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); 96ecfbd733SMike Kravetz static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma); 977ca02d0aSMike Kravetz 981d88433bSMiaohe Lin static inline bool subpool_is_free(struct hugepage_subpool *spool) 991d88433bSMiaohe Lin { 1001d88433bSMiaohe Lin if (spool->count) 1011d88433bSMiaohe Lin return false; 1021d88433bSMiaohe Lin if (spool->max_hpages != -1) 1031d88433bSMiaohe Lin return spool->used_hpages == 0; 1041d88433bSMiaohe Lin if (spool->min_hpages != -1) 1051d88433bSMiaohe Lin return spool->rsv_hpages == spool->min_hpages; 1061d88433bSMiaohe Lin 1071d88433bSMiaohe Lin return true; 1081d88433bSMiaohe Lin } 1091d88433bSMiaohe Lin 110db71ef79SMike Kravetz static inline void unlock_or_release_subpool(struct hugepage_subpool *spool, 111db71ef79SMike Kravetz unsigned long irq_flags) 11290481622SDavid Gibson { 113db71ef79SMike Kravetz spin_unlock_irqrestore(&spool->lock, irq_flags); 11490481622SDavid Gibson 11590481622SDavid Gibson /* If no pages are used, and no other handles to the subpool 1167c8de358SEthon Paul * remain, give up any reservations based on minimum size and 1177ca02d0aSMike Kravetz * free the subpool */ 1181d88433bSMiaohe Lin if (subpool_is_free(spool)) { 1197ca02d0aSMike Kravetz if (spool->min_hpages != -1) 1207ca02d0aSMike Kravetz hugetlb_acct_memory(spool->hstate, 1217ca02d0aSMike Kravetz -spool->min_hpages); 12290481622SDavid Gibson kfree(spool); 12390481622SDavid Gibson } 1247ca02d0aSMike Kravetz } 12590481622SDavid Gibson 1267ca02d0aSMike Kravetz struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, 1277ca02d0aSMike Kravetz long min_hpages) 12890481622SDavid Gibson { 12990481622SDavid Gibson struct hugepage_subpool *spool; 13090481622SDavid Gibson 131c6a91820SMike Kravetz spool = kzalloc(sizeof(*spool), GFP_KERNEL); 13290481622SDavid Gibson if (!spool) 13390481622SDavid Gibson return NULL; 13490481622SDavid Gibson 13590481622SDavid Gibson spin_lock_init(&spool->lock); 13690481622SDavid Gibson spool->count = 1; 1377ca02d0aSMike Kravetz spool->max_hpages = max_hpages; 1387ca02d0aSMike Kravetz spool->hstate = h; 1397ca02d0aSMike Kravetz spool->min_hpages = min_hpages; 1407ca02d0aSMike Kravetz 1417ca02d0aSMike Kravetz if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { 1427ca02d0aSMike Kravetz kfree(spool); 1437ca02d0aSMike Kravetz return NULL; 1447ca02d0aSMike Kravetz } 1457ca02d0aSMike Kravetz spool->rsv_hpages = min_hpages; 14690481622SDavid Gibson 14790481622SDavid Gibson return spool; 14890481622SDavid Gibson } 14990481622SDavid Gibson 15090481622SDavid Gibson void hugepage_put_subpool(struct hugepage_subpool *spool) 15190481622SDavid Gibson { 152db71ef79SMike Kravetz unsigned long flags; 153db71ef79SMike Kravetz 154db71ef79SMike Kravetz spin_lock_irqsave(&spool->lock, flags); 15590481622SDavid Gibson BUG_ON(!spool->count); 15690481622SDavid Gibson spool->count--; 157db71ef79SMike Kravetz unlock_or_release_subpool(spool, flags); 15890481622SDavid Gibson } 15990481622SDavid Gibson 1601c5ecae3SMike Kravetz /* 1611c5ecae3SMike Kravetz * Subpool accounting for allocating and reserving pages. 1621c5ecae3SMike Kravetz * Return -ENOMEM if there are not enough resources to satisfy the 1639e7ee400SRandy Dunlap * request. Otherwise, return the number of pages by which the 1641c5ecae3SMike Kravetz * global pools must be adjusted (upward). The returned value may 1651c5ecae3SMike Kravetz * only be different than the passed value (delta) in the case where 1667c8de358SEthon Paul * a subpool minimum size must be maintained. 1671c5ecae3SMike Kravetz */ 1681c5ecae3SMike Kravetz static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, 16990481622SDavid Gibson long delta) 17090481622SDavid Gibson { 1711c5ecae3SMike Kravetz long ret = delta; 17290481622SDavid Gibson 17390481622SDavid Gibson if (!spool) 1741c5ecae3SMike Kravetz return ret; 17590481622SDavid Gibson 176db71ef79SMike Kravetz spin_lock_irq(&spool->lock); 17790481622SDavid Gibson 1781c5ecae3SMike Kravetz if (spool->max_hpages != -1) { /* maximum size accounting */ 1791c5ecae3SMike Kravetz if ((spool->used_hpages + delta) <= spool->max_hpages) 1801c5ecae3SMike Kravetz spool->used_hpages += delta; 1811c5ecae3SMike Kravetz else { 1821c5ecae3SMike Kravetz ret = -ENOMEM; 1831c5ecae3SMike Kravetz goto unlock_ret; 1841c5ecae3SMike Kravetz } 1851c5ecae3SMike Kravetz } 1861c5ecae3SMike Kravetz 18709a95e29SMike Kravetz /* minimum size accounting */ 18809a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->rsv_hpages) { 1891c5ecae3SMike Kravetz if (delta > spool->rsv_hpages) { 1901c5ecae3SMike Kravetz /* 1911c5ecae3SMike Kravetz * Asking for more reserves than those already taken on 1921c5ecae3SMike Kravetz * behalf of subpool. Return difference. 1931c5ecae3SMike Kravetz */ 1941c5ecae3SMike Kravetz ret = delta - spool->rsv_hpages; 1951c5ecae3SMike Kravetz spool->rsv_hpages = 0; 1961c5ecae3SMike Kravetz } else { 1971c5ecae3SMike Kravetz ret = 0; /* reserves already accounted for */ 1981c5ecae3SMike Kravetz spool->rsv_hpages -= delta; 1991c5ecae3SMike Kravetz } 2001c5ecae3SMike Kravetz } 2011c5ecae3SMike Kravetz 2021c5ecae3SMike Kravetz unlock_ret: 203db71ef79SMike Kravetz spin_unlock_irq(&spool->lock); 20490481622SDavid Gibson return ret; 20590481622SDavid Gibson } 20690481622SDavid Gibson 2071c5ecae3SMike Kravetz /* 2081c5ecae3SMike Kravetz * Subpool accounting for freeing and unreserving pages. 2091c5ecae3SMike Kravetz * Return the number of global page reservations that must be dropped. 2101c5ecae3SMike Kravetz * The return value may only be different than the passed value (delta) 2111c5ecae3SMike Kravetz * in the case where a subpool minimum size must be maintained. 2121c5ecae3SMike Kravetz */ 2131c5ecae3SMike Kravetz static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, 21490481622SDavid Gibson long delta) 21590481622SDavid Gibson { 2161c5ecae3SMike Kravetz long ret = delta; 217db71ef79SMike Kravetz unsigned long flags; 2181c5ecae3SMike Kravetz 21990481622SDavid Gibson if (!spool) 2201c5ecae3SMike Kravetz return delta; 22190481622SDavid Gibson 222db71ef79SMike Kravetz spin_lock_irqsave(&spool->lock, flags); 2231c5ecae3SMike Kravetz 2241c5ecae3SMike Kravetz if (spool->max_hpages != -1) /* maximum size accounting */ 22590481622SDavid Gibson spool->used_hpages -= delta; 2261c5ecae3SMike Kravetz 22709a95e29SMike Kravetz /* minimum size accounting */ 22809a95e29SMike Kravetz if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { 2291c5ecae3SMike Kravetz if (spool->rsv_hpages + delta <= spool->min_hpages) 2301c5ecae3SMike Kravetz ret = 0; 2311c5ecae3SMike Kravetz else 2321c5ecae3SMike Kravetz ret = spool->rsv_hpages + delta - spool->min_hpages; 2331c5ecae3SMike Kravetz 2341c5ecae3SMike Kravetz spool->rsv_hpages += delta; 2351c5ecae3SMike Kravetz if (spool->rsv_hpages > spool->min_hpages) 2361c5ecae3SMike Kravetz spool->rsv_hpages = spool->min_hpages; 2371c5ecae3SMike Kravetz } 2381c5ecae3SMike Kravetz 2391c5ecae3SMike Kravetz /* 2401c5ecae3SMike Kravetz * If hugetlbfs_put_super couldn't free spool due to an outstanding 2411c5ecae3SMike Kravetz * quota reference, free it now. 2421c5ecae3SMike Kravetz */ 243db71ef79SMike Kravetz unlock_or_release_subpool(spool, flags); 2441c5ecae3SMike Kravetz 2451c5ecae3SMike Kravetz return ret; 24690481622SDavid Gibson } 24790481622SDavid Gibson 24890481622SDavid Gibson static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 24990481622SDavid Gibson { 25090481622SDavid Gibson return HUGETLBFS_SB(inode->i_sb)->spool; 25190481622SDavid Gibson } 25290481622SDavid Gibson 25390481622SDavid Gibson static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 25490481622SDavid Gibson { 255496ad9aaSAl Viro return subpool_inode(file_inode(vma->vm_file)); 25690481622SDavid Gibson } 25790481622SDavid Gibson 2580db9d74eSMina Almasry /* Helper that removes a struct file_region from the resv_map cache and returns 2590db9d74eSMina Almasry * it for use. 2600db9d74eSMina Almasry */ 2610db9d74eSMina Almasry static struct file_region * 2620db9d74eSMina Almasry get_file_region_entry_from_cache(struct resv_map *resv, long from, long to) 2630db9d74eSMina Almasry { 2643259914fSXU pengfei struct file_region *nrg; 2650db9d74eSMina Almasry 2660db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count <= 0); 2670db9d74eSMina Almasry 2680db9d74eSMina Almasry resv->region_cache_count--; 2690db9d74eSMina Almasry nrg = list_first_entry(&resv->region_cache, struct file_region, link); 2700db9d74eSMina Almasry list_del(&nrg->link); 2710db9d74eSMina Almasry 2720db9d74eSMina Almasry nrg->from = from; 2730db9d74eSMina Almasry nrg->to = to; 2740db9d74eSMina Almasry 2750db9d74eSMina Almasry return nrg; 2760db9d74eSMina Almasry } 2770db9d74eSMina Almasry 278075a61d0SMina Almasry static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, 279075a61d0SMina Almasry struct file_region *rg) 280075a61d0SMina Almasry { 281075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 282075a61d0SMina Almasry nrg->reservation_counter = rg->reservation_counter; 283075a61d0SMina Almasry nrg->css = rg->css; 284075a61d0SMina Almasry if (rg->css) 285075a61d0SMina Almasry css_get(rg->css); 286075a61d0SMina Almasry #endif 287075a61d0SMina Almasry } 288075a61d0SMina Almasry 289075a61d0SMina Almasry /* Helper that records hugetlb_cgroup uncharge info. */ 290075a61d0SMina Almasry static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, 291075a61d0SMina Almasry struct hstate *h, 292075a61d0SMina Almasry struct resv_map *resv, 293075a61d0SMina Almasry struct file_region *nrg) 294075a61d0SMina Almasry { 295075a61d0SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 296075a61d0SMina Almasry if (h_cg) { 297075a61d0SMina Almasry nrg->reservation_counter = 298075a61d0SMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 299075a61d0SMina Almasry nrg->css = &h_cg->css; 300d85aecf2SMiaohe Lin /* 301d85aecf2SMiaohe Lin * The caller will hold exactly one h_cg->css reference for the 302d85aecf2SMiaohe Lin * whole contiguous reservation region. But this area might be 303d85aecf2SMiaohe Lin * scattered when there are already some file_regions reside in 304d85aecf2SMiaohe Lin * it. As a result, many file_regions may share only one css 305d85aecf2SMiaohe Lin * reference. In order to ensure that one file_region must hold 306d85aecf2SMiaohe Lin * exactly one h_cg->css reference, we should do css_get for 307d85aecf2SMiaohe Lin * each file_region and leave the reference held by caller 308d85aecf2SMiaohe Lin * untouched. 309d85aecf2SMiaohe Lin */ 310d85aecf2SMiaohe Lin css_get(&h_cg->css); 311075a61d0SMina Almasry if (!resv->pages_per_hpage) 312075a61d0SMina Almasry resv->pages_per_hpage = pages_per_huge_page(h); 313075a61d0SMina Almasry /* pages_per_hpage should be the same for all entries in 314075a61d0SMina Almasry * a resv_map. 315075a61d0SMina Almasry */ 316075a61d0SMina Almasry VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); 317075a61d0SMina Almasry } else { 318075a61d0SMina Almasry nrg->reservation_counter = NULL; 319075a61d0SMina Almasry nrg->css = NULL; 320075a61d0SMina Almasry } 321075a61d0SMina Almasry #endif 322075a61d0SMina Almasry } 323075a61d0SMina Almasry 324d85aecf2SMiaohe Lin static void put_uncharge_info(struct file_region *rg) 325d85aecf2SMiaohe Lin { 326d85aecf2SMiaohe Lin #ifdef CONFIG_CGROUP_HUGETLB 327d85aecf2SMiaohe Lin if (rg->css) 328d85aecf2SMiaohe Lin css_put(rg->css); 329d85aecf2SMiaohe Lin #endif 330d85aecf2SMiaohe Lin } 331d85aecf2SMiaohe Lin 332a9b3f867SMina Almasry static bool has_same_uncharge_info(struct file_region *rg, 333a9b3f867SMina Almasry struct file_region *org) 334a9b3f867SMina Almasry { 335a9b3f867SMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 3360739eb43SBaolin Wang return rg->reservation_counter == org->reservation_counter && 337a9b3f867SMina Almasry rg->css == org->css; 338a9b3f867SMina Almasry 339a9b3f867SMina Almasry #else 340a9b3f867SMina Almasry return true; 341a9b3f867SMina Almasry #endif 342a9b3f867SMina Almasry } 343a9b3f867SMina Almasry 344a9b3f867SMina Almasry static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) 345a9b3f867SMina Almasry { 3463259914fSXU pengfei struct file_region *nrg, *prg; 347a9b3f867SMina Almasry 348a9b3f867SMina Almasry prg = list_prev_entry(rg, link); 349a9b3f867SMina Almasry if (&prg->link != &resv->regions && prg->to == rg->from && 350a9b3f867SMina Almasry has_same_uncharge_info(prg, rg)) { 351a9b3f867SMina Almasry prg->to = rg->to; 352a9b3f867SMina Almasry 353a9b3f867SMina Almasry list_del(&rg->link); 354d85aecf2SMiaohe Lin put_uncharge_info(rg); 355a9b3f867SMina Almasry kfree(rg); 356a9b3f867SMina Almasry 3577db5e7b6SWei Yang rg = prg; 358a9b3f867SMina Almasry } 359a9b3f867SMina Almasry 360a9b3f867SMina Almasry nrg = list_next_entry(rg, link); 361a9b3f867SMina Almasry if (&nrg->link != &resv->regions && nrg->from == rg->to && 362a9b3f867SMina Almasry has_same_uncharge_info(nrg, rg)) { 363a9b3f867SMina Almasry nrg->from = rg->from; 364a9b3f867SMina Almasry 365a9b3f867SMina Almasry list_del(&rg->link); 366d85aecf2SMiaohe Lin put_uncharge_info(rg); 367a9b3f867SMina Almasry kfree(rg); 368a9b3f867SMina Almasry } 369a9b3f867SMina Almasry } 370a9b3f867SMina Almasry 3712103cf9cSPeter Xu static inline long 37284448c8eSJakob Koschel hugetlb_resv_map_add(struct resv_map *map, struct list_head *rg, long from, 3732103cf9cSPeter Xu long to, struct hstate *h, struct hugetlb_cgroup *cg, 3742103cf9cSPeter Xu long *regions_needed) 3752103cf9cSPeter Xu { 3762103cf9cSPeter Xu struct file_region *nrg; 3772103cf9cSPeter Xu 3782103cf9cSPeter Xu if (!regions_needed) { 3792103cf9cSPeter Xu nrg = get_file_region_entry_from_cache(map, from, to); 3802103cf9cSPeter Xu record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg); 38184448c8eSJakob Koschel list_add(&nrg->link, rg); 3822103cf9cSPeter Xu coalesce_file_region(map, nrg); 3832103cf9cSPeter Xu } else 3842103cf9cSPeter Xu *regions_needed += 1; 3852103cf9cSPeter Xu 3862103cf9cSPeter Xu return to - from; 3872103cf9cSPeter Xu } 3882103cf9cSPeter Xu 389972a3da3SWei Yang /* 390972a3da3SWei Yang * Must be called with resv->lock held. 391972a3da3SWei Yang * 392972a3da3SWei Yang * Calling this with regions_needed != NULL will count the number of pages 393972a3da3SWei Yang * to be added but will not modify the linked list. And regions_needed will 394972a3da3SWei Yang * indicate the number of file_regions needed in the cache to carry out to add 395972a3da3SWei Yang * the regions for this range. 396d75c6af9SMina Almasry */ 397d75c6af9SMina Almasry static long add_reservation_in_range(struct resv_map *resv, long f, long t, 398075a61d0SMina Almasry struct hugetlb_cgroup *h_cg, 399972a3da3SWei Yang struct hstate *h, long *regions_needed) 400d75c6af9SMina Almasry { 4010db9d74eSMina Almasry long add = 0; 402d75c6af9SMina Almasry struct list_head *head = &resv->regions; 4030db9d74eSMina Almasry long last_accounted_offset = f; 40484448c8eSJakob Koschel struct file_region *iter, *trg = NULL; 40584448c8eSJakob Koschel struct list_head *rg = NULL; 406d75c6af9SMina Almasry 4070db9d74eSMina Almasry if (regions_needed) 4080db9d74eSMina Almasry *regions_needed = 0; 409d75c6af9SMina Almasry 4100db9d74eSMina Almasry /* In this loop, we essentially handle an entry for the range 41184448c8eSJakob Koschel * [last_accounted_offset, iter->from), at every iteration, with some 4120db9d74eSMina Almasry * bounds checking. 4130db9d74eSMina Almasry */ 41484448c8eSJakob Koschel list_for_each_entry_safe(iter, trg, head, link) { 4150db9d74eSMina Almasry /* Skip irrelevant regions that start before our range. */ 41684448c8eSJakob Koschel if (iter->from < f) { 4170db9d74eSMina Almasry /* If this region ends after the last accounted offset, 4180db9d74eSMina Almasry * then we need to update last_accounted_offset. 4190db9d74eSMina Almasry */ 42084448c8eSJakob Koschel if (iter->to > last_accounted_offset) 42184448c8eSJakob Koschel last_accounted_offset = iter->to; 4220db9d74eSMina Almasry continue; 4230db9d74eSMina Almasry } 424d75c6af9SMina Almasry 4250db9d74eSMina Almasry /* When we find a region that starts beyond our range, we've 4260db9d74eSMina Almasry * finished. 4270db9d74eSMina Almasry */ 42884448c8eSJakob Koschel if (iter->from >= t) { 42984448c8eSJakob Koschel rg = iter->link.prev; 430d75c6af9SMina Almasry break; 43184448c8eSJakob Koschel } 432d75c6af9SMina Almasry 43384448c8eSJakob Koschel /* Add an entry for last_accounted_offset -> iter->from, and 4340db9d74eSMina Almasry * update last_accounted_offset. 435d75c6af9SMina Almasry */ 43684448c8eSJakob Koschel if (iter->from > last_accounted_offset) 43784448c8eSJakob Koschel add += hugetlb_resv_map_add(resv, iter->link.prev, 4382103cf9cSPeter Xu last_accounted_offset, 43984448c8eSJakob Koschel iter->from, h, h_cg, 4402103cf9cSPeter Xu regions_needed); 441d75c6af9SMina Almasry 44284448c8eSJakob Koschel last_accounted_offset = iter->to; 4430db9d74eSMina Almasry } 4440db9d74eSMina Almasry 4450db9d74eSMina Almasry /* Handle the case where our range extends beyond 4460db9d74eSMina Almasry * last_accounted_offset. 4470db9d74eSMina Almasry */ 44884448c8eSJakob Koschel if (!rg) 44984448c8eSJakob Koschel rg = head->prev; 4502103cf9cSPeter Xu if (last_accounted_offset < t) 4512103cf9cSPeter Xu add += hugetlb_resv_map_add(resv, rg, last_accounted_offset, 4522103cf9cSPeter Xu t, h, h_cg, regions_needed); 4530db9d74eSMina Almasry 4540db9d74eSMina Almasry return add; 4550db9d74eSMina Almasry } 4560db9d74eSMina Almasry 4570db9d74eSMina Almasry /* Must be called with resv->lock acquired. Will drop lock to allocate entries. 4580db9d74eSMina Almasry */ 4590db9d74eSMina Almasry static int allocate_file_region_entries(struct resv_map *resv, 4600db9d74eSMina Almasry int regions_needed) 4610db9d74eSMina Almasry __must_hold(&resv->lock) 4620db9d74eSMina Almasry { 46334665341SMiaohe Lin LIST_HEAD(allocated_regions); 4640db9d74eSMina Almasry int to_allocate = 0, i = 0; 4650db9d74eSMina Almasry struct file_region *trg = NULL, *rg = NULL; 4660db9d74eSMina Almasry 4670db9d74eSMina Almasry VM_BUG_ON(regions_needed < 0); 4680db9d74eSMina Almasry 4690db9d74eSMina Almasry /* 4700db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 4710db9d74eSMina Almasry * the number of in progress add operations plus regions_needed. 4720db9d74eSMina Almasry * 4730db9d74eSMina Almasry * This is a while loop because when we drop the lock, some other call 4740db9d74eSMina Almasry * to region_add or region_del may have consumed some region_entries, 4750db9d74eSMina Almasry * so we keep looping here until we finally have enough entries for 4760db9d74eSMina Almasry * (adds_in_progress + regions_needed). 4770db9d74eSMina Almasry */ 4780db9d74eSMina Almasry while (resv->region_cache_count < 4790db9d74eSMina Almasry (resv->adds_in_progress + regions_needed)) { 4800db9d74eSMina Almasry to_allocate = resv->adds_in_progress + regions_needed - 4810db9d74eSMina Almasry resv->region_cache_count; 4820db9d74eSMina Almasry 4830db9d74eSMina Almasry /* At this point, we should have enough entries in the cache 484f0953a1bSIngo Molnar * for all the existing adds_in_progress. We should only be 4850db9d74eSMina Almasry * needing to allocate for regions_needed. 4860db9d74eSMina Almasry */ 4870db9d74eSMina Almasry VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); 4880db9d74eSMina Almasry 4890db9d74eSMina Almasry spin_unlock(&resv->lock); 4900db9d74eSMina Almasry for (i = 0; i < to_allocate; i++) { 4910db9d74eSMina Almasry trg = kmalloc(sizeof(*trg), GFP_KERNEL); 4920db9d74eSMina Almasry if (!trg) 4930db9d74eSMina Almasry goto out_of_memory; 4940db9d74eSMina Almasry list_add(&trg->link, &allocated_regions); 4950db9d74eSMina Almasry } 4960db9d74eSMina Almasry 4970db9d74eSMina Almasry spin_lock(&resv->lock); 4980db9d74eSMina Almasry 499d3ec7b6eSWei Yang list_splice(&allocated_regions, &resv->region_cache); 500d3ec7b6eSWei Yang resv->region_cache_count += to_allocate; 5010db9d74eSMina Almasry } 5020db9d74eSMina Almasry 5030db9d74eSMina Almasry return 0; 5040db9d74eSMina Almasry 5050db9d74eSMina Almasry out_of_memory: 5060db9d74eSMina Almasry list_for_each_entry_safe(rg, trg, &allocated_regions, link) { 507d75c6af9SMina Almasry list_del(&rg->link); 508d75c6af9SMina Almasry kfree(rg); 509d75c6af9SMina Almasry } 5100db9d74eSMina Almasry return -ENOMEM; 511d75c6af9SMina Almasry } 512d75c6af9SMina Almasry 5131dd308a7SMike Kravetz /* 5141dd308a7SMike Kravetz * Add the huge page range represented by [f, t) to the reserve 5150db9d74eSMina Almasry * map. Regions will be taken from the cache to fill in this range. 5160db9d74eSMina Almasry * Sufficient regions should exist in the cache due to the previous 5170db9d74eSMina Almasry * call to region_chg with the same range, but in some cases the cache will not 5180db9d74eSMina Almasry * have sufficient entries due to races with other code doing region_add or 5190db9d74eSMina Almasry * region_del. The extra needed entries will be allocated. 520cf3ad20bSMike Kravetz * 5210db9d74eSMina Almasry * regions_needed is the out value provided by a previous call to region_chg. 5220db9d74eSMina Almasry * 5230db9d74eSMina Almasry * Return the number of new huge pages added to the map. This number is greater 5240db9d74eSMina Almasry * than or equal to zero. If file_region entries needed to be allocated for 5257c8de358SEthon Paul * this operation and we were not able to allocate, it returns -ENOMEM. 5260db9d74eSMina Almasry * region_add of regions of length 1 never allocate file_regions and cannot 5270db9d74eSMina Almasry * fail; region_chg will always allocate at least 1 entry and a region_add for 5280db9d74eSMina Almasry * 1 page will only require at most 1 entry. 5291dd308a7SMike Kravetz */ 5300db9d74eSMina Almasry static long region_add(struct resv_map *resv, long f, long t, 531075a61d0SMina Almasry long in_regions_needed, struct hstate *h, 532075a61d0SMina Almasry struct hugetlb_cgroup *h_cg) 53396822904SAndy Whitcroft { 5340db9d74eSMina Almasry long add = 0, actual_regions_needed = 0; 53596822904SAndy Whitcroft 5367b24d861SDavidlohr Bueso spin_lock(&resv->lock); 5370db9d74eSMina Almasry retry: 5380db9d74eSMina Almasry 5390db9d74eSMina Almasry /* Count how many regions are actually needed to execute this add. */ 540972a3da3SWei Yang add_reservation_in_range(resv, f, t, NULL, NULL, 541972a3da3SWei Yang &actual_regions_needed); 54296822904SAndy Whitcroft 5435e911373SMike Kravetz /* 5440db9d74eSMina Almasry * Check for sufficient descriptors in the cache to accommodate 5450db9d74eSMina Almasry * this add operation. Note that actual_regions_needed may be greater 5460db9d74eSMina Almasry * than in_regions_needed, as the resv_map may have been modified since 5470db9d74eSMina Almasry * the region_chg call. In this case, we need to make sure that we 5480db9d74eSMina Almasry * allocate extra entries, such that we have enough for all the 5490db9d74eSMina Almasry * existing adds_in_progress, plus the excess needed for this 5500db9d74eSMina Almasry * operation. 5515e911373SMike Kravetz */ 5520db9d74eSMina Almasry if (actual_regions_needed > in_regions_needed && 5530db9d74eSMina Almasry resv->region_cache_count < 5540db9d74eSMina Almasry resv->adds_in_progress + 5550db9d74eSMina Almasry (actual_regions_needed - in_regions_needed)) { 5560db9d74eSMina Almasry /* region_add operation of range 1 should never need to 5570db9d74eSMina Almasry * allocate file_region entries. 5580db9d74eSMina Almasry */ 5590db9d74eSMina Almasry VM_BUG_ON(t - f <= 1); 5605e911373SMike Kravetz 5610db9d74eSMina Almasry if (allocate_file_region_entries( 5620db9d74eSMina Almasry resv, actual_regions_needed - in_regions_needed)) { 5630db9d74eSMina Almasry return -ENOMEM; 5645e911373SMike Kravetz } 5655e911373SMike Kravetz 5660db9d74eSMina Almasry goto retry; 5670db9d74eSMina Almasry } 568cf3ad20bSMike Kravetz 569972a3da3SWei Yang add = add_reservation_in_range(resv, f, t, h_cg, h, NULL); 5700db9d74eSMina Almasry 5710db9d74eSMina Almasry resv->adds_in_progress -= in_regions_needed; 5720db9d74eSMina Almasry 5737b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 574cf3ad20bSMike Kravetz return add; 57596822904SAndy Whitcroft } 57696822904SAndy Whitcroft 5771dd308a7SMike Kravetz /* 5781dd308a7SMike Kravetz * Examine the existing reserve map and determine how many 5791dd308a7SMike Kravetz * huge pages in the specified range [f, t) are NOT currently 5801dd308a7SMike Kravetz * represented. This routine is called before a subsequent 5811dd308a7SMike Kravetz * call to region_add that will actually modify the reserve 5821dd308a7SMike Kravetz * map to add the specified range [f, t). region_chg does 5831dd308a7SMike Kravetz * not change the number of huge pages represented by the 5840db9d74eSMina Almasry * map. A number of new file_region structures is added to the cache as a 5850db9d74eSMina Almasry * placeholder, for the subsequent region_add call to use. At least 1 5860db9d74eSMina Almasry * file_region structure is added. 5870db9d74eSMina Almasry * 5880db9d74eSMina Almasry * out_regions_needed is the number of regions added to the 5890db9d74eSMina Almasry * resv->adds_in_progress. This value needs to be provided to a follow up call 5900db9d74eSMina Almasry * to region_add or region_abort for proper accounting. 5915e911373SMike Kravetz * 5925e911373SMike Kravetz * Returns the number of huge pages that need to be added to the existing 5935e911373SMike Kravetz * reservation map for the range [f, t). This number is greater or equal to 5945e911373SMike Kravetz * zero. -ENOMEM is returned if a new file_region structure or cache entry 5955e911373SMike Kravetz * is needed and can not be allocated. 5961dd308a7SMike Kravetz */ 5970db9d74eSMina Almasry static long region_chg(struct resv_map *resv, long f, long t, 5980db9d74eSMina Almasry long *out_regions_needed) 59996822904SAndy Whitcroft { 60096822904SAndy Whitcroft long chg = 0; 60196822904SAndy Whitcroft 6027b24d861SDavidlohr Bueso spin_lock(&resv->lock); 6035e911373SMike Kravetz 604972a3da3SWei Yang /* Count how many hugepages in this range are NOT represented. */ 605075a61d0SMina Almasry chg = add_reservation_in_range(resv, f, t, NULL, NULL, 606972a3da3SWei Yang out_regions_needed); 6075e911373SMike Kravetz 6080db9d74eSMina Almasry if (*out_regions_needed == 0) 6090db9d74eSMina Almasry *out_regions_needed = 1; 6105e911373SMike Kravetz 6110db9d74eSMina Almasry if (allocate_file_region_entries(resv, *out_regions_needed)) 6125e911373SMike Kravetz return -ENOMEM; 6135e911373SMike Kravetz 6140db9d74eSMina Almasry resv->adds_in_progress += *out_regions_needed; 61596822904SAndy Whitcroft 6167b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 61796822904SAndy Whitcroft return chg; 61896822904SAndy Whitcroft } 61996822904SAndy Whitcroft 6201dd308a7SMike Kravetz /* 6215e911373SMike Kravetz * Abort the in progress add operation. The adds_in_progress field 6225e911373SMike Kravetz * of the resv_map keeps track of the operations in progress between 6235e911373SMike Kravetz * calls to region_chg and region_add. Operations are sometimes 6245e911373SMike Kravetz * aborted after the call to region_chg. In such cases, region_abort 6250db9d74eSMina Almasry * is called to decrement the adds_in_progress counter. regions_needed 6260db9d74eSMina Almasry * is the value returned by the region_chg call, it is used to decrement 6270db9d74eSMina Almasry * the adds_in_progress counter. 6285e911373SMike Kravetz * 6295e911373SMike Kravetz * NOTE: The range arguments [f, t) are not needed or used in this 6305e911373SMike Kravetz * routine. They are kept to make reading the calling code easier as 6315e911373SMike Kravetz * arguments will match the associated region_chg call. 6325e911373SMike Kravetz */ 6330db9d74eSMina Almasry static void region_abort(struct resv_map *resv, long f, long t, 6340db9d74eSMina Almasry long regions_needed) 6355e911373SMike Kravetz { 6365e911373SMike Kravetz spin_lock(&resv->lock); 6375e911373SMike Kravetz VM_BUG_ON(!resv->region_cache_count); 6380db9d74eSMina Almasry resv->adds_in_progress -= regions_needed; 6395e911373SMike Kravetz spin_unlock(&resv->lock); 6405e911373SMike Kravetz } 6415e911373SMike Kravetz 6425e911373SMike Kravetz /* 643feba16e2SMike Kravetz * Delete the specified range [f, t) from the reserve map. If the 644feba16e2SMike Kravetz * t parameter is LONG_MAX, this indicates that ALL regions after f 645feba16e2SMike Kravetz * should be deleted. Locate the regions which intersect [f, t) 646feba16e2SMike Kravetz * and either trim, delete or split the existing regions. 647feba16e2SMike Kravetz * 648feba16e2SMike Kravetz * Returns the number of huge pages deleted from the reserve map. 649feba16e2SMike Kravetz * In the normal case, the return value is zero or more. In the 650feba16e2SMike Kravetz * case where a region must be split, a new region descriptor must 651feba16e2SMike Kravetz * be allocated. If the allocation fails, -ENOMEM will be returned. 652feba16e2SMike Kravetz * NOTE: If the parameter t == LONG_MAX, then we will never split 653feba16e2SMike Kravetz * a region and possibly return -ENOMEM. Callers specifying 654feba16e2SMike Kravetz * t == LONG_MAX do not need to check for -ENOMEM error. 6551dd308a7SMike Kravetz */ 656feba16e2SMike Kravetz static long region_del(struct resv_map *resv, long f, long t) 65796822904SAndy Whitcroft { 6581406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 65996822904SAndy Whitcroft struct file_region *rg, *trg; 660feba16e2SMike Kravetz struct file_region *nrg = NULL; 661feba16e2SMike Kravetz long del = 0; 66296822904SAndy Whitcroft 663feba16e2SMike Kravetz retry: 6647b24d861SDavidlohr Bueso spin_lock(&resv->lock); 665feba16e2SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 666dbe409e4SMike Kravetz /* 667dbe409e4SMike Kravetz * Skip regions before the range to be deleted. file_region 668dbe409e4SMike Kravetz * ranges are normally of the form [from, to). However, there 669dbe409e4SMike Kravetz * may be a "placeholder" entry in the map which is of the form 670dbe409e4SMike Kravetz * (from, to) with from == to. Check for placeholder entries 671dbe409e4SMike Kravetz * at the beginning of the range to be deleted. 672dbe409e4SMike Kravetz */ 673dbe409e4SMike Kravetz if (rg->to <= f && (rg->to != rg->from || rg->to != f)) 674feba16e2SMike Kravetz continue; 675dbe409e4SMike Kravetz 676feba16e2SMike Kravetz if (rg->from >= t) 67796822904SAndy Whitcroft break; 67896822904SAndy Whitcroft 679feba16e2SMike Kravetz if (f > rg->from && t < rg->to) { /* Must split region */ 680feba16e2SMike Kravetz /* 681feba16e2SMike Kravetz * Check for an entry in the cache before dropping 682feba16e2SMike Kravetz * lock and attempting allocation. 683feba16e2SMike Kravetz */ 684feba16e2SMike Kravetz if (!nrg && 685feba16e2SMike Kravetz resv->region_cache_count > resv->adds_in_progress) { 686feba16e2SMike Kravetz nrg = list_first_entry(&resv->region_cache, 687feba16e2SMike Kravetz struct file_region, 688feba16e2SMike Kravetz link); 689feba16e2SMike Kravetz list_del(&nrg->link); 690feba16e2SMike Kravetz resv->region_cache_count--; 69196822904SAndy Whitcroft } 69296822904SAndy Whitcroft 693feba16e2SMike Kravetz if (!nrg) { 694feba16e2SMike Kravetz spin_unlock(&resv->lock); 695feba16e2SMike Kravetz nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 696feba16e2SMike Kravetz if (!nrg) 697feba16e2SMike Kravetz return -ENOMEM; 698feba16e2SMike Kravetz goto retry; 699feba16e2SMike Kravetz } 700feba16e2SMike Kravetz 701feba16e2SMike Kravetz del += t - f; 70279aa925bSMike Kravetz hugetlb_cgroup_uncharge_file_region( 703d85aecf2SMiaohe Lin resv, rg, t - f, false); 704feba16e2SMike Kravetz 705feba16e2SMike Kravetz /* New entry for end of split region */ 706feba16e2SMike Kravetz nrg->from = t; 707feba16e2SMike Kravetz nrg->to = rg->to; 708075a61d0SMina Almasry 709075a61d0SMina Almasry copy_hugetlb_cgroup_uncharge_info(nrg, rg); 710075a61d0SMina Almasry 711feba16e2SMike Kravetz INIT_LIST_HEAD(&nrg->link); 712feba16e2SMike Kravetz 713feba16e2SMike Kravetz /* Original entry is trimmed */ 714feba16e2SMike Kravetz rg->to = f; 715feba16e2SMike Kravetz 716feba16e2SMike Kravetz list_add(&nrg->link, &rg->link); 717feba16e2SMike Kravetz nrg = NULL; 71896822904SAndy Whitcroft break; 719feba16e2SMike Kravetz } 720feba16e2SMike Kravetz 721feba16e2SMike Kravetz if (f <= rg->from && t >= rg->to) { /* Remove entire region */ 722feba16e2SMike Kravetz del += rg->to - rg->from; 723075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 724d85aecf2SMiaohe Lin rg->to - rg->from, true); 72596822904SAndy Whitcroft list_del(&rg->link); 72696822904SAndy Whitcroft kfree(rg); 727feba16e2SMike Kravetz continue; 72896822904SAndy Whitcroft } 7297b24d861SDavidlohr Bueso 730feba16e2SMike Kravetz if (f <= rg->from) { /* Trim beginning of region */ 731075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 732d85aecf2SMiaohe Lin t - rg->from, false); 733075a61d0SMina Almasry 73479aa925bSMike Kravetz del += t - rg->from; 73579aa925bSMike Kravetz rg->from = t; 73679aa925bSMike Kravetz } else { /* Trim end of region */ 737075a61d0SMina Almasry hugetlb_cgroup_uncharge_file_region(resv, rg, 738d85aecf2SMiaohe Lin rg->to - f, false); 73979aa925bSMike Kravetz 74079aa925bSMike Kravetz del += rg->to - f; 74179aa925bSMike Kravetz rg->to = f; 742feba16e2SMike Kravetz } 743feba16e2SMike Kravetz } 744feba16e2SMike Kravetz 7457b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 746feba16e2SMike Kravetz kfree(nrg); 747feba16e2SMike Kravetz return del; 74896822904SAndy Whitcroft } 74996822904SAndy Whitcroft 7501dd308a7SMike Kravetz /* 751b5cec28dSMike Kravetz * A rare out of memory error was encountered which prevented removal of 752b5cec28dSMike Kravetz * the reserve map region for a page. The huge page itself was free'ed 753b5cec28dSMike Kravetz * and removed from the page cache. This routine will adjust the subpool 754b5cec28dSMike Kravetz * usage count, and the global reserve count if needed. By incrementing 755b5cec28dSMike Kravetz * these counts, the reserve map entry which could not be deleted will 756b5cec28dSMike Kravetz * appear as a "reserved" entry instead of simply dangling with incorrect 757b5cec28dSMike Kravetz * counts. 758b5cec28dSMike Kravetz */ 75972e2936cSzhong jiang void hugetlb_fix_reserve_counts(struct inode *inode) 760b5cec28dSMike Kravetz { 761b5cec28dSMike Kravetz struct hugepage_subpool *spool = subpool_inode(inode); 762b5cec28dSMike Kravetz long rsv_adjust; 763da56388cSMiaohe Lin bool reserved = false; 764b5cec28dSMike Kravetz 765b5cec28dSMike Kravetz rsv_adjust = hugepage_subpool_get_pages(spool, 1); 766da56388cSMiaohe Lin if (rsv_adjust > 0) { 767b5cec28dSMike Kravetz struct hstate *h = hstate_inode(inode); 768b5cec28dSMike Kravetz 769da56388cSMiaohe Lin if (!hugetlb_acct_memory(h, 1)) 770da56388cSMiaohe Lin reserved = true; 771da56388cSMiaohe Lin } else if (!rsv_adjust) { 772da56388cSMiaohe Lin reserved = true; 773b5cec28dSMike Kravetz } 774da56388cSMiaohe Lin 775da56388cSMiaohe Lin if (!reserved) 776da56388cSMiaohe Lin pr_warn("hugetlb: Huge Page Reserved count may go negative.\n"); 777b5cec28dSMike Kravetz } 778b5cec28dSMike Kravetz 779b5cec28dSMike Kravetz /* 7801dd308a7SMike Kravetz * Count and return the number of huge pages in the reserve map 7811dd308a7SMike Kravetz * that intersect with the range [f, t). 7821dd308a7SMike Kravetz */ 7831406ec9bSJoonsoo Kim static long region_count(struct resv_map *resv, long f, long t) 78484afd99bSAndy Whitcroft { 7851406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 78684afd99bSAndy Whitcroft struct file_region *rg; 78784afd99bSAndy Whitcroft long chg = 0; 78884afd99bSAndy Whitcroft 7897b24d861SDavidlohr Bueso spin_lock(&resv->lock); 79084afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 79184afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 792f2135a4aSWang Sheng-Hui long seg_from; 793f2135a4aSWang Sheng-Hui long seg_to; 79484afd99bSAndy Whitcroft 79584afd99bSAndy Whitcroft if (rg->to <= f) 79684afd99bSAndy Whitcroft continue; 79784afd99bSAndy Whitcroft if (rg->from >= t) 79884afd99bSAndy Whitcroft break; 79984afd99bSAndy Whitcroft 80084afd99bSAndy Whitcroft seg_from = max(rg->from, f); 80184afd99bSAndy Whitcroft seg_to = min(rg->to, t); 80284afd99bSAndy Whitcroft 80384afd99bSAndy Whitcroft chg += seg_to - seg_from; 80484afd99bSAndy Whitcroft } 8057b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 80684afd99bSAndy Whitcroft 80784afd99bSAndy Whitcroft return chg; 80884afd99bSAndy Whitcroft } 80984afd99bSAndy Whitcroft 81096822904SAndy Whitcroft /* 811e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 812e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 813e7c4b0bfSAndy Whitcroft */ 814a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 815a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 816e7c4b0bfSAndy Whitcroft { 817a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 818a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 819e7c4b0bfSAndy Whitcroft } 820e7c4b0bfSAndy Whitcroft 8210fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 8220fe6e20bSNaoya Horiguchi unsigned long address) 8230fe6e20bSNaoya Horiguchi { 8240fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 8250fe6e20bSNaoya Horiguchi } 826dee41079SDan Williams EXPORT_SYMBOL_GPL(linear_hugepage_index); 8270fe6e20bSNaoya Horiguchi 82884afd99bSAndy Whitcroft /* 82908fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 83008fba699SMel Gorman * cases this will be same size as used by the page table entries. 83108fba699SMel Gorman */ 83208fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 83308fba699SMel Gorman { 83405ea8860SDan Williams if (vma->vm_ops && vma->vm_ops->pagesize) 83505ea8860SDan Williams return vma->vm_ops->pagesize(vma); 83608fba699SMel Gorman return PAGE_SIZE; 83708fba699SMel Gorman } 838f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 83908fba699SMel Gorman 84008fba699SMel Gorman /* 8413340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 8423340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 84309135cc5SDan Williams * architectures where it differs, an architecture-specific 'strong' 84409135cc5SDan Williams * version of this symbol is required. 8453340289dSMel Gorman */ 84609135cc5SDan Williams __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 8473340289dSMel Gorman { 8483340289dSMel Gorman return vma_kernel_pagesize(vma); 8493340289dSMel Gorman } 8503340289dSMel Gorman 8513340289dSMel Gorman /* 85284afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 85384afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 85484afd99bSAndy Whitcroft * alignment. 85584afd99bSAndy Whitcroft */ 85684afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 85784afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 85804f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 85984afd99bSAndy Whitcroft 860a1e78772SMel Gorman /* 861a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 862a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 863a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 864a1e78772SMel Gorman * 8658d9bfb26SMike Kravetz * With the exception of hugetlb_dup_vma_private() which is called at fork(), 866a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 867a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 868a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 86984afd99bSAndy Whitcroft * 87084afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 87184afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 87284afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 87384afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 87484afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 87584afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 87684afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 87784afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 878a1e78772SMel Gorman */ 879e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 880e7c4b0bfSAndy Whitcroft { 881e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 882e7c4b0bfSAndy Whitcroft } 883e7c4b0bfSAndy Whitcroft 884e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 885e7c4b0bfSAndy Whitcroft unsigned long value) 886e7c4b0bfSAndy Whitcroft { 887e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 888e7c4b0bfSAndy Whitcroft } 889e7c4b0bfSAndy Whitcroft 890e9fe92aeSMina Almasry static void 891e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, 892e9fe92aeSMina Almasry struct hugetlb_cgroup *h_cg, 893e9fe92aeSMina Almasry struct hstate *h) 894e9fe92aeSMina Almasry { 895e9fe92aeSMina Almasry #ifdef CONFIG_CGROUP_HUGETLB 896e9fe92aeSMina Almasry if (!h_cg || !h) { 897e9fe92aeSMina Almasry resv_map->reservation_counter = NULL; 898e9fe92aeSMina Almasry resv_map->pages_per_hpage = 0; 899e9fe92aeSMina Almasry resv_map->css = NULL; 900e9fe92aeSMina Almasry } else { 901e9fe92aeSMina Almasry resv_map->reservation_counter = 902e9fe92aeSMina Almasry &h_cg->rsvd_hugepage[hstate_index(h)]; 903e9fe92aeSMina Almasry resv_map->pages_per_hpage = pages_per_huge_page(h); 904e9fe92aeSMina Almasry resv_map->css = &h_cg->css; 905e9fe92aeSMina Almasry } 906e9fe92aeSMina Almasry #endif 907e9fe92aeSMina Almasry } 908e9fe92aeSMina Almasry 9099119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 91084afd99bSAndy Whitcroft { 91184afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 9125e911373SMike Kravetz struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 9135e911373SMike Kravetz 9145e911373SMike Kravetz if (!resv_map || !rg) { 9155e911373SMike Kravetz kfree(resv_map); 9165e911373SMike Kravetz kfree(rg); 91784afd99bSAndy Whitcroft return NULL; 9185e911373SMike Kravetz } 91984afd99bSAndy Whitcroft 92084afd99bSAndy Whitcroft kref_init(&resv_map->refs); 9217b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 92284afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 92384afd99bSAndy Whitcroft 9245e911373SMike Kravetz resv_map->adds_in_progress = 0; 925e9fe92aeSMina Almasry /* 926e9fe92aeSMina Almasry * Initialize these to 0. On shared mappings, 0's here indicate these 927e9fe92aeSMina Almasry * fields don't do cgroup accounting. On private mappings, these will be 928e9fe92aeSMina Almasry * re-initialized to the proper values, to indicate that hugetlb cgroup 929e9fe92aeSMina Almasry * reservations are to be un-charged from here. 930e9fe92aeSMina Almasry */ 931e9fe92aeSMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); 9325e911373SMike Kravetz 9335e911373SMike Kravetz INIT_LIST_HEAD(&resv_map->region_cache); 9345e911373SMike Kravetz list_add(&rg->link, &resv_map->region_cache); 9355e911373SMike Kravetz resv_map->region_cache_count = 1; 9365e911373SMike Kravetz 93784afd99bSAndy Whitcroft return resv_map; 93884afd99bSAndy Whitcroft } 93984afd99bSAndy Whitcroft 9409119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 94184afd99bSAndy Whitcroft { 94284afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 9435e911373SMike Kravetz struct list_head *head = &resv_map->region_cache; 9445e911373SMike Kravetz struct file_region *rg, *trg; 94584afd99bSAndy Whitcroft 94684afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 947feba16e2SMike Kravetz region_del(resv_map, 0, LONG_MAX); 9485e911373SMike Kravetz 9495e911373SMike Kravetz /* ... and any entries left in the cache */ 9505e911373SMike Kravetz list_for_each_entry_safe(rg, trg, head, link) { 9515e911373SMike Kravetz list_del(&rg->link); 9525e911373SMike Kravetz kfree(rg); 9535e911373SMike Kravetz } 9545e911373SMike Kravetz 9555e911373SMike Kravetz VM_BUG_ON(resv_map->adds_in_progress); 9565e911373SMike Kravetz 95784afd99bSAndy Whitcroft kfree(resv_map); 95884afd99bSAndy Whitcroft } 95984afd99bSAndy Whitcroft 9604e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 9614e35f483SJoonsoo Kim { 962f27a5136SMike Kravetz /* 963f27a5136SMike Kravetz * At inode evict time, i_mapping may not point to the original 964f27a5136SMike Kravetz * address space within the inode. This original address space 965f27a5136SMike Kravetz * contains the pointer to the resv_map. So, always use the 966f27a5136SMike Kravetz * address space embedded within the inode. 967f27a5136SMike Kravetz * The VERY common case is inode->mapping == &inode->i_data but, 968f27a5136SMike Kravetz * this may not be true for device special inodes. 969f27a5136SMike Kravetz */ 970f27a5136SMike Kravetz return (struct resv_map *)(&inode->i_data)->private_data; 9714e35f483SJoonsoo Kim } 9724e35f483SJoonsoo Kim 97384afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 974a1e78772SMel Gorman { 97581d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 9764e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 9774e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 9784e35f483SJoonsoo Kim struct inode *inode = mapping->host; 9794e35f483SJoonsoo Kim 9804e35f483SJoonsoo Kim return inode_resv_map(inode); 9814e35f483SJoonsoo Kim 9824e35f483SJoonsoo Kim } else { 98384afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 98484afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 9854e35f483SJoonsoo Kim } 986a1e78772SMel Gorman } 987a1e78772SMel Gorman 98884afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 989a1e78772SMel Gorman { 99081d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 99181d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 992a1e78772SMel Gorman 99384afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 99484afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 99504f2cbe3SMel Gorman } 99604f2cbe3SMel Gorman 99704f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 99804f2cbe3SMel Gorman { 99981d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 100081d1b09cSSasha Levin VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 1001e7c4b0bfSAndy Whitcroft 1002e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 100304f2cbe3SMel Gorman } 100404f2cbe3SMel Gorman 100504f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 100604f2cbe3SMel Gorman { 100781d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 1008e7c4b0bfSAndy Whitcroft 1009e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 1010a1e78772SMel Gorman } 1011a1e78772SMel Gorman 10128d9bfb26SMike Kravetz void hugetlb_dup_vma_private(struct vm_area_struct *vma) 1013a1e78772SMel Gorman { 101481d1b09cSSasha Levin VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 10158d9bfb26SMike Kravetz /* 10168d9bfb26SMike Kravetz * Clear vm_private_data 10178d9bfb26SMike Kravetz * - For MAP_PRIVATE mappings, this is the reserve map which does 10188d9bfb26SMike Kravetz * not apply to children. Faults generated by the children are 10198d9bfb26SMike Kravetz * not guaranteed to succeed, even if read-only. 10208d9bfb26SMike Kravetz * - For shared mappings this is a per-vma semaphore that may be 10218d9bfb26SMike Kravetz * allocated in a subsequent call to hugetlb_vm_op_open. 10228d9bfb26SMike Kravetz */ 1023a1e78772SMel Gorman vma->vm_private_data = (void *)0; 10248d9bfb26SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 10258d9bfb26SMike Kravetz return; 1026a1e78772SMel Gorman } 1027a1e78772SMel Gorman 1028550a7d60SMina Almasry /* 1029550a7d60SMina Almasry * Reset and decrement one ref on hugepage private reservation. 1030550a7d60SMina Almasry * Called with mm->mmap_sem writer semaphore held. 1031550a7d60SMina Almasry * This function should be only used by move_vma() and operate on 1032550a7d60SMina Almasry * same sized vma. It should never come here with last ref on the 1033550a7d60SMina Almasry * reservation. 1034550a7d60SMina Almasry */ 1035550a7d60SMina Almasry void clear_vma_resv_huge_pages(struct vm_area_struct *vma) 1036550a7d60SMina Almasry { 1037550a7d60SMina Almasry /* 1038550a7d60SMina Almasry * Clear the old hugetlb private page reservation. 1039550a7d60SMina Almasry * It has already been transferred to new_vma. 1040550a7d60SMina Almasry * 1041550a7d60SMina Almasry * During a mremap() operation of a hugetlb vma we call move_vma() 1042550a7d60SMina Almasry * which copies vma into new_vma and unmaps vma. After the copy 1043550a7d60SMina Almasry * operation both new_vma and vma share a reference to the resv_map 1044550a7d60SMina Almasry * struct, and at that point vma is about to be unmapped. We don't 1045550a7d60SMina Almasry * want to return the reservation to the pool at unmap of vma because 1046550a7d60SMina Almasry * the reservation still lives on in new_vma, so simply decrement the 1047550a7d60SMina Almasry * ref here and remove the resv_map reference from this vma. 1048550a7d60SMina Almasry */ 1049550a7d60SMina Almasry struct resv_map *reservations = vma_resv_map(vma); 1050550a7d60SMina Almasry 1051afe041c2SBui Quang Minh if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 1052afe041c2SBui Quang Minh resv_map_put_hugetlb_cgroup_uncharge_info(reservations); 1053550a7d60SMina Almasry kref_put(&reservations->refs, resv_map_release); 1054afe041c2SBui Quang Minh } 1055550a7d60SMina Almasry 10568d9bfb26SMike Kravetz hugetlb_dup_vma_private(vma); 1057550a7d60SMina Almasry } 1058550a7d60SMina Almasry 1059a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 1060559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 1061a1e78772SMel Gorman { 1062af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 1063af0ed73eSJoonsoo Kim /* 1064af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 1065af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 1066af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 1067af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 1068af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 1069af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 1070af0ed73eSJoonsoo Kim * properly, so add work-around here. 1071af0ed73eSJoonsoo Kim */ 1072af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 1073559ec2f8SNicholas Krause return true; 1074af0ed73eSJoonsoo Kim else 1075559ec2f8SNicholas Krause return false; 1076af0ed73eSJoonsoo Kim } 1077a63884e9SJoonsoo Kim 1078a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 10791fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 10801fb1b0e9SMike Kravetz /* 10811fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 10821fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 10831fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 10847c8de358SEthon Paul * fallocate. In this case, there really are no reserves to 10851fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 10861fb1b0e9SMike Kravetz */ 10871fb1b0e9SMike Kravetz if (chg) 10881fb1b0e9SMike Kravetz return false; 10891fb1b0e9SMike Kravetz else 1090559ec2f8SNicholas Krause return true; 10911fb1b0e9SMike Kravetz } 1092a63884e9SJoonsoo Kim 1093a63884e9SJoonsoo Kim /* 1094a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 1095a63884e9SJoonsoo Kim * private mappings. 1096a63884e9SJoonsoo Kim */ 109767961f9dSMike Kravetz if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 109867961f9dSMike Kravetz /* 109967961f9dSMike Kravetz * Like the shared case above, a hole punch or truncate 110067961f9dSMike Kravetz * could have been performed on the private mapping. 110167961f9dSMike Kravetz * Examine the value of chg to determine if reserves 110267961f9dSMike Kravetz * actually exist or were previously consumed. 110367961f9dSMike Kravetz * Very Subtle - The value of chg comes from a previous 110467961f9dSMike Kravetz * call to vma_needs_reserves(). The reserve map for 110567961f9dSMike Kravetz * private mappings has different (opposite) semantics 110667961f9dSMike Kravetz * than that of shared mappings. vma_needs_reserves() 110767961f9dSMike Kravetz * has already taken this difference in semantics into 110867961f9dSMike Kravetz * account. Therefore, the meaning of chg is the same 110967961f9dSMike Kravetz * as in the shared case above. Code could easily be 111067961f9dSMike Kravetz * combined, but keeping it separate draws attention to 111167961f9dSMike Kravetz * subtle differences. 111267961f9dSMike Kravetz */ 111367961f9dSMike Kravetz if (chg) 111467961f9dSMike Kravetz return false; 111567961f9dSMike Kravetz else 1116559ec2f8SNicholas Krause return true; 111767961f9dSMike Kravetz } 1118a63884e9SJoonsoo Kim 1119559ec2f8SNicholas Krause return false; 1120a1e78772SMel Gorman } 1121a1e78772SMel Gorman 1122a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 11231da177e4SLinus Torvalds { 11241da177e4SLinus Torvalds int nid = page_to_nid(page); 11259487ca60SMike Kravetz 11269487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1127b65a4edaSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1128b65a4edaSMike Kravetz 11290edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 1130a5516438SAndi Kleen h->free_huge_pages++; 1131a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 11326c037149SMike Kravetz SetHPageFreed(page); 11331da177e4SLinus Torvalds } 11341da177e4SLinus Torvalds 113594310cbcSAnshuman Khandual static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) 1136bf50bab2SNaoya Horiguchi { 1137bf50bab2SNaoya Horiguchi struct page *page; 11381a08ae36SPavel Tatashin bool pin = !!(current->flags & PF_MEMALLOC_PIN); 1139bf50bab2SNaoya Horiguchi 11409487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1141bbe88753SJoonsoo Kim list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { 11426077c943SAlex Sierra if (pin && !is_longterm_pinnable_page(page)) 1143bbe88753SJoonsoo Kim continue; 1144bbe88753SJoonsoo Kim 11456664bfc8SWei Yang if (PageHWPoison(page)) 11466664bfc8SWei Yang continue; 1147bbe88753SJoonsoo Kim 11480edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 1149a9869b83SNaoya Horiguchi set_page_refcounted(page); 11506c037149SMike Kravetz ClearHPageFreed(page); 1151bf50bab2SNaoya Horiguchi h->free_huge_pages--; 1152bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 1153bf50bab2SNaoya Horiguchi return page; 1154bf50bab2SNaoya Horiguchi } 1155bf50bab2SNaoya Horiguchi 11566664bfc8SWei Yang return NULL; 11576664bfc8SWei Yang } 11586664bfc8SWei Yang 11593e59fcb0SMichal Hocko static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, 11603e59fcb0SMichal Hocko nodemask_t *nmask) 116194310cbcSAnshuman Khandual { 11623e59fcb0SMichal Hocko unsigned int cpuset_mems_cookie; 11633e59fcb0SMichal Hocko struct zonelist *zonelist; 11643e59fcb0SMichal Hocko struct zone *zone; 11653e59fcb0SMichal Hocko struct zoneref *z; 116698fa15f3SAnshuman Khandual int node = NUMA_NO_NODE; 11673e59fcb0SMichal Hocko 11683e59fcb0SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 11693e59fcb0SMichal Hocko 11703e59fcb0SMichal Hocko retry_cpuset: 11713e59fcb0SMichal Hocko cpuset_mems_cookie = read_mems_allowed_begin(); 11723e59fcb0SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 117394310cbcSAnshuman Khandual struct page *page; 117494310cbcSAnshuman Khandual 11753e59fcb0SMichal Hocko if (!cpuset_zone_allowed(zone, gfp_mask)) 11763e59fcb0SMichal Hocko continue; 11773e59fcb0SMichal Hocko /* 11783e59fcb0SMichal Hocko * no need to ask again on the same node. Pool is node rather than 11793e59fcb0SMichal Hocko * zone aware 11803e59fcb0SMichal Hocko */ 11813e59fcb0SMichal Hocko if (zone_to_nid(zone) == node) 11823e59fcb0SMichal Hocko continue; 11833e59fcb0SMichal Hocko node = zone_to_nid(zone); 118494310cbcSAnshuman Khandual 118594310cbcSAnshuman Khandual page = dequeue_huge_page_node_exact(h, node); 118694310cbcSAnshuman Khandual if (page) 118794310cbcSAnshuman Khandual return page; 118894310cbcSAnshuman Khandual } 11893e59fcb0SMichal Hocko if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 11903e59fcb0SMichal Hocko goto retry_cpuset; 11913e59fcb0SMichal Hocko 119294310cbcSAnshuman Khandual return NULL; 119394310cbcSAnshuman Khandual } 119494310cbcSAnshuman Khandual 11958346d69dSXin Hao static unsigned long available_huge_pages(struct hstate *h) 11968346d69dSXin Hao { 11978346d69dSXin Hao return h->free_huge_pages - h->resv_huge_pages; 11988346d69dSXin Hao } 11998346d69dSXin Hao 1200a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 1201a5516438SAndi Kleen struct vm_area_struct *vma, 1202af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 1203af0ed73eSJoonsoo Kim long chg) 12041da177e4SLinus Torvalds { 1205cfcaa66fSBen Widawsky struct page *page = NULL; 1206480eccf9SLee Schermerhorn struct mempolicy *mpol; 120704ec6264SVlastimil Babka gfp_t gfp_mask; 12083e59fcb0SMichal Hocko nodemask_t *nodemask; 120904ec6264SVlastimil Babka int nid; 12101da177e4SLinus Torvalds 1211a1e78772SMel Gorman /* 1212a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 1213a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 1214a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 1215a1e78772SMel Gorman */ 12168346d69dSXin Hao if (!vma_has_reserves(vma, chg) && !available_huge_pages(h)) 1217c0ff7453SMiao Xie goto err; 1218a1e78772SMel Gorman 121904f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 12208346d69dSXin Hao if (avoid_reserve && !available_huge_pages(h)) 12216eab04a8SJustin P. Mattock goto err; 122204f2cbe3SMel Gorman 122304ec6264SVlastimil Babka gfp_mask = htlb_alloc_mask(h); 122404ec6264SVlastimil Babka nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 1225cfcaa66fSBen Widawsky 1226cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 12273e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 1228cfcaa66fSBen Widawsky 1229cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 1230cfcaa66fSBen Widawsky nodemask = NULL; 1231cfcaa66fSBen Widawsky } 1232cfcaa66fSBen Widawsky 1233cfcaa66fSBen Widawsky if (!page) 1234cfcaa66fSBen Widawsky page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 1235cfcaa66fSBen Widawsky 12363e59fcb0SMichal Hocko if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { 1237d6995da3SMike Kravetz SetHPageRestoreReserve(page); 1238a63884e9SJoonsoo Kim h->resv_huge_pages--; 1239bf50bab2SNaoya Horiguchi } 1240cc9a6c87SMel Gorman 1241cc9a6c87SMel Gorman mpol_cond_put(mpol); 1242cc9a6c87SMel Gorman return page; 1243cc9a6c87SMel Gorman 1244c0ff7453SMiao Xie err: 1245cc9a6c87SMel Gorman return NULL; 12461da177e4SLinus Torvalds } 12471da177e4SLinus Torvalds 12481cac6f2cSLuiz Capitulino /* 12491cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 12501cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 12511cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 12521cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 12531cac6f2cSLuiz Capitulino * node for alloc or free. 12541cac6f2cSLuiz Capitulino */ 12551cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 12561cac6f2cSLuiz Capitulino { 12570edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 12581cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 12591cac6f2cSLuiz Capitulino 12601cac6f2cSLuiz Capitulino return nid; 12611cac6f2cSLuiz Capitulino } 12621cac6f2cSLuiz Capitulino 12631cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 12641cac6f2cSLuiz Capitulino { 12651cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 12661cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 12671cac6f2cSLuiz Capitulino return nid; 12681cac6f2cSLuiz Capitulino } 12691cac6f2cSLuiz Capitulino 12701cac6f2cSLuiz Capitulino /* 12711cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 12721cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 12731cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 12741cac6f2cSLuiz Capitulino * mask. 12751cac6f2cSLuiz Capitulino */ 12761cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 12771cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 12781cac6f2cSLuiz Capitulino { 12791cac6f2cSLuiz Capitulino int nid; 12801cac6f2cSLuiz Capitulino 12811cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 12821cac6f2cSLuiz Capitulino 12831cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 12841cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 12851cac6f2cSLuiz Capitulino 12861cac6f2cSLuiz Capitulino return nid; 12871cac6f2cSLuiz Capitulino } 12881cac6f2cSLuiz Capitulino 12891cac6f2cSLuiz Capitulino /* 129010c6ec49SMike Kravetz * helper for remove_pool_huge_page() - return the previously saved 12911cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 12921cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 12931cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 12941cac6f2cSLuiz Capitulino */ 12951cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 12961cac6f2cSLuiz Capitulino { 12971cac6f2cSLuiz Capitulino int nid; 12981cac6f2cSLuiz Capitulino 12991cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 13001cac6f2cSLuiz Capitulino 13011cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 13021cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 13031cac6f2cSLuiz Capitulino 13041cac6f2cSLuiz Capitulino return nid; 13051cac6f2cSLuiz Capitulino } 13061cac6f2cSLuiz Capitulino 13071cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 13081cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 13091cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 13101cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 13111cac6f2cSLuiz Capitulino nr_nodes--) 13121cac6f2cSLuiz Capitulino 13131cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 13141cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 13151cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 13161cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 13171cac6f2cSLuiz Capitulino nr_nodes--) 13181cac6f2cSLuiz Capitulino 13198531fc6fSMike Kravetz /* used to demote non-gigantic_huge pages as well */ 132034d9e35bSMike Kravetz static void __destroy_compound_gigantic_page(struct page *page, 132134d9e35bSMike Kravetz unsigned int order, bool demote) 1322944d9fecSLuiz Capitulino { 1323944d9fecSLuiz Capitulino int i; 1324944d9fecSLuiz Capitulino int nr_pages = 1 << order; 132514455eabSCheng Li struct page *p; 1326944d9fecSLuiz Capitulino 1327c8cc708aSGerald Schaefer atomic_set(compound_mapcount_ptr(page), 0); 132847e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 132947e29d32SJohn Hubbard 133014455eabSCheng Li for (i = 1; i < nr_pages; i++) { 133114455eabSCheng Li p = nth_page(page, i); 1332a01f4390SMike Kravetz p->mapping = NULL; 13331d798ca3SKirill A. Shutemov clear_compound_head(p); 133434d9e35bSMike Kravetz if (!demote) 1335944d9fecSLuiz Capitulino set_page_refcounted(p); 1336944d9fecSLuiz Capitulino } 1337944d9fecSLuiz Capitulino 1338944d9fecSLuiz Capitulino set_compound_order(page, 0); 13395232c63fSMatthew Wilcox (Oracle) #ifdef CONFIG_64BIT 1340ba9c1201SGerald Schaefer page[1].compound_nr = 0; 13415232c63fSMatthew Wilcox (Oracle) #endif 1342944d9fecSLuiz Capitulino __ClearPageHead(page); 1343944d9fecSLuiz Capitulino } 1344944d9fecSLuiz Capitulino 13458531fc6fSMike Kravetz static void destroy_compound_hugetlb_page_for_demote(struct page *page, 13468531fc6fSMike Kravetz unsigned int order) 13478531fc6fSMike Kravetz { 13488531fc6fSMike Kravetz __destroy_compound_gigantic_page(page, order, true); 13498531fc6fSMike Kravetz } 13508531fc6fSMike Kravetz 13518531fc6fSMike Kravetz #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 135234d9e35bSMike Kravetz static void destroy_compound_gigantic_page(struct page *page, 135334d9e35bSMike Kravetz unsigned int order) 135434d9e35bSMike Kravetz { 135534d9e35bSMike Kravetz __destroy_compound_gigantic_page(page, order, false); 135634d9e35bSMike Kravetz } 135734d9e35bSMike Kravetz 1358d00181b9SKirill A. Shutemov static void free_gigantic_page(struct page *page, unsigned int order) 1359944d9fecSLuiz Capitulino { 1360cf11e85fSRoman Gushchin /* 1361cf11e85fSRoman Gushchin * If the page isn't allocated using the cma allocator, 1362cf11e85fSRoman Gushchin * cma_release() returns false. 1363cf11e85fSRoman Gushchin */ 1364dbda8feaSBarry Song #ifdef CONFIG_CMA 1365dbda8feaSBarry Song if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order)) 1366cf11e85fSRoman Gushchin return; 1367dbda8feaSBarry Song #endif 1368cf11e85fSRoman Gushchin 1369944d9fecSLuiz Capitulino free_contig_range(page_to_pfn(page), 1 << order); 1370944d9fecSLuiz Capitulino } 1371944d9fecSLuiz Capitulino 13724eb0716eSAlexandre Ghiti #ifdef CONFIG_CONTIG_ALLOC 1373d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 1374d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) 1375944d9fecSLuiz Capitulino { 137604adbc3fSMiaohe Lin unsigned long nr_pages = pages_per_huge_page(h); 1377953f064aSLi Xinhai if (nid == NUMA_NO_NODE) 1378953f064aSLi Xinhai nid = numa_mem_id(); 1379944d9fecSLuiz Capitulino 1380dbda8feaSBarry Song #ifdef CONFIG_CMA 1381dbda8feaSBarry Song { 1382cf11e85fSRoman Gushchin struct page *page; 1383cf11e85fSRoman Gushchin int node; 1384cf11e85fSRoman Gushchin 1385953f064aSLi Xinhai if (hugetlb_cma[nid]) { 1386953f064aSLi Xinhai page = cma_alloc(hugetlb_cma[nid], nr_pages, 1387953f064aSLi Xinhai huge_page_order(h), true); 1388953f064aSLi Xinhai if (page) 1389953f064aSLi Xinhai return page; 1390953f064aSLi Xinhai } 1391953f064aSLi Xinhai 1392953f064aSLi Xinhai if (!(gfp_mask & __GFP_THISNODE)) { 1393cf11e85fSRoman Gushchin for_each_node_mask(node, *nodemask) { 1394953f064aSLi Xinhai if (node == nid || !hugetlb_cma[node]) 1395cf11e85fSRoman Gushchin continue; 1396cf11e85fSRoman Gushchin 1397cf11e85fSRoman Gushchin page = cma_alloc(hugetlb_cma[node], nr_pages, 1398cf11e85fSRoman Gushchin huge_page_order(h), true); 1399cf11e85fSRoman Gushchin if (page) 1400cf11e85fSRoman Gushchin return page; 1401cf11e85fSRoman Gushchin } 1402cf11e85fSRoman Gushchin } 1403953f064aSLi Xinhai } 1404dbda8feaSBarry Song #endif 1405cf11e85fSRoman Gushchin 14065e27a2dfSAnshuman Khandual return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); 1407944d9fecSLuiz Capitulino } 1408944d9fecSLuiz Capitulino 14094eb0716eSAlexandre Ghiti #else /* !CONFIG_CONTIG_ALLOC */ 14104eb0716eSAlexandre Ghiti static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 14114eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 14124eb0716eSAlexandre Ghiti { 14134eb0716eSAlexandre Ghiti return NULL; 14144eb0716eSAlexandre Ghiti } 14154eb0716eSAlexandre Ghiti #endif /* CONFIG_CONTIG_ALLOC */ 1416944d9fecSLuiz Capitulino 1417e1073d1eSAneesh Kumar K.V #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 1418d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 14194eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 14204eb0716eSAlexandre Ghiti { 14214eb0716eSAlexandre Ghiti return NULL; 14224eb0716eSAlexandre Ghiti } 1423d00181b9SKirill A. Shutemov static inline void free_gigantic_page(struct page *page, unsigned int order) { } 1424944d9fecSLuiz Capitulino static inline void destroy_compound_gigantic_page(struct page *page, 1425d00181b9SKirill A. Shutemov unsigned int order) { } 1426944d9fecSLuiz Capitulino #endif 1427944d9fecSLuiz Capitulino 14286eb4e88aSMike Kravetz /* 14296eb4e88aSMike Kravetz * Remove hugetlb page from lists, and update dtor so that page appears 143034d9e35bSMike Kravetz * as just a compound page. 143134d9e35bSMike Kravetz * 143234d9e35bSMike Kravetz * A reference is held on the page, except in the case of demote. 14336eb4e88aSMike Kravetz * 14346eb4e88aSMike Kravetz * Must be called with hugetlb lock held. 14356eb4e88aSMike Kravetz */ 143634d9e35bSMike Kravetz static void __remove_hugetlb_page(struct hstate *h, struct page *page, 143734d9e35bSMike Kravetz bool adjust_surplus, 143834d9e35bSMike Kravetz bool demote) 14396eb4e88aSMike Kravetz { 14406eb4e88aSMike Kravetz int nid = page_to_nid(page); 14416eb4e88aSMike Kravetz 14426eb4e88aSMike Kravetz VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 14436eb4e88aSMike Kravetz VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page); 14446eb4e88aSMike Kravetz 14459487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 14466eb4e88aSMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 14476eb4e88aSMike Kravetz return; 14486eb4e88aSMike Kravetz 14496eb4e88aSMike Kravetz list_del(&page->lru); 14506eb4e88aSMike Kravetz 14516eb4e88aSMike Kravetz if (HPageFreed(page)) { 14526eb4e88aSMike Kravetz h->free_huge_pages--; 14536eb4e88aSMike Kravetz h->free_huge_pages_node[nid]--; 14546eb4e88aSMike Kravetz } 14556eb4e88aSMike Kravetz if (adjust_surplus) { 14566eb4e88aSMike Kravetz h->surplus_huge_pages--; 14576eb4e88aSMike Kravetz h->surplus_huge_pages_node[nid]--; 14586eb4e88aSMike Kravetz } 14596eb4e88aSMike Kravetz 1460e32d20c0SMike Kravetz /* 1461e32d20c0SMike Kravetz * Very subtle 1462e32d20c0SMike Kravetz * 1463e32d20c0SMike Kravetz * For non-gigantic pages set the destructor to the normal compound 1464e32d20c0SMike Kravetz * page dtor. This is needed in case someone takes an additional 1465e32d20c0SMike Kravetz * temporary ref to the page, and freeing is delayed until they drop 1466e32d20c0SMike Kravetz * their reference. 1467e32d20c0SMike Kravetz * 1468e32d20c0SMike Kravetz * For gigantic pages set the destructor to the null dtor. This 1469e32d20c0SMike Kravetz * destructor will never be called. Before freeing the gigantic 1470e32d20c0SMike Kravetz * page destroy_compound_gigantic_page will turn the compound page 1471e32d20c0SMike Kravetz * into a simple group of pages. After this the destructor does not 1472e32d20c0SMike Kravetz * apply. 1473e32d20c0SMike Kravetz * 1474e32d20c0SMike Kravetz * This handles the case where more than one ref is held when and 1475e32d20c0SMike Kravetz * after update_and_free_page is called. 147634d9e35bSMike Kravetz * 147734d9e35bSMike Kravetz * In the case of demote we do not ref count the page as it will soon 147834d9e35bSMike Kravetz * be turned into a page of smaller size. 1479e32d20c0SMike Kravetz */ 148034d9e35bSMike Kravetz if (!demote) 14816eb4e88aSMike Kravetz set_page_refcounted(page); 1482e32d20c0SMike Kravetz if (hstate_is_gigantic(h)) 14836eb4e88aSMike Kravetz set_compound_page_dtor(page, NULL_COMPOUND_DTOR); 1484e32d20c0SMike Kravetz else 1485e32d20c0SMike Kravetz set_compound_page_dtor(page, COMPOUND_PAGE_DTOR); 14866eb4e88aSMike Kravetz 14876eb4e88aSMike Kravetz h->nr_huge_pages--; 14886eb4e88aSMike Kravetz h->nr_huge_pages_node[nid]--; 14896eb4e88aSMike Kravetz } 14906eb4e88aSMike Kravetz 149134d9e35bSMike Kravetz static void remove_hugetlb_page(struct hstate *h, struct page *page, 149234d9e35bSMike Kravetz bool adjust_surplus) 149334d9e35bSMike Kravetz { 149434d9e35bSMike Kravetz __remove_hugetlb_page(h, page, adjust_surplus, false); 149534d9e35bSMike Kravetz } 149634d9e35bSMike Kravetz 14978531fc6fSMike Kravetz static void remove_hugetlb_page_for_demote(struct hstate *h, struct page *page, 14988531fc6fSMike Kravetz bool adjust_surplus) 14998531fc6fSMike Kravetz { 15008531fc6fSMike Kravetz __remove_hugetlb_page(h, page, adjust_surplus, true); 15018531fc6fSMike Kravetz } 15028531fc6fSMike Kravetz 1503ad2fa371SMuchun Song static void add_hugetlb_page(struct hstate *h, struct page *page, 1504ad2fa371SMuchun Song bool adjust_surplus) 1505ad2fa371SMuchun Song { 1506ad2fa371SMuchun Song int zeroed; 1507ad2fa371SMuchun Song int nid = page_to_nid(page); 1508ad2fa371SMuchun Song 1509ad2fa371SMuchun Song VM_BUG_ON_PAGE(!HPageVmemmapOptimized(page), page); 1510ad2fa371SMuchun Song 1511ad2fa371SMuchun Song lockdep_assert_held(&hugetlb_lock); 1512ad2fa371SMuchun Song 1513ad2fa371SMuchun Song INIT_LIST_HEAD(&page->lru); 1514ad2fa371SMuchun Song h->nr_huge_pages++; 1515ad2fa371SMuchun Song h->nr_huge_pages_node[nid]++; 1516ad2fa371SMuchun Song 1517ad2fa371SMuchun Song if (adjust_surplus) { 1518ad2fa371SMuchun Song h->surplus_huge_pages++; 1519ad2fa371SMuchun Song h->surplus_huge_pages_node[nid]++; 1520ad2fa371SMuchun Song } 1521ad2fa371SMuchun Song 1522ad2fa371SMuchun Song set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1523ad2fa371SMuchun Song set_page_private(page, 0); 1524a9e1eab2SMiaohe Lin /* 1525a9e1eab2SMiaohe Lin * We have to set HPageVmemmapOptimized again as above 1526a9e1eab2SMiaohe Lin * set_page_private(page, 0) cleared it. 1527a9e1eab2SMiaohe Lin */ 1528ad2fa371SMuchun Song SetHPageVmemmapOptimized(page); 1529ad2fa371SMuchun Song 1530ad2fa371SMuchun Song /* 1531b65a4edaSMike Kravetz * This page is about to be managed by the hugetlb allocator and 1532b65a4edaSMike Kravetz * should have no users. Drop our reference, and check for others 1533b65a4edaSMike Kravetz * just in case. 1534ad2fa371SMuchun Song */ 1535ad2fa371SMuchun Song zeroed = put_page_testzero(page); 1536b65a4edaSMike Kravetz if (!zeroed) 1537b65a4edaSMike Kravetz /* 1538b65a4edaSMike Kravetz * It is VERY unlikely soneone else has taken a ref on 1539b65a4edaSMike Kravetz * the page. In this case, we simply return as the 1540b65a4edaSMike Kravetz * hugetlb destructor (free_huge_page) will be called 1541b65a4edaSMike Kravetz * when this other ref is dropped. 1542b65a4edaSMike Kravetz */ 1543b65a4edaSMike Kravetz return; 1544b65a4edaSMike Kravetz 1545ad2fa371SMuchun Song arch_clear_hugepage_flags(page); 1546ad2fa371SMuchun Song enqueue_huge_page(h, page); 1547ad2fa371SMuchun Song } 1548ad2fa371SMuchun Song 1549b65d4adbSMuchun Song static void __update_and_free_page(struct hstate *h, struct page *page) 15506af2acb6SAdam Litke { 15516af2acb6SAdam Litke int i; 155214455eabSCheng Li struct page *subpage; 1553a5516438SAndi Kleen 15544eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1555944d9fecSLuiz Capitulino return; 155618229df5SAndy Whitcroft 1557161df60eSNaoya Horiguchi /* 1558161df60eSNaoya Horiguchi * If we don't know which subpages are hwpoisoned, we can't free 1559161df60eSNaoya Horiguchi * the hugepage, so it's leaked intentionally. 1560161df60eSNaoya Horiguchi */ 1561161df60eSNaoya Horiguchi if (HPageRawHwpUnreliable(page)) 1562161df60eSNaoya Horiguchi return; 1563161df60eSNaoya Horiguchi 15646213834cSMuchun Song if (hugetlb_vmemmap_restore(h, page)) { 1565ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 1566ad2fa371SMuchun Song /* 1567ad2fa371SMuchun Song * If we cannot allocate vmemmap pages, just refuse to free the 1568ad2fa371SMuchun Song * page and put the page back on the hugetlb free list and treat 1569ad2fa371SMuchun Song * as a surplus page. 1570ad2fa371SMuchun Song */ 1571ad2fa371SMuchun Song add_hugetlb_page(h, page, true); 1572ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 1573ad2fa371SMuchun Song return; 1574ad2fa371SMuchun Song } 1575ad2fa371SMuchun Song 1576161df60eSNaoya Horiguchi /* 1577161df60eSNaoya Horiguchi * Move PageHWPoison flag from head page to the raw error pages, 1578161df60eSNaoya Horiguchi * which makes any healthy subpages reusable. 1579161df60eSNaoya Horiguchi */ 1580161df60eSNaoya Horiguchi if (unlikely(PageHWPoison(page))) 1581161df60eSNaoya Horiguchi hugetlb_clear_page_hwpoison(page); 1582161df60eSNaoya Horiguchi 158314455eabSCheng Li for (i = 0; i < pages_per_huge_page(h); i++) { 158414455eabSCheng Li subpage = nth_page(page, i); 1585dbfee5aeSMike Kravetz subpage->flags &= ~(1 << PG_locked | 1 << PG_error | 158632f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1587a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1588a7407a27SLuiz Capitulino 1 << PG_writeback); 15896af2acb6SAdam Litke } 1590a01f4390SMike Kravetz 1591a01f4390SMike Kravetz /* 1592a01f4390SMike Kravetz * Non-gigantic pages demoted from CMA allocated gigantic pages 1593a01f4390SMike Kravetz * need to be given back to CMA in free_gigantic_page. 1594a01f4390SMike Kravetz */ 1595a01f4390SMike Kravetz if (hstate_is_gigantic(h) || 1596a01f4390SMike Kravetz hugetlb_cma_page(page, huge_page_order(h))) { 1597944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 1598944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 1599944d9fecSLuiz Capitulino } else { 1600a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 16016af2acb6SAdam Litke } 1602944d9fecSLuiz Capitulino } 16036af2acb6SAdam Litke 1604b65d4adbSMuchun Song /* 1605b65d4adbSMuchun Song * As update_and_free_page() can be called under any context, so we cannot 1606b65d4adbSMuchun Song * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the 1607b65d4adbSMuchun Song * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate 1608b65d4adbSMuchun Song * the vmemmap pages. 1609b65d4adbSMuchun Song * 1610b65d4adbSMuchun Song * free_hpage_workfn() locklessly retrieves the linked list of pages to be 1611b65d4adbSMuchun Song * freed and frees them one-by-one. As the page->mapping pointer is going 1612b65d4adbSMuchun Song * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node 1613b65d4adbSMuchun Song * structure of a lockless linked list of huge pages to be freed. 1614b65d4adbSMuchun Song */ 1615b65d4adbSMuchun Song static LLIST_HEAD(hpage_freelist); 1616b65d4adbSMuchun Song 1617b65d4adbSMuchun Song static void free_hpage_workfn(struct work_struct *work) 1618b65d4adbSMuchun Song { 1619b65d4adbSMuchun Song struct llist_node *node; 1620b65d4adbSMuchun Song 1621b65d4adbSMuchun Song node = llist_del_all(&hpage_freelist); 1622b65d4adbSMuchun Song 1623b65d4adbSMuchun Song while (node) { 1624b65d4adbSMuchun Song struct page *page; 1625b65d4adbSMuchun Song struct hstate *h; 1626b65d4adbSMuchun Song 1627b65d4adbSMuchun Song page = container_of((struct address_space **)node, 1628b65d4adbSMuchun Song struct page, mapping); 1629b65d4adbSMuchun Song node = node->next; 1630b65d4adbSMuchun Song page->mapping = NULL; 1631b65d4adbSMuchun Song /* 1632b65d4adbSMuchun Song * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate() 1633b65d4adbSMuchun Song * is going to trigger because a previous call to 1634b65d4adbSMuchun Song * remove_hugetlb_page() will set_compound_page_dtor(page, 1635b65d4adbSMuchun Song * NULL_COMPOUND_DTOR), so do not use page_hstate() directly. 1636b65d4adbSMuchun Song */ 1637b65d4adbSMuchun Song h = size_to_hstate(page_size(page)); 1638b65d4adbSMuchun Song 1639b65d4adbSMuchun Song __update_and_free_page(h, page); 1640b65d4adbSMuchun Song 1641b65d4adbSMuchun Song cond_resched(); 1642b65d4adbSMuchun Song } 1643b65d4adbSMuchun Song } 1644b65d4adbSMuchun Song static DECLARE_WORK(free_hpage_work, free_hpage_workfn); 1645b65d4adbSMuchun Song 1646b65d4adbSMuchun Song static inline void flush_free_hpage_work(struct hstate *h) 1647b65d4adbSMuchun Song { 16486213834cSMuchun Song if (hugetlb_vmemmap_optimizable(h)) 1649b65d4adbSMuchun Song flush_work(&free_hpage_work); 1650b65d4adbSMuchun Song } 1651b65d4adbSMuchun Song 1652b65d4adbSMuchun Song static void update_and_free_page(struct hstate *h, struct page *page, 1653b65d4adbSMuchun Song bool atomic) 1654b65d4adbSMuchun Song { 1655ad2fa371SMuchun Song if (!HPageVmemmapOptimized(page) || !atomic) { 1656b65d4adbSMuchun Song __update_and_free_page(h, page); 1657b65d4adbSMuchun Song return; 1658b65d4adbSMuchun Song } 1659b65d4adbSMuchun Song 1660b65d4adbSMuchun Song /* 1661b65d4adbSMuchun Song * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages. 1662b65d4adbSMuchun Song * 1663b65d4adbSMuchun Song * Only call schedule_work() if hpage_freelist is previously 1664b65d4adbSMuchun Song * empty. Otherwise, schedule_work() had been called but the workfn 1665b65d4adbSMuchun Song * hasn't retrieved the list yet. 1666b65d4adbSMuchun Song */ 1667b65d4adbSMuchun Song if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist)) 1668b65d4adbSMuchun Song schedule_work(&free_hpage_work); 1669b65d4adbSMuchun Song } 1670b65d4adbSMuchun Song 167110c6ec49SMike Kravetz static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list) 167210c6ec49SMike Kravetz { 167310c6ec49SMike Kravetz struct page *page, *t_page; 167410c6ec49SMike Kravetz 167510c6ec49SMike Kravetz list_for_each_entry_safe(page, t_page, list, lru) { 1676b65d4adbSMuchun Song update_and_free_page(h, page, false); 167710c6ec49SMike Kravetz cond_resched(); 167810c6ec49SMike Kravetz } 167910c6ec49SMike Kravetz } 168010c6ec49SMike Kravetz 1681e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1682e5ff2159SAndi Kleen { 1683e5ff2159SAndi Kleen struct hstate *h; 1684e5ff2159SAndi Kleen 1685e5ff2159SAndi Kleen for_each_hstate(h) { 1686e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1687e5ff2159SAndi Kleen return h; 1688e5ff2159SAndi Kleen } 1689e5ff2159SAndi Kleen return NULL; 1690e5ff2159SAndi Kleen } 1691e5ff2159SAndi Kleen 1692db71ef79SMike Kravetz void free_huge_page(struct page *page) 169327a85ef1SDavid Gibson { 1694a5516438SAndi Kleen /* 1695a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1696a5516438SAndi Kleen * compound page destructor. 1697a5516438SAndi Kleen */ 1698e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 16997893d1d5SAdam Litke int nid = page_to_nid(page); 1700d6995da3SMike Kravetz struct hugepage_subpool *spool = hugetlb_page_subpool(page); 170107443a85SJoonsoo Kim bool restore_reserve; 1702db71ef79SMike Kravetz unsigned long flags; 170327a85ef1SDavid Gibson 1704b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1705b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_mapcount(page), page); 17068ace22bcSYongkai Wu 1707d6995da3SMike Kravetz hugetlb_set_page_subpool(page, NULL); 170878fbe906SDavid Hildenbrand if (PageAnon(page)) 170978fbe906SDavid Hildenbrand __ClearPageAnonExclusive(page); 17108ace22bcSYongkai Wu page->mapping = NULL; 1711d6995da3SMike Kravetz restore_reserve = HPageRestoreReserve(page); 1712d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 171327a85ef1SDavid Gibson 17141c5ecae3SMike Kravetz /* 1715d6995da3SMike Kravetz * If HPageRestoreReserve was set on page, page allocation consumed a 17160919e1b6SMike Kravetz * reservation. If the page was associated with a subpool, there 17170919e1b6SMike Kravetz * would have been a page reserved in the subpool before allocation 17180919e1b6SMike Kravetz * via hugepage_subpool_get_pages(). Since we are 'restoring' the 17196c26d310SMiaohe Lin * reservation, do not call hugepage_subpool_put_pages() as this will 17200919e1b6SMike Kravetz * remove the reserved page from the subpool. 17210919e1b6SMike Kravetz */ 17220919e1b6SMike Kravetz if (!restore_reserve) { 17230919e1b6SMike Kravetz /* 17240919e1b6SMike Kravetz * A return code of zero implies that the subpool will be 17250919e1b6SMike Kravetz * under its minimum size if the reservation is not restored 17260919e1b6SMike Kravetz * after page is free. Therefore, force restore_reserve 17270919e1b6SMike Kravetz * operation. 17281c5ecae3SMike Kravetz */ 17291c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 17301c5ecae3SMike Kravetz restore_reserve = true; 17310919e1b6SMike Kravetz } 17321c5ecae3SMike Kravetz 1733db71ef79SMike Kravetz spin_lock_irqsave(&hugetlb_lock, flags); 17348f251a3dSMike Kravetz ClearHPageMigratable(page); 17356d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 17366d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 173708cf9fafSMina Almasry hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 173808cf9fafSMina Almasry pages_per_huge_page(h), page); 173907443a85SJoonsoo Kim if (restore_reserve) 174007443a85SJoonsoo Kim h->resv_huge_pages++; 174107443a85SJoonsoo Kim 17429157c311SMike Kravetz if (HPageTemporary(page)) { 17436eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 1744db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1745b65d4adbSMuchun Song update_and_free_page(h, page, true); 1746ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 17470edaecfaSAneesh Kumar K.V /* remove the page from active list */ 17486eb4e88aSMike Kravetz remove_hugetlb_page(h, page, true); 1749db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1750b65d4adbSMuchun Song update_and_free_page(h, page, true); 17517893d1d5SAdam Litke } else { 17525d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1753a5516438SAndi Kleen enqueue_huge_page(h, page); 1754db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 175527a85ef1SDavid Gibson } 17561121828aSMike Kravetz } 175727a85ef1SDavid Gibson 1758d3d99fccSOscar Salvador /* 1759d3d99fccSOscar Salvador * Must be called with the hugetlb lock held 1760d3d99fccSOscar Salvador */ 1761d3d99fccSOscar Salvador static void __prep_account_new_huge_page(struct hstate *h, int nid) 1762d3d99fccSOscar Salvador { 1763d3d99fccSOscar Salvador lockdep_assert_held(&hugetlb_lock); 1764d3d99fccSOscar Salvador h->nr_huge_pages++; 1765d3d99fccSOscar Salvador h->nr_huge_pages_node[nid]++; 1766d3d99fccSOscar Salvador } 1767d3d99fccSOscar Salvador 1768f41f2ed4SMuchun Song static void __prep_new_huge_page(struct hstate *h, struct page *page) 1769b7ba30c6SAndi Kleen { 17706213834cSMuchun Song hugetlb_vmemmap_optimize(h, page); 17710edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1772f1e61557SKirill A. Shutemov set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1773ff546117SMike Kravetz hugetlb_set_page_subpool(page, NULL); 17749dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 17751adc4d41SMina Almasry set_hugetlb_cgroup_rsvd(page, NULL); 1776d3d99fccSOscar Salvador } 1777d3d99fccSOscar Salvador 1778d3d99fccSOscar Salvador static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1779d3d99fccSOscar Salvador { 1780f41f2ed4SMuchun Song __prep_new_huge_page(h, page); 1781db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 1782d3d99fccSOscar Salvador __prep_account_new_huge_page(h, nid); 1783db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1784b7ba30c6SAndi Kleen } 1785b7ba30c6SAndi Kleen 178634d9e35bSMike Kravetz static bool __prep_compound_gigantic_page(struct page *page, unsigned int order, 178734d9e35bSMike Kravetz bool demote) 178820a0307cSWu Fengguang { 17897118fc29SMike Kravetz int i, j; 179020a0307cSWu Fengguang int nr_pages = 1 << order; 179114455eabSCheng Li struct page *p; 179220a0307cSWu Fengguang 179320a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 179420a0307cSWu Fengguang set_compound_order(page, order); 1795de09d31dSKirill A. Shutemov __SetPageHead(page); 17962b21624fSMike Kravetz for (i = 0; i < nr_pages; i++) { 179714455eabSCheng Li p = nth_page(page, i); 179814455eabSCheng Li 1799ef5a22beSAndrea Arcangeli /* 1800ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1801ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1802ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 18037c8de358SEthon Paul * too. Otherwise drivers using get_user_pages() to access tail 1804ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1805ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1806ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1807ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1808ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1809ef5a22beSAndrea Arcangeli * after get_user_pages(). 1810ef5a22beSAndrea Arcangeli */ 1811ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 18127118fc29SMike Kravetz /* 18137118fc29SMike Kravetz * Subtle and very unlikely 18147118fc29SMike Kravetz * 18157118fc29SMike Kravetz * Gigantic 'page allocators' such as memblock or cma will 18167118fc29SMike Kravetz * return a set of pages with each page ref counted. We need 18177118fc29SMike Kravetz * to turn this set of pages into a compound page with tail 18187118fc29SMike Kravetz * page ref counts set to zero. Code such as speculative page 18197118fc29SMike Kravetz * cache adding could take a ref on a 'to be' tail page. 18207118fc29SMike Kravetz * We need to respect any increased ref count, and only set 18217118fc29SMike Kravetz * the ref count to zero if count is currently 1. If count 1822416d85edSMike Kravetz * is not 1, we return an error. An error return indicates 1823416d85edSMike Kravetz * the set of pages can not be converted to a gigantic page. 1824416d85edSMike Kravetz * The caller who allocated the pages should then discard the 1825416d85edSMike Kravetz * pages using the appropriate free interface. 182634d9e35bSMike Kravetz * 182734d9e35bSMike Kravetz * In the case of demote, the ref count will be zero. 18287118fc29SMike Kravetz */ 182934d9e35bSMike Kravetz if (!demote) { 18307118fc29SMike Kravetz if (!page_ref_freeze(p, 1)) { 1831416d85edSMike Kravetz pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n"); 18327118fc29SMike Kravetz goto out_error; 18337118fc29SMike Kravetz } 183434d9e35bSMike Kravetz } else { 183534d9e35bSMike Kravetz VM_BUG_ON_PAGE(page_count(p), p); 183634d9e35bSMike Kravetz } 18372b21624fSMike Kravetz if (i != 0) 18381d798ca3SKirill A. Shutemov set_compound_head(p, page); 183920a0307cSWu Fengguang } 1840b4330afbSMike Kravetz atomic_set(compound_mapcount_ptr(page), -1); 184147e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 18427118fc29SMike Kravetz return true; 18437118fc29SMike Kravetz 18447118fc29SMike Kravetz out_error: 18452b21624fSMike Kravetz /* undo page modifications made above */ 18462b21624fSMike Kravetz for (j = 0; j < i; j++) { 184714455eabSCheng Li p = nth_page(page, j); 18482b21624fSMike Kravetz if (j != 0) 18497118fc29SMike Kravetz clear_compound_head(p); 18507118fc29SMike Kravetz set_page_refcounted(p); 18517118fc29SMike Kravetz } 18527118fc29SMike Kravetz /* need to clear PG_reserved on remaining tail pages */ 185314455eabSCheng Li for (; j < nr_pages; j++) { 185414455eabSCheng Li p = nth_page(page, j); 18557118fc29SMike Kravetz __ClearPageReserved(p); 185614455eabSCheng Li } 18577118fc29SMike Kravetz set_compound_order(page, 0); 18585232c63fSMatthew Wilcox (Oracle) #ifdef CONFIG_64BIT 18597118fc29SMike Kravetz page[1].compound_nr = 0; 18605232c63fSMatthew Wilcox (Oracle) #endif 18617118fc29SMike Kravetz __ClearPageHead(page); 18627118fc29SMike Kravetz return false; 186320a0307cSWu Fengguang } 186420a0307cSWu Fengguang 186534d9e35bSMike Kravetz static bool prep_compound_gigantic_page(struct page *page, unsigned int order) 186634d9e35bSMike Kravetz { 186734d9e35bSMike Kravetz return __prep_compound_gigantic_page(page, order, false); 186834d9e35bSMike Kravetz } 186934d9e35bSMike Kravetz 18708531fc6fSMike Kravetz static bool prep_compound_gigantic_page_for_demote(struct page *page, 18718531fc6fSMike Kravetz unsigned int order) 18728531fc6fSMike Kravetz { 18738531fc6fSMike Kravetz return __prep_compound_gigantic_page(page, order, true); 18748531fc6fSMike Kravetz } 18758531fc6fSMike Kravetz 18767795912cSAndrew Morton /* 18777795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 18787795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 18797795912cSAndrew Morton * details. 18807795912cSAndrew Morton */ 188120a0307cSWu Fengguang int PageHuge(struct page *page) 188220a0307cSWu Fengguang { 188320a0307cSWu Fengguang if (!PageCompound(page)) 188420a0307cSWu Fengguang return 0; 188520a0307cSWu Fengguang 188620a0307cSWu Fengguang page = compound_head(page); 1887f1e61557SKirill A. Shutemov return page[1].compound_dtor == HUGETLB_PAGE_DTOR; 188820a0307cSWu Fengguang } 188943131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 189043131e14SNaoya Horiguchi 189127c73ae7SAndrea Arcangeli /* 189227c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 189327c73ae7SAndrea Arcangeli * normal or transparent huge pages. 189427c73ae7SAndrea Arcangeli */ 189527c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 189627c73ae7SAndrea Arcangeli { 189727c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 189827c73ae7SAndrea Arcangeli return 0; 189927c73ae7SAndrea Arcangeli 1900d4af73e3SVlastimil Babka return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; 190127c73ae7SAndrea Arcangeli } 19024e936eccSDavid Howells EXPORT_SYMBOL_GPL(PageHeadHuge); 190327c73ae7SAndrea Arcangeli 1904c0d0381aSMike Kravetz /* 1905c0d0381aSMike Kravetz * Find and lock address space (mapping) in write mode. 1906c0d0381aSMike Kravetz * 1907336bf30eSMike Kravetz * Upon entry, the page is locked which means that page_mapping() is 1908336bf30eSMike Kravetz * stable. Due to locking order, we can only trylock_write. If we can 1909336bf30eSMike Kravetz * not get the lock, simply return NULL to caller. 1910c0d0381aSMike Kravetz */ 1911c0d0381aSMike Kravetz struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) 1912c0d0381aSMike Kravetz { 1913336bf30eSMike Kravetz struct address_space *mapping = page_mapping(hpage); 1914c0d0381aSMike Kravetz 1915c0d0381aSMike Kravetz if (!mapping) 1916c0d0381aSMike Kravetz return mapping; 1917c0d0381aSMike Kravetz 1918c0d0381aSMike Kravetz if (i_mmap_trylock_write(mapping)) 1919c0d0381aSMike Kravetz return mapping; 1920c0d0381aSMike Kravetz 1921c0d0381aSMike Kravetz return NULL; 1922c0d0381aSMike Kravetz } 1923c0d0381aSMike Kravetz 1924fe19bd3dSHugh Dickins pgoff_t hugetlb_basepage_index(struct page *page) 192513d60f4bSZhang Yi { 192613d60f4bSZhang Yi struct page *page_head = compound_head(page); 192713d60f4bSZhang Yi pgoff_t index = page_index(page_head); 192813d60f4bSZhang Yi unsigned long compound_idx; 192913d60f4bSZhang Yi 193013d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 193113d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 193213d60f4bSZhang Yi else 193313d60f4bSZhang Yi compound_idx = page - page_head; 193413d60f4bSZhang Yi 193513d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 193613d60f4bSZhang Yi } 193713d60f4bSZhang Yi 19380c397daeSMichal Hocko static struct page *alloc_buddy_huge_page(struct hstate *h, 1939f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1940f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 19411da177e4SLinus Torvalds { 1942af0fb9dfSMichal Hocko int order = huge_page_order(h); 19431da177e4SLinus Torvalds struct page *page; 1944f60858f9SMike Kravetz bool alloc_try_hard = true; 19452b21624fSMike Kravetz bool retry = true; 1946f96efd58SJoe Jin 1947f60858f9SMike Kravetz /* 1948f60858f9SMike Kravetz * By default we always try hard to allocate the page with 1949f60858f9SMike Kravetz * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in 1950f60858f9SMike Kravetz * a loop (to adjust global huge page counts) and previous allocation 1951f60858f9SMike Kravetz * failed, do not continue to try hard on the same node. Use the 1952f60858f9SMike Kravetz * node_alloc_noretry bitmap to manage this state information. 1953f60858f9SMike Kravetz */ 1954f60858f9SMike Kravetz if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) 1955f60858f9SMike Kravetz alloc_try_hard = false; 1956f60858f9SMike Kravetz gfp_mask |= __GFP_COMP|__GFP_NOWARN; 1957f60858f9SMike Kravetz if (alloc_try_hard) 1958f60858f9SMike Kravetz gfp_mask |= __GFP_RETRY_MAYFAIL; 1959af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 1960af0fb9dfSMichal Hocko nid = numa_mem_id(); 19612b21624fSMike Kravetz retry: 196284172f4bSMatthew Wilcox (Oracle) page = __alloc_pages(gfp_mask, order, nid, nmask); 19632b21624fSMike Kravetz 19642b21624fSMike Kravetz /* Freeze head page */ 19652b21624fSMike Kravetz if (page && !page_ref_freeze(page, 1)) { 19662b21624fSMike Kravetz __free_pages(page, order); 19672b21624fSMike Kravetz if (retry) { /* retry once */ 19682b21624fSMike Kravetz retry = false; 19692b21624fSMike Kravetz goto retry; 19702b21624fSMike Kravetz } 19712b21624fSMike Kravetz /* WOW! twice in a row. */ 19722b21624fSMike Kravetz pr_warn("HugeTLB head page unexpected inflated ref count\n"); 19732b21624fSMike Kravetz page = NULL; 19742b21624fSMike Kravetz } 19752b21624fSMike Kravetz 1976af0fb9dfSMichal Hocko if (page) 1977af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC); 1978af0fb9dfSMichal Hocko else 1979af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 198063b4613cSNishanth Aravamudan 1981f60858f9SMike Kravetz /* 1982f60858f9SMike Kravetz * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this 1983f60858f9SMike Kravetz * indicates an overall state change. Clear bit so that we resume 1984f60858f9SMike Kravetz * normal 'try hard' allocations. 1985f60858f9SMike Kravetz */ 1986f60858f9SMike Kravetz if (node_alloc_noretry && page && !alloc_try_hard) 1987f60858f9SMike Kravetz node_clear(nid, *node_alloc_noretry); 1988f60858f9SMike Kravetz 1989f60858f9SMike Kravetz /* 1990f60858f9SMike Kravetz * If we tried hard to get a page but failed, set bit so that 1991f60858f9SMike Kravetz * subsequent attempts will not try as hard until there is an 1992f60858f9SMike Kravetz * overall state change. 1993f60858f9SMike Kravetz */ 1994f60858f9SMike Kravetz if (node_alloc_noretry && !page && alloc_try_hard) 1995f60858f9SMike Kravetz node_set(nid, *node_alloc_noretry); 1996f60858f9SMike Kravetz 199763b4613cSNishanth Aravamudan return page; 199863b4613cSNishanth Aravamudan } 199963b4613cSNishanth Aravamudan 2000af0fb9dfSMichal Hocko /* 20010c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 20020c397daeSMichal Hocko * should use this function to get new hugetlb pages 20032b21624fSMike Kravetz * 20042b21624fSMike Kravetz * Note that returned page is 'frozen': ref count of head page and all tail 20052b21624fSMike Kravetz * pages is zero. 20060c397daeSMichal Hocko */ 20070c397daeSMichal Hocko static struct page *alloc_fresh_huge_page(struct hstate *h, 2008f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 2009f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 20100c397daeSMichal Hocko { 20110c397daeSMichal Hocko struct page *page; 20127118fc29SMike Kravetz bool retry = false; 20130c397daeSMichal Hocko 20147118fc29SMike Kravetz retry: 20150c397daeSMichal Hocko if (hstate_is_gigantic(h)) 20160c397daeSMichal Hocko page = alloc_gigantic_page(h, gfp_mask, nid, nmask); 20170c397daeSMichal Hocko else 20180c397daeSMichal Hocko page = alloc_buddy_huge_page(h, gfp_mask, 2019f60858f9SMike Kravetz nid, nmask, node_alloc_noretry); 20200c397daeSMichal Hocko if (!page) 20210c397daeSMichal Hocko return NULL; 20220c397daeSMichal Hocko 20237118fc29SMike Kravetz if (hstate_is_gigantic(h)) { 20247118fc29SMike Kravetz if (!prep_compound_gigantic_page(page, huge_page_order(h))) { 20257118fc29SMike Kravetz /* 20267118fc29SMike Kravetz * Rare failure to convert pages to compound page. 20277118fc29SMike Kravetz * Free pages and try again - ONCE! 20287118fc29SMike Kravetz */ 20297118fc29SMike Kravetz free_gigantic_page(page, huge_page_order(h)); 20307118fc29SMike Kravetz if (!retry) { 20317118fc29SMike Kravetz retry = true; 20327118fc29SMike Kravetz goto retry; 20337118fc29SMike Kravetz } 20347118fc29SMike Kravetz return NULL; 20357118fc29SMike Kravetz } 20367118fc29SMike Kravetz } 20370c397daeSMichal Hocko prep_new_huge_page(h, page, page_to_nid(page)); 20380c397daeSMichal Hocko 20390c397daeSMichal Hocko return page; 20400c397daeSMichal Hocko } 20410c397daeSMichal Hocko 20420c397daeSMichal Hocko /* 2043af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 2044af0fb9dfSMichal Hocko * manner. 2045af0fb9dfSMichal Hocko */ 2046f60858f9SMike Kravetz static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 2047f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 2048b2261026SJoonsoo Kim { 2049b2261026SJoonsoo Kim struct page *page; 2050b2261026SJoonsoo Kim int nr_nodes, node; 2051af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2052b2261026SJoonsoo Kim 2053b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 2054f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, 2055f60858f9SMike Kravetz node_alloc_noretry); 2056af0fb9dfSMichal Hocko if (page) 2057b2261026SJoonsoo Kim break; 2058b2261026SJoonsoo Kim } 2059b2261026SJoonsoo Kim 2060af0fb9dfSMichal Hocko if (!page) 2061af0fb9dfSMichal Hocko return 0; 2062b2261026SJoonsoo Kim 20632b21624fSMike Kravetz free_huge_page(page); /* free it into the hugepage allocator */ 2064af0fb9dfSMichal Hocko 2065af0fb9dfSMichal Hocko return 1; 2066b2261026SJoonsoo Kim } 2067b2261026SJoonsoo Kim 2068e8c5c824SLee Schermerhorn /* 206910c6ec49SMike Kravetz * Remove huge page from pool from next node to free. Attempt to keep 207010c6ec49SMike Kravetz * persistent huge pages more or less balanced over allowed nodes. 207110c6ec49SMike Kravetz * This routine only 'removes' the hugetlb page. The caller must make 207210c6ec49SMike Kravetz * an additional call to free the page to low level allocators. 2073e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 2074e8c5c824SLee Schermerhorn */ 207510c6ec49SMike Kravetz static struct page *remove_pool_huge_page(struct hstate *h, 207610c6ec49SMike Kravetz nodemask_t *nodes_allowed, 20776ae11b27SLee Schermerhorn bool acct_surplus) 2078e8c5c824SLee Schermerhorn { 2079b2261026SJoonsoo Kim int nr_nodes, node; 208010c6ec49SMike Kravetz struct page *page = NULL; 2081e8c5c824SLee Schermerhorn 20829487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2083b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2084685f3457SLee Schermerhorn /* 2085685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 2086685f3457SLee Schermerhorn * nodes with surplus pages. 2087685f3457SLee Schermerhorn */ 2088b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 2089b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 209010c6ec49SMike Kravetz page = list_entry(h->hugepage_freelists[node].next, 2091e8c5c824SLee Schermerhorn struct page, lru); 20926eb4e88aSMike Kravetz remove_hugetlb_page(h, page, acct_surplus); 20939a76db09SLee Schermerhorn break; 2094e8c5c824SLee Schermerhorn } 2095b2261026SJoonsoo Kim } 2096e8c5c824SLee Schermerhorn 209710c6ec49SMike Kravetz return page; 2098e8c5c824SLee Schermerhorn } 2099e8c5c824SLee Schermerhorn 2100c8721bbbSNaoya Horiguchi /* 2101c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 2102faf53defSNaoya Horiguchi * nothing for in-use hugepages and non-hugepages. 2103faf53defSNaoya Horiguchi * This function returns values like below: 2104faf53defSNaoya Horiguchi * 2105ad2fa371SMuchun Song * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages 2106ad2fa371SMuchun Song * when the system is under memory pressure and the feature of 2107ad2fa371SMuchun Song * freeing unused vmemmap pages associated with each hugetlb page 2108ad2fa371SMuchun Song * is enabled. 2109faf53defSNaoya Horiguchi * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 2110faf53defSNaoya Horiguchi * (allocated or reserved.) 2111faf53defSNaoya Horiguchi * 0: successfully dissolved free hugepages or the page is not a 2112faf53defSNaoya Horiguchi * hugepage (considered as already dissolved) 2113c8721bbbSNaoya Horiguchi */ 2114c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 2115c8721bbbSNaoya Horiguchi { 21166bc9b564SNaoya Horiguchi int rc = -EBUSY; 2117082d5b6bSGerald Schaefer 21187ffddd49SMuchun Song retry: 2119faf53defSNaoya Horiguchi /* Not to disrupt normal path by vainly holding hugetlb_lock */ 2120faf53defSNaoya Horiguchi if (!PageHuge(page)) 2121faf53defSNaoya Horiguchi return 0; 2122faf53defSNaoya Horiguchi 2123db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2124faf53defSNaoya Horiguchi if (!PageHuge(page)) { 2125faf53defSNaoya Horiguchi rc = 0; 2126faf53defSNaoya Horiguchi goto out; 2127faf53defSNaoya Horiguchi } 2128faf53defSNaoya Horiguchi 2129faf53defSNaoya Horiguchi if (!page_count(page)) { 21302247bb33SGerald Schaefer struct page *head = compound_head(page); 21312247bb33SGerald Schaefer struct hstate *h = page_hstate(head); 21328346d69dSXin Hao if (!available_huge_pages(h)) 2133082d5b6bSGerald Schaefer goto out; 21347ffddd49SMuchun Song 21357ffddd49SMuchun Song /* 21367ffddd49SMuchun Song * We should make sure that the page is already on the free list 21377ffddd49SMuchun Song * when it is dissolved. 21387ffddd49SMuchun Song */ 21396c037149SMike Kravetz if (unlikely(!HPageFreed(head))) { 2140db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 21417ffddd49SMuchun Song cond_resched(); 21427ffddd49SMuchun Song 21437ffddd49SMuchun Song /* 21447ffddd49SMuchun Song * Theoretically, we should return -EBUSY when we 21457ffddd49SMuchun Song * encounter this race. In fact, we have a chance 21467ffddd49SMuchun Song * to successfully dissolve the page if we do a 21477ffddd49SMuchun Song * retry. Because the race window is quite small. 21487ffddd49SMuchun Song * If we seize this opportunity, it is an optimization 21497ffddd49SMuchun Song * for increasing the success rate of dissolving page. 21507ffddd49SMuchun Song */ 21517ffddd49SMuchun Song goto retry; 21527ffddd49SMuchun Song } 21537ffddd49SMuchun Song 2154ad2fa371SMuchun Song remove_hugetlb_page(h, head, false); 2155ad2fa371SMuchun Song h->max_huge_pages--; 2156ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2157ad2fa371SMuchun Song 2158c3114a84SAnshuman Khandual /* 2159ad2fa371SMuchun Song * Normally update_and_free_page will allocate required vmemmmap 2160ad2fa371SMuchun Song * before freeing the page. update_and_free_page will fail to 2161ad2fa371SMuchun Song * free the page if it can not allocate required vmemmap. We 2162ad2fa371SMuchun Song * need to adjust max_huge_pages if the page is not freed. 2163ad2fa371SMuchun Song * Attempt to allocate vmemmmap here so that we can take 2164ad2fa371SMuchun Song * appropriate action on failure. 2165ad2fa371SMuchun Song */ 21666213834cSMuchun Song rc = hugetlb_vmemmap_restore(h, head); 2167ad2fa371SMuchun Song if (!rc) { 2168b65d4adbSMuchun Song update_and_free_page(h, head, false); 2169ad2fa371SMuchun Song } else { 2170ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 2171ad2fa371SMuchun Song add_hugetlb_page(h, head, false); 2172ad2fa371SMuchun Song h->max_huge_pages++; 2173ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2174ad2fa371SMuchun Song } 2175ad2fa371SMuchun Song 2176ad2fa371SMuchun Song return rc; 2177c8721bbbSNaoya Horiguchi } 2178082d5b6bSGerald Schaefer out: 2179db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2180082d5b6bSGerald Schaefer return rc; 2181c8721bbbSNaoya Horiguchi } 2182c8721bbbSNaoya Horiguchi 2183c8721bbbSNaoya Horiguchi /* 2184c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 2185c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 21862247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 21872247bb33SGerald Schaefer * part of it lies within the given range. 2188082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 2189082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 2190c8721bbbSNaoya Horiguchi */ 2191082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 2192c8721bbbSNaoya Horiguchi { 2193c8721bbbSNaoya Horiguchi unsigned long pfn; 2194eb03aa00SGerald Schaefer struct page *page; 2195082d5b6bSGerald Schaefer int rc = 0; 2196dc2628f3SMuchun Song unsigned int order; 2197dc2628f3SMuchun Song struct hstate *h; 2198c8721bbbSNaoya Horiguchi 2199d0177639SLi Zhong if (!hugepages_supported()) 2200082d5b6bSGerald Schaefer return rc; 2201d0177639SLi Zhong 2202dc2628f3SMuchun Song order = huge_page_order(&default_hstate); 2203dc2628f3SMuchun Song for_each_hstate(h) 2204dc2628f3SMuchun Song order = min(order, huge_page_order(h)); 2205dc2628f3SMuchun Song 2206dc2628f3SMuchun Song for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) { 2207eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 2208eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 2209eb03aa00SGerald Schaefer if (rc) 2210082d5b6bSGerald Schaefer break; 2211eb03aa00SGerald Schaefer } 2212082d5b6bSGerald Schaefer 2213082d5b6bSGerald Schaefer return rc; 2214c8721bbbSNaoya Horiguchi } 2215c8721bbbSNaoya Horiguchi 2216ab5ac90aSMichal Hocko /* 2217ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 2218ab5ac90aSMichal Hocko */ 22190c397daeSMichal Hocko static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, 22202b21624fSMike Kravetz int nid, nodemask_t *nmask) 22217893d1d5SAdam Litke { 22229980d744SMichal Hocko struct page *page = NULL; 22237893d1d5SAdam Litke 2224bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2225aa888a74SAndi Kleen return NULL; 2226aa888a74SAndi Kleen 2227db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 22289980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 22299980d744SMichal Hocko goto out_unlock; 2230db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2231d1c3fb1fSNishanth Aravamudan 2232f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 22339980d744SMichal Hocko if (!page) 22340c397daeSMichal Hocko return NULL; 2235d1c3fb1fSNishanth Aravamudan 2236db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 22379980d744SMichal Hocko /* 22389980d744SMichal Hocko * We could have raced with the pool size change. 22399980d744SMichal Hocko * Double check that and simply deallocate the new page 22409980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 22419980d744SMichal Hocko * temporary page to workaround the nasty free_huge_page 22429980d744SMichal Hocko * codeflow 22439980d744SMichal Hocko */ 22449980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 22459157c311SMike Kravetz SetHPageTemporary(page); 2246db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 22472b21624fSMike Kravetz free_huge_page(page); 22482bf753e6SKai Shen return NULL; 2249b65a4edaSMike Kravetz } 2250b65a4edaSMike Kravetz 22519980d744SMichal Hocko h->surplus_huge_pages++; 22524704dea3SMichal Hocko h->surplus_huge_pages_node[page_to_nid(page)]++; 22539980d744SMichal Hocko 22549980d744SMichal Hocko out_unlock: 2255db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 22567893d1d5SAdam Litke 22577893d1d5SAdam Litke return page; 22587893d1d5SAdam Litke } 22597893d1d5SAdam Litke 2260bbe88753SJoonsoo Kim static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, 2261ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 2262ab5ac90aSMichal Hocko { 2263ab5ac90aSMichal Hocko struct page *page; 2264ab5ac90aSMichal Hocko 2265ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 2266ab5ac90aSMichal Hocko return NULL; 2267ab5ac90aSMichal Hocko 2268f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 2269ab5ac90aSMichal Hocko if (!page) 2270ab5ac90aSMichal Hocko return NULL; 2271ab5ac90aSMichal Hocko 22722b21624fSMike Kravetz /* fresh huge pages are frozen */ 22732b21624fSMike Kravetz set_page_refcounted(page); 22742b21624fSMike Kravetz 2275ab5ac90aSMichal Hocko /* 2276ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 2277ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 2278ab5ac90aSMichal Hocko */ 22799157c311SMike Kravetz SetHPageTemporary(page); 2280ab5ac90aSMichal Hocko 2281ab5ac90aSMichal Hocko return page; 2282ab5ac90aSMichal Hocko } 2283ab5ac90aSMichal Hocko 2284e4e574b7SAdam Litke /* 2285099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 2286099730d6SDave Hansen */ 2287e0ec90eeSDave Hansen static 22880c397daeSMichal Hocko struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, 2289099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 2290099730d6SDave Hansen { 2291cfcaa66fSBen Widawsky struct page *page = NULL; 2292aaf14e40SMichal Hocko struct mempolicy *mpol; 2293aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 2294aaf14e40SMichal Hocko int nid; 2295aaf14e40SMichal Hocko nodemask_t *nodemask; 2296aaf14e40SMichal Hocko 2297aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 2298cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 2299cfcaa66fSBen Widawsky gfp_t gfp = gfp_mask | __GFP_NOWARN; 2300cfcaa66fSBen Widawsky 2301cfcaa66fSBen Widawsky gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL); 23022b21624fSMike Kravetz page = alloc_surplus_huge_page(h, gfp, nid, nodemask); 2303cfcaa66fSBen Widawsky 2304cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 2305cfcaa66fSBen Widawsky nodemask = NULL; 2306cfcaa66fSBen Widawsky } 2307cfcaa66fSBen Widawsky 2308cfcaa66fSBen Widawsky if (!page) 23092b21624fSMike Kravetz page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); 2310aaf14e40SMichal Hocko mpol_cond_put(mpol); 2311aaf14e40SMichal Hocko return page; 2312099730d6SDave Hansen } 2313099730d6SDave Hansen 2314ab5ac90aSMichal Hocko /* page migration callback function */ 23153e59fcb0SMichal Hocko struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, 2316d92bbc27SJoonsoo Kim nodemask_t *nmask, gfp_t gfp_mask) 23174db9b2efSMichal Hocko { 2318db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 23198346d69dSXin Hao if (available_huge_pages(h)) { 23203e59fcb0SMichal Hocko struct page *page; 23213e59fcb0SMichal Hocko 23223e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); 23233e59fcb0SMichal Hocko if (page) { 2324db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 23253e59fcb0SMichal Hocko return page; 23264db9b2efSMichal Hocko } 23274db9b2efSMichal Hocko } 2328db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 23294db9b2efSMichal Hocko 23300c397daeSMichal Hocko return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); 23314db9b2efSMichal Hocko } 23324db9b2efSMichal Hocko 2333ebd63723SMichal Hocko /* mempolicy aware migration callback */ 2334389c8178SMichal Hocko struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, 2335389c8178SMichal Hocko unsigned long address) 2336ebd63723SMichal Hocko { 2337ebd63723SMichal Hocko struct mempolicy *mpol; 2338ebd63723SMichal Hocko nodemask_t *nodemask; 2339ebd63723SMichal Hocko struct page *page; 2340ebd63723SMichal Hocko gfp_t gfp_mask; 2341ebd63723SMichal Hocko int node; 2342ebd63723SMichal Hocko 2343ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 2344ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 2345d92bbc27SJoonsoo Kim page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); 2346ebd63723SMichal Hocko mpol_cond_put(mpol); 2347ebd63723SMichal Hocko 2348ebd63723SMichal Hocko return page; 2349ebd63723SMichal Hocko } 2350ebd63723SMichal Hocko 2351bf50bab2SNaoya Horiguchi /* 235225985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 2353e4e574b7SAdam Litke * of size 'delta'. 2354e4e574b7SAdam Litke */ 23550a4f3d1bSLiu Xiang static int gather_surplus_pages(struct hstate *h, long delta) 23561b2a1e7bSJules Irenge __must_hold(&hugetlb_lock) 2357e4e574b7SAdam Litke { 235834665341SMiaohe Lin LIST_HEAD(surplus_list); 2359e4e574b7SAdam Litke struct page *page, *tmp; 23600a4f3d1bSLiu Xiang int ret; 23610a4f3d1bSLiu Xiang long i; 23620a4f3d1bSLiu Xiang long needed, allocated; 236328073b02SHillf Danton bool alloc_ok = true; 2364e4e574b7SAdam Litke 23659487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2366a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 2367ac09b3a1SAdam Litke if (needed <= 0) { 2368a5516438SAndi Kleen h->resv_huge_pages += delta; 2369e4e574b7SAdam Litke return 0; 2370ac09b3a1SAdam Litke } 2371e4e574b7SAdam Litke 2372e4e574b7SAdam Litke allocated = 0; 2373e4e574b7SAdam Litke 2374e4e574b7SAdam Litke ret = -ENOMEM; 2375e4e574b7SAdam Litke retry: 2376db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2377e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 23780c397daeSMichal Hocko page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), 23792b21624fSMike Kravetz NUMA_NO_NODE, NULL); 238028073b02SHillf Danton if (!page) { 238128073b02SHillf Danton alloc_ok = false; 238228073b02SHillf Danton break; 238328073b02SHillf Danton } 2384e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 238569ed779aSDavid Rientjes cond_resched(); 2386e4e574b7SAdam Litke } 238728073b02SHillf Danton allocated += i; 2388e4e574b7SAdam Litke 2389e4e574b7SAdam Litke /* 2390e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 2391e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 2392e4e574b7SAdam Litke */ 2393db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2394a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 2395a5516438SAndi Kleen (h->free_huge_pages + allocated); 239628073b02SHillf Danton if (needed > 0) { 239728073b02SHillf Danton if (alloc_ok) 2398e4e574b7SAdam Litke goto retry; 239928073b02SHillf Danton /* 240028073b02SHillf Danton * We were not able to allocate enough pages to 240128073b02SHillf Danton * satisfy the entire reservation so we free what 240228073b02SHillf Danton * we've allocated so far. 240328073b02SHillf Danton */ 240428073b02SHillf Danton goto free; 240528073b02SHillf Danton } 2406e4e574b7SAdam Litke /* 2407e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 240825985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 2409e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 2410ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 2411ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 2412ac09b3a1SAdam Litke * before they are reserved. 2413e4e574b7SAdam Litke */ 2414e4e574b7SAdam Litke needed += allocated; 2415a5516438SAndi Kleen h->resv_huge_pages += delta; 2416e4e574b7SAdam Litke ret = 0; 2417a9869b83SNaoya Horiguchi 241819fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 241919fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 242019fc3f0aSAdam Litke if ((--needed) < 0) 242119fc3f0aSAdam Litke break; 2422b65a4edaSMike Kravetz /* Add the page to the hugetlb allocator */ 2423a5516438SAndi Kleen enqueue_huge_page(h, page); 242419fc3f0aSAdam Litke } 242528073b02SHillf Danton free: 2426db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 242719fc3f0aSAdam Litke 2428b65a4edaSMike Kravetz /* 2429b65a4edaSMike Kravetz * Free unnecessary surplus pages to the buddy allocator. 2430b65a4edaSMike Kravetz * Pages have no ref count, call free_huge_page directly. 2431b65a4edaSMike Kravetz */ 2432c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 2433b65a4edaSMike Kravetz free_huge_page(page); 2434db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2435e4e574b7SAdam Litke 2436e4e574b7SAdam Litke return ret; 2437e4e574b7SAdam Litke } 2438e4e574b7SAdam Litke 2439e4e574b7SAdam Litke /* 2440e5bbc8a6SMike Kravetz * This routine has two main purposes: 2441e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 2442e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 2443e5bbc8a6SMike Kravetz * to the associated reservation map. 2444e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 2445e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 2446e4e574b7SAdam Litke */ 2447a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 2448a5516438SAndi Kleen unsigned long unused_resv_pages) 2449e4e574b7SAdam Litke { 2450e4e574b7SAdam Litke unsigned long nr_pages; 245110c6ec49SMike Kravetz struct page *page; 245210c6ec49SMike Kravetz LIST_HEAD(page_list); 245310c6ec49SMike Kravetz 24549487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 245510c6ec49SMike Kravetz /* Uncommit the reservation */ 245610c6ec49SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 2457e4e574b7SAdam Litke 2458c0531714SNaoya Horiguchi if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 2459e5bbc8a6SMike Kravetz goto out; 2460aa888a74SAndi Kleen 2461e5bbc8a6SMike Kravetz /* 2462e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 2463e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 2464e5bbc8a6SMike Kravetz */ 2465a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 2466e4e574b7SAdam Litke 2467685f3457SLee Schermerhorn /* 2468685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 24699b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 24709b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 24719b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 247210c6ec49SMike Kravetz * remove_pool_huge_page() will balance the freed pages across the 24739b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 2474685f3457SLee Schermerhorn */ 2475685f3457SLee Schermerhorn while (nr_pages--) { 247610c6ec49SMike Kravetz page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1); 247710c6ec49SMike Kravetz if (!page) 2478e5bbc8a6SMike Kravetz goto out; 247910c6ec49SMike Kravetz 248010c6ec49SMike Kravetz list_add(&page->lru, &page_list); 2481e4e574b7SAdam Litke } 2482e5bbc8a6SMike Kravetz 2483e5bbc8a6SMike Kravetz out: 2484db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 248510c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 2486db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2487e4e574b7SAdam Litke } 2488e4e574b7SAdam Litke 24895e911373SMike Kravetz 2490c37f9fb1SAndy Whitcroft /* 2491feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 24925e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 2493cf3ad20bSMike Kravetz * 2494cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 2495cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 2496cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 2497cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 2498cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 2499feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 2500feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 2501feba16e2SMike Kravetz * is called in such cases. 2502cf3ad20bSMike Kravetz * 2503cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 2504cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 2505cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 2506cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 2507cf3ad20bSMike Kravetz * take appropriate action. 250896b96a96SMike Kravetz * 250996b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 251096b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 251196b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 251296b96a96SMike Kravetz * reservation exists. 2513846be085SMike Kravetz * 2514846be085SMike Kravetz * vma_del_reservation is used in error paths where an entry in the reserve 2515846be085SMike Kravetz * map was created during huge page allocation and must be removed. It is to 2516846be085SMike Kravetz * be called after calling vma_needs_reservation to determine if a reservation 2517846be085SMike Kravetz * exists. 2518c37f9fb1SAndy Whitcroft */ 25195e911373SMike Kravetz enum vma_resv_mode { 25205e911373SMike Kravetz VMA_NEEDS_RESV, 25215e911373SMike Kravetz VMA_COMMIT_RESV, 2522feba16e2SMike Kravetz VMA_END_RESV, 252396b96a96SMike Kravetz VMA_ADD_RESV, 2524846be085SMike Kravetz VMA_DEL_RESV, 25255e911373SMike Kravetz }; 2526cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 2527cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 25285e911373SMike Kravetz enum vma_resv_mode mode) 2529c37f9fb1SAndy Whitcroft { 25304e35f483SJoonsoo Kim struct resv_map *resv; 25314e35f483SJoonsoo Kim pgoff_t idx; 2532cf3ad20bSMike Kravetz long ret; 25330db9d74eSMina Almasry long dummy_out_regions_needed; 2534c37f9fb1SAndy Whitcroft 25354e35f483SJoonsoo Kim resv = vma_resv_map(vma); 25364e35f483SJoonsoo Kim if (!resv) 2537c37f9fb1SAndy Whitcroft return 1; 2538c37f9fb1SAndy Whitcroft 25394e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 25405e911373SMike Kravetz switch (mode) { 25415e911373SMike Kravetz case VMA_NEEDS_RESV: 25420db9d74eSMina Almasry ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); 25430db9d74eSMina Almasry /* We assume that vma_reservation_* routines always operate on 25440db9d74eSMina Almasry * 1 page, and that adding to resv map a 1 page entry can only 25450db9d74eSMina Almasry * ever require 1 region. 25460db9d74eSMina Almasry */ 25470db9d74eSMina Almasry VM_BUG_ON(dummy_out_regions_needed != 1); 25485e911373SMike Kravetz break; 25495e911373SMike Kravetz case VMA_COMMIT_RESV: 2550075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 25510db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 25520db9d74eSMina Almasry VM_BUG_ON(ret < 0); 25535e911373SMike Kravetz break; 2554feba16e2SMike Kravetz case VMA_END_RESV: 25550db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 25565e911373SMike Kravetz ret = 0; 25575e911373SMike Kravetz break; 255896b96a96SMike Kravetz case VMA_ADD_RESV: 25590db9d74eSMina Almasry if (vma->vm_flags & VM_MAYSHARE) { 2560075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 25610db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 25620db9d74eSMina Almasry VM_BUG_ON(ret < 0); 25630db9d74eSMina Almasry } else { 25640db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 256596b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 256696b96a96SMike Kravetz } 256796b96a96SMike Kravetz break; 2568846be085SMike Kravetz case VMA_DEL_RESV: 2569846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 2570846be085SMike Kravetz region_abort(resv, idx, idx + 1, 1); 2571846be085SMike Kravetz ret = region_del(resv, idx, idx + 1); 2572846be085SMike Kravetz } else { 2573846be085SMike Kravetz ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 2574846be085SMike Kravetz /* region_add calls of range 1 should never fail. */ 2575846be085SMike Kravetz VM_BUG_ON(ret < 0); 2576846be085SMike Kravetz } 2577846be085SMike Kravetz break; 25785e911373SMike Kravetz default: 25795e911373SMike Kravetz BUG(); 25805e911373SMike Kravetz } 258184afd99bSAndy Whitcroft 2582846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV) 2583cf3ad20bSMike Kravetz return ret; 258467961f9dSMike Kravetz /* 2585bf3d12b9SMiaohe Lin * We know private mapping must have HPAGE_RESV_OWNER set. 2586bf3d12b9SMiaohe Lin * 258767961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 258867961f9dSMike Kravetz * However, a file associated with mapping could have been 258967961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 259067961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 259167961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 259267961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 259367961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 259467961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 259567961f9dSMike Kravetz * reservation has already been consumed. As a result, the 259667961f9dSMike Kravetz * return value of this routine is the opposite of the 259767961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 259867961f9dSMike Kravetz */ 2599bf3d12b9SMiaohe Lin if (ret > 0) 260067961f9dSMike Kravetz return 0; 2601bf3d12b9SMiaohe Lin if (ret == 0) 260267961f9dSMike Kravetz return 1; 2603bf3d12b9SMiaohe Lin return ret; 260484afd99bSAndy Whitcroft } 2605cf3ad20bSMike Kravetz 2606cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 2607a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 2608c37f9fb1SAndy Whitcroft { 26095e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 2610cf3ad20bSMike Kravetz } 2611c37f9fb1SAndy Whitcroft 2612cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 2613cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 2614cf3ad20bSMike Kravetz { 26155e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 26165e911373SMike Kravetz } 26175e911373SMike Kravetz 2618feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 26195e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 26205e911373SMike Kravetz { 2621feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 2622c37f9fb1SAndy Whitcroft } 2623c37f9fb1SAndy Whitcroft 262496b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 262596b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 262696b96a96SMike Kravetz { 262796b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 262896b96a96SMike Kravetz } 262996b96a96SMike Kravetz 2630846be085SMike Kravetz static long vma_del_reservation(struct hstate *h, 2631846be085SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 263296b96a96SMike Kravetz { 2633846be085SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV); 2634846be085SMike Kravetz } 2635846be085SMike Kravetz 2636846be085SMike Kravetz /* 2637846be085SMike Kravetz * This routine is called to restore reservation information on error paths. 2638846be085SMike Kravetz * It should ONLY be called for pages allocated via alloc_huge_page(), and 2639846be085SMike Kravetz * the hugetlb mutex should remain held when calling this routine. 2640846be085SMike Kravetz * 2641846be085SMike Kravetz * It handles two specific cases: 2642846be085SMike Kravetz * 1) A reservation was in place and the page consumed the reservation. 2643846be085SMike Kravetz * HPageRestoreReserve is set in the page. 2644846be085SMike Kravetz * 2) No reservation was in place for the page, so HPageRestoreReserve is 2645846be085SMike Kravetz * not set. However, alloc_huge_page always updates the reserve map. 2646846be085SMike Kravetz * 2647846be085SMike Kravetz * In case 1, free_huge_page later in the error path will increment the 2648846be085SMike Kravetz * global reserve count. But, free_huge_page does not have enough context 2649846be085SMike Kravetz * to adjust the reservation map. This case deals primarily with private 2650846be085SMike Kravetz * mappings. Adjust the reserve map here to be consistent with global 2651846be085SMike Kravetz * reserve count adjustments to be made by free_huge_page. Make sure the 2652846be085SMike Kravetz * reserve map indicates there is a reservation present. 2653846be085SMike Kravetz * 2654846be085SMike Kravetz * In case 2, simply undo reserve map modifications done by alloc_huge_page. 2655846be085SMike Kravetz */ 2656846be085SMike Kravetz void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma, 2657846be085SMike Kravetz unsigned long address, struct page *page) 2658846be085SMike Kravetz { 265996b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 266096b96a96SMike Kravetz 2661846be085SMike Kravetz if (HPageRestoreReserve(page)) { 2662846be085SMike Kravetz if (unlikely(rc < 0)) 266396b96a96SMike Kravetz /* 266496b96a96SMike Kravetz * Rare out of memory condition in reserve map 2665d6995da3SMike Kravetz * manipulation. Clear HPageRestoreReserve so that 266696b96a96SMike Kravetz * global reserve count will not be incremented 266796b96a96SMike Kravetz * by free_huge_page. This will make it appear 266896b96a96SMike Kravetz * as though the reservation for this page was 266996b96a96SMike Kravetz * consumed. This may prevent the task from 267096b96a96SMike Kravetz * faulting in the page at a later time. This 267196b96a96SMike Kravetz * is better than inconsistent global huge page 267296b96a96SMike Kravetz * accounting of reserve counts. 267396b96a96SMike Kravetz */ 2674d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 2675846be085SMike Kravetz else if (rc) 2676846be085SMike Kravetz (void)vma_add_reservation(h, vma, address); 2677846be085SMike Kravetz else 2678846be085SMike Kravetz vma_end_reservation(h, vma, address); 2679846be085SMike Kravetz } else { 2680846be085SMike Kravetz if (!rc) { 268196b96a96SMike Kravetz /* 2682846be085SMike Kravetz * This indicates there is an entry in the reserve map 2683c7b1850dSMike Kravetz * not added by alloc_huge_page. We know it was added 2684846be085SMike Kravetz * before the alloc_huge_page call, otherwise 2685846be085SMike Kravetz * HPageRestoreReserve would be set on the page. 2686846be085SMike Kravetz * Remove the entry so that a subsequent allocation 2687846be085SMike Kravetz * does not consume a reservation. 268896b96a96SMike Kravetz */ 2689846be085SMike Kravetz rc = vma_del_reservation(h, vma, address); 2690846be085SMike Kravetz if (rc < 0) 2691846be085SMike Kravetz /* 2692846be085SMike Kravetz * VERY rare out of memory condition. Since 2693846be085SMike Kravetz * we can not delete the entry, set 2694846be085SMike Kravetz * HPageRestoreReserve so that the reserve 2695846be085SMike Kravetz * count will be incremented when the page 2696846be085SMike Kravetz * is freed. This reserve will be consumed 2697846be085SMike Kravetz * on a subsequent allocation. 2698846be085SMike Kravetz */ 2699846be085SMike Kravetz SetHPageRestoreReserve(page); 2700846be085SMike Kravetz } else if (rc < 0) { 2701846be085SMike Kravetz /* 2702846be085SMike Kravetz * Rare out of memory condition from 2703846be085SMike Kravetz * vma_needs_reservation call. Memory allocation is 2704846be085SMike Kravetz * only attempted if a new entry is needed. Therefore, 2705846be085SMike Kravetz * this implies there is not an entry in the 2706846be085SMike Kravetz * reserve map. 2707846be085SMike Kravetz * 2708846be085SMike Kravetz * For shared mappings, no entry in the map indicates 2709846be085SMike Kravetz * no reservation. We are done. 2710846be085SMike Kravetz */ 2711846be085SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 2712846be085SMike Kravetz /* 2713846be085SMike Kravetz * For private mappings, no entry indicates 2714846be085SMike Kravetz * a reservation is present. Since we can 2715846be085SMike Kravetz * not add an entry, set SetHPageRestoreReserve 2716846be085SMike Kravetz * on the page so reserve count will be 2717846be085SMike Kravetz * incremented when freed. This reserve will 2718846be085SMike Kravetz * be consumed on a subsequent allocation. 2719846be085SMike Kravetz */ 2720846be085SMike Kravetz SetHPageRestoreReserve(page); 272196b96a96SMike Kravetz } else 2722846be085SMike Kravetz /* 2723846be085SMike Kravetz * No reservation present, do nothing 2724846be085SMike Kravetz */ 272596b96a96SMike Kravetz vma_end_reservation(h, vma, address); 272696b96a96SMike Kravetz } 272796b96a96SMike Kravetz } 272896b96a96SMike Kravetz 2729369fa227SOscar Salvador /* 2730369fa227SOscar Salvador * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one 2731369fa227SOscar Salvador * @h: struct hstate old page belongs to 2732369fa227SOscar Salvador * @old_page: Old page to dissolve 2733ae37c7ffSOscar Salvador * @list: List to isolate the page in case we need to 2734369fa227SOscar Salvador * Returns 0 on success, otherwise negated error. 2735369fa227SOscar Salvador */ 2736ae37c7ffSOscar Salvador static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page, 2737ae37c7ffSOscar Salvador struct list_head *list) 2738369fa227SOscar Salvador { 2739369fa227SOscar Salvador gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2740369fa227SOscar Salvador int nid = page_to_nid(old_page); 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 */ 2751369fa227SOscar Salvador new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL); 2752369fa227SOscar Salvador if (!new_page) 2753369fa227SOscar Salvador return -ENOMEM; 2754f41f2ed4SMuchun Song __prep_new_huge_page(h, new_page); 2755369fa227SOscar Salvador 2756369fa227SOscar Salvador retry: 2757369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2758369fa227SOscar Salvador if (!PageHuge(old_page)) { 2759369fa227SOscar Salvador /* 2760369fa227SOscar Salvador * Freed from under us. Drop new_page too. 2761369fa227SOscar Salvador */ 2762369fa227SOscar Salvador goto free_new; 2763369fa227SOscar Salvador } else if (page_count(old_page)) { 2764369fa227SOscar Salvador /* 2765ae37c7ffSOscar Salvador * Someone has grabbed the page, try to isolate it here. 2766ae37c7ffSOscar Salvador * Fail with -EBUSY if not possible. 2767369fa227SOscar Salvador */ 2768ae37c7ffSOscar Salvador spin_unlock_irq(&hugetlb_lock); 27697ce82f4cSMiaohe Lin ret = isolate_hugetlb(old_page, list); 2770ae37c7ffSOscar Salvador spin_lock_irq(&hugetlb_lock); 2771369fa227SOscar Salvador goto free_new; 2772369fa227SOscar Salvador } else if (!HPageFreed(old_page)) { 2773369fa227SOscar Salvador /* 2774369fa227SOscar Salvador * Page's refcount is 0 but it has not been enqueued in the 2775369fa227SOscar Salvador * freelist yet. Race window is small, so we can succeed here if 2776369fa227SOscar Salvador * we retry. 2777369fa227SOscar Salvador */ 2778369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2779369fa227SOscar Salvador cond_resched(); 2780369fa227SOscar Salvador goto retry; 2781369fa227SOscar Salvador } else { 2782369fa227SOscar Salvador /* 2783369fa227SOscar Salvador * Ok, old_page is still a genuine free hugepage. Remove it from 2784369fa227SOscar Salvador * the freelist and decrease the counters. These will be 2785369fa227SOscar Salvador * incremented again when calling __prep_account_new_huge_page() 2786369fa227SOscar Salvador * and enqueue_huge_page() for new_page. The counters will remain 2787369fa227SOscar Salvador * stable since this happens under the lock. 2788369fa227SOscar Salvador */ 2789369fa227SOscar Salvador remove_hugetlb_page(h, old_page, false); 2790369fa227SOscar Salvador 2791369fa227SOscar Salvador /* 2792b65a4edaSMike Kravetz * Ref count on new page is already zero as it was dropped 2793b65a4edaSMike Kravetz * earlier. It can be directly added to the pool free list. 2794369fa227SOscar Salvador */ 2795369fa227SOscar Salvador __prep_account_new_huge_page(h, nid); 2796369fa227SOscar Salvador enqueue_huge_page(h, new_page); 2797369fa227SOscar Salvador 2798369fa227SOscar Salvador /* 2799369fa227SOscar Salvador * Pages have been replaced, we can safely free the old one. 2800369fa227SOscar Salvador */ 2801369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2802b65d4adbSMuchun Song update_and_free_page(h, old_page, false); 2803369fa227SOscar Salvador } 2804369fa227SOscar Salvador 2805369fa227SOscar Salvador return ret; 2806369fa227SOscar Salvador 2807369fa227SOscar Salvador free_new: 2808369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2809b65a4edaSMike Kravetz /* Page has a zero ref count, but needs a ref to be freed */ 2810b65a4edaSMike Kravetz set_page_refcounted(new_page); 2811b65d4adbSMuchun Song update_and_free_page(h, new_page, false); 2812369fa227SOscar Salvador 2813369fa227SOscar Salvador return ret; 2814369fa227SOscar Salvador } 2815369fa227SOscar Salvador 2816ae37c7ffSOscar Salvador int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list) 2817369fa227SOscar Salvador { 2818369fa227SOscar Salvador struct hstate *h; 2819369fa227SOscar Salvador struct page *head; 2820ae37c7ffSOscar Salvador int ret = -EBUSY; 2821369fa227SOscar Salvador 2822369fa227SOscar Salvador /* 2823369fa227SOscar Salvador * The page might have been dissolved from under our feet, so make sure 2824369fa227SOscar Salvador * to carefully check the state under the lock. 2825369fa227SOscar Salvador * Return success when racing as if we dissolved the page ourselves. 2826369fa227SOscar Salvador */ 2827369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2828369fa227SOscar Salvador if (PageHuge(page)) { 2829369fa227SOscar Salvador head = compound_head(page); 2830369fa227SOscar Salvador h = page_hstate(head); 2831369fa227SOscar Salvador } else { 2832369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2833369fa227SOscar Salvador return 0; 2834369fa227SOscar Salvador } 2835369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2836369fa227SOscar Salvador 2837369fa227SOscar Salvador /* 2838369fa227SOscar Salvador * Fence off gigantic pages as there is a cyclic dependency between 2839369fa227SOscar Salvador * alloc_contig_range and them. Return -ENOMEM as this has the effect 2840369fa227SOscar Salvador * of bailing out right away without further retrying. 2841369fa227SOscar Salvador */ 2842369fa227SOscar Salvador if (hstate_is_gigantic(h)) 2843369fa227SOscar Salvador return -ENOMEM; 2844369fa227SOscar Salvador 28457ce82f4cSMiaohe Lin if (page_count(head) && !isolate_hugetlb(head, list)) 2846ae37c7ffSOscar Salvador ret = 0; 2847ae37c7ffSOscar Salvador else if (!page_count(head)) 2848ae37c7ffSOscar Salvador ret = alloc_and_dissolve_huge_page(h, head, list); 2849ae37c7ffSOscar Salvador 2850ae37c7ffSOscar Salvador return ret; 2851369fa227SOscar Salvador } 2852369fa227SOscar Salvador 285370c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 285404f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 2855348ea204SAdam Litke { 285690481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 2857a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2858348ea204SAdam Litke struct page *page; 2859d85f69b0SMike Kravetz long map_chg, map_commit; 2860d85f69b0SMike Kravetz long gbl_chg; 28616d76dcf4SAneesh Kumar K.V int ret, idx; 28626d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 286308cf9fafSMina Almasry bool deferred_reserve; 28642fc39cecSAdam Litke 28656d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 2866a1e78772SMel Gorman /* 2867d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 2868d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 2869d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 2870a1e78772SMel Gorman */ 2871d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 2872d85f69b0SMike Kravetz if (map_chg < 0) 287376dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 2874d85f69b0SMike Kravetz 2875d85f69b0SMike Kravetz /* 2876d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 2877d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 2878d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 2879d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 2880d85f69b0SMike Kravetz * checked against any subpool limit. 2881d85f69b0SMike Kravetz */ 2882d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 2883d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 2884d85f69b0SMike Kravetz if (gbl_chg < 0) { 2885feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 288676dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 28875e911373SMike Kravetz } 288890d8b7e6SAdam Litke 2889d85f69b0SMike Kravetz /* 2890d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 2891d85f69b0SMike Kravetz * map, there could be reservations associated with the 2892d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 2893d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 2894d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 2895d85f69b0SMike Kravetz * the subpool reservations. 2896d85f69b0SMike Kravetz */ 2897d85f69b0SMike Kravetz if (avoid_reserve) 2898d85f69b0SMike Kravetz gbl_chg = 1; 2899d85f69b0SMike Kravetz } 2900d85f69b0SMike Kravetz 290108cf9fafSMina Almasry /* If this allocation is not consuming a reservation, charge it now. 290208cf9fafSMina Almasry */ 29036501fe5fSMiaohe Lin deferred_reserve = map_chg || avoid_reserve; 290408cf9fafSMina Almasry if (deferred_reserve) { 290508cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 290608cf9fafSMina Almasry idx, pages_per_huge_page(h), &h_cg); 29078f34af6fSJianyu Zhan if (ret) 29088f34af6fSJianyu Zhan goto out_subpool_put; 290908cf9fafSMina Almasry } 291008cf9fafSMina Almasry 291108cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 291208cf9fafSMina Almasry if (ret) 291308cf9fafSMina Almasry goto out_uncharge_cgroup_reservation; 29148f34af6fSJianyu Zhan 2915db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2916d85f69b0SMike Kravetz /* 2917d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 2918d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 2919d85f69b0SMike Kravetz * a reservation exists for the allocation. 2920d85f69b0SMike Kravetz */ 2921d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 292281a6fcaeSJoonsoo Kim if (!page) { 2923db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 29240c397daeSMichal Hocko page = alloc_buddy_huge_page_with_mpol(h, vma, addr); 29258f34af6fSJianyu Zhan if (!page) 29268f34af6fSJianyu Zhan goto out_uncharge_cgroup; 2927a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 2928d6995da3SMike Kravetz SetHPageRestoreReserve(page); 2929a88c7695SNaoya Horiguchi h->resv_huge_pages--; 2930a88c7695SNaoya Horiguchi } 2931db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 293215a8d68eSWei Yang list_add(&page->lru, &h->hugepage_activelist); 29332b21624fSMike Kravetz set_page_refcounted(page); 293481a6fcaeSJoonsoo Kim /* Fall through */ 2935a1e78772SMel Gorman } 293681a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 293708cf9fafSMina Almasry /* If allocation is not consuming a reservation, also store the 293808cf9fafSMina Almasry * hugetlb_cgroup pointer on the page. 293908cf9fafSMina Almasry */ 294008cf9fafSMina Almasry if (deferred_reserve) { 294108cf9fafSMina Almasry hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), 294208cf9fafSMina Almasry h_cg, page); 294308cf9fafSMina Almasry } 294408cf9fafSMina Almasry 2945db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2946a1e78772SMel Gorman 2947d6995da3SMike Kravetz hugetlb_set_page_subpool(page, spool); 2948a1e78772SMel Gorman 2949d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 2950d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 295133039678SMike Kravetz /* 295233039678SMike Kravetz * The page was added to the reservation map between 295333039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 295433039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 295533039678SMike Kravetz * Adjust for the subpool count incremented above AND 295633039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 295733039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 295833039678SMike Kravetz * no longer applies. 295933039678SMike Kravetz */ 296033039678SMike Kravetz long rsv_adjust; 296133039678SMike Kravetz 296233039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 296333039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 296479aa925bSMike Kravetz if (deferred_reserve) 296579aa925bSMike Kravetz hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 296679aa925bSMike Kravetz pages_per_huge_page(h), page); 296733039678SMike Kravetz } 29687893d1d5SAdam Litke return page; 29698f34af6fSJianyu Zhan 29708f34af6fSJianyu Zhan out_uncharge_cgroup: 29718f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 297208cf9fafSMina Almasry out_uncharge_cgroup_reservation: 297308cf9fafSMina Almasry if (deferred_reserve) 297408cf9fafSMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), 297508cf9fafSMina Almasry h_cg); 29768f34af6fSJianyu Zhan out_subpool_put: 2977d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 29788f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 2979feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 29808f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 2981b45b5bd6SDavid Gibson } 2982b45b5bd6SDavid Gibson 2983b5389086SZhenguo Yao int alloc_bootmem_huge_page(struct hstate *h, int nid) 2984e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 2985b5389086SZhenguo Yao int __alloc_bootmem_huge_page(struct hstate *h, int nid) 2986aa888a74SAndi Kleen { 2987b5389086SZhenguo Yao struct huge_bootmem_page *m = NULL; /* initialize for clang */ 2988b2261026SJoonsoo Kim int nr_nodes, node; 2989aa888a74SAndi Kleen 2990b5389086SZhenguo Yao /* do node specific alloc */ 2991b5389086SZhenguo Yao if (nid != NUMA_NO_NODE) { 2992b5389086SZhenguo Yao m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h), 2993b5389086SZhenguo Yao 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid); 2994b5389086SZhenguo Yao if (!m) 2995b5389086SZhenguo Yao return 0; 2996b5389086SZhenguo Yao goto found; 2997b5389086SZhenguo Yao } 2998b5389086SZhenguo Yao /* allocate from next node when distributing huge pages */ 2999b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 3000b5389086SZhenguo Yao m = memblock_alloc_try_nid_raw( 30018b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 300297ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 3003aa888a74SAndi Kleen /* 3004aa888a74SAndi Kleen * Use the beginning of the huge page to store the 3005aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 3006aa888a74SAndi Kleen * puts them into the mem_map). 3007aa888a74SAndi Kleen */ 3008b5389086SZhenguo Yao if (!m) 3009b5389086SZhenguo Yao return 0; 3010aa888a74SAndi Kleen goto found; 3011aa888a74SAndi Kleen } 3012aa888a74SAndi Kleen 3013aa888a74SAndi Kleen found: 3014aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 3015330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 3016aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 3017aa888a74SAndi Kleen m->hstate = h; 3018aa888a74SAndi Kleen return 1; 3019aa888a74SAndi Kleen } 3020aa888a74SAndi Kleen 302148b8d744SMike Kravetz /* 302248b8d744SMike Kravetz * Put bootmem huge pages into the standard lists after mem_map is up. 302348b8d744SMike Kravetz * Note: This only applies to gigantic (order > MAX_ORDER) pages. 302448b8d744SMike Kravetz */ 3025aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 3026aa888a74SAndi Kleen { 3027aa888a74SAndi Kleen struct huge_bootmem_page *m; 3028aa888a74SAndi Kleen 3029aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 303040d18ebfSMike Kravetz struct page *page = virt_to_page(m); 3031aa888a74SAndi Kleen struct hstate *h = m->hstate; 3032ee8f248dSBecky Bruce 303348b8d744SMike Kravetz VM_BUG_ON(!hstate_is_gigantic(h)); 3034aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 30357118fc29SMike Kravetz if (prep_compound_gigantic_page(page, huge_page_order(h))) { 3036ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 3037aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 30382b21624fSMike Kravetz free_huge_page(page); /* add to the hugepage allocator */ 30397118fc29SMike Kravetz } else { 3040416d85edSMike Kravetz /* VERY unlikely inflated ref count on a tail page */ 30417118fc29SMike Kravetz free_gigantic_page(page, huge_page_order(h)); 30427118fc29SMike Kravetz } 3043af0fb9dfSMichal Hocko 3044b0320c7bSRafael Aquini /* 304548b8d744SMike Kravetz * We need to restore the 'stolen' pages to totalram_pages 304648b8d744SMike Kravetz * in order to fix confusing memory reports from free(1) and 304748b8d744SMike Kravetz * other side-effects, like CommitLimit going negative. 3048b0320c7bSRafael Aquini */ 3049c78a7f36SMiaohe Lin adjust_managed_page_count(page, pages_per_huge_page(h)); 3050520495feSCannon Matthews cond_resched(); 3051aa888a74SAndi Kleen } 3052aa888a74SAndi Kleen } 3053b5389086SZhenguo Yao static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid) 3054b5389086SZhenguo Yao { 3055b5389086SZhenguo Yao unsigned long i; 3056b5389086SZhenguo Yao char buf[32]; 3057b5389086SZhenguo Yao 3058b5389086SZhenguo Yao for (i = 0; i < h->max_huge_pages_node[nid]; ++i) { 3059b5389086SZhenguo Yao if (hstate_is_gigantic(h)) { 3060b5389086SZhenguo Yao if (!alloc_bootmem_huge_page(h, nid)) 3061b5389086SZhenguo Yao break; 3062b5389086SZhenguo Yao } else { 3063b5389086SZhenguo Yao struct page *page; 3064b5389086SZhenguo Yao gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 3065b5389086SZhenguo Yao 3066b5389086SZhenguo Yao page = alloc_fresh_huge_page(h, gfp_mask, nid, 3067b5389086SZhenguo Yao &node_states[N_MEMORY], NULL); 3068b5389086SZhenguo Yao if (!page) 3069b5389086SZhenguo Yao break; 30702b21624fSMike Kravetz free_huge_page(page); /* free it into the hugepage allocator */ 3071b5389086SZhenguo Yao } 3072b5389086SZhenguo Yao cond_resched(); 3073b5389086SZhenguo Yao } 3074b5389086SZhenguo Yao if (i == h->max_huge_pages_node[nid]) 3075b5389086SZhenguo Yao return; 3076b5389086SZhenguo Yao 3077b5389086SZhenguo Yao string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3078b5389086SZhenguo Yao pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n", 3079b5389086SZhenguo Yao h->max_huge_pages_node[nid], buf, nid, i); 3080b5389086SZhenguo Yao h->max_huge_pages -= (h->max_huge_pages_node[nid] - i); 3081b5389086SZhenguo Yao h->max_huge_pages_node[nid] = i; 3082b5389086SZhenguo Yao } 3083aa888a74SAndi Kleen 30848faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 30851da177e4SLinus Torvalds { 30861da177e4SLinus Torvalds unsigned long i; 3087f60858f9SMike Kravetz nodemask_t *node_alloc_noretry; 3088b5389086SZhenguo Yao bool node_specific_alloc = false; 3089f60858f9SMike Kravetz 3090b5389086SZhenguo Yao /* skip gigantic hugepages allocation if hugetlb_cma enabled */ 3091b5389086SZhenguo Yao if (hstate_is_gigantic(h) && hugetlb_cma_size) { 3092b5389086SZhenguo Yao pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); 3093b5389086SZhenguo Yao return; 3094b5389086SZhenguo Yao } 3095b5389086SZhenguo Yao 3096b5389086SZhenguo Yao /* do node specific alloc */ 30970a7a0f6fSPeng Liu for_each_online_node(i) { 3098b5389086SZhenguo Yao if (h->max_huge_pages_node[i] > 0) { 3099b5389086SZhenguo Yao hugetlb_hstate_alloc_pages_onenode(h, i); 3100b5389086SZhenguo Yao node_specific_alloc = true; 3101b5389086SZhenguo Yao } 3102b5389086SZhenguo Yao } 3103b5389086SZhenguo Yao 3104b5389086SZhenguo Yao if (node_specific_alloc) 3105b5389086SZhenguo Yao return; 3106b5389086SZhenguo Yao 3107b5389086SZhenguo Yao /* below will do all node balanced alloc */ 3108f60858f9SMike Kravetz if (!hstate_is_gigantic(h)) { 3109f60858f9SMike Kravetz /* 3110f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3111f60858f9SMike Kravetz * Ignore errors as lower level routines can deal with 3112f60858f9SMike Kravetz * node_alloc_noretry == NULL. If this kmalloc fails at boot 3113f60858f9SMike Kravetz * time, we are likely in bigger trouble. 3114f60858f9SMike Kravetz */ 3115f60858f9SMike Kravetz node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), 3116f60858f9SMike Kravetz GFP_KERNEL); 3117f60858f9SMike Kravetz } else { 3118f60858f9SMike Kravetz /* allocations done at boot time */ 3119f60858f9SMike Kravetz node_alloc_noretry = NULL; 3120f60858f9SMike Kravetz } 3121f60858f9SMike Kravetz 3122f60858f9SMike Kravetz /* bit mask controlling how hard we retry per-node allocations */ 3123f60858f9SMike Kravetz if (node_alloc_noretry) 3124f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 31251da177e4SLinus Torvalds 3126e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 3127bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 3128b5389086SZhenguo Yao if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE)) 3129aa888a74SAndi Kleen break; 31300c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 3131f60858f9SMike Kravetz &node_states[N_MEMORY], 3132f60858f9SMike Kravetz node_alloc_noretry)) 31331da177e4SLinus Torvalds break; 313469ed779aSDavid Rientjes cond_resched(); 31351da177e4SLinus Torvalds } 3136d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 3137d715cf80SLiam R. Howlett char buf[32]; 3138d715cf80SLiam R. Howlett 3139c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3140d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 3141d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 31428faa8b07SAndi Kleen h->max_huge_pages = i; 3143e5ff2159SAndi Kleen } 3144f60858f9SMike Kravetz kfree(node_alloc_noretry); 3145d715cf80SLiam R. Howlett } 3146e5ff2159SAndi Kleen 3147e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 3148e5ff2159SAndi Kleen { 314979dfc695SMike Kravetz struct hstate *h, *h2; 3150e5ff2159SAndi Kleen 3151e5ff2159SAndi Kleen for_each_hstate(h) { 31528faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 3153bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 31548faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 315579dfc695SMike Kravetz 315679dfc695SMike Kravetz /* 315779dfc695SMike Kravetz * Set demote order for each hstate. Note that 315879dfc695SMike Kravetz * h->demote_order is initially 0. 315979dfc695SMike Kravetz * - We can not demote gigantic pages if runtime freeing 316079dfc695SMike Kravetz * is not supported, so skip this. 3161a01f4390SMike Kravetz * - If CMA allocation is possible, we can not demote 3162a01f4390SMike Kravetz * HUGETLB_PAGE_ORDER or smaller size pages. 316379dfc695SMike Kravetz */ 316479dfc695SMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 316579dfc695SMike Kravetz continue; 3166a01f4390SMike Kravetz if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER) 3167a01f4390SMike Kravetz continue; 316879dfc695SMike Kravetz for_each_hstate(h2) { 316979dfc695SMike Kravetz if (h2 == h) 317079dfc695SMike Kravetz continue; 317179dfc695SMike Kravetz if (h2->order < h->order && 317279dfc695SMike Kravetz h2->order > h->demote_order) 317379dfc695SMike Kravetz h->demote_order = h2->order; 317479dfc695SMike Kravetz } 3175e5ff2159SAndi Kleen } 3176e5ff2159SAndi Kleen } 3177e5ff2159SAndi Kleen 3178e5ff2159SAndi Kleen static void __init report_hugepages(void) 3179e5ff2159SAndi Kleen { 3180e5ff2159SAndi Kleen struct hstate *h; 3181e5ff2159SAndi Kleen 3182e5ff2159SAndi Kleen for_each_hstate(h) { 31834abd32dbSAndi Kleen char buf[32]; 3184c6247f72SMatthew Wilcox 3185c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 31866213834cSMuchun Song pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n", 3187c6247f72SMatthew Wilcox buf, h->free_huge_pages); 31886213834cSMuchun Song pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n", 31896213834cSMuchun Song hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf); 3190e5ff2159SAndi Kleen } 3191e5ff2159SAndi Kleen } 3192e5ff2159SAndi Kleen 31931da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 31946ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 31956ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 31961da177e4SLinus Torvalds { 31974415cc8dSChristoph Lameter int i; 31981121828aSMike Kravetz LIST_HEAD(page_list); 31994415cc8dSChristoph Lameter 32009487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 3201bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3202aa888a74SAndi Kleen return; 3203aa888a74SAndi Kleen 32041121828aSMike Kravetz /* 32051121828aSMike Kravetz * Collect pages to be freed on a list, and free after dropping lock 32061121828aSMike Kravetz */ 32076ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 320810c6ec49SMike Kravetz struct page *page, *next; 3209a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 3210a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 3211a5516438SAndi Kleen if (count >= h->nr_huge_pages) 32121121828aSMike Kravetz goto out; 32131da177e4SLinus Torvalds if (PageHighMem(page)) 32141da177e4SLinus Torvalds continue; 32156eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 32161121828aSMike Kravetz list_add(&page->lru, &page_list); 32171121828aSMike Kravetz } 32181121828aSMike Kravetz } 32191121828aSMike Kravetz 32201121828aSMike Kravetz out: 3221db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 322210c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3223db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 32241da177e4SLinus Torvalds } 32251da177e4SLinus Torvalds #else 32266ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 32276ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 32281da177e4SLinus Torvalds { 32291da177e4SLinus Torvalds } 32301da177e4SLinus Torvalds #endif 32311da177e4SLinus Torvalds 323220a0307cSWu Fengguang /* 323320a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 323420a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 323520a0307cSWu Fengguang * Returns 1 if an adjustment was made. 323620a0307cSWu Fengguang */ 32376ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 32386ae11b27SLee Schermerhorn int delta) 323920a0307cSWu Fengguang { 3240b2261026SJoonsoo Kim int nr_nodes, node; 324120a0307cSWu Fengguang 32429487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 324320a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 324420a0307cSWu Fengguang 3245e8c5c824SLee Schermerhorn if (delta < 0) { 3246b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 3247b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 3248b2261026SJoonsoo Kim goto found; 3249b2261026SJoonsoo Kim } 3250b2261026SJoonsoo Kim } else { 3251b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 3252b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 3253b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 3254b2261026SJoonsoo Kim goto found; 3255e8c5c824SLee Schermerhorn } 32569a76db09SLee Schermerhorn } 3257b2261026SJoonsoo Kim return 0; 325820a0307cSWu Fengguang 3259b2261026SJoonsoo Kim found: 326020a0307cSWu Fengguang h->surplus_huge_pages += delta; 3261b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 3262b2261026SJoonsoo Kim return 1; 326320a0307cSWu Fengguang } 326420a0307cSWu Fengguang 3265a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 3266fd875dcaSMike Kravetz static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 32676ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 32681da177e4SLinus Torvalds { 32697893d1d5SAdam Litke unsigned long min_count, ret; 327010c6ec49SMike Kravetz struct page *page; 327110c6ec49SMike Kravetz LIST_HEAD(page_list); 3272f60858f9SMike Kravetz NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); 3273f60858f9SMike Kravetz 3274f60858f9SMike Kravetz /* 3275f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3276f60858f9SMike Kravetz * If we can not allocate the bit mask, do not attempt to allocate 3277f60858f9SMike Kravetz * the requested huge pages. 3278f60858f9SMike Kravetz */ 3279f60858f9SMike Kravetz if (node_alloc_noretry) 3280f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 3281f60858f9SMike Kravetz else 3282f60858f9SMike Kravetz return -ENOMEM; 32831da177e4SLinus Torvalds 328429383967SMike Kravetz /* 328529383967SMike Kravetz * resize_lock mutex prevents concurrent adjustments to number of 328629383967SMike Kravetz * pages in hstate via the proc/sysfs interfaces. 328729383967SMike Kravetz */ 328829383967SMike Kravetz mutex_lock(&h->resize_lock); 3289b65d4adbSMuchun Song flush_free_hpage_work(h); 3290db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 32914eb0716eSAlexandre Ghiti 32924eb0716eSAlexandre Ghiti /* 3293fd875dcaSMike Kravetz * Check for a node specific request. 3294fd875dcaSMike Kravetz * Changing node specific huge page count may require a corresponding 3295fd875dcaSMike Kravetz * change to the global count. In any case, the passed node mask 3296fd875dcaSMike Kravetz * (nodes_allowed) will restrict alloc/free to the specified node. 3297fd875dcaSMike Kravetz */ 3298fd875dcaSMike Kravetz if (nid != NUMA_NO_NODE) { 3299fd875dcaSMike Kravetz unsigned long old_count = count; 3300fd875dcaSMike Kravetz 3301fd875dcaSMike Kravetz count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 3302fd875dcaSMike Kravetz /* 3303fd875dcaSMike Kravetz * User may have specified a large count value which caused the 3304fd875dcaSMike Kravetz * above calculation to overflow. In this case, they wanted 3305fd875dcaSMike Kravetz * to allocate as many huge pages as possible. Set count to 3306fd875dcaSMike Kravetz * largest possible value to align with their intention. 3307fd875dcaSMike Kravetz */ 3308fd875dcaSMike Kravetz if (count < old_count) 3309fd875dcaSMike Kravetz count = ULONG_MAX; 3310fd875dcaSMike Kravetz } 3311fd875dcaSMike Kravetz 3312fd875dcaSMike Kravetz /* 33134eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 33144eb0716eSAlexandre Ghiti * page range allocation. 33154eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 33164eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 33174eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 33184eb0716eSAlexandre Ghiti */ 33194eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 33204eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 3321db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 332229383967SMike Kravetz mutex_unlock(&h->resize_lock); 3323f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 33244eb0716eSAlexandre Ghiti return -EINVAL; 33254eb0716eSAlexandre Ghiti } 33264eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 33274eb0716eSAlexandre Ghiti } 3328aa888a74SAndi Kleen 33297893d1d5SAdam Litke /* 33307893d1d5SAdam Litke * Increase the pool size 33317893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 33327893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 3333d1c3fb1fSNishanth Aravamudan * 33340c397daeSMichal Hocko * We might race with alloc_surplus_huge_page() here and be unable 3335d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 3336d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 3337d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 3338d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 33397893d1d5SAdam Litke */ 3340a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 33416ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 33427893d1d5SAdam Litke break; 33437893d1d5SAdam Litke } 33447893d1d5SAdam Litke 3345a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 33467893d1d5SAdam Litke /* 33477893d1d5SAdam Litke * If this allocation races such that we no longer need the 33487893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 33497893d1d5SAdam Litke * and reducing the surplus. 33507893d1d5SAdam Litke */ 3351db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3352649920c6SJia He 3353649920c6SJia He /* yield cpu to avoid soft lockup */ 3354649920c6SJia He cond_resched(); 3355649920c6SJia He 3356f60858f9SMike Kravetz ret = alloc_pool_huge_page(h, nodes_allowed, 3357f60858f9SMike Kravetz node_alloc_noretry); 3358db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 33597893d1d5SAdam Litke if (!ret) 33607893d1d5SAdam Litke goto out; 33617893d1d5SAdam Litke 3362536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 3363536240f2SMel Gorman if (signal_pending(current)) 3364536240f2SMel Gorman goto out; 33657893d1d5SAdam Litke } 33667893d1d5SAdam Litke 33677893d1d5SAdam Litke /* 33687893d1d5SAdam Litke * Decrease the pool size 33697893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 33707893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 33717893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 33727893d1d5SAdam Litke * to the desired size as pages become free. 3373d1c3fb1fSNishanth Aravamudan * 3374d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 3375d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 3376d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 33770c397daeSMichal Hocko * alloc_surplus_huge_page() is checking the global counter, 3378d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 3379d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 3380d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 33817893d1d5SAdam Litke */ 3382a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 33836b0c880dSAdam Litke min_count = max(count, min_count); 33846ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 338510c6ec49SMike Kravetz 338610c6ec49SMike Kravetz /* 338710c6ec49SMike Kravetz * Collect pages to be removed on list without dropping lock 338810c6ec49SMike Kravetz */ 3389a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 339010c6ec49SMike Kravetz page = remove_pool_huge_page(h, nodes_allowed, 0); 339110c6ec49SMike Kravetz if (!page) 33921da177e4SLinus Torvalds break; 339310c6ec49SMike Kravetz 339410c6ec49SMike Kravetz list_add(&page->lru, &page_list); 33951da177e4SLinus Torvalds } 339610c6ec49SMike Kravetz /* free the pages after dropping lock */ 3397db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 339810c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3399b65d4adbSMuchun Song flush_free_hpage_work(h); 3400db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 340110c6ec49SMike Kravetz 3402a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 34036ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 34047893d1d5SAdam Litke break; 34057893d1d5SAdam Litke } 34067893d1d5SAdam Litke out: 34074eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 3408db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 340929383967SMike Kravetz mutex_unlock(&h->resize_lock); 34104eb0716eSAlexandre Ghiti 3411f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 3412f60858f9SMike Kravetz 34134eb0716eSAlexandre Ghiti return 0; 34141da177e4SLinus Torvalds } 34151da177e4SLinus Torvalds 34168531fc6fSMike Kravetz static int demote_free_huge_page(struct hstate *h, struct page *page) 34178531fc6fSMike Kravetz { 34188531fc6fSMike Kravetz int i, nid = page_to_nid(page); 34198531fc6fSMike Kravetz struct hstate *target_hstate; 342031731452SDoug Berger struct page *subpage; 34218531fc6fSMike Kravetz int rc = 0; 34228531fc6fSMike Kravetz 34238531fc6fSMike Kravetz target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order); 34248531fc6fSMike Kravetz 34258531fc6fSMike Kravetz remove_hugetlb_page_for_demote(h, page, false); 34268531fc6fSMike Kravetz spin_unlock_irq(&hugetlb_lock); 34278531fc6fSMike Kravetz 34286213834cSMuchun Song rc = hugetlb_vmemmap_restore(h, page); 34298531fc6fSMike Kravetz if (rc) { 34308531fc6fSMike Kravetz /* Allocation of vmemmmap failed, we can not demote page */ 34318531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 34328531fc6fSMike Kravetz set_page_refcounted(page); 34338531fc6fSMike Kravetz add_hugetlb_page(h, page, false); 34348531fc6fSMike Kravetz return rc; 34358531fc6fSMike Kravetz } 34368531fc6fSMike Kravetz 34378531fc6fSMike Kravetz /* 34388531fc6fSMike Kravetz * Use destroy_compound_hugetlb_page_for_demote for all huge page 34398531fc6fSMike Kravetz * sizes as it will not ref count pages. 34408531fc6fSMike Kravetz */ 34418531fc6fSMike Kravetz destroy_compound_hugetlb_page_for_demote(page, huge_page_order(h)); 34428531fc6fSMike Kravetz 34438531fc6fSMike Kravetz /* 34448531fc6fSMike Kravetz * Taking target hstate mutex synchronizes with set_max_huge_pages. 34458531fc6fSMike Kravetz * Without the mutex, pages added to target hstate could be marked 34468531fc6fSMike Kravetz * as surplus. 34478531fc6fSMike Kravetz * 34488531fc6fSMike Kravetz * Note that we already hold h->resize_lock. To prevent deadlock, 34498531fc6fSMike Kravetz * use the convention of always taking larger size hstate mutex first. 34508531fc6fSMike Kravetz */ 34518531fc6fSMike Kravetz mutex_lock(&target_hstate->resize_lock); 34528531fc6fSMike Kravetz for (i = 0; i < pages_per_huge_page(h); 34538531fc6fSMike Kravetz i += pages_per_huge_page(target_hstate)) { 345431731452SDoug Berger subpage = nth_page(page, i); 34558531fc6fSMike Kravetz if (hstate_is_gigantic(target_hstate)) 345631731452SDoug Berger prep_compound_gigantic_page_for_demote(subpage, 34578531fc6fSMike Kravetz target_hstate->order); 34588531fc6fSMike Kravetz else 345931731452SDoug Berger prep_compound_page(subpage, target_hstate->order); 346031731452SDoug Berger set_page_private(subpage, 0); 346131731452SDoug Berger prep_new_huge_page(target_hstate, subpage, nid); 34622b21624fSMike Kravetz free_huge_page(subpage); 34638531fc6fSMike Kravetz } 34648531fc6fSMike Kravetz mutex_unlock(&target_hstate->resize_lock); 34658531fc6fSMike Kravetz 34668531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 34678531fc6fSMike Kravetz 34688531fc6fSMike Kravetz /* 34698531fc6fSMike Kravetz * Not absolutely necessary, but for consistency update max_huge_pages 34708531fc6fSMike Kravetz * based on pool changes for the demoted page. 34718531fc6fSMike Kravetz */ 34728531fc6fSMike Kravetz h->max_huge_pages--; 3473a43a83c7SMiaohe Lin target_hstate->max_huge_pages += 3474a43a83c7SMiaohe Lin pages_per_huge_page(h) / pages_per_huge_page(target_hstate); 34758531fc6fSMike Kravetz 34768531fc6fSMike Kravetz return rc; 34778531fc6fSMike Kravetz } 34788531fc6fSMike Kravetz 347979dfc695SMike Kravetz static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 348079dfc695SMike Kravetz __must_hold(&hugetlb_lock) 348179dfc695SMike Kravetz { 34828531fc6fSMike Kravetz int nr_nodes, node; 34838531fc6fSMike Kravetz struct page *page; 348479dfc695SMike Kravetz 348579dfc695SMike Kravetz lockdep_assert_held(&hugetlb_lock); 348679dfc695SMike Kravetz 348779dfc695SMike Kravetz /* We should never get here if no demote order */ 348879dfc695SMike Kravetz if (!h->demote_order) { 348979dfc695SMike Kravetz pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n"); 349079dfc695SMike Kravetz return -EINVAL; /* internal error */ 349179dfc695SMike Kravetz } 349279dfc695SMike Kravetz 34938531fc6fSMike Kravetz for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 34945a317412SMike Kravetz list_for_each_entry(page, &h->hugepage_freelists[node], lru) { 34955a317412SMike Kravetz if (PageHWPoison(page)) 34965a317412SMike Kravetz continue; 34975a317412SMike Kravetz 34985a317412SMike Kravetz return demote_free_huge_page(h, page); 34998531fc6fSMike Kravetz } 35008531fc6fSMike Kravetz } 35018531fc6fSMike Kravetz 35025a317412SMike Kravetz /* 35035a317412SMike Kravetz * Only way to get here is if all pages on free lists are poisoned. 35045a317412SMike Kravetz * Return -EBUSY so that caller will not retry. 35055a317412SMike Kravetz */ 35065a317412SMike Kravetz return -EBUSY; 350779dfc695SMike Kravetz } 350879dfc695SMike Kravetz 3509a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 3510a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 3511a3437870SNishanth Aravamudan 351279dfc695SMike Kravetz #define HSTATE_ATTR_WO(_name) \ 351379dfc695SMike Kravetz static struct kobj_attribute _name##_attr = __ATTR_WO(_name) 351479dfc695SMike Kravetz 3515a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 351698bc26acSMiaohe Lin static struct kobj_attribute _name##_attr = __ATTR_RW(_name) 3517a3437870SNishanth Aravamudan 3518a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 3519a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 3520a3437870SNishanth Aravamudan 35219a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 35229a305230SLee Schermerhorn 35239a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 3524a3437870SNishanth Aravamudan { 3525a3437870SNishanth Aravamudan int i; 35269a305230SLee Schermerhorn 3527a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 35289a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 35299a305230SLee Schermerhorn if (nidp) 35309a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 3531a3437870SNishanth Aravamudan return &hstates[i]; 35329a305230SLee Schermerhorn } 35339a305230SLee Schermerhorn 35349a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 3535a3437870SNishanth Aravamudan } 3536a3437870SNishanth Aravamudan 353706808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 3538a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3539a3437870SNishanth Aravamudan { 35409a305230SLee Schermerhorn struct hstate *h; 35419a305230SLee Schermerhorn unsigned long nr_huge_pages; 35429a305230SLee Schermerhorn int nid; 35439a305230SLee Schermerhorn 35449a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 35459a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 35469a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 35479a305230SLee Schermerhorn else 35489a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 35499a305230SLee Schermerhorn 3550ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", nr_huge_pages); 3551a3437870SNishanth Aravamudan } 3552adbe8726SEric B Munson 3553238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 3554238d3c13SDavid Rientjes struct hstate *h, int nid, 3555238d3c13SDavid Rientjes unsigned long count, size_t len) 3556a3437870SNishanth Aravamudan { 3557a3437870SNishanth Aravamudan int err; 35582d0adf7eSOscar Salvador nodemask_t nodes_allowed, *n_mask; 3559a3437870SNishanth Aravamudan 35602d0adf7eSOscar Salvador if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 35612d0adf7eSOscar Salvador return -EINVAL; 3562adbe8726SEric B Munson 35639a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 35649a305230SLee Schermerhorn /* 35659a305230SLee Schermerhorn * global hstate attribute 35669a305230SLee Schermerhorn */ 35679a305230SLee Schermerhorn if (!(obey_mempolicy && 35682d0adf7eSOscar Salvador init_nodemask_of_mempolicy(&nodes_allowed))) 35692d0adf7eSOscar Salvador n_mask = &node_states[N_MEMORY]; 35702d0adf7eSOscar Salvador else 35712d0adf7eSOscar Salvador n_mask = &nodes_allowed; 35722d0adf7eSOscar Salvador } else { 35739a305230SLee Schermerhorn /* 3574fd875dcaSMike Kravetz * Node specific request. count adjustment happens in 3575fd875dcaSMike Kravetz * set_max_huge_pages() after acquiring hugetlb_lock. 35769a305230SLee Schermerhorn */ 35772d0adf7eSOscar Salvador init_nodemask_of_node(&nodes_allowed, nid); 35782d0adf7eSOscar Salvador n_mask = &nodes_allowed; 3579fd875dcaSMike Kravetz } 35809a305230SLee Schermerhorn 35812d0adf7eSOscar Salvador err = set_max_huge_pages(h, count, nid, n_mask); 358206808b08SLee Schermerhorn 35834eb0716eSAlexandre Ghiti return err ? err : len; 358406808b08SLee Schermerhorn } 358506808b08SLee Schermerhorn 3586238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 3587238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 3588238d3c13SDavid Rientjes size_t len) 3589238d3c13SDavid Rientjes { 3590238d3c13SDavid Rientjes struct hstate *h; 3591238d3c13SDavid Rientjes unsigned long count; 3592238d3c13SDavid Rientjes int nid; 3593238d3c13SDavid Rientjes int err; 3594238d3c13SDavid Rientjes 3595238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 3596238d3c13SDavid Rientjes if (err) 3597238d3c13SDavid Rientjes return err; 3598238d3c13SDavid Rientjes 3599238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 3600238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 3601238d3c13SDavid Rientjes } 3602238d3c13SDavid Rientjes 360306808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 360406808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 360506808b08SLee Schermerhorn { 360606808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 360706808b08SLee Schermerhorn } 360806808b08SLee Schermerhorn 360906808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 361006808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 361106808b08SLee Schermerhorn { 3612238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 3613a3437870SNishanth Aravamudan } 3614a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 3615a3437870SNishanth Aravamudan 361606808b08SLee Schermerhorn #ifdef CONFIG_NUMA 361706808b08SLee Schermerhorn 361806808b08SLee Schermerhorn /* 361906808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 362006808b08SLee Schermerhorn * huge page alloc/free. 362106808b08SLee Schermerhorn */ 362206808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 3623ae7a927dSJoe Perches struct kobj_attribute *attr, 3624ae7a927dSJoe Perches char *buf) 362506808b08SLee Schermerhorn { 362606808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 362706808b08SLee Schermerhorn } 362806808b08SLee Schermerhorn 362906808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 363006808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 363106808b08SLee Schermerhorn { 3632238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 363306808b08SLee Schermerhorn } 363406808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 363506808b08SLee Schermerhorn #endif 363606808b08SLee Schermerhorn 363706808b08SLee Schermerhorn 3638a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 3639a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3640a3437870SNishanth Aravamudan { 36419a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3642ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); 3643a3437870SNishanth Aravamudan } 3644adbe8726SEric B Munson 3645a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 3646a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 3647a3437870SNishanth Aravamudan { 3648a3437870SNishanth Aravamudan int err; 3649a3437870SNishanth Aravamudan unsigned long input; 36509a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3651a3437870SNishanth Aravamudan 3652bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3653adbe8726SEric B Munson return -EINVAL; 3654adbe8726SEric B Munson 36553dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 3656a3437870SNishanth Aravamudan if (err) 365773ae31e5SEric B Munson return err; 3658a3437870SNishanth Aravamudan 3659db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3660a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 3661db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3662a3437870SNishanth Aravamudan 3663a3437870SNishanth Aravamudan return count; 3664a3437870SNishanth Aravamudan } 3665a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 3666a3437870SNishanth Aravamudan 3667a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 3668a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3669a3437870SNishanth Aravamudan { 36709a305230SLee Schermerhorn struct hstate *h; 36719a305230SLee Schermerhorn unsigned long free_huge_pages; 36729a305230SLee Schermerhorn int nid; 36739a305230SLee Schermerhorn 36749a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 36759a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 36769a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 36779a305230SLee Schermerhorn else 36789a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 36799a305230SLee Schermerhorn 3680ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", free_huge_pages); 3681a3437870SNishanth Aravamudan } 3682a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 3683a3437870SNishanth Aravamudan 3684a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 3685a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3686a3437870SNishanth Aravamudan { 36879a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3688ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); 3689a3437870SNishanth Aravamudan } 3690a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 3691a3437870SNishanth Aravamudan 3692a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 3693a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3694a3437870SNishanth Aravamudan { 36959a305230SLee Schermerhorn struct hstate *h; 36969a305230SLee Schermerhorn unsigned long surplus_huge_pages; 36979a305230SLee Schermerhorn int nid; 36989a305230SLee Schermerhorn 36999a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 37009a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 37019a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 37029a305230SLee Schermerhorn else 37039a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 37049a305230SLee Schermerhorn 3705ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", surplus_huge_pages); 3706a3437870SNishanth Aravamudan } 3707a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 3708a3437870SNishanth Aravamudan 370979dfc695SMike Kravetz static ssize_t demote_store(struct kobject *kobj, 371079dfc695SMike Kravetz struct kobj_attribute *attr, const char *buf, size_t len) 371179dfc695SMike Kravetz { 371279dfc695SMike Kravetz unsigned long nr_demote; 371379dfc695SMike Kravetz unsigned long nr_available; 371479dfc695SMike Kravetz nodemask_t nodes_allowed, *n_mask; 371579dfc695SMike Kravetz struct hstate *h; 37168eeda55fSLi zeming int err; 371779dfc695SMike Kravetz int nid; 371879dfc695SMike Kravetz 371979dfc695SMike Kravetz err = kstrtoul(buf, 10, &nr_demote); 372079dfc695SMike Kravetz if (err) 372179dfc695SMike Kravetz return err; 372279dfc695SMike Kravetz h = kobj_to_hstate(kobj, &nid); 372379dfc695SMike Kravetz 372479dfc695SMike Kravetz if (nid != NUMA_NO_NODE) { 372579dfc695SMike Kravetz init_nodemask_of_node(&nodes_allowed, nid); 372679dfc695SMike Kravetz n_mask = &nodes_allowed; 372779dfc695SMike Kravetz } else { 372879dfc695SMike Kravetz n_mask = &node_states[N_MEMORY]; 372979dfc695SMike Kravetz } 373079dfc695SMike Kravetz 373179dfc695SMike Kravetz /* Synchronize with other sysfs operations modifying huge pages */ 373279dfc695SMike Kravetz mutex_lock(&h->resize_lock); 373379dfc695SMike Kravetz spin_lock_irq(&hugetlb_lock); 373479dfc695SMike Kravetz 373579dfc695SMike Kravetz while (nr_demote) { 373679dfc695SMike Kravetz /* 373779dfc695SMike Kravetz * Check for available pages to demote each time thorough the 373879dfc695SMike Kravetz * loop as demote_pool_huge_page will drop hugetlb_lock. 373979dfc695SMike Kravetz */ 374079dfc695SMike Kravetz if (nid != NUMA_NO_NODE) 374179dfc695SMike Kravetz nr_available = h->free_huge_pages_node[nid]; 374279dfc695SMike Kravetz else 374379dfc695SMike Kravetz nr_available = h->free_huge_pages; 374479dfc695SMike Kravetz nr_available -= h->resv_huge_pages; 374579dfc695SMike Kravetz if (!nr_available) 374679dfc695SMike Kravetz break; 374779dfc695SMike Kravetz 374879dfc695SMike Kravetz err = demote_pool_huge_page(h, n_mask); 374979dfc695SMike Kravetz if (err) 375079dfc695SMike Kravetz break; 375179dfc695SMike Kravetz 375279dfc695SMike Kravetz nr_demote--; 375379dfc695SMike Kravetz } 375479dfc695SMike Kravetz 375579dfc695SMike Kravetz spin_unlock_irq(&hugetlb_lock); 375679dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 375779dfc695SMike Kravetz 375879dfc695SMike Kravetz if (err) 375979dfc695SMike Kravetz return err; 376079dfc695SMike Kravetz return len; 376179dfc695SMike Kravetz } 376279dfc695SMike Kravetz HSTATE_ATTR_WO(demote); 376379dfc695SMike Kravetz 376479dfc695SMike Kravetz static ssize_t demote_size_show(struct kobject *kobj, 376579dfc695SMike Kravetz struct kobj_attribute *attr, char *buf) 376679dfc695SMike Kravetz { 376712658abfSMiaohe Lin struct hstate *h = kobj_to_hstate(kobj, NULL); 376879dfc695SMike Kravetz unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K; 376979dfc695SMike Kravetz 377079dfc695SMike Kravetz return sysfs_emit(buf, "%lukB\n", demote_size); 377179dfc695SMike Kravetz } 377279dfc695SMike Kravetz 377379dfc695SMike Kravetz static ssize_t demote_size_store(struct kobject *kobj, 377479dfc695SMike Kravetz struct kobj_attribute *attr, 377579dfc695SMike Kravetz const char *buf, size_t count) 377679dfc695SMike Kravetz { 377779dfc695SMike Kravetz struct hstate *h, *demote_hstate; 377879dfc695SMike Kravetz unsigned long demote_size; 377979dfc695SMike Kravetz unsigned int demote_order; 378079dfc695SMike Kravetz 378179dfc695SMike Kravetz demote_size = (unsigned long)memparse(buf, NULL); 378279dfc695SMike Kravetz 378379dfc695SMike Kravetz demote_hstate = size_to_hstate(demote_size); 378479dfc695SMike Kravetz if (!demote_hstate) 378579dfc695SMike Kravetz return -EINVAL; 378679dfc695SMike Kravetz demote_order = demote_hstate->order; 3787a01f4390SMike Kravetz if (demote_order < HUGETLB_PAGE_ORDER) 3788a01f4390SMike Kravetz return -EINVAL; 378979dfc695SMike Kravetz 379079dfc695SMike Kravetz /* demote order must be smaller than hstate order */ 379112658abfSMiaohe Lin h = kobj_to_hstate(kobj, NULL); 379279dfc695SMike Kravetz if (demote_order >= h->order) 379379dfc695SMike Kravetz return -EINVAL; 379479dfc695SMike Kravetz 379579dfc695SMike Kravetz /* resize_lock synchronizes access to demote size and writes */ 379679dfc695SMike Kravetz mutex_lock(&h->resize_lock); 379779dfc695SMike Kravetz h->demote_order = demote_order; 379879dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 379979dfc695SMike Kravetz 380079dfc695SMike Kravetz return count; 380179dfc695SMike Kravetz } 380279dfc695SMike Kravetz HSTATE_ATTR(demote_size); 380379dfc695SMike Kravetz 3804a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 3805a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 3806a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 3807a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 3808a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 3809a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 381006808b08SLee Schermerhorn #ifdef CONFIG_NUMA 381106808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 381206808b08SLee Schermerhorn #endif 3813a3437870SNishanth Aravamudan NULL, 3814a3437870SNishanth Aravamudan }; 3815a3437870SNishanth Aravamudan 381667e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 3817a3437870SNishanth Aravamudan .attrs = hstate_attrs, 3818a3437870SNishanth Aravamudan }; 3819a3437870SNishanth Aravamudan 382079dfc695SMike Kravetz static struct attribute *hstate_demote_attrs[] = { 382179dfc695SMike Kravetz &demote_size_attr.attr, 382279dfc695SMike Kravetz &demote_attr.attr, 382379dfc695SMike Kravetz NULL, 382479dfc695SMike Kravetz }; 382579dfc695SMike Kravetz 382679dfc695SMike Kravetz static const struct attribute_group hstate_demote_attr_group = { 382779dfc695SMike Kravetz .attrs = hstate_demote_attrs, 382879dfc695SMike Kravetz }; 382979dfc695SMike Kravetz 3830094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 38319a305230SLee Schermerhorn struct kobject **hstate_kobjs, 383267e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 3833a3437870SNishanth Aravamudan { 3834a3437870SNishanth Aravamudan int retval; 3835972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 3836a3437870SNishanth Aravamudan 38379a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 38389a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 3839a3437870SNishanth Aravamudan return -ENOMEM; 3840a3437870SNishanth Aravamudan 38419a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 3842cc2205a6SMiaohe Lin if (retval) { 38439a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 3844cc2205a6SMiaohe Lin hstate_kobjs[hi] = NULL; 38453a6bdda0SMiaohe Lin return retval; 3846cc2205a6SMiaohe Lin } 3847a3437870SNishanth Aravamudan 384879dfc695SMike Kravetz if (h->demote_order) { 384901088a60SMiaohe Lin retval = sysfs_create_group(hstate_kobjs[hi], 385001088a60SMiaohe Lin &hstate_demote_attr_group); 385101088a60SMiaohe Lin if (retval) { 385279dfc695SMike Kravetz pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name); 385301088a60SMiaohe Lin sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group); 385401088a60SMiaohe Lin kobject_put(hstate_kobjs[hi]); 385501088a60SMiaohe Lin hstate_kobjs[hi] = NULL; 385601088a60SMiaohe Lin return retval; 385701088a60SMiaohe Lin } 385879dfc695SMike Kravetz } 385979dfc695SMike Kravetz 386001088a60SMiaohe Lin return 0; 3861a3437870SNishanth Aravamudan } 3862a3437870SNishanth Aravamudan 38639a305230SLee Schermerhorn #ifdef CONFIG_NUMA 3864a4a00b45SMuchun Song static bool hugetlb_sysfs_initialized __ro_after_init; 38659a305230SLee Schermerhorn 38669a305230SLee Schermerhorn /* 38679a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 386810fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 386910fbcf4cSKay Sievers * index of a node device or _hstate == node id. 387010fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 38719a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 38729a305230SLee Schermerhorn */ 38739a305230SLee Schermerhorn struct node_hstate { 38749a305230SLee Schermerhorn struct kobject *hugepages_kobj; 38759a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 38769a305230SLee Schermerhorn }; 3877b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 38789a305230SLee Schermerhorn 38799a305230SLee Schermerhorn /* 388010fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 38819a305230SLee Schermerhorn */ 38829a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 38839a305230SLee Schermerhorn &nr_hugepages_attr.attr, 38849a305230SLee Schermerhorn &free_hugepages_attr.attr, 38859a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 38869a305230SLee Schermerhorn NULL, 38879a305230SLee Schermerhorn }; 38889a305230SLee Schermerhorn 388967e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 38909a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 38919a305230SLee Schermerhorn }; 38929a305230SLee Schermerhorn 38939a305230SLee Schermerhorn /* 389410fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 38959a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 38969a305230SLee Schermerhorn */ 38979a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 38989a305230SLee Schermerhorn { 38999a305230SLee Schermerhorn int nid; 39009a305230SLee Schermerhorn 39019a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 39029a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 39039a305230SLee Schermerhorn int i; 39049a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 39059a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 39069a305230SLee Schermerhorn if (nidp) 39079a305230SLee Schermerhorn *nidp = nid; 39089a305230SLee Schermerhorn return &hstates[i]; 39099a305230SLee Schermerhorn } 39109a305230SLee Schermerhorn } 39119a305230SLee Schermerhorn 39129a305230SLee Schermerhorn BUG(); 39139a305230SLee Schermerhorn return NULL; 39149a305230SLee Schermerhorn } 39159a305230SLee Schermerhorn 39169a305230SLee Schermerhorn /* 391710fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 39189a305230SLee Schermerhorn * No-op if no hstate attributes attached. 39199a305230SLee Schermerhorn */ 3920a4a00b45SMuchun Song void hugetlb_unregister_node(struct node *node) 39219a305230SLee Schermerhorn { 39229a305230SLee Schermerhorn struct hstate *h; 392310fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 39249a305230SLee Schermerhorn 39259a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 39269b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 39279a305230SLee Schermerhorn 3928972dc4deSAneesh Kumar K.V for_each_hstate(h) { 3929972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 393001088a60SMiaohe Lin struct kobject *hstate_kobj = nhs->hstate_kobjs[idx]; 393101088a60SMiaohe Lin 393201088a60SMiaohe Lin if (!hstate_kobj) 393301088a60SMiaohe Lin continue; 393401088a60SMiaohe Lin if (h->demote_order) 393501088a60SMiaohe Lin sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group); 393601088a60SMiaohe Lin sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group); 393701088a60SMiaohe Lin kobject_put(hstate_kobj); 3938972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 3939972dc4deSAneesh Kumar K.V } 39409a305230SLee Schermerhorn 39419a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 39429a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 39439a305230SLee Schermerhorn } 39449a305230SLee Schermerhorn 39459a305230SLee Schermerhorn 39469a305230SLee Schermerhorn /* 394710fbcf4cSKay Sievers * Register hstate attributes for a single node device. 39489a305230SLee Schermerhorn * No-op if attributes already registered. 39499a305230SLee Schermerhorn */ 3950a4a00b45SMuchun Song void hugetlb_register_node(struct node *node) 39519a305230SLee Schermerhorn { 39529a305230SLee Schermerhorn struct hstate *h; 395310fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 39549a305230SLee Schermerhorn int err; 39559a305230SLee Schermerhorn 3956a4a00b45SMuchun Song if (!hugetlb_sysfs_initialized) 3957a4a00b45SMuchun Song return; 3958a4a00b45SMuchun Song 39599a305230SLee Schermerhorn if (nhs->hugepages_kobj) 39609a305230SLee Schermerhorn return; /* already allocated */ 39619a305230SLee Schermerhorn 39629a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 396310fbcf4cSKay Sievers &node->dev.kobj); 39649a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 39659a305230SLee Schermerhorn return; 39669a305230SLee Schermerhorn 39679a305230SLee Schermerhorn for_each_hstate(h) { 39689a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 39699a305230SLee Schermerhorn nhs->hstate_kobjs, 39709a305230SLee Schermerhorn &per_node_hstate_attr_group); 39719a305230SLee Schermerhorn if (err) { 3972282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s for node %d\n", 397310fbcf4cSKay Sievers h->name, node->dev.id); 39749a305230SLee Schermerhorn hugetlb_unregister_node(node); 39759a305230SLee Schermerhorn break; 39769a305230SLee Schermerhorn } 39779a305230SLee Schermerhorn } 39789a305230SLee Schermerhorn } 39799a305230SLee Schermerhorn 39809a305230SLee Schermerhorn /* 39819b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 398210fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 398310fbcf4cSKay Sievers * registered their associated device by this time. 39849a305230SLee Schermerhorn */ 39857d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 39869a305230SLee Schermerhorn { 39879a305230SLee Schermerhorn int nid; 39889a305230SLee Schermerhorn 3989a4a00b45SMuchun Song for_each_online_node(nid) 3990b958d4d0SMuchun Song hugetlb_register_node(node_devices[nid]); 39919a305230SLee Schermerhorn } 39929a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 39939a305230SLee Schermerhorn 39949a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 39959a305230SLee Schermerhorn { 39969a305230SLee Schermerhorn BUG(); 39979a305230SLee Schermerhorn if (nidp) 39989a305230SLee Schermerhorn *nidp = -1; 39999a305230SLee Schermerhorn return NULL; 40009a305230SLee Schermerhorn } 40019a305230SLee Schermerhorn 40029a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 40039a305230SLee Schermerhorn 40049a305230SLee Schermerhorn #endif 40059a305230SLee Schermerhorn 4006263b8998SMiaohe Lin #ifdef CONFIG_CMA 4007263b8998SMiaohe Lin static void __init hugetlb_cma_check(void); 4008263b8998SMiaohe Lin #else 4009263b8998SMiaohe Lin static inline __init void hugetlb_cma_check(void) 4010263b8998SMiaohe Lin { 4011263b8998SMiaohe Lin } 4012263b8998SMiaohe Lin #endif 4013263b8998SMiaohe Lin 4014a4a00b45SMuchun Song static void __init hugetlb_sysfs_init(void) 4015a4a00b45SMuchun Song { 4016a4a00b45SMuchun Song struct hstate *h; 4017a4a00b45SMuchun Song int err; 4018a4a00b45SMuchun Song 4019a4a00b45SMuchun Song hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 4020a4a00b45SMuchun Song if (!hugepages_kobj) 4021a4a00b45SMuchun Song return; 4022a4a00b45SMuchun Song 4023a4a00b45SMuchun Song for_each_hstate(h) { 4024a4a00b45SMuchun Song err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 4025a4a00b45SMuchun Song hstate_kobjs, &hstate_attr_group); 4026a4a00b45SMuchun Song if (err) 4027a4a00b45SMuchun Song pr_err("HugeTLB: Unable to add hstate %s", h->name); 4028a4a00b45SMuchun Song } 4029a4a00b45SMuchun Song 4030a4a00b45SMuchun Song #ifdef CONFIG_NUMA 4031a4a00b45SMuchun Song hugetlb_sysfs_initialized = true; 4032a4a00b45SMuchun Song #endif 4033a4a00b45SMuchun Song hugetlb_register_all_nodes(); 4034a4a00b45SMuchun Song } 4035a4a00b45SMuchun Song 4036a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 4037a3437870SNishanth Aravamudan { 40388382d914SDavidlohr Bueso int i; 40398382d914SDavidlohr Bueso 4040d6995da3SMike Kravetz BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < 4041d6995da3SMike Kravetz __NR_HPAGEFLAGS); 4042d6995da3SMike Kravetz 4043c2833a5bSMike Kravetz if (!hugepages_supported()) { 4044c2833a5bSMike Kravetz if (hugetlb_max_hstate || default_hstate_max_huge_pages) 4045c2833a5bSMike Kravetz pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); 40460ef89d25SBenjamin Herrenschmidt return 0; 4047d715cf80SLiam R. Howlett } 4048d715cf80SLiam R. Howlett 4049282f4214SMike Kravetz /* 4050282f4214SMike Kravetz * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some 4051282f4214SMike Kravetz * architectures depend on setup being done here. 4052282f4214SMike Kravetz */ 4053a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 4054282f4214SMike Kravetz if (!parsed_default_hugepagesz) { 4055282f4214SMike Kravetz /* 4056282f4214SMike Kravetz * If we did not parse a default huge page size, set 4057282f4214SMike Kravetz * default_hstate_idx to HPAGE_SIZE hstate. And, if the 4058282f4214SMike Kravetz * number of huge pages for this default size was implicitly 4059282f4214SMike Kravetz * specified, set that here as well. 4060282f4214SMike Kravetz * Note that the implicit setting will overwrite an explicit 4061282f4214SMike Kravetz * setting. A warning will be printed in this case. 4062282f4214SMike Kravetz */ 4063282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); 4064f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 4065282f4214SMike Kravetz if (default_hstate.max_huge_pages) { 4066282f4214SMike Kravetz char buf[32]; 4067282f4214SMike Kravetz 4068282f4214SMike Kravetz string_get_size(huge_page_size(&default_hstate), 4069282f4214SMike Kravetz 1, STRING_UNITS_2, buf, 32); 4070282f4214SMike Kravetz pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", 4071282f4214SMike Kravetz default_hstate.max_huge_pages, buf); 4072282f4214SMike Kravetz pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", 4073282f4214SMike Kravetz default_hstate_max_huge_pages); 4074282f4214SMike Kravetz } 4075282f4214SMike Kravetz default_hstate.max_huge_pages = 4076282f4214SMike Kravetz default_hstate_max_huge_pages; 4077b5389086SZhenguo Yao 40780a7a0f6fSPeng Liu for_each_online_node(i) 4079b5389086SZhenguo Yao default_hstate.max_huge_pages_node[i] = 4080b5389086SZhenguo Yao default_hugepages_in_node[i]; 4081282f4214SMike Kravetz } 4082f8b74815SVaishali Thakkar } 4083a3437870SNishanth Aravamudan 4084cf11e85fSRoman Gushchin hugetlb_cma_check(); 4085a3437870SNishanth Aravamudan hugetlb_init_hstates(); 4086aa888a74SAndi Kleen gather_bootmem_prealloc(); 4087a3437870SNishanth Aravamudan report_hugepages(); 4088a3437870SNishanth Aravamudan 4089a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 40907179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 40919a305230SLee Schermerhorn 40928382d914SDavidlohr Bueso #ifdef CONFIG_SMP 40938382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 40948382d914SDavidlohr Bueso #else 40958382d914SDavidlohr Bueso num_fault_mutexes = 1; 40968382d914SDavidlohr Bueso #endif 4097c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 40986da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 40996da2ec56SKees Cook GFP_KERNEL); 4100c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 41018382d914SDavidlohr Bueso 41028382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 4103c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 4104a3437870SNishanth Aravamudan return 0; 4105a3437870SNishanth Aravamudan } 41063e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 4107a3437870SNishanth Aravamudan 4108ae94da89SMike Kravetz /* Overwritten by architectures with more huge page sizes */ 4109ae94da89SMike Kravetz bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) 41109fee021dSVaishali Thakkar { 4111ae94da89SMike Kravetz return size == HPAGE_SIZE; 41129fee021dSVaishali Thakkar } 41139fee021dSVaishali Thakkar 4114d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 4115a3437870SNishanth Aravamudan { 4116a3437870SNishanth Aravamudan struct hstate *h; 41178faa8b07SAndi Kleen unsigned long i; 41188faa8b07SAndi Kleen 4119a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 4120a3437870SNishanth Aravamudan return; 4121a3437870SNishanth Aravamudan } 412247d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 4123a3437870SNishanth Aravamudan BUG_ON(order == 0); 412447d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 412529383967SMike Kravetz mutex_init(&h->resize_lock); 4126a3437870SNishanth Aravamudan h->order = order; 4127aca78307SMiaohe Lin h->mask = ~(huge_page_size(h) - 1); 41288faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 41298faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 41300edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 413154f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 413254f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 4133a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 4134c2c3a60aSMiaohe Lin huge_page_size(h)/SZ_1K); 41358faa8b07SAndi Kleen 4136a3437870SNishanth Aravamudan parsed_hstate = h; 4137a3437870SNishanth Aravamudan } 4138a3437870SNishanth Aravamudan 4139b5389086SZhenguo Yao bool __init __weak hugetlb_node_alloc_supported(void) 4140b5389086SZhenguo Yao { 4141b5389086SZhenguo Yao return true; 4142b5389086SZhenguo Yao } 4143f87442f4SPeng Liu 4144f87442f4SPeng Liu static void __init hugepages_clear_pages_in_node(void) 4145f87442f4SPeng Liu { 4146f87442f4SPeng Liu if (!hugetlb_max_hstate) { 4147f87442f4SPeng Liu default_hstate_max_huge_pages = 0; 4148f87442f4SPeng Liu memset(default_hugepages_in_node, 0, 414910395680SMiaohe Lin sizeof(default_hugepages_in_node)); 4150f87442f4SPeng Liu } else { 4151f87442f4SPeng Liu parsed_hstate->max_huge_pages = 0; 4152f87442f4SPeng Liu memset(parsed_hstate->max_huge_pages_node, 0, 415310395680SMiaohe Lin sizeof(parsed_hstate->max_huge_pages_node)); 4154f87442f4SPeng Liu } 4155f87442f4SPeng Liu } 4156f87442f4SPeng Liu 4157282f4214SMike Kravetz /* 4158282f4214SMike Kravetz * hugepages command line processing 4159282f4214SMike Kravetz * hugepages normally follows a valid hugepagsz or default_hugepagsz 4160282f4214SMike Kravetz * specification. If not, ignore the hugepages value. hugepages can also 4161282f4214SMike Kravetz * be the first huge page command line option in which case it implicitly 4162282f4214SMike Kravetz * specifies the number of huge pages for the default size. 4163282f4214SMike Kravetz */ 4164282f4214SMike Kravetz static int __init hugepages_setup(char *s) 4165a3437870SNishanth Aravamudan { 4166a3437870SNishanth Aravamudan unsigned long *mhp; 41678faa8b07SAndi Kleen static unsigned long *last_mhp; 4168b5389086SZhenguo Yao int node = NUMA_NO_NODE; 4169b5389086SZhenguo Yao int count; 4170b5389086SZhenguo Yao unsigned long tmp; 4171b5389086SZhenguo Yao char *p = s; 4172a3437870SNishanth Aravamudan 41739fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 4174282f4214SMike Kravetz pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); 41759fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 4176f81f6e4bSPeng Liu return 1; 41779fee021dSVaishali Thakkar } 4178282f4214SMike Kravetz 4179a3437870SNishanth Aravamudan /* 4180282f4214SMike Kravetz * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter 4181282f4214SMike Kravetz * yet, so this hugepages= parameter goes to the "default hstate". 4182282f4214SMike Kravetz * Otherwise, it goes with the previously parsed hugepagesz or 4183282f4214SMike Kravetz * default_hugepagesz. 4184a3437870SNishanth Aravamudan */ 41859fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 4186a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 4187a3437870SNishanth Aravamudan else 4188a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 4189a3437870SNishanth Aravamudan 41908faa8b07SAndi Kleen if (mhp == last_mhp) { 4191282f4214SMike Kravetz pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); 4192f81f6e4bSPeng Liu return 1; 41938faa8b07SAndi Kleen } 41948faa8b07SAndi Kleen 4195b5389086SZhenguo Yao while (*p) { 4196b5389086SZhenguo Yao count = 0; 4197b5389086SZhenguo Yao if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4198b5389086SZhenguo Yao goto invalid; 4199b5389086SZhenguo Yao /* Parameter is node format */ 4200b5389086SZhenguo Yao if (p[count] == ':') { 4201b5389086SZhenguo Yao if (!hugetlb_node_alloc_supported()) { 4202b5389086SZhenguo Yao pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n"); 4203f81f6e4bSPeng Liu return 1; 4204b5389086SZhenguo Yao } 42050a7a0f6fSPeng Liu if (tmp >= MAX_NUMNODES || !node_online(tmp)) 4206e79ce983SLiu Yuntao goto invalid; 42070a7a0f6fSPeng Liu node = array_index_nospec(tmp, MAX_NUMNODES); 4208b5389086SZhenguo Yao p += count + 1; 4209b5389086SZhenguo Yao /* Parse hugepages */ 4210b5389086SZhenguo Yao if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4211b5389086SZhenguo Yao goto invalid; 4212b5389086SZhenguo Yao if (!hugetlb_max_hstate) 4213b5389086SZhenguo Yao default_hugepages_in_node[node] = tmp; 4214b5389086SZhenguo Yao else 4215b5389086SZhenguo Yao parsed_hstate->max_huge_pages_node[node] = tmp; 4216b5389086SZhenguo Yao *mhp += tmp; 4217b5389086SZhenguo Yao /* Go to parse next node*/ 4218b5389086SZhenguo Yao if (p[count] == ',') 4219b5389086SZhenguo Yao p += count + 1; 4220b5389086SZhenguo Yao else 4221b5389086SZhenguo Yao break; 4222b5389086SZhenguo Yao } else { 4223b5389086SZhenguo Yao if (p != s) 4224b5389086SZhenguo Yao goto invalid; 4225b5389086SZhenguo Yao *mhp = tmp; 4226b5389086SZhenguo Yao break; 4227b5389086SZhenguo Yao } 4228b5389086SZhenguo Yao } 4229a3437870SNishanth Aravamudan 42308faa8b07SAndi Kleen /* 42318faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 423204adbc3fSMiaohe Lin * But we need to allocate gigantic hstates here early to still 42338faa8b07SAndi Kleen * use the bootmem allocator. 42348faa8b07SAndi Kleen */ 423504adbc3fSMiaohe Lin if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) 42368faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 42378faa8b07SAndi Kleen 42388faa8b07SAndi Kleen last_mhp = mhp; 42398faa8b07SAndi Kleen 4240a3437870SNishanth Aravamudan return 1; 4241b5389086SZhenguo Yao 4242b5389086SZhenguo Yao invalid: 4243b5389086SZhenguo Yao pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p); 4244f87442f4SPeng Liu hugepages_clear_pages_in_node(); 4245f81f6e4bSPeng Liu return 1; 4246a3437870SNishanth Aravamudan } 4247282f4214SMike Kravetz __setup("hugepages=", hugepages_setup); 4248e11bfbfcSNick Piggin 4249282f4214SMike Kravetz /* 4250282f4214SMike Kravetz * hugepagesz command line processing 4251282f4214SMike Kravetz * A specific huge page size can only be specified once with hugepagesz. 4252282f4214SMike Kravetz * hugepagesz is followed by hugepages on the command line. The global 4253282f4214SMike Kravetz * variable 'parsed_valid_hugepagesz' is used to determine if prior 4254282f4214SMike Kravetz * hugepagesz argument was valid. 4255282f4214SMike Kravetz */ 4256359f2544SMike Kravetz static int __init hugepagesz_setup(char *s) 4257e11bfbfcSNick Piggin { 4258359f2544SMike Kravetz unsigned long size; 4259282f4214SMike Kravetz struct hstate *h; 4260282f4214SMike Kravetz 4261282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4262359f2544SMike Kravetz size = (unsigned long)memparse(s, NULL); 4263359f2544SMike Kravetz 4264359f2544SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4265282f4214SMike Kravetz pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); 4266f81f6e4bSPeng Liu return 1; 4267359f2544SMike Kravetz } 4268359f2544SMike Kravetz 4269282f4214SMike Kravetz h = size_to_hstate(size); 4270282f4214SMike Kravetz if (h) { 4271282f4214SMike Kravetz /* 4272282f4214SMike Kravetz * hstate for this size already exists. This is normally 4273282f4214SMike Kravetz * an error, but is allowed if the existing hstate is the 4274282f4214SMike Kravetz * default hstate. More specifically, it is only allowed if 4275282f4214SMike Kravetz * the number of huge pages for the default hstate was not 4276282f4214SMike Kravetz * previously specified. 4277282f4214SMike Kravetz */ 4278282f4214SMike Kravetz if (!parsed_default_hugepagesz || h != &default_hstate || 4279282f4214SMike Kravetz default_hstate.max_huge_pages) { 4280282f4214SMike Kravetz pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); 4281f81f6e4bSPeng Liu return 1; 428238237830SMike Kravetz } 428338237830SMike Kravetz 4284282f4214SMike Kravetz /* 4285282f4214SMike Kravetz * No need to call hugetlb_add_hstate() as hstate already 4286282f4214SMike Kravetz * exists. But, do set parsed_hstate so that a following 4287282f4214SMike Kravetz * hugepages= parameter will be applied to this hstate. 4288282f4214SMike Kravetz */ 4289282f4214SMike Kravetz parsed_hstate = h; 4290282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4291e11bfbfcSNick Piggin return 1; 4292e11bfbfcSNick Piggin } 4293282f4214SMike Kravetz 4294359f2544SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4295282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4296359f2544SMike Kravetz return 1; 4297359f2544SMike Kravetz } 4298359f2544SMike Kravetz __setup("hugepagesz=", hugepagesz_setup); 4299359f2544SMike Kravetz 4300282f4214SMike Kravetz /* 4301282f4214SMike Kravetz * default_hugepagesz command line input 4302282f4214SMike Kravetz * Only one instance of default_hugepagesz allowed on command line. 4303282f4214SMike Kravetz */ 4304ae94da89SMike Kravetz static int __init default_hugepagesz_setup(char *s) 4305e11bfbfcSNick Piggin { 4306ae94da89SMike Kravetz unsigned long size; 4307b5389086SZhenguo Yao int i; 4308ae94da89SMike Kravetz 4309282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4310282f4214SMike Kravetz if (parsed_default_hugepagesz) { 4311282f4214SMike Kravetz pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); 4312f81f6e4bSPeng Liu return 1; 4313282f4214SMike Kravetz } 4314282f4214SMike Kravetz 4315282f4214SMike Kravetz size = (unsigned long)memparse(s, NULL); 4316282f4214SMike Kravetz 4317282f4214SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4318282f4214SMike Kravetz pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); 4319f81f6e4bSPeng Liu return 1; 4320282f4214SMike Kravetz } 4321282f4214SMike Kravetz 4322282f4214SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4323282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4324282f4214SMike Kravetz parsed_default_hugepagesz = true; 4325282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(size)); 4326282f4214SMike Kravetz 4327282f4214SMike Kravetz /* 4328282f4214SMike Kravetz * The number of default huge pages (for this size) could have been 4329282f4214SMike Kravetz * specified as the first hugetlb parameter: hugepages=X. If so, 4330282f4214SMike Kravetz * then default_hstate_max_huge_pages is set. If the default huge 4331282f4214SMike Kravetz * page size is gigantic (>= MAX_ORDER), then the pages must be 4332282f4214SMike Kravetz * allocated here from bootmem allocator. 4333282f4214SMike Kravetz */ 4334282f4214SMike Kravetz if (default_hstate_max_huge_pages) { 4335282f4214SMike Kravetz default_hstate.max_huge_pages = default_hstate_max_huge_pages; 43360a7a0f6fSPeng Liu for_each_online_node(i) 4337b5389086SZhenguo Yao default_hstate.max_huge_pages_node[i] = 4338b5389086SZhenguo Yao default_hugepages_in_node[i]; 4339282f4214SMike Kravetz if (hstate_is_gigantic(&default_hstate)) 4340282f4214SMike Kravetz hugetlb_hstate_alloc_pages(&default_hstate); 4341282f4214SMike Kravetz default_hstate_max_huge_pages = 0; 4342282f4214SMike Kravetz } 4343282f4214SMike Kravetz 4344e11bfbfcSNick Piggin return 1; 4345e11bfbfcSNick Piggin } 4346ae94da89SMike Kravetz __setup("default_hugepagesz=", default_hugepagesz_setup); 4347a3437870SNishanth Aravamudan 4348d2226ebdSFeng Tang static nodemask_t *policy_mbind_nodemask(gfp_t gfp) 4349d2226ebdSFeng Tang { 4350d2226ebdSFeng Tang #ifdef CONFIG_NUMA 4351d2226ebdSFeng Tang struct mempolicy *mpol = get_task_policy(current); 4352d2226ebdSFeng Tang 4353d2226ebdSFeng Tang /* 4354d2226ebdSFeng Tang * Only enforce MPOL_BIND policy which overlaps with cpuset policy 4355d2226ebdSFeng Tang * (from policy_nodemask) specifically for hugetlb case 4356d2226ebdSFeng Tang */ 4357d2226ebdSFeng Tang if (mpol->mode == MPOL_BIND && 4358d2226ebdSFeng Tang (apply_policy_zone(mpol, gfp_zone(gfp)) && 4359d2226ebdSFeng Tang cpuset_nodemask_valid_mems_allowed(&mpol->nodes))) 4360d2226ebdSFeng Tang return &mpol->nodes; 4361d2226ebdSFeng Tang #endif 4362d2226ebdSFeng Tang return NULL; 4363d2226ebdSFeng Tang } 4364d2226ebdSFeng Tang 43658ca39e68SMuchun Song static unsigned int allowed_mems_nr(struct hstate *h) 43668a213460SNishanth Aravamudan { 43678a213460SNishanth Aravamudan int node; 43688a213460SNishanth Aravamudan unsigned int nr = 0; 4369d2226ebdSFeng Tang nodemask_t *mbind_nodemask; 43708ca39e68SMuchun Song unsigned int *array = h->free_huge_pages_node; 43718ca39e68SMuchun Song gfp_t gfp_mask = htlb_alloc_mask(h); 43728a213460SNishanth Aravamudan 4373d2226ebdSFeng Tang mbind_nodemask = policy_mbind_nodemask(gfp_mask); 43748ca39e68SMuchun Song for_each_node_mask(node, cpuset_current_mems_allowed) { 4375d2226ebdSFeng Tang if (!mbind_nodemask || node_isset(node, *mbind_nodemask)) 43768a213460SNishanth Aravamudan nr += array[node]; 43778ca39e68SMuchun Song } 43788a213460SNishanth Aravamudan 43798a213460SNishanth Aravamudan return nr; 43808a213460SNishanth Aravamudan } 43818a213460SNishanth Aravamudan 43828a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 438317743798SMuchun Song static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, 438417743798SMuchun Song void *buffer, size_t *length, 438517743798SMuchun Song loff_t *ppos, unsigned long *out) 438617743798SMuchun Song { 438717743798SMuchun Song struct ctl_table dup_table; 438817743798SMuchun Song 438917743798SMuchun Song /* 439017743798SMuchun Song * In order to avoid races with __do_proc_doulongvec_minmax(), we 439117743798SMuchun Song * can duplicate the @table and alter the duplicate of it. 439217743798SMuchun Song */ 439317743798SMuchun Song dup_table = *table; 439417743798SMuchun Song dup_table.data = out; 439517743798SMuchun Song 439617743798SMuchun Song return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); 439717743798SMuchun Song } 439817743798SMuchun Song 439906808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 440006808b08SLee Schermerhorn struct ctl_table *table, int write, 440132927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 44021da177e4SLinus Torvalds { 4403e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 4404238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 440508d4a246SMichal Hocko int ret; 4406e5ff2159SAndi Kleen 4407457c1b27SNishanth Aravamudan if (!hugepages_supported()) 440886613628SJan Stancek return -EOPNOTSUPP; 4409457c1b27SNishanth Aravamudan 441017743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 441117743798SMuchun Song &tmp); 441208d4a246SMichal Hocko if (ret) 441308d4a246SMichal Hocko goto out; 4414e5ff2159SAndi Kleen 4415238d3c13SDavid Rientjes if (write) 4416238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 4417238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 441808d4a246SMichal Hocko out: 441908d4a246SMichal Hocko return ret; 44201da177e4SLinus Torvalds } 4421396faf03SMel Gorman 442206808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 442332927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 442406808b08SLee Schermerhorn { 442506808b08SLee Schermerhorn 442606808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 442706808b08SLee Schermerhorn buffer, length, ppos); 442806808b08SLee Schermerhorn } 442906808b08SLee Schermerhorn 443006808b08SLee Schermerhorn #ifdef CONFIG_NUMA 443106808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 443232927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 443306808b08SLee Schermerhorn { 443406808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 443506808b08SLee Schermerhorn buffer, length, ppos); 443606808b08SLee Schermerhorn } 443706808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 443806808b08SLee Schermerhorn 4439a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 444032927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 4441a3d0c6aaSNishanth Aravamudan { 4442a5516438SAndi Kleen struct hstate *h = &default_hstate; 4443e5ff2159SAndi Kleen unsigned long tmp; 444408d4a246SMichal Hocko int ret; 4445e5ff2159SAndi Kleen 4446457c1b27SNishanth Aravamudan if (!hugepages_supported()) 444786613628SJan Stancek return -EOPNOTSUPP; 4448457c1b27SNishanth Aravamudan 4449e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 4450e5ff2159SAndi Kleen 4451bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 4452adbe8726SEric B Munson return -EINVAL; 4453adbe8726SEric B Munson 445417743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 445517743798SMuchun Song &tmp); 445608d4a246SMichal Hocko if (ret) 445708d4a246SMichal Hocko goto out; 4458e5ff2159SAndi Kleen 4459e5ff2159SAndi Kleen if (write) { 4460db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4461e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 4462db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4463e5ff2159SAndi Kleen } 446408d4a246SMichal Hocko out: 446508d4a246SMichal Hocko return ret; 4466a3d0c6aaSNishanth Aravamudan } 4467a3d0c6aaSNishanth Aravamudan 44681da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 44691da177e4SLinus Torvalds 4470e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 44711da177e4SLinus Torvalds { 4472fcb2b0c5SRoman Gushchin struct hstate *h; 4473fcb2b0c5SRoman Gushchin unsigned long total = 0; 4474fcb2b0c5SRoman Gushchin 4475457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4476457c1b27SNishanth Aravamudan return; 4477fcb2b0c5SRoman Gushchin 4478fcb2b0c5SRoman Gushchin for_each_hstate(h) { 4479fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 4480fcb2b0c5SRoman Gushchin 4481aca78307SMiaohe Lin total += huge_page_size(h) * count; 4482fcb2b0c5SRoman Gushchin 4483fcb2b0c5SRoman Gushchin if (h == &default_hstate) 4484e1759c21SAlexey Dobriyan seq_printf(m, 44851da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 44861da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 4487b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 44887893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 44894f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 4490fcb2b0c5SRoman Gushchin count, 4491a5516438SAndi Kleen h->free_huge_pages, 4492a5516438SAndi Kleen h->resv_huge_pages, 4493a5516438SAndi Kleen h->surplus_huge_pages, 4494aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4495fcb2b0c5SRoman Gushchin } 4496fcb2b0c5SRoman Gushchin 4497aca78307SMiaohe Lin seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); 44981da177e4SLinus Torvalds } 44991da177e4SLinus Torvalds 45007981593bSJoe Perches int hugetlb_report_node_meminfo(char *buf, int len, int nid) 45011da177e4SLinus Torvalds { 4502a5516438SAndi Kleen struct hstate *h = &default_hstate; 45037981593bSJoe Perches 4504457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4505457c1b27SNishanth Aravamudan return 0; 45067981593bSJoe Perches 45077981593bSJoe Perches return sysfs_emit_at(buf, len, 45081da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 4509a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 4510a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 4511a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 4512a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 4513a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 45141da177e4SLinus Torvalds } 45151da177e4SLinus Torvalds 4516dcadcf1cSGang Li void hugetlb_show_meminfo_node(int nid) 4517949f7ec5SDavid Rientjes { 4518949f7ec5SDavid Rientjes struct hstate *h; 4519949f7ec5SDavid Rientjes 4520457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4521457c1b27SNishanth Aravamudan return; 4522457c1b27SNishanth Aravamudan 4523949f7ec5SDavid Rientjes for_each_hstate(h) 4524dcadcf1cSGang Li printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 4525949f7ec5SDavid Rientjes nid, 4526949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 4527949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 4528949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 4529aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4530949f7ec5SDavid Rientjes } 4531949f7ec5SDavid Rientjes 45325d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 45335d317b2bSNaoya Horiguchi { 45345d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 45355d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 45365d317b2bSNaoya Horiguchi } 45375d317b2bSNaoya Horiguchi 45381da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 45391da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 45401da177e4SLinus Torvalds { 4541d0028588SWanpeng Li struct hstate *h; 4542d0028588SWanpeng Li unsigned long nr_total_pages = 0; 4543d0028588SWanpeng Li 4544d0028588SWanpeng Li for_each_hstate(h) 4545d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 4546d0028588SWanpeng Li return nr_total_pages; 45471da177e4SLinus Torvalds } 45481da177e4SLinus Torvalds 4549a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 4550fc1b8a73SMel Gorman { 4551fc1b8a73SMel Gorman int ret = -ENOMEM; 4552fc1b8a73SMel Gorman 45530aa7f354SMiaohe Lin if (!delta) 45540aa7f354SMiaohe Lin return 0; 45550aa7f354SMiaohe Lin 4556db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4557fc1b8a73SMel Gorman /* 4558fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 4559fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 4560fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 4561fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 4562fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 4563fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 4564fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 4565fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 4566fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 4567fc1b8a73SMel Gorman * 4568fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 4569fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 4570fc1b8a73SMel Gorman * we fall back to check against current free page availability as 4571fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 4572fc1b8a73SMel Gorman * semantics that cpuset has. 45738ca39e68SMuchun Song * 45748ca39e68SMuchun Song * Apart from cpuset, we also have memory policy mechanism that 45758ca39e68SMuchun Song * also determines from which node the kernel will allocate memory 45768ca39e68SMuchun Song * in a NUMA system. So similar to cpuset, we also should consider 45778ca39e68SMuchun Song * the memory policy of the current task. Similar to the description 45788ca39e68SMuchun Song * above. 4579fc1b8a73SMel Gorman */ 4580fc1b8a73SMel Gorman if (delta > 0) { 4581a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 4582fc1b8a73SMel Gorman goto out; 4583fc1b8a73SMel Gorman 45848ca39e68SMuchun Song if (delta > allowed_mems_nr(h)) { 4585a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 4586fc1b8a73SMel Gorman goto out; 4587fc1b8a73SMel Gorman } 4588fc1b8a73SMel Gorman } 4589fc1b8a73SMel Gorman 4590fc1b8a73SMel Gorman ret = 0; 4591fc1b8a73SMel Gorman if (delta < 0) 4592a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 4593fc1b8a73SMel Gorman 4594fc1b8a73SMel Gorman out: 4595db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4596fc1b8a73SMel Gorman return ret; 4597fc1b8a73SMel Gorman } 4598fc1b8a73SMel Gorman 459984afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 460084afd99bSAndy Whitcroft { 4601f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 460284afd99bSAndy Whitcroft 460384afd99bSAndy Whitcroft /* 460484afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 460584afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 460684afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 460725985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 460884afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 460984afd99bSAndy Whitcroft * new reference here without additional locking. 461084afd99bSAndy Whitcroft */ 461109a26e83SMike Kravetz if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 461209a26e83SMike Kravetz resv_map_dup_hugetlb_cgroup_uncharge_info(resv); 4613f522c3acSJoonsoo Kim kref_get(&resv->refs); 461484afd99bSAndy Whitcroft } 46158d9bfb26SMike Kravetz 4616131a79b4SMike Kravetz /* 4617131a79b4SMike Kravetz * vma_lock structure for sharable mappings is vma specific. 4618131a79b4SMike Kravetz * Clear old pointer (if copied via vm_area_dup) and create new. 4619131a79b4SMike Kravetz */ 4620131a79b4SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 4621131a79b4SMike Kravetz vma->vm_private_data = NULL; 46228d9bfb26SMike Kravetz hugetlb_vma_lock_alloc(vma); 462309a26e83SMike Kravetz } 4624131a79b4SMike Kravetz } 462584afd99bSAndy Whitcroft 4626a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 4627a1e78772SMel Gorman { 4628a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 46298d9bfb26SMike Kravetz struct resv_map *resv; 463090481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 46314e35f483SJoonsoo Kim unsigned long reserve, start, end; 46321c5ecae3SMike Kravetz long gbl_reserve; 463384afd99bSAndy Whitcroft 46348d9bfb26SMike Kravetz hugetlb_vma_lock_free(vma); 46358d9bfb26SMike Kravetz 46368d9bfb26SMike Kravetz resv = vma_resv_map(vma); 46374e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 46384e35f483SJoonsoo Kim return; 46394e35f483SJoonsoo Kim 4640a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 4641a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 464284afd99bSAndy Whitcroft 46434e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 4644e9fe92aeSMina Almasry hugetlb_cgroup_uncharge_counter(resv, start, end); 46457251ff78SAdam Litke if (reserve) { 46461c5ecae3SMike Kravetz /* 46471c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 46481c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 46491c5ecae3SMike Kravetz */ 46501c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 46511c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 46527251ff78SAdam Litke } 4653e9fe92aeSMina Almasry 4654e9fe92aeSMina Almasry kref_put(&resv->refs, resv_map_release); 4655a1e78772SMel Gorman } 4656a1e78772SMel Gorman 465731383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 465831383c68SDan Williams { 465931383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 466031383c68SDan Williams return -EINVAL; 466131383c68SDan Williams return 0; 466231383c68SDan Williams } 466331383c68SDan Williams 466405ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 466505ea8860SDan Williams { 4666aca78307SMiaohe Lin return huge_page_size(hstate_vma(vma)); 466705ea8860SDan Williams } 466805ea8860SDan Williams 46691da177e4SLinus Torvalds /* 46701da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 46711da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 46726c26d310SMiaohe Lin * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get 46731da177e4SLinus Torvalds * this far. 46741da177e4SLinus Torvalds */ 4675b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 46761da177e4SLinus Torvalds { 46771da177e4SLinus Torvalds BUG(); 4678d0217ac0SNick Piggin return 0; 46791da177e4SLinus Torvalds } 46801da177e4SLinus Torvalds 4681eec3636aSJane Chu /* 4682eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 4683eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 4684eec3636aSJane Chu * This is because under System V memory model, mappings created via 4685eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 4686eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 4687eec3636aSJane Chu */ 4688f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 4689d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 469084afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 4691a1e78772SMel Gorman .close = hugetlb_vm_op_close, 4692dd3b614fSDmitry Safonov .may_split = hugetlb_vm_op_split, 469305ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 46941da177e4SLinus Torvalds }; 46951da177e4SLinus Torvalds 46961e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 46971e8f889bSDavid Gibson int writable) 469863551ae0SDavid Gibson { 469963551ae0SDavid Gibson pte_t entry; 470079c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 470163551ae0SDavid Gibson 47021e8f889bSDavid Gibson if (writable) { 4703106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 4704106c992aSGerald Schaefer vma->vm_page_prot))); 470563551ae0SDavid Gibson } else { 4706106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 4707106c992aSGerald Schaefer vma->vm_page_prot)); 470863551ae0SDavid Gibson } 470963551ae0SDavid Gibson entry = pte_mkyoung(entry); 471079c1c594SChristophe Leroy entry = arch_make_huge_pte(entry, shift, vma->vm_flags); 471163551ae0SDavid Gibson 471263551ae0SDavid Gibson return entry; 471363551ae0SDavid Gibson } 471463551ae0SDavid Gibson 47151e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 47161e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 47171e8f889bSDavid Gibson { 47181e8f889bSDavid Gibson pte_t entry; 47191e8f889bSDavid Gibson 4720106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 472132f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 47224b3073e1SRussell King update_mmu_cache(vma, address, ptep); 47231e8f889bSDavid Gibson } 47241e8f889bSDavid Gibson 4725d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 47264a705fefSNaoya Horiguchi { 47274a705fefSNaoya Horiguchi swp_entry_t swp; 47284a705fefSNaoya Horiguchi 47294a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 4730d5ed7444SAneesh Kumar K.V return false; 47314a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4732d79d176aSBaoquan He if (is_migration_entry(swp)) 4733d5ed7444SAneesh Kumar K.V return true; 47344a705fefSNaoya Horiguchi else 4735d5ed7444SAneesh Kumar K.V return false; 47364a705fefSNaoya Horiguchi } 47374a705fefSNaoya Horiguchi 47383e5c3600SBaoquan He static bool is_hugetlb_entry_hwpoisoned(pte_t pte) 47394a705fefSNaoya Horiguchi { 47404a705fefSNaoya Horiguchi swp_entry_t swp; 47414a705fefSNaoya Horiguchi 47424a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 47433e5c3600SBaoquan He return false; 47444a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4745d79d176aSBaoquan He if (is_hwpoison_entry(swp)) 47463e5c3600SBaoquan He return true; 47474a705fefSNaoya Horiguchi else 47483e5c3600SBaoquan He return false; 47494a705fefSNaoya Horiguchi } 47501e8f889bSDavid Gibson 47514eae4efaSPeter Xu static void 47524eae4efaSPeter Xu hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, 47534eae4efaSPeter Xu struct page *new_page) 47544eae4efaSPeter Xu { 47554eae4efaSPeter Xu __SetPageUptodate(new_page); 47564eae4efaSPeter Xu hugepage_add_new_anon_rmap(new_page, vma, addr); 47571eba86c0SPasha Tatashin set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1)); 47584eae4efaSPeter Xu hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); 47594eae4efaSPeter Xu ClearHPageRestoreReserve(new_page); 47604eae4efaSPeter Xu SetHPageMigratable(new_page); 47614eae4efaSPeter Xu } 47624eae4efaSPeter Xu 476363551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 4764bc70fbf2SPeter Xu struct vm_area_struct *dst_vma, 4765bc70fbf2SPeter Xu struct vm_area_struct *src_vma) 476663551ae0SDavid Gibson { 47673aa4ed80SMiaohe Lin pte_t *src_pte, *dst_pte, entry; 476863551ae0SDavid Gibson struct page *ptepage; 47691c59827dSHugh Dickins unsigned long addr; 4770bc70fbf2SPeter Xu bool cow = is_cow_mapping(src_vma->vm_flags); 4771bc70fbf2SPeter Xu struct hstate *h = hstate_vma(src_vma); 4772a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 47734eae4efaSPeter Xu unsigned long npages = pages_per_huge_page(h); 4774ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 4775e95a9851SMike Kravetz unsigned long last_addr_mask; 4776e8569dd2SAndreas Sandberg int ret = 0; 47771e8f889bSDavid Gibson 4778ac46d4f3SJérôme Glisse if (cow) { 4779bc70fbf2SPeter Xu mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src_vma, src, 4780bc70fbf2SPeter Xu src_vma->vm_start, 4781bc70fbf2SPeter Xu src_vma->vm_end); 4782ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4783623a1ddfSDavid Hildenbrand mmap_assert_write_locked(src); 4784623a1ddfSDavid Hildenbrand raw_write_seqcount_begin(&src->write_protect_seq); 478540549ba8SMike Kravetz } else { 478640549ba8SMike Kravetz /* 478740549ba8SMike Kravetz * For shared mappings the vma lock must be held before 478840549ba8SMike Kravetz * calling huge_pte_offset in the src vma. Otherwise, the 478940549ba8SMike Kravetz * returned ptep could go away if part of a shared pmd and 479040549ba8SMike Kravetz * another thread calls huge_pmd_unshare. 479140549ba8SMike Kravetz */ 479240549ba8SMike Kravetz hugetlb_vma_lock_read(src_vma); 4793ac46d4f3SJérôme Glisse } 4794e8569dd2SAndreas Sandberg 4795e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 4796bc70fbf2SPeter Xu for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) { 4797cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 47987868a208SPunit Agrawal src_pte = huge_pte_offset(src, addr, sz); 4799e95a9851SMike Kravetz if (!src_pte) { 4800e95a9851SMike Kravetz addr |= last_addr_mask; 4801c74df32cSHugh Dickins continue; 4802e95a9851SMike Kravetz } 4803bc70fbf2SPeter Xu dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz); 4804e8569dd2SAndreas Sandberg if (!dst_pte) { 4805e8569dd2SAndreas Sandberg ret = -ENOMEM; 4806e8569dd2SAndreas Sandberg break; 4807e8569dd2SAndreas Sandberg } 4808c5c99429SLarry Woodman 48095e41540cSMike Kravetz /* 48105e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 48115e41540cSMike Kravetz * 48123aa4ed80SMiaohe Lin * dst_pte == src_pte is the common case of src/dest sharing. 48135e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 48143aa4ed80SMiaohe Lin * another vma. So page_count of ptep page is checked instead 48153aa4ed80SMiaohe Lin * to reliably determine whether pte is shared. 48165e41540cSMike Kravetz */ 48173aa4ed80SMiaohe Lin if (page_count(virt_to_page(dst_pte)) > 1) { 4818e95a9851SMike Kravetz addr |= last_addr_mask; 4819c5c99429SLarry Woodman continue; 4820e95a9851SMike Kravetz } 4821c5c99429SLarry Woodman 4822cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 4823cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 4824cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 48254a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 48264eae4efaSPeter Xu again: 48273aa4ed80SMiaohe Lin if (huge_pte_none(entry)) { 48285e41540cSMike Kravetz /* 48293aa4ed80SMiaohe Lin * Skip if src entry none. 48305e41540cSMike Kravetz */ 48314a705fefSNaoya Horiguchi ; 4832c2cb0dccSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) { 4833c2cb0dccSNaoya Horiguchi bool uffd_wp = huge_pte_uffd_wp(entry); 4834c2cb0dccSNaoya Horiguchi 4835c2cb0dccSNaoya Horiguchi if (!userfaultfd_wp(dst_vma) && uffd_wp) 4836c2cb0dccSNaoya Horiguchi entry = huge_pte_clear_uffd_wp(entry); 4837c2cb0dccSNaoya Horiguchi set_huge_pte_at(dst, addr, dst_pte, entry); 4838c2cb0dccSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry))) { 48394a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 4840bc70fbf2SPeter Xu bool uffd_wp = huge_pte_uffd_wp(entry); 48414a705fefSNaoya Horiguchi 48426c287605SDavid Hildenbrand if (!is_readable_migration_entry(swp_entry) && cow) { 48434a705fefSNaoya Horiguchi /* 48444a705fefSNaoya Horiguchi * COW mappings require pages in both 48454a705fefSNaoya Horiguchi * parent and child to be set to read. 48464a705fefSNaoya Horiguchi */ 48474dd845b5SAlistair Popple swp_entry = make_readable_migration_entry( 48484dd845b5SAlistair Popple swp_offset(swp_entry)); 48494a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 4850bc70fbf2SPeter Xu if (userfaultfd_wp(src_vma) && uffd_wp) 4851bc70fbf2SPeter Xu entry = huge_pte_mkuffd_wp(entry); 485218f39629SQi Zheng set_huge_pte_at(src, addr, src_pte, entry); 48534a705fefSNaoya Horiguchi } 4854bc70fbf2SPeter Xu if (!userfaultfd_wp(dst_vma) && uffd_wp) 4855bc70fbf2SPeter Xu entry = huge_pte_clear_uffd_wp(entry); 485618f39629SQi Zheng set_huge_pte_at(dst, addr, dst_pte, entry); 4857bc70fbf2SPeter Xu } else if (unlikely(is_pte_marker(entry))) { 4858bc70fbf2SPeter Xu /* 4859bc70fbf2SPeter Xu * We copy the pte marker only if the dst vma has 4860bc70fbf2SPeter Xu * uffd-wp enabled. 4861bc70fbf2SPeter Xu */ 4862bc70fbf2SPeter Xu if (userfaultfd_wp(dst_vma)) 4863bc70fbf2SPeter Xu set_huge_pte_at(dst, addr, dst_pte, entry); 48644a705fefSNaoya Horiguchi } else { 48654eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 48664eae4efaSPeter Xu ptepage = pte_page(entry); 48674eae4efaSPeter Xu get_page(ptepage); 48684eae4efaSPeter Xu 48694eae4efaSPeter Xu /* 4870fb3d824dSDavid Hildenbrand * Failing to duplicate the anon rmap is a rare case 4871fb3d824dSDavid Hildenbrand * where we see pinned hugetlb pages while they're 4872fb3d824dSDavid Hildenbrand * prone to COW. We need to do the COW earlier during 4873fb3d824dSDavid Hildenbrand * fork. 48744eae4efaSPeter Xu * 48754eae4efaSPeter Xu * When pre-allocating the page or copying data, we 48764eae4efaSPeter Xu * need to be without the pgtable locks since we could 48774eae4efaSPeter Xu * sleep during the process. 48784eae4efaSPeter Xu */ 4879fb3d824dSDavid Hildenbrand if (!PageAnon(ptepage)) { 4880fb3d824dSDavid Hildenbrand page_dup_file_rmap(ptepage, true); 4881bc70fbf2SPeter Xu } else if (page_try_dup_anon_rmap(ptepage, true, 4882bc70fbf2SPeter Xu src_vma)) { 48834eae4efaSPeter Xu pte_t src_pte_old = entry; 48844eae4efaSPeter Xu struct page *new; 48854eae4efaSPeter Xu 48864eae4efaSPeter Xu spin_unlock(src_ptl); 48874eae4efaSPeter Xu spin_unlock(dst_ptl); 48884eae4efaSPeter Xu /* Do not use reserve as it's private owned */ 4889bc70fbf2SPeter Xu new = alloc_huge_page(dst_vma, addr, 1); 48904eae4efaSPeter Xu if (IS_ERR(new)) { 48914eae4efaSPeter Xu put_page(ptepage); 48924eae4efaSPeter Xu ret = PTR_ERR(new); 48934eae4efaSPeter Xu break; 48944eae4efaSPeter Xu } 4895bc70fbf2SPeter Xu copy_user_huge_page(new, ptepage, addr, dst_vma, 48964eae4efaSPeter Xu npages); 48974eae4efaSPeter Xu put_page(ptepage); 48984eae4efaSPeter Xu 48994eae4efaSPeter Xu /* Install the new huge page if src pte stable */ 49004eae4efaSPeter Xu dst_ptl = huge_pte_lock(h, dst, dst_pte); 49014eae4efaSPeter Xu src_ptl = huge_pte_lockptr(h, src, src_pte); 49024eae4efaSPeter Xu spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 49034eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 49044eae4efaSPeter Xu if (!pte_same(src_pte_old, entry)) { 4905bc70fbf2SPeter Xu restore_reserve_on_error(h, dst_vma, addr, 4906846be085SMike Kravetz new); 49074eae4efaSPeter Xu put_page(new); 49083aa4ed80SMiaohe Lin /* huge_ptep of dst_pte won't change as in child */ 49094eae4efaSPeter Xu goto again; 49104eae4efaSPeter Xu } 4911bc70fbf2SPeter Xu hugetlb_install_page(dst_vma, dst_pte, addr, new); 49124eae4efaSPeter Xu spin_unlock(src_ptl); 49134eae4efaSPeter Xu spin_unlock(dst_ptl); 49144eae4efaSPeter Xu continue; 49154eae4efaSPeter Xu } 49164eae4efaSPeter Xu 491734ee645eSJoerg Roedel if (cow) { 49180f10851eSJérôme Glisse /* 49190f10851eSJérôme Glisse * No need to notify as we are downgrading page 49200f10851eSJérôme Glisse * table protection not changing it to point 49210f10851eSJérôme Glisse * to a new page. 49220f10851eSJérôme Glisse * 4923ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 49240f10851eSJérôme Glisse */ 49257f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 492684894e1cSPeter Xu entry = huge_pte_wrprotect(entry); 492734ee645eSJoerg Roedel } 49284eae4efaSPeter Xu 492963551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 49304eae4efaSPeter Xu hugetlb_count_add(npages, dst); 49311c59827dSHugh Dickins } 4932cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 4933cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 493463551ae0SDavid Gibson } 493563551ae0SDavid Gibson 4936623a1ddfSDavid Hildenbrand if (cow) { 4937623a1ddfSDavid Hildenbrand raw_write_seqcount_end(&src->write_protect_seq); 4938ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 493940549ba8SMike Kravetz } else { 494040549ba8SMike Kravetz hugetlb_vma_unlock_read(src_vma); 4941623a1ddfSDavid Hildenbrand } 4942e8569dd2SAndreas Sandberg 4943e8569dd2SAndreas Sandberg return ret; 494463551ae0SDavid Gibson } 494563551ae0SDavid Gibson 4946550a7d60SMina Almasry static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr, 4947db110a99SAneesh Kumar K.V unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte) 4948550a7d60SMina Almasry { 4949550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 4950550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 4951550a7d60SMina Almasry spinlock_t *src_ptl, *dst_ptl; 4952db110a99SAneesh Kumar K.V pte_t pte; 4953550a7d60SMina Almasry 4954550a7d60SMina Almasry dst_ptl = huge_pte_lock(h, mm, dst_pte); 4955550a7d60SMina Almasry src_ptl = huge_pte_lockptr(h, mm, src_pte); 4956550a7d60SMina Almasry 4957550a7d60SMina Almasry /* 4958550a7d60SMina Almasry * We don't have to worry about the ordering of src and dst ptlocks 4959550a7d60SMina Almasry * because exclusive mmap_sem (or the i_mmap_lock) prevents deadlock. 4960550a7d60SMina Almasry */ 4961550a7d60SMina Almasry if (src_ptl != dst_ptl) 4962550a7d60SMina Almasry spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 4963550a7d60SMina Almasry 4964550a7d60SMina Almasry pte = huge_ptep_get_and_clear(mm, old_addr, src_pte); 4965550a7d60SMina Almasry set_huge_pte_at(mm, new_addr, dst_pte, pte); 4966550a7d60SMina Almasry 4967550a7d60SMina Almasry if (src_ptl != dst_ptl) 4968550a7d60SMina Almasry spin_unlock(src_ptl); 4969550a7d60SMina Almasry spin_unlock(dst_ptl); 4970550a7d60SMina Almasry } 4971550a7d60SMina Almasry 4972550a7d60SMina Almasry int move_hugetlb_page_tables(struct vm_area_struct *vma, 4973550a7d60SMina Almasry struct vm_area_struct *new_vma, 4974550a7d60SMina Almasry unsigned long old_addr, unsigned long new_addr, 4975550a7d60SMina Almasry unsigned long len) 4976550a7d60SMina Almasry { 4977550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 4978550a7d60SMina Almasry struct address_space *mapping = vma->vm_file->f_mapping; 4979550a7d60SMina Almasry unsigned long sz = huge_page_size(h); 4980550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 4981550a7d60SMina Almasry unsigned long old_end = old_addr + len; 4982e95a9851SMike Kravetz unsigned long last_addr_mask; 4983550a7d60SMina Almasry pte_t *src_pte, *dst_pte; 4984550a7d60SMina Almasry struct mmu_notifier_range range; 49853d0b95cdSBaolin Wang bool shared_pmd = false; 4986550a7d60SMina Almasry 4987550a7d60SMina Almasry mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, old_addr, 4988550a7d60SMina Almasry old_end); 4989550a7d60SMina Almasry adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 49903d0b95cdSBaolin Wang /* 49913d0b95cdSBaolin Wang * In case of shared PMDs, we should cover the maximum possible 49923d0b95cdSBaolin Wang * range. 49933d0b95cdSBaolin Wang */ 49943d0b95cdSBaolin Wang flush_cache_range(vma, range.start, range.end); 49953d0b95cdSBaolin Wang 4996550a7d60SMina Almasry mmu_notifier_invalidate_range_start(&range); 4997e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 4998550a7d60SMina Almasry /* Prevent race with file truncation */ 499940549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 5000550a7d60SMina Almasry i_mmap_lock_write(mapping); 5001550a7d60SMina Almasry for (; old_addr < old_end; old_addr += sz, new_addr += sz) { 5002550a7d60SMina Almasry src_pte = huge_pte_offset(mm, old_addr, sz); 5003e95a9851SMike Kravetz if (!src_pte) { 5004e95a9851SMike Kravetz old_addr |= last_addr_mask; 5005e95a9851SMike Kravetz new_addr |= last_addr_mask; 5006550a7d60SMina Almasry continue; 5007e95a9851SMike Kravetz } 5008550a7d60SMina Almasry if (huge_pte_none(huge_ptep_get(src_pte))) 5009550a7d60SMina Almasry continue; 5010550a7d60SMina Almasry 50114ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) { 50123d0b95cdSBaolin Wang shared_pmd = true; 50134ddb4d91SMike Kravetz old_addr |= last_addr_mask; 50144ddb4d91SMike Kravetz new_addr |= last_addr_mask; 5015550a7d60SMina Almasry continue; 50163d0b95cdSBaolin Wang } 5017550a7d60SMina Almasry 5018550a7d60SMina Almasry dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz); 5019550a7d60SMina Almasry if (!dst_pte) 5020550a7d60SMina Almasry break; 5021550a7d60SMina Almasry 5022db110a99SAneesh Kumar K.V move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte); 5023550a7d60SMina Almasry } 50243d0b95cdSBaolin Wang 50253d0b95cdSBaolin Wang if (shared_pmd) 50263d0b95cdSBaolin Wang flush_tlb_range(vma, range.start, range.end); 50273d0b95cdSBaolin Wang else 5028550a7d60SMina Almasry flush_tlb_range(vma, old_end - len, old_end); 5029550a7d60SMina Almasry mmu_notifier_invalidate_range_end(&range); 503013e4ad2cSNadav Amit i_mmap_unlock_write(mapping); 503140549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 5032550a7d60SMina Almasry 5033550a7d60SMina Almasry return len + old_addr - old_end; 5034550a7d60SMina Almasry } 5035550a7d60SMina Almasry 503673c54763SPeter Xu static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 503724669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 503805e90bd0SPeter Xu struct page *ref_page, zap_flags_t zap_flags) 503963551ae0SDavid Gibson { 504063551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 504163551ae0SDavid Gibson unsigned long address; 5042c7546f8fSDavid Gibson pte_t *ptep; 504363551ae0SDavid Gibson pte_t pte; 5044cb900f41SKirill A. Shutemov spinlock_t *ptl; 504563551ae0SDavid Gibson struct page *page; 5046a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 5047a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 5048ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 5049e95a9851SMike Kravetz unsigned long last_addr_mask; 5050a4a118f2SNadav Amit bool force_flush = false; 5051a5516438SAndi Kleen 505263551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 5053a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 5054a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 505563551ae0SDavid Gibson 505607e32661SAneesh Kumar K.V /* 505707e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 505807e32661SAneesh Kumar K.V * to huge page. 505907e32661SAneesh Kumar K.V */ 5060ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 506124669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 5062dff11abeSMike Kravetz 5063dff11abeSMike Kravetz /* 5064dff11abeSMike Kravetz * If sharing possible, alert mmu notifiers of worst case. 5065dff11abeSMike Kravetz */ 50666f4f13e8SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, 50676f4f13e8SJérôme Glisse end); 5068ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 5069ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 5070e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 5071569f48b8SHillf Danton address = start; 5072569f48b8SHillf Danton for (; address < end; address += sz) { 50737868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, sz); 5074e95a9851SMike Kravetz if (!ptep) { 5075e95a9851SMike Kravetz address |= last_addr_mask; 5076c7546f8fSDavid Gibson continue; 5077e95a9851SMike Kravetz } 5078c7546f8fSDavid Gibson 5079cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 50804ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, address, ptep)) { 508131d49da5SAneesh Kumar K.V spin_unlock(ptl); 5082a4a118f2SNadav Amit tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE); 5083a4a118f2SNadav Amit force_flush = true; 50844ddb4d91SMike Kravetz address |= last_addr_mask; 508531d49da5SAneesh Kumar K.V continue; 508631d49da5SAneesh Kumar K.V } 508739dde65cSChen, Kenneth W 50886629326bSHillf Danton pte = huge_ptep_get(ptep); 508931d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 509031d49da5SAneesh Kumar K.V spin_unlock(ptl); 509131d49da5SAneesh Kumar K.V continue; 509231d49da5SAneesh Kumar K.V } 50936629326bSHillf Danton 50946629326bSHillf Danton /* 50959fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 50969fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 50976629326bSHillf Danton */ 50989fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 509905e90bd0SPeter Xu /* 510005e90bd0SPeter Xu * If the pte was wr-protected by uffd-wp in any of the 510105e90bd0SPeter Xu * swap forms, meanwhile the caller does not want to 510205e90bd0SPeter Xu * drop the uffd-wp bit in this zap, then replace the 510305e90bd0SPeter Xu * pte with a marker. 510405e90bd0SPeter Xu */ 510505e90bd0SPeter Xu if (pte_swp_uffd_wp_any(pte) && 510605e90bd0SPeter Xu !(zap_flags & ZAP_FLAG_DROP_MARKER)) 510705e90bd0SPeter Xu set_huge_pte_at(mm, address, ptep, 510805e90bd0SPeter Xu make_pte_marker(PTE_MARKER_UFFD_WP)); 510905e90bd0SPeter Xu else 51109386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 511131d49da5SAneesh Kumar K.V spin_unlock(ptl); 511231d49da5SAneesh Kumar K.V continue; 51138c4894c6SNaoya Horiguchi } 51146629326bSHillf Danton 51156629326bSHillf Danton page = pte_page(pte); 511604f2cbe3SMel Gorman /* 511704f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 511804f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 511904f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 512004f2cbe3SMel Gorman */ 512104f2cbe3SMel Gorman if (ref_page) { 512231d49da5SAneesh Kumar K.V if (page != ref_page) { 512331d49da5SAneesh Kumar K.V spin_unlock(ptl); 512431d49da5SAneesh Kumar K.V continue; 512531d49da5SAneesh Kumar K.V } 512604f2cbe3SMel Gorman /* 512704f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 512804f2cbe3SMel Gorman * future faults in this VMA will fail rather than 512904f2cbe3SMel Gorman * looking like data was lost 513004f2cbe3SMel Gorman */ 513104f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 513204f2cbe3SMel Gorman } 513304f2cbe3SMel Gorman 5134c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 5135b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 5136106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 51376649a386SKen Chen set_page_dirty(page); 513805e90bd0SPeter Xu /* Leave a uffd-wp pte marker if needed */ 513905e90bd0SPeter Xu if (huge_pte_uffd_wp(pte) && 514005e90bd0SPeter Xu !(zap_flags & ZAP_FLAG_DROP_MARKER)) 514105e90bd0SPeter Xu set_huge_pte_at(mm, address, ptep, 514205e90bd0SPeter Xu make_pte_marker(PTE_MARKER_UFFD_WP)); 51435d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 5144cea86fe2SHugh Dickins page_remove_rmap(page, vma, true); 514531d49da5SAneesh Kumar K.V 5146cb900f41SKirill A. Shutemov spin_unlock(ptl); 5147e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 514824669e58SAneesh Kumar K.V /* 514931d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 515024669e58SAneesh Kumar K.V */ 515131d49da5SAneesh Kumar K.V if (ref_page) 515231d49da5SAneesh Kumar K.V break; 5153fe1668aeSChen, Kenneth W } 5154ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 515524669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 5156a4a118f2SNadav Amit 5157a4a118f2SNadav Amit /* 5158a4a118f2SNadav Amit * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We 5159a4a118f2SNadav Amit * could defer the flush until now, since by holding i_mmap_rwsem we 5160a4a118f2SNadav Amit * guaranteed that the last refernece would not be dropped. But we must 5161a4a118f2SNadav Amit * do the flushing before we return, as otherwise i_mmap_rwsem will be 5162a4a118f2SNadav Amit * dropped and the last reference to the shared PMDs page might be 5163a4a118f2SNadav Amit * dropped as well. 5164a4a118f2SNadav Amit * 5165a4a118f2SNadav Amit * In theory we could defer the freeing of the PMD pages as well, but 5166a4a118f2SNadav Amit * huge_pmd_unshare() relies on the exact page_count for the PMD page to 5167a4a118f2SNadav Amit * detect sharing, so we cannot defer the release of the page either. 5168a4a118f2SNadav Amit * Instead, do flush now. 5169a4a118f2SNadav Amit */ 5170a4a118f2SNadav Amit if (force_flush) 5171a4a118f2SNadav Amit tlb_flush_mmu_tlbonly(tlb); 51721da177e4SLinus Torvalds } 517363551ae0SDavid Gibson 5174d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 5175d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 517605e90bd0SPeter Xu unsigned long end, struct page *ref_page, 517705e90bd0SPeter Xu zap_flags_t zap_flags) 5178d833352aSMel Gorman { 5179131a79b4SMike Kravetz hugetlb_vma_lock_write(vma); 5180131a79b4SMike Kravetz i_mmap_lock_write(vma->vm_file->f_mapping); 5181131a79b4SMike Kravetz 518205e90bd0SPeter Xu __unmap_hugepage_range(tlb, vma, start, end, ref_page, zap_flags); 5183d833352aSMel Gorman 5184d833352aSMel Gorman /* 5185131a79b4SMike Kravetz * Unlock and free the vma lock before releasing i_mmap_rwsem. When 5186131a79b4SMike Kravetz * the vma_lock is freed, this makes the vma ineligible for pmd 5187131a79b4SMike Kravetz * sharing. And, i_mmap_rwsem is required to set up pmd sharing. 5188131a79b4SMike Kravetz * This is important as page tables for this unmapped range will 5189131a79b4SMike Kravetz * be asynchrously deleted. If the page tables are shared, there 5190131a79b4SMike Kravetz * will be issues when accessed by someone else. 5191d833352aSMel Gorman */ 5192ecfbd733SMike Kravetz __hugetlb_vma_unlock_write_free(vma); 5193131a79b4SMike Kravetz 5194131a79b4SMike Kravetz i_mmap_unlock_write(vma->vm_file->f_mapping); 5195d833352aSMel Gorman } 5196d833352aSMel Gorman 5197502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 519805e90bd0SPeter Xu unsigned long end, struct page *ref_page, 519905e90bd0SPeter Xu zap_flags_t zap_flags) 5200502717f4SChen, Kenneth W { 520124669e58SAneesh Kumar K.V struct mmu_gather tlb; 5202dff11abeSMike Kravetz 5203a72afd87SWill Deacon tlb_gather_mmu(&tlb, vma->vm_mm); 520405e90bd0SPeter Xu __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags); 5205ae8eba8bSWill Deacon tlb_finish_mmu(&tlb); 5206502717f4SChen, Kenneth W } 5207502717f4SChen, Kenneth W 520804f2cbe3SMel Gorman /* 520904f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 5210578b7725SZhiyuan Dai * mapping it owns the reserve page for. The intention is to unmap the page 521104f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 521204f2cbe3SMel Gorman * same region. 521304f2cbe3SMel Gorman */ 52142f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 52152a4b3dedSHarvey Harrison struct page *page, unsigned long address) 521604f2cbe3SMel Gorman { 52177526674dSAdam Litke struct hstate *h = hstate_vma(vma); 521804f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 521904f2cbe3SMel Gorman struct address_space *mapping; 522004f2cbe3SMel Gorman pgoff_t pgoff; 522104f2cbe3SMel Gorman 522204f2cbe3SMel Gorman /* 522304f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 522404f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 522504f2cbe3SMel Gorman */ 52267526674dSAdam Litke address = address & huge_page_mask(h); 522736e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 522836e4f20aSMichal Hocko vma->vm_pgoff; 522993c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 523004f2cbe3SMel Gorman 52314eb2b1dcSMel Gorman /* 52324eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 52334eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 52344eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 52354eb2b1dcSMel Gorman */ 523683cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 52376b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 523804f2cbe3SMel Gorman /* Do not unmap the current VMA */ 523904f2cbe3SMel Gorman if (iter_vma == vma) 524004f2cbe3SMel Gorman continue; 524104f2cbe3SMel Gorman 524204f2cbe3SMel Gorman /* 52432f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 52442f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 52452f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 52462f84a899SMel Gorman */ 52472f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 52482f84a899SMel Gorman continue; 52492f84a899SMel Gorman 52502f84a899SMel Gorman /* 525104f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 525204f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 525304f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 525404f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 525504f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 525604f2cbe3SMel Gorman */ 525704f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 525824669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 525905e90bd0SPeter Xu address + huge_page_size(h), page, 0); 526004f2cbe3SMel Gorman } 526183cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 526204f2cbe3SMel Gorman } 526304f2cbe3SMel Gorman 52640fe6e20bSNaoya Horiguchi /* 5265c89357e2SDavid Hildenbrand * hugetlb_wp() should be called with page lock of the original hugepage held. 5266aa6d2e8cSBaolin Wang * Called with hugetlb_fault_mutex_table held and pte_page locked so we 5267ef009b25SMichal Hocko * cannot race with other handlers or page migration. 5268ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 52690fe6e20bSNaoya Horiguchi */ 5270c89357e2SDavid Hildenbrand static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma, 5271c89357e2SDavid Hildenbrand unsigned long address, pte_t *ptep, unsigned int flags, 5272cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 52731e8f889bSDavid Gibson { 5274c89357e2SDavid Hildenbrand const bool unshare = flags & FAULT_FLAG_UNSHARE; 52753999f52eSAneesh Kumar K.V pte_t pte; 5276a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 52771e8f889bSDavid Gibson struct page *old_page, *new_page; 52782b740303SSouptick Joarder int outside_reserve = 0; 52792b740303SSouptick Joarder vm_fault_t ret = 0; 5280974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 5281ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 52821e8f889bSDavid Gibson 5283c89357e2SDavid Hildenbrand VM_BUG_ON(unshare && (flags & FOLL_WRITE)); 5284c89357e2SDavid Hildenbrand VM_BUG_ON(!unshare && !(flags & FOLL_WRITE)); 5285c89357e2SDavid Hildenbrand 52861d8d1464SDavid Hildenbrand /* 52871d8d1464SDavid Hildenbrand * hugetlb does not support FOLL_FORCE-style write faults that keep the 52881d8d1464SDavid Hildenbrand * PTE mapped R/O such as maybe_mkwrite() would do. 52891d8d1464SDavid Hildenbrand */ 52901d8d1464SDavid Hildenbrand if (WARN_ON_ONCE(!unshare && !(vma->vm_flags & VM_WRITE))) 52911d8d1464SDavid Hildenbrand return VM_FAULT_SIGSEGV; 52921d8d1464SDavid Hildenbrand 52931d8d1464SDavid Hildenbrand /* Let's take out MAP_SHARED mappings first. */ 52941d8d1464SDavid Hildenbrand if (vma->vm_flags & VM_MAYSHARE) { 52951d8d1464SDavid Hildenbrand if (unlikely(unshare)) 52961d8d1464SDavid Hildenbrand return 0; 52971d8d1464SDavid Hildenbrand set_huge_ptep_writable(vma, haddr, ptep); 52981d8d1464SDavid Hildenbrand return 0; 52991d8d1464SDavid Hildenbrand } 53001d8d1464SDavid Hildenbrand 53013999f52eSAneesh Kumar K.V pte = huge_ptep_get(ptep); 53021e8f889bSDavid Gibson old_page = pte_page(pte); 53031e8f889bSDavid Gibson 5304662ce1dcSYang Yang delayacct_wpcopy_start(); 5305662ce1dcSYang Yang 530604f2cbe3SMel Gorman retry_avoidcopy: 5307c89357e2SDavid Hildenbrand /* 5308c89357e2SDavid Hildenbrand * If no-one else is actually using this page, we're the exclusive 5309c89357e2SDavid Hildenbrand * owner and can reuse this page. 5310c89357e2SDavid Hildenbrand */ 531137a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 5312c89357e2SDavid Hildenbrand if (!PageAnonExclusive(old_page)) 53135a49973dSHugh Dickins page_move_anon_rmap(old_page, vma); 5314c89357e2SDavid Hildenbrand if (likely(!unshare)) 53155b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 5316662ce1dcSYang Yang 5317662ce1dcSYang Yang delayacct_wpcopy_end(); 531883c54070SNick Piggin return 0; 53191e8f889bSDavid Gibson } 53206c287605SDavid Hildenbrand VM_BUG_ON_PAGE(PageAnon(old_page) && PageAnonExclusive(old_page), 53216c287605SDavid Hildenbrand old_page); 53221e8f889bSDavid Gibson 532304f2cbe3SMel Gorman /* 532404f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 532504f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 532604f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 532704f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 532804f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 532904f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 533004f2cbe3SMel Gorman * of the full address range. 533104f2cbe3SMel Gorman */ 53325944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 533304f2cbe3SMel Gorman old_page != pagecache_page) 533404f2cbe3SMel Gorman outside_reserve = 1; 533504f2cbe3SMel Gorman 533609cbfeafSKirill A. Shutemov get_page(old_page); 5337b76c8cfbSLarry Woodman 5338ad4404a2SDavidlohr Bueso /* 5339ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 5340ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 5341ad4404a2SDavidlohr Bueso */ 5342cb900f41SKirill A. Shutemov spin_unlock(ptl); 53435b7a1d40SHuang Ying new_page = alloc_huge_page(vma, haddr, outside_reserve); 53441e8f889bSDavid Gibson 53452fc39cecSAdam Litke if (IS_ERR(new_page)) { 534604f2cbe3SMel Gorman /* 534704f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 534804f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 534904f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 535004f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 535104f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 535204f2cbe3SMel Gorman */ 535304f2cbe3SMel Gorman if (outside_reserve) { 535440549ba8SMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 535540549ba8SMike Kravetz pgoff_t idx; 535640549ba8SMike Kravetz u32 hash; 535740549ba8SMike Kravetz 535809cbfeafSKirill A. Shutemov put_page(old_page); 535940549ba8SMike Kravetz /* 536040549ba8SMike Kravetz * Drop hugetlb_fault_mutex and vma_lock before 536140549ba8SMike Kravetz * unmapping. unmapping needs to hold vma_lock 536240549ba8SMike Kravetz * in write mode. Dropping vma_lock in read mode 536340549ba8SMike Kravetz * here is OK as COW mappings do not interact with 536440549ba8SMike Kravetz * PMD sharing. 536540549ba8SMike Kravetz * 536640549ba8SMike Kravetz * Reacquire both after unmap operation. 536740549ba8SMike Kravetz */ 536840549ba8SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 536940549ba8SMike Kravetz hash = hugetlb_fault_mutex_hash(mapping, idx); 537040549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 537140549ba8SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 537240549ba8SMike Kravetz 53735b7a1d40SHuang Ying unmap_ref_private(mm, vma, old_page, haddr); 537440549ba8SMike Kravetz 537540549ba8SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 537640549ba8SMike Kravetz hugetlb_vma_lock_read(vma); 5377cb900f41SKirill A. Shutemov spin_lock(ptl); 53785b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 5379a9af0c5dSNaoya Horiguchi if (likely(ptep && 5380a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 538104f2cbe3SMel Gorman goto retry_avoidcopy; 5382a734bcc8SHillf Danton /* 5383cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 5384cb900f41SKirill A. Shutemov * lock, and our job is done. 5385a734bcc8SHillf Danton */ 5386662ce1dcSYang Yang delayacct_wpcopy_end(); 5387a734bcc8SHillf Danton return 0; 538804f2cbe3SMel Gorman } 538904f2cbe3SMel Gorman 53902b740303SSouptick Joarder ret = vmf_error(PTR_ERR(new_page)); 5391ad4404a2SDavidlohr Bueso goto out_release_old; 53921e8f889bSDavid Gibson } 53931e8f889bSDavid Gibson 53940fe6e20bSNaoya Horiguchi /* 53950fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 53960fe6e20bSNaoya Horiguchi * anon_vma prepared. 53970fe6e20bSNaoya Horiguchi */ 539844e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 5399ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 5400ad4404a2SDavidlohr Bueso goto out_release_all; 540144e2aa93SDean Nelson } 54020fe6e20bSNaoya Horiguchi 5403974e6d66SHuang Ying copy_user_huge_page(new_page, old_page, address, vma, 540447ad8475SAndrea Arcangeli pages_per_huge_page(h)); 54050ed361deSNick Piggin __SetPageUptodate(new_page); 54061e8f889bSDavid Gibson 54077269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, 54086f4f13e8SJérôme Glisse haddr + huge_page_size(h)); 5409ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 5410ad4404a2SDavidlohr Bueso 5411b76c8cfbSLarry Woodman /* 5412cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 5413b76c8cfbSLarry Woodman * before the page tables are altered 5414b76c8cfbSLarry Woodman */ 5415cb900f41SKirill A. Shutemov spin_lock(ptl); 54165b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 5417a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 5418d6995da3SMike Kravetz ClearHPageRestoreReserve(new_page); 541907443a85SJoonsoo Kim 5420c89357e2SDavid Hildenbrand /* Break COW or unshare */ 54215b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 5422ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, range.end); 5423cea86fe2SHugh Dickins page_remove_rmap(old_page, vma, true); 54245b7a1d40SHuang Ying hugepage_add_new_anon_rmap(new_page, vma, haddr); 54251eba86c0SPasha Tatashin set_huge_pte_at(mm, haddr, ptep, 5426c89357e2SDavid Hildenbrand make_huge_pte(vma, new_page, !unshare)); 54278f251a3dSMike Kravetz SetHPageMigratable(new_page); 54281e8f889bSDavid Gibson /* Make the old page be freed below */ 54291e8f889bSDavid Gibson new_page = old_page; 54301e8f889bSDavid Gibson } 5431cb900f41SKirill A. Shutemov spin_unlock(ptl); 5432ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 5433ad4404a2SDavidlohr Bueso out_release_all: 5434c89357e2SDavid Hildenbrand /* 5435c89357e2SDavid Hildenbrand * No restore in case of successful pagetable update (Break COW or 5436c89357e2SDavid Hildenbrand * unshare) 5437c89357e2SDavid Hildenbrand */ 5438c7b1850dSMike Kravetz if (new_page != old_page) 54395b7a1d40SHuang Ying restore_reserve_on_error(h, vma, haddr, new_page); 544009cbfeafSKirill A. Shutemov put_page(new_page); 5441ad4404a2SDavidlohr Bueso out_release_old: 544209cbfeafSKirill A. Shutemov put_page(old_page); 54438312034fSJoonsoo Kim 5444ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 5445662ce1dcSYang Yang 5446662ce1dcSYang Yang delayacct_wpcopy_end(); 5447ad4404a2SDavidlohr Bueso return ret; 54481e8f889bSDavid Gibson } 54491e8f889bSDavid Gibson 54503ae77f43SHugh Dickins /* 54513ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 54523ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 54533ae77f43SHugh Dickins */ 54543ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 54552a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 54562a15efc9SHugh Dickins { 54572a15efc9SHugh Dickins struct address_space *mapping; 54582a15efc9SHugh Dickins pgoff_t idx; 54592a15efc9SHugh Dickins struct page *page; 54602a15efc9SHugh Dickins 54612a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 54622a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 54632a15efc9SHugh Dickins 54642a15efc9SHugh Dickins page = find_get_page(mapping, idx); 54652a15efc9SHugh Dickins if (page) 54662a15efc9SHugh Dickins put_page(page); 54672a15efc9SHugh Dickins return page != NULL; 54682a15efc9SHugh Dickins } 54692a15efc9SHugh Dickins 54707e1813d4SMike Kravetz int hugetlb_add_to_page_cache(struct page *page, struct address_space *mapping, 5471ab76ad54SMike Kravetz pgoff_t idx) 5472ab76ad54SMike Kravetz { 5473d9ef44deSMatthew Wilcox (Oracle) struct folio *folio = page_folio(page); 5474ab76ad54SMike Kravetz struct inode *inode = mapping->host; 5475ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 5476d9ef44deSMatthew Wilcox (Oracle) int err; 5477ab76ad54SMike Kravetz 5478d9ef44deSMatthew Wilcox (Oracle) __folio_set_locked(folio); 5479d9ef44deSMatthew Wilcox (Oracle) err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL); 5480d9ef44deSMatthew Wilcox (Oracle) 5481d9ef44deSMatthew Wilcox (Oracle) if (unlikely(err)) { 5482d9ef44deSMatthew Wilcox (Oracle) __folio_clear_locked(folio); 5483ab76ad54SMike Kravetz return err; 5484d9ef44deSMatthew Wilcox (Oracle) } 5485d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 5486ab76ad54SMike Kravetz 548722146c3cSMike Kravetz /* 5488d9ef44deSMatthew Wilcox (Oracle) * mark folio dirty so that it will not be removed from cache/file 548922146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 549022146c3cSMike Kravetz */ 5491d9ef44deSMatthew Wilcox (Oracle) folio_mark_dirty(folio); 549222146c3cSMike Kravetz 5493ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 5494ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 5495ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 5496ab76ad54SMike Kravetz return 0; 5497ab76ad54SMike Kravetz } 5498ab76ad54SMike Kravetz 54997677f7fdSAxel Rasmussen static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma, 55007677f7fdSAxel Rasmussen struct address_space *mapping, 55017677f7fdSAxel Rasmussen pgoff_t idx, 55027677f7fdSAxel Rasmussen unsigned int flags, 55037677f7fdSAxel Rasmussen unsigned long haddr, 5504824ddc60SNadav Amit unsigned long addr, 55057677f7fdSAxel Rasmussen unsigned long reason) 55067677f7fdSAxel Rasmussen { 55077677f7fdSAxel Rasmussen u32 hash; 55087677f7fdSAxel Rasmussen struct vm_fault vmf = { 55097677f7fdSAxel Rasmussen .vma = vma, 55107677f7fdSAxel Rasmussen .address = haddr, 5511824ddc60SNadav Amit .real_address = addr, 55127677f7fdSAxel Rasmussen .flags = flags, 55137677f7fdSAxel Rasmussen 55147677f7fdSAxel Rasmussen /* 55157677f7fdSAxel Rasmussen * Hard to debug if it ends up being 55167677f7fdSAxel Rasmussen * used by a callee that assumes 55177677f7fdSAxel Rasmussen * something about the other 55187677f7fdSAxel Rasmussen * uninitialized fields... same as in 55197677f7fdSAxel Rasmussen * memory.c 55207677f7fdSAxel Rasmussen */ 55217677f7fdSAxel Rasmussen }; 55227677f7fdSAxel Rasmussen 55237677f7fdSAxel Rasmussen /* 5524958f32ceSLiu Shixin * vma_lock and hugetlb_fault_mutex must be dropped before handling 5525958f32ceSLiu Shixin * userfault. Also mmap_lock could be dropped due to handling 5526958f32ceSLiu Shixin * userfault, any vma operation should be careful from here. 55277677f7fdSAxel Rasmussen */ 552840549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 55297677f7fdSAxel Rasmussen hash = hugetlb_fault_mutex_hash(mapping, idx); 55307677f7fdSAxel Rasmussen mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5531958f32ceSLiu Shixin return handle_userfault(&vmf, reason); 55327677f7fdSAxel Rasmussen } 55337677f7fdSAxel Rasmussen 55342b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 55352b740303SSouptick Joarder struct vm_area_struct *vma, 55368382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 5537c64e912cSPeter Xu unsigned long address, pte_t *ptep, 5538c64e912cSPeter Xu pte_t old_pte, unsigned int flags) 5539ac9b9c66SHugh Dickins { 5540a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 55412b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 5542409eb8c2SHillf Danton int anon_rmap = 0; 55434c887265SAdam Litke unsigned long size; 55444c887265SAdam Litke struct page *page; 55451e8f889bSDavid Gibson pte_t new_pte; 5546cb900f41SKirill A. Shutemov spinlock_t *ptl; 5547285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 5548c7b1850dSMike Kravetz bool new_page, new_pagecache_page = false; 5549958f32ceSLiu Shixin u32 hash = hugetlb_fault_mutex_hash(mapping, idx); 55504c887265SAdam Litke 555104f2cbe3SMel Gorman /* 555204f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 555304f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 5554c89357e2SDavid Hildenbrand * COW/unsharing. Warn that such a situation has occurred as it may not 5555c89357e2SDavid Hildenbrand * be obvious. 555604f2cbe3SMel Gorman */ 555704f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 5558910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 555904f2cbe3SMel Gorman current->pid); 5560958f32ceSLiu Shixin goto out; 556104f2cbe3SMel Gorman } 556204f2cbe3SMel Gorman 55634c887265SAdam Litke /* 5564188a3972SMike Kravetz * Use page lock to guard against racing truncation 5565188a3972SMike Kravetz * before we get page_table_lock. 55664c887265SAdam Litke */ 5567c7b1850dSMike Kravetz new_page = false; 556887bf91d3SMike Kravetz page = find_lock_page(mapping, idx); 556987bf91d3SMike Kravetz if (!page) { 5570188a3972SMike Kravetz size = i_size_read(mapping->host) >> huge_page_shift(h); 5571188a3972SMike Kravetz if (idx >= size) 5572188a3972SMike Kravetz goto out; 55737677f7fdSAxel Rasmussen /* Check for page in userfault range */ 5574958f32ceSLiu Shixin if (userfaultfd_missing(vma)) 5575958f32ceSLiu Shixin return hugetlb_handle_userfault(vma, mapping, idx, 5576824ddc60SNadav Amit flags, haddr, address, 55777677f7fdSAxel Rasmussen VM_UFFD_MISSING); 55781a1aad8aSMike Kravetz 5579285b8dcaSHuang Ying page = alloc_huge_page(vma, haddr, 0); 55802fc39cecSAdam Litke if (IS_ERR(page)) { 55814643d67eSMike Kravetz /* 55824643d67eSMike Kravetz * Returning error will result in faulting task being 55834643d67eSMike Kravetz * sent SIGBUS. The hugetlb fault mutex prevents two 55844643d67eSMike Kravetz * tasks from racing to fault in the same page which 55854643d67eSMike Kravetz * could result in false unable to allocate errors. 55864643d67eSMike Kravetz * Page migration does not take the fault mutex, but 55874643d67eSMike Kravetz * does a clear then write of pte's under page table 55884643d67eSMike Kravetz * lock. Page fault code could race with migration, 55894643d67eSMike Kravetz * notice the clear pte and try to allocate a page 55904643d67eSMike Kravetz * here. Before returning error, get ptl and make 55914643d67eSMike Kravetz * sure there really is no pte entry. 55924643d67eSMike Kravetz */ 55934643d67eSMike Kravetz ptl = huge_pte_lock(h, mm, ptep); 55944643d67eSMike Kravetz ret = 0; 5595d83e6c8aSMiaohe Lin if (huge_pte_none(huge_ptep_get(ptep))) 55962b740303SSouptick Joarder ret = vmf_error(PTR_ERR(page)); 5597d83e6c8aSMiaohe Lin spin_unlock(ptl); 55986bda666aSChristoph Lameter goto out; 55996bda666aSChristoph Lameter } 560047ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 56010ed361deSNick Piggin __SetPageUptodate(page); 5602cb6acd01SMike Kravetz new_page = true; 5603ac9b9c66SHugh Dickins 5604f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 56057e1813d4SMike Kravetz int err = hugetlb_add_to_page_cache(page, mapping, idx); 56066bda666aSChristoph Lameter if (err) { 56073a5497a2SMiaohe Lin /* 56083a5497a2SMiaohe Lin * err can't be -EEXIST which implies someone 56093a5497a2SMiaohe Lin * else consumed the reservation since hugetlb 56103a5497a2SMiaohe Lin * fault mutex is held when add a hugetlb page 56113a5497a2SMiaohe Lin * to the page cache. So it's safe to call 56123a5497a2SMiaohe Lin * restore_reserve_on_error() here. 56133a5497a2SMiaohe Lin */ 56143a5497a2SMiaohe Lin restore_reserve_on_error(h, vma, haddr, page); 56156bda666aSChristoph Lameter put_page(page); 56166bda666aSChristoph Lameter goto out; 56176bda666aSChristoph Lameter } 5618c7b1850dSMike Kravetz new_pagecache_page = true; 561923be7468SMel Gorman } else { 56206bda666aSChristoph Lameter lock_page(page); 56210fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 56220fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 56230fe6e20bSNaoya Horiguchi goto backout_unlocked; 562423be7468SMel Gorman } 5625409eb8c2SHillf Danton anon_rmap = 1; 56260fe6e20bSNaoya Horiguchi } 56270fe6e20bSNaoya Horiguchi } else { 562857303d80SAndy Whitcroft /* 5629998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 5630998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 5631998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 5632fd6a03edSNaoya Horiguchi */ 5633fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 56340eb98f15SMiaohe Lin ret = VM_FAULT_HWPOISON_LARGE | 5635972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 5636fd6a03edSNaoya Horiguchi goto backout_unlocked; 56376bda666aSChristoph Lameter } 56387677f7fdSAxel Rasmussen 56397677f7fdSAxel Rasmussen /* Check for page in userfault range. */ 56407677f7fdSAxel Rasmussen if (userfaultfd_minor(vma)) { 56417677f7fdSAxel Rasmussen unlock_page(page); 56427677f7fdSAxel Rasmussen put_page(page); 5643958f32ceSLiu Shixin return hugetlb_handle_userfault(vma, mapping, idx, 5644824ddc60SNadav Amit flags, haddr, address, 56457677f7fdSAxel Rasmussen VM_UFFD_MINOR); 56467677f7fdSAxel Rasmussen } 5647998b4382SNaoya Horiguchi } 56481e8f889bSDavid Gibson 564957303d80SAndy Whitcroft /* 565057303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 565157303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 565257303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 565357303d80SAndy Whitcroft * the spinlock. 565457303d80SAndy Whitcroft */ 56555e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 5656285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 56572b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 56582b26736cSAndy Whitcroft goto backout_unlocked; 56592b26736cSAndy Whitcroft } 56605e911373SMike Kravetz /* Just decrements count, does not deallocate */ 5661285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 56625e911373SMike Kravetz } 566357303d80SAndy Whitcroft 56648bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 566583c54070SNick Piggin ret = 0; 5666c64e912cSPeter Xu /* If pte changed from under us, retry */ 5667c64e912cSPeter Xu if (!pte_same(huge_ptep_get(ptep), old_pte)) 56684c887265SAdam Litke goto backout; 56694c887265SAdam Litke 567007443a85SJoonsoo Kim if (anon_rmap) { 5671d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 5672285b8dcaSHuang Ying hugepage_add_new_anon_rmap(page, vma, haddr); 5673ac714904SChoi Gi-yong } else 5674fb3d824dSDavid Hildenbrand page_dup_file_rmap(page, true); 56751e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 56761e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 5677c64e912cSPeter Xu /* 5678c64e912cSPeter Xu * If this pte was previously wr-protected, keep it wr-protected even 5679c64e912cSPeter Xu * if populated. 5680c64e912cSPeter Xu */ 5681c64e912cSPeter Xu if (unlikely(pte_marker_uffd_wp(old_pte))) 5682c64e912cSPeter Xu new_pte = huge_pte_wrprotect(huge_pte_mkuffd_wp(new_pte)); 5683285b8dcaSHuang Ying set_huge_pte_at(mm, haddr, ptep, new_pte); 56841e8f889bSDavid Gibson 56855d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 5686788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 56871e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 5688c89357e2SDavid Hildenbrand ret = hugetlb_wp(mm, vma, address, ptep, flags, page, ptl); 56891e8f889bSDavid Gibson } 56901e8f889bSDavid Gibson 5691cb900f41SKirill A. Shutemov spin_unlock(ptl); 5692cb6acd01SMike Kravetz 5693cb6acd01SMike Kravetz /* 56948f251a3dSMike Kravetz * Only set HPageMigratable in newly allocated pages. Existing pages 56958f251a3dSMike Kravetz * found in the pagecache may not have HPageMigratableset if they have 56968f251a3dSMike Kravetz * been isolated for migration. 5697cb6acd01SMike Kravetz */ 5698cb6acd01SMike Kravetz if (new_page) 56998f251a3dSMike Kravetz SetHPageMigratable(page); 5700cb6acd01SMike Kravetz 57014c887265SAdam Litke unlock_page(page); 57024c887265SAdam Litke out: 5703958f32ceSLiu Shixin hugetlb_vma_unlock_read(vma); 5704958f32ceSLiu Shixin mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5705ac9b9c66SHugh Dickins return ret; 57064c887265SAdam Litke 57074c887265SAdam Litke backout: 5708cb900f41SKirill A. Shutemov spin_unlock(ptl); 57092b26736cSAndy Whitcroft backout_unlocked: 5710c7b1850dSMike Kravetz if (new_page && !new_pagecache_page) 5711285b8dcaSHuang Ying restore_reserve_on_error(h, vma, haddr, page); 5712fa27759aSMike Kravetz 5713fa27759aSMike Kravetz unlock_page(page); 57144c887265SAdam Litke put_page(page); 57154c887265SAdam Litke goto out; 5716ac9b9c66SHugh Dickins } 5717ac9b9c66SHugh Dickins 57188382d914SDavidlohr Bueso #ifdef CONFIG_SMP 5719188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 57208382d914SDavidlohr Bueso { 57218382d914SDavidlohr Bueso unsigned long key[2]; 57228382d914SDavidlohr Bueso u32 hash; 57238382d914SDavidlohr Bueso 57248382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 57258382d914SDavidlohr Bueso key[1] = idx; 57268382d914SDavidlohr Bueso 572755254636SMike Kravetz hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); 57288382d914SDavidlohr Bueso 57298382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 57308382d914SDavidlohr Bueso } 57318382d914SDavidlohr Bueso #else 57328382d914SDavidlohr Bueso /* 57336c26d310SMiaohe Lin * For uniprocessor systems we always use a single mutex, so just 57348382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 57358382d914SDavidlohr Bueso */ 5736188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 57378382d914SDavidlohr Bueso { 57388382d914SDavidlohr Bueso return 0; 57398382d914SDavidlohr Bueso } 57408382d914SDavidlohr Bueso #endif 57418382d914SDavidlohr Bueso 57422b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 5743788c7df4SHugh Dickins unsigned long address, unsigned int flags) 574486e5216fSAdam Litke { 57458382d914SDavidlohr Bueso pte_t *ptep, entry; 5746cb900f41SKirill A. Shutemov spinlock_t *ptl; 57472b740303SSouptick Joarder vm_fault_t ret; 57488382d914SDavidlohr Bueso u32 hash; 57498382d914SDavidlohr Bueso pgoff_t idx; 57500fe6e20bSNaoya Horiguchi struct page *page = NULL; 575157303d80SAndy Whitcroft struct page *pagecache_page = NULL; 5752a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 57538382d914SDavidlohr Bueso struct address_space *mapping; 57540f792cf9SNaoya Horiguchi int need_wait_lock = 0; 5755285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 575686e5216fSAdam Litke 5757285b8dcaSHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 5758fd6a03edSNaoya Horiguchi if (ptep) { 575940549ba8SMike Kravetz /* 576040549ba8SMike Kravetz * Since we hold no locks, ptep could be stale. That is 576140549ba8SMike Kravetz * OK as we are only making decisions based on content and 576240549ba8SMike Kravetz * not actually modifying content here. 576340549ba8SMike Kravetz */ 5764fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 5765290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 5766ad1ac596SMiaohe Lin migration_entry_wait_huge(vma, ptep); 5767290408d4SNaoya Horiguchi return 0; 5768290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 5769aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 5770972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 5771c0d0381aSMike Kravetz } 5772ddeaab32SMike Kravetz 57733935baa9SDavid Gibson /* 57743935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 57753935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 57763935baa9SDavid Gibson * the same page in the page cache. 57773935baa9SDavid Gibson */ 577840549ba8SMike Kravetz mapping = vma->vm_file->f_mapping; 577940549ba8SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 5780188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 5781c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 57828382d914SDavidlohr Bueso 578340549ba8SMike Kravetz /* 578440549ba8SMike Kravetz * Acquire vma lock before calling huge_pte_alloc and hold 578540549ba8SMike Kravetz * until finished with ptep. This prevents huge_pmd_unshare from 578640549ba8SMike Kravetz * being called elsewhere and making the ptep no longer valid. 578740549ba8SMike Kravetz * 578840549ba8SMike Kravetz * ptep could have already be assigned via huge_pte_offset. That 578940549ba8SMike Kravetz * is OK, as huge_pte_alloc will return the same value unless 579040549ba8SMike Kravetz * something has changed. 579140549ba8SMike Kravetz */ 579240549ba8SMike Kravetz hugetlb_vma_lock_read(vma); 579340549ba8SMike Kravetz ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); 579440549ba8SMike Kravetz if (!ptep) { 579540549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 579640549ba8SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 579740549ba8SMike Kravetz return VM_FAULT_OOM; 579840549ba8SMike Kravetz } 579940549ba8SMike Kravetz 58007f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 5801c64e912cSPeter Xu /* PTE markers should be handled the same way as none pte */ 5802958f32ceSLiu Shixin if (huge_pte_none_mostly(entry)) 5803958f32ceSLiu Shixin /* 5804958f32ceSLiu Shixin * hugetlb_no_page will drop vma lock and hugetlb fault 5805958f32ceSLiu Shixin * mutex internally, which make us return immediately. 5806958f32ceSLiu Shixin */ 5807958f32ceSLiu Shixin return hugetlb_no_page(mm, vma, mapping, idx, address, ptep, 5808c64e912cSPeter Xu entry, flags); 580986e5216fSAdam Litke 581083c54070SNick Piggin ret = 0; 58111e8f889bSDavid Gibson 581257303d80SAndy Whitcroft /* 58130f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 58140f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 58157c8de358SEthon Paul * an active hugepage in pagecache. This goto expects the 2nd page 58167c8de358SEthon Paul * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will 58177c8de358SEthon Paul * properly handle it. 58180f792cf9SNaoya Horiguchi */ 58190f792cf9SNaoya Horiguchi if (!pte_present(entry)) 58200f792cf9SNaoya Horiguchi goto out_mutex; 58210f792cf9SNaoya Horiguchi 58220f792cf9SNaoya Horiguchi /* 5823c89357e2SDavid Hildenbrand * If we are going to COW/unshare the mapping later, we examine the 5824c89357e2SDavid Hildenbrand * pending reservations for this page now. This will ensure that any 582557303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 58261d8d1464SDavid Hildenbrand * spinlock. Also lookup the pagecache page now as it is used to 58271d8d1464SDavid Hildenbrand * determine if a reservation has been consumed. 582857303d80SAndy Whitcroft */ 5829c89357e2SDavid Hildenbrand if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) && 58301d8d1464SDavid Hildenbrand !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(entry)) { 5831285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 58322b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 5833b4d1d99fSDavid Gibson goto out_mutex; 58342b26736cSAndy Whitcroft } 58355e911373SMike Kravetz /* Just decrements count, does not deallocate */ 5836285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 583757303d80SAndy Whitcroft 583829be8426SMiaohe Lin pagecache_page = find_lock_page(mapping, idx); 583957303d80SAndy Whitcroft } 584057303d80SAndy Whitcroft 58410f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 58420fe6e20bSNaoya Horiguchi 5843c89357e2SDavid Hildenbrand /* Check for a racing update before calling hugetlb_wp() */ 5844b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 5845cb900f41SKirill A. Shutemov goto out_ptl; 5846b4d1d99fSDavid Gibson 5847166f3eccSPeter Xu /* Handle userfault-wp first, before trying to lock more pages */ 5848166f3eccSPeter Xu if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) && 5849166f3eccSPeter Xu (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 5850166f3eccSPeter Xu struct vm_fault vmf = { 5851166f3eccSPeter Xu .vma = vma, 5852166f3eccSPeter Xu .address = haddr, 5853166f3eccSPeter Xu .real_address = address, 5854166f3eccSPeter Xu .flags = flags, 5855166f3eccSPeter Xu }; 5856166f3eccSPeter Xu 5857166f3eccSPeter Xu spin_unlock(ptl); 5858166f3eccSPeter Xu if (pagecache_page) { 5859166f3eccSPeter Xu unlock_page(pagecache_page); 5860166f3eccSPeter Xu put_page(pagecache_page); 5861166f3eccSPeter Xu } 586240549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 5863166f3eccSPeter Xu mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5864166f3eccSPeter Xu return handle_userfault(&vmf, VM_UFFD_WP); 5865166f3eccSPeter Xu } 5866166f3eccSPeter Xu 58670f792cf9SNaoya Horiguchi /* 5868c89357e2SDavid Hildenbrand * hugetlb_wp() requires page locks of pte_page(entry) and 58690f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 58700f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 58710f792cf9SNaoya Horiguchi */ 58720f792cf9SNaoya Horiguchi page = pte_page(entry); 58730f792cf9SNaoya Horiguchi if (page != pagecache_page) 58740f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 58750f792cf9SNaoya Horiguchi need_wait_lock = 1; 58760f792cf9SNaoya Horiguchi goto out_ptl; 58770f792cf9SNaoya Horiguchi } 58780f792cf9SNaoya Horiguchi 58790f792cf9SNaoya Horiguchi get_page(page); 5880b4d1d99fSDavid Gibson 5881c89357e2SDavid Hildenbrand if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) { 5882106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 5883c89357e2SDavid Hildenbrand ret = hugetlb_wp(mm, vma, address, ptep, flags, 5884cb900f41SKirill A. Shutemov pagecache_page, ptl); 58850f792cf9SNaoya Horiguchi goto out_put_page; 5886c89357e2SDavid Hildenbrand } else if (likely(flags & FAULT_FLAG_WRITE)) { 5887106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 5888b4d1d99fSDavid Gibson } 5889c89357e2SDavid Hildenbrand } 5890b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 5891285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 5892788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 5893285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 58940f792cf9SNaoya Horiguchi out_put_page: 58950f792cf9SNaoya Horiguchi if (page != pagecache_page) 58960f792cf9SNaoya Horiguchi unlock_page(page); 58970f792cf9SNaoya Horiguchi put_page(page); 5898cb900f41SKirill A. Shutemov out_ptl: 5899cb900f41SKirill A. Shutemov spin_unlock(ptl); 590057303d80SAndy Whitcroft 590157303d80SAndy Whitcroft if (pagecache_page) { 590257303d80SAndy Whitcroft unlock_page(pagecache_page); 590357303d80SAndy Whitcroft put_page(pagecache_page); 590457303d80SAndy Whitcroft } 5905b4d1d99fSDavid Gibson out_mutex: 590640549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 5907c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 59080f792cf9SNaoya Horiguchi /* 59090f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 59100f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 59110f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 59120f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 59130f792cf9SNaoya Horiguchi * here without taking refcount. 59140f792cf9SNaoya Horiguchi */ 59150f792cf9SNaoya Horiguchi if (need_wait_lock) 59160f792cf9SNaoya Horiguchi wait_on_page_locked(page); 59171e8f889bSDavid Gibson return ret; 591886e5216fSAdam Litke } 591986e5216fSAdam Litke 5920714c1891SAxel Rasmussen #ifdef CONFIG_USERFAULTFD 59218fb5debcSMike Kravetz /* 59228fb5debcSMike Kravetz * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with 59238fb5debcSMike Kravetz * modifications for huge pages. 59248fb5debcSMike Kravetz */ 59258fb5debcSMike Kravetz int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, 59268fb5debcSMike Kravetz pte_t *dst_pte, 59278fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 59288fb5debcSMike Kravetz unsigned long dst_addr, 59298fb5debcSMike Kravetz unsigned long src_addr, 5930f6191471SAxel Rasmussen enum mcopy_atomic_mode mode, 59316041c691SPeter Xu struct page **pagep, 59326041c691SPeter Xu bool wp_copy) 59338fb5debcSMike Kravetz { 5934f6191471SAxel Rasmussen bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE); 59358cc5fcbbSMina Almasry struct hstate *h = hstate_vma(dst_vma); 59368cc5fcbbSMina Almasry struct address_space *mapping = dst_vma->vm_file->f_mapping; 59378cc5fcbbSMina Almasry pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr); 59381e392147SAndrea Arcangeli unsigned long size; 59391c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 59408fb5debcSMike Kravetz pte_t _dst_pte; 59418fb5debcSMike Kravetz spinlock_t *ptl; 59428cc5fcbbSMina Almasry int ret = -ENOMEM; 59438fb5debcSMike Kravetz struct page *page; 5944f6191471SAxel Rasmussen int writable; 5945cc30042dSMina Almasry bool page_in_pagecache = false; 59468fb5debcSMike Kravetz 5947f6191471SAxel Rasmussen if (is_continue) { 5948f6191471SAxel Rasmussen ret = -EFAULT; 5949f6191471SAxel Rasmussen page = find_lock_page(mapping, idx); 5950f6191471SAxel Rasmussen if (!page) 5951f6191471SAxel Rasmussen goto out; 5952cc30042dSMina Almasry page_in_pagecache = true; 5953f6191471SAxel Rasmussen } else if (!*pagep) { 5954d84cf06eSMina Almasry /* If a page already exists, then it's UFFDIO_COPY for 5955d84cf06eSMina Almasry * a non-missing case. Return -EEXIST. 5956d84cf06eSMina Almasry */ 5957d84cf06eSMina Almasry if (vm_shared && 5958d84cf06eSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 5959d84cf06eSMina Almasry ret = -EEXIST; 59608fb5debcSMike Kravetz goto out; 5961d84cf06eSMina Almasry } 5962d84cf06eSMina Almasry 5963d84cf06eSMina Almasry page = alloc_huge_page(dst_vma, dst_addr, 0); 5964d84cf06eSMina Almasry if (IS_ERR(page)) { 5965d84cf06eSMina Almasry ret = -ENOMEM; 5966d84cf06eSMina Almasry goto out; 5967d84cf06eSMina Almasry } 59688fb5debcSMike Kravetz 59698fb5debcSMike Kravetz ret = copy_huge_page_from_user(page, 59708fb5debcSMike Kravetz (const void __user *) src_addr, 5971810a56b9SMike Kravetz pages_per_huge_page(h), false); 59728fb5debcSMike Kravetz 5973c1e8d7c6SMichel Lespinasse /* fallback to copy_from_user outside mmap_lock */ 59748fb5debcSMike Kravetz if (unlikely(ret)) { 59759e368259SAndrea Arcangeli ret = -ENOENT; 59768cc5fcbbSMina Almasry /* Free the allocated page which may have 59778cc5fcbbSMina Almasry * consumed a reservation. 59788cc5fcbbSMina Almasry */ 59798cc5fcbbSMina Almasry restore_reserve_on_error(h, dst_vma, dst_addr, page); 59808cc5fcbbSMina Almasry put_page(page); 59818cc5fcbbSMina Almasry 59828cc5fcbbSMina Almasry /* Allocate a temporary page to hold the copied 59838cc5fcbbSMina Almasry * contents. 59848cc5fcbbSMina Almasry */ 59858cc5fcbbSMina Almasry page = alloc_huge_page_vma(h, dst_vma, dst_addr); 59868cc5fcbbSMina Almasry if (!page) { 59878cc5fcbbSMina Almasry ret = -ENOMEM; 59888cc5fcbbSMina Almasry goto out; 59898cc5fcbbSMina Almasry } 59908fb5debcSMike Kravetz *pagep = page; 59918cc5fcbbSMina Almasry /* Set the outparam pagep and return to the caller to 59928cc5fcbbSMina Almasry * copy the contents outside the lock. Don't free the 59938cc5fcbbSMina Almasry * page. 59948cc5fcbbSMina Almasry */ 59958fb5debcSMike Kravetz goto out; 59968fb5debcSMike Kravetz } 59978fb5debcSMike Kravetz } else { 59988cc5fcbbSMina Almasry if (vm_shared && 59998cc5fcbbSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 60008cc5fcbbSMina Almasry put_page(*pagep); 60018cc5fcbbSMina Almasry ret = -EEXIST; 60028cc5fcbbSMina Almasry *pagep = NULL; 60038cc5fcbbSMina Almasry goto out; 60048cc5fcbbSMina Almasry } 60058cc5fcbbSMina Almasry 60068cc5fcbbSMina Almasry page = alloc_huge_page(dst_vma, dst_addr, 0); 60078cc5fcbbSMina Almasry if (IS_ERR(page)) { 6008da9a298fSMiaohe Lin put_page(*pagep); 60098cc5fcbbSMina Almasry ret = -ENOMEM; 60108cc5fcbbSMina Almasry *pagep = NULL; 60118cc5fcbbSMina Almasry goto out; 60128cc5fcbbSMina Almasry } 601334892366SMuchun Song copy_user_huge_page(page, *pagep, dst_addr, dst_vma, 601434892366SMuchun Song pages_per_huge_page(h)); 60158cc5fcbbSMina Almasry put_page(*pagep); 60168fb5debcSMike Kravetz *pagep = NULL; 60178fb5debcSMike Kravetz } 60188fb5debcSMike Kravetz 60198fb5debcSMike Kravetz /* 60208fb5debcSMike Kravetz * The memory barrier inside __SetPageUptodate makes sure that 60218fb5debcSMike Kravetz * preceding stores to the page contents become visible before 60228fb5debcSMike Kravetz * the set_pte_at() write. 60238fb5debcSMike Kravetz */ 60248fb5debcSMike Kravetz __SetPageUptodate(page); 60258fb5debcSMike Kravetz 6026f6191471SAxel Rasmussen /* Add shared, newly allocated pages to the page cache. */ 6027f6191471SAxel Rasmussen if (vm_shared && !is_continue) { 60281e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 60291e392147SAndrea Arcangeli ret = -EFAULT; 60301e392147SAndrea Arcangeli if (idx >= size) 60311e392147SAndrea Arcangeli goto out_release_nounlock; 60321c9e8defSMike Kravetz 60331e392147SAndrea Arcangeli /* 60341e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 60357e1813d4SMike Kravetz * hugetlb_add_to_page_cache() below happens through the 60361e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 60371e392147SAndrea Arcangeli * the caller. 60381e392147SAndrea Arcangeli */ 60397e1813d4SMike Kravetz ret = hugetlb_add_to_page_cache(page, mapping, idx); 60401c9e8defSMike Kravetz if (ret) 60411c9e8defSMike Kravetz goto out_release_nounlock; 6042cc30042dSMina Almasry page_in_pagecache = true; 60431c9e8defSMike Kravetz } 60441c9e8defSMike Kravetz 6045bcc66543SMiaohe Lin ptl = huge_pte_lock(h, dst_mm, dst_pte); 60468fb5debcSMike Kravetz 60471e392147SAndrea Arcangeli /* 60486041c691SPeter Xu * We allow to overwrite a pte marker: consider when both MISSING|WP 60496041c691SPeter Xu * registered, we firstly wr-protect a none pte which has no page cache 60506041c691SPeter Xu * page backing it, then access the page. 60516041c691SPeter Xu */ 6052fa27759aSMike Kravetz ret = -EEXIST; 60536041c691SPeter Xu if (!huge_pte_none_mostly(huge_ptep_get(dst_pte))) 60548fb5debcSMike Kravetz goto out_release_unlock; 60558fb5debcSMike Kravetz 6056ab74ef70SMiaohe Lin if (page_in_pagecache) { 6057fb3d824dSDavid Hildenbrand page_dup_file_rmap(page, true); 60581c9e8defSMike Kravetz } else { 6059d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 60608fb5debcSMike Kravetz hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); 60611c9e8defSMike Kravetz } 60628fb5debcSMike Kravetz 60636041c691SPeter Xu /* 60646041c691SPeter Xu * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY 60656041c691SPeter Xu * with wp flag set, don't set pte write bit. 60666041c691SPeter Xu */ 60676041c691SPeter Xu if (wp_copy || (is_continue && !vm_shared)) 6068f6191471SAxel Rasmussen writable = 0; 6069f6191471SAxel Rasmussen else 6070f6191471SAxel Rasmussen writable = dst_vma->vm_flags & VM_WRITE; 6071f6191471SAxel Rasmussen 6072f6191471SAxel Rasmussen _dst_pte = make_huge_pte(dst_vma, page, writable); 60736041c691SPeter Xu /* 60746041c691SPeter Xu * Always mark UFFDIO_COPY page dirty; note that this may not be 60756041c691SPeter Xu * extremely important for hugetlbfs for now since swapping is not 60766041c691SPeter Xu * supported, but we should still be clear in that this page cannot be 60776041c691SPeter Xu * thrown away at will, even if write bit not set. 60786041c691SPeter Xu */ 60798fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 60808fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 60818fb5debcSMike Kravetz 60826041c691SPeter Xu if (wp_copy) 60836041c691SPeter Xu _dst_pte = huge_pte_mkuffd_wp(_dst_pte); 60846041c691SPeter Xu 60858fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 60868fb5debcSMike Kravetz 60878fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 60888fb5debcSMike Kravetz 60898fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 60908fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 60918fb5debcSMike Kravetz 60928fb5debcSMike Kravetz spin_unlock(ptl); 6093f6191471SAxel Rasmussen if (!is_continue) 60948f251a3dSMike Kravetz SetHPageMigratable(page); 6095f6191471SAxel Rasmussen if (vm_shared || is_continue) 60961c9e8defSMike Kravetz unlock_page(page); 60978fb5debcSMike Kravetz ret = 0; 60988fb5debcSMike Kravetz out: 60998fb5debcSMike Kravetz return ret; 61008fb5debcSMike Kravetz out_release_unlock: 61018fb5debcSMike Kravetz spin_unlock(ptl); 6102f6191471SAxel Rasmussen if (vm_shared || is_continue) 61031c9e8defSMike Kravetz unlock_page(page); 61045af10dfdSAndrea Arcangeli out_release_nounlock: 6105cc30042dSMina Almasry if (!page_in_pagecache) 6106846be085SMike Kravetz restore_reserve_on_error(h, dst_vma, dst_addr, page); 61078fb5debcSMike Kravetz put_page(page); 61088fb5debcSMike Kravetz goto out; 61098fb5debcSMike Kravetz } 6110714c1891SAxel Rasmussen #endif /* CONFIG_USERFAULTFD */ 61118fb5debcSMike Kravetz 611282e5d378SJoao Martins static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma, 611382e5d378SJoao Martins int refs, struct page **pages, 611482e5d378SJoao Martins struct vm_area_struct **vmas) 611582e5d378SJoao Martins { 611682e5d378SJoao Martins int nr; 611782e5d378SJoao Martins 611882e5d378SJoao Martins for (nr = 0; nr < refs; nr++) { 611982e5d378SJoao Martins if (likely(pages)) 612014455eabSCheng Li pages[nr] = nth_page(page, nr); 612182e5d378SJoao Martins if (vmas) 612282e5d378SJoao Martins vmas[nr] = vma; 612382e5d378SJoao Martins } 612482e5d378SJoao Martins } 612582e5d378SJoao Martins 6126a7f22660SDavid Hildenbrand static inline bool __follow_hugetlb_must_fault(unsigned int flags, pte_t *pte, 6127a7f22660SDavid Hildenbrand bool *unshare) 6128a7f22660SDavid Hildenbrand { 6129a7f22660SDavid Hildenbrand pte_t pteval = huge_ptep_get(pte); 6130a7f22660SDavid Hildenbrand 6131a7f22660SDavid Hildenbrand *unshare = false; 6132a7f22660SDavid Hildenbrand if (is_swap_pte(pteval)) 6133a7f22660SDavid Hildenbrand return true; 6134a7f22660SDavid Hildenbrand if (huge_pte_write(pteval)) 6135a7f22660SDavid Hildenbrand return false; 6136a7f22660SDavid Hildenbrand if (flags & FOLL_WRITE) 6137a7f22660SDavid Hildenbrand return true; 6138a7f22660SDavid Hildenbrand if (gup_must_unshare(flags, pte_page(pteval))) { 6139a7f22660SDavid Hildenbrand *unshare = true; 6140a7f22660SDavid Hildenbrand return true; 6141a7f22660SDavid Hildenbrand } 6142a7f22660SDavid Hildenbrand return false; 6143a7f22660SDavid Hildenbrand } 6144a7f22660SDavid Hildenbrand 614528a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 614663551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 614728a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 61484f6da934SPeter Xu long i, unsigned int flags, int *locked) 614963551ae0SDavid Gibson { 6150d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 6151d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 615228a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 6153a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 61540fa5bc40SJoao Martins int err = -EFAULT, refs; 615563551ae0SDavid Gibson 615663551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 615763551ae0SDavid Gibson pte_t *pte; 6158cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 6159a7f22660SDavid Hildenbrand bool unshare = false; 61602a15efc9SHugh Dickins int absent; 616163551ae0SDavid Gibson struct page *page; 616263551ae0SDavid Gibson 61634c887265SAdam Litke /* 616402057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 616502057967SDavid Rientjes * potentially allocating memory. 616602057967SDavid Rientjes */ 6167fa45f116SDavidlohr Bueso if (fatal_signal_pending(current)) { 616802057967SDavid Rientjes remainder = 0; 616902057967SDavid Rientjes break; 617002057967SDavid Rientjes } 617102057967SDavid Rientjes 617202057967SDavid Rientjes /* 61734c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 61742a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 61754c887265SAdam Litke * first, for the page indexing below to work. 6176cb900f41SKirill A. Shutemov * 6177cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 61784c887265SAdam Litke */ 61797868a208SPunit Agrawal pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), 61807868a208SPunit Agrawal huge_page_size(h)); 6181cb900f41SKirill A. Shutemov if (pte) 6182cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 61832a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 618463551ae0SDavid Gibson 61852a15efc9SHugh Dickins /* 61862a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 61873ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 61883ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 61893ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 61903ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 61912a15efc9SHugh Dickins */ 61923ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 61933ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 6194cb900f41SKirill A. Shutemov if (pte) 6195cb900f41SKirill A. Shutemov spin_unlock(ptl); 61962a15efc9SHugh Dickins remainder = 0; 61972a15efc9SHugh Dickins break; 61982a15efc9SHugh Dickins } 61992a15efc9SHugh Dickins 62009cc3a5bdSNaoya Horiguchi /* 62019cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 62029cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 62039cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 62049cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 62059cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 62069cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 62079cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 62089cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 62099cc3a5bdSNaoya Horiguchi */ 6210a7f22660SDavid Hildenbrand if (absent || 6211a7f22660SDavid Hildenbrand __follow_hugetlb_must_fault(flags, pte, &unshare)) { 62122b740303SSouptick Joarder vm_fault_t ret; 621387ffc118SAndrea Arcangeli unsigned int fault_flags = 0; 62144c887265SAdam Litke 6215cb900f41SKirill A. Shutemov if (pte) 6216cb900f41SKirill A. Shutemov spin_unlock(ptl); 621787ffc118SAndrea Arcangeli if (flags & FOLL_WRITE) 621887ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_WRITE; 6219a7f22660SDavid Hildenbrand else if (unshare) 6220a7f22660SDavid Hildenbrand fault_flags |= FAULT_FLAG_UNSHARE; 62214f6da934SPeter Xu if (locked) 622271335f37SPeter Xu fault_flags |= FAULT_FLAG_ALLOW_RETRY | 622371335f37SPeter Xu FAULT_FLAG_KILLABLE; 622487ffc118SAndrea Arcangeli if (flags & FOLL_NOWAIT) 622587ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY | 622687ffc118SAndrea Arcangeli FAULT_FLAG_RETRY_NOWAIT; 622787ffc118SAndrea Arcangeli if (flags & FOLL_TRIED) { 62284426e945SPeter Xu /* 62294426e945SPeter Xu * Note: FAULT_FLAG_ALLOW_RETRY and 62304426e945SPeter Xu * FAULT_FLAG_TRIED can co-exist 62314426e945SPeter Xu */ 623287ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_TRIED; 623387ffc118SAndrea Arcangeli } 623487ffc118SAndrea Arcangeli ret = hugetlb_fault(mm, vma, vaddr, fault_flags); 623587ffc118SAndrea Arcangeli if (ret & VM_FAULT_ERROR) { 62362be7cfedSDaniel Jordan err = vm_fault_to_errno(ret, flags); 62371c59827dSHugh Dickins remainder = 0; 62381c59827dSHugh Dickins break; 62391c59827dSHugh Dickins } 624087ffc118SAndrea Arcangeli if (ret & VM_FAULT_RETRY) { 62414f6da934SPeter Xu if (locked && 62421ac25013SAndrea Arcangeli !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) 62434f6da934SPeter Xu *locked = 0; 624487ffc118SAndrea Arcangeli *nr_pages = 0; 624587ffc118SAndrea Arcangeli /* 624687ffc118SAndrea Arcangeli * VM_FAULT_RETRY must not return an 624787ffc118SAndrea Arcangeli * error, it will return zero 624887ffc118SAndrea Arcangeli * instead. 624987ffc118SAndrea Arcangeli * 625087ffc118SAndrea Arcangeli * No need to update "position" as the 625187ffc118SAndrea Arcangeli * caller will not check it after 625287ffc118SAndrea Arcangeli * *nr_pages is set to 0. 625387ffc118SAndrea Arcangeli */ 625487ffc118SAndrea Arcangeli return i; 625587ffc118SAndrea Arcangeli } 625687ffc118SAndrea Arcangeli continue; 625787ffc118SAndrea Arcangeli } 625863551ae0SDavid Gibson 6259a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 62607f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 62618fde12caSLinus Torvalds 6262b6a2619cSDavid Hildenbrand VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) && 6263b6a2619cSDavid Hildenbrand !PageAnonExclusive(page), page); 6264b6a2619cSDavid Hildenbrand 62658fde12caSLinus Torvalds /* 6266acbfb087SZhigang Lu * If subpage information not requested, update counters 6267acbfb087SZhigang Lu * and skip the same_page loop below. 6268acbfb087SZhigang Lu */ 6269acbfb087SZhigang Lu if (!pages && !vmas && !pfn_offset && 6270acbfb087SZhigang Lu (vaddr + huge_page_size(h) < vma->vm_end) && 6271acbfb087SZhigang Lu (remainder >= pages_per_huge_page(h))) { 6272acbfb087SZhigang Lu vaddr += huge_page_size(h); 6273acbfb087SZhigang Lu remainder -= pages_per_huge_page(h); 6274acbfb087SZhigang Lu i += pages_per_huge_page(h); 6275acbfb087SZhigang Lu spin_unlock(ptl); 6276acbfb087SZhigang Lu continue; 6277acbfb087SZhigang Lu } 6278acbfb087SZhigang Lu 6279d08af0a5SJoao Martins /* vaddr may not be aligned to PAGE_SIZE */ 6280d08af0a5SJoao Martins refs = min3(pages_per_huge_page(h) - pfn_offset, remainder, 6281d08af0a5SJoao Martins (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT); 62820fa5bc40SJoao Martins 628382e5d378SJoao Martins if (pages || vmas) 628414455eabSCheng Li record_subpages_vmas(nth_page(page, pfn_offset), 628582e5d378SJoao Martins vma, refs, 628682e5d378SJoao Martins likely(pages) ? pages + i : NULL, 628782e5d378SJoao Martins vmas ? vmas + i : NULL); 628863551ae0SDavid Gibson 628982e5d378SJoao Martins if (pages) { 62900fa5bc40SJoao Martins /* 6291822951d8SMatthew Wilcox (Oracle) * try_grab_folio() should always succeed here, 62920fa5bc40SJoao Martins * because: a) we hold the ptl lock, and b) we've just 62930fa5bc40SJoao Martins * checked that the huge page is present in the page 62940fa5bc40SJoao Martins * tables. If the huge page is present, then the tail 62950fa5bc40SJoao Martins * pages must also be present. The ptl prevents the 62960fa5bc40SJoao Martins * head page and tail pages from being rearranged in 62970fa5bc40SJoao Martins * any way. So this page must be available at this 62980fa5bc40SJoao Martins * point, unless the page refcount overflowed: 62990fa5bc40SJoao Martins */ 6300822951d8SMatthew Wilcox (Oracle) if (WARN_ON_ONCE(!try_grab_folio(pages[i], refs, 63010fa5bc40SJoao Martins flags))) { 63020fa5bc40SJoao Martins spin_unlock(ptl); 63030fa5bc40SJoao Martins remainder = 0; 63040fa5bc40SJoao Martins err = -ENOMEM; 63050fa5bc40SJoao Martins break; 63060fa5bc40SJoao Martins } 6307d5d4b0aaSChen, Kenneth W } 630882e5d378SJoao Martins 630982e5d378SJoao Martins vaddr += (refs << PAGE_SHIFT); 631082e5d378SJoao Martins remainder -= refs; 631182e5d378SJoao Martins i += refs; 631282e5d378SJoao Martins 6313cb900f41SKirill A. Shutemov spin_unlock(ptl); 631463551ae0SDavid Gibson } 631528a35716SMichel Lespinasse *nr_pages = remainder; 631687ffc118SAndrea Arcangeli /* 631787ffc118SAndrea Arcangeli * setting position is actually required only if remainder is 631887ffc118SAndrea Arcangeli * not zero but it's faster not to add a "if (remainder)" 631987ffc118SAndrea Arcangeli * branch. 632087ffc118SAndrea Arcangeli */ 632163551ae0SDavid Gibson *position = vaddr; 632263551ae0SDavid Gibson 63232be7cfedSDaniel Jordan return i ? i : err; 632463551ae0SDavid Gibson } 63258f860591SZhang, Yanmin 63267da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 63275a90d5a1SPeter Xu unsigned long address, unsigned long end, 63285a90d5a1SPeter Xu pgprot_t newprot, unsigned long cp_flags) 63298f860591SZhang, Yanmin { 63308f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 63318f860591SZhang, Yanmin unsigned long start = address; 63328f860591SZhang, Yanmin pte_t *ptep; 63338f860591SZhang, Yanmin pte_t pte; 6334a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 633560dfaad6SPeter Xu unsigned long pages = 0, psize = huge_page_size(h); 6336dff11abeSMike Kravetz bool shared_pmd = false; 6337ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 6338e95a9851SMike Kravetz unsigned long last_addr_mask; 63395a90d5a1SPeter Xu bool uffd_wp = cp_flags & MM_CP_UFFD_WP; 63405a90d5a1SPeter Xu bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE; 6341dff11abeSMike Kravetz 6342dff11abeSMike Kravetz /* 6343dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 6344ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 6345dff11abeSMike Kravetz * range if PMD sharing is possible. 6346dff11abeSMike Kravetz */ 63477269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 63487269f999SJérôme Glisse 0, vma, mm, start, end); 6349ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 63508f860591SZhang, Yanmin 63518f860591SZhang, Yanmin BUG_ON(address >= end); 6352ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 63538f860591SZhang, Yanmin 6354ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 635540549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 635683cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 635740549ba8SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 635860dfaad6SPeter Xu for (; address < end; address += psize) { 6359cb900f41SKirill A. Shutemov spinlock_t *ptl; 636060dfaad6SPeter Xu ptep = huge_pte_offset(mm, address, psize); 6361e95a9851SMike Kravetz if (!ptep) { 6362e95a9851SMike Kravetz address |= last_addr_mask; 63638f860591SZhang, Yanmin continue; 6364e95a9851SMike Kravetz } 6365cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 63664ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, address, ptep)) { 636760dfaad6SPeter Xu /* 636860dfaad6SPeter Xu * When uffd-wp is enabled on the vma, unshare 636960dfaad6SPeter Xu * shouldn't happen at all. Warn about it if it 637060dfaad6SPeter Xu * happened due to some reason. 637160dfaad6SPeter Xu */ 637260dfaad6SPeter Xu WARN_ON_ONCE(uffd_wp || uffd_wp_resolve); 63737da4d641SPeter Zijlstra pages++; 6374cb900f41SKirill A. Shutemov spin_unlock(ptl); 6375dff11abeSMike Kravetz shared_pmd = true; 63764ddb4d91SMike Kravetz address |= last_addr_mask; 637739dde65cSChen, Kenneth W continue; 63787da4d641SPeter Zijlstra } 6379a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 6380a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 6381a8bda28dSNaoya Horiguchi spin_unlock(ptl); 6382a8bda28dSNaoya Horiguchi continue; 6383a8bda28dSNaoya Horiguchi } 6384a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 6385a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 63866c287605SDavid Hildenbrand struct page *page = pfn_swap_entry_to_page(entry); 6387a8bda28dSNaoya Horiguchi 63886c287605SDavid Hildenbrand if (!is_readable_migration_entry(entry)) { 6389a8bda28dSNaoya Horiguchi pte_t newpte; 6390a8bda28dSNaoya Horiguchi 63916c287605SDavid Hildenbrand if (PageAnon(page)) 63926c287605SDavid Hildenbrand entry = make_readable_exclusive_migration_entry( 63936c287605SDavid Hildenbrand swp_offset(entry)); 63946c287605SDavid Hildenbrand else 63954dd845b5SAlistair Popple entry = make_readable_migration_entry( 63964dd845b5SAlistair Popple swp_offset(entry)); 6397a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 63985a90d5a1SPeter Xu if (uffd_wp) 63995a90d5a1SPeter Xu newpte = pte_swp_mkuffd_wp(newpte); 64005a90d5a1SPeter Xu else if (uffd_wp_resolve) 64015a90d5a1SPeter Xu newpte = pte_swp_clear_uffd_wp(newpte); 640218f39629SQi Zheng set_huge_pte_at(mm, address, ptep, newpte); 6403a8bda28dSNaoya Horiguchi pages++; 6404a8bda28dSNaoya Horiguchi } 6405a8bda28dSNaoya Horiguchi spin_unlock(ptl); 6406a8bda28dSNaoya Horiguchi continue; 6407a8bda28dSNaoya Horiguchi } 640860dfaad6SPeter Xu if (unlikely(pte_marker_uffd_wp(pte))) { 640960dfaad6SPeter Xu /* 641060dfaad6SPeter Xu * This is changing a non-present pte into a none pte, 641160dfaad6SPeter Xu * no need for huge_ptep_modify_prot_start/commit(). 641260dfaad6SPeter Xu */ 641360dfaad6SPeter Xu if (uffd_wp_resolve) 641460dfaad6SPeter Xu huge_pte_clear(mm, address, ptep, psize); 641560dfaad6SPeter Xu } 6416a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 6417023bdd00SAneesh Kumar K.V pte_t old_pte; 641879c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 6419023bdd00SAneesh Kumar K.V 6420023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 642116785bd7SAnshuman Khandual pte = huge_pte_modify(old_pte, newprot); 642279c1c594SChristophe Leroy pte = arch_make_huge_pte(pte, shift, vma->vm_flags); 64235a90d5a1SPeter Xu if (uffd_wp) 64245a90d5a1SPeter Xu pte = huge_pte_mkuffd_wp(huge_pte_wrprotect(pte)); 64255a90d5a1SPeter Xu else if (uffd_wp_resolve) 64265a90d5a1SPeter Xu pte = huge_pte_clear_uffd_wp(pte); 6427023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 64287da4d641SPeter Zijlstra pages++; 642960dfaad6SPeter Xu } else { 643060dfaad6SPeter Xu /* None pte */ 643160dfaad6SPeter Xu if (unlikely(uffd_wp)) 643260dfaad6SPeter Xu /* Safe to modify directly (none->non-present). */ 643360dfaad6SPeter Xu set_huge_pte_at(mm, address, ptep, 643460dfaad6SPeter Xu make_pte_marker(PTE_MARKER_UFFD_WP)); 64358f860591SZhang, Yanmin } 6436cb900f41SKirill A. Shutemov spin_unlock(ptl); 64378f860591SZhang, Yanmin } 6438d833352aSMel Gorman /* 6439c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 6440d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 6441c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 6442dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 6443dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 6444d833352aSMel Gorman */ 6445dff11abeSMike Kravetz if (shared_pmd) 6446ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 6447dff11abeSMike Kravetz else 64485491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 64490f10851eSJérôme Glisse /* 64500f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 64510f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 64520f10851eSJérôme Glisse * 6453ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 64540f10851eSJérôme Glisse */ 645583cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 645640549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 6457ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 64587da4d641SPeter Zijlstra 64597da4d641SPeter Zijlstra return pages << h->order; 64608f860591SZhang, Yanmin } 64618f860591SZhang, Yanmin 646233b8f84aSMike Kravetz /* Return true if reservation was successful, false otherwise. */ 646333b8f84aSMike Kravetz bool hugetlb_reserve_pages(struct inode *inode, 6464a1e78772SMel Gorman long from, long to, 64655a6fe125SMel Gorman struct vm_area_struct *vma, 6466ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 6467e4e574b7SAdam Litke { 646833b8f84aSMike Kravetz long chg, add = -1; 6469a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 647090481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 64719119a41eSJoonsoo Kim struct resv_map *resv_map; 6472075a61d0SMina Almasry struct hugetlb_cgroup *h_cg = NULL; 64730db9d74eSMina Almasry long gbl_reserve, regions_needed = 0; 6474e4e574b7SAdam Litke 647563489f8eSMike Kravetz /* This should never happen */ 647663489f8eSMike Kravetz if (from > to) { 647763489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 647833b8f84aSMike Kravetz return false; 647963489f8eSMike Kravetz } 648063489f8eSMike Kravetz 6481a1e78772SMel Gorman /* 64828d9bfb26SMike Kravetz * vma specific semaphore used for pmd sharing synchronization 64838d9bfb26SMike Kravetz */ 64848d9bfb26SMike Kravetz hugetlb_vma_lock_alloc(vma); 64858d9bfb26SMike Kravetz 64868d9bfb26SMike Kravetz /* 648717c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 648817c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 648990481622SDavid Gibson * without using reserves 649017c9d12eSMel Gorman */ 6491ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 649233b8f84aSMike Kravetz return true; 649317c9d12eSMel Gorman 649417c9d12eSMel Gorman /* 6495a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 6496a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 6497a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 6498a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 6499a1e78772SMel Gorman */ 65009119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 6501f27a5136SMike Kravetz /* 6502f27a5136SMike Kravetz * resv_map can not be NULL as hugetlb_reserve_pages is only 6503f27a5136SMike Kravetz * called for inodes for which resv_maps were created (see 6504f27a5136SMike Kravetz * hugetlbfs_get_inode). 6505f27a5136SMike Kravetz */ 65064e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 65079119a41eSJoonsoo Kim 65080db9d74eSMina Almasry chg = region_chg(resv_map, from, to, ®ions_needed); 65099119a41eSJoonsoo Kim } else { 6510e9fe92aeSMina Almasry /* Private mapping. */ 65119119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 65125a6fe125SMel Gorman if (!resv_map) 65138d9bfb26SMike Kravetz goto out_err; 65145a6fe125SMel Gorman 651517c9d12eSMel Gorman chg = to - from; 651617c9d12eSMel Gorman 65175a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 65185a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 65195a6fe125SMel Gorman } 65205a6fe125SMel Gorman 652133b8f84aSMike Kravetz if (chg < 0) 6522c50ac050SDave Hansen goto out_err; 652317c9d12eSMel Gorman 652433b8f84aSMike Kravetz if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), 652533b8f84aSMike Kravetz chg * pages_per_huge_page(h), &h_cg) < 0) 6526075a61d0SMina Almasry goto out_err; 6527075a61d0SMina Almasry 6528075a61d0SMina Almasry if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { 6529075a61d0SMina Almasry /* For private mappings, the hugetlb_cgroup uncharge info hangs 6530075a61d0SMina Almasry * of the resv_map. 6531075a61d0SMina Almasry */ 6532075a61d0SMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); 6533075a61d0SMina Almasry } 6534075a61d0SMina Almasry 65351c5ecae3SMike Kravetz /* 65361c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 65371c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 65381c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 65391c5ecae3SMike Kravetz */ 65401c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 654133b8f84aSMike Kravetz if (gbl_reserve < 0) 6542075a61d0SMina Almasry goto out_uncharge_cgroup; 654317c9d12eSMel Gorman 654417c9d12eSMel Gorman /* 654517c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 654690481622SDavid Gibson * Hand the pages back to the subpool if there are not 654717c9d12eSMel Gorman */ 654833b8f84aSMike Kravetz if (hugetlb_acct_memory(h, gbl_reserve) < 0) 6549075a61d0SMina Almasry goto out_put_pages; 655017c9d12eSMel Gorman 655117c9d12eSMel Gorman /* 655217c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 655317c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 655417c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 655517c9d12eSMel Gorman * the reservation was and the page cache tells how much of 655617c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 655717c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 655817c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 655917c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 656017c9d12eSMel Gorman * else has to be done for private mappings here 656117c9d12eSMel Gorman */ 656233039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 6563075a61d0SMina Almasry add = region_add(resv_map, from, to, regions_needed, h, h_cg); 656433039678SMike Kravetz 65650db9d74eSMina Almasry if (unlikely(add < 0)) { 65660db9d74eSMina Almasry hugetlb_acct_memory(h, -gbl_reserve); 6567075a61d0SMina Almasry goto out_put_pages; 65680db9d74eSMina Almasry } else if (unlikely(chg > add)) { 656933039678SMike Kravetz /* 657033039678SMike Kravetz * pages in this range were added to the reserve 657133039678SMike Kravetz * map between region_chg and region_add. This 657233039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 657333039678SMike Kravetz * the subpool and reserve counts modified above 657433039678SMike Kravetz * based on the difference. 657533039678SMike Kravetz */ 657633039678SMike Kravetz long rsv_adjust; 657733039678SMike Kravetz 6578d85aecf2SMiaohe Lin /* 6579d85aecf2SMiaohe Lin * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the 6580d85aecf2SMiaohe Lin * reference to h_cg->css. See comment below for detail. 6581d85aecf2SMiaohe Lin */ 6582075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd( 6583075a61d0SMina Almasry hstate_index(h), 6584075a61d0SMina Almasry (chg - add) * pages_per_huge_page(h), h_cg); 6585075a61d0SMina Almasry 658633039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 658733039678SMike Kravetz chg - add); 658833039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 6589d85aecf2SMiaohe Lin } else if (h_cg) { 6590d85aecf2SMiaohe Lin /* 6591d85aecf2SMiaohe Lin * The file_regions will hold their own reference to 6592d85aecf2SMiaohe Lin * h_cg->css. So we should release the reference held 6593d85aecf2SMiaohe Lin * via hugetlb_cgroup_charge_cgroup_rsvd() when we are 6594d85aecf2SMiaohe Lin * done. 6595d85aecf2SMiaohe Lin */ 6596d85aecf2SMiaohe Lin hugetlb_cgroup_put_rsvd_cgroup(h_cg); 659733039678SMike Kravetz } 659833039678SMike Kravetz } 659933b8f84aSMike Kravetz return true; 660033b8f84aSMike Kravetz 6601075a61d0SMina Almasry out_put_pages: 6602075a61d0SMina Almasry /* put back original number of pages, chg */ 6603075a61d0SMina Almasry (void)hugepage_subpool_put_pages(spool, chg); 6604075a61d0SMina Almasry out_uncharge_cgroup: 6605075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), 6606075a61d0SMina Almasry chg * pages_per_huge_page(h), h_cg); 6607c50ac050SDave Hansen out_err: 66088d9bfb26SMike Kravetz hugetlb_vma_lock_free(vma); 66095e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 66100db9d74eSMina Almasry /* Only call region_abort if the region_chg succeeded but the 66110db9d74eSMina Almasry * region_add failed or didn't run. 66120db9d74eSMina Almasry */ 66130db9d74eSMina Almasry if (chg >= 0 && add < 0) 66140db9d74eSMina Almasry region_abort(resv_map, from, to, regions_needed); 6615f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 6616f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 661733b8f84aSMike Kravetz return false; 6618a43a8c39SChen, Kenneth W } 6619a43a8c39SChen, Kenneth W 6620b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 6621b5cec28dSMike Kravetz long freed) 6622a43a8c39SChen, Kenneth W { 6623a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 66244e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 66259119a41eSJoonsoo Kim long chg = 0; 662690481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 66271c5ecae3SMike Kravetz long gbl_reserve; 662845c682a6SKen Chen 6629f27a5136SMike Kravetz /* 6630f27a5136SMike Kravetz * Since this routine can be called in the evict inode path for all 6631f27a5136SMike Kravetz * hugetlbfs inodes, resv_map could be NULL. 6632f27a5136SMike Kravetz */ 6633b5cec28dSMike Kravetz if (resv_map) { 6634b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 6635b5cec28dSMike Kravetz /* 6636b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 6637b5cec28dSMike Kravetz * must be split and another region descriptor can not be 6638b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 6639b5cec28dSMike Kravetz */ 6640b5cec28dSMike Kravetz if (chg < 0) 6641b5cec28dSMike Kravetz return chg; 6642b5cec28dSMike Kravetz } 6643b5cec28dSMike Kravetz 664445c682a6SKen Chen spin_lock(&inode->i_lock); 6645e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 664645c682a6SKen Chen spin_unlock(&inode->i_lock); 664745c682a6SKen Chen 66481c5ecae3SMike Kravetz /* 66491c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 66501c5ecae3SMike Kravetz * reservations to be released may be adjusted. 6651dddf31a4SMiaohe Lin * 6652dddf31a4SMiaohe Lin * Note that !resv_map implies freed == 0. So (chg - freed) 6653dddf31a4SMiaohe Lin * won't go negative. 66541c5ecae3SMike Kravetz */ 66551c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 66561c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 6657b5cec28dSMike Kravetz 6658b5cec28dSMike Kravetz return 0; 6659a43a8c39SChen, Kenneth W } 666093f70f90SNaoya Horiguchi 66613212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 66623212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 66633212b535SSteve Capper struct vm_area_struct *vma, 66643212b535SSteve Capper unsigned long addr, pgoff_t idx) 66653212b535SSteve Capper { 66663212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 66673212b535SSteve Capper svma->vm_start; 66683212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 66693212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 66703212b535SSteve Capper 66713212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 6672de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 6673de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 66743212b535SSteve Capper 66753212b535SSteve Capper /* 66763212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 66773212b535SSteve Capper * page table page. 6678131a79b4SMike Kravetz * 6679131a79b4SMike Kravetz * Also, vma_lock (vm_private_data) is required for sharing. 66803212b535SSteve Capper */ 66813212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 66823212b535SSteve Capper vm_flags != svm_flags || 6683131a79b4SMike Kravetz !range_in_vma(svma, sbase, s_end) || 6684131a79b4SMike Kravetz !svma->vm_private_data) 66853212b535SSteve Capper return 0; 66863212b535SSteve Capper 66873212b535SSteve Capper return saddr; 66883212b535SSteve Capper } 66893212b535SSteve Capper 6690*bbff39ccSMike Kravetz bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 66913212b535SSteve Capper { 6692*bbff39ccSMike Kravetz unsigned long start = addr & PUD_MASK; 6693*bbff39ccSMike Kravetz unsigned long end = start + PUD_SIZE; 6694*bbff39ccSMike Kravetz 66958d9bfb26SMike Kravetz #ifdef CONFIG_USERFAULTFD 66968d9bfb26SMike Kravetz if (uffd_disable_huge_pmd_share(vma)) 66978d9bfb26SMike Kravetz return false; 66988d9bfb26SMike Kravetz #endif 66993212b535SSteve Capper /* 67003212b535SSteve Capper * check on proper vm_flags and page table alignment 67013212b535SSteve Capper */ 67028d9bfb26SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 670331aafb45SNicholas Krause return false; 6704*bbff39ccSMike Kravetz if (!vma->vm_private_data) /* vma lock required for sharing */ 67058d9bfb26SMike Kravetz return false; 67068d9bfb26SMike Kravetz if (!range_in_vma(vma, start, end)) 67078d9bfb26SMike Kravetz return false; 67088d9bfb26SMike Kravetz return true; 67098d9bfb26SMike Kravetz } 67108d9bfb26SMike Kravetz 67113212b535SSteve Capper /* 6712017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 6713017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 6714017b1660SMike Kravetz * shared pmd mappings. 6715017b1660SMike Kravetz */ 6716017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 6717017b1660SMike Kravetz unsigned long *start, unsigned long *end) 6718017b1660SMike Kravetz { 6719a1ba9da8SLi Xinhai unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), 6720a1ba9da8SLi Xinhai v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 6721017b1660SMike Kravetz 6722a1ba9da8SLi Xinhai /* 6723f0953a1bSIngo Molnar * vma needs to span at least one aligned PUD size, and the range 6724f0953a1bSIngo Molnar * must be at least partially within in. 6725a1ba9da8SLi Xinhai */ 6726a1ba9da8SLi Xinhai if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || 6727a1ba9da8SLi Xinhai (*end <= v_start) || (*start >= v_end)) 6728017b1660SMike Kravetz return; 6729017b1660SMike Kravetz 673075802ca6SPeter Xu /* Extend the range to be PUD aligned for a worst case scenario */ 6731a1ba9da8SLi Xinhai if (*start > v_start) 6732a1ba9da8SLi Xinhai *start = ALIGN_DOWN(*start, PUD_SIZE); 6733017b1660SMike Kravetz 6734a1ba9da8SLi Xinhai if (*end < v_end) 6735a1ba9da8SLi Xinhai *end = ALIGN(*end, PUD_SIZE); 6736017b1660SMike Kravetz } 6737017b1660SMike Kravetz 67388d9bfb26SMike Kravetz static bool __vma_shareable_flags_pmd(struct vm_area_struct *vma) 67398d9bfb26SMike Kravetz { 67408d9bfb26SMike Kravetz return vma->vm_flags & (VM_MAYSHARE | VM_SHARED) && 67418d9bfb26SMike Kravetz vma->vm_private_data; 67428d9bfb26SMike Kravetz } 67438d9bfb26SMike Kravetz 67448d9bfb26SMike Kravetz void hugetlb_vma_lock_read(struct vm_area_struct *vma) 67458d9bfb26SMike Kravetz { 67468d9bfb26SMike Kravetz if (__vma_shareable_flags_pmd(vma)) { 67478d9bfb26SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 67488d9bfb26SMike Kravetz 67498d9bfb26SMike Kravetz down_read(&vma_lock->rw_sema); 67508d9bfb26SMike Kravetz } 67518d9bfb26SMike Kravetz } 67528d9bfb26SMike Kravetz 67538d9bfb26SMike Kravetz void hugetlb_vma_unlock_read(struct vm_area_struct *vma) 67548d9bfb26SMike Kravetz { 67558d9bfb26SMike Kravetz if (__vma_shareable_flags_pmd(vma)) { 67568d9bfb26SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 67578d9bfb26SMike Kravetz 67588d9bfb26SMike Kravetz up_read(&vma_lock->rw_sema); 67598d9bfb26SMike Kravetz } 67608d9bfb26SMike Kravetz } 67618d9bfb26SMike Kravetz 67628d9bfb26SMike Kravetz void hugetlb_vma_lock_write(struct vm_area_struct *vma) 67638d9bfb26SMike Kravetz { 67648d9bfb26SMike Kravetz if (__vma_shareable_flags_pmd(vma)) { 67658d9bfb26SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 67668d9bfb26SMike Kravetz 67678d9bfb26SMike Kravetz down_write(&vma_lock->rw_sema); 67688d9bfb26SMike Kravetz } 67698d9bfb26SMike Kravetz } 67708d9bfb26SMike Kravetz 67718d9bfb26SMike Kravetz void hugetlb_vma_unlock_write(struct vm_area_struct *vma) 67728d9bfb26SMike Kravetz { 67738d9bfb26SMike Kravetz if (__vma_shareable_flags_pmd(vma)) { 67748d9bfb26SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 67758d9bfb26SMike Kravetz 67768d9bfb26SMike Kravetz up_write(&vma_lock->rw_sema); 67778d9bfb26SMike Kravetz } 67788d9bfb26SMike Kravetz } 67798d9bfb26SMike Kravetz 67808d9bfb26SMike Kravetz int hugetlb_vma_trylock_write(struct vm_area_struct *vma) 67818d9bfb26SMike Kravetz { 67828d9bfb26SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 67838d9bfb26SMike Kravetz 67848d9bfb26SMike Kravetz if (!__vma_shareable_flags_pmd(vma)) 67858d9bfb26SMike Kravetz return 1; 67868d9bfb26SMike Kravetz 67878d9bfb26SMike Kravetz return down_write_trylock(&vma_lock->rw_sema); 67888d9bfb26SMike Kravetz } 67898d9bfb26SMike Kravetz 67908d9bfb26SMike Kravetz void hugetlb_vma_assert_locked(struct vm_area_struct *vma) 67918d9bfb26SMike Kravetz { 67928d9bfb26SMike Kravetz if (__vma_shareable_flags_pmd(vma)) { 67938d9bfb26SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 67948d9bfb26SMike Kravetz 67958d9bfb26SMike Kravetz lockdep_assert_held(&vma_lock->rw_sema); 67968d9bfb26SMike Kravetz } 67978d9bfb26SMike Kravetz } 67988d9bfb26SMike Kravetz 67998d9bfb26SMike Kravetz void hugetlb_vma_lock_release(struct kref *kref) 68008d9bfb26SMike Kravetz { 68018d9bfb26SMike Kravetz struct hugetlb_vma_lock *vma_lock = container_of(kref, 68028d9bfb26SMike Kravetz struct hugetlb_vma_lock, refs); 68038d9bfb26SMike Kravetz 68048d9bfb26SMike Kravetz kfree(vma_lock); 68058d9bfb26SMike Kravetz } 68068d9bfb26SMike Kravetz 6807ecfbd733SMike Kravetz void __hugetlb_vma_unlock_write_put(struct hugetlb_vma_lock *vma_lock) 6808ecfbd733SMike Kravetz { 6809ecfbd733SMike Kravetz struct vm_area_struct *vma = vma_lock->vma; 6810ecfbd733SMike Kravetz 6811ecfbd733SMike Kravetz /* 6812ecfbd733SMike Kravetz * vma_lock structure may or not be released as a result of put, 6813ecfbd733SMike Kravetz * it certainly will no longer be attached to vma so clear pointer. 6814ecfbd733SMike Kravetz * Semaphore synchronizes access to vma_lock->vma field. 6815ecfbd733SMike Kravetz */ 6816ecfbd733SMike Kravetz vma_lock->vma = NULL; 6817ecfbd733SMike Kravetz vma->vm_private_data = NULL; 6818ecfbd733SMike Kravetz up_write(&vma_lock->rw_sema); 6819ecfbd733SMike Kravetz kref_put(&vma_lock->refs, hugetlb_vma_lock_release); 6820ecfbd733SMike Kravetz } 6821ecfbd733SMike Kravetz 6822ecfbd733SMike Kravetz static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma) 6823ecfbd733SMike Kravetz { 6824ecfbd733SMike Kravetz if (__vma_shareable_flags_pmd(vma)) { 6825ecfbd733SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 6826ecfbd733SMike Kravetz 6827ecfbd733SMike Kravetz __hugetlb_vma_unlock_write_put(vma_lock); 6828ecfbd733SMike Kravetz } 6829ecfbd733SMike Kravetz } 6830ecfbd733SMike Kravetz 68318d9bfb26SMike Kravetz static void hugetlb_vma_lock_free(struct vm_area_struct *vma) 68328d9bfb26SMike Kravetz { 68338d9bfb26SMike Kravetz /* 6834131a79b4SMike Kravetz * Only present in sharable vmas. 68358d9bfb26SMike Kravetz */ 6836131a79b4SMike Kravetz if (!vma || !__vma_shareable_flags_pmd(vma)) 68378d9bfb26SMike Kravetz return; 68388d9bfb26SMike Kravetz 68398d9bfb26SMike Kravetz if (vma->vm_private_data) { 68408d9bfb26SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 68418d9bfb26SMike Kravetz 6842ecfbd733SMike Kravetz down_write(&vma_lock->rw_sema); 6843ecfbd733SMike Kravetz __hugetlb_vma_unlock_write_put(vma_lock); 68448d9bfb26SMike Kravetz } 68458d9bfb26SMike Kravetz } 68468d9bfb26SMike Kravetz 68478d9bfb26SMike Kravetz static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma) 68488d9bfb26SMike Kravetz { 68498d9bfb26SMike Kravetz struct hugetlb_vma_lock *vma_lock; 68508d9bfb26SMike Kravetz 68518d9bfb26SMike Kravetz /* Only establish in (flags) sharable vmas */ 68528d9bfb26SMike Kravetz if (!vma || !(vma->vm_flags & VM_MAYSHARE)) 68538d9bfb26SMike Kravetz return; 68548d9bfb26SMike Kravetz 68558d9bfb26SMike Kravetz /* Should never get here with non-NULL vm_private_data */ 68568d9bfb26SMike Kravetz if (vma->vm_private_data) 68578d9bfb26SMike Kravetz return; 68588d9bfb26SMike Kravetz 68598d9bfb26SMike Kravetz vma_lock = kmalloc(sizeof(*vma_lock), GFP_KERNEL); 6860*bbff39ccSMike Kravetz if (!vma_lock) { 68618d9bfb26SMike Kravetz /* 68628d9bfb26SMike Kravetz * If we can not allocate structure, then vma can not 6863*bbff39ccSMike Kravetz * participate in pmd sharing. This is only a possible 6864*bbff39ccSMike Kravetz * performance enhancement and memory saving issue. 6865*bbff39ccSMike Kravetz * However, the lock is also used to synchronize page 6866*bbff39ccSMike Kravetz * faults with truncation. If the lock is not present, 6867*bbff39ccSMike Kravetz * unlikely races could leave pages in a file past i_size 6868*bbff39ccSMike Kravetz * until the file is removed. Warn in the unlikely case of 6869*bbff39ccSMike Kravetz * allocation failure. 68708d9bfb26SMike Kravetz */ 6871*bbff39ccSMike Kravetz pr_warn_once("HugeTLB: unable to allocate vma specific lock\n"); 68728d9bfb26SMike Kravetz return; 6873*bbff39ccSMike Kravetz } 68748d9bfb26SMike Kravetz 68758d9bfb26SMike Kravetz kref_init(&vma_lock->refs); 68768d9bfb26SMike Kravetz init_rwsem(&vma_lock->rw_sema); 68778d9bfb26SMike Kravetz vma_lock->vma = vma; 68788d9bfb26SMike Kravetz vma->vm_private_data = vma_lock; 68798d9bfb26SMike Kravetz } 68808d9bfb26SMike Kravetz 6881017b1660SMike Kravetz /* 68823212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 68833212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 68843212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 68853a47c54fSMike Kravetz * code much cleaner. pmd allocation is essential for the shared case because 68863a47c54fSMike Kravetz * pud has to be populated inside the same i_mmap_rwsem section - otherwise 68873a47c54fSMike Kravetz * racing tasks could either miss the sharing (see huge_pte_offset) or select a 68883a47c54fSMike Kravetz * bad pmd for sharing. 68893212b535SSteve Capper */ 6890aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 6891aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 68923212b535SSteve Capper { 68933212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 68943212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 68953212b535SSteve Capper vma->vm_pgoff; 68963212b535SSteve Capper struct vm_area_struct *svma; 68973212b535SSteve Capper unsigned long saddr; 68983212b535SSteve Capper pte_t *spte = NULL; 68993212b535SSteve Capper pte_t *pte; 6900cb900f41SKirill A. Shutemov spinlock_t *ptl; 69013212b535SSteve Capper 69023a47c54fSMike Kravetz i_mmap_lock_read(mapping); 69033212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 69043212b535SSteve Capper if (svma == vma) 69053212b535SSteve Capper continue; 69063212b535SSteve Capper 69073212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 69083212b535SSteve Capper if (saddr) { 69097868a208SPunit Agrawal spte = huge_pte_offset(svma->vm_mm, saddr, 69107868a208SPunit Agrawal vma_mmu_pagesize(svma)); 69113212b535SSteve Capper if (spte) { 69123212b535SSteve Capper get_page(virt_to_page(spte)); 69133212b535SSteve Capper break; 69143212b535SSteve Capper } 69153212b535SSteve Capper } 69163212b535SSteve Capper } 69173212b535SSteve Capper 69183212b535SSteve Capper if (!spte) 69193212b535SSteve Capper goto out; 69203212b535SSteve Capper 69218bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 6922dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 69233212b535SSteve Capper pud_populate(mm, pud, 69243212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 6925c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 6926dc6c9a35SKirill A. Shutemov } else { 69273212b535SSteve Capper put_page(virt_to_page(spte)); 6928dc6c9a35SKirill A. Shutemov } 6929cb900f41SKirill A. Shutemov spin_unlock(ptl); 69303212b535SSteve Capper out: 69313212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 69323a47c54fSMike Kravetz i_mmap_unlock_read(mapping); 69333212b535SSteve Capper return pte; 69343212b535SSteve Capper } 69353212b535SSteve Capper 69363212b535SSteve Capper /* 69373212b535SSteve Capper * unmap huge page backed by shared pte. 69383212b535SSteve Capper * 69393212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 69403212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 69413212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 69423212b535SSteve Capper * 69433a47c54fSMike Kravetz * Called with page table lock held. 69443212b535SSteve Capper * 69453212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 69463212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 69473212b535SSteve Capper */ 694834ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 69494ddb4d91SMike Kravetz unsigned long addr, pte_t *ptep) 69503212b535SSteve Capper { 69514ddb4d91SMike Kravetz pgd_t *pgd = pgd_offset(mm, addr); 69524ddb4d91SMike Kravetz p4d_t *p4d = p4d_offset(pgd, addr); 69534ddb4d91SMike Kravetz pud_t *pud = pud_offset(p4d, addr); 69543212b535SSteve Capper 695534ae204fSMike Kravetz i_mmap_assert_write_locked(vma->vm_file->f_mapping); 695640549ba8SMike Kravetz hugetlb_vma_assert_locked(vma); 69573212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 69583212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 69593212b535SSteve Capper return 0; 69603212b535SSteve Capper 69613212b535SSteve Capper pud_clear(pud); 69623212b535SSteve Capper put_page(virt_to_page(ptep)); 6963dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 69643212b535SSteve Capper return 1; 69653212b535SSteve Capper } 6966c1991e07SPeter Xu 69679e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 69688d9bfb26SMike Kravetz 696940549ba8SMike Kravetz void hugetlb_vma_lock_read(struct vm_area_struct *vma) 697040549ba8SMike Kravetz { 697140549ba8SMike Kravetz } 697240549ba8SMike Kravetz 697340549ba8SMike Kravetz void hugetlb_vma_unlock_read(struct vm_area_struct *vma) 697440549ba8SMike Kravetz { 697540549ba8SMike Kravetz } 697640549ba8SMike Kravetz 697740549ba8SMike Kravetz void hugetlb_vma_lock_write(struct vm_area_struct *vma) 697840549ba8SMike Kravetz { 697940549ba8SMike Kravetz } 698040549ba8SMike Kravetz 698140549ba8SMike Kravetz void hugetlb_vma_unlock_write(struct vm_area_struct *vma) 698240549ba8SMike Kravetz { 698340549ba8SMike Kravetz } 698440549ba8SMike Kravetz 698540549ba8SMike Kravetz int hugetlb_vma_trylock_write(struct vm_area_struct *vma) 698640549ba8SMike Kravetz { 698740549ba8SMike Kravetz return 1; 698840549ba8SMike Kravetz } 698940549ba8SMike Kravetz 699040549ba8SMike Kravetz void hugetlb_vma_assert_locked(struct vm_area_struct *vma) 699140549ba8SMike Kravetz { 699240549ba8SMike Kravetz } 699340549ba8SMike Kravetz 69948d9bfb26SMike Kravetz void hugetlb_vma_lock_release(struct kref *kref) 69958d9bfb26SMike Kravetz { 69968d9bfb26SMike Kravetz } 69978d9bfb26SMike Kravetz 6998ecfbd733SMike Kravetz static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma) 6999ecfbd733SMike Kravetz { 7000ecfbd733SMike Kravetz } 7001ecfbd733SMike Kravetz 70028d9bfb26SMike Kravetz static void hugetlb_vma_lock_free(struct vm_area_struct *vma) 70038d9bfb26SMike Kravetz { 70048d9bfb26SMike Kravetz } 70058d9bfb26SMike Kravetz 70068d9bfb26SMike Kravetz static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma) 70078d9bfb26SMike Kravetz { 70088d9bfb26SMike Kravetz } 70098d9bfb26SMike Kravetz 7010aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 7011aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 70129e5fc74cSSteve Capper { 70139e5fc74cSSteve Capper return NULL; 70149e5fc74cSSteve Capper } 7015e81f2d22SZhang Zhen 701634ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 70174ddb4d91SMike Kravetz unsigned long addr, pte_t *ptep) 7018e81f2d22SZhang Zhen { 7019e81f2d22SZhang Zhen return 0; 7020e81f2d22SZhang Zhen } 7021017b1660SMike Kravetz 7022017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 7023017b1660SMike Kravetz unsigned long *start, unsigned long *end) 7024017b1660SMike Kravetz { 7025017b1660SMike Kravetz } 7026c1991e07SPeter Xu 7027c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 7028c1991e07SPeter Xu { 7029c1991e07SPeter Xu return false; 7030c1991e07SPeter Xu } 70313212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 70323212b535SSteve Capper 70339e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 7034aec44e0fSPeter Xu pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, 70359e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 70369e5fc74cSSteve Capper { 70379e5fc74cSSteve Capper pgd_t *pgd; 7038c2febafcSKirill A. Shutemov p4d_t *p4d; 70399e5fc74cSSteve Capper pud_t *pud; 70409e5fc74cSSteve Capper pte_t *pte = NULL; 70419e5fc74cSSteve Capper 70429e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 7043f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 7044f4f0a3d8SKirill A. Shutemov if (!p4d) 7045f4f0a3d8SKirill A. Shutemov return NULL; 7046c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 70479e5fc74cSSteve Capper if (pud) { 70489e5fc74cSSteve Capper if (sz == PUD_SIZE) { 70499e5fc74cSSteve Capper pte = (pte_t *)pud; 70509e5fc74cSSteve Capper } else { 70519e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 7052c1991e07SPeter Xu if (want_pmd_share(vma, addr) && pud_none(*pud)) 7053aec44e0fSPeter Xu pte = huge_pmd_share(mm, vma, addr, pud); 70549e5fc74cSSteve Capper else 70559e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 70569e5fc74cSSteve Capper } 70579e5fc74cSSteve Capper } 70584e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 70599e5fc74cSSteve Capper 70609e5fc74cSSteve Capper return pte; 70619e5fc74cSSteve Capper } 70629e5fc74cSSteve Capper 70639b19df29SPunit Agrawal /* 70649b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 70659b19df29SPunit Agrawal * entry at address @addr 70669b19df29SPunit Agrawal * 70678ac0b81aSLi Xinhai * Return: Pointer to page table entry (PUD or PMD) for 70688ac0b81aSLi Xinhai * address @addr, or NULL if a !p*d_present() entry is encountered and the 70699b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 70709b19df29SPunit Agrawal * table. 70719b19df29SPunit Agrawal */ 70727868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 70737868a208SPunit Agrawal unsigned long addr, unsigned long sz) 70749e5fc74cSSteve Capper { 70759e5fc74cSSteve Capper pgd_t *pgd; 7076c2febafcSKirill A. Shutemov p4d_t *p4d; 70778ac0b81aSLi Xinhai pud_t *pud; 70788ac0b81aSLi Xinhai pmd_t *pmd; 70799e5fc74cSSteve Capper 70809e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 7081c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 7082c2febafcSKirill A. Shutemov return NULL; 7083c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 7084c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 7085c2febafcSKirill A. Shutemov return NULL; 70869b19df29SPunit Agrawal 7087c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 70888ac0b81aSLi Xinhai if (sz == PUD_SIZE) 70898ac0b81aSLi Xinhai /* must be pud huge, non-present or none */ 70909e5fc74cSSteve Capper return (pte_t *)pud; 70918ac0b81aSLi Xinhai if (!pud_present(*pud)) 70928ac0b81aSLi Xinhai return NULL; 70938ac0b81aSLi Xinhai /* must have a valid entry and size to go further */ 70949b19df29SPunit Agrawal 70959e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 70968ac0b81aSLi Xinhai /* must be pmd huge, non-present or none */ 70979e5fc74cSSteve Capper return (pte_t *)pmd; 70989e5fc74cSSteve Capper } 70999e5fc74cSSteve Capper 7100e95a9851SMike Kravetz /* 7101e95a9851SMike Kravetz * Return a mask that can be used to update an address to the last huge 7102e95a9851SMike Kravetz * page in a page table page mapping size. Used to skip non-present 7103e95a9851SMike Kravetz * page table entries when linearly scanning address ranges. Architectures 7104e95a9851SMike Kravetz * with unique huge page to page table relationships can define their own 7105e95a9851SMike Kravetz * version of this routine. 7106e95a9851SMike Kravetz */ 7107e95a9851SMike Kravetz unsigned long hugetlb_mask_last_page(struct hstate *h) 7108e95a9851SMike Kravetz { 7109e95a9851SMike Kravetz unsigned long hp_size = huge_page_size(h); 7110e95a9851SMike Kravetz 7111e95a9851SMike Kravetz if (hp_size == PUD_SIZE) 7112e95a9851SMike Kravetz return P4D_SIZE - PUD_SIZE; 7113e95a9851SMike Kravetz else if (hp_size == PMD_SIZE) 7114e95a9851SMike Kravetz return PUD_SIZE - PMD_SIZE; 7115e95a9851SMike Kravetz else 7116e95a9851SMike Kravetz return 0UL; 7117e95a9851SMike Kravetz } 7118e95a9851SMike Kravetz 7119e95a9851SMike Kravetz #else 7120e95a9851SMike Kravetz 7121e95a9851SMike Kravetz /* See description above. Architectures can provide their own version. */ 7122e95a9851SMike Kravetz __weak unsigned long hugetlb_mask_last_page(struct hstate *h) 7123e95a9851SMike Kravetz { 71244ddb4d91SMike Kravetz #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 71254ddb4d91SMike Kravetz if (huge_page_size(h) == PMD_SIZE) 71264ddb4d91SMike Kravetz return PUD_SIZE - PMD_SIZE; 71274ddb4d91SMike Kravetz #endif 7128e95a9851SMike Kravetz return 0UL; 7129e95a9851SMike Kravetz } 7130e95a9851SMike Kravetz 713161f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 713261f77edaSNaoya Horiguchi 713361f77edaSNaoya Horiguchi /* 713461f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 713561f77edaSNaoya Horiguchi * behavior. 713661f77edaSNaoya Horiguchi */ 713761f77edaSNaoya Horiguchi struct page * __weak 713861f77edaSNaoya Horiguchi follow_huge_addr(struct mm_struct *mm, unsigned long address, 713961f77edaSNaoya Horiguchi int write) 714061f77edaSNaoya Horiguchi { 714161f77edaSNaoya Horiguchi return ERR_PTR(-EINVAL); 714261f77edaSNaoya Horiguchi } 714361f77edaSNaoya Horiguchi 714461f77edaSNaoya Horiguchi struct page * __weak 71454dc71451SAneesh Kumar K.V follow_huge_pd(struct vm_area_struct *vma, 71464dc71451SAneesh Kumar K.V unsigned long address, hugepd_t hpd, int flags, int pdshift) 71474dc71451SAneesh Kumar K.V { 71484dc71451SAneesh Kumar K.V WARN(1, "hugepd follow called with no support for hugepage directory format\n"); 71494dc71451SAneesh Kumar K.V return NULL; 71504dc71451SAneesh Kumar K.V } 71514dc71451SAneesh Kumar K.V 71524dc71451SAneesh Kumar K.V struct page * __weak 71539e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 7154e66f17ffSNaoya Horiguchi pmd_t *pmd, int flags) 71559e5fc74cSSteve Capper { 7156e66f17ffSNaoya Horiguchi struct page *page = NULL; 7157e66f17ffSNaoya Horiguchi spinlock_t *ptl; 7158c9d398faSNaoya Horiguchi pte_t pte; 71593faa52c0SJohn Hubbard 71608909691bSDavid Hildenbrand /* 71618909691bSDavid Hildenbrand * FOLL_PIN is not supported for follow_page(). Ordinary GUP goes via 71628909691bSDavid Hildenbrand * follow_hugetlb_page(). 71638909691bSDavid Hildenbrand */ 71648909691bSDavid Hildenbrand if (WARN_ON_ONCE(flags & FOLL_PIN)) 71653faa52c0SJohn Hubbard return NULL; 71663faa52c0SJohn Hubbard 7167e66f17ffSNaoya Horiguchi retry: 7168e66f17ffSNaoya Horiguchi ptl = pmd_lockptr(mm, pmd); 7169e66f17ffSNaoya Horiguchi spin_lock(ptl); 7170e66f17ffSNaoya Horiguchi /* 7171e66f17ffSNaoya Horiguchi * make sure that the address range covered by this pmd is not 7172e66f17ffSNaoya Horiguchi * unmapped from other threads. 7173e66f17ffSNaoya Horiguchi */ 7174e66f17ffSNaoya Horiguchi if (!pmd_huge(*pmd)) 7175e66f17ffSNaoya Horiguchi goto out; 7176c9d398faSNaoya Horiguchi pte = huge_ptep_get((pte_t *)pmd); 7177c9d398faSNaoya Horiguchi if (pte_present(pte)) { 717897534127SGerald Schaefer page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 71793faa52c0SJohn Hubbard /* 71803faa52c0SJohn Hubbard * try_grab_page() should always succeed here, because: a) we 71813faa52c0SJohn Hubbard * hold the pmd (ptl) lock, and b) we've just checked that the 71823faa52c0SJohn Hubbard * huge pmd (head) page is present in the page tables. The ptl 71833faa52c0SJohn Hubbard * prevents the head page and tail pages from being rearranged 71843faa52c0SJohn Hubbard * in any way. So this page must be available at this point, 71853faa52c0SJohn Hubbard * unless the page refcount overflowed: 71863faa52c0SJohn Hubbard */ 71873faa52c0SJohn Hubbard if (WARN_ON_ONCE(!try_grab_page(page, flags))) { 71883faa52c0SJohn Hubbard page = NULL; 71893faa52c0SJohn Hubbard goto out; 71903faa52c0SJohn Hubbard } 7191e66f17ffSNaoya Horiguchi } else { 7192c9d398faSNaoya Horiguchi if (is_hugetlb_entry_migration(pte)) { 7193e66f17ffSNaoya Horiguchi spin_unlock(ptl); 7194ad1ac596SMiaohe Lin __migration_entry_wait_huge((pte_t *)pmd, ptl); 7195e66f17ffSNaoya Horiguchi goto retry; 7196e66f17ffSNaoya Horiguchi } 7197e66f17ffSNaoya Horiguchi /* 7198e66f17ffSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 7199e66f17ffSNaoya Horiguchi * follow_page_mask(). 7200e66f17ffSNaoya Horiguchi */ 7201e66f17ffSNaoya Horiguchi } 7202e66f17ffSNaoya Horiguchi out: 7203e66f17ffSNaoya Horiguchi spin_unlock(ptl); 72049e5fc74cSSteve Capper return page; 72059e5fc74cSSteve Capper } 72069e5fc74cSSteve Capper 720761f77edaSNaoya Horiguchi struct page * __weak 72089e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 7209e66f17ffSNaoya Horiguchi pud_t *pud, int flags) 72109e5fc74cSSteve Capper { 72113a194f3fSNaoya Horiguchi struct page *page = NULL; 72123a194f3fSNaoya Horiguchi spinlock_t *ptl; 72133a194f3fSNaoya Horiguchi pte_t pte; 72143a194f3fSNaoya Horiguchi 72153a194f3fSNaoya Horiguchi if (WARN_ON_ONCE(flags & FOLL_PIN)) 7216e66f17ffSNaoya Horiguchi return NULL; 72179e5fc74cSSteve Capper 72183a194f3fSNaoya Horiguchi retry: 72193a194f3fSNaoya Horiguchi ptl = huge_pte_lock(hstate_sizelog(PUD_SHIFT), mm, (pte_t *)pud); 72203a194f3fSNaoya Horiguchi if (!pud_huge(*pud)) 72213a194f3fSNaoya Horiguchi goto out; 72223a194f3fSNaoya Horiguchi pte = huge_ptep_get((pte_t *)pud); 72233a194f3fSNaoya Horiguchi if (pte_present(pte)) { 72243a194f3fSNaoya Horiguchi page = pud_page(*pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 72253a194f3fSNaoya Horiguchi if (WARN_ON_ONCE(!try_grab_page(page, flags))) { 72263a194f3fSNaoya Horiguchi page = NULL; 72273a194f3fSNaoya Horiguchi goto out; 72283a194f3fSNaoya Horiguchi } 72293a194f3fSNaoya Horiguchi } else { 72303a194f3fSNaoya Horiguchi if (is_hugetlb_entry_migration(pte)) { 72313a194f3fSNaoya Horiguchi spin_unlock(ptl); 72323a194f3fSNaoya Horiguchi __migration_entry_wait(mm, (pte_t *)pud, ptl); 72333a194f3fSNaoya Horiguchi goto retry; 72343a194f3fSNaoya Horiguchi } 72353a194f3fSNaoya Horiguchi /* 72363a194f3fSNaoya Horiguchi * hwpoisoned entry is treated as no_page_table in 72373a194f3fSNaoya Horiguchi * follow_page_mask(). 72383a194f3fSNaoya Horiguchi */ 72393a194f3fSNaoya Horiguchi } 72403a194f3fSNaoya Horiguchi out: 72413a194f3fSNaoya Horiguchi spin_unlock(ptl); 72423a194f3fSNaoya Horiguchi return page; 72439e5fc74cSSteve Capper } 72449e5fc74cSSteve Capper 7245faaa5b62SAnshuman Khandual struct page * __weak 7246faaa5b62SAnshuman Khandual follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) 7247faaa5b62SAnshuman Khandual { 72483faa52c0SJohn Hubbard if (flags & (FOLL_GET | FOLL_PIN)) 7249faaa5b62SAnshuman Khandual return NULL; 7250faaa5b62SAnshuman Khandual 7251faaa5b62SAnshuman Khandual return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); 7252faaa5b62SAnshuman Khandual } 7253faaa5b62SAnshuman Khandual 72547ce82f4cSMiaohe Lin int isolate_hugetlb(struct page *page, struct list_head *list) 725531caf665SNaoya Horiguchi { 72567ce82f4cSMiaohe Lin int ret = 0; 7257bcc54222SNaoya Horiguchi 7258db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 72598f251a3dSMike Kravetz if (!PageHeadHuge(page) || 72608f251a3dSMike Kravetz !HPageMigratable(page) || 72610eb2df2bSMuchun Song !get_page_unless_zero(page)) { 72627ce82f4cSMiaohe Lin ret = -EBUSY; 7263bcc54222SNaoya Horiguchi goto unlock; 7264bcc54222SNaoya Horiguchi } 72658f251a3dSMike Kravetz ClearHPageMigratable(page); 726631caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 7267bcc54222SNaoya Horiguchi unlock: 7268db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7269bcc54222SNaoya Horiguchi return ret; 727031caf665SNaoya Horiguchi } 727131caf665SNaoya Horiguchi 727225182f05SNaoya Horiguchi int get_hwpoison_huge_page(struct page *page, bool *hugetlb) 727325182f05SNaoya Horiguchi { 727425182f05SNaoya Horiguchi int ret = 0; 727525182f05SNaoya Horiguchi 727625182f05SNaoya Horiguchi *hugetlb = false; 727725182f05SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 727825182f05SNaoya Horiguchi if (PageHeadHuge(page)) { 727925182f05SNaoya Horiguchi *hugetlb = true; 7280b283d983SNaoya Horiguchi if (HPageFreed(page)) 7281b283d983SNaoya Horiguchi ret = 0; 7282b283d983SNaoya Horiguchi else if (HPageMigratable(page)) 728325182f05SNaoya Horiguchi ret = get_page_unless_zero(page); 72840ed950d1SNaoya Horiguchi else 72850ed950d1SNaoya Horiguchi ret = -EBUSY; 728625182f05SNaoya Horiguchi } 728725182f05SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 728825182f05SNaoya Horiguchi return ret; 728925182f05SNaoya Horiguchi } 729025182f05SNaoya Horiguchi 7291405ce051SNaoya Horiguchi int get_huge_page_for_hwpoison(unsigned long pfn, int flags) 7292405ce051SNaoya Horiguchi { 7293405ce051SNaoya Horiguchi int ret; 7294405ce051SNaoya Horiguchi 7295405ce051SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 7296405ce051SNaoya Horiguchi ret = __get_huge_page_for_hwpoison(pfn, flags); 7297405ce051SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 7298405ce051SNaoya Horiguchi return ret; 7299405ce051SNaoya Horiguchi } 7300405ce051SNaoya Horiguchi 730131caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 730231caf665SNaoya Horiguchi { 7303db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 73048f251a3dSMike Kravetz SetHPageMigratable(page); 730531caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 7306db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 730731caf665SNaoya Horiguchi put_page(page); 730831caf665SNaoya Horiguchi } 7309ab5ac90aSMichal Hocko 7310ab5ac90aSMichal Hocko void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) 7311ab5ac90aSMichal Hocko { 7312ab5ac90aSMichal Hocko struct hstate *h = page_hstate(oldpage); 7313ab5ac90aSMichal Hocko 7314ab5ac90aSMichal Hocko hugetlb_cgroup_migrate(oldpage, newpage); 7315ab5ac90aSMichal Hocko set_page_owner_migrate_reason(newpage, reason); 7316ab5ac90aSMichal Hocko 7317ab5ac90aSMichal Hocko /* 7318ab5ac90aSMichal Hocko * transfer temporary state of the new huge page. This is 7319ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 7320ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 7321ab5ac90aSMichal Hocko * the temporary status. 7322ab5ac90aSMichal Hocko * 7323ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 7324ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 7325ab5ac90aSMichal Hocko * the per-node's. 7326ab5ac90aSMichal Hocko */ 73279157c311SMike Kravetz if (HPageTemporary(newpage)) { 7328ab5ac90aSMichal Hocko int old_nid = page_to_nid(oldpage); 7329ab5ac90aSMichal Hocko int new_nid = page_to_nid(newpage); 7330ab5ac90aSMichal Hocko 73319157c311SMike Kravetz SetHPageTemporary(oldpage); 73329157c311SMike Kravetz ClearHPageTemporary(newpage); 7333ab5ac90aSMichal Hocko 73345af1ab1dSMiaohe Lin /* 73355af1ab1dSMiaohe Lin * There is no need to transfer the per-node surplus state 73365af1ab1dSMiaohe Lin * when we do not cross the node. 73375af1ab1dSMiaohe Lin */ 73385af1ab1dSMiaohe Lin if (new_nid == old_nid) 73395af1ab1dSMiaohe Lin return; 7340db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 7341ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 7342ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 7343ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 7344ab5ac90aSMichal Hocko } 7345db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7346ab5ac90aSMichal Hocko } 7347ab5ac90aSMichal Hocko } 7348cf11e85fSRoman Gushchin 73496dfeaff9SPeter Xu /* 73506dfeaff9SPeter Xu * This function will unconditionally remove all the shared pmd pgtable entries 73516dfeaff9SPeter Xu * within the specific vma for a hugetlbfs memory range. 73526dfeaff9SPeter Xu */ 73536dfeaff9SPeter Xu void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) 73546dfeaff9SPeter Xu { 73556dfeaff9SPeter Xu struct hstate *h = hstate_vma(vma); 73566dfeaff9SPeter Xu unsigned long sz = huge_page_size(h); 73576dfeaff9SPeter Xu struct mm_struct *mm = vma->vm_mm; 73586dfeaff9SPeter Xu struct mmu_notifier_range range; 73596dfeaff9SPeter Xu unsigned long address, start, end; 73606dfeaff9SPeter Xu spinlock_t *ptl; 73616dfeaff9SPeter Xu pte_t *ptep; 73626dfeaff9SPeter Xu 73636dfeaff9SPeter Xu if (!(vma->vm_flags & VM_MAYSHARE)) 73646dfeaff9SPeter Xu return; 73656dfeaff9SPeter Xu 73666dfeaff9SPeter Xu start = ALIGN(vma->vm_start, PUD_SIZE); 73676dfeaff9SPeter Xu end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 73686dfeaff9SPeter Xu 73696dfeaff9SPeter Xu if (start >= end) 73706dfeaff9SPeter Xu return; 73716dfeaff9SPeter Xu 73729c8bbfacSBaolin Wang flush_cache_range(vma, start, end); 73736dfeaff9SPeter Xu /* 73746dfeaff9SPeter Xu * No need to call adjust_range_if_pmd_sharing_possible(), because 73756dfeaff9SPeter Xu * we have already done the PUD_SIZE alignment. 73766dfeaff9SPeter Xu */ 73776dfeaff9SPeter Xu mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, 73786dfeaff9SPeter Xu start, end); 73796dfeaff9SPeter Xu mmu_notifier_invalidate_range_start(&range); 738040549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 73816dfeaff9SPeter Xu i_mmap_lock_write(vma->vm_file->f_mapping); 73826dfeaff9SPeter Xu for (address = start; address < end; address += PUD_SIZE) { 73836dfeaff9SPeter Xu ptep = huge_pte_offset(mm, address, sz); 73846dfeaff9SPeter Xu if (!ptep) 73856dfeaff9SPeter Xu continue; 73866dfeaff9SPeter Xu ptl = huge_pte_lock(h, mm, ptep); 73874ddb4d91SMike Kravetz huge_pmd_unshare(mm, vma, address, ptep); 73886dfeaff9SPeter Xu spin_unlock(ptl); 73896dfeaff9SPeter Xu } 73906dfeaff9SPeter Xu flush_hugetlb_tlb_range(vma, start, end); 73916dfeaff9SPeter Xu i_mmap_unlock_write(vma->vm_file->f_mapping); 739240549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 73936dfeaff9SPeter Xu /* 73946dfeaff9SPeter Xu * No need to call mmu_notifier_invalidate_range(), see 7395ee65728eSMike Rapoport * Documentation/mm/mmu_notifier.rst. 73966dfeaff9SPeter Xu */ 73976dfeaff9SPeter Xu mmu_notifier_invalidate_range_end(&range); 73986dfeaff9SPeter Xu } 73996dfeaff9SPeter Xu 7400cf11e85fSRoman Gushchin #ifdef CONFIG_CMA 7401cf11e85fSRoman Gushchin static bool cma_reserve_called __initdata; 7402cf11e85fSRoman Gushchin 7403cf11e85fSRoman Gushchin static int __init cmdline_parse_hugetlb_cma(char *p) 7404cf11e85fSRoman Gushchin { 740538e719abSBaolin Wang int nid, count = 0; 740638e719abSBaolin Wang unsigned long tmp; 740738e719abSBaolin Wang char *s = p; 740838e719abSBaolin Wang 740938e719abSBaolin Wang while (*s) { 741038e719abSBaolin Wang if (sscanf(s, "%lu%n", &tmp, &count) != 1) 741138e719abSBaolin Wang break; 741238e719abSBaolin Wang 741338e719abSBaolin Wang if (s[count] == ':') { 7414f9317f77SMike Kravetz if (tmp >= MAX_NUMNODES) 741538e719abSBaolin Wang break; 7416f9317f77SMike Kravetz nid = array_index_nospec(tmp, MAX_NUMNODES); 741738e719abSBaolin Wang 741838e719abSBaolin Wang s += count + 1; 741938e719abSBaolin Wang tmp = memparse(s, &s); 742038e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = tmp; 742138e719abSBaolin Wang hugetlb_cma_size += tmp; 742238e719abSBaolin Wang 742338e719abSBaolin Wang /* 742438e719abSBaolin Wang * Skip the separator if have one, otherwise 742538e719abSBaolin Wang * break the parsing. 742638e719abSBaolin Wang */ 742738e719abSBaolin Wang if (*s == ',') 742838e719abSBaolin Wang s++; 742938e719abSBaolin Wang else 743038e719abSBaolin Wang break; 743138e719abSBaolin Wang } else { 7432cf11e85fSRoman Gushchin hugetlb_cma_size = memparse(p, &p); 743338e719abSBaolin Wang break; 743438e719abSBaolin Wang } 743538e719abSBaolin Wang } 743638e719abSBaolin Wang 7437cf11e85fSRoman Gushchin return 0; 7438cf11e85fSRoman Gushchin } 7439cf11e85fSRoman Gushchin 7440cf11e85fSRoman Gushchin early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); 7441cf11e85fSRoman Gushchin 7442cf11e85fSRoman Gushchin void __init hugetlb_cma_reserve(int order) 7443cf11e85fSRoman Gushchin { 7444cf11e85fSRoman Gushchin unsigned long size, reserved, per_node; 744538e719abSBaolin Wang bool node_specific_cma_alloc = false; 7446cf11e85fSRoman Gushchin int nid; 7447cf11e85fSRoman Gushchin 7448cf11e85fSRoman Gushchin cma_reserve_called = true; 7449cf11e85fSRoman Gushchin 7450cf11e85fSRoman Gushchin if (!hugetlb_cma_size) 7451cf11e85fSRoman Gushchin return; 7452cf11e85fSRoman Gushchin 745338e719abSBaolin Wang for (nid = 0; nid < MAX_NUMNODES; nid++) { 745438e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 745538e719abSBaolin Wang continue; 745638e719abSBaolin Wang 745730a51400SPeng Liu if (!node_online(nid)) { 745838e719abSBaolin Wang pr_warn("hugetlb_cma: invalid node %d specified\n", nid); 745938e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 746038e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 746138e719abSBaolin Wang continue; 746238e719abSBaolin Wang } 746338e719abSBaolin Wang 746438e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) { 746538e719abSBaolin Wang pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n", 746638e719abSBaolin Wang nid, (PAGE_SIZE << order) / SZ_1M); 746738e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 746838e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 746938e719abSBaolin Wang } else { 747038e719abSBaolin Wang node_specific_cma_alloc = true; 747138e719abSBaolin Wang } 747238e719abSBaolin Wang } 747338e719abSBaolin Wang 747438e719abSBaolin Wang /* Validate the CMA size again in case some invalid nodes specified. */ 747538e719abSBaolin Wang if (!hugetlb_cma_size) 747638e719abSBaolin Wang return; 747738e719abSBaolin Wang 7478cf11e85fSRoman Gushchin if (hugetlb_cma_size < (PAGE_SIZE << order)) { 7479cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", 7480cf11e85fSRoman Gushchin (PAGE_SIZE << order) / SZ_1M); 7481a01f4390SMike Kravetz hugetlb_cma_size = 0; 7482cf11e85fSRoman Gushchin return; 7483cf11e85fSRoman Gushchin } 7484cf11e85fSRoman Gushchin 748538e719abSBaolin Wang if (!node_specific_cma_alloc) { 7486cf11e85fSRoman Gushchin /* 7487cf11e85fSRoman Gushchin * If 3 GB area is requested on a machine with 4 numa nodes, 7488cf11e85fSRoman Gushchin * let's allocate 1 GB on first three nodes and ignore the last one. 7489cf11e85fSRoman Gushchin */ 7490cf11e85fSRoman Gushchin per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); 7491cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", 7492cf11e85fSRoman Gushchin hugetlb_cma_size / SZ_1M, per_node / SZ_1M); 749338e719abSBaolin Wang } 7494cf11e85fSRoman Gushchin 7495cf11e85fSRoman Gushchin reserved = 0; 749630a51400SPeng Liu for_each_online_node(nid) { 7497cf11e85fSRoman Gushchin int res; 74982281f797SBarry Song char name[CMA_MAX_NAME]; 7499cf11e85fSRoman Gushchin 750038e719abSBaolin Wang if (node_specific_cma_alloc) { 750138e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 750238e719abSBaolin Wang continue; 750338e719abSBaolin Wang 750438e719abSBaolin Wang size = hugetlb_cma_size_in_node[nid]; 750538e719abSBaolin Wang } else { 7506cf11e85fSRoman Gushchin size = min(per_node, hugetlb_cma_size - reserved); 750738e719abSBaolin Wang } 750838e719abSBaolin Wang 7509cf11e85fSRoman Gushchin size = round_up(size, PAGE_SIZE << order); 7510cf11e85fSRoman Gushchin 75112281f797SBarry Song snprintf(name, sizeof(name), "hugetlb%d", nid); 7512a01f4390SMike Kravetz /* 7513a01f4390SMike Kravetz * Note that 'order per bit' is based on smallest size that 7514a01f4390SMike Kravetz * may be returned to CMA allocator in the case of 7515a01f4390SMike Kravetz * huge page demotion. 7516a01f4390SMike Kravetz */ 7517a01f4390SMike Kravetz res = cma_declare_contiguous_nid(0, size, 0, 7518a01f4390SMike Kravetz PAGE_SIZE << HUGETLB_PAGE_ORDER, 751929d0f41dSBarry Song 0, false, name, 7520cf11e85fSRoman Gushchin &hugetlb_cma[nid], nid); 7521cf11e85fSRoman Gushchin if (res) { 7522cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: reservation failed: err %d, node %d", 7523cf11e85fSRoman Gushchin res, nid); 7524cf11e85fSRoman Gushchin continue; 7525cf11e85fSRoman Gushchin } 7526cf11e85fSRoman Gushchin 7527cf11e85fSRoman Gushchin reserved += size; 7528cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", 7529cf11e85fSRoman Gushchin size / SZ_1M, nid); 7530cf11e85fSRoman Gushchin 7531cf11e85fSRoman Gushchin if (reserved >= hugetlb_cma_size) 7532cf11e85fSRoman Gushchin break; 7533cf11e85fSRoman Gushchin } 7534a01f4390SMike Kravetz 7535a01f4390SMike Kravetz if (!reserved) 7536a01f4390SMike Kravetz /* 7537a01f4390SMike Kravetz * hugetlb_cma_size is used to determine if allocations from 7538a01f4390SMike Kravetz * cma are possible. Set to zero if no cma regions are set up. 7539a01f4390SMike Kravetz */ 7540a01f4390SMike Kravetz hugetlb_cma_size = 0; 7541cf11e85fSRoman Gushchin } 7542cf11e85fSRoman Gushchin 7543263b8998SMiaohe Lin static void __init hugetlb_cma_check(void) 7544cf11e85fSRoman Gushchin { 7545cf11e85fSRoman Gushchin if (!hugetlb_cma_size || cma_reserve_called) 7546cf11e85fSRoman Gushchin return; 7547cf11e85fSRoman Gushchin 7548cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); 7549cf11e85fSRoman Gushchin } 7550cf11e85fSRoman Gushchin 7551cf11e85fSRoman Gushchin #endif /* CONFIG_CMA */ 7552