = true; 6809cf11e85fSRoman Gushchin 6810cf11e85fSRoman Gushchin if (!hugetlb_cma_size) 6811cf11e85fSRoman Gushchin return; 6812cf11e85fSRoman Gushchin 681338e719abSBaolin Wang for (nid = 0; nid < MAX_NUMNODES; nid++) { 681438e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 681538e719abSBaolin Wang continue; 681638e719abSBaolin Wang 681738e719abSBaolin Wang if (!node_state(nid, N_ONLINE)) { 681838e719abSBaolin Wang pr_warn("hugetlb_cma: invalid node %d specified\n", nid); 681938e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 682038e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 682138e719abSBaolin Wang continue; 682238e719abSBaolin Wang } 682338e719abSBaolin Wang 682438e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) { 682538e719abSBaolin Wang pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n", 682638e719abSBaolin Wang nid, (PAGE_SIZE << order) / SZ_1M); 682738e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 682838e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 682938e719abSBaolin Wang } else { 683038e719abSBaolin Wang node_specific_cma_alloc = true; 683138e719abSBaolin Wang } 683238e719abSBaolin Wang } 683338e719abSBaolin Wang 683438e719abSBaolin Wang /* Validate the CMA size again in case some invalid nodes specified. */ 683538e719abSBaolin Wang if (!hugetlb_cma_size) 683638e719abSBaolin Wang return; 683738e719abSBaolin Wang 6838cf11e85fSRoman Gushchin if (hugetlb_cma_size < (PAGE_SIZE << order)) { 6839cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", 6840cf11e85fSRoman Gushchin (PAGE_SIZE << order) / SZ_1M); 6841a01f4390SMike Kravetz hugetlb_cma_size = 0; 6842cf11e85fSRoman Gushchin return; 6843cf11e85fSRoman Gushchin } 6844cf11e85fSRoman Gushchin 684538e719abSBaolin Wang if (!node_specific_cma_alloc) { 6846cf11e85fSRoman Gushchin /* 6847cf11e85fSRoman Gushchin * If 3 GB area is requested on a machine with 4 numa nodes, 6848cf11e85fSRoman Gushchin * let's allocate 1 GB on first three nodes and ignore the last one. 6849cf11e85fSRoman Gushchin */ 6850cf11e85fSRoman Gushchin per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); 6851cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", 6852cf11e85fSRoman Gushchin hugetlb_cma_size / SZ_1M, per_node / SZ_1M); 685338e719abSBaolin Wang } 6854cf11e85fSRoman Gushchin 6855cf11e85fSRoman Gushchin reserved = 0; 6856cf11e85fSRoman Gushchin for_each_node_state(nid, N_ONLINE) { 6857cf11e85fSRoman Gushchin int res; 68582281f797SBarry Song char name[CMA_MAX_NAME]; 6859cf11e85fSRoman Gushchin 686038e719abSBaolin Wang if (node_specific_cma_alloc) { 686138e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 686238e719abSBaolin Wang continue; 686338e719abSBaolin Wang 686438e719abSBaolin Wang size = hugetlb_cma_size_in_node[nid]; 686538e719abSBaolin Wang } else { 6866cf11e85fSRoman Gushchin size = min(per_node, hugetlb_cma_size - reserved); 686738e719abSBaolin Wang } 686838e719abSBaolin Wang 6869cf11e85fSRoman Gushchin size = round_up(size, PAGE_SIZE << order); 6870cf11e85fSRoman Gushchin 68712281f797SBarry Song snprintf(name, sizeof(name), "hugetlb%d", nid); 6872a01f4390SMike Kravetz /* 6873a01f4390SMike Kravetz * Note that 'order per bit' is based on smallest size that 6874a01f4390SMike Kravetz * may be returned to CMA allocator in the case of 6875a01f4390SMike Kravetz * huge page demotion. 6876a01f4390SMike Kravetz */ 6877a01f4390SMike Kravetz res = cma_declare_contiguous_nid(0, size, 0, 6878a01f4390SMike Kravetz PAGE_SIZE << HUGETLB_PAGE_ORDER, 687929d0f41dSBarry Song 0, false, name, 6880cf11e85fSRoman Gushchin &hugetlb_cma[nid], nid); 6881cf11e85fSRoman Gushchin if (res) { 6882cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: reservation failed: err %d, node %d", 6883cf11e85fSRoman Gushchin res, nid); 6884cf11e85fSRoman Gushchin continue; 6885cf11e85fSRoman Gushchin } 6886cf11e85fSRoman Gushchin 6887cf11e85fSRoman Gushchin reserved += size; 6888cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", 6889cf11e85fSRoman Gushchin size / SZ_1M, nid); 6890cf11e85fSRoman Gushchin 6891cf11e85fSRoman Gushchin if (reserved >= hugetlb_cma_size) 6892cf11e85fSRoman Gushchin break; 6893cf11e85fSRoman Gushchin } 6894a01f4390SMike Kravetz 6895a01f4390SMike Kravetz if (!reserved) 6896a01f4390SMike Kravetz /* 6897a01f4390SMike Kravetz * hugetlb_cma_size is used to determine if allocations from 6898a01f4390SMike Kravetz * cma are possible. Set to zero if no cma regions are set up. 6899a01f4390SMike Kravetz */ 6900a01f4390SMike Kravetz hugetlb_cma_size = 0; 6901cf11e85fSRoman Gushchin } 6902cf11e85fSRoman Gushchin 6903cf11e85fSRoman Gushchin void __init hugetlb_cma_check(void) 6904cf11e85fSRoman Gushchin { 6905cf11e85fSRoman Gushchin if (!hugetlb_cma_size || cma_reserve_called) 6906cf11e85fSRoman Gushchin return; 6907cf11e85fSRoman Gushchin 6908cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); 6909cf11e85fSRoman Gushchin } 6910cf11e85fSRoman Gushchin 6911cf11e85fSRoman Gushchin #endif /* CONFIG_CMA */ 6912