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 1017612b8a31SMike Kravetz * - For shared mappings this is a per-vma semaphore that may be 1018612b8a31SMike Kravetz * allocated in a subsequent call to hugetlb_vm_op_open. 1019612b8a31SMike Kravetz * Before clearing, make sure pointer is not associated with vma 1020612b8a31SMike Kravetz * as this will leak the structure. This is the case when called 1021612b8a31SMike Kravetz * via clear_vma_resv_huge_pages() and hugetlb_vm_op_open has already 1022612b8a31SMike Kravetz * been called to allocate a new structure. 10238d9bfb26SMike Kravetz * - For MAP_PRIVATE mappings, this is the reserve map which does 10248d9bfb26SMike Kravetz * not apply to children. Faults generated by the children are 10258d9bfb26SMike Kravetz * not guaranteed to succeed, even if read-only. 10268d9bfb26SMike Kravetz */ 1027612b8a31SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 1028612b8a31SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 1029612b8a31SMike Kravetz 1030612b8a31SMike Kravetz if (vma_lock && vma_lock->vma != vma) 1031612b8a31SMike Kravetz vma->vm_private_data = NULL; 1032612b8a31SMike Kravetz } else 1033612b8a31SMike Kravetz vma->vm_private_data = NULL; 1034a1e78772SMel Gorman } 1035a1e78772SMel Gorman 1036550a7d60SMina Almasry /* 1037550a7d60SMina Almasry * Reset and decrement one ref on hugepage private reservation. 1038550a7d60SMina Almasry * Called with mm->mmap_sem writer semaphore held. 1039550a7d60SMina Almasry * This function should be only used by move_vma() and operate on 1040550a7d60SMina Almasry * same sized vma. It should never come here with last ref on the 1041550a7d60SMina Almasry * reservation. 1042550a7d60SMina Almasry */ 1043550a7d60SMina Almasry void clear_vma_resv_huge_pages(struct vm_area_struct *vma) 1044550a7d60SMina Almasry { 1045550a7d60SMina Almasry /* 1046550a7d60SMina Almasry * Clear the old hugetlb private page reservation. 1047550a7d60SMina Almasry * It has already been transferred to new_vma. 1048550a7d60SMina Almasry * 1049550a7d60SMina Almasry * During a mremap() operation of a hugetlb vma we call move_vma() 1050550a7d60SMina Almasry * which copies vma into new_vma and unmaps vma. After the copy 1051550a7d60SMina Almasry * operation both new_vma and vma share a reference to the resv_map 1052550a7d60SMina Almasry * struct, and at that point vma is about to be unmapped. We don't 1053550a7d60SMina Almasry * want to return the reservation to the pool at unmap of vma because 1054550a7d60SMina Almasry * the reservation still lives on in new_vma, so simply decrement the 1055550a7d60SMina Almasry * ref here and remove the resv_map reference from this vma. 1056550a7d60SMina Almasry */ 1057550a7d60SMina Almasry struct resv_map *reservations = vma_resv_map(vma); 1058550a7d60SMina Almasry 1059afe041c2SBui Quang Minh if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 1060afe041c2SBui Quang Minh resv_map_put_hugetlb_cgroup_uncharge_info(reservations); 1061550a7d60SMina Almasry kref_put(&reservations->refs, resv_map_release); 1062afe041c2SBui Quang Minh } 1063550a7d60SMina Almasry 10648d9bfb26SMike Kravetz hugetlb_dup_vma_private(vma); 1065550a7d60SMina Almasry } 1066550a7d60SMina Almasry 1067a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 1068559ec2f8SNicholas Krause static bool vma_has_reserves(struct vm_area_struct *vma, long chg) 1069a1e78772SMel Gorman { 1070af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 1071af0ed73eSJoonsoo Kim /* 1072af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 1073af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 1074af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 1075af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 1076af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 1077af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 1078af0ed73eSJoonsoo Kim * properly, so add work-around here. 1079af0ed73eSJoonsoo Kim */ 1080af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 1081559ec2f8SNicholas Krause return true; 1082af0ed73eSJoonsoo Kim else 1083559ec2f8SNicholas Krause return false; 1084af0ed73eSJoonsoo Kim } 1085a63884e9SJoonsoo Kim 1086a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 10871fb1b0e9SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 10881fb1b0e9SMike Kravetz /* 10891fb1b0e9SMike Kravetz * We know VM_NORESERVE is not set. Therefore, there SHOULD 10901fb1b0e9SMike Kravetz * be a region map for all pages. The only situation where 10911fb1b0e9SMike Kravetz * there is no region map is if a hole was punched via 10927c8de358SEthon Paul * fallocate. In this case, there really are no reserves to 10931fb1b0e9SMike Kravetz * use. This situation is indicated if chg != 0. 10941fb1b0e9SMike Kravetz */ 10951fb1b0e9SMike Kravetz if (chg) 10961fb1b0e9SMike Kravetz return false; 10971fb1b0e9SMike Kravetz else 1098559ec2f8SNicholas Krause return true; 10991fb1b0e9SMike Kravetz } 1100a63884e9SJoonsoo Kim 1101a63884e9SJoonsoo Kim /* 1102a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 1103a63884e9SJoonsoo Kim * private mappings. 1104a63884e9SJoonsoo Kim */ 110567961f9dSMike Kravetz if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 110667961f9dSMike Kravetz /* 110767961f9dSMike Kravetz * Like the shared case above, a hole punch or truncate 110867961f9dSMike Kravetz * could have been performed on the private mapping. 110967961f9dSMike Kravetz * Examine the value of chg to determine if reserves 111067961f9dSMike Kravetz * actually exist or were previously consumed. 111167961f9dSMike Kravetz * Very Subtle - The value of chg comes from a previous 111267961f9dSMike Kravetz * call to vma_needs_reserves(). The reserve map for 111367961f9dSMike Kravetz * private mappings has different (opposite) semantics 111467961f9dSMike Kravetz * than that of shared mappings. vma_needs_reserves() 111567961f9dSMike Kravetz * has already taken this difference in semantics into 111667961f9dSMike Kravetz * account. Therefore, the meaning of chg is the same 111767961f9dSMike Kravetz * as in the shared case above. Code could easily be 111867961f9dSMike Kravetz * combined, but keeping it separate draws attention to 111967961f9dSMike Kravetz * subtle differences. 112067961f9dSMike Kravetz */ 112167961f9dSMike Kravetz if (chg) 112267961f9dSMike Kravetz return false; 112367961f9dSMike Kravetz else 1124559ec2f8SNicholas Krause return true; 112567961f9dSMike Kravetz } 1126a63884e9SJoonsoo Kim 1127559ec2f8SNicholas Krause return false; 1128a1e78772SMel Gorman } 1129a1e78772SMel Gorman 1130a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 11311da177e4SLinus Torvalds { 11321da177e4SLinus Torvalds int nid = page_to_nid(page); 11339487ca60SMike Kravetz 11349487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1135b65a4edaSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1136b65a4edaSMike Kravetz 11370edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 1138a5516438SAndi Kleen h->free_huge_pages++; 1139a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 11406c037149SMike Kravetz SetHPageFreed(page); 11411da177e4SLinus Torvalds } 11421da177e4SLinus Torvalds 114394310cbcSAnshuman Khandual static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) 1144bf50bab2SNaoya Horiguchi { 1145bf50bab2SNaoya Horiguchi struct page *page; 11461a08ae36SPavel Tatashin bool pin = !!(current->flags & PF_MEMALLOC_PIN); 1147bf50bab2SNaoya Horiguchi 11489487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 1149bbe88753SJoonsoo Kim list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { 11506077c943SAlex Sierra if (pin && !is_longterm_pinnable_page(page)) 1151bbe88753SJoonsoo Kim continue; 1152bbe88753SJoonsoo Kim 11536664bfc8SWei Yang if (PageHWPoison(page)) 11546664bfc8SWei Yang continue; 1155bbe88753SJoonsoo Kim 11560edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 1157a9869b83SNaoya Horiguchi set_page_refcounted(page); 11586c037149SMike Kravetz ClearHPageFreed(page); 1159bf50bab2SNaoya Horiguchi h->free_huge_pages--; 1160bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 1161bf50bab2SNaoya Horiguchi return page; 1162bf50bab2SNaoya Horiguchi } 1163bf50bab2SNaoya Horiguchi 11646664bfc8SWei Yang return NULL; 11656664bfc8SWei Yang } 11666664bfc8SWei Yang 11673e59fcb0SMichal Hocko static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, 11683e59fcb0SMichal Hocko nodemask_t *nmask) 116994310cbcSAnshuman Khandual { 11703e59fcb0SMichal Hocko unsigned int cpuset_mems_cookie; 11713e59fcb0SMichal Hocko struct zonelist *zonelist; 11723e59fcb0SMichal Hocko struct zone *zone; 11733e59fcb0SMichal Hocko struct zoneref *z; 117498fa15f3SAnshuman Khandual int node = NUMA_NO_NODE; 11753e59fcb0SMichal Hocko 11763e59fcb0SMichal Hocko zonelist = node_zonelist(nid, gfp_mask); 11773e59fcb0SMichal Hocko 11783e59fcb0SMichal Hocko retry_cpuset: 11793e59fcb0SMichal Hocko cpuset_mems_cookie = read_mems_allowed_begin(); 11803e59fcb0SMichal Hocko for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 118194310cbcSAnshuman Khandual struct page *page; 118294310cbcSAnshuman Khandual 11833e59fcb0SMichal Hocko if (!cpuset_zone_allowed(zone, gfp_mask)) 11843e59fcb0SMichal Hocko continue; 11853e59fcb0SMichal Hocko /* 11863e59fcb0SMichal Hocko * no need to ask again on the same node. Pool is node rather than 11873e59fcb0SMichal Hocko * zone aware 11883e59fcb0SMichal Hocko */ 11893e59fcb0SMichal Hocko if (zone_to_nid(zone) == node) 11903e59fcb0SMichal Hocko continue; 11913e59fcb0SMichal Hocko node = zone_to_nid(zone); 119294310cbcSAnshuman Khandual 119394310cbcSAnshuman Khandual page = dequeue_huge_page_node_exact(h, node); 119494310cbcSAnshuman Khandual if (page) 119594310cbcSAnshuman Khandual return page; 119694310cbcSAnshuman Khandual } 11973e59fcb0SMichal Hocko if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 11983e59fcb0SMichal Hocko goto retry_cpuset; 11993e59fcb0SMichal Hocko 120094310cbcSAnshuman Khandual return NULL; 120194310cbcSAnshuman Khandual } 120294310cbcSAnshuman Khandual 12038346d69dSXin Hao static unsigned long available_huge_pages(struct hstate *h) 12048346d69dSXin Hao { 12058346d69dSXin Hao return h->free_huge_pages - h->resv_huge_pages; 12068346d69dSXin Hao } 12078346d69dSXin Hao 1208a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 1209a5516438SAndi Kleen struct vm_area_struct *vma, 1210af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 1211af0ed73eSJoonsoo Kim long chg) 12121da177e4SLinus Torvalds { 1213cfcaa66fSBen Widawsky struct page *page = NULL; 1214480eccf9SLee Schermerhorn struct mempolicy *mpol; 121504ec6264SVlastimil Babka gfp_t gfp_mask; 12163e59fcb0SMichal Hocko nodemask_t *nodemask; 121704ec6264SVlastimil Babka int nid; 12181da177e4SLinus Torvalds 1219a1e78772SMel Gorman /* 1220a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 1221a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 1222a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 1223a1e78772SMel Gorman */ 12248346d69dSXin Hao if (!vma_has_reserves(vma, chg) && !available_huge_pages(h)) 1225c0ff7453SMiao Xie goto err; 1226a1e78772SMel Gorman 122704f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 12288346d69dSXin Hao if (avoid_reserve && !available_huge_pages(h)) 12296eab04a8SJustin P. Mattock goto err; 123004f2cbe3SMel Gorman 123104ec6264SVlastimil Babka gfp_mask = htlb_alloc_mask(h); 123204ec6264SVlastimil Babka nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 1233cfcaa66fSBen Widawsky 1234cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 12353e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 1236cfcaa66fSBen Widawsky 1237cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 1238cfcaa66fSBen Widawsky nodemask = NULL; 1239cfcaa66fSBen Widawsky } 1240cfcaa66fSBen Widawsky 1241cfcaa66fSBen Widawsky if (!page) 1242cfcaa66fSBen Widawsky page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); 1243cfcaa66fSBen Widawsky 12443e59fcb0SMichal Hocko if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { 1245d6995da3SMike Kravetz SetHPageRestoreReserve(page); 1246a63884e9SJoonsoo Kim h->resv_huge_pages--; 1247bf50bab2SNaoya Horiguchi } 1248cc9a6c87SMel Gorman 1249cc9a6c87SMel Gorman mpol_cond_put(mpol); 1250cc9a6c87SMel Gorman return page; 1251cc9a6c87SMel Gorman 1252c0ff7453SMiao Xie err: 1253cc9a6c87SMel Gorman return NULL; 12541da177e4SLinus Torvalds } 12551da177e4SLinus Torvalds 12561cac6f2cSLuiz Capitulino /* 12571cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 12581cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 12591cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 12601cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 12611cac6f2cSLuiz Capitulino * node for alloc or free. 12621cac6f2cSLuiz Capitulino */ 12631cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 12641cac6f2cSLuiz Capitulino { 12650edaf86cSAndrew Morton nid = next_node_in(nid, *nodes_allowed); 12661cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 12671cac6f2cSLuiz Capitulino 12681cac6f2cSLuiz Capitulino return nid; 12691cac6f2cSLuiz Capitulino } 12701cac6f2cSLuiz Capitulino 12711cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 12721cac6f2cSLuiz Capitulino { 12731cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 12741cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 12751cac6f2cSLuiz Capitulino return nid; 12761cac6f2cSLuiz Capitulino } 12771cac6f2cSLuiz Capitulino 12781cac6f2cSLuiz Capitulino /* 12791cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 12801cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 12811cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 12821cac6f2cSLuiz Capitulino * mask. 12831cac6f2cSLuiz Capitulino */ 12841cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 12851cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 12861cac6f2cSLuiz Capitulino { 12871cac6f2cSLuiz Capitulino int nid; 12881cac6f2cSLuiz Capitulino 12891cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 12901cac6f2cSLuiz Capitulino 12911cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 12921cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 12931cac6f2cSLuiz Capitulino 12941cac6f2cSLuiz Capitulino return nid; 12951cac6f2cSLuiz Capitulino } 12961cac6f2cSLuiz Capitulino 12971cac6f2cSLuiz Capitulino /* 129810c6ec49SMike Kravetz * helper for remove_pool_huge_page() - return the previously saved 12991cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 13001cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 13011cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 13021cac6f2cSLuiz Capitulino */ 13031cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 13041cac6f2cSLuiz Capitulino { 13051cac6f2cSLuiz Capitulino int nid; 13061cac6f2cSLuiz Capitulino 13071cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 13081cac6f2cSLuiz Capitulino 13091cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 13101cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 13111cac6f2cSLuiz Capitulino 13121cac6f2cSLuiz Capitulino return nid; 13131cac6f2cSLuiz Capitulino } 13141cac6f2cSLuiz Capitulino 13151cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 13161cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 13171cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 13181cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 13191cac6f2cSLuiz Capitulino nr_nodes--) 13201cac6f2cSLuiz Capitulino 13211cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 13221cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 13231cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 13241cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 13251cac6f2cSLuiz Capitulino nr_nodes--) 13261cac6f2cSLuiz Capitulino 13278531fc6fSMike Kravetz /* used to demote non-gigantic_huge pages as well */ 132834d9e35bSMike Kravetz static void __destroy_compound_gigantic_page(struct page *page, 132934d9e35bSMike Kravetz unsigned int order, bool demote) 1330944d9fecSLuiz Capitulino { 1331944d9fecSLuiz Capitulino int i; 1332944d9fecSLuiz Capitulino int nr_pages = 1 << order; 133314455eabSCheng Li struct page *p; 1334944d9fecSLuiz Capitulino 1335c8cc708aSGerald Schaefer atomic_set(compound_mapcount_ptr(page), 0); 133647e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 133747e29d32SJohn Hubbard 133814455eabSCheng Li for (i = 1; i < nr_pages; i++) { 133914455eabSCheng Li p = nth_page(page, i); 1340a01f4390SMike Kravetz p->mapping = NULL; 13411d798ca3SKirill A. Shutemov clear_compound_head(p); 134234d9e35bSMike Kravetz if (!demote) 1343944d9fecSLuiz Capitulino set_page_refcounted(p); 1344944d9fecSLuiz Capitulino } 1345944d9fecSLuiz Capitulino 1346944d9fecSLuiz Capitulino set_compound_order(page, 0); 13475232c63fSMatthew Wilcox (Oracle) #ifdef CONFIG_64BIT 1348ba9c1201SGerald Schaefer page[1].compound_nr = 0; 13495232c63fSMatthew Wilcox (Oracle) #endif 1350944d9fecSLuiz Capitulino __ClearPageHead(page); 1351944d9fecSLuiz Capitulino } 1352944d9fecSLuiz Capitulino 13538531fc6fSMike Kravetz static void destroy_compound_hugetlb_page_for_demote(struct page *page, 13548531fc6fSMike Kravetz unsigned int order) 13558531fc6fSMike Kravetz { 13568531fc6fSMike Kravetz __destroy_compound_gigantic_page(page, order, true); 13578531fc6fSMike Kravetz } 13588531fc6fSMike Kravetz 13598531fc6fSMike Kravetz #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 136034d9e35bSMike Kravetz static void destroy_compound_gigantic_page(struct page *page, 136134d9e35bSMike Kravetz unsigned int order) 136234d9e35bSMike Kravetz { 136334d9e35bSMike Kravetz __destroy_compound_gigantic_page(page, order, false); 136434d9e35bSMike Kravetz } 136534d9e35bSMike Kravetz 1366d00181b9SKirill A. Shutemov static void free_gigantic_page(struct page *page, unsigned int order) 1367944d9fecSLuiz Capitulino { 1368cf11e85fSRoman Gushchin /* 1369cf11e85fSRoman Gushchin * If the page isn't allocated using the cma allocator, 1370cf11e85fSRoman Gushchin * cma_release() returns false. 1371cf11e85fSRoman Gushchin */ 1372dbda8feaSBarry Song #ifdef CONFIG_CMA 1373dbda8feaSBarry Song if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order)) 1374cf11e85fSRoman Gushchin return; 1375dbda8feaSBarry Song #endif 1376cf11e85fSRoman Gushchin 1377944d9fecSLuiz Capitulino free_contig_range(page_to_pfn(page), 1 << order); 1378944d9fecSLuiz Capitulino } 1379944d9fecSLuiz Capitulino 13804eb0716eSAlexandre Ghiti #ifdef CONFIG_CONTIG_ALLOC 1381d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 1382d9cc948fSMichal Hocko int nid, nodemask_t *nodemask) 1383944d9fecSLuiz Capitulino { 138404adbc3fSMiaohe Lin unsigned long nr_pages = pages_per_huge_page(h); 1385953f064aSLi Xinhai if (nid == NUMA_NO_NODE) 1386953f064aSLi Xinhai nid = numa_mem_id(); 1387944d9fecSLuiz Capitulino 1388dbda8feaSBarry Song #ifdef CONFIG_CMA 1389dbda8feaSBarry Song { 1390cf11e85fSRoman Gushchin struct page *page; 1391cf11e85fSRoman Gushchin int node; 1392cf11e85fSRoman Gushchin 1393953f064aSLi Xinhai if (hugetlb_cma[nid]) { 1394953f064aSLi Xinhai page = cma_alloc(hugetlb_cma[nid], nr_pages, 1395953f064aSLi Xinhai huge_page_order(h), true); 1396953f064aSLi Xinhai if (page) 1397953f064aSLi Xinhai return page; 1398953f064aSLi Xinhai } 1399953f064aSLi Xinhai 1400953f064aSLi Xinhai if (!(gfp_mask & __GFP_THISNODE)) { 1401cf11e85fSRoman Gushchin for_each_node_mask(node, *nodemask) { 1402953f064aSLi Xinhai if (node == nid || !hugetlb_cma[node]) 1403cf11e85fSRoman Gushchin continue; 1404cf11e85fSRoman Gushchin 1405cf11e85fSRoman Gushchin page = cma_alloc(hugetlb_cma[node], nr_pages, 1406cf11e85fSRoman Gushchin huge_page_order(h), true); 1407cf11e85fSRoman Gushchin if (page) 1408cf11e85fSRoman Gushchin return page; 1409cf11e85fSRoman Gushchin } 1410cf11e85fSRoman Gushchin } 1411953f064aSLi Xinhai } 1412dbda8feaSBarry Song #endif 1413cf11e85fSRoman Gushchin 14145e27a2dfSAnshuman Khandual return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); 1415944d9fecSLuiz Capitulino } 1416944d9fecSLuiz Capitulino 14174eb0716eSAlexandre Ghiti #else /* !CONFIG_CONTIG_ALLOC */ 14184eb0716eSAlexandre Ghiti 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 } 14234eb0716eSAlexandre Ghiti #endif /* CONFIG_CONTIG_ALLOC */ 1424944d9fecSLuiz Capitulino 1425e1073d1eSAneesh Kumar K.V #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 1426d9cc948fSMichal Hocko static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, 14274eb0716eSAlexandre Ghiti int nid, nodemask_t *nodemask) 14284eb0716eSAlexandre Ghiti { 14294eb0716eSAlexandre Ghiti return NULL; 14304eb0716eSAlexandre Ghiti } 1431d00181b9SKirill A. Shutemov static inline void free_gigantic_page(struct page *page, unsigned int order) { } 1432944d9fecSLuiz Capitulino static inline void destroy_compound_gigantic_page(struct page *page, 1433d00181b9SKirill A. Shutemov unsigned int order) { } 1434944d9fecSLuiz Capitulino #endif 1435944d9fecSLuiz Capitulino 14366eb4e88aSMike Kravetz /* 14376eb4e88aSMike Kravetz * Remove hugetlb page from lists, and update dtor so that page appears 143834d9e35bSMike Kravetz * as just a compound page. 143934d9e35bSMike Kravetz * 144034d9e35bSMike Kravetz * A reference is held on the page, except in the case of demote. 14416eb4e88aSMike Kravetz * 14426eb4e88aSMike Kravetz * Must be called with hugetlb lock held. 14436eb4e88aSMike Kravetz */ 144434d9e35bSMike Kravetz static void __remove_hugetlb_page(struct hstate *h, struct page *page, 144534d9e35bSMike Kravetz bool adjust_surplus, 144634d9e35bSMike Kravetz bool demote) 14476eb4e88aSMike Kravetz { 14486eb4e88aSMike Kravetz int nid = page_to_nid(page); 1449f074732dSSidhartha Kumar struct folio *folio = page_folio(page); 14506eb4e88aSMike Kravetz 1451f074732dSSidhartha Kumar VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio(folio), folio); 1452f074732dSSidhartha Kumar VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio_rsvd(folio), folio); 14536eb4e88aSMike Kravetz 14549487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 14556eb4e88aSMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 14566eb4e88aSMike Kravetz return; 14576eb4e88aSMike Kravetz 14586eb4e88aSMike Kravetz list_del(&page->lru); 14596eb4e88aSMike Kravetz 14606eb4e88aSMike Kravetz if (HPageFreed(page)) { 14616eb4e88aSMike Kravetz h->free_huge_pages--; 14626eb4e88aSMike Kravetz h->free_huge_pages_node[nid]--; 14636eb4e88aSMike Kravetz } 14646eb4e88aSMike Kravetz if (adjust_surplus) { 14656eb4e88aSMike Kravetz h->surplus_huge_pages--; 14666eb4e88aSMike Kravetz h->surplus_huge_pages_node[nid]--; 14676eb4e88aSMike Kravetz } 14686eb4e88aSMike Kravetz 1469e32d20c0SMike Kravetz /* 1470e32d20c0SMike Kravetz * Very subtle 1471e32d20c0SMike Kravetz * 1472e32d20c0SMike Kravetz * For non-gigantic pages set the destructor to the normal compound 1473e32d20c0SMike Kravetz * page dtor. This is needed in case someone takes an additional 1474e32d20c0SMike Kravetz * temporary ref to the page, and freeing is delayed until they drop 1475e32d20c0SMike Kravetz * their reference. 1476e32d20c0SMike Kravetz * 1477e32d20c0SMike Kravetz * For gigantic pages set the destructor to the null dtor. This 1478e32d20c0SMike Kravetz * destructor will never be called. Before freeing the gigantic 1479e32d20c0SMike Kravetz * page destroy_compound_gigantic_page will turn the compound page 1480e32d20c0SMike Kravetz * into a simple group of pages. After this the destructor does not 1481e32d20c0SMike Kravetz * apply. 1482e32d20c0SMike Kravetz * 1483e32d20c0SMike Kravetz * This handles the case where more than one ref is held when and 1484e32d20c0SMike Kravetz * after update_and_free_page is called. 148534d9e35bSMike Kravetz * 148634d9e35bSMike Kravetz * In the case of demote we do not ref count the page as it will soon 148734d9e35bSMike Kravetz * be turned into a page of smaller size. 1488e32d20c0SMike Kravetz */ 148934d9e35bSMike Kravetz if (!demote) 14906eb4e88aSMike Kravetz set_page_refcounted(page); 1491e32d20c0SMike Kravetz if (hstate_is_gigantic(h)) 14926eb4e88aSMike Kravetz set_compound_page_dtor(page, NULL_COMPOUND_DTOR); 1493e32d20c0SMike Kravetz else 1494e32d20c0SMike Kravetz set_compound_page_dtor(page, COMPOUND_PAGE_DTOR); 14956eb4e88aSMike Kravetz 14966eb4e88aSMike Kravetz h->nr_huge_pages--; 14976eb4e88aSMike Kravetz h->nr_huge_pages_node[nid]--; 14986eb4e88aSMike Kravetz } 14996eb4e88aSMike Kravetz 150034d9e35bSMike Kravetz static void remove_hugetlb_page(struct hstate *h, struct page *page, 150134d9e35bSMike Kravetz bool adjust_surplus) 150234d9e35bSMike Kravetz { 150334d9e35bSMike Kravetz __remove_hugetlb_page(h, page, adjust_surplus, false); 150434d9e35bSMike Kravetz } 150534d9e35bSMike Kravetz 15068531fc6fSMike Kravetz static void remove_hugetlb_page_for_demote(struct hstate *h, struct page *page, 15078531fc6fSMike Kravetz bool adjust_surplus) 15088531fc6fSMike Kravetz { 15098531fc6fSMike Kravetz __remove_hugetlb_page(h, page, adjust_surplus, true); 15108531fc6fSMike Kravetz } 15118531fc6fSMike Kravetz 1512ad2fa371SMuchun Song static void add_hugetlb_page(struct hstate *h, struct page *page, 1513ad2fa371SMuchun Song bool adjust_surplus) 1514ad2fa371SMuchun Song { 1515ad2fa371SMuchun Song int zeroed; 1516ad2fa371SMuchun Song int nid = page_to_nid(page); 1517ad2fa371SMuchun Song 1518ad2fa371SMuchun Song VM_BUG_ON_PAGE(!HPageVmemmapOptimized(page), page); 1519ad2fa371SMuchun Song 1520ad2fa371SMuchun Song lockdep_assert_held(&hugetlb_lock); 1521ad2fa371SMuchun Song 1522ad2fa371SMuchun Song INIT_LIST_HEAD(&page->lru); 1523ad2fa371SMuchun Song h->nr_huge_pages++; 1524ad2fa371SMuchun Song h->nr_huge_pages_node[nid]++; 1525ad2fa371SMuchun Song 1526ad2fa371SMuchun Song if (adjust_surplus) { 1527ad2fa371SMuchun Song h->surplus_huge_pages++; 1528ad2fa371SMuchun Song h->surplus_huge_pages_node[nid]++; 1529ad2fa371SMuchun Song } 1530ad2fa371SMuchun Song 1531ad2fa371SMuchun Song set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); 1532ad2fa371SMuchun Song set_page_private(page, 0); 1533a9e1eab2SMiaohe Lin /* 1534a9e1eab2SMiaohe Lin * We have to set HPageVmemmapOptimized again as above 1535a9e1eab2SMiaohe Lin * set_page_private(page, 0) cleared it. 1536a9e1eab2SMiaohe Lin */ 1537ad2fa371SMuchun Song SetHPageVmemmapOptimized(page); 1538ad2fa371SMuchun Song 1539ad2fa371SMuchun Song /* 1540b65a4edaSMike Kravetz * This page is about to be managed by the hugetlb allocator and 1541b65a4edaSMike Kravetz * should have no users. Drop our reference, and check for others 1542b65a4edaSMike Kravetz * just in case. 1543ad2fa371SMuchun Song */ 1544ad2fa371SMuchun Song zeroed = put_page_testzero(page); 1545b65a4edaSMike Kravetz if (!zeroed) 1546b65a4edaSMike Kravetz /* 1547b65a4edaSMike Kravetz * It is VERY unlikely soneone else has taken a ref on 1548b65a4edaSMike Kravetz * the page. In this case, we simply return as the 1549b65a4edaSMike Kravetz * hugetlb destructor (free_huge_page) will be called 1550b65a4edaSMike Kravetz * when this other ref is dropped. 1551b65a4edaSMike Kravetz */ 1552b65a4edaSMike Kravetz return; 1553b65a4edaSMike Kravetz 1554ad2fa371SMuchun Song arch_clear_hugepage_flags(page); 1555ad2fa371SMuchun Song enqueue_huge_page(h, page); 1556ad2fa371SMuchun Song } 1557ad2fa371SMuchun Song 1558b65d4adbSMuchun Song static void __update_and_free_page(struct hstate *h, struct page *page) 15596af2acb6SAdam Litke { 15606af2acb6SAdam Litke int i; 156114455eabSCheng Li struct page *subpage; 1562a5516438SAndi Kleen 15634eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1564944d9fecSLuiz Capitulino return; 156518229df5SAndy Whitcroft 1566161df60eSNaoya Horiguchi /* 1567161df60eSNaoya Horiguchi * If we don't know which subpages are hwpoisoned, we can't free 1568161df60eSNaoya Horiguchi * the hugepage, so it's leaked intentionally. 1569161df60eSNaoya Horiguchi */ 1570161df60eSNaoya Horiguchi if (HPageRawHwpUnreliable(page)) 1571161df60eSNaoya Horiguchi return; 1572161df60eSNaoya Horiguchi 15736213834cSMuchun Song if (hugetlb_vmemmap_restore(h, page)) { 1574ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 1575ad2fa371SMuchun Song /* 1576ad2fa371SMuchun Song * If we cannot allocate vmemmap pages, just refuse to free the 1577ad2fa371SMuchun Song * page and put the page back on the hugetlb free list and treat 1578ad2fa371SMuchun Song * as a surplus page. 1579ad2fa371SMuchun Song */ 1580ad2fa371SMuchun Song add_hugetlb_page(h, page, true); 1581ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 1582ad2fa371SMuchun Song return; 1583ad2fa371SMuchun Song } 1584ad2fa371SMuchun Song 1585161df60eSNaoya Horiguchi /* 1586161df60eSNaoya Horiguchi * Move PageHWPoison flag from head page to the raw error pages, 1587161df60eSNaoya Horiguchi * which makes any healthy subpages reusable. 1588161df60eSNaoya Horiguchi */ 1589161df60eSNaoya Horiguchi if (unlikely(PageHWPoison(page))) 1590161df60eSNaoya Horiguchi hugetlb_clear_page_hwpoison(page); 1591161df60eSNaoya Horiguchi 159214455eabSCheng Li for (i = 0; i < pages_per_huge_page(h); i++) { 159314455eabSCheng Li subpage = nth_page(page, i); 1594dbfee5aeSMike Kravetz subpage->flags &= ~(1 << PG_locked | 1 << PG_error | 159532f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 1596a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 1597a7407a27SLuiz Capitulino 1 << PG_writeback); 15986af2acb6SAdam Litke } 1599a01f4390SMike Kravetz 1600a01f4390SMike Kravetz /* 1601a01f4390SMike Kravetz * Non-gigantic pages demoted from CMA allocated gigantic pages 1602a01f4390SMike Kravetz * need to be given back to CMA in free_gigantic_page. 1603a01f4390SMike Kravetz */ 1604a01f4390SMike Kravetz if (hstate_is_gigantic(h) || 1605a01f4390SMike Kravetz hugetlb_cma_page(page, huge_page_order(h))) { 1606944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 1607944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 1608944d9fecSLuiz Capitulino } else { 1609a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 16106af2acb6SAdam Litke } 1611944d9fecSLuiz Capitulino } 16126af2acb6SAdam Litke 1613b65d4adbSMuchun Song /* 1614b65d4adbSMuchun Song * As update_and_free_page() can be called under any context, so we cannot 1615b65d4adbSMuchun Song * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the 1616b65d4adbSMuchun Song * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate 1617b65d4adbSMuchun Song * the vmemmap pages. 1618b65d4adbSMuchun Song * 1619b65d4adbSMuchun Song * free_hpage_workfn() locklessly retrieves the linked list of pages to be 1620b65d4adbSMuchun Song * freed and frees them one-by-one. As the page->mapping pointer is going 1621b65d4adbSMuchun Song * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node 1622b65d4adbSMuchun Song * structure of a lockless linked list of huge pages to be freed. 1623b65d4adbSMuchun Song */ 1624b65d4adbSMuchun Song static LLIST_HEAD(hpage_freelist); 1625b65d4adbSMuchun Song 1626b65d4adbSMuchun Song static void free_hpage_workfn(struct work_struct *work) 1627b65d4adbSMuchun Song { 1628b65d4adbSMuchun Song struct llist_node *node; 1629b65d4adbSMuchun Song 1630b65d4adbSMuchun Song node = llist_del_all(&hpage_freelist); 1631b65d4adbSMuchun Song 1632b65d4adbSMuchun Song while (node) { 1633b65d4adbSMuchun Song struct page *page; 1634b65d4adbSMuchun Song struct hstate *h; 1635b65d4adbSMuchun Song 1636b65d4adbSMuchun Song page = container_of((struct address_space **)node, 1637b65d4adbSMuchun Song struct page, mapping); 1638b65d4adbSMuchun Song node = node->next; 1639b65d4adbSMuchun Song page->mapping = NULL; 1640b65d4adbSMuchun Song /* 1641b65d4adbSMuchun Song * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate() 1642b65d4adbSMuchun Song * is going to trigger because a previous call to 1643b65d4adbSMuchun Song * remove_hugetlb_page() will set_compound_page_dtor(page, 1644b65d4adbSMuchun Song * NULL_COMPOUND_DTOR), so do not use page_hstate() directly. 1645b65d4adbSMuchun Song */ 1646b65d4adbSMuchun Song h = size_to_hstate(page_size(page)); 1647b65d4adbSMuchun Song 1648b65d4adbSMuchun Song __update_and_free_page(h, page); 1649b65d4adbSMuchun Song 1650b65d4adbSMuchun Song cond_resched(); 1651b65d4adbSMuchun Song } 1652b65d4adbSMuchun Song } 1653b65d4adbSMuchun Song static DECLARE_WORK(free_hpage_work, free_hpage_workfn); 1654b65d4adbSMuchun Song 1655b65d4adbSMuchun Song static inline void flush_free_hpage_work(struct hstate *h) 1656b65d4adbSMuchun Song { 16576213834cSMuchun Song if (hugetlb_vmemmap_optimizable(h)) 1658b65d4adbSMuchun Song flush_work(&free_hpage_work); 1659b65d4adbSMuchun Song } 1660b65d4adbSMuchun Song 1661b65d4adbSMuchun Song static void update_and_free_page(struct hstate *h, struct page *page, 1662b65d4adbSMuchun Song bool atomic) 1663b65d4adbSMuchun Song { 1664ad2fa371SMuchun Song if (!HPageVmemmapOptimized(page) || !atomic) { 1665b65d4adbSMuchun Song __update_and_free_page(h, page); 1666b65d4adbSMuchun Song return; 1667b65d4adbSMuchun Song } 1668b65d4adbSMuchun Song 1669b65d4adbSMuchun Song /* 1670b65d4adbSMuchun Song * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages. 1671b65d4adbSMuchun Song * 1672b65d4adbSMuchun Song * Only call schedule_work() if hpage_freelist is previously 1673b65d4adbSMuchun Song * empty. Otherwise, schedule_work() had been called but the workfn 1674b65d4adbSMuchun Song * hasn't retrieved the list yet. 1675b65d4adbSMuchun Song */ 1676b65d4adbSMuchun Song if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist)) 1677b65d4adbSMuchun Song schedule_work(&free_hpage_work); 1678b65d4adbSMuchun Song } 1679b65d4adbSMuchun Song 168010c6ec49SMike Kravetz static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list) 168110c6ec49SMike Kravetz { 168210c6ec49SMike Kravetz struct page *page, *t_page; 168310c6ec49SMike Kravetz 168410c6ec49SMike Kravetz list_for_each_entry_safe(page, t_page, list, lru) { 1685b65d4adbSMuchun Song update_and_free_page(h, page, false); 168610c6ec49SMike Kravetz cond_resched(); 168710c6ec49SMike Kravetz } 168810c6ec49SMike Kravetz } 168910c6ec49SMike Kravetz 1690e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 1691e5ff2159SAndi Kleen { 1692e5ff2159SAndi Kleen struct hstate *h; 1693e5ff2159SAndi Kleen 1694e5ff2159SAndi Kleen for_each_hstate(h) { 1695e5ff2159SAndi Kleen if (huge_page_size(h) == size) 1696e5ff2159SAndi Kleen return h; 1697e5ff2159SAndi Kleen } 1698e5ff2159SAndi Kleen return NULL; 1699e5ff2159SAndi Kleen } 1700e5ff2159SAndi Kleen 1701db71ef79SMike Kravetz void free_huge_page(struct page *page) 170227a85ef1SDavid Gibson { 1703a5516438SAndi Kleen /* 1704a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 1705a5516438SAndi Kleen * compound page destructor. 1706a5516438SAndi Kleen */ 1707e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 17087893d1d5SAdam Litke int nid = page_to_nid(page); 1709d6995da3SMike Kravetz struct hugepage_subpool *spool = hugetlb_page_subpool(page); 171007443a85SJoonsoo Kim bool restore_reserve; 1711db71ef79SMike Kravetz unsigned long flags; 171227a85ef1SDavid Gibson 1713b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_count(page), page); 1714b4330afbSMike Kravetz VM_BUG_ON_PAGE(page_mapcount(page), page); 17158ace22bcSYongkai Wu 1716d6995da3SMike Kravetz hugetlb_set_page_subpool(page, NULL); 171778fbe906SDavid Hildenbrand if (PageAnon(page)) 171878fbe906SDavid Hildenbrand __ClearPageAnonExclusive(page); 17198ace22bcSYongkai Wu page->mapping = NULL; 1720d6995da3SMike Kravetz restore_reserve = HPageRestoreReserve(page); 1721d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 172227a85ef1SDavid Gibson 17231c5ecae3SMike Kravetz /* 1724d6995da3SMike Kravetz * If HPageRestoreReserve was set on page, page allocation consumed a 17250919e1b6SMike Kravetz * reservation. If the page was associated with a subpool, there 17260919e1b6SMike Kravetz * would have been a page reserved in the subpool before allocation 17270919e1b6SMike Kravetz * via hugepage_subpool_get_pages(). Since we are 'restoring' the 17286c26d310SMiaohe Lin * reservation, do not call hugepage_subpool_put_pages() as this will 17290919e1b6SMike Kravetz * remove the reserved page from the subpool. 17300919e1b6SMike Kravetz */ 17310919e1b6SMike Kravetz if (!restore_reserve) { 17320919e1b6SMike Kravetz /* 17330919e1b6SMike Kravetz * A return code of zero implies that the subpool will be 17340919e1b6SMike Kravetz * under its minimum size if the reservation is not restored 17350919e1b6SMike Kravetz * after page is free. Therefore, force restore_reserve 17360919e1b6SMike Kravetz * operation. 17371c5ecae3SMike Kravetz */ 17381c5ecae3SMike Kravetz if (hugepage_subpool_put_pages(spool, 1) == 0) 17391c5ecae3SMike Kravetz restore_reserve = true; 17400919e1b6SMike Kravetz } 17411c5ecae3SMike Kravetz 1742db71ef79SMike Kravetz spin_lock_irqsave(&hugetlb_lock, flags); 17438f251a3dSMike Kravetz ClearHPageMigratable(page); 17446d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 17456d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 174608cf9fafSMina Almasry hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 174708cf9fafSMina Almasry pages_per_huge_page(h), page); 174807443a85SJoonsoo Kim if (restore_reserve) 174907443a85SJoonsoo Kim h->resv_huge_pages++; 175007443a85SJoonsoo Kim 17519157c311SMike Kravetz if (HPageTemporary(page)) { 17526eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 1753db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1754b65d4adbSMuchun Song update_and_free_page(h, page, true); 1755ab5ac90aSMichal Hocko } else if (h->surplus_huge_pages_node[nid]) { 17560edaecfaSAneesh Kumar K.V /* remove the page from active list */ 17576eb4e88aSMike Kravetz remove_hugetlb_page(h, page, true); 1758db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 1759b65d4adbSMuchun Song update_and_free_page(h, page, true); 17607893d1d5SAdam Litke } else { 17615d3a551cSWill Deacon arch_clear_hugepage_flags(page); 1762a5516438SAndi Kleen enqueue_huge_page(h, page); 1763db71ef79SMike Kravetz spin_unlock_irqrestore(&hugetlb_lock, flags); 176427a85ef1SDavid Gibson } 17651121828aSMike Kravetz } 176627a85ef1SDavid Gibson 1767d3d99fccSOscar Salvador /* 1768d3d99fccSOscar Salvador * Must be called with the hugetlb lock held 1769d3d99fccSOscar Salvador */ 1770d3d99fccSOscar Salvador static void __prep_account_new_huge_page(struct hstate *h, int nid) 1771d3d99fccSOscar Salvador { 1772d3d99fccSOscar Salvador lockdep_assert_held(&hugetlb_lock); 1773d3d99fccSOscar Salvador h->nr_huge_pages++; 1774d3d99fccSOscar Salvador h->nr_huge_pages_node[nid]++; 1775d3d99fccSOscar Salvador } 1776d3d99fccSOscar Salvador 1777de656ed3SSidhartha Kumar static void __prep_new_hugetlb_folio(struct hstate *h, struct folio *folio) 1778b7ba30c6SAndi Kleen { 1779de656ed3SSidhartha Kumar hugetlb_vmemmap_optimize(h, &folio->page); 1780de656ed3SSidhartha Kumar INIT_LIST_HEAD(&folio->lru); 1781de656ed3SSidhartha Kumar folio->_folio_dtor = HUGETLB_PAGE_DTOR; 1782de656ed3SSidhartha Kumar hugetlb_set_folio_subpool(folio, NULL); 1783de656ed3SSidhartha Kumar set_hugetlb_cgroup(folio, NULL); 1784de656ed3SSidhartha Kumar set_hugetlb_cgroup_rsvd(folio, NULL); 1785d3d99fccSOscar Salvador } 1786d3d99fccSOscar Salvador 1787d3d99fccSOscar Salvador static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1788d3d99fccSOscar Salvador { 1789de656ed3SSidhartha Kumar struct folio *folio = page_folio(page); 1790de656ed3SSidhartha Kumar 1791de656ed3SSidhartha Kumar __prep_new_hugetlb_folio(h, folio); 1792db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 1793d3d99fccSOscar Salvador __prep_account_new_huge_page(h, nid); 1794db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 1795b7ba30c6SAndi Kleen } 1796b7ba30c6SAndi Kleen 179734d9e35bSMike Kravetz static bool __prep_compound_gigantic_page(struct page *page, unsigned int order, 179834d9e35bSMike Kravetz bool demote) 179920a0307cSWu Fengguang { 18007118fc29SMike Kravetz int i, j; 180120a0307cSWu Fengguang int nr_pages = 1 << order; 180214455eabSCheng Li struct page *p; 180320a0307cSWu Fengguang 180420a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 180520a0307cSWu Fengguang set_compound_order(page, order); 18067fb0728aSMike Kravetz __ClearPageReserved(page); 1807de09d31dSKirill A. Shutemov __SetPageHead(page); 18082b21624fSMike Kravetz for (i = 0; i < nr_pages; i++) { 180914455eabSCheng Li p = nth_page(page, i); 181014455eabSCheng Li 1811ef5a22beSAndrea Arcangeli /* 1812ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 1813ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 1814ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 18157c8de358SEthon Paul * too. Otherwise drivers using get_user_pages() to access tail 1816ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 1817ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 1818ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 1819ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 1820ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 1821ef5a22beSAndrea Arcangeli * after get_user_pages(). 1822ef5a22beSAndrea Arcangeli */ 18237fb0728aSMike Kravetz if (i != 0) /* head page cleared above */ 1824ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 18257118fc29SMike Kravetz /* 18267118fc29SMike Kravetz * Subtle and very unlikely 18277118fc29SMike Kravetz * 18287118fc29SMike Kravetz * Gigantic 'page allocators' such as memblock or cma will 18297118fc29SMike Kravetz * return a set of pages with each page ref counted. We need 18307118fc29SMike Kravetz * to turn this set of pages into a compound page with tail 18317118fc29SMike Kravetz * page ref counts set to zero. Code such as speculative page 18327118fc29SMike Kravetz * cache adding could take a ref on a 'to be' tail page. 18337118fc29SMike Kravetz * We need to respect any increased ref count, and only set 18347118fc29SMike Kravetz * the ref count to zero if count is currently 1. If count 1835416d85edSMike Kravetz * is not 1, we return an error. An error return indicates 1836416d85edSMike Kravetz * the set of pages can not be converted to a gigantic page. 1837416d85edSMike Kravetz * The caller who allocated the pages should then discard the 1838416d85edSMike Kravetz * pages using the appropriate free interface. 183934d9e35bSMike Kravetz * 184034d9e35bSMike Kravetz * In the case of demote, the ref count will be zero. 18417118fc29SMike Kravetz */ 184234d9e35bSMike Kravetz if (!demote) { 18437118fc29SMike Kravetz if (!page_ref_freeze(p, 1)) { 1844416d85edSMike Kravetz pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n"); 18457118fc29SMike Kravetz goto out_error; 18467118fc29SMike Kravetz } 184734d9e35bSMike Kravetz } else { 184834d9e35bSMike Kravetz VM_BUG_ON_PAGE(page_count(p), p); 184934d9e35bSMike Kravetz } 18502b21624fSMike Kravetz if (i != 0) 18511d798ca3SKirill A. Shutemov set_compound_head(p, page); 185220a0307cSWu Fengguang } 1853b4330afbSMike Kravetz atomic_set(compound_mapcount_ptr(page), -1); 185447e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 18557118fc29SMike Kravetz return true; 18567118fc29SMike Kravetz 18577118fc29SMike Kravetz out_error: 18582b21624fSMike Kravetz /* undo page modifications made above */ 18592b21624fSMike Kravetz for (j = 0; j < i; j++) { 186014455eabSCheng Li p = nth_page(page, j); 18612b21624fSMike Kravetz if (j != 0) 18627118fc29SMike Kravetz clear_compound_head(p); 18637118fc29SMike Kravetz set_page_refcounted(p); 18647118fc29SMike Kravetz } 18657118fc29SMike Kravetz /* need to clear PG_reserved on remaining tail pages */ 186614455eabSCheng Li for (; j < nr_pages; j++) { 186714455eabSCheng Li p = nth_page(page, j); 18687118fc29SMike Kravetz __ClearPageReserved(p); 186914455eabSCheng Li } 18707118fc29SMike Kravetz set_compound_order(page, 0); 18715232c63fSMatthew Wilcox (Oracle) #ifdef CONFIG_64BIT 18727118fc29SMike Kravetz page[1].compound_nr = 0; 18735232c63fSMatthew Wilcox (Oracle) #endif 18747118fc29SMike Kravetz __ClearPageHead(page); 18757118fc29SMike Kravetz return false; 187620a0307cSWu Fengguang } 187720a0307cSWu Fengguang 187834d9e35bSMike Kravetz static bool prep_compound_gigantic_page(struct page *page, unsigned int order) 187934d9e35bSMike Kravetz { 188034d9e35bSMike Kravetz return __prep_compound_gigantic_page(page, order, false); 188134d9e35bSMike Kravetz } 188234d9e35bSMike Kravetz 18838531fc6fSMike Kravetz static bool prep_compound_gigantic_page_for_demote(struct page *page, 18848531fc6fSMike Kravetz unsigned int order) 18858531fc6fSMike Kravetz { 18868531fc6fSMike Kravetz return __prep_compound_gigantic_page(page, order, true); 18878531fc6fSMike Kravetz } 18888531fc6fSMike Kravetz 18897795912cSAndrew Morton /* 18907795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 18917795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 18927795912cSAndrew Morton * details. 18937795912cSAndrew Morton */ 189420a0307cSWu Fengguang int PageHuge(struct page *page) 189520a0307cSWu Fengguang { 189620a0307cSWu Fengguang if (!PageCompound(page)) 189720a0307cSWu Fengguang return 0; 189820a0307cSWu Fengguang 189920a0307cSWu Fengguang page = compound_head(page); 1900f1e61557SKirill A. Shutemov return page[1].compound_dtor == HUGETLB_PAGE_DTOR; 190120a0307cSWu Fengguang } 190243131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 190343131e14SNaoya Horiguchi 190427c73ae7SAndrea Arcangeli /* 190527c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 190627c73ae7SAndrea Arcangeli * normal or transparent huge pages. 190727c73ae7SAndrea Arcangeli */ 190827c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 190927c73ae7SAndrea Arcangeli { 191027c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 191127c73ae7SAndrea Arcangeli return 0; 191227c73ae7SAndrea Arcangeli 1913d4af73e3SVlastimil Babka return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; 191427c73ae7SAndrea Arcangeli } 19154e936eccSDavid Howells EXPORT_SYMBOL_GPL(PageHeadHuge); 191627c73ae7SAndrea Arcangeli 1917c0d0381aSMike Kravetz /* 1918c0d0381aSMike Kravetz * Find and lock address space (mapping) in write mode. 1919c0d0381aSMike Kravetz * 1920336bf30eSMike Kravetz * Upon entry, the page is locked which means that page_mapping() is 1921336bf30eSMike Kravetz * stable. Due to locking order, we can only trylock_write. If we can 1922336bf30eSMike Kravetz * not get the lock, simply return NULL to caller. 1923c0d0381aSMike Kravetz */ 1924c0d0381aSMike Kravetz struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) 1925c0d0381aSMike Kravetz { 1926336bf30eSMike Kravetz struct address_space *mapping = page_mapping(hpage); 1927c0d0381aSMike Kravetz 1928c0d0381aSMike Kravetz if (!mapping) 1929c0d0381aSMike Kravetz return mapping; 1930c0d0381aSMike Kravetz 1931c0d0381aSMike Kravetz if (i_mmap_trylock_write(mapping)) 1932c0d0381aSMike Kravetz return mapping; 1933c0d0381aSMike Kravetz 1934c0d0381aSMike Kravetz return NULL; 1935c0d0381aSMike Kravetz } 1936c0d0381aSMike Kravetz 1937fe19bd3dSHugh Dickins pgoff_t hugetlb_basepage_index(struct page *page) 193813d60f4bSZhang Yi { 193913d60f4bSZhang Yi struct page *page_head = compound_head(page); 194013d60f4bSZhang Yi pgoff_t index = page_index(page_head); 194113d60f4bSZhang Yi unsigned long compound_idx; 194213d60f4bSZhang Yi 194313d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 194413d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 194513d60f4bSZhang Yi else 194613d60f4bSZhang Yi compound_idx = page - page_head; 194713d60f4bSZhang Yi 194813d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 194913d60f4bSZhang Yi } 195013d60f4bSZhang Yi 19510c397daeSMichal Hocko static struct page *alloc_buddy_huge_page(struct hstate *h, 1952f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 1953f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 19541da177e4SLinus Torvalds { 1955af0fb9dfSMichal Hocko int order = huge_page_order(h); 19561da177e4SLinus Torvalds struct page *page; 1957f60858f9SMike Kravetz bool alloc_try_hard = true; 19582b21624fSMike Kravetz bool retry = true; 1959f96efd58SJoe Jin 1960f60858f9SMike Kravetz /* 1961f60858f9SMike Kravetz * By default we always try hard to allocate the page with 1962f60858f9SMike Kravetz * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in 1963f60858f9SMike Kravetz * a loop (to adjust global huge page counts) and previous allocation 1964f60858f9SMike Kravetz * failed, do not continue to try hard on the same node. Use the 1965f60858f9SMike Kravetz * node_alloc_noretry bitmap to manage this state information. 1966f60858f9SMike Kravetz */ 1967f60858f9SMike Kravetz if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) 1968f60858f9SMike Kravetz alloc_try_hard = false; 1969f60858f9SMike Kravetz gfp_mask |= __GFP_COMP|__GFP_NOWARN; 1970f60858f9SMike Kravetz if (alloc_try_hard) 1971f60858f9SMike Kravetz gfp_mask |= __GFP_RETRY_MAYFAIL; 1972af0fb9dfSMichal Hocko if (nid == NUMA_NO_NODE) 1973af0fb9dfSMichal Hocko nid = numa_mem_id(); 19742b21624fSMike Kravetz retry: 197584172f4bSMatthew Wilcox (Oracle) page = __alloc_pages(gfp_mask, order, nid, nmask); 19762b21624fSMike Kravetz 19772b21624fSMike Kravetz /* Freeze head page */ 19782b21624fSMike Kravetz if (page && !page_ref_freeze(page, 1)) { 19792b21624fSMike Kravetz __free_pages(page, order); 19802b21624fSMike Kravetz if (retry) { /* retry once */ 19812b21624fSMike Kravetz retry = false; 19822b21624fSMike Kravetz goto retry; 19832b21624fSMike Kravetz } 19842b21624fSMike Kravetz /* WOW! twice in a row. */ 19852b21624fSMike Kravetz pr_warn("HugeTLB head page unexpected inflated ref count\n"); 19862b21624fSMike Kravetz page = NULL; 19872b21624fSMike Kravetz } 19882b21624fSMike Kravetz 1989af0fb9dfSMichal Hocko if (page) 1990af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC); 1991af0fb9dfSMichal Hocko else 1992af0fb9dfSMichal Hocko __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 199363b4613cSNishanth Aravamudan 1994f60858f9SMike Kravetz /* 1995f60858f9SMike Kravetz * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this 1996f60858f9SMike Kravetz * indicates an overall state change. Clear bit so that we resume 1997f60858f9SMike Kravetz * normal 'try hard' allocations. 1998f60858f9SMike Kravetz */ 1999f60858f9SMike Kravetz if (node_alloc_noretry && page && !alloc_try_hard) 2000f60858f9SMike Kravetz node_clear(nid, *node_alloc_noretry); 2001f60858f9SMike Kravetz 2002f60858f9SMike Kravetz /* 2003f60858f9SMike Kravetz * If we tried hard to get a page but failed, set bit so that 2004f60858f9SMike Kravetz * subsequent attempts will not try as hard until there is an 2005f60858f9SMike Kravetz * overall state change. 2006f60858f9SMike Kravetz */ 2007f60858f9SMike Kravetz if (node_alloc_noretry && !page && alloc_try_hard) 2008f60858f9SMike Kravetz node_set(nid, *node_alloc_noretry); 2009f60858f9SMike Kravetz 201063b4613cSNishanth Aravamudan return page; 201163b4613cSNishanth Aravamudan } 201263b4613cSNishanth Aravamudan 2013af0fb9dfSMichal Hocko /* 20140c397daeSMichal Hocko * Common helper to allocate a fresh hugetlb page. All specific allocators 20150c397daeSMichal Hocko * should use this function to get new hugetlb pages 20162b21624fSMike Kravetz * 20172b21624fSMike Kravetz * Note that returned page is 'frozen': ref count of head page and all tail 20182b21624fSMike Kravetz * pages is zero. 20190c397daeSMichal Hocko */ 20200c397daeSMichal Hocko static struct page *alloc_fresh_huge_page(struct hstate *h, 2021f60858f9SMike Kravetz gfp_t gfp_mask, int nid, nodemask_t *nmask, 2022f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 20230c397daeSMichal Hocko { 20240c397daeSMichal Hocko struct page *page; 20257118fc29SMike Kravetz bool retry = false; 20260c397daeSMichal Hocko 20277118fc29SMike Kravetz retry: 20280c397daeSMichal Hocko if (hstate_is_gigantic(h)) 20290c397daeSMichal Hocko page = alloc_gigantic_page(h, gfp_mask, nid, nmask); 20300c397daeSMichal Hocko else 20310c397daeSMichal Hocko page = alloc_buddy_huge_page(h, gfp_mask, 2032f60858f9SMike Kravetz nid, nmask, node_alloc_noretry); 20330c397daeSMichal Hocko if (!page) 20340c397daeSMichal Hocko return NULL; 20350c397daeSMichal Hocko 20367118fc29SMike Kravetz if (hstate_is_gigantic(h)) { 20377118fc29SMike Kravetz if (!prep_compound_gigantic_page(page, huge_page_order(h))) { 20387118fc29SMike Kravetz /* 20397118fc29SMike Kravetz * Rare failure to convert pages to compound page. 20407118fc29SMike Kravetz * Free pages and try again - ONCE! 20417118fc29SMike Kravetz */ 20427118fc29SMike Kravetz free_gigantic_page(page, huge_page_order(h)); 20437118fc29SMike Kravetz if (!retry) { 20447118fc29SMike Kravetz retry = true; 20457118fc29SMike Kravetz goto retry; 20467118fc29SMike Kravetz } 20477118fc29SMike Kravetz return NULL; 20487118fc29SMike Kravetz } 20497118fc29SMike Kravetz } 20500c397daeSMichal Hocko prep_new_huge_page(h, page, page_to_nid(page)); 20510c397daeSMichal Hocko 20520c397daeSMichal Hocko return page; 20530c397daeSMichal Hocko } 20540c397daeSMichal Hocko 20550c397daeSMichal Hocko /* 2056af0fb9dfSMichal Hocko * Allocates a fresh page to the hugetlb allocator pool in the node interleaved 2057af0fb9dfSMichal Hocko * manner. 2058af0fb9dfSMichal Hocko */ 2059f60858f9SMike Kravetz static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 2060f60858f9SMike Kravetz nodemask_t *node_alloc_noretry) 2061b2261026SJoonsoo Kim { 2062b2261026SJoonsoo Kim struct page *page; 2063b2261026SJoonsoo Kim int nr_nodes, node; 2064af0fb9dfSMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2065b2261026SJoonsoo Kim 2066b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 2067f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, 2068f60858f9SMike Kravetz node_alloc_noretry); 2069af0fb9dfSMichal Hocko if (page) 2070b2261026SJoonsoo Kim break; 2071b2261026SJoonsoo Kim } 2072b2261026SJoonsoo Kim 2073af0fb9dfSMichal Hocko if (!page) 2074af0fb9dfSMichal Hocko return 0; 2075b2261026SJoonsoo Kim 20762b21624fSMike Kravetz free_huge_page(page); /* free it into the hugepage allocator */ 2077af0fb9dfSMichal Hocko 2078af0fb9dfSMichal Hocko return 1; 2079b2261026SJoonsoo Kim } 2080b2261026SJoonsoo Kim 2081e8c5c824SLee Schermerhorn /* 208210c6ec49SMike Kravetz * Remove huge page from pool from next node to free. Attempt to keep 208310c6ec49SMike Kravetz * persistent huge pages more or less balanced over allowed nodes. 208410c6ec49SMike Kravetz * This routine only 'removes' the hugetlb page. The caller must make 208510c6ec49SMike Kravetz * an additional call to free the page to low level allocators. 2086e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 2087e8c5c824SLee Schermerhorn */ 208810c6ec49SMike Kravetz static struct page *remove_pool_huge_page(struct hstate *h, 208910c6ec49SMike Kravetz nodemask_t *nodes_allowed, 20906ae11b27SLee Schermerhorn bool acct_surplus) 2091e8c5c824SLee Schermerhorn { 2092b2261026SJoonsoo Kim int nr_nodes, node; 209310c6ec49SMike Kravetz struct page *page = NULL; 2094e8c5c824SLee Schermerhorn 20959487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2096b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2097685f3457SLee Schermerhorn /* 2098685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 2099685f3457SLee Schermerhorn * nodes with surplus pages. 2100685f3457SLee Schermerhorn */ 2101b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 2102b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 210310c6ec49SMike Kravetz page = list_entry(h->hugepage_freelists[node].next, 2104e8c5c824SLee Schermerhorn struct page, lru); 21056eb4e88aSMike Kravetz remove_hugetlb_page(h, page, acct_surplus); 21069a76db09SLee Schermerhorn break; 2107e8c5c824SLee Schermerhorn } 2108b2261026SJoonsoo Kim } 2109e8c5c824SLee Schermerhorn 211010c6ec49SMike Kravetz return page; 2111e8c5c824SLee Schermerhorn } 2112e8c5c824SLee Schermerhorn 2113c8721bbbSNaoya Horiguchi /* 2114c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 2115faf53defSNaoya Horiguchi * nothing for in-use hugepages and non-hugepages. 2116faf53defSNaoya Horiguchi * This function returns values like below: 2117faf53defSNaoya Horiguchi * 2118ad2fa371SMuchun Song * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages 2119ad2fa371SMuchun Song * when the system is under memory pressure and the feature of 2120ad2fa371SMuchun Song * freeing unused vmemmap pages associated with each hugetlb page 2121ad2fa371SMuchun Song * is enabled. 2122faf53defSNaoya Horiguchi * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 2123faf53defSNaoya Horiguchi * (allocated or reserved.) 2124faf53defSNaoya Horiguchi * 0: successfully dissolved free hugepages or the page is not a 2125faf53defSNaoya Horiguchi * hugepage (considered as already dissolved) 2126c8721bbbSNaoya Horiguchi */ 2127c3114a84SAnshuman Khandual int dissolve_free_huge_page(struct page *page) 2128c8721bbbSNaoya Horiguchi { 21296bc9b564SNaoya Horiguchi int rc = -EBUSY; 2130082d5b6bSGerald Schaefer 21317ffddd49SMuchun Song retry: 2132faf53defSNaoya Horiguchi /* Not to disrupt normal path by vainly holding hugetlb_lock */ 2133faf53defSNaoya Horiguchi if (!PageHuge(page)) 2134faf53defSNaoya Horiguchi return 0; 2135faf53defSNaoya Horiguchi 2136db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2137faf53defSNaoya Horiguchi if (!PageHuge(page)) { 2138faf53defSNaoya Horiguchi rc = 0; 2139faf53defSNaoya Horiguchi goto out; 2140faf53defSNaoya Horiguchi } 2141faf53defSNaoya Horiguchi 2142faf53defSNaoya Horiguchi if (!page_count(page)) { 21432247bb33SGerald Schaefer struct page *head = compound_head(page); 21442247bb33SGerald Schaefer struct hstate *h = page_hstate(head); 21458346d69dSXin Hao if (!available_huge_pages(h)) 2146082d5b6bSGerald Schaefer goto out; 21477ffddd49SMuchun Song 21487ffddd49SMuchun Song /* 21497ffddd49SMuchun Song * We should make sure that the page is already on the free list 21507ffddd49SMuchun Song * when it is dissolved. 21517ffddd49SMuchun Song */ 21526c037149SMike Kravetz if (unlikely(!HPageFreed(head))) { 2153db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 21547ffddd49SMuchun Song cond_resched(); 21557ffddd49SMuchun Song 21567ffddd49SMuchun Song /* 21577ffddd49SMuchun Song * Theoretically, we should return -EBUSY when we 21587ffddd49SMuchun Song * encounter this race. In fact, we have a chance 21597ffddd49SMuchun Song * to successfully dissolve the page if we do a 21607ffddd49SMuchun Song * retry. Because the race window is quite small. 21617ffddd49SMuchun Song * If we seize this opportunity, it is an optimization 21627ffddd49SMuchun Song * for increasing the success rate of dissolving page. 21637ffddd49SMuchun Song */ 21647ffddd49SMuchun Song goto retry; 21657ffddd49SMuchun Song } 21667ffddd49SMuchun Song 2167ad2fa371SMuchun Song remove_hugetlb_page(h, head, false); 2168ad2fa371SMuchun Song h->max_huge_pages--; 2169ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2170ad2fa371SMuchun Song 2171c3114a84SAnshuman Khandual /* 2172ad2fa371SMuchun Song * Normally update_and_free_page will allocate required vmemmmap 2173ad2fa371SMuchun Song * before freeing the page. update_and_free_page will fail to 2174ad2fa371SMuchun Song * free the page if it can not allocate required vmemmap. We 2175ad2fa371SMuchun Song * need to adjust max_huge_pages if the page is not freed. 2176ad2fa371SMuchun Song * Attempt to allocate vmemmmap here so that we can take 2177ad2fa371SMuchun Song * appropriate action on failure. 2178ad2fa371SMuchun Song */ 21796213834cSMuchun Song rc = hugetlb_vmemmap_restore(h, head); 2180ad2fa371SMuchun Song if (!rc) { 2181b65d4adbSMuchun Song update_and_free_page(h, head, false); 2182ad2fa371SMuchun Song } else { 2183ad2fa371SMuchun Song spin_lock_irq(&hugetlb_lock); 2184ad2fa371SMuchun Song add_hugetlb_page(h, head, false); 2185ad2fa371SMuchun Song h->max_huge_pages++; 2186ad2fa371SMuchun Song spin_unlock_irq(&hugetlb_lock); 2187ad2fa371SMuchun Song } 2188ad2fa371SMuchun Song 2189ad2fa371SMuchun Song return rc; 2190c8721bbbSNaoya Horiguchi } 2191082d5b6bSGerald Schaefer out: 2192db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2193082d5b6bSGerald Schaefer return rc; 2194c8721bbbSNaoya Horiguchi } 2195c8721bbbSNaoya Horiguchi 2196c8721bbbSNaoya Horiguchi /* 2197c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 2198c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 21992247bb33SGerald Schaefer * Note that this will dissolve a free gigantic hugepage completely, if any 22002247bb33SGerald Schaefer * part of it lies within the given range. 2201082d5b6bSGerald Schaefer * Also note that if dissolve_free_huge_page() returns with an error, all 2202082d5b6bSGerald Schaefer * free hugepages that were dissolved before that error are lost. 2203c8721bbbSNaoya Horiguchi */ 2204082d5b6bSGerald Schaefer int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 2205c8721bbbSNaoya Horiguchi { 2206c8721bbbSNaoya Horiguchi unsigned long pfn; 2207eb03aa00SGerald Schaefer struct page *page; 2208082d5b6bSGerald Schaefer int rc = 0; 2209dc2628f3SMuchun Song unsigned int order; 2210dc2628f3SMuchun Song struct hstate *h; 2211c8721bbbSNaoya Horiguchi 2212d0177639SLi Zhong if (!hugepages_supported()) 2213082d5b6bSGerald Schaefer return rc; 2214d0177639SLi Zhong 2215dc2628f3SMuchun Song order = huge_page_order(&default_hstate); 2216dc2628f3SMuchun Song for_each_hstate(h) 2217dc2628f3SMuchun Song order = min(order, huge_page_order(h)); 2218dc2628f3SMuchun Song 2219dc2628f3SMuchun Song for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) { 2220eb03aa00SGerald Schaefer page = pfn_to_page(pfn); 2221eb03aa00SGerald Schaefer rc = dissolve_free_huge_page(page); 2222eb03aa00SGerald Schaefer if (rc) 2223082d5b6bSGerald Schaefer break; 2224eb03aa00SGerald Schaefer } 2225082d5b6bSGerald Schaefer 2226082d5b6bSGerald Schaefer return rc; 2227c8721bbbSNaoya Horiguchi } 2228c8721bbbSNaoya Horiguchi 2229ab5ac90aSMichal Hocko /* 2230ab5ac90aSMichal Hocko * Allocates a fresh surplus page from the page allocator. 2231ab5ac90aSMichal Hocko */ 22320c397daeSMichal Hocko static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, 22332b21624fSMike Kravetz int nid, nodemask_t *nmask) 22347893d1d5SAdam Litke { 22359980d744SMichal Hocko struct page *page = NULL; 22367893d1d5SAdam Litke 2237bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 2238aa888a74SAndi Kleen return NULL; 2239aa888a74SAndi Kleen 2240db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 22419980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 22429980d744SMichal Hocko goto out_unlock; 2243db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2244d1c3fb1fSNishanth Aravamudan 2245f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 22469980d744SMichal Hocko if (!page) 22470c397daeSMichal Hocko return NULL; 2248d1c3fb1fSNishanth Aravamudan 2249db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 22509980d744SMichal Hocko /* 22519980d744SMichal Hocko * We could have raced with the pool size change. 22529980d744SMichal Hocko * Double check that and simply deallocate the new page 22539980d744SMichal Hocko * if we would end up overcommiting the surpluses. Abuse 22549980d744SMichal Hocko * temporary page to workaround the nasty free_huge_page 22559980d744SMichal Hocko * codeflow 22569980d744SMichal Hocko */ 22579980d744SMichal Hocko if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 22589157c311SMike Kravetz SetHPageTemporary(page); 2259db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 22602b21624fSMike Kravetz free_huge_page(page); 22612bf753e6SKai Shen return NULL; 2262b65a4edaSMike Kravetz } 2263b65a4edaSMike Kravetz 22649980d744SMichal Hocko h->surplus_huge_pages++; 22654704dea3SMichal Hocko h->surplus_huge_pages_node[page_to_nid(page)]++; 22669980d744SMichal Hocko 22679980d744SMichal Hocko out_unlock: 2268db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 22697893d1d5SAdam Litke 22707893d1d5SAdam Litke return page; 22717893d1d5SAdam Litke } 22727893d1d5SAdam Litke 2273bbe88753SJoonsoo Kim static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, 2274ab5ac90aSMichal Hocko int nid, nodemask_t *nmask) 2275ab5ac90aSMichal Hocko { 2276ab5ac90aSMichal Hocko struct page *page; 2277ab5ac90aSMichal Hocko 2278ab5ac90aSMichal Hocko if (hstate_is_gigantic(h)) 2279ab5ac90aSMichal Hocko return NULL; 2280ab5ac90aSMichal Hocko 2281f60858f9SMike Kravetz page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); 2282ab5ac90aSMichal Hocko if (!page) 2283ab5ac90aSMichal Hocko return NULL; 2284ab5ac90aSMichal Hocko 22852b21624fSMike Kravetz /* fresh huge pages are frozen */ 22862b21624fSMike Kravetz set_page_refcounted(page); 22872b21624fSMike Kravetz 2288ab5ac90aSMichal Hocko /* 2289ab5ac90aSMichal Hocko * We do not account these pages as surplus because they are only 2290ab5ac90aSMichal Hocko * temporary and will be released properly on the last reference 2291ab5ac90aSMichal Hocko */ 22929157c311SMike Kravetz SetHPageTemporary(page); 2293ab5ac90aSMichal Hocko 2294ab5ac90aSMichal Hocko return page; 2295ab5ac90aSMichal Hocko } 2296ab5ac90aSMichal Hocko 2297e4e574b7SAdam Litke /* 2298099730d6SDave Hansen * Use the VMA's mpolicy to allocate a huge page from the buddy. 2299099730d6SDave Hansen */ 2300e0ec90eeSDave Hansen static 23010c397daeSMichal Hocko struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, 2302099730d6SDave Hansen struct vm_area_struct *vma, unsigned long addr) 2303099730d6SDave Hansen { 2304cfcaa66fSBen Widawsky struct page *page = NULL; 2305aaf14e40SMichal Hocko struct mempolicy *mpol; 2306aaf14e40SMichal Hocko gfp_t gfp_mask = htlb_alloc_mask(h); 2307aaf14e40SMichal Hocko int nid; 2308aaf14e40SMichal Hocko nodemask_t *nodemask; 2309aaf14e40SMichal Hocko 2310aaf14e40SMichal Hocko nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 2311cfcaa66fSBen Widawsky if (mpol_is_preferred_many(mpol)) { 2312cfcaa66fSBen Widawsky gfp_t gfp = gfp_mask | __GFP_NOWARN; 2313cfcaa66fSBen Widawsky 2314cfcaa66fSBen Widawsky gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL); 23152b21624fSMike Kravetz page = alloc_surplus_huge_page(h, gfp, nid, nodemask); 2316cfcaa66fSBen Widawsky 2317cfcaa66fSBen Widawsky /* Fallback to all nodes if page==NULL */ 2318cfcaa66fSBen Widawsky nodemask = NULL; 2319cfcaa66fSBen Widawsky } 2320cfcaa66fSBen Widawsky 2321cfcaa66fSBen Widawsky if (!page) 23222b21624fSMike Kravetz page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); 2323aaf14e40SMichal Hocko mpol_cond_put(mpol); 2324aaf14e40SMichal Hocko return page; 2325099730d6SDave Hansen } 2326099730d6SDave Hansen 2327ab5ac90aSMichal Hocko /* page migration callback function */ 23283e59fcb0SMichal Hocko struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, 2329d92bbc27SJoonsoo Kim nodemask_t *nmask, gfp_t gfp_mask) 23304db9b2efSMichal Hocko { 2331db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 23328346d69dSXin Hao if (available_huge_pages(h)) { 23333e59fcb0SMichal Hocko struct page *page; 23343e59fcb0SMichal Hocko 23353e59fcb0SMichal Hocko page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); 23363e59fcb0SMichal Hocko if (page) { 2337db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 23383e59fcb0SMichal Hocko return page; 23394db9b2efSMichal Hocko } 23404db9b2efSMichal Hocko } 2341db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 23424db9b2efSMichal Hocko 23430c397daeSMichal Hocko return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); 23444db9b2efSMichal Hocko } 23454db9b2efSMichal Hocko 2346ebd63723SMichal Hocko /* mempolicy aware migration callback */ 2347389c8178SMichal Hocko struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, 2348389c8178SMichal Hocko unsigned long address) 2349ebd63723SMichal Hocko { 2350ebd63723SMichal Hocko struct mempolicy *mpol; 2351ebd63723SMichal Hocko nodemask_t *nodemask; 2352ebd63723SMichal Hocko struct page *page; 2353ebd63723SMichal Hocko gfp_t gfp_mask; 2354ebd63723SMichal Hocko int node; 2355ebd63723SMichal Hocko 2356ebd63723SMichal Hocko gfp_mask = htlb_alloc_mask(h); 2357ebd63723SMichal Hocko node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 2358d92bbc27SJoonsoo Kim page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); 2359ebd63723SMichal Hocko mpol_cond_put(mpol); 2360ebd63723SMichal Hocko 2361ebd63723SMichal Hocko return page; 2362ebd63723SMichal Hocko } 2363ebd63723SMichal Hocko 2364bf50bab2SNaoya Horiguchi /* 236525985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 2366e4e574b7SAdam Litke * of size 'delta'. 2367e4e574b7SAdam Litke */ 23680a4f3d1bSLiu Xiang static int gather_surplus_pages(struct hstate *h, long delta) 23691b2a1e7bSJules Irenge __must_hold(&hugetlb_lock) 2370e4e574b7SAdam Litke { 237134665341SMiaohe Lin LIST_HEAD(surplus_list); 2372e4e574b7SAdam Litke struct page *page, *tmp; 23730a4f3d1bSLiu Xiang int ret; 23740a4f3d1bSLiu Xiang long i; 23750a4f3d1bSLiu Xiang long needed, allocated; 237628073b02SHillf Danton bool alloc_ok = true; 2377e4e574b7SAdam Litke 23789487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 2379a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 2380ac09b3a1SAdam Litke if (needed <= 0) { 2381a5516438SAndi Kleen h->resv_huge_pages += delta; 2382e4e574b7SAdam Litke return 0; 2383ac09b3a1SAdam Litke } 2384e4e574b7SAdam Litke 2385e4e574b7SAdam Litke allocated = 0; 2386e4e574b7SAdam Litke 2387e4e574b7SAdam Litke ret = -ENOMEM; 2388e4e574b7SAdam Litke retry: 2389db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2390e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 23910c397daeSMichal Hocko page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), 23922b21624fSMike Kravetz NUMA_NO_NODE, NULL); 239328073b02SHillf Danton if (!page) { 239428073b02SHillf Danton alloc_ok = false; 239528073b02SHillf Danton break; 239628073b02SHillf Danton } 2397e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 239869ed779aSDavid Rientjes cond_resched(); 2399e4e574b7SAdam Litke } 240028073b02SHillf Danton allocated += i; 2401e4e574b7SAdam Litke 2402e4e574b7SAdam Litke /* 2403e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 2404e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 2405e4e574b7SAdam Litke */ 2406db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2407a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 2408a5516438SAndi Kleen (h->free_huge_pages + allocated); 240928073b02SHillf Danton if (needed > 0) { 241028073b02SHillf Danton if (alloc_ok) 2411e4e574b7SAdam Litke goto retry; 241228073b02SHillf Danton /* 241328073b02SHillf Danton * We were not able to allocate enough pages to 241428073b02SHillf Danton * satisfy the entire reservation so we free what 241528073b02SHillf Danton * we've allocated so far. 241628073b02SHillf Danton */ 241728073b02SHillf Danton goto free; 241828073b02SHillf Danton } 2419e4e574b7SAdam Litke /* 2420e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 242125985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 2422e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 2423ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 2424ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 2425ac09b3a1SAdam Litke * before they are reserved. 2426e4e574b7SAdam Litke */ 2427e4e574b7SAdam Litke needed += allocated; 2428a5516438SAndi Kleen h->resv_huge_pages += delta; 2429e4e574b7SAdam Litke ret = 0; 2430a9869b83SNaoya Horiguchi 243119fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 243219fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 243319fc3f0aSAdam Litke if ((--needed) < 0) 243419fc3f0aSAdam Litke break; 2435b65a4edaSMike Kravetz /* Add the page to the hugetlb allocator */ 2436a5516438SAndi Kleen enqueue_huge_page(h, page); 243719fc3f0aSAdam Litke } 243828073b02SHillf Danton free: 2439db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 244019fc3f0aSAdam Litke 2441b65a4edaSMike Kravetz /* 2442b65a4edaSMike Kravetz * Free unnecessary surplus pages to the buddy allocator. 2443b65a4edaSMike Kravetz * Pages have no ref count, call free_huge_page directly. 2444b65a4edaSMike Kravetz */ 2445c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 2446b65a4edaSMike Kravetz free_huge_page(page); 2447db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2448e4e574b7SAdam Litke 2449e4e574b7SAdam Litke return ret; 2450e4e574b7SAdam Litke } 2451e4e574b7SAdam Litke 2452e4e574b7SAdam Litke /* 2453e5bbc8a6SMike Kravetz * This routine has two main purposes: 2454e5bbc8a6SMike Kravetz * 1) Decrement the reservation count (resv_huge_pages) by the value passed 2455e5bbc8a6SMike Kravetz * in unused_resv_pages. This corresponds to the prior adjustments made 2456e5bbc8a6SMike Kravetz * to the associated reservation map. 2457e5bbc8a6SMike Kravetz * 2) Free any unused surplus pages that may have been allocated to satisfy 2458e5bbc8a6SMike Kravetz * the reservation. As many as unused_resv_pages may be freed. 2459e4e574b7SAdam Litke */ 2460a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 2461a5516438SAndi Kleen unsigned long unused_resv_pages) 2462e4e574b7SAdam Litke { 2463e4e574b7SAdam Litke unsigned long nr_pages; 246410c6ec49SMike Kravetz struct page *page; 246510c6ec49SMike Kravetz LIST_HEAD(page_list); 246610c6ec49SMike Kravetz 24679487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 246810c6ec49SMike Kravetz /* Uncommit the reservation */ 246910c6ec49SMike Kravetz h->resv_huge_pages -= unused_resv_pages; 2470e4e574b7SAdam Litke 2471c0531714SNaoya Horiguchi if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 2472e5bbc8a6SMike Kravetz goto out; 2473aa888a74SAndi Kleen 2474e5bbc8a6SMike Kravetz /* 2475e5bbc8a6SMike Kravetz * Part (or even all) of the reservation could have been backed 2476e5bbc8a6SMike Kravetz * by pre-allocated pages. Only free surplus pages. 2477e5bbc8a6SMike Kravetz */ 2478a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 2479e4e574b7SAdam Litke 2480685f3457SLee Schermerhorn /* 2481685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 24829b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 24839b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 24849b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 248510c6ec49SMike Kravetz * remove_pool_huge_page() will balance the freed pages across the 24869b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 2487685f3457SLee Schermerhorn */ 2488685f3457SLee Schermerhorn while (nr_pages--) { 248910c6ec49SMike Kravetz page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1); 249010c6ec49SMike Kravetz if (!page) 2491e5bbc8a6SMike Kravetz goto out; 249210c6ec49SMike Kravetz 249310c6ec49SMike Kravetz list_add(&page->lru, &page_list); 2494e4e574b7SAdam Litke } 2495e5bbc8a6SMike Kravetz 2496e5bbc8a6SMike Kravetz out: 2497db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 249810c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 2499db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2500e4e574b7SAdam Litke } 2501e4e574b7SAdam Litke 25025e911373SMike Kravetz 2503c37f9fb1SAndy Whitcroft /* 2504feba16e2SMike Kravetz * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 25055e911373SMike Kravetz * are used by the huge page allocation routines to manage reservations. 2506cf3ad20bSMike Kravetz * 2507cf3ad20bSMike Kravetz * vma_needs_reservation is called to determine if the huge page at addr 2508cf3ad20bSMike Kravetz * within the vma has an associated reservation. If a reservation is 2509cf3ad20bSMike Kravetz * needed, the value 1 is returned. The caller is then responsible for 2510cf3ad20bSMike Kravetz * managing the global reservation and subpool usage counts. After 2511cf3ad20bSMike Kravetz * the huge page has been allocated, vma_commit_reservation is called 2512feba16e2SMike Kravetz * to add the page to the reservation map. If the page allocation fails, 2513feba16e2SMike Kravetz * the reservation must be ended instead of committed. vma_end_reservation 2514feba16e2SMike Kravetz * is called in such cases. 2515cf3ad20bSMike Kravetz * 2516cf3ad20bSMike Kravetz * In the normal case, vma_commit_reservation returns the same value 2517cf3ad20bSMike Kravetz * as the preceding vma_needs_reservation call. The only time this 2518cf3ad20bSMike Kravetz * is not the case is if a reserve map was changed between calls. It 2519cf3ad20bSMike Kravetz * is the responsibility of the caller to notice the difference and 2520cf3ad20bSMike Kravetz * take appropriate action. 252196b96a96SMike Kravetz * 252296b96a96SMike Kravetz * vma_add_reservation is used in error paths where a reservation must 252396b96a96SMike Kravetz * be restored when a newly allocated huge page must be freed. It is 252496b96a96SMike Kravetz * to be called after calling vma_needs_reservation to determine if a 252596b96a96SMike Kravetz * reservation exists. 2526846be085SMike Kravetz * 2527846be085SMike Kravetz * vma_del_reservation is used in error paths where an entry in the reserve 2528846be085SMike Kravetz * map was created during huge page allocation and must be removed. It is to 2529846be085SMike Kravetz * be called after calling vma_needs_reservation to determine if a reservation 2530846be085SMike Kravetz * exists. 2531c37f9fb1SAndy Whitcroft */ 25325e911373SMike Kravetz enum vma_resv_mode { 25335e911373SMike Kravetz VMA_NEEDS_RESV, 25345e911373SMike Kravetz VMA_COMMIT_RESV, 2535feba16e2SMike Kravetz VMA_END_RESV, 253696b96a96SMike Kravetz VMA_ADD_RESV, 2537846be085SMike Kravetz VMA_DEL_RESV, 25385e911373SMike Kravetz }; 2539cf3ad20bSMike Kravetz static long __vma_reservation_common(struct hstate *h, 2540cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr, 25415e911373SMike Kravetz enum vma_resv_mode mode) 2542c37f9fb1SAndy Whitcroft { 25434e35f483SJoonsoo Kim struct resv_map *resv; 25444e35f483SJoonsoo Kim pgoff_t idx; 2545cf3ad20bSMike Kravetz long ret; 25460db9d74eSMina Almasry long dummy_out_regions_needed; 2547c37f9fb1SAndy Whitcroft 25484e35f483SJoonsoo Kim resv = vma_resv_map(vma); 25494e35f483SJoonsoo Kim if (!resv) 2550c37f9fb1SAndy Whitcroft return 1; 2551c37f9fb1SAndy Whitcroft 25524e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 25535e911373SMike Kravetz switch (mode) { 25545e911373SMike Kravetz case VMA_NEEDS_RESV: 25550db9d74eSMina Almasry ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); 25560db9d74eSMina Almasry /* We assume that vma_reservation_* routines always operate on 25570db9d74eSMina Almasry * 1 page, and that adding to resv map a 1 page entry can only 25580db9d74eSMina Almasry * ever require 1 region. 25590db9d74eSMina Almasry */ 25600db9d74eSMina Almasry VM_BUG_ON(dummy_out_regions_needed != 1); 25615e911373SMike Kravetz break; 25625e911373SMike Kravetz case VMA_COMMIT_RESV: 2563075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 25640db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 25650db9d74eSMina Almasry VM_BUG_ON(ret < 0); 25665e911373SMike Kravetz break; 2567feba16e2SMike Kravetz case VMA_END_RESV: 25680db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 25695e911373SMike Kravetz ret = 0; 25705e911373SMike Kravetz break; 257196b96a96SMike Kravetz case VMA_ADD_RESV: 25720db9d74eSMina Almasry if (vma->vm_flags & VM_MAYSHARE) { 2573075a61d0SMina Almasry ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 25740db9d74eSMina Almasry /* region_add calls of range 1 should never fail. */ 25750db9d74eSMina Almasry VM_BUG_ON(ret < 0); 25760db9d74eSMina Almasry } else { 25770db9d74eSMina Almasry region_abort(resv, idx, idx + 1, 1); 257896b96a96SMike Kravetz ret = region_del(resv, idx, idx + 1); 257996b96a96SMike Kravetz } 258096b96a96SMike Kravetz break; 2581846be085SMike Kravetz case VMA_DEL_RESV: 2582846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 2583846be085SMike Kravetz region_abort(resv, idx, idx + 1, 1); 2584846be085SMike Kravetz ret = region_del(resv, idx, idx + 1); 2585846be085SMike Kravetz } else { 2586846be085SMike Kravetz ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 2587846be085SMike Kravetz /* region_add calls of range 1 should never fail. */ 2588846be085SMike Kravetz VM_BUG_ON(ret < 0); 2589846be085SMike Kravetz } 2590846be085SMike Kravetz break; 25915e911373SMike Kravetz default: 25925e911373SMike Kravetz BUG(); 25935e911373SMike Kravetz } 259484afd99bSAndy Whitcroft 2595846be085SMike Kravetz if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV) 2596cf3ad20bSMike Kravetz return ret; 259767961f9dSMike Kravetz /* 2598bf3d12b9SMiaohe Lin * We know private mapping must have HPAGE_RESV_OWNER set. 2599bf3d12b9SMiaohe Lin * 260067961f9dSMike Kravetz * In most cases, reserves always exist for private mappings. 260167961f9dSMike Kravetz * However, a file associated with mapping could have been 260267961f9dSMike Kravetz * hole punched or truncated after reserves were consumed. 260367961f9dSMike Kravetz * As subsequent fault on such a range will not use reserves. 260467961f9dSMike Kravetz * Subtle - The reserve map for private mappings has the 260567961f9dSMike Kravetz * opposite meaning than that of shared mappings. If NO 260667961f9dSMike Kravetz * entry is in the reserve map, it means a reservation exists. 260767961f9dSMike Kravetz * If an entry exists in the reserve map, it means the 260867961f9dSMike Kravetz * reservation has already been consumed. As a result, the 260967961f9dSMike Kravetz * return value of this routine is the opposite of the 261067961f9dSMike Kravetz * value returned from reserve map manipulation routines above. 261167961f9dSMike Kravetz */ 2612bf3d12b9SMiaohe Lin if (ret > 0) 261367961f9dSMike Kravetz return 0; 2614bf3d12b9SMiaohe Lin if (ret == 0) 261567961f9dSMike Kravetz return 1; 2616bf3d12b9SMiaohe Lin return ret; 261784afd99bSAndy Whitcroft } 2618cf3ad20bSMike Kravetz 2619cf3ad20bSMike Kravetz static long vma_needs_reservation(struct hstate *h, 2620a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 2621c37f9fb1SAndy Whitcroft { 26225e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 2623cf3ad20bSMike Kravetz } 2624c37f9fb1SAndy Whitcroft 2625cf3ad20bSMike Kravetz static long vma_commit_reservation(struct hstate *h, 2626cf3ad20bSMike Kravetz struct vm_area_struct *vma, unsigned long addr) 2627cf3ad20bSMike Kravetz { 26285e911373SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 26295e911373SMike Kravetz } 26305e911373SMike Kravetz 2631feba16e2SMike Kravetz static void vma_end_reservation(struct hstate *h, 26325e911373SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 26335e911373SMike Kravetz { 2634feba16e2SMike Kravetz (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 2635c37f9fb1SAndy Whitcroft } 2636c37f9fb1SAndy Whitcroft 263796b96a96SMike Kravetz static long vma_add_reservation(struct hstate *h, 263896b96a96SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 263996b96a96SMike Kravetz { 264096b96a96SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 264196b96a96SMike Kravetz } 264296b96a96SMike Kravetz 2643846be085SMike Kravetz static long vma_del_reservation(struct hstate *h, 2644846be085SMike Kravetz struct vm_area_struct *vma, unsigned long addr) 264596b96a96SMike Kravetz { 2646846be085SMike Kravetz return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV); 2647846be085SMike Kravetz } 2648846be085SMike Kravetz 2649846be085SMike Kravetz /* 2650846be085SMike Kravetz * This routine is called to restore reservation information on error paths. 2651846be085SMike Kravetz * It should ONLY be called for pages allocated via alloc_huge_page(), and 2652846be085SMike Kravetz * the hugetlb mutex should remain held when calling this routine. 2653846be085SMike Kravetz * 2654846be085SMike Kravetz * It handles two specific cases: 2655846be085SMike Kravetz * 1) A reservation was in place and the page consumed the reservation. 2656846be085SMike Kravetz * HPageRestoreReserve is set in the page. 2657846be085SMike Kravetz * 2) No reservation was in place for the page, so HPageRestoreReserve is 2658846be085SMike Kravetz * not set. However, alloc_huge_page always updates the reserve map. 2659846be085SMike Kravetz * 2660846be085SMike Kravetz * In case 1, free_huge_page later in the error path will increment the 2661846be085SMike Kravetz * global reserve count. But, free_huge_page does not have enough context 2662846be085SMike Kravetz * to adjust the reservation map. This case deals primarily with private 2663846be085SMike Kravetz * mappings. Adjust the reserve map here to be consistent with global 2664846be085SMike Kravetz * reserve count adjustments to be made by free_huge_page. Make sure the 2665846be085SMike Kravetz * reserve map indicates there is a reservation present. 2666846be085SMike Kravetz * 2667846be085SMike Kravetz * In case 2, simply undo reserve map modifications done by alloc_huge_page. 2668846be085SMike Kravetz */ 2669846be085SMike Kravetz void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma, 2670846be085SMike Kravetz unsigned long address, struct page *page) 2671846be085SMike Kravetz { 267296b96a96SMike Kravetz long rc = vma_needs_reservation(h, vma, address); 267396b96a96SMike Kravetz 2674846be085SMike Kravetz if (HPageRestoreReserve(page)) { 2675846be085SMike Kravetz if (unlikely(rc < 0)) 267696b96a96SMike Kravetz /* 267796b96a96SMike Kravetz * Rare out of memory condition in reserve map 2678d6995da3SMike Kravetz * manipulation. Clear HPageRestoreReserve so that 267996b96a96SMike Kravetz * global reserve count will not be incremented 268096b96a96SMike Kravetz * by free_huge_page. This will make it appear 268196b96a96SMike Kravetz * as though the reservation for this page was 268296b96a96SMike Kravetz * consumed. This may prevent the task from 268396b96a96SMike Kravetz * faulting in the page at a later time. This 268496b96a96SMike Kravetz * is better than inconsistent global huge page 268596b96a96SMike Kravetz * accounting of reserve counts. 268696b96a96SMike Kravetz */ 2687d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 2688846be085SMike Kravetz else if (rc) 2689846be085SMike Kravetz (void)vma_add_reservation(h, vma, address); 2690846be085SMike Kravetz else 2691846be085SMike Kravetz vma_end_reservation(h, vma, address); 2692846be085SMike Kravetz } else { 2693846be085SMike Kravetz if (!rc) { 269496b96a96SMike Kravetz /* 2695846be085SMike Kravetz * This indicates there is an entry in the reserve map 2696c7b1850dSMike Kravetz * not added by alloc_huge_page. We know it was added 2697846be085SMike Kravetz * before the alloc_huge_page call, otherwise 2698846be085SMike Kravetz * HPageRestoreReserve would be set on the page. 2699846be085SMike Kravetz * Remove the entry so that a subsequent allocation 2700846be085SMike Kravetz * does not consume a reservation. 270196b96a96SMike Kravetz */ 2702846be085SMike Kravetz rc = vma_del_reservation(h, vma, address); 2703846be085SMike Kravetz if (rc < 0) 2704846be085SMike Kravetz /* 2705846be085SMike Kravetz * VERY rare out of memory condition. Since 2706846be085SMike Kravetz * we can not delete the entry, set 2707846be085SMike Kravetz * HPageRestoreReserve so that the reserve 2708846be085SMike Kravetz * count will be incremented when the page 2709846be085SMike Kravetz * is freed. This reserve will be consumed 2710846be085SMike Kravetz * on a subsequent allocation. 2711846be085SMike Kravetz */ 2712846be085SMike Kravetz SetHPageRestoreReserve(page); 2713846be085SMike Kravetz } else if (rc < 0) { 2714846be085SMike Kravetz /* 2715846be085SMike Kravetz * Rare out of memory condition from 2716846be085SMike Kravetz * vma_needs_reservation call. Memory allocation is 2717846be085SMike Kravetz * only attempted if a new entry is needed. Therefore, 2718846be085SMike Kravetz * this implies there is not an entry in the 2719846be085SMike Kravetz * reserve map. 2720846be085SMike Kravetz * 2721846be085SMike Kravetz * For shared mappings, no entry in the map indicates 2722846be085SMike Kravetz * no reservation. We are done. 2723846be085SMike Kravetz */ 2724846be085SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 2725846be085SMike Kravetz /* 2726846be085SMike Kravetz * For private mappings, no entry indicates 2727846be085SMike Kravetz * a reservation is present. Since we can 2728846be085SMike Kravetz * not add an entry, set SetHPageRestoreReserve 2729846be085SMike Kravetz * on the page so reserve count will be 2730846be085SMike Kravetz * incremented when freed. This reserve will 2731846be085SMike Kravetz * be consumed on a subsequent allocation. 2732846be085SMike Kravetz */ 2733846be085SMike Kravetz SetHPageRestoreReserve(page); 273496b96a96SMike Kravetz } else 2735846be085SMike Kravetz /* 2736846be085SMike Kravetz * No reservation present, do nothing 2737846be085SMike Kravetz */ 273896b96a96SMike Kravetz vma_end_reservation(h, vma, address); 273996b96a96SMike Kravetz } 274096b96a96SMike Kravetz } 274196b96a96SMike Kravetz 2742369fa227SOscar Salvador /* 2743369fa227SOscar Salvador * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one 2744369fa227SOscar Salvador * @h: struct hstate old page belongs to 2745369fa227SOscar Salvador * @old_page: Old page to dissolve 2746ae37c7ffSOscar Salvador * @list: List to isolate the page in case we need to 2747369fa227SOscar Salvador * Returns 0 on success, otherwise negated error. 2748369fa227SOscar Salvador */ 2749ae37c7ffSOscar Salvador static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page, 2750ae37c7ffSOscar Salvador struct list_head *list) 2751369fa227SOscar Salvador { 2752369fa227SOscar Salvador gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2753de656ed3SSidhartha Kumar struct folio *old_folio = page_folio(old_page); 2754de656ed3SSidhartha Kumar int nid = folio_nid(old_folio); 2755369fa227SOscar Salvador struct page *new_page; 2756de656ed3SSidhartha Kumar struct folio *new_folio; 2757369fa227SOscar Salvador int ret = 0; 2758369fa227SOscar Salvador 2759369fa227SOscar Salvador /* 2760369fa227SOscar Salvador * Before dissolving the page, we need to allocate a new one for the 2761f41f2ed4SMuchun Song * pool to remain stable. Here, we allocate the page and 'prep' it 2762f41f2ed4SMuchun Song * by doing everything but actually updating counters and adding to 2763f41f2ed4SMuchun Song * the pool. This simplifies and let us do most of the processing 2764f41f2ed4SMuchun Song * under the lock. 2765369fa227SOscar Salvador */ 2766369fa227SOscar Salvador new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL); 2767369fa227SOscar Salvador if (!new_page) 2768369fa227SOscar Salvador return -ENOMEM; 2769de656ed3SSidhartha Kumar new_folio = page_folio(new_page); 2770de656ed3SSidhartha Kumar __prep_new_hugetlb_folio(h, new_folio); 2771369fa227SOscar Salvador 2772369fa227SOscar Salvador retry: 2773369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2774de656ed3SSidhartha Kumar if (!folio_test_hugetlb(old_folio)) { 2775369fa227SOscar Salvador /* 2776369fa227SOscar Salvador * Freed from under us. Drop new_page too. 2777369fa227SOscar Salvador */ 2778369fa227SOscar Salvador goto free_new; 2779de656ed3SSidhartha Kumar } else if (folio_ref_count(old_folio)) { 2780369fa227SOscar Salvador /* 2781ae37c7ffSOscar Salvador * Someone has grabbed the page, try to isolate it here. 2782ae37c7ffSOscar Salvador * Fail with -EBUSY if not possible. 2783369fa227SOscar Salvador */ 2784ae37c7ffSOscar Salvador spin_unlock_irq(&hugetlb_lock); 27857ce82f4cSMiaohe Lin ret = isolate_hugetlb(old_page, list); 2786ae37c7ffSOscar Salvador spin_lock_irq(&hugetlb_lock); 2787369fa227SOscar Salvador goto free_new; 2788de656ed3SSidhartha Kumar } else if (!folio_test_hugetlb_freed(old_folio)) { 2789369fa227SOscar Salvador /* 2790369fa227SOscar Salvador * Page's refcount is 0 but it has not been enqueued in the 2791369fa227SOscar Salvador * freelist yet. Race window is small, so we can succeed here if 2792369fa227SOscar Salvador * we retry. 2793369fa227SOscar Salvador */ 2794369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2795369fa227SOscar Salvador cond_resched(); 2796369fa227SOscar Salvador goto retry; 2797369fa227SOscar Salvador } else { 2798369fa227SOscar Salvador /* 2799369fa227SOscar Salvador * Ok, old_page is still a genuine free hugepage. Remove it from 2800369fa227SOscar Salvador * the freelist and decrease the counters. These will be 2801369fa227SOscar Salvador * incremented again when calling __prep_account_new_huge_page() 2802369fa227SOscar Salvador * and enqueue_huge_page() for new_page. The counters will remain 2803369fa227SOscar Salvador * stable since this happens under the lock. 2804369fa227SOscar Salvador */ 2805369fa227SOscar Salvador remove_hugetlb_page(h, old_page, false); 2806369fa227SOscar Salvador 2807369fa227SOscar Salvador /* 2808b65a4edaSMike Kravetz * Ref count on new page is already zero as it was dropped 2809b65a4edaSMike Kravetz * earlier. It can be directly added to the pool free list. 2810369fa227SOscar Salvador */ 2811369fa227SOscar Salvador __prep_account_new_huge_page(h, nid); 2812369fa227SOscar Salvador enqueue_huge_page(h, new_page); 2813369fa227SOscar Salvador 2814369fa227SOscar Salvador /* 2815369fa227SOscar Salvador * Pages have been replaced, we can safely free the old one. 2816369fa227SOscar Salvador */ 2817369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2818b65d4adbSMuchun Song update_and_free_page(h, old_page, false); 2819369fa227SOscar Salvador } 2820369fa227SOscar Salvador 2821369fa227SOscar Salvador return ret; 2822369fa227SOscar Salvador 2823369fa227SOscar Salvador free_new: 2824369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2825b65a4edaSMike Kravetz /* Page has a zero ref count, but needs a ref to be freed */ 2826de656ed3SSidhartha Kumar folio_ref_unfreeze(new_folio, 1); 2827b65d4adbSMuchun Song update_and_free_page(h, new_page, false); 2828369fa227SOscar Salvador 2829369fa227SOscar Salvador return ret; 2830369fa227SOscar Salvador } 2831369fa227SOscar Salvador 2832ae37c7ffSOscar Salvador int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list) 2833369fa227SOscar Salvador { 2834369fa227SOscar Salvador struct hstate *h; 2835*d5e33bd8SSidhartha Kumar struct folio *folio = page_folio(page); 2836ae37c7ffSOscar Salvador int ret = -EBUSY; 2837369fa227SOscar Salvador 2838369fa227SOscar Salvador /* 2839369fa227SOscar Salvador * The page might have been dissolved from under our feet, so make sure 2840369fa227SOscar Salvador * to carefully check the state under the lock. 2841369fa227SOscar Salvador * Return success when racing as if we dissolved the page ourselves. 2842369fa227SOscar Salvador */ 2843369fa227SOscar Salvador spin_lock_irq(&hugetlb_lock); 2844*d5e33bd8SSidhartha Kumar if (folio_test_hugetlb(folio)) { 2845*d5e33bd8SSidhartha Kumar h = folio_hstate(folio); 2846369fa227SOscar Salvador } else { 2847369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2848369fa227SOscar Salvador return 0; 2849369fa227SOscar Salvador } 2850369fa227SOscar Salvador spin_unlock_irq(&hugetlb_lock); 2851369fa227SOscar Salvador 2852369fa227SOscar Salvador /* 2853369fa227SOscar Salvador * Fence off gigantic pages as there is a cyclic dependency between 2854369fa227SOscar Salvador * alloc_contig_range and them. Return -ENOMEM as this has the effect 2855369fa227SOscar Salvador * of bailing out right away without further retrying. 2856369fa227SOscar Salvador */ 2857369fa227SOscar Salvador if (hstate_is_gigantic(h)) 2858369fa227SOscar Salvador return -ENOMEM; 2859369fa227SOscar Salvador 2860*d5e33bd8SSidhartha Kumar if (folio_ref_count(folio) && !isolate_hugetlb(&folio->page, list)) 2861ae37c7ffSOscar Salvador ret = 0; 2862*d5e33bd8SSidhartha Kumar else if (!folio_ref_count(folio)) 2863*d5e33bd8SSidhartha Kumar ret = alloc_and_dissolve_huge_page(h, &folio->page, list); 2864ae37c7ffSOscar Salvador 2865ae37c7ffSOscar Salvador return ret; 2866369fa227SOscar Salvador } 2867369fa227SOscar Salvador 286870c3547eSMike Kravetz struct page *alloc_huge_page(struct vm_area_struct *vma, 286904f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 2870348ea204SAdam Litke { 287190481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 2872a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2873348ea204SAdam Litke struct page *page; 2874d85f69b0SMike Kravetz long map_chg, map_commit; 2875d85f69b0SMike Kravetz long gbl_chg; 28766d76dcf4SAneesh Kumar K.V int ret, idx; 28776d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 287808cf9fafSMina Almasry bool deferred_reserve; 28792fc39cecSAdam Litke 28806d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 2881a1e78772SMel Gorman /* 2882d85f69b0SMike Kravetz * Examine the region/reserve map to determine if the process 2883d85f69b0SMike Kravetz * has a reservation for the page to be allocated. A return 2884d85f69b0SMike Kravetz * code of zero indicates a reservation exists (no change). 2885a1e78772SMel Gorman */ 2886d85f69b0SMike Kravetz map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); 2887d85f69b0SMike Kravetz if (map_chg < 0) 288876dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 2889d85f69b0SMike Kravetz 2890d85f69b0SMike Kravetz /* 2891d85f69b0SMike Kravetz * Processes that did not create the mapping will have no 2892d85f69b0SMike Kravetz * reserves as indicated by the region/reserve map. Check 2893d85f69b0SMike Kravetz * that the allocation will not exceed the subpool limit. 2894d85f69b0SMike Kravetz * Allocations for MAP_NORESERVE mappings also need to be 2895d85f69b0SMike Kravetz * checked against any subpool limit. 2896d85f69b0SMike Kravetz */ 2897d85f69b0SMike Kravetz if (map_chg || avoid_reserve) { 2898d85f69b0SMike Kravetz gbl_chg = hugepage_subpool_get_pages(spool, 1); 2899d85f69b0SMike Kravetz if (gbl_chg < 0) { 2900feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 290176dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 29025e911373SMike Kravetz } 290390d8b7e6SAdam Litke 2904d85f69b0SMike Kravetz /* 2905d85f69b0SMike Kravetz * Even though there was no reservation in the region/reserve 2906d85f69b0SMike Kravetz * map, there could be reservations associated with the 2907d85f69b0SMike Kravetz * subpool that can be used. This would be indicated if the 2908d85f69b0SMike Kravetz * return value of hugepage_subpool_get_pages() is zero. 2909d85f69b0SMike Kravetz * However, if avoid_reserve is specified we still avoid even 2910d85f69b0SMike Kravetz * the subpool reservations. 2911d85f69b0SMike Kravetz */ 2912d85f69b0SMike Kravetz if (avoid_reserve) 2913d85f69b0SMike Kravetz gbl_chg = 1; 2914d85f69b0SMike Kravetz } 2915d85f69b0SMike Kravetz 291608cf9fafSMina Almasry /* If this allocation is not consuming a reservation, charge it now. 291708cf9fafSMina Almasry */ 29186501fe5fSMiaohe Lin deferred_reserve = map_chg || avoid_reserve; 291908cf9fafSMina Almasry if (deferred_reserve) { 292008cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup_rsvd( 292108cf9fafSMina Almasry idx, pages_per_huge_page(h), &h_cg); 29228f34af6fSJianyu Zhan if (ret) 29238f34af6fSJianyu Zhan goto out_subpool_put; 292408cf9fafSMina Almasry } 292508cf9fafSMina Almasry 292608cf9fafSMina Almasry ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 292708cf9fafSMina Almasry if (ret) 292808cf9fafSMina Almasry goto out_uncharge_cgroup_reservation; 29298f34af6fSJianyu Zhan 2930db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 2931d85f69b0SMike Kravetz /* 2932d85f69b0SMike Kravetz * glb_chg is passed to indicate whether or not a page must be taken 2933d85f69b0SMike Kravetz * from the global free pool (global change). gbl_chg == 0 indicates 2934d85f69b0SMike Kravetz * a reservation exists for the allocation. 2935d85f69b0SMike Kravetz */ 2936d85f69b0SMike Kravetz page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); 293781a6fcaeSJoonsoo Kim if (!page) { 2938db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 29390c397daeSMichal Hocko page = alloc_buddy_huge_page_with_mpol(h, vma, addr); 29408f34af6fSJianyu Zhan if (!page) 29418f34af6fSJianyu Zhan goto out_uncharge_cgroup; 294212df140fSRik van Riel spin_lock_irq(&hugetlb_lock); 2943a88c7695SNaoya Horiguchi if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { 2944d6995da3SMike Kravetz SetHPageRestoreReserve(page); 2945a88c7695SNaoya Horiguchi h->resv_huge_pages--; 2946a88c7695SNaoya Horiguchi } 294715a8d68eSWei Yang list_add(&page->lru, &h->hugepage_activelist); 29482b21624fSMike Kravetz set_page_refcounted(page); 294981a6fcaeSJoonsoo Kim /* Fall through */ 2950a1e78772SMel Gorman } 295181a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 295208cf9fafSMina Almasry /* If allocation is not consuming a reservation, also store the 295308cf9fafSMina Almasry * hugetlb_cgroup pointer on the page. 295408cf9fafSMina Almasry */ 295508cf9fafSMina Almasry if (deferred_reserve) { 295608cf9fafSMina Almasry hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), 295708cf9fafSMina Almasry h_cg, page); 295808cf9fafSMina Almasry } 295908cf9fafSMina Almasry 2960db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 2961a1e78772SMel Gorman 2962d6995da3SMike Kravetz hugetlb_set_page_subpool(page, spool); 2963a1e78772SMel Gorman 2964d85f69b0SMike Kravetz map_commit = vma_commit_reservation(h, vma, addr); 2965d85f69b0SMike Kravetz if (unlikely(map_chg > map_commit)) { 296633039678SMike Kravetz /* 296733039678SMike Kravetz * The page was added to the reservation map between 296833039678SMike Kravetz * vma_needs_reservation and vma_commit_reservation. 296933039678SMike Kravetz * This indicates a race with hugetlb_reserve_pages. 297033039678SMike Kravetz * Adjust for the subpool count incremented above AND 297133039678SMike Kravetz * in hugetlb_reserve_pages for the same page. Also, 297233039678SMike Kravetz * the reservation count added in hugetlb_reserve_pages 297333039678SMike Kravetz * no longer applies. 297433039678SMike Kravetz */ 297533039678SMike Kravetz long rsv_adjust; 297633039678SMike Kravetz 297733039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 1); 297833039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 297979aa925bSMike Kravetz if (deferred_reserve) 298079aa925bSMike Kravetz hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), 298179aa925bSMike Kravetz pages_per_huge_page(h), page); 298233039678SMike Kravetz } 29837893d1d5SAdam Litke return page; 29848f34af6fSJianyu Zhan 29858f34af6fSJianyu Zhan out_uncharge_cgroup: 29868f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 298708cf9fafSMina Almasry out_uncharge_cgroup_reservation: 298808cf9fafSMina Almasry if (deferred_reserve) 298908cf9fafSMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), 299008cf9fafSMina Almasry h_cg); 29918f34af6fSJianyu Zhan out_subpool_put: 2992d85f69b0SMike Kravetz if (map_chg || avoid_reserve) 29938f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 2994feba16e2SMike Kravetz vma_end_reservation(h, vma, addr); 29958f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 2996b45b5bd6SDavid Gibson } 2997b45b5bd6SDavid Gibson 2998b5389086SZhenguo Yao int alloc_bootmem_huge_page(struct hstate *h, int nid) 2999e24a1307SAneesh Kumar K.V __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 3000b5389086SZhenguo Yao int __alloc_bootmem_huge_page(struct hstate *h, int nid) 3001aa888a74SAndi Kleen { 3002b5389086SZhenguo Yao struct huge_bootmem_page *m = NULL; /* initialize for clang */ 3003b2261026SJoonsoo Kim int nr_nodes, node; 3004aa888a74SAndi Kleen 3005b5389086SZhenguo Yao /* do node specific alloc */ 3006b5389086SZhenguo Yao if (nid != NUMA_NO_NODE) { 3007b5389086SZhenguo Yao m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h), 3008b5389086SZhenguo Yao 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid); 3009b5389086SZhenguo Yao if (!m) 3010b5389086SZhenguo Yao return 0; 3011b5389086SZhenguo Yao goto found; 3012b5389086SZhenguo Yao } 3013b5389086SZhenguo Yao /* allocate from next node when distributing huge pages */ 3014b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 3015b5389086SZhenguo Yao m = memblock_alloc_try_nid_raw( 30168b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 301797ad1087SMike Rapoport 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 3018aa888a74SAndi Kleen /* 3019aa888a74SAndi Kleen * Use the beginning of the huge page to store the 3020aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 3021aa888a74SAndi Kleen * puts them into the mem_map). 3022aa888a74SAndi Kleen */ 3023b5389086SZhenguo Yao if (!m) 3024b5389086SZhenguo Yao return 0; 3025aa888a74SAndi Kleen goto found; 3026aa888a74SAndi Kleen } 3027aa888a74SAndi Kleen 3028aa888a74SAndi Kleen found: 3029aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 3030330d6e48SCannon Matthews INIT_LIST_HEAD(&m->list); 3031aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 3032aa888a74SAndi Kleen m->hstate = h; 3033aa888a74SAndi Kleen return 1; 3034aa888a74SAndi Kleen } 3035aa888a74SAndi Kleen 303648b8d744SMike Kravetz /* 303748b8d744SMike Kravetz * Put bootmem huge pages into the standard lists after mem_map is up. 303848b8d744SMike Kravetz * Note: This only applies to gigantic (order > MAX_ORDER) pages. 303948b8d744SMike Kravetz */ 3040aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 3041aa888a74SAndi Kleen { 3042aa888a74SAndi Kleen struct huge_bootmem_page *m; 3043aa888a74SAndi Kleen 3044aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 304540d18ebfSMike Kravetz struct page *page = virt_to_page(m); 3046aa888a74SAndi Kleen struct hstate *h = m->hstate; 3047ee8f248dSBecky Bruce 304848b8d744SMike Kravetz VM_BUG_ON(!hstate_is_gigantic(h)); 3049aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 30507118fc29SMike Kravetz if (prep_compound_gigantic_page(page, huge_page_order(h))) { 3051ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 3052aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 30532b21624fSMike Kravetz free_huge_page(page); /* add to the hugepage allocator */ 30547118fc29SMike Kravetz } else { 3055416d85edSMike Kravetz /* VERY unlikely inflated ref count on a tail page */ 30567118fc29SMike Kravetz free_gigantic_page(page, huge_page_order(h)); 30577118fc29SMike Kravetz } 3058af0fb9dfSMichal Hocko 3059b0320c7bSRafael Aquini /* 306048b8d744SMike Kravetz * We need to restore the 'stolen' pages to totalram_pages 306148b8d744SMike Kravetz * in order to fix confusing memory reports from free(1) and 306248b8d744SMike Kravetz * other side-effects, like CommitLimit going negative. 3063b0320c7bSRafael Aquini */ 3064c78a7f36SMiaohe Lin adjust_managed_page_count(page, pages_per_huge_page(h)); 3065520495feSCannon Matthews cond_resched(); 3066aa888a74SAndi Kleen } 3067aa888a74SAndi Kleen } 3068b5389086SZhenguo Yao static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid) 3069b5389086SZhenguo Yao { 3070b5389086SZhenguo Yao unsigned long i; 3071b5389086SZhenguo Yao char buf[32]; 3072b5389086SZhenguo Yao 3073b5389086SZhenguo Yao for (i = 0; i < h->max_huge_pages_node[nid]; ++i) { 3074b5389086SZhenguo Yao if (hstate_is_gigantic(h)) { 3075b5389086SZhenguo Yao if (!alloc_bootmem_huge_page(h, nid)) 3076b5389086SZhenguo Yao break; 3077b5389086SZhenguo Yao } else { 3078b5389086SZhenguo Yao struct page *page; 3079b5389086SZhenguo Yao gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 3080b5389086SZhenguo Yao 3081b5389086SZhenguo Yao page = alloc_fresh_huge_page(h, gfp_mask, nid, 3082b5389086SZhenguo Yao &node_states[N_MEMORY], NULL); 3083b5389086SZhenguo Yao if (!page) 3084b5389086SZhenguo Yao break; 30852b21624fSMike Kravetz free_huge_page(page); /* free it into the hugepage allocator */ 3086b5389086SZhenguo Yao } 3087b5389086SZhenguo Yao cond_resched(); 3088b5389086SZhenguo Yao } 3089b5389086SZhenguo Yao if (i == h->max_huge_pages_node[nid]) 3090b5389086SZhenguo Yao return; 3091b5389086SZhenguo Yao 3092b5389086SZhenguo Yao string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3093b5389086SZhenguo Yao pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n", 3094b5389086SZhenguo Yao h->max_huge_pages_node[nid], buf, nid, i); 3095b5389086SZhenguo Yao h->max_huge_pages -= (h->max_huge_pages_node[nid] - i); 3096b5389086SZhenguo Yao h->max_huge_pages_node[nid] = i; 3097b5389086SZhenguo Yao } 3098aa888a74SAndi Kleen 30998faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 31001da177e4SLinus Torvalds { 31011da177e4SLinus Torvalds unsigned long i; 3102f60858f9SMike Kravetz nodemask_t *node_alloc_noretry; 3103b5389086SZhenguo Yao bool node_specific_alloc = false; 3104f60858f9SMike Kravetz 3105b5389086SZhenguo Yao /* skip gigantic hugepages allocation if hugetlb_cma enabled */ 3106b5389086SZhenguo Yao if (hstate_is_gigantic(h) && hugetlb_cma_size) { 3107b5389086SZhenguo Yao pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); 3108b5389086SZhenguo Yao return; 3109b5389086SZhenguo Yao } 3110b5389086SZhenguo Yao 3111b5389086SZhenguo Yao /* do node specific alloc */ 31120a7a0f6fSPeng Liu for_each_online_node(i) { 3113b5389086SZhenguo Yao if (h->max_huge_pages_node[i] > 0) { 3114b5389086SZhenguo Yao hugetlb_hstate_alloc_pages_onenode(h, i); 3115b5389086SZhenguo Yao node_specific_alloc = true; 3116b5389086SZhenguo Yao } 3117b5389086SZhenguo Yao } 3118b5389086SZhenguo Yao 3119b5389086SZhenguo Yao if (node_specific_alloc) 3120b5389086SZhenguo Yao return; 3121b5389086SZhenguo Yao 3122b5389086SZhenguo Yao /* below will do all node balanced alloc */ 3123f60858f9SMike Kravetz if (!hstate_is_gigantic(h)) { 3124f60858f9SMike Kravetz /* 3125f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3126f60858f9SMike Kravetz * Ignore errors as lower level routines can deal with 3127f60858f9SMike Kravetz * node_alloc_noretry == NULL. If this kmalloc fails at boot 3128f60858f9SMike Kravetz * time, we are likely in bigger trouble. 3129f60858f9SMike Kravetz */ 3130f60858f9SMike Kravetz node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), 3131f60858f9SMike Kravetz GFP_KERNEL); 3132f60858f9SMike Kravetz } else { 3133f60858f9SMike Kravetz /* allocations done at boot time */ 3134f60858f9SMike Kravetz node_alloc_noretry = NULL; 3135f60858f9SMike Kravetz } 3136f60858f9SMike Kravetz 3137f60858f9SMike Kravetz /* bit mask controlling how hard we retry per-node allocations */ 3138f60858f9SMike Kravetz if (node_alloc_noretry) 3139f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 31401da177e4SLinus Torvalds 3141e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 3142bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 3143b5389086SZhenguo Yao if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE)) 3144aa888a74SAndi Kleen break; 31450c397daeSMichal Hocko } else if (!alloc_pool_huge_page(h, 3146f60858f9SMike Kravetz &node_states[N_MEMORY], 3147f60858f9SMike Kravetz node_alloc_noretry)) 31481da177e4SLinus Torvalds break; 314969ed779aSDavid Rientjes cond_resched(); 31501da177e4SLinus Torvalds } 3151d715cf80SLiam R. Howlett if (i < h->max_huge_pages) { 3152d715cf80SLiam R. Howlett char buf[32]; 3153d715cf80SLiam R. Howlett 3154c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3155d715cf80SLiam R. Howlett pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 3156d715cf80SLiam R. Howlett h->max_huge_pages, buf, i); 31578faa8b07SAndi Kleen h->max_huge_pages = i; 3158e5ff2159SAndi Kleen } 3159f60858f9SMike Kravetz kfree(node_alloc_noretry); 3160d715cf80SLiam R. Howlett } 3161e5ff2159SAndi Kleen 3162e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 3163e5ff2159SAndi Kleen { 316479dfc695SMike Kravetz struct hstate *h, *h2; 3165e5ff2159SAndi Kleen 3166e5ff2159SAndi Kleen for_each_hstate(h) { 31678faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 3168bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 31698faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 317079dfc695SMike Kravetz 317179dfc695SMike Kravetz /* 317279dfc695SMike Kravetz * Set demote order for each hstate. Note that 317379dfc695SMike Kravetz * h->demote_order is initially 0. 317479dfc695SMike Kravetz * - We can not demote gigantic pages if runtime freeing 317579dfc695SMike Kravetz * is not supported, so skip this. 3176a01f4390SMike Kravetz * - If CMA allocation is possible, we can not demote 3177a01f4390SMike Kravetz * HUGETLB_PAGE_ORDER or smaller size pages. 317879dfc695SMike Kravetz */ 317979dfc695SMike Kravetz if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 318079dfc695SMike Kravetz continue; 3181a01f4390SMike Kravetz if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER) 3182a01f4390SMike Kravetz continue; 318379dfc695SMike Kravetz for_each_hstate(h2) { 318479dfc695SMike Kravetz if (h2 == h) 318579dfc695SMike Kravetz continue; 318679dfc695SMike Kravetz if (h2->order < h->order && 318779dfc695SMike Kravetz h2->order > h->demote_order) 318879dfc695SMike Kravetz h->demote_order = h2->order; 318979dfc695SMike Kravetz } 3190e5ff2159SAndi Kleen } 3191e5ff2159SAndi Kleen } 3192e5ff2159SAndi Kleen 3193e5ff2159SAndi Kleen static void __init report_hugepages(void) 3194e5ff2159SAndi Kleen { 3195e5ff2159SAndi Kleen struct hstate *h; 3196e5ff2159SAndi Kleen 3197e5ff2159SAndi Kleen for_each_hstate(h) { 31984abd32dbSAndi Kleen char buf[32]; 3199c6247f72SMatthew Wilcox 3200c6247f72SMatthew Wilcox string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 32016213834cSMuchun Song pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n", 3202c6247f72SMatthew Wilcox buf, h->free_huge_pages); 32036213834cSMuchun Song pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n", 32046213834cSMuchun Song hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf); 3205e5ff2159SAndi Kleen } 3206e5ff2159SAndi Kleen } 3207e5ff2159SAndi Kleen 32081da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 32096ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 32106ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 32111da177e4SLinus Torvalds { 32124415cc8dSChristoph Lameter int i; 32131121828aSMike Kravetz LIST_HEAD(page_list); 32144415cc8dSChristoph Lameter 32159487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 3216bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3217aa888a74SAndi Kleen return; 3218aa888a74SAndi Kleen 32191121828aSMike Kravetz /* 32201121828aSMike Kravetz * Collect pages to be freed on a list, and free after dropping lock 32211121828aSMike Kravetz */ 32226ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 322310c6ec49SMike Kravetz struct page *page, *next; 3224a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 3225a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 3226a5516438SAndi Kleen if (count >= h->nr_huge_pages) 32271121828aSMike Kravetz goto out; 32281da177e4SLinus Torvalds if (PageHighMem(page)) 32291da177e4SLinus Torvalds continue; 32306eb4e88aSMike Kravetz remove_hugetlb_page(h, page, false); 32311121828aSMike Kravetz list_add(&page->lru, &page_list); 32321121828aSMike Kravetz } 32331121828aSMike Kravetz } 32341121828aSMike Kravetz 32351121828aSMike Kravetz out: 3236db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 323710c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3238db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 32391da177e4SLinus Torvalds } 32401da177e4SLinus Torvalds #else 32416ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 32426ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 32431da177e4SLinus Torvalds { 32441da177e4SLinus Torvalds } 32451da177e4SLinus Torvalds #endif 32461da177e4SLinus Torvalds 324720a0307cSWu Fengguang /* 324820a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 324920a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 325020a0307cSWu Fengguang * Returns 1 if an adjustment was made. 325120a0307cSWu Fengguang */ 32526ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 32536ae11b27SLee Schermerhorn int delta) 325420a0307cSWu Fengguang { 3255b2261026SJoonsoo Kim int nr_nodes, node; 325620a0307cSWu Fengguang 32579487ca60SMike Kravetz lockdep_assert_held(&hugetlb_lock); 325820a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 325920a0307cSWu Fengguang 3260e8c5c824SLee Schermerhorn if (delta < 0) { 3261b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 3262b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 3263b2261026SJoonsoo Kim goto found; 3264b2261026SJoonsoo Kim } 3265b2261026SJoonsoo Kim } else { 3266b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 3267b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 3268b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 3269b2261026SJoonsoo Kim goto found; 3270e8c5c824SLee Schermerhorn } 32719a76db09SLee Schermerhorn } 3272b2261026SJoonsoo Kim return 0; 327320a0307cSWu Fengguang 3274b2261026SJoonsoo Kim found: 327520a0307cSWu Fengguang h->surplus_huge_pages += delta; 3276b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 3277b2261026SJoonsoo Kim return 1; 327820a0307cSWu Fengguang } 327920a0307cSWu Fengguang 3280a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 3281fd875dcaSMike Kravetz static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 32826ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 32831da177e4SLinus Torvalds { 32847893d1d5SAdam Litke unsigned long min_count, ret; 328510c6ec49SMike Kravetz struct page *page; 328610c6ec49SMike Kravetz LIST_HEAD(page_list); 3287f60858f9SMike Kravetz NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); 3288f60858f9SMike Kravetz 3289f60858f9SMike Kravetz /* 3290f60858f9SMike Kravetz * Bit mask controlling how hard we retry per-node allocations. 3291f60858f9SMike Kravetz * If we can not allocate the bit mask, do not attempt to allocate 3292f60858f9SMike Kravetz * the requested huge pages. 3293f60858f9SMike Kravetz */ 3294f60858f9SMike Kravetz if (node_alloc_noretry) 3295f60858f9SMike Kravetz nodes_clear(*node_alloc_noretry); 3296f60858f9SMike Kravetz else 3297f60858f9SMike Kravetz return -ENOMEM; 32981da177e4SLinus Torvalds 329929383967SMike Kravetz /* 330029383967SMike Kravetz * resize_lock mutex prevents concurrent adjustments to number of 330129383967SMike Kravetz * pages in hstate via the proc/sysfs interfaces. 330229383967SMike Kravetz */ 330329383967SMike Kravetz mutex_lock(&h->resize_lock); 3304b65d4adbSMuchun Song flush_free_hpage_work(h); 3305db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 33064eb0716eSAlexandre Ghiti 33074eb0716eSAlexandre Ghiti /* 3308fd875dcaSMike Kravetz * Check for a node specific request. 3309fd875dcaSMike Kravetz * Changing node specific huge page count may require a corresponding 3310fd875dcaSMike Kravetz * change to the global count. In any case, the passed node mask 3311fd875dcaSMike Kravetz * (nodes_allowed) will restrict alloc/free to the specified node. 3312fd875dcaSMike Kravetz */ 3313fd875dcaSMike Kravetz if (nid != NUMA_NO_NODE) { 3314fd875dcaSMike Kravetz unsigned long old_count = count; 3315fd875dcaSMike Kravetz 3316fd875dcaSMike Kravetz count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 3317fd875dcaSMike Kravetz /* 3318fd875dcaSMike Kravetz * User may have specified a large count value which caused the 3319fd875dcaSMike Kravetz * above calculation to overflow. In this case, they wanted 3320fd875dcaSMike Kravetz * to allocate as many huge pages as possible. Set count to 3321fd875dcaSMike Kravetz * largest possible value to align with their intention. 3322fd875dcaSMike Kravetz */ 3323fd875dcaSMike Kravetz if (count < old_count) 3324fd875dcaSMike Kravetz count = ULONG_MAX; 3325fd875dcaSMike Kravetz } 3326fd875dcaSMike Kravetz 3327fd875dcaSMike Kravetz /* 33284eb0716eSAlexandre Ghiti * Gigantic pages runtime allocation depend on the capability for large 33294eb0716eSAlexandre Ghiti * page range allocation. 33304eb0716eSAlexandre Ghiti * If the system does not provide this feature, return an error when 33314eb0716eSAlexandre Ghiti * the user tries to allocate gigantic pages but let the user free the 33324eb0716eSAlexandre Ghiti * boottime allocated gigantic pages. 33334eb0716eSAlexandre Ghiti */ 33344eb0716eSAlexandre Ghiti if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 33354eb0716eSAlexandre Ghiti if (count > persistent_huge_pages(h)) { 3336db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 333729383967SMike Kravetz mutex_unlock(&h->resize_lock); 3338f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 33394eb0716eSAlexandre Ghiti return -EINVAL; 33404eb0716eSAlexandre Ghiti } 33414eb0716eSAlexandre Ghiti /* Fall through to decrease pool */ 33424eb0716eSAlexandre Ghiti } 3343aa888a74SAndi Kleen 33447893d1d5SAdam Litke /* 33457893d1d5SAdam Litke * Increase the pool size 33467893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 33477893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 3348d1c3fb1fSNishanth Aravamudan * 33490c397daeSMichal Hocko * We might race with alloc_surplus_huge_page() here and be unable 3350d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 3351d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 3352d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 3353d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 33547893d1d5SAdam Litke */ 3355a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 33566ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 33577893d1d5SAdam Litke break; 33587893d1d5SAdam Litke } 33597893d1d5SAdam Litke 3360a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 33617893d1d5SAdam Litke /* 33627893d1d5SAdam Litke * If this allocation races such that we no longer need the 33637893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 33647893d1d5SAdam Litke * and reducing the surplus. 33657893d1d5SAdam Litke */ 3366db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3367649920c6SJia He 3368649920c6SJia He /* yield cpu to avoid soft lockup */ 3369649920c6SJia He cond_resched(); 3370649920c6SJia He 3371f60858f9SMike Kravetz ret = alloc_pool_huge_page(h, nodes_allowed, 3372f60858f9SMike Kravetz node_alloc_noretry); 3373db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 33747893d1d5SAdam Litke if (!ret) 33757893d1d5SAdam Litke goto out; 33767893d1d5SAdam Litke 3377536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 3378536240f2SMel Gorman if (signal_pending(current)) 3379536240f2SMel Gorman goto out; 33807893d1d5SAdam Litke } 33817893d1d5SAdam Litke 33827893d1d5SAdam Litke /* 33837893d1d5SAdam Litke * Decrease the pool size 33847893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 33857893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 33867893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 33877893d1d5SAdam Litke * to the desired size as pages become free. 3388d1c3fb1fSNishanth Aravamudan * 3389d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 3390d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 3391d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 33920c397daeSMichal Hocko * alloc_surplus_huge_page() is checking the global counter, 3393d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 3394d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 3395d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 33967893d1d5SAdam Litke */ 3397a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 33986b0c880dSAdam Litke min_count = max(count, min_count); 33996ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 340010c6ec49SMike Kravetz 340110c6ec49SMike Kravetz /* 340210c6ec49SMike Kravetz * Collect pages to be removed on list without dropping lock 340310c6ec49SMike Kravetz */ 3404a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 340510c6ec49SMike Kravetz page = remove_pool_huge_page(h, nodes_allowed, 0); 340610c6ec49SMike Kravetz if (!page) 34071da177e4SLinus Torvalds break; 340810c6ec49SMike Kravetz 340910c6ec49SMike Kravetz list_add(&page->lru, &page_list); 34101da177e4SLinus Torvalds } 341110c6ec49SMike Kravetz /* free the pages after dropping lock */ 3412db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 341310c6ec49SMike Kravetz update_and_free_pages_bulk(h, &page_list); 3414b65d4adbSMuchun Song flush_free_hpage_work(h); 3415db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 341610c6ec49SMike Kravetz 3417a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 34186ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 34197893d1d5SAdam Litke break; 34207893d1d5SAdam Litke } 34217893d1d5SAdam Litke out: 34224eb0716eSAlexandre Ghiti h->max_huge_pages = persistent_huge_pages(h); 3423db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 342429383967SMike Kravetz mutex_unlock(&h->resize_lock); 34254eb0716eSAlexandre Ghiti 3426f60858f9SMike Kravetz NODEMASK_FREE(node_alloc_noretry); 3427f60858f9SMike Kravetz 34284eb0716eSAlexandre Ghiti return 0; 34291da177e4SLinus Torvalds } 34301da177e4SLinus Torvalds 34318531fc6fSMike Kravetz static int demote_free_huge_page(struct hstate *h, struct page *page) 34328531fc6fSMike Kravetz { 34338531fc6fSMike Kravetz int i, nid = page_to_nid(page); 34348531fc6fSMike Kravetz struct hstate *target_hstate; 343531731452SDoug Berger struct page *subpage; 34368531fc6fSMike Kravetz int rc = 0; 34378531fc6fSMike Kravetz 34388531fc6fSMike Kravetz target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order); 34398531fc6fSMike Kravetz 34408531fc6fSMike Kravetz remove_hugetlb_page_for_demote(h, page, false); 34418531fc6fSMike Kravetz spin_unlock_irq(&hugetlb_lock); 34428531fc6fSMike Kravetz 34436213834cSMuchun Song rc = hugetlb_vmemmap_restore(h, page); 34448531fc6fSMike Kravetz if (rc) { 34458531fc6fSMike Kravetz /* Allocation of vmemmmap failed, we can not demote page */ 34468531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 34478531fc6fSMike Kravetz set_page_refcounted(page); 34488531fc6fSMike Kravetz add_hugetlb_page(h, page, false); 34498531fc6fSMike Kravetz return rc; 34508531fc6fSMike Kravetz } 34518531fc6fSMike Kravetz 34528531fc6fSMike Kravetz /* 34538531fc6fSMike Kravetz * Use destroy_compound_hugetlb_page_for_demote for all huge page 34548531fc6fSMike Kravetz * sizes as it will not ref count pages. 34558531fc6fSMike Kravetz */ 34568531fc6fSMike Kravetz destroy_compound_hugetlb_page_for_demote(page, huge_page_order(h)); 34578531fc6fSMike Kravetz 34588531fc6fSMike Kravetz /* 34598531fc6fSMike Kravetz * Taking target hstate mutex synchronizes with set_max_huge_pages. 34608531fc6fSMike Kravetz * Without the mutex, pages added to target hstate could be marked 34618531fc6fSMike Kravetz * as surplus. 34628531fc6fSMike Kravetz * 34638531fc6fSMike Kravetz * Note that we already hold h->resize_lock. To prevent deadlock, 34648531fc6fSMike Kravetz * use the convention of always taking larger size hstate mutex first. 34658531fc6fSMike Kravetz */ 34668531fc6fSMike Kravetz mutex_lock(&target_hstate->resize_lock); 34678531fc6fSMike Kravetz for (i = 0; i < pages_per_huge_page(h); 34688531fc6fSMike Kravetz i += pages_per_huge_page(target_hstate)) { 346931731452SDoug Berger subpage = nth_page(page, i); 34708531fc6fSMike Kravetz if (hstate_is_gigantic(target_hstate)) 347131731452SDoug Berger prep_compound_gigantic_page_for_demote(subpage, 34728531fc6fSMike Kravetz target_hstate->order); 34738531fc6fSMike Kravetz else 347431731452SDoug Berger prep_compound_page(subpage, target_hstate->order); 347531731452SDoug Berger set_page_private(subpage, 0); 347631731452SDoug Berger prep_new_huge_page(target_hstate, subpage, nid); 34772b21624fSMike Kravetz free_huge_page(subpage); 34788531fc6fSMike Kravetz } 34798531fc6fSMike Kravetz mutex_unlock(&target_hstate->resize_lock); 34808531fc6fSMike Kravetz 34818531fc6fSMike Kravetz spin_lock_irq(&hugetlb_lock); 34828531fc6fSMike Kravetz 34838531fc6fSMike Kravetz /* 34848531fc6fSMike Kravetz * Not absolutely necessary, but for consistency update max_huge_pages 34858531fc6fSMike Kravetz * based on pool changes for the demoted page. 34868531fc6fSMike Kravetz */ 34878531fc6fSMike Kravetz h->max_huge_pages--; 3488a43a83c7SMiaohe Lin target_hstate->max_huge_pages += 3489a43a83c7SMiaohe Lin pages_per_huge_page(h) / pages_per_huge_page(target_hstate); 34908531fc6fSMike Kravetz 34918531fc6fSMike Kravetz return rc; 34928531fc6fSMike Kravetz } 34938531fc6fSMike Kravetz 349479dfc695SMike Kravetz static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 349579dfc695SMike Kravetz __must_hold(&hugetlb_lock) 349679dfc695SMike Kravetz { 34978531fc6fSMike Kravetz int nr_nodes, node; 34988531fc6fSMike Kravetz struct page *page; 349979dfc695SMike Kravetz 350079dfc695SMike Kravetz lockdep_assert_held(&hugetlb_lock); 350179dfc695SMike Kravetz 350279dfc695SMike Kravetz /* We should never get here if no demote order */ 350379dfc695SMike Kravetz if (!h->demote_order) { 350479dfc695SMike Kravetz pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n"); 350579dfc695SMike Kravetz return -EINVAL; /* internal error */ 350679dfc695SMike Kravetz } 350779dfc695SMike Kravetz 35088531fc6fSMike Kravetz for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 35095a317412SMike Kravetz list_for_each_entry(page, &h->hugepage_freelists[node], lru) { 35105a317412SMike Kravetz if (PageHWPoison(page)) 35115a317412SMike Kravetz continue; 35125a317412SMike Kravetz 35135a317412SMike Kravetz return demote_free_huge_page(h, page); 35148531fc6fSMike Kravetz } 35158531fc6fSMike Kravetz } 35168531fc6fSMike Kravetz 35175a317412SMike Kravetz /* 35185a317412SMike Kravetz * Only way to get here is if all pages on free lists are poisoned. 35195a317412SMike Kravetz * Return -EBUSY so that caller will not retry. 35205a317412SMike Kravetz */ 35215a317412SMike Kravetz return -EBUSY; 352279dfc695SMike Kravetz } 352379dfc695SMike Kravetz 3524a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 3525a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 3526a3437870SNishanth Aravamudan 352779dfc695SMike Kravetz #define HSTATE_ATTR_WO(_name) \ 352879dfc695SMike Kravetz static struct kobj_attribute _name##_attr = __ATTR_WO(_name) 352979dfc695SMike Kravetz 3530a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 353198bc26acSMiaohe Lin static struct kobj_attribute _name##_attr = __ATTR_RW(_name) 3532a3437870SNishanth Aravamudan 3533a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 3534a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 3535a3437870SNishanth Aravamudan 35369a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 35379a305230SLee Schermerhorn 35389a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 3539a3437870SNishanth Aravamudan { 3540a3437870SNishanth Aravamudan int i; 35419a305230SLee Schermerhorn 3542a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 35439a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 35449a305230SLee Schermerhorn if (nidp) 35459a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 3546a3437870SNishanth Aravamudan return &hstates[i]; 35479a305230SLee Schermerhorn } 35489a305230SLee Schermerhorn 35499a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 3550a3437870SNishanth Aravamudan } 3551a3437870SNishanth Aravamudan 355206808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 3553a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3554a3437870SNishanth Aravamudan { 35559a305230SLee Schermerhorn struct hstate *h; 35569a305230SLee Schermerhorn unsigned long nr_huge_pages; 35579a305230SLee Schermerhorn int nid; 35589a305230SLee Schermerhorn 35599a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 35609a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 35619a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 35629a305230SLee Schermerhorn else 35639a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 35649a305230SLee Schermerhorn 3565ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", nr_huge_pages); 3566a3437870SNishanth Aravamudan } 3567adbe8726SEric B Munson 3568238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 3569238d3c13SDavid Rientjes struct hstate *h, int nid, 3570238d3c13SDavid Rientjes unsigned long count, size_t len) 3571a3437870SNishanth Aravamudan { 3572a3437870SNishanth Aravamudan int err; 35732d0adf7eSOscar Salvador nodemask_t nodes_allowed, *n_mask; 3574a3437870SNishanth Aravamudan 35752d0adf7eSOscar Salvador if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 35762d0adf7eSOscar Salvador return -EINVAL; 3577adbe8726SEric B Munson 35789a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 35799a305230SLee Schermerhorn /* 35809a305230SLee Schermerhorn * global hstate attribute 35819a305230SLee Schermerhorn */ 35829a305230SLee Schermerhorn if (!(obey_mempolicy && 35832d0adf7eSOscar Salvador init_nodemask_of_mempolicy(&nodes_allowed))) 35842d0adf7eSOscar Salvador n_mask = &node_states[N_MEMORY]; 35852d0adf7eSOscar Salvador else 35862d0adf7eSOscar Salvador n_mask = &nodes_allowed; 35872d0adf7eSOscar Salvador } else { 35889a305230SLee Schermerhorn /* 3589fd875dcaSMike Kravetz * Node specific request. count adjustment happens in 3590fd875dcaSMike Kravetz * set_max_huge_pages() after acquiring hugetlb_lock. 35919a305230SLee Schermerhorn */ 35922d0adf7eSOscar Salvador init_nodemask_of_node(&nodes_allowed, nid); 35932d0adf7eSOscar Salvador n_mask = &nodes_allowed; 3594fd875dcaSMike Kravetz } 35959a305230SLee Schermerhorn 35962d0adf7eSOscar Salvador err = set_max_huge_pages(h, count, nid, n_mask); 359706808b08SLee Schermerhorn 35984eb0716eSAlexandre Ghiti return err ? err : len; 359906808b08SLee Schermerhorn } 360006808b08SLee Schermerhorn 3601238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 3602238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 3603238d3c13SDavid Rientjes size_t len) 3604238d3c13SDavid Rientjes { 3605238d3c13SDavid Rientjes struct hstate *h; 3606238d3c13SDavid Rientjes unsigned long count; 3607238d3c13SDavid Rientjes int nid; 3608238d3c13SDavid Rientjes int err; 3609238d3c13SDavid Rientjes 3610238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 3611238d3c13SDavid Rientjes if (err) 3612238d3c13SDavid Rientjes return err; 3613238d3c13SDavid Rientjes 3614238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 3615238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 3616238d3c13SDavid Rientjes } 3617238d3c13SDavid Rientjes 361806808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 361906808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 362006808b08SLee Schermerhorn { 362106808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 362206808b08SLee Schermerhorn } 362306808b08SLee Schermerhorn 362406808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 362506808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 362606808b08SLee Schermerhorn { 3627238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 3628a3437870SNishanth Aravamudan } 3629a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 3630a3437870SNishanth Aravamudan 363106808b08SLee Schermerhorn #ifdef CONFIG_NUMA 363206808b08SLee Schermerhorn 363306808b08SLee Schermerhorn /* 363406808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 363506808b08SLee Schermerhorn * huge page alloc/free. 363606808b08SLee Schermerhorn */ 363706808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 3638ae7a927dSJoe Perches struct kobj_attribute *attr, 3639ae7a927dSJoe Perches char *buf) 364006808b08SLee Schermerhorn { 364106808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 364206808b08SLee Schermerhorn } 364306808b08SLee Schermerhorn 364406808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 364506808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 364606808b08SLee Schermerhorn { 3647238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 364806808b08SLee Schermerhorn } 364906808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 365006808b08SLee Schermerhorn #endif 365106808b08SLee Schermerhorn 365206808b08SLee Schermerhorn 3653a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 3654a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3655a3437870SNishanth Aravamudan { 36569a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3657ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); 3658a3437870SNishanth Aravamudan } 3659adbe8726SEric B Munson 3660a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 3661a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 3662a3437870SNishanth Aravamudan { 3663a3437870SNishanth Aravamudan int err; 3664a3437870SNishanth Aravamudan unsigned long input; 36659a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3666a3437870SNishanth Aravamudan 3667bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 3668adbe8726SEric B Munson return -EINVAL; 3669adbe8726SEric B Munson 36703dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 3671a3437870SNishanth Aravamudan if (err) 367273ae31e5SEric B Munson return err; 3673a3437870SNishanth Aravamudan 3674db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 3675a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 3676db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 3677a3437870SNishanth Aravamudan 3678a3437870SNishanth Aravamudan return count; 3679a3437870SNishanth Aravamudan } 3680a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 3681a3437870SNishanth Aravamudan 3682a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 3683a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3684a3437870SNishanth Aravamudan { 36859a305230SLee Schermerhorn struct hstate *h; 36869a305230SLee Schermerhorn unsigned long free_huge_pages; 36879a305230SLee Schermerhorn int nid; 36889a305230SLee Schermerhorn 36899a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 36909a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 36919a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 36929a305230SLee Schermerhorn else 36939a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 36949a305230SLee Schermerhorn 3695ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", free_huge_pages); 3696a3437870SNishanth Aravamudan } 3697a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 3698a3437870SNishanth Aravamudan 3699a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 3700a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3701a3437870SNishanth Aravamudan { 37029a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 3703ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); 3704a3437870SNishanth Aravamudan } 3705a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 3706a3437870SNishanth Aravamudan 3707a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 3708a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 3709a3437870SNishanth Aravamudan { 37109a305230SLee Schermerhorn struct hstate *h; 37119a305230SLee Schermerhorn unsigned long surplus_huge_pages; 37129a305230SLee Schermerhorn int nid; 37139a305230SLee Schermerhorn 37149a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 37159a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 37169a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 37179a305230SLee Schermerhorn else 37189a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 37199a305230SLee Schermerhorn 3720ae7a927dSJoe Perches return sysfs_emit(buf, "%lu\n", surplus_huge_pages); 3721a3437870SNishanth Aravamudan } 3722a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 3723a3437870SNishanth Aravamudan 372479dfc695SMike Kravetz static ssize_t demote_store(struct kobject *kobj, 372579dfc695SMike Kravetz struct kobj_attribute *attr, const char *buf, size_t len) 372679dfc695SMike Kravetz { 372779dfc695SMike Kravetz unsigned long nr_demote; 372879dfc695SMike Kravetz unsigned long nr_available; 372979dfc695SMike Kravetz nodemask_t nodes_allowed, *n_mask; 373079dfc695SMike Kravetz struct hstate *h; 37318eeda55fSLi zeming int err; 373279dfc695SMike Kravetz int nid; 373379dfc695SMike Kravetz 373479dfc695SMike Kravetz err = kstrtoul(buf, 10, &nr_demote); 373579dfc695SMike Kravetz if (err) 373679dfc695SMike Kravetz return err; 373779dfc695SMike Kravetz h = kobj_to_hstate(kobj, &nid); 373879dfc695SMike Kravetz 373979dfc695SMike Kravetz if (nid != NUMA_NO_NODE) { 374079dfc695SMike Kravetz init_nodemask_of_node(&nodes_allowed, nid); 374179dfc695SMike Kravetz n_mask = &nodes_allowed; 374279dfc695SMike Kravetz } else { 374379dfc695SMike Kravetz n_mask = &node_states[N_MEMORY]; 374479dfc695SMike Kravetz } 374579dfc695SMike Kravetz 374679dfc695SMike Kravetz /* Synchronize with other sysfs operations modifying huge pages */ 374779dfc695SMike Kravetz mutex_lock(&h->resize_lock); 374879dfc695SMike Kravetz spin_lock_irq(&hugetlb_lock); 374979dfc695SMike Kravetz 375079dfc695SMike Kravetz while (nr_demote) { 375179dfc695SMike Kravetz /* 375279dfc695SMike Kravetz * Check for available pages to demote each time thorough the 375379dfc695SMike Kravetz * loop as demote_pool_huge_page will drop hugetlb_lock. 375479dfc695SMike Kravetz */ 375579dfc695SMike Kravetz if (nid != NUMA_NO_NODE) 375679dfc695SMike Kravetz nr_available = h->free_huge_pages_node[nid]; 375779dfc695SMike Kravetz else 375879dfc695SMike Kravetz nr_available = h->free_huge_pages; 375979dfc695SMike Kravetz nr_available -= h->resv_huge_pages; 376079dfc695SMike Kravetz if (!nr_available) 376179dfc695SMike Kravetz break; 376279dfc695SMike Kravetz 376379dfc695SMike Kravetz err = demote_pool_huge_page(h, n_mask); 376479dfc695SMike Kravetz if (err) 376579dfc695SMike Kravetz break; 376679dfc695SMike Kravetz 376779dfc695SMike Kravetz nr_demote--; 376879dfc695SMike Kravetz } 376979dfc695SMike Kravetz 377079dfc695SMike Kravetz spin_unlock_irq(&hugetlb_lock); 377179dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 377279dfc695SMike Kravetz 377379dfc695SMike Kravetz if (err) 377479dfc695SMike Kravetz return err; 377579dfc695SMike Kravetz return len; 377679dfc695SMike Kravetz } 377779dfc695SMike Kravetz HSTATE_ATTR_WO(demote); 377879dfc695SMike Kravetz 377979dfc695SMike Kravetz static ssize_t demote_size_show(struct kobject *kobj, 378079dfc695SMike Kravetz struct kobj_attribute *attr, char *buf) 378179dfc695SMike Kravetz { 378212658abfSMiaohe Lin struct hstate *h = kobj_to_hstate(kobj, NULL); 378379dfc695SMike Kravetz unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K; 378479dfc695SMike Kravetz 378579dfc695SMike Kravetz return sysfs_emit(buf, "%lukB\n", demote_size); 378679dfc695SMike Kravetz } 378779dfc695SMike Kravetz 378879dfc695SMike Kravetz static ssize_t demote_size_store(struct kobject *kobj, 378979dfc695SMike Kravetz struct kobj_attribute *attr, 379079dfc695SMike Kravetz const char *buf, size_t count) 379179dfc695SMike Kravetz { 379279dfc695SMike Kravetz struct hstate *h, *demote_hstate; 379379dfc695SMike Kravetz unsigned long demote_size; 379479dfc695SMike Kravetz unsigned int demote_order; 379579dfc695SMike Kravetz 379679dfc695SMike Kravetz demote_size = (unsigned long)memparse(buf, NULL); 379779dfc695SMike Kravetz 379879dfc695SMike Kravetz demote_hstate = size_to_hstate(demote_size); 379979dfc695SMike Kravetz if (!demote_hstate) 380079dfc695SMike Kravetz return -EINVAL; 380179dfc695SMike Kravetz demote_order = demote_hstate->order; 3802a01f4390SMike Kravetz if (demote_order < HUGETLB_PAGE_ORDER) 3803a01f4390SMike Kravetz return -EINVAL; 380479dfc695SMike Kravetz 380579dfc695SMike Kravetz /* demote order must be smaller than hstate order */ 380612658abfSMiaohe Lin h = kobj_to_hstate(kobj, NULL); 380779dfc695SMike Kravetz if (demote_order >= h->order) 380879dfc695SMike Kravetz return -EINVAL; 380979dfc695SMike Kravetz 381079dfc695SMike Kravetz /* resize_lock synchronizes access to demote size and writes */ 381179dfc695SMike Kravetz mutex_lock(&h->resize_lock); 381279dfc695SMike Kravetz h->demote_order = demote_order; 381379dfc695SMike Kravetz mutex_unlock(&h->resize_lock); 381479dfc695SMike Kravetz 381579dfc695SMike Kravetz return count; 381679dfc695SMike Kravetz } 381779dfc695SMike Kravetz HSTATE_ATTR(demote_size); 381879dfc695SMike Kravetz 3819a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 3820a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 3821a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 3822a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 3823a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 3824a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 382506808b08SLee Schermerhorn #ifdef CONFIG_NUMA 382606808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 382706808b08SLee Schermerhorn #endif 3828a3437870SNishanth Aravamudan NULL, 3829a3437870SNishanth Aravamudan }; 3830a3437870SNishanth Aravamudan 383167e5ed96SArvind Yadav static const struct attribute_group hstate_attr_group = { 3832a3437870SNishanth Aravamudan .attrs = hstate_attrs, 3833a3437870SNishanth Aravamudan }; 3834a3437870SNishanth Aravamudan 383579dfc695SMike Kravetz static struct attribute *hstate_demote_attrs[] = { 383679dfc695SMike Kravetz &demote_size_attr.attr, 383779dfc695SMike Kravetz &demote_attr.attr, 383879dfc695SMike Kravetz NULL, 383979dfc695SMike Kravetz }; 384079dfc695SMike Kravetz 384179dfc695SMike Kravetz static const struct attribute_group hstate_demote_attr_group = { 384279dfc695SMike Kravetz .attrs = hstate_demote_attrs, 384379dfc695SMike Kravetz }; 384479dfc695SMike Kravetz 3845094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 38469a305230SLee Schermerhorn struct kobject **hstate_kobjs, 384767e5ed96SArvind Yadav const struct attribute_group *hstate_attr_group) 3848a3437870SNishanth Aravamudan { 3849a3437870SNishanth Aravamudan int retval; 3850972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 3851a3437870SNishanth Aravamudan 38529a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 38539a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 3854a3437870SNishanth Aravamudan return -ENOMEM; 3855a3437870SNishanth Aravamudan 38569a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 3857cc2205a6SMiaohe Lin if (retval) { 38589a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 3859cc2205a6SMiaohe Lin hstate_kobjs[hi] = NULL; 38603a6bdda0SMiaohe Lin return retval; 3861cc2205a6SMiaohe Lin } 3862a3437870SNishanth Aravamudan 386379dfc695SMike Kravetz if (h->demote_order) { 386401088a60SMiaohe Lin retval = sysfs_create_group(hstate_kobjs[hi], 386501088a60SMiaohe Lin &hstate_demote_attr_group); 386601088a60SMiaohe Lin if (retval) { 386779dfc695SMike Kravetz pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name); 386801088a60SMiaohe Lin sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group); 386901088a60SMiaohe Lin kobject_put(hstate_kobjs[hi]); 387001088a60SMiaohe Lin hstate_kobjs[hi] = NULL; 387101088a60SMiaohe Lin return retval; 387201088a60SMiaohe Lin } 387379dfc695SMike Kravetz } 387479dfc695SMike Kravetz 387501088a60SMiaohe Lin return 0; 3876a3437870SNishanth Aravamudan } 3877a3437870SNishanth Aravamudan 38789a305230SLee Schermerhorn #ifdef CONFIG_NUMA 3879a4a00b45SMuchun Song static bool hugetlb_sysfs_initialized __ro_after_init; 38809a305230SLee Schermerhorn 38819a305230SLee Schermerhorn /* 38829a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 388310fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 388410fbcf4cSKay Sievers * index of a node device or _hstate == node id. 388510fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 38869a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 38879a305230SLee Schermerhorn */ 38889a305230SLee Schermerhorn struct node_hstate { 38899a305230SLee Schermerhorn struct kobject *hugepages_kobj; 38909a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 38919a305230SLee Schermerhorn }; 3892b4e289a6SAlexander Kuleshov static struct node_hstate node_hstates[MAX_NUMNODES]; 38939a305230SLee Schermerhorn 38949a305230SLee Schermerhorn /* 389510fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 38969a305230SLee Schermerhorn */ 38979a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 38989a305230SLee Schermerhorn &nr_hugepages_attr.attr, 38999a305230SLee Schermerhorn &free_hugepages_attr.attr, 39009a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 39019a305230SLee Schermerhorn NULL, 39029a305230SLee Schermerhorn }; 39039a305230SLee Schermerhorn 390467e5ed96SArvind Yadav static const struct attribute_group per_node_hstate_attr_group = { 39059a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 39069a305230SLee Schermerhorn }; 39079a305230SLee Schermerhorn 39089a305230SLee Schermerhorn /* 390910fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 39109a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 39119a305230SLee Schermerhorn */ 39129a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 39139a305230SLee Schermerhorn { 39149a305230SLee Schermerhorn int nid; 39159a305230SLee Schermerhorn 39169a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 39179a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 39189a305230SLee Schermerhorn int i; 39199a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 39209a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 39219a305230SLee Schermerhorn if (nidp) 39229a305230SLee Schermerhorn *nidp = nid; 39239a305230SLee Schermerhorn return &hstates[i]; 39249a305230SLee Schermerhorn } 39259a305230SLee Schermerhorn } 39269a305230SLee Schermerhorn 39279a305230SLee Schermerhorn BUG(); 39289a305230SLee Schermerhorn return NULL; 39299a305230SLee Schermerhorn } 39309a305230SLee Schermerhorn 39319a305230SLee Schermerhorn /* 393210fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 39339a305230SLee Schermerhorn * No-op if no hstate attributes attached. 39349a305230SLee Schermerhorn */ 3935a4a00b45SMuchun Song void hugetlb_unregister_node(struct node *node) 39369a305230SLee Schermerhorn { 39379a305230SLee Schermerhorn struct hstate *h; 393810fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 39399a305230SLee Schermerhorn 39409a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 39419b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 39429a305230SLee Schermerhorn 3943972dc4deSAneesh Kumar K.V for_each_hstate(h) { 3944972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 394501088a60SMiaohe Lin struct kobject *hstate_kobj = nhs->hstate_kobjs[idx]; 394601088a60SMiaohe Lin 394701088a60SMiaohe Lin if (!hstate_kobj) 394801088a60SMiaohe Lin continue; 394901088a60SMiaohe Lin if (h->demote_order) 395001088a60SMiaohe Lin sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group); 395101088a60SMiaohe Lin sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group); 395201088a60SMiaohe Lin kobject_put(hstate_kobj); 3953972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 3954972dc4deSAneesh Kumar K.V } 39559a305230SLee Schermerhorn 39569a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 39579a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 39589a305230SLee Schermerhorn } 39599a305230SLee Schermerhorn 39609a305230SLee Schermerhorn 39619a305230SLee Schermerhorn /* 396210fbcf4cSKay Sievers * Register hstate attributes for a single node device. 39639a305230SLee Schermerhorn * No-op if attributes already registered. 39649a305230SLee Schermerhorn */ 3965a4a00b45SMuchun Song void hugetlb_register_node(struct node *node) 39669a305230SLee Schermerhorn { 39679a305230SLee Schermerhorn struct hstate *h; 396810fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 39699a305230SLee Schermerhorn int err; 39709a305230SLee Schermerhorn 3971a4a00b45SMuchun Song if (!hugetlb_sysfs_initialized) 3972a4a00b45SMuchun Song return; 3973a4a00b45SMuchun Song 39749a305230SLee Schermerhorn if (nhs->hugepages_kobj) 39759a305230SLee Schermerhorn return; /* already allocated */ 39769a305230SLee Schermerhorn 39779a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 397810fbcf4cSKay Sievers &node->dev.kobj); 39799a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 39809a305230SLee Schermerhorn return; 39819a305230SLee Schermerhorn 39829a305230SLee Schermerhorn for_each_hstate(h) { 39839a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 39849a305230SLee Schermerhorn nhs->hstate_kobjs, 39859a305230SLee Schermerhorn &per_node_hstate_attr_group); 39869a305230SLee Schermerhorn if (err) { 3987282f4214SMike Kravetz pr_err("HugeTLB: Unable to add hstate %s for node %d\n", 398810fbcf4cSKay Sievers h->name, node->dev.id); 39899a305230SLee Schermerhorn hugetlb_unregister_node(node); 39909a305230SLee Schermerhorn break; 39919a305230SLee Schermerhorn } 39929a305230SLee Schermerhorn } 39939a305230SLee Schermerhorn } 39949a305230SLee Schermerhorn 39959a305230SLee Schermerhorn /* 39969b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 399710fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 399810fbcf4cSKay Sievers * registered their associated device by this time. 39999a305230SLee Schermerhorn */ 40007d9ca000SLuiz Capitulino static void __init hugetlb_register_all_nodes(void) 40019a305230SLee Schermerhorn { 40029a305230SLee Schermerhorn int nid; 40039a305230SLee Schermerhorn 4004a4a00b45SMuchun Song for_each_online_node(nid) 4005b958d4d0SMuchun Song hugetlb_register_node(node_devices[nid]); 40069a305230SLee Schermerhorn } 40079a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 40089a305230SLee Schermerhorn 40099a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 40109a305230SLee Schermerhorn { 40119a305230SLee Schermerhorn BUG(); 40129a305230SLee Schermerhorn if (nidp) 40139a305230SLee Schermerhorn *nidp = -1; 40149a305230SLee Schermerhorn return NULL; 40159a305230SLee Schermerhorn } 40169a305230SLee Schermerhorn 40179a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 40189a305230SLee Schermerhorn 40199a305230SLee Schermerhorn #endif 40209a305230SLee Schermerhorn 4021263b8998SMiaohe Lin #ifdef CONFIG_CMA 4022263b8998SMiaohe Lin static void __init hugetlb_cma_check(void); 4023263b8998SMiaohe Lin #else 4024263b8998SMiaohe Lin static inline __init void hugetlb_cma_check(void) 4025263b8998SMiaohe Lin { 4026263b8998SMiaohe Lin } 4027263b8998SMiaohe Lin #endif 4028263b8998SMiaohe Lin 4029a4a00b45SMuchun Song static void __init hugetlb_sysfs_init(void) 4030a4a00b45SMuchun Song { 4031a4a00b45SMuchun Song struct hstate *h; 4032a4a00b45SMuchun Song int err; 4033a4a00b45SMuchun Song 4034a4a00b45SMuchun Song hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 4035a4a00b45SMuchun Song if (!hugepages_kobj) 4036a4a00b45SMuchun Song return; 4037a4a00b45SMuchun Song 4038a4a00b45SMuchun Song for_each_hstate(h) { 4039a4a00b45SMuchun Song err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 4040a4a00b45SMuchun Song hstate_kobjs, &hstate_attr_group); 4041a4a00b45SMuchun Song if (err) 4042a4a00b45SMuchun Song pr_err("HugeTLB: Unable to add hstate %s", h->name); 4043a4a00b45SMuchun Song } 4044a4a00b45SMuchun Song 4045a4a00b45SMuchun Song #ifdef CONFIG_NUMA 4046a4a00b45SMuchun Song hugetlb_sysfs_initialized = true; 4047a4a00b45SMuchun Song #endif 4048a4a00b45SMuchun Song hugetlb_register_all_nodes(); 4049a4a00b45SMuchun Song } 4050a4a00b45SMuchun Song 4051a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 4052a3437870SNishanth Aravamudan { 40538382d914SDavidlohr Bueso int i; 40548382d914SDavidlohr Bueso 4055d6995da3SMike Kravetz BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < 4056d6995da3SMike Kravetz __NR_HPAGEFLAGS); 4057d6995da3SMike Kravetz 4058c2833a5bSMike Kravetz if (!hugepages_supported()) { 4059c2833a5bSMike Kravetz if (hugetlb_max_hstate || default_hstate_max_huge_pages) 4060c2833a5bSMike Kravetz pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); 40610ef89d25SBenjamin Herrenschmidt return 0; 4062d715cf80SLiam R. Howlett } 4063d715cf80SLiam R. Howlett 4064282f4214SMike Kravetz /* 4065282f4214SMike Kravetz * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some 4066282f4214SMike Kravetz * architectures depend on setup being done here. 4067282f4214SMike Kravetz */ 4068a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 4069282f4214SMike Kravetz if (!parsed_default_hugepagesz) { 4070282f4214SMike Kravetz /* 4071282f4214SMike Kravetz * If we did not parse a default huge page size, set 4072282f4214SMike Kravetz * default_hstate_idx to HPAGE_SIZE hstate. And, if the 4073282f4214SMike Kravetz * number of huge pages for this default size was implicitly 4074282f4214SMike Kravetz * specified, set that here as well. 4075282f4214SMike Kravetz * Note that the implicit setting will overwrite an explicit 4076282f4214SMike Kravetz * setting. A warning will be printed in this case. 4077282f4214SMike Kravetz */ 4078282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); 4079f8b74815SVaishali Thakkar if (default_hstate_max_huge_pages) { 4080282f4214SMike Kravetz if (default_hstate.max_huge_pages) { 4081282f4214SMike Kravetz char buf[32]; 4082282f4214SMike Kravetz 4083282f4214SMike Kravetz string_get_size(huge_page_size(&default_hstate), 4084282f4214SMike Kravetz 1, STRING_UNITS_2, buf, 32); 4085282f4214SMike Kravetz pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", 4086282f4214SMike Kravetz default_hstate.max_huge_pages, buf); 4087282f4214SMike Kravetz pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", 4088282f4214SMike Kravetz default_hstate_max_huge_pages); 4089282f4214SMike Kravetz } 4090282f4214SMike Kravetz default_hstate.max_huge_pages = 4091282f4214SMike Kravetz default_hstate_max_huge_pages; 4092b5389086SZhenguo Yao 40930a7a0f6fSPeng Liu for_each_online_node(i) 4094b5389086SZhenguo Yao default_hstate.max_huge_pages_node[i] = 4095b5389086SZhenguo Yao default_hugepages_in_node[i]; 4096282f4214SMike Kravetz } 4097f8b74815SVaishali Thakkar } 4098a3437870SNishanth Aravamudan 4099cf11e85fSRoman Gushchin hugetlb_cma_check(); 4100a3437870SNishanth Aravamudan hugetlb_init_hstates(); 4101aa888a74SAndi Kleen gather_bootmem_prealloc(); 4102a3437870SNishanth Aravamudan report_hugepages(); 4103a3437870SNishanth Aravamudan 4104a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 41057179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 41069a305230SLee Schermerhorn 41078382d914SDavidlohr Bueso #ifdef CONFIG_SMP 41088382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 41098382d914SDavidlohr Bueso #else 41108382d914SDavidlohr Bueso num_fault_mutexes = 1; 41118382d914SDavidlohr Bueso #endif 4112c672c7f2SMike Kravetz hugetlb_fault_mutex_table = 41136da2ec56SKees Cook kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 41146da2ec56SKees Cook GFP_KERNEL); 4115c672c7f2SMike Kravetz BUG_ON(!hugetlb_fault_mutex_table); 41168382d914SDavidlohr Bueso 41178382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 4118c672c7f2SMike Kravetz mutex_init(&hugetlb_fault_mutex_table[i]); 4119a3437870SNishanth Aravamudan return 0; 4120a3437870SNishanth Aravamudan } 41213e89e1c5SPaul Gortmaker subsys_initcall(hugetlb_init); 4122a3437870SNishanth Aravamudan 4123ae94da89SMike Kravetz /* Overwritten by architectures with more huge page sizes */ 4124ae94da89SMike Kravetz bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) 41259fee021dSVaishali Thakkar { 4126ae94da89SMike Kravetz return size == HPAGE_SIZE; 41279fee021dSVaishali Thakkar } 41289fee021dSVaishali Thakkar 4129d00181b9SKirill A. Shutemov void __init hugetlb_add_hstate(unsigned int order) 4130a3437870SNishanth Aravamudan { 4131a3437870SNishanth Aravamudan struct hstate *h; 41328faa8b07SAndi Kleen unsigned long i; 41338faa8b07SAndi Kleen 4134a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 4135a3437870SNishanth Aravamudan return; 4136a3437870SNishanth Aravamudan } 413747d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 4138a3437870SNishanth Aravamudan BUG_ON(order == 0); 413947d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 414029383967SMike Kravetz mutex_init(&h->resize_lock); 4141a3437870SNishanth Aravamudan h->order = order; 4142aca78307SMiaohe Lin h->mask = ~(huge_page_size(h) - 1); 41438faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 41448faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 41450edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 414654f18d35SAndrew Morton h->next_nid_to_alloc = first_memory_node; 414754f18d35SAndrew Morton h->next_nid_to_free = first_memory_node; 4148a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 4149c2c3a60aSMiaohe Lin huge_page_size(h)/SZ_1K); 41508faa8b07SAndi Kleen 4151a3437870SNishanth Aravamudan parsed_hstate = h; 4152a3437870SNishanth Aravamudan } 4153a3437870SNishanth Aravamudan 4154b5389086SZhenguo Yao bool __init __weak hugetlb_node_alloc_supported(void) 4155b5389086SZhenguo Yao { 4156b5389086SZhenguo Yao return true; 4157b5389086SZhenguo Yao } 4158f87442f4SPeng Liu 4159f87442f4SPeng Liu static void __init hugepages_clear_pages_in_node(void) 4160f87442f4SPeng Liu { 4161f87442f4SPeng Liu if (!hugetlb_max_hstate) { 4162f87442f4SPeng Liu default_hstate_max_huge_pages = 0; 4163f87442f4SPeng Liu memset(default_hugepages_in_node, 0, 416410395680SMiaohe Lin sizeof(default_hugepages_in_node)); 4165f87442f4SPeng Liu } else { 4166f87442f4SPeng Liu parsed_hstate->max_huge_pages = 0; 4167f87442f4SPeng Liu memset(parsed_hstate->max_huge_pages_node, 0, 416810395680SMiaohe Lin sizeof(parsed_hstate->max_huge_pages_node)); 4169f87442f4SPeng Liu } 4170f87442f4SPeng Liu } 4171f87442f4SPeng Liu 4172282f4214SMike Kravetz /* 4173282f4214SMike Kravetz * hugepages command line processing 4174282f4214SMike Kravetz * hugepages normally follows a valid hugepagsz or default_hugepagsz 4175282f4214SMike Kravetz * specification. If not, ignore the hugepages value. hugepages can also 4176282f4214SMike Kravetz * be the first huge page command line option in which case it implicitly 4177282f4214SMike Kravetz * specifies the number of huge pages for the default size. 4178282f4214SMike Kravetz */ 4179282f4214SMike Kravetz static int __init hugepages_setup(char *s) 4180a3437870SNishanth Aravamudan { 4181a3437870SNishanth Aravamudan unsigned long *mhp; 41828faa8b07SAndi Kleen static unsigned long *last_mhp; 4183b5389086SZhenguo Yao int node = NUMA_NO_NODE; 4184b5389086SZhenguo Yao int count; 4185b5389086SZhenguo Yao unsigned long tmp; 4186b5389086SZhenguo Yao char *p = s; 4187a3437870SNishanth Aravamudan 41889fee021dSVaishali Thakkar if (!parsed_valid_hugepagesz) { 4189282f4214SMike Kravetz pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); 41909fee021dSVaishali Thakkar parsed_valid_hugepagesz = true; 4191f81f6e4bSPeng Liu return 1; 41929fee021dSVaishali Thakkar } 4193282f4214SMike Kravetz 4194a3437870SNishanth Aravamudan /* 4195282f4214SMike Kravetz * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter 4196282f4214SMike Kravetz * yet, so this hugepages= parameter goes to the "default hstate". 4197282f4214SMike Kravetz * Otherwise, it goes with the previously parsed hugepagesz or 4198282f4214SMike Kravetz * default_hugepagesz. 4199a3437870SNishanth Aravamudan */ 42009fee021dSVaishali Thakkar else if (!hugetlb_max_hstate) 4201a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 4202a3437870SNishanth Aravamudan else 4203a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 4204a3437870SNishanth Aravamudan 42058faa8b07SAndi Kleen if (mhp == last_mhp) { 4206282f4214SMike Kravetz pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); 4207f81f6e4bSPeng Liu return 1; 42088faa8b07SAndi Kleen } 42098faa8b07SAndi Kleen 4210b5389086SZhenguo Yao while (*p) { 4211b5389086SZhenguo Yao count = 0; 4212b5389086SZhenguo Yao if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4213b5389086SZhenguo Yao goto invalid; 4214b5389086SZhenguo Yao /* Parameter is node format */ 4215b5389086SZhenguo Yao if (p[count] == ':') { 4216b5389086SZhenguo Yao if (!hugetlb_node_alloc_supported()) { 4217b5389086SZhenguo Yao pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n"); 4218f81f6e4bSPeng Liu return 1; 4219b5389086SZhenguo Yao } 42200a7a0f6fSPeng Liu if (tmp >= MAX_NUMNODES || !node_online(tmp)) 4221e79ce983SLiu Yuntao goto invalid; 42220a7a0f6fSPeng Liu node = array_index_nospec(tmp, MAX_NUMNODES); 4223b5389086SZhenguo Yao p += count + 1; 4224b5389086SZhenguo Yao /* Parse hugepages */ 4225b5389086SZhenguo Yao if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4226b5389086SZhenguo Yao goto invalid; 4227b5389086SZhenguo Yao if (!hugetlb_max_hstate) 4228b5389086SZhenguo Yao default_hugepages_in_node[node] = tmp; 4229b5389086SZhenguo Yao else 4230b5389086SZhenguo Yao parsed_hstate->max_huge_pages_node[node] = tmp; 4231b5389086SZhenguo Yao *mhp += tmp; 4232b5389086SZhenguo Yao /* Go to parse next node*/ 4233b5389086SZhenguo Yao if (p[count] == ',') 4234b5389086SZhenguo Yao p += count + 1; 4235b5389086SZhenguo Yao else 4236b5389086SZhenguo Yao break; 4237b5389086SZhenguo Yao } else { 4238b5389086SZhenguo Yao if (p != s) 4239b5389086SZhenguo Yao goto invalid; 4240b5389086SZhenguo Yao *mhp = tmp; 4241b5389086SZhenguo Yao break; 4242b5389086SZhenguo Yao } 4243b5389086SZhenguo Yao } 4244a3437870SNishanth Aravamudan 42458faa8b07SAndi Kleen /* 42468faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 424704adbc3fSMiaohe Lin * But we need to allocate gigantic hstates here early to still 42488faa8b07SAndi Kleen * use the bootmem allocator. 42498faa8b07SAndi Kleen */ 425004adbc3fSMiaohe Lin if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) 42518faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 42528faa8b07SAndi Kleen 42538faa8b07SAndi Kleen last_mhp = mhp; 42548faa8b07SAndi Kleen 4255a3437870SNishanth Aravamudan return 1; 4256b5389086SZhenguo Yao 4257b5389086SZhenguo Yao invalid: 4258b5389086SZhenguo Yao pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p); 4259f87442f4SPeng Liu hugepages_clear_pages_in_node(); 4260f81f6e4bSPeng Liu return 1; 4261a3437870SNishanth Aravamudan } 4262282f4214SMike Kravetz __setup("hugepages=", hugepages_setup); 4263e11bfbfcSNick Piggin 4264282f4214SMike Kravetz /* 4265282f4214SMike Kravetz * hugepagesz command line processing 4266282f4214SMike Kravetz * A specific huge page size can only be specified once with hugepagesz. 4267282f4214SMike Kravetz * hugepagesz is followed by hugepages on the command line. The global 4268282f4214SMike Kravetz * variable 'parsed_valid_hugepagesz' is used to determine if prior 4269282f4214SMike Kravetz * hugepagesz argument was valid. 4270282f4214SMike Kravetz */ 4271359f2544SMike Kravetz static int __init hugepagesz_setup(char *s) 4272e11bfbfcSNick Piggin { 4273359f2544SMike Kravetz unsigned long size; 4274282f4214SMike Kravetz struct hstate *h; 4275282f4214SMike Kravetz 4276282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4277359f2544SMike Kravetz size = (unsigned long)memparse(s, NULL); 4278359f2544SMike Kravetz 4279359f2544SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4280282f4214SMike Kravetz pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); 4281f81f6e4bSPeng Liu return 1; 4282359f2544SMike Kravetz } 4283359f2544SMike Kravetz 4284282f4214SMike Kravetz h = size_to_hstate(size); 4285282f4214SMike Kravetz if (h) { 4286282f4214SMike Kravetz /* 4287282f4214SMike Kravetz * hstate for this size already exists. This is normally 4288282f4214SMike Kravetz * an error, but is allowed if the existing hstate is the 4289282f4214SMike Kravetz * default hstate. More specifically, it is only allowed if 4290282f4214SMike Kravetz * the number of huge pages for the default hstate was not 4291282f4214SMike Kravetz * previously specified. 4292282f4214SMike Kravetz */ 4293282f4214SMike Kravetz if (!parsed_default_hugepagesz || h != &default_hstate || 4294282f4214SMike Kravetz default_hstate.max_huge_pages) { 4295282f4214SMike Kravetz pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); 4296f81f6e4bSPeng Liu return 1; 429738237830SMike Kravetz } 429838237830SMike Kravetz 4299282f4214SMike Kravetz /* 4300282f4214SMike Kravetz * No need to call hugetlb_add_hstate() as hstate already 4301282f4214SMike Kravetz * exists. But, do set parsed_hstate so that a following 4302282f4214SMike Kravetz * hugepages= parameter will be applied to this hstate. 4303282f4214SMike Kravetz */ 4304282f4214SMike Kravetz parsed_hstate = h; 4305282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4306e11bfbfcSNick Piggin return 1; 4307e11bfbfcSNick Piggin } 4308282f4214SMike Kravetz 4309359f2544SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4310282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4311359f2544SMike Kravetz return 1; 4312359f2544SMike Kravetz } 4313359f2544SMike Kravetz __setup("hugepagesz=", hugepagesz_setup); 4314359f2544SMike Kravetz 4315282f4214SMike Kravetz /* 4316282f4214SMike Kravetz * default_hugepagesz command line input 4317282f4214SMike Kravetz * Only one instance of default_hugepagesz allowed on command line. 4318282f4214SMike Kravetz */ 4319ae94da89SMike Kravetz static int __init default_hugepagesz_setup(char *s) 4320e11bfbfcSNick Piggin { 4321ae94da89SMike Kravetz unsigned long size; 4322b5389086SZhenguo Yao int i; 4323ae94da89SMike Kravetz 4324282f4214SMike Kravetz parsed_valid_hugepagesz = false; 4325282f4214SMike Kravetz if (parsed_default_hugepagesz) { 4326282f4214SMike Kravetz pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); 4327f81f6e4bSPeng Liu return 1; 4328282f4214SMike Kravetz } 4329282f4214SMike Kravetz 4330282f4214SMike Kravetz size = (unsigned long)memparse(s, NULL); 4331282f4214SMike Kravetz 4332282f4214SMike Kravetz if (!arch_hugetlb_valid_size(size)) { 4333282f4214SMike Kravetz pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); 4334f81f6e4bSPeng Liu return 1; 4335282f4214SMike Kravetz } 4336282f4214SMike Kravetz 4337282f4214SMike Kravetz hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4338282f4214SMike Kravetz parsed_valid_hugepagesz = true; 4339282f4214SMike Kravetz parsed_default_hugepagesz = true; 4340282f4214SMike Kravetz default_hstate_idx = hstate_index(size_to_hstate(size)); 4341282f4214SMike Kravetz 4342282f4214SMike Kravetz /* 4343282f4214SMike Kravetz * The number of default huge pages (for this size) could have been 4344282f4214SMike Kravetz * specified as the first hugetlb parameter: hugepages=X. If so, 4345282f4214SMike Kravetz * then default_hstate_max_huge_pages is set. If the default huge 4346282f4214SMike Kravetz * page size is gigantic (>= MAX_ORDER), then the pages must be 4347282f4214SMike Kravetz * allocated here from bootmem allocator. 4348282f4214SMike Kravetz */ 4349282f4214SMike Kravetz if (default_hstate_max_huge_pages) { 4350282f4214SMike Kravetz default_hstate.max_huge_pages = default_hstate_max_huge_pages; 43510a7a0f6fSPeng Liu for_each_online_node(i) 4352b5389086SZhenguo Yao default_hstate.max_huge_pages_node[i] = 4353b5389086SZhenguo Yao default_hugepages_in_node[i]; 4354282f4214SMike Kravetz if (hstate_is_gigantic(&default_hstate)) 4355282f4214SMike Kravetz hugetlb_hstate_alloc_pages(&default_hstate); 4356282f4214SMike Kravetz default_hstate_max_huge_pages = 0; 4357282f4214SMike Kravetz } 4358282f4214SMike Kravetz 4359e11bfbfcSNick Piggin return 1; 4360e11bfbfcSNick Piggin } 4361ae94da89SMike Kravetz __setup("default_hugepagesz=", default_hugepagesz_setup); 4362a3437870SNishanth Aravamudan 4363d2226ebdSFeng Tang static nodemask_t *policy_mbind_nodemask(gfp_t gfp) 4364d2226ebdSFeng Tang { 4365d2226ebdSFeng Tang #ifdef CONFIG_NUMA 4366d2226ebdSFeng Tang struct mempolicy *mpol = get_task_policy(current); 4367d2226ebdSFeng Tang 4368d2226ebdSFeng Tang /* 4369d2226ebdSFeng Tang * Only enforce MPOL_BIND policy which overlaps with cpuset policy 4370d2226ebdSFeng Tang * (from policy_nodemask) specifically for hugetlb case 4371d2226ebdSFeng Tang */ 4372d2226ebdSFeng Tang if (mpol->mode == MPOL_BIND && 4373d2226ebdSFeng Tang (apply_policy_zone(mpol, gfp_zone(gfp)) && 4374d2226ebdSFeng Tang cpuset_nodemask_valid_mems_allowed(&mpol->nodes))) 4375d2226ebdSFeng Tang return &mpol->nodes; 4376d2226ebdSFeng Tang #endif 4377d2226ebdSFeng Tang return NULL; 4378d2226ebdSFeng Tang } 4379d2226ebdSFeng Tang 43808ca39e68SMuchun Song static unsigned int allowed_mems_nr(struct hstate *h) 43818a213460SNishanth Aravamudan { 43828a213460SNishanth Aravamudan int node; 43838a213460SNishanth Aravamudan unsigned int nr = 0; 4384d2226ebdSFeng Tang nodemask_t *mbind_nodemask; 43858ca39e68SMuchun Song unsigned int *array = h->free_huge_pages_node; 43868ca39e68SMuchun Song gfp_t gfp_mask = htlb_alloc_mask(h); 43878a213460SNishanth Aravamudan 4388d2226ebdSFeng Tang mbind_nodemask = policy_mbind_nodemask(gfp_mask); 43898ca39e68SMuchun Song for_each_node_mask(node, cpuset_current_mems_allowed) { 4390d2226ebdSFeng Tang if (!mbind_nodemask || node_isset(node, *mbind_nodemask)) 43918a213460SNishanth Aravamudan nr += array[node]; 43928ca39e68SMuchun Song } 43938a213460SNishanth Aravamudan 43948a213460SNishanth Aravamudan return nr; 43958a213460SNishanth Aravamudan } 43968a213460SNishanth Aravamudan 43978a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 439817743798SMuchun Song static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, 439917743798SMuchun Song void *buffer, size_t *length, 440017743798SMuchun Song loff_t *ppos, unsigned long *out) 440117743798SMuchun Song { 440217743798SMuchun Song struct ctl_table dup_table; 440317743798SMuchun Song 440417743798SMuchun Song /* 440517743798SMuchun Song * In order to avoid races with __do_proc_doulongvec_minmax(), we 440617743798SMuchun Song * can duplicate the @table and alter the duplicate of it. 440717743798SMuchun Song */ 440817743798SMuchun Song dup_table = *table; 440917743798SMuchun Song dup_table.data = out; 441017743798SMuchun Song 441117743798SMuchun Song return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); 441217743798SMuchun Song } 441317743798SMuchun Song 441406808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 441506808b08SLee Schermerhorn struct ctl_table *table, int write, 441632927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 44171da177e4SLinus Torvalds { 4418e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 4419238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 442008d4a246SMichal Hocko int ret; 4421e5ff2159SAndi Kleen 4422457c1b27SNishanth Aravamudan if (!hugepages_supported()) 442386613628SJan Stancek return -EOPNOTSUPP; 4424457c1b27SNishanth Aravamudan 442517743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 442617743798SMuchun Song &tmp); 442708d4a246SMichal Hocko if (ret) 442808d4a246SMichal Hocko goto out; 4429e5ff2159SAndi Kleen 4430238d3c13SDavid Rientjes if (write) 4431238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 4432238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 443308d4a246SMichal Hocko out: 443408d4a246SMichal Hocko return ret; 44351da177e4SLinus Torvalds } 4436396faf03SMel Gorman 443706808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 443832927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 443906808b08SLee Schermerhorn { 444006808b08SLee Schermerhorn 444106808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 444206808b08SLee Schermerhorn buffer, length, ppos); 444306808b08SLee Schermerhorn } 444406808b08SLee Schermerhorn 444506808b08SLee Schermerhorn #ifdef CONFIG_NUMA 444606808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 444732927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 444806808b08SLee Schermerhorn { 444906808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 445006808b08SLee Schermerhorn buffer, length, ppos); 445106808b08SLee Schermerhorn } 445206808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 445306808b08SLee Schermerhorn 4454a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 445532927393SChristoph Hellwig void *buffer, size_t *length, loff_t *ppos) 4456a3d0c6aaSNishanth Aravamudan { 4457a5516438SAndi Kleen struct hstate *h = &default_hstate; 4458e5ff2159SAndi Kleen unsigned long tmp; 445908d4a246SMichal Hocko int ret; 4460e5ff2159SAndi Kleen 4461457c1b27SNishanth Aravamudan if (!hugepages_supported()) 446286613628SJan Stancek return -EOPNOTSUPP; 4463457c1b27SNishanth Aravamudan 4464e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 4465e5ff2159SAndi Kleen 4466bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 4467adbe8726SEric B Munson return -EINVAL; 4468adbe8726SEric B Munson 446917743798SMuchun Song ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 447017743798SMuchun Song &tmp); 447108d4a246SMichal Hocko if (ret) 447208d4a246SMichal Hocko goto out; 4473e5ff2159SAndi Kleen 4474e5ff2159SAndi Kleen if (write) { 4475db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4476e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 4477db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4478e5ff2159SAndi Kleen } 447908d4a246SMichal Hocko out: 448008d4a246SMichal Hocko return ret; 4481a3d0c6aaSNishanth Aravamudan } 4482a3d0c6aaSNishanth Aravamudan 44831da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 44841da177e4SLinus Torvalds 4485e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 44861da177e4SLinus Torvalds { 4487fcb2b0c5SRoman Gushchin struct hstate *h; 4488fcb2b0c5SRoman Gushchin unsigned long total = 0; 4489fcb2b0c5SRoman Gushchin 4490457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4491457c1b27SNishanth Aravamudan return; 4492fcb2b0c5SRoman Gushchin 4493fcb2b0c5SRoman Gushchin for_each_hstate(h) { 4494fcb2b0c5SRoman Gushchin unsigned long count = h->nr_huge_pages; 4495fcb2b0c5SRoman Gushchin 4496aca78307SMiaohe Lin total += huge_page_size(h) * count; 4497fcb2b0c5SRoman Gushchin 4498fcb2b0c5SRoman Gushchin if (h == &default_hstate) 4499e1759c21SAlexey Dobriyan seq_printf(m, 45001da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 45011da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 4502b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 45037893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 45044f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 4505fcb2b0c5SRoman Gushchin count, 4506a5516438SAndi Kleen h->free_huge_pages, 4507a5516438SAndi Kleen h->resv_huge_pages, 4508a5516438SAndi Kleen h->surplus_huge_pages, 4509aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4510fcb2b0c5SRoman Gushchin } 4511fcb2b0c5SRoman Gushchin 4512aca78307SMiaohe Lin seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); 45131da177e4SLinus Torvalds } 45141da177e4SLinus Torvalds 45157981593bSJoe Perches int hugetlb_report_node_meminfo(char *buf, int len, int nid) 45161da177e4SLinus Torvalds { 4517a5516438SAndi Kleen struct hstate *h = &default_hstate; 45187981593bSJoe Perches 4519457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4520457c1b27SNishanth Aravamudan return 0; 45217981593bSJoe Perches 45227981593bSJoe Perches return sysfs_emit_at(buf, len, 45231da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 4524a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 4525a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 4526a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 4527a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 4528a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 45291da177e4SLinus Torvalds } 45301da177e4SLinus Torvalds 4531dcadcf1cSGang Li void hugetlb_show_meminfo_node(int nid) 4532949f7ec5SDavid Rientjes { 4533949f7ec5SDavid Rientjes struct hstate *h; 4534949f7ec5SDavid Rientjes 4535457c1b27SNishanth Aravamudan if (!hugepages_supported()) 4536457c1b27SNishanth Aravamudan return; 4537457c1b27SNishanth Aravamudan 4538949f7ec5SDavid Rientjes for_each_hstate(h) 4539dcadcf1cSGang Li printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 4540949f7ec5SDavid Rientjes nid, 4541949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 4542949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 4543949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 4544aca78307SMiaohe Lin huge_page_size(h) / SZ_1K); 4545949f7ec5SDavid Rientjes } 4546949f7ec5SDavid Rientjes 45475d317b2bSNaoya Horiguchi void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 45485d317b2bSNaoya Horiguchi { 45495d317b2bSNaoya Horiguchi seq_printf(m, "HugetlbPages:\t%8lu kB\n", 45505d317b2bSNaoya Horiguchi atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); 45515d317b2bSNaoya Horiguchi } 45525d317b2bSNaoya Horiguchi 45531da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 45541da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 45551da177e4SLinus Torvalds { 4556d0028588SWanpeng Li struct hstate *h; 4557d0028588SWanpeng Li unsigned long nr_total_pages = 0; 4558d0028588SWanpeng Li 4559d0028588SWanpeng Li for_each_hstate(h) 4560d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 4561d0028588SWanpeng Li return nr_total_pages; 45621da177e4SLinus Torvalds } 45631da177e4SLinus Torvalds 4564a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 4565fc1b8a73SMel Gorman { 4566fc1b8a73SMel Gorman int ret = -ENOMEM; 4567fc1b8a73SMel Gorman 45680aa7f354SMiaohe Lin if (!delta) 45690aa7f354SMiaohe Lin return 0; 45700aa7f354SMiaohe Lin 4571db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 4572fc1b8a73SMel Gorman /* 4573fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 4574fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 4575fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 4576fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 4577fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 4578fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 4579fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 4580fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 4581fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 4582fc1b8a73SMel Gorman * 4583fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 4584fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 4585fc1b8a73SMel Gorman * we fall back to check against current free page availability as 4586fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 4587fc1b8a73SMel Gorman * semantics that cpuset has. 45888ca39e68SMuchun Song * 45898ca39e68SMuchun Song * Apart from cpuset, we also have memory policy mechanism that 45908ca39e68SMuchun Song * also determines from which node the kernel will allocate memory 45918ca39e68SMuchun Song * in a NUMA system. So similar to cpuset, we also should consider 45928ca39e68SMuchun Song * the memory policy of the current task. Similar to the description 45938ca39e68SMuchun Song * above. 4594fc1b8a73SMel Gorman */ 4595fc1b8a73SMel Gorman if (delta > 0) { 4596a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 4597fc1b8a73SMel Gorman goto out; 4598fc1b8a73SMel Gorman 45998ca39e68SMuchun Song if (delta > allowed_mems_nr(h)) { 4600a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 4601fc1b8a73SMel Gorman goto out; 4602fc1b8a73SMel Gorman } 4603fc1b8a73SMel Gorman } 4604fc1b8a73SMel Gorman 4605fc1b8a73SMel Gorman ret = 0; 4606fc1b8a73SMel Gorman if (delta < 0) 4607a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 4608fc1b8a73SMel Gorman 4609fc1b8a73SMel Gorman out: 4610db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 4611fc1b8a73SMel Gorman return ret; 4612fc1b8a73SMel Gorman } 4613fc1b8a73SMel Gorman 461484afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 461584afd99bSAndy Whitcroft { 4616f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 461784afd99bSAndy Whitcroft 461884afd99bSAndy Whitcroft /* 4619612b8a31SMike Kravetz * HPAGE_RESV_OWNER indicates a private mapping. 462084afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 462184afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 462284afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 462325985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 462484afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 462584afd99bSAndy Whitcroft * new reference here without additional locking. 462684afd99bSAndy Whitcroft */ 462709a26e83SMike Kravetz if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 462809a26e83SMike Kravetz resv_map_dup_hugetlb_cgroup_uncharge_info(resv); 4629f522c3acSJoonsoo Kim kref_get(&resv->refs); 463084afd99bSAndy Whitcroft } 46318d9bfb26SMike Kravetz 4632131a79b4SMike Kravetz /* 4633131a79b4SMike Kravetz * vma_lock structure for sharable mappings is vma specific. 4634612b8a31SMike Kravetz * Clear old pointer (if copied via vm_area_dup) and allocate 4635612b8a31SMike Kravetz * new structure. Before clearing, make sure vma_lock is not 4636612b8a31SMike Kravetz * for this vma. 4637131a79b4SMike Kravetz */ 4638131a79b4SMike Kravetz if (vma->vm_flags & VM_MAYSHARE) { 4639612b8a31SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 4640612b8a31SMike Kravetz 4641612b8a31SMike Kravetz if (vma_lock) { 4642612b8a31SMike Kravetz if (vma_lock->vma != vma) { 4643131a79b4SMike Kravetz vma->vm_private_data = NULL; 46448d9bfb26SMike Kravetz hugetlb_vma_lock_alloc(vma); 4645612b8a31SMike Kravetz } else 4646612b8a31SMike Kravetz pr_warn("HugeTLB: vma_lock already exists in %s.\n", __func__); 4647612b8a31SMike Kravetz } else 4648612b8a31SMike Kravetz hugetlb_vma_lock_alloc(vma); 464909a26e83SMike Kravetz } 4650131a79b4SMike Kravetz } 465184afd99bSAndy Whitcroft 4652a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 4653a1e78772SMel Gorman { 4654a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 46558d9bfb26SMike Kravetz struct resv_map *resv; 465690481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 46574e35f483SJoonsoo Kim unsigned long reserve, start, end; 46581c5ecae3SMike Kravetz long gbl_reserve; 465984afd99bSAndy Whitcroft 46608d9bfb26SMike Kravetz hugetlb_vma_lock_free(vma); 46618d9bfb26SMike Kravetz 46628d9bfb26SMike Kravetz resv = vma_resv_map(vma); 46634e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 46644e35f483SJoonsoo Kim return; 46654e35f483SJoonsoo Kim 4666a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 4667a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 466884afd99bSAndy Whitcroft 46694e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 4670e9fe92aeSMina Almasry hugetlb_cgroup_uncharge_counter(resv, start, end); 46717251ff78SAdam Litke if (reserve) { 46721c5ecae3SMike Kravetz /* 46731c5ecae3SMike Kravetz * Decrement reserve counts. The global reserve count may be 46741c5ecae3SMike Kravetz * adjusted if the subpool has a minimum size. 46751c5ecae3SMike Kravetz */ 46761c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 46771c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 46787251ff78SAdam Litke } 4679e9fe92aeSMina Almasry 4680e9fe92aeSMina Almasry kref_put(&resv->refs, resv_map_release); 4681a1e78772SMel Gorman } 4682a1e78772SMel Gorman 468331383c68SDan Williams static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 468431383c68SDan Williams { 468531383c68SDan Williams if (addr & ~(huge_page_mask(hstate_vma(vma)))) 468631383c68SDan Williams return -EINVAL; 468731383c68SDan Williams return 0; 468831383c68SDan Williams } 468931383c68SDan Williams 469005ea8860SDan Williams static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 469105ea8860SDan Williams { 4692aca78307SMiaohe Lin return huge_page_size(hstate_vma(vma)); 469305ea8860SDan Williams } 469405ea8860SDan Williams 46951da177e4SLinus Torvalds /* 46961da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 46971da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 46986c26d310SMiaohe Lin * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get 46991da177e4SLinus Torvalds * this far. 47001da177e4SLinus Torvalds */ 4701b3ec9f33SSouptick Joarder static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 47021da177e4SLinus Torvalds { 47031da177e4SLinus Torvalds BUG(); 4704d0217ac0SNick Piggin return 0; 47051da177e4SLinus Torvalds } 47061da177e4SLinus Torvalds 4707eec3636aSJane Chu /* 4708eec3636aSJane Chu * When a new function is introduced to vm_operations_struct and added 4709eec3636aSJane Chu * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 4710eec3636aSJane Chu * This is because under System V memory model, mappings created via 4711eec3636aSJane Chu * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 4712eec3636aSJane Chu * their original vm_ops are overwritten with shm_vm_ops. 4713eec3636aSJane Chu */ 4714f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 4715d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 471684afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 4717a1e78772SMel Gorman .close = hugetlb_vm_op_close, 4718dd3b614fSDmitry Safonov .may_split = hugetlb_vm_op_split, 471905ea8860SDan Williams .pagesize = hugetlb_vm_op_pagesize, 47201da177e4SLinus Torvalds }; 47211da177e4SLinus Torvalds 47221e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 47231e8f889bSDavid Gibson int writable) 472463551ae0SDavid Gibson { 472563551ae0SDavid Gibson pte_t entry; 472679c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 472763551ae0SDavid Gibson 47281e8f889bSDavid Gibson if (writable) { 4729106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 4730106c992aSGerald Schaefer vma->vm_page_prot))); 473163551ae0SDavid Gibson } else { 4732106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 4733106c992aSGerald Schaefer vma->vm_page_prot)); 473463551ae0SDavid Gibson } 473563551ae0SDavid Gibson entry = pte_mkyoung(entry); 473679c1c594SChristophe Leroy entry = arch_make_huge_pte(entry, shift, vma->vm_flags); 473763551ae0SDavid Gibson 473863551ae0SDavid Gibson return entry; 473963551ae0SDavid Gibson } 474063551ae0SDavid Gibson 47411e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 47421e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 47431e8f889bSDavid Gibson { 47441e8f889bSDavid Gibson pte_t entry; 47451e8f889bSDavid Gibson 4746106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 474732f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 47484b3073e1SRussell King update_mmu_cache(vma, address, ptep); 47491e8f889bSDavid Gibson } 47501e8f889bSDavid Gibson 4751d5ed7444SAneesh Kumar K.V bool is_hugetlb_entry_migration(pte_t pte) 47524a705fefSNaoya Horiguchi { 47534a705fefSNaoya Horiguchi swp_entry_t swp; 47544a705fefSNaoya Horiguchi 47554a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 4756d5ed7444SAneesh Kumar K.V return false; 47574a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4758d79d176aSBaoquan He if (is_migration_entry(swp)) 4759d5ed7444SAneesh Kumar K.V return true; 47604a705fefSNaoya Horiguchi else 4761d5ed7444SAneesh Kumar K.V return false; 47624a705fefSNaoya Horiguchi } 47634a705fefSNaoya Horiguchi 47643e5c3600SBaoquan He static bool is_hugetlb_entry_hwpoisoned(pte_t pte) 47654a705fefSNaoya Horiguchi { 47664a705fefSNaoya Horiguchi swp_entry_t swp; 47674a705fefSNaoya Horiguchi 47684a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 47693e5c3600SBaoquan He return false; 47704a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 4771d79d176aSBaoquan He if (is_hwpoison_entry(swp)) 47723e5c3600SBaoquan He return true; 47734a705fefSNaoya Horiguchi else 47743e5c3600SBaoquan He return false; 47754a705fefSNaoya Horiguchi } 47761e8f889bSDavid Gibson 47774eae4efaSPeter Xu static void 47784eae4efaSPeter Xu hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, 47794eae4efaSPeter Xu struct page *new_page) 47804eae4efaSPeter Xu { 47814eae4efaSPeter Xu __SetPageUptodate(new_page); 47824eae4efaSPeter Xu hugepage_add_new_anon_rmap(new_page, vma, addr); 47831eba86c0SPasha Tatashin set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1)); 47844eae4efaSPeter Xu hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); 47854eae4efaSPeter Xu SetHPageMigratable(new_page); 47864eae4efaSPeter Xu } 47874eae4efaSPeter Xu 478863551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 4789bc70fbf2SPeter Xu struct vm_area_struct *dst_vma, 4790bc70fbf2SPeter Xu struct vm_area_struct *src_vma) 479163551ae0SDavid Gibson { 47923aa4ed80SMiaohe Lin pte_t *src_pte, *dst_pte, entry; 479363551ae0SDavid Gibson struct page *ptepage; 47941c59827dSHugh Dickins unsigned long addr; 4795bc70fbf2SPeter Xu bool cow = is_cow_mapping(src_vma->vm_flags); 4796bc70fbf2SPeter Xu struct hstate *h = hstate_vma(src_vma); 4797a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 47984eae4efaSPeter Xu unsigned long npages = pages_per_huge_page(h); 4799ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 4800e95a9851SMike Kravetz unsigned long last_addr_mask; 4801e8569dd2SAndreas Sandberg int ret = 0; 48021e8f889bSDavid Gibson 4803ac46d4f3SJérôme Glisse if (cow) { 4804bc70fbf2SPeter Xu mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src_vma, src, 4805bc70fbf2SPeter Xu src_vma->vm_start, 4806bc70fbf2SPeter Xu src_vma->vm_end); 4807ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 4808623a1ddfSDavid Hildenbrand mmap_assert_write_locked(src); 4809623a1ddfSDavid Hildenbrand raw_write_seqcount_begin(&src->write_protect_seq); 481040549ba8SMike Kravetz } else { 481140549ba8SMike Kravetz /* 481240549ba8SMike Kravetz * For shared mappings the vma lock must be held before 481340549ba8SMike Kravetz * calling huge_pte_offset in the src vma. Otherwise, the 481440549ba8SMike Kravetz * returned ptep could go away if part of a shared pmd and 481540549ba8SMike Kravetz * another thread calls huge_pmd_unshare. 481640549ba8SMike Kravetz */ 481740549ba8SMike Kravetz hugetlb_vma_lock_read(src_vma); 4818ac46d4f3SJérôme Glisse } 4819e8569dd2SAndreas Sandberg 4820e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 4821bc70fbf2SPeter Xu for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) { 4822cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 48237868a208SPunit Agrawal src_pte = huge_pte_offset(src, addr, sz); 4824e95a9851SMike Kravetz if (!src_pte) { 4825e95a9851SMike Kravetz addr |= last_addr_mask; 4826c74df32cSHugh Dickins continue; 4827e95a9851SMike Kravetz } 4828bc70fbf2SPeter Xu dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz); 4829e8569dd2SAndreas Sandberg if (!dst_pte) { 4830e8569dd2SAndreas Sandberg ret = -ENOMEM; 4831e8569dd2SAndreas Sandberg break; 4832e8569dd2SAndreas Sandberg } 4833c5c99429SLarry Woodman 48345e41540cSMike Kravetz /* 48355e41540cSMike Kravetz * If the pagetables are shared don't copy or take references. 48365e41540cSMike Kravetz * 48373aa4ed80SMiaohe Lin * dst_pte == src_pte is the common case of src/dest sharing. 48385e41540cSMike Kravetz * However, src could have 'unshared' and dst shares with 48393aa4ed80SMiaohe Lin * another vma. So page_count of ptep page is checked instead 48403aa4ed80SMiaohe Lin * to reliably determine whether pte is shared. 48415e41540cSMike Kravetz */ 48423aa4ed80SMiaohe Lin if (page_count(virt_to_page(dst_pte)) > 1) { 4843e95a9851SMike Kravetz addr |= last_addr_mask; 4844c5c99429SLarry Woodman continue; 4845e95a9851SMike Kravetz } 4846c5c99429SLarry Woodman 4847cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 4848cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 4849cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 48504a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 48514eae4efaSPeter Xu again: 48523aa4ed80SMiaohe Lin if (huge_pte_none(entry)) { 48535e41540cSMike Kravetz /* 48543aa4ed80SMiaohe Lin * Skip if src entry none. 48555e41540cSMike Kravetz */ 48564a705fefSNaoya Horiguchi ; 4857c2cb0dccSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) { 4858c2cb0dccSNaoya Horiguchi bool uffd_wp = huge_pte_uffd_wp(entry); 4859c2cb0dccSNaoya Horiguchi 4860c2cb0dccSNaoya Horiguchi if (!userfaultfd_wp(dst_vma) && uffd_wp) 4861c2cb0dccSNaoya Horiguchi entry = huge_pte_clear_uffd_wp(entry); 4862c2cb0dccSNaoya Horiguchi set_huge_pte_at(dst, addr, dst_pte, entry); 4863c2cb0dccSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry))) { 48644a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 4865bc70fbf2SPeter Xu bool uffd_wp = huge_pte_uffd_wp(entry); 48664a705fefSNaoya Horiguchi 48676c287605SDavid Hildenbrand if (!is_readable_migration_entry(swp_entry) && cow) { 48684a705fefSNaoya Horiguchi /* 48694a705fefSNaoya Horiguchi * COW mappings require pages in both 48704a705fefSNaoya Horiguchi * parent and child to be set to read. 48714a705fefSNaoya Horiguchi */ 48724dd845b5SAlistair Popple swp_entry = make_readable_migration_entry( 48734dd845b5SAlistair Popple swp_offset(swp_entry)); 48744a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 4875bc70fbf2SPeter Xu if (userfaultfd_wp(src_vma) && uffd_wp) 4876bc70fbf2SPeter Xu entry = huge_pte_mkuffd_wp(entry); 487718f39629SQi Zheng set_huge_pte_at(src, addr, src_pte, entry); 48784a705fefSNaoya Horiguchi } 4879bc70fbf2SPeter Xu if (!userfaultfd_wp(dst_vma) && uffd_wp) 4880bc70fbf2SPeter Xu entry = huge_pte_clear_uffd_wp(entry); 488118f39629SQi Zheng set_huge_pte_at(dst, addr, dst_pte, entry); 4882bc70fbf2SPeter Xu } else if (unlikely(is_pte_marker(entry))) { 4883bc70fbf2SPeter Xu /* 4884bc70fbf2SPeter Xu * We copy the pte marker only if the dst vma has 4885bc70fbf2SPeter Xu * uffd-wp enabled. 4886bc70fbf2SPeter Xu */ 4887bc70fbf2SPeter Xu if (userfaultfd_wp(dst_vma)) 4888bc70fbf2SPeter Xu set_huge_pte_at(dst, addr, dst_pte, entry); 48894a705fefSNaoya Horiguchi } else { 48904eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 48914eae4efaSPeter Xu ptepage = pte_page(entry); 48924eae4efaSPeter Xu get_page(ptepage); 48934eae4efaSPeter Xu 48944eae4efaSPeter Xu /* 4895fb3d824dSDavid Hildenbrand * Failing to duplicate the anon rmap is a rare case 4896fb3d824dSDavid Hildenbrand * where we see pinned hugetlb pages while they're 4897fb3d824dSDavid Hildenbrand * prone to COW. We need to do the COW earlier during 4898fb3d824dSDavid Hildenbrand * fork. 48994eae4efaSPeter Xu * 49004eae4efaSPeter Xu * When pre-allocating the page or copying data, we 49014eae4efaSPeter Xu * need to be without the pgtable locks since we could 49024eae4efaSPeter Xu * sleep during the process. 49034eae4efaSPeter Xu */ 4904fb3d824dSDavid Hildenbrand if (!PageAnon(ptepage)) { 4905fb3d824dSDavid Hildenbrand page_dup_file_rmap(ptepage, true); 4906bc70fbf2SPeter Xu } else if (page_try_dup_anon_rmap(ptepage, true, 4907bc70fbf2SPeter Xu src_vma)) { 49084eae4efaSPeter Xu pte_t src_pte_old = entry; 49094eae4efaSPeter Xu struct page *new; 49104eae4efaSPeter Xu 49114eae4efaSPeter Xu spin_unlock(src_ptl); 49124eae4efaSPeter Xu spin_unlock(dst_ptl); 49134eae4efaSPeter Xu /* Do not use reserve as it's private owned */ 4914bc70fbf2SPeter Xu new = alloc_huge_page(dst_vma, addr, 1); 49154eae4efaSPeter Xu if (IS_ERR(new)) { 49164eae4efaSPeter Xu put_page(ptepage); 49174eae4efaSPeter Xu ret = PTR_ERR(new); 49184eae4efaSPeter Xu break; 49194eae4efaSPeter Xu } 4920bc70fbf2SPeter Xu copy_user_huge_page(new, ptepage, addr, dst_vma, 49214eae4efaSPeter Xu npages); 49224eae4efaSPeter Xu put_page(ptepage); 49234eae4efaSPeter Xu 49244eae4efaSPeter Xu /* Install the new huge page if src pte stable */ 49254eae4efaSPeter Xu dst_ptl = huge_pte_lock(h, dst, dst_pte); 49264eae4efaSPeter Xu src_ptl = huge_pte_lockptr(h, src, src_pte); 49274eae4efaSPeter Xu spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 49284eae4efaSPeter Xu entry = huge_ptep_get(src_pte); 49294eae4efaSPeter Xu if (!pte_same(src_pte_old, entry)) { 4930bc70fbf2SPeter Xu restore_reserve_on_error(h, dst_vma, addr, 4931846be085SMike Kravetz new); 49324eae4efaSPeter Xu put_page(new); 49333aa4ed80SMiaohe Lin /* huge_ptep of dst_pte won't change as in child */ 49344eae4efaSPeter Xu goto again; 49354eae4efaSPeter Xu } 4936bc70fbf2SPeter Xu hugetlb_install_page(dst_vma, dst_pte, addr, new); 49374eae4efaSPeter Xu spin_unlock(src_ptl); 49384eae4efaSPeter Xu spin_unlock(dst_ptl); 49394eae4efaSPeter Xu continue; 49404eae4efaSPeter Xu } 49414eae4efaSPeter Xu 494234ee645eSJoerg Roedel if (cow) { 49430f10851eSJérôme Glisse /* 49440f10851eSJérôme Glisse * No need to notify as we are downgrading page 49450f10851eSJérôme Glisse * table protection not changing it to point 49460f10851eSJérôme Glisse * to a new page. 49470f10851eSJérôme Glisse * 4948ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 49490f10851eSJérôme Glisse */ 49507f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 495184894e1cSPeter Xu entry = huge_pte_wrprotect(entry); 495234ee645eSJoerg Roedel } 49534eae4efaSPeter Xu 495463551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 49554eae4efaSPeter Xu hugetlb_count_add(npages, dst); 49561c59827dSHugh Dickins } 4957cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 4958cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 495963551ae0SDavid Gibson } 496063551ae0SDavid Gibson 4961623a1ddfSDavid Hildenbrand if (cow) { 4962623a1ddfSDavid Hildenbrand raw_write_seqcount_end(&src->write_protect_seq); 4963ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 496440549ba8SMike Kravetz } else { 496540549ba8SMike Kravetz hugetlb_vma_unlock_read(src_vma); 4966623a1ddfSDavid Hildenbrand } 4967e8569dd2SAndreas Sandberg 4968e8569dd2SAndreas Sandberg return ret; 496963551ae0SDavid Gibson } 497063551ae0SDavid Gibson 4971550a7d60SMina Almasry static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr, 4972db110a99SAneesh Kumar K.V unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte) 4973550a7d60SMina Almasry { 4974550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 4975550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 4976550a7d60SMina Almasry spinlock_t *src_ptl, *dst_ptl; 4977db110a99SAneesh Kumar K.V pte_t pte; 4978550a7d60SMina Almasry 4979550a7d60SMina Almasry dst_ptl = huge_pte_lock(h, mm, dst_pte); 4980550a7d60SMina Almasry src_ptl = huge_pte_lockptr(h, mm, src_pte); 4981550a7d60SMina Almasry 4982550a7d60SMina Almasry /* 4983550a7d60SMina Almasry * We don't have to worry about the ordering of src and dst ptlocks 4984550a7d60SMina Almasry * because exclusive mmap_sem (or the i_mmap_lock) prevents deadlock. 4985550a7d60SMina Almasry */ 4986550a7d60SMina Almasry if (src_ptl != dst_ptl) 4987550a7d60SMina Almasry spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 4988550a7d60SMina Almasry 4989550a7d60SMina Almasry pte = huge_ptep_get_and_clear(mm, old_addr, src_pte); 4990550a7d60SMina Almasry set_huge_pte_at(mm, new_addr, dst_pte, pte); 4991550a7d60SMina Almasry 4992550a7d60SMina Almasry if (src_ptl != dst_ptl) 4993550a7d60SMina Almasry spin_unlock(src_ptl); 4994550a7d60SMina Almasry spin_unlock(dst_ptl); 4995550a7d60SMina Almasry } 4996550a7d60SMina Almasry 4997550a7d60SMina Almasry int move_hugetlb_page_tables(struct vm_area_struct *vma, 4998550a7d60SMina Almasry struct vm_area_struct *new_vma, 4999550a7d60SMina Almasry unsigned long old_addr, unsigned long new_addr, 5000550a7d60SMina Almasry unsigned long len) 5001550a7d60SMina Almasry { 5002550a7d60SMina Almasry struct hstate *h = hstate_vma(vma); 5003550a7d60SMina Almasry struct address_space *mapping = vma->vm_file->f_mapping; 5004550a7d60SMina Almasry unsigned long sz = huge_page_size(h); 5005550a7d60SMina Almasry struct mm_struct *mm = vma->vm_mm; 5006550a7d60SMina Almasry unsigned long old_end = old_addr + len; 5007e95a9851SMike Kravetz unsigned long last_addr_mask; 5008550a7d60SMina Almasry pte_t *src_pte, *dst_pte; 5009550a7d60SMina Almasry struct mmu_notifier_range range; 50103d0b95cdSBaolin Wang bool shared_pmd = false; 5011550a7d60SMina Almasry 5012550a7d60SMina Almasry mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, old_addr, 5013550a7d60SMina Almasry old_end); 5014550a7d60SMina Almasry adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 50153d0b95cdSBaolin Wang /* 50163d0b95cdSBaolin Wang * In case of shared PMDs, we should cover the maximum possible 50173d0b95cdSBaolin Wang * range. 50183d0b95cdSBaolin Wang */ 50193d0b95cdSBaolin Wang flush_cache_range(vma, range.start, range.end); 50203d0b95cdSBaolin Wang 5021550a7d60SMina Almasry mmu_notifier_invalidate_range_start(&range); 5022e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 5023550a7d60SMina Almasry /* Prevent race with file truncation */ 502440549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 5025550a7d60SMina Almasry i_mmap_lock_write(mapping); 5026550a7d60SMina Almasry for (; old_addr < old_end; old_addr += sz, new_addr += sz) { 5027550a7d60SMina Almasry src_pte = huge_pte_offset(mm, old_addr, sz); 5028e95a9851SMike Kravetz if (!src_pte) { 5029e95a9851SMike Kravetz old_addr |= last_addr_mask; 5030e95a9851SMike Kravetz new_addr |= last_addr_mask; 5031550a7d60SMina Almasry continue; 5032e95a9851SMike Kravetz } 5033550a7d60SMina Almasry if (huge_pte_none(huge_ptep_get(src_pte))) 5034550a7d60SMina Almasry continue; 5035550a7d60SMina Almasry 50364ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) { 50373d0b95cdSBaolin Wang shared_pmd = true; 50384ddb4d91SMike Kravetz old_addr |= last_addr_mask; 50394ddb4d91SMike Kravetz new_addr |= last_addr_mask; 5040550a7d60SMina Almasry continue; 50413d0b95cdSBaolin Wang } 5042550a7d60SMina Almasry 5043550a7d60SMina Almasry dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz); 5044550a7d60SMina Almasry if (!dst_pte) 5045550a7d60SMina Almasry break; 5046550a7d60SMina Almasry 5047db110a99SAneesh Kumar K.V move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte); 5048550a7d60SMina Almasry } 50493d0b95cdSBaolin Wang 50503d0b95cdSBaolin Wang if (shared_pmd) 50513d0b95cdSBaolin Wang flush_tlb_range(vma, range.start, range.end); 50523d0b95cdSBaolin Wang else 5053550a7d60SMina Almasry flush_tlb_range(vma, old_end - len, old_end); 5054550a7d60SMina Almasry mmu_notifier_invalidate_range_end(&range); 505513e4ad2cSNadav Amit i_mmap_unlock_write(mapping); 505640549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 5057550a7d60SMina Almasry 5058550a7d60SMina Almasry return len + old_addr - old_end; 5059550a7d60SMina Almasry } 5060550a7d60SMina Almasry 506173c54763SPeter Xu static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 506224669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 506305e90bd0SPeter Xu struct page *ref_page, zap_flags_t zap_flags) 506463551ae0SDavid Gibson { 506563551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 506663551ae0SDavid Gibson unsigned long address; 5067c7546f8fSDavid Gibson pte_t *ptep; 506863551ae0SDavid Gibson pte_t pte; 5069cb900f41SKirill A. Shutemov spinlock_t *ptl; 507063551ae0SDavid Gibson struct page *page; 5071a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 5072a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 5073ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 5074e95a9851SMike Kravetz unsigned long last_addr_mask; 5075a4a118f2SNadav Amit bool force_flush = false; 5076a5516438SAndi Kleen 507763551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 5078a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 5079a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 508063551ae0SDavid Gibson 508107e32661SAneesh Kumar K.V /* 508207e32661SAneesh Kumar K.V * This is a hugetlb vma, all the pte entries should point 508307e32661SAneesh Kumar K.V * to huge page. 508407e32661SAneesh Kumar K.V */ 5085ed6a7935SPeter Zijlstra tlb_change_page_size(tlb, sz); 508624669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 5087dff11abeSMike Kravetz 5088dff11abeSMike Kravetz /* 5089dff11abeSMike Kravetz * If sharing possible, alert mmu notifiers of worst case. 5090dff11abeSMike Kravetz */ 50916f4f13e8SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, 50926f4f13e8SJérôme Glisse end); 5093ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 5094ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 5095e95a9851SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 5096569f48b8SHillf Danton address = start; 5097569f48b8SHillf Danton for (; address < end; address += sz) { 50987868a208SPunit Agrawal ptep = huge_pte_offset(mm, address, sz); 5099e95a9851SMike Kravetz if (!ptep) { 5100e95a9851SMike Kravetz address |= last_addr_mask; 5101c7546f8fSDavid Gibson continue; 5102e95a9851SMike Kravetz } 5103c7546f8fSDavid Gibson 5104cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 51054ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, address, ptep)) { 510631d49da5SAneesh Kumar K.V spin_unlock(ptl); 5107a4a118f2SNadav Amit tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE); 5108a4a118f2SNadav Amit force_flush = true; 51094ddb4d91SMike Kravetz address |= last_addr_mask; 511031d49da5SAneesh Kumar K.V continue; 511131d49da5SAneesh Kumar K.V } 511239dde65cSChen, Kenneth W 51136629326bSHillf Danton pte = huge_ptep_get(ptep); 511431d49da5SAneesh Kumar K.V if (huge_pte_none(pte)) { 511531d49da5SAneesh Kumar K.V spin_unlock(ptl); 511631d49da5SAneesh Kumar K.V continue; 511731d49da5SAneesh Kumar K.V } 51186629326bSHillf Danton 51196629326bSHillf Danton /* 51209fbc1f63SNaoya Horiguchi * Migrating hugepage or HWPoisoned hugepage is already 51219fbc1f63SNaoya Horiguchi * unmapped and its refcount is dropped, so just clear pte here. 51226629326bSHillf Danton */ 51239fbc1f63SNaoya Horiguchi if (unlikely(!pte_present(pte))) { 512405e90bd0SPeter Xu /* 512505e90bd0SPeter Xu * If the pte was wr-protected by uffd-wp in any of the 512605e90bd0SPeter Xu * swap forms, meanwhile the caller does not want to 512705e90bd0SPeter Xu * drop the uffd-wp bit in this zap, then replace the 512805e90bd0SPeter Xu * pte with a marker. 512905e90bd0SPeter Xu */ 513005e90bd0SPeter Xu if (pte_swp_uffd_wp_any(pte) && 513105e90bd0SPeter Xu !(zap_flags & ZAP_FLAG_DROP_MARKER)) 513205e90bd0SPeter Xu set_huge_pte_at(mm, address, ptep, 513305e90bd0SPeter Xu make_pte_marker(PTE_MARKER_UFFD_WP)); 513405e90bd0SPeter Xu else 51359386fac3SPunit Agrawal huge_pte_clear(mm, address, ptep, sz); 513631d49da5SAneesh Kumar K.V spin_unlock(ptl); 513731d49da5SAneesh Kumar K.V continue; 51388c4894c6SNaoya Horiguchi } 51396629326bSHillf Danton 51406629326bSHillf Danton page = pte_page(pte); 514104f2cbe3SMel Gorman /* 514204f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 514304f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 514404f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 514504f2cbe3SMel Gorman */ 514604f2cbe3SMel Gorman if (ref_page) { 514731d49da5SAneesh Kumar K.V if (page != ref_page) { 514831d49da5SAneesh Kumar K.V spin_unlock(ptl); 514931d49da5SAneesh Kumar K.V continue; 515031d49da5SAneesh Kumar K.V } 515104f2cbe3SMel Gorman /* 515204f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 515304f2cbe3SMel Gorman * future faults in this VMA will fail rather than 515404f2cbe3SMel Gorman * looking like data was lost 515504f2cbe3SMel Gorman */ 515604f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 515704f2cbe3SMel Gorman } 515804f2cbe3SMel Gorman 5159c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 5160b528e4b6SAneesh Kumar K.V tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 5161106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 51626649a386SKen Chen set_page_dirty(page); 516305e90bd0SPeter Xu /* Leave a uffd-wp pte marker if needed */ 516405e90bd0SPeter Xu if (huge_pte_uffd_wp(pte) && 516505e90bd0SPeter Xu !(zap_flags & ZAP_FLAG_DROP_MARKER)) 516605e90bd0SPeter Xu set_huge_pte_at(mm, address, ptep, 516705e90bd0SPeter Xu make_pte_marker(PTE_MARKER_UFFD_WP)); 51685d317b2bSNaoya Horiguchi hugetlb_count_sub(pages_per_huge_page(h), mm); 5169cea86fe2SHugh Dickins page_remove_rmap(page, vma, true); 517031d49da5SAneesh Kumar K.V 5171cb900f41SKirill A. Shutemov spin_unlock(ptl); 5172e77b0852SAneesh Kumar K.V tlb_remove_page_size(tlb, page, huge_page_size(h)); 517324669e58SAneesh Kumar K.V /* 517431d49da5SAneesh Kumar K.V * Bail out after unmapping reference page if supplied 517524669e58SAneesh Kumar K.V */ 517631d49da5SAneesh Kumar K.V if (ref_page) 517731d49da5SAneesh Kumar K.V break; 5178fe1668aeSChen, Kenneth W } 5179ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 518024669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 5181a4a118f2SNadav Amit 5182a4a118f2SNadav Amit /* 5183a4a118f2SNadav Amit * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We 5184a4a118f2SNadav Amit * could defer the flush until now, since by holding i_mmap_rwsem we 5185a4a118f2SNadav Amit * guaranteed that the last refernece would not be dropped. But we must 5186a4a118f2SNadav Amit * do the flushing before we return, as otherwise i_mmap_rwsem will be 5187a4a118f2SNadav Amit * dropped and the last reference to the shared PMDs page might be 5188a4a118f2SNadav Amit * dropped as well. 5189a4a118f2SNadav Amit * 5190a4a118f2SNadav Amit * In theory we could defer the freeing of the PMD pages as well, but 5191a4a118f2SNadav Amit * huge_pmd_unshare() relies on the exact page_count for the PMD page to 5192a4a118f2SNadav Amit * detect sharing, so we cannot defer the release of the page either. 5193a4a118f2SNadav Amit * Instead, do flush now. 5194a4a118f2SNadav Amit */ 5195a4a118f2SNadav Amit if (force_flush) 5196a4a118f2SNadav Amit tlb_flush_mmu_tlbonly(tlb); 51971da177e4SLinus Torvalds } 519863551ae0SDavid Gibson 5199d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 5200d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 520105e90bd0SPeter Xu unsigned long end, struct page *ref_page, 520205e90bd0SPeter Xu zap_flags_t zap_flags) 5203d833352aSMel Gorman { 5204131a79b4SMike Kravetz hugetlb_vma_lock_write(vma); 5205131a79b4SMike Kravetz i_mmap_lock_write(vma->vm_file->f_mapping); 5206131a79b4SMike Kravetz 520705e90bd0SPeter Xu __unmap_hugepage_range(tlb, vma, start, end, ref_page, zap_flags); 5208d833352aSMel Gorman 520904ada095SMike Kravetz if (zap_flags & ZAP_FLAG_UNMAP) { /* final unmap */ 5210d833352aSMel Gorman /* 521104ada095SMike Kravetz * Unlock and free the vma lock before releasing i_mmap_rwsem. 521204ada095SMike Kravetz * When the vma_lock is freed, this makes the vma ineligible 521304ada095SMike Kravetz * for pmd sharing. And, i_mmap_rwsem is required to set up 521404ada095SMike Kravetz * pmd sharing. This is important as page tables for this 521504ada095SMike Kravetz * unmapped range will be asynchrously deleted. If the page 521604ada095SMike Kravetz * tables are shared, there will be issues when accessed by 521704ada095SMike Kravetz * someone else. 5218d833352aSMel Gorman */ 5219ecfbd733SMike Kravetz __hugetlb_vma_unlock_write_free(vma); 5220131a79b4SMike Kravetz i_mmap_unlock_write(vma->vm_file->f_mapping); 522104ada095SMike Kravetz } else { 522204ada095SMike Kravetz i_mmap_unlock_write(vma->vm_file->f_mapping); 522304ada095SMike Kravetz hugetlb_vma_unlock_write(vma); 522404ada095SMike Kravetz } 5225d833352aSMel Gorman } 5226d833352aSMel Gorman 5227502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 522805e90bd0SPeter Xu unsigned long end, struct page *ref_page, 522905e90bd0SPeter Xu zap_flags_t zap_flags) 5230502717f4SChen, Kenneth W { 523124669e58SAneesh Kumar K.V struct mmu_gather tlb; 5232dff11abeSMike Kravetz 5233a72afd87SWill Deacon tlb_gather_mmu(&tlb, vma->vm_mm); 523405e90bd0SPeter Xu __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags); 5235ae8eba8bSWill Deacon tlb_finish_mmu(&tlb); 5236502717f4SChen, Kenneth W } 5237502717f4SChen, Kenneth W 523804f2cbe3SMel Gorman /* 523904f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 5240578b7725SZhiyuan Dai * mapping it owns the reserve page for. The intention is to unmap the page 524104f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 524204f2cbe3SMel Gorman * same region. 524304f2cbe3SMel Gorman */ 52442f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 52452a4b3dedSHarvey Harrison struct page *page, unsigned long address) 524604f2cbe3SMel Gorman { 52477526674dSAdam Litke struct hstate *h = hstate_vma(vma); 524804f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 524904f2cbe3SMel Gorman struct address_space *mapping; 525004f2cbe3SMel Gorman pgoff_t pgoff; 525104f2cbe3SMel Gorman 525204f2cbe3SMel Gorman /* 525304f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 525404f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 525504f2cbe3SMel Gorman */ 52567526674dSAdam Litke address = address & huge_page_mask(h); 525736e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 525836e4f20aSMichal Hocko vma->vm_pgoff; 525993c76a3dSAl Viro mapping = vma->vm_file->f_mapping; 526004f2cbe3SMel Gorman 52614eb2b1dcSMel Gorman /* 52624eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 52634eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 52644eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 52654eb2b1dcSMel Gorman */ 526683cde9e8SDavidlohr Bueso i_mmap_lock_write(mapping); 52676b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 526804f2cbe3SMel Gorman /* Do not unmap the current VMA */ 526904f2cbe3SMel Gorman if (iter_vma == vma) 527004f2cbe3SMel Gorman continue; 527104f2cbe3SMel Gorman 527204f2cbe3SMel Gorman /* 52732f84a899SMel Gorman * Shared VMAs have their own reserves and do not affect 52742f84a899SMel Gorman * MAP_PRIVATE accounting but it is possible that a shared 52752f84a899SMel Gorman * VMA is using the same page so check and skip such VMAs. 52762f84a899SMel Gorman */ 52772f84a899SMel Gorman if (iter_vma->vm_flags & VM_MAYSHARE) 52782f84a899SMel Gorman continue; 52792f84a899SMel Gorman 52802f84a899SMel Gorman /* 528104f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 528204f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 528304f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 528404f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 528504f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 528604f2cbe3SMel Gorman */ 528704f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 528824669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 528905e90bd0SPeter Xu address + huge_page_size(h), page, 0); 529004f2cbe3SMel Gorman } 529183cde9e8SDavidlohr Bueso i_mmap_unlock_write(mapping); 529204f2cbe3SMel Gorman } 529304f2cbe3SMel Gorman 52940fe6e20bSNaoya Horiguchi /* 5295c89357e2SDavid Hildenbrand * hugetlb_wp() should be called with page lock of the original hugepage held. 5296aa6d2e8cSBaolin Wang * Called with hugetlb_fault_mutex_table held and pte_page locked so we 5297ef009b25SMichal Hocko * cannot race with other handlers or page migration. 5298ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 52990fe6e20bSNaoya Horiguchi */ 5300c89357e2SDavid Hildenbrand static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma, 5301c89357e2SDavid Hildenbrand unsigned long address, pte_t *ptep, unsigned int flags, 5302cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 53031e8f889bSDavid Gibson { 5304c89357e2SDavid Hildenbrand const bool unshare = flags & FAULT_FLAG_UNSHARE; 53053999f52eSAneesh Kumar K.V pte_t pte; 5306a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 53071e8f889bSDavid Gibson struct page *old_page, *new_page; 53082b740303SSouptick Joarder int outside_reserve = 0; 53092b740303SSouptick Joarder vm_fault_t ret = 0; 5310974e6d66SHuang Ying unsigned long haddr = address & huge_page_mask(h); 5311ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 53121e8f889bSDavid Gibson 5313c89357e2SDavid Hildenbrand VM_BUG_ON(unshare && (flags & FOLL_WRITE)); 5314c89357e2SDavid Hildenbrand VM_BUG_ON(!unshare && !(flags & FOLL_WRITE)); 5315c89357e2SDavid Hildenbrand 53161d8d1464SDavid Hildenbrand /* 53171d8d1464SDavid Hildenbrand * hugetlb does not support FOLL_FORCE-style write faults that keep the 53181d8d1464SDavid Hildenbrand * PTE mapped R/O such as maybe_mkwrite() would do. 53191d8d1464SDavid Hildenbrand */ 53201d8d1464SDavid Hildenbrand if (WARN_ON_ONCE(!unshare && !(vma->vm_flags & VM_WRITE))) 53211d8d1464SDavid Hildenbrand return VM_FAULT_SIGSEGV; 53221d8d1464SDavid Hildenbrand 53231d8d1464SDavid Hildenbrand /* Let's take out MAP_SHARED mappings first. */ 53241d8d1464SDavid Hildenbrand if (vma->vm_flags & VM_MAYSHARE) { 53251d8d1464SDavid Hildenbrand if (unlikely(unshare)) 53261d8d1464SDavid Hildenbrand return 0; 53271d8d1464SDavid Hildenbrand set_huge_ptep_writable(vma, haddr, ptep); 53281d8d1464SDavid Hildenbrand return 0; 53291d8d1464SDavid Hildenbrand } 53301d8d1464SDavid Hildenbrand 53313999f52eSAneesh Kumar K.V pte = huge_ptep_get(ptep); 53321e8f889bSDavid Gibson old_page = pte_page(pte); 53331e8f889bSDavid Gibson 5334662ce1dcSYang Yang delayacct_wpcopy_start(); 5335662ce1dcSYang Yang 533604f2cbe3SMel Gorman retry_avoidcopy: 5337c89357e2SDavid Hildenbrand /* 5338c89357e2SDavid Hildenbrand * If no-one else is actually using this page, we're the exclusive 5339c89357e2SDavid Hildenbrand * owner and can reuse this page. 5340c89357e2SDavid Hildenbrand */ 534137a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 5342c89357e2SDavid Hildenbrand if (!PageAnonExclusive(old_page)) 53435a49973dSHugh Dickins page_move_anon_rmap(old_page, vma); 5344c89357e2SDavid Hildenbrand if (likely(!unshare)) 53455b7a1d40SHuang Ying set_huge_ptep_writable(vma, haddr, ptep); 5346662ce1dcSYang Yang 5347662ce1dcSYang Yang delayacct_wpcopy_end(); 534883c54070SNick Piggin return 0; 53491e8f889bSDavid Gibson } 53506c287605SDavid Hildenbrand VM_BUG_ON_PAGE(PageAnon(old_page) && PageAnonExclusive(old_page), 53516c287605SDavid Hildenbrand old_page); 53521e8f889bSDavid Gibson 535304f2cbe3SMel Gorman /* 535404f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 535504f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 535604f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 535704f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 535804f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 535904f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 536004f2cbe3SMel Gorman * of the full address range. 536104f2cbe3SMel Gorman */ 53625944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 536304f2cbe3SMel Gorman old_page != pagecache_page) 536404f2cbe3SMel Gorman outside_reserve = 1; 536504f2cbe3SMel Gorman 536609cbfeafSKirill A. Shutemov get_page(old_page); 5367b76c8cfbSLarry Woodman 5368ad4404a2SDavidlohr Bueso /* 5369ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 5370ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 5371ad4404a2SDavidlohr Bueso */ 5372cb900f41SKirill A. Shutemov spin_unlock(ptl); 53735b7a1d40SHuang Ying new_page = alloc_huge_page(vma, haddr, outside_reserve); 53741e8f889bSDavid Gibson 53752fc39cecSAdam Litke if (IS_ERR(new_page)) { 537604f2cbe3SMel Gorman /* 537704f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 537804f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 537904f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 538004f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 538104f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 538204f2cbe3SMel Gorman */ 538304f2cbe3SMel Gorman if (outside_reserve) { 538440549ba8SMike Kravetz struct address_space *mapping = vma->vm_file->f_mapping; 538540549ba8SMike Kravetz pgoff_t idx; 538640549ba8SMike Kravetz u32 hash; 538740549ba8SMike Kravetz 538809cbfeafSKirill A. Shutemov put_page(old_page); 538940549ba8SMike Kravetz /* 539040549ba8SMike Kravetz * Drop hugetlb_fault_mutex and vma_lock before 539140549ba8SMike Kravetz * unmapping. unmapping needs to hold vma_lock 539240549ba8SMike Kravetz * in write mode. Dropping vma_lock in read mode 539340549ba8SMike Kravetz * here is OK as COW mappings do not interact with 539440549ba8SMike Kravetz * PMD sharing. 539540549ba8SMike Kravetz * 539640549ba8SMike Kravetz * Reacquire both after unmap operation. 539740549ba8SMike Kravetz */ 539840549ba8SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 539940549ba8SMike Kravetz hash = hugetlb_fault_mutex_hash(mapping, idx); 540040549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 540140549ba8SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 540240549ba8SMike Kravetz 54035b7a1d40SHuang Ying unmap_ref_private(mm, vma, old_page, haddr); 540440549ba8SMike Kravetz 540540549ba8SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 540640549ba8SMike Kravetz hugetlb_vma_lock_read(vma); 5407cb900f41SKirill A. Shutemov spin_lock(ptl); 54085b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 5409a9af0c5dSNaoya Horiguchi if (likely(ptep && 5410a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 541104f2cbe3SMel Gorman goto retry_avoidcopy; 5412a734bcc8SHillf Danton /* 5413cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 5414cb900f41SKirill A. Shutemov * lock, and our job is done. 5415a734bcc8SHillf Danton */ 5416662ce1dcSYang Yang delayacct_wpcopy_end(); 5417a734bcc8SHillf Danton return 0; 541804f2cbe3SMel Gorman } 541904f2cbe3SMel Gorman 54202b740303SSouptick Joarder ret = vmf_error(PTR_ERR(new_page)); 5421ad4404a2SDavidlohr Bueso goto out_release_old; 54221e8f889bSDavid Gibson } 54231e8f889bSDavid Gibson 54240fe6e20bSNaoya Horiguchi /* 54250fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 54260fe6e20bSNaoya Horiguchi * anon_vma prepared. 54270fe6e20bSNaoya Horiguchi */ 542844e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 5429ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 5430ad4404a2SDavidlohr Bueso goto out_release_all; 543144e2aa93SDean Nelson } 54320fe6e20bSNaoya Horiguchi 5433974e6d66SHuang Ying copy_user_huge_page(new_page, old_page, address, vma, 543447ad8475SAndrea Arcangeli pages_per_huge_page(h)); 54350ed361deSNick Piggin __SetPageUptodate(new_page); 54361e8f889bSDavid Gibson 54377269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, 54386f4f13e8SJérôme Glisse haddr + huge_page_size(h)); 5439ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 5440ad4404a2SDavidlohr Bueso 5441b76c8cfbSLarry Woodman /* 5442cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 5443b76c8cfbSLarry Woodman * before the page tables are altered 5444b76c8cfbSLarry Woodman */ 5445cb900f41SKirill A. Shutemov spin_lock(ptl); 54465b7a1d40SHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 5447a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 5448c89357e2SDavid Hildenbrand /* Break COW or unshare */ 54495b7a1d40SHuang Ying huge_ptep_clear_flush(vma, haddr, ptep); 5450ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, range.end); 5451cea86fe2SHugh Dickins page_remove_rmap(old_page, vma, true); 54525b7a1d40SHuang Ying hugepage_add_new_anon_rmap(new_page, vma, haddr); 54531eba86c0SPasha Tatashin set_huge_pte_at(mm, haddr, ptep, 5454c89357e2SDavid Hildenbrand make_huge_pte(vma, new_page, !unshare)); 54558f251a3dSMike Kravetz SetHPageMigratable(new_page); 54561e8f889bSDavid Gibson /* Make the old page be freed below */ 54571e8f889bSDavid Gibson new_page = old_page; 54581e8f889bSDavid Gibson } 5459cb900f41SKirill A. Shutemov spin_unlock(ptl); 5460ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 5461ad4404a2SDavidlohr Bueso out_release_all: 5462c89357e2SDavid Hildenbrand /* 5463c89357e2SDavid Hildenbrand * No restore in case of successful pagetable update (Break COW or 5464c89357e2SDavid Hildenbrand * unshare) 5465c89357e2SDavid Hildenbrand */ 5466c7b1850dSMike Kravetz if (new_page != old_page) 54675b7a1d40SHuang Ying restore_reserve_on_error(h, vma, haddr, new_page); 546809cbfeafSKirill A. Shutemov put_page(new_page); 5469ad4404a2SDavidlohr Bueso out_release_old: 547009cbfeafSKirill A. Shutemov put_page(old_page); 54718312034fSJoonsoo Kim 5472ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 5473662ce1dcSYang Yang 5474662ce1dcSYang Yang delayacct_wpcopy_end(); 5475ad4404a2SDavidlohr Bueso return ret; 54761e8f889bSDavid Gibson } 54771e8f889bSDavid Gibson 54783ae77f43SHugh Dickins /* 54793ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 54803ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 54813ae77f43SHugh Dickins */ 54823ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 54832a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 54842a15efc9SHugh Dickins { 54852a15efc9SHugh Dickins struct address_space *mapping; 54862a15efc9SHugh Dickins pgoff_t idx; 54872a15efc9SHugh Dickins struct page *page; 54882a15efc9SHugh Dickins 54892a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 54902a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 54912a15efc9SHugh Dickins 54922a15efc9SHugh Dickins page = find_get_page(mapping, idx); 54932a15efc9SHugh Dickins if (page) 54942a15efc9SHugh Dickins put_page(page); 54952a15efc9SHugh Dickins return page != NULL; 54962a15efc9SHugh Dickins } 54972a15efc9SHugh Dickins 54987e1813d4SMike Kravetz int hugetlb_add_to_page_cache(struct page *page, struct address_space *mapping, 5499ab76ad54SMike Kravetz pgoff_t idx) 5500ab76ad54SMike Kravetz { 5501d9ef44deSMatthew Wilcox (Oracle) struct folio *folio = page_folio(page); 5502ab76ad54SMike Kravetz struct inode *inode = mapping->host; 5503ab76ad54SMike Kravetz struct hstate *h = hstate_inode(inode); 5504d9ef44deSMatthew Wilcox (Oracle) int err; 5505ab76ad54SMike Kravetz 5506d9ef44deSMatthew Wilcox (Oracle) __folio_set_locked(folio); 5507d9ef44deSMatthew Wilcox (Oracle) err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL); 5508d9ef44deSMatthew Wilcox (Oracle) 5509d9ef44deSMatthew Wilcox (Oracle) if (unlikely(err)) { 5510d9ef44deSMatthew Wilcox (Oracle) __folio_clear_locked(folio); 5511ab76ad54SMike Kravetz return err; 5512d9ef44deSMatthew Wilcox (Oracle) } 5513d6995da3SMike Kravetz ClearHPageRestoreReserve(page); 5514ab76ad54SMike Kravetz 551522146c3cSMike Kravetz /* 5516d9ef44deSMatthew Wilcox (Oracle) * mark folio dirty so that it will not be removed from cache/file 551722146c3cSMike Kravetz * by non-hugetlbfs specific code paths. 551822146c3cSMike Kravetz */ 5519d9ef44deSMatthew Wilcox (Oracle) folio_mark_dirty(folio); 552022146c3cSMike Kravetz 5521ab76ad54SMike Kravetz spin_lock(&inode->i_lock); 5522ab76ad54SMike Kravetz inode->i_blocks += blocks_per_huge_page(h); 5523ab76ad54SMike Kravetz spin_unlock(&inode->i_lock); 5524ab76ad54SMike Kravetz return 0; 5525ab76ad54SMike Kravetz } 5526ab76ad54SMike Kravetz 55277677f7fdSAxel Rasmussen static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma, 55287677f7fdSAxel Rasmussen struct address_space *mapping, 55297677f7fdSAxel Rasmussen pgoff_t idx, 55307677f7fdSAxel Rasmussen unsigned int flags, 55317677f7fdSAxel Rasmussen unsigned long haddr, 5532824ddc60SNadav Amit unsigned long addr, 55337677f7fdSAxel Rasmussen unsigned long reason) 55347677f7fdSAxel Rasmussen { 55357677f7fdSAxel Rasmussen u32 hash; 55367677f7fdSAxel Rasmussen struct vm_fault vmf = { 55377677f7fdSAxel Rasmussen .vma = vma, 55387677f7fdSAxel Rasmussen .address = haddr, 5539824ddc60SNadav Amit .real_address = addr, 55407677f7fdSAxel Rasmussen .flags = flags, 55417677f7fdSAxel Rasmussen 55427677f7fdSAxel Rasmussen /* 55437677f7fdSAxel Rasmussen * Hard to debug if it ends up being 55447677f7fdSAxel Rasmussen * used by a callee that assumes 55457677f7fdSAxel Rasmussen * something about the other 55467677f7fdSAxel Rasmussen * uninitialized fields... same as in 55477677f7fdSAxel Rasmussen * memory.c 55487677f7fdSAxel Rasmussen */ 55497677f7fdSAxel Rasmussen }; 55507677f7fdSAxel Rasmussen 55517677f7fdSAxel Rasmussen /* 5552958f32ceSLiu Shixin * vma_lock and hugetlb_fault_mutex must be dropped before handling 5553958f32ceSLiu Shixin * userfault. Also mmap_lock could be dropped due to handling 5554958f32ceSLiu Shixin * userfault, any vma operation should be careful from here. 55557677f7fdSAxel Rasmussen */ 555640549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 55577677f7fdSAxel Rasmussen hash = hugetlb_fault_mutex_hash(mapping, idx); 55587677f7fdSAxel Rasmussen mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5559958f32ceSLiu Shixin return handle_userfault(&vmf, reason); 55607677f7fdSAxel Rasmussen } 55617677f7fdSAxel Rasmussen 55622ea7ff1eSPeter Xu /* 55632ea7ff1eSPeter Xu * Recheck pte with pgtable lock. Returns true if pte didn't change, or 55642ea7ff1eSPeter Xu * false if pte changed or is changing. 55652ea7ff1eSPeter Xu */ 55662ea7ff1eSPeter Xu static bool hugetlb_pte_stable(struct hstate *h, struct mm_struct *mm, 55672ea7ff1eSPeter Xu pte_t *ptep, pte_t old_pte) 55682ea7ff1eSPeter Xu { 55692ea7ff1eSPeter Xu spinlock_t *ptl; 55702ea7ff1eSPeter Xu bool same; 55712ea7ff1eSPeter Xu 55722ea7ff1eSPeter Xu ptl = huge_pte_lock(h, mm, ptep); 55732ea7ff1eSPeter Xu same = pte_same(huge_ptep_get(ptep), old_pte); 55742ea7ff1eSPeter Xu spin_unlock(ptl); 55752ea7ff1eSPeter Xu 55762ea7ff1eSPeter Xu return same; 55772ea7ff1eSPeter Xu } 55782ea7ff1eSPeter Xu 55792b740303SSouptick Joarder static vm_fault_t hugetlb_no_page(struct mm_struct *mm, 55802b740303SSouptick Joarder struct vm_area_struct *vma, 55818382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 5582c64e912cSPeter Xu unsigned long address, pte_t *ptep, 5583c64e912cSPeter Xu pte_t old_pte, unsigned int flags) 5584ac9b9c66SHugh Dickins { 5585a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 55862b740303SSouptick Joarder vm_fault_t ret = VM_FAULT_SIGBUS; 5587409eb8c2SHillf Danton int anon_rmap = 0; 55884c887265SAdam Litke unsigned long size; 55894c887265SAdam Litke struct page *page; 55901e8f889bSDavid Gibson pte_t new_pte; 5591cb900f41SKirill A. Shutemov spinlock_t *ptl; 5592285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 5593c7b1850dSMike Kravetz bool new_page, new_pagecache_page = false; 5594958f32ceSLiu Shixin u32 hash = hugetlb_fault_mutex_hash(mapping, idx); 55954c887265SAdam Litke 559604f2cbe3SMel Gorman /* 559704f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 559804f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 5599c89357e2SDavid Hildenbrand * COW/unsharing. Warn that such a situation has occurred as it may not 5600c89357e2SDavid Hildenbrand * be obvious. 560104f2cbe3SMel Gorman */ 560204f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 5603910154d5SGeoffrey Thomas pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 560404f2cbe3SMel Gorman current->pid); 5605958f32ceSLiu Shixin goto out; 560604f2cbe3SMel Gorman } 560704f2cbe3SMel Gorman 56084c887265SAdam Litke /* 5609188a3972SMike Kravetz * Use page lock to guard against racing truncation 5610188a3972SMike Kravetz * before we get page_table_lock. 56114c887265SAdam Litke */ 5612c7b1850dSMike Kravetz new_page = false; 561387bf91d3SMike Kravetz page = find_lock_page(mapping, idx); 561487bf91d3SMike Kravetz if (!page) { 5615188a3972SMike Kravetz size = i_size_read(mapping->host) >> huge_page_shift(h); 5616188a3972SMike Kravetz if (idx >= size) 5617188a3972SMike Kravetz goto out; 56187677f7fdSAxel Rasmussen /* Check for page in userfault range */ 56192ea7ff1eSPeter Xu if (userfaultfd_missing(vma)) { 56202ea7ff1eSPeter Xu /* 56212ea7ff1eSPeter Xu * Since hugetlb_no_page() was examining pte 56222ea7ff1eSPeter Xu * without pgtable lock, we need to re-test under 56232ea7ff1eSPeter Xu * lock because the pte may not be stable and could 56242ea7ff1eSPeter Xu * have changed from under us. Try to detect 56252ea7ff1eSPeter Xu * either changed or during-changing ptes and retry 56262ea7ff1eSPeter Xu * properly when needed. 56272ea7ff1eSPeter Xu * 56282ea7ff1eSPeter Xu * Note that userfaultfd is actually fine with 56292ea7ff1eSPeter Xu * false positives (e.g. caused by pte changed), 56302ea7ff1eSPeter Xu * but not wrong logical events (e.g. caused by 56312ea7ff1eSPeter Xu * reading a pte during changing). The latter can 56322ea7ff1eSPeter Xu * confuse the userspace, so the strictness is very 56332ea7ff1eSPeter Xu * much preferred. E.g., MISSING event should 56342ea7ff1eSPeter Xu * never happen on the page after UFFDIO_COPY has 56352ea7ff1eSPeter Xu * correctly installed the page and returned. 56362ea7ff1eSPeter Xu */ 56372ea7ff1eSPeter Xu if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) { 56382ea7ff1eSPeter Xu ret = 0; 56392ea7ff1eSPeter Xu goto out; 56402ea7ff1eSPeter Xu } 56412ea7ff1eSPeter Xu 56422ea7ff1eSPeter Xu return hugetlb_handle_userfault(vma, mapping, idx, flags, 56432ea7ff1eSPeter Xu haddr, address, 56447677f7fdSAxel Rasmussen VM_UFFD_MISSING); 56452ea7ff1eSPeter Xu } 56461a1aad8aSMike Kravetz 5647285b8dcaSHuang Ying page = alloc_huge_page(vma, haddr, 0); 56482fc39cecSAdam Litke if (IS_ERR(page)) { 56494643d67eSMike Kravetz /* 56504643d67eSMike Kravetz * Returning error will result in faulting task being 56514643d67eSMike Kravetz * sent SIGBUS. The hugetlb fault mutex prevents two 56524643d67eSMike Kravetz * tasks from racing to fault in the same page which 56534643d67eSMike Kravetz * could result in false unable to allocate errors. 56544643d67eSMike Kravetz * Page migration does not take the fault mutex, but 56554643d67eSMike Kravetz * does a clear then write of pte's under page table 56564643d67eSMike Kravetz * lock. Page fault code could race with migration, 56574643d67eSMike Kravetz * notice the clear pte and try to allocate a page 56584643d67eSMike Kravetz * here. Before returning error, get ptl and make 56594643d67eSMike Kravetz * sure there really is no pte entry. 56604643d67eSMike Kravetz */ 5661f9bf6c03SPeter Xu if (hugetlb_pte_stable(h, mm, ptep, old_pte)) 56622b740303SSouptick Joarder ret = vmf_error(PTR_ERR(page)); 5663f9bf6c03SPeter Xu else 5664f9bf6c03SPeter Xu ret = 0; 56656bda666aSChristoph Lameter goto out; 56666bda666aSChristoph Lameter } 566747ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 56680ed361deSNick Piggin __SetPageUptodate(page); 5669cb6acd01SMike Kravetz new_page = true; 5670ac9b9c66SHugh Dickins 5671f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 56727e1813d4SMike Kravetz int err = hugetlb_add_to_page_cache(page, mapping, idx); 56736bda666aSChristoph Lameter if (err) { 56743a5497a2SMiaohe Lin /* 56753a5497a2SMiaohe Lin * err can't be -EEXIST which implies someone 56763a5497a2SMiaohe Lin * else consumed the reservation since hugetlb 56773a5497a2SMiaohe Lin * fault mutex is held when add a hugetlb page 56783a5497a2SMiaohe Lin * to the page cache. So it's safe to call 56793a5497a2SMiaohe Lin * restore_reserve_on_error() here. 56803a5497a2SMiaohe Lin */ 56813a5497a2SMiaohe Lin restore_reserve_on_error(h, vma, haddr, page); 56826bda666aSChristoph Lameter put_page(page); 56836bda666aSChristoph Lameter goto out; 56846bda666aSChristoph Lameter } 5685c7b1850dSMike Kravetz new_pagecache_page = true; 568623be7468SMel Gorman } else { 56876bda666aSChristoph Lameter lock_page(page); 56880fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 56890fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 56900fe6e20bSNaoya Horiguchi goto backout_unlocked; 569123be7468SMel Gorman } 5692409eb8c2SHillf Danton anon_rmap = 1; 56930fe6e20bSNaoya Horiguchi } 56940fe6e20bSNaoya Horiguchi } else { 569557303d80SAndy Whitcroft /* 5696998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 5697998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 5698998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 5699fd6a03edSNaoya Horiguchi */ 5700fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 57010eb98f15SMiaohe Lin ret = VM_FAULT_HWPOISON_LARGE | 5702972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 5703fd6a03edSNaoya Horiguchi goto backout_unlocked; 57046bda666aSChristoph Lameter } 57057677f7fdSAxel Rasmussen 57067677f7fdSAxel Rasmussen /* Check for page in userfault range. */ 57077677f7fdSAxel Rasmussen if (userfaultfd_minor(vma)) { 57087677f7fdSAxel Rasmussen unlock_page(page); 57097677f7fdSAxel Rasmussen put_page(page); 57102ea7ff1eSPeter Xu /* See comment in userfaultfd_missing() block above */ 57112ea7ff1eSPeter Xu if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) { 57122ea7ff1eSPeter Xu ret = 0; 57132ea7ff1eSPeter Xu goto out; 57142ea7ff1eSPeter Xu } 57152ea7ff1eSPeter Xu return hugetlb_handle_userfault(vma, mapping, idx, flags, 57162ea7ff1eSPeter Xu haddr, address, 57177677f7fdSAxel Rasmussen VM_UFFD_MINOR); 57187677f7fdSAxel Rasmussen } 5719998b4382SNaoya Horiguchi } 57201e8f889bSDavid Gibson 572157303d80SAndy Whitcroft /* 572257303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 572357303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 572457303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 572557303d80SAndy Whitcroft * the spinlock. 572657303d80SAndy Whitcroft */ 57275e911373SMike Kravetz if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 5728285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 57292b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 57302b26736cSAndy Whitcroft goto backout_unlocked; 57312b26736cSAndy Whitcroft } 57325e911373SMike Kravetz /* Just decrements count, does not deallocate */ 5733285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 57345e911373SMike Kravetz } 573557303d80SAndy Whitcroft 57368bea8052SAneesh Kumar K.V ptl = huge_pte_lock(h, mm, ptep); 573783c54070SNick Piggin ret = 0; 5738c64e912cSPeter Xu /* If pte changed from under us, retry */ 5739c64e912cSPeter Xu if (!pte_same(huge_ptep_get(ptep), old_pte)) 57404c887265SAdam Litke goto backout; 57414c887265SAdam Litke 57424781593dSPeter Xu if (anon_rmap) 5743285b8dcaSHuang Ying hugepage_add_new_anon_rmap(page, vma, haddr); 57444781593dSPeter Xu else 5745fb3d824dSDavid Hildenbrand page_dup_file_rmap(page, true); 57461e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 57471e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 5748c64e912cSPeter Xu /* 5749c64e912cSPeter Xu * If this pte was previously wr-protected, keep it wr-protected even 5750c64e912cSPeter Xu * if populated. 5751c64e912cSPeter Xu */ 5752c64e912cSPeter Xu if (unlikely(pte_marker_uffd_wp(old_pte))) 5753c64e912cSPeter Xu new_pte = huge_pte_wrprotect(huge_pte_mkuffd_wp(new_pte)); 5754285b8dcaSHuang Ying set_huge_pte_at(mm, haddr, ptep, new_pte); 57551e8f889bSDavid Gibson 57565d317b2bSNaoya Horiguchi hugetlb_count_add(pages_per_huge_page(h), mm); 5757788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 57581e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 5759c89357e2SDavid Hildenbrand ret = hugetlb_wp(mm, vma, address, ptep, flags, page, ptl); 57601e8f889bSDavid Gibson } 57611e8f889bSDavid Gibson 5762cb900f41SKirill A. Shutemov spin_unlock(ptl); 5763cb6acd01SMike Kravetz 5764cb6acd01SMike Kravetz /* 57658f251a3dSMike Kravetz * Only set HPageMigratable in newly allocated pages. Existing pages 57668f251a3dSMike Kravetz * found in the pagecache may not have HPageMigratableset if they have 57678f251a3dSMike Kravetz * been isolated for migration. 5768cb6acd01SMike Kravetz */ 5769cb6acd01SMike Kravetz if (new_page) 57708f251a3dSMike Kravetz SetHPageMigratable(page); 5771cb6acd01SMike Kravetz 57724c887265SAdam Litke unlock_page(page); 57734c887265SAdam Litke out: 5774958f32ceSLiu Shixin hugetlb_vma_unlock_read(vma); 5775958f32ceSLiu Shixin mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5776ac9b9c66SHugh Dickins return ret; 57774c887265SAdam Litke 57784c887265SAdam Litke backout: 5779cb900f41SKirill A. Shutemov spin_unlock(ptl); 57802b26736cSAndy Whitcroft backout_unlocked: 5781c7b1850dSMike Kravetz if (new_page && !new_pagecache_page) 5782285b8dcaSHuang Ying restore_reserve_on_error(h, vma, haddr, page); 5783fa27759aSMike Kravetz 5784fa27759aSMike Kravetz unlock_page(page); 57854c887265SAdam Litke put_page(page); 57864c887265SAdam Litke goto out; 5787ac9b9c66SHugh Dickins } 5788ac9b9c66SHugh Dickins 57898382d914SDavidlohr Bueso #ifdef CONFIG_SMP 5790188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 57918382d914SDavidlohr Bueso { 57928382d914SDavidlohr Bueso unsigned long key[2]; 57938382d914SDavidlohr Bueso u32 hash; 57948382d914SDavidlohr Bueso 57958382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 57968382d914SDavidlohr Bueso key[1] = idx; 57978382d914SDavidlohr Bueso 579855254636SMike Kravetz hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); 57998382d914SDavidlohr Bueso 58008382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 58018382d914SDavidlohr Bueso } 58028382d914SDavidlohr Bueso #else 58038382d914SDavidlohr Bueso /* 58046c26d310SMiaohe Lin * For uniprocessor systems we always use a single mutex, so just 58058382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 58068382d914SDavidlohr Bueso */ 5807188b04a7SWei Yang u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 58088382d914SDavidlohr Bueso { 58098382d914SDavidlohr Bueso return 0; 58108382d914SDavidlohr Bueso } 58118382d914SDavidlohr Bueso #endif 58128382d914SDavidlohr Bueso 58132b740303SSouptick Joarder vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 5814788c7df4SHugh Dickins unsigned long address, unsigned int flags) 581586e5216fSAdam Litke { 58168382d914SDavidlohr Bueso pte_t *ptep, entry; 5817cb900f41SKirill A. Shutemov spinlock_t *ptl; 58182b740303SSouptick Joarder vm_fault_t ret; 58198382d914SDavidlohr Bueso u32 hash; 58208382d914SDavidlohr Bueso pgoff_t idx; 58210fe6e20bSNaoya Horiguchi struct page *page = NULL; 582257303d80SAndy Whitcroft struct page *pagecache_page = NULL; 5823a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 58248382d914SDavidlohr Bueso struct address_space *mapping; 58250f792cf9SNaoya Horiguchi int need_wait_lock = 0; 5826285b8dcaSHuang Ying unsigned long haddr = address & huge_page_mask(h); 582786e5216fSAdam Litke 5828285b8dcaSHuang Ying ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); 5829fd6a03edSNaoya Horiguchi if (ptep) { 583040549ba8SMike Kravetz /* 583140549ba8SMike Kravetz * Since we hold no locks, ptep could be stale. That is 583240549ba8SMike Kravetz * OK as we are only making decisions based on content and 583340549ba8SMike Kravetz * not actually modifying content here. 583440549ba8SMike Kravetz */ 5835fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 5836290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 5837ad1ac596SMiaohe Lin migration_entry_wait_huge(vma, ptep); 5838290408d4SNaoya Horiguchi return 0; 5839290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 5840aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 5841972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 5842c0d0381aSMike Kravetz } 5843ddeaab32SMike Kravetz 58443935baa9SDavid Gibson /* 58453935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 58463935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 58473935baa9SDavid Gibson * the same page in the page cache. 58483935baa9SDavid Gibson */ 584940549ba8SMike Kravetz mapping = vma->vm_file->f_mapping; 585040549ba8SMike Kravetz idx = vma_hugecache_offset(h, vma, haddr); 5851188b04a7SWei Yang hash = hugetlb_fault_mutex_hash(mapping, idx); 5852c672c7f2SMike Kravetz mutex_lock(&hugetlb_fault_mutex_table[hash]); 58538382d914SDavidlohr Bueso 585440549ba8SMike Kravetz /* 585540549ba8SMike Kravetz * Acquire vma lock before calling huge_pte_alloc and hold 585640549ba8SMike Kravetz * until finished with ptep. This prevents huge_pmd_unshare from 585740549ba8SMike Kravetz * being called elsewhere and making the ptep no longer valid. 585840549ba8SMike Kravetz * 585940549ba8SMike Kravetz * ptep could have already be assigned via huge_pte_offset. That 586040549ba8SMike Kravetz * is OK, as huge_pte_alloc will return the same value unless 586140549ba8SMike Kravetz * something has changed. 586240549ba8SMike Kravetz */ 586340549ba8SMike Kravetz hugetlb_vma_lock_read(vma); 586440549ba8SMike Kravetz ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); 586540549ba8SMike Kravetz if (!ptep) { 586640549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 586740549ba8SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 586840549ba8SMike Kravetz return VM_FAULT_OOM; 586940549ba8SMike Kravetz } 587040549ba8SMike Kravetz 58717f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 5872c64e912cSPeter Xu /* PTE markers should be handled the same way as none pte */ 5873958f32ceSLiu Shixin if (huge_pte_none_mostly(entry)) 5874958f32ceSLiu Shixin /* 5875958f32ceSLiu Shixin * hugetlb_no_page will drop vma lock and hugetlb fault 5876958f32ceSLiu Shixin * mutex internally, which make us return immediately. 5877958f32ceSLiu Shixin */ 5878958f32ceSLiu Shixin return hugetlb_no_page(mm, vma, mapping, idx, address, ptep, 5879c64e912cSPeter Xu entry, flags); 588086e5216fSAdam Litke 588183c54070SNick Piggin ret = 0; 58821e8f889bSDavid Gibson 588357303d80SAndy Whitcroft /* 58840f792cf9SNaoya Horiguchi * entry could be a migration/hwpoison entry at this point, so this 58850f792cf9SNaoya Horiguchi * check prevents the kernel from going below assuming that we have 58867c8de358SEthon Paul * an active hugepage in pagecache. This goto expects the 2nd page 58877c8de358SEthon Paul * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will 58887c8de358SEthon Paul * properly handle it. 58890f792cf9SNaoya Horiguchi */ 58900f792cf9SNaoya Horiguchi if (!pte_present(entry)) 58910f792cf9SNaoya Horiguchi goto out_mutex; 58920f792cf9SNaoya Horiguchi 58930f792cf9SNaoya Horiguchi /* 5894c89357e2SDavid Hildenbrand * If we are going to COW/unshare the mapping later, we examine the 5895c89357e2SDavid Hildenbrand * pending reservations for this page now. This will ensure that any 589657303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 58971d8d1464SDavid Hildenbrand * spinlock. Also lookup the pagecache page now as it is used to 58981d8d1464SDavid Hildenbrand * determine if a reservation has been consumed. 589957303d80SAndy Whitcroft */ 5900c89357e2SDavid Hildenbrand if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) && 59011d8d1464SDavid Hildenbrand !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(entry)) { 5902285b8dcaSHuang Ying if (vma_needs_reservation(h, vma, haddr) < 0) { 59032b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 5904b4d1d99fSDavid Gibson goto out_mutex; 59052b26736cSAndy Whitcroft } 59065e911373SMike Kravetz /* Just decrements count, does not deallocate */ 5907285b8dcaSHuang Ying vma_end_reservation(h, vma, haddr); 590857303d80SAndy Whitcroft 590929be8426SMiaohe Lin pagecache_page = find_lock_page(mapping, idx); 591057303d80SAndy Whitcroft } 591157303d80SAndy Whitcroft 59120f792cf9SNaoya Horiguchi ptl = huge_pte_lock(h, mm, ptep); 59130fe6e20bSNaoya Horiguchi 5914c89357e2SDavid Hildenbrand /* Check for a racing update before calling hugetlb_wp() */ 5915b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 5916cb900f41SKirill A. Shutemov goto out_ptl; 5917b4d1d99fSDavid Gibson 5918166f3eccSPeter Xu /* Handle userfault-wp first, before trying to lock more pages */ 5919166f3eccSPeter Xu if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) && 5920166f3eccSPeter Xu (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 5921166f3eccSPeter Xu struct vm_fault vmf = { 5922166f3eccSPeter Xu .vma = vma, 5923166f3eccSPeter Xu .address = haddr, 5924166f3eccSPeter Xu .real_address = address, 5925166f3eccSPeter Xu .flags = flags, 5926166f3eccSPeter Xu }; 5927166f3eccSPeter Xu 5928166f3eccSPeter Xu spin_unlock(ptl); 5929166f3eccSPeter Xu if (pagecache_page) { 5930166f3eccSPeter Xu unlock_page(pagecache_page); 5931166f3eccSPeter Xu put_page(pagecache_page); 5932166f3eccSPeter Xu } 593340549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 5934166f3eccSPeter Xu mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5935166f3eccSPeter Xu return handle_userfault(&vmf, VM_UFFD_WP); 5936166f3eccSPeter Xu } 5937166f3eccSPeter Xu 59380f792cf9SNaoya Horiguchi /* 5939c89357e2SDavid Hildenbrand * hugetlb_wp() requires page locks of pte_page(entry) and 59400f792cf9SNaoya Horiguchi * pagecache_page, so here we need take the former one 59410f792cf9SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 59420f792cf9SNaoya Horiguchi */ 59430f792cf9SNaoya Horiguchi page = pte_page(entry); 59440f792cf9SNaoya Horiguchi if (page != pagecache_page) 59450f792cf9SNaoya Horiguchi if (!trylock_page(page)) { 59460f792cf9SNaoya Horiguchi need_wait_lock = 1; 59470f792cf9SNaoya Horiguchi goto out_ptl; 59480f792cf9SNaoya Horiguchi } 59490f792cf9SNaoya Horiguchi 59500f792cf9SNaoya Horiguchi get_page(page); 5951b4d1d99fSDavid Gibson 5952c89357e2SDavid Hildenbrand if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) { 5953106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 5954c89357e2SDavid Hildenbrand ret = hugetlb_wp(mm, vma, address, ptep, flags, 5955cb900f41SKirill A. Shutemov pagecache_page, ptl); 59560f792cf9SNaoya Horiguchi goto out_put_page; 5957c89357e2SDavid Hildenbrand } else if (likely(flags & FAULT_FLAG_WRITE)) { 5958106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 5959b4d1d99fSDavid Gibson } 5960c89357e2SDavid Hildenbrand } 5961b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 5962285b8dcaSHuang Ying if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, 5963788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 5964285b8dcaSHuang Ying update_mmu_cache(vma, haddr, ptep); 59650f792cf9SNaoya Horiguchi out_put_page: 59660f792cf9SNaoya Horiguchi if (page != pagecache_page) 59670f792cf9SNaoya Horiguchi unlock_page(page); 59680f792cf9SNaoya Horiguchi put_page(page); 5969cb900f41SKirill A. Shutemov out_ptl: 5970cb900f41SKirill A. Shutemov spin_unlock(ptl); 597157303d80SAndy Whitcroft 597257303d80SAndy Whitcroft if (pagecache_page) { 597357303d80SAndy Whitcroft unlock_page(pagecache_page); 597457303d80SAndy Whitcroft put_page(pagecache_page); 597557303d80SAndy Whitcroft } 5976b4d1d99fSDavid Gibson out_mutex: 597740549ba8SMike Kravetz hugetlb_vma_unlock_read(vma); 5978c672c7f2SMike Kravetz mutex_unlock(&hugetlb_fault_mutex_table[hash]); 59790f792cf9SNaoya Horiguchi /* 59800f792cf9SNaoya Horiguchi * Generally it's safe to hold refcount during waiting page lock. But 59810f792cf9SNaoya Horiguchi * here we just wait to defer the next page fault to avoid busy loop and 59820f792cf9SNaoya Horiguchi * the page is not used after unlocked before returning from the current 59830f792cf9SNaoya Horiguchi * page fault. So we are safe from accessing freed page, even if we wait 59840f792cf9SNaoya Horiguchi * here without taking refcount. 59850f792cf9SNaoya Horiguchi */ 59860f792cf9SNaoya Horiguchi if (need_wait_lock) 59870f792cf9SNaoya Horiguchi wait_on_page_locked(page); 59881e8f889bSDavid Gibson return ret; 598986e5216fSAdam Litke } 599086e5216fSAdam Litke 5991714c1891SAxel Rasmussen #ifdef CONFIG_USERFAULTFD 59928fb5debcSMike Kravetz /* 59938fb5debcSMike Kravetz * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with 59948fb5debcSMike Kravetz * modifications for huge pages. 59958fb5debcSMike Kravetz */ 59968fb5debcSMike Kravetz int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, 59978fb5debcSMike Kravetz pte_t *dst_pte, 59988fb5debcSMike Kravetz struct vm_area_struct *dst_vma, 59998fb5debcSMike Kravetz unsigned long dst_addr, 60008fb5debcSMike Kravetz unsigned long src_addr, 6001f6191471SAxel Rasmussen enum mcopy_atomic_mode mode, 60026041c691SPeter Xu struct page **pagep, 60036041c691SPeter Xu bool wp_copy) 60048fb5debcSMike Kravetz { 6005f6191471SAxel Rasmussen bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE); 60068cc5fcbbSMina Almasry struct hstate *h = hstate_vma(dst_vma); 60078cc5fcbbSMina Almasry struct address_space *mapping = dst_vma->vm_file->f_mapping; 60088cc5fcbbSMina Almasry pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr); 60091e392147SAndrea Arcangeli unsigned long size; 60101c9e8defSMike Kravetz int vm_shared = dst_vma->vm_flags & VM_SHARED; 60118fb5debcSMike Kravetz pte_t _dst_pte; 60128fb5debcSMike Kravetz spinlock_t *ptl; 60138cc5fcbbSMina Almasry int ret = -ENOMEM; 60148fb5debcSMike Kravetz struct page *page; 6015f6191471SAxel Rasmussen int writable; 6016cc30042dSMina Almasry bool page_in_pagecache = false; 60178fb5debcSMike Kravetz 6018f6191471SAxel Rasmussen if (is_continue) { 6019f6191471SAxel Rasmussen ret = -EFAULT; 6020f6191471SAxel Rasmussen page = find_lock_page(mapping, idx); 6021f6191471SAxel Rasmussen if (!page) 6022f6191471SAxel Rasmussen goto out; 6023cc30042dSMina Almasry page_in_pagecache = true; 6024f6191471SAxel Rasmussen } else if (!*pagep) { 6025d84cf06eSMina Almasry /* If a page already exists, then it's UFFDIO_COPY for 6026d84cf06eSMina Almasry * a non-missing case. Return -EEXIST. 6027d84cf06eSMina Almasry */ 6028d84cf06eSMina Almasry if (vm_shared && 6029d84cf06eSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 6030d84cf06eSMina Almasry ret = -EEXIST; 60318fb5debcSMike Kravetz goto out; 6032d84cf06eSMina Almasry } 6033d84cf06eSMina Almasry 6034d84cf06eSMina Almasry page = alloc_huge_page(dst_vma, dst_addr, 0); 6035d84cf06eSMina Almasry if (IS_ERR(page)) { 6036d84cf06eSMina Almasry ret = -ENOMEM; 6037d84cf06eSMina Almasry goto out; 6038d84cf06eSMina Almasry } 60398fb5debcSMike Kravetz 60408fb5debcSMike Kravetz ret = copy_huge_page_from_user(page, 60418fb5debcSMike Kravetz (const void __user *) src_addr, 6042810a56b9SMike Kravetz pages_per_huge_page(h), false); 60438fb5debcSMike Kravetz 6044c1e8d7c6SMichel Lespinasse /* fallback to copy_from_user outside mmap_lock */ 60458fb5debcSMike Kravetz if (unlikely(ret)) { 60469e368259SAndrea Arcangeli ret = -ENOENT; 60478cc5fcbbSMina Almasry /* Free the allocated page which may have 60488cc5fcbbSMina Almasry * consumed a reservation. 60498cc5fcbbSMina Almasry */ 60508cc5fcbbSMina Almasry restore_reserve_on_error(h, dst_vma, dst_addr, page); 60518cc5fcbbSMina Almasry put_page(page); 60528cc5fcbbSMina Almasry 60538cc5fcbbSMina Almasry /* Allocate a temporary page to hold the copied 60548cc5fcbbSMina Almasry * contents. 60558cc5fcbbSMina Almasry */ 60568cc5fcbbSMina Almasry page = alloc_huge_page_vma(h, dst_vma, dst_addr); 60578cc5fcbbSMina Almasry if (!page) { 60588cc5fcbbSMina Almasry ret = -ENOMEM; 60598cc5fcbbSMina Almasry goto out; 60608cc5fcbbSMina Almasry } 60618fb5debcSMike Kravetz *pagep = page; 60628cc5fcbbSMina Almasry /* Set the outparam pagep and return to the caller to 60638cc5fcbbSMina Almasry * copy the contents outside the lock. Don't free the 60648cc5fcbbSMina Almasry * page. 60658cc5fcbbSMina Almasry */ 60668fb5debcSMike Kravetz goto out; 60678fb5debcSMike Kravetz } 60688fb5debcSMike Kravetz } else { 60698cc5fcbbSMina Almasry if (vm_shared && 60708cc5fcbbSMina Almasry hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 60718cc5fcbbSMina Almasry put_page(*pagep); 60728cc5fcbbSMina Almasry ret = -EEXIST; 60738cc5fcbbSMina Almasry *pagep = NULL; 60748cc5fcbbSMina Almasry goto out; 60758cc5fcbbSMina Almasry } 60768cc5fcbbSMina Almasry 60778cc5fcbbSMina Almasry page = alloc_huge_page(dst_vma, dst_addr, 0); 60788cc5fcbbSMina Almasry if (IS_ERR(page)) { 6079da9a298fSMiaohe Lin put_page(*pagep); 60808cc5fcbbSMina Almasry ret = -ENOMEM; 60818cc5fcbbSMina Almasry *pagep = NULL; 60828cc5fcbbSMina Almasry goto out; 60838cc5fcbbSMina Almasry } 608434892366SMuchun Song copy_user_huge_page(page, *pagep, dst_addr, dst_vma, 608534892366SMuchun Song pages_per_huge_page(h)); 60868cc5fcbbSMina Almasry put_page(*pagep); 60878fb5debcSMike Kravetz *pagep = NULL; 60888fb5debcSMike Kravetz } 60898fb5debcSMike Kravetz 60908fb5debcSMike Kravetz /* 60918fb5debcSMike Kravetz * The memory barrier inside __SetPageUptodate makes sure that 60928fb5debcSMike Kravetz * preceding stores to the page contents become visible before 60938fb5debcSMike Kravetz * the set_pte_at() write. 60948fb5debcSMike Kravetz */ 60958fb5debcSMike Kravetz __SetPageUptodate(page); 60968fb5debcSMike Kravetz 6097f6191471SAxel Rasmussen /* Add shared, newly allocated pages to the page cache. */ 6098f6191471SAxel Rasmussen if (vm_shared && !is_continue) { 60991e392147SAndrea Arcangeli size = i_size_read(mapping->host) >> huge_page_shift(h); 61001e392147SAndrea Arcangeli ret = -EFAULT; 61011e392147SAndrea Arcangeli if (idx >= size) 61021e392147SAndrea Arcangeli goto out_release_nounlock; 61031c9e8defSMike Kravetz 61041e392147SAndrea Arcangeli /* 61051e392147SAndrea Arcangeli * Serialization between remove_inode_hugepages() and 61067e1813d4SMike Kravetz * hugetlb_add_to_page_cache() below happens through the 61071e392147SAndrea Arcangeli * hugetlb_fault_mutex_table that here must be hold by 61081e392147SAndrea Arcangeli * the caller. 61091e392147SAndrea Arcangeli */ 61107e1813d4SMike Kravetz ret = hugetlb_add_to_page_cache(page, mapping, idx); 61111c9e8defSMike Kravetz if (ret) 61121c9e8defSMike Kravetz goto out_release_nounlock; 6113cc30042dSMina Almasry page_in_pagecache = true; 61141c9e8defSMike Kravetz } 61151c9e8defSMike Kravetz 6116bcc66543SMiaohe Lin ptl = huge_pte_lock(h, dst_mm, dst_pte); 61178fb5debcSMike Kravetz 61188625147cSJames Houghton ret = -EIO; 61198625147cSJames Houghton if (PageHWPoison(page)) 61208625147cSJames Houghton goto out_release_unlock; 61218625147cSJames Houghton 61221e392147SAndrea Arcangeli /* 61236041c691SPeter Xu * We allow to overwrite a pte marker: consider when both MISSING|WP 61246041c691SPeter Xu * registered, we firstly wr-protect a none pte which has no page cache 61256041c691SPeter Xu * page backing it, then access the page. 61266041c691SPeter Xu */ 6127fa27759aSMike Kravetz ret = -EEXIST; 61286041c691SPeter Xu if (!huge_pte_none_mostly(huge_ptep_get(dst_pte))) 61298fb5debcSMike Kravetz goto out_release_unlock; 61308fb5debcSMike Kravetz 61314781593dSPeter Xu if (page_in_pagecache) 6132fb3d824dSDavid Hildenbrand page_dup_file_rmap(page, true); 61334781593dSPeter Xu else 61348fb5debcSMike Kravetz hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); 61358fb5debcSMike Kravetz 61366041c691SPeter Xu /* 61376041c691SPeter Xu * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY 61386041c691SPeter Xu * with wp flag set, don't set pte write bit. 61396041c691SPeter Xu */ 61406041c691SPeter Xu if (wp_copy || (is_continue && !vm_shared)) 6141f6191471SAxel Rasmussen writable = 0; 6142f6191471SAxel Rasmussen else 6143f6191471SAxel Rasmussen writable = dst_vma->vm_flags & VM_WRITE; 6144f6191471SAxel Rasmussen 6145f6191471SAxel Rasmussen _dst_pte = make_huge_pte(dst_vma, page, writable); 61466041c691SPeter Xu /* 61476041c691SPeter Xu * Always mark UFFDIO_COPY page dirty; note that this may not be 61486041c691SPeter Xu * extremely important for hugetlbfs for now since swapping is not 61496041c691SPeter Xu * supported, but we should still be clear in that this page cannot be 61506041c691SPeter Xu * thrown away at will, even if write bit not set. 61516041c691SPeter Xu */ 61528fb5debcSMike Kravetz _dst_pte = huge_pte_mkdirty(_dst_pte); 61538fb5debcSMike Kravetz _dst_pte = pte_mkyoung(_dst_pte); 61548fb5debcSMike Kravetz 61556041c691SPeter Xu if (wp_copy) 61566041c691SPeter Xu _dst_pte = huge_pte_mkuffd_wp(_dst_pte); 61576041c691SPeter Xu 61588fb5debcSMike Kravetz set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); 61598fb5debcSMike Kravetz 61608fb5debcSMike Kravetz hugetlb_count_add(pages_per_huge_page(h), dst_mm); 61618fb5debcSMike Kravetz 61628fb5debcSMike Kravetz /* No need to invalidate - it was non-present before */ 61638fb5debcSMike Kravetz update_mmu_cache(dst_vma, dst_addr, dst_pte); 61648fb5debcSMike Kravetz 61658fb5debcSMike Kravetz spin_unlock(ptl); 6166f6191471SAxel Rasmussen if (!is_continue) 61678f251a3dSMike Kravetz SetHPageMigratable(page); 6168f6191471SAxel Rasmussen if (vm_shared || is_continue) 61691c9e8defSMike Kravetz unlock_page(page); 61708fb5debcSMike Kravetz ret = 0; 61718fb5debcSMike Kravetz out: 61728fb5debcSMike Kravetz return ret; 61738fb5debcSMike Kravetz out_release_unlock: 61748fb5debcSMike Kravetz spin_unlock(ptl); 6175f6191471SAxel Rasmussen if (vm_shared || is_continue) 61761c9e8defSMike Kravetz unlock_page(page); 61775af10dfdSAndrea Arcangeli out_release_nounlock: 6178cc30042dSMina Almasry if (!page_in_pagecache) 6179846be085SMike Kravetz restore_reserve_on_error(h, dst_vma, dst_addr, page); 61808fb5debcSMike Kravetz put_page(page); 61818fb5debcSMike Kravetz goto out; 61828fb5debcSMike Kravetz } 6183714c1891SAxel Rasmussen #endif /* CONFIG_USERFAULTFD */ 61848fb5debcSMike Kravetz 618582e5d378SJoao Martins static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma, 618682e5d378SJoao Martins int refs, struct page **pages, 618782e5d378SJoao Martins struct vm_area_struct **vmas) 618882e5d378SJoao Martins { 618982e5d378SJoao Martins int nr; 619082e5d378SJoao Martins 619182e5d378SJoao Martins for (nr = 0; nr < refs; nr++) { 619282e5d378SJoao Martins if (likely(pages)) 619314455eabSCheng Li pages[nr] = nth_page(page, nr); 619482e5d378SJoao Martins if (vmas) 619582e5d378SJoao Martins vmas[nr] = vma; 619682e5d378SJoao Martins } 619782e5d378SJoao Martins } 619882e5d378SJoao Martins 6199a7f22660SDavid Hildenbrand static inline bool __follow_hugetlb_must_fault(unsigned int flags, pte_t *pte, 6200a7f22660SDavid Hildenbrand bool *unshare) 6201a7f22660SDavid Hildenbrand { 6202a7f22660SDavid Hildenbrand pte_t pteval = huge_ptep_get(pte); 6203a7f22660SDavid Hildenbrand 6204a7f22660SDavid Hildenbrand *unshare = false; 6205a7f22660SDavid Hildenbrand if (is_swap_pte(pteval)) 6206a7f22660SDavid Hildenbrand return true; 6207a7f22660SDavid Hildenbrand if (huge_pte_write(pteval)) 6208a7f22660SDavid Hildenbrand return false; 6209a7f22660SDavid Hildenbrand if (flags & FOLL_WRITE) 6210a7f22660SDavid Hildenbrand return true; 6211a7f22660SDavid Hildenbrand if (gup_must_unshare(flags, pte_page(pteval))) { 6212a7f22660SDavid Hildenbrand *unshare = true; 6213a7f22660SDavid Hildenbrand return true; 6214a7f22660SDavid Hildenbrand } 6215a7f22660SDavid Hildenbrand return false; 6216a7f22660SDavid Hildenbrand } 6217a7f22660SDavid Hildenbrand 621857a196a5SMike Kravetz struct page *hugetlb_follow_page_mask(struct vm_area_struct *vma, 621957a196a5SMike Kravetz unsigned long address, unsigned int flags) 622057a196a5SMike Kravetz { 622157a196a5SMike Kravetz struct hstate *h = hstate_vma(vma); 622257a196a5SMike Kravetz struct mm_struct *mm = vma->vm_mm; 622357a196a5SMike Kravetz unsigned long haddr = address & huge_page_mask(h); 622457a196a5SMike Kravetz struct page *page = NULL; 622557a196a5SMike Kravetz spinlock_t *ptl; 622657a196a5SMike Kravetz pte_t *pte, entry; 622757a196a5SMike Kravetz 622857a196a5SMike Kravetz /* 622957a196a5SMike Kravetz * FOLL_PIN is not supported for follow_page(). Ordinary GUP goes via 623057a196a5SMike Kravetz * follow_hugetlb_page(). 623157a196a5SMike Kravetz */ 623257a196a5SMike Kravetz if (WARN_ON_ONCE(flags & FOLL_PIN)) 623357a196a5SMike Kravetz return NULL; 623457a196a5SMike Kravetz 623557a196a5SMike Kravetz retry: 623657a196a5SMike Kravetz pte = huge_pte_offset(mm, haddr, huge_page_size(h)); 623757a196a5SMike Kravetz if (!pte) 623857a196a5SMike Kravetz return NULL; 623957a196a5SMike Kravetz 624057a196a5SMike Kravetz ptl = huge_pte_lock(h, mm, pte); 624157a196a5SMike Kravetz entry = huge_ptep_get(pte); 624257a196a5SMike Kravetz if (pte_present(entry)) { 624357a196a5SMike Kravetz page = pte_page(entry) + 624457a196a5SMike Kravetz ((address & ~huge_page_mask(h)) >> PAGE_SHIFT); 624557a196a5SMike Kravetz /* 624657a196a5SMike Kravetz * Note that page may be a sub-page, and with vmemmap 624757a196a5SMike Kravetz * optimizations the page struct may be read only. 624857a196a5SMike Kravetz * try_grab_page() will increase the ref count on the 624957a196a5SMike Kravetz * head page, so this will be OK. 625057a196a5SMike Kravetz * 625157a196a5SMike Kravetz * try_grab_page() should always succeed here, because we hold 625257a196a5SMike Kravetz * the ptl lock and have verified pte_present(). 625357a196a5SMike Kravetz */ 625457a196a5SMike Kravetz if (WARN_ON_ONCE(!try_grab_page(page, flags))) { 625557a196a5SMike Kravetz page = NULL; 625657a196a5SMike Kravetz goto out; 625757a196a5SMike Kravetz } 625857a196a5SMike Kravetz } else { 625957a196a5SMike Kravetz if (is_hugetlb_entry_migration(entry)) { 626057a196a5SMike Kravetz spin_unlock(ptl); 626157a196a5SMike Kravetz __migration_entry_wait_huge(pte, ptl); 626257a196a5SMike Kravetz goto retry; 626357a196a5SMike Kravetz } 626457a196a5SMike Kravetz /* 626557a196a5SMike Kravetz * hwpoisoned entry is treated as no_page_table in 626657a196a5SMike Kravetz * follow_page_mask(). 626757a196a5SMike Kravetz */ 626857a196a5SMike Kravetz } 626957a196a5SMike Kravetz out: 627057a196a5SMike Kravetz spin_unlock(ptl); 627157a196a5SMike Kravetz return page; 627257a196a5SMike Kravetz } 627357a196a5SMike Kravetz 627428a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 627563551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 627628a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 62774f6da934SPeter Xu long i, unsigned int flags, int *locked) 627863551ae0SDavid Gibson { 6279d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 6280d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 628128a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 6282a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 62830fa5bc40SJoao Martins int err = -EFAULT, refs; 628463551ae0SDavid Gibson 628563551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 628663551ae0SDavid Gibson pte_t *pte; 6287cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 6288a7f22660SDavid Hildenbrand bool unshare = false; 62892a15efc9SHugh Dickins int absent; 629063551ae0SDavid Gibson struct page *page; 629163551ae0SDavid Gibson 62924c887265SAdam Litke /* 629302057967SDavid Rientjes * If we have a pending SIGKILL, don't keep faulting pages and 629402057967SDavid Rientjes * potentially allocating memory. 629502057967SDavid Rientjes */ 6296fa45f116SDavidlohr Bueso if (fatal_signal_pending(current)) { 629702057967SDavid Rientjes remainder = 0; 629802057967SDavid Rientjes break; 629902057967SDavid Rientjes } 630002057967SDavid Rientjes 630102057967SDavid Rientjes /* 63024c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 63032a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 63044c887265SAdam Litke * first, for the page indexing below to work. 6305cb900f41SKirill A. Shutemov * 6306cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 63074c887265SAdam Litke */ 63087868a208SPunit Agrawal pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), 63097868a208SPunit Agrawal huge_page_size(h)); 6310cb900f41SKirill A. Shutemov if (pte) 6311cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 63122a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 631363551ae0SDavid Gibson 63142a15efc9SHugh Dickins /* 63152a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 63163ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 63173ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 63183ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 63193ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 63202a15efc9SHugh Dickins */ 63213ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 63223ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 6323cb900f41SKirill A. Shutemov if (pte) 6324cb900f41SKirill A. Shutemov spin_unlock(ptl); 63252a15efc9SHugh Dickins remainder = 0; 63262a15efc9SHugh Dickins break; 63272a15efc9SHugh Dickins } 63282a15efc9SHugh Dickins 63299cc3a5bdSNaoya Horiguchi /* 63309cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 63319cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 63329cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 63339cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 63349cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 63359cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 63369cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 63379cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 63389cc3a5bdSNaoya Horiguchi */ 6339a7f22660SDavid Hildenbrand if (absent || 6340a7f22660SDavid Hildenbrand __follow_hugetlb_must_fault(flags, pte, &unshare)) { 63412b740303SSouptick Joarder vm_fault_t ret; 634287ffc118SAndrea Arcangeli unsigned int fault_flags = 0; 63434c887265SAdam Litke 6344cb900f41SKirill A. Shutemov if (pte) 6345cb900f41SKirill A. Shutemov spin_unlock(ptl); 634687ffc118SAndrea Arcangeli if (flags & FOLL_WRITE) 634787ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_WRITE; 6348a7f22660SDavid Hildenbrand else if (unshare) 6349a7f22660SDavid Hildenbrand fault_flags |= FAULT_FLAG_UNSHARE; 63504f6da934SPeter Xu if (locked) 635171335f37SPeter Xu fault_flags |= FAULT_FLAG_ALLOW_RETRY | 635271335f37SPeter Xu FAULT_FLAG_KILLABLE; 635387ffc118SAndrea Arcangeli if (flags & FOLL_NOWAIT) 635487ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_ALLOW_RETRY | 635587ffc118SAndrea Arcangeli FAULT_FLAG_RETRY_NOWAIT; 635687ffc118SAndrea Arcangeli if (flags & FOLL_TRIED) { 63574426e945SPeter Xu /* 63584426e945SPeter Xu * Note: FAULT_FLAG_ALLOW_RETRY and 63594426e945SPeter Xu * FAULT_FLAG_TRIED can co-exist 63604426e945SPeter Xu */ 636187ffc118SAndrea Arcangeli fault_flags |= FAULT_FLAG_TRIED; 636287ffc118SAndrea Arcangeli } 636387ffc118SAndrea Arcangeli ret = hugetlb_fault(mm, vma, vaddr, fault_flags); 636487ffc118SAndrea Arcangeli if (ret & VM_FAULT_ERROR) { 63652be7cfedSDaniel Jordan err = vm_fault_to_errno(ret, flags); 63661c59827dSHugh Dickins remainder = 0; 63671c59827dSHugh Dickins break; 63681c59827dSHugh Dickins } 636987ffc118SAndrea Arcangeli if (ret & VM_FAULT_RETRY) { 63704f6da934SPeter Xu if (locked && 63711ac25013SAndrea Arcangeli !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) 63724f6da934SPeter Xu *locked = 0; 637387ffc118SAndrea Arcangeli *nr_pages = 0; 637487ffc118SAndrea Arcangeli /* 637587ffc118SAndrea Arcangeli * VM_FAULT_RETRY must not return an 637687ffc118SAndrea Arcangeli * error, it will return zero 637787ffc118SAndrea Arcangeli * instead. 637887ffc118SAndrea Arcangeli * 637987ffc118SAndrea Arcangeli * No need to update "position" as the 638087ffc118SAndrea Arcangeli * caller will not check it after 638187ffc118SAndrea Arcangeli * *nr_pages is set to 0. 638287ffc118SAndrea Arcangeli */ 638387ffc118SAndrea Arcangeli return i; 638487ffc118SAndrea Arcangeli } 638587ffc118SAndrea Arcangeli continue; 638687ffc118SAndrea Arcangeli } 638763551ae0SDavid Gibson 6388a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 63897f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 63908fde12caSLinus Torvalds 6391b6a2619cSDavid Hildenbrand VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) && 6392b6a2619cSDavid Hildenbrand !PageAnonExclusive(page), page); 6393b6a2619cSDavid Hildenbrand 63948fde12caSLinus Torvalds /* 6395acbfb087SZhigang Lu * If subpage information not requested, update counters 6396acbfb087SZhigang Lu * and skip the same_page loop below. 6397acbfb087SZhigang Lu */ 6398acbfb087SZhigang Lu if (!pages && !vmas && !pfn_offset && 6399acbfb087SZhigang Lu (vaddr + huge_page_size(h) < vma->vm_end) && 6400acbfb087SZhigang Lu (remainder >= pages_per_huge_page(h))) { 6401acbfb087SZhigang Lu vaddr += huge_page_size(h); 6402acbfb087SZhigang Lu remainder -= pages_per_huge_page(h); 6403acbfb087SZhigang Lu i += pages_per_huge_page(h); 6404acbfb087SZhigang Lu spin_unlock(ptl); 6405acbfb087SZhigang Lu continue; 6406acbfb087SZhigang Lu } 6407acbfb087SZhigang Lu 6408d08af0a5SJoao Martins /* vaddr may not be aligned to PAGE_SIZE */ 6409d08af0a5SJoao Martins refs = min3(pages_per_huge_page(h) - pfn_offset, remainder, 6410d08af0a5SJoao Martins (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT); 64110fa5bc40SJoao Martins 641282e5d378SJoao Martins if (pages || vmas) 641314455eabSCheng Li record_subpages_vmas(nth_page(page, pfn_offset), 641482e5d378SJoao Martins vma, refs, 641582e5d378SJoao Martins likely(pages) ? pages + i : NULL, 641682e5d378SJoao Martins vmas ? vmas + i : NULL); 641763551ae0SDavid Gibson 641882e5d378SJoao Martins if (pages) { 64190fa5bc40SJoao Martins /* 6420822951d8SMatthew Wilcox (Oracle) * try_grab_folio() should always succeed here, 64210fa5bc40SJoao Martins * because: a) we hold the ptl lock, and b) we've just 64220fa5bc40SJoao Martins * checked that the huge page is present in the page 64230fa5bc40SJoao Martins * tables. If the huge page is present, then the tail 64240fa5bc40SJoao Martins * pages must also be present. The ptl prevents the 64250fa5bc40SJoao Martins * head page and tail pages from being rearranged in 64260fa5bc40SJoao Martins * any way. So this page must be available at this 64270fa5bc40SJoao Martins * point, unless the page refcount overflowed: 64280fa5bc40SJoao Martins */ 6429822951d8SMatthew Wilcox (Oracle) if (WARN_ON_ONCE(!try_grab_folio(pages[i], refs, 64300fa5bc40SJoao Martins flags))) { 64310fa5bc40SJoao Martins spin_unlock(ptl); 64320fa5bc40SJoao Martins remainder = 0; 64330fa5bc40SJoao Martins err = -ENOMEM; 64340fa5bc40SJoao Martins break; 64350fa5bc40SJoao Martins } 6436d5d4b0aaSChen, Kenneth W } 643782e5d378SJoao Martins 643882e5d378SJoao Martins vaddr += (refs << PAGE_SHIFT); 643982e5d378SJoao Martins remainder -= refs; 644082e5d378SJoao Martins i += refs; 644182e5d378SJoao Martins 6442cb900f41SKirill A. Shutemov spin_unlock(ptl); 644363551ae0SDavid Gibson } 644428a35716SMichel Lespinasse *nr_pages = remainder; 644587ffc118SAndrea Arcangeli /* 644687ffc118SAndrea Arcangeli * setting position is actually required only if remainder is 644787ffc118SAndrea Arcangeli * not zero but it's faster not to add a "if (remainder)" 644887ffc118SAndrea Arcangeli * branch. 644987ffc118SAndrea Arcangeli */ 645063551ae0SDavid Gibson *position = vaddr; 645163551ae0SDavid Gibson 64522be7cfedSDaniel Jordan return i ? i : err; 645363551ae0SDavid Gibson } 64548f860591SZhang, Yanmin 64557da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 64565a90d5a1SPeter Xu unsigned long address, unsigned long end, 64575a90d5a1SPeter Xu pgprot_t newprot, unsigned long cp_flags) 64588f860591SZhang, Yanmin { 64598f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 64608f860591SZhang, Yanmin unsigned long start = address; 64618f860591SZhang, Yanmin pte_t *ptep; 64628f860591SZhang, Yanmin pte_t pte; 6463a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 646460dfaad6SPeter Xu unsigned long pages = 0, psize = huge_page_size(h); 6465dff11abeSMike Kravetz bool shared_pmd = false; 6466ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 6467e95a9851SMike Kravetz unsigned long last_addr_mask; 64685a90d5a1SPeter Xu bool uffd_wp = cp_flags & MM_CP_UFFD_WP; 64695a90d5a1SPeter Xu bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE; 6470dff11abeSMike Kravetz 6471dff11abeSMike Kravetz /* 6472dff11abeSMike Kravetz * In the case of shared PMDs, the area to flush could be beyond 6473ac46d4f3SJérôme Glisse * start/end. Set range.start/range.end to cover the maximum possible 6474dff11abeSMike Kravetz * range if PMD sharing is possible. 6475dff11abeSMike Kravetz */ 64767269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 64777269f999SJérôme Glisse 0, vma, mm, start, end); 6478ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 64798f860591SZhang, Yanmin 64808f860591SZhang, Yanmin BUG_ON(address >= end); 6481ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 64828f860591SZhang, Yanmin 6483ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 648440549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 648583cde9e8SDavidlohr Bueso i_mmap_lock_write(vma->vm_file->f_mapping); 648640549ba8SMike Kravetz last_addr_mask = hugetlb_mask_last_page(h); 648760dfaad6SPeter Xu for (; address < end; address += psize) { 6488cb900f41SKirill A. Shutemov spinlock_t *ptl; 648960dfaad6SPeter Xu ptep = huge_pte_offset(mm, address, psize); 6490e95a9851SMike Kravetz if (!ptep) { 6491e95a9851SMike Kravetz address |= last_addr_mask; 64928f860591SZhang, Yanmin continue; 6493e95a9851SMike Kravetz } 6494cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 64954ddb4d91SMike Kravetz if (huge_pmd_unshare(mm, vma, address, ptep)) { 649660dfaad6SPeter Xu /* 649760dfaad6SPeter Xu * When uffd-wp is enabled on the vma, unshare 649860dfaad6SPeter Xu * shouldn't happen at all. Warn about it if it 649960dfaad6SPeter Xu * happened due to some reason. 650060dfaad6SPeter Xu */ 650160dfaad6SPeter Xu WARN_ON_ONCE(uffd_wp || uffd_wp_resolve); 65027da4d641SPeter Zijlstra pages++; 6503cb900f41SKirill A. Shutemov spin_unlock(ptl); 6504dff11abeSMike Kravetz shared_pmd = true; 65054ddb4d91SMike Kravetz address |= last_addr_mask; 650639dde65cSChen, Kenneth W continue; 65077da4d641SPeter Zijlstra } 6508a8bda28dSNaoya Horiguchi pte = huge_ptep_get(ptep); 6509a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 6510a8bda28dSNaoya Horiguchi spin_unlock(ptl); 6511a8bda28dSNaoya Horiguchi continue; 6512a8bda28dSNaoya Horiguchi } 6513a8bda28dSNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(pte))) { 6514a8bda28dSNaoya Horiguchi swp_entry_t entry = pte_to_swp_entry(pte); 65156c287605SDavid Hildenbrand struct page *page = pfn_swap_entry_to_page(entry); 6516a8bda28dSNaoya Horiguchi 65176c287605SDavid Hildenbrand if (!is_readable_migration_entry(entry)) { 6518a8bda28dSNaoya Horiguchi pte_t newpte; 6519a8bda28dSNaoya Horiguchi 65206c287605SDavid Hildenbrand if (PageAnon(page)) 65216c287605SDavid Hildenbrand entry = make_readable_exclusive_migration_entry( 65226c287605SDavid Hildenbrand swp_offset(entry)); 65236c287605SDavid Hildenbrand else 65244dd845b5SAlistair Popple entry = make_readable_migration_entry( 65254dd845b5SAlistair Popple swp_offset(entry)); 6526a8bda28dSNaoya Horiguchi newpte = swp_entry_to_pte(entry); 65275a90d5a1SPeter Xu if (uffd_wp) 65285a90d5a1SPeter Xu newpte = pte_swp_mkuffd_wp(newpte); 65295a90d5a1SPeter Xu else if (uffd_wp_resolve) 65305a90d5a1SPeter Xu newpte = pte_swp_clear_uffd_wp(newpte); 653118f39629SQi Zheng set_huge_pte_at(mm, address, ptep, newpte); 6532a8bda28dSNaoya Horiguchi pages++; 6533a8bda28dSNaoya Horiguchi } 6534a8bda28dSNaoya Horiguchi spin_unlock(ptl); 6535a8bda28dSNaoya Horiguchi continue; 6536a8bda28dSNaoya Horiguchi } 653760dfaad6SPeter Xu if (unlikely(pte_marker_uffd_wp(pte))) { 653860dfaad6SPeter Xu /* 653960dfaad6SPeter Xu * This is changing a non-present pte into a none pte, 654060dfaad6SPeter Xu * no need for huge_ptep_modify_prot_start/commit(). 654160dfaad6SPeter Xu */ 654260dfaad6SPeter Xu if (uffd_wp_resolve) 654360dfaad6SPeter Xu huge_pte_clear(mm, address, ptep, psize); 654460dfaad6SPeter Xu } 6545a8bda28dSNaoya Horiguchi if (!huge_pte_none(pte)) { 6546023bdd00SAneesh Kumar K.V pte_t old_pte; 654779c1c594SChristophe Leroy unsigned int shift = huge_page_shift(hstate_vma(vma)); 6548023bdd00SAneesh Kumar K.V 6549023bdd00SAneesh Kumar K.V old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 655016785bd7SAnshuman Khandual pte = huge_pte_modify(old_pte, newprot); 655179c1c594SChristophe Leroy pte = arch_make_huge_pte(pte, shift, vma->vm_flags); 65525a90d5a1SPeter Xu if (uffd_wp) 65535a90d5a1SPeter Xu pte = huge_pte_mkuffd_wp(huge_pte_wrprotect(pte)); 65545a90d5a1SPeter Xu else if (uffd_wp_resolve) 65555a90d5a1SPeter Xu pte = huge_pte_clear_uffd_wp(pte); 6556023bdd00SAneesh Kumar K.V huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 65577da4d641SPeter Zijlstra pages++; 655860dfaad6SPeter Xu } else { 655960dfaad6SPeter Xu /* None pte */ 656060dfaad6SPeter Xu if (unlikely(uffd_wp)) 656160dfaad6SPeter Xu /* Safe to modify directly (none->non-present). */ 656260dfaad6SPeter Xu set_huge_pte_at(mm, address, ptep, 656360dfaad6SPeter Xu make_pte_marker(PTE_MARKER_UFFD_WP)); 65648f860591SZhang, Yanmin } 6565cb900f41SKirill A. Shutemov spin_unlock(ptl); 65668f860591SZhang, Yanmin } 6567d833352aSMel Gorman /* 6568c8c06efaSDavidlohr Bueso * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 6569d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 6570c8c06efaSDavidlohr Bueso * once we release i_mmap_rwsem, another task can do the final put_page 6571dff11abeSMike Kravetz * and that page table be reused and filled with junk. If we actually 6572dff11abeSMike Kravetz * did unshare a page of pmds, flush the range corresponding to the pud. 6573d833352aSMel Gorman */ 6574dff11abeSMike Kravetz if (shared_pmd) 6575ac46d4f3SJérôme Glisse flush_hugetlb_tlb_range(vma, range.start, range.end); 6576dff11abeSMike Kravetz else 65775491ae7bSAneesh Kumar K.V flush_hugetlb_tlb_range(vma, start, end); 65780f10851eSJérôme Glisse /* 65790f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() we are downgrading 65800f10851eSJérôme Glisse * page table protection not changing it to point to a new page. 65810f10851eSJérôme Glisse * 6582ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 65830f10851eSJérôme Glisse */ 658483cde9e8SDavidlohr Bueso i_mmap_unlock_write(vma->vm_file->f_mapping); 658540549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 6586ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 65877da4d641SPeter Zijlstra 65887da4d641SPeter Zijlstra return pages << h->order; 65898f860591SZhang, Yanmin } 65908f860591SZhang, Yanmin 659133b8f84aSMike Kravetz /* Return true if reservation was successful, false otherwise. */ 659233b8f84aSMike Kravetz bool hugetlb_reserve_pages(struct inode *inode, 6593a1e78772SMel Gorman long from, long to, 65945a6fe125SMel Gorman struct vm_area_struct *vma, 6595ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 6596e4e574b7SAdam Litke { 659733b8f84aSMike Kravetz long chg, add = -1; 6598a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 659990481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 66009119a41eSJoonsoo Kim struct resv_map *resv_map; 6601075a61d0SMina Almasry struct hugetlb_cgroup *h_cg = NULL; 66020db9d74eSMina Almasry long gbl_reserve, regions_needed = 0; 6603e4e574b7SAdam Litke 660463489f8eSMike Kravetz /* This should never happen */ 660563489f8eSMike Kravetz if (from > to) { 660663489f8eSMike Kravetz VM_WARN(1, "%s called with a negative range\n", __func__); 660733b8f84aSMike Kravetz return false; 660863489f8eSMike Kravetz } 660963489f8eSMike Kravetz 6610a1e78772SMel Gorman /* 66118d9bfb26SMike Kravetz * vma specific semaphore used for pmd sharing synchronization 66128d9bfb26SMike Kravetz */ 66138d9bfb26SMike Kravetz hugetlb_vma_lock_alloc(vma); 66148d9bfb26SMike Kravetz 66158d9bfb26SMike Kravetz /* 661617c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 661717c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 661890481622SDavid Gibson * without using reserves 661917c9d12eSMel Gorman */ 6620ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 662133b8f84aSMike Kravetz return true; 662217c9d12eSMel Gorman 662317c9d12eSMel Gorman /* 6624a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 6625a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 6626a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 6627a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 6628a1e78772SMel Gorman */ 66299119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 6630f27a5136SMike Kravetz /* 6631f27a5136SMike Kravetz * resv_map can not be NULL as hugetlb_reserve_pages is only 6632f27a5136SMike Kravetz * called for inodes for which resv_maps were created (see 6633f27a5136SMike Kravetz * hugetlbfs_get_inode). 6634f27a5136SMike Kravetz */ 66354e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 66369119a41eSJoonsoo Kim 66370db9d74eSMina Almasry chg = region_chg(resv_map, from, to, ®ions_needed); 66389119a41eSJoonsoo Kim } else { 6639e9fe92aeSMina Almasry /* Private mapping. */ 66409119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 66415a6fe125SMel Gorman if (!resv_map) 66428d9bfb26SMike Kravetz goto out_err; 66435a6fe125SMel Gorman 664417c9d12eSMel Gorman chg = to - from; 664517c9d12eSMel Gorman 66465a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 66475a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 66485a6fe125SMel Gorman } 66495a6fe125SMel Gorman 665033b8f84aSMike Kravetz if (chg < 0) 6651c50ac050SDave Hansen goto out_err; 665217c9d12eSMel Gorman 665333b8f84aSMike Kravetz if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), 665433b8f84aSMike Kravetz chg * pages_per_huge_page(h), &h_cg) < 0) 6655075a61d0SMina Almasry goto out_err; 6656075a61d0SMina Almasry 6657075a61d0SMina Almasry if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { 6658075a61d0SMina Almasry /* For private mappings, the hugetlb_cgroup uncharge info hangs 6659075a61d0SMina Almasry * of the resv_map. 6660075a61d0SMina Almasry */ 6661075a61d0SMina Almasry resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); 6662075a61d0SMina Almasry } 6663075a61d0SMina Almasry 66641c5ecae3SMike Kravetz /* 66651c5ecae3SMike Kravetz * There must be enough pages in the subpool for the mapping. If 66661c5ecae3SMike Kravetz * the subpool has a minimum size, there may be some global 66671c5ecae3SMike Kravetz * reservations already in place (gbl_reserve). 66681c5ecae3SMike Kravetz */ 66691c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_get_pages(spool, chg); 667033b8f84aSMike Kravetz if (gbl_reserve < 0) 6671075a61d0SMina Almasry goto out_uncharge_cgroup; 667217c9d12eSMel Gorman 667317c9d12eSMel Gorman /* 667417c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 667590481622SDavid Gibson * Hand the pages back to the subpool if there are not 667617c9d12eSMel Gorman */ 667733b8f84aSMike Kravetz if (hugetlb_acct_memory(h, gbl_reserve) < 0) 6678075a61d0SMina Almasry goto out_put_pages; 667917c9d12eSMel Gorman 668017c9d12eSMel Gorman /* 668117c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 668217c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 668317c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 668417c9d12eSMel Gorman * the reservation was and the page cache tells how much of 668517c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 668617c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 668717c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 668817c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 668917c9d12eSMel Gorman * else has to be done for private mappings here 669017c9d12eSMel Gorman */ 669133039678SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) { 6692075a61d0SMina Almasry add = region_add(resv_map, from, to, regions_needed, h, h_cg); 669333039678SMike Kravetz 66940db9d74eSMina Almasry if (unlikely(add < 0)) { 66950db9d74eSMina Almasry hugetlb_acct_memory(h, -gbl_reserve); 6696075a61d0SMina Almasry goto out_put_pages; 66970db9d74eSMina Almasry } else if (unlikely(chg > add)) { 669833039678SMike Kravetz /* 669933039678SMike Kravetz * pages in this range were added to the reserve 670033039678SMike Kravetz * map between region_chg and region_add. This 670133039678SMike Kravetz * indicates a race with alloc_huge_page. Adjust 670233039678SMike Kravetz * the subpool and reserve counts modified above 670333039678SMike Kravetz * based on the difference. 670433039678SMike Kravetz */ 670533039678SMike Kravetz long rsv_adjust; 670633039678SMike Kravetz 6707d85aecf2SMiaohe Lin /* 6708d85aecf2SMiaohe Lin * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the 6709d85aecf2SMiaohe Lin * reference to h_cg->css. See comment below for detail. 6710d85aecf2SMiaohe Lin */ 6711075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd( 6712075a61d0SMina Almasry hstate_index(h), 6713075a61d0SMina Almasry (chg - add) * pages_per_huge_page(h), h_cg); 6714075a61d0SMina Almasry 671533039678SMike Kravetz rsv_adjust = hugepage_subpool_put_pages(spool, 671633039678SMike Kravetz chg - add); 671733039678SMike Kravetz hugetlb_acct_memory(h, -rsv_adjust); 6718d85aecf2SMiaohe Lin } else if (h_cg) { 6719d85aecf2SMiaohe Lin /* 6720d85aecf2SMiaohe Lin * The file_regions will hold their own reference to 6721d85aecf2SMiaohe Lin * h_cg->css. So we should release the reference held 6722d85aecf2SMiaohe Lin * via hugetlb_cgroup_charge_cgroup_rsvd() when we are 6723d85aecf2SMiaohe Lin * done. 6724d85aecf2SMiaohe Lin */ 6725d85aecf2SMiaohe Lin hugetlb_cgroup_put_rsvd_cgroup(h_cg); 672633039678SMike Kravetz } 672733039678SMike Kravetz } 672833b8f84aSMike Kravetz return true; 672933b8f84aSMike Kravetz 6730075a61d0SMina Almasry out_put_pages: 6731075a61d0SMina Almasry /* put back original number of pages, chg */ 6732075a61d0SMina Almasry (void)hugepage_subpool_put_pages(spool, chg); 6733075a61d0SMina Almasry out_uncharge_cgroup: 6734075a61d0SMina Almasry hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), 6735075a61d0SMina Almasry chg * pages_per_huge_page(h), h_cg); 6736c50ac050SDave Hansen out_err: 67378d9bfb26SMike Kravetz hugetlb_vma_lock_free(vma); 67385e911373SMike Kravetz if (!vma || vma->vm_flags & VM_MAYSHARE) 67390db9d74eSMina Almasry /* Only call region_abort if the region_chg succeeded but the 67400db9d74eSMina Almasry * region_add failed or didn't run. 67410db9d74eSMina Almasry */ 67420db9d74eSMina Almasry if (chg >= 0 && add < 0) 67430db9d74eSMina Almasry region_abort(resv_map, from, to, regions_needed); 6744f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 6745f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 674633b8f84aSMike Kravetz return false; 6747a43a8c39SChen, Kenneth W } 6748a43a8c39SChen, Kenneth W 6749b5cec28dSMike Kravetz long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 6750b5cec28dSMike Kravetz long freed) 6751a43a8c39SChen, Kenneth W { 6752a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 67534e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 67549119a41eSJoonsoo Kim long chg = 0; 675590481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 67561c5ecae3SMike Kravetz long gbl_reserve; 675745c682a6SKen Chen 6758f27a5136SMike Kravetz /* 6759f27a5136SMike Kravetz * Since this routine can be called in the evict inode path for all 6760f27a5136SMike Kravetz * hugetlbfs inodes, resv_map could be NULL. 6761f27a5136SMike Kravetz */ 6762b5cec28dSMike Kravetz if (resv_map) { 6763b5cec28dSMike Kravetz chg = region_del(resv_map, start, end); 6764b5cec28dSMike Kravetz /* 6765b5cec28dSMike Kravetz * region_del() can fail in the rare case where a region 6766b5cec28dSMike Kravetz * must be split and another region descriptor can not be 6767b5cec28dSMike Kravetz * allocated. If end == LONG_MAX, it will not fail. 6768b5cec28dSMike Kravetz */ 6769b5cec28dSMike Kravetz if (chg < 0) 6770b5cec28dSMike Kravetz return chg; 6771b5cec28dSMike Kravetz } 6772b5cec28dSMike Kravetz 677345c682a6SKen Chen spin_lock(&inode->i_lock); 6774e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 677545c682a6SKen Chen spin_unlock(&inode->i_lock); 677645c682a6SKen Chen 67771c5ecae3SMike Kravetz /* 67781c5ecae3SMike Kravetz * If the subpool has a minimum size, the number of global 67791c5ecae3SMike Kravetz * reservations to be released may be adjusted. 6780dddf31a4SMiaohe Lin * 6781dddf31a4SMiaohe Lin * Note that !resv_map implies freed == 0. So (chg - freed) 6782dddf31a4SMiaohe Lin * won't go negative. 67831c5ecae3SMike Kravetz */ 67841c5ecae3SMike Kravetz gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 67851c5ecae3SMike Kravetz hugetlb_acct_memory(h, -gbl_reserve); 6786b5cec28dSMike Kravetz 6787b5cec28dSMike Kravetz return 0; 6788a43a8c39SChen, Kenneth W } 678993f70f90SNaoya Horiguchi 67903212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 67913212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 67923212b535SSteve Capper struct vm_area_struct *vma, 67933212b535SSteve Capper unsigned long addr, pgoff_t idx) 67943212b535SSteve Capper { 67953212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 67963212b535SSteve Capper svma->vm_start; 67973212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 67983212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 67993212b535SSteve Capper 68003212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 6801de60f5f1SEric B Munson unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; 6802de60f5f1SEric B Munson unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; 68033212b535SSteve Capper 68043212b535SSteve Capper /* 68053212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 68063212b535SSteve Capper * page table page. 6807131a79b4SMike Kravetz * 6808131a79b4SMike Kravetz * Also, vma_lock (vm_private_data) is required for sharing. 68093212b535SSteve Capper */ 68103212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 68113212b535SSteve Capper vm_flags != svm_flags || 6812131a79b4SMike Kravetz !range_in_vma(svma, sbase, s_end) || 6813131a79b4SMike Kravetz !svma->vm_private_data) 68143212b535SSteve Capper return 0; 68153212b535SSteve Capper 68163212b535SSteve Capper return saddr; 68173212b535SSteve Capper } 68183212b535SSteve Capper 6819bbff39ccSMike Kravetz bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 68203212b535SSteve Capper { 6821bbff39ccSMike Kravetz unsigned long start = addr & PUD_MASK; 6822bbff39ccSMike Kravetz unsigned long end = start + PUD_SIZE; 6823bbff39ccSMike Kravetz 68248d9bfb26SMike Kravetz #ifdef CONFIG_USERFAULTFD 68258d9bfb26SMike Kravetz if (uffd_disable_huge_pmd_share(vma)) 68268d9bfb26SMike Kravetz return false; 68278d9bfb26SMike Kravetz #endif 68283212b535SSteve Capper /* 68293212b535SSteve Capper * check on proper vm_flags and page table alignment 68303212b535SSteve Capper */ 68318d9bfb26SMike Kravetz if (!(vma->vm_flags & VM_MAYSHARE)) 683231aafb45SNicholas Krause return false; 6833bbff39ccSMike Kravetz if (!vma->vm_private_data) /* vma lock required for sharing */ 68348d9bfb26SMike Kravetz return false; 68358d9bfb26SMike Kravetz if (!range_in_vma(vma, start, end)) 68368d9bfb26SMike Kravetz return false; 68378d9bfb26SMike Kravetz return true; 68388d9bfb26SMike Kravetz } 68398d9bfb26SMike Kravetz 68403212b535SSteve Capper /* 6841017b1660SMike Kravetz * Determine if start,end range within vma could be mapped by shared pmd. 6842017b1660SMike Kravetz * If yes, adjust start and end to cover range associated with possible 6843017b1660SMike Kravetz * shared pmd mappings. 6844017b1660SMike Kravetz */ 6845017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 6846017b1660SMike Kravetz unsigned long *start, unsigned long *end) 6847017b1660SMike Kravetz { 6848a1ba9da8SLi Xinhai unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), 6849a1ba9da8SLi Xinhai v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 6850017b1660SMike Kravetz 6851a1ba9da8SLi Xinhai /* 6852f0953a1bSIngo Molnar * vma needs to span at least one aligned PUD size, and the range 6853f0953a1bSIngo Molnar * must be at least partially within in. 6854a1ba9da8SLi Xinhai */ 6855a1ba9da8SLi Xinhai if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || 6856a1ba9da8SLi Xinhai (*end <= v_start) || (*start >= v_end)) 6857017b1660SMike Kravetz return; 6858017b1660SMike Kravetz 685975802ca6SPeter Xu /* Extend the range to be PUD aligned for a worst case scenario */ 6860a1ba9da8SLi Xinhai if (*start > v_start) 6861a1ba9da8SLi Xinhai *start = ALIGN_DOWN(*start, PUD_SIZE); 6862017b1660SMike Kravetz 6863a1ba9da8SLi Xinhai if (*end < v_end) 6864a1ba9da8SLi Xinhai *end = ALIGN(*end, PUD_SIZE); 6865017b1660SMike Kravetz } 6866017b1660SMike Kravetz 68678d9bfb26SMike Kravetz static bool __vma_shareable_flags_pmd(struct vm_area_struct *vma) 68688d9bfb26SMike Kravetz { 68698d9bfb26SMike Kravetz return vma->vm_flags & (VM_MAYSHARE | VM_SHARED) && 68708d9bfb26SMike Kravetz vma->vm_private_data; 68718d9bfb26SMike Kravetz } 68728d9bfb26SMike Kravetz 68738d9bfb26SMike Kravetz void hugetlb_vma_lock_read(struct vm_area_struct *vma) 68748d9bfb26SMike Kravetz { 68758d9bfb26SMike Kravetz if (__vma_shareable_flags_pmd(vma)) { 68768d9bfb26SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 68778d9bfb26SMike Kravetz 68788d9bfb26SMike Kravetz down_read(&vma_lock->rw_sema); 68798d9bfb26SMike Kravetz } 68808d9bfb26SMike Kravetz } 68818d9bfb26SMike Kravetz 68828d9bfb26SMike Kravetz void hugetlb_vma_unlock_read(struct vm_area_struct *vma) 68838d9bfb26SMike Kravetz { 68848d9bfb26SMike Kravetz if (__vma_shareable_flags_pmd(vma)) { 68858d9bfb26SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 68868d9bfb26SMike Kravetz 68878d9bfb26SMike Kravetz up_read(&vma_lock->rw_sema); 68888d9bfb26SMike Kravetz } 68898d9bfb26SMike Kravetz } 68908d9bfb26SMike Kravetz 68918d9bfb26SMike Kravetz void hugetlb_vma_lock_write(struct vm_area_struct *vma) 68928d9bfb26SMike Kravetz { 68938d9bfb26SMike Kravetz if (__vma_shareable_flags_pmd(vma)) { 68948d9bfb26SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 68958d9bfb26SMike Kravetz 68968d9bfb26SMike Kravetz down_write(&vma_lock->rw_sema); 68978d9bfb26SMike Kravetz } 68988d9bfb26SMike Kravetz } 68998d9bfb26SMike Kravetz 69008d9bfb26SMike Kravetz void hugetlb_vma_unlock_write(struct vm_area_struct *vma) 69018d9bfb26SMike Kravetz { 69028d9bfb26SMike Kravetz if (__vma_shareable_flags_pmd(vma)) { 69038d9bfb26SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 69048d9bfb26SMike Kravetz 69058d9bfb26SMike Kravetz up_write(&vma_lock->rw_sema); 69068d9bfb26SMike Kravetz } 69078d9bfb26SMike Kravetz } 69088d9bfb26SMike Kravetz 69098d9bfb26SMike Kravetz int hugetlb_vma_trylock_write(struct vm_area_struct *vma) 69108d9bfb26SMike Kravetz { 69118d9bfb26SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 69128d9bfb26SMike Kravetz 69138d9bfb26SMike Kravetz if (!__vma_shareable_flags_pmd(vma)) 69148d9bfb26SMike Kravetz return 1; 69158d9bfb26SMike Kravetz 69168d9bfb26SMike Kravetz return down_write_trylock(&vma_lock->rw_sema); 69178d9bfb26SMike Kravetz } 69188d9bfb26SMike Kravetz 69198d9bfb26SMike Kravetz void hugetlb_vma_assert_locked(struct vm_area_struct *vma) 69208d9bfb26SMike Kravetz { 69218d9bfb26SMike Kravetz if (__vma_shareable_flags_pmd(vma)) { 69228d9bfb26SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 69238d9bfb26SMike Kravetz 69248d9bfb26SMike Kravetz lockdep_assert_held(&vma_lock->rw_sema); 69258d9bfb26SMike Kravetz } 69268d9bfb26SMike Kravetz } 69278d9bfb26SMike Kravetz 69288d9bfb26SMike Kravetz void hugetlb_vma_lock_release(struct kref *kref) 69298d9bfb26SMike Kravetz { 69308d9bfb26SMike Kravetz struct hugetlb_vma_lock *vma_lock = container_of(kref, 69318d9bfb26SMike Kravetz struct hugetlb_vma_lock, refs); 69328d9bfb26SMike Kravetz 69338d9bfb26SMike Kravetz kfree(vma_lock); 69348d9bfb26SMike Kravetz } 69358d9bfb26SMike Kravetz 6936acfac378SAndrew Morton static void __hugetlb_vma_unlock_write_put(struct hugetlb_vma_lock *vma_lock) 6937ecfbd733SMike Kravetz { 6938ecfbd733SMike Kravetz struct vm_area_struct *vma = vma_lock->vma; 6939ecfbd733SMike Kravetz 6940ecfbd733SMike Kravetz /* 6941ecfbd733SMike Kravetz * vma_lock structure may or not be released as a result of put, 6942ecfbd733SMike Kravetz * it certainly will no longer be attached to vma so clear pointer. 6943ecfbd733SMike Kravetz * Semaphore synchronizes access to vma_lock->vma field. 6944ecfbd733SMike Kravetz */ 6945ecfbd733SMike Kravetz vma_lock->vma = NULL; 6946ecfbd733SMike Kravetz vma->vm_private_data = NULL; 6947ecfbd733SMike Kravetz up_write(&vma_lock->rw_sema); 6948ecfbd733SMike Kravetz kref_put(&vma_lock->refs, hugetlb_vma_lock_release); 6949ecfbd733SMike Kravetz } 6950ecfbd733SMike Kravetz 6951ecfbd733SMike Kravetz static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma) 6952ecfbd733SMike Kravetz { 6953ecfbd733SMike Kravetz if (__vma_shareable_flags_pmd(vma)) { 6954ecfbd733SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 6955ecfbd733SMike Kravetz 6956ecfbd733SMike Kravetz __hugetlb_vma_unlock_write_put(vma_lock); 6957ecfbd733SMike Kravetz } 6958ecfbd733SMike Kravetz } 6959ecfbd733SMike Kravetz 69608d9bfb26SMike Kravetz static void hugetlb_vma_lock_free(struct vm_area_struct *vma) 69618d9bfb26SMike Kravetz { 69628d9bfb26SMike Kravetz /* 6963131a79b4SMike Kravetz * Only present in sharable vmas. 69648d9bfb26SMike Kravetz */ 6965131a79b4SMike Kravetz if (!vma || !__vma_shareable_flags_pmd(vma)) 69668d9bfb26SMike Kravetz return; 69678d9bfb26SMike Kravetz 69688d9bfb26SMike Kravetz if (vma->vm_private_data) { 69698d9bfb26SMike Kravetz struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 69708d9bfb26SMike Kravetz 6971ecfbd733SMike Kravetz down_write(&vma_lock->rw_sema); 6972ecfbd733SMike Kravetz __hugetlb_vma_unlock_write_put(vma_lock); 69738d9bfb26SMike Kravetz } 69748d9bfb26SMike Kravetz } 69758d9bfb26SMike Kravetz 69768d9bfb26SMike Kravetz static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma) 69778d9bfb26SMike Kravetz { 69788d9bfb26SMike Kravetz struct hugetlb_vma_lock *vma_lock; 69798d9bfb26SMike Kravetz 69808d9bfb26SMike Kravetz /* Only establish in (flags) sharable vmas */ 69818d9bfb26SMike Kravetz if (!vma || !(vma->vm_flags & VM_MAYSHARE)) 69828d9bfb26SMike Kravetz return; 69838d9bfb26SMike Kravetz 69848d9bfb26SMike Kravetz /* Should never get here with non-NULL vm_private_data */ 69858d9bfb26SMike Kravetz if (vma->vm_private_data) 69868d9bfb26SMike Kravetz return; 69878d9bfb26SMike Kravetz 69888d9bfb26SMike Kravetz vma_lock = kmalloc(sizeof(*vma_lock), GFP_KERNEL); 6989bbff39ccSMike Kravetz if (!vma_lock) { 69908d9bfb26SMike Kravetz /* 69918d9bfb26SMike Kravetz * If we can not allocate structure, then vma can not 6992bbff39ccSMike Kravetz * participate in pmd sharing. This is only a possible 6993bbff39ccSMike Kravetz * performance enhancement and memory saving issue. 6994bbff39ccSMike Kravetz * However, the lock is also used to synchronize page 6995bbff39ccSMike Kravetz * faults with truncation. If the lock is not present, 6996bbff39ccSMike Kravetz * unlikely races could leave pages in a file past i_size 6997bbff39ccSMike Kravetz * until the file is removed. Warn in the unlikely case of 6998bbff39ccSMike Kravetz * allocation failure. 69998d9bfb26SMike Kravetz */ 7000bbff39ccSMike Kravetz pr_warn_once("HugeTLB: unable to allocate vma specific lock\n"); 70018d9bfb26SMike Kravetz return; 7002bbff39ccSMike Kravetz } 70038d9bfb26SMike Kravetz 70048d9bfb26SMike Kravetz kref_init(&vma_lock->refs); 70058d9bfb26SMike Kravetz init_rwsem(&vma_lock->rw_sema); 70068d9bfb26SMike Kravetz vma_lock->vma = vma; 70078d9bfb26SMike Kravetz vma->vm_private_data = vma_lock; 70088d9bfb26SMike Kravetz } 70098d9bfb26SMike Kravetz 7010017b1660SMike Kravetz /* 70113212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 70123212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 70133212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 70143a47c54fSMike Kravetz * code much cleaner. pmd allocation is essential for the shared case because 70153a47c54fSMike Kravetz * pud has to be populated inside the same i_mmap_rwsem section - otherwise 70163a47c54fSMike Kravetz * racing tasks could either miss the sharing (see huge_pte_offset) or select a 70173a47c54fSMike Kravetz * bad pmd for sharing. 70183212b535SSteve Capper */ 7019aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 7020aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 70213212b535SSteve Capper { 70223212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 70233212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 70243212b535SSteve Capper vma->vm_pgoff; 70253212b535SSteve Capper struct vm_area_struct *svma; 70263212b535SSteve Capper unsigned long saddr; 70273212b535SSteve Capper pte_t *spte = NULL; 70283212b535SSteve Capper pte_t *pte; 7029cb900f41SKirill A. Shutemov spinlock_t *ptl; 70303212b535SSteve Capper 70313a47c54fSMike Kravetz i_mmap_lock_read(mapping); 70323212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 70333212b535SSteve Capper if (svma == vma) 70343212b535SSteve Capper continue; 70353212b535SSteve Capper 70363212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 70373212b535SSteve Capper if (saddr) { 70387868a208SPunit Agrawal spte = huge_pte_offset(svma->vm_mm, saddr, 70397868a208SPunit Agrawal vma_mmu_pagesize(svma)); 70403212b535SSteve Capper if (spte) { 70413212b535SSteve Capper get_page(virt_to_page(spte)); 70423212b535SSteve Capper break; 70433212b535SSteve Capper } 70443212b535SSteve Capper } 70453212b535SSteve Capper } 70463212b535SSteve Capper 70473212b535SSteve Capper if (!spte) 70483212b535SSteve Capper goto out; 70493212b535SSteve Capper 70508bea8052SAneesh Kumar K.V ptl = huge_pte_lock(hstate_vma(vma), mm, spte); 7051dc6c9a35SKirill A. Shutemov if (pud_none(*pud)) { 70523212b535SSteve Capper pud_populate(mm, pud, 70533212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 7054c17b1f42SKirill A. Shutemov mm_inc_nr_pmds(mm); 7055dc6c9a35SKirill A. Shutemov } else { 70563212b535SSteve Capper put_page(virt_to_page(spte)); 7057dc6c9a35SKirill A. Shutemov } 7058cb900f41SKirill A. Shutemov spin_unlock(ptl); 70593212b535SSteve Capper out: 70603212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 70613a47c54fSMike Kravetz i_mmap_unlock_read(mapping); 70623212b535SSteve Capper return pte; 70633212b535SSteve Capper } 70643212b535SSteve Capper 70653212b535SSteve Capper /* 70663212b535SSteve Capper * unmap huge page backed by shared pte. 70673212b535SSteve Capper * 70683212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 70693212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 70703212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 70713212b535SSteve Capper * 70723a47c54fSMike Kravetz * Called with page table lock held. 70733212b535SSteve Capper * 70743212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 70753212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 70763212b535SSteve Capper */ 707734ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 70784ddb4d91SMike Kravetz unsigned long addr, pte_t *ptep) 70793212b535SSteve Capper { 70804ddb4d91SMike Kravetz pgd_t *pgd = pgd_offset(mm, addr); 70814ddb4d91SMike Kravetz p4d_t *p4d = p4d_offset(pgd, addr); 70824ddb4d91SMike Kravetz pud_t *pud = pud_offset(p4d, addr); 70833212b535SSteve Capper 708434ae204fSMike Kravetz i_mmap_assert_write_locked(vma->vm_file->f_mapping); 708540549ba8SMike Kravetz hugetlb_vma_assert_locked(vma); 70863212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 70873212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 70883212b535SSteve Capper return 0; 70893212b535SSteve Capper 70903212b535SSteve Capper pud_clear(pud); 70913212b535SSteve Capper put_page(virt_to_page(ptep)); 7092dc6c9a35SKirill A. Shutemov mm_dec_nr_pmds(mm); 70933212b535SSteve Capper return 1; 70943212b535SSteve Capper } 7095c1991e07SPeter Xu 70969e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 70978d9bfb26SMike Kravetz 709840549ba8SMike Kravetz void hugetlb_vma_lock_read(struct vm_area_struct *vma) 709940549ba8SMike Kravetz { 710040549ba8SMike Kravetz } 710140549ba8SMike Kravetz 710240549ba8SMike Kravetz void hugetlb_vma_unlock_read(struct vm_area_struct *vma) 710340549ba8SMike Kravetz { 710440549ba8SMike Kravetz } 710540549ba8SMike Kravetz 710640549ba8SMike Kravetz void hugetlb_vma_lock_write(struct vm_area_struct *vma) 710740549ba8SMike Kravetz { 710840549ba8SMike Kravetz } 710940549ba8SMike Kravetz 711040549ba8SMike Kravetz void hugetlb_vma_unlock_write(struct vm_area_struct *vma) 711140549ba8SMike Kravetz { 711240549ba8SMike Kravetz } 711340549ba8SMike Kravetz 711440549ba8SMike Kravetz int hugetlb_vma_trylock_write(struct vm_area_struct *vma) 711540549ba8SMike Kravetz { 711640549ba8SMike Kravetz return 1; 711740549ba8SMike Kravetz } 711840549ba8SMike Kravetz 711940549ba8SMike Kravetz void hugetlb_vma_assert_locked(struct vm_area_struct *vma) 712040549ba8SMike Kravetz { 712140549ba8SMike Kravetz } 712240549ba8SMike Kravetz 71238d9bfb26SMike Kravetz void hugetlb_vma_lock_release(struct kref *kref) 71248d9bfb26SMike Kravetz { 71258d9bfb26SMike Kravetz } 71268d9bfb26SMike Kravetz 7127ecfbd733SMike Kravetz static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma) 7128ecfbd733SMike Kravetz { 7129ecfbd733SMike Kravetz } 7130ecfbd733SMike Kravetz 71318d9bfb26SMike Kravetz static void hugetlb_vma_lock_free(struct vm_area_struct *vma) 71328d9bfb26SMike Kravetz { 71338d9bfb26SMike Kravetz } 71348d9bfb26SMike Kravetz 71358d9bfb26SMike Kravetz static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma) 71368d9bfb26SMike Kravetz { 71378d9bfb26SMike Kravetz } 71388d9bfb26SMike Kravetz 7139aec44e0fSPeter Xu pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 7140aec44e0fSPeter Xu unsigned long addr, pud_t *pud) 71419e5fc74cSSteve Capper { 71429e5fc74cSSteve Capper return NULL; 71439e5fc74cSSteve Capper } 7144e81f2d22SZhang Zhen 714534ae204fSMike Kravetz int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 71464ddb4d91SMike Kravetz unsigned long addr, pte_t *ptep) 7147e81f2d22SZhang Zhen { 7148e81f2d22SZhang Zhen return 0; 7149e81f2d22SZhang Zhen } 7150017b1660SMike Kravetz 7151017b1660SMike Kravetz void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 7152017b1660SMike Kravetz unsigned long *start, unsigned long *end) 7153017b1660SMike Kravetz { 7154017b1660SMike Kravetz } 7155c1991e07SPeter Xu 7156c1991e07SPeter Xu bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 7157c1991e07SPeter Xu { 7158c1991e07SPeter Xu return false; 7159c1991e07SPeter Xu } 71603212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 71613212b535SSteve Capper 71629e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 7163aec44e0fSPeter Xu pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, 71649e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 71659e5fc74cSSteve Capper { 71669e5fc74cSSteve Capper pgd_t *pgd; 7167c2febafcSKirill A. Shutemov p4d_t *p4d; 71689e5fc74cSSteve Capper pud_t *pud; 71699e5fc74cSSteve Capper pte_t *pte = NULL; 71709e5fc74cSSteve Capper 71719e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 7172f4f0a3d8SKirill A. Shutemov p4d = p4d_alloc(mm, pgd, addr); 7173f4f0a3d8SKirill A. Shutemov if (!p4d) 7174f4f0a3d8SKirill A. Shutemov return NULL; 7175c2febafcSKirill A. Shutemov pud = pud_alloc(mm, p4d, addr); 71769e5fc74cSSteve Capper if (pud) { 71779e5fc74cSSteve Capper if (sz == PUD_SIZE) { 71789e5fc74cSSteve Capper pte = (pte_t *)pud; 71799e5fc74cSSteve Capper } else { 71809e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 7181c1991e07SPeter Xu if (want_pmd_share(vma, addr) && pud_none(*pud)) 7182aec44e0fSPeter Xu pte = huge_pmd_share(mm, vma, addr, pud); 71839e5fc74cSSteve Capper else 71849e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 71859e5fc74cSSteve Capper } 71869e5fc74cSSteve Capper } 71874e666314SMichal Hocko BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); 71889e5fc74cSSteve Capper 71899e5fc74cSSteve Capper return pte; 71909e5fc74cSSteve Capper } 71919e5fc74cSSteve Capper 71929b19df29SPunit Agrawal /* 71939b19df29SPunit Agrawal * huge_pte_offset() - Walk the page table to resolve the hugepage 71949b19df29SPunit Agrawal * entry at address @addr 71959b19df29SPunit Agrawal * 71968ac0b81aSLi Xinhai * Return: Pointer to page table entry (PUD or PMD) for 71978ac0b81aSLi Xinhai * address @addr, or NULL if a !p*d_present() entry is encountered and the 71989b19df29SPunit Agrawal * size @sz doesn't match the hugepage size at this level of the page 71999b19df29SPunit Agrawal * table. 72009b19df29SPunit Agrawal */ 72017868a208SPunit Agrawal pte_t *huge_pte_offset(struct mm_struct *mm, 72027868a208SPunit Agrawal unsigned long addr, unsigned long sz) 72039e5fc74cSSteve Capper { 72049e5fc74cSSteve Capper pgd_t *pgd; 7205c2febafcSKirill A. Shutemov p4d_t *p4d; 72068ac0b81aSLi Xinhai pud_t *pud; 72078ac0b81aSLi Xinhai pmd_t *pmd; 72089e5fc74cSSteve Capper 72099e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 7210c2febafcSKirill A. Shutemov if (!pgd_present(*pgd)) 7211c2febafcSKirill A. Shutemov return NULL; 7212c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, addr); 7213c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 7214c2febafcSKirill A. Shutemov return NULL; 72159b19df29SPunit Agrawal 7216c2febafcSKirill A. Shutemov pud = pud_offset(p4d, addr); 72178ac0b81aSLi Xinhai if (sz == PUD_SIZE) 72188ac0b81aSLi Xinhai /* must be pud huge, non-present or none */ 72199e5fc74cSSteve Capper return (pte_t *)pud; 72208ac0b81aSLi Xinhai if (!pud_present(*pud)) 72218ac0b81aSLi Xinhai return NULL; 72228ac0b81aSLi Xinhai /* must have a valid entry and size to go further */ 72239b19df29SPunit Agrawal 72249e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 72258ac0b81aSLi Xinhai /* must be pmd huge, non-present or none */ 72269e5fc74cSSteve Capper return (pte_t *)pmd; 72279e5fc74cSSteve Capper } 72289e5fc74cSSteve Capper 7229e95a9851SMike Kravetz /* 7230e95a9851SMike Kravetz * Return a mask that can be used to update an address to the last huge 7231e95a9851SMike Kravetz * page in a page table page mapping size. Used to skip non-present 7232e95a9851SMike Kravetz * page table entries when linearly scanning address ranges. Architectures 7233e95a9851SMike Kravetz * with unique huge page to page table relationships can define their own 7234e95a9851SMike Kravetz * version of this routine. 7235e95a9851SMike Kravetz */ 7236e95a9851SMike Kravetz unsigned long hugetlb_mask_last_page(struct hstate *h) 7237e95a9851SMike Kravetz { 7238e95a9851SMike Kravetz unsigned long hp_size = huge_page_size(h); 7239e95a9851SMike Kravetz 7240e95a9851SMike Kravetz if (hp_size == PUD_SIZE) 7241e95a9851SMike Kravetz return P4D_SIZE - PUD_SIZE; 7242e95a9851SMike Kravetz else if (hp_size == PMD_SIZE) 7243e95a9851SMike Kravetz return PUD_SIZE - PMD_SIZE; 7244e95a9851SMike Kravetz else 7245e95a9851SMike Kravetz return 0UL; 7246e95a9851SMike Kravetz } 7247e95a9851SMike Kravetz 7248e95a9851SMike Kravetz #else 7249e95a9851SMike Kravetz 7250e95a9851SMike Kravetz /* See description above. Architectures can provide their own version. */ 7251e95a9851SMike Kravetz __weak unsigned long hugetlb_mask_last_page(struct hstate *h) 7252e95a9851SMike Kravetz { 72534ddb4d91SMike Kravetz #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 72544ddb4d91SMike Kravetz if (huge_page_size(h) == PMD_SIZE) 72554ddb4d91SMike Kravetz return PUD_SIZE - PMD_SIZE; 72564ddb4d91SMike Kravetz #endif 7257e95a9851SMike Kravetz return 0UL; 7258e95a9851SMike Kravetz } 7259e95a9851SMike Kravetz 726061f77edaSNaoya Horiguchi #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 726161f77edaSNaoya Horiguchi 726261f77edaSNaoya Horiguchi /* 726361f77edaSNaoya Horiguchi * These functions are overwritable if your architecture needs its own 726461f77edaSNaoya Horiguchi * behavior. 726561f77edaSNaoya Horiguchi */ 72667ce82f4cSMiaohe Lin int isolate_hugetlb(struct page *page, struct list_head *list) 726731caf665SNaoya Horiguchi { 72687ce82f4cSMiaohe Lin int ret = 0; 7269bcc54222SNaoya Horiguchi 7270db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 72718f251a3dSMike Kravetz if (!PageHeadHuge(page) || 72728f251a3dSMike Kravetz !HPageMigratable(page) || 72730eb2df2bSMuchun Song !get_page_unless_zero(page)) { 72747ce82f4cSMiaohe Lin ret = -EBUSY; 7275bcc54222SNaoya Horiguchi goto unlock; 7276bcc54222SNaoya Horiguchi } 72778f251a3dSMike Kravetz ClearHPageMigratable(page); 727831caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 7279bcc54222SNaoya Horiguchi unlock: 7280db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7281bcc54222SNaoya Horiguchi return ret; 728231caf665SNaoya Horiguchi } 728331caf665SNaoya Horiguchi 7284e591ef7dSNaoya Horiguchi int get_hwpoison_huge_page(struct page *page, bool *hugetlb, bool unpoison) 728525182f05SNaoya Horiguchi { 728625182f05SNaoya Horiguchi int ret = 0; 728725182f05SNaoya Horiguchi 728825182f05SNaoya Horiguchi *hugetlb = false; 728925182f05SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 729025182f05SNaoya Horiguchi if (PageHeadHuge(page)) { 729125182f05SNaoya Horiguchi *hugetlb = true; 7292b283d983SNaoya Horiguchi if (HPageFreed(page)) 7293b283d983SNaoya Horiguchi ret = 0; 7294e591ef7dSNaoya Horiguchi else if (HPageMigratable(page) || unpoison) 729525182f05SNaoya Horiguchi ret = get_page_unless_zero(page); 72960ed950d1SNaoya Horiguchi else 72970ed950d1SNaoya Horiguchi ret = -EBUSY; 729825182f05SNaoya Horiguchi } 729925182f05SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 730025182f05SNaoya Horiguchi return ret; 730125182f05SNaoya Horiguchi } 730225182f05SNaoya Horiguchi 7303e591ef7dSNaoya Horiguchi int get_huge_page_for_hwpoison(unsigned long pfn, int flags, 7304e591ef7dSNaoya Horiguchi bool *migratable_cleared) 7305405ce051SNaoya Horiguchi { 7306405ce051SNaoya Horiguchi int ret; 7307405ce051SNaoya Horiguchi 7308405ce051SNaoya Horiguchi spin_lock_irq(&hugetlb_lock); 7309e591ef7dSNaoya Horiguchi ret = __get_huge_page_for_hwpoison(pfn, flags, migratable_cleared); 7310405ce051SNaoya Horiguchi spin_unlock_irq(&hugetlb_lock); 7311405ce051SNaoya Horiguchi return ret; 7312405ce051SNaoya Horiguchi } 7313405ce051SNaoya Horiguchi 731431caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 731531caf665SNaoya Horiguchi { 7316db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 73178f251a3dSMike Kravetz SetHPageMigratable(page); 731831caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 7319db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 732031caf665SNaoya Horiguchi put_page(page); 732131caf665SNaoya Horiguchi } 7322ab5ac90aSMichal Hocko 7323ab5ac90aSMichal Hocko void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) 7324ab5ac90aSMichal Hocko { 7325ab5ac90aSMichal Hocko struct hstate *h = page_hstate(oldpage); 7326ab5ac90aSMichal Hocko 732729f39430SSidhartha Kumar hugetlb_cgroup_migrate(page_folio(oldpage), page_folio(newpage)); 7328ab5ac90aSMichal Hocko set_page_owner_migrate_reason(newpage, reason); 7329ab5ac90aSMichal Hocko 7330ab5ac90aSMichal Hocko /* 7331ab5ac90aSMichal Hocko * transfer temporary state of the new huge page. This is 7332ab5ac90aSMichal Hocko * reverse to other transitions because the newpage is going to 7333ab5ac90aSMichal Hocko * be final while the old one will be freed so it takes over 7334ab5ac90aSMichal Hocko * the temporary status. 7335ab5ac90aSMichal Hocko * 7336ab5ac90aSMichal Hocko * Also note that we have to transfer the per-node surplus state 7337ab5ac90aSMichal Hocko * here as well otherwise the global surplus count will not match 7338ab5ac90aSMichal Hocko * the per-node's. 7339ab5ac90aSMichal Hocko */ 73409157c311SMike Kravetz if (HPageTemporary(newpage)) { 7341ab5ac90aSMichal Hocko int old_nid = page_to_nid(oldpage); 7342ab5ac90aSMichal Hocko int new_nid = page_to_nid(newpage); 7343ab5ac90aSMichal Hocko 73449157c311SMike Kravetz SetHPageTemporary(oldpage); 73459157c311SMike Kravetz ClearHPageTemporary(newpage); 7346ab5ac90aSMichal Hocko 73475af1ab1dSMiaohe Lin /* 73485af1ab1dSMiaohe Lin * There is no need to transfer the per-node surplus state 73495af1ab1dSMiaohe Lin * when we do not cross the node. 73505af1ab1dSMiaohe Lin */ 73515af1ab1dSMiaohe Lin if (new_nid == old_nid) 73525af1ab1dSMiaohe Lin return; 7353db71ef79SMike Kravetz spin_lock_irq(&hugetlb_lock); 7354ab5ac90aSMichal Hocko if (h->surplus_huge_pages_node[old_nid]) { 7355ab5ac90aSMichal Hocko h->surplus_huge_pages_node[old_nid]--; 7356ab5ac90aSMichal Hocko h->surplus_huge_pages_node[new_nid]++; 7357ab5ac90aSMichal Hocko } 7358db71ef79SMike Kravetz spin_unlock_irq(&hugetlb_lock); 7359ab5ac90aSMichal Hocko } 7360ab5ac90aSMichal Hocko } 7361cf11e85fSRoman Gushchin 73626dfeaff9SPeter Xu /* 73636dfeaff9SPeter Xu * This function will unconditionally remove all the shared pmd pgtable entries 73646dfeaff9SPeter Xu * within the specific vma for a hugetlbfs memory range. 73656dfeaff9SPeter Xu */ 73666dfeaff9SPeter Xu void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) 73676dfeaff9SPeter Xu { 73686dfeaff9SPeter Xu struct hstate *h = hstate_vma(vma); 73696dfeaff9SPeter Xu unsigned long sz = huge_page_size(h); 73706dfeaff9SPeter Xu struct mm_struct *mm = vma->vm_mm; 73716dfeaff9SPeter Xu struct mmu_notifier_range range; 73726dfeaff9SPeter Xu unsigned long address, start, end; 73736dfeaff9SPeter Xu spinlock_t *ptl; 73746dfeaff9SPeter Xu pte_t *ptep; 73756dfeaff9SPeter Xu 73766dfeaff9SPeter Xu if (!(vma->vm_flags & VM_MAYSHARE)) 73776dfeaff9SPeter Xu return; 73786dfeaff9SPeter Xu 73796dfeaff9SPeter Xu start = ALIGN(vma->vm_start, PUD_SIZE); 73806dfeaff9SPeter Xu end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 73816dfeaff9SPeter Xu 73826dfeaff9SPeter Xu if (start >= end) 73836dfeaff9SPeter Xu return; 73846dfeaff9SPeter Xu 73859c8bbfacSBaolin Wang flush_cache_range(vma, start, end); 73866dfeaff9SPeter Xu /* 73876dfeaff9SPeter Xu * No need to call adjust_range_if_pmd_sharing_possible(), because 73886dfeaff9SPeter Xu * we have already done the PUD_SIZE alignment. 73896dfeaff9SPeter Xu */ 73906dfeaff9SPeter Xu mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, 73916dfeaff9SPeter Xu start, end); 73926dfeaff9SPeter Xu mmu_notifier_invalidate_range_start(&range); 739340549ba8SMike Kravetz hugetlb_vma_lock_write(vma); 73946dfeaff9SPeter Xu i_mmap_lock_write(vma->vm_file->f_mapping); 73956dfeaff9SPeter Xu for (address = start; address < end; address += PUD_SIZE) { 73966dfeaff9SPeter Xu ptep = huge_pte_offset(mm, address, sz); 73976dfeaff9SPeter Xu if (!ptep) 73986dfeaff9SPeter Xu continue; 73996dfeaff9SPeter Xu ptl = huge_pte_lock(h, mm, ptep); 74004ddb4d91SMike Kravetz huge_pmd_unshare(mm, vma, address, ptep); 74016dfeaff9SPeter Xu spin_unlock(ptl); 74026dfeaff9SPeter Xu } 74036dfeaff9SPeter Xu flush_hugetlb_tlb_range(vma, start, end); 74046dfeaff9SPeter Xu i_mmap_unlock_write(vma->vm_file->f_mapping); 740540549ba8SMike Kravetz hugetlb_vma_unlock_write(vma); 74066dfeaff9SPeter Xu /* 74076dfeaff9SPeter Xu * No need to call mmu_notifier_invalidate_range(), see 7408ee65728eSMike Rapoport * Documentation/mm/mmu_notifier.rst. 74096dfeaff9SPeter Xu */ 74106dfeaff9SPeter Xu mmu_notifier_invalidate_range_end(&range); 74116dfeaff9SPeter Xu } 74126dfeaff9SPeter Xu 7413cf11e85fSRoman Gushchin #ifdef CONFIG_CMA 7414cf11e85fSRoman Gushchin static bool cma_reserve_called __initdata; 7415cf11e85fSRoman Gushchin 7416cf11e85fSRoman Gushchin static int __init cmdline_parse_hugetlb_cma(char *p) 7417cf11e85fSRoman Gushchin { 741838e719abSBaolin Wang int nid, count = 0; 741938e719abSBaolin Wang unsigned long tmp; 742038e719abSBaolin Wang char *s = p; 742138e719abSBaolin Wang 742238e719abSBaolin Wang while (*s) { 742338e719abSBaolin Wang if (sscanf(s, "%lu%n", &tmp, &count) != 1) 742438e719abSBaolin Wang break; 742538e719abSBaolin Wang 742638e719abSBaolin Wang if (s[count] == ':') { 7427f9317f77SMike Kravetz if (tmp >= MAX_NUMNODES) 742838e719abSBaolin Wang break; 7429f9317f77SMike Kravetz nid = array_index_nospec(tmp, MAX_NUMNODES); 743038e719abSBaolin Wang 743138e719abSBaolin Wang s += count + 1; 743238e719abSBaolin Wang tmp = memparse(s, &s); 743338e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = tmp; 743438e719abSBaolin Wang hugetlb_cma_size += tmp; 743538e719abSBaolin Wang 743638e719abSBaolin Wang /* 743738e719abSBaolin Wang * Skip the separator if have one, otherwise 743838e719abSBaolin Wang * break the parsing. 743938e719abSBaolin Wang */ 744038e719abSBaolin Wang if (*s == ',') 744138e719abSBaolin Wang s++; 744238e719abSBaolin Wang else 744338e719abSBaolin Wang break; 744438e719abSBaolin Wang } else { 7445cf11e85fSRoman Gushchin hugetlb_cma_size = memparse(p, &p); 744638e719abSBaolin Wang break; 744738e719abSBaolin Wang } 744838e719abSBaolin Wang } 744938e719abSBaolin Wang 7450cf11e85fSRoman Gushchin return 0; 7451cf11e85fSRoman Gushchin } 7452cf11e85fSRoman Gushchin 7453cf11e85fSRoman Gushchin early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); 7454cf11e85fSRoman Gushchin 7455cf11e85fSRoman Gushchin void __init hugetlb_cma_reserve(int order) 7456cf11e85fSRoman Gushchin { 7457cf11e85fSRoman Gushchin unsigned long size, reserved, per_node; 745838e719abSBaolin Wang bool node_specific_cma_alloc = false; 7459cf11e85fSRoman Gushchin int nid; 7460cf11e85fSRoman Gushchin 7461cf11e85fSRoman Gushchin cma_reserve_called = true; 7462cf11e85fSRoman Gushchin 7463cf11e85fSRoman Gushchin if (!hugetlb_cma_size) 7464cf11e85fSRoman Gushchin return; 7465cf11e85fSRoman Gushchin 746638e719abSBaolin Wang for (nid = 0; nid < MAX_NUMNODES; nid++) { 746738e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 746838e719abSBaolin Wang continue; 746938e719abSBaolin Wang 747030a51400SPeng Liu if (!node_online(nid)) { 747138e719abSBaolin Wang pr_warn("hugetlb_cma: invalid node %d specified\n", nid); 747238e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 747338e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 747438e719abSBaolin Wang continue; 747538e719abSBaolin Wang } 747638e719abSBaolin Wang 747738e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) { 747838e719abSBaolin Wang pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n", 747938e719abSBaolin Wang nid, (PAGE_SIZE << order) / SZ_1M); 748038e719abSBaolin Wang hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 748138e719abSBaolin Wang hugetlb_cma_size_in_node[nid] = 0; 748238e719abSBaolin Wang } else { 748338e719abSBaolin Wang node_specific_cma_alloc = true; 748438e719abSBaolin Wang } 748538e719abSBaolin Wang } 748638e719abSBaolin Wang 748738e719abSBaolin Wang /* Validate the CMA size again in case some invalid nodes specified. */ 748838e719abSBaolin Wang if (!hugetlb_cma_size) 748938e719abSBaolin Wang return; 749038e719abSBaolin Wang 7491cf11e85fSRoman Gushchin if (hugetlb_cma_size < (PAGE_SIZE << order)) { 7492cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", 7493cf11e85fSRoman Gushchin (PAGE_SIZE << order) / SZ_1M); 7494a01f4390SMike Kravetz hugetlb_cma_size = 0; 7495cf11e85fSRoman Gushchin return; 7496cf11e85fSRoman Gushchin } 7497cf11e85fSRoman Gushchin 749838e719abSBaolin Wang if (!node_specific_cma_alloc) { 7499cf11e85fSRoman Gushchin /* 7500cf11e85fSRoman Gushchin * If 3 GB area is requested on a machine with 4 numa nodes, 7501cf11e85fSRoman Gushchin * let's allocate 1 GB on first three nodes and ignore the last one. 7502cf11e85fSRoman Gushchin */ 7503cf11e85fSRoman Gushchin per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); 7504cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", 7505cf11e85fSRoman Gushchin hugetlb_cma_size / SZ_1M, per_node / SZ_1M); 750638e719abSBaolin Wang } 7507cf11e85fSRoman Gushchin 7508cf11e85fSRoman Gushchin reserved = 0; 750930a51400SPeng Liu for_each_online_node(nid) { 7510cf11e85fSRoman Gushchin int res; 75112281f797SBarry Song char name[CMA_MAX_NAME]; 7512cf11e85fSRoman Gushchin 751338e719abSBaolin Wang if (node_specific_cma_alloc) { 751438e719abSBaolin Wang if (hugetlb_cma_size_in_node[nid] == 0) 751538e719abSBaolin Wang continue; 751638e719abSBaolin Wang 751738e719abSBaolin Wang size = hugetlb_cma_size_in_node[nid]; 751838e719abSBaolin Wang } else { 7519cf11e85fSRoman Gushchin size = min(per_node, hugetlb_cma_size - reserved); 752038e719abSBaolin Wang } 752138e719abSBaolin Wang 7522cf11e85fSRoman Gushchin size = round_up(size, PAGE_SIZE << order); 7523cf11e85fSRoman Gushchin 75242281f797SBarry Song snprintf(name, sizeof(name), "hugetlb%d", nid); 7525a01f4390SMike Kravetz /* 7526a01f4390SMike Kravetz * Note that 'order per bit' is based on smallest size that 7527a01f4390SMike Kravetz * may be returned to CMA allocator in the case of 7528a01f4390SMike Kravetz * huge page demotion. 7529a01f4390SMike Kravetz */ 7530a01f4390SMike Kravetz res = cma_declare_contiguous_nid(0, size, 0, 7531a01f4390SMike Kravetz PAGE_SIZE << HUGETLB_PAGE_ORDER, 753229d0f41dSBarry Song 0, false, name, 7533cf11e85fSRoman Gushchin &hugetlb_cma[nid], nid); 7534cf11e85fSRoman Gushchin if (res) { 7535cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: reservation failed: err %d, node %d", 7536cf11e85fSRoman Gushchin res, nid); 7537cf11e85fSRoman Gushchin continue; 7538cf11e85fSRoman Gushchin } 7539cf11e85fSRoman Gushchin 7540cf11e85fSRoman Gushchin reserved += size; 7541cf11e85fSRoman Gushchin pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", 7542cf11e85fSRoman Gushchin size / SZ_1M, nid); 7543cf11e85fSRoman Gushchin 7544cf11e85fSRoman Gushchin if (reserved >= hugetlb_cma_size) 7545cf11e85fSRoman Gushchin break; 7546cf11e85fSRoman Gushchin } 7547a01f4390SMike Kravetz 7548a01f4390SMike Kravetz if (!reserved) 7549a01f4390SMike Kravetz /* 7550a01f4390SMike Kravetz * hugetlb_cma_size is used to determine if allocations from 7551a01f4390SMike Kravetz * cma are possible. Set to zero if no cma regions are set up. 7552a01f4390SMike Kravetz */ 7553a01f4390SMike Kravetz hugetlb_cma_size = 0; 7554cf11e85fSRoman Gushchin } 7555cf11e85fSRoman Gushchin 7556263b8998SMiaohe Lin static void __init hugetlb_cma_check(void) 7557cf11e85fSRoman Gushchin { 7558cf11e85fSRoman Gushchin if (!hugetlb_cma_size || cma_reserve_called) 7559cf11e85fSRoman Gushchin return; 7560cf11e85fSRoman Gushchin 7561cf11e85fSRoman Gushchin pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); 7562cf11e85fSRoman Gushchin } 7563cf11e85fSRoman Gushchin 7564cf11e85fSRoman Gushchin #endif /* CONFIG_CMA */ 7565