1457c8996SThomas Gleixner // SPDX-License-Identifier: GPL-2.0-only 21da177e4SLinus Torvalds /* 31da177e4SLinus Torvalds * Generic hugetlb support. 46d49e352SNadia Yvette Chambers * (C) Nadia Yvette Chambers, April 2004 51da177e4SLinus Torvalds */ 61da177e4SLinus Torvalds #include <linux/list.h> 71da177e4SLinus Torvalds #include <linux/init.h> 81da177e4SLinus Torvalds #include <linux/mm.h> 9e1759c21SAlexey Dobriyan #include <linux/seq_file.h> 101da177e4SLinus Torvalds #include <linux/sysctl.h> 111da177e4SLinus Torvalds #include <linux/highmem.h> 12cddb8a5cSAndrea Arcangeli #include <linux/mmu_notifier.h> 131da177e4SLinus Torvalds #include <linux/nodemask.h> 1463551ae0SDavid Gibson #include <linux/pagemap.h> 155da7ca86SChristoph Lameter #include <linux/mempolicy.h> 163b32123dSGideon Israel Dsouza #include <linux/compiler.h> 17aea47ff3SChristoph Lameter #include <linux/cpuset.h> 183935baa9SDavid Gibson #include <linux/mutex.h> 1997ad1087SMike Rapoport #include <linux/memblock.h> 20a3437870SNishanth Aravamudan #include <linux/sysfs.h> 215a0e3ad6STejun Heo #include <linux/slab.h> 22bbe88753SJoonsoo Kim #include <linux/sched/mm.h> 2363489f8eSMike Kravetz #include <linux/mmdebug.h> 24174cd4b1SIngo Molnar #include <linux/sched/signal.h> 250fe6e20bSNaoya Horiguchi #include <linux/rmap.h> 26c6247f72SMatthew Wilcox #include <linux/string_helpers.h> 27fd6a03edSNaoya Horiguchi #include <linux/swap.h> 28fd6a03edSNaoya Horiguchi #include <linux/swapops.h> 298382d914SDavidlohr Bueso #include <linux/jhash.h> 3098fa15f3SAnshuman Khandual #include <linux/numa.h> 31c77c0a8aSWaiman Long #include <linux/llist.h> 32cf11e85fSRoman Gushchin #include <linux/cma.h> 338cc5fcbbSMina Almasry #include <linux/migrate.h> 34f9317f77SMike Kravetz #include <linux/nospec.h> 35662ce1dcSYang Yang #include <linux/delayacct.h> 36b958d4d0SMuchun Song #include <linux/memory.h> 37d6606683SLinus Torvalds 3863551ae0SDavid Gibson #include <asm/page.h> 39ca15ca40SMike Rapoport #include <asm/pgalloc.h> 4024669e58SAneesh Kumar K.V #include <asm/tlb.h> 4163551ae0SDavid Gibson 4224669e58SAneesh Kumar K.V #include <linux/io.h> 4363551ae0SDavid Gibson #include <linux/hugetlb.h> 449dd540e2SAneesh Kumar K.V #include <linux/hugetlb_cgroup.h> 459a305230SLee Schermerhorn #include <linux/node.h> 46ab5ac90aSMichal Hocko #include <linux/page_owner.h> 477835e98bSNick Piggin #include "internal.h" 48f41f2ed4SMuchun Song #include "hugetlb_vmemmap.h" 491da177e4SLinus Torvalds 50c3f38a38SAneesh Kumar K.V int hugetlb_max_hstate __read_mostly; 51e5ff2159SAndi Kleen unsigned int default_hstate_idx; 52e5ff2159SAndi Kleen struct hstate hstates[HUGE_MAX_HSTATE]; 53cf11e85fSRoman Gushchin 54dbda8feaSBarry Song #ifdef CONFIG_CMA 55cf11e85fSRoman Gushchin static struct cma *hugetlb_cma[MAX_NUMNODES]; 5638e719abSBaolin Wang static unsigned long hugetlb_cma_size_in_node[MAX_NUMNODES] __initdata; 57a01f4390SMike Kravetz static bool hugetlb_cma_page(struct page *page, unsigned int order) 58a01f4390SMike Kravetz { 59a01f4390SMike Kravetz return cma_pages_valid(hugetlb_cma[page_to_nid(page)], page, 60a01f4390SMike Kravetz 1 << order); 61a01f4390SMike Kravetz } 62a01f4390SMike Kravetz #else 63a01f4390SMike Kravetz static bool hugetlb_cma_page(struct page *page, unsigned int order) 64a01f4390SMike Kravetz { 65a01f4390SMike Kravetz return false; 66a01f4390SMike Kravetz } 67dbda8feaSBarry Song #endif 68dbda8feaSBarry Song static unsigned long hugetlb_cma_size __initdata; 69cf11e85fSRoman Gushchin 7053ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 7153ba51d2SJon Tollefson 72e5ff2159SAndi Kleen /* for command line parsing */ 73e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 74e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 759fee021dSVaishali Thakkar static bool __initdata parsed_valid_hugepagesz = true; 76282f4214SMike Kravetz static bool __initdata parsed_default_hugepagesz; 77b5389086SZhenguo Yao static unsigned int default_hugepages_in_node[MAX_NUMNODES] __initdata; 78e5ff2159SAndi Kleen 793935baa9SDavid Gibson /* 8031caf665SNaoya Horiguchi * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, 8131caf665SNaoya Horiguchi * free_huge_pages, and surplus_huge_pages. 823935baa9SDavid Gibson */ 83c3f38a38SAneesh Kumar K.V DEFINE_SPINLOCK(hugetlb_lock); 840bd0f9fbSEric Paris 858382d914SDavidlohr Bueso /* 868382d914SDavidlohr Bueso * Serializes faults on the same logical page. This is used to 878382d914SDavidlohr Bueso * prevent spurious OOMs when the hugepage pool is fully utilized. 888382d914SDavidlohr Bueso */ 898382d914SDavidlohr Bueso static int num_fault_mutexes; 90c672c7f2SMike Kravetz struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; 918382d914SDavidlohr Bueso 927ca02d0aSMike Kravetz /* Forward declaration */ 937ca02d0aSMike Kravetz static int hugetlb_acct_memory(struct hstate *h, long delta); 948d9bfb26SMike Kravetz static void hugetlb_vma_lock_free(struct vm_area_struct *vma); 958d9bfb26SMike Kravetz static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma); 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 { 2633259914fSXU pengfei struct file_region *nrg; 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 3350739eb43SBaolin 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 { 3453259914fSXU pengfei struct file_region *nrg, *prg; 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 37184448c8eSJakob Koschel hugetlb_resv_map_add(struct resv_map *map, struct list_head *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); 38084448c8eSJakob Koschel list_add(&nrg->link, rg); 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; 40384448c8eSJakob Koschel struct file_region *iter, *trg = NULL; 40484448c8eSJakob Koschel struct list_head *rg = NULL; 405d75c6af9SMina Almasry 4060db9d74eSMina Almasry if (regions_needed) 4070db9d74eSMina Almasry *regions_needed = 0; 408d75c6af9SMina Almasry 4090db9d74eSMina Almasry /* In this loop, we essentially handle an entry for the range 41084448c8eSJakob Koschel * [last_accounted_offset, iter->from), at every iteration, with some 4110db9d74eSMina Almasry * bounds checking. 4120db9d74eSMina Almasry */ 41384448c8eSJakob Koschel list_for_each_entry_safe(iter, trg, head, link) { 4140db9d74eSMina Almasry /* Skip irrelevant regions that start before our range. */ 41584448c8eSJakob Koschel if (iter->from < f) { 4160db9d74eSMina Almasry /* If this region ends after the last accounted offset, 4170db9d74eSMina Almasry * then we need to update last_accounted_offset. 4180db9d74eSMina Almasry */ 41984448c8eSJakob Koschel if (iter->to > last_accounted_offset) 42084448c8eSJakob Koschel last_accounted_offset = iter->to; 4210db9d74eSMina Almasry continue; 4220db9d74eSMina Almasry } 423d75c6af9SMina Almasry 4240db9d74eSMina Almasry /* When we find a region that starts beyond our range, we've 4250db9d74eSMina Almasry * finished. 4260db9d74eSMina Almasry */ 42784448c8eSJakob Koschel if (iter->from >= t) { 42884448c8eSJakob Koschel rg = iter->link.prev; 429d75c6af9SMina Almasry break; 43084448c8eSJakob Koschel } 431d75c6af9SMina Almasry 43284448c8eSJakob Koschel /* Add an entry for last_accounted_offset -> iter->from, and 4330db9d74eSMina Almasry * update last_accounted_offset. 434d75c6af9SMina Almasry */ 43584448c8eSJakob Koschel if (iter->from > last_accounted_offset) 43684448c8eSJakob Koschel add += hugetlb_resv_map_add(resv, iter->link.prev, 4372103cf9cSPeter Xu last_accounted_offset, 43884448c8eSJakob Koschel iter->from, h, h_cg, 4392103cf9cSPeter Xu regions_needed); 440d75c6af9SMina Almasry 44184448c8eSJakob Koschel last_accounted_offset = iter->to; 4420db9d74eSMina Almasry } 4430db9d74eSMina Almasry 4440db9d74eSMina Almasry /* Handle the case where our range extends beyond 4450db9d74eSMina Almasry * last_accounted_offset. 4460db9d74eSMina Almasry */ 44784448c8eSJakob Koschel if (!rg) 44884448c8eSJakob Koschel rg = head->prev; 4492103cf9cSPeter Xu if (last_accounted_offset < t) 4502103cf9cSPeter Xu add += hugetlb_resv_map_add(resv, rg, last_accounted_offset, 4512103cf9cSPeter Xu t, h, h_cg, regions_needed); 4520db9d74eSMina Almasry 4530db9d74eSMina Almasry return add; 4540db9d74eSMina Almasry } 4550db9d74eSMina Almasry 4560db9d74eSMina Almasry /* Must be called with resv->lock acquired. Will drop lock to allocate entries. 4570db9d74eSMina Almasry */ 4580db9d74eSMina Almasry static int allocate_file_region_entries(struct resv_map *resv, 4590db9d74eSMina Almasry int regions_needed) 4600db9d74eSMina Almasry __must_hold(&resv->lock) 4610db9d74eSMina Almasry { 46234665341SMiaohe Lin LIST_HEAD(allocated_regions); 4630db9d74eSMina Almasry int to_allocate = 0, i = 0; 4640db9d74eSMina Almasry struct file_region *trg = NULL, *rg = NULL; 4650db9d74eSMina Almasry 4660db9d74eSMina Almasry VM_BUG_ON(regions_needed < 0); 4670db9d74eSMina Almasry 4680db9d74eSMina Almasry /* 4690db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 4700db9d74eSMina Almasry * the number of in progress add operations plus regions_needed. 4710db9d74eSMina Almasry * 4720db9d74eSMina Almasry * This is a while loop because when we drop the lock, some other call 4730db9d74eSMina Almasry * to region_add or region_del may have consumed some region_entries, 4740db9d74eSMina Almasry * so we keep looping here until we finally have enough entries for 4750db9d74eSMina Almasry * (adds_in_progress + regions_needed). 4760db9d74eSMina Almasry */ 4770db9d74eSMina Almasry while (resv->region_cache_count < 4780db9d74eSMina Almasry (resv->adds_in_progress + regions_needed)) { 4790db9d74eSMina Almasry to_allocate = resv->adds_in_progress + regions_needed - 4800db9d74eSMina Almasry resv->region_cache_count; 4810db9d74eSMina Almasry 4820db9d74eSMina Almasry /* At this point, we should have enough entries in the cache 483f0953a1bSIngo Molnar * for all the existing adds_in_progress. We should only be 4840db9d74eSMina Almasry * needing to allocate for regions_needed. 4850db9d74eSMina Almasry */ 4860db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); 4870db9d74eSMina Almasry 4880db9d74eSMina Almasry spin_unlock(&resv->lock); 4890db9d74eSMina Almasry for (i = 0; i < to_allocate; i++) { 4900db9d74eSMina Almasry trg = kmalloc(sizeof(*trg), GFP_KERNEL); 4910db9d74eSMina Almasry if (!trg) 4920db9d74eSMina Almasry goto out_of_memory; 4930db9d74eSMina Almasry list_add(&trg->link, &allocated_regions); 4940db9d74eSMina Almasry } 4950db9d74eSMina Almasry 4960db9d74eSMina Almasry spin_lock(&resv->lock); 4970db9d74eSMina Almasry 498d3ec7b6eSWei Yang list_splice(&allocated_regions, &resv->region_cache); 499d3ec7b6eSWei Yang resv->region_cache_count += to_allocate; 5000db9d74eSMina Almasry } 5010db9d74eSMina Almasry 5020db9d74eSMina Almasry return 0; 5030db9d74eSMina Almasry 5040db9d74eSMina Almasry out_of_memory: 5050db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, &allocated_regions, link) { 506d75c6af9SMina Almasry list_del(&rg->link); 507d75c6af9SMina Almasry kfree(rg); 508d75c6af9SMina Almasry } 5090db9d74eSMina Almasry return -ENOMEM; 510d75c6af9SMina Almasry } 511d75c6af9SMina Almasry 5121dd308a7SMike Kravetz /* 5131dd308a7SMike Kravetz * Add the huge page range represented by [f, t) to the reserve 5140db9d74eSMina Almasry * map. Regions will be taken from the cache to fill in this range. 5150db9d74eSMina Almasry * Sufficient regions should exist in the cache due to the previous 5160db9d74eSMina Almasry * call to region_chg with the same range, but in some cases the cache will not 5170db9d74eSMina Almasry * have sufficient entries due to races with other code doing region_add or 5180db9d74eSMina Almasry * region_del. The extra needed entries will be allocated. 519cf3ad20bSMike Kravetz * 5200db9d74eSMina Almasry * regions_needed is the out value provided by a previous call to region_chg. 5210db9d74eSMina Almasry * 5220db9d74eSMina Almasry * Return the number of new huge pages added to the map. This number is greater 5230db9d74eSMina Almasry * than or equal to zero. If file_region entries needed to be allocated for 5247c8de358SEthon Paul * this operation and we were not able to allocate, it returns -ENOMEM. 5250db9d74eSMina Almasry * region_add of regions of length 1 never allocate file_regions and cannot 5260db9d74eSMina Almasry * fail; region_chg will always allocate at least 1 entry and a region_add for 5270db9d74eSMina Almasry * 1 page will only require at most 1 entry. 5281dd308a7SMike Kravetz */ 5290db9d74eSMina Almasry static long region_add(struct resv_map *resv, long f, long t, 530075a61d0SMina Almasry long in_regions_needed, struct hstate *h, 531075a61d0SMina Almasry struct hugetlb_cgroup *h_cg) 53296822904SAndy Whitcroft { 5330db9d74eSMina Almasry long add = 0, actual_regions_needed = 0; 53496822904SAndy Whitcroft 5357b24d861SDavidlohr Bueso spin_lock(&resv->lock); 5360db9d74eSMina Almasry retry: 5370db9d74eSMina Almasry 5380db9d74eSMina Almasry /* Count how many regions are actually needed to execute this add. */ 539972a3da3SWei Yang add_reservation_in_range(resv, f, t, NULL, NULL, 540972a3da3SWei Yang &actual_regions_needed); 54196822904SAndy Whitcroft 5425e911373SMike Kravetz /* 5430db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 5440db9d74eSMina Almasry * this add operation. Note that actual_regions_needed may be greater 5450db9d74eSMina Almasry * than in_regions_needed, as the resv_map may have been modified since 5460db9d74eSMina Almasry * the region_chg call. In this case, we need to make sure that we 5470db9d74eSMina Almasry * allocate extra entries, such that we have enough for all the 5480db9d74eSMina Almasry * existing adds_in_progress, plus the excess needed for this 5490db9d74eSMina Almasry * operation. 5505e911373SMike Kravetz */ 5510db9d74eSMina Almasry if (actual_regions_needed > in_regions_needed && 5520db9d74eSMina Almasry resv->region_cache_count < 5530db9d74eSMina Almasry resv->adds_in_progress + 5540db9d74eSMina Almasry (actual_regions_needed - in_regions_needed)) { 5550db9d74eSMina Almasry /* region_add operation of range 1 should never need to 5560db9d74eSMina Almasry * allocate file_region entries. 5570db9d74eSMina Almasry */ 5580db9d74eSMina Almasry VM_BUG_ON(t - f <= 1); 5595e911373SMike Kravetz 5600db9d74eSMina Almasry if (allocate_file_region_entries( 5610db9d74eSMina Almasry resv, actual_regions_needed - in_regions_needed)) { 5620db9d74eSMina Almasry return -ENOMEM; 5635e911373SMike Kravetz } 5645e911373SMike Kravetz 5650db9d74eSMina Almasry goto retry; 5660db9d74eSMina Almasry } 567cf3ad20bSMike Kravetz 568972a3da3SWei Yang add = add_reservation_in_range(resv, f, t, h_cg, h, NULL); 5690db9d74eSMina Almasry 5700db9d74eSMina Almasry resv->adds_in_progress -= in_regions_needed; 5710db9d74eSMina Almasry 5727b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 573cf3ad20bSMike Kravetz return add; 57496822904SAndy Whitcroft } 57596822904SAndy Whitcroft 5761dd308a7SMike Kravetz /* 5771dd308a7SMike Kravetz * Examine the existing reserve map and determine how many 5781dd308a7SMike Kravetz * huge pages in the specified range [f, t) are NOT currently 5791dd308a7SMike Kravetz * represented. This routine is called before a subsequent 5801dd308a7SMike Kravetz * call to region_add that will actually modify the reserve 5811dd308a7SMike Kravetz * map to add the specified range [f, t). region_chg does 5821dd308a7SMike Kravetz * not change the number of huge pages represented by the 5830db9d74eSMina Almasry * map. A number of new file_region structures is added to the cache as a 5840db9d74eSMina Almasry * placeholder, for the subsequent region_add call to use. At least 1 5850db9d74eSMina Almasry * file_region structure is added. 5860db9d74eSMina Almasry * 5870db9d74eSMina Almasry * out_regions_needed is the number of regions added to the 5880db9d74eSMina Almasry * resv->adds_in_progress. This value needs to be provided to a follow up call 5890db9d74eSMina Almasry * to region_add or region_abort for proper accounting. 5905e911373SMike Kravetz * 5915e911373SMike Kravetz * Returns the number of huge pages that need to be added to the existing 5925e911373SMike Kravetz * reservation map for the range [f, t). This number is greater or equal to 5935e911373SMike Kravetz * zero. -ENOMEM is returned if a new file_region structure or cache entry 5945e911373SMike Kravetz * is needed and can not be allocated. 5951dd308a7SMike Kravetz */ 5960db9d74eSMina Almasry static long region_chg(struct resv_map *resv, long f, long t, 5970db9d74eSMina Almasry long *out_regions_needed) 59896822904SAndy Whitcroft { 59996822904SAndy Whitcroft long chg = 0; 60096822904SAndy Whitcroft 6017b24d861SDavidlohr Bueso spin_lock(&resv->lock); 6025e911373SMike Kravetz 603972a3da3SWei Yang /* Count how many hugepages in this range are NOT represented. */ 604075a61d0SMina Almasry chg = add_reservation_in_range(resv, f, t, NULL, NULL, 605972a3da3SWei Yang out_regions_needed); 6065e911373SMike Kravetz 6070db9d74eSMina Almasry if (*out_regions_needed == 0) 6080db9d74eSMina Almasry *out_regions_needed = 1; 6095e911373SMike Kravetz 6100db9d74eSMina Almasry if (allocate_file_region_entries(resv, *out_regions_needed)) 6115e911373SMike Kravetz return -ENOMEM; 6125e911373SMike Kravetz 6130db9d74eSMina Almasry resv->adds_in_progress += *out_regions_needed; 61496822904SAndy Whitcroft 6157b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 61696822904SAndy Whitcroft return chg; 61796822904SAndy Whitcroft } 61896822904SAndy Whitcroft 6191dd308a7SMike Kravetz /* 6205e911373SMike Kravetz * Abort the in progress add operation. The adds_in_progress field 6215e911373SMike Kravetz * of the resv_map keeps track of the operations in progress between 6225e911373SMike Kravetz * calls to region_chg and region_add. Operations are sometimes 6235e911373SMike Kravetz * aborted after the call to region_chg. In such cases, region_abort 6240db9d74eSMina Almasry * is called to decrement the adds_in_progress counter. regions_needed 6250db9d74eSMina Almasry * is the value returned by the region_chg call, it is used to decrement 6260db9d74eSMina Almasry * the adds_in_progress counter. 6275e911373SMike Kravetz * 6285e911373SMike Kravetz * NOTE: The range arguments [f, t) are not needed or used in this 6295e911373SMike Kravetz * routine. They are kept to make reading the calling code easier as 6305e911373SMike Kravetz * arguments will match the associated region_chg call. 6315e911373SMike Kravetz */ 6320db9d74eSMina Almasry static void region_abort(struct resv_map *resv, long f, long t, 6330db9d74eSMina Almasry long regions_needed) 6345e911373SMike Kravetz { 6355e911373SMike Kravetz spin_lock(&resv->lock); 6365e911373SMike Kravetz VM_BUG_ON(!resv->region_cache_count); 6370db9d74eSMina Almasry resv->adds_in_progress -= regions_needed; 6385e911373SMike Kravetz spin_unlock(&resv->lock); 6395e911373SMike Kravetz } 6405e911373SMike Kravetz 6415e911373SMike Kravetz /* 642feba16e2SMike Kravetz * Delete the specified range [f, t) from the reserve map. If the 643feba16e2SMike Kravetz * t parameter is LONG_MAX, this indicates that ALL regions after f 644feba16e2SMike Kravetz * should be deleted. Locate the regions which intersect [f, t) 645feba16e2SMike Kravetz * and either trim, delete or split the existing regions. 646feba16e2SMike Kravetz * 647feba16e2SMike Kravetz * Returns the number of huge pages deleted from the reserve map. 648feba16e2SMike Kravetz * In the normal case, the return value is zero or more. In the 649feba16e2SMike Kravetz * case where a region must be split, a new region descriptor must 650feba16e2SMike Kravetz * be allocated. If the allocation fails, -ENOMEM will be returned. 651feba16e2SMike Kravetz * NOTE: If the parameter t == LONG_MAX, then we will never split 652feba16e2SMike Kravetz * a region and possibly return -ENOMEM. Callers specifying 653feba16e2SMike Kravetz * t == LONG_MAX do not need to check for -ENOMEM error. 6541dd308a7SMike Kravetz */ 655feba16e2SMike Kravetz static long region_del(struct resv_map *resv, long f, long t) 65696822904SAndy Whitcroft { 6571406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 65896822904SAndy Whitcroft struct file_region *rg, *trg; 659feba16e2SMike Kravetz struct file_region *nrg = NULL; 660feba16e2SMike Kravetz long del = 0; 66196822904SAndy Whitcroft 662feba16e2SMike Kravetz retry: 6637b24d861SDavidlohr Bueso spin_lock(&resv->lock); 664feba16e2SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 665dbe409e4SMike Kravetz /* 666dbe409e4SMike Kravetz * Skip regions before the range to be deleted. file_region 667dbe409e4SMike Kravetz * ranges are normally of the form [from, to). However, there 668dbe409e4SMike Kravetz * may be a "placeholder" entry in the map which is of the form 669dbe409e4SMike Kravetz * (from, to) with from == to. Check for placeholder entries 670dbe409e4SMike Kravetz * at the beginning of the range to be deleted. 671dbe409e4SMike Kravetz */ 672dbe409e4SMike Kravetz if (rg->to <= f && (rg->to != rg->from || rg->to != f)) 673feba16e2SMike Kravetz continue; 674dbe409e4SMike Kravetz 675feba16e2SMike Kravetz if (rg->from >= t) 67696822904SAndy Whitcroft break; 67796822904SAndy Whitcroft 678feba16e2SMike Kravetz if (f > rg->from && t < rg->to) { /* Must split region */ 679feba16e2SMike Kravetz /* 680feba16e2SMike Kravetz * Check for an entry in the cache before dropping 681feba16e2SMike Kravetz * lock and attempting allocation. 682feba16e2SMike Kravetz */ 683feba16e2SMike Kravetz if (!nrg && 684feba16e2SMike Kravetz resv->region_cache_count > resv->adds_in_progress) { 685feba16e2SMike Kravetz nrg = list_first_entry(&resv->region_cache, 686feba16e2SMike Kravetz struct file_region, 687feba16e2SMike Kravetz link); 688feba16e2SMike Kravetz list_del(&nrg->link); 689feba16e2SMike Kravetz resv->region_cache_count--; 69096822904SAndy Whitcroft } 69196822904SAndy Whitcroft 692feba16e2SMike Kravetz if (!nrg) { 693feba16e2SMike Kravetz spin_unlock(&resv->lock); 694feba16e2SMike Kravetz nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 695feba16e2SMike Kravetz if (!nrg) 696feba16e2SMike Kravetz return -ENOMEM; 697feba16e2SMike Kravetz goto retry; 698feba16e2SMike Kravetz } 699feba16e2SMike Kravetz 700feba16e2SMike Kravetz del += t - f; 70179aa925bSMike Kravetz hugetlb_cgroup_uncharge_file_region( 702d85aecf2SMiaohe Lin resv, rg, t - f, false); 703feba16e2SMike Kravetz 704feba16e2SMike Kravetz /* New entry for end of split region */ 705feba16e2SMike Kravetz nrg->from = t; 706feba16e2SMike Kravetz nrg->to = rg->to; 707075a61d0SMina Almasry 708075a61d0SMina Almasry copy_hugetlb_cgroup_uncharge_info(nrg, rg); 709075a61d0SMina Almasry 710feba16e2SMike Kravetz INIT_LIST_HEAD(&nrg->link); 711feba16e2SMike Kravetz 712feba16e2SMike Kravetz /* Original entry is trimmed */ 713feba16e2SMike Kravetz rg->to = f; 714feba16e2SMike Kravetz 715feba16e2SMike Kravetz list_add(&nrg->link, &rg->link); 716feba16e2SMike Kravetz nrg = NULL; 71796822904SAndy Whitcroft break; 718feba16e2SMike Kravetz } 719feba16e2SMike Kravetz 720feba16e2SMike Kravetz if (f <= rg->from && t >= rg->to) { /* Remove entire region */ 721feba16e2SMike Kravetz del += rg->to - rg->from; 722075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 723d85aecf2SMiaohe Lin rg->to - rg->from, true); 72496822904SAndy Whitcroft list_del(&rg->link); 72596822904SAndy Whitcroft kfree(rg); 726feba16e2SMike Kravetz continue; 72796822904SAndy Whitcroft } 7287b24d861SDavidlohr Bueso 729feba16e2SMike Kravetz if (f <= rg->from) { /* Trim beginning of region */ 730075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 731d85aecf2SMiaohe Lin t - rg->from, false); 732075a61d0SMina Almasry 73379aa925bSMike Kravetz del += t - rg->from; 73479aa925bSMike Kravetz rg->from = t; 73579aa925bSMike Kravetz } else { /* Trim end of region */ 736075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 737d85aecf2SMiaohe Lin rg->to - f, false); 73879aa925bSMike Kravetz 73979aa925bSMike Kravetz del += rg->to - f; 74079aa925bSMike Kravetz rg->to = f; 741feba16e2SMike Kravetz } 742feba16e2SMike Kravetz } 743feba16e2SMike Kravetz 7447b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 745feba16e2SMike Kravetz kfree(nrg); 746feba16e2SMike Kravetz return del; 74796822904SAndy Whitcroft } 74896822904SAndy Whitcroft 7491dd308a7SMike Kravetz /* 750b5cec28dSMike Kravetz * A rare out of memory error was encountered which prevented removal of 751b5cec28dSMike Kravetz * the reserve map region for a page. The huge page itself was free'ed 752b5cec28dSMike Kravetz * and removed from the page cache. This routine will adjust the subpool 753b5cec28dSMike Kravetz * usage count, and the global reserve count if needed. By incrementing 754b5cec28dSMike Kravetz * these counts, the reserve map entry which could not be deleted will 755b5cec28dSMike Kravetz * appear as a "reserved" entry instead of simply dangling with incorrect 756b5cec28dSMike Kravetz * counts. 757b5cec28dSMike Kravetz */ 75872e2936cSzhong jiang void hugetlb_fix_reserve_counts(struct inode *inode) 759b5cec28dSMike Kravetz { 760b5cec28dSMike Kravetz struct hugepage_subpool *spool = subpool_inode(inode); 761b5cec28dSMike Kravetz long rsv_adjust; 762da56388cSMiaohe Lin bool reserved = false; 763b5cec28dSMike Kravetz 764b5cec28dSMike Kravetz rsv_adjust = hugepage_subpool_get_pages(spool, 1); 765da56388cSMiaohe Lin if (rsv_adjust > 0) { 766b5cec28dSMike Kravetz struct hstate *h = hstate_inode(inode); 767b5cec28dSMike Kravetz 768da56388cSMiaohe Lin if (!hugetlb_acct_memory(h, 1)) 769da56388cSMiaohe Lin reserved = true; 770da56388cSMiaohe Lin } else if (!rsv_adjust) { 771da56388cSMiaohe Lin reserved = true; 772b5cec28dSMike Kravetz } 773da56388cSMiaohe Lin 774da56388cSMiaohe Lin if (!reserved) 775da56388cSMiaohe Lin pr_warn("hugetlb: Huge Page Reserved count may go negative.\n"); 776b5cec28dSMike Kravetz } 777b5cec28dSMike Kravetz 778b5cec28dSMike Kravetz /* 7791dd308a7SMike Kravetz * Count and return the number of huge pages in the reserve map 7801dd308a7SMike Kravetz * that intersect with the range [f, t). 7811dd308a7SMike Kravetz */ 7821406ec9bSJoonsoo Kim static long region_count(struct resv_map *resv, long f, long t) 78384afd99bSAndy Whitcroft { 7841406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 78584afd99bSAndy Whitcroft struct file_region *rg; 78684afd99bSAndy Whitcroft long chg = 0; 78784afd99bSAndy Whitcroft 7887b24d861SDavidlohr Bueso spin_lock(&resv->lock); 78984afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 79084afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 791f2135a4aSWang Sheng-Hui long seg_from; 792f2135a4aSWang Sheng-Hui long seg_to; 79384afd99bSAndy Whitcroft 79484afd99bSAndy Whitcroft if (rg->to <= f) 79584afd99bSAndy Whitcroft continue; 79684afd99bSAndy Whitcroft if (rg->from >= t) 79784afd99bSAndy Whitcroft break; 79884afd99bSAndy Whitcroft 79984afd99bSAndy Whitcroft seg_from = max(rg->from, f); 80084afd99bSAndy Whitcroft seg_to = min(rg->to, t); 80184afd99bSAndy Whitcroft 80284afd99bSAndy Whitcroft chg += seg_to - seg_from; 80384afd99bSAndy Whitcroft } 8047b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 80584afd99bSAndy Whitcroft 80684afd99bSAndy Whitcroft return chg; 80784afd99bSAndy Whitcroft } 80884afd99bSAndy Whitcroft 80996822904SAndy Whitcroft /* 810e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 811e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 812e7c4b0bfSAndy Whitcroft */ 813a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 814a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 815e7c4b0bfSAndy Whitcroft { 816a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 817a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 818e7c4b0bfSAndy Whitcroft } 819e7c4b0bfSAndy Whitcroft 8200fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 8210fe6e20bSNaoya Horiguchi unsigned long address) 8220fe6e20bSNaoya Horiguchi { 8230fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 8240fe6e20bSNaoya Horiguchi } 825dee41079SDan Williams EXPORT_SYMBOL_GPL(linear_hugepage_index); 8260fe6e20bSNaoya Horiguchi 82784afd99bSAndy Whitcroft /* 82808fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 82908fba699SMel Gorman * cases this will be same size as used by the page table entries. 83008fba699SMel Gorman */ 83108fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 83208fba699SMel Gorman { 83305ea8860SDan Williams if (vma->vm_ops && vma->vm_ops->pagesize) 83405ea8860SDan Williams return vma->vm_ops->pagesize(vma); 83508fba699SMel Gorman return PAGE_SIZE; 83608fba699SMel Gorman } 837f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 83808fba699SMel Gorman 83908fba699SMel Gorman /* 8403340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 8413340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 84209135cc5SDan Williams * architectures where it differs, an architecture-specific 'strong' 84309135cc5SDan Williams * version of this symbol is required. 8443340289dSMel Gorman */ 84509135cc5SDan Williams __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 8463340289dSMel Gorman { 8473340289dSMel Gorman return vma_kernel_pagesize(vma); 8483340289dSMel Gorman } 8493340289dSMel Gorman 8503340289dSMel Gorman /* 85184afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 85284afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 85384afd99bSAndy Whitcroft * alignment. 85484afd99bSAndy Whitcroft */ 85584afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 85684afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 85704f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 85884afd99bSAndy Whitcroft 859a1e78772SMel Gorman /* 860a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 861a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 862a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 863a1e78772SMel Gorman * 8648d9bfb26SMike Kravetz * With the exception of hugetlb_dup_vma_private() which is called at fork(), 865a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 866a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 867a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 86884afd99bSAndy Whitcroft * 86984afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 87084afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 87184afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 87284afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 87384afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 87484afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 87584afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 87684afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 877a1e78772SMel Gorman */ 878e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 879e7c4b0bfSAndy Whitcroft { 880e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 881e7c4b0bfSAndy Whitcroft } 882e7c4b0bfSAndy Whitcroft 883e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 884e7c4b0bfSAndy Whitcroft unsigned long value) 885e7c4b0bfSAndy Whitcroft { 886e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 887e7c4b0bfSAndy Whitcroft } 888e7c4b0bfSAndy Whitcroft 889e9fe92aeSMina Almasry static void 890e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, 891e9fe92aeSMina Almasry struct hugetlb_cgroup *h_cg, 892e9fe92aeSMina Almasry struct hstate *h) 893e9fe92aeSMina Almasry { 894e9fe92aeSMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 895e9fe92aeSMina Almasry if (!h_cg || !h) { 896e9fe92aeSMina Almasry resv_map->reservation_counter = NULL; 897e9fe92aeSMina Almasry resv_map->pages_per_hpage = 0; 898e9fe92aeSMina Almasry resv_map->css = NULL; 899e9fe92aeSMina Almasry } else { 900e9fe92aeSMina Almasry resv_map->reservation_counter = 901e9fe92aeSMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 902e9fe92aeSMina Almasry resv_map->pages_per_hpage = pages_per_huge_page(h); 903e9fe92aeSMina Almasry resv_map->css = &h_cg->css; 904e9fe92aeSMina Almasry } 905e9fe92aeSMina Almasry #endif 906e9fe92aeSMina Almasry } 907e9fe92aeSMina Almasry 9089119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 90984afd99bSAndy Whitcroft { 91084afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 9115e911373SMike Kravetz struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 9125e911373SMike Kravetz 9135e911373SMike Kravetz if (!resv_map || !rg) { 9145e911373SMike Kravetz kfree(resv_map); 9155e911373SMike Kravetz kfree(rg); 91684afd99bSAndy Whitcroft return NULL; 9175e911373SMike Kravetz } 91884afd99bSAndy Whitcroft 91984afd99bSAndy Whitcroft kref_init(&resv_map->refs); 9207b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 92184afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 92284afd99bSAndy Whitcroft 9235e911373SMike Kravetz resv_map->adds_in_progress = 0; 924e9fe92aeSMina Almasry /* 925e9fe92aeSMina Almasry * Initialize these to 0. On shared mappings, 0's here indicate these 926e9fe92aeSMina Almasry * fields don't do cgroup accounting. On private mappings, these will be 927e9fe92aeSMina Almasry * re-initialized to the proper values, to indicate that hugetlb cgroup 928e9fe92aeSMina Almasry * reservations are to be un-charged from here. 929e9fe92aeSMina Almasry */ 930e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); 9315e911373SMike Kravetz 9325e911373SMike Kravetz INIT_LIST_HEAD(&resv_map->region_cache); 9335e911373SMike Kravetz list_add(&rg->link, &resv_map->region_cache); 9345e911373SMike Kravetz resv_map->region_cache_count = 1; 9355e911373SMike Kravetz 93684afd99bSAndy Whitcroft return resv_map; 93784afd99bSAndy Whitcroft } 93884afd99bSAndy Whitcroft 9399119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 94084afd99bSAndy Whitcroft { 94184afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 9425e911373SMike Kravetz struct list_head *head = &resv_map->region_cache; 9435e911373SMike Kravetz struct file_region *rg, *trg; 94484afd99bSAndy Whitcroft 94584afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 946feba16e2SMike Kravetz region_del(resv_map, 0, LONG_MAX); 9475e911373SMike Kravetz 9485e911373SMike Kravetz /* ... and any entries left in the cache */ 9495e911373SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 9505e911373SMike Kravetz list_del(&rg->link); 9515e911373SMike Kravetz kfree(rg); 9525e911373SMike Kravetz } 9535e911373SMike Kravetz 9545e911373SMike Kravetz VM_BUG_ON(resv_map->adds_in_progress); 9555e911373SMike Kravetz 95684afd99bSAndy Whitcroft kfree(resv_map); 95784afd99bSAndy Whitcroft } 95884afd99bSAndy Whitcroft 9594e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 9604e35f483SJoonsoo Kim { 961f27a5136SMike Kravetz /* 962f27a5136SMike Kravetz * At inode evict time, i_mapping may not point to the original 963f27a5136SMike Kravetz * address space within the inode. This original address space 964f27a5136SMike Kravetz * contains the pointer to the resv_map. So, always use the 965f27a5136SMike Kravetz * address space embedded within the inode. 966f27a5136SMike Kravetz * The VERY common case is inode->mapping == &inode->i_data but, 967f27a5136SMike Kravetz * this may not be true for device special inodes. 968f27a5136SMike Kravetz */ 969f27a5136SMike Kravetz return (struct resv_map *)(&inode->i_data)->private_data; 9704e35f483SJoonsoo Kim } 9714e35f483SJoonsoo Kim 97284afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 973a1e78772SMel Gorman { 97481d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 9754e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 9764e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 9774e35f483SJoonsoo Kim struct inode *inode = mapping->host; 9784e35f483SJoonsoo Kim 9794e35f483SJoonsoo Kim return inode_resv_map(inode); 9804e35f483SJoonsoo Kim 9814e35f483SJoonsoo Kim } else { 98284afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 98384afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 9844e35f483SJoonsoo Kim } 985a1e78772SMel Gorman } 986a1e78772SMel Gorman 98784afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 988a1e78772SMel Gorman { 98981d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 99081d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 991a1e78772SMel Gorman 99284afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 99384afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 99404f2cbe3SMel Gorman } 99504f2cbe3SMel Gorman 99604f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 99704f2cbe3SMel Gorman { 99881d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 99981d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 1000e7c4b0bfSAndy Whitcroft 1001e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 100204f2cbe3SMel Gorman } 100304f2cbe3SMel Gorman 100404f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 100504f2cbe3SMel Gorman { 100681d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 1007e7c4b0bfSAndy Whitcroft 1008e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 1009a1e78772SMel Gorman } 1010a1e78772SMel Gorman 10118d9bfb26SMike Kravetz void hugetlb_dup_vma_private(struct vm_area_struct *vma) 1012a1e78772SMel Gorman { 101381d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 10148d9bfb26SMike Kravetz /* 10158d9bfb26SMike Kravetz * Clear vm_private_data 10168d9bfb26SMike Kravetz * - For MAP_PRIVATE mappings, this is the reserve map which does 10178d9bfb26SMike Kravetz * not apply to children. Faults generated by the children are 10188d9bfb26SMike Kravetz * not guaranteed to succeed, even if read-only. 10198d9bfb26SMike Kravetz * - For shared mappings this is a per-vma semaphore that may be 10208d9bfb26SMike Kravetz * allocated in a subsequent call to hugetlb_vm_op_open. 10218d9bfb26SMike Kravetz */ 1022a1e78772SMel Gorman vma->vm_private_data = (void *)0; 10238d9bfb26SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 10248d9bfb26SMike Kravetz return; 1025a1e78772SMel Gorman } 1026a1e78772SMel Gorman 1027550a7d60SMina Almasry /* 1028550a7d60SMina Almasry * Reset and decrement one ref on hugepage private reservation. 1029550a7d60SMina Almasry * Called with mm->mmap_sem writer semaphore held. 1030550a7d60SMina Almasry * This function should be only used by move_vma() and operate on 1031550a7d60SMina Almasry * same sized vma. It should never come here with last ref on the 1032550a7d60SMina Almasry * reservation. 1033550a7d60SMina Almasry */ 1034550a7d60SMina Almasry void clear_vma_resv_huge_pages(struct vm_area_struct *vma) 1035550a7d60SMina Almasry { 1036550a7d60SMina Almasry /* 1037550a7d60SMina Almasry * Clear the old hugetlb private page reservation. 1038550a7d60SMina Almasry * It has already been transferred to new_vma. 1039550a7d60SMina Almasry * 1040550a7d60SMina Almasry * During a mremap() operation of a hugetlb vma we call move_vma() 1041550a7d60SMina Almasry * which copies vma into new_vma and unmaps vma. After the copy 1042550a7d60SMina Almasry * operation both new_vma and vma share a reference to the resv_map 1043550a7d60SMina Almasry * struct, and at that point vma is about to be unmapped. We don't 1044550a7d60SMina Almasry * want to return the reservation to the pool at unmap of vma because 1045550a7d60SMina Almasry * the reservation still lives on in new_vma, so simply decrement the 1046550a7d60SMina Almasry * ref here and remove the resv_map reference from this vma. 1047550a7d60SMina Almasry */ 1048550a7d60SMina Almasry struct resv_map *reservations = vma_resv_map(vma); 1049550a7d60SMina Almasry 1050afe041c2SBui Quang Minh if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 1051afe041c2SBui Quang Minh resv_map_put_hugetlb_cgroup_uncharge_info(reservations); 1052550a7d60SMina Almasry kref_put(&reservations->refs, resv_map_release); 1053afe041c2SBui Quang Minh } 1054550a7d60SMina Almasry 10558d9bfb26SMike Kravetz hugetlb_dup_vma_private(vma); 1056550a7d60SMina Almasry } 1057550a7d60SMina Almasry 1058a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 1059559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 1060a1e78772SMel Gorman { 1061af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 1062af0ed73eSJoonsoo Kim /* 1063af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 1064af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 1065af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 1066af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 1067af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 1068af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 1069af0ed73eSJoonsoo Kim * properly, so add work-around here. 1070af0ed73eSJoonsoo Kim */ 1071af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 1072559ec2f8SNicholas Krause return true; 1073af0ed73eSJoonsoo Kim else 1074559ec2f8SNicholas Krause return false; 1075af0ed73eSJoonsoo Kim } 1076a63884e9SJoonsoo Kim 1077a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 10781fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 10791fb1b0e9SMike Kravetz /* 10801fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 10811fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 10821fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 10837c8de358SEthon Paul * fallocate. In this case, there really are no reserves to 10841fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 10851fb1b0e9SMike Kravetz */ 10861fb1b0e9SMike Kravetz if (chg) 10871fb1b0e9SMike Kravetz return false; 10881fb1b0e9SMike Kravetz else 1089559ec2f8SNicholas Krause return true; 10901fb1b0e9SMike Kravetz } 1091a63884e9SJoonsoo Kim 1092a63884e9SJoonsoo Kim /* 1093a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 1094a63884e9SJoonsoo Kim * private mappings. 1095a63884e9SJoonsoo Kim */ 109667961f9dSMike Kravetz if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 109767961f9dSMike Kravetz /* 109867961f9dSMike Kravetz * Like the shared case above, a hole punch or truncate 109967961f9dSMike Kravetz * could have been performed on the private mapping. 110067961f9dSMike Kravetz * Examine the value of chg to determine if reserves 110167961f9dSMike Kravetz * actually exist or were previously consumed. 110267961f9dSMike Kravetz * Very Subtle - The value of chg comes from a previous 110367961f9dSMike Kravetz * call to vma_needs_reserves(). The reserve map for 110467961f9dSMike Kravetz * private mappings has different (opposite) semantics 110567961f9dSMike Kravetz * than that of shared mappings. vma_needs_reserves() 110667961f9dSMike Kravetz * has already taken this difference in semantics into 110767961f9dSMike Kravetz * account. Therefore, the meaning of chg is the same 110867961f9dSMike Kravetz * as in the shared case above. Code could easily be 110967961f9dSMike Kravetz * combined, but keeping it separate draws attention to 111067961f9dSMike Kravetz * subtle differences. 111167961f9dSMike Kravetz */ 111267961f9dSMike Kravetz if (chg) 111367961f9dSMike Kravetz return false; 111467961f9dSMike Kravetz else 1115559ec2f8SNicholas Krause return true; 111667961f9dSMike Kravetz } 1117a63884e9SJoonsoo Kim 1118559ec2f8SNicholas Krause return false; 1119a1e78772SMel Gorman } 1120a1e78772SMel Gorman 1121a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 11221da177e4SLinus Torvalds { 11231da177e4SLinus Torvalds int nid = page_to_nid(page); 11249487ca60SMike Kravetz 11259487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1126b65a4edaSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1127b65a4edaSMike Kravetz 11280edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 1129a5516438SAndi Kleen h->free_huge_pages++; 1130a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 11316c037149SMike Kravetz SetHPageFreed(page); 11321da177e4SLinus Torvalds } 11331da177e4SLinus Torvalds 113494310cbcSAnshuman Khandual static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) 1135bf50bab2SNaoya Horiguchi { 1136bf50bab2SNaoya Horiguchi struct page *page; 11371a08ae36SPavel Tatashin bool pin = !!(current->flags & PF_MEMALLOC_PIN); 1138bf50bab2SNaoya Horiguchi 11399487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1140bbe88753SJoonsoo Kim list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { 11416077c943SAlex Sierra if (pin && !is_longterm_pinnable_page(page)) 1142bbe88753SJoonsoo Kim continue; 1143bbe88753SJoonsoo Kim 11446664bfc8SWei Yang if (PageHWPoison(page)) 11456664bfc8SWei Yang continue; 1146bbe88753SJoonsoo Kim 11470edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 1148a9869b83SNaoya Horiguchi set_page_refcounted(page); 11496c037149SMike Kravetz ClearHPageFreed(page); 1150bf50bab2SNaoya Horiguchi h->free_huge_pages--; 1151bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 1152bf50bab2SNaoya Horiguchi return page; 1153bf50bab2SNaoya Horiguchi } 1154bf50bab2SNaoya Horiguchi 11556664bfc8SWei Yang return NULL; 11566664bfc8SWei Yang } 11576664bfc8SWei Yang 11583e59fcb0SMichal Hocko static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, 11593e59fcb0SMichal Hocko nodemask_t *nmask) 116094310cbcSAnshuman Khandual { 11613e59fcb0SMichal Hocko unsigned int cpuset_mems_cookie; 11623e59fcb0SMichal Hocko struct zonelist *zonelist; 11633e59fcb0SMichal Hocko struct zone *zone; 11643e59fcb0SMichal Hocko struct zoneref *z; 116598fa15f3SAnshuman Khandual int node = NUMA_NO_NODE; 11663e59fcb0SMichal Hocko 11673e59fcb0SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 11683e59fcb0SMichal Hocko 11693e59fcb0SMichal Hocko retry_cpuset: 11703e59fcb0SMichal Hocko cpuset_mems_cookie = read_mems_allowed_begin(); 11713e59fcb0SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 117294310cbcSAnshuman Khandual struct page *page; 117394310cbcSAnshuman Khandual 11743e59fcb0SMichal Hocko if (!cpuset_zone_allowed(zone, gfp_mask)) 11753e59fcb0SMichal Hocko continue; 11763e59fcb0SMichal Hocko /* 11773e59fcb0SMichal Hocko * no need to ask again on the same node. Pool is node rather than 11783e59fcb0SMichal Hocko * zone aware 11793e59fcb0SMichal Hocko */ 11803e59fcb0SMichal Hocko if (zone_to_nid(zone) == node) 11813e59fcb0SMichal Hocko continue; 11823e59fcb0SMichal Hocko node = zone_to_nid(zone); 118394310cbcSAnshuman Khandual 118494310cbcSAnshuman Khandual page = dequeue_huge_page_node_exact(h, node); 118594310cbcSAnshuman Khandual if (page) 118694310cbcSAnshuman Khandual return page; 118794310cbcSAnshuman Khandual } 11883e59fcb0SMichal Hocko if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 11893e59fcb0SMichal Hocko goto retry_cpuset; 11903e59fcb0SMichal Hocko 119194310cbcSAnshuman Khandual return NULL; 119294310cbcSAnshuman Khandual } 119394310cbcSAnshuman Khandual 1194a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 1195a5516438SAndi Kleen struct vm_area_struct *vma, 1196af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 1197af0ed73eSJoonsoo Kim long chg) 11981da177e4SLinus Torvalds { 1199cfcaa66fSBen Widawsky struct page *page = NULL; 1200480eccf9SLee Schermerhorn struct mempolicy *mpol; 120104ec6264SVlastimil Babka gfp_t gfp_mask; 12023e59fcb0SMichal Hocko nodemask_t *nodemask; 120304ec6264SVlastimil Babka int nid; 12041da177e4SLinus Torvalds 1205a1e78772SMel Gorman /* 1206a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 1207a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 1208a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 1209a1e78772SMel Gorman */ 1210af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg) && 1211a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 1212c0ff7453SMiao Xie goto err; 1213a1e78772SMel Gorman 121404f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 1215a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 12166eab04a8SJustin P. Mattock goto err; 121704f2cbe3SMel Gorman 121804ec6264SVlastimil Babka gfp_mask = htlb_alloc_mask(h); 121904ec6264SVlastimil Babka nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 1220cfcaa66fSBen Widawsky 1221cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 12223e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 1223cfcaa66fSBen Widawsky 1224cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 1225cfcaa66fSBen Widawsky nodemask = NULL; 1226cfcaa66fSBen Widawsky } 1227cfcaa66fSBen Widawsky 1228cfcaa66fSBen Widawsky if (!page) 1229cfcaa66fSBen Widawsky page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 1230cfcaa66fSBen Widawsky 12313e59fcb0SMichal Hocko if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { 1232d6995da3SMike Kravetz SetHPageRestoreReserve(page); 1233a63884e9SJoonsoo Kim h->resv_huge_pages--; 1234bf50bab2SNaoya Horiguchi } 1235cc9a6c87SMel Gorman 1236cc9a6c87SMel Gorman mpol_cond_put(mpol); 1237cc9a6c87SMel Gorman return page; 1238cc9a6c87SMel Gorman 1239c0ff7453SMiao Xie err: 1240cc9a6c87SMel Gorman return NULL; 12411da177e4SLinus Torvalds } 12421da177e4SLinus Torvalds 12431cac6f2cSLuiz Capitulino /* 12441cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 12451cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 12461cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 12471cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 12481cac6f2cSLuiz Capitulino * node for alloc or free. 12491cac6f2cSLuiz Capitulino */ 12501cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 12511cac6f2cSLuiz Capitulino { 12520edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 12531cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 12541cac6f2cSLuiz Capitulino 12551cac6f2cSLuiz Capitulino return nid; 12561cac6f2cSLuiz Capitulino } 12571cac6f2cSLuiz Capitulino 12581cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 12591cac6f2cSLuiz Capitulino { 12601cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 12611cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 12621cac6f2cSLuiz Capitulino return nid; 12631cac6f2cSLuiz Capitulino } 12641cac6f2cSLuiz Capitulino 12651cac6f2cSLuiz Capitulino /* 12661cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 12671cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 12681cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 12691cac6f2cSLuiz Capitulino * mask. 12701cac6f2cSLuiz Capitulino */ 12711cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 12721cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 12731cac6f2cSLuiz Capitulino { 12741cac6f2cSLuiz Capitulino int nid; 12751cac6f2cSLuiz Capitulino 12761cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 12771cac6f2cSLuiz Capitulino 12781cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 12791cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 12801cac6f2cSLuiz Capitulino 12811cac6f2cSLuiz Capitulino return nid; 12821cac6f2cSLuiz Capitulino } 12831cac6f2cSLuiz Capitulino 12841cac6f2cSLuiz Capitulino /* 128510c6ec49SMike Kravetz * helper for remove_pool_huge_page() - return the previously saved 12861cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 12871cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 12881cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 12891cac6f2cSLuiz Capitulino */ 12901cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 12911cac6f2cSLuiz Capitulino { 12921cac6f2cSLuiz Capitulino int nid; 12931cac6f2cSLuiz Capitulino 12941cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 12951cac6f2cSLuiz Capitulino 12961cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 12971cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 12981cac6f2cSLuiz Capitulino 12991cac6f2cSLuiz Capitulino return nid; 13001cac6f2cSLuiz Capitulino } 13011cac6f2cSLuiz Capitulino 13021cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 13031cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 13041cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 13051cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 13061cac6f2cSLuiz Capitulino nr_nodes--) 13071cac6f2cSLuiz Capitulino 13081cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 13091cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 13101cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 13111cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 13121cac6f2cSLuiz Capitulino nr_nodes--) 13131cac6f2cSLuiz Capitulino 13148531fc6fSMike Kravetz /* used to demote non-gigantic_huge pages as well */ 131534d9e35bSMike Kravetz static void __destroy_compound_gigantic_page(struct page *page, 131634d9e35bSMike Kravetz unsigned int order, bool demote) 1317944d9fecSLuiz Capitulino { 1318944d9fecSLuiz Capitulino int i; 1319944d9fecSLuiz Capitulino int nr_pages = 1 << order; 132014455eabSCheng Li struct page *p; 1321944d9fecSLuiz Capitulino 1322c8cc708aSGerald Schaefer atomic_set(compound_mapcount_ptr(page), 0); 132347e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 132447e29d32SJohn Hubbard 132514455eabSCheng Li for (i = 1; i < nr_pages; i++) { 132614455eabSCheng Li p = nth_page(page, i); 1327a01f4390SMike Kravetz p->mapping = NULL; 13281d798ca3SKirill A. Shutemov clear_compound_head(p); 132934d9e35bSMike Kravetz if (!demote) 1330944d9fecSLuiz Capitulino set_page_refcounted(p); 1331944d9fecSLuiz Capitulino } 1332944d9fecSLuiz Capitulino 1333944d9fecSLuiz Capitulino set_compound_order(page, 0); 13345232c63fSMatthew Wilcox (Oracle) #ifdef CONFIG_64BIT 1335ba9c1201SGerald Schaefer page[1].compound_nr = 0; 13365232c63fSMatthew Wilcox (Oracle) #endif 1337944d9fecSLuiz Capitulino __ClearPageHead(page); 1338944d9fecSLuiz Capitulino } 1339944d9fecSLuiz Capitulino 13408531fc6fSMike Kravetz static void destroy_compound_hugetlb_page_for_demote(struct page *page, 13418531fc6fSMike Kravetz unsigned int order) 13428531fc6fSMike Kravetz { 13438531fc6fSMike Kravetz __destroy_compound_gigantic_page(page, order, true); 13448531fc6fSMike Kravetz } 13458531fc6fSMike Kravetz 13468531fc6fSMike Kravetz #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 134734d9e35bSMike Kravetz static void destroy_compound_gigantic_page(struct page *page, 134834d9e35bSMike Kravetz unsigned int order) 134934d9e35bSMike Kravetz { 135034d9e35bSMike Kravetz __destroy_compound_gigantic_page(page, order, false); 135134d9e35bSMike Kravetz } 135234d9e35bSMike Kravetz 1353d00181b9SKirill A. Shutemov static void free_gigantic_page(struct page *page, unsigned int order) 1354944d9fecSLuiz Capitulino { 1355cf11e85fSRoman Gushchin /* 1356cf11e85fSRoman Gushchin * If the page isn't allocated using the cma allocator, 1357cf11e85fSRoman Gushchin * cma_release() returns false. 1358cf11e85fSRoman Gushchin */ 1359dbda8feaSBarry Song #ifdef CONFIG_CMA 1360dbda8feaSBarry Song if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order)) 1361cf11e85fSRoman Gushchin return; 1362dbda8feaSBarry Song #endif 1363cf11e85fSRoman Gushchin 1364944d9fecSLuiz Capitulino free_contig_range(page_to_pfn(page), 1 << order); 1365944d9fecSLuiz Capitulino } 1366944d9fecSLuiz Capitulino 13674eb0716eSAlexandre Ghiti #ifdef CONFIG_CONTIG_ALLOC 1368d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 1369d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) 1370944d9fecSLuiz Capitulino { 137104adbc3fSMiaohe Lin unsigned long nr_pages = pages_per_huge_page(h); 1372953f064aSLi Xinhai if (nid == NUMA_NO_NODE) 1373953f064aSLi Xinhai nid = numa_mem_id(); 1374944d9fecSLuiz Capitulino 1375dbda8feaSBarry Song #ifdef CONFIG_CMA 1376dbda8feaSBarry Song { 1377cf11e85fSRoman Gushchin struct page *page; 1378cf11e85fSRoman Gushchin int node; 1379cf11e85fSRoman Gushchin 1380953f064aSLi Xinhai if (hugetlb_cma[nid]) { 1381953f064aSLi Xinhai page = cma_alloc(hugetlb_cma[nid], nr_pages, 1382953f064aSLi Xinhai huge_page_order(h), true); 1383953f064aSLi Xinhai if (page) 1384953f064aSLi Xinhai return page; 1385953f064aSLi Xinhai } 1386953f064aSLi Xinhai 1387953f064aSLi Xinhai if (!(gfp_mask & __GFP_THISNODE)) { 1388cf11e85fSRoman Gushchin for_each_node_mask(node, *nodemask) { 1389953f064aSLi Xinhai if (node == nid || !hugetlb_cma[node]) 1390cf11e85fSRoman Gushchin continue; 1391cf11e85fSRoman Gushchin 1392cf11e85fSRoman Gushchin page = cma_alloc(hugetlb_cma[node], nr_pages, 1393cf11e85fSRoman Gushchin huge_page_order(h), true); 1394cf11e85fSRoman Gushchin if (page) 1395cf11e85fSRoman Gushchin return page; 1396cf11e85fSRoman Gushchin } 1397cf11e85fSRoman Gushchin } 1398953f064aSLi Xinhai } 1399dbda8feaSBarry Song #endif 1400cf11e85fSRoman Gushchin 14015e27a2dfSAnshuman Khandual return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); 1402944d9fecSLuiz Capitulino } 1403944d9fecSLuiz Capitulino 14044eb0716eSAlexandre Ghiti #else /* !CONFIG_CONTIG_ALLOC */ 14054eb0716eSAlexandre Ghiti static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 14064eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 14074eb0716eSAlexandre Ghiti { 14084eb0716eSAlexandre Ghiti return NULL; 14094eb0716eSAlexandre Ghiti } 14104eb0716eSAlexandre Ghiti #endif /* CONFIG_CONTIG_ALLOC */ 1411944d9fecSLuiz Capitulino 1412e1073d1eSAneesh Kumar K.V #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 1413d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 14144eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 14154eb0716eSAlexandre Ghiti { 14164eb0716eSAlexandre Ghiti return NULL; 14174eb0716eSAlexandre Ghiti } 1418d00181b9SKirill A. Shutemov static inline void free_gigantic_page(struct page *page, unsigned int order) { } 1419944d9fecSLuiz Capitulino static inline void destroy_compound_gigantic_page(struct page *page, 1420d00181b9SKirill A. Shutemov unsigned int order) { } 1421944d9fecSLuiz Capitulino #endif 1422944d9fecSLuiz Capitulino 14236eb4e88aSMike Kravetz /* 14246eb4e88aSMike Kravetz * Remove hugetlb page from lists, and update dtor so that page appears 142534d9e35bSMike Kravetz * as just a compound page. 142634d9e35bSMike Kravetz * 142734d9e35bSMike Kravetz * A reference is held on the page, except in the case of demote. 14286eb4e88aSMike Kravetz * 14296eb4e88aSMike Kravetz * Must be called with hugetlb lock held. 14306eb4e88aSMike Kravetz */ 143134d9e35bSMike Kravetz static void __remove_hugetlb_page(struct hstate *h, struct page *page, 143234d9e35bSMike Kravetz bool adjust_surplus, 143334d9e35bSMike Kravetz bool demote) 14346eb4e88aSMike Kravetz { 14356eb4e88aSMike Kravetz int nid = page_to_nid(page); 14366eb4e88aSMike Kravetz 14376eb4e88aSMike Kravetz VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 14386eb4e88aSMike Kravetz VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page); 14396eb4e88aSMike Kravetz 14409487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 14416eb4e88aSMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 14426eb4e88aSMike Kravetz return; 14436eb4e88aSMike Kravetz 14446eb4e88aSMike Kravetz list_del(&page->lru); 14456eb4e88aSMike Kravetz 14466eb4e88aSMike Kravetz if (HPageFreed(page)) { 14476eb4e88aSMike Kravetz h->free_huge_pages--; 14486eb4e88aSMike Kravetz h->free_huge_pages_node[nid]--; 14496eb4e88aSMike Kravetz } 14506eb4e88aSMike Kravetz if (adjust_surplus) { 14516eb4e88aSMike Kravetz h->surplus_huge_pages--; 14526eb4e88aSMike Kravetz h->surplus_huge_pages_node[nid]--; 14536eb4e88aSMike Kravetz } 14546eb4e88aSMike Kravetz 1455e32d20c0SMike Kravetz /* 1456e32d20c0SMike Kravetz * Very subtle 1457e32d20c0SMike Kravetz * 1458e32d20c0SMike Kravetz * For non-gigantic pages set the destructor to the normal compound 1459e32d20c0SMike Kravetz * page dtor. This is needed in case someone takes an additional 1460e32d20c0SMike Kravetz * temporary ref to the page, and freeing is delayed until they drop 1461e32d20c0SMike Kravetz * their reference. 1462e32d20c0SMike Kravetz * 1463e32d20c0SMike Kravetz * For gigantic pages set the destructor to the null dtor. This 1464e32d20c0SMike Kravetz * destructor will never be called. Before freeing the gigantic 1465e32d20c0SMike Kravetz * page destroy_compound_gigantic_page will turn the compound page 1466e32d20c0SMike Kravetz * into a simple group of pages. After this the destructor does not 1467e32d20c0SMike Kravetz * apply. 1468e32d20c0SMike Kravetz * 1469e32d20c0SMike Kravetz * This handles the case where more than one ref is held when and 1470e32d20c0SMike Kravetz * after update_and_free_page is called. 147134d9e35bSMike Kravetz * 147234d9e35bSMike Kravetz * In the case of demote we do not ref count the page as it will soon 147334d9e35bSMike Kravetz * be turned into a page of smaller size. 1474e32d20c0SMike Kravetz */ 147534d9e35bSMike Kravetz if (!demote) 14766eb4e88aSMike Kravetz set_page_refcounted(page); 1477e32d20c0SMike Kravetz if (hstate_is_gigantic(h)) 14786eb4e88aSMike Kravetz set_compound_page_dtor(page, NULL_COMPOUND_DTOR); 1479e32d20c0SMike Kravetz else 1480e32d20c0SMike Kravetz set_compound_page_dtor(page, COMPOUND_PAGE_DTOR); 14816eb4e88aSMike Kravetz 14826eb4e88aSMike Kravetz h->nr_huge_pages--; 14836eb4e88aSMike Kravetz h->nr_huge_pages_node[nid]--; 14846eb4e88aSMike Kravetz } 14856eb4e88aSMike Kravetz 148634d9e35bSMike Kravetz static void remove_hugetlb_page(struct hstate *h, struct page *page, 148734d9e35bSMike Kravetz bool adjust_surplus) 148834d9e35bSMike Kravetz { 148934d9e35bSMike Kravetz __remove_hugetlb_page(h, page, adjust_surplus, false); 149034d9e35bSMike Kravetz } 149134d9e35bSMike Kravetz 14928531fc6fSMike Kravetz static void remove_hugetlb_page_for_demote(struct hstate *h, struct page *page, 14938531fc6fSMike Kravetz bool adjust_surplus) 14948531fc6fSMike Kravetz { 14958531fc6fSMike Kravetz __remove_hugetlb_page(h, page, adjust_surplus, true); 14968531fc6fSMike Kravetz } 14978531fc6fSMike Kravetz 1498ad2fa371SMuchun Song static void add_hugetlb_page(struct hstate *h, struct page *page, 1499ad2fa371SMuchun Song bool adjust_surplus) 1500ad2fa371SMuchun Song { 1501ad2fa371SMuchun Song int zeroed; 1502ad2fa371SMuchun Song int nid = page_to_nid(page); 1503ad2fa371SMuchun Song 1504ad2fa371SMuchun Song VM_BUG_ON_PAGE(!HPageVmemmapOptimized(page), page); 1505ad2fa371SMuchun Song 1506ad2fa371SMuchun Song lockdep_assert_held(&hugetlb_lock); 1507ad2fa371SMuchun Song 1508ad2fa371SMuchun Song INIT_LIST_HEAD(&page->lru); 1509ad2fa371SMuchun Song h->nr_huge_pages++; 1510ad2fa371SMuchun Song h->nr_huge_pages_node[nid]++; 1511ad2fa371SMuchun Song 1512ad2fa371SMuchun Song if (adjust_surplus) { 1513ad2fa371SMuchun Song h->surplus_huge_pages++; 1514ad2fa371SMuchun Song h->surplus_huge_pages_node[nid]++; 1515ad2fa371SMuchun Song } 1516ad2fa371SMuchun Song 1517ad2fa371SMuchun Song set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1518ad2fa371SMuchun Song set_page_private(page, 0); 1519a9e1eab2SMiaohe Lin /* 1520a9e1eab2SMiaohe Lin * We have to set HPageVmemmapOptimized again as above 1521a9e1eab2SMiaohe Lin * set_page_private(page, 0) cleared it. 1522a9e1eab2SMiaohe Lin */ 1523ad2fa371SMuchun Song SetHPageVmemmapOptimized(page); 1524ad2fa371SMuchun Song 1525ad2fa371SMuchun Song /* 1526b65a4edaSMike Kravetz * This page is about to be managed by the hugetlb allocator and 1527b65a4edaSMike Kravetz * should have no users. Drop our reference, and check for others 1528b65a4edaSMike Kravetz * just in case. 1529ad2fa371SMuchun Song */ 1530ad2fa371SMuchun Song zeroed = put_page_testzero(page); 1531b65a4edaSMike Kravetz if (!zeroed) 1532b65a4edaSMike Kravetz /* 1533b65a4edaSMike Kravetz * It is VERY unlikely soneone else has taken a ref on 1534b65a4edaSMike Kravetz * the page. In this case, we simply return as the 1535b65a4edaSMike Kravetz * hugetlb destructor (free_huge_page) will be called 1536b65a4edaSMike Kravetz * when this other ref is dropped. 1537b65a4edaSMike Kravetz */ 1538b65a4edaSMike Kravetz return; 1539b65a4edaSMike Kravetz 1540ad2fa371SMuchun Song arch_clear_hugepage_flags(page); 1541ad2fa371SMuchun Song enqueue_huge_page(h, page); 1542ad2fa371SMuchun Song } 1543ad2fa371SMuchun Song 1544b65d4adbSMuchun Song static void __update_and_free_page(struct hstate *h, struct page *page) 15456af2acb6SAdam Litke { 15466af2acb6SAdam Litke int i; 154714455eabSCheng Li struct page *subpage; 1548a5516438SAndi Kleen 15494eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1550944d9fecSLuiz Capitulino return; 155118229df5SAndy Whitcroft 1552161df60eSNaoya Horiguchi /* 1553161df60eSNaoya Horiguchi * If we don't know which subpages are hwpoisoned, we can't free 1554161df60eSNaoya Horiguchi * the hugepage, so it's leaked intentionally. 1555161df60eSNaoya Horiguchi */ 1556161df60eSNaoya Horiguchi if (HPageRawHwpUnreliable(page)) 1557161df60eSNaoya Horiguchi return; 1558161df60eSNaoya Horiguchi 15596213834cSMuchun Song if (hugetlb_vmemmap_restore(h, page)) { 1560ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 1561ad2fa371SMuchun Song /* 1562ad2fa371SMuchun Song * If we cannot allocate vmemmap pages, just refuse to free the 1563ad2fa371SMuchun Song * page and put the page back on the hugetlb free list and treat 1564ad2fa371SMuchun Song * as a surplus page. 1565ad2fa371SMuchun Song */ 1566ad2fa371SMuchun Song add_hugetlb_page(h, page, true); 1567ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 1568ad2fa371SMuchun Song return; 1569ad2fa371SMuchun Song } 1570ad2fa371SMuchun Song 1571161df60eSNaoya Horiguchi /* 1572161df60eSNaoya Horiguchi * Move PageHWPoison flag from head page to the raw error pages, 1573161df60eSNaoya Horiguchi * which makes any healthy subpages reusable. 1574161df60eSNaoya Horiguchi */ 1575161df60eSNaoya Horiguchi if (unlikely(PageHWPoison(page))) 1576161df60eSNaoya Horiguchi hugetlb_clear_page_hwpoison(page); 1577161df60eSNaoya Horiguchi 157814455eabSCheng Li for (i = 0; i < pages_per_huge_page(h); i++) { 157914455eabSCheng Li subpage = nth_page(page, i); 1580dbfee5aeSMike Kravetz subpage->flags &= ~(1 << PG_locked | 1 << PG_error | 158132f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1582a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1583a7407a27SLuiz Capitulino 1 << PG_writeback); 15846af2acb6SAdam Litke } 1585a01f4390SMike Kravetz 1586a01f4390SMike Kravetz /* 1587a01f4390SMike Kravetz * Non-gigantic pages demoted from CMA allocated gigantic pages 1588a01f4390SMike Kravetz * need to be given back to CMA in free_gigantic_page. 1589a01f4390SMike Kravetz */ 1590a01f4390SMike Kravetz if (hstate_is_gigantic(h) || 1591a01f4390SMike Kravetz hugetlb_cma_page(page, huge_page_order(h))) { 1592944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 1593944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 1594944d9fecSLuiz Capitulino } else { 1595a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 15966af2acb6SAdam Litke } 1597944d9fecSLuiz Capitulino } 15986af2acb6SAdam Litke 1599b65d4adbSMuchun Song /* 1600b65d4adbSMuchun Song * As update_and_free_page() can be called under any context, so we cannot 1601b65d4adbSMuchun Song * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the 1602b65d4adbSMuchun Song * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate 1603b65d4adbSMuchun Song * the vmemmap pages. 1604b65d4adbSMuchun Song * 1605b65d4adbSMuchun Song * free_hpage_workfn() locklessly retrieves the linked list of pages to be 1606b65d4adbSMuchun Song * freed and frees them one-by-one. As the page->mapping pointer is going 1607b65d4adbSMuchun Song * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node 1608b65d4adbSMuchun Song * structure of a lockless linked list of huge pages to be freed. 1609b65d4adbSMuchun Song */ 1610b65d4adbSMuchun Song static LLIST_HEAD(hpage_freelist); 1611b65d4adbSMuchun Song 1612b65d4adbSMuchun Song static void free_hpage_workfn(struct work_struct *work) 1613b65d4adbSMuchun Song { 1614b65d4adbSMuchun Song struct llist_node *node; 1615b65d4adbSMuchun Song 1616b65d4adbSMuchun Song node = llist_del_all(&hpage_freelist); 1617b65d4adbSMuchun Song 1618b65d4adbSMuchun Song while (node) { 1619b65d4adbSMuchun Song struct page *page; 1620b65d4adbSMuchun Song struct hstate *h; 1621b65d4adbSMuchun Song 1622b65d4adbSMuchun Song page = container_of((struct address_space **)node, 1623b65d4adbSMuchun Song struct page, mapping); 1624b65d4adbSMuchun Song node = node->next; 1625b65d4adbSMuchun Song page->mapping = NULL; 1626b65d4adbSMuchun Song /* 1627b65d4adbSMuchun Song * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate() 1628b65d4adbSMuchun Song * is going to trigger because a previous call to 1629b65d4adbSMuchun Song * remove_hugetlb_page() will set_compound_page_dtor(page, 1630b65d4adbSMuchun Song * NULL_COMPOUND_DTOR), so do not use page_hstate() directly. 1631b65d4adbSMuchun Song */ 1632b65d4adbSMuchun Song h = size_to_hstate(page_size(page)); 1633b65d4adbSMuchun Song 1634b65d4adbSMuchun Song __update_and_free_page(h, page); 1635b65d4adbSMuchun Song 1636b65d4adbSMuchun Song cond_resched(); 1637b65d4adbSMuchun Song } 1638b65d4adbSMuchun Song } 1639b65d4adbSMuchun Song static DECLARE_WORK(free_hpage_work, free_hpage_workfn); 1640b65d4adbSMuchun Song 1641b65d4adbSMuchun Song static inline void flush_free_hpage_work(struct hstate *h) 1642b65d4adbSMuchun Song { 16436213834cSMuchun Song if (hugetlb_vmemmap_optimizable(h)) 1644b65d4adbSMuchun Song flush_work(&free_hpage_work); 1645b65d4adbSMuchun Song } 1646b65d4adbSMuchun Song 1647b65d4adbSMuchun Song static void update_and_free_page(struct hstate *h, struct page *page, 1648b65d4adbSMuchun Song bool atomic) 1649b65d4adbSMuchun Song { 1650ad2fa371SMuchun Song if (!HPageVmemmapOptimized(page) || !atomic) { 1651b65d4adbSMuchun Song __update_and_free_page(h, page); 1652b65d4adbSMuchun Song return; 1653b65d4adbSMuchun Song } 1654b65d4adbSMuchun Song 1655b65d4adbSMuchun Song /* 1656b65d4adbSMuchun Song * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages. 1657b65d4adbSMuchun Song * 1658b65d4adbSMuchun Song * Only call schedule_work() if hpage_freelist is previously 1659b65d4adbSMuchun Song * empty. Otherwise, schedule_work() had been called but the workfn 1660b65d4adbSMuchun Song * hasn't retrieved the list yet. 1661b65d4adbSMuchun Song */ 1662b65d4adbSMuchun Song if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist)) 1663b65d4adbSMuchun Song schedule_work(&free_hpage_work); 1664b65d4adbSMuchun Song } 1665b65d4adbSMuchun Song 166610c6ec49SMike Kravetz static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list) 166710c6ec49SMike Kravetz { 166810c6ec49SMike Kravetz struct page *page, *t_page; 166910c6ec49SMike Kravetz 167010c6ec49SMike Kravetz list_for_each_entry_safe(page, t_page, list, lru) { 1671b65d4adbSMuchun Song update_and_free_page(h, page, false); 167210c6ec49SMike Kravetz cond_resched(); 167310c6ec49SMike Kravetz } 167410c6ec49SMike Kravetz } 167510c6ec49SMike Kravetz 1676e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1677e5ff2159SAndi Kleen { 1678e5ff2159SAndi Kleen struct hstate *h; 1679e5ff2159SAndi Kleen 1680e5ff2159SAndi Kleen for_each_hstate(h) { 1681e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1682e5ff2159SAndi Kleen return h; 1683e5ff2159SAndi Kleen } 1684e5ff2159SAndi Kleen return NULL; 1685e5ff2159SAndi Kleen } 1686e5ff2159SAndi Kleen 1687db71ef79SMike Kravetz void free_huge_page(struct page *page) 168827a85ef1SDavid Gibson { 1689a5516438SAndi Kleen /* 1690a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1691a5516438SAndi Kleen * compound page destructor. 1692a5516438SAndi Kleen */ 1693e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 16947893d1d5SAdam Litke int nid = page_to_nid(page); 1695d6995da3SMike Kravetz struct hugepage_subpool *spool = hugetlb_page_subpool(page); 169607443a85SJoonsoo Kim bool restore_reserve; 1697db71ef79SMike Kravetz unsigned long flags; 169827a85ef1SDavid Gibson 1699b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1700b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_mapcount(page), page); 17018ace22bcSYongkai Wu 1702d6995da3SMike Kravetz hugetlb_set_page_subpool(page, NULL); 170378fbe906SDavid Hildenbrand if (PageAnon(page)) 170478fbe906SDavid Hildenbrand __ClearPageAnonExclusive(page); 17058ace22bcSYongkai Wu page->mapping = NULL; 1706d6995da3SMike Kravetz restore_reserve = HPageRestoreReserve(page); 1707d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 170827a85ef1SDavid Gibson 17091c5ecae3SMike Kravetz /* 1710d6995da3SMike Kravetz * If HPageRestoreReserve was set on page, page allocation consumed a 17110919e1b6SMike Kravetz * reservation. If the page was associated with a subpool, there 17120919e1b6SMike Kravetz * would have been a page reserved in the subpool before allocation 17130919e1b6SMike Kravetz * via hugepage_subpool_get_pages(). Since we are 'restoring' the 17146c26d310SMiaohe Lin * reservation, do not call hugepage_subpool_put_pages() as this will 17150919e1b6SMike Kravetz * remove the reserved page from the subpool. 17160919e1b6SMike Kravetz */ 17170919e1b6SMike Kravetz if (!restore_reserve) { 17180919e1b6SMike Kravetz /* 17190919e1b6SMike Kravetz * A return code of zero implies that the subpool will be 17200919e1b6SMike Kravetz * under its minimum size if the reservation is not restored 17210919e1b6SMike Kravetz * after page is free. Therefore, force restore_reserve 17220919e1b6SMike Kravetz * operation. 17231c5ecae3SMike Kravetz */ 17241c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 17251c5ecae3SMike Kravetz restore_reserve = true; 17260919e1b6SMike Kravetz } 17271c5ecae3SMike Kravetz 1728db71ef79SMike Kravetz spin_lock_irqsave(&hugetlb_lock, flags); 17298f251a3dSMike Kravetz ClearHPageMigratable(page); 17306d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 17316d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 173208cf9fafSMina Almasry hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 173308cf9fafSMina Almasry pages_per_huge_page(h), page); 173407443a85SJoonsoo Kim if (restore_reserve) 173507443a85SJoonsoo Kim h->resv_huge_pages++; 173607443a85SJoonsoo Kim 17379157c311SMike Kravetz if (HPageTemporary(page)) { 17386eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 1739db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1740b65d4adbSMuchun Song update_and_free_page(h, page, true); 1741ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 17420edaecfaSAneesh Kumar K.V /* remove the page from active list */ 17436eb4e88aSMike Kravetz remove_hugetlb_page(h, page, true); 1744db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1745b65d4adbSMuchun Song update_and_free_page(h, page, true); 17467893d1d5SAdam Litke } else { 17475d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1748a5516438SAndi Kleen enqueue_huge_page(h, page); 1749db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 175027a85ef1SDavid Gibson } 17511121828aSMike Kravetz } 175227a85ef1SDavid Gibson 1753d3d99fccSOscar Salvador /* 1754d3d99fccSOscar Salvador * Must be called with the hugetlb lock held 1755d3d99fccSOscar Salvador */ 1756d3d99fccSOscar Salvador static void __prep_account_new_huge_page(struct hstate *h, int nid) 1757d3d99fccSOscar Salvador { 1758d3d99fccSOscar Salvador lockdep_assert_held(&hugetlb_lock); 1759d3d99fccSOscar Salvador h->nr_huge_pages++; 1760d3d99fccSOscar Salvador h->nr_huge_pages_node[nid]++; 1761d3d99fccSOscar Salvador } 1762d3d99fccSOscar Salvador 1763f41f2ed4SMuchun Song static void __prep_new_huge_page(struct hstate *h, struct page *page) 1764b7ba30c6SAndi Kleen { 17656213834cSMuchun Song hugetlb_vmemmap_optimize(h, page); 17660edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1767f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1768ff546117SMike Kravetz hugetlb_set_page_subpool(page, NULL); 17699dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 17701adc4d41SMina Almasry set_hugetlb_cgroup_rsvd(page, NULL); 1771d3d99fccSOscar Salvador } 1772d3d99fccSOscar Salvador 1773d3d99fccSOscar Salvador static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1774d3d99fccSOscar Salvador { 1775f41f2ed4SMuchun Song __prep_new_huge_page(h, page); 1776db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 1777d3d99fccSOscar Salvador __prep_account_new_huge_page(h, nid); 1778db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1779b7ba30c6SAndi Kleen } 1780b7ba30c6SAndi Kleen 178134d9e35bSMike Kravetz static bool __prep_compound_gigantic_page(struct page *page, unsigned int order, 178234d9e35bSMike Kravetz bool demote) 178320a0307cSWu Fengguang { 17847118fc29SMike Kravetz int i, j; 178520a0307cSWu Fengguang int nr_pages = 1 << order; 178614455eabSCheng Li struct page *p; 178720a0307cSWu Fengguang 178820a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 178920a0307cSWu Fengguang set_compound_order(page, order); 1790ef5a22beSAndrea Arcangeli __ClearPageReserved(page); 1791de09d31dSKirill A. Shutemov __SetPageHead(page); 179214455eabSCheng Li for (i = 1; i < nr_pages; i++) { 179314455eabSCheng Li p = nth_page(page, i); 179414455eabSCheng Li 1795ef5a22beSAndrea Arcangeli /* 1796ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1797ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1798ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 17997c8de358SEthon Paul * too. Otherwise drivers using get_user_pages() to access tail 1800ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1801ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1802ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1803ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1804ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1805ef5a22beSAndrea Arcangeli * after get_user_pages(). 1806ef5a22beSAndrea Arcangeli */ 1807ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 18087118fc29SMike Kravetz /* 18097118fc29SMike Kravetz * Subtle and very unlikely 18107118fc29SMike Kravetz * 18117118fc29SMike Kravetz * Gigantic 'page allocators' such as memblock or cma will 18127118fc29SMike Kravetz * return a set of pages with each page ref counted. We need 18137118fc29SMike Kravetz * to turn this set of pages into a compound page with tail 18147118fc29SMike Kravetz * page ref counts set to zero. Code such as speculative page 18157118fc29SMike Kravetz * cache adding could take a ref on a 'to be' tail page. 18167118fc29SMike Kravetz * We need to respect any increased ref count, and only set 18177118fc29SMike Kravetz * the ref count to zero if count is currently 1. If count 1818416d85edSMike Kravetz * is not 1, we return an error. An error return indicates 1819416d85edSMike Kravetz * the set of pages can not be converted to a gigantic page. 1820416d85edSMike Kravetz * The caller who allocated the pages should then discard the 1821416d85edSMike Kravetz * pages using the appropriate free interface. 182234d9e35bSMike Kravetz * 182334d9e35bSMike Kravetz * In the case of demote, the ref count will be zero. 18247118fc29SMike Kravetz */ 182534d9e35bSMike Kravetz if (!demote) { 18267118fc29SMike Kravetz if (!page_ref_freeze(p, 1)) { 1827416d85edSMike Kravetz pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n"); 18287118fc29SMike Kravetz goto out_error; 18297118fc29SMike Kravetz } 183034d9e35bSMike Kravetz } else { 183134d9e35bSMike Kravetz VM_BUG_ON_PAGE(page_count(p), p); 183234d9e35bSMike Kravetz } 18331d798ca3SKirill A. Shutemov set_compound_head(p, page); 183420a0307cSWu Fengguang } 1835b4330afbSMike Kravetz atomic_set(compound_mapcount_ptr(page), -1); 183647e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 18377118fc29SMike Kravetz return true; 18387118fc29SMike Kravetz 18397118fc29SMike Kravetz out_error: 18407118fc29SMike Kravetz /* undo tail page modifications made above */ 184114455eabSCheng Li for (j = 1; j < i; j++) { 184214455eabSCheng Li p = nth_page(page, j); 18437118fc29SMike Kravetz clear_compound_head(p); 18447118fc29SMike Kravetz set_page_refcounted(p); 18457118fc29SMike Kravetz } 18467118fc29SMike Kravetz /* need to clear PG_reserved on remaining tail pages */ 184714455eabSCheng Li for (; j < nr_pages; j++) { 184814455eabSCheng Li p = nth_page(page, j); 18497118fc29SMike Kravetz __ClearPageReserved(p); 185014455eabSCheng Li } 18517118fc29SMike Kravetz set_compound_order(page, 0); 18525232c63fSMatthew Wilcox (Oracle) #ifdef CONFIG_64BIT 18537118fc29SMike Kravetz page[1].compound_nr = 0; 18545232c63fSMatthew Wilcox (Oracle) #endif 18557118fc29SMike Kravetz __ClearPageHead(page); 18567118fc29SMike Kravetz return false; 185720a0307cSWu Fengguang } 185820a0307cSWu Fengguang 185934d9e35bSMike Kravetz static bool prep_compound_gigantic_page(struct page *page, unsigned int order) 186034d9e35bSMike Kravetz { 186134d9e35bSMike Kravetz return __prep_compound_gigantic_page(page, order, false); 186234d9e35bSMike Kravetz } 186334d9e35bSMike Kravetz 18648531fc6fSMike Kravetz static bool prep_compound_gigantic_page_for_demote(struct page *page, 18658531fc6fSMike Kravetz unsigned int order) 18668531fc6fSMike Kravetz { 18678531fc6fSMike Kravetz return __prep_compound_gigantic_page(page, order, true); 18688531fc6fSMike Kravetz } 18698531fc6fSMike Kravetz 18707795912cSAndrew Morton /* 18717795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 18727795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 18737795912cSAndrew Morton * details. 18747795912cSAndrew Morton */ 187520a0307cSWu Fengguang int PageHuge(struct page *page) 187620a0307cSWu Fengguang { 187720a0307cSWu Fengguang if (!PageCompound(page)) 187820a0307cSWu Fengguang return 0; 187920a0307cSWu Fengguang 188020a0307cSWu Fengguang page = compound_head(page); 1881f1e61557SKirill A. Shutemov return page[1].compound_dtor == HUGETLB_PAGE_DTOR; 188220a0307cSWu Fengguang } 188343131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 188443131e14SNaoya Horiguchi 188527c73ae7SAndrea Arcangeli /* 188627c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 188727c73ae7SAndrea Arcangeli * normal or transparent huge pages. 188827c73ae7SAndrea Arcangeli */ 188927c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 189027c73ae7SAndrea Arcangeli { 189127c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 189227c73ae7SAndrea Arcangeli return 0; 189327c73ae7SAndrea Arcangeli 1894d4af73e3SVlastimil Babka return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; 189527c73ae7SAndrea Arcangeli } 18964e936eccSDavid Howells EXPORT_SYMBOL_GPL(PageHeadHuge); 189727c73ae7SAndrea Arcangeli 1898c0d0381aSMike Kravetz /* 1899c0d0381aSMike Kravetz * Find and lock address space (mapping) in write mode. 1900c0d0381aSMike Kravetz * 1901336bf30eSMike Kravetz * Upon entry, the page is locked which means that page_mapping() is 1902336bf30eSMike Kravetz * stable. Due to locking order, we can only trylock_write. If we can 1903336bf30eSMike Kravetz * not get the lock, simply return NULL to caller. 1904c0d0381aSMike Kravetz */ 1905c0d0381aSMike Kravetz struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) 1906c0d0381aSMike Kravetz { 1907336bf30eSMike Kravetz struct address_space *mapping = page_mapping(hpage); 1908c0d0381aSMike Kravetz 1909c0d0381aSMike Kravetz if (!mapping) 1910c0d0381aSMike Kravetz return mapping; 1911c0d0381aSMike Kravetz 1912c0d0381aSMike Kravetz if (i_mmap_trylock_write(mapping)) 1913c0d0381aSMike Kravetz return mapping; 1914c0d0381aSMike Kravetz 1915c0d0381aSMike Kravetz return NULL; 1916c0d0381aSMike Kravetz } 1917c0d0381aSMike Kravetz 1918fe19bd3dSHugh Dickins pgoff_t hugetlb_basepage_index(struct page *page) 191913d60f4bSZhang Yi { 192013d60f4bSZhang Yi struct page *page_head = compound_head(page); 192113d60f4bSZhang Yi pgoff_t index = page_index(page_head); 192213d60f4bSZhang Yi unsigned long compound_idx; 192313d60f4bSZhang Yi 192413d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 192513d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 192613d60f4bSZhang Yi else 192713d60f4bSZhang Yi compound_idx = page - page_head; 192813d60f4bSZhang Yi 192913d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 193013d60f4bSZhang Yi } 193113d60f4bSZhang Yi 19320c397daeSMichal Hocko static struct page *alloc_buddy_huge_page(struct hstate *h, 1933f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1934f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 19351da177e4SLinus Torvalds { 1936af0fb9dfSMichal Hocko int order = huge_page_order(h); 19371da177e4SLinus Torvalds struct page *page; 1938f60858f9SMike Kravetz bool alloc_try_hard = true; 1939f96efd58SJoe Jin 1940f60858f9SMike Kravetz /* 1941f60858f9SMike Kravetz * By default we always try hard to allocate the page with 1942f60858f9SMike Kravetz * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in 1943f60858f9SMike Kravetz * a loop (to adjust global huge page counts) and previous allocation 1944f60858f9SMike Kravetz * failed, do not continue to try hard on the same node. Use the 1945f60858f9SMike Kravetz * node_alloc_noretry bitmap to manage this state information. 1946f60858f9SMike Kravetz */ 1947f60858f9SMike Kravetz if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) 1948f60858f9SMike Kravetz alloc_try_hard = false; 1949f60858f9SMike Kravetz gfp_mask |= __GFP_COMP|__GFP_NOWARN; 1950f60858f9SMike Kravetz if (alloc_try_hard) 1951f60858f9SMike Kravetz gfp_mask |= __GFP_RETRY_MAYFAIL; 1952af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 1953af0fb9dfSMichal Hocko nid = numa_mem_id(); 195484172f4bSMatthew Wilcox (Oracle) page = __alloc_pages(gfp_mask, order, nid, nmask); 1955af0fb9dfSMichal Hocko if (page) 1956af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC); 1957af0fb9dfSMichal Hocko else 1958af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 195963b4613cSNishanth Aravamudan 1960f60858f9SMike Kravetz /* 1961f60858f9SMike Kravetz * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this 1962f60858f9SMike Kravetz * indicates an overall state change. Clear bit so that we resume 1963f60858f9SMike Kravetz * normal 'try hard' allocations. 1964f60858f9SMike Kravetz */ 1965f60858f9SMike Kravetz if (node_alloc_noretry && page && !alloc_try_hard) 1966f60858f9SMike Kravetz node_clear(nid, *node_alloc_noretry); 1967f60858f9SMike Kravetz 1968f60858f9SMike Kravetz /* 1969f60858f9SMike Kravetz * If we tried hard to get a page but failed, set bit so that 1970f60858f9SMike Kravetz * subsequent attempts will not try as hard until there is an 1971f60858f9SMike Kravetz * overall state change. 1972f60858f9SMike Kravetz */ 1973f60858f9SMike Kravetz if (node_alloc_noretry && !page && alloc_try_hard) 1974f60858f9SMike Kravetz node_set(nid, *node_alloc_noretry); 1975f60858f9SMike Kravetz 197663b4613cSNishanth Aravamudan return page; 197763b4613cSNishanth Aravamudan } 197863b4613cSNishanth Aravamudan 1979af0fb9dfSMichal Hocko /* 19800c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 19810c397daeSMichal Hocko * should use this function to get new hugetlb pages 19820c397daeSMichal Hocko */ 19830c397daeSMichal Hocko static struct page *alloc_fresh_huge_page(struct hstate *h, 1984f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1985f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 19860c397daeSMichal Hocko { 19870c397daeSMichal Hocko struct page *page; 19887118fc29SMike Kravetz bool retry = false; 19890c397daeSMichal Hocko 19907118fc29SMike Kravetz retry: 19910c397daeSMichal Hocko if (hstate_is_gigantic(h)) 19920c397daeSMichal Hocko page = alloc_gigantic_page(h, gfp_mask, nid, nmask); 19930c397daeSMichal Hocko else 19940c397daeSMichal Hocko page = alloc_buddy_huge_page(h, gfp_mask, 1995f60858f9SMike Kravetz nid, nmask, node_alloc_noretry); 19960c397daeSMichal Hocko if (!page) 19970c397daeSMichal Hocko return NULL; 19980c397daeSMichal Hocko 19997118fc29SMike Kravetz if (hstate_is_gigantic(h)) { 20007118fc29SMike Kravetz if (!prep_compound_gigantic_page(page, huge_page_order(h))) { 20017118fc29SMike Kravetz /* 20027118fc29SMike Kravetz * Rare failure to convert pages to compound page. 20037118fc29SMike Kravetz * Free pages and try again - ONCE! 20047118fc29SMike Kravetz */ 20057118fc29SMike Kravetz free_gigantic_page(page, huge_page_order(h)); 20067118fc29SMike Kravetz if (!retry) { 20077118fc29SMike Kravetz retry = true; 20087118fc29SMike Kravetz goto retry; 20097118fc29SMike Kravetz } 20107118fc29SMike Kravetz return NULL; 20117118fc29SMike Kravetz } 20127118fc29SMike Kravetz } 20130c397daeSMichal Hocko prep_new_huge_page(h, page, page_to_nid(page)); 20140c397daeSMichal Hocko 20150c397daeSMichal Hocko return page; 20160c397daeSMichal Hocko } 20170c397daeSMichal Hocko 20180c397daeSMichal Hocko /* 2019af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 2020af0fb9dfSMichal Hocko * manner. 2021af0fb9dfSMichal Hocko */ 2022f60858f9SMike Kravetz static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 2023f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 2024b2261026SJoonsoo Kim { 2025b2261026SJoonsoo Kim struct page *page; 2026b2261026SJoonsoo Kim int nr_nodes, node; 2027af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2028b2261026SJoonsoo Kim 2029b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 2030f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, 2031f60858f9SMike Kravetz node_alloc_noretry); 2032af0fb9dfSMichal Hocko if (page) 2033b2261026SJoonsoo Kim break; 2034b2261026SJoonsoo Kim } 2035b2261026SJoonsoo Kim 2036af0fb9dfSMichal Hocko if (!page) 2037af0fb9dfSMichal Hocko return 0; 2038b2261026SJoonsoo Kim 2039af0fb9dfSMichal Hocko put_page(page); /* free it into the hugepage allocator */ 2040af0fb9dfSMichal Hocko 2041af0fb9dfSMichal Hocko return 1; 2042b2261026SJoonsoo Kim } 2043b2261026SJoonsoo Kim 2044e8c5c824SLee Schermerhorn /* 204510c6ec49SMike Kravetz * Remove huge page from pool from next node to free. Attempt to keep 204610c6ec49SMike Kravetz * persistent huge pages more or less balanced over allowed nodes. 204710c6ec49SMike Kravetz * This routine only 'removes' the hugetlb page. The caller must make 204810c6ec49SMike Kravetz * an additional call to free the page to low level allocators. 2049e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 2050e8c5c824SLee Schermerhorn */ 205110c6ec49SMike Kravetz static struct page *remove_pool_huge_page(struct hstate *h, 205210c6ec49SMike Kravetz nodemask_t *nodes_allowed, 20536ae11b27SLee Schermerhorn bool acct_surplus) 2054e8c5c824SLee Schermerhorn { 2055b2261026SJoonsoo Kim int nr_nodes, node; 205610c6ec49SMike Kravetz struct page *page = NULL; 2057e8c5c824SLee Schermerhorn 20589487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2059b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2060685f3457SLee Schermerhorn /* 2061685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 2062685f3457SLee Schermerhorn * nodes with surplus pages. 2063685f3457SLee Schermerhorn */ 2064b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 2065b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 206610c6ec49SMike Kravetz page = list_entry(h->hugepage_freelists[node].next, 2067e8c5c824SLee Schermerhorn struct page, lru); 20686eb4e88aSMike Kravetz remove_hugetlb_page(h, page, acct_surplus); 20699a76db09SLee Schermerhorn break; 2070e8c5c824SLee Schermerhorn } 2071b2261026SJoonsoo Kim } 2072e8c5c824SLee Schermerhorn 207310c6ec49SMike Kravetz return page; 2074e8c5c824SLee Schermerhorn } 2075e8c5c824SLee Schermerhorn 2076c8721bbbSNaoya Horiguchi /* 2077c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 2078faf53defSNaoya Horiguchi * nothing for in-use hugepages and non-hugepages. 2079faf53defSNaoya Horiguchi * This function returns values like below: 2080faf53defSNaoya Horiguchi * 2081ad2fa371SMuchun Song * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages 2082ad2fa371SMuchun Song * when the system is under memory pressure and the feature of 2083ad2fa371SMuchun Song * freeing unused vmemmap pages associated with each hugetlb page 2084ad2fa371SMuchun Song * is enabled. 2085faf53defSNaoya Horiguchi * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 2086faf53defSNaoya Horiguchi * (allocated or reserved.) 2087faf53defSNaoya Horiguchi * 0: successfully dissolved free hugepages or the page is not a 2088faf53defSNaoya Horiguchi * hugepage (considered as already dissolved) 2089c8721bbbSNaoya Horiguchi */ 2090c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 2091c8721bbbSNaoya Horiguchi { 20926bc9b564SNaoya Horiguchi int rc = -EBUSY; 2093082d5b6bSGerald Schaefer 20947ffddd49SMuchun Song retry: 2095faf53defSNaoya Horiguchi /* Not to disrupt normal path by vainly holding hugetlb_lock */ 2096faf53defSNaoya Horiguchi if (!PageHuge(page)) 2097faf53defSNaoya Horiguchi return 0; 2098faf53defSNaoya Horiguchi 2099db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2100faf53defSNaoya Horiguchi if (!PageHuge(page)) { 2101faf53defSNaoya Horiguchi rc = 0; 2102faf53defSNaoya Horiguchi goto out; 2103faf53defSNaoya Horiguchi } 2104faf53defSNaoya Horiguchi 2105faf53defSNaoya Horiguchi if (!page_count(page)) { 21062247bb33SGerald Schaefer struct page *head = compound_head(page); 21072247bb33SGerald Schaefer struct hstate *h = page_hstate(head); 21086bc9b564SNaoya Horiguchi if (h->free_huge_pages - h->resv_huge_pages == 0) 2109082d5b6bSGerald Schaefer goto out; 21107ffddd49SMuchun Song 21117ffddd49SMuchun Song /* 21127ffddd49SMuchun Song * We should make sure that the page is already on the free list 21137ffddd49SMuchun Song * when it is dissolved. 21147ffddd49SMuchun Song */ 21156c037149SMike Kravetz if (unlikely(!HPageFreed(head))) { 2116db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 21177ffddd49SMuchun Song cond_resched(); 21187ffddd49SMuchun Song 21197ffddd49SMuchun Song /* 21207ffddd49SMuchun Song * Theoretically, we should return -EBUSY when we 21217ffddd49SMuchun Song * encounter this race. In fact, we have a chance 21227ffddd49SMuchun Song * to successfully dissolve the page if we do a 21237ffddd49SMuchun Song * retry. Because the race window is quite small. 21247ffddd49SMuchun Song * If we seize this opportunity, it is an optimization 21257ffddd49SMuchun Song * for increasing the success rate of dissolving page. 21267ffddd49SMuchun Song */ 21277ffddd49SMuchun Song goto retry; 21287ffddd49SMuchun Song } 21297ffddd49SMuchun Song 2130ad2fa371SMuchun Song remove_hugetlb_page(h, head, false); 2131ad2fa371SMuchun Song h->max_huge_pages--; 2132ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2133ad2fa371SMuchun Song 2134c3114a84SAnshuman Khandual /* 2135ad2fa371SMuchun Song * Normally update_and_free_page will allocate required vmemmmap 2136ad2fa371SMuchun Song * before freeing the page. update_and_free_page will fail to 2137ad2fa371SMuchun Song * free the page if it can not allocate required vmemmap. We 2138ad2fa371SMuchun Song * need to adjust max_huge_pages if the page is not freed. 2139ad2fa371SMuchun Song * Attempt to allocate vmemmmap here so that we can take 2140ad2fa371SMuchun Song * appropriate action on failure. 2141ad2fa371SMuchun Song */ 21426213834cSMuchun Song rc = hugetlb_vmemmap_restore(h, head); 2143ad2fa371SMuchun Song if (!rc) { 2144b65d4adbSMuchun Song update_and_free_page(h, head, false); 2145ad2fa371SMuchun Song } else { 2146ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 2147ad2fa371SMuchun Song add_hugetlb_page(h, head, false); 2148ad2fa371SMuchun Song h->max_huge_pages++; 2149ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2150ad2fa371SMuchun Song } 2151ad2fa371SMuchun Song 2152ad2fa371SMuchun Song return rc; 2153c8721bbbSNaoya Horiguchi } 2154082d5b6bSGerald Schaefer out: 2155db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2156082d5b6bSGerald Schaefer return rc; 2157c8721bbbSNaoya Horiguchi } 2158c8721bbbSNaoya Horiguchi 2159c8721bbbSNaoya Horiguchi /* 2160c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 2161c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 21622247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 21632247bb33SGerald Schaefer * part of it lies within the given range. 2164082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 2165082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 2166c8721bbbSNaoya Horiguchi */ 2167082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 2168c8721bbbSNaoya Horiguchi { 2169c8721bbbSNaoya Horiguchi unsigned long pfn; 2170eb03aa00SGerald Schaefer struct page *page; 2171082d5b6bSGerald Schaefer int rc = 0; 2172dc2628f3SMuchun Song unsigned int order; 2173dc2628f3SMuchun Song struct hstate *h; 2174c8721bbbSNaoya Horiguchi 2175d0177639SLi Zhong if (!hugepages_supported()) 2176082d5b6bSGerald Schaefer return rc; 2177d0177639SLi Zhong 2178dc2628f3SMuchun Song order = huge_page_order(&default_hstate); 2179dc2628f3SMuchun Song for_each_hstate(h) 2180dc2628f3SMuchun Song order = min(order, huge_page_order(h)); 2181dc2628f3SMuchun Song 2182dc2628f3SMuchun Song for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) { 2183eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 2184eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 2185eb03aa00SGerald Schaefer if (rc) 2186082d5b6bSGerald Schaefer break; 2187eb03aa00SGerald Schaefer } 2188082d5b6bSGerald Schaefer 2189082d5b6bSGerald Schaefer return rc; 2190c8721bbbSNaoya Horiguchi } 2191c8721bbbSNaoya Horiguchi 2192ab5ac90aSMichal Hocko /* 2193ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 2194ab5ac90aSMichal Hocko */ 21950c397daeSMichal Hocko static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, 2196b65a4edaSMike Kravetz int nid, nodemask_t *nmask, bool zero_ref) 21977893d1d5SAdam Litke { 21989980d744SMichal Hocko struct page *page = NULL; 2199b65a4edaSMike Kravetz bool retry = false; 22007893d1d5SAdam Litke 2201bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2202aa888a74SAndi Kleen return NULL; 2203aa888a74SAndi Kleen 2204db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 22059980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 22069980d744SMichal Hocko goto out_unlock; 2207db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2208d1c3fb1fSNishanth Aravamudan 2209b65a4edaSMike Kravetz retry: 2210f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 22119980d744SMichal Hocko if (!page) 22120c397daeSMichal Hocko return NULL; 2213d1c3fb1fSNishanth Aravamudan 2214db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 22159980d744SMichal Hocko /* 22169980d744SMichal Hocko * We could have raced with the pool size change. 22179980d744SMichal Hocko * Double check that and simply deallocate the new page 22189980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 22199980d744SMichal Hocko * temporary page to workaround the nasty free_huge_page 22209980d744SMichal Hocko * codeflow 22219980d744SMichal Hocko */ 22229980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 22239157c311SMike Kravetz SetHPageTemporary(page); 2224db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 22259980d744SMichal Hocko put_page(page); 22262bf753e6SKai Shen return NULL; 2227b65a4edaSMike Kravetz } 2228b65a4edaSMike Kravetz 2229b65a4edaSMike Kravetz if (zero_ref) { 2230b65a4edaSMike Kravetz /* 2231b65a4edaSMike Kravetz * Caller requires a page with zero ref count. 2232b65a4edaSMike Kravetz * We will drop ref count here. If someone else is holding 2233b65a4edaSMike Kravetz * a ref, the page will be freed when they drop it. Abuse 2234b65a4edaSMike Kravetz * temporary page flag to accomplish this. 2235b65a4edaSMike Kravetz */ 2236b65a4edaSMike Kravetz SetHPageTemporary(page); 2237b65a4edaSMike Kravetz if (!put_page_testzero(page)) { 2238b65a4edaSMike Kravetz /* 2239b65a4edaSMike Kravetz * Unexpected inflated ref count on freshly allocated 2240b65a4edaSMike Kravetz * huge. Retry once. 2241b65a4edaSMike Kravetz */ 2242b65a4edaSMike Kravetz pr_info("HugeTLB unexpected inflated ref count on freshly allocated page\n"); 2243b65a4edaSMike Kravetz spin_unlock_irq(&hugetlb_lock); 2244b65a4edaSMike Kravetz if (retry) 2245b65a4edaSMike Kravetz return NULL; 2246b65a4edaSMike Kravetz 2247b65a4edaSMike Kravetz retry = true; 2248b65a4edaSMike Kravetz goto retry; 2249b65a4edaSMike Kravetz } 2250b65a4edaSMike Kravetz ClearHPageTemporary(page); 2251b65a4edaSMike Kravetz } 2252b65a4edaSMike Kravetz 22539980d744SMichal Hocko h->surplus_huge_pages++; 22544704dea3SMichal Hocko h->surplus_huge_pages_node[page_to_nid(page)]++; 22559980d744SMichal Hocko 22569980d744SMichal Hocko out_unlock: 2257db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 22587893d1d5SAdam Litke 22597893d1d5SAdam Litke return page; 22607893d1d5SAdam Litke } 22617893d1d5SAdam Litke 2262bbe88753SJoonsoo Kim static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, 2263ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 2264ab5ac90aSMichal Hocko { 2265ab5ac90aSMichal Hocko struct page *page; 2266ab5ac90aSMichal Hocko 2267ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 2268ab5ac90aSMichal Hocko return NULL; 2269ab5ac90aSMichal Hocko 2270f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 2271ab5ac90aSMichal Hocko if (!page) 2272ab5ac90aSMichal Hocko return NULL; 2273ab5ac90aSMichal Hocko 2274ab5ac90aSMichal Hocko /* 2275ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 2276ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 2277ab5ac90aSMichal Hocko */ 22789157c311SMike Kravetz SetHPageTemporary(page); 2279ab5ac90aSMichal Hocko 2280ab5ac90aSMichal Hocko return page; 2281ab5ac90aSMichal Hocko } 2282ab5ac90aSMichal Hocko 2283e4e574b7SAdam Litke /* 2284099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 2285099730d6SDave Hansen */ 2286e0ec90eeSDave Hansen static 22870c397daeSMichal Hocko struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, 2288099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 2289099730d6SDave Hansen { 2290cfcaa66fSBen Widawsky struct page *page = NULL; 2291aaf14e40SMichal Hocko struct mempolicy *mpol; 2292aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 2293aaf14e40SMichal Hocko int nid; 2294aaf14e40SMichal Hocko nodemask_t *nodemask; 2295aaf14e40SMichal Hocko 2296aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 2297cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 2298cfcaa66fSBen Widawsky gfp_t gfp = gfp_mask | __GFP_NOWARN; 2299cfcaa66fSBen Widawsky 2300cfcaa66fSBen Widawsky gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL); 2301cfcaa66fSBen Widawsky page = alloc_surplus_huge_page(h, gfp, nid, nodemask, false); 2302cfcaa66fSBen Widawsky 2303cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 2304cfcaa66fSBen Widawsky nodemask = NULL; 2305cfcaa66fSBen Widawsky } 2306cfcaa66fSBen Widawsky 2307cfcaa66fSBen Widawsky if (!page) 2308b65a4edaSMike Kravetz page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask, false); 2309aaf14e40SMichal Hocko mpol_cond_put(mpol); 2310aaf14e40SMichal Hocko return page; 2311099730d6SDave Hansen } 2312099730d6SDave Hansen 2313ab5ac90aSMichal Hocko /* page migration callback function */ 23143e59fcb0SMichal Hocko struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, 2315d92bbc27SJoonsoo Kim nodemask_t *nmask, gfp_t gfp_mask) 23164db9b2efSMichal Hocko { 2317db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 23184db9b2efSMichal Hocko if (h->free_huge_pages - h->resv_huge_pages > 0) { 23193e59fcb0SMichal Hocko struct page *page; 23203e59fcb0SMichal Hocko 23213e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); 23223e59fcb0SMichal Hocko if (page) { 2323db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 23243e59fcb0SMichal Hocko return page; 23254db9b2efSMichal Hocko } 23264db9b2efSMichal Hocko } 2327db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 23284db9b2efSMichal Hocko 23290c397daeSMichal Hocko return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); 23304db9b2efSMichal Hocko } 23314db9b2efSMichal Hocko 2332ebd63723SMichal Hocko /* mempolicy aware migration callback */ 2333389c8178SMichal Hocko struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, 2334389c8178SMichal Hocko unsigned long address) 2335ebd63723SMichal Hocko { 2336ebd63723SMichal Hocko struct mempolicy *mpol; 2337ebd63723SMichal Hocko nodemask_t *nodemask; 2338ebd63723SMichal Hocko struct page *page; 2339ebd63723SMichal Hocko gfp_t gfp_mask; 2340ebd63723SMichal Hocko int node; 2341ebd63723SMichal Hocko 2342ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 2343ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 2344d92bbc27SJoonsoo Kim page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); 2345ebd63723SMichal Hocko mpol_cond_put(mpol); 2346ebd63723SMichal Hocko 2347ebd63723SMichal Hocko return page; 2348ebd63723SMichal Hocko } 2349ebd63723SMichal Hocko 2350bf50bab2SNaoya Horiguchi /* 235125985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 2352e4e574b7SAdam Litke * of size 'delta'. 2353e4e574b7SAdam Litke */ 23540a4f3d1bSLiu Xiang static int gather_surplus_pages(struct hstate *h, long delta) 23551b2a1e7bSJules Irenge __must_hold(&hugetlb_lock) 2356e4e574b7SAdam Litke { 235734665341SMiaohe Lin LIST_HEAD(surplus_list); 2358e4e574b7SAdam Litke struct page *page, *tmp; 23590a4f3d1bSLiu Xiang int ret; 23600a4f3d1bSLiu Xiang long i; 23610a4f3d1bSLiu Xiang long needed, allocated; 236228073b02SHillf Danton bool alloc_ok = true; 2363e4e574b7SAdam Litke 23649487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2365a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 2366ac09b3a1SAdam Litke if (needed <= 0) { 2367a5516438SAndi Kleen h->resv_huge_pages += delta; 2368e4e574b7SAdam Litke return 0; 2369ac09b3a1SAdam Litke } 2370e4e574b7SAdam Litke 2371e4e574b7SAdam Litke allocated = 0; 2372e4e574b7SAdam Litke 2373e4e574b7SAdam Litke ret = -ENOMEM; 2374e4e574b7SAdam Litke retry: 2375db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2376e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 23770c397daeSMichal Hocko page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), 2378b65a4edaSMike Kravetz NUMA_NO_NODE, NULL, true); 237928073b02SHillf Danton if (!page) { 238028073b02SHillf Danton alloc_ok = false; 238128073b02SHillf Danton break; 238228073b02SHillf Danton } 2383e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 238469ed779aSDavid Rientjes cond_resched(); 2385e4e574b7SAdam Litke } 238628073b02SHillf Danton allocated += i; 2387e4e574b7SAdam Litke 2388e4e574b7SAdam Litke /* 2389e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 2390e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 2391e4e574b7SAdam Litke */ 2392db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2393a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 2394a5516438SAndi Kleen (h->free_huge_pages + allocated); 239528073b02SHillf Danton if (needed > 0) { 239628073b02SHillf Danton if (alloc_ok) 2397e4e574b7SAdam Litke goto retry; 239828073b02SHillf Danton /* 239928073b02SHillf Danton * We were not able to allocate enough pages to 240028073b02SHillf Danton * satisfy the entire reservation so we free what 240128073b02SHillf Danton * we've allocated so far. 240228073b02SHillf Danton */ 240328073b02SHillf Danton goto free; 240428073b02SHillf Danton } 2405e4e574b7SAdam Litke /* 2406e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 240725985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 2408e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 2409ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 2410ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 2411ac09b3a1SAdam Litke * before they are reserved. 2412e4e574b7SAdam Litke */ 2413e4e574b7SAdam Litke needed += allocated; 2414a5516438SAndi Kleen h->resv_huge_pages += delta; 2415e4e574b7SAdam Litke ret = 0; 2416a9869b83SNaoya Horiguchi 241719fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 241819fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 241919fc3f0aSAdam Litke if ((--needed) < 0) 242019fc3f0aSAdam Litke break; 2421b65a4edaSMike Kravetz /* Add the page to the hugetlb allocator */ 2422a5516438SAndi Kleen enqueue_huge_page(h, page); 242319fc3f0aSAdam Litke } 242428073b02SHillf Danton free: 2425db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 242619fc3f0aSAdam Litke 2427b65a4edaSMike Kravetz /* 2428b65a4edaSMike Kravetz * Free unnecessary surplus pages to the buddy allocator. 2429b65a4edaSMike Kravetz * Pages have no ref count, call free_huge_page directly. 2430b65a4edaSMike Kravetz */ 2431c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 2432b65a4edaSMike Kravetz free_huge_page(page); 2433db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2434e4e574b7SAdam Litke 2435e4e574b7SAdam Litke return ret; 2436e4e574b7SAdam Litke } 2437e4e574b7SAdam Litke 2438e4e574b7SAdam Litke /* 2439e5bbc8a6SMike Kravetz * This routine has two main purposes: 2440e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 2441e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 2442e5bbc8a6SMike Kravetz * to the associated reservation map. 2443e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 2444e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 2445e4e574b7SAdam Litke */ 2446a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 2447a5516438SAndi Kleen unsigned long unused_resv_pages) 2448e4e574b7SAdam Litke { 2449e4e574b7SAdam Litke unsigned long nr_pages; 245010c6ec49SMike Kravetz struct page *page; 245110c6ec49SMike Kravetz LIST_HEAD(page_list); 245210c6ec49SMike Kravetz 24539487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 245410c6ec49SMike Kravetz /* Uncommit the reservation */ 245510c6ec49SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 2456e4e574b7SAdam Litke 2457c0531714SNaoya Horiguchi if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 2458e5bbc8a6SMike Kravetz goto out; 2459aa888a74SAndi Kleen 2460e5bbc8a6SMike Kravetz /* 2461e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 2462e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 2463e5bbc8a6SMike Kravetz */ 2464a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 2465e4e574b7SAdam Litke 2466685f3457SLee Schermerhorn /* 2467685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 24689b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 24699b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 24709b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 247110c6ec49SMike Kravetz * remove_pool_huge_page() will balance the freed pages across the 24729b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 2473685f3457SLee Schermerhorn */ 2474685f3457SLee Schermerhorn while (nr_pages--) { 247510c6ec49SMike Kravetz page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1); 247610c6ec49SMike Kravetz if (!page) 2477e5bbc8a6SMike Kravetz goto out; 247810c6ec49SMike Kravetz 247910c6ec49SMike Kravetz list_add(&page->lru, &page_list); 2480e4e574b7SAdam Litke } 2481e5bbc8a6SMike Kravetz 2482e5bbc8a6SMike Kravetz out: 2483db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 248410c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 2485db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2486e4e574b7SAdam Litke } 2487e4e574b7SAdam Litke 24885e911373SMike Kravetz 2489c37f9fb1SAndy Whitcroft /* 2490feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 24915e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 2492cf3ad20bSMike Kravetz * 2493cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 2494cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 2495cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 2496cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 2497cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 2498feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 2499feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 2500feba16e2SMike Kravetz * is called in such cases. 2501cf3ad20bSMike Kravetz * 2502cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 2503cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 2504cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 2505cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 2506cf3ad20bSMike Kravetz * take appropriate action. 250796b96a96SMike Kravetz * 250896b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 250996b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 251096b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 251196b96a96SMike Kravetz * reservation exists. 2512846be085SMike Kravetz * 2513846be085SMike Kravetz * vma_del_reservation is used in error paths where an entry in the reserve 2514846be085SMike Kravetz * map was created during huge page allocation and must be removed. It is to 2515846be085SMike Kravetz * be called after calling vma_needs_reservation to determine if a reservation 2516846be085SMike Kravetz * exists. 2517c37f9fb1SAndy Whitcroft */ 25185e911373SMike Kravetz enum vma_resv_mode { 25195e911373SMike Kravetz VMA_NEEDS_RESV, 25205e911373SMike Kravetz VMA_COMMIT_RESV, 2521feba16e2SMike Kravetz VMA_END_RESV, 252296b96a96SMike Kravetz VMA_ADD_RESV, 2523846be085SMike Kravetz VMA_DEL_RESV, 25245e911373SMike Kravetz }; 2525cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 2526cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 25275e911373SMike Kravetz enum vma_resv_mode mode) 2528c37f9fb1SAndy Whitcroft { 25294e35f483SJoonsoo Kim struct resv_map *resv; 25304e35f483SJoonsoo Kim pgoff_t idx; 2531cf3ad20bSMike Kravetz long ret; 25320db9d74eSMina Almasry long dummy_out_regions_needed; 2533c37f9fb1SAndy Whitcroft 25344e35f483SJoonsoo Kim resv = vma_resv_map(vma); 25354e35f483SJoonsoo Kim if (!resv) 2536c37f9fb1SAndy Whitcroft return 1; 2537c37f9fb1SAndy Whitcroft 25384e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 25395e911373SMike Kravetz switch (mode) { 25405e911373SMike Kravetz case VMA_NEEDS_RESV: 25410db9d74eSMina Almasry ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); 25420db9d74eSMina Almasry /* We assume that vma_reservation_* routines always operate on 25430db9d74eSMina Almasry * 1 page, and that adding to resv map a 1 page entry can only 25440db9d74eSMina Almasry * ever require 1 region. 25450db9d74eSMina Almasry */ 25460db9d74eSMina Almasry VM_BUG_ON(dummy_out_regions_needed != 1); 25475e911373SMike Kravetz break; 25485e911373SMike Kravetz case VMA_COMMIT_RESV: 2549075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 25500db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 25510db9d74eSMina Almasry VM_BUG_ON(ret < 0); 25525e911373SMike Kravetz break; 2553feba16e2SMike Kravetz case VMA_END_RESV: 25540db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 25555e911373SMike Kravetz ret = 0; 25565e911373SMike Kravetz break; 255796b96a96SMike Kravetz case VMA_ADD_RESV: 25580db9d74eSMina Almasry if (vma->vm_flags & VM_MAYSHARE) { 2559075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 25600db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 25610db9d74eSMina Almasry VM_BUG_ON(ret < 0); 25620db9d74eSMina Almasry } else { 25630db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 256496b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 256596b96a96SMike Kravetz } 256696b96a96SMike Kravetz break; 2567846be085SMike Kravetz case VMA_DEL_RESV: 2568846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 2569846be085SMike Kravetz region_abort(resv, idx, idx + 1, 1); 2570846be085SMike Kravetz ret = region_del(resv, idx, idx + 1); 2571846be085SMike Kravetz } else { 2572846be085SMike Kravetz ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 2573846be085SMike Kravetz /* region_add calls of range 1 should never fail. */ 2574846be085SMike Kravetz VM_BUG_ON(ret < 0); 2575846be085SMike Kravetz } 2576846be085SMike Kravetz break; 25775e911373SMike Kravetz default: 25785e911373SMike Kravetz BUG(); 25795e911373SMike Kravetz } 258084afd99bSAndy Whitcroft 2581846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV) 2582cf3ad20bSMike Kravetz return ret; 258367961f9dSMike Kravetz /* 2584bf3d12b9SMiaohe Lin * We know private mapping must have HPAGE_RESV_OWNER set. 2585bf3d12b9SMiaohe Lin * 258667961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 258767961f9dSMike Kravetz * However, a file associated with mapping could have been 258867961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 258967961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 259067961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 259167961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 259267961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 259367961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 259467961f9dSMike Kravetz * reservation has already been consumed. As a result, the 259567961f9dSMike Kravetz * return value of this routine is the opposite of the 259667961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 259767961f9dSMike Kravetz */ 2598bf3d12b9SMiaohe Lin if (ret > 0) 259967961f9dSMike Kravetz return 0; 2600bf3d12b9SMiaohe Lin if (ret == 0) 260167961f9dSMike Kravetz return 1; 2602bf3d12b9SMiaohe Lin return ret; 260384afd99bSAndy Whitcroft } 2604cf3ad20bSMike Kravetz 2605cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 2606a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 2607c37f9fb1SAndy Whitcroft { 26085e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 2609cf3ad20bSMike Kravetz } 2610c37f9fb1SAndy Whitcroft 2611cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 2612cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 2613cf3ad20bSMike Kravetz { 26145e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 26155e911373SMike Kravetz } 26165e911373SMike Kravetz 2617feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 26185e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 26195e911373SMike Kravetz { 2620feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 2621c37f9fb1SAndy Whitcroft } 2622c37f9fb1SAndy Whitcroft 262396b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 262496b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 262596b96a96SMike Kravetz { 262696b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 262796b96a96SMike Kravetz } 262896b96a96SMike Kravetz 2629846be085SMike Kravetz static long vma_del_reservation(struct hstate *h, 2630846be085SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 263196b96a96SMike Kravetz { 2632846be085SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV); 2633846be085SMike Kravetz } 2634846be085SMike Kravetz 2635846be085SMike Kravetz /* 2636846be085SMike Kravetz * This routine is called to restore reservation information on error paths. 2637846be085SMike Kravetz * It should ONLY be called for pages allocated via alloc_huge_page(), and 2638846be085SMike Kravetz * the hugetlb mutex should remain held when calling this routine. 2639846be085SMike Kravetz * 2640846be085SMike Kravetz * It handles two specific cases: 2641846be085SMike Kravetz * 1) A reservation was in place and the page consumed the reservation. 2642846be085SMike Kravetz * HPageRestoreReserve is set in the page. 2643846be085SMike Kravetz * 2) No reservation was in place for the page, so HPageRestoreReserve is 2644846be085SMike Kravetz * not set. However, alloc_huge_page always updates the reserve map. 2645846be085SMike Kravetz * 2646846be085SMike Kravetz * In case 1, free_huge_page later in the error path will increment the 2647846be085SMike Kravetz * global reserve count. But, free_huge_page does not have enough context 2648846be085SMike Kravetz * to adjust the reservation map. This case deals primarily with private 2649846be085SMike Kravetz * mappings. Adjust the reserve map here to be consistent with global 2650846be085SMike Kravetz * reserve count adjustments to be made by free_huge_page. Make sure the 2651846be085SMike Kravetz * reserve map indicates there is a reservation present. 2652846be085SMike Kravetz * 2653846be085SMike Kravetz * In case 2, simply undo reserve map modifications done by alloc_huge_page. 2654846be085SMike Kravetz */ 2655846be085SMike Kravetz void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma, 2656846be085SMike Kravetz unsigned long address, struct page *page) 2657846be085SMike Kravetz { 265896b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 265996b96a96SMike Kravetz 2660846be085SMike Kravetz if (HPageRestoreReserve(page)) { 2661846be085SMike Kravetz if (unlikely(rc < 0)) 266296b96a96SMike Kravetz /* 266396b96a96SMike Kravetz * Rare out of memory condition in reserve map 2664d6995da3SMike Kravetz * manipulation. Clear HPageRestoreReserve so that 266596b96a96SMike Kravetz * global reserve count will not be incremented 266696b96a96SMike Kravetz * by free_huge_page. This will make it appear 266796b96a96SMike Kravetz * as though the reservation for this page was 266896b96a96SMike Kravetz * consumed. This may prevent the task from 266996b96a96SMike Kravetz * faulting in the page at a later time. This 267096b96a96SMike Kravetz * is better than inconsistent global huge page 267196b96a96SMike Kravetz * accounting of reserve counts. 267296b96a96SMike Kravetz */ 2673d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 2674846be085SMike Kravetz else if (rc) 2675846be085SMike Kravetz (void)vma_add_reservation(h, vma, address); 2676846be085SMike Kravetz else 2677846be085SMike Kravetz vma_end_reservation(h, vma, address); 2678846be085SMike Kravetz } else { 2679846be085SMike Kravetz if (!rc) { 268096b96a96SMike Kravetz /* 2681846be085SMike Kravetz * This indicates there is an entry in the reserve map 2682c7b1850dSMike Kravetz * not added by alloc_huge_page. We know it was added 2683846be085SMike Kravetz * before the alloc_huge_page call, otherwise 2684846be085SMike Kravetz * HPageRestoreReserve would be set on the page. 2685846be085SMike Kravetz * Remove the entry so that a subsequent allocation 2686846be085SMike Kravetz * does not consume a reservation. 268796b96a96SMike Kravetz */ 2688846be085SMike Kravetz rc = vma_del_reservation(h, vma, address); 2689846be085SMike Kravetz if (rc < 0) 2690846be085SMike Kravetz /* 2691846be085SMike Kravetz * VERY rare out of memory condition. Since 2692846be085SMike Kravetz * we can not delete the entry, set 2693846be085SMike Kravetz * HPageRestoreReserve so that the reserve 2694846be085SMike Kravetz * count will be incremented when the page 2695846be085SMike Kravetz * is freed. This reserve will be consumed 2696846be085SMike Kravetz * on a subsequent allocation. 2697846be085SMike Kravetz */ 2698846be085SMike Kravetz SetHPageRestoreReserve(page); 2699846be085SMike Kravetz } else if (rc < 0) { 2700846be085SMike Kravetz /* 2701846be085SMike Kravetz * Rare out of memory condition from 2702846be085SMike Kravetz * vma_needs_reservation call. Memory allocation is 2703846be085SMike Kravetz * only attempted if a new entry is needed. Therefore, 2704846be085SMike Kravetz * this implies there is not an entry in the 2705846be085SMike Kravetz * reserve map. 2706846be085SMike Kravetz * 2707846be085SMike Kravetz * For shared mappings, no entry in the map indicates 2708846be085SMike Kravetz * no reservation. We are done. 2709846be085SMike Kravetz */ 2710846be085SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 2711846be085SMike Kravetz /* 2712846be085SMike Kravetz * For private mappings, no entry indicates 2713846be085SMike Kravetz * a reservation is present. Since we can 2714846be085SMike Kravetz * not add an entry, set SetHPageRestoreReserve 2715846be085SMike Kravetz * on the page so reserve count will be 2716846be085SMike Kravetz * incremented when freed. This reserve will 2717846be085SMike Kravetz * be consumed on a subsequent allocation. 2718846be085SMike Kravetz */ 2719846be085SMike Kravetz SetHPageRestoreReserve(page); 272096b96a96SMike Kravetz } else 2721846be085SMike Kravetz /* 2722846be085SMike Kravetz * No reservation present, do nothing 2723846be085SMike Kravetz */ 272496b96a96SMike Kravetz vma_end_reservation(h, vma, address); 272596b96a96SMike Kravetz } 272696b96a96SMike Kravetz } 272796b96a96SMike Kravetz 2728369fa227SOscar Salvador /* 2729369fa227SOscar Salvador * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one 2730369fa227SOscar Salvador * @h: struct hstate old page belongs to 2731369fa227SOscar Salvador * @old_page: Old page to dissolve 2732ae37c7ffSOscar Salvador * @list: List to isolate the page in case we need to 2733369fa227SOscar Salvador * Returns 0 on success, otherwise negated error. 2734369fa227SOscar Salvador */ 2735ae37c7ffSOscar Salvador static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page, 2736ae37c7ffSOscar Salvador struct list_head *list) 2737369fa227SOscar Salvador { 2738369fa227SOscar Salvador gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2739369fa227SOscar Salvador int nid = page_to_nid(old_page); 2740b65a4edaSMike Kravetz bool alloc_retry = false; 2741369fa227SOscar Salvador struct page *new_page; 2742369fa227SOscar Salvador int ret = 0; 2743369fa227SOscar Salvador 2744369fa227SOscar Salvador /* 2745369fa227SOscar Salvador * Before dissolving the page, we need to allocate a new one for the 2746f41f2ed4SMuchun Song * pool to remain stable. Here, we allocate the page and 'prep' it 2747f41f2ed4SMuchun Song * by doing everything but actually updating counters and adding to 2748f41f2ed4SMuchun Song * the pool. This simplifies and let us do most of the processing 2749f41f2ed4SMuchun Song * under the lock. 2750369fa227SOscar Salvador */ 2751b65a4edaSMike Kravetz alloc_retry: 2752369fa227SOscar Salvador new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL); 2753369fa227SOscar Salvador if (!new_page) 2754369fa227SOscar Salvador return -ENOMEM; 2755b65a4edaSMike Kravetz /* 2756b65a4edaSMike Kravetz * If all goes well, this page will be directly added to the free 2757b65a4edaSMike Kravetz * list in the pool. For this the ref count needs to be zero. 2758b65a4edaSMike Kravetz * Attempt to drop now, and retry once if needed. It is VERY 2759b65a4edaSMike Kravetz * unlikely there is another ref on the page. 2760b65a4edaSMike Kravetz * 2761b65a4edaSMike Kravetz * If someone else has a reference to the page, it will be freed 2762b65a4edaSMike Kravetz * when they drop their ref. Abuse temporary page flag to accomplish 2763b65a4edaSMike Kravetz * this. Retry once if there is an inflated ref count. 2764b65a4edaSMike Kravetz */ 2765b65a4edaSMike Kravetz SetHPageTemporary(new_page); 2766b65a4edaSMike Kravetz if (!put_page_testzero(new_page)) { 2767b65a4edaSMike Kravetz if (alloc_retry) 2768b65a4edaSMike Kravetz return -EBUSY; 2769b65a4edaSMike Kravetz 2770b65a4edaSMike Kravetz alloc_retry = true; 2771b65a4edaSMike Kravetz goto alloc_retry; 2772b65a4edaSMike Kravetz } 2773b65a4edaSMike Kravetz ClearHPageTemporary(new_page); 2774b65a4edaSMike Kravetz 2775f41f2ed4SMuchun Song __prep_new_huge_page(h, new_page); 2776369fa227SOscar Salvador 2777369fa227SOscar Salvador retry: 2778369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2779369fa227SOscar Salvador if (!PageHuge(old_page)) { 2780369fa227SOscar Salvador /* 2781369fa227SOscar Salvador * Freed from under us. Drop new_page too. 2782369fa227SOscar Salvador */ 2783369fa227SOscar Salvador goto free_new; 2784369fa227SOscar Salvador } else if (page_count(old_page)) { 2785369fa227SOscar Salvador /* 2786ae37c7ffSOscar Salvador * Someone has grabbed the page, try to isolate it here. 2787ae37c7ffSOscar Salvador * Fail with -EBUSY if not possible. 2788369fa227SOscar Salvador */ 2789ae37c7ffSOscar Salvador spin_unlock_irq(&hugetlb_lock); 27907ce82f4cSMiaohe Lin ret = isolate_hugetlb(old_page, list); 2791ae37c7ffSOscar Salvador spin_lock_irq(&hugetlb_lock); 2792369fa227SOscar Salvador goto free_new; 2793369fa227SOscar Salvador } else if (!HPageFreed(old_page)) { 2794369fa227SOscar Salvador /* 2795369fa227SOscar Salvador * Page's refcount is 0 but it has not been enqueued in the 2796369fa227SOscar Salvador * freelist yet. Race window is small, so we can succeed here if 2797369fa227SOscar Salvador * we retry. 2798369fa227SOscar Salvador */ 2799369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2800369fa227SOscar Salvador cond_resched(); 2801369fa227SOscar Salvador goto retry; 2802369fa227SOscar Salvador } else { 2803369fa227SOscar Salvador /* 2804369fa227SOscar Salvador * Ok, old_page is still a genuine free hugepage. Remove it from 2805369fa227SOscar Salvador * the freelist and decrease the counters. These will be 2806369fa227SOscar Salvador * incremented again when calling __prep_account_new_huge_page() 2807369fa227SOscar Salvador * and enqueue_huge_page() for new_page. The counters will remain 2808369fa227SOscar Salvador * stable since this happens under the lock. 2809369fa227SOscar Salvador */ 2810369fa227SOscar Salvador remove_hugetlb_page(h, old_page, false); 2811369fa227SOscar Salvador 2812369fa227SOscar Salvador /* 2813b65a4edaSMike Kravetz * Ref count on new page is already zero as it was dropped 2814b65a4edaSMike Kravetz * earlier. It can be directly added to the pool free list. 2815369fa227SOscar Salvador */ 2816369fa227SOscar Salvador __prep_account_new_huge_page(h, nid); 2817369fa227SOscar Salvador enqueue_huge_page(h, new_page); 2818369fa227SOscar Salvador 2819369fa227SOscar Salvador /* 2820369fa227SOscar Salvador * Pages have been replaced, we can safely free the old one. 2821369fa227SOscar Salvador */ 2822369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2823b65d4adbSMuchun Song update_and_free_page(h, old_page, false); 2824369fa227SOscar Salvador } 2825369fa227SOscar Salvador 2826369fa227SOscar Salvador return ret; 2827369fa227SOscar Salvador 2828369fa227SOscar Salvador free_new: 2829369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2830b65a4edaSMike Kravetz /* Page has a zero ref count, but needs a ref to be freed */ 2831b65a4edaSMike Kravetz set_page_refcounted(new_page); 2832b65d4adbSMuchun Song update_and_free_page(h, new_page, false); 2833369fa227SOscar Salvador 2834369fa227SOscar Salvador return ret; 2835369fa227SOscar Salvador } 2836369fa227SOscar Salvador 2837ae37c7ffSOscar Salvador int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list) 2838369fa227SOscar Salvador { 2839369fa227SOscar Salvador struct hstate *h; 2840369fa227SOscar Salvador struct page *head; 2841ae37c7ffSOscar Salvador int ret = -EBUSY; 2842369fa227SOscar Salvador 2843369fa227SOscar Salvador /* 2844369fa227SOscar Salvador * The page might have been dissolved from under our feet, so make sure 2845369fa227SOscar Salvador * to carefully check the state under the lock. 2846369fa227SOscar Salvador * Return success when racing as if we dissolved the page ourselves. 2847369fa227SOscar Salvador */ 2848369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2849369fa227SOscar Salvador if (PageHuge(page)) { 2850369fa227SOscar Salvador head = compound_head(page); 2851369fa227SOscar Salvador h = page_hstate(head); 2852369fa227SOscar Salvador } else { 2853369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2854369fa227SOscar Salvador return 0; 2855369fa227SOscar Salvador } 2856369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2857369fa227SOscar Salvador 2858369fa227SOscar Salvador /* 2859369fa227SOscar Salvador * Fence off gigantic pages as there is a cyclic dependency between 2860369fa227SOscar Salvador * alloc_contig_range and them. Return -ENOMEM as this has the effect 2861369fa227SOscar Salvador * of bailing out right away without further retrying. 2862369fa227SOscar Salvador */ 2863369fa227SOscar Salvador if (hstate_is_gigantic(h)) 2864369fa227SOscar Salvador return -ENOMEM; 2865369fa227SOscar Salvador 28667ce82f4cSMiaohe Lin if (page_count(head) && !isolate_hugetlb(head, list)) 2867ae37c7ffSOscar Salvador ret = 0; 2868ae37c7ffSOscar Salvador else if (!page_count(head)) 2869ae37c7ffSOscar Salvador ret = alloc_and_dissolve_huge_page(h, head, list); 2870ae37c7ffSOscar Salvador 2871ae37c7ffSOscar Salvador return ret; 2872369fa227SOscar Salvador } 2873369fa227SOscar Salvador 287470c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 287504f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 2876348ea204SAdam Litke { 287790481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 2878a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2879348ea204SAdam Litke struct page *page; 2880d85f69b0SMike Kravetz long map_chg, map_commit; 2881d85f69b0SMike Kravetz long gbl_chg; 28826d76dcf4SAneesh Kumar K.V int ret, idx; 28836d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 288408cf9fafSMina Almasry bool deferred_reserve; 28852fc39cecSAdam Litke 28866d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 2887a1e78772SMel Gorman /* 2888d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 2889d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 2890d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 2891a1e78772SMel Gorman */ 2892d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 2893d85f69b0SMike Kravetz if (map_chg < 0) 289476dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 2895d85f69b0SMike Kravetz 2896d85f69b0SMike Kravetz /* 2897d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 2898d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 2899d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 2900d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 2901d85f69b0SMike Kravetz * checked against any subpool limit. 2902d85f69b0SMike Kravetz */ 2903d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 2904d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 2905d85f69b0SMike Kravetz if (gbl_chg < 0) { 2906feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 290776dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 29085e911373SMike Kravetz } 290990d8b7e6SAdam Litke 2910d85f69b0SMike Kravetz /* 2911d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 2912d85f69b0SMike Kravetz * map, there could be reservations associated with the 2913d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 2914d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 2915d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 2916d85f69b0SMike Kravetz * the subpool reservations. 2917d85f69b0SMike Kravetz */ 2918d85f69b0SMike Kravetz if (avoid_reserve) 2919d85f69b0SMike Kravetz gbl_chg = 1; 2920d85f69b0SMike Kravetz } 2921d85f69b0SMike Kravetz 292208cf9fafSMina Almasry /* If this allocation is not consuming a reservation, charge it now. 292308cf9fafSMina Almasry */ 29246501fe5fSMiaohe Lin deferred_reserve = map_chg || avoid_reserve; 292508cf9fafSMina Almasry if (deferred_reserve) { 292608cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 292708cf9fafSMina Almasry idx, pages_per_huge_page(h), &h_cg); 29288f34af6fSJianyu Zhan if (ret) 29298f34af6fSJianyu Zhan goto out_subpool_put; 293008cf9fafSMina Almasry } 293108cf9fafSMina Almasry 293208cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 293308cf9fafSMina Almasry if (ret) 293408cf9fafSMina Almasry goto out_uncharge_cgroup_reservation; 29358f34af6fSJianyu Zhan 2936db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2937d85f69b0SMike Kravetz /* 2938d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 2939d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 2940d85f69b0SMike Kravetz * a reservation exists for the allocation. 2941d85f69b0SMike Kravetz */ 2942d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 294381a6fcaeSJoonsoo Kim if (!page) { 2944db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 29450c397daeSMichal Hocko page = alloc_buddy_huge_page_with_mpol(h, vma, addr); 29468f34af6fSJianyu Zhan if (!page) 29478f34af6fSJianyu Zhan goto out_uncharge_cgroup; 2948a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 2949d6995da3SMike Kravetz SetHPageRestoreReserve(page); 2950a88c7695SNaoya Horiguchi h->resv_huge_pages--; 2951a88c7695SNaoya Horiguchi } 2952db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 295315a8d68eSWei Yang list_add(&page->lru, &h->hugepage_activelist); 295481a6fcaeSJoonsoo Kim /* Fall through */ 2955a1e78772SMel Gorman } 295681a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 295708cf9fafSMina Almasry /* If allocation is not consuming a reservation, also store the 295808cf9fafSMina Almasry * hugetlb_cgroup pointer on the page. 295908cf9fafSMina Almasry */ 296008cf9fafSMina Almasry if (deferred_reserve) { 296108cf9fafSMina Almasry hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), 296208cf9fafSMina Almasry h_cg, page); 296308cf9fafSMina Almasry } 296408cf9fafSMina Almasry 2965db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2966a1e78772SMel Gorman 2967d6995da3SMike Kravetz hugetlb_set_page_subpool(page, spool); 2968a1e78772SMel Gorman 2969d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 2970d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 297133039678SMike Kravetz /* 297233039678SMike Kravetz * The page was added to the reservation map between 297333039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 297433039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 297533039678SMike Kravetz * Adjust for the subpool count incremented above AND 297633039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 297733039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 297833039678SMike Kravetz * no longer applies. 297933039678SMike Kravetz */ 298033039678SMike Kravetz long rsv_adjust; 298133039678SMike Kravetz 298233039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 298333039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 298479aa925bSMike Kravetz if (deferred_reserve) 298579aa925bSMike Kravetz hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 298679aa925bSMike Kravetz pages_per_huge_page(h), page); 298733039678SMike Kravetz } 29887893d1d5SAdam Litke return page; 29898f34af6fSJianyu Zhan 29908f34af6fSJianyu Zhan out_uncharge_cgroup: 29918f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 299208cf9fafSMina Almasry out_uncharge_cgroup_reservation: 299308cf9fafSMina Almasry if (deferred_reserve) 299408cf9fafSMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), 299508cf9fafSMina Almasry h_cg); 29968f34af6fSJianyu Zhan out_subpool_put: 2997d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 29988f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 2999feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 30008f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 3001b45b5bd6SDavid Gibson } 3002b45b5bd6SDavid Gibson 3003b5389086SZhenguo Yao int alloc_bootmem_huge_page(struct hstate *h, int nid) 3004e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 3005b5389086SZhenguo Yao int __alloc_bootmem_huge_page(struct hstate *h, int nid) 3006aa888a74SAndi Kleen { 3007b5389086SZhenguo Yao struct huge_bootmem_page *m = NULL; /* initialize for clang */ 3008b2261026SJoonsoo Kim int nr_nodes, node; 3009aa888a74SAndi Kleen 3010b5389086SZhenguo Yao /* do node specific alloc */ 3011b5389086SZhenguo Yao if (nid != NUMA_NO_NODE) { 3012b5389086SZhenguo Yao m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h), 3013b5389086SZhenguo Yao 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid); 3014b5389086SZhenguo Yao if (!m) 3015b5389086SZhenguo Yao return 0; 3016b5389086SZhenguo Yao goto found; 3017b5389086SZhenguo Yao } 3018b5389086SZhenguo Yao /* allocate from next node when distributing huge pages */ 3019b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 3020b5389086SZhenguo Yao m = memblock_alloc_try_nid_raw( 30218b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 302297ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 3023aa888a74SAndi Kleen /* 3024aa888a74SAndi Kleen * Use the beginning of the huge page to store the 3025aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 3026aa888a74SAndi Kleen * puts them into the mem_map). 3027aa888a74SAndi Kleen */ 3028b5389086SZhenguo Yao if (!m) 3029b5389086SZhenguo Yao return 0; 3030aa888a74SAndi Kleen goto found; 3031aa888a74SAndi Kleen } 3032aa888a74SAndi Kleen 3033aa888a74SAndi Kleen found: 3034aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 3035330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 3036aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 3037aa888a74SAndi Kleen m->hstate = h; 3038aa888a74SAndi Kleen return 1; 3039aa888a74SAndi Kleen } 3040aa888a74SAndi Kleen 304148b8d744SMike Kravetz /* 304248b8d744SMike Kravetz * Put bootmem huge pages into the standard lists after mem_map is up. 304348b8d744SMike Kravetz * Note: This only applies to gigantic (order > MAX_ORDER) pages. 304448b8d744SMike Kravetz */ 3045aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 3046aa888a74SAndi Kleen { 3047aa888a74SAndi Kleen struct huge_bootmem_page *m; 3048aa888a74SAndi Kleen 3049aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 305040d18ebfSMike Kravetz struct page *page = virt_to_page(m); 3051aa888a74SAndi Kleen struct hstate *h = m->hstate; 3052ee8f248dSBecky Bruce 305348b8d744SMike Kravetz VM_BUG_ON(!hstate_is_gigantic(h)); 3054aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 30557118fc29SMike Kravetz if (prep_compound_gigantic_page(page, huge_page_order(h))) { 3056ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 3057aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 30587118fc29SMike Kravetz put_page(page); /* add to the hugepage allocator */ 30597118fc29SMike Kravetz } else { 3060416d85edSMike Kravetz /* VERY unlikely inflated ref count on a tail page */ 30617118fc29SMike Kravetz free_gigantic_page(page, huge_page_order(h)); 30627118fc29SMike Kravetz } 3063af0fb9dfSMichal Hocko 3064b0320c7bSRafael Aquini /* 306548b8d744SMike Kravetz * We need to restore the 'stolen' pages to totalram_pages 306648b8d744SMike Kravetz * in order to fix confusing memory reports from free(1) and 306748b8d744SMike Kravetz * other side-effects, like CommitLimit going negative. 3068b0320c7bSRafael Aquini */ 3069c78a7f36SMiaohe Lin adjust_managed_page_count(page, pages_per_huge_page(h)); 3070520495feSCannon Matthews cond_resched(); 3071aa888a74SAndi Kleen } 3072aa888a74SAndi Kleen } 3073b5389086SZhenguo Yao static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid) 3074b5389086SZhenguo Yao { 3075b5389086SZhenguo Yao unsigned long i; 3076b5389086SZhenguo Yao char buf[32]; 3077b5389086SZhenguo Yao 3078b5389086SZhenguo Yao for (i = 0; i < h->max_huge_pages_node[nid]; ++i) { 3079b5389086SZhenguo Yao if (hstate_is_gigantic(h)) { 3080b5389086SZhenguo Yao if (!alloc_bootmem_huge_page(h, nid)) 3081b5389086SZhenguo Yao break; 3082b5389086SZhenguo Yao } else { 3083b5389086SZhenguo Yao struct page *page; 3084b5389086SZhenguo Yao gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 3085b5389086SZhenguo Yao 3086b5389086SZhenguo Yao page = alloc_fresh_huge_page(h, gfp_mask, nid, 3087b5389086SZhenguo Yao &node_states[N_MEMORY], NULL); 3088b5389086SZhenguo Yao if (!page) 3089b5389086SZhenguo Yao break; 3090b5389086SZhenguo Yao put_page(page); /* free it into the hugepage allocator */ 3091b5389086SZhenguo Yao } 3092b5389086SZhenguo Yao cond_resched(); 3093b5389086SZhenguo Yao } 3094b5389086SZhenguo Yao if (i == h->max_huge_pages_node[nid]) 3095b5389086SZhenguo Yao return; 3096b5389086SZhenguo Yao 3097b5389086SZhenguo Yao string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3098b5389086SZhenguo Yao pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n", 3099b5389086SZhenguo Yao h->max_huge_pages_node[nid], buf, nid, i); 3100b5389086SZhenguo Yao h->max_huge_pages -= (h->max_huge_pages_node[nid] - i); 3101b5389086SZhenguo Yao h->max_huge_pages_node[nid] = i; 3102b5389086SZhenguo Yao } 3103aa888a74SAndi Kleen 31048faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 31051da177e4SLinus Torvalds { 31061da177e4SLinus Torvalds unsigned long i; 3107f60858f9SMike Kravetz nodemask_t *node_alloc_noretry; 3108b5389086SZhenguo Yao bool node_specific_alloc = false; 3109f60858f9SMike Kravetz 3110b5389086SZhenguo Yao /* skip gigantic hugepages allocation if hugetlb_cma enabled */ 3111b5389086SZhenguo Yao if (hstate_is_gigantic(h) && hugetlb_cma_size) { 3112b5389086SZhenguo Yao pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); 3113b5389086SZhenguo Yao return; 3114b5389086SZhenguo Yao } 3115b5389086SZhenguo Yao 3116b5389086SZhenguo Yao /* do node specific alloc */ 31170a7a0f6fSPeng Liu for_each_online_node(i) { 3118b5389086SZhenguo Yao if (h->max_huge_pages_node[i] > 0) { 3119b5389086SZhenguo Yao hugetlb_hstate_alloc_pages_onenode(h, i); 3120b5389086SZhenguo Yao node_specific_alloc = true; 3121b5389086SZhenguo Yao } 3122b5389086SZhenguo Yao } 3123b5389086SZhenguo Yao 3124b5389086SZhenguo Yao if (node_specific_alloc) 3125b5389086SZhenguo Yao return; 3126b5389086SZhenguo Yao 3127b5389086SZhenguo Yao /* below will do all node balanced alloc */ 3128f60858f9SMike Kravetz if (!hstate_is_gigantic(h)) { 3129f60858f9SMike Kravetz /* 3130f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3131f60858f9SMike Kravetz * Ignore errors as lower level routines can deal with 3132f60858f9SMike Kravetz * node_alloc_noretry == NULL. If this kmalloc fails at boot 3133f60858f9SMike Kravetz * time, we are likely in bigger trouble. 3134f60858f9SMike Kravetz */ 3135f60858f9SMike Kravetz node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), 3136f60858f9SMike Kravetz GFP_KERNEL); 3137f60858f9SMike Kravetz } else { 3138f60858f9SMike Kravetz /* allocations done at boot time */ 3139f60858f9SMike Kravetz node_alloc_noretry = NULL; 3140f60858f9SMike Kravetz } 3141f60858f9SMike Kravetz 3142f60858f9SMike Kravetz /* bit mask controlling how hard we retry per-node allocations */ 3143f60858f9SMike Kravetz if (node_alloc_noretry) 3144f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 31451da177e4SLinus Torvalds 3146e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 3147bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 3148b5389086SZhenguo Yao if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE)) 3149aa888a74SAndi Kleen break; 31500c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 3151f60858f9SMike Kravetz &node_states[N_MEMORY], 3152f60858f9SMike Kravetz node_alloc_noretry)) 31531da177e4SLinus Torvalds break; 315469ed779aSDavid Rientjes cond_resched(); 31551da177e4SLinus Torvalds } 3156d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 3157d715cf80SLiam R. Howlett char buf[32]; 3158d715cf80SLiam R. Howlett 3159c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3160d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 3161d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 31628faa8b07SAndi Kleen h->max_huge_pages = i; 3163e5ff2159SAndi Kleen } 3164f60858f9SMike Kravetz kfree(node_alloc_noretry); 3165d715cf80SLiam R. Howlett } 3166e5ff2159SAndi Kleen 3167e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 3168e5ff2159SAndi Kleen { 316979dfc695SMike Kravetz struct hstate *h, *h2; 3170e5ff2159SAndi Kleen 3171e5ff2159SAndi Kleen for_each_hstate(h) { 31728faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 3173bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 31748faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 317579dfc695SMike Kravetz 317679dfc695SMike Kravetz /* 317779dfc695SMike Kravetz * Set demote order for each hstate. Note that 317879dfc695SMike Kravetz * h->demote_order is initially 0. 317979dfc695SMike Kravetz * - We can not demote gigantic pages if runtime freeing 318079dfc695SMike Kravetz * is not supported, so skip this. 3181a01f4390SMike Kravetz * - If CMA allocation is possible, we can not demote 3182a01f4390SMike Kravetz * HUGETLB_PAGE_ORDER or smaller size pages. 318379dfc695SMike Kravetz */ 318479dfc695SMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 318579dfc695SMike Kravetz continue; 3186a01f4390SMike Kravetz if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER) 3187a01f4390SMike Kravetz continue; 318879dfc695SMike Kravetz for_each_hstate(h2) { 318979dfc695SMike Kravetz if (h2 == h) 319079dfc695SMike Kravetz continue; 319179dfc695SMike Kravetz if (h2->order < h->order && 319279dfc695SMike Kravetz h2->order > h->demote_order) 319379dfc695SMike Kravetz h->demote_order = h2->order; 319479dfc695SMike Kravetz } 3195e5ff2159SAndi Kleen } 3196e5ff2159SAndi Kleen } 3197e5ff2159SAndi Kleen 3198e5ff2159SAndi Kleen static void __init report_hugepages(void) 3199e5ff2159SAndi Kleen { 3200e5ff2159SAndi Kleen struct hstate *h; 3201e5ff2159SAndi Kleen 3202e5ff2159SAndi Kleen for_each_hstate(h) { 32034abd32dbSAndi Kleen char buf[32]; 3204c6247f72SMatthew Wilcox 3205c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 32066213834cSMuchun Song pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n", 3207c6247f72SMatthew Wilcox buf, h->free_huge_pages); 32086213834cSMuchun Song pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n", 32096213834cSMuchun Song hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf); 3210e5ff2159SAndi Kleen } 3211e5ff2159SAndi Kleen } 3212e5ff2159SAndi Kleen 32131da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 32146ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 32156ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 32161da177e4SLinus Torvalds { 32174415cc8dSChristoph Lameter int i; 32181121828aSMike Kravetz LIST_HEAD(page_list); 32194415cc8dSChristoph Lameter 32209487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 3221bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3222aa888a74SAndi Kleen return; 3223aa888a74SAndi Kleen 32241121828aSMike Kravetz /* 32251121828aSMike Kravetz * Collect pages to be freed on a list, and free after dropping lock 32261121828aSMike Kravetz */ 32276ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 322810c6ec49SMike Kravetz struct page *page, *next; 3229a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 3230a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 3231a5516438SAndi Kleen if (count >= h->nr_huge_pages) 32321121828aSMike Kravetz goto out; 32331da177e4SLinus Torvalds if (PageHighMem(page)) 32341da177e4SLinus Torvalds continue; 32356eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 32361121828aSMike Kravetz list_add(&page->lru, &page_list); 32371121828aSMike Kravetz } 32381121828aSMike Kravetz } 32391121828aSMike Kravetz 32401121828aSMike Kravetz out: 3241db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 324210c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3243db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 32441da177e4SLinus Torvalds } 32451da177e4SLinus Torvalds #else 32466ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 32476ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 32481da177e4SLinus Torvalds { 32491da177e4SLinus Torvalds } 32501da177e4SLinus Torvalds #endif 32511da177e4SLinus Torvalds 325220a0307cSWu Fengguang /* 325320a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 325420a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 325520a0307cSWu Fengguang * Returns 1 if an adjustment was made. 325620a0307cSWu Fengguang */ 32576ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 32586ae11b27SLee Schermerhorn int delta) 325920a0307cSWu Fengguang { 3260b2261026SJoonsoo Kim int nr_nodes, node; 326120a0307cSWu Fengguang 32629487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 326320a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 326420a0307cSWu Fengguang 3265e8c5c824SLee Schermerhorn if (delta < 0) { 3266b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 3267b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 3268b2261026SJoonsoo Kim goto found; 3269b2261026SJoonsoo Kim } 3270b2261026SJoonsoo Kim } else { 3271b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 3272b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 3273b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 3274b2261026SJoonsoo Kim goto found; 3275e8c5c824SLee Schermerhorn } 32769a76db09SLee Schermerhorn } 3277b2261026SJoonsoo Kim return 0; 327820a0307cSWu Fengguang 3279b2261026SJoonsoo Kim found: 328020a0307cSWu Fengguang h->surplus_huge_pages += delta; 3281b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 3282b2261026SJoonsoo Kim return 1; 328320a0307cSWu Fengguang } 328420a0307cSWu Fengguang 3285a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 3286fd875dcaSMike Kravetz static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 32876ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 32881da177e4SLinus Torvalds { 32897893d1d5SAdam Litke unsigned long min_count, ret; 329010c6ec49SMike Kravetz struct page *page; 329110c6ec49SMike Kravetz LIST_HEAD(page_list); 3292f60858f9SMike Kravetz NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); 3293f60858f9SMike Kravetz 3294f60858f9SMike Kravetz /* 3295f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3296f60858f9SMike Kravetz * If we can not allocate the bit mask, do not attempt to allocate 3297f60858f9SMike Kravetz * the requested huge pages. 3298f60858f9SMike Kravetz */ 3299f60858f9SMike Kravetz if (node_alloc_noretry) 3300f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 3301f60858f9SMike Kravetz else 3302f60858f9SMike Kravetz return -ENOMEM; 33031da177e4SLinus Torvalds 330429383967SMike Kravetz /* 330529383967SMike Kravetz * resize_lock mutex prevents concurrent adjustments to number of 330629383967SMike Kravetz * pages in hstate via the proc/sysfs interfaces. 330729383967SMike Kravetz */ 330829383967SMike Kravetz mutex_lock(&h->resize_lock); 3309b65d4adbSMuchun Song flush_free_hpage_work(h); 3310db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 33114eb0716eSAlexandre Ghiti 33124eb0716eSAlexandre Ghiti /* 3313fd875dcaSMike Kravetz * Check for a node specific request. 3314fd875dcaSMike Kravetz * Changing node specific huge page count may require a corresponding 3315fd875dcaSMike Kravetz * change to the global count. In any case, the passed node mask 3316fd875dcaSMike Kravetz * (nodes_allowed) will restrict alloc/free to the specified node. 3317fd875dcaSMike Kravetz */ 3318fd875dcaSMike Kravetz if (nid != NUMA_NO_NODE) { 3319fd875dcaSMike Kravetz unsigned long old_count = count; 3320fd875dcaSMike Kravetz 3321fd875dcaSMike Kravetz count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 3322fd875dcaSMike Kravetz /* 3323fd875dcaSMike Kravetz * User may have specified a large count value which caused the 3324fd875dcaSMike Kravetz * above calculation to overflow. In this case, they wanted 3325fd875dcaSMike Kravetz * to allocate as many huge pages as possible. Set count to 3326fd875dcaSMike Kravetz * largest possible value to align with their intention. 3327fd875dcaSMike Kravetz */ 3328fd875dcaSMike Kravetz if (count < old_count) 3329fd875dcaSMike Kravetz count = ULONG_MAX; 3330fd875dcaSMike Kravetz } 3331fd875dcaSMike Kravetz 3332fd875dcaSMike Kravetz /* 33334eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 33344eb0716eSAlexandre Ghiti * page range allocation. 33354eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 33364eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 33374eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 33384eb0716eSAlexandre Ghiti */ 33394eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 33404eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 3341db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 334229383967SMike Kravetz mutex_unlock(&h->resize_lock); 3343f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 33444eb0716eSAlexandre Ghiti return -EINVAL; 33454eb0716eSAlexandre Ghiti } 33464eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 33474eb0716eSAlexandre Ghiti } 3348aa888a74SAndi Kleen 33497893d1d5SAdam Litke /* 33507893d1d5SAdam Litke * Increase the pool size 33517893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 33527893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 3353d1c3fb1fSNishanth Aravamudan * 33540c397daeSMichal Hocko * We might race with alloc_surplus_huge_page() here and be unable 3355d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 3356d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 3357d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 3358d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 33597893d1d5SAdam Litke */ 3360a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 33616ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 33627893d1d5SAdam Litke break; 33637893d1d5SAdam Litke } 33647893d1d5SAdam Litke 3365a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 33667893d1d5SAdam Litke /* 33677893d1d5SAdam Litke * If this allocation races such that we no longer need the 33687893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 33697893d1d5SAdam Litke * and reducing the surplus. 33707893d1d5SAdam Litke */ 3371db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3372649920c6SJia He 3373649920c6SJia He /* yield cpu to avoid soft lockup */ 3374649920c6SJia He cond_resched(); 3375649920c6SJia He 3376f60858f9SMike Kravetz ret = alloc_pool_huge_page(h, nodes_allowed, 3377f60858f9SMike Kravetz node_alloc_noretry); 3378db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 33797893d1d5SAdam Litke if (!ret) 33807893d1d5SAdam Litke goto out; 33817893d1d5SAdam Litke 3382536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 3383536240f2SMel Gorman if (signal_pending(current)) 3384536240f2SMel Gorman goto out; 33857893d1d5SAdam Litke } 33867893d1d5SAdam Litke 33877893d1d5SAdam Litke /* 33887893d1d5SAdam Litke * Decrease the pool size 33897893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 33907893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 33917893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 33927893d1d5SAdam Litke * to the desired size as pages become free. 3393d1c3fb1fSNishanth Aravamudan * 3394d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 3395d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 3396d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 33970c397daeSMichal Hocko * alloc_surplus_huge_page() is checking the global counter, 3398d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 3399d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 3400d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 34017893d1d5SAdam Litke */ 3402a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 34036b0c880dSAdam Litke min_count = max(count, min_count); 34046ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 340510c6ec49SMike Kravetz 340610c6ec49SMike Kravetz /* 340710c6ec49SMike Kravetz * Collect pages to be removed on list without dropping lock 340810c6ec49SMike Kravetz */ 3409a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 341010c6ec49SMike Kravetz page = remove_pool_huge_page(h, nodes_allowed, 0); 341110c6ec49SMike Kravetz if (!page) 34121da177e4SLinus Torvalds break; 341310c6ec49SMike Kravetz 341410c6ec49SMike Kravetz list_add(&page->lru, &page_list); 34151da177e4SLinus Torvalds } 341610c6ec49SMike Kravetz /* free the pages after dropping lock */ 3417db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 341810c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3419b65d4adbSMuchun Song flush_free_hpage_work(h); 3420db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 342110c6ec49SMike Kravetz 3422a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 34236ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 34247893d1d5SAdam Litke break; 34257893d1d5SAdam Litke } 34267893d1d5SAdam Litke out: 34274eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 3428db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 342929383967SMike Kravetz mutex_unlock(&h->resize_lock); 34304eb0716eSAlexandre Ghiti 3431f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 3432f60858f9SMike Kravetz 34334eb0716eSAlexandre Ghiti return 0; 34341da177e4SLinus Torvalds } 34351da177e4SLinus Torvalds 34368531fc6fSMike Kravetz static int demote_free_huge_page(struct hstate *h, struct page *page) 34378531fc6fSMike Kravetz { 34388531fc6fSMike Kravetz int i, nid = page_to_nid(page); 34398531fc6fSMike Kravetz struct hstate *target_hstate; 344031731452SDoug Berger struct page *subpage; 34418531fc6fSMike Kravetz int rc = 0; 34428531fc6fSMike Kravetz 34438531fc6fSMike Kravetz target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order); 34448531fc6fSMike Kravetz 34458531fc6fSMike Kravetz remove_hugetlb_page_for_demote(h, page, false); 34468531fc6fSMike Kravetz spin_unlock_irq(&hugetlb_lock); 34478531fc6fSMike Kravetz 34486213834cSMuchun Song rc = hugetlb_vmemmap_restore(h, page); 34498531fc6fSMike Kravetz if (rc) { 34508531fc6fSMike Kravetz /* Allocation of vmemmmap failed, we can not demote page */ 34518531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 34528531fc6fSMike Kravetz set_page_refcounted(page); 34538531fc6fSMike Kravetz add_hugetlb_page(h, page, false); 34548531fc6fSMike Kravetz return rc; 34558531fc6fSMike Kravetz } 34568531fc6fSMike Kravetz 34578531fc6fSMike Kravetz /* 34588531fc6fSMike Kravetz * Use destroy_compound_hugetlb_page_for_demote for all huge page 34598531fc6fSMike Kravetz * sizes as it will not ref count pages. 34608531fc6fSMike Kravetz */ 34618531fc6fSMike Kravetz destroy_compound_hugetlb_page_for_demote(page, huge_page_order(h)); 34628531fc6fSMike Kravetz 34638531fc6fSMike Kravetz /* 34648531fc6fSMike Kravetz * Taking target hstate mutex synchronizes with set_max_huge_pages. 34658531fc6fSMike Kravetz * Without the mutex, pages added to target hstate could be marked 34668531fc6fSMike Kravetz * as surplus. 34678531fc6fSMike Kravetz * 34688531fc6fSMike Kravetz * Note that we already hold h->resize_lock. To prevent deadlock, 34698531fc6fSMike Kravetz * use the convention of always taking larger size hstate mutex first. 34708531fc6fSMike Kravetz */ 34718531fc6fSMike Kravetz mutex_lock(&target_hstate->resize_lock); 34728531fc6fSMike Kravetz for (i = 0; i < pages_per_huge_page(h); 34738531fc6fSMike Kravetz i += pages_per_huge_page(target_hstate)) { 347431731452SDoug Berger subpage = nth_page(page, i); 34758531fc6fSMike Kravetz if (hstate_is_gigantic(target_hstate)) 347631731452SDoug Berger prep_compound_gigantic_page_for_demote(subpage, 34778531fc6fSMike Kravetz target_hstate->order); 34788531fc6fSMike Kravetz else 347931731452SDoug Berger prep_compound_page(subpage, target_hstate->order); 348031731452SDoug Berger set_page_private(subpage, 0); 348131731452SDoug Berger set_page_refcounted(subpage); 348231731452SDoug Berger prep_new_huge_page(target_hstate, subpage, nid); 348331731452SDoug Berger put_page(subpage); 34848531fc6fSMike Kravetz } 34858531fc6fSMike Kravetz mutex_unlock(&target_hstate->resize_lock); 34868531fc6fSMike Kravetz 34878531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 34888531fc6fSMike Kravetz 34898531fc6fSMike Kravetz /* 34908531fc6fSMike Kravetz * Not absolutely necessary, but for consistency update max_huge_pages 34918531fc6fSMike Kravetz * based on pool changes for the demoted page. 34928531fc6fSMike Kravetz */ 34938531fc6fSMike Kravetz h->max_huge_pages--; 3494a43a83c7SMiaohe Lin target_hstate->max_huge_pages += 3495a43a83c7SMiaohe Lin pages_per_huge_page(h) / pages_per_huge_page(target_hstate); 34968531fc6fSMike Kravetz 34978531fc6fSMike Kravetz return rc; 34988531fc6fSMike Kravetz } 34998531fc6fSMike Kravetz 350079dfc695SMike Kravetz static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 350179dfc695SMike Kravetz __must_hold(&hugetlb_lock) 350279dfc695SMike Kravetz { 35038531fc6fSMike Kravetz int nr_nodes, node; 35048531fc6fSMike Kravetz struct page *page; 350579dfc695SMike Kravetz 350679dfc695SMike Kravetz lockdep_assert_held(&hugetlb_lock); 350779dfc695SMike Kravetz 350879dfc695SMike Kravetz /* We should never get here if no demote order */ 350979dfc695SMike Kravetz if (!h->demote_order) { 351079dfc695SMike Kravetz pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n"); 351179dfc695SMike Kravetz return -EINVAL; /* internal error */ 351279dfc695SMike Kravetz } 351379dfc695SMike Kravetz 35148531fc6fSMike Kravetz for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 35155a317412SMike Kravetz list_for_each_entry(page, &h->hugepage_freelists[node], lru) { 35165a317412SMike Kravetz if (PageHWPoison(page)) 35175a317412SMike Kravetz continue; 35185a317412SMike Kravetz 35195a317412SMike Kravetz return demote_free_huge_page(h, page); 35208531fc6fSMike Kravetz } 35218531fc6fSMike Kravetz } 35228531fc6fSMike Kravetz 35235a317412SMike Kravetz /* 35245a317412SMike Kravetz * Only way to get here is if all pages on free lists are poisoned. 35255a317412SMike Kravetz * Return -EBUSY so that caller will not retry. 35265a317412SMike Kravetz */ 35275a317412SMike Kravetz return -EBUSY; 352879dfc695SMike Kravetz } 352979dfc695SMike Kravetz 3530a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 3531a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 3532a3437870SNishanth Aravamudan 353379dfc695SMike Kravetz #define HSTATE_ATTR_WO(_name) \ 353479dfc695SMike Kravetz static struct kobj_attribute _name##_attr = __ATTR_WO(_name) 353579dfc695SMike Kravetz 3536a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 353798bc26acSMiaohe Lin static struct kobj_attribute _name##_attr = __ATTR_RW(_name) 3538a3437870SNishanth Aravamudan 3539a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 3540a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 3541a3437870SNishanth Aravamudan 35429a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 35439a305230SLee Schermerhorn 35449a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 3545a3437870SNishanth Aravamudan { 3546a3437870SNishanth Aravamudan int i; 35479a305230SLee Schermerhorn 3548a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 35499a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 35509a305230SLee Schermerhorn if (nidp) 35519a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 3552a3437870SNishanth Aravamudan return &hstates[i]; 35539a305230SLee Schermerhorn } 35549a305230SLee Schermerhorn 35559a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 3556a3437870SNishanth Aravamudan } 3557a3437870SNishanth Aravamudan 355806808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 3559a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3560a3437870SNishanth Aravamudan { 35619a305230SLee Schermerhorn struct hstate *h; 35629a305230SLee Schermerhorn unsigned long nr_huge_pages; 35639a305230SLee Schermerhorn int nid; 35649a305230SLee Schermerhorn 35659a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 35669a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 35679a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 35689a305230SLee Schermerhorn else 35699a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 35709a305230SLee Schermerhorn 3571ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", nr_huge_pages); 3572a3437870SNishanth Aravamudan } 3573adbe8726SEric B Munson 3574238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 3575238d3c13SDavid Rientjes struct hstate *h, int nid, 3576238d3c13SDavid Rientjes unsigned long count, size_t len) 3577a3437870SNishanth Aravamudan { 3578a3437870SNishanth Aravamudan int err; 35792d0adf7eSOscar Salvador nodemask_t nodes_allowed, *n_mask; 3580a3437870SNishanth Aravamudan 35812d0adf7eSOscar Salvador if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 35822d0adf7eSOscar Salvador return -EINVAL; 3583adbe8726SEric B Munson 35849a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 35859a305230SLee Schermerhorn /* 35869a305230SLee Schermerhorn * global hstate attribute 35879a305230SLee Schermerhorn */ 35889a305230SLee Schermerhorn if (!(obey_mempolicy && 35892d0adf7eSOscar Salvador init_nodemask_of_mempolicy(&nodes_allowed))) 35902d0adf7eSOscar Salvador n_mask = &node_states[N_MEMORY]; 35912d0adf7eSOscar Salvador else 35922d0adf7eSOscar Salvador n_mask = &nodes_allowed; 35932d0adf7eSOscar Salvador } else { 35949a305230SLee Schermerhorn /* 3595fd875dcaSMike Kravetz * Node specific request. count adjustment happens in 3596fd875dcaSMike Kravetz * set_max_huge_pages() after acquiring hugetlb_lock. 35979a305230SLee Schermerhorn */ 35982d0adf7eSOscar Salvador init_nodemask_of_node(&nodes_allowed, nid); 35992d0adf7eSOscar Salvador n_mask = &nodes_allowed; 3600fd875dcaSMike Kravetz } 36019a305230SLee Schermerhorn 36022d0adf7eSOscar Salvador err = set_max_huge_pages(h, count, nid, n_mask); 360306808b08SLee Schermerhorn 36044eb0716eSAlexandre Ghiti return err ? err : len; 360506808b08SLee Schermerhorn } 360606808b08SLee Schermerhorn 3607238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 3608238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 3609238d3c13SDavid Rientjes size_t len) 3610238d3c13SDavid Rientjes { 3611238d3c13SDavid Rientjes struct hstate *h; 3612238d3c13SDavid Rientjes unsigned long count; 3613238d3c13SDavid Rientjes int nid; 3614238d3c13SDavid Rientjes int err; 3615238d3c13SDavid Rientjes 3616238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 3617238d3c13SDavid Rientjes if (err) 3618238d3c13SDavid Rientjes return err; 3619238d3c13SDavid Rientjes 3620238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 3621238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 3622238d3c13SDavid Rientjes } 3623238d3c13SDavid Rientjes 362406808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 362506808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 362606808b08SLee Schermerhorn { 362706808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 362806808b08SLee Schermerhorn } 362906808b08SLee Schermerhorn 363006808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 363106808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 363206808b08SLee Schermerhorn { 3633238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 3634a3437870SNishanth Aravamudan } 3635a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 3636a3437870SNishanth Aravamudan 363706808b08SLee Schermerhorn #ifdef CONFIG_NUMA 363806808b08SLee Schermerhorn 363906808b08SLee Schermerhorn /* 364006808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 364106808b08SLee Schermerhorn * huge page alloc/free. 364206808b08SLee Schermerhorn */ 364306808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 3644ae7a927dSJoe Perches struct kobj_attribute *attr, 3645ae7a927dSJoe Perches char *buf) 364606808b08SLee Schermerhorn { 364706808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 364806808b08SLee Schermerhorn } 364906808b08SLee Schermerhorn 365006808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 365106808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 365206808b08SLee Schermerhorn { 3653238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 365406808b08SLee Schermerhorn } 365506808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 365606808b08SLee Schermerhorn #endif 365706808b08SLee Schermerhorn 365806808b08SLee Schermerhorn 3659a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 3660a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3661a3437870SNishanth Aravamudan { 36629a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3663ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); 3664a3437870SNishanth Aravamudan } 3665adbe8726SEric B Munson 3666a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 3667a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 3668a3437870SNishanth Aravamudan { 3669a3437870SNishanth Aravamudan int err; 3670a3437870SNishanth Aravamudan unsigned long input; 36719a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3672a3437870SNishanth Aravamudan 3673bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3674adbe8726SEric B Munson return -EINVAL; 3675adbe8726SEric B Munson 36763dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 3677a3437870SNishanth Aravamudan if (err) 367873ae31e5SEric B Munson return err; 3679a3437870SNishanth Aravamudan 3680db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3681a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 3682db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3683a3437870SNishanth Aravamudan 3684a3437870SNishanth Aravamudan return count; 3685a3437870SNishanth Aravamudan } 3686a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 3687a3437870SNishanth Aravamudan 3688a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 3689a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3690a3437870SNishanth Aravamudan { 36919a305230SLee Schermerhorn struct hstate *h; 36929a305230SLee Schermerhorn unsigned long free_huge_pages; 36939a305230SLee Schermerhorn int nid; 36949a305230SLee Schermerhorn 36959a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 36969a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 36979a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 36989a305230SLee Schermerhorn else 36999a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 37009a305230SLee Schermerhorn 3701ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", free_huge_pages); 3702a3437870SNishanth Aravamudan } 3703a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 3704a3437870SNishanth Aravamudan 3705a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 3706a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3707a3437870SNishanth Aravamudan { 37089a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3709ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); 3710a3437870SNishanth Aravamudan } 3711a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 3712a3437870SNishanth Aravamudan 3713a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 3714a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3715a3437870SNishanth Aravamudan { 37169a305230SLee Schermerhorn struct hstate *h; 37179a305230SLee Schermerhorn unsigned long surplus_huge_pages; 37189a305230SLee Schermerhorn int nid; 37199a305230SLee Schermerhorn 37209a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 37219a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 37229a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 37239a305230SLee Schermerhorn else 37249a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 37259a305230SLee Schermerhorn 3726ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", surplus_huge_pages); 3727a3437870SNishanth Aravamudan } 3728a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 3729a3437870SNishanth Aravamudan 373079dfc695SMike Kravetz static ssize_t demote_store(struct kobject *kobj, 373179dfc695SMike Kravetz struct kobj_attribute *attr, const char *buf, size_t len) 373279dfc695SMike Kravetz { 373379dfc695SMike Kravetz unsigned long nr_demote; 373479dfc695SMike Kravetz unsigned long nr_available; 373579dfc695SMike Kravetz nodemask_t nodes_allowed, *n_mask; 373679dfc695SMike Kravetz struct hstate *h; 37378eeda55fSLi zeming int err; 373879dfc695SMike Kravetz int nid; 373979dfc695SMike Kravetz 374079dfc695SMike Kravetz err = kstrtoul(buf, 10, &nr_demote); 374179dfc695SMike Kravetz if (err) 374279dfc695SMike Kravetz return err; 374379dfc695SMike Kravetz h = kobj_to_hstate(kobj, &nid); 374479dfc695SMike Kravetz 374579dfc695SMike Kravetz if (nid != NUMA_NO_NODE) { 374679dfc695SMike Kravetz init_nodemask_of_node(&nodes_allowed, nid); 374779dfc695SMike Kravetz n_mask = &nodes_allowed; 374879dfc695SMike Kravetz } else { 374979dfc695SMike Kravetz n_mask = &node_states[N_MEMORY]; 375079dfc695SMike Kravetz } 375179dfc695SMike Kravetz 375279dfc695SMike Kravetz /* Synchronize with other sysfs operations modifying huge pages */ 375379dfc695SMike Kravetz mutex_lock(&h->resize_lock); 375479dfc695SMike Kravetz spin_lock_irq(&hugetlb_lock); 375579dfc695SMike Kravetz 375679dfc695SMike Kravetz while (nr_demote) { 375779dfc695SMike Kravetz /* 375879dfc695SMike Kravetz * Check for available pages to demote each time thorough the 375979dfc695SMike Kravetz * loop as demote_pool_huge_page will drop hugetlb_lock. 376079dfc695SMike Kravetz */ 376179dfc695SMike Kravetz if (nid != NUMA_NO_NODE) 376279dfc695SMike Kravetz nr_available = h->free_huge_pages_node[nid]; 376379dfc695SMike Kravetz else 376479dfc695SMike Kravetz nr_available = h->free_huge_pages; 376579dfc695SMike Kravetz nr_available -= h->resv_huge_pages; 376679dfc695SMike Kravetz if (!nr_available) 376779dfc695SMike Kravetz break; 376879dfc695SMike Kravetz 376979dfc695SMike Kravetz err = demote_pool_huge_page(h, n_mask); 377079dfc695SMike Kravetz if (err) 377179dfc695SMike Kravetz break; 377279dfc695SMike Kravetz 377379dfc695SMike Kravetz nr_demote--; 377479dfc695SMike Kravetz } 377579dfc695SMike Kravetz 377679dfc695SMike Kravetz spin_unlock_irq(&hugetlb_lock); 377779dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 377879dfc695SMike Kravetz 377979dfc695SMike Kravetz if (err) 378079dfc695SMike Kravetz return err; 378179dfc695SMike Kravetz return len; 378279dfc695SMike Kravetz } 378379dfc695SMike Kravetz HSTATE_ATTR_WO(demote); 378479dfc695SMike Kravetz 378579dfc695SMike Kravetz static ssize_t demote_size_show(struct kobject *kobj, 378679dfc695SMike Kravetz struct kobj_attribute *attr, char *buf) 378779dfc695SMike Kravetz { 378812658abfSMiaohe Lin struct hstate *h = kobj_to_hstate(kobj, NULL); 378979dfc695SMike Kravetz unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K; 379079dfc695SMike Kravetz 379179dfc695SMike Kravetz return sysfs_emit(buf, "%lukB\n", demote_size); 379279dfc695SMike Kravetz } 379379dfc695SMike Kravetz 379479dfc695SMike Kravetz static ssize_t demote_size_store(struct kobject *kobj, 379579dfc695SMike Kravetz struct kobj_attribute *attr, 379679dfc695SMike Kravetz const char *buf, size_t count) 379779dfc695SMike Kravetz { 379879dfc695SMike Kravetz struct hstate *h, *demote_hstate; 379979dfc695SMike Kravetz unsigned long demote_size; 380079dfc695SMike Kravetz unsigned int demote_order; 380179dfc695SMike Kravetz 380279dfc695SMike Kravetz demote_size = (unsigned long)memparse(buf, NULL); 380379dfc695SMike Kravetz 380479dfc695SMike Kravetz demote_hstate = size_to_hstate(demote_size); 380579dfc695SMike Kravetz if (!demote_hstate) 380679dfc695SMike Kravetz return -EINVAL; 380779dfc695SMike Kravetz demote_order = demote_hstate->order; 3808a01f4390SMike Kravetz if (demote_order < HUGETLB_PAGE_ORDER) 3809a01f4390SMike Kravetz return -EINVAL; 381079dfc695SMike Kravetz 381179dfc695SMike Kravetz /* demote order must be smaller than hstate order */ 381212658abfSMiaohe Lin h = kobj_to_hstate(kobj, NULL); 381379dfc695SMike Kravetz if (demote_order >= h->order) 381479dfc695SMike Kravetz return -EINVAL; 381579dfc695SMike Kravetz 381679dfc695SMike Kravetz /* resize_lock synchronizes access to demote size and writes */ 381779dfc695SMike Kravetz mutex_lock(&h->resize_lock); 381879dfc695SMike Kravetz h->demote_order = demote_order; 381979dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 382079dfc695SMike Kravetz 382179dfc695SMike Kravetz return count; 382279dfc695SMike Kravetz } 382379dfc695SMike Kravetz HSTATE_ATTR(demote_size); 382479dfc695SMike Kravetz 3825a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 3826a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 3827a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 3828a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 3829a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 3830a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 383106808b08SLee Schermerhorn #ifdef CONFIG_NUMA 383206808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 383306808b08SLee Schermerhorn #endif 3834a3437870SNishanth Aravamudan NULL, 3835a3437870SNishanth Aravamudan }; 3836a3437870SNishanth Aravamudan 383767e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 3838a3437870SNishanth Aravamudan .attrs = hstate_attrs, 3839a3437870SNishanth Aravamudan }; 3840a3437870SNishanth Aravamudan 384179dfc695SMike Kravetz static struct attribute *hstate_demote_attrs[] = { 384279dfc695SMike Kravetz &demote_size_attr.attr, 384379dfc695SMike Kravetz &demote_attr.attr, 384479dfc695SMike Kravetz NULL, 384579dfc695SMike Kravetz }; 384679dfc695SMike Kravetz 384779dfc695SMike Kravetz static const struct attribute_group hstate_demote_attr_group = { 384879dfc695SMike Kravetz .attrs = hstate_demote_attrs, 384979dfc695SMike Kravetz }; 385079dfc695SMike Kravetz 3851094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 38529a305230SLee Schermerhorn struct kobject **hstate_kobjs, 385367e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 3854a3437870SNishanth Aravamudan { 3855a3437870SNishanth Aravamudan int retval; 3856972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 3857a3437870SNishanth Aravamudan 38589a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 38599a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 3860a3437870SNishanth Aravamudan return -ENOMEM; 3861a3437870SNishanth Aravamudan 38629a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 3863cc2205a6SMiaohe Lin if (retval) { 38649a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 3865cc2205a6SMiaohe Lin hstate_kobjs[hi] = NULL; 38663a6bdda0SMiaohe Lin return retval; 3867cc2205a6SMiaohe Lin } 3868a3437870SNishanth Aravamudan 386979dfc695SMike Kravetz if (h->demote_order) { 387001088a60SMiaohe Lin retval = sysfs_create_group(hstate_kobjs[hi], 387101088a60SMiaohe Lin &hstate_demote_attr_group); 387201088a60SMiaohe Lin if (retval) { 387379dfc695SMike Kravetz pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name); 387401088a60SMiaohe Lin sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group); 387501088a60SMiaohe Lin kobject_put(hstate_kobjs[hi]); 387601088a60SMiaohe Lin hstate_kobjs[hi] = NULL; 387701088a60SMiaohe Lin return retval; 387801088a60SMiaohe Lin } 387979dfc695SMike Kravetz } 388079dfc695SMike Kravetz 388101088a60SMiaohe Lin return 0; 3882a3437870SNishanth Aravamudan } 3883a3437870SNishanth Aravamudan 38849a305230SLee Schermerhorn #ifdef CONFIG_NUMA 3885a4a00b45SMuchun Song static bool hugetlb_sysfs_initialized __ro_after_init; 38869a305230SLee Schermerhorn 38879a305230SLee Schermerhorn /* 38889a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 388910fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 389010fbcf4cSKay Sievers * index of a node device or _hstate == node id. 389110fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 38929a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 38939a305230SLee Schermerhorn */ 38949a305230SLee Schermerhorn struct node_hstate { 38959a305230SLee Schermerhorn struct kobject *hugepages_kobj; 38969a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 38979a305230SLee Schermerhorn }; 3898b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 38999a305230SLee Schermerhorn 39009a305230SLee Schermerhorn /* 390110fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 39029a305230SLee Schermerhorn */ 39039a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 39049a305230SLee Schermerhorn &nr_hugepages_attr.attr, 39059a305230SLee Schermerhorn &free_hugepages_attr.attr, 39069a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 39079a305230SLee Schermerhorn NULL, 39089a305230SLee Schermerhorn }; 39099a305230SLee Schermerhorn 391067e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 39119a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 39129a305230SLee Schermerhorn }; 39139a305230SLee Schermerhorn 39149a305230SLee Schermerhorn /* 391510fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 39169a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 39179a305230SLee Schermerhorn */ 39189a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 39199a305230SLee Schermerhorn { 39209a305230SLee Schermerhorn int nid; 39219a305230SLee Schermerhorn 39229a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 39239a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 39249a305230SLee Schermerhorn int i; 39259a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 39269a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 39279a305230SLee Schermerhorn if (nidp) 39289a305230SLee Schermerhorn *nidp = nid; 39299a305230SLee Schermerhorn return &hstates[i]; 39309a305230SLee Schermerhorn } 39319a305230SLee Schermerhorn } 39329a305230SLee Schermerhorn 39339a305230SLee Schermerhorn BUG(); 39349a305230SLee Schermerhorn return NULL; 39359a305230SLee Schermerhorn } 39369a305230SLee Schermerhorn 39379a305230SLee Schermerhorn /* 393810fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 39399a305230SLee Schermerhorn * No-op if no hstate attributes attached. 39409a305230SLee Schermerhorn */ 3941a4a00b45SMuchun Song void hugetlb_unregister_node(struct node *node) 39429a305230SLee Schermerhorn { 39439a305230SLee Schermerhorn struct hstate *h; 394410fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 39459a305230SLee Schermerhorn 39469a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 39479b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 39489a305230SLee Schermerhorn 3949972dc4deSAneesh Kumar K.V for_each_hstate(h) { 3950972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 395101088a60SMiaohe Lin struct kobject *hstate_kobj = nhs->hstate_kobjs[idx]; 395201088a60SMiaohe Lin 395301088a60SMiaohe Lin if (!hstate_kobj) 395401088a60SMiaohe Lin continue; 395501088a60SMiaohe Lin if (h->demote_order) 395601088a60SMiaohe Lin sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group); 395701088a60SMiaohe Lin sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group); 395801088a60SMiaohe Lin kobject_put(hstate_kobj); 3959972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 3960972dc4deSAneesh Kumar K.V } 39619a305230SLee Schermerhorn 39629a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 39639a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 39649a305230SLee Schermerhorn } 39659a305230SLee Schermerhorn 39669a305230SLee Schermerhorn 39679a305230SLee Schermerhorn /* 396810fbcf4cSKay Sievers * Register hstate attributes for a single node device. 39699a305230SLee Schermerhorn * No-op if attributes already registered. 39709a305230SLee Schermerhorn */ 3971a4a00b45SMuchun Song void hugetlb_register_node(struct node *node) 39729a305230SLee Schermerhorn { 39739a305230SLee Schermerhorn struct hstate *h; 397410fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 39759a305230SLee Schermerhorn int err; 39769a305230SLee Schermerhorn 3977a4a00b45SMuchun Song if (!hugetlb_sysfs_initialized) 3978a4a00b45SMuchun Song return; 3979a4a00b45SMuchun Song 39809a305230SLee Schermerhorn if (nhs->hugepages_kobj) 39819a305230SLee Schermerhorn return; /* already allocated */ 39829a305230SLee Schermerhorn 39839a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 398410fbcf4cSKay Sievers &node->dev.kobj); 39859a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 39869a305230SLee Schermerhorn return; 39879a305230SLee Schermerhorn 39889a305230SLee Schermerhorn for_each_hstate(h) { 39899a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 39909a305230SLee Schermerhorn nhs->hstate_kobjs, 39919a305230SLee Schermerhorn &per_node_hstate_attr_group); 39929a305230SLee Schermerhorn if (err) { 3993282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s for node %d\n", 399410fbcf4cSKay Sievers h->name, node->dev.id); 39959a305230SLee Schermerhorn hugetlb_unregister_node(node); 39969a305230SLee Schermerhorn break; 39979a305230SLee Schermerhorn } 39989a305230SLee Schermerhorn } 39999a305230SLee Schermerhorn } 40009a305230SLee Schermerhorn 40019a305230SLee Schermerhorn /* 40029b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 400310fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 400410fbcf4cSKay Sievers * registered their associated device by this time. 40059a305230SLee Schermerhorn */ 40067d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 40079a305230SLee Schermerhorn { 40089a305230SLee Schermerhorn int nid; 40099a305230SLee Schermerhorn 4010a4a00b45SMuchun Song for_each_online_node(nid) 4011b958d4d0SMuchun Song hugetlb_register_node(node_devices[nid]); 40129a305230SLee Schermerhorn } 40139a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 40149a305230SLee Schermerhorn 40159a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 40169a305230SLee Schermerhorn { 40179a305230SLee Schermerhorn BUG(); 40189a305230SLee Schermerhorn if (nidp) 40199a305230SLee Schermerhorn *nidp = -1; 40209a305230SLee Schermerhorn return NULL; 40219a305230SLee Schermerhorn } 40229a305230SLee Schermerhorn 40239a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 40249a305230SLee Schermerhorn 40259a305230SLee Schermerhorn #endif 40269a305230SLee Schermerhorn 4027263b8998SMiaohe Lin #ifdef CONFIG_CMA 4028263b8998SMiaohe Lin static void __init hugetlb_cma_check(void); 4029263b8998SMiaohe Lin #else 4030263b8998SMiaohe Lin static inline __init void hugetlb_cma_check(void) 4031263b8998SMiaohe Lin { 4032263b8998SMiaohe Lin } 4033263b8998SMiaohe Lin #endif 4034263b8998SMiaohe Lin 4035a4a00b45SMuchun Song static void __init hugetlb_sysfs_init(void) 4036a4a00b45SMuchun Song { 4037a4a00b45SMuchun Song struct hstate *h; 4038a4a00b45SMuchun Song int err; 4039a4a00b45SMuchun Song 4040a4a00b45SMuchun Song hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 4041a4a00b45SMuchun Song if (!hugepages_kobj) 4042a4a00b45SMuchun Song return; 4043a4a00b45SMuchun Song 4044a4a00b45SMuchun Song for_each_hstate(h) { 4045a4a00b45SMuchun Song err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 4046a4a00b45SMuchun Song hstate_kobjs, &hstate_attr_group); 4047a4a00b45SMuchun Song if (err) 4048a4a00b45SMuchun Song pr_err("HugeTLB: Unable to add hstate %s", h->name); 4049a4a00b45SMuchun Song } 4050a4a00b45SMuchun Song 4051a4a00b45SMuchun Song #ifdef CONFIG_NUMA 4052a4a00b45SMuchun Song hugetlb_sysfs_initialized = true; 4053a4a00b45SMuchun Song #endif 4054a4a00b45SMuchun Song hugetlb_register_all_nodes(); 4055a4a00b45SMuchun Song } 4056a4a00b45SMuchun Song 4057a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 4058a3437870SNishanth Aravamudan { 40598382d914SDavidlohr Bueso int i; 40608382d914SDavidlohr Bueso 4061d6995da3SMike Kravetz BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < 4062d6995da3SMike Kravetz __NR_HPAGEFLAGS); 4063d6995da3SMike Kravetz 4064c2833a5bSMike Kravetz if (!hugepages_supported()) { 4065c2833a5bSMike Kravetz if (hugetlb_max_hstate || default_hstate_max_huge_pages) 4066c2833a5bSMike Kravetz pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); 40670ef89d25SBenjamin Herrenschmidt return 0; 4068d715cf80SLiam R. Howlett } 4069d715cf80SLiam R. Howlett 4070282f4214SMike Kravetz /* 4071282f4214SMike Kravetz * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some 4072282f4214SMike Kravetz * architectures depend on setup being done here. 4073282f4214SMike Kravetz */ 4074a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 4075282f4214SMike Kravetz if (!parsed_default_hugepagesz) { 4076282f4214SMike Kravetz /* 4077282f4214SMike Kravetz * If we did not parse a default huge page size, set 4078282f4214SMike Kravetz * default_hstate_idx to HPAGE_SIZE hstate. And, if the 4079282f4214SMike Kravetz * number of huge pages for this default size was implicitly 4080282f4214SMike Kravetz * specified, set that here as well. 4081282f4214SMike Kravetz * Note that the implicit setting will overwrite an explicit 4082282f4214SMike Kravetz * setting. A warning will be printed in this case. 4083282f4214SMike Kravetz */ 4084282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); 4085f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 4086282f4214SMike Kravetz if (default_hstate.max_huge_pages) { 4087282f4214SMike Kravetz char buf[32]; 4088282f4214SMike Kravetz 4089282f4214SMike Kravetz string_get_size(huge_page_size(&default_hstate), 4090282f4214SMike Kravetz 1, STRING_UNITS_2, buf, 32); 4091282f4214SMike Kravetz pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", 4092282f4214SMike Kravetz default_hstate.max_huge_pages, buf); 4093282f4214SMike Kravetz pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", 4094282f4214SMike Kravetz default_hstate_max_huge_pages); 4095282f4214SMike Kravetz } 4096282f4214SMike Kravetz default_hstate.max_huge_pages = 4097282f4214SMike Kravetz default_hstate_max_huge_pages; 4098b5389086SZhenguo Yao 40990a7a0f6fSPeng Liu for_each_online_node(i) 4100b5389086SZhenguo Yao default_hstate.max_huge_pages_node[i] = 4101b5389086SZhenguo Yao default_hugepages_in_node[i]; 4102282f4214SMike Kravetz } 4103f8b74815SVaishali Thakkar } 4104a3437870SNishanth Aravamudan 4105cf11e85fSRoman Gushchin hugetlb_cma_check(); 4106a3437870SNishanth Aravamudan hugetlb_init_hstates(); 4107aa888a74SAndi Kleen gather_bootmem_prealloc(); 4108a3437870SNishanth Aravamudan report_hugepages(); 4109a3437870SNishanth Aravamudan 4110a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 41117179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 41129a305230SLee Schermerhorn 41138382d914SDavidlohr Bueso #ifdef CONFIG_SMP 41148382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 41158382d914SDavidlohr Bueso #else 41168382d914SDavidlohr Bueso num_fault_mutexes = 1; 41178382d914SDavidlohr Bueso #endif 4118c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 41196da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 41206da2ec56SKees Cook GFP_KERNEL); 4121c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 41228382d914SDavidlohr Bueso 41238382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 4124c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 4125a3437870SNishanth Aravamudan return 0; 4126a3437870SNishanth Aravamudan } 41273e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 4128a3437870SNishanth Aravamudan 4129ae94da89SMike Kravetz /* Overwritten by architectures with more huge page sizes */ 4130ae94da89SMike Kravetz bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) 41319fee021dSVaishali Thakkar { 4132ae94da89SMike Kravetz return size == HPAGE_SIZE; 41339fee021dSVaishali Thakkar } 41349fee021dSVaishali Thakkar 4135d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 4136a3437870SNishanth Aravamudan { 4137a3437870SNishanth Aravamudan struct hstate *h; 41388faa8b07SAndi Kleen unsigned long i; 41398faa8b07SAndi Kleen 4140a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 4141a3437870SNishanth Aravamudan return; 4142a3437870SNishanth Aravamudan } 414347d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 4144a3437870SNishanth Aravamudan BUG_ON(order == 0); 414547d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 414629383967SMike Kravetz mutex_init(&h->resize_lock); 4147a3437870SNishanth Aravamudan h->order = order; 4148aca78307SMiaohe Lin h->mask = ~(huge_page_size(h) - 1); 41498faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 41508faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 41510edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 415254f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 415354f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 4154a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 4155c2c3a60aSMiaohe Lin huge_page_size(h)/SZ_1K); 41568faa8b07SAndi Kleen 4157a3437870SNishanth Aravamudan parsed_hstate = h; 4158a3437870SNishanth Aravamudan } 4159a3437870SNishanth Aravamudan 4160b5389086SZhenguo Yao bool __init __weak hugetlb_node_alloc_supported(void) 4161b5389086SZhenguo Yao { 4162b5389086SZhenguo Yao return true; 4163b5389086SZhenguo Yao } 4164f87442f4SPeng Liu 4165f87442f4SPeng Liu static void __init hugepages_clear_pages_in_node(void) 4166f87442f4SPeng Liu { 4167f87442f4SPeng Liu if (!hugetlb_max_hstate) { 4168f87442f4SPeng Liu default_hstate_max_huge_pages = 0; 4169f87442f4SPeng Liu memset(default_hugepages_in_node, 0, 417010395680SMiaohe Lin sizeof(default_hugepages_in_node)); 4171f87442f4SPeng Liu } else { 4172f87442f4SPeng Liu parsed_hstate->max_huge_pages = 0; 4173f87442f4SPeng Liu memset(parsed_hstate->max_huge_pages_node, 0, 417410395680SMiaohe Lin sizeof(parsed_hstate->max_huge_pages_node)); 4175f87442f4SPeng Liu } 4176f87442f4SPeng Liu } 4177f87442f4SPeng Liu 4178282f4214SMike Kravetz /* 4179282f4214SMike Kravetz * hugepages command line processing 4180282f4214SMike Kravetz * hugepages normally follows a valid hugepagsz or default_hugepagsz 4181282f4214SMike Kravetz * specification. If not, ignore the hugepages value. hugepages can also 4182282f4214SMike Kravetz * be the first huge page command line option in which case it implicitly 4183282f4214SMike Kravetz * specifies the number of huge pages for the default size. 4184282f4214SMike Kravetz */ 4185282f4214SMike Kravetz static int __init hugepages_setup(char *s) 4186a3437870SNishanth Aravamudan { 4187a3437870SNishanth Aravamudan unsigned long *mhp; 41888faa8b07SAndi Kleen static unsigned long *last_mhp; 4189b5389086SZhenguo Yao int node = NUMA_NO_NODE; 4190b5389086SZhenguo Yao int count; 4191b5389086SZhenguo Yao unsigned long tmp; 4192b5389086SZhenguo Yao char *p = s; 4193a3437870SNishanth Aravamudan 41949fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 4195282f4214SMike Kravetz pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); 41969fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 4197f81f6e4bSPeng Liu return 1; 41989fee021dSVaishali Thakkar } 4199282f4214SMike Kravetz 4200a3437870SNishanth Aravamudan /* 4201282f4214SMike Kravetz * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter 4202282f4214SMike Kravetz * yet, so this hugepages= parameter goes to the "default hstate". 4203282f4214SMike Kravetz * Otherwise, it goes with the previously parsed hugepagesz or 4204282f4214SMike Kravetz * default_hugepagesz. 4205a3437870SNishanth Aravamudan */ 42069fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 4207a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 4208a3437870SNishanth Aravamudan else 4209a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 4210a3437870SNishanth Aravamudan 42118faa8b07SAndi Kleen if (mhp == last_mhp) { 4212282f4214SMike Kravetz pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); 4213f81f6e4bSPeng Liu return 1; 42148faa8b07SAndi Kleen } 42158faa8b07SAndi Kleen 4216b5389086SZhenguo Yao while (*p) { 4217b5389086SZhenguo Yao count = 0; 4218b5389086SZhenguo Yao if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4219b5389086SZhenguo Yao goto invalid; 4220b5389086SZhenguo Yao /* Parameter is node format */ 4221b5389086SZhenguo Yao if (p[count] == ':') { 4222b5389086SZhenguo Yao if (!hugetlb_node_alloc_supported()) { 4223b5389086SZhenguo Yao pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n"); 4224f81f6e4bSPeng Liu return 1; 4225b5389086SZhenguo Yao } 42260a7a0f6fSPeng Liu if (tmp >= MAX_NUMNODES || !node_online(tmp)) 4227e79ce983SLiu Yuntao goto invalid; 42280a7a0f6fSPeng Liu node = array_index_nospec(tmp, MAX_NUMNODES); 4229b5389086SZhenguo Yao p += count + 1; 4230b5389086SZhenguo Yao /* Parse hugepages */ 4231b5389086SZhenguo Yao if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4232b5389086SZhenguo Yao goto invalid; 4233b5389086SZhenguo Yao if (!hugetlb_max_hstate) 4234b5389086SZhenguo Yao default_hugepages_in_node[node] = tmp; 4235b5389086SZhenguo Yao else 4236b5389086SZhenguo Yao parsed_hstate->max_huge_pages_node[node] = tmp; 4237b5389086SZhenguo Yao *mhp += tmp; 4238b5389086SZhenguo Yao /* Go to parse next node*/ 4239b5389086SZhenguo Yao if (p[count] == ',') 4240b5389086SZhenguo Yao p += count + 1; 4241b5389086SZhenguo Yao else 4242b5389086SZhenguo Yao break; 4243b5389086SZhenguo Yao } else { 4244b5389086SZhenguo Yao if (p != s) 4245b5389086SZhenguo Yao goto invalid; 4246b5389086SZhenguo Yao *mhp = tmp; 4247b5389086SZhenguo Yao break; 4248b5389086SZhenguo Yao } 4249b5389086SZhenguo Yao } 4250a3437870SNishanth Aravamudan 42518faa8b07SAndi Kleen /* 42528faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 425304adbc3fSMiaohe Lin * But we need to allocate gigantic hstates here early to still 42548faa8b07SAndi Kleen * use the bootmem allocator. 42558faa8b07SAndi Kleen */ 425604adbc3fSMiaohe Lin if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) 42578faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 42588faa8b07SAndi Kleen 42598faa8b07SAndi Kleen last_mhp = mhp; 42608faa8b07SAndi Kleen 4261a3437870SNishanth Aravamudan return 1; 4262b5389086SZhenguo Yao 4263b5389086SZhenguo Yao invalid: 4264b5389086SZhenguo Yao pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p); 4265f87442f4SPeng Liu hugepages_clear_pages_in_node(); 4266f81f6e4bSPeng Liu return 1; 4267a3437870SNishanth Aravamudan } 4268282f4214SMike Kravetz __setup("hugepages=", hugepages_setup); 4269e11bfbfcSNick Piggin 4270282f4214SMike Kravetz /* 4271282f4214SMike Kravetz * hugepagesz command line processing 4272282f4214SMike Kravetz * A specific huge page size can only be specified once with hugepagesz. 4273282f4214SMike Kravetz * hugepagesz is followed by hugepages on the command line. The global 4274282f4214SMike Kravetz * variable 'parsed_valid_hugepagesz' is used to determine if prior 4275282f4214SMike Kravetz * hugepagesz argument was valid. 4276282f4214SMike Kravetz */ 4277359f2544SMike Kravetz static int __init hugepagesz_setup(char *s) 4278e11bfbfcSNick Piggin { 4279359f2544SMike Kravetz unsigned long size; 4280282f4214SMike Kravetz struct hstate *h; 4281282f4214SMike Kravetz 4282282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4283359f2544SMike Kravetz size = (unsigned long)memparse(s, NULL); 4284359f2544SMike Kravetz 4285359f2544SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4286282f4214SMike Kravetz pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); 4287f81f6e4bSPeng Liu return 1; 4288359f2544SMike Kravetz } 4289359f2544SMike Kravetz 4290282f4214SMike Kravetz h = size_to_hstate(size); 4291282f4214SMike Kravetz if (h) { 4292282f4214SMike Kravetz /* 4293282f4214SMike Kravetz * hstate for this size already exists. This is normally 4294282f4214SMike Kravetz * an error, but is allowed if the existing hstate is the 4295282f4214SMike Kravetz * default hstate. More specifically, it is only allowed if 4296282f4214SMike Kravetz * the number of huge pages for the default hstate was not 4297282f4214SMike Kravetz * previously specified. 4298282f4214SMike Kravetz */ 4299282f4214SMike Kravetz if (!parsed_default_hugepagesz || h != &default_hstate || 4300282f4214SMike Kravetz default_hstate.max_huge_pages) { 4301282f4214SMike Kravetz pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); 4302f81f6e4bSPeng Liu return 1; 430338237830SMike Kravetz } 430438237830SMike Kravetz 4305282f4214SMike Kravetz /* 4306282f4214SMike Kravetz * No need to call hugetlb_add_hstate() as hstate already 4307282f4214SMike Kravetz * exists. But, do set parsed_hstate so that a following 4308282f4214SMike Kravetz * hugepages= parameter will be applied to this hstate. 4309282f4214SMike Kravetz */ 4310282f4214SMike Kravetz parsed_hstate = h; 4311282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4312e11bfbfcSNick Piggin return 1; 4313e11bfbfcSNick Piggin } 4314282f4214SMike Kravetz 4315359f2544SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4316282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4317359f2544SMike Kravetz return 1; 4318359f2544SMike Kravetz } 4319359f2544SMike Kravetz __setup("hugepagesz=", hugepagesz_setup); 4320359f2544SMike Kravetz 4321282f4214SMike Kravetz /* 4322282f4214SMike Kravetz * default_hugepagesz command line input 4323282f4214SMike Kravetz * Only one instance of default_hugepagesz allowed on command line. 4324282f4214SMike Kravetz */ 4325ae94da89SMike Kravetz static int __init default_hugepagesz_setup(char *s) 4326e11bfbfcSNick Piggin { 4327ae94da89SMike Kravetz unsigned long size; 4328b5389086SZhenguo Yao int i; 4329ae94da89SMike Kravetz 4330282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4331282f4214SMike Kravetz if (parsed_default_hugepagesz) { 4332282f4214SMike Kravetz pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); 4333f81f6e4bSPeng Liu return 1; 4334282f4214SMike Kravetz } 4335282f4214SMike Kravetz 4336282f4214SMike Kravetz size = (unsigned long)memparse(s, NULL); 4337282f4214SMike Kravetz 4338282f4214SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4339282f4214SMike Kravetz pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); 4340f81f6e4bSPeng Liu return 1; 4341282f4214SMike Kravetz } 4342282f4214SMike Kravetz 4343282f4214SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4344282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4345282f4214SMike Kravetz parsed_default_hugepagesz = true; 4346282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(size)); 4347282f4214SMike Kravetz 4348282f4214SMike Kravetz /* 4349282f4214SMike Kravetz * The number of default huge pages (for this size) could have been 4350282f4214SMike Kravetz * specified as the first hugetlb parameter: hugepages=X. If so, 4351282f4214SMike Kravetz * then default_hstate_max_huge_pages is set. If the default huge 4352282f4214SMike Kravetz * page size is gigantic (>= MAX_ORDER), then the pages must be 4353282f4214SMike Kravetz * allocated here from bootmem allocator. 4354282f4214SMike Kravetz */ 4355282f4214SMike Kravetz if (default_hstate_max_huge_pages) { 4356282f4214SMike Kravetz default_hstate.max_huge_pages = default_hstate_max_huge_pages; 43570a7a0f6fSPeng Liu for_each_online_node(i) 4358b5389086SZhenguo Yao default_hstate.max_huge_pages_node[i] = 4359b5389086SZhenguo Yao default_hugepages_in_node[i]; 4360282f4214SMike Kravetz if (hstate_is_gigantic(&default_hstate)) 4361282f4214SMike Kravetz hugetlb_hstate_alloc_pages(&default_hstate); 4362282f4214SMike Kravetz default_hstate_max_huge_pages = 0; 4363282f4214SMike Kravetz } 4364282f4214SMike Kravetz 4365e11bfbfcSNick Piggin return 1; 4366e11bfbfcSNick Piggin } 4367ae94da89SMike Kravetz __setup("default_hugepagesz=", default_hugepagesz_setup); 4368a3437870SNishanth Aravamudan 4369d2226ebdSFeng Tang static nodemask_t *policy_mbind_nodemask(gfp_t gfp) 4370d2226ebdSFeng Tang { 4371d2226ebdSFeng Tang #ifdef CONFIG_NUMA 4372d2226ebdSFeng Tang struct mempolicy *mpol = get_task_policy(current); 4373d2226ebdSFeng Tang 4374d2226ebdSFeng Tang /* 4375d2226ebdSFeng Tang * Only enforce MPOL_BIND policy which overlaps with cpuset policy 4376d2226ebdSFeng Tang * (from policy_nodemask) specifically for hugetlb case 4377d2226ebdSFeng Tang */ 4378d2226ebdSFeng Tang if (mpol->mode == MPOL_BIND && 4379d2226ebdSFeng Tang (apply_policy_zone(mpol, gfp_zone(gfp)) && 4380d2226ebdSFeng Tang cpuset_nodemask_valid_mems_allowed(&mpol->nodes))) 4381d2226ebdSFeng Tang return &mpol->nodes; 4382d2226ebdSFeng Tang #endif 4383d2226ebdSFeng Tang return NULL; 4384d2226ebdSFeng Tang } 4385d2226ebdSFeng Tang 43868ca39e68SMuchun Song static unsigned int allowed_mems_nr(struct hstate *h) 43878a213460SNishanth Aravamudan { 43888a213460SNishanth Aravamudan int node; 43898a213460SNishanth Aravamudan unsigned int nr = 0; 4390d2226ebdSFeng Tang nodemask_t *mbind_nodemask; 43918ca39e68SMuchun Song unsigned int *array = h->free_huge_pages_node; 43928ca39e68SMuchun Song gfp_t gfp_mask = htlb_alloc_mask(h); 43938a213460SNishanth Aravamudan 4394d2226ebdSFeng Tang mbind_nodemask = policy_mbind_nodemask(gfp_mask); 43958ca39e68SMuchun Song for_each_node_mask(node, cpuset_current_mems_allowed) { 4396d2226ebdSFeng Tang if (!mbind_nodemask || node_isset(node, *mbind_nodemask)) 43978a213460SNishanth Aravamudan nr += array[node]; 43988ca39e68SMuchun Song } 43998a213460SNishanth Aravamudan 44008a213460SNishanth Aravamudan return nr; 44018a213460SNishanth Aravamudan } 44028a213460SNishanth Aravamudan 44038a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 440417743798SMuchun Song static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, 440517743798SMuchun Song void *buffer, size_t *length, 440617743798SMuchun Song loff_t *ppos, unsigned long *out) 440717743798SMuchun Song { 440817743798SMuchun Song struct ctl_table dup_table; 440917743798SMuchun Song 441017743798SMuchun Song /* 441117743798SMuchun Song * In order to avoid races with __do_proc_doulongvec_minmax(), we 441217743798SMuchun Song * can duplicate the @table and alter the duplicate of it. 441317743798SMuchun Song */ 441417743798SMuchun Song dup_table = *table; 441517743798SMuchun Song dup_table.data = out; 441617743798SMuchun Song 441717743798SMuchun Song return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); 441817743798SMuchun Song } 441917743798SMuchun Song 442006808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 442106808b08SLee Schermerhorn struct ctl_table *table, int write, 442232927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 44231da177e4SLinus Torvalds { 4424e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 4425238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 442608d4a246SMichal Hocko int ret; 4427e5ff2159SAndi Kleen 4428457c1b27SNishanth Aravamudan if (!hugepages_supported()) 442986613628SJan Stancek return -EOPNOTSUPP; 4430457c1b27SNishanth Aravamudan 443117743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 443217743798SMuchun Song &tmp); 443308d4a246SMichal Hocko if (ret) 443408d4a246SMichal Hocko goto out; 4435e5ff2159SAndi Kleen 4436238d3c13SDavid Rientjes if (write) 4437238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 4438238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 443908d4a246SMichal Hocko out: 444008d4a246SMichal Hocko return ret; 44411da177e4SLinus Torvalds } 4442396faf03SMel Gorman 444306808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 444432927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 444506808b08SLee Schermerhorn { 444606808b08SLee Schermerhorn 444706808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 444806808b08SLee Schermerhorn buffer, length, ppos); 444906808b08SLee Schermerhorn } 445006808b08SLee Schermerhorn 445106808b08SLee Schermerhorn #ifdef CONFIG_NUMA 445206808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 445332927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 445406808b08SLee Schermerhorn { 445506808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 445606808b08SLee Schermerhorn buffer, length, ppos); 445706808b08SLee Schermerhorn } 445806808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 445906808b08SLee Schermerhorn 4460a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 446132927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 4462a3d0c6aaSNishanth Aravamudan { 4463a5516438SAndi Kleen struct hstate *h = &default_hstate; 4464e5ff2159SAndi Kleen unsigned long tmp; 446508d4a246SMichal Hocko int ret; 4466e5ff2159SAndi Kleen 4467457c1b27SNishanth Aravamudan if (!hugepages_supported()) 446886613628SJan Stancek return -EOPNOTSUPP; 4469457c1b27SNishanth Aravamudan 4470e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 4471e5ff2159SAndi Kleen 4472bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 4473adbe8726SEric B Munson return -EINVAL; 4474adbe8726SEric B Munson 447517743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 447617743798SMuchun Song &tmp); 447708d4a246SMichal Hocko if (ret) 447808d4a246SMichal Hocko goto out; 4479e5ff2159SAndi Kleen 4480e5ff2159SAndi Kleen if (write) { 4481db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4482e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 4483db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4484e5ff2159SAndi Kleen } 448508d4a246SMichal Hocko out: 448608d4a246SMichal Hocko return ret; 4487a3d0c6aaSNishanth Aravamudan } 4488a3d0c6aaSNishanth Aravamudan 44891da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 44901da177e4SLinus Torvalds 4491e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 44921da177e4SLinus Torvalds { 4493fcb2b0c5SRoman Gushchin struct hstate *h; 4494fcb2b0c5SRoman Gushchin unsigned long total = 0; 4495fcb2b0c5SRoman Gushchin 4496457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4497457c1b27SNishanth Aravamudan return; 4498fcb2b0c5SRoman Gushchin 4499fcb2b0c5SRoman Gushchin for_each_hstate(h) { 4500fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 4501fcb2b0c5SRoman Gushchin 4502aca78307SMiaohe Lin total += huge_page_size(h) * count; 4503fcb2b0c5SRoman Gushchin 4504fcb2b0c5SRoman Gushchin if (h == &default_hstate) 4505e1759c21SAlexey Dobriyan seq_printf(m, 45061da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 45071da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 4508b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 45097893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 45104f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 4511fcb2b0c5SRoman Gushchin count, 4512a5516438SAndi Kleen h->free_huge_pages, 4513a5516438SAndi Kleen h->resv_huge_pages, 4514a5516438SAndi Kleen h->surplus_huge_pages, 4515aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4516fcb2b0c5SRoman Gushchin } 4517fcb2b0c5SRoman Gushchin 4518aca78307SMiaohe Lin seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); 45191da177e4SLinus Torvalds } 45201da177e4SLinus Torvalds 45217981593bSJoe Perches int hugetlb_report_node_meminfo(char *buf, int len, int nid) 45221da177e4SLinus Torvalds { 4523a5516438SAndi Kleen struct hstate *h = &default_hstate; 45247981593bSJoe Perches 4525457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4526457c1b27SNishanth Aravamudan return 0; 45277981593bSJoe Perches 45287981593bSJoe Perches return sysfs_emit_at(buf, len, 45291da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 4530a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 4531a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 4532a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 4533a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 4534a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 45351da177e4SLinus Torvalds } 45361da177e4SLinus Torvalds 4537dcadcf1cSGang Li void hugetlb_show_meminfo_node(int nid) 4538949f7ec5SDavid Rientjes { 4539949f7ec5SDavid Rientjes struct hstate *h; 4540949f7ec5SDavid Rientjes 4541457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4542457c1b27SNishanth Aravamudan return; 4543457c1b27SNishanth Aravamudan 4544949f7ec5SDavid Rientjes for_each_hstate(h) 4545dcadcf1cSGang Li printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 4546949f7ec5SDavid Rientjes nid, 4547949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 4548949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 4549949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 4550aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4551949f7ec5SDavid Rientjes } 4552949f7ec5SDavid Rientjes 45535d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 45545d317b2bSNaoya Horiguchi { 45555d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 45565d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 45575d317b2bSNaoya Horiguchi } 45585d317b2bSNaoya Horiguchi 45591da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 45601da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 45611da177e4SLinus Torvalds { 4562d0028588SWanpeng Li struct hstate *h; 4563d0028588SWanpeng Li unsigned long nr_total_pages = 0; 4564d0028588SWanpeng Li 4565d0028588SWanpeng Li for_each_hstate(h) 4566d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 4567d0028588SWanpeng Li return nr_total_pages; 45681da177e4SLinus Torvalds } 45691da177e4SLinus Torvalds 4570a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 4571fc1b8a73SMel Gorman { 4572fc1b8a73SMel Gorman int ret = -ENOMEM; 4573fc1b8a73SMel Gorman 45740aa7f354SMiaohe Lin if (!delta) 45750aa7f354SMiaohe Lin return 0; 45760aa7f354SMiaohe Lin 4577db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4578fc1b8a73SMel Gorman /* 4579fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 4580fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 4581fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 4582fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 4583fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 4584fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 4585fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 4586fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 4587fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 4588fc1b8a73SMel Gorman * 4589fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 4590fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 4591fc1b8a73SMel Gorman * we fall back to check against current free page availability as 4592fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 4593fc1b8a73SMel Gorman * semantics that cpuset has. 45948ca39e68SMuchun Song * 45958ca39e68SMuchun Song * Apart from cpuset, we also have memory policy mechanism that 45968ca39e68SMuchun Song * also determines from which node the kernel will allocate memory 45978ca39e68SMuchun Song * in a NUMA system. So similar to cpuset, we also should consider 45988ca39e68SMuchun Song * the memory policy of the current task. Similar to the description 45998ca39e68SMuchun Song * above. 4600fc1b8a73SMel Gorman */ 4601fc1b8a73SMel Gorman if (delta > 0) { 4602a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 4603fc1b8a73SMel Gorman goto out; 4604fc1b8a73SMel Gorman 46058ca39e68SMuchun Song if (delta > allowed_mems_nr(h)) { 4606a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 4607fc1b8a73SMel Gorman goto out; 4608fc1b8a73SMel Gorman } 4609fc1b8a73SMel Gorman } 4610fc1b8a73SMel Gorman 4611fc1b8a73SMel Gorman ret = 0; 4612fc1b8a73SMel Gorman if (delta < 0) 4613a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 4614fc1b8a73SMel Gorman 4615fc1b8a73SMel Gorman out: 4616db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4617fc1b8a73SMel Gorman return ret; 4618fc1b8a73SMel Gorman } 4619fc1b8a73SMel Gorman 462084afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 462184afd99bSAndy Whitcroft { 4622f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 462384afd99bSAndy Whitcroft 462484afd99bSAndy Whitcroft /* 462584afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 462684afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 462784afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 462825985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 462984afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 463084afd99bSAndy Whitcroft * new reference here without additional locking. 463184afd99bSAndy Whitcroft */ 463209a26e83SMike Kravetz if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 463309a26e83SMike Kravetz resv_map_dup_hugetlb_cgroup_uncharge_info(resv); 4634f522c3acSJoonsoo Kim kref_get(&resv->refs); 463584afd99bSAndy Whitcroft } 46368d9bfb26SMike Kravetz 46378d9bfb26SMike Kravetz hugetlb_vma_lock_alloc(vma); 463809a26e83SMike Kravetz } 463984afd99bSAndy Whitcroft 4640a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 4641a1e78772SMel Gorman { 4642a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 46438d9bfb26SMike Kravetz struct resv_map *resv; 464490481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 46454e35f483SJoonsoo Kim unsigned long reserve, start, end; 46461c5ecae3SMike Kravetz long gbl_reserve; 464784afd99bSAndy Whitcroft 46488d9bfb26SMike Kravetz hugetlb_vma_lock_free(vma); 46498d9bfb26SMike Kravetz 46508d9bfb26SMike Kravetz resv = vma_resv_map(vma); 46514e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 46524e35f483SJoonsoo Kim return; 46534e35f483SJoonsoo Kim 4654a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 4655a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 465684afd99bSAndy Whitcroft 46574e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 4658e9fe92aeSMina Almasry hugetlb_cgroup_uncharge_counter(resv, start, end); 46597251ff78SAdam Litke if (reserve) { 46601c5ecae3SMike Kravetz /* 46611c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 46621c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 46631c5ecae3SMike Kravetz */ 46641c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 46651c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 46667251ff78SAdam Litke } 4667e9fe92aeSMina Almasry 4668e9fe92aeSMina Almasry kref_put(&resv->refs, resv_map_release); 4669a1e78772SMel Gorman } 4670a1e78772SMel Gorman 467131383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 467231383c68SDan Williams { 467331383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 467431383c68SDan Williams return -EINVAL; 467531383c68SDan Williams return 0; 467631383c68SDan Williams } 467731383c68SDan Williams 467805ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 467905ea8860SDan Williams { 4680aca78307SMiaohe Lin return huge_page_size(hstate_vma(vma)); 468105ea8860SDan Williams } 468205ea8860SDan Williams 46831da177e4SLinus Torvalds /* 46841da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 46851da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 46866c26d310SMiaohe Lin * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get 46871da177e4SLinus Torvalds * this far. 46881da177e4SLinus Torvalds */ 4689b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 46901da177e4SLinus Torvalds { 46911da177e4SLinus Torvalds BUG(); 4692d0217ac0SNick Piggin return 0; 46931da177e4SLinus Torvalds } 46941da177e4SLinus Torvalds 4695eec3636aSJane Chu /* 4696eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 4697eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 4698eec3636aSJane Chu * This is because under System V memory model, mappings created via 4699eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 4700eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 4701eec3636aSJane Chu */ 4702f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 4703d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 470484afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 4705a1e78772SMel Gorman .close = hugetlb_vm_op_close, 4706dd3b614fSDmitry Safonov .may_split = hugetlb_vm_op_split, 470705ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 47081da177e4SLinus Torvalds }; 47091da177e4SLinus Torvalds 47101e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 47111e8f889bSDavid Gibson int writable) 471263551ae0SDavid Gibson { 471363551ae0SDavid Gibson pte_t entry; 471479c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 471563551ae0SDavid Gibson 47161e8f889bSDavid Gibson if (writable) { 4717106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 4718106c992aSGerald Schaefer vma->vm_page_prot))); 471963551ae0SDavid Gibson } else { 4720106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 4721106c992aSGerald Schaefer vma->vm_page_prot)); 472263551ae0SDavid Gibson } 472363551ae0SDavid Gibson entry = pte_mkyoung(entry); 472479c1c594SChristophe Leroy entry = arch_make_huge_pte(entry, shift, vma->vm_flags); 472563551ae0SDavid Gibson 472663551ae0SDavid Gibson return entry; 472763551ae0SDavid Gibson } 472863551ae0SDavid Gibson 47291e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 47301e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 47311e8f889bSDavid Gibson { 47321e8f889bSDavid Gibson pte_t entry; 47331e8f889bSDavid Gibson 4734106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 473532f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 47364b3073e1SRussell King update_mmu_cache(vma, address, ptep); 47371e8f889bSDavid Gibson } 47381e8f889bSDavid Gibson 4739d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 47404a705fefSNaoya Horiguchi { 47414a705fefSNaoya Horiguchi swp_entry_t swp; 47424a705fefSNaoya Horiguchi 47434a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 4744d5ed7444SAneesh Kumar K.V return false; 47454a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4746d79d176aSBaoquan He if (is_migration_entry(swp)) 4747d5ed7444SAneesh Kumar K.V return true; 47484a705fefSNaoya Horiguchi else 4749d5ed7444SAneesh Kumar K.V return false; 47504a705fefSNaoya Horiguchi } 47514a705fefSNaoya Horiguchi 47523e5c3600SBaoquan He static bool is_hugetlb_entry_hwpoisoned(pte_t pte) 47534a705fefSNaoya Horiguchi { 47544a705fefSNaoya Horiguchi swp_entry_t swp; 47554a705fefSNaoya Horiguchi 47564a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 47573e5c3600SBaoquan He return false; 47584a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4759d79d176aSBaoquan He if (is_hwpoison_entry(swp)) 47603e5c3600SBaoquan He return true; 47614a705fefSNaoya Horiguchi else 47623e5c3600SBaoquan He return false; 47634a705fefSNaoya Horiguchi } 47641e8f889bSDavid Gibson 47654eae4efaSPeter Xu static void 47664eae4efaSPeter Xu hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, 47674eae4efaSPeter Xu struct page *new_page) 47684eae4efaSPeter Xu { 47694eae4efaSPeter Xu __SetPageUptodate(new_page); 47704eae4efaSPeter Xu hugepage_add_new_anon_rmap(new_page, vma, addr); 47711eba86c0SPasha Tatashin set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1)); 47724eae4efaSPeter Xu hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); 47734eae4efaSPeter Xu ClearHPageRestoreReserve(new_page); 47744eae4efaSPeter Xu SetHPageMigratable(new_page); 47754eae4efaSPeter Xu } 47764eae4efaSPeter Xu 477763551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 4778bc70fbf2SPeter Xu struct vm_area_struct *dst_vma, 4779bc70fbf2SPeter Xu struct vm_area_struct *src_vma) 478063551ae0SDavid Gibson { 47813aa4ed80SMiaohe Lin pte_t *src_pte, *dst_pte, entry; 478263551ae0SDavid Gibson struct page *ptepage; 47831c59827dSHugh Dickins unsigned long addr; 4784bc70fbf2SPeter Xu bool cow = is_cow_mapping(src_vma->vm_flags); 4785bc70fbf2SPeter Xu struct hstate *h = hstate_vma(src_vma); 4786a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 47874eae4efaSPeter Xu unsigned long npages = pages_per_huge_page(h); 4788ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 4789e95a9851SMike Kravetz unsigned long last_addr_mask; 4790e8569dd2SAndreas Sandberg int ret = 0; 47911e8f889bSDavid Gibson 4792ac46d4f3SJérôme Glisse if (cow) { 4793bc70fbf2SPeter Xu mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src_vma, src, 4794bc70fbf2SPeter Xu src_vma->vm_start, 4795bc70fbf2SPeter Xu src_vma->vm_end); 4796ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4797623a1ddfSDavid Hildenbrand mmap_assert_write_locked(src); 4798623a1ddfSDavid Hildenbrand raw_write_seqcount_begin(&src->write_protect_seq); 4799*40549ba8SMike Kravetz } else { 4800*40549ba8SMike Kravetz /* 4801*40549ba8SMike Kravetz * For shared mappings the vma lock must be held before 4802*40549ba8SMike Kravetz * calling huge_pte_offset in the src vma. Otherwise, the 4803*40549ba8SMike Kravetz * returned ptep could go away if part of a shared pmd and 4804*40549ba8SMike Kravetz * another thread calls huge_pmd_unshare. 4805*40549ba8SMike Kravetz */ 4806*40549ba8SMike Kravetz hugetlb_vma_lock_read(src_vma); 4807ac46d4f3SJérôme Glisse } 4808e8569dd2SAndreas Sandberg 4809e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 4810bc70fbf2SPeter Xu for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) { 4811cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 48127868a208SPunit Agrawal src_pte = huge_pte_offset(src, addr, sz); 4813e95a9851SMike Kravetz if (!src_pte) { 4814e95a9851SMike Kravetz addr |= last_addr_mask; 4815c74df32cSHugh Dickins continue; 4816e95a9851SMike Kravetz } 4817bc70fbf2SPeter Xu dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz); 4818e8569dd2SAndreas Sandberg if (!dst_pte) { 4819e8569dd2SAndreas Sandberg ret = -ENOMEM; 4820e8569dd2SAndreas Sandberg break; 4821e8569dd2SAndreas Sandberg } 4822c5c99429SLarry Woodman 48235e41540cSMike Kravetz /* 48245e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 48255e41540cSMike Kravetz * 48263aa4ed80SMiaohe Lin * dst_pte == src_pte is the common case of src/dest sharing. 48275e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 48283aa4ed80SMiaohe Lin * another vma. So page_count of ptep page is checked instead 48293aa4ed80SMiaohe Lin * to reliably determine whether pte is shared. 48305e41540cSMike Kravetz */ 48313aa4ed80SMiaohe Lin if (page_count(virt_to_page(dst_pte)) > 1) { 4832e95a9851SMike Kravetz addr |= last_addr_mask; 4833c5c99429SLarry Woodman continue; 4834e95a9851SMike Kravetz } 4835c5c99429SLarry Woodman 4836cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 4837cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 4838cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 48394a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 48404eae4efaSPeter Xu again: 48413aa4ed80SMiaohe Lin if (huge_pte_none(entry)) { 48425e41540cSMike Kravetz /* 48433aa4ed80SMiaohe Lin * Skip if src entry none. 48445e41540cSMike Kravetz */ 48454a705fefSNaoya Horiguchi ; 4846c2cb0dccSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) { 4847c2cb0dccSNaoya Horiguchi bool uffd_wp = huge_pte_uffd_wp(entry); 4848c2cb0dccSNaoya Horiguchi 4849c2cb0dccSNaoya Horiguchi if (!userfaultfd_wp(dst_vma) && uffd_wp) 4850c2cb0dccSNaoya Horiguchi entry = huge_pte_clear_uffd_wp(entry); 4851c2cb0dccSNaoya Horiguchi set_huge_pte_at(dst, addr, dst_pte, entry); 4852c2cb0dccSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry))) { 48534a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 4854bc70fbf2SPeter Xu bool uffd_wp = huge_pte_uffd_wp(entry); 48554a705fefSNaoya Horiguchi 48566c287605SDavid Hildenbrand if (!is_readable_migration_entry(swp_entry) && cow) { 48574a705fefSNaoya Horiguchi /* 48584a705fefSNaoya Horiguchi * COW mappings require pages in both 48594a705fefSNaoya Horiguchi * parent and child to be set to read. 48604a705fefSNaoya Horiguchi */ 48614dd845b5SAlistair Popple swp_entry = make_readable_migration_entry( 48624dd845b5SAlistair Popple swp_offset(swp_entry)); 48634a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 4864bc70fbf2SPeter Xu if (userfaultfd_wp(src_vma) && uffd_wp) 4865bc70fbf2SPeter Xu entry = huge_pte_mkuffd_wp(entry); 486618f39629SQi Zheng set_huge_pte_at(src, addr, src_pte, entry); 48674a705fefSNaoya Horiguchi } 4868bc70fbf2SPeter Xu if (!userfaultfd_wp(dst_vma) && uffd_wp) 4869bc70fbf2SPeter Xu entry = huge_pte_clear_uffd_wp(entry); 487018f39629SQi Zheng set_huge_pte_at(dst, addr, dst_pte, entry); 4871bc70fbf2SPeter Xu } else if (unlikely(is_pte_marker(entry))) { 4872bc70fbf2SPeter Xu /* 4873bc70fbf2SPeter Xu * We copy the pte marker only if the dst vma has 4874bc70fbf2SPeter Xu * uffd-wp enabled. 4875bc70fbf2SPeter Xu */ 4876bc70fbf2SPeter Xu if (userfaultfd_wp(dst_vma)) 4877bc70fbf2SPeter Xu set_huge_pte_at(dst, addr, dst_pte, entry); 48784a705fefSNaoya Horiguchi } else { 48794eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 48804eae4efaSPeter Xu ptepage = pte_page(entry); 48814eae4efaSPeter Xu get_page(ptepage); 48824eae4efaSPeter Xu 48834eae4efaSPeter Xu /* 4884fb3d824dSDavid Hildenbrand * Failing to duplicate the anon rmap is a rare case 4885fb3d824dSDavid Hildenbrand * where we see pinned hugetlb pages while they're 4886fb3d824dSDavid Hildenbrand * prone to COW. We need to do the COW earlier during 4887fb3d824dSDavid Hildenbrand * fork. 48884eae4efaSPeter Xu * 48894eae4efaSPeter Xu * When pre-allocating the page or copying data, we 48904eae4efaSPeter Xu * need to be without the pgtable locks since we could 48914eae4efaSPeter Xu * sleep during the process. 48924eae4efaSPeter Xu */ 4893fb3d824dSDavid Hildenbrand if (!PageAnon(ptepage)) { 4894fb3d824dSDavid Hildenbrand page_dup_file_rmap(ptepage, true); 4895bc70fbf2SPeter Xu } else if (page_try_dup_anon_rmap(ptepage, true, 4896bc70fbf2SPeter Xu src_vma)) { 48974eae4efaSPeter Xu pte_t src_pte_old = entry; 48984eae4efaSPeter Xu struct page *new; 48994eae4efaSPeter Xu 49004eae4efaSPeter Xu spin_unlock(src_ptl); 49014eae4efaSPeter Xu spin_unlock(dst_ptl); 49024eae4efaSPeter Xu /* Do not use reserve as it's private owned */ 4903bc70fbf2SPeter Xu new = alloc_huge_page(dst_vma, addr, 1); 49044eae4efaSPeter Xu if (IS_ERR(new)) { 49054eae4efaSPeter Xu put_page(ptepage); 49064eae4efaSPeter Xu ret = PTR_ERR(new); 49074eae4efaSPeter Xu break; 49084eae4efaSPeter Xu } 4909bc70fbf2SPeter Xu copy_user_huge_page(new, ptepage, addr, dst_vma, 49104eae4efaSPeter Xu npages); 49114eae4efaSPeter Xu put_page(ptepage); 49124eae4efaSPeter Xu 49134eae4efaSPeter Xu /* Install the new huge page if src pte stable */ 49144eae4efaSPeter Xu dst_ptl = huge_pte_lock(h, dst, dst_pte); 49154eae4efaSPeter Xu src_ptl = huge_pte_lockptr(h, src, src_pte); 49164eae4efaSPeter Xu spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 49174eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 49184eae4efaSPeter Xu if (!pte_same(src_pte_old, entry)) { 4919bc70fbf2SPeter Xu restore_reserve_on_error(h, dst_vma, addr, 4920846be085SMike Kravetz new); 49214eae4efaSPeter Xu put_page(new); 49223aa4ed80SMiaohe Lin /* huge_ptep of dst_pte won't change as in child */ 49234eae4efaSPeter Xu goto again; 49244eae4efaSPeter Xu } 4925bc70fbf2SPeter Xu hugetlb_install_page(dst_vma, dst_pte, addr, new); 49264eae4efaSPeter Xu spin_unlock(src_ptl); 49274eae4efaSPeter Xu spin_unlock(dst_ptl); 49284eae4efaSPeter Xu continue; 49294eae4efaSPeter Xu } 49304eae4efaSPeter Xu 493134ee645eSJoerg Roedel if (cow) { 49320f10851eSJérôme Glisse /* 49330f10851eSJérôme Glisse * No need to notify as we are downgrading page 49340f10851eSJérôme Glisse * table protection not changing it to point 49350f10851eSJérôme Glisse * to a new page. 49360f10851eSJérôme Glisse * 4937ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 49380f10851eSJérôme Glisse */ 49397f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 494084894e1cSPeter Xu entry = huge_pte_wrprotect(entry); 494134ee645eSJoerg Roedel } 49424eae4efaSPeter Xu 494363551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 49444eae4efaSPeter Xu hugetlb_count_add(npages, dst); 49451c59827dSHugh Dickins } 4946cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 4947cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 494863551ae0SDavid Gibson } 494963551ae0SDavid Gibson 4950623a1ddfSDavid Hildenbrand if (cow) { 4951623a1ddfSDavid Hildenbrand raw_write_seqcount_end(&src->write_protect_seq); 4952ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 4953*40549ba8SMike Kravetz } else { 4954*40549ba8SMike Kravetz hugetlb_vma_unlock_read(src_vma); 4955623a1ddfSDavid Hildenbrand } 4956e8569dd2SAndreas Sandberg 4957e8569dd2SAndreas Sandberg return ret; 495863551ae0SDavid Gibson } 495963551ae0SDavid Gibson 4960550a7d60SMina Almasry static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr, 4961db110a99SAneesh Kumar K.V unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte) 4962550a7d60SMina Almasry { 4963550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 4964550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 4965550a7d60SMina Almasry spinlock_t *src_ptl, *dst_ptl; 4966db110a99SAneesh Kumar K.V pte_t pte; 4967550a7d60SMina Almasry 4968550a7d60SMina Almasry dst_ptl = huge_pte_lock(h, mm, dst_pte); 4969550a7d60SMina Almasry src_ptl = huge_pte_lockptr(h, mm, src_pte); 4970550a7d60SMina Almasry 4971550a7d60SMina Almasry /* 4972550a7d60SMina Almasry * We don't have to worry about the ordering of src and dst ptlocks 4973550a7d60SMina Almasry * because exclusive mmap_sem (or the i_mmap_lock) prevents deadlock. 4974550a7d60SMina Almasry */ 4975550a7d60SMina Almasry if (src_ptl != dst_ptl) 4976550a7d60SMina Almasry spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 4977550a7d60SMina Almasry 4978550a7d60SMina Almasry pte = huge_ptep_get_and_clear(mm, old_addr, src_pte); 4979550a7d60SMina Almasry set_huge_pte_at(mm, new_addr, dst_pte, pte); 4980550a7d60SMina Almasry 4981550a7d60SMina Almasry if (src_ptl != dst_ptl) 4982550a7d60SMina Almasry spin_unlock(src_ptl); 4983550a7d60SMina Almasry spin_unlock(dst_ptl); 4984550a7d60SMina Almasry } 4985550a7d60SMina Almasry 4986550a7d60SMina Almasry int move_hugetlb_page_tables(struct vm_area_struct *vma, 4987550a7d60SMina Almasry struct vm_area_struct *new_vma, 4988550a7d60SMina Almasry unsigned long old_addr, unsigned long new_addr, 4989550a7d60SMina Almasry unsigned long len) 4990550a7d60SMina Almasry { 4991550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 4992550a7d60SMina Almasry struct address_space *mapping = vma->vm_file->f_mapping; 4993550a7d60SMina Almasry unsigned long sz = huge_page_size(h); 4994550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 4995550a7d60SMina Almasry unsigned long old_end = old_addr + len; 4996e95a9851SMike Kravetz unsigned long last_addr_mask; 4997550a7d60SMina Almasry pte_t *src_pte, *dst_pte; 4998550a7d60SMina Almasry struct mmu_notifier_range range; 49993d0b95cdSBaolin Wang bool shared_pmd = false; 5000550a7d60SMina Almasry 5001550a7d60SMina Almasry mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, old_addr, 5002550a7d60SMina Almasry old_end); 5003550a7d60SMina Almasry adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 50043d0b95cdSBaolin Wang /* 50053d0b95cdSBaolin Wang * In case of shared PMDs, we should cover the maximum possible 50063d0b95cdSBaolin Wang * range. 50073d0b95cdSBaolin Wang */ 50083d0b95cdSBaolin Wang flush_cache_range(vma, range.start, range.end); 50093d0b95cdSBaolin Wang 5010550a7d60SMina Almasry mmu_notifier_invalidate_range_start(&range); 5011e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 5012550a7d60SMina Almasry /* Prevent race with file truncation */ 5013*40549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 5014550a7d60SMina Almasry i_mmap_lock_write(mapping); 5015550a7d60SMina Almasry for (; old_addr < old_end; old_addr += sz, new_addr += sz) { 5016550a7d60SMina Almasry src_pte = huge_pte_offset(mm, old_addr, sz); 5017e95a9851SMike Kravetz if (!src_pte) { 5018e95a9851SMike Kravetz old_addr |= last_addr_mask; 5019e95a9851SMike Kravetz new_addr |= last_addr_mask; 5020550a7d60SMina Almasry continue; 5021e95a9851SMike Kravetz } 5022550a7d60SMina Almasry if (huge_pte_none(huge_ptep_get(src_pte))) 5023550a7d60SMina Almasry continue; 5024550a7d60SMina Almasry 50254ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) { 50263d0b95cdSBaolin Wang shared_pmd = true; 50274ddb4d91SMike Kravetz old_addr |= last_addr_mask; 50284ddb4d91SMike Kravetz new_addr |= last_addr_mask; 5029550a7d60SMina Almasry continue; 50303d0b95cdSBaolin Wang } 5031550a7d60SMina Almasry 5032550a7d60SMina Almasry dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz); 5033550a7d60SMina Almasry if (!dst_pte) 5034550a7d60SMina Almasry break; 5035550a7d60SMina Almasry 5036db110a99SAneesh Kumar K.V move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte); 5037550a7d60SMina Almasry } 50383d0b95cdSBaolin Wang 50393d0b95cdSBaolin Wang if (shared_pmd) 50403d0b95cdSBaolin Wang flush_tlb_range(vma, range.start, range.end); 50413d0b95cdSBaolin Wang else 5042550a7d60SMina Almasry flush_tlb_range(vma, old_end - len, old_end); 5043550a7d60SMina Almasry mmu_notifier_invalidate_range_end(&range); 504413e4ad2cSNadav Amit i_mmap_unlock_write(mapping); 5045*40549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 5046550a7d60SMina Almasry 5047550a7d60SMina Almasry return len + old_addr - old_end; 5048550a7d60SMina Almasry } 5049550a7d60SMina Almasry 505073c54763SPeter Xu static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 505124669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 505205e90bd0SPeter Xu struct page *ref_page, zap_flags_t zap_flags) 505363551ae0SDavid Gibson { 505463551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 505563551ae0SDavid Gibson unsigned long address; 5056c7546f8fSDavid Gibson pte_t *ptep; 505763551ae0SDavid Gibson pte_t pte; 5058cb900f41SKirill A. Shutemov spinlock_t *ptl; 505963551ae0SDavid Gibson struct page *page; 5060a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 5061a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 5062ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 5063e95a9851SMike Kravetz unsigned long last_addr_mask; 5064a4a118f2SNadav Amit bool force_flush = false; 5065a5516438SAndi Kleen 506663551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 5067a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 5068a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 506963551ae0SDavid Gibson 507007e32661SAneesh Kumar K.V /* 507107e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 507207e32661SAneesh Kumar K.V * to huge page. 507307e32661SAneesh Kumar K.V */ 5074ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 507524669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 5076dff11abeSMike Kravetz 5077dff11abeSMike Kravetz /* 5078dff11abeSMike Kravetz * If sharing possible, alert mmu notifiers of worst case. 5079dff11abeSMike Kravetz */ 50806f4f13e8SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, 50816f4f13e8SJérôme Glisse end); 5082ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 5083ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 5084e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 5085569f48b8SHillf Danton address = start; 5086569f48b8SHillf Danton for (; address < end; address += sz) { 50877868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, sz); 5088e95a9851SMike Kravetz if (!ptep) { 5089e95a9851SMike Kravetz address |= last_addr_mask; 5090c7546f8fSDavid Gibson continue; 5091e95a9851SMike Kravetz } 5092c7546f8fSDavid Gibson 5093cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 50944ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, address, ptep)) { 509531d49da5SAneesh Kumar K.V spin_unlock(ptl); 5096a4a118f2SNadav Amit tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE); 5097a4a118f2SNadav Amit force_flush = true; 50984ddb4d91SMike Kravetz address |= last_addr_mask; 509931d49da5SAneesh Kumar K.V continue; 510031d49da5SAneesh Kumar K.V } 510139dde65cSChen, Kenneth W 51026629326bSHillf Danton pte = huge_ptep_get(ptep); 510331d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 510431d49da5SAneesh Kumar K.V spin_unlock(ptl); 510531d49da5SAneesh Kumar K.V continue; 510631d49da5SAneesh Kumar K.V } 51076629326bSHillf Danton 51086629326bSHillf Danton /* 51099fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 51109fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 51116629326bSHillf Danton */ 51129fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 511305e90bd0SPeter Xu /* 511405e90bd0SPeter Xu * If the pte was wr-protected by uffd-wp in any of the 511505e90bd0SPeter Xu * swap forms, meanwhile the caller does not want to 511605e90bd0SPeter Xu * drop the uffd-wp bit in this zap, then replace the 511705e90bd0SPeter Xu * pte with a marker. 511805e90bd0SPeter Xu */ 511905e90bd0SPeter Xu if (pte_swp_uffd_wp_any(pte) && 512005e90bd0SPeter Xu !(zap_flags & ZAP_FLAG_DROP_MARKER)) 512105e90bd0SPeter Xu set_huge_pte_at(mm, address, ptep, 512205e90bd0SPeter Xu make_pte_marker(PTE_MARKER_UFFD_WP)); 512305e90bd0SPeter Xu else 51249386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 512531d49da5SAneesh Kumar K.V spin_unlock(ptl); 512631d49da5SAneesh Kumar K.V continue; 51278c4894c6SNaoya Horiguchi } 51286629326bSHillf Danton 51296629326bSHillf Danton page = pte_page(pte); 513004f2cbe3SMel Gorman /* 513104f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 513204f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 513304f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 513404f2cbe3SMel Gorman */ 513504f2cbe3SMel Gorman if (ref_page) { 513631d49da5SAneesh Kumar K.V if (page != ref_page) { 513731d49da5SAneesh Kumar K.V spin_unlock(ptl); 513831d49da5SAneesh Kumar K.V continue; 513931d49da5SAneesh Kumar K.V } 514004f2cbe3SMel Gorman /* 514104f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 514204f2cbe3SMel Gorman * future faults in this VMA will fail rather than 514304f2cbe3SMel Gorman * looking like data was lost 514404f2cbe3SMel Gorman */ 514504f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 514604f2cbe3SMel Gorman } 514704f2cbe3SMel Gorman 5148c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 5149b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 5150106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 51516649a386SKen Chen set_page_dirty(page); 515205e90bd0SPeter Xu /* Leave a uffd-wp pte marker if needed */ 515305e90bd0SPeter Xu if (huge_pte_uffd_wp(pte) && 515405e90bd0SPeter Xu !(zap_flags & ZAP_FLAG_DROP_MARKER)) 515505e90bd0SPeter Xu set_huge_pte_at(mm, address, ptep, 515605e90bd0SPeter Xu make_pte_marker(PTE_MARKER_UFFD_WP)); 51575d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 5158cea86fe2SHugh Dickins page_remove_rmap(page, vma, true); 515931d49da5SAneesh Kumar K.V 5160cb900f41SKirill A. Shutemov spin_unlock(ptl); 5161e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 516224669e58SAneesh Kumar K.V /* 516331d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 516424669e58SAneesh Kumar K.V */ 516531d49da5SAneesh Kumar K.V if (ref_page) 516631d49da5SAneesh Kumar K.V break; 5167fe1668aeSChen, Kenneth W } 5168ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 516924669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 5170a4a118f2SNadav Amit 5171a4a118f2SNadav Amit /* 5172a4a118f2SNadav Amit * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We 5173a4a118f2SNadav Amit * could defer the flush until now, since by holding i_mmap_rwsem we 5174a4a118f2SNadav Amit * guaranteed that the last refernece would not be dropped. But we must 5175a4a118f2SNadav Amit * do the flushing before we return, as otherwise i_mmap_rwsem will be 5176a4a118f2SNadav Amit * dropped and the last reference to the shared PMDs page might be 5177a4a118f2SNadav Amit * dropped as well. 5178a4a118f2SNadav Amit * 5179a4a118f2SNadav Amit * In theory we could defer the freeing of the PMD pages as well, but 5180a4a118f2SNadav Amit * huge_pmd_unshare() relies on the exact page_count for the PMD page to 5181a4a118f2SNadav Amit * detect sharing, so we cannot defer the release of the page either. 5182a4a118f2SNadav Amit * Instead, do flush now. 5183a4a118f2SNadav Amit */ 5184a4a118f2SNadav Amit if (force_flush) 5185a4a118f2SNadav Amit tlb_flush_mmu_tlbonly(tlb); 51861da177e4SLinus Torvalds } 518763551ae0SDavid Gibson 5188d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 5189d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 519005e90bd0SPeter Xu unsigned long end, struct page *ref_page, 519105e90bd0SPeter Xu zap_flags_t zap_flags) 5192d833352aSMel Gorman { 519305e90bd0SPeter Xu __unmap_hugepage_range(tlb, vma, start, end, ref_page, zap_flags); 5194d833352aSMel Gorman 5195d833352aSMel Gorman /* 5196d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 5197d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 5198d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 5199d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 5200d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 5201c8c06efaSDavidlohr Bueso * is to clear it before releasing the i_mmap_rwsem. This works 5202d833352aSMel Gorman * because in the context this is called, the VMA is about to be 5203c8c06efaSDavidlohr Bueso * destroyed and the i_mmap_rwsem is held. 5204d833352aSMel Gorman */ 5205d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 5206d833352aSMel Gorman } 5207d833352aSMel Gorman 5208502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 520905e90bd0SPeter Xu unsigned long end, struct page *ref_page, 521005e90bd0SPeter Xu zap_flags_t zap_flags) 5211502717f4SChen, Kenneth W { 521224669e58SAneesh Kumar K.V struct mmu_gather tlb; 5213dff11abeSMike Kravetz 5214a72afd87SWill Deacon tlb_gather_mmu(&tlb, vma->vm_mm); 521505e90bd0SPeter Xu __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags); 5216ae8eba8bSWill Deacon tlb_finish_mmu(&tlb); 5217502717f4SChen, Kenneth W } 5218502717f4SChen, Kenneth W 521904f2cbe3SMel Gorman /* 522004f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 5221578b7725SZhiyuan Dai * mapping it owns the reserve page for. The intention is to unmap the page 522204f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 522304f2cbe3SMel Gorman * same region. 522404f2cbe3SMel Gorman */ 52252f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 52262a4b3dedSHarvey Harrison struct page *page, unsigned long address) 522704f2cbe3SMel Gorman { 52287526674dSAdam Litke struct hstate *h = hstate_vma(vma); 522904f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 523004f2cbe3SMel Gorman struct address_space *mapping; 523104f2cbe3SMel Gorman pgoff_t pgoff; 523204f2cbe3SMel Gorman 523304f2cbe3SMel Gorman /* 523404f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 523504f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 523604f2cbe3SMel Gorman */ 52377526674dSAdam Litke address = address & huge_page_mask(h); 523836e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 523936e4f20aSMichal Hocko vma->vm_pgoff; 524093c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 524104f2cbe3SMel Gorman 52424eb2b1dcSMel Gorman /* 52434eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 52444eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 52454eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 52464eb2b1dcSMel Gorman */ 524783cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 52486b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 524904f2cbe3SMel Gorman /* Do not unmap the current VMA */ 525004f2cbe3SMel Gorman if (iter_vma == vma) 525104f2cbe3SMel Gorman continue; 525204f2cbe3SMel Gorman 525304f2cbe3SMel Gorman /* 52542f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 52552f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 52562f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 52572f84a899SMel Gorman */ 52582f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 52592f84a899SMel Gorman continue; 52602f84a899SMel Gorman 52612f84a899SMel Gorman /* 526204f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 526304f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 526404f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 526504f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 526604f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 526704f2cbe3SMel Gorman */ 526804f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 526924669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 527005e90bd0SPeter Xu address + huge_page_size(h), page, 0); 527104f2cbe3SMel Gorman } 527283cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 527304f2cbe3SMel Gorman } 527404f2cbe3SMel Gorman 52750fe6e20bSNaoya Horiguchi /* 5276c89357e2SDavid Hildenbrand * hugetlb_wp() should be called with page lock of the original hugepage held. 5277aa6d2e8cSBaolin Wang * Called with hugetlb_fault_mutex_table held and pte_page locked so we 5278ef009b25SMichal Hocko * cannot race with other handlers or page migration. 5279ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 52800fe6e20bSNaoya Horiguchi */ 5281c89357e2SDavid Hildenbrand static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma, 5282c89357e2SDavid Hildenbrand unsigned long address, pte_t *ptep, unsigned int flags, 5283cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 52841e8f889bSDavid Gibson { 5285c89357e2SDavid Hildenbrand const bool unshare = flags & FAULT_FLAG_UNSHARE; 52863999f52eSAneesh Kumar K.V pte_t pte; 5287a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 52881e8f889bSDavid Gibson struct page *old_page, *new_page; 52892b740303SSouptick Joarder int outside_reserve = 0; 52902b740303SSouptick Joarder vm_fault_t ret = 0; 5291974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 5292ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 52931e8f889bSDavid Gibson 5294c89357e2SDavid Hildenbrand VM_BUG_ON(unshare && (flags & FOLL_WRITE)); 5295c89357e2SDavid Hildenbrand VM_BUG_ON(!unshare && !(flags & FOLL_WRITE)); 5296c89357e2SDavid Hildenbrand 52971d8d1464SDavid Hildenbrand /* 52981d8d1464SDavid Hildenbrand * hugetlb does not support FOLL_FORCE-style write faults that keep the 52991d8d1464SDavid Hildenbrand * PTE mapped R/O such as maybe_mkwrite() would do. 53001d8d1464SDavid Hildenbrand */ 53011d8d1464SDavid Hildenbrand if (WARN_ON_ONCE(!unshare && !(vma->vm_flags & VM_WRITE))) 53021d8d1464SDavid Hildenbrand return VM_FAULT_SIGSEGV; 53031d8d1464SDavid Hildenbrand 53041d8d1464SDavid Hildenbrand /* Let's take out MAP_SHARED mappings first. */ 53051d8d1464SDavid Hildenbrand if (vma->vm_flags & VM_MAYSHARE) { 53061d8d1464SDavid Hildenbrand if (unlikely(unshare)) 53071d8d1464SDavid Hildenbrand return 0; 53081d8d1464SDavid Hildenbrand set_huge_ptep_writable(vma, haddr, ptep); 53091d8d1464SDavid Hildenbrand return 0; 53101d8d1464SDavid Hildenbrand } 53111d8d1464SDavid Hildenbrand 53123999f52eSAneesh Kumar K.V pte = huge_ptep_get(ptep); 53131e8f889bSDavid Gibson old_page = pte_page(pte); 53141e8f889bSDavid Gibson 5315662ce1dcSYang Yang delayacct_wpcopy_start(); 5316662ce1dcSYang Yang 531704f2cbe3SMel Gorman retry_avoidcopy: 5318c89357e2SDavid Hildenbrand /* 5319c89357e2SDavid Hildenbrand * If no-one else is actually using this page, we're the exclusive 5320c89357e2SDavid Hildenbrand * owner and can reuse this page. 5321c89357e2SDavid Hildenbrand */ 532237a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 5323c89357e2SDavid Hildenbrand if (!PageAnonExclusive(old_page)) 53245a49973dSHugh Dickins page_move_anon_rmap(old_page, vma); 5325c89357e2SDavid Hildenbrand if (likely(!unshare)) 53265b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 5327662ce1dcSYang Yang 5328662ce1dcSYang Yang delayacct_wpcopy_end(); 532983c54070SNick Piggin return 0; 53301e8f889bSDavid Gibson } 53316c287605SDavid Hildenbrand VM_BUG_ON_PAGE(PageAnon(old_page) && PageAnonExclusive(old_page), 53326c287605SDavid Hildenbrand old_page); 53331e8f889bSDavid Gibson 533404f2cbe3SMel Gorman /* 533504f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 533604f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 533704f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 533804f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 533904f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 534004f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 534104f2cbe3SMel Gorman * of the full address range. 534204f2cbe3SMel Gorman */ 53435944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 534404f2cbe3SMel Gorman old_page != pagecache_page) 534504f2cbe3SMel Gorman outside_reserve = 1; 534604f2cbe3SMel Gorman 534709cbfeafSKirill A. Shutemov get_page(old_page); 5348b76c8cfbSLarry Woodman 5349ad4404a2SDavidlohr Bueso /* 5350ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 5351ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 5352ad4404a2SDavidlohr Bueso */ 5353cb900f41SKirill A. Shutemov spin_unlock(ptl); 53545b7a1d40SHuang Ying new_page = alloc_huge_page(vma, haddr, outside_reserve); 53551e8f889bSDavid Gibson 53562fc39cecSAdam Litke if (IS_ERR(new_page)) { 535704f2cbe3SMel Gorman /* 535804f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 535904f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 536004f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 536104f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 536204f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 536304f2cbe3SMel Gorman */ 536404f2cbe3SMel Gorman if (outside_reserve) { 5365*40549ba8SMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 5366*40549ba8SMike Kravetz pgoff_t idx; 5367*40549ba8SMike Kravetz u32 hash; 5368*40549ba8SMike Kravetz 536909cbfeafSKirill A. Shutemov put_page(old_page); 5370*40549ba8SMike Kravetz /* 5371*40549ba8SMike Kravetz * Drop hugetlb_fault_mutex and vma_lock before 5372*40549ba8SMike Kravetz * unmapping. unmapping needs to hold vma_lock 5373*40549ba8SMike Kravetz * in write mode. Dropping vma_lock in read mode 5374*40549ba8SMike Kravetz * here is OK as COW mappings do not interact with 5375*40549ba8SMike Kravetz * PMD sharing. 5376*40549ba8SMike Kravetz * 5377*40549ba8SMike Kravetz * Reacquire both after unmap operation. 5378*40549ba8SMike Kravetz */ 5379*40549ba8SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 5380*40549ba8SMike Kravetz hash = hugetlb_fault_mutex_hash(mapping, idx); 5381*40549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 5382*40549ba8SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5383*40549ba8SMike Kravetz 53845b7a1d40SHuang Ying unmap_ref_private(mm, vma, old_page, haddr); 5385*40549ba8SMike Kravetz 5386*40549ba8SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 5387*40549ba8SMike Kravetz hugetlb_vma_lock_read(vma); 5388cb900f41SKirill A. Shutemov spin_lock(ptl); 53895b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 5390a9af0c5dSNaoya Horiguchi if (likely(ptep && 5391a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 539204f2cbe3SMel Gorman goto retry_avoidcopy; 5393a734bcc8SHillf Danton /* 5394cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 5395cb900f41SKirill A. Shutemov * lock, and our job is done. 5396a734bcc8SHillf Danton */ 5397662ce1dcSYang Yang delayacct_wpcopy_end(); 5398a734bcc8SHillf Danton return 0; 539904f2cbe3SMel Gorman } 540004f2cbe3SMel Gorman 54012b740303SSouptick Joarder ret = vmf_error(PTR_ERR(new_page)); 5402ad4404a2SDavidlohr Bueso goto out_release_old; 54031e8f889bSDavid Gibson } 54041e8f889bSDavid Gibson 54050fe6e20bSNaoya Horiguchi /* 54060fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 54070fe6e20bSNaoya Horiguchi * anon_vma prepared. 54080fe6e20bSNaoya Horiguchi */ 540944e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 5410ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 5411ad4404a2SDavidlohr Bueso goto out_release_all; 541244e2aa93SDean Nelson } 54130fe6e20bSNaoya Horiguchi 5414974e6d66SHuang Ying copy_user_huge_page(new_page, old_page, address, vma, 541547ad8475SAndrea Arcangeli pages_per_huge_page(h)); 54160ed361deSNick Piggin __SetPageUptodate(new_page); 54171e8f889bSDavid Gibson 54187269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, 54196f4f13e8SJérôme Glisse haddr + huge_page_size(h)); 5420ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 5421ad4404a2SDavidlohr Bueso 5422b76c8cfbSLarry Woodman /* 5423cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 5424b76c8cfbSLarry Woodman * before the page tables are altered 5425b76c8cfbSLarry Woodman */ 5426cb900f41SKirill A. Shutemov spin_lock(ptl); 54275b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 5428a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 5429d6995da3SMike Kravetz ClearHPageRestoreReserve(new_page); 543007443a85SJoonsoo Kim 5431c89357e2SDavid Hildenbrand /* Break COW or unshare */ 54325b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 5433ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, range.end); 5434cea86fe2SHugh Dickins page_remove_rmap(old_page, vma, true); 54355b7a1d40SHuang Ying hugepage_add_new_anon_rmap(new_page, vma, haddr); 54361eba86c0SPasha Tatashin set_huge_pte_at(mm, haddr, ptep, 5437c89357e2SDavid Hildenbrand make_huge_pte(vma, new_page, !unshare)); 54388f251a3dSMike Kravetz SetHPageMigratable(new_page); 54391e8f889bSDavid Gibson /* Make the old page be freed below */ 54401e8f889bSDavid Gibson new_page = old_page; 54411e8f889bSDavid Gibson } 5442cb900f41SKirill A. Shutemov spin_unlock(ptl); 5443ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 5444ad4404a2SDavidlohr Bueso out_release_all: 5445c89357e2SDavid Hildenbrand /* 5446c89357e2SDavid Hildenbrand * No restore in case of successful pagetable update (Break COW or 5447c89357e2SDavid Hildenbrand * unshare) 5448c89357e2SDavid Hildenbrand */ 5449c7b1850dSMike Kravetz if (new_page != old_page) 54505b7a1d40SHuang Ying restore_reserve_on_error(h, vma, haddr, new_page); 545109cbfeafSKirill A. Shutemov put_page(new_page); 5452ad4404a2SDavidlohr Bueso out_release_old: 545309cbfeafSKirill A. Shutemov put_page(old_page); 54548312034fSJoonsoo Kim 5455ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 5456662ce1dcSYang Yang 5457662ce1dcSYang Yang delayacct_wpcopy_end(); 5458ad4404a2SDavidlohr Bueso return ret; 54591e8f889bSDavid Gibson } 54601e8f889bSDavid Gibson 54613ae77f43SHugh Dickins /* 54623ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 54633ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 54643ae77f43SHugh Dickins */ 54653ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 54662a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 54672a15efc9SHugh Dickins { 54682a15efc9SHugh Dickins struct address_space *mapping; 54692a15efc9SHugh Dickins pgoff_t idx; 54702a15efc9SHugh Dickins struct page *page; 54712a15efc9SHugh Dickins 54722a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 54732a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 54742a15efc9SHugh Dickins 54752a15efc9SHugh Dickins page = find_get_page(mapping, idx); 54762a15efc9SHugh Dickins if (page) 54772a15efc9SHugh Dickins put_page(page); 54782a15efc9SHugh Dickins return page != NULL; 54792a15efc9SHugh Dickins } 54802a15efc9SHugh Dickins 54817e1813d4SMike Kravetz int hugetlb_add_to_page_cache(struct page *page, struct address_space *mapping, 5482ab76ad54SMike Kravetz pgoff_t idx) 5483ab76ad54SMike Kravetz { 5484d9ef44deSMatthew Wilcox (Oracle) struct folio *folio = page_folio(page); 5485ab76ad54SMike Kravetz struct inode *inode = mapping->host; 5486ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 5487d9ef44deSMatthew Wilcox (Oracle) int err; 5488ab76ad54SMike Kravetz 5489d9ef44deSMatthew Wilcox (Oracle) __folio_set_locked(folio); 5490d9ef44deSMatthew Wilcox (Oracle) err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL); 5491d9ef44deSMatthew Wilcox (Oracle) 5492d9ef44deSMatthew Wilcox (Oracle) if (unlikely(err)) { 5493d9ef44deSMatthew Wilcox (Oracle) __folio_clear_locked(folio); 5494ab76ad54SMike Kravetz return err; 5495d9ef44deSMatthew Wilcox (Oracle) } 5496d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 5497ab76ad54SMike Kravetz 549822146c3cSMike Kravetz /* 5499d9ef44deSMatthew Wilcox (Oracle) * mark folio dirty so that it will not be removed from cache/file 550022146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 550122146c3cSMike Kravetz */ 5502d9ef44deSMatthew Wilcox (Oracle) folio_mark_dirty(folio); 550322146c3cSMike Kravetz 5504ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 5505ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 5506ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 5507ab76ad54SMike Kravetz return 0; 5508ab76ad54SMike Kravetz } 5509ab76ad54SMike Kravetz 55107677f7fdSAxel Rasmussen static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma, 55117677f7fdSAxel Rasmussen struct address_space *mapping, 55127677f7fdSAxel Rasmussen pgoff_t idx, 55137677f7fdSAxel Rasmussen unsigned int flags, 55147677f7fdSAxel Rasmussen unsigned long haddr, 5515824ddc60SNadav Amit unsigned long addr, 55167677f7fdSAxel Rasmussen unsigned long reason) 55177677f7fdSAxel Rasmussen { 55187677f7fdSAxel Rasmussen vm_fault_t ret; 55197677f7fdSAxel Rasmussen u32 hash; 55207677f7fdSAxel Rasmussen struct vm_fault vmf = { 55217677f7fdSAxel Rasmussen .vma = vma, 55227677f7fdSAxel Rasmussen .address = haddr, 5523824ddc60SNadav Amit .real_address = addr, 55247677f7fdSAxel Rasmussen .flags = flags, 55257677f7fdSAxel Rasmussen 55267677f7fdSAxel Rasmussen /* 55277677f7fdSAxel Rasmussen * Hard to debug if it ends up being 55287677f7fdSAxel Rasmussen * used by a callee that assumes 55297677f7fdSAxel Rasmussen * something about the other 55307677f7fdSAxel Rasmussen * uninitialized fields... same as in 55317677f7fdSAxel Rasmussen * memory.c 55327677f7fdSAxel Rasmussen */ 55337677f7fdSAxel Rasmussen }; 55347677f7fdSAxel Rasmussen 55357677f7fdSAxel Rasmussen /* 5536*40549ba8SMike Kravetz * vma_lock and hugetlb_fault_mutex must be 55377677f7fdSAxel Rasmussen * dropped before handling userfault. Reacquire 55387677f7fdSAxel Rasmussen * after handling fault to make calling code simpler. 55397677f7fdSAxel Rasmussen */ 5540*40549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 55417677f7fdSAxel Rasmussen hash = hugetlb_fault_mutex_hash(mapping, idx); 55427677f7fdSAxel Rasmussen mutex_unlock(&hugetlb_fault_mutex_table[hash]); 55437677f7fdSAxel Rasmussen ret = handle_userfault(&vmf, reason); 55447677f7fdSAxel Rasmussen mutex_lock(&hugetlb_fault_mutex_table[hash]); 5545*40549ba8SMike Kravetz hugetlb_vma_lock_read(vma); 55467677f7fdSAxel Rasmussen 55477677f7fdSAxel Rasmussen return ret; 55487677f7fdSAxel Rasmussen } 55497677f7fdSAxel Rasmussen 55502b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 55512b740303SSouptick Joarder struct vm_area_struct *vma, 55528382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 5553c64e912cSPeter Xu unsigned long address, pte_t *ptep, 5554c64e912cSPeter Xu pte_t old_pte, unsigned int flags) 5555ac9b9c66SHugh Dickins { 5556a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 55572b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 5558409eb8c2SHillf Danton int anon_rmap = 0; 55594c887265SAdam Litke unsigned long size; 55604c887265SAdam Litke struct page *page; 55611e8f889bSDavid Gibson pte_t new_pte; 5562cb900f41SKirill A. Shutemov spinlock_t *ptl; 5563285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 5564c7b1850dSMike Kravetz bool new_page, new_pagecache_page = false; 55654c887265SAdam Litke 556604f2cbe3SMel Gorman /* 556704f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 556804f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 5569c89357e2SDavid Hildenbrand * COW/unsharing. Warn that such a situation has occurred as it may not 5570c89357e2SDavid Hildenbrand * be obvious. 557104f2cbe3SMel Gorman */ 557204f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 5573910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 557404f2cbe3SMel Gorman current->pid); 557504f2cbe3SMel Gorman return ret; 557604f2cbe3SMel Gorman } 557704f2cbe3SMel Gorman 55784c887265SAdam Litke /* 5579188a3972SMike Kravetz * Use page lock to guard against racing truncation 5580188a3972SMike Kravetz * before we get page_table_lock. 55814c887265SAdam Litke */ 5582c7b1850dSMike Kravetz new_page = false; 558387bf91d3SMike Kravetz page = find_lock_page(mapping, idx); 558487bf91d3SMike Kravetz if (!page) { 5585188a3972SMike Kravetz size = i_size_read(mapping->host) >> huge_page_shift(h); 5586188a3972SMike Kravetz if (idx >= size) 5587188a3972SMike Kravetz goto out; 55887677f7fdSAxel Rasmussen /* Check for page in userfault range */ 55891a1aad8aSMike Kravetz if (userfaultfd_missing(vma)) { 55907677f7fdSAxel Rasmussen ret = hugetlb_handle_userfault(vma, mapping, idx, 5591824ddc60SNadav Amit flags, haddr, address, 55927677f7fdSAxel Rasmussen VM_UFFD_MISSING); 55931a1aad8aSMike Kravetz goto out; 55941a1aad8aSMike Kravetz } 55951a1aad8aSMike Kravetz 5596285b8dcaSHuang Ying page = alloc_huge_page(vma, haddr, 0); 55972fc39cecSAdam Litke if (IS_ERR(page)) { 55984643d67eSMike Kravetz /* 55994643d67eSMike Kravetz * Returning error will result in faulting task being 56004643d67eSMike Kravetz * sent SIGBUS. The hugetlb fault mutex prevents two 56014643d67eSMike Kravetz * tasks from racing to fault in the same page which 56024643d67eSMike Kravetz * could result in false unable to allocate errors. 56034643d67eSMike Kravetz * Page migration does not take the fault mutex, but 56044643d67eSMike Kravetz * does a clear then write of pte's under page table 56054643d67eSMike Kravetz * lock. Page fault code could race with migration, 56064643d67eSMike Kravetz * notice the clear pte and try to allocate a page 56074643d67eSMike Kravetz * here. Before returning error, get ptl and make 56084643d67eSMike Kravetz * sure there really is no pte entry. 56094643d67eSMike Kravetz */ 56104643d67eSMike Kravetz ptl = huge_pte_lock(h, mm, ptep); 56114643d67eSMike Kravetz ret = 0; 5612d83e6c8aSMiaohe Lin if (huge_pte_none(huge_ptep_get(ptep))) 56132b740303SSouptick Joarder ret = vmf_error(PTR_ERR(page)); 5614d83e6c8aSMiaohe Lin spin_unlock(ptl); 56156bda666aSChristoph Lameter goto out; 56166bda666aSChristoph Lameter } 561747ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 56180ed361deSNick Piggin __SetPageUptodate(page); 5619cb6acd01SMike Kravetz new_page = true; 5620ac9b9c66SHugh Dickins 5621f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 56227e1813d4SMike Kravetz int err = hugetlb_add_to_page_cache(page, mapping, idx); 56236bda666aSChristoph Lameter if (err) { 56243a5497a2SMiaohe Lin /* 56253a5497a2SMiaohe Lin * err can't be -EEXIST which implies someone 56263a5497a2SMiaohe Lin * else consumed the reservation since hugetlb 56273a5497a2SMiaohe Lin * fault mutex is held when add a hugetlb page 56283a5497a2SMiaohe Lin * to the page cache. So it's safe to call 56293a5497a2SMiaohe Lin * restore_reserve_on_error() here. 56303a5497a2SMiaohe Lin */ 56313a5497a2SMiaohe Lin restore_reserve_on_error(h, vma, haddr, page); 56326bda666aSChristoph Lameter put_page(page); 56336bda666aSChristoph Lameter goto out; 56346bda666aSChristoph Lameter } 5635c7b1850dSMike Kravetz new_pagecache_page = true; 563623be7468SMel Gorman } else { 56376bda666aSChristoph Lameter lock_page(page); 56380fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 56390fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 56400fe6e20bSNaoya Horiguchi goto backout_unlocked; 564123be7468SMel Gorman } 5642409eb8c2SHillf Danton anon_rmap = 1; 56430fe6e20bSNaoya Horiguchi } 56440fe6e20bSNaoya Horiguchi } else { 564557303d80SAndy Whitcroft /* 5646998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 5647998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 5648998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 5649fd6a03edSNaoya Horiguchi */ 5650fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 56510eb98f15SMiaohe Lin ret = VM_FAULT_HWPOISON_LARGE | 5652972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 5653fd6a03edSNaoya Horiguchi goto backout_unlocked; 56546bda666aSChristoph Lameter } 56557677f7fdSAxel Rasmussen 56567677f7fdSAxel Rasmussen /* Check for page in userfault range. */ 56577677f7fdSAxel Rasmussen if (userfaultfd_minor(vma)) { 56587677f7fdSAxel Rasmussen unlock_page(page); 56597677f7fdSAxel Rasmussen put_page(page); 56607677f7fdSAxel Rasmussen ret = hugetlb_handle_userfault(vma, mapping, idx, 5661824ddc60SNadav Amit flags, haddr, address, 56627677f7fdSAxel Rasmussen VM_UFFD_MINOR); 56637677f7fdSAxel Rasmussen goto out; 56647677f7fdSAxel Rasmussen } 5665998b4382SNaoya Horiguchi } 56661e8f889bSDavid Gibson 566757303d80SAndy Whitcroft /* 566857303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 566957303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 567057303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 567157303d80SAndy Whitcroft * the spinlock. 567257303d80SAndy Whitcroft */ 56735e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 5674285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 56752b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 56762b26736cSAndy Whitcroft goto backout_unlocked; 56772b26736cSAndy Whitcroft } 56785e911373SMike Kravetz /* Just decrements count, does not deallocate */ 5679285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 56805e911373SMike Kravetz } 568157303d80SAndy Whitcroft 56828bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 5683188a3972SMike Kravetz size = i_size_read(mapping->host) >> huge_page_shift(h); 5684188a3972SMike Kravetz if (idx >= size) 5685188a3972SMike Kravetz goto backout; 5686188a3972SMike Kravetz 568783c54070SNick Piggin ret = 0; 5688c64e912cSPeter Xu /* If pte changed from under us, retry */ 5689c64e912cSPeter Xu if (!pte_same(huge_ptep_get(ptep), old_pte)) 56904c887265SAdam Litke goto backout; 56914c887265SAdam Litke 569207443a85SJoonsoo Kim if (anon_rmap) { 5693d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 5694285b8dcaSHuang Ying hugepage_add_new_anon_rmap(page, vma, haddr); 5695ac714904SChoi Gi-yong } else 5696fb3d824dSDavid Hildenbrand page_dup_file_rmap(page, true); 56971e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 56981e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 5699c64e912cSPeter Xu /* 5700c64e912cSPeter Xu * If this pte was previously wr-protected, keep it wr-protected even 5701c64e912cSPeter Xu * if populated. 5702c64e912cSPeter Xu */ 5703c64e912cSPeter Xu if (unlikely(pte_marker_uffd_wp(old_pte))) 5704c64e912cSPeter Xu new_pte = huge_pte_wrprotect(huge_pte_mkuffd_wp(new_pte)); 5705285b8dcaSHuang Ying set_huge_pte_at(mm, haddr, ptep, new_pte); 57061e8f889bSDavid Gibson 57075d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 5708788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 57091e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 5710c89357e2SDavid Hildenbrand ret = hugetlb_wp(mm, vma, address, ptep, flags, page, ptl); 57111e8f889bSDavid Gibson } 57121e8f889bSDavid Gibson 5713cb900f41SKirill A. Shutemov spin_unlock(ptl); 5714cb6acd01SMike Kravetz 5715cb6acd01SMike Kravetz /* 57168f251a3dSMike Kravetz * Only set HPageMigratable in newly allocated pages. Existing pages 57178f251a3dSMike Kravetz * found in the pagecache may not have HPageMigratableset if they have 57188f251a3dSMike Kravetz * been isolated for migration. 5719cb6acd01SMike Kravetz */ 5720cb6acd01SMike Kravetz if (new_page) 57218f251a3dSMike Kravetz SetHPageMigratable(page); 5722cb6acd01SMike Kravetz 57234c887265SAdam Litke unlock_page(page); 57244c887265SAdam Litke out: 5725ac9b9c66SHugh Dickins return ret; 57264c887265SAdam Litke 57274c887265SAdam Litke backout: 5728cb900f41SKirill A. Shutemov spin_unlock(ptl); 57292b26736cSAndy Whitcroft backout_unlocked: 57304c887265SAdam Litke unlock_page(page); 5731c7b1850dSMike Kravetz /* restore reserve for newly allocated pages not in page cache */ 5732c7b1850dSMike Kravetz if (new_page && !new_pagecache_page) 5733285b8dcaSHuang Ying restore_reserve_on_error(h, vma, haddr, page); 57344c887265SAdam Litke put_page(page); 57354c887265SAdam Litke goto out; 5736ac9b9c66SHugh Dickins } 5737ac9b9c66SHugh Dickins 57388382d914SDavidlohr Bueso #ifdef CONFIG_SMP 5739188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 57408382d914SDavidlohr Bueso { 57418382d914SDavidlohr Bueso unsigned long key[2]; 57428382d914SDavidlohr Bueso u32 hash; 57438382d914SDavidlohr Bueso 57448382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 57458382d914SDavidlohr Bueso key[1] = idx; 57468382d914SDavidlohr Bueso 574755254636SMike Kravetz hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); 57488382d914SDavidlohr Bueso 57498382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 57508382d914SDavidlohr Bueso } 57518382d914SDavidlohr Bueso #else 57528382d914SDavidlohr Bueso /* 57536c26d310SMiaohe Lin * For uniprocessor systems we always use a single mutex, so just 57548382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 57558382d914SDavidlohr Bueso */ 5756188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 57578382d914SDavidlohr Bueso { 57588382d914SDavidlohr Bueso return 0; 57598382d914SDavidlohr Bueso } 57608382d914SDavidlohr Bueso #endif 57618382d914SDavidlohr Bueso 57622b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 5763788c7df4SHugh Dickins unsigned long address, unsigned int flags) 576486e5216fSAdam Litke { 57658382d914SDavidlohr Bueso pte_t *ptep, entry; 5766cb900f41SKirill A. Shutemov spinlock_t *ptl; 57672b740303SSouptick Joarder vm_fault_t ret; 57688382d914SDavidlohr Bueso u32 hash; 57698382d914SDavidlohr Bueso pgoff_t idx; 57700fe6e20bSNaoya Horiguchi struct page *page = NULL; 577157303d80SAndy Whitcroft struct page *pagecache_page = NULL; 5772a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 57738382d914SDavidlohr Bueso struct address_space *mapping; 57740f792cf9SNaoya Horiguchi int need_wait_lock = 0; 5775285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 577686e5216fSAdam Litke 5777285b8dcaSHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 5778fd6a03edSNaoya Horiguchi if (ptep) { 5779*40549ba8SMike Kravetz /* 5780*40549ba8SMike Kravetz * Since we hold no locks, ptep could be stale. That is 5781*40549ba8SMike Kravetz * OK as we are only making decisions based on content and 5782*40549ba8SMike Kravetz * not actually modifying content here. 5783*40549ba8SMike Kravetz */ 5784fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 5785290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 5786ad1ac596SMiaohe Lin migration_entry_wait_huge(vma, ptep); 5787290408d4SNaoya Horiguchi return 0; 5788290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 5789aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 5790972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 5791c0d0381aSMike Kravetz } 5792ddeaab32SMike Kravetz 57933935baa9SDavid Gibson /* 57943935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 57953935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 57963935baa9SDavid Gibson * the same page in the page cache. 57973935baa9SDavid Gibson */ 5798*40549ba8SMike Kravetz mapping = vma->vm_file->f_mapping; 5799*40549ba8SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 5800188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 5801c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 58028382d914SDavidlohr Bueso 5803*40549ba8SMike Kravetz /* 5804*40549ba8SMike Kravetz * Acquire vma lock before calling huge_pte_alloc and hold 5805*40549ba8SMike Kravetz * until finished with ptep. This prevents huge_pmd_unshare from 5806*40549ba8SMike Kravetz * being called elsewhere and making the ptep no longer valid. 5807*40549ba8SMike Kravetz * 5808*40549ba8SMike Kravetz * ptep could have already be assigned via huge_pte_offset. That 5809*40549ba8SMike Kravetz * is OK, as huge_pte_alloc will return the same value unless 5810*40549ba8SMike Kravetz * something has changed. 5811*40549ba8SMike Kravetz */ 5812*40549ba8SMike Kravetz hugetlb_vma_lock_read(vma); 5813*40549ba8SMike Kravetz ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); 5814*40549ba8SMike Kravetz if (!ptep) { 5815*40549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 5816*40549ba8SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5817*40549ba8SMike Kravetz return VM_FAULT_OOM; 5818*40549ba8SMike Kravetz } 5819*40549ba8SMike Kravetz 58207f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 5821c64e912cSPeter Xu /* PTE markers should be handled the same way as none pte */ 5822c64e912cSPeter Xu if (huge_pte_none_mostly(entry)) { 5823c64e912cSPeter Xu ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, 5824c64e912cSPeter Xu entry, flags); 5825b4d1d99fSDavid Gibson goto out_mutex; 58263935baa9SDavid Gibson } 582786e5216fSAdam Litke 582883c54070SNick Piggin ret = 0; 58291e8f889bSDavid Gibson 583057303d80SAndy Whitcroft /* 58310f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 58320f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 58337c8de358SEthon Paul * an active hugepage in pagecache. This goto expects the 2nd page 58347c8de358SEthon Paul * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will 58357c8de358SEthon Paul * properly handle it. 58360f792cf9SNaoya Horiguchi */ 58370f792cf9SNaoya Horiguchi if (!pte_present(entry)) 58380f792cf9SNaoya Horiguchi goto out_mutex; 58390f792cf9SNaoya Horiguchi 58400f792cf9SNaoya Horiguchi /* 5841c89357e2SDavid Hildenbrand * If we are going to COW/unshare the mapping later, we examine the 5842c89357e2SDavid Hildenbrand * pending reservations for this page now. This will ensure that any 584357303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 58441d8d1464SDavid Hildenbrand * spinlock. Also lookup the pagecache page now as it is used to 58451d8d1464SDavid Hildenbrand * determine if a reservation has been consumed. 584657303d80SAndy Whitcroft */ 5847c89357e2SDavid Hildenbrand if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) && 58481d8d1464SDavid Hildenbrand !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(entry)) { 5849285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 58502b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 5851b4d1d99fSDavid Gibson goto out_mutex; 58522b26736cSAndy Whitcroft } 58535e911373SMike Kravetz /* Just decrements count, does not deallocate */ 5854285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 585557303d80SAndy Whitcroft 585629be8426SMiaohe Lin pagecache_page = find_lock_page(mapping, idx); 585757303d80SAndy Whitcroft } 585857303d80SAndy Whitcroft 58590f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 58600fe6e20bSNaoya Horiguchi 5861c89357e2SDavid Hildenbrand /* Check for a racing update before calling hugetlb_wp() */ 5862b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 5863cb900f41SKirill A. Shutemov goto out_ptl; 5864b4d1d99fSDavid Gibson 5865166f3eccSPeter Xu /* Handle userfault-wp first, before trying to lock more pages */ 5866166f3eccSPeter Xu if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) && 5867166f3eccSPeter Xu (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 5868166f3eccSPeter Xu struct vm_fault vmf = { 5869166f3eccSPeter Xu .vma = vma, 5870166f3eccSPeter Xu .address = haddr, 5871166f3eccSPeter Xu .real_address = address, 5872166f3eccSPeter Xu .flags = flags, 5873166f3eccSPeter Xu }; 5874166f3eccSPeter Xu 5875166f3eccSPeter Xu spin_unlock(ptl); 5876166f3eccSPeter Xu if (pagecache_page) { 5877166f3eccSPeter Xu unlock_page(pagecache_page); 5878166f3eccSPeter Xu put_page(pagecache_page); 5879166f3eccSPeter Xu } 5880*40549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 5881166f3eccSPeter Xu mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5882166f3eccSPeter Xu return handle_userfault(&vmf, VM_UFFD_WP); 5883166f3eccSPeter Xu } 5884166f3eccSPeter Xu 58850f792cf9SNaoya Horiguchi /* 5886c89357e2SDavid Hildenbrand * hugetlb_wp() requires page locks of pte_page(entry) and 58870f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 58880f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 58890f792cf9SNaoya Horiguchi */ 58900f792cf9SNaoya Horiguchi page = pte_page(entry); 58910f792cf9SNaoya Horiguchi if (page != pagecache_page) 58920f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 58930f792cf9SNaoya Horiguchi need_wait_lock = 1; 58940f792cf9SNaoya Horiguchi goto out_ptl; 58950f792cf9SNaoya Horiguchi } 58960f792cf9SNaoya Horiguchi 58970f792cf9SNaoya Horiguchi get_page(page); 5898b4d1d99fSDavid Gibson 5899c89357e2SDavid Hildenbrand if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) { 5900106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 5901c89357e2SDavid Hildenbrand ret = hugetlb_wp(mm, vma, address, ptep, flags, 5902cb900f41SKirill A. Shutemov pagecache_page, ptl); 59030f792cf9SNaoya Horiguchi goto out_put_page; 5904c89357e2SDavid Hildenbrand } else if (likely(flags & FAULT_FLAG_WRITE)) { 5905106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 5906b4d1d99fSDavid Gibson } 5907c89357e2SDavid Hildenbrand } 5908b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 5909285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 5910788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 5911285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 59120f792cf9SNaoya Horiguchi out_put_page: 59130f792cf9SNaoya Horiguchi if (page != pagecache_page) 59140f792cf9SNaoya Horiguchi unlock_page(page); 59150f792cf9SNaoya Horiguchi put_page(page); 5916cb900f41SKirill A. Shutemov out_ptl: 5917cb900f41SKirill A. Shutemov spin_unlock(ptl); 591857303d80SAndy Whitcroft 591957303d80SAndy Whitcroft if (pagecache_page) { 592057303d80SAndy Whitcroft unlock_page(pagecache_page); 592157303d80SAndy Whitcroft put_page(pagecache_page); 592257303d80SAndy Whitcroft } 5923b4d1d99fSDavid Gibson out_mutex: 5924*40549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 5925c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 59260f792cf9SNaoya Horiguchi /* 59270f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 59280f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 59290f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 59300f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 59310f792cf9SNaoya Horiguchi * here without taking refcount. 59320f792cf9SNaoya Horiguchi */ 59330f792cf9SNaoya Horiguchi if (need_wait_lock) 59340f792cf9SNaoya Horiguchi wait_on_page_locked(page); 59351e8f889bSDavid Gibson return ret; 593686e5216fSAdam Litke } 593786e5216fSAdam Litke 5938714c1891SAxel Rasmussen #ifdef CONFIG_USERFAULTFD 59398fb5debcSMike Kravetz /* 59408fb5debcSMike Kravetz * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with 59418fb5debcSMike Kravetz * modifications for huge pages. 59428fb5debcSMike Kravetz */ 59438fb5debcSMike Kravetz int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, 59448fb5debcSMike Kravetz pte_t *dst_pte, 59458fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 59468fb5debcSMike Kravetz unsigned long dst_addr, 59478fb5debcSMike Kravetz unsigned long src_addr, 5948f6191471SAxel Rasmussen enum mcopy_atomic_mode mode, 59496041c691SPeter Xu struct page **pagep, 59506041c691SPeter Xu bool wp_copy) 59518fb5debcSMike Kravetz { 5952f6191471SAxel Rasmussen bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE); 59538cc5fcbbSMina Almasry struct hstate *h = hstate_vma(dst_vma); 59548cc5fcbbSMina Almasry struct address_space *mapping = dst_vma->vm_file->f_mapping; 59558cc5fcbbSMina Almasry pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr); 59561e392147SAndrea Arcangeli unsigned long size; 59571c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 59588fb5debcSMike Kravetz pte_t _dst_pte; 59598fb5debcSMike Kravetz spinlock_t *ptl; 59608cc5fcbbSMina Almasry int ret = -ENOMEM; 59618fb5debcSMike Kravetz struct page *page; 5962f6191471SAxel Rasmussen int writable; 5963cc30042dSMina Almasry bool page_in_pagecache = false; 59648fb5debcSMike Kravetz 5965f6191471SAxel Rasmussen if (is_continue) { 5966f6191471SAxel Rasmussen ret = -EFAULT; 5967f6191471SAxel Rasmussen page = find_lock_page(mapping, idx); 5968f6191471SAxel Rasmussen if (!page) 5969f6191471SAxel Rasmussen goto out; 5970cc30042dSMina Almasry page_in_pagecache = true; 5971f6191471SAxel Rasmussen } else if (!*pagep) { 5972d84cf06eSMina Almasry /* If a page already exists, then it's UFFDIO_COPY for 5973d84cf06eSMina Almasry * a non-missing case. Return -EEXIST. 5974d84cf06eSMina Almasry */ 5975d84cf06eSMina Almasry if (vm_shared && 5976d84cf06eSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 5977d84cf06eSMina Almasry ret = -EEXIST; 59788fb5debcSMike Kravetz goto out; 5979d84cf06eSMina Almasry } 5980d84cf06eSMina Almasry 5981d84cf06eSMina Almasry page = alloc_huge_page(dst_vma, dst_addr, 0); 5982d84cf06eSMina Almasry if (IS_ERR(page)) { 5983d84cf06eSMina Almasry ret = -ENOMEM; 5984d84cf06eSMina Almasry goto out; 5985d84cf06eSMina Almasry } 59868fb5debcSMike Kravetz 59878fb5debcSMike Kravetz ret = copy_huge_page_from_user(page, 59888fb5debcSMike Kravetz (const void __user *) src_addr, 5989810a56b9SMike Kravetz pages_per_huge_page(h), false); 59908fb5debcSMike Kravetz 5991c1e8d7c6SMichel Lespinasse /* fallback to copy_from_user outside mmap_lock */ 59928fb5debcSMike Kravetz if (unlikely(ret)) { 59939e368259SAndrea Arcangeli ret = -ENOENT; 59948cc5fcbbSMina Almasry /* Free the allocated page which may have 59958cc5fcbbSMina Almasry * consumed a reservation. 59968cc5fcbbSMina Almasry */ 59978cc5fcbbSMina Almasry restore_reserve_on_error(h, dst_vma, dst_addr, page); 59988cc5fcbbSMina Almasry put_page(page); 59998cc5fcbbSMina Almasry 60008cc5fcbbSMina Almasry /* Allocate a temporary page to hold the copied 60018cc5fcbbSMina Almasry * contents. 60028cc5fcbbSMina Almasry */ 60038cc5fcbbSMina Almasry page = alloc_huge_page_vma(h, dst_vma, dst_addr); 60048cc5fcbbSMina Almasry if (!page) { 60058cc5fcbbSMina Almasry ret = -ENOMEM; 60068cc5fcbbSMina Almasry goto out; 60078cc5fcbbSMina Almasry } 60088fb5debcSMike Kravetz *pagep = page; 60098cc5fcbbSMina Almasry /* Set the outparam pagep and return to the caller to 60108cc5fcbbSMina Almasry * copy the contents outside the lock. Don't free the 60118cc5fcbbSMina Almasry * page. 60128cc5fcbbSMina Almasry */ 60138fb5debcSMike Kravetz goto out; 60148fb5debcSMike Kravetz } 60158fb5debcSMike Kravetz } else { 60168cc5fcbbSMina Almasry if (vm_shared && 60178cc5fcbbSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 60188cc5fcbbSMina Almasry put_page(*pagep); 60198cc5fcbbSMina Almasry ret = -EEXIST; 60208cc5fcbbSMina Almasry *pagep = NULL; 60218cc5fcbbSMina Almasry goto out; 60228cc5fcbbSMina Almasry } 60238cc5fcbbSMina Almasry 60248cc5fcbbSMina Almasry page = alloc_huge_page(dst_vma, dst_addr, 0); 60258cc5fcbbSMina Almasry if (IS_ERR(page)) { 6026da9a298fSMiaohe Lin put_page(*pagep); 60278cc5fcbbSMina Almasry ret = -ENOMEM; 60288cc5fcbbSMina Almasry *pagep = NULL; 60298cc5fcbbSMina Almasry goto out; 60308cc5fcbbSMina Almasry } 603134892366SMuchun Song copy_user_huge_page(page, *pagep, dst_addr, dst_vma, 603234892366SMuchun Song pages_per_huge_page(h)); 60338cc5fcbbSMina Almasry put_page(*pagep); 60348fb5debcSMike Kravetz *pagep = NULL; 60358fb5debcSMike Kravetz } 60368fb5debcSMike Kravetz 60378fb5debcSMike Kravetz /* 60388fb5debcSMike Kravetz * The memory barrier inside __SetPageUptodate makes sure that 60398fb5debcSMike Kravetz * preceding stores to the page contents become visible before 60408fb5debcSMike Kravetz * the set_pte_at() write. 60418fb5debcSMike Kravetz */ 60428fb5debcSMike Kravetz __SetPageUptodate(page); 60438fb5debcSMike Kravetz 6044f6191471SAxel Rasmussen /* Add shared, newly allocated pages to the page cache. */ 6045f6191471SAxel Rasmussen if (vm_shared && !is_continue) { 60461e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 60471e392147SAndrea Arcangeli ret = -EFAULT; 60481e392147SAndrea Arcangeli if (idx >= size) 60491e392147SAndrea Arcangeli goto out_release_nounlock; 60501c9e8defSMike Kravetz 60511e392147SAndrea Arcangeli /* 60521e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 60537e1813d4SMike Kravetz * hugetlb_add_to_page_cache() below happens through the 60541e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 60551e392147SAndrea Arcangeli * the caller. 60561e392147SAndrea Arcangeli */ 60577e1813d4SMike Kravetz ret = hugetlb_add_to_page_cache(page, mapping, idx); 60581c9e8defSMike Kravetz if (ret) 60591c9e8defSMike Kravetz goto out_release_nounlock; 6060cc30042dSMina Almasry page_in_pagecache = true; 60611c9e8defSMike Kravetz } 60621c9e8defSMike Kravetz 6063bcc66543SMiaohe Lin ptl = huge_pte_lock(h, dst_mm, dst_pte); 60648fb5debcSMike Kravetz 60651e392147SAndrea Arcangeli /* 60661e392147SAndrea Arcangeli * Recheck the i_size after holding PT lock to make sure not 60671e392147SAndrea Arcangeli * to leave any page mapped (as page_mapped()) beyond the end 60681e392147SAndrea Arcangeli * of the i_size (remove_inode_hugepages() is strict about 60691e392147SAndrea Arcangeli * enforcing that). If we bail out here, we'll also leave a 60701e392147SAndrea Arcangeli * page in the radix tree in the vm_shared case beyond the end 60711e392147SAndrea Arcangeli * of the i_size, but remove_inode_hugepages() will take care 60721e392147SAndrea Arcangeli * of it as soon as we drop the hugetlb_fault_mutex_table. 60731e392147SAndrea Arcangeli */ 60741e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 60751e392147SAndrea Arcangeli ret = -EFAULT; 60761e392147SAndrea Arcangeli if (idx >= size) 60771e392147SAndrea Arcangeli goto out_release_unlock; 60781e392147SAndrea Arcangeli 60798fb5debcSMike Kravetz ret = -EEXIST; 60806041c691SPeter Xu /* 60816041c691SPeter Xu * We allow to overwrite a pte marker: consider when both MISSING|WP 60826041c691SPeter Xu * registered, we firstly wr-protect a none pte which has no page cache 60836041c691SPeter Xu * page backing it, then access the page. 60846041c691SPeter Xu */ 60856041c691SPeter Xu if (!huge_pte_none_mostly(huge_ptep_get(dst_pte))) 60868fb5debcSMike Kravetz goto out_release_unlock; 60878fb5debcSMike Kravetz 6088ab74ef70SMiaohe Lin if (page_in_pagecache) { 6089fb3d824dSDavid Hildenbrand page_dup_file_rmap(page, true); 60901c9e8defSMike Kravetz } else { 6091d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 60928fb5debcSMike Kravetz hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); 60931c9e8defSMike Kravetz } 60948fb5debcSMike Kravetz 60956041c691SPeter Xu /* 60966041c691SPeter Xu * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY 60976041c691SPeter Xu * with wp flag set, don't set pte write bit. 60986041c691SPeter Xu */ 60996041c691SPeter Xu if (wp_copy || (is_continue && !vm_shared)) 6100f6191471SAxel Rasmussen writable = 0; 6101f6191471SAxel Rasmussen else 6102f6191471SAxel Rasmussen writable = dst_vma->vm_flags & VM_WRITE; 6103f6191471SAxel Rasmussen 6104f6191471SAxel Rasmussen _dst_pte = make_huge_pte(dst_vma, page, writable); 61056041c691SPeter Xu /* 61066041c691SPeter Xu * Always mark UFFDIO_COPY page dirty; note that this may not be 61076041c691SPeter Xu * extremely important for hugetlbfs for now since swapping is not 61086041c691SPeter Xu * supported, but we should still be clear in that this page cannot be 61096041c691SPeter Xu * thrown away at will, even if write bit not set. 61106041c691SPeter Xu */ 61118fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 61128fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 61138fb5debcSMike Kravetz 61146041c691SPeter Xu if (wp_copy) 61156041c691SPeter Xu _dst_pte = huge_pte_mkuffd_wp(_dst_pte); 61166041c691SPeter Xu 61178fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 61188fb5debcSMike Kravetz 61198fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 61208fb5debcSMike Kravetz 61218fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 61228fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 61238fb5debcSMike Kravetz 61248fb5debcSMike Kravetz spin_unlock(ptl); 6125f6191471SAxel Rasmussen if (!is_continue) 61268f251a3dSMike Kravetz SetHPageMigratable(page); 6127f6191471SAxel Rasmussen if (vm_shared || is_continue) 61281c9e8defSMike Kravetz unlock_page(page); 61298fb5debcSMike Kravetz ret = 0; 61308fb5debcSMike Kravetz out: 61318fb5debcSMike Kravetz return ret; 61328fb5debcSMike Kravetz out_release_unlock: 61338fb5debcSMike Kravetz spin_unlock(ptl); 6134f6191471SAxel Rasmussen if (vm_shared || is_continue) 61351c9e8defSMike Kravetz unlock_page(page); 61365af10dfdSAndrea Arcangeli out_release_nounlock: 6137cc30042dSMina Almasry if (!page_in_pagecache) 6138846be085SMike Kravetz restore_reserve_on_error(h, dst_vma, dst_addr, page); 61398fb5debcSMike Kravetz put_page(page); 61408fb5debcSMike Kravetz goto out; 61418fb5debcSMike Kravetz } 6142714c1891SAxel Rasmussen #endif /* CONFIG_USERFAULTFD */ 61438fb5debcSMike Kravetz 614482e5d378SJoao Martins static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma, 614582e5d378SJoao Martins int refs, struct page **pages, 614682e5d378SJoao Martins struct vm_area_struct **vmas) 614782e5d378SJoao Martins { 614882e5d378SJoao Martins int nr; 614982e5d378SJoao Martins 615082e5d378SJoao Martins for (nr = 0; nr < refs; nr++) { 615182e5d378SJoao Martins if (likely(pages)) 615214455eabSCheng Li pages[nr] = nth_page(page, nr); 615382e5d378SJoao Martins if (vmas) 615482e5d378SJoao Martins vmas[nr] = vma; 615582e5d378SJoao Martins } 615682e5d378SJoao Martins } 615782e5d378SJoao Martins 6158a7f22660SDavid Hildenbrand static inline bool __follow_hugetlb_must_fault(unsigned int flags, pte_t *pte, 6159a7f22660SDavid Hildenbrand bool *unshare) 6160a7f22660SDavid Hildenbrand { 6161a7f22660SDavid Hildenbrand pte_t pteval = huge_ptep_get(pte); 6162a7f22660SDavid Hildenbrand 6163a7f22660SDavid Hildenbrand *unshare = false; 6164a7f22660SDavid Hildenbrand if (is_swap_pte(pteval)) 6165a7f22660SDavid Hildenbrand return true; 6166a7f22660SDavid Hildenbrand if (huge_pte_write(pteval)) 6167a7f22660SDavid Hildenbrand return false; 6168a7f22660SDavid Hildenbrand if (flags & FOLL_WRITE) 6169a7f22660SDavid Hildenbrand return true; 6170a7f22660SDavid Hildenbrand if (gup_must_unshare(flags, pte_page(pteval))) { 6171a7f22660SDavid Hildenbrand *unshare = true; 6172a7f22660SDavid Hildenbrand return true; 6173a7f22660SDavid Hildenbrand } 6174a7f22660SDavid Hildenbrand return false; 6175a7f22660SDavid Hildenbrand } 6176a7f22660SDavid Hildenbrand 617728a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 617863551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 617928a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 61804f6da934SPeter Xu long i, unsigned int flags, int *locked) 618163551ae0SDavid Gibson { 6182d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 6183d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 618428a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 6185a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 61860fa5bc40SJoao Martins int err = -EFAULT, refs; 618763551ae0SDavid Gibson 618863551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 618963551ae0SDavid Gibson pte_t *pte; 6190cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 6191a7f22660SDavid Hildenbrand bool unshare = false; 61922a15efc9SHugh Dickins int absent; 619363551ae0SDavid Gibson struct page *page; 619463551ae0SDavid Gibson 61954c887265SAdam Litke /* 619602057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 619702057967SDavid Rientjes * potentially allocating memory. 619802057967SDavid Rientjes */ 6199fa45f116SDavidlohr Bueso if (fatal_signal_pending(current)) { 620002057967SDavid Rientjes remainder = 0; 620102057967SDavid Rientjes break; 620202057967SDavid Rientjes } 620302057967SDavid Rientjes 620402057967SDavid Rientjes /* 62054c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 62062a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 62074c887265SAdam Litke * first, for the page indexing below to work. 6208cb900f41SKirill A. Shutemov * 6209cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 62104c887265SAdam Litke */ 62117868a208SPunit Agrawal pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), 62127868a208SPunit Agrawal huge_page_size(h)); 6213cb900f41SKirill A. Shutemov if (pte) 6214cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 62152a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 621663551ae0SDavid Gibson 62172a15efc9SHugh Dickins /* 62182a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 62193ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 62203ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 62213ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 62223ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 62232a15efc9SHugh Dickins */ 62243ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 62253ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 6226cb900f41SKirill A. Shutemov if (pte) 6227cb900f41SKirill A. Shutemov spin_unlock(ptl); 62282a15efc9SHugh Dickins remainder = 0; 62292a15efc9SHugh Dickins break; 62302a15efc9SHugh Dickins } 62312a15efc9SHugh Dickins 62329cc3a5bdSNaoya Horiguchi /* 62339cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 62349cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 62359cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 62369cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 62379cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 62389cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 62399cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 62409cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 62419cc3a5bdSNaoya Horiguchi */ 6242a7f22660SDavid Hildenbrand if (absent || 6243a7f22660SDavid Hildenbrand __follow_hugetlb_must_fault(flags, pte, &unshare)) { 62442b740303SSouptick Joarder vm_fault_t ret; 624587ffc118SAndrea Arcangeli unsigned int fault_flags = 0; 62464c887265SAdam Litke 6247cb900f41SKirill A. Shutemov if (pte) 6248cb900f41SKirill A. Shutemov spin_unlock(ptl); 624987ffc118SAndrea Arcangeli if (flags & FOLL_WRITE) 625087ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_WRITE; 6251a7f22660SDavid Hildenbrand else if (unshare) 6252a7f22660SDavid Hildenbrand fault_flags |= FAULT_FLAG_UNSHARE; 62534f6da934SPeter Xu if (locked) 625471335f37SPeter Xu fault_flags |= FAULT_FLAG_ALLOW_RETRY | 625571335f37SPeter Xu FAULT_FLAG_KILLABLE; 625687ffc118SAndrea Arcangeli if (flags & FOLL_NOWAIT) 625787ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY | 625887ffc118SAndrea Arcangeli FAULT_FLAG_RETRY_NOWAIT; 625987ffc118SAndrea Arcangeli if (flags & FOLL_TRIED) { 62604426e945SPeter Xu /* 62614426e945SPeter Xu * Note: FAULT_FLAG_ALLOW_RETRY and 62624426e945SPeter Xu * FAULT_FLAG_TRIED can co-exist 62634426e945SPeter Xu */ 626487ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_TRIED; 626587ffc118SAndrea Arcangeli } 626687ffc118SAndrea Arcangeli ret = hugetlb_fault(mm, vma, vaddr, fault_flags); 626787ffc118SAndrea Arcangeli if (ret & VM_FAULT_ERROR) { 62682be7cfedSDaniel Jordan err = vm_fault_to_errno(ret, flags); 62691c59827dSHugh Dickins remainder = 0; 62701c59827dSHugh Dickins break; 62711c59827dSHugh Dickins } 627287ffc118SAndrea Arcangeli if (ret & VM_FAULT_RETRY) { 62734f6da934SPeter Xu if (locked && 62741ac25013SAndrea Arcangeli !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) 62754f6da934SPeter Xu *locked = 0; 627687ffc118SAndrea Arcangeli *nr_pages = 0; 627787ffc118SAndrea Arcangeli /* 627887ffc118SAndrea Arcangeli * VM_FAULT_RETRY must not return an 627987ffc118SAndrea Arcangeli * error, it will return zero 628087ffc118SAndrea Arcangeli * instead. 628187ffc118SAndrea Arcangeli * 628287ffc118SAndrea Arcangeli * No need to update "position" as the 628387ffc118SAndrea Arcangeli * caller will not check it after 628487ffc118SAndrea Arcangeli * *nr_pages is set to 0. 628587ffc118SAndrea Arcangeli */ 628687ffc118SAndrea Arcangeli return i; 628787ffc118SAndrea Arcangeli } 628887ffc118SAndrea Arcangeli continue; 628987ffc118SAndrea Arcangeli } 629063551ae0SDavid Gibson 6291a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 62927f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 62938fde12caSLinus Torvalds 6294b6a2619cSDavid Hildenbrand VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) && 6295b6a2619cSDavid Hildenbrand !PageAnonExclusive(page), page); 6296b6a2619cSDavid Hildenbrand 62978fde12caSLinus Torvalds /* 6298acbfb087SZhigang Lu * If subpage information not requested, update counters 6299acbfb087SZhigang Lu * and skip the same_page loop below. 6300acbfb087SZhigang Lu */ 6301acbfb087SZhigang Lu if (!pages && !vmas && !pfn_offset && 6302acbfb087SZhigang Lu (vaddr + huge_page_size(h) < vma->vm_end) && 6303acbfb087SZhigang Lu (remainder >= pages_per_huge_page(h))) { 6304acbfb087SZhigang Lu vaddr += huge_page_size(h); 6305acbfb087SZhigang Lu remainder -= pages_per_huge_page(h); 6306acbfb087SZhigang Lu i += pages_per_huge_page(h); 6307acbfb087SZhigang Lu spin_unlock(ptl); 6308acbfb087SZhigang Lu continue; 6309acbfb087SZhigang Lu } 6310acbfb087SZhigang Lu 6311d08af0a5SJoao Martins /* vaddr may not be aligned to PAGE_SIZE */ 6312d08af0a5SJoao Martins refs = min3(pages_per_huge_page(h) - pfn_offset, remainder, 6313d08af0a5SJoao Martins (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT); 63140fa5bc40SJoao Martins 631582e5d378SJoao Martins if (pages || vmas) 631614455eabSCheng Li record_subpages_vmas(nth_page(page, pfn_offset), 631782e5d378SJoao Martins vma, refs, 631882e5d378SJoao Martins likely(pages) ? pages + i : NULL, 631982e5d378SJoao Martins vmas ? vmas + i : NULL); 632063551ae0SDavid Gibson 632182e5d378SJoao Martins if (pages) { 63220fa5bc40SJoao Martins /* 6323822951d8SMatthew Wilcox (Oracle) * try_grab_folio() should always succeed here, 63240fa5bc40SJoao Martins * because: a) we hold the ptl lock, and b) we've just 63250fa5bc40SJoao Martins * checked that the huge page is present in the page 63260fa5bc40SJoao Martins * tables. If the huge page is present, then the tail 63270fa5bc40SJoao Martins * pages must also be present. The ptl prevents the 63280fa5bc40SJoao Martins * head page and tail pages from being rearranged in 63290fa5bc40SJoao Martins * any way. So this page must be available at this 63300fa5bc40SJoao Martins * point, unless the page refcount overflowed: 63310fa5bc40SJoao Martins */ 6332822951d8SMatthew Wilcox (Oracle) if (WARN_ON_ONCE(!try_grab_folio(pages[i], refs, 63330fa5bc40SJoao Martins flags))) { 63340fa5bc40SJoao Martins spin_unlock(ptl); 63350fa5bc40SJoao Martins remainder = 0; 63360fa5bc40SJoao Martins err = -ENOMEM; 63370fa5bc40SJoao Martins break; 63380fa5bc40SJoao Martins } 6339d5d4b0aaSChen, Kenneth W } 634082e5d378SJoao Martins 634182e5d378SJoao Martins vaddr += (refs << PAGE_SHIFT); 634282e5d378SJoao Martins remainder -= refs; 634382e5d378SJoao Martins i += refs; 634482e5d378SJoao Martins 6345cb900f41SKirill A. Shutemov spin_unlock(ptl); 634663551ae0SDavid Gibson } 634728a35716SMichel Lespinasse *nr_pages = remainder; 634887ffc118SAndrea Arcangeli /* 634987ffc118SAndrea Arcangeli * setting position is actually required only if remainder is 635087ffc118SAndrea Arcangeli * not zero but it's faster not to add a "if (remainder)" 635187ffc118SAndrea Arcangeli * branch. 635287ffc118SAndrea Arcangeli */ 635363551ae0SDavid Gibson *position = vaddr; 635463551ae0SDavid Gibson 63552be7cfedSDaniel Jordan return i ? i : err; 635663551ae0SDavid Gibson } 63578f860591SZhang, Yanmin 63587da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 63595a90d5a1SPeter Xu unsigned long address, unsigned long end, 63605a90d5a1SPeter Xu pgprot_t newprot, unsigned long cp_flags) 63618f860591SZhang, Yanmin { 63628f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 63638f860591SZhang, Yanmin unsigned long start = address; 63648f860591SZhang, Yanmin pte_t *ptep; 63658f860591SZhang, Yanmin pte_t pte; 6366a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 636760dfaad6SPeter Xu unsigned long pages = 0, psize = huge_page_size(h); 6368dff11abeSMike Kravetz bool shared_pmd = false; 6369ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 6370e95a9851SMike Kravetz unsigned long last_addr_mask; 63715a90d5a1SPeter Xu bool uffd_wp = cp_flags & MM_CP_UFFD_WP; 63725a90d5a1SPeter Xu bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE; 6373dff11abeSMike Kravetz 6374dff11abeSMike Kravetz /* 6375dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 6376ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 6377dff11abeSMike Kravetz * range if PMD sharing is possible. 6378dff11abeSMike Kravetz */ 63797269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 63807269f999SJérôme Glisse 0, vma, mm, start, end); 6381ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 63828f860591SZhang, Yanmin 63838f860591SZhang, Yanmin BUG_ON(address >= end); 6384ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 63858f860591SZhang, Yanmin 6386ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 6387*40549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 638883cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 6389*40549ba8SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 639060dfaad6SPeter Xu for (; address < end; address += psize) { 6391cb900f41SKirill A. Shutemov spinlock_t *ptl; 639260dfaad6SPeter Xu ptep = huge_pte_offset(mm, address, psize); 6393e95a9851SMike Kravetz if (!ptep) { 6394e95a9851SMike Kravetz address |= last_addr_mask; 63958f860591SZhang, Yanmin continue; 6396e95a9851SMike Kravetz } 6397cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 63984ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, address, ptep)) { 639960dfaad6SPeter Xu /* 640060dfaad6SPeter Xu * When uffd-wp is enabled on the vma, unshare 640160dfaad6SPeter Xu * shouldn't happen at all. Warn about it if it 640260dfaad6SPeter Xu * happened due to some reason. 640360dfaad6SPeter Xu */ 640460dfaad6SPeter Xu WARN_ON_ONCE(uffd_wp || uffd_wp_resolve); 64057da4d641SPeter Zijlstra pages++; 6406cb900f41SKirill A. Shutemov spin_unlock(ptl); 6407dff11abeSMike Kravetz shared_pmd = true; 64084ddb4d91SMike Kravetz address |= last_addr_mask; 640939dde65cSChen, Kenneth W continue; 64107da4d641SPeter Zijlstra } 6411a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 6412a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 6413a8bda28dSNaoya Horiguchi spin_unlock(ptl); 6414a8bda28dSNaoya Horiguchi continue; 6415a8bda28dSNaoya Horiguchi } 6416a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 6417a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 64186c287605SDavid Hildenbrand struct page *page = pfn_swap_entry_to_page(entry); 6419a8bda28dSNaoya Horiguchi 64206c287605SDavid Hildenbrand if (!is_readable_migration_entry(entry)) { 6421a8bda28dSNaoya Horiguchi pte_t newpte; 6422a8bda28dSNaoya Horiguchi 64236c287605SDavid Hildenbrand if (PageAnon(page)) 64246c287605SDavid Hildenbrand entry = make_readable_exclusive_migration_entry( 64256c287605SDavid Hildenbrand swp_offset(entry)); 64266c287605SDavid Hildenbrand else 64274dd845b5SAlistair Popple entry = make_readable_migration_entry( 64284dd845b5SAlistair Popple swp_offset(entry)); 6429a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 64305a90d5a1SPeter Xu if (uffd_wp) 64315a90d5a1SPeter Xu newpte = pte_swp_mkuffd_wp(newpte); 64325a90d5a1SPeter Xu else if (uffd_wp_resolve) 64335a90d5a1SPeter Xu newpte = pte_swp_clear_uffd_wp(newpte); 643418f39629SQi Zheng set_huge_pte_at(mm, address, ptep, newpte); 6435a8bda28dSNaoya Horiguchi pages++; 6436a8bda28dSNaoya Horiguchi } 6437a8bda28dSNaoya Horiguchi spin_unlock(ptl); 6438a8bda28dSNaoya Horiguchi continue; 6439a8bda28dSNaoya Horiguchi } 644060dfaad6SPeter Xu if (unlikely(pte_marker_uffd_wp(pte))) { 644160dfaad6SPeter Xu /* 644260dfaad6SPeter Xu * This is changing a non-present pte into a none pte, 644360dfaad6SPeter Xu * no need for huge_ptep_modify_prot_start/commit(). 644460dfaad6SPeter Xu */ 644560dfaad6SPeter Xu if (uffd_wp_resolve) 644660dfaad6SPeter Xu huge_pte_clear(mm, address, ptep, psize); 644760dfaad6SPeter Xu } 6448a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 6449023bdd00SAneesh Kumar K.V pte_t old_pte; 645079c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 6451023bdd00SAneesh Kumar K.V 6452023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 645316785bd7SAnshuman Khandual pte = huge_pte_modify(old_pte, newprot); 645479c1c594SChristophe Leroy pte = arch_make_huge_pte(pte, shift, vma->vm_flags); 64555a90d5a1SPeter Xu if (uffd_wp) 64565a90d5a1SPeter Xu pte = huge_pte_mkuffd_wp(huge_pte_wrprotect(pte)); 64575a90d5a1SPeter Xu else if (uffd_wp_resolve) 64585a90d5a1SPeter Xu pte = huge_pte_clear_uffd_wp(pte); 6459023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 64607da4d641SPeter Zijlstra pages++; 646160dfaad6SPeter Xu } else { 646260dfaad6SPeter Xu /* None pte */ 646360dfaad6SPeter Xu if (unlikely(uffd_wp)) 646460dfaad6SPeter Xu /* Safe to modify directly (none->non-present). */ 646560dfaad6SPeter Xu set_huge_pte_at(mm, address, ptep, 646660dfaad6SPeter Xu make_pte_marker(PTE_MARKER_UFFD_WP)); 64678f860591SZhang, Yanmin } 6468cb900f41SKirill A. Shutemov spin_unlock(ptl); 64698f860591SZhang, Yanmin } 6470d833352aSMel Gorman /* 6471c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 6472d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 6473c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 6474dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 6475dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 6476d833352aSMel Gorman */ 6477dff11abeSMike Kravetz if (shared_pmd) 6478ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 6479dff11abeSMike Kravetz else 64805491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 64810f10851eSJérôme Glisse /* 64820f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 64830f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 64840f10851eSJérôme Glisse * 6485ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 64860f10851eSJérôme Glisse */ 648783cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 6488*40549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 6489ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 64907da4d641SPeter Zijlstra 64917da4d641SPeter Zijlstra return pages << h->order; 64928f860591SZhang, Yanmin } 64938f860591SZhang, Yanmin 649433b8f84aSMike Kravetz /* Return true if reservation was successful, false otherwise. */ 649533b8f84aSMike Kravetz bool hugetlb_reserve_pages(struct inode *inode, 6496a1e78772SMel Gorman long from, long to, 64975a6fe125SMel Gorman struct vm_area_struct *vma, 6498ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 6499e4e574b7SAdam Litke { 650033b8f84aSMike Kravetz long chg, add = -1; 6501a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 650290481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 65039119a41eSJoonsoo Kim struct resv_map *resv_map; 6504075a61d0SMina Almasry struct hugetlb_cgroup *h_cg = NULL; 65050db9d74eSMina Almasry long gbl_reserve, regions_needed = 0; 6506e4e574b7SAdam Litke 650763489f8eSMike Kravetz /* This should never happen */ 650863489f8eSMike Kravetz if (from > to) { 650963489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 651033b8f84aSMike Kravetz return false; 651163489f8eSMike Kravetz } 651263489f8eSMike Kravetz 6513a1e78772SMel Gorman /* 65148d9bfb26SMike Kravetz * vma specific semaphore used for pmd sharing synchronization 65158d9bfb26SMike Kravetz */ 65168d9bfb26SMike Kravetz hugetlb_vma_lock_alloc(vma); 65178d9bfb26SMike Kravetz 65188d9bfb26SMike Kravetz /* 651917c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 652017c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 652190481622SDavid Gibson * without using reserves 652217c9d12eSMel Gorman */ 6523ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 652433b8f84aSMike Kravetz return true; 652517c9d12eSMel Gorman 652617c9d12eSMel Gorman /* 6527a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 6528a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 6529a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 6530a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 6531a1e78772SMel Gorman */ 65329119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 6533f27a5136SMike Kravetz /* 6534f27a5136SMike Kravetz * resv_map can not be NULL as hugetlb_reserve_pages is only 6535f27a5136SMike Kravetz * called for inodes for which resv_maps were created (see 6536f27a5136SMike Kravetz * hugetlbfs_get_inode). 6537f27a5136SMike Kravetz */ 65384e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 65399119a41eSJoonsoo Kim 65400db9d74eSMina Almasry chg = region_chg(resv_map, from, to, ®ions_needed); 65419119a41eSJoonsoo Kim } else { 6542e9fe92aeSMina Almasry /* Private mapping. */ 65439119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 65445a6fe125SMel Gorman if (!resv_map) 65458d9bfb26SMike Kravetz goto out_err; 65465a6fe125SMel Gorman 654717c9d12eSMel Gorman chg = to - from; 654817c9d12eSMel Gorman 65495a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 65505a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 65515a6fe125SMel Gorman } 65525a6fe125SMel Gorman 655333b8f84aSMike Kravetz if (chg < 0) 6554c50ac050SDave Hansen goto out_err; 655517c9d12eSMel Gorman 655633b8f84aSMike Kravetz if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), 655733b8f84aSMike Kravetz chg * pages_per_huge_page(h), &h_cg) < 0) 6558075a61d0SMina Almasry goto out_err; 6559075a61d0SMina Almasry 6560075a61d0SMina Almasry if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { 6561075a61d0SMina Almasry /* For private mappings, the hugetlb_cgroup uncharge info hangs 6562075a61d0SMina Almasry * of the resv_map. 6563075a61d0SMina Almasry */ 6564075a61d0SMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); 6565075a61d0SMina Almasry } 6566075a61d0SMina Almasry 65671c5ecae3SMike Kravetz /* 65681c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 65691c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 65701c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 65711c5ecae3SMike Kravetz */ 65721c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 657333b8f84aSMike Kravetz if (gbl_reserve < 0) 6574075a61d0SMina Almasry goto out_uncharge_cgroup; 657517c9d12eSMel Gorman 657617c9d12eSMel Gorman /* 657717c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 657890481622SDavid Gibson * Hand the pages back to the subpool if there are not 657917c9d12eSMel Gorman */ 658033b8f84aSMike Kravetz if (hugetlb_acct_memory(h, gbl_reserve) < 0) 6581075a61d0SMina Almasry goto out_put_pages; 658217c9d12eSMel Gorman 658317c9d12eSMel Gorman /* 658417c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 658517c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 658617c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 658717c9d12eSMel Gorman * the reservation was and the page cache tells how much of 658817c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 658917c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 659017c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 659117c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 659217c9d12eSMel Gorman * else has to be done for private mappings here 659317c9d12eSMel Gorman */ 659433039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 6595075a61d0SMina Almasry add = region_add(resv_map, from, to, regions_needed, h, h_cg); 659633039678SMike Kravetz 65970db9d74eSMina Almasry if (unlikely(add < 0)) { 65980db9d74eSMina Almasry hugetlb_acct_memory(h, -gbl_reserve); 6599075a61d0SMina Almasry goto out_put_pages; 66000db9d74eSMina Almasry } else if (unlikely(chg > add)) { 660133039678SMike Kravetz /* 660233039678SMike Kravetz * pages in this range were added to the reserve 660333039678SMike Kravetz * map between region_chg and region_add. This 660433039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 660533039678SMike Kravetz * the subpool and reserve counts modified above 660633039678SMike Kravetz * based on the difference. 660733039678SMike Kravetz */ 660833039678SMike Kravetz long rsv_adjust; 660933039678SMike Kravetz 6610d85aecf2SMiaohe Lin /* 6611d85aecf2SMiaohe Lin * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the 6612d85aecf2SMiaohe Lin * reference to h_cg->css. See comment below for detail. 6613d85aecf2SMiaohe Lin */ 6614075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd( 6615075a61d0SMina Almasry hstate_index(h), 6616075a61d0SMina Almasry (chg - add) * pages_per_huge_page(h), h_cg); 6617075a61d0SMina Almasry 661833039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 661933039678SMike Kravetz chg - add); 662033039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 6621d85aecf2SMiaohe Lin } else if (h_cg) { 6622d85aecf2SMiaohe Lin /* 6623d85aecf2SMiaohe Lin * The file_regions will hold their own reference to 6624d85aecf2SMiaohe Lin * h_cg->css. So we should release the reference held 6625d85aecf2SMiaohe Lin * via hugetlb_cgroup_charge_cgroup_rsvd() when we are 6626d85aecf2SMiaohe Lin * done. 6627d85aecf2SMiaohe Lin */ 6628d85aecf2SMiaohe Lin hugetlb_cgroup_put_rsvd_cgroup(h_cg); 662933039678SMike Kravetz } 663033039678SMike Kravetz } 663133b8f84aSMike Kravetz return true; 663233b8f84aSMike Kravetz 6633075a61d0SMina Almasry out_put_pages: 6634075a61d0SMina Almasry /* put back original number of pages, chg */ 6635075a61d0SMina Almasry (void)hugepage_subpool_put_pages(spool, chg); 6636075a61d0SMina Almasry out_uncharge_cgroup: 6637075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), 6638075a61d0SMina Almasry chg * pages_per_huge_page(h), h_cg); 6639c50ac050SDave Hansen out_err: 66408d9bfb26SMike Kravetz hugetlb_vma_lock_free(vma); 66415e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 66420db9d74eSMina Almasry /* Only call region_abort if the region_chg succeeded but the 66430db9d74eSMina Almasry * region_add failed or didn't run. 66440db9d74eSMina Almasry */ 66450db9d74eSMina Almasry if (chg >= 0 && add < 0) 66460db9d74eSMina Almasry region_abort(resv_map, from, to, regions_needed); 6647f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 6648f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 664933b8f84aSMike Kravetz return false; 6650a43a8c39SChen, Kenneth W } 6651a43a8c39SChen, Kenneth W 6652b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 6653b5cec28dSMike Kravetz long freed) 6654a43a8c39SChen, Kenneth W { 6655a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 66564e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 66579119a41eSJoonsoo Kim long chg = 0; 665890481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 66591c5ecae3SMike Kravetz long gbl_reserve; 666045c682a6SKen Chen 6661f27a5136SMike Kravetz /* 6662f27a5136SMike Kravetz * Since this routine can be called in the evict inode path for all 6663f27a5136SMike Kravetz * hugetlbfs inodes, resv_map could be NULL. 6664f27a5136SMike Kravetz */ 6665b5cec28dSMike Kravetz if (resv_map) { 6666b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 6667b5cec28dSMike Kravetz /* 6668b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 6669b5cec28dSMike Kravetz * must be split and another region descriptor can not be 6670b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 6671b5cec28dSMike Kravetz */ 6672b5cec28dSMike Kravetz if (chg < 0) 6673b5cec28dSMike Kravetz return chg; 6674b5cec28dSMike Kravetz } 6675b5cec28dSMike Kravetz 667645c682a6SKen Chen spin_lock(&inode->i_lock); 6677e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 667845c682a6SKen Chen spin_unlock(&inode->i_lock); 667945c682a6SKen Chen 66801c5ecae3SMike Kravetz /* 66811c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 66821c5ecae3SMike Kravetz * reservations to be released may be adjusted. 6683dddf31a4SMiaohe Lin * 6684dddf31a4SMiaohe Lin * Note that !resv_map implies freed == 0. So (chg - freed) 6685dddf31a4SMiaohe Lin * won't go negative. 66861c5ecae3SMike Kravetz */ 66871c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 66881c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 6689b5cec28dSMike Kravetz 6690b5cec28dSMike Kravetz return 0; 6691a43a8c39SChen, Kenneth W } 669293f70f90SNaoya Horiguchi 66933212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 66943212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 66953212b535SSteve Capper struct vm_area_struct *vma, 66963212b535SSteve Capper unsigned long addr, pgoff_t idx) 66973212b535SSteve Capper { 66983212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 66993212b535SSteve Capper svma->vm_start; 67003212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 67013212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 67023212b535SSteve Capper 67033212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 6704de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 6705de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 67063212b535SSteve Capper 67073212b535SSteve Capper /* 67083212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 67093212b535SSteve Capper * page table page. 67103212b535SSteve Capper */ 67113212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 67123212b535SSteve Capper vm_flags != svm_flags || 671307e51edfSMiaohe Lin !range_in_vma(svma, sbase, s_end)) 67143212b535SSteve Capper return 0; 67153212b535SSteve Capper 67163212b535SSteve Capper return saddr; 67173212b535SSteve Capper } 67183212b535SSteve Capper 671912710fd6SMike Kravetz static bool __vma_aligned_range_pmd_shareable(struct vm_area_struct *vma, 67208d9bfb26SMike Kravetz unsigned long start, unsigned long end, 67218d9bfb26SMike Kravetz bool check_vma_lock) 67223212b535SSteve Capper { 67238d9bfb26SMike Kravetz #ifdef CONFIG_USERFAULTFD 67248d9bfb26SMike Kravetz if (uffd_disable_huge_pmd_share(vma)) 67258d9bfb26SMike Kravetz return false; 67268d9bfb26SMike Kravetz #endif 67273212b535SSteve Capper /* 67283212b535SSteve Capper * check on proper vm_flags and page table alignment 67293212b535SSteve Capper */ 67308d9bfb26SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 673131aafb45SNicholas Krause return false; 67328d9bfb26SMike Kravetz if (check_vma_lock && !vma->vm_private_data) 67338d9bfb26SMike Kravetz return false; 67348d9bfb26SMike Kravetz if (!range_in_vma(vma, start, end)) 67358d9bfb26SMike Kravetz return false; 67368d9bfb26SMike Kravetz return true; 67378d9bfb26SMike Kravetz } 67388d9bfb26SMike Kravetz 67398d9bfb26SMike Kravetz static bool vma_pmd_shareable(struct vm_area_struct *vma) 67408d9bfb26SMike Kravetz { 67418d9bfb26SMike Kravetz unsigned long start = ALIGN(vma->vm_start, PUD_SIZE), 67428d9bfb26SMike Kravetz end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 67438d9bfb26SMike Kravetz 67448d9bfb26SMike Kravetz if (start >= end) 67458d9bfb26SMike Kravetz return false; 67468d9bfb26SMike Kravetz 67478d9bfb26SMike Kravetz return __vma_aligned_range_pmd_shareable(vma, start, end, false); 67483212b535SSteve Capper } 67493212b535SSteve Capper 675012710fd6SMike Kravetz static bool vma_addr_pmd_shareable(struct vm_area_struct *vma, 675112710fd6SMike Kravetz unsigned long addr) 675212710fd6SMike Kravetz { 675312710fd6SMike Kravetz unsigned long start = addr & PUD_MASK; 675412710fd6SMike Kravetz unsigned long end = start + PUD_SIZE; 675512710fd6SMike Kravetz 67568d9bfb26SMike Kravetz return __vma_aligned_range_pmd_shareable(vma, start, end, true); 675712710fd6SMike Kravetz } 675812710fd6SMike Kravetz 6759c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 6760c1991e07SPeter Xu { 676112710fd6SMike Kravetz return vma_addr_pmd_shareable(vma, addr); 6762c1991e07SPeter Xu } 6763c1991e07SPeter Xu 67643212b535SSteve Capper /* 6765017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 6766017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 6767017b1660SMike Kravetz * shared pmd mappings. 6768017b1660SMike Kravetz */ 6769017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 6770017b1660SMike Kravetz unsigned long *start, unsigned long *end) 6771017b1660SMike Kravetz { 6772a1ba9da8SLi Xinhai unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), 6773a1ba9da8SLi Xinhai v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 6774017b1660SMike Kravetz 6775a1ba9da8SLi Xinhai /* 6776f0953a1bSIngo Molnar * vma needs to span at least one aligned PUD size, and the range 6777f0953a1bSIngo Molnar * must be at least partially within in. 6778a1ba9da8SLi Xinhai */ 6779a1ba9da8SLi Xinhai if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || 6780a1ba9da8SLi Xinhai (*end <= v_start) || (*start >= v_end)) 6781017b1660SMike Kravetz return; 6782017b1660SMike Kravetz 678375802ca6SPeter Xu /* Extend the range to be PUD aligned for a worst case scenario */ 6784a1ba9da8SLi Xinhai if (*start > v_start) 6785a1ba9da8SLi Xinhai *start = ALIGN_DOWN(*start, PUD_SIZE); 6786017b1660SMike Kravetz 6787a1ba9da8SLi Xinhai if (*end < v_end) 6788a1ba9da8SLi Xinhai *end = ALIGN(*end, PUD_SIZE); 6789017b1660SMike Kravetz } 6790017b1660SMike Kravetz 67918d9bfb26SMike Kravetz static bool __vma_shareable_flags_pmd(struct vm_area_struct *vma) 67928d9bfb26SMike Kravetz { 67938d9bfb26SMike Kravetz return vma->vm_flags & (VM_MAYSHARE | VM_SHARED) && 67948d9bfb26SMike Kravetz vma->vm_private_data; 67958d9bfb26SMike Kravetz } 67968d9bfb26SMike Kravetz 67978d9bfb26SMike Kravetz void hugetlb_vma_lock_read(struct vm_area_struct *vma) 67988d9bfb26SMike Kravetz { 67998d9bfb26SMike Kravetz if (__vma_shareable_flags_pmd(vma)) { 68008d9bfb26SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 68018d9bfb26SMike Kravetz 68028d9bfb26SMike Kravetz down_read(&vma_lock->rw_sema); 68038d9bfb26SMike Kravetz } 68048d9bfb26SMike Kravetz } 68058d9bfb26SMike Kravetz 68068d9bfb26SMike Kravetz void hugetlb_vma_unlock_read(struct vm_area_struct *vma) 68078d9bfb26SMike Kravetz { 68088d9bfb26SMike Kravetz if (__vma_shareable_flags_pmd(vma)) { 68098d9bfb26SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 68108d9bfb26SMike Kravetz 68118d9bfb26SMike Kravetz up_read(&vma_lock->rw_sema); 68128d9bfb26SMike Kravetz } 68138d9bfb26SMike Kravetz } 68148d9bfb26SMike Kravetz 68158d9bfb26SMike Kravetz void hugetlb_vma_lock_write(struct vm_area_struct *vma) 68168d9bfb26SMike Kravetz { 68178d9bfb26SMike Kravetz if (__vma_shareable_flags_pmd(vma)) { 68188d9bfb26SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 68198d9bfb26SMike Kravetz 68208d9bfb26SMike Kravetz down_write(&vma_lock->rw_sema); 68218d9bfb26SMike Kravetz } 68228d9bfb26SMike Kravetz } 68238d9bfb26SMike Kravetz 68248d9bfb26SMike Kravetz void hugetlb_vma_unlock_write(struct vm_area_struct *vma) 68258d9bfb26SMike Kravetz { 68268d9bfb26SMike Kravetz if (__vma_shareable_flags_pmd(vma)) { 68278d9bfb26SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 68288d9bfb26SMike Kravetz 68298d9bfb26SMike Kravetz up_write(&vma_lock->rw_sema); 68308d9bfb26SMike Kravetz } 68318d9bfb26SMike Kravetz } 68328d9bfb26SMike Kravetz 68338d9bfb26SMike Kravetz int hugetlb_vma_trylock_write(struct vm_area_struct *vma) 68348d9bfb26SMike Kravetz { 68358d9bfb26SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 68368d9bfb26SMike Kravetz 68378d9bfb26SMike Kravetz if (!__vma_shareable_flags_pmd(vma)) 68388d9bfb26SMike Kravetz return 1; 68398d9bfb26SMike Kravetz 68408d9bfb26SMike Kravetz return down_write_trylock(&vma_lock->rw_sema); 68418d9bfb26SMike Kravetz } 68428d9bfb26SMike Kravetz 68438d9bfb26SMike Kravetz void hugetlb_vma_assert_locked(struct vm_area_struct *vma) 68448d9bfb26SMike Kravetz { 68458d9bfb26SMike Kravetz if (__vma_shareable_flags_pmd(vma)) { 68468d9bfb26SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 68478d9bfb26SMike Kravetz 68488d9bfb26SMike Kravetz lockdep_assert_held(&vma_lock->rw_sema); 68498d9bfb26SMike Kravetz } 68508d9bfb26SMike Kravetz } 68518d9bfb26SMike Kravetz 68528d9bfb26SMike Kravetz void hugetlb_vma_lock_release(struct kref *kref) 68538d9bfb26SMike Kravetz { 68548d9bfb26SMike Kravetz struct hugetlb_vma_lock *vma_lock = container_of(kref, 68558d9bfb26SMike Kravetz struct hugetlb_vma_lock, refs); 68568d9bfb26SMike Kravetz 68578d9bfb26SMike Kravetz kfree(vma_lock); 68588d9bfb26SMike Kravetz } 68598d9bfb26SMike Kravetz 68608d9bfb26SMike Kravetz static void hugetlb_vma_lock_free(struct vm_area_struct *vma) 68618d9bfb26SMike Kravetz { 68628d9bfb26SMike Kravetz /* 68638d9bfb26SMike Kravetz * Only present in sharable vmas. See comment in 68648d9bfb26SMike Kravetz * __unmap_hugepage_range_final about how VM_SHARED could 68658d9bfb26SMike Kravetz * be set without VM_MAYSHARE. As a result, we need to 68668d9bfb26SMike Kravetz * check if either is set in the free path. 68678d9bfb26SMike Kravetz */ 68688d9bfb26SMike Kravetz if (!vma || !(vma->vm_flags & (VM_MAYSHARE | VM_SHARED))) 68698d9bfb26SMike Kravetz return; 68708d9bfb26SMike Kravetz 68718d9bfb26SMike Kravetz if (vma->vm_private_data) { 68728d9bfb26SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 68738d9bfb26SMike Kravetz 68748d9bfb26SMike Kravetz /* 68758d9bfb26SMike Kravetz * vma_lock structure may or not be released, but it 68768d9bfb26SMike Kravetz * certainly will no longer be attached to vma so clear 68778d9bfb26SMike Kravetz * pointer. 68788d9bfb26SMike Kravetz */ 68798d9bfb26SMike Kravetz vma_lock->vma = NULL; 68808d9bfb26SMike Kravetz kref_put(&vma_lock->refs, hugetlb_vma_lock_release); 68818d9bfb26SMike Kravetz vma->vm_private_data = NULL; 68828d9bfb26SMike Kravetz } 68838d9bfb26SMike Kravetz } 68848d9bfb26SMike Kravetz 68858d9bfb26SMike Kravetz static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma) 68868d9bfb26SMike Kravetz { 68878d9bfb26SMike Kravetz struct hugetlb_vma_lock *vma_lock; 68888d9bfb26SMike Kravetz 68898d9bfb26SMike Kravetz /* Only establish in (flags) sharable vmas */ 68908d9bfb26SMike Kravetz if (!vma || !(vma->vm_flags & VM_MAYSHARE)) 68918d9bfb26SMike Kravetz return; 68928d9bfb26SMike Kravetz 68938d9bfb26SMike Kravetz /* Should never get here with non-NULL vm_private_data */ 68948d9bfb26SMike Kravetz if (vma->vm_private_data) 68958d9bfb26SMike Kravetz return; 68968d9bfb26SMike Kravetz 68978d9bfb26SMike Kravetz /* Check size/alignment for pmd sharing possible */ 68988d9bfb26SMike Kravetz if (!vma_pmd_shareable(vma)) 68998d9bfb26SMike Kravetz return; 69008d9bfb26SMike Kravetz 69018d9bfb26SMike Kravetz vma_lock = kmalloc(sizeof(*vma_lock), GFP_KERNEL); 69028d9bfb26SMike Kravetz if (!vma_lock) 69038d9bfb26SMike Kravetz /* 69048d9bfb26SMike Kravetz * If we can not allocate structure, then vma can not 69058d9bfb26SMike Kravetz * participate in pmd sharing. 69068d9bfb26SMike Kravetz */ 69078d9bfb26SMike Kravetz return; 69088d9bfb26SMike Kravetz 69098d9bfb26SMike Kravetz kref_init(&vma_lock->refs); 69108d9bfb26SMike Kravetz init_rwsem(&vma_lock->rw_sema); 69118d9bfb26SMike Kravetz vma_lock->vma = vma; 69128d9bfb26SMike Kravetz vma->vm_private_data = vma_lock; 69138d9bfb26SMike Kravetz } 69148d9bfb26SMike Kravetz 6915017b1660SMike Kravetz /* 69163212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 69173212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 69183212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 69193a47c54fSMike Kravetz * code much cleaner. pmd allocation is essential for the shared case because 69203a47c54fSMike Kravetz * pud has to be populated inside the same i_mmap_rwsem section - otherwise 69213a47c54fSMike Kravetz * racing tasks could either miss the sharing (see huge_pte_offset) or select a 69223a47c54fSMike Kravetz * bad pmd for sharing. 69233212b535SSteve Capper */ 6924aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 6925aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 69263212b535SSteve Capper { 69273212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 69283212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 69293212b535SSteve Capper vma->vm_pgoff; 69303212b535SSteve Capper struct vm_area_struct *svma; 69313212b535SSteve Capper unsigned long saddr; 69323212b535SSteve Capper pte_t *spte = NULL; 69333212b535SSteve Capper pte_t *pte; 6934cb900f41SKirill A. Shutemov spinlock_t *ptl; 69353212b535SSteve Capper 69363a47c54fSMike Kravetz i_mmap_lock_read(mapping); 69373212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 69383212b535SSteve Capper if (svma == vma) 69393212b535SSteve Capper continue; 69403212b535SSteve Capper 69413212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 69423212b535SSteve Capper if (saddr) { 69437868a208SPunit Agrawal spte = huge_pte_offset(svma->vm_mm, saddr, 69447868a208SPunit Agrawal vma_mmu_pagesize(svma)); 69453212b535SSteve Capper if (spte) { 69463212b535SSteve Capper get_page(virt_to_page(spte)); 69473212b535SSteve Capper break; 69483212b535SSteve Capper } 69493212b535SSteve Capper } 69503212b535SSteve Capper } 69513212b535SSteve Capper 69523212b535SSteve Capper if (!spte) 69533212b535SSteve Capper goto out; 69543212b535SSteve Capper 69558bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 6956dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 69573212b535SSteve Capper pud_populate(mm, pud, 69583212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 6959c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 6960dc6c9a35SKirill A. Shutemov } else { 69613212b535SSteve Capper put_page(virt_to_page(spte)); 6962dc6c9a35SKirill A. Shutemov } 6963cb900f41SKirill A. Shutemov spin_unlock(ptl); 69643212b535SSteve Capper out: 69653212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 69663a47c54fSMike Kravetz i_mmap_unlock_read(mapping); 69673212b535SSteve Capper return pte; 69683212b535SSteve Capper } 69693212b535SSteve Capper 69703212b535SSteve Capper /* 69713212b535SSteve Capper * unmap huge page backed by shared pte. 69723212b535SSteve Capper * 69733212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 69743212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 69753212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 69763212b535SSteve Capper * 69773a47c54fSMike Kravetz * Called with page table lock held. 69783212b535SSteve Capper * 69793212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 69803212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 69813212b535SSteve Capper */ 698234ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 69834ddb4d91SMike Kravetz unsigned long addr, pte_t *ptep) 69843212b535SSteve Capper { 69854ddb4d91SMike Kravetz pgd_t *pgd = pgd_offset(mm, addr); 69864ddb4d91SMike Kravetz p4d_t *p4d = p4d_offset(pgd, addr); 69874ddb4d91SMike Kravetz pud_t *pud = pud_offset(p4d, addr); 69883212b535SSteve Capper 698934ae204fSMike Kravetz i_mmap_assert_write_locked(vma->vm_file->f_mapping); 6990*40549ba8SMike Kravetz hugetlb_vma_assert_locked(vma); 69913212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 69923212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 69933212b535SSteve Capper return 0; 69943212b535SSteve Capper 69953212b535SSteve Capper pud_clear(pud); 69963212b535SSteve Capper put_page(virt_to_page(ptep)); 6997dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 69983212b535SSteve Capper return 1; 69993212b535SSteve Capper } 7000c1991e07SPeter Xu 70019e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 70028d9bfb26SMike Kravetz 7003*40549ba8SMike Kravetz void hugetlb_vma_lock_read(struct vm_area_struct *vma) 7004*40549ba8SMike Kravetz { 7005*40549ba8SMike Kravetz } 7006*40549ba8SMike Kravetz 7007*40549ba8SMike Kravetz void hugetlb_vma_unlock_read(struct vm_area_struct *vma) 7008*40549ba8SMike Kravetz { 7009*40549ba8SMike Kravetz } 7010*40549ba8SMike Kravetz 7011*40549ba8SMike Kravetz void hugetlb_vma_lock_write(struct vm_area_struct *vma) 7012*40549ba8SMike Kravetz { 7013*40549ba8SMike Kravetz } 7014*40549ba8SMike Kravetz 7015*40549ba8SMike Kravetz void hugetlb_vma_unlock_write(struct vm_area_struct *vma) 7016*40549ba8SMike Kravetz { 7017*40549ba8SMike Kravetz } 7018*40549ba8SMike Kravetz 7019*40549ba8SMike Kravetz int hugetlb_vma_trylock_write(struct vm_area_struct *vma) 7020*40549ba8SMike Kravetz { 7021*40549ba8SMike Kravetz return 1; 7022*40549ba8SMike Kravetz } 7023*40549ba8SMike Kravetz 7024*40549ba8SMike Kravetz void hugetlb_vma_assert_locked(struct vm_area_struct *vma) 7025*40549ba8SMike Kravetz { 7026*40549ba8SMike Kravetz } 7027*40549ba8SMike Kravetz 70288d9bfb26SMike Kravetz void hugetlb_vma_lock_release(struct kref *kref) 70298d9bfb26SMike Kravetz { 70308d9bfb26SMike Kravetz } 70318d9bfb26SMike Kravetz 70328d9bfb26SMike Kravetz static void hugetlb_vma_lock_free(struct vm_area_struct *vma) 70338d9bfb26SMike Kravetz { 70348d9bfb26SMike Kravetz } 70358d9bfb26SMike Kravetz 70368d9bfb26SMike Kravetz static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma) 70378d9bfb26SMike Kravetz { 70388d9bfb26SMike Kravetz } 70398d9bfb26SMike Kravetz 7040aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 7041aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 70429e5fc74cSSteve Capper { 70439e5fc74cSSteve Capper return NULL; 70449e5fc74cSSteve Capper } 7045e81f2d22SZhang Zhen 704634ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 70474ddb4d91SMike Kravetz unsigned long addr, pte_t *ptep) 7048e81f2d22SZhang Zhen { 7049e81f2d22SZhang Zhen return 0; 7050e81f2d22SZhang Zhen } 7051017b1660SMike Kravetz 7052017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 7053017b1660SMike Kravetz unsigned long *start, unsigned long *end) 7054017b1660SMike Kravetz { 7055017b1660SMike Kravetz } 7056c1991e07SPeter Xu 7057c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 7058c1991e07SPeter Xu { 7059c1991e07SPeter Xu return false; 7060c1991e07SPeter Xu } 70613212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 70623212b535SSteve Capper 70639e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 7064aec44e0fSPeter Xu pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, 70659e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 70669e5fc74cSSteve Capper { 70679e5fc74cSSteve Capper pgd_t *pgd; 7068c2febafcSKirill A. Shutemov p4d_t *p4d; 70699e5fc74cSSteve Capper pud_t *pud; 70709e5fc74cSSteve Capper pte_t *pte = NULL; 70719e5fc74cSSteve Capper 70729e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 7073f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 7074f4f0a3d8SKirill A. Shutemov if (!p4d) 7075f4f0a3d8SKirill A. Shutemov return NULL; 7076c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 70779e5fc74cSSteve Capper if (pud) { 70789e5fc74cSSteve Capper if (sz == PUD_SIZE) { 70799e5fc74cSSteve Capper pte = (pte_t *)pud; 70809e5fc74cSSteve Capper } else { 70819e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 7082c1991e07SPeter Xu if (want_pmd_share(vma, addr) && pud_none(*pud)) 7083aec44e0fSPeter Xu pte = huge_pmd_share(mm, vma, addr, pud); 70849e5fc74cSSteve Capper else 70859e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 70869e5fc74cSSteve Capper } 70879e5fc74cSSteve Capper } 70884e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 70899e5fc74cSSteve Capper 70909e5fc74cSSteve Capper return pte; 70919e5fc74cSSteve Capper } 70929e5fc74cSSteve Capper 70939b19df29SPunit Agrawal /* 70949b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 70959b19df29SPunit Agrawal * entry at address @addr 70969b19df29SPunit Agrawal * 70978ac0b81aSLi Xinhai * Return: Pointer to page table entry (PUD or PMD) for 70988ac0b81aSLi Xinhai * address @addr, or NULL if a !p*d_present() entry is encountered and the 70999b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 71009b19df29SPunit Agrawal * table. 71019b19df29SPunit Agrawal */ 71027868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 71037868a208SPunit Agrawal unsigned long addr, unsigned long sz) 71049e5fc74cSSteve Capper { 71059e5fc74cSSteve Capper pgd_t *pgd; 7106c2febafcSKirill A. Shutemov p4d_t *p4d; 71078ac0b81aSLi Xinhai pud_t *pud; 71088ac0b81aSLi Xinhai pmd_t *pmd; 71099e5fc74cSSteve Capper 71109e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 7111c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 7112c2febafcSKirill A. Shutemov return NULL; 7113c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 7114c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 7115c2febafcSKirill A. Shutemov return NULL; 71169b19df29SPunit Agrawal 7117c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 71188ac0b81aSLi Xinhai if (sz == PUD_SIZE) 71198ac0b81aSLi Xinhai /* must be pud huge, non-present or none */ 71209e5fc74cSSteve Capper return (pte_t *)pud; 71218ac0b81aSLi Xinhai if (!pud_present(*pud)) 71228ac0b81aSLi Xinhai return NULL; 71238ac0b81aSLi Xinhai /* must have a valid entry and size to go further */ 71249b19df29SPunit Agrawal 71259e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 71268ac0b81aSLi Xinhai /* must be pmd huge, non-present or none */ 71279e5fc74cSSteve Capper return (pte_t *)pmd; 71289e5fc74cSSteve Capper } 71299e5fc74cSSteve Capper 7130e95a9851SMike Kravetz /* 7131e95a9851SMike Kravetz * Return a mask that can be used to update an address to the last huge 7132e95a9851SMike Kravetz * page in a page table page mapping size. Used to skip non-present 7133e95a9851SMike Kravetz * page table entries when linearly scanning address ranges. Architectures 7134e95a9851SMike Kravetz * with unique huge page to page table relationships can define their own 7135e95a9851SMike Kravetz * version of this routine. 7136e95a9851SMike Kravetz */ 7137e95a9851SMike Kravetz unsigned long hugetlb_mask_last_page(struct hstate *h) 7138e95a9851SMike Kravetz { 7139e95a9851SMike Kravetz unsigned long hp_size = huge_page_size(h); 7140e95a9851SMike Kravetz 7141e95a9851SMike Kravetz if (hp_size == PUD_SIZE) 7142e95a9851SMike Kravetz return P4D_SIZE - PUD_SIZE; 7143e95a9851SMike Kravetz else if (hp_size == PMD_SIZE) 7144e95a9851SMike Kravetz return PUD_SIZE - PMD_SIZE; 7145e95a9851SMike Kravetz else 7146e95a9851SMike Kravetz return 0UL; 7147e95a9851SMike Kravetz } 7148e95a9851SMike Kravetz 7149e95a9851SMike Kravetz #else 7150e95a9851SMike Kravetz 7151e95a9851SMike Kravetz /* See description above. Architectures can provide their own version. */ 7152e95a9851SMike Kravetz __weak unsigned long hugetlb_mask_last_page(struct hstate *h) 7153e95a9851SMike Kravetz { 71544ddb4d91SMike Kravetz #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 71554ddb4d91SMike Kravetz if (huge_page_size(h) == PMD_SIZE) 71564ddb4d91SMike Kravetz return PUD_SIZE - PMD_SIZE; 71574ddb4d91SMike Kravetz #endif 7158e95a9851SMike Kravetz return 0UL; 7159e95a9851SMike Kravetz } 7160e95a9851SMike Kravetz 716161f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 716261f77edaSNaoya Horiguchi 716361f77edaSNaoya Horiguchi /* 716461f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 716561f77edaSNaoya Horiguchi * behavior. 716661f77edaSNaoya Horiguchi */ 716761f77edaSNaoya Horiguchi struct page * __weak 716861f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 716961f77edaSNaoya Horiguchi int write) 717061f77edaSNaoya Horiguchi { 717161f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 717261f77edaSNaoya Horiguchi } 717361f77edaSNaoya Horiguchi 717461f77edaSNaoya Horiguchi struct page * __weak 71754dc71451SAneesh Kumar K.V follow_huge_pd(struct vm_area_struct *vma, 71764dc71451SAneesh Kumar K.V unsigned long address, hugepd_t hpd, int flags, int pdshift) 71774dc71451SAneesh Kumar K.V { 71784dc71451SAneesh Kumar K.V WARN(1, "hugepd follow called with no support for hugepage directory format\n"); 71794dc71451SAneesh Kumar K.V return NULL; 71804dc71451SAneesh Kumar K.V } 71814dc71451SAneesh Kumar K.V 71824dc71451SAneesh Kumar K.V struct page * __weak 71839e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 7184e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 71859e5fc74cSSteve Capper { 7186e66f17ffSNaoya Horiguchi struct page *page = NULL; 7187e66f17ffSNaoya Horiguchi spinlock_t *ptl; 7188c9d398faSNaoya Horiguchi pte_t pte; 71893faa52c0SJohn Hubbard 71908909691bSDavid Hildenbrand /* 71918909691bSDavid Hildenbrand * FOLL_PIN is not supported for follow_page(). Ordinary GUP goes via 71928909691bSDavid Hildenbrand * follow_hugetlb_page(). 71938909691bSDavid Hildenbrand */ 71948909691bSDavid Hildenbrand if (WARN_ON_ONCE(flags & FOLL_PIN)) 71953faa52c0SJohn Hubbard return NULL; 71963faa52c0SJohn Hubbard 7197e66f17ffSNaoya Horiguchi retry: 7198e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 7199e66f17ffSNaoya Horiguchi spin_lock(ptl); 7200e66f17ffSNaoya Horiguchi /* 7201e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 7202e66f17ffSNaoya Horiguchi * unmapped from other threads. 7203e66f17ffSNaoya Horiguchi */ 7204e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 7205e66f17ffSNaoya Horiguchi goto out; 7206c9d398faSNaoya Horiguchi pte = huge_ptep_get((pte_t *)pmd); 7207c9d398faSNaoya Horiguchi if (pte_present(pte)) { 720897534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 72093faa52c0SJohn Hubbard /* 72103faa52c0SJohn Hubbard * try_grab_page() should always succeed here, because: a) we 72113faa52c0SJohn Hubbard * hold the pmd (ptl) lock, and b) we've just checked that the 72123faa52c0SJohn Hubbard * huge pmd (head) page is present in the page tables. The ptl 72133faa52c0SJohn Hubbard * prevents the head page and tail pages from being rearranged 72143faa52c0SJohn Hubbard * in any way. So this page must be available at this point, 72153faa52c0SJohn Hubbard * unless the page refcount overflowed: 72163faa52c0SJohn Hubbard */ 72173faa52c0SJohn Hubbard if (WARN_ON_ONCE(!try_grab_page(page, flags))) { 72183faa52c0SJohn Hubbard page = NULL; 72193faa52c0SJohn Hubbard goto out; 72203faa52c0SJohn Hubbard } 7221e66f17ffSNaoya Horiguchi } else { 7222c9d398faSNaoya Horiguchi if (is_hugetlb_entry_migration(pte)) { 7223e66f17ffSNaoya Horiguchi spin_unlock(ptl); 7224ad1ac596SMiaohe Lin __migration_entry_wait_huge((pte_t *)pmd, ptl); 7225e66f17ffSNaoya Horiguchi goto retry; 7226e66f17ffSNaoya Horiguchi } 7227e66f17ffSNaoya Horiguchi /* 7228e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 7229e66f17ffSNaoya Horiguchi * follow_page_mask(). 7230e66f17ffSNaoya Horiguchi */ 7231e66f17ffSNaoya Horiguchi } 7232e66f17ffSNaoya Horiguchi out: 7233e66f17ffSNaoya Horiguchi spin_unlock(ptl); 72349e5fc74cSSteve Capper return page; 72359e5fc74cSSteve Capper } 72369e5fc74cSSteve Capper 723761f77edaSNaoya Horiguchi struct page * __weak 72389e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 7239e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 72409e5fc74cSSteve Capper { 72413a194f3fSNaoya Horiguchi struct page *page = NULL; 72423a194f3fSNaoya Horiguchi spinlock_t *ptl; 72433a194f3fSNaoya Horiguchi pte_t pte; 72443a194f3fSNaoya Horiguchi 72453a194f3fSNaoya Horiguchi if (WARN_ON_ONCE(flags & FOLL_PIN)) 7246e66f17ffSNaoya Horiguchi return NULL; 72479e5fc74cSSteve Capper 72483a194f3fSNaoya Horiguchi retry: 72493a194f3fSNaoya Horiguchi ptl = huge_pte_lock(hstate_sizelog(PUD_SHIFT), mm, (pte_t *)pud); 72503a194f3fSNaoya Horiguchi if (!pud_huge(*pud)) 72513a194f3fSNaoya Horiguchi goto out; 72523a194f3fSNaoya Horiguchi pte = huge_ptep_get((pte_t *)pud); 72533a194f3fSNaoya Horiguchi if (pte_present(pte)) { 72543a194f3fSNaoya Horiguchi page = pud_page(*pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 72553a194f3fSNaoya Horiguchi if (WARN_ON_ONCE(!try_grab_page(page, flags))) { 72563a194f3fSNaoya Horiguchi page = NULL; 72573a194f3fSNaoya Horiguchi goto out; 72583a194f3fSNaoya Horiguchi } 72593a194f3fSNaoya Horiguchi } else { 72603a194f3fSNaoya Horiguchi if (is_hugetlb_entry_migration(pte)) { 72613a194f3fSNaoya Horiguchi spin_unlock(ptl); 72623a194f3fSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pud, ptl); 72633a194f3fSNaoya Horiguchi goto retry; 72643a194f3fSNaoya Horiguchi } 72653a194f3fSNaoya Horiguchi /* 72663a194f3fSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 72673a194f3fSNaoya Horiguchi * follow_page_mask(). 72683a194f3fSNaoya Horiguchi */ 72693a194f3fSNaoya Horiguchi } 72703a194f3fSNaoya Horiguchi out: 72713a194f3fSNaoya Horiguchi spin_unlock(ptl); 72723a194f3fSNaoya Horiguchi return page; 72739e5fc74cSSteve Capper } 72749e5fc74cSSteve Capper 7275faaa5b62SAnshuman Khandual struct page * __weak 7276faaa5b62SAnshuman Khandual follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) 7277faaa5b62SAnshuman Khandual { 72783faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 7279faaa5b62SAnshuman Khandual return NULL; 7280faaa5b62SAnshuman Khandual 7281faaa5b62SAnshuman Khandual return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); 7282faaa5b62SAnshuman Khandual } 7283faaa5b62SAnshuman Khandual 72847ce82f4cSMiaohe Lin int isolate_hugetlb(struct page *page, struct list_head *list) 728531caf665SNaoya Horiguchi { 72867ce82f4cSMiaohe Lin int ret = 0; 7287bcc54222SNaoya Horiguchi 7288db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 72898f251a3dSMike Kravetz if (!PageHeadHuge(page) || 72908f251a3dSMike Kravetz !HPageMigratable(page) || 72910eb2df2bSMuchun Song !get_page_unless_zero(page)) { 72927ce82f4cSMiaohe Lin ret = -EBUSY; 7293bcc54222SNaoya Horiguchi goto unlock; 7294bcc54222SNaoya Horiguchi } 72958f251a3dSMike Kravetz ClearHPageMigratable(page); 729631caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 7297bcc54222SNaoya Horiguchi unlock: 7298db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7299bcc54222SNaoya Horiguchi return ret; 730031caf665SNaoya Horiguchi } 730131caf665SNaoya Horiguchi 730225182f05SNaoya Horiguchi int get_hwpoison_huge_page(struct page *page, bool *hugetlb) 730325182f05SNaoya Horiguchi { 730425182f05SNaoya Horiguchi int ret = 0; 730525182f05SNaoya Horiguchi 730625182f05SNaoya Horiguchi *hugetlb = false; 730725182f05SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 730825182f05SNaoya Horiguchi if (PageHeadHuge(page)) { 730925182f05SNaoya Horiguchi *hugetlb = true; 7310b283d983SNaoya Horiguchi if (HPageFreed(page)) 7311b283d983SNaoya Horiguchi ret = 0; 7312b283d983SNaoya Horiguchi else if (HPageMigratable(page)) 731325182f05SNaoya Horiguchi ret = get_page_unless_zero(page); 73140ed950d1SNaoya Horiguchi else 73150ed950d1SNaoya Horiguchi ret = -EBUSY; 731625182f05SNaoya Horiguchi } 731725182f05SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 731825182f05SNaoya Horiguchi return ret; 731925182f05SNaoya Horiguchi } 732025182f05SNaoya Horiguchi 7321405ce051SNaoya Horiguchi int get_huge_page_for_hwpoison(unsigned long pfn, int flags) 7322405ce051SNaoya Horiguchi { 7323405ce051SNaoya Horiguchi int ret; 7324405ce051SNaoya Horiguchi 7325405ce051SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 7326405ce051SNaoya Horiguchi ret = __get_huge_page_for_hwpoison(pfn, flags); 7327405ce051SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 7328405ce051SNaoya Horiguchi return ret; 7329405ce051SNaoya Horiguchi } 7330405ce051SNaoya Horiguchi 733131caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 733231caf665SNaoya Horiguchi { 7333db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 73348f251a3dSMike Kravetz SetHPageMigratable(page); 733531caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 7336db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 733731caf665SNaoya Horiguchi put_page(page); 733831caf665SNaoya Horiguchi } 7339ab5ac90aSMichal Hocko 7340ab5ac90aSMichal Hocko void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) 7341ab5ac90aSMichal Hocko { 7342ab5ac90aSMichal Hocko struct hstate *h = page_hstate(oldpage); 7343ab5ac90aSMichal Hocko 7344ab5ac90aSMichal Hocko hugetlb_cgroup_migrate(oldpage, newpage); 7345ab5ac90aSMichal Hocko set_page_owner_migrate_reason(newpage, reason); 7346ab5ac90aSMichal Hocko 7347ab5ac90aSMichal Hocko /* 7348ab5ac90aSMichal Hocko * transfer temporary state of the new huge page. This is 7349ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 7350ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 7351ab5ac90aSMichal Hocko * the temporary status. 7352ab5ac90aSMichal Hocko * 7353ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 7354ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 7355ab5ac90aSMichal Hocko * the per-node's. 7356ab5ac90aSMichal Hocko */ 73579157c311SMike Kravetz if (HPageTemporary(newpage)) { 7358ab5ac90aSMichal Hocko int old_nid = page_to_nid(oldpage); 7359ab5ac90aSMichal Hocko int new_nid = page_to_nid(newpage); 7360ab5ac90aSMichal Hocko 73619157c311SMike Kravetz SetHPageTemporary(oldpage); 73629157c311SMike Kravetz ClearHPageTemporary(newpage); 7363ab5ac90aSMichal Hocko 73645af1ab1dSMiaohe Lin /* 73655af1ab1dSMiaohe Lin * There is no need to transfer the per-node surplus state 73665af1ab1dSMiaohe Lin * when we do not cross the node. 73675af1ab1dSMiaohe Lin */ 73685af1ab1dSMiaohe Lin if (new_nid == old_nid) 73695af1ab1dSMiaohe Lin return; 7370db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 7371ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 7372ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 7373ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 7374ab5ac90aSMichal Hocko } 7375db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7376ab5ac90aSMichal Hocko } 7377ab5ac90aSMichal Hocko } 7378cf11e85fSRoman Gushchin 73796dfeaff9SPeter Xu /* 73806dfeaff9SPeter Xu * This function will unconditionally remove all the shared pmd pgtable entries 73816dfeaff9SPeter Xu * within the specific vma for a hugetlbfs memory range. 73826dfeaff9SPeter Xu */ 73836dfeaff9SPeter Xu void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) 73846dfeaff9SPeter Xu { 73856dfeaff9SPeter Xu struct hstate *h = hstate_vma(vma); 73866dfeaff9SPeter Xu unsigned long sz = huge_page_size(h); 73876dfeaff9SPeter Xu struct mm_struct *mm = vma->vm_mm; 73886dfeaff9SPeter Xu struct mmu_notifier_range range; 73896dfeaff9SPeter Xu unsigned long address, start, end; 73906dfeaff9SPeter Xu spinlock_t *ptl; 73916dfeaff9SPeter Xu pte_t *ptep; 73926dfeaff9SPeter Xu 73936dfeaff9SPeter Xu if (!(vma->vm_flags & VM_MAYSHARE)) 73946dfeaff9SPeter Xu return; 73956dfeaff9SPeter Xu 73966dfeaff9SPeter Xu start = ALIGN(vma->vm_start, PUD_SIZE); 73976dfeaff9SPeter Xu end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 73986dfeaff9SPeter Xu 73996dfeaff9SPeter Xu if (start >= end) 74006dfeaff9SPeter Xu return; 74016dfeaff9SPeter Xu 74029c8bbfacSBaolin Wang flush_cache_range(vma, start, end); 74036dfeaff9SPeter Xu /* 74046dfeaff9SPeter Xu * No need to call adjust_range_if_pmd_sharing_possible(), because 74056dfeaff9SPeter Xu * we have already done the PUD_SIZE alignment. 74066dfeaff9SPeter Xu */ 74076dfeaff9SPeter Xu mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, 74086dfeaff9SPeter Xu start, end); 74096dfeaff9SPeter Xu mmu_notifier_invalidate_range_start(&range); 7410*40549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 74116dfeaff9SPeter Xu i_mmap_lock_write(vma->vm_file->f_mapping); 74126dfeaff9SPeter Xu for (address = start; address < end; address += PUD_SIZE) { 74136dfeaff9SPeter Xu ptep = huge_pte_offset(mm, address, sz); 74146dfeaff9SPeter Xu if (!ptep) 74156dfeaff9SPeter Xu continue; 74166dfeaff9SPeter Xu ptl = huge_pte_lock(h, mm, ptep); 74174ddb4d91SMike Kravetz huge_pmd_unshare(mm, vma, address, ptep); 74186dfeaff9SPeter Xu spin_unlock(ptl); 74196dfeaff9SPeter Xu } 74206dfeaff9SPeter Xu flush_hugetlb_tlb_range(vma, start, end); 74216dfeaff9SPeter Xu i_mmap_unlock_write(vma->vm_file->f_mapping); 7422*40549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 74236dfeaff9SPeter Xu /* 74246dfeaff9SPeter Xu * No need to call mmu_notifier_invalidate_range(), see 7425ee65728eSMike Rapoport * Documentation/mm/mmu_notifier.rst. 74266dfeaff9SPeter Xu */ 74276dfeaff9SPeter Xu mmu_notifier_invalidate_range_end(&range); 74286dfeaff9SPeter Xu } 74296dfeaff9SPeter Xu 7430cf11e85fSRoman Gushchin #ifdef CONFIG_CMA 7431cf11e85fSRoman Gushchin static bool cma_reserve_called __initdata; 7432cf11e85fSRoman Gushchin 7433cf11e85fSRoman Gushchin static int __init cmdline_parse_hugetlb_cma(char *p) 7434cf11e85fSRoman Gushchin { 743538e719abSBaolin Wang int nid, count = 0; 743638e719abSBaolin Wang unsigned long tmp; 743738e719abSBaolin Wang char *s = p; 743838e719abSBaolin Wang 743938e719abSBaolin Wang while (*s) { 744038e719abSBaolin Wang if (sscanf(s, "%lu%n", &tmp, &count) != 1) 744138e719abSBaolin Wang break; 744238e719abSBaolin Wang 744338e719abSBaolin Wang if (s[count] == ':') { 7444f9317f77SMike Kravetz if (tmp >= MAX_NUMNODES) 744538e719abSBaolin Wang break; 7446f9317f77SMike Kravetz nid = array_index_nospec(tmp, MAX_NUMNODES); 744738e719abSBaolin Wang 744838e719abSBaolin Wang s += count + 1; 744938e719abSBaolin Wang tmp = memparse(s, &s); 745038e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = tmp; 745138e719abSBaolin Wang hugetlb_cma_size += tmp; 745238e719abSBaolin Wang 745338e719abSBaolin Wang /* 745438e719abSBaolin Wang * Skip the separator if have one, otherwise 745538e719abSBaolin Wang * break the parsing. 745638e719abSBaolin Wang */ 745738e719abSBaolin Wang if (*s == ',') 745838e719abSBaolin Wang s++; 745938e719abSBaolin Wang else 746038e719abSBaolin Wang break; 746138e719abSBaolin Wang } else { 7462cf11e85fSRoman Gushchin hugetlb_cma_size = memparse(p, &p); 746338e719abSBaolin Wang break; 746438e719abSBaolin Wang } 746538e719abSBaolin Wang } 746638e719abSBaolin Wang 7467cf11e85fSRoman Gushchin return 0; 7468cf11e85fSRoman Gushchin } 7469cf11e85fSRoman Gushchin 7470cf11e85fSRoman Gushchin early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); 7471cf11e85fSRoman Gushchin 7472cf11e85fSRoman Gushchin void __init hugetlb_cma_reserve(int order) 7473cf11e85fSRoman Gushchin { 7474cf11e85fSRoman Gushchin unsigned long size, reserved, per_node; 747538e719abSBaolin Wang bool node_specific_cma_alloc = false; 7476cf11e85fSRoman Gushchin int nid; 7477cf11e85fSRoman Gushchin 7478cf11e85fSRoman Gushchin cma_reserve_called = true; 7479cf11e85fSRoman Gushchin 7480cf11e85fSRoman Gushchin if (!hugetlb_cma_size) 7481cf11e85fSRoman Gushchin return; 7482cf11e85fSRoman Gushchin 748338e719abSBaolin Wang for (nid = 0; nid < MAX_NUMNODES; nid++) { 748438e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 748538e719abSBaolin Wang continue; 748638e719abSBaolin Wang 748730a51400SPeng Liu if (!node_online(nid)) { 748838e719abSBaolin Wang pr_warn("hugetlb_cma: invalid node %d specified\n", nid); 748938e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 749038e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 749138e719abSBaolin Wang continue; 749238e719abSBaolin Wang } 749338e719abSBaolin Wang 749438e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) { 749538e719abSBaolin Wang pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n", 749638e719abSBaolin Wang nid, (PAGE_SIZE << order) / SZ_1M); 749738e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 749838e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 749938e719abSBaolin Wang } else { 750038e719abSBaolin Wang node_specific_cma_alloc = true; 750138e719abSBaolin Wang } 750238e719abSBaolin Wang } 750338e719abSBaolin Wang 750438e719abSBaolin Wang /* Validate the CMA size again in case some invalid nodes specified. */ 750538e719abSBaolin Wang if (!hugetlb_cma_size) 750638e719abSBaolin Wang return; 750738e719abSBaolin Wang 7508cf11e85fSRoman Gushchin if (hugetlb_cma_size < (PAGE_SIZE << order)) { 7509cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", 7510cf11e85fSRoman Gushchin (PAGE_SIZE << order) / SZ_1M); 7511a01f4390SMike Kravetz hugetlb_cma_size = 0; 7512cf11e85fSRoman Gushchin return; 7513cf11e85fSRoman Gushchin } 7514cf11e85fSRoman Gushchin 751538e719abSBaolin Wang if (!node_specific_cma_alloc) { 7516cf11e85fSRoman Gushchin /* 7517cf11e85fSRoman Gushchin * If 3 GB area is requested on a machine with 4 numa nodes, 7518cf11e85fSRoman Gushchin * let's allocate 1 GB on first three nodes and ignore the last one. 7519cf11e85fSRoman Gushchin */ 7520cf11e85fSRoman Gushchin per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); 7521cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", 7522cf11e85fSRoman Gushchin hugetlb_cma_size / SZ_1M, per_node / SZ_1M); 752338e719abSBaolin Wang } 7524cf11e85fSRoman Gushchin 7525cf11e85fSRoman Gushchin reserved = 0; 752630a51400SPeng Liu for_each_online_node(nid) { 7527cf11e85fSRoman Gushchin int res; 75282281f797SBarry Song char name[CMA_MAX_NAME]; 7529cf11e85fSRoman Gushchin 753038e719abSBaolin Wang if (node_specific_cma_alloc) { 753138e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 753238e719abSBaolin Wang continue; 753338e719abSBaolin Wang 753438e719abSBaolin Wang size = hugetlb_cma_size_in_node[nid]; 753538e719abSBaolin Wang } else { 7536cf11e85fSRoman Gushchin size = min(per_node, hugetlb_cma_size - reserved); 753738e719abSBaolin Wang } 753838e719abSBaolin Wang 7539cf11e85fSRoman Gushchin size = round_up(size, PAGE_SIZE << order); 7540cf11e85fSRoman Gushchin 75412281f797SBarry Song snprintf(name, sizeof(name), "hugetlb%d", nid); 7542a01f4390SMike Kravetz /* 7543a01f4390SMike Kravetz * Note that 'order per bit' is based on smallest size that 7544a01f4390SMike Kravetz * may be returned to CMA allocator in the case of 7545a01f4390SMike Kravetz * huge page demotion. 7546a01f4390SMike Kravetz */ 7547a01f4390SMike Kravetz res = cma_declare_contiguous_nid(0, size, 0, 7548a01f4390SMike Kravetz PAGE_SIZE << HUGETLB_PAGE_ORDER, 754929d0f41dSBarry Song 0, false, name, 7550cf11e85fSRoman Gushchin &hugetlb_cma[nid], nid); 7551cf11e85fSRoman Gushchin if (res) { 7552cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: reservation failed: err %d, node %d", 7553cf11e85fSRoman Gushchin res, nid); 7554cf11e85fSRoman Gushchin continue; 7555cf11e85fSRoman Gushchin } 7556cf11e85fSRoman Gushchin 7557cf11e85fSRoman Gushchin reserved += size; 7558cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", 7559cf11e85fSRoman Gushchin size / SZ_1M, nid); 7560cf11e85fSRoman Gushchin 7561cf11e85fSRoman Gushchin if (reserved >= hugetlb_cma_size) 7562cf11e85fSRoman Gushchin break; 7563cf11e85fSRoman Gushchin } 7564a01f4390SMike Kravetz 7565a01f4390SMike Kravetz if (!reserved) 7566a01f4390SMike Kravetz /* 7567a01f4390SMike Kravetz * hugetlb_cma_size is used to determine if allocations from 7568a01f4390SMike Kravetz * cma are possible. Set to zero if no cma regions are set up. 7569a01f4390SMike Kravetz */ 7570a01f4390SMike Kravetz hugetlb_cma_size = 0; 7571cf11e85fSRoman Gushchin } 7572cf11e85fSRoman Gushchin 7573263b8998SMiaohe Lin static void __init hugetlb_cma_check(void) 7574cf11e85fSRoman Gushchin { 7575cf11e85fSRoman Gushchin if (!hugetlb_cma_size || cma_reserve_called) 7576cf11e85fSRoman Gushchin return; 7577cf11e85fSRoman Gushchin 7578cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); 7579cf11e85fSRoman Gushchin } 7580cf11e85fSRoman Gushchin 7581cf11e85fSRoman Gushchin #endif /* CONFIG_CMA */ 7582